[PATCH v4 3/5] drm/xe: Do not wait on TLB invalidations in page fault binds

[Matthew Brost <matthew.brost@intel.com>]

The migrate queue is shared by all processes using a device, thus is
possible while servicing a page fault another process uses the migrate
queue resulting in a TLB invalidation. In case of page fault binds, this
TLB invalidation has nothing to do with the current bind so there is not
need to wait on it. Teach the bind pipeline to be able to skip waits on
TLB invalidations.

Signed-off-by: Matthew Brost <matthew.brost@intel.com>
---
 drivers/gpu/drm/xe/xe_vm.c       | 14 ++++++++++++--
 drivers/gpu/drm/xe/xe_vm_types.h |  1 +
 2 files changed, 13 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
index 2f181c44b8b7..df0a44d9eb46 100644
--- a/drivers/gpu/drm/xe/xe_vm.c
+++ b/drivers/gpu/drm/xe/xe_vm.c
@@ -755,6 +755,7 @@ struct dma_fence *xe_vma_rebind(struct xe_vm *vm, struct xe_vma *vma, u8 tile_ma
 	xe_assert(vm->xe, xe_vm_in_fault_mode(vm));
 
 	xe_vma_ops_init(&vops, vm, NULL, NULL, 0);
+	vops.flags |= XE_VMA_OPS_FLAG_SKIP_TLB_WAIT;
 	for_each_tile(tile, vm->xe, id) {
 		vops.pt_update_ops[id].wait_vm_bookkeep = true;
 		vops.pt_update_ops[tile->id].q =
@@ -845,6 +846,7 @@ struct dma_fence *xe_vm_range_rebind(struct xe_vm *vm,
 	xe_assert(vm->xe, xe_vma_is_cpu_addr_mirror(vma));
 
 	xe_vma_ops_init(&vops, vm, NULL, NULL, 0);
+	vops.flags |= XE_VMA_OPS_FLAG_SKIP_TLB_WAIT;
 	for_each_tile(tile, vm->xe, id) {
 		vops.pt_update_ops[id].wait_vm_bookkeep = true;
 		vops.pt_update_ops[tile->id].q =
@@ -3111,8 +3113,13 @@ static struct dma_fence *ops_execute(struct xe_vm *vm,
 	if (number_tiles == 0)
 		return ERR_PTR(-ENODATA);
 
-	for_each_tile(tile, vm->xe, id)
-		n_fence += (1 + XE_MAX_GT_PER_TILE);
+	if (vops->flags & XE_VMA_OPS_FLAG_SKIP_TLB_WAIT) {
+		for_each_tile(tile, vm->xe, id)
+			++n_fence;
+	} else {
+		for_each_tile(tile, vm->xe, id)
+			n_fence += (1 + XE_MAX_GT_PER_TILE);
+	}
 
 	fences = kmalloc_array(n_fence, sizeof(*fences), GFP_KERNEL);
 	if (!fences) {
@@ -3153,6 +3160,9 @@ static struct dma_fence *ops_execute(struct xe_vm *vm,
 
 collect_fences:
 		fences[current_fence++] = fence ?: dma_fence_get_stub();
+		if (vops->flags & XE_VMA_OPS_FLAG_SKIP_TLB_WAIT)
+			continue;
+
 		xe_migrate_job_lock(tile->migrate, q);
 		for_each_tlb_inval(i)
 			fences[current_fence++] =
diff --git a/drivers/gpu/drm/xe/xe_vm_types.h b/drivers/gpu/drm/xe/xe_vm_types.h
index 542dbe2f9310..3766dc37b3ad 100644
--- a/drivers/gpu/drm/xe/xe_vm_types.h
+++ b/drivers/gpu/drm/xe/xe_vm_types.h
@@ -466,6 +466,7 @@ struct xe_vma_ops {
 #define XE_VMA_OPS_FLAG_HAS_SVM_PREFETCH BIT(0)
 #define XE_VMA_OPS_FLAG_MADVISE          BIT(1)
 #define XE_VMA_OPS_ARRAY_OF_BINDS	 BIT(2)
+#define XE_VMA_OPS_FLAG_SKIP_TLB_WAIT	 BIT(3)
 	u32 flags;
 #ifdef TEST_VM_OPS_ERROR
 	/** @inject_error: inject error to test error handling */
-- 
2.34.1