1 /*
2  * Copyright 2007 Dave Airlied
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * VA LINUX SYSTEMS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  */
24 /*
25  * Authors: Dave Airlied <airlied@linux.ie>
26  *	    Ben Skeggs   <darktama@iinet.net.au>
27  *	    Jeremy Kolb  <jkolb@brandeis.edu>
28  */
29 
30 #include <linux/dma-mapping.h>
31 #include <linux/swiotlb.h>
32 
33 #include "nouveau_drm.h"
34 #include "nouveau_dma.h"
35 #include "nouveau_fence.h"
36 
37 #include "nouveau_bo.h"
38 #include "nouveau_ttm.h"
39 #include "nouveau_gem.h"
40 
41 /*
42  * NV10-NV40 tiling helpers
43  */
44 
45 static void
46 nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
47 			   u32 addr, u32 size, u32 pitch, u32 flags)
48 {
49 	struct nouveau_drm *drm = nouveau_drm(dev);
50 	int i = reg - drm->tile.reg;
51 	struct nouveau_fb *pfb = nvkm_fb(&drm->device);
52 	struct nouveau_fb_tile *tile = &pfb->tile.region[i];
53 	struct nouveau_engine *engine;
54 
55 	nouveau_fence_unref(&reg->fence);
56 
57 	if (tile->pitch)
58 		pfb->tile.fini(pfb, i, tile);
59 
60 	if (pitch)
61 		pfb->tile.init(pfb, i, addr, size, pitch, flags, tile);
62 
63 	pfb->tile.prog(pfb, i, tile);
64 
65 	if ((engine = nouveau_engine(pfb, NVDEV_ENGINE_GR)))
66 		engine->tile_prog(engine, i);
67 	if ((engine = nouveau_engine(pfb, NVDEV_ENGINE_MPEG)))
68 		engine->tile_prog(engine, i);
69 }
70 
71 static struct nouveau_drm_tile *
72 nv10_bo_get_tile_region(struct drm_device *dev, int i)
73 {
74 	struct nouveau_drm *drm = nouveau_drm(dev);
75 	struct nouveau_drm_tile *tile = &drm->tile.reg[i];
76 
77 	spin_lock(&drm->tile.lock);
78 
79 	if (!tile->used &&
80 	    (!tile->fence || nouveau_fence_done(tile->fence)))
81 		tile->used = true;
82 	else
83 		tile = NULL;
84 
85 	spin_unlock(&drm->tile.lock);
86 	return tile;
87 }
88 
89 static void
90 nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
91 			struct nouveau_fence *fence)
92 {
93 	struct nouveau_drm *drm = nouveau_drm(dev);
94 
95 	if (tile) {
96 		spin_lock(&drm->tile.lock);
97 		tile->fence = nouveau_fence_ref(fence);
98 		tile->used = false;
99 		spin_unlock(&drm->tile.lock);
100 	}
101 }
102 
103 static struct nouveau_drm_tile *
104 nv10_bo_set_tiling(struct drm_device *dev, u32 addr,
105 		   u32 size, u32 pitch, u32 flags)
106 {
107 	struct nouveau_drm *drm = nouveau_drm(dev);
108 	struct nouveau_fb *pfb = nvkm_fb(&drm->device);
109 	struct nouveau_drm_tile *tile, *found = NULL;
110 	int i;
111 
112 	for (i = 0; i < pfb->tile.regions; i++) {
113 		tile = nv10_bo_get_tile_region(dev, i);
114 
115 		if (pitch && !found) {
116 			found = tile;
117 			continue;
118 
119 		} else if (tile && pfb->tile.region[i].pitch) {
120 			/* Kill an unused tile region. */
121 			nv10_bo_update_tile_region(dev, tile, 0, 0, 0, 0);
122 		}
123 
124 		nv10_bo_put_tile_region(dev, tile, NULL);
125 	}
126 
127 	if (found)
128 		nv10_bo_update_tile_region(dev, found, addr, size,
129 					    pitch, flags);
130 	return found;
131 }
132 
133 static void
134 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
135 {
136 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
137 	struct drm_device *dev = drm->dev;
138 	struct nouveau_bo *nvbo = nouveau_bo(bo);
139 
140 	if (unlikely(nvbo->gem.filp))
141 		DRM_ERROR("bo %p still attached to GEM object\n", bo);
142 	WARN_ON(nvbo->pin_refcnt > 0);
143 	nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
144 	kfree(nvbo);
145 }
146 
147 static void
148 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
149 		       int *align, int *size)
150 {
151 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
152 	struct nvif_device *device = &drm->device;
153 
154 	if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
155 		if (nvbo->tile_mode) {
156 			if (device->info.chipset >= 0x40) {
157 				*align = 65536;
158 				*size = roundup(*size, 64 * nvbo->tile_mode);
159 
160 			} else if (device->info.chipset >= 0x30) {
161 				*align = 32768;
162 				*size = roundup(*size, 64 * nvbo->tile_mode);
163 
164 			} else if (device->info.chipset >= 0x20) {
165 				*align = 16384;
166 				*size = roundup(*size, 64 * nvbo->tile_mode);
167 
168 			} else if (device->info.chipset >= 0x10) {
169 				*align = 16384;
170 				*size = roundup(*size, 32 * nvbo->tile_mode);
171 			}
172 		}
173 	} else {
174 		*size = roundup(*size, (1 << nvbo->page_shift));
175 		*align = max((1 <<  nvbo->page_shift), *align);
176 	}
177 
178 	*size = roundup(*size, PAGE_SIZE);
179 }
180 
181 int
182 nouveau_bo_new(struct drm_device *dev, int size, int align,
183 	       uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
184 	       struct sg_table *sg,
185 	       struct nouveau_bo **pnvbo)
186 {
187 	struct nouveau_drm *drm = nouveau_drm(dev);
188 	struct nouveau_bo *nvbo;
189 	size_t acc_size;
190 	int ret;
191 	int type = ttm_bo_type_device;
192 	int lpg_shift = 12;
193 	int max_size;
194 
195 	if (drm->client.vm)
196 		lpg_shift = drm->client.vm->vmm->lpg_shift;
197 	max_size = INT_MAX & ~((1 << lpg_shift) - 1);
198 
199 	if (size <= 0 || size > max_size) {
200 		NV_WARN(drm, "skipped size %x\n", (u32)size);
201 		return -EINVAL;
202 	}
203 
204 	if (sg)
205 		type = ttm_bo_type_sg;
206 
207 	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
208 	if (!nvbo)
209 		return -ENOMEM;
210 	INIT_LIST_HEAD(&nvbo->head);
211 	INIT_LIST_HEAD(&nvbo->entry);
212 	INIT_LIST_HEAD(&nvbo->vma_list);
213 	nvbo->tile_mode = tile_mode;
214 	nvbo->tile_flags = tile_flags;
215 	nvbo->bo.bdev = &drm->ttm.bdev;
216 
217 	nvbo->page_shift = 12;
218 	if (drm->client.vm) {
219 		if (!(flags & TTM_PL_FLAG_TT) && size > 256 * 1024)
220 			nvbo->page_shift = drm->client.vm->vmm->lpg_shift;
221 	}
222 
223 	nouveau_bo_fixup_align(nvbo, flags, &align, &size);
224 	nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
225 	nouveau_bo_placement_set(nvbo, flags, 0);
226 
227 	acc_size = ttm_bo_dma_acc_size(&drm->ttm.bdev, size,
228 				       sizeof(struct nouveau_bo));
229 
230 	ret = ttm_bo_init(&drm->ttm.bdev, &nvbo->bo, size,
231 			  type, &nvbo->placement,
232 			  align >> PAGE_SHIFT, false, NULL, acc_size, sg,
233 			  nouveau_bo_del_ttm);
234 	if (ret) {
235 		/* ttm will call nouveau_bo_del_ttm if it fails.. */
236 		return ret;
237 	}
238 
239 	*pnvbo = nvbo;
240 	return 0;
241 }
242 
243 static void
244 set_placement_list(uint32_t *pl, unsigned *n, uint32_t type, uint32_t flags)
245 {
246 	*n = 0;
247 
248 	if (type & TTM_PL_FLAG_VRAM)
249 		pl[(*n)++] = TTM_PL_FLAG_VRAM | flags;
250 	if (type & TTM_PL_FLAG_TT)
251 		pl[(*n)++] = TTM_PL_FLAG_TT | flags;
252 	if (type & TTM_PL_FLAG_SYSTEM)
253 		pl[(*n)++] = TTM_PL_FLAG_SYSTEM | flags;
254 }
255 
256 static void
257 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
258 {
259 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
260 	u32 vram_pages = drm->device.info.ram_size >> PAGE_SHIFT;
261 
262 	if (drm->device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
263 	    nvbo->tile_mode && (type & TTM_PL_FLAG_VRAM) &&
264 	    nvbo->bo.mem.num_pages < vram_pages / 4) {
265 		/*
266 		 * Make sure that the color and depth buffers are handled
267 		 * by independent memory controller units. Up to a 9x
268 		 * speed up when alpha-blending and depth-test are enabled
269 		 * at the same time.
270 		 */
271 		if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
272 			nvbo->placement.fpfn = vram_pages / 2;
273 			nvbo->placement.lpfn = ~0;
274 		} else {
275 			nvbo->placement.fpfn = 0;
276 			nvbo->placement.lpfn = vram_pages / 2;
277 		}
278 	}
279 }
280 
281 void
282 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
283 {
284 	struct ttm_placement *pl = &nvbo->placement;
285 	uint32_t flags = TTM_PL_MASK_CACHING |
286 		(nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
287 
288 	pl->placement = nvbo->placements;
289 	set_placement_list(nvbo->placements, &pl->num_placement,
290 			   type, flags);
291 
292 	pl->busy_placement = nvbo->busy_placements;
293 	set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
294 			   type | busy, flags);
295 
296 	set_placement_range(nvbo, type);
297 }
298 
299 int
300 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype)
301 {
302 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
303 	struct ttm_buffer_object *bo = &nvbo->bo;
304 	int ret;
305 
306 	ret = ttm_bo_reserve(bo, false, false, false, NULL);
307 	if (ret)
308 		goto out;
309 
310 	if (nvbo->pin_refcnt && !(memtype & (1 << bo->mem.mem_type))) {
311 		NV_ERROR(drm, "bo %p pinned elsewhere: 0x%08x vs 0x%08x\n", bo,
312 			 1 << bo->mem.mem_type, memtype);
313 		ret = -EINVAL;
314 		goto out;
315 	}
316 
317 	if (nvbo->pin_refcnt++)
318 		goto out;
319 
320 	nouveau_bo_placement_set(nvbo, memtype, 0);
321 
322 	ret = nouveau_bo_validate(nvbo, false, false);
323 	if (ret == 0) {
324 		switch (bo->mem.mem_type) {
325 		case TTM_PL_VRAM:
326 			drm->gem.vram_available -= bo->mem.size;
327 			break;
328 		case TTM_PL_TT:
329 			drm->gem.gart_available -= bo->mem.size;
330 			break;
331 		default:
332 			break;
333 		}
334 	}
335 out:
336 	ttm_bo_unreserve(bo);
337 	return ret;
338 }
339 
340 int
341 nouveau_bo_unpin(struct nouveau_bo *nvbo)
342 {
343 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
344 	struct ttm_buffer_object *bo = &nvbo->bo;
345 	int ret, ref;
346 
347 	ret = ttm_bo_reserve(bo, false, false, false, NULL);
348 	if (ret)
349 		return ret;
350 
351 	ref = --nvbo->pin_refcnt;
352 	WARN_ON_ONCE(ref < 0);
353 	if (ref)
354 		goto out;
355 
356 	nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
357 
358 	ret = nouveau_bo_validate(nvbo, false, false);
359 	if (ret == 0) {
360 		switch (bo->mem.mem_type) {
361 		case TTM_PL_VRAM:
362 			drm->gem.vram_available += bo->mem.size;
363 			break;
364 		case TTM_PL_TT:
365 			drm->gem.gart_available += bo->mem.size;
366 			break;
367 		default:
368 			break;
369 		}
370 	}
371 
372 out:
373 	ttm_bo_unreserve(bo);
374 	return ret;
375 }
376 
377 int
378 nouveau_bo_map(struct nouveau_bo *nvbo)
379 {
380 	int ret;
381 
382 	ret = ttm_bo_reserve(&nvbo->bo, false, false, false, NULL);
383 	if (ret)
384 		return ret;
385 
386 	ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
387 	ttm_bo_unreserve(&nvbo->bo);
388 	return ret;
389 }
390 
391 void
392 nouveau_bo_unmap(struct nouveau_bo *nvbo)
393 {
394 	if (nvbo)
395 		ttm_bo_kunmap(&nvbo->kmap);
396 }
397 
398 int
399 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
400 		    bool no_wait_gpu)
401 {
402 	int ret;
403 
404 	ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement,
405 			      interruptible, no_wait_gpu);
406 	if (ret)
407 		return ret;
408 
409 	return 0;
410 }
411 
412 u16
413 nouveau_bo_rd16(struct nouveau_bo *nvbo, unsigned index)
414 {
415 	bool is_iomem;
416 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
417 	mem = &mem[index];
418 	if (is_iomem)
419 		return ioread16_native((void __force __iomem *)mem);
420 	else
421 		return *mem;
422 }
423 
424 void
425 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
426 {
427 	bool is_iomem;
428 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
429 	mem = &mem[index];
430 	if (is_iomem)
431 		iowrite16_native(val, (void __force __iomem *)mem);
432 	else
433 		*mem = val;
434 }
435 
436 u32
437 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
438 {
439 	bool is_iomem;
440 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
441 	mem = &mem[index];
442 	if (is_iomem)
443 		return ioread32_native((void __force __iomem *)mem);
444 	else
445 		return *mem;
446 }
447 
448 void
449 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
450 {
451 	bool is_iomem;
452 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
453 	mem = &mem[index];
454 	if (is_iomem)
455 		iowrite32_native(val, (void __force __iomem *)mem);
456 	else
457 		*mem = val;
458 }
459 
460 static struct ttm_tt *
461 nouveau_ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
462 		      uint32_t page_flags, struct page *dummy_read)
463 {
464 #if __OS_HAS_AGP
465 	struct nouveau_drm *drm = nouveau_bdev(bdev);
466 	struct drm_device *dev = drm->dev;
467 
468 	if (drm->agp.stat == ENABLED) {
469 		return ttm_agp_tt_create(bdev, dev->agp->bridge, size,
470 					 page_flags, dummy_read);
471 	}
472 #endif
473 
474 	return nouveau_sgdma_create_ttm(bdev, size, page_flags, dummy_read);
475 }
476 
477 static int
478 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
479 {
480 	/* We'll do this from user space. */
481 	return 0;
482 }
483 
484 static int
485 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
486 			 struct ttm_mem_type_manager *man)
487 {
488 	struct nouveau_drm *drm = nouveau_bdev(bdev);
489 
490 	switch (type) {
491 	case TTM_PL_SYSTEM:
492 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
493 		man->available_caching = TTM_PL_MASK_CACHING;
494 		man->default_caching = TTM_PL_FLAG_CACHED;
495 		break;
496 	case TTM_PL_VRAM:
497 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
498 			     TTM_MEMTYPE_FLAG_MAPPABLE;
499 		man->available_caching = TTM_PL_FLAG_UNCACHED |
500 					 TTM_PL_FLAG_WC;
501 		man->default_caching = TTM_PL_FLAG_WC;
502 
503 		if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
504 			/* Some BARs do not support being ioremapped WC */
505 			if (nvkm_bar(&drm->device)->iomap_uncached) {
506 				man->available_caching = TTM_PL_FLAG_UNCACHED;
507 				man->default_caching = TTM_PL_FLAG_UNCACHED;
508 			}
509 
510 			man->func = &nouveau_vram_manager;
511 			man->io_reserve_fastpath = false;
512 			man->use_io_reserve_lru = true;
513 		} else {
514 			man->func = &ttm_bo_manager_func;
515 		}
516 		break;
517 	case TTM_PL_TT:
518 		if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA)
519 			man->func = &nouveau_gart_manager;
520 		else
521 		if (drm->agp.stat != ENABLED)
522 			man->func = &nv04_gart_manager;
523 		else
524 			man->func = &ttm_bo_manager_func;
525 
526 		if (drm->agp.stat == ENABLED) {
527 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
528 			man->available_caching = TTM_PL_FLAG_UNCACHED |
529 				TTM_PL_FLAG_WC;
530 			man->default_caching = TTM_PL_FLAG_WC;
531 		} else {
532 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
533 				     TTM_MEMTYPE_FLAG_CMA;
534 			man->available_caching = TTM_PL_MASK_CACHING;
535 			man->default_caching = TTM_PL_FLAG_CACHED;
536 		}
537 
538 		break;
539 	default:
540 		return -EINVAL;
541 	}
542 	return 0;
543 }
544 
545 static void
546 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
547 {
548 	struct nouveau_bo *nvbo = nouveau_bo(bo);
549 
550 	switch (bo->mem.mem_type) {
551 	case TTM_PL_VRAM:
552 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
553 					 TTM_PL_FLAG_SYSTEM);
554 		break;
555 	default:
556 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
557 		break;
558 	}
559 
560 	*pl = nvbo->placement;
561 }
562 
563 
564 static int
565 nve0_bo_move_init(struct nouveau_channel *chan, u32 handle)
566 {
567 	int ret = RING_SPACE(chan, 2);
568 	if (ret == 0) {
569 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
570 		OUT_RING  (chan, handle & 0x0000ffff);
571 		FIRE_RING (chan);
572 	}
573 	return ret;
574 }
575 
576 static int
577 nve0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
578 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
579 {
580 	struct nouveau_mem *node = old_mem->mm_node;
581 	int ret = RING_SPACE(chan, 10);
582 	if (ret == 0) {
583 		BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
584 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
585 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
586 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
587 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
588 		OUT_RING  (chan, PAGE_SIZE);
589 		OUT_RING  (chan, PAGE_SIZE);
590 		OUT_RING  (chan, PAGE_SIZE);
591 		OUT_RING  (chan, new_mem->num_pages);
592 		BEGIN_IMC0(chan, NvSubCopy, 0x0300, 0x0386);
593 	}
594 	return ret;
595 }
596 
597 static int
598 nvc0_bo_move_init(struct nouveau_channel *chan, u32 handle)
599 {
600 	int ret = RING_SPACE(chan, 2);
601 	if (ret == 0) {
602 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
603 		OUT_RING  (chan, handle);
604 	}
605 	return ret;
606 }
607 
608 static int
609 nvc0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
610 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
611 {
612 	struct nouveau_mem *node = old_mem->mm_node;
613 	u64 src_offset = node->vma[0].offset;
614 	u64 dst_offset = node->vma[1].offset;
615 	u32 page_count = new_mem->num_pages;
616 	int ret;
617 
618 	page_count = new_mem->num_pages;
619 	while (page_count) {
620 		int line_count = (page_count > 8191) ? 8191 : page_count;
621 
622 		ret = RING_SPACE(chan, 11);
623 		if (ret)
624 			return ret;
625 
626 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 8);
627 		OUT_RING  (chan, upper_32_bits(src_offset));
628 		OUT_RING  (chan, lower_32_bits(src_offset));
629 		OUT_RING  (chan, upper_32_bits(dst_offset));
630 		OUT_RING  (chan, lower_32_bits(dst_offset));
631 		OUT_RING  (chan, PAGE_SIZE);
632 		OUT_RING  (chan, PAGE_SIZE);
633 		OUT_RING  (chan, PAGE_SIZE);
634 		OUT_RING  (chan, line_count);
635 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
636 		OUT_RING  (chan, 0x00000110);
637 
638 		page_count -= line_count;
639 		src_offset += (PAGE_SIZE * line_count);
640 		dst_offset += (PAGE_SIZE * line_count);
641 	}
642 
643 	return 0;
644 }
645 
646 static int
647 nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
648 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
649 {
650 	struct nouveau_mem *node = old_mem->mm_node;
651 	u64 src_offset = node->vma[0].offset;
652 	u64 dst_offset = node->vma[1].offset;
653 	u32 page_count = new_mem->num_pages;
654 	int ret;
655 
656 	page_count = new_mem->num_pages;
657 	while (page_count) {
658 		int line_count = (page_count > 2047) ? 2047 : page_count;
659 
660 		ret = RING_SPACE(chan, 12);
661 		if (ret)
662 			return ret;
663 
664 		BEGIN_NVC0(chan, NvSubCopy, 0x0238, 2);
665 		OUT_RING  (chan, upper_32_bits(dst_offset));
666 		OUT_RING  (chan, lower_32_bits(dst_offset));
667 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 6);
668 		OUT_RING  (chan, upper_32_bits(src_offset));
669 		OUT_RING  (chan, lower_32_bits(src_offset));
670 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
671 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
672 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
673 		OUT_RING  (chan, line_count);
674 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
675 		OUT_RING  (chan, 0x00100110);
676 
677 		page_count -= line_count;
678 		src_offset += (PAGE_SIZE * line_count);
679 		dst_offset += (PAGE_SIZE * line_count);
680 	}
681 
682 	return 0;
683 }
684 
685 static int
686 nva3_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
687 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
688 {
689 	struct nouveau_mem *node = old_mem->mm_node;
690 	u64 src_offset = node->vma[0].offset;
691 	u64 dst_offset = node->vma[1].offset;
692 	u32 page_count = new_mem->num_pages;
693 	int ret;
694 
695 	page_count = new_mem->num_pages;
696 	while (page_count) {
697 		int line_count = (page_count > 8191) ? 8191 : page_count;
698 
699 		ret = RING_SPACE(chan, 11);
700 		if (ret)
701 			return ret;
702 
703 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
704 		OUT_RING  (chan, upper_32_bits(src_offset));
705 		OUT_RING  (chan, lower_32_bits(src_offset));
706 		OUT_RING  (chan, upper_32_bits(dst_offset));
707 		OUT_RING  (chan, lower_32_bits(dst_offset));
708 		OUT_RING  (chan, PAGE_SIZE);
709 		OUT_RING  (chan, PAGE_SIZE);
710 		OUT_RING  (chan, PAGE_SIZE);
711 		OUT_RING  (chan, line_count);
712 		BEGIN_NV04(chan, NvSubCopy, 0x0300, 1);
713 		OUT_RING  (chan, 0x00000110);
714 
715 		page_count -= line_count;
716 		src_offset += (PAGE_SIZE * line_count);
717 		dst_offset += (PAGE_SIZE * line_count);
718 	}
719 
720 	return 0;
721 }
722 
723 static int
724 nv98_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
725 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
726 {
727 	struct nouveau_mem *node = old_mem->mm_node;
728 	int ret = RING_SPACE(chan, 7);
729 	if (ret == 0) {
730 		BEGIN_NV04(chan, NvSubCopy, 0x0320, 6);
731 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
732 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
733 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
734 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
735 		OUT_RING  (chan, 0x00000000 /* COPY */);
736 		OUT_RING  (chan, new_mem->num_pages << PAGE_SHIFT);
737 	}
738 	return ret;
739 }
740 
741 static int
742 nv84_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
743 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
744 {
745 	struct nouveau_mem *node = old_mem->mm_node;
746 	int ret = RING_SPACE(chan, 7);
747 	if (ret == 0) {
748 		BEGIN_NV04(chan, NvSubCopy, 0x0304, 6);
749 		OUT_RING  (chan, new_mem->num_pages << PAGE_SHIFT);
750 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
751 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
752 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
753 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
754 		OUT_RING  (chan, 0x00000000 /* MODE_COPY, QUERY_NONE */);
755 	}
756 	return ret;
757 }
758 
759 static int
760 nv50_bo_move_init(struct nouveau_channel *chan, u32 handle)
761 {
762 	int ret = RING_SPACE(chan, 6);
763 	if (ret == 0) {
764 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
765 		OUT_RING  (chan, handle);
766 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 3);
767 		OUT_RING  (chan, chan->drm->ntfy.handle);
768 		OUT_RING  (chan, chan->vram.handle);
769 		OUT_RING  (chan, chan->vram.handle);
770 	}
771 
772 	return ret;
773 }
774 
775 static int
776 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
777 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
778 {
779 	struct nouveau_mem *node = old_mem->mm_node;
780 	u64 length = (new_mem->num_pages << PAGE_SHIFT);
781 	u64 src_offset = node->vma[0].offset;
782 	u64 dst_offset = node->vma[1].offset;
783 	int src_tiled = !!node->memtype;
784 	int dst_tiled = !!((struct nouveau_mem *)new_mem->mm_node)->memtype;
785 	int ret;
786 
787 	while (length) {
788 		u32 amount, stride, height;
789 
790 		ret = RING_SPACE(chan, 18 + 6 * (src_tiled + dst_tiled));
791 		if (ret)
792 			return ret;
793 
794 		amount  = min(length, (u64)(4 * 1024 * 1024));
795 		stride  = 16 * 4;
796 		height  = amount / stride;
797 
798 		if (src_tiled) {
799 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 7);
800 			OUT_RING  (chan, 0);
801 			OUT_RING  (chan, 0);
802 			OUT_RING  (chan, stride);
803 			OUT_RING  (chan, height);
804 			OUT_RING  (chan, 1);
805 			OUT_RING  (chan, 0);
806 			OUT_RING  (chan, 0);
807 		} else {
808 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 1);
809 			OUT_RING  (chan, 1);
810 		}
811 		if (dst_tiled) {
812 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 7);
813 			OUT_RING  (chan, 0);
814 			OUT_RING  (chan, 0);
815 			OUT_RING  (chan, stride);
816 			OUT_RING  (chan, height);
817 			OUT_RING  (chan, 1);
818 			OUT_RING  (chan, 0);
819 			OUT_RING  (chan, 0);
820 		} else {
821 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 1);
822 			OUT_RING  (chan, 1);
823 		}
824 
825 		BEGIN_NV04(chan, NvSubCopy, 0x0238, 2);
826 		OUT_RING  (chan, upper_32_bits(src_offset));
827 		OUT_RING  (chan, upper_32_bits(dst_offset));
828 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
829 		OUT_RING  (chan, lower_32_bits(src_offset));
830 		OUT_RING  (chan, lower_32_bits(dst_offset));
831 		OUT_RING  (chan, stride);
832 		OUT_RING  (chan, stride);
833 		OUT_RING  (chan, stride);
834 		OUT_RING  (chan, height);
835 		OUT_RING  (chan, 0x00000101);
836 		OUT_RING  (chan, 0x00000000);
837 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
838 		OUT_RING  (chan, 0);
839 
840 		length -= amount;
841 		src_offset += amount;
842 		dst_offset += amount;
843 	}
844 
845 	return 0;
846 }
847 
848 static int
849 nv04_bo_move_init(struct nouveau_channel *chan, u32 handle)
850 {
851 	int ret = RING_SPACE(chan, 4);
852 	if (ret == 0) {
853 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
854 		OUT_RING  (chan, handle);
855 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 1);
856 		OUT_RING  (chan, chan->drm->ntfy.handle);
857 	}
858 
859 	return ret;
860 }
861 
862 static inline uint32_t
863 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
864 		      struct nouveau_channel *chan, struct ttm_mem_reg *mem)
865 {
866 	if (mem->mem_type == TTM_PL_TT)
867 		return NvDmaTT;
868 	return chan->vram.handle;
869 }
870 
871 static int
872 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
873 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
874 {
875 	u32 src_offset = old_mem->start << PAGE_SHIFT;
876 	u32 dst_offset = new_mem->start << PAGE_SHIFT;
877 	u32 page_count = new_mem->num_pages;
878 	int ret;
879 
880 	ret = RING_SPACE(chan, 3);
881 	if (ret)
882 		return ret;
883 
884 	BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
885 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
886 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
887 
888 	page_count = new_mem->num_pages;
889 	while (page_count) {
890 		int line_count = (page_count > 2047) ? 2047 : page_count;
891 
892 		ret = RING_SPACE(chan, 11);
893 		if (ret)
894 			return ret;
895 
896 		BEGIN_NV04(chan, NvSubCopy,
897 				 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
898 		OUT_RING  (chan, src_offset);
899 		OUT_RING  (chan, dst_offset);
900 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
901 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
902 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
903 		OUT_RING  (chan, line_count);
904 		OUT_RING  (chan, 0x00000101);
905 		OUT_RING  (chan, 0x00000000);
906 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
907 		OUT_RING  (chan, 0);
908 
909 		page_count -= line_count;
910 		src_offset += (PAGE_SIZE * line_count);
911 		dst_offset += (PAGE_SIZE * line_count);
912 	}
913 
914 	return 0;
915 }
916 
917 static int
918 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
919 		     struct ttm_mem_reg *mem)
920 {
921 	struct nouveau_mem *old_node = bo->mem.mm_node;
922 	struct nouveau_mem *new_node = mem->mm_node;
923 	u64 size = (u64)mem->num_pages << PAGE_SHIFT;
924 	int ret;
925 
926 	ret = nouveau_vm_get(drm->client.vm, size, old_node->page_shift,
927 			     NV_MEM_ACCESS_RW, &old_node->vma[0]);
928 	if (ret)
929 		return ret;
930 
931 	ret = nouveau_vm_get(drm->client.vm, size, new_node->page_shift,
932 			     NV_MEM_ACCESS_RW, &old_node->vma[1]);
933 	if (ret) {
934 		nouveau_vm_put(&old_node->vma[0]);
935 		return ret;
936 	}
937 
938 	nouveau_vm_map(&old_node->vma[0], old_node);
939 	nouveau_vm_map(&old_node->vma[1], new_node);
940 	return 0;
941 }
942 
943 static int
944 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
945 		     bool no_wait_gpu, struct ttm_mem_reg *new_mem)
946 {
947 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
948 	struct nouveau_channel *chan = drm->ttm.chan;
949 	struct nouveau_cli *cli = (void *)nvif_client(&chan->device->base);
950 	struct nouveau_fence *fence;
951 	int ret;
952 
953 	/* create temporary vmas for the transfer and attach them to the
954 	 * old nouveau_mem node, these will get cleaned up after ttm has
955 	 * destroyed the ttm_mem_reg
956 	 */
957 	if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
958 		ret = nouveau_bo_move_prep(drm, bo, new_mem);
959 		if (ret)
960 			return ret;
961 	}
962 
963 	mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
964 	ret = nouveau_fence_sync(bo->sync_obj, chan);
965 	if (ret == 0) {
966 		ret = drm->ttm.move(chan, bo, &bo->mem, new_mem);
967 		if (ret == 0) {
968 			ret = nouveau_fence_new(chan, false, &fence);
969 			if (ret == 0) {
970 				ret = ttm_bo_move_accel_cleanup(bo, fence,
971 								evict,
972 								no_wait_gpu,
973 								new_mem);
974 				nouveau_fence_unref(&fence);
975 			}
976 		}
977 	}
978 	mutex_unlock(&cli->mutex);
979 	return ret;
980 }
981 
982 void
983 nouveau_bo_move_init(struct nouveau_drm *drm)
984 {
985 	static const struct {
986 		const char *name;
987 		int engine;
988 		u32 oclass;
989 		int (*exec)(struct nouveau_channel *,
990 			    struct ttm_buffer_object *,
991 			    struct ttm_mem_reg *, struct ttm_mem_reg *);
992 		int (*init)(struct nouveau_channel *, u32 handle);
993 	} _methods[] = {
994 		{  "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
995 		{  "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
996 		{ "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
997 		{ "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
998 		{  "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
999 		{ "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
1000 		{  "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
1001 		{  "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
1002 		{  "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
1003 		{},
1004 		{ "CRYPT", 0, 0x88b4, nv98_bo_move_exec, nv50_bo_move_init },
1005 	}, *mthd = _methods;
1006 	const char *name = "CPU";
1007 	int ret;
1008 
1009 	do {
1010 		struct nouveau_channel *chan;
1011 
1012 		if (mthd->engine)
1013 			chan = drm->cechan;
1014 		else
1015 			chan = drm->channel;
1016 		if (chan == NULL)
1017 			continue;
1018 
1019 		ret = nvif_object_init(chan->object, NULL,
1020 				       mthd->oclass | (mthd->engine << 16),
1021 				       mthd->oclass, NULL, 0,
1022 				       &drm->ttm.copy);
1023 		if (ret == 0) {
1024 			ret = mthd->init(chan, drm->ttm.copy.handle);
1025 			if (ret) {
1026 				nvif_object_fini(&drm->ttm.copy);
1027 				continue;
1028 			}
1029 
1030 			drm->ttm.move = mthd->exec;
1031 			drm->ttm.chan = chan;
1032 			name = mthd->name;
1033 			break;
1034 		}
1035 	} while ((++mthd)->exec);
1036 
1037 	NV_INFO(drm, "MM: using %s for buffer copies\n", name);
1038 }
1039 
1040 static int
1041 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
1042 		      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1043 {
1044 	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
1045 	struct ttm_placement placement;
1046 	struct ttm_mem_reg tmp_mem;
1047 	int ret;
1048 
1049 	placement.fpfn = placement.lpfn = 0;
1050 	placement.num_placement = placement.num_busy_placement = 1;
1051 	placement.placement = placement.busy_placement = &placement_memtype;
1052 
1053 	tmp_mem = *new_mem;
1054 	tmp_mem.mm_node = NULL;
1055 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1056 	if (ret)
1057 		return ret;
1058 
1059 	ret = ttm_tt_bind(bo->ttm, &tmp_mem);
1060 	if (ret)
1061 		goto out;
1062 
1063 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, &tmp_mem);
1064 	if (ret)
1065 		goto out;
1066 
1067 	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
1068 out:
1069 	ttm_bo_mem_put(bo, &tmp_mem);
1070 	return ret;
1071 }
1072 
1073 static int
1074 nouveau_bo_move_flips(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_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
1093 	if (ret)
1094 		goto out;
1095 
1096 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_mem);
1097 	if (ret)
1098 		goto out;
1099 
1100 out:
1101 	ttm_bo_mem_put(bo, &tmp_mem);
1102 	return ret;
1103 }
1104 
1105 static void
1106 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem)
1107 {
1108 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1109 	struct nouveau_vma *vma;
1110 
1111 	/* ttm can now (stupidly) pass the driver bos it didn't create... */
1112 	if (bo->destroy != nouveau_bo_del_ttm)
1113 		return;
1114 
1115 	list_for_each_entry(vma, &nvbo->vma_list, head) {
1116 		if (new_mem && new_mem->mem_type != TTM_PL_SYSTEM &&
1117 			      (new_mem->mem_type == TTM_PL_VRAM ||
1118 			       nvbo->page_shift != vma->vm->vmm->lpg_shift)) {
1119 			nouveau_vm_map(vma, new_mem->mm_node);
1120 		} else {
1121 			nouveau_vm_unmap(vma);
1122 		}
1123 	}
1124 }
1125 
1126 static int
1127 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
1128 		   struct nouveau_drm_tile **new_tile)
1129 {
1130 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1131 	struct drm_device *dev = drm->dev;
1132 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1133 	u64 offset = new_mem->start << PAGE_SHIFT;
1134 
1135 	*new_tile = NULL;
1136 	if (new_mem->mem_type != TTM_PL_VRAM)
1137 		return 0;
1138 
1139 	if (drm->device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
1140 		*new_tile = nv10_bo_set_tiling(dev, offset, new_mem->size,
1141 						nvbo->tile_mode,
1142 						nvbo->tile_flags);
1143 	}
1144 
1145 	return 0;
1146 }
1147 
1148 static void
1149 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1150 		      struct nouveau_drm_tile *new_tile,
1151 		      struct nouveau_drm_tile **old_tile)
1152 {
1153 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1154 	struct drm_device *dev = drm->dev;
1155 
1156 	nv10_bo_put_tile_region(dev, *old_tile, bo->sync_obj);
1157 	*old_tile = new_tile;
1158 }
1159 
1160 static int
1161 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
1162 		bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1163 {
1164 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1165 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1166 	struct ttm_mem_reg *old_mem = &bo->mem;
1167 	struct nouveau_drm_tile *new_tile = NULL;
1168 	int ret = 0;
1169 
1170 	if (drm->device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1171 		ret = nouveau_bo_vm_bind(bo, new_mem, &new_tile);
1172 		if (ret)
1173 			return ret;
1174 	}
1175 
1176 	/* Fake bo copy. */
1177 	if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
1178 		BUG_ON(bo->mem.mm_node != NULL);
1179 		bo->mem = *new_mem;
1180 		new_mem->mm_node = NULL;
1181 		goto out;
1182 	}
1183 
1184 	/* Hardware assisted copy. */
1185 	if (drm->ttm.move) {
1186 		if (new_mem->mem_type == TTM_PL_SYSTEM)
1187 			ret = nouveau_bo_move_flipd(bo, evict, intr,
1188 						    no_wait_gpu, new_mem);
1189 		else if (old_mem->mem_type == TTM_PL_SYSTEM)
1190 			ret = nouveau_bo_move_flips(bo, evict, intr,
1191 						    no_wait_gpu, new_mem);
1192 		else
1193 			ret = nouveau_bo_move_m2mf(bo, evict, intr,
1194 						   no_wait_gpu, new_mem);
1195 		if (!ret)
1196 			goto out;
1197 	}
1198 
1199 	/* Fallback to software copy. */
1200 	spin_lock(&bo->bdev->fence_lock);
1201 	ret = ttm_bo_wait(bo, true, intr, no_wait_gpu);
1202 	spin_unlock(&bo->bdev->fence_lock);
1203 	if (ret == 0)
1204 		ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
1205 
1206 out:
1207 	if (drm->device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1208 		if (ret)
1209 			nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1210 		else
1211 			nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1212 	}
1213 
1214 	return ret;
1215 }
1216 
1217 static int
1218 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1219 {
1220 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1221 
1222 	return drm_vma_node_verify_access(&nvbo->gem.vma_node, filp);
1223 }
1224 
1225 static int
1226 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1227 {
1228 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1229 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1230 	struct nouveau_mem *node = mem->mm_node;
1231 	int ret;
1232 
1233 	mem->bus.addr = NULL;
1234 	mem->bus.offset = 0;
1235 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
1236 	mem->bus.base = 0;
1237 	mem->bus.is_iomem = false;
1238 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
1239 		return -EINVAL;
1240 	switch (mem->mem_type) {
1241 	case TTM_PL_SYSTEM:
1242 		/* System memory */
1243 		return 0;
1244 	case TTM_PL_TT:
1245 #if __OS_HAS_AGP
1246 		if (drm->agp.stat == ENABLED) {
1247 			mem->bus.offset = mem->start << PAGE_SHIFT;
1248 			mem->bus.base = drm->agp.base;
1249 			mem->bus.is_iomem = !drm->dev->agp->cant_use_aperture;
1250 		}
1251 #endif
1252 		if (drm->device.info.family < NV_DEVICE_INFO_V0_TESLA || !node->memtype)
1253 			/* untiled */
1254 			break;
1255 		/* fallthrough, tiled memory */
1256 	case TTM_PL_VRAM:
1257 		mem->bus.offset = mem->start << PAGE_SHIFT;
1258 		mem->bus.base = nv_device_resource_start(nvkm_device(&drm->device), 1);
1259 		mem->bus.is_iomem = true;
1260 		if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
1261 			struct nouveau_bar *bar = nvkm_bar(&drm->device);
1262 
1263 			ret = bar->umap(bar, node, NV_MEM_ACCESS_RW,
1264 					&node->bar_vma);
1265 			if (ret)
1266 				return ret;
1267 
1268 			mem->bus.offset = node->bar_vma.offset;
1269 		}
1270 		break;
1271 	default:
1272 		return -EINVAL;
1273 	}
1274 	return 0;
1275 }
1276 
1277 static void
1278 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1279 {
1280 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1281 	struct nouveau_bar *bar = nvkm_bar(&drm->device);
1282 	struct nouveau_mem *node = mem->mm_node;
1283 
1284 	if (!node->bar_vma.node)
1285 		return;
1286 
1287 	bar->unmap(bar, &node->bar_vma);
1288 }
1289 
1290 static int
1291 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1292 {
1293 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1294 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1295 	struct nvif_device *device = &drm->device;
1296 	u32 mappable = nv_device_resource_len(nvkm_device(device), 1) >> PAGE_SHIFT;
1297 	int ret;
1298 
1299 	/* as long as the bo isn't in vram, and isn't tiled, we've got
1300 	 * nothing to do here.
1301 	 */
1302 	if (bo->mem.mem_type != TTM_PL_VRAM) {
1303 		if (drm->device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1304 		    !nouveau_bo_tile_layout(nvbo))
1305 			return 0;
1306 
1307 		if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1308 			nouveau_bo_placement_set(nvbo, TTM_PL_TT, 0);
1309 
1310 			ret = nouveau_bo_validate(nvbo, false, false);
1311 			if (ret)
1312 				return ret;
1313 		}
1314 		return 0;
1315 	}
1316 
1317 	/* make sure bo is in mappable vram */
1318 	if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1319 	    bo->mem.start + bo->mem.num_pages < mappable)
1320 		return 0;
1321 
1322 
1323 	nvbo->placement.fpfn = 0;
1324 	nvbo->placement.lpfn = mappable;
1325 	nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_VRAM, 0);
1326 	return nouveau_bo_validate(nvbo, false, false);
1327 }
1328 
1329 static int
1330 nouveau_ttm_tt_populate(struct ttm_tt *ttm)
1331 {
1332 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1333 	struct nouveau_drm *drm;
1334 	struct nouveau_device *device;
1335 	struct drm_device *dev;
1336 	struct device *pdev;
1337 	unsigned i;
1338 	int r;
1339 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1340 
1341 	if (ttm->state != tt_unpopulated)
1342 		return 0;
1343 
1344 	if (slave && ttm->sg) {
1345 		/* make userspace faulting work */
1346 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1347 						 ttm_dma->dma_address, ttm->num_pages);
1348 		ttm->state = tt_unbound;
1349 		return 0;
1350 	}
1351 
1352 	drm = nouveau_bdev(ttm->bdev);
1353 	device = nvkm_device(&drm->device);
1354 	dev = drm->dev;
1355 	pdev = nv_device_base(device);
1356 
1357 #if __OS_HAS_AGP
1358 	if (drm->agp.stat == ENABLED) {
1359 		return ttm_agp_tt_populate(ttm);
1360 	}
1361 #endif
1362 
1363 #ifdef CONFIG_SWIOTLB
1364 	if (swiotlb_nr_tbl()) {
1365 		return ttm_dma_populate((void *)ttm, dev->dev);
1366 	}
1367 #endif
1368 
1369 	r = ttm_pool_populate(ttm);
1370 	if (r) {
1371 		return r;
1372 	}
1373 
1374 	for (i = 0; i < ttm->num_pages; i++) {
1375 		dma_addr_t addr;
1376 
1377 		addr = dma_map_page(pdev, ttm->pages[i], 0, PAGE_SIZE,
1378 				    DMA_BIDIRECTIONAL);
1379 
1380 		if (dma_mapping_error(pdev, addr)) {
1381 			while (--i) {
1382 				dma_unmap_page(pdev, ttm_dma->dma_address[i],
1383 					       PAGE_SIZE, DMA_BIDIRECTIONAL);
1384 				ttm_dma->dma_address[i] = 0;
1385 			}
1386 			ttm_pool_unpopulate(ttm);
1387 			return -EFAULT;
1388 		}
1389 
1390 		ttm_dma->dma_address[i] = addr;
1391 	}
1392 	return 0;
1393 }
1394 
1395 static void
1396 nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
1397 {
1398 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1399 	struct nouveau_drm *drm;
1400 	struct nouveau_device *device;
1401 	struct drm_device *dev;
1402 	struct device *pdev;
1403 	unsigned i;
1404 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1405 
1406 	if (slave)
1407 		return;
1408 
1409 	drm = nouveau_bdev(ttm->bdev);
1410 	device = nvkm_device(&drm->device);
1411 	dev = drm->dev;
1412 	pdev = nv_device_base(device);
1413 
1414 #if __OS_HAS_AGP
1415 	if (drm->agp.stat == ENABLED) {
1416 		ttm_agp_tt_unpopulate(ttm);
1417 		return;
1418 	}
1419 #endif
1420 
1421 #ifdef CONFIG_SWIOTLB
1422 	if (swiotlb_nr_tbl()) {
1423 		ttm_dma_unpopulate((void *)ttm, dev->dev);
1424 		return;
1425 	}
1426 #endif
1427 
1428 	for (i = 0; i < ttm->num_pages; i++) {
1429 		if (ttm_dma->dma_address[i]) {
1430 			dma_unmap_page(pdev, ttm_dma->dma_address[i], PAGE_SIZE,
1431 				       DMA_BIDIRECTIONAL);
1432 		}
1433 	}
1434 
1435 	ttm_pool_unpopulate(ttm);
1436 }
1437 
1438 void
1439 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
1440 {
1441 	struct nouveau_fence *new_fence = nouveau_fence_ref(fence);
1442 	struct nouveau_fence *old_fence = NULL;
1443 
1444 	spin_lock(&nvbo->bo.bdev->fence_lock);
1445 	old_fence = nvbo->bo.sync_obj;
1446 	nvbo->bo.sync_obj = new_fence;
1447 	spin_unlock(&nvbo->bo.bdev->fence_lock);
1448 
1449 	nouveau_fence_unref(&old_fence);
1450 }
1451 
1452 static void
1453 nouveau_bo_fence_unref(void **sync_obj)
1454 {
1455 	nouveau_fence_unref((struct nouveau_fence **)sync_obj);
1456 }
1457 
1458 static void *
1459 nouveau_bo_fence_ref(void *sync_obj)
1460 {
1461 	return nouveau_fence_ref(sync_obj);
1462 }
1463 
1464 static bool
1465 nouveau_bo_fence_signalled(void *sync_obj)
1466 {
1467 	return nouveau_fence_done(sync_obj);
1468 }
1469 
1470 static int
1471 nouveau_bo_fence_wait(void *sync_obj, bool lazy, bool intr)
1472 {
1473 	return nouveau_fence_wait(sync_obj, lazy, intr);
1474 }
1475 
1476 static int
1477 nouveau_bo_fence_flush(void *sync_obj)
1478 {
1479 	return 0;
1480 }
1481 
1482 struct ttm_bo_driver nouveau_bo_driver = {
1483 	.ttm_tt_create = &nouveau_ttm_tt_create,
1484 	.ttm_tt_populate = &nouveau_ttm_tt_populate,
1485 	.ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1486 	.invalidate_caches = nouveau_bo_invalidate_caches,
1487 	.init_mem_type = nouveau_bo_init_mem_type,
1488 	.evict_flags = nouveau_bo_evict_flags,
1489 	.move_notify = nouveau_bo_move_ntfy,
1490 	.move = nouveau_bo_move,
1491 	.verify_access = nouveau_bo_verify_access,
1492 	.sync_obj_signaled = nouveau_bo_fence_signalled,
1493 	.sync_obj_wait = nouveau_bo_fence_wait,
1494 	.sync_obj_flush = nouveau_bo_fence_flush,
1495 	.sync_obj_unref = nouveau_bo_fence_unref,
1496 	.sync_obj_ref = nouveau_bo_fence_ref,
1497 	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1498 	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1499 	.io_mem_free = &nouveau_ttm_io_mem_free,
1500 };
1501 
1502 struct nouveau_vma *
1503 nouveau_bo_vma_find(struct nouveau_bo *nvbo, struct nouveau_vm *vm)
1504 {
1505 	struct nouveau_vma *vma;
1506 	list_for_each_entry(vma, &nvbo->vma_list, head) {
1507 		if (vma->vm == vm)
1508 			return vma;
1509 	}
1510 
1511 	return NULL;
1512 }
1513 
1514 int
1515 nouveau_bo_vma_add(struct nouveau_bo *nvbo, struct nouveau_vm *vm,
1516 		   struct nouveau_vma *vma)
1517 {
1518 	const u32 size = nvbo->bo.mem.num_pages << PAGE_SHIFT;
1519 	int ret;
1520 
1521 	ret = nouveau_vm_get(vm, size, nvbo->page_shift,
1522 			     NV_MEM_ACCESS_RW, vma);
1523 	if (ret)
1524 		return ret;
1525 
1526 	if ( nvbo->bo.mem.mem_type != TTM_PL_SYSTEM &&
1527 	    (nvbo->bo.mem.mem_type == TTM_PL_VRAM ||
1528 	     nvbo->page_shift != vma->vm->vmm->lpg_shift))
1529 		nouveau_vm_map(vma, nvbo->bo.mem.mm_node);
1530 
1531 	list_add_tail(&vma->head, &nvbo->vma_list);
1532 	vma->refcount = 1;
1533 	return 0;
1534 }
1535 
1536 void
1537 nouveau_bo_vma_del(struct nouveau_bo *nvbo, struct nouveau_vma *vma)
1538 {
1539 	if (vma->node) {
1540 		if (nvbo->bo.mem.mem_type != TTM_PL_SYSTEM)
1541 			nouveau_vm_unmap(vma);
1542 		nouveau_vm_put(vma);
1543 		list_del(&vma->head);
1544 	}
1545 }
1546