Merge tag 'drm-xe-fixes-2024-04-04' of https://gitlab.freedesktop.org/drm/xe/kernel into drm-fixes

- Stop using system_unbound_wq for preempt fences,
  as this can cause starvation when reaching more
  than max_active defined by workqueue
- Fix saving unordered rebinding fences by attaching
  them as kernel feces to the vm's resv
- Fix TLB invalidation fences completing out of order
- Move rebind TLB invalidation to the ring ops to reduce
  the latency

Signed-off-by: Dave Airlie <airlied@redhat.com>

From: Lucas De Marchi <lucas.demarchi@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/tizan6wdpxu4ayudeikjglxdgzmnhdzj3li3z2pgkierjtozzw@lbfddeg43a7h

Author: airlied

drivers/gpu/drm/xe/xe_device.c

@@ -193,6 +193,9 @@ static void xe_device_destroy(struct drm_device *dev, void *dummy)
 {
 	struct xe_device *xe = to_xe_device(dev);
 
+	if (xe->preempt_fence_wq)
+		destroy_workqueue(xe->preempt_fence_wq);
+
 	if (xe->ordered_wq)
 		destroy_workqueue(xe->ordered_wq);
 
@@ -258,9 +261,15 @@ struct xe_device *xe_device_create(struct pci_dev *pdev,
 	INIT_LIST_HEAD(&xe->pinned.external_vram);
 	INIT_LIST_HEAD(&xe->pinned.evicted);
 
+	xe->preempt_fence_wq = alloc_ordered_workqueue("xe-preempt-fence-wq", 0);
 	xe->ordered_wq = alloc_ordered_workqueue("xe-ordered-wq", 0);
 	xe->unordered_wq = alloc_workqueue("xe-unordered-wq", 0, 0);
-	if (!xe->ordered_wq || !xe->unordered_wq) {
+	if (!xe->ordered_wq || !xe->unordered_wq ||
+	    !xe->preempt_fence_wq) {
+		/*
+		 * Cleanup done in xe_device_destroy via
+		 * drmm_add_action_or_reset register above
+		 */
 		drm_err(&xe->drm, "Failed to allocate xe workqueues\n");
 		err = -ENOMEM;
 		goto err;

drivers/gpu/drm/xe/xe_device_types.h

@@ -363,6 +363,9 @@ struct xe_device {
 	/** @ufence_wq: user fence wait queue */
 	wait_queue_head_t ufence_wq;
 
+	/** @preempt_fence_wq: used to serialize preempt fences */
+	struct workqueue_struct *preempt_fence_wq;
+
 	/** @ordered_wq: used to serialize compute mode resume */
 	struct workqueue_struct *ordered_wq;
 

drivers/gpu/drm/xe/xe_exec.c

@@ -94,48 +94,16 @@
  *	Unlock all
  */
 
+/*
+ * Add validation and rebinding to the drm_exec locking loop, since both can
+ * trigger eviction which may require sleeping dma_resv locks.
+ */
 static int xe_exec_fn(struct drm_gpuvm_exec *vm_exec)
 {
 	struct xe_vm *vm = container_of(vm_exec->vm, struct xe_vm, gpuvm);
-	struct drm_gem_object *obj;
-	unsigned long index;
-	int num_fences;
-	int ret;
-
-	ret = drm_gpuvm_validate(vm_exec->vm, &vm_exec->exec);
-	if (ret)
-		return ret;
-
-	/*
-	 * 1 fence slot for the final submit, and 1 more for every per-tile for
-	 * GPU bind and 1 extra for CPU bind. Note that there are potentially
-	 * many vma per object/dma-resv, however the fence slot will just be
-	 * re-used, since they are largely the same timeline and the seqno
-	 * should be in order. In the case of CPU bind there is dummy fence used
-	 * for all CPU binds, so no need to have a per-tile slot for that.
-	 */
-	num_fences = 1 + 1 + vm->xe->info.tile_count;
 
-	/*
-	 * We don't know upfront exactly how many fence slots we will need at
-	 * the start of the exec, since the TTM bo_validate above can consume
-	 * numerous fence slots. Also due to how the dma_resv_reserve_fences()
-	 * works it only ensures that at least that many fence slots are
-	 * available i.e if there are already 10 slots available and we reserve
-	 * two more, it can just noop without reserving anything.  With this it
-	 * is quite possible that TTM steals some of the fence slots and then
-	 * when it comes time to do the vma binding and final exec stage we are
-	 * lacking enough fence slots, leading to some nasty BUG_ON() when
-	 * adding the fences. Hence just add our own fences here, after the
-	 * validate stage.
-	 */
-	drm_exec_for_each_locked_object(&vm_exec->exec, index, obj) {
-		ret = dma_resv_reserve_fences(obj->resv, num_fences);
-		if (ret)
-			return ret;
-	}
-
-	return 0;
+	/* The fence slot added here is intended for the exec sched job. */
+	return xe_vm_validate_rebind(vm, &vm_exec->exec, 1);
 }
 
 int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
@@ -152,7 +120,6 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 	struct drm_exec *exec = &vm_exec.exec;
 	u32 i, num_syncs = 0, num_ufence = 0;
 	struct xe_sched_job *job;
-	struct dma_fence *rebind_fence;
 	struct xe_vm *vm;
 	bool write_locked, skip_retry = false;
 	ktime_t end = 0;
@@ -290,39 +257,7 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 		goto err_exec;
 	}
 
-	/*
-	 * Rebind any invalidated userptr or evicted BOs in the VM, non-compute
-	 * VM mode only.
-	 */
-	rebind_fence = xe_vm_rebind(vm, false);
-	if (IS_ERR(rebind_fence)) {
-		err = PTR_ERR(rebind_fence);
-		goto err_put_job;
-	}
-
-	/*
-	 * We store the rebind_fence in the VM so subsequent execs don't get
-	 * scheduled before the rebinds of userptrs / evicted BOs is complete.
-	 */
-	if (rebind_fence) {
-		dma_fence_put(vm->rebind_fence);
-		vm->rebind_fence = rebind_fence;
-	}
-	if (vm->rebind_fence) {
-		if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
-			     &vm->rebind_fence->flags)) {
-			dma_fence_put(vm->rebind_fence);
-			vm->rebind_fence = NULL;
-		} else {
-			dma_fence_get(vm->rebind_fence);
-			err = drm_sched_job_add_dependency(&job->drm,
-							   vm->rebind_fence);
-			if (err)
-				goto err_put_job;
-		}
-	}
-
-	/* Wait behind munmap style rebinds */
+	/* Wait behind rebinds */
 	if (!xe_vm_in_lr_mode(vm)) {
 		err = drm_sched_job_add_resv_dependencies(&job->drm,
 							  xe_vm_resv(vm),

drivers/gpu/drm/xe/xe_exec_queue_types.h

@@ -148,6 +148,11 @@ struct xe_exec_queue {
 	const struct xe_ring_ops *ring_ops;
 	/** @entity: DRM sched entity for this exec queue (1 to 1 relationship) */
 	struct drm_sched_entity *entity;
+	/**
+	 * @tlb_flush_seqno: The seqno of the last rebind tlb flush performed
+	 * Protected by @vm's resv. Unused if @vm == NULL.
+	 */
+	u64 tlb_flush_seqno;
 	/** @lrc: logical ring context for this exec queue */
 	struct xe_lrc lrc[];
 };

drivers/gpu/drm/xe/xe_gt_pagefault.c

@@ -100,10 +100,9 @@ static int xe_pf_begin(struct drm_exec *exec, struct xe_vma *vma,
 {
 	struct xe_bo *bo = xe_vma_bo(vma);
 	struct xe_vm *vm = xe_vma_vm(vma);
-	unsigned int num_shared = 2; /* slots for bind + move */
 	int err;
 
-	err = xe_vm_prepare_vma(exec, vma, num_shared);
+	err = xe_vm_lock_vma(exec, vma);
 	if (err)
 		return err;
 

drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c

@@ -61,7 +61,6 @@ int xe_gt_tlb_invalidation_init(struct xe_gt *gt)
 	INIT_LIST_HEAD(&gt->tlb_invalidation.pending_fences);
 	spin_lock_init(&gt->tlb_invalidation.pending_lock);
 	spin_lock_init(&gt->tlb_invalidation.lock);
-	gt->tlb_invalidation.fence_context = dma_fence_context_alloc(1);
 	INIT_DELAYED_WORK(&gt->tlb_invalidation.fence_tdr,
 			  xe_gt_tlb_fence_timeout);
 

drivers/gpu/drm/xe/xe_gt_types.h

@@ -177,13 +177,6 @@ struct xe_gt {
 		 * xe_gt_tlb_fence_timeout after the timeut interval is over.
 		 */
 		struct delayed_work fence_tdr;
-		/** @tlb_invalidation.fence_context: context for TLB invalidation fences */
-		u64 fence_context;
-		/**
-		 * @tlb_invalidation.fence_seqno: seqno to TLB invalidation fences, protected by
-		 * tlb_invalidation.lock
-		 */
-		u32 fence_seqno;
 		/** @tlb_invalidation.lock: protects TLB invalidation fences */
 		spinlock_t lock;
 	} tlb_invalidation;

drivers/gpu/drm/xe/xe_preempt_fence.c

@@ -49,7 +49,7 @@ static bool preempt_fence_enable_signaling(struct dma_fence *fence)
 	struct xe_exec_queue *q = pfence->q;
 
 	pfence->error = q->ops->suspend(q);
-	queue_work(system_unbound_wq, &pfence->preempt_work);
+	queue_work(q->vm->xe->preempt_fence_wq, &pfence->preempt_work);
 	return true;
 }
 

drivers/gpu/drm/xe/xe_pt.c

@@ -1135,8 +1135,7 @@ static int invalidation_fence_init(struct xe_gt *gt,
 	spin_lock_irq(&gt->tlb_invalidation.lock);
 	dma_fence_init(&ifence->base.base, &invalidation_fence_ops,
 		       &gt->tlb_invalidation.lock,
-		       gt->tlb_invalidation.fence_context,
-		       ++gt->tlb_invalidation.fence_seqno);
+		       dma_fence_context_alloc(1), 1);
 	spin_unlock_irq(&gt->tlb_invalidation.lock);
 
 	INIT_LIST_HEAD(&ifence->base.link);
@@ -1236,6 +1235,13 @@ __xe_pt_bind_vma(struct xe_tile *tile, struct xe_vma *vma, struct xe_exec_queue
 	err = xe_pt_prepare_bind(tile, vma, entries, &num_entries);
 	if (err)
 		goto err;
+
+	err = dma_resv_reserve_fences(xe_vm_resv(vm), 1);
+	if (!err && !xe_vma_has_no_bo(vma) && !xe_vma_bo(vma)->vm)
+		err = dma_resv_reserve_fences(xe_vma_bo(vma)->ttm.base.resv, 1);
+	if (err)
+		goto err;
+
 	xe_tile_assert(tile, num_entries <= ARRAY_SIZE(entries));
 
 	xe_vm_dbg_print_entries(tile_to_xe(tile), entries, num_entries);
@@ -1254,11 +1260,13 @@ __xe_pt_bind_vma(struct xe_tile *tile, struct xe_vma *vma, struct xe_exec_queue
 	 * non-faulting LR, in particular on user-space batch buffer chaining,
 	 * it needs to be done here.
 	 */
-	if ((rebind && !xe_vm_in_lr_mode(vm) && !vm->batch_invalidate_tlb) ||
-	    (!rebind && xe_vm_has_scratch(vm) && xe_vm_in_preempt_fence_mode(vm))) {
+	if ((!rebind && xe_vm_has_scratch(vm) && xe_vm_in_preempt_fence_mode(vm))) {
 		ifence = kzalloc(sizeof(*ifence), GFP_KERNEL);
 		if (!ifence)
 			return ERR_PTR(-ENOMEM);
+	} else if (rebind && !xe_vm_in_lr_mode(vm)) {
+		/* We bump also if batch_invalidate_tlb is true */
+		vm->tlb_flush_seqno++;
 	}
 
 	rfence = kzalloc(sizeof(*rfence), GFP_KERNEL);
@@ -1297,7 +1305,7 @@ __xe_pt_bind_vma(struct xe_tile *tile, struct xe_vma *vma, struct xe_exec_queue
 		}
 
 		/* add shared fence now for pagetable delayed destroy */
-		dma_resv_add_fence(xe_vm_resv(vm), fence, !rebind &&
+		dma_resv_add_fence(xe_vm_resv(vm), fence, rebind ||
 				   last_munmap_rebind ?
 				   DMA_RESV_USAGE_KERNEL :
 				   DMA_RESV_USAGE_BOOKKEEP);
@@ -1576,6 +1584,7 @@ __xe_pt_unbind_vma(struct xe_tile *tile, struct xe_vma *vma, struct xe_exec_queu
 	struct dma_fence *fence = NULL;
 	struct invalidation_fence *ifence;
 	struct xe_range_fence *rfence;
+	int err;
 
 	LLIST_HEAD(deferred);
 
@@ -1593,6 +1602,12 @@ __xe_pt_unbind_vma(struct xe_tile *tile, struct xe_vma *vma, struct xe_exec_queu
 	xe_pt_calc_rfence_interval(vma, &unbind_pt_update, entries,
 				   num_entries);
 
+	err = dma_resv_reserve_fences(xe_vm_resv(vm), 1);
+	if (!err && !xe_vma_has_no_bo(vma) && !xe_vma_bo(vma)->vm)
+		err = dma_resv_reserve_fences(xe_vma_bo(vma)->ttm.base.resv, 1);
+	if (err)
+		return ERR_PTR(err);
+
 	ifence = kzalloc(sizeof(*ifence), GFP_KERNEL);
 	if (!ifence)
 		return ERR_PTR(-ENOMEM);

drivers/gpu/drm/xe/xe_ring_ops.c

@@ -219,10 +219,9 @@ static void __emit_job_gen12_simple(struct xe_sched_job *job, struct xe_lrc *lrc
 {
 	u32 dw[MAX_JOB_SIZE_DW], i = 0;
 	u32 ppgtt_flag = get_ppgtt_flag(job);
-	struct xe_vm *vm = job->q->vm;
 	struct xe_gt *gt = job->q->gt;
 
-	if (vm && vm->batch_invalidate_tlb) {
+	if (job->ring_ops_flush_tlb) {
 		dw[i++] = preparser_disable(true);
 		i = emit_flush_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
 					seqno, true, dw, i);
@@ -270,7 +269,6 @@ static void __emit_job_gen12_video(struct xe_sched_job *job, struct xe_lrc *lrc,
 	struct xe_gt *gt = job->q->gt;
 	struct xe_device *xe = gt_to_xe(gt);
 	bool decode = job->q->class == XE_ENGINE_CLASS_VIDEO_DECODE;
-	struct xe_vm *vm = job->q->vm;
 
 	dw[i++] = preparser_disable(true);
 
@@ -282,13 +280,13 @@ static void __emit_job_gen12_video(struct xe_sched_job *job, struct xe_lrc *lrc,
 			i = emit_aux_table_inv(gt, VE0_AUX_INV, dw, i);
 	}
 
-	if (vm && vm->batch_invalidate_tlb)
+	if (job->ring_ops_flush_tlb)
 		i = emit_flush_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
 					seqno, true, dw, i);
 
 	dw[i++] = preparser_disable(false);
 
-	if (!vm || !vm->batch_invalidate_tlb)
+	if (!job->ring_ops_flush_tlb)
 		i = emit_store_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
 					seqno, dw, i);
 
@@ -317,7 +315,6 @@ static void __emit_job_gen12_render_compute(struct xe_sched_job *job,
 	struct xe_gt *gt = job->q->gt;
 	struct xe_device *xe = gt_to_xe(gt);
 	bool lacks_render = !(gt->info.engine_mask & XE_HW_ENGINE_RCS_MASK);
-	struct xe_vm *vm = job->q->vm;
 	u32 mask_flags = 0;
 
 	dw[i++] = preparser_disable(true);
@@ -327,7 +324,7 @@ static void __emit_job_gen12_render_compute(struct xe_sched_job *job,
 		mask_flags = PIPE_CONTROL_3D_ENGINE_FLAGS;
 
 	/* See __xe_pt_bind_vma() for a discussion on TLB invalidations. */
-	i = emit_pipe_invalidate(mask_flags, vm && vm->batch_invalidate_tlb, dw, i);
+	i = emit_pipe_invalidate(mask_flags, job->ring_ops_flush_tlb, dw, i);
 
 	/* hsdes: 1809175790 */
 	if (has_aux_ccs(xe))