Am 15.11.23 um 00:32 schrieb Felix Kuehling:
DMABuf imports in compute VMs are not wrapped in a kgd_mem object on the
process_info->kfd_bo_list. There is no explicit KFD API call to validate
them or add eviction fences to them.
This patch automatically validates and fences dymanic DMABuf imports when
they are added to a compute VM. Revalidation after evictions is handled
in the VM code.
Signed-off-by: Felix Kuehling <Felix.Kuehling@xxxxxxx>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 3 +
.../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 45 ++++---
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 6 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c | 4 +
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 26 ++++
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 122 +++++++++++++++++-
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 12 +-
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 12 +-
8 files changed, 196 insertions(+), 34 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index fcf8a98ad15e..68d534a89942 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -178,6 +178,9 @@ int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
struct mm_struct *mm,
struct svm_range_bo *svm_bo);
+int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo,
+ uint32_t domain,
+ struct dma_fence *fence);
#if defined(CONFIG_DEBUG_FS)
int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data);
#endif
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 2e302956a279..bb83ab7a0256 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -423,9 +423,9 @@ static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
return ret;
}
-static int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo,
- uint32_t domain,
- struct dma_fence *fence)
+int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo,
+ uint32_t domain,
+ struct dma_fence *fence)
{
int ret = amdgpu_bo_reserve(bo, false);
@@ -461,13 +461,16 @@ static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo)
* again. Page directories are only updated after updating page
* tables.
*/
-static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm)
+static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm,
+ struct ww_acquire_ctx *ticket)
{
struct amdgpu_bo *pd = vm->root.bo;
struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
int ret;
- ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate_vm_bo, NULL);
+ ret = amdgpu_vm_validate_evicted_bos(adev, vm, ticket,
+ amdgpu_amdkfd_validate_vm_bo,
+ NULL);
if (ret) {
pr_err("failed to validate PT BOs\n");
return ret;
@@ -1347,14 +1350,15 @@ static int map_bo_to_gpuvm(struct kgd_mem *mem,
return ret;
}
-static int process_validate_vms(struct amdkfd_process_info *process_info)
+static int process_validate_vms(struct amdkfd_process_info *process_info,
+ struct ww_acquire_ctx *ticket)
{
struct amdgpu_vm *peer_vm;
int ret;
list_for_each_entry(peer_vm, &process_info->vm_list_head,
vm_list_node) {
- ret = vm_validate_pt_pd_bos(peer_vm);
+ ret = vm_validate_pt_pd_bos(peer_vm, ticket);
if (ret)
return ret;
}
@@ -1440,7 +1444,7 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
ret = amdgpu_bo_reserve(vm->root.bo, true);
if (ret)
goto reserve_pd_fail;
- ret = vm_validate_pt_pd_bos(vm);
+ ret = vm_validate_pt_pd_bos(vm, NULL);
if (ret) {
pr_err("validate_pt_pd_bos() failed\n");
goto validate_pd_fail;
@@ -2075,7 +2079,7 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
is_invalid_userptr = true;
- ret = vm_validate_pt_pd_bos(avm);
+ ret = vm_validate_pt_pd_bos(avm, NULL);
if (unlikely(ret))
goto out_unreserve;
@@ -2154,7 +2158,7 @@ int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
goto unreserve_out;
}
- ret = vm_validate_pt_pd_bos(avm);
+ ret = vm_validate_pt_pd_bos(avm, NULL);
if (unlikely(ret))
goto unreserve_out;
@@ -2635,7 +2639,7 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
amdgpu_sync_create(&sync);
- ret = process_validate_vms(process_info);
+ ret = process_validate_vms(process_info, NULL);
if (ret)
goto unreserve_out;
@@ -2887,11 +2891,6 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
amdgpu_sync_create(&sync_obj);
- /* Validate PDs and PTs */
- ret = process_validate_vms(process_info);
- if (ret)
- goto validate_map_fail;
-
/* Validate BOs and map them to GPUVM (update VM page tables). */
list_for_each_entry(mem, &process_info->kfd_bo_list,
validate_list.head) {
@@ -2942,6 +2941,13 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
if (failed_size)
pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
+ /* Validate PDs, PTs and evicted DMABuf imports last. Otherwise BO
+ * validations above would invalidate DMABuf imports again.
+ */
+ ret = process_validate_vms(process_info, &ctx.ticket);
+ if (ret)
+ goto validate_map_fail;
+
/* Update mappings not managed by KFD */
list_for_each_entry(peer_vm, &process_info->vm_list_head,
vm_list_node) {
@@ -3001,7 +3007,7 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
&process_info->eviction_fence->base,
DMA_RESV_USAGE_BOOKKEEP);
}
- /* Attach eviction fence to PD / PT BOs */
+ /* Attach eviction fence to PD / PT BOs and DMABuf imports */
list_for_each_entry(peer_vm, &process_info->vm_list_head,
vm_list_node) {
struct amdgpu_bo *bo = peer_vm->root.bo;
@@ -3009,6 +3015,11 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
dma_resv_add_fence(bo->tbo.base.resv,
&process_info->eviction_fence->base,
DMA_RESV_USAGE_BOOKKEEP);
+
+ ret = amdgpu_vm_fence_imports(peer_vm, &ctx.ticket,
+ &process_info->eviction_fence->base);
+ if (ret)
+ break;
}
validate_map_fail:
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
index 1d4e80f72573..7d78a5841db4 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
@@ -968,10 +968,10 @@ static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
p->bytes_moved = 0;
p->bytes_moved_vis = 0;
- r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
- amdgpu_cs_bo_validate, p);
+ r = amdgpu_vm_validate_evicted_bos(p->adev, &fpriv->vm, NULL,
+ amdgpu_cs_bo_validate, p);
if (r) {
- DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
+ DRM_ERROR("amdgpu_vm_validate_evicted_bos() failed.\n");
goto error_validate;
}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c
index e7e87a3b2601..8b556669cec0 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c
@@ -373,6 +373,10 @@ amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
struct amdgpu_vm_bo_base *bo_base;
int r;
+ /* FIXME: This should be after the "if", but needs a fix to make sure
+ * DMABuf imports are initialized in the right VM list.
+ */
+ amdgpu_vm_bo_invalidate(adev, bo, false);
if (!bo->tbo.resource || bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
return;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
index 849fffbb367d..fd43921844ed 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
@@ -186,6 +186,32 @@ static int amdgpu_gem_object_open(struct drm_gem_object *obj,
else
++bo_va->ref_count;
amdgpu_bo_unreserve(abo);
+
+ /* Validate and add eviction fence to DMABuf imports with dynamic
+ * attachment in compute VMs. Re-validation will be done by
+ * amdgpu_vm_validate_evicted_bos and the fence will be updated by
+ * amdgpu_vm_fence_imports in amdgpu_amdkfd_gpuvm_restore_process_bos.
+ *
+ * Nested locking below for the case that a GEM object is opened in
+ * kfd_mem_export_dmabuf. Since the lock below is only taken for imports,
+ * but not for export, this is a different lock class that cannot lead to
+ * circular lock dependencies.
+ */
+ if (!vm->is_compute_context || !vm->process_info)
+ return 0;
+ if (!obj->import_attach ||
+ !dma_buf_is_dynamic(obj->import_attach->dmabuf))
+ return 0;
+ mutex_lock_nested(&vm->process_info->lock, 1);
+ if (!WARN_ON(!vm->process_info->eviction_fence)) {
+ r = amdgpu_amdkfd_bo_validate_and_fence(abo, AMDGPU_GEM_DOMAIN_GTT,
+ &vm->process_info->eviction_fence->base);
+ if (r)
+ dev_warn(adev->dev, "%d: validate_and_fence failed: %d\n",
+ vm->task_info.pid, r);
Why not returning the error code instead of just printing a message?
+ }
+ mutex_unlock(&vm->process_info->lock);
+
return 0;
}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 1c9397198083..9d3fb161cc0a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -368,24 +368,31 @@ void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
}
/**
- * amdgpu_vm_validate_pt_bos - validate the page table BOs
+ * amdgpu_vm_validate_evicted_bos - validate evicted BOs tracked in the VM
*
* @adev: amdgpu device pointer
* @vm: vm providing the BOs
+ * @ticket: optional reservation ticket used to reserve the VM
* @validate: callback to do the validation
* @param: parameter for the validation callback
*
- * Validate the page table BOs on command submission if neccessary.
+ * Validate the page table BOs and per-VM BOs on command submission if
+ * necessary. If a ticket is given, also try to validate evicted user queue
+ * BOs. They must already be reserved with the given ticket.
*
* Returns:
* Validation result.
*/
-int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
- int (*validate)(void *p, struct amdgpu_bo *bo),
- void *param)
+int amdgpu_vm_validate_evicted_bos(struct amdgpu_device *adev,
Maybe just use amdgpu_vm_validate() for the name here. That the
validated BOs are evicted should be obvious.
+ struct amdgpu_vm *vm,
+ struct ww_acquire_ctx *ticket,
+ int (*validate)(void *p,
+ struct amdgpu_bo *bo),
+ void *param)
{
struct amdgpu_vm_bo_base *bo_base;
struct amdgpu_bo *shadow;
+ struct dma_resv *resv;
struct amdgpu_bo *bo;
int r;
@@ -416,6 +423,28 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
}
spin_lock(&vm->status_lock);
}
+ while (ticket && !list_empty(&vm->evicted_user)) {
+ bo_base = list_first_entry(&vm->evicted_user,
+ struct amdgpu_vm_bo_base,
+ vm_status);
+ spin_unlock(&vm->status_lock);
+
+ bo = bo_base->bo;
+ resv = bo->tbo.base.resv;
+
+ if (dma_resv_locking_ctx(resv) != ticket) {
+ pr_warn_ratelimited("Evicted user BO is not reserved in pid %d\n",
+ vm->task_info.pid);
+ return -EINVAL;
+ }
+
+ r = validate(param, bo);
+ if (r)
+ return r;
+
+ amdgpu_vm_bo_invalidated(bo_base);
+ spin_lock(&vm->status_lock);
+ }
I see two main differences to the evicted state here:
1. We now handle BOs which don't use the shared reservation object and
needs to be locked manually with the ticket before calling this function.
2. The BOs are then moved into a different state after validating them.
That could be handled with just an additional if if I'm not completely
mistaken which would make the extra state superfluous.
spin_unlock(&vm->status_lock);
amdgpu_vm_eviction_lock(vm);
@@ -1358,11 +1387,21 @@ int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
}
r = amdgpu_vm_bo_update(adev, bo_va, clear);
- if (r)
- return r;
if (unlock)
dma_resv_unlock(resv);
+ if (r)
+ return r;
+
+ /* Remember evicted DMABuf imports in compute VMs for later
+ * validation
+ */
+ if (vm->is_compute_context && bo_va->base.bo &&
+ bo_va->base.bo->tbo.base.import_attach &&
+ (!bo_va->base.bo->tbo.resource ||
+ bo_va->base.bo->tbo.resource->mem_type == TTM_PL_SYSTEM))
+ amdgpu_vm_bo_evicted_user(&bo_va->base);
+
spin_lock(&vm->status_lock);
}
spin_unlock(&vm->status_lock);
@@ -1370,6 +1409,75 @@ int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
return 0;
}
+/**
+ * amdgpu_vm_fence_imports - add fence to valid DMABuf imports
+ *
+ * @vm: requested vm
+ * @ticket: optional reservation ticket used to reserve the VM
+ * @fence: fence to add
+ *
+ * Add the specified fence to all dynamic DMABuf imports that are valid.
+ *
+ * Returns:
+ * 0 for success.
+ */
+int amdgpu_vm_fence_imports(struct amdgpu_vm *vm,
+ struct ww_acquire_ctx *ticket,
+ struct dma_fence *fence)
+{
+ struct amdgpu_bo_va *bo_va, *tmp;
+ struct dma_resv *resv;
+ LIST_HEAD(imports);
+ bool unlock;
+ int ret = 0;
+
+ if (!vm->is_compute_context)
+ return 0;
+
+ /* Move all the DMABuf imports to a private list so I can reserve
+ * them while not holding te status_lock.
+ */
+ spin_lock(&vm->status_lock);
+ list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
+ if (bo_va->base.bo && bo_va->base.bo->tbo.base.import_attach &&
+ dma_buf_is_dynamic(bo_va->base.bo->tbo.base.import_attach->dmabuf))
+ list_move(&bo_va->base.vm_status, &imports);
+ }
+ spin_unlock(&vm->status_lock);
That looks a lot like it won't work.
What's preventing the status of DMA-buf BOs from changing while doing this?
Regards,
Christian.
+
+ list_for_each_entry(bo_va, &imports, base.vm_status) {
+ resv = bo_va->base.bo->tbo.base.resv;
+
+ /* Try to reserve the BO */
+ if (dma_resv_trylock(resv)) {
+ unlock = true;
+ /* The caller is already holding the reservation lock */
+ } else if (ticket && dma_resv_locking_ctx(resv) == ticket) {
+ unlock = false;
+ } else {
+ WARN_ONCE(1, "Failed to reserve DMABuf import");
+ ret = -EBUSY;
+ break;
+ }
+
+ ret = dma_resv_reserve_fences(resv, 1);
+ if (!ret)
+ dma_resv_add_fence(resv, fence,
+ DMA_RESV_USAGE_BOOKKEEP);
+
+ if (unlock)
+ dma_resv_unlock(resv);
+ if (ret)
+ break;
+ }
+
+ spin_lock(&vm->status_lock);
+ list_splice(&imports, &vm->idle);
+ spin_unlock(&vm->status_lock);
+
+ return ret;
+}
+
/**
* amdgpu_vm_flush_compute_tlb - Flush TLB on compute VM
*
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
index 939d0c2219c0..a380c4acf8e8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
@@ -432,9 +432,12 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
struct list_head *validated,
struct amdgpu_bo_list_entry *entry);
bool amdgpu_vm_ready(struct amdgpu_vm *vm);
-int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
- int (*callback)(void *p, struct amdgpu_bo *bo),
- void *param);
+int amdgpu_vm_validate_evicted_bos(struct amdgpu_device *adev,
+ struct amdgpu_vm *vm,
+ struct ww_acquire_ctx *ticket,
+ int (*callback)(void *p,
+ struct amdgpu_bo *bo),
+ void *param);
int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync);
int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
struct amdgpu_vm *vm, bool immediate);
@@ -444,6 +447,9 @@ int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
struct amdgpu_vm *vm,
struct ww_acquire_ctx *ticket);
+int amdgpu_vm_fence_imports(struct amdgpu_vm *vm,
+ struct ww_acquire_ctx *ticket,
+ struct dma_fence *fence);
int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
struct amdgpu_vm *vm,
uint32_t flush_type,
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index b23ba92a794c..6eec0f4ddde4 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1558,9 +1558,11 @@ static int svm_range_reserve_bos(struct svm_validate_context *ctx)
goto unreserve_out;
}
- r = amdgpu_vm_validate_pt_bos(pdd->dev->adev,
- drm_priv_to_vm(pdd->drm_priv),
- svm_range_bo_validate, NULL);
+ r = amdgpu_vm_validate_evicted_bos(pdd->dev->adev,
+ drm_priv_to_vm(pdd->drm_priv),
+ NULL,
+ svm_range_bo_validate,
+ NULL);
if (r) {
pr_debug("failed %d validate pt bos\n", r);
goto unreserve_out;
@@ -1677,7 +1679,9 @@ static int svm_range_validate_and_map(struct mm_struct *mm,
goto free_ctx;
}
- svm_range_reserve_bos(ctx);
+ r = svm_range_reserve_bos(ctx);
+ if (r)
+ goto free_ctx;
p = container_of(prange->svms, struct kfd_process, svms);
owner = kfd_svm_page_owner(p, find_first_bit(ctx->bitmap,
