1 /*
2  * Copyright 2007 Dave Airlied
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * VA LINUX SYSTEMS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  */
24 /*
25  * Authors: Dave Airlied <airlied@linux.ie>
26  *	    Ben Skeggs   <darktama@iinet.net.au>
27  *	    Jeremy Kolb  <jkolb@brandeis.edu>
28  */
29 
30 #include <linux/dma-mapping.h>
31 #include <linux/swiotlb.h>
32 
33 #include "nouveau_drv.h"
34 #include "nouveau_dma.h"
35 #include "nouveau_fence.h"
36 
37 #include "nouveau_bo.h"
38 #include "nouveau_ttm.h"
39 #include "nouveau_gem.h"
40 #include "nouveau_mem.h"
41 #include "nouveau_vmm.h"
42 
43 #include <nvif/class.h>
44 #include <nvif/if500b.h>
45 #include <nvif/if900b.h>
46 
47 /*
48  * NV10-NV40 tiling helpers
49  */
50 
51 static void
52 nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
53 			   u32 addr, u32 size, u32 pitch, u32 flags)
54 {
55 	struct nouveau_drm *drm = nouveau_drm(dev);
56 	int i = reg - drm->tile.reg;
57 	struct nvkm_fb *fb = nvxx_fb(&drm->client.device);
58 	struct nvkm_fb_tile *tile = &fb->tile.region[i];
59 
60 	nouveau_fence_unref(&reg->fence);
61 
62 	if (tile->pitch)
63 		nvkm_fb_tile_fini(fb, i, tile);
64 
65 	if (pitch)
66 		nvkm_fb_tile_init(fb, i, addr, size, pitch, flags, tile);
67 
68 	nvkm_fb_tile_prog(fb, i, tile);
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 dma_fence *fence)
92 {
93 	struct nouveau_drm *drm = nouveau_drm(dev);
94 
95 	if (tile) {
96 		spin_lock(&drm->tile.lock);
97 		tile->fence = (struct nouveau_fence *)dma_fence_get(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 zeta)
106 {
107 	struct nouveau_drm *drm = nouveau_drm(dev);
108 	struct nvkm_fb *fb = nvxx_fb(&drm->client.device);
109 	struct nouveau_drm_tile *tile, *found = NULL;
110 	int i;
111 
112 	for (i = 0; i < fb->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 && fb->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, pitch, zeta);
129 	return found;
130 }
131 
132 static void
133 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
134 {
135 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
136 	struct drm_device *dev = drm->dev;
137 	struct nouveau_bo *nvbo = nouveau_bo(bo);
138 
139 	if (unlikely(nvbo->gem.filp))
140 		DRM_ERROR("bo %p still attached to GEM object\n", bo);
141 	WARN_ON(nvbo->pin_refcnt > 0);
142 	nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
143 	kfree(nvbo);
144 }
145 
146 static inline u64
147 roundup_64(u64 x, u32 y)
148 {
149 	x += y - 1;
150 	do_div(x, y);
151 	return x * y;
152 }
153 
154 static void
155 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
156 		       int *align, u64 *size)
157 {
158 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
159 	struct nvif_device *device = &drm->client.device;
160 
161 	if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
162 		if (nvbo->mode) {
163 			if (device->info.chipset >= 0x40) {
164 				*align = 65536;
165 				*size = roundup_64(*size, 64 * nvbo->mode);
166 
167 			} else if (device->info.chipset >= 0x30) {
168 				*align = 32768;
169 				*size = roundup_64(*size, 64 * nvbo->mode);
170 
171 			} else if (device->info.chipset >= 0x20) {
172 				*align = 16384;
173 				*size = roundup_64(*size, 64 * nvbo->mode);
174 
175 			} else if (device->info.chipset >= 0x10) {
176 				*align = 16384;
177 				*size = roundup_64(*size, 32 * nvbo->mode);
178 			}
179 		}
180 	} else {
181 		*size = roundup_64(*size, (1 << nvbo->page));
182 		*align = max((1 <<  nvbo->page), *align);
183 	}
184 
185 	*size = roundup_64(*size, PAGE_SIZE);
186 }
187 
188 int
189 nouveau_bo_new(struct nouveau_cli *cli, u64 size, int align,
190 	       uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
191 	       struct sg_table *sg, struct reservation_object *robj,
192 	       struct nouveau_bo **pnvbo)
193 {
194 	struct nouveau_drm *drm = cli->drm;
195 	struct nouveau_bo *nvbo;
196 	struct nvif_mmu *mmu = &cli->mmu;
197 	struct nvif_vmm *vmm = &cli->vmm.vmm;
198 	size_t acc_size;
199 	int type = ttm_bo_type_device;
200 	int ret, i, pi = -1;
201 
202 	if (!size) {
203 		NV_WARN(drm, "skipped size %016llx\n", size);
204 		return -EINVAL;
205 	}
206 
207 	if (sg)
208 		type = ttm_bo_type_sg;
209 
210 	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
211 	if (!nvbo)
212 		return -ENOMEM;
213 	INIT_LIST_HEAD(&nvbo->head);
214 	INIT_LIST_HEAD(&nvbo->entry);
215 	INIT_LIST_HEAD(&nvbo->vma_list);
216 	nvbo->bo.bdev = &drm->ttm.bdev;
217 	nvbo->cli = cli;
218 
219 	/* This is confusing, and doesn't actually mean we want an uncached
220 	 * mapping, but is what NOUVEAU_GEM_DOMAIN_COHERENT gets translated
221 	 * into in nouveau_gem_new().
222 	 */
223 	if (flags & TTM_PL_FLAG_UNCACHED) {
224 		/* Determine if we can get a cache-coherent map, forcing
225 		 * uncached mapping if we can't.
226 		 */
227 		if (mmu->type[drm->ttm.type_host].type & NVIF_MEM_UNCACHED)
228 			nvbo->force_coherent = true;
229 	}
230 
231 	if (cli->device.info.family >= NV_DEVICE_INFO_V0_FERMI) {
232 		nvbo->kind = (tile_flags & 0x0000ff00) >> 8;
233 		if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
234 			kfree(nvbo);
235 			return -EINVAL;
236 		}
237 
238 		nvbo->comp = mmu->kind[nvbo->kind] != nvbo->kind;
239 	} else
240 	if (cli->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
241 		nvbo->kind = (tile_flags & 0x00007f00) >> 8;
242 		nvbo->comp = (tile_flags & 0x00030000) >> 16;
243 		if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
244 			kfree(nvbo);
245 			return -EINVAL;
246 		}
247 	} else {
248 		nvbo->zeta = (tile_flags & 0x00000007);
249 	}
250 	nvbo->mode = tile_mode;
251 	nvbo->contig = !(tile_flags & NOUVEAU_GEM_TILE_NONCONTIG);
252 
253 	/* Determine the desirable target GPU page size for the buffer. */
254 	for (i = 0; i < vmm->page_nr; i++) {
255 		/* Because we cannot currently allow VMM maps to fail
256 		 * during buffer migration, we need to determine page
257 		 * size for the buffer up-front, and pre-allocate its
258 		 * page tables.
259 		 *
260 		 * Skip page sizes that can't support needed domains.
261 		 */
262 		if (cli->device.info.family > NV_DEVICE_INFO_V0_CURIE &&
263 		    (flags & TTM_PL_FLAG_VRAM) && !vmm->page[i].vram)
264 			continue;
265 		if ((flags & TTM_PL_FLAG_TT  ) && !vmm->page[i].host)
266 			continue;
267 
268 		/* Select this page size if it's the first that supports
269 		 * the potential memory domains, or when it's compatible
270 		 * with the requested compression settings.
271 		 */
272 		if (pi < 0 || !nvbo->comp || vmm->page[i].comp)
273 			pi = i;
274 
275 		/* Stop once the buffer is larger than the current page size. */
276 		if (size >= 1ULL << vmm->page[i].shift)
277 			break;
278 	}
279 
280 	if (WARN_ON(pi < 0))
281 		return -EINVAL;
282 
283 	/* Disable compression if suitable settings couldn't be found. */
284 	if (nvbo->comp && !vmm->page[pi].comp) {
285 		if (mmu->object.oclass >= NVIF_CLASS_MMU_GF100)
286 			nvbo->kind = mmu->kind[nvbo->kind];
287 		nvbo->comp = 0;
288 	}
289 	nvbo->page = vmm->page[pi].shift;
290 
291 	nouveau_bo_fixup_align(nvbo, flags, &align, &size);
292 	nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
293 	nouveau_bo_placement_set(nvbo, flags, 0);
294 
295 	acc_size = ttm_bo_dma_acc_size(&drm->ttm.bdev, size,
296 				       sizeof(struct nouveau_bo));
297 
298 	ret = ttm_bo_init(&drm->ttm.bdev, &nvbo->bo, size,
299 			  type, &nvbo->placement,
300 			  align >> PAGE_SHIFT, false, NULL, acc_size, sg,
301 			  robj, nouveau_bo_del_ttm);
302 	if (ret) {
303 		/* ttm will call nouveau_bo_del_ttm if it fails.. */
304 		return ret;
305 	}
306 
307 	*pnvbo = nvbo;
308 	return 0;
309 }
310 
311 static void
312 set_placement_list(struct ttm_place *pl, unsigned *n, uint32_t type, uint32_t flags)
313 {
314 	*n = 0;
315 
316 	if (type & TTM_PL_FLAG_VRAM)
317 		pl[(*n)++].flags = TTM_PL_FLAG_VRAM | flags;
318 	if (type & TTM_PL_FLAG_TT)
319 		pl[(*n)++].flags = TTM_PL_FLAG_TT | flags;
320 	if (type & TTM_PL_FLAG_SYSTEM)
321 		pl[(*n)++].flags = TTM_PL_FLAG_SYSTEM | flags;
322 }
323 
324 static void
325 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
326 {
327 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
328 	u32 vram_pages = drm->client.device.info.ram_size >> PAGE_SHIFT;
329 	unsigned i, fpfn, lpfn;
330 
331 	if (drm->client.device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
332 	    nvbo->mode && (type & TTM_PL_FLAG_VRAM) &&
333 	    nvbo->bo.mem.num_pages < vram_pages / 4) {
334 		/*
335 		 * Make sure that the color and depth buffers are handled
336 		 * by independent memory controller units. Up to a 9x
337 		 * speed up when alpha-blending and depth-test are enabled
338 		 * at the same time.
339 		 */
340 		if (nvbo->zeta) {
341 			fpfn = vram_pages / 2;
342 			lpfn = ~0;
343 		} else {
344 			fpfn = 0;
345 			lpfn = vram_pages / 2;
346 		}
347 		for (i = 0; i < nvbo->placement.num_placement; ++i) {
348 			nvbo->placements[i].fpfn = fpfn;
349 			nvbo->placements[i].lpfn = lpfn;
350 		}
351 		for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
352 			nvbo->busy_placements[i].fpfn = fpfn;
353 			nvbo->busy_placements[i].lpfn = lpfn;
354 		}
355 	}
356 }
357 
358 void
359 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
360 {
361 	struct ttm_placement *pl = &nvbo->placement;
362 	uint32_t flags = (nvbo->force_coherent ? TTM_PL_FLAG_UNCACHED :
363 						 TTM_PL_MASK_CACHING) |
364 			 (nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
365 
366 	pl->placement = nvbo->placements;
367 	set_placement_list(nvbo->placements, &pl->num_placement,
368 			   type, flags);
369 
370 	pl->busy_placement = nvbo->busy_placements;
371 	set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
372 			   type | busy, flags);
373 
374 	set_placement_range(nvbo, type);
375 }
376 
377 int
378 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype, bool contig)
379 {
380 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
381 	struct ttm_buffer_object *bo = &nvbo->bo;
382 	bool force = false, evict = false;
383 	int ret;
384 
385 	ret = ttm_bo_reserve(bo, false, false, NULL);
386 	if (ret)
387 		return ret;
388 
389 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
390 	    memtype == TTM_PL_FLAG_VRAM && contig) {
391 		if (!nvbo->contig) {
392 			nvbo->contig = true;
393 			force = true;
394 			evict = true;
395 		}
396 	}
397 
398 	if (nvbo->pin_refcnt) {
399 		if (!(memtype & (1 << bo->mem.mem_type)) || evict) {
400 			NV_ERROR(drm, "bo %p pinned elsewhere: "
401 				      "0x%08x vs 0x%08x\n", bo,
402 				 1 << bo->mem.mem_type, memtype);
403 			ret = -EBUSY;
404 		}
405 		nvbo->pin_refcnt++;
406 		goto out;
407 	}
408 
409 	if (evict) {
410 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT, 0);
411 		ret = nouveau_bo_validate(nvbo, false, false);
412 		if (ret)
413 			goto out;
414 	}
415 
416 	nvbo->pin_refcnt++;
417 	nouveau_bo_placement_set(nvbo, memtype, 0);
418 
419 	/* drop pin_refcnt temporarily, so we don't trip the assertion
420 	 * in nouveau_bo_move() that makes sure we're not trying to
421 	 * move a pinned buffer
422 	 */
423 	nvbo->pin_refcnt--;
424 	ret = nouveau_bo_validate(nvbo, false, false);
425 	if (ret)
426 		goto out;
427 	nvbo->pin_refcnt++;
428 
429 	switch (bo->mem.mem_type) {
430 	case TTM_PL_VRAM:
431 		drm->gem.vram_available -= bo->mem.size;
432 		break;
433 	case TTM_PL_TT:
434 		drm->gem.gart_available -= bo->mem.size;
435 		break;
436 	default:
437 		break;
438 	}
439 
440 out:
441 	if (force && ret)
442 		nvbo->contig = false;
443 	ttm_bo_unreserve(bo);
444 	return ret;
445 }
446 
447 int
448 nouveau_bo_unpin(struct nouveau_bo *nvbo)
449 {
450 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
451 	struct ttm_buffer_object *bo = &nvbo->bo;
452 	int ret, ref;
453 
454 	ret = ttm_bo_reserve(bo, false, false, NULL);
455 	if (ret)
456 		return ret;
457 
458 	ref = --nvbo->pin_refcnt;
459 	WARN_ON_ONCE(ref < 0);
460 	if (ref)
461 		goto out;
462 
463 	nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
464 
465 	ret = nouveau_bo_validate(nvbo, false, false);
466 	if (ret == 0) {
467 		switch (bo->mem.mem_type) {
468 		case TTM_PL_VRAM:
469 			drm->gem.vram_available += bo->mem.size;
470 			break;
471 		case TTM_PL_TT:
472 			drm->gem.gart_available += bo->mem.size;
473 			break;
474 		default:
475 			break;
476 		}
477 	}
478 
479 out:
480 	ttm_bo_unreserve(bo);
481 	return ret;
482 }
483 
484 int
485 nouveau_bo_map(struct nouveau_bo *nvbo)
486 {
487 	int ret;
488 
489 	ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
490 	if (ret)
491 		return ret;
492 
493 	ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
494 
495 	ttm_bo_unreserve(&nvbo->bo);
496 	return ret;
497 }
498 
499 void
500 nouveau_bo_unmap(struct nouveau_bo *nvbo)
501 {
502 	if (!nvbo)
503 		return;
504 
505 	ttm_bo_kunmap(&nvbo->kmap);
506 }
507 
508 void
509 nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
510 {
511 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
512 	struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
513 	int i;
514 
515 	if (!ttm_dma)
516 		return;
517 
518 	/* Don't waste time looping if the object is coherent */
519 	if (nvbo->force_coherent)
520 		return;
521 
522 	for (i = 0; i < ttm_dma->ttm.num_pages; i++)
523 		dma_sync_single_for_device(drm->dev->dev,
524 					   ttm_dma->dma_address[i],
525 					   PAGE_SIZE, DMA_TO_DEVICE);
526 }
527 
528 void
529 nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
530 {
531 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
532 	struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
533 	int i;
534 
535 	if (!ttm_dma)
536 		return;
537 
538 	/* Don't waste time looping if the object is coherent */
539 	if (nvbo->force_coherent)
540 		return;
541 
542 	for (i = 0; i < ttm_dma->ttm.num_pages; i++)
543 		dma_sync_single_for_cpu(drm->dev->dev, ttm_dma->dma_address[i],
544 					PAGE_SIZE, DMA_FROM_DEVICE);
545 }
546 
547 int
548 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
549 		    bool no_wait_gpu)
550 {
551 	int ret;
552 
553 	ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement,
554 			      interruptible, no_wait_gpu);
555 	if (ret)
556 		return ret;
557 
558 	nouveau_bo_sync_for_device(nvbo);
559 
560 	return 0;
561 }
562 
563 void
564 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
565 {
566 	bool is_iomem;
567 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
568 
569 	mem += index;
570 
571 	if (is_iomem)
572 		iowrite16_native(val, (void __force __iomem *)mem);
573 	else
574 		*mem = val;
575 }
576 
577 u32
578 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
579 {
580 	bool is_iomem;
581 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
582 
583 	mem += index;
584 
585 	if (is_iomem)
586 		return ioread32_native((void __force __iomem *)mem);
587 	else
588 		return *mem;
589 }
590 
591 void
592 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
593 {
594 	bool is_iomem;
595 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
596 
597 	mem += index;
598 
599 	if (is_iomem)
600 		iowrite32_native(val, (void __force __iomem *)mem);
601 	else
602 		*mem = val;
603 }
604 
605 static struct ttm_tt *
606 nouveau_ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
607 		      uint32_t page_flags, struct page *dummy_read)
608 {
609 #if IS_ENABLED(CONFIG_AGP)
610 	struct nouveau_drm *drm = nouveau_bdev(bdev);
611 
612 	if (drm->agp.bridge) {
613 		return ttm_agp_tt_create(bdev, drm->agp.bridge, size,
614 					 page_flags, dummy_read);
615 	}
616 #endif
617 
618 	return nouveau_sgdma_create_ttm(bdev, size, page_flags, dummy_read);
619 }
620 
621 static int
622 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
623 {
624 	/* We'll do this from user space. */
625 	return 0;
626 }
627 
628 static int
629 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
630 			 struct ttm_mem_type_manager *man)
631 {
632 	struct nouveau_drm *drm = nouveau_bdev(bdev);
633 	struct nvif_mmu *mmu = &drm->client.mmu;
634 
635 	switch (type) {
636 	case TTM_PL_SYSTEM:
637 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
638 		man->available_caching = TTM_PL_MASK_CACHING;
639 		man->default_caching = TTM_PL_FLAG_CACHED;
640 		break;
641 	case TTM_PL_VRAM:
642 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
643 			     TTM_MEMTYPE_FLAG_MAPPABLE;
644 		man->available_caching = TTM_PL_FLAG_UNCACHED |
645 					 TTM_PL_FLAG_WC;
646 		man->default_caching = TTM_PL_FLAG_WC;
647 
648 		if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
649 			/* Some BARs do not support being ioremapped WC */
650 			const u8 type = mmu->type[drm->ttm.type_vram].type;
651 			if (type & NVIF_MEM_UNCACHED) {
652 				man->available_caching = TTM_PL_FLAG_UNCACHED;
653 				man->default_caching = TTM_PL_FLAG_UNCACHED;
654 			}
655 
656 			man->func = &nouveau_vram_manager;
657 			man->io_reserve_fastpath = false;
658 			man->use_io_reserve_lru = true;
659 		} else {
660 			man->func = &ttm_bo_manager_func;
661 		}
662 		break;
663 	case TTM_PL_TT:
664 		if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA)
665 			man->func = &nouveau_gart_manager;
666 		else
667 		if (!drm->agp.bridge)
668 			man->func = &nv04_gart_manager;
669 		else
670 			man->func = &ttm_bo_manager_func;
671 
672 		if (drm->agp.bridge) {
673 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
674 			man->available_caching = TTM_PL_FLAG_UNCACHED |
675 				TTM_PL_FLAG_WC;
676 			man->default_caching = TTM_PL_FLAG_WC;
677 		} else {
678 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
679 				     TTM_MEMTYPE_FLAG_CMA;
680 			man->available_caching = TTM_PL_MASK_CACHING;
681 			man->default_caching = TTM_PL_FLAG_CACHED;
682 		}
683 
684 		break;
685 	default:
686 		return -EINVAL;
687 	}
688 	return 0;
689 }
690 
691 static void
692 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
693 {
694 	struct nouveau_bo *nvbo = nouveau_bo(bo);
695 
696 	switch (bo->mem.mem_type) {
697 	case TTM_PL_VRAM:
698 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
699 					 TTM_PL_FLAG_SYSTEM);
700 		break;
701 	default:
702 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
703 		break;
704 	}
705 
706 	*pl = nvbo->placement;
707 }
708 
709 
710 static int
711 nve0_bo_move_init(struct nouveau_channel *chan, u32 handle)
712 {
713 	int ret = RING_SPACE(chan, 2);
714 	if (ret == 0) {
715 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
716 		OUT_RING  (chan, handle & 0x0000ffff);
717 		FIRE_RING (chan);
718 	}
719 	return ret;
720 }
721 
722 static int
723 nve0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
724 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
725 {
726 	struct nouveau_mem *mem = nouveau_mem(old_reg);
727 	int ret = RING_SPACE(chan, 10);
728 	if (ret == 0) {
729 		BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
730 		OUT_RING  (chan, upper_32_bits(mem->vma[0].addr));
731 		OUT_RING  (chan, lower_32_bits(mem->vma[0].addr));
732 		OUT_RING  (chan, upper_32_bits(mem->vma[1].addr));
733 		OUT_RING  (chan, lower_32_bits(mem->vma[1].addr));
734 		OUT_RING  (chan, PAGE_SIZE);
735 		OUT_RING  (chan, PAGE_SIZE);
736 		OUT_RING  (chan, PAGE_SIZE);
737 		OUT_RING  (chan, new_reg->num_pages);
738 		BEGIN_IMC0(chan, NvSubCopy, 0x0300, 0x0386);
739 	}
740 	return ret;
741 }
742 
743 static int
744 nvc0_bo_move_init(struct nouveau_channel *chan, u32 handle)
745 {
746 	int ret = RING_SPACE(chan, 2);
747 	if (ret == 0) {
748 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
749 		OUT_RING  (chan, handle);
750 	}
751 	return ret;
752 }
753 
754 static int
755 nvc0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
756 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
757 {
758 	struct nouveau_mem *mem = nouveau_mem(old_reg);
759 	u64 src_offset = mem->vma[0].addr;
760 	u64 dst_offset = mem->vma[1].addr;
761 	u32 page_count = new_reg->num_pages;
762 	int ret;
763 
764 	page_count = new_reg->num_pages;
765 	while (page_count) {
766 		int line_count = (page_count > 8191) ? 8191 : page_count;
767 
768 		ret = RING_SPACE(chan, 11);
769 		if (ret)
770 			return ret;
771 
772 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 8);
773 		OUT_RING  (chan, upper_32_bits(src_offset));
774 		OUT_RING  (chan, lower_32_bits(src_offset));
775 		OUT_RING  (chan, upper_32_bits(dst_offset));
776 		OUT_RING  (chan, lower_32_bits(dst_offset));
777 		OUT_RING  (chan, PAGE_SIZE);
778 		OUT_RING  (chan, PAGE_SIZE);
779 		OUT_RING  (chan, PAGE_SIZE);
780 		OUT_RING  (chan, line_count);
781 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
782 		OUT_RING  (chan, 0x00000110);
783 
784 		page_count -= line_count;
785 		src_offset += (PAGE_SIZE * line_count);
786 		dst_offset += (PAGE_SIZE * line_count);
787 	}
788 
789 	return 0;
790 }
791 
792 static int
793 nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
794 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
795 {
796 	struct nouveau_mem *mem = nouveau_mem(old_reg);
797 	u64 src_offset = mem->vma[0].addr;
798 	u64 dst_offset = mem->vma[1].addr;
799 	u32 page_count = new_reg->num_pages;
800 	int ret;
801 
802 	page_count = new_reg->num_pages;
803 	while (page_count) {
804 		int line_count = (page_count > 2047) ? 2047 : page_count;
805 
806 		ret = RING_SPACE(chan, 12);
807 		if (ret)
808 			return ret;
809 
810 		BEGIN_NVC0(chan, NvSubCopy, 0x0238, 2);
811 		OUT_RING  (chan, upper_32_bits(dst_offset));
812 		OUT_RING  (chan, lower_32_bits(dst_offset));
813 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 6);
814 		OUT_RING  (chan, upper_32_bits(src_offset));
815 		OUT_RING  (chan, lower_32_bits(src_offset));
816 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
817 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
818 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
819 		OUT_RING  (chan, line_count);
820 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
821 		OUT_RING  (chan, 0x00100110);
822 
823 		page_count -= line_count;
824 		src_offset += (PAGE_SIZE * line_count);
825 		dst_offset += (PAGE_SIZE * line_count);
826 	}
827 
828 	return 0;
829 }
830 
831 static int
832 nva3_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
833 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
834 {
835 	struct nouveau_mem *mem = nouveau_mem(old_reg);
836 	u64 src_offset = mem->vma[0].addr;
837 	u64 dst_offset = mem->vma[1].addr;
838 	u32 page_count = new_reg->num_pages;
839 	int ret;
840 
841 	page_count = new_reg->num_pages;
842 	while (page_count) {
843 		int line_count = (page_count > 8191) ? 8191 : page_count;
844 
845 		ret = RING_SPACE(chan, 11);
846 		if (ret)
847 			return ret;
848 
849 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
850 		OUT_RING  (chan, upper_32_bits(src_offset));
851 		OUT_RING  (chan, lower_32_bits(src_offset));
852 		OUT_RING  (chan, upper_32_bits(dst_offset));
853 		OUT_RING  (chan, lower_32_bits(dst_offset));
854 		OUT_RING  (chan, PAGE_SIZE);
855 		OUT_RING  (chan, PAGE_SIZE);
856 		OUT_RING  (chan, PAGE_SIZE);
857 		OUT_RING  (chan, line_count);
858 		BEGIN_NV04(chan, NvSubCopy, 0x0300, 1);
859 		OUT_RING  (chan, 0x00000110);
860 
861 		page_count -= line_count;
862 		src_offset += (PAGE_SIZE * line_count);
863 		dst_offset += (PAGE_SIZE * line_count);
864 	}
865 
866 	return 0;
867 }
868 
869 static int
870 nv98_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
871 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
872 {
873 	struct nouveau_mem *mem = nouveau_mem(old_reg);
874 	int ret = RING_SPACE(chan, 7);
875 	if (ret == 0) {
876 		BEGIN_NV04(chan, NvSubCopy, 0x0320, 6);
877 		OUT_RING  (chan, upper_32_bits(mem->vma[0].addr));
878 		OUT_RING  (chan, lower_32_bits(mem->vma[0].addr));
879 		OUT_RING  (chan, upper_32_bits(mem->vma[1].addr));
880 		OUT_RING  (chan, lower_32_bits(mem->vma[1].addr));
881 		OUT_RING  (chan, 0x00000000 /* COPY */);
882 		OUT_RING  (chan, new_reg->num_pages << PAGE_SHIFT);
883 	}
884 	return ret;
885 }
886 
887 static int
888 nv84_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
889 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
890 {
891 	struct nouveau_mem *mem = nouveau_mem(old_reg);
892 	int ret = RING_SPACE(chan, 7);
893 	if (ret == 0) {
894 		BEGIN_NV04(chan, NvSubCopy, 0x0304, 6);
895 		OUT_RING  (chan, new_reg->num_pages << PAGE_SHIFT);
896 		OUT_RING  (chan, upper_32_bits(mem->vma[0].addr));
897 		OUT_RING  (chan, lower_32_bits(mem->vma[0].addr));
898 		OUT_RING  (chan, upper_32_bits(mem->vma[1].addr));
899 		OUT_RING  (chan, lower_32_bits(mem->vma[1].addr));
900 		OUT_RING  (chan, 0x00000000 /* MODE_COPY, QUERY_NONE */);
901 	}
902 	return ret;
903 }
904 
905 static int
906 nv50_bo_move_init(struct nouveau_channel *chan, u32 handle)
907 {
908 	int ret = RING_SPACE(chan, 6);
909 	if (ret == 0) {
910 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
911 		OUT_RING  (chan, handle);
912 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 3);
913 		OUT_RING  (chan, chan->drm->ntfy.handle);
914 		OUT_RING  (chan, chan->vram.handle);
915 		OUT_RING  (chan, chan->vram.handle);
916 	}
917 
918 	return ret;
919 }
920 
921 static int
922 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
923 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
924 {
925 	struct nouveau_mem *mem = nouveau_mem(old_reg);
926 	u64 length = (new_reg->num_pages << PAGE_SHIFT);
927 	u64 src_offset = mem->vma[0].addr;
928 	u64 dst_offset = mem->vma[1].addr;
929 	int src_tiled = !!mem->kind;
930 	int dst_tiled = !!nouveau_mem(new_reg)->kind;
931 	int ret;
932 
933 	while (length) {
934 		u32 amount, stride, height;
935 
936 		ret = RING_SPACE(chan, 18 + 6 * (src_tiled + dst_tiled));
937 		if (ret)
938 			return ret;
939 
940 		amount  = min(length, (u64)(4 * 1024 * 1024));
941 		stride  = 16 * 4;
942 		height  = amount / stride;
943 
944 		if (src_tiled) {
945 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 7);
946 			OUT_RING  (chan, 0);
947 			OUT_RING  (chan, 0);
948 			OUT_RING  (chan, stride);
949 			OUT_RING  (chan, height);
950 			OUT_RING  (chan, 1);
951 			OUT_RING  (chan, 0);
952 			OUT_RING  (chan, 0);
953 		} else {
954 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 1);
955 			OUT_RING  (chan, 1);
956 		}
957 		if (dst_tiled) {
958 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 7);
959 			OUT_RING  (chan, 0);
960 			OUT_RING  (chan, 0);
961 			OUT_RING  (chan, stride);
962 			OUT_RING  (chan, height);
963 			OUT_RING  (chan, 1);
964 			OUT_RING  (chan, 0);
965 			OUT_RING  (chan, 0);
966 		} else {
967 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 1);
968 			OUT_RING  (chan, 1);
969 		}
970 
971 		BEGIN_NV04(chan, NvSubCopy, 0x0238, 2);
972 		OUT_RING  (chan, upper_32_bits(src_offset));
973 		OUT_RING  (chan, upper_32_bits(dst_offset));
974 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
975 		OUT_RING  (chan, lower_32_bits(src_offset));
976 		OUT_RING  (chan, lower_32_bits(dst_offset));
977 		OUT_RING  (chan, stride);
978 		OUT_RING  (chan, stride);
979 		OUT_RING  (chan, stride);
980 		OUT_RING  (chan, height);
981 		OUT_RING  (chan, 0x00000101);
982 		OUT_RING  (chan, 0x00000000);
983 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
984 		OUT_RING  (chan, 0);
985 
986 		length -= amount;
987 		src_offset += amount;
988 		dst_offset += amount;
989 	}
990 
991 	return 0;
992 }
993 
994 static int
995 nv04_bo_move_init(struct nouveau_channel *chan, u32 handle)
996 {
997 	int ret = RING_SPACE(chan, 4);
998 	if (ret == 0) {
999 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
1000 		OUT_RING  (chan, handle);
1001 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 1);
1002 		OUT_RING  (chan, chan->drm->ntfy.handle);
1003 	}
1004 
1005 	return ret;
1006 }
1007 
1008 static inline uint32_t
1009 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
1010 		      struct nouveau_channel *chan, struct ttm_mem_reg *reg)
1011 {
1012 	if (reg->mem_type == TTM_PL_TT)
1013 		return NvDmaTT;
1014 	return chan->vram.handle;
1015 }
1016 
1017 static int
1018 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
1019 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
1020 {
1021 	u32 src_offset = old_reg->start << PAGE_SHIFT;
1022 	u32 dst_offset = new_reg->start << PAGE_SHIFT;
1023 	u32 page_count = new_reg->num_pages;
1024 	int ret;
1025 
1026 	ret = RING_SPACE(chan, 3);
1027 	if (ret)
1028 		return ret;
1029 
1030 	BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
1031 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_reg));
1032 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_reg));
1033 
1034 	page_count = new_reg->num_pages;
1035 	while (page_count) {
1036 		int line_count = (page_count > 2047) ? 2047 : page_count;
1037 
1038 		ret = RING_SPACE(chan, 11);
1039 		if (ret)
1040 			return ret;
1041 
1042 		BEGIN_NV04(chan, NvSubCopy,
1043 				 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
1044 		OUT_RING  (chan, src_offset);
1045 		OUT_RING  (chan, dst_offset);
1046 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
1047 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
1048 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
1049 		OUT_RING  (chan, line_count);
1050 		OUT_RING  (chan, 0x00000101);
1051 		OUT_RING  (chan, 0x00000000);
1052 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
1053 		OUT_RING  (chan, 0);
1054 
1055 		page_count -= line_count;
1056 		src_offset += (PAGE_SIZE * line_count);
1057 		dst_offset += (PAGE_SIZE * line_count);
1058 	}
1059 
1060 	return 0;
1061 }
1062 
1063 static int
1064 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
1065 		     struct ttm_mem_reg *reg)
1066 {
1067 	struct nouveau_mem *old_mem = nouveau_mem(&bo->mem);
1068 	struct nouveau_mem *new_mem = nouveau_mem(reg);
1069 	struct nvif_vmm *vmm = &drm->client.vmm.vmm;
1070 	int ret;
1071 
1072 	ret = nvif_vmm_get(vmm, LAZY, false, old_mem->mem.page, 0,
1073 			   old_mem->mem.size, &old_mem->vma[0]);
1074 	if (ret)
1075 		return ret;
1076 
1077 	ret = nvif_vmm_get(vmm, LAZY, false, new_mem->mem.page, 0,
1078 			   new_mem->mem.size, &old_mem->vma[1]);
1079 	if (ret)
1080 		goto done;
1081 
1082 	ret = nouveau_mem_map(old_mem, vmm, &old_mem->vma[0]);
1083 	if (ret)
1084 		goto done;
1085 
1086 	ret = nouveau_mem_map(new_mem, vmm, &old_mem->vma[1]);
1087 done:
1088 	if (ret) {
1089 		nvif_vmm_put(vmm, &old_mem->vma[1]);
1090 		nvif_vmm_put(vmm, &old_mem->vma[0]);
1091 	}
1092 	return 0;
1093 }
1094 
1095 static int
1096 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
1097 		     bool no_wait_gpu, struct ttm_mem_reg *new_reg)
1098 {
1099 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1100 	struct nouveau_channel *chan = drm->ttm.chan;
1101 	struct nouveau_cli *cli = (void *)chan->user.client;
1102 	struct nouveau_fence *fence;
1103 	int ret;
1104 
1105 	/* create temporary vmas for the transfer and attach them to the
1106 	 * old nvkm_mem node, these will get cleaned up after ttm has
1107 	 * destroyed the ttm_mem_reg
1108 	 */
1109 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
1110 		ret = nouveau_bo_move_prep(drm, bo, new_reg);
1111 		if (ret)
1112 			return ret;
1113 	}
1114 
1115 	mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
1116 	ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, intr);
1117 	if (ret == 0) {
1118 		ret = drm->ttm.move(chan, bo, &bo->mem, new_reg);
1119 		if (ret == 0) {
1120 			ret = nouveau_fence_new(chan, false, &fence);
1121 			if (ret == 0) {
1122 				ret = ttm_bo_move_accel_cleanup(bo,
1123 								&fence->base,
1124 								evict,
1125 								new_reg);
1126 				nouveau_fence_unref(&fence);
1127 			}
1128 		}
1129 	}
1130 	mutex_unlock(&cli->mutex);
1131 	return ret;
1132 }
1133 
1134 void
1135 nouveau_bo_move_init(struct nouveau_drm *drm)
1136 {
1137 	static const struct {
1138 		const char *name;
1139 		int engine;
1140 		s32 oclass;
1141 		int (*exec)(struct nouveau_channel *,
1142 			    struct ttm_buffer_object *,
1143 			    struct ttm_mem_reg *, struct ttm_mem_reg *);
1144 		int (*init)(struct nouveau_channel *, u32 handle);
1145 	} _methods[] = {
1146 		{  "COPY", 4, 0xc1b5, nve0_bo_move_copy, nve0_bo_move_init },
1147 		{  "GRCE", 0, 0xc1b5, nve0_bo_move_copy, nvc0_bo_move_init },
1148 		{  "COPY", 4, 0xc0b5, nve0_bo_move_copy, nve0_bo_move_init },
1149 		{  "GRCE", 0, 0xc0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1150 		{  "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
1151 		{  "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1152 		{  "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
1153 		{  "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1154 		{ "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
1155 		{ "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
1156 		{  "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
1157 		{ "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
1158 		{  "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
1159 		{  "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
1160 		{  "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
1161 		{},
1162 		{ "CRYPT", 0, 0x88b4, nv98_bo_move_exec, nv50_bo_move_init },
1163 	}, *mthd = _methods;
1164 	const char *name = "CPU";
1165 	int ret;
1166 
1167 	do {
1168 		struct nouveau_channel *chan;
1169 
1170 		if (mthd->engine)
1171 			chan = drm->cechan;
1172 		else
1173 			chan = drm->channel;
1174 		if (chan == NULL)
1175 			continue;
1176 
1177 		ret = nvif_object_init(&chan->user,
1178 				       mthd->oclass | (mthd->engine << 16),
1179 				       mthd->oclass, NULL, 0,
1180 				       &drm->ttm.copy);
1181 		if (ret == 0) {
1182 			ret = mthd->init(chan, drm->ttm.copy.handle);
1183 			if (ret) {
1184 				nvif_object_fini(&drm->ttm.copy);
1185 				continue;
1186 			}
1187 
1188 			drm->ttm.move = mthd->exec;
1189 			drm->ttm.chan = chan;
1190 			name = mthd->name;
1191 			break;
1192 		}
1193 	} while ((++mthd)->exec);
1194 
1195 	NV_INFO(drm, "MM: using %s for buffer copies\n", name);
1196 }
1197 
1198 static int
1199 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
1200 		      bool no_wait_gpu, struct ttm_mem_reg *new_reg)
1201 {
1202 	struct ttm_place placement_memtype = {
1203 		.fpfn = 0,
1204 		.lpfn = 0,
1205 		.flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
1206 	};
1207 	struct ttm_placement placement;
1208 	struct ttm_mem_reg tmp_reg;
1209 	int ret;
1210 
1211 	placement.num_placement = placement.num_busy_placement = 1;
1212 	placement.placement = placement.busy_placement = &placement_memtype;
1213 
1214 	tmp_reg = *new_reg;
1215 	tmp_reg.mm_node = NULL;
1216 	ret = ttm_bo_mem_space(bo, &placement, &tmp_reg, intr, no_wait_gpu);
1217 	if (ret)
1218 		return ret;
1219 
1220 	ret = ttm_tt_bind(bo->ttm, &tmp_reg);
1221 	if (ret)
1222 		goto out;
1223 
1224 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, &tmp_reg);
1225 	if (ret)
1226 		goto out;
1227 
1228 	ret = ttm_bo_move_ttm(bo, intr, no_wait_gpu, new_reg);
1229 out:
1230 	ttm_bo_mem_put(bo, &tmp_reg);
1231 	return ret;
1232 }
1233 
1234 static int
1235 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
1236 		      bool no_wait_gpu, struct ttm_mem_reg *new_reg)
1237 {
1238 	struct ttm_place placement_memtype = {
1239 		.fpfn = 0,
1240 		.lpfn = 0,
1241 		.flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
1242 	};
1243 	struct ttm_placement placement;
1244 	struct ttm_mem_reg tmp_reg;
1245 	int ret;
1246 
1247 	placement.num_placement = placement.num_busy_placement = 1;
1248 	placement.placement = placement.busy_placement = &placement_memtype;
1249 
1250 	tmp_reg = *new_reg;
1251 	tmp_reg.mm_node = NULL;
1252 	ret = ttm_bo_mem_space(bo, &placement, &tmp_reg, intr, no_wait_gpu);
1253 	if (ret)
1254 		return ret;
1255 
1256 	ret = ttm_bo_move_ttm(bo, intr, no_wait_gpu, &tmp_reg);
1257 	if (ret)
1258 		goto out;
1259 
1260 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_reg);
1261 	if (ret)
1262 		goto out;
1263 
1264 out:
1265 	ttm_bo_mem_put(bo, &tmp_reg);
1266 	return ret;
1267 }
1268 
1269 static void
1270 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, bool evict,
1271 		     struct ttm_mem_reg *new_reg)
1272 {
1273 	struct nouveau_mem *mem = new_reg ? nouveau_mem(new_reg) : NULL;
1274 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1275 	struct nouveau_vma *vma;
1276 
1277 	/* ttm can now (stupidly) pass the driver bos it didn't create... */
1278 	if (bo->destroy != nouveau_bo_del_ttm)
1279 		return;
1280 
1281 	if (mem && new_reg->mem_type != TTM_PL_SYSTEM &&
1282 	    mem->mem.page == nvbo->page) {
1283 		list_for_each_entry(vma, &nvbo->vma_list, head) {
1284 			nouveau_vma_map(vma, mem);
1285 		}
1286 	} else {
1287 		list_for_each_entry(vma, &nvbo->vma_list, head) {
1288 			WARN_ON(ttm_bo_wait(bo, false, false));
1289 			nouveau_vma_unmap(vma);
1290 		}
1291 	}
1292 }
1293 
1294 static int
1295 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_reg,
1296 		   struct nouveau_drm_tile **new_tile)
1297 {
1298 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1299 	struct drm_device *dev = drm->dev;
1300 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1301 	u64 offset = new_reg->start << PAGE_SHIFT;
1302 
1303 	*new_tile = NULL;
1304 	if (new_reg->mem_type != TTM_PL_VRAM)
1305 		return 0;
1306 
1307 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
1308 		*new_tile = nv10_bo_set_tiling(dev, offset, new_reg->size,
1309 					       nvbo->mode, nvbo->zeta);
1310 	}
1311 
1312 	return 0;
1313 }
1314 
1315 static void
1316 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1317 		      struct nouveau_drm_tile *new_tile,
1318 		      struct nouveau_drm_tile **old_tile)
1319 {
1320 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1321 	struct drm_device *dev = drm->dev;
1322 	struct dma_fence *fence = reservation_object_get_excl(bo->resv);
1323 
1324 	nv10_bo_put_tile_region(dev, *old_tile, fence);
1325 	*old_tile = new_tile;
1326 }
1327 
1328 static int
1329 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
1330 		bool no_wait_gpu, struct ttm_mem_reg *new_reg)
1331 {
1332 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1333 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1334 	struct ttm_mem_reg *old_reg = &bo->mem;
1335 	struct nouveau_drm_tile *new_tile = NULL;
1336 	int ret = 0;
1337 
1338 	ret = ttm_bo_wait(bo, intr, no_wait_gpu);
1339 	if (ret)
1340 		return ret;
1341 
1342 	if (nvbo->pin_refcnt)
1343 		NV_WARN(drm, "Moving pinned object %p!\n", nvbo);
1344 
1345 	if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1346 		ret = nouveau_bo_vm_bind(bo, new_reg, &new_tile);
1347 		if (ret)
1348 			return ret;
1349 	}
1350 
1351 	/* Fake bo copy. */
1352 	if (old_reg->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
1353 		BUG_ON(bo->mem.mm_node != NULL);
1354 		bo->mem = *new_reg;
1355 		new_reg->mm_node = NULL;
1356 		goto out;
1357 	}
1358 
1359 	/* Hardware assisted copy. */
1360 	if (drm->ttm.move) {
1361 		if (new_reg->mem_type == TTM_PL_SYSTEM)
1362 			ret = nouveau_bo_move_flipd(bo, evict, intr,
1363 						    no_wait_gpu, new_reg);
1364 		else if (old_reg->mem_type == TTM_PL_SYSTEM)
1365 			ret = nouveau_bo_move_flips(bo, evict, intr,
1366 						    no_wait_gpu, new_reg);
1367 		else
1368 			ret = nouveau_bo_move_m2mf(bo, evict, intr,
1369 						   no_wait_gpu, new_reg);
1370 		if (!ret)
1371 			goto out;
1372 	}
1373 
1374 	/* Fallback to software copy. */
1375 	ret = ttm_bo_wait(bo, intr, no_wait_gpu);
1376 	if (ret == 0)
1377 		ret = ttm_bo_move_memcpy(bo, intr, no_wait_gpu, new_reg);
1378 
1379 out:
1380 	if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1381 		if (ret)
1382 			nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1383 		else
1384 			nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1385 	}
1386 
1387 	return ret;
1388 }
1389 
1390 static int
1391 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1392 {
1393 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1394 
1395 	return drm_vma_node_verify_access(&nvbo->gem.vma_node,
1396 					  filp->private_data);
1397 }
1398 
1399 static int
1400 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *reg)
1401 {
1402 	struct ttm_mem_type_manager *man = &bdev->man[reg->mem_type];
1403 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1404 	struct nvkm_device *device = nvxx_device(&drm->client.device);
1405 	struct nouveau_mem *mem = nouveau_mem(reg);
1406 
1407 	reg->bus.addr = NULL;
1408 	reg->bus.offset = 0;
1409 	reg->bus.size = reg->num_pages << PAGE_SHIFT;
1410 	reg->bus.base = 0;
1411 	reg->bus.is_iomem = false;
1412 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
1413 		return -EINVAL;
1414 	switch (reg->mem_type) {
1415 	case TTM_PL_SYSTEM:
1416 		/* System memory */
1417 		return 0;
1418 	case TTM_PL_TT:
1419 #if IS_ENABLED(CONFIG_AGP)
1420 		if (drm->agp.bridge) {
1421 			reg->bus.offset = reg->start << PAGE_SHIFT;
1422 			reg->bus.base = drm->agp.base;
1423 			reg->bus.is_iomem = !drm->agp.cma;
1424 		}
1425 #endif
1426 		if (drm->client.mem->oclass < NVIF_CLASS_MEM_NV50 || !mem->kind)
1427 			/* untiled */
1428 			break;
1429 		/* fallthrough, tiled memory */
1430 	case TTM_PL_VRAM:
1431 		reg->bus.offset = reg->start << PAGE_SHIFT;
1432 		reg->bus.base = device->func->resource_addr(device, 1);
1433 		reg->bus.is_iomem = true;
1434 		if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1435 			union {
1436 				struct nv50_mem_map_v0 nv50;
1437 				struct gf100_mem_map_v0 gf100;
1438 			} args;
1439 			u64 handle, length;
1440 			u32 argc = 0;
1441 			int ret;
1442 
1443 			switch (mem->mem.object.oclass) {
1444 			case NVIF_CLASS_MEM_NV50:
1445 				args.nv50.version = 0;
1446 				args.nv50.ro = 0;
1447 				args.nv50.kind = mem->kind;
1448 				args.nv50.comp = mem->comp;
1449 				break;
1450 			case NVIF_CLASS_MEM_GF100:
1451 				args.gf100.version = 0;
1452 				args.gf100.ro = 0;
1453 				args.gf100.kind = mem->kind;
1454 				break;
1455 			default:
1456 				WARN_ON(1);
1457 				break;
1458 			}
1459 
1460 			ret = nvif_object_map_handle(&mem->mem.object,
1461 						     &argc, argc,
1462 						     &handle, &length);
1463 			if (ret != 1)
1464 				return ret ? ret : -EINVAL;
1465 
1466 			reg->bus.base = 0;
1467 			reg->bus.offset = handle;
1468 		}
1469 		break;
1470 	default:
1471 		return -EINVAL;
1472 	}
1473 	return 0;
1474 }
1475 
1476 static void
1477 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *reg)
1478 {
1479 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1480 	struct nouveau_mem *mem = nouveau_mem(reg);
1481 
1482 	if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1483 		switch (reg->mem_type) {
1484 		case TTM_PL_TT:
1485 			if (mem->kind)
1486 				nvif_object_unmap_handle(&mem->mem.object);
1487 			break;
1488 		case TTM_PL_VRAM:
1489 			nvif_object_unmap_handle(&mem->mem.object);
1490 			break;
1491 		default:
1492 			break;
1493 		}
1494 	}
1495 }
1496 
1497 static int
1498 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1499 {
1500 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1501 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1502 	struct nvkm_device *device = nvxx_device(&drm->client.device);
1503 	u32 mappable = device->func->resource_size(device, 1) >> PAGE_SHIFT;
1504 	int i, ret;
1505 
1506 	/* as long as the bo isn't in vram, and isn't tiled, we've got
1507 	 * nothing to do here.
1508 	 */
1509 	if (bo->mem.mem_type != TTM_PL_VRAM) {
1510 		if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1511 		    !nvbo->kind)
1512 			return 0;
1513 
1514 		if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1515 			nouveau_bo_placement_set(nvbo, TTM_PL_TT, 0);
1516 
1517 			ret = nouveau_bo_validate(nvbo, false, false);
1518 			if (ret)
1519 				return ret;
1520 		}
1521 		return 0;
1522 	}
1523 
1524 	/* make sure bo is in mappable vram */
1525 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1526 	    bo->mem.start + bo->mem.num_pages < mappable)
1527 		return 0;
1528 
1529 	for (i = 0; i < nvbo->placement.num_placement; ++i) {
1530 		nvbo->placements[i].fpfn = 0;
1531 		nvbo->placements[i].lpfn = mappable;
1532 	}
1533 
1534 	for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
1535 		nvbo->busy_placements[i].fpfn = 0;
1536 		nvbo->busy_placements[i].lpfn = mappable;
1537 	}
1538 
1539 	nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_VRAM, 0);
1540 	return nouveau_bo_validate(nvbo, false, false);
1541 }
1542 
1543 static int
1544 nouveau_ttm_tt_populate(struct ttm_tt *ttm)
1545 {
1546 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1547 	struct nouveau_drm *drm;
1548 	struct device *dev;
1549 	unsigned i;
1550 	int r;
1551 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1552 
1553 	if (ttm->state != tt_unpopulated)
1554 		return 0;
1555 
1556 	if (slave && ttm->sg) {
1557 		/* make userspace faulting work */
1558 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1559 						 ttm_dma->dma_address, ttm->num_pages);
1560 		ttm->state = tt_unbound;
1561 		return 0;
1562 	}
1563 
1564 	drm = nouveau_bdev(ttm->bdev);
1565 	dev = drm->dev->dev;
1566 
1567 #if IS_ENABLED(CONFIG_AGP)
1568 	if (drm->agp.bridge) {
1569 		return ttm_agp_tt_populate(ttm);
1570 	}
1571 #endif
1572 
1573 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1574 	if (swiotlb_nr_tbl()) {
1575 		return ttm_dma_populate((void *)ttm, dev);
1576 	}
1577 #endif
1578 
1579 	r = ttm_pool_populate(ttm);
1580 	if (r) {
1581 		return r;
1582 	}
1583 
1584 	for (i = 0; i < ttm->num_pages; i++) {
1585 		dma_addr_t addr;
1586 
1587 		addr = dma_map_page(dev, ttm->pages[i], 0, PAGE_SIZE,
1588 				    DMA_BIDIRECTIONAL);
1589 
1590 		if (dma_mapping_error(dev, addr)) {
1591 			while (i--) {
1592 				dma_unmap_page(dev, ttm_dma->dma_address[i],
1593 					       PAGE_SIZE, DMA_BIDIRECTIONAL);
1594 				ttm_dma->dma_address[i] = 0;
1595 			}
1596 			ttm_pool_unpopulate(ttm);
1597 			return -EFAULT;
1598 		}
1599 
1600 		ttm_dma->dma_address[i] = addr;
1601 	}
1602 	return 0;
1603 }
1604 
1605 static void
1606 nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
1607 {
1608 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1609 	struct nouveau_drm *drm;
1610 	struct device *dev;
1611 	unsigned i;
1612 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1613 
1614 	if (slave)
1615 		return;
1616 
1617 	drm = nouveau_bdev(ttm->bdev);
1618 	dev = drm->dev->dev;
1619 
1620 #if IS_ENABLED(CONFIG_AGP)
1621 	if (drm->agp.bridge) {
1622 		ttm_agp_tt_unpopulate(ttm);
1623 		return;
1624 	}
1625 #endif
1626 
1627 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1628 	if (swiotlb_nr_tbl()) {
1629 		ttm_dma_unpopulate((void *)ttm, dev);
1630 		return;
1631 	}
1632 #endif
1633 
1634 	for (i = 0; i < ttm->num_pages; i++) {
1635 		if (ttm_dma->dma_address[i]) {
1636 			dma_unmap_page(dev, ttm_dma->dma_address[i], PAGE_SIZE,
1637 				       DMA_BIDIRECTIONAL);
1638 		}
1639 	}
1640 
1641 	ttm_pool_unpopulate(ttm);
1642 }
1643 
1644 void
1645 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1646 {
1647 	struct reservation_object *resv = nvbo->bo.resv;
1648 
1649 	if (exclusive)
1650 		reservation_object_add_excl_fence(resv, &fence->base);
1651 	else if (fence)
1652 		reservation_object_add_shared_fence(resv, &fence->base);
1653 }
1654 
1655 struct ttm_bo_driver nouveau_bo_driver = {
1656 	.ttm_tt_create = &nouveau_ttm_tt_create,
1657 	.ttm_tt_populate = &nouveau_ttm_tt_populate,
1658 	.ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1659 	.invalidate_caches = nouveau_bo_invalidate_caches,
1660 	.init_mem_type = nouveau_bo_init_mem_type,
1661 	.eviction_valuable = ttm_bo_eviction_valuable,
1662 	.evict_flags = nouveau_bo_evict_flags,
1663 	.move_notify = nouveau_bo_move_ntfy,
1664 	.move = nouveau_bo_move,
1665 	.verify_access = nouveau_bo_verify_access,
1666 	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1667 	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1668 	.io_mem_free = &nouveau_ttm_io_mem_free,
1669 	.io_mem_pfn = ttm_bo_default_io_mem_pfn,
1670 };
1671