xref: /linux/drivers/gpu/drm/nouveau/nouveau_bo.c (revision 5ba0a3be6ecc3a0b0d52c2a818b05564c6b42510)
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 
32 #include <subdev/fb.h>
33 #include <subdev/vm.h>
34 #include <subdev/bar.h>
35 
36 #include "nouveau_drm.h"
37 #include "nouveau_dma.h"
38 #include "nouveau_fence.h"
39 
40 #include "nouveau_bo.h"
41 #include "nouveau_ttm.h"
42 #include "nouveau_gem.h"
43 
44 /*
45  * NV10-NV40 tiling helpers
46  */
47 
48 static void
49 nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
50 			   u32 addr, u32 size, u32 pitch, u32 flags)
51 {
52 	struct nouveau_drm *drm = nouveau_drm(dev);
53 	int i = reg - drm->tile.reg;
54 	struct nouveau_fb *pfb = nouveau_fb(drm->device);
55 	struct nouveau_fb_tile *tile = &pfb->tile.region[i];
56 	struct nouveau_engine *engine;
57 
58 	nouveau_fence_unref(&reg->fence);
59 
60 	if (tile->pitch)
61 		pfb->tile.fini(pfb, i, tile);
62 
63 	if (pitch)
64 		pfb->tile.init(pfb, i, addr, size, pitch, flags, tile);
65 
66 	pfb->tile.prog(pfb, i, tile);
67 
68 	if ((engine = nouveau_engine(pfb, NVDEV_ENGINE_GR)))
69 		engine->tile_prog(engine, i);
70 	if ((engine = nouveau_engine(pfb, NVDEV_ENGINE_MPEG)))
71 		engine->tile_prog(engine, i);
72 }
73 
74 static struct nouveau_drm_tile *
75 nv10_bo_get_tile_region(struct drm_device *dev, int i)
76 {
77 	struct nouveau_drm *drm = nouveau_drm(dev);
78 	struct nouveau_drm_tile *tile = &drm->tile.reg[i];
79 
80 	spin_lock(&drm->tile.lock);
81 
82 	if (!tile->used &&
83 	    (!tile->fence || nouveau_fence_done(tile->fence)))
84 		tile->used = true;
85 	else
86 		tile = NULL;
87 
88 	spin_unlock(&drm->tile.lock);
89 	return tile;
90 }
91 
92 static void
93 nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
94 			struct nouveau_fence *fence)
95 {
96 	struct nouveau_drm *drm = nouveau_drm(dev);
97 
98 	if (tile) {
99 		spin_lock(&drm->tile.lock);
100 		if (fence) {
101 			/* Mark it as pending. */
102 			tile->fence = fence;
103 			nouveau_fence_ref(fence);
104 		}
105 
106 		tile->used = false;
107 		spin_unlock(&drm->tile.lock);
108 	}
109 }
110 
111 static struct nouveau_drm_tile *
112 nv10_bo_set_tiling(struct drm_device *dev, u32 addr,
113 		   u32 size, u32 pitch, u32 flags)
114 {
115 	struct nouveau_drm *drm = nouveau_drm(dev);
116 	struct nouveau_fb *pfb = nouveau_fb(drm->device);
117 	struct nouveau_drm_tile *tile, *found = NULL;
118 	int i;
119 
120 	for (i = 0; i < pfb->tile.regions; i++) {
121 		tile = nv10_bo_get_tile_region(dev, i);
122 
123 		if (pitch && !found) {
124 			found = tile;
125 			continue;
126 
127 		} else if (tile && pfb->tile.region[i].pitch) {
128 			/* Kill an unused tile region. */
129 			nv10_bo_update_tile_region(dev, tile, 0, 0, 0, 0);
130 		}
131 
132 		nv10_bo_put_tile_region(dev, tile, NULL);
133 	}
134 
135 	if (found)
136 		nv10_bo_update_tile_region(dev, found, addr, size,
137 					    pitch, flags);
138 	return found;
139 }
140 
141 static void
142 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
143 {
144 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
145 	struct drm_device *dev = drm->dev;
146 	struct nouveau_bo *nvbo = nouveau_bo(bo);
147 
148 	if (unlikely(nvbo->gem))
149 		DRM_ERROR("bo %p still attached to GEM object\n", bo);
150 	nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
151 	kfree(nvbo);
152 }
153 
154 static void
155 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
156 		       int *align, int *size)
157 {
158 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
159 	struct nouveau_device *device = nv_device(drm->device);
160 
161 	if (device->card_type < NV_50) {
162 		if (nvbo->tile_mode) {
163 			if (device->chipset >= 0x40) {
164 				*align = 65536;
165 				*size = roundup(*size, 64 * nvbo->tile_mode);
166 
167 			} else if (device->chipset >= 0x30) {
168 				*align = 32768;
169 				*size = roundup(*size, 64 * nvbo->tile_mode);
170 
171 			} else if (device->chipset >= 0x20) {
172 				*align = 16384;
173 				*size = roundup(*size, 64 * nvbo->tile_mode);
174 
175 			} else if (device->chipset >= 0x10) {
176 				*align = 16384;
177 				*size = roundup(*size, 32 * nvbo->tile_mode);
178 			}
179 		}
180 	} else {
181 		*size = roundup(*size, (1 << nvbo->page_shift));
182 		*align = max((1 <<  nvbo->page_shift), *align);
183 	}
184 
185 	*size = roundup(*size, PAGE_SIZE);
186 }
187 
188 int
189 nouveau_bo_new(struct drm_device *dev, int size, int align,
190 	       uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
191 	       struct sg_table *sg,
192 	       struct nouveau_bo **pnvbo)
193 {
194 	struct nouveau_drm *drm = nouveau_drm(dev);
195 	struct nouveau_bo *nvbo;
196 	size_t acc_size;
197 	int ret;
198 	int type = ttm_bo_type_device;
199 
200 	if (sg)
201 		type = ttm_bo_type_sg;
202 
203 	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
204 	if (!nvbo)
205 		return -ENOMEM;
206 	INIT_LIST_HEAD(&nvbo->head);
207 	INIT_LIST_HEAD(&nvbo->entry);
208 	INIT_LIST_HEAD(&nvbo->vma_list);
209 	nvbo->tile_mode = tile_mode;
210 	nvbo->tile_flags = tile_flags;
211 	nvbo->bo.bdev = &drm->ttm.bdev;
212 
213 	nvbo->page_shift = 12;
214 	if (drm->client.base.vm) {
215 		if (!(flags & TTM_PL_FLAG_TT) && size > 256 * 1024)
216 			nvbo->page_shift = drm->client.base.vm->vmm->lpg_shift;
217 	}
218 
219 	nouveau_bo_fixup_align(nvbo, flags, &align, &size);
220 	nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
221 	nouveau_bo_placement_set(nvbo, flags, 0);
222 
223 	acc_size = ttm_bo_dma_acc_size(&drm->ttm.bdev, size,
224 				       sizeof(struct nouveau_bo));
225 
226 	ret = ttm_bo_init(&drm->ttm.bdev, &nvbo->bo, size,
227 			  type, &nvbo->placement,
228 			  align >> PAGE_SHIFT, false, NULL, acc_size, sg,
229 			  nouveau_bo_del_ttm);
230 	if (ret) {
231 		/* ttm will call nouveau_bo_del_ttm if it fails.. */
232 		return ret;
233 	}
234 
235 	*pnvbo = nvbo;
236 	return 0;
237 }
238 
239 static void
240 set_placement_list(uint32_t *pl, unsigned *n, uint32_t type, uint32_t flags)
241 {
242 	*n = 0;
243 
244 	if (type & TTM_PL_FLAG_VRAM)
245 		pl[(*n)++] = TTM_PL_FLAG_VRAM | flags;
246 	if (type & TTM_PL_FLAG_TT)
247 		pl[(*n)++] = TTM_PL_FLAG_TT | flags;
248 	if (type & TTM_PL_FLAG_SYSTEM)
249 		pl[(*n)++] = TTM_PL_FLAG_SYSTEM | flags;
250 }
251 
252 static void
253 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
254 {
255 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
256 	struct nouveau_fb *pfb = nouveau_fb(drm->device);
257 	u32 vram_pages = pfb->ram.size >> PAGE_SHIFT;
258 
259 	if (nv_device(drm->device)->card_type == NV_10 &&
260 	    nvbo->tile_mode && (type & TTM_PL_FLAG_VRAM) &&
261 	    nvbo->bo.mem.num_pages < vram_pages / 4) {
262 		/*
263 		 * Make sure that the color and depth buffers are handled
264 		 * by independent memory controller units. Up to a 9x
265 		 * speed up when alpha-blending and depth-test are enabled
266 		 * at the same time.
267 		 */
268 		if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
269 			nvbo->placement.fpfn = vram_pages / 2;
270 			nvbo->placement.lpfn = ~0;
271 		} else {
272 			nvbo->placement.fpfn = 0;
273 			nvbo->placement.lpfn = vram_pages / 2;
274 		}
275 	}
276 }
277 
278 void
279 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
280 {
281 	struct ttm_placement *pl = &nvbo->placement;
282 	uint32_t flags = TTM_PL_MASK_CACHING |
283 		(nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
284 
285 	pl->placement = nvbo->placements;
286 	set_placement_list(nvbo->placements, &pl->num_placement,
287 			   type, flags);
288 
289 	pl->busy_placement = nvbo->busy_placements;
290 	set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
291 			   type | busy, flags);
292 
293 	set_placement_range(nvbo, type);
294 }
295 
296 int
297 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype)
298 {
299 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
300 	struct ttm_buffer_object *bo = &nvbo->bo;
301 	int ret;
302 
303 	if (nvbo->pin_refcnt && !(memtype & (1 << bo->mem.mem_type))) {
304 		NV_ERROR(drm, "bo %p pinned elsewhere: 0x%08x vs 0x%08x\n", bo,
305 			 1 << bo->mem.mem_type, memtype);
306 		return -EINVAL;
307 	}
308 
309 	if (nvbo->pin_refcnt++)
310 		return 0;
311 
312 	ret = ttm_bo_reserve(bo, false, false, false, 0);
313 	if (ret)
314 		goto out;
315 
316 	nouveau_bo_placement_set(nvbo, memtype, 0);
317 
318 	ret = nouveau_bo_validate(nvbo, false, false);
319 	if (ret == 0) {
320 		switch (bo->mem.mem_type) {
321 		case TTM_PL_VRAM:
322 			drm->gem.vram_available -= bo->mem.size;
323 			break;
324 		case TTM_PL_TT:
325 			drm->gem.gart_available -= bo->mem.size;
326 			break;
327 		default:
328 			break;
329 		}
330 	}
331 	ttm_bo_unreserve(bo);
332 out:
333 	if (unlikely(ret))
334 		nvbo->pin_refcnt--;
335 	return ret;
336 }
337 
338 int
339 nouveau_bo_unpin(struct nouveau_bo *nvbo)
340 {
341 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
342 	struct ttm_buffer_object *bo = &nvbo->bo;
343 	int ret;
344 
345 	if (--nvbo->pin_refcnt)
346 		return 0;
347 
348 	ret = ttm_bo_reserve(bo, false, false, false, 0);
349 	if (ret)
350 		return ret;
351 
352 	nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
353 
354 	ret = nouveau_bo_validate(nvbo, false, false);
355 	if (ret == 0) {
356 		switch (bo->mem.mem_type) {
357 		case TTM_PL_VRAM:
358 			drm->gem.vram_available += bo->mem.size;
359 			break;
360 		case TTM_PL_TT:
361 			drm->gem.gart_available += bo->mem.size;
362 			break;
363 		default:
364 			break;
365 		}
366 	}
367 
368 	ttm_bo_unreserve(bo);
369 	return ret;
370 }
371 
372 int
373 nouveau_bo_map(struct nouveau_bo *nvbo)
374 {
375 	int ret;
376 
377 	ret = ttm_bo_reserve(&nvbo->bo, false, false, false, 0);
378 	if (ret)
379 		return ret;
380 
381 	ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
382 	ttm_bo_unreserve(&nvbo->bo);
383 	return ret;
384 }
385 
386 void
387 nouveau_bo_unmap(struct nouveau_bo *nvbo)
388 {
389 	if (nvbo)
390 		ttm_bo_kunmap(&nvbo->kmap);
391 }
392 
393 int
394 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
395 		    bool no_wait_gpu)
396 {
397 	int ret;
398 
399 	ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement,
400 			      interruptible, no_wait_gpu);
401 	if (ret)
402 		return ret;
403 
404 	return 0;
405 }
406 
407 u16
408 nouveau_bo_rd16(struct nouveau_bo *nvbo, unsigned index)
409 {
410 	bool is_iomem;
411 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
412 	mem = &mem[index];
413 	if (is_iomem)
414 		return ioread16_native((void __force __iomem *)mem);
415 	else
416 		return *mem;
417 }
418 
419 void
420 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
421 {
422 	bool is_iomem;
423 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
424 	mem = &mem[index];
425 	if (is_iomem)
426 		iowrite16_native(val, (void __force __iomem *)mem);
427 	else
428 		*mem = val;
429 }
430 
431 u32
432 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
433 {
434 	bool is_iomem;
435 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
436 	mem = &mem[index];
437 	if (is_iomem)
438 		return ioread32_native((void __force __iomem *)mem);
439 	else
440 		return *mem;
441 }
442 
443 void
444 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
445 {
446 	bool is_iomem;
447 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
448 	mem = &mem[index];
449 	if (is_iomem)
450 		iowrite32_native(val, (void __force __iomem *)mem);
451 	else
452 		*mem = val;
453 }
454 
455 static struct ttm_tt *
456 nouveau_ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
457 		      uint32_t page_flags, struct page *dummy_read)
458 {
459 #if __OS_HAS_AGP
460 	struct nouveau_drm *drm = nouveau_bdev(bdev);
461 	struct drm_device *dev = drm->dev;
462 
463 	if (drm->agp.stat == ENABLED) {
464 		return ttm_agp_tt_create(bdev, dev->agp->bridge, size,
465 					 page_flags, dummy_read);
466 	}
467 #endif
468 
469 	return nouveau_sgdma_create_ttm(bdev, size, page_flags, dummy_read);
470 }
471 
472 static int
473 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
474 {
475 	/* We'll do this from user space. */
476 	return 0;
477 }
478 
479 static int
480 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
481 			 struct ttm_mem_type_manager *man)
482 {
483 	struct nouveau_drm *drm = nouveau_bdev(bdev);
484 
485 	switch (type) {
486 	case TTM_PL_SYSTEM:
487 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
488 		man->available_caching = TTM_PL_MASK_CACHING;
489 		man->default_caching = TTM_PL_FLAG_CACHED;
490 		break;
491 	case TTM_PL_VRAM:
492 		if (nv_device(drm->device)->card_type >= NV_50) {
493 			man->func = &nouveau_vram_manager;
494 			man->io_reserve_fastpath = false;
495 			man->use_io_reserve_lru = true;
496 		} else {
497 			man->func = &ttm_bo_manager_func;
498 		}
499 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
500 			     TTM_MEMTYPE_FLAG_MAPPABLE;
501 		man->available_caching = TTM_PL_FLAG_UNCACHED |
502 					 TTM_PL_FLAG_WC;
503 		man->default_caching = TTM_PL_FLAG_WC;
504 		break;
505 	case TTM_PL_TT:
506 		if (nv_device(drm->device)->card_type >= NV_50)
507 			man->func = &nouveau_gart_manager;
508 		else
509 		if (drm->agp.stat != ENABLED)
510 			man->func = &nv04_gart_manager;
511 		else
512 			man->func = &ttm_bo_manager_func;
513 
514 		if (drm->agp.stat == ENABLED) {
515 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
516 			man->available_caching = TTM_PL_FLAG_UNCACHED |
517 				TTM_PL_FLAG_WC;
518 			man->default_caching = TTM_PL_FLAG_WC;
519 		} else {
520 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
521 				     TTM_MEMTYPE_FLAG_CMA;
522 			man->available_caching = TTM_PL_MASK_CACHING;
523 			man->default_caching = TTM_PL_FLAG_CACHED;
524 		}
525 
526 		break;
527 	default:
528 		return -EINVAL;
529 	}
530 	return 0;
531 }
532 
533 static void
534 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
535 {
536 	struct nouveau_bo *nvbo = nouveau_bo(bo);
537 
538 	switch (bo->mem.mem_type) {
539 	case TTM_PL_VRAM:
540 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
541 					 TTM_PL_FLAG_SYSTEM);
542 		break;
543 	default:
544 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
545 		break;
546 	}
547 
548 	*pl = nvbo->placement;
549 }
550 
551 
552 /* GPU-assisted copy using NV_MEMORY_TO_MEMORY_FORMAT, can access
553  * TTM_PL_{VRAM,TT} directly.
554  */
555 
556 static int
557 nouveau_bo_move_accel_cleanup(struct nouveau_channel *chan,
558 			      struct nouveau_bo *nvbo, bool evict,
559 			      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
560 {
561 	struct nouveau_fence *fence = NULL;
562 	int ret;
563 
564 	ret = nouveau_fence_new(chan, &fence);
565 	if (ret)
566 		return ret;
567 
568 	ret = ttm_bo_move_accel_cleanup(&nvbo->bo, fence, evict,
569 					no_wait_gpu, new_mem);
570 	nouveau_fence_unref(&fence);
571 	return ret;
572 }
573 
574 static int
575 nve0_bo_move_init(struct nouveau_channel *chan, u32 handle)
576 {
577 	int ret = RING_SPACE(chan, 2);
578 	if (ret == 0) {
579 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
580 		OUT_RING  (chan, handle);
581 		FIRE_RING (chan);
582 	}
583 	return ret;
584 }
585 
586 static int
587 nve0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
588 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
589 {
590 	struct nouveau_mem *node = old_mem->mm_node;
591 	int ret = RING_SPACE(chan, 10);
592 	if (ret == 0) {
593 		BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
594 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
595 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
596 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
597 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
598 		OUT_RING  (chan, PAGE_SIZE);
599 		OUT_RING  (chan, PAGE_SIZE);
600 		OUT_RING  (chan, PAGE_SIZE);
601 		OUT_RING  (chan, new_mem->num_pages);
602 		BEGIN_IMC0(chan, NvSubCopy, 0x0300, 0x0386);
603 	}
604 	return ret;
605 }
606 
607 static int
608 nvc0_bo_move_init(struct nouveau_channel *chan, u32 handle)
609 {
610 	int ret = RING_SPACE(chan, 2);
611 	if (ret == 0) {
612 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
613 		OUT_RING  (chan, handle);
614 	}
615 	return ret;
616 }
617 
618 static int
619 nvc0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
620 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
621 {
622 	struct nouveau_mem *node = old_mem->mm_node;
623 	u64 src_offset = node->vma[0].offset;
624 	u64 dst_offset = node->vma[1].offset;
625 	u32 page_count = new_mem->num_pages;
626 	int ret;
627 
628 	page_count = new_mem->num_pages;
629 	while (page_count) {
630 		int line_count = (page_count > 8191) ? 8191 : page_count;
631 
632 		ret = RING_SPACE(chan, 11);
633 		if (ret)
634 			return ret;
635 
636 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 8);
637 		OUT_RING  (chan, upper_32_bits(src_offset));
638 		OUT_RING  (chan, lower_32_bits(src_offset));
639 		OUT_RING  (chan, upper_32_bits(dst_offset));
640 		OUT_RING  (chan, lower_32_bits(dst_offset));
641 		OUT_RING  (chan, PAGE_SIZE);
642 		OUT_RING  (chan, PAGE_SIZE);
643 		OUT_RING  (chan, PAGE_SIZE);
644 		OUT_RING  (chan, line_count);
645 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
646 		OUT_RING  (chan, 0x00000110);
647 
648 		page_count -= line_count;
649 		src_offset += (PAGE_SIZE * line_count);
650 		dst_offset += (PAGE_SIZE * line_count);
651 	}
652 
653 	return 0;
654 }
655 
656 static int
657 nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
658 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
659 {
660 	struct nouveau_mem *node = old_mem->mm_node;
661 	u64 src_offset = node->vma[0].offset;
662 	u64 dst_offset = node->vma[1].offset;
663 	u32 page_count = new_mem->num_pages;
664 	int ret;
665 
666 	page_count = new_mem->num_pages;
667 	while (page_count) {
668 		int line_count = (page_count > 2047) ? 2047 : page_count;
669 
670 		ret = RING_SPACE(chan, 12);
671 		if (ret)
672 			return ret;
673 
674 		BEGIN_NVC0(chan, NvSubCopy, 0x0238, 2);
675 		OUT_RING  (chan, upper_32_bits(dst_offset));
676 		OUT_RING  (chan, lower_32_bits(dst_offset));
677 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 6);
678 		OUT_RING  (chan, upper_32_bits(src_offset));
679 		OUT_RING  (chan, lower_32_bits(src_offset));
680 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
681 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
682 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
683 		OUT_RING  (chan, line_count);
684 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
685 		OUT_RING  (chan, 0x00100110);
686 
687 		page_count -= line_count;
688 		src_offset += (PAGE_SIZE * line_count);
689 		dst_offset += (PAGE_SIZE * line_count);
690 	}
691 
692 	return 0;
693 }
694 
695 static int
696 nva3_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
697 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
698 {
699 	struct nouveau_mem *node = old_mem->mm_node;
700 	u64 src_offset = node->vma[0].offset;
701 	u64 dst_offset = node->vma[1].offset;
702 	u32 page_count = new_mem->num_pages;
703 	int ret;
704 
705 	page_count = new_mem->num_pages;
706 	while (page_count) {
707 		int line_count = (page_count > 8191) ? 8191 : page_count;
708 
709 		ret = RING_SPACE(chan, 11);
710 		if (ret)
711 			return ret;
712 
713 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
714 		OUT_RING  (chan, upper_32_bits(src_offset));
715 		OUT_RING  (chan, lower_32_bits(src_offset));
716 		OUT_RING  (chan, upper_32_bits(dst_offset));
717 		OUT_RING  (chan, lower_32_bits(dst_offset));
718 		OUT_RING  (chan, PAGE_SIZE);
719 		OUT_RING  (chan, PAGE_SIZE);
720 		OUT_RING  (chan, PAGE_SIZE);
721 		OUT_RING  (chan, line_count);
722 		BEGIN_NV04(chan, NvSubCopy, 0x0300, 1);
723 		OUT_RING  (chan, 0x00000110);
724 
725 		page_count -= line_count;
726 		src_offset += (PAGE_SIZE * line_count);
727 		dst_offset += (PAGE_SIZE * line_count);
728 	}
729 
730 	return 0;
731 }
732 
733 static int
734 nv98_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
735 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
736 {
737 	struct nouveau_mem *node = old_mem->mm_node;
738 	int ret = RING_SPACE(chan, 7);
739 	if (ret == 0) {
740 		BEGIN_NV04(chan, NvSubCopy, 0x0320, 6);
741 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
742 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
743 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
744 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
745 		OUT_RING  (chan, 0x00000000 /* COPY */);
746 		OUT_RING  (chan, new_mem->num_pages << PAGE_SHIFT);
747 	}
748 	return ret;
749 }
750 
751 static int
752 nv84_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
753 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
754 {
755 	struct nouveau_mem *node = old_mem->mm_node;
756 	int ret = RING_SPACE(chan, 7);
757 	if (ret == 0) {
758 		BEGIN_NV04(chan, NvSubCopy, 0x0304, 6);
759 		OUT_RING  (chan, new_mem->num_pages << PAGE_SHIFT);
760 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
761 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
762 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
763 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
764 		OUT_RING  (chan, 0x00000000 /* MODE_COPY, QUERY_NONE */);
765 	}
766 	return ret;
767 }
768 
769 static int
770 nv50_bo_move_init(struct nouveau_channel *chan, u32 handle)
771 {
772 	int ret = RING_SPACE(chan, 6);
773 	if (ret == 0) {
774 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
775 		OUT_RING  (chan, handle);
776 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 3);
777 		OUT_RING  (chan, NvNotify0);
778 		OUT_RING  (chan, NvDmaFB);
779 		OUT_RING  (chan, NvDmaFB);
780 	}
781 
782 	return ret;
783 }
784 
785 static int
786 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
787 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
788 {
789 	struct nouveau_mem *node = old_mem->mm_node;
790 	struct nouveau_bo *nvbo = nouveau_bo(bo);
791 	u64 length = (new_mem->num_pages << PAGE_SHIFT);
792 	u64 src_offset = node->vma[0].offset;
793 	u64 dst_offset = node->vma[1].offset;
794 	int ret;
795 
796 	while (length) {
797 		u32 amount, stride, height;
798 
799 		amount  = min(length, (u64)(4 * 1024 * 1024));
800 		stride  = 16 * 4;
801 		height  = amount / stride;
802 
803 		if (new_mem->mem_type == TTM_PL_VRAM &&
804 		    nouveau_bo_tile_layout(nvbo)) {
805 			ret = RING_SPACE(chan, 8);
806 			if (ret)
807 				return ret;
808 
809 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 7);
810 			OUT_RING  (chan, 0);
811 			OUT_RING  (chan, 0);
812 			OUT_RING  (chan, stride);
813 			OUT_RING  (chan, height);
814 			OUT_RING  (chan, 1);
815 			OUT_RING  (chan, 0);
816 			OUT_RING  (chan, 0);
817 		} else {
818 			ret = RING_SPACE(chan, 2);
819 			if (ret)
820 				return ret;
821 
822 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 1);
823 			OUT_RING  (chan, 1);
824 		}
825 		if (old_mem->mem_type == TTM_PL_VRAM &&
826 		    nouveau_bo_tile_layout(nvbo)) {
827 			ret = RING_SPACE(chan, 8);
828 			if (ret)
829 				return ret;
830 
831 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 7);
832 			OUT_RING  (chan, 0);
833 			OUT_RING  (chan, 0);
834 			OUT_RING  (chan, stride);
835 			OUT_RING  (chan, height);
836 			OUT_RING  (chan, 1);
837 			OUT_RING  (chan, 0);
838 			OUT_RING  (chan, 0);
839 		} else {
840 			ret = RING_SPACE(chan, 2);
841 			if (ret)
842 				return ret;
843 
844 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 1);
845 			OUT_RING  (chan, 1);
846 		}
847 
848 		ret = RING_SPACE(chan, 14);
849 		if (ret)
850 			return ret;
851 
852 		BEGIN_NV04(chan, NvSubCopy, 0x0238, 2);
853 		OUT_RING  (chan, upper_32_bits(src_offset));
854 		OUT_RING  (chan, upper_32_bits(dst_offset));
855 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
856 		OUT_RING  (chan, lower_32_bits(src_offset));
857 		OUT_RING  (chan, lower_32_bits(dst_offset));
858 		OUT_RING  (chan, stride);
859 		OUT_RING  (chan, stride);
860 		OUT_RING  (chan, stride);
861 		OUT_RING  (chan, height);
862 		OUT_RING  (chan, 0x00000101);
863 		OUT_RING  (chan, 0x00000000);
864 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
865 		OUT_RING  (chan, 0);
866 
867 		length -= amount;
868 		src_offset += amount;
869 		dst_offset += amount;
870 	}
871 
872 	return 0;
873 }
874 
875 static int
876 nv04_bo_move_init(struct nouveau_channel *chan, u32 handle)
877 {
878 	int ret = RING_SPACE(chan, 4);
879 	if (ret == 0) {
880 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
881 		OUT_RING  (chan, handle);
882 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 1);
883 		OUT_RING  (chan, NvNotify0);
884 	}
885 
886 	return ret;
887 }
888 
889 static inline uint32_t
890 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
891 		      struct nouveau_channel *chan, struct ttm_mem_reg *mem)
892 {
893 	if (mem->mem_type == TTM_PL_TT)
894 		return NvDmaTT;
895 	return NvDmaFB;
896 }
897 
898 static int
899 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
900 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
901 {
902 	u32 src_offset = old_mem->start << PAGE_SHIFT;
903 	u32 dst_offset = new_mem->start << PAGE_SHIFT;
904 	u32 page_count = new_mem->num_pages;
905 	int ret;
906 
907 	ret = RING_SPACE(chan, 3);
908 	if (ret)
909 		return ret;
910 
911 	BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
912 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
913 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
914 
915 	page_count = new_mem->num_pages;
916 	while (page_count) {
917 		int line_count = (page_count > 2047) ? 2047 : page_count;
918 
919 		ret = RING_SPACE(chan, 11);
920 		if (ret)
921 			return ret;
922 
923 		BEGIN_NV04(chan, NvSubCopy,
924 				 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
925 		OUT_RING  (chan, src_offset);
926 		OUT_RING  (chan, dst_offset);
927 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
928 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
929 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
930 		OUT_RING  (chan, line_count);
931 		OUT_RING  (chan, 0x00000101);
932 		OUT_RING  (chan, 0x00000000);
933 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
934 		OUT_RING  (chan, 0);
935 
936 		page_count -= line_count;
937 		src_offset += (PAGE_SIZE * line_count);
938 		dst_offset += (PAGE_SIZE * line_count);
939 	}
940 
941 	return 0;
942 }
943 
944 static int
945 nouveau_vma_getmap(struct nouveau_channel *chan, struct nouveau_bo *nvbo,
946 		   struct ttm_mem_reg *mem, struct nouveau_vma *vma)
947 {
948 	struct nouveau_mem *node = mem->mm_node;
949 	int ret;
950 
951 	ret = nouveau_vm_get(nv_client(chan->cli)->vm, mem->num_pages <<
952 			     PAGE_SHIFT, node->page_shift,
953 			     NV_MEM_ACCESS_RW, vma);
954 	if (ret)
955 		return ret;
956 
957 	if (mem->mem_type == TTM_PL_VRAM)
958 		nouveau_vm_map(vma, node);
959 	else
960 		nouveau_vm_map_sg(vma, 0, mem->num_pages << PAGE_SHIFT, node);
961 
962 	return 0;
963 }
964 
965 static int
966 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
967 		     bool no_wait_gpu, struct ttm_mem_reg *new_mem)
968 {
969 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
970 	struct nouveau_channel *chan = chan = drm->channel;
971 	struct nouveau_bo *nvbo = nouveau_bo(bo);
972 	struct ttm_mem_reg *old_mem = &bo->mem;
973 	int ret;
974 
975 	mutex_lock(&chan->cli->mutex);
976 
977 	/* create temporary vmas for the transfer and attach them to the
978 	 * old nouveau_mem node, these will get cleaned up after ttm has
979 	 * destroyed the ttm_mem_reg
980 	 */
981 	if (nv_device(drm->device)->card_type >= NV_50) {
982 		struct nouveau_mem *node = old_mem->mm_node;
983 
984 		ret = nouveau_vma_getmap(chan, nvbo, old_mem, &node->vma[0]);
985 		if (ret)
986 			goto out;
987 
988 		ret = nouveau_vma_getmap(chan, nvbo, new_mem, &node->vma[1]);
989 		if (ret)
990 			goto out;
991 	}
992 
993 	ret = drm->ttm.move(chan, bo, &bo->mem, new_mem);
994 	if (ret == 0) {
995 		ret = nouveau_bo_move_accel_cleanup(chan, nvbo, evict,
996 						    no_wait_gpu, new_mem);
997 	}
998 
999 out:
1000 	mutex_unlock(&chan->cli->mutex);
1001 	return ret;
1002 }
1003 
1004 void
1005 nouveau_bo_move_init(struct nouveau_drm *drm)
1006 {
1007 	static const struct {
1008 		const char *name;
1009 		int engine;
1010 		u32 oclass;
1011 		int (*exec)(struct nouveau_channel *,
1012 			    struct ttm_buffer_object *,
1013 			    struct ttm_mem_reg *, struct ttm_mem_reg *);
1014 		int (*init)(struct nouveau_channel *, u32 handle);
1015 	} _methods[] = {
1016 		{  "COPY", 0, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
1017 		{  "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1018 		{ "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
1019 		{ "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
1020 		{  "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
1021 		{ "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
1022 		{  "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
1023 		{  "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
1024 		{  "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
1025 		{},
1026 		{ "CRYPT", 0, 0x88b4, nv98_bo_move_exec, nv50_bo_move_init },
1027 	}, *mthd = _methods;
1028 	const char *name = "CPU";
1029 	int ret;
1030 
1031 	do {
1032 		struct nouveau_object *object;
1033 		struct nouveau_channel *chan;
1034 		u32 handle = (mthd->engine << 16) | mthd->oclass;
1035 
1036 		if (mthd->init == nve0_bo_move_init)
1037 			chan = drm->cechan;
1038 		else
1039 			chan = drm->channel;
1040 		if (chan == NULL)
1041 			continue;
1042 
1043 		ret = nouveau_object_new(nv_object(drm), chan->handle, handle,
1044 					 mthd->oclass, NULL, 0, &object);
1045 		if (ret == 0) {
1046 			ret = mthd->init(chan, handle);
1047 			if (ret) {
1048 				nouveau_object_del(nv_object(drm),
1049 						   chan->handle, handle);
1050 				continue;
1051 			}
1052 
1053 			drm->ttm.move = mthd->exec;
1054 			name = mthd->name;
1055 			break;
1056 		}
1057 	} while ((++mthd)->exec);
1058 
1059 	NV_INFO(drm, "MM: using %s for buffer copies\n", name);
1060 }
1061 
1062 static int
1063 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
1064 		      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1065 {
1066 	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
1067 	struct ttm_placement placement;
1068 	struct ttm_mem_reg tmp_mem;
1069 	int ret;
1070 
1071 	placement.fpfn = placement.lpfn = 0;
1072 	placement.num_placement = placement.num_busy_placement = 1;
1073 	placement.placement = placement.busy_placement = &placement_memtype;
1074 
1075 	tmp_mem = *new_mem;
1076 	tmp_mem.mm_node = NULL;
1077 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1078 	if (ret)
1079 		return ret;
1080 
1081 	ret = ttm_tt_bind(bo->ttm, &tmp_mem);
1082 	if (ret)
1083 		goto out;
1084 
1085 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, &tmp_mem);
1086 	if (ret)
1087 		goto out;
1088 
1089 	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
1090 out:
1091 	ttm_bo_mem_put(bo, &tmp_mem);
1092 	return ret;
1093 }
1094 
1095 static int
1096 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
1097 		      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1098 {
1099 	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
1100 	struct ttm_placement placement;
1101 	struct ttm_mem_reg tmp_mem;
1102 	int ret;
1103 
1104 	placement.fpfn = placement.lpfn = 0;
1105 	placement.num_placement = placement.num_busy_placement = 1;
1106 	placement.placement = placement.busy_placement = &placement_memtype;
1107 
1108 	tmp_mem = *new_mem;
1109 	tmp_mem.mm_node = NULL;
1110 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1111 	if (ret)
1112 		return ret;
1113 
1114 	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
1115 	if (ret)
1116 		goto out;
1117 
1118 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_mem);
1119 	if (ret)
1120 		goto out;
1121 
1122 out:
1123 	ttm_bo_mem_put(bo, &tmp_mem);
1124 	return ret;
1125 }
1126 
1127 static void
1128 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem)
1129 {
1130 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1131 	struct nouveau_vma *vma;
1132 
1133 	/* ttm can now (stupidly) pass the driver bos it didn't create... */
1134 	if (bo->destroy != nouveau_bo_del_ttm)
1135 		return;
1136 
1137 	list_for_each_entry(vma, &nvbo->vma_list, head) {
1138 		if (new_mem && new_mem->mem_type == TTM_PL_VRAM) {
1139 			nouveau_vm_map(vma, new_mem->mm_node);
1140 		} else
1141 		if (new_mem && new_mem->mem_type == TTM_PL_TT &&
1142 		    nvbo->page_shift == vma->vm->vmm->spg_shift) {
1143 			if (((struct nouveau_mem *)new_mem->mm_node)->sg)
1144 				nouveau_vm_map_sg_table(vma, 0, new_mem->
1145 						  num_pages << PAGE_SHIFT,
1146 						  new_mem->mm_node);
1147 			else
1148 				nouveau_vm_map_sg(vma, 0, new_mem->
1149 						  num_pages << PAGE_SHIFT,
1150 						  new_mem->mm_node);
1151 		} else {
1152 			nouveau_vm_unmap(vma);
1153 		}
1154 	}
1155 }
1156 
1157 static int
1158 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
1159 		   struct nouveau_drm_tile **new_tile)
1160 {
1161 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1162 	struct drm_device *dev = drm->dev;
1163 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1164 	u64 offset = new_mem->start << PAGE_SHIFT;
1165 
1166 	*new_tile = NULL;
1167 	if (new_mem->mem_type != TTM_PL_VRAM)
1168 		return 0;
1169 
1170 	if (nv_device(drm->device)->card_type >= NV_10) {
1171 		*new_tile = nv10_bo_set_tiling(dev, offset, new_mem->size,
1172 						nvbo->tile_mode,
1173 						nvbo->tile_flags);
1174 	}
1175 
1176 	return 0;
1177 }
1178 
1179 static void
1180 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1181 		      struct nouveau_drm_tile *new_tile,
1182 		      struct nouveau_drm_tile **old_tile)
1183 {
1184 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1185 	struct drm_device *dev = drm->dev;
1186 
1187 	nv10_bo_put_tile_region(dev, *old_tile, bo->sync_obj);
1188 	*old_tile = new_tile;
1189 }
1190 
1191 static int
1192 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
1193 		bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1194 {
1195 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1196 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1197 	struct ttm_mem_reg *old_mem = &bo->mem;
1198 	struct nouveau_drm_tile *new_tile = NULL;
1199 	int ret = 0;
1200 
1201 	if (nv_device(drm->device)->card_type < NV_50) {
1202 		ret = nouveau_bo_vm_bind(bo, new_mem, &new_tile);
1203 		if (ret)
1204 			return ret;
1205 	}
1206 
1207 	/* Fake bo copy. */
1208 	if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
1209 		BUG_ON(bo->mem.mm_node != NULL);
1210 		bo->mem = *new_mem;
1211 		new_mem->mm_node = NULL;
1212 		goto out;
1213 	}
1214 
1215 	/* CPU copy if we have no accelerated method available */
1216 	if (!drm->ttm.move) {
1217 		ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
1218 		goto out;
1219 	}
1220 
1221 	/* Hardware assisted copy. */
1222 	if (new_mem->mem_type == TTM_PL_SYSTEM)
1223 		ret = nouveau_bo_move_flipd(bo, evict, intr,
1224 					    no_wait_gpu, new_mem);
1225 	else if (old_mem->mem_type == TTM_PL_SYSTEM)
1226 		ret = nouveau_bo_move_flips(bo, evict, intr,
1227 					    no_wait_gpu, new_mem);
1228 	else
1229 		ret = nouveau_bo_move_m2mf(bo, evict, intr,
1230 					   no_wait_gpu, new_mem);
1231 
1232 	if (!ret)
1233 		goto out;
1234 
1235 	/* Fallback to software copy. */
1236 	ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
1237 
1238 out:
1239 	if (nv_device(drm->device)->card_type < NV_50) {
1240 		if (ret)
1241 			nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1242 		else
1243 			nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1244 	}
1245 
1246 	return ret;
1247 }
1248 
1249 static int
1250 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1251 {
1252 	return 0;
1253 }
1254 
1255 static int
1256 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1257 {
1258 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1259 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1260 	struct drm_device *dev = drm->dev;
1261 	int ret;
1262 
1263 	mem->bus.addr = NULL;
1264 	mem->bus.offset = 0;
1265 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
1266 	mem->bus.base = 0;
1267 	mem->bus.is_iomem = false;
1268 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
1269 		return -EINVAL;
1270 	switch (mem->mem_type) {
1271 	case TTM_PL_SYSTEM:
1272 		/* System memory */
1273 		return 0;
1274 	case TTM_PL_TT:
1275 #if __OS_HAS_AGP
1276 		if (drm->agp.stat == ENABLED) {
1277 			mem->bus.offset = mem->start << PAGE_SHIFT;
1278 			mem->bus.base = drm->agp.base;
1279 			mem->bus.is_iomem = !dev->agp->cant_use_aperture;
1280 		}
1281 #endif
1282 		break;
1283 	case TTM_PL_VRAM:
1284 		mem->bus.offset = mem->start << PAGE_SHIFT;
1285 		mem->bus.base = pci_resource_start(dev->pdev, 1);
1286 		mem->bus.is_iomem = true;
1287 		if (nv_device(drm->device)->card_type >= NV_50) {
1288 			struct nouveau_bar *bar = nouveau_bar(drm->device);
1289 			struct nouveau_mem *node = mem->mm_node;
1290 
1291 			ret = bar->umap(bar, node, NV_MEM_ACCESS_RW,
1292 					&node->bar_vma);
1293 			if (ret)
1294 				return ret;
1295 
1296 			mem->bus.offset = node->bar_vma.offset;
1297 		}
1298 		break;
1299 	default:
1300 		return -EINVAL;
1301 	}
1302 	return 0;
1303 }
1304 
1305 static void
1306 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1307 {
1308 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1309 	struct nouveau_bar *bar = nouveau_bar(drm->device);
1310 	struct nouveau_mem *node = mem->mm_node;
1311 
1312 	if (!node->bar_vma.node)
1313 		return;
1314 
1315 	bar->unmap(bar, &node->bar_vma);
1316 }
1317 
1318 static int
1319 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1320 {
1321 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1322 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1323 	struct nouveau_device *device = nv_device(drm->device);
1324 	u32 mappable = pci_resource_len(device->pdev, 1) >> PAGE_SHIFT;
1325 
1326 	/* as long as the bo isn't in vram, and isn't tiled, we've got
1327 	 * nothing to do here.
1328 	 */
1329 	if (bo->mem.mem_type != TTM_PL_VRAM) {
1330 		if (nv_device(drm->device)->card_type < NV_50 ||
1331 		    !nouveau_bo_tile_layout(nvbo))
1332 			return 0;
1333 	}
1334 
1335 	/* make sure bo is in mappable vram */
1336 	if (bo->mem.start + bo->mem.num_pages < mappable)
1337 		return 0;
1338 
1339 
1340 	nvbo->placement.fpfn = 0;
1341 	nvbo->placement.lpfn = mappable;
1342 	nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_VRAM, 0);
1343 	return nouveau_bo_validate(nvbo, false, false);
1344 }
1345 
1346 static int
1347 nouveau_ttm_tt_populate(struct ttm_tt *ttm)
1348 {
1349 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1350 	struct nouveau_drm *drm;
1351 	struct drm_device *dev;
1352 	unsigned i;
1353 	int r;
1354 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1355 
1356 	if (ttm->state != tt_unpopulated)
1357 		return 0;
1358 
1359 	if (slave && ttm->sg) {
1360 		/* make userspace faulting work */
1361 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1362 						 ttm_dma->dma_address, ttm->num_pages);
1363 		ttm->state = tt_unbound;
1364 		return 0;
1365 	}
1366 
1367 	drm = nouveau_bdev(ttm->bdev);
1368 	dev = drm->dev;
1369 
1370 #if __OS_HAS_AGP
1371 	if (drm->agp.stat == ENABLED) {
1372 		return ttm_agp_tt_populate(ttm);
1373 	}
1374 #endif
1375 
1376 #ifdef CONFIG_SWIOTLB
1377 	if (swiotlb_nr_tbl()) {
1378 		return ttm_dma_populate((void *)ttm, dev->dev);
1379 	}
1380 #endif
1381 
1382 	r = ttm_pool_populate(ttm);
1383 	if (r) {
1384 		return r;
1385 	}
1386 
1387 	for (i = 0; i < ttm->num_pages; i++) {
1388 		ttm_dma->dma_address[i] = pci_map_page(dev->pdev, ttm->pages[i],
1389 						   0, PAGE_SIZE,
1390 						   PCI_DMA_BIDIRECTIONAL);
1391 		if (pci_dma_mapping_error(dev->pdev, ttm_dma->dma_address[i])) {
1392 			while (--i) {
1393 				pci_unmap_page(dev->pdev, ttm_dma->dma_address[i],
1394 					       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
1395 				ttm_dma->dma_address[i] = 0;
1396 			}
1397 			ttm_pool_unpopulate(ttm);
1398 			return -EFAULT;
1399 		}
1400 	}
1401 	return 0;
1402 }
1403 
1404 static void
1405 nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
1406 {
1407 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1408 	struct nouveau_drm *drm;
1409 	struct drm_device *dev;
1410 	unsigned i;
1411 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1412 
1413 	if (slave)
1414 		return;
1415 
1416 	drm = nouveau_bdev(ttm->bdev);
1417 	dev = drm->dev;
1418 
1419 #if __OS_HAS_AGP
1420 	if (drm->agp.stat == ENABLED) {
1421 		ttm_agp_tt_unpopulate(ttm);
1422 		return;
1423 	}
1424 #endif
1425 
1426 #ifdef CONFIG_SWIOTLB
1427 	if (swiotlb_nr_tbl()) {
1428 		ttm_dma_unpopulate((void *)ttm, dev->dev);
1429 		return;
1430 	}
1431 #endif
1432 
1433 	for (i = 0; i < ttm->num_pages; i++) {
1434 		if (ttm_dma->dma_address[i]) {
1435 			pci_unmap_page(dev->pdev, ttm_dma->dma_address[i],
1436 				       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
1437 		}
1438 	}
1439 
1440 	ttm_pool_unpopulate(ttm);
1441 }
1442 
1443 void
1444 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
1445 {
1446 	struct nouveau_fence *old_fence = NULL;
1447 
1448 	if (likely(fence))
1449 		nouveau_fence_ref(fence);
1450 
1451 	spin_lock(&nvbo->bo.bdev->fence_lock);
1452 	old_fence = nvbo->bo.sync_obj;
1453 	nvbo->bo.sync_obj = fence;
1454 	spin_unlock(&nvbo->bo.bdev->fence_lock);
1455 
1456 	nouveau_fence_unref(&old_fence);
1457 }
1458 
1459 static void
1460 nouveau_bo_fence_unref(void **sync_obj)
1461 {
1462 	nouveau_fence_unref((struct nouveau_fence **)sync_obj);
1463 }
1464 
1465 static void *
1466 nouveau_bo_fence_ref(void *sync_obj)
1467 {
1468 	return nouveau_fence_ref(sync_obj);
1469 }
1470 
1471 static bool
1472 nouveau_bo_fence_signalled(void *sync_obj)
1473 {
1474 	return nouveau_fence_done(sync_obj);
1475 }
1476 
1477 static int
1478 nouveau_bo_fence_wait(void *sync_obj, bool lazy, bool intr)
1479 {
1480 	return nouveau_fence_wait(sync_obj, lazy, intr);
1481 }
1482 
1483 static int
1484 nouveau_bo_fence_flush(void *sync_obj)
1485 {
1486 	return 0;
1487 }
1488 
1489 struct ttm_bo_driver nouveau_bo_driver = {
1490 	.ttm_tt_create = &nouveau_ttm_tt_create,
1491 	.ttm_tt_populate = &nouveau_ttm_tt_populate,
1492 	.ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1493 	.invalidate_caches = nouveau_bo_invalidate_caches,
1494 	.init_mem_type = nouveau_bo_init_mem_type,
1495 	.evict_flags = nouveau_bo_evict_flags,
1496 	.move_notify = nouveau_bo_move_ntfy,
1497 	.move = nouveau_bo_move,
1498 	.verify_access = nouveau_bo_verify_access,
1499 	.sync_obj_signaled = nouveau_bo_fence_signalled,
1500 	.sync_obj_wait = nouveau_bo_fence_wait,
1501 	.sync_obj_flush = nouveau_bo_fence_flush,
1502 	.sync_obj_unref = nouveau_bo_fence_unref,
1503 	.sync_obj_ref = nouveau_bo_fence_ref,
1504 	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1505 	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1506 	.io_mem_free = &nouveau_ttm_io_mem_free,
1507 };
1508 
1509 struct nouveau_vma *
1510 nouveau_bo_vma_find(struct nouveau_bo *nvbo, struct nouveau_vm *vm)
1511 {
1512 	struct nouveau_vma *vma;
1513 	list_for_each_entry(vma, &nvbo->vma_list, head) {
1514 		if (vma->vm == vm)
1515 			return vma;
1516 	}
1517 
1518 	return NULL;
1519 }
1520 
1521 int
1522 nouveau_bo_vma_add(struct nouveau_bo *nvbo, struct nouveau_vm *vm,
1523 		   struct nouveau_vma *vma)
1524 {
1525 	const u32 size = nvbo->bo.mem.num_pages << PAGE_SHIFT;
1526 	struct nouveau_mem *node = nvbo->bo.mem.mm_node;
1527 	int ret;
1528 
1529 	ret = nouveau_vm_get(vm, size, nvbo->page_shift,
1530 			     NV_MEM_ACCESS_RW, vma);
1531 	if (ret)
1532 		return ret;
1533 
1534 	if (nvbo->bo.mem.mem_type == TTM_PL_VRAM)
1535 		nouveau_vm_map(vma, nvbo->bo.mem.mm_node);
1536 	else if (nvbo->bo.mem.mem_type == TTM_PL_TT) {
1537 		if (node->sg)
1538 			nouveau_vm_map_sg_table(vma, 0, size, node);
1539 		else
1540 			nouveau_vm_map_sg(vma, 0, size, node);
1541 	}
1542 
1543 	list_add_tail(&vma->head, &nvbo->vma_list);
1544 	vma->refcount = 1;
1545 	return 0;
1546 }
1547 
1548 void
1549 nouveau_bo_vma_del(struct nouveau_bo *nvbo, struct nouveau_vma *vma)
1550 {
1551 	if (vma->node) {
1552 		if (nvbo->bo.mem.mem_type != TTM_PL_SYSTEM) {
1553 			spin_lock(&nvbo->bo.bdev->fence_lock);
1554 			ttm_bo_wait(&nvbo->bo, false, false, false);
1555 			spin_unlock(&nvbo->bo.bdev->fence_lock);
1556 			nouveau_vm_unmap(vma);
1557 		}
1558 
1559 		nouveau_vm_put(vma);
1560 		list_del(&vma->head);
1561 	}
1562 }
1563