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 * @ring: ring we want to submit job to
131 * @vm: vm to allocate id for
133 * Allocate an id for the vm (cayman+).
134 * Returns the fence we need to sync to (if any).
136 * Global and local mutex must be locked!
138 struct amdgpu_fence *amdgpu_vm_grab_id(struct amdgpu_ring *ring,
139 struct amdgpu_vm *vm)
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 vm_id->id = choices[i];
176 trace_amdgpu_vm_grab_id(choices[i], ring->idx);
177 return adev->vm_manager.active[choices[i]];
181 /* should never happen */
187 * amdgpu_vm_flush - hardware flush the vm
189 * @ring: ring to use for flush
190 * @vm: vm we want to flush
191 * @updates: last vm update that we waited for
193 * Flush the vm (cayman+).
195 * Global and local mutex must be locked!
197 void amdgpu_vm_flush(struct amdgpu_ring *ring,
198 struct amdgpu_vm *vm,
199 struct amdgpu_fence *updates)
201 uint64_t pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
202 struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx];
203 struct amdgpu_fence *flushed_updates = vm_id->flushed_updates;
205 if (pd_addr != vm_id->pd_gpu_addr || !flushed_updates ||
206 (updates && amdgpu_fence_is_earlier(flushed_updates, updates))) {
208 trace_amdgpu_vm_flush(pd_addr, ring->idx, vm_id->id);
209 vm_id->flushed_updates = amdgpu_fence_ref(
210 amdgpu_fence_later(flushed_updates, updates));
211 amdgpu_fence_unref(&flushed_updates);
212 vm_id->pd_gpu_addr = pd_addr;
213 amdgpu_ring_emit_vm_flush(ring, vm_id->id, vm_id->pd_gpu_addr);
218 * amdgpu_vm_fence - remember fence for vm
220 * @adev: amdgpu_device pointer
221 * @vm: vm we want to fence
222 * @fence: fence to remember
224 * Fence the vm (cayman+).
225 * Set the fence used to protect page table and id.
227 * Global and local mutex must be locked!
229 void amdgpu_vm_fence(struct amdgpu_device *adev,
230 struct amdgpu_vm *vm,
231 struct amdgpu_fence *fence)
233 unsigned ridx = fence->ring->idx;
234 unsigned vm_id = vm->ids[ridx].id;
236 amdgpu_fence_unref(&adev->vm_manager.active[vm_id]);
237 adev->vm_manager.active[vm_id] = amdgpu_fence_ref(fence);
239 amdgpu_fence_unref(&vm->ids[ridx].last_id_use);
240 vm->ids[ridx].last_id_use = amdgpu_fence_ref(fence);
244 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
247 * @bo: requested buffer object
249 * Find @bo inside the requested vm (cayman+).
250 * Search inside the @bos vm list for the requested vm
251 * Returns the found bo_va or NULL if none is found
253 * Object has to be reserved!
255 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
256 struct amdgpu_bo *bo)
258 struct amdgpu_bo_va *bo_va;
260 list_for_each_entry(bo_va, &bo->va, bo_list) {
261 if (bo_va->vm == vm) {
269 * amdgpu_vm_update_pages - helper to call the right asic function
271 * @adev: amdgpu_device pointer
272 * @ib: indirect buffer to fill with commands
273 * @pe: addr of the page entry
274 * @addr: dst addr to write into pe
275 * @count: number of page entries to update
276 * @incr: increase next addr by incr bytes
277 * @flags: hw access flags
278 * @gtt_flags: GTT hw access flags
280 * Traces the parameters and calls the right asic functions
281 * to setup the page table using the DMA.
283 static void amdgpu_vm_update_pages(struct amdgpu_device *adev,
284 struct amdgpu_ib *ib,
285 uint64_t pe, uint64_t addr,
286 unsigned count, uint32_t incr,
287 uint32_t flags, uint32_t gtt_flags)
289 trace_amdgpu_vm_set_page(pe, addr, count, incr, flags);
291 if ((flags & AMDGPU_PTE_SYSTEM) && (flags == gtt_flags)) {
292 uint64_t src = adev->gart.table_addr + (addr >> 12) * 8;
293 amdgpu_vm_copy_pte(adev, ib, pe, src, count);
295 } else if ((flags & AMDGPU_PTE_SYSTEM) || (count < 3)) {
296 amdgpu_vm_write_pte(adev, ib, pe, addr,
300 amdgpu_vm_set_pte_pde(adev, ib, pe, addr,
306 * amdgpu_vm_clear_bo - initially clear the page dir/table
308 * @adev: amdgpu_device pointer
311 static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
312 struct amdgpu_bo *bo)
314 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
320 r = amdgpu_bo_reserve(bo, false);
324 r = reservation_object_reserve_shared(bo->tbo.resv);
328 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
330 goto error_unreserve;
332 addr = amdgpu_bo_gpu_offset(bo);
333 entries = amdgpu_bo_size(bo) / 8;
335 r = amdgpu_ib_get(ring, NULL, entries * 2 + 64, &ib);
337 goto error_unreserve;
341 amdgpu_vm_update_pages(adev, &ib, addr, 0, entries, 0, 0, 0);
342 amdgpu_vm_pad_ib(adev, &ib);
343 WARN_ON(ib.length_dw > 64);
345 r = amdgpu_ib_schedule(adev, 1, &ib, AMDGPU_FENCE_OWNER_VM);
349 amdgpu_bo_fence(bo, ib.fence, true);
352 amdgpu_ib_free(adev, &ib);
355 amdgpu_bo_unreserve(bo);
360 * amdgpu_vm_map_gart - get the physical address of a gart page
362 * @adev: amdgpu_device pointer
363 * @addr: the unmapped addr
365 * Look up the physical address of the page that the pte resolves
367 * Returns the physical address of the page.
369 uint64_t amdgpu_vm_map_gart(struct amdgpu_device *adev, uint64_t addr)
373 /* page table offset */
374 result = adev->gart.pages_addr[addr >> PAGE_SHIFT];
376 /* in case cpu page size != gpu page size*/
377 result |= addr & (~PAGE_MASK);
383 * amdgpu_vm_update_pdes - make sure that page directory is valid
385 * @adev: amdgpu_device pointer
387 * @start: start of GPU address range
388 * @end: end of GPU address range
390 * Allocates new page tables if necessary
391 * and updates the page directory (cayman+).
392 * Returns 0 for success, error for failure.
394 * Global and local mutex must be locked!
396 int amdgpu_vm_update_page_directory(struct amdgpu_device *adev,
397 struct amdgpu_vm *vm)
399 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
400 struct amdgpu_bo *pd = vm->page_directory;
401 uint64_t pd_addr = amdgpu_bo_gpu_offset(pd);
402 uint32_t incr = AMDGPU_VM_PTE_COUNT * 8;
403 uint64_t last_pde = ~0, last_pt = ~0;
404 unsigned count = 0, pt_idx, ndw;
411 /* assume the worst case */
412 ndw += vm->max_pde_used * 6;
414 /* update too big for an IB */
418 r = amdgpu_ib_get(ring, NULL, ndw * 4, &ib);
423 /* walk over the address space and update the page directory */
424 for (pt_idx = 0; pt_idx <= vm->max_pde_used; ++pt_idx) {
425 struct amdgpu_bo *bo = vm->page_tables[pt_idx].bo;
431 pt = amdgpu_bo_gpu_offset(bo);
432 if (vm->page_tables[pt_idx].addr == pt)
434 vm->page_tables[pt_idx].addr = pt;
436 pde = pd_addr + pt_idx * 8;
437 if (((last_pde + 8 * count) != pde) ||
438 ((last_pt + incr * count) != pt)) {
441 amdgpu_vm_update_pages(adev, &ib, last_pde,
442 last_pt, count, incr,
443 AMDGPU_PTE_VALID, 0);
455 amdgpu_vm_update_pages(adev, &ib, last_pde, last_pt, count,
456 incr, AMDGPU_PTE_VALID, 0);
458 if (ib.length_dw != 0) {
459 amdgpu_vm_pad_ib(adev, &ib);
460 amdgpu_sync_resv(adev, &ib.sync, pd->tbo.resv, AMDGPU_FENCE_OWNER_VM);
461 WARN_ON(ib.length_dw > ndw);
462 r = amdgpu_ib_schedule(adev, 1, &ib, AMDGPU_FENCE_OWNER_VM);
464 amdgpu_ib_free(adev, &ib);
467 amdgpu_bo_fence(pd, ib.fence, true);
469 amdgpu_ib_free(adev, &ib);
475 * amdgpu_vm_frag_ptes - add fragment information to PTEs
477 * @adev: amdgpu_device pointer
478 * @ib: IB for the update
479 * @pe_start: first PTE to handle
480 * @pe_end: last PTE to handle
481 * @addr: addr those PTEs should point to
482 * @flags: hw mapping flags
483 * @gtt_flags: GTT hw mapping flags
485 * Global and local mutex must be locked!
487 static void amdgpu_vm_frag_ptes(struct amdgpu_device *adev,
488 struct amdgpu_ib *ib,
489 uint64_t pe_start, uint64_t pe_end,
490 uint64_t addr, uint32_t flags,
494 * The MC L1 TLB supports variable sized pages, based on a fragment
495 * field in the PTE. When this field is set to a non-zero value, page
496 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
497 * flags are considered valid for all PTEs within the fragment range
498 * and corresponding mappings are assumed to be physically contiguous.
500 * The L1 TLB can store a single PTE for the whole fragment,
501 * significantly increasing the space available for translation
502 * caching. This leads to large improvements in throughput when the
503 * TLB is under pressure.
505 * The L2 TLB distributes small and large fragments into two
506 * asymmetric partitions. The large fragment cache is significantly
507 * larger. Thus, we try to use large fragments wherever possible.
508 * Userspace can support this by aligning virtual base address and
509 * allocation size to the fragment size.
512 /* SI and newer are optimized for 64KB */
513 uint64_t frag_flags = AMDGPU_PTE_FRAG_64KB;
514 uint64_t frag_align = 0x80;
516 uint64_t frag_start = ALIGN(pe_start, frag_align);
517 uint64_t frag_end = pe_end & ~(frag_align - 1);
521 /* system pages are non continuously */
522 if ((flags & AMDGPU_PTE_SYSTEM) || !(flags & AMDGPU_PTE_VALID) ||
523 (frag_start >= frag_end)) {
525 count = (pe_end - pe_start) / 8;
526 amdgpu_vm_update_pages(adev, ib, pe_start, addr, count,
527 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
531 /* handle the 4K area at the beginning */
532 if (pe_start != frag_start) {
533 count = (frag_start - pe_start) / 8;
534 amdgpu_vm_update_pages(adev, ib, pe_start, addr, count,
535 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
536 addr += AMDGPU_GPU_PAGE_SIZE * count;
539 /* handle the area in the middle */
540 count = (frag_end - frag_start) / 8;
541 amdgpu_vm_update_pages(adev, ib, frag_start, addr, count,
542 AMDGPU_GPU_PAGE_SIZE, flags | frag_flags,
545 /* handle the 4K area at the end */
546 if (frag_end != pe_end) {
547 addr += AMDGPU_GPU_PAGE_SIZE * count;
548 count = (pe_end - frag_end) / 8;
549 amdgpu_vm_update_pages(adev, ib, frag_end, addr, count,
550 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
555 * amdgpu_vm_update_ptes - make sure that page tables are valid
557 * @adev: amdgpu_device pointer
559 * @start: start of GPU address range
560 * @end: end of GPU address range
561 * @dst: destination address to map to
562 * @flags: mapping flags
564 * Update the page tables in the range @start - @end (cayman+).
566 * Global and local mutex must be locked!
568 static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
569 struct amdgpu_vm *vm,
570 struct amdgpu_ib *ib,
571 uint64_t start, uint64_t end,
572 uint64_t dst, uint32_t flags,
575 uint64_t mask = AMDGPU_VM_PTE_COUNT - 1;
576 uint64_t last_pte = ~0, last_dst = ~0;
580 /* walk over the address space and update the page tables */
581 for (addr = start; addr < end; ) {
582 uint64_t pt_idx = addr >> amdgpu_vm_block_size;
583 struct amdgpu_bo *pt = vm->page_tables[pt_idx].bo;
588 amdgpu_sync_resv(adev, &ib->sync, pt->tbo.resv,
589 AMDGPU_FENCE_OWNER_VM);
590 r = reservation_object_reserve_shared(pt->tbo.resv);
594 if ((addr & ~mask) == (end & ~mask))
597 nptes = AMDGPU_VM_PTE_COUNT - (addr & mask);
599 pte = amdgpu_bo_gpu_offset(pt);
600 pte += (addr & mask) * 8;
602 if ((last_pte + 8 * count) != pte) {
605 amdgpu_vm_frag_ptes(adev, ib, last_pte,
606 last_pte + 8 * count,
619 dst += nptes * AMDGPU_GPU_PAGE_SIZE;
623 amdgpu_vm_frag_ptes(adev, ib, last_pte,
624 last_pte + 8 * count,
625 last_dst, flags, gtt_flags);
632 * amdgpu_vm_fence_pts - fence page tables after an update
635 * @start: start of GPU address range
636 * @end: end of GPU address range
637 * @fence: fence to use
639 * Fence the page tables in the range @start - @end (cayman+).
641 * Global and local mutex must be locked!
643 static void amdgpu_vm_fence_pts(struct amdgpu_vm *vm,
644 uint64_t start, uint64_t end,
645 struct amdgpu_fence *fence)
649 start >>= amdgpu_vm_block_size;
650 end >>= amdgpu_vm_block_size;
652 for (i = start; i <= end; ++i)
653 amdgpu_bo_fence(vm->page_tables[i].bo, fence, true);
657 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
659 * @adev: amdgpu_device pointer
661 * @mapping: mapped range and flags to use for the update
662 * @addr: addr to set the area to
663 * @gtt_flags: flags as they are used for GTT
664 * @fence: optional resulting fence
666 * Fill in the page table entries for @mapping.
667 * Returns 0 for success, -EINVAL for failure.
669 * Object have to be reserved and mutex must be locked!
671 static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
672 struct amdgpu_vm *vm,
673 struct amdgpu_bo_va_mapping *mapping,
674 uint64_t addr, uint32_t gtt_flags,
675 struct amdgpu_fence **fence)
677 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
678 unsigned nptes, ncmds, ndw;
679 uint32_t flags = gtt_flags;
683 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
684 * but in case of something, we filter the flags in first place
686 if (!(mapping->flags & AMDGPU_PTE_READABLE))
687 flags &= ~AMDGPU_PTE_READABLE;
688 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
689 flags &= ~AMDGPU_PTE_WRITEABLE;
691 trace_amdgpu_vm_bo_update(mapping);
693 nptes = mapping->it.last - mapping->it.start + 1;
696 * reserve space for one command every (1 << BLOCK_SIZE)
697 * entries or 2k dwords (whatever is smaller)
699 ncmds = (nptes >> min(amdgpu_vm_block_size, 11)) + 1;
704 if ((flags & AMDGPU_PTE_SYSTEM) && (flags == gtt_flags)) {
705 /* only copy commands needed */
708 } else if (flags & AMDGPU_PTE_SYSTEM) {
709 /* header for write data commands */
712 /* body of write data command */
716 /* set page commands needed */
719 /* two extra commands for begin/end of fragment */
723 /* update too big for an IB */
727 r = amdgpu_ib_get(ring, NULL, ndw * 4, &ib);
732 if (!(flags & AMDGPU_PTE_VALID)) {
735 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
736 struct amdgpu_fence *f = vm->ids[i].last_id_use;
737 r = amdgpu_sync_fence(adev, &ib.sync, &f->base);
743 r = amdgpu_vm_update_ptes(adev, vm, &ib, mapping->it.start,
744 mapping->it.last + 1, addr + mapping->offset,
748 amdgpu_ib_free(adev, &ib);
752 amdgpu_vm_pad_ib(adev, &ib);
753 WARN_ON(ib.length_dw > ndw);
755 r = amdgpu_ib_schedule(adev, 1, &ib, AMDGPU_FENCE_OWNER_VM);
757 amdgpu_ib_free(adev, &ib);
760 amdgpu_vm_fence_pts(vm, mapping->it.start,
761 mapping->it.last + 1, ib.fence);
763 amdgpu_fence_unref(fence);
764 *fence = amdgpu_fence_ref(ib.fence);
766 amdgpu_ib_free(adev, &ib);
772 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
774 * @adev: amdgpu_device pointer
775 * @bo_va: requested BO and VM object
778 * Fill in the page table entries for @bo_va.
779 * Returns 0 for success, -EINVAL for failure.
781 * Object have to be reserved and mutex must be locked!
783 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
784 struct amdgpu_bo_va *bo_va,
785 struct ttm_mem_reg *mem)
787 struct amdgpu_vm *vm = bo_va->vm;
788 struct amdgpu_bo_va_mapping *mapping;
794 addr = mem->start << PAGE_SHIFT;
795 if (mem->mem_type != TTM_PL_TT)
796 addr += adev->vm_manager.vram_base_offset;
801 if (addr == bo_va->addr)
804 flags = amdgpu_ttm_tt_pte_flags(adev, bo_va->bo->tbo.ttm, mem);
806 list_for_each_entry(mapping, &bo_va->mappings, list) {
807 r = amdgpu_vm_bo_update_mapping(adev, vm, mapping, addr,
808 flags, &bo_va->last_pt_update);
814 spin_lock(&vm->status_lock);
815 list_del_init(&bo_va->vm_status);
816 spin_unlock(&vm->status_lock);
822 * amdgpu_vm_clear_freed - clear freed BOs in the PT
824 * @adev: amdgpu_device pointer
827 * Make sure all freed BOs are cleared in the PT.
828 * Returns 0 for success.
830 * PTs have to be reserved and mutex must be locked!
832 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
833 struct amdgpu_vm *vm)
835 struct amdgpu_bo_va_mapping *mapping;
838 while (!list_empty(&vm->freed)) {
839 mapping = list_first_entry(&vm->freed,
840 struct amdgpu_bo_va_mapping, list);
841 list_del(&mapping->list);
843 r = amdgpu_vm_bo_update_mapping(adev, vm, mapping, 0, 0, NULL);
854 * amdgpu_vm_clear_invalids - clear invalidated BOs in the PT
856 * @adev: amdgpu_device pointer
859 * Make sure all invalidated BOs are cleared in the PT.
860 * Returns 0 for success.
862 * PTs have to be reserved and mutex must be locked!
864 int amdgpu_vm_clear_invalids(struct amdgpu_device *adev,
865 struct amdgpu_vm *vm, struct amdgpu_sync *sync)
867 struct amdgpu_bo_va *bo_va = NULL;
870 spin_lock(&vm->status_lock);
871 while (!list_empty(&vm->invalidated)) {
872 bo_va = list_first_entry(&vm->invalidated,
873 struct amdgpu_bo_va, vm_status);
874 spin_unlock(&vm->status_lock);
876 r = amdgpu_vm_bo_update(adev, bo_va, NULL);
880 spin_lock(&vm->status_lock);
882 spin_unlock(&vm->status_lock);
885 r = amdgpu_sync_fence(adev, sync, &bo_va->last_pt_update->base);
891 * amdgpu_vm_bo_add - add a bo to a specific vm
893 * @adev: amdgpu_device pointer
895 * @bo: amdgpu buffer object
897 * Add @bo into the requested vm (cayman+).
898 * Add @bo to the list of bos associated with the vm
899 * Returns newly added bo_va or NULL for failure
901 * Object has to be reserved!
903 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
904 struct amdgpu_vm *vm,
905 struct amdgpu_bo *bo)
907 struct amdgpu_bo_va *bo_va;
909 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
916 bo_va->ref_count = 1;
917 INIT_LIST_HEAD(&bo_va->bo_list);
918 INIT_LIST_HEAD(&bo_va->mappings);
919 INIT_LIST_HEAD(&bo_va->vm_status);
921 mutex_lock(&vm->mutex);
922 list_add_tail(&bo_va->bo_list, &bo->va);
923 mutex_unlock(&vm->mutex);
929 * amdgpu_vm_bo_map - map bo inside a vm
931 * @adev: amdgpu_device pointer
932 * @bo_va: bo_va to store the address
933 * @saddr: where to map the BO
934 * @offset: requested offset in the BO
935 * @flags: attributes of pages (read/write/valid/etc.)
937 * Add a mapping of the BO at the specefied addr into the VM.
938 * Returns 0 for success, error for failure.
940 * Object has to be reserved and gets unreserved by this function!
942 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
943 struct amdgpu_bo_va *bo_va,
944 uint64_t saddr, uint64_t offset,
945 uint64_t size, uint32_t flags)
947 struct amdgpu_bo_va_mapping *mapping;
948 struct amdgpu_vm *vm = bo_va->vm;
949 struct interval_tree_node *it;
950 unsigned last_pfn, pt_idx;
954 /* validate the parameters */
955 if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
956 size == 0 || size & AMDGPU_GPU_PAGE_MASK) {
957 amdgpu_bo_unreserve(bo_va->bo);
961 /* make sure object fit at this offset */
962 eaddr = saddr + size;
963 if ((saddr >= eaddr) || (offset + size > amdgpu_bo_size(bo_va->bo))) {
964 amdgpu_bo_unreserve(bo_va->bo);
968 last_pfn = eaddr / AMDGPU_GPU_PAGE_SIZE;
969 if (last_pfn > adev->vm_manager.max_pfn) {
970 dev_err(adev->dev, "va above limit (0x%08X > 0x%08X)\n",
971 last_pfn, adev->vm_manager.max_pfn);
972 amdgpu_bo_unreserve(bo_va->bo);
976 mutex_lock(&vm->mutex);
978 saddr /= AMDGPU_GPU_PAGE_SIZE;
979 eaddr /= AMDGPU_GPU_PAGE_SIZE;
981 it = interval_tree_iter_first(&vm->va, saddr, eaddr - 1);
983 struct amdgpu_bo_va_mapping *tmp;
984 tmp = container_of(it, struct amdgpu_bo_va_mapping, it);
985 /* bo and tmp overlap, invalid addr */
986 dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
987 "0x%010lx-0x%010lx\n", bo_va->bo, saddr, eaddr,
988 tmp->it.start, tmp->it.last + 1);
989 amdgpu_bo_unreserve(bo_va->bo);
994 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
996 amdgpu_bo_unreserve(bo_va->bo);
1001 INIT_LIST_HEAD(&mapping->list);
1002 mapping->it.start = saddr;
1003 mapping->it.last = eaddr - 1;
1004 mapping->offset = offset;
1005 mapping->flags = flags;
1007 list_add(&mapping->list, &bo_va->mappings);
1008 interval_tree_insert(&mapping->it, &vm->va);
1009 trace_amdgpu_vm_bo_map(bo_va, mapping);
1013 /* Make sure the page tables are allocated */
1014 saddr >>= amdgpu_vm_block_size;
1015 eaddr >>= amdgpu_vm_block_size;
1017 BUG_ON(eaddr >= amdgpu_vm_num_pdes(adev));
1019 if (eaddr > vm->max_pde_used)
1020 vm->max_pde_used = eaddr;
1022 amdgpu_bo_unreserve(bo_va->bo);
1024 /* walk over the address space and allocate the page tables */
1025 for (pt_idx = saddr; pt_idx <= eaddr; ++pt_idx) {
1026 struct amdgpu_bo *pt;
1028 if (vm->page_tables[pt_idx].bo)
1031 /* drop mutex to allocate and clear page table */
1032 mutex_unlock(&vm->mutex);
1034 r = amdgpu_bo_create(adev, AMDGPU_VM_PTE_COUNT * 8,
1035 AMDGPU_GPU_PAGE_SIZE, true,
1036 AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &pt);
1040 r = amdgpu_vm_clear_bo(adev, pt);
1042 amdgpu_bo_unref(&pt);
1046 /* aquire mutex again */
1047 mutex_lock(&vm->mutex);
1048 if (vm->page_tables[pt_idx].bo) {
1049 /* someone else allocated the pt in the meantime */
1050 mutex_unlock(&vm->mutex);
1051 amdgpu_bo_unref(&pt);
1052 mutex_lock(&vm->mutex);
1056 vm->page_tables[pt_idx].addr = 0;
1057 vm->page_tables[pt_idx].bo = pt;
1060 mutex_unlock(&vm->mutex);
1064 mutex_lock(&vm->mutex);
1065 list_del(&mapping->list);
1066 interval_tree_remove(&mapping->it, &vm->va);
1067 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1071 mutex_unlock(&vm->mutex);
1076 * amdgpu_vm_bo_unmap - remove bo mapping from vm
1078 * @adev: amdgpu_device pointer
1079 * @bo_va: bo_va to remove the address from
1080 * @saddr: where to the BO is mapped
1082 * Remove a mapping of the BO at the specefied addr from the VM.
1083 * Returns 0 for success, error for failure.
1085 * Object has to be reserved and gets unreserved by this function!
1087 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
1088 struct amdgpu_bo_va *bo_va,
1091 struct amdgpu_bo_va_mapping *mapping;
1092 struct amdgpu_vm *vm = bo_va->vm;
1094 saddr /= AMDGPU_GPU_PAGE_SIZE;
1096 list_for_each_entry(mapping, &bo_va->mappings, list) {
1097 if (mapping->it.start == saddr)
1101 if (&mapping->list == &bo_va->mappings) {
1102 amdgpu_bo_unreserve(bo_va->bo);
1106 mutex_lock(&vm->mutex);
1107 list_del(&mapping->list);
1108 interval_tree_remove(&mapping->it, &vm->va);
1109 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1112 /* clear the old address */
1113 list_add(&mapping->list, &vm->freed);
1117 mutex_unlock(&vm->mutex);
1118 amdgpu_bo_unreserve(bo_va->bo);
1124 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
1126 * @adev: amdgpu_device pointer
1127 * @bo_va: requested bo_va
1129 * Remove @bo_va->bo from the requested vm (cayman+).
1131 * Object have to be reserved!
1133 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
1134 struct amdgpu_bo_va *bo_va)
1136 struct amdgpu_bo_va_mapping *mapping, *next;
1137 struct amdgpu_vm *vm = bo_va->vm;
1139 list_del(&bo_va->bo_list);
1141 mutex_lock(&vm->mutex);
1143 spin_lock(&vm->status_lock);
1144 list_del(&bo_va->vm_status);
1145 spin_unlock(&vm->status_lock);
1147 list_for_each_entry_safe(mapping, next, &bo_va->mappings, list) {
1148 list_del(&mapping->list);
1149 interval_tree_remove(&mapping->it, &vm->va);
1150 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1152 list_add(&mapping->list, &vm->freed);
1156 amdgpu_fence_unref(&bo_va->last_pt_update);
1159 mutex_unlock(&vm->mutex);
1163 * amdgpu_vm_bo_invalidate - mark the bo as invalid
1165 * @adev: amdgpu_device pointer
1167 * @bo: amdgpu buffer object
1169 * Mark @bo as invalid (cayman+).
1171 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
1172 struct amdgpu_bo *bo)
1174 struct amdgpu_bo_va *bo_va;
1176 list_for_each_entry(bo_va, &bo->va, bo_list) {
1178 spin_lock(&bo_va->vm->status_lock);
1179 list_del(&bo_va->vm_status);
1180 list_add(&bo_va->vm_status, &bo_va->vm->invalidated);
1181 spin_unlock(&bo_va->vm->status_lock);
1187 * amdgpu_vm_init - initialize a vm instance
1189 * @adev: amdgpu_device pointer
1192 * Init @vm fields (cayman+).
1194 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1196 const unsigned align = min(AMDGPU_VM_PTB_ALIGN_SIZE,
1197 AMDGPU_VM_PTE_COUNT * 8);
1198 unsigned pd_size, pd_entries, pts_size;
1201 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1203 vm->ids[i].flushed_updates = NULL;
1204 vm->ids[i].last_id_use = NULL;
1206 mutex_init(&vm->mutex);
1208 spin_lock_init(&vm->status_lock);
1209 INIT_LIST_HEAD(&vm->invalidated);
1210 INIT_LIST_HEAD(&vm->freed);
1212 pd_size = amdgpu_vm_directory_size(adev);
1213 pd_entries = amdgpu_vm_num_pdes(adev);
1215 /* allocate page table array */
1216 pts_size = pd_entries * sizeof(struct amdgpu_vm_pt);
1217 vm->page_tables = kzalloc(pts_size, GFP_KERNEL);
1218 if (vm->page_tables == NULL) {
1219 DRM_ERROR("Cannot allocate memory for page table array\n");
1223 r = amdgpu_bo_create(adev, pd_size, align, true,
1224 AMDGPU_GEM_DOMAIN_VRAM, 0,
1225 NULL, &vm->page_directory);
1229 r = amdgpu_vm_clear_bo(adev, vm->page_directory);
1231 amdgpu_bo_unref(&vm->page_directory);
1232 vm->page_directory = NULL;
1240 * amdgpu_vm_fini - tear down a vm instance
1242 * @adev: amdgpu_device pointer
1245 * Tear down @vm (cayman+).
1246 * Unbind the VM and remove all bos from the vm bo list
1248 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1250 struct amdgpu_bo_va_mapping *mapping, *tmp;
1253 if (!RB_EMPTY_ROOT(&vm->va)) {
1254 dev_err(adev->dev, "still active bo inside vm\n");
1256 rbtree_postorder_for_each_entry_safe(mapping, tmp, &vm->va, it.rb) {
1257 list_del(&mapping->list);
1258 interval_tree_remove(&mapping->it, &vm->va);
1261 list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
1262 list_del(&mapping->list);
1266 for (i = 0; i < amdgpu_vm_num_pdes(adev); i++)
1267 amdgpu_bo_unref(&vm->page_tables[i].bo);
1268 kfree(vm->page_tables);
1270 amdgpu_bo_unref(&vm->page_directory);
1272 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1273 amdgpu_fence_unref(&vm->ids[i].flushed_updates);
1274 amdgpu_fence_unref(&vm->ids[i].last_id_use);
1277 mutex_destroy(&vm->mutex);