[PATCH] drm/i915: improve the catch-all evict to handle lock contention
Matthew Auld
matthew.auld at intel.com
Wed Nov 23 18:58:03 UTC 2022
The catch-all evict can fail due to object lock contention, since it
only goes as far as trylocking the object, due to us already holding the
vm->mutex. Doing a full object lock can deadlock the system, since the
vm->mutex is always our inner lock. Add another execbuf pass which drops
the vm->mutex and then tries to grab the object will the full lock,
before then retrying the eviction.
Testcase: igt at igem_ppgtt@shrink-vs-evict-*
References: https://gitlab.freedesktop.org/drm/intel/-/issues/7570
Signed-off-by: Matthew Auld <matthew.auld at intel.com>
---
.../gpu/drm/i915/gem/i915_gem_execbuffer.c | 24 +++++++++++--
drivers/gpu/drm/i915/gem/i915_gem_mman.c | 2 +-
drivers/gpu/drm/i915/i915_gem_evict.c | 36 +++++++++++++------
drivers/gpu/drm/i915/i915_gem_evict.h | 4 ++-
drivers/gpu/drm/i915/i915_vma.c | 2 +-
.../gpu/drm/i915/selftests/i915_gem_evict.c | 4 +--
6 files changed, 55 insertions(+), 17 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
index 29e9e8d5b6fe..a9b5fab79806 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
@@ -742,7 +742,7 @@ static int eb_reserve(struct i915_execbuffer *eb)
*
* Defragmenting is skipped if all objects are pinned at a fixed location.
*/
- for (pass = 0; pass <= 2; pass++) {
+ for (pass = 0; pass <= 3; pass++) {
int pin_flags = PIN_USER | PIN_VALIDATE;
if (pass == 0)
@@ -754,13 +754,33 @@ static int eb_reserve(struct i915_execbuffer *eb)
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) {
+ do {
+ 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) {
+ if (busy_bo) {
+ err = i915_gem_object_lock(busy_bo, &eb->ww);
+ i915_gem_object_put(busy_bo);
+ }
+ }
+ }
+ if (err)
+ return err;
+ } while (1);
+ }
+
list_for_each_entry(ev, &eb->unbound, bind_link) {
err = eb_reserve_vma(eb, ev, pin_flags);
if (err)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_mman.c b/drivers/gpu/drm/i915/gem/i915_gem_mman.c
index c29efdef8313..0ad44f3868de 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_mman.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_mman.c
@@ -369,7 +369,7 @@ static vm_fault_t vm_fault_gtt(struct vm_fault *vmf)
if (vma == ERR_PTR(-ENOSPC)) {
ret = mutex_lock_interruptible(&ggtt->vm.mutex);
if (!ret) {
- ret = i915_gem_evict_vm(&ggtt->vm, &ww);
+ ret = i915_gem_evict_vm(&ggtt->vm, &ww, NULL);
mutex_unlock(&ggtt->vm.mutex);
}
if (ret)
diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c
index f025ee4fa526..ba50dba1e213 100644
--- a/drivers/gpu/drm/i915/i915_gem_evict.c
+++ b/drivers/gpu/drm/i915/i915_gem_evict.c
@@ -416,6 +416,11 @@ int i915_gem_evict_for_node(struct i915_address_space *vm,
* @vm: Address space to cleanse
* @ww: An optional struct i915_gem_ww_ctx. If not NULL, i915_gem_evict_vm
* will be able to evict vma's locked by the ww as well.
+ * @busy_bo: Optional pointer to struct drm_i915_gem_object. If not NULL, then
+ * in the event i915_gem_evict_vm() is unable to trylock an object for eviction,
+ * then @busy_bo will point to it. -EBUSY is also returned. The caller must drop
+ * the vm->mutex, before trying again to acquire the contended lock. The caller
+ * also owns a reference to the object.
*
* This function evicts all vmas from a vm.
*
@@ -425,7 +430,8 @@ int i915_gem_evict_for_node(struct i915_address_space *vm,
* To clarify: This is for freeing up virtual address space, not for freeing
* memory in e.g. the shrinker.
*/
-int i915_gem_evict_vm(struct i915_address_space *vm, struct i915_gem_ww_ctx *ww)
+int i915_gem_evict_vm(struct i915_address_space *vm, struct i915_gem_ww_ctx *ww,
+ struct drm_i915_gem_object **busy_bo)
{
int ret = 0;
@@ -457,15 +463,21 @@ int i915_gem_evict_vm(struct i915_address_space *vm, struct i915_gem_ww_ctx *ww)
* the resv is shared among multiple objects, we still
* need the object ref.
*/
- if (dying_vma(vma) ||
+ if (!i915_gem_object_get_rcu(vma->obj) ||
(ww && (dma_resv_locking_ctx(vma->obj->base.resv) == &ww->ctx))) {
__i915_vma_pin(vma);
list_add(&vma->evict_link, &locked_eviction_list);
continue;
}
- if (!i915_gem_object_trylock(vma->obj, ww))
+ if (!i915_gem_object_trylock(vma->obj, ww)) {
+ if (busy_bo) {
+ *busy_bo = vma->obj; /* holds ref */
+ ret = -EBUSY;
+ break;
+ }
continue;
+ }
__i915_vma_pin(vma);
list_add(&vma->evict_link, &eviction_list);
@@ -473,25 +485,29 @@ int i915_gem_evict_vm(struct i915_address_space *vm, struct i915_gem_ww_ctx *ww)
if (list_empty(&eviction_list) && list_empty(&locked_eviction_list))
break;
- ret = 0;
/* Unbind locked objects first, before unlocking the eviction_list */
list_for_each_entry_safe(vma, vn, &locked_eviction_list, evict_link) {
__i915_vma_unpin(vma);
- if (ret == 0)
+ if (ret == 0) {
ret = __i915_vma_unbind(vma);
- if (ret != -EINTR) /* "Get me out of here!" */
- ret = 0;
+ if (ret != -EINTR) /* "Get me out of here!" */
+ ret = 0;
+ }
+ if (!dying_vma(vma))
+ i915_gem_object_put(vma->obj);
}
list_for_each_entry_safe(vma, vn, &eviction_list, evict_link) {
__i915_vma_unpin(vma);
- if (ret == 0)
+ if (ret == 0) {
ret = __i915_vma_unbind(vma);
- if (ret != -EINTR) /* "Get me out of here!" */
- ret = 0;
+ if (ret != -EINTR) /* "Get me out of here!" */
+ ret = 0;
+ }
i915_gem_object_unlock(vma->obj);
+ i915_gem_object_put(vma->obj);
}
} while (ret == 0);
diff --git a/drivers/gpu/drm/i915/i915_gem_evict.h b/drivers/gpu/drm/i915/i915_gem_evict.h
index e593c530f9bd..bf0ee0e4fe60 100644
--- a/drivers/gpu/drm/i915/i915_gem_evict.h
+++ b/drivers/gpu/drm/i915/i915_gem_evict.h
@@ -11,6 +11,7 @@
struct drm_mm_node;
struct i915_address_space;
struct i915_gem_ww_ctx;
+struct drm_i915_gem_object;
int __must_check i915_gem_evict_something(struct i915_address_space *vm,
struct i915_gem_ww_ctx *ww,
@@ -23,6 +24,7 @@ int __must_check i915_gem_evict_for_node(struct i915_address_space *vm,
struct drm_mm_node *node,
unsigned int flags);
int i915_gem_evict_vm(struct i915_address_space *vm,
- struct i915_gem_ww_ctx *ww);
+ struct i915_gem_ww_ctx *ww,
+ struct drm_i915_gem_object **busy_bo);
#endif /* __I915_GEM_EVICT_H__ */
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
index 703fee6b5f75..3a33be5401ed 100644
--- a/drivers/gpu/drm/i915/i915_vma.c
+++ b/drivers/gpu/drm/i915/i915_vma.c
@@ -1566,7 +1566,7 @@ static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
* locked objects when called from execbuf when pinning
* is removed. This would probably regress badly.
*/
- i915_gem_evict_vm(vm, NULL);
+ i915_gem_evict_vm(vm, NULL, NULL);
mutex_unlock(&vm->mutex);
}
} while (1);
diff --git a/drivers/gpu/drm/i915/selftests/i915_gem_evict.c b/drivers/gpu/drm/i915/selftests/i915_gem_evict.c
index 8c6517d29b8e..37068542aafe 100644
--- a/drivers/gpu/drm/i915/selftests/i915_gem_evict.c
+++ b/drivers/gpu/drm/i915/selftests/i915_gem_evict.c
@@ -344,7 +344,7 @@ static int igt_evict_vm(void *arg)
/* Everything is pinned, nothing should happen */
mutex_lock(&ggtt->vm.mutex);
- err = i915_gem_evict_vm(&ggtt->vm, NULL);
+ err = i915_gem_evict_vm(&ggtt->vm, NULL, NULL);
mutex_unlock(&ggtt->vm.mutex);
if (err) {
pr_err("i915_gem_evict_vm on a full GGTT returned err=%d]\n",
@@ -356,7 +356,7 @@ static int igt_evict_vm(void *arg)
for_i915_gem_ww(&ww, err, false) {
mutex_lock(&ggtt->vm.mutex);
- err = i915_gem_evict_vm(&ggtt->vm, &ww);
+ err = i915_gem_evict_vm(&ggtt->vm, &ww, NULL);
mutex_unlock(&ggtt->vm.mutex);
}
--
2.38.1
More information about the Intel-gfx-trybot
mailing list