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 <airlied@linux.ie>
26 * Ben Skeggs <darktama@iinet.net.au>
27 * Jeremy Kolb <jkolb@brandeis.edu>
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 void
nouveau_bo_unpin_del(struct nouveau_bo ** pnvbo)405 nouveau_bo_unpin_del(struct nouveau_bo **pnvbo)
406 {
407 struct nouveau_bo *nvbo = *pnvbo;
408
409 if (!nvbo)
410 return;
411
412 nouveau_bo_unmap(nvbo);
413 nouveau_bo_unpin(nvbo);
414 nouveau_bo_fini(nvbo);
415
416 *pnvbo = NULL;
417 }
418
419 int
nouveau_bo_new_pin(struct nouveau_cli * cli,u32 domain,u32 size,struct nouveau_bo ** pnvbo)420 nouveau_bo_new_pin(struct nouveau_cli *cli, u32 domain, u32 size, struct nouveau_bo **pnvbo)
421 {
422 struct nouveau_bo *nvbo;
423 int ret;
424
425 ret = nouveau_bo_new(cli, size, 0, domain, 0, 0, NULL, NULL, &nvbo);
426 if (ret)
427 return ret;
428
429 ret = nouveau_bo_pin(nvbo, domain, false);
430 if (ret) {
431 nouveau_bo_fini(nvbo);
432 return ret;
433 }
434
435 *pnvbo = nvbo;
436 return 0;
437 }
438
439 int
nouveau_bo_new_map(struct nouveau_cli * cli,u32 domain,u32 size,struct nouveau_bo ** pnvbo)440 nouveau_bo_new_map(struct nouveau_cli *cli, u32 domain, u32 size, struct nouveau_bo **pnvbo)
441 {
442 struct nouveau_bo *nvbo;
443 int ret;
444
445 ret = nouveau_bo_new_pin(cli, domain, size, &nvbo);
446 if (ret)
447 return ret;
448
449 ret = nouveau_bo_map(nvbo);
450 if (ret) {
451 nouveau_bo_unpin_del(&nvbo);
452 return ret;
453 }
454
455 *pnvbo = nvbo;
456 return 0;
457 }
458
459 int
nouveau_bo_new_map_gpu(struct nouveau_cli * cli,u32 domain,u32 size,struct nouveau_bo ** pnvbo,struct nouveau_vma ** pvma)460 nouveau_bo_new_map_gpu(struct nouveau_cli *cli, u32 domain, u32 size,
461 struct nouveau_bo **pnvbo, struct nouveau_vma **pvma)
462 {
463 struct nouveau_vmm *vmm = nouveau_cli_vmm(cli);
464 struct nouveau_bo *nvbo;
465 int ret;
466
467 ret = nouveau_bo_new_map(cli, domain, size, &nvbo);
468 if (ret)
469 return ret;
470
471 ret = nouveau_vma_new(nvbo, vmm, pvma);
472 if (ret) {
473 nouveau_bo_unpin_del(&nvbo);
474 return ret;
475 }
476
477 *pnvbo = nvbo;
478 return 0;
479 }
480
481 static void
set_placement_range(struct nouveau_bo * nvbo,uint32_t domain)482 set_placement_range(struct nouveau_bo *nvbo, uint32_t domain)
483 {
484 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
485 u64 vram_size = drm->client.device.info.ram_size;
486 unsigned i, fpfn, lpfn;
487
488 if (drm->client.device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
489 nvbo->mode && (domain & NOUVEAU_GEM_DOMAIN_VRAM) &&
490 nvbo->bo.base.size < vram_size / 4) {
491 /*
492 * Make sure that the color and depth buffers are handled
493 * by independent memory controller units. Up to a 9x
494 * speed up when alpha-blending and depth-test are enabled
495 * at the same time.
496 */
497 if (nvbo->zeta) {
498 fpfn = (vram_size / 2) >> PAGE_SHIFT;
499 lpfn = ~0;
500 } else {
501 fpfn = 0;
502 lpfn = (vram_size / 2) >> PAGE_SHIFT;
503 }
504 for (i = 0; i < nvbo->placement.num_placement; ++i) {
505 nvbo->placements[i].fpfn = fpfn;
506 nvbo->placements[i].lpfn = lpfn;
507 }
508 }
509 }
510
511 void
nouveau_bo_placement_set(struct nouveau_bo * nvbo,uint32_t domain,uint32_t busy)512 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t domain,
513 uint32_t busy)
514 {
515 unsigned int *n = &nvbo->placement.num_placement;
516 struct ttm_place *pl = nvbo->placements;
517
518 domain |= busy;
519
520 *n = 0;
521 if (domain & NOUVEAU_GEM_DOMAIN_VRAM) {
522 pl[*n].mem_type = TTM_PL_VRAM;
523 pl[*n].flags = busy & NOUVEAU_GEM_DOMAIN_VRAM ?
524 TTM_PL_FLAG_FALLBACK : 0;
525 (*n)++;
526 }
527 if (domain & NOUVEAU_GEM_DOMAIN_GART) {
528 pl[*n].mem_type = TTM_PL_TT;
529 pl[*n].flags = busy & NOUVEAU_GEM_DOMAIN_GART ?
530 TTM_PL_FLAG_FALLBACK : 0;
531 (*n)++;
532 }
533 if (domain & NOUVEAU_GEM_DOMAIN_CPU) {
534 pl[*n].mem_type = TTM_PL_SYSTEM;
535 pl[*n].flags = busy & NOUVEAU_GEM_DOMAIN_CPU ?
536 TTM_PL_FLAG_FALLBACK : 0;
537 (*n)++;
538 }
539
540 nvbo->placement.placement = nvbo->placements;
541 set_placement_range(nvbo, domain);
542 }
543
nouveau_bo_pin_locked(struct nouveau_bo * nvbo,uint32_t domain,bool contig)544 int nouveau_bo_pin_locked(struct nouveau_bo *nvbo, uint32_t domain, bool contig)
545 {
546 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
547 struct ttm_buffer_object *bo = &nvbo->bo;
548 bool force = false, evict = false;
549 int ret = 0;
550
551 dma_resv_assert_held(bo->base.resv);
552
553 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
554 domain == NOUVEAU_GEM_DOMAIN_VRAM && contig) {
555 if (!nvbo->contig) {
556 nvbo->contig = true;
557 force = true;
558 evict = true;
559 }
560 }
561
562 if (nvbo->bo.pin_count) {
563 bool error = evict;
564
565 switch (bo->resource->mem_type) {
566 case TTM_PL_VRAM:
567 error |= !(domain & NOUVEAU_GEM_DOMAIN_VRAM);
568 break;
569 case TTM_PL_TT:
570 error |= !(domain & NOUVEAU_GEM_DOMAIN_GART);
571 break;
572 default:
573 break;
574 }
575
576 if (error) {
577 NV_ERROR(drm, "bo %p pinned elsewhere: "
578 "0x%08x vs 0x%08x\n", bo,
579 bo->resource->mem_type, domain);
580 ret = -EBUSY;
581 }
582 ttm_bo_pin(&nvbo->bo);
583 goto out;
584 }
585
586 if (evict) {
587 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART, 0);
588 ret = nouveau_bo_validate(nvbo, false, false);
589 if (ret)
590 goto out;
591 }
592
593 nouveau_bo_placement_set(nvbo, domain, 0);
594 ret = nouveau_bo_validate(nvbo, false, false);
595 if (ret)
596 goto out;
597
598 ttm_bo_pin(&nvbo->bo);
599
600 switch (bo->resource->mem_type) {
601 case TTM_PL_VRAM:
602 drm->gem.vram_available -= bo->base.size;
603 break;
604 case TTM_PL_TT:
605 drm->gem.gart_available -= bo->base.size;
606 break;
607 default:
608 break;
609 }
610
611 out:
612 if (force && ret)
613 nvbo->contig = false;
614 return ret;
615 }
616
nouveau_bo_unpin_locked(struct nouveau_bo * nvbo)617 void nouveau_bo_unpin_locked(struct nouveau_bo *nvbo)
618 {
619 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
620 struct ttm_buffer_object *bo = &nvbo->bo;
621
622 dma_resv_assert_held(bo->base.resv);
623
624 ttm_bo_unpin(&nvbo->bo);
625 if (!nvbo->bo.pin_count) {
626 switch (bo->resource->mem_type) {
627 case TTM_PL_VRAM:
628 drm->gem.vram_available += bo->base.size;
629 break;
630 case TTM_PL_TT:
631 drm->gem.gart_available += bo->base.size;
632 break;
633 default:
634 break;
635 }
636 }
637 }
638
nouveau_bo_pin(struct nouveau_bo * nvbo,uint32_t domain,bool contig)639 int nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t domain, bool contig)
640 {
641 struct ttm_buffer_object *bo = &nvbo->bo;
642 int ret;
643
644 ret = ttm_bo_reserve(bo, false, false, NULL);
645 if (ret)
646 return ret;
647 ret = nouveau_bo_pin_locked(nvbo, domain, contig);
648 ttm_bo_unreserve(bo);
649
650 return ret;
651 }
652
nouveau_bo_unpin(struct nouveau_bo * nvbo)653 int nouveau_bo_unpin(struct nouveau_bo *nvbo)
654 {
655 struct ttm_buffer_object *bo = &nvbo->bo;
656 int ret;
657
658 ret = ttm_bo_reserve(bo, false, false, NULL);
659 if (ret)
660 return ret;
661 nouveau_bo_unpin_locked(nvbo);
662 ttm_bo_unreserve(bo);
663
664 return 0;
665 }
666
667 int
nouveau_bo_map(struct nouveau_bo * nvbo)668 nouveau_bo_map(struct nouveau_bo *nvbo)
669 {
670 int ret;
671
672 ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
673 if (ret)
674 return ret;
675
676 ret = ttm_bo_kmap(&nvbo->bo, 0, PFN_UP(nvbo->bo.base.size), &nvbo->kmap);
677
678 ttm_bo_unreserve(&nvbo->bo);
679 return ret;
680 }
681
682 void
nouveau_bo_unmap(struct nouveau_bo * nvbo)683 nouveau_bo_unmap(struct nouveau_bo *nvbo)
684 {
685 if (!nvbo)
686 return;
687
688 ttm_bo_kunmap(&nvbo->kmap);
689 }
690
691 void
nouveau_bo_sync_for_device(struct nouveau_bo * nvbo)692 nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
693 {
694 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
695 struct ttm_tt *ttm_dma = (struct ttm_tt *)nvbo->bo.ttm;
696 int i, j;
697
698 if (!ttm_dma || !ttm_dma->dma_address)
699 return;
700 if (!ttm_dma->pages) {
701 NV_DEBUG(drm, "ttm_dma 0x%p: pages NULL\n", ttm_dma);
702 return;
703 }
704
705 /* Don't waste time looping if the object is coherent */
706 if (nvbo->force_coherent)
707 return;
708
709 i = 0;
710 while (i < ttm_dma->num_pages) {
711 struct page *p = ttm_dma->pages[i];
712 size_t num_pages = 1;
713
714 for (j = i + 1; j < ttm_dma->num_pages; ++j) {
715 if (++p != ttm_dma->pages[j])
716 break;
717
718 ++num_pages;
719 }
720 dma_sync_single_for_device(drm->dev->dev,
721 ttm_dma->dma_address[i],
722 num_pages * PAGE_SIZE, DMA_TO_DEVICE);
723 i += num_pages;
724 }
725 }
726
727 void
nouveau_bo_sync_for_cpu(struct nouveau_bo * nvbo)728 nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
729 {
730 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
731 struct ttm_tt *ttm_dma = (struct ttm_tt *)nvbo->bo.ttm;
732 int i, j;
733
734 if (!ttm_dma || !ttm_dma->dma_address)
735 return;
736 if (!ttm_dma->pages) {
737 NV_DEBUG(drm, "ttm_dma 0x%p: pages NULL\n", ttm_dma);
738 return;
739 }
740
741 /* Don't waste time looping if the object is coherent */
742 if (nvbo->force_coherent)
743 return;
744
745 i = 0;
746 while (i < ttm_dma->num_pages) {
747 struct page *p = ttm_dma->pages[i];
748 size_t num_pages = 1;
749
750 for (j = i + 1; j < ttm_dma->num_pages; ++j) {
751 if (++p != ttm_dma->pages[j])
752 break;
753
754 ++num_pages;
755 }
756
757 dma_sync_single_for_cpu(drm->dev->dev, ttm_dma->dma_address[i],
758 num_pages * PAGE_SIZE, DMA_FROM_DEVICE);
759 i += num_pages;
760 }
761 }
762
nouveau_bo_add_io_reserve_lru(struct ttm_buffer_object * bo)763 void nouveau_bo_add_io_reserve_lru(struct ttm_buffer_object *bo)
764 {
765 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
766 struct nouveau_bo *nvbo = nouveau_bo(bo);
767
768 mutex_lock(&drm->ttm.io_reserve_mutex);
769 list_move_tail(&nvbo->io_reserve_lru, &drm->ttm.io_reserve_lru);
770 mutex_unlock(&drm->ttm.io_reserve_mutex);
771 }
772
nouveau_bo_del_io_reserve_lru(struct ttm_buffer_object * bo)773 void nouveau_bo_del_io_reserve_lru(struct ttm_buffer_object *bo)
774 {
775 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
776 struct nouveau_bo *nvbo = nouveau_bo(bo);
777
778 mutex_lock(&drm->ttm.io_reserve_mutex);
779 list_del_init(&nvbo->io_reserve_lru);
780 mutex_unlock(&drm->ttm.io_reserve_mutex);
781 }
782
783 int
nouveau_bo_validate(struct nouveau_bo * nvbo,bool interruptible,bool no_wait_gpu)784 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
785 bool no_wait_gpu)
786 {
787 struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
788 int ret;
789
790 ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, &ctx);
791 if (ret)
792 return ret;
793
794 nouveau_bo_sync_for_device(nvbo);
795
796 return 0;
797 }
798
799 void
nouveau_bo_wr16(struct nouveau_bo * nvbo,unsigned index,u16 val)800 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
801 {
802 bool is_iomem;
803 u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
804
805 mem += index;
806
807 if (is_iomem)
808 iowrite16_native(val, (void __force __iomem *)mem);
809 else
810 *mem = val;
811 }
812
813 u32
nouveau_bo_rd32(struct nouveau_bo * nvbo,unsigned index)814 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
815 {
816 bool is_iomem;
817 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
818
819 mem += index;
820
821 if (is_iomem)
822 return ioread32_native((void __force __iomem *)mem);
823 else
824 return *mem;
825 }
826
827 void
nouveau_bo_wr32(struct nouveau_bo * nvbo,unsigned index,u32 val)828 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
829 {
830 bool is_iomem;
831 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
832
833 mem += index;
834
835 if (is_iomem)
836 iowrite32_native(val, (void __force __iomem *)mem);
837 else
838 *mem = val;
839 }
840
841 static struct ttm_tt *
nouveau_ttm_tt_create(struct ttm_buffer_object * bo,uint32_t page_flags)842 nouveau_ttm_tt_create(struct ttm_buffer_object *bo, uint32_t page_flags)
843 {
844 #if IS_ENABLED(CONFIG_AGP)
845 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
846
847 if (drm->agp.bridge) {
848 return ttm_agp_tt_create(bo, drm->agp.bridge, page_flags);
849 }
850 #endif
851
852 return nouveau_sgdma_create_ttm(bo, page_flags);
853 }
854
855 static int
nouveau_ttm_tt_bind(struct ttm_device * bdev,struct ttm_tt * ttm,struct ttm_resource * reg)856 nouveau_ttm_tt_bind(struct ttm_device *bdev, struct ttm_tt *ttm,
857 struct ttm_resource *reg)
858 {
859 #if IS_ENABLED(CONFIG_AGP)
860 struct nouveau_drm *drm = nouveau_bdev(bdev);
861 #endif
862 if (!reg)
863 return -EINVAL;
864 #if IS_ENABLED(CONFIG_AGP)
865 if (drm->agp.bridge)
866 return ttm_agp_bind(ttm, reg);
867 #endif
868 return nouveau_sgdma_bind(bdev, ttm, reg);
869 }
870
871 static void
nouveau_ttm_tt_unbind(struct ttm_device * bdev,struct ttm_tt * ttm)872 nouveau_ttm_tt_unbind(struct ttm_device *bdev, struct ttm_tt *ttm)
873 {
874 #if IS_ENABLED(CONFIG_AGP)
875 struct nouveau_drm *drm = nouveau_bdev(bdev);
876
877 if (drm->agp.bridge) {
878 ttm_agp_unbind(ttm);
879 return;
880 }
881 #endif
882 nouveau_sgdma_unbind(bdev, ttm);
883 }
884
885 static void
nouveau_bo_evict_flags(struct ttm_buffer_object * bo,struct ttm_placement * pl)886 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
887 {
888 struct nouveau_bo *nvbo = nouveau_bo(bo);
889
890 switch (bo->resource->mem_type) {
891 case TTM_PL_VRAM:
892 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART,
893 NOUVEAU_GEM_DOMAIN_CPU);
894 break;
895 default:
896 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_CPU, 0);
897 break;
898 }
899
900 *pl = nvbo->placement;
901 }
902
903 static int
nouveau_bo_move_prep(struct nouveau_drm * drm,struct ttm_buffer_object * bo,struct ttm_resource * reg)904 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
905 struct ttm_resource *reg)
906 {
907 struct nouveau_mem *old_mem = nouveau_mem(bo->resource);
908 struct nouveau_mem *new_mem = nouveau_mem(reg);
909 struct nvif_vmm *vmm = &drm->client.vmm.vmm;
910 int ret;
911
912 ret = nvif_vmm_get(vmm, LAZY, false, old_mem->mem.page, 0,
913 old_mem->mem.size, &old_mem->vma[0]);
914 if (ret)
915 return ret;
916
917 ret = nvif_vmm_get(vmm, LAZY, false, new_mem->mem.page, 0,
918 new_mem->mem.size, &old_mem->vma[1]);
919 if (ret)
920 goto done;
921
922 ret = nouveau_mem_map(old_mem, vmm, &old_mem->vma[0]);
923 if (ret)
924 goto done;
925
926 ret = nouveau_mem_map(new_mem, vmm, &old_mem->vma[1]);
927 done:
928 if (ret) {
929 nvif_vmm_put(vmm, &old_mem->vma[1]);
930 nvif_vmm_put(vmm, &old_mem->vma[0]);
931 }
932 return 0;
933 }
934
935 static int
nouveau_bo_move_m2mf(struct ttm_buffer_object * bo,int evict,struct ttm_operation_ctx * ctx,struct ttm_resource * new_reg)936 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict,
937 struct ttm_operation_ctx *ctx,
938 struct ttm_resource *new_reg)
939 {
940 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
941 struct nouveau_channel *chan = drm->ttm.chan;
942 struct nouveau_cli *cli = chan->cli;
943 struct nouveau_fence *fence;
944 int ret;
945
946 /* create temporary vmas for the transfer and attach them to the
947 * old nvkm_mem node, these will get cleaned up after ttm has
948 * destroyed the ttm_resource
949 */
950 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
951 ret = nouveau_bo_move_prep(drm, bo, new_reg);
952 if (ret)
953 return ret;
954 }
955
956 if (drm_drv_uses_atomic_modeset(drm->dev))
957 mutex_lock(&cli->mutex);
958 else
959 mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
960
961 ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, ctx->interruptible);
962 if (ret)
963 goto out_unlock;
964
965 ret = drm->ttm.move(chan, bo, bo->resource, new_reg);
966 if (ret)
967 goto out_unlock;
968
969 ret = nouveau_fence_new(&fence, chan);
970 if (ret)
971 goto out_unlock;
972
973 /* TODO: figure out a better solution here
974 *
975 * wait on the fence here explicitly as going through
976 * ttm_bo_move_accel_cleanup somehow doesn't seem to do it.
977 *
978 * Without this the operation can timeout and we'll fallback to a
979 * software copy, which might take several minutes to finish.
980 */
981 nouveau_fence_wait(fence, false, false);
982 ret = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, false,
983 new_reg);
984 nouveau_fence_unref(&fence);
985
986 out_unlock:
987 mutex_unlock(&cli->mutex);
988 return ret;
989 }
990
991 void
nouveau_bo_move_init(struct nouveau_drm * drm)992 nouveau_bo_move_init(struct nouveau_drm *drm)
993 {
994 static const struct _method_table {
995 const char *name;
996 int engine;
997 s32 oclass;
998 int (*exec)(struct nouveau_channel *,
999 struct ttm_buffer_object *,
1000 struct ttm_resource *, struct ttm_resource *);
1001 int (*init)(struct nouveau_channel *, u32 handle);
1002 } _methods[] = {
1003 { "COPY", 4, 0xcab5, nve0_bo_move_copy, nve0_bo_move_init },
1004 { "COPY", 4, 0xc9b5, nve0_bo_move_copy, nve0_bo_move_init },
1005 { "COPY", 4, 0xc8b5, nve0_bo_move_copy, nve0_bo_move_init },
1006 { "COPY", 4, 0xc7b5, nve0_bo_move_copy, nve0_bo_move_init },
1007 { "GRCE", 0, 0xc7b5, nve0_bo_move_copy, nvc0_bo_move_init },
1008 { "COPY", 4, 0xc6b5, nve0_bo_move_copy, nve0_bo_move_init },
1009 { "GRCE", 0, 0xc6b5, nve0_bo_move_copy, nvc0_bo_move_init },
1010 { "COPY", 4, 0xc5b5, nve0_bo_move_copy, nve0_bo_move_init },
1011 { "GRCE", 0, 0xc5b5, nve0_bo_move_copy, nvc0_bo_move_init },
1012 { "COPY", 4, 0xc3b5, nve0_bo_move_copy, nve0_bo_move_init },
1013 { "GRCE", 0, 0xc3b5, nve0_bo_move_copy, nvc0_bo_move_init },
1014 { "COPY", 4, 0xc1b5, nve0_bo_move_copy, nve0_bo_move_init },
1015 { "GRCE", 0, 0xc1b5, nve0_bo_move_copy, nvc0_bo_move_init },
1016 { "COPY", 4, 0xc0b5, nve0_bo_move_copy, nve0_bo_move_init },
1017 { "GRCE", 0, 0xc0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1018 { "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
1019 { "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1020 { "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
1021 { "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1022 { "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
1023 { "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
1024 { "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
1025 { "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
1026 { "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
1027 { "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
1028 { "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
1029 {},
1030 };
1031 const struct _method_table *mthd = _methods;
1032 const char *name = "CPU";
1033 int ret;
1034
1035 do {
1036 struct nouveau_channel *chan;
1037
1038 if (mthd->engine)
1039 chan = drm->cechan;
1040 else
1041 chan = drm->channel;
1042 if (chan == NULL)
1043 continue;
1044
1045 ret = nvif_object_ctor(&chan->user, "ttmBoMove",
1046 mthd->oclass | (mthd->engine << 16),
1047 mthd->oclass, NULL, 0,
1048 &drm->ttm.copy);
1049 if (ret == 0) {
1050 ret = mthd->init(chan, drm->ttm.copy.handle);
1051 if (ret) {
1052 nvif_object_dtor(&drm->ttm.copy);
1053 continue;
1054 }
1055
1056 drm->ttm.move = mthd->exec;
1057 drm->ttm.chan = chan;
1058 name = mthd->name;
1059 break;
1060 }
1061 } while ((++mthd)->exec);
1062
1063 NV_INFO(drm, "MM: using %s for buffer copies\n", name);
1064 }
1065
nouveau_bo_move_ntfy(struct ttm_buffer_object * bo,struct ttm_resource * new_reg)1066 static void nouveau_bo_move_ntfy(struct ttm_buffer_object *bo,
1067 struct ttm_resource *new_reg)
1068 {
1069 struct nouveau_mem *mem = new_reg ? nouveau_mem(new_reg) : NULL;
1070 struct nouveau_bo *nvbo = nouveau_bo(bo);
1071 struct nouveau_vma *vma;
1072 long ret;
1073
1074 /* ttm can now (stupidly) pass the driver bos it didn't create... */
1075 if (bo->destroy != nouveau_bo_del_ttm)
1076 return;
1077
1078 nouveau_bo_del_io_reserve_lru(bo);
1079
1080 if (mem && new_reg->mem_type != TTM_PL_SYSTEM &&
1081 mem->mem.page == nvbo->page) {
1082 list_for_each_entry(vma, &nvbo->vma_list, head) {
1083 nouveau_vma_map(vma, mem);
1084 }
1085 nouveau_uvmm_bo_map_all(nvbo, mem);
1086 } else {
1087 list_for_each_entry(vma, &nvbo->vma_list, head) {
1088 ret = dma_resv_wait_timeout(bo->base.resv,
1089 DMA_RESV_USAGE_BOOKKEEP,
1090 false, 15 * HZ);
1091 WARN_ON(ret <= 0);
1092 nouveau_vma_unmap(vma);
1093 }
1094 nouveau_uvmm_bo_unmap_all(nvbo);
1095 }
1096
1097 if (new_reg)
1098 nvbo->offset = (new_reg->start << PAGE_SHIFT);
1099
1100 }
1101
1102 static int
nouveau_bo_vm_bind(struct ttm_buffer_object * bo,struct ttm_resource * new_reg,struct nouveau_drm_tile ** new_tile)1103 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_resource *new_reg,
1104 struct nouveau_drm_tile **new_tile)
1105 {
1106 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1107 struct drm_device *dev = drm->dev;
1108 struct nouveau_bo *nvbo = nouveau_bo(bo);
1109 u64 offset = new_reg->start << PAGE_SHIFT;
1110
1111 *new_tile = NULL;
1112 if (new_reg->mem_type != TTM_PL_VRAM)
1113 return 0;
1114
1115 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
1116 *new_tile = nv10_bo_set_tiling(dev, offset, bo->base.size,
1117 nvbo->mode, nvbo->zeta);
1118 }
1119
1120 return 0;
1121 }
1122
1123 static void
nouveau_bo_vm_cleanup(struct ttm_buffer_object * bo,struct nouveau_drm_tile * new_tile,struct nouveau_drm_tile ** old_tile)1124 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1125 struct nouveau_drm_tile *new_tile,
1126 struct nouveau_drm_tile **old_tile)
1127 {
1128 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1129 struct drm_device *dev = drm->dev;
1130 struct dma_fence *fence;
1131 int ret;
1132
1133 ret = dma_resv_get_singleton(bo->base.resv, DMA_RESV_USAGE_WRITE,
1134 &fence);
1135 if (ret)
1136 dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_WRITE,
1137 false, MAX_SCHEDULE_TIMEOUT);
1138
1139 nv10_bo_put_tile_region(dev, *old_tile, fence);
1140 *old_tile = new_tile;
1141 }
1142
1143 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)1144 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict,
1145 struct ttm_operation_ctx *ctx,
1146 struct ttm_resource *new_reg,
1147 struct ttm_place *hop)
1148 {
1149 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1150 struct nouveau_bo *nvbo = nouveau_bo(bo);
1151 struct drm_gem_object *obj = &bo->base;
1152 struct ttm_resource *old_reg = bo->resource;
1153 struct nouveau_drm_tile *new_tile = NULL;
1154 int ret = 0;
1155
1156 if (new_reg->mem_type == TTM_PL_TT) {
1157 ret = nouveau_ttm_tt_bind(bo->bdev, bo->ttm, new_reg);
1158 if (ret)
1159 return ret;
1160 }
1161
1162 drm_gpuvm_bo_gem_evict(obj, evict);
1163 nouveau_bo_move_ntfy(bo, new_reg);
1164 ret = ttm_bo_wait_ctx(bo, ctx);
1165 if (ret)
1166 goto out_ntfy;
1167
1168 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1169 ret = nouveau_bo_vm_bind(bo, new_reg, &new_tile);
1170 if (ret)
1171 goto out_ntfy;
1172 }
1173
1174 /* Fake bo copy. */
1175 if (!old_reg || (old_reg->mem_type == TTM_PL_SYSTEM &&
1176 !bo->ttm)) {
1177 ttm_bo_move_null(bo, new_reg);
1178 goto out;
1179 }
1180
1181 if (old_reg->mem_type == TTM_PL_SYSTEM &&
1182 new_reg->mem_type == TTM_PL_TT) {
1183 ttm_bo_move_null(bo, new_reg);
1184 goto out;
1185 }
1186
1187 if (old_reg->mem_type == TTM_PL_TT &&
1188 new_reg->mem_type == TTM_PL_SYSTEM) {
1189 nouveau_ttm_tt_unbind(bo->bdev, bo->ttm);
1190 ttm_resource_free(bo, &bo->resource);
1191 ttm_bo_assign_mem(bo, new_reg);
1192 goto out;
1193 }
1194
1195 /* Hardware assisted copy. */
1196 if (drm->ttm.move) {
1197 if ((old_reg->mem_type == TTM_PL_SYSTEM &&
1198 new_reg->mem_type == TTM_PL_VRAM) ||
1199 (old_reg->mem_type == TTM_PL_VRAM &&
1200 new_reg->mem_type == TTM_PL_SYSTEM)) {
1201 hop->fpfn = 0;
1202 hop->lpfn = 0;
1203 hop->mem_type = TTM_PL_TT;
1204 hop->flags = 0;
1205 return -EMULTIHOP;
1206 }
1207 ret = nouveau_bo_move_m2mf(bo, evict, ctx,
1208 new_reg);
1209 } else
1210 ret = -ENODEV;
1211
1212 if (ret) {
1213 /* Fallback to software copy. */
1214 ret = ttm_bo_move_memcpy(bo, ctx, new_reg);
1215 }
1216
1217 out:
1218 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1219 if (ret)
1220 nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1221 else
1222 nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1223 }
1224 out_ntfy:
1225 if (ret) {
1226 nouveau_bo_move_ntfy(bo, bo->resource);
1227 drm_gpuvm_bo_gem_evict(obj, !evict);
1228 }
1229 return ret;
1230 }
1231
1232 static void
nouveau_ttm_io_mem_free_locked(struct nouveau_drm * drm,struct ttm_resource * reg)1233 nouveau_ttm_io_mem_free_locked(struct nouveau_drm *drm,
1234 struct ttm_resource *reg)
1235 {
1236 struct nouveau_mem *mem = nouveau_mem(reg);
1237
1238 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1239 switch (reg->mem_type) {
1240 case TTM_PL_TT:
1241 if (mem->kind)
1242 nvif_object_unmap_handle(&mem->mem.object);
1243 break;
1244 case TTM_PL_VRAM:
1245 nvif_object_unmap_handle(&mem->mem.object);
1246 break;
1247 default:
1248 break;
1249 }
1250 }
1251 }
1252
1253 static int
nouveau_ttm_io_mem_reserve(struct ttm_device * bdev,struct ttm_resource * reg)1254 nouveau_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *reg)
1255 {
1256 struct nouveau_drm *drm = nouveau_bdev(bdev);
1257 struct nvkm_device *device = nvxx_device(drm);
1258 struct nouveau_mem *mem = nouveau_mem(reg);
1259 struct nvif_mmu *mmu = &drm->client.mmu;
1260 int ret;
1261
1262 mutex_lock(&drm->ttm.io_reserve_mutex);
1263 retry:
1264 switch (reg->mem_type) {
1265 case TTM_PL_SYSTEM:
1266 /* System memory */
1267 ret = 0;
1268 goto out;
1269 case TTM_PL_TT:
1270 #if IS_ENABLED(CONFIG_AGP)
1271 if (drm->agp.bridge) {
1272 reg->bus.offset = (reg->start << PAGE_SHIFT) +
1273 drm->agp.base;
1274 reg->bus.is_iomem = !drm->agp.cma;
1275 reg->bus.caching = ttm_write_combined;
1276 }
1277 #endif
1278 if (drm->client.mem->oclass < NVIF_CLASS_MEM_NV50 ||
1279 !mem->kind) {
1280 /* untiled */
1281 ret = 0;
1282 break;
1283 }
1284 fallthrough; /* tiled memory */
1285 case TTM_PL_VRAM:
1286 reg->bus.offset = (reg->start << PAGE_SHIFT) +
1287 device->func->resource_addr(device, NVKM_BAR1_FB);
1288 reg->bus.is_iomem = true;
1289
1290 /* Some BARs do not support being ioremapped WC */
1291 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
1292 mmu->type[drm->ttm.type_vram].type & NVIF_MEM_UNCACHED)
1293 reg->bus.caching = ttm_uncached;
1294 else
1295 reg->bus.caching = ttm_write_combined;
1296
1297 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1298 union {
1299 struct nv50_mem_map_v0 nv50;
1300 struct gf100_mem_map_v0 gf100;
1301 } args;
1302 u64 handle, length;
1303 u32 argc = 0;
1304
1305 switch (mem->mem.object.oclass) {
1306 case NVIF_CLASS_MEM_NV50:
1307 args.nv50.version = 0;
1308 args.nv50.ro = 0;
1309 args.nv50.kind = mem->kind;
1310 args.nv50.comp = mem->comp;
1311 argc = sizeof(args.nv50);
1312 break;
1313 case NVIF_CLASS_MEM_GF100:
1314 args.gf100.version = 0;
1315 args.gf100.ro = 0;
1316 args.gf100.kind = mem->kind;
1317 argc = sizeof(args.gf100);
1318 break;
1319 default:
1320 WARN_ON(1);
1321 break;
1322 }
1323
1324 ret = nvif_object_map_handle(&mem->mem.object,
1325 &args, argc,
1326 &handle, &length);
1327 if (ret != 1) {
1328 if (WARN_ON(ret == 0))
1329 ret = -EINVAL;
1330 goto out;
1331 }
1332
1333 reg->bus.offset = handle;
1334 }
1335 ret = 0;
1336 break;
1337 default:
1338 ret = -EINVAL;
1339 }
1340
1341 out:
1342 if (ret == -ENOSPC) {
1343 struct nouveau_bo *nvbo;
1344
1345 nvbo = list_first_entry_or_null(&drm->ttm.io_reserve_lru,
1346 typeof(*nvbo),
1347 io_reserve_lru);
1348 if (nvbo) {
1349 list_del_init(&nvbo->io_reserve_lru);
1350 drm_vma_node_unmap(&nvbo->bo.base.vma_node,
1351 bdev->dev_mapping);
1352 nouveau_ttm_io_mem_free_locked(drm, nvbo->bo.resource);
1353 nvbo->bo.resource->bus.offset = 0;
1354 nvbo->bo.resource->bus.addr = NULL;
1355 goto retry;
1356 }
1357
1358 }
1359 mutex_unlock(&drm->ttm.io_reserve_mutex);
1360 return ret;
1361 }
1362
1363 static void
nouveau_ttm_io_mem_free(struct ttm_device * bdev,struct ttm_resource * reg)1364 nouveau_ttm_io_mem_free(struct ttm_device *bdev, struct ttm_resource *reg)
1365 {
1366 struct nouveau_drm *drm = nouveau_bdev(bdev);
1367
1368 mutex_lock(&drm->ttm.io_reserve_mutex);
1369 nouveau_ttm_io_mem_free_locked(drm, reg);
1370 mutex_unlock(&drm->ttm.io_reserve_mutex);
1371 }
1372
nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object * bo)1373 vm_fault_t nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1374 {
1375 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1376 struct nouveau_bo *nvbo = nouveau_bo(bo);
1377 struct nvkm_device *device = nvxx_device(drm);
1378 u32 mappable = device->func->resource_size(device, NVKM_BAR1_FB) >> PAGE_SHIFT;
1379 int i, ret;
1380
1381 /* as long as the bo isn't in vram, and isn't tiled, we've got
1382 * nothing to do here.
1383 */
1384 if (bo->resource->mem_type != TTM_PL_VRAM) {
1385 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1386 !nvbo->kind)
1387 return 0;
1388
1389 if (bo->resource->mem_type != TTM_PL_SYSTEM)
1390 return 0;
1391
1392 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART, 0);
1393
1394 } else {
1395 /* make sure bo is in mappable vram */
1396 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1397 bo->resource->start + PFN_UP(bo->resource->size) < mappable)
1398 return 0;
1399
1400 for (i = 0; i < nvbo->placement.num_placement; ++i) {
1401 nvbo->placements[i].fpfn = 0;
1402 nvbo->placements[i].lpfn = mappable;
1403 }
1404
1405 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_VRAM, 0);
1406 }
1407
1408 ret = nouveau_bo_validate(nvbo, false, false);
1409 if (unlikely(ret == -EBUSY || ret == -ERESTARTSYS))
1410 return VM_FAULT_NOPAGE;
1411 else if (unlikely(ret))
1412 return VM_FAULT_SIGBUS;
1413
1414 ttm_bo_move_to_lru_tail_unlocked(bo);
1415 return 0;
1416 }
1417
1418 static int
nouveau_ttm_tt_populate(struct ttm_device * bdev,struct ttm_tt * ttm,struct ttm_operation_ctx * ctx)1419 nouveau_ttm_tt_populate(struct ttm_device *bdev,
1420 struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1421 {
1422 struct ttm_tt *ttm_dma = (void *)ttm;
1423 struct nouveau_drm *drm;
1424 bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL);
1425
1426 if (ttm_tt_is_populated(ttm))
1427 return 0;
1428
1429 if (slave && ttm->sg) {
1430 drm_prime_sg_to_dma_addr_array(ttm->sg, ttm_dma->dma_address,
1431 ttm->num_pages);
1432 return 0;
1433 }
1434
1435 drm = nouveau_bdev(bdev);
1436
1437 return ttm_pool_alloc(&drm->ttm.bdev.pool, ttm, ctx);
1438 }
1439
1440 static void
nouveau_ttm_tt_unpopulate(struct ttm_device * bdev,struct ttm_tt * ttm)1441 nouveau_ttm_tt_unpopulate(struct ttm_device *bdev,
1442 struct ttm_tt *ttm)
1443 {
1444 struct nouveau_drm *drm;
1445 bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL);
1446
1447 if (slave)
1448 return;
1449
1450 nouveau_ttm_tt_unbind(bdev, ttm);
1451
1452 drm = nouveau_bdev(bdev);
1453
1454 return ttm_pool_free(&drm->ttm.bdev.pool, ttm);
1455 }
1456
1457 static void
nouveau_ttm_tt_destroy(struct ttm_device * bdev,struct ttm_tt * ttm)1458 nouveau_ttm_tt_destroy(struct ttm_device *bdev,
1459 struct ttm_tt *ttm)
1460 {
1461 #if IS_ENABLED(CONFIG_AGP)
1462 struct nouveau_drm *drm = nouveau_bdev(bdev);
1463 if (drm->agp.bridge) {
1464 ttm_agp_destroy(ttm);
1465 return;
1466 }
1467 #endif
1468 nouveau_sgdma_destroy(bdev, ttm);
1469 }
1470
1471 void
nouveau_bo_fence(struct nouveau_bo * nvbo,struct nouveau_fence * fence,bool exclusive)1472 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1473 {
1474 struct dma_resv *resv = nvbo->bo.base.resv;
1475
1476 if (!fence)
1477 return;
1478
1479 dma_resv_add_fence(resv, &fence->base, exclusive ?
1480 DMA_RESV_USAGE_WRITE : DMA_RESV_USAGE_READ);
1481 }
1482
1483 static void
nouveau_bo_delete_mem_notify(struct ttm_buffer_object * bo)1484 nouveau_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1485 {
1486 nouveau_bo_move_ntfy(bo, NULL);
1487 }
1488
1489 struct ttm_device_funcs nouveau_bo_driver = {
1490 .ttm_tt_create = &nouveau_ttm_tt_create,
1491 .ttm_tt_populate = &nouveau_ttm_tt_populate,
1492 .ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1493 .ttm_tt_destroy = &nouveau_ttm_tt_destroy,
1494 .eviction_valuable = ttm_bo_eviction_valuable,
1495 .evict_flags = nouveau_bo_evict_flags,
1496 .delete_mem_notify = nouveau_bo_delete_mem_notify,
1497 .move = nouveau_bo_move,
1498 .io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1499 .io_mem_free = &nouveau_ttm_io_mem_free,
1500 };
1501