OpenCloudOS-Kernel/drivers/dma-buf/dma-buf.c

1318 lines
36 KiB
C
Raw Permalink Normal View History

// SPDX-License-Identifier: GPL-2.0-only
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
/*
* Framework for buffer objects that can be shared across devices/subsystems.
*
* Copyright(C) 2011 Linaro Limited. All rights reserved.
* Author: Sumit Semwal <sumit.semwal@ti.com>
*
* Many thanks to linaro-mm-sig list, and specially
* Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and
* Daniel Vetter <daniel@ffwll.ch> for their support in creation and
* refining of this idea.
*/
#include <linux/fs.h>
#include <linux/slab.h>
#include <linux/dma-buf.h>
dma-buf: Rename struct fence to dma_fence I plan to usurp the short name of struct fence for a core kernel struct, and so I need to rename the specialised fence/timeline for DMA operations to make room. A consensus was reached in https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html that making clear this fence applies to DMA operations was a good thing. Since then the patch has grown a bit as usage increases, so hopefully it remains a good thing! (v2...: rebase, rerun spatch) v3: Compile on msm, spotted a manual fixup that I broke. v4: Try again for msm, sorry Daniel coccinelle script: @@ @@ - struct fence + struct dma_fence @@ @@ - struct fence_ops + struct dma_fence_ops @@ @@ - struct fence_cb + struct dma_fence_cb @@ @@ - struct fence_array + struct dma_fence_array @@ @@ - enum fence_flag_bits + enum dma_fence_flag_bits @@ @@ ( - fence_init + dma_fence_init | - fence_release + dma_fence_release | - fence_free + dma_fence_free | - fence_get + dma_fence_get | - fence_get_rcu + dma_fence_get_rcu | - fence_put + dma_fence_put | - fence_signal + dma_fence_signal | - fence_signal_locked + dma_fence_signal_locked | - fence_default_wait + dma_fence_default_wait | - fence_add_callback + dma_fence_add_callback | - fence_remove_callback + dma_fence_remove_callback | - fence_enable_sw_signaling + dma_fence_enable_sw_signaling | - fence_is_signaled_locked + dma_fence_is_signaled_locked | - fence_is_signaled + dma_fence_is_signaled | - fence_is_later + dma_fence_is_later | - fence_later + dma_fence_later | - fence_wait_timeout + dma_fence_wait_timeout | - fence_wait_any_timeout + dma_fence_wait_any_timeout | - fence_wait + dma_fence_wait | - fence_context_alloc + dma_fence_context_alloc | - fence_array_create + dma_fence_array_create | - to_fence_array + to_dma_fence_array | - fence_is_array + dma_fence_is_array | - trace_fence_emit + trace_dma_fence_emit | - FENCE_TRACE + DMA_FENCE_TRACE | - FENCE_WARN + DMA_FENCE_WARN | - FENCE_ERR + DMA_FENCE_ERR ) ( ... ) Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk> Acked-by: Sumit Semwal <sumit.semwal@linaro.org> Acked-by: Christian König <christian.koenig@amd.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk
2016-10-25 20:00:45 +08:00
#include <linux/dma-fence.h>
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
#include <linux/anon_inodes.h>
#include <linux/export.h>
#include <linux/debugfs.h>
#include <linux/module.h>
#include <linux/seq_file.h>
#include <linux/poll.h>
#include <linux/dma-resv.h>
#include <linux/mm.h>
#include <linux/mount.h>
#include <linux/pseudo_fs.h>
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
dma-buf: Add ioctls to allow userspace to flush The userspace might need some sort of cache coherency management e.g. when CPU and GPU domains are being accessed through dma-buf at the same time. To circumvent this problem there are begin/end coherency markers, that forward directly to existing dma-buf device drivers vfunc hooks. Userspace can make use of those markers through the DMA_BUF_IOCTL_SYNC ioctl. The sequence would be used like following: - mmap dma-buf fd - for each drawing/upload cycle in CPU 1. SYNC_START ioctl, 2. read/write to mmap area 3. SYNC_END ioctl. This can be repeated as often as you want (with the new data being consumed by the GPU or say scanout device) - munmap once you don't need the buffer any more v2 (Tiago): Fix header file type names (u64 -> __u64) v3 (Tiago): Add documentation. Use enum dma_buf_sync_flags to the begin/end dma-buf functions. Check for overflows in start/length. v4 (Tiago): use 2d regions for sync. v5 (Tiago): forget about 2d regions (v4); use _IOW in DMA_BUF_IOCTL_SYNC and remove range information from struct dma_buf_sync. v6 (Tiago): use __u64 structured padded flags instead enum. Adjust documentation about the recommendation on using sync ioctls. v7 (Tiago): Alex' nit on flags definition and being even more wording in the doc about sync usage. v9 (Tiago): remove useless is_dma_buf_file check. Fix sync.flags conditionals and its mask order check. Add <linux/types.h> include in dma-buf.h. Cc: Ville Syrjälä <ville.syrjala@linux.intel.com> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Reviewed-by: Stéphane Marchesin <marcheu@chromium.org> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> Signed-off-by: Tiago Vignatti <tiago.vignatti@intel.com> Reviewed-by: David Herrmann <dh.herrmann@gmail.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1455228291-29640-1-git-send-email-tiago.vignatti@intel.com
2016-02-12 06:04:51 +08:00
#include <uapi/linux/dma-buf.h>
#include <uapi/linux/magic.h>
dma-buf: Add ioctls to allow userspace to flush The userspace might need some sort of cache coherency management e.g. when CPU and GPU domains are being accessed through dma-buf at the same time. To circumvent this problem there are begin/end coherency markers, that forward directly to existing dma-buf device drivers vfunc hooks. Userspace can make use of those markers through the DMA_BUF_IOCTL_SYNC ioctl. The sequence would be used like following: - mmap dma-buf fd - for each drawing/upload cycle in CPU 1. SYNC_START ioctl, 2. read/write to mmap area 3. SYNC_END ioctl. This can be repeated as often as you want (with the new data being consumed by the GPU or say scanout device) - munmap once you don't need the buffer any more v2 (Tiago): Fix header file type names (u64 -> __u64) v3 (Tiago): Add documentation. Use enum dma_buf_sync_flags to the begin/end dma-buf functions. Check for overflows in start/length. v4 (Tiago): use 2d regions for sync. v5 (Tiago): forget about 2d regions (v4); use _IOW in DMA_BUF_IOCTL_SYNC and remove range information from struct dma_buf_sync. v6 (Tiago): use __u64 structured padded flags instead enum. Adjust documentation about the recommendation on using sync ioctls. v7 (Tiago): Alex' nit on flags definition and being even more wording in the doc about sync usage. v9 (Tiago): remove useless is_dma_buf_file check. Fix sync.flags conditionals and its mask order check. Add <linux/types.h> include in dma-buf.h. Cc: Ville Syrjälä <ville.syrjala@linux.intel.com> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Reviewed-by: Stéphane Marchesin <marcheu@chromium.org> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> Signed-off-by: Tiago Vignatti <tiago.vignatti@intel.com> Reviewed-by: David Herrmann <dh.herrmann@gmail.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1455228291-29640-1-git-send-email-tiago.vignatti@intel.com
2016-02-12 06:04:51 +08:00
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
static inline int is_dma_buf_file(struct file *);
struct dma_buf_list {
struct list_head head;
struct mutex lock;
};
static struct dma_buf_list db_list;
static char *dmabuffs_dname(struct dentry *dentry, char *buffer, int buflen)
{
struct dma_buf *dmabuf;
char name[DMA_BUF_NAME_LEN];
size_t ret = 0;
dmabuf = dentry->d_fsdata;
spin_lock(&dmabuf->name_lock);
if (dmabuf->name)
ret = strlcpy(name, dmabuf->name, DMA_BUF_NAME_LEN);
spin_unlock(&dmabuf->name_lock);
return dynamic_dname(dentry, buffer, buflen, "/%s:%s",
dentry->d_name.name, ret > 0 ? name : "");
}
static void dma_buf_release(struct dentry *dentry)
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
{
struct dma_buf *dmabuf;
dmabuf = dentry->d_fsdata;
if (unlikely(!dmabuf))
return;
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
BUG_ON(dmabuf->vmapping_counter);
/*
* Any fences that a dma-buf poll can wait on should be signaled
* before releasing dma-buf. This is the responsibility of each
* driver that uses the reservation objects.
*
* If you hit this BUG() it means someone dropped their ref to the
* dma-buf while still having pending operation to the buffer.
*/
BUG_ON(dmabuf->cb_shared.active || dmabuf->cb_excl.active);
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
dmabuf->ops->release(dmabuf);
if (dmabuf->resv == (struct dma_resv *)&dmabuf[1])
dma_resv_fini(dmabuf->resv);
WARN_ON(!list_empty(&dmabuf->attachments));
module_put(dmabuf->owner);
kfree(dmabuf->name);
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
kfree(dmabuf);
}
static int dma_buf_file_release(struct inode *inode, struct file *file)
{
struct dma_buf *dmabuf;
if (!is_dma_buf_file(file))
return -EINVAL;
dmabuf = file->private_data;
mutex_lock(&db_list.lock);
list_del(&dmabuf->list_node);
mutex_unlock(&db_list.lock);
return 0;
}
static const struct dentry_operations dma_buf_dentry_ops = {
.d_dname = dmabuffs_dname,
.d_release = dma_buf_release,
};
static struct vfsmount *dma_buf_mnt;
static int dma_buf_fs_init_context(struct fs_context *fc)
{
struct pseudo_fs_context *ctx;
ctx = init_pseudo(fc, DMA_BUF_MAGIC);
if (!ctx)
return -ENOMEM;
ctx->dops = &dma_buf_dentry_ops;
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
return 0;
}
static struct file_system_type dma_buf_fs_type = {
.name = "dmabuf",
.init_fs_context = dma_buf_fs_init_context,
.kill_sb = kill_anon_super,
};
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
static int dma_buf_mmap_internal(struct file *file, struct vm_area_struct *vma)
{
struct dma_buf *dmabuf;
if (!is_dma_buf_file(file))
return -EINVAL;
dmabuf = file->private_data;
/* check if buffer supports mmap */
if (!dmabuf->ops->mmap)
return -EINVAL;
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
/* check for overflowing the buffer's size */
if (vma->vm_pgoff + vma_pages(vma) >
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
dmabuf->size >> PAGE_SHIFT)
return -EINVAL;
return dmabuf->ops->mmap(dmabuf, vma);
}
static loff_t dma_buf_llseek(struct file *file, loff_t offset, int whence)
{
struct dma_buf *dmabuf;
loff_t base;
if (!is_dma_buf_file(file))
return -EBADF;
dmabuf = file->private_data;
/* only support discovering the end of the buffer,
but also allow SEEK_SET to maintain the idiomatic
SEEK_END(0), SEEK_CUR(0) pattern */
if (whence == SEEK_END)
base = dmabuf->size;
else if (whence == SEEK_SET)
base = 0;
else
return -EINVAL;
if (offset != 0)
return -EINVAL;
return base + offset;
}
/**
* DOC: fence polling
*
* To support cross-device and cross-driver synchronization of buffer access
* implicit fences (represented internally in the kernel with &struct fence) can
* be attached to a &dma_buf. The glue for that and a few related things are
* provided in the &dma_resv structure.
*
* Userspace can query the state of these implicitly tracked fences using poll()
* and related system calls:
*
* - Checking for EPOLLIN, i.e. read access, can be use to query the state of the
* most recent write or exclusive fence.
*
* - Checking for EPOLLOUT, i.e. write access, can be used to query the state of
* all attached fences, shared and exclusive ones.
*
* Note that this only signals the completion of the respective fences, i.e. the
* DMA transfers are complete. Cache flushing and any other necessary
* preparations before CPU access can begin still need to happen.
*/
dma-buf: Rename struct fence to dma_fence I plan to usurp the short name of struct fence for a core kernel struct, and so I need to rename the specialised fence/timeline for DMA operations to make room. A consensus was reached in https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html that making clear this fence applies to DMA operations was a good thing. Since then the patch has grown a bit as usage increases, so hopefully it remains a good thing! (v2...: rebase, rerun spatch) v3: Compile on msm, spotted a manual fixup that I broke. v4: Try again for msm, sorry Daniel coccinelle script: @@ @@ - struct fence + struct dma_fence @@ @@ - struct fence_ops + struct dma_fence_ops @@ @@ - struct fence_cb + struct dma_fence_cb @@ @@ - struct fence_array + struct dma_fence_array @@ @@ - enum fence_flag_bits + enum dma_fence_flag_bits @@ @@ ( - fence_init + dma_fence_init | - fence_release + dma_fence_release | - fence_free + dma_fence_free | - fence_get + dma_fence_get | - fence_get_rcu + dma_fence_get_rcu | - fence_put + dma_fence_put | - fence_signal + dma_fence_signal | - fence_signal_locked + dma_fence_signal_locked | - fence_default_wait + dma_fence_default_wait | - fence_add_callback + dma_fence_add_callback | - fence_remove_callback + dma_fence_remove_callback | - fence_enable_sw_signaling + dma_fence_enable_sw_signaling | - fence_is_signaled_locked + dma_fence_is_signaled_locked | - fence_is_signaled + dma_fence_is_signaled | - fence_is_later + dma_fence_is_later | - fence_later + dma_fence_later | - fence_wait_timeout + dma_fence_wait_timeout | - fence_wait_any_timeout + dma_fence_wait_any_timeout | - fence_wait + dma_fence_wait | - fence_context_alloc + dma_fence_context_alloc | - fence_array_create + dma_fence_array_create | - to_fence_array + to_dma_fence_array | - fence_is_array + dma_fence_is_array | - trace_fence_emit + trace_dma_fence_emit | - FENCE_TRACE + DMA_FENCE_TRACE | - FENCE_WARN + DMA_FENCE_WARN | - FENCE_ERR + DMA_FENCE_ERR ) ( ... ) Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk> Acked-by: Sumit Semwal <sumit.semwal@linaro.org> Acked-by: Christian König <christian.koenig@amd.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk
2016-10-25 20:00:45 +08:00
static void dma_buf_poll_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
{
struct dma_buf_poll_cb_t *dcb = (struct dma_buf_poll_cb_t *)cb;
unsigned long flags;
spin_lock_irqsave(&dcb->poll->lock, flags);
wake_up_locked_poll(dcb->poll, dcb->active);
dcb->active = 0;
spin_unlock_irqrestore(&dcb->poll->lock, flags);
}
static __poll_t dma_buf_poll(struct file *file, poll_table *poll)
{
struct dma_buf *dmabuf;
struct dma_resv *resv;
struct dma_resv_list *fobj;
dma-buf: Rename struct fence to dma_fence I plan to usurp the short name of struct fence for a core kernel struct, and so I need to rename the specialised fence/timeline for DMA operations to make room. A consensus was reached in https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html that making clear this fence applies to DMA operations was a good thing. Since then the patch has grown a bit as usage increases, so hopefully it remains a good thing! (v2...: rebase, rerun spatch) v3: Compile on msm, spotted a manual fixup that I broke. v4: Try again for msm, sorry Daniel coccinelle script: @@ @@ - struct fence + struct dma_fence @@ @@ - struct fence_ops + struct dma_fence_ops @@ @@ - struct fence_cb + struct dma_fence_cb @@ @@ - struct fence_array + struct dma_fence_array @@ @@ - enum fence_flag_bits + enum dma_fence_flag_bits @@ @@ ( - fence_init + dma_fence_init | - fence_release + dma_fence_release | - fence_free + dma_fence_free | - fence_get + dma_fence_get | - fence_get_rcu + dma_fence_get_rcu | - fence_put + dma_fence_put | - fence_signal + dma_fence_signal | - fence_signal_locked + dma_fence_signal_locked | - fence_default_wait + dma_fence_default_wait | - fence_add_callback + dma_fence_add_callback | - fence_remove_callback + dma_fence_remove_callback | - fence_enable_sw_signaling + dma_fence_enable_sw_signaling | - fence_is_signaled_locked + dma_fence_is_signaled_locked | - fence_is_signaled + dma_fence_is_signaled | - fence_is_later + dma_fence_is_later | - fence_later + dma_fence_later | - fence_wait_timeout + dma_fence_wait_timeout | - fence_wait_any_timeout + dma_fence_wait_any_timeout | - fence_wait + dma_fence_wait | - fence_context_alloc + dma_fence_context_alloc | - fence_array_create + dma_fence_array_create | - to_fence_array + to_dma_fence_array | - fence_is_array + dma_fence_is_array | - trace_fence_emit + trace_dma_fence_emit | - FENCE_TRACE + DMA_FENCE_TRACE | - FENCE_WARN + DMA_FENCE_WARN | - FENCE_ERR + DMA_FENCE_ERR ) ( ... ) Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk> Acked-by: Sumit Semwal <sumit.semwal@linaro.org> Acked-by: Christian König <christian.koenig@amd.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk
2016-10-25 20:00:45 +08:00
struct dma_fence *fence_excl;
__poll_t events;
unsigned shared_count, seq;
dmabuf = file->private_data;
if (!dmabuf || !dmabuf->resv)
return EPOLLERR;
resv = dmabuf->resv;
poll_wait(file, &dmabuf->poll, poll);
events = poll_requested_events(poll) & (EPOLLIN | EPOLLOUT);
if (!events)
return 0;
retry:
seq = read_seqcount_begin(&resv->seq);
rcu_read_lock();
fobj = rcu_dereference(resv->fence);
if (fobj)
shared_count = fobj->shared_count;
else
shared_count = 0;
fence_excl = rcu_dereference(resv->fence_excl);
if (read_seqcount_retry(&resv->seq, seq)) {
rcu_read_unlock();
goto retry;
}
if (fence_excl && (!(events & EPOLLOUT) || shared_count == 0)) {
struct dma_buf_poll_cb_t *dcb = &dmabuf->cb_excl;
__poll_t pevents = EPOLLIN;
if (shared_count == 0)
pevents |= EPOLLOUT;
spin_lock_irq(&dmabuf->poll.lock);
if (dcb->active) {
dcb->active |= pevents;
events &= ~pevents;
} else
dcb->active = pevents;
spin_unlock_irq(&dmabuf->poll.lock);
if (events & pevents) {
dma-buf: Rename struct fence to dma_fence I plan to usurp the short name of struct fence for a core kernel struct, and so I need to rename the specialised fence/timeline for DMA operations to make room. A consensus was reached in https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html that making clear this fence applies to DMA operations was a good thing. Since then the patch has grown a bit as usage increases, so hopefully it remains a good thing! (v2...: rebase, rerun spatch) v3: Compile on msm, spotted a manual fixup that I broke. v4: Try again for msm, sorry Daniel coccinelle script: @@ @@ - struct fence + struct dma_fence @@ @@ - struct fence_ops + struct dma_fence_ops @@ @@ - struct fence_cb + struct dma_fence_cb @@ @@ - struct fence_array + struct dma_fence_array @@ @@ - enum fence_flag_bits + enum dma_fence_flag_bits @@ @@ ( - fence_init + dma_fence_init | - fence_release + dma_fence_release | - fence_free + dma_fence_free | - fence_get + dma_fence_get | - fence_get_rcu + dma_fence_get_rcu | - fence_put + dma_fence_put | - fence_signal + dma_fence_signal | - fence_signal_locked + dma_fence_signal_locked | - fence_default_wait + dma_fence_default_wait | - fence_add_callback + dma_fence_add_callback | - fence_remove_callback + dma_fence_remove_callback | - fence_enable_sw_signaling + dma_fence_enable_sw_signaling | - fence_is_signaled_locked + dma_fence_is_signaled_locked | - fence_is_signaled + dma_fence_is_signaled | - fence_is_later + dma_fence_is_later | - fence_later + dma_fence_later | - fence_wait_timeout + dma_fence_wait_timeout | - fence_wait_any_timeout + dma_fence_wait_any_timeout | - fence_wait + dma_fence_wait | - fence_context_alloc + dma_fence_context_alloc | - fence_array_create + dma_fence_array_create | - to_fence_array + to_dma_fence_array | - fence_is_array + dma_fence_is_array | - trace_fence_emit + trace_dma_fence_emit | - FENCE_TRACE + DMA_FENCE_TRACE | - FENCE_WARN + DMA_FENCE_WARN | - FENCE_ERR + DMA_FENCE_ERR ) ( ... ) Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk> Acked-by: Sumit Semwal <sumit.semwal@linaro.org> Acked-by: Christian König <christian.koenig@amd.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk
2016-10-25 20:00:45 +08:00
if (!dma_fence_get_rcu(fence_excl)) {
/* force a recheck */
events &= ~pevents;
dma_buf_poll_cb(NULL, &dcb->cb);
dma-buf: Rename struct fence to dma_fence I plan to usurp the short name of struct fence for a core kernel struct, and so I need to rename the specialised fence/timeline for DMA operations to make room. A consensus was reached in https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html that making clear this fence applies to DMA operations was a good thing. Since then the patch has grown a bit as usage increases, so hopefully it remains a good thing! (v2...: rebase, rerun spatch) v3: Compile on msm, spotted a manual fixup that I broke. v4: Try again for msm, sorry Daniel coccinelle script: @@ @@ - struct fence + struct dma_fence @@ @@ - struct fence_ops + struct dma_fence_ops @@ @@ - struct fence_cb + struct dma_fence_cb @@ @@ - struct fence_array + struct dma_fence_array @@ @@ - enum fence_flag_bits + enum dma_fence_flag_bits @@ @@ ( - fence_init + dma_fence_init | - fence_release + dma_fence_release | - fence_free + dma_fence_free | - fence_get + dma_fence_get | - fence_get_rcu + dma_fence_get_rcu | - fence_put + dma_fence_put | - fence_signal + dma_fence_signal | - fence_signal_locked + dma_fence_signal_locked | - fence_default_wait + dma_fence_default_wait | - fence_add_callback + dma_fence_add_callback | - fence_remove_callback + dma_fence_remove_callback | - fence_enable_sw_signaling + dma_fence_enable_sw_signaling | - fence_is_signaled_locked + dma_fence_is_signaled_locked | - fence_is_signaled + dma_fence_is_signaled | - fence_is_later + dma_fence_is_later | - fence_later + dma_fence_later | - fence_wait_timeout + dma_fence_wait_timeout | - fence_wait_any_timeout + dma_fence_wait_any_timeout | - fence_wait + dma_fence_wait | - fence_context_alloc + dma_fence_context_alloc | - fence_array_create + dma_fence_array_create | - to_fence_array + to_dma_fence_array | - fence_is_array + dma_fence_is_array | - trace_fence_emit + trace_dma_fence_emit | - FENCE_TRACE + DMA_FENCE_TRACE | - FENCE_WARN + DMA_FENCE_WARN | - FENCE_ERR + DMA_FENCE_ERR ) ( ... ) Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk> Acked-by: Sumit Semwal <sumit.semwal@linaro.org> Acked-by: Christian König <christian.koenig@amd.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk
2016-10-25 20:00:45 +08:00
} else if (!dma_fence_add_callback(fence_excl, &dcb->cb,
dma_buf_poll_cb)) {
events &= ~pevents;
dma-buf: Rename struct fence to dma_fence I plan to usurp the short name of struct fence for a core kernel struct, and so I need to rename the specialised fence/timeline for DMA operations to make room. A consensus was reached in https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html that making clear this fence applies to DMA operations was a good thing. Since then the patch has grown a bit as usage increases, so hopefully it remains a good thing! (v2...: rebase, rerun spatch) v3: Compile on msm, spotted a manual fixup that I broke. v4: Try again for msm, sorry Daniel coccinelle script: @@ @@ - struct fence + struct dma_fence @@ @@ - struct fence_ops + struct dma_fence_ops @@ @@ - struct fence_cb + struct dma_fence_cb @@ @@ - struct fence_array + struct dma_fence_array @@ @@ - enum fence_flag_bits + enum dma_fence_flag_bits @@ @@ ( - fence_init + dma_fence_init | - fence_release + dma_fence_release | - fence_free + dma_fence_free | - fence_get + dma_fence_get | - fence_get_rcu + dma_fence_get_rcu | - fence_put + dma_fence_put | - fence_signal + dma_fence_signal | - fence_signal_locked + dma_fence_signal_locked | - fence_default_wait + dma_fence_default_wait | - fence_add_callback + dma_fence_add_callback | - fence_remove_callback + dma_fence_remove_callback | - fence_enable_sw_signaling + dma_fence_enable_sw_signaling | - fence_is_signaled_locked + dma_fence_is_signaled_locked | - fence_is_signaled + dma_fence_is_signaled | - fence_is_later + dma_fence_is_later | - fence_later + dma_fence_later | - fence_wait_timeout + dma_fence_wait_timeout | - fence_wait_any_timeout + dma_fence_wait_any_timeout | - fence_wait + dma_fence_wait | - fence_context_alloc + dma_fence_context_alloc | - fence_array_create + dma_fence_array_create | - to_fence_array + to_dma_fence_array | - fence_is_array + dma_fence_is_array | - trace_fence_emit + trace_dma_fence_emit | - FENCE_TRACE + DMA_FENCE_TRACE | - FENCE_WARN + DMA_FENCE_WARN | - FENCE_ERR + DMA_FENCE_ERR ) ( ... ) Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk> Acked-by: Sumit Semwal <sumit.semwal@linaro.org> Acked-by: Christian König <christian.koenig@amd.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk
2016-10-25 20:00:45 +08:00
dma_fence_put(fence_excl);
} else {
/*
* No callback queued, wake up any additional
* waiters.
*/
dma-buf: Rename struct fence to dma_fence I plan to usurp the short name of struct fence for a core kernel struct, and so I need to rename the specialised fence/timeline for DMA operations to make room. A consensus was reached in https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html that making clear this fence applies to DMA operations was a good thing. Since then the patch has grown a bit as usage increases, so hopefully it remains a good thing! (v2...: rebase, rerun spatch) v3: Compile on msm, spotted a manual fixup that I broke. v4: Try again for msm, sorry Daniel coccinelle script: @@ @@ - struct fence + struct dma_fence @@ @@ - struct fence_ops + struct dma_fence_ops @@ @@ - struct fence_cb + struct dma_fence_cb @@ @@ - struct fence_array + struct dma_fence_array @@ @@ - enum fence_flag_bits + enum dma_fence_flag_bits @@ @@ ( - fence_init + dma_fence_init | - fence_release + dma_fence_release | - fence_free + dma_fence_free | - fence_get + dma_fence_get | - fence_get_rcu + dma_fence_get_rcu | - fence_put + dma_fence_put | - fence_signal + dma_fence_signal | - fence_signal_locked + dma_fence_signal_locked | - fence_default_wait + dma_fence_default_wait | - fence_add_callback + dma_fence_add_callback | - fence_remove_callback + dma_fence_remove_callback | - fence_enable_sw_signaling + dma_fence_enable_sw_signaling | - fence_is_signaled_locked + dma_fence_is_signaled_locked | - fence_is_signaled + dma_fence_is_signaled | - fence_is_later + dma_fence_is_later | - fence_later + dma_fence_later | - fence_wait_timeout + dma_fence_wait_timeout | - fence_wait_any_timeout + dma_fence_wait_any_timeout | - fence_wait + dma_fence_wait | - fence_context_alloc + dma_fence_context_alloc | - fence_array_create + dma_fence_array_create | - to_fence_array + to_dma_fence_array | - fence_is_array + dma_fence_is_array | - trace_fence_emit + trace_dma_fence_emit | - FENCE_TRACE + DMA_FENCE_TRACE | - FENCE_WARN + DMA_FENCE_WARN | - FENCE_ERR + DMA_FENCE_ERR ) ( ... ) Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk> Acked-by: Sumit Semwal <sumit.semwal@linaro.org> Acked-by: Christian König <christian.koenig@amd.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk
2016-10-25 20:00:45 +08:00
dma_fence_put(fence_excl);
dma_buf_poll_cb(NULL, &dcb->cb);
}
}
}
if ((events & EPOLLOUT) && shared_count > 0) {
struct dma_buf_poll_cb_t *dcb = &dmabuf->cb_shared;
int i;
/* Only queue a new callback if no event has fired yet */
spin_lock_irq(&dmabuf->poll.lock);
if (dcb->active)
events &= ~EPOLLOUT;
else
dcb->active = EPOLLOUT;
spin_unlock_irq(&dmabuf->poll.lock);
if (!(events & EPOLLOUT))
goto out;
for (i = 0; i < shared_count; ++i) {
dma-buf: Rename struct fence to dma_fence I plan to usurp the short name of struct fence for a core kernel struct, and so I need to rename the specialised fence/timeline for DMA operations to make room. A consensus was reached in https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html that making clear this fence applies to DMA operations was a good thing. Since then the patch has grown a bit as usage increases, so hopefully it remains a good thing! (v2...: rebase, rerun spatch) v3: Compile on msm, spotted a manual fixup that I broke. v4: Try again for msm, sorry Daniel coccinelle script: @@ @@ - struct fence + struct dma_fence @@ @@ - struct fence_ops + struct dma_fence_ops @@ @@ - struct fence_cb + struct dma_fence_cb @@ @@ - struct fence_array + struct dma_fence_array @@ @@ - enum fence_flag_bits + enum dma_fence_flag_bits @@ @@ ( - fence_init + dma_fence_init | - fence_release + dma_fence_release | - fence_free + dma_fence_free | - fence_get + dma_fence_get | - fence_get_rcu + dma_fence_get_rcu | - fence_put + dma_fence_put | - fence_signal + dma_fence_signal | - fence_signal_locked + dma_fence_signal_locked | - fence_default_wait + dma_fence_default_wait | - fence_add_callback + dma_fence_add_callback | - fence_remove_callback + dma_fence_remove_callback | - fence_enable_sw_signaling + dma_fence_enable_sw_signaling | - fence_is_signaled_locked + dma_fence_is_signaled_locked | - fence_is_signaled + dma_fence_is_signaled | - fence_is_later + dma_fence_is_later | - fence_later + dma_fence_later | - fence_wait_timeout + dma_fence_wait_timeout | - fence_wait_any_timeout + dma_fence_wait_any_timeout | - fence_wait + dma_fence_wait | - fence_context_alloc + dma_fence_context_alloc | - fence_array_create + dma_fence_array_create | - to_fence_array + to_dma_fence_array | - fence_is_array + dma_fence_is_array | - trace_fence_emit + trace_dma_fence_emit | - FENCE_TRACE + DMA_FENCE_TRACE | - FENCE_WARN + DMA_FENCE_WARN | - FENCE_ERR + DMA_FENCE_ERR ) ( ... ) Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk> Acked-by: Sumit Semwal <sumit.semwal@linaro.org> Acked-by: Christian König <christian.koenig@amd.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk
2016-10-25 20:00:45 +08:00
struct dma_fence *fence = rcu_dereference(fobj->shared[i]);
dma-buf: Rename struct fence to dma_fence I plan to usurp the short name of struct fence for a core kernel struct, and so I need to rename the specialised fence/timeline for DMA operations to make room. A consensus was reached in https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html that making clear this fence applies to DMA operations was a good thing. Since then the patch has grown a bit as usage increases, so hopefully it remains a good thing! (v2...: rebase, rerun spatch) v3: Compile on msm, spotted a manual fixup that I broke. v4: Try again for msm, sorry Daniel coccinelle script: @@ @@ - struct fence + struct dma_fence @@ @@ - struct fence_ops + struct dma_fence_ops @@ @@ - struct fence_cb + struct dma_fence_cb @@ @@ - struct fence_array + struct dma_fence_array @@ @@ - enum fence_flag_bits + enum dma_fence_flag_bits @@ @@ ( - fence_init + dma_fence_init | - fence_release + dma_fence_release | - fence_free + dma_fence_free | - fence_get + dma_fence_get | - fence_get_rcu + dma_fence_get_rcu | - fence_put + dma_fence_put | - fence_signal + dma_fence_signal | - fence_signal_locked + dma_fence_signal_locked | - fence_default_wait + dma_fence_default_wait | - fence_add_callback + dma_fence_add_callback | - fence_remove_callback + dma_fence_remove_callback | - fence_enable_sw_signaling + dma_fence_enable_sw_signaling | - fence_is_signaled_locked + dma_fence_is_signaled_locked | - fence_is_signaled + dma_fence_is_signaled | - fence_is_later + dma_fence_is_later | - fence_later + dma_fence_later | - fence_wait_timeout + dma_fence_wait_timeout | - fence_wait_any_timeout + dma_fence_wait_any_timeout | - fence_wait + dma_fence_wait | - fence_context_alloc + dma_fence_context_alloc | - fence_array_create + dma_fence_array_create | - to_fence_array + to_dma_fence_array | - fence_is_array + dma_fence_is_array | - trace_fence_emit + trace_dma_fence_emit | - FENCE_TRACE + DMA_FENCE_TRACE | - FENCE_WARN + DMA_FENCE_WARN | - FENCE_ERR + DMA_FENCE_ERR ) ( ... ) Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk> Acked-by: Sumit Semwal <sumit.semwal@linaro.org> Acked-by: Christian König <christian.koenig@amd.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk
2016-10-25 20:00:45 +08:00
if (!dma_fence_get_rcu(fence)) {
/*
* fence refcount dropped to zero, this means
* that fobj has been freed
*
* call dma_buf_poll_cb and force a recheck!
*/
events &= ~EPOLLOUT;
dma_buf_poll_cb(NULL, &dcb->cb);
break;
}
dma-buf: Rename struct fence to dma_fence I plan to usurp the short name of struct fence for a core kernel struct, and so I need to rename the specialised fence/timeline for DMA operations to make room. A consensus was reached in https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html that making clear this fence applies to DMA operations was a good thing. Since then the patch has grown a bit as usage increases, so hopefully it remains a good thing! (v2...: rebase, rerun spatch) v3: Compile on msm, spotted a manual fixup that I broke. v4: Try again for msm, sorry Daniel coccinelle script: @@ @@ - struct fence + struct dma_fence @@ @@ - struct fence_ops + struct dma_fence_ops @@ @@ - struct fence_cb + struct dma_fence_cb @@ @@ - struct fence_array + struct dma_fence_array @@ @@ - enum fence_flag_bits + enum dma_fence_flag_bits @@ @@ ( - fence_init + dma_fence_init | - fence_release + dma_fence_release | - fence_free + dma_fence_free | - fence_get + dma_fence_get | - fence_get_rcu + dma_fence_get_rcu | - fence_put + dma_fence_put | - fence_signal + dma_fence_signal | - fence_signal_locked + dma_fence_signal_locked | - fence_default_wait + dma_fence_default_wait | - fence_add_callback + dma_fence_add_callback | - fence_remove_callback + dma_fence_remove_callback | - fence_enable_sw_signaling + dma_fence_enable_sw_signaling | - fence_is_signaled_locked + dma_fence_is_signaled_locked | - fence_is_signaled + dma_fence_is_signaled | - fence_is_later + dma_fence_is_later | - fence_later + dma_fence_later | - fence_wait_timeout + dma_fence_wait_timeout | - fence_wait_any_timeout + dma_fence_wait_any_timeout | - fence_wait + dma_fence_wait | - fence_context_alloc + dma_fence_context_alloc | - fence_array_create + dma_fence_array_create | - to_fence_array + to_dma_fence_array | - fence_is_array + dma_fence_is_array | - trace_fence_emit + trace_dma_fence_emit | - FENCE_TRACE + DMA_FENCE_TRACE | - FENCE_WARN + DMA_FENCE_WARN | - FENCE_ERR + DMA_FENCE_ERR ) ( ... ) Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk> Acked-by: Sumit Semwal <sumit.semwal@linaro.org> Acked-by: Christian König <christian.koenig@amd.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk
2016-10-25 20:00:45 +08:00
if (!dma_fence_add_callback(fence, &dcb->cb,
dma_buf_poll_cb)) {
dma_fence_put(fence);
events &= ~EPOLLOUT;
break;
}
dma-buf: Rename struct fence to dma_fence I plan to usurp the short name of struct fence for a core kernel struct, and so I need to rename the specialised fence/timeline for DMA operations to make room. A consensus was reached in https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html that making clear this fence applies to DMA operations was a good thing. Since then the patch has grown a bit as usage increases, so hopefully it remains a good thing! (v2...: rebase, rerun spatch) v3: Compile on msm, spotted a manual fixup that I broke. v4: Try again for msm, sorry Daniel coccinelle script: @@ @@ - struct fence + struct dma_fence @@ @@ - struct fence_ops + struct dma_fence_ops @@ @@ - struct fence_cb + struct dma_fence_cb @@ @@ - struct fence_array + struct dma_fence_array @@ @@ - enum fence_flag_bits + enum dma_fence_flag_bits @@ @@ ( - fence_init + dma_fence_init | - fence_release + dma_fence_release | - fence_free + dma_fence_free | - fence_get + dma_fence_get | - fence_get_rcu + dma_fence_get_rcu | - fence_put + dma_fence_put | - fence_signal + dma_fence_signal | - fence_signal_locked + dma_fence_signal_locked | - fence_default_wait + dma_fence_default_wait | - fence_add_callback + dma_fence_add_callback | - fence_remove_callback + dma_fence_remove_callback | - fence_enable_sw_signaling + dma_fence_enable_sw_signaling | - fence_is_signaled_locked + dma_fence_is_signaled_locked | - fence_is_signaled + dma_fence_is_signaled | - fence_is_later + dma_fence_is_later | - fence_later + dma_fence_later | - fence_wait_timeout + dma_fence_wait_timeout | - fence_wait_any_timeout + dma_fence_wait_any_timeout | - fence_wait + dma_fence_wait | - fence_context_alloc + dma_fence_context_alloc | - fence_array_create + dma_fence_array_create | - to_fence_array + to_dma_fence_array | - fence_is_array + dma_fence_is_array | - trace_fence_emit + trace_dma_fence_emit | - FENCE_TRACE + DMA_FENCE_TRACE | - FENCE_WARN + DMA_FENCE_WARN | - FENCE_ERR + DMA_FENCE_ERR ) ( ... ) Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk> Acked-by: Sumit Semwal <sumit.semwal@linaro.org> Acked-by: Christian König <christian.koenig@amd.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk
2016-10-25 20:00:45 +08:00
dma_fence_put(fence);
}
/* No callback queued, wake up any additional waiters. */
if (i == shared_count)
dma_buf_poll_cb(NULL, &dcb->cb);
}
out:
rcu_read_unlock();
return events;
}
/**
* dma_buf_set_name - Set a name to a specific dma_buf to track the usage.
* The name of the dma-buf buffer can only be set when the dma-buf is not
* attached to any devices. It could theoritically support changing the
* name of the dma-buf if the same piece of memory is used for multiple
* purpose between different devices.
*
* @dmabuf [in] dmabuf buffer that will be renamed.
* @buf: [in] A piece of userspace memory that contains the name of
* the dma-buf.
*
* Returns 0 on success. If the dma-buf buffer is already attached to
* devices, return -EBUSY.
*
*/
static long dma_buf_set_name(struct dma_buf *dmabuf, const char __user *buf)
{
char *name = strndup_user(buf, DMA_BUF_NAME_LEN);
long ret = 0;
if (IS_ERR(name))
return PTR_ERR(name);
mutex_lock(&dmabuf->lock);
if (!list_empty(&dmabuf->attachments)) {
ret = -EBUSY;
kfree(name);
goto out_unlock;
}
spin_lock(&dmabuf->name_lock);
kfree(dmabuf->name);
dmabuf->name = name;
spin_unlock(&dmabuf->name_lock);
out_unlock:
mutex_unlock(&dmabuf->lock);
return ret;
}
dma-buf: Add ioctls to allow userspace to flush The userspace might need some sort of cache coherency management e.g. when CPU and GPU domains are being accessed through dma-buf at the same time. To circumvent this problem there are begin/end coherency markers, that forward directly to existing dma-buf device drivers vfunc hooks. Userspace can make use of those markers through the DMA_BUF_IOCTL_SYNC ioctl. The sequence would be used like following: - mmap dma-buf fd - for each drawing/upload cycle in CPU 1. SYNC_START ioctl, 2. read/write to mmap area 3. SYNC_END ioctl. This can be repeated as often as you want (with the new data being consumed by the GPU or say scanout device) - munmap once you don't need the buffer any more v2 (Tiago): Fix header file type names (u64 -> __u64) v3 (Tiago): Add documentation. Use enum dma_buf_sync_flags to the begin/end dma-buf functions. Check for overflows in start/length. v4 (Tiago): use 2d regions for sync. v5 (Tiago): forget about 2d regions (v4); use _IOW in DMA_BUF_IOCTL_SYNC and remove range information from struct dma_buf_sync. v6 (Tiago): use __u64 structured padded flags instead enum. Adjust documentation about the recommendation on using sync ioctls. v7 (Tiago): Alex' nit on flags definition and being even more wording in the doc about sync usage. v9 (Tiago): remove useless is_dma_buf_file check. Fix sync.flags conditionals and its mask order check. Add <linux/types.h> include in dma-buf.h. Cc: Ville Syrjälä <ville.syrjala@linux.intel.com> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Reviewed-by: Stéphane Marchesin <marcheu@chromium.org> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> Signed-off-by: Tiago Vignatti <tiago.vignatti@intel.com> Reviewed-by: David Herrmann <dh.herrmann@gmail.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1455228291-29640-1-git-send-email-tiago.vignatti@intel.com
2016-02-12 06:04:51 +08:00
static long dma_buf_ioctl(struct file *file,
unsigned int cmd, unsigned long arg)
{
struct dma_buf *dmabuf;
struct dma_buf_sync sync;
enum dma_data_direction direction;
dma-buf, drm, ion: Propagate error code from dma_buf_start_cpu_access() Drivers, especially i915.ko, can fail during the initial migration of a dma-buf for CPU access. However, the error code from the driver was not being propagated back to ioctl and so userspace was blissfully ignorant of the failure. Rendering corruption ensues. Whilst fixing the ioctl to return the error code from dma_buf_start_cpu_access(), also do the same for dma_buf_end_cpu_access(). For most drivers, dma_buf_end_cpu_access() cannot fail. i915.ko however, as most drivers would, wants to avoid being uninterruptible (as would be required to guarrantee no failure when flushing the buffer to the device). As userspace already has to handle errors from the SYNC_IOCTL, take advantage of this to be able to restart the syscall across signals. This fixes a coherency issue for i915.ko as well as reducing the uninterruptible hold upon its BKL, the struct_mutex. Fixes commit c11e391da2a8fe973c3c2398452000bed505851e Author: Daniel Vetter <daniel.vetter@ffwll.ch> Date: Thu Feb 11 20:04:51 2016 -0200 dma-buf: Add ioctls to allow userspace to flush Testcase: igt/gem_concurrent_blit/*dmabuf*interruptible Testcase: igt/prime_mmap_coherency/ioctl-errors Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tiago Vignatti <tiago.vignatti@intel.com> Cc: Stéphane Marchesin <marcheu@chromium.org> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Daniel Vetter <daniel.vetter@intel.com> CC: linux-media@vger.kernel.org Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: intel-gfx@lists.freedesktop.org Cc: devel@driverdev.osuosl.org Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1458331359-2634-1-git-send-email-chris@chris-wilson.co.uk
2016-03-19 04:02:39 +08:00
int ret;
dma-buf: Add ioctls to allow userspace to flush The userspace might need some sort of cache coherency management e.g. when CPU and GPU domains are being accessed through dma-buf at the same time. To circumvent this problem there are begin/end coherency markers, that forward directly to existing dma-buf device drivers vfunc hooks. Userspace can make use of those markers through the DMA_BUF_IOCTL_SYNC ioctl. The sequence would be used like following: - mmap dma-buf fd - for each drawing/upload cycle in CPU 1. SYNC_START ioctl, 2. read/write to mmap area 3. SYNC_END ioctl. This can be repeated as often as you want (with the new data being consumed by the GPU or say scanout device) - munmap once you don't need the buffer any more v2 (Tiago): Fix header file type names (u64 -> __u64) v3 (Tiago): Add documentation. Use enum dma_buf_sync_flags to the begin/end dma-buf functions. Check for overflows in start/length. v4 (Tiago): use 2d regions for sync. v5 (Tiago): forget about 2d regions (v4); use _IOW in DMA_BUF_IOCTL_SYNC and remove range information from struct dma_buf_sync. v6 (Tiago): use __u64 structured padded flags instead enum. Adjust documentation about the recommendation on using sync ioctls. v7 (Tiago): Alex' nit on flags definition and being even more wording in the doc about sync usage. v9 (Tiago): remove useless is_dma_buf_file check. Fix sync.flags conditionals and its mask order check. Add <linux/types.h> include in dma-buf.h. Cc: Ville Syrjälä <ville.syrjala@linux.intel.com> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Reviewed-by: Stéphane Marchesin <marcheu@chromium.org> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> Signed-off-by: Tiago Vignatti <tiago.vignatti@intel.com> Reviewed-by: David Herrmann <dh.herrmann@gmail.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1455228291-29640-1-git-send-email-tiago.vignatti@intel.com
2016-02-12 06:04:51 +08:00
dmabuf = file->private_data;
switch (cmd) {
case DMA_BUF_IOCTL_SYNC:
if (copy_from_user(&sync, (void __user *) arg, sizeof(sync)))
return -EFAULT;
if (sync.flags & ~DMA_BUF_SYNC_VALID_FLAGS_MASK)
return -EINVAL;
switch (sync.flags & DMA_BUF_SYNC_RW) {
case DMA_BUF_SYNC_READ:
direction = DMA_FROM_DEVICE;
break;
case DMA_BUF_SYNC_WRITE:
direction = DMA_TO_DEVICE;
break;
case DMA_BUF_SYNC_RW:
direction = DMA_BIDIRECTIONAL;
break;
default:
return -EINVAL;
}
if (sync.flags & DMA_BUF_SYNC_END)
dma-buf, drm, ion: Propagate error code from dma_buf_start_cpu_access() Drivers, especially i915.ko, can fail during the initial migration of a dma-buf for CPU access. However, the error code from the driver was not being propagated back to ioctl and so userspace was blissfully ignorant of the failure. Rendering corruption ensues. Whilst fixing the ioctl to return the error code from dma_buf_start_cpu_access(), also do the same for dma_buf_end_cpu_access(). For most drivers, dma_buf_end_cpu_access() cannot fail. i915.ko however, as most drivers would, wants to avoid being uninterruptible (as would be required to guarrantee no failure when flushing the buffer to the device). As userspace already has to handle errors from the SYNC_IOCTL, take advantage of this to be able to restart the syscall across signals. This fixes a coherency issue for i915.ko as well as reducing the uninterruptible hold upon its BKL, the struct_mutex. Fixes commit c11e391da2a8fe973c3c2398452000bed505851e Author: Daniel Vetter <daniel.vetter@ffwll.ch> Date: Thu Feb 11 20:04:51 2016 -0200 dma-buf: Add ioctls to allow userspace to flush Testcase: igt/gem_concurrent_blit/*dmabuf*interruptible Testcase: igt/prime_mmap_coherency/ioctl-errors Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tiago Vignatti <tiago.vignatti@intel.com> Cc: Stéphane Marchesin <marcheu@chromium.org> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Daniel Vetter <daniel.vetter@intel.com> CC: linux-media@vger.kernel.org Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: intel-gfx@lists.freedesktop.org Cc: devel@driverdev.osuosl.org Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1458331359-2634-1-git-send-email-chris@chris-wilson.co.uk
2016-03-19 04:02:39 +08:00
ret = dma_buf_end_cpu_access(dmabuf, direction);
dma-buf: Add ioctls to allow userspace to flush The userspace might need some sort of cache coherency management e.g. when CPU and GPU domains are being accessed through dma-buf at the same time. To circumvent this problem there are begin/end coherency markers, that forward directly to existing dma-buf device drivers vfunc hooks. Userspace can make use of those markers through the DMA_BUF_IOCTL_SYNC ioctl. The sequence would be used like following: - mmap dma-buf fd - for each drawing/upload cycle in CPU 1. SYNC_START ioctl, 2. read/write to mmap area 3. SYNC_END ioctl. This can be repeated as often as you want (with the new data being consumed by the GPU or say scanout device) - munmap once you don't need the buffer any more v2 (Tiago): Fix header file type names (u64 -> __u64) v3 (Tiago): Add documentation. Use enum dma_buf_sync_flags to the begin/end dma-buf functions. Check for overflows in start/length. v4 (Tiago): use 2d regions for sync. v5 (Tiago): forget about 2d regions (v4); use _IOW in DMA_BUF_IOCTL_SYNC and remove range information from struct dma_buf_sync. v6 (Tiago): use __u64 structured padded flags instead enum. Adjust documentation about the recommendation on using sync ioctls. v7 (Tiago): Alex' nit on flags definition and being even more wording in the doc about sync usage. v9 (Tiago): remove useless is_dma_buf_file check. Fix sync.flags conditionals and its mask order check. Add <linux/types.h> include in dma-buf.h. Cc: Ville Syrjälä <ville.syrjala@linux.intel.com> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Reviewed-by: Stéphane Marchesin <marcheu@chromium.org> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> Signed-off-by: Tiago Vignatti <tiago.vignatti@intel.com> Reviewed-by: David Herrmann <dh.herrmann@gmail.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1455228291-29640-1-git-send-email-tiago.vignatti@intel.com
2016-02-12 06:04:51 +08:00
else
dma-buf, drm, ion: Propagate error code from dma_buf_start_cpu_access() Drivers, especially i915.ko, can fail during the initial migration of a dma-buf for CPU access. However, the error code from the driver was not being propagated back to ioctl and so userspace was blissfully ignorant of the failure. Rendering corruption ensues. Whilst fixing the ioctl to return the error code from dma_buf_start_cpu_access(), also do the same for dma_buf_end_cpu_access(). For most drivers, dma_buf_end_cpu_access() cannot fail. i915.ko however, as most drivers would, wants to avoid being uninterruptible (as would be required to guarrantee no failure when flushing the buffer to the device). As userspace already has to handle errors from the SYNC_IOCTL, take advantage of this to be able to restart the syscall across signals. This fixes a coherency issue for i915.ko as well as reducing the uninterruptible hold upon its BKL, the struct_mutex. Fixes commit c11e391da2a8fe973c3c2398452000bed505851e Author: Daniel Vetter <daniel.vetter@ffwll.ch> Date: Thu Feb 11 20:04:51 2016 -0200 dma-buf: Add ioctls to allow userspace to flush Testcase: igt/gem_concurrent_blit/*dmabuf*interruptible Testcase: igt/prime_mmap_coherency/ioctl-errors Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tiago Vignatti <tiago.vignatti@intel.com> Cc: Stéphane Marchesin <marcheu@chromium.org> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Daniel Vetter <daniel.vetter@intel.com> CC: linux-media@vger.kernel.org Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: intel-gfx@lists.freedesktop.org Cc: devel@driverdev.osuosl.org Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1458331359-2634-1-git-send-email-chris@chris-wilson.co.uk
2016-03-19 04:02:39 +08:00
ret = dma_buf_begin_cpu_access(dmabuf, direction);
dma-buf: Add ioctls to allow userspace to flush The userspace might need some sort of cache coherency management e.g. when CPU and GPU domains are being accessed through dma-buf at the same time. To circumvent this problem there are begin/end coherency markers, that forward directly to existing dma-buf device drivers vfunc hooks. Userspace can make use of those markers through the DMA_BUF_IOCTL_SYNC ioctl. The sequence would be used like following: - mmap dma-buf fd - for each drawing/upload cycle in CPU 1. SYNC_START ioctl, 2. read/write to mmap area 3. SYNC_END ioctl. This can be repeated as often as you want (with the new data being consumed by the GPU or say scanout device) - munmap once you don't need the buffer any more v2 (Tiago): Fix header file type names (u64 -> __u64) v3 (Tiago): Add documentation. Use enum dma_buf_sync_flags to the begin/end dma-buf functions. Check for overflows in start/length. v4 (Tiago): use 2d regions for sync. v5 (Tiago): forget about 2d regions (v4); use _IOW in DMA_BUF_IOCTL_SYNC and remove range information from struct dma_buf_sync. v6 (Tiago): use __u64 structured padded flags instead enum. Adjust documentation about the recommendation on using sync ioctls. v7 (Tiago): Alex' nit on flags definition and being even more wording in the doc about sync usage. v9 (Tiago): remove useless is_dma_buf_file check. Fix sync.flags conditionals and its mask order check. Add <linux/types.h> include in dma-buf.h. Cc: Ville Syrjälä <ville.syrjala@linux.intel.com> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Reviewed-by: Stéphane Marchesin <marcheu@chromium.org> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> Signed-off-by: Tiago Vignatti <tiago.vignatti@intel.com> Reviewed-by: David Herrmann <dh.herrmann@gmail.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1455228291-29640-1-git-send-email-tiago.vignatti@intel.com
2016-02-12 06:04:51 +08:00
dma-buf, drm, ion: Propagate error code from dma_buf_start_cpu_access() Drivers, especially i915.ko, can fail during the initial migration of a dma-buf for CPU access. However, the error code from the driver was not being propagated back to ioctl and so userspace was blissfully ignorant of the failure. Rendering corruption ensues. Whilst fixing the ioctl to return the error code from dma_buf_start_cpu_access(), also do the same for dma_buf_end_cpu_access(). For most drivers, dma_buf_end_cpu_access() cannot fail. i915.ko however, as most drivers would, wants to avoid being uninterruptible (as would be required to guarrantee no failure when flushing the buffer to the device). As userspace already has to handle errors from the SYNC_IOCTL, take advantage of this to be able to restart the syscall across signals. This fixes a coherency issue for i915.ko as well as reducing the uninterruptible hold upon its BKL, the struct_mutex. Fixes commit c11e391da2a8fe973c3c2398452000bed505851e Author: Daniel Vetter <daniel.vetter@ffwll.ch> Date: Thu Feb 11 20:04:51 2016 -0200 dma-buf: Add ioctls to allow userspace to flush Testcase: igt/gem_concurrent_blit/*dmabuf*interruptible Testcase: igt/prime_mmap_coherency/ioctl-errors Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tiago Vignatti <tiago.vignatti@intel.com> Cc: Stéphane Marchesin <marcheu@chromium.org> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Daniel Vetter <daniel.vetter@intel.com> CC: linux-media@vger.kernel.org Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: intel-gfx@lists.freedesktop.org Cc: devel@driverdev.osuosl.org Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1458331359-2634-1-git-send-email-chris@chris-wilson.co.uk
2016-03-19 04:02:39 +08:00
return ret;
case DMA_BUF_SET_NAME_A:
case DMA_BUF_SET_NAME_B:
return dma_buf_set_name(dmabuf, (const char __user *)arg);
dma-buf: Add ioctls to allow userspace to flush The userspace might need some sort of cache coherency management e.g. when CPU and GPU domains are being accessed through dma-buf at the same time. To circumvent this problem there are begin/end coherency markers, that forward directly to existing dma-buf device drivers vfunc hooks. Userspace can make use of those markers through the DMA_BUF_IOCTL_SYNC ioctl. The sequence would be used like following: - mmap dma-buf fd - for each drawing/upload cycle in CPU 1. SYNC_START ioctl, 2. read/write to mmap area 3. SYNC_END ioctl. This can be repeated as often as you want (with the new data being consumed by the GPU or say scanout device) - munmap once you don't need the buffer any more v2 (Tiago): Fix header file type names (u64 -> __u64) v3 (Tiago): Add documentation. Use enum dma_buf_sync_flags to the begin/end dma-buf functions. Check for overflows in start/length. v4 (Tiago): use 2d regions for sync. v5 (Tiago): forget about 2d regions (v4); use _IOW in DMA_BUF_IOCTL_SYNC and remove range information from struct dma_buf_sync. v6 (Tiago): use __u64 structured padded flags instead enum. Adjust documentation about the recommendation on using sync ioctls. v7 (Tiago): Alex' nit on flags definition and being even more wording in the doc about sync usage. v9 (Tiago): remove useless is_dma_buf_file check. Fix sync.flags conditionals and its mask order check. Add <linux/types.h> include in dma-buf.h. Cc: Ville Syrjälä <ville.syrjala@linux.intel.com> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Reviewed-by: Stéphane Marchesin <marcheu@chromium.org> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> Signed-off-by: Tiago Vignatti <tiago.vignatti@intel.com> Reviewed-by: David Herrmann <dh.herrmann@gmail.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1455228291-29640-1-git-send-email-tiago.vignatti@intel.com
2016-02-12 06:04:51 +08:00
default:
return -ENOTTY;
}
}
static void dma_buf_show_fdinfo(struct seq_file *m, struct file *file)
{
struct dma_buf *dmabuf = file->private_data;
seq_printf(m, "size:\t%zu\n", dmabuf->size);
/* Don't count the temporary reference taken inside procfs seq_show */
seq_printf(m, "count:\t%ld\n", file_count(dmabuf->file) - 1);
seq_printf(m, "exp_name:\t%s\n", dmabuf->exp_name);
spin_lock(&dmabuf->name_lock);
if (dmabuf->name)
seq_printf(m, "name:\t%s\n", dmabuf->name);
spin_unlock(&dmabuf->name_lock);
}
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
static const struct file_operations dma_buf_fops = {
.release = dma_buf_file_release,
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
.mmap = dma_buf_mmap_internal,
.llseek = dma_buf_llseek,
.poll = dma_buf_poll,
dma-buf: Add ioctls to allow userspace to flush The userspace might need some sort of cache coherency management e.g. when CPU and GPU domains are being accessed through dma-buf at the same time. To circumvent this problem there are begin/end coherency markers, that forward directly to existing dma-buf device drivers vfunc hooks. Userspace can make use of those markers through the DMA_BUF_IOCTL_SYNC ioctl. The sequence would be used like following: - mmap dma-buf fd - for each drawing/upload cycle in CPU 1. SYNC_START ioctl, 2. read/write to mmap area 3. SYNC_END ioctl. This can be repeated as often as you want (with the new data being consumed by the GPU or say scanout device) - munmap once you don't need the buffer any more v2 (Tiago): Fix header file type names (u64 -> __u64) v3 (Tiago): Add documentation. Use enum dma_buf_sync_flags to the begin/end dma-buf functions. Check for overflows in start/length. v4 (Tiago): use 2d regions for sync. v5 (Tiago): forget about 2d regions (v4); use _IOW in DMA_BUF_IOCTL_SYNC and remove range information from struct dma_buf_sync. v6 (Tiago): use __u64 structured padded flags instead enum. Adjust documentation about the recommendation on using sync ioctls. v7 (Tiago): Alex' nit on flags definition and being even more wording in the doc about sync usage. v9 (Tiago): remove useless is_dma_buf_file check. Fix sync.flags conditionals and its mask order check. Add <linux/types.h> include in dma-buf.h. Cc: Ville Syrjälä <ville.syrjala@linux.intel.com> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Reviewed-by: Stéphane Marchesin <marcheu@chromium.org> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> Signed-off-by: Tiago Vignatti <tiago.vignatti@intel.com> Reviewed-by: David Herrmann <dh.herrmann@gmail.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1455228291-29640-1-git-send-email-tiago.vignatti@intel.com
2016-02-12 06:04:51 +08:00
.unlocked_ioctl = dma_buf_ioctl,
#ifdef CONFIG_COMPAT
.compat_ioctl = dma_buf_ioctl,
#endif
.show_fdinfo = dma_buf_show_fdinfo,
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
};
/*
* is_dma_buf_file - Check if struct file* is associated with dma_buf
*/
static inline int is_dma_buf_file(struct file *file)
{
return file->f_op == &dma_buf_fops;
}
static struct file *dma_buf_getfile(struct dma_buf *dmabuf, int flags)
{
struct file *file;
struct inode *inode = alloc_anon_inode(dma_buf_mnt->mnt_sb);
if (IS_ERR(inode))
return ERR_CAST(inode);
inode->i_size = dmabuf->size;
inode_set_bytes(inode, dmabuf->size);
file = alloc_file_pseudo(inode, dma_buf_mnt, "dmabuf",
flags, &dma_buf_fops);
if (IS_ERR(file))
goto err_alloc_file;
file->f_flags = flags & (O_ACCMODE | O_NONBLOCK);
file->private_data = dmabuf;
file->f_path.dentry->d_fsdata = dmabuf;
return file;
err_alloc_file:
iput(inode);
return file;
}
/**
* DOC: dma buf device access
*
* For device DMA access to a shared DMA buffer the usual sequence of operations
* is fairly simple:
*
* 1. The exporter defines his exporter instance using
* DEFINE_DMA_BUF_EXPORT_INFO() and calls dma_buf_export() to wrap a private
* buffer object into a &dma_buf. It then exports that &dma_buf to userspace
* as a file descriptor by calling dma_buf_fd().
*
* 2. Userspace passes this file-descriptors to all drivers it wants this buffer
* to share with: First the filedescriptor is converted to a &dma_buf using
* dma_buf_get(). Then the buffer is attached to the device using
* dma_buf_attach().
*
* Up to this stage the exporter is still free to migrate or reallocate the
* backing storage.
*
* 3. Once the buffer is attached to all devices userspace can initiate DMA
* access to the shared buffer. In the kernel this is done by calling
* dma_buf_map_attachment() and dma_buf_unmap_attachment().
*
* 4. Once a driver is done with a shared buffer it needs to call
* dma_buf_detach() (after cleaning up any mappings) and then release the
* reference acquired with dma_buf_get by calling dma_buf_put().
*
* For the detailed semantics exporters are expected to implement see
* &dma_buf_ops.
*/
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
/**
* dma_buf_export - Creates a new dma_buf, and associates an anon file
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
* with this buffer, so it can be exported.
* Also connect the allocator specific data and ops to the buffer.
* Additionally, provide a name string for exporter; useful in debugging.
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
*
* @exp_info: [in] holds all the export related information provided
* by the exporter. see &struct dma_buf_export_info
* for further details.
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
*
* Returns, on success, a newly created dma_buf object, which wraps the
* supplied private data and operations for dma_buf_ops. On either missing
* ops, or error in allocating struct dma_buf, will return negative error.
*
* For most cases the easiest way to create @exp_info is through the
* %DEFINE_DMA_BUF_EXPORT_INFO macro.
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
*/
struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info)
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
{
struct dma_buf *dmabuf;
struct dma_resv *resv = exp_info->resv;
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
struct file *file;
size_t alloc_size = sizeof(struct dma_buf);
int ret;
if (!exp_info->resv)
alloc_size += sizeof(struct dma_resv);
else
/* prevent &dma_buf[1] == dma_buf->resv */
alloc_size += 1;
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
if (WARN_ON(!exp_info->priv
|| !exp_info->ops
|| !exp_info->ops->map_dma_buf
|| !exp_info->ops->unmap_dma_buf
|| !exp_info->ops->release)) {
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
return ERR_PTR(-EINVAL);
}
if (!try_module_get(exp_info->owner))
return ERR_PTR(-ENOENT);
dmabuf = kzalloc(alloc_size, GFP_KERNEL);
if (!dmabuf) {
ret = -ENOMEM;
goto err_module;
}
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
dmabuf->priv = exp_info->priv;
dmabuf->ops = exp_info->ops;
dmabuf->size = exp_info->size;
dmabuf->exp_name = exp_info->exp_name;
dmabuf->owner = exp_info->owner;
spin_lock_init(&dmabuf->name_lock);
init_waitqueue_head(&dmabuf->poll);
dmabuf->cb_excl.poll = dmabuf->cb_shared.poll = &dmabuf->poll;
dmabuf->cb_excl.active = dmabuf->cb_shared.active = 0;
if (!resv) {
resv = (struct dma_resv *)&dmabuf[1];
dma_resv_init(resv);
}
dmabuf->resv = resv;
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
file = dma_buf_getfile(dmabuf, exp_info->flags);
if (IS_ERR(file)) {
ret = PTR_ERR(file);
goto err_dmabuf;
}
file->f_mode |= FMODE_LSEEK;
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
dmabuf->file = file;
mutex_init(&dmabuf->lock);
INIT_LIST_HEAD(&dmabuf->attachments);
mutex_lock(&db_list.lock);
list_add(&dmabuf->list_node, &db_list.head);
mutex_unlock(&db_list.lock);
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
return dmabuf;
err_dmabuf:
kfree(dmabuf);
err_module:
module_put(exp_info->owner);
return ERR_PTR(ret);
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
}
EXPORT_SYMBOL_GPL(dma_buf_export);
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
/**
* dma_buf_fd - returns a file descriptor for the given dma_buf
* @dmabuf: [in] pointer to dma_buf for which fd is required.
* @flags: [in] flags to give to fd
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
*
* On success, returns an associated 'fd'. Else, returns error.
*/
int dma_buf_fd(struct dma_buf *dmabuf, int flags)
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
{
int fd;
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
if (!dmabuf || !dmabuf->file)
return -EINVAL;
fd = get_unused_fd_flags(flags);
if (fd < 0)
return fd;
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
fd_install(fd, dmabuf->file);
return fd;
}
EXPORT_SYMBOL_GPL(dma_buf_fd);
/**
* dma_buf_get - returns the dma_buf structure related to an fd
* @fd: [in] fd associated with the dma_buf to be returned
*
* On success, returns the dma_buf structure associated with an fd; uses
* file's refcounting done by fget to increase refcount. returns ERR_PTR
* otherwise.
*/
struct dma_buf *dma_buf_get(int fd)
{
struct file *file;
file = fget(fd);
if (!file)
return ERR_PTR(-EBADF);
if (!is_dma_buf_file(file)) {
fput(file);
return ERR_PTR(-EINVAL);
}
return file->private_data;
}
EXPORT_SYMBOL_GPL(dma_buf_get);
/**
* dma_buf_put - decreases refcount of the buffer
* @dmabuf: [in] buffer to reduce refcount of
*
* Uses file's refcounting done implicitly by fput().
*
* If, as a result of this call, the refcount becomes 0, the 'release' file
* operation related to this fd is called. It calls &dma_buf_ops.release vfunc
* in turn, and frees the memory allocated for dmabuf when exported.
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
*/
void dma_buf_put(struct dma_buf *dmabuf)
{
if (WARN_ON(!dmabuf || !dmabuf->file))
return;
fput(dmabuf->file);
}
EXPORT_SYMBOL_GPL(dma_buf_put);
/**
* dma_buf_attach - Add the device to dma_buf's attachments list; optionally,
* calls attach() of dma_buf_ops to allow device-specific attach functionality
* @dmabuf: [in] buffer to attach device to.
* @dev: [in] device to be attached.
*
* Returns struct dma_buf_attachment pointer for this attachment. Attachments
* must be cleaned up by calling dma_buf_detach().
*
* Returns:
*
* A pointer to newly created &dma_buf_attachment on success, or a negative
* error code wrapped into a pointer on failure.
*
* Note that this can fail if the backing storage of @dmabuf is in a place not
* accessible to @dev, and cannot be moved to a more suitable place. This is
* indicated with the error code -EBUSY.
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
*/
struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
struct device *dev)
{
struct dma_buf_attachment *attach;
int ret;
if (WARN_ON(!dmabuf || !dev))
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
return ERR_PTR(-EINVAL);
attach = kzalloc(sizeof(*attach), GFP_KERNEL);
if (!attach)
return ERR_PTR(-ENOMEM);
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
attach->dev = dev;
attach->dmabuf = dmabuf;
mutex_lock(&dmabuf->lock);
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
if (dmabuf->ops->attach) {
ret = dmabuf->ops->attach(dmabuf, attach);
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
if (ret)
goto err_attach;
}
list_add(&attach->node, &dmabuf->attachments);
mutex_unlock(&dmabuf->lock);
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
return attach;
err_attach:
kfree(attach);
mutex_unlock(&dmabuf->lock);
return ERR_PTR(ret);
}
EXPORT_SYMBOL_GPL(dma_buf_attach);
/**
* dma_buf_detach - Remove the given attachment from dmabuf's attachments list;
* optionally calls detach() of dma_buf_ops for device-specific detach
* @dmabuf: [in] buffer to detach from.
* @attach: [in] attachment to be detached; is free'd after this call.
*
* Clean up a device attachment obtained by calling dma_buf_attach().
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
*/
void dma_buf_detach(struct dma_buf *dmabuf, struct dma_buf_attachment *attach)
{
if (WARN_ON(!dmabuf || !attach))
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
return;
if (attach->sgt)
dmabuf->ops->unmap_dma_buf(attach, attach->sgt, attach->dir);
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
mutex_lock(&dmabuf->lock);
list_del(&attach->node);
if (dmabuf->ops->detach)
dmabuf->ops->detach(dmabuf, attach);
mutex_unlock(&dmabuf->lock);
kfree(attach);
}
EXPORT_SYMBOL_GPL(dma_buf_detach);
/**
* dma_buf_map_attachment - Returns the scatterlist table of the attachment;
* mapped into _device_ address space. Is a wrapper for map_dma_buf() of the
* dma_buf_ops.
* @attach: [in] attachment whose scatterlist is to be returned
* @direction: [in] direction of DMA transfer
*
* Returns sg_table containing the scatterlist to be returned; returns ERR_PTR
* on error. May return -EINTR if it is interrupted by a signal.
*
* A mapping must be unmapped by using dma_buf_unmap_attachment(). Note that
* the underlying backing storage is pinned for as long as a mapping exists,
* therefore users/importers should not hold onto a mapping for undue amounts of
* time.
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
*/
struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *attach,
enum dma_data_direction direction)
{
struct sg_table *sg_table;
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
might_sleep();
if (WARN_ON(!attach || !attach->dmabuf))
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
return ERR_PTR(-EINVAL);
if (attach->sgt) {
/*
* Two mappings with different directions for the same
* attachment are not allowed.
*/
if (attach->dir != direction &&
attach->dir != DMA_BIDIRECTIONAL)
return ERR_PTR(-EBUSY);
return attach->sgt;
}
sg_table = attach->dmabuf->ops->map_dma_buf(attach, direction);
if (!sg_table)
sg_table = ERR_PTR(-ENOMEM);
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
if (!IS_ERR(sg_table) && attach->dmabuf->ops->cache_sgt_mapping) {
attach->sgt = sg_table;
attach->dir = direction;
}
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
return sg_table;
}
EXPORT_SYMBOL_GPL(dma_buf_map_attachment);
/**
* dma_buf_unmap_attachment - unmaps and decreases usecount of the buffer;might
* deallocate the scatterlist associated. Is a wrapper for unmap_dma_buf() of
* dma_buf_ops.
* @attach: [in] attachment to unmap buffer from
* @sg_table: [in] scatterlist info of the buffer to unmap
* @direction: [in] direction of DMA transfer
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
*
* This unmaps a DMA mapping for @attached obtained by dma_buf_map_attachment().
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
*/
void dma_buf_unmap_attachment(struct dma_buf_attachment *attach,
struct sg_table *sg_table,
enum dma_data_direction direction)
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
{
might_sleep();
if (WARN_ON(!attach || !attach->dmabuf || !sg_table))
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
return;
if (attach->sgt == sg_table)
return;
attach->dmabuf->ops->unmap_dma_buf(attach, sg_table, direction);
dma-buf: Introduce dma buffer sharing mechanism This is the first step in defining a dma buffer sharing mechanism. A new buffer object dma_buf is added, with operations and API to allow easy sharing of this buffer object across devices. The framework allows: - creation of a buffer object, its association with a file pointer, and associated allocator-defined operations on that buffer. This operation is called the 'export' operation. - different devices to 'attach' themselves to this exported buffer object, to facilitate backing storage negotiation, using dma_buf_attach() API. - the exported buffer object to be shared with the other entity by asking for its 'file-descriptor (fd)', and sharing the fd across. - a received fd to get the buffer object back, where it can be accessed using the associated exporter-defined operations. - the exporter and user to share the scatterlist associated with this buffer object using map_dma_buf and unmap_dma_buf operations. Atleast one 'attach()' call is required to be made prior to calling the map_dma_buf() operation. Couple of building blocks in map_dma_buf() are added to ease introduction of sync'ing across exporter and users, and late allocation by the exporter. For this first version, this framework will work with certain conditions: - *ONLY* exporter will be allowed to mmap to userspace (outside of this framework - mmap is not a buffer object operation), - currently, *ONLY* users that do not need CPU access to the buffer are allowed. More details are there in the documentation patch. This is based on design suggestions from many people at the mini-summits[1], most notably from Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and Daniel Vetter <daniel@ffwll.ch>. The implementation is inspired from proof-of-concept patch-set from Tomasz Stanislawski <t.stanislaws@samsung.com>, who demonstrated buffer sharing between two v4l2 devices. [2] [1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement [2]: http://lwn.net/Articles/454389 Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org> Signed-off-by: Sumit Semwal <sumit.semwal@ti.com> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Dave Airlie <airlied@redhat.com> Reviewed-and-Tested-by: Rob Clark <rob.clark@linaro.org> Signed-off-by: Dave Airlie <airlied@redhat.com>
2011-12-26 17:23:15 +08:00
}
EXPORT_SYMBOL_GPL(dma_buf_unmap_attachment);
/**
* DOC: cpu access
*
* There are mutliple reasons for supporting CPU access to a dma buffer object:
*
* - Fallback operations in the kernel, for example when a device is connected
* over USB and the kernel needs to shuffle the data around first before
* sending it away. Cache coherency is handled by braketing any transactions
* with calls to dma_buf_begin_cpu_access() and dma_buf_end_cpu_access()
* access.
*
* To support dma_buf objects residing in highmem cpu access is page-based
* using an api similar to kmap. Accessing a dma_buf is done in aligned chunks
* of PAGE_SIZE size. Before accessing a chunk it needs to be mapped, which
* returns a pointer in kernel virtual address space. Afterwards the chunk
* needs to be unmapped again. There is no limit on how often a given chunk
* can be mapped and unmapped, i.e. the importer does not need to call
* begin_cpu_access again before mapping the same chunk again.
*
* Interfaces::
* void \*dma_buf_kmap(struct dma_buf \*, unsigned long);
* void dma_buf_kunmap(struct dma_buf \*, unsigned long, void \*);
*
* Implementing the functions is optional for exporters and for importers all
* the restrictions of using kmap apply.
*
* dma_buf kmap calls outside of the range specified in begin_cpu_access are
* undefined. If the range is not PAGE_SIZE aligned, kmap needs to succeed on
* the partial chunks at the beginning and end but may return stale or bogus
* data outside of the range (in these partial chunks).
*
* For some cases the overhead of kmap can be too high, a vmap interface
* is introduced. This interface should be used very carefully, as vmalloc
* space is a limited resources on many architectures.
*
* Interfaces::
* void \*dma_buf_vmap(struct dma_buf \*dmabuf)
* void dma_buf_vunmap(struct dma_buf \*dmabuf, void \*vaddr)
*
* The vmap call can fail if there is no vmap support in the exporter, or if
* it runs out of vmalloc space. Fallback to kmap should be implemented. Note
* that the dma-buf layer keeps a reference count for all vmap access and
* calls down into the exporter's vmap function only when no vmapping exists,
* and only unmaps it once. Protection against concurrent vmap/vunmap calls is
* provided by taking the dma_buf->lock mutex.
*
* - For full compatibility on the importer side with existing userspace
* interfaces, which might already support mmap'ing buffers. This is needed in
* many processing pipelines (e.g. feeding a software rendered image into a
* hardware pipeline, thumbnail creation, snapshots, ...). Also, Android's ION
* framework already supported this and for DMA buffer file descriptors to
* replace ION buffers mmap support was needed.
*
* There is no special interfaces, userspace simply calls mmap on the dma-buf
* fd. But like for CPU access there's a need to braket the actual access,
* which is handled by the ioctl (DMA_BUF_IOCTL_SYNC). Note that
* DMA_BUF_IOCTL_SYNC can fail with -EAGAIN or -EINTR, in which case it must
* be restarted.
*
* Some systems might need some sort of cache coherency management e.g. when
* CPU and GPU domains are being accessed through dma-buf at the same time.
* To circumvent this problem there are begin/end coherency markers, that
* forward directly to existing dma-buf device drivers vfunc hooks. Userspace
* can make use of those markers through the DMA_BUF_IOCTL_SYNC ioctl. The
* sequence would be used like following:
*
* - mmap dma-buf fd
* - for each drawing/upload cycle in CPU 1. SYNC_START ioctl, 2. read/write
* to mmap area 3. SYNC_END ioctl. This can be repeated as often as you
* want (with the new data being consumed by say the GPU or the scanout
* device)
* - munmap once you don't need the buffer any more
*
* For correctness and optimal performance, it is always required to use
* SYNC_START and SYNC_END before and after, respectively, when accessing the
* mapped address. Userspace cannot rely on coherent access, even when there
* are systems where it just works without calling these ioctls.
*
* - And as a CPU fallback in userspace processing pipelines.
*
* Similar to the motivation for kernel cpu access it is again important that
* the userspace code of a given importing subsystem can use the same
* interfaces with a imported dma-buf buffer object as with a native buffer
* object. This is especially important for drm where the userspace part of
* contemporary OpenGL, X, and other drivers is huge, and reworking them to
* use a different way to mmap a buffer rather invasive.
*
* The assumption in the current dma-buf interfaces is that redirecting the
* initial mmap is all that's needed. A survey of some of the existing
* subsystems shows that no driver seems to do any nefarious thing like
* syncing up with outstanding asynchronous processing on the device or
* allocating special resources at fault time. So hopefully this is good
* enough, since adding interfaces to intercept pagefaults and allow pte
* shootdowns would increase the complexity quite a bit.
*
* Interface::
* int dma_buf_mmap(struct dma_buf \*, struct vm_area_struct \*,
* unsigned long);
*
* If the importing subsystem simply provides a special-purpose mmap call to
* set up a mapping in userspace, calling do_mmap with dma_buf->file will
* equally achieve that for a dma-buf object.
*/
static int __dma_buf_begin_cpu_access(struct dma_buf *dmabuf,
enum dma_data_direction direction)
{
bool write = (direction == DMA_BIDIRECTIONAL ||
direction == DMA_TO_DEVICE);
struct dma_resv *resv = dmabuf->resv;
long ret;
/* Wait on any implicit rendering fences */
ret = dma_resv_wait_timeout_rcu(resv, write, true,
MAX_SCHEDULE_TIMEOUT);
if (ret < 0)
return ret;
return 0;
}
/**
* dma_buf_begin_cpu_access - Must be called before accessing a dma_buf from the
* cpu in the kernel context. Calls begin_cpu_access to allow exporter-specific
* preparations. Coherency is only guaranteed in the specified range for the
* specified access direction.
* @dmabuf: [in] buffer to prepare cpu access for.
* @direction: [in] length of range for cpu access.
*
* After the cpu access is complete the caller should call
* dma_buf_end_cpu_access(). Only when cpu access is braketed by both calls is
* it guaranteed to be coherent with other DMA access.
*
* Can return negative error values, returns 0 on success.
*/
int dma_buf_begin_cpu_access(struct dma_buf *dmabuf,
enum dma_data_direction direction)
{
int ret = 0;
if (WARN_ON(!dmabuf))
return -EINVAL;
if (dmabuf->ops->begin_cpu_access)
ret = dmabuf->ops->begin_cpu_access(dmabuf, direction);
/* Ensure that all fences are waited upon - but we first allow
* the native handler the chance to do so more efficiently if it
* chooses. A double invocation here will be reasonably cheap no-op.
*/
if (ret == 0)
ret = __dma_buf_begin_cpu_access(dmabuf, direction);
return ret;
}
EXPORT_SYMBOL_GPL(dma_buf_begin_cpu_access);
/**
* dma_buf_end_cpu_access - Must be called after accessing a dma_buf from the
* cpu in the kernel context. Calls end_cpu_access to allow exporter-specific
* actions. Coherency is only guaranteed in the specified range for the
* specified access direction.
* @dmabuf: [in] buffer to complete cpu access for.
* @direction: [in] length of range for cpu access.
*
* This terminates CPU access started with dma_buf_begin_cpu_access().
*
* Can return negative error values, returns 0 on success.
*/
dma-buf, drm, ion: Propagate error code from dma_buf_start_cpu_access() Drivers, especially i915.ko, can fail during the initial migration of a dma-buf for CPU access. However, the error code from the driver was not being propagated back to ioctl and so userspace was blissfully ignorant of the failure. Rendering corruption ensues. Whilst fixing the ioctl to return the error code from dma_buf_start_cpu_access(), also do the same for dma_buf_end_cpu_access(). For most drivers, dma_buf_end_cpu_access() cannot fail. i915.ko however, as most drivers would, wants to avoid being uninterruptible (as would be required to guarrantee no failure when flushing the buffer to the device). As userspace already has to handle errors from the SYNC_IOCTL, take advantage of this to be able to restart the syscall across signals. This fixes a coherency issue for i915.ko as well as reducing the uninterruptible hold upon its BKL, the struct_mutex. Fixes commit c11e391da2a8fe973c3c2398452000bed505851e Author: Daniel Vetter <daniel.vetter@ffwll.ch> Date: Thu Feb 11 20:04:51 2016 -0200 dma-buf: Add ioctls to allow userspace to flush Testcase: igt/gem_concurrent_blit/*dmabuf*interruptible Testcase: igt/prime_mmap_coherency/ioctl-errors Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tiago Vignatti <tiago.vignatti@intel.com> Cc: Stéphane Marchesin <marcheu@chromium.org> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Daniel Vetter <daniel.vetter@intel.com> CC: linux-media@vger.kernel.org Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: intel-gfx@lists.freedesktop.org Cc: devel@driverdev.osuosl.org Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1458331359-2634-1-git-send-email-chris@chris-wilson.co.uk
2016-03-19 04:02:39 +08:00
int dma_buf_end_cpu_access(struct dma_buf *dmabuf,
enum dma_data_direction direction)
{
dma-buf, drm, ion: Propagate error code from dma_buf_start_cpu_access() Drivers, especially i915.ko, can fail during the initial migration of a dma-buf for CPU access. However, the error code from the driver was not being propagated back to ioctl and so userspace was blissfully ignorant of the failure. Rendering corruption ensues. Whilst fixing the ioctl to return the error code from dma_buf_start_cpu_access(), also do the same for dma_buf_end_cpu_access(). For most drivers, dma_buf_end_cpu_access() cannot fail. i915.ko however, as most drivers would, wants to avoid being uninterruptible (as would be required to guarrantee no failure when flushing the buffer to the device). As userspace already has to handle errors from the SYNC_IOCTL, take advantage of this to be able to restart the syscall across signals. This fixes a coherency issue for i915.ko as well as reducing the uninterruptible hold upon its BKL, the struct_mutex. Fixes commit c11e391da2a8fe973c3c2398452000bed505851e Author: Daniel Vetter <daniel.vetter@ffwll.ch> Date: Thu Feb 11 20:04:51 2016 -0200 dma-buf: Add ioctls to allow userspace to flush Testcase: igt/gem_concurrent_blit/*dmabuf*interruptible Testcase: igt/prime_mmap_coherency/ioctl-errors Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tiago Vignatti <tiago.vignatti@intel.com> Cc: Stéphane Marchesin <marcheu@chromium.org> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Daniel Vetter <daniel.vetter@intel.com> CC: linux-media@vger.kernel.org Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: intel-gfx@lists.freedesktop.org Cc: devel@driverdev.osuosl.org Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1458331359-2634-1-git-send-email-chris@chris-wilson.co.uk
2016-03-19 04:02:39 +08:00
int ret = 0;
WARN_ON(!dmabuf);
if (dmabuf->ops->end_cpu_access)
dma-buf, drm, ion: Propagate error code from dma_buf_start_cpu_access() Drivers, especially i915.ko, can fail during the initial migration of a dma-buf for CPU access. However, the error code from the driver was not being propagated back to ioctl and so userspace was blissfully ignorant of the failure. Rendering corruption ensues. Whilst fixing the ioctl to return the error code from dma_buf_start_cpu_access(), also do the same for dma_buf_end_cpu_access(). For most drivers, dma_buf_end_cpu_access() cannot fail. i915.ko however, as most drivers would, wants to avoid being uninterruptible (as would be required to guarrantee no failure when flushing the buffer to the device). As userspace already has to handle errors from the SYNC_IOCTL, take advantage of this to be able to restart the syscall across signals. This fixes a coherency issue for i915.ko as well as reducing the uninterruptible hold upon its BKL, the struct_mutex. Fixes commit c11e391da2a8fe973c3c2398452000bed505851e Author: Daniel Vetter <daniel.vetter@ffwll.ch> Date: Thu Feb 11 20:04:51 2016 -0200 dma-buf: Add ioctls to allow userspace to flush Testcase: igt/gem_concurrent_blit/*dmabuf*interruptible Testcase: igt/prime_mmap_coherency/ioctl-errors Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tiago Vignatti <tiago.vignatti@intel.com> Cc: Stéphane Marchesin <marcheu@chromium.org> Cc: David Herrmann <dh.herrmann@gmail.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Daniel Vetter <daniel.vetter@intel.com> CC: linux-media@vger.kernel.org Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: intel-gfx@lists.freedesktop.org Cc: devel@driverdev.osuosl.org Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Link: http://patchwork.freedesktop.org/patch/msgid/1458331359-2634-1-git-send-email-chris@chris-wilson.co.uk
2016-03-19 04:02:39 +08:00
ret = dmabuf->ops->end_cpu_access(dmabuf, direction);
return ret;
}
EXPORT_SYMBOL_GPL(dma_buf_end_cpu_access);
/**
* dma_buf_kmap - Map a page of the buffer object into kernel address space. The
* same restrictions as for kmap and friends apply.
* @dmabuf: [in] buffer to map page from.
* @page_num: [in] page in PAGE_SIZE units to map.
*
* This call must always succeed, any necessary preparations that might fail
* need to be done in begin_cpu_access.
*/
void *dma_buf_kmap(struct dma_buf *dmabuf, unsigned long page_num)
{
WARN_ON(!dmabuf);
if (!dmabuf->ops->map)
return NULL;
return dmabuf->ops->map(dmabuf, page_num);
}
EXPORT_SYMBOL_GPL(dma_buf_kmap);
/**
* dma_buf_kunmap - Unmap a page obtained by dma_buf_kmap.
* @dmabuf: [in] buffer to unmap page from.
* @page_num: [in] page in PAGE_SIZE units to unmap.
* @vaddr: [in] kernel space pointer obtained from dma_buf_kmap.
*
* This call must always succeed.
*/
void dma_buf_kunmap(struct dma_buf *dmabuf, unsigned long page_num,
void *vaddr)
{
WARN_ON(!dmabuf);
if (dmabuf->ops->unmap)
dmabuf->ops->unmap(dmabuf, page_num, vaddr);
}
EXPORT_SYMBOL_GPL(dma_buf_kunmap);
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
/**
* dma_buf_mmap - Setup up a userspace mmap with the given vma
* @dmabuf: [in] buffer that should back the vma
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
* @vma: [in] vma for the mmap
* @pgoff: [in] offset in pages where this mmap should start within the
* dma-buf buffer.
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
*
* This function adjusts the passed in vma so that it points at the file of the
* dma_buf operation. It also adjusts the starting pgoff and does bounds
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
* checking on the size of the vma. Then it calls the exporters mmap function to
* set up the mapping.
*
* Can return negative error values, returns 0 on success.
*/
int dma_buf_mmap(struct dma_buf *dmabuf, struct vm_area_struct *vma,
unsigned long pgoff)
{
struct file *oldfile;
int ret;
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
if (WARN_ON(!dmabuf || !vma))
return -EINVAL;
/* check if buffer supports mmap */
if (!dmabuf->ops->mmap)
return -EINVAL;
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
/* check for offset overflow */
if (pgoff + vma_pages(vma) < pgoff)
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
return -EOVERFLOW;
/* check for overflowing the buffer's size */
if (pgoff + vma_pages(vma) >
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
dmabuf->size >> PAGE_SHIFT)
return -EINVAL;
/* readjust the vma */
get_file(dmabuf->file);
oldfile = vma->vm_file;
vma->vm_file = dmabuf->file;
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
vma->vm_pgoff = pgoff;
ret = dmabuf->ops->mmap(dmabuf, vma);
if (ret) {
/* restore old parameters on failure */
vma->vm_file = oldfile;
fput(dmabuf->file);
} else {
if (oldfile)
fput(oldfile);
}
return ret;
dma-buf: mmap support Compared to Rob Clark's RFC I've ditched the prepare/finish hooks and corresponding ioctls on the dma_buf file. The major reason for that is that many people seem to be under the impression that this is also for synchronization with outstanding asynchronous processsing. I'm pretty massively opposed to this because: - It boils down reinventing a new rather general-purpose userspace synchronization interface. If we look at things like futexes, this is hard to get right. - Furthermore a lot of kernel code has to interact with this synchronization primitive. This smells a look like the dri1 hw_lock, a horror show I prefer not to reinvent. - Even more fun is that multiple different subsystems would interact here, so we have plenty of opportunities to create funny deadlock scenarios. I think synchronization is a wholesale different problem from data sharing and should be tackled as an orthogonal problem. Now we could demand that prepare/finish may only ensure cache coherency (as Rob intended), but that runs up into the next problem: We not only need mmap support to facilitate sw-only processing nodes in a pipeline (without jumping through hoops by importing the dma_buf into some sw-access only importer), which allows for a nicer ION->dma-buf upgrade path for existing Android userspace. We also need mmap support for existing importing subsystems to support existing userspace libraries. And a loot of these subsystems are expected to export coherent userspace mappings. So prepare/finish can only ever be optional and the exporter /needs/ to support coherent mappings. Given that mmap access is always somewhat fallback-y in nature I've decided to drop this optimization, instead of just making it optional. If we demonstrate a clear need for this, supported by benchmark results, we can always add it in again later as an optional extension. Other differences compared to Rob's RFC is the above mentioned support for mapping a dma-buf through facilities provided by the importer. Which results in mmap support no longer being optional. Note that this dma-buf mmap patch does _not_ support every possible insanity an existing subsystem could pull of with mmap: Because it does not allow to intercept pagefaults and shoot down ptes importing subsystems can't add some magic of their own at these points (e.g. to automatically synchronize with outstanding rendering or set up some special resources). I've done a cursory read through a few mmap implementions of various subsytems and I'm hopeful that we can avoid this (and the complexity it'd bring with it). Additonally I've extended the documentation a bit to explain the hows and whys of this mmap extension. In case we ever want to add support for explicitly cache maneged userspace mmap with a prepare/finish ioctl pair, we could specify that userspace needs to mmap a different part of the dma_buf, e.g. the range starting at dma_buf->size up to dma_buf->size*2. This works because the size of a dma_buf is invariant over it's lifetime. The exporter would obviously need to fall back to coherent mappings for both ranges if a legacy clients maps the coherent range and the architecture cannot suppor conflicting caching policies. Also, this would obviously be optional and userspace needs to be able to fall back to coherent mappings. v2: - Spelling fixes from Rob Clark. - Compile fix for !DMA_BUF from Rob Clark. - Extend commit message to explain how explicitly cache managed mmap support could be added later. - Extend the documentation with implementations notes for exporters that need to manually fake coherency. v3: - dma_buf pointer initialization goof-up noticed by Rebecca Schultz Zavin. Cc: Rob Clark <rob.clark@linaro.org> Cc: Rebecca Schultz Zavin <rebecca@android.com> Acked-by: Rob Clark <rob.clark@linaro.org> Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Sumit Semwal <sumit.semwal@linaro.org>
2012-04-24 17:08:52 +08:00
}
EXPORT_SYMBOL_GPL(dma_buf_mmap);
/**
* dma_buf_vmap - Create virtual mapping for the buffer object into kernel
* address space. Same restrictions as for vmap and friends apply.
* @dmabuf: [in] buffer to vmap
*
* This call may fail due to lack of virtual mapping address space.
* These calls are optional in drivers. The intended use for them
* is for mapping objects linear in kernel space for high use objects.
* Please attempt to use kmap/kunmap before thinking about these interfaces.
*
* Returns NULL on error.
*/
void *dma_buf_vmap(struct dma_buf *dmabuf)
{
void *ptr;
if (WARN_ON(!dmabuf))
return NULL;
if (!dmabuf->ops->vmap)
return NULL;
mutex_lock(&dmabuf->lock);
if (dmabuf->vmapping_counter) {
dmabuf->vmapping_counter++;
BUG_ON(!dmabuf->vmap_ptr);
ptr = dmabuf->vmap_ptr;
goto out_unlock;
}
BUG_ON(dmabuf->vmap_ptr);
ptr = dmabuf->ops->vmap(dmabuf);
if (WARN_ON_ONCE(IS_ERR(ptr)))
ptr = NULL;
if (!ptr)
goto out_unlock;
dmabuf->vmap_ptr = ptr;
dmabuf->vmapping_counter = 1;
out_unlock:
mutex_unlock(&dmabuf->lock);
return ptr;
}
EXPORT_SYMBOL_GPL(dma_buf_vmap);
/**
* dma_buf_vunmap - Unmap a vmap obtained by dma_buf_vmap.
* @dmabuf: [in] buffer to vunmap
* @vaddr: [in] vmap to vunmap
*/
void dma_buf_vunmap(struct dma_buf *dmabuf, void *vaddr)
{
if (WARN_ON(!dmabuf))
return;
BUG_ON(!dmabuf->vmap_ptr);
BUG_ON(dmabuf->vmapping_counter == 0);
BUG_ON(dmabuf->vmap_ptr != vaddr);
mutex_lock(&dmabuf->lock);
if (--dmabuf->vmapping_counter == 0) {
if (dmabuf->ops->vunmap)
dmabuf->ops->vunmap(dmabuf, vaddr);
dmabuf->vmap_ptr = NULL;
}
mutex_unlock(&dmabuf->lock);
}
EXPORT_SYMBOL_GPL(dma_buf_vunmap);
#ifdef CONFIG_DEBUG_FS
static int dma_buf_debug_show(struct seq_file *s, void *unused)
{
int ret;
struct dma_buf *buf_obj;
struct dma_buf_attachment *attach_obj;
struct dma_resv *robj;
struct dma_resv_list *fobj;
struct dma_fence *fence;
unsigned seq;
int count = 0, attach_count, shared_count, i;
size_t size = 0;
ret = mutex_lock_interruptible(&db_list.lock);
if (ret)
return ret;
seq_puts(s, "\nDma-buf Objects:\n");
seq_printf(s, "%-8s\t%-8s\t%-8s\t%-8s\texp_name\t%-8s\n",
"size", "flags", "mode", "count", "ino");
list_for_each_entry(buf_obj, &db_list.head, list_node) {
ret = mutex_lock_interruptible(&buf_obj->lock);
if (ret) {
seq_puts(s,
"\tERROR locking buffer object: skipping\n");
continue;
}
seq_printf(s, "%08zu\t%08x\t%08x\t%08ld\t%s\t%08lu\t%s\n",
buf_obj->size,
buf_obj->file->f_flags, buf_obj->file->f_mode,
file_count(buf_obj->file),
buf_obj->exp_name,
file_inode(buf_obj->file)->i_ino,
buf_obj->name ?: "");
robj = buf_obj->resv;
while (true) {
seq = read_seqcount_begin(&robj->seq);
rcu_read_lock();
fobj = rcu_dereference(robj->fence);
shared_count = fobj ? fobj->shared_count : 0;
fence = rcu_dereference(robj->fence_excl);
if (!read_seqcount_retry(&robj->seq, seq))
break;
rcu_read_unlock();
}
if (fence)
seq_printf(s, "\tExclusive fence: %s %s %ssignalled\n",
fence->ops->get_driver_name(fence),
fence->ops->get_timeline_name(fence),
dma_fence_is_signaled(fence) ? "" : "un");
for (i = 0; i < shared_count; i++) {
fence = rcu_dereference(fobj->shared[i]);
if (!dma_fence_get_rcu(fence))
continue;
seq_printf(s, "\tShared fence: %s %s %ssignalled\n",
fence->ops->get_driver_name(fence),
fence->ops->get_timeline_name(fence),
dma_fence_is_signaled(fence) ? "" : "un");
dma_fence_put(fence);
}
rcu_read_unlock();
seq_puts(s, "\tAttached Devices:\n");
attach_count = 0;
list_for_each_entry(attach_obj, &buf_obj->attachments, node) {
seq_printf(s, "\t%s\n", dev_name(attach_obj->dev));
attach_count++;
}
seq_printf(s, "Total %d devices attached\n\n",
attach_count);
count++;
size += buf_obj->size;
mutex_unlock(&buf_obj->lock);
}
seq_printf(s, "\nTotal %d objects, %zu bytes\n", count, size);
mutex_unlock(&db_list.lock);
return 0;
}
DEFINE_SHOW_ATTRIBUTE(dma_buf_debug);
static struct dentry *dma_buf_debugfs_dir;
static int dma_buf_init_debugfs(void)
{
struct dentry *d;
int err = 0;
d = debugfs_create_dir("dma_buf", NULL);
if (IS_ERR(d))
return PTR_ERR(d);
dma_buf_debugfs_dir = d;
d = debugfs_create_file("bufinfo", S_IRUGO, dma_buf_debugfs_dir,
NULL, &dma_buf_debug_fops);
if (IS_ERR(d)) {
pr_debug("dma_buf: debugfs: failed to create node bufinfo\n");
debugfs_remove_recursive(dma_buf_debugfs_dir);
dma_buf_debugfs_dir = NULL;
err = PTR_ERR(d);
}
return err;
}
static void dma_buf_uninit_debugfs(void)
{
debugfs_remove_recursive(dma_buf_debugfs_dir);
}
#else
static inline int dma_buf_init_debugfs(void)
{
return 0;
}
static inline void dma_buf_uninit_debugfs(void)
{
}
#endif
static int __init dma_buf_init(void)
{
dma_buf_mnt = kern_mount(&dma_buf_fs_type);
if (IS_ERR(dma_buf_mnt))
return PTR_ERR(dma_buf_mnt);
mutex_init(&db_list.lock);
INIT_LIST_HEAD(&db_list.head);
dma_buf_init_debugfs();
return 0;
}
subsys_initcall(dma_buf_init);
static void __exit dma_buf_deinit(void)
{
dma_buf_uninit_debugfs();
kern_unmount(dma_buf_mnt);
}
__exitcall(dma_buf_deinit);