drm/xe: Use DRM GPUVM helpers for external- and evicted objects

Adapt to the DRM_GPUVM helpers moving removing a lot of complicated
driver-specific code.

For now this uses fine-grained locking for the evict list and external
object list, which may incur a slight performance penalty in some
situations.

v2:
- Don't lock all bos and validate on LR exec submissions (Matthew Brost)
- Add some kerneldoc
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Acked-by: Matthew Brost <matthew.brost@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231212100144.6833-2-thomas.hellstrom@linux.intel.comSigned-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>

drivers/gpu/drm/xe/xe_bo.c

@@ -468,9 +468,9 @@ static int xe_bo_trigger_rebind(struct xe_device *xe, struct xe_bo *bo,
 {
 	struct dma_resv_iter cursor;
 	struct dma_fence *fence;
-	struct drm_gpuva *gpuva;
 	struct drm_gem_object *obj = &bo->ttm.base;
 	struct drm_gpuvm_bo *vm_bo;
+	bool idle = false;
 	int ret = 0;
 
 	dma_resv_assert_held(bo->ttm.base.resv);
@@ -484,14 +484,15 @@ static int xe_bo_trigger_rebind(struct xe_device *xe, struct xe_bo *bo,
 	}
 
 	drm_gem_for_each_gpuvm_bo(vm_bo, obj) {
-		drm_gpuvm_bo_for_each_va(gpuva, vm_bo) {
-			struct xe_vma *vma = gpuva_to_vma(gpuva);
-			struct xe_vm *vm = xe_vma_vm(vma);
+		struct xe_vm *vm = gpuvm_to_vm(vm_bo->vm);
+		struct drm_gpuva *gpuva;
 
-			trace_xe_vma_evict(vma);
+		if (!xe_vm_in_fault_mode(vm)) {
+			drm_gpuvm_bo_evict(vm_bo, true);
+			continue;
+		}
 
-		if (xe_vm_in_fault_mode(vm)) {
-			/* Wait for pending binds / unbinds. */
+		if (!idle) {
 			long timeout;
 
 			if (ctx->no_wait_gpu &&
@@ -503,45 +504,21 @@ static int xe_bo_trigger_rebind(struct xe_device *xe, struct xe_bo *bo,
 							DMA_RESV_USAGE_BOOKKEEP,
 							ctx->interruptible,
 							MAX_SCHEDULE_TIMEOUT);
-			if (timeout > 0) {
-				ret = xe_vm_invalidate_vma(vma);
-				XE_WARN_ON(ret);
-			} else if (!timeout) {
-				ret = -ETIME;
-			} else {
-				ret = timeout;
-			}
-
-		} else {
-			bool vm_resv_locked = false;
+			if (!timeout)
+				return -ETIME;
+			if (timeout < 0)
+				return timeout;
 
-			/*
-			 * We need to put the vma on the vm's rebind_list,
-			 * but need the vm resv to do so. If we can't verify
-			 * that we indeed have it locked, put the vma an the
-			 * vm's notifier.rebind_list instead and scoop later.
-			 */
-			if (dma_resv_trylock(xe_vm_resv(vm)))
-				vm_resv_locked = true;
-			else if (ctx->resv != xe_vm_resv(vm)) {
-				spin_lock(&vm->notifier.list_lock);
-				if (!(vma->gpuva.flags & XE_VMA_DESTROYED))
-					list_move_tail(&vma->notifier.rebind_link,
-						       &vm->notifier.rebind_list);
-				spin_unlock(&vm->notifier.list_lock);
-				continue;
-			}
+			idle = true;
+		}
 
-			xe_vm_assert_held(vm);
-			if (vma->tile_present &&
-			    !(vma->gpuva.flags & XE_VMA_DESTROYED) &&
-			    list_empty(&vma->combined_links.rebind))
-				list_add_tail(&vma->combined_links.rebind,
-					      &vm->rebind_list);
+		drm_gpuvm_bo_for_each_va(gpuva, vm_bo) {
+			struct xe_vma *vma = gpuva_to_vma(gpuva);
 
-			if (vm_resv_locked)
-				dma_resv_unlock(xe_vm_resv(vm));
-		}
+			trace_xe_vma_evict(vma);
+			ret = xe_vm_invalidate_vma(vma);
+			if (XE_WARN_ON(ret))
+				return ret;
 		}
 	}
 

drivers/gpu/drm/xe/xe_exec.c

@@ -94,40 +94,9 @@
  *	Unlock all
  */
 
-static int xe_exec_begin(struct drm_exec *exec, struct xe_vm *vm)
+static int xe_exec_fn(struct drm_gpuvm_exec *vm_exec)
 {
-	struct xe_vma *vma;
-	LIST_HEAD(dups);
-	int err = 0;
-
-	if (xe_vm_in_lr_mode(vm))
-		return 0;
-
-	/*
-	 * 1 fence for job from exec plus a fence for each tile from a possible
-	 * rebind
-	 */
-	err = xe_vm_lock_dma_resv(vm, exec, 1 + vm->xe->info.tile_count, true);
-	if (err)
-		return err;
-
-	/*
-	 * Validate BOs that have been evicted (i.e. make sure the
-	 * BOs have valid placements possibly moving an evicted BO back
-	 * to a location where the GPU can access it).
-	 */
-	list_for_each_entry(vma, &vm->rebind_list, combined_links.rebind) {
-		xe_assert(vm->xe, !xe_vma_is_null(vma));
-
-		if (xe_vma_is_userptr(vma))
-			continue;
-
-		err = xe_bo_validate(xe_vma_bo(vma), vm, false);
-		if (err)
-			break;
-	}
-
-	return err;
+	return drm_gpuvm_validate(vm_exec->vm, &vm_exec->exec);
 }
 
 int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
@@ -140,7 +109,8 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 	struct xe_exec_queue *q;
 	struct xe_sync_entry *syncs = NULL;
 	u64 addresses[XE_HW_ENGINE_MAX_INSTANCE];
-	struct drm_exec exec;
+	struct drm_gpuvm_exec vm_exec = {.extra.fn = xe_exec_fn};
+	struct drm_exec *exec = &vm_exec.exec;
 	u32 i, num_syncs = 0;
 	struct xe_sched_job *job;
 	struct dma_fence *rebind_fence;
@@ -216,16 +186,18 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 			goto err_unlock_list;
 	}
 
-	drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT);
-	drm_exec_until_all_locked(&exec) {
-		err = xe_exec_begin(&exec, vm);
-		drm_exec_retry_on_contention(&exec);
-		if (err && xe_vm_validate_should_retry(&exec, err, &end)) {
-			err = -EAGAIN;
+	vm_exec.vm = &vm->gpuvm;
+	vm_exec.num_fences = 1 + vm->xe->info.tile_count;
+	vm_exec.flags = DRM_EXEC_INTERRUPTIBLE_WAIT;
+	if (xe_vm_in_lr_mode(vm)) {
+		drm_exec_init(exec, vm_exec.flags);
+	} else {
+		err = drm_gpuvm_exec_lock(&vm_exec);
+		if (err) {
+			if (xe_vm_validate_should_retry(exec, err, &end))
+				err = -EAGAIN;
 			goto err_unlock_list;
 		}
-		if (err)
-			goto err_exec;
 	}
 
 	if (xe_vm_is_closed_or_banned(q->vm)) {
@@ -307,19 +279,9 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 	 * the job and let the DRM scheduler / backend clean up the job.
 	 */
 	xe_sched_job_arm(job);
-	if (!xe_vm_in_lr_mode(vm)) {
-		/* Block userptr invalidations / BO eviction */
-		dma_resv_add_fence(xe_vm_resv(vm),
-				   &job->drm.s_fence->finished,
-				   DMA_RESV_USAGE_BOOKKEEP);
-
-		/*
-		 * Make implicit sync work across drivers, assuming all external
-		 * BOs are written as we don't pass in a read / write list.
-		 */
-		xe_vm_fence_all_extobjs(vm, &job->drm.s_fence->finished,
-					DMA_RESV_USAGE_WRITE);
-	}
+	if (!xe_vm_in_lr_mode(vm))
+		drm_gpuvm_resv_add_fence(&vm->gpuvm, exec, &job->drm.s_fence->finished,
+					 DMA_RESV_USAGE_BOOKKEEP, DMA_RESV_USAGE_WRITE);
 
 	for (i = 0; i < num_syncs; i++)
 		xe_sync_entry_signal(&syncs[i], job,
@@ -343,7 +305,7 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 	if (err)
 		xe_sched_job_put(job);
 err_exec:
-	drm_exec_fini(&exec);
+	drm_exec_fini(exec);
 err_unlock_list:
 	if (write_locked)
 		up_write(&vm->lock);

drivers/gpu/drm/xe/xe_vm.h

@@ -74,9 +74,20 @@ static inline bool xe_vm_has_scratch(const struct xe_vm *vm)
 	return vm->flags & XE_VM_FLAG_SCRATCH_PAGE;
 }
 
+/**
+ * gpuvm_to_vm() - Return the embedding xe_vm from a struct drm_gpuvm pointer
+ * @gpuvm: The struct drm_gpuvm pointer
+ *
+ * Return: Pointer to the embedding struct xe_vm.
+ */
+static inline struct xe_vm *gpuvm_to_vm(struct drm_gpuvm *gpuvm)
+{
+	return container_of(gpuvm, struct xe_vm, gpuvm);
+}
+
 static inline struct xe_vm *gpuva_to_vm(struct drm_gpuva *gpuva)
 {
-	return container_of(gpuva->vm, struct xe_vm, gpuvm);
+	return gpuvm_to_vm(gpuva->vm);
 }
 
 static inline struct xe_vma *gpuva_to_vma(struct drm_gpuva *gpuva)
@@ -219,12 +230,6 @@ int xe_vma_userptr_check_repin(struct xe_vma *vma);
 
 bool xe_vm_validate_should_retry(struct drm_exec *exec, int err, ktime_t *end);
 
-int xe_vm_lock_dma_resv(struct xe_vm *vm, struct drm_exec *exec,
-			unsigned int num_shared, bool lock_vm);
-
-void xe_vm_fence_all_extobjs(struct xe_vm *vm, struct dma_fence *fence,
-			     enum dma_resv_usage usage);
-
 int xe_analyze_vm(struct drm_printer *p, struct xe_vm *vm, int gt_id);
 
 int xe_vm_prepare_vma(struct drm_exec *exec, struct xe_vma *vma,

drivers/gpu/drm/xe/xe_vm_types.h

@@ -62,26 +62,17 @@ struct xe_vma {
 	/** @gpuva: Base GPUVA object */
 	struct drm_gpuva gpuva;
 
-	/** @combined_links: links into lists which are mutually exclusive */
+	/**
+	 * @combined_links: links into lists which are mutually exclusive.
+	 * Locking: vm lock in write mode OR vm lock in read mode and the vm's
+	 * resv.
+	 */
 	union {
-		/**
-		 * @userptr: link into VM repin list if userptr. Protected by
-		 * vm->lock in write mode.
-		 */
+		/** @userptr: link into VM repin list if userptr. */
 		struct list_head userptr;
-		/**
-		 * @rebind: link into VM if this VMA needs rebinding, and
-		 * if it's a bo (not userptr) needs validation after a possible
-		 * eviction. Protected by the vm's resv lock and typically
-		 * vm->lock is also held in write mode. The only place where
-		 * vm->lock isn't held is the BO eviction path which has
-		 * mutually exclusive execution with userptr.
-		 */
+		/** @rebind: link into VM if this VMA needs rebinding. */
 		struct list_head rebind;
-		/**
-		 * @destroy: link to contested list when VM is being closed.
-		 * Protected by vm->lock in write mode and vm's resv lock.
-		 */
+		/** @destroy: link to contested list when VM is being closed. */
 		struct list_head destroy;
 	} combined_links;
 
@@ -115,18 +106,6 @@ struct xe_vma {
 	 */
 	u16 pat_index;
 
-	struct {
-		struct list_head rebind_link;
-	} notifier;
-
-	struct {
-		/**
-		 * @extobj.link: Link into vm's external object list.
-		 * protected by the vm lock.
-		 */
-		struct list_head link;
-	} extobj;
-
 	/**
 	 * @userptr: user pointer state, only allocated for VMAs that are
 	 * user pointers
@@ -180,9 +159,9 @@ struct xe_vm {
 	struct rw_semaphore lock;
 
 	/**
-	 * @rebind_list: list of VMAs that need rebinding, and if they are
-	 * bos (not userptr), need validation after a possible eviction. The
-	 * list is protected by @resv.
+	 * @rebind_list: list of VMAs that need rebinding. Protected by the
+	 * vm->lock in write mode, OR (the vm->lock in read mode and the
+	 * vm resv).
 	 */
 	struct list_head rebind_list;
 
@@ -202,14 +181,6 @@ struct xe_vm {
 	 */
 	struct xe_range_fence_tree rftree[XE_MAX_TILES_PER_DEVICE];
 
-	/** @extobj: bookkeeping for external objects. Protected by the vm lock */
-	struct {
-		/** @enties: number of external BOs attached this VM */
-		u32 entries;
-		/** @list: list of vmas with external bos attached */
-		struct list_head list;
-	} extobj;
-
 	/** @async_ops: async VM operations (bind / unbinds) */
 	struct {
 		/** @list: list of pending async VM ops */
@@ -299,22 +270,6 @@ struct xe_vm {
 		struct xe_vma *last_fault_vma;
 	} usm;
 
-	/**
-	 * @notifier: Lists and locks for temporary usage within notifiers where
-	 * we either can't grab the vm lock or the vm resv.
-	 */
-	struct {
-		/** @notifier.list_lock: lock protecting @rebind_list */
-		spinlock_t list_lock;
-		/**
-		 * @notifier.rebind_list: list of vmas that we want to put on the
-		 * main @rebind_list. This list is protected for writing by both
-		 * notifier.list_lock, and the resv of the bo the vma points to,
-		 * and for reading by the notifier.list_lock only.
-		 */
-		struct list_head rebind_list;
-	} notifier;
-
 	/** @error_capture: allow to track errors */
 	struct {
 		/** @capture_once: capture only one error per VM */