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 * @validated: head of validation list
82 * @duplicates: head of duplicates list
84 * Add the page directory to the list of BOs to
85 * validate for command submission (cayman+).
87 struct amdgpu_bo_list_entry *amdgpu_vm_get_bos(struct amdgpu_device *adev,
89 struct list_head *validated,
90 struct list_head *duplicates)
92 struct amdgpu_bo_list_entry *list;
95 list = drm_malloc_ab(vm->max_pde_used + 2,
96 sizeof(struct amdgpu_bo_list_entry));
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, validated);
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, duplicates);
127 * amdgpu_vm_grab_id - allocate the next free VMID
129 * @vm: vm to allocate id for
130 * @ring: ring we want to submit job to
131 * @sync: sync object where we add dependencies
133 * Allocate an id for the vm, adding fences to the sync obj as necessary.
135 * Global mutex must be locked!
137 int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
138 struct amdgpu_sync *sync)
140 struct fence *best[AMDGPU_MAX_RINGS] = {};
141 struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx];
142 struct amdgpu_device *adev = ring->adev;
144 unsigned choices[2] = {};
147 /* check if the id is still valid */
149 unsigned id = vm_id->id;
152 owner = atomic_long_read(&adev->vm_manager.ids[id].owner);
153 if (owner == (long)vm) {
154 trace_amdgpu_vm_grab_id(vm_id->id, ring->idx);
159 /* we definately need to flush */
160 vm_id->pd_gpu_addr = ~0ll;
162 /* skip over VMID 0, since it is the system VM */
163 for (i = 1; i < adev->vm_manager.nvm; ++i) {
164 struct fence *fence = adev->vm_manager.ids[i].active;
165 struct amdgpu_ring *fring;
168 /* found a free one */
170 trace_amdgpu_vm_grab_id(i, ring->idx);
174 fring = amdgpu_ring_from_fence(fence);
175 if (best[fring->idx] == NULL ||
176 fence_is_later(best[fring->idx], fence)) {
177 best[fring->idx] = fence;
178 choices[fring == ring ? 0 : 1] = i;
182 for (i = 0; i < 2; ++i) {
186 fence = adev->vm_manager.ids[choices[i]].active;
187 vm_id->id = choices[i];
189 trace_amdgpu_vm_grab_id(choices[i], ring->idx);
190 return amdgpu_sync_fence(ring->adev, sync, fence);
194 /* should never happen */
200 * amdgpu_vm_flush - hardware flush the vm
202 * @ring: ring to use for flush
203 * @vm: vm we want to flush
204 * @updates: last vm update that we waited for
206 * Flush the vm (cayman+).
208 * Global and local mutex must be locked!
210 void amdgpu_vm_flush(struct amdgpu_ring *ring,
211 struct amdgpu_vm *vm,
212 struct fence *updates)
214 uint64_t pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
215 struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx];
216 struct fence *flushed_updates = vm_id->flushed_updates;
219 if (!flushed_updates)
224 is_later = fence_is_later(updates, flushed_updates);
226 if (pd_addr != vm_id->pd_gpu_addr || is_later) {
227 trace_amdgpu_vm_flush(pd_addr, ring->idx, vm_id->id);
229 vm_id->flushed_updates = fence_get(updates);
230 fence_put(flushed_updates);
232 vm_id->pd_gpu_addr = pd_addr;
233 amdgpu_ring_emit_vm_flush(ring, vm_id->id, vm_id->pd_gpu_addr);
238 * amdgpu_vm_fence - remember fence for vm
240 * @adev: amdgpu_device pointer
241 * @vm: vm we want to fence
242 * @fence: fence to remember
244 * Fence the vm (cayman+).
245 * Set the fence used to protect page table and id.
247 * Global and local mutex must be locked!
249 void amdgpu_vm_fence(struct amdgpu_device *adev,
250 struct amdgpu_vm *vm,
253 struct amdgpu_ring *ring = amdgpu_ring_from_fence(fence);
254 unsigned vm_id = vm->ids[ring->idx].id;
256 fence_put(adev->vm_manager.ids[vm_id].active);
257 adev->vm_manager.ids[vm_id].active = fence_get(fence);
258 atomic_long_set(&adev->vm_manager.ids[vm_id].owner, (long)vm);
262 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
265 * @bo: requested buffer object
267 * Find @bo inside the requested vm (cayman+).
268 * Search inside the @bos vm list for the requested vm
269 * Returns the found bo_va or NULL if none is found
271 * Object has to be reserved!
273 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
274 struct amdgpu_bo *bo)
276 struct amdgpu_bo_va *bo_va;
278 list_for_each_entry(bo_va, &bo->va, bo_list) {
279 if (bo_va->vm == vm) {
287 * amdgpu_vm_update_pages - helper to call the right asic function
289 * @adev: amdgpu_device pointer
290 * @ib: indirect buffer to fill with commands
291 * @pe: addr of the page entry
292 * @addr: dst addr to write into pe
293 * @count: number of page entries to update
294 * @incr: increase next addr by incr bytes
295 * @flags: hw access flags
296 * @gtt_flags: GTT hw access flags
298 * Traces the parameters and calls the right asic functions
299 * to setup the page table using the DMA.
301 static void amdgpu_vm_update_pages(struct amdgpu_device *adev,
302 struct amdgpu_ib *ib,
303 uint64_t pe, uint64_t addr,
304 unsigned count, uint32_t incr,
305 uint32_t flags, uint32_t gtt_flags)
307 trace_amdgpu_vm_set_page(pe, addr, count, incr, flags);
309 if ((flags & AMDGPU_PTE_SYSTEM) && (flags == gtt_flags)) {
310 uint64_t src = adev->gart.table_addr + (addr >> 12) * 8;
311 amdgpu_vm_copy_pte(adev, ib, pe, src, count);
313 } else if ((flags & AMDGPU_PTE_SYSTEM) || (count < 3)) {
314 amdgpu_vm_write_pte(adev, ib, pe, addr,
318 amdgpu_vm_set_pte_pde(adev, ib, pe, addr,
323 int amdgpu_vm_free_job(struct amdgpu_job *job)
326 for (i = 0; i < job->num_ibs; i++)
327 amdgpu_ib_free(job->adev, &job->ibs[i]);
333 * amdgpu_vm_clear_bo - initially clear the page dir/table
335 * @adev: amdgpu_device pointer
338 * need to reserve bo first before calling it.
340 static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
341 struct amdgpu_bo *bo)
343 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
344 struct fence *fence = NULL;
345 struct amdgpu_ib *ib;
350 r = reservation_object_reserve_shared(bo->tbo.resv);
354 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
358 addr = amdgpu_bo_gpu_offset(bo);
359 entries = amdgpu_bo_size(bo) / 8;
361 ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
365 r = amdgpu_ib_get(ring, NULL, entries * 2 + 64, ib);
371 amdgpu_vm_update_pages(adev, ib, addr, 0, entries, 0, 0, 0);
372 amdgpu_vm_pad_ib(adev, ib);
373 WARN_ON(ib->length_dw > 64);
374 r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
376 AMDGPU_FENCE_OWNER_VM,
379 amdgpu_bo_fence(bo, fence, true);
381 if (amdgpu_enable_scheduler)
385 amdgpu_ib_free(adev, ib);
393 * amdgpu_vm_map_gart - get the physical address of a gart page
395 * @adev: amdgpu_device pointer
396 * @addr: the unmapped addr
398 * Look up the physical address of the page that the pte resolves
400 * Returns the physical address of the page.
402 uint64_t amdgpu_vm_map_gart(struct amdgpu_device *adev, uint64_t addr)
406 /* page table offset */
407 result = adev->gart.pages_addr[addr >> PAGE_SHIFT];
409 /* in case cpu page size != gpu page size*/
410 result |= addr & (~PAGE_MASK);
416 * amdgpu_vm_update_pdes - make sure that page directory is valid
418 * @adev: amdgpu_device pointer
420 * @start: start of GPU address range
421 * @end: end of GPU address range
423 * Allocates new page tables if necessary
424 * and updates the page directory (cayman+).
425 * Returns 0 for success, error for failure.
427 * Global and local mutex must be locked!
429 int amdgpu_vm_update_page_directory(struct amdgpu_device *adev,
430 struct amdgpu_vm *vm)
432 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
433 struct amdgpu_bo *pd = vm->page_directory;
434 uint64_t pd_addr = amdgpu_bo_gpu_offset(pd);
435 uint32_t incr = AMDGPU_VM_PTE_COUNT * 8;
436 uint64_t last_pde = ~0, last_pt = ~0;
437 unsigned count = 0, pt_idx, ndw;
438 struct amdgpu_ib *ib;
439 struct fence *fence = NULL;
446 /* assume the worst case */
447 ndw += vm->max_pde_used * 6;
449 /* update too big for an IB */
453 ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
457 r = amdgpu_ib_get(ring, NULL, ndw * 4, ib);
464 /* walk over the address space and update the page directory */
465 for (pt_idx = 0; pt_idx <= vm->max_pde_used; ++pt_idx) {
466 struct amdgpu_bo *bo = vm->page_tables[pt_idx].bo;
472 pt = amdgpu_bo_gpu_offset(bo);
473 if (vm->page_tables[pt_idx].addr == pt)
475 vm->page_tables[pt_idx].addr = pt;
477 pde = pd_addr + pt_idx * 8;
478 if (((last_pde + 8 * count) != pde) ||
479 ((last_pt + incr * count) != pt)) {
482 amdgpu_vm_update_pages(adev, ib, last_pde,
483 last_pt, count, incr,
484 AMDGPU_PTE_VALID, 0);
496 amdgpu_vm_update_pages(adev, ib, last_pde, last_pt, count,
497 incr, AMDGPU_PTE_VALID, 0);
499 if (ib->length_dw != 0) {
500 amdgpu_vm_pad_ib(adev, ib);
501 amdgpu_sync_resv(adev, &ib->sync, pd->tbo.resv, AMDGPU_FENCE_OWNER_VM);
502 WARN_ON(ib->length_dw > ndw);
503 r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
505 AMDGPU_FENCE_OWNER_VM,
510 amdgpu_bo_fence(pd, fence, true);
511 fence_put(vm->page_directory_fence);
512 vm->page_directory_fence = fence_get(fence);
516 if (!amdgpu_enable_scheduler || ib->length_dw == 0) {
517 amdgpu_ib_free(adev, ib);
524 amdgpu_ib_free(adev, ib);
530 * amdgpu_vm_frag_ptes - add fragment information to PTEs
532 * @adev: amdgpu_device pointer
533 * @ib: IB for the update
534 * @pe_start: first PTE to handle
535 * @pe_end: last PTE to handle
536 * @addr: addr those PTEs should point to
537 * @flags: hw mapping flags
538 * @gtt_flags: GTT hw mapping flags
540 * Global and local mutex must be locked!
542 static void amdgpu_vm_frag_ptes(struct amdgpu_device *adev,
543 struct amdgpu_ib *ib,
544 uint64_t pe_start, uint64_t pe_end,
545 uint64_t addr, uint32_t flags,
549 * The MC L1 TLB supports variable sized pages, based on a fragment
550 * field in the PTE. When this field is set to a non-zero value, page
551 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
552 * flags are considered valid for all PTEs within the fragment range
553 * and corresponding mappings are assumed to be physically contiguous.
555 * The L1 TLB can store a single PTE for the whole fragment,
556 * significantly increasing the space available for translation
557 * caching. This leads to large improvements in throughput when the
558 * TLB is under pressure.
560 * The L2 TLB distributes small and large fragments into two
561 * asymmetric partitions. The large fragment cache is significantly
562 * larger. Thus, we try to use large fragments wherever possible.
563 * Userspace can support this by aligning virtual base address and
564 * allocation size to the fragment size.
567 /* SI and newer are optimized for 64KB */
568 uint64_t frag_flags = AMDGPU_PTE_FRAG_64KB;
569 uint64_t frag_align = 0x80;
571 uint64_t frag_start = ALIGN(pe_start, frag_align);
572 uint64_t frag_end = pe_end & ~(frag_align - 1);
576 /* system pages are non continuously */
577 if ((flags & AMDGPU_PTE_SYSTEM) || !(flags & AMDGPU_PTE_VALID) ||
578 (frag_start >= frag_end)) {
580 count = (pe_end - pe_start) / 8;
581 amdgpu_vm_update_pages(adev, ib, pe_start, addr, count,
582 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
586 /* handle the 4K area at the beginning */
587 if (pe_start != frag_start) {
588 count = (frag_start - pe_start) / 8;
589 amdgpu_vm_update_pages(adev, ib, pe_start, addr, count,
590 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
591 addr += AMDGPU_GPU_PAGE_SIZE * count;
594 /* handle the area in the middle */
595 count = (frag_end - frag_start) / 8;
596 amdgpu_vm_update_pages(adev, ib, frag_start, addr, count,
597 AMDGPU_GPU_PAGE_SIZE, flags | frag_flags,
600 /* handle the 4K area at the end */
601 if (frag_end != pe_end) {
602 addr += AMDGPU_GPU_PAGE_SIZE * count;
603 count = (pe_end - frag_end) / 8;
604 amdgpu_vm_update_pages(adev, ib, frag_end, addr, count,
605 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
610 * amdgpu_vm_update_ptes - make sure that page tables are valid
612 * @adev: amdgpu_device pointer
614 * @start: start of GPU address range
615 * @end: end of GPU address range
616 * @dst: destination address to map to
617 * @flags: mapping flags
619 * Update the page tables in the range @start - @end (cayman+).
621 * Global and local mutex must be locked!
623 static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
624 struct amdgpu_vm *vm,
625 struct amdgpu_ib *ib,
626 uint64_t start, uint64_t end,
627 uint64_t dst, uint32_t flags,
630 uint64_t mask = AMDGPU_VM_PTE_COUNT - 1;
631 uint64_t last_pte = ~0, last_dst = ~0;
632 void *owner = AMDGPU_FENCE_OWNER_VM;
636 /* sync to everything on unmapping */
637 if (!(flags & AMDGPU_PTE_VALID))
638 owner = AMDGPU_FENCE_OWNER_UNDEFINED;
640 /* walk over the address space and update the page tables */
641 for (addr = start; addr < end; ) {
642 uint64_t pt_idx = addr >> amdgpu_vm_block_size;
643 struct amdgpu_bo *pt = vm->page_tables[pt_idx].bo;
648 amdgpu_sync_resv(adev, &ib->sync, pt->tbo.resv, owner);
649 r = reservation_object_reserve_shared(pt->tbo.resv);
653 if ((addr & ~mask) == (end & ~mask))
656 nptes = AMDGPU_VM_PTE_COUNT - (addr & mask);
658 pte = amdgpu_bo_gpu_offset(pt);
659 pte += (addr & mask) * 8;
661 if ((last_pte + 8 * count) != pte) {
664 amdgpu_vm_frag_ptes(adev, ib, last_pte,
665 last_pte + 8 * count,
678 dst += nptes * AMDGPU_GPU_PAGE_SIZE;
682 amdgpu_vm_frag_ptes(adev, ib, last_pte,
683 last_pte + 8 * count,
684 last_dst, flags, gtt_flags);
691 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
693 * @adev: amdgpu_device pointer
695 * @mapping: mapped range and flags to use for the update
696 * @addr: addr to set the area to
697 * @gtt_flags: flags as they are used for GTT
698 * @fence: optional resulting fence
700 * Fill in the page table entries for @mapping.
701 * Returns 0 for success, -EINVAL for failure.
703 * Object have to be reserved and mutex must be locked!
705 static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
706 struct amdgpu_vm *vm,
707 struct amdgpu_bo_va_mapping *mapping,
708 uint64_t addr, uint32_t gtt_flags,
709 struct fence **fence)
711 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
712 unsigned nptes, ncmds, ndw;
713 uint32_t flags = gtt_flags;
714 struct amdgpu_ib *ib;
715 struct fence *f = NULL;
718 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
719 * but in case of something, we filter the flags in first place
721 if (!(mapping->flags & AMDGPU_PTE_READABLE))
722 flags &= ~AMDGPU_PTE_READABLE;
723 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
724 flags &= ~AMDGPU_PTE_WRITEABLE;
726 trace_amdgpu_vm_bo_update(mapping);
728 nptes = mapping->it.last - mapping->it.start + 1;
731 * reserve space for one command every (1 << BLOCK_SIZE)
732 * entries or 2k dwords (whatever is smaller)
734 ncmds = (nptes >> min(amdgpu_vm_block_size, 11)) + 1;
739 if ((flags & AMDGPU_PTE_SYSTEM) && (flags == gtt_flags)) {
740 /* only copy commands needed */
743 } else if (flags & AMDGPU_PTE_SYSTEM) {
744 /* header for write data commands */
747 /* body of write data command */
751 /* set page commands needed */
754 /* two extra commands for begin/end of fragment */
758 /* update too big for an IB */
762 ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
766 r = amdgpu_ib_get(ring, NULL, ndw * 4, ib);
774 r = amdgpu_vm_update_ptes(adev, vm, ib, mapping->it.start,
775 mapping->it.last + 1, addr + mapping->offset,
779 amdgpu_ib_free(adev, ib);
784 amdgpu_vm_pad_ib(adev, ib);
785 WARN_ON(ib->length_dw > ndw);
786 r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
788 AMDGPU_FENCE_OWNER_VM,
793 amdgpu_bo_fence(vm->page_directory, f, true);
796 *fence = fence_get(f);
799 if (!amdgpu_enable_scheduler) {
800 amdgpu_ib_free(adev, ib);
806 amdgpu_ib_free(adev, ib);
812 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
814 * @adev: amdgpu_device pointer
815 * @bo_va: requested BO and VM object
818 * Fill in the page table entries for @bo_va.
819 * Returns 0 for success, -EINVAL for failure.
821 * Object have to be reserved and mutex must be locked!
823 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
824 struct amdgpu_bo_va *bo_va,
825 struct ttm_mem_reg *mem)
827 struct amdgpu_vm *vm = bo_va->vm;
828 struct amdgpu_bo_va_mapping *mapping;
834 addr = (u64)mem->start << PAGE_SHIFT;
835 if (mem->mem_type != TTM_PL_TT)
836 addr += adev->vm_manager.vram_base_offset;
841 flags = amdgpu_ttm_tt_pte_flags(adev, bo_va->bo->tbo.ttm, mem);
843 spin_lock(&vm->status_lock);
844 if (!list_empty(&bo_va->vm_status))
845 list_splice_init(&bo_va->valids, &bo_va->invalids);
846 spin_unlock(&vm->status_lock);
848 list_for_each_entry(mapping, &bo_va->invalids, list) {
849 r = amdgpu_vm_bo_update_mapping(adev, vm, mapping, addr,
850 flags, &bo_va->last_pt_update);
855 if (trace_amdgpu_vm_bo_mapping_enabled()) {
856 list_for_each_entry(mapping, &bo_va->valids, list)
857 trace_amdgpu_vm_bo_mapping(mapping);
859 list_for_each_entry(mapping, &bo_va->invalids, list)
860 trace_amdgpu_vm_bo_mapping(mapping);
863 spin_lock(&vm->status_lock);
864 list_splice_init(&bo_va->invalids, &bo_va->valids);
865 list_del_init(&bo_va->vm_status);
867 list_add(&bo_va->vm_status, &vm->cleared);
868 spin_unlock(&vm->status_lock);
874 * amdgpu_vm_clear_freed - clear freed BOs in the PT
876 * @adev: amdgpu_device pointer
879 * Make sure all freed BOs are cleared in the PT.
880 * Returns 0 for success.
882 * PTs have to be reserved and mutex must be locked!
884 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
885 struct amdgpu_vm *vm)
887 struct amdgpu_bo_va_mapping *mapping;
890 spin_lock(&vm->freed_lock);
891 while (!list_empty(&vm->freed)) {
892 mapping = list_first_entry(&vm->freed,
893 struct amdgpu_bo_va_mapping, list);
894 list_del(&mapping->list);
895 spin_unlock(&vm->freed_lock);
896 r = amdgpu_vm_bo_update_mapping(adev, vm, mapping, 0, 0, NULL);
901 spin_lock(&vm->freed_lock);
903 spin_unlock(&vm->freed_lock);
910 * amdgpu_vm_clear_invalids - clear invalidated BOs in the PT
912 * @adev: amdgpu_device pointer
915 * Make sure all invalidated BOs are cleared in the PT.
916 * Returns 0 for success.
918 * PTs have to be reserved and mutex must be locked!
920 int amdgpu_vm_clear_invalids(struct amdgpu_device *adev,
921 struct amdgpu_vm *vm, struct amdgpu_sync *sync)
923 struct amdgpu_bo_va *bo_va = NULL;
926 spin_lock(&vm->status_lock);
927 while (!list_empty(&vm->invalidated)) {
928 bo_va = list_first_entry(&vm->invalidated,
929 struct amdgpu_bo_va, vm_status);
930 spin_unlock(&vm->status_lock);
931 mutex_lock(&bo_va->mutex);
932 r = amdgpu_vm_bo_update(adev, bo_va, NULL);
933 mutex_unlock(&bo_va->mutex);
937 spin_lock(&vm->status_lock);
939 spin_unlock(&vm->status_lock);
942 r = amdgpu_sync_fence(adev, sync, bo_va->last_pt_update);
948 * amdgpu_vm_bo_add - add a bo to a specific vm
950 * @adev: amdgpu_device pointer
952 * @bo: amdgpu buffer object
954 * Add @bo into the requested vm (cayman+).
955 * Add @bo to the list of bos associated with the vm
956 * Returns newly added bo_va or NULL for failure
958 * Object has to be reserved!
960 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
961 struct amdgpu_vm *vm,
962 struct amdgpu_bo *bo)
964 struct amdgpu_bo_va *bo_va;
966 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
972 bo_va->ref_count = 1;
973 INIT_LIST_HEAD(&bo_va->bo_list);
974 INIT_LIST_HEAD(&bo_va->valids);
975 INIT_LIST_HEAD(&bo_va->invalids);
976 INIT_LIST_HEAD(&bo_va->vm_status);
977 mutex_init(&bo_va->mutex);
978 list_add_tail(&bo_va->bo_list, &bo->va);
984 * amdgpu_vm_bo_map - map bo inside a vm
986 * @adev: amdgpu_device pointer
987 * @bo_va: bo_va to store the address
988 * @saddr: where to map the BO
989 * @offset: requested offset in the BO
990 * @flags: attributes of pages (read/write/valid/etc.)
992 * Add a mapping of the BO at the specefied addr into the VM.
993 * Returns 0 for success, error for failure.
995 * Object has to be reserved and unreserved outside!
997 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
998 struct amdgpu_bo_va *bo_va,
999 uint64_t saddr, uint64_t offset,
1000 uint64_t size, uint32_t flags)
1002 struct amdgpu_bo_va_mapping *mapping;
1003 struct amdgpu_vm *vm = bo_va->vm;
1004 struct interval_tree_node *it;
1005 unsigned last_pfn, pt_idx;
1009 /* validate the parameters */
1010 if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
1011 size == 0 || size & AMDGPU_GPU_PAGE_MASK)
1014 /* make sure object fit at this offset */
1015 eaddr = saddr + size;
1016 if ((saddr >= eaddr) || (offset + size > amdgpu_bo_size(bo_va->bo)))
1019 last_pfn = eaddr / AMDGPU_GPU_PAGE_SIZE;
1020 if (last_pfn > adev->vm_manager.max_pfn) {
1021 dev_err(adev->dev, "va above limit (0x%08X > 0x%08X)\n",
1022 last_pfn, adev->vm_manager.max_pfn);
1026 saddr /= AMDGPU_GPU_PAGE_SIZE;
1027 eaddr /= AMDGPU_GPU_PAGE_SIZE;
1029 spin_lock(&vm->it_lock);
1030 it = interval_tree_iter_first(&vm->va, saddr, eaddr - 1);
1031 spin_unlock(&vm->it_lock);
1033 struct amdgpu_bo_va_mapping *tmp;
1034 tmp = container_of(it, struct amdgpu_bo_va_mapping, it);
1035 /* bo and tmp overlap, invalid addr */
1036 dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
1037 "0x%010lx-0x%010lx\n", bo_va->bo, saddr, eaddr,
1038 tmp->it.start, tmp->it.last + 1);
1043 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1049 INIT_LIST_HEAD(&mapping->list);
1050 mapping->it.start = saddr;
1051 mapping->it.last = eaddr - 1;
1052 mapping->offset = offset;
1053 mapping->flags = flags;
1055 mutex_lock(&bo_va->mutex);
1056 list_add(&mapping->list, &bo_va->invalids);
1057 mutex_unlock(&bo_va->mutex);
1058 spin_lock(&vm->it_lock);
1059 interval_tree_insert(&mapping->it, &vm->va);
1060 spin_unlock(&vm->it_lock);
1061 trace_amdgpu_vm_bo_map(bo_va, mapping);
1063 /* Make sure the page tables are allocated */
1064 saddr >>= amdgpu_vm_block_size;
1065 eaddr >>= amdgpu_vm_block_size;
1067 BUG_ON(eaddr >= amdgpu_vm_num_pdes(adev));
1069 if (eaddr > vm->max_pde_used)
1070 vm->max_pde_used = eaddr;
1072 /* walk over the address space and allocate the page tables */
1073 for (pt_idx = saddr; pt_idx <= eaddr; ++pt_idx) {
1074 struct reservation_object *resv = vm->page_directory->tbo.resv;
1075 struct amdgpu_bo *pt;
1077 if (vm->page_tables[pt_idx].bo)
1080 r = amdgpu_bo_create(adev, AMDGPU_VM_PTE_COUNT * 8,
1081 AMDGPU_GPU_PAGE_SIZE, true,
1082 AMDGPU_GEM_DOMAIN_VRAM,
1083 AMDGPU_GEM_CREATE_NO_CPU_ACCESS,
1088 r = amdgpu_vm_clear_bo(adev, pt);
1090 amdgpu_bo_unref(&pt);
1094 vm->page_tables[pt_idx].addr = 0;
1095 vm->page_tables[pt_idx].bo = pt;
1101 list_del(&mapping->list);
1102 spin_lock(&vm->it_lock);
1103 interval_tree_remove(&mapping->it, &vm->va);
1104 spin_unlock(&vm->it_lock);
1105 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1113 * amdgpu_vm_bo_unmap - remove bo mapping from vm
1115 * @adev: amdgpu_device pointer
1116 * @bo_va: bo_va to remove the address from
1117 * @saddr: where to the BO is mapped
1119 * Remove a mapping of the BO at the specefied addr from the VM.
1120 * Returns 0 for success, error for failure.
1122 * Object has to be reserved and unreserved outside!
1124 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
1125 struct amdgpu_bo_va *bo_va,
1128 struct amdgpu_bo_va_mapping *mapping;
1129 struct amdgpu_vm *vm = bo_va->vm;
1132 saddr /= AMDGPU_GPU_PAGE_SIZE;
1133 mutex_lock(&bo_va->mutex);
1134 list_for_each_entry(mapping, &bo_va->valids, list) {
1135 if (mapping->it.start == saddr)
1139 if (&mapping->list == &bo_va->valids) {
1142 list_for_each_entry(mapping, &bo_va->invalids, list) {
1143 if (mapping->it.start == saddr)
1147 if (&mapping->list == &bo_va->invalids) {
1148 mutex_unlock(&bo_va->mutex);
1152 mutex_unlock(&bo_va->mutex);
1153 list_del(&mapping->list);
1154 spin_lock(&vm->it_lock);
1155 interval_tree_remove(&mapping->it, &vm->va);
1156 spin_unlock(&vm->it_lock);
1157 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1160 spin_lock(&vm->freed_lock);
1161 list_add(&mapping->list, &vm->freed);
1162 spin_unlock(&vm->freed_lock);
1171 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
1173 * @adev: amdgpu_device pointer
1174 * @bo_va: requested bo_va
1176 * Remove @bo_va->bo from the requested vm (cayman+).
1178 * Object have to be reserved!
1180 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
1181 struct amdgpu_bo_va *bo_va)
1183 struct amdgpu_bo_va_mapping *mapping, *next;
1184 struct amdgpu_vm *vm = bo_va->vm;
1186 list_del(&bo_va->bo_list);
1188 spin_lock(&vm->status_lock);
1189 list_del(&bo_va->vm_status);
1190 spin_unlock(&vm->status_lock);
1192 list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
1193 list_del(&mapping->list);
1194 spin_lock(&vm->it_lock);
1195 interval_tree_remove(&mapping->it, &vm->va);
1196 spin_unlock(&vm->it_lock);
1197 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1198 spin_lock(&vm->freed_lock);
1199 list_add(&mapping->list, &vm->freed);
1200 spin_unlock(&vm->freed_lock);
1202 list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
1203 list_del(&mapping->list);
1204 spin_lock(&vm->it_lock);
1205 interval_tree_remove(&mapping->it, &vm->va);
1206 spin_unlock(&vm->it_lock);
1209 fence_put(bo_va->last_pt_update);
1210 mutex_destroy(&bo_va->mutex);
1215 * amdgpu_vm_bo_invalidate - mark the bo as invalid
1217 * @adev: amdgpu_device pointer
1219 * @bo: amdgpu buffer object
1221 * Mark @bo as invalid (cayman+).
1223 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
1224 struct amdgpu_bo *bo)
1226 struct amdgpu_bo_va *bo_va;
1228 list_for_each_entry(bo_va, &bo->va, bo_list) {
1229 spin_lock(&bo_va->vm->status_lock);
1230 if (list_empty(&bo_va->vm_status))
1231 list_add(&bo_va->vm_status, &bo_va->vm->invalidated);
1232 spin_unlock(&bo_va->vm->status_lock);
1237 * amdgpu_vm_init - initialize a vm instance
1239 * @adev: amdgpu_device pointer
1242 * Init @vm fields (cayman+).
1244 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1246 const unsigned align = min(AMDGPU_VM_PTB_ALIGN_SIZE,
1247 AMDGPU_VM_PTE_COUNT * 8);
1248 unsigned pd_size, pd_entries, pts_size;
1251 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1253 vm->ids[i].flushed_updates = NULL;
1256 spin_lock_init(&vm->status_lock);
1257 INIT_LIST_HEAD(&vm->invalidated);
1258 INIT_LIST_HEAD(&vm->cleared);
1259 INIT_LIST_HEAD(&vm->freed);
1260 spin_lock_init(&vm->it_lock);
1261 spin_lock_init(&vm->freed_lock);
1262 pd_size = amdgpu_vm_directory_size(adev);
1263 pd_entries = amdgpu_vm_num_pdes(adev);
1265 /* allocate page table array */
1266 pts_size = pd_entries * sizeof(struct amdgpu_vm_pt);
1267 vm->page_tables = kzalloc(pts_size, GFP_KERNEL);
1268 if (vm->page_tables == NULL) {
1269 DRM_ERROR("Cannot allocate memory for page table array\n");
1273 vm->page_directory_fence = NULL;
1275 r = amdgpu_bo_create(adev, pd_size, align, true,
1276 AMDGPU_GEM_DOMAIN_VRAM,
1277 AMDGPU_GEM_CREATE_NO_CPU_ACCESS,
1278 NULL, NULL, &vm->page_directory);
1281 r = amdgpu_bo_reserve(vm->page_directory, false);
1283 amdgpu_bo_unref(&vm->page_directory);
1284 vm->page_directory = NULL;
1287 r = amdgpu_vm_clear_bo(adev, vm->page_directory);
1288 amdgpu_bo_unreserve(vm->page_directory);
1290 amdgpu_bo_unref(&vm->page_directory);
1291 vm->page_directory = NULL;
1299 * amdgpu_vm_fini - tear down a vm instance
1301 * @adev: amdgpu_device pointer
1304 * Tear down @vm (cayman+).
1305 * Unbind the VM and remove all bos from the vm bo list
1307 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1309 struct amdgpu_bo_va_mapping *mapping, *tmp;
1312 if (!RB_EMPTY_ROOT(&vm->va)) {
1313 dev_err(adev->dev, "still active bo inside vm\n");
1315 rbtree_postorder_for_each_entry_safe(mapping, tmp, &vm->va, it.rb) {
1316 list_del(&mapping->list);
1317 interval_tree_remove(&mapping->it, &vm->va);
1320 list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
1321 list_del(&mapping->list);
1325 for (i = 0; i < amdgpu_vm_num_pdes(adev); i++)
1326 amdgpu_bo_unref(&vm->page_tables[i].bo);
1327 kfree(vm->page_tables);
1329 amdgpu_bo_unref(&vm->page_directory);
1330 fence_put(vm->page_directory_fence);
1331 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1332 unsigned id = vm->ids[i].id;
1334 atomic_long_cmpxchg(&adev->vm_manager.ids[id].owner,
1336 fence_put(vm->ids[i].flushed_updates);
1342 * amdgpu_vm_manager_fini - cleanup VM manager
1344 * @adev: amdgpu_device pointer
1346 * Cleanup the VM manager and free resources.
1348 void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
1352 for (i = 0; i < AMDGPU_NUM_VM; ++i)
1353 fence_put(adev->vm_manager.ids[i].active);