const struct i915_vma *vma,
unsigned int flags)
{
- if (vma->node.size < entry->pad_to_size)
+ const u64 start = i915_vma_offset(vma);
+ const u64 size = i915_vma_size(vma);
+
+ if (size < entry->pad_to_size)
return true;
- if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment))
+ if (entry->alignment && !IS_ALIGNED(start, entry->alignment))
return true;
if (flags & EXEC_OBJECT_PINNED &&
- vma->node.start != entry->offset)
+ start != entry->offset)
return true;
if (flags & __EXEC_OBJECT_NEEDS_BIAS &&
- vma->node.start < BATCH_OFFSET_BIAS)
+ start < BATCH_OFFSET_BIAS)
return true;
if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) &&
- (vma->node.start + vma->node.size + 4095) >> 32)
+ (start + size + 4095) >> 32)
return true;
if (flags & __EXEC_OBJECT_NEEDS_MAP &&
int err;
if (vma->node.size)
- pin_flags = vma->node.start;
+ pin_flags = __i915_vma_offset(vma);
else
pin_flags = entry->offset & PIN_OFFSET_MASK;
if (err)
return err;
- if (entry->offset != vma->node.start) {
- entry->offset = vma->node.start | UPDATE;
+ if (entry->offset != i915_vma_offset(vma)) {
+ entry->offset = i915_vma_offset(vma) | UPDATE;
eb->args->flags |= __EXEC_HAS_RELOC;
}
bool unpinned;
/*
- * Attempt to pin all of the buffers into the GTT.
- * This is done in 2 phases:
+ * We have one more buffers that we couldn't bind, which could be due to
+ * various reasons. To resolve this we have 4 passes, with every next
+ * level turning the screws tighter:
+ *
+ * 0. Unbind all objects that do not match the GTT constraints for the
+ * execbuffer (fenceable, mappable, alignment etc). Bind all new
+ * objects. This avoids unnecessary unbinding of later objects in order
+ * to make room for the earlier objects *unless* we need to defragment.
*
- * 1. Unbind all objects that do not match the GTT constraints for
- * the execbuffer (fenceable, mappable, alignment etc).
- * 2. Bind new objects.
+ * 1. Reorder the buffers, where objects with the most restrictive
+ * placement requirements go first (ignoring fixed location buffers for
+ * now). For example, objects needing the mappable aperture (the first
+ * 256M of GTT), should go first vs objects that can be placed just
+ * about anywhere. Repeat the previous pass.
*
- * This avoid unnecessary unbinding of later objects in order to make
- * room for the earlier objects *unless* we need to defragment.
+ * 2. Consider buffers that are pinned at a fixed location. Also try to
+ * evict the entire VM this time, leaving only objects that we were
+ * unable to lock. Try again to bind the buffers. (still using the new
+ * buffer order).
*
- * Defragmenting is skipped if all objects are pinned at a fixed location.
+ * 3. We likely have object lock contention for one or more stubborn
+ * objects in the VM, for which we need to evict to make forward
+ * progress (perhaps we are fighting the shrinker?). When evicting the
+ * VM this time around, anything that we can't lock we now track using
+ * the busy_bo, using the full lock (after dropping the vm->mutex to
+ * prevent deadlocks), instead of trylock. We then continue to evict the
+ * VM, this time with the stubborn object locked, which we can now
+ * hopefully unbind (if still bound in the VM). Repeat until the VM is
+ * evicted. Finally we should be able bind everything.
*/
- for (pass = 0; pass <= 2; pass++) {
+ for (pass = 0; pass <= 3; pass++) {
int pin_flags = PIN_USER | PIN_VALIDATE;
if (pass == 0)
pin_flags |= PIN_NONBLOCK;
if (pass >= 1)
- unpinned = eb_unbind(eb, pass == 2);
+ unpinned = eb_unbind(eb, pass >= 2);
if (pass == 2) {
err = mutex_lock_interruptible(&eb->context->vm->mutex);
if (!err) {
- err = i915_gem_evict_vm(eb->context->vm, &eb->ww);
+ err = i915_gem_evict_vm(eb->context->vm, &eb->ww, NULL);
mutex_unlock(&eb->context->vm->mutex);
}
if (err)
return err;
}
+ if (pass == 3) {
+retry:
+ err = mutex_lock_interruptible(&eb->context->vm->mutex);
+ if (!err) {
+ struct drm_i915_gem_object *busy_bo = NULL;
+
+ err = i915_gem_evict_vm(eb->context->vm, &eb->ww, &busy_bo);
+ mutex_unlock(&eb->context->vm->mutex);
+ if (err && busy_bo) {
+ err = i915_gem_object_lock(busy_bo, &eb->ww);
+ i915_gem_object_put(busy_bo);
+ if (!err)
+ goto retry;
+ }
+ }
+ if (err)
+ return err;
+ }
+
list_for_each_entry(ev, &eb->unbound, bind_link) {
err = eb_reserve_vma(eb, ev, pin_flags);
if (err)
*/
if (i915_gem_context_uses_protected_content(eb->gem_context) &&
i915_gem_object_is_protected(obj)) {
- err = intel_pxp_key_check(&vm->gt->pxp, obj, true);
+ err = intel_pxp_key_check(eb->i915->pxp, obj, true);
if (err) {
i915_gem_object_put(obj);
return ERR_PTR(err);
return err;
if (!err) {
- if (entry->offset != vma->node.start) {
- entry->offset = vma->node.start | UPDATE;
+ if (entry->offset != i915_vma_offset(vma)) {
+ entry->offset = i915_vma_offset(vma) | UPDATE;
eb->args->flags |= __EXEC_HAS_RELOC;
}
} else {
relocation_target(const struct drm_i915_gem_relocation_entry *reloc,
const struct i915_vma *target)
{
- return gen8_canonical_addr((int)reloc->delta + target->node.start);
+ return gen8_canonical_addr((int)reloc->delta + i915_vma_offset(target));
}
static void reloc_cache_init(struct reloc_cache *cache,
if (err) /* no inactive aperture space, use cpu reloc */
return NULL;
} else {
- cache->node.start = vma->node.start;
+ cache->node.start = i915_ggtt_offset(vma);
cache->node.mm = (void *)vma;
}
}
* more work needs to be done.
*/
if (!DBG_FORCE_RELOC &&
- gen8_canonical_addr(target->vma->node.start) == reloc->presumed_offset)
+ gen8_canonical_addr(i915_vma_offset(target->vma)) == reloc->presumed_offset)
return 0;
/* Check that the relocation address is valid... */
}
err = rq->context->engine->emit_bb_start(rq,
- batch->node.start +
+ i915_vma_offset(batch) +
eb->batch_start_offset,
batch_len,
eb->batch_flags);
GEM_BUG_ON(intel_context_is_parallel(rq->context));
GEM_BUG_ON(eb->batch_start_offset);
err = rq->context->engine->emit_bb_start(rq,
- eb->trampoline->node.start +
+ i915_vma_offset(eb->trampoline) +
batch_len, 0, 0);
if (err)
return err;