2 * Copyright 2008 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
24 * Authors: Dave Airlie
29 #include <drm/amdgpu_drm.h>
31 #include "amdgpu_trace.h"
35 * GPUVM is similar to the legacy gart on older asics, however
36 * rather than there being a single global gart table
37 * for the entire GPU, there are multiple VM page tables active
38 * at any given time. The VM page tables can contain a mix
39 * vram pages and system memory pages and system memory pages
40 * can be mapped as snooped (cached system pages) or unsnooped
41 * (uncached system pages).
42 * Each VM has an ID associated with it and there is a page table
43 * associated with each VMID. When execting a command buffer,
44 * the kernel tells the the ring what VMID to use for that command
45 * buffer. VMIDs are allocated dynamically as commands are submitted.
46 * The userspace drivers maintain their own address space and the kernel
47 * sets up their pages tables accordingly when they submit their
48 * command buffers and a VMID is assigned.
49 * Cayman/Trinity support up to 8 active VMs at any given time;
53 /* Special value that no flush is necessary */
54 #define AMDGPU_VM_NO_FLUSH (~0ll)
57 * amdgpu_vm_num_pde - return the number of page directory entries
59 * @adev: amdgpu_device pointer
61 * Calculate the number of page directory entries.
63 static unsigned amdgpu_vm_num_pdes(struct amdgpu_device *adev)
65 return adev->vm_manager.max_pfn >> amdgpu_vm_block_size;
69 * amdgpu_vm_directory_size - returns the size of the page directory in bytes
71 * @adev: amdgpu_device pointer
73 * Calculate the size of the page directory in bytes.
75 static unsigned amdgpu_vm_directory_size(struct amdgpu_device *adev)
77 return AMDGPU_GPU_PAGE_ALIGN(amdgpu_vm_num_pdes(adev) * 8);
81 * amdgpu_vm_get_pd_bo - add the VM PD to a validation list
83 * @vm: vm providing the BOs
84 * @validated: head of validation list
85 * @entry: entry to add
87 * Add the page directory to the list of BOs to
88 * validate for command submission.
90 void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
91 struct list_head *validated,
92 struct amdgpu_bo_list_entry *entry)
94 entry->robj = vm->page_directory;
96 entry->tv.bo = &vm->page_directory->tbo;
97 entry->tv.shared = true;
98 entry->user_pages = NULL;
99 list_add(&entry->tv.head, validated);
103 * amdgpu_vm_get_bos - add the vm BOs to a duplicates list
105 * @vm: vm providing the BOs
106 * @duplicates: head of duplicates list
108 * Add the page directory to the BO duplicates list
109 * for command submission.
111 void amdgpu_vm_get_pt_bos(struct amdgpu_vm *vm, struct list_head *duplicates)
115 /* add the vm page table to the list */
116 for (i = 0; i <= vm->max_pde_used; ++i) {
117 struct amdgpu_bo_list_entry *entry = &vm->page_tables[i].entry;
122 list_add(&entry->tv.head, duplicates);
128 * amdgpu_vm_move_pt_bos_in_lru - move the PT BOs to the LRU tail
130 * @adev: amdgpu device instance
131 * @vm: vm providing the BOs
133 * Move the PT BOs to the tail of the LRU.
135 void amdgpu_vm_move_pt_bos_in_lru(struct amdgpu_device *adev,
136 struct amdgpu_vm *vm)
138 struct ttm_bo_global *glob = adev->mman.bdev.glob;
141 spin_lock(&glob->lru_lock);
142 for (i = 0; i <= vm->max_pde_used; ++i) {
143 struct amdgpu_bo_list_entry *entry = &vm->page_tables[i].entry;
148 ttm_bo_move_to_lru_tail(&entry->robj->tbo);
150 spin_unlock(&glob->lru_lock);
154 * amdgpu_vm_grab_id - allocate the next free VMID
156 * @vm: vm to allocate id for
157 * @ring: ring we want to submit job to
158 * @sync: sync object where we add dependencies
159 * @fence: fence protecting ID from reuse
161 * Allocate an id for the vm, adding fences to the sync obj as necessary.
163 int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
164 struct amdgpu_sync *sync, struct fence *fence,
165 unsigned *vm_id, uint64_t *vm_pd_addr)
167 uint64_t pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
168 struct amdgpu_device *adev = ring->adev;
169 struct amdgpu_vm_id *id = &vm->ids[ring->idx];
170 struct fence *updates = sync->last_vm_update;
173 mutex_lock(&adev->vm_manager.lock);
175 /* check if the id is still valid */
177 struct fence *flushed = id->flushed_updates;
186 is_later = fence_is_later(updates, flushed);
188 owner = atomic_long_read(&id->mgr_id->owner);
189 if (!is_later && owner == (long)id &&
190 pd_addr == id->pd_gpu_addr) {
192 r = amdgpu_sync_fence(ring->adev, sync,
195 mutex_unlock(&adev->vm_manager.lock);
199 fence_put(id->mgr_id->active);
200 id->mgr_id->active = fence_get(fence);
202 list_move_tail(&id->mgr_id->list,
203 &adev->vm_manager.ids_lru);
205 *vm_id = id->mgr_id - adev->vm_manager.ids;
206 *vm_pd_addr = AMDGPU_VM_NO_FLUSH;
207 trace_amdgpu_vm_grab_id(vm, ring->idx, *vm_id,
210 mutex_unlock(&adev->vm_manager.lock);
215 id->mgr_id = list_first_entry(&adev->vm_manager.ids_lru,
216 struct amdgpu_vm_manager_id,
219 if (id->mgr_id->active && !fence_is_signaled(id->mgr_id->active)) {
220 struct amdgpu_vm_manager_id *mgr_id, *tmp;
221 struct list_head *head = &adev->vm_manager.ids_lru;
222 list_for_each_entry_safe(mgr_id, tmp, &adev->vm_manager.ids_lru, list) {
223 if (mgr_id->active && fence_is_signaled(mgr_id->active)) {
224 list_move(&mgr_id->list, head);
225 head = &mgr_id->list;
228 id->mgr_id = list_first_entry(&adev->vm_manager.ids_lru,
229 struct amdgpu_vm_manager_id,
233 r = amdgpu_sync_fence(ring->adev, sync, id->mgr_id->active);
235 fence_put(id->mgr_id->active);
236 id->mgr_id->active = fence_get(fence);
238 fence_put(id->flushed_updates);
239 id->flushed_updates = fence_get(updates);
241 id->pd_gpu_addr = pd_addr;
243 list_move_tail(&id->mgr_id->list, &adev->vm_manager.ids_lru);
244 atomic_long_set(&id->mgr_id->owner, (long)id);
246 *vm_id = id->mgr_id - adev->vm_manager.ids;
247 *vm_pd_addr = pd_addr;
248 trace_amdgpu_vm_grab_id(vm, ring->idx, *vm_id, *vm_pd_addr);
251 mutex_unlock(&adev->vm_manager.lock);
256 * amdgpu_vm_flush - hardware flush the vm
258 * @ring: ring to use for flush
259 * @vm_id: vmid number to use
260 * @pd_addr: address of the page directory
262 * Emit a VM flush when it is necessary.
264 void amdgpu_vm_flush(struct amdgpu_ring *ring,
265 unsigned vm_id, uint64_t pd_addr,
266 uint32_t gds_base, uint32_t gds_size,
267 uint32_t gws_base, uint32_t gws_size,
268 uint32_t oa_base, uint32_t oa_size)
270 struct amdgpu_device *adev = ring->adev;
271 struct amdgpu_vm_manager_id *mgr_id = &adev->vm_manager.ids[vm_id];
272 bool gds_switch_needed = ring->funcs->emit_gds_switch && (
273 mgr_id->gds_base != gds_base ||
274 mgr_id->gds_size != gds_size ||
275 mgr_id->gws_base != gws_base ||
276 mgr_id->gws_size != gws_size ||
277 mgr_id->oa_base != oa_base ||
278 mgr_id->oa_size != oa_size);
280 if (ring->funcs->emit_pipeline_sync && (
281 pd_addr != AMDGPU_VM_NO_FLUSH || gds_switch_needed))
282 amdgpu_ring_emit_pipeline_sync(ring);
284 if (pd_addr != AMDGPU_VM_NO_FLUSH) {
285 trace_amdgpu_vm_flush(pd_addr, ring->idx, vm_id);
286 amdgpu_ring_emit_vm_flush(ring, vm_id, pd_addr);
289 if (gds_switch_needed) {
290 mgr_id->gds_base = gds_base;
291 mgr_id->gds_size = gds_size;
292 mgr_id->gws_base = gws_base;
293 mgr_id->gws_size = gws_size;
294 mgr_id->oa_base = oa_base;
295 mgr_id->oa_size = oa_size;
296 amdgpu_ring_emit_gds_switch(ring, vm_id,
304 * amdgpu_vm_reset_id - reset VMID to zero
306 * @adev: amdgpu device structure
307 * @vm_id: vmid number to use
309 * Reset saved GDW, GWS and OA to force switch on next flush.
311 void amdgpu_vm_reset_id(struct amdgpu_device *adev, unsigned vm_id)
313 struct amdgpu_vm_manager_id *mgr_id = &adev->vm_manager.ids[vm_id];
315 mgr_id->gds_base = 0;
316 mgr_id->gds_size = 0;
317 mgr_id->gws_base = 0;
318 mgr_id->gws_size = 0;
324 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
327 * @bo: requested buffer object
329 * Find @bo inside the requested vm.
330 * Search inside the @bos vm list for the requested vm
331 * Returns the found bo_va or NULL if none is found
333 * Object has to be reserved!
335 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
336 struct amdgpu_bo *bo)
338 struct amdgpu_bo_va *bo_va;
340 list_for_each_entry(bo_va, &bo->va, bo_list) {
341 if (bo_va->vm == vm) {
349 * amdgpu_vm_update_pages - helper to call the right asic function
351 * @adev: amdgpu_device pointer
352 * @gtt: GART instance to use for mapping
353 * @gtt_flags: GTT hw access flags
354 * @ib: indirect buffer to fill with commands
355 * @pe: addr of the page entry
356 * @addr: dst addr to write into pe
357 * @count: number of page entries to update
358 * @incr: increase next addr by incr bytes
359 * @flags: hw access flags
361 * Traces the parameters and calls the right asic functions
362 * to setup the page table using the DMA.
364 static void amdgpu_vm_update_pages(struct amdgpu_device *adev,
365 struct amdgpu_gart *gtt,
367 struct amdgpu_ib *ib,
368 uint64_t pe, uint64_t addr,
369 unsigned count, uint32_t incr,
372 trace_amdgpu_vm_set_page(pe, addr, count, incr, flags);
374 if ((gtt == &adev->gart) && (flags == gtt_flags)) {
375 uint64_t src = gtt->table_addr + (addr >> 12) * 8;
376 amdgpu_vm_copy_pte(adev, ib, pe, src, count);
379 dma_addr_t *pages_addr = gtt->pages_addr;
380 amdgpu_vm_write_pte(adev, ib, pages_addr, pe, addr,
383 } else if (count < 3) {
384 amdgpu_vm_write_pte(adev, ib, NULL, pe, addr,
388 amdgpu_vm_set_pte_pde(adev, ib, pe, addr,
394 * amdgpu_vm_clear_bo - initially clear the page dir/table
396 * @adev: amdgpu_device pointer
399 * need to reserve bo first before calling it.
401 static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
402 struct amdgpu_vm *vm,
403 struct amdgpu_bo *bo)
405 struct amdgpu_ring *ring;
406 struct fence *fence = NULL;
407 struct amdgpu_job *job;
412 ring = container_of(vm->entity.sched, struct amdgpu_ring, sched);
414 r = reservation_object_reserve_shared(bo->tbo.resv);
418 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
422 addr = amdgpu_bo_gpu_offset(bo);
423 entries = amdgpu_bo_size(bo) / 8;
425 r = amdgpu_job_alloc_with_ib(adev, 64, &job);
429 amdgpu_vm_update_pages(adev, NULL, 0, &job->ibs[0], addr, 0, entries,
431 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
433 WARN_ON(job->ibs[0].length_dw > 64);
434 r = amdgpu_job_submit(job, ring, &vm->entity,
435 AMDGPU_FENCE_OWNER_VM, &fence);
439 amdgpu_bo_fence(bo, fence, true);
444 amdgpu_job_free(job);
451 * amdgpu_vm_map_gart - Resolve gart mapping of addr
453 * @pages_addr: optional DMA address to use for lookup
454 * @addr: the unmapped addr
456 * Look up the physical address of the page that the pte resolves
457 * to and return the pointer for the page table entry.
459 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
464 /* page table offset */
465 result = pages_addr[addr >> PAGE_SHIFT];
467 /* in case cpu page size != gpu page size*/
468 result |= addr & (~PAGE_MASK);
471 /* No mapping required */
475 result &= 0xFFFFFFFFFFFFF000ULL;
481 * amdgpu_vm_update_pdes - make sure that page directory is valid
483 * @adev: amdgpu_device pointer
485 * @start: start of GPU address range
486 * @end: end of GPU address range
488 * Allocates new page tables if necessary
489 * and updates the page directory.
490 * Returns 0 for success, error for failure.
492 int amdgpu_vm_update_page_directory(struct amdgpu_device *adev,
493 struct amdgpu_vm *vm)
495 struct amdgpu_ring *ring;
496 struct amdgpu_bo *pd = vm->page_directory;
497 uint64_t pd_addr = amdgpu_bo_gpu_offset(pd);
498 uint32_t incr = AMDGPU_VM_PTE_COUNT * 8;
499 uint64_t last_pde = ~0, last_pt = ~0;
500 unsigned count = 0, pt_idx, ndw;
501 struct amdgpu_job *job;
502 struct amdgpu_ib *ib;
503 struct fence *fence = NULL;
507 ring = container_of(vm->entity.sched, struct amdgpu_ring, sched);
512 /* assume the worst case */
513 ndw += vm->max_pde_used * 6;
515 r = amdgpu_job_alloc_with_ib(adev, ndw * 4, &job);
521 /* walk over the address space and update the page directory */
522 for (pt_idx = 0; pt_idx <= vm->max_pde_used; ++pt_idx) {
523 struct amdgpu_bo *bo = vm->page_tables[pt_idx].entry.robj;
529 pt = amdgpu_bo_gpu_offset(bo);
530 if (vm->page_tables[pt_idx].addr == pt)
532 vm->page_tables[pt_idx].addr = pt;
534 pde = pd_addr + pt_idx * 8;
535 if (((last_pde + 8 * count) != pde) ||
536 ((last_pt + incr * count) != pt)) {
539 amdgpu_vm_update_pages(adev, NULL, 0, ib,
554 amdgpu_vm_update_pages(adev, NULL, 0, ib, last_pde, last_pt,
555 count, incr, AMDGPU_PTE_VALID);
557 if (ib->length_dw != 0) {
558 amdgpu_ring_pad_ib(ring, ib);
559 amdgpu_sync_resv(adev, &job->sync, pd->tbo.resv,
560 AMDGPU_FENCE_OWNER_VM);
561 WARN_ON(ib->length_dw > ndw);
562 r = amdgpu_job_submit(job, ring, &vm->entity,
563 AMDGPU_FENCE_OWNER_VM, &fence);
567 amdgpu_bo_fence(pd, fence, true);
568 fence_put(vm->page_directory_fence);
569 vm->page_directory_fence = fence_get(fence);
573 amdgpu_job_free(job);
579 amdgpu_job_free(job);
584 * amdgpu_vm_frag_ptes - add fragment information to PTEs
586 * @adev: amdgpu_device pointer
587 * @gtt: GART instance to use for mapping
588 * @gtt_flags: GTT hw mapping flags
589 * @ib: IB for the update
590 * @pe_start: first PTE to handle
591 * @pe_end: last PTE to handle
592 * @addr: addr those PTEs should point to
593 * @flags: hw mapping flags
595 static void amdgpu_vm_frag_ptes(struct amdgpu_device *adev,
596 struct amdgpu_gart *gtt,
598 struct amdgpu_ib *ib,
599 uint64_t pe_start, uint64_t pe_end,
600 uint64_t addr, uint32_t flags)
603 * The MC L1 TLB supports variable sized pages, based on a fragment
604 * field in the PTE. When this field is set to a non-zero value, page
605 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
606 * flags are considered valid for all PTEs within the fragment range
607 * and corresponding mappings are assumed to be physically contiguous.
609 * The L1 TLB can store a single PTE for the whole fragment,
610 * significantly increasing the space available for translation
611 * caching. This leads to large improvements in throughput when the
612 * TLB is under pressure.
614 * The L2 TLB distributes small and large fragments into two
615 * asymmetric partitions. The large fragment cache is significantly
616 * larger. Thus, we try to use large fragments wherever possible.
617 * Userspace can support this by aligning virtual base address and
618 * allocation size to the fragment size.
621 /* SI and newer are optimized for 64KB */
622 uint64_t frag_flags = AMDGPU_PTE_FRAG_64KB;
623 uint64_t frag_align = 0x80;
625 uint64_t frag_start = ALIGN(pe_start, frag_align);
626 uint64_t frag_end = pe_end & ~(frag_align - 1);
630 /* Abort early if there isn't anything to do */
631 if (pe_start == pe_end)
634 /* system pages are non continuously */
635 if (gtt || !(flags & AMDGPU_PTE_VALID) || (frag_start >= frag_end)) {
637 count = (pe_end - pe_start) / 8;
638 amdgpu_vm_update_pages(adev, gtt, gtt_flags, ib, pe_start,
639 addr, count, AMDGPU_GPU_PAGE_SIZE,
644 /* handle the 4K area at the beginning */
645 if (pe_start != frag_start) {
646 count = (frag_start - pe_start) / 8;
647 amdgpu_vm_update_pages(adev, NULL, 0, ib, pe_start, addr,
648 count, AMDGPU_GPU_PAGE_SIZE, flags);
649 addr += AMDGPU_GPU_PAGE_SIZE * count;
652 /* handle the area in the middle */
653 count = (frag_end - frag_start) / 8;
654 amdgpu_vm_update_pages(adev, NULL, 0, ib, frag_start, addr, count,
655 AMDGPU_GPU_PAGE_SIZE, flags | frag_flags);
657 /* handle the 4K area at the end */
658 if (frag_end != pe_end) {
659 addr += AMDGPU_GPU_PAGE_SIZE * count;
660 count = (pe_end - frag_end) / 8;
661 amdgpu_vm_update_pages(adev, NULL, 0, ib, frag_end, addr,
662 count, AMDGPU_GPU_PAGE_SIZE, flags);
667 * amdgpu_vm_update_ptes - make sure that page tables are valid
669 * @adev: amdgpu_device pointer
670 * @gtt: GART instance to use for mapping
671 * @gtt_flags: GTT hw mapping flags
673 * @start: start of GPU address range
674 * @end: end of GPU address range
675 * @dst: destination address to map to
676 * @flags: mapping flags
678 * Update the page tables in the range @start - @end.
680 static void amdgpu_vm_update_ptes(struct amdgpu_device *adev,
681 struct amdgpu_gart *gtt,
683 struct amdgpu_vm *vm,
684 struct amdgpu_ib *ib,
685 uint64_t start, uint64_t end,
686 uint64_t dst, uint32_t flags)
688 const uint64_t mask = AMDGPU_VM_PTE_COUNT - 1;
690 uint64_t last_pe_start = ~0, last_pe_end = ~0, last_dst = ~0;
693 /* walk over the address space and update the page tables */
694 for (addr = start; addr < end; ) {
695 uint64_t pt_idx = addr >> amdgpu_vm_block_size;
696 struct amdgpu_bo *pt = vm->page_tables[pt_idx].entry.robj;
700 if ((addr & ~mask) == (end & ~mask))
703 nptes = AMDGPU_VM_PTE_COUNT - (addr & mask);
705 pe_start = amdgpu_bo_gpu_offset(pt);
706 pe_start += (addr & mask) * 8;
708 if (last_pe_end != pe_start) {
710 amdgpu_vm_frag_ptes(adev, gtt, gtt_flags, ib,
711 last_pe_start, last_pe_end,
714 last_pe_start = pe_start;
715 last_pe_end = pe_start + 8 * nptes;
718 last_pe_end += 8 * nptes;
722 dst += nptes * AMDGPU_GPU_PAGE_SIZE;
725 amdgpu_vm_frag_ptes(adev, gtt, gtt_flags, ib,
726 last_pe_start, last_pe_end,
731 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
733 * @adev: amdgpu_device pointer
734 * @gtt: GART instance to use for mapping
735 * @gtt_flags: flags as they are used for GTT
737 * @start: start of mapped range
738 * @last: last mapped entry
739 * @flags: flags for the entries
740 * @addr: addr to set the area to
741 * @fence: optional resulting fence
743 * Fill in the page table entries between @start and @last.
744 * Returns 0 for success, -EINVAL for failure.
746 static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
747 struct amdgpu_gart *gtt,
749 struct amdgpu_vm *vm,
750 uint64_t start, uint64_t last,
751 uint32_t flags, uint64_t addr,
752 struct fence **fence)
754 struct amdgpu_ring *ring;
755 void *owner = AMDGPU_FENCE_OWNER_VM;
756 unsigned nptes, ncmds, ndw;
757 struct amdgpu_job *job;
758 struct amdgpu_ib *ib;
759 struct fence *f = NULL;
762 ring = container_of(vm->entity.sched, struct amdgpu_ring, sched);
764 /* sync to everything on unmapping */
765 if (!(flags & AMDGPU_PTE_VALID))
766 owner = AMDGPU_FENCE_OWNER_UNDEFINED;
768 nptes = last - start + 1;
771 * reserve space for one command every (1 << BLOCK_SIZE)
772 * entries or 2k dwords (whatever is smaller)
774 ncmds = (nptes >> min(amdgpu_vm_block_size, 11)) + 1;
779 if ((gtt == &adev->gart) && (flags == gtt_flags)) {
780 /* only copy commands needed */
784 /* header for write data commands */
787 /* body of write data command */
791 /* set page commands needed */
794 /* two extra commands for begin/end of fragment */
798 r = amdgpu_job_alloc_with_ib(adev, ndw * 4, &job);
804 r = amdgpu_sync_resv(adev, &job->sync, vm->page_directory->tbo.resv,
809 r = reservation_object_reserve_shared(vm->page_directory->tbo.resv);
813 amdgpu_vm_update_ptes(adev, gtt, gtt_flags, vm, ib, start, last + 1,
816 amdgpu_ring_pad_ib(ring, ib);
817 WARN_ON(ib->length_dw > ndw);
818 r = amdgpu_job_submit(job, ring, &vm->entity,
819 AMDGPU_FENCE_OWNER_VM, &f);
823 amdgpu_bo_fence(vm->page_directory, f, true);
826 *fence = fence_get(f);
832 amdgpu_job_free(job);
837 * amdgpu_vm_bo_split_mapping - split a mapping into smaller chunks
839 * @adev: amdgpu_device pointer
840 * @gtt: GART instance to use for mapping
842 * @mapping: mapped range and flags to use for the update
843 * @addr: addr to set the area to
844 * @gtt_flags: flags as they are used for GTT
845 * @fence: optional resulting fence
847 * Split the mapping into smaller chunks so that each update fits
849 * Returns 0 for success, -EINVAL for failure.
851 static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev,
852 struct amdgpu_gart *gtt,
854 struct amdgpu_vm *vm,
855 struct amdgpu_bo_va_mapping *mapping,
856 uint64_t addr, struct fence **fence)
858 const uint64_t max_size = 64ULL * 1024ULL * 1024ULL / AMDGPU_GPU_PAGE_SIZE;
860 uint64_t start = mapping->it.start;
861 uint32_t flags = gtt_flags;
864 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
865 * but in case of something, we filter the flags in first place
867 if (!(mapping->flags & AMDGPU_PTE_READABLE))
868 flags &= ~AMDGPU_PTE_READABLE;
869 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
870 flags &= ~AMDGPU_PTE_WRITEABLE;
872 trace_amdgpu_vm_bo_update(mapping);
874 addr += mapping->offset;
876 if (!gtt || ((gtt == &adev->gart) && (flags == gtt_flags)))
877 return amdgpu_vm_bo_update_mapping(adev, gtt, gtt_flags, vm,
878 start, mapping->it.last,
881 while (start != mapping->it.last + 1) {
884 last = min((uint64_t)mapping->it.last, start + max_size - 1);
885 r = amdgpu_vm_bo_update_mapping(adev, gtt, gtt_flags, vm,
886 start, last, flags, addr,
892 addr += max_size * AMDGPU_GPU_PAGE_SIZE;
899 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
901 * @adev: amdgpu_device pointer
902 * @bo_va: requested BO and VM object
905 * Fill in the page table entries for @bo_va.
906 * Returns 0 for success, -EINVAL for failure.
908 * Object have to be reserved and mutex must be locked!
910 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
911 struct amdgpu_bo_va *bo_va,
912 struct ttm_mem_reg *mem)
914 struct amdgpu_vm *vm = bo_va->vm;
915 struct amdgpu_bo_va_mapping *mapping;
916 struct amdgpu_gart *gtt = NULL;
922 addr = (u64)mem->start << PAGE_SHIFT;
923 switch (mem->mem_type) {
925 gtt = &bo_va->bo->adev->gart;
929 addr += adev->vm_manager.vram_base_offset;
939 flags = amdgpu_ttm_tt_pte_flags(adev, bo_va->bo->tbo.ttm, mem);
941 spin_lock(&vm->status_lock);
942 if (!list_empty(&bo_va->vm_status))
943 list_splice_init(&bo_va->valids, &bo_va->invalids);
944 spin_unlock(&vm->status_lock);
946 list_for_each_entry(mapping, &bo_va->invalids, list) {
947 r = amdgpu_vm_bo_split_mapping(adev, gtt, flags, vm, mapping, addr,
948 &bo_va->last_pt_update);
953 if (trace_amdgpu_vm_bo_mapping_enabled()) {
954 list_for_each_entry(mapping, &bo_va->valids, list)
955 trace_amdgpu_vm_bo_mapping(mapping);
957 list_for_each_entry(mapping, &bo_va->invalids, list)
958 trace_amdgpu_vm_bo_mapping(mapping);
961 spin_lock(&vm->status_lock);
962 list_splice_init(&bo_va->invalids, &bo_va->valids);
963 list_del_init(&bo_va->vm_status);
965 list_add(&bo_va->vm_status, &vm->cleared);
966 spin_unlock(&vm->status_lock);
972 * amdgpu_vm_clear_freed - clear freed BOs in the PT
974 * @adev: amdgpu_device pointer
977 * Make sure all freed BOs are cleared in the PT.
978 * Returns 0 for success.
980 * PTs have to be reserved and mutex must be locked!
982 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
983 struct amdgpu_vm *vm)
985 struct amdgpu_bo_va_mapping *mapping;
988 while (!list_empty(&vm->freed)) {
989 mapping = list_first_entry(&vm->freed,
990 struct amdgpu_bo_va_mapping, list);
991 list_del(&mapping->list);
993 r = amdgpu_vm_bo_split_mapping(adev, NULL, 0, vm, mapping,
1005 * amdgpu_vm_clear_invalids - clear invalidated BOs in the PT
1007 * @adev: amdgpu_device pointer
1010 * Make sure all invalidated BOs are cleared in the PT.
1011 * Returns 0 for success.
1013 * PTs have to be reserved and mutex must be locked!
1015 int amdgpu_vm_clear_invalids(struct amdgpu_device *adev,
1016 struct amdgpu_vm *vm, struct amdgpu_sync *sync)
1018 struct amdgpu_bo_va *bo_va = NULL;
1021 spin_lock(&vm->status_lock);
1022 while (!list_empty(&vm->invalidated)) {
1023 bo_va = list_first_entry(&vm->invalidated,
1024 struct amdgpu_bo_va, vm_status);
1025 spin_unlock(&vm->status_lock);
1027 r = amdgpu_vm_bo_update(adev, bo_va, NULL);
1031 spin_lock(&vm->status_lock);
1033 spin_unlock(&vm->status_lock);
1036 r = amdgpu_sync_fence(adev, sync, bo_va->last_pt_update);
1042 * amdgpu_vm_bo_add - add a bo to a specific vm
1044 * @adev: amdgpu_device pointer
1046 * @bo: amdgpu buffer object
1048 * Add @bo into the requested vm.
1049 * Add @bo to the list of bos associated with the vm
1050 * Returns newly added bo_va or NULL for failure
1052 * Object has to be reserved!
1054 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
1055 struct amdgpu_vm *vm,
1056 struct amdgpu_bo *bo)
1058 struct amdgpu_bo_va *bo_va;
1060 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
1061 if (bo_va == NULL) {
1066 bo_va->ref_count = 1;
1067 INIT_LIST_HEAD(&bo_va->bo_list);
1068 INIT_LIST_HEAD(&bo_va->valids);
1069 INIT_LIST_HEAD(&bo_va->invalids);
1070 INIT_LIST_HEAD(&bo_va->vm_status);
1072 list_add_tail(&bo_va->bo_list, &bo->va);
1078 * amdgpu_vm_bo_map - map bo inside a vm
1080 * @adev: amdgpu_device pointer
1081 * @bo_va: bo_va to store the address
1082 * @saddr: where to map the BO
1083 * @offset: requested offset in the BO
1084 * @flags: attributes of pages (read/write/valid/etc.)
1086 * Add a mapping of the BO at the specefied addr into the VM.
1087 * Returns 0 for success, error for failure.
1089 * Object has to be reserved and unreserved outside!
1091 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
1092 struct amdgpu_bo_va *bo_va,
1093 uint64_t saddr, uint64_t offset,
1094 uint64_t size, uint32_t flags)
1096 struct amdgpu_bo_va_mapping *mapping;
1097 struct amdgpu_vm *vm = bo_va->vm;
1098 struct interval_tree_node *it;
1099 unsigned last_pfn, pt_idx;
1103 /* validate the parameters */
1104 if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
1105 size == 0 || size & AMDGPU_GPU_PAGE_MASK)
1108 /* make sure object fit at this offset */
1109 eaddr = saddr + size - 1;
1110 if ((saddr >= eaddr) || (offset + size > amdgpu_bo_size(bo_va->bo)))
1113 last_pfn = eaddr / AMDGPU_GPU_PAGE_SIZE;
1114 if (last_pfn >= adev->vm_manager.max_pfn) {
1115 dev_err(adev->dev, "va above limit (0x%08X >= 0x%08X)\n",
1116 last_pfn, adev->vm_manager.max_pfn);
1120 saddr /= AMDGPU_GPU_PAGE_SIZE;
1121 eaddr /= AMDGPU_GPU_PAGE_SIZE;
1123 it = interval_tree_iter_first(&vm->va, saddr, eaddr);
1125 struct amdgpu_bo_va_mapping *tmp;
1126 tmp = container_of(it, struct amdgpu_bo_va_mapping, it);
1127 /* bo and tmp overlap, invalid addr */
1128 dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
1129 "0x%010lx-0x%010lx\n", bo_va->bo, saddr, eaddr,
1130 tmp->it.start, tmp->it.last + 1);
1135 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1141 INIT_LIST_HEAD(&mapping->list);
1142 mapping->it.start = saddr;
1143 mapping->it.last = eaddr;
1144 mapping->offset = offset;
1145 mapping->flags = flags;
1147 list_add(&mapping->list, &bo_va->invalids);
1148 interval_tree_insert(&mapping->it, &vm->va);
1150 /* Make sure the page tables are allocated */
1151 saddr >>= amdgpu_vm_block_size;
1152 eaddr >>= amdgpu_vm_block_size;
1154 BUG_ON(eaddr >= amdgpu_vm_num_pdes(adev));
1156 if (eaddr > vm->max_pde_used)
1157 vm->max_pde_used = eaddr;
1159 /* walk over the address space and allocate the page tables */
1160 for (pt_idx = saddr; pt_idx <= eaddr; ++pt_idx) {
1161 struct reservation_object *resv = vm->page_directory->tbo.resv;
1162 struct amdgpu_bo_list_entry *entry;
1163 struct amdgpu_bo *pt;
1165 entry = &vm->page_tables[pt_idx].entry;
1169 r = amdgpu_bo_create(adev, AMDGPU_VM_PTE_COUNT * 8,
1170 AMDGPU_GPU_PAGE_SIZE, true,
1171 AMDGPU_GEM_DOMAIN_VRAM,
1172 AMDGPU_GEM_CREATE_NO_CPU_ACCESS,
1177 /* Keep a reference to the page table to avoid freeing
1178 * them up in the wrong order.
1180 pt->parent = amdgpu_bo_ref(vm->page_directory);
1182 r = amdgpu_vm_clear_bo(adev, vm, pt);
1184 amdgpu_bo_unref(&pt);
1189 entry->priority = 0;
1190 entry->tv.bo = &entry->robj->tbo;
1191 entry->tv.shared = true;
1192 entry->user_pages = NULL;
1193 vm->page_tables[pt_idx].addr = 0;
1199 list_del(&mapping->list);
1200 interval_tree_remove(&mapping->it, &vm->va);
1201 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1209 * amdgpu_vm_bo_unmap - remove bo mapping from vm
1211 * @adev: amdgpu_device pointer
1212 * @bo_va: bo_va to remove the address from
1213 * @saddr: where to the BO is mapped
1215 * Remove a mapping of the BO at the specefied addr from the VM.
1216 * Returns 0 for success, error for failure.
1218 * Object has to be reserved and unreserved outside!
1220 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
1221 struct amdgpu_bo_va *bo_va,
1224 struct amdgpu_bo_va_mapping *mapping;
1225 struct amdgpu_vm *vm = bo_va->vm;
1228 saddr /= AMDGPU_GPU_PAGE_SIZE;
1230 list_for_each_entry(mapping, &bo_va->valids, list) {
1231 if (mapping->it.start == saddr)
1235 if (&mapping->list == &bo_va->valids) {
1238 list_for_each_entry(mapping, &bo_va->invalids, list) {
1239 if (mapping->it.start == saddr)
1243 if (&mapping->list == &bo_va->invalids)
1247 list_del(&mapping->list);
1248 interval_tree_remove(&mapping->it, &vm->va);
1249 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1252 list_add(&mapping->list, &vm->freed);
1260 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
1262 * @adev: amdgpu_device pointer
1263 * @bo_va: requested bo_va
1265 * Remove @bo_va->bo from the requested vm.
1267 * Object have to be reserved!
1269 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
1270 struct amdgpu_bo_va *bo_va)
1272 struct amdgpu_bo_va_mapping *mapping, *next;
1273 struct amdgpu_vm *vm = bo_va->vm;
1275 list_del(&bo_va->bo_list);
1277 spin_lock(&vm->status_lock);
1278 list_del(&bo_va->vm_status);
1279 spin_unlock(&vm->status_lock);
1281 list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
1282 list_del(&mapping->list);
1283 interval_tree_remove(&mapping->it, &vm->va);
1284 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1285 list_add(&mapping->list, &vm->freed);
1287 list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
1288 list_del(&mapping->list);
1289 interval_tree_remove(&mapping->it, &vm->va);
1293 fence_put(bo_va->last_pt_update);
1298 * amdgpu_vm_bo_invalidate - mark the bo as invalid
1300 * @adev: amdgpu_device pointer
1302 * @bo: amdgpu buffer object
1304 * Mark @bo as invalid.
1306 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
1307 struct amdgpu_bo *bo)
1309 struct amdgpu_bo_va *bo_va;
1311 list_for_each_entry(bo_va, &bo->va, bo_list) {
1312 spin_lock(&bo_va->vm->status_lock);
1313 if (list_empty(&bo_va->vm_status))
1314 list_add(&bo_va->vm_status, &bo_va->vm->invalidated);
1315 spin_unlock(&bo_va->vm->status_lock);
1320 * amdgpu_vm_init - initialize a vm instance
1322 * @adev: amdgpu_device pointer
1327 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1329 const unsigned align = min(AMDGPU_VM_PTB_ALIGN_SIZE,
1330 AMDGPU_VM_PTE_COUNT * 8);
1331 unsigned pd_size, pd_entries;
1332 unsigned ring_instance;
1333 struct amdgpu_ring *ring;
1334 struct amd_sched_rq *rq;
1337 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1338 vm->ids[i].mgr_id = NULL;
1339 vm->ids[i].flushed_updates = NULL;
1342 spin_lock_init(&vm->status_lock);
1343 INIT_LIST_HEAD(&vm->invalidated);
1344 INIT_LIST_HEAD(&vm->cleared);
1345 INIT_LIST_HEAD(&vm->freed);
1347 pd_size = amdgpu_vm_directory_size(adev);
1348 pd_entries = amdgpu_vm_num_pdes(adev);
1350 /* allocate page table array */
1351 vm->page_tables = drm_calloc_large(pd_entries, sizeof(struct amdgpu_vm_pt));
1352 if (vm->page_tables == NULL) {
1353 DRM_ERROR("Cannot allocate memory for page table array\n");
1357 /* create scheduler entity for page table updates */
1359 ring_instance = atomic_inc_return(&adev->vm_manager.vm_pte_next_ring);
1360 ring_instance %= adev->vm_manager.vm_pte_num_rings;
1361 ring = adev->vm_manager.vm_pte_rings[ring_instance];
1362 rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_KERNEL];
1363 r = amd_sched_entity_init(&ring->sched, &vm->entity,
1364 rq, amdgpu_sched_jobs);
1368 vm->page_directory_fence = NULL;
1370 r = amdgpu_bo_create(adev, pd_size, align, true,
1371 AMDGPU_GEM_DOMAIN_VRAM,
1372 AMDGPU_GEM_CREATE_NO_CPU_ACCESS,
1373 NULL, NULL, &vm->page_directory);
1375 goto error_free_sched_entity;
1377 r = amdgpu_bo_reserve(vm->page_directory, false);
1379 goto error_free_page_directory;
1381 r = amdgpu_vm_clear_bo(adev, vm, vm->page_directory);
1382 amdgpu_bo_unreserve(vm->page_directory);
1384 goto error_free_page_directory;
1388 error_free_page_directory:
1389 amdgpu_bo_unref(&vm->page_directory);
1390 vm->page_directory = NULL;
1392 error_free_sched_entity:
1393 amd_sched_entity_fini(&ring->sched, &vm->entity);
1399 * amdgpu_vm_fini - tear down a vm instance
1401 * @adev: amdgpu_device pointer
1405 * Unbind the VM and remove all bos from the vm bo list
1407 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1409 struct amdgpu_bo_va_mapping *mapping, *tmp;
1412 amd_sched_entity_fini(vm->entity.sched, &vm->entity);
1414 if (!RB_EMPTY_ROOT(&vm->va)) {
1415 dev_err(adev->dev, "still active bo inside vm\n");
1417 rbtree_postorder_for_each_entry_safe(mapping, tmp, &vm->va, it.rb) {
1418 list_del(&mapping->list);
1419 interval_tree_remove(&mapping->it, &vm->va);
1422 list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
1423 list_del(&mapping->list);
1427 for (i = 0; i < amdgpu_vm_num_pdes(adev); i++)
1428 amdgpu_bo_unref(&vm->page_tables[i].entry.robj);
1429 drm_free_large(vm->page_tables);
1431 amdgpu_bo_unref(&vm->page_directory);
1432 fence_put(vm->page_directory_fence);
1434 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1435 struct amdgpu_vm_id *id = &vm->ids[i];
1438 atomic_long_cmpxchg(&id->mgr_id->owner,
1440 fence_put(id->flushed_updates);
1445 * amdgpu_vm_manager_init - init the VM manager
1447 * @adev: amdgpu_device pointer
1449 * Initialize the VM manager structures
1451 void amdgpu_vm_manager_init(struct amdgpu_device *adev)
1455 INIT_LIST_HEAD(&adev->vm_manager.ids_lru);
1457 /* skip over VMID 0, since it is the system VM */
1458 for (i = 1; i < adev->vm_manager.num_ids; ++i) {
1459 amdgpu_vm_reset_id(adev, i);
1460 list_add_tail(&adev->vm_manager.ids[i].list,
1461 &adev->vm_manager.ids_lru);
1464 atomic_set(&adev->vm_manager.vm_pte_next_ring, 0);
1468 * amdgpu_vm_manager_fini - cleanup VM manager
1470 * @adev: amdgpu_device pointer
1472 * Cleanup the VM manager and free resources.
1474 void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
1478 for (i = 0; i < AMDGPU_NUM_VM; ++i)
1479 fence_put(adev->vm_manager.ids[i].active);