Commit 9332f3b1 authored by Chris Wilson's avatar Chris Wilson

drm/i915: Combine loops within i915_gem_evict_something

Slight micro-optimise to produce combine loops so that gcc is able to
optimise the inner-loops concisely. Since we are reviewing the loops, we
can update the comments to describe the current state of affairs, in
particular the distinction between evicting from the global GTT (which
may contain untracked items and transient global pins) and the
per-process GTT.
Signed-off-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: default avatarJoonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1470324762-2545-1-git-send-email-chris@chris-wilson.co.uk
parent d838a110
...@@ -33,6 +33,19 @@ ...@@ -33,6 +33,19 @@
#include "intel_drv.h" #include "intel_drv.h"
#include "i915_trace.h" #include "i915_trace.h"
static bool
gpu_is_idle(struct drm_i915_private *dev_priv)
{
struct intel_engine_cs *engine;
for_each_engine(engine, dev_priv) {
if (!list_empty(&engine->request_list))
return false;
}
return true;
}
static bool static bool
mark_free(struct i915_vma *vma, struct list_head *unwind) mark_free(struct i915_vma *vma, struct list_head *unwind)
{ {
...@@ -76,37 +89,31 @@ i915_gem_evict_something(struct drm_device *dev, struct i915_address_space *vm, ...@@ -76,37 +89,31 @@ i915_gem_evict_something(struct drm_device *dev, struct i915_address_space *vm,
unsigned long start, unsigned long end, unsigned long start, unsigned long end,
unsigned flags) unsigned flags)
{ {
struct list_head eviction_list, unwind_list; struct drm_i915_private *dev_priv = to_i915(dev);
struct i915_vma *vma; struct list_head eviction_list;
int ret = 0; struct list_head *phases[] = {
int pass = 0; &vm->inactive_list,
&vm->active_list,
NULL,
}, **phase;
struct i915_vma *vma, *next;
int ret;
trace_i915_gem_evict(dev, min_size, alignment, flags); trace_i915_gem_evict(dev, min_size, alignment, flags);
/* /*
* The goal is to evict objects and amalgamate space in LRU order. * The goal is to evict objects and amalgamate space in LRU order.
* The oldest idle objects reside on the inactive list, which is in * The oldest idle objects reside on the inactive list, which is in
* retirement order. The next objects to retire are those on the (per * retirement order. The next objects to retire are those in flight,
* ring) active list that do not have an outstanding flush. Once the * on the active list, again in retirement order.
* 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: * The retirement sequence is thus:
* 1. Inactive objects (already retired) * 1. Inactive objects (already retired)
* 2. Clean active objects * 2. Active objects (will stall on unbinding)
* 3. Flushing list
* 4. Dirty active objects.
* *
* On each list, the oldest objects lie at the HEAD with the freshest * On each list, the oldest objects lie at the HEAD with the freshest
* object on the TAIL. * object on the TAIL.
*/ */
INIT_LIST_HEAD(&unwind_list);
if (start != 0 || end != vm->total) { if (start != 0 || end != vm->total) {
drm_mm_init_scan_with_range(&vm->mm, min_size, drm_mm_init_scan_with_range(&vm->mm, min_size,
alignment, cache_level, alignment, cache_level,
...@@ -114,49 +121,51 @@ i915_gem_evict_something(struct drm_device *dev, struct i915_address_space *vm, ...@@ -114,49 +121,51 @@ i915_gem_evict_something(struct drm_device *dev, struct i915_address_space *vm,
} else } else
drm_mm_init_scan(&vm->mm, min_size, alignment, cache_level); drm_mm_init_scan(&vm->mm, min_size, alignment, cache_level);
search_again:
/* First see if there is a large enough contiguous idle region... */
list_for_each_entry(vma, &vm->inactive_list, vm_link) {
if (mark_free(vma, &unwind_list))
goto found;
}
if (flags & PIN_NONBLOCK) if (flags & PIN_NONBLOCK)
goto none; phases[1] = NULL;
/* Now merge in the soon-to-be-expired objects... */ search_again:
list_for_each_entry(vma, &vm->active_list, vm_link) { INIT_LIST_HEAD(&eviction_list);
if (mark_free(vma, &unwind_list)) phase = phases;
do {
list_for_each_entry(vma, *phase, vm_link)
if (mark_free(vma, &eviction_list))
goto found; goto found;
} } while (*++phase);
none:
/* Nothing found, clean up and bail out! */ /* Nothing found, clean up and bail out! */
while (!list_empty(&unwind_list)) { list_for_each_entry_safe(vma, next, &eviction_list, exec_list) {
vma = list_first_entry(&unwind_list,
struct i915_vma,
exec_list);
ret = drm_mm_scan_remove_block(&vma->node); ret = drm_mm_scan_remove_block(&vma->node);
BUG_ON(ret); BUG_ON(ret);
list_del_init(&vma->exec_list); INIT_LIST_HEAD(&vma->exec_list);
} }
/* Can we unpin some objects such as idle hw contents, /* Can we unpin some objects such as idle hw contents,
* or pending flips? * 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.
*/ */
if (flags & PIN_NONBLOCK) if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK)
return -ENOSPC; return -ENOSPC;
/* Only idle the GPU and repeat the search once */ if (gpu_is_idle(dev_priv)) {
if (pass++ == 0) { /* If we still have pending pageflip completions, drop
struct drm_i915_private *dev_priv = to_i915(dev); * back to userspace to give our workqueues time to
* acquire our locks and unpin the old scanouts.
*/
return intel_has_pending_fb_unpin(dev) ? -EAGAIN : -ENOSPC;
}
if (i915_is_ggtt(vm)) { /* 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.
*/
ret = i915_gem_switch_to_kernel_context(dev_priv); ret = i915_gem_switch_to_kernel_context(dev_priv);
if (ret) if (ret)
return ret; return ret;
}
ret = i915_gem_wait_for_idle(dev_priv); ret = i915_gem_wait_for_idle(dev_priv);
if (ret) if (ret)
...@@ -164,28 +173,18 @@ i915_gem_evict_something(struct drm_device *dev, struct i915_address_space *vm, ...@@ -164,28 +173,18 @@ i915_gem_evict_something(struct drm_device *dev, struct i915_address_space *vm,
i915_gem_retire_requests(dev_priv); i915_gem_retire_requests(dev_priv);
goto search_again; goto search_again;
}
/* 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) ? -EAGAIN : -ENOSPC;
found: found:
/* drm_mm doesn't allow any other other operations while /* drm_mm doesn't allow any other other operations while
* scanning, therefore store to be evicted objects on a * scanning, therefore store to-be-evicted objects on a
* temporary list. */ * temporary list and take a reference for all before
INIT_LIST_HEAD(&eviction_list); * calling unbind (which may remove the active reference
while (!list_empty(&unwind_list)) { * of any of our objects, thus corrupting the list).
vma = list_first_entry(&unwind_list, */
struct i915_vma, list_for_each_entry_safe(vma, next, &eviction_list, exec_list) {
exec_list); if (drm_mm_scan_remove_block(&vma->node))
if (drm_mm_scan_remove_block(&vma->node)) {
vma->pin_count++; vma->pin_count++;
list_move(&vma->exec_list, &eviction_list); else
continue;
}
list_del_init(&vma->exec_list); list_del_init(&vma->exec_list);
} }
...@@ -200,7 +199,6 @@ i915_gem_evict_something(struct drm_device *dev, struct i915_address_space *vm, ...@@ -200,7 +199,6 @@ i915_gem_evict_something(struct drm_device *dev, struct i915_address_space *vm,
if (ret == 0) if (ret == 0)
ret = i915_vma_unbind(vma); ret = i915_vma_unbind(vma);
} }
return ret; return ret;
} }
...@@ -279,7 +277,6 @@ int i915_gem_evict_vm(struct i915_address_space *vm, bool do_idle) ...@@ -279,7 +277,6 @@ int i915_gem_evict_vm(struct i915_address_space *vm, bool do_idle)
return ret; return ret;
i915_gem_retire_requests(dev_priv); i915_gem_retire_requests(dev_priv);
WARN_ON(!list_empty(&vm->active_list)); WARN_ON(!list_empty(&vm->active_list));
} }
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment