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