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;
54 * amdgpu_vm_num_pde - return the number of page directory entries
56 * @adev: amdgpu_device pointer
58 * Calculate the number of page directory entries (cayman+).
60 static unsigned amdgpu_vm_num_pdes(struct amdgpu_device *adev)
62 return adev->vm_manager.max_pfn >> amdgpu_vm_block_size;
66 * amdgpu_vm_directory_size - returns the size of the page directory in bytes
68 * @adev: amdgpu_device pointer
70 * Calculate the size of the page directory in bytes (cayman+).
72 static unsigned amdgpu_vm_directory_size(struct amdgpu_device *adev)
74 return AMDGPU_GPU_PAGE_ALIGN(amdgpu_vm_num_pdes(adev) * 8);
78 * amdgpu_vm_get_bos - add the vm BOs to a validation list
80 * @vm: vm providing the BOs
81 * @head: head of validation list
83 * Add the page directory to the list of BOs to
84 * validate for command submission (cayman+).
86 struct amdgpu_bo_list_entry *amdgpu_vm_get_bos(struct amdgpu_device *adev,
88 struct list_head *head)
90 struct amdgpu_bo_list_entry *list;
93 mutex_lock(&vm->mutex);
94 list = drm_malloc_ab(vm->max_pde_used + 2,
95 sizeof(struct amdgpu_bo_list_entry));
97 mutex_unlock(&vm->mutex);
101 /* add the vm page table to the list */
102 list[0].robj = vm->page_directory;
103 list[0].prefered_domains = AMDGPU_GEM_DOMAIN_VRAM;
104 list[0].allowed_domains = AMDGPU_GEM_DOMAIN_VRAM;
105 list[0].priority = 0;
106 list[0].tv.bo = &vm->page_directory->tbo;
107 list[0].tv.shared = true;
108 list_add(&list[0].tv.head, head);
110 for (i = 0, idx = 1; i <= vm->max_pde_used; i++) {
111 if (!vm->page_tables[i].bo)
114 list[idx].robj = vm->page_tables[i].bo;
115 list[idx].prefered_domains = AMDGPU_GEM_DOMAIN_VRAM;
116 list[idx].allowed_domains = AMDGPU_GEM_DOMAIN_VRAM;
117 list[idx].priority = 0;
118 list[idx].tv.bo = &list[idx].robj->tbo;
119 list[idx].tv.shared = true;
120 list_add(&list[idx++].tv.head, head);
122 mutex_unlock(&vm->mutex);
128 * amdgpu_vm_grab_id - allocate the next free VMID
130 * @vm: vm to allocate id for
131 * @ring: ring we want to submit job to
132 * @sync: sync object where we add dependencies
134 * Allocate an id for the vm, adding fences to the sync obj as necessary.
136 * Global mutex must be locked!
138 int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
139 struct amdgpu_sync *sync)
141 struct amdgpu_fence *best[AMDGPU_MAX_RINGS] = {};
142 struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx];
143 struct amdgpu_device *adev = ring->adev;
145 unsigned choices[2] = {};
148 /* check if the id is still valid */
149 if (vm_id->id && vm_id->last_id_use &&
150 vm_id->last_id_use == adev->vm_manager.active[vm_id->id])
153 /* we definately need to flush */
154 vm_id->pd_gpu_addr = ~0ll;
156 /* skip over VMID 0, since it is the system VM */
157 for (i = 1; i < adev->vm_manager.nvm; ++i) {
158 struct amdgpu_fence *fence = adev->vm_manager.active[i];
161 /* found a free one */
163 trace_amdgpu_vm_grab_id(i, ring->idx);
167 if (amdgpu_fence_is_earlier(fence, best[fence->ring->idx])) {
168 best[fence->ring->idx] = fence;
169 choices[fence->ring == ring ? 0 : 1] = i;
173 for (i = 0; i < 2; ++i) {
175 struct amdgpu_fence *fence;
177 fence = adev->vm_manager.active[choices[i]];
178 vm_id->id = choices[i];
180 trace_amdgpu_vm_grab_id(choices[i], ring->idx);
181 return amdgpu_sync_fence(ring->adev, sync, &fence->base);
185 /* should never happen */
191 * amdgpu_vm_flush - hardware flush the vm
193 * @ring: ring to use for flush
194 * @vm: vm we want to flush
195 * @updates: last vm update that we waited for
197 * Flush the vm (cayman+).
199 * Global and local mutex must be locked!
201 void amdgpu_vm_flush(struct amdgpu_ring *ring,
202 struct amdgpu_vm *vm,
203 struct amdgpu_fence *updates)
205 uint64_t pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
206 struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx];
207 struct amdgpu_fence *flushed_updates = vm_id->flushed_updates;
209 if (pd_addr != vm_id->pd_gpu_addr || !flushed_updates ||
210 (updates && amdgpu_fence_is_earlier(flushed_updates, updates))) {
212 trace_amdgpu_vm_flush(pd_addr, ring->idx, vm_id->id);
213 vm_id->flushed_updates = amdgpu_fence_ref(
214 amdgpu_fence_later(flushed_updates, updates));
215 amdgpu_fence_unref(&flushed_updates);
216 vm_id->pd_gpu_addr = pd_addr;
217 amdgpu_ring_emit_vm_flush(ring, vm_id->id, vm_id->pd_gpu_addr);
222 * amdgpu_vm_fence - remember fence for vm
224 * @adev: amdgpu_device pointer
225 * @vm: vm we want to fence
226 * @fence: fence to remember
228 * Fence the vm (cayman+).
229 * Set the fence used to protect page table and id.
231 * Global and local mutex must be locked!
233 void amdgpu_vm_fence(struct amdgpu_device *adev,
234 struct amdgpu_vm *vm,
235 struct amdgpu_fence *fence)
237 unsigned ridx = fence->ring->idx;
238 unsigned vm_id = vm->ids[ridx].id;
240 amdgpu_fence_unref(&adev->vm_manager.active[vm_id]);
241 adev->vm_manager.active[vm_id] = amdgpu_fence_ref(fence);
243 amdgpu_fence_unref(&vm->ids[ridx].last_id_use);
244 vm->ids[ridx].last_id_use = amdgpu_fence_ref(fence);
248 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
251 * @bo: requested buffer object
253 * Find @bo inside the requested vm (cayman+).
254 * Search inside the @bos vm list for the requested vm
255 * Returns the found bo_va or NULL if none is found
257 * Object has to be reserved!
259 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
260 struct amdgpu_bo *bo)
262 struct amdgpu_bo_va *bo_va;
264 list_for_each_entry(bo_va, &bo->va, bo_list) {
265 if (bo_va->vm == vm) {
273 * amdgpu_vm_update_pages - helper to call the right asic function
275 * @adev: amdgpu_device pointer
276 * @ib: indirect buffer to fill with commands
277 * @pe: addr of the page entry
278 * @addr: dst addr to write into pe
279 * @count: number of page entries to update
280 * @incr: increase next addr by incr bytes
281 * @flags: hw access flags
282 * @gtt_flags: GTT hw access flags
284 * Traces the parameters and calls the right asic functions
285 * to setup the page table using the DMA.
287 static void amdgpu_vm_update_pages(struct amdgpu_device *adev,
288 struct amdgpu_ib *ib,
289 uint64_t pe, uint64_t addr,
290 unsigned count, uint32_t incr,
291 uint32_t flags, uint32_t gtt_flags)
293 trace_amdgpu_vm_set_page(pe, addr, count, incr, flags);
295 if ((flags & AMDGPU_PTE_SYSTEM) && (flags == gtt_flags)) {
296 uint64_t src = adev->gart.table_addr + (addr >> 12) * 8;
297 amdgpu_vm_copy_pte(adev, ib, pe, src, count);
299 } else if ((flags & AMDGPU_PTE_SYSTEM) || (count < 3)) {
300 amdgpu_vm_write_pte(adev, ib, pe, addr,
304 amdgpu_vm_set_pte_pde(adev, ib, pe, addr,
309 static int amdgpu_vm_free_job(
310 struct amdgpu_cs_parser *sched_job)
313 for (i = 0; i < sched_job->num_ibs; i++)
314 amdgpu_ib_free(sched_job->adev, &sched_job->ibs[i]);
315 kfree(sched_job->ibs);
319 static int amdgpu_vm_run_job(
320 struct amdgpu_cs_parser *sched_job)
322 amdgpu_bo_fence(sched_job->job_param.vm.bo,
323 &sched_job->ibs[sched_job->num_ibs -1].fence->base, true);
328 * amdgpu_vm_clear_bo - initially clear the page dir/table
330 * @adev: amdgpu_device pointer
333 static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
334 struct amdgpu_bo *bo)
336 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
337 struct amdgpu_cs_parser *sched_job = NULL;
338 struct amdgpu_ib *ib;
343 r = amdgpu_bo_reserve(bo, false);
347 r = reservation_object_reserve_shared(bo->tbo.resv);
351 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
353 goto error_unreserve;
355 addr = amdgpu_bo_gpu_offset(bo);
356 entries = amdgpu_bo_size(bo) / 8;
358 ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
360 goto error_unreserve;
362 r = amdgpu_ib_get(ring, NULL, entries * 2 + 64, ib);
368 amdgpu_vm_update_pages(adev, ib, addr, 0, entries, 0, 0, 0);
369 amdgpu_vm_pad_ib(adev, ib);
370 WARN_ON(ib->length_dw > 64);
372 if (amdgpu_enable_scheduler) {
374 sched_job = amdgpu_cs_parser_create(adev, AMDGPU_FENCE_OWNER_VM,
375 adev->kernel_ctx, ib, 1);
378 sched_job->job_param.vm.bo = bo;
379 sched_job->run_job = amdgpu_vm_run_job;
380 sched_job->free_job = amdgpu_vm_free_job;
381 ib->sequence = amd_sched_push_job(ring->scheduler,
382 &adev->kernel_ctx->rings[ring->idx].c_entity,
384 r = amd_sched_wait_emit(&adev->kernel_ctx->rings[ring->idx].c_entity,
385 ib->sequence, false, -1);
387 DRM_ERROR("emit timeout\n");
389 amdgpu_bo_unreserve(bo);
392 r = amdgpu_ib_schedule(adev, 1, ib, AMDGPU_FENCE_OWNER_VM);
395 amdgpu_bo_fence(bo, &ib->fence->base, true);
399 amdgpu_ib_free(adev, ib);
403 amdgpu_bo_unreserve(bo);
408 * amdgpu_vm_map_gart - get the physical address of a gart page
410 * @adev: amdgpu_device pointer
411 * @addr: the unmapped addr
413 * Look up the physical address of the page that the pte resolves
415 * Returns the physical address of the page.
417 uint64_t amdgpu_vm_map_gart(struct amdgpu_device *adev, uint64_t addr)
421 /* page table offset */
422 result = adev->gart.pages_addr[addr >> PAGE_SHIFT];
424 /* in case cpu page size != gpu page size*/
425 result |= addr & (~PAGE_MASK);
431 * amdgpu_vm_update_pdes - make sure that page directory is valid
433 * @adev: amdgpu_device pointer
435 * @start: start of GPU address range
436 * @end: end of GPU address range
438 * Allocates new page tables if necessary
439 * and updates the page directory (cayman+).
440 * Returns 0 for success, error for failure.
442 * Global and local mutex must be locked!
444 int amdgpu_vm_update_page_directory(struct amdgpu_device *adev,
445 struct amdgpu_vm *vm)
447 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
448 struct amdgpu_bo *pd = vm->page_directory;
449 uint64_t pd_addr = amdgpu_bo_gpu_offset(pd);
450 uint32_t incr = AMDGPU_VM_PTE_COUNT * 8;
451 uint64_t last_pde = ~0, last_pt = ~0;
452 unsigned count = 0, pt_idx, ndw;
453 struct amdgpu_ib *ib;
454 struct amdgpu_cs_parser *sched_job = NULL;
461 /* assume the worst case */
462 ndw += vm->max_pde_used * 6;
464 /* update too big for an IB */
468 ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
472 r = amdgpu_ib_get(ring, NULL, ndw * 4, ib);
477 /* walk over the address space and update the page directory */
478 for (pt_idx = 0; pt_idx <= vm->max_pde_used; ++pt_idx) {
479 struct amdgpu_bo *bo = vm->page_tables[pt_idx].bo;
485 pt = amdgpu_bo_gpu_offset(bo);
486 if (vm->page_tables[pt_idx].addr == pt)
488 vm->page_tables[pt_idx].addr = pt;
490 pde = pd_addr + pt_idx * 8;
491 if (((last_pde + 8 * count) != pde) ||
492 ((last_pt + incr * count) != pt)) {
495 amdgpu_vm_update_pages(adev, ib, last_pde,
496 last_pt, count, incr,
497 AMDGPU_PTE_VALID, 0);
509 amdgpu_vm_update_pages(adev, ib, last_pde, last_pt, count,
510 incr, AMDGPU_PTE_VALID, 0);
512 if (ib->length_dw != 0) {
513 amdgpu_vm_pad_ib(adev, ib);
514 amdgpu_sync_resv(adev, &ib->sync, pd->tbo.resv, AMDGPU_FENCE_OWNER_VM);
515 WARN_ON(ib->length_dw > ndw);
517 if (amdgpu_enable_scheduler) {
519 sched_job = amdgpu_cs_parser_create(adev, AMDGPU_FENCE_OWNER_VM,
524 sched_job->job_param.vm.bo = pd;
525 sched_job->run_job = amdgpu_vm_run_job;
526 sched_job->free_job = amdgpu_vm_free_job;
527 ib->sequence = amd_sched_push_job(ring->scheduler,
528 &adev->kernel_ctx->rings[ring->idx].c_entity,
530 r = amd_sched_wait_emit(&adev->kernel_ctx->rings[ring->idx].c_entity,
531 ib->sequence, false, -1);
533 DRM_ERROR("emit timeout\n");
535 r = amdgpu_ib_schedule(adev, 1, ib, AMDGPU_FENCE_OWNER_VM);
537 amdgpu_ib_free(adev, ib);
540 amdgpu_bo_fence(pd, &ib->fence->base, true);
544 if (!amdgpu_enable_scheduler || ib->length_dw == 0) {
545 amdgpu_ib_free(adev, ib);
554 amdgpu_ib_free(adev, ib);
560 * amdgpu_vm_frag_ptes - add fragment information to PTEs
562 * @adev: amdgpu_device pointer
563 * @ib: IB for the update
564 * @pe_start: first PTE to handle
565 * @pe_end: last PTE to handle
566 * @addr: addr those PTEs should point to
567 * @flags: hw mapping flags
568 * @gtt_flags: GTT hw mapping flags
570 * Global and local mutex must be locked!
572 static void amdgpu_vm_frag_ptes(struct amdgpu_device *adev,
573 struct amdgpu_ib *ib,
574 uint64_t pe_start, uint64_t pe_end,
575 uint64_t addr, uint32_t flags,
579 * The MC L1 TLB supports variable sized pages, based on a fragment
580 * field in the PTE. When this field is set to a non-zero value, page
581 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
582 * flags are considered valid for all PTEs within the fragment range
583 * and corresponding mappings are assumed to be physically contiguous.
585 * The L1 TLB can store a single PTE for the whole fragment,
586 * significantly increasing the space available for translation
587 * caching. This leads to large improvements in throughput when the
588 * TLB is under pressure.
590 * The L2 TLB distributes small and large fragments into two
591 * asymmetric partitions. The large fragment cache is significantly
592 * larger. Thus, we try to use large fragments wherever possible.
593 * Userspace can support this by aligning virtual base address and
594 * allocation size to the fragment size.
597 /* SI and newer are optimized for 64KB */
598 uint64_t frag_flags = AMDGPU_PTE_FRAG_64KB;
599 uint64_t frag_align = 0x80;
601 uint64_t frag_start = ALIGN(pe_start, frag_align);
602 uint64_t frag_end = pe_end & ~(frag_align - 1);
606 /* system pages are non continuously */
607 if ((flags & AMDGPU_PTE_SYSTEM) || !(flags & AMDGPU_PTE_VALID) ||
608 (frag_start >= frag_end)) {
610 count = (pe_end - pe_start) / 8;
611 amdgpu_vm_update_pages(adev, ib, pe_start, addr, count,
612 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
616 /* handle the 4K area at the beginning */
617 if (pe_start != frag_start) {
618 count = (frag_start - pe_start) / 8;
619 amdgpu_vm_update_pages(adev, ib, pe_start, addr, count,
620 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
621 addr += AMDGPU_GPU_PAGE_SIZE * count;
624 /* handle the area in the middle */
625 count = (frag_end - frag_start) / 8;
626 amdgpu_vm_update_pages(adev, ib, frag_start, addr, count,
627 AMDGPU_GPU_PAGE_SIZE, flags | frag_flags,
630 /* handle the 4K area at the end */
631 if (frag_end != pe_end) {
632 addr += AMDGPU_GPU_PAGE_SIZE * count;
633 count = (pe_end - frag_end) / 8;
634 amdgpu_vm_update_pages(adev, ib, frag_end, addr, count,
635 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
640 * amdgpu_vm_update_ptes - make sure that page tables are valid
642 * @adev: amdgpu_device pointer
644 * @start: start of GPU address range
645 * @end: end of GPU address range
646 * @dst: destination address to map to
647 * @flags: mapping flags
649 * Update the page tables in the range @start - @end (cayman+).
651 * Global and local mutex must be locked!
653 static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
654 struct amdgpu_vm *vm,
655 struct amdgpu_ib *ib,
656 uint64_t start, uint64_t end,
657 uint64_t dst, uint32_t flags,
660 uint64_t mask = AMDGPU_VM_PTE_COUNT - 1;
661 uint64_t last_pte = ~0, last_dst = ~0;
665 /* walk over the address space and update the page tables */
666 for (addr = start; addr < end; ) {
667 uint64_t pt_idx = addr >> amdgpu_vm_block_size;
668 struct amdgpu_bo *pt = vm->page_tables[pt_idx].bo;
673 amdgpu_sync_resv(adev, &ib->sync, pt->tbo.resv,
674 AMDGPU_FENCE_OWNER_VM);
675 r = reservation_object_reserve_shared(pt->tbo.resv);
679 if ((addr & ~mask) == (end & ~mask))
682 nptes = AMDGPU_VM_PTE_COUNT - (addr & mask);
684 pte = amdgpu_bo_gpu_offset(pt);
685 pte += (addr & mask) * 8;
687 if ((last_pte + 8 * count) != pte) {
690 amdgpu_vm_frag_ptes(adev, ib, last_pte,
691 last_pte + 8 * count,
704 dst += nptes * AMDGPU_GPU_PAGE_SIZE;
708 amdgpu_vm_frag_ptes(adev, ib, last_pte,
709 last_pte + 8 * count,
710 last_dst, flags, gtt_flags);
717 * amdgpu_vm_fence_pts - fence page tables after an update
720 * @start: start of GPU address range
721 * @end: end of GPU address range
722 * @fence: fence to use
724 * Fence the page tables in the range @start - @end (cayman+).
726 * Global and local mutex must be locked!
728 static void amdgpu_vm_fence_pts(struct amdgpu_vm *vm,
729 uint64_t start, uint64_t end,
734 start >>= amdgpu_vm_block_size;
735 end >>= amdgpu_vm_block_size;
737 for (i = start; i <= end; ++i)
738 amdgpu_bo_fence(vm->page_tables[i].bo, fence, true);
741 static int amdgpu_vm_bo_update_mapping_run_job(
742 struct amdgpu_cs_parser *sched_job)
744 struct fence **fence = sched_job->job_param.vm_mapping.fence;
745 amdgpu_vm_fence_pts(sched_job->job_param.vm_mapping.vm,
746 sched_job->job_param.vm_mapping.start,
747 sched_job->job_param.vm_mapping.last + 1,
748 &sched_job->ibs[sched_job->num_ibs -1].fence->base);
751 *fence = fence_get(&sched_job->ibs[sched_job->num_ibs -1].fence->base);
756 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
758 * @adev: amdgpu_device pointer
760 * @mapping: mapped range and flags to use for the update
761 * @addr: addr to set the area to
762 * @gtt_flags: flags as they are used for GTT
763 * @fence: optional resulting fence
765 * Fill in the page table entries for @mapping.
766 * Returns 0 for success, -EINVAL for failure.
768 * Object have to be reserved and mutex must be locked!
770 static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
771 struct amdgpu_vm *vm,
772 struct amdgpu_bo_va_mapping *mapping,
773 uint64_t addr, uint32_t gtt_flags,
774 struct fence **fence)
776 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
777 unsigned nptes, ncmds, ndw;
778 uint32_t flags = gtt_flags;
779 struct amdgpu_ib *ib;
780 struct amdgpu_cs_parser *sched_job = NULL;
783 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
784 * but in case of something, we filter the flags in first place
786 if (!(mapping->flags & AMDGPU_PTE_READABLE))
787 flags &= ~AMDGPU_PTE_READABLE;
788 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
789 flags &= ~AMDGPU_PTE_WRITEABLE;
791 trace_amdgpu_vm_bo_update(mapping);
793 nptes = mapping->it.last - mapping->it.start + 1;
796 * reserve space for one command every (1 << BLOCK_SIZE)
797 * entries or 2k dwords (whatever is smaller)
799 ncmds = (nptes >> min(amdgpu_vm_block_size, 11)) + 1;
804 if ((flags & AMDGPU_PTE_SYSTEM) && (flags == gtt_flags)) {
805 /* only copy commands needed */
808 } else if (flags & AMDGPU_PTE_SYSTEM) {
809 /* header for write data commands */
812 /* body of write data command */
816 /* set page commands needed */
819 /* two extra commands for begin/end of fragment */
823 /* update too big for an IB */
827 ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
831 r = amdgpu_ib_get(ring, NULL, ndw * 4, ib);
839 if (!(flags & AMDGPU_PTE_VALID)) {
842 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
843 struct amdgpu_fence *f = vm->ids[i].last_id_use;
844 r = amdgpu_sync_fence(adev, &ib->sync, &f->base);
850 r = amdgpu_vm_update_ptes(adev, vm, ib, mapping->it.start,
851 mapping->it.last + 1, addr + mapping->offset,
855 amdgpu_ib_free(adev, ib);
860 amdgpu_vm_pad_ib(adev, ib);
861 WARN_ON(ib->length_dw > ndw);
863 if (amdgpu_enable_scheduler) {
865 sched_job = amdgpu_cs_parser_create(adev, AMDGPU_FENCE_OWNER_VM,
866 adev->kernel_ctx, ib, 1);
869 sched_job->job_param.vm_mapping.vm = vm;
870 sched_job->job_param.vm_mapping.start = mapping->it.start;
871 sched_job->job_param.vm_mapping.last = mapping->it.last;
872 sched_job->job_param.vm_mapping.fence = fence;
873 sched_job->run_job = amdgpu_vm_bo_update_mapping_run_job;
874 sched_job->free_job = amdgpu_vm_free_job;
875 ib->sequence = amd_sched_push_job(ring->scheduler,
876 &adev->kernel_ctx->rings[ring->idx].c_entity,
878 r = amd_sched_wait_emit(&adev->kernel_ctx->rings[ring->idx].c_entity,
879 ib->sequence, false, -1);
881 DRM_ERROR("emit timeout\n");
883 r = amdgpu_ib_schedule(adev, 1, ib, AMDGPU_FENCE_OWNER_VM);
885 amdgpu_ib_free(adev, ib);
889 amdgpu_vm_fence_pts(vm, mapping->it.start,
890 mapping->it.last + 1, &ib->fence->base);
893 *fence = fence_get(&ib->fence->base);
896 amdgpu_ib_free(adev, ib);
904 amdgpu_ib_free(adev, ib);
910 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
912 * @adev: amdgpu_device pointer
913 * @bo_va: requested BO and VM object
916 * Fill in the page table entries for @bo_va.
917 * Returns 0 for success, -EINVAL for failure.
919 * Object have to be reserved and mutex must be locked!
921 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
922 struct amdgpu_bo_va *bo_va,
923 struct ttm_mem_reg *mem)
925 struct amdgpu_vm *vm = bo_va->vm;
926 struct amdgpu_bo_va_mapping *mapping;
932 addr = mem->start << PAGE_SHIFT;
933 if (mem->mem_type != TTM_PL_TT)
934 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_update_mapping(adev, vm, mapping, addr,
948 flags, &bo_va->last_pt_update);
953 spin_lock(&vm->status_lock);
954 list_del_init(&bo_va->vm_status);
956 list_add(&bo_va->vm_status, &vm->cleared);
957 spin_unlock(&vm->status_lock);
963 * amdgpu_vm_clear_freed - clear freed BOs in the PT
965 * @adev: amdgpu_device pointer
968 * Make sure all freed BOs are cleared in the PT.
969 * Returns 0 for success.
971 * PTs have to be reserved and mutex must be locked!
973 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
974 struct amdgpu_vm *vm)
976 struct amdgpu_bo_va_mapping *mapping;
979 while (!list_empty(&vm->freed)) {
980 mapping = list_first_entry(&vm->freed,
981 struct amdgpu_bo_va_mapping, list);
982 list_del(&mapping->list);
984 r = amdgpu_vm_bo_update_mapping(adev, vm, mapping, 0, 0, NULL);
995 * amdgpu_vm_clear_invalids - clear invalidated BOs in the PT
997 * @adev: amdgpu_device pointer
1000 * Make sure all invalidated BOs are cleared in the PT.
1001 * Returns 0 for success.
1003 * PTs have to be reserved and mutex must be locked!
1005 int amdgpu_vm_clear_invalids(struct amdgpu_device *adev,
1006 struct amdgpu_vm *vm, struct amdgpu_sync *sync)
1008 struct amdgpu_bo_va *bo_va = NULL;
1011 spin_lock(&vm->status_lock);
1012 while (!list_empty(&vm->invalidated)) {
1013 bo_va = list_first_entry(&vm->invalidated,
1014 struct amdgpu_bo_va, vm_status);
1015 spin_unlock(&vm->status_lock);
1017 r = amdgpu_vm_bo_update(adev, bo_va, NULL);
1021 spin_lock(&vm->status_lock);
1023 spin_unlock(&vm->status_lock);
1026 r = amdgpu_sync_fence(adev, sync, bo_va->last_pt_update);
1032 * amdgpu_vm_bo_add - add a bo to a specific vm
1034 * @adev: amdgpu_device pointer
1036 * @bo: amdgpu buffer object
1038 * Add @bo into the requested vm (cayman+).
1039 * Add @bo to the list of bos associated with the vm
1040 * Returns newly added bo_va or NULL for failure
1042 * Object has to be reserved!
1044 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
1045 struct amdgpu_vm *vm,
1046 struct amdgpu_bo *bo)
1048 struct amdgpu_bo_va *bo_va;
1050 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
1051 if (bo_va == NULL) {
1056 bo_va->ref_count = 1;
1057 INIT_LIST_HEAD(&bo_va->bo_list);
1058 INIT_LIST_HEAD(&bo_va->valids);
1059 INIT_LIST_HEAD(&bo_va->invalids);
1060 INIT_LIST_HEAD(&bo_va->vm_status);
1062 mutex_lock(&vm->mutex);
1063 list_add_tail(&bo_va->bo_list, &bo->va);
1064 mutex_unlock(&vm->mutex);
1070 * amdgpu_vm_bo_map - map bo inside a vm
1072 * @adev: amdgpu_device pointer
1073 * @bo_va: bo_va to store the address
1074 * @saddr: where to map the BO
1075 * @offset: requested offset in the BO
1076 * @flags: attributes of pages (read/write/valid/etc.)
1078 * Add a mapping of the BO at the specefied addr into the VM.
1079 * Returns 0 for success, error for failure.
1081 * Object has to be reserved and gets unreserved by this function!
1083 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
1084 struct amdgpu_bo_va *bo_va,
1085 uint64_t saddr, uint64_t offset,
1086 uint64_t size, uint32_t flags)
1088 struct amdgpu_bo_va_mapping *mapping;
1089 struct amdgpu_vm *vm = bo_va->vm;
1090 struct interval_tree_node *it;
1091 unsigned last_pfn, pt_idx;
1095 /* validate the parameters */
1096 if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
1097 size == 0 || size & AMDGPU_GPU_PAGE_MASK) {
1098 amdgpu_bo_unreserve(bo_va->bo);
1102 /* make sure object fit at this offset */
1103 eaddr = saddr + size;
1104 if ((saddr >= eaddr) || (offset + size > amdgpu_bo_size(bo_va->bo))) {
1105 amdgpu_bo_unreserve(bo_va->bo);
1109 last_pfn = eaddr / AMDGPU_GPU_PAGE_SIZE;
1110 if (last_pfn > adev->vm_manager.max_pfn) {
1111 dev_err(adev->dev, "va above limit (0x%08X > 0x%08X)\n",
1112 last_pfn, adev->vm_manager.max_pfn);
1113 amdgpu_bo_unreserve(bo_va->bo);
1117 mutex_lock(&vm->mutex);
1119 saddr /= AMDGPU_GPU_PAGE_SIZE;
1120 eaddr /= AMDGPU_GPU_PAGE_SIZE;
1122 it = interval_tree_iter_first(&vm->va, saddr, eaddr - 1);
1124 struct amdgpu_bo_va_mapping *tmp;
1125 tmp = container_of(it, struct amdgpu_bo_va_mapping, it);
1126 /* bo and tmp overlap, invalid addr */
1127 dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
1128 "0x%010lx-0x%010lx\n", bo_va->bo, saddr, eaddr,
1129 tmp->it.start, tmp->it.last + 1);
1130 amdgpu_bo_unreserve(bo_va->bo);
1135 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1137 amdgpu_bo_unreserve(bo_va->bo);
1142 INIT_LIST_HEAD(&mapping->list);
1143 mapping->it.start = saddr;
1144 mapping->it.last = eaddr - 1;
1145 mapping->offset = offset;
1146 mapping->flags = flags;
1148 list_add(&mapping->list, &bo_va->invalids);
1149 interval_tree_insert(&mapping->it, &vm->va);
1150 trace_amdgpu_vm_bo_map(bo_va, mapping);
1152 /* Make sure the page tables are allocated */
1153 saddr >>= amdgpu_vm_block_size;
1154 eaddr >>= amdgpu_vm_block_size;
1156 BUG_ON(eaddr >= amdgpu_vm_num_pdes(adev));
1158 if (eaddr > vm->max_pde_used)
1159 vm->max_pde_used = eaddr;
1161 amdgpu_bo_unreserve(bo_va->bo);
1163 /* walk over the address space and allocate the page tables */
1164 for (pt_idx = saddr; pt_idx <= eaddr; ++pt_idx) {
1165 struct amdgpu_bo *pt;
1167 if (vm->page_tables[pt_idx].bo)
1170 /* drop mutex to allocate and clear page table */
1171 mutex_unlock(&vm->mutex);
1173 r = amdgpu_bo_create(adev, AMDGPU_VM_PTE_COUNT * 8,
1174 AMDGPU_GPU_PAGE_SIZE, true,
1175 AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &pt);
1179 r = amdgpu_vm_clear_bo(adev, pt);
1181 amdgpu_bo_unref(&pt);
1185 /* aquire mutex again */
1186 mutex_lock(&vm->mutex);
1187 if (vm->page_tables[pt_idx].bo) {
1188 /* someone else allocated the pt in the meantime */
1189 mutex_unlock(&vm->mutex);
1190 amdgpu_bo_unref(&pt);
1191 mutex_lock(&vm->mutex);
1195 vm->page_tables[pt_idx].addr = 0;
1196 vm->page_tables[pt_idx].bo = pt;
1199 mutex_unlock(&vm->mutex);
1203 mutex_lock(&vm->mutex);
1204 list_del(&mapping->list);
1205 interval_tree_remove(&mapping->it, &vm->va);
1206 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1210 mutex_unlock(&vm->mutex);
1215 * amdgpu_vm_bo_unmap - remove bo mapping from vm
1217 * @adev: amdgpu_device pointer
1218 * @bo_va: bo_va to remove the address from
1219 * @saddr: where to the BO is mapped
1221 * Remove a mapping of the BO at the specefied addr from the VM.
1222 * Returns 0 for success, error for failure.
1224 * Object has to be reserved and gets unreserved by this function!
1226 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
1227 struct amdgpu_bo_va *bo_va,
1230 struct amdgpu_bo_va_mapping *mapping;
1231 struct amdgpu_vm *vm = bo_va->vm;
1234 saddr /= AMDGPU_GPU_PAGE_SIZE;
1236 list_for_each_entry(mapping, &bo_va->valids, list) {
1237 if (mapping->it.start == saddr)
1241 if (&mapping->list == &bo_va->valids) {
1244 list_for_each_entry(mapping, &bo_va->invalids, list) {
1245 if (mapping->it.start == saddr)
1249 if (&mapping->list == &bo_va->invalids) {
1250 amdgpu_bo_unreserve(bo_va->bo);
1255 mutex_lock(&vm->mutex);
1256 list_del(&mapping->list);
1257 interval_tree_remove(&mapping->it, &vm->va);
1258 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1261 list_add(&mapping->list, &vm->freed);
1264 mutex_unlock(&vm->mutex);
1265 amdgpu_bo_unreserve(bo_va->bo);
1271 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
1273 * @adev: amdgpu_device pointer
1274 * @bo_va: requested bo_va
1276 * Remove @bo_va->bo from the requested vm (cayman+).
1278 * Object have to be reserved!
1280 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
1281 struct amdgpu_bo_va *bo_va)
1283 struct amdgpu_bo_va_mapping *mapping, *next;
1284 struct amdgpu_vm *vm = bo_va->vm;
1286 list_del(&bo_va->bo_list);
1288 mutex_lock(&vm->mutex);
1290 spin_lock(&vm->status_lock);
1291 list_del(&bo_va->vm_status);
1292 spin_unlock(&vm->status_lock);
1294 list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
1295 list_del(&mapping->list);
1296 interval_tree_remove(&mapping->it, &vm->va);
1297 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1298 list_add(&mapping->list, &vm->freed);
1300 list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
1301 list_del(&mapping->list);
1302 interval_tree_remove(&mapping->it, &vm->va);
1306 fence_put(bo_va->last_pt_update);
1309 mutex_unlock(&vm->mutex);
1313 * amdgpu_vm_bo_invalidate - mark the bo as invalid
1315 * @adev: amdgpu_device pointer
1317 * @bo: amdgpu buffer object
1319 * Mark @bo as invalid (cayman+).
1321 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
1322 struct amdgpu_bo *bo)
1324 struct amdgpu_bo_va *bo_va;
1326 list_for_each_entry(bo_va, &bo->va, bo_list) {
1327 spin_lock(&bo_va->vm->status_lock);
1328 if (list_empty(&bo_va->vm_status))
1329 list_add(&bo_va->vm_status, &bo_va->vm->invalidated);
1330 spin_unlock(&bo_va->vm->status_lock);
1335 * amdgpu_vm_init - initialize a vm instance
1337 * @adev: amdgpu_device pointer
1340 * Init @vm fields (cayman+).
1342 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1344 const unsigned align = min(AMDGPU_VM_PTB_ALIGN_SIZE,
1345 AMDGPU_VM_PTE_COUNT * 8);
1346 unsigned pd_size, pd_entries, pts_size;
1349 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1351 vm->ids[i].flushed_updates = NULL;
1352 vm->ids[i].last_id_use = NULL;
1354 mutex_init(&vm->mutex);
1356 spin_lock_init(&vm->status_lock);
1357 INIT_LIST_HEAD(&vm->invalidated);
1358 INIT_LIST_HEAD(&vm->cleared);
1359 INIT_LIST_HEAD(&vm->freed);
1361 pd_size = amdgpu_vm_directory_size(adev);
1362 pd_entries = amdgpu_vm_num_pdes(adev);
1364 /* allocate page table array */
1365 pts_size = pd_entries * sizeof(struct amdgpu_vm_pt);
1366 vm->page_tables = kzalloc(pts_size, GFP_KERNEL);
1367 if (vm->page_tables == NULL) {
1368 DRM_ERROR("Cannot allocate memory for page table array\n");
1372 r = amdgpu_bo_create(adev, pd_size, align, true,
1373 AMDGPU_GEM_DOMAIN_VRAM, 0,
1374 NULL, &vm->page_directory);
1378 r = amdgpu_vm_clear_bo(adev, vm->page_directory);
1380 amdgpu_bo_unref(&vm->page_directory);
1381 vm->page_directory = NULL;
1389 * amdgpu_vm_fini - tear down a vm instance
1391 * @adev: amdgpu_device pointer
1394 * Tear down @vm (cayman+).
1395 * Unbind the VM and remove all bos from the vm bo list
1397 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1399 struct amdgpu_bo_va_mapping *mapping, *tmp;
1402 if (!RB_EMPTY_ROOT(&vm->va)) {
1403 dev_err(adev->dev, "still active bo inside vm\n");
1405 rbtree_postorder_for_each_entry_safe(mapping, tmp, &vm->va, it.rb) {
1406 list_del(&mapping->list);
1407 interval_tree_remove(&mapping->it, &vm->va);
1410 list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
1411 list_del(&mapping->list);
1415 for (i = 0; i < amdgpu_vm_num_pdes(adev); i++)
1416 amdgpu_bo_unref(&vm->page_tables[i].bo);
1417 kfree(vm->page_tables);
1419 amdgpu_bo_unref(&vm->page_directory);
1421 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1422 amdgpu_fence_unref(&vm->ids[i].flushed_updates);
1423 amdgpu_fence_unref(&vm->ids[i].last_id_use);
1426 mutex_destroy(&vm->mutex);