1 /*
2 * Copyright 2007 Dave Airlied
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 "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * VA LINUX SYSTEMS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
23 */
24 /*
25 * Authors: Dave Airlied <[email protected]>
26 * Ben Skeggs <[email protected]>
27 * Jeremy Kolb <[email protected]>
28 */
29
30 #include <linux/dma-mapping.h>
31 #include <drm/ttm/ttm_tt.h>
32
33 #include "nouveau_drv.h"
34 #include "nouveau_chan.h"
35 #include "nouveau_fence.h"
36
37 #include "nouveau_bo.h"
38 #include "nouveau_ttm.h"
39 #include "nouveau_gem.h"
40 #include "nouveau_mem.h"
41 #include "nouveau_vmm.h"
42
43 #include <nvif/class.h>
44 #include <nvif/if500b.h>
45 #include <nvif/if900b.h>
46
47 static int nouveau_ttm_tt_bind(struct ttm_device *bdev, struct ttm_tt *ttm,
48 struct ttm_resource *reg);
49 static void nouveau_ttm_tt_unbind(struct ttm_device *bdev, struct ttm_tt *ttm);
50
51 /*
52 * NV10-NV40 tiling helpers
53 */
54
55 static void
nv10_bo_update_tile_region(struct drm_device * dev,struct nouveau_drm_tile * reg,u32 addr,u32 size,u32 pitch,u32 flags)56 nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
57 u32 addr, u32 size, u32 pitch, u32 flags)
58 {
59 struct nouveau_drm *drm = nouveau_drm(dev);
60 int i = reg - drm->tile.reg;
61 struct nvkm_fb *fb = nvxx_fb(drm);
62 struct nvkm_fb_tile *tile = &fb->tile.region[i];
63
64 nouveau_fence_unref(®->fence);
65
66 if (tile->pitch)
67 nvkm_fb_tile_fini(fb, i, tile);
68
69 if (pitch)
70 nvkm_fb_tile_init(fb, i, addr, size, pitch, flags, tile);
71
72 nvkm_fb_tile_prog(fb, i, tile);
73 }
74
75 static struct nouveau_drm_tile *
nv10_bo_get_tile_region(struct drm_device * dev,int i)76 nv10_bo_get_tile_region(struct drm_device *dev, int i)
77 {
78 struct nouveau_drm *drm = nouveau_drm(dev);
79 struct nouveau_drm_tile *tile = &drm->tile.reg[i];
80
81 spin_lock(&drm->tile.lock);
82
83 if (!tile->used &&
84 (!tile->fence || nouveau_fence_done(tile->fence)))
85 tile->used = true;
86 else
87 tile = NULL;
88
89 spin_unlock(&drm->tile.lock);
90 return tile;
91 }
92
93 static void
nv10_bo_put_tile_region(struct drm_device * dev,struct nouveau_drm_tile * tile,struct dma_fence * fence)94 nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
95 struct dma_fence *fence)
96 {
97 struct nouveau_drm *drm = nouveau_drm(dev);
98
99 if (tile) {
100 spin_lock(&drm->tile.lock);
101 tile->fence = (struct nouveau_fence *)dma_fence_get(fence);
102 tile->used = false;
103 spin_unlock(&drm->tile.lock);
104 }
105 }
106
107 static struct nouveau_drm_tile *
nv10_bo_set_tiling(struct drm_device * dev,u32 addr,u32 size,u32 pitch,u32 zeta)108 nv10_bo_set_tiling(struct drm_device *dev, u32 addr,
109 u32 size, u32 pitch, u32 zeta)
110 {
111 struct nouveau_drm *drm = nouveau_drm(dev);
112 struct nvkm_fb *fb = nvxx_fb(drm);
113 struct nouveau_drm_tile *tile, *found = NULL;
114 int i;
115
116 for (i = 0; i < fb->tile.regions; i++) {
117 tile = nv10_bo_get_tile_region(dev, i);
118
119 if (pitch && !found) {
120 found = tile;
121 continue;
122
123 } else if (tile && fb->tile.region[i].pitch) {
124 /* Kill an unused tile region. */
125 nv10_bo_update_tile_region(dev, tile, 0, 0, 0, 0);
126 }
127
128 nv10_bo_put_tile_region(dev, tile, NULL);
129 }
130
131 if (found)
132 nv10_bo_update_tile_region(dev, found, addr, size, pitch, zeta);
133 return found;
134 }
135
136 static void
nouveau_bo_del_ttm(struct ttm_buffer_object * bo)137 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
138 {
139 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
140 struct drm_device *dev = drm->dev;
141 struct nouveau_bo *nvbo = nouveau_bo(bo);
142
143 WARN_ON(nvbo->bo.pin_count > 0);
144 nouveau_bo_del_io_reserve_lru(bo);
145 nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
146
147 if (bo->base.import_attach)
148 drm_prime_gem_destroy(&bo->base, bo->sg);
149
150 /*
151 * If nouveau_bo_new() allocated this buffer, the GEM object was never
152 * initialized, so don't attempt to release it.
153 */
154 if (bo->base.dev) {
155 /* Gem objects not being shared with other VMs get their
156 * dma_resv from a root GEM object.
157 */
158 if (nvbo->no_share)
159 drm_gem_object_put(nvbo->r_obj);
160
161 drm_gem_object_release(&bo->base);
162 } else {
163 dma_resv_fini(&bo->base._resv);
164 }
165
166 kfree(nvbo);
167 }
168
169 static inline u64
roundup_64(u64 x,u32 y)170 roundup_64(u64 x, u32 y)
171 {
172 x += y - 1;
173 do_div(x, y);
174 return x * y;
175 }
176
177 static void
nouveau_bo_fixup_align(struct nouveau_bo * nvbo,int * align,u64 * size)178 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, int *align, u64 *size)
179 {
180 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
181 struct nvif_device *device = &drm->client.device;
182
183 if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
184 if (nvbo->mode) {
185 if (device->info.chipset >= 0x40) {
186 *align = 65536;
187 *size = roundup_64(*size, 64 * nvbo->mode);
188
189 } else if (device->info.chipset >= 0x30) {
190 *align = 32768;
191 *size = roundup_64(*size, 64 * nvbo->mode);
192
193 } else if (device->info.chipset >= 0x20) {
194 *align = 16384;
195 *size = roundup_64(*size, 64 * nvbo->mode);
196
197 } else if (device->info.chipset >= 0x10) {
198 *align = 16384;
199 *size = roundup_64(*size, 32 * nvbo->mode);
200 }
201 }
202 } else {
203 *size = roundup_64(*size, (1 << nvbo->page));
204 *align = max((1 << nvbo->page), *align);
205 }
206
207 *size = roundup_64(*size, PAGE_SIZE);
208 }
209
210 struct nouveau_bo *
nouveau_bo_alloc(struct nouveau_cli * cli,u64 * size,int * align,u32 domain,u32 tile_mode,u32 tile_flags,bool internal)211 nouveau_bo_alloc(struct nouveau_cli *cli, u64 *size, int *align, u32 domain,
212 u32 tile_mode, u32 tile_flags, bool internal)
213 {
214 struct nouveau_drm *drm = cli->drm;
215 struct nouveau_bo *nvbo;
216 struct nvif_mmu *mmu = &cli->mmu;
217 struct nvif_vmm *vmm = &nouveau_cli_vmm(cli)->vmm;
218 int i, pi = -1;
219
220 if (!*size) {
221 NV_WARN(drm, "skipped size %016llx\n", *size);
222 return ERR_PTR(-EINVAL);
223 }
224
225 nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
226 if (!nvbo)
227 return ERR_PTR(-ENOMEM);
228
229 INIT_LIST_HEAD(&nvbo->head);
230 INIT_LIST_HEAD(&nvbo->entry);
231 INIT_LIST_HEAD(&nvbo->vma_list);
232 nvbo->bo.bdev = &drm->ttm.bdev;
233
234 /* This is confusing, and doesn't actually mean we want an uncached
235 * mapping, but is what NOUVEAU_GEM_DOMAIN_COHERENT gets translated
236 * into in nouveau_gem_new().
237 */
238 if (domain & NOUVEAU_GEM_DOMAIN_COHERENT) {
239 /* Determine if we can get a cache-coherent map, forcing
240 * uncached mapping if we can't.
241 */
242 if (!nouveau_drm_use_coherent_gpu_mapping(drm))
243 nvbo->force_coherent = true;
244 }
245
246 nvbo->contig = !(tile_flags & NOUVEAU_GEM_TILE_NONCONTIG);
247
248 if (cli->device.info.family >= NV_DEVICE_INFO_V0_FERMI) {
249 nvbo->kind = (tile_flags & 0x0000ff00) >> 8;
250 if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
251 kfree(nvbo);
252 return ERR_PTR(-EINVAL);
253 }
254
255 nvbo->comp = mmu->kind[nvbo->kind] != nvbo->kind;
256 } else if (cli->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
257 nvbo->kind = (tile_flags & 0x00007f00) >> 8;
258 nvbo->comp = (tile_flags & 0x00030000) >> 16;
259 if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
260 kfree(nvbo);
261 return ERR_PTR(-EINVAL);
262 }
263 } else {
264 nvbo->zeta = (tile_flags & 0x00000007);
265 }
266 nvbo->mode = tile_mode;
267
268 if (!nouveau_cli_uvmm(cli) || internal) {
269 /* Determine the desirable target GPU page size for the buffer. */
270 for (i = 0; i < vmm->page_nr; i++) {
271 /* Because we cannot currently allow VMM maps to fail
272 * during buffer migration, we need to determine page
273 * size for the buffer up-front, and pre-allocate its
274 * page tables.
275 *
276 * Skip page sizes that can't support needed domains.
277 */
278 if (cli->device.info.family > NV_DEVICE_INFO_V0_CURIE &&
279 (domain & NOUVEAU_GEM_DOMAIN_VRAM) && !vmm->page[i].vram)
280 continue;
281 if ((domain & NOUVEAU_GEM_DOMAIN_GART) &&
282 (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
283 continue;
284
285 /* Select this page size if it's the first that supports
286 * the potential memory domains, or when it's compatible
287 * with the requested compression settings.
288 */
289 if (pi < 0 || !nvbo->comp || vmm->page[i].comp)
290 pi = i;
291
292 /* Stop once the buffer is larger than the current page size. */
293 if (*size >= 1ULL << vmm->page[i].shift)
294 break;
295 }
296
297 if (WARN_ON(pi < 0)) {
298 kfree(nvbo);
299 return ERR_PTR(-EINVAL);
300 }
301
302 /* Disable compression if suitable settings couldn't be found. */
303 if (nvbo->comp && !vmm->page[pi].comp) {
304 if (mmu->object.oclass >= NVIF_CLASS_MMU_GF100)
305 nvbo->kind = mmu->kind[nvbo->kind];
306 nvbo->comp = 0;
307 }
308 nvbo->page = vmm->page[pi].shift;
309 } else {
310 /* Determine the desirable target GPU page size for the buffer. */
311 for (i = 0; i < vmm->page_nr; i++) {
312 /* Because we cannot currently allow VMM maps to fail
313 * during buffer migration, we need to determine page
314 * size for the buffer up-front, and pre-allocate its
315 * page tables.
316 *
317 * Skip page sizes that can't support needed domains.
318 */
319 if ((domain & NOUVEAU_GEM_DOMAIN_VRAM) && !vmm->page[i].vram)
320 continue;
321 if ((domain & NOUVEAU_GEM_DOMAIN_GART) &&
322 (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
323 continue;
324
325 /* pick the last one as it will be smallest. */
326 pi = i;
327
328 /* Stop once the buffer is larger than the current page size. */
329 if (*size >= 1ULL << vmm->page[i].shift)
330 break;
331 }
332 if (WARN_ON(pi < 0)) {
333 kfree(nvbo);
334 return ERR_PTR(-EINVAL);
335 }
336 nvbo->page = vmm->page[pi].shift;
337 }
338
339 nouveau_bo_fixup_align(nvbo, align, size);
340
341 return nvbo;
342 }
343
344 int
nouveau_bo_init(struct nouveau_bo * nvbo,u64 size,int align,u32 domain,struct sg_table * sg,struct dma_resv * robj)345 nouveau_bo_init(struct nouveau_bo *nvbo, u64 size, int align, u32 domain,
346 struct sg_table *sg, struct dma_resv *robj)
347 {
348 int type = sg ? ttm_bo_type_sg : ttm_bo_type_device;
349 int ret;
350 struct ttm_operation_ctx ctx = {
351 .interruptible = false,
352 .no_wait_gpu = false,
353 .resv = robj,
354 };
355
356 nouveau_bo_placement_set(nvbo, domain, 0);
357 INIT_LIST_HEAD(&nvbo->io_reserve_lru);
358
359 ret = ttm_bo_init_reserved(nvbo->bo.bdev, &nvbo->bo, type,
360 &nvbo->placement, align >> PAGE_SHIFT, &ctx,
361 sg, robj, nouveau_bo_del_ttm);
362 if (ret) {
363 /* ttm will call nouveau_bo_del_ttm if it fails.. */
364 return ret;
365 }
366
367 if (!robj)
368 ttm_bo_unreserve(&nvbo->bo);
369
370 return 0;
371 }
372
373 int
nouveau_bo_new(struct nouveau_cli * cli,u64 size,int align,uint32_t domain,uint32_t tile_mode,uint32_t tile_flags,struct sg_table * sg,struct dma_resv * robj,struct nouveau_bo ** pnvbo)374 nouveau_bo_new(struct nouveau_cli *cli, u64 size, int align,
375 uint32_t domain, uint32_t tile_mode, uint32_t tile_flags,
376 struct sg_table *sg, struct dma_resv *robj,
377 struct nouveau_bo **pnvbo)
378 {
379 struct nouveau_bo *nvbo;
380 int ret;
381
382 nvbo = nouveau_bo_alloc(cli, &size, &align, domain, tile_mode,
383 tile_flags, true);
384 if (IS_ERR(nvbo))
385 return PTR_ERR(nvbo);
386
387 nvbo->bo.base.size = size;
388 dma_resv_init(&nvbo->bo.base._resv);
389 drm_vma_node_reset(&nvbo->bo.base.vma_node);
390
391 /* This must be called before ttm_bo_init_reserved(). Subsequent
392 * bo_move() callbacks might already iterate the GEMs GPUVA list.
393 */
394 drm_gem_gpuva_init(&nvbo->bo.base);
395
396 ret = nouveau_bo_init(nvbo, size, align, domain, sg, robj);
397 if (ret)
398 return ret;
399
400 *pnvbo = nvbo;
401 return 0;
402 }
403
404 static void
set_placement_range(struct nouveau_bo * nvbo,uint32_t domain)405 set_placement_range(struct nouveau_bo *nvbo, uint32_t domain)
406 {
407 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
408 u64 vram_size = drm->client.device.info.ram_size;
409 unsigned i, fpfn, lpfn;
410
411 if (drm->client.device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
412 nvbo->mode && (domain & NOUVEAU_GEM_DOMAIN_VRAM) &&
413 nvbo->bo.base.size < vram_size / 4) {
414 /*
415 * Make sure that the color and depth buffers are handled
416 * by independent memory controller units. Up to a 9x
417 * speed up when alpha-blending and depth-test are enabled
418 * at the same time.
419 */
420 if (nvbo->zeta) {
421 fpfn = (vram_size / 2) >> PAGE_SHIFT;
422 lpfn = ~0;
423 } else {
424 fpfn = 0;
425 lpfn = (vram_size / 2) >> PAGE_SHIFT;
426 }
427 for (i = 0; i < nvbo->placement.num_placement; ++i) {
428 nvbo->placements[i].fpfn = fpfn;
429 nvbo->placements[i].lpfn = lpfn;
430 }
431 }
432 }
433
434 void
nouveau_bo_placement_set(struct nouveau_bo * nvbo,uint32_t domain,uint32_t busy)435 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t domain,
436 uint32_t busy)
437 {
438 unsigned int *n = &nvbo->placement.num_placement;
439 struct ttm_place *pl = nvbo->placements;
440
441 domain |= busy;
442
443 *n = 0;
444 if (domain & NOUVEAU_GEM_DOMAIN_VRAM) {
445 pl[*n].mem_type = TTM_PL_VRAM;
446 pl[*n].flags = busy & NOUVEAU_GEM_DOMAIN_VRAM ?
447 TTM_PL_FLAG_FALLBACK : 0;
448 (*n)++;
449 }
450 if (domain & NOUVEAU_GEM_DOMAIN_GART) {
451 pl[*n].mem_type = TTM_PL_TT;
452 pl[*n].flags = busy & NOUVEAU_GEM_DOMAIN_GART ?
453 TTM_PL_FLAG_FALLBACK : 0;
454 (*n)++;
455 }
456 if (domain & NOUVEAU_GEM_DOMAIN_CPU) {
457 pl[*n].mem_type = TTM_PL_SYSTEM;
458 pl[*n].flags = busy & NOUVEAU_GEM_DOMAIN_CPU ?
459 TTM_PL_FLAG_FALLBACK : 0;
460 (*n)++;
461 }
462
463 nvbo->placement.placement = nvbo->placements;
464 set_placement_range(nvbo, domain);
465 }
466
nouveau_bo_pin_locked(struct nouveau_bo * nvbo,uint32_t domain,bool contig)467 int nouveau_bo_pin_locked(struct nouveau_bo *nvbo, uint32_t domain, bool contig)
468 {
469 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
470 struct ttm_buffer_object *bo = &nvbo->bo;
471 bool force = false, evict = false;
472 int ret = 0;
473
474 dma_resv_assert_held(bo->base.resv);
475
476 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
477 domain == NOUVEAU_GEM_DOMAIN_VRAM && contig) {
478 if (!nvbo->contig) {
479 nvbo->contig = true;
480 force = true;
481 evict = true;
482 }
483 }
484
485 if (nvbo->bo.pin_count) {
486 bool error = evict;
487
488 switch (bo->resource->mem_type) {
489 case TTM_PL_VRAM:
490 error |= !(domain & NOUVEAU_GEM_DOMAIN_VRAM);
491 break;
492 case TTM_PL_TT:
493 error |= !(domain & NOUVEAU_GEM_DOMAIN_GART);
494 break;
495 default:
496 break;
497 }
498
499 if (error) {
500 NV_ERROR(drm, "bo %p pinned elsewhere: "
501 "0x%08x vs 0x%08x\n", bo,
502 bo->resource->mem_type, domain);
503 ret = -EBUSY;
504 }
505 ttm_bo_pin(&nvbo->bo);
506 goto out;
507 }
508
509 if (evict) {
510 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART, 0);
511 ret = nouveau_bo_validate(nvbo, false, false);
512 if (ret)
513 goto out;
514 }
515
516 nouveau_bo_placement_set(nvbo, domain, 0);
517 ret = nouveau_bo_validate(nvbo, false, false);
518 if (ret)
519 goto out;
520
521 ttm_bo_pin(&nvbo->bo);
522
523 switch (bo->resource->mem_type) {
524 case TTM_PL_VRAM:
525 drm->gem.vram_available -= bo->base.size;
526 break;
527 case TTM_PL_TT:
528 drm->gem.gart_available -= bo->base.size;
529 break;
530 default:
531 break;
532 }
533
534 out:
535 if (force && ret)
536 nvbo->contig = false;
537 return ret;
538 }
539
nouveau_bo_unpin_locked(struct nouveau_bo * nvbo)540 void nouveau_bo_unpin_locked(struct nouveau_bo *nvbo)
541 {
542 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
543 struct ttm_buffer_object *bo = &nvbo->bo;
544
545 dma_resv_assert_held(bo->base.resv);
546
547 ttm_bo_unpin(&nvbo->bo);
548 if (!nvbo->bo.pin_count) {
549 switch (bo->resource->mem_type) {
550 case TTM_PL_VRAM:
551 drm->gem.vram_available += bo->base.size;
552 break;
553 case TTM_PL_TT:
554 drm->gem.gart_available += bo->base.size;
555 break;
556 default:
557 break;
558 }
559 }
560 }
561
nouveau_bo_pin(struct nouveau_bo * nvbo,uint32_t domain,bool contig)562 int nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t domain, bool contig)
563 {
564 struct ttm_buffer_object *bo = &nvbo->bo;
565 int ret;
566
567 ret = ttm_bo_reserve(bo, false, false, NULL);
568 if (ret)
569 return ret;
570 ret = nouveau_bo_pin_locked(nvbo, domain, contig);
571 ttm_bo_unreserve(bo);
572
573 return ret;
574 }
575
nouveau_bo_unpin(struct nouveau_bo * nvbo)576 int nouveau_bo_unpin(struct nouveau_bo *nvbo)
577 {
578 struct ttm_buffer_object *bo = &nvbo->bo;
579 int ret;
580
581 ret = ttm_bo_reserve(bo, false, false, NULL);
582 if (ret)
583 return ret;
584 nouveau_bo_unpin_locked(nvbo);
585 ttm_bo_unreserve(bo);
586
587 return 0;
588 }
589
590 int
nouveau_bo_map(struct nouveau_bo * nvbo)591 nouveau_bo_map(struct nouveau_bo *nvbo)
592 {
593 int ret;
594
595 ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
596 if (ret)
597 return ret;
598
599 ret = ttm_bo_kmap(&nvbo->bo, 0, PFN_UP(nvbo->bo.base.size), &nvbo->kmap);
600
601 ttm_bo_unreserve(&nvbo->bo);
602 return ret;
603 }
604
605 void
nouveau_bo_unmap(struct nouveau_bo * nvbo)606 nouveau_bo_unmap(struct nouveau_bo *nvbo)
607 {
608 if (!nvbo)
609 return;
610
611 ttm_bo_kunmap(&nvbo->kmap);
612 }
613
614 void
nouveau_bo_sync_for_device(struct nouveau_bo * nvbo)615 nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
616 {
617 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
618 struct ttm_tt *ttm_dma = (struct ttm_tt *)nvbo->bo.ttm;
619 int i, j;
620
621 if (!ttm_dma || !ttm_dma->dma_address)
622 return;
623 if (!ttm_dma->pages) {
624 NV_DEBUG(drm, "ttm_dma 0x%p: pages NULL\n", ttm_dma);
625 return;
626 }
627
628 /* Don't waste time looping if the object is coherent */
629 if (nvbo->force_coherent)
630 return;
631
632 i = 0;
633 while (i < ttm_dma->num_pages) {
634 struct page *p = ttm_dma->pages[i];
635 size_t num_pages = 1;
636
637 for (j = i + 1; j < ttm_dma->num_pages; ++j) {
638 if (++p != ttm_dma->pages[j])
639 break;
640
641 ++num_pages;
642 }
643 dma_sync_single_for_device(drm->dev->dev,
644 ttm_dma->dma_address[i],
645 num_pages * PAGE_SIZE, DMA_TO_DEVICE);
646 i += num_pages;
647 }
648 }
649
650 void
nouveau_bo_sync_for_cpu(struct nouveau_bo * nvbo)651 nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
652 {
653 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
654 struct ttm_tt *ttm_dma = (struct ttm_tt *)nvbo->bo.ttm;
655 int i, j;
656
657 if (!ttm_dma || !ttm_dma->dma_address)
658 return;
659 if (!ttm_dma->pages) {
660 NV_DEBUG(drm, "ttm_dma 0x%p: pages NULL\n", ttm_dma);
661 return;
662 }
663
664 /* Don't waste time looping if the object is coherent */
665 if (nvbo->force_coherent)
666 return;
667
668 i = 0;
669 while (i < ttm_dma->num_pages) {
670 struct page *p = ttm_dma->pages[i];
671 size_t num_pages = 1;
672
673 for (j = i + 1; j < ttm_dma->num_pages; ++j) {
674 if (++p != ttm_dma->pages[j])
675 break;
676
677 ++num_pages;
678 }
679
680 dma_sync_single_for_cpu(drm->dev->dev, ttm_dma->dma_address[i],
681 num_pages * PAGE_SIZE, DMA_FROM_DEVICE);
682 i += num_pages;
683 }
684 }
685
nouveau_bo_add_io_reserve_lru(struct ttm_buffer_object * bo)686 void nouveau_bo_add_io_reserve_lru(struct ttm_buffer_object *bo)
687 {
688 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
689 struct nouveau_bo *nvbo = nouveau_bo(bo);
690
691 mutex_lock(&drm->ttm.io_reserve_mutex);
692 list_move_tail(&nvbo->io_reserve_lru, &drm->ttm.io_reserve_lru);
693 mutex_unlock(&drm->ttm.io_reserve_mutex);
694 }
695
nouveau_bo_del_io_reserve_lru(struct ttm_buffer_object * bo)696 void nouveau_bo_del_io_reserve_lru(struct ttm_buffer_object *bo)
697 {
698 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
699 struct nouveau_bo *nvbo = nouveau_bo(bo);
700
701 mutex_lock(&drm->ttm.io_reserve_mutex);
702 list_del_init(&nvbo->io_reserve_lru);
703 mutex_unlock(&drm->ttm.io_reserve_mutex);
704 }
705
706 int
nouveau_bo_validate(struct nouveau_bo * nvbo,bool interruptible,bool no_wait_gpu)707 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
708 bool no_wait_gpu)
709 {
710 struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
711 int ret;
712
713 ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, &ctx);
714 if (ret)
715 return ret;
716
717 nouveau_bo_sync_for_device(nvbo);
718
719 return 0;
720 }
721
722 void
nouveau_bo_wr16(struct nouveau_bo * nvbo,unsigned index,u16 val)723 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
724 {
725 bool is_iomem;
726 u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
727
728 mem += index;
729
730 if (is_iomem)
731 iowrite16_native(val, (void __force __iomem *)mem);
732 else
733 *mem = val;
734 }
735
736 u32
nouveau_bo_rd32(struct nouveau_bo * nvbo,unsigned index)737 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
738 {
739 bool is_iomem;
740 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
741
742 mem += index;
743
744 if (is_iomem)
745 return ioread32_native((void __force __iomem *)mem);
746 else
747 return *mem;
748 }
749
750 void
nouveau_bo_wr32(struct nouveau_bo * nvbo,unsigned index,u32 val)751 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
752 {
753 bool is_iomem;
754 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
755
756 mem += index;
757
758 if (is_iomem)
759 iowrite32_native(val, (void __force __iomem *)mem);
760 else
761 *mem = val;
762 }
763
764 static struct ttm_tt *
nouveau_ttm_tt_create(struct ttm_buffer_object * bo,uint32_t page_flags)765 nouveau_ttm_tt_create(struct ttm_buffer_object *bo, uint32_t page_flags)
766 {
767 #if IS_ENABLED(CONFIG_AGP)
768 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
769
770 if (drm->agp.bridge) {
771 return ttm_agp_tt_create(bo, drm->agp.bridge, page_flags);
772 }
773 #endif
774
775 return nouveau_sgdma_create_ttm(bo, page_flags);
776 }
777
778 static int
nouveau_ttm_tt_bind(struct ttm_device * bdev,struct ttm_tt * ttm,struct ttm_resource * reg)779 nouveau_ttm_tt_bind(struct ttm_device *bdev, struct ttm_tt *ttm,
780 struct ttm_resource *reg)
781 {
782 #if IS_ENABLED(CONFIG_AGP)
783 struct nouveau_drm *drm = nouveau_bdev(bdev);
784 #endif
785 if (!reg)
786 return -EINVAL;
787 #if IS_ENABLED(CONFIG_AGP)
788 if (drm->agp.bridge)
789 return ttm_agp_bind(ttm, reg);
790 #endif
791 return nouveau_sgdma_bind(bdev, ttm, reg);
792 }
793
794 static void
nouveau_ttm_tt_unbind(struct ttm_device * bdev,struct ttm_tt * ttm)795 nouveau_ttm_tt_unbind(struct ttm_device *bdev, struct ttm_tt *ttm)
796 {
797 #if IS_ENABLED(CONFIG_AGP)
798 struct nouveau_drm *drm = nouveau_bdev(bdev);
799
800 if (drm->agp.bridge) {
801 ttm_agp_unbind(ttm);
802 return;
803 }
804 #endif
805 nouveau_sgdma_unbind(bdev, ttm);
806 }
807
808 static void
nouveau_bo_evict_flags(struct ttm_buffer_object * bo,struct ttm_placement * pl)809 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
810 {
811 struct nouveau_bo *nvbo = nouveau_bo(bo);
812
813 switch (bo->resource->mem_type) {
814 case TTM_PL_VRAM:
815 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART,
816 NOUVEAU_GEM_DOMAIN_CPU);
817 break;
818 default:
819 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_CPU, 0);
820 break;
821 }
822
823 *pl = nvbo->placement;
824 }
825
826 static int
nouveau_bo_move_prep(struct nouveau_drm * drm,struct ttm_buffer_object * bo,struct ttm_resource * reg)827 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
828 struct ttm_resource *reg)
829 {
830 struct nouveau_mem *old_mem = nouveau_mem(bo->resource);
831 struct nouveau_mem *new_mem = nouveau_mem(reg);
832 struct nvif_vmm *vmm = &drm->client.vmm.vmm;
833 int ret;
834
835 ret = nvif_vmm_get(vmm, LAZY, false, old_mem->mem.page, 0,
836 old_mem->mem.size, &old_mem->vma[0]);
837 if (ret)
838 return ret;
839
840 ret = nvif_vmm_get(vmm, LAZY, false, new_mem->mem.page, 0,
841 new_mem->mem.size, &old_mem->vma[1]);
842 if (ret)
843 goto done;
844
845 ret = nouveau_mem_map(old_mem, vmm, &old_mem->vma[0]);
846 if (ret)
847 goto done;
848
849 ret = nouveau_mem_map(new_mem, vmm, &old_mem->vma[1]);
850 done:
851 if (ret) {
852 nvif_vmm_put(vmm, &old_mem->vma[1]);
853 nvif_vmm_put(vmm, &old_mem->vma[0]);
854 }
855 return 0;
856 }
857
858 static int
nouveau_bo_move_m2mf(struct ttm_buffer_object * bo,int evict,struct ttm_operation_ctx * ctx,struct ttm_resource * new_reg)859 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict,
860 struct ttm_operation_ctx *ctx,
861 struct ttm_resource *new_reg)
862 {
863 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
864 struct nouveau_channel *chan = drm->ttm.chan;
865 struct nouveau_cli *cli = chan->cli;
866 struct nouveau_fence *fence;
867 int ret;
868
869 /* create temporary vmas for the transfer and attach them to the
870 * old nvkm_mem node, these will get cleaned up after ttm has
871 * destroyed the ttm_resource
872 */
873 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
874 ret = nouveau_bo_move_prep(drm, bo, new_reg);
875 if (ret)
876 return ret;
877 }
878
879 if (drm_drv_uses_atomic_modeset(drm->dev))
880 mutex_lock(&cli->mutex);
881 else
882 mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
883
884 ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, ctx->interruptible);
885 if (ret)
886 goto out_unlock;
887
888 ret = drm->ttm.move(chan, bo, bo->resource, new_reg);
889 if (ret)
890 goto out_unlock;
891
892 ret = nouveau_fence_new(&fence, chan);
893 if (ret)
894 goto out_unlock;
895
896 /* TODO: figure out a better solution here
897 *
898 * wait on the fence here explicitly as going through
899 * ttm_bo_move_accel_cleanup somehow doesn't seem to do it.
900 *
901 * Without this the operation can timeout and we'll fallback to a
902 * software copy, which might take several minutes to finish.
903 */
904 nouveau_fence_wait(fence, false, false);
905 ret = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, false,
906 new_reg);
907 nouveau_fence_unref(&fence);
908
909 out_unlock:
910 mutex_unlock(&cli->mutex);
911 return ret;
912 }
913
914 void
nouveau_bo_move_init(struct nouveau_drm * drm)915 nouveau_bo_move_init(struct nouveau_drm *drm)
916 {
917 static const struct _method_table {
918 const char *name;
919 int engine;
920 s32 oclass;
921 int (*exec)(struct nouveau_channel *,
922 struct ttm_buffer_object *,
923 struct ttm_resource *, struct ttm_resource *);
924 int (*init)(struct nouveau_channel *, u32 handle);
925 } _methods[] = {
926 { "COPY", 4, 0xc7b5, nve0_bo_move_copy, nve0_bo_move_init },
927 { "GRCE", 0, 0xc7b5, nve0_bo_move_copy, nvc0_bo_move_init },
928 { "COPY", 4, 0xc6b5, nve0_bo_move_copy, nve0_bo_move_init },
929 { "GRCE", 0, 0xc6b5, nve0_bo_move_copy, nvc0_bo_move_init },
930 { "COPY", 4, 0xc5b5, nve0_bo_move_copy, nve0_bo_move_init },
931 { "GRCE", 0, 0xc5b5, nve0_bo_move_copy, nvc0_bo_move_init },
932 { "COPY", 4, 0xc3b5, nve0_bo_move_copy, nve0_bo_move_init },
933 { "GRCE", 0, 0xc3b5, nve0_bo_move_copy, nvc0_bo_move_init },
934 { "COPY", 4, 0xc1b5, nve0_bo_move_copy, nve0_bo_move_init },
935 { "GRCE", 0, 0xc1b5, nve0_bo_move_copy, nvc0_bo_move_init },
936 { "COPY", 4, 0xc0b5, nve0_bo_move_copy, nve0_bo_move_init },
937 { "GRCE", 0, 0xc0b5, nve0_bo_move_copy, nvc0_bo_move_init },
938 { "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
939 { "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
940 { "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
941 { "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
942 { "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
943 { "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
944 { "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
945 { "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
946 { "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
947 { "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
948 { "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
949 {},
950 };
951 const struct _method_table *mthd = _methods;
952 const char *name = "CPU";
953 int ret;
954
955 do {
956 struct nouveau_channel *chan;
957
958 if (mthd->engine)
959 chan = drm->cechan;
960 else
961 chan = drm->channel;
962 if (chan == NULL)
963 continue;
964
965 ret = nvif_object_ctor(&chan->user, "ttmBoMove",
966 mthd->oclass | (mthd->engine << 16),
967 mthd->oclass, NULL, 0,
968 &drm->ttm.copy);
969 if (ret == 0) {
970 ret = mthd->init(chan, drm->ttm.copy.handle);
971 if (ret) {
972 nvif_object_dtor(&drm->ttm.copy);
973 continue;
974 }
975
976 drm->ttm.move = mthd->exec;
977 drm->ttm.chan = chan;
978 name = mthd->name;
979 break;
980 }
981 } while ((++mthd)->exec);
982
983 NV_INFO(drm, "MM: using %s for buffer copies\n", name);
984 }
985
nouveau_bo_move_ntfy(struct ttm_buffer_object * bo,struct ttm_resource * new_reg)986 static void nouveau_bo_move_ntfy(struct ttm_buffer_object *bo,
987 struct ttm_resource *new_reg)
988 {
989 struct nouveau_mem *mem = new_reg ? nouveau_mem(new_reg) : NULL;
990 struct nouveau_bo *nvbo = nouveau_bo(bo);
991 struct nouveau_vma *vma;
992 long ret;
993
994 /* ttm can now (stupidly) pass the driver bos it didn't create... */
995 if (bo->destroy != nouveau_bo_del_ttm)
996 return;
997
998 nouveau_bo_del_io_reserve_lru(bo);
999
1000 if (mem && new_reg->mem_type != TTM_PL_SYSTEM &&
1001 mem->mem.page == nvbo->page) {
1002 list_for_each_entry(vma, &nvbo->vma_list, head) {
1003 nouveau_vma_map(vma, mem);
1004 }
1005 nouveau_uvmm_bo_map_all(nvbo, mem);
1006 } else {
1007 list_for_each_entry(vma, &nvbo->vma_list, head) {
1008 ret = dma_resv_wait_timeout(bo->base.resv,
1009 DMA_RESV_USAGE_BOOKKEEP,
1010 false, 15 * HZ);
1011 WARN_ON(ret <= 0);
1012 nouveau_vma_unmap(vma);
1013 }
1014 nouveau_uvmm_bo_unmap_all(nvbo);
1015 }
1016
1017 if (new_reg)
1018 nvbo->offset = (new_reg->start << PAGE_SHIFT);
1019
1020 }
1021
1022 static int
nouveau_bo_vm_bind(struct ttm_buffer_object * bo,struct ttm_resource * new_reg,struct nouveau_drm_tile ** new_tile)1023 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_resource *new_reg,
1024 struct nouveau_drm_tile **new_tile)
1025 {
1026 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1027 struct drm_device *dev = drm->dev;
1028 struct nouveau_bo *nvbo = nouveau_bo(bo);
1029 u64 offset = new_reg->start << PAGE_SHIFT;
1030
1031 *new_tile = NULL;
1032 if (new_reg->mem_type != TTM_PL_VRAM)
1033 return 0;
1034
1035 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
1036 *new_tile = nv10_bo_set_tiling(dev, offset, bo->base.size,
1037 nvbo->mode, nvbo->zeta);
1038 }
1039
1040 return 0;
1041 }
1042
1043 static void
nouveau_bo_vm_cleanup(struct ttm_buffer_object * bo,struct nouveau_drm_tile * new_tile,struct nouveau_drm_tile ** old_tile)1044 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1045 struct nouveau_drm_tile *new_tile,
1046 struct nouveau_drm_tile **old_tile)
1047 {
1048 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1049 struct drm_device *dev = drm->dev;
1050 struct dma_fence *fence;
1051 int ret;
1052
1053 ret = dma_resv_get_singleton(bo->base.resv, DMA_RESV_USAGE_WRITE,
1054 &fence);
1055 if (ret)
1056 dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_WRITE,
1057 false, MAX_SCHEDULE_TIMEOUT);
1058
1059 nv10_bo_put_tile_region(dev, *old_tile, fence);
1060 *old_tile = new_tile;
1061 }
1062
1063 static int
nouveau_bo_move(struct ttm_buffer_object * bo,bool evict,struct ttm_operation_ctx * ctx,struct ttm_resource * new_reg,struct ttm_place * hop)1064 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict,
1065 struct ttm_operation_ctx *ctx,
1066 struct ttm_resource *new_reg,
1067 struct ttm_place *hop)
1068 {
1069 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1070 struct nouveau_bo *nvbo = nouveau_bo(bo);
1071 struct drm_gem_object *obj = &bo->base;
1072 struct ttm_resource *old_reg = bo->resource;
1073 struct nouveau_drm_tile *new_tile = NULL;
1074 int ret = 0;
1075
1076 if (new_reg->mem_type == TTM_PL_TT) {
1077 ret = nouveau_ttm_tt_bind(bo->bdev, bo->ttm, new_reg);
1078 if (ret)
1079 return ret;
1080 }
1081
1082 drm_gpuvm_bo_gem_evict(obj, evict);
1083 nouveau_bo_move_ntfy(bo, new_reg);
1084 ret = ttm_bo_wait_ctx(bo, ctx);
1085 if (ret)
1086 goto out_ntfy;
1087
1088 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1089 ret = nouveau_bo_vm_bind(bo, new_reg, &new_tile);
1090 if (ret)
1091 goto out_ntfy;
1092 }
1093
1094 /* Fake bo copy. */
1095 if (!old_reg || (old_reg->mem_type == TTM_PL_SYSTEM &&
1096 !bo->ttm)) {
1097 ttm_bo_move_null(bo, new_reg);
1098 goto out;
1099 }
1100
1101 if (old_reg->mem_type == TTM_PL_SYSTEM &&
1102 new_reg->mem_type == TTM_PL_TT) {
1103 ttm_bo_move_null(bo, new_reg);
1104 goto out;
1105 }
1106
1107 if (old_reg->mem_type == TTM_PL_TT &&
1108 new_reg->mem_type == TTM_PL_SYSTEM) {
1109 nouveau_ttm_tt_unbind(bo->bdev, bo->ttm);
1110 ttm_resource_free(bo, &bo->resource);
1111 ttm_bo_assign_mem(bo, new_reg);
1112 goto out;
1113 }
1114
1115 /* Hardware assisted copy. */
1116 if (drm->ttm.move) {
1117 if ((old_reg->mem_type == TTM_PL_SYSTEM &&
1118 new_reg->mem_type == TTM_PL_VRAM) ||
1119 (old_reg->mem_type == TTM_PL_VRAM &&
1120 new_reg->mem_type == TTM_PL_SYSTEM)) {
1121 hop->fpfn = 0;
1122 hop->lpfn = 0;
1123 hop->mem_type = TTM_PL_TT;
1124 hop->flags = 0;
1125 return -EMULTIHOP;
1126 }
1127 ret = nouveau_bo_move_m2mf(bo, evict, ctx,
1128 new_reg);
1129 } else
1130 ret = -ENODEV;
1131
1132 if (ret) {
1133 /* Fallback to software copy. */
1134 ret = ttm_bo_move_memcpy(bo, ctx, new_reg);
1135 }
1136
1137 out:
1138 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1139 if (ret)
1140 nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1141 else
1142 nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1143 }
1144 out_ntfy:
1145 if (ret) {
1146 nouveau_bo_move_ntfy(bo, bo->resource);
1147 drm_gpuvm_bo_gem_evict(obj, !evict);
1148 }
1149 return ret;
1150 }
1151
1152 static void
nouveau_ttm_io_mem_free_locked(struct nouveau_drm * drm,struct ttm_resource * reg)1153 nouveau_ttm_io_mem_free_locked(struct nouveau_drm *drm,
1154 struct ttm_resource *reg)
1155 {
1156 struct nouveau_mem *mem = nouveau_mem(reg);
1157
1158 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1159 switch (reg->mem_type) {
1160 case TTM_PL_TT:
1161 if (mem->kind)
1162 nvif_object_unmap_handle(&mem->mem.object);
1163 break;
1164 case TTM_PL_VRAM:
1165 nvif_object_unmap_handle(&mem->mem.object);
1166 break;
1167 default:
1168 break;
1169 }
1170 }
1171 }
1172
1173 static int
nouveau_ttm_io_mem_reserve(struct ttm_device * bdev,struct ttm_resource * reg)1174 nouveau_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *reg)
1175 {
1176 struct nouveau_drm *drm = nouveau_bdev(bdev);
1177 struct nvkm_device *device = nvxx_device(drm);
1178 struct nouveau_mem *mem = nouveau_mem(reg);
1179 struct nvif_mmu *mmu = &drm->client.mmu;
1180 int ret;
1181
1182 mutex_lock(&drm->ttm.io_reserve_mutex);
1183 retry:
1184 switch (reg->mem_type) {
1185 case TTM_PL_SYSTEM:
1186 /* System memory */
1187 ret = 0;
1188 goto out;
1189 case TTM_PL_TT:
1190 #if IS_ENABLED(CONFIG_AGP)
1191 if (drm->agp.bridge) {
1192 reg->bus.offset = (reg->start << PAGE_SHIFT) +
1193 drm->agp.base;
1194 reg->bus.is_iomem = !drm->agp.cma;
1195 reg->bus.caching = ttm_write_combined;
1196 }
1197 #endif
1198 if (drm->client.mem->oclass < NVIF_CLASS_MEM_NV50 ||
1199 !mem->kind) {
1200 /* untiled */
1201 ret = 0;
1202 break;
1203 }
1204 fallthrough; /* tiled memory */
1205 case TTM_PL_VRAM:
1206 reg->bus.offset = (reg->start << PAGE_SHIFT) +
1207 device->func->resource_addr(device, 1);
1208 reg->bus.is_iomem = true;
1209
1210 /* Some BARs do not support being ioremapped WC */
1211 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
1212 mmu->type[drm->ttm.type_vram].type & NVIF_MEM_UNCACHED)
1213 reg->bus.caching = ttm_uncached;
1214 else
1215 reg->bus.caching = ttm_write_combined;
1216
1217 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1218 union {
1219 struct nv50_mem_map_v0 nv50;
1220 struct gf100_mem_map_v0 gf100;
1221 } args;
1222 u64 handle, length;
1223 u32 argc = 0;
1224
1225 switch (mem->mem.object.oclass) {
1226 case NVIF_CLASS_MEM_NV50:
1227 args.nv50.version = 0;
1228 args.nv50.ro = 0;
1229 args.nv50.kind = mem->kind;
1230 args.nv50.comp = mem->comp;
1231 argc = sizeof(args.nv50);
1232 break;
1233 case NVIF_CLASS_MEM_GF100:
1234 args.gf100.version = 0;
1235 args.gf100.ro = 0;
1236 args.gf100.kind = mem->kind;
1237 argc = sizeof(args.gf100);
1238 break;
1239 default:
1240 WARN_ON(1);
1241 break;
1242 }
1243
1244 ret = nvif_object_map_handle(&mem->mem.object,
1245 &args, argc,
1246 &handle, &length);
1247 if (ret != 1) {
1248 if (WARN_ON(ret == 0))
1249 ret = -EINVAL;
1250 goto out;
1251 }
1252
1253 reg->bus.offset = handle;
1254 }
1255 ret = 0;
1256 break;
1257 default:
1258 ret = -EINVAL;
1259 }
1260
1261 out:
1262 if (ret == -ENOSPC) {
1263 struct nouveau_bo *nvbo;
1264
1265 nvbo = list_first_entry_or_null(&drm->ttm.io_reserve_lru,
1266 typeof(*nvbo),
1267 io_reserve_lru);
1268 if (nvbo) {
1269 list_del_init(&nvbo->io_reserve_lru);
1270 drm_vma_node_unmap(&nvbo->bo.base.vma_node,
1271 bdev->dev_mapping);
1272 nouveau_ttm_io_mem_free_locked(drm, nvbo->bo.resource);
1273 nvbo->bo.resource->bus.offset = 0;
1274 nvbo->bo.resource->bus.addr = NULL;
1275 goto retry;
1276 }
1277
1278 }
1279 mutex_unlock(&drm->ttm.io_reserve_mutex);
1280 return ret;
1281 }
1282
1283 static void
nouveau_ttm_io_mem_free(struct ttm_device * bdev,struct ttm_resource * reg)1284 nouveau_ttm_io_mem_free(struct ttm_device *bdev, struct ttm_resource *reg)
1285 {
1286 struct nouveau_drm *drm = nouveau_bdev(bdev);
1287
1288 mutex_lock(&drm->ttm.io_reserve_mutex);
1289 nouveau_ttm_io_mem_free_locked(drm, reg);
1290 mutex_unlock(&drm->ttm.io_reserve_mutex);
1291 }
1292
nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object * bo)1293 vm_fault_t nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1294 {
1295 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1296 struct nouveau_bo *nvbo = nouveau_bo(bo);
1297 struct nvkm_device *device = nvxx_device(drm);
1298 u32 mappable = device->func->resource_size(device, 1) >> PAGE_SHIFT;
1299 int i, ret;
1300
1301 /* as long as the bo isn't in vram, and isn't tiled, we've got
1302 * nothing to do here.
1303 */
1304 if (bo->resource->mem_type != TTM_PL_VRAM) {
1305 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1306 !nvbo->kind)
1307 return 0;
1308
1309 if (bo->resource->mem_type != TTM_PL_SYSTEM)
1310 return 0;
1311
1312 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART, 0);
1313
1314 } else {
1315 /* make sure bo is in mappable vram */
1316 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1317 bo->resource->start + PFN_UP(bo->resource->size) < mappable)
1318 return 0;
1319
1320 for (i = 0; i < nvbo->placement.num_placement; ++i) {
1321 nvbo->placements[i].fpfn = 0;
1322 nvbo->placements[i].lpfn = mappable;
1323 }
1324
1325 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_VRAM, 0);
1326 }
1327
1328 ret = nouveau_bo_validate(nvbo, false, false);
1329 if (unlikely(ret == -EBUSY || ret == -ERESTARTSYS))
1330 return VM_FAULT_NOPAGE;
1331 else if (unlikely(ret))
1332 return VM_FAULT_SIGBUS;
1333
1334 ttm_bo_move_to_lru_tail_unlocked(bo);
1335 return 0;
1336 }
1337
1338 static int
nouveau_ttm_tt_populate(struct ttm_device * bdev,struct ttm_tt * ttm,struct ttm_operation_ctx * ctx)1339 nouveau_ttm_tt_populate(struct ttm_device *bdev,
1340 struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1341 {
1342 struct ttm_tt *ttm_dma = (void *)ttm;
1343 struct nouveau_drm *drm;
1344 bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL);
1345
1346 if (ttm_tt_is_populated(ttm))
1347 return 0;
1348
1349 if (slave && ttm->sg) {
1350 drm_prime_sg_to_dma_addr_array(ttm->sg, ttm_dma->dma_address,
1351 ttm->num_pages);
1352 return 0;
1353 }
1354
1355 drm = nouveau_bdev(bdev);
1356
1357 return ttm_pool_alloc(&drm->ttm.bdev.pool, ttm, ctx);
1358 }
1359
1360 static void
nouveau_ttm_tt_unpopulate(struct ttm_device * bdev,struct ttm_tt * ttm)1361 nouveau_ttm_tt_unpopulate(struct ttm_device *bdev,
1362 struct ttm_tt *ttm)
1363 {
1364 struct nouveau_drm *drm;
1365 bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL);
1366
1367 if (slave)
1368 return;
1369
1370 nouveau_ttm_tt_unbind(bdev, ttm);
1371
1372 drm = nouveau_bdev(bdev);
1373
1374 return ttm_pool_free(&drm->ttm.bdev.pool, ttm);
1375 }
1376
1377 static void
nouveau_ttm_tt_destroy(struct ttm_device * bdev,struct ttm_tt * ttm)1378 nouveau_ttm_tt_destroy(struct ttm_device *bdev,
1379 struct ttm_tt *ttm)
1380 {
1381 #if IS_ENABLED(CONFIG_AGP)
1382 struct nouveau_drm *drm = nouveau_bdev(bdev);
1383 if (drm->agp.bridge) {
1384 ttm_agp_destroy(ttm);
1385 return;
1386 }
1387 #endif
1388 nouveau_sgdma_destroy(bdev, ttm);
1389 }
1390
1391 void
nouveau_bo_fence(struct nouveau_bo * nvbo,struct nouveau_fence * fence,bool exclusive)1392 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1393 {
1394 struct dma_resv *resv = nvbo->bo.base.resv;
1395
1396 if (!fence)
1397 return;
1398
1399 dma_resv_add_fence(resv, &fence->base, exclusive ?
1400 DMA_RESV_USAGE_WRITE : DMA_RESV_USAGE_READ);
1401 }
1402
1403 static void
nouveau_bo_delete_mem_notify(struct ttm_buffer_object * bo)1404 nouveau_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1405 {
1406 nouveau_bo_move_ntfy(bo, NULL);
1407 }
1408
1409 struct ttm_device_funcs nouveau_bo_driver = {
1410 .ttm_tt_create = &nouveau_ttm_tt_create,
1411 .ttm_tt_populate = &nouveau_ttm_tt_populate,
1412 .ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1413 .ttm_tt_destroy = &nouveau_ttm_tt_destroy,
1414 .eviction_valuable = ttm_bo_eviction_valuable,
1415 .evict_flags = nouveau_bo_evict_flags,
1416 .delete_mem_notify = nouveau_bo_delete_mem_notify,
1417 .move = nouveau_bo_move,
1418 .io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1419 .io_mem_free = &nouveau_ttm_io_mem_free,
1420 };
1421