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 list = drm_malloc_ab(vm->max_pde_used + 2,
94 sizeof(struct amdgpu_bo_list_entry));
99 /* add the vm page table to the list */
100 list[0].robj = vm->page_directory;
101 list[0].prefered_domains = AMDGPU_GEM_DOMAIN_VRAM;
102 list[0].allowed_domains = AMDGPU_GEM_DOMAIN_VRAM;
103 list[0].priority = 0;
104 list[0].tv.bo = &vm->page_directory->tbo;
105 list[0].tv.shared = true;
106 list_add(&list[0].tv.head, head);
108 for (i = 0, idx = 1; i <= vm->max_pde_used; i++) {
109 if (!vm->page_tables[i].bo)
112 list[idx].robj = vm->page_tables[i].bo;
113 list[idx].prefered_domains = AMDGPU_GEM_DOMAIN_VRAM;
114 list[idx].allowed_domains = AMDGPU_GEM_DOMAIN_VRAM;
115 list[idx].priority = 0;
116 list[idx].tv.bo = &list[idx].robj->tbo;
117 list[idx].tv.shared = true;
118 list_add(&list[idx++].tv.head, head);
125 * amdgpu_vm_grab_id - allocate the next free VMID
127 * @vm: vm to allocate id for
128 * @ring: ring we want to submit job to
129 * @sync: sync object where we add dependencies
131 * Allocate an id for the vm, adding fences to the sync obj as necessary.
133 * Global mutex must be locked!
135 int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
136 struct amdgpu_sync *sync)
138 struct fence *best[AMDGPU_MAX_RINGS] = {};
139 struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx];
140 struct amdgpu_device *adev = ring->adev;
142 unsigned choices[2] = {};
145 /* check if the id is still valid */
146 if (vm_id->id && vm_id->last_id_use &&
147 vm_id->last_id_use == adev->vm_manager.active[vm_id->id]) {
148 trace_amdgpu_vm_grab_id(vm_id->id, ring->idx);
152 /* we definately need to flush */
153 vm_id->pd_gpu_addr = ~0ll;
155 /* skip over VMID 0, since it is the system VM */
156 for (i = 1; i < adev->vm_manager.nvm; ++i) {
157 struct fence *fence = adev->vm_manager.active[i];
158 struct amdgpu_ring *fring;
161 /* found a free one */
163 trace_amdgpu_vm_grab_id(i, ring->idx);
167 fring = amdgpu_ring_from_fence(fence);
168 if (best[fring->idx] == NULL ||
169 fence_is_later(best[fring->idx], fence)) {
170 best[fring->idx] = fence;
171 choices[fring == ring ? 0 : 1] = i;
175 for (i = 0; i < 2; ++i) {
179 fence = adev->vm_manager.active[choices[i]];
180 vm_id->id = choices[i];
182 trace_amdgpu_vm_grab_id(choices[i], ring->idx);
183 return amdgpu_sync_fence(ring->adev, sync, fence);
187 /* should never happen */
193 * amdgpu_vm_flush - hardware flush the vm
195 * @ring: ring to use for flush
196 * @vm: vm we want to flush
197 * @updates: last vm update that we waited for
199 * Flush the vm (cayman+).
201 * Global and local mutex must be locked!
203 void amdgpu_vm_flush(struct amdgpu_ring *ring,
204 struct amdgpu_vm *vm,
205 struct fence *updates)
207 uint64_t pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
208 struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx];
209 struct fence *flushed_updates = vm_id->flushed_updates;
212 if (!flushed_updates)
217 is_later = fence_is_later(updates, flushed_updates);
219 if (pd_addr != vm_id->pd_gpu_addr || is_later) {
220 trace_amdgpu_vm_flush(pd_addr, ring->idx, vm_id->id);
222 vm_id->flushed_updates = fence_get(updates);
223 fence_put(flushed_updates);
225 vm_id->pd_gpu_addr = pd_addr;
226 amdgpu_ring_emit_vm_flush(ring, vm_id->id, vm_id->pd_gpu_addr);
231 * amdgpu_vm_fence - remember fence for vm
233 * @adev: amdgpu_device pointer
234 * @vm: vm we want to fence
235 * @fence: fence to remember
237 * Fence the vm (cayman+).
238 * Set the fence used to protect page table and id.
240 * Global and local mutex must be locked!
242 void amdgpu_vm_fence(struct amdgpu_device *adev,
243 struct amdgpu_vm *vm,
244 struct amdgpu_fence *fence)
246 unsigned ridx = fence->ring->idx;
247 unsigned vm_id = vm->ids[ridx].id;
249 fence_put(adev->vm_manager.active[vm_id]);
250 adev->vm_manager.active[vm_id] = fence_get(&fence->base);
252 fence_put(vm->ids[ridx].last_id_use);
253 vm->ids[ridx].last_id_use = fence_get(&fence->base);
257 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
260 * @bo: requested buffer object
262 * Find @bo inside the requested vm (cayman+).
263 * Search inside the @bos vm list for the requested vm
264 * Returns the found bo_va or NULL if none is found
266 * Object has to be reserved!
268 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
269 struct amdgpu_bo *bo)
271 struct amdgpu_bo_va *bo_va;
273 list_for_each_entry(bo_va, &bo->va, bo_list) {
274 if (bo_va->vm == vm) {
282 * amdgpu_vm_update_pages - helper to call the right asic function
284 * @adev: amdgpu_device pointer
285 * @ib: indirect buffer to fill with commands
286 * @pe: addr of the page entry
287 * @addr: dst addr to write into pe
288 * @count: number of page entries to update
289 * @incr: increase next addr by incr bytes
290 * @flags: hw access flags
291 * @gtt_flags: GTT hw access flags
293 * Traces the parameters and calls the right asic functions
294 * to setup the page table using the DMA.
296 static void amdgpu_vm_update_pages(struct amdgpu_device *adev,
297 struct amdgpu_ib *ib,
298 uint64_t pe, uint64_t addr,
299 unsigned count, uint32_t incr,
300 uint32_t flags, uint32_t gtt_flags)
302 trace_amdgpu_vm_set_page(pe, addr, count, incr, flags);
304 if ((flags & AMDGPU_PTE_SYSTEM) && (flags == gtt_flags)) {
305 uint64_t src = adev->gart.table_addr + (addr >> 12) * 8;
306 amdgpu_vm_copy_pte(adev, ib, pe, src, count);
308 } else if ((flags & AMDGPU_PTE_SYSTEM) || (count < 3)) {
309 amdgpu_vm_write_pte(adev, ib, pe, addr,
313 amdgpu_vm_set_pte_pde(adev, ib, pe, addr,
318 int amdgpu_vm_free_job(struct amdgpu_job *job)
321 for (i = 0; i < job->num_ibs; i++)
322 amdgpu_ib_free(job->adev, &job->ibs[i]);
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 fence *fence = 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);
371 r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
373 AMDGPU_FENCE_OWNER_VM,
376 amdgpu_bo_fence(bo, fence, true);
378 if (amdgpu_enable_scheduler) {
379 amdgpu_bo_unreserve(bo);
383 amdgpu_ib_free(adev, ib);
387 amdgpu_bo_unreserve(bo);
392 * amdgpu_vm_map_gart - get the physical address of a gart page
394 * @adev: amdgpu_device pointer
395 * @addr: the unmapped addr
397 * Look up the physical address of the page that the pte resolves
399 * Returns the physical address of the page.
401 uint64_t amdgpu_vm_map_gart(struct amdgpu_device *adev, uint64_t addr)
405 /* page table offset */
406 result = adev->gart.pages_addr[addr >> PAGE_SHIFT];
408 /* in case cpu page size != gpu page size*/
409 result |= addr & (~PAGE_MASK);
415 * amdgpu_vm_update_pdes - make sure that page directory is valid
417 * @adev: amdgpu_device pointer
419 * @start: start of GPU address range
420 * @end: end of GPU address range
422 * Allocates new page tables if necessary
423 * and updates the page directory (cayman+).
424 * Returns 0 for success, error for failure.
426 * Global and local mutex must be locked!
428 int amdgpu_vm_update_page_directory(struct amdgpu_device *adev,
429 struct amdgpu_vm *vm)
431 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
432 struct amdgpu_bo *pd = vm->page_directory;
433 uint64_t pd_addr = amdgpu_bo_gpu_offset(pd);
434 uint32_t incr = AMDGPU_VM_PTE_COUNT * 8;
435 uint64_t last_pde = ~0, last_pt = ~0;
436 unsigned count = 0, pt_idx, ndw;
437 struct amdgpu_ib *ib;
438 struct fence *fence = NULL;
445 /* assume the worst case */
446 ndw += vm->max_pde_used * 6;
448 /* update too big for an IB */
452 ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
456 r = amdgpu_ib_get(ring, NULL, ndw * 4, ib);
463 /* walk over the address space and update the page directory */
464 for (pt_idx = 0; pt_idx <= vm->max_pde_used; ++pt_idx) {
465 struct amdgpu_bo *bo = vm->page_tables[pt_idx].bo;
471 pt = amdgpu_bo_gpu_offset(bo);
472 if (vm->page_tables[pt_idx].addr == pt)
474 vm->page_tables[pt_idx].addr = pt;
476 pde = pd_addr + pt_idx * 8;
477 if (((last_pde + 8 * count) != pde) ||
478 ((last_pt + incr * count) != pt)) {
481 amdgpu_vm_update_pages(adev, ib, last_pde,
482 last_pt, count, incr,
483 AMDGPU_PTE_VALID, 0);
495 amdgpu_vm_update_pages(adev, ib, last_pde, last_pt, count,
496 incr, AMDGPU_PTE_VALID, 0);
498 if (ib->length_dw != 0) {
499 amdgpu_vm_pad_ib(adev, ib);
500 amdgpu_sync_resv(adev, &ib->sync, pd->tbo.resv, AMDGPU_FENCE_OWNER_VM);
501 WARN_ON(ib->length_dw > ndw);
502 r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
504 AMDGPU_FENCE_OWNER_VM,
509 amdgpu_bo_fence(pd, fence, true);
510 fence_put(vm->page_directory_fence);
511 vm->page_directory_fence = fence_get(fence);
515 if (!amdgpu_enable_scheduler || ib->length_dw == 0) {
516 amdgpu_ib_free(adev, ib);
523 amdgpu_ib_free(adev, ib);
529 * amdgpu_vm_frag_ptes - add fragment information to PTEs
531 * @adev: amdgpu_device pointer
532 * @ib: IB for the update
533 * @pe_start: first PTE to handle
534 * @pe_end: last PTE to handle
535 * @addr: addr those PTEs should point to
536 * @flags: hw mapping flags
537 * @gtt_flags: GTT hw mapping flags
539 * Global and local mutex must be locked!
541 static void amdgpu_vm_frag_ptes(struct amdgpu_device *adev,
542 struct amdgpu_ib *ib,
543 uint64_t pe_start, uint64_t pe_end,
544 uint64_t addr, uint32_t flags,
548 * The MC L1 TLB supports variable sized pages, based on a fragment
549 * field in the PTE. When this field is set to a non-zero value, page
550 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
551 * flags are considered valid for all PTEs within the fragment range
552 * and corresponding mappings are assumed to be physically contiguous.
554 * The L1 TLB can store a single PTE for the whole fragment,
555 * significantly increasing the space available for translation
556 * caching. This leads to large improvements in throughput when the
557 * TLB is under pressure.
559 * The L2 TLB distributes small and large fragments into two
560 * asymmetric partitions. The large fragment cache is significantly
561 * larger. Thus, we try to use large fragments wherever possible.
562 * Userspace can support this by aligning virtual base address and
563 * allocation size to the fragment size.
566 /* SI and newer are optimized for 64KB */
567 uint64_t frag_flags = AMDGPU_PTE_FRAG_64KB;
568 uint64_t frag_align = 0x80;
570 uint64_t frag_start = ALIGN(pe_start, frag_align);
571 uint64_t frag_end = pe_end & ~(frag_align - 1);
575 /* system pages are non continuously */
576 if ((flags & AMDGPU_PTE_SYSTEM) || !(flags & AMDGPU_PTE_VALID) ||
577 (frag_start >= frag_end)) {
579 count = (pe_end - pe_start) / 8;
580 amdgpu_vm_update_pages(adev, ib, pe_start, addr, count,
581 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
585 /* handle the 4K area at the beginning */
586 if (pe_start != frag_start) {
587 count = (frag_start - pe_start) / 8;
588 amdgpu_vm_update_pages(adev, ib, pe_start, addr, count,
589 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
590 addr += AMDGPU_GPU_PAGE_SIZE * count;
593 /* handle the area in the middle */
594 count = (frag_end - frag_start) / 8;
595 amdgpu_vm_update_pages(adev, ib, frag_start, addr, count,
596 AMDGPU_GPU_PAGE_SIZE, flags | frag_flags,
599 /* handle the 4K area at the end */
600 if (frag_end != pe_end) {
601 addr += AMDGPU_GPU_PAGE_SIZE * count;
602 count = (pe_end - frag_end) / 8;
603 amdgpu_vm_update_pages(adev, ib, frag_end, addr, count,
604 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
609 * amdgpu_vm_update_ptes - make sure that page tables are valid
611 * @adev: amdgpu_device pointer
613 * @start: start of GPU address range
614 * @end: end of GPU address range
615 * @dst: destination address to map to
616 * @flags: mapping flags
618 * Update the page tables in the range @start - @end (cayman+).
620 * Global and local mutex must be locked!
622 static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
623 struct amdgpu_vm *vm,
624 struct amdgpu_ib *ib,
625 uint64_t start, uint64_t end,
626 uint64_t dst, uint32_t flags,
629 uint64_t mask = AMDGPU_VM_PTE_COUNT - 1;
630 uint64_t last_pte = ~0, last_dst = ~0;
631 void *owner = AMDGPU_FENCE_OWNER_VM;
635 /* sync to everything on unmapping */
636 if (!(flags & AMDGPU_PTE_VALID))
637 owner = AMDGPU_FENCE_OWNER_UNDEFINED;
639 /* walk over the address space and update the page tables */
640 for (addr = start; addr < end; ) {
641 uint64_t pt_idx = addr >> amdgpu_vm_block_size;
642 struct amdgpu_bo *pt = vm->page_tables[pt_idx].bo;
647 amdgpu_sync_resv(adev, &ib->sync, pt->tbo.resv, owner);
648 r = reservation_object_reserve_shared(pt->tbo.resv);
652 if ((addr & ~mask) == (end & ~mask))
655 nptes = AMDGPU_VM_PTE_COUNT - (addr & mask);
657 pte = amdgpu_bo_gpu_offset(pt);
658 pte += (addr & mask) * 8;
660 if ((last_pte + 8 * count) != pte) {
663 amdgpu_vm_frag_ptes(adev, ib, last_pte,
664 last_pte + 8 * count,
677 dst += nptes * AMDGPU_GPU_PAGE_SIZE;
681 amdgpu_vm_frag_ptes(adev, ib, last_pte,
682 last_pte + 8 * count,
683 last_dst, flags, gtt_flags);
690 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
692 * @adev: amdgpu_device pointer
694 * @mapping: mapped range and flags to use for the update
695 * @addr: addr to set the area to
696 * @gtt_flags: flags as they are used for GTT
697 * @fence: optional resulting fence
699 * Fill in the page table entries for @mapping.
700 * Returns 0 for success, -EINVAL for failure.
702 * Object have to be reserved and mutex must be locked!
704 static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
705 struct amdgpu_vm *vm,
706 struct amdgpu_bo_va_mapping *mapping,
707 uint64_t addr, uint32_t gtt_flags,
708 struct fence **fence)
710 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
711 unsigned nptes, ncmds, ndw;
712 uint32_t flags = gtt_flags;
713 struct amdgpu_ib *ib;
714 struct fence *f = NULL;
717 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
718 * but in case of something, we filter the flags in first place
720 if (!(mapping->flags & AMDGPU_PTE_READABLE))
721 flags &= ~AMDGPU_PTE_READABLE;
722 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
723 flags &= ~AMDGPU_PTE_WRITEABLE;
725 trace_amdgpu_vm_bo_update(mapping);
727 nptes = mapping->it.last - mapping->it.start + 1;
730 * reserve space for one command every (1 << BLOCK_SIZE)
731 * entries or 2k dwords (whatever is smaller)
733 ncmds = (nptes >> min(amdgpu_vm_block_size, 11)) + 1;
738 if ((flags & AMDGPU_PTE_SYSTEM) && (flags == gtt_flags)) {
739 /* only copy commands needed */
742 } else if (flags & AMDGPU_PTE_SYSTEM) {
743 /* header for write data commands */
746 /* body of write data command */
750 /* set page commands needed */
753 /* two extra commands for begin/end of fragment */
757 /* update too big for an IB */
761 ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
765 r = amdgpu_ib_get(ring, NULL, ndw * 4, ib);
773 r = amdgpu_vm_update_ptes(adev, vm, ib, mapping->it.start,
774 mapping->it.last + 1, addr + mapping->offset,
778 amdgpu_ib_free(adev, ib);
783 amdgpu_vm_pad_ib(adev, ib);
784 WARN_ON(ib->length_dw > ndw);
785 r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
787 AMDGPU_FENCE_OWNER_VM,
792 amdgpu_bo_fence(vm->page_directory, f, true);
795 *fence = fence_get(f);
798 if (!amdgpu_enable_scheduler) {
799 amdgpu_ib_free(adev, ib);
805 amdgpu_ib_free(adev, ib);
811 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
813 * @adev: amdgpu_device pointer
814 * @bo_va: requested BO and VM object
817 * Fill in the page table entries for @bo_va.
818 * Returns 0 for success, -EINVAL for failure.
820 * Object have to be reserved and mutex must be locked!
822 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
823 struct amdgpu_bo_va *bo_va,
824 struct ttm_mem_reg *mem)
826 struct amdgpu_vm *vm = bo_va->vm;
827 struct amdgpu_bo_va_mapping *mapping;
833 addr = (u64)mem->start << PAGE_SHIFT;
834 if (mem->mem_type != TTM_PL_TT)
835 addr += adev->vm_manager.vram_base_offset;
840 flags = amdgpu_ttm_tt_pte_flags(adev, bo_va->bo->tbo.ttm, mem);
842 spin_lock(&vm->status_lock);
843 if (!list_empty(&bo_va->vm_status))
844 list_splice_init(&bo_va->valids, &bo_va->invalids);
845 spin_unlock(&vm->status_lock);
847 list_for_each_entry(mapping, &bo_va->invalids, list) {
848 r = amdgpu_vm_bo_update_mapping(adev, vm, mapping, addr,
849 flags, &bo_va->last_pt_update);
854 if (trace_amdgpu_vm_bo_mapping_enabled()) {
855 list_for_each_entry(mapping, &bo_va->valids, list)
856 trace_amdgpu_vm_bo_mapping(mapping);
858 list_for_each_entry(mapping, &bo_va->invalids, list)
859 trace_amdgpu_vm_bo_mapping(mapping);
862 spin_lock(&vm->status_lock);
863 list_splice_init(&bo_va->invalids, &bo_va->valids);
864 list_del_init(&bo_va->vm_status);
866 list_add(&bo_va->vm_status, &vm->cleared);
867 spin_unlock(&vm->status_lock);
873 * amdgpu_vm_clear_freed - clear freed BOs in the PT
875 * @adev: amdgpu_device pointer
878 * Make sure all freed BOs are cleared in the PT.
879 * Returns 0 for success.
881 * PTs have to be reserved and mutex must be locked!
883 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
884 struct amdgpu_vm *vm)
886 struct amdgpu_bo_va_mapping *mapping;
889 while (!list_empty(&vm->freed)) {
890 mapping = list_first_entry(&vm->freed,
891 struct amdgpu_bo_va_mapping, list);
892 list_del(&mapping->list);
894 r = amdgpu_vm_bo_update_mapping(adev, vm, mapping, 0, 0, NULL);
905 * amdgpu_vm_clear_invalids - clear invalidated BOs in the PT
907 * @adev: amdgpu_device pointer
910 * Make sure all invalidated BOs are cleared in the PT.
911 * Returns 0 for success.
913 * PTs have to be reserved and mutex must be locked!
915 int amdgpu_vm_clear_invalids(struct amdgpu_device *adev,
916 struct amdgpu_vm *vm, struct amdgpu_sync *sync)
918 struct amdgpu_bo_va *bo_va = NULL;
921 spin_lock(&vm->status_lock);
922 while (!list_empty(&vm->invalidated)) {
923 bo_va = list_first_entry(&vm->invalidated,
924 struct amdgpu_bo_va, vm_status);
925 spin_unlock(&vm->status_lock);
927 r = amdgpu_vm_bo_update(adev, bo_va, NULL);
931 spin_lock(&vm->status_lock);
933 spin_unlock(&vm->status_lock);
936 r = amdgpu_sync_fence(adev, sync, bo_va->last_pt_update);
942 * amdgpu_vm_bo_add - add a bo to a specific vm
944 * @adev: amdgpu_device pointer
946 * @bo: amdgpu buffer object
948 * Add @bo into the requested vm (cayman+).
949 * Add @bo to the list of bos associated with the vm
950 * Returns newly added bo_va or NULL for failure
952 * Object has to be reserved!
954 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
955 struct amdgpu_vm *vm,
956 struct amdgpu_bo *bo)
958 struct amdgpu_bo_va *bo_va;
960 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
966 bo_va->ref_count = 1;
967 INIT_LIST_HEAD(&bo_va->bo_list);
968 INIT_LIST_HEAD(&bo_va->valids);
969 INIT_LIST_HEAD(&bo_va->invalids);
970 INIT_LIST_HEAD(&bo_va->vm_status);
972 list_add_tail(&bo_va->bo_list, &bo->va);
978 * amdgpu_vm_bo_map - map bo inside a vm
980 * @adev: amdgpu_device pointer
981 * @bo_va: bo_va to store the address
982 * @saddr: where to map the BO
983 * @offset: requested offset in the BO
984 * @flags: attributes of pages (read/write/valid/etc.)
986 * Add a mapping of the BO at the specefied addr into the VM.
987 * Returns 0 for success, error for failure.
989 * Object has to be reserved and gets unreserved by this function!
991 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
992 struct amdgpu_bo_va *bo_va,
993 uint64_t saddr, uint64_t offset,
994 uint64_t size, uint32_t flags)
996 struct amdgpu_bo_va_mapping *mapping;
997 struct amdgpu_vm *vm = bo_va->vm;
998 struct interval_tree_node *it;
999 unsigned last_pfn, pt_idx;
1003 /* validate the parameters */
1004 if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
1005 size == 0 || size & AMDGPU_GPU_PAGE_MASK) {
1006 amdgpu_bo_unreserve(bo_va->bo);
1010 /* make sure object fit at this offset */
1011 eaddr = saddr + size;
1012 if ((saddr >= eaddr) || (offset + size > amdgpu_bo_size(bo_va->bo))) {
1013 amdgpu_bo_unreserve(bo_va->bo);
1017 last_pfn = eaddr / AMDGPU_GPU_PAGE_SIZE;
1018 if (last_pfn > adev->vm_manager.max_pfn) {
1019 dev_err(adev->dev, "va above limit (0x%08X > 0x%08X)\n",
1020 last_pfn, adev->vm_manager.max_pfn);
1021 amdgpu_bo_unreserve(bo_va->bo);
1025 saddr /= AMDGPU_GPU_PAGE_SIZE;
1026 eaddr /= AMDGPU_GPU_PAGE_SIZE;
1028 it = interval_tree_iter_first(&vm->va, saddr, eaddr - 1);
1030 struct amdgpu_bo_va_mapping *tmp;
1031 tmp = container_of(it, struct amdgpu_bo_va_mapping, it);
1032 /* bo and tmp overlap, invalid addr */
1033 dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
1034 "0x%010lx-0x%010lx\n", bo_va->bo, saddr, eaddr,
1035 tmp->it.start, tmp->it.last + 1);
1036 amdgpu_bo_unreserve(bo_va->bo);
1041 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1043 amdgpu_bo_unreserve(bo_va->bo);
1048 INIT_LIST_HEAD(&mapping->list);
1049 mapping->it.start = saddr;
1050 mapping->it.last = eaddr - 1;
1051 mapping->offset = offset;
1052 mapping->flags = flags;
1054 list_add(&mapping->list, &bo_va->invalids);
1055 interval_tree_insert(&mapping->it, &vm->va);
1056 trace_amdgpu_vm_bo_map(bo_va, mapping);
1058 /* Make sure the page tables are allocated */
1059 saddr >>= amdgpu_vm_block_size;
1060 eaddr >>= amdgpu_vm_block_size;
1062 BUG_ON(eaddr >= amdgpu_vm_num_pdes(adev));
1064 if (eaddr > vm->max_pde_used)
1065 vm->max_pde_used = eaddr;
1067 amdgpu_bo_unreserve(bo_va->bo);
1069 /* walk over the address space and allocate the page tables */
1070 for (pt_idx = saddr; pt_idx <= eaddr; ++pt_idx) {
1071 struct reservation_object *resv = vm->page_directory->tbo.resv;
1072 struct amdgpu_bo *pt;
1074 if (vm->page_tables[pt_idx].bo)
1077 ww_mutex_lock(&resv->lock, NULL);
1078 r = amdgpu_bo_create(adev, AMDGPU_VM_PTE_COUNT * 8,
1079 AMDGPU_GPU_PAGE_SIZE, true,
1080 AMDGPU_GEM_DOMAIN_VRAM,
1081 AMDGPU_GEM_CREATE_NO_CPU_ACCESS,
1083 ww_mutex_unlock(&resv->lock);
1087 r = amdgpu_vm_clear_bo(adev, pt);
1089 amdgpu_bo_unref(&pt);
1093 vm->page_tables[pt_idx].addr = 0;
1094 vm->page_tables[pt_idx].bo = pt;
1100 list_del(&mapping->list);
1101 interval_tree_remove(&mapping->it, &vm->va);
1102 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1110 * amdgpu_vm_bo_unmap - remove bo mapping from vm
1112 * @adev: amdgpu_device pointer
1113 * @bo_va: bo_va to remove the address from
1114 * @saddr: where to the BO is mapped
1116 * Remove a mapping of the BO at the specefied addr from the VM.
1117 * Returns 0 for success, error for failure.
1119 * Object has to be reserved and gets unreserved by this function!
1121 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
1122 struct amdgpu_bo_va *bo_va,
1125 struct amdgpu_bo_va_mapping *mapping;
1126 struct amdgpu_vm *vm = bo_va->vm;
1129 saddr /= AMDGPU_GPU_PAGE_SIZE;
1131 list_for_each_entry(mapping, &bo_va->valids, list) {
1132 if (mapping->it.start == saddr)
1136 if (&mapping->list == &bo_va->valids) {
1139 list_for_each_entry(mapping, &bo_va->invalids, list) {
1140 if (mapping->it.start == saddr)
1144 if (&mapping->list == &bo_va->invalids) {
1145 amdgpu_bo_unreserve(bo_va->bo);
1150 list_del(&mapping->list);
1151 interval_tree_remove(&mapping->it, &vm->va);
1152 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1155 list_add(&mapping->list, &vm->freed);
1158 amdgpu_bo_unreserve(bo_va->bo);
1164 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
1166 * @adev: amdgpu_device pointer
1167 * @bo_va: requested bo_va
1169 * Remove @bo_va->bo from the requested vm (cayman+).
1171 * Object have to be reserved!
1173 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
1174 struct amdgpu_bo_va *bo_va)
1176 struct amdgpu_bo_va_mapping *mapping, *next;
1177 struct amdgpu_vm *vm = bo_va->vm;
1179 list_del(&bo_va->bo_list);
1181 spin_lock(&vm->status_lock);
1182 list_del(&bo_va->vm_status);
1183 spin_unlock(&vm->status_lock);
1185 list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
1186 list_del(&mapping->list);
1187 interval_tree_remove(&mapping->it, &vm->va);
1188 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1189 list_add(&mapping->list, &vm->freed);
1191 list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
1192 list_del(&mapping->list);
1193 interval_tree_remove(&mapping->it, &vm->va);
1197 fence_put(bo_va->last_pt_update);
1202 * amdgpu_vm_bo_invalidate - mark the bo as invalid
1204 * @adev: amdgpu_device pointer
1206 * @bo: amdgpu buffer object
1208 * Mark @bo as invalid (cayman+).
1210 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
1211 struct amdgpu_bo *bo)
1213 struct amdgpu_bo_va *bo_va;
1215 list_for_each_entry(bo_va, &bo->va, bo_list) {
1216 spin_lock(&bo_va->vm->status_lock);
1217 if (list_empty(&bo_va->vm_status))
1218 list_add(&bo_va->vm_status, &bo_va->vm->invalidated);
1219 spin_unlock(&bo_va->vm->status_lock);
1224 * amdgpu_vm_init - initialize a vm instance
1226 * @adev: amdgpu_device pointer
1229 * Init @vm fields (cayman+).
1231 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1233 const unsigned align = min(AMDGPU_VM_PTB_ALIGN_SIZE,
1234 AMDGPU_VM_PTE_COUNT * 8);
1235 unsigned pd_size, pd_entries, pts_size;
1238 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1240 vm->ids[i].flushed_updates = NULL;
1241 vm->ids[i].last_id_use = NULL;
1243 mutex_init(&vm->mutex);
1245 spin_lock_init(&vm->status_lock);
1246 INIT_LIST_HEAD(&vm->invalidated);
1247 INIT_LIST_HEAD(&vm->cleared);
1248 INIT_LIST_HEAD(&vm->freed);
1250 pd_size = amdgpu_vm_directory_size(adev);
1251 pd_entries = amdgpu_vm_num_pdes(adev);
1253 /* allocate page table array */
1254 pts_size = pd_entries * sizeof(struct amdgpu_vm_pt);
1255 vm->page_tables = kzalloc(pts_size, GFP_KERNEL);
1256 if (vm->page_tables == NULL) {
1257 DRM_ERROR("Cannot allocate memory for page table array\n");
1261 vm->page_directory_fence = NULL;
1263 r = amdgpu_bo_create(adev, pd_size, align, true,
1264 AMDGPU_GEM_DOMAIN_VRAM,
1265 AMDGPU_GEM_CREATE_NO_CPU_ACCESS,
1266 NULL, NULL, &vm->page_directory);
1270 r = amdgpu_vm_clear_bo(adev, vm->page_directory);
1272 amdgpu_bo_unref(&vm->page_directory);
1273 vm->page_directory = NULL;
1281 * amdgpu_vm_fini - tear down a vm instance
1283 * @adev: amdgpu_device pointer
1286 * Tear down @vm (cayman+).
1287 * Unbind the VM and remove all bos from the vm bo list
1289 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1291 struct amdgpu_bo_va_mapping *mapping, *tmp;
1294 if (!RB_EMPTY_ROOT(&vm->va)) {
1295 dev_err(adev->dev, "still active bo inside vm\n");
1297 rbtree_postorder_for_each_entry_safe(mapping, tmp, &vm->va, it.rb) {
1298 list_del(&mapping->list);
1299 interval_tree_remove(&mapping->it, &vm->va);
1302 list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
1303 list_del(&mapping->list);
1307 for (i = 0; i < amdgpu_vm_num_pdes(adev); i++)
1308 amdgpu_bo_unref(&vm->page_tables[i].bo);
1309 kfree(vm->page_tables);
1311 amdgpu_bo_unref(&vm->page_directory);
1312 fence_put(vm->page_directory_fence);
1314 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1315 fence_put(vm->ids[i].flushed_updates);
1316 fence_put(vm->ids[i].last_id_use);
1319 mutex_destroy(&vm->mutex);