2 * Copyright 2009 Jerome Glisse.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
28 * Jerome Glisse <glisse@freedesktop.org>
29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
33 #include <linux/dma-mapping.h>
34 #include <linux/iommu.h>
35 #include <linux/hmm.h>
36 #include <linux/pagemap.h>
37 #include <linux/sched/task.h>
38 #include <linux/sched/mm.h>
39 #include <linux/seq_file.h>
40 #include <linux/slab.h>
41 #include <linux/swap.h>
42 #include <linux/swiotlb.h>
43 #include <linux/dma-buf.h>
44 #include <linux/sizes.h>
46 #include <drm/ttm/ttm_bo_api.h>
47 #include <drm/ttm/ttm_bo_driver.h>
48 #include <drm/ttm/ttm_placement.h>
49 #include <drm/ttm/ttm_module.h>
51 #include <drm/drm_debugfs.h>
52 #include <drm/amdgpu_drm.h>
55 #include "amdgpu_object.h"
56 #include "amdgpu_trace.h"
57 #include "amdgpu_amdkfd.h"
58 #include "amdgpu_sdma.h"
59 #include "amdgpu_ras.h"
60 #include "amdgpu_atomfirmware.h"
61 #include "bif/bif_4_1_d.h"
63 #define AMDGPU_TTM_VRAM_MAX_DW_READ (size_t)128
65 static int amdgpu_ttm_backend_bind(struct ttm_bo_device *bdev,
67 struct ttm_resource *bo_mem);
68 static void amdgpu_ttm_backend_unbind(struct ttm_bo_device *bdev,
71 static int amdgpu_ttm_init_on_chip(struct amdgpu_device *adev,
73 uint64_t size_in_page)
75 return ttm_range_man_init(&adev->mman.bdev, type,
80 * amdgpu_evict_flags - Compute placement flags
82 * @bo: The buffer object to evict
83 * @placement: Possible destination(s) for evicted BO
85 * Fill in placement data when ttm_bo_evict() is called
87 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
88 struct ttm_placement *placement)
90 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
91 struct amdgpu_bo *abo;
92 static const struct ttm_place placements = {
95 .mem_type = TTM_PL_SYSTEM,
99 /* Don't handle scatter gather BOs */
100 if (bo->type == ttm_bo_type_sg) {
101 placement->num_placement = 0;
102 placement->num_busy_placement = 0;
106 /* Object isn't an AMDGPU object so ignore */
107 if (!amdgpu_bo_is_amdgpu_bo(bo)) {
108 placement->placement = &placements;
109 placement->busy_placement = &placements;
110 placement->num_placement = 1;
111 placement->num_busy_placement = 1;
115 abo = ttm_to_amdgpu_bo(bo);
116 switch (bo->mem.mem_type) {
120 placement->num_placement = 0;
121 placement->num_busy_placement = 0;
125 if (!adev->mman.buffer_funcs_enabled) {
126 /* Move to system memory */
127 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
128 } else if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
129 !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) &&
130 amdgpu_bo_in_cpu_visible_vram(abo)) {
132 /* Try evicting to the CPU inaccessible part of VRAM
133 * first, but only set GTT as busy placement, so this
134 * BO will be evicted to GTT rather than causing other
135 * BOs to be evicted from VRAM
137 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
138 AMDGPU_GEM_DOMAIN_GTT);
139 abo->placements[0].fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
140 abo->placements[0].lpfn = 0;
141 abo->placement.busy_placement = &abo->placements[1];
142 abo->placement.num_busy_placement = 1;
144 /* Move to GTT memory */
145 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT);
150 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
153 *placement = abo->placement;
157 * amdgpu_verify_access - Verify access for a mmap call
159 * @bo: The buffer object to map
160 * @filp: The file pointer from the process performing the mmap
162 * This is called by ttm_bo_mmap() to verify whether a process
163 * has the right to mmap a BO to their process space.
165 static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
167 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
170 * Don't verify access for KFD BOs. They don't have a GEM
171 * object associated with them.
176 if (amdgpu_ttm_tt_get_usermm(bo->ttm))
178 return drm_vma_node_verify_access(&abo->tbo.base.vma_node,
183 * amdgpu_mm_node_addr - Compute the GPU relative offset of a GTT buffer.
185 * @bo: The bo to assign the memory to.
186 * @mm_node: Memory manager node for drm allocator.
187 * @mem: The region where the bo resides.
190 static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
191 struct drm_mm_node *mm_node,
192 struct ttm_resource *mem)
196 if (mm_node->start != AMDGPU_BO_INVALID_OFFSET) {
197 addr = mm_node->start << PAGE_SHIFT;
198 addr += amdgpu_ttm_domain_start(amdgpu_ttm_adev(bo->bdev),
205 * amdgpu_find_mm_node - Helper function finds the drm_mm_node corresponding to
206 * @offset. It also modifies the offset to be within the drm_mm_node returned
208 * @mem: The region where the bo resides.
209 * @offset: The offset that drm_mm_node is used for finding.
212 static struct drm_mm_node *amdgpu_find_mm_node(struct ttm_resource *mem,
215 struct drm_mm_node *mm_node = mem->mm_node;
217 while (*offset >= (mm_node->size << PAGE_SHIFT)) {
218 *offset -= (mm_node->size << PAGE_SHIFT);
225 * amdgpu_ttm_map_buffer - Map memory into the GART windows
226 * @bo: buffer object to map
227 * @mem: memory object to map
228 * @mm_node: drm_mm node object to map
229 * @num_pages: number of pages to map
230 * @offset: offset into @mm_node where to start
231 * @window: which GART window to use
232 * @ring: DMA ring to use for the copy
233 * @tmz: if we should setup a TMZ enabled mapping
234 * @addr: resulting address inside the MC address space
236 * Setup one of the GART windows to access a specific piece of memory or return
237 * the physical address for local memory.
239 static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo,
240 struct ttm_resource *mem,
241 struct drm_mm_node *mm_node,
242 unsigned num_pages, uint64_t offset,
243 unsigned window, struct amdgpu_ring *ring,
244 bool tmz, uint64_t *addr)
246 struct amdgpu_device *adev = ring->adev;
247 struct amdgpu_job *job;
248 unsigned num_dw, num_bytes;
249 struct dma_fence *fence;
250 uint64_t src_addr, dst_addr;
256 BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
257 AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
259 /* Map only what can't be accessed directly */
260 if (!tmz && mem->start != AMDGPU_BO_INVALID_OFFSET) {
261 *addr = amdgpu_mm_node_addr(bo, mm_node, mem) + offset;
265 *addr = adev->gmc.gart_start;
266 *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE *
267 AMDGPU_GPU_PAGE_SIZE;
268 *addr += offset & ~PAGE_MASK;
270 num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
271 num_bytes = num_pages * 8;
273 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes,
274 AMDGPU_IB_POOL_DELAYED, &job);
278 src_addr = num_dw * 4;
279 src_addr += job->ibs[0].gpu_addr;
281 dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
282 dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
283 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
284 dst_addr, num_bytes, false);
286 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
287 WARN_ON(job->ibs[0].length_dw > num_dw);
289 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, mem);
291 flags |= AMDGPU_PTE_TMZ;
293 cpu_addr = &job->ibs[0].ptr[num_dw];
295 if (mem->mem_type == TTM_PL_TT) {
296 dma_addr_t *dma_address;
298 dma_address = &bo->ttm->dma_address[offset >> PAGE_SHIFT];
299 r = amdgpu_gart_map(adev, 0, num_pages, dma_address, flags,
304 dma_addr_t dma_address;
306 dma_address = (mm_node->start << PAGE_SHIFT) + offset;
307 dma_address += adev->vm_manager.vram_base_offset;
309 for (i = 0; i < num_pages; ++i) {
310 r = amdgpu_gart_map(adev, i << PAGE_SHIFT, 1,
311 &dma_address, flags, cpu_addr);
315 dma_address += PAGE_SIZE;
319 r = amdgpu_job_submit(job, &adev->mman.entity,
320 AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
324 dma_fence_put(fence);
329 amdgpu_job_free(job);
334 * amdgpu_copy_ttm_mem_to_mem - Helper function for copy
335 * @adev: amdgpu device
336 * @src: buffer/address where to read from
337 * @dst: buffer/address where to write to
338 * @size: number of bytes to copy
339 * @tmz: if a secure copy should be used
340 * @resv: resv object to sync to
341 * @f: Returns the last fence if multiple jobs are submitted.
343 * The function copies @size bytes from {src->mem + src->offset} to
344 * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a
345 * move and different for a BO to BO copy.
348 int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev,
349 const struct amdgpu_copy_mem *src,
350 const struct amdgpu_copy_mem *dst,
351 uint64_t size, bool tmz,
352 struct dma_resv *resv,
353 struct dma_fence **f)
355 const uint32_t GTT_MAX_BYTES = (AMDGPU_GTT_MAX_TRANSFER_SIZE *
356 AMDGPU_GPU_PAGE_SIZE);
358 uint64_t src_node_size, dst_node_size, src_offset, dst_offset;
359 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
360 struct drm_mm_node *src_mm, *dst_mm;
361 struct dma_fence *fence = NULL;
364 if (!adev->mman.buffer_funcs_enabled) {
365 DRM_ERROR("Trying to move memory with ring turned off.\n");
369 src_offset = src->offset;
370 if (src->mem->mm_node) {
371 src_mm = amdgpu_find_mm_node(src->mem, &src_offset);
372 src_node_size = (src_mm->size << PAGE_SHIFT) - src_offset;
375 src_node_size = ULLONG_MAX;
378 dst_offset = dst->offset;
379 if (dst->mem->mm_node) {
380 dst_mm = amdgpu_find_mm_node(dst->mem, &dst_offset);
381 dst_node_size = (dst_mm->size << PAGE_SHIFT) - dst_offset;
384 dst_node_size = ULLONG_MAX;
387 mutex_lock(&adev->mman.gtt_window_lock);
390 uint32_t src_page_offset = src_offset & ~PAGE_MASK;
391 uint32_t dst_page_offset = dst_offset & ~PAGE_MASK;
392 struct dma_fence *next;
396 /* Copy size cannot exceed GTT_MAX_BYTES. So if src or dst
397 * begins at an offset, then adjust the size accordingly
399 cur_size = max(src_page_offset, dst_page_offset);
400 cur_size = min(min3(src_node_size, dst_node_size, size),
401 (uint64_t)(GTT_MAX_BYTES - cur_size));
403 /* Map src to window 0 and dst to window 1. */
404 r = amdgpu_ttm_map_buffer(src->bo, src->mem, src_mm,
405 PFN_UP(cur_size + src_page_offset),
406 src_offset, 0, ring, tmz, &from);
410 r = amdgpu_ttm_map_buffer(dst->bo, dst->mem, dst_mm,
411 PFN_UP(cur_size + dst_page_offset),
412 dst_offset, 1, ring, tmz, &to);
416 r = amdgpu_copy_buffer(ring, from, to, cur_size,
417 resv, &next, false, true, tmz);
421 dma_fence_put(fence);
428 src_node_size -= cur_size;
429 if (!src_node_size) {
431 src_node_size = src_mm->size << PAGE_SHIFT;
434 src_offset += cur_size;
437 dst_node_size -= cur_size;
438 if (!dst_node_size) {
440 dst_node_size = dst_mm->size << PAGE_SHIFT;
443 dst_offset += cur_size;
447 mutex_unlock(&adev->mman.gtt_window_lock);
449 *f = dma_fence_get(fence);
450 dma_fence_put(fence);
455 * amdgpu_move_blit - Copy an entire buffer to another buffer
457 * This is a helper called by amdgpu_bo_move() and amdgpu_move_vram_ram() to
458 * help move buffers to and from VRAM.
460 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
462 struct ttm_resource *new_mem,
463 struct ttm_resource *old_mem)
465 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
466 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
467 struct amdgpu_copy_mem src, dst;
468 struct dma_fence *fence = NULL;
478 r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
479 new_mem->num_pages << PAGE_SHIFT,
480 amdgpu_bo_encrypted(abo),
481 bo->base.resv, &fence);
485 /* clear the space being freed */
486 if (old_mem->mem_type == TTM_PL_VRAM &&
487 (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) {
488 struct dma_fence *wipe_fence = NULL;
490 r = amdgpu_fill_buffer(ttm_to_amdgpu_bo(bo), AMDGPU_POISON,
494 } else if (wipe_fence) {
495 dma_fence_put(fence);
500 /* Always block for VM page tables before committing the new location */
501 if (bo->type == ttm_bo_type_kernel)
502 r = ttm_bo_move_accel_cleanup(bo, fence, true, false, new_mem);
504 r = ttm_bo_move_accel_cleanup(bo, fence, evict, true, new_mem);
505 dma_fence_put(fence);
510 dma_fence_wait(fence, false);
511 dma_fence_put(fence);
516 * amdgpu_mem_visible - Check that memory can be accessed by ttm_bo_move_memcpy
518 * Called by amdgpu_bo_move()
520 static bool amdgpu_mem_visible(struct amdgpu_device *adev,
521 struct ttm_resource *mem)
523 struct drm_mm_node *nodes = mem->mm_node;
525 if (mem->mem_type == TTM_PL_SYSTEM ||
526 mem->mem_type == TTM_PL_TT)
528 if (mem->mem_type != TTM_PL_VRAM)
531 /* ttm_resource_ioremap only supports contiguous memory */
532 if (nodes->size != mem->num_pages)
535 return ((nodes->start + nodes->size) << PAGE_SHIFT)
536 <= adev->gmc.visible_vram_size;
540 * amdgpu_bo_move - Move a buffer object to a new memory location
542 * Called by ttm_bo_handle_move_mem()
544 static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
545 struct ttm_operation_ctx *ctx,
546 struct ttm_resource *new_mem,
547 struct ttm_place *hop)
549 struct amdgpu_device *adev;
550 struct amdgpu_bo *abo;
551 struct ttm_resource *old_mem = &bo->mem;
554 if (new_mem->mem_type == TTM_PL_TT) {
555 r = amdgpu_ttm_backend_bind(bo->bdev, bo->ttm, new_mem);
560 /* Can't move a pinned BO */
561 abo = ttm_to_amdgpu_bo(bo);
562 if (WARN_ON_ONCE(abo->tbo.pin_count > 0))
565 adev = amdgpu_ttm_adev(bo->bdev);
567 if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
568 ttm_bo_move_null(bo, new_mem);
571 if (old_mem->mem_type == TTM_PL_SYSTEM &&
572 new_mem->mem_type == TTM_PL_TT) {
573 ttm_bo_move_null(bo, new_mem);
576 if (old_mem->mem_type == TTM_PL_TT &&
577 new_mem->mem_type == TTM_PL_SYSTEM) {
578 r = ttm_bo_wait_ctx(bo, ctx);
582 amdgpu_ttm_backend_unbind(bo->bdev, bo->ttm);
583 ttm_resource_free(bo, &bo->mem);
584 ttm_bo_assign_mem(bo, new_mem);
588 if (old_mem->mem_type == AMDGPU_PL_GDS ||
589 old_mem->mem_type == AMDGPU_PL_GWS ||
590 old_mem->mem_type == AMDGPU_PL_OA ||
591 new_mem->mem_type == AMDGPU_PL_GDS ||
592 new_mem->mem_type == AMDGPU_PL_GWS ||
593 new_mem->mem_type == AMDGPU_PL_OA) {
594 /* Nothing to save here */
595 ttm_bo_move_null(bo, new_mem);
599 if (adev->mman.buffer_funcs_enabled) {
600 if (((old_mem->mem_type == TTM_PL_SYSTEM &&
601 new_mem->mem_type == TTM_PL_VRAM) ||
602 (old_mem->mem_type == TTM_PL_VRAM &&
603 new_mem->mem_type == TTM_PL_SYSTEM))) {
606 hop->mem_type = TTM_PL_TT;
611 r = amdgpu_move_blit(bo, evict, new_mem, old_mem);
617 /* Check that all memory is CPU accessible */
618 if (!amdgpu_mem_visible(adev, old_mem) ||
619 !amdgpu_mem_visible(adev, new_mem)) {
620 pr_err("Move buffer fallback to memcpy unavailable\n");
624 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
629 if (bo->type == ttm_bo_type_device &&
630 new_mem->mem_type == TTM_PL_VRAM &&
631 old_mem->mem_type != TTM_PL_VRAM) {
632 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
633 * accesses the BO after it's moved.
635 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
639 /* update statistics */
640 atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved);
641 amdgpu_bo_move_notify(bo, evict, new_mem);
646 * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault
648 * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault()
650 static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_resource *mem)
652 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
653 struct drm_mm_node *mm_node = mem->mm_node;
654 size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT;
656 switch (mem->mem_type) {
663 mem->bus.offset = mem->start << PAGE_SHIFT;
664 /* check if it's visible */
665 if ((mem->bus.offset + bus_size) > adev->gmc.visible_vram_size)
667 /* Only physically contiguous buffers apply. In a contiguous
668 * buffer, size of the first mm_node would match the number of
669 * pages in ttm_resource.
671 if (adev->mman.aper_base_kaddr &&
672 (mm_node->size == mem->num_pages))
673 mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
676 mem->bus.offset += adev->gmc.aper_base;
677 mem->bus.is_iomem = true;
678 mem->bus.caching = ttm_write_combined;
686 static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
687 unsigned long page_offset)
689 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
690 uint64_t offset = (page_offset << PAGE_SHIFT);
691 struct drm_mm_node *mm;
693 mm = amdgpu_find_mm_node(&bo->mem, &offset);
694 offset += adev->gmc.aper_base;
695 return mm->start + (offset >> PAGE_SHIFT);
699 * amdgpu_ttm_domain_start - Returns GPU start address
700 * @adev: amdgpu device object
701 * @type: type of the memory
704 * GPU start address of a memory domain
707 uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type)
711 return adev->gmc.gart_start;
713 return adev->gmc.vram_start;
720 * TTM backend functions.
722 struct amdgpu_ttm_tt {
724 struct drm_gem_object *gobj;
727 struct task_struct *usertask;
730 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
731 struct hmm_range *range;
735 #ifdef CONFIG_DRM_AMDGPU_USERPTR
737 * amdgpu_ttm_tt_get_user_pages - get device accessible pages that back user
738 * memory and start HMM tracking CPU page table update
740 * Calling function must call amdgpu_ttm_tt_userptr_range_done() once and only
741 * once afterwards to stop HMM tracking
743 int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages)
745 struct ttm_tt *ttm = bo->tbo.ttm;
746 struct amdgpu_ttm_tt *gtt = (void *)ttm;
747 unsigned long start = gtt->userptr;
748 struct vm_area_struct *vma;
749 struct hmm_range *range;
750 unsigned long timeout;
751 struct mm_struct *mm;
755 mm = bo->notifier.mm;
757 DRM_DEBUG_DRIVER("BO is not registered?\n");
761 /* Another get_user_pages is running at the same time?? */
762 if (WARN_ON(gtt->range))
765 if (!mmget_not_zero(mm)) /* Happens during process shutdown */
768 range = kzalloc(sizeof(*range), GFP_KERNEL);
769 if (unlikely(!range)) {
773 range->notifier = &bo->notifier;
774 range->start = bo->notifier.interval_tree.start;
775 range->end = bo->notifier.interval_tree.last + 1;
776 range->default_flags = HMM_PFN_REQ_FAULT;
777 if (!amdgpu_ttm_tt_is_readonly(ttm))
778 range->default_flags |= HMM_PFN_REQ_WRITE;
780 range->hmm_pfns = kvmalloc_array(ttm->num_pages,
781 sizeof(*range->hmm_pfns), GFP_KERNEL);
782 if (unlikely(!range->hmm_pfns)) {
784 goto out_free_ranges;
788 vma = find_vma(mm, start);
789 if (unlikely(!vma || start < vma->vm_start)) {
793 if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) &&
798 mmap_read_unlock(mm);
799 timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
802 range->notifier_seq = mmu_interval_read_begin(&bo->notifier);
805 r = hmm_range_fault(range);
806 mmap_read_unlock(mm);
809 * FIXME: This timeout should encompass the retry from
810 * mmu_interval_read_retry() as well.
812 if (r == -EBUSY && !time_after(jiffies, timeout))
818 * Due to default_flags, all pages are HMM_PFN_VALID or
819 * hmm_range_fault() fails. FIXME: The pages cannot be touched outside
820 * the notifier_lock, and mmu_interval_read_retry() must be done first.
822 for (i = 0; i < ttm->num_pages; i++)
823 pages[i] = hmm_pfn_to_page(range->hmm_pfns[i]);
831 mmap_read_unlock(mm);
833 kvfree(range->hmm_pfns);
842 * amdgpu_ttm_tt_userptr_range_done - stop HMM track the CPU page table change
843 * Check if the pages backing this ttm range have been invalidated
845 * Returns: true if pages are still valid
847 bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm)
849 struct amdgpu_ttm_tt *gtt = (void *)ttm;
852 if (!gtt || !gtt->userptr)
855 DRM_DEBUG_DRIVER("user_pages_done 0x%llx pages 0x%x\n",
856 gtt->userptr, ttm->num_pages);
858 WARN_ONCE(!gtt->range || !gtt->range->hmm_pfns,
859 "No user pages to check\n");
863 * FIXME: Must always hold notifier_lock for this, and must
864 * not ignore the return code.
866 r = mmu_interval_read_retry(gtt->range->notifier,
867 gtt->range->notifier_seq);
868 kvfree(gtt->range->hmm_pfns);
878 * amdgpu_ttm_tt_set_user_pages - Copy pages in, putting old pages as necessary.
880 * Called by amdgpu_cs_list_validate(). This creates the page list
881 * that backs user memory and will ultimately be mapped into the device
884 void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages)
888 for (i = 0; i < ttm->num_pages; ++i)
889 ttm->pages[i] = pages ? pages[i] : NULL;
893 * amdgpu_ttm_tt_pin_userptr - prepare the sg table with the user pages
895 * Called by amdgpu_ttm_backend_bind()
897 static int amdgpu_ttm_tt_pin_userptr(struct ttm_bo_device *bdev,
900 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
901 struct amdgpu_ttm_tt *gtt = (void *)ttm;
904 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
905 enum dma_data_direction direction = write ?
906 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
908 /* Allocate an SG array and squash pages into it */
909 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
910 ttm->num_pages << PAGE_SHIFT,
915 /* Map SG to device */
916 r = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
920 /* convert SG to linear array of pages and dma addresses */
921 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
922 gtt->ttm.dma_address, ttm->num_pages);
933 * amdgpu_ttm_tt_unpin_userptr - Unpin and unmap userptr pages
935 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_bo_device *bdev,
938 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
939 struct amdgpu_ttm_tt *gtt = (void *)ttm;
941 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
942 enum dma_data_direction direction = write ?
943 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
945 /* double check that we don't free the table twice */
949 /* unmap the pages mapped to the device */
950 dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
951 sg_free_table(ttm->sg);
953 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
957 for (i = 0; i < ttm->num_pages; i++) {
959 hmm_pfn_to_page(gtt->range->hmm_pfns[i]))
963 WARN((i == ttm->num_pages), "Missing get_user_page_done\n");
968 static int amdgpu_ttm_gart_bind(struct amdgpu_device *adev,
969 struct ttm_buffer_object *tbo,
972 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(tbo);
973 struct ttm_tt *ttm = tbo->ttm;
974 struct amdgpu_ttm_tt *gtt = (void *)ttm;
977 if (amdgpu_bo_encrypted(abo))
978 flags |= AMDGPU_PTE_TMZ;
980 if (abo->flags & AMDGPU_GEM_CREATE_CP_MQD_GFX9) {
981 uint64_t page_idx = 1;
983 r = amdgpu_gart_bind(adev, gtt->offset, page_idx,
984 ttm->pages, gtt->ttm.dma_address, flags);
988 /* The memory type of the first page defaults to UC. Now
989 * modify the memory type to NC from the second page of
992 flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
993 flags |= AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC);
995 r = amdgpu_gart_bind(adev,
996 gtt->offset + (page_idx << PAGE_SHIFT),
997 ttm->num_pages - page_idx,
998 &ttm->pages[page_idx],
999 &(gtt->ttm.dma_address[page_idx]), flags);
1001 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
1002 ttm->pages, gtt->ttm.dma_address, flags);
1007 DRM_ERROR("failed to bind %u pages at 0x%08llX\n",
1008 ttm->num_pages, gtt->offset);
1014 * amdgpu_ttm_backend_bind - Bind GTT memory
1016 * Called by ttm_tt_bind() on behalf of ttm_bo_handle_move_mem().
1017 * This handles binding GTT memory to the device address space.
1019 static int amdgpu_ttm_backend_bind(struct ttm_bo_device *bdev,
1021 struct ttm_resource *bo_mem)
1023 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1024 struct amdgpu_ttm_tt *gtt = (void*)ttm;
1035 r = amdgpu_ttm_tt_pin_userptr(bdev, ttm);
1037 DRM_ERROR("failed to pin userptr\n");
1041 if (!ttm->num_pages) {
1042 WARN(1, "nothing to bind %u pages for mreg %p back %p!\n",
1043 ttm->num_pages, bo_mem, ttm);
1046 if (bo_mem->mem_type == AMDGPU_PL_GDS ||
1047 bo_mem->mem_type == AMDGPU_PL_GWS ||
1048 bo_mem->mem_type == AMDGPU_PL_OA)
1051 if (!amdgpu_gtt_mgr_has_gart_addr(bo_mem)) {
1052 gtt->offset = AMDGPU_BO_INVALID_OFFSET;
1056 /* compute PTE flags relevant to this BO memory */
1057 flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
1059 /* bind pages into GART page tables */
1060 gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
1061 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
1062 ttm->pages, gtt->ttm.dma_address, flags);
1065 DRM_ERROR("failed to bind %u pages at 0x%08llX\n",
1066 ttm->num_pages, gtt->offset);
1072 * amdgpu_ttm_alloc_gart - Make sure buffer object is accessible either
1073 * through AGP or GART aperture.
1075 * If bo is accessible through AGP aperture, then use AGP aperture
1076 * to access bo; otherwise allocate logical space in GART aperture
1077 * and map bo to GART aperture.
1079 int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
1081 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1082 struct ttm_operation_ctx ctx = { false, false };
1083 struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
1084 struct ttm_resource tmp;
1085 struct ttm_placement placement;
1086 struct ttm_place placements;
1087 uint64_t addr, flags;
1090 if (bo->mem.start != AMDGPU_BO_INVALID_OFFSET)
1093 addr = amdgpu_gmc_agp_addr(bo);
1094 if (addr != AMDGPU_BO_INVALID_OFFSET) {
1095 bo->mem.start = addr >> PAGE_SHIFT;
1098 /* allocate GART space */
1101 placement.num_placement = 1;
1102 placement.placement = &placements;
1103 placement.num_busy_placement = 1;
1104 placement.busy_placement = &placements;
1105 placements.fpfn = 0;
1106 placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
1107 placements.mem_type = TTM_PL_TT;
1108 placements.flags = bo->mem.placement;
1110 r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx);
1114 /* compute PTE flags for this buffer object */
1115 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, &tmp);
1118 gtt->offset = (u64)tmp.start << PAGE_SHIFT;
1119 r = amdgpu_ttm_gart_bind(adev, bo, flags);
1121 ttm_resource_free(bo, &tmp);
1125 ttm_resource_free(bo, &bo->mem);
1133 * amdgpu_ttm_recover_gart - Rebind GTT pages
1135 * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to
1136 * rebind GTT pages during a GPU reset.
1138 int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
1140 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
1147 flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, &tbo->mem);
1148 r = amdgpu_ttm_gart_bind(adev, tbo, flags);
1154 * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages
1156 * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and
1159 static void amdgpu_ttm_backend_unbind(struct ttm_bo_device *bdev,
1162 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1163 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1169 /* if the pages have userptr pinning then clear that first */
1171 amdgpu_ttm_tt_unpin_userptr(bdev, ttm);
1173 if (gtt->offset == AMDGPU_BO_INVALID_OFFSET)
1176 /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
1177 r = amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages);
1179 DRM_ERROR("failed to unbind %u pages at 0x%08llX\n",
1180 gtt->ttm.num_pages, gtt->offset);
1184 static void amdgpu_ttm_backend_destroy(struct ttm_bo_device *bdev,
1187 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1189 amdgpu_ttm_backend_unbind(bdev, ttm);
1190 ttm_tt_destroy_common(bdev, ttm);
1192 put_task_struct(gtt->usertask);
1194 ttm_tt_fini(>t->ttm);
1199 * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO
1201 * @bo: The buffer object to create a GTT ttm_tt object around
1203 * Called by ttm_tt_create().
1205 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
1206 uint32_t page_flags)
1208 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1209 struct amdgpu_ttm_tt *gtt;
1210 enum ttm_caching caching;
1212 gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
1216 gtt->gobj = &bo->base;
1218 if (abo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
1219 caching = ttm_write_combined;
1221 caching = ttm_cached;
1223 /* allocate space for the uninitialized page entries */
1224 if (ttm_sg_tt_init(>t->ttm, bo, page_flags, caching)) {
1232 * amdgpu_ttm_tt_populate - Map GTT pages visible to the device
1234 * Map the pages of a ttm_tt object to an address space visible
1235 * to the underlying device.
1237 static int amdgpu_ttm_tt_populate(struct ttm_bo_device *bdev,
1239 struct ttm_operation_ctx *ctx)
1241 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1242 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1244 /* user pages are bound by amdgpu_ttm_tt_pin_userptr() */
1245 if (gtt && gtt->userptr) {
1246 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
1250 ttm->page_flags |= TTM_PAGE_FLAG_SG;
1254 if (ttm->page_flags & TTM_PAGE_FLAG_SG) {
1256 struct dma_buf_attachment *attach;
1257 struct sg_table *sgt;
1259 attach = gtt->gobj->import_attach;
1260 sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
1262 return PTR_ERR(sgt);
1267 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1268 gtt->ttm.dma_address,
1273 return ttm_pool_alloc(&adev->mman.bdev.pool, ttm, ctx);
1277 * amdgpu_ttm_tt_unpopulate - unmap GTT pages and unpopulate page arrays
1279 * Unmaps pages of a ttm_tt object from the device address space and
1280 * unpopulates the page array backing it.
1282 static void amdgpu_ttm_tt_unpopulate(struct ttm_bo_device *bdev,
1285 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1286 struct amdgpu_device *adev;
1288 if (gtt && gtt->userptr) {
1289 amdgpu_ttm_tt_set_user_pages(ttm, NULL);
1291 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
1295 if (ttm->sg && gtt->gobj->import_attach) {
1296 struct dma_buf_attachment *attach;
1298 attach = gtt->gobj->import_attach;
1299 dma_buf_unmap_attachment(attach, ttm->sg, DMA_BIDIRECTIONAL);
1304 if (ttm->page_flags & TTM_PAGE_FLAG_SG)
1307 adev = amdgpu_ttm_adev(bdev);
1308 return ttm_pool_free(&adev->mman.bdev.pool, ttm);
1312 * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt for the current
1315 * @bo: The ttm_buffer_object to bind this userptr to
1316 * @addr: The address in the current tasks VM space to use
1317 * @flags: Requirements of userptr object.
1319 * Called by amdgpu_gem_userptr_ioctl() to bind userptr pages
1322 int amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object *bo,
1323 uint64_t addr, uint32_t flags)
1325 struct amdgpu_ttm_tt *gtt;
1328 /* TODO: We want a separate TTM object type for userptrs */
1329 bo->ttm = amdgpu_ttm_tt_create(bo, 0);
1330 if (bo->ttm == NULL)
1334 gtt = (void *)bo->ttm;
1335 gtt->userptr = addr;
1336 gtt->userflags = flags;
1339 put_task_struct(gtt->usertask);
1340 gtt->usertask = current->group_leader;
1341 get_task_struct(gtt->usertask);
1347 * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object
1349 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1351 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1356 if (gtt->usertask == NULL)
1359 return gtt->usertask->mm;
1363 * amdgpu_ttm_tt_affect_userptr - Determine if a ttm_tt object lays inside an
1364 * address range for the current task.
1367 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1370 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1373 if (gtt == NULL || !gtt->userptr)
1376 /* Return false if no part of the ttm_tt object lies within
1379 size = (unsigned long)gtt->ttm.num_pages * PAGE_SIZE;
1380 if (gtt->userptr > end || gtt->userptr + size <= start)
1387 * amdgpu_ttm_tt_is_userptr - Have the pages backing by userptr?
1389 bool amdgpu_ttm_tt_is_userptr(struct ttm_tt *ttm)
1391 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1393 if (gtt == NULL || !gtt->userptr)
1400 * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only?
1402 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1404 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1409 return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1413 * amdgpu_ttm_tt_pde_flags - Compute PDE flags for ttm_tt object
1415 * @ttm: The ttm_tt object to compute the flags for
1416 * @mem: The memory registry backing this ttm_tt object
1418 * Figure out the flags to use for a VM PDE (Page Directory Entry).
1420 uint64_t amdgpu_ttm_tt_pde_flags(struct ttm_tt *ttm, struct ttm_resource *mem)
1424 if (mem && mem->mem_type != TTM_PL_SYSTEM)
1425 flags |= AMDGPU_PTE_VALID;
1427 if (mem && mem->mem_type == TTM_PL_TT) {
1428 flags |= AMDGPU_PTE_SYSTEM;
1430 if (ttm->caching == ttm_cached)
1431 flags |= AMDGPU_PTE_SNOOPED;
1438 * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object
1440 * @ttm: The ttm_tt object to compute the flags for
1441 * @mem: The memory registry backing this ttm_tt object
1443 * Figure out the flags to use for a VM PTE (Page Table Entry).
1445 uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1446 struct ttm_resource *mem)
1448 uint64_t flags = amdgpu_ttm_tt_pde_flags(ttm, mem);
1450 flags |= adev->gart.gart_pte_flags;
1451 flags |= AMDGPU_PTE_READABLE;
1453 if (!amdgpu_ttm_tt_is_readonly(ttm))
1454 flags |= AMDGPU_PTE_WRITEABLE;
1460 * amdgpu_ttm_bo_eviction_valuable - Check to see if we can evict a buffer
1463 * Return true if eviction is sensible. Called by ttm_mem_evict_first() on
1464 * behalf of ttm_bo_mem_force_space() which tries to evict buffer objects until
1465 * it can find space for a new object and by ttm_bo_force_list_clean() which is
1466 * used to clean out a memory space.
1468 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1469 const struct ttm_place *place)
1471 unsigned long num_pages = bo->mem.num_pages;
1472 struct drm_mm_node *node = bo->mem.mm_node;
1473 struct dma_resv_list *flist;
1474 struct dma_fence *f;
1477 if (bo->type == ttm_bo_type_kernel &&
1478 !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo)))
1481 /* If bo is a KFD BO, check if the bo belongs to the current process.
1482 * If true, then return false as any KFD process needs all its BOs to
1483 * be resident to run successfully
1485 flist = dma_resv_get_list(bo->base.resv);
1487 for (i = 0; i < flist->shared_count; ++i) {
1488 f = rcu_dereference_protected(flist->shared[i],
1489 dma_resv_held(bo->base.resv));
1490 if (amdkfd_fence_check_mm(f, current->mm))
1495 switch (bo->mem.mem_type) {
1497 if (amdgpu_bo_is_amdgpu_bo(bo) &&
1498 amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo)))
1503 /* Check each drm MM node individually */
1505 if (place->fpfn < (node->start + node->size) &&
1506 !(place->lpfn && place->lpfn <= node->start))
1509 num_pages -= node->size;
1518 return ttm_bo_eviction_valuable(bo, place);
1522 * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object.
1524 * @bo: The buffer object to read/write
1525 * @offset: Offset into buffer object
1526 * @buf: Secondary buffer to write/read from
1527 * @len: Length in bytes of access
1528 * @write: true if writing
1530 * This is used to access VRAM that backs a buffer object via MMIO
1531 * access for debugging purposes.
1533 static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1534 unsigned long offset,
1535 void *buf, int len, int write)
1537 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1538 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1539 struct drm_mm_node *nodes;
1543 unsigned long flags;
1545 if (bo->mem.mem_type != TTM_PL_VRAM)
1549 nodes = amdgpu_find_mm_node(&abo->tbo.mem, &pos);
1550 pos += (nodes->start << PAGE_SHIFT);
1552 while (len && pos < adev->gmc.mc_vram_size) {
1553 uint64_t aligned_pos = pos & ~(uint64_t)3;
1554 uint64_t bytes = 4 - (pos & 3);
1555 uint32_t shift = (pos & 3) * 8;
1556 uint32_t mask = 0xffffffff << shift;
1559 mask &= 0xffffffff >> (bytes - len) * 8;
1563 if (mask != 0xffffffff) {
1564 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1565 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)aligned_pos) | 0x80000000);
1566 WREG32_NO_KIQ(mmMM_INDEX_HI, aligned_pos >> 31);
1567 if (!write || mask != 0xffffffff)
1568 value = RREG32_NO_KIQ(mmMM_DATA);
1571 value |= (*(uint32_t *)buf << shift) & mask;
1572 WREG32_NO_KIQ(mmMM_DATA, value);
1574 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1576 value = (value & mask) >> shift;
1577 memcpy(buf, &value, bytes);
1580 bytes = (nodes->start + nodes->size) << PAGE_SHIFT;
1581 bytes = min(bytes - pos, (uint64_t)len & ~0x3ull);
1583 amdgpu_device_vram_access(adev, pos, (uint32_t *)buf,
1588 buf = (uint8_t *)buf + bytes;
1591 if (pos >= (nodes->start + nodes->size) << PAGE_SHIFT) {
1593 pos = (nodes->start << PAGE_SHIFT);
1601 amdgpu_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1603 amdgpu_bo_move_notify(bo, false, NULL);
1606 static struct ttm_bo_driver amdgpu_bo_driver = {
1607 .ttm_tt_create = &amdgpu_ttm_tt_create,
1608 .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1609 .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1610 .ttm_tt_destroy = &amdgpu_ttm_backend_destroy,
1611 .eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1612 .evict_flags = &amdgpu_evict_flags,
1613 .move = &amdgpu_bo_move,
1614 .verify_access = &amdgpu_verify_access,
1615 .delete_mem_notify = &amdgpu_bo_delete_mem_notify,
1616 .release_notify = &amdgpu_bo_release_notify,
1617 .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1618 .io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1619 .access_memory = &amdgpu_ttm_access_memory,
1620 .del_from_lru_notify = &amdgpu_vm_del_from_lru_notify
1624 * Firmware Reservation functions
1627 * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram
1629 * @adev: amdgpu_device pointer
1631 * free fw reserved vram if it has been reserved.
1633 static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
1635 amdgpu_bo_free_kernel(&adev->mman.fw_vram_usage_reserved_bo,
1636 NULL, &adev->mman.fw_vram_usage_va);
1640 * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
1642 * @adev: amdgpu_device pointer
1644 * create bo vram reservation from fw.
1646 static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
1648 uint64_t vram_size = adev->gmc.visible_vram_size;
1650 adev->mman.fw_vram_usage_va = NULL;
1651 adev->mman.fw_vram_usage_reserved_bo = NULL;
1653 if (adev->mman.fw_vram_usage_size == 0 ||
1654 adev->mman.fw_vram_usage_size > vram_size)
1657 return amdgpu_bo_create_kernel_at(adev,
1658 adev->mman.fw_vram_usage_start_offset,
1659 adev->mman.fw_vram_usage_size,
1660 AMDGPU_GEM_DOMAIN_VRAM,
1661 &adev->mman.fw_vram_usage_reserved_bo,
1662 &adev->mman.fw_vram_usage_va);
1666 * Memoy training reservation functions
1670 * amdgpu_ttm_training_reserve_vram_fini - free memory training reserved vram
1672 * @adev: amdgpu_device pointer
1674 * free memory training reserved vram if it has been reserved.
1676 static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev)
1678 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1680 ctx->init = PSP_MEM_TRAIN_NOT_SUPPORT;
1681 amdgpu_bo_free_kernel(&ctx->c2p_bo, NULL, NULL);
1687 static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev)
1689 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1691 memset(ctx, 0, sizeof(*ctx));
1693 ctx->c2p_train_data_offset =
1694 ALIGN((adev->gmc.mc_vram_size - adev->mman.discovery_tmr_size - SZ_1M), SZ_1M);
1695 ctx->p2c_train_data_offset =
1696 (adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET);
1697 ctx->train_data_size =
1698 GDDR6_MEM_TRAINING_DATA_SIZE_IN_BYTES;
1700 DRM_DEBUG("train_data_size:%llx,p2c_train_data_offset:%llx,c2p_train_data_offset:%llx.\n",
1701 ctx->train_data_size,
1702 ctx->p2c_train_data_offset,
1703 ctx->c2p_train_data_offset);
1707 * reserve TMR memory at the top of VRAM which holds
1708 * IP Discovery data and is protected by PSP.
1710 static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
1713 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1714 bool mem_train_support = false;
1716 if (!amdgpu_sriov_vf(adev)) {
1717 ret = amdgpu_mem_train_support(adev);
1719 mem_train_support = true;
1723 DRM_DEBUG("memory training does not support!\n");
1727 * Query reserved tmr size through atom firmwareinfo for Sienna_Cichlid and onwards for all
1728 * the use cases (IP discovery/G6 memory training/profiling/diagnostic data.etc)
1730 * Otherwise, fallback to legacy approach to check and reserve tmr block for ip
1731 * discovery data and G6 memory training data respectively
1733 adev->mman.discovery_tmr_size =
1734 amdgpu_atomfirmware_get_fw_reserved_fb_size(adev);
1735 if (!adev->mman.discovery_tmr_size)
1736 adev->mman.discovery_tmr_size = DISCOVERY_TMR_OFFSET;
1738 if (mem_train_support) {
1739 /* reserve vram for mem train according to TMR location */
1740 amdgpu_ttm_training_data_block_init(adev);
1741 ret = amdgpu_bo_create_kernel_at(adev,
1742 ctx->c2p_train_data_offset,
1743 ctx->train_data_size,
1744 AMDGPU_GEM_DOMAIN_VRAM,
1748 DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret);
1749 amdgpu_ttm_training_reserve_vram_fini(adev);
1752 ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS;
1755 ret = amdgpu_bo_create_kernel_at(adev,
1756 adev->gmc.real_vram_size - adev->mman.discovery_tmr_size,
1757 adev->mman.discovery_tmr_size,
1758 AMDGPU_GEM_DOMAIN_VRAM,
1759 &adev->mman.discovery_memory,
1762 DRM_ERROR("alloc tmr failed(%d)!\n", ret);
1763 amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL);
1771 * amdgpu_ttm_init - Init the memory management (ttm) as well as various
1772 * gtt/vram related fields.
1774 * This initializes all of the memory space pools that the TTM layer
1775 * will need such as the GTT space (system memory mapped to the device),
1776 * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which
1777 * can be mapped per VMID.
1779 int amdgpu_ttm_init(struct amdgpu_device *adev)
1785 mutex_init(&adev->mman.gtt_window_lock);
1787 /* No others user of address space so set it to 0 */
1788 r = ttm_bo_device_init(&adev->mman.bdev, &amdgpu_bo_driver, adev->dev,
1789 adev_to_drm(adev)->anon_inode->i_mapping,
1790 adev_to_drm(adev)->vma_offset_manager,
1792 dma_addressing_limited(adev->dev));
1794 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1797 adev->mman.initialized = true;
1799 /* Initialize VRAM pool with all of VRAM divided into pages */
1800 r = amdgpu_vram_mgr_init(adev);
1802 DRM_ERROR("Failed initializing VRAM heap.\n");
1806 /* Reduce size of CPU-visible VRAM if requested */
1807 vis_vram_limit = (u64)amdgpu_vis_vram_limit * 1024 * 1024;
1808 if (amdgpu_vis_vram_limit > 0 &&
1809 vis_vram_limit <= adev->gmc.visible_vram_size)
1810 adev->gmc.visible_vram_size = vis_vram_limit;
1812 /* Change the size here instead of the init above so only lpfn is affected */
1813 amdgpu_ttm_set_buffer_funcs_status(adev, false);
1815 adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
1816 adev->gmc.visible_vram_size);
1820 *The reserved vram for firmware must be pinned to the specified
1821 *place on the VRAM, so reserve it early.
1823 r = amdgpu_ttm_fw_reserve_vram_init(adev);
1829 * only NAVI10 and onwards ASIC support for IP discovery.
1830 * If IP discovery enabled, a block of memory should be
1831 * reserved for IP discovey.
1833 if (adev->mman.discovery_bin) {
1834 r = amdgpu_ttm_reserve_tmr(adev);
1839 /* allocate memory as required for VGA
1840 * This is used for VGA emulation and pre-OS scanout buffers to
1841 * avoid display artifacts while transitioning between pre-OS
1843 r = amdgpu_bo_create_kernel_at(adev, 0, adev->mman.stolen_vga_size,
1844 AMDGPU_GEM_DOMAIN_VRAM,
1845 &adev->mman.stolen_vga_memory,
1849 r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size,
1850 adev->mman.stolen_extended_size,
1851 AMDGPU_GEM_DOMAIN_VRAM,
1852 &adev->mman.stolen_extended_memory,
1857 DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1858 (unsigned) (adev->gmc.real_vram_size / (1024 * 1024)));
1860 /* Compute GTT size, either bsaed on 3/4th the size of RAM size
1861 * or whatever the user passed on module init */
1862 if (amdgpu_gtt_size == -1) {
1866 gtt_size = min(max((AMDGPU_DEFAULT_GTT_SIZE_MB << 20),
1867 adev->gmc.mc_vram_size),
1868 ((uint64_t)si.totalram * si.mem_unit * 3/4));
1871 gtt_size = (uint64_t)amdgpu_gtt_size << 20;
1873 /* Initialize GTT memory pool */
1874 r = amdgpu_gtt_mgr_init(adev, gtt_size);
1876 DRM_ERROR("Failed initializing GTT heap.\n");
1879 DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1880 (unsigned)(gtt_size / (1024 * 1024)));
1882 /* Initialize various on-chip memory pools */
1883 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GDS, adev->gds.gds_size);
1885 DRM_ERROR("Failed initializing GDS heap.\n");
1889 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GWS, adev->gds.gws_size);
1891 DRM_ERROR("Failed initializing gws heap.\n");
1895 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_OA, adev->gds.oa_size);
1897 DRM_ERROR("Failed initializing oa heap.\n");
1905 * amdgpu_ttm_late_init - Handle any late initialization for amdgpu_ttm
1907 void amdgpu_ttm_late_init(struct amdgpu_device *adev)
1909 /* return the VGA stolen memory (if any) back to VRAM */
1910 if (!adev->mman.keep_stolen_vga_memory)
1911 amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL);
1912 amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL);
1916 * amdgpu_ttm_fini - De-initialize the TTM memory pools
1918 void amdgpu_ttm_fini(struct amdgpu_device *adev)
1920 if (!adev->mman.initialized)
1923 amdgpu_ttm_training_reserve_vram_fini(adev);
1924 /* return the stolen vga memory back to VRAM */
1925 if (adev->mman.keep_stolen_vga_memory)
1926 amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL);
1927 /* return the IP Discovery TMR memory back to VRAM */
1928 amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL);
1929 amdgpu_ttm_fw_reserve_vram_fini(adev);
1931 if (adev->mman.aper_base_kaddr)
1932 iounmap(adev->mman.aper_base_kaddr);
1933 adev->mman.aper_base_kaddr = NULL;
1935 amdgpu_vram_mgr_fini(adev);
1936 amdgpu_gtt_mgr_fini(adev);
1937 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GDS);
1938 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GWS);
1939 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_OA);
1940 ttm_bo_device_release(&adev->mman.bdev);
1941 adev->mman.initialized = false;
1942 DRM_INFO("amdgpu: ttm finalized\n");
1946 * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions
1948 * @adev: amdgpu_device pointer
1949 * @enable: true when we can use buffer functions.
1951 * Enable/disable use of buffer functions during suspend/resume. This should
1952 * only be called at bootup or when userspace isn't running.
1954 void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable)
1956 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
1960 if (!adev->mman.initialized || amdgpu_in_reset(adev) ||
1961 adev->mman.buffer_funcs_enabled == enable)
1965 struct amdgpu_ring *ring;
1966 struct drm_gpu_scheduler *sched;
1968 ring = adev->mman.buffer_funcs_ring;
1969 sched = &ring->sched;
1970 r = drm_sched_entity_init(&adev->mman.entity,
1971 DRM_SCHED_PRIORITY_KERNEL, &sched,
1974 DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
1979 drm_sched_entity_destroy(&adev->mman.entity);
1980 dma_fence_put(man->move);
1984 /* this just adjusts TTM size idea, which sets lpfn to the correct value */
1986 size = adev->gmc.real_vram_size;
1988 size = adev->gmc.visible_vram_size;
1989 man->size = size >> PAGE_SHIFT;
1990 adev->mman.buffer_funcs_enabled = enable;
1993 static vm_fault_t amdgpu_ttm_fault(struct vm_fault *vmf)
1995 struct ttm_buffer_object *bo = vmf->vma->vm_private_data;
1998 ret = ttm_bo_vm_reserve(bo, vmf);
2002 ret = amdgpu_bo_fault_reserve_notify(bo);
2006 ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot,
2007 TTM_BO_VM_NUM_PREFAULT, 1);
2008 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
2012 dma_resv_unlock(bo->base.resv);
2016 static struct vm_operations_struct amdgpu_ttm_vm_ops = {
2017 .fault = amdgpu_ttm_fault,
2018 .open = ttm_bo_vm_open,
2019 .close = ttm_bo_vm_close,
2020 .access = ttm_bo_vm_access
2023 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
2025 struct drm_file *file_priv = filp->private_data;
2026 struct amdgpu_device *adev = drm_to_adev(file_priv->minor->dev);
2029 r = ttm_bo_mmap(filp, vma, &adev->mman.bdev);
2030 if (unlikely(r != 0))
2033 vma->vm_ops = &amdgpu_ttm_vm_ops;
2037 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
2038 uint64_t dst_offset, uint32_t byte_count,
2039 struct dma_resv *resv,
2040 struct dma_fence **fence, bool direct_submit,
2041 bool vm_needs_flush, bool tmz)
2043 enum amdgpu_ib_pool_type pool = direct_submit ? AMDGPU_IB_POOL_DIRECT :
2044 AMDGPU_IB_POOL_DELAYED;
2045 struct amdgpu_device *adev = ring->adev;
2046 struct amdgpu_job *job;
2049 unsigned num_loops, num_dw;
2053 if (direct_submit && !ring->sched.ready) {
2054 DRM_ERROR("Trying to move memory with ring turned off.\n");
2058 max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
2059 num_loops = DIV_ROUND_UP(byte_count, max_bytes);
2060 num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8);
2062 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, pool, &job);
2066 if (vm_needs_flush) {
2067 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
2068 job->vm_needs_flush = true;
2071 r = amdgpu_sync_resv(adev, &job->sync, resv,
2073 AMDGPU_FENCE_OWNER_UNDEFINED);
2075 DRM_ERROR("sync failed (%d).\n", r);
2080 for (i = 0; i < num_loops; i++) {
2081 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
2083 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
2084 dst_offset, cur_size_in_bytes, tmz);
2086 src_offset += cur_size_in_bytes;
2087 dst_offset += cur_size_in_bytes;
2088 byte_count -= cur_size_in_bytes;
2091 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2092 WARN_ON(job->ibs[0].length_dw > num_dw);
2094 r = amdgpu_job_submit_direct(job, ring, fence);
2096 r = amdgpu_job_submit(job, &adev->mman.entity,
2097 AMDGPU_FENCE_OWNER_UNDEFINED, fence);
2104 amdgpu_job_free(job);
2105 DRM_ERROR("Error scheduling IBs (%d)\n", r);
2109 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
2111 struct dma_resv *resv,
2112 struct dma_fence **fence)
2114 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
2115 uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
2116 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2118 struct drm_mm_node *mm_node;
2119 unsigned long num_pages;
2120 unsigned int num_loops, num_dw;
2122 struct amdgpu_job *job;
2125 if (!adev->mman.buffer_funcs_enabled) {
2126 DRM_ERROR("Trying to clear memory with ring turned off.\n");
2130 if (bo->tbo.mem.mem_type == TTM_PL_TT) {
2131 r = amdgpu_ttm_alloc_gart(&bo->tbo);
2136 num_pages = bo->tbo.num_pages;
2137 mm_node = bo->tbo.mem.mm_node;
2140 uint64_t byte_count = mm_node->size << PAGE_SHIFT;
2142 num_loops += DIV_ROUND_UP_ULL(byte_count, max_bytes);
2143 num_pages -= mm_node->size;
2146 num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw;
2148 /* for IB padding */
2151 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, AMDGPU_IB_POOL_DELAYED,
2157 r = amdgpu_sync_resv(adev, &job->sync, resv,
2159 AMDGPU_FENCE_OWNER_UNDEFINED);
2161 DRM_ERROR("sync failed (%d).\n", r);
2166 num_pages = bo->tbo.num_pages;
2167 mm_node = bo->tbo.mem.mm_node;
2170 uint64_t byte_count = mm_node->size << PAGE_SHIFT;
2173 dst_addr = amdgpu_mm_node_addr(&bo->tbo, mm_node, &bo->tbo.mem);
2174 while (byte_count) {
2175 uint32_t cur_size_in_bytes = min_t(uint64_t, byte_count,
2178 amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data,
2179 dst_addr, cur_size_in_bytes);
2181 dst_addr += cur_size_in_bytes;
2182 byte_count -= cur_size_in_bytes;
2185 num_pages -= mm_node->size;
2189 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2190 WARN_ON(job->ibs[0].length_dw > num_dw);
2191 r = amdgpu_job_submit(job, &adev->mman.entity,
2192 AMDGPU_FENCE_OWNER_UNDEFINED, fence);
2199 amdgpu_job_free(job);
2203 #if defined(CONFIG_DEBUG_FS)
2205 static int amdgpu_mm_dump_table(struct seq_file *m, void *data)
2207 struct drm_info_node *node = (struct drm_info_node *)m->private;
2208 unsigned ttm_pl = (uintptr_t)node->info_ent->data;
2209 struct drm_device *dev = node->minor->dev;
2210 struct amdgpu_device *adev = drm_to_adev(dev);
2211 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, ttm_pl);
2212 struct drm_printer p = drm_seq_file_printer(m);
2214 man->func->debug(man, &p);
2218 static int amdgpu_ttm_pool_debugfs(struct seq_file *m, void *data)
2220 struct drm_info_node *node = (struct drm_info_node *)m->private;
2221 struct drm_device *dev = node->minor->dev;
2222 struct amdgpu_device *adev = drm_to_adev(dev);
2224 return ttm_pool_debugfs(&adev->mman.bdev.pool, m);
2227 static const struct drm_info_list amdgpu_ttm_debugfs_list[] = {
2228 {"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, (void *)TTM_PL_VRAM},
2229 {"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, (void *)TTM_PL_TT},
2230 {"amdgpu_gds_mm", amdgpu_mm_dump_table, 0, (void *)AMDGPU_PL_GDS},
2231 {"amdgpu_gws_mm", amdgpu_mm_dump_table, 0, (void *)AMDGPU_PL_GWS},
2232 {"amdgpu_oa_mm", amdgpu_mm_dump_table, 0, (void *)AMDGPU_PL_OA},
2233 {"ttm_page_pool", amdgpu_ttm_pool_debugfs, 0, NULL},
2237 * amdgpu_ttm_vram_read - Linear read access to VRAM
2239 * Accesses VRAM via MMIO for debugging purposes.
2241 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
2242 size_t size, loff_t *pos)
2244 struct amdgpu_device *adev = file_inode(f)->i_private;
2247 if (size & 0x3 || *pos & 0x3)
2250 if (*pos >= adev->gmc.mc_vram_size)
2253 size = min(size, (size_t)(adev->gmc.mc_vram_size - *pos));
2255 size_t bytes = min(size, AMDGPU_TTM_VRAM_MAX_DW_READ * 4);
2256 uint32_t value[AMDGPU_TTM_VRAM_MAX_DW_READ];
2258 amdgpu_device_vram_access(adev, *pos, value, bytes, false);
2259 if (copy_to_user(buf, value, bytes))
2272 * amdgpu_ttm_vram_write - Linear write access to VRAM
2274 * Accesses VRAM via MMIO for debugging purposes.
2276 static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf,
2277 size_t size, loff_t *pos)
2279 struct amdgpu_device *adev = file_inode(f)->i_private;
2283 if (size & 0x3 || *pos & 0x3)
2286 if (*pos >= adev->gmc.mc_vram_size)
2290 unsigned long flags;
2293 if (*pos >= adev->gmc.mc_vram_size)
2296 r = get_user(value, (uint32_t *)buf);
2300 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
2301 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
2302 WREG32_NO_KIQ(mmMM_INDEX_HI, *pos >> 31);
2303 WREG32_NO_KIQ(mmMM_DATA, value);
2304 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
2315 static const struct file_operations amdgpu_ttm_vram_fops = {
2316 .owner = THIS_MODULE,
2317 .read = amdgpu_ttm_vram_read,
2318 .write = amdgpu_ttm_vram_write,
2319 .llseek = default_llseek,
2322 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
2325 * amdgpu_ttm_gtt_read - Linear read access to GTT memory
2327 static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
2328 size_t size, loff_t *pos)
2330 struct amdgpu_device *adev = file_inode(f)->i_private;
2335 loff_t p = *pos / PAGE_SIZE;
2336 unsigned off = *pos & ~PAGE_MASK;
2337 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
2341 if (p >= adev->gart.num_cpu_pages)
2344 page = adev->gart.pages[p];
2349 r = copy_to_user(buf, ptr, cur_size);
2350 kunmap(adev->gart.pages[p]);
2352 r = clear_user(buf, cur_size);
2366 static const struct file_operations amdgpu_ttm_gtt_fops = {
2367 .owner = THIS_MODULE,
2368 .read = amdgpu_ttm_gtt_read,
2369 .llseek = default_llseek
2375 * amdgpu_iomem_read - Virtual read access to GPU mapped memory
2377 * This function is used to read memory that has been mapped to the
2378 * GPU and the known addresses are not physical addresses but instead
2379 * bus addresses (e.g., what you'd put in an IB or ring buffer).
2381 static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
2382 size_t size, loff_t *pos)
2384 struct amdgpu_device *adev = file_inode(f)->i_private;
2385 struct iommu_domain *dom;
2389 /* retrieve the IOMMU domain if any for this device */
2390 dom = iommu_get_domain_for_dev(adev->dev);
2393 phys_addr_t addr = *pos & PAGE_MASK;
2394 loff_t off = *pos & ~PAGE_MASK;
2395 size_t bytes = PAGE_SIZE - off;
2400 bytes = bytes < size ? bytes : size;
2402 /* Translate the bus address to a physical address. If
2403 * the domain is NULL it means there is no IOMMU active
2404 * and the address translation is the identity
2406 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2408 pfn = addr >> PAGE_SHIFT;
2409 if (!pfn_valid(pfn))
2412 p = pfn_to_page(pfn);
2413 if (p->mapping != adev->mman.bdev.dev_mapping)
2417 r = copy_to_user(buf, ptr + off, bytes);
2431 * amdgpu_iomem_write - Virtual write access to GPU mapped memory
2433 * This function is used to write memory that has been mapped to the
2434 * GPU and the known addresses are not physical addresses but instead
2435 * bus addresses (e.g., what you'd put in an IB or ring buffer).
2437 static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
2438 size_t size, loff_t *pos)
2440 struct amdgpu_device *adev = file_inode(f)->i_private;
2441 struct iommu_domain *dom;
2445 dom = iommu_get_domain_for_dev(adev->dev);
2448 phys_addr_t addr = *pos & PAGE_MASK;
2449 loff_t off = *pos & ~PAGE_MASK;
2450 size_t bytes = PAGE_SIZE - off;
2455 bytes = bytes < size ? bytes : size;
2457 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2459 pfn = addr >> PAGE_SHIFT;
2460 if (!pfn_valid(pfn))
2463 p = pfn_to_page(pfn);
2464 if (p->mapping != adev->mman.bdev.dev_mapping)
2468 r = copy_from_user(ptr + off, buf, bytes);
2481 static const struct file_operations amdgpu_ttm_iomem_fops = {
2482 .owner = THIS_MODULE,
2483 .read = amdgpu_iomem_read,
2484 .write = amdgpu_iomem_write,
2485 .llseek = default_llseek
2488 static const struct {
2490 const struct file_operations *fops;
2492 } ttm_debugfs_entries[] = {
2493 { "amdgpu_vram", &amdgpu_ttm_vram_fops, TTM_PL_VRAM },
2494 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
2495 { "amdgpu_gtt", &amdgpu_ttm_gtt_fops, TTM_PL_TT },
2497 { "amdgpu_iomem", &amdgpu_ttm_iomem_fops, TTM_PL_SYSTEM },
2502 int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
2504 #if defined(CONFIG_DEBUG_FS)
2507 struct drm_minor *minor = adev_to_drm(adev)->primary;
2508 struct dentry *ent, *root = minor->debugfs_root;
2510 for (count = 0; count < ARRAY_SIZE(ttm_debugfs_entries); count++) {
2511 ent = debugfs_create_file(
2512 ttm_debugfs_entries[count].name,
2513 S_IFREG | S_IRUGO, root,
2515 ttm_debugfs_entries[count].fops);
2517 return PTR_ERR(ent);
2518 if (ttm_debugfs_entries[count].domain == TTM_PL_VRAM)
2519 i_size_write(ent->d_inode, adev->gmc.mc_vram_size);
2520 else if (ttm_debugfs_entries[count].domain == TTM_PL_TT)
2521 i_size_write(ent->d_inode, adev->gmc.gart_size);
2522 adev->mman.debugfs_entries[count] = ent;
2525 count = ARRAY_SIZE(amdgpu_ttm_debugfs_list);
2526 return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count);