xref: /linux/drivers/gpu/drm/nouveau/nouveau_bo.c (revision 31a1b26f16e822577def5402ffc79cfe4aed2db9)
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 <core/engine.h>
31 #include <linux/swiotlb.h>
32 
33 #include <subdev/fb.h>
34 #include <subdev/vm.h>
35 #include <subdev/bar.h>
36 
37 #include "nouveau_drm.h"
38 #include "nouveau_dma.h"
39 #include "nouveau_fence.h"
40 
41 #include "nouveau_bo.h"
42 #include "nouveau_ttm.h"
43 #include "nouveau_gem.h"
44 
45 /*
46  * NV10-NV40 tiling helpers
47  */
48 
49 static void
50 nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
51 			   u32 addr, u32 size, u32 pitch, u32 flags)
52 {
53 	struct nouveau_drm *drm = nouveau_drm(dev);
54 	int i = reg - drm->tile.reg;
55 	struct nouveau_fb *pfb = nouveau_fb(drm->device);
56 	struct nouveau_fb_tile *tile = &pfb->tile.region[i];
57 	struct nouveau_engine *engine;
58 
59 	nouveau_fence_unref(&reg->fence);
60 
61 	if (tile->pitch)
62 		pfb->tile.fini(pfb, i, tile);
63 
64 	if (pitch)
65 		pfb->tile.init(pfb, i, addr, size, pitch, flags, tile);
66 
67 	pfb->tile.prog(pfb, i, tile);
68 
69 	if ((engine = nouveau_engine(pfb, NVDEV_ENGINE_GR)))
70 		engine->tile_prog(engine, i);
71 	if ((engine = nouveau_engine(pfb, NVDEV_ENGINE_MPEG)))
72 		engine->tile_prog(engine, i);
73 }
74 
75 static struct nouveau_drm_tile *
76 nv10_bo_get_tile_region(struct drm_device *dev, int i)
77 {
78 	struct nouveau_drm *drm = nouveau_drm(dev);
79 	struct nouveau_drm_tile *tile = &drm->tile.reg[i];
80 
81 	spin_lock(&drm->tile.lock);
82 
83 	if (!tile->used &&
84 	    (!tile->fence || nouveau_fence_done(tile->fence)))
85 		tile->used = true;
86 	else
87 		tile = NULL;
88 
89 	spin_unlock(&drm->tile.lock);
90 	return tile;
91 }
92 
93 static void
94 nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
95 			struct nouveau_fence *fence)
96 {
97 	struct nouveau_drm *drm = nouveau_drm(dev);
98 
99 	if (tile) {
100 		spin_lock(&drm->tile.lock);
101 		if (fence) {
102 			/* Mark it as pending. */
103 			tile->fence = fence;
104 			nouveau_fence_ref(fence);
105 		}
106 
107 		tile->used = false;
108 		spin_unlock(&drm->tile.lock);
109 	}
110 }
111 
112 static struct nouveau_drm_tile *
113 nv10_bo_set_tiling(struct drm_device *dev, u32 addr,
114 		   u32 size, u32 pitch, u32 flags)
115 {
116 	struct nouveau_drm *drm = nouveau_drm(dev);
117 	struct nouveau_fb *pfb = nouveau_fb(drm->device);
118 	struct nouveau_drm_tile *tile, *found = NULL;
119 	int i;
120 
121 	for (i = 0; i < pfb->tile.regions; i++) {
122 		tile = nv10_bo_get_tile_region(dev, i);
123 
124 		if (pitch && !found) {
125 			found = tile;
126 			continue;
127 
128 		} else if (tile && pfb->tile.region[i].pitch) {
129 			/* Kill an unused tile region. */
130 			nv10_bo_update_tile_region(dev, tile, 0, 0, 0, 0);
131 		}
132 
133 		nv10_bo_put_tile_region(dev, tile, NULL);
134 	}
135 
136 	if (found)
137 		nv10_bo_update_tile_region(dev, found, addr, size,
138 					    pitch, flags);
139 	return found;
140 }
141 
142 static void
143 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
144 {
145 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
146 	struct drm_device *dev = drm->dev;
147 	struct nouveau_bo *nvbo = nouveau_bo(bo);
148 
149 	if (unlikely(nvbo->gem))
150 		DRM_ERROR("bo %p still attached to GEM object\n", bo);
151 	WARN_ON(nvbo->pin_refcnt > 0);
152 	nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
153 	kfree(nvbo);
154 }
155 
156 static void
157 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
158 		       int *align, int *size)
159 {
160 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
161 	struct nouveau_device *device = nv_device(drm->device);
162 
163 	if (device->card_type < NV_50) {
164 		if (nvbo->tile_mode) {
165 			if (device->chipset >= 0x40) {
166 				*align = 65536;
167 				*size = roundup(*size, 64 * nvbo->tile_mode);
168 
169 			} else if (device->chipset >= 0x30) {
170 				*align = 32768;
171 				*size = roundup(*size, 64 * nvbo->tile_mode);
172 
173 			} else if (device->chipset >= 0x20) {
174 				*align = 16384;
175 				*size = roundup(*size, 64 * nvbo->tile_mode);
176 
177 			} else if (device->chipset >= 0x10) {
178 				*align = 16384;
179 				*size = roundup(*size, 32 * nvbo->tile_mode);
180 			}
181 		}
182 	} else {
183 		*size = roundup(*size, (1 << nvbo->page_shift));
184 		*align = max((1 <<  nvbo->page_shift), *align);
185 	}
186 
187 	*size = roundup(*size, PAGE_SIZE);
188 }
189 
190 int
191 nouveau_bo_new(struct drm_device *dev, int size, int align,
192 	       uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
193 	       struct sg_table *sg,
194 	       struct nouveau_bo **pnvbo)
195 {
196 	struct nouveau_drm *drm = nouveau_drm(dev);
197 	struct nouveau_bo *nvbo;
198 	size_t acc_size;
199 	int ret;
200 	int type = ttm_bo_type_device;
201 	int max_size = INT_MAX & ~((1 << drm->client.base.vm->vmm->lpg_shift) - 1);
202 
203 	if (size <= 0 || size > max_size) {
204 		nv_warn(drm, "skipped size %x\n", (u32)size);
205 		return -EINVAL;
206 	}
207 
208 	if (sg)
209 		type = ttm_bo_type_sg;
210 
211 	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
212 	if (!nvbo)
213 		return -ENOMEM;
214 	INIT_LIST_HEAD(&nvbo->head);
215 	INIT_LIST_HEAD(&nvbo->entry);
216 	INIT_LIST_HEAD(&nvbo->vma_list);
217 	nvbo->tile_mode = tile_mode;
218 	nvbo->tile_flags = tile_flags;
219 	nvbo->bo.bdev = &drm->ttm.bdev;
220 
221 	nvbo->page_shift = 12;
222 	if (drm->client.base.vm) {
223 		if (!(flags & TTM_PL_FLAG_TT) && size > 256 * 1024)
224 			nvbo->page_shift = drm->client.base.vm->vmm->lpg_shift;
225 	}
226 
227 	nouveau_bo_fixup_align(nvbo, flags, &align, &size);
228 	nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
229 	nouveau_bo_placement_set(nvbo, flags, 0);
230 
231 	acc_size = ttm_bo_dma_acc_size(&drm->ttm.bdev, size,
232 				       sizeof(struct nouveau_bo));
233 
234 	ret = ttm_bo_init(&drm->ttm.bdev, &nvbo->bo, size,
235 			  type, &nvbo->placement,
236 			  align >> PAGE_SHIFT, false, NULL, acc_size, sg,
237 			  nouveau_bo_del_ttm);
238 	if (ret) {
239 		/* ttm will call nouveau_bo_del_ttm if it fails.. */
240 		return ret;
241 	}
242 
243 	*pnvbo = nvbo;
244 	return 0;
245 }
246 
247 static void
248 set_placement_list(uint32_t *pl, unsigned *n, uint32_t type, uint32_t flags)
249 {
250 	*n = 0;
251 
252 	if (type & TTM_PL_FLAG_VRAM)
253 		pl[(*n)++] = TTM_PL_FLAG_VRAM | flags;
254 	if (type & TTM_PL_FLAG_TT)
255 		pl[(*n)++] = TTM_PL_FLAG_TT | flags;
256 	if (type & TTM_PL_FLAG_SYSTEM)
257 		pl[(*n)++] = TTM_PL_FLAG_SYSTEM | flags;
258 }
259 
260 static void
261 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
262 {
263 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
264 	struct nouveau_fb *pfb = nouveau_fb(drm->device);
265 	u32 vram_pages = pfb->ram->size >> PAGE_SHIFT;
266 
267 	if (nv_device(drm->device)->card_type == NV_10 &&
268 	    nvbo->tile_mode && (type & TTM_PL_FLAG_VRAM) &&
269 	    nvbo->bo.mem.num_pages < vram_pages / 4) {
270 		/*
271 		 * Make sure that the color and depth buffers are handled
272 		 * by independent memory controller units. Up to a 9x
273 		 * speed up when alpha-blending and depth-test are enabled
274 		 * at the same time.
275 		 */
276 		if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
277 			nvbo->placement.fpfn = vram_pages / 2;
278 			nvbo->placement.lpfn = ~0;
279 		} else {
280 			nvbo->placement.fpfn = 0;
281 			nvbo->placement.lpfn = vram_pages / 2;
282 		}
283 	}
284 }
285 
286 void
287 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
288 {
289 	struct ttm_placement *pl = &nvbo->placement;
290 	uint32_t flags = TTM_PL_MASK_CACHING |
291 		(nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
292 
293 	pl->placement = nvbo->placements;
294 	set_placement_list(nvbo->placements, &pl->num_placement,
295 			   type, flags);
296 
297 	pl->busy_placement = nvbo->busy_placements;
298 	set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
299 			   type | busy, flags);
300 
301 	set_placement_range(nvbo, type);
302 }
303 
304 int
305 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype)
306 {
307 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
308 	struct ttm_buffer_object *bo = &nvbo->bo;
309 	int ret;
310 
311 	ret = ttm_bo_reserve(bo, false, false, false, 0);
312 	if (ret)
313 		goto out;
314 
315 	if (nvbo->pin_refcnt && !(memtype & (1 << bo->mem.mem_type))) {
316 		NV_ERROR(drm, "bo %p pinned elsewhere: 0x%08x vs 0x%08x\n", bo,
317 			 1 << bo->mem.mem_type, memtype);
318 		ret = -EINVAL;
319 		goto out;
320 	}
321 
322 	if (nvbo->pin_refcnt++)
323 		goto out;
324 
325 	nouveau_bo_placement_set(nvbo, memtype, 0);
326 
327 	ret = nouveau_bo_validate(nvbo, false, false);
328 	if (ret == 0) {
329 		switch (bo->mem.mem_type) {
330 		case TTM_PL_VRAM:
331 			drm->gem.vram_available -= bo->mem.size;
332 			break;
333 		case TTM_PL_TT:
334 			drm->gem.gart_available -= bo->mem.size;
335 			break;
336 		default:
337 			break;
338 		}
339 	}
340 out:
341 	ttm_bo_unreserve(bo);
342 	return ret;
343 }
344 
345 int
346 nouveau_bo_unpin(struct nouveau_bo *nvbo)
347 {
348 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
349 	struct ttm_buffer_object *bo = &nvbo->bo;
350 	int ret, ref;
351 
352 	ret = ttm_bo_reserve(bo, false, false, false, 0);
353 	if (ret)
354 		return ret;
355 
356 	ref = --nvbo->pin_refcnt;
357 	WARN_ON_ONCE(ref < 0);
358 	if (ref)
359 		goto out;
360 
361 	nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
362 
363 	ret = nouveau_bo_validate(nvbo, false, false);
364 	if (ret == 0) {
365 		switch (bo->mem.mem_type) {
366 		case TTM_PL_VRAM:
367 			drm->gem.vram_available += bo->mem.size;
368 			break;
369 		case TTM_PL_TT:
370 			drm->gem.gart_available += bo->mem.size;
371 			break;
372 		default:
373 			break;
374 		}
375 	}
376 
377 out:
378 	ttm_bo_unreserve(bo);
379 	return ret;
380 }
381 
382 int
383 nouveau_bo_map(struct nouveau_bo *nvbo)
384 {
385 	int ret;
386 
387 	ret = ttm_bo_reserve(&nvbo->bo, false, false, false, 0);
388 	if (ret)
389 		return ret;
390 
391 	ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
392 	ttm_bo_unreserve(&nvbo->bo);
393 	return ret;
394 }
395 
396 void
397 nouveau_bo_unmap(struct nouveau_bo *nvbo)
398 {
399 	if (nvbo)
400 		ttm_bo_kunmap(&nvbo->kmap);
401 }
402 
403 int
404 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
405 		    bool no_wait_gpu)
406 {
407 	int ret;
408 
409 	ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement,
410 			      interruptible, no_wait_gpu);
411 	if (ret)
412 		return ret;
413 
414 	return 0;
415 }
416 
417 u16
418 nouveau_bo_rd16(struct nouveau_bo *nvbo, unsigned index)
419 {
420 	bool is_iomem;
421 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
422 	mem = &mem[index];
423 	if (is_iomem)
424 		return ioread16_native((void __force __iomem *)mem);
425 	else
426 		return *mem;
427 }
428 
429 void
430 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
431 {
432 	bool is_iomem;
433 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
434 	mem = &mem[index];
435 	if (is_iomem)
436 		iowrite16_native(val, (void __force __iomem *)mem);
437 	else
438 		*mem = val;
439 }
440 
441 u32
442 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
443 {
444 	bool is_iomem;
445 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
446 	mem = &mem[index];
447 	if (is_iomem)
448 		return ioread32_native((void __force __iomem *)mem);
449 	else
450 		return *mem;
451 }
452 
453 void
454 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
455 {
456 	bool is_iomem;
457 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
458 	mem = &mem[index];
459 	if (is_iomem)
460 		iowrite32_native(val, (void __force __iomem *)mem);
461 	else
462 		*mem = val;
463 }
464 
465 static struct ttm_tt *
466 nouveau_ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
467 		      uint32_t page_flags, struct page *dummy_read)
468 {
469 #if __OS_HAS_AGP
470 	struct nouveau_drm *drm = nouveau_bdev(bdev);
471 	struct drm_device *dev = drm->dev;
472 
473 	if (drm->agp.stat == ENABLED) {
474 		return ttm_agp_tt_create(bdev, dev->agp->bridge, size,
475 					 page_flags, dummy_read);
476 	}
477 #endif
478 
479 	return nouveau_sgdma_create_ttm(bdev, size, page_flags, dummy_read);
480 }
481 
482 static int
483 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
484 {
485 	/* We'll do this from user space. */
486 	return 0;
487 }
488 
489 static int
490 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
491 			 struct ttm_mem_type_manager *man)
492 {
493 	struct nouveau_drm *drm = nouveau_bdev(bdev);
494 
495 	switch (type) {
496 	case TTM_PL_SYSTEM:
497 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
498 		man->available_caching = TTM_PL_MASK_CACHING;
499 		man->default_caching = TTM_PL_FLAG_CACHED;
500 		break;
501 	case TTM_PL_VRAM:
502 		if (nv_device(drm->device)->card_type >= NV_50) {
503 			man->func = &nouveau_vram_manager;
504 			man->io_reserve_fastpath = false;
505 			man->use_io_reserve_lru = true;
506 		} else {
507 			man->func = &ttm_bo_manager_func;
508 		}
509 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
510 			     TTM_MEMTYPE_FLAG_MAPPABLE;
511 		man->available_caching = TTM_PL_FLAG_UNCACHED |
512 					 TTM_PL_FLAG_WC;
513 		man->default_caching = TTM_PL_FLAG_WC;
514 		break;
515 	case TTM_PL_TT:
516 		if (nv_device(drm->device)->card_type >= NV_50)
517 			man->func = &nouveau_gart_manager;
518 		else
519 		if (drm->agp.stat != ENABLED)
520 			man->func = &nv04_gart_manager;
521 		else
522 			man->func = &ttm_bo_manager_func;
523 
524 		if (drm->agp.stat == ENABLED) {
525 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
526 			man->available_caching = TTM_PL_FLAG_UNCACHED |
527 				TTM_PL_FLAG_WC;
528 			man->default_caching = TTM_PL_FLAG_WC;
529 		} else {
530 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
531 				     TTM_MEMTYPE_FLAG_CMA;
532 			man->available_caching = TTM_PL_MASK_CACHING;
533 			man->default_caching = TTM_PL_FLAG_CACHED;
534 		}
535 
536 		break;
537 	default:
538 		return -EINVAL;
539 	}
540 	return 0;
541 }
542 
543 static void
544 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
545 {
546 	struct nouveau_bo *nvbo = nouveau_bo(bo);
547 
548 	switch (bo->mem.mem_type) {
549 	case TTM_PL_VRAM:
550 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
551 					 TTM_PL_FLAG_SYSTEM);
552 		break;
553 	default:
554 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
555 		break;
556 	}
557 
558 	*pl = nvbo->placement;
559 }
560 
561 
562 /* GPU-assisted copy using NV_MEMORY_TO_MEMORY_FORMAT, can access
563  * TTM_PL_{VRAM,TT} directly.
564  */
565 
566 static int
567 nouveau_bo_move_accel_cleanup(struct nouveau_channel *chan,
568 			      struct nouveau_bo *nvbo, bool evict,
569 			      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
570 {
571 	struct nouveau_fence *fence = NULL;
572 	int ret;
573 
574 	ret = nouveau_fence_new(chan, false, &fence);
575 	if (ret)
576 		return ret;
577 
578 	ret = ttm_bo_move_accel_cleanup(&nvbo->bo, fence, evict,
579 					no_wait_gpu, new_mem);
580 	nouveau_fence_unref(&fence);
581 	return ret;
582 }
583 
584 static int
585 nve0_bo_move_init(struct nouveau_channel *chan, u32 handle)
586 {
587 	int ret = RING_SPACE(chan, 2);
588 	if (ret == 0) {
589 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
590 		OUT_RING  (chan, handle & 0x0000ffff);
591 		FIRE_RING (chan);
592 	}
593 	return ret;
594 }
595 
596 static int
597 nve0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
598 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
599 {
600 	struct nouveau_mem *node = old_mem->mm_node;
601 	int ret = RING_SPACE(chan, 10);
602 	if (ret == 0) {
603 		BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
604 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
605 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
606 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
607 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
608 		OUT_RING  (chan, PAGE_SIZE);
609 		OUT_RING  (chan, PAGE_SIZE);
610 		OUT_RING  (chan, PAGE_SIZE);
611 		OUT_RING  (chan, new_mem->num_pages);
612 		BEGIN_IMC0(chan, NvSubCopy, 0x0300, 0x0386);
613 	}
614 	return ret;
615 }
616 
617 static int
618 nvc0_bo_move_init(struct nouveau_channel *chan, u32 handle)
619 {
620 	int ret = RING_SPACE(chan, 2);
621 	if (ret == 0) {
622 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
623 		OUT_RING  (chan, handle);
624 	}
625 	return ret;
626 }
627 
628 static int
629 nvc0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
630 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
631 {
632 	struct nouveau_mem *node = old_mem->mm_node;
633 	u64 src_offset = node->vma[0].offset;
634 	u64 dst_offset = node->vma[1].offset;
635 	u32 page_count = new_mem->num_pages;
636 	int ret;
637 
638 	page_count = new_mem->num_pages;
639 	while (page_count) {
640 		int line_count = (page_count > 8191) ? 8191 : page_count;
641 
642 		ret = RING_SPACE(chan, 11);
643 		if (ret)
644 			return ret;
645 
646 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 8);
647 		OUT_RING  (chan, upper_32_bits(src_offset));
648 		OUT_RING  (chan, lower_32_bits(src_offset));
649 		OUT_RING  (chan, upper_32_bits(dst_offset));
650 		OUT_RING  (chan, lower_32_bits(dst_offset));
651 		OUT_RING  (chan, PAGE_SIZE);
652 		OUT_RING  (chan, PAGE_SIZE);
653 		OUT_RING  (chan, PAGE_SIZE);
654 		OUT_RING  (chan, line_count);
655 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
656 		OUT_RING  (chan, 0x00000110);
657 
658 		page_count -= line_count;
659 		src_offset += (PAGE_SIZE * line_count);
660 		dst_offset += (PAGE_SIZE * line_count);
661 	}
662 
663 	return 0;
664 }
665 
666 static int
667 nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
668 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
669 {
670 	struct nouveau_mem *node = old_mem->mm_node;
671 	u64 src_offset = node->vma[0].offset;
672 	u64 dst_offset = node->vma[1].offset;
673 	u32 page_count = new_mem->num_pages;
674 	int ret;
675 
676 	page_count = new_mem->num_pages;
677 	while (page_count) {
678 		int line_count = (page_count > 2047) ? 2047 : page_count;
679 
680 		ret = RING_SPACE(chan, 12);
681 		if (ret)
682 			return ret;
683 
684 		BEGIN_NVC0(chan, NvSubCopy, 0x0238, 2);
685 		OUT_RING  (chan, upper_32_bits(dst_offset));
686 		OUT_RING  (chan, lower_32_bits(dst_offset));
687 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 6);
688 		OUT_RING  (chan, upper_32_bits(src_offset));
689 		OUT_RING  (chan, lower_32_bits(src_offset));
690 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
691 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
692 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
693 		OUT_RING  (chan, line_count);
694 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
695 		OUT_RING  (chan, 0x00100110);
696 
697 		page_count -= line_count;
698 		src_offset += (PAGE_SIZE * line_count);
699 		dst_offset += (PAGE_SIZE * line_count);
700 	}
701 
702 	return 0;
703 }
704 
705 static int
706 nva3_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
707 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
708 {
709 	struct nouveau_mem *node = old_mem->mm_node;
710 	u64 src_offset = node->vma[0].offset;
711 	u64 dst_offset = node->vma[1].offset;
712 	u32 page_count = new_mem->num_pages;
713 	int ret;
714 
715 	page_count = new_mem->num_pages;
716 	while (page_count) {
717 		int line_count = (page_count > 8191) ? 8191 : page_count;
718 
719 		ret = RING_SPACE(chan, 11);
720 		if (ret)
721 			return ret;
722 
723 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
724 		OUT_RING  (chan, upper_32_bits(src_offset));
725 		OUT_RING  (chan, lower_32_bits(src_offset));
726 		OUT_RING  (chan, upper_32_bits(dst_offset));
727 		OUT_RING  (chan, lower_32_bits(dst_offset));
728 		OUT_RING  (chan, PAGE_SIZE);
729 		OUT_RING  (chan, PAGE_SIZE);
730 		OUT_RING  (chan, PAGE_SIZE);
731 		OUT_RING  (chan, line_count);
732 		BEGIN_NV04(chan, NvSubCopy, 0x0300, 1);
733 		OUT_RING  (chan, 0x00000110);
734 
735 		page_count -= line_count;
736 		src_offset += (PAGE_SIZE * line_count);
737 		dst_offset += (PAGE_SIZE * line_count);
738 	}
739 
740 	return 0;
741 }
742 
743 static int
744 nv98_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
745 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
746 {
747 	struct nouveau_mem *node = old_mem->mm_node;
748 	int ret = RING_SPACE(chan, 7);
749 	if (ret == 0) {
750 		BEGIN_NV04(chan, NvSubCopy, 0x0320, 6);
751 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
752 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
753 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
754 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
755 		OUT_RING  (chan, 0x00000000 /* COPY */);
756 		OUT_RING  (chan, new_mem->num_pages << PAGE_SHIFT);
757 	}
758 	return ret;
759 }
760 
761 static int
762 nv84_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
763 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
764 {
765 	struct nouveau_mem *node = old_mem->mm_node;
766 	int ret = RING_SPACE(chan, 7);
767 	if (ret == 0) {
768 		BEGIN_NV04(chan, NvSubCopy, 0x0304, 6);
769 		OUT_RING  (chan, new_mem->num_pages << PAGE_SHIFT);
770 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
771 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
772 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
773 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
774 		OUT_RING  (chan, 0x00000000 /* MODE_COPY, QUERY_NONE */);
775 	}
776 	return ret;
777 }
778 
779 static int
780 nv50_bo_move_init(struct nouveau_channel *chan, u32 handle)
781 {
782 	int ret = RING_SPACE(chan, 6);
783 	if (ret == 0) {
784 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
785 		OUT_RING  (chan, handle);
786 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 3);
787 		OUT_RING  (chan, NvNotify0);
788 		OUT_RING  (chan, NvDmaFB);
789 		OUT_RING  (chan, NvDmaFB);
790 	}
791 
792 	return ret;
793 }
794 
795 static int
796 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
797 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
798 {
799 	struct nouveau_mem *node = old_mem->mm_node;
800 	struct nouveau_bo *nvbo = nouveau_bo(bo);
801 	u64 length = (new_mem->num_pages << PAGE_SHIFT);
802 	u64 src_offset = node->vma[0].offset;
803 	u64 dst_offset = node->vma[1].offset;
804 	int ret;
805 
806 	while (length) {
807 		u32 amount, stride, height;
808 
809 		amount  = min(length, (u64)(4 * 1024 * 1024));
810 		stride  = 16 * 4;
811 		height  = amount / stride;
812 
813 		if (old_mem->mem_type == TTM_PL_VRAM &&
814 		    nouveau_bo_tile_layout(nvbo)) {
815 			ret = RING_SPACE(chan, 8);
816 			if (ret)
817 				return ret;
818 
819 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 7);
820 			OUT_RING  (chan, 0);
821 			OUT_RING  (chan, 0);
822 			OUT_RING  (chan, stride);
823 			OUT_RING  (chan, height);
824 			OUT_RING  (chan, 1);
825 			OUT_RING  (chan, 0);
826 			OUT_RING  (chan, 0);
827 		} else {
828 			ret = RING_SPACE(chan, 2);
829 			if (ret)
830 				return ret;
831 
832 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 1);
833 			OUT_RING  (chan, 1);
834 		}
835 		if (new_mem->mem_type == TTM_PL_VRAM &&
836 		    nouveau_bo_tile_layout(nvbo)) {
837 			ret = RING_SPACE(chan, 8);
838 			if (ret)
839 				return ret;
840 
841 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 7);
842 			OUT_RING  (chan, 0);
843 			OUT_RING  (chan, 0);
844 			OUT_RING  (chan, stride);
845 			OUT_RING  (chan, height);
846 			OUT_RING  (chan, 1);
847 			OUT_RING  (chan, 0);
848 			OUT_RING  (chan, 0);
849 		} else {
850 			ret = RING_SPACE(chan, 2);
851 			if (ret)
852 				return ret;
853 
854 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 1);
855 			OUT_RING  (chan, 1);
856 		}
857 
858 		ret = RING_SPACE(chan, 14);
859 		if (ret)
860 			return ret;
861 
862 		BEGIN_NV04(chan, NvSubCopy, 0x0238, 2);
863 		OUT_RING  (chan, upper_32_bits(src_offset));
864 		OUT_RING  (chan, upper_32_bits(dst_offset));
865 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
866 		OUT_RING  (chan, lower_32_bits(src_offset));
867 		OUT_RING  (chan, lower_32_bits(dst_offset));
868 		OUT_RING  (chan, stride);
869 		OUT_RING  (chan, stride);
870 		OUT_RING  (chan, stride);
871 		OUT_RING  (chan, height);
872 		OUT_RING  (chan, 0x00000101);
873 		OUT_RING  (chan, 0x00000000);
874 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
875 		OUT_RING  (chan, 0);
876 
877 		length -= amount;
878 		src_offset += amount;
879 		dst_offset += amount;
880 	}
881 
882 	return 0;
883 }
884 
885 static int
886 nv04_bo_move_init(struct nouveau_channel *chan, u32 handle)
887 {
888 	int ret = RING_SPACE(chan, 4);
889 	if (ret == 0) {
890 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
891 		OUT_RING  (chan, handle);
892 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 1);
893 		OUT_RING  (chan, NvNotify0);
894 	}
895 
896 	return ret;
897 }
898 
899 static inline uint32_t
900 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
901 		      struct nouveau_channel *chan, struct ttm_mem_reg *mem)
902 {
903 	if (mem->mem_type == TTM_PL_TT)
904 		return NvDmaTT;
905 	return NvDmaFB;
906 }
907 
908 static int
909 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
910 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
911 {
912 	u32 src_offset = old_mem->start << PAGE_SHIFT;
913 	u32 dst_offset = new_mem->start << PAGE_SHIFT;
914 	u32 page_count = new_mem->num_pages;
915 	int ret;
916 
917 	ret = RING_SPACE(chan, 3);
918 	if (ret)
919 		return ret;
920 
921 	BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
922 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
923 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
924 
925 	page_count = new_mem->num_pages;
926 	while (page_count) {
927 		int line_count = (page_count > 2047) ? 2047 : page_count;
928 
929 		ret = RING_SPACE(chan, 11);
930 		if (ret)
931 			return ret;
932 
933 		BEGIN_NV04(chan, NvSubCopy,
934 				 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
935 		OUT_RING  (chan, src_offset);
936 		OUT_RING  (chan, dst_offset);
937 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
938 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
939 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
940 		OUT_RING  (chan, line_count);
941 		OUT_RING  (chan, 0x00000101);
942 		OUT_RING  (chan, 0x00000000);
943 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
944 		OUT_RING  (chan, 0);
945 
946 		page_count -= line_count;
947 		src_offset += (PAGE_SIZE * line_count);
948 		dst_offset += (PAGE_SIZE * line_count);
949 	}
950 
951 	return 0;
952 }
953 
954 static int
955 nouveau_vma_getmap(struct nouveau_channel *chan, struct nouveau_bo *nvbo,
956 		   struct ttm_mem_reg *mem, struct nouveau_vma *vma)
957 {
958 	struct nouveau_mem *node = mem->mm_node;
959 	int ret;
960 
961 	ret = nouveau_vm_get(nv_client(chan->cli)->vm, mem->num_pages <<
962 			     PAGE_SHIFT, node->page_shift,
963 			     NV_MEM_ACCESS_RW, vma);
964 	if (ret)
965 		return ret;
966 
967 	if (mem->mem_type == TTM_PL_VRAM)
968 		nouveau_vm_map(vma, node);
969 	else
970 		nouveau_vm_map_sg(vma, 0, mem->num_pages << PAGE_SHIFT, node);
971 
972 	return 0;
973 }
974 
975 static int
976 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
977 		     bool no_wait_gpu, struct ttm_mem_reg *new_mem)
978 {
979 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
980 	struct nouveau_channel *chan = chan = drm->ttm.chan;
981 	struct nouveau_bo *nvbo = nouveau_bo(bo);
982 	struct ttm_mem_reg *old_mem = &bo->mem;
983 	int ret;
984 
985 	mutex_lock_nested(&chan->cli->mutex, SINGLE_DEPTH_NESTING);
986 
987 	/* create temporary vmas for the transfer and attach them to the
988 	 * old nouveau_mem node, these will get cleaned up after ttm has
989 	 * destroyed the ttm_mem_reg
990 	 */
991 	if (nv_device(drm->device)->card_type >= NV_50) {
992 		struct nouveau_mem *node = old_mem->mm_node;
993 
994 		ret = nouveau_vma_getmap(chan, nvbo, old_mem, &node->vma[0]);
995 		if (ret)
996 			goto out;
997 
998 		ret = nouveau_vma_getmap(chan, nvbo, new_mem, &node->vma[1]);
999 		if (ret)
1000 			goto out;
1001 	}
1002 
1003 	ret = drm->ttm.move(chan, bo, &bo->mem, new_mem);
1004 	if (ret == 0) {
1005 		ret = nouveau_bo_move_accel_cleanup(chan, nvbo, evict,
1006 						    no_wait_gpu, new_mem);
1007 	}
1008 
1009 out:
1010 	mutex_unlock(&chan->cli->mutex);
1011 	return ret;
1012 }
1013 
1014 void
1015 nouveau_bo_move_init(struct nouveau_drm *drm)
1016 {
1017 	static const struct {
1018 		const char *name;
1019 		int engine;
1020 		u32 oclass;
1021 		int (*exec)(struct nouveau_channel *,
1022 			    struct ttm_buffer_object *,
1023 			    struct ttm_mem_reg *, struct ttm_mem_reg *);
1024 		int (*init)(struct nouveau_channel *, u32 handle);
1025 	} _methods[] = {
1026 		{  "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
1027 		{  "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1028 		{ "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
1029 		{ "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
1030 		{  "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
1031 		{ "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
1032 		{  "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
1033 		{  "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
1034 		{  "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
1035 		{},
1036 		{ "CRYPT", 0, 0x88b4, nv98_bo_move_exec, nv50_bo_move_init },
1037 	}, *mthd = _methods;
1038 	const char *name = "CPU";
1039 	int ret;
1040 
1041 	do {
1042 		struct nouveau_object *object;
1043 		struct nouveau_channel *chan;
1044 		u32 handle = (mthd->engine << 16) | mthd->oclass;
1045 
1046 		if (mthd->engine)
1047 			chan = drm->cechan;
1048 		else
1049 			chan = drm->channel;
1050 		if (chan == NULL)
1051 			continue;
1052 
1053 		ret = nouveau_object_new(nv_object(drm), chan->handle, handle,
1054 					 mthd->oclass, NULL, 0, &object);
1055 		if (ret == 0) {
1056 			ret = mthd->init(chan, handle);
1057 			if (ret) {
1058 				nouveau_object_del(nv_object(drm),
1059 						   chan->handle, handle);
1060 				continue;
1061 			}
1062 
1063 			drm->ttm.move = mthd->exec;
1064 			drm->ttm.chan = chan;
1065 			name = mthd->name;
1066 			break;
1067 		}
1068 	} while ((++mthd)->exec);
1069 
1070 	NV_INFO(drm, "MM: using %s for buffer copies\n", name);
1071 }
1072 
1073 static int
1074 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
1075 		      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1076 {
1077 	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
1078 	struct ttm_placement placement;
1079 	struct ttm_mem_reg tmp_mem;
1080 	int ret;
1081 
1082 	placement.fpfn = placement.lpfn = 0;
1083 	placement.num_placement = placement.num_busy_placement = 1;
1084 	placement.placement = placement.busy_placement = &placement_memtype;
1085 
1086 	tmp_mem = *new_mem;
1087 	tmp_mem.mm_node = NULL;
1088 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1089 	if (ret)
1090 		return ret;
1091 
1092 	ret = ttm_tt_bind(bo->ttm, &tmp_mem);
1093 	if (ret)
1094 		goto out;
1095 
1096 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, &tmp_mem);
1097 	if (ret)
1098 		goto out;
1099 
1100 	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
1101 out:
1102 	ttm_bo_mem_put(bo, &tmp_mem);
1103 	return ret;
1104 }
1105 
1106 static int
1107 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
1108 		      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1109 {
1110 	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
1111 	struct ttm_placement placement;
1112 	struct ttm_mem_reg tmp_mem;
1113 	int ret;
1114 
1115 	placement.fpfn = placement.lpfn = 0;
1116 	placement.num_placement = placement.num_busy_placement = 1;
1117 	placement.placement = placement.busy_placement = &placement_memtype;
1118 
1119 	tmp_mem = *new_mem;
1120 	tmp_mem.mm_node = NULL;
1121 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1122 	if (ret)
1123 		return ret;
1124 
1125 	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
1126 	if (ret)
1127 		goto out;
1128 
1129 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_mem);
1130 	if (ret)
1131 		goto out;
1132 
1133 out:
1134 	ttm_bo_mem_put(bo, &tmp_mem);
1135 	return ret;
1136 }
1137 
1138 static void
1139 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem)
1140 {
1141 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1142 	struct nouveau_vma *vma;
1143 
1144 	/* ttm can now (stupidly) pass the driver bos it didn't create... */
1145 	if (bo->destroy != nouveau_bo_del_ttm)
1146 		return;
1147 
1148 	list_for_each_entry(vma, &nvbo->vma_list, head) {
1149 		if (new_mem && new_mem->mem_type == TTM_PL_VRAM) {
1150 			nouveau_vm_map(vma, new_mem->mm_node);
1151 		} else
1152 		if (new_mem && new_mem->mem_type == TTM_PL_TT &&
1153 		    nvbo->page_shift == vma->vm->vmm->spg_shift) {
1154 			if (((struct nouveau_mem *)new_mem->mm_node)->sg)
1155 				nouveau_vm_map_sg_table(vma, 0, new_mem->
1156 						  num_pages << PAGE_SHIFT,
1157 						  new_mem->mm_node);
1158 			else
1159 				nouveau_vm_map_sg(vma, 0, new_mem->
1160 						  num_pages << PAGE_SHIFT,
1161 						  new_mem->mm_node);
1162 		} else {
1163 			nouveau_vm_unmap(vma);
1164 		}
1165 	}
1166 }
1167 
1168 static int
1169 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
1170 		   struct nouveau_drm_tile **new_tile)
1171 {
1172 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1173 	struct drm_device *dev = drm->dev;
1174 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1175 	u64 offset = new_mem->start << PAGE_SHIFT;
1176 
1177 	*new_tile = NULL;
1178 	if (new_mem->mem_type != TTM_PL_VRAM)
1179 		return 0;
1180 
1181 	if (nv_device(drm->device)->card_type >= NV_10) {
1182 		*new_tile = nv10_bo_set_tiling(dev, offset, new_mem->size,
1183 						nvbo->tile_mode,
1184 						nvbo->tile_flags);
1185 	}
1186 
1187 	return 0;
1188 }
1189 
1190 static void
1191 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1192 		      struct nouveau_drm_tile *new_tile,
1193 		      struct nouveau_drm_tile **old_tile)
1194 {
1195 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1196 	struct drm_device *dev = drm->dev;
1197 
1198 	nv10_bo_put_tile_region(dev, *old_tile, bo->sync_obj);
1199 	*old_tile = new_tile;
1200 }
1201 
1202 static int
1203 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
1204 		bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1205 {
1206 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1207 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1208 	struct ttm_mem_reg *old_mem = &bo->mem;
1209 	struct nouveau_drm_tile *new_tile = NULL;
1210 	int ret = 0;
1211 
1212 	if (nv_device(drm->device)->card_type < NV_50) {
1213 		ret = nouveau_bo_vm_bind(bo, new_mem, &new_tile);
1214 		if (ret)
1215 			return ret;
1216 	}
1217 
1218 	/* Fake bo copy. */
1219 	if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
1220 		BUG_ON(bo->mem.mm_node != NULL);
1221 		bo->mem = *new_mem;
1222 		new_mem->mm_node = NULL;
1223 		goto out;
1224 	}
1225 
1226 	/* CPU copy if we have no accelerated method available */
1227 	if (!drm->ttm.move) {
1228 		ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
1229 		goto out;
1230 	}
1231 
1232 	/* Hardware assisted copy. */
1233 	if (new_mem->mem_type == TTM_PL_SYSTEM)
1234 		ret = nouveau_bo_move_flipd(bo, evict, intr,
1235 					    no_wait_gpu, new_mem);
1236 	else if (old_mem->mem_type == TTM_PL_SYSTEM)
1237 		ret = nouveau_bo_move_flips(bo, evict, intr,
1238 					    no_wait_gpu, new_mem);
1239 	else
1240 		ret = nouveau_bo_move_m2mf(bo, evict, intr,
1241 					   no_wait_gpu, new_mem);
1242 
1243 	if (!ret)
1244 		goto out;
1245 
1246 	/* Fallback to software copy. */
1247 	ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
1248 
1249 out:
1250 	if (nv_device(drm->device)->card_type < NV_50) {
1251 		if (ret)
1252 			nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1253 		else
1254 			nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1255 	}
1256 
1257 	return ret;
1258 }
1259 
1260 static int
1261 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1262 {
1263 	return 0;
1264 }
1265 
1266 static int
1267 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1268 {
1269 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1270 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1271 	struct drm_device *dev = drm->dev;
1272 	int ret;
1273 
1274 	mem->bus.addr = NULL;
1275 	mem->bus.offset = 0;
1276 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
1277 	mem->bus.base = 0;
1278 	mem->bus.is_iomem = false;
1279 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
1280 		return -EINVAL;
1281 	switch (mem->mem_type) {
1282 	case TTM_PL_SYSTEM:
1283 		/* System memory */
1284 		return 0;
1285 	case TTM_PL_TT:
1286 #if __OS_HAS_AGP
1287 		if (drm->agp.stat == ENABLED) {
1288 			mem->bus.offset = mem->start << PAGE_SHIFT;
1289 			mem->bus.base = drm->agp.base;
1290 			mem->bus.is_iomem = !dev->agp->cant_use_aperture;
1291 		}
1292 #endif
1293 		break;
1294 	case TTM_PL_VRAM:
1295 		mem->bus.offset = mem->start << PAGE_SHIFT;
1296 		mem->bus.base = pci_resource_start(dev->pdev, 1);
1297 		mem->bus.is_iomem = true;
1298 		if (nv_device(drm->device)->card_type >= NV_50) {
1299 			struct nouveau_bar *bar = nouveau_bar(drm->device);
1300 			struct nouveau_mem *node = mem->mm_node;
1301 
1302 			ret = bar->umap(bar, node, NV_MEM_ACCESS_RW,
1303 					&node->bar_vma);
1304 			if (ret)
1305 				return ret;
1306 
1307 			mem->bus.offset = node->bar_vma.offset;
1308 		}
1309 		break;
1310 	default:
1311 		return -EINVAL;
1312 	}
1313 	return 0;
1314 }
1315 
1316 static void
1317 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1318 {
1319 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1320 	struct nouveau_bar *bar = nouveau_bar(drm->device);
1321 	struct nouveau_mem *node = mem->mm_node;
1322 
1323 	if (!node->bar_vma.node)
1324 		return;
1325 
1326 	bar->unmap(bar, &node->bar_vma);
1327 }
1328 
1329 static int
1330 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1331 {
1332 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1333 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1334 	struct nouveau_device *device = nv_device(drm->device);
1335 	u32 mappable = pci_resource_len(device->pdev, 1) >> PAGE_SHIFT;
1336 
1337 	/* as long as the bo isn't in vram, and isn't tiled, we've got
1338 	 * nothing to do here.
1339 	 */
1340 	if (bo->mem.mem_type != TTM_PL_VRAM) {
1341 		if (nv_device(drm->device)->card_type < NV_50 ||
1342 		    !nouveau_bo_tile_layout(nvbo))
1343 			return 0;
1344 	}
1345 
1346 	/* make sure bo is in mappable vram */
1347 	if (bo->mem.start + bo->mem.num_pages < mappable)
1348 		return 0;
1349 
1350 
1351 	nvbo->placement.fpfn = 0;
1352 	nvbo->placement.lpfn = mappable;
1353 	nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_VRAM, 0);
1354 	return nouveau_bo_validate(nvbo, false, false);
1355 }
1356 
1357 static int
1358 nouveau_ttm_tt_populate(struct ttm_tt *ttm)
1359 {
1360 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1361 	struct nouveau_drm *drm;
1362 	struct drm_device *dev;
1363 	unsigned i;
1364 	int r;
1365 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1366 
1367 	if (ttm->state != tt_unpopulated)
1368 		return 0;
1369 
1370 	if (slave && ttm->sg) {
1371 		/* make userspace faulting work */
1372 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1373 						 ttm_dma->dma_address, ttm->num_pages);
1374 		ttm->state = tt_unbound;
1375 		return 0;
1376 	}
1377 
1378 	drm = nouveau_bdev(ttm->bdev);
1379 	dev = drm->dev;
1380 
1381 #if __OS_HAS_AGP
1382 	if (drm->agp.stat == ENABLED) {
1383 		return ttm_agp_tt_populate(ttm);
1384 	}
1385 #endif
1386 
1387 #ifdef CONFIG_SWIOTLB
1388 	if (swiotlb_nr_tbl()) {
1389 		return ttm_dma_populate((void *)ttm, dev->dev);
1390 	}
1391 #endif
1392 
1393 	r = ttm_pool_populate(ttm);
1394 	if (r) {
1395 		return r;
1396 	}
1397 
1398 	for (i = 0; i < ttm->num_pages; i++) {
1399 		ttm_dma->dma_address[i] = pci_map_page(dev->pdev, ttm->pages[i],
1400 						   0, PAGE_SIZE,
1401 						   PCI_DMA_BIDIRECTIONAL);
1402 		if (pci_dma_mapping_error(dev->pdev, ttm_dma->dma_address[i])) {
1403 			while (--i) {
1404 				pci_unmap_page(dev->pdev, ttm_dma->dma_address[i],
1405 					       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
1406 				ttm_dma->dma_address[i] = 0;
1407 			}
1408 			ttm_pool_unpopulate(ttm);
1409 			return -EFAULT;
1410 		}
1411 	}
1412 	return 0;
1413 }
1414 
1415 static void
1416 nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
1417 {
1418 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1419 	struct nouveau_drm *drm;
1420 	struct drm_device *dev;
1421 	unsigned i;
1422 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1423 
1424 	if (slave)
1425 		return;
1426 
1427 	drm = nouveau_bdev(ttm->bdev);
1428 	dev = drm->dev;
1429 
1430 #if __OS_HAS_AGP
1431 	if (drm->agp.stat == ENABLED) {
1432 		ttm_agp_tt_unpopulate(ttm);
1433 		return;
1434 	}
1435 #endif
1436 
1437 #ifdef CONFIG_SWIOTLB
1438 	if (swiotlb_nr_tbl()) {
1439 		ttm_dma_unpopulate((void *)ttm, dev->dev);
1440 		return;
1441 	}
1442 #endif
1443 
1444 	for (i = 0; i < ttm->num_pages; i++) {
1445 		if (ttm_dma->dma_address[i]) {
1446 			pci_unmap_page(dev->pdev, ttm_dma->dma_address[i],
1447 				       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
1448 		}
1449 	}
1450 
1451 	ttm_pool_unpopulate(ttm);
1452 }
1453 
1454 void
1455 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
1456 {
1457 	struct nouveau_fence *old_fence = NULL;
1458 
1459 	if (likely(fence))
1460 		nouveau_fence_ref(fence);
1461 
1462 	spin_lock(&nvbo->bo.bdev->fence_lock);
1463 	old_fence = nvbo->bo.sync_obj;
1464 	nvbo->bo.sync_obj = fence;
1465 	spin_unlock(&nvbo->bo.bdev->fence_lock);
1466 
1467 	nouveau_fence_unref(&old_fence);
1468 }
1469 
1470 static void
1471 nouveau_bo_fence_unref(void **sync_obj)
1472 {
1473 	nouveau_fence_unref((struct nouveau_fence **)sync_obj);
1474 }
1475 
1476 static void *
1477 nouveau_bo_fence_ref(void *sync_obj)
1478 {
1479 	return nouveau_fence_ref(sync_obj);
1480 }
1481 
1482 static bool
1483 nouveau_bo_fence_signalled(void *sync_obj)
1484 {
1485 	return nouveau_fence_done(sync_obj);
1486 }
1487 
1488 static int
1489 nouveau_bo_fence_wait(void *sync_obj, bool lazy, bool intr)
1490 {
1491 	return nouveau_fence_wait(sync_obj, lazy, intr);
1492 }
1493 
1494 static int
1495 nouveau_bo_fence_flush(void *sync_obj)
1496 {
1497 	return 0;
1498 }
1499 
1500 struct ttm_bo_driver nouveau_bo_driver = {
1501 	.ttm_tt_create = &nouveau_ttm_tt_create,
1502 	.ttm_tt_populate = &nouveau_ttm_tt_populate,
1503 	.ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1504 	.invalidate_caches = nouveau_bo_invalidate_caches,
1505 	.init_mem_type = nouveau_bo_init_mem_type,
1506 	.evict_flags = nouveau_bo_evict_flags,
1507 	.move_notify = nouveau_bo_move_ntfy,
1508 	.move = nouveau_bo_move,
1509 	.verify_access = nouveau_bo_verify_access,
1510 	.sync_obj_signaled = nouveau_bo_fence_signalled,
1511 	.sync_obj_wait = nouveau_bo_fence_wait,
1512 	.sync_obj_flush = nouveau_bo_fence_flush,
1513 	.sync_obj_unref = nouveau_bo_fence_unref,
1514 	.sync_obj_ref = nouveau_bo_fence_ref,
1515 	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1516 	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1517 	.io_mem_free = &nouveau_ttm_io_mem_free,
1518 };
1519 
1520 struct nouveau_vma *
1521 nouveau_bo_vma_find(struct nouveau_bo *nvbo, struct nouveau_vm *vm)
1522 {
1523 	struct nouveau_vma *vma;
1524 	list_for_each_entry(vma, &nvbo->vma_list, head) {
1525 		if (vma->vm == vm)
1526 			return vma;
1527 	}
1528 
1529 	return NULL;
1530 }
1531 
1532 int
1533 nouveau_bo_vma_add(struct nouveau_bo *nvbo, struct nouveau_vm *vm,
1534 		   struct nouveau_vma *vma)
1535 {
1536 	const u32 size = nvbo->bo.mem.num_pages << PAGE_SHIFT;
1537 	struct nouveau_mem *node = nvbo->bo.mem.mm_node;
1538 	int ret;
1539 
1540 	ret = nouveau_vm_get(vm, size, nvbo->page_shift,
1541 			     NV_MEM_ACCESS_RW, vma);
1542 	if (ret)
1543 		return ret;
1544 
1545 	if (nvbo->bo.mem.mem_type == TTM_PL_VRAM)
1546 		nouveau_vm_map(vma, nvbo->bo.mem.mm_node);
1547 	else if (nvbo->bo.mem.mem_type == TTM_PL_TT) {
1548 		if (node->sg)
1549 			nouveau_vm_map_sg_table(vma, 0, size, node);
1550 		else
1551 			nouveau_vm_map_sg(vma, 0, size, node);
1552 	}
1553 
1554 	list_add_tail(&vma->head, &nvbo->vma_list);
1555 	vma->refcount = 1;
1556 	return 0;
1557 }
1558 
1559 void
1560 nouveau_bo_vma_del(struct nouveau_bo *nvbo, struct nouveau_vma *vma)
1561 {
1562 	if (vma->node) {
1563 		if (nvbo->bo.mem.mem_type != TTM_PL_SYSTEM)
1564 			nouveau_vm_unmap(vma);
1565 		nouveau_vm_put(vma);
1566 		list_del(&vma->head);
1567 	}
1568 }
1569