xref: /linux/drivers/gpu/drm/nouveau/nouveau_bo.c (revision 071bf69a0220253a44acb8b2a27f7a262b9a46bf)
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 
41 /*
42  * NV10-NV40 tiling helpers
43  */
44 
45 static void
46 nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
47 			   u32 addr, u32 size, u32 pitch, u32 flags)
48 {
49 	struct nouveau_drm *drm = nouveau_drm(dev);
50 	int i = reg - drm->tile.reg;
51 	struct nvkm_device *device = nvxx_device(&drm->device);
52 	struct nvkm_fb *fb = device->fb;
53 	struct nvkm_fb_tile *tile = &fb->tile.region[i];
54 
55 	nouveau_fence_unref(&reg->fence);
56 
57 	if (tile->pitch)
58 		nvkm_fb_tile_fini(fb, i, tile);
59 
60 	if (pitch)
61 		nvkm_fb_tile_init(fb, i, addr, size, pitch, flags, tile);
62 
63 	nvkm_fb_tile_prog(fb, i, tile);
64 }
65 
66 static struct nouveau_drm_tile *
67 nv10_bo_get_tile_region(struct drm_device *dev, int i)
68 {
69 	struct nouveau_drm *drm = nouveau_drm(dev);
70 	struct nouveau_drm_tile *tile = &drm->tile.reg[i];
71 
72 	spin_lock(&drm->tile.lock);
73 
74 	if (!tile->used &&
75 	    (!tile->fence || nouveau_fence_done(tile->fence)))
76 		tile->used = true;
77 	else
78 		tile = NULL;
79 
80 	spin_unlock(&drm->tile.lock);
81 	return tile;
82 }
83 
84 static void
85 nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
86 			struct fence *fence)
87 {
88 	struct nouveau_drm *drm = nouveau_drm(dev);
89 
90 	if (tile) {
91 		spin_lock(&drm->tile.lock);
92 		tile->fence = (struct nouveau_fence *)fence_get(fence);
93 		tile->used = false;
94 		spin_unlock(&drm->tile.lock);
95 	}
96 }
97 
98 static struct nouveau_drm_tile *
99 nv10_bo_set_tiling(struct drm_device *dev, u32 addr,
100 		   u32 size, u32 pitch, u32 flags)
101 {
102 	struct nouveau_drm *drm = nouveau_drm(dev);
103 	struct nvkm_fb *fb = nvxx_fb(&drm->device);
104 	struct nouveau_drm_tile *tile, *found = NULL;
105 	int i;
106 
107 	for (i = 0; i < fb->tile.regions; i++) {
108 		tile = nv10_bo_get_tile_region(dev, i);
109 
110 		if (pitch && !found) {
111 			found = tile;
112 			continue;
113 
114 		} else if (tile && fb->tile.region[i].pitch) {
115 			/* Kill an unused tile region. */
116 			nv10_bo_update_tile_region(dev, tile, 0, 0, 0, 0);
117 		}
118 
119 		nv10_bo_put_tile_region(dev, tile, NULL);
120 	}
121 
122 	if (found)
123 		nv10_bo_update_tile_region(dev, found, addr, size,
124 					    pitch, flags);
125 	return found;
126 }
127 
128 static void
129 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
130 {
131 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
132 	struct drm_device *dev = drm->dev;
133 	struct nouveau_bo *nvbo = nouveau_bo(bo);
134 
135 	if (unlikely(nvbo->gem.filp))
136 		DRM_ERROR("bo %p still attached to GEM object\n", bo);
137 	WARN_ON(nvbo->pin_refcnt > 0);
138 	nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
139 	kfree(nvbo);
140 }
141 
142 static void
143 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
144 		       int *align, int *size)
145 {
146 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
147 	struct nvif_device *device = &drm->device;
148 
149 	if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
150 		if (nvbo->tile_mode) {
151 			if (device->info.chipset >= 0x40) {
152 				*align = 65536;
153 				*size = roundup(*size, 64 * nvbo->tile_mode);
154 
155 			} else if (device->info.chipset >= 0x30) {
156 				*align = 32768;
157 				*size = roundup(*size, 64 * nvbo->tile_mode);
158 
159 			} else if (device->info.chipset >= 0x20) {
160 				*align = 16384;
161 				*size = roundup(*size, 64 * nvbo->tile_mode);
162 
163 			} else if (device->info.chipset >= 0x10) {
164 				*align = 16384;
165 				*size = roundup(*size, 32 * nvbo->tile_mode);
166 			}
167 		}
168 	} else {
169 		*size = roundup(*size, (1 << nvbo->page_shift));
170 		*align = max((1 <<  nvbo->page_shift), *align);
171 	}
172 
173 	*size = roundup(*size, PAGE_SIZE);
174 }
175 
176 int
177 nouveau_bo_new(struct drm_device *dev, int size, int align,
178 	       uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
179 	       struct sg_table *sg, struct reservation_object *robj,
180 	       struct nouveau_bo **pnvbo)
181 {
182 	struct nouveau_drm *drm = nouveau_drm(dev);
183 	struct nouveau_bo *nvbo;
184 	size_t acc_size;
185 	int ret;
186 	int type = ttm_bo_type_device;
187 	int lpg_shift = 12;
188 	int max_size;
189 
190 	if (drm->client.vm)
191 		lpg_shift = drm->client.vm->mmu->lpg_shift;
192 	max_size = INT_MAX & ~((1 << lpg_shift) - 1);
193 
194 	if (size <= 0 || size > max_size) {
195 		NV_WARN(drm, "skipped size %x\n", (u32)size);
196 		return -EINVAL;
197 	}
198 
199 	if (sg)
200 		type = ttm_bo_type_sg;
201 
202 	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
203 	if (!nvbo)
204 		return -ENOMEM;
205 	INIT_LIST_HEAD(&nvbo->head);
206 	INIT_LIST_HEAD(&nvbo->entry);
207 	INIT_LIST_HEAD(&nvbo->vma_list);
208 	nvbo->tile_mode = tile_mode;
209 	nvbo->tile_flags = tile_flags;
210 	nvbo->bo.bdev = &drm->ttm.bdev;
211 
212 	nvbo->force_coherent = flags & TTM_PL_FLAG_UNCACHED;
213 
214 	nvbo->page_shift = 12;
215 	if (drm->client.vm) {
216 		if (!(flags & TTM_PL_FLAG_TT) && size > 256 * 1024)
217 			nvbo->page_shift = drm->client.vm->mmu->lpg_shift;
218 	}
219 
220 	nouveau_bo_fixup_align(nvbo, flags, &align, &size);
221 	nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
222 	nouveau_bo_placement_set(nvbo, flags, 0);
223 
224 	acc_size = ttm_bo_dma_acc_size(&drm->ttm.bdev, size,
225 				       sizeof(struct nouveau_bo));
226 
227 	ret = ttm_bo_init(&drm->ttm.bdev, &nvbo->bo, size,
228 			  type, &nvbo->placement,
229 			  align >> PAGE_SHIFT, false, NULL, acc_size, sg,
230 			  robj, nouveau_bo_del_ttm);
231 	if (ret) {
232 		/* ttm will call nouveau_bo_del_ttm if it fails.. */
233 		return ret;
234 	}
235 
236 	*pnvbo = nvbo;
237 	return 0;
238 }
239 
240 static void
241 set_placement_list(struct ttm_place *pl, unsigned *n, uint32_t type, uint32_t flags)
242 {
243 	*n = 0;
244 
245 	if (type & TTM_PL_FLAG_VRAM)
246 		pl[(*n)++].flags = TTM_PL_FLAG_VRAM | flags;
247 	if (type & TTM_PL_FLAG_TT)
248 		pl[(*n)++].flags = TTM_PL_FLAG_TT | flags;
249 	if (type & TTM_PL_FLAG_SYSTEM)
250 		pl[(*n)++].flags = TTM_PL_FLAG_SYSTEM | flags;
251 }
252 
253 static void
254 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
255 {
256 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
257 	u32 vram_pages = drm->device.info.ram_size >> PAGE_SHIFT;
258 	unsigned i, fpfn, lpfn;
259 
260 	if (drm->device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
261 	    nvbo->tile_mode && (type & TTM_PL_FLAG_VRAM) &&
262 	    nvbo->bo.mem.num_pages < vram_pages / 4) {
263 		/*
264 		 * Make sure that the color and depth buffers are handled
265 		 * by independent memory controller units. Up to a 9x
266 		 * speed up when alpha-blending and depth-test are enabled
267 		 * at the same time.
268 		 */
269 		if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
270 			fpfn = vram_pages / 2;
271 			lpfn = ~0;
272 		} else {
273 			fpfn = 0;
274 			lpfn = vram_pages / 2;
275 		}
276 		for (i = 0; i < nvbo->placement.num_placement; ++i) {
277 			nvbo->placements[i].fpfn = fpfn;
278 			nvbo->placements[i].lpfn = lpfn;
279 		}
280 		for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
281 			nvbo->busy_placements[i].fpfn = fpfn;
282 			nvbo->busy_placements[i].lpfn = lpfn;
283 		}
284 	}
285 }
286 
287 void
288 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
289 {
290 	struct ttm_placement *pl = &nvbo->placement;
291 	uint32_t flags = (nvbo->force_coherent ? TTM_PL_FLAG_UNCACHED :
292 						 TTM_PL_MASK_CACHING) |
293 			 (nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
294 
295 	pl->placement = nvbo->placements;
296 	set_placement_list(nvbo->placements, &pl->num_placement,
297 			   type, flags);
298 
299 	pl->busy_placement = nvbo->busy_placements;
300 	set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
301 			   type | busy, flags);
302 
303 	set_placement_range(nvbo, type);
304 }
305 
306 int
307 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype, bool contig)
308 {
309 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
310 	struct ttm_buffer_object *bo = &nvbo->bo;
311 	bool force = false, evict = false;
312 	int ret;
313 
314 	ret = ttm_bo_reserve(bo, false, false, NULL);
315 	if (ret)
316 		return ret;
317 
318 	if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
319 	    memtype == TTM_PL_FLAG_VRAM && contig) {
320 		if (nvbo->tile_flags & NOUVEAU_GEM_TILE_NONCONTIG) {
321 			if (bo->mem.mem_type == TTM_PL_VRAM) {
322 				struct nvkm_mem *mem = bo->mem.mm_node;
323 				if (!list_is_singular(&mem->regions))
324 					evict = true;
325 			}
326 			nvbo->tile_flags &= ~NOUVEAU_GEM_TILE_NONCONTIG;
327 			force = true;
328 		}
329 	}
330 
331 	if (nvbo->pin_refcnt) {
332 		if (!(memtype & (1 << bo->mem.mem_type)) || evict) {
333 			NV_ERROR(drm, "bo %p pinned elsewhere: "
334 				      "0x%08x vs 0x%08x\n", bo,
335 				 1 << bo->mem.mem_type, memtype);
336 			ret = -EBUSY;
337 		}
338 		nvbo->pin_refcnt++;
339 		goto out;
340 	}
341 
342 	if (evict) {
343 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT, 0);
344 		ret = nouveau_bo_validate(nvbo, false, false);
345 		if (ret)
346 			goto out;
347 	}
348 
349 	nvbo->pin_refcnt++;
350 	nouveau_bo_placement_set(nvbo, memtype, 0);
351 
352 	/* drop pin_refcnt temporarily, so we don't trip the assertion
353 	 * in nouveau_bo_move() that makes sure we're not trying to
354 	 * move a pinned buffer
355 	 */
356 	nvbo->pin_refcnt--;
357 	ret = nouveau_bo_validate(nvbo, false, false);
358 	if (ret)
359 		goto out;
360 	nvbo->pin_refcnt++;
361 
362 	switch (bo->mem.mem_type) {
363 	case TTM_PL_VRAM:
364 		drm->gem.vram_available -= bo->mem.size;
365 		break;
366 	case TTM_PL_TT:
367 		drm->gem.gart_available -= bo->mem.size;
368 		break;
369 	default:
370 		break;
371 	}
372 
373 out:
374 	if (force && ret)
375 		nvbo->tile_flags |= NOUVEAU_GEM_TILE_NONCONTIG;
376 	ttm_bo_unreserve(bo);
377 	return ret;
378 }
379 
380 int
381 nouveau_bo_unpin(struct nouveau_bo *nvbo)
382 {
383 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
384 	struct ttm_buffer_object *bo = &nvbo->bo;
385 	int ret, ref;
386 
387 	ret = ttm_bo_reserve(bo, false, false, NULL);
388 	if (ret)
389 		return ret;
390 
391 	ref = --nvbo->pin_refcnt;
392 	WARN_ON_ONCE(ref < 0);
393 	if (ref)
394 		goto out;
395 
396 	nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
397 
398 	ret = nouveau_bo_validate(nvbo, false, false);
399 	if (ret == 0) {
400 		switch (bo->mem.mem_type) {
401 		case TTM_PL_VRAM:
402 			drm->gem.vram_available += bo->mem.size;
403 			break;
404 		case TTM_PL_TT:
405 			drm->gem.gart_available += bo->mem.size;
406 			break;
407 		default:
408 			break;
409 		}
410 	}
411 
412 out:
413 	ttm_bo_unreserve(bo);
414 	return ret;
415 }
416 
417 int
418 nouveau_bo_map(struct nouveau_bo *nvbo)
419 {
420 	int ret;
421 
422 	ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
423 	if (ret)
424 		return ret;
425 
426 	ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
427 
428 	ttm_bo_unreserve(&nvbo->bo);
429 	return ret;
430 }
431 
432 void
433 nouveau_bo_unmap(struct nouveau_bo *nvbo)
434 {
435 	if (!nvbo)
436 		return;
437 
438 	ttm_bo_kunmap(&nvbo->kmap);
439 }
440 
441 void
442 nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
443 {
444 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
445 	struct nvkm_device *device = nvxx_device(&drm->device);
446 	struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
447 	int i;
448 
449 	if (!ttm_dma)
450 		return;
451 
452 	/* Don't waste time looping if the object is coherent */
453 	if (nvbo->force_coherent)
454 		return;
455 
456 	for (i = 0; i < ttm_dma->ttm.num_pages; i++)
457 		dma_sync_single_for_device(device->dev, ttm_dma->dma_address[i],
458 					   PAGE_SIZE, DMA_TO_DEVICE);
459 }
460 
461 void
462 nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
463 {
464 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
465 	struct nvkm_device *device = nvxx_device(&drm->device);
466 	struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
467 	int i;
468 
469 	if (!ttm_dma)
470 		return;
471 
472 	/* Don't waste time looping if the object is coherent */
473 	if (nvbo->force_coherent)
474 		return;
475 
476 	for (i = 0; i < ttm_dma->ttm.num_pages; i++)
477 		dma_sync_single_for_cpu(device->dev, ttm_dma->dma_address[i],
478 					PAGE_SIZE, DMA_FROM_DEVICE);
479 }
480 
481 int
482 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
483 		    bool no_wait_gpu)
484 {
485 	int ret;
486 
487 	ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement,
488 			      interruptible, no_wait_gpu);
489 	if (ret)
490 		return ret;
491 
492 	nouveau_bo_sync_for_device(nvbo);
493 
494 	return 0;
495 }
496 
497 void
498 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
499 {
500 	bool is_iomem;
501 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
502 
503 	mem += index;
504 
505 	if (is_iomem)
506 		iowrite16_native(val, (void __force __iomem *)mem);
507 	else
508 		*mem = val;
509 }
510 
511 u32
512 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
513 {
514 	bool is_iomem;
515 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
516 
517 	mem += index;
518 
519 	if (is_iomem)
520 		return ioread32_native((void __force __iomem *)mem);
521 	else
522 		return *mem;
523 }
524 
525 void
526 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
527 {
528 	bool is_iomem;
529 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
530 
531 	mem += index;
532 
533 	if (is_iomem)
534 		iowrite32_native(val, (void __force __iomem *)mem);
535 	else
536 		*mem = val;
537 }
538 
539 static struct ttm_tt *
540 nouveau_ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
541 		      uint32_t page_flags, struct page *dummy_read)
542 {
543 #if IS_ENABLED(CONFIG_AGP)
544 	struct nouveau_drm *drm = nouveau_bdev(bdev);
545 
546 	if (drm->agp.bridge) {
547 		return ttm_agp_tt_create(bdev, drm->agp.bridge, size,
548 					 page_flags, dummy_read);
549 	}
550 #endif
551 
552 	return nouveau_sgdma_create_ttm(bdev, size, page_flags, dummy_read);
553 }
554 
555 static int
556 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
557 {
558 	/* We'll do this from user space. */
559 	return 0;
560 }
561 
562 static int
563 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
564 			 struct ttm_mem_type_manager *man)
565 {
566 	struct nouveau_drm *drm = nouveau_bdev(bdev);
567 
568 	switch (type) {
569 	case TTM_PL_SYSTEM:
570 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
571 		man->available_caching = TTM_PL_MASK_CACHING;
572 		man->default_caching = TTM_PL_FLAG_CACHED;
573 		break;
574 	case TTM_PL_VRAM:
575 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
576 			     TTM_MEMTYPE_FLAG_MAPPABLE;
577 		man->available_caching = TTM_PL_FLAG_UNCACHED |
578 					 TTM_PL_FLAG_WC;
579 		man->default_caching = TTM_PL_FLAG_WC;
580 
581 		if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
582 			/* Some BARs do not support being ioremapped WC */
583 			if (nvxx_bar(&drm->device)->iomap_uncached) {
584 				man->available_caching = TTM_PL_FLAG_UNCACHED;
585 				man->default_caching = TTM_PL_FLAG_UNCACHED;
586 			}
587 
588 			man->func = &nouveau_vram_manager;
589 			man->io_reserve_fastpath = false;
590 			man->use_io_reserve_lru = true;
591 		} else {
592 			man->func = &ttm_bo_manager_func;
593 		}
594 		break;
595 	case TTM_PL_TT:
596 		if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA)
597 			man->func = &nouveau_gart_manager;
598 		else
599 		if (!drm->agp.bridge)
600 			man->func = &nv04_gart_manager;
601 		else
602 			man->func = &ttm_bo_manager_func;
603 
604 		if (drm->agp.bridge) {
605 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
606 			man->available_caching = TTM_PL_FLAG_UNCACHED |
607 				TTM_PL_FLAG_WC;
608 			man->default_caching = TTM_PL_FLAG_WC;
609 		} else {
610 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
611 				     TTM_MEMTYPE_FLAG_CMA;
612 			man->available_caching = TTM_PL_MASK_CACHING;
613 			man->default_caching = TTM_PL_FLAG_CACHED;
614 		}
615 
616 		break;
617 	default:
618 		return -EINVAL;
619 	}
620 	return 0;
621 }
622 
623 static void
624 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
625 {
626 	struct nouveau_bo *nvbo = nouveau_bo(bo);
627 
628 	switch (bo->mem.mem_type) {
629 	case TTM_PL_VRAM:
630 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
631 					 TTM_PL_FLAG_SYSTEM);
632 		break;
633 	default:
634 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
635 		break;
636 	}
637 
638 	*pl = nvbo->placement;
639 }
640 
641 
642 static int
643 nve0_bo_move_init(struct nouveau_channel *chan, u32 handle)
644 {
645 	int ret = RING_SPACE(chan, 2);
646 	if (ret == 0) {
647 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
648 		OUT_RING  (chan, handle & 0x0000ffff);
649 		FIRE_RING (chan);
650 	}
651 	return ret;
652 }
653 
654 static int
655 nve0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
656 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
657 {
658 	struct nvkm_mem *node = old_mem->mm_node;
659 	int ret = RING_SPACE(chan, 10);
660 	if (ret == 0) {
661 		BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
662 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
663 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
664 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
665 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
666 		OUT_RING  (chan, PAGE_SIZE);
667 		OUT_RING  (chan, PAGE_SIZE);
668 		OUT_RING  (chan, PAGE_SIZE);
669 		OUT_RING  (chan, new_mem->num_pages);
670 		BEGIN_IMC0(chan, NvSubCopy, 0x0300, 0x0386);
671 	}
672 	return ret;
673 }
674 
675 static int
676 nvc0_bo_move_init(struct nouveau_channel *chan, u32 handle)
677 {
678 	int ret = RING_SPACE(chan, 2);
679 	if (ret == 0) {
680 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
681 		OUT_RING  (chan, handle);
682 	}
683 	return ret;
684 }
685 
686 static int
687 nvc0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
688 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
689 {
690 	struct nvkm_mem *node = old_mem->mm_node;
691 	u64 src_offset = node->vma[0].offset;
692 	u64 dst_offset = node->vma[1].offset;
693 	u32 page_count = new_mem->num_pages;
694 	int ret;
695 
696 	page_count = new_mem->num_pages;
697 	while (page_count) {
698 		int line_count = (page_count > 8191) ? 8191 : page_count;
699 
700 		ret = RING_SPACE(chan, 11);
701 		if (ret)
702 			return ret;
703 
704 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 8);
705 		OUT_RING  (chan, upper_32_bits(src_offset));
706 		OUT_RING  (chan, lower_32_bits(src_offset));
707 		OUT_RING  (chan, upper_32_bits(dst_offset));
708 		OUT_RING  (chan, lower_32_bits(dst_offset));
709 		OUT_RING  (chan, PAGE_SIZE);
710 		OUT_RING  (chan, PAGE_SIZE);
711 		OUT_RING  (chan, PAGE_SIZE);
712 		OUT_RING  (chan, line_count);
713 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
714 		OUT_RING  (chan, 0x00000110);
715 
716 		page_count -= line_count;
717 		src_offset += (PAGE_SIZE * line_count);
718 		dst_offset += (PAGE_SIZE * line_count);
719 	}
720 
721 	return 0;
722 }
723 
724 static int
725 nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
726 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
727 {
728 	struct nvkm_mem *node = old_mem->mm_node;
729 	u64 src_offset = node->vma[0].offset;
730 	u64 dst_offset = node->vma[1].offset;
731 	u32 page_count = new_mem->num_pages;
732 	int ret;
733 
734 	page_count = new_mem->num_pages;
735 	while (page_count) {
736 		int line_count = (page_count > 2047) ? 2047 : page_count;
737 
738 		ret = RING_SPACE(chan, 12);
739 		if (ret)
740 			return ret;
741 
742 		BEGIN_NVC0(chan, NvSubCopy, 0x0238, 2);
743 		OUT_RING  (chan, upper_32_bits(dst_offset));
744 		OUT_RING  (chan, lower_32_bits(dst_offset));
745 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 6);
746 		OUT_RING  (chan, upper_32_bits(src_offset));
747 		OUT_RING  (chan, lower_32_bits(src_offset));
748 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
749 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
750 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
751 		OUT_RING  (chan, line_count);
752 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
753 		OUT_RING  (chan, 0x00100110);
754 
755 		page_count -= line_count;
756 		src_offset += (PAGE_SIZE * line_count);
757 		dst_offset += (PAGE_SIZE * line_count);
758 	}
759 
760 	return 0;
761 }
762 
763 static int
764 nva3_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
765 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
766 {
767 	struct nvkm_mem *node = old_mem->mm_node;
768 	u64 src_offset = node->vma[0].offset;
769 	u64 dst_offset = node->vma[1].offset;
770 	u32 page_count = new_mem->num_pages;
771 	int ret;
772 
773 	page_count = new_mem->num_pages;
774 	while (page_count) {
775 		int line_count = (page_count > 8191) ? 8191 : page_count;
776 
777 		ret = RING_SPACE(chan, 11);
778 		if (ret)
779 			return ret;
780 
781 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
782 		OUT_RING  (chan, upper_32_bits(src_offset));
783 		OUT_RING  (chan, lower_32_bits(src_offset));
784 		OUT_RING  (chan, upper_32_bits(dst_offset));
785 		OUT_RING  (chan, lower_32_bits(dst_offset));
786 		OUT_RING  (chan, PAGE_SIZE);
787 		OUT_RING  (chan, PAGE_SIZE);
788 		OUT_RING  (chan, PAGE_SIZE);
789 		OUT_RING  (chan, line_count);
790 		BEGIN_NV04(chan, NvSubCopy, 0x0300, 1);
791 		OUT_RING  (chan, 0x00000110);
792 
793 		page_count -= line_count;
794 		src_offset += (PAGE_SIZE * line_count);
795 		dst_offset += (PAGE_SIZE * line_count);
796 	}
797 
798 	return 0;
799 }
800 
801 static int
802 nv98_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
803 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
804 {
805 	struct nvkm_mem *node = old_mem->mm_node;
806 	int ret = RING_SPACE(chan, 7);
807 	if (ret == 0) {
808 		BEGIN_NV04(chan, NvSubCopy, 0x0320, 6);
809 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
810 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
811 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
812 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
813 		OUT_RING  (chan, 0x00000000 /* COPY */);
814 		OUT_RING  (chan, new_mem->num_pages << PAGE_SHIFT);
815 	}
816 	return ret;
817 }
818 
819 static int
820 nv84_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
821 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
822 {
823 	struct nvkm_mem *node = old_mem->mm_node;
824 	int ret = RING_SPACE(chan, 7);
825 	if (ret == 0) {
826 		BEGIN_NV04(chan, NvSubCopy, 0x0304, 6);
827 		OUT_RING  (chan, new_mem->num_pages << PAGE_SHIFT);
828 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
829 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
830 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
831 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
832 		OUT_RING  (chan, 0x00000000 /* MODE_COPY, QUERY_NONE */);
833 	}
834 	return ret;
835 }
836 
837 static int
838 nv50_bo_move_init(struct nouveau_channel *chan, u32 handle)
839 {
840 	int ret = RING_SPACE(chan, 6);
841 	if (ret == 0) {
842 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
843 		OUT_RING  (chan, handle);
844 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 3);
845 		OUT_RING  (chan, chan->drm->ntfy.handle);
846 		OUT_RING  (chan, chan->vram.handle);
847 		OUT_RING  (chan, chan->vram.handle);
848 	}
849 
850 	return ret;
851 }
852 
853 static int
854 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
855 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
856 {
857 	struct nvkm_mem *node = old_mem->mm_node;
858 	u64 length = (new_mem->num_pages << PAGE_SHIFT);
859 	u64 src_offset = node->vma[0].offset;
860 	u64 dst_offset = node->vma[1].offset;
861 	int src_tiled = !!node->memtype;
862 	int dst_tiled = !!((struct nvkm_mem *)new_mem->mm_node)->memtype;
863 	int ret;
864 
865 	while (length) {
866 		u32 amount, stride, height;
867 
868 		ret = RING_SPACE(chan, 18 + 6 * (src_tiled + dst_tiled));
869 		if (ret)
870 			return ret;
871 
872 		amount  = min(length, (u64)(4 * 1024 * 1024));
873 		stride  = 16 * 4;
874 		height  = amount / stride;
875 
876 		if (src_tiled) {
877 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 7);
878 			OUT_RING  (chan, 0);
879 			OUT_RING  (chan, 0);
880 			OUT_RING  (chan, stride);
881 			OUT_RING  (chan, height);
882 			OUT_RING  (chan, 1);
883 			OUT_RING  (chan, 0);
884 			OUT_RING  (chan, 0);
885 		} else {
886 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 1);
887 			OUT_RING  (chan, 1);
888 		}
889 		if (dst_tiled) {
890 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 7);
891 			OUT_RING  (chan, 0);
892 			OUT_RING  (chan, 0);
893 			OUT_RING  (chan, stride);
894 			OUT_RING  (chan, height);
895 			OUT_RING  (chan, 1);
896 			OUT_RING  (chan, 0);
897 			OUT_RING  (chan, 0);
898 		} else {
899 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 1);
900 			OUT_RING  (chan, 1);
901 		}
902 
903 		BEGIN_NV04(chan, NvSubCopy, 0x0238, 2);
904 		OUT_RING  (chan, upper_32_bits(src_offset));
905 		OUT_RING  (chan, upper_32_bits(dst_offset));
906 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
907 		OUT_RING  (chan, lower_32_bits(src_offset));
908 		OUT_RING  (chan, lower_32_bits(dst_offset));
909 		OUT_RING  (chan, stride);
910 		OUT_RING  (chan, stride);
911 		OUT_RING  (chan, stride);
912 		OUT_RING  (chan, height);
913 		OUT_RING  (chan, 0x00000101);
914 		OUT_RING  (chan, 0x00000000);
915 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
916 		OUT_RING  (chan, 0);
917 
918 		length -= amount;
919 		src_offset += amount;
920 		dst_offset += amount;
921 	}
922 
923 	return 0;
924 }
925 
926 static int
927 nv04_bo_move_init(struct nouveau_channel *chan, u32 handle)
928 {
929 	int ret = RING_SPACE(chan, 4);
930 	if (ret == 0) {
931 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
932 		OUT_RING  (chan, handle);
933 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 1);
934 		OUT_RING  (chan, chan->drm->ntfy.handle);
935 	}
936 
937 	return ret;
938 }
939 
940 static inline uint32_t
941 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
942 		      struct nouveau_channel *chan, struct ttm_mem_reg *mem)
943 {
944 	if (mem->mem_type == TTM_PL_TT)
945 		return NvDmaTT;
946 	return chan->vram.handle;
947 }
948 
949 static int
950 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
951 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
952 {
953 	u32 src_offset = old_mem->start << PAGE_SHIFT;
954 	u32 dst_offset = new_mem->start << PAGE_SHIFT;
955 	u32 page_count = new_mem->num_pages;
956 	int ret;
957 
958 	ret = RING_SPACE(chan, 3);
959 	if (ret)
960 		return ret;
961 
962 	BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
963 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
964 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
965 
966 	page_count = new_mem->num_pages;
967 	while (page_count) {
968 		int line_count = (page_count > 2047) ? 2047 : page_count;
969 
970 		ret = RING_SPACE(chan, 11);
971 		if (ret)
972 			return ret;
973 
974 		BEGIN_NV04(chan, NvSubCopy,
975 				 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
976 		OUT_RING  (chan, src_offset);
977 		OUT_RING  (chan, dst_offset);
978 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
979 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
980 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
981 		OUT_RING  (chan, line_count);
982 		OUT_RING  (chan, 0x00000101);
983 		OUT_RING  (chan, 0x00000000);
984 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
985 		OUT_RING  (chan, 0);
986 
987 		page_count -= line_count;
988 		src_offset += (PAGE_SIZE * line_count);
989 		dst_offset += (PAGE_SIZE * line_count);
990 	}
991 
992 	return 0;
993 }
994 
995 static int
996 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
997 		     struct ttm_mem_reg *mem)
998 {
999 	struct nvkm_mem *old_node = bo->mem.mm_node;
1000 	struct nvkm_mem *new_node = mem->mm_node;
1001 	u64 size = (u64)mem->num_pages << PAGE_SHIFT;
1002 	int ret;
1003 
1004 	ret = nvkm_vm_get(drm->client.vm, size, old_node->page_shift,
1005 			  NV_MEM_ACCESS_RW, &old_node->vma[0]);
1006 	if (ret)
1007 		return ret;
1008 
1009 	ret = nvkm_vm_get(drm->client.vm, size, new_node->page_shift,
1010 			  NV_MEM_ACCESS_RW, &old_node->vma[1]);
1011 	if (ret) {
1012 		nvkm_vm_put(&old_node->vma[0]);
1013 		return ret;
1014 	}
1015 
1016 	nvkm_vm_map(&old_node->vma[0], old_node);
1017 	nvkm_vm_map(&old_node->vma[1], new_node);
1018 	return 0;
1019 }
1020 
1021 static int
1022 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
1023 		     bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1024 {
1025 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1026 	struct nouveau_channel *chan = drm->ttm.chan;
1027 	struct nouveau_cli *cli = (void *)chan->user.client;
1028 	struct nouveau_fence *fence;
1029 	int ret;
1030 
1031 	/* create temporary vmas for the transfer and attach them to the
1032 	 * old nvkm_mem node, these will get cleaned up after ttm has
1033 	 * destroyed the ttm_mem_reg
1034 	 */
1035 	if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
1036 		ret = nouveau_bo_move_prep(drm, bo, new_mem);
1037 		if (ret)
1038 			return ret;
1039 	}
1040 
1041 	mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
1042 	ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, intr);
1043 	if (ret == 0) {
1044 		ret = drm->ttm.move(chan, bo, &bo->mem, new_mem);
1045 		if (ret == 0) {
1046 			ret = nouveau_fence_new(chan, false, &fence);
1047 			if (ret == 0) {
1048 				ret = ttm_bo_move_accel_cleanup(bo,
1049 								&fence->base,
1050 								evict,
1051 								new_mem);
1052 				nouveau_fence_unref(&fence);
1053 			}
1054 		}
1055 	}
1056 	mutex_unlock(&cli->mutex);
1057 	return ret;
1058 }
1059 
1060 void
1061 nouveau_bo_move_init(struct nouveau_drm *drm)
1062 {
1063 	static const struct {
1064 		const char *name;
1065 		int engine;
1066 		s32 oclass;
1067 		int (*exec)(struct nouveau_channel *,
1068 			    struct ttm_buffer_object *,
1069 			    struct ttm_mem_reg *, struct ttm_mem_reg *);
1070 		int (*init)(struct nouveau_channel *, u32 handle);
1071 	} _methods[] = {
1072 		{  "COPY", 4, 0xc1b5, nve0_bo_move_copy, nve0_bo_move_init },
1073 		{  "GRCE", 0, 0xc1b5, nve0_bo_move_copy, nvc0_bo_move_init },
1074 		{  "COPY", 4, 0xc0b5, nve0_bo_move_copy, nve0_bo_move_init },
1075 		{  "GRCE", 0, 0xc0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1076 		{  "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
1077 		{  "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1078 		{  "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
1079 		{  "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1080 		{ "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
1081 		{ "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
1082 		{  "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
1083 		{ "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
1084 		{  "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
1085 		{  "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
1086 		{  "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
1087 		{},
1088 		{ "CRYPT", 0, 0x88b4, nv98_bo_move_exec, nv50_bo_move_init },
1089 	}, *mthd = _methods;
1090 	const char *name = "CPU";
1091 	int ret;
1092 
1093 	do {
1094 		struct nouveau_channel *chan;
1095 
1096 		if (mthd->engine)
1097 			chan = drm->cechan;
1098 		else
1099 			chan = drm->channel;
1100 		if (chan == NULL)
1101 			continue;
1102 
1103 		ret = nvif_object_init(&chan->user,
1104 				       mthd->oclass | (mthd->engine << 16),
1105 				       mthd->oclass, NULL, 0,
1106 				       &drm->ttm.copy);
1107 		if (ret == 0) {
1108 			ret = mthd->init(chan, drm->ttm.copy.handle);
1109 			if (ret) {
1110 				nvif_object_fini(&drm->ttm.copy);
1111 				continue;
1112 			}
1113 
1114 			drm->ttm.move = mthd->exec;
1115 			drm->ttm.chan = chan;
1116 			name = mthd->name;
1117 			break;
1118 		}
1119 	} while ((++mthd)->exec);
1120 
1121 	NV_INFO(drm, "MM: using %s for buffer copies\n", name);
1122 }
1123 
1124 static int
1125 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
1126 		      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1127 {
1128 	struct ttm_place placement_memtype = {
1129 		.fpfn = 0,
1130 		.lpfn = 0,
1131 		.flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
1132 	};
1133 	struct ttm_placement placement;
1134 	struct ttm_mem_reg tmp_mem;
1135 	int ret;
1136 
1137 	placement.num_placement = placement.num_busy_placement = 1;
1138 	placement.placement = placement.busy_placement = &placement_memtype;
1139 
1140 	tmp_mem = *new_mem;
1141 	tmp_mem.mm_node = NULL;
1142 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1143 	if (ret)
1144 		return ret;
1145 
1146 	ret = ttm_tt_bind(bo->ttm, &tmp_mem);
1147 	if (ret)
1148 		goto out;
1149 
1150 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, &tmp_mem);
1151 	if (ret)
1152 		goto out;
1153 
1154 	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
1155 out:
1156 	ttm_bo_mem_put(bo, &tmp_mem);
1157 	return ret;
1158 }
1159 
1160 static int
1161 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
1162 		      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1163 {
1164 	struct ttm_place placement_memtype = {
1165 		.fpfn = 0,
1166 		.lpfn = 0,
1167 		.flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
1168 	};
1169 	struct ttm_placement placement;
1170 	struct ttm_mem_reg tmp_mem;
1171 	int ret;
1172 
1173 	placement.num_placement = placement.num_busy_placement = 1;
1174 	placement.placement = placement.busy_placement = &placement_memtype;
1175 
1176 	tmp_mem = *new_mem;
1177 	tmp_mem.mm_node = NULL;
1178 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1179 	if (ret)
1180 		return ret;
1181 
1182 	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
1183 	if (ret)
1184 		goto out;
1185 
1186 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_mem);
1187 	if (ret)
1188 		goto out;
1189 
1190 out:
1191 	ttm_bo_mem_put(bo, &tmp_mem);
1192 	return ret;
1193 }
1194 
1195 static void
1196 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem)
1197 {
1198 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1199 	struct nvkm_vma *vma;
1200 
1201 	/* ttm can now (stupidly) pass the driver bos it didn't create... */
1202 	if (bo->destroy != nouveau_bo_del_ttm)
1203 		return;
1204 
1205 	list_for_each_entry(vma, &nvbo->vma_list, head) {
1206 		if (new_mem && new_mem->mem_type != TTM_PL_SYSTEM &&
1207 			      (new_mem->mem_type == TTM_PL_VRAM ||
1208 			       nvbo->page_shift != vma->vm->mmu->lpg_shift)) {
1209 			nvkm_vm_map(vma, new_mem->mm_node);
1210 		} else {
1211 			nvkm_vm_unmap(vma);
1212 		}
1213 	}
1214 }
1215 
1216 static int
1217 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
1218 		   struct nouveau_drm_tile **new_tile)
1219 {
1220 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1221 	struct drm_device *dev = drm->dev;
1222 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1223 	u64 offset = new_mem->start << PAGE_SHIFT;
1224 
1225 	*new_tile = NULL;
1226 	if (new_mem->mem_type != TTM_PL_VRAM)
1227 		return 0;
1228 
1229 	if (drm->device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
1230 		*new_tile = nv10_bo_set_tiling(dev, offset, new_mem->size,
1231 						nvbo->tile_mode,
1232 						nvbo->tile_flags);
1233 	}
1234 
1235 	return 0;
1236 }
1237 
1238 static void
1239 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1240 		      struct nouveau_drm_tile *new_tile,
1241 		      struct nouveau_drm_tile **old_tile)
1242 {
1243 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1244 	struct drm_device *dev = drm->dev;
1245 	struct fence *fence = reservation_object_get_excl(bo->resv);
1246 
1247 	nv10_bo_put_tile_region(dev, *old_tile, fence);
1248 	*old_tile = new_tile;
1249 }
1250 
1251 static int
1252 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
1253 		bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1254 {
1255 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1256 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1257 	struct ttm_mem_reg *old_mem = &bo->mem;
1258 	struct nouveau_drm_tile *new_tile = NULL;
1259 	int ret = 0;
1260 
1261 	ret = ttm_bo_wait(bo, intr, no_wait_gpu);
1262 	if (ret)
1263 		return ret;
1264 
1265 	if (nvbo->pin_refcnt)
1266 		NV_WARN(drm, "Moving pinned object %p!\n", nvbo);
1267 
1268 	if (drm->device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1269 		ret = nouveau_bo_vm_bind(bo, new_mem, &new_tile);
1270 		if (ret)
1271 			return ret;
1272 	}
1273 
1274 	/* Fake bo copy. */
1275 	if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
1276 		BUG_ON(bo->mem.mm_node != NULL);
1277 		bo->mem = *new_mem;
1278 		new_mem->mm_node = NULL;
1279 		goto out;
1280 	}
1281 
1282 	/* Hardware assisted copy. */
1283 	if (drm->ttm.move) {
1284 		if (new_mem->mem_type == TTM_PL_SYSTEM)
1285 			ret = nouveau_bo_move_flipd(bo, evict, intr,
1286 						    no_wait_gpu, new_mem);
1287 		else if (old_mem->mem_type == TTM_PL_SYSTEM)
1288 			ret = nouveau_bo_move_flips(bo, evict, intr,
1289 						    no_wait_gpu, new_mem);
1290 		else
1291 			ret = nouveau_bo_move_m2mf(bo, evict, intr,
1292 						   no_wait_gpu, new_mem);
1293 		if (!ret)
1294 			goto out;
1295 	}
1296 
1297 	/* Fallback to software copy. */
1298 	ret = ttm_bo_wait(bo, intr, no_wait_gpu);
1299 	if (ret == 0)
1300 		ret = ttm_bo_move_memcpy(bo, evict, intr, no_wait_gpu, new_mem);
1301 
1302 out:
1303 	if (drm->device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1304 		if (ret)
1305 			nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1306 		else
1307 			nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1308 	}
1309 
1310 	return ret;
1311 }
1312 
1313 static int
1314 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1315 {
1316 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1317 
1318 	return drm_vma_node_verify_access(&nvbo->gem.vma_node, filp);
1319 }
1320 
1321 static int
1322 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1323 {
1324 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1325 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1326 	struct nvkm_device *device = nvxx_device(&drm->device);
1327 	struct nvkm_mem *node = mem->mm_node;
1328 	int ret;
1329 
1330 	mem->bus.addr = NULL;
1331 	mem->bus.offset = 0;
1332 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
1333 	mem->bus.base = 0;
1334 	mem->bus.is_iomem = false;
1335 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
1336 		return -EINVAL;
1337 	switch (mem->mem_type) {
1338 	case TTM_PL_SYSTEM:
1339 		/* System memory */
1340 		return 0;
1341 	case TTM_PL_TT:
1342 #if IS_ENABLED(CONFIG_AGP)
1343 		if (drm->agp.bridge) {
1344 			mem->bus.offset = mem->start << PAGE_SHIFT;
1345 			mem->bus.base = drm->agp.base;
1346 			mem->bus.is_iomem = !drm->agp.cma;
1347 		}
1348 #endif
1349 		if (drm->device.info.family < NV_DEVICE_INFO_V0_TESLA || !node->memtype)
1350 			/* untiled */
1351 			break;
1352 		/* fallthrough, tiled memory */
1353 	case TTM_PL_VRAM:
1354 		mem->bus.offset = mem->start << PAGE_SHIFT;
1355 		mem->bus.base = device->func->resource_addr(device, 1);
1356 		mem->bus.is_iomem = true;
1357 		if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
1358 			struct nvkm_bar *bar = nvxx_bar(&drm->device);
1359 			int page_shift = 12;
1360 			if (drm->device.info.family >= NV_DEVICE_INFO_V0_FERMI)
1361 				page_shift = node->page_shift;
1362 
1363 			ret = nvkm_bar_umap(bar, node->size << 12, page_shift,
1364 					    &node->bar_vma);
1365 			if (ret)
1366 				return ret;
1367 
1368 			nvkm_vm_map(&node->bar_vma, node);
1369 			mem->bus.offset = node->bar_vma.offset;
1370 		}
1371 		break;
1372 	default:
1373 		return -EINVAL;
1374 	}
1375 	return 0;
1376 }
1377 
1378 static void
1379 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1380 {
1381 	struct nvkm_mem *node = mem->mm_node;
1382 
1383 	if (!node->bar_vma.node)
1384 		return;
1385 
1386 	nvkm_vm_unmap(&node->bar_vma);
1387 	nvkm_vm_put(&node->bar_vma);
1388 }
1389 
1390 static int
1391 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1392 {
1393 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1394 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1395 	struct nvkm_device *device = nvxx_device(&drm->device);
1396 	u32 mappable = device->func->resource_size(device, 1) >> PAGE_SHIFT;
1397 	int i, ret;
1398 
1399 	/* as long as the bo isn't in vram, and isn't tiled, we've got
1400 	 * nothing to do here.
1401 	 */
1402 	if (bo->mem.mem_type != TTM_PL_VRAM) {
1403 		if (drm->device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1404 		    !nouveau_bo_tile_layout(nvbo))
1405 			return 0;
1406 
1407 		if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1408 			nouveau_bo_placement_set(nvbo, TTM_PL_TT, 0);
1409 
1410 			ret = nouveau_bo_validate(nvbo, false, false);
1411 			if (ret)
1412 				return ret;
1413 		}
1414 		return 0;
1415 	}
1416 
1417 	/* make sure bo is in mappable vram */
1418 	if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1419 	    bo->mem.start + bo->mem.num_pages < mappable)
1420 		return 0;
1421 
1422 	for (i = 0; i < nvbo->placement.num_placement; ++i) {
1423 		nvbo->placements[i].fpfn = 0;
1424 		nvbo->placements[i].lpfn = mappable;
1425 	}
1426 
1427 	for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
1428 		nvbo->busy_placements[i].fpfn = 0;
1429 		nvbo->busy_placements[i].lpfn = mappable;
1430 	}
1431 
1432 	nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_VRAM, 0);
1433 	return nouveau_bo_validate(nvbo, false, false);
1434 }
1435 
1436 static int
1437 nouveau_ttm_tt_populate(struct ttm_tt *ttm)
1438 {
1439 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1440 	struct nouveau_drm *drm;
1441 	struct nvkm_device *device;
1442 	struct drm_device *dev;
1443 	struct device *pdev;
1444 	unsigned i;
1445 	int r;
1446 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1447 
1448 	if (ttm->state != tt_unpopulated)
1449 		return 0;
1450 
1451 	if (slave && ttm->sg) {
1452 		/* make userspace faulting work */
1453 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1454 						 ttm_dma->dma_address, ttm->num_pages);
1455 		ttm->state = tt_unbound;
1456 		return 0;
1457 	}
1458 
1459 	drm = nouveau_bdev(ttm->bdev);
1460 	device = nvxx_device(&drm->device);
1461 	dev = drm->dev;
1462 	pdev = device->dev;
1463 
1464 #if IS_ENABLED(CONFIG_AGP)
1465 	if (drm->agp.bridge) {
1466 		return ttm_agp_tt_populate(ttm);
1467 	}
1468 #endif
1469 
1470 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1471 	if (swiotlb_nr_tbl()) {
1472 		return ttm_dma_populate((void *)ttm, dev->dev);
1473 	}
1474 #endif
1475 
1476 	r = ttm_pool_populate(ttm);
1477 	if (r) {
1478 		return r;
1479 	}
1480 
1481 	for (i = 0; i < ttm->num_pages; i++) {
1482 		dma_addr_t addr;
1483 
1484 		addr = dma_map_page(pdev, ttm->pages[i], 0, PAGE_SIZE,
1485 				    DMA_BIDIRECTIONAL);
1486 
1487 		if (dma_mapping_error(pdev, addr)) {
1488 			while (i--) {
1489 				dma_unmap_page(pdev, ttm_dma->dma_address[i],
1490 					       PAGE_SIZE, DMA_BIDIRECTIONAL);
1491 				ttm_dma->dma_address[i] = 0;
1492 			}
1493 			ttm_pool_unpopulate(ttm);
1494 			return -EFAULT;
1495 		}
1496 
1497 		ttm_dma->dma_address[i] = addr;
1498 	}
1499 	return 0;
1500 }
1501 
1502 static void
1503 nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
1504 {
1505 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1506 	struct nouveau_drm *drm;
1507 	struct nvkm_device *device;
1508 	struct drm_device *dev;
1509 	struct device *pdev;
1510 	unsigned i;
1511 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1512 
1513 	if (slave)
1514 		return;
1515 
1516 	drm = nouveau_bdev(ttm->bdev);
1517 	device = nvxx_device(&drm->device);
1518 	dev = drm->dev;
1519 	pdev = device->dev;
1520 
1521 #if IS_ENABLED(CONFIG_AGP)
1522 	if (drm->agp.bridge) {
1523 		ttm_agp_tt_unpopulate(ttm);
1524 		return;
1525 	}
1526 #endif
1527 
1528 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1529 	if (swiotlb_nr_tbl()) {
1530 		ttm_dma_unpopulate((void *)ttm, dev->dev);
1531 		return;
1532 	}
1533 #endif
1534 
1535 	for (i = 0; i < ttm->num_pages; i++) {
1536 		if (ttm_dma->dma_address[i]) {
1537 			dma_unmap_page(pdev, ttm_dma->dma_address[i], PAGE_SIZE,
1538 				       DMA_BIDIRECTIONAL);
1539 		}
1540 	}
1541 
1542 	ttm_pool_unpopulate(ttm);
1543 }
1544 
1545 void
1546 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1547 {
1548 	struct reservation_object *resv = nvbo->bo.resv;
1549 
1550 	if (exclusive)
1551 		reservation_object_add_excl_fence(resv, &fence->base);
1552 	else if (fence)
1553 		reservation_object_add_shared_fence(resv, &fence->base);
1554 }
1555 
1556 struct ttm_bo_driver nouveau_bo_driver = {
1557 	.ttm_tt_create = &nouveau_ttm_tt_create,
1558 	.ttm_tt_populate = &nouveau_ttm_tt_populate,
1559 	.ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1560 	.invalidate_caches = nouveau_bo_invalidate_caches,
1561 	.init_mem_type = nouveau_bo_init_mem_type,
1562 	.evict_flags = nouveau_bo_evict_flags,
1563 	.move_notify = nouveau_bo_move_ntfy,
1564 	.move = nouveau_bo_move,
1565 	.verify_access = nouveau_bo_verify_access,
1566 	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1567 	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1568 	.io_mem_free = &nouveau_ttm_io_mem_free,
1569 	.lru_tail = &ttm_bo_default_lru_tail,
1570 	.swap_lru_tail = &ttm_bo_default_swap_lru_tail,
1571 };
1572 
1573 struct nvkm_vma *
1574 nouveau_bo_vma_find(struct nouveau_bo *nvbo, struct nvkm_vm *vm)
1575 {
1576 	struct nvkm_vma *vma;
1577 	list_for_each_entry(vma, &nvbo->vma_list, head) {
1578 		if (vma->vm == vm)
1579 			return vma;
1580 	}
1581 
1582 	return NULL;
1583 }
1584 
1585 int
1586 nouveau_bo_vma_add(struct nouveau_bo *nvbo, struct nvkm_vm *vm,
1587 		   struct nvkm_vma *vma)
1588 {
1589 	const u32 size = nvbo->bo.mem.num_pages << PAGE_SHIFT;
1590 	int ret;
1591 
1592 	ret = nvkm_vm_get(vm, size, nvbo->page_shift,
1593 			     NV_MEM_ACCESS_RW, vma);
1594 	if (ret)
1595 		return ret;
1596 
1597 	if ( nvbo->bo.mem.mem_type != TTM_PL_SYSTEM &&
1598 	    (nvbo->bo.mem.mem_type == TTM_PL_VRAM ||
1599 	     nvbo->page_shift != vma->vm->mmu->lpg_shift))
1600 		nvkm_vm_map(vma, nvbo->bo.mem.mm_node);
1601 
1602 	list_add_tail(&vma->head, &nvbo->vma_list);
1603 	vma->refcount = 1;
1604 	return 0;
1605 }
1606 
1607 void
1608 nouveau_bo_vma_del(struct nouveau_bo *nvbo, struct nvkm_vma *vma)
1609 {
1610 	if (vma->node) {
1611 		if (nvbo->bo.mem.mem_type != TTM_PL_SYSTEM)
1612 			nvkm_vm_unmap(vma);
1613 		nvkm_vm_put(vma);
1614 		list_del(&vma->head);
1615 	}
1616 }
1617