2010-08-07 18:01:23 +08:00
|
|
|
/*
|
|
|
|
* Copyright © 2008-2010 Intel Corporation
|
|
|
|
*
|
|
|
|
* Permission is hereby granted, free of charge, to any person obtaining a
|
|
|
|
* copy of this software and associated documentation files (the "Software"),
|
|
|
|
* to deal in the Software without restriction, including without limitation
|
|
|
|
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
|
|
|
|
* and/or sell copies of the Software, and to permit persons to whom the
|
|
|
|
* Software is furnished to do so, subject to the following conditions:
|
|
|
|
*
|
|
|
|
* The above copyright notice and this permission notice (including the next
|
|
|
|
* paragraph) shall be included in all copies or substantial portions of the
|
|
|
|
* Software.
|
|
|
|
*
|
|
|
|
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
|
|
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
|
|
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
|
|
|
|
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
|
|
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
|
|
|
|
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
|
|
|
|
* IN THE SOFTWARE.
|
|
|
|
*
|
|
|
|
* Authors:
|
|
|
|
* Eric Anholt <eric@anholt.net>
|
|
|
|
* Chris Wilson <chris@chris-wilson.co.uuk>
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
2012-10-03 01:01:07 +08:00
|
|
|
#include <drm/drmP.h>
|
|
|
|
#include <drm/i915_drm.h>
|
2014-01-20 18:17:37 +08:00
|
|
|
|
|
|
|
#include "i915_drv.h"
|
|
|
|
#include "intel_drv.h"
|
2011-02-03 19:57:46 +08:00
|
|
|
#include "i915_trace.h"
|
2010-08-07 18:01:23 +08:00
|
|
|
|
2017-10-12 20:57:26 +08:00
|
|
|
I915_SELFTEST_DECLARE(static struct igt_evict_ctl {
|
|
|
|
bool fail_if_busy:1;
|
|
|
|
} igt_evict_ctl;)
|
|
|
|
|
2017-10-12 20:57:24 +08:00
|
|
|
static bool ggtt_is_idle(struct drm_i915_private *i915)
|
2016-08-04 23:32:17 +08:00
|
|
|
{
|
2017-10-12 20:57:24 +08:00
|
|
|
struct intel_engine_cs *engine;
|
|
|
|
enum intel_engine_id id;
|
2016-08-04 23:32:17 +08:00
|
|
|
|
2017-10-12 20:57:24 +08:00
|
|
|
if (i915->gt.active_requests)
|
|
|
|
return false;
|
2016-10-28 20:58:58 +08:00
|
|
|
|
2017-10-12 20:57:24 +08:00
|
|
|
for_each_engine(engine, i915, id) {
|
2017-10-25 06:08:55 +08:00
|
|
|
if (!intel_engine_has_kernel_context(engine))
|
2017-10-12 20:57:24 +08:00
|
|
|
return false;
|
|
|
|
}
|
2016-08-04 23:32:17 +08:00
|
|
|
|
2017-10-12 20:57:24 +08:00
|
|
|
return true;
|
2016-08-04 23:32:17 +08:00
|
|
|
}
|
|
|
|
|
drm/i915: Eliminate lots of iterations over the execobjects array
The major scaling bottleneck in execbuffer is the processing of the
execobjects. Creating an auxiliary list is inefficient when compared to
using the execobject array we already have allocated.
Reservation is then split into phases. As we lookup up the VMA, we
try and bind it back into active location. Only if that fails, do we add
it to the unbound list for phase 2. In phase 2, we try and add all those
objects that could not fit into their previous location, with fallback
to retrying all objects and evicting the VM in case of severe
fragmentation. (This is the same as before, except that phase 1 is now
done inline with looking up the VMA to avoid an iteration over the
execobject array. In the ideal case, we eliminate the separate reservation
phase). During the reservation phase, we only evict from the VM between
passes (rather than currently as we try to fit every new VMA). In
testing with Unreal Engine's Atlantis demo which stresses the eviction
logic on gen7 class hardware, this speed up the framerate by a factor of
2.
The second loop amalgamation is between move_to_gpu and move_to_active.
As we always submit the request, even if incomplete, we can use the
current request to track active VMA as we perform the flushes and
synchronisation required.
The next big advancement is to avoid copying back to the user any
execobjects and relocations that are not changed.
v2: Add a Theory of Operation spiel.
v3: Fall back to slow relocations in preparation for flushing userptrs.
v4: Document struct members, factor out eb_validate_vma(), add a few
more comments to explain some magic and hide other magic behind macros.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2017-06-16 22:05:19 +08:00
|
|
|
static int ggtt_flush(struct drm_i915_private *i915)
|
|
|
|
{
|
|
|
|
int err;
|
|
|
|
|
|
|
|
/* Not everything in the GGTT is tracked via vma (otherwise we
|
|
|
|
* could evict as required with minimal stalling) so we are forced
|
|
|
|
* to idle the GPU and explicitly retire outstanding requests in
|
|
|
|
* the hopes that we can then remove contexts and the like only
|
|
|
|
* bound by their active reference.
|
|
|
|
*/
|
|
|
|
err = i915_gem_switch_to_kernel_context(i915);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
|
|
|
|
err = i915_gem_wait_for_idle(i915,
|
|
|
|
I915_WAIT_INTERRUPTIBLE |
|
|
|
|
I915_WAIT_LOCKED);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
|
2017-10-25 06:08:55 +08:00
|
|
|
GEM_BUG_ON(!ggtt_is_idle(i915));
|
drm/i915: Eliminate lots of iterations over the execobjects array
The major scaling bottleneck in execbuffer is the processing of the
execobjects. Creating an auxiliary list is inefficient when compared to
using the execobject array we already have allocated.
Reservation is then split into phases. As we lookup up the VMA, we
try and bind it back into active location. Only if that fails, do we add
it to the unbound list for phase 2. In phase 2, we try and add all those
objects that could not fit into their previous location, with fallback
to retrying all objects and evicting the VM in case of severe
fragmentation. (This is the same as before, except that phase 1 is now
done inline with looking up the VMA to avoid an iteration over the
execobject array. In the ideal case, we eliminate the separate reservation
phase). During the reservation phase, we only evict from the VM between
passes (rather than currently as we try to fit every new VMA). In
testing with Unreal Engine's Atlantis demo which stresses the eviction
logic on gen7 class hardware, this speed up the framerate by a factor of
2.
The second loop amalgamation is between move_to_gpu and move_to_active.
As we always submit the request, even if incomplete, we can use the
current request to track active VMA as we perform the flushes and
synchronisation required.
The next big advancement is to avoid copying back to the user any
execobjects and relocations that are not changed.
v2: Add a Theory of Operation spiel.
v3: Fall back to slow relocations in preparation for flushing userptrs.
v4: Document struct members, factor out eb_validate_vma(), add a few
more comments to explain some magic and hide other magic behind macros.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2017-06-16 22:05:19 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2010-08-07 18:01:24 +08:00
|
|
|
static bool
|
2016-12-22 16:36:29 +08:00
|
|
|
mark_free(struct drm_mm_scan *scan,
|
|
|
|
struct i915_vma *vma,
|
|
|
|
unsigned int flags,
|
|
|
|
struct list_head *unwind)
|
2010-08-07 18:01:23 +08:00
|
|
|
{
|
2016-08-04 23:32:30 +08:00
|
|
|
if (i915_vma_is_pinned(vma))
|
2012-04-24 22:47:30 +08:00
|
|
|
return false;
|
|
|
|
|
2017-10-09 16:43:57 +08:00
|
|
|
if (flags & PIN_NONFAULT && i915_vma_has_userfault(vma))
|
2016-08-19 00:17:05 +08:00
|
|
|
return false;
|
|
|
|
|
2017-06-15 16:14:35 +08:00
|
|
|
list_add(&vma->evict_link, unwind);
|
2016-12-22 16:36:29 +08:00
|
|
|
return drm_mm_scan_add_block(scan, &vma->node);
|
2010-08-07 18:01:23 +08:00
|
|
|
}
|
|
|
|
|
2014-01-30 05:07:11 +08:00
|
|
|
/**
|
|
|
|
* i915_gem_evict_something - Evict vmas to make room for binding a new one
|
|
|
|
* @vm: address space to evict from
|
2015-01-05 21:36:59 +08:00
|
|
|
* @min_size: size of the desired free space
|
2014-01-30 05:07:11 +08:00
|
|
|
* @alignment: alignment constraint of the desired free space
|
|
|
|
* @cache_level: cache_level for the desired space
|
2015-01-05 21:36:59 +08:00
|
|
|
* @start: start (inclusive) of the range from which to evict objects
|
|
|
|
* @end: end (exclusive) of the range from which to evict objects
|
|
|
|
* @flags: additional flags to control the eviction algorithm
|
2014-01-30 05:07:11 +08:00
|
|
|
*
|
|
|
|
* This function will try to evict vmas until a free space satisfying the
|
|
|
|
* requirements is found. Callers must check first whether any such hole exists
|
|
|
|
* already before calling this function.
|
|
|
|
*
|
|
|
|
* This function is used by the object/vma binding code.
|
|
|
|
*
|
2015-03-18 21:47:59 +08:00
|
|
|
* Since this function is only used to free up virtual address space it only
|
|
|
|
* ignores pinned vmas, and not object where the backing storage itself is
|
|
|
|
* pinned. Hence obj->pages_pin_count does not protect against eviction.
|
|
|
|
*
|
2014-01-30 05:07:11 +08:00
|
|
|
* To clarify: This is for freeing up virtual address space, not for freeing
|
|
|
|
* memory in e.g. the shrinker.
|
|
|
|
*/
|
2010-08-07 18:01:23 +08:00
|
|
|
int
|
2016-08-04 23:32:18 +08:00
|
|
|
i915_gem_evict_something(struct i915_address_space *vm,
|
2016-08-04 23:32:22 +08:00
|
|
|
u64 min_size, u64 alignment,
|
|
|
|
unsigned cache_level,
|
|
|
|
u64 start, u64 end,
|
2014-02-14 21:01:11 +08:00
|
|
|
unsigned flags)
|
2010-08-07 18:01:23 +08:00
|
|
|
{
|
2016-11-29 17:50:08 +08:00
|
|
|
struct drm_i915_private *dev_priv = vm->i915;
|
2016-12-22 16:36:29 +08:00
|
|
|
struct drm_mm_scan scan;
|
2016-08-04 23:32:17 +08:00
|
|
|
struct list_head eviction_list;
|
|
|
|
struct list_head *phases[] = {
|
|
|
|
&vm->inactive_list,
|
|
|
|
&vm->active_list,
|
|
|
|
NULL,
|
|
|
|
}, **phase;
|
|
|
|
struct i915_vma *vma, *next;
|
2016-12-22 16:36:36 +08:00
|
|
|
struct drm_mm_node *node;
|
2017-02-03 05:04:38 +08:00
|
|
|
enum drm_mm_insert_mode mode;
|
2016-08-04 23:32:17 +08:00
|
|
|
int ret;
|
2010-08-07 18:01:23 +08:00
|
|
|
|
2016-11-29 17:50:08 +08:00
|
|
|
lockdep_assert_held(&vm->i915->drm.struct_mutex);
|
2016-08-04 23:32:18 +08:00
|
|
|
trace_i915_gem_evict(vm, min_size, alignment, flags);
|
2011-02-03 19:57:46 +08:00
|
|
|
|
2010-08-07 18:01:24 +08:00
|
|
|
/*
|
|
|
|
* The goal is to evict objects and amalgamate space in LRU order.
|
|
|
|
* The oldest idle objects reside on the inactive list, which is in
|
2016-08-04 23:32:17 +08:00
|
|
|
* retirement order. The next objects to retire are those in flight,
|
|
|
|
* on the active list, again in retirement order.
|
2010-08-07 18:01:24 +08:00
|
|
|
*
|
|
|
|
* The retirement sequence is thus:
|
|
|
|
* 1. Inactive objects (already retired)
|
2016-08-04 23:32:17 +08:00
|
|
|
* 2. Active objects (will stall on unbinding)
|
2010-08-07 18:01:24 +08:00
|
|
|
*
|
|
|
|
* On each list, the oldest objects lie at the HEAD with the freshest
|
|
|
|
* object on the TAIL.
|
|
|
|
*/
|
2017-02-03 05:04:38 +08:00
|
|
|
mode = DRM_MM_INSERT_BEST;
|
|
|
|
if (flags & PIN_HIGH)
|
|
|
|
mode = DRM_MM_INSERT_HIGH;
|
|
|
|
if (flags & PIN_MAPPABLE)
|
|
|
|
mode = DRM_MM_INSERT_LOW;
|
2016-12-22 16:36:31 +08:00
|
|
|
drm_mm_scan_init_with_range(&scan, &vm->mm,
|
|
|
|
min_size, alignment, cache_level,
|
2017-02-03 05:04:38 +08:00
|
|
|
start, end, mode);
|
2010-08-07 18:01:24 +08:00
|
|
|
|
2017-10-12 20:57:24 +08:00
|
|
|
/*
|
|
|
|
* Retire before we search the active list. Although we have
|
2016-12-09 23:05:55 +08:00
|
|
|
* reasonable accuracy in our retirement lists, we may have
|
|
|
|
* a stray pin (preventing eviction) that can only be resolved by
|
|
|
|
* retiring.
|
|
|
|
*/
|
|
|
|
if (!(flags & PIN_NONBLOCK))
|
|
|
|
i915_gem_retire_requests(dev_priv);
|
|
|
|
else
|
2016-08-04 23:32:17 +08:00
|
|
|
phases[1] = NULL;
|
2010-08-07 18:01:23 +08:00
|
|
|
|
2016-08-04 23:32:17 +08:00
|
|
|
search_again:
|
|
|
|
INIT_LIST_HEAD(&eviction_list);
|
|
|
|
phase = phases;
|
|
|
|
do {
|
|
|
|
list_for_each_entry(vma, *phase, vm_link)
|
2016-12-22 16:36:29 +08:00
|
|
|
if (mark_free(&scan, vma, flags, &eviction_list))
|
2016-08-04 23:32:17 +08:00
|
|
|
goto found;
|
|
|
|
} while (*++phase);
|
2010-08-07 18:01:24 +08:00
|
|
|
|
|
|
|
/* Nothing found, clean up and bail out! */
|
2017-06-15 16:14:35 +08:00
|
|
|
list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
|
2016-12-22 16:36:29 +08:00
|
|
|
ret = drm_mm_scan_remove_block(&scan, &vma->node);
|
2010-08-07 18:01:24 +08:00
|
|
|
BUG_ON(ret);
|
|
|
|
}
|
|
|
|
|
2017-10-12 20:57:24 +08:00
|
|
|
/*
|
|
|
|
* Can we unpin some objects such as idle hw contents,
|
2016-08-04 23:32:17 +08:00
|
|
|
* or pending flips? But since only the GGTT has global entries
|
|
|
|
* such as scanouts, rinbuffers and contexts, we can skip the
|
|
|
|
* purge when inspecting per-process local address spaces.
|
2010-08-07 18:01:24 +08:00
|
|
|
*/
|
2016-08-04 23:32:17 +08:00
|
|
|
if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK)
|
2014-01-20 18:17:37 +08:00
|
|
|
return -ENOSPC;
|
2013-12-09 18:37:24 +08:00
|
|
|
|
2017-10-12 20:57:24 +08:00
|
|
|
/*
|
|
|
|
* Not everything in the GGTT is tracked via VMA using
|
|
|
|
* i915_vma_move_to_active(), otherwise we could evict as required
|
|
|
|
* with minimal stalling. Instead we are forced to idle the GPU and
|
|
|
|
* explicitly retire outstanding requests which will then remove
|
|
|
|
* the pinning for active objects such as contexts and ring,
|
|
|
|
* enabling us to evict them on the next iteration.
|
|
|
|
*
|
|
|
|
* To ensure that all user contexts are evictable, we perform
|
|
|
|
* a switch to the perma-pinned kernel context. This all also gives
|
|
|
|
* us a termination condition, when the last retired context is
|
|
|
|
* the kernel's there is no more we can evict.
|
|
|
|
*/
|
|
|
|
if (!ggtt_is_idle(dev_priv)) {
|
2017-10-12 20:57:26 +08:00
|
|
|
if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy))
|
|
|
|
return -EBUSY;
|
|
|
|
|
2017-10-12 20:57:24 +08:00
|
|
|
ret = ggtt_flush(dev_priv);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
2014-01-20 18:17:37 +08:00
|
|
|
|
2017-10-25 04:50:53 +08:00
|
|
|
cond_resched();
|
2017-10-12 20:57:24 +08:00
|
|
|
goto search_again;
|
|
|
|
}
|
2016-08-04 23:32:17 +08:00
|
|
|
|
2017-10-12 20:57:24 +08:00
|
|
|
/*
|
|
|
|
* If we still have pending pageflip completions, drop
|
|
|
|
* back to userspace to give our workqueues time to
|
|
|
|
* acquire our locks and unpin the old scanouts.
|
|
|
|
*/
|
|
|
|
return intel_has_pending_fb_unpin(dev_priv) ? -EAGAIN : -ENOSPC;
|
2010-08-07 18:01:24 +08:00
|
|
|
|
|
|
|
found:
|
2010-09-30 05:23:05 +08:00
|
|
|
/* drm_mm doesn't allow any other other operations while
|
2016-08-04 23:32:17 +08:00
|
|
|
* scanning, therefore store to-be-evicted objects on a
|
|
|
|
* temporary list and take a reference for all before
|
|
|
|
* calling unbind (which may remove the active reference
|
|
|
|
* of any of our objects, thus corrupting the list).
|
|
|
|
*/
|
2017-06-15 16:14:35 +08:00
|
|
|
list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
|
2016-12-22 16:36:29 +08:00
|
|
|
if (drm_mm_scan_remove_block(&scan, &vma->node))
|
2016-08-04 23:32:30 +08:00
|
|
|
__i915_vma_pin(vma);
|
2016-08-04 23:32:17 +08:00
|
|
|
else
|
2017-06-15 16:14:35 +08:00
|
|
|
list_del(&vma->evict_link);
|
2010-08-07 18:01:24 +08:00
|
|
|
}
|
2010-08-07 18:01:23 +08:00
|
|
|
|
2010-08-07 18:01:24 +08:00
|
|
|
/* Unbinding will emit any required flushes */
|
2017-01-05 23:59:40 +08:00
|
|
|
ret = 0;
|
2017-06-15 16:14:35 +08:00
|
|
|
list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
|
2016-08-04 23:32:30 +08:00
|
|
|
__i915_vma_unpin(vma);
|
2010-09-30 05:23:05 +08:00
|
|
|
if (ret == 0)
|
2013-08-14 17:38:34 +08:00
|
|
|
ret = i915_vma_unbind(vma);
|
2010-08-07 18:01:23 +08:00
|
|
|
}
|
2016-12-22 16:36:36 +08:00
|
|
|
|
|
|
|
while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) {
|
|
|
|
vma = container_of(node, struct i915_vma, node);
|
|
|
|
ret = i915_vma_unbind(vma);
|
|
|
|
}
|
|
|
|
|
2010-09-30 05:23:05 +08:00
|
|
|
return ret;
|
2010-08-07 18:01:23 +08:00
|
|
|
}
|
|
|
|
|
2016-12-05 22:29:37 +08:00
|
|
|
/**
|
|
|
|
* i915_gem_evict_for_vma - Evict vmas to make room for binding a new one
|
2017-01-11 19:23:11 +08:00
|
|
|
* @vm: address space to evict from
|
|
|
|
* @target: range (and color) to evict for
|
2016-12-05 22:29:37 +08:00
|
|
|
* @flags: additional flags to control the eviction algorithm
|
|
|
|
*
|
|
|
|
* This function will try to evict vmas that overlap the target node.
|
|
|
|
*
|
|
|
|
* To clarify: This is for freeing up virtual address space, not for freeing
|
|
|
|
* memory in e.g. the shrinker.
|
|
|
|
*/
|
2017-01-11 19:23:11 +08:00
|
|
|
int i915_gem_evict_for_node(struct i915_address_space *vm,
|
|
|
|
struct drm_mm_node *target,
|
|
|
|
unsigned int flags)
|
2015-12-08 19:55:07 +08:00
|
|
|
{
|
2016-12-05 22:29:37 +08:00
|
|
|
LIST_HEAD(eviction_list);
|
|
|
|
struct drm_mm_node *node;
|
2017-01-11 19:23:11 +08:00
|
|
|
u64 start = target->start;
|
|
|
|
u64 end = start + target->size;
|
2016-12-05 22:29:37 +08:00
|
|
|
struct i915_vma *vma, *next;
|
|
|
|
bool check_color;
|
|
|
|
int ret = 0;
|
2015-12-08 19:55:07 +08:00
|
|
|
|
2017-01-11 19:23:11 +08:00
|
|
|
lockdep_assert_held(&vm->i915->drm.struct_mutex);
|
2017-02-06 16:45:47 +08:00
|
|
|
GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
|
|
|
|
GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
|
|
|
|
|
2017-01-11 19:23:11 +08:00
|
|
|
trace_i915_gem_evict_node(vm, target, flags);
|
2016-12-05 22:29:37 +08:00
|
|
|
|
2016-12-09 23:05:55 +08:00
|
|
|
/* Retire before we search the active list. Although we have
|
|
|
|
* reasonable accuracy in our retirement lists, we may have
|
|
|
|
* a stray pin (preventing eviction) that can only be resolved by
|
|
|
|
* retiring.
|
|
|
|
*/
|
|
|
|
if (!(flags & PIN_NONBLOCK))
|
2017-01-11 19:23:11 +08:00
|
|
|
i915_gem_retire_requests(vm->i915);
|
2016-12-09 23:05:55 +08:00
|
|
|
|
2017-01-11 19:23:11 +08:00
|
|
|
check_color = vm->mm.color_adjust;
|
2016-12-05 22:29:37 +08:00
|
|
|
if (check_color) {
|
|
|
|
/* Expand search to cover neighbouring guard pages (or lack!) */
|
2017-02-15 16:43:54 +08:00
|
|
|
if (start)
|
2017-01-10 22:47:34 +08:00
|
|
|
start -= I915_GTT_PAGE_SIZE;
|
2017-02-06 16:45:47 +08:00
|
|
|
|
|
|
|
/* Always look at the page afterwards to avoid the end-of-GTT */
|
|
|
|
end += I915_GTT_PAGE_SIZE;
|
2016-12-05 22:29:37 +08:00
|
|
|
}
|
2017-02-06 16:45:47 +08:00
|
|
|
GEM_BUG_ON(start >= end);
|
2016-10-28 20:58:32 +08:00
|
|
|
|
2017-01-11 19:23:11 +08:00
|
|
|
drm_mm_for_each_node_in_range(node, &vm->mm, start, end) {
|
2016-12-05 22:29:37 +08:00
|
|
|
/* If we find any non-objects (!vma), we cannot evict them */
|
|
|
|
if (node->color == I915_COLOR_UNEVICTABLE) {
|
|
|
|
ret = -ENOSPC;
|
2015-12-08 19:55:07 +08:00
|
|
|
break;
|
2016-12-05 22:29:37 +08:00
|
|
|
}
|
2015-12-08 19:55:07 +08:00
|
|
|
|
2017-02-06 16:45:47 +08:00
|
|
|
GEM_BUG_ON(!node->allocated);
|
2015-12-08 19:55:07 +08:00
|
|
|
vma = container_of(node, typeof(*vma), node);
|
|
|
|
|
2016-12-05 22:29:37 +08:00
|
|
|
/* If we are using coloring to insert guard pages between
|
|
|
|
* different cache domains within the address space, we have
|
|
|
|
* to check whether the objects on either side of our range
|
|
|
|
* abutt and conflict. If they are in conflict, then we evict
|
|
|
|
* those as well to make room for our guard pages.
|
|
|
|
*/
|
|
|
|
if (check_color) {
|
2017-03-07 07:54:01 +08:00
|
|
|
if (node->start + node->size == target->start) {
|
|
|
|
if (node->color == target->color)
|
2016-12-05 22:29:37 +08:00
|
|
|
continue;
|
|
|
|
}
|
2017-03-07 07:54:01 +08:00
|
|
|
if (node->start == target->start + target->size) {
|
|
|
|
if (node->color == target->color)
|
2016-12-05 22:29:37 +08:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
2015-12-08 19:55:07 +08:00
|
|
|
|
2016-12-05 22:29:37 +08:00
|
|
|
if (flags & PIN_NONBLOCK &&
|
|
|
|
(i915_vma_is_pinned(vma) || i915_vma_is_active(vma))) {
|
|
|
|
ret = -ENOSPC;
|
|
|
|
break;
|
|
|
|
}
|
2015-12-08 19:55:07 +08:00
|
|
|
|
2017-10-09 16:43:58 +08:00
|
|
|
if (flags & PIN_NONFAULT && i915_vma_has_userfault(vma)) {
|
|
|
|
ret = -ENOSPC;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2016-12-05 22:29:37 +08:00
|
|
|
/* Overlap of objects in the same batch? */
|
2017-06-15 16:14:34 +08:00
|
|
|
if (i915_vma_is_pinned(vma)) {
|
2016-12-05 22:29:37 +08:00
|
|
|
ret = -ENOSPC;
|
2017-08-16 16:52:06 +08:00
|
|
|
if (vma->exec_flags &&
|
|
|
|
*vma->exec_flags & EXEC_OBJECT_PINNED)
|
2016-12-05 22:29:37 +08:00
|
|
|
ret = -EINVAL;
|
|
|
|
break;
|
2015-12-08 19:55:07 +08:00
|
|
|
}
|
|
|
|
|
2016-12-05 22:29:37 +08:00
|
|
|
/* Never show fear in the face of dragons!
|
|
|
|
*
|
|
|
|
* We cannot directly remove this node from within this
|
|
|
|
* iterator and as with i915_gem_evict_something() we employ
|
|
|
|
* the vma pin_count in order to prevent the action of
|
|
|
|
* unbinding one vma from freeing (by dropping its active
|
|
|
|
* reference) another in our eviction list.
|
|
|
|
*/
|
|
|
|
__i915_vma_pin(vma);
|
2017-06-15 16:14:35 +08:00
|
|
|
list_add(&vma->evict_link, &eviction_list);
|
2016-12-05 22:29:37 +08:00
|
|
|
}
|
|
|
|
|
2017-06-15 16:14:35 +08:00
|
|
|
list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
|
2016-12-05 22:29:37 +08:00
|
|
|
__i915_vma_unpin(vma);
|
|
|
|
if (ret == 0)
|
|
|
|
ret = i915_vma_unbind(vma);
|
2015-12-08 19:55:07 +08:00
|
|
|
}
|
|
|
|
|
2016-12-05 22:29:37 +08:00
|
|
|
return ret;
|
2015-12-08 19:55:07 +08:00
|
|
|
}
|
|
|
|
|
2013-09-12 05:57:50 +08:00
|
|
|
/**
|
2014-01-30 05:07:11 +08:00
|
|
|
* i915_gem_evict_vm - Evict all idle vmas from a vm
|
|
|
|
* @vm: Address space to cleanse
|
2013-09-12 05:57:50 +08:00
|
|
|
*
|
drm/i915: Eliminate lots of iterations over the execobjects array
The major scaling bottleneck in execbuffer is the processing of the
execobjects. Creating an auxiliary list is inefficient when compared to
using the execobject array we already have allocated.
Reservation is then split into phases. As we lookup up the VMA, we
try and bind it back into active location. Only if that fails, do we add
it to the unbound list for phase 2. In phase 2, we try and add all those
objects that could not fit into their previous location, with fallback
to retrying all objects and evicting the VM in case of severe
fragmentation. (This is the same as before, except that phase 1 is now
done inline with looking up the VMA to avoid an iteration over the
execobject array. In the ideal case, we eliminate the separate reservation
phase). During the reservation phase, we only evict from the VM between
passes (rather than currently as we try to fit every new VMA). In
testing with Unreal Engine's Atlantis demo which stresses the eviction
logic on gen7 class hardware, this speed up the framerate by a factor of
2.
The second loop amalgamation is between move_to_gpu and move_to_active.
As we always submit the request, even if incomplete, we can use the
current request to track active VMA as we perform the flushes and
synchronisation required.
The next big advancement is to avoid copying back to the user any
execobjects and relocations that are not changed.
v2: Add a Theory of Operation spiel.
v3: Fall back to slow relocations in preparation for flushing userptrs.
v4: Document struct members, factor out eb_validate_vma(), add a few
more comments to explain some magic and hide other magic behind macros.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2017-06-16 22:05:19 +08:00
|
|
|
* This function evicts all vmas from a vm.
|
2013-09-12 05:57:50 +08:00
|
|
|
*
|
2014-01-30 05:07:11 +08:00
|
|
|
* This is used by the execbuf code as a last-ditch effort to defragment the
|
|
|
|
* address space.
|
|
|
|
*
|
|
|
|
* To clarify: This is for freeing up virtual address space, not for freeing
|
|
|
|
* memory in e.g. the shrinker.
|
2013-09-12 05:57:50 +08:00
|
|
|
*/
|
drm/i915: Eliminate lots of iterations over the execobjects array
The major scaling bottleneck in execbuffer is the processing of the
execobjects. Creating an auxiliary list is inefficient when compared to
using the execobject array we already have allocated.
Reservation is then split into phases. As we lookup up the VMA, we
try and bind it back into active location. Only if that fails, do we add
it to the unbound list for phase 2. In phase 2, we try and add all those
objects that could not fit into their previous location, with fallback
to retrying all objects and evicting the VM in case of severe
fragmentation. (This is the same as before, except that phase 1 is now
done inline with looking up the VMA to avoid an iteration over the
execobject array. In the ideal case, we eliminate the separate reservation
phase). During the reservation phase, we only evict from the VM between
passes (rather than currently as we try to fit every new VMA). In
testing with Unreal Engine's Atlantis demo which stresses the eviction
logic on gen7 class hardware, this speed up the framerate by a factor of
2.
The second loop amalgamation is between move_to_gpu and move_to_active.
As we always submit the request, even if incomplete, we can use the
current request to track active VMA as we perform the flushes and
synchronisation required.
The next big advancement is to avoid copying back to the user any
execobjects and relocations that are not changed.
v2: Add a Theory of Operation spiel.
v3: Fall back to slow relocations in preparation for flushing userptrs.
v4: Document struct members, factor out eb_validate_vma(), add a few
more comments to explain some magic and hide other magic behind macros.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2017-06-16 22:05:19 +08:00
|
|
|
int i915_gem_evict_vm(struct i915_address_space *vm)
|
2013-09-12 05:57:49 +08:00
|
|
|
{
|
drm/i915: Eliminate lots of iterations over the execobjects array
The major scaling bottleneck in execbuffer is the processing of the
execobjects. Creating an auxiliary list is inefficient when compared to
using the execobject array we already have allocated.
Reservation is then split into phases. As we lookup up the VMA, we
try and bind it back into active location. Only if that fails, do we add
it to the unbound list for phase 2. In phase 2, we try and add all those
objects that could not fit into their previous location, with fallback
to retrying all objects and evicting the VM in case of severe
fragmentation. (This is the same as before, except that phase 1 is now
done inline with looking up the VMA to avoid an iteration over the
execobject array. In the ideal case, we eliminate the separate reservation
phase). During the reservation phase, we only evict from the VM between
passes (rather than currently as we try to fit every new VMA). In
testing with Unreal Engine's Atlantis demo which stresses the eviction
logic on gen7 class hardware, this speed up the framerate by a factor of
2.
The second loop amalgamation is between move_to_gpu and move_to_active.
As we always submit the request, even if incomplete, we can use the
current request to track active VMA as we perform the flushes and
synchronisation required.
The next big advancement is to avoid copying back to the user any
execobjects and relocations that are not changed.
v2: Add a Theory of Operation spiel.
v3: Fall back to slow relocations in preparation for flushing userptrs.
v4: Document struct members, factor out eb_validate_vma(), add a few
more comments to explain some magic and hide other magic behind macros.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2017-06-16 22:05:19 +08:00
|
|
|
struct list_head *phases[] = {
|
|
|
|
&vm->inactive_list,
|
|
|
|
&vm->active_list,
|
|
|
|
NULL
|
|
|
|
}, **phase;
|
|
|
|
struct list_head eviction_list;
|
2013-09-12 05:57:49 +08:00
|
|
|
struct i915_vma *vma, *next;
|
|
|
|
int ret;
|
|
|
|
|
2016-11-29 17:50:08 +08:00
|
|
|
lockdep_assert_held(&vm->i915->drm.struct_mutex);
|
2013-09-25 00:57:56 +08:00
|
|
|
trace_i915_gem_evict_vm(vm);
|
|
|
|
|
drm/i915: Eliminate lots of iterations over the execobjects array
The major scaling bottleneck in execbuffer is the processing of the
execobjects. Creating an auxiliary list is inefficient when compared to
using the execobject array we already have allocated.
Reservation is then split into phases. As we lookup up the VMA, we
try and bind it back into active location. Only if that fails, do we add
it to the unbound list for phase 2. In phase 2, we try and add all those
objects that could not fit into their previous location, with fallback
to retrying all objects and evicting the VM in case of severe
fragmentation. (This is the same as before, except that phase 1 is now
done inline with looking up the VMA to avoid an iteration over the
execobject array. In the ideal case, we eliminate the separate reservation
phase). During the reservation phase, we only evict from the VM between
passes (rather than currently as we try to fit every new VMA). In
testing with Unreal Engine's Atlantis demo which stresses the eviction
logic on gen7 class hardware, this speed up the framerate by a factor of
2.
The second loop amalgamation is between move_to_gpu and move_to_active.
As we always submit the request, even if incomplete, we can use the
current request to track active VMA as we perform the flushes and
synchronisation required.
The next big advancement is to avoid copying back to the user any
execobjects and relocations that are not changed.
v2: Add a Theory of Operation spiel.
v3: Fall back to slow relocations in preparation for flushing userptrs.
v4: Document struct members, factor out eb_validate_vma(), add a few
more comments to explain some magic and hide other magic behind macros.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2017-06-16 22:05:19 +08:00
|
|
|
/* Switch back to the default context in order to unpin
|
|
|
|
* the existing context objects. However, such objects only
|
|
|
|
* pin themselves inside the global GTT and performing the
|
|
|
|
* switch otherwise is ineffective.
|
|
|
|
*/
|
|
|
|
if (i915_is_ggtt(vm)) {
|
|
|
|
ret = ggtt_flush(vm->i915);
|
2013-09-12 05:57:49 +08:00
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
drm/i915: Eliminate lots of iterations over the execobjects array
The major scaling bottleneck in execbuffer is the processing of the
execobjects. Creating an auxiliary list is inefficient when compared to
using the execobject array we already have allocated.
Reservation is then split into phases. As we lookup up the VMA, we
try and bind it back into active location. Only if that fails, do we add
it to the unbound list for phase 2. In phase 2, we try and add all those
objects that could not fit into their previous location, with fallback
to retrying all objects and evicting the VM in case of severe
fragmentation. (This is the same as before, except that phase 1 is now
done inline with looking up the VMA to avoid an iteration over the
execobject array. In the ideal case, we eliminate the separate reservation
phase). During the reservation phase, we only evict from the VM between
passes (rather than currently as we try to fit every new VMA). In
testing with Unreal Engine's Atlantis demo which stresses the eviction
logic on gen7 class hardware, this speed up the framerate by a factor of
2.
The second loop amalgamation is between move_to_gpu and move_to_active.
As we always submit the request, even if incomplete, we can use the
current request to track active VMA as we perform the flushes and
synchronisation required.
The next big advancement is to avoid copying back to the user any
execobjects and relocations that are not changed.
v2: Add a Theory of Operation spiel.
v3: Fall back to slow relocations in preparation for flushing userptrs.
v4: Document struct members, factor out eb_validate_vma(), add a few
more comments to explain some magic and hide other magic behind macros.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2017-06-16 22:05:19 +08:00
|
|
|
INIT_LIST_HEAD(&eviction_list);
|
|
|
|
phase = phases;
|
|
|
|
do {
|
|
|
|
list_for_each_entry(vma, *phase, vm_link) {
|
|
|
|
if (i915_vma_is_pinned(vma))
|
|
|
|
continue;
|
2013-09-12 05:57:49 +08:00
|
|
|
|
drm/i915: Eliminate lots of iterations over the execobjects array
The major scaling bottleneck in execbuffer is the processing of the
execobjects. Creating an auxiliary list is inefficient when compared to
using the execobject array we already have allocated.
Reservation is then split into phases. As we lookup up the VMA, we
try and bind it back into active location. Only if that fails, do we add
it to the unbound list for phase 2. In phase 2, we try and add all those
objects that could not fit into their previous location, with fallback
to retrying all objects and evicting the VM in case of severe
fragmentation. (This is the same as before, except that phase 1 is now
done inline with looking up the VMA to avoid an iteration over the
execobject array. In the ideal case, we eliminate the separate reservation
phase). During the reservation phase, we only evict from the VM between
passes (rather than currently as we try to fit every new VMA). In
testing with Unreal Engine's Atlantis demo which stresses the eviction
logic on gen7 class hardware, this speed up the framerate by a factor of
2.
The second loop amalgamation is between move_to_gpu and move_to_active.
As we always submit the request, even if incomplete, we can use the
current request to track active VMA as we perform the flushes and
synchronisation required.
The next big advancement is to avoid copying back to the user any
execobjects and relocations that are not changed.
v2: Add a Theory of Operation spiel.
v3: Fall back to slow relocations in preparation for flushing userptrs.
v4: Document struct members, factor out eb_validate_vma(), add a few
more comments to explain some magic and hide other magic behind macros.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2017-06-16 22:05:19 +08:00
|
|
|
__i915_vma_pin(vma);
|
|
|
|
list_add(&vma->evict_link, &eviction_list);
|
|
|
|
}
|
|
|
|
} while (*++phase);
|
|
|
|
|
|
|
|
ret = 0;
|
|
|
|
list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
|
|
|
|
__i915_vma_unpin(vma);
|
|
|
|
if (ret == 0)
|
|
|
|
ret = i915_vma_unbind(vma);
|
|
|
|
}
|
|
|
|
return ret;
|
2013-09-12 05:57:49 +08:00
|
|
|
}
|
2017-02-14 01:15:52 +08:00
|
|
|
|
|
|
|
#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
|
|
|
|
#include "selftests/i915_gem_evict.c"
|
|
|
|
#endif
|