1 // SPDX-License-Identifier: MIT
3 * Copyright © 2021 Intel Corporation
8 #include <linux/dma-fence-array.h>
10 #include <drm/drm_print.h>
11 #include <drm/ttm/ttm_execbuf_util.h>
12 #include <drm/ttm/ttm_tt.h>
13 #include <drm/xe_drm.h>
14 #include <linux/delay.h>
15 #include <linux/kthread.h>
17 #include <linux/swap.h>
20 #include "xe_device.h"
21 #include "xe_engine.h"
23 #include "xe_gt_pagefault.h"
24 #include "xe_gt_tlb_invalidation.h"
25 #include "xe_migrate.h"
27 #include "xe_preempt_fence.h"
29 #include "xe_res_cursor.h"
33 #define TEST_VM_ASYNC_OPS_ERROR
36 * xe_vma_userptr_check_repin() - Advisory check for repin needed
37 * @vma: The userptr vma
39 * Check if the userptr vma has been invalidated since last successful
40 * repin. The check is advisory only and can the function can be called
41 * without the vm->userptr.notifier_lock held. There is no guarantee that the
42 * vma userptr will remain valid after a lockless check, so typically
43 * the call needs to be followed by a proper check under the notifier_lock.
45 * Return: 0 if userptr vma is valid, -EAGAIN otherwise; repin recommended.
47 int xe_vma_userptr_check_repin(struct xe_vma *vma)
49 return mmu_interval_check_retry(&vma->userptr.notifier,
50 vma->userptr.notifier_seq) ?
54 int xe_vma_userptr_pin_pages(struct xe_vma *vma)
56 struct xe_vm *vm = vma->vm;
57 struct xe_device *xe = vm->xe;
58 const unsigned long num_pages =
59 (vma->end - vma->start + 1) >> PAGE_SHIFT;
61 bool in_kthread = !current->mm;
62 unsigned long notifier_seq;
64 bool read_only = vma->pte_flags & XE_PTE_READ_ONLY;
66 lockdep_assert_held(&vm->lock);
67 XE_BUG_ON(!xe_vma_is_userptr(vma));
72 notifier_seq = mmu_interval_read_begin(&vma->userptr.notifier);
73 if (notifier_seq == vma->userptr.notifier_seq)
76 pages = kvmalloc_array(num_pages, sizeof(*pages), GFP_KERNEL);
80 if (vma->userptr.sg) {
81 dma_unmap_sgtable(xe->drm.dev,
83 read_only ? DMA_TO_DEVICE :
84 DMA_BIDIRECTIONAL, 0);
85 sg_free_table(vma->userptr.sg);
86 vma->userptr.sg = NULL;
91 if (!mmget_not_zero(vma->userptr.notifier.mm)) {
95 kthread_use_mm(vma->userptr.notifier.mm);
98 while (pinned < num_pages) {
99 ret = get_user_pages_fast(vma->userptr.ptr + pinned * PAGE_SIZE,
101 read_only ? 0 : FOLL_WRITE,
114 kthread_unuse_mm(vma->userptr.notifier.mm);
115 mmput(vma->userptr.notifier.mm);
121 ret = sg_alloc_table_from_pages_segment(&vma->userptr.sgt, pages,
123 (u64)pinned << PAGE_SHIFT,
124 xe_sg_segment_size(xe->drm.dev),
127 vma->userptr.sg = NULL;
130 vma->userptr.sg = &vma->userptr.sgt;
132 ret = dma_map_sgtable(xe->drm.dev, vma->userptr.sg,
133 read_only ? DMA_TO_DEVICE :
135 DMA_ATTR_SKIP_CPU_SYNC |
136 DMA_ATTR_NO_KERNEL_MAPPING);
138 sg_free_table(vma->userptr.sg);
139 vma->userptr.sg = NULL;
143 for (i = 0; i < pinned; ++i) {
146 set_page_dirty(pages[i]);
147 unlock_page(pages[i]);
150 mark_page_accessed(pages[i]);
154 release_pages(pages, pinned);
158 vma->userptr.notifier_seq = notifier_seq;
159 if (xe_vma_userptr_check_repin(vma) == -EAGAIN)
163 return ret < 0 ? ret : 0;
166 static bool preempt_fences_waiting(struct xe_vm *vm)
170 lockdep_assert_held(&vm->lock);
171 xe_vm_assert_held(vm);
173 list_for_each_entry(e, &vm->preempt.engines, compute.link) {
174 if (!e->compute.pfence || (e->compute.pfence &&
175 test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
176 &e->compute.pfence->flags))) {
184 static void free_preempt_fences(struct list_head *list)
186 struct list_head *link, *next;
188 list_for_each_safe(link, next, list)
189 xe_preempt_fence_free(to_preempt_fence_from_link(link));
192 static int alloc_preempt_fences(struct xe_vm *vm, struct list_head *list,
195 lockdep_assert_held(&vm->lock);
196 xe_vm_assert_held(vm);
198 if (*count >= vm->preempt.num_engines)
201 for (; *count < vm->preempt.num_engines; ++(*count)) {
202 struct xe_preempt_fence *pfence = xe_preempt_fence_alloc();
205 return PTR_ERR(pfence);
207 list_move_tail(xe_preempt_fence_link(pfence), list);
213 static int wait_for_existing_preempt_fences(struct xe_vm *vm)
217 xe_vm_assert_held(vm);
219 list_for_each_entry(e, &vm->preempt.engines, compute.link) {
220 if (e->compute.pfence) {
221 long timeout = dma_fence_wait(e->compute.pfence, false);
225 dma_fence_put(e->compute.pfence);
226 e->compute.pfence = NULL;
233 static bool xe_vm_is_idle(struct xe_vm *vm)
237 xe_vm_assert_held(vm);
238 list_for_each_entry(e, &vm->preempt.engines, compute.link) {
239 if (!xe_engine_is_idle(e))
246 static void arm_preempt_fences(struct xe_vm *vm, struct list_head *list)
248 struct list_head *link;
251 list_for_each_entry(e, &vm->preempt.engines, compute.link) {
252 struct dma_fence *fence;
255 XE_BUG_ON(link == list);
257 fence = xe_preempt_fence_arm(to_preempt_fence_from_link(link),
258 e, e->compute.context,
260 dma_fence_put(e->compute.pfence);
261 e->compute.pfence = fence;
265 static int add_preempt_fences(struct xe_vm *vm, struct xe_bo *bo)
268 struct ww_acquire_ctx ww;
271 err = xe_bo_lock(bo, &ww, vm->preempt.num_engines, true);
275 list_for_each_entry(e, &vm->preempt.engines, compute.link)
276 if (e->compute.pfence) {
277 dma_resv_add_fence(bo->ttm.base.resv,
279 DMA_RESV_USAGE_BOOKKEEP);
282 xe_bo_unlock(bo, &ww);
287 * xe_vm_fence_all_extobjs() - Add a fence to vm's external objects' resv
289 * @fence: The fence to add.
290 * @usage: The resv usage for the fence.
292 * Loops over all of the vm's external object bindings and adds a @fence
293 * with the given @usage to all of the external object's reservation
296 void xe_vm_fence_all_extobjs(struct xe_vm *vm, struct dma_fence *fence,
297 enum dma_resv_usage usage)
301 list_for_each_entry(vma, &vm->extobj.list, extobj.link)
302 dma_resv_add_fence(vma->bo->ttm.base.resv, fence, usage);
305 static void resume_and_reinstall_preempt_fences(struct xe_vm *vm)
309 lockdep_assert_held(&vm->lock);
310 xe_vm_assert_held(vm);
312 list_for_each_entry(e, &vm->preempt.engines, compute.link) {
315 dma_resv_add_fence(&vm->resv, e->compute.pfence,
316 DMA_RESV_USAGE_BOOKKEEP);
317 xe_vm_fence_all_extobjs(vm, e->compute.pfence,
318 DMA_RESV_USAGE_BOOKKEEP);
322 int xe_vm_add_compute_engine(struct xe_vm *vm, struct xe_engine *e)
324 struct ttm_validate_buffer tv_onstack[XE_ONSTACK_TV];
325 struct ttm_validate_buffer *tv;
326 struct ww_acquire_ctx ww;
327 struct list_head objs;
328 struct dma_fence *pfence;
332 XE_BUG_ON(!xe_vm_in_compute_mode(vm));
334 down_write(&vm->lock);
336 err = xe_vm_lock_dma_resv(vm, &ww, tv_onstack, &tv, &objs, true, 1);
338 goto out_unlock_outer;
340 pfence = xe_preempt_fence_create(e, e->compute.context,
347 list_add(&e->compute.link, &vm->preempt.engines);
348 ++vm->preempt.num_engines;
349 e->compute.pfence = pfence;
351 down_read(&vm->userptr.notifier_lock);
353 dma_resv_add_fence(&vm->resv, pfence,
354 DMA_RESV_USAGE_BOOKKEEP);
356 xe_vm_fence_all_extobjs(vm, pfence, DMA_RESV_USAGE_BOOKKEEP);
359 * Check to see if a preemption on VM is in flight or userptr
360 * invalidation, if so trigger this preempt fence to sync state with
361 * other preempt fences on the VM.
363 wait = __xe_vm_userptr_needs_repin(vm) || preempt_fences_waiting(vm);
365 dma_fence_enable_sw_signaling(pfence);
367 up_read(&vm->userptr.notifier_lock);
370 xe_vm_unlock_dma_resv(vm, tv_onstack, tv, &ww, &objs);
378 * __xe_vm_userptr_needs_repin() - Check whether the VM does have userptrs
379 * that need repinning.
382 * This function checks for whether the VM has userptrs that need repinning,
383 * and provides a release-type barrier on the userptr.notifier_lock after
386 * Return: 0 if there are no userptrs needing repinning, -EAGAIN if there are.
388 int __xe_vm_userptr_needs_repin(struct xe_vm *vm)
390 lockdep_assert_held_read(&vm->userptr.notifier_lock);
392 return (list_empty(&vm->userptr.repin_list) &&
393 list_empty(&vm->userptr.invalidated)) ? 0 : -EAGAIN;
397 * xe_vm_lock_dma_resv() - Lock the vm dma_resv object and the dma_resv
398 * objects of the vm's external buffer objects.
400 * @ww: Pointer to a struct ww_acquire_ctx locking context.
401 * @tv_onstack: Array size XE_ONSTACK_TV of storage for the struct
402 * ttm_validate_buffers used for locking.
403 * @tv: Pointer to a pointer that on output contains the actual storage used.
404 * @objs: List head for the buffer objects locked.
405 * @intr: Whether to lock interruptible.
406 * @num_shared: Number of dma-fence slots to reserve in the locked objects.
408 * Locks the vm dma-resv objects and all the dma-resv objects of the
409 * buffer objects on the vm external object list. The TTM utilities require
410 * a list of struct ttm_validate_buffers pointing to the actual buffer
411 * objects to lock. Storage for those struct ttm_validate_buffers should
412 * be provided in @tv_onstack, and is typically reserved on the stack
413 * of the caller. If the size of @tv_onstack isn't sufficient, then
414 * storage will be allocated internally using kvmalloc().
416 * The function performs deadlock handling internally, and after a
417 * successful return the ww locking transaction should be considered
420 * Return: 0 on success, Negative error code on error. In particular if
421 * @intr is set to true, -EINTR or -ERESTARTSYS may be returned. In case
422 * of error, any locking performed has been reverted.
424 int xe_vm_lock_dma_resv(struct xe_vm *vm, struct ww_acquire_ctx *ww,
425 struct ttm_validate_buffer *tv_onstack,
426 struct ttm_validate_buffer **tv,
427 struct list_head *objs,
429 unsigned int num_shared)
431 struct ttm_validate_buffer *tv_vm, *tv_bo;
432 struct xe_vma *vma, *next;
436 lockdep_assert_held(&vm->lock);
438 if (vm->extobj.entries < XE_ONSTACK_TV) {
441 tv_vm = kvmalloc_array(vm->extobj.entries + 1, sizeof(*tv_vm),
448 INIT_LIST_HEAD(objs);
449 list_for_each_entry(vma, &vm->extobj.list, extobj.link) {
450 tv_bo->num_shared = num_shared;
451 tv_bo->bo = &vma->bo->ttm;
453 list_add_tail(&tv_bo->head, objs);
456 tv_vm->num_shared = num_shared;
457 tv_vm->bo = xe_vm_ttm_bo(vm);
458 list_add_tail(&tv_vm->head, objs);
459 err = ttm_eu_reserve_buffers(ww, objs, intr, &dups);
463 spin_lock(&vm->notifier.list_lock);
464 list_for_each_entry_safe(vma, next, &vm->notifier.rebind_list,
465 notifier.rebind_link) {
466 xe_bo_assert_held(vma->bo);
468 list_del_init(&vma->notifier.rebind_link);
469 if (vma->tile_present && !vma->destroyed)
470 list_move_tail(&vma->rebind_link, &vm->rebind_list);
472 spin_unlock(&vm->notifier.list_lock);
478 if (tv_vm != tv_onstack)
485 * xe_vm_unlock_dma_resv() - Unlock reservation objects locked by
486 * xe_vm_lock_dma_resv()
488 * @tv_onstack: The @tv_onstack array given to xe_vm_lock_dma_resv().
489 * @tv: The value of *@tv given by xe_vm_lock_dma_resv().
490 * @ww: The ww_acquire_context used for locking.
491 * @objs: The list returned from xe_vm_lock_dma_resv().
493 * Unlocks the reservation objects and frees any memory allocated by
494 * xe_vm_lock_dma_resv().
496 void xe_vm_unlock_dma_resv(struct xe_vm *vm,
497 struct ttm_validate_buffer *tv_onstack,
498 struct ttm_validate_buffer *tv,
499 struct ww_acquire_ctx *ww,
500 struct list_head *objs)
503 * Nothing should've been able to enter the list while we were locked,
504 * since we've held the dma-resvs of all the vm's external objects,
505 * and holding the dma_resv of an object is required for list
506 * addition, and we shouldn't add ourselves.
508 XE_WARN_ON(!list_empty(&vm->notifier.rebind_list));
510 ttm_eu_backoff_reservation(ww, objs);
511 if (tv && tv != tv_onstack)
515 #define XE_VM_REBIND_RETRY_TIMEOUT_MS 1000
517 static void preempt_rebind_work_func(struct work_struct *w)
519 struct xe_vm *vm = container_of(w, struct xe_vm, preempt.rebind_work);
521 struct ttm_validate_buffer tv_onstack[XE_ONSTACK_TV];
522 struct ttm_validate_buffer *tv;
523 struct ww_acquire_ctx ww;
524 struct list_head objs;
525 struct dma_fence *rebind_fence;
526 unsigned int fence_count = 0;
527 LIST_HEAD(preempt_fences);
531 int __maybe_unused tries = 0;
533 XE_BUG_ON(!xe_vm_in_compute_mode(vm));
534 trace_xe_vm_rebind_worker_enter(vm);
536 if (xe_vm_is_closed(vm)) {
537 trace_xe_vm_rebind_worker_exit(vm);
541 down_write(&vm->lock);
544 if (vm->async_ops.error)
545 goto out_unlock_outer;
548 * Extreme corner where we exit a VM error state with a munmap style VM
549 * unbind inflight which requires a rebind. In this case the rebind
550 * needs to install some fences into the dma-resv slots. The worker to
551 * do this queued, let that worker make progress by dropping vm->lock
552 * and trying this again.
554 if (vm->async_ops.munmap_rebind_inflight) {
556 flush_work(&vm->async_ops.work);
560 if (xe_vm_userptr_check_repin(vm)) {
561 err = xe_vm_userptr_pin(vm);
563 goto out_unlock_outer;
566 err = xe_vm_lock_dma_resv(vm, &ww, tv_onstack, &tv, &objs,
567 false, vm->preempt.num_engines);
569 goto out_unlock_outer;
571 if (xe_vm_is_idle(vm)) {
572 vm->preempt.rebind_deactivated = true;
576 /* Fresh preempt fences already installed. Everyting is running. */
577 if (!preempt_fences_waiting(vm))
581 * This makes sure vm is completely suspended and also balances
582 * xe_engine suspend- and resume; we resume *all* vm engines below.
584 err = wait_for_existing_preempt_fences(vm);
588 err = alloc_preempt_fences(vm, &preempt_fences, &fence_count);
592 list_for_each_entry(vma, &vm->rebind_list, rebind_link) {
593 if (xe_vma_is_userptr(vma) || vma->destroyed)
596 err = xe_bo_validate(vma->bo, vm, false);
601 rebind_fence = xe_vm_rebind(vm, true);
602 if (IS_ERR(rebind_fence)) {
603 err = PTR_ERR(rebind_fence);
608 dma_fence_wait(rebind_fence, false);
609 dma_fence_put(rebind_fence);
612 /* Wait on munmap style VM unbinds */
613 wait = dma_resv_wait_timeout(&vm->resv,
614 DMA_RESV_USAGE_KERNEL,
615 false, MAX_SCHEDULE_TIMEOUT);
621 #define retry_required(__tries, __vm) \
622 (IS_ENABLED(CONFIG_DRM_XE_USERPTR_INVAL_INJECT) ? \
623 (!(__tries)++ || __xe_vm_userptr_needs_repin(__vm)) : \
624 __xe_vm_userptr_needs_repin(__vm))
626 down_read(&vm->userptr.notifier_lock);
627 if (retry_required(tries, vm)) {
628 up_read(&vm->userptr.notifier_lock);
633 #undef retry_required
635 /* Point of no return. */
636 arm_preempt_fences(vm, &preempt_fences);
637 resume_and_reinstall_preempt_fences(vm);
638 up_read(&vm->userptr.notifier_lock);
641 xe_vm_unlock_dma_resv(vm, tv_onstack, tv, &ww, &objs);
643 if (err == -EAGAIN) {
644 trace_xe_vm_rebind_worker_retry(vm);
649 * With multiple active VMs, under memory pressure, it is possible that
650 * ttm_bo_validate() run into -EDEADLK and in such case returns -ENOMEM.
651 * Until ttm properly handles locking in such scenarios, best thing the
652 * driver can do is retry with a timeout. Killing the VM or putting it
653 * in error state after timeout or other error scenarios is still TBD.
655 if (err == -ENOMEM) {
656 ktime_t cur = ktime_get();
658 end = end ? : ktime_add_ms(cur, XE_VM_REBIND_RETRY_TIMEOUT_MS);
659 if (ktime_before(cur, end)) {
661 trace_xe_vm_rebind_worker_retry(vm);
667 free_preempt_fences(&preempt_fences);
669 XE_WARN_ON(err < 0); /* TODO: Kill VM or put in error state */
670 trace_xe_vm_rebind_worker_exit(vm);
673 struct async_op_fence;
674 static int __xe_vm_bind(struct xe_vm *vm, struct xe_vma *vma,
675 struct xe_engine *e, struct xe_sync_entry *syncs,
676 u32 num_syncs, struct async_op_fence *afence);
678 static bool vma_userptr_invalidate(struct mmu_interval_notifier *mni,
679 const struct mmu_notifier_range *range,
680 unsigned long cur_seq)
682 struct xe_vma *vma = container_of(mni, struct xe_vma, userptr.notifier);
683 struct xe_vm *vm = vma->vm;
684 struct dma_resv_iter cursor;
685 struct dma_fence *fence;
688 XE_BUG_ON(!xe_vma_is_userptr(vma));
689 trace_xe_vma_userptr_invalidate(vma);
691 if (!mmu_notifier_range_blockable(range))
694 down_write(&vm->userptr.notifier_lock);
695 mmu_interval_set_seq(mni, cur_seq);
697 /* No need to stop gpu access if the userptr is not yet bound. */
698 if (!vma->userptr.initial_bind) {
699 up_write(&vm->userptr.notifier_lock);
704 * Tell exec and rebind worker they need to repin and rebind this
707 if (!xe_vm_in_fault_mode(vm) && !vma->destroyed && vma->tile_present) {
708 spin_lock(&vm->userptr.invalidated_lock);
709 list_move_tail(&vma->userptr.invalidate_link,
710 &vm->userptr.invalidated);
711 spin_unlock(&vm->userptr.invalidated_lock);
714 up_write(&vm->userptr.notifier_lock);
717 * Preempt fences turn into schedule disables, pipeline these.
718 * Note that even in fault mode, we need to wait for binds and
719 * unbinds to complete, and those are attached as BOOKMARK fences
722 dma_resv_iter_begin(&cursor, &vm->resv,
723 DMA_RESV_USAGE_BOOKKEEP);
724 dma_resv_for_each_fence_unlocked(&cursor, fence)
725 dma_fence_enable_sw_signaling(fence);
726 dma_resv_iter_end(&cursor);
728 err = dma_resv_wait_timeout(&vm->resv,
729 DMA_RESV_USAGE_BOOKKEEP,
730 false, MAX_SCHEDULE_TIMEOUT);
731 XE_WARN_ON(err <= 0);
733 if (xe_vm_in_fault_mode(vm)) {
734 err = xe_vm_invalidate_vma(vma);
738 trace_xe_vma_userptr_invalidate_complete(vma);
743 static const struct mmu_interval_notifier_ops vma_userptr_notifier_ops = {
744 .invalidate = vma_userptr_invalidate,
747 int xe_vm_userptr_pin(struct xe_vm *vm)
749 struct xe_vma *vma, *next;
751 LIST_HEAD(tmp_evict);
753 lockdep_assert_held_write(&vm->lock);
755 /* Collect invalidated userptrs */
756 spin_lock(&vm->userptr.invalidated_lock);
757 list_for_each_entry_safe(vma, next, &vm->userptr.invalidated,
758 userptr.invalidate_link) {
759 list_del_init(&vma->userptr.invalidate_link);
760 list_move_tail(&vma->userptr_link, &vm->userptr.repin_list);
762 spin_unlock(&vm->userptr.invalidated_lock);
764 /* Pin and move to temporary list */
765 list_for_each_entry_safe(vma, next, &vm->userptr.repin_list, userptr_link) {
766 err = xe_vma_userptr_pin_pages(vma);
770 list_move_tail(&vma->userptr_link, &tmp_evict);
773 /* Take lock and move to rebind_list for rebinding. */
774 err = dma_resv_lock_interruptible(&vm->resv, NULL);
778 list_for_each_entry_safe(vma, next, &tmp_evict, userptr_link) {
779 list_del_init(&vma->userptr_link);
780 list_move_tail(&vma->rebind_link, &vm->rebind_list);
783 dma_resv_unlock(&vm->resv);
788 list_splice_tail(&tmp_evict, &vm->userptr.repin_list);
794 * xe_vm_userptr_check_repin() - Check whether the VM might have userptrs
795 * that need repinning.
798 * This function does an advisory check for whether the VM has userptrs that
801 * Return: 0 if there are no indications of userptrs needing repinning,
802 * -EAGAIN if there are.
804 int xe_vm_userptr_check_repin(struct xe_vm *vm)
806 return (list_empty_careful(&vm->userptr.repin_list) &&
807 list_empty_careful(&vm->userptr.invalidated)) ? 0 : -EAGAIN;
810 static struct dma_fence *
811 xe_vm_bind_vma(struct xe_vma *vma, struct xe_engine *e,
812 struct xe_sync_entry *syncs, u32 num_syncs);
814 struct dma_fence *xe_vm_rebind(struct xe_vm *vm, bool rebind_worker)
816 struct dma_fence *fence = NULL;
817 struct xe_vma *vma, *next;
819 lockdep_assert_held(&vm->lock);
820 if (xe_vm_no_dma_fences(vm) && !rebind_worker)
823 xe_vm_assert_held(vm);
824 list_for_each_entry_safe(vma, next, &vm->rebind_list, rebind_link) {
825 XE_WARN_ON(!vma->tile_present);
827 list_del_init(&vma->rebind_link);
828 dma_fence_put(fence);
830 trace_xe_vma_rebind_worker(vma);
832 trace_xe_vma_rebind_exec(vma);
833 fence = xe_vm_bind_vma(vma, NULL, NULL, 0);
841 static struct xe_vma *xe_vma_create(struct xe_vm *vm,
843 u64 bo_offset_or_userptr,
849 struct xe_tile *tile;
852 XE_BUG_ON(start >= end);
853 XE_BUG_ON(end >= vm->size);
855 vma = kzalloc(sizeof(*vma), GFP_KERNEL);
857 vma = ERR_PTR(-ENOMEM);
861 INIT_LIST_HEAD(&vma->rebind_link);
862 INIT_LIST_HEAD(&vma->unbind_link);
863 INIT_LIST_HEAD(&vma->userptr_link);
864 INIT_LIST_HEAD(&vma->userptr.invalidate_link);
865 INIT_LIST_HEAD(&vma->notifier.rebind_link);
866 INIT_LIST_HEAD(&vma->extobj.link);
872 vma->pte_flags = XE_PTE_READ_ONLY;
875 vma->tile_mask = tile_mask;
877 for_each_tile(tile, vm->xe, id)
878 vma->tile_mask |= 0x1 << id;
881 if (vm->xe->info.platform == XE_PVC)
882 vma->use_atomic_access_pte_bit = true;
885 xe_bo_assert_held(bo);
886 vma->bo_offset = bo_offset_or_userptr;
887 vma->bo = xe_bo_get(bo);
888 list_add_tail(&vma->bo_link, &bo->vmas);
889 } else /* userptr */ {
890 u64 size = end - start + 1;
893 vma->userptr.ptr = bo_offset_or_userptr;
895 err = mmu_interval_notifier_insert(&vma->userptr.notifier,
897 vma->userptr.ptr, size,
898 &vma_userptr_notifier_ops);
905 vma->userptr.notifier_seq = LONG_MAX;
912 static bool vm_remove_extobj(struct xe_vma *vma)
914 if (!list_empty(&vma->extobj.link)) {
915 vma->vm->extobj.entries--;
916 list_del_init(&vma->extobj.link);
922 static void xe_vma_destroy_late(struct xe_vma *vma)
924 struct xe_vm *vm = vma->vm;
925 struct xe_device *xe = vm->xe;
926 bool read_only = vma->pte_flags & XE_PTE_READ_ONLY;
928 if (xe_vma_is_userptr(vma)) {
929 if (vma->userptr.sg) {
930 dma_unmap_sgtable(xe->drm.dev,
932 read_only ? DMA_TO_DEVICE :
933 DMA_BIDIRECTIONAL, 0);
934 sg_free_table(vma->userptr.sg);
935 vma->userptr.sg = NULL;
939 * Since userptr pages are not pinned, we can't remove
940 * the notifer until we're sure the GPU is not accessing
943 mmu_interval_notifier_remove(&vma->userptr.notifier);
952 static void vma_destroy_work_func(struct work_struct *w)
955 container_of(w, struct xe_vma, destroy_work);
957 xe_vma_destroy_late(vma);
960 static struct xe_vma *
961 bo_has_vm_references_locked(struct xe_bo *bo, struct xe_vm *vm,
962 struct xe_vma *ignore)
966 list_for_each_entry(vma, &bo->vmas, bo_link) {
967 if (vma != ignore && vma->vm == vm)
974 static bool bo_has_vm_references(struct xe_bo *bo, struct xe_vm *vm,
975 struct xe_vma *ignore)
977 struct ww_acquire_ctx ww;
980 xe_bo_lock(bo, &ww, 0, false);
981 ret = !!bo_has_vm_references_locked(bo, vm, ignore);
982 xe_bo_unlock(bo, &ww);
987 static void __vm_insert_extobj(struct xe_vm *vm, struct xe_vma *vma)
989 list_add(&vma->extobj.link, &vm->extobj.list);
990 vm->extobj.entries++;
993 static void vm_insert_extobj(struct xe_vm *vm, struct xe_vma *vma)
995 struct xe_bo *bo = vma->bo;
997 lockdep_assert_held_write(&vm->lock);
999 if (bo_has_vm_references(bo, vm, vma))
1002 __vm_insert_extobj(vm, vma);
1005 static void vma_destroy_cb(struct dma_fence *fence,
1006 struct dma_fence_cb *cb)
1008 struct xe_vma *vma = container_of(cb, struct xe_vma, destroy_cb);
1010 INIT_WORK(&vma->destroy_work, vma_destroy_work_func);
1011 queue_work(system_unbound_wq, &vma->destroy_work);
1014 static void xe_vma_destroy(struct xe_vma *vma, struct dma_fence *fence)
1016 struct xe_vm *vm = vma->vm;
1018 lockdep_assert_held_write(&vm->lock);
1019 XE_BUG_ON(!list_empty(&vma->unbind_link));
1021 if (xe_vma_is_userptr(vma)) {
1022 XE_WARN_ON(!vma->destroyed);
1023 spin_lock(&vm->userptr.invalidated_lock);
1024 list_del_init(&vma->userptr.invalidate_link);
1025 spin_unlock(&vm->userptr.invalidated_lock);
1026 list_del(&vma->userptr_link);
1028 xe_bo_assert_held(vma->bo);
1029 list_del(&vma->bo_link);
1031 spin_lock(&vm->notifier.list_lock);
1032 list_del(&vma->notifier.rebind_link);
1033 spin_unlock(&vm->notifier.list_lock);
1035 if (!vma->bo->vm && vm_remove_extobj(vma)) {
1036 struct xe_vma *other;
1038 other = bo_has_vm_references_locked(vma->bo, vm, NULL);
1041 __vm_insert_extobj(vm, other);
1045 xe_vm_assert_held(vm);
1046 if (!list_empty(&vma->rebind_link))
1047 list_del(&vma->rebind_link);
1050 int ret = dma_fence_add_callback(fence, &vma->destroy_cb,
1054 XE_WARN_ON(ret != -ENOENT);
1055 xe_vma_destroy_late(vma);
1058 xe_vma_destroy_late(vma);
1062 static void xe_vma_destroy_unlocked(struct xe_vma *vma)
1064 struct ttm_validate_buffer tv[2];
1065 struct ww_acquire_ctx ww;
1066 struct xe_bo *bo = vma->bo;
1071 memset(tv, 0, sizeof(tv));
1072 tv[0].bo = xe_vm_ttm_bo(vma->vm);
1073 list_add(&tv[0].head, &objs);
1076 tv[1].bo = &xe_bo_get(bo)->ttm;
1077 list_add(&tv[1].head, &objs);
1079 err = ttm_eu_reserve_buffers(&ww, &objs, false, &dups);
1082 xe_vma_destroy(vma, NULL);
1084 ttm_eu_backoff_reservation(&ww, &objs);
1089 static struct xe_vma *to_xe_vma(const struct rb_node *node)
1091 BUILD_BUG_ON(offsetof(struct xe_vma, vm_node) != 0);
1092 return (struct xe_vma *)node;
1095 static int xe_vma_cmp(const struct xe_vma *a, const struct xe_vma *b)
1097 if (a->end < b->start) {
1099 } else if (b->end < a->start) {
1106 static bool xe_vma_less_cb(struct rb_node *a, const struct rb_node *b)
1108 return xe_vma_cmp(to_xe_vma(a), to_xe_vma(b)) < 0;
1111 int xe_vma_cmp_vma_cb(const void *key, const struct rb_node *node)
1113 struct xe_vma *cmp = to_xe_vma(node);
1114 const struct xe_vma *own = key;
1116 if (own->start > cmp->end)
1119 if (own->end < cmp->start)
1126 xe_vm_find_overlapping_vma(struct xe_vm *vm, const struct xe_vma *vma)
1128 struct rb_node *node;
1130 if (xe_vm_is_closed(vm))
1133 XE_BUG_ON(vma->end >= vm->size);
1134 lockdep_assert_held(&vm->lock);
1136 node = rb_find(vma, &vm->vmas, xe_vma_cmp_vma_cb);
1138 return node ? to_xe_vma(node) : NULL;
1141 static void xe_vm_insert_vma(struct xe_vm *vm, struct xe_vma *vma)
1143 XE_BUG_ON(vma->vm != vm);
1144 lockdep_assert_held(&vm->lock);
1146 rb_add(&vma->vm_node, &vm->vmas, xe_vma_less_cb);
1149 static void xe_vm_remove_vma(struct xe_vm *vm, struct xe_vma *vma)
1151 XE_BUG_ON(vma->vm != vm);
1152 lockdep_assert_held(&vm->lock);
1154 rb_erase(&vma->vm_node, &vm->vmas);
1155 if (vm->usm.last_fault_vma == vma)
1156 vm->usm.last_fault_vma = NULL;
1159 static void async_op_work_func(struct work_struct *w);
1160 static void vm_destroy_work_func(struct work_struct *w);
1162 struct xe_vm *xe_vm_create(struct xe_device *xe, u32 flags)
1165 int err, i = 0, number_tiles = 0;
1166 struct xe_tile *tile;
1169 vm = kzalloc(sizeof(*vm), GFP_KERNEL);
1171 return ERR_PTR(-ENOMEM);
1174 kref_init(&vm->refcount);
1175 dma_resv_init(&vm->resv);
1177 vm->size = 1ull << xe_pt_shift(xe->info.vm_max_level + 1);
1182 init_rwsem(&vm->lock);
1184 INIT_LIST_HEAD(&vm->rebind_list);
1186 INIT_LIST_HEAD(&vm->userptr.repin_list);
1187 INIT_LIST_HEAD(&vm->userptr.invalidated);
1188 init_rwsem(&vm->userptr.notifier_lock);
1189 spin_lock_init(&vm->userptr.invalidated_lock);
1191 INIT_LIST_HEAD(&vm->notifier.rebind_list);
1192 spin_lock_init(&vm->notifier.list_lock);
1194 INIT_LIST_HEAD(&vm->async_ops.pending);
1195 INIT_WORK(&vm->async_ops.work, async_op_work_func);
1196 spin_lock_init(&vm->async_ops.lock);
1198 INIT_WORK(&vm->destroy_work, vm_destroy_work_func);
1200 INIT_LIST_HEAD(&vm->preempt.engines);
1201 vm->preempt.min_run_period_ms = 10; /* FIXME: Wire up to uAPI */
1203 INIT_LIST_HEAD(&vm->extobj.list);
1205 if (!(flags & XE_VM_FLAG_MIGRATION)) {
1206 /* We need to immeditatelly exit from any D3 state */
1207 xe_pm_runtime_get(xe);
1208 xe_device_mem_access_get(xe);
1211 err = dma_resv_lock_interruptible(&vm->resv, NULL);
1215 if (IS_DGFX(xe) && xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K)
1216 vm->flags |= XE_VM_FLAGS_64K;
1218 for_each_tile(tile, xe, id) {
1219 if (flags & XE_VM_FLAG_MIGRATION &&
1220 tile->id != XE_VM_FLAG_GT_ID(flags))
1223 vm->pt_root[id] = xe_pt_create(vm, tile, xe->info.vm_max_level);
1224 if (IS_ERR(vm->pt_root[id])) {
1225 err = PTR_ERR(vm->pt_root[id]);
1226 vm->pt_root[id] = NULL;
1227 goto err_destroy_root;
1231 if (flags & XE_VM_FLAG_SCRATCH_PAGE) {
1232 for_each_tile(tile, xe, id) {
1233 if (!vm->pt_root[id])
1236 err = xe_pt_create_scratch(xe, tile, vm);
1238 goto err_scratch_pt;
1240 vm->batch_invalidate_tlb = true;
1243 if (flags & DRM_XE_VM_CREATE_COMPUTE_MODE) {
1244 INIT_WORK(&vm->preempt.rebind_work, preempt_rebind_work_func);
1245 vm->flags |= XE_VM_FLAG_COMPUTE_MODE;
1246 vm->batch_invalidate_tlb = false;
1249 if (flags & DRM_XE_VM_CREATE_ASYNC_BIND_OPS) {
1250 vm->async_ops.fence.context = dma_fence_context_alloc(1);
1251 vm->flags |= XE_VM_FLAG_ASYNC_BIND_OPS;
1254 /* Fill pt_root after allocating scratch tables */
1255 for_each_tile(tile, xe, id) {
1256 if (!vm->pt_root[id])
1259 xe_pt_populate_empty(tile, vm, vm->pt_root[id]);
1261 dma_resv_unlock(&vm->resv);
1263 /* Kernel migration VM shouldn't have a circular loop.. */
1264 if (!(flags & XE_VM_FLAG_MIGRATION)) {
1265 for_each_tile(tile, xe, id) {
1266 struct xe_gt *gt = tile->primary_gt;
1267 struct xe_vm *migrate_vm;
1268 struct xe_engine *eng;
1270 if (!vm->pt_root[id])
1273 migrate_vm = xe_migrate_get_vm(tile->migrate);
1274 eng = xe_engine_create_class(xe, gt, migrate_vm,
1275 XE_ENGINE_CLASS_COPY,
1277 xe_vm_put(migrate_vm);
1279 xe_vm_close_and_put(vm);
1280 return ERR_CAST(eng);
1287 if (number_tiles > 1)
1288 vm->composite_fence_ctx = dma_fence_context_alloc(1);
1290 mutex_lock(&xe->usm.lock);
1291 if (flags & XE_VM_FLAG_FAULT_MODE)
1292 xe->usm.num_vm_in_fault_mode++;
1293 else if (!(flags & XE_VM_FLAG_MIGRATION))
1294 xe->usm.num_vm_in_non_fault_mode++;
1295 mutex_unlock(&xe->usm.lock);
1297 trace_xe_vm_create(vm);
1302 for_each_tile(tile, xe, id) {
1303 if (!vm->pt_root[id])
1306 i = vm->pt_root[id]->level;
1308 if (vm->scratch_pt[id][--i])
1309 xe_pt_destroy(vm->scratch_pt[id][i],
1311 xe_bo_unpin(vm->scratch_bo[id]);
1312 xe_bo_put(vm->scratch_bo[id]);
1315 for_each_tile(tile, xe, id) {
1316 if (vm->pt_root[id])
1317 xe_pt_destroy(vm->pt_root[id], vm->flags, NULL);
1319 dma_resv_unlock(&vm->resv);
1321 dma_resv_fini(&vm->resv);
1323 if (!(flags & XE_VM_FLAG_MIGRATION)) {
1324 xe_device_mem_access_put(xe);
1325 xe_pm_runtime_put(xe);
1327 return ERR_PTR(err);
1330 static void flush_async_ops(struct xe_vm *vm)
1332 queue_work(system_unbound_wq, &vm->async_ops.work);
1333 flush_work(&vm->async_ops.work);
1336 static void vm_error_capture(struct xe_vm *vm, int err,
1337 u32 op, u64 addr, u64 size)
1339 struct drm_xe_vm_bind_op_error_capture capture;
1340 u64 __user *address =
1341 u64_to_user_ptr(vm->async_ops.error_capture.addr);
1342 bool in_kthread = !current->mm;
1344 capture.error = err;
1346 capture.addr = addr;
1347 capture.size = size;
1350 if (!mmget_not_zero(vm->async_ops.error_capture.mm))
1352 kthread_use_mm(vm->async_ops.error_capture.mm);
1355 if (copy_to_user(address, &capture, sizeof(capture)))
1356 XE_WARN_ON("Copy to user failed");
1359 kthread_unuse_mm(vm->async_ops.error_capture.mm);
1360 mmput(vm->async_ops.error_capture.mm);
1364 wake_up_all(&vm->async_ops.error_capture.wq);
1367 void xe_vm_close_and_put(struct xe_vm *vm)
1369 struct rb_root contested = RB_ROOT;
1370 struct ww_acquire_ctx ww;
1371 struct xe_device *xe = vm->xe;
1372 struct xe_tile *tile;
1375 XE_BUG_ON(vm->preempt.num_engines);
1379 flush_async_ops(vm);
1380 if (xe_vm_in_compute_mode(vm))
1381 flush_work(&vm->preempt.rebind_work);
1383 for_each_tile(tile, xe, id) {
1385 xe_engine_kill(vm->eng[id]);
1386 xe_engine_put(vm->eng[id]);
1391 down_write(&vm->lock);
1392 xe_vm_lock(vm, &ww, 0, false);
1393 while (vm->vmas.rb_node) {
1394 struct xe_vma *vma = to_xe_vma(vm->vmas.rb_node);
1396 if (xe_vma_is_userptr(vma)) {
1397 down_read(&vm->userptr.notifier_lock);
1398 vma->destroyed = true;
1399 up_read(&vm->userptr.notifier_lock);
1402 rb_erase(&vma->vm_node, &vm->vmas);
1404 /* easy case, remove from VMA? */
1405 if (xe_vma_is_userptr(vma) || vma->bo->vm) {
1406 xe_vma_destroy(vma, NULL);
1410 rb_add(&vma->vm_node, &contested, xe_vma_less_cb);
1414 * All vm operations will add shared fences to resv.
1415 * The only exception is eviction for a shared object,
1416 * but even so, the unbind when evicted would still
1417 * install a fence to resv. Hence it's safe to
1418 * destroy the pagetables immediately.
1420 for_each_tile(tile, xe, id) {
1421 if (vm->scratch_bo[id]) {
1424 xe_bo_unpin(vm->scratch_bo[id]);
1425 xe_bo_put(vm->scratch_bo[id]);
1426 for (i = 0; i < vm->pt_root[id]->level; i++)
1427 xe_pt_destroy(vm->scratch_pt[id][i], vm->flags,
1431 xe_vm_unlock(vm, &ww);
1433 if (contested.rb_node) {
1436 * VM is now dead, cannot re-add nodes to vm->vmas if it's NULL
1437 * Since we hold a refcount to the bo, we can remove and free
1438 * the members safely without locking.
1440 while (contested.rb_node) {
1441 struct xe_vma *vma = to_xe_vma(contested.rb_node);
1443 rb_erase(&vma->vm_node, &contested);
1444 xe_vma_destroy_unlocked(vma);
1448 if (vm->async_ops.error_capture.addr)
1449 wake_up_all(&vm->async_ops.error_capture.wq);
1451 XE_WARN_ON(!list_empty(&vm->extobj.list));
1452 up_write(&vm->lock);
1454 mutex_lock(&xe->usm.lock);
1455 if (vm->flags & XE_VM_FLAG_FAULT_MODE)
1456 xe->usm.num_vm_in_fault_mode--;
1457 else if (!(vm->flags & XE_VM_FLAG_MIGRATION))
1458 xe->usm.num_vm_in_non_fault_mode--;
1459 mutex_unlock(&xe->usm.lock);
1464 static void vm_destroy_work_func(struct work_struct *w)
1467 container_of(w, struct xe_vm, destroy_work);
1468 struct ww_acquire_ctx ww;
1469 struct xe_device *xe = vm->xe;
1470 struct xe_tile *tile;
1474 /* xe_vm_close_and_put was not called? */
1475 XE_WARN_ON(vm->size);
1477 if (!(vm->flags & XE_VM_FLAG_MIGRATION)) {
1478 xe_device_mem_access_put(xe);
1479 xe_pm_runtime_put(xe);
1481 if (xe->info.has_asid) {
1482 mutex_lock(&xe->usm.lock);
1483 lookup = xa_erase(&xe->usm.asid_to_vm, vm->usm.asid);
1484 XE_WARN_ON(lookup != vm);
1485 mutex_unlock(&xe->usm.lock);
1490 * XXX: We delay destroying the PT root until the VM if freed as PT root
1491 * is needed for xe_vm_lock to work. If we remove that dependency this
1492 * can be moved to xe_vm_close_and_put.
1494 xe_vm_lock(vm, &ww, 0, false);
1495 for_each_tile(tile, xe, id) {
1496 if (vm->pt_root[id]) {
1497 xe_pt_destroy(vm->pt_root[id], vm->flags, NULL);
1498 vm->pt_root[id] = NULL;
1501 xe_vm_unlock(vm, &ww);
1503 trace_xe_vm_free(vm);
1504 dma_fence_put(vm->rebind_fence);
1505 dma_resv_fini(&vm->resv);
1509 void xe_vm_free(struct kref *ref)
1511 struct xe_vm *vm = container_of(ref, struct xe_vm, refcount);
1513 /* To destroy the VM we need to be able to sleep */
1514 queue_work(system_unbound_wq, &vm->destroy_work);
1517 struct xe_vm *xe_vm_lookup(struct xe_file *xef, u32 id)
1521 mutex_lock(&xef->vm.lock);
1522 vm = xa_load(&xef->vm.xa, id);
1523 mutex_unlock(&xef->vm.lock);
1531 u64 xe_vm_pdp4_descriptor(struct xe_vm *vm, struct xe_tile *tile)
1533 return gen8_pde_encode(vm->pt_root[tile->id]->bo, 0,
1537 static struct dma_fence *
1538 xe_vm_unbind_vma(struct xe_vma *vma, struct xe_engine *e,
1539 struct xe_sync_entry *syncs, u32 num_syncs)
1541 struct xe_tile *tile;
1542 struct dma_fence *fence = NULL;
1543 struct dma_fence **fences = NULL;
1544 struct dma_fence_array *cf = NULL;
1545 struct xe_vm *vm = vma->vm;
1546 int cur_fence = 0, i;
1547 int number_tiles = hweight_long(vma->tile_present);
1551 trace_xe_vma_unbind(vma);
1553 if (number_tiles > 1) {
1554 fences = kmalloc_array(number_tiles, sizeof(*fences),
1557 return ERR_PTR(-ENOMEM);
1560 for_each_tile(tile, vm->xe, id) {
1561 if (!(vma->tile_present & BIT(id)))
1564 fence = __xe_pt_unbind_vma(tile, vma, e, syncs, num_syncs);
1565 if (IS_ERR(fence)) {
1566 err = PTR_ERR(fence);
1571 fences[cur_fence++] = fence;
1574 if (e && vm->pt_root[id] && !list_empty(&e->multi_gt_list))
1575 e = list_next_entry(e, multi_gt_list);
1579 cf = dma_fence_array_create(number_tiles, fences,
1580 vm->composite_fence_ctx,
1581 vm->composite_fence_seqno++,
1584 --vm->composite_fence_seqno;
1590 for (i = 0; i < num_syncs; i++)
1591 xe_sync_entry_signal(&syncs[i], NULL, cf ? &cf->base : fence);
1593 return cf ? &cf->base : !fence ? dma_fence_get_stub() : fence;
1598 /* FIXME: Rewind the previous binds? */
1599 dma_fence_put(fences[--cur_fence]);
1604 return ERR_PTR(err);
1607 static struct dma_fence *
1608 xe_vm_bind_vma(struct xe_vma *vma, struct xe_engine *e,
1609 struct xe_sync_entry *syncs, u32 num_syncs)
1611 struct xe_tile *tile;
1612 struct dma_fence *fence;
1613 struct dma_fence **fences = NULL;
1614 struct dma_fence_array *cf = NULL;
1615 struct xe_vm *vm = vma->vm;
1616 int cur_fence = 0, i;
1617 int number_tiles = hweight_long(vma->tile_mask);
1621 trace_xe_vma_bind(vma);
1623 if (number_tiles > 1) {
1624 fences = kmalloc_array(number_tiles, sizeof(*fences),
1627 return ERR_PTR(-ENOMEM);
1630 for_each_tile(tile, vm->xe, id) {
1631 if (!(vma->tile_mask & BIT(id)))
1634 fence = __xe_pt_bind_vma(tile, vma, e, syncs, num_syncs,
1635 vma->tile_present & BIT(id));
1636 if (IS_ERR(fence)) {
1637 err = PTR_ERR(fence);
1642 fences[cur_fence++] = fence;
1645 if (e && vm->pt_root[id] && !list_empty(&e->multi_gt_list))
1646 e = list_next_entry(e, multi_gt_list);
1650 cf = dma_fence_array_create(number_tiles, fences,
1651 vm->composite_fence_ctx,
1652 vm->composite_fence_seqno++,
1655 --vm->composite_fence_seqno;
1661 for (i = 0; i < num_syncs; i++)
1662 xe_sync_entry_signal(&syncs[i], NULL, cf ? &cf->base : fence);
1664 return cf ? &cf->base : fence;
1669 /* FIXME: Rewind the previous binds? */
1670 dma_fence_put(fences[--cur_fence]);
1675 return ERR_PTR(err);
1678 struct async_op_fence {
1679 struct dma_fence fence;
1680 struct dma_fence *wait_fence;
1681 struct dma_fence_cb cb;
1683 wait_queue_head_t wq;
1687 static const char *async_op_fence_get_driver_name(struct dma_fence *dma_fence)
1693 async_op_fence_get_timeline_name(struct dma_fence *dma_fence)
1695 return "async_op_fence";
1698 static const struct dma_fence_ops async_op_fence_ops = {
1699 .get_driver_name = async_op_fence_get_driver_name,
1700 .get_timeline_name = async_op_fence_get_timeline_name,
1703 static void async_op_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
1705 struct async_op_fence *afence =
1706 container_of(cb, struct async_op_fence, cb);
1708 afence->fence.error = afence->wait_fence->error;
1709 dma_fence_signal(&afence->fence);
1710 xe_vm_put(afence->vm);
1711 dma_fence_put(afence->wait_fence);
1712 dma_fence_put(&afence->fence);
1715 static void add_async_op_fence_cb(struct xe_vm *vm,
1716 struct dma_fence *fence,
1717 struct async_op_fence *afence)
1721 if (!xe_vm_no_dma_fences(vm)) {
1722 afence->started = true;
1724 wake_up_all(&afence->wq);
1727 afence->wait_fence = dma_fence_get(fence);
1728 afence->vm = xe_vm_get(vm);
1729 dma_fence_get(&afence->fence);
1730 ret = dma_fence_add_callback(fence, &afence->cb, async_op_fence_cb);
1731 if (ret == -ENOENT) {
1732 afence->fence.error = afence->wait_fence->error;
1733 dma_fence_signal(&afence->fence);
1737 dma_fence_put(afence->wait_fence);
1738 dma_fence_put(&afence->fence);
1740 XE_WARN_ON(ret && ret != -ENOENT);
1743 int xe_vm_async_fence_wait_start(struct dma_fence *fence)
1745 if (fence->ops == &async_op_fence_ops) {
1746 struct async_op_fence *afence =
1747 container_of(fence, struct async_op_fence, fence);
1749 XE_BUG_ON(xe_vm_no_dma_fences(afence->vm));
1752 return wait_event_interruptible(afence->wq, afence->started);
1758 static int __xe_vm_bind(struct xe_vm *vm, struct xe_vma *vma,
1759 struct xe_engine *e, struct xe_sync_entry *syncs,
1760 u32 num_syncs, struct async_op_fence *afence)
1762 struct dma_fence *fence;
1764 xe_vm_assert_held(vm);
1766 fence = xe_vm_bind_vma(vma, e, syncs, num_syncs);
1768 return PTR_ERR(fence);
1770 add_async_op_fence_cb(vm, fence, afence);
1772 dma_fence_put(fence);
1776 static int xe_vm_bind(struct xe_vm *vm, struct xe_vma *vma, struct xe_engine *e,
1777 struct xe_bo *bo, struct xe_sync_entry *syncs,
1778 u32 num_syncs, struct async_op_fence *afence)
1782 xe_vm_assert_held(vm);
1783 xe_bo_assert_held(bo);
1786 err = xe_bo_validate(bo, vm, true);
1791 return __xe_vm_bind(vm, vma, e, syncs, num_syncs, afence);
1794 static int xe_vm_unbind(struct xe_vm *vm, struct xe_vma *vma,
1795 struct xe_engine *e, struct xe_sync_entry *syncs,
1796 u32 num_syncs, struct async_op_fence *afence)
1798 struct dma_fence *fence;
1800 xe_vm_assert_held(vm);
1801 xe_bo_assert_held(vma->bo);
1803 fence = xe_vm_unbind_vma(vma, e, syncs, num_syncs);
1805 return PTR_ERR(fence);
1807 add_async_op_fence_cb(vm, fence, afence);
1809 xe_vma_destroy(vma, fence);
1810 dma_fence_put(fence);
1815 static int vm_set_error_capture_address(struct xe_device *xe, struct xe_vm *vm,
1818 if (XE_IOCTL_ERR(xe, !value))
1821 if (XE_IOCTL_ERR(xe, !(vm->flags & XE_VM_FLAG_ASYNC_BIND_OPS)))
1824 if (XE_IOCTL_ERR(xe, vm->async_ops.error_capture.addr))
1827 vm->async_ops.error_capture.mm = current->mm;
1828 vm->async_ops.error_capture.addr = value;
1829 init_waitqueue_head(&vm->async_ops.error_capture.wq);
1834 typedef int (*xe_vm_set_property_fn)(struct xe_device *xe, struct xe_vm *vm,
1837 static const xe_vm_set_property_fn vm_set_property_funcs[] = {
1838 [XE_VM_PROPERTY_BIND_OP_ERROR_CAPTURE_ADDRESS] =
1839 vm_set_error_capture_address,
1842 static int vm_user_ext_set_property(struct xe_device *xe, struct xe_vm *vm,
1845 u64 __user *address = u64_to_user_ptr(extension);
1846 struct drm_xe_ext_vm_set_property ext;
1849 err = __copy_from_user(&ext, address, sizeof(ext));
1850 if (XE_IOCTL_ERR(xe, err))
1853 if (XE_IOCTL_ERR(xe, ext.property >=
1854 ARRAY_SIZE(vm_set_property_funcs)) ||
1855 XE_IOCTL_ERR(xe, ext.pad) ||
1856 XE_IOCTL_ERR(xe, ext.reserved[0] || ext.reserved[1]))
1859 return vm_set_property_funcs[ext.property](xe, vm, ext.value);
1862 typedef int (*xe_vm_user_extension_fn)(struct xe_device *xe, struct xe_vm *vm,
1865 static const xe_vm_set_property_fn vm_user_extension_funcs[] = {
1866 [XE_VM_EXTENSION_SET_PROPERTY] = vm_user_ext_set_property,
1869 #define MAX_USER_EXTENSIONS 16
1870 static int vm_user_extensions(struct xe_device *xe, struct xe_vm *vm,
1871 u64 extensions, int ext_number)
1873 u64 __user *address = u64_to_user_ptr(extensions);
1874 struct xe_user_extension ext;
1877 if (XE_IOCTL_ERR(xe, ext_number >= MAX_USER_EXTENSIONS))
1880 err = __copy_from_user(&ext, address, sizeof(ext));
1881 if (XE_IOCTL_ERR(xe, err))
1884 if (XE_IOCTL_ERR(xe, ext.pad) ||
1885 XE_IOCTL_ERR(xe, ext.name >=
1886 ARRAY_SIZE(vm_user_extension_funcs)))
1889 err = vm_user_extension_funcs[ext.name](xe, vm, extensions);
1890 if (XE_IOCTL_ERR(xe, err))
1893 if (ext.next_extension)
1894 return vm_user_extensions(xe, vm, ext.next_extension,
1900 #define ALL_DRM_XE_VM_CREATE_FLAGS (DRM_XE_VM_CREATE_SCRATCH_PAGE | \
1901 DRM_XE_VM_CREATE_COMPUTE_MODE | \
1902 DRM_XE_VM_CREATE_ASYNC_BIND_OPS | \
1903 DRM_XE_VM_CREATE_FAULT_MODE)
1905 int xe_vm_create_ioctl(struct drm_device *dev, void *data,
1906 struct drm_file *file)
1908 struct xe_device *xe = to_xe_device(dev);
1909 struct xe_file *xef = to_xe_file(file);
1910 struct drm_xe_vm_create *args = data;
1916 if (XE_IOCTL_ERR(xe, args->reserved[0] || args->reserved[1]))
1919 if (XE_IOCTL_ERR(xe, args->flags & ~ALL_DRM_XE_VM_CREATE_FLAGS))
1922 if (XE_IOCTL_ERR(xe, args->flags & DRM_XE_VM_CREATE_SCRATCH_PAGE &&
1923 args->flags & DRM_XE_VM_CREATE_FAULT_MODE))
1926 if (XE_IOCTL_ERR(xe, args->flags & DRM_XE_VM_CREATE_COMPUTE_MODE &&
1927 args->flags & DRM_XE_VM_CREATE_FAULT_MODE))
1930 if (XE_IOCTL_ERR(xe, args->flags & DRM_XE_VM_CREATE_FAULT_MODE &&
1931 xe_device_in_non_fault_mode(xe)))
1934 if (XE_IOCTL_ERR(xe, !(args->flags & DRM_XE_VM_CREATE_FAULT_MODE) &&
1935 xe_device_in_fault_mode(xe)))
1938 if (XE_IOCTL_ERR(xe, args->flags & DRM_XE_VM_CREATE_FAULT_MODE &&
1939 !xe->info.supports_usm))
1942 if (args->flags & DRM_XE_VM_CREATE_SCRATCH_PAGE)
1943 flags |= XE_VM_FLAG_SCRATCH_PAGE;
1944 if (args->flags & DRM_XE_VM_CREATE_COMPUTE_MODE)
1945 flags |= XE_VM_FLAG_COMPUTE_MODE;
1946 if (args->flags & DRM_XE_VM_CREATE_ASYNC_BIND_OPS)
1947 flags |= XE_VM_FLAG_ASYNC_BIND_OPS;
1948 if (args->flags & DRM_XE_VM_CREATE_FAULT_MODE)
1949 flags |= XE_VM_FLAG_FAULT_MODE;
1951 vm = xe_vm_create(xe, flags);
1955 if (args->extensions) {
1956 err = vm_user_extensions(xe, vm, args->extensions, 0);
1957 if (XE_IOCTL_ERR(xe, err)) {
1958 xe_vm_close_and_put(vm);
1963 mutex_lock(&xef->vm.lock);
1964 err = xa_alloc(&xef->vm.xa, &id, vm, xa_limit_32b, GFP_KERNEL);
1965 mutex_unlock(&xef->vm.lock);
1967 xe_vm_close_and_put(vm);
1971 if (xe->info.has_asid) {
1972 mutex_lock(&xe->usm.lock);
1973 err = xa_alloc_cyclic(&xe->usm.asid_to_vm, &asid, vm,
1974 XA_LIMIT(0, XE_MAX_ASID - 1),
1975 &xe->usm.next_asid, GFP_KERNEL);
1976 mutex_unlock(&xe->usm.lock);
1978 xe_vm_close_and_put(vm);
1981 vm->usm.asid = asid;
1986 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG_MEM)
1987 /* Warning: Security issue - never enable by default */
1988 args->reserved[0] = xe_bo_main_addr(vm->pt_root[0]->bo, XE_PAGE_SIZE);
1994 int xe_vm_destroy_ioctl(struct drm_device *dev, void *data,
1995 struct drm_file *file)
1997 struct xe_device *xe = to_xe_device(dev);
1998 struct xe_file *xef = to_xe_file(file);
1999 struct drm_xe_vm_destroy *args = data;
2002 if (XE_IOCTL_ERR(xe, args->pad) ||
2003 XE_IOCTL_ERR(xe, args->reserved[0] || args->reserved[1]))
2006 vm = xe_vm_lookup(xef, args->vm_id);
2007 if (XE_IOCTL_ERR(xe, !vm))
2011 /* FIXME: Extend this check to non-compute mode VMs */
2012 if (XE_IOCTL_ERR(xe, vm->preempt.num_engines))
2015 mutex_lock(&xef->vm.lock);
2016 xa_erase(&xef->vm.xa, args->vm_id);
2017 mutex_unlock(&xef->vm.lock);
2019 xe_vm_close_and_put(vm);
2024 static const u32 region_to_mem_type[] = {
2030 static int xe_vm_prefetch(struct xe_vm *vm, struct xe_vma *vma,
2031 struct xe_engine *e, u32 region,
2032 struct xe_sync_entry *syncs, u32 num_syncs,
2033 struct async_op_fence *afence)
2037 XE_BUG_ON(region > ARRAY_SIZE(region_to_mem_type));
2039 if (!xe_vma_is_userptr(vma)) {
2040 err = xe_bo_migrate(vma->bo, region_to_mem_type[region]);
2045 if (vma->tile_mask != (vma->tile_present & ~vma->usm.tile_invalidated)) {
2046 return xe_vm_bind(vm, vma, e, vma->bo, syncs, num_syncs,
2051 /* Nothing to do, signal fences now */
2052 for (i = 0; i < num_syncs; i++)
2053 xe_sync_entry_signal(&syncs[i], NULL,
2054 dma_fence_get_stub());
2056 dma_fence_signal(&afence->fence);
2061 #define VM_BIND_OP(op) (op & 0xffff)
2063 static int __vm_bind_ioctl(struct xe_vm *vm, struct xe_vma *vma,
2064 struct xe_engine *e, struct xe_bo *bo, u32 op,
2065 u32 region, struct xe_sync_entry *syncs,
2066 u32 num_syncs, struct async_op_fence *afence)
2068 switch (VM_BIND_OP(op)) {
2069 case XE_VM_BIND_OP_MAP:
2070 return xe_vm_bind(vm, vma, e, bo, syncs, num_syncs, afence);
2071 case XE_VM_BIND_OP_UNMAP:
2072 case XE_VM_BIND_OP_UNMAP_ALL:
2073 return xe_vm_unbind(vm, vma, e, syncs, num_syncs, afence);
2074 case XE_VM_BIND_OP_MAP_USERPTR:
2075 return xe_vm_bind(vm, vma, e, NULL, syncs, num_syncs, afence);
2076 case XE_VM_BIND_OP_PREFETCH:
2077 return xe_vm_prefetch(vm, vma, e, region, syncs, num_syncs,
2081 XE_BUG_ON("NOT POSSIBLE");
2086 struct ttm_buffer_object *xe_vm_ttm_bo(struct xe_vm *vm)
2088 int idx = vm->flags & XE_VM_FLAG_MIGRATION ?
2089 XE_VM_FLAG_GT_ID(vm->flags) : 0;
2091 /* Safe to use index 0 as all BO in the VM share a single dma-resv lock */
2092 return &vm->pt_root[idx]->bo->ttm;
2095 static void xe_vm_tv_populate(struct xe_vm *vm, struct ttm_validate_buffer *tv)
2098 tv->bo = xe_vm_ttm_bo(vm);
2101 static bool is_map_op(u32 op)
2103 return VM_BIND_OP(op) == XE_VM_BIND_OP_MAP ||
2104 VM_BIND_OP(op) == XE_VM_BIND_OP_MAP_USERPTR;
2107 static bool is_unmap_op(u32 op)
2109 return VM_BIND_OP(op) == XE_VM_BIND_OP_UNMAP ||
2110 VM_BIND_OP(op) == XE_VM_BIND_OP_UNMAP_ALL;
2113 static int vm_bind_ioctl(struct xe_vm *vm, struct xe_vma *vma,
2114 struct xe_engine *e, struct xe_bo *bo,
2115 struct drm_xe_vm_bind_op *bind_op,
2116 struct xe_sync_entry *syncs, u32 num_syncs,
2117 struct async_op_fence *afence)
2121 struct ttm_validate_buffer tv_bo, tv_vm;
2122 struct ww_acquire_ctx ww;
2126 lockdep_assert_held(&vm->lock);
2127 XE_BUG_ON(!list_empty(&vma->unbind_link));
2129 /* Binds deferred to faults, signal fences now */
2130 if (xe_vm_in_fault_mode(vm) && is_map_op(bind_op->op) &&
2131 !(bind_op->op & XE_VM_BIND_FLAG_IMMEDIATE)) {
2132 for (i = 0; i < num_syncs; i++)
2133 xe_sync_entry_signal(&syncs[i], NULL,
2134 dma_fence_get_stub());
2136 dma_fence_signal(&afence->fence);
2140 xe_vm_tv_populate(vm, &tv_vm);
2141 list_add_tail(&tv_vm.head, &objs);
2145 * An unbind can drop the last reference to the BO and
2146 * the BO is needed for ttm_eu_backoff_reservation so
2147 * take a reference here.
2151 tv_bo.bo = &vbo->ttm;
2152 tv_bo.num_shared = 1;
2153 list_add(&tv_bo.head, &objs);
2157 err = ttm_eu_reserve_buffers(&ww, &objs, true, &dups);
2159 err = __vm_bind_ioctl(vm, vma, e, bo,
2160 bind_op->op, bind_op->region, syncs,
2162 ttm_eu_backoff_reservation(&ww, &objs);
2163 if (err == -EAGAIN && xe_vma_is_userptr(vma)) {
2164 lockdep_assert_held_write(&vm->lock);
2165 err = xe_vma_userptr_pin_pages(vma);
2177 struct xe_engine *engine;
2179 struct drm_xe_vm_bind_op bind_op;
2180 struct xe_sync_entry *syncs;
2182 struct list_head link;
2183 struct async_op_fence *fence;
2186 static void async_op_cleanup(struct xe_vm *vm, struct async_op *op)
2188 while (op->num_syncs--)
2189 xe_sync_entry_cleanup(&op->syncs[op->num_syncs]);
2193 xe_engine_put(op->engine);
2196 dma_fence_put(&op->fence->fence);
2200 static struct async_op *next_async_op(struct xe_vm *vm)
2202 return list_first_entry_or_null(&vm->async_ops.pending,
2203 struct async_op, link);
2206 static void vm_set_async_error(struct xe_vm *vm, int err)
2208 lockdep_assert_held(&vm->lock);
2209 vm->async_ops.error = err;
2212 static void async_op_work_func(struct work_struct *w)
2214 struct xe_vm *vm = container_of(w, struct xe_vm, async_ops.work);
2217 struct async_op *op;
2220 if (vm->async_ops.error && !xe_vm_is_closed(vm))
2223 spin_lock_irq(&vm->async_ops.lock);
2224 op = next_async_op(vm);
2226 list_del_init(&op->link);
2227 spin_unlock_irq(&vm->async_ops.lock);
2232 if (!xe_vm_is_closed(vm)) {
2235 down_write(&vm->lock);
2237 first = op->vma->first_munmap_rebind;
2238 last = op->vma->last_munmap_rebind;
2239 #ifdef TEST_VM_ASYNC_OPS_ERROR
2240 #define FORCE_ASYNC_OP_ERROR BIT(31)
2241 if (!(op->bind_op.op & FORCE_ASYNC_OP_ERROR)) {
2242 err = vm_bind_ioctl(vm, op->vma, op->engine,
2243 op->bo, &op->bind_op,
2244 op->syncs, op->num_syncs,
2248 op->bind_op.op &= ~FORCE_ASYNC_OP_ERROR;
2251 err = vm_bind_ioctl(vm, op->vma, op->engine, op->bo,
2252 &op->bind_op, op->syncs,
2253 op->num_syncs, op->fence);
2256 * In order for the fencing to work (stall behind
2257 * existing jobs / prevent new jobs from running) all
2258 * the dma-resv slots need to be programmed in a batch
2259 * relative to execs / the rebind worker. The vm->lock
2262 if (!err && ((first && VM_BIND_OP(op->bind_op.op) ==
2263 XE_VM_BIND_OP_UNMAP) ||
2264 vm->async_ops.munmap_rebind_inflight)) {
2266 op->vma->last_munmap_rebind = false;
2267 vm->async_ops.munmap_rebind_inflight =
2270 vm->async_ops.munmap_rebind_inflight =
2273 async_op_cleanup(vm, op);
2275 spin_lock_irq(&vm->async_ops.lock);
2276 op = next_async_op(vm);
2278 list_del_init(&op->link);
2279 spin_unlock_irq(&vm->async_ops.lock);
2285 trace_xe_vma_fail(op->vma);
2286 drm_warn(&vm->xe->drm, "Async VM op(%d) failed with %d",
2287 VM_BIND_OP(op->bind_op.op),
2290 spin_lock_irq(&vm->async_ops.lock);
2291 list_add(&op->link, &vm->async_ops.pending);
2292 spin_unlock_irq(&vm->async_ops.lock);
2294 vm_set_async_error(vm, err);
2295 up_write(&vm->lock);
2297 if (vm->async_ops.error_capture.addr)
2298 vm_error_capture(vm, err,
2304 up_write(&vm->lock);
2306 trace_xe_vma_flush(op->vma);
2308 if (is_unmap_op(op->bind_op.op)) {
2309 down_write(&vm->lock);
2310 xe_vma_destroy_unlocked(op->vma);
2311 up_write(&vm->lock);
2314 if (op->fence && !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
2315 &op->fence->fence.flags)) {
2316 if (!xe_vm_no_dma_fences(vm)) {
2317 op->fence->started = true;
2319 wake_up_all(&op->fence->wq);
2321 dma_fence_signal(&op->fence->fence);
2325 async_op_cleanup(vm, op);
2329 static int __vm_bind_ioctl_async(struct xe_vm *vm, struct xe_vma *vma,
2330 struct xe_engine *e, struct xe_bo *bo,
2331 struct drm_xe_vm_bind_op *bind_op,
2332 struct xe_sync_entry *syncs, u32 num_syncs)
2334 struct async_op *op;
2335 bool installed = false;
2339 lockdep_assert_held(&vm->lock);
2341 op = kmalloc(sizeof(*op), GFP_KERNEL);
2347 op->fence = kmalloc(sizeof(*op->fence), GFP_KERNEL);
2353 seqno = e ? ++e->bind.fence_seqno : ++vm->async_ops.fence.seqno;
2354 dma_fence_init(&op->fence->fence, &async_op_fence_ops,
2355 &vm->async_ops.lock, e ? e->bind.fence_ctx :
2356 vm->async_ops.fence.context, seqno);
2358 if (!xe_vm_no_dma_fences(vm)) {
2360 op->fence->started = false;
2361 init_waitqueue_head(&op->fence->wq);
2369 op->bind_op = *bind_op;
2371 op->num_syncs = num_syncs;
2372 INIT_LIST_HEAD(&op->link);
2374 for (i = 0; i < num_syncs; i++)
2375 installed |= xe_sync_entry_signal(&syncs[i], NULL,
2378 if (!installed && op->fence)
2379 dma_fence_signal(&op->fence->fence);
2381 spin_lock_irq(&vm->async_ops.lock);
2382 list_add_tail(&op->link, &vm->async_ops.pending);
2383 spin_unlock_irq(&vm->async_ops.lock);
2385 if (!vm->async_ops.error)
2386 queue_work(system_unbound_wq, &vm->async_ops.work);
2391 static int vm_bind_ioctl_async(struct xe_vm *vm, struct xe_vma *vma,
2392 struct xe_engine *e, struct xe_bo *bo,
2393 struct drm_xe_vm_bind_op *bind_op,
2394 struct xe_sync_entry *syncs, u32 num_syncs)
2396 struct xe_vma *__vma, *next;
2397 struct list_head rebind_list;
2398 struct xe_sync_entry *in_syncs = NULL, *out_syncs = NULL;
2399 u32 num_in_syncs = 0, num_out_syncs = 0;
2400 bool first = true, last;
2404 lockdep_assert_held(&vm->lock);
2406 /* Not a linked list of unbinds + rebinds, easy */
2407 if (list_empty(&vma->unbind_link))
2408 return __vm_bind_ioctl_async(vm, vma, e, bo, bind_op,
2412 * Linked list of unbinds + rebinds, decompose syncs into 'in / out'
2413 * passing the 'in' to the first operation and 'out' to the last. Also
2414 * the reference counting is a little tricky, increment the VM / bind
2415 * engine ref count on all but the last operation and increment the BOs
2416 * ref count on each rebind.
2419 XE_BUG_ON(VM_BIND_OP(bind_op->op) != XE_VM_BIND_OP_UNMAP &&
2420 VM_BIND_OP(bind_op->op) != XE_VM_BIND_OP_UNMAP_ALL &&
2421 VM_BIND_OP(bind_op->op) != XE_VM_BIND_OP_PREFETCH);
2423 /* Decompose syncs */
2425 in_syncs = kmalloc(sizeof(*in_syncs) * num_syncs, GFP_KERNEL);
2426 out_syncs = kmalloc(sizeof(*out_syncs) * num_syncs, GFP_KERNEL);
2427 if (!in_syncs || !out_syncs) {
2432 for (i = 0; i < num_syncs; ++i) {
2433 bool signal = syncs[i].flags & DRM_XE_SYNC_SIGNAL;
2436 out_syncs[num_out_syncs++] = syncs[i];
2438 in_syncs[num_in_syncs++] = syncs[i];
2442 /* Do unbinds + move rebinds to new list */
2443 INIT_LIST_HEAD(&rebind_list);
2444 list_for_each_entry_safe(__vma, next, &vma->unbind_link, unbind_link) {
2445 if (__vma->destroyed ||
2446 VM_BIND_OP(bind_op->op) == XE_VM_BIND_OP_PREFETCH) {
2447 list_del_init(&__vma->unbind_link);
2449 err = __vm_bind_ioctl_async(xe_vm_get(vm), __vma,
2450 e ? xe_engine_get(e) : NULL,
2451 bo, bind_op, first ?
2453 first ? num_in_syncs : 0);
2464 list_move_tail(&__vma->unbind_link, &rebind_list);
2467 last = list_empty(&rebind_list);
2473 err = __vm_bind_ioctl_async(vm, vma, e,
2476 last ? out_syncs : NULL,
2477 first ? num_in_syncs :
2478 last ? num_out_syncs : 0);
2490 list_for_each_entry_safe(__vma, next, &rebind_list, unbind_link) {
2491 list_del_init(&__vma->unbind_link);
2492 last = list_empty(&rebind_list);
2494 if (xe_vma_is_userptr(__vma)) {
2495 bind_op->op = XE_VM_BIND_FLAG_ASYNC |
2496 XE_VM_BIND_OP_MAP_USERPTR;
2498 bind_op->op = XE_VM_BIND_FLAG_ASYNC |
2500 xe_bo_get(__vma->bo);
2509 err = __vm_bind_ioctl_async(vm, __vma, e,
2510 __vma->bo, bind_op, last ?
2512 last ? num_out_syncs : 0);
2534 static int __vm_bind_ioctl_lookup_vma(struct xe_vm *vm, struct xe_bo *bo,
2535 u64 addr, u64 range, u32 op)
2537 struct xe_device *xe = vm->xe;
2538 struct xe_vma *vma, lookup;
2539 bool async = !!(op & XE_VM_BIND_FLAG_ASYNC);
2541 lockdep_assert_held(&vm->lock);
2543 lookup.start = addr;
2544 lookup.end = addr + range - 1;
2546 switch (VM_BIND_OP(op)) {
2547 case XE_VM_BIND_OP_MAP:
2548 case XE_VM_BIND_OP_MAP_USERPTR:
2549 vma = xe_vm_find_overlapping_vma(vm, &lookup);
2550 if (XE_IOCTL_ERR(xe, vma))
2553 case XE_VM_BIND_OP_UNMAP:
2554 case XE_VM_BIND_OP_PREFETCH:
2555 vma = xe_vm_find_overlapping_vma(vm, &lookup);
2556 if (XE_IOCTL_ERR(xe, !vma) ||
2557 XE_IOCTL_ERR(xe, (vma->start != addr ||
2558 vma->end != addr + range - 1) && !async))
2561 case XE_VM_BIND_OP_UNMAP_ALL:
2564 XE_BUG_ON("NOT POSSIBLE");
2571 static void prep_vma_destroy(struct xe_vm *vm, struct xe_vma *vma)
2573 down_read(&vm->userptr.notifier_lock);
2574 vma->destroyed = true;
2575 up_read(&vm->userptr.notifier_lock);
2576 xe_vm_remove_vma(vm, vma);
2579 static int prep_replacement_vma(struct xe_vm *vm, struct xe_vma *vma)
2583 if (vma->bo && !vma->bo->vm) {
2584 vm_insert_extobj(vm, vma);
2585 err = add_preempt_fences(vm, vma->bo);
2594 * Find all overlapping VMAs in lookup range and add to a list in the returned
2595 * VMA, all of VMAs found will be unbound. Also possibly add 2 new VMAs that
2596 * need to be bound if first / last VMAs are not fully unbound. This is akin to
2599 static struct xe_vma *vm_unbind_lookup_vmas(struct xe_vm *vm,
2600 struct xe_vma *lookup)
2602 struct xe_vma *vma = xe_vm_find_overlapping_vma(vm, lookup);
2603 struct rb_node *node;
2604 struct xe_vma *first = vma, *last = vma, *new_first = NULL,
2605 *new_last = NULL, *__vma, *next;
2607 bool first_munmap_rebind = false;
2609 lockdep_assert_held(&vm->lock);
2612 node = &vma->vm_node;
2613 while ((node = rb_next(node))) {
2614 if (!xe_vma_cmp_vma_cb(lookup, node)) {
2615 __vma = to_xe_vma(node);
2616 list_add_tail(&__vma->unbind_link, &vma->unbind_link);
2623 node = &vma->vm_node;
2624 while ((node = rb_prev(node))) {
2625 if (!xe_vma_cmp_vma_cb(lookup, node)) {
2626 __vma = to_xe_vma(node);
2627 list_add(&__vma->unbind_link, &vma->unbind_link);
2634 if (first->start != lookup->start) {
2635 struct ww_acquire_ctx ww;
2638 err = xe_bo_lock(first->bo, &ww, 0, true);
2641 new_first = xe_vma_create(first->vm, first->bo,
2642 first->bo ? first->bo_offset :
2646 (first->pte_flags & XE_PTE_READ_ONLY),
2649 xe_bo_unlock(first->bo, &ww);
2655 err = xe_vma_userptr_pin_pages(new_first);
2659 err = prep_replacement_vma(vm, new_first);
2664 if (last->end != lookup->end) {
2665 struct ww_acquire_ctx ww;
2666 u64 chunk = lookup->end + 1 - last->start;
2669 err = xe_bo_lock(last->bo, &ww, 0, true);
2672 new_last = xe_vma_create(last->vm, last->bo,
2673 last->bo ? last->bo_offset + chunk :
2674 last->userptr.ptr + chunk,
2675 last->start + chunk,
2677 (last->pte_flags & XE_PTE_READ_ONLY),
2680 xe_bo_unlock(last->bo, &ww);
2686 err = xe_vma_userptr_pin_pages(new_last);
2690 err = prep_replacement_vma(vm, new_last);
2695 prep_vma_destroy(vm, vma);
2696 if (list_empty(&vma->unbind_link) && (new_first || new_last))
2697 vma->first_munmap_rebind = true;
2698 list_for_each_entry(__vma, &vma->unbind_link, unbind_link) {
2699 if ((new_first || new_last) && !first_munmap_rebind) {
2700 __vma->first_munmap_rebind = true;
2701 first_munmap_rebind = true;
2703 prep_vma_destroy(vm, __vma);
2706 xe_vm_insert_vma(vm, new_first);
2707 list_add_tail(&new_first->unbind_link, &vma->unbind_link);
2709 new_first->last_munmap_rebind = true;
2712 xe_vm_insert_vma(vm, new_last);
2713 list_add_tail(&new_last->unbind_link, &vma->unbind_link);
2714 new_last->last_munmap_rebind = true;
2720 list_for_each_entry_safe(__vma, next, &vma->unbind_link, unbind_link)
2721 list_del_init(&__vma->unbind_link);
2723 prep_vma_destroy(vm, new_last);
2724 xe_vma_destroy_unlocked(new_last);
2727 prep_vma_destroy(vm, new_first);
2728 xe_vma_destroy_unlocked(new_first);
2731 return ERR_PTR(err);
2735 * Similar to vm_unbind_lookup_vmas, find all VMAs in lookup range to prefetch
2737 static struct xe_vma *vm_prefetch_lookup_vmas(struct xe_vm *vm,
2738 struct xe_vma *lookup,
2741 struct xe_vma *vma = xe_vm_find_overlapping_vma(vm, lookup), *__vma,
2743 struct rb_node *node;
2745 if (!xe_vma_is_userptr(vma)) {
2746 if (!xe_bo_can_migrate(vma->bo, region_to_mem_type[region]))
2747 return ERR_PTR(-EINVAL);
2750 node = &vma->vm_node;
2751 while ((node = rb_next(node))) {
2752 if (!xe_vma_cmp_vma_cb(lookup, node)) {
2753 __vma = to_xe_vma(node);
2754 if (!xe_vma_is_userptr(__vma)) {
2755 if (!xe_bo_can_migrate(__vma->bo, region_to_mem_type[region]))
2758 list_add_tail(&__vma->unbind_link, &vma->unbind_link);
2764 node = &vma->vm_node;
2765 while ((node = rb_prev(node))) {
2766 if (!xe_vma_cmp_vma_cb(lookup, node)) {
2767 __vma = to_xe_vma(node);
2768 if (!xe_vma_is_userptr(__vma)) {
2769 if (!xe_bo_can_migrate(__vma->bo, region_to_mem_type[region]))
2772 list_add(&__vma->unbind_link, &vma->unbind_link);
2781 list_for_each_entry_safe(__vma, next, &vma->unbind_link,
2783 list_del_init(&__vma->unbind_link);
2785 return ERR_PTR(-EINVAL);
2788 static struct xe_vma *vm_unbind_all_lookup_vmas(struct xe_vm *vm,
2791 struct xe_vma *first = NULL, *vma;
2793 lockdep_assert_held(&vm->lock);
2794 xe_bo_assert_held(bo);
2796 list_for_each_entry(vma, &bo->vmas, bo_link) {
2800 prep_vma_destroy(vm, vma);
2804 list_add_tail(&vma->unbind_link, &first->unbind_link);
2810 static struct xe_vma *vm_bind_ioctl_lookup_vma(struct xe_vm *vm,
2812 u64 bo_offset_or_userptr,
2813 u64 addr, u64 range, u32 op,
2814 u64 tile_mask, u32 region)
2816 struct ww_acquire_ctx ww;
2817 struct xe_vma *vma, lookup;
2820 lockdep_assert_held(&vm->lock);
2822 lookup.start = addr;
2823 lookup.end = addr + range - 1;
2825 switch (VM_BIND_OP(op)) {
2826 case XE_VM_BIND_OP_MAP:
2829 err = xe_bo_lock(bo, &ww, 0, true);
2831 return ERR_PTR(err);
2832 vma = xe_vma_create(vm, bo, bo_offset_or_userptr, addr,
2834 op & XE_VM_BIND_FLAG_READONLY,
2836 xe_bo_unlock(bo, &ww);
2838 return ERR_PTR(-ENOMEM);
2840 xe_vm_insert_vma(vm, vma);
2842 vm_insert_extobj(vm, vma);
2843 err = add_preempt_fences(vm, bo);
2845 prep_vma_destroy(vm, vma);
2846 xe_vma_destroy_unlocked(vma);
2848 return ERR_PTR(err);
2852 case XE_VM_BIND_OP_UNMAP:
2853 vma = vm_unbind_lookup_vmas(vm, &lookup);
2855 case XE_VM_BIND_OP_PREFETCH:
2856 vma = vm_prefetch_lookup_vmas(vm, &lookup, region);
2858 case XE_VM_BIND_OP_UNMAP_ALL:
2861 err = xe_bo_lock(bo, &ww, 0, true);
2863 return ERR_PTR(err);
2864 vma = vm_unbind_all_lookup_vmas(vm, bo);
2866 vma = ERR_PTR(-EINVAL);
2867 xe_bo_unlock(bo, &ww);
2869 case XE_VM_BIND_OP_MAP_USERPTR:
2872 vma = xe_vma_create(vm, NULL, bo_offset_or_userptr, addr,
2874 op & XE_VM_BIND_FLAG_READONLY,
2877 return ERR_PTR(-ENOMEM);
2879 err = xe_vma_userptr_pin_pages(vma);
2881 prep_vma_destroy(vm, vma);
2882 xe_vma_destroy_unlocked(vma);
2884 return ERR_PTR(err);
2886 xe_vm_insert_vma(vm, vma);
2890 XE_BUG_ON("NOT POSSIBLE");
2891 vma = ERR_PTR(-EINVAL);
2897 #ifdef TEST_VM_ASYNC_OPS_ERROR
2898 #define SUPPORTED_FLAGS \
2899 (FORCE_ASYNC_OP_ERROR | XE_VM_BIND_FLAG_ASYNC | \
2900 XE_VM_BIND_FLAG_READONLY | XE_VM_BIND_FLAG_IMMEDIATE | 0xffff)
2902 #define SUPPORTED_FLAGS \
2903 (XE_VM_BIND_FLAG_ASYNC | XE_VM_BIND_FLAG_READONLY | \
2904 XE_VM_BIND_FLAG_IMMEDIATE | 0xffff)
2906 #define XE_64K_PAGE_MASK 0xffffull
2908 #define MAX_BINDS 512 /* FIXME: Picking random upper limit */
2910 static int vm_bind_ioctl_check_args(struct xe_device *xe,
2911 struct drm_xe_vm_bind *args,
2912 struct drm_xe_vm_bind_op **bind_ops,
2918 if (XE_IOCTL_ERR(xe, args->extensions) ||
2919 XE_IOCTL_ERR(xe, args->pad || args->pad2) ||
2920 XE_IOCTL_ERR(xe, args->reserved[0] || args->reserved[1]) ||
2921 XE_IOCTL_ERR(xe, !args->num_binds) ||
2922 XE_IOCTL_ERR(xe, args->num_binds > MAX_BINDS))
2925 if (args->num_binds > 1) {
2926 u64 __user *bind_user =
2927 u64_to_user_ptr(args->vector_of_binds);
2929 *bind_ops = kmalloc(sizeof(struct drm_xe_vm_bind_op) *
2930 args->num_binds, GFP_KERNEL);
2934 err = __copy_from_user(*bind_ops, bind_user,
2935 sizeof(struct drm_xe_vm_bind_op) *
2937 if (XE_IOCTL_ERR(xe, err)) {
2942 *bind_ops = &args->bind;
2945 for (i = 0; i < args->num_binds; ++i) {
2946 u64 range = (*bind_ops)[i].range;
2947 u64 addr = (*bind_ops)[i].addr;
2948 u32 op = (*bind_ops)[i].op;
2949 u32 obj = (*bind_ops)[i].obj;
2950 u64 obj_offset = (*bind_ops)[i].obj_offset;
2951 u32 region = (*bind_ops)[i].region;
2953 if (XE_IOCTL_ERR(xe, (*bind_ops)[i].pad) ||
2954 XE_IOCTL_ERR(xe, (*bind_ops)[i].reserved[0] ||
2955 (*bind_ops)[i].reserved[1])) {
2961 *async = !!(op & XE_VM_BIND_FLAG_ASYNC);
2962 } else if (XE_IOCTL_ERR(xe, !*async) ||
2963 XE_IOCTL_ERR(xe, !(op & XE_VM_BIND_FLAG_ASYNC)) ||
2964 XE_IOCTL_ERR(xe, VM_BIND_OP(op) ==
2965 XE_VM_BIND_OP_RESTART)) {
2970 if (XE_IOCTL_ERR(xe, !*async &&
2971 VM_BIND_OP(op) == XE_VM_BIND_OP_UNMAP_ALL)) {
2976 if (XE_IOCTL_ERR(xe, !*async &&
2977 VM_BIND_OP(op) == XE_VM_BIND_OP_PREFETCH)) {
2982 if (XE_IOCTL_ERR(xe, VM_BIND_OP(op) >
2983 XE_VM_BIND_OP_PREFETCH) ||
2984 XE_IOCTL_ERR(xe, op & ~SUPPORTED_FLAGS) ||
2985 XE_IOCTL_ERR(xe, !obj &&
2986 VM_BIND_OP(op) == XE_VM_BIND_OP_MAP) ||
2987 XE_IOCTL_ERR(xe, !obj &&
2988 VM_BIND_OP(op) == XE_VM_BIND_OP_UNMAP_ALL) ||
2989 XE_IOCTL_ERR(xe, addr &&
2990 VM_BIND_OP(op) == XE_VM_BIND_OP_UNMAP_ALL) ||
2991 XE_IOCTL_ERR(xe, range &&
2992 VM_BIND_OP(op) == XE_VM_BIND_OP_UNMAP_ALL) ||
2993 XE_IOCTL_ERR(xe, obj &&
2994 VM_BIND_OP(op) == XE_VM_BIND_OP_MAP_USERPTR) ||
2995 XE_IOCTL_ERR(xe, obj &&
2996 VM_BIND_OP(op) == XE_VM_BIND_OP_PREFETCH) ||
2997 XE_IOCTL_ERR(xe, region &&
2998 VM_BIND_OP(op) != XE_VM_BIND_OP_PREFETCH) ||
2999 XE_IOCTL_ERR(xe, !(BIT(region) &
3000 xe->info.mem_region_mask)) ||
3001 XE_IOCTL_ERR(xe, obj &&
3002 VM_BIND_OP(op) == XE_VM_BIND_OP_UNMAP)) {
3007 if (XE_IOCTL_ERR(xe, obj_offset & ~PAGE_MASK) ||
3008 XE_IOCTL_ERR(xe, addr & ~PAGE_MASK) ||
3009 XE_IOCTL_ERR(xe, range & ~PAGE_MASK) ||
3010 XE_IOCTL_ERR(xe, !range && VM_BIND_OP(op) !=
3011 XE_VM_BIND_OP_RESTART &&
3012 VM_BIND_OP(op) != XE_VM_BIND_OP_UNMAP_ALL)) {
3021 if (args->num_binds > 1)
3026 int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
3028 struct xe_device *xe = to_xe_device(dev);
3029 struct xe_file *xef = to_xe_file(file);
3030 struct drm_xe_vm_bind *args = data;
3031 struct drm_xe_sync __user *syncs_user;
3032 struct xe_bo **bos = NULL;
3033 struct xe_vma **vmas = NULL;
3035 struct xe_engine *e = NULL;
3037 struct xe_sync_entry *syncs = NULL;
3038 struct drm_xe_vm_bind_op *bind_ops;
3043 err = vm_bind_ioctl_check_args(xe, args, &bind_ops, &async);
3047 vm = xe_vm_lookup(xef, args->vm_id);
3048 if (XE_IOCTL_ERR(xe, !vm)) {
3053 if (XE_IOCTL_ERR(xe, xe_vm_is_closed(vm))) {
3054 drm_err(dev, "VM closed while we began looking up?\n");
3059 if (args->engine_id) {
3060 e = xe_engine_lookup(xef, args->engine_id);
3061 if (XE_IOCTL_ERR(xe, !e)) {
3065 if (XE_IOCTL_ERR(xe, !(e->flags & ENGINE_FLAG_VM))) {
3071 if (VM_BIND_OP(bind_ops[0].op) == XE_VM_BIND_OP_RESTART) {
3072 if (XE_IOCTL_ERR(xe, !(vm->flags & XE_VM_FLAG_ASYNC_BIND_OPS)))
3074 if (XE_IOCTL_ERR(xe, !err && args->num_syncs))
3076 if (XE_IOCTL_ERR(xe, !err && !vm->async_ops.error))
3080 down_write(&vm->lock);
3081 trace_xe_vm_restart(vm);
3082 vm_set_async_error(vm, 0);
3083 up_write(&vm->lock);
3085 queue_work(system_unbound_wq, &vm->async_ops.work);
3087 /* Rebinds may have been blocked, give worker a kick */
3088 if (xe_vm_in_compute_mode(vm))
3089 queue_work(vm->xe->ordered_wq,
3090 &vm->preempt.rebind_work);
3096 if (XE_IOCTL_ERR(xe, !vm->async_ops.error &&
3097 async != !!(vm->flags & XE_VM_FLAG_ASYNC_BIND_OPS))) {
3102 for (i = 0; i < args->num_binds; ++i) {
3103 u64 range = bind_ops[i].range;
3104 u64 addr = bind_ops[i].addr;
3106 if (XE_IOCTL_ERR(xe, range > vm->size) ||
3107 XE_IOCTL_ERR(xe, addr > vm->size - range)) {
3112 if (bind_ops[i].tile_mask) {
3113 u64 valid_tiles = BIT(xe->info.tile_count) - 1;
3115 if (XE_IOCTL_ERR(xe, bind_ops[i].tile_mask &
3123 bos = kzalloc(sizeof(*bos) * args->num_binds, GFP_KERNEL);
3129 vmas = kzalloc(sizeof(*vmas) * args->num_binds, GFP_KERNEL);
3135 for (i = 0; i < args->num_binds; ++i) {
3136 struct drm_gem_object *gem_obj;
3137 u64 range = bind_ops[i].range;
3138 u64 addr = bind_ops[i].addr;
3139 u32 obj = bind_ops[i].obj;
3140 u64 obj_offset = bind_ops[i].obj_offset;
3145 gem_obj = drm_gem_object_lookup(file, obj);
3146 if (XE_IOCTL_ERR(xe, !gem_obj)) {
3150 bos[i] = gem_to_xe_bo(gem_obj);
3152 if (XE_IOCTL_ERR(xe, range > bos[i]->size) ||
3153 XE_IOCTL_ERR(xe, obj_offset >
3154 bos[i]->size - range)) {
3159 if (bos[i]->flags & XE_BO_INTERNAL_64K) {
3160 if (XE_IOCTL_ERR(xe, obj_offset &
3161 XE_64K_PAGE_MASK) ||
3162 XE_IOCTL_ERR(xe, addr & XE_64K_PAGE_MASK) ||
3163 XE_IOCTL_ERR(xe, range & XE_64K_PAGE_MASK)) {
3170 if (args->num_syncs) {
3171 syncs = kcalloc(args->num_syncs, sizeof(*syncs), GFP_KERNEL);
3178 syncs_user = u64_to_user_ptr(args->syncs);
3179 for (num_syncs = 0; num_syncs < args->num_syncs; num_syncs++) {
3180 err = xe_sync_entry_parse(xe, xef, &syncs[num_syncs],
3181 &syncs_user[num_syncs], false,
3182 xe_vm_in_fault_mode(vm));
3187 err = down_write_killable(&vm->lock);
3191 /* Do some error checking first to make the unwind easier */
3192 for (i = 0; i < args->num_binds; ++i) {
3193 u64 range = bind_ops[i].range;
3194 u64 addr = bind_ops[i].addr;
3195 u32 op = bind_ops[i].op;
3197 err = __vm_bind_ioctl_lookup_vma(vm, bos[i], addr, range, op);
3199 goto release_vm_lock;
3202 for (i = 0; i < args->num_binds; ++i) {
3203 u64 range = bind_ops[i].range;
3204 u64 addr = bind_ops[i].addr;
3205 u32 op = bind_ops[i].op;
3206 u64 obj_offset = bind_ops[i].obj_offset;
3207 u64 tile_mask = bind_ops[i].tile_mask;
3208 u32 region = bind_ops[i].region;
3210 vmas[i] = vm_bind_ioctl_lookup_vma(vm, bos[i], obj_offset,
3211 addr, range, op, tile_mask,
3213 if (IS_ERR(vmas[i])) {
3214 err = PTR_ERR(vmas[i]);
3220 for (j = 0; j < args->num_binds; ++j) {
3221 struct xe_sync_entry *__syncs;
3222 u32 __num_syncs = 0;
3223 bool first_or_last = j == 0 || j == args->num_binds - 1;
3225 if (args->num_binds == 1) {
3226 __num_syncs = num_syncs;
3228 } else if (first_or_last && num_syncs) {
3229 bool first = j == 0;
3231 __syncs = kmalloc(sizeof(*__syncs) * num_syncs,
3238 /* in-syncs on first bind, out-syncs on last bind */
3239 for (i = 0; i < num_syncs; ++i) {
3240 bool signal = syncs[i].flags &
3243 if ((first && !signal) || (!first && signal))
3244 __syncs[__num_syncs++] = syncs[i];
3252 bool last = j == args->num_binds - 1;
3255 * Each pass of async worker drops the ref, take a ref
3256 * here, 1 set of refs taken above
3264 err = vm_bind_ioctl_async(vm, vmas[j], e, bos[j],
3265 bind_ops + j, __syncs,
3275 XE_BUG_ON(j != 0); /* Not supported */
3276 err = vm_bind_ioctl(vm, vmas[j], e, bos[j],
3277 bind_ops + j, __syncs,
3279 break; /* Needed so cleanup loops work */
3283 /* Most of cleanup owned by the async bind worker */
3284 if (async && !err) {
3285 up_write(&vm->lock);
3286 if (args->num_binds > 1)
3292 for (i = j; err && i < args->num_binds; ++i) {
3293 u32 op = bind_ops[i].op;
3294 struct xe_vma *vma, *next;
3299 list_for_each_entry_safe(vma, next, &vma->unbind_link,
3301 list_del_init(&vma->unbind_link);
3302 if (!vma->destroyed) {
3303 prep_vma_destroy(vm, vma);
3304 xe_vma_destroy_unlocked(vma);
3308 switch (VM_BIND_OP(op)) {
3309 case XE_VM_BIND_OP_MAP:
3310 prep_vma_destroy(vm, vmas[i]);
3311 xe_vma_destroy_unlocked(vmas[i]);
3313 case XE_VM_BIND_OP_MAP_USERPTR:
3314 prep_vma_destroy(vm, vmas[i]);
3315 xe_vma_destroy_unlocked(vmas[i]);
3320 up_write(&vm->lock);
3322 while (num_syncs--) {
3324 !(syncs[num_syncs].flags & DRM_XE_SYNC_SIGNAL))
3325 continue; /* Still in async worker */
3326 xe_sync_entry_cleanup(&syncs[num_syncs]);
3331 for (i = j; i < args->num_binds; ++i)
3341 if (args->num_binds > 1)
3347 * XXX: Using the TTM wrappers for now, likely can call into dma-resv code
3348 * directly to optimize. Also this likely should be an inline function.
3350 int xe_vm_lock(struct xe_vm *vm, struct ww_acquire_ctx *ww,
3351 int num_resv, bool intr)
3353 struct ttm_validate_buffer tv_vm;
3359 tv_vm.num_shared = num_resv;
3360 tv_vm.bo = xe_vm_ttm_bo(vm);;
3361 list_add_tail(&tv_vm.head, &objs);
3363 return ttm_eu_reserve_buffers(ww, &objs, intr, &dups);
3366 void xe_vm_unlock(struct xe_vm *vm, struct ww_acquire_ctx *ww)
3368 dma_resv_unlock(&vm->resv);
3369 ww_acquire_fini(ww);
3373 * xe_vm_invalidate_vma - invalidate GPU mappings for VMA without a lock
3374 * @vma: VMA to invalidate
3376 * Walks a list of page tables leaves which it memset the entries owned by this
3377 * VMA to zero, invalidates the TLBs, and block until TLBs invalidation is
3380 * Returns 0 for success, negative error code otherwise.
3382 int xe_vm_invalidate_vma(struct xe_vma *vma)
3384 struct xe_device *xe = vma->vm->xe;
3385 struct xe_tile *tile;
3386 u32 tile_needs_invalidate = 0;
3387 int seqno[XE_MAX_TILES_PER_DEVICE];
3391 XE_BUG_ON(!xe_vm_in_fault_mode(vma->vm));
3392 trace_xe_vma_usm_invalidate(vma);
3394 /* Check that we don't race with page-table updates */
3395 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
3396 if (xe_vma_is_userptr(vma)) {
3397 WARN_ON_ONCE(!mmu_interval_check_retry
3398 (&vma->userptr.notifier,
3399 vma->userptr.notifier_seq));
3400 WARN_ON_ONCE(!dma_resv_test_signaled(&vma->vm->resv,
3401 DMA_RESV_USAGE_BOOKKEEP));
3404 xe_bo_assert_held(vma->bo);
3408 for_each_tile(tile, xe, id) {
3409 if (xe_pt_zap_ptes(tile, vma)) {
3410 tile_needs_invalidate |= BIT(id);
3413 * FIXME: We potentially need to invalidate multiple
3414 * GTs within the tile
3416 seqno[id] = xe_gt_tlb_invalidation_vma(tile->primary_gt, NULL, vma);
3422 for_each_tile(tile, xe, id) {
3423 if (tile_needs_invalidate & BIT(id)) {
3424 ret = xe_gt_tlb_invalidation_wait(tile->primary_gt, seqno[id]);
3430 vma->usm.tile_invalidated = vma->tile_mask;
3435 int xe_analyze_vm(struct drm_printer *p, struct xe_vm *vm, int gt_id)
3437 struct rb_node *node;
3441 if (!down_read_trylock(&vm->lock)) {
3442 drm_printf(p, " Failed to acquire VM lock to dump capture");
3445 if (vm->pt_root[gt_id]) {
3446 addr = xe_bo_addr(vm->pt_root[gt_id]->bo, 0, XE_PAGE_SIZE,
3448 drm_printf(p, " VM root: A:0x%llx %s\n", addr, is_vram ? "VRAM" : "SYS");
3451 for (node = rb_first(&vm->vmas); node; node = rb_next(node)) {
3452 struct xe_vma *vma = to_xe_vma(node);
3453 bool is_userptr = xe_vma_is_userptr(vma);
3456 struct xe_res_cursor cur;
3458 if (vma->userptr.sg) {
3459 xe_res_first_sg(vma->userptr.sg, 0, XE_PAGE_SIZE,
3461 addr = xe_res_dma(&cur);
3466 addr = __xe_bo_addr(vma->bo, 0, XE_PAGE_SIZE, &is_vram);
3468 drm_printf(p, " [%016llx-%016llx] S:0x%016llx A:%016llx %s\n",
3469 vma->start, vma->end, vma->end - vma->start + 1ull,
3470 addr, is_userptr ? "USR" : is_vram ? "VRAM" : "SYS");