xref: /linux/drivers/gpu/drm/nouveau/nouveau_bo.c (revision 959314c4)
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->client.device);
62 	struct nvkm_fb_tile *tile = &fb->tile.region[i];
63 
64 	nouveau_fence_unref(&reg->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->client.device);
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 	/*
148 	 * If nouveau_bo_new() allocated this buffer, the GEM object was never
149 	 * initialized, so don't attempt to release it.
150 	 */
151 	if (bo->base.dev) {
152 		/* Gem objects not being shared with other VMs get their
153 		 * dma_resv from a root GEM object.
154 		 */
155 		if (nvbo->no_share)
156 			drm_gem_object_put(nvbo->r_obj);
157 
158 		drm_gem_object_release(&bo->base);
159 	} else {
160 		dma_resv_fini(&bo->base._resv);
161 	}
162 
163 	kfree(nvbo);
164 }
165 
166 static inline u64
roundup_64(u64 x,u32 y)167 roundup_64(u64 x, u32 y)
168 {
169 	x += y - 1;
170 	do_div(x, y);
171 	return x * y;
172 }
173 
174 static void
nouveau_bo_fixup_align(struct nouveau_bo * nvbo,int * align,u64 * size)175 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, int *align, u64 *size)
176 {
177 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
178 	struct nvif_device *device = &drm->client.device;
179 
180 	if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
181 		if (nvbo->mode) {
182 			if (device->info.chipset >= 0x40) {
183 				*align = 65536;
184 				*size = roundup_64(*size, 64 * nvbo->mode);
185 
186 			} else if (device->info.chipset >= 0x30) {
187 				*align = 32768;
188 				*size = roundup_64(*size, 64 * nvbo->mode);
189 
190 			} else if (device->info.chipset >= 0x20) {
191 				*align = 16384;
192 				*size = roundup_64(*size, 64 * nvbo->mode);
193 
194 			} else if (device->info.chipset >= 0x10) {
195 				*align = 16384;
196 				*size = roundup_64(*size, 32 * nvbo->mode);
197 			}
198 		}
199 	} else {
200 		*size = roundup_64(*size, (1 << nvbo->page));
201 		*align = max((1 <<  nvbo->page), *align);
202 	}
203 
204 	*size = roundup_64(*size, PAGE_SIZE);
205 }
206 
207 struct nouveau_bo *
nouveau_bo_alloc(struct nouveau_cli * cli,u64 * size,int * align,u32 domain,u32 tile_mode,u32 tile_flags,bool internal)208 nouveau_bo_alloc(struct nouveau_cli *cli, u64 *size, int *align, u32 domain,
209 		 u32 tile_mode, u32 tile_flags, bool internal)
210 {
211 	struct nouveau_drm *drm = cli->drm;
212 	struct nouveau_bo *nvbo;
213 	struct nvif_mmu *mmu = &cli->mmu;
214 	struct nvif_vmm *vmm = &nouveau_cli_vmm(cli)->vmm;
215 	int i, pi = -1;
216 
217 	if (!*size) {
218 		NV_WARN(drm, "skipped size %016llx\n", *size);
219 		return ERR_PTR(-EINVAL);
220 	}
221 
222 	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
223 	if (!nvbo)
224 		return ERR_PTR(-ENOMEM);
225 
226 	INIT_LIST_HEAD(&nvbo->head);
227 	INIT_LIST_HEAD(&nvbo->entry);
228 	INIT_LIST_HEAD(&nvbo->vma_list);
229 	nvbo->bo.bdev = &drm->ttm.bdev;
230 
231 	/* This is confusing, and doesn't actually mean we want an uncached
232 	 * mapping, but is what NOUVEAU_GEM_DOMAIN_COHERENT gets translated
233 	 * into in nouveau_gem_new().
234 	 */
235 	if (domain & NOUVEAU_GEM_DOMAIN_COHERENT) {
236 		/* Determine if we can get a cache-coherent map, forcing
237 		 * uncached mapping if we can't.
238 		 */
239 		if (!nouveau_drm_use_coherent_gpu_mapping(drm))
240 			nvbo->force_coherent = true;
241 	}
242 
243 	nvbo->contig = !(tile_flags & NOUVEAU_GEM_TILE_NONCONTIG);
244 
245 	if (cli->device.info.family >= NV_DEVICE_INFO_V0_FERMI) {
246 		nvbo->kind = (tile_flags & 0x0000ff00) >> 8;
247 		if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
248 			kfree(nvbo);
249 			return ERR_PTR(-EINVAL);
250 		}
251 
252 		nvbo->comp = mmu->kind[nvbo->kind] != nvbo->kind;
253 	} else if (cli->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
254 		nvbo->kind = (tile_flags & 0x00007f00) >> 8;
255 		nvbo->comp = (tile_flags & 0x00030000) >> 16;
256 		if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
257 			kfree(nvbo);
258 			return ERR_PTR(-EINVAL);
259 		}
260 	} else {
261 		nvbo->zeta = (tile_flags & 0x00000007);
262 	}
263 	nvbo->mode = tile_mode;
264 
265 	if (!nouveau_cli_uvmm(cli) || internal) {
266 		/* Determine the desirable target GPU page size for the buffer. */
267 		for (i = 0; i < vmm->page_nr; i++) {
268 			/* Because we cannot currently allow VMM maps to fail
269 			 * during buffer migration, we need to determine page
270 			 * size for the buffer up-front, and pre-allocate its
271 			 * page tables.
272 			 *
273 			 * Skip page sizes that can't support needed domains.
274 			 */
275 			if (cli->device.info.family > NV_DEVICE_INFO_V0_CURIE &&
276 			    (domain & NOUVEAU_GEM_DOMAIN_VRAM) && !vmm->page[i].vram)
277 				continue;
278 			if ((domain & NOUVEAU_GEM_DOMAIN_GART) &&
279 			    (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
280 				continue;
281 
282 			/* Select this page size if it's the first that supports
283 			 * the potential memory domains, or when it's compatible
284 			 * with the requested compression settings.
285 			 */
286 			if (pi < 0 || !nvbo->comp || vmm->page[i].comp)
287 				pi = i;
288 
289 			/* Stop once the buffer is larger than the current page size. */
290 			if (*size >= 1ULL << vmm->page[i].shift)
291 				break;
292 		}
293 
294 		if (WARN_ON(pi < 0)) {
295 			kfree(nvbo);
296 			return ERR_PTR(-EINVAL);
297 		}
298 
299 		/* Disable compression if suitable settings couldn't be found. */
300 		if (nvbo->comp && !vmm->page[pi].comp) {
301 			if (mmu->object.oclass >= NVIF_CLASS_MMU_GF100)
302 				nvbo->kind = mmu->kind[nvbo->kind];
303 			nvbo->comp = 0;
304 		}
305 		nvbo->page = vmm->page[pi].shift;
306 	} else {
307 		/* Determine the desirable target GPU page size for the buffer. */
308 		for (i = 0; i < vmm->page_nr; i++) {
309 			/* Because we cannot currently allow VMM maps to fail
310 			 * during buffer migration, we need to determine page
311 			 * size for the buffer up-front, and pre-allocate its
312 			 * page tables.
313 			 *
314 			 * Skip page sizes that can't support needed domains.
315 			 */
316 			if ((domain & NOUVEAU_GEM_DOMAIN_VRAM) && !vmm->page[i].vram)
317 				continue;
318 			if ((domain & NOUVEAU_GEM_DOMAIN_GART) &&
319 			    (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
320 				continue;
321 
322 			/* pick the last one as it will be smallest. */
323 			pi = i;
324 
325 			/* Stop once the buffer is larger than the current page size. */
326 			if (*size >= 1ULL << vmm->page[i].shift)
327 				break;
328 		}
329 		if (WARN_ON(pi < 0)) {
330 			kfree(nvbo);
331 			return ERR_PTR(-EINVAL);
332 		}
333 		nvbo->page = vmm->page[pi].shift;
334 	}
335 
336 	nouveau_bo_fixup_align(nvbo, align, size);
337 
338 	return nvbo;
339 }
340 
341 int
nouveau_bo_init(struct nouveau_bo * nvbo,u64 size,int align,u32 domain,struct sg_table * sg,struct dma_resv * robj)342 nouveau_bo_init(struct nouveau_bo *nvbo, u64 size, int align, u32 domain,
343 		struct sg_table *sg, struct dma_resv *robj)
344 {
345 	int type = sg ? ttm_bo_type_sg : ttm_bo_type_device;
346 	int ret;
347 	struct ttm_operation_ctx ctx = {
348 		.interruptible = false,
349 		.no_wait_gpu = false,
350 		.resv = robj,
351 	};
352 
353 	nouveau_bo_placement_set(nvbo, domain, 0);
354 	INIT_LIST_HEAD(&nvbo->io_reserve_lru);
355 
356 	ret = ttm_bo_init_reserved(nvbo->bo.bdev, &nvbo->bo, type,
357 				   &nvbo->placement, align >> PAGE_SHIFT, &ctx,
358 				   sg, robj, nouveau_bo_del_ttm);
359 	if (ret) {
360 		/* ttm will call nouveau_bo_del_ttm if it fails.. */
361 		return ret;
362 	}
363 
364 	if (!robj)
365 		ttm_bo_unreserve(&nvbo->bo);
366 
367 	return 0;
368 }
369 
370 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)371 nouveau_bo_new(struct nouveau_cli *cli, u64 size, int align,
372 	       uint32_t domain, uint32_t tile_mode, uint32_t tile_flags,
373 	       struct sg_table *sg, struct dma_resv *robj,
374 	       struct nouveau_bo **pnvbo)
375 {
376 	struct nouveau_bo *nvbo;
377 	int ret;
378 
379 	nvbo = nouveau_bo_alloc(cli, &size, &align, domain, tile_mode,
380 				tile_flags, true);
381 	if (IS_ERR(nvbo))
382 		return PTR_ERR(nvbo);
383 
384 	nvbo->bo.base.size = size;
385 	dma_resv_init(&nvbo->bo.base._resv);
386 	drm_vma_node_reset(&nvbo->bo.base.vma_node);
387 
388 	/* This must be called before ttm_bo_init_reserved(). Subsequent
389 	 * bo_move() callbacks might already iterate the GEMs GPUVA list.
390 	 */
391 	drm_gem_gpuva_init(&nvbo->bo.base);
392 
393 	ret = nouveau_bo_init(nvbo, size, align, domain, sg, robj);
394 	if (ret)
395 		return ret;
396 
397 	*pnvbo = nvbo;
398 	return 0;
399 }
400 
401 static void
set_placement_range(struct nouveau_bo * nvbo,uint32_t domain)402 set_placement_range(struct nouveau_bo *nvbo, uint32_t domain)
403 {
404 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
405 	u64 vram_size = drm->client.device.info.ram_size;
406 	unsigned i, fpfn, lpfn;
407 
408 	if (drm->client.device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
409 	    nvbo->mode && (domain & NOUVEAU_GEM_DOMAIN_VRAM) &&
410 	    nvbo->bo.base.size < vram_size / 4) {
411 		/*
412 		 * Make sure that the color and depth buffers are handled
413 		 * by independent memory controller units. Up to a 9x
414 		 * speed up when alpha-blending and depth-test are enabled
415 		 * at the same time.
416 		 */
417 		if (nvbo->zeta) {
418 			fpfn = (vram_size / 2) >> PAGE_SHIFT;
419 			lpfn = ~0;
420 		} else {
421 			fpfn = 0;
422 			lpfn = (vram_size / 2) >> PAGE_SHIFT;
423 		}
424 		for (i = 0; i < nvbo->placement.num_placement; ++i) {
425 			nvbo->placements[i].fpfn = fpfn;
426 			nvbo->placements[i].lpfn = lpfn;
427 		}
428 	}
429 }
430 
431 void
nouveau_bo_placement_set(struct nouveau_bo * nvbo,uint32_t domain,uint32_t busy)432 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t domain,
433 			 uint32_t busy)
434 {
435 	unsigned int *n = &nvbo->placement.num_placement;
436 	struct ttm_place *pl = nvbo->placements;
437 
438 	domain |= busy;
439 
440 	*n = 0;
441 	if (domain & NOUVEAU_GEM_DOMAIN_VRAM) {
442 		pl[*n].mem_type = TTM_PL_VRAM;
443 		pl[*n].flags = busy & NOUVEAU_GEM_DOMAIN_VRAM ?
444 			TTM_PL_FLAG_FALLBACK : 0;
445 		(*n)++;
446 	}
447 	if (domain & NOUVEAU_GEM_DOMAIN_GART) {
448 		pl[*n].mem_type = TTM_PL_TT;
449 		pl[*n].flags = busy & NOUVEAU_GEM_DOMAIN_GART ?
450 			TTM_PL_FLAG_FALLBACK : 0;
451 		(*n)++;
452 	}
453 	if (domain & NOUVEAU_GEM_DOMAIN_CPU) {
454 		pl[*n].mem_type = TTM_PL_SYSTEM;
455 		pl[*n].flags = busy & NOUVEAU_GEM_DOMAIN_CPU ?
456 			TTM_PL_FLAG_FALLBACK : 0;
457 		(*n)++;
458 	}
459 
460 	nvbo->placement.placement = nvbo->placements;
461 	set_placement_range(nvbo, domain);
462 }
463 
nouveau_bo_pin_locked(struct nouveau_bo * nvbo,uint32_t domain,bool contig)464 int nouveau_bo_pin_locked(struct nouveau_bo *nvbo, uint32_t domain, bool contig)
465 {
466 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
467 	struct ttm_buffer_object *bo = &nvbo->bo;
468 	bool force = false, evict = false;
469 	int ret = 0;
470 
471 	dma_resv_assert_held(bo->base.resv);
472 
473 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
474 	    domain == NOUVEAU_GEM_DOMAIN_VRAM && contig) {
475 		if (!nvbo->contig) {
476 			nvbo->contig = true;
477 			force = true;
478 			evict = true;
479 		}
480 	}
481 
482 	if (nvbo->bo.pin_count) {
483 		bool error = evict;
484 
485 		switch (bo->resource->mem_type) {
486 		case TTM_PL_VRAM:
487 			error |= !(domain & NOUVEAU_GEM_DOMAIN_VRAM);
488 			break;
489 		case TTM_PL_TT:
490 			error |= !(domain & NOUVEAU_GEM_DOMAIN_GART);
491 			break;
492 		default:
493 			break;
494 		}
495 
496 		if (error) {
497 			NV_ERROR(drm, "bo %p pinned elsewhere: "
498 				      "0x%08x vs 0x%08x\n", bo,
499 				 bo->resource->mem_type, domain);
500 			ret = -EBUSY;
501 		}
502 		ttm_bo_pin(&nvbo->bo);
503 		goto out;
504 	}
505 
506 	if (evict) {
507 		nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART, 0);
508 		ret = nouveau_bo_validate(nvbo, false, false);
509 		if (ret)
510 			goto out;
511 	}
512 
513 	nouveau_bo_placement_set(nvbo, domain, 0);
514 	ret = nouveau_bo_validate(nvbo, false, false);
515 	if (ret)
516 		goto out;
517 
518 	ttm_bo_pin(&nvbo->bo);
519 
520 	switch (bo->resource->mem_type) {
521 	case TTM_PL_VRAM:
522 		drm->gem.vram_available -= bo->base.size;
523 		break;
524 	case TTM_PL_TT:
525 		drm->gem.gart_available -= bo->base.size;
526 		break;
527 	default:
528 		break;
529 	}
530 
531 out:
532 	if (force && ret)
533 		nvbo->contig = false;
534 	return ret;
535 }
536 
nouveau_bo_unpin_locked(struct nouveau_bo * nvbo)537 void nouveau_bo_unpin_locked(struct nouveau_bo *nvbo)
538 {
539 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
540 	struct ttm_buffer_object *bo = &nvbo->bo;
541 
542 	dma_resv_assert_held(bo->base.resv);
543 
544 	ttm_bo_unpin(&nvbo->bo);
545 	if (!nvbo->bo.pin_count) {
546 		switch (bo->resource->mem_type) {
547 		case TTM_PL_VRAM:
548 			drm->gem.vram_available += bo->base.size;
549 			break;
550 		case TTM_PL_TT:
551 			drm->gem.gart_available += bo->base.size;
552 			break;
553 		default:
554 			break;
555 		}
556 	}
557 }
558 
nouveau_bo_pin(struct nouveau_bo * nvbo,uint32_t domain,bool contig)559 int nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t domain, bool contig)
560 {
561 	struct ttm_buffer_object *bo = &nvbo->bo;
562 	int ret;
563 
564 	ret = ttm_bo_reserve(bo, false, false, NULL);
565 	if (ret)
566 		return ret;
567 	ret = nouveau_bo_pin_locked(nvbo, domain, contig);
568 	ttm_bo_unreserve(bo);
569 
570 	return ret;
571 }
572 
nouveau_bo_unpin(struct nouveau_bo * nvbo)573 int nouveau_bo_unpin(struct nouveau_bo *nvbo)
574 {
575 	struct ttm_buffer_object *bo = &nvbo->bo;
576 	int ret;
577 
578 	ret = ttm_bo_reserve(bo, false, false, NULL);
579 	if (ret)
580 		return ret;
581 	nouveau_bo_unpin_locked(nvbo);
582 	ttm_bo_unreserve(bo);
583 
584 	return 0;
585 }
586 
587 int
nouveau_bo_map(struct nouveau_bo * nvbo)588 nouveau_bo_map(struct nouveau_bo *nvbo)
589 {
590 	int ret;
591 
592 	ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
593 	if (ret)
594 		return ret;
595 
596 	ret = ttm_bo_kmap(&nvbo->bo, 0, PFN_UP(nvbo->bo.base.size), &nvbo->kmap);
597 
598 	ttm_bo_unreserve(&nvbo->bo);
599 	return ret;
600 }
601 
602 void
nouveau_bo_unmap(struct nouveau_bo * nvbo)603 nouveau_bo_unmap(struct nouveau_bo *nvbo)
604 {
605 	if (!nvbo)
606 		return;
607 
608 	ttm_bo_kunmap(&nvbo->kmap);
609 }
610 
611 void
nouveau_bo_sync_for_device(struct nouveau_bo * nvbo)612 nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
613 {
614 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
615 	struct ttm_tt *ttm_dma = (struct ttm_tt *)nvbo->bo.ttm;
616 	int i, j;
617 
618 	if (!ttm_dma || !ttm_dma->dma_address)
619 		return;
620 	if (!ttm_dma->pages) {
621 		NV_DEBUG(drm, "ttm_dma 0x%p: pages NULL\n", ttm_dma);
622 		return;
623 	}
624 
625 	/* Don't waste time looping if the object is coherent */
626 	if (nvbo->force_coherent)
627 		return;
628 
629 	i = 0;
630 	while (i < ttm_dma->num_pages) {
631 		struct page *p = ttm_dma->pages[i];
632 		size_t num_pages = 1;
633 
634 		for (j = i + 1; j < ttm_dma->num_pages; ++j) {
635 			if (++p != ttm_dma->pages[j])
636 				break;
637 
638 			++num_pages;
639 		}
640 		dma_sync_single_for_device(drm->dev->dev,
641 					   ttm_dma->dma_address[i],
642 					   num_pages * PAGE_SIZE, DMA_TO_DEVICE);
643 		i += num_pages;
644 	}
645 }
646 
647 void
nouveau_bo_sync_for_cpu(struct nouveau_bo * nvbo)648 nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
649 {
650 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
651 	struct ttm_tt *ttm_dma = (struct ttm_tt *)nvbo->bo.ttm;
652 	int i, j;
653 
654 	if (!ttm_dma || !ttm_dma->dma_address)
655 		return;
656 	if (!ttm_dma->pages) {
657 		NV_DEBUG(drm, "ttm_dma 0x%p: pages NULL\n", ttm_dma);
658 		return;
659 	}
660 
661 	/* Don't waste time looping if the object is coherent */
662 	if (nvbo->force_coherent)
663 		return;
664 
665 	i = 0;
666 	while (i < ttm_dma->num_pages) {
667 		struct page *p = ttm_dma->pages[i];
668 		size_t num_pages = 1;
669 
670 		for (j = i + 1; j < ttm_dma->num_pages; ++j) {
671 			if (++p != ttm_dma->pages[j])
672 				break;
673 
674 			++num_pages;
675 		}
676 
677 		dma_sync_single_for_cpu(drm->dev->dev, ttm_dma->dma_address[i],
678 					num_pages * PAGE_SIZE, DMA_FROM_DEVICE);
679 		i += num_pages;
680 	}
681 }
682 
nouveau_bo_add_io_reserve_lru(struct ttm_buffer_object * bo)683 void nouveau_bo_add_io_reserve_lru(struct ttm_buffer_object *bo)
684 {
685 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
686 	struct nouveau_bo *nvbo = nouveau_bo(bo);
687 
688 	mutex_lock(&drm->ttm.io_reserve_mutex);
689 	list_move_tail(&nvbo->io_reserve_lru, &drm->ttm.io_reserve_lru);
690 	mutex_unlock(&drm->ttm.io_reserve_mutex);
691 }
692 
nouveau_bo_del_io_reserve_lru(struct ttm_buffer_object * bo)693 void nouveau_bo_del_io_reserve_lru(struct ttm_buffer_object *bo)
694 {
695 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
696 	struct nouveau_bo *nvbo = nouveau_bo(bo);
697 
698 	mutex_lock(&drm->ttm.io_reserve_mutex);
699 	list_del_init(&nvbo->io_reserve_lru);
700 	mutex_unlock(&drm->ttm.io_reserve_mutex);
701 }
702 
703 int
nouveau_bo_validate(struct nouveau_bo * nvbo,bool interruptible,bool no_wait_gpu)704 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
705 		    bool no_wait_gpu)
706 {
707 	struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
708 	int ret;
709 
710 	ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, &ctx);
711 	if (ret)
712 		return ret;
713 
714 	nouveau_bo_sync_for_device(nvbo);
715 
716 	return 0;
717 }
718 
719 void
nouveau_bo_wr16(struct nouveau_bo * nvbo,unsigned index,u16 val)720 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
721 {
722 	bool is_iomem;
723 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
724 
725 	mem += index;
726 
727 	if (is_iomem)
728 		iowrite16_native(val, (void __force __iomem *)mem);
729 	else
730 		*mem = val;
731 }
732 
733 u32
nouveau_bo_rd32(struct nouveau_bo * nvbo,unsigned index)734 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
735 {
736 	bool is_iomem;
737 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
738 
739 	mem += index;
740 
741 	if (is_iomem)
742 		return ioread32_native((void __force __iomem *)mem);
743 	else
744 		return *mem;
745 }
746 
747 void
nouveau_bo_wr32(struct nouveau_bo * nvbo,unsigned index,u32 val)748 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
749 {
750 	bool is_iomem;
751 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
752 
753 	mem += index;
754 
755 	if (is_iomem)
756 		iowrite32_native(val, (void __force __iomem *)mem);
757 	else
758 		*mem = val;
759 }
760 
761 static struct ttm_tt *
nouveau_ttm_tt_create(struct ttm_buffer_object * bo,uint32_t page_flags)762 nouveau_ttm_tt_create(struct ttm_buffer_object *bo, uint32_t page_flags)
763 {
764 #if IS_ENABLED(CONFIG_AGP)
765 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
766 
767 	if (drm->agp.bridge) {
768 		return ttm_agp_tt_create(bo, drm->agp.bridge, page_flags);
769 	}
770 #endif
771 
772 	return nouveau_sgdma_create_ttm(bo, page_flags);
773 }
774 
775 static int
nouveau_ttm_tt_bind(struct ttm_device * bdev,struct ttm_tt * ttm,struct ttm_resource * reg)776 nouveau_ttm_tt_bind(struct ttm_device *bdev, struct ttm_tt *ttm,
777 		    struct ttm_resource *reg)
778 {
779 #if IS_ENABLED(CONFIG_AGP)
780 	struct nouveau_drm *drm = nouveau_bdev(bdev);
781 #endif
782 	if (!reg)
783 		return -EINVAL;
784 #if IS_ENABLED(CONFIG_AGP)
785 	if (drm->agp.bridge)
786 		return ttm_agp_bind(ttm, reg);
787 #endif
788 	return nouveau_sgdma_bind(bdev, ttm, reg);
789 }
790 
791 static void
nouveau_ttm_tt_unbind(struct ttm_device * bdev,struct ttm_tt * ttm)792 nouveau_ttm_tt_unbind(struct ttm_device *bdev, struct ttm_tt *ttm)
793 {
794 #if IS_ENABLED(CONFIG_AGP)
795 	struct nouveau_drm *drm = nouveau_bdev(bdev);
796 
797 	if (drm->agp.bridge) {
798 		ttm_agp_unbind(ttm);
799 		return;
800 	}
801 #endif
802 	nouveau_sgdma_unbind(bdev, ttm);
803 }
804 
805 static void
nouveau_bo_evict_flags(struct ttm_buffer_object * bo,struct ttm_placement * pl)806 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
807 {
808 	struct nouveau_bo *nvbo = nouveau_bo(bo);
809 
810 	switch (bo->resource->mem_type) {
811 	case TTM_PL_VRAM:
812 		nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART,
813 					 NOUVEAU_GEM_DOMAIN_CPU);
814 		break;
815 	default:
816 		nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_CPU, 0);
817 		break;
818 	}
819 
820 	*pl = nvbo->placement;
821 }
822 
823 static int
nouveau_bo_move_prep(struct nouveau_drm * drm,struct ttm_buffer_object * bo,struct ttm_resource * reg)824 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
825 		     struct ttm_resource *reg)
826 {
827 	struct nouveau_mem *old_mem = nouveau_mem(bo->resource);
828 	struct nouveau_mem *new_mem = nouveau_mem(reg);
829 	struct nvif_vmm *vmm = &drm->client.vmm.vmm;
830 	int ret;
831 
832 	ret = nvif_vmm_get(vmm, LAZY, false, old_mem->mem.page, 0,
833 			   old_mem->mem.size, &old_mem->vma[0]);
834 	if (ret)
835 		return ret;
836 
837 	ret = nvif_vmm_get(vmm, LAZY, false, new_mem->mem.page, 0,
838 			   new_mem->mem.size, &old_mem->vma[1]);
839 	if (ret)
840 		goto done;
841 
842 	ret = nouveau_mem_map(old_mem, vmm, &old_mem->vma[0]);
843 	if (ret)
844 		goto done;
845 
846 	ret = nouveau_mem_map(new_mem, vmm, &old_mem->vma[1]);
847 done:
848 	if (ret) {
849 		nvif_vmm_put(vmm, &old_mem->vma[1]);
850 		nvif_vmm_put(vmm, &old_mem->vma[0]);
851 	}
852 	return 0;
853 }
854 
855 static int
nouveau_bo_move_m2mf(struct ttm_buffer_object * bo,int evict,struct ttm_operation_ctx * ctx,struct ttm_resource * new_reg)856 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict,
857 		     struct ttm_operation_ctx *ctx,
858 		     struct ttm_resource *new_reg)
859 {
860 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
861 	struct nouveau_channel *chan = drm->ttm.chan;
862 	struct nouveau_cli *cli = (void *)chan->user.client;
863 	struct nouveau_fence *fence;
864 	int ret;
865 
866 	/* create temporary vmas for the transfer and attach them to the
867 	 * old nvkm_mem node, these will get cleaned up after ttm has
868 	 * destroyed the ttm_resource
869 	 */
870 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
871 		ret = nouveau_bo_move_prep(drm, bo, new_reg);
872 		if (ret)
873 			return ret;
874 	}
875 
876 	if (drm_drv_uses_atomic_modeset(drm->dev))
877 		mutex_lock(&cli->mutex);
878 	else
879 		mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
880 
881 	ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, ctx->interruptible);
882 	if (ret)
883 		goto out_unlock;
884 
885 	ret = drm->ttm.move(chan, bo, bo->resource, new_reg);
886 	if (ret)
887 		goto out_unlock;
888 
889 	ret = nouveau_fence_new(&fence, chan);
890 	if (ret)
891 		goto out_unlock;
892 
893 	/* TODO: figure out a better solution here
894 	 *
895 	 * wait on the fence here explicitly as going through
896 	 * ttm_bo_move_accel_cleanup somehow doesn't seem to do it.
897 	 *
898 	 * Without this the operation can timeout and we'll fallback to a
899 	 * software copy, which might take several minutes to finish.
900 	 */
901 	nouveau_fence_wait(fence, false, false);
902 	ret = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, false,
903 					new_reg);
904 	nouveau_fence_unref(&fence);
905 
906 out_unlock:
907 	mutex_unlock(&cli->mutex);
908 	return ret;
909 }
910 
911 void
nouveau_bo_move_init(struct nouveau_drm * drm)912 nouveau_bo_move_init(struct nouveau_drm *drm)
913 {
914 	static const struct _method_table {
915 		const char *name;
916 		int engine;
917 		s32 oclass;
918 		int (*exec)(struct nouveau_channel *,
919 			    struct ttm_buffer_object *,
920 			    struct ttm_resource *, struct ttm_resource *);
921 		int (*init)(struct nouveau_channel *, u32 handle);
922 	} _methods[] = {
923 		{  "COPY", 4, 0xc7b5, nve0_bo_move_copy, nve0_bo_move_init },
924 		{  "GRCE", 0, 0xc7b5, nve0_bo_move_copy, nvc0_bo_move_init },
925 		{  "COPY", 4, 0xc6b5, nve0_bo_move_copy, nve0_bo_move_init },
926 		{  "GRCE", 0, 0xc6b5, nve0_bo_move_copy, nvc0_bo_move_init },
927 		{  "COPY", 4, 0xc5b5, nve0_bo_move_copy, nve0_bo_move_init },
928 		{  "GRCE", 0, 0xc5b5, nve0_bo_move_copy, nvc0_bo_move_init },
929 		{  "COPY", 4, 0xc3b5, nve0_bo_move_copy, nve0_bo_move_init },
930 		{  "GRCE", 0, 0xc3b5, nve0_bo_move_copy, nvc0_bo_move_init },
931 		{  "COPY", 4, 0xc1b5, nve0_bo_move_copy, nve0_bo_move_init },
932 		{  "GRCE", 0, 0xc1b5, nve0_bo_move_copy, nvc0_bo_move_init },
933 		{  "COPY", 4, 0xc0b5, nve0_bo_move_copy, nve0_bo_move_init },
934 		{  "GRCE", 0, 0xc0b5, nve0_bo_move_copy, nvc0_bo_move_init },
935 		{  "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
936 		{  "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
937 		{  "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
938 		{  "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
939 		{ "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
940 		{ "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
941 		{  "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
942 		{ "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
943 		{  "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
944 		{  "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
945 		{  "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
946 		{},
947 	};
948 	const struct _method_table *mthd = _methods;
949 	const char *name = "CPU";
950 	int ret;
951 
952 	do {
953 		struct nouveau_channel *chan;
954 
955 		if (mthd->engine)
956 			chan = drm->cechan;
957 		else
958 			chan = drm->channel;
959 		if (chan == NULL)
960 			continue;
961 
962 		ret = nvif_object_ctor(&chan->user, "ttmBoMove",
963 				       mthd->oclass | (mthd->engine << 16),
964 				       mthd->oclass, NULL, 0,
965 				       &drm->ttm.copy);
966 		if (ret == 0) {
967 			ret = mthd->init(chan, drm->ttm.copy.handle);
968 			if (ret) {
969 				nvif_object_dtor(&drm->ttm.copy);
970 				continue;
971 			}
972 
973 			drm->ttm.move = mthd->exec;
974 			drm->ttm.chan = chan;
975 			name = mthd->name;
976 			break;
977 		}
978 	} while ((++mthd)->exec);
979 
980 	NV_INFO(drm, "MM: using %s for buffer copies\n", name);
981 }
982 
nouveau_bo_move_ntfy(struct ttm_buffer_object * bo,struct ttm_resource * new_reg)983 static void nouveau_bo_move_ntfy(struct ttm_buffer_object *bo,
984 				 struct ttm_resource *new_reg)
985 {
986 	struct nouveau_mem *mem = new_reg ? nouveau_mem(new_reg) : NULL;
987 	struct nouveau_bo *nvbo = nouveau_bo(bo);
988 	struct nouveau_vma *vma;
989 	long ret;
990 
991 	/* ttm can now (stupidly) pass the driver bos it didn't create... */
992 	if (bo->destroy != nouveau_bo_del_ttm)
993 		return;
994 
995 	nouveau_bo_del_io_reserve_lru(bo);
996 
997 	if (mem && new_reg->mem_type != TTM_PL_SYSTEM &&
998 	    mem->mem.page == nvbo->page) {
999 		list_for_each_entry(vma, &nvbo->vma_list, head) {
1000 			nouveau_vma_map(vma, mem);
1001 		}
1002 		nouveau_uvmm_bo_map_all(nvbo, mem);
1003 	} else {
1004 		list_for_each_entry(vma, &nvbo->vma_list, head) {
1005 			ret = dma_resv_wait_timeout(bo->base.resv,
1006 						    DMA_RESV_USAGE_BOOKKEEP,
1007 						    false, 15 * HZ);
1008 			WARN_ON(ret <= 0);
1009 			nouveau_vma_unmap(vma);
1010 		}
1011 		nouveau_uvmm_bo_unmap_all(nvbo);
1012 	}
1013 
1014 	if (new_reg)
1015 		nvbo->offset = (new_reg->start << PAGE_SHIFT);
1016 
1017 }
1018 
1019 static int
nouveau_bo_vm_bind(struct ttm_buffer_object * bo,struct ttm_resource * new_reg,struct nouveau_drm_tile ** new_tile)1020 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_resource *new_reg,
1021 		   struct nouveau_drm_tile **new_tile)
1022 {
1023 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1024 	struct drm_device *dev = drm->dev;
1025 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1026 	u64 offset = new_reg->start << PAGE_SHIFT;
1027 
1028 	*new_tile = NULL;
1029 	if (new_reg->mem_type != TTM_PL_VRAM)
1030 		return 0;
1031 
1032 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
1033 		*new_tile = nv10_bo_set_tiling(dev, offset, bo->base.size,
1034 					       nvbo->mode, nvbo->zeta);
1035 	}
1036 
1037 	return 0;
1038 }
1039 
1040 static void
nouveau_bo_vm_cleanup(struct ttm_buffer_object * bo,struct nouveau_drm_tile * new_tile,struct nouveau_drm_tile ** old_tile)1041 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1042 		      struct nouveau_drm_tile *new_tile,
1043 		      struct nouveau_drm_tile **old_tile)
1044 {
1045 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1046 	struct drm_device *dev = drm->dev;
1047 	struct dma_fence *fence;
1048 	int ret;
1049 
1050 	ret = dma_resv_get_singleton(bo->base.resv, DMA_RESV_USAGE_WRITE,
1051 				     &fence);
1052 	if (ret)
1053 		dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_WRITE,
1054 				      false, MAX_SCHEDULE_TIMEOUT);
1055 
1056 	nv10_bo_put_tile_region(dev, *old_tile, fence);
1057 	*old_tile = new_tile;
1058 }
1059 
1060 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)1061 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict,
1062 		struct ttm_operation_ctx *ctx,
1063 		struct ttm_resource *new_reg,
1064 		struct ttm_place *hop)
1065 {
1066 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1067 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1068 	struct drm_gem_object *obj = &bo->base;
1069 	struct ttm_resource *old_reg = bo->resource;
1070 	struct nouveau_drm_tile *new_tile = NULL;
1071 	int ret = 0;
1072 
1073 	if (new_reg->mem_type == TTM_PL_TT) {
1074 		ret = nouveau_ttm_tt_bind(bo->bdev, bo->ttm, new_reg);
1075 		if (ret)
1076 			return ret;
1077 	}
1078 
1079 	drm_gpuvm_bo_gem_evict(obj, evict);
1080 	nouveau_bo_move_ntfy(bo, new_reg);
1081 	ret = ttm_bo_wait_ctx(bo, ctx);
1082 	if (ret)
1083 		goto out_ntfy;
1084 
1085 	if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1086 		ret = nouveau_bo_vm_bind(bo, new_reg, &new_tile);
1087 		if (ret)
1088 			goto out_ntfy;
1089 	}
1090 
1091 	/* Fake bo copy. */
1092 	if (!old_reg || (old_reg->mem_type == TTM_PL_SYSTEM &&
1093 			 !bo->ttm)) {
1094 		ttm_bo_move_null(bo, new_reg);
1095 		goto out;
1096 	}
1097 
1098 	if (old_reg->mem_type == TTM_PL_SYSTEM &&
1099 	    new_reg->mem_type == TTM_PL_TT) {
1100 		ttm_bo_move_null(bo, new_reg);
1101 		goto out;
1102 	}
1103 
1104 	if (old_reg->mem_type == TTM_PL_TT &&
1105 	    new_reg->mem_type == TTM_PL_SYSTEM) {
1106 		nouveau_ttm_tt_unbind(bo->bdev, bo->ttm);
1107 		ttm_resource_free(bo, &bo->resource);
1108 		ttm_bo_assign_mem(bo, new_reg);
1109 		goto out;
1110 	}
1111 
1112 	/* Hardware assisted copy. */
1113 	if (drm->ttm.move) {
1114 		if ((old_reg->mem_type == TTM_PL_SYSTEM &&
1115 		     new_reg->mem_type == TTM_PL_VRAM) ||
1116 		    (old_reg->mem_type == TTM_PL_VRAM &&
1117 		     new_reg->mem_type == TTM_PL_SYSTEM)) {
1118 			hop->fpfn = 0;
1119 			hop->lpfn = 0;
1120 			hop->mem_type = TTM_PL_TT;
1121 			hop->flags = 0;
1122 			return -EMULTIHOP;
1123 		}
1124 		ret = nouveau_bo_move_m2mf(bo, evict, ctx,
1125 					   new_reg);
1126 	} else
1127 		ret = -ENODEV;
1128 
1129 	if (ret) {
1130 		/* Fallback to software copy. */
1131 		ret = ttm_bo_move_memcpy(bo, ctx, new_reg);
1132 	}
1133 
1134 out:
1135 	if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1136 		if (ret)
1137 			nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1138 		else
1139 			nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1140 	}
1141 out_ntfy:
1142 	if (ret) {
1143 		nouveau_bo_move_ntfy(bo, bo->resource);
1144 		drm_gpuvm_bo_gem_evict(obj, !evict);
1145 	}
1146 	return ret;
1147 }
1148 
1149 static void
nouveau_ttm_io_mem_free_locked(struct nouveau_drm * drm,struct ttm_resource * reg)1150 nouveau_ttm_io_mem_free_locked(struct nouveau_drm *drm,
1151 			       struct ttm_resource *reg)
1152 {
1153 	struct nouveau_mem *mem = nouveau_mem(reg);
1154 
1155 	if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1156 		switch (reg->mem_type) {
1157 		case TTM_PL_TT:
1158 			if (mem->kind)
1159 				nvif_object_unmap_handle(&mem->mem.object);
1160 			break;
1161 		case TTM_PL_VRAM:
1162 			nvif_object_unmap_handle(&mem->mem.object);
1163 			break;
1164 		default:
1165 			break;
1166 		}
1167 	}
1168 }
1169 
1170 static int
nouveau_ttm_io_mem_reserve(struct ttm_device * bdev,struct ttm_resource * reg)1171 nouveau_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *reg)
1172 {
1173 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1174 	struct nvkm_device *device = nvxx_device(&drm->client.device);
1175 	struct nouveau_mem *mem = nouveau_mem(reg);
1176 	struct nvif_mmu *mmu = &drm->client.mmu;
1177 	int ret;
1178 
1179 	mutex_lock(&drm->ttm.io_reserve_mutex);
1180 retry:
1181 	switch (reg->mem_type) {
1182 	case TTM_PL_SYSTEM:
1183 		/* System memory */
1184 		ret = 0;
1185 		goto out;
1186 	case TTM_PL_TT:
1187 #if IS_ENABLED(CONFIG_AGP)
1188 		if (drm->agp.bridge) {
1189 			reg->bus.offset = (reg->start << PAGE_SHIFT) +
1190 				drm->agp.base;
1191 			reg->bus.is_iomem = !drm->agp.cma;
1192 			reg->bus.caching = ttm_write_combined;
1193 		}
1194 #endif
1195 		if (drm->client.mem->oclass < NVIF_CLASS_MEM_NV50 ||
1196 		    !mem->kind) {
1197 			/* untiled */
1198 			ret = 0;
1199 			break;
1200 		}
1201 		fallthrough;	/* tiled memory */
1202 	case TTM_PL_VRAM:
1203 		reg->bus.offset = (reg->start << PAGE_SHIFT) +
1204 			device->func->resource_addr(device, 1);
1205 		reg->bus.is_iomem = true;
1206 
1207 		/* Some BARs do not support being ioremapped WC */
1208 		if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
1209 		    mmu->type[drm->ttm.type_vram].type & NVIF_MEM_UNCACHED)
1210 			reg->bus.caching = ttm_uncached;
1211 		else
1212 			reg->bus.caching = ttm_write_combined;
1213 
1214 		if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1215 			union {
1216 				struct nv50_mem_map_v0 nv50;
1217 				struct gf100_mem_map_v0 gf100;
1218 			} args;
1219 			u64 handle, length;
1220 			u32 argc = 0;
1221 
1222 			switch (mem->mem.object.oclass) {
1223 			case NVIF_CLASS_MEM_NV50:
1224 				args.nv50.version = 0;
1225 				args.nv50.ro = 0;
1226 				args.nv50.kind = mem->kind;
1227 				args.nv50.comp = mem->comp;
1228 				argc = sizeof(args.nv50);
1229 				break;
1230 			case NVIF_CLASS_MEM_GF100:
1231 				args.gf100.version = 0;
1232 				args.gf100.ro = 0;
1233 				args.gf100.kind = mem->kind;
1234 				argc = sizeof(args.gf100);
1235 				break;
1236 			default:
1237 				WARN_ON(1);
1238 				break;
1239 			}
1240 
1241 			ret = nvif_object_map_handle(&mem->mem.object,
1242 						     &args, argc,
1243 						     &handle, &length);
1244 			if (ret != 1) {
1245 				if (WARN_ON(ret == 0))
1246 					ret = -EINVAL;
1247 				goto out;
1248 			}
1249 
1250 			reg->bus.offset = handle;
1251 		}
1252 		ret = 0;
1253 		break;
1254 	default:
1255 		ret = -EINVAL;
1256 	}
1257 
1258 out:
1259 	if (ret == -ENOSPC) {
1260 		struct nouveau_bo *nvbo;
1261 
1262 		nvbo = list_first_entry_or_null(&drm->ttm.io_reserve_lru,
1263 						typeof(*nvbo),
1264 						io_reserve_lru);
1265 		if (nvbo) {
1266 			list_del_init(&nvbo->io_reserve_lru);
1267 			drm_vma_node_unmap(&nvbo->bo.base.vma_node,
1268 					   bdev->dev_mapping);
1269 			nouveau_ttm_io_mem_free_locked(drm, nvbo->bo.resource);
1270 			nvbo->bo.resource->bus.offset = 0;
1271 			nvbo->bo.resource->bus.addr = NULL;
1272 			goto retry;
1273 		}
1274 
1275 	}
1276 	mutex_unlock(&drm->ttm.io_reserve_mutex);
1277 	return ret;
1278 }
1279 
1280 static void
nouveau_ttm_io_mem_free(struct ttm_device * bdev,struct ttm_resource * reg)1281 nouveau_ttm_io_mem_free(struct ttm_device *bdev, struct ttm_resource *reg)
1282 {
1283 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1284 
1285 	mutex_lock(&drm->ttm.io_reserve_mutex);
1286 	nouveau_ttm_io_mem_free_locked(drm, reg);
1287 	mutex_unlock(&drm->ttm.io_reserve_mutex);
1288 }
1289 
nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object * bo)1290 vm_fault_t nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1291 {
1292 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1293 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1294 	struct nvkm_device *device = nvxx_device(&drm->client.device);
1295 	u32 mappable = device->func->resource_size(device, 1) >> PAGE_SHIFT;
1296 	int i, ret;
1297 
1298 	/* as long as the bo isn't in vram, and isn't tiled, we've got
1299 	 * nothing to do here.
1300 	 */
1301 	if (bo->resource->mem_type != TTM_PL_VRAM) {
1302 		if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1303 		    !nvbo->kind)
1304 			return 0;
1305 
1306 		if (bo->resource->mem_type != TTM_PL_SYSTEM)
1307 			return 0;
1308 
1309 		nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART, 0);
1310 
1311 	} else {
1312 		/* make sure bo is in mappable vram */
1313 		if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1314 		    bo->resource->start + PFN_UP(bo->resource->size) < mappable)
1315 			return 0;
1316 
1317 		for (i = 0; i < nvbo->placement.num_placement; ++i) {
1318 			nvbo->placements[i].fpfn = 0;
1319 			nvbo->placements[i].lpfn = mappable;
1320 		}
1321 
1322 		nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_VRAM, 0);
1323 	}
1324 
1325 	ret = nouveau_bo_validate(nvbo, false, false);
1326 	if (unlikely(ret == -EBUSY || ret == -ERESTARTSYS))
1327 		return VM_FAULT_NOPAGE;
1328 	else if (unlikely(ret))
1329 		return VM_FAULT_SIGBUS;
1330 
1331 	ttm_bo_move_to_lru_tail_unlocked(bo);
1332 	return 0;
1333 }
1334 
1335 static int
nouveau_ttm_tt_populate(struct ttm_device * bdev,struct ttm_tt * ttm,struct ttm_operation_ctx * ctx)1336 nouveau_ttm_tt_populate(struct ttm_device *bdev,
1337 			struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1338 {
1339 	struct ttm_tt *ttm_dma = (void *)ttm;
1340 	struct nouveau_drm *drm;
1341 	bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL);
1342 
1343 	if (ttm_tt_is_populated(ttm))
1344 		return 0;
1345 
1346 	if (slave && ttm->sg) {
1347 		drm_prime_sg_to_dma_addr_array(ttm->sg, ttm_dma->dma_address,
1348 					       ttm->num_pages);
1349 		return 0;
1350 	}
1351 
1352 	drm = nouveau_bdev(bdev);
1353 
1354 	return ttm_pool_alloc(&drm->ttm.bdev.pool, ttm, ctx);
1355 }
1356 
1357 static void
nouveau_ttm_tt_unpopulate(struct ttm_device * bdev,struct ttm_tt * ttm)1358 nouveau_ttm_tt_unpopulate(struct ttm_device *bdev,
1359 			  struct ttm_tt *ttm)
1360 {
1361 	struct nouveau_drm *drm;
1362 	bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL);
1363 
1364 	if (slave)
1365 		return;
1366 
1367 	nouveau_ttm_tt_unbind(bdev, ttm);
1368 
1369 	drm = nouveau_bdev(bdev);
1370 
1371 	return ttm_pool_free(&drm->ttm.bdev.pool, ttm);
1372 }
1373 
1374 static void
nouveau_ttm_tt_destroy(struct ttm_device * bdev,struct ttm_tt * ttm)1375 nouveau_ttm_tt_destroy(struct ttm_device *bdev,
1376 		       struct ttm_tt *ttm)
1377 {
1378 #if IS_ENABLED(CONFIG_AGP)
1379 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1380 	if (drm->agp.bridge) {
1381 		ttm_agp_destroy(ttm);
1382 		return;
1383 	}
1384 #endif
1385 	nouveau_sgdma_destroy(bdev, ttm);
1386 }
1387 
1388 void
nouveau_bo_fence(struct nouveau_bo * nvbo,struct nouveau_fence * fence,bool exclusive)1389 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1390 {
1391 	struct dma_resv *resv = nvbo->bo.base.resv;
1392 
1393 	if (!fence)
1394 		return;
1395 
1396 	dma_resv_add_fence(resv, &fence->base, exclusive ?
1397 			   DMA_RESV_USAGE_WRITE : DMA_RESV_USAGE_READ);
1398 }
1399 
1400 static void
nouveau_bo_delete_mem_notify(struct ttm_buffer_object * bo)1401 nouveau_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1402 {
1403 	nouveau_bo_move_ntfy(bo, NULL);
1404 }
1405 
1406 struct ttm_device_funcs nouveau_bo_driver = {
1407 	.ttm_tt_create = &nouveau_ttm_tt_create,
1408 	.ttm_tt_populate = &nouveau_ttm_tt_populate,
1409 	.ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1410 	.ttm_tt_destroy = &nouveau_ttm_tt_destroy,
1411 	.eviction_valuable = ttm_bo_eviction_valuable,
1412 	.evict_flags = nouveau_bo_evict_flags,
1413 	.delete_mem_notify = nouveau_bo_delete_mem_notify,
1414 	.move = nouveau_bo_move,
1415 	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1416 	.io_mem_free = &nouveau_ttm_io_mem_free,
1417 };
1418