summaryrefslogtreecommitdiff
path: root/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c')
-rw-r--r--drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c197
1 files changed, 143 insertions, 54 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 41fbc4fd0fac..d17b2452cb1f 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -25,6 +25,7 @@
#include <linux/pagemap.h>
#include <linux/sched/mm.h>
#include <linux/sched/task.h>
+#include <linux/fdtable.h>
#include <drm/ttm/ttm_tt.h>
#include <drm/drm_exec.h>
@@ -806,13 +807,22 @@ kfd_mem_dmaunmap_attachment(struct kgd_mem *mem,
static int kfd_mem_export_dmabuf(struct kgd_mem *mem)
{
if (!mem->dmabuf) {
- struct dma_buf *ret = amdgpu_gem_prime_export(
- &mem->bo->tbo.base,
+ struct amdgpu_device *bo_adev;
+ struct dma_buf *dmabuf;
+ int r, fd;
+
+ bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
+ r = drm_gem_prime_handle_to_fd(&bo_adev->ddev, bo_adev->kfd.client.file,
+ mem->gem_handle,
mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
- DRM_RDWR : 0);
- if (IS_ERR(ret))
- return PTR_ERR(ret);
- mem->dmabuf = ret;
+ DRM_RDWR : 0, &fd);
+ if (r)
+ return r;
+ dmabuf = dma_buf_get(fd);
+ close_fd(fd);
+ if (WARN_ON_ONCE(IS_ERR(dmabuf)))
+ return PTR_ERR(dmabuf);
+ mem->dmabuf = dmabuf;
}
return 0;
@@ -1137,7 +1147,7 @@ static int reserve_bo_and_vm(struct kgd_mem *mem,
ctx->n_vms = 1;
ctx->sync = &mem->sync;
- drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT);
+ drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
drm_exec_until_all_locked(&ctx->exec) {
ret = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
drm_exec_retry_on_contention(&ctx->exec);
@@ -1176,7 +1186,7 @@ static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
int ret;
ctx->sync = &mem->sync;
- drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT);
+ drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
drm_exec_until_all_locked(&ctx->exec) {
ctx->n_vms = 0;
list_for_each_entry(entry, &mem->attachments, list) {
@@ -1384,7 +1394,6 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
amdgpu_amdkfd_restore_userptr_worker);
*process_info = info;
- *ef = dma_fence_get(&info->eviction_fence->base);
}
vm->process_info = *process_info;
@@ -1415,6 +1424,8 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
list_add_tail(&vm->vm_list_node,
&(vm->process_info->vm_list_head));
vm->process_info->n_vms++;
+
+ *ef = dma_fence_get(&vm->process_info->eviction_fence->base);
mutex_unlock(&vm->process_info->lock);
return 0;
@@ -1426,10 +1437,7 @@ validate_pd_fail:
reserve_pd_fail:
vm->process_info = NULL;
if (info) {
- /* Two fence references: one in info and one in *ef */
dma_fence_put(&info->eviction_fence->base);
- dma_fence_put(*ef);
- *ef = NULL;
*process_info = NULL;
put_pid(info->pid);
create_evict_fence_fail:
@@ -1623,7 +1631,8 @@ int amdgpu_amdkfd_criu_resume(void *p)
goto out_unlock;
}
WRITE_ONCE(pinfo->block_mmu_notifications, false);
- schedule_delayed_work(&pinfo->restore_userptr_work, 0);
+ queue_delayed_work(system_freezable_wq,
+ &pinfo->restore_userptr_work, 0);
out_unlock:
mutex_unlock(&pinfo->lock);
@@ -1779,6 +1788,9 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
pr_debug("Failed to allow vma node access. ret %d\n", ret);
goto err_node_allow;
}
+ ret = drm_gem_handle_create(adev->kfd.client.file, gobj, &(*mem)->gem_handle);
+ if (ret)
+ goto err_gem_handle_create;
bo = gem_to_amdgpu_bo(gobj);
if (bo_type == ttm_bo_type_sg) {
bo->tbo.sg = sg;
@@ -1830,6 +1842,8 @@ allocate_init_user_pages_failed:
err_pin_bo:
err_validate_bo:
remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
+ drm_gem_handle_delete(adev->kfd.client.file, (*mem)->gem_handle);
+err_gem_handle_create:
drm_vma_node_revoke(&gobj->vma_node, drm_priv);
err_node_allow:
/* Don't unreserve system mem limit twice */
@@ -1942,8 +1956,11 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
/* Free the BO*/
drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
- if (mem->dmabuf)
+ drm_gem_handle_delete(adev->kfd.client.file, mem->gem_handle);
+ if (mem->dmabuf) {
dma_buf_put(mem->dmabuf);
+ mem->dmabuf = NULL;
+ }
mutex_destroy(&mem->lock);
/* If this releases the last reference, it will end up calling
@@ -2295,34 +2312,26 @@ int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
return 0;
}
-int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
- struct dma_buf *dma_buf,
- uint64_t va, void *drm_priv,
- struct kgd_mem **mem, uint64_t *size,
- uint64_t *mmap_offset)
+static int import_obj_create(struct amdgpu_device *adev,
+ struct dma_buf *dma_buf,
+ struct drm_gem_object *obj,
+ uint64_t va, void *drm_priv,
+ struct kgd_mem **mem, uint64_t *size,
+ uint64_t *mmap_offset)
{
struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
- struct drm_gem_object *obj;
struct amdgpu_bo *bo;
int ret;
- obj = amdgpu_gem_prime_import(adev_to_drm(adev), dma_buf);
- if (IS_ERR(obj))
- return PTR_ERR(obj);
-
bo = gem_to_amdgpu_bo(obj);
if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
- AMDGPU_GEM_DOMAIN_GTT))) {
+ AMDGPU_GEM_DOMAIN_GTT)))
/* Only VRAM and GTT BOs are supported */
- ret = -EINVAL;
- goto err_put_obj;
- }
+ return -EINVAL;
*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
- if (!*mem) {
- ret = -ENOMEM;
- goto err_put_obj;
- }
+ if (!*mem)
+ return -ENOMEM;
ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
if (ret)
@@ -2372,8 +2381,41 @@ err_remove_mem:
drm_vma_node_revoke(&obj->vma_node, drm_priv);
err_free_mem:
kfree(*mem);
+ return ret;
+}
+
+int amdgpu_amdkfd_gpuvm_import_dmabuf_fd(struct amdgpu_device *adev, int fd,
+ uint64_t va, void *drm_priv,
+ struct kgd_mem **mem, uint64_t *size,
+ uint64_t *mmap_offset)
+{
+ struct drm_gem_object *obj;
+ uint32_t handle;
+ int ret;
+
+ ret = drm_gem_prime_fd_to_handle(&adev->ddev, adev->kfd.client.file, fd,
+ &handle);
+ if (ret)
+ return ret;
+ obj = drm_gem_object_lookup(adev->kfd.client.file, handle);
+ if (!obj) {
+ ret = -EINVAL;
+ goto err_release_handle;
+ }
+
+ ret = import_obj_create(adev, obj->dma_buf, obj, va, drm_priv, mem, size,
+ mmap_offset);
+ if (ret)
+ goto err_put_obj;
+
+ (*mem)->gem_handle = handle;
+
+ return 0;
+
err_put_obj:
drm_gem_object_put(obj);
+err_release_handle:
+ drm_gem_handle_delete(adev->kfd.client.file, handle);
return ret;
}
@@ -2426,7 +2468,8 @@ int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
if (r)
pr_err("Failed to quiesce KFD\n");
- schedule_delayed_work(&process_info->restore_userptr_work,
+ queue_delayed_work(system_freezable_wq,
+ &process_info->restore_userptr_work,
msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
}
mutex_unlock(&process_info->notifier_lock);
@@ -2552,7 +2595,7 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
amdgpu_sync_create(&sync);
- drm_exec_init(&exec, 0);
+ drm_exec_init(&exec, 0, 0);
/* Reserve all BOs and page tables for validation */
drm_exec_until_all_locked(&exec) {
/* Reserve all the page directories */
@@ -2749,7 +2792,8 @@ unlock_out:
/* If validation failed, reschedule another attempt */
if (evicted_bos) {
- schedule_delayed_work(&process_info->restore_userptr_work,
+ queue_delayed_work(system_freezable_wq,
+ &process_info->restore_userptr_work,
msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
kfd_smi_event_queue_restore_rescheduled(mm);
@@ -2758,6 +2802,23 @@ unlock_out:
put_task_struct(usertask);
}
+static void replace_eviction_fence(struct dma_fence **ef,
+ struct dma_fence *new_ef)
+{
+ struct dma_fence *old_ef = rcu_replace_pointer(*ef, new_ef, true
+ /* protected by process_info->lock */);
+
+ /* If we're replacing an unsignaled eviction fence, that fence will
+ * never be signaled, and if anyone is still waiting on that fence,
+ * they will hang forever. This should never happen. We should only
+ * replace the fence in restore_work that only gets scheduled after
+ * eviction work signaled the fence.
+ */
+ WARN_ONCE(!dma_fence_is_signaled(old_ef),
+ "Replacing unsignaled eviction fence");
+ dma_fence_put(old_ef);
+}
+
/** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
* KFD process identified by process_info
*
@@ -2781,7 +2842,6 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
struct amdkfd_process_info *process_info = info;
struct amdgpu_vm *peer_vm;
struct kgd_mem *mem;
- struct amdgpu_amdkfd_fence *new_fence;
struct list_head duplicate_save;
struct amdgpu_sync sync_obj;
unsigned long failed_size = 0;
@@ -2793,7 +2853,7 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
mutex_lock(&process_info->lock);
- drm_exec_init(&exec, 0);
+ drm_exec_init(&exec, 0, 0);
drm_exec_until_all_locked(&exec) {
list_for_each_entry(peer_vm, &process_info->vm_list_head,
vm_list_node) {
@@ -2825,12 +2885,6 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
if (ret)
goto validate_map_fail;
- ret = process_sync_pds_resv(process_info, &sync_obj);
- if (ret) {
- pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
- 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) {
@@ -2881,6 +2935,19 @@ 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);
+ /* Update mappings not managed by KFD */
+ list_for_each_entry(peer_vm, &process_info->vm_list_head,
+ vm_list_node) {
+ struct amdgpu_device *adev = amdgpu_ttm_adev(
+ peer_vm->root.bo->tbo.bdev);
+
+ ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket);
+ if (ret) {
+ pr_debug("Memory eviction: handle moved failed. Try again\n");
+ goto validate_map_fail;
+ }
+ }
+
/* Update page directories */
ret = process_update_pds(process_info, &sync_obj);
if (ret) {
@@ -2888,25 +2955,47 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
goto validate_map_fail;
}
+ /* Sync with fences on all the page tables. They implicitly depend on any
+ * move fences from amdgpu_vm_handle_moved above.
+ */
+ ret = process_sync_pds_resv(process_info, &sync_obj);
+ if (ret) {
+ pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
+ goto validate_map_fail;
+ }
+
/* Wait for validate and PT updates to finish */
amdgpu_sync_wait(&sync_obj, false);
- /* Release old eviction fence and create new one, because fence only
- * goes from unsignaled to signaled, fence cannot be reused.
- * Use context and mm from the old fence.
+ /* The old eviction fence may be unsignaled if restore happens
+ * after a GPU reset or suspend/resume. Keep the old fence in that
+ * case. Otherwise release the old eviction fence and create new
+ * one, because fence only goes from unsignaled to signaled once
+ * and cannot be reused. Use context and mm from the old fence.
+ *
+ * If an old eviction fence signals after this check, that's OK.
+ * Anyone signaling an eviction fence must stop the queues first
+ * and schedule another restore worker.
*/
- new_fence = amdgpu_amdkfd_fence_create(
+ if (dma_fence_is_signaled(&process_info->eviction_fence->base)) {
+ struct amdgpu_amdkfd_fence *new_fence =
+ amdgpu_amdkfd_fence_create(
process_info->eviction_fence->base.context,
process_info->eviction_fence->mm,
NULL);
- if (!new_fence) {
- pr_err("Failed to create eviction fence\n");
- ret = -ENOMEM;
- goto validate_map_fail;
+
+ if (!new_fence) {
+ pr_err("Failed to create eviction fence\n");
+ ret = -ENOMEM;
+ goto validate_map_fail;
+ }
+ dma_fence_put(&process_info->eviction_fence->base);
+ process_info->eviction_fence = new_fence;
+ replace_eviction_fence(ef, dma_fence_get(&new_fence->base));
+ } else {
+ WARN_ONCE(*ef != &process_info->eviction_fence->base,
+ "KFD eviction fence doesn't match KGD process_info");
}
- dma_fence_put(&process_info->eviction_fence->base);
- process_info->eviction_fence = new_fence;
- *ef = dma_fence_get(&new_fence->base);
/* Attach new eviction fence to all BOs except pinned ones */
list_for_each_entry(mem, &process_info->kfd_bo_list, validate_list) {
generated by cgit v1.2.3-70-g09d2 (git 2.51.1) at 2025-11-04 01:48:43 +0000