OpenCloudOS-Kernel/drivers/gpu/drm/radeon/radeon_atombios.c

2953 lines
95 KiB
C
Raw Normal View History

drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
/*
* Copyright 2007-8 Advanced Micro Devices, Inc.
* Copyright 2008 Red Hat Inc.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
* OTHER DEALINGS IN THE SOFTWARE.
*
* Authors: Dave Airlie
* Alex Deucher
*/
#include "drmP.h"
#include "radeon_drm.h"
#include "radeon.h"
#include "atom.h"
#include "atom-bits.h"
/* from radeon_encoder.c */
extern uint32_t
radeon_get_encoder_enum(struct drm_device *dev, uint32_t supported_device,
uint8_t dac);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
extern void radeon_link_encoder_connector(struct drm_device *dev);
extern void
radeon_add_atom_encoder(struct drm_device *dev, uint32_t encoder_enum,
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
uint32_t supported_device);
/* from radeon_connector.c */
extern void
radeon_add_atom_connector(struct drm_device *dev,
uint32_t connector_id,
uint32_t supported_device,
int connector_type,
struct radeon_i2c_bus_rec *i2c_bus,
uint32_t igp_lane_info,
uint16_t connector_object_id,
struct radeon_hpd *hpd,
struct radeon_router *router);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
/* from radeon_legacy_encoder.c */
extern void
radeon_add_legacy_encoder(struct drm_device *dev, uint32_t encoder_enum,
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
uint32_t supported_device);
union atom_supported_devices {
struct _ATOM_SUPPORTED_DEVICES_INFO info;
struct _ATOM_SUPPORTED_DEVICES_INFO_2 info_2;
struct _ATOM_SUPPORTED_DEVICES_INFO_2d1 info_2d1;
};
static inline struct radeon_i2c_bus_rec radeon_lookup_i2c_gpio(struct radeon_device *rdev,
uint8_t id)
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
{
struct atom_context *ctx = rdev->mode_info.atom_context;
ATOM_GPIO_I2C_ASSIGMENT *gpio;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
struct radeon_i2c_bus_rec i2c;
int index = GetIndexIntoMasterTable(DATA, GPIO_I2C_Info);
struct _ATOM_GPIO_I2C_INFO *i2c_info;
uint16_t data_offset, size;
int i, num_indices;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
memset(&i2c, 0, sizeof(struct radeon_i2c_bus_rec));
i2c.valid = false;
if (atom_parse_data_header(ctx, index, &size, NULL, NULL, &data_offset)) {
i2c_info = (struct _ATOM_GPIO_I2C_INFO *)(ctx->bios + data_offset);
num_indices = (size - sizeof(ATOM_COMMON_TABLE_HEADER)) /
sizeof(ATOM_GPIO_I2C_ASSIGMENT);
for (i = 0; i < num_indices; i++) {
gpio = &i2c_info->asGPIO_Info[i];
/* some evergreen boards have bad data for this entry */
if (ASIC_IS_DCE4(rdev)) {
if ((i == 7) &&
(gpio->usClkMaskRegisterIndex == 0x1936) &&
(gpio->sucI2cId.ucAccess == 0)) {
gpio->sucI2cId.ucAccess = 0x97;
gpio->ucDataMaskShift = 8;
gpio->ucDataEnShift = 8;
gpio->ucDataY_Shift = 8;
gpio->ucDataA_Shift = 8;
}
}
/* some DCE3 boards have bad data for this entry */
if (ASIC_IS_DCE3(rdev)) {
if ((i == 4) &&
(gpio->usClkMaskRegisterIndex == 0x1fda) &&
(gpio->sucI2cId.ucAccess == 0x94))
gpio->sucI2cId.ucAccess = 0x14;
}
if (gpio->sucI2cId.ucAccess == id) {
i2c.mask_clk_reg = le16_to_cpu(gpio->usClkMaskRegisterIndex) * 4;
i2c.mask_data_reg = le16_to_cpu(gpio->usDataMaskRegisterIndex) * 4;
i2c.en_clk_reg = le16_to_cpu(gpio->usClkEnRegisterIndex) * 4;
i2c.en_data_reg = le16_to_cpu(gpio->usDataEnRegisterIndex) * 4;
i2c.y_clk_reg = le16_to_cpu(gpio->usClkY_RegisterIndex) * 4;
i2c.y_data_reg = le16_to_cpu(gpio->usDataY_RegisterIndex) * 4;
i2c.a_clk_reg = le16_to_cpu(gpio->usClkA_RegisterIndex) * 4;
i2c.a_data_reg = le16_to_cpu(gpio->usDataA_RegisterIndex) * 4;
i2c.mask_clk_mask = (1 << gpio->ucClkMaskShift);
i2c.mask_data_mask = (1 << gpio->ucDataMaskShift);
i2c.en_clk_mask = (1 << gpio->ucClkEnShift);
i2c.en_data_mask = (1 << gpio->ucDataEnShift);
i2c.y_clk_mask = (1 << gpio->ucClkY_Shift);
i2c.y_data_mask = (1 << gpio->ucDataY_Shift);
i2c.a_clk_mask = (1 << gpio->ucClkA_Shift);
i2c.a_data_mask = (1 << gpio->ucDataA_Shift);
if (gpio->sucI2cId.sbfAccess.bfHW_Capable)
i2c.hw_capable = true;
else
i2c.hw_capable = false;
if (gpio->sucI2cId.ucAccess == 0xa0)
i2c.mm_i2c = true;
else
i2c.mm_i2c = false;
i2c.i2c_id = gpio->sucI2cId.ucAccess;
if (i2c.mask_clk_reg)
i2c.valid = true;
break;
}
}
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
return i2c;
}
void radeon_atombios_i2c_init(struct radeon_device *rdev)
{
struct atom_context *ctx = rdev->mode_info.atom_context;
ATOM_GPIO_I2C_ASSIGMENT *gpio;
struct radeon_i2c_bus_rec i2c;
int index = GetIndexIntoMasterTable(DATA, GPIO_I2C_Info);
struct _ATOM_GPIO_I2C_INFO *i2c_info;
uint16_t data_offset, size;
int i, num_indices;
char stmp[32];
memset(&i2c, 0, sizeof(struct radeon_i2c_bus_rec));
if (atom_parse_data_header(ctx, index, &size, NULL, NULL, &data_offset)) {
i2c_info = (struct _ATOM_GPIO_I2C_INFO *)(ctx->bios + data_offset);
num_indices = (size - sizeof(ATOM_COMMON_TABLE_HEADER)) /
sizeof(ATOM_GPIO_I2C_ASSIGMENT);
for (i = 0; i < num_indices; i++) {
gpio = &i2c_info->asGPIO_Info[i];
i2c.valid = false;
/* some evergreen boards have bad data for this entry */
if (ASIC_IS_DCE4(rdev)) {
if ((i == 7) &&
(gpio->usClkMaskRegisterIndex == 0x1936) &&
(gpio->sucI2cId.ucAccess == 0)) {
gpio->sucI2cId.ucAccess = 0x97;
gpio->ucDataMaskShift = 8;
gpio->ucDataEnShift = 8;
gpio->ucDataY_Shift = 8;
gpio->ucDataA_Shift = 8;
}
}
/* some DCE3 boards have bad data for this entry */
if (ASIC_IS_DCE3(rdev)) {
if ((i == 4) &&
(gpio->usClkMaskRegisterIndex == 0x1fda) &&
(gpio->sucI2cId.ucAccess == 0x94))
gpio->sucI2cId.ucAccess = 0x14;
}
i2c.mask_clk_reg = le16_to_cpu(gpio->usClkMaskRegisterIndex) * 4;
i2c.mask_data_reg = le16_to_cpu(gpio->usDataMaskRegisterIndex) * 4;
i2c.en_clk_reg = le16_to_cpu(gpio->usClkEnRegisterIndex) * 4;
i2c.en_data_reg = le16_to_cpu(gpio->usDataEnRegisterIndex) * 4;
i2c.y_clk_reg = le16_to_cpu(gpio->usClkY_RegisterIndex) * 4;
i2c.y_data_reg = le16_to_cpu(gpio->usDataY_RegisterIndex) * 4;
i2c.a_clk_reg = le16_to_cpu(gpio->usClkA_RegisterIndex) * 4;
i2c.a_data_reg = le16_to_cpu(gpio->usDataA_RegisterIndex) * 4;
i2c.mask_clk_mask = (1 << gpio->ucClkMaskShift);
i2c.mask_data_mask = (1 << gpio->ucDataMaskShift);
i2c.en_clk_mask = (1 << gpio->ucClkEnShift);
i2c.en_data_mask = (1 << gpio->ucDataEnShift);
i2c.y_clk_mask = (1 << gpio->ucClkY_Shift);
i2c.y_data_mask = (1 << gpio->ucDataY_Shift);
i2c.a_clk_mask = (1 << gpio->ucClkA_Shift);
i2c.a_data_mask = (1 << gpio->ucDataA_Shift);
if (gpio->sucI2cId.sbfAccess.bfHW_Capable)
i2c.hw_capable = true;
else
i2c.hw_capable = false;
if (gpio->sucI2cId.ucAccess == 0xa0)
i2c.mm_i2c = true;
else
i2c.mm_i2c = false;
i2c.i2c_id = gpio->sucI2cId.ucAccess;
if (i2c.mask_clk_reg) {
i2c.valid = true;
sprintf(stmp, "0x%x", i2c.i2c_id);
rdev->i2c_bus[i] = radeon_i2c_create(rdev->ddev, &i2c, stmp);
}
}
}
}
static inline struct radeon_gpio_rec radeon_lookup_gpio(struct radeon_device *rdev,
u8 id)
{
struct atom_context *ctx = rdev->mode_info.atom_context;
struct radeon_gpio_rec gpio;
int index = GetIndexIntoMasterTable(DATA, GPIO_Pin_LUT);
struct _ATOM_GPIO_PIN_LUT *gpio_info;
ATOM_GPIO_PIN_ASSIGNMENT *pin;
u16 data_offset, size;
int i, num_indices;
memset(&gpio, 0, sizeof(struct radeon_gpio_rec));
gpio.valid = false;
if (atom_parse_data_header(ctx, index, &size, NULL, NULL, &data_offset)) {
gpio_info = (struct _ATOM_GPIO_PIN_LUT *)(ctx->bios + data_offset);
num_indices = (size - sizeof(ATOM_COMMON_TABLE_HEADER)) /
sizeof(ATOM_GPIO_PIN_ASSIGNMENT);
for (i = 0; i < num_indices; i++) {
pin = &gpio_info->asGPIO_Pin[i];
if (id == pin->ucGPIO_ID) {
gpio.id = pin->ucGPIO_ID;
gpio.reg = pin->usGpioPin_AIndex * 4;
gpio.mask = (1 << pin->ucGpioPinBitShift);
gpio.valid = true;
break;
}
}
}
return gpio;
}
static struct radeon_hpd radeon_atom_get_hpd_info_from_gpio(struct radeon_device *rdev,
struct radeon_gpio_rec *gpio)
{
struct radeon_hpd hpd;
u32 reg;
memset(&hpd, 0, sizeof(struct radeon_hpd));
if (ASIC_IS_DCE4(rdev))
reg = EVERGREEN_DC_GPIO_HPD_A;
else
reg = AVIVO_DC_GPIO_HPD_A;
hpd.gpio = *gpio;
if (gpio->reg == reg) {
switch(gpio->mask) {
case (1 << 0):
hpd.hpd = RADEON_HPD_1;
break;
case (1 << 8):
hpd.hpd = RADEON_HPD_2;
break;
case (1 << 16):
hpd.hpd = RADEON_HPD_3;
break;
case (1 << 24):
hpd.hpd = RADEON_HPD_4;
break;
case (1 << 26):
hpd.hpd = RADEON_HPD_5;
break;
case (1 << 28):
hpd.hpd = RADEON_HPD_6;
break;
default:
hpd.hpd = RADEON_HPD_NONE;
break;
}
} else
hpd.hpd = RADEON_HPD_NONE;
return hpd;
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
static bool radeon_atom_apply_quirks(struct drm_device *dev,
uint32_t supported_device,
int *connector_type,
struct radeon_i2c_bus_rec *i2c_bus,
uint16_t *line_mux,
struct radeon_hpd *hpd)
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
{
/* Asus M2A-VM HDMI board lists the DVI port as HDMI */
if ((dev->pdev->device == 0x791e) &&
(dev->pdev->subsystem_vendor == 0x1043) &&
(dev->pdev->subsystem_device == 0x826d)) {
if ((*connector_type == DRM_MODE_CONNECTOR_HDMIA) &&
(supported_device == ATOM_DEVICE_DFP3_SUPPORT))
*connector_type = DRM_MODE_CONNECTOR_DVID;
}
/* Asrock RS600 board lists the DVI port as HDMI */
if ((dev->pdev->device == 0x7941) &&
(dev->pdev->subsystem_vendor == 0x1849) &&
(dev->pdev->subsystem_device == 0x7941)) {
if ((*connector_type == DRM_MODE_CONNECTOR_HDMIA) &&
(supported_device == ATOM_DEVICE_DFP3_SUPPORT))
*connector_type = DRM_MODE_CONNECTOR_DVID;
}
/* MSI K9A2GM V2/V3 board has no HDMI or DVI */
if ((dev->pdev->device == 0x796e) &&
(dev->pdev->subsystem_vendor == 0x1462) &&
(dev->pdev->subsystem_device == 0x7302)) {
if ((supported_device == ATOM_DEVICE_DFP2_SUPPORT) ||
(supported_device == ATOM_DEVICE_DFP3_SUPPORT))
return false;
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
/* a-bit f-i90hd - ciaranm on #radeonhd - this board has no DVI */
if ((dev->pdev->device == 0x7941) &&
(dev->pdev->subsystem_vendor == 0x147b) &&
(dev->pdev->subsystem_device == 0x2412)) {
if (*connector_type == DRM_MODE_CONNECTOR_DVII)
return false;
}
/* Falcon NW laptop lists vga ddc line for LVDS */
if ((dev->pdev->device == 0x5653) &&
(dev->pdev->subsystem_vendor == 0x1462) &&
(dev->pdev->subsystem_device == 0x0291)) {
if (*connector_type == DRM_MODE_CONNECTOR_LVDS) {
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
i2c_bus->valid = false;
*line_mux = 53;
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
/* HIS X1300 is DVI+VGA, not DVI+DVI */
if ((dev->pdev->device == 0x7146) &&
(dev->pdev->subsystem_vendor == 0x17af) &&
(dev->pdev->subsystem_device == 0x2058)) {
if (supported_device == ATOM_DEVICE_DFP1_SUPPORT)
return false;
}
/* Gigabyte X1300 is DVI+VGA, not DVI+DVI */
if ((dev->pdev->device == 0x7142) &&
(dev->pdev->subsystem_vendor == 0x1458) &&
(dev->pdev->subsystem_device == 0x2134)) {
if (supported_device == ATOM_DEVICE_DFP1_SUPPORT)
return false;
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
/* Funky macbooks */
if ((dev->pdev->device == 0x71C5) &&
(dev->pdev->subsystem_vendor == 0x106b) &&
(dev->pdev->subsystem_device == 0x0080)) {
if ((supported_device == ATOM_DEVICE_CRT1_SUPPORT) ||
(supported_device == ATOM_DEVICE_DFP2_SUPPORT))
return false;
if (supported_device == ATOM_DEVICE_CRT2_SUPPORT)
*line_mux = 0x90;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
/* mac rv630 */
if ((dev->pdev->device == 0x9588) &&
(dev->pdev->subsystem_vendor == 0x106b) &&
(dev->pdev->subsystem_device == 0x00a6)) {
if ((supported_device == ATOM_DEVICE_TV1_SUPPORT) &&
(*connector_type == DRM_MODE_CONNECTOR_DVII)) {
*connector_type = DRM_MODE_CONNECTOR_9PinDIN;
*line_mux = CONNECTOR_7PIN_DIN_ENUM_ID1;
}
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
/* ASUS HD 3600 XT board lists the DVI port as HDMI */
if ((dev->pdev->device == 0x9598) &&
(dev->pdev->subsystem_vendor == 0x1043) &&
(dev->pdev->subsystem_device == 0x01da)) {
if (*connector_type == DRM_MODE_CONNECTOR_HDMIA) {
*connector_type = DRM_MODE_CONNECTOR_DVII;
}
}
/* ASUS HD 3600 board lists the DVI port as HDMI */
if ((dev->pdev->device == 0x9598) &&
(dev->pdev->subsystem_vendor == 0x1043) &&
(dev->pdev->subsystem_device == 0x01e4)) {
if (*connector_type == DRM_MODE_CONNECTOR_HDMIA) {
*connector_type = DRM_MODE_CONNECTOR_DVII;
}
}
/* ASUS HD 3450 board lists the DVI port as HDMI */
if ((dev->pdev->device == 0x95C5) &&
(dev->pdev->subsystem_vendor == 0x1043) &&
(dev->pdev->subsystem_device == 0x01e2)) {
if (*connector_type == DRM_MODE_CONNECTOR_HDMIA) {
*connector_type = DRM_MODE_CONNECTOR_DVII;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
}
/* some BIOSes seem to report DAC on HDMI - usually this is a board with
* HDMI + VGA reporting as HDMI
*/
if (*connector_type == DRM_MODE_CONNECTOR_HDMIA) {
if (supported_device & (ATOM_DEVICE_CRT_SUPPORT)) {
*connector_type = DRM_MODE_CONNECTOR_VGA;
*line_mux = 0;
}
}
/* Acer laptop (Acer TravelMate 5730G) has an HDMI port
* on the laptop and a DVI port on the docking station and
* both share the same encoder, hpd pin, and ddc line.
* So while the bios table is technically correct,
* we drop the DVI port here since xrandr has no concept of
* encoders and will try and drive both connectors
* with different crtcs which isn't possible on the hardware
* side and leaves no crtcs for LVDS or VGA.
*/
if ((dev->pdev->device == 0x95c4) &&
(dev->pdev->subsystem_vendor == 0x1025) &&
(dev->pdev->subsystem_device == 0x013c)) {
if ((*connector_type == DRM_MODE_CONNECTOR_DVII) &&
(supported_device == ATOM_DEVICE_DFP1_SUPPORT)) {
/* actually it's a DVI-D port not DVI-I */
*connector_type = DRM_MODE_CONNECTOR_DVID;
return false;
}
}
/* XFX Pine Group device rv730 reports no VGA DDC lines
* even though they are wired up to record 0x93
*/
if ((dev->pdev->device == 0x9498) &&
(dev->pdev->subsystem_vendor == 0x1682) &&
(dev->pdev->subsystem_device == 0x2452)) {
struct radeon_device *rdev = dev->dev_private;
*i2c_bus = radeon_lookup_i2c_gpio(rdev, 0x93);
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
return true;
}
const int supported_devices_connector_convert[] = {
DRM_MODE_CONNECTOR_Unknown,
DRM_MODE_CONNECTOR_VGA,
DRM_MODE_CONNECTOR_DVII,
DRM_MODE_CONNECTOR_DVID,
DRM_MODE_CONNECTOR_DVIA,
DRM_MODE_CONNECTOR_SVIDEO,
DRM_MODE_CONNECTOR_Composite,
DRM_MODE_CONNECTOR_LVDS,
DRM_MODE_CONNECTOR_Unknown,
DRM_MODE_CONNECTOR_Unknown,
DRM_MODE_CONNECTOR_HDMIA,
DRM_MODE_CONNECTOR_HDMIB,
DRM_MODE_CONNECTOR_Unknown,
DRM_MODE_CONNECTOR_Unknown,
DRM_MODE_CONNECTOR_9PinDIN,
DRM_MODE_CONNECTOR_DisplayPort
};
const uint16_t supported_devices_connector_object_id_convert[] = {
CONNECTOR_OBJECT_ID_NONE,
CONNECTOR_OBJECT_ID_VGA,
CONNECTOR_OBJECT_ID_DUAL_LINK_DVI_I, /* not all boards support DL */
CONNECTOR_OBJECT_ID_DUAL_LINK_DVI_D, /* not all boards support DL */
CONNECTOR_OBJECT_ID_VGA, /* technically DVI-A */
CONNECTOR_OBJECT_ID_COMPOSITE,
CONNECTOR_OBJECT_ID_SVIDEO,
CONNECTOR_OBJECT_ID_LVDS,
CONNECTOR_OBJECT_ID_9PIN_DIN,
CONNECTOR_OBJECT_ID_9PIN_DIN,
CONNECTOR_OBJECT_ID_DISPLAYPORT,
CONNECTOR_OBJECT_ID_HDMI_TYPE_A,
CONNECTOR_OBJECT_ID_HDMI_TYPE_B,
CONNECTOR_OBJECT_ID_SVIDEO
};
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
const int object_connector_convert[] = {
DRM_MODE_CONNECTOR_Unknown,
DRM_MODE_CONNECTOR_DVII,
DRM_MODE_CONNECTOR_DVII,
DRM_MODE_CONNECTOR_DVID,
DRM_MODE_CONNECTOR_DVID,
DRM_MODE_CONNECTOR_VGA,
DRM_MODE_CONNECTOR_Composite,
DRM_MODE_CONNECTOR_SVIDEO,
DRM_MODE_CONNECTOR_Unknown,
DRM_MODE_CONNECTOR_Unknown,
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
DRM_MODE_CONNECTOR_9PinDIN,
DRM_MODE_CONNECTOR_Unknown,
DRM_MODE_CONNECTOR_HDMIA,
DRM_MODE_CONNECTOR_HDMIB,
DRM_MODE_CONNECTOR_LVDS,
DRM_MODE_CONNECTOR_9PinDIN,
DRM_MODE_CONNECTOR_Unknown,
DRM_MODE_CONNECTOR_Unknown,
DRM_MODE_CONNECTOR_Unknown,
DRM_MODE_CONNECTOR_DisplayPort,
DRM_MODE_CONNECTOR_eDP,
DRM_MODE_CONNECTOR_Unknown
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
};
bool radeon_get_atom_connector_info_from_object_table(struct drm_device *dev)
{
struct radeon_device *rdev = dev->dev_private;
struct radeon_mode_info *mode_info = &rdev->mode_info;
struct atom_context *ctx = mode_info->atom_context;
int index = GetIndexIntoMasterTable(DATA, Object_Header);
u16 size, data_offset;
u8 frev, crev;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
ATOM_CONNECTOR_OBJECT_TABLE *con_obj;
ATOM_OBJECT_TABLE *router_obj;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
ATOM_DISPLAY_OBJECT_PATH_TABLE *path_obj;
ATOM_OBJECT_HEADER *obj_header;
int i, j, k, path_size, device_support;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
int connector_type;
u16 igp_lane_info, conn_id, connector_object_id;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
struct radeon_i2c_bus_rec ddc_bus;
struct radeon_router router;
struct radeon_gpio_rec gpio;
struct radeon_hpd hpd;
if (!atom_parse_data_header(ctx, index, &size, &frev, &crev, &data_offset))
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
return false;
if (crev < 2)
return false;
obj_header = (ATOM_OBJECT_HEADER *) (ctx->bios + data_offset);
path_obj = (ATOM_DISPLAY_OBJECT_PATH_TABLE *)
(ctx->bios + data_offset +
le16_to_cpu(obj_header->usDisplayPathTableOffset));
con_obj = (ATOM_CONNECTOR_OBJECT_TABLE *)
(ctx->bios + data_offset +
le16_to_cpu(obj_header->usConnectorObjectTableOffset));
router_obj = (ATOM_OBJECT_TABLE *)
(ctx->bios + data_offset +
le16_to_cpu(obj_header->usRouterObjectTableOffset));
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
device_support = le16_to_cpu(obj_header->usDeviceSupport);
path_size = 0;
for (i = 0; i < path_obj->ucNumOfDispPath; i++) {
uint8_t *addr = (uint8_t *) path_obj->asDispPath;
ATOM_DISPLAY_OBJECT_PATH *path;
addr += path_size;
path = (ATOM_DISPLAY_OBJECT_PATH *) addr;
path_size += le16_to_cpu(path->usSize);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
if (device_support & le16_to_cpu(path->usDeviceTag)) {
uint8_t con_obj_id, con_obj_num, con_obj_type;
con_obj_id =
(le16_to_cpu(path->usConnObjectId) & OBJECT_ID_MASK)
>> OBJECT_ID_SHIFT;
con_obj_num =
(le16_to_cpu(path->usConnObjectId) & ENUM_ID_MASK)
>> ENUM_ID_SHIFT;
con_obj_type =
(le16_to_cpu(path->usConnObjectId) &
OBJECT_TYPE_MASK) >> OBJECT_TYPE_SHIFT;
/* TODO CV support */
if (le16_to_cpu(path->usDeviceTag) ==
ATOM_DEVICE_CV_SUPPORT)
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
continue;
/* IGP chips */
if ((rdev->flags & RADEON_IS_IGP) &&
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
(con_obj_id ==
CONNECTOR_OBJECT_ID_PCIE_CONNECTOR)) {
uint16_t igp_offset = 0;
ATOM_INTEGRATED_SYSTEM_INFO_V2 *igp_obj;
index =
GetIndexIntoMasterTable(DATA,
IntegratedSystemInfo);
if (atom_parse_data_header(ctx, index, &size, &frev,
&crev, &igp_offset)) {
if (crev >= 2) {
igp_obj =
(ATOM_INTEGRATED_SYSTEM_INFO_V2
*) (ctx->bios + igp_offset);
if (igp_obj) {
uint32_t slot_config, ct;
if (con_obj_num == 1)
slot_config =
igp_obj->
ulDDISlot1Config;
else
slot_config =
igp_obj->
ulDDISlot2Config;
ct = (slot_config >> 16) & 0xff;
connector_type =
object_connector_convert
[ct];
connector_object_id = ct;
igp_lane_info =
slot_config & 0xffff;
} else
continue;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
} else
continue;
} else {
igp_lane_info = 0;
connector_type =
object_connector_convert[con_obj_id];
connector_object_id = con_obj_id;
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
} else {
igp_lane_info = 0;
connector_type =
object_connector_convert[con_obj_id];
connector_object_id = con_obj_id;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
if (connector_type == DRM_MODE_CONNECTOR_Unknown)
continue;
router.ddc_valid = false;
router.cd_valid = false;
for (j = 0; j < ((le16_to_cpu(path->usSize) - 8) / 2); j++) {
uint8_t grph_obj_id, grph_obj_num, grph_obj_type;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
grph_obj_id =
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
(le16_to_cpu(path->usGraphicObjIds[j]) &
OBJECT_ID_MASK) >> OBJECT_ID_SHIFT;
grph_obj_num =
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
(le16_to_cpu(path->usGraphicObjIds[j]) &
ENUM_ID_MASK) >> ENUM_ID_SHIFT;
grph_obj_type =
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
(le16_to_cpu(path->usGraphicObjIds[j]) &
OBJECT_TYPE_MASK) >> OBJECT_TYPE_SHIFT;
if (grph_obj_type == GRAPH_OBJECT_TYPE_ENCODER) {
u16 encoder_obj = le16_to_cpu(path->usGraphicObjIds[j]);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
radeon_add_atom_encoder(dev,
encoder_obj,
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
le16_to_cpu
(path->
usDeviceTag));
} else if (grph_obj_type == GRAPH_OBJECT_TYPE_ROUTER) {
for (k = 0; k < router_obj->ucNumberOfObjects; k++) {
u16 router_obj_id = le16_to_cpu(router_obj->asObjects[k].usObjectID);
if (le16_to_cpu(path->usGraphicObjIds[j]) == router_obj_id) {
ATOM_COMMON_RECORD_HEADER *record = (ATOM_COMMON_RECORD_HEADER *)
(ctx->bios + data_offset +
le16_to_cpu(router_obj->asObjects[k].usRecordOffset));
ATOM_I2C_RECORD *i2c_record;
ATOM_I2C_ID_CONFIG_ACCESS *i2c_config;
ATOM_ROUTER_DDC_PATH_SELECT_RECORD *ddc_path;
ATOM_ROUTER_DATA_CLOCK_PATH_SELECT_RECORD *cd_path;
ATOM_SRC_DST_TABLE_FOR_ONE_OBJECT *router_src_dst_table =
(ATOM_SRC_DST_TABLE_FOR_ONE_OBJECT *)
(ctx->bios + data_offset +
le16_to_cpu(router_obj->asObjects[k].usSrcDstTableOffset));
int enum_id;
router.router_id = router_obj_id;
for (enum_id = 0; enum_id < router_src_dst_table->ucNumberOfDst;
enum_id++) {
if (le16_to_cpu(path->usConnObjectId) ==
le16_to_cpu(router_src_dst_table->usDstObjectID[enum_id]))
break;
}
while (record->ucRecordType > 0 &&
record->ucRecordType <= ATOM_MAX_OBJECT_RECORD_NUMBER) {
switch (record->ucRecordType) {
case ATOM_I2C_RECORD_TYPE:
i2c_record =
(ATOM_I2C_RECORD *)
record;
i2c_config =
(ATOM_I2C_ID_CONFIG_ACCESS *)
&i2c_record->sucI2cId;
router.i2c_info =
radeon_lookup_i2c_gpio(rdev,
i2c_config->
ucAccess);
router.i2c_addr = i2c_record->ucI2CAddr >> 1;
break;
case ATOM_ROUTER_DDC_PATH_SELECT_RECORD_TYPE:
ddc_path = (ATOM_ROUTER_DDC_PATH_SELECT_RECORD *)
record;
router.ddc_valid = true;
router.ddc_mux_type = ddc_path->ucMuxType;
router.ddc_mux_control_pin = ddc_path->ucMuxControlPin;
router.ddc_mux_state = ddc_path->ucMuxState[enum_id];
break;
case ATOM_ROUTER_DATA_CLOCK_PATH_SELECT_RECORD_TYPE:
cd_path = (ATOM_ROUTER_DATA_CLOCK_PATH_SELECT_RECORD *)
record;
router.cd_valid = true;
router.cd_mux_type = cd_path->ucMuxType;
router.cd_mux_control_pin = cd_path->ucMuxControlPin;
router.cd_mux_state = cd_path->ucMuxState[enum_id];
break;
}
record = (ATOM_COMMON_RECORD_HEADER *)
((char *)record + record->ucRecordSize);
}
}
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
}
/* look up gpio for ddc, hpd */
ddc_bus.valid = false;
hpd.hpd = RADEON_HPD_NONE;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
if ((le16_to_cpu(path->usDeviceTag) &
(ATOM_DEVICE_TV_SUPPORT | ATOM_DEVICE_CV_SUPPORT)) == 0) {
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
for (j = 0; j < con_obj->ucNumberOfObjects; j++) {
if (le16_to_cpu(path->usConnObjectId) ==
le16_to_cpu(con_obj->asObjects[j].
usObjectID)) {
ATOM_COMMON_RECORD_HEADER
*record =
(ATOM_COMMON_RECORD_HEADER
*)
(ctx->bios + data_offset +
le16_to_cpu(con_obj->
asObjects[j].
usRecordOffset));
ATOM_I2C_RECORD *i2c_record;
ATOM_HPD_INT_RECORD *hpd_record;
ATOM_I2C_ID_CONFIG_ACCESS *i2c_config;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
while (record->ucRecordType > 0
&& record->
ucRecordType <=
ATOM_MAX_OBJECT_RECORD_NUMBER) {
switch (record->ucRecordType) {
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
case ATOM_I2C_RECORD_TYPE:
i2c_record =
(ATOM_I2C_RECORD *)
record;
i2c_config =
(ATOM_I2C_ID_CONFIG_ACCESS *)
&i2c_record->sucI2cId;
ddc_bus = radeon_lookup_i2c_gpio(rdev,
i2c_config->
ucAccess);
break;
case ATOM_HPD_INT_RECORD_TYPE:
hpd_record =
(ATOM_HPD_INT_RECORD *)
record;
gpio = radeon_lookup_gpio(rdev,
hpd_record->ucHPDIntGPIOID);
hpd = radeon_atom_get_hpd_info_from_gpio(rdev, &gpio);
hpd.plugged_state = hpd_record->ucPlugged_PinState;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
break;
}
record =
(ATOM_COMMON_RECORD_HEADER
*) ((char *)record
+
record->
ucRecordSize);
}
break;
}
}
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
/* needed for aux chan transactions */
ddc_bus.hpd = hpd.hpd;
conn_id = le16_to_cpu(path->usConnObjectId);
if (!radeon_atom_apply_quirks
(dev, le16_to_cpu(path->usDeviceTag), &connector_type,
&ddc_bus, &conn_id, &hpd))
continue;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
radeon_add_atom_connector(dev,
conn_id,
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
le16_to_cpu(path->
usDeviceTag),
connector_type, &ddc_bus,
igp_lane_info,
connector_object_id,
&hpd,
&router);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
}
radeon_link_encoder_connector(dev);
return true;
}
static uint16_t atombios_get_connector_object_id(struct drm_device *dev,
int connector_type,
uint16_t devices)
{
struct radeon_device *rdev = dev->dev_private;
if (rdev->flags & RADEON_IS_IGP) {
return supported_devices_connector_object_id_convert
[connector_type];
} else if (((connector_type == DRM_MODE_CONNECTOR_DVII) ||
(connector_type == DRM_MODE_CONNECTOR_DVID)) &&
(devices & ATOM_DEVICE_DFP2_SUPPORT)) {
struct radeon_mode_info *mode_info = &rdev->mode_info;
struct atom_context *ctx = mode_info->atom_context;
int index = GetIndexIntoMasterTable(DATA, XTMDS_Info);
uint16_t size, data_offset;
uint8_t frev, crev;
ATOM_XTMDS_INFO *xtmds;
if (atom_parse_data_header(ctx, index, &size, &frev, &crev, &data_offset)) {
xtmds = (ATOM_XTMDS_INFO *)(ctx->bios + data_offset);
if (xtmds->ucSupportedLink & ATOM_XTMDS_SUPPORTED_DUALLINK) {
if (connector_type == DRM_MODE_CONNECTOR_DVII)
return CONNECTOR_OBJECT_ID_DUAL_LINK_DVI_I;
else
return CONNECTOR_OBJECT_ID_DUAL_LINK_DVI_D;
} else {
if (connector_type == DRM_MODE_CONNECTOR_DVII)
return CONNECTOR_OBJECT_ID_SINGLE_LINK_DVI_I;
else
return CONNECTOR_OBJECT_ID_SINGLE_LINK_DVI_D;
}
} else
return supported_devices_connector_object_id_convert
[connector_type];
} else {
return supported_devices_connector_object_id_convert
[connector_type];
}
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
struct bios_connector {
bool valid;
uint16_t line_mux;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
uint16_t devices;
int connector_type;
struct radeon_i2c_bus_rec ddc_bus;
struct radeon_hpd hpd;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
};
bool radeon_get_atom_connector_info_from_supported_devices_table(struct
drm_device
*dev)
{
struct radeon_device *rdev = dev->dev_private;
struct radeon_mode_info *mode_info = &rdev->mode_info;
struct atom_context *ctx = mode_info->atom_context;
int index = GetIndexIntoMasterTable(DATA, SupportedDevicesInfo);
uint16_t size, data_offset;
uint8_t frev, crev;
uint16_t device_support;
uint8_t dac;
union atom_supported_devices *supported_devices;
int i, j, max_device;
struct bios_connector *bios_connectors;
size_t bc_size = sizeof(*bios_connectors) * ATOM_MAX_SUPPORTED_DEVICE;
struct radeon_router router;
router.ddc_valid = false;
router.cd_valid = false;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_connectors = kzalloc(bc_size, GFP_KERNEL);
if (!bios_connectors)
return false;
if (!atom_parse_data_header(ctx, index, &size, &frev, &crev,
&data_offset)) {
kfree(bios_connectors);
return false;
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
supported_devices =
(union atom_supported_devices *)(ctx->bios + data_offset);
device_support = le16_to_cpu(supported_devices->info.usDeviceSupport);
if (frev > 1)
max_device = ATOM_MAX_SUPPORTED_DEVICE;
else
max_device = ATOM_MAX_SUPPORTED_DEVICE_INFO;
for (i = 0; i < max_device; i++) {
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
ATOM_CONNECTOR_INFO_I2C ci =
supported_devices->info.asConnInfo[i];
bios_connectors[i].valid = false;
if (!(device_support & (1 << i))) {
continue;
}
if (i == ATOM_DEVICE_CV_INDEX) {
DRM_DEBUG_KMS("Skipping Component Video\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
continue;
}
bios_connectors[i].connector_type =
supported_devices_connector_convert[ci.sucConnectorInfo.
sbfAccess.
bfConnectorType];
if (bios_connectors[i].connector_type ==
DRM_MODE_CONNECTOR_Unknown)
continue;
dac = ci.sucConnectorInfo.sbfAccess.bfAssociatedDAC;
bios_connectors[i].line_mux =
ci.sucI2cId.ucAccess;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
/* give tv unique connector ids */
if (i == ATOM_DEVICE_TV1_INDEX) {
bios_connectors[i].ddc_bus.valid = false;
bios_connectors[i].line_mux = 50;
} else if (i == ATOM_DEVICE_TV2_INDEX) {
bios_connectors[i].ddc_bus.valid = false;
bios_connectors[i].line_mux = 51;
} else if (i == ATOM_DEVICE_CV_INDEX) {
bios_connectors[i].ddc_bus.valid = false;
bios_connectors[i].line_mux = 52;
} else
bios_connectors[i].ddc_bus =
radeon_lookup_i2c_gpio(rdev,
bios_connectors[i].line_mux);
if ((crev > 1) && (frev > 1)) {
u8 isb = supported_devices->info_2d1.asIntSrcInfo[i].ucIntSrcBitmap;
switch (isb) {
case 0x4:
bios_connectors[i].hpd.hpd = RADEON_HPD_1;
break;
case 0xa:
bios_connectors[i].hpd.hpd = RADEON_HPD_2;
break;
default:
bios_connectors[i].hpd.hpd = RADEON_HPD_NONE;
break;
}
} else {
if (i == ATOM_DEVICE_DFP1_INDEX)
bios_connectors[i].hpd.hpd = RADEON_HPD_1;
else if (i == ATOM_DEVICE_DFP2_INDEX)
bios_connectors[i].hpd.hpd = RADEON_HPD_2;
else
bios_connectors[i].hpd.hpd = RADEON_HPD_NONE;
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
/* Always set the connector type to VGA for CRT1/CRT2. if they are
* shared with a DVI port, we'll pick up the DVI connector when we
* merge the outputs. Some bioses incorrectly list VGA ports as DVI.
*/
if (i == ATOM_DEVICE_CRT1_INDEX || i == ATOM_DEVICE_CRT2_INDEX)
bios_connectors[i].connector_type =
DRM_MODE_CONNECTOR_VGA;
if (!radeon_atom_apply_quirks
(dev, (1 << i), &bios_connectors[i].connector_type,
&bios_connectors[i].ddc_bus, &bios_connectors[i].line_mux,
&bios_connectors[i].hpd))
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
continue;
bios_connectors[i].valid = true;
bios_connectors[i].devices = (1 << i);
if (ASIC_IS_AVIVO(rdev) || radeon_r4xx_atom)
radeon_add_atom_encoder(dev,
radeon_get_encoder_enum(dev,
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
(1 << i),
dac),
(1 << i));
else
radeon_add_legacy_encoder(dev,
radeon_get_encoder_enum(dev,
(1 << i),
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
dac),
(1 << i));
}
/* combine shared connectors */
for (i = 0; i < max_device; i++) {
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
if (bios_connectors[i].valid) {
for (j = 0; j < max_device; j++) {
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
if (bios_connectors[j].valid && (i != j)) {
if (bios_connectors[i].line_mux ==
bios_connectors[j].line_mux) {
/* make sure not to combine LVDS */
if (bios_connectors[i].devices & (ATOM_DEVICE_LCD_SUPPORT)) {
bios_connectors[i].line_mux = 53;
bios_connectors[i].ddc_bus.valid = false;
continue;
}
if (bios_connectors[j].devices & (ATOM_DEVICE_LCD_SUPPORT)) {
bios_connectors[j].line_mux = 53;
bios_connectors[j].ddc_bus.valid = false;
continue;
}
/* combine analog and digital for DVI-I */
if (((bios_connectors[i].devices & (ATOM_DEVICE_DFP_SUPPORT)) &&
(bios_connectors[j].devices & (ATOM_DEVICE_CRT_SUPPORT))) ||
((bios_connectors[j].devices & (ATOM_DEVICE_DFP_SUPPORT)) &&
(bios_connectors[i].devices & (ATOM_DEVICE_CRT_SUPPORT)))) {
bios_connectors[i].devices |=
bios_connectors[j].devices;
bios_connectors[i].connector_type =
DRM_MODE_CONNECTOR_DVII;
if (bios_connectors[j].devices & (ATOM_DEVICE_DFP_SUPPORT))
bios_connectors[i].hpd =
bios_connectors[j].hpd;
bios_connectors[j].valid = false;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
}
}
}
}
}
/* add the connectors */
for (i = 0; i < max_device; i++) {
if (bios_connectors[i].valid) {
uint16_t connector_object_id =
atombios_get_connector_object_id(dev,
bios_connectors[i].connector_type,
bios_connectors[i].devices);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
radeon_add_atom_connector(dev,
bios_connectors[i].line_mux,
bios_connectors[i].devices,
bios_connectors[i].
connector_type,
&bios_connectors[i].ddc_bus,
0,
connector_object_id,
&bios_connectors[i].hpd,
&router);
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
radeon_link_encoder_connector(dev);
kfree(bios_connectors);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
return true;
}
union firmware_info {
ATOM_FIRMWARE_INFO info;
ATOM_FIRMWARE_INFO_V1_2 info_12;
ATOM_FIRMWARE_INFO_V1_3 info_13;
ATOM_FIRMWARE_INFO_V1_4 info_14;
ATOM_FIRMWARE_INFO_V2_1 info_21;
ATOM_FIRMWARE_INFO_V2_2 info_22;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
};
bool radeon_atom_get_clock_info(struct drm_device *dev)
{
struct radeon_device *rdev = dev->dev_private;
struct radeon_mode_info *mode_info = &rdev->mode_info;
int index = GetIndexIntoMasterTable(DATA, FirmwareInfo);
union firmware_info *firmware_info;
uint8_t frev, crev;
struct radeon_pll *p1pll = &rdev->clock.p1pll;
struct radeon_pll *p2pll = &rdev->clock.p2pll;
struct radeon_pll *dcpll = &rdev->clock.dcpll;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
struct radeon_pll *spll = &rdev->clock.spll;
struct radeon_pll *mpll = &rdev->clock.mpll;
uint16_t data_offset;
if (atom_parse_data_header(mode_info->atom_context, index, NULL,
&frev, &crev, &data_offset)) {
firmware_info =
(union firmware_info *)(mode_info->atom_context->bios +
data_offset);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
/* pixel clocks */
p1pll->reference_freq =
le16_to_cpu(firmware_info->info.usReferenceClock);
p1pll->reference_div = 0;
if (crev < 2)
p1pll->pll_out_min =
le16_to_cpu(firmware_info->info.usMinPixelClockPLL_Output);
else
p1pll->pll_out_min =
le32_to_cpu(firmware_info->info_12.ulMinPixelClockPLL_Output);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
p1pll->pll_out_max =
le32_to_cpu(firmware_info->info.ulMaxPixelClockPLL_Output);
if (crev >= 4) {
p1pll->lcd_pll_out_min =
le16_to_cpu(firmware_info->info_14.usLcdMinPixelClockPLL_Output) * 100;
if (p1pll->lcd_pll_out_min == 0)
p1pll->lcd_pll_out_min = p1pll->pll_out_min;
p1pll->lcd_pll_out_max =
le16_to_cpu(firmware_info->info_14.usLcdMaxPixelClockPLL_Output) * 100;
if (p1pll->lcd_pll_out_max == 0)
p1pll->lcd_pll_out_max = p1pll->pll_out_max;
} else {
p1pll->lcd_pll_out_min = p1pll->pll_out_min;
p1pll->lcd_pll_out_max = p1pll->pll_out_max;
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
if (p1pll->pll_out_min == 0) {
if (ASIC_IS_AVIVO(rdev))
p1pll->pll_out_min = 64800;
else
p1pll->pll_out_min = 20000;
} else if (p1pll->pll_out_min > 64800) {
/* Limiting the pll output range is a good thing generally as
* it limits the number of possible pll combinations for a given
* frequency presumably to the ones that work best on each card.
* However, certain duallink DVI monitors seem to like
* pll combinations that would be limited by this at least on
* pre-DCE 3.0 r6xx hardware. This might need to be adjusted per
* family.
*/
p1pll->pll_out_min = 64800;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
p1pll->pll_in_min =
le16_to_cpu(firmware_info->info.usMinPixelClockPLL_Input);
p1pll->pll_in_max =
le16_to_cpu(firmware_info->info.usMaxPixelClockPLL_Input);
*p2pll = *p1pll;
/* system clock */
if (ASIC_IS_DCE4(rdev))
spll->reference_freq =
le16_to_cpu(firmware_info->info_21.usCoreReferenceClock);
else
spll->reference_freq =
le16_to_cpu(firmware_info->info.usReferenceClock);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
spll->reference_div = 0;
spll->pll_out_min =
le16_to_cpu(firmware_info->info.usMinEngineClockPLL_Output);
spll->pll_out_max =
le32_to_cpu(firmware_info->info.ulMaxEngineClockPLL_Output);
/* ??? */
if (spll->pll_out_min == 0) {
if (ASIC_IS_AVIVO(rdev))
spll->pll_out_min = 64800;
else
spll->pll_out_min = 20000;
}
spll->pll_in_min =
le16_to_cpu(firmware_info->info.usMinEngineClockPLL_Input);
spll->pll_in_max =
le16_to_cpu(firmware_info->info.usMaxEngineClockPLL_Input);
/* memory clock */
if (ASIC_IS_DCE4(rdev))
mpll->reference_freq =
le16_to_cpu(firmware_info->info_21.usMemoryReferenceClock);
else
mpll->reference_freq =
le16_to_cpu(firmware_info->info.usReferenceClock);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
mpll->reference_div = 0;
mpll->pll_out_min =
le16_to_cpu(firmware_info->info.usMinMemoryClockPLL_Output);
mpll->pll_out_max =
le32_to_cpu(firmware_info->info.ulMaxMemoryClockPLL_Output);
/* ??? */
if (mpll->pll_out_min == 0) {
if (ASIC_IS_AVIVO(rdev))
mpll->pll_out_min = 64800;
else
mpll->pll_out_min = 20000;
}
mpll->pll_in_min =
le16_to_cpu(firmware_info->info.usMinMemoryClockPLL_Input);
mpll->pll_in_max =
le16_to_cpu(firmware_info->info.usMaxMemoryClockPLL_Input);
rdev->clock.default_sclk =
le32_to_cpu(firmware_info->info.ulDefaultEngineClock);
rdev->clock.default_mclk =
le32_to_cpu(firmware_info->info.ulDefaultMemoryClock);
if (ASIC_IS_DCE4(rdev)) {
rdev->clock.default_dispclk =
le32_to_cpu(firmware_info->info_21.ulDefaultDispEngineClkFreq);
if (rdev->clock.default_dispclk == 0) {
if (ASIC_IS_DCE5(rdev))
rdev->clock.default_dispclk = 54000; /* 540 Mhz */
else
rdev->clock.default_dispclk = 60000; /* 600 Mhz */
}
rdev->clock.dp_extclk =
le16_to_cpu(firmware_info->info_21.usUniphyDPModeExtClkFreq);
}
*dcpll = *p1pll;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
return true;
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
return false;
}
union igp_info {
struct _ATOM_INTEGRATED_SYSTEM_INFO info;
struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 info_2;
};
bool radeon_atombios_sideport_present(struct radeon_device *rdev)
{
struct radeon_mode_info *mode_info = &rdev->mode_info;
int index = GetIndexIntoMasterTable(DATA, IntegratedSystemInfo);
union igp_info *igp_info;
u8 frev, crev;
u16 data_offset;
/* sideport is AMD only */
if (rdev->family == CHIP_RS600)
return false;
if (atom_parse_data_header(mode_info->atom_context, index, NULL,
&frev, &crev, &data_offset)) {
igp_info = (union igp_info *)(mode_info->atom_context->bios +
data_offset);
switch (crev) {
case 1:
if (igp_info->info.ulBootUpMemoryClock)
return true;
break;
case 2:
if (igp_info->info_2.ulBootUpSidePortClock)
return true;
break;
default:
DRM_ERROR("Unsupported IGP table: %d %d\n", frev, crev);
break;
}
}
return false;
}
bool radeon_atombios_get_tmds_info(struct radeon_encoder *encoder,
struct radeon_encoder_int_tmds *tmds)
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
{
struct drm_device *dev = encoder->base.dev;
struct radeon_device *rdev = dev->dev_private;
struct radeon_mode_info *mode_info = &rdev->mode_info;
int index = GetIndexIntoMasterTable(DATA, TMDS_Info);
uint16_t data_offset;
struct _ATOM_TMDS_INFO *tmds_info;
uint8_t frev, crev;
uint16_t maxfreq;
int i;
if (atom_parse_data_header(mode_info->atom_context, index, NULL,
&frev, &crev, &data_offset)) {
tmds_info =
(struct _ATOM_TMDS_INFO *)(mode_info->atom_context->bios +
data_offset);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
maxfreq = le16_to_cpu(tmds_info->usMaxFrequency);
for (i = 0; i < 4; i++) {
tmds->tmds_pll[i].freq =
le16_to_cpu(tmds_info->asMiscInfo[i].usFrequency);
tmds->tmds_pll[i].value =
tmds_info->asMiscInfo[i].ucPLL_ChargePump & 0x3f;
tmds->tmds_pll[i].value |=
(tmds_info->asMiscInfo[i].
ucPLL_VCO_Gain & 0x3f) << 6;
tmds->tmds_pll[i].value |=
(tmds_info->asMiscInfo[i].
ucPLL_DutyCycle & 0xf) << 12;
tmds->tmds_pll[i].value |=
(tmds_info->asMiscInfo[i].
ucPLL_VoltageSwing & 0xf) << 16;
DRM_DEBUG_KMS("TMDS PLL From ATOMBIOS %u %x\n",
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
tmds->tmds_pll[i].freq,
tmds->tmds_pll[i].value);
if (maxfreq == tmds->tmds_pll[i].freq) {
tmds->tmds_pll[i].freq = 0xffffffff;
break;
}
}
return true;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
return false;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
bool radeon_atombios_get_ppll_ss_info(struct radeon_device *rdev,
struct radeon_atom_ss *ss,
int id)
{
struct radeon_mode_info *mode_info = &rdev->mode_info;
int index = GetIndexIntoMasterTable(DATA, PPLL_SS_Info);
uint16_t data_offset, size;
struct _ATOM_SPREAD_SPECTRUM_INFO *ss_info;
uint8_t frev, crev;
int i, num_indices;
memset(ss, 0, sizeof(struct radeon_atom_ss));
if (atom_parse_data_header(mode_info->atom_context, index, &size,
&frev, &crev, &data_offset)) {
ss_info =
(struct _ATOM_SPREAD_SPECTRUM_INFO *)(mode_info->atom_context->bios + data_offset);
num_indices = (size - sizeof(ATOM_COMMON_TABLE_HEADER)) /
sizeof(ATOM_SPREAD_SPECTRUM_ASSIGNMENT);
for (i = 0; i < num_indices; i++) {
if (ss_info->asSS_Info[i].ucSS_Id == id) {
ss->percentage =
le16_to_cpu(ss_info->asSS_Info[i].usSpreadSpectrumPercentage);
ss->type = ss_info->asSS_Info[i].ucSpreadSpectrumType;
ss->step = ss_info->asSS_Info[i].ucSS_Step;
ss->delay = ss_info->asSS_Info[i].ucSS_Delay;
ss->range = ss_info->asSS_Info[i].ucSS_Range;
ss->refdiv = ss_info->asSS_Info[i].ucRecommendedRef_Div;
return true;
}
}
}
return false;
}
static void radeon_atombios_get_igp_ss_overrides(struct radeon_device *rdev,
struct radeon_atom_ss *ss,
int id)
{
struct radeon_mode_info *mode_info = &rdev->mode_info;
int index = GetIndexIntoMasterTable(DATA, IntegratedSystemInfo);
u16 data_offset, size;
struct _ATOM_INTEGRATED_SYSTEM_INFO_V6 *igp_info;
u8 frev, crev;
u16 percentage = 0, rate = 0;
/* get any igp specific overrides */
if (atom_parse_data_header(mode_info->atom_context, index, &size,
&frev, &crev, &data_offset)) {
igp_info = (struct _ATOM_INTEGRATED_SYSTEM_INFO_V6 *)
(mode_info->atom_context->bios + data_offset);
switch (id) {
case ASIC_INTERNAL_SS_ON_TMDS:
percentage = le16_to_cpu(igp_info->usDVISSPercentage);
rate = le16_to_cpu(igp_info->usDVISSpreadRateIn10Hz);
break;
case ASIC_INTERNAL_SS_ON_HDMI:
percentage = le16_to_cpu(igp_info->usHDMISSPercentage);
rate = le16_to_cpu(igp_info->usHDMISSpreadRateIn10Hz);
break;
case ASIC_INTERNAL_SS_ON_LVDS:
percentage = le16_to_cpu(igp_info->usLvdsSSPercentage);
rate = le16_to_cpu(igp_info->usLvdsSSpreadRateIn10Hz);
break;
}
if (percentage)
ss->percentage = percentage;
if (rate)
ss->rate = rate;
}
}
union asic_ss_info {
struct _ATOM_ASIC_INTERNAL_SS_INFO info;
struct _ATOM_ASIC_INTERNAL_SS_INFO_V2 info_2;
struct _ATOM_ASIC_INTERNAL_SS_INFO_V3 info_3;
};
bool radeon_atombios_get_asic_ss_info(struct radeon_device *rdev,
struct radeon_atom_ss *ss,
int id, u32 clock)
{
struct radeon_mode_info *mode_info = &rdev->mode_info;
int index = GetIndexIntoMasterTable(DATA, ASIC_InternalSS_Info);
uint16_t data_offset, size;
union asic_ss_info *ss_info;
uint8_t frev, crev;
int i, num_indices;
memset(ss, 0, sizeof(struct radeon_atom_ss));
if (atom_parse_data_header(mode_info->atom_context, index, &size,
&frev, &crev, &data_offset)) {
ss_info =
(union asic_ss_info *)(mode_info->atom_context->bios + data_offset);
switch (frev) {
case 1:
num_indices = (size - sizeof(ATOM_COMMON_TABLE_HEADER)) /
sizeof(ATOM_ASIC_SS_ASSIGNMENT);
for (i = 0; i < num_indices; i++) {
if ((ss_info->info.asSpreadSpectrum[i].ucClockIndication == id) &&
(clock <= ss_info->info.asSpreadSpectrum[i].ulTargetClockRange)) {
ss->percentage =
le16_to_cpu(ss_info->info.asSpreadSpectrum[i].usSpreadSpectrumPercentage);
ss->type = ss_info->info.asSpreadSpectrum[i].ucSpreadSpectrumMode;
ss->rate = le16_to_cpu(ss_info->info.asSpreadSpectrum[i].usSpreadRateInKhz);
return true;
}
}
break;
case 2:
num_indices = (size - sizeof(ATOM_COMMON_TABLE_HEADER)) /
sizeof(ATOM_ASIC_SS_ASSIGNMENT_V2);
for (i = 0; i < num_indices; i++) {
if ((ss_info->info_2.asSpreadSpectrum[i].ucClockIndication == id) &&
(clock <= ss_info->info_2.asSpreadSpectrum[i].ulTargetClockRange)) {
ss->percentage =
le16_to_cpu(ss_info->info_2.asSpreadSpectrum[i].usSpreadSpectrumPercentage);
ss->type = ss_info->info_2.asSpreadSpectrum[i].ucSpreadSpectrumMode;
ss->rate = le16_to_cpu(ss_info->info_2.asSpreadSpectrum[i].usSpreadRateIn10Hz);
return true;
}
}
break;
case 3:
num_indices = (size - sizeof(ATOM_COMMON_TABLE_HEADER)) /
sizeof(ATOM_ASIC_SS_ASSIGNMENT_V3);
for (i = 0; i < num_indices; i++) {
if ((ss_info->info_3.asSpreadSpectrum[i].ucClockIndication == id) &&
(clock <= ss_info->info_3.asSpreadSpectrum[i].ulTargetClockRange)) {
ss->percentage =
le16_to_cpu(ss_info->info_3.asSpreadSpectrum[i].usSpreadSpectrumPercentage);
ss->type = ss_info->info_3.asSpreadSpectrum[i].ucSpreadSpectrumMode;
ss->rate = le16_to_cpu(ss_info->info_3.asSpreadSpectrum[i].usSpreadRateIn10Hz);
if (rdev->flags & RADEON_IS_IGP)
radeon_atombios_get_igp_ss_overrides(rdev, ss, id);
return true;
}
}
break;
default:
DRM_ERROR("Unsupported ASIC_InternalSS_Info table: %d %d\n", frev, crev);
break;
}
}
return false;
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
union lvds_info {
struct _ATOM_LVDS_INFO info;
struct _ATOM_LVDS_INFO_V12 info_12;
};
struct radeon_encoder_atom_dig *radeon_atombios_get_lvds_info(struct
radeon_encoder
*encoder)
{
struct drm_device *dev = encoder->base.dev;
struct radeon_device *rdev = dev->dev_private;
struct radeon_mode_info *mode_info = &rdev->mode_info;
int index = GetIndexIntoMasterTable(DATA, LVDS_Info);
uint16_t data_offset, misc;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
union lvds_info *lvds_info;
uint8_t frev, crev;
struct radeon_encoder_atom_dig *lvds = NULL;
int encoder_enum = (encoder->encoder_enum & ENUM_ID_MASK) >> ENUM_ID_SHIFT;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
if (atom_parse_data_header(mode_info->atom_context, index, NULL,
&frev, &crev, &data_offset)) {
lvds_info =
(union lvds_info *)(mode_info->atom_context->bios + data_offset);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
lvds =
kzalloc(sizeof(struct radeon_encoder_atom_dig), GFP_KERNEL);
if (!lvds)
return NULL;
lvds->native_mode.clock =
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
le16_to_cpu(lvds_info->info.sLCDTiming.usPixClk) * 10;
lvds->native_mode.hdisplay =
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
le16_to_cpu(lvds_info->info.sLCDTiming.usHActive);
lvds->native_mode.vdisplay =
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
le16_to_cpu(lvds_info->info.sLCDTiming.usVActive);
lvds->native_mode.htotal = lvds->native_mode.hdisplay +
le16_to_cpu(lvds_info->info.sLCDTiming.usHBlanking_Time);
lvds->native_mode.hsync_start = lvds->native_mode.hdisplay +
le16_to_cpu(lvds_info->info.sLCDTiming.usHSyncOffset);
lvds->native_mode.hsync_end = lvds->native_mode.hsync_start +
le16_to_cpu(lvds_info->info.sLCDTiming.usHSyncWidth);
lvds->native_mode.vtotal = lvds->native_mode.vdisplay +
le16_to_cpu(lvds_info->info.sLCDTiming.usVBlanking_Time);
lvds->native_mode.vsync_start = lvds->native_mode.vdisplay +
le16_to_cpu(lvds_info->info.sLCDTiming.usVSyncOffset);
lvds->native_mode.vsync_end = lvds->native_mode.vsync_start +
le16_to_cpu(lvds_info->info.sLCDTiming.usVSyncWidth);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
lvds->panel_pwr_delay =
le16_to_cpu(lvds_info->info.usOffDelayInMs);
lvds->lcd_misc = lvds_info->info.ucLVDS_Misc;
misc = le16_to_cpu(lvds_info->info.sLCDTiming.susModeMiscInfo.usAccess);
if (misc & ATOM_VSYNC_POLARITY)
lvds->native_mode.flags |= DRM_MODE_FLAG_NVSYNC;
if (misc & ATOM_HSYNC_POLARITY)
lvds->native_mode.flags |= DRM_MODE_FLAG_NHSYNC;
if (misc & ATOM_COMPOSITESYNC)
lvds->native_mode.flags |= DRM_MODE_FLAG_CSYNC;
if (misc & ATOM_INTERLACE)
lvds->native_mode.flags |= DRM_MODE_FLAG_INTERLACE;
if (misc & ATOM_DOUBLE_CLOCK_MODE)
lvds->native_mode.flags |= DRM_MODE_FLAG_DBLSCAN;
lvds->native_mode.width_mm = lvds_info->info.sLCDTiming.usImageHSize;
lvds->native_mode.height_mm = lvds_info->info.sLCDTiming.usImageVSize;
/* set crtc values */
drm_mode_set_crtcinfo(&lvds->native_mode, CRTC_INTERLACE_HALVE_V);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
lvds->lcd_ss_id = lvds_info->info.ucSS_Id;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
encoder->native_mode = lvds->native_mode;
if (encoder_enum == 2)
lvds->linkb = true;
else
lvds->linkb = false;
/* parse the lcd record table */
if (lvds_info->info.usModePatchTableOffset) {
ATOM_FAKE_EDID_PATCH_RECORD *fake_edid_record;
ATOM_PANEL_RESOLUTION_PATCH_RECORD *panel_res_record;
bool bad_record = false;
u8 *record = (u8 *)(mode_info->atom_context->bios +
data_offset +
lvds_info->info.usModePatchTableOffset);
while (*record != ATOM_RECORD_END_TYPE) {
switch (*record) {
case LCD_MODE_PATCH_RECORD_MODE_TYPE:
record += sizeof(ATOM_PATCH_RECORD_MODE);
break;
case LCD_RTS_RECORD_TYPE:
record += sizeof(ATOM_LCD_RTS_RECORD);
break;
case LCD_CAP_RECORD_TYPE:
record += sizeof(ATOM_LCD_MODE_CONTROL_CAP);
break;
case LCD_FAKE_EDID_PATCH_RECORD_TYPE:
fake_edid_record = (ATOM_FAKE_EDID_PATCH_RECORD *)record;
if (fake_edid_record->ucFakeEDIDLength) {
struct edid *edid;
int edid_size =
max((int)EDID_LENGTH, (int)fake_edid_record->ucFakeEDIDLength);
edid = kmalloc(edid_size, GFP_KERNEL);
if (edid) {
memcpy((u8 *)edid, (u8 *)&fake_edid_record->ucFakeEDIDString[0],
fake_edid_record->ucFakeEDIDLength);
if (drm_edid_is_valid(edid))
rdev->mode_info.bios_hardcoded_edid = edid;
else
kfree(edid);
}
}
record += sizeof(ATOM_FAKE_EDID_PATCH_RECORD);
break;
case LCD_PANEL_RESOLUTION_RECORD_TYPE:
panel_res_record = (ATOM_PANEL_RESOLUTION_PATCH_RECORD *)record;
lvds->native_mode.width_mm = panel_res_record->usHSize;
lvds->native_mode.height_mm = panel_res_record->usVSize;
record += sizeof(ATOM_PANEL_RESOLUTION_PATCH_RECORD);
break;
default:
DRM_ERROR("Bad LCD record %d\n", *record);
bad_record = true;
break;
}
if (bad_record)
break;
}
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
return lvds;
}
struct radeon_encoder_primary_dac *
radeon_atombios_get_primary_dac_info(struct radeon_encoder *encoder)
{
struct drm_device *dev = encoder->base.dev;
struct radeon_device *rdev = dev->dev_private;
struct radeon_mode_info *mode_info = &rdev->mode_info;
int index = GetIndexIntoMasterTable(DATA, CompassionateData);
uint16_t data_offset;
struct _COMPASSIONATE_DATA *dac_info;
uint8_t frev, crev;
uint8_t bg, dac;
struct radeon_encoder_primary_dac *p_dac = NULL;
if (atom_parse_data_header(mode_info->atom_context, index, NULL,
&frev, &crev, &data_offset)) {
dac_info = (struct _COMPASSIONATE_DATA *)
(mode_info->atom_context->bios + data_offset);
p_dac = kzalloc(sizeof(struct radeon_encoder_primary_dac), GFP_KERNEL);
if (!p_dac)
return NULL;
bg = dac_info->ucDAC1_BG_Adjustment;
dac = dac_info->ucDAC1_DAC_Adjustment;
p_dac->ps2_pdac_adj = (bg << 8) | (dac);
}
return p_dac;
}
bool radeon_atom_get_tv_timings(struct radeon_device *rdev, int index,
struct drm_display_mode *mode)
{
struct radeon_mode_info *mode_info = &rdev->mode_info;
ATOM_ANALOG_TV_INFO *tv_info;
ATOM_ANALOG_TV_INFO_V1_2 *tv_info_v1_2;
ATOM_DTD_FORMAT *dtd_timings;
int data_index = GetIndexIntoMasterTable(DATA, AnalogTV_Info);
u8 frev, crev;
u16 data_offset, misc;
if (!atom_parse_data_header(mode_info->atom_context, data_index, NULL,
&frev, &crev, &data_offset))
return false;
switch (crev) {
case 1:
tv_info = (ATOM_ANALOG_TV_INFO *)(mode_info->atom_context->bios + data_offset);
if (index >= MAX_SUPPORTED_TV_TIMING)
return false;
mode->crtc_htotal = le16_to_cpu(tv_info->aModeTimings[index].usCRTC_H_Total);
mode->crtc_hdisplay = le16_to_cpu(tv_info->aModeTimings[index].usCRTC_H_Disp);
mode->crtc_hsync_start = le16_to_cpu(tv_info->aModeTimings[index].usCRTC_H_SyncStart);
mode->crtc_hsync_end = le16_to_cpu(tv_info->aModeTimings[index].usCRTC_H_SyncStart) +
le16_to_cpu(tv_info->aModeTimings[index].usCRTC_H_SyncWidth);
mode->crtc_vtotal = le16_to_cpu(tv_info->aModeTimings[index].usCRTC_V_Total);
mode->crtc_vdisplay = le16_to_cpu(tv_info->aModeTimings[index].usCRTC_V_Disp);
mode->crtc_vsync_start = le16_to_cpu(tv_info->aModeTimings[index].usCRTC_V_SyncStart);
mode->crtc_vsync_end = le16_to_cpu(tv_info->aModeTimings[index].usCRTC_V_SyncStart) +
le16_to_cpu(tv_info->aModeTimings[index].usCRTC_V_SyncWidth);
mode->flags = 0;
misc = le16_to_cpu(tv_info->aModeTimings[index].susModeMiscInfo.usAccess);
if (misc & ATOM_VSYNC_POLARITY)
mode->flags |= DRM_MODE_FLAG_NVSYNC;
if (misc & ATOM_HSYNC_POLARITY)
mode->flags |= DRM_MODE_FLAG_NHSYNC;
if (misc & ATOM_COMPOSITESYNC)
mode->flags |= DRM_MODE_FLAG_CSYNC;
if (misc & ATOM_INTERLACE)
mode->flags |= DRM_MODE_FLAG_INTERLACE;
if (misc & ATOM_DOUBLE_CLOCK_MODE)
mode->flags |= DRM_MODE_FLAG_DBLSCAN;
mode->clock = le16_to_cpu(tv_info->aModeTimings[index].usPixelClock) * 10;
if (index == 1) {
/* PAL timings appear to have wrong values for totals */
mode->crtc_htotal -= 1;
mode->crtc_vtotal -= 1;
}
break;
case 2:
tv_info_v1_2 = (ATOM_ANALOG_TV_INFO_V1_2 *)(mode_info->atom_context->bios + data_offset);
if (index >= MAX_SUPPORTED_TV_TIMING_V1_2)
return false;
dtd_timings = &tv_info_v1_2->aModeTimings[index];
mode->crtc_htotal = le16_to_cpu(dtd_timings->usHActive) +
le16_to_cpu(dtd_timings->usHBlanking_Time);
mode->crtc_hdisplay = le16_to_cpu(dtd_timings->usHActive);
mode->crtc_hsync_start = le16_to_cpu(dtd_timings->usHActive) +
le16_to_cpu(dtd_timings->usHSyncOffset);
mode->crtc_hsync_end = mode->crtc_hsync_start +
le16_to_cpu(dtd_timings->usHSyncWidth);
mode->crtc_vtotal = le16_to_cpu(dtd_timings->usVActive) +
le16_to_cpu(dtd_timings->usVBlanking_Time);
mode->crtc_vdisplay = le16_to_cpu(dtd_timings->usVActive);
mode->crtc_vsync_start = le16_to_cpu(dtd_timings->usVActive) +
le16_to_cpu(dtd_timings->usVSyncOffset);
mode->crtc_vsync_end = mode->crtc_vsync_start +
le16_to_cpu(dtd_timings->usVSyncWidth);
mode->flags = 0;
misc = le16_to_cpu(dtd_timings->susModeMiscInfo.usAccess);
if (misc & ATOM_VSYNC_POLARITY)
mode->flags |= DRM_MODE_FLAG_NVSYNC;
if (misc & ATOM_HSYNC_POLARITY)
mode->flags |= DRM_MODE_FLAG_NHSYNC;
if (misc & ATOM_COMPOSITESYNC)
mode->flags |= DRM_MODE_FLAG_CSYNC;
if (misc & ATOM_INTERLACE)
mode->flags |= DRM_MODE_FLAG_INTERLACE;
if (misc & ATOM_DOUBLE_CLOCK_MODE)
mode->flags |= DRM_MODE_FLAG_DBLSCAN;
mode->clock = le16_to_cpu(dtd_timings->usPixClk) * 10;
break;
}
return true;
}
enum radeon_tv_std
radeon_atombios_get_tv_info(struct radeon_device *rdev)
{
struct radeon_mode_info *mode_info = &rdev->mode_info;
int index = GetIndexIntoMasterTable(DATA, AnalogTV_Info);
uint16_t data_offset;
uint8_t frev, crev;
struct _ATOM_ANALOG_TV_INFO *tv_info;
enum radeon_tv_std tv_std = TV_STD_NTSC;
if (atom_parse_data_header(mode_info->atom_context, index, NULL,
&frev, &crev, &data_offset)) {
tv_info = (struct _ATOM_ANALOG_TV_INFO *)
(mode_info->atom_context->bios + data_offset);
switch (tv_info->ucTV_BootUpDefaultStandard) {
case ATOM_TV_NTSC:
tv_std = TV_STD_NTSC;
DRM_DEBUG_KMS("Default TV standard: NTSC\n");
break;
case ATOM_TV_NTSCJ:
tv_std = TV_STD_NTSC_J;
DRM_DEBUG_KMS("Default TV standard: NTSC-J\n");
break;
case ATOM_TV_PAL:
tv_std = TV_STD_PAL;
DRM_DEBUG_KMS("Default TV standard: PAL\n");
break;
case ATOM_TV_PALM:
tv_std = TV_STD_PAL_M;
DRM_DEBUG_KMS("Default TV standard: PAL-M\n");
break;
case ATOM_TV_PALN:
tv_std = TV_STD_PAL_N;
DRM_DEBUG_KMS("Default TV standard: PAL-N\n");
break;
case ATOM_TV_PALCN:
tv_std = TV_STD_PAL_CN;
DRM_DEBUG_KMS("Default TV standard: PAL-CN\n");
break;
case ATOM_TV_PAL60:
tv_std = TV_STD_PAL_60;
DRM_DEBUG_KMS("Default TV standard: PAL-60\n");
break;
case ATOM_TV_SECAM:
tv_std = TV_STD_SECAM;
DRM_DEBUG_KMS("Default TV standard: SECAM\n");
break;
default:
tv_std = TV_STD_NTSC;
DRM_DEBUG_KMS("Unknown TV standard; defaulting to NTSC\n");
break;
}
}
return tv_std;
}
struct radeon_encoder_tv_dac *
radeon_atombios_get_tv_dac_info(struct radeon_encoder *encoder)
{
struct drm_device *dev = encoder->base.dev;
struct radeon_device *rdev = dev->dev_private;
struct radeon_mode_info *mode_info = &rdev->mode_info;
int index = GetIndexIntoMasterTable(DATA, CompassionateData);
uint16_t data_offset;
struct _COMPASSIONATE_DATA *dac_info;
uint8_t frev, crev;
uint8_t bg, dac;
struct radeon_encoder_tv_dac *tv_dac = NULL;
if (atom_parse_data_header(mode_info->atom_context, index, NULL,
&frev, &crev, &data_offset)) {
dac_info = (struct _COMPASSIONATE_DATA *)
(mode_info->atom_context->bios + data_offset);
tv_dac = kzalloc(sizeof(struct radeon_encoder_tv_dac), GFP_KERNEL);
if (!tv_dac)
return NULL;
bg = dac_info->ucDAC2_CRT2_BG_Adjustment;
dac = dac_info->ucDAC2_CRT2_DAC_Adjustment;
tv_dac->ps2_tvdac_adj = (bg << 16) | (dac << 20);
bg = dac_info->ucDAC2_PAL_BG_Adjustment;
dac = dac_info->ucDAC2_PAL_DAC_Adjustment;
tv_dac->pal_tvdac_adj = (bg << 16) | (dac << 20);
bg = dac_info->ucDAC2_NTSC_BG_Adjustment;
dac = dac_info->ucDAC2_NTSC_DAC_Adjustment;
tv_dac->ntsc_tvdac_adj = (bg << 16) | (dac << 20);
tv_dac->tv_std = radeon_atombios_get_tv_info(rdev);
}
return tv_dac;
}
static const char *thermal_controller_names[] = {
"NONE",
"lm63",
"adm1032",
"adm1030",
"max6649",
"lm64",
"f75375",
"asc7xxx",
};
static const char *pp_lib_thermal_controller_names[] = {
"NONE",
"lm63",
"adm1032",
"adm1030",
"max6649",
"lm64",
"f75375",
"RV6xx",
"RV770",
"adt7473",
"NONE",
"External GPIO",
"Evergreen",
"emc2103",
"Sumo",
};
union power_info {
struct _ATOM_POWERPLAY_INFO info;
struct _ATOM_POWERPLAY_INFO_V2 info_2;
struct _ATOM_POWERPLAY_INFO_V3 info_3;
struct _ATOM_PPLIB_POWERPLAYTABLE pplib;
struct _ATOM_PPLIB_POWERPLAYTABLE2 pplib2;
struct _ATOM_PPLIB_POWERPLAYTABLE3 pplib3;
};
union pplib_clock_info {
struct _ATOM_PPLIB_R600_CLOCK_INFO r600;
struct _ATOM_PPLIB_RS780_CLOCK_INFO rs780;
struct _ATOM_PPLIB_EVERGREEN_CLOCK_INFO evergreen;
struct _ATOM_PPLIB_SUMO_CLOCK_INFO sumo;
};
union pplib_power_state {
struct _ATOM_PPLIB_STATE v1;
struct _ATOM_PPLIB_STATE_V2 v2;
};
static void radeon_atombios_parse_misc_flags_1_3(struct radeon_device *rdev,
int state_index,
u32 misc, u32 misc2)
{
rdev->pm.power_state[state_index].misc = misc;
rdev->pm.power_state[state_index].misc2 = misc2;
/* order matters! */
if (misc & ATOM_PM_MISCINFO_POWER_SAVING_MODE)
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_POWERSAVE;
if (misc & ATOM_PM_MISCINFO_DEFAULT_DC_STATE_ENTRY_TRUE)
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_BATTERY;
if (misc & ATOM_PM_MISCINFO_DEFAULT_LOW_DC_STATE_ENTRY_TRUE)
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_BATTERY;
if (misc & ATOM_PM_MISCINFO_LOAD_BALANCE_EN)
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_BALANCED;
if (misc & ATOM_PM_MISCINFO_3D_ACCELERATION_EN) {
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_PERFORMANCE;
rdev->pm.power_state[state_index].flags &=
~RADEON_PM_STATE_SINGLE_DISPLAY_ONLY;
}
if (misc2 & ATOM_PM_MISCINFO2_SYSTEM_AC_LITE_MODE)
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_BALANCED;
if (misc & ATOM_PM_MISCINFO_DRIVER_DEFAULT_MODE) {
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_DEFAULT;
rdev->pm.default_power_state_index = state_index;
rdev->pm.power_state[state_index].default_clock_mode =
&rdev->pm.power_state[state_index].clock_info[0];
} else if (state_index == 0) {
rdev->pm.power_state[state_index].clock_info[0].flags |=
RADEON_PM_MODE_NO_DISPLAY;
}
}
static int radeon_atombios_parse_power_table_1_3(struct radeon_device *rdev)
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
{
struct radeon_mode_info *mode_info = &rdev->mode_info;
u32 misc, misc2 = 0;
int num_modes = 0, i;
int state_index = 0;
struct radeon_i2c_bus_rec i2c_bus;
union power_info *power_info;
int index = GetIndexIntoMasterTable(DATA, PowerPlayInfo);
u16 data_offset;
u8 frev, crev;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
if (!atom_parse_data_header(mode_info->atom_context, index, NULL,
&frev, &crev, &data_offset))
return state_index;
power_info = (union power_info *)(mode_info->atom_context->bios + data_offset);
/* add the i2c bus for thermal/fan chip */
if (power_info->info.ucOverdriveThermalController > 0) {
DRM_INFO("Possible %s thermal controller at 0x%02x\n",
thermal_controller_names[power_info->info.ucOverdriveThermalController],
power_info->info.ucOverdriveControllerAddress >> 1);
i2c_bus = radeon_lookup_i2c_gpio(rdev, power_info->info.ucOverdriveI2cLine);
rdev->pm.i2c_bus = radeon_i2c_lookup(rdev, &i2c_bus);
if (rdev->pm.i2c_bus) {
struct i2c_board_info info = { };
const char *name = thermal_controller_names[power_info->info.
ucOverdriveThermalController];
info.addr = power_info->info.ucOverdriveControllerAddress >> 1;
strlcpy(info.type, name, sizeof(info.type));
i2c_new_device(&rdev->pm.i2c_bus->adapter, &info);
}
}
num_modes = power_info->info.ucNumOfPowerModeEntries;
if (num_modes > ATOM_MAX_NUMBEROF_POWER_BLOCK)
num_modes = ATOM_MAX_NUMBEROF_POWER_BLOCK;
/* last mode is usually default, array is low to high */
for (i = 0; i < num_modes; i++) {
rdev->pm.power_state[state_index].clock_info[0].voltage.type = VOLTAGE_NONE;
switch (frev) {
case 1:
rdev->pm.power_state[state_index].num_clock_modes = 1;
rdev->pm.power_state[state_index].clock_info[0].mclk =
le16_to_cpu(power_info->info.asPowerPlayInfo[i].usMemoryClock);
rdev->pm.power_state[state_index].clock_info[0].sclk =
le16_to_cpu(power_info->info.asPowerPlayInfo[i].usEngineClock);
/* skip invalid modes */
if ((rdev->pm.power_state[state_index].clock_info[0].mclk == 0) ||
(rdev->pm.power_state[state_index].clock_info[0].sclk == 0))
continue;
rdev->pm.power_state[state_index].pcie_lanes =
power_info->info.asPowerPlayInfo[i].ucNumPciELanes;
misc = le32_to_cpu(power_info->info.asPowerPlayInfo[i].ulMiscInfo);
if ((misc & ATOM_PM_MISCINFO_VOLTAGE_DROP_SUPPORT) ||
(misc & ATOM_PM_MISCINFO_VOLTAGE_DROP_ACTIVE_HIGH)) {
rdev->pm.power_state[state_index].clock_info[0].voltage.type =
VOLTAGE_GPIO;
rdev->pm.power_state[state_index].clock_info[0].voltage.gpio =
radeon_lookup_gpio(rdev,
power_info->info.asPowerPlayInfo[i].ucVoltageDropIndex);
if (misc & ATOM_PM_MISCINFO_VOLTAGE_DROP_ACTIVE_HIGH)
rdev->pm.power_state[state_index].clock_info[0].voltage.active_high =
true;
else
rdev->pm.power_state[state_index].clock_info[0].voltage.active_high =
false;
} else if (misc & ATOM_PM_MISCINFO_PROGRAM_VOLTAGE) {
rdev->pm.power_state[state_index].clock_info[0].voltage.type =
VOLTAGE_VDDC;
rdev->pm.power_state[state_index].clock_info[0].voltage.vddc_id =
power_info->info.asPowerPlayInfo[i].ucVoltageDropIndex;
}
rdev->pm.power_state[state_index].flags = RADEON_PM_STATE_SINGLE_DISPLAY_ONLY;
radeon_atombios_parse_misc_flags_1_3(rdev, state_index, misc, 0);
state_index++;
break;
case 2:
rdev->pm.power_state[state_index].num_clock_modes = 1;
rdev->pm.power_state[state_index].clock_info[0].mclk =
le32_to_cpu(power_info->info_2.asPowerPlayInfo[i].ulMemoryClock);
rdev->pm.power_state[state_index].clock_info[0].sclk =
le32_to_cpu(power_info->info_2.asPowerPlayInfo[i].ulEngineClock);
/* skip invalid modes */
if ((rdev->pm.power_state[state_index].clock_info[0].mclk == 0) ||
(rdev->pm.power_state[state_index].clock_info[0].sclk == 0))
continue;
rdev->pm.power_state[state_index].pcie_lanes =
power_info->info_2.asPowerPlayInfo[i].ucNumPciELanes;
misc = le32_to_cpu(power_info->info_2.asPowerPlayInfo[i].ulMiscInfo);
misc2 = le32_to_cpu(power_info->info_2.asPowerPlayInfo[i].ulMiscInfo2);
if ((misc & ATOM_PM_MISCINFO_VOLTAGE_DROP_SUPPORT) ||
(misc & ATOM_PM_MISCINFO_VOLTAGE_DROP_ACTIVE_HIGH)) {
rdev->pm.power_state[state_index].clock_info[0].voltage.type =
VOLTAGE_GPIO;
rdev->pm.power_state[state_index].clock_info[0].voltage.gpio =
radeon_lookup_gpio(rdev,
power_info->info_2.asPowerPlayInfo[i].ucVoltageDropIndex);
if (misc & ATOM_PM_MISCINFO_VOLTAGE_DROP_ACTIVE_HIGH)
rdev->pm.power_state[state_index].clock_info[0].voltage.active_high =
true;
else
rdev->pm.power_state[state_index].clock_info[0].voltage.active_high =
false;
} else if (misc & ATOM_PM_MISCINFO_PROGRAM_VOLTAGE) {
rdev->pm.power_state[state_index].clock_info[0].voltage.type =
VOLTAGE_VDDC;
rdev->pm.power_state[state_index].clock_info[0].voltage.vddc_id =
power_info->info_2.asPowerPlayInfo[i].ucVoltageDropIndex;
}
rdev->pm.power_state[state_index].flags = RADEON_PM_STATE_SINGLE_DISPLAY_ONLY;
radeon_atombios_parse_misc_flags_1_3(rdev, state_index, misc, misc2);
state_index++;
break;
case 3:
rdev->pm.power_state[state_index].num_clock_modes = 1;
rdev->pm.power_state[state_index].clock_info[0].mclk =
le32_to_cpu(power_info->info_3.asPowerPlayInfo[i].ulMemoryClock);
rdev->pm.power_state[state_index].clock_info[0].sclk =
le32_to_cpu(power_info->info_3.asPowerPlayInfo[i].ulEngineClock);
/* skip invalid modes */
if ((rdev->pm.power_state[state_index].clock_info[0].mclk == 0) ||
(rdev->pm.power_state[state_index].clock_info[0].sclk == 0))
continue;
rdev->pm.power_state[state_index].pcie_lanes =
power_info->info_3.asPowerPlayInfo[i].ucNumPciELanes;
misc = le32_to_cpu(power_info->info_3.asPowerPlayInfo[i].ulMiscInfo);
misc2 = le32_to_cpu(power_info->info_3.asPowerPlayInfo[i].ulMiscInfo2);
if ((misc & ATOM_PM_MISCINFO_VOLTAGE_DROP_SUPPORT) ||
(misc & ATOM_PM_MISCINFO_VOLTAGE_DROP_ACTIVE_HIGH)) {
rdev->pm.power_state[state_index].clock_info[0].voltage.type =
VOLTAGE_GPIO;
rdev->pm.power_state[state_index].clock_info[0].voltage.gpio =
radeon_lookup_gpio(rdev,
power_info->info_3.asPowerPlayInfo[i].ucVoltageDropIndex);
if (misc & ATOM_PM_MISCINFO_VOLTAGE_DROP_ACTIVE_HIGH)
rdev->pm.power_state[state_index].clock_info[0].voltage.active_high =
true;
else
rdev->pm.power_state[state_index].clock_info[0].voltage.active_high =
false;
} else if (misc & ATOM_PM_MISCINFO_PROGRAM_VOLTAGE) {
rdev->pm.power_state[state_index].clock_info[0].voltage.type =
VOLTAGE_VDDC;
rdev->pm.power_state[state_index].clock_info[0].voltage.vddc_id =
power_info->info_3.asPowerPlayInfo[i].ucVoltageDropIndex;
if (misc2 & ATOM_PM_MISCINFO2_VDDCI_DYNAMIC_VOLTAGE_EN) {
rdev->pm.power_state[state_index].clock_info[0].voltage.vddci_enabled =
true;
rdev->pm.power_state[state_index].clock_info[0].voltage.vddci_id =
power_info->info_3.asPowerPlayInfo[i].ucVDDCI_VoltageDropIndex;
}
}
rdev->pm.power_state[state_index].flags = RADEON_PM_STATE_SINGLE_DISPLAY_ONLY;
radeon_atombios_parse_misc_flags_1_3(rdev, state_index, misc, misc2);
state_index++;
break;
}
}
/* last mode is usually default */
if (rdev->pm.default_power_state_index == -1) {
rdev->pm.power_state[state_index - 1].type =
POWER_STATE_TYPE_DEFAULT;
rdev->pm.default_power_state_index = state_index - 1;
rdev->pm.power_state[state_index - 1].default_clock_mode =
&rdev->pm.power_state[state_index - 1].clock_info[0];
rdev->pm.power_state[state_index].flags &=
~RADEON_PM_STATE_SINGLE_DISPLAY_ONLY;
rdev->pm.power_state[state_index].misc = 0;
rdev->pm.power_state[state_index].misc2 = 0;
}
return state_index;
}
static void radeon_atombios_add_pplib_thermal_controller(struct radeon_device *rdev,
ATOM_PPLIB_THERMALCONTROLLER *controller)
{
struct radeon_i2c_bus_rec i2c_bus;
/* add the i2c bus for thermal/fan chip */
if (controller->ucType > 0) {
if (controller->ucType == ATOM_PP_THERMALCONTROLLER_RV6xx) {
DRM_INFO("Internal thermal controller %s fan control\n",
(controller->ucFanParameters &
ATOM_PP_FANPARAMETERS_NOFAN) ? "without" : "with");
rdev->pm.int_thermal_type = THERMAL_TYPE_RV6XX;
} else if (controller->ucType == ATOM_PP_THERMALCONTROLLER_RV770) {
DRM_INFO("Internal thermal controller %s fan control\n",
(controller->ucFanParameters &
ATOM_PP_FANPARAMETERS_NOFAN) ? "without" : "with");
rdev->pm.int_thermal_type = THERMAL_TYPE_RV770;
} else if (controller->ucType == ATOM_PP_THERMALCONTROLLER_EVERGREEN) {
DRM_INFO("Internal thermal controller %s fan control\n",
(controller->ucFanParameters &
ATOM_PP_FANPARAMETERS_NOFAN) ? "without" : "with");
rdev->pm.int_thermal_type = THERMAL_TYPE_EVERGREEN;
} else if (controller->ucType == ATOM_PP_THERMALCONTROLLER_SUMO) {
DRM_INFO("Internal thermal controller %s fan control\n",
(controller->ucFanParameters &
ATOM_PP_FANPARAMETERS_NOFAN) ? "without" : "with");
rdev->pm.int_thermal_type = THERMAL_TYPE_SUMO;
} else if ((controller->ucType ==
ATOM_PP_THERMALCONTROLLER_EXTERNAL_GPIO) ||
(controller->ucType ==
ATOM_PP_THERMALCONTROLLER_ADT7473_WITH_INTERNAL) ||
(controller->ucType ==
ATOM_PP_THERMALCONTROLLER_EMC2103_WITH_INTERNAL)) {
DRM_INFO("Special thermal controller config\n");
} else {
DRM_INFO("Possible %s thermal controller at 0x%02x %s fan control\n",
pp_lib_thermal_controller_names[controller->ucType],
controller->ucI2cAddress >> 1,
(controller->ucFanParameters &
ATOM_PP_FANPARAMETERS_NOFAN) ? "without" : "with");
i2c_bus = radeon_lookup_i2c_gpio(rdev, controller->ucI2cLine);
rdev->pm.i2c_bus = radeon_i2c_lookup(rdev, &i2c_bus);
if (rdev->pm.i2c_bus) {
struct i2c_board_info info = { };
const char *name = pp_lib_thermal_controller_names[controller->ucType];
info.addr = controller->ucI2cAddress >> 1;
strlcpy(info.type, name, sizeof(info.type));
i2c_new_device(&rdev->pm.i2c_bus->adapter, &info);
}
}
}
}
static u16 radeon_atombios_get_default_vddc(struct radeon_device *rdev)
{
struct radeon_mode_info *mode_info = &rdev->mode_info;
int index = GetIndexIntoMasterTable(DATA, FirmwareInfo);
u8 frev, crev;
u16 data_offset;
union firmware_info *firmware_info;
u16 vddc = 0;
if (atom_parse_data_header(mode_info->atom_context, index, NULL,
&frev, &crev, &data_offset)) {
firmware_info =
(union firmware_info *)(mode_info->atom_context->bios +
data_offset);
vddc = firmware_info->info_14.usBootUpVDDCVoltage;
}
return vddc;
}
static void radeon_atombios_parse_pplib_non_clock_info(struct radeon_device *rdev,
int state_index, int mode_index,
struct _ATOM_PPLIB_NONCLOCK_INFO *non_clock_info)
{
int j;
u32 misc = le32_to_cpu(non_clock_info->ulCapsAndSettings);
u32 misc2 = le16_to_cpu(non_clock_info->usClassification);
u16 vddc = radeon_atombios_get_default_vddc(rdev);
rdev->pm.power_state[state_index].misc = misc;
rdev->pm.power_state[state_index].misc2 = misc2;
rdev->pm.power_state[state_index].pcie_lanes =
((misc & ATOM_PPLIB_PCIE_LINK_WIDTH_MASK) >>
ATOM_PPLIB_PCIE_LINK_WIDTH_SHIFT) + 1;
switch (misc2 & ATOM_PPLIB_CLASSIFICATION_UI_MASK) {
case ATOM_PPLIB_CLASSIFICATION_UI_BATTERY:
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_BATTERY;
break;
case ATOM_PPLIB_CLASSIFICATION_UI_BALANCED:
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_BALANCED;
break;
case ATOM_PPLIB_CLASSIFICATION_UI_PERFORMANCE:
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_PERFORMANCE;
break;
case ATOM_PPLIB_CLASSIFICATION_UI_NONE:
if (misc2 & ATOM_PPLIB_CLASSIFICATION_3DPERFORMANCE)
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_PERFORMANCE;
break;
}
rdev->pm.power_state[state_index].flags = 0;
if (misc & ATOM_PPLIB_SINGLE_DISPLAY_ONLY)
rdev->pm.power_state[state_index].flags |=
RADEON_PM_STATE_SINGLE_DISPLAY_ONLY;
if (misc2 & ATOM_PPLIB_CLASSIFICATION_BOOT) {
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_DEFAULT;
rdev->pm.default_power_state_index = state_index;
rdev->pm.power_state[state_index].default_clock_mode =
&rdev->pm.power_state[state_index].clock_info[mode_index - 1];
/* patch the table values with the default slck/mclk from firmware info */
for (j = 0; j < mode_index; j++) {
rdev->pm.power_state[state_index].clock_info[j].mclk =
rdev->clock.default_mclk;
rdev->pm.power_state[state_index].clock_info[j].sclk =
rdev->clock.default_sclk;
if (vddc)
rdev->pm.power_state[state_index].clock_info[j].voltage.voltage =
vddc;
}
}
}
static bool radeon_atombios_parse_pplib_clock_info(struct radeon_device *rdev,
int state_index, int mode_index,
union pplib_clock_info *clock_info)
{
u32 sclk, mclk;
if (rdev->flags & RADEON_IS_IGP) {
if (rdev->family >= CHIP_PALM) {
sclk = le16_to_cpu(clock_info->sumo.usEngineClockLow);
sclk |= clock_info->sumo.ucEngineClockHigh << 16;
rdev->pm.power_state[state_index].clock_info[mode_index].sclk = sclk;
} else {
sclk = le16_to_cpu(clock_info->rs780.usLowEngineClockLow);
sclk |= clock_info->rs780.ucLowEngineClockHigh << 16;
rdev->pm.power_state[state_index].clock_info[mode_index].sclk = sclk;
}
} else if (ASIC_IS_DCE4(rdev)) {
sclk = le16_to_cpu(clock_info->evergreen.usEngineClockLow);
sclk |= clock_info->evergreen.ucEngineClockHigh << 16;
mclk = le16_to_cpu(clock_info->evergreen.usMemoryClockLow);
mclk |= clock_info->evergreen.ucMemoryClockHigh << 16;
rdev->pm.power_state[state_index].clock_info[mode_index].mclk = mclk;
rdev->pm.power_state[state_index].clock_info[mode_index].sclk = sclk;
rdev->pm.power_state[state_index].clock_info[mode_index].voltage.type =
VOLTAGE_SW;
rdev->pm.power_state[state_index].clock_info[mode_index].voltage.voltage =
clock_info->evergreen.usVDDC;
} else {
sclk = le16_to_cpu(clock_info->r600.usEngineClockLow);
sclk |= clock_info->r600.ucEngineClockHigh << 16;
mclk = le16_to_cpu(clock_info->r600.usMemoryClockLow);
mclk |= clock_info->r600.ucMemoryClockHigh << 16;
rdev->pm.power_state[state_index].clock_info[mode_index].mclk = mclk;
rdev->pm.power_state[state_index].clock_info[mode_index].sclk = sclk;
rdev->pm.power_state[state_index].clock_info[mode_index].voltage.type =
VOLTAGE_SW;
rdev->pm.power_state[state_index].clock_info[mode_index].voltage.voltage =
clock_info->r600.usVDDC;
}
if (rdev->flags & RADEON_IS_IGP) {
/* skip invalid modes */
if (rdev->pm.power_state[state_index].clock_info[mode_index].sclk == 0)
return false;
} else {
/* skip invalid modes */
if ((rdev->pm.power_state[state_index].clock_info[mode_index].mclk == 0) ||
(rdev->pm.power_state[state_index].clock_info[mode_index].sclk == 0))
return false;
}
return true;
}
static int radeon_atombios_parse_power_table_4_5(struct radeon_device *rdev)
{
struct radeon_mode_info *mode_info = &rdev->mode_info;
struct _ATOM_PPLIB_NONCLOCK_INFO *non_clock_info;
union pplib_power_state *power_state;
int i, j;
int state_index = 0, mode_index = 0;
union pplib_clock_info *clock_info;
bool valid;
union power_info *power_info;
int index = GetIndexIntoMasterTable(DATA, PowerPlayInfo);
u16 data_offset;
u8 frev, crev;
if (!atom_parse_data_header(mode_info->atom_context, index, NULL,
&frev, &crev, &data_offset))
return state_index;
power_info = (union power_info *)(mode_info->atom_context->bios + data_offset);
radeon_atombios_add_pplib_thermal_controller(rdev, &power_info->pplib.sThermalController);
/* first mode is usually default, followed by low to high */
for (i = 0; i < power_info->pplib.ucNumStates; i++) {
mode_index = 0;
power_state = (union pplib_power_state *)
(mode_info->atom_context->bios + data_offset +
le16_to_cpu(power_info->pplib.usStateArrayOffset) +
i * power_info->pplib.ucStateEntrySize);
non_clock_info = (struct _ATOM_PPLIB_NONCLOCK_INFO *)
(mode_info->atom_context->bios + data_offset +
le16_to_cpu(power_info->pplib.usNonClockInfoArrayOffset) +
(power_state->v1.ucNonClockStateIndex *
power_info->pplib.ucNonClockSize));
for (j = 0; j < (power_info->pplib.ucStateEntrySize - 1); j++) {
clock_info = (union pplib_clock_info *)
(mode_info->atom_context->bios + data_offset +
le16_to_cpu(power_info->pplib.usClockInfoArrayOffset) +
(power_state->v1.ucClockStateIndices[j] *
power_info->pplib.ucClockInfoSize));
valid = radeon_atombios_parse_pplib_clock_info(rdev,
state_index, mode_index,
clock_info);
if (valid)
mode_index++;
}
rdev->pm.power_state[state_index].num_clock_modes = mode_index;
if (mode_index) {
radeon_atombios_parse_pplib_non_clock_info(rdev, state_index, mode_index,
non_clock_info);
state_index++;
}
}
/* if multiple clock modes, mark the lowest as no display */
for (i = 0; i < state_index; i++) {
if (rdev->pm.power_state[i].num_clock_modes > 1)
rdev->pm.power_state[i].clock_info[0].flags |=
RADEON_PM_MODE_NO_DISPLAY;
}
/* first mode is usually default */
if (rdev->pm.default_power_state_index == -1) {
rdev->pm.power_state[0].type =
POWER_STATE_TYPE_DEFAULT;
rdev->pm.default_power_state_index = 0;
rdev->pm.power_state[0].default_clock_mode =
&rdev->pm.power_state[0].clock_info[0];
}
return state_index;
}
static int radeon_atombios_parse_power_table_6(struct radeon_device *rdev)
{
struct radeon_mode_info *mode_info = &rdev->mode_info;
struct _ATOM_PPLIB_NONCLOCK_INFO *non_clock_info;
union pplib_power_state *power_state;
int i, j, non_clock_array_index, clock_array_index;
int state_index = 0, mode_index = 0;
union pplib_clock_info *clock_info;
struct StateArray *state_array;
struct ClockInfoArray *clock_info_array;
struct NonClockInfoArray *non_clock_info_array;
bool valid;
union power_info *power_info;
int index = GetIndexIntoMasterTable(DATA, PowerPlayInfo);
u16 data_offset;
u8 frev, crev;
if (!atom_parse_data_header(mode_info->atom_context, index, NULL,
&frev, &crev, &data_offset))
return state_index;
power_info = (union power_info *)(mode_info->atom_context->bios + data_offset);
radeon_atombios_add_pplib_thermal_controller(rdev, &power_info->pplib.sThermalController);
state_array = (struct StateArray *)
(mode_info->atom_context->bios + data_offset +
power_info->pplib.usStateArrayOffset);
clock_info_array = (struct ClockInfoArray *)
(mode_info->atom_context->bios + data_offset +
power_info->pplib.usClockInfoArrayOffset);
non_clock_info_array = (struct NonClockInfoArray *)
(mode_info->atom_context->bios + data_offset +
power_info->pplib.usNonClockInfoArrayOffset);
for (i = 0; i < state_array->ucNumEntries; i++) {
mode_index = 0;
power_state = (union pplib_power_state *)&state_array->states[i];
/* XXX this might be an inagua bug... */
non_clock_array_index = i; /* power_state->v2.nonClockInfoIndex */
non_clock_info = (struct _ATOM_PPLIB_NONCLOCK_INFO *)
&non_clock_info_array->nonClockInfo[non_clock_array_index];
for (j = 0; j < power_state->v2.ucNumDPMLevels; j++) {
clock_array_index = power_state->v2.clockInfoIndex[j];
/* XXX this might be an inagua bug... */
if (clock_array_index >= clock_info_array->ucNumEntries)
continue;
clock_info = (union pplib_clock_info *)
&clock_info_array->clockInfo[clock_array_index];
valid = radeon_atombios_parse_pplib_clock_info(rdev,
state_index, mode_index,
clock_info);
if (valid)
mode_index++;
}
rdev->pm.power_state[state_index].num_clock_modes = mode_index;
if (mode_index) {
radeon_atombios_parse_pplib_non_clock_info(rdev, state_index, mode_index,
non_clock_info);
state_index++;
}
}
/* if multiple clock modes, mark the lowest as no display */
for (i = 0; i < state_index; i++) {
if (rdev->pm.power_state[i].num_clock_modes > 1)
rdev->pm.power_state[i].clock_info[0].flags |=
RADEON_PM_MODE_NO_DISPLAY;
}
/* first mode is usually default */
if (rdev->pm.default_power_state_index == -1) {
rdev->pm.power_state[0].type =
POWER_STATE_TYPE_DEFAULT;
rdev->pm.default_power_state_index = 0;
rdev->pm.power_state[0].default_clock_mode =
&rdev->pm.power_state[0].clock_info[0];
}
return state_index;
}
void radeon_atombios_get_power_modes(struct radeon_device *rdev)
{
struct radeon_mode_info *mode_info = &rdev->mode_info;
int index = GetIndexIntoMasterTable(DATA, PowerPlayInfo);
u16 data_offset;
u8 frev, crev;
int state_index = 0;
rdev->pm.default_power_state_index = -1;
if (atom_parse_data_header(mode_info->atom_context, index, NULL,
&frev, &crev, &data_offset)) {
switch (frev) {
case 1:
case 2:
case 3:
state_index = radeon_atombios_parse_power_table_1_3(rdev);
break;
case 4:
case 5:
state_index = radeon_atombios_parse_power_table_4_5(rdev);
break;
case 6:
state_index = radeon_atombios_parse_power_table_6(rdev);
break;
default:
break;
}
} else {
/* add the default mode */
rdev->pm.power_state[state_index].type =
POWER_STATE_TYPE_DEFAULT;
rdev->pm.power_state[state_index].num_clock_modes = 1;
rdev->pm.power_state[state_index].clock_info[0].mclk = rdev->clock.default_mclk;
rdev->pm.power_state[state_index].clock_info[0].sclk = rdev->clock.default_sclk;
rdev->pm.power_state[state_index].default_clock_mode =
&rdev->pm.power_state[state_index].clock_info[0];
rdev->pm.power_state[state_index].clock_info[0].voltage.type = VOLTAGE_NONE;
rdev->pm.power_state[state_index].pcie_lanes = 16;
rdev->pm.default_power_state_index = state_index;
rdev->pm.power_state[state_index].flags = 0;
state_index++;
}
rdev->pm.num_power_states = state_index;
rdev->pm.current_power_state_index = rdev->pm.default_power_state_index;
rdev->pm.current_clock_mode_index = 0;
rdev->pm.current_vddc = rdev->pm.power_state[rdev->pm.default_power_state_index].clock_info[0].voltage.voltage;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
}
void radeon_atom_set_clock_gating(struct radeon_device *rdev, int enable)
{
DYNAMIC_CLOCK_GATING_PS_ALLOCATION args;
int index = GetIndexIntoMasterTable(COMMAND, DynamicClockGating);
args.ucEnable = enable;
atom_execute_table(rdev->mode_info.atom_context, index, (uint32_t *)&args);
}
uint32_t radeon_atom_get_engine_clock(struct radeon_device *rdev)
{
GET_ENGINE_CLOCK_PS_ALLOCATION args;
int index = GetIndexIntoMasterTable(COMMAND, GetEngineClock);
atom_execute_table(rdev->mode_info.atom_context, index, (uint32_t *)&args);
return args.ulReturnEngineClock;
}
uint32_t radeon_atom_get_memory_clock(struct radeon_device *rdev)
{
GET_MEMORY_CLOCK_PS_ALLOCATION args;
int index = GetIndexIntoMasterTable(COMMAND, GetMemoryClock);
atom_execute_table(rdev->mode_info.atom_context, index, (uint32_t *)&args);
return args.ulReturnMemoryClock;
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
void radeon_atom_set_engine_clock(struct radeon_device *rdev,
uint32_t eng_clock)
{
SET_ENGINE_CLOCK_PS_ALLOCATION args;
int index = GetIndexIntoMasterTable(COMMAND, SetEngineClock);
args.ulTargetEngineClock = eng_clock; /* 10 khz */
atom_execute_table(rdev->mode_info.atom_context, index, (uint32_t *)&args);
}
void radeon_atom_set_memory_clock(struct radeon_device *rdev,
uint32_t mem_clock)
{
SET_MEMORY_CLOCK_PS_ALLOCATION args;
int index = GetIndexIntoMasterTable(COMMAND, SetMemoryClock);
if (rdev->flags & RADEON_IS_IGP)
return;
args.ulTargetMemoryClock = mem_clock; /* 10 khz */
atom_execute_table(rdev->mode_info.atom_context, index, (uint32_t *)&args);
}
union set_voltage {
struct _SET_VOLTAGE_PS_ALLOCATION alloc;
struct _SET_VOLTAGE_PARAMETERS v1;
struct _SET_VOLTAGE_PARAMETERS_V2 v2;
};
void radeon_atom_set_voltage(struct radeon_device *rdev, u16 level)
{
union set_voltage args;
int index = GetIndexIntoMasterTable(COMMAND, SetVoltage);
u8 frev, crev, volt_index = level;
if (!atom_parse_cmd_header(rdev->mode_info.atom_context, index, &frev, &crev))
return;
switch (crev) {
case 1:
args.v1.ucVoltageType = SET_VOLTAGE_TYPE_ASIC_VDDC;
args.v1.ucVoltageMode = SET_ASIC_VOLTAGE_MODE_ALL_SOURCE;
args.v1.ucVoltageIndex = volt_index;
break;
case 2:
args.v2.ucVoltageType = SET_VOLTAGE_TYPE_ASIC_VDDC;
args.v2.ucVoltageMode = SET_ASIC_VOLTAGE_MODE_SET_VOLTAGE;
args.v2.usVoltageLevel = cpu_to_le16(level);
break;
default:
DRM_ERROR("Unknown table version %d, %d\n", frev, crev);
return;
}
atom_execute_table(rdev->mode_info.atom_context, index, (uint32_t *)&args);
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
void radeon_atom_initialize_bios_scratch_regs(struct drm_device *dev)
{
struct radeon_device *rdev = dev->dev_private;
uint32_t bios_2_scratch, bios_6_scratch;
if (rdev->family >= CHIP_R600) {
bios_2_scratch = RREG32(R600_BIOS_2_SCRATCH);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_6_scratch = RREG32(R600_BIOS_6_SCRATCH);
} else {
bios_2_scratch = RREG32(RADEON_BIOS_2_SCRATCH);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_6_scratch = RREG32(RADEON_BIOS_6_SCRATCH);
}
/* let the bios control the backlight */
bios_2_scratch &= ~ATOM_S2_VRI_BRIGHT_ENABLE;
/* tell the bios not to handle mode switching */
bios_6_scratch |= (ATOM_S6_ACC_BLOCK_DISPLAY_SWITCH | ATOM_S6_ACC_MODE);
if (rdev->family >= CHIP_R600) {
WREG32(R600_BIOS_2_SCRATCH, bios_2_scratch);
WREG32(R600_BIOS_6_SCRATCH, bios_6_scratch);
} else {
WREG32(RADEON_BIOS_2_SCRATCH, bios_2_scratch);
WREG32(RADEON_BIOS_6_SCRATCH, bios_6_scratch);
}
}
void radeon_save_bios_scratch_regs(struct radeon_device *rdev)
{
uint32_t scratch_reg;
int i;
if (rdev->family >= CHIP_R600)
scratch_reg = R600_BIOS_0_SCRATCH;
else
scratch_reg = RADEON_BIOS_0_SCRATCH;
for (i = 0; i < RADEON_BIOS_NUM_SCRATCH; i++)
rdev->bios_scratch[i] = RREG32(scratch_reg + (i * 4));
}
void radeon_restore_bios_scratch_regs(struct radeon_device *rdev)
{
uint32_t scratch_reg;
int i;
if (rdev->family >= CHIP_R600)
scratch_reg = R600_BIOS_0_SCRATCH;
else
scratch_reg = RADEON_BIOS_0_SCRATCH;
for (i = 0; i < RADEON_BIOS_NUM_SCRATCH; i++)
WREG32(scratch_reg + (i * 4), rdev->bios_scratch[i]);
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
void radeon_atom_output_lock(struct drm_encoder *encoder, bool lock)
{
struct drm_device *dev = encoder->dev;
struct radeon_device *rdev = dev->dev_private;
uint32_t bios_6_scratch;
if (rdev->family >= CHIP_R600)
bios_6_scratch = RREG32(R600_BIOS_6_SCRATCH);
else
bios_6_scratch = RREG32(RADEON_BIOS_6_SCRATCH);
if (lock)
bios_6_scratch |= ATOM_S6_CRITICAL_STATE;
else
bios_6_scratch &= ~ATOM_S6_CRITICAL_STATE;
if (rdev->family >= CHIP_R600)
WREG32(R600_BIOS_6_SCRATCH, bios_6_scratch);
else
WREG32(RADEON_BIOS_6_SCRATCH, bios_6_scratch);
}
/* at some point we may want to break this out into individual functions */
void
radeon_atombios_connected_scratch_regs(struct drm_connector *connector,
struct drm_encoder *encoder,
bool connected)
{
struct drm_device *dev = connector->dev;
struct radeon_device *rdev = dev->dev_private;
struct radeon_connector *radeon_connector =
to_radeon_connector(connector);
struct radeon_encoder *radeon_encoder = to_radeon_encoder(encoder);
uint32_t bios_0_scratch, bios_3_scratch, bios_6_scratch;
if (rdev->family >= CHIP_R600) {
bios_0_scratch = RREG32(R600_BIOS_0_SCRATCH);
bios_3_scratch = RREG32(R600_BIOS_3_SCRATCH);
bios_6_scratch = RREG32(R600_BIOS_6_SCRATCH);
} else {
bios_0_scratch = RREG32(RADEON_BIOS_0_SCRATCH);
bios_3_scratch = RREG32(RADEON_BIOS_3_SCRATCH);
bios_6_scratch = RREG32(RADEON_BIOS_6_SCRATCH);
}
if ((radeon_encoder->devices & ATOM_DEVICE_TV1_SUPPORT) &&
(radeon_connector->devices & ATOM_DEVICE_TV1_SUPPORT)) {
if (connected) {
DRM_DEBUG_KMS("TV1 connected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_3_scratch |= ATOM_S3_TV1_ACTIVE;
bios_6_scratch |= ATOM_S6_ACC_REQ_TV1;
} else {
DRM_DEBUG_KMS("TV1 disconnected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch &= ~ATOM_S0_TV1_MASK;
bios_3_scratch &= ~ATOM_S3_TV1_ACTIVE;
bios_6_scratch &= ~ATOM_S6_ACC_REQ_TV1;
}
}
if ((radeon_encoder->devices & ATOM_DEVICE_CV_SUPPORT) &&
(radeon_connector->devices & ATOM_DEVICE_CV_SUPPORT)) {
if (connected) {
DRM_DEBUG_KMS("CV connected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_3_scratch |= ATOM_S3_CV_ACTIVE;
bios_6_scratch |= ATOM_S6_ACC_REQ_CV;
} else {
DRM_DEBUG_KMS("CV disconnected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch &= ~ATOM_S0_CV_MASK;
bios_3_scratch &= ~ATOM_S3_CV_ACTIVE;
bios_6_scratch &= ~ATOM_S6_ACC_REQ_CV;
}
}
if ((radeon_encoder->devices & ATOM_DEVICE_LCD1_SUPPORT) &&
(radeon_connector->devices & ATOM_DEVICE_LCD1_SUPPORT)) {
if (connected) {
DRM_DEBUG_KMS("LCD1 connected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch |= ATOM_S0_LCD1;
bios_3_scratch |= ATOM_S3_LCD1_ACTIVE;
bios_6_scratch |= ATOM_S6_ACC_REQ_LCD1;
} else {
DRM_DEBUG_KMS("LCD1 disconnected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch &= ~ATOM_S0_LCD1;
bios_3_scratch &= ~ATOM_S3_LCD1_ACTIVE;
bios_6_scratch &= ~ATOM_S6_ACC_REQ_LCD1;
}
}
if ((radeon_encoder->devices & ATOM_DEVICE_CRT1_SUPPORT) &&
(radeon_connector->devices & ATOM_DEVICE_CRT1_SUPPORT)) {
if (connected) {
DRM_DEBUG_KMS("CRT1 connected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch |= ATOM_S0_CRT1_COLOR;
bios_3_scratch |= ATOM_S3_CRT1_ACTIVE;
bios_6_scratch |= ATOM_S6_ACC_REQ_CRT1;
} else {
DRM_DEBUG_KMS("CRT1 disconnected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch &= ~ATOM_S0_CRT1_MASK;
bios_3_scratch &= ~ATOM_S3_CRT1_ACTIVE;
bios_6_scratch &= ~ATOM_S6_ACC_REQ_CRT1;
}
}
if ((radeon_encoder->devices & ATOM_DEVICE_CRT2_SUPPORT) &&
(radeon_connector->devices & ATOM_DEVICE_CRT2_SUPPORT)) {
if (connected) {
DRM_DEBUG_KMS("CRT2 connected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch |= ATOM_S0_CRT2_COLOR;
bios_3_scratch |= ATOM_S3_CRT2_ACTIVE;
bios_6_scratch |= ATOM_S6_ACC_REQ_CRT2;
} else {
DRM_DEBUG_KMS("CRT2 disconnected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch &= ~ATOM_S0_CRT2_MASK;
bios_3_scratch &= ~ATOM_S3_CRT2_ACTIVE;
bios_6_scratch &= ~ATOM_S6_ACC_REQ_CRT2;
}
}
if ((radeon_encoder->devices & ATOM_DEVICE_DFP1_SUPPORT) &&
(radeon_connector->devices & ATOM_DEVICE_DFP1_SUPPORT)) {
if (connected) {
DRM_DEBUG_KMS("DFP1 connected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch |= ATOM_S0_DFP1;
bios_3_scratch |= ATOM_S3_DFP1_ACTIVE;
bios_6_scratch |= ATOM_S6_ACC_REQ_DFP1;
} else {
DRM_DEBUG_KMS("DFP1 disconnected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch &= ~ATOM_S0_DFP1;
bios_3_scratch &= ~ATOM_S3_DFP1_ACTIVE;
bios_6_scratch &= ~ATOM_S6_ACC_REQ_DFP1;
}
}
if ((radeon_encoder->devices & ATOM_DEVICE_DFP2_SUPPORT) &&
(radeon_connector->devices & ATOM_DEVICE_DFP2_SUPPORT)) {
if (connected) {
DRM_DEBUG_KMS("DFP2 connected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch |= ATOM_S0_DFP2;
bios_3_scratch |= ATOM_S3_DFP2_ACTIVE;
bios_6_scratch |= ATOM_S6_ACC_REQ_DFP2;
} else {
DRM_DEBUG_KMS("DFP2 disconnected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch &= ~ATOM_S0_DFP2;
bios_3_scratch &= ~ATOM_S3_DFP2_ACTIVE;
bios_6_scratch &= ~ATOM_S6_ACC_REQ_DFP2;
}
}
if ((radeon_encoder->devices & ATOM_DEVICE_DFP3_SUPPORT) &&
(radeon_connector->devices & ATOM_DEVICE_DFP3_SUPPORT)) {
if (connected) {
DRM_DEBUG_KMS("DFP3 connected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch |= ATOM_S0_DFP3;
bios_3_scratch |= ATOM_S3_DFP3_ACTIVE;
bios_6_scratch |= ATOM_S6_ACC_REQ_DFP3;
} else {
DRM_DEBUG_KMS("DFP3 disconnected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch &= ~ATOM_S0_DFP3;
bios_3_scratch &= ~ATOM_S3_DFP3_ACTIVE;
bios_6_scratch &= ~ATOM_S6_ACC_REQ_DFP3;
}
}
if ((radeon_encoder->devices & ATOM_DEVICE_DFP4_SUPPORT) &&
(radeon_connector->devices & ATOM_DEVICE_DFP4_SUPPORT)) {
if (connected) {
DRM_DEBUG_KMS("DFP4 connected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch |= ATOM_S0_DFP4;
bios_3_scratch |= ATOM_S3_DFP4_ACTIVE;
bios_6_scratch |= ATOM_S6_ACC_REQ_DFP4;
} else {
DRM_DEBUG_KMS("DFP4 disconnected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch &= ~ATOM_S0_DFP4;
bios_3_scratch &= ~ATOM_S3_DFP4_ACTIVE;
bios_6_scratch &= ~ATOM_S6_ACC_REQ_DFP4;
}
}
if ((radeon_encoder->devices & ATOM_DEVICE_DFP5_SUPPORT) &&
(radeon_connector->devices & ATOM_DEVICE_DFP5_SUPPORT)) {
if (connected) {
DRM_DEBUG_KMS("DFP5 connected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch |= ATOM_S0_DFP5;
bios_3_scratch |= ATOM_S3_DFP5_ACTIVE;
bios_6_scratch |= ATOM_S6_ACC_REQ_DFP5;
} else {
DRM_DEBUG_KMS("DFP5 disconnected\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 20:42:42 +08:00
bios_0_scratch &= ~ATOM_S0_DFP5;
bios_3_scratch &= ~ATOM_S3_DFP5_ACTIVE;
bios_6_scratch &= ~ATOM_S6_ACC_REQ_DFP5;
}
}
if (rdev->family >= CHIP_R600) {
WREG32(R600_BIOS_0_SCRATCH, bios_0_scratch);
WREG32(R600_BIOS_3_SCRATCH, bios_3_scratch);
WREG32(R600_BIOS_6_SCRATCH, bios_6_scratch);
} else {
WREG32(RADEON_BIOS_0_SCRATCH, bios_0_scratch);
WREG32(RADEON_BIOS_3_SCRATCH, bios_3_scratch);
WREG32(RADEON_BIOS_6_SCRATCH, bios_6_scratch);
}
}
void
radeon_atombios_encoder_crtc_scratch_regs(struct drm_encoder *encoder, int crtc)
{
struct drm_device *dev = encoder->dev;
struct radeon_device *rdev = dev->dev_private;
struct radeon_encoder *radeon_encoder = to_radeon_encoder(encoder);
uint32_t bios_3_scratch;
if (rdev->family >= CHIP_R600)
bios_3_scratch = RREG32(R600_BIOS_3_SCRATCH);
else
bios_3_scratch = RREG32(RADEON_BIOS_3_SCRATCH);
if (radeon_encoder->devices & ATOM_DEVICE_TV1_SUPPORT) {
bios_3_scratch &= ~ATOM_S3_TV1_CRTC_ACTIVE;
bios_3_scratch |= (crtc << 18);
}
if (radeon_encoder->devices & ATOM_DEVICE_CV_SUPPORT) {
bios_3_scratch &= ~ATOM_S3_CV_CRTC_ACTIVE;
bios_3_scratch |= (crtc << 24);
}
if (radeon_encoder->devices & ATOM_DEVICE_CRT1_SUPPORT) {
bios_3_scratch &= ~ATOM_S3_CRT1_CRTC_ACTIVE;
bios_3_scratch |= (crtc << 16);
}
if (radeon_encoder->devices & ATOM_DEVICE_CRT2_SUPPORT) {
bios_3_scratch &= ~ATOM_S3_CRT2_CRTC_ACTIVE;
bios_3_scratch |= (crtc << 20);
}
if (radeon_encoder->devices & ATOM_DEVICE_LCD1_SUPPORT) {
bios_3_scratch &= ~ATOM_S3_LCD1_CRTC_ACTIVE;
bios_3_scratch |= (crtc << 17);
}
if (radeon_encoder->devices & ATOM_DEVICE_DFP1_SUPPORT) {
bios_3_scratch &= ~ATOM_S3_DFP1_CRTC_ACTIVE;
bios_3_scratch |= (crtc << 19);
}
if (radeon_encoder->devices & ATOM_DEVICE_DFP2_SUPPORT) {
bios_3_scratch &= ~ATOM_S3_DFP2_CRTC_ACTIVE;
bios_3_scratch |= (crtc << 23);
}
if (radeon_encoder->devices & ATOM_DEVICE_DFP3_SUPPORT) {
bios_3_scratch &= ~ATOM_S3_DFP3_CRTC_ACTIVE;
bios_3_scratch |= (crtc << 25);
}
if (rdev->family >= CHIP_R600)
WREG32(R600_BIOS_3_SCRATCH, bios_3_scratch);
else
WREG32(RADEON_BIOS_3_SCRATCH, bios_3_scratch);
}
void
radeon_atombios_encoder_dpms_scratch_regs(struct drm_encoder *encoder, bool on)
{
struct drm_device *dev = encoder->dev;
struct radeon_device *rdev = dev->dev_private;
struct radeon_encoder *radeon_encoder = to_radeon_encoder(encoder);
uint32_t bios_2_scratch;
if (rdev->family >= CHIP_R600)
bios_2_scratch = RREG32(R600_BIOS_2_SCRATCH);
else
bios_2_scratch = RREG32(RADEON_BIOS_2_SCRATCH);
if (radeon_encoder->devices & ATOM_DEVICE_TV1_SUPPORT) {
if (on)
bios_2_scratch &= ~ATOM_S2_TV1_DPMS_STATE;
else
bios_2_scratch |= ATOM_S2_TV1_DPMS_STATE;
}
if (radeon_encoder->devices & ATOM_DEVICE_CV_SUPPORT) {
if (on)
bios_2_scratch &= ~ATOM_S2_CV_DPMS_STATE;
else
bios_2_scratch |= ATOM_S2_CV_DPMS_STATE;
}
if (radeon_encoder->devices & ATOM_DEVICE_CRT1_SUPPORT) {
if (on)
bios_2_scratch &= ~ATOM_S2_CRT1_DPMS_STATE;
else
bios_2_scratch |= ATOM_S2_CRT1_DPMS_STATE;
}
if (radeon_encoder->devices & ATOM_DEVICE_CRT2_SUPPORT) {
if (on)
bios_2_scratch &= ~ATOM_S2_CRT2_DPMS_STATE;
else
bios_2_scratch |= ATOM_S2_CRT2_DPMS_STATE;
}
if (radeon_encoder->devices & ATOM_DEVICE_LCD1_SUPPORT) {
if (on)
bios_2_scratch &= ~ATOM_S2_LCD1_DPMS_STATE;
else
bios_2_scratch |= ATOM_S2_LCD1_DPMS_STATE;
}
if (radeon_encoder->devices & ATOM_DEVICE_DFP1_SUPPORT) {
if (on)
bios_2_scratch &= ~ATOM_S2_DFP1_DPMS_STATE;
else
bios_2_scratch |= ATOM_S2_DFP1_DPMS_STATE;
}
if (radeon_encoder->devices & ATOM_DEVICE_DFP2_SUPPORT) {
if (on)
bios_2_scratch &= ~ATOM_S2_DFP2_DPMS_STATE;
else
bios_2_scratch |= ATOM_S2_DFP2_DPMS_STATE;
}
if (radeon_encoder->devices & ATOM_DEVICE_DFP3_SUPPORT) {
if (on)
bios_2_scratch &= ~ATOM_S2_DFP3_DPMS_STATE;
else
bios_2_scratch |= ATOM_S2_DFP3_DPMS_STATE;
}
if (radeon_encoder->devices & ATOM_DEVICE_DFP4_SUPPORT) {
if (on)
bios_2_scratch &= ~ATOM_S2_DFP4_DPMS_STATE;
else
bios_2_scratch |= ATOM_S2_DFP4_DPMS_STATE;
}
if (radeon_encoder->devices & ATOM_DEVICE_DFP5_SUPPORT) {
if (on)
bios_2_scratch &= ~ATOM_S2_DFP5_DPMS_STATE;
else
bios_2_scratch |= ATOM_S2_DFP5_DPMS_STATE;
}
if (rdev->family >= CHIP_R600)
WREG32(R600_BIOS_2_SCRATCH, bios_2_scratch);
else
WREG32(RADEON_BIOS_2_SCRATCH, bios_2_scratch);
}