1 /*
2  * Copyright 2007 Dave Airlied
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * VA LINUX SYSTEMS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  */
24 /*
25  * Authors: Dave Airlied <airlied@linux.ie>
26  *	    Ben Skeggs   <darktama@iinet.net.au>
27  *	    Jeremy Kolb  <jkolb@brandeis.edu>
28  */
29 
30 #include <core/engine.h>
31 #include <linux/swiotlb.h>
32 
33 #include <subdev/fb.h>
34 #include <subdev/vm.h>
35 #include <subdev/bar.h>
36 
37 #include "nouveau_drm.h"
38 #include "nouveau_dma.h"
39 #include "nouveau_fence.h"
40 
41 #include "nouveau_bo.h"
42 #include "nouveau_ttm.h"
43 #include "nouveau_gem.h"
44 
45 /*
46  * NV10-NV40 tiling helpers
47  */
48 
49 static void
50 nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
51 			   u32 addr, u32 size, u32 pitch, u32 flags)
52 {
53 	struct nouveau_drm *drm = nouveau_drm(dev);
54 	int i = reg - drm->tile.reg;
55 	struct nouveau_fb *pfb = nouveau_fb(drm->device);
56 	struct nouveau_fb_tile *tile = &pfb->tile.region[i];
57 	struct nouveau_engine *engine;
58 
59 	nouveau_fence_unref(&reg->fence);
60 
61 	if (tile->pitch)
62 		pfb->tile.fini(pfb, i, tile);
63 
64 	if (pitch)
65 		pfb->tile.init(pfb, i, addr, size, pitch, flags, tile);
66 
67 	pfb->tile.prog(pfb, i, tile);
68 
69 	if ((engine = nouveau_engine(pfb, NVDEV_ENGINE_GR)))
70 		engine->tile_prog(engine, i);
71 	if ((engine = nouveau_engine(pfb, NVDEV_ENGINE_MPEG)))
72 		engine->tile_prog(engine, i);
73 }
74 
75 static struct nouveau_drm_tile *
76 nv10_bo_get_tile_region(struct drm_device *dev, int i)
77 {
78 	struct nouveau_drm *drm = nouveau_drm(dev);
79 	struct nouveau_drm_tile *tile = &drm->tile.reg[i];
80 
81 	spin_lock(&drm->tile.lock);
82 
83 	if (!tile->used &&
84 	    (!tile->fence || nouveau_fence_done(tile->fence)))
85 		tile->used = true;
86 	else
87 		tile = NULL;
88 
89 	spin_unlock(&drm->tile.lock);
90 	return tile;
91 }
92 
93 static void
94 nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
95 			struct nouveau_fence *fence)
96 {
97 	struct nouveau_drm *drm = nouveau_drm(dev);
98 
99 	if (tile) {
100 		spin_lock(&drm->tile.lock);
101 		tile->fence = nouveau_fence_ref(fence);
102 		tile->used = false;
103 		spin_unlock(&drm->tile.lock);
104 	}
105 }
106 
107 static struct nouveau_drm_tile *
108 nv10_bo_set_tiling(struct drm_device *dev, u32 addr,
109 		   u32 size, u32 pitch, u32 flags)
110 {
111 	struct nouveau_drm *drm = nouveau_drm(dev);
112 	struct nouveau_fb *pfb = nouveau_fb(drm->device);
113 	struct nouveau_drm_tile *tile, *found = NULL;
114 	int i;
115 
116 	for (i = 0; i < pfb->tile.regions; i++) {
117 		tile = nv10_bo_get_tile_region(dev, i);
118 
119 		if (pitch && !found) {
120 			found = tile;
121 			continue;
122 
123 		} else if (tile && pfb->tile.region[i].pitch) {
124 			/* Kill an unused tile region. */
125 			nv10_bo_update_tile_region(dev, tile, 0, 0, 0, 0);
126 		}
127 
128 		nv10_bo_put_tile_region(dev, tile, NULL);
129 	}
130 
131 	if (found)
132 		nv10_bo_update_tile_region(dev, found, addr, size,
133 					    pitch, flags);
134 	return found;
135 }
136 
137 static void
138 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
139 {
140 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
141 	struct drm_device *dev = drm->dev;
142 	struct nouveau_bo *nvbo = nouveau_bo(bo);
143 
144 	if (unlikely(nvbo->gem.filp))
145 		DRM_ERROR("bo %p still attached to GEM object\n", bo);
146 	WARN_ON(nvbo->pin_refcnt > 0);
147 	nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
148 	kfree(nvbo);
149 }
150 
151 static void
152 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
153 		       int *align, int *size)
154 {
155 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
156 	struct nouveau_device *device = nv_device(drm->device);
157 
158 	if (device->card_type < NV_50) {
159 		if (nvbo->tile_mode) {
160 			if (device->chipset >= 0x40) {
161 				*align = 65536;
162 				*size = roundup(*size, 64 * nvbo->tile_mode);
163 
164 			} else if (device->chipset >= 0x30) {
165 				*align = 32768;
166 				*size = roundup(*size, 64 * nvbo->tile_mode);
167 
168 			} else if (device->chipset >= 0x20) {
169 				*align = 16384;
170 				*size = roundup(*size, 64 * nvbo->tile_mode);
171 
172 			} else if (device->chipset >= 0x10) {
173 				*align = 16384;
174 				*size = roundup(*size, 32 * nvbo->tile_mode);
175 			}
176 		}
177 	} else {
178 		*size = roundup(*size, (1 << nvbo->page_shift));
179 		*align = max((1 <<  nvbo->page_shift), *align);
180 	}
181 
182 	*size = roundup(*size, PAGE_SIZE);
183 }
184 
185 int
186 nouveau_bo_new(struct drm_device *dev, int size, int align,
187 	       uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
188 	       struct sg_table *sg,
189 	       struct nouveau_bo **pnvbo)
190 {
191 	struct nouveau_drm *drm = nouveau_drm(dev);
192 	struct nouveau_bo *nvbo;
193 	size_t acc_size;
194 	int ret;
195 	int type = ttm_bo_type_device;
196 	int lpg_shift = 12;
197 	int max_size;
198 
199 	if (drm->client.base.vm)
200 		lpg_shift = drm->client.base.vm->vmm->lpg_shift;
201 	max_size = INT_MAX & ~((1 << lpg_shift) - 1);
202 
203 	if (size <= 0 || size > max_size) {
204 		nv_warn(drm, "skipped size %x\n", (u32)size);
205 		return -EINVAL;
206 	}
207 
208 	if (sg)
209 		type = ttm_bo_type_sg;
210 
211 	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
212 	if (!nvbo)
213 		return -ENOMEM;
214 	INIT_LIST_HEAD(&nvbo->head);
215 	INIT_LIST_HEAD(&nvbo->entry);
216 	INIT_LIST_HEAD(&nvbo->vma_list);
217 	nvbo->tile_mode = tile_mode;
218 	nvbo->tile_flags = tile_flags;
219 	nvbo->bo.bdev = &drm->ttm.bdev;
220 
221 	nvbo->page_shift = 12;
222 	if (drm->client.base.vm) {
223 		if (!(flags & TTM_PL_FLAG_TT) && size > 256 * 1024)
224 			nvbo->page_shift = drm->client.base.vm->vmm->lpg_shift;
225 	}
226 
227 	nouveau_bo_fixup_align(nvbo, flags, &align, &size);
228 	nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
229 	nouveau_bo_placement_set(nvbo, flags, 0);
230 
231 	acc_size = ttm_bo_dma_acc_size(&drm->ttm.bdev, size,
232 				       sizeof(struct nouveau_bo));
233 
234 	ret = ttm_bo_init(&drm->ttm.bdev, &nvbo->bo, size,
235 			  type, &nvbo->placement,
236 			  align >> PAGE_SHIFT, false, NULL, acc_size, sg,
237 			  nouveau_bo_del_ttm);
238 	if (ret) {
239 		/* ttm will call nouveau_bo_del_ttm if it fails.. */
240 		return ret;
241 	}
242 
243 	*pnvbo = nvbo;
244 	return 0;
245 }
246 
247 static void
248 set_placement_list(uint32_t *pl, unsigned *n, uint32_t type, uint32_t flags)
249 {
250 	*n = 0;
251 
252 	if (type & TTM_PL_FLAG_VRAM)
253 		pl[(*n)++] = TTM_PL_FLAG_VRAM | flags;
254 	if (type & TTM_PL_FLAG_TT)
255 		pl[(*n)++] = TTM_PL_FLAG_TT | flags;
256 	if (type & TTM_PL_FLAG_SYSTEM)
257 		pl[(*n)++] = TTM_PL_FLAG_SYSTEM | flags;
258 }
259 
260 static void
261 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
262 {
263 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
264 	struct nouveau_fb *pfb = nouveau_fb(drm->device);
265 	u32 vram_pages = pfb->ram->size >> PAGE_SHIFT;
266 
267 	if ((nv_device(drm->device)->card_type == NV_10 ||
268 	     nv_device(drm->device)->card_type == NV_11) &&
269 	    nvbo->tile_mode && (type & TTM_PL_FLAG_VRAM) &&
270 	    nvbo->bo.mem.num_pages < vram_pages / 4) {
271 		/*
272 		 * Make sure that the color and depth buffers are handled
273 		 * by independent memory controller units. Up to a 9x
274 		 * speed up when alpha-blending and depth-test are enabled
275 		 * at the same time.
276 		 */
277 		if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
278 			nvbo->placement.fpfn = vram_pages / 2;
279 			nvbo->placement.lpfn = ~0;
280 		} else {
281 			nvbo->placement.fpfn = 0;
282 			nvbo->placement.lpfn = vram_pages / 2;
283 		}
284 	}
285 }
286 
287 void
288 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
289 {
290 	struct ttm_placement *pl = &nvbo->placement;
291 	uint32_t flags = TTM_PL_MASK_CACHING |
292 		(nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
293 
294 	pl->placement = nvbo->placements;
295 	set_placement_list(nvbo->placements, &pl->num_placement,
296 			   type, flags);
297 
298 	pl->busy_placement = nvbo->busy_placements;
299 	set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
300 			   type | busy, flags);
301 
302 	set_placement_range(nvbo, type);
303 }
304 
305 int
306 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype)
307 {
308 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
309 	struct ttm_buffer_object *bo = &nvbo->bo;
310 	int ret;
311 
312 	ret = ttm_bo_reserve(bo, false, false, false, 0);
313 	if (ret)
314 		goto out;
315 
316 	if (nvbo->pin_refcnt && !(memtype & (1 << bo->mem.mem_type))) {
317 		NV_ERROR(drm, "bo %p pinned elsewhere: 0x%08x vs 0x%08x\n", bo,
318 			 1 << bo->mem.mem_type, memtype);
319 		ret = -EINVAL;
320 		goto out;
321 	}
322 
323 	if (nvbo->pin_refcnt++)
324 		goto out;
325 
326 	nouveau_bo_placement_set(nvbo, memtype, 0);
327 
328 	ret = nouveau_bo_validate(nvbo, false, false);
329 	if (ret == 0) {
330 		switch (bo->mem.mem_type) {
331 		case TTM_PL_VRAM:
332 			drm->gem.vram_available -= bo->mem.size;
333 			break;
334 		case TTM_PL_TT:
335 			drm->gem.gart_available -= bo->mem.size;
336 			break;
337 		default:
338 			break;
339 		}
340 	}
341 out:
342 	ttm_bo_unreserve(bo);
343 	return ret;
344 }
345 
346 int
347 nouveau_bo_unpin(struct nouveau_bo *nvbo)
348 {
349 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
350 	struct ttm_buffer_object *bo = &nvbo->bo;
351 	int ret, ref;
352 
353 	ret = ttm_bo_reserve(bo, false, false, false, 0);
354 	if (ret)
355 		return ret;
356 
357 	ref = --nvbo->pin_refcnt;
358 	WARN_ON_ONCE(ref < 0);
359 	if (ref)
360 		goto out;
361 
362 	nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
363 
364 	ret = nouveau_bo_validate(nvbo, false, false);
365 	if (ret == 0) {
366 		switch (bo->mem.mem_type) {
367 		case TTM_PL_VRAM:
368 			drm->gem.vram_available += bo->mem.size;
369 			break;
370 		case TTM_PL_TT:
371 			drm->gem.gart_available += bo->mem.size;
372 			break;
373 		default:
374 			break;
375 		}
376 	}
377 
378 out:
379 	ttm_bo_unreserve(bo);
380 	return ret;
381 }
382 
383 int
384 nouveau_bo_map(struct nouveau_bo *nvbo)
385 {
386 	int ret;
387 
388 	ret = ttm_bo_reserve(&nvbo->bo, false, false, false, 0);
389 	if (ret)
390 		return ret;
391 
392 	ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
393 	ttm_bo_unreserve(&nvbo->bo);
394 	return ret;
395 }
396 
397 void
398 nouveau_bo_unmap(struct nouveau_bo *nvbo)
399 {
400 	if (nvbo)
401 		ttm_bo_kunmap(&nvbo->kmap);
402 }
403 
404 int
405 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
406 		    bool no_wait_gpu)
407 {
408 	int ret;
409 
410 	ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement,
411 			      interruptible, no_wait_gpu);
412 	if (ret)
413 		return ret;
414 
415 	return 0;
416 }
417 
418 u16
419 nouveau_bo_rd16(struct nouveau_bo *nvbo, unsigned index)
420 {
421 	bool is_iomem;
422 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
423 	mem = &mem[index];
424 	if (is_iomem)
425 		return ioread16_native((void __force __iomem *)mem);
426 	else
427 		return *mem;
428 }
429 
430 void
431 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
432 {
433 	bool is_iomem;
434 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
435 	mem = &mem[index];
436 	if (is_iomem)
437 		iowrite16_native(val, (void __force __iomem *)mem);
438 	else
439 		*mem = val;
440 }
441 
442 u32
443 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
444 {
445 	bool is_iomem;
446 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
447 	mem = &mem[index];
448 	if (is_iomem)
449 		return ioread32_native((void __force __iomem *)mem);
450 	else
451 		return *mem;
452 }
453 
454 void
455 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
456 {
457 	bool is_iomem;
458 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
459 	mem = &mem[index];
460 	if (is_iomem)
461 		iowrite32_native(val, (void __force __iomem *)mem);
462 	else
463 		*mem = val;
464 }
465 
466 static struct ttm_tt *
467 nouveau_ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
468 		      uint32_t page_flags, struct page *dummy_read)
469 {
470 #if __OS_HAS_AGP
471 	struct nouveau_drm *drm = nouveau_bdev(bdev);
472 	struct drm_device *dev = drm->dev;
473 
474 	if (drm->agp.stat == ENABLED) {
475 		return ttm_agp_tt_create(bdev, dev->agp->bridge, size,
476 					 page_flags, dummy_read);
477 	}
478 #endif
479 
480 	return nouveau_sgdma_create_ttm(bdev, size, page_flags, dummy_read);
481 }
482 
483 static int
484 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
485 {
486 	/* We'll do this from user space. */
487 	return 0;
488 }
489 
490 static int
491 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
492 			 struct ttm_mem_type_manager *man)
493 {
494 	struct nouveau_drm *drm = nouveau_bdev(bdev);
495 
496 	switch (type) {
497 	case TTM_PL_SYSTEM:
498 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
499 		man->available_caching = TTM_PL_MASK_CACHING;
500 		man->default_caching = TTM_PL_FLAG_CACHED;
501 		break;
502 	case TTM_PL_VRAM:
503 		if (nv_device(drm->device)->card_type >= NV_50) {
504 			man->func = &nouveau_vram_manager;
505 			man->io_reserve_fastpath = false;
506 			man->use_io_reserve_lru = true;
507 		} else {
508 			man->func = &ttm_bo_manager_func;
509 		}
510 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
511 			     TTM_MEMTYPE_FLAG_MAPPABLE;
512 		man->available_caching = TTM_PL_FLAG_UNCACHED |
513 					 TTM_PL_FLAG_WC;
514 		man->default_caching = TTM_PL_FLAG_WC;
515 		break;
516 	case TTM_PL_TT:
517 		if (nv_device(drm->device)->card_type >= NV_50)
518 			man->func = &nouveau_gart_manager;
519 		else
520 		if (drm->agp.stat != ENABLED)
521 			man->func = &nv04_gart_manager;
522 		else
523 			man->func = &ttm_bo_manager_func;
524 
525 		if (drm->agp.stat == ENABLED) {
526 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
527 			man->available_caching = TTM_PL_FLAG_UNCACHED |
528 				TTM_PL_FLAG_WC;
529 			man->default_caching = TTM_PL_FLAG_WC;
530 		} else {
531 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
532 				     TTM_MEMTYPE_FLAG_CMA;
533 			man->available_caching = TTM_PL_MASK_CACHING;
534 			man->default_caching = TTM_PL_FLAG_CACHED;
535 		}
536 
537 		break;
538 	default:
539 		return -EINVAL;
540 	}
541 	return 0;
542 }
543 
544 static void
545 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
546 {
547 	struct nouveau_bo *nvbo = nouveau_bo(bo);
548 
549 	switch (bo->mem.mem_type) {
550 	case TTM_PL_VRAM:
551 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
552 					 TTM_PL_FLAG_SYSTEM);
553 		break;
554 	default:
555 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
556 		break;
557 	}
558 
559 	*pl = nvbo->placement;
560 }
561 
562 
563 static int
564 nve0_bo_move_init(struct nouveau_channel *chan, u32 handle)
565 {
566 	int ret = RING_SPACE(chan, 2);
567 	if (ret == 0) {
568 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
569 		OUT_RING  (chan, handle & 0x0000ffff);
570 		FIRE_RING (chan);
571 	}
572 	return ret;
573 }
574 
575 static int
576 nve0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
577 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
578 {
579 	struct nouveau_mem *node = old_mem->mm_node;
580 	int ret = RING_SPACE(chan, 10);
581 	if (ret == 0) {
582 		BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
583 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
584 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
585 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
586 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
587 		OUT_RING  (chan, PAGE_SIZE);
588 		OUT_RING  (chan, PAGE_SIZE);
589 		OUT_RING  (chan, PAGE_SIZE);
590 		OUT_RING  (chan, new_mem->num_pages);
591 		BEGIN_IMC0(chan, NvSubCopy, 0x0300, 0x0386);
592 	}
593 	return ret;
594 }
595 
596 static int
597 nvc0_bo_move_init(struct nouveau_channel *chan, u32 handle)
598 {
599 	int ret = RING_SPACE(chan, 2);
600 	if (ret == 0) {
601 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
602 		OUT_RING  (chan, handle);
603 	}
604 	return ret;
605 }
606 
607 static int
608 nvc0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
609 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
610 {
611 	struct nouveau_mem *node = old_mem->mm_node;
612 	u64 src_offset = node->vma[0].offset;
613 	u64 dst_offset = node->vma[1].offset;
614 	u32 page_count = new_mem->num_pages;
615 	int ret;
616 
617 	page_count = new_mem->num_pages;
618 	while (page_count) {
619 		int line_count = (page_count > 8191) ? 8191 : page_count;
620 
621 		ret = RING_SPACE(chan, 11);
622 		if (ret)
623 			return ret;
624 
625 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 8);
626 		OUT_RING  (chan, upper_32_bits(src_offset));
627 		OUT_RING  (chan, lower_32_bits(src_offset));
628 		OUT_RING  (chan, upper_32_bits(dst_offset));
629 		OUT_RING  (chan, lower_32_bits(dst_offset));
630 		OUT_RING  (chan, PAGE_SIZE);
631 		OUT_RING  (chan, PAGE_SIZE);
632 		OUT_RING  (chan, PAGE_SIZE);
633 		OUT_RING  (chan, line_count);
634 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
635 		OUT_RING  (chan, 0x00000110);
636 
637 		page_count -= line_count;
638 		src_offset += (PAGE_SIZE * line_count);
639 		dst_offset += (PAGE_SIZE * line_count);
640 	}
641 
642 	return 0;
643 }
644 
645 static int
646 nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
647 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
648 {
649 	struct nouveau_mem *node = old_mem->mm_node;
650 	u64 src_offset = node->vma[0].offset;
651 	u64 dst_offset = node->vma[1].offset;
652 	u32 page_count = new_mem->num_pages;
653 	int ret;
654 
655 	page_count = new_mem->num_pages;
656 	while (page_count) {
657 		int line_count = (page_count > 2047) ? 2047 : page_count;
658 
659 		ret = RING_SPACE(chan, 12);
660 		if (ret)
661 			return ret;
662 
663 		BEGIN_NVC0(chan, NvSubCopy, 0x0238, 2);
664 		OUT_RING  (chan, upper_32_bits(dst_offset));
665 		OUT_RING  (chan, lower_32_bits(dst_offset));
666 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 6);
667 		OUT_RING  (chan, upper_32_bits(src_offset));
668 		OUT_RING  (chan, lower_32_bits(src_offset));
669 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
670 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
671 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
672 		OUT_RING  (chan, line_count);
673 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
674 		OUT_RING  (chan, 0x00100110);
675 
676 		page_count -= line_count;
677 		src_offset += (PAGE_SIZE * line_count);
678 		dst_offset += (PAGE_SIZE * line_count);
679 	}
680 
681 	return 0;
682 }
683 
684 static int
685 nva3_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
686 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
687 {
688 	struct nouveau_mem *node = old_mem->mm_node;
689 	u64 src_offset = node->vma[0].offset;
690 	u64 dst_offset = node->vma[1].offset;
691 	u32 page_count = new_mem->num_pages;
692 	int ret;
693 
694 	page_count = new_mem->num_pages;
695 	while (page_count) {
696 		int line_count = (page_count > 8191) ? 8191 : page_count;
697 
698 		ret = RING_SPACE(chan, 11);
699 		if (ret)
700 			return ret;
701 
702 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
703 		OUT_RING  (chan, upper_32_bits(src_offset));
704 		OUT_RING  (chan, lower_32_bits(src_offset));
705 		OUT_RING  (chan, upper_32_bits(dst_offset));
706 		OUT_RING  (chan, lower_32_bits(dst_offset));
707 		OUT_RING  (chan, PAGE_SIZE);
708 		OUT_RING  (chan, PAGE_SIZE);
709 		OUT_RING  (chan, PAGE_SIZE);
710 		OUT_RING  (chan, line_count);
711 		BEGIN_NV04(chan, NvSubCopy, 0x0300, 1);
712 		OUT_RING  (chan, 0x00000110);
713 
714 		page_count -= line_count;
715 		src_offset += (PAGE_SIZE * line_count);
716 		dst_offset += (PAGE_SIZE * line_count);
717 	}
718 
719 	return 0;
720 }
721 
722 static int
723 nv98_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
724 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
725 {
726 	struct nouveau_mem *node = old_mem->mm_node;
727 	int ret = RING_SPACE(chan, 7);
728 	if (ret == 0) {
729 		BEGIN_NV04(chan, NvSubCopy, 0x0320, 6);
730 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
731 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
732 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
733 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
734 		OUT_RING  (chan, 0x00000000 /* COPY */);
735 		OUT_RING  (chan, new_mem->num_pages << PAGE_SHIFT);
736 	}
737 	return ret;
738 }
739 
740 static int
741 nv84_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
742 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
743 {
744 	struct nouveau_mem *node = old_mem->mm_node;
745 	int ret = RING_SPACE(chan, 7);
746 	if (ret == 0) {
747 		BEGIN_NV04(chan, NvSubCopy, 0x0304, 6);
748 		OUT_RING  (chan, new_mem->num_pages << PAGE_SHIFT);
749 		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
750 		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
751 		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
752 		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
753 		OUT_RING  (chan, 0x00000000 /* MODE_COPY, QUERY_NONE */);
754 	}
755 	return ret;
756 }
757 
758 static int
759 nv50_bo_move_init(struct nouveau_channel *chan, u32 handle)
760 {
761 	int ret = RING_SPACE(chan, 6);
762 	if (ret == 0) {
763 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
764 		OUT_RING  (chan, handle);
765 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 3);
766 		OUT_RING  (chan, NvNotify0);
767 		OUT_RING  (chan, NvDmaFB);
768 		OUT_RING  (chan, NvDmaFB);
769 	}
770 
771 	return ret;
772 }
773 
774 static int
775 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
776 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
777 {
778 	struct nouveau_mem *node = old_mem->mm_node;
779 	u64 length = (new_mem->num_pages << PAGE_SHIFT);
780 	u64 src_offset = node->vma[0].offset;
781 	u64 dst_offset = node->vma[1].offset;
782 	int src_tiled = !!node->memtype;
783 	int dst_tiled = !!((struct nouveau_mem *)new_mem->mm_node)->memtype;
784 	int ret;
785 
786 	while (length) {
787 		u32 amount, stride, height;
788 
789 		ret = RING_SPACE(chan, 18 + 6 * (src_tiled + dst_tiled));
790 		if (ret)
791 			return ret;
792 
793 		amount  = min(length, (u64)(4 * 1024 * 1024));
794 		stride  = 16 * 4;
795 		height  = amount / stride;
796 
797 		if (src_tiled) {
798 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 7);
799 			OUT_RING  (chan, 0);
800 			OUT_RING  (chan, 0);
801 			OUT_RING  (chan, stride);
802 			OUT_RING  (chan, height);
803 			OUT_RING  (chan, 1);
804 			OUT_RING  (chan, 0);
805 			OUT_RING  (chan, 0);
806 		} else {
807 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 1);
808 			OUT_RING  (chan, 1);
809 		}
810 		if (dst_tiled) {
811 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 7);
812 			OUT_RING  (chan, 0);
813 			OUT_RING  (chan, 0);
814 			OUT_RING  (chan, stride);
815 			OUT_RING  (chan, height);
816 			OUT_RING  (chan, 1);
817 			OUT_RING  (chan, 0);
818 			OUT_RING  (chan, 0);
819 		} else {
820 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 1);
821 			OUT_RING  (chan, 1);
822 		}
823 
824 		BEGIN_NV04(chan, NvSubCopy, 0x0238, 2);
825 		OUT_RING  (chan, upper_32_bits(src_offset));
826 		OUT_RING  (chan, upper_32_bits(dst_offset));
827 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
828 		OUT_RING  (chan, lower_32_bits(src_offset));
829 		OUT_RING  (chan, lower_32_bits(dst_offset));
830 		OUT_RING  (chan, stride);
831 		OUT_RING  (chan, stride);
832 		OUT_RING  (chan, stride);
833 		OUT_RING  (chan, height);
834 		OUT_RING  (chan, 0x00000101);
835 		OUT_RING  (chan, 0x00000000);
836 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
837 		OUT_RING  (chan, 0);
838 
839 		length -= amount;
840 		src_offset += amount;
841 		dst_offset += amount;
842 	}
843 
844 	return 0;
845 }
846 
847 static int
848 nv04_bo_move_init(struct nouveau_channel *chan, u32 handle)
849 {
850 	int ret = RING_SPACE(chan, 4);
851 	if (ret == 0) {
852 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
853 		OUT_RING  (chan, handle);
854 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 1);
855 		OUT_RING  (chan, NvNotify0);
856 	}
857 
858 	return ret;
859 }
860 
861 static inline uint32_t
862 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
863 		      struct nouveau_channel *chan, struct ttm_mem_reg *mem)
864 {
865 	if (mem->mem_type == TTM_PL_TT)
866 		return NvDmaTT;
867 	return NvDmaFB;
868 }
869 
870 static int
871 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
872 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
873 {
874 	u32 src_offset = old_mem->start << PAGE_SHIFT;
875 	u32 dst_offset = new_mem->start << PAGE_SHIFT;
876 	u32 page_count = new_mem->num_pages;
877 	int ret;
878 
879 	ret = RING_SPACE(chan, 3);
880 	if (ret)
881 		return ret;
882 
883 	BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
884 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
885 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
886 
887 	page_count = new_mem->num_pages;
888 	while (page_count) {
889 		int line_count = (page_count > 2047) ? 2047 : page_count;
890 
891 		ret = RING_SPACE(chan, 11);
892 		if (ret)
893 			return ret;
894 
895 		BEGIN_NV04(chan, NvSubCopy,
896 				 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
897 		OUT_RING  (chan, src_offset);
898 		OUT_RING  (chan, dst_offset);
899 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
900 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
901 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
902 		OUT_RING  (chan, line_count);
903 		OUT_RING  (chan, 0x00000101);
904 		OUT_RING  (chan, 0x00000000);
905 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
906 		OUT_RING  (chan, 0);
907 
908 		page_count -= line_count;
909 		src_offset += (PAGE_SIZE * line_count);
910 		dst_offset += (PAGE_SIZE * line_count);
911 	}
912 
913 	return 0;
914 }
915 
916 static int
917 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
918 		     struct ttm_mem_reg *mem)
919 {
920 	struct nouveau_mem *old_node = bo->mem.mm_node;
921 	struct nouveau_mem *new_node = mem->mm_node;
922 	u64 size = (u64)mem->num_pages << PAGE_SHIFT;
923 	int ret;
924 
925 	ret = nouveau_vm_get(nv_client(drm)->vm, size, old_node->page_shift,
926 			     NV_MEM_ACCESS_RW, &old_node->vma[0]);
927 	if (ret)
928 		return ret;
929 
930 	ret = nouveau_vm_get(nv_client(drm)->vm, size, new_node->page_shift,
931 			     NV_MEM_ACCESS_RW, &old_node->vma[1]);
932 	if (ret) {
933 		nouveau_vm_put(&old_node->vma[0]);
934 		return ret;
935 	}
936 
937 	nouveau_vm_map(&old_node->vma[0], old_node);
938 	nouveau_vm_map(&old_node->vma[1], new_node);
939 	return 0;
940 }
941 
942 static int
943 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
944 		     bool no_wait_gpu, struct ttm_mem_reg *new_mem)
945 {
946 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
947 	struct nouveau_channel *chan = drm->ttm.chan;
948 	struct nouveau_fence *fence;
949 	int ret;
950 
951 	/* create temporary vmas for the transfer and attach them to the
952 	 * old nouveau_mem node, these will get cleaned up after ttm has
953 	 * destroyed the ttm_mem_reg
954 	 */
955 	if (nv_device(drm->device)->card_type >= NV_50) {
956 		ret = nouveau_bo_move_prep(drm, bo, new_mem);
957 		if (ret)
958 			return ret;
959 	}
960 
961 	mutex_lock_nested(&chan->cli->mutex, SINGLE_DEPTH_NESTING);
962 	ret = nouveau_fence_sync(bo->sync_obj, chan);
963 	if (ret == 0) {
964 		ret = drm->ttm.move(chan, bo, &bo->mem, new_mem);
965 		if (ret == 0) {
966 			ret = nouveau_fence_new(chan, false, &fence);
967 			if (ret == 0) {
968 				ret = ttm_bo_move_accel_cleanup(bo, fence,
969 								evict,
970 								no_wait_gpu,
971 								new_mem);
972 				nouveau_fence_unref(&fence);
973 			}
974 		}
975 	}
976 	mutex_unlock(&chan->cli->mutex);
977 	return ret;
978 }
979 
980 void
981 nouveau_bo_move_init(struct nouveau_drm *drm)
982 {
983 	static const struct {
984 		const char *name;
985 		int engine;
986 		u32 oclass;
987 		int (*exec)(struct nouveau_channel *,
988 			    struct ttm_buffer_object *,
989 			    struct ttm_mem_reg *, struct ttm_mem_reg *);
990 		int (*init)(struct nouveau_channel *, u32 handle);
991 	} _methods[] = {
992 		{  "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
993 		{  "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
994 		{ "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
995 		{ "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
996 		{  "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
997 		{ "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
998 		{  "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
999 		{  "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
1000 		{  "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
1001 		{},
1002 		{ "CRYPT", 0, 0x88b4, nv98_bo_move_exec, nv50_bo_move_init },
1003 	}, *mthd = _methods;
1004 	const char *name = "CPU";
1005 	int ret;
1006 
1007 	do {
1008 		struct nouveau_object *object;
1009 		struct nouveau_channel *chan;
1010 		u32 handle = (mthd->engine << 16) | mthd->oclass;
1011 
1012 		if (mthd->engine)
1013 			chan = drm->cechan;
1014 		else
1015 			chan = drm->channel;
1016 		if (chan == NULL)
1017 			continue;
1018 
1019 		ret = nouveau_object_new(nv_object(drm), chan->handle, handle,
1020 					 mthd->oclass, NULL, 0, &object);
1021 		if (ret == 0) {
1022 			ret = mthd->init(chan, handle);
1023 			if (ret) {
1024 				nouveau_object_del(nv_object(drm),
1025 						   chan->handle, handle);
1026 				continue;
1027 			}
1028 
1029 			drm->ttm.move = mthd->exec;
1030 			drm->ttm.chan = chan;
1031 			name = mthd->name;
1032 			break;
1033 		}
1034 	} while ((++mthd)->exec);
1035 
1036 	NV_INFO(drm, "MM: using %s for buffer copies\n", name);
1037 }
1038 
1039 static int
1040 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
1041 		      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1042 {
1043 	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
1044 	struct ttm_placement placement;
1045 	struct ttm_mem_reg tmp_mem;
1046 	int ret;
1047 
1048 	placement.fpfn = placement.lpfn = 0;
1049 	placement.num_placement = placement.num_busy_placement = 1;
1050 	placement.placement = placement.busy_placement = &placement_memtype;
1051 
1052 	tmp_mem = *new_mem;
1053 	tmp_mem.mm_node = NULL;
1054 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1055 	if (ret)
1056 		return ret;
1057 
1058 	ret = ttm_tt_bind(bo->ttm, &tmp_mem);
1059 	if (ret)
1060 		goto out;
1061 
1062 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, &tmp_mem);
1063 	if (ret)
1064 		goto out;
1065 
1066 	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
1067 out:
1068 	ttm_bo_mem_put(bo, &tmp_mem);
1069 	return ret;
1070 }
1071 
1072 static int
1073 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
1074 		      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1075 {
1076 	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
1077 	struct ttm_placement placement;
1078 	struct ttm_mem_reg tmp_mem;
1079 	int ret;
1080 
1081 	placement.fpfn = placement.lpfn = 0;
1082 	placement.num_placement = placement.num_busy_placement = 1;
1083 	placement.placement = placement.busy_placement = &placement_memtype;
1084 
1085 	tmp_mem = *new_mem;
1086 	tmp_mem.mm_node = NULL;
1087 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1088 	if (ret)
1089 		return ret;
1090 
1091 	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
1092 	if (ret)
1093 		goto out;
1094 
1095 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_mem);
1096 	if (ret)
1097 		goto out;
1098 
1099 out:
1100 	ttm_bo_mem_put(bo, &tmp_mem);
1101 	return ret;
1102 }
1103 
1104 static void
1105 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem)
1106 {
1107 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1108 	struct nouveau_vma *vma;
1109 
1110 	/* ttm can now (stupidly) pass the driver bos it didn't create... */
1111 	if (bo->destroy != nouveau_bo_del_ttm)
1112 		return;
1113 
1114 	list_for_each_entry(vma, &nvbo->vma_list, head) {
1115 		if (new_mem && new_mem->mem_type != TTM_PL_SYSTEM &&
1116 			      (new_mem->mem_type == TTM_PL_VRAM ||
1117 			       nvbo->page_shift != vma->vm->vmm->lpg_shift)) {
1118 			nouveau_vm_map(vma, new_mem->mm_node);
1119 		} else {
1120 			nouveau_vm_unmap(vma);
1121 		}
1122 	}
1123 }
1124 
1125 static int
1126 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
1127 		   struct nouveau_drm_tile **new_tile)
1128 {
1129 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1130 	struct drm_device *dev = drm->dev;
1131 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1132 	u64 offset = new_mem->start << PAGE_SHIFT;
1133 
1134 	*new_tile = NULL;
1135 	if (new_mem->mem_type != TTM_PL_VRAM)
1136 		return 0;
1137 
1138 	if (nv_device(drm->device)->card_type >= NV_10) {
1139 		*new_tile = nv10_bo_set_tiling(dev, offset, new_mem->size,
1140 						nvbo->tile_mode,
1141 						nvbo->tile_flags);
1142 	}
1143 
1144 	return 0;
1145 }
1146 
1147 static void
1148 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1149 		      struct nouveau_drm_tile *new_tile,
1150 		      struct nouveau_drm_tile **old_tile)
1151 {
1152 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1153 	struct drm_device *dev = drm->dev;
1154 
1155 	nv10_bo_put_tile_region(dev, *old_tile, bo->sync_obj);
1156 	*old_tile = new_tile;
1157 }
1158 
1159 static int
1160 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
1161 		bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1162 {
1163 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1164 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1165 	struct ttm_mem_reg *old_mem = &bo->mem;
1166 	struct nouveau_drm_tile *new_tile = NULL;
1167 	int ret = 0;
1168 
1169 	if (nv_device(drm->device)->card_type < NV_50) {
1170 		ret = nouveau_bo_vm_bind(bo, new_mem, &new_tile);
1171 		if (ret)
1172 			return ret;
1173 	}
1174 
1175 	/* Fake bo copy. */
1176 	if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
1177 		BUG_ON(bo->mem.mm_node != NULL);
1178 		bo->mem = *new_mem;
1179 		new_mem->mm_node = NULL;
1180 		goto out;
1181 	}
1182 
1183 	/* Hardware assisted copy. */
1184 	if (drm->ttm.move) {
1185 		if (new_mem->mem_type == TTM_PL_SYSTEM)
1186 			ret = nouveau_bo_move_flipd(bo, evict, intr,
1187 						    no_wait_gpu, new_mem);
1188 		else if (old_mem->mem_type == TTM_PL_SYSTEM)
1189 			ret = nouveau_bo_move_flips(bo, evict, intr,
1190 						    no_wait_gpu, new_mem);
1191 		else
1192 			ret = nouveau_bo_move_m2mf(bo, evict, intr,
1193 						   no_wait_gpu, new_mem);
1194 		if (!ret)
1195 			goto out;
1196 	}
1197 
1198 	/* Fallback to software copy. */
1199 	spin_lock(&bo->bdev->fence_lock);
1200 	ret = ttm_bo_wait(bo, true, intr, no_wait_gpu);
1201 	spin_unlock(&bo->bdev->fence_lock);
1202 	if (ret == 0)
1203 		ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
1204 
1205 out:
1206 	if (nv_device(drm->device)->card_type < NV_50) {
1207 		if (ret)
1208 			nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1209 		else
1210 			nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1211 	}
1212 
1213 	return ret;
1214 }
1215 
1216 static int
1217 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1218 {
1219 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1220 
1221 	return drm_vma_node_verify_access(&nvbo->gem.vma_node, filp);
1222 }
1223 
1224 static int
1225 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1226 {
1227 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1228 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1229 	struct nouveau_mem *node = mem->mm_node;
1230 	struct drm_device *dev = drm->dev;
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 = !dev->agp->cant_use_aperture;
1250 		}
1251 #endif
1252 		if (nv_device(drm->device)->card_type < NV_50 || !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(nouveau_dev(dev), 1);
1259 		mem->bus.is_iomem = true;
1260 		if (nv_device(drm->device)->card_type >= NV_50) {
1261 			struct nouveau_bar *bar = nouveau_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 = nouveau_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 nouveau_device *device = nv_device(drm->device);
1296 	u32 mappable = nv_device_resource_len(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 (nv_device(drm->device)->card_type < NV_50 ||
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 (nv_device(drm->device)->card_type >= NV_50 ||
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 	unsigned i;
1337 	int r;
1338 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1339 
1340 	if (ttm->state != tt_unpopulated)
1341 		return 0;
1342 
1343 	if (slave && ttm->sg) {
1344 		/* make userspace faulting work */
1345 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1346 						 ttm_dma->dma_address, ttm->num_pages);
1347 		ttm->state = tt_unbound;
1348 		return 0;
1349 	}
1350 
1351 	drm = nouveau_bdev(ttm->bdev);
1352 	device = nv_device(drm->device);
1353 	dev = drm->dev;
1354 
1355 #if __OS_HAS_AGP
1356 	if (drm->agp.stat == ENABLED) {
1357 		return ttm_agp_tt_populate(ttm);
1358 	}
1359 #endif
1360 
1361 #ifdef CONFIG_SWIOTLB
1362 	if (swiotlb_nr_tbl()) {
1363 		return ttm_dma_populate((void *)ttm, dev->dev);
1364 	}
1365 #endif
1366 
1367 	r = ttm_pool_populate(ttm);
1368 	if (r) {
1369 		return r;
1370 	}
1371 
1372 	for (i = 0; i < ttm->num_pages; i++) {
1373 		ttm_dma->dma_address[i] = nv_device_map_page(device,
1374 							     ttm->pages[i]);
1375 		if (!ttm_dma->dma_address[i]) {
1376 			while (--i) {
1377 				nv_device_unmap_page(device,
1378 						     ttm_dma->dma_address[i]);
1379 				ttm_dma->dma_address[i] = 0;
1380 			}
1381 			ttm_pool_unpopulate(ttm);
1382 			return -EFAULT;
1383 		}
1384 	}
1385 	return 0;
1386 }
1387 
1388 static void
1389 nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
1390 {
1391 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1392 	struct nouveau_drm *drm;
1393 	struct nouveau_device *device;
1394 	struct drm_device *dev;
1395 	unsigned i;
1396 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1397 
1398 	if (slave)
1399 		return;
1400 
1401 	drm = nouveau_bdev(ttm->bdev);
1402 	device = nv_device(drm->device);
1403 	dev = drm->dev;
1404 
1405 #if __OS_HAS_AGP
1406 	if (drm->agp.stat == ENABLED) {
1407 		ttm_agp_tt_unpopulate(ttm);
1408 		return;
1409 	}
1410 #endif
1411 
1412 #ifdef CONFIG_SWIOTLB
1413 	if (swiotlb_nr_tbl()) {
1414 		ttm_dma_unpopulate((void *)ttm, dev->dev);
1415 		return;
1416 	}
1417 #endif
1418 
1419 	for (i = 0; i < ttm->num_pages; i++) {
1420 		if (ttm_dma->dma_address[i]) {
1421 			nv_device_unmap_page(device, ttm_dma->dma_address[i]);
1422 		}
1423 	}
1424 
1425 	ttm_pool_unpopulate(ttm);
1426 }
1427 
1428 void
1429 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
1430 {
1431 	struct nouveau_fence *new_fence = nouveau_fence_ref(fence);
1432 	struct nouveau_fence *old_fence = NULL;
1433 
1434 	spin_lock(&nvbo->bo.bdev->fence_lock);
1435 	old_fence = nvbo->bo.sync_obj;
1436 	nvbo->bo.sync_obj = new_fence;
1437 	spin_unlock(&nvbo->bo.bdev->fence_lock);
1438 
1439 	nouveau_fence_unref(&old_fence);
1440 }
1441 
1442 static void
1443 nouveau_bo_fence_unref(void **sync_obj)
1444 {
1445 	nouveau_fence_unref((struct nouveau_fence **)sync_obj);
1446 }
1447 
1448 static void *
1449 nouveau_bo_fence_ref(void *sync_obj)
1450 {
1451 	return nouveau_fence_ref(sync_obj);
1452 }
1453 
1454 static bool
1455 nouveau_bo_fence_signalled(void *sync_obj)
1456 {
1457 	return nouveau_fence_done(sync_obj);
1458 }
1459 
1460 static int
1461 nouveau_bo_fence_wait(void *sync_obj, bool lazy, bool intr)
1462 {
1463 	return nouveau_fence_wait(sync_obj, lazy, intr);
1464 }
1465 
1466 static int
1467 nouveau_bo_fence_flush(void *sync_obj)
1468 {
1469 	return 0;
1470 }
1471 
1472 struct ttm_bo_driver nouveau_bo_driver = {
1473 	.ttm_tt_create = &nouveau_ttm_tt_create,
1474 	.ttm_tt_populate = &nouveau_ttm_tt_populate,
1475 	.ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1476 	.invalidate_caches = nouveau_bo_invalidate_caches,
1477 	.init_mem_type = nouveau_bo_init_mem_type,
1478 	.evict_flags = nouveau_bo_evict_flags,
1479 	.move_notify = nouveau_bo_move_ntfy,
1480 	.move = nouveau_bo_move,
1481 	.verify_access = nouveau_bo_verify_access,
1482 	.sync_obj_signaled = nouveau_bo_fence_signalled,
1483 	.sync_obj_wait = nouveau_bo_fence_wait,
1484 	.sync_obj_flush = nouveau_bo_fence_flush,
1485 	.sync_obj_unref = nouveau_bo_fence_unref,
1486 	.sync_obj_ref = nouveau_bo_fence_ref,
1487 	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1488 	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1489 	.io_mem_free = &nouveau_ttm_io_mem_free,
1490 };
1491 
1492 struct nouveau_vma *
1493 nouveau_bo_vma_find(struct nouveau_bo *nvbo, struct nouveau_vm *vm)
1494 {
1495 	struct nouveau_vma *vma;
1496 	list_for_each_entry(vma, &nvbo->vma_list, head) {
1497 		if (vma->vm == vm)
1498 			return vma;
1499 	}
1500 
1501 	return NULL;
1502 }
1503 
1504 int
1505 nouveau_bo_vma_add(struct nouveau_bo *nvbo, struct nouveau_vm *vm,
1506 		   struct nouveau_vma *vma)
1507 {
1508 	const u32 size = nvbo->bo.mem.num_pages << PAGE_SHIFT;
1509 	int ret;
1510 
1511 	ret = nouveau_vm_get(vm, size, nvbo->page_shift,
1512 			     NV_MEM_ACCESS_RW, vma);
1513 	if (ret)
1514 		return ret;
1515 
1516 	if ( nvbo->bo.mem.mem_type != TTM_PL_SYSTEM &&
1517 	    (nvbo->bo.mem.mem_type == TTM_PL_VRAM ||
1518 	     nvbo->page_shift != vma->vm->vmm->lpg_shift))
1519 		nouveau_vm_map(vma, nvbo->bo.mem.mm_node);
1520 
1521 	list_add_tail(&vma->head, &nvbo->vma_list);
1522 	vma->refcount = 1;
1523 	return 0;
1524 }
1525 
1526 void
1527 nouveau_bo_vma_del(struct nouveau_bo *nvbo, struct nouveau_vma *vma)
1528 {
1529 	if (vma->node) {
1530 		if (nvbo->bo.mem.mem_type != TTM_PL_SYSTEM)
1531 			nouveau_vm_unmap(vma);
1532 		nouveau_vm_put(vma);
1533 		list_del(&vma->head);
1534 	}
1535 }
1536