2 * Copyright 2008 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
24 * Authors: Dave Airlie
29 #include <drm/amdgpu_drm.h>
31 #include "amdgpu_trace.h"
35 * GPUVM is similar to the legacy gart on older asics, however
36 * rather than there being a single global gart table
37 * for the entire GPU, there are multiple VM page tables active
38 * at any given time. The VM page tables can contain a mix
39 * vram pages and system memory pages and system memory pages
40 * can be mapped as snooped (cached system pages) or unsnooped
41 * (uncached system pages).
42 * Each VM has an ID associated with it and there is a page table
43 * associated with each VMID. When execting a command buffer,
44 * the kernel tells the the ring what VMID to use for that command
45 * buffer. VMIDs are allocated dynamically as commands are submitted.
46 * The userspace drivers maintain their own address space and the kernel
47 * sets up their pages tables accordingly when they submit their
48 * command buffers and a VMID is assigned.
49 * Cayman/Trinity support up to 8 active VMs at any given time;
54 * amdgpu_vm_num_pde - return the number of page directory entries
56 * @adev: amdgpu_device pointer
58 * Calculate the number of page directory entries (cayman+).
60 static unsigned amdgpu_vm_num_pdes(struct amdgpu_device *adev)
62 return adev->vm_manager.max_pfn >> amdgpu_vm_block_size;
66 * amdgpu_vm_directory_size - returns the size of the page directory in bytes
68 * @adev: amdgpu_device pointer
70 * Calculate the size of the page directory in bytes (cayman+).
72 static unsigned amdgpu_vm_directory_size(struct amdgpu_device *adev)
74 return AMDGPU_GPU_PAGE_ALIGN(amdgpu_vm_num_pdes(adev) * 8);
78 * amdgpu_vm_get_bos - add the vm BOs to a validation list
80 * @vm: vm providing the BOs
81 * @head: head of validation list
83 * Add the page directory to the list of BOs to
84 * validate for command submission (cayman+).
86 struct amdgpu_bo_list_entry *amdgpu_vm_get_bos(struct amdgpu_device *adev,
88 struct list_head *head)
90 struct amdgpu_bo_list_entry *list;
93 mutex_lock(&vm->mutex);
94 list = drm_malloc_ab(vm->max_pde_used + 2,
95 sizeof(struct amdgpu_bo_list_entry));
97 mutex_unlock(&vm->mutex);
101 /* add the vm page table to the list */
102 list[0].robj = vm->page_directory;
103 list[0].prefered_domains = AMDGPU_GEM_DOMAIN_VRAM;
104 list[0].allowed_domains = AMDGPU_GEM_DOMAIN_VRAM;
105 list[0].priority = 0;
106 list[0].tv.bo = &vm->page_directory->tbo;
107 list[0].tv.shared = true;
108 list_add(&list[0].tv.head, head);
110 for (i = 0, idx = 1; i <= vm->max_pde_used; i++) {
111 if (!vm->page_tables[i].bo)
114 list[idx].robj = vm->page_tables[i].bo;
115 list[idx].prefered_domains = AMDGPU_GEM_DOMAIN_VRAM;
116 list[idx].allowed_domains = AMDGPU_GEM_DOMAIN_VRAM;
117 list[idx].priority = 0;
118 list[idx].tv.bo = &list[idx].robj->tbo;
119 list[idx].tv.shared = true;
120 list_add(&list[idx++].tv.head, head);
122 mutex_unlock(&vm->mutex);
128 * amdgpu_vm_grab_id - allocate the next free VMID
130 * @vm: vm to allocate id for
131 * @ring: ring we want to submit job to
132 * @sync: sync object where we add dependencies
134 * Allocate an id for the vm, adding fences to the sync obj as necessary.
136 * Global mutex must be locked!
138 int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
139 struct amdgpu_sync *sync)
141 struct amdgpu_fence *best[AMDGPU_MAX_RINGS] = {};
142 struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx];
143 struct amdgpu_device *adev = ring->adev;
145 unsigned choices[2] = {};
148 /* check if the id is still valid */
149 if (vm_id->id && vm_id->last_id_use &&
150 vm_id->last_id_use == adev->vm_manager.active[vm_id->id])
153 /* we definately need to flush */
154 vm_id->pd_gpu_addr = ~0ll;
156 /* skip over VMID 0, since it is the system VM */
157 for (i = 1; i < adev->vm_manager.nvm; ++i) {
158 struct amdgpu_fence *fence = adev->vm_manager.active[i];
161 /* found a free one */
163 trace_amdgpu_vm_grab_id(i, ring->idx);
167 if (amdgpu_fence_is_earlier(fence, best[fence->ring->idx])) {
168 best[fence->ring->idx] = fence;
169 choices[fence->ring == ring ? 0 : 1] = i;
173 for (i = 0; i < 2; ++i) {
175 struct amdgpu_fence *fence;
177 fence = adev->vm_manager.active[choices[i]];
178 vm_id->id = choices[i];
180 trace_amdgpu_vm_grab_id(choices[i], ring->idx);
181 return amdgpu_sync_fence(ring->adev, sync, &fence->base);
185 /* should never happen */
191 * amdgpu_vm_flush - hardware flush the vm
193 * @ring: ring to use for flush
194 * @vm: vm we want to flush
195 * @updates: last vm update that we waited for
197 * Flush the vm (cayman+).
199 * Global and local mutex must be locked!
201 void amdgpu_vm_flush(struct amdgpu_ring *ring,
202 struct amdgpu_vm *vm,
203 struct fence *updates)
205 uint64_t pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
206 struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx];
207 struct fence *flushed_updates = vm_id->flushed_updates;
208 bool is_earlier = false;
210 if (flushed_updates && updates) {
211 BUG_ON(flushed_updates->context != updates->context);
212 is_earlier = (updates->seqno - flushed_updates->seqno <=
213 INT_MAX) ? true : false;
216 if (pd_addr != vm_id->pd_gpu_addr || !flushed_updates ||
219 trace_amdgpu_vm_flush(pd_addr, ring->idx, vm_id->id);
221 vm_id->flushed_updates = fence_get(updates);
222 fence_put(flushed_updates);
224 if (!flushed_updates)
225 vm_id->flushed_updates = fence_get(updates);
226 vm_id->pd_gpu_addr = pd_addr;
227 amdgpu_ring_emit_vm_flush(ring, vm_id->id, vm_id->pd_gpu_addr);
232 * amdgpu_vm_fence - remember fence for vm
234 * @adev: amdgpu_device pointer
235 * @vm: vm we want to fence
236 * @fence: fence to remember
238 * Fence the vm (cayman+).
239 * Set the fence used to protect page table and id.
241 * Global and local mutex must be locked!
243 void amdgpu_vm_fence(struct amdgpu_device *adev,
244 struct amdgpu_vm *vm,
245 struct amdgpu_fence *fence)
247 unsigned ridx = fence->ring->idx;
248 unsigned vm_id = vm->ids[ridx].id;
250 amdgpu_fence_unref(&adev->vm_manager.active[vm_id]);
251 adev->vm_manager.active[vm_id] = amdgpu_fence_ref(fence);
253 amdgpu_fence_unref(&vm->ids[ridx].last_id_use);
254 vm->ids[ridx].last_id_use = amdgpu_fence_ref(fence);
258 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
261 * @bo: requested buffer object
263 * Find @bo inside the requested vm (cayman+).
264 * Search inside the @bos vm list for the requested vm
265 * Returns the found bo_va or NULL if none is found
267 * Object has to be reserved!
269 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
270 struct amdgpu_bo *bo)
272 struct amdgpu_bo_va *bo_va;
274 list_for_each_entry(bo_va, &bo->va, bo_list) {
275 if (bo_va->vm == vm) {
283 * amdgpu_vm_update_pages - helper to call the right asic function
285 * @adev: amdgpu_device pointer
286 * @ib: indirect buffer to fill with commands
287 * @pe: addr of the page entry
288 * @addr: dst addr to write into pe
289 * @count: number of page entries to update
290 * @incr: increase next addr by incr bytes
291 * @flags: hw access flags
292 * @gtt_flags: GTT hw access flags
294 * Traces the parameters and calls the right asic functions
295 * to setup the page table using the DMA.
297 static void amdgpu_vm_update_pages(struct amdgpu_device *adev,
298 struct amdgpu_ib *ib,
299 uint64_t pe, uint64_t addr,
300 unsigned count, uint32_t incr,
301 uint32_t flags, uint32_t gtt_flags)
303 trace_amdgpu_vm_set_page(pe, addr, count, incr, flags);
305 if ((flags & AMDGPU_PTE_SYSTEM) && (flags == gtt_flags)) {
306 uint64_t src = adev->gart.table_addr + (addr >> 12) * 8;
307 amdgpu_vm_copy_pte(adev, ib, pe, src, count);
309 } else if ((flags & AMDGPU_PTE_SYSTEM) || (count < 3)) {
310 amdgpu_vm_write_pte(adev, ib, pe, addr,
314 amdgpu_vm_set_pte_pde(adev, ib, pe, addr,
319 int amdgpu_vm_free_job(struct amdgpu_job *sched_job)
322 for (i = 0; i < sched_job->num_ibs; i++)
323 amdgpu_ib_free(sched_job->adev, &sched_job->ibs[i]);
324 kfree(sched_job->ibs);
329 * amdgpu_vm_clear_bo - initially clear the page dir/table
331 * @adev: amdgpu_device pointer
334 static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
335 struct amdgpu_bo *bo)
337 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
338 struct fence *fence = NULL;
339 struct amdgpu_ib *ib;
344 r = amdgpu_bo_reserve(bo, false);
348 r = reservation_object_reserve_shared(bo->tbo.resv);
352 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
354 goto error_unreserve;
356 addr = amdgpu_bo_gpu_offset(bo);
357 entries = amdgpu_bo_size(bo) / 8;
359 ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
361 goto error_unreserve;
363 r = amdgpu_ib_get(ring, NULL, entries * 2 + 64, ib);
369 amdgpu_vm_update_pages(adev, ib, addr, 0, entries, 0, 0, 0);
370 amdgpu_vm_pad_ib(adev, ib);
371 WARN_ON(ib->length_dw > 64);
372 r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
374 AMDGPU_FENCE_OWNER_VM,
377 amdgpu_bo_fence(bo, fence, true);
379 if (amdgpu_enable_scheduler) {
380 amdgpu_bo_unreserve(bo);
384 amdgpu_ib_free(adev, ib);
388 amdgpu_bo_unreserve(bo);
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);
462 /* walk over the address space and update the page directory */
463 for (pt_idx = 0; pt_idx <= vm->max_pde_used; ++pt_idx) {
464 struct amdgpu_bo *bo = vm->page_tables[pt_idx].bo;
470 pt = amdgpu_bo_gpu_offset(bo);
471 if (vm->page_tables[pt_idx].addr == pt)
473 vm->page_tables[pt_idx].addr = pt;
475 pde = pd_addr + pt_idx * 8;
476 if (((last_pde + 8 * count) != pde) ||
477 ((last_pt + incr * count) != pt)) {
480 amdgpu_vm_update_pages(adev, ib, last_pde,
481 last_pt, count, incr,
482 AMDGPU_PTE_VALID, 0);
494 amdgpu_vm_update_pages(adev, ib, last_pde, last_pt, count,
495 incr, AMDGPU_PTE_VALID, 0);
497 if (ib->length_dw != 0) {
498 amdgpu_vm_pad_ib(adev, ib);
499 amdgpu_sync_resv(adev, &ib->sync, pd->tbo.resv, AMDGPU_FENCE_OWNER_VM);
500 WARN_ON(ib->length_dw > ndw);
501 r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
503 AMDGPU_FENCE_OWNER_VM,
508 amdgpu_bo_fence(pd, fence, true);
509 fence_put(vm->page_directory_fence);
510 vm->page_directory_fence = fence_get(fence);
514 if (!amdgpu_enable_scheduler || ib->length_dw == 0) {
515 amdgpu_ib_free(adev, ib);
522 amdgpu_ib_free(adev, ib);
528 * amdgpu_vm_frag_ptes - add fragment information to PTEs
530 * @adev: amdgpu_device pointer
531 * @ib: IB for the update
532 * @pe_start: first PTE to handle
533 * @pe_end: last PTE to handle
534 * @addr: addr those PTEs should point to
535 * @flags: hw mapping flags
536 * @gtt_flags: GTT hw mapping flags
538 * Global and local mutex must be locked!
540 static void amdgpu_vm_frag_ptes(struct amdgpu_device *adev,
541 struct amdgpu_ib *ib,
542 uint64_t pe_start, uint64_t pe_end,
543 uint64_t addr, uint32_t flags,
547 * The MC L1 TLB supports variable sized pages, based on a fragment
548 * field in the PTE. When this field is set to a non-zero value, page
549 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
550 * flags are considered valid for all PTEs within the fragment range
551 * and corresponding mappings are assumed to be physically contiguous.
553 * The L1 TLB can store a single PTE for the whole fragment,
554 * significantly increasing the space available for translation
555 * caching. This leads to large improvements in throughput when the
556 * TLB is under pressure.
558 * The L2 TLB distributes small and large fragments into two
559 * asymmetric partitions. The large fragment cache is significantly
560 * larger. Thus, we try to use large fragments wherever possible.
561 * Userspace can support this by aligning virtual base address and
562 * allocation size to the fragment size.
565 /* SI and newer are optimized for 64KB */
566 uint64_t frag_flags = AMDGPU_PTE_FRAG_64KB;
567 uint64_t frag_align = 0x80;
569 uint64_t frag_start = ALIGN(pe_start, frag_align);
570 uint64_t frag_end = pe_end & ~(frag_align - 1);
574 /* system pages are non continuously */
575 if ((flags & AMDGPU_PTE_SYSTEM) || !(flags & AMDGPU_PTE_VALID) ||
576 (frag_start >= frag_end)) {
578 count = (pe_end - pe_start) / 8;
579 amdgpu_vm_update_pages(adev, ib, pe_start, addr, count,
580 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
584 /* handle the 4K area at the beginning */
585 if (pe_start != frag_start) {
586 count = (frag_start - pe_start) / 8;
587 amdgpu_vm_update_pages(adev, ib, pe_start, addr, count,
588 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
589 addr += AMDGPU_GPU_PAGE_SIZE * count;
592 /* handle the area in the middle */
593 count = (frag_end - frag_start) / 8;
594 amdgpu_vm_update_pages(adev, ib, frag_start, addr, count,
595 AMDGPU_GPU_PAGE_SIZE, flags | frag_flags,
598 /* handle the 4K area at the end */
599 if (frag_end != pe_end) {
600 addr += AMDGPU_GPU_PAGE_SIZE * count;
601 count = (pe_end - frag_end) / 8;
602 amdgpu_vm_update_pages(adev, ib, frag_end, addr, count,
603 AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
608 * amdgpu_vm_update_ptes - make sure that page tables are valid
610 * @adev: amdgpu_device pointer
612 * @start: start of GPU address range
613 * @end: end of GPU address range
614 * @dst: destination address to map to
615 * @flags: mapping flags
617 * Update the page tables in the range @start - @end (cayman+).
619 * Global and local mutex must be locked!
621 static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
622 struct amdgpu_vm *vm,
623 struct amdgpu_ib *ib,
624 uint64_t start, uint64_t end,
625 uint64_t dst, uint32_t flags,
628 uint64_t mask = AMDGPU_VM_PTE_COUNT - 1;
629 uint64_t last_pte = ~0, last_dst = ~0;
633 /* walk over the address space and update the page tables */
634 for (addr = start; addr < end; ) {
635 uint64_t pt_idx = addr >> amdgpu_vm_block_size;
636 struct amdgpu_bo *pt = vm->page_tables[pt_idx].bo;
641 amdgpu_sync_resv(adev, &ib->sync, pt->tbo.resv,
642 AMDGPU_FENCE_OWNER_VM);
643 r = reservation_object_reserve_shared(pt->tbo.resv);
647 if ((addr & ~mask) == (end & ~mask))
650 nptes = AMDGPU_VM_PTE_COUNT - (addr & mask);
652 pte = amdgpu_bo_gpu_offset(pt);
653 pte += (addr & mask) * 8;
655 if ((last_pte + 8 * count) != pte) {
658 amdgpu_vm_frag_ptes(adev, ib, last_pte,
659 last_pte + 8 * count,
672 dst += nptes * AMDGPU_GPU_PAGE_SIZE;
676 amdgpu_vm_frag_ptes(adev, ib, last_pte,
677 last_pte + 8 * count,
678 last_dst, flags, gtt_flags);
685 * amdgpu_vm_fence_pts - fence page tables after an update
688 * @start: start of GPU address range
689 * @end: end of GPU address range
690 * @fence: fence to use
692 * Fence the page tables in the range @start - @end (cayman+).
694 * Global and local mutex must be locked!
696 static void amdgpu_vm_fence_pts(struct amdgpu_vm *vm,
697 uint64_t start, uint64_t end,
702 start >>= amdgpu_vm_block_size;
703 end >>= amdgpu_vm_block_size;
705 for (i = start; i <= end; ++i)
706 amdgpu_bo_fence(vm->page_tables[i].bo, fence, true);
710 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
712 * @adev: amdgpu_device pointer
714 * @mapping: mapped range and flags to use for the update
715 * @addr: addr to set the area to
716 * @gtt_flags: flags as they are used for GTT
717 * @fence: optional resulting fence
719 * Fill in the page table entries for @mapping.
720 * Returns 0 for success, -EINVAL for failure.
722 * Object have to be reserved and mutex must be locked!
724 static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
725 struct amdgpu_vm *vm,
726 struct amdgpu_bo_va_mapping *mapping,
727 uint64_t addr, uint32_t gtt_flags,
728 struct fence **fence)
730 struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
731 unsigned nptes, ncmds, ndw;
732 uint32_t flags = gtt_flags;
733 struct amdgpu_ib *ib;
734 struct fence *f = NULL;
737 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
738 * but in case of something, we filter the flags in first place
740 if (!(mapping->flags & AMDGPU_PTE_READABLE))
741 flags &= ~AMDGPU_PTE_READABLE;
742 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
743 flags &= ~AMDGPU_PTE_WRITEABLE;
745 trace_amdgpu_vm_bo_update(mapping);
747 nptes = mapping->it.last - mapping->it.start + 1;
750 * reserve space for one command every (1 << BLOCK_SIZE)
751 * entries or 2k dwords (whatever is smaller)
753 ncmds = (nptes >> min(amdgpu_vm_block_size, 11)) + 1;
758 if ((flags & AMDGPU_PTE_SYSTEM) && (flags == gtt_flags)) {
759 /* only copy commands needed */
762 } else if (flags & AMDGPU_PTE_SYSTEM) {
763 /* header for write data commands */
766 /* body of write data command */
770 /* set page commands needed */
773 /* two extra commands for begin/end of fragment */
777 /* update too big for an IB */
781 ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
785 r = amdgpu_ib_get(ring, NULL, ndw * 4, ib);
793 if (!(flags & AMDGPU_PTE_VALID)) {
796 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
797 struct amdgpu_fence *f = vm->ids[i].last_id_use;
798 r = amdgpu_sync_fence(adev, &ib->sync, &f->base);
804 r = amdgpu_vm_update_ptes(adev, vm, ib, mapping->it.start,
805 mapping->it.last + 1, addr + mapping->offset,
809 amdgpu_ib_free(adev, ib);
814 amdgpu_vm_pad_ib(adev, ib);
815 WARN_ON(ib->length_dw > ndw);
816 r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
818 AMDGPU_FENCE_OWNER_VM,
823 amdgpu_vm_fence_pts(vm, mapping->it.start,
824 mapping->it.last + 1, f);
827 *fence = fence_get(f);
830 if (!amdgpu_enable_scheduler) {
831 amdgpu_ib_free(adev, ib);
837 amdgpu_ib_free(adev, ib);
843 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
845 * @adev: amdgpu_device pointer
846 * @bo_va: requested BO and VM object
849 * Fill in the page table entries for @bo_va.
850 * Returns 0 for success, -EINVAL for failure.
852 * Object have to be reserved and mutex must be locked!
854 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
855 struct amdgpu_bo_va *bo_va,
856 struct ttm_mem_reg *mem)
858 struct amdgpu_vm *vm = bo_va->vm;
859 struct amdgpu_bo_va_mapping *mapping;
865 addr = mem->start << PAGE_SHIFT;
866 if (mem->mem_type != TTM_PL_TT)
867 addr += adev->vm_manager.vram_base_offset;
872 flags = amdgpu_ttm_tt_pte_flags(adev, bo_va->bo->tbo.ttm, mem);
874 spin_lock(&vm->status_lock);
875 if (!list_empty(&bo_va->vm_status))
876 list_splice_init(&bo_va->valids, &bo_va->invalids);
877 spin_unlock(&vm->status_lock);
879 list_for_each_entry(mapping, &bo_va->invalids, list) {
880 r = amdgpu_vm_bo_update_mapping(adev, vm, mapping, addr,
881 flags, &bo_va->last_pt_update);
886 spin_lock(&vm->status_lock);
887 list_splice_init(&bo_va->invalids, &bo_va->valids);
888 list_del_init(&bo_va->vm_status);
890 list_add(&bo_va->vm_status, &vm->cleared);
891 spin_unlock(&vm->status_lock);
897 * amdgpu_vm_clear_freed - clear freed BOs in the PT
899 * @adev: amdgpu_device pointer
902 * Make sure all freed BOs are cleared in the PT.
903 * Returns 0 for success.
905 * PTs have to be reserved and mutex must be locked!
907 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
908 struct amdgpu_vm *vm)
910 struct amdgpu_bo_va_mapping *mapping;
913 while (!list_empty(&vm->freed)) {
914 mapping = list_first_entry(&vm->freed,
915 struct amdgpu_bo_va_mapping, list);
916 list_del(&mapping->list);
918 r = amdgpu_vm_bo_update_mapping(adev, vm, mapping, 0, 0, NULL);
929 * amdgpu_vm_clear_invalids - clear invalidated BOs in the PT
931 * @adev: amdgpu_device pointer
934 * Make sure all invalidated BOs are cleared in the PT.
935 * Returns 0 for success.
937 * PTs have to be reserved and mutex must be locked!
939 int amdgpu_vm_clear_invalids(struct amdgpu_device *adev,
940 struct amdgpu_vm *vm, struct amdgpu_sync *sync)
942 struct amdgpu_bo_va *bo_va = NULL;
945 spin_lock(&vm->status_lock);
946 while (!list_empty(&vm->invalidated)) {
947 bo_va = list_first_entry(&vm->invalidated,
948 struct amdgpu_bo_va, vm_status);
949 spin_unlock(&vm->status_lock);
951 r = amdgpu_vm_bo_update(adev, bo_va, NULL);
955 spin_lock(&vm->status_lock);
957 spin_unlock(&vm->status_lock);
960 r = amdgpu_sync_fence(adev, sync, bo_va->last_pt_update);
966 * amdgpu_vm_bo_add - add a bo to a specific vm
968 * @adev: amdgpu_device pointer
970 * @bo: amdgpu buffer object
972 * Add @bo into the requested vm (cayman+).
973 * Add @bo to the list of bos associated with the vm
974 * Returns newly added bo_va or NULL for failure
976 * Object has to be reserved!
978 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
979 struct amdgpu_vm *vm,
980 struct amdgpu_bo *bo)
982 struct amdgpu_bo_va *bo_va;
984 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
990 bo_va->ref_count = 1;
991 INIT_LIST_HEAD(&bo_va->bo_list);
992 INIT_LIST_HEAD(&bo_va->valids);
993 INIT_LIST_HEAD(&bo_va->invalids);
994 INIT_LIST_HEAD(&bo_va->vm_status);
996 mutex_lock(&vm->mutex);
997 list_add_tail(&bo_va->bo_list, &bo->va);
998 mutex_unlock(&vm->mutex);
1004 * amdgpu_vm_bo_map - map bo inside a vm
1006 * @adev: amdgpu_device pointer
1007 * @bo_va: bo_va to store the address
1008 * @saddr: where to map the BO
1009 * @offset: requested offset in the BO
1010 * @flags: attributes of pages (read/write/valid/etc.)
1012 * Add a mapping of the BO at the specefied addr into the VM.
1013 * Returns 0 for success, error for failure.
1015 * Object has to be reserved and gets unreserved by this function!
1017 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
1018 struct amdgpu_bo_va *bo_va,
1019 uint64_t saddr, uint64_t offset,
1020 uint64_t size, uint32_t flags)
1022 struct amdgpu_bo_va_mapping *mapping;
1023 struct amdgpu_vm *vm = bo_va->vm;
1024 struct interval_tree_node *it;
1025 unsigned last_pfn, pt_idx;
1029 /* validate the parameters */
1030 if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
1031 size == 0 || size & AMDGPU_GPU_PAGE_MASK) {
1032 amdgpu_bo_unreserve(bo_va->bo);
1036 /* make sure object fit at this offset */
1037 eaddr = saddr + size;
1038 if ((saddr >= eaddr) || (offset + size > amdgpu_bo_size(bo_va->bo))) {
1039 amdgpu_bo_unreserve(bo_va->bo);
1043 last_pfn = eaddr / AMDGPU_GPU_PAGE_SIZE;
1044 if (last_pfn > adev->vm_manager.max_pfn) {
1045 dev_err(adev->dev, "va above limit (0x%08X > 0x%08X)\n",
1046 last_pfn, adev->vm_manager.max_pfn);
1047 amdgpu_bo_unreserve(bo_va->bo);
1051 mutex_lock(&vm->mutex);
1053 saddr /= AMDGPU_GPU_PAGE_SIZE;
1054 eaddr /= AMDGPU_GPU_PAGE_SIZE;
1056 it = interval_tree_iter_first(&vm->va, saddr, eaddr - 1);
1058 struct amdgpu_bo_va_mapping *tmp;
1059 tmp = container_of(it, struct amdgpu_bo_va_mapping, it);
1060 /* bo and tmp overlap, invalid addr */
1061 dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
1062 "0x%010lx-0x%010lx\n", bo_va->bo, saddr, eaddr,
1063 tmp->it.start, tmp->it.last + 1);
1064 amdgpu_bo_unreserve(bo_va->bo);
1069 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1071 amdgpu_bo_unreserve(bo_va->bo);
1076 INIT_LIST_HEAD(&mapping->list);
1077 mapping->it.start = saddr;
1078 mapping->it.last = eaddr - 1;
1079 mapping->offset = offset;
1080 mapping->flags = flags;
1082 list_add(&mapping->list, &bo_va->invalids);
1083 interval_tree_insert(&mapping->it, &vm->va);
1084 trace_amdgpu_vm_bo_map(bo_va, mapping);
1086 /* Make sure the page tables are allocated */
1087 saddr >>= amdgpu_vm_block_size;
1088 eaddr >>= amdgpu_vm_block_size;
1090 BUG_ON(eaddr >= amdgpu_vm_num_pdes(adev));
1092 if (eaddr > vm->max_pde_used)
1093 vm->max_pde_used = eaddr;
1095 amdgpu_bo_unreserve(bo_va->bo);
1097 /* walk over the address space and allocate the page tables */
1098 for (pt_idx = saddr; pt_idx <= eaddr; ++pt_idx) {
1099 struct amdgpu_bo *pt;
1101 if (vm->page_tables[pt_idx].bo)
1104 /* drop mutex to allocate and clear page table */
1105 mutex_unlock(&vm->mutex);
1107 r = amdgpu_bo_create(adev, AMDGPU_VM_PTE_COUNT * 8,
1108 AMDGPU_GPU_PAGE_SIZE, true,
1109 AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &pt);
1113 r = amdgpu_vm_clear_bo(adev, pt);
1115 amdgpu_bo_unref(&pt);
1119 /* aquire mutex again */
1120 mutex_lock(&vm->mutex);
1121 if (vm->page_tables[pt_idx].bo) {
1122 /* someone else allocated the pt in the meantime */
1123 mutex_unlock(&vm->mutex);
1124 amdgpu_bo_unref(&pt);
1125 mutex_lock(&vm->mutex);
1129 vm->page_tables[pt_idx].addr = 0;
1130 vm->page_tables[pt_idx].bo = pt;
1133 mutex_unlock(&vm->mutex);
1137 mutex_lock(&vm->mutex);
1138 list_del(&mapping->list);
1139 interval_tree_remove(&mapping->it, &vm->va);
1140 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1144 mutex_unlock(&vm->mutex);
1149 * amdgpu_vm_bo_unmap - remove bo mapping from vm
1151 * @adev: amdgpu_device pointer
1152 * @bo_va: bo_va to remove the address from
1153 * @saddr: where to the BO is mapped
1155 * Remove a mapping of the BO at the specefied addr from the VM.
1156 * Returns 0 for success, error for failure.
1158 * Object has to be reserved and gets unreserved by this function!
1160 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
1161 struct amdgpu_bo_va *bo_va,
1164 struct amdgpu_bo_va_mapping *mapping;
1165 struct amdgpu_vm *vm = bo_va->vm;
1168 saddr /= AMDGPU_GPU_PAGE_SIZE;
1170 list_for_each_entry(mapping, &bo_va->valids, list) {
1171 if (mapping->it.start == saddr)
1175 if (&mapping->list == &bo_va->valids) {
1178 list_for_each_entry(mapping, &bo_va->invalids, list) {
1179 if (mapping->it.start == saddr)
1183 if (&mapping->list == &bo_va->invalids) {
1184 amdgpu_bo_unreserve(bo_va->bo);
1189 mutex_lock(&vm->mutex);
1190 list_del(&mapping->list);
1191 interval_tree_remove(&mapping->it, &vm->va);
1192 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1195 list_add(&mapping->list, &vm->freed);
1198 mutex_unlock(&vm->mutex);
1199 amdgpu_bo_unreserve(bo_va->bo);
1205 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
1207 * @adev: amdgpu_device pointer
1208 * @bo_va: requested bo_va
1210 * Remove @bo_va->bo from the requested vm (cayman+).
1212 * Object have to be reserved!
1214 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
1215 struct amdgpu_bo_va *bo_va)
1217 struct amdgpu_bo_va_mapping *mapping, *next;
1218 struct amdgpu_vm *vm = bo_va->vm;
1220 list_del(&bo_va->bo_list);
1222 mutex_lock(&vm->mutex);
1224 spin_lock(&vm->status_lock);
1225 list_del(&bo_va->vm_status);
1226 spin_unlock(&vm->status_lock);
1228 list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
1229 list_del(&mapping->list);
1230 interval_tree_remove(&mapping->it, &vm->va);
1231 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1232 list_add(&mapping->list, &vm->freed);
1234 list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
1235 list_del(&mapping->list);
1236 interval_tree_remove(&mapping->it, &vm->va);
1240 fence_put(bo_va->last_pt_update);
1243 mutex_unlock(&vm->mutex);
1247 * amdgpu_vm_bo_invalidate - mark the bo as invalid
1249 * @adev: amdgpu_device pointer
1251 * @bo: amdgpu buffer object
1253 * Mark @bo as invalid (cayman+).
1255 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
1256 struct amdgpu_bo *bo)
1258 struct amdgpu_bo_va *bo_va;
1260 list_for_each_entry(bo_va, &bo->va, bo_list) {
1261 spin_lock(&bo_va->vm->status_lock);
1262 if (list_empty(&bo_va->vm_status))
1263 list_add(&bo_va->vm_status, &bo_va->vm->invalidated);
1264 spin_unlock(&bo_va->vm->status_lock);
1269 * amdgpu_vm_init - initialize a vm instance
1271 * @adev: amdgpu_device pointer
1274 * Init @vm fields (cayman+).
1276 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1278 const unsigned align = min(AMDGPU_VM_PTB_ALIGN_SIZE,
1279 AMDGPU_VM_PTE_COUNT * 8);
1280 unsigned pd_size, pd_entries, pts_size;
1283 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1285 vm->ids[i].flushed_updates = NULL;
1286 vm->ids[i].last_id_use = NULL;
1288 mutex_init(&vm->mutex);
1290 spin_lock_init(&vm->status_lock);
1291 INIT_LIST_HEAD(&vm->invalidated);
1292 INIT_LIST_HEAD(&vm->cleared);
1293 INIT_LIST_HEAD(&vm->freed);
1295 pd_size = amdgpu_vm_directory_size(adev);
1296 pd_entries = amdgpu_vm_num_pdes(adev);
1298 /* allocate page table array */
1299 pts_size = pd_entries * sizeof(struct amdgpu_vm_pt);
1300 vm->page_tables = kzalloc(pts_size, GFP_KERNEL);
1301 if (vm->page_tables == NULL) {
1302 DRM_ERROR("Cannot allocate memory for page table array\n");
1306 vm->page_directory_fence = NULL;
1308 r = amdgpu_bo_create(adev, pd_size, align, true,
1309 AMDGPU_GEM_DOMAIN_VRAM, 0,
1310 NULL, &vm->page_directory);
1314 r = amdgpu_vm_clear_bo(adev, vm->page_directory);
1316 amdgpu_bo_unref(&vm->page_directory);
1317 vm->page_directory = NULL;
1325 * amdgpu_vm_fini - tear down a vm instance
1327 * @adev: amdgpu_device pointer
1330 * Tear down @vm (cayman+).
1331 * Unbind the VM and remove all bos from the vm bo list
1333 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1335 struct amdgpu_bo_va_mapping *mapping, *tmp;
1338 if (!RB_EMPTY_ROOT(&vm->va)) {
1339 dev_err(adev->dev, "still active bo inside vm\n");
1341 rbtree_postorder_for_each_entry_safe(mapping, tmp, &vm->va, it.rb) {
1342 list_del(&mapping->list);
1343 interval_tree_remove(&mapping->it, &vm->va);
1346 list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
1347 list_del(&mapping->list);
1351 for (i = 0; i < amdgpu_vm_num_pdes(adev); i++)
1352 amdgpu_bo_unref(&vm->page_tables[i].bo);
1353 kfree(vm->page_tables);
1355 amdgpu_bo_unref(&vm->page_directory);
1356 fence_put(vm->page_directory_fence);
1358 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1359 fence_put(vm->ids[i].flushed_updates);
1360 amdgpu_fence_unref(&vm->ids[i].last_id_use);
1363 mutex_destroy(&vm->mutex);