2010-08-07 17:01:23 +07:00
|
|
|
/*
|
|
|
|
* 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>
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include "drmP.h"
|
|
|
|
#include "drm.h"
|
|
|
|
#include "i915_drv.h"
|
|
|
|
#include "i915_drm.h"
|
2011-02-03 18:57:46 +07:00
|
|
|
#include "i915_trace.h"
|
2010-08-07 17:01:23 +07:00
|
|
|
|
2010-08-07 17:01:24 +07:00
|
|
|
static bool
|
2010-11-09 02:18:58 +07:00
|
|
|
mark_free(struct drm_i915_gem_object *obj, struct list_head *unwind)
|
2010-08-07 17:01:23 +07:00
|
|
|
{
|
2012-04-24 21:47:30 +07:00
|
|
|
if (obj->pin_count)
|
|
|
|
return false;
|
|
|
|
|
2010-11-26 02:32:06 +07:00
|
|
|
list_add(&obj->exec_list, unwind);
|
2010-11-09 02:18:58 +07:00
|
|
|
return drm_mm_scan_add_block(obj->gtt_space);
|
2010-08-07 17:01:23 +07:00
|
|
|
}
|
|
|
|
|
|
|
|
int
|
2010-09-16 20:45:15 +07:00
|
|
|
i915_gem_evict_something(struct drm_device *dev, int min_size,
|
2012-07-26 17:49:32 +07:00
|
|
|
unsigned alignment, unsigned cache_level,
|
2012-08-11 21:41:04 +07:00
|
|
|
bool mappable, bool nonblocking)
|
2010-08-07 17:01:23 +07:00
|
|
|
{
|
|
|
|
drm_i915_private_t *dev_priv = dev->dev_private;
|
2010-08-07 17:01:24 +07:00
|
|
|
struct list_head eviction_list, unwind_list;
|
2010-11-09 02:18:58 +07:00
|
|
|
struct drm_i915_gem_object *obj;
|
2010-08-07 17:01:24 +07:00
|
|
|
int ret = 0;
|
2010-08-07 17:01:23 +07:00
|
|
|
|
2011-02-03 18:57:46 +07:00
|
|
|
trace_i915_gem_evict(dev, min_size, alignment, mappable);
|
|
|
|
|
2010-08-07 17:01:24 +07:00
|
|
|
/*
|
|
|
|
* The goal is to evict objects and amalgamate space in LRU order.
|
|
|
|
* The oldest idle objects reside on the inactive list, which is in
|
|
|
|
* retirement order. The next objects to retire are those on the (per
|
|
|
|
* ring) active list that do not have an outstanding flush. Once the
|
|
|
|
* hardware reports completion (the seqno is updated after the
|
|
|
|
* batchbuffer has been finished) the clean buffer objects would
|
|
|
|
* be retired to the inactive list. Any dirty objects would be added
|
|
|
|
* to the tail of the flushing list. So after processing the clean
|
|
|
|
* active objects we need to emit a MI_FLUSH to retire the flushing
|
|
|
|
* list, hence the retirement order of the flushing list is in
|
|
|
|
* advance of the dirty objects on the active lists.
|
|
|
|
*
|
|
|
|
* The retirement sequence is thus:
|
|
|
|
* 1. Inactive objects (already retired)
|
|
|
|
* 2. Clean active objects
|
|
|
|
* 3. Flushing list
|
|
|
|
* 4. Dirty active objects.
|
|
|
|
*
|
|
|
|
* On each list, the oldest objects lie at the HEAD with the freshest
|
|
|
|
* object on the TAIL.
|
|
|
|
*/
|
|
|
|
|
|
|
|
INIT_LIST_HEAD(&unwind_list);
|
2010-09-16 20:45:15 +07:00
|
|
|
if (mappable)
|
2012-07-10 17:15:23 +07:00
|
|
|
drm_mm_init_scan_with_range(&dev_priv->mm.gtt_space,
|
2012-07-26 17:49:32 +07:00
|
|
|
min_size, alignment, cache_level,
|
2012-07-10 17:15:23 +07:00
|
|
|
0, dev_priv->mm.gtt_mappable_end);
|
2010-09-16 20:45:15 +07:00
|
|
|
else
|
2012-07-10 17:15:23 +07:00
|
|
|
drm_mm_init_scan(&dev_priv->mm.gtt_space,
|
2012-07-26 17:49:32 +07:00
|
|
|
min_size, alignment, cache_level);
|
2010-08-07 17:01:24 +07:00
|
|
|
|
|
|
|
/* First see if there is a large enough contiguous idle region... */
|
2010-11-09 02:18:58 +07:00
|
|
|
list_for_each_entry(obj, &dev_priv->mm.inactive_list, mm_list) {
|
|
|
|
if (mark_free(obj, &unwind_list))
|
2010-08-07 17:01:24 +07:00
|
|
|
goto found;
|
|
|
|
}
|
2010-08-07 17:01:23 +07:00
|
|
|
|
2012-08-11 21:41:04 +07:00
|
|
|
if (nonblocking)
|
|
|
|
goto none;
|
|
|
|
|
2010-08-07 17:01:24 +07:00
|
|
|
/* Now merge in the soon-to-be-expired objects... */
|
2010-11-09 02:18:58 +07:00
|
|
|
list_for_each_entry(obj, &dev_priv->mm.active_list, mm_list) {
|
|
|
|
if (mark_free(obj, &unwind_list))
|
2010-08-07 17:01:24 +07:00
|
|
|
goto found;
|
|
|
|
}
|
|
|
|
|
2012-08-11 21:41:04 +07:00
|
|
|
none:
|
2010-08-07 17:01:24 +07:00
|
|
|
/* Nothing found, clean up and bail out! */
|
2011-01-10 21:21:05 +07:00
|
|
|
while (!list_empty(&unwind_list)) {
|
|
|
|
obj = list_first_entry(&unwind_list,
|
|
|
|
struct drm_i915_gem_object,
|
|
|
|
exec_list);
|
|
|
|
|
2010-11-09 02:18:58 +07:00
|
|
|
ret = drm_mm_scan_remove_block(obj->gtt_space);
|
2010-08-07 17:01:24 +07:00
|
|
|
BUG_ON(ret);
|
2011-01-10 21:21:05 +07:00
|
|
|
|
|
|
|
list_del_init(&obj->exec_list);
|
2010-08-07 17:01:24 +07:00
|
|
|
}
|
|
|
|
|
|
|
|
/* We expect the caller to unpin, evict all and try again, or give up.
|
|
|
|
* So calling i915_gem_evict_everything() is unnecessary.
|
|
|
|
*/
|
|
|
|
return -ENOSPC;
|
|
|
|
|
|
|
|
found:
|
2010-09-30 04:23:05 +07:00
|
|
|
/* drm_mm doesn't allow any other other operations while
|
|
|
|
* scanning, therefore store to be evicted objects on a
|
|
|
|
* temporary list. */
|
2010-08-07 17:01:24 +07:00
|
|
|
INIT_LIST_HEAD(&eviction_list);
|
2010-09-30 04:23:05 +07:00
|
|
|
while (!list_empty(&unwind_list)) {
|
2010-11-09 02:18:58 +07:00
|
|
|
obj = list_first_entry(&unwind_list,
|
|
|
|
struct drm_i915_gem_object,
|
2010-11-26 02:32:06 +07:00
|
|
|
exec_list);
|
2010-11-09 02:18:58 +07:00
|
|
|
if (drm_mm_scan_remove_block(obj->gtt_space)) {
|
2010-11-26 02:32:06 +07:00
|
|
|
list_move(&obj->exec_list, &eviction_list);
|
2012-02-24 07:27:21 +07:00
|
|
|
drm_gem_object_reference(&obj->base);
|
2010-09-30 04:23:05 +07:00
|
|
|
continue;
|
|
|
|
}
|
2010-11-26 02:32:06 +07:00
|
|
|
list_del_init(&obj->exec_list);
|
2010-08-07 17:01:24 +07:00
|
|
|
}
|
2010-08-07 17:01:23 +07:00
|
|
|
|
2010-08-07 17:01:24 +07:00
|
|
|
/* Unbinding will emit any required flushes */
|
2010-09-30 04:23:05 +07:00
|
|
|
while (!list_empty(&eviction_list)) {
|
2010-11-09 02:18:58 +07:00
|
|
|
obj = list_first_entry(&eviction_list,
|
|
|
|
struct drm_i915_gem_object,
|
2010-11-26 02:32:06 +07:00
|
|
|
exec_list);
|
2010-09-30 04:23:05 +07:00
|
|
|
if (ret == 0)
|
2010-11-09 02:18:58 +07:00
|
|
|
ret = i915_gem_object_unbind(obj);
|
2011-01-10 21:21:05 +07:00
|
|
|
|
2010-11-26 02:32:06 +07:00
|
|
|
list_del_init(&obj->exec_list);
|
2010-11-09 02:18:58 +07:00
|
|
|
drm_gem_object_unreference(&obj->base);
|
2010-08-07 17:01:23 +07:00
|
|
|
}
|
2010-08-07 17:01:24 +07:00
|
|
|
|
2010-09-30 04:23:05 +07:00
|
|
|
return ret;
|
2010-08-07 17:01:23 +07:00
|
|
|
}
|
|
|
|
|
|
|
|
int
|
drm/i915: Track unbound pages
When dealing with a working set larger than the GATT, or even the
mappable aperture when touching through the GTT, we end up with evicting
objects only to rebind them at a new offset again later. Moving an
object into and out of the GTT requires clflushing the pages, thus
causing a double-clflush penalty for rebinding.
To avoid having to clflush on rebinding, we can track the pages as they
are evicted from the GTT and only relinquish those pages on memory
pressure.
As usual, if it were not for the handling of out-of-memory condition and
having to manually shrink our own bo caches, it would be a net reduction
of code. Alas.
Note: The patch also contains a few changes to the last-hope
evict_everything logic in i916_gem_execbuffer.c - we no longer try to
only evict the purgeable stuff in a first try (since that's superflous
and only helps in OOM corner-cases, not fragmented-gtt trashing
situations).
Also, the extraction of the get_pages retry loop from bind_to_gtt (and
other callsites) to get_pages should imo have been a separate patch.
v2: Ditch the newly added put_pages (for unbound objects only) in
i915_gem_reset. A quick irc discussion hasn't revealed any important
reason for this, so if we need this, I'd like to have a git blame'able
explanation for it.
v3: Undo the s/drm_malloc_ab/kmalloc/ in get_pages that Chris noticed.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Split out code movements and rant a bit in the commit message
with a few Notes. Done v2]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-08-20 16:40:46 +07:00
|
|
|
i915_gem_evict_everything(struct drm_device *dev)
|
2010-08-07 17:01:23 +07:00
|
|
|
{
|
|
|
|
drm_i915_private_t *dev_priv = dev->dev_private;
|
2012-04-25 00:22:52 +07:00
|
|
|
struct drm_i915_gem_object *obj, *next;
|
2010-08-07 17:01:23 +07:00
|
|
|
bool lists_empty;
|
2012-05-11 20:29:30 +07:00
|
|
|
int ret;
|
2010-08-07 17:01:23 +07:00
|
|
|
|
|
|
|
lists_empty = (list_empty(&dev_priv->mm.inactive_list) &&
|
2010-10-29 03:28:46 +07:00
|
|
|
list_empty(&dev_priv->mm.active_list));
|
2010-08-07 17:01:23 +07:00
|
|
|
if (lists_empty)
|
|
|
|
return -ENOSPC;
|
|
|
|
|
drm/i915: Track unbound pages
When dealing with a working set larger than the GATT, or even the
mappable aperture when touching through the GTT, we end up with evicting
objects only to rebind them at a new offset again later. Moving an
object into and out of the GTT requires clflushing the pages, thus
causing a double-clflush penalty for rebinding.
To avoid having to clflush on rebinding, we can track the pages as they
are evicted from the GTT and only relinquish those pages on memory
pressure.
As usual, if it were not for the handling of out-of-memory condition and
having to manually shrink our own bo caches, it would be a net reduction
of code. Alas.
Note: The patch also contains a few changes to the last-hope
evict_everything logic in i916_gem_execbuffer.c - we no longer try to
only evict the purgeable stuff in a first try (since that's superflous
and only helps in OOM corner-cases, not fragmented-gtt trashing
situations).
Also, the extraction of the get_pages retry loop from bind_to_gtt (and
other callsites) to get_pages should imo have been a separate patch.
v2: Ditch the newly added put_pages (for unbound objects only) in
i915_gem_reset. A quick irc discussion hasn't revealed any important
reason for this, so if we need this, I'd like to have a git blame'able
explanation for it.
v3: Undo the s/drm_malloc_ab/kmalloc/ in get_pages that Chris noticed.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Split out code movements and rant a bit in the commit message
with a few Notes. Done v2]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-08-20 16:40:46 +07:00
|
|
|
trace_i915_gem_evict_everything(dev);
|
2011-02-03 18:57:46 +07:00
|
|
|
|
2012-04-27 06:02:58 +07:00
|
|
|
/* The gpu_idle will flush everything in the write domain to the
|
|
|
|
* active list. Then we must move everything off the active list
|
|
|
|
* with retire requests.
|
|
|
|
*/
|
2012-05-11 20:29:30 +07:00
|
|
|
ret = i915_gpu_idle(dev);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
2012-04-27 06:02:58 +07:00
|
|
|
|
|
|
|
i915_gem_retire_requests(dev);
|
|
|
|
|
2012-04-25 00:22:52 +07:00
|
|
|
/* Having flushed everything, unbind() should never raise an error */
|
2010-10-31 15:49:47 +07:00
|
|
|
list_for_each_entry_safe(obj, next,
|
drm/i915: Track unbound pages
When dealing with a working set larger than the GATT, or even the
mappable aperture when touching through the GTT, we end up with evicting
objects only to rebind them at a new offset again later. Moving an
object into and out of the GTT requires clflushing the pages, thus
causing a double-clflush penalty for rebinding.
To avoid having to clflush on rebinding, we can track the pages as they
are evicted from the GTT and only relinquish those pages on memory
pressure.
As usual, if it were not for the handling of out-of-memory condition and
having to manually shrink our own bo caches, it would be a net reduction
of code. Alas.
Note: The patch also contains a few changes to the last-hope
evict_everything logic in i916_gem_execbuffer.c - we no longer try to
only evict the purgeable stuff in a first try (since that's superflous
and only helps in OOM corner-cases, not fragmented-gtt trashing
situations).
Also, the extraction of the get_pages retry loop from bind_to_gtt (and
other callsites) to get_pages should imo have been a separate patch.
v2: Ditch the newly added put_pages (for unbound objects only) in
i915_gem_reset. A quick irc discussion hasn't revealed any important
reason for this, so if we need this, I'd like to have a git blame'able
explanation for it.
v3: Undo the s/drm_malloc_ab/kmalloc/ in get_pages that Chris noticed.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Split out code movements and rant a bit in the commit message
with a few Notes. Done v2]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-08-20 16:40:46 +07:00
|
|
|
&dev_priv->mm.inactive_list, mm_list)
|
|
|
|
if (obj->pin_count == 0)
|
|
|
|
WARN_ON(i915_gem_object_unbind(obj));
|
2010-08-07 17:01:23 +07:00
|
|
|
|
2012-05-11 20:29:30 +07:00
|
|
|
return 0;
|
2010-08-07 17:01:23 +07:00
|
|
|
}
|