xref: /linux/drivers/gpu/drm/nouveau/nouveau_bo.c (revision bd628c1bed7902ec1f24ba0fe70758949146abbe)
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 <linux/swiotlb.h>
32 
33 #include "nouveau_drv.h"
34 #include "nouveau_dma.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 /*
48  * NV10-NV40 tiling helpers
49  */
50 
51 static void
52 nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
53 			   u32 addr, u32 size, u32 pitch, u32 flags)
54 {
55 	struct nouveau_drm *drm = nouveau_drm(dev);
56 	int i = reg - drm->tile.reg;
57 	struct nvkm_fb *fb = nvxx_fb(&drm->client.device);
58 	struct nvkm_fb_tile *tile = &fb->tile.region[i];
59 
60 	nouveau_fence_unref(&reg->fence);
61 
62 	if (tile->pitch)
63 		nvkm_fb_tile_fini(fb, i, tile);
64 
65 	if (pitch)
66 		nvkm_fb_tile_init(fb, i, addr, size, pitch, flags, tile);
67 
68 	nvkm_fb_tile_prog(fb, i, tile);
69 }
70 
71 static struct nouveau_drm_tile *
72 nv10_bo_get_tile_region(struct drm_device *dev, int i)
73 {
74 	struct nouveau_drm *drm = nouveau_drm(dev);
75 	struct nouveau_drm_tile *tile = &drm->tile.reg[i];
76 
77 	spin_lock(&drm->tile.lock);
78 
79 	if (!tile->used &&
80 	    (!tile->fence || nouveau_fence_done(tile->fence)))
81 		tile->used = true;
82 	else
83 		tile = NULL;
84 
85 	spin_unlock(&drm->tile.lock);
86 	return tile;
87 }
88 
89 static void
90 nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
91 			struct dma_fence *fence)
92 {
93 	struct nouveau_drm *drm = nouveau_drm(dev);
94 
95 	if (tile) {
96 		spin_lock(&drm->tile.lock);
97 		tile->fence = (struct nouveau_fence *)dma_fence_get(fence);
98 		tile->used = false;
99 		spin_unlock(&drm->tile.lock);
100 	}
101 }
102 
103 static struct nouveau_drm_tile *
104 nv10_bo_set_tiling(struct drm_device *dev, u32 addr,
105 		   u32 size, u32 pitch, u32 zeta)
106 {
107 	struct nouveau_drm *drm = nouveau_drm(dev);
108 	struct nvkm_fb *fb = nvxx_fb(&drm->client.device);
109 	struct nouveau_drm_tile *tile, *found = NULL;
110 	int i;
111 
112 	for (i = 0; i < fb->tile.regions; i++) {
113 		tile = nv10_bo_get_tile_region(dev, i);
114 
115 		if (pitch && !found) {
116 			found = tile;
117 			continue;
118 
119 		} else if (tile && fb->tile.region[i].pitch) {
120 			/* Kill an unused tile region. */
121 			nv10_bo_update_tile_region(dev, tile, 0, 0, 0, 0);
122 		}
123 
124 		nv10_bo_put_tile_region(dev, tile, NULL);
125 	}
126 
127 	if (found)
128 		nv10_bo_update_tile_region(dev, found, addr, size, pitch, zeta);
129 	return found;
130 }
131 
132 static void
133 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
134 {
135 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
136 	struct drm_device *dev = drm->dev;
137 	struct nouveau_bo *nvbo = nouveau_bo(bo);
138 
139 	if (unlikely(nvbo->gem.filp))
140 		DRM_ERROR("bo %p still attached to GEM object\n", bo);
141 	WARN_ON(nvbo->pin_refcnt > 0);
142 	nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
143 	kfree(nvbo);
144 }
145 
146 static inline u64
147 roundup_64(u64 x, u32 y)
148 {
149 	x += y - 1;
150 	do_div(x, y);
151 	return x * y;
152 }
153 
154 static void
155 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
156 		       int *align, u64 *size)
157 {
158 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
159 	struct nvif_device *device = &drm->client.device;
160 
161 	if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
162 		if (nvbo->mode) {
163 			if (device->info.chipset >= 0x40) {
164 				*align = 65536;
165 				*size = roundup_64(*size, 64 * nvbo->mode);
166 
167 			} else if (device->info.chipset >= 0x30) {
168 				*align = 32768;
169 				*size = roundup_64(*size, 64 * nvbo->mode);
170 
171 			} else if (device->info.chipset >= 0x20) {
172 				*align = 16384;
173 				*size = roundup_64(*size, 64 * nvbo->mode);
174 
175 			} else if (device->info.chipset >= 0x10) {
176 				*align = 16384;
177 				*size = roundup_64(*size, 32 * nvbo->mode);
178 			}
179 		}
180 	} else {
181 		*size = roundup_64(*size, (1 << nvbo->page));
182 		*align = max((1 <<  nvbo->page), *align);
183 	}
184 
185 	*size = roundup_64(*size, PAGE_SIZE);
186 }
187 
188 int
189 nouveau_bo_new(struct nouveau_cli *cli, u64 size, int align,
190 	       uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
191 	       struct sg_table *sg, struct reservation_object *robj,
192 	       struct nouveau_bo **pnvbo)
193 {
194 	struct nouveau_drm *drm = cli->drm;
195 	struct nouveau_bo *nvbo;
196 	struct nvif_mmu *mmu = &cli->mmu;
197 	struct nvif_vmm *vmm = &cli->vmm.vmm;
198 	size_t acc_size;
199 	int type = ttm_bo_type_device;
200 	int ret, i, pi = -1;
201 
202 	if (!size) {
203 		NV_WARN(drm, "skipped size %016llx\n", size);
204 		return -EINVAL;
205 	}
206 
207 	if (sg)
208 		type = ttm_bo_type_sg;
209 
210 	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
211 	if (!nvbo)
212 		return -ENOMEM;
213 	INIT_LIST_HEAD(&nvbo->head);
214 	INIT_LIST_HEAD(&nvbo->entry);
215 	INIT_LIST_HEAD(&nvbo->vma_list);
216 	nvbo->bo.bdev = &drm->ttm.bdev;
217 
218 	/* This is confusing, and doesn't actually mean we want an uncached
219 	 * mapping, but is what NOUVEAU_GEM_DOMAIN_COHERENT gets translated
220 	 * into in nouveau_gem_new().
221 	 */
222 	if (flags & TTM_PL_FLAG_UNCACHED) {
223 		/* Determine if we can get a cache-coherent map, forcing
224 		 * uncached mapping if we can't.
225 		 */
226 		if (!nouveau_drm_use_coherent_gpu_mapping(drm))
227 			nvbo->force_coherent = true;
228 	}
229 
230 	if (cli->device.info.family >= NV_DEVICE_INFO_V0_FERMI) {
231 		nvbo->kind = (tile_flags & 0x0000ff00) >> 8;
232 		if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
233 			kfree(nvbo);
234 			return -EINVAL;
235 		}
236 
237 		nvbo->comp = mmu->kind[nvbo->kind] != nvbo->kind;
238 	} else
239 	if (cli->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
240 		nvbo->kind = (tile_flags & 0x00007f00) >> 8;
241 		nvbo->comp = (tile_flags & 0x00030000) >> 16;
242 		if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
243 			kfree(nvbo);
244 			return -EINVAL;
245 		}
246 	} else {
247 		nvbo->zeta = (tile_flags & 0x00000007);
248 	}
249 	nvbo->mode = tile_mode;
250 	nvbo->contig = !(tile_flags & NOUVEAU_GEM_TILE_NONCONTIG);
251 
252 	/* Determine the desirable target GPU page size for the buffer. */
253 	for (i = 0; i < vmm->page_nr; i++) {
254 		/* Because we cannot currently allow VMM maps to fail
255 		 * during buffer migration, we need to determine page
256 		 * size for the buffer up-front, and pre-allocate its
257 		 * page tables.
258 		 *
259 		 * Skip page sizes that can't support needed domains.
260 		 */
261 		if (cli->device.info.family > NV_DEVICE_INFO_V0_CURIE &&
262 		    (flags & TTM_PL_FLAG_VRAM) && !vmm->page[i].vram)
263 			continue;
264 		if ((flags & TTM_PL_FLAG_TT) &&
265 		    (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
266 			continue;
267 
268 		/* Select this page size if it's the first that supports
269 		 * the potential memory domains, or when it's compatible
270 		 * with the requested compression settings.
271 		 */
272 		if (pi < 0 || !nvbo->comp || vmm->page[i].comp)
273 			pi = i;
274 
275 		/* Stop once the buffer is larger than the current page size. */
276 		if (size >= 1ULL << vmm->page[i].shift)
277 			break;
278 	}
279 
280 	if (WARN_ON(pi < 0))
281 		return -EINVAL;
282 
283 	/* Disable compression if suitable settings couldn't be found. */
284 	if (nvbo->comp && !vmm->page[pi].comp) {
285 		if (mmu->object.oclass >= NVIF_CLASS_MMU_GF100)
286 			nvbo->kind = mmu->kind[nvbo->kind];
287 		nvbo->comp = 0;
288 	}
289 	nvbo->page = vmm->page[pi].shift;
290 
291 	nouveau_bo_fixup_align(nvbo, flags, &align, &size);
292 	nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
293 	nouveau_bo_placement_set(nvbo, flags, 0);
294 
295 	acc_size = ttm_bo_dma_acc_size(&drm->ttm.bdev, size,
296 				       sizeof(struct nouveau_bo));
297 
298 	ret = ttm_bo_init(&drm->ttm.bdev, &nvbo->bo, size,
299 			  type, &nvbo->placement,
300 			  align >> PAGE_SHIFT, false, acc_size, sg,
301 			  robj, nouveau_bo_del_ttm);
302 	if (ret) {
303 		/* ttm will call nouveau_bo_del_ttm if it fails.. */
304 		return ret;
305 	}
306 
307 	*pnvbo = nvbo;
308 	return 0;
309 }
310 
311 static void
312 set_placement_list(struct ttm_place *pl, unsigned *n, uint32_t type, uint32_t flags)
313 {
314 	*n = 0;
315 
316 	if (type & TTM_PL_FLAG_VRAM)
317 		pl[(*n)++].flags = TTM_PL_FLAG_VRAM | flags;
318 	if (type & TTM_PL_FLAG_TT)
319 		pl[(*n)++].flags = TTM_PL_FLAG_TT | flags;
320 	if (type & TTM_PL_FLAG_SYSTEM)
321 		pl[(*n)++].flags = TTM_PL_FLAG_SYSTEM | flags;
322 }
323 
324 static void
325 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
326 {
327 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
328 	u32 vram_pages = drm->client.device.info.ram_size >> PAGE_SHIFT;
329 	unsigned i, fpfn, lpfn;
330 
331 	if (drm->client.device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
332 	    nvbo->mode && (type & TTM_PL_FLAG_VRAM) &&
333 	    nvbo->bo.mem.num_pages < vram_pages / 4) {
334 		/*
335 		 * Make sure that the color and depth buffers are handled
336 		 * by independent memory controller units. Up to a 9x
337 		 * speed up when alpha-blending and depth-test are enabled
338 		 * at the same time.
339 		 */
340 		if (nvbo->zeta) {
341 			fpfn = vram_pages / 2;
342 			lpfn = ~0;
343 		} else {
344 			fpfn = 0;
345 			lpfn = vram_pages / 2;
346 		}
347 		for (i = 0; i < nvbo->placement.num_placement; ++i) {
348 			nvbo->placements[i].fpfn = fpfn;
349 			nvbo->placements[i].lpfn = lpfn;
350 		}
351 		for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
352 			nvbo->busy_placements[i].fpfn = fpfn;
353 			nvbo->busy_placements[i].lpfn = lpfn;
354 		}
355 	}
356 }
357 
358 void
359 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
360 {
361 	struct ttm_placement *pl = &nvbo->placement;
362 	uint32_t flags = (nvbo->force_coherent ? TTM_PL_FLAG_UNCACHED :
363 						 TTM_PL_MASK_CACHING) |
364 			 (nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
365 
366 	pl->placement = nvbo->placements;
367 	set_placement_list(nvbo->placements, &pl->num_placement,
368 			   type, flags);
369 
370 	pl->busy_placement = nvbo->busy_placements;
371 	set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
372 			   type | busy, flags);
373 
374 	set_placement_range(nvbo, type);
375 }
376 
377 int
378 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype, bool contig)
379 {
380 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
381 	struct ttm_buffer_object *bo = &nvbo->bo;
382 	bool force = false, evict = false;
383 	int ret;
384 
385 	ret = ttm_bo_reserve(bo, false, false, NULL);
386 	if (ret)
387 		return ret;
388 
389 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
390 	    memtype == TTM_PL_FLAG_VRAM && contig) {
391 		if (!nvbo->contig) {
392 			nvbo->contig = true;
393 			force = true;
394 			evict = true;
395 		}
396 	}
397 
398 	if (nvbo->pin_refcnt) {
399 		if (!(memtype & (1 << bo->mem.mem_type)) || evict) {
400 			NV_ERROR(drm, "bo %p pinned elsewhere: "
401 				      "0x%08x vs 0x%08x\n", bo,
402 				 1 << bo->mem.mem_type, memtype);
403 			ret = -EBUSY;
404 		}
405 		nvbo->pin_refcnt++;
406 		goto out;
407 	}
408 
409 	if (evict) {
410 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT, 0);
411 		ret = nouveau_bo_validate(nvbo, false, false);
412 		if (ret)
413 			goto out;
414 	}
415 
416 	nvbo->pin_refcnt++;
417 	nouveau_bo_placement_set(nvbo, memtype, 0);
418 
419 	/* drop pin_refcnt temporarily, so we don't trip the assertion
420 	 * in nouveau_bo_move() that makes sure we're not trying to
421 	 * move a pinned buffer
422 	 */
423 	nvbo->pin_refcnt--;
424 	ret = nouveau_bo_validate(nvbo, false, false);
425 	if (ret)
426 		goto out;
427 	nvbo->pin_refcnt++;
428 
429 	switch (bo->mem.mem_type) {
430 	case TTM_PL_VRAM:
431 		drm->gem.vram_available -= bo->mem.size;
432 		break;
433 	case TTM_PL_TT:
434 		drm->gem.gart_available -= bo->mem.size;
435 		break;
436 	default:
437 		break;
438 	}
439 
440 out:
441 	if (force && ret)
442 		nvbo->contig = false;
443 	ttm_bo_unreserve(bo);
444 	return ret;
445 }
446 
447 int
448 nouveau_bo_unpin(struct nouveau_bo *nvbo)
449 {
450 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
451 	struct ttm_buffer_object *bo = &nvbo->bo;
452 	int ret, ref;
453 
454 	ret = ttm_bo_reserve(bo, false, false, NULL);
455 	if (ret)
456 		return ret;
457 
458 	ref = --nvbo->pin_refcnt;
459 	WARN_ON_ONCE(ref < 0);
460 	if (ref)
461 		goto out;
462 
463 	nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
464 
465 	ret = nouveau_bo_validate(nvbo, false, false);
466 	if (ret == 0) {
467 		switch (bo->mem.mem_type) {
468 		case TTM_PL_VRAM:
469 			drm->gem.vram_available += bo->mem.size;
470 			break;
471 		case TTM_PL_TT:
472 			drm->gem.gart_available += bo->mem.size;
473 			break;
474 		default:
475 			break;
476 		}
477 	}
478 
479 out:
480 	ttm_bo_unreserve(bo);
481 	return ret;
482 }
483 
484 int
485 nouveau_bo_map(struct nouveau_bo *nvbo)
486 {
487 	int ret;
488 
489 	ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
490 	if (ret)
491 		return ret;
492 
493 	ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
494 
495 	ttm_bo_unreserve(&nvbo->bo);
496 	return ret;
497 }
498 
499 void
500 nouveau_bo_unmap(struct nouveau_bo *nvbo)
501 {
502 	if (!nvbo)
503 		return;
504 
505 	ttm_bo_kunmap(&nvbo->kmap);
506 }
507 
508 void
509 nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
510 {
511 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
512 	struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
513 	int i;
514 
515 	if (!ttm_dma)
516 		return;
517 
518 	/* Don't waste time looping if the object is coherent */
519 	if (nvbo->force_coherent)
520 		return;
521 
522 	for (i = 0; i < ttm_dma->ttm.num_pages; i++)
523 		dma_sync_single_for_device(drm->dev->dev,
524 					   ttm_dma->dma_address[i],
525 					   PAGE_SIZE, DMA_TO_DEVICE);
526 }
527 
528 void
529 nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
530 {
531 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
532 	struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
533 	int i;
534 
535 	if (!ttm_dma)
536 		return;
537 
538 	/* Don't waste time looping if the object is coherent */
539 	if (nvbo->force_coherent)
540 		return;
541 
542 	for (i = 0; i < ttm_dma->ttm.num_pages; i++)
543 		dma_sync_single_for_cpu(drm->dev->dev, ttm_dma->dma_address[i],
544 					PAGE_SIZE, DMA_FROM_DEVICE);
545 }
546 
547 int
548 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
549 		    bool no_wait_gpu)
550 {
551 	struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
552 	int ret;
553 
554 	ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, &ctx);
555 	if (ret)
556 		return ret;
557 
558 	nouveau_bo_sync_for_device(nvbo);
559 
560 	return 0;
561 }
562 
563 void
564 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
565 {
566 	bool is_iomem;
567 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
568 
569 	mem += index;
570 
571 	if (is_iomem)
572 		iowrite16_native(val, (void __force __iomem *)mem);
573 	else
574 		*mem = val;
575 }
576 
577 u32
578 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
579 {
580 	bool is_iomem;
581 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
582 
583 	mem += index;
584 
585 	if (is_iomem)
586 		return ioread32_native((void __force __iomem *)mem);
587 	else
588 		return *mem;
589 }
590 
591 void
592 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
593 {
594 	bool is_iomem;
595 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
596 
597 	mem += index;
598 
599 	if (is_iomem)
600 		iowrite32_native(val, (void __force __iomem *)mem);
601 	else
602 		*mem = val;
603 }
604 
605 static struct ttm_tt *
606 nouveau_ttm_tt_create(struct ttm_buffer_object *bo, uint32_t page_flags)
607 {
608 #if IS_ENABLED(CONFIG_AGP)
609 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
610 
611 	if (drm->agp.bridge) {
612 		return ttm_agp_tt_create(bo, drm->agp.bridge, page_flags);
613 	}
614 #endif
615 
616 	return nouveau_sgdma_create_ttm(bo, page_flags);
617 }
618 
619 static int
620 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
621 {
622 	/* We'll do this from user space. */
623 	return 0;
624 }
625 
626 static int
627 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
628 			 struct ttm_mem_type_manager *man)
629 {
630 	struct nouveau_drm *drm = nouveau_bdev(bdev);
631 	struct nvif_mmu *mmu = &drm->client.mmu;
632 
633 	switch (type) {
634 	case TTM_PL_SYSTEM:
635 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
636 		man->available_caching = TTM_PL_MASK_CACHING;
637 		man->default_caching = TTM_PL_FLAG_CACHED;
638 		break;
639 	case TTM_PL_VRAM:
640 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
641 			     TTM_MEMTYPE_FLAG_MAPPABLE;
642 		man->available_caching = TTM_PL_FLAG_UNCACHED |
643 					 TTM_PL_FLAG_WC;
644 		man->default_caching = TTM_PL_FLAG_WC;
645 
646 		if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
647 			/* Some BARs do not support being ioremapped WC */
648 			const u8 type = mmu->type[drm->ttm.type_vram].type;
649 			if (type & NVIF_MEM_UNCACHED) {
650 				man->available_caching = TTM_PL_FLAG_UNCACHED;
651 				man->default_caching = TTM_PL_FLAG_UNCACHED;
652 			}
653 
654 			man->func = &nouveau_vram_manager;
655 			man->io_reserve_fastpath = false;
656 			man->use_io_reserve_lru = true;
657 		} else {
658 			man->func = &ttm_bo_manager_func;
659 		}
660 		break;
661 	case TTM_PL_TT:
662 		if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA)
663 			man->func = &nouveau_gart_manager;
664 		else
665 		if (!drm->agp.bridge)
666 			man->func = &nv04_gart_manager;
667 		else
668 			man->func = &ttm_bo_manager_func;
669 
670 		if (drm->agp.bridge) {
671 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
672 			man->available_caching = TTM_PL_FLAG_UNCACHED |
673 				TTM_PL_FLAG_WC;
674 			man->default_caching = TTM_PL_FLAG_WC;
675 		} else {
676 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
677 				     TTM_MEMTYPE_FLAG_CMA;
678 			man->available_caching = TTM_PL_MASK_CACHING;
679 			man->default_caching = TTM_PL_FLAG_CACHED;
680 		}
681 
682 		break;
683 	default:
684 		return -EINVAL;
685 	}
686 	return 0;
687 }
688 
689 static void
690 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
691 {
692 	struct nouveau_bo *nvbo = nouveau_bo(bo);
693 
694 	switch (bo->mem.mem_type) {
695 	case TTM_PL_VRAM:
696 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
697 					 TTM_PL_FLAG_SYSTEM);
698 		break;
699 	default:
700 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
701 		break;
702 	}
703 
704 	*pl = nvbo->placement;
705 }
706 
707 
708 static int
709 nve0_bo_move_init(struct nouveau_channel *chan, u32 handle)
710 {
711 	int ret = RING_SPACE(chan, 2);
712 	if (ret == 0) {
713 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
714 		OUT_RING  (chan, handle & 0x0000ffff);
715 		FIRE_RING (chan);
716 	}
717 	return ret;
718 }
719 
720 static int
721 nve0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
722 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
723 {
724 	struct nouveau_mem *mem = nouveau_mem(old_reg);
725 	int ret = RING_SPACE(chan, 10);
726 	if (ret == 0) {
727 		BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
728 		OUT_RING  (chan, upper_32_bits(mem->vma[0].addr));
729 		OUT_RING  (chan, lower_32_bits(mem->vma[0].addr));
730 		OUT_RING  (chan, upper_32_bits(mem->vma[1].addr));
731 		OUT_RING  (chan, lower_32_bits(mem->vma[1].addr));
732 		OUT_RING  (chan, PAGE_SIZE);
733 		OUT_RING  (chan, PAGE_SIZE);
734 		OUT_RING  (chan, PAGE_SIZE);
735 		OUT_RING  (chan, new_reg->num_pages);
736 		BEGIN_IMC0(chan, NvSubCopy, 0x0300, 0x0386);
737 	}
738 	return ret;
739 }
740 
741 static int
742 nvc0_bo_move_init(struct nouveau_channel *chan, u32 handle)
743 {
744 	int ret = RING_SPACE(chan, 2);
745 	if (ret == 0) {
746 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
747 		OUT_RING  (chan, handle);
748 	}
749 	return ret;
750 }
751 
752 static int
753 nvc0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
754 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
755 {
756 	struct nouveau_mem *mem = nouveau_mem(old_reg);
757 	u64 src_offset = mem->vma[0].addr;
758 	u64 dst_offset = mem->vma[1].addr;
759 	u32 page_count = new_reg->num_pages;
760 	int ret;
761 
762 	page_count = new_reg->num_pages;
763 	while (page_count) {
764 		int line_count = (page_count > 8191) ? 8191 : page_count;
765 
766 		ret = RING_SPACE(chan, 11);
767 		if (ret)
768 			return ret;
769 
770 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 8);
771 		OUT_RING  (chan, upper_32_bits(src_offset));
772 		OUT_RING  (chan, lower_32_bits(src_offset));
773 		OUT_RING  (chan, upper_32_bits(dst_offset));
774 		OUT_RING  (chan, lower_32_bits(dst_offset));
775 		OUT_RING  (chan, PAGE_SIZE);
776 		OUT_RING  (chan, PAGE_SIZE);
777 		OUT_RING  (chan, PAGE_SIZE);
778 		OUT_RING  (chan, line_count);
779 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
780 		OUT_RING  (chan, 0x00000110);
781 
782 		page_count -= line_count;
783 		src_offset += (PAGE_SIZE * line_count);
784 		dst_offset += (PAGE_SIZE * line_count);
785 	}
786 
787 	return 0;
788 }
789 
790 static int
791 nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
792 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
793 {
794 	struct nouveau_mem *mem = nouveau_mem(old_reg);
795 	u64 src_offset = mem->vma[0].addr;
796 	u64 dst_offset = mem->vma[1].addr;
797 	u32 page_count = new_reg->num_pages;
798 	int ret;
799 
800 	page_count = new_reg->num_pages;
801 	while (page_count) {
802 		int line_count = (page_count > 2047) ? 2047 : page_count;
803 
804 		ret = RING_SPACE(chan, 12);
805 		if (ret)
806 			return ret;
807 
808 		BEGIN_NVC0(chan, NvSubCopy, 0x0238, 2);
809 		OUT_RING  (chan, upper_32_bits(dst_offset));
810 		OUT_RING  (chan, lower_32_bits(dst_offset));
811 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 6);
812 		OUT_RING  (chan, upper_32_bits(src_offset));
813 		OUT_RING  (chan, lower_32_bits(src_offset));
814 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
815 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
816 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
817 		OUT_RING  (chan, line_count);
818 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
819 		OUT_RING  (chan, 0x00100110);
820 
821 		page_count -= line_count;
822 		src_offset += (PAGE_SIZE * line_count);
823 		dst_offset += (PAGE_SIZE * line_count);
824 	}
825 
826 	return 0;
827 }
828 
829 static int
830 nva3_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
831 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
832 {
833 	struct nouveau_mem *mem = nouveau_mem(old_reg);
834 	u64 src_offset = mem->vma[0].addr;
835 	u64 dst_offset = mem->vma[1].addr;
836 	u32 page_count = new_reg->num_pages;
837 	int ret;
838 
839 	page_count = new_reg->num_pages;
840 	while (page_count) {
841 		int line_count = (page_count > 8191) ? 8191 : page_count;
842 
843 		ret = RING_SPACE(chan, 11);
844 		if (ret)
845 			return ret;
846 
847 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
848 		OUT_RING  (chan, upper_32_bits(src_offset));
849 		OUT_RING  (chan, lower_32_bits(src_offset));
850 		OUT_RING  (chan, upper_32_bits(dst_offset));
851 		OUT_RING  (chan, lower_32_bits(dst_offset));
852 		OUT_RING  (chan, PAGE_SIZE);
853 		OUT_RING  (chan, PAGE_SIZE);
854 		OUT_RING  (chan, PAGE_SIZE);
855 		OUT_RING  (chan, line_count);
856 		BEGIN_NV04(chan, NvSubCopy, 0x0300, 1);
857 		OUT_RING  (chan, 0x00000110);
858 
859 		page_count -= line_count;
860 		src_offset += (PAGE_SIZE * line_count);
861 		dst_offset += (PAGE_SIZE * line_count);
862 	}
863 
864 	return 0;
865 }
866 
867 static int
868 nv98_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
869 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
870 {
871 	struct nouveau_mem *mem = nouveau_mem(old_reg);
872 	int ret = RING_SPACE(chan, 7);
873 	if (ret == 0) {
874 		BEGIN_NV04(chan, NvSubCopy, 0x0320, 6);
875 		OUT_RING  (chan, upper_32_bits(mem->vma[0].addr));
876 		OUT_RING  (chan, lower_32_bits(mem->vma[0].addr));
877 		OUT_RING  (chan, upper_32_bits(mem->vma[1].addr));
878 		OUT_RING  (chan, lower_32_bits(mem->vma[1].addr));
879 		OUT_RING  (chan, 0x00000000 /* COPY */);
880 		OUT_RING  (chan, new_reg->num_pages << PAGE_SHIFT);
881 	}
882 	return ret;
883 }
884 
885 static int
886 nv84_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
887 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
888 {
889 	struct nouveau_mem *mem = nouveau_mem(old_reg);
890 	int ret = RING_SPACE(chan, 7);
891 	if (ret == 0) {
892 		BEGIN_NV04(chan, NvSubCopy, 0x0304, 6);
893 		OUT_RING  (chan, new_reg->num_pages << PAGE_SHIFT);
894 		OUT_RING  (chan, upper_32_bits(mem->vma[0].addr));
895 		OUT_RING  (chan, lower_32_bits(mem->vma[0].addr));
896 		OUT_RING  (chan, upper_32_bits(mem->vma[1].addr));
897 		OUT_RING  (chan, lower_32_bits(mem->vma[1].addr));
898 		OUT_RING  (chan, 0x00000000 /* MODE_COPY, QUERY_NONE */);
899 	}
900 	return ret;
901 }
902 
903 static int
904 nv50_bo_move_init(struct nouveau_channel *chan, u32 handle)
905 {
906 	int ret = RING_SPACE(chan, 6);
907 	if (ret == 0) {
908 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
909 		OUT_RING  (chan, handle);
910 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 3);
911 		OUT_RING  (chan, chan->drm->ntfy.handle);
912 		OUT_RING  (chan, chan->vram.handle);
913 		OUT_RING  (chan, chan->vram.handle);
914 	}
915 
916 	return ret;
917 }
918 
919 static int
920 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
921 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
922 {
923 	struct nouveau_mem *mem = nouveau_mem(old_reg);
924 	u64 length = (new_reg->num_pages << PAGE_SHIFT);
925 	u64 src_offset = mem->vma[0].addr;
926 	u64 dst_offset = mem->vma[1].addr;
927 	int src_tiled = !!mem->kind;
928 	int dst_tiled = !!nouveau_mem(new_reg)->kind;
929 	int ret;
930 
931 	while (length) {
932 		u32 amount, stride, height;
933 
934 		ret = RING_SPACE(chan, 18 + 6 * (src_tiled + dst_tiled));
935 		if (ret)
936 			return ret;
937 
938 		amount  = min(length, (u64)(4 * 1024 * 1024));
939 		stride  = 16 * 4;
940 		height  = amount / stride;
941 
942 		if (src_tiled) {
943 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 7);
944 			OUT_RING  (chan, 0);
945 			OUT_RING  (chan, 0);
946 			OUT_RING  (chan, stride);
947 			OUT_RING  (chan, height);
948 			OUT_RING  (chan, 1);
949 			OUT_RING  (chan, 0);
950 			OUT_RING  (chan, 0);
951 		} else {
952 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 1);
953 			OUT_RING  (chan, 1);
954 		}
955 		if (dst_tiled) {
956 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 7);
957 			OUT_RING  (chan, 0);
958 			OUT_RING  (chan, 0);
959 			OUT_RING  (chan, stride);
960 			OUT_RING  (chan, height);
961 			OUT_RING  (chan, 1);
962 			OUT_RING  (chan, 0);
963 			OUT_RING  (chan, 0);
964 		} else {
965 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 1);
966 			OUT_RING  (chan, 1);
967 		}
968 
969 		BEGIN_NV04(chan, NvSubCopy, 0x0238, 2);
970 		OUT_RING  (chan, upper_32_bits(src_offset));
971 		OUT_RING  (chan, upper_32_bits(dst_offset));
972 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
973 		OUT_RING  (chan, lower_32_bits(src_offset));
974 		OUT_RING  (chan, lower_32_bits(dst_offset));
975 		OUT_RING  (chan, stride);
976 		OUT_RING  (chan, stride);
977 		OUT_RING  (chan, stride);
978 		OUT_RING  (chan, height);
979 		OUT_RING  (chan, 0x00000101);
980 		OUT_RING  (chan, 0x00000000);
981 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
982 		OUT_RING  (chan, 0);
983 
984 		length -= amount;
985 		src_offset += amount;
986 		dst_offset += amount;
987 	}
988 
989 	return 0;
990 }
991 
992 static int
993 nv04_bo_move_init(struct nouveau_channel *chan, u32 handle)
994 {
995 	int ret = RING_SPACE(chan, 4);
996 	if (ret == 0) {
997 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
998 		OUT_RING  (chan, handle);
999 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 1);
1000 		OUT_RING  (chan, chan->drm->ntfy.handle);
1001 	}
1002 
1003 	return ret;
1004 }
1005 
1006 static inline uint32_t
1007 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
1008 		      struct nouveau_channel *chan, struct ttm_mem_reg *reg)
1009 {
1010 	if (reg->mem_type == TTM_PL_TT)
1011 		return NvDmaTT;
1012 	return chan->vram.handle;
1013 }
1014 
1015 static int
1016 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
1017 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
1018 {
1019 	u32 src_offset = old_reg->start << PAGE_SHIFT;
1020 	u32 dst_offset = new_reg->start << PAGE_SHIFT;
1021 	u32 page_count = new_reg->num_pages;
1022 	int ret;
1023 
1024 	ret = RING_SPACE(chan, 3);
1025 	if (ret)
1026 		return ret;
1027 
1028 	BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
1029 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_reg));
1030 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_reg));
1031 
1032 	page_count = new_reg->num_pages;
1033 	while (page_count) {
1034 		int line_count = (page_count > 2047) ? 2047 : page_count;
1035 
1036 		ret = RING_SPACE(chan, 11);
1037 		if (ret)
1038 			return ret;
1039 
1040 		BEGIN_NV04(chan, NvSubCopy,
1041 				 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
1042 		OUT_RING  (chan, src_offset);
1043 		OUT_RING  (chan, dst_offset);
1044 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
1045 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
1046 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
1047 		OUT_RING  (chan, line_count);
1048 		OUT_RING  (chan, 0x00000101);
1049 		OUT_RING  (chan, 0x00000000);
1050 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
1051 		OUT_RING  (chan, 0);
1052 
1053 		page_count -= line_count;
1054 		src_offset += (PAGE_SIZE * line_count);
1055 		dst_offset += (PAGE_SIZE * line_count);
1056 	}
1057 
1058 	return 0;
1059 }
1060 
1061 static int
1062 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
1063 		     struct ttm_mem_reg *reg)
1064 {
1065 	struct nouveau_mem *old_mem = nouveau_mem(&bo->mem);
1066 	struct nouveau_mem *new_mem = nouveau_mem(reg);
1067 	struct nvif_vmm *vmm = &drm->client.vmm.vmm;
1068 	int ret;
1069 
1070 	ret = nvif_vmm_get(vmm, LAZY, false, old_mem->mem.page, 0,
1071 			   old_mem->mem.size, &old_mem->vma[0]);
1072 	if (ret)
1073 		return ret;
1074 
1075 	ret = nvif_vmm_get(vmm, LAZY, false, new_mem->mem.page, 0,
1076 			   new_mem->mem.size, &old_mem->vma[1]);
1077 	if (ret)
1078 		goto done;
1079 
1080 	ret = nouveau_mem_map(old_mem, vmm, &old_mem->vma[0]);
1081 	if (ret)
1082 		goto done;
1083 
1084 	ret = nouveau_mem_map(new_mem, vmm, &old_mem->vma[1]);
1085 done:
1086 	if (ret) {
1087 		nvif_vmm_put(vmm, &old_mem->vma[1]);
1088 		nvif_vmm_put(vmm, &old_mem->vma[0]);
1089 	}
1090 	return 0;
1091 }
1092 
1093 static int
1094 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
1095 		     bool no_wait_gpu, struct ttm_mem_reg *new_reg)
1096 {
1097 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1098 	struct nouveau_channel *chan = drm->ttm.chan;
1099 	struct nouveau_cli *cli = (void *)chan->user.client;
1100 	struct nouveau_fence *fence;
1101 	int ret;
1102 
1103 	/* create temporary vmas for the transfer and attach them to the
1104 	 * old nvkm_mem node, these will get cleaned up after ttm has
1105 	 * destroyed the ttm_mem_reg
1106 	 */
1107 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
1108 		ret = nouveau_bo_move_prep(drm, bo, new_reg);
1109 		if (ret)
1110 			return ret;
1111 	}
1112 
1113 	mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
1114 	ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, intr);
1115 	if (ret == 0) {
1116 		ret = drm->ttm.move(chan, bo, &bo->mem, new_reg);
1117 		if (ret == 0) {
1118 			ret = nouveau_fence_new(chan, false, &fence);
1119 			if (ret == 0) {
1120 				ret = ttm_bo_move_accel_cleanup(bo,
1121 								&fence->base,
1122 								evict,
1123 								new_reg);
1124 				nouveau_fence_unref(&fence);
1125 			}
1126 		}
1127 	}
1128 	mutex_unlock(&cli->mutex);
1129 	return ret;
1130 }
1131 
1132 void
1133 nouveau_bo_move_init(struct nouveau_drm *drm)
1134 {
1135 	static const struct {
1136 		const char *name;
1137 		int engine;
1138 		s32 oclass;
1139 		int (*exec)(struct nouveau_channel *,
1140 			    struct ttm_buffer_object *,
1141 			    struct ttm_mem_reg *, struct ttm_mem_reg *);
1142 		int (*init)(struct nouveau_channel *, u32 handle);
1143 	} _methods[] = {
1144 		{  "COPY", 4, 0xc5b5, nve0_bo_move_copy, nve0_bo_move_init },
1145 		{  "GRCE", 0, 0xc5b5, nve0_bo_move_copy, nvc0_bo_move_init },
1146 		{  "COPY", 4, 0xc3b5, nve0_bo_move_copy, nve0_bo_move_init },
1147 		{  "GRCE", 0, 0xc3b5, nve0_bo_move_copy, nvc0_bo_move_init },
1148 		{  "COPY", 4, 0xc1b5, nve0_bo_move_copy, nve0_bo_move_init },
1149 		{  "GRCE", 0, 0xc1b5, nve0_bo_move_copy, nvc0_bo_move_init },
1150 		{  "COPY", 4, 0xc0b5, nve0_bo_move_copy, nve0_bo_move_init },
1151 		{  "GRCE", 0, 0xc0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1152 		{  "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
1153 		{  "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1154 		{  "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
1155 		{  "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1156 		{ "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
1157 		{ "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
1158 		{  "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
1159 		{ "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
1160 		{  "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
1161 		{  "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
1162 		{  "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
1163 		{},
1164 		{ "CRYPT", 0, 0x88b4, nv98_bo_move_exec, nv50_bo_move_init },
1165 	}, *mthd = _methods;
1166 	const char *name = "CPU";
1167 	int ret;
1168 
1169 	do {
1170 		struct nouveau_channel *chan;
1171 
1172 		if (mthd->engine)
1173 			chan = drm->cechan;
1174 		else
1175 			chan = drm->channel;
1176 		if (chan == NULL)
1177 			continue;
1178 
1179 		ret = nvif_object_init(&chan->user,
1180 				       mthd->oclass | (mthd->engine << 16),
1181 				       mthd->oclass, NULL, 0,
1182 				       &drm->ttm.copy);
1183 		if (ret == 0) {
1184 			ret = mthd->init(chan, drm->ttm.copy.handle);
1185 			if (ret) {
1186 				nvif_object_fini(&drm->ttm.copy);
1187 				continue;
1188 			}
1189 
1190 			drm->ttm.move = mthd->exec;
1191 			drm->ttm.chan = chan;
1192 			name = mthd->name;
1193 			break;
1194 		}
1195 	} while ((++mthd)->exec);
1196 
1197 	NV_INFO(drm, "MM: using %s for buffer copies\n", name);
1198 }
1199 
1200 static int
1201 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
1202 		      bool no_wait_gpu, struct ttm_mem_reg *new_reg)
1203 {
1204 	struct ttm_operation_ctx ctx = { intr, no_wait_gpu };
1205 	struct ttm_place placement_memtype = {
1206 		.fpfn = 0,
1207 		.lpfn = 0,
1208 		.flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
1209 	};
1210 	struct ttm_placement placement;
1211 	struct ttm_mem_reg tmp_reg;
1212 	int ret;
1213 
1214 	placement.num_placement = placement.num_busy_placement = 1;
1215 	placement.placement = placement.busy_placement = &placement_memtype;
1216 
1217 	tmp_reg = *new_reg;
1218 	tmp_reg.mm_node = NULL;
1219 	ret = ttm_bo_mem_space(bo, &placement, &tmp_reg, &ctx);
1220 	if (ret)
1221 		return ret;
1222 
1223 	ret = ttm_tt_bind(bo->ttm, &tmp_reg, &ctx);
1224 	if (ret)
1225 		goto out;
1226 
1227 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, &tmp_reg);
1228 	if (ret)
1229 		goto out;
1230 
1231 	ret = ttm_bo_move_ttm(bo, &ctx, new_reg);
1232 out:
1233 	ttm_bo_mem_put(bo, &tmp_reg);
1234 	return ret;
1235 }
1236 
1237 static int
1238 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
1239 		      bool no_wait_gpu, struct ttm_mem_reg *new_reg)
1240 {
1241 	struct ttm_operation_ctx ctx = { intr, no_wait_gpu };
1242 	struct ttm_place placement_memtype = {
1243 		.fpfn = 0,
1244 		.lpfn = 0,
1245 		.flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
1246 	};
1247 	struct ttm_placement placement;
1248 	struct ttm_mem_reg tmp_reg;
1249 	int ret;
1250 
1251 	placement.num_placement = placement.num_busy_placement = 1;
1252 	placement.placement = placement.busy_placement = &placement_memtype;
1253 
1254 	tmp_reg = *new_reg;
1255 	tmp_reg.mm_node = NULL;
1256 	ret = ttm_bo_mem_space(bo, &placement, &tmp_reg, &ctx);
1257 	if (ret)
1258 		return ret;
1259 
1260 	ret = ttm_bo_move_ttm(bo, &ctx, &tmp_reg);
1261 	if (ret)
1262 		goto out;
1263 
1264 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_reg);
1265 	if (ret)
1266 		goto out;
1267 
1268 out:
1269 	ttm_bo_mem_put(bo, &tmp_reg);
1270 	return ret;
1271 }
1272 
1273 static void
1274 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, bool evict,
1275 		     struct ttm_mem_reg *new_reg)
1276 {
1277 	struct nouveau_mem *mem = new_reg ? nouveau_mem(new_reg) : NULL;
1278 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1279 	struct nouveau_vma *vma;
1280 
1281 	/* ttm can now (stupidly) pass the driver bos it didn't create... */
1282 	if (bo->destroy != nouveau_bo_del_ttm)
1283 		return;
1284 
1285 	if (mem && new_reg->mem_type != TTM_PL_SYSTEM &&
1286 	    mem->mem.page == nvbo->page) {
1287 		list_for_each_entry(vma, &nvbo->vma_list, head) {
1288 			nouveau_vma_map(vma, mem);
1289 		}
1290 	} else {
1291 		list_for_each_entry(vma, &nvbo->vma_list, head) {
1292 			WARN_ON(ttm_bo_wait(bo, false, false));
1293 			nouveau_vma_unmap(vma);
1294 		}
1295 	}
1296 }
1297 
1298 static int
1299 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_reg,
1300 		   struct nouveau_drm_tile **new_tile)
1301 {
1302 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1303 	struct drm_device *dev = drm->dev;
1304 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1305 	u64 offset = new_reg->start << PAGE_SHIFT;
1306 
1307 	*new_tile = NULL;
1308 	if (new_reg->mem_type != TTM_PL_VRAM)
1309 		return 0;
1310 
1311 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
1312 		*new_tile = nv10_bo_set_tiling(dev, offset, new_reg->size,
1313 					       nvbo->mode, nvbo->zeta);
1314 	}
1315 
1316 	return 0;
1317 }
1318 
1319 static void
1320 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1321 		      struct nouveau_drm_tile *new_tile,
1322 		      struct nouveau_drm_tile **old_tile)
1323 {
1324 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1325 	struct drm_device *dev = drm->dev;
1326 	struct dma_fence *fence = reservation_object_get_excl(bo->resv);
1327 
1328 	nv10_bo_put_tile_region(dev, *old_tile, fence);
1329 	*old_tile = new_tile;
1330 }
1331 
1332 static int
1333 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict,
1334 		struct ttm_operation_ctx *ctx,
1335 		struct ttm_mem_reg *new_reg)
1336 {
1337 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1338 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1339 	struct ttm_mem_reg *old_reg = &bo->mem;
1340 	struct nouveau_drm_tile *new_tile = NULL;
1341 	int ret = 0;
1342 
1343 	ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
1344 	if (ret)
1345 		return ret;
1346 
1347 	if (nvbo->pin_refcnt)
1348 		NV_WARN(drm, "Moving pinned object %p!\n", nvbo);
1349 
1350 	if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1351 		ret = nouveau_bo_vm_bind(bo, new_reg, &new_tile);
1352 		if (ret)
1353 			return ret;
1354 	}
1355 
1356 	/* Fake bo copy. */
1357 	if (old_reg->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
1358 		BUG_ON(bo->mem.mm_node != NULL);
1359 		bo->mem = *new_reg;
1360 		new_reg->mm_node = NULL;
1361 		goto out;
1362 	}
1363 
1364 	/* Hardware assisted copy. */
1365 	if (drm->ttm.move) {
1366 		if (new_reg->mem_type == TTM_PL_SYSTEM)
1367 			ret = nouveau_bo_move_flipd(bo, evict,
1368 						    ctx->interruptible,
1369 						    ctx->no_wait_gpu, new_reg);
1370 		else if (old_reg->mem_type == TTM_PL_SYSTEM)
1371 			ret = nouveau_bo_move_flips(bo, evict,
1372 						    ctx->interruptible,
1373 						    ctx->no_wait_gpu, new_reg);
1374 		else
1375 			ret = nouveau_bo_move_m2mf(bo, evict,
1376 						   ctx->interruptible,
1377 						   ctx->no_wait_gpu, new_reg);
1378 		if (!ret)
1379 			goto out;
1380 	}
1381 
1382 	/* Fallback to software copy. */
1383 	ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
1384 	if (ret == 0)
1385 		ret = ttm_bo_move_memcpy(bo, ctx, new_reg);
1386 
1387 out:
1388 	if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1389 		if (ret)
1390 			nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1391 		else
1392 			nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1393 	}
1394 
1395 	return ret;
1396 }
1397 
1398 static int
1399 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1400 {
1401 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1402 
1403 	return drm_vma_node_verify_access(&nvbo->gem.vma_node,
1404 					  filp->private_data);
1405 }
1406 
1407 static int
1408 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *reg)
1409 {
1410 	struct ttm_mem_type_manager *man = &bdev->man[reg->mem_type];
1411 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1412 	struct nvkm_device *device = nvxx_device(&drm->client.device);
1413 	struct nouveau_mem *mem = nouveau_mem(reg);
1414 
1415 	reg->bus.addr = NULL;
1416 	reg->bus.offset = 0;
1417 	reg->bus.size = reg->num_pages << PAGE_SHIFT;
1418 	reg->bus.base = 0;
1419 	reg->bus.is_iomem = false;
1420 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
1421 		return -EINVAL;
1422 	switch (reg->mem_type) {
1423 	case TTM_PL_SYSTEM:
1424 		/* System memory */
1425 		return 0;
1426 	case TTM_PL_TT:
1427 #if IS_ENABLED(CONFIG_AGP)
1428 		if (drm->agp.bridge) {
1429 			reg->bus.offset = reg->start << PAGE_SHIFT;
1430 			reg->bus.base = drm->agp.base;
1431 			reg->bus.is_iomem = !drm->agp.cma;
1432 		}
1433 #endif
1434 		if (drm->client.mem->oclass < NVIF_CLASS_MEM_NV50 || !mem->kind)
1435 			/* untiled */
1436 			break;
1437 		/* fallthrough, tiled memory */
1438 	case TTM_PL_VRAM:
1439 		reg->bus.offset = reg->start << PAGE_SHIFT;
1440 		reg->bus.base = device->func->resource_addr(device, 1);
1441 		reg->bus.is_iomem = true;
1442 		if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1443 			union {
1444 				struct nv50_mem_map_v0 nv50;
1445 				struct gf100_mem_map_v0 gf100;
1446 			} args;
1447 			u64 handle, length;
1448 			u32 argc = 0;
1449 			int ret;
1450 
1451 			switch (mem->mem.object.oclass) {
1452 			case NVIF_CLASS_MEM_NV50:
1453 				args.nv50.version = 0;
1454 				args.nv50.ro = 0;
1455 				args.nv50.kind = mem->kind;
1456 				args.nv50.comp = mem->comp;
1457 				argc = sizeof(args.nv50);
1458 				break;
1459 			case NVIF_CLASS_MEM_GF100:
1460 				args.gf100.version = 0;
1461 				args.gf100.ro = 0;
1462 				args.gf100.kind = mem->kind;
1463 				argc = sizeof(args.gf100);
1464 				break;
1465 			default:
1466 				WARN_ON(1);
1467 				break;
1468 			}
1469 
1470 			ret = nvif_object_map_handle(&mem->mem.object,
1471 						     &args, argc,
1472 						     &handle, &length);
1473 			if (ret != 1)
1474 				return ret ? ret : -EINVAL;
1475 
1476 			reg->bus.base = 0;
1477 			reg->bus.offset = handle;
1478 		}
1479 		break;
1480 	default:
1481 		return -EINVAL;
1482 	}
1483 	return 0;
1484 }
1485 
1486 static void
1487 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *reg)
1488 {
1489 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1490 	struct nouveau_mem *mem = nouveau_mem(reg);
1491 
1492 	if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1493 		switch (reg->mem_type) {
1494 		case TTM_PL_TT:
1495 			if (mem->kind)
1496 				nvif_object_unmap_handle(&mem->mem.object);
1497 			break;
1498 		case TTM_PL_VRAM:
1499 			nvif_object_unmap_handle(&mem->mem.object);
1500 			break;
1501 		default:
1502 			break;
1503 		}
1504 	}
1505 }
1506 
1507 static int
1508 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1509 {
1510 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1511 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1512 	struct nvkm_device *device = nvxx_device(&drm->client.device);
1513 	u32 mappable = device->func->resource_size(device, 1) >> PAGE_SHIFT;
1514 	int i, ret;
1515 
1516 	/* as long as the bo isn't in vram, and isn't tiled, we've got
1517 	 * nothing to do here.
1518 	 */
1519 	if (bo->mem.mem_type != TTM_PL_VRAM) {
1520 		if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1521 		    !nvbo->kind)
1522 			return 0;
1523 
1524 		if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1525 			nouveau_bo_placement_set(nvbo, TTM_PL_TT, 0);
1526 
1527 			ret = nouveau_bo_validate(nvbo, false, false);
1528 			if (ret)
1529 				return ret;
1530 		}
1531 		return 0;
1532 	}
1533 
1534 	/* make sure bo is in mappable vram */
1535 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1536 	    bo->mem.start + bo->mem.num_pages < mappable)
1537 		return 0;
1538 
1539 	for (i = 0; i < nvbo->placement.num_placement; ++i) {
1540 		nvbo->placements[i].fpfn = 0;
1541 		nvbo->placements[i].lpfn = mappable;
1542 	}
1543 
1544 	for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
1545 		nvbo->busy_placements[i].fpfn = 0;
1546 		nvbo->busy_placements[i].lpfn = mappable;
1547 	}
1548 
1549 	nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_VRAM, 0);
1550 	return nouveau_bo_validate(nvbo, false, false);
1551 }
1552 
1553 static int
1554 nouveau_ttm_tt_populate(struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1555 {
1556 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1557 	struct nouveau_drm *drm;
1558 	struct device *dev;
1559 	unsigned i;
1560 	int r;
1561 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1562 
1563 	if (ttm->state != tt_unpopulated)
1564 		return 0;
1565 
1566 	if (slave && ttm->sg) {
1567 		/* make userspace faulting work */
1568 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1569 						 ttm_dma->dma_address, ttm->num_pages);
1570 		ttm->state = tt_unbound;
1571 		return 0;
1572 	}
1573 
1574 	drm = nouveau_bdev(ttm->bdev);
1575 	dev = drm->dev->dev;
1576 
1577 #if IS_ENABLED(CONFIG_AGP)
1578 	if (drm->agp.bridge) {
1579 		return ttm_agp_tt_populate(ttm, ctx);
1580 	}
1581 #endif
1582 
1583 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1584 	if (swiotlb_nr_tbl()) {
1585 		return ttm_dma_populate((void *)ttm, dev, ctx);
1586 	}
1587 #endif
1588 
1589 	r = ttm_pool_populate(ttm, ctx);
1590 	if (r) {
1591 		return r;
1592 	}
1593 
1594 	for (i = 0; i < ttm->num_pages; i++) {
1595 		dma_addr_t addr;
1596 
1597 		addr = dma_map_page(dev, ttm->pages[i], 0, PAGE_SIZE,
1598 				    DMA_BIDIRECTIONAL);
1599 
1600 		if (dma_mapping_error(dev, addr)) {
1601 			while (i--) {
1602 				dma_unmap_page(dev, ttm_dma->dma_address[i],
1603 					       PAGE_SIZE, DMA_BIDIRECTIONAL);
1604 				ttm_dma->dma_address[i] = 0;
1605 			}
1606 			ttm_pool_unpopulate(ttm);
1607 			return -EFAULT;
1608 		}
1609 
1610 		ttm_dma->dma_address[i] = addr;
1611 	}
1612 	return 0;
1613 }
1614 
1615 static void
1616 nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
1617 {
1618 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1619 	struct nouveau_drm *drm;
1620 	struct device *dev;
1621 	unsigned i;
1622 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1623 
1624 	if (slave)
1625 		return;
1626 
1627 	drm = nouveau_bdev(ttm->bdev);
1628 	dev = drm->dev->dev;
1629 
1630 #if IS_ENABLED(CONFIG_AGP)
1631 	if (drm->agp.bridge) {
1632 		ttm_agp_tt_unpopulate(ttm);
1633 		return;
1634 	}
1635 #endif
1636 
1637 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1638 	if (swiotlb_nr_tbl()) {
1639 		ttm_dma_unpopulate((void *)ttm, dev);
1640 		return;
1641 	}
1642 #endif
1643 
1644 	for (i = 0; i < ttm->num_pages; i++) {
1645 		if (ttm_dma->dma_address[i]) {
1646 			dma_unmap_page(dev, ttm_dma->dma_address[i], PAGE_SIZE,
1647 				       DMA_BIDIRECTIONAL);
1648 		}
1649 	}
1650 
1651 	ttm_pool_unpopulate(ttm);
1652 }
1653 
1654 void
1655 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1656 {
1657 	struct reservation_object *resv = nvbo->bo.resv;
1658 
1659 	if (exclusive)
1660 		reservation_object_add_excl_fence(resv, &fence->base);
1661 	else if (fence)
1662 		reservation_object_add_shared_fence(resv, &fence->base);
1663 }
1664 
1665 struct ttm_bo_driver nouveau_bo_driver = {
1666 	.ttm_tt_create = &nouveau_ttm_tt_create,
1667 	.ttm_tt_populate = &nouveau_ttm_tt_populate,
1668 	.ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1669 	.invalidate_caches = nouveau_bo_invalidate_caches,
1670 	.init_mem_type = nouveau_bo_init_mem_type,
1671 	.eviction_valuable = ttm_bo_eviction_valuable,
1672 	.evict_flags = nouveau_bo_evict_flags,
1673 	.move_notify = nouveau_bo_move_ntfy,
1674 	.move = nouveau_bo_move,
1675 	.verify_access = nouveau_bo_verify_access,
1676 	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1677 	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1678 	.io_mem_free = &nouveau_ttm_io_mem_free,
1679 };
1680