Loading...
1/*
2 * Copyright 2009 Jerome Glisse.
3 * All Rights Reserved.
4 *
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:
12 *
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.
20 *
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
23 * of the Software.
24 *
25 */
26/*
27 * Authors:
28 * Jerome Glisse <glisse@freedesktop.org>
29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
30 * Dave Airlie
31 */
32
33#include <linux/dma-mapping.h>
34#include <linux/iommu.h>
35#include <linux/pagemap.h>
36#include <linux/sched/task.h>
37#include <linux/sched/mm.h>
38#include <linux/seq_file.h>
39#include <linux/slab.h>
40#include <linux/swap.h>
41#include <linux/dma-buf.h>
42#include <linux/sizes.h>
43#include <linux/module.h>
44
45#include <drm/drm_drv.h>
46#include <drm/ttm/ttm_bo.h>
47#include <drm/ttm/ttm_placement.h>
48#include <drm/ttm/ttm_range_manager.h>
49#include <drm/ttm/ttm_tt.h>
50
51#include <drm/amdgpu_drm.h>
52
53#include "amdgpu.h"
54#include "amdgpu_object.h"
55#include "amdgpu_trace.h"
56#include "amdgpu_amdkfd.h"
57#include "amdgpu_sdma.h"
58#include "amdgpu_ras.h"
59#include "amdgpu_hmm.h"
60#include "amdgpu_atomfirmware.h"
61#include "amdgpu_res_cursor.h"
62#include "bif/bif_4_1_d.h"
63
64MODULE_IMPORT_NS(DMA_BUF);
65
66#define AMDGPU_TTM_VRAM_MAX_DW_READ ((size_t)128)
67
68static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
69 struct ttm_tt *ttm,
70 struct ttm_resource *bo_mem);
71static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev,
72 struct ttm_tt *ttm);
73
74static int amdgpu_ttm_init_on_chip(struct amdgpu_device *adev,
75 unsigned int type,
76 uint64_t size_in_page)
77{
78 return ttm_range_man_init(&adev->mman.bdev, type,
79 false, size_in_page);
80}
81
82/**
83 * amdgpu_evict_flags - Compute placement flags
84 *
85 * @bo: The buffer object to evict
86 * @placement: Possible destination(s) for evicted BO
87 *
88 * Fill in placement data when ttm_bo_evict() is called
89 */
90static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
91 struct ttm_placement *placement)
92{
93 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
94 struct amdgpu_bo *abo;
95 static const struct ttm_place placements = {
96 .fpfn = 0,
97 .lpfn = 0,
98 .mem_type = TTM_PL_SYSTEM,
99 .flags = 0
100 };
101
102 /* Don't handle scatter gather BOs */
103 if (bo->type == ttm_bo_type_sg) {
104 placement->num_placement = 0;
105 return;
106 }
107
108 /* Object isn't an AMDGPU object so ignore */
109 if (!amdgpu_bo_is_amdgpu_bo(bo)) {
110 placement->placement = &placements;
111 placement->num_placement = 1;
112 return;
113 }
114
115 abo = ttm_to_amdgpu_bo(bo);
116 if (abo->flags & AMDGPU_GEM_CREATE_DISCARDABLE) {
117 placement->num_placement = 0;
118 return;
119 }
120
121 switch (bo->resource->mem_type) {
122 case AMDGPU_PL_GDS:
123 case AMDGPU_PL_GWS:
124 case AMDGPU_PL_OA:
125 case AMDGPU_PL_DOORBELL:
126 placement->num_placement = 0;
127 return;
128
129 case TTM_PL_VRAM:
130 if (!adev->mman.buffer_funcs_enabled) {
131 /* Move to system memory */
132 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
133
134 } else if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
135 !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) &&
136 amdgpu_res_cpu_visible(adev, bo->resource)) {
137
138 /* Try evicting to the CPU inaccessible part of VRAM
139 * first, but only set GTT as busy placement, so this
140 * BO will be evicted to GTT rather than causing other
141 * BOs to be evicted from VRAM
142 */
143 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
144 AMDGPU_GEM_DOMAIN_GTT |
145 AMDGPU_GEM_DOMAIN_CPU);
146 abo->placements[0].fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
147 abo->placements[0].lpfn = 0;
148 abo->placements[0].flags |= TTM_PL_FLAG_DESIRED;
149 } else {
150 /* Move to GTT memory */
151 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT |
152 AMDGPU_GEM_DOMAIN_CPU);
153 }
154 break;
155 case TTM_PL_TT:
156 case AMDGPU_PL_PREEMPT:
157 default:
158 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
159 break;
160 }
161 *placement = abo->placement;
162}
163
164/**
165 * amdgpu_ttm_map_buffer - Map memory into the GART windows
166 * @bo: buffer object to map
167 * @mem: memory object to map
168 * @mm_cur: range to map
169 * @window: which GART window to use
170 * @ring: DMA ring to use for the copy
171 * @tmz: if we should setup a TMZ enabled mapping
172 * @size: in number of bytes to map, out number of bytes mapped
173 * @addr: resulting address inside the MC address space
174 *
175 * Setup one of the GART windows to access a specific piece of memory or return
176 * the physical address for local memory.
177 */
178static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo,
179 struct ttm_resource *mem,
180 struct amdgpu_res_cursor *mm_cur,
181 unsigned int window, struct amdgpu_ring *ring,
182 bool tmz, uint64_t *size, uint64_t *addr)
183{
184 struct amdgpu_device *adev = ring->adev;
185 unsigned int offset, num_pages, num_dw, num_bytes;
186 uint64_t src_addr, dst_addr;
187 struct amdgpu_job *job;
188 void *cpu_addr;
189 uint64_t flags;
190 unsigned int i;
191 int r;
192
193 BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
194 AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
195
196 if (WARN_ON(mem->mem_type == AMDGPU_PL_PREEMPT))
197 return -EINVAL;
198
199 /* Map only what can't be accessed directly */
200 if (!tmz && mem->start != AMDGPU_BO_INVALID_OFFSET) {
201 *addr = amdgpu_ttm_domain_start(adev, mem->mem_type) +
202 mm_cur->start;
203 return 0;
204 }
205
206
207 /*
208 * If start begins at an offset inside the page, then adjust the size
209 * and addr accordingly
210 */
211 offset = mm_cur->start & ~PAGE_MASK;
212
213 num_pages = PFN_UP(*size + offset);
214 num_pages = min_t(uint32_t, num_pages, AMDGPU_GTT_MAX_TRANSFER_SIZE);
215
216 *size = min(*size, (uint64_t)num_pages * PAGE_SIZE - offset);
217
218 *addr = adev->gmc.gart_start;
219 *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE *
220 AMDGPU_GPU_PAGE_SIZE;
221 *addr += offset;
222
223 num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
224 num_bytes = num_pages * 8 * AMDGPU_GPU_PAGES_IN_CPU_PAGE;
225
226 r = amdgpu_job_alloc_with_ib(adev, &adev->mman.high_pr,
227 AMDGPU_FENCE_OWNER_UNDEFINED,
228 num_dw * 4 + num_bytes,
229 AMDGPU_IB_POOL_DELAYED, &job);
230 if (r)
231 return r;
232
233 src_addr = num_dw * 4;
234 src_addr += job->ibs[0].gpu_addr;
235
236 dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
237 dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
238 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
239 dst_addr, num_bytes, false);
240
241 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
242 WARN_ON(job->ibs[0].length_dw > num_dw);
243
244 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, mem);
245 if (tmz)
246 flags |= AMDGPU_PTE_TMZ;
247
248 cpu_addr = &job->ibs[0].ptr[num_dw];
249
250 if (mem->mem_type == TTM_PL_TT) {
251 dma_addr_t *dma_addr;
252
253 dma_addr = &bo->ttm->dma_address[mm_cur->start >> PAGE_SHIFT];
254 amdgpu_gart_map(adev, 0, num_pages, dma_addr, flags, cpu_addr);
255 } else {
256 dma_addr_t dma_address;
257
258 dma_address = mm_cur->start;
259 dma_address += adev->vm_manager.vram_base_offset;
260
261 for (i = 0; i < num_pages; ++i) {
262 amdgpu_gart_map(adev, i << PAGE_SHIFT, 1, &dma_address,
263 flags, cpu_addr);
264 dma_address += PAGE_SIZE;
265 }
266 }
267
268 dma_fence_put(amdgpu_job_submit(job));
269 return 0;
270}
271
272/**
273 * amdgpu_ttm_copy_mem_to_mem - Helper function for copy
274 * @adev: amdgpu device
275 * @src: buffer/address where to read from
276 * @dst: buffer/address where to write to
277 * @size: number of bytes to copy
278 * @tmz: if a secure copy should be used
279 * @resv: resv object to sync to
280 * @f: Returns the last fence if multiple jobs are submitted.
281 *
282 * The function copies @size bytes from {src->mem + src->offset} to
283 * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a
284 * move and different for a BO to BO copy.
285 *
286 */
287int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev,
288 const struct amdgpu_copy_mem *src,
289 const struct amdgpu_copy_mem *dst,
290 uint64_t size, bool tmz,
291 struct dma_resv *resv,
292 struct dma_fence **f)
293{
294 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
295 struct amdgpu_res_cursor src_mm, dst_mm;
296 struct dma_fence *fence = NULL;
297 int r = 0;
298
299 if (!adev->mman.buffer_funcs_enabled) {
300 DRM_ERROR("Trying to move memory with ring turned off.\n");
301 return -EINVAL;
302 }
303
304 amdgpu_res_first(src->mem, src->offset, size, &src_mm);
305 amdgpu_res_first(dst->mem, dst->offset, size, &dst_mm);
306
307 mutex_lock(&adev->mman.gtt_window_lock);
308 while (src_mm.remaining) {
309 uint64_t from, to, cur_size;
310 struct dma_fence *next;
311
312 /* Never copy more than 256MiB at once to avoid a timeout */
313 cur_size = min3(src_mm.size, dst_mm.size, 256ULL << 20);
314
315 /* Map src to window 0 and dst to window 1. */
316 r = amdgpu_ttm_map_buffer(src->bo, src->mem, &src_mm,
317 0, ring, tmz, &cur_size, &from);
318 if (r)
319 goto error;
320
321 r = amdgpu_ttm_map_buffer(dst->bo, dst->mem, &dst_mm,
322 1, ring, tmz, &cur_size, &to);
323 if (r)
324 goto error;
325
326 r = amdgpu_copy_buffer(ring, from, to, cur_size,
327 resv, &next, false, true, tmz);
328 if (r)
329 goto error;
330
331 dma_fence_put(fence);
332 fence = next;
333
334 amdgpu_res_next(&src_mm, cur_size);
335 amdgpu_res_next(&dst_mm, cur_size);
336 }
337error:
338 mutex_unlock(&adev->mman.gtt_window_lock);
339 if (f)
340 *f = dma_fence_get(fence);
341 dma_fence_put(fence);
342 return r;
343}
344
345/*
346 * amdgpu_move_blit - Copy an entire buffer to another buffer
347 *
348 * This is a helper called by amdgpu_bo_move() and amdgpu_move_vram_ram() to
349 * help move buffers to and from VRAM.
350 */
351static int amdgpu_move_blit(struct ttm_buffer_object *bo,
352 bool evict,
353 struct ttm_resource *new_mem,
354 struct ttm_resource *old_mem)
355{
356 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
357 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
358 struct amdgpu_copy_mem src, dst;
359 struct dma_fence *fence = NULL;
360 int r;
361
362 src.bo = bo;
363 dst.bo = bo;
364 src.mem = old_mem;
365 dst.mem = new_mem;
366 src.offset = 0;
367 dst.offset = 0;
368
369 r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
370 new_mem->size,
371 amdgpu_bo_encrypted(abo),
372 bo->base.resv, &fence);
373 if (r)
374 goto error;
375
376 /* clear the space being freed */
377 if (old_mem->mem_type == TTM_PL_VRAM &&
378 (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) {
379 struct dma_fence *wipe_fence = NULL;
380
381 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, NULL, &wipe_fence,
382 false);
383 if (r) {
384 goto error;
385 } else if (wipe_fence) {
386 dma_fence_put(fence);
387 fence = wipe_fence;
388 }
389 }
390
391 /* Always block for VM page tables before committing the new location */
392 if (bo->type == ttm_bo_type_kernel)
393 r = ttm_bo_move_accel_cleanup(bo, fence, true, false, new_mem);
394 else
395 r = ttm_bo_move_accel_cleanup(bo, fence, evict, true, new_mem);
396 dma_fence_put(fence);
397 return r;
398
399error:
400 if (fence)
401 dma_fence_wait(fence, false);
402 dma_fence_put(fence);
403 return r;
404}
405
406/**
407 * amdgpu_res_cpu_visible - Check that resource can be accessed by CPU
408 * @adev: amdgpu device
409 * @res: the resource to check
410 *
411 * Returns: true if the full resource is CPU visible, false otherwise.
412 */
413bool amdgpu_res_cpu_visible(struct amdgpu_device *adev,
414 struct ttm_resource *res)
415{
416 struct amdgpu_res_cursor cursor;
417
418 if (!res)
419 return false;
420
421 if (res->mem_type == TTM_PL_SYSTEM || res->mem_type == TTM_PL_TT ||
422 res->mem_type == AMDGPU_PL_PREEMPT || res->mem_type == AMDGPU_PL_DOORBELL)
423 return true;
424
425 if (res->mem_type != TTM_PL_VRAM)
426 return false;
427
428 amdgpu_res_first(res, 0, res->size, &cursor);
429 while (cursor.remaining) {
430 if ((cursor.start + cursor.size) > adev->gmc.visible_vram_size)
431 return false;
432 amdgpu_res_next(&cursor, cursor.size);
433 }
434
435 return true;
436}
437
438/*
439 * amdgpu_res_copyable - Check that memory can be accessed by ttm_bo_move_memcpy
440 *
441 * Called by amdgpu_bo_move()
442 */
443static bool amdgpu_res_copyable(struct amdgpu_device *adev,
444 struct ttm_resource *mem)
445{
446 if (!amdgpu_res_cpu_visible(adev, mem))
447 return false;
448
449 /* ttm_resource_ioremap only supports contiguous memory */
450 if (mem->mem_type == TTM_PL_VRAM &&
451 !(mem->placement & TTM_PL_FLAG_CONTIGUOUS))
452 return false;
453
454 return true;
455}
456
457/*
458 * amdgpu_bo_move - Move a buffer object to a new memory location
459 *
460 * Called by ttm_bo_handle_move_mem()
461 */
462static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
463 struct ttm_operation_ctx *ctx,
464 struct ttm_resource *new_mem,
465 struct ttm_place *hop)
466{
467 struct amdgpu_device *adev;
468 struct amdgpu_bo *abo;
469 struct ttm_resource *old_mem = bo->resource;
470 int r;
471
472 if (new_mem->mem_type == TTM_PL_TT ||
473 new_mem->mem_type == AMDGPU_PL_PREEMPT) {
474 r = amdgpu_ttm_backend_bind(bo->bdev, bo->ttm, new_mem);
475 if (r)
476 return r;
477 }
478
479 abo = ttm_to_amdgpu_bo(bo);
480 adev = amdgpu_ttm_adev(bo->bdev);
481
482 if (!old_mem || (old_mem->mem_type == TTM_PL_SYSTEM &&
483 bo->ttm == NULL)) {
484 amdgpu_bo_move_notify(bo, evict, new_mem);
485 ttm_bo_move_null(bo, new_mem);
486 return 0;
487 }
488 if (old_mem->mem_type == TTM_PL_SYSTEM &&
489 (new_mem->mem_type == TTM_PL_TT ||
490 new_mem->mem_type == AMDGPU_PL_PREEMPT)) {
491 amdgpu_bo_move_notify(bo, evict, new_mem);
492 ttm_bo_move_null(bo, new_mem);
493 return 0;
494 }
495 if ((old_mem->mem_type == TTM_PL_TT ||
496 old_mem->mem_type == AMDGPU_PL_PREEMPT) &&
497 new_mem->mem_type == TTM_PL_SYSTEM) {
498 r = ttm_bo_wait_ctx(bo, ctx);
499 if (r)
500 return r;
501
502 amdgpu_ttm_backend_unbind(bo->bdev, bo->ttm);
503 amdgpu_bo_move_notify(bo, evict, new_mem);
504 ttm_resource_free(bo, &bo->resource);
505 ttm_bo_assign_mem(bo, new_mem);
506 return 0;
507 }
508
509 if (old_mem->mem_type == AMDGPU_PL_GDS ||
510 old_mem->mem_type == AMDGPU_PL_GWS ||
511 old_mem->mem_type == AMDGPU_PL_OA ||
512 old_mem->mem_type == AMDGPU_PL_DOORBELL ||
513 new_mem->mem_type == AMDGPU_PL_GDS ||
514 new_mem->mem_type == AMDGPU_PL_GWS ||
515 new_mem->mem_type == AMDGPU_PL_OA ||
516 new_mem->mem_type == AMDGPU_PL_DOORBELL) {
517 /* Nothing to save here */
518 amdgpu_bo_move_notify(bo, evict, new_mem);
519 ttm_bo_move_null(bo, new_mem);
520 return 0;
521 }
522
523 if (bo->type == ttm_bo_type_device &&
524 new_mem->mem_type == TTM_PL_VRAM &&
525 old_mem->mem_type != TTM_PL_VRAM) {
526 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
527 * accesses the BO after it's moved.
528 */
529 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
530 }
531
532 if (adev->mman.buffer_funcs_enabled &&
533 ((old_mem->mem_type == TTM_PL_SYSTEM &&
534 new_mem->mem_type == TTM_PL_VRAM) ||
535 (old_mem->mem_type == TTM_PL_VRAM &&
536 new_mem->mem_type == TTM_PL_SYSTEM))) {
537 hop->fpfn = 0;
538 hop->lpfn = 0;
539 hop->mem_type = TTM_PL_TT;
540 hop->flags = TTM_PL_FLAG_TEMPORARY;
541 return -EMULTIHOP;
542 }
543
544 amdgpu_bo_move_notify(bo, evict, new_mem);
545 if (adev->mman.buffer_funcs_enabled)
546 r = amdgpu_move_blit(bo, evict, new_mem, old_mem);
547 else
548 r = -ENODEV;
549
550 if (r) {
551 /* Check that all memory is CPU accessible */
552 if (!amdgpu_res_copyable(adev, old_mem) ||
553 !amdgpu_res_copyable(adev, new_mem)) {
554 pr_err("Move buffer fallback to memcpy unavailable\n");
555 return r;
556 }
557
558 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
559 if (r)
560 return r;
561 }
562
563 /* update statistics after the move */
564 if (evict)
565 atomic64_inc(&adev->num_evictions);
566 atomic64_add(bo->base.size, &adev->num_bytes_moved);
567 return 0;
568}
569
570/*
571 * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault
572 *
573 * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault()
574 */
575static int amdgpu_ttm_io_mem_reserve(struct ttm_device *bdev,
576 struct ttm_resource *mem)
577{
578 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
579
580 switch (mem->mem_type) {
581 case TTM_PL_SYSTEM:
582 /* system memory */
583 return 0;
584 case TTM_PL_TT:
585 case AMDGPU_PL_PREEMPT:
586 break;
587 case TTM_PL_VRAM:
588 mem->bus.offset = mem->start << PAGE_SHIFT;
589
590 if (adev->mman.aper_base_kaddr &&
591 mem->placement & TTM_PL_FLAG_CONTIGUOUS)
592 mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
593 mem->bus.offset;
594
595 mem->bus.offset += adev->gmc.aper_base;
596 mem->bus.is_iomem = true;
597 break;
598 case AMDGPU_PL_DOORBELL:
599 mem->bus.offset = mem->start << PAGE_SHIFT;
600 mem->bus.offset += adev->doorbell.base;
601 mem->bus.is_iomem = true;
602 mem->bus.caching = ttm_uncached;
603 break;
604 default:
605 return -EINVAL;
606 }
607 return 0;
608}
609
610static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
611 unsigned long page_offset)
612{
613 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
614 struct amdgpu_res_cursor cursor;
615
616 amdgpu_res_first(bo->resource, (u64)page_offset << PAGE_SHIFT, 0,
617 &cursor);
618
619 if (bo->resource->mem_type == AMDGPU_PL_DOORBELL)
620 return ((uint64_t)(adev->doorbell.base + cursor.start)) >> PAGE_SHIFT;
621
622 return (adev->gmc.aper_base + cursor.start) >> PAGE_SHIFT;
623}
624
625/**
626 * amdgpu_ttm_domain_start - Returns GPU start address
627 * @adev: amdgpu device object
628 * @type: type of the memory
629 *
630 * Returns:
631 * GPU start address of a memory domain
632 */
633
634uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type)
635{
636 switch (type) {
637 case TTM_PL_TT:
638 return adev->gmc.gart_start;
639 case TTM_PL_VRAM:
640 return adev->gmc.vram_start;
641 }
642
643 return 0;
644}
645
646/*
647 * TTM backend functions.
648 */
649struct amdgpu_ttm_tt {
650 struct ttm_tt ttm;
651 struct drm_gem_object *gobj;
652 u64 offset;
653 uint64_t userptr;
654 struct task_struct *usertask;
655 uint32_t userflags;
656 bool bound;
657 int32_t pool_id;
658};
659
660#define ttm_to_amdgpu_ttm_tt(ptr) container_of(ptr, struct amdgpu_ttm_tt, ttm)
661
662#ifdef CONFIG_DRM_AMDGPU_USERPTR
663/*
664 * amdgpu_ttm_tt_get_user_pages - get device accessible pages that back user
665 * memory and start HMM tracking CPU page table update
666 *
667 * Calling function must call amdgpu_ttm_tt_userptr_range_done() once and only
668 * once afterwards to stop HMM tracking
669 */
670int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages,
671 struct hmm_range **range)
672{
673 struct ttm_tt *ttm = bo->tbo.ttm;
674 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
675 unsigned long start = gtt->userptr;
676 struct vm_area_struct *vma;
677 struct mm_struct *mm;
678 bool readonly;
679 int r = 0;
680
681 /* Make sure get_user_pages_done() can cleanup gracefully */
682 *range = NULL;
683
684 mm = bo->notifier.mm;
685 if (unlikely(!mm)) {
686 DRM_DEBUG_DRIVER("BO is not registered?\n");
687 return -EFAULT;
688 }
689
690 if (!mmget_not_zero(mm)) /* Happens during process shutdown */
691 return -ESRCH;
692
693 mmap_read_lock(mm);
694 vma = vma_lookup(mm, start);
695 if (unlikely(!vma)) {
696 r = -EFAULT;
697 goto out_unlock;
698 }
699 if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) &&
700 vma->vm_file)) {
701 r = -EPERM;
702 goto out_unlock;
703 }
704
705 readonly = amdgpu_ttm_tt_is_readonly(ttm);
706 r = amdgpu_hmm_range_get_pages(&bo->notifier, start, ttm->num_pages,
707 readonly, NULL, pages, range);
708out_unlock:
709 mmap_read_unlock(mm);
710 if (r)
711 pr_debug("failed %d to get user pages 0x%lx\n", r, start);
712
713 mmput(mm);
714
715 return r;
716}
717
718/* amdgpu_ttm_tt_discard_user_pages - Discard range and pfn array allocations
719 */
720void amdgpu_ttm_tt_discard_user_pages(struct ttm_tt *ttm,
721 struct hmm_range *range)
722{
723 struct amdgpu_ttm_tt *gtt = (void *)ttm;
724
725 if (gtt && gtt->userptr && range)
726 amdgpu_hmm_range_get_pages_done(range);
727}
728
729/*
730 * amdgpu_ttm_tt_get_user_pages_done - stop HMM track the CPU page table change
731 * Check if the pages backing this ttm range have been invalidated
732 *
733 * Returns: true if pages are still valid
734 */
735bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm,
736 struct hmm_range *range)
737{
738 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
739
740 if (!gtt || !gtt->userptr || !range)
741 return false;
742
743 DRM_DEBUG_DRIVER("user_pages_done 0x%llx pages 0x%x\n",
744 gtt->userptr, ttm->num_pages);
745
746 WARN_ONCE(!range->hmm_pfns, "No user pages to check\n");
747
748 return !amdgpu_hmm_range_get_pages_done(range);
749}
750#endif
751
752/*
753 * amdgpu_ttm_tt_set_user_pages - Copy pages in, putting old pages as necessary.
754 *
755 * Called by amdgpu_cs_list_validate(). This creates the page list
756 * that backs user memory and will ultimately be mapped into the device
757 * address space.
758 */
759void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages)
760{
761 unsigned long i;
762
763 for (i = 0; i < ttm->num_pages; ++i)
764 ttm->pages[i] = pages ? pages[i] : NULL;
765}
766
767/*
768 * amdgpu_ttm_tt_pin_userptr - prepare the sg table with the user pages
769 *
770 * Called by amdgpu_ttm_backend_bind()
771 **/
772static int amdgpu_ttm_tt_pin_userptr(struct ttm_device *bdev,
773 struct ttm_tt *ttm)
774{
775 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
776 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
777 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
778 enum dma_data_direction direction = write ?
779 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
780 int r;
781
782 /* Allocate an SG array and squash pages into it */
783 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
784 (u64)ttm->num_pages << PAGE_SHIFT,
785 GFP_KERNEL);
786 if (r)
787 goto release_sg;
788
789 /* Map SG to device */
790 r = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
791 if (r)
792 goto release_sg;
793
794 /* convert SG to linear array of pages and dma addresses */
795 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
796 ttm->num_pages);
797
798 return 0;
799
800release_sg:
801 kfree(ttm->sg);
802 ttm->sg = NULL;
803 return r;
804}
805
806/*
807 * amdgpu_ttm_tt_unpin_userptr - Unpin and unmap userptr pages
808 */
809static void amdgpu_ttm_tt_unpin_userptr(struct ttm_device *bdev,
810 struct ttm_tt *ttm)
811{
812 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
813 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
814 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
815 enum dma_data_direction direction = write ?
816 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
817
818 /* double check that we don't free the table twice */
819 if (!ttm->sg || !ttm->sg->sgl)
820 return;
821
822 /* unmap the pages mapped to the device */
823 dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
824 sg_free_table(ttm->sg);
825}
826
827/*
828 * total_pages is constructed as MQD0+CtrlStack0 + MQD1+CtrlStack1 + ...
829 * MQDn+CtrlStackn where n is the number of XCCs per partition.
830 * pages_per_xcc is the size of one MQD+CtrlStack. The first page is MQD
831 * and uses memory type default, UC. The rest of pages_per_xcc are
832 * Ctrl stack and modify their memory type to NC.
833 */
834static void amdgpu_ttm_gart_bind_gfx9_mqd(struct amdgpu_device *adev,
835 struct ttm_tt *ttm, uint64_t flags)
836{
837 struct amdgpu_ttm_tt *gtt = (void *)ttm;
838 uint64_t total_pages = ttm->num_pages;
839 int num_xcc = max(1U, adev->gfx.num_xcc_per_xcp);
840 uint64_t page_idx, pages_per_xcc;
841 int i;
842 uint64_t ctrl_flags = (flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) |
843 AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC);
844
845 pages_per_xcc = total_pages;
846 do_div(pages_per_xcc, num_xcc);
847
848 for (i = 0, page_idx = 0; i < num_xcc; i++, page_idx += pages_per_xcc) {
849 /* MQD page: use default flags */
850 amdgpu_gart_bind(adev,
851 gtt->offset + (page_idx << PAGE_SHIFT),
852 1, >t->ttm.dma_address[page_idx], flags);
853 /*
854 * Ctrl pages - modify the memory type to NC (ctrl_flags) from
855 * the second page of the BO onward.
856 */
857 amdgpu_gart_bind(adev,
858 gtt->offset + ((page_idx + 1) << PAGE_SHIFT),
859 pages_per_xcc - 1,
860 >t->ttm.dma_address[page_idx + 1],
861 ctrl_flags);
862 }
863}
864
865static void amdgpu_ttm_gart_bind(struct amdgpu_device *adev,
866 struct ttm_buffer_object *tbo,
867 uint64_t flags)
868{
869 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(tbo);
870 struct ttm_tt *ttm = tbo->ttm;
871 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
872
873 if (amdgpu_bo_encrypted(abo))
874 flags |= AMDGPU_PTE_TMZ;
875
876 if (abo->flags & AMDGPU_GEM_CREATE_CP_MQD_GFX9) {
877 amdgpu_ttm_gart_bind_gfx9_mqd(adev, ttm, flags);
878 } else {
879 amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
880 gtt->ttm.dma_address, flags);
881 }
882 gtt->bound = true;
883}
884
885/*
886 * amdgpu_ttm_backend_bind - Bind GTT memory
887 *
888 * Called by ttm_tt_bind() on behalf of ttm_bo_handle_move_mem().
889 * This handles binding GTT memory to the device address space.
890 */
891static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
892 struct ttm_tt *ttm,
893 struct ttm_resource *bo_mem)
894{
895 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
896 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
897 uint64_t flags;
898 int r;
899
900 if (!bo_mem)
901 return -EINVAL;
902
903 if (gtt->bound)
904 return 0;
905
906 if (gtt->userptr) {
907 r = amdgpu_ttm_tt_pin_userptr(bdev, ttm);
908 if (r) {
909 DRM_ERROR("failed to pin userptr\n");
910 return r;
911 }
912 } else if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL) {
913 if (!ttm->sg) {
914 struct dma_buf_attachment *attach;
915 struct sg_table *sgt;
916
917 attach = gtt->gobj->import_attach;
918 sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
919 if (IS_ERR(sgt))
920 return PTR_ERR(sgt);
921
922 ttm->sg = sgt;
923 }
924
925 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
926 ttm->num_pages);
927 }
928
929 if (!ttm->num_pages) {
930 WARN(1, "nothing to bind %u pages for mreg %p back %p!\n",
931 ttm->num_pages, bo_mem, ttm);
932 }
933
934 if (bo_mem->mem_type != TTM_PL_TT ||
935 !amdgpu_gtt_mgr_has_gart_addr(bo_mem)) {
936 gtt->offset = AMDGPU_BO_INVALID_OFFSET;
937 return 0;
938 }
939
940 /* compute PTE flags relevant to this BO memory */
941 flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
942
943 /* bind pages into GART page tables */
944 gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
945 amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
946 gtt->ttm.dma_address, flags);
947 gtt->bound = true;
948 return 0;
949}
950
951/*
952 * amdgpu_ttm_alloc_gart - Make sure buffer object is accessible either
953 * through AGP or GART aperture.
954 *
955 * If bo is accessible through AGP aperture, then use AGP aperture
956 * to access bo; otherwise allocate logical space in GART aperture
957 * and map bo to GART aperture.
958 */
959int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
960{
961 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
962 struct ttm_operation_ctx ctx = { false, false };
963 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(bo->ttm);
964 struct ttm_placement placement;
965 struct ttm_place placements;
966 struct ttm_resource *tmp;
967 uint64_t addr, flags;
968 int r;
969
970 if (bo->resource->start != AMDGPU_BO_INVALID_OFFSET)
971 return 0;
972
973 addr = amdgpu_gmc_agp_addr(bo);
974 if (addr != AMDGPU_BO_INVALID_OFFSET)
975 return 0;
976
977 /* allocate GART space */
978 placement.num_placement = 1;
979 placement.placement = &placements;
980 placements.fpfn = 0;
981 placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
982 placements.mem_type = TTM_PL_TT;
983 placements.flags = bo->resource->placement;
984
985 r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx);
986 if (unlikely(r))
987 return r;
988
989 /* compute PTE flags for this buffer object */
990 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, tmp);
991
992 /* Bind pages */
993 gtt->offset = (u64)tmp->start << PAGE_SHIFT;
994 amdgpu_ttm_gart_bind(adev, bo, flags);
995 amdgpu_gart_invalidate_tlb(adev);
996 ttm_resource_free(bo, &bo->resource);
997 ttm_bo_assign_mem(bo, tmp);
998
999 return 0;
1000}
1001
1002/*
1003 * amdgpu_ttm_recover_gart - Rebind GTT pages
1004 *
1005 * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to
1006 * rebind GTT pages during a GPU reset.
1007 */
1008void amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
1009{
1010 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
1011 uint64_t flags;
1012
1013 if (!tbo->ttm)
1014 return;
1015
1016 flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, tbo->resource);
1017 amdgpu_ttm_gart_bind(adev, tbo, flags);
1018}
1019
1020/*
1021 * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages
1022 *
1023 * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and
1024 * ttm_tt_destroy().
1025 */
1026static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev,
1027 struct ttm_tt *ttm)
1028{
1029 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1030 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1031
1032 /* if the pages have userptr pinning then clear that first */
1033 if (gtt->userptr) {
1034 amdgpu_ttm_tt_unpin_userptr(bdev, ttm);
1035 } else if (ttm->sg && gtt->gobj->import_attach) {
1036 struct dma_buf_attachment *attach;
1037
1038 attach = gtt->gobj->import_attach;
1039 dma_buf_unmap_attachment(attach, ttm->sg, DMA_BIDIRECTIONAL);
1040 ttm->sg = NULL;
1041 }
1042
1043 if (!gtt->bound)
1044 return;
1045
1046 if (gtt->offset == AMDGPU_BO_INVALID_OFFSET)
1047 return;
1048
1049 /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
1050 amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages);
1051 gtt->bound = false;
1052}
1053
1054static void amdgpu_ttm_backend_destroy(struct ttm_device *bdev,
1055 struct ttm_tt *ttm)
1056{
1057 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1058
1059 if (gtt->usertask)
1060 put_task_struct(gtt->usertask);
1061
1062 ttm_tt_fini(>t->ttm);
1063 kfree(gtt);
1064}
1065
1066/**
1067 * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO
1068 *
1069 * @bo: The buffer object to create a GTT ttm_tt object around
1070 * @page_flags: Page flags to be added to the ttm_tt object
1071 *
1072 * Called by ttm_tt_create().
1073 */
1074static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
1075 uint32_t page_flags)
1076{
1077 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1078 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1079 struct amdgpu_ttm_tt *gtt;
1080 enum ttm_caching caching;
1081
1082 gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
1083 if (!gtt)
1084 return NULL;
1085
1086 gtt->gobj = &bo->base;
1087 if (adev->gmc.mem_partitions && abo->xcp_id >= 0)
1088 gtt->pool_id = KFD_XCP_MEM_ID(adev, abo->xcp_id);
1089 else
1090 gtt->pool_id = abo->xcp_id;
1091
1092 if (abo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
1093 caching = ttm_write_combined;
1094 else
1095 caching = ttm_cached;
1096
1097 /* allocate space for the uninitialized page entries */
1098 if (ttm_sg_tt_init(>t->ttm, bo, page_flags, caching)) {
1099 kfree(gtt);
1100 return NULL;
1101 }
1102 return >t->ttm;
1103}
1104
1105/*
1106 * amdgpu_ttm_tt_populate - Map GTT pages visible to the device
1107 *
1108 * Map the pages of a ttm_tt object to an address space visible
1109 * to the underlying device.
1110 */
1111static int amdgpu_ttm_tt_populate(struct ttm_device *bdev,
1112 struct ttm_tt *ttm,
1113 struct ttm_operation_ctx *ctx)
1114{
1115 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1116 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1117 struct ttm_pool *pool;
1118 pgoff_t i;
1119 int ret;
1120
1121 /* user pages are bound by amdgpu_ttm_tt_pin_userptr() */
1122 if (gtt->userptr) {
1123 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
1124 if (!ttm->sg)
1125 return -ENOMEM;
1126 return 0;
1127 }
1128
1129 if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL)
1130 return 0;
1131
1132 if (adev->mman.ttm_pools && gtt->pool_id >= 0)
1133 pool = &adev->mman.ttm_pools[gtt->pool_id];
1134 else
1135 pool = &adev->mman.bdev.pool;
1136 ret = ttm_pool_alloc(pool, ttm, ctx);
1137 if (ret)
1138 return ret;
1139
1140 for (i = 0; i < ttm->num_pages; ++i)
1141 ttm->pages[i]->mapping = bdev->dev_mapping;
1142
1143 return 0;
1144}
1145
1146/*
1147 * amdgpu_ttm_tt_unpopulate - unmap GTT pages and unpopulate page arrays
1148 *
1149 * Unmaps pages of a ttm_tt object from the device address space and
1150 * unpopulates the page array backing it.
1151 */
1152static void amdgpu_ttm_tt_unpopulate(struct ttm_device *bdev,
1153 struct ttm_tt *ttm)
1154{
1155 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1156 struct amdgpu_device *adev;
1157 struct ttm_pool *pool;
1158 pgoff_t i;
1159
1160 amdgpu_ttm_backend_unbind(bdev, ttm);
1161
1162 if (gtt->userptr) {
1163 amdgpu_ttm_tt_set_user_pages(ttm, NULL);
1164 kfree(ttm->sg);
1165 ttm->sg = NULL;
1166 return;
1167 }
1168
1169 if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL)
1170 return;
1171
1172 for (i = 0; i < ttm->num_pages; ++i)
1173 ttm->pages[i]->mapping = NULL;
1174
1175 adev = amdgpu_ttm_adev(bdev);
1176
1177 if (adev->mman.ttm_pools && gtt->pool_id >= 0)
1178 pool = &adev->mman.ttm_pools[gtt->pool_id];
1179 else
1180 pool = &adev->mman.bdev.pool;
1181
1182 return ttm_pool_free(pool, ttm);
1183}
1184
1185/**
1186 * amdgpu_ttm_tt_get_userptr - Return the userptr GTT ttm_tt for the current
1187 * task
1188 *
1189 * @tbo: The ttm_buffer_object that contains the userptr
1190 * @user_addr: The returned value
1191 */
1192int amdgpu_ttm_tt_get_userptr(const struct ttm_buffer_object *tbo,
1193 uint64_t *user_addr)
1194{
1195 struct amdgpu_ttm_tt *gtt;
1196
1197 if (!tbo->ttm)
1198 return -EINVAL;
1199
1200 gtt = (void *)tbo->ttm;
1201 *user_addr = gtt->userptr;
1202 return 0;
1203}
1204
1205/**
1206 * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt for the current
1207 * task
1208 *
1209 * @bo: The ttm_buffer_object to bind this userptr to
1210 * @addr: The address in the current tasks VM space to use
1211 * @flags: Requirements of userptr object.
1212 *
1213 * Called by amdgpu_gem_userptr_ioctl() and kfd_ioctl_alloc_memory_of_gpu() to
1214 * bind userptr pages to current task and by kfd_ioctl_acquire_vm() to
1215 * initialize GPU VM for a KFD process.
1216 */
1217int amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object *bo,
1218 uint64_t addr, uint32_t flags)
1219{
1220 struct amdgpu_ttm_tt *gtt;
1221
1222 if (!bo->ttm) {
1223 /* TODO: We want a separate TTM object type for userptrs */
1224 bo->ttm = amdgpu_ttm_tt_create(bo, 0);
1225 if (bo->ttm == NULL)
1226 return -ENOMEM;
1227 }
1228
1229 /* Set TTM_TT_FLAG_EXTERNAL before populate but after create. */
1230 bo->ttm->page_flags |= TTM_TT_FLAG_EXTERNAL;
1231
1232 gtt = ttm_to_amdgpu_ttm_tt(bo->ttm);
1233 gtt->userptr = addr;
1234 gtt->userflags = flags;
1235
1236 if (gtt->usertask)
1237 put_task_struct(gtt->usertask);
1238 gtt->usertask = current->group_leader;
1239 get_task_struct(gtt->usertask);
1240
1241 return 0;
1242}
1243
1244/*
1245 * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object
1246 */
1247struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1248{
1249 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1250
1251 if (gtt == NULL)
1252 return NULL;
1253
1254 if (gtt->usertask == NULL)
1255 return NULL;
1256
1257 return gtt->usertask->mm;
1258}
1259
1260/*
1261 * amdgpu_ttm_tt_affect_userptr - Determine if a ttm_tt object lays inside an
1262 * address range for the current task.
1263 *
1264 */
1265bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1266 unsigned long end, unsigned long *userptr)
1267{
1268 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1269 unsigned long size;
1270
1271 if (gtt == NULL || !gtt->userptr)
1272 return false;
1273
1274 /* Return false if no part of the ttm_tt object lies within
1275 * the range
1276 */
1277 size = (unsigned long)gtt->ttm.num_pages * PAGE_SIZE;
1278 if (gtt->userptr > end || gtt->userptr + size <= start)
1279 return false;
1280
1281 if (userptr)
1282 *userptr = gtt->userptr;
1283 return true;
1284}
1285
1286/*
1287 * amdgpu_ttm_tt_is_userptr - Have the pages backing by userptr?
1288 */
1289bool amdgpu_ttm_tt_is_userptr(struct ttm_tt *ttm)
1290{
1291 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1292
1293 if (gtt == NULL || !gtt->userptr)
1294 return false;
1295
1296 return true;
1297}
1298
1299/*
1300 * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only?
1301 */
1302bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1303{
1304 struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1305
1306 if (gtt == NULL)
1307 return false;
1308
1309 return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1310}
1311
1312/**
1313 * amdgpu_ttm_tt_pde_flags - Compute PDE flags for ttm_tt object
1314 *
1315 * @ttm: The ttm_tt object to compute the flags for
1316 * @mem: The memory registry backing this ttm_tt object
1317 *
1318 * Figure out the flags to use for a VM PDE (Page Directory Entry).
1319 */
1320uint64_t amdgpu_ttm_tt_pde_flags(struct ttm_tt *ttm, struct ttm_resource *mem)
1321{
1322 uint64_t flags = 0;
1323
1324 if (mem && mem->mem_type != TTM_PL_SYSTEM)
1325 flags |= AMDGPU_PTE_VALID;
1326
1327 if (mem && (mem->mem_type == TTM_PL_TT ||
1328 mem->mem_type == AMDGPU_PL_DOORBELL ||
1329 mem->mem_type == AMDGPU_PL_PREEMPT)) {
1330 flags |= AMDGPU_PTE_SYSTEM;
1331
1332 if (ttm->caching == ttm_cached)
1333 flags |= AMDGPU_PTE_SNOOPED;
1334 }
1335
1336 if (mem && mem->mem_type == TTM_PL_VRAM &&
1337 mem->bus.caching == ttm_cached)
1338 flags |= AMDGPU_PTE_SNOOPED;
1339
1340 return flags;
1341}
1342
1343/**
1344 * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object
1345 *
1346 * @adev: amdgpu_device pointer
1347 * @ttm: The ttm_tt object to compute the flags for
1348 * @mem: The memory registry backing this ttm_tt object
1349 *
1350 * Figure out the flags to use for a VM PTE (Page Table Entry).
1351 */
1352uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1353 struct ttm_resource *mem)
1354{
1355 uint64_t flags = amdgpu_ttm_tt_pde_flags(ttm, mem);
1356
1357 flags |= adev->gart.gart_pte_flags;
1358 flags |= AMDGPU_PTE_READABLE;
1359
1360 if (!amdgpu_ttm_tt_is_readonly(ttm))
1361 flags |= AMDGPU_PTE_WRITEABLE;
1362
1363 return flags;
1364}
1365
1366/*
1367 * amdgpu_ttm_bo_eviction_valuable - Check to see if we can evict a buffer
1368 * object.
1369 *
1370 * Return true if eviction is sensible. Called by ttm_mem_evict_first() on
1371 * behalf of ttm_bo_mem_force_space() which tries to evict buffer objects until
1372 * it can find space for a new object and by ttm_bo_force_list_clean() which is
1373 * used to clean out a memory space.
1374 */
1375static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1376 const struct ttm_place *place)
1377{
1378 struct dma_resv_iter resv_cursor;
1379 struct dma_fence *f;
1380
1381 if (!amdgpu_bo_is_amdgpu_bo(bo))
1382 return ttm_bo_eviction_valuable(bo, place);
1383
1384 /* Swapout? */
1385 if (bo->resource->mem_type == TTM_PL_SYSTEM)
1386 return true;
1387
1388 if (bo->type == ttm_bo_type_kernel &&
1389 !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo)))
1390 return false;
1391
1392 /* If bo is a KFD BO, check if the bo belongs to the current process.
1393 * If true, then return false as any KFD process needs all its BOs to
1394 * be resident to run successfully
1395 */
1396 dma_resv_for_each_fence(&resv_cursor, bo->base.resv,
1397 DMA_RESV_USAGE_BOOKKEEP, f) {
1398 if (amdkfd_fence_check_mm(f, current->mm))
1399 return false;
1400 }
1401
1402 /* Preemptible BOs don't own system resources managed by the
1403 * driver (pages, VRAM, GART space). They point to resources
1404 * owned by someone else (e.g. pageable memory in user mode
1405 * or a DMABuf). They are used in a preemptible context so we
1406 * can guarantee no deadlocks and good QoS in case of MMU
1407 * notifiers or DMABuf move notifiers from the resource owner.
1408 */
1409 if (bo->resource->mem_type == AMDGPU_PL_PREEMPT)
1410 return false;
1411
1412 if (bo->resource->mem_type == TTM_PL_TT &&
1413 amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo)))
1414 return false;
1415
1416 return ttm_bo_eviction_valuable(bo, place);
1417}
1418
1419static void amdgpu_ttm_vram_mm_access(struct amdgpu_device *adev, loff_t pos,
1420 void *buf, size_t size, bool write)
1421{
1422 while (size) {
1423 uint64_t aligned_pos = ALIGN_DOWN(pos, 4);
1424 uint64_t bytes = 4 - (pos & 0x3);
1425 uint32_t shift = (pos & 0x3) * 8;
1426 uint32_t mask = 0xffffffff << shift;
1427 uint32_t value = 0;
1428
1429 if (size < bytes) {
1430 mask &= 0xffffffff >> (bytes - size) * 8;
1431 bytes = size;
1432 }
1433
1434 if (mask != 0xffffffff) {
1435 amdgpu_device_mm_access(adev, aligned_pos, &value, 4, false);
1436 if (write) {
1437 value &= ~mask;
1438 value |= (*(uint32_t *)buf << shift) & mask;
1439 amdgpu_device_mm_access(adev, aligned_pos, &value, 4, true);
1440 } else {
1441 value = (value & mask) >> shift;
1442 memcpy(buf, &value, bytes);
1443 }
1444 } else {
1445 amdgpu_device_mm_access(adev, aligned_pos, buf, 4, write);
1446 }
1447
1448 pos += bytes;
1449 buf += bytes;
1450 size -= bytes;
1451 }
1452}
1453
1454static int amdgpu_ttm_access_memory_sdma(struct ttm_buffer_object *bo,
1455 unsigned long offset, void *buf,
1456 int len, int write)
1457{
1458 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1459 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1460 struct amdgpu_res_cursor src_mm;
1461 struct amdgpu_job *job;
1462 struct dma_fence *fence;
1463 uint64_t src_addr, dst_addr;
1464 unsigned int num_dw;
1465 int r, idx;
1466
1467 if (len != PAGE_SIZE)
1468 return -EINVAL;
1469
1470 if (!adev->mman.sdma_access_ptr)
1471 return -EACCES;
1472
1473 if (!drm_dev_enter(adev_to_drm(adev), &idx))
1474 return -ENODEV;
1475
1476 if (write)
1477 memcpy(adev->mman.sdma_access_ptr, buf, len);
1478
1479 num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
1480 r = amdgpu_job_alloc_with_ib(adev, &adev->mman.high_pr,
1481 AMDGPU_FENCE_OWNER_UNDEFINED,
1482 num_dw * 4, AMDGPU_IB_POOL_DELAYED,
1483 &job);
1484 if (r)
1485 goto out;
1486
1487 amdgpu_res_first(abo->tbo.resource, offset, len, &src_mm);
1488 src_addr = amdgpu_ttm_domain_start(adev, bo->resource->mem_type) +
1489 src_mm.start;
1490 dst_addr = amdgpu_bo_gpu_offset(adev->mman.sdma_access_bo);
1491 if (write)
1492 swap(src_addr, dst_addr);
1493
1494 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, dst_addr,
1495 PAGE_SIZE, false);
1496
1497 amdgpu_ring_pad_ib(adev->mman.buffer_funcs_ring, &job->ibs[0]);
1498 WARN_ON(job->ibs[0].length_dw > num_dw);
1499
1500 fence = amdgpu_job_submit(job);
1501
1502 if (!dma_fence_wait_timeout(fence, false, adev->sdma_timeout))
1503 r = -ETIMEDOUT;
1504 dma_fence_put(fence);
1505
1506 if (!(r || write))
1507 memcpy(buf, adev->mman.sdma_access_ptr, len);
1508out:
1509 drm_dev_exit(idx);
1510 return r;
1511}
1512
1513/**
1514 * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object.
1515 *
1516 * @bo: The buffer object to read/write
1517 * @offset: Offset into buffer object
1518 * @buf: Secondary buffer to write/read from
1519 * @len: Length in bytes of access
1520 * @write: true if writing
1521 *
1522 * This is used to access VRAM that backs a buffer object via MMIO
1523 * access for debugging purposes.
1524 */
1525static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1526 unsigned long offset, void *buf, int len,
1527 int write)
1528{
1529 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1530 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1531 struct amdgpu_res_cursor cursor;
1532 int ret = 0;
1533
1534 if (bo->resource->mem_type != TTM_PL_VRAM)
1535 return -EIO;
1536
1537 if (amdgpu_device_has_timeouts_enabled(adev) &&
1538 !amdgpu_ttm_access_memory_sdma(bo, offset, buf, len, write))
1539 return len;
1540
1541 amdgpu_res_first(bo->resource, offset, len, &cursor);
1542 while (cursor.remaining) {
1543 size_t count, size = cursor.size;
1544 loff_t pos = cursor.start;
1545
1546 count = amdgpu_device_aper_access(adev, pos, buf, size, write);
1547 size -= count;
1548 if (size) {
1549 /* using MM to access rest vram and handle un-aligned address */
1550 pos += count;
1551 buf += count;
1552 amdgpu_ttm_vram_mm_access(adev, pos, buf, size, write);
1553 }
1554
1555 ret += cursor.size;
1556 buf += cursor.size;
1557 amdgpu_res_next(&cursor, cursor.size);
1558 }
1559
1560 return ret;
1561}
1562
1563static void
1564amdgpu_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1565{
1566 amdgpu_bo_move_notify(bo, false, NULL);
1567}
1568
1569static struct ttm_device_funcs amdgpu_bo_driver = {
1570 .ttm_tt_create = &amdgpu_ttm_tt_create,
1571 .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1572 .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1573 .ttm_tt_destroy = &amdgpu_ttm_backend_destroy,
1574 .eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1575 .evict_flags = &amdgpu_evict_flags,
1576 .move = &amdgpu_bo_move,
1577 .delete_mem_notify = &amdgpu_bo_delete_mem_notify,
1578 .release_notify = &amdgpu_bo_release_notify,
1579 .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1580 .io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1581 .access_memory = &amdgpu_ttm_access_memory,
1582};
1583
1584/*
1585 * Firmware Reservation functions
1586 */
1587/**
1588 * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram
1589 *
1590 * @adev: amdgpu_device pointer
1591 *
1592 * free fw reserved vram if it has been reserved.
1593 */
1594static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
1595{
1596 amdgpu_bo_free_kernel(&adev->mman.fw_vram_usage_reserved_bo,
1597 NULL, &adev->mman.fw_vram_usage_va);
1598}
1599
1600/*
1601 * Driver Reservation functions
1602 */
1603/**
1604 * amdgpu_ttm_drv_reserve_vram_fini - free drv reserved vram
1605 *
1606 * @adev: amdgpu_device pointer
1607 *
1608 * free drv reserved vram if it has been reserved.
1609 */
1610static void amdgpu_ttm_drv_reserve_vram_fini(struct amdgpu_device *adev)
1611{
1612 amdgpu_bo_free_kernel(&adev->mman.drv_vram_usage_reserved_bo,
1613 NULL,
1614 &adev->mman.drv_vram_usage_va);
1615}
1616
1617/**
1618 * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
1619 *
1620 * @adev: amdgpu_device pointer
1621 *
1622 * create bo vram reservation from fw.
1623 */
1624static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
1625{
1626 uint64_t vram_size = adev->gmc.visible_vram_size;
1627
1628 adev->mman.fw_vram_usage_va = NULL;
1629 adev->mman.fw_vram_usage_reserved_bo = NULL;
1630
1631 if (adev->mman.fw_vram_usage_size == 0 ||
1632 adev->mman.fw_vram_usage_size > vram_size)
1633 return 0;
1634
1635 return amdgpu_bo_create_kernel_at(adev,
1636 adev->mman.fw_vram_usage_start_offset,
1637 adev->mman.fw_vram_usage_size,
1638 &adev->mman.fw_vram_usage_reserved_bo,
1639 &adev->mman.fw_vram_usage_va);
1640}
1641
1642/**
1643 * amdgpu_ttm_drv_reserve_vram_init - create bo vram reservation from driver
1644 *
1645 * @adev: amdgpu_device pointer
1646 *
1647 * create bo vram reservation from drv.
1648 */
1649static int amdgpu_ttm_drv_reserve_vram_init(struct amdgpu_device *adev)
1650{
1651 u64 vram_size = adev->gmc.visible_vram_size;
1652
1653 adev->mman.drv_vram_usage_va = NULL;
1654 adev->mman.drv_vram_usage_reserved_bo = NULL;
1655
1656 if (adev->mman.drv_vram_usage_size == 0 ||
1657 adev->mman.drv_vram_usage_size > vram_size)
1658 return 0;
1659
1660 return amdgpu_bo_create_kernel_at(adev,
1661 adev->mman.drv_vram_usage_start_offset,
1662 adev->mman.drv_vram_usage_size,
1663 &adev->mman.drv_vram_usage_reserved_bo,
1664 &adev->mman.drv_vram_usage_va);
1665}
1666
1667/*
1668 * Memoy training reservation functions
1669 */
1670
1671/**
1672 * amdgpu_ttm_training_reserve_vram_fini - free memory training reserved vram
1673 *
1674 * @adev: amdgpu_device pointer
1675 *
1676 * free memory training reserved vram if it has been reserved.
1677 */
1678static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev)
1679{
1680 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1681
1682 ctx->init = PSP_MEM_TRAIN_NOT_SUPPORT;
1683 amdgpu_bo_free_kernel(&ctx->c2p_bo, NULL, NULL);
1684 ctx->c2p_bo = NULL;
1685
1686 return 0;
1687}
1688
1689static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev,
1690 uint32_t reserve_size)
1691{
1692 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1693
1694 memset(ctx, 0, sizeof(*ctx));
1695
1696 ctx->c2p_train_data_offset =
1697 ALIGN((adev->gmc.mc_vram_size - reserve_size - SZ_1M), SZ_1M);
1698 ctx->p2c_train_data_offset =
1699 (adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET);
1700 ctx->train_data_size =
1701 GDDR6_MEM_TRAINING_DATA_SIZE_IN_BYTES;
1702
1703 DRM_DEBUG("train_data_size:%llx,p2c_train_data_offset:%llx,c2p_train_data_offset:%llx.\n",
1704 ctx->train_data_size,
1705 ctx->p2c_train_data_offset,
1706 ctx->c2p_train_data_offset);
1707}
1708
1709/*
1710 * reserve TMR memory at the top of VRAM which holds
1711 * IP Discovery data and is protected by PSP.
1712 */
1713static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
1714{
1715 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1716 bool mem_train_support = false;
1717 uint32_t reserve_size = 0;
1718 int ret;
1719
1720 if (adev->bios && !amdgpu_sriov_vf(adev)) {
1721 if (amdgpu_atomfirmware_mem_training_supported(adev))
1722 mem_train_support = true;
1723 else
1724 DRM_DEBUG("memory training does not support!\n");
1725 }
1726
1727 /*
1728 * Query reserved tmr size through atom firmwareinfo for Sienna_Cichlid and onwards for all
1729 * the use cases (IP discovery/G6 memory training/profiling/diagnostic data.etc)
1730 *
1731 * Otherwise, fallback to legacy approach to check and reserve tmr block for ip
1732 * discovery data and G6 memory training data respectively
1733 */
1734 if (adev->bios)
1735 reserve_size =
1736 amdgpu_atomfirmware_get_fw_reserved_fb_size(adev);
1737
1738 if (!adev->bios &&
1739 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3))
1740 reserve_size = max(reserve_size, (uint32_t)280 << 20);
1741 else if (!reserve_size)
1742 reserve_size = DISCOVERY_TMR_OFFSET;
1743
1744 if (mem_train_support) {
1745 /* reserve vram for mem train according to TMR location */
1746 amdgpu_ttm_training_data_block_init(adev, reserve_size);
1747 ret = amdgpu_bo_create_kernel_at(adev,
1748 ctx->c2p_train_data_offset,
1749 ctx->train_data_size,
1750 &ctx->c2p_bo,
1751 NULL);
1752 if (ret) {
1753 DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret);
1754 amdgpu_ttm_training_reserve_vram_fini(adev);
1755 return ret;
1756 }
1757 ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS;
1758 }
1759
1760 if (!adev->gmc.is_app_apu) {
1761 ret = amdgpu_bo_create_kernel_at(
1762 adev, adev->gmc.real_vram_size - reserve_size,
1763 reserve_size, &adev->mman.fw_reserved_memory, NULL);
1764 if (ret) {
1765 DRM_ERROR("alloc tmr failed(%d)!\n", ret);
1766 amdgpu_bo_free_kernel(&adev->mman.fw_reserved_memory,
1767 NULL, NULL);
1768 return ret;
1769 }
1770 } else {
1771 DRM_DEBUG_DRIVER("backdoor fw loading path for PSP TMR, no reservation needed\n");
1772 }
1773
1774 return 0;
1775}
1776
1777static int amdgpu_ttm_pools_init(struct amdgpu_device *adev)
1778{
1779 int i;
1780
1781 if (!adev->gmc.is_app_apu || !adev->gmc.num_mem_partitions)
1782 return 0;
1783
1784 adev->mman.ttm_pools = kcalloc(adev->gmc.num_mem_partitions,
1785 sizeof(*adev->mman.ttm_pools),
1786 GFP_KERNEL);
1787 if (!adev->mman.ttm_pools)
1788 return -ENOMEM;
1789
1790 for (i = 0; i < adev->gmc.num_mem_partitions; i++) {
1791 ttm_pool_init(&adev->mman.ttm_pools[i], adev->dev,
1792 adev->gmc.mem_partitions[i].numa.node,
1793 false, false);
1794 }
1795 return 0;
1796}
1797
1798static void amdgpu_ttm_pools_fini(struct amdgpu_device *adev)
1799{
1800 int i;
1801
1802 if (!adev->gmc.is_app_apu || !adev->mman.ttm_pools)
1803 return;
1804
1805 for (i = 0; i < adev->gmc.num_mem_partitions; i++)
1806 ttm_pool_fini(&adev->mman.ttm_pools[i]);
1807
1808 kfree(adev->mman.ttm_pools);
1809 adev->mman.ttm_pools = NULL;
1810}
1811
1812/*
1813 * amdgpu_ttm_init - Init the memory management (ttm) as well as various
1814 * gtt/vram related fields.
1815 *
1816 * This initializes all of the memory space pools that the TTM layer
1817 * will need such as the GTT space (system memory mapped to the device),
1818 * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which
1819 * can be mapped per VMID.
1820 */
1821int amdgpu_ttm_init(struct amdgpu_device *adev)
1822{
1823 uint64_t gtt_size;
1824 int r;
1825
1826 mutex_init(&adev->mman.gtt_window_lock);
1827
1828 /* No others user of address space so set it to 0 */
1829 r = ttm_device_init(&adev->mman.bdev, &amdgpu_bo_driver, adev->dev,
1830 adev_to_drm(adev)->anon_inode->i_mapping,
1831 adev_to_drm(adev)->vma_offset_manager,
1832 adev->need_swiotlb,
1833 dma_addressing_limited(adev->dev));
1834 if (r) {
1835 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1836 return r;
1837 }
1838
1839 r = amdgpu_ttm_pools_init(adev);
1840 if (r) {
1841 DRM_ERROR("failed to init ttm pools(%d).\n", r);
1842 return r;
1843 }
1844 adev->mman.initialized = true;
1845
1846 /* Initialize VRAM pool with all of VRAM divided into pages */
1847 r = amdgpu_vram_mgr_init(adev);
1848 if (r) {
1849 DRM_ERROR("Failed initializing VRAM heap.\n");
1850 return r;
1851 }
1852
1853 /* Change the size here instead of the init above so only lpfn is affected */
1854 amdgpu_ttm_set_buffer_funcs_status(adev, false);
1855#ifdef CONFIG_64BIT
1856#ifdef CONFIG_X86
1857 if (adev->gmc.xgmi.connected_to_cpu)
1858 adev->mman.aper_base_kaddr = ioremap_cache(adev->gmc.aper_base,
1859 adev->gmc.visible_vram_size);
1860
1861 else if (adev->gmc.is_app_apu)
1862 DRM_DEBUG_DRIVER(
1863 "No need to ioremap when real vram size is 0\n");
1864 else
1865#endif
1866 adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
1867 adev->gmc.visible_vram_size);
1868#endif
1869
1870 /*
1871 *The reserved vram for firmware must be pinned to the specified
1872 *place on the VRAM, so reserve it early.
1873 */
1874 r = amdgpu_ttm_fw_reserve_vram_init(adev);
1875 if (r)
1876 return r;
1877
1878 /*
1879 *The reserved vram for driver must be pinned to the specified
1880 *place on the VRAM, so reserve it early.
1881 */
1882 r = amdgpu_ttm_drv_reserve_vram_init(adev);
1883 if (r)
1884 return r;
1885
1886 /*
1887 * only NAVI10 and onwards ASIC support for IP discovery.
1888 * If IP discovery enabled, a block of memory should be
1889 * reserved for IP discovey.
1890 */
1891 if (adev->mman.discovery_bin) {
1892 r = amdgpu_ttm_reserve_tmr(adev);
1893 if (r)
1894 return r;
1895 }
1896
1897 /* allocate memory as required for VGA
1898 * This is used for VGA emulation and pre-OS scanout buffers to
1899 * avoid display artifacts while transitioning between pre-OS
1900 * and driver.
1901 */
1902 if (!adev->gmc.is_app_apu) {
1903 r = amdgpu_bo_create_kernel_at(adev, 0,
1904 adev->mman.stolen_vga_size,
1905 &adev->mman.stolen_vga_memory,
1906 NULL);
1907 if (r)
1908 return r;
1909
1910 r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size,
1911 adev->mman.stolen_extended_size,
1912 &adev->mman.stolen_extended_memory,
1913 NULL);
1914
1915 if (r)
1916 return r;
1917
1918 r = amdgpu_bo_create_kernel_at(adev,
1919 adev->mman.stolen_reserved_offset,
1920 adev->mman.stolen_reserved_size,
1921 &adev->mman.stolen_reserved_memory,
1922 NULL);
1923 if (r)
1924 return r;
1925 } else {
1926 DRM_DEBUG_DRIVER("Skipped stolen memory reservation\n");
1927 }
1928
1929 DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1930 (unsigned int)(adev->gmc.real_vram_size / (1024 * 1024)));
1931
1932 /* Compute GTT size, either based on TTM limit
1933 * or whatever the user passed on module init.
1934 */
1935 if (amdgpu_gtt_size == -1)
1936 gtt_size = ttm_tt_pages_limit() << PAGE_SHIFT;
1937 else
1938 gtt_size = (uint64_t)amdgpu_gtt_size << 20;
1939
1940 /* Initialize GTT memory pool */
1941 r = amdgpu_gtt_mgr_init(adev, gtt_size);
1942 if (r) {
1943 DRM_ERROR("Failed initializing GTT heap.\n");
1944 return r;
1945 }
1946 DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1947 (unsigned int)(gtt_size / (1024 * 1024)));
1948
1949 /* Initiailize doorbell pool on PCI BAR */
1950 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_DOORBELL, adev->doorbell.size / PAGE_SIZE);
1951 if (r) {
1952 DRM_ERROR("Failed initializing doorbell heap.\n");
1953 return r;
1954 }
1955
1956 /* Create a boorbell page for kernel usages */
1957 r = amdgpu_doorbell_create_kernel_doorbells(adev);
1958 if (r) {
1959 DRM_ERROR("Failed to initialize kernel doorbells.\n");
1960 return r;
1961 }
1962
1963 /* Initialize preemptible memory pool */
1964 r = amdgpu_preempt_mgr_init(adev);
1965 if (r) {
1966 DRM_ERROR("Failed initializing PREEMPT heap.\n");
1967 return r;
1968 }
1969
1970 /* Initialize various on-chip memory pools */
1971 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GDS, adev->gds.gds_size);
1972 if (r) {
1973 DRM_ERROR("Failed initializing GDS heap.\n");
1974 return r;
1975 }
1976
1977 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GWS, adev->gds.gws_size);
1978 if (r) {
1979 DRM_ERROR("Failed initializing gws heap.\n");
1980 return r;
1981 }
1982
1983 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_OA, adev->gds.oa_size);
1984 if (r) {
1985 DRM_ERROR("Failed initializing oa heap.\n");
1986 return r;
1987 }
1988 if (amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
1989 AMDGPU_GEM_DOMAIN_GTT,
1990 &adev->mman.sdma_access_bo, NULL,
1991 &adev->mman.sdma_access_ptr))
1992 DRM_WARN("Debug VRAM access will use slowpath MM access\n");
1993
1994 return 0;
1995}
1996
1997/*
1998 * amdgpu_ttm_fini - De-initialize the TTM memory pools
1999 */
2000void amdgpu_ttm_fini(struct amdgpu_device *adev)
2001{
2002 int idx;
2003
2004 if (!adev->mman.initialized)
2005 return;
2006
2007 amdgpu_ttm_pools_fini(adev);
2008
2009 amdgpu_ttm_training_reserve_vram_fini(adev);
2010 /* return the stolen vga memory back to VRAM */
2011 if (!adev->gmc.is_app_apu) {
2012 amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL);
2013 amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL);
2014 /* return the FW reserved memory back to VRAM */
2015 amdgpu_bo_free_kernel(&adev->mman.fw_reserved_memory, NULL,
2016 NULL);
2017 if (adev->mman.stolen_reserved_size)
2018 amdgpu_bo_free_kernel(&adev->mman.stolen_reserved_memory,
2019 NULL, NULL);
2020 }
2021 amdgpu_bo_free_kernel(&adev->mman.sdma_access_bo, NULL,
2022 &adev->mman.sdma_access_ptr);
2023 amdgpu_ttm_fw_reserve_vram_fini(adev);
2024 amdgpu_ttm_drv_reserve_vram_fini(adev);
2025
2026 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
2027
2028 if (adev->mman.aper_base_kaddr)
2029 iounmap(adev->mman.aper_base_kaddr);
2030 adev->mman.aper_base_kaddr = NULL;
2031
2032 drm_dev_exit(idx);
2033 }
2034
2035 amdgpu_vram_mgr_fini(adev);
2036 amdgpu_gtt_mgr_fini(adev);
2037 amdgpu_preempt_mgr_fini(adev);
2038 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GDS);
2039 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GWS);
2040 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_OA);
2041 ttm_device_fini(&adev->mman.bdev);
2042 adev->mman.initialized = false;
2043 DRM_INFO("amdgpu: ttm finalized\n");
2044}
2045
2046/**
2047 * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions
2048 *
2049 * @adev: amdgpu_device pointer
2050 * @enable: true when we can use buffer functions.
2051 *
2052 * Enable/disable use of buffer functions during suspend/resume. This should
2053 * only be called at bootup or when userspace isn't running.
2054 */
2055void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable)
2056{
2057 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
2058 uint64_t size;
2059 int r;
2060
2061 if (!adev->mman.initialized || amdgpu_in_reset(adev) ||
2062 adev->mman.buffer_funcs_enabled == enable || adev->gmc.is_app_apu)
2063 return;
2064
2065 if (enable) {
2066 struct amdgpu_ring *ring;
2067 struct drm_gpu_scheduler *sched;
2068
2069 ring = adev->mman.buffer_funcs_ring;
2070 sched = &ring->sched;
2071 r = drm_sched_entity_init(&adev->mman.high_pr,
2072 DRM_SCHED_PRIORITY_KERNEL, &sched,
2073 1, NULL);
2074 if (r) {
2075 DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
2076 r);
2077 return;
2078 }
2079
2080 r = drm_sched_entity_init(&adev->mman.low_pr,
2081 DRM_SCHED_PRIORITY_NORMAL, &sched,
2082 1, NULL);
2083 if (r) {
2084 DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
2085 r);
2086 goto error_free_entity;
2087 }
2088 } else {
2089 drm_sched_entity_destroy(&adev->mman.high_pr);
2090 drm_sched_entity_destroy(&adev->mman.low_pr);
2091 dma_fence_put(man->move);
2092 man->move = NULL;
2093 }
2094
2095 /* this just adjusts TTM size idea, which sets lpfn to the correct value */
2096 if (enable)
2097 size = adev->gmc.real_vram_size;
2098 else
2099 size = adev->gmc.visible_vram_size;
2100 man->size = size;
2101 adev->mman.buffer_funcs_enabled = enable;
2102
2103 return;
2104
2105error_free_entity:
2106 drm_sched_entity_destroy(&adev->mman.high_pr);
2107}
2108
2109static int amdgpu_ttm_prepare_job(struct amdgpu_device *adev,
2110 bool direct_submit,
2111 unsigned int num_dw,
2112 struct dma_resv *resv,
2113 bool vm_needs_flush,
2114 struct amdgpu_job **job,
2115 bool delayed)
2116{
2117 enum amdgpu_ib_pool_type pool = direct_submit ?
2118 AMDGPU_IB_POOL_DIRECT :
2119 AMDGPU_IB_POOL_DELAYED;
2120 int r;
2121 struct drm_sched_entity *entity = delayed ? &adev->mman.low_pr :
2122 &adev->mman.high_pr;
2123 r = amdgpu_job_alloc_with_ib(adev, entity,
2124 AMDGPU_FENCE_OWNER_UNDEFINED,
2125 num_dw * 4, pool, job);
2126 if (r)
2127 return r;
2128
2129 if (vm_needs_flush) {
2130 (*job)->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gmc.pdb0_bo ?
2131 adev->gmc.pdb0_bo :
2132 adev->gart.bo);
2133 (*job)->vm_needs_flush = true;
2134 }
2135 if (!resv)
2136 return 0;
2137
2138 return drm_sched_job_add_resv_dependencies(&(*job)->base, resv,
2139 DMA_RESV_USAGE_BOOKKEEP);
2140}
2141
2142int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
2143 uint64_t dst_offset, uint32_t byte_count,
2144 struct dma_resv *resv,
2145 struct dma_fence **fence, bool direct_submit,
2146 bool vm_needs_flush, bool tmz)
2147{
2148 struct amdgpu_device *adev = ring->adev;
2149 unsigned int num_loops, num_dw;
2150 struct amdgpu_job *job;
2151 uint32_t max_bytes;
2152 unsigned int i;
2153 int r;
2154
2155 if (!direct_submit && !ring->sched.ready) {
2156 DRM_ERROR("Trying to move memory with ring turned off.\n");
2157 return -EINVAL;
2158 }
2159
2160 max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
2161 num_loops = DIV_ROUND_UP(byte_count, max_bytes);
2162 num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8);
2163 r = amdgpu_ttm_prepare_job(adev, direct_submit, num_dw,
2164 resv, vm_needs_flush, &job, false);
2165 if (r)
2166 return r;
2167
2168 for (i = 0; i < num_loops; i++) {
2169 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
2170
2171 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
2172 dst_offset, cur_size_in_bytes, tmz);
2173
2174 src_offset += cur_size_in_bytes;
2175 dst_offset += cur_size_in_bytes;
2176 byte_count -= cur_size_in_bytes;
2177 }
2178
2179 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2180 WARN_ON(job->ibs[0].length_dw > num_dw);
2181 if (direct_submit)
2182 r = amdgpu_job_submit_direct(job, ring, fence);
2183 else
2184 *fence = amdgpu_job_submit(job);
2185 if (r)
2186 goto error_free;
2187
2188 return r;
2189
2190error_free:
2191 amdgpu_job_free(job);
2192 DRM_ERROR("Error scheduling IBs (%d)\n", r);
2193 return r;
2194}
2195
2196static int amdgpu_ttm_fill_mem(struct amdgpu_ring *ring, uint32_t src_data,
2197 uint64_t dst_addr, uint32_t byte_count,
2198 struct dma_resv *resv,
2199 struct dma_fence **fence,
2200 bool vm_needs_flush, bool delayed)
2201{
2202 struct amdgpu_device *adev = ring->adev;
2203 unsigned int num_loops, num_dw;
2204 struct amdgpu_job *job;
2205 uint32_t max_bytes;
2206 unsigned int i;
2207 int r;
2208
2209 max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
2210 num_loops = DIV_ROUND_UP_ULL(byte_count, max_bytes);
2211 num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->fill_num_dw, 8);
2212 r = amdgpu_ttm_prepare_job(adev, false, num_dw, resv, vm_needs_flush,
2213 &job, delayed);
2214 if (r)
2215 return r;
2216
2217 for (i = 0; i < num_loops; i++) {
2218 uint32_t cur_size = min(byte_count, max_bytes);
2219
2220 amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data, dst_addr,
2221 cur_size);
2222
2223 dst_addr += cur_size;
2224 byte_count -= cur_size;
2225 }
2226
2227 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2228 WARN_ON(job->ibs[0].length_dw > num_dw);
2229 *fence = amdgpu_job_submit(job);
2230 return 0;
2231}
2232
2233int amdgpu_fill_buffer(struct amdgpu_bo *bo,
2234 uint32_t src_data,
2235 struct dma_resv *resv,
2236 struct dma_fence **f,
2237 bool delayed)
2238{
2239 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
2240 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2241 struct dma_fence *fence = NULL;
2242 struct amdgpu_res_cursor dst;
2243 int r;
2244
2245 if (!adev->mman.buffer_funcs_enabled) {
2246 DRM_ERROR("Trying to clear memory with ring turned off.\n");
2247 return -EINVAL;
2248 }
2249
2250 amdgpu_res_first(bo->tbo.resource, 0, amdgpu_bo_size(bo), &dst);
2251
2252 mutex_lock(&adev->mman.gtt_window_lock);
2253 while (dst.remaining) {
2254 struct dma_fence *next;
2255 uint64_t cur_size, to;
2256
2257 /* Never fill more than 256MiB at once to avoid timeouts */
2258 cur_size = min(dst.size, 256ULL << 20);
2259
2260 r = amdgpu_ttm_map_buffer(&bo->tbo, bo->tbo.resource, &dst,
2261 1, ring, false, &cur_size, &to);
2262 if (r)
2263 goto error;
2264
2265 r = amdgpu_ttm_fill_mem(ring, src_data, to, cur_size, resv,
2266 &next, true, delayed);
2267 if (r)
2268 goto error;
2269
2270 dma_fence_put(fence);
2271 fence = next;
2272
2273 amdgpu_res_next(&dst, cur_size);
2274 }
2275error:
2276 mutex_unlock(&adev->mman.gtt_window_lock);
2277 if (f)
2278 *f = dma_fence_get(fence);
2279 dma_fence_put(fence);
2280 return r;
2281}
2282
2283/**
2284 * amdgpu_ttm_evict_resources - evict memory buffers
2285 * @adev: amdgpu device object
2286 * @mem_type: evicted BO's memory type
2287 *
2288 * Evicts all @mem_type buffers on the lru list of the memory type.
2289 *
2290 * Returns:
2291 * 0 for success or a negative error code on failure.
2292 */
2293int amdgpu_ttm_evict_resources(struct amdgpu_device *adev, int mem_type)
2294{
2295 struct ttm_resource_manager *man;
2296
2297 switch (mem_type) {
2298 case TTM_PL_VRAM:
2299 case TTM_PL_TT:
2300 case AMDGPU_PL_GWS:
2301 case AMDGPU_PL_GDS:
2302 case AMDGPU_PL_OA:
2303 man = ttm_manager_type(&adev->mman.bdev, mem_type);
2304 break;
2305 default:
2306 DRM_ERROR("Trying to evict invalid memory type\n");
2307 return -EINVAL;
2308 }
2309
2310 return ttm_resource_manager_evict_all(&adev->mman.bdev, man);
2311}
2312
2313#if defined(CONFIG_DEBUG_FS)
2314
2315static int amdgpu_ttm_page_pool_show(struct seq_file *m, void *unused)
2316{
2317 struct amdgpu_device *adev = m->private;
2318
2319 return ttm_pool_debugfs(&adev->mman.bdev.pool, m);
2320}
2321
2322DEFINE_SHOW_ATTRIBUTE(amdgpu_ttm_page_pool);
2323
2324/*
2325 * amdgpu_ttm_vram_read - Linear read access to VRAM
2326 *
2327 * Accesses VRAM via MMIO for debugging purposes.
2328 */
2329static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
2330 size_t size, loff_t *pos)
2331{
2332 struct amdgpu_device *adev = file_inode(f)->i_private;
2333 ssize_t result = 0;
2334
2335 if (size & 0x3 || *pos & 0x3)
2336 return -EINVAL;
2337
2338 if (*pos >= adev->gmc.mc_vram_size)
2339 return -ENXIO;
2340
2341 size = min(size, (size_t)(adev->gmc.mc_vram_size - *pos));
2342 while (size) {
2343 size_t bytes = min(size, AMDGPU_TTM_VRAM_MAX_DW_READ * 4);
2344 uint32_t value[AMDGPU_TTM_VRAM_MAX_DW_READ];
2345
2346 amdgpu_device_vram_access(adev, *pos, value, bytes, false);
2347 if (copy_to_user(buf, value, bytes))
2348 return -EFAULT;
2349
2350 result += bytes;
2351 buf += bytes;
2352 *pos += bytes;
2353 size -= bytes;
2354 }
2355
2356 return result;
2357}
2358
2359/*
2360 * amdgpu_ttm_vram_write - Linear write access to VRAM
2361 *
2362 * Accesses VRAM via MMIO for debugging purposes.
2363 */
2364static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf,
2365 size_t size, loff_t *pos)
2366{
2367 struct amdgpu_device *adev = file_inode(f)->i_private;
2368 ssize_t result = 0;
2369 int r;
2370
2371 if (size & 0x3 || *pos & 0x3)
2372 return -EINVAL;
2373
2374 if (*pos >= adev->gmc.mc_vram_size)
2375 return -ENXIO;
2376
2377 while (size) {
2378 uint32_t value;
2379
2380 if (*pos >= adev->gmc.mc_vram_size)
2381 return result;
2382
2383 r = get_user(value, (uint32_t *)buf);
2384 if (r)
2385 return r;
2386
2387 amdgpu_device_mm_access(adev, *pos, &value, 4, true);
2388
2389 result += 4;
2390 buf += 4;
2391 *pos += 4;
2392 size -= 4;
2393 }
2394
2395 return result;
2396}
2397
2398static const struct file_operations amdgpu_ttm_vram_fops = {
2399 .owner = THIS_MODULE,
2400 .read = amdgpu_ttm_vram_read,
2401 .write = amdgpu_ttm_vram_write,
2402 .llseek = default_llseek,
2403};
2404
2405/*
2406 * amdgpu_iomem_read - Virtual read access to GPU mapped memory
2407 *
2408 * This function is used to read memory that has been mapped to the
2409 * GPU and the known addresses are not physical addresses but instead
2410 * bus addresses (e.g., what you'd put in an IB or ring buffer).
2411 */
2412static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
2413 size_t size, loff_t *pos)
2414{
2415 struct amdgpu_device *adev = file_inode(f)->i_private;
2416 struct iommu_domain *dom;
2417 ssize_t result = 0;
2418 int r;
2419
2420 /* retrieve the IOMMU domain if any for this device */
2421 dom = iommu_get_domain_for_dev(adev->dev);
2422
2423 while (size) {
2424 phys_addr_t addr = *pos & PAGE_MASK;
2425 loff_t off = *pos & ~PAGE_MASK;
2426 size_t bytes = PAGE_SIZE - off;
2427 unsigned long pfn;
2428 struct page *p;
2429 void *ptr;
2430
2431 bytes = min(bytes, size);
2432
2433 /* Translate the bus address to a physical address. If
2434 * the domain is NULL it means there is no IOMMU active
2435 * and the address translation is the identity
2436 */
2437 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2438
2439 pfn = addr >> PAGE_SHIFT;
2440 if (!pfn_valid(pfn))
2441 return -EPERM;
2442
2443 p = pfn_to_page(pfn);
2444 if (p->mapping != adev->mman.bdev.dev_mapping)
2445 return -EPERM;
2446
2447 ptr = kmap_local_page(p);
2448 r = copy_to_user(buf, ptr + off, bytes);
2449 kunmap_local(ptr);
2450 if (r)
2451 return -EFAULT;
2452
2453 size -= bytes;
2454 *pos += bytes;
2455 result += bytes;
2456 }
2457
2458 return result;
2459}
2460
2461/*
2462 * amdgpu_iomem_write - Virtual write access to GPU mapped memory
2463 *
2464 * This function is used to write memory that has been mapped to the
2465 * GPU and the known addresses are not physical addresses but instead
2466 * bus addresses (e.g., what you'd put in an IB or ring buffer).
2467 */
2468static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
2469 size_t size, loff_t *pos)
2470{
2471 struct amdgpu_device *adev = file_inode(f)->i_private;
2472 struct iommu_domain *dom;
2473 ssize_t result = 0;
2474 int r;
2475
2476 dom = iommu_get_domain_for_dev(adev->dev);
2477
2478 while (size) {
2479 phys_addr_t addr = *pos & PAGE_MASK;
2480 loff_t off = *pos & ~PAGE_MASK;
2481 size_t bytes = PAGE_SIZE - off;
2482 unsigned long pfn;
2483 struct page *p;
2484 void *ptr;
2485
2486 bytes = min(bytes, size);
2487
2488 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2489
2490 pfn = addr >> PAGE_SHIFT;
2491 if (!pfn_valid(pfn))
2492 return -EPERM;
2493
2494 p = pfn_to_page(pfn);
2495 if (p->mapping != adev->mman.bdev.dev_mapping)
2496 return -EPERM;
2497
2498 ptr = kmap_local_page(p);
2499 r = copy_from_user(ptr + off, buf, bytes);
2500 kunmap_local(ptr);
2501 if (r)
2502 return -EFAULT;
2503
2504 size -= bytes;
2505 *pos += bytes;
2506 result += bytes;
2507 }
2508
2509 return result;
2510}
2511
2512static const struct file_operations amdgpu_ttm_iomem_fops = {
2513 .owner = THIS_MODULE,
2514 .read = amdgpu_iomem_read,
2515 .write = amdgpu_iomem_write,
2516 .llseek = default_llseek
2517};
2518
2519#endif
2520
2521void amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
2522{
2523#if defined(CONFIG_DEBUG_FS)
2524 struct drm_minor *minor = adev_to_drm(adev)->primary;
2525 struct dentry *root = minor->debugfs_root;
2526
2527 debugfs_create_file_size("amdgpu_vram", 0444, root, adev,
2528 &amdgpu_ttm_vram_fops, adev->gmc.mc_vram_size);
2529 debugfs_create_file("amdgpu_iomem", 0444, root, adev,
2530 &amdgpu_ttm_iomem_fops);
2531 debugfs_create_file("ttm_page_pool", 0444, root, adev,
2532 &amdgpu_ttm_page_pool_fops);
2533 ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2534 TTM_PL_VRAM),
2535 root, "amdgpu_vram_mm");
2536 ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2537 TTM_PL_TT),
2538 root, "amdgpu_gtt_mm");
2539 ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2540 AMDGPU_PL_GDS),
2541 root, "amdgpu_gds_mm");
2542 ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2543 AMDGPU_PL_GWS),
2544 root, "amdgpu_gws_mm");
2545 ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2546 AMDGPU_PL_OA),
2547 root, "amdgpu_oa_mm");
2548
2549#endif
2550}
1/*
2 * Copyright 2009 Jerome Glisse.
3 * All Rights Reserved.
4 *
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:
12 *
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.
20 *
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
23 * of the Software.
24 *
25 */
26/*
27 * Authors:
28 * Jerome Glisse <glisse@freedesktop.org>
29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
30 * Dave Airlie
31 */
32#include <drm/ttm/ttm_bo_api.h>
33#include <drm/ttm/ttm_bo_driver.h>
34#include <drm/ttm/ttm_placement.h>
35#include <drm/ttm/ttm_module.h>
36#include <drm/ttm/ttm_page_alloc.h>
37#include <drm/drmP.h>
38#include <drm/amdgpu_drm.h>
39#include <linux/seq_file.h>
40#include <linux/slab.h>
41#include <linux/swiotlb.h>
42#include <linux/swap.h>
43#include <linux/pagemap.h>
44#include <linux/debugfs.h>
45#include <linux/iommu.h>
46#include "amdgpu.h"
47#include "amdgpu_object.h"
48#include "amdgpu_trace.h"
49#include "amdgpu_amdkfd.h"
50#include "bif/bif_4_1_d.h"
51
52#define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
53
54static int amdgpu_map_buffer(struct ttm_buffer_object *bo,
55 struct ttm_mem_reg *mem, unsigned num_pages,
56 uint64_t offset, unsigned window,
57 struct amdgpu_ring *ring,
58 uint64_t *addr);
59
60static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev);
61static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev);
62
63/*
64 * Global memory.
65 */
66static int amdgpu_ttm_mem_global_init(struct drm_global_reference *ref)
67{
68 return ttm_mem_global_init(ref->object);
69}
70
71static void amdgpu_ttm_mem_global_release(struct drm_global_reference *ref)
72{
73 ttm_mem_global_release(ref->object);
74}
75
76static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
77{
78 struct drm_global_reference *global_ref;
79 struct amdgpu_ring *ring;
80 struct drm_sched_rq *rq;
81 int r;
82
83 adev->mman.mem_global_referenced = false;
84 global_ref = &adev->mman.mem_global_ref;
85 global_ref->global_type = DRM_GLOBAL_TTM_MEM;
86 global_ref->size = sizeof(struct ttm_mem_global);
87 global_ref->init = &amdgpu_ttm_mem_global_init;
88 global_ref->release = &amdgpu_ttm_mem_global_release;
89 r = drm_global_item_ref(global_ref);
90 if (r) {
91 DRM_ERROR("Failed setting up TTM memory accounting "
92 "subsystem.\n");
93 goto error_mem;
94 }
95
96 adev->mman.bo_global_ref.mem_glob =
97 adev->mman.mem_global_ref.object;
98 global_ref = &adev->mman.bo_global_ref.ref;
99 global_ref->global_type = DRM_GLOBAL_TTM_BO;
100 global_ref->size = sizeof(struct ttm_bo_global);
101 global_ref->init = &ttm_bo_global_init;
102 global_ref->release = &ttm_bo_global_release;
103 r = drm_global_item_ref(global_ref);
104 if (r) {
105 DRM_ERROR("Failed setting up TTM BO subsystem.\n");
106 goto error_bo;
107 }
108
109 mutex_init(&adev->mman.gtt_window_lock);
110
111 ring = adev->mman.buffer_funcs_ring;
112 rq = &ring->sched.sched_rq[DRM_SCHED_PRIORITY_KERNEL];
113 r = drm_sched_entity_init(&ring->sched, &adev->mman.entity,
114 rq, amdgpu_sched_jobs, NULL);
115 if (r) {
116 DRM_ERROR("Failed setting up TTM BO move run queue.\n");
117 goto error_entity;
118 }
119
120 adev->mman.mem_global_referenced = true;
121
122 return 0;
123
124error_entity:
125 drm_global_item_unref(&adev->mman.bo_global_ref.ref);
126error_bo:
127 drm_global_item_unref(&adev->mman.mem_global_ref);
128error_mem:
129 return r;
130}
131
132static void amdgpu_ttm_global_fini(struct amdgpu_device *adev)
133{
134 if (adev->mman.mem_global_referenced) {
135 drm_sched_entity_fini(adev->mman.entity.sched,
136 &adev->mman.entity);
137 mutex_destroy(&adev->mman.gtt_window_lock);
138 drm_global_item_unref(&adev->mman.bo_global_ref.ref);
139 drm_global_item_unref(&adev->mman.mem_global_ref);
140 adev->mman.mem_global_referenced = false;
141 }
142}
143
144static int amdgpu_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
145{
146 return 0;
147}
148
149static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
150 struct ttm_mem_type_manager *man)
151{
152 struct amdgpu_device *adev;
153
154 adev = amdgpu_ttm_adev(bdev);
155
156 switch (type) {
157 case TTM_PL_SYSTEM:
158 /* System memory */
159 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
160 man->available_caching = TTM_PL_MASK_CACHING;
161 man->default_caching = TTM_PL_FLAG_CACHED;
162 break;
163 case TTM_PL_TT:
164 man->func = &amdgpu_gtt_mgr_func;
165 man->gpu_offset = adev->gmc.gart_start;
166 man->available_caching = TTM_PL_MASK_CACHING;
167 man->default_caching = TTM_PL_FLAG_CACHED;
168 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
169 break;
170 case TTM_PL_VRAM:
171 /* "On-card" video ram */
172 man->func = &amdgpu_vram_mgr_func;
173 man->gpu_offset = adev->gmc.vram_start;
174 man->flags = TTM_MEMTYPE_FLAG_FIXED |
175 TTM_MEMTYPE_FLAG_MAPPABLE;
176 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
177 man->default_caching = TTM_PL_FLAG_WC;
178 break;
179 case AMDGPU_PL_GDS:
180 case AMDGPU_PL_GWS:
181 case AMDGPU_PL_OA:
182 /* On-chip GDS memory*/
183 man->func = &ttm_bo_manager_func;
184 man->gpu_offset = 0;
185 man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_CMA;
186 man->available_caching = TTM_PL_FLAG_UNCACHED;
187 man->default_caching = TTM_PL_FLAG_UNCACHED;
188 break;
189 default:
190 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
191 return -EINVAL;
192 }
193 return 0;
194}
195
196static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
197 struct ttm_placement *placement)
198{
199 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
200 struct amdgpu_bo *abo;
201 static const struct ttm_place placements = {
202 .fpfn = 0,
203 .lpfn = 0,
204 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
205 };
206
207 if (bo->type == ttm_bo_type_sg) {
208 placement->num_placement = 0;
209 placement->num_busy_placement = 0;
210 return;
211 }
212
213 if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
214 placement->placement = &placements;
215 placement->busy_placement = &placements;
216 placement->num_placement = 1;
217 placement->num_busy_placement = 1;
218 return;
219 }
220 abo = ttm_to_amdgpu_bo(bo);
221 switch (bo->mem.mem_type) {
222 case TTM_PL_VRAM:
223 if (!adev->mman.buffer_funcs_enabled) {
224 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
225 } else if (adev->gmc.visible_vram_size < adev->gmc.real_vram_size &&
226 !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
227 unsigned fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
228 struct drm_mm_node *node = bo->mem.mm_node;
229 unsigned long pages_left;
230
231 for (pages_left = bo->mem.num_pages;
232 pages_left;
233 pages_left -= node->size, node++) {
234 if (node->start < fpfn)
235 break;
236 }
237
238 if (!pages_left)
239 goto gtt;
240
241 /* Try evicting to the CPU inaccessible part of VRAM
242 * first, but only set GTT as busy placement, so this
243 * BO will be evicted to GTT rather than causing other
244 * BOs to be evicted from VRAM
245 */
246 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
247 AMDGPU_GEM_DOMAIN_GTT);
248 abo->placements[0].fpfn = fpfn;
249 abo->placements[0].lpfn = 0;
250 abo->placement.busy_placement = &abo->placements[1];
251 abo->placement.num_busy_placement = 1;
252 } else {
253gtt:
254 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT);
255 }
256 break;
257 case TTM_PL_TT:
258 default:
259 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
260 }
261 *placement = abo->placement;
262}
263
264static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
265{
266 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
267
268 /*
269 * Don't verify access for KFD BOs. They don't have a GEM
270 * object associated with them.
271 */
272 if (abo->kfd_bo)
273 return 0;
274
275 if (amdgpu_ttm_tt_get_usermm(bo->ttm))
276 return -EPERM;
277 return drm_vma_node_verify_access(&abo->gem_base.vma_node,
278 filp->private_data);
279}
280
281static void amdgpu_move_null(struct ttm_buffer_object *bo,
282 struct ttm_mem_reg *new_mem)
283{
284 struct ttm_mem_reg *old_mem = &bo->mem;
285
286 BUG_ON(old_mem->mm_node != NULL);
287 *old_mem = *new_mem;
288 new_mem->mm_node = NULL;
289}
290
291static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
292 struct drm_mm_node *mm_node,
293 struct ttm_mem_reg *mem)
294{
295 uint64_t addr = 0;
296
297 if (mem->mem_type != TTM_PL_TT || amdgpu_gtt_mgr_has_gart_addr(mem)) {
298 addr = mm_node->start << PAGE_SHIFT;
299 addr += bo->bdev->man[mem->mem_type].gpu_offset;
300 }
301 return addr;
302}
303
304/**
305 * amdgpu_find_mm_node - Helper function finds the drm_mm_node
306 * corresponding to @offset. It also modifies the offset to be
307 * within the drm_mm_node returned
308 */
309static struct drm_mm_node *amdgpu_find_mm_node(struct ttm_mem_reg *mem,
310 unsigned long *offset)
311{
312 struct drm_mm_node *mm_node = mem->mm_node;
313
314 while (*offset >= (mm_node->size << PAGE_SHIFT)) {
315 *offset -= (mm_node->size << PAGE_SHIFT);
316 ++mm_node;
317 }
318 return mm_node;
319}
320
321/**
322 * amdgpu_copy_ttm_mem_to_mem - Helper function for copy
323 *
324 * The function copies @size bytes from {src->mem + src->offset} to
325 * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a
326 * move and different for a BO to BO copy.
327 *
328 * @f: Returns the last fence if multiple jobs are submitted.
329 */
330int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev,
331 struct amdgpu_copy_mem *src,
332 struct amdgpu_copy_mem *dst,
333 uint64_t size,
334 struct reservation_object *resv,
335 struct dma_fence **f)
336{
337 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
338 struct drm_mm_node *src_mm, *dst_mm;
339 uint64_t src_node_start, dst_node_start, src_node_size,
340 dst_node_size, src_page_offset, dst_page_offset;
341 struct dma_fence *fence = NULL;
342 int r = 0;
343 const uint64_t GTT_MAX_BYTES = (AMDGPU_GTT_MAX_TRANSFER_SIZE *
344 AMDGPU_GPU_PAGE_SIZE);
345
346 if (!adev->mman.buffer_funcs_enabled) {
347 DRM_ERROR("Trying to move memory with ring turned off.\n");
348 return -EINVAL;
349 }
350
351 src_mm = amdgpu_find_mm_node(src->mem, &src->offset);
352 src_node_start = amdgpu_mm_node_addr(src->bo, src_mm, src->mem) +
353 src->offset;
354 src_node_size = (src_mm->size << PAGE_SHIFT) - src->offset;
355 src_page_offset = src_node_start & (PAGE_SIZE - 1);
356
357 dst_mm = amdgpu_find_mm_node(dst->mem, &dst->offset);
358 dst_node_start = amdgpu_mm_node_addr(dst->bo, dst_mm, dst->mem) +
359 dst->offset;
360 dst_node_size = (dst_mm->size << PAGE_SHIFT) - dst->offset;
361 dst_page_offset = dst_node_start & (PAGE_SIZE - 1);
362
363 mutex_lock(&adev->mman.gtt_window_lock);
364
365 while (size) {
366 unsigned long cur_size;
367 uint64_t from = src_node_start, to = dst_node_start;
368 struct dma_fence *next;
369
370 /* Copy size cannot exceed GTT_MAX_BYTES. So if src or dst
371 * begins at an offset, then adjust the size accordingly
372 */
373 cur_size = min3(min(src_node_size, dst_node_size), size,
374 GTT_MAX_BYTES);
375 if (cur_size + src_page_offset > GTT_MAX_BYTES ||
376 cur_size + dst_page_offset > GTT_MAX_BYTES)
377 cur_size -= max(src_page_offset, dst_page_offset);
378
379 /* Map only what needs to be accessed. Map src to window 0 and
380 * dst to window 1
381 */
382 if (src->mem->mem_type == TTM_PL_TT &&
383 !amdgpu_gtt_mgr_has_gart_addr(src->mem)) {
384 r = amdgpu_map_buffer(src->bo, src->mem,
385 PFN_UP(cur_size + src_page_offset),
386 src_node_start, 0, ring,
387 &from);
388 if (r)
389 goto error;
390 /* Adjust the offset because amdgpu_map_buffer returns
391 * start of mapped page
392 */
393 from += src_page_offset;
394 }
395
396 if (dst->mem->mem_type == TTM_PL_TT &&
397 !amdgpu_gtt_mgr_has_gart_addr(dst->mem)) {
398 r = amdgpu_map_buffer(dst->bo, dst->mem,
399 PFN_UP(cur_size + dst_page_offset),
400 dst_node_start, 1, ring,
401 &to);
402 if (r)
403 goto error;
404 to += dst_page_offset;
405 }
406
407 r = amdgpu_copy_buffer(ring, from, to, cur_size,
408 resv, &next, false, true);
409 if (r)
410 goto error;
411
412 dma_fence_put(fence);
413 fence = next;
414
415 size -= cur_size;
416 if (!size)
417 break;
418
419 src_node_size -= cur_size;
420 if (!src_node_size) {
421 src_node_start = amdgpu_mm_node_addr(src->bo, ++src_mm,
422 src->mem);
423 src_node_size = (src_mm->size << PAGE_SHIFT);
424 } else {
425 src_node_start += cur_size;
426 src_page_offset = src_node_start & (PAGE_SIZE - 1);
427 }
428 dst_node_size -= cur_size;
429 if (!dst_node_size) {
430 dst_node_start = amdgpu_mm_node_addr(dst->bo, ++dst_mm,
431 dst->mem);
432 dst_node_size = (dst_mm->size << PAGE_SHIFT);
433 } else {
434 dst_node_start += cur_size;
435 dst_page_offset = dst_node_start & (PAGE_SIZE - 1);
436 }
437 }
438error:
439 mutex_unlock(&adev->mman.gtt_window_lock);
440 if (f)
441 *f = dma_fence_get(fence);
442 dma_fence_put(fence);
443 return r;
444}
445
446
447static int amdgpu_move_blit(struct ttm_buffer_object *bo,
448 bool evict, bool no_wait_gpu,
449 struct ttm_mem_reg *new_mem,
450 struct ttm_mem_reg *old_mem)
451{
452 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
453 struct amdgpu_copy_mem src, dst;
454 struct dma_fence *fence = NULL;
455 int r;
456
457 src.bo = bo;
458 dst.bo = bo;
459 src.mem = old_mem;
460 dst.mem = new_mem;
461 src.offset = 0;
462 dst.offset = 0;
463
464 r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
465 new_mem->num_pages << PAGE_SHIFT,
466 bo->resv, &fence);
467 if (r)
468 goto error;
469
470 r = ttm_bo_pipeline_move(bo, fence, evict, new_mem);
471 dma_fence_put(fence);
472 return r;
473
474error:
475 if (fence)
476 dma_fence_wait(fence, false);
477 dma_fence_put(fence);
478 return r;
479}
480
481static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo, bool evict,
482 struct ttm_operation_ctx *ctx,
483 struct ttm_mem_reg *new_mem)
484{
485 struct amdgpu_device *adev;
486 struct ttm_mem_reg *old_mem = &bo->mem;
487 struct ttm_mem_reg tmp_mem;
488 struct ttm_place placements;
489 struct ttm_placement placement;
490 int r;
491
492 adev = amdgpu_ttm_adev(bo->bdev);
493 tmp_mem = *new_mem;
494 tmp_mem.mm_node = NULL;
495 placement.num_placement = 1;
496 placement.placement = &placements;
497 placement.num_busy_placement = 1;
498 placement.busy_placement = &placements;
499 placements.fpfn = 0;
500 placements.lpfn = 0;
501 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
502 r = ttm_bo_mem_space(bo, &placement, &tmp_mem, ctx);
503 if (unlikely(r)) {
504 return r;
505 }
506
507 r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
508 if (unlikely(r)) {
509 goto out_cleanup;
510 }
511
512 r = ttm_tt_bind(bo->ttm, &tmp_mem, ctx);
513 if (unlikely(r)) {
514 goto out_cleanup;
515 }
516 r = amdgpu_move_blit(bo, true, ctx->no_wait_gpu, &tmp_mem, old_mem);
517 if (unlikely(r)) {
518 goto out_cleanup;
519 }
520 r = ttm_bo_move_ttm(bo, ctx, new_mem);
521out_cleanup:
522 ttm_bo_mem_put(bo, &tmp_mem);
523 return r;
524}
525
526static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo, bool evict,
527 struct ttm_operation_ctx *ctx,
528 struct ttm_mem_reg *new_mem)
529{
530 struct amdgpu_device *adev;
531 struct ttm_mem_reg *old_mem = &bo->mem;
532 struct ttm_mem_reg tmp_mem;
533 struct ttm_placement placement;
534 struct ttm_place placements;
535 int r;
536
537 adev = amdgpu_ttm_adev(bo->bdev);
538 tmp_mem = *new_mem;
539 tmp_mem.mm_node = NULL;
540 placement.num_placement = 1;
541 placement.placement = &placements;
542 placement.num_busy_placement = 1;
543 placement.busy_placement = &placements;
544 placements.fpfn = 0;
545 placements.lpfn = 0;
546 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
547 r = ttm_bo_mem_space(bo, &placement, &tmp_mem, ctx);
548 if (unlikely(r)) {
549 return r;
550 }
551 r = ttm_bo_move_ttm(bo, ctx, &tmp_mem);
552 if (unlikely(r)) {
553 goto out_cleanup;
554 }
555 r = amdgpu_move_blit(bo, true, ctx->no_wait_gpu, new_mem, old_mem);
556 if (unlikely(r)) {
557 goto out_cleanup;
558 }
559out_cleanup:
560 ttm_bo_mem_put(bo, &tmp_mem);
561 return r;
562}
563
564static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
565 struct ttm_operation_ctx *ctx,
566 struct ttm_mem_reg *new_mem)
567{
568 struct amdgpu_device *adev;
569 struct amdgpu_bo *abo;
570 struct ttm_mem_reg *old_mem = &bo->mem;
571 int r;
572
573 /* Can't move a pinned BO */
574 abo = ttm_to_amdgpu_bo(bo);
575 if (WARN_ON_ONCE(abo->pin_count > 0))
576 return -EINVAL;
577
578 adev = amdgpu_ttm_adev(bo->bdev);
579
580 if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
581 amdgpu_move_null(bo, new_mem);
582 return 0;
583 }
584 if ((old_mem->mem_type == TTM_PL_TT &&
585 new_mem->mem_type == TTM_PL_SYSTEM) ||
586 (old_mem->mem_type == TTM_PL_SYSTEM &&
587 new_mem->mem_type == TTM_PL_TT)) {
588 /* bind is enough */
589 amdgpu_move_null(bo, new_mem);
590 return 0;
591 }
592
593 if (!adev->mman.buffer_funcs_enabled)
594 goto memcpy;
595
596 if (old_mem->mem_type == TTM_PL_VRAM &&
597 new_mem->mem_type == TTM_PL_SYSTEM) {
598 r = amdgpu_move_vram_ram(bo, evict, ctx, new_mem);
599 } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
600 new_mem->mem_type == TTM_PL_VRAM) {
601 r = amdgpu_move_ram_vram(bo, evict, ctx, new_mem);
602 } else {
603 r = amdgpu_move_blit(bo, evict, ctx->no_wait_gpu,
604 new_mem, old_mem);
605 }
606
607 if (r) {
608memcpy:
609 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
610 if (r) {
611 return r;
612 }
613 }
614
615 if (bo->type == ttm_bo_type_device &&
616 new_mem->mem_type == TTM_PL_VRAM &&
617 old_mem->mem_type != TTM_PL_VRAM) {
618 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
619 * accesses the BO after it's moved.
620 */
621 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
622 }
623
624 /* update statistics */
625 atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved);
626 return 0;
627}
628
629static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
630{
631 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
632 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
633 struct drm_mm_node *mm_node = mem->mm_node;
634
635 mem->bus.addr = NULL;
636 mem->bus.offset = 0;
637 mem->bus.size = mem->num_pages << PAGE_SHIFT;
638 mem->bus.base = 0;
639 mem->bus.is_iomem = false;
640 if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
641 return -EINVAL;
642 switch (mem->mem_type) {
643 case TTM_PL_SYSTEM:
644 /* system memory */
645 return 0;
646 case TTM_PL_TT:
647 break;
648 case TTM_PL_VRAM:
649 mem->bus.offset = mem->start << PAGE_SHIFT;
650 /* check if it's visible */
651 if ((mem->bus.offset + mem->bus.size) > adev->gmc.visible_vram_size)
652 return -EINVAL;
653 /* Only physically contiguous buffers apply. In a contiguous
654 * buffer, size of the first mm_node would match the number of
655 * pages in ttm_mem_reg.
656 */
657 if (adev->mman.aper_base_kaddr &&
658 (mm_node->size == mem->num_pages))
659 mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
660 mem->bus.offset;
661
662 mem->bus.base = adev->gmc.aper_base;
663 mem->bus.is_iomem = true;
664 break;
665 default:
666 return -EINVAL;
667 }
668 return 0;
669}
670
671static void amdgpu_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
672{
673}
674
675static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
676 unsigned long page_offset)
677{
678 struct drm_mm_node *mm;
679 unsigned long offset = (page_offset << PAGE_SHIFT);
680
681 mm = amdgpu_find_mm_node(&bo->mem, &offset);
682 return (bo->mem.bus.base >> PAGE_SHIFT) + mm->start +
683 (offset >> PAGE_SHIFT);
684}
685
686/*
687 * TTM backend functions.
688 */
689struct amdgpu_ttm_gup_task_list {
690 struct list_head list;
691 struct task_struct *task;
692};
693
694struct amdgpu_ttm_tt {
695 struct ttm_dma_tt ttm;
696 u64 offset;
697 uint64_t userptr;
698 struct mm_struct *usermm;
699 uint32_t userflags;
700 spinlock_t guptasklock;
701 struct list_head guptasks;
702 atomic_t mmu_invalidations;
703 uint32_t last_set_pages;
704};
705
706int amdgpu_ttm_tt_get_user_pages(struct ttm_tt *ttm, struct page **pages)
707{
708 struct amdgpu_ttm_tt *gtt = (void *)ttm;
709 unsigned int flags = 0;
710 unsigned pinned = 0;
711 int r;
712
713 if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
714 flags |= FOLL_WRITE;
715
716 down_read(¤t->mm->mmap_sem);
717
718 if (gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) {
719 /* check that we only use anonymous memory
720 to prevent problems with writeback */
721 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
722 struct vm_area_struct *vma;
723
724 vma = find_vma(gtt->usermm, gtt->userptr);
725 if (!vma || vma->vm_file || vma->vm_end < end) {
726 up_read(¤t->mm->mmap_sem);
727 return -EPERM;
728 }
729 }
730
731 do {
732 unsigned num_pages = ttm->num_pages - pinned;
733 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
734 struct page **p = pages + pinned;
735 struct amdgpu_ttm_gup_task_list guptask;
736
737 guptask.task = current;
738 spin_lock(>t->guptasklock);
739 list_add(&guptask.list, >t->guptasks);
740 spin_unlock(>t->guptasklock);
741
742 r = get_user_pages(userptr, num_pages, flags, p, NULL);
743
744 spin_lock(>t->guptasklock);
745 list_del(&guptask.list);
746 spin_unlock(>t->guptasklock);
747
748 if (r < 0)
749 goto release_pages;
750
751 pinned += r;
752
753 } while (pinned < ttm->num_pages);
754
755 up_read(¤t->mm->mmap_sem);
756 return 0;
757
758release_pages:
759 release_pages(pages, pinned);
760 up_read(¤t->mm->mmap_sem);
761 return r;
762}
763
764void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages)
765{
766 struct amdgpu_ttm_tt *gtt = (void *)ttm;
767 unsigned i;
768
769 gtt->last_set_pages = atomic_read(>t->mmu_invalidations);
770 for (i = 0; i < ttm->num_pages; ++i) {
771 if (ttm->pages[i])
772 put_page(ttm->pages[i]);
773
774 ttm->pages[i] = pages ? pages[i] : NULL;
775 }
776}
777
778void amdgpu_ttm_tt_mark_user_pages(struct ttm_tt *ttm)
779{
780 struct amdgpu_ttm_tt *gtt = (void *)ttm;
781 unsigned i;
782
783 for (i = 0; i < ttm->num_pages; ++i) {
784 struct page *page = ttm->pages[i];
785
786 if (!page)
787 continue;
788
789 if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
790 set_page_dirty(page);
791
792 mark_page_accessed(page);
793 }
794}
795
796/* prepare the sg table with the user pages */
797static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)
798{
799 struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
800 struct amdgpu_ttm_tt *gtt = (void *)ttm;
801 unsigned nents;
802 int r;
803
804 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
805 enum dma_data_direction direction = write ?
806 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
807
808 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
809 ttm->num_pages << PAGE_SHIFT,
810 GFP_KERNEL);
811 if (r)
812 goto release_sg;
813
814 r = -ENOMEM;
815 nents = dma_map_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
816 if (nents != ttm->sg->nents)
817 goto release_sg;
818
819 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
820 gtt->ttm.dma_address, ttm->num_pages);
821
822 return 0;
823
824release_sg:
825 kfree(ttm->sg);
826 return r;
827}
828
829static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
830{
831 struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
832 struct amdgpu_ttm_tt *gtt = (void *)ttm;
833
834 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
835 enum dma_data_direction direction = write ?
836 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
837
838 /* double check that we don't free the table twice */
839 if (!ttm->sg->sgl)
840 return;
841
842 /* free the sg table and pages again */
843 dma_unmap_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
844
845 amdgpu_ttm_tt_mark_user_pages(ttm);
846
847 sg_free_table(ttm->sg);
848}
849
850static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
851 struct ttm_mem_reg *bo_mem)
852{
853 struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
854 struct amdgpu_ttm_tt *gtt = (void*)ttm;
855 uint64_t flags;
856 int r = 0;
857
858 if (gtt->userptr) {
859 r = amdgpu_ttm_tt_pin_userptr(ttm);
860 if (r) {
861 DRM_ERROR("failed to pin userptr\n");
862 return r;
863 }
864 }
865 if (!ttm->num_pages) {
866 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
867 ttm->num_pages, bo_mem, ttm);
868 }
869
870 if (bo_mem->mem_type == AMDGPU_PL_GDS ||
871 bo_mem->mem_type == AMDGPU_PL_GWS ||
872 bo_mem->mem_type == AMDGPU_PL_OA)
873 return -EINVAL;
874
875 if (!amdgpu_gtt_mgr_has_gart_addr(bo_mem)) {
876 gtt->offset = AMDGPU_BO_INVALID_OFFSET;
877 return 0;
878 }
879
880 flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
881 gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
882 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
883 ttm->pages, gtt->ttm.dma_address, flags);
884
885 if (r)
886 DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
887 ttm->num_pages, gtt->offset);
888 return r;
889}
890
891int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
892{
893 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
894 struct ttm_operation_ctx ctx = { false, false };
895 struct amdgpu_ttm_tt *gtt = (void*)bo->ttm;
896 struct ttm_mem_reg tmp;
897 struct ttm_placement placement;
898 struct ttm_place placements;
899 uint64_t flags;
900 int r;
901
902 if (bo->mem.mem_type != TTM_PL_TT ||
903 amdgpu_gtt_mgr_has_gart_addr(&bo->mem))
904 return 0;
905
906 tmp = bo->mem;
907 tmp.mm_node = NULL;
908 placement.num_placement = 1;
909 placement.placement = &placements;
910 placement.num_busy_placement = 1;
911 placement.busy_placement = &placements;
912 placements.fpfn = 0;
913 placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
914 placements.flags = (bo->mem.placement & ~TTM_PL_MASK_MEM) |
915 TTM_PL_FLAG_TT;
916
917 r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx);
918 if (unlikely(r))
919 return r;
920
921 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, &tmp);
922 gtt->offset = (u64)tmp.start << PAGE_SHIFT;
923 r = amdgpu_gart_bind(adev, gtt->offset, bo->ttm->num_pages,
924 bo->ttm->pages, gtt->ttm.dma_address, flags);
925 if (unlikely(r)) {
926 ttm_bo_mem_put(bo, &tmp);
927 return r;
928 }
929
930 ttm_bo_mem_put(bo, &bo->mem);
931 bo->mem = tmp;
932 bo->offset = (bo->mem.start << PAGE_SHIFT) +
933 bo->bdev->man[bo->mem.mem_type].gpu_offset;
934
935 return 0;
936}
937
938int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
939{
940 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
941 struct amdgpu_ttm_tt *gtt = (void *)tbo->ttm;
942 uint64_t flags;
943 int r;
944
945 if (!gtt)
946 return 0;
947
948 flags = amdgpu_ttm_tt_pte_flags(adev, >t->ttm.ttm, &tbo->mem);
949 r = amdgpu_gart_bind(adev, gtt->offset, gtt->ttm.ttm.num_pages,
950 gtt->ttm.ttm.pages, gtt->ttm.dma_address, flags);
951 if (r)
952 DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
953 gtt->ttm.ttm.num_pages, gtt->offset);
954 return r;
955}
956
957static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
958{
959 struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
960 struct amdgpu_ttm_tt *gtt = (void *)ttm;
961 int r;
962
963 if (gtt->userptr)
964 amdgpu_ttm_tt_unpin_userptr(ttm);
965
966 if (gtt->offset == AMDGPU_BO_INVALID_OFFSET)
967 return 0;
968
969 /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
970 r = amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages);
971 if (r)
972 DRM_ERROR("failed to unbind %lu pages at 0x%08llX\n",
973 gtt->ttm.ttm.num_pages, gtt->offset);
974 return r;
975}
976
977static void amdgpu_ttm_backend_destroy(struct ttm_tt *ttm)
978{
979 struct amdgpu_ttm_tt *gtt = (void *)ttm;
980
981 ttm_dma_tt_fini(>t->ttm);
982 kfree(gtt);
983}
984
985static struct ttm_backend_func amdgpu_backend_func = {
986 .bind = &amdgpu_ttm_backend_bind,
987 .unbind = &amdgpu_ttm_backend_unbind,
988 .destroy = &amdgpu_ttm_backend_destroy,
989};
990
991static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
992 uint32_t page_flags)
993{
994 struct amdgpu_device *adev;
995 struct amdgpu_ttm_tt *gtt;
996
997 adev = amdgpu_ttm_adev(bo->bdev);
998
999 gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
1000 if (gtt == NULL) {
1001 return NULL;
1002 }
1003 gtt->ttm.ttm.func = &amdgpu_backend_func;
1004 if (ttm_sg_tt_init(>t->ttm, bo, page_flags)) {
1005 kfree(gtt);
1006 return NULL;
1007 }
1008 return >t->ttm.ttm;
1009}
1010
1011static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm,
1012 struct ttm_operation_ctx *ctx)
1013{
1014 struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
1015 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1016 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1017
1018 if (gtt && gtt->userptr) {
1019 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
1020 if (!ttm->sg)
1021 return -ENOMEM;
1022
1023 ttm->page_flags |= TTM_PAGE_FLAG_SG;
1024 ttm->state = tt_unbound;
1025 return 0;
1026 }
1027
1028 if (slave && ttm->sg) {
1029 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1030 gtt->ttm.dma_address,
1031 ttm->num_pages);
1032 ttm->state = tt_unbound;
1033 return 0;
1034 }
1035
1036#ifdef CONFIG_SWIOTLB
1037 if (adev->need_swiotlb && swiotlb_nr_tbl()) {
1038 return ttm_dma_populate(>t->ttm, adev->dev, ctx);
1039 }
1040#endif
1041
1042 return ttm_populate_and_map_pages(adev->dev, >t->ttm, ctx);
1043}
1044
1045static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm)
1046{
1047 struct amdgpu_device *adev;
1048 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1049 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1050
1051 if (gtt && gtt->userptr) {
1052 amdgpu_ttm_tt_set_user_pages(ttm, NULL);
1053 kfree(ttm->sg);
1054 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
1055 return;
1056 }
1057
1058 if (slave)
1059 return;
1060
1061 adev = amdgpu_ttm_adev(ttm->bdev);
1062
1063#ifdef CONFIG_SWIOTLB
1064 if (adev->need_swiotlb && swiotlb_nr_tbl()) {
1065 ttm_dma_unpopulate(>t->ttm, adev->dev);
1066 return;
1067 }
1068#endif
1069
1070 ttm_unmap_and_unpopulate_pages(adev->dev, >t->ttm);
1071}
1072
1073int amdgpu_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
1074 uint32_t flags)
1075{
1076 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1077
1078 if (gtt == NULL)
1079 return -EINVAL;
1080
1081 gtt->userptr = addr;
1082 gtt->usermm = current->mm;
1083 gtt->userflags = flags;
1084 spin_lock_init(>t->guptasklock);
1085 INIT_LIST_HEAD(>t->guptasks);
1086 atomic_set(>t->mmu_invalidations, 0);
1087 gtt->last_set_pages = 0;
1088
1089 return 0;
1090}
1091
1092struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1093{
1094 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1095
1096 if (gtt == NULL)
1097 return NULL;
1098
1099 return gtt->usermm;
1100}
1101
1102bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1103 unsigned long end)
1104{
1105 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1106 struct amdgpu_ttm_gup_task_list *entry;
1107 unsigned long size;
1108
1109 if (gtt == NULL || !gtt->userptr)
1110 return false;
1111
1112 size = (unsigned long)gtt->ttm.ttm.num_pages * PAGE_SIZE;
1113 if (gtt->userptr > end || gtt->userptr + size <= start)
1114 return false;
1115
1116 spin_lock(>t->guptasklock);
1117 list_for_each_entry(entry, >t->guptasks, list) {
1118 if (entry->task == current) {
1119 spin_unlock(>t->guptasklock);
1120 return false;
1121 }
1122 }
1123 spin_unlock(>t->guptasklock);
1124
1125 atomic_inc(>t->mmu_invalidations);
1126
1127 return true;
1128}
1129
1130bool amdgpu_ttm_tt_userptr_invalidated(struct ttm_tt *ttm,
1131 int *last_invalidated)
1132{
1133 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1134 int prev_invalidated = *last_invalidated;
1135
1136 *last_invalidated = atomic_read(>t->mmu_invalidations);
1137 return prev_invalidated != *last_invalidated;
1138}
1139
1140bool amdgpu_ttm_tt_userptr_needs_pages(struct ttm_tt *ttm)
1141{
1142 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1143
1144 if (gtt == NULL || !gtt->userptr)
1145 return false;
1146
1147 return atomic_read(>t->mmu_invalidations) != gtt->last_set_pages;
1148}
1149
1150bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1151{
1152 struct amdgpu_ttm_tt *gtt = (void *)ttm;
1153
1154 if (gtt == NULL)
1155 return false;
1156
1157 return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1158}
1159
1160uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1161 struct ttm_mem_reg *mem)
1162{
1163 uint64_t flags = 0;
1164
1165 if (mem && mem->mem_type != TTM_PL_SYSTEM)
1166 flags |= AMDGPU_PTE_VALID;
1167
1168 if (mem && mem->mem_type == TTM_PL_TT) {
1169 flags |= AMDGPU_PTE_SYSTEM;
1170
1171 if (ttm->caching_state == tt_cached)
1172 flags |= AMDGPU_PTE_SNOOPED;
1173 }
1174
1175 flags |= adev->gart.gart_pte_flags;
1176 flags |= AMDGPU_PTE_READABLE;
1177
1178 if (!amdgpu_ttm_tt_is_readonly(ttm))
1179 flags |= AMDGPU_PTE_WRITEABLE;
1180
1181 return flags;
1182}
1183
1184static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1185 const struct ttm_place *place)
1186{
1187 unsigned long num_pages = bo->mem.num_pages;
1188 struct drm_mm_node *node = bo->mem.mm_node;
1189 struct reservation_object_list *flist;
1190 struct dma_fence *f;
1191 int i;
1192
1193 /* If bo is a KFD BO, check if the bo belongs to the current process.
1194 * If true, then return false as any KFD process needs all its BOs to
1195 * be resident to run successfully
1196 */
1197 flist = reservation_object_get_list(bo->resv);
1198 if (flist) {
1199 for (i = 0; i < flist->shared_count; ++i) {
1200 f = rcu_dereference_protected(flist->shared[i],
1201 reservation_object_held(bo->resv));
1202 if (amdkfd_fence_check_mm(f, current->mm))
1203 return false;
1204 }
1205 }
1206
1207 switch (bo->mem.mem_type) {
1208 case TTM_PL_TT:
1209 return true;
1210
1211 case TTM_PL_VRAM:
1212 /* Check each drm MM node individually */
1213 while (num_pages) {
1214 if (place->fpfn < (node->start + node->size) &&
1215 !(place->lpfn && place->lpfn <= node->start))
1216 return true;
1217
1218 num_pages -= node->size;
1219 ++node;
1220 }
1221 return false;
1222
1223 default:
1224 break;
1225 }
1226
1227 return ttm_bo_eviction_valuable(bo, place);
1228}
1229
1230static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1231 unsigned long offset,
1232 void *buf, int len, int write)
1233{
1234 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1235 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1236 struct drm_mm_node *nodes;
1237 uint32_t value = 0;
1238 int ret = 0;
1239 uint64_t pos;
1240 unsigned long flags;
1241
1242 if (bo->mem.mem_type != TTM_PL_VRAM)
1243 return -EIO;
1244
1245 nodes = amdgpu_find_mm_node(&abo->tbo.mem, &offset);
1246 pos = (nodes->start << PAGE_SHIFT) + offset;
1247
1248 while (len && pos < adev->gmc.mc_vram_size) {
1249 uint64_t aligned_pos = pos & ~(uint64_t)3;
1250 uint32_t bytes = 4 - (pos & 3);
1251 uint32_t shift = (pos & 3) * 8;
1252 uint32_t mask = 0xffffffff << shift;
1253
1254 if (len < bytes) {
1255 mask &= 0xffffffff >> (bytes - len) * 8;
1256 bytes = len;
1257 }
1258
1259 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1260 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)aligned_pos) | 0x80000000);
1261 WREG32_NO_KIQ(mmMM_INDEX_HI, aligned_pos >> 31);
1262 if (!write || mask != 0xffffffff)
1263 value = RREG32_NO_KIQ(mmMM_DATA);
1264 if (write) {
1265 value &= ~mask;
1266 value |= (*(uint32_t *)buf << shift) & mask;
1267 WREG32_NO_KIQ(mmMM_DATA, value);
1268 }
1269 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1270 if (!write) {
1271 value = (value & mask) >> shift;
1272 memcpy(buf, &value, bytes);
1273 }
1274
1275 ret += bytes;
1276 buf = (uint8_t *)buf + bytes;
1277 pos += bytes;
1278 len -= bytes;
1279 if (pos >= (nodes->start + nodes->size) << PAGE_SHIFT) {
1280 ++nodes;
1281 pos = (nodes->start << PAGE_SHIFT);
1282 }
1283 }
1284
1285 return ret;
1286}
1287
1288static struct ttm_bo_driver amdgpu_bo_driver = {
1289 .ttm_tt_create = &amdgpu_ttm_tt_create,
1290 .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1291 .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1292 .invalidate_caches = &amdgpu_invalidate_caches,
1293 .init_mem_type = &amdgpu_init_mem_type,
1294 .eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1295 .evict_flags = &amdgpu_evict_flags,
1296 .move = &amdgpu_bo_move,
1297 .verify_access = &amdgpu_verify_access,
1298 .move_notify = &amdgpu_bo_move_notify,
1299 .fault_reserve_notify = &amdgpu_bo_fault_reserve_notify,
1300 .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1301 .io_mem_free = &amdgpu_ttm_io_mem_free,
1302 .io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1303 .access_memory = &amdgpu_ttm_access_memory
1304};
1305
1306/*
1307 * Firmware Reservation functions
1308 */
1309/**
1310 * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram
1311 *
1312 * @adev: amdgpu_device pointer
1313 *
1314 * free fw reserved vram if it has been reserved.
1315 */
1316static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
1317{
1318 amdgpu_bo_free_kernel(&adev->fw_vram_usage.reserved_bo,
1319 NULL, &adev->fw_vram_usage.va);
1320}
1321
1322/**
1323 * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
1324 *
1325 * @adev: amdgpu_device pointer
1326 *
1327 * create bo vram reservation from fw.
1328 */
1329static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
1330{
1331 struct ttm_operation_ctx ctx = { false, false };
1332 int r = 0;
1333 int i;
1334 u64 vram_size = adev->gmc.visible_vram_size;
1335 u64 offset = adev->fw_vram_usage.start_offset;
1336 u64 size = adev->fw_vram_usage.size;
1337 struct amdgpu_bo *bo;
1338
1339 adev->fw_vram_usage.va = NULL;
1340 adev->fw_vram_usage.reserved_bo = NULL;
1341
1342 if (adev->fw_vram_usage.size > 0 &&
1343 adev->fw_vram_usage.size <= vram_size) {
1344
1345 r = amdgpu_bo_create(adev, adev->fw_vram_usage.size, PAGE_SIZE,
1346 AMDGPU_GEM_DOMAIN_VRAM,
1347 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
1348 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
1349 ttm_bo_type_kernel, NULL,
1350 &adev->fw_vram_usage.reserved_bo);
1351 if (r)
1352 goto error_create;
1353
1354 r = amdgpu_bo_reserve(adev->fw_vram_usage.reserved_bo, false);
1355 if (r)
1356 goto error_reserve;
1357
1358 /* remove the original mem node and create a new one at the
1359 * request position
1360 */
1361 bo = adev->fw_vram_usage.reserved_bo;
1362 offset = ALIGN(offset, PAGE_SIZE);
1363 for (i = 0; i < bo->placement.num_placement; ++i) {
1364 bo->placements[i].fpfn = offset >> PAGE_SHIFT;
1365 bo->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
1366 }
1367
1368 ttm_bo_mem_put(&bo->tbo, &bo->tbo.mem);
1369 r = ttm_bo_mem_space(&bo->tbo, &bo->placement,
1370 &bo->tbo.mem, &ctx);
1371 if (r)
1372 goto error_pin;
1373
1374 r = amdgpu_bo_pin_restricted(adev->fw_vram_usage.reserved_bo,
1375 AMDGPU_GEM_DOMAIN_VRAM,
1376 adev->fw_vram_usage.start_offset,
1377 (adev->fw_vram_usage.start_offset +
1378 adev->fw_vram_usage.size), NULL);
1379 if (r)
1380 goto error_pin;
1381 r = amdgpu_bo_kmap(adev->fw_vram_usage.reserved_bo,
1382 &adev->fw_vram_usage.va);
1383 if (r)
1384 goto error_kmap;
1385
1386 amdgpu_bo_unreserve(adev->fw_vram_usage.reserved_bo);
1387 }
1388 return r;
1389
1390error_kmap:
1391 amdgpu_bo_unpin(adev->fw_vram_usage.reserved_bo);
1392error_pin:
1393 amdgpu_bo_unreserve(adev->fw_vram_usage.reserved_bo);
1394error_reserve:
1395 amdgpu_bo_unref(&adev->fw_vram_usage.reserved_bo);
1396error_create:
1397 adev->fw_vram_usage.va = NULL;
1398 adev->fw_vram_usage.reserved_bo = NULL;
1399 return r;
1400}
1401
1402int amdgpu_ttm_init(struct amdgpu_device *adev)
1403{
1404 uint64_t gtt_size;
1405 int r;
1406 u64 vis_vram_limit;
1407
1408 r = amdgpu_ttm_global_init(adev);
1409 if (r) {
1410 return r;
1411 }
1412 /* No others user of address space so set it to 0 */
1413 r = ttm_bo_device_init(&adev->mman.bdev,
1414 adev->mman.bo_global_ref.ref.object,
1415 &amdgpu_bo_driver,
1416 adev->ddev->anon_inode->i_mapping,
1417 DRM_FILE_PAGE_OFFSET,
1418 adev->need_dma32);
1419 if (r) {
1420 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1421 return r;
1422 }
1423 adev->mman.initialized = true;
1424
1425 /* We opt to avoid OOM on system pages allocations */
1426 adev->mman.bdev.no_retry = true;
1427
1428 r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
1429 adev->gmc.real_vram_size >> PAGE_SHIFT);
1430 if (r) {
1431 DRM_ERROR("Failed initializing VRAM heap.\n");
1432 return r;
1433 }
1434
1435 /* Reduce size of CPU-visible VRAM if requested */
1436 vis_vram_limit = (u64)amdgpu_vis_vram_limit * 1024 * 1024;
1437 if (amdgpu_vis_vram_limit > 0 &&
1438 vis_vram_limit <= adev->gmc.visible_vram_size)
1439 adev->gmc.visible_vram_size = vis_vram_limit;
1440
1441 /* Change the size here instead of the init above so only lpfn is affected */
1442 amdgpu_ttm_set_buffer_funcs_status(adev, false);
1443#ifdef CONFIG_64BIT
1444 adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
1445 adev->gmc.visible_vram_size);
1446#endif
1447
1448 /*
1449 *The reserved vram for firmware must be pinned to the specified
1450 *place on the VRAM, so reserve it early.
1451 */
1452 r = amdgpu_ttm_fw_reserve_vram_init(adev);
1453 if (r) {
1454 return r;
1455 }
1456
1457 r = amdgpu_bo_create_kernel(adev, adev->gmc.stolen_size, PAGE_SIZE,
1458 AMDGPU_GEM_DOMAIN_VRAM,
1459 &adev->stolen_vga_memory,
1460 NULL, NULL);
1461 if (r)
1462 return r;
1463 DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1464 (unsigned) (adev->gmc.real_vram_size / (1024 * 1024)));
1465
1466 if (amdgpu_gtt_size == -1) {
1467 struct sysinfo si;
1468
1469 si_meminfo(&si);
1470 gtt_size = min(max((AMDGPU_DEFAULT_GTT_SIZE_MB << 20),
1471 adev->gmc.mc_vram_size),
1472 ((uint64_t)si.totalram * si.mem_unit * 3/4));
1473 }
1474 else
1475 gtt_size = (uint64_t)amdgpu_gtt_size << 20;
1476 r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_TT, gtt_size >> PAGE_SHIFT);
1477 if (r) {
1478 DRM_ERROR("Failed initializing GTT heap.\n");
1479 return r;
1480 }
1481 DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1482 (unsigned)(gtt_size / (1024 * 1024)));
1483
1484 adev->gds.mem.total_size = adev->gds.mem.total_size << AMDGPU_GDS_SHIFT;
1485 adev->gds.mem.gfx_partition_size = adev->gds.mem.gfx_partition_size << AMDGPU_GDS_SHIFT;
1486 adev->gds.mem.cs_partition_size = adev->gds.mem.cs_partition_size << AMDGPU_GDS_SHIFT;
1487 adev->gds.gws.total_size = adev->gds.gws.total_size << AMDGPU_GWS_SHIFT;
1488 adev->gds.gws.gfx_partition_size = adev->gds.gws.gfx_partition_size << AMDGPU_GWS_SHIFT;
1489 adev->gds.gws.cs_partition_size = adev->gds.gws.cs_partition_size << AMDGPU_GWS_SHIFT;
1490 adev->gds.oa.total_size = adev->gds.oa.total_size << AMDGPU_OA_SHIFT;
1491 adev->gds.oa.gfx_partition_size = adev->gds.oa.gfx_partition_size << AMDGPU_OA_SHIFT;
1492 adev->gds.oa.cs_partition_size = adev->gds.oa.cs_partition_size << AMDGPU_OA_SHIFT;
1493 /* GDS Memory */
1494 if (adev->gds.mem.total_size) {
1495 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GDS,
1496 adev->gds.mem.total_size >> PAGE_SHIFT);
1497 if (r) {
1498 DRM_ERROR("Failed initializing GDS heap.\n");
1499 return r;
1500 }
1501 }
1502
1503 /* GWS */
1504 if (adev->gds.gws.total_size) {
1505 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GWS,
1506 adev->gds.gws.total_size >> PAGE_SHIFT);
1507 if (r) {
1508 DRM_ERROR("Failed initializing gws heap.\n");
1509 return r;
1510 }
1511 }
1512
1513 /* OA */
1514 if (adev->gds.oa.total_size) {
1515 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_OA,
1516 adev->gds.oa.total_size >> PAGE_SHIFT);
1517 if (r) {
1518 DRM_ERROR("Failed initializing oa heap.\n");
1519 return r;
1520 }
1521 }
1522
1523 r = amdgpu_ttm_debugfs_init(adev);
1524 if (r) {
1525 DRM_ERROR("Failed to init debugfs\n");
1526 return r;
1527 }
1528 return 0;
1529}
1530
1531void amdgpu_ttm_fini(struct amdgpu_device *adev)
1532{
1533 if (!adev->mman.initialized)
1534 return;
1535
1536 amdgpu_ttm_debugfs_fini(adev);
1537 amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, NULL);
1538 amdgpu_ttm_fw_reserve_vram_fini(adev);
1539 if (adev->mman.aper_base_kaddr)
1540 iounmap(adev->mman.aper_base_kaddr);
1541 adev->mman.aper_base_kaddr = NULL;
1542
1543 ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM);
1544 ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT);
1545 if (adev->gds.mem.total_size)
1546 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GDS);
1547 if (adev->gds.gws.total_size)
1548 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GWS);
1549 if (adev->gds.oa.total_size)
1550 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_OA);
1551 ttm_bo_device_release(&adev->mman.bdev);
1552 amdgpu_ttm_global_fini(adev);
1553 adev->mman.initialized = false;
1554 DRM_INFO("amdgpu: ttm finalized\n");
1555}
1556
1557/**
1558 * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions
1559 *
1560 * @adev: amdgpu_device pointer
1561 * @enable: true when we can use buffer functions.
1562 *
1563 * Enable/disable use of buffer functions during suspend/resume. This should
1564 * only be called at bootup or when userspace isn't running.
1565 */
1566void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable)
1567{
1568 struct ttm_mem_type_manager *man = &adev->mman.bdev.man[TTM_PL_VRAM];
1569 uint64_t size;
1570
1571 if (!adev->mman.initialized || adev->in_gpu_reset)
1572 return;
1573
1574 /* this just adjusts TTM size idea, which sets lpfn to the correct value */
1575 if (enable)
1576 size = adev->gmc.real_vram_size;
1577 else
1578 size = adev->gmc.visible_vram_size;
1579 man->size = size >> PAGE_SHIFT;
1580 adev->mman.buffer_funcs_enabled = enable;
1581}
1582
1583int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
1584{
1585 struct drm_file *file_priv;
1586 struct amdgpu_device *adev;
1587
1588 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET))
1589 return -EINVAL;
1590
1591 file_priv = filp->private_data;
1592 adev = file_priv->minor->dev->dev_private;
1593 if (adev == NULL)
1594 return -EINVAL;
1595
1596 return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
1597}
1598
1599static int amdgpu_map_buffer(struct ttm_buffer_object *bo,
1600 struct ttm_mem_reg *mem, unsigned num_pages,
1601 uint64_t offset, unsigned window,
1602 struct amdgpu_ring *ring,
1603 uint64_t *addr)
1604{
1605 struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
1606 struct amdgpu_device *adev = ring->adev;
1607 struct ttm_tt *ttm = bo->ttm;
1608 struct amdgpu_job *job;
1609 unsigned num_dw, num_bytes;
1610 dma_addr_t *dma_address;
1611 struct dma_fence *fence;
1612 uint64_t src_addr, dst_addr;
1613 uint64_t flags;
1614 int r;
1615
1616 BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
1617 AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
1618
1619 *addr = adev->gmc.gart_start;
1620 *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE *
1621 AMDGPU_GPU_PAGE_SIZE;
1622
1623 num_dw = adev->mman.buffer_funcs->copy_num_dw;
1624 while (num_dw & 0x7)
1625 num_dw++;
1626
1627 num_bytes = num_pages * 8;
1628
1629 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes, &job);
1630 if (r)
1631 return r;
1632
1633 src_addr = num_dw * 4;
1634 src_addr += job->ibs[0].gpu_addr;
1635
1636 dst_addr = adev->gart.table_addr;
1637 dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
1638 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
1639 dst_addr, num_bytes);
1640
1641 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1642 WARN_ON(job->ibs[0].length_dw > num_dw);
1643
1644 dma_address = >t->ttm.dma_address[offset >> PAGE_SHIFT];
1645 flags = amdgpu_ttm_tt_pte_flags(adev, ttm, mem);
1646 r = amdgpu_gart_map(adev, 0, num_pages, dma_address, flags,
1647 &job->ibs[0].ptr[num_dw]);
1648 if (r)
1649 goto error_free;
1650
1651 r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1652 AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
1653 if (r)
1654 goto error_free;
1655
1656 dma_fence_put(fence);
1657
1658 return r;
1659
1660error_free:
1661 amdgpu_job_free(job);
1662 return r;
1663}
1664
1665int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
1666 uint64_t dst_offset, uint32_t byte_count,
1667 struct reservation_object *resv,
1668 struct dma_fence **fence, bool direct_submit,
1669 bool vm_needs_flush)
1670{
1671 struct amdgpu_device *adev = ring->adev;
1672 struct amdgpu_job *job;
1673
1674 uint32_t max_bytes;
1675 unsigned num_loops, num_dw;
1676 unsigned i;
1677 int r;
1678
1679 if (direct_submit && !ring->ready) {
1680 DRM_ERROR("Trying to move memory with ring turned off.\n");
1681 return -EINVAL;
1682 }
1683
1684 max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
1685 num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1686 num_dw = num_loops * adev->mman.buffer_funcs->copy_num_dw;
1687
1688 /* for IB padding */
1689 while (num_dw & 0x7)
1690 num_dw++;
1691
1692 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1693 if (r)
1694 return r;
1695
1696 job->vm_needs_flush = vm_needs_flush;
1697 if (resv) {
1698 r = amdgpu_sync_resv(adev, &job->sync, resv,
1699 AMDGPU_FENCE_OWNER_UNDEFINED,
1700 false);
1701 if (r) {
1702 DRM_ERROR("sync failed (%d).\n", r);
1703 goto error_free;
1704 }
1705 }
1706
1707 for (i = 0; i < num_loops; i++) {
1708 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1709
1710 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
1711 dst_offset, cur_size_in_bytes);
1712
1713 src_offset += cur_size_in_bytes;
1714 dst_offset += cur_size_in_bytes;
1715 byte_count -= cur_size_in_bytes;
1716 }
1717
1718 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1719 WARN_ON(job->ibs[0].length_dw > num_dw);
1720 if (direct_submit) {
1721 r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
1722 NULL, fence);
1723 job->fence = dma_fence_get(*fence);
1724 if (r)
1725 DRM_ERROR("Error scheduling IBs (%d)\n", r);
1726 amdgpu_job_free(job);
1727 } else {
1728 r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1729 AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1730 if (r)
1731 goto error_free;
1732 }
1733
1734 return r;
1735
1736error_free:
1737 amdgpu_job_free(job);
1738 return r;
1739}
1740
1741int amdgpu_fill_buffer(struct amdgpu_bo *bo,
1742 uint32_t src_data,
1743 struct reservation_object *resv,
1744 struct dma_fence **fence)
1745{
1746 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1747 uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
1748 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
1749
1750 struct drm_mm_node *mm_node;
1751 unsigned long num_pages;
1752 unsigned int num_loops, num_dw;
1753
1754 struct amdgpu_job *job;
1755 int r;
1756
1757 if (!adev->mman.buffer_funcs_enabled) {
1758 DRM_ERROR("Trying to clear memory with ring turned off.\n");
1759 return -EINVAL;
1760 }
1761
1762 if (bo->tbo.mem.mem_type == TTM_PL_TT) {
1763 r = amdgpu_ttm_alloc_gart(&bo->tbo);
1764 if (r)
1765 return r;
1766 }
1767
1768 num_pages = bo->tbo.num_pages;
1769 mm_node = bo->tbo.mem.mm_node;
1770 num_loops = 0;
1771 while (num_pages) {
1772 uint32_t byte_count = mm_node->size << PAGE_SHIFT;
1773
1774 num_loops += DIV_ROUND_UP(byte_count, max_bytes);
1775 num_pages -= mm_node->size;
1776 ++mm_node;
1777 }
1778 num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw;
1779
1780 /* for IB padding */
1781 num_dw += 64;
1782
1783 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1784 if (r)
1785 return r;
1786
1787 if (resv) {
1788 r = amdgpu_sync_resv(adev, &job->sync, resv,
1789 AMDGPU_FENCE_OWNER_UNDEFINED, false);
1790 if (r) {
1791 DRM_ERROR("sync failed (%d).\n", r);
1792 goto error_free;
1793 }
1794 }
1795
1796 num_pages = bo->tbo.num_pages;
1797 mm_node = bo->tbo.mem.mm_node;
1798
1799 while (num_pages) {
1800 uint32_t byte_count = mm_node->size << PAGE_SHIFT;
1801 uint64_t dst_addr;
1802
1803 dst_addr = amdgpu_mm_node_addr(&bo->tbo, mm_node, &bo->tbo.mem);
1804 while (byte_count) {
1805 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1806
1807 amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data,
1808 dst_addr, cur_size_in_bytes);
1809
1810 dst_addr += cur_size_in_bytes;
1811 byte_count -= cur_size_in_bytes;
1812 }
1813
1814 num_pages -= mm_node->size;
1815 ++mm_node;
1816 }
1817
1818 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1819 WARN_ON(job->ibs[0].length_dw > num_dw);
1820 r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1821 AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1822 if (r)
1823 goto error_free;
1824
1825 return 0;
1826
1827error_free:
1828 amdgpu_job_free(job);
1829 return r;
1830}
1831
1832#if defined(CONFIG_DEBUG_FS)
1833
1834static int amdgpu_mm_dump_table(struct seq_file *m, void *data)
1835{
1836 struct drm_info_node *node = (struct drm_info_node *)m->private;
1837 unsigned ttm_pl = *(int *)node->info_ent->data;
1838 struct drm_device *dev = node->minor->dev;
1839 struct amdgpu_device *adev = dev->dev_private;
1840 struct ttm_mem_type_manager *man = &adev->mman.bdev.man[ttm_pl];
1841 struct drm_printer p = drm_seq_file_printer(m);
1842
1843 man->func->debug(man, &p);
1844 return 0;
1845}
1846
1847static int ttm_pl_vram = TTM_PL_VRAM;
1848static int ttm_pl_tt = TTM_PL_TT;
1849
1850static const struct drm_info_list amdgpu_ttm_debugfs_list[] = {
1851 {"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, &ttm_pl_vram},
1852 {"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, &ttm_pl_tt},
1853 {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1854#ifdef CONFIG_SWIOTLB
1855 {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1856#endif
1857};
1858
1859static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
1860 size_t size, loff_t *pos)
1861{
1862 struct amdgpu_device *adev = file_inode(f)->i_private;
1863 ssize_t result = 0;
1864 int r;
1865
1866 if (size & 0x3 || *pos & 0x3)
1867 return -EINVAL;
1868
1869 if (*pos >= adev->gmc.mc_vram_size)
1870 return -ENXIO;
1871
1872 while (size) {
1873 unsigned long flags;
1874 uint32_t value;
1875
1876 if (*pos >= adev->gmc.mc_vram_size)
1877 return result;
1878
1879 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1880 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
1881 WREG32_NO_KIQ(mmMM_INDEX_HI, *pos >> 31);
1882 value = RREG32_NO_KIQ(mmMM_DATA);
1883 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1884
1885 r = put_user(value, (uint32_t *)buf);
1886 if (r)
1887 return r;
1888
1889 result += 4;
1890 buf += 4;
1891 *pos += 4;
1892 size -= 4;
1893 }
1894
1895 return result;
1896}
1897
1898static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf,
1899 size_t size, loff_t *pos)
1900{
1901 struct amdgpu_device *adev = file_inode(f)->i_private;
1902 ssize_t result = 0;
1903 int r;
1904
1905 if (size & 0x3 || *pos & 0x3)
1906 return -EINVAL;
1907
1908 if (*pos >= adev->gmc.mc_vram_size)
1909 return -ENXIO;
1910
1911 while (size) {
1912 unsigned long flags;
1913 uint32_t value;
1914
1915 if (*pos >= adev->gmc.mc_vram_size)
1916 return result;
1917
1918 r = get_user(value, (uint32_t *)buf);
1919 if (r)
1920 return r;
1921
1922 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1923 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
1924 WREG32_NO_KIQ(mmMM_INDEX_HI, *pos >> 31);
1925 WREG32_NO_KIQ(mmMM_DATA, value);
1926 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1927
1928 result += 4;
1929 buf += 4;
1930 *pos += 4;
1931 size -= 4;
1932 }
1933
1934 return result;
1935}
1936
1937static const struct file_operations amdgpu_ttm_vram_fops = {
1938 .owner = THIS_MODULE,
1939 .read = amdgpu_ttm_vram_read,
1940 .write = amdgpu_ttm_vram_write,
1941 .llseek = default_llseek,
1942};
1943
1944#ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
1945
1946static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
1947 size_t size, loff_t *pos)
1948{
1949 struct amdgpu_device *adev = file_inode(f)->i_private;
1950 ssize_t result = 0;
1951 int r;
1952
1953 while (size) {
1954 loff_t p = *pos / PAGE_SIZE;
1955 unsigned off = *pos & ~PAGE_MASK;
1956 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1957 struct page *page;
1958 void *ptr;
1959
1960 if (p >= adev->gart.num_cpu_pages)
1961 return result;
1962
1963 page = adev->gart.pages[p];
1964 if (page) {
1965 ptr = kmap(page);
1966 ptr += off;
1967
1968 r = copy_to_user(buf, ptr, cur_size);
1969 kunmap(adev->gart.pages[p]);
1970 } else
1971 r = clear_user(buf, cur_size);
1972
1973 if (r)
1974 return -EFAULT;
1975
1976 result += cur_size;
1977 buf += cur_size;
1978 *pos += cur_size;
1979 size -= cur_size;
1980 }
1981
1982 return result;
1983}
1984
1985static const struct file_operations amdgpu_ttm_gtt_fops = {
1986 .owner = THIS_MODULE,
1987 .read = amdgpu_ttm_gtt_read,
1988 .llseek = default_llseek
1989};
1990
1991#endif
1992
1993static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
1994 size_t size, loff_t *pos)
1995{
1996 struct amdgpu_device *adev = file_inode(f)->i_private;
1997 struct iommu_domain *dom;
1998 ssize_t result = 0;
1999 int r;
2000
2001 dom = iommu_get_domain_for_dev(adev->dev);
2002
2003 while (size) {
2004 phys_addr_t addr = *pos & PAGE_MASK;
2005 loff_t off = *pos & ~PAGE_MASK;
2006 size_t bytes = PAGE_SIZE - off;
2007 unsigned long pfn;
2008 struct page *p;
2009 void *ptr;
2010
2011 bytes = bytes < size ? bytes : size;
2012
2013 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2014
2015 pfn = addr >> PAGE_SHIFT;
2016 if (!pfn_valid(pfn))
2017 return -EPERM;
2018
2019 p = pfn_to_page(pfn);
2020 if (p->mapping != adev->mman.bdev.dev_mapping)
2021 return -EPERM;
2022
2023 ptr = kmap(p);
2024 r = copy_to_user(buf, ptr + off, bytes);
2025 kunmap(p);
2026 if (r)
2027 return -EFAULT;
2028
2029 size -= bytes;
2030 *pos += bytes;
2031 result += bytes;
2032 }
2033
2034 return result;
2035}
2036
2037static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
2038 size_t size, loff_t *pos)
2039{
2040 struct amdgpu_device *adev = file_inode(f)->i_private;
2041 struct iommu_domain *dom;
2042 ssize_t result = 0;
2043 int r;
2044
2045 dom = iommu_get_domain_for_dev(adev->dev);
2046
2047 while (size) {
2048 phys_addr_t addr = *pos & PAGE_MASK;
2049 loff_t off = *pos & ~PAGE_MASK;
2050 size_t bytes = PAGE_SIZE - off;
2051 unsigned long pfn;
2052 struct page *p;
2053 void *ptr;
2054
2055 bytes = bytes < size ? bytes : size;
2056
2057 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2058
2059 pfn = addr >> PAGE_SHIFT;
2060 if (!pfn_valid(pfn))
2061 return -EPERM;
2062
2063 p = pfn_to_page(pfn);
2064 if (p->mapping != adev->mman.bdev.dev_mapping)
2065 return -EPERM;
2066
2067 ptr = kmap(p);
2068 r = copy_from_user(ptr + off, buf, bytes);
2069 kunmap(p);
2070 if (r)
2071 return -EFAULT;
2072
2073 size -= bytes;
2074 *pos += bytes;
2075 result += bytes;
2076 }
2077
2078 return result;
2079}
2080
2081static const struct file_operations amdgpu_ttm_iomem_fops = {
2082 .owner = THIS_MODULE,
2083 .read = amdgpu_iomem_read,
2084 .write = amdgpu_iomem_write,
2085 .llseek = default_llseek
2086};
2087
2088static const struct {
2089 char *name;
2090 const struct file_operations *fops;
2091 int domain;
2092} ttm_debugfs_entries[] = {
2093 { "amdgpu_vram", &amdgpu_ttm_vram_fops, TTM_PL_VRAM },
2094#ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
2095 { "amdgpu_gtt", &amdgpu_ttm_gtt_fops, TTM_PL_TT },
2096#endif
2097 { "amdgpu_iomem", &amdgpu_ttm_iomem_fops, TTM_PL_SYSTEM },
2098};
2099
2100#endif
2101
2102static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
2103{
2104#if defined(CONFIG_DEBUG_FS)
2105 unsigned count;
2106
2107 struct drm_minor *minor = adev->ddev->primary;
2108 struct dentry *ent, *root = minor->debugfs_root;
2109
2110 for (count = 0; count < ARRAY_SIZE(ttm_debugfs_entries); count++) {
2111 ent = debugfs_create_file(
2112 ttm_debugfs_entries[count].name,
2113 S_IFREG | S_IRUGO, root,
2114 adev,
2115 ttm_debugfs_entries[count].fops);
2116 if (IS_ERR(ent))
2117 return PTR_ERR(ent);
2118 if (ttm_debugfs_entries[count].domain == TTM_PL_VRAM)
2119 i_size_write(ent->d_inode, adev->gmc.mc_vram_size);
2120 else if (ttm_debugfs_entries[count].domain == TTM_PL_TT)
2121 i_size_write(ent->d_inode, adev->gmc.gart_size);
2122 adev->mman.debugfs_entries[count] = ent;
2123 }
2124
2125 count = ARRAY_SIZE(amdgpu_ttm_debugfs_list);
2126
2127#ifdef CONFIG_SWIOTLB
2128 if (!(adev->need_swiotlb && swiotlb_nr_tbl()))
2129 --count;
2130#endif
2131
2132 return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count);
2133#else
2134 return 0;
2135#endif
2136}
2137
2138static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev)
2139{
2140#if defined(CONFIG_DEBUG_FS)
2141 unsigned i;
2142
2143 for (i = 0; i < ARRAY_SIZE(ttm_debugfs_entries); i++)
2144 debugfs_remove(adev->mman.debugfs_entries[i]);
2145#endif
2146}