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 (!nouveau_drm_use_coherent_gpu_mapping(drm))
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) &&
266 		    (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
267 			continue;
268 
269 		/* Select this page size if it's the first that supports
270 		 * the potential memory domains, or when it's compatible
271 		 * with the requested compression settings.
272 		 */
273 		if (pi < 0 || !nvbo->comp || vmm->page[i].comp)
274 			pi = i;
275 
276 		/* Stop once the buffer is larger than the current page size. */
277 		if (size >= 1ULL << vmm->page[i].shift)
278 			break;
279 	}
280 
281 	if (WARN_ON(pi < 0))
282 		return -EINVAL;
283 
284 	/* Disable compression if suitable settings couldn't be found. */
285 	if (nvbo->comp && !vmm->page[pi].comp) {
286 		if (mmu->object.oclass >= NVIF_CLASS_MMU_GF100)
287 			nvbo->kind = mmu->kind[nvbo->kind];
288 		nvbo->comp = 0;
289 	}
290 	nvbo->page = vmm->page[pi].shift;
291 
292 	nouveau_bo_fixup_align(nvbo, flags, &align, &size);
293 	nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
294 	nouveau_bo_placement_set(nvbo, flags, 0);
295 
296 	acc_size = ttm_bo_dma_acc_size(&drm->ttm.bdev, size,
297 				       sizeof(struct nouveau_bo));
298 
299 	ret = ttm_bo_init(&drm->ttm.bdev, &nvbo->bo, size,
300 			  type, &nvbo->placement,
301 			  align >> PAGE_SHIFT, false, acc_size, sg,
302 			  robj, nouveau_bo_del_ttm);
303 	if (ret) {
304 		/* ttm will call nouveau_bo_del_ttm if it fails.. */
305 		return ret;
306 	}
307 
308 	*pnvbo = nvbo;
309 	return 0;
310 }
311 
312 static void
313 set_placement_list(struct ttm_place *pl, unsigned *n, uint32_t type, uint32_t flags)
314 {
315 	*n = 0;
316 
317 	if (type & TTM_PL_FLAG_VRAM)
318 		pl[(*n)++].flags = TTM_PL_FLAG_VRAM | flags;
319 	if (type & TTM_PL_FLAG_TT)
320 		pl[(*n)++].flags = TTM_PL_FLAG_TT | flags;
321 	if (type & TTM_PL_FLAG_SYSTEM)
322 		pl[(*n)++].flags = TTM_PL_FLAG_SYSTEM | flags;
323 }
324 
325 static void
326 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
327 {
328 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
329 	u32 vram_pages = drm->client.device.info.ram_size >> PAGE_SHIFT;
330 	unsigned i, fpfn, lpfn;
331 
332 	if (drm->client.device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
333 	    nvbo->mode && (type & TTM_PL_FLAG_VRAM) &&
334 	    nvbo->bo.mem.num_pages < vram_pages / 4) {
335 		/*
336 		 * Make sure that the color and depth buffers are handled
337 		 * by independent memory controller units. Up to a 9x
338 		 * speed up when alpha-blending and depth-test are enabled
339 		 * at the same time.
340 		 */
341 		if (nvbo->zeta) {
342 			fpfn = vram_pages / 2;
343 			lpfn = ~0;
344 		} else {
345 			fpfn = 0;
346 			lpfn = vram_pages / 2;
347 		}
348 		for (i = 0; i < nvbo->placement.num_placement; ++i) {
349 			nvbo->placements[i].fpfn = fpfn;
350 			nvbo->placements[i].lpfn = lpfn;
351 		}
352 		for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
353 			nvbo->busy_placements[i].fpfn = fpfn;
354 			nvbo->busy_placements[i].lpfn = lpfn;
355 		}
356 	}
357 }
358 
359 void
360 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
361 {
362 	struct ttm_placement *pl = &nvbo->placement;
363 	uint32_t flags = (nvbo->force_coherent ? TTM_PL_FLAG_UNCACHED :
364 						 TTM_PL_MASK_CACHING) |
365 			 (nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
366 
367 	pl->placement = nvbo->placements;
368 	set_placement_list(nvbo->placements, &pl->num_placement,
369 			   type, flags);
370 
371 	pl->busy_placement = nvbo->busy_placements;
372 	set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
373 			   type | busy, flags);
374 
375 	set_placement_range(nvbo, type);
376 }
377 
378 int
379 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype, bool contig)
380 {
381 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
382 	struct ttm_buffer_object *bo = &nvbo->bo;
383 	bool force = false, evict = false;
384 	int ret;
385 
386 	ret = ttm_bo_reserve(bo, false, false, NULL);
387 	if (ret)
388 		return ret;
389 
390 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
391 	    memtype == TTM_PL_FLAG_VRAM && contig) {
392 		if (!nvbo->contig) {
393 			nvbo->contig = true;
394 			force = true;
395 			evict = true;
396 		}
397 	}
398 
399 	if (nvbo->pin_refcnt) {
400 		if (!(memtype & (1 << bo->mem.mem_type)) || evict) {
401 			NV_ERROR(drm, "bo %p pinned elsewhere: "
402 				      "0x%08x vs 0x%08x\n", bo,
403 				 1 << bo->mem.mem_type, memtype);
404 			ret = -EBUSY;
405 		}
406 		nvbo->pin_refcnt++;
407 		goto out;
408 	}
409 
410 	if (evict) {
411 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT, 0);
412 		ret = nouveau_bo_validate(nvbo, false, false);
413 		if (ret)
414 			goto out;
415 	}
416 
417 	nvbo->pin_refcnt++;
418 	nouveau_bo_placement_set(nvbo, memtype, 0);
419 
420 	/* drop pin_refcnt temporarily, so we don't trip the assertion
421 	 * in nouveau_bo_move() that makes sure we're not trying to
422 	 * move a pinned buffer
423 	 */
424 	nvbo->pin_refcnt--;
425 	ret = nouveau_bo_validate(nvbo, false, false);
426 	if (ret)
427 		goto out;
428 	nvbo->pin_refcnt++;
429 
430 	switch (bo->mem.mem_type) {
431 	case TTM_PL_VRAM:
432 		drm->gem.vram_available -= bo->mem.size;
433 		break;
434 	case TTM_PL_TT:
435 		drm->gem.gart_available -= bo->mem.size;
436 		break;
437 	default:
438 		break;
439 	}
440 
441 out:
442 	if (force && ret)
443 		nvbo->contig = false;
444 	ttm_bo_unreserve(bo);
445 	return ret;
446 }
447 
448 int
449 nouveau_bo_unpin(struct nouveau_bo *nvbo)
450 {
451 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
452 	struct ttm_buffer_object *bo = &nvbo->bo;
453 	int ret, ref;
454 
455 	ret = ttm_bo_reserve(bo, false, false, NULL);
456 	if (ret)
457 		return ret;
458 
459 	ref = --nvbo->pin_refcnt;
460 	WARN_ON_ONCE(ref < 0);
461 	if (ref)
462 		goto out;
463 
464 	nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
465 
466 	ret = nouveau_bo_validate(nvbo, false, false);
467 	if (ret == 0) {
468 		switch (bo->mem.mem_type) {
469 		case TTM_PL_VRAM:
470 			drm->gem.vram_available += bo->mem.size;
471 			break;
472 		case TTM_PL_TT:
473 			drm->gem.gart_available += bo->mem.size;
474 			break;
475 		default:
476 			break;
477 		}
478 	}
479 
480 out:
481 	ttm_bo_unreserve(bo);
482 	return ret;
483 }
484 
485 int
486 nouveau_bo_map(struct nouveau_bo *nvbo)
487 {
488 	int ret;
489 
490 	ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
491 	if (ret)
492 		return ret;
493 
494 	ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
495 
496 	ttm_bo_unreserve(&nvbo->bo);
497 	return ret;
498 }
499 
500 void
501 nouveau_bo_unmap(struct nouveau_bo *nvbo)
502 {
503 	if (!nvbo)
504 		return;
505 
506 	ttm_bo_kunmap(&nvbo->kmap);
507 }
508 
509 void
510 nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
511 {
512 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
513 	struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
514 	int i;
515 
516 	if (!ttm_dma)
517 		return;
518 
519 	/* Don't waste time looping if the object is coherent */
520 	if (nvbo->force_coherent)
521 		return;
522 
523 	for (i = 0; i < ttm_dma->ttm.num_pages; i++)
524 		dma_sync_single_for_device(drm->dev->dev,
525 					   ttm_dma->dma_address[i],
526 					   PAGE_SIZE, DMA_TO_DEVICE);
527 }
528 
529 void
530 nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
531 {
532 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
533 	struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
534 	int i;
535 
536 	if (!ttm_dma)
537 		return;
538 
539 	/* Don't waste time looping if the object is coherent */
540 	if (nvbo->force_coherent)
541 		return;
542 
543 	for (i = 0; i < ttm_dma->ttm.num_pages; i++)
544 		dma_sync_single_for_cpu(drm->dev->dev, ttm_dma->dma_address[i],
545 					PAGE_SIZE, DMA_FROM_DEVICE);
546 }
547 
548 int
549 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
550 		    bool no_wait_gpu)
551 {
552 	struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
553 	int ret;
554 
555 	ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, &ctx);
556 	if (ret)
557 		return ret;
558 
559 	nouveau_bo_sync_for_device(nvbo);
560 
561 	return 0;
562 }
563 
564 void
565 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
566 {
567 	bool is_iomem;
568 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
569 
570 	mem += index;
571 
572 	if (is_iomem)
573 		iowrite16_native(val, (void __force __iomem *)mem);
574 	else
575 		*mem = val;
576 }
577 
578 u32
579 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
580 {
581 	bool is_iomem;
582 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
583 
584 	mem += index;
585 
586 	if (is_iomem)
587 		return ioread32_native((void __force __iomem *)mem);
588 	else
589 		return *mem;
590 }
591 
592 void
593 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
594 {
595 	bool is_iomem;
596 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
597 
598 	mem += index;
599 
600 	if (is_iomem)
601 		iowrite32_native(val, (void __force __iomem *)mem);
602 	else
603 		*mem = val;
604 }
605 
606 static struct ttm_tt *
607 nouveau_ttm_tt_create(struct ttm_buffer_object *bo, uint32_t page_flags)
608 {
609 #if IS_ENABLED(CONFIG_AGP)
610 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
611 
612 	if (drm->agp.bridge) {
613 		return ttm_agp_tt_create(bo, drm->agp.bridge, page_flags);
614 	}
615 #endif
616 
617 	return nouveau_sgdma_create_ttm(bo, page_flags);
618 }
619 
620 static int
621 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
622 {
623 	/* We'll do this from user space. */
624 	return 0;
625 }
626 
627 static int
628 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
629 			 struct ttm_mem_type_manager *man)
630 {
631 	struct nouveau_drm *drm = nouveau_bdev(bdev);
632 	struct nvif_mmu *mmu = &drm->client.mmu;
633 
634 	switch (type) {
635 	case TTM_PL_SYSTEM:
636 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
637 		man->available_caching = TTM_PL_MASK_CACHING;
638 		man->default_caching = TTM_PL_FLAG_CACHED;
639 		break;
640 	case TTM_PL_VRAM:
641 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
642 			     TTM_MEMTYPE_FLAG_MAPPABLE;
643 		man->available_caching = TTM_PL_FLAG_UNCACHED |
644 					 TTM_PL_FLAG_WC;
645 		man->default_caching = TTM_PL_FLAG_WC;
646 
647 		if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
648 			/* Some BARs do not support being ioremapped WC */
649 			const u8 type = mmu->type[drm->ttm.type_vram].type;
650 			if (type & NVIF_MEM_UNCACHED) {
651 				man->available_caching = TTM_PL_FLAG_UNCACHED;
652 				man->default_caching = TTM_PL_FLAG_UNCACHED;
653 			}
654 
655 			man->func = &nouveau_vram_manager;
656 			man->io_reserve_fastpath = false;
657 			man->use_io_reserve_lru = true;
658 		} else {
659 			man->func = &ttm_bo_manager_func;
660 		}
661 		break;
662 	case TTM_PL_TT:
663 		if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA)
664 			man->func = &nouveau_gart_manager;
665 		else
666 		if (!drm->agp.bridge)
667 			man->func = &nv04_gart_manager;
668 		else
669 			man->func = &ttm_bo_manager_func;
670 
671 		if (drm->agp.bridge) {
672 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
673 			man->available_caching = TTM_PL_FLAG_UNCACHED |
674 				TTM_PL_FLAG_WC;
675 			man->default_caching = TTM_PL_FLAG_WC;
676 		} else {
677 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
678 				     TTM_MEMTYPE_FLAG_CMA;
679 			man->available_caching = TTM_PL_MASK_CACHING;
680 			man->default_caching = TTM_PL_FLAG_CACHED;
681 		}
682 
683 		break;
684 	default:
685 		return -EINVAL;
686 	}
687 	return 0;
688 }
689 
690 static void
691 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
692 {
693 	struct nouveau_bo *nvbo = nouveau_bo(bo);
694 
695 	switch (bo->mem.mem_type) {
696 	case TTM_PL_VRAM:
697 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
698 					 TTM_PL_FLAG_SYSTEM);
699 		break;
700 	default:
701 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
702 		break;
703 	}
704 
705 	*pl = nvbo->placement;
706 }
707 
708 
709 static int
710 nve0_bo_move_init(struct nouveau_channel *chan, u32 handle)
711 {
712 	int ret = RING_SPACE(chan, 2);
713 	if (ret == 0) {
714 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
715 		OUT_RING  (chan, handle & 0x0000ffff);
716 		FIRE_RING (chan);
717 	}
718 	return ret;
719 }
720 
721 static int
722 nve0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
723 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
724 {
725 	struct nouveau_mem *mem = nouveau_mem(old_reg);
726 	int ret = RING_SPACE(chan, 10);
727 	if (ret == 0) {
728 		BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
729 		OUT_RING  (chan, upper_32_bits(mem->vma[0].addr));
730 		OUT_RING  (chan, lower_32_bits(mem->vma[0].addr));
731 		OUT_RING  (chan, upper_32_bits(mem->vma[1].addr));
732 		OUT_RING  (chan, lower_32_bits(mem->vma[1].addr));
733 		OUT_RING  (chan, PAGE_SIZE);
734 		OUT_RING  (chan, PAGE_SIZE);
735 		OUT_RING  (chan, PAGE_SIZE);
736 		OUT_RING  (chan, new_reg->num_pages);
737 		BEGIN_IMC0(chan, NvSubCopy, 0x0300, 0x0386);
738 	}
739 	return ret;
740 }
741 
742 static int
743 nvc0_bo_move_init(struct nouveau_channel *chan, u32 handle)
744 {
745 	int ret = RING_SPACE(chan, 2);
746 	if (ret == 0) {
747 		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
748 		OUT_RING  (chan, handle);
749 	}
750 	return ret;
751 }
752 
753 static int
754 nvc0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
755 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
756 {
757 	struct nouveau_mem *mem = nouveau_mem(old_reg);
758 	u64 src_offset = mem->vma[0].addr;
759 	u64 dst_offset = mem->vma[1].addr;
760 	u32 page_count = new_reg->num_pages;
761 	int ret;
762 
763 	page_count = new_reg->num_pages;
764 	while (page_count) {
765 		int line_count = (page_count > 8191) ? 8191 : page_count;
766 
767 		ret = RING_SPACE(chan, 11);
768 		if (ret)
769 			return ret;
770 
771 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 8);
772 		OUT_RING  (chan, upper_32_bits(src_offset));
773 		OUT_RING  (chan, lower_32_bits(src_offset));
774 		OUT_RING  (chan, upper_32_bits(dst_offset));
775 		OUT_RING  (chan, lower_32_bits(dst_offset));
776 		OUT_RING  (chan, PAGE_SIZE);
777 		OUT_RING  (chan, PAGE_SIZE);
778 		OUT_RING  (chan, PAGE_SIZE);
779 		OUT_RING  (chan, line_count);
780 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
781 		OUT_RING  (chan, 0x00000110);
782 
783 		page_count -= line_count;
784 		src_offset += (PAGE_SIZE * line_count);
785 		dst_offset += (PAGE_SIZE * line_count);
786 	}
787 
788 	return 0;
789 }
790 
791 static int
792 nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
793 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
794 {
795 	struct nouveau_mem *mem = nouveau_mem(old_reg);
796 	u64 src_offset = mem->vma[0].addr;
797 	u64 dst_offset = mem->vma[1].addr;
798 	u32 page_count = new_reg->num_pages;
799 	int ret;
800 
801 	page_count = new_reg->num_pages;
802 	while (page_count) {
803 		int line_count = (page_count > 2047) ? 2047 : page_count;
804 
805 		ret = RING_SPACE(chan, 12);
806 		if (ret)
807 			return ret;
808 
809 		BEGIN_NVC0(chan, NvSubCopy, 0x0238, 2);
810 		OUT_RING  (chan, upper_32_bits(dst_offset));
811 		OUT_RING  (chan, lower_32_bits(dst_offset));
812 		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 6);
813 		OUT_RING  (chan, upper_32_bits(src_offset));
814 		OUT_RING  (chan, lower_32_bits(src_offset));
815 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
816 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
817 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
818 		OUT_RING  (chan, line_count);
819 		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
820 		OUT_RING  (chan, 0x00100110);
821 
822 		page_count -= line_count;
823 		src_offset += (PAGE_SIZE * line_count);
824 		dst_offset += (PAGE_SIZE * line_count);
825 	}
826 
827 	return 0;
828 }
829 
830 static int
831 nva3_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
832 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
833 {
834 	struct nouveau_mem *mem = nouveau_mem(old_reg);
835 	u64 src_offset = mem->vma[0].addr;
836 	u64 dst_offset = mem->vma[1].addr;
837 	u32 page_count = new_reg->num_pages;
838 	int ret;
839 
840 	page_count = new_reg->num_pages;
841 	while (page_count) {
842 		int line_count = (page_count > 8191) ? 8191 : page_count;
843 
844 		ret = RING_SPACE(chan, 11);
845 		if (ret)
846 			return ret;
847 
848 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
849 		OUT_RING  (chan, upper_32_bits(src_offset));
850 		OUT_RING  (chan, lower_32_bits(src_offset));
851 		OUT_RING  (chan, upper_32_bits(dst_offset));
852 		OUT_RING  (chan, lower_32_bits(dst_offset));
853 		OUT_RING  (chan, PAGE_SIZE);
854 		OUT_RING  (chan, PAGE_SIZE);
855 		OUT_RING  (chan, PAGE_SIZE);
856 		OUT_RING  (chan, line_count);
857 		BEGIN_NV04(chan, NvSubCopy, 0x0300, 1);
858 		OUT_RING  (chan, 0x00000110);
859 
860 		page_count -= line_count;
861 		src_offset += (PAGE_SIZE * line_count);
862 		dst_offset += (PAGE_SIZE * line_count);
863 	}
864 
865 	return 0;
866 }
867 
868 static int
869 nv98_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
870 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
871 {
872 	struct nouveau_mem *mem = nouveau_mem(old_reg);
873 	int ret = RING_SPACE(chan, 7);
874 	if (ret == 0) {
875 		BEGIN_NV04(chan, NvSubCopy, 0x0320, 6);
876 		OUT_RING  (chan, upper_32_bits(mem->vma[0].addr));
877 		OUT_RING  (chan, lower_32_bits(mem->vma[0].addr));
878 		OUT_RING  (chan, upper_32_bits(mem->vma[1].addr));
879 		OUT_RING  (chan, lower_32_bits(mem->vma[1].addr));
880 		OUT_RING  (chan, 0x00000000 /* COPY */);
881 		OUT_RING  (chan, new_reg->num_pages << PAGE_SHIFT);
882 	}
883 	return ret;
884 }
885 
886 static int
887 nv84_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
888 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
889 {
890 	struct nouveau_mem *mem = nouveau_mem(old_reg);
891 	int ret = RING_SPACE(chan, 7);
892 	if (ret == 0) {
893 		BEGIN_NV04(chan, NvSubCopy, 0x0304, 6);
894 		OUT_RING  (chan, new_reg->num_pages << PAGE_SHIFT);
895 		OUT_RING  (chan, upper_32_bits(mem->vma[0].addr));
896 		OUT_RING  (chan, lower_32_bits(mem->vma[0].addr));
897 		OUT_RING  (chan, upper_32_bits(mem->vma[1].addr));
898 		OUT_RING  (chan, lower_32_bits(mem->vma[1].addr));
899 		OUT_RING  (chan, 0x00000000 /* MODE_COPY, QUERY_NONE */);
900 	}
901 	return ret;
902 }
903 
904 static int
905 nv50_bo_move_init(struct nouveau_channel *chan, u32 handle)
906 {
907 	int ret = RING_SPACE(chan, 6);
908 	if (ret == 0) {
909 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
910 		OUT_RING  (chan, handle);
911 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 3);
912 		OUT_RING  (chan, chan->drm->ntfy.handle);
913 		OUT_RING  (chan, chan->vram.handle);
914 		OUT_RING  (chan, chan->vram.handle);
915 	}
916 
917 	return ret;
918 }
919 
920 static int
921 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
922 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
923 {
924 	struct nouveau_mem *mem = nouveau_mem(old_reg);
925 	u64 length = (new_reg->num_pages << PAGE_SHIFT);
926 	u64 src_offset = mem->vma[0].addr;
927 	u64 dst_offset = mem->vma[1].addr;
928 	int src_tiled = !!mem->kind;
929 	int dst_tiled = !!nouveau_mem(new_reg)->kind;
930 	int ret;
931 
932 	while (length) {
933 		u32 amount, stride, height;
934 
935 		ret = RING_SPACE(chan, 18 + 6 * (src_tiled + dst_tiled));
936 		if (ret)
937 			return ret;
938 
939 		amount  = min(length, (u64)(4 * 1024 * 1024));
940 		stride  = 16 * 4;
941 		height  = amount / stride;
942 
943 		if (src_tiled) {
944 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 7);
945 			OUT_RING  (chan, 0);
946 			OUT_RING  (chan, 0);
947 			OUT_RING  (chan, stride);
948 			OUT_RING  (chan, height);
949 			OUT_RING  (chan, 1);
950 			OUT_RING  (chan, 0);
951 			OUT_RING  (chan, 0);
952 		} else {
953 			BEGIN_NV04(chan, NvSubCopy, 0x0200, 1);
954 			OUT_RING  (chan, 1);
955 		}
956 		if (dst_tiled) {
957 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 7);
958 			OUT_RING  (chan, 0);
959 			OUT_RING  (chan, 0);
960 			OUT_RING  (chan, stride);
961 			OUT_RING  (chan, height);
962 			OUT_RING  (chan, 1);
963 			OUT_RING  (chan, 0);
964 			OUT_RING  (chan, 0);
965 		} else {
966 			BEGIN_NV04(chan, NvSubCopy, 0x021c, 1);
967 			OUT_RING  (chan, 1);
968 		}
969 
970 		BEGIN_NV04(chan, NvSubCopy, 0x0238, 2);
971 		OUT_RING  (chan, upper_32_bits(src_offset));
972 		OUT_RING  (chan, upper_32_bits(dst_offset));
973 		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
974 		OUT_RING  (chan, lower_32_bits(src_offset));
975 		OUT_RING  (chan, lower_32_bits(dst_offset));
976 		OUT_RING  (chan, stride);
977 		OUT_RING  (chan, stride);
978 		OUT_RING  (chan, stride);
979 		OUT_RING  (chan, height);
980 		OUT_RING  (chan, 0x00000101);
981 		OUT_RING  (chan, 0x00000000);
982 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
983 		OUT_RING  (chan, 0);
984 
985 		length -= amount;
986 		src_offset += amount;
987 		dst_offset += amount;
988 	}
989 
990 	return 0;
991 }
992 
993 static int
994 nv04_bo_move_init(struct nouveau_channel *chan, u32 handle)
995 {
996 	int ret = RING_SPACE(chan, 4);
997 	if (ret == 0) {
998 		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
999 		OUT_RING  (chan, handle);
1000 		BEGIN_NV04(chan, NvSubCopy, 0x0180, 1);
1001 		OUT_RING  (chan, chan->drm->ntfy.handle);
1002 	}
1003 
1004 	return ret;
1005 }
1006 
1007 static inline uint32_t
1008 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
1009 		      struct nouveau_channel *chan, struct ttm_mem_reg *reg)
1010 {
1011 	if (reg->mem_type == TTM_PL_TT)
1012 		return NvDmaTT;
1013 	return chan->vram.handle;
1014 }
1015 
1016 static int
1017 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
1018 		  struct ttm_mem_reg *old_reg, struct ttm_mem_reg *new_reg)
1019 {
1020 	u32 src_offset = old_reg->start << PAGE_SHIFT;
1021 	u32 dst_offset = new_reg->start << PAGE_SHIFT;
1022 	u32 page_count = new_reg->num_pages;
1023 	int ret;
1024 
1025 	ret = RING_SPACE(chan, 3);
1026 	if (ret)
1027 		return ret;
1028 
1029 	BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
1030 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_reg));
1031 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_reg));
1032 
1033 	page_count = new_reg->num_pages;
1034 	while (page_count) {
1035 		int line_count = (page_count > 2047) ? 2047 : page_count;
1036 
1037 		ret = RING_SPACE(chan, 11);
1038 		if (ret)
1039 			return ret;
1040 
1041 		BEGIN_NV04(chan, NvSubCopy,
1042 				 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
1043 		OUT_RING  (chan, src_offset);
1044 		OUT_RING  (chan, dst_offset);
1045 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
1046 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
1047 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
1048 		OUT_RING  (chan, line_count);
1049 		OUT_RING  (chan, 0x00000101);
1050 		OUT_RING  (chan, 0x00000000);
1051 		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
1052 		OUT_RING  (chan, 0);
1053 
1054 		page_count -= line_count;
1055 		src_offset += (PAGE_SIZE * line_count);
1056 		dst_offset += (PAGE_SIZE * line_count);
1057 	}
1058 
1059 	return 0;
1060 }
1061 
1062 static int
1063 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
1064 		     struct ttm_mem_reg *reg)
1065 {
1066 	struct nouveau_mem *old_mem = nouveau_mem(&bo->mem);
1067 	struct nouveau_mem *new_mem = nouveau_mem(reg);
1068 	struct nvif_vmm *vmm = &drm->client.vmm.vmm;
1069 	int ret;
1070 
1071 	ret = nvif_vmm_get(vmm, LAZY, false, old_mem->mem.page, 0,
1072 			   old_mem->mem.size, &old_mem->vma[0]);
1073 	if (ret)
1074 		return ret;
1075 
1076 	ret = nvif_vmm_get(vmm, LAZY, false, new_mem->mem.page, 0,
1077 			   new_mem->mem.size, &old_mem->vma[1]);
1078 	if (ret)
1079 		goto done;
1080 
1081 	ret = nouveau_mem_map(old_mem, vmm, &old_mem->vma[0]);
1082 	if (ret)
1083 		goto done;
1084 
1085 	ret = nouveau_mem_map(new_mem, vmm, &old_mem->vma[1]);
1086 done:
1087 	if (ret) {
1088 		nvif_vmm_put(vmm, &old_mem->vma[1]);
1089 		nvif_vmm_put(vmm, &old_mem->vma[0]);
1090 	}
1091 	return 0;
1092 }
1093 
1094 static int
1095 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
1096 		     bool no_wait_gpu, struct ttm_mem_reg *new_reg)
1097 {
1098 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1099 	struct nouveau_channel *chan = drm->ttm.chan;
1100 	struct nouveau_cli *cli = (void *)chan->user.client;
1101 	struct nouveau_fence *fence;
1102 	int ret;
1103 
1104 	/* create temporary vmas for the transfer and attach them to the
1105 	 * old nvkm_mem node, these will get cleaned up after ttm has
1106 	 * destroyed the ttm_mem_reg
1107 	 */
1108 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
1109 		ret = nouveau_bo_move_prep(drm, bo, new_reg);
1110 		if (ret)
1111 			return ret;
1112 	}
1113 
1114 	mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
1115 	ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, intr);
1116 	if (ret == 0) {
1117 		ret = drm->ttm.move(chan, bo, &bo->mem, new_reg);
1118 		if (ret == 0) {
1119 			ret = nouveau_fence_new(chan, false, &fence);
1120 			if (ret == 0) {
1121 				ret = ttm_bo_move_accel_cleanup(bo,
1122 								&fence->base,
1123 								evict,
1124 								new_reg);
1125 				nouveau_fence_unref(&fence);
1126 			}
1127 		}
1128 	}
1129 	mutex_unlock(&cli->mutex);
1130 	return ret;
1131 }
1132 
1133 void
1134 nouveau_bo_move_init(struct nouveau_drm *drm)
1135 {
1136 	static const struct {
1137 		const char *name;
1138 		int engine;
1139 		s32 oclass;
1140 		int (*exec)(struct nouveau_channel *,
1141 			    struct ttm_buffer_object *,
1142 			    struct ttm_mem_reg *, struct ttm_mem_reg *);
1143 		int (*init)(struct nouveau_channel *, u32 handle);
1144 	} _methods[] = {
1145 		{  "COPY", 4, 0xc1b5, nve0_bo_move_copy, nve0_bo_move_init },
1146 		{  "GRCE", 0, 0xc1b5, nve0_bo_move_copy, nvc0_bo_move_init },
1147 		{  "COPY", 4, 0xc0b5, nve0_bo_move_copy, nve0_bo_move_init },
1148 		{  "GRCE", 0, 0xc0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1149 		{  "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
1150 		{  "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1151 		{  "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
1152 		{  "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1153 		{ "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
1154 		{ "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
1155 		{  "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
1156 		{ "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
1157 		{  "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
1158 		{  "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
1159 		{  "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
1160 		{},
1161 		{ "CRYPT", 0, 0x88b4, nv98_bo_move_exec, nv50_bo_move_init },
1162 	}, *mthd = _methods;
1163 	const char *name = "CPU";
1164 	int ret;
1165 
1166 	do {
1167 		struct nouveau_channel *chan;
1168 
1169 		if (mthd->engine)
1170 			chan = drm->cechan;
1171 		else
1172 			chan = drm->channel;
1173 		if (chan == NULL)
1174 			continue;
1175 
1176 		ret = nvif_object_init(&chan->user,
1177 				       mthd->oclass | (mthd->engine << 16),
1178 				       mthd->oclass, NULL, 0,
1179 				       &drm->ttm.copy);
1180 		if (ret == 0) {
1181 			ret = mthd->init(chan, drm->ttm.copy.handle);
1182 			if (ret) {
1183 				nvif_object_fini(&drm->ttm.copy);
1184 				continue;
1185 			}
1186 
1187 			drm->ttm.move = mthd->exec;
1188 			drm->ttm.chan = chan;
1189 			name = mthd->name;
1190 			break;
1191 		}
1192 	} while ((++mthd)->exec);
1193 
1194 	NV_INFO(drm, "MM: using %s for buffer copies\n", name);
1195 }
1196 
1197 static int
1198 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
1199 		      bool no_wait_gpu, struct ttm_mem_reg *new_reg)
1200 {
1201 	struct ttm_operation_ctx ctx = { intr, no_wait_gpu };
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, &ctx);
1217 	if (ret)
1218 		return ret;
1219 
1220 	ret = ttm_tt_bind(bo->ttm, &tmp_reg, &ctx);
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, &ctx, 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_operation_ctx ctx = { intr, no_wait_gpu };
1239 	struct ttm_place placement_memtype = {
1240 		.fpfn = 0,
1241 		.lpfn = 0,
1242 		.flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
1243 	};
1244 	struct ttm_placement placement;
1245 	struct ttm_mem_reg tmp_reg;
1246 	int ret;
1247 
1248 	placement.num_placement = placement.num_busy_placement = 1;
1249 	placement.placement = placement.busy_placement = &placement_memtype;
1250 
1251 	tmp_reg = *new_reg;
1252 	tmp_reg.mm_node = NULL;
1253 	ret = ttm_bo_mem_space(bo, &placement, &tmp_reg, &ctx);
1254 	if (ret)
1255 		return ret;
1256 
1257 	ret = ttm_bo_move_ttm(bo, &ctx, &tmp_reg);
1258 	if (ret)
1259 		goto out;
1260 
1261 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_reg);
1262 	if (ret)
1263 		goto out;
1264 
1265 out:
1266 	ttm_bo_mem_put(bo, &tmp_reg);
1267 	return ret;
1268 }
1269 
1270 static void
1271 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, bool evict,
1272 		     struct ttm_mem_reg *new_reg)
1273 {
1274 	struct nouveau_mem *mem = new_reg ? nouveau_mem(new_reg) : NULL;
1275 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1276 	struct nouveau_vma *vma;
1277 
1278 	/* ttm can now (stupidly) pass the driver bos it didn't create... */
1279 	if (bo->destroy != nouveau_bo_del_ttm)
1280 		return;
1281 
1282 	if (mem && new_reg->mem_type != TTM_PL_SYSTEM &&
1283 	    mem->mem.page == nvbo->page) {
1284 		list_for_each_entry(vma, &nvbo->vma_list, head) {
1285 			nouveau_vma_map(vma, mem);
1286 		}
1287 	} else {
1288 		list_for_each_entry(vma, &nvbo->vma_list, head) {
1289 			WARN_ON(ttm_bo_wait(bo, false, false));
1290 			nouveau_vma_unmap(vma);
1291 		}
1292 	}
1293 }
1294 
1295 static int
1296 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_reg,
1297 		   struct nouveau_drm_tile **new_tile)
1298 {
1299 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1300 	struct drm_device *dev = drm->dev;
1301 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1302 	u64 offset = new_reg->start << PAGE_SHIFT;
1303 
1304 	*new_tile = NULL;
1305 	if (new_reg->mem_type != TTM_PL_VRAM)
1306 		return 0;
1307 
1308 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
1309 		*new_tile = nv10_bo_set_tiling(dev, offset, new_reg->size,
1310 					       nvbo->mode, nvbo->zeta);
1311 	}
1312 
1313 	return 0;
1314 }
1315 
1316 static void
1317 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1318 		      struct nouveau_drm_tile *new_tile,
1319 		      struct nouveau_drm_tile **old_tile)
1320 {
1321 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1322 	struct drm_device *dev = drm->dev;
1323 	struct dma_fence *fence = reservation_object_get_excl(bo->resv);
1324 
1325 	nv10_bo_put_tile_region(dev, *old_tile, fence);
1326 	*old_tile = new_tile;
1327 }
1328 
1329 static int
1330 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict,
1331 		struct ttm_operation_ctx *ctx,
1332 		struct ttm_mem_reg *new_reg)
1333 {
1334 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1335 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1336 	struct ttm_mem_reg *old_reg = &bo->mem;
1337 	struct nouveau_drm_tile *new_tile = NULL;
1338 	int ret = 0;
1339 
1340 	ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
1341 	if (ret)
1342 		return ret;
1343 
1344 	if (nvbo->pin_refcnt)
1345 		NV_WARN(drm, "Moving pinned object %p!\n", nvbo);
1346 
1347 	if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1348 		ret = nouveau_bo_vm_bind(bo, new_reg, &new_tile);
1349 		if (ret)
1350 			return ret;
1351 	}
1352 
1353 	/* Fake bo copy. */
1354 	if (old_reg->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
1355 		BUG_ON(bo->mem.mm_node != NULL);
1356 		bo->mem = *new_reg;
1357 		new_reg->mm_node = NULL;
1358 		goto out;
1359 	}
1360 
1361 	/* Hardware assisted copy. */
1362 	if (drm->ttm.move) {
1363 		if (new_reg->mem_type == TTM_PL_SYSTEM)
1364 			ret = nouveau_bo_move_flipd(bo, evict,
1365 						    ctx->interruptible,
1366 						    ctx->no_wait_gpu, new_reg);
1367 		else if (old_reg->mem_type == TTM_PL_SYSTEM)
1368 			ret = nouveau_bo_move_flips(bo, evict,
1369 						    ctx->interruptible,
1370 						    ctx->no_wait_gpu, new_reg);
1371 		else
1372 			ret = nouveau_bo_move_m2mf(bo, evict,
1373 						   ctx->interruptible,
1374 						   ctx->no_wait_gpu, new_reg);
1375 		if (!ret)
1376 			goto out;
1377 	}
1378 
1379 	/* Fallback to software copy. */
1380 	ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
1381 	if (ret == 0)
1382 		ret = ttm_bo_move_memcpy(bo, ctx, new_reg);
1383 
1384 out:
1385 	if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1386 		if (ret)
1387 			nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1388 		else
1389 			nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1390 	}
1391 
1392 	return ret;
1393 }
1394 
1395 static int
1396 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1397 {
1398 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1399 
1400 	return drm_vma_node_verify_access(&nvbo->gem.vma_node,
1401 					  filp->private_data);
1402 }
1403 
1404 static int
1405 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *reg)
1406 {
1407 	struct ttm_mem_type_manager *man = &bdev->man[reg->mem_type];
1408 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1409 	struct nvkm_device *device = nvxx_device(&drm->client.device);
1410 	struct nouveau_mem *mem = nouveau_mem(reg);
1411 
1412 	reg->bus.addr = NULL;
1413 	reg->bus.offset = 0;
1414 	reg->bus.size = reg->num_pages << PAGE_SHIFT;
1415 	reg->bus.base = 0;
1416 	reg->bus.is_iomem = false;
1417 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
1418 		return -EINVAL;
1419 	switch (reg->mem_type) {
1420 	case TTM_PL_SYSTEM:
1421 		/* System memory */
1422 		return 0;
1423 	case TTM_PL_TT:
1424 #if IS_ENABLED(CONFIG_AGP)
1425 		if (drm->agp.bridge) {
1426 			reg->bus.offset = reg->start << PAGE_SHIFT;
1427 			reg->bus.base = drm->agp.base;
1428 			reg->bus.is_iomem = !drm->agp.cma;
1429 		}
1430 #endif
1431 		if (drm->client.mem->oclass < NVIF_CLASS_MEM_NV50 || !mem->kind)
1432 			/* untiled */
1433 			break;
1434 		/* fallthrough, tiled memory */
1435 	case TTM_PL_VRAM:
1436 		reg->bus.offset = reg->start << PAGE_SHIFT;
1437 		reg->bus.base = device->func->resource_addr(device, 1);
1438 		reg->bus.is_iomem = true;
1439 		if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1440 			union {
1441 				struct nv50_mem_map_v0 nv50;
1442 				struct gf100_mem_map_v0 gf100;
1443 			} args;
1444 			u64 handle, length;
1445 			u32 argc = 0;
1446 			int ret;
1447 
1448 			switch (mem->mem.object.oclass) {
1449 			case NVIF_CLASS_MEM_NV50:
1450 				args.nv50.version = 0;
1451 				args.nv50.ro = 0;
1452 				args.nv50.kind = mem->kind;
1453 				args.nv50.comp = mem->comp;
1454 				argc = sizeof(args.nv50);
1455 				break;
1456 			case NVIF_CLASS_MEM_GF100:
1457 				args.gf100.version = 0;
1458 				args.gf100.ro = 0;
1459 				args.gf100.kind = mem->kind;
1460 				argc = sizeof(args.gf100);
1461 				break;
1462 			default:
1463 				WARN_ON(1);
1464 				break;
1465 			}
1466 
1467 			ret = nvif_object_map_handle(&mem->mem.object,
1468 						     &args, argc,
1469 						     &handle, &length);
1470 			if (ret != 1)
1471 				return ret ? ret : -EINVAL;
1472 
1473 			reg->bus.base = 0;
1474 			reg->bus.offset = handle;
1475 		}
1476 		break;
1477 	default:
1478 		return -EINVAL;
1479 	}
1480 	return 0;
1481 }
1482 
1483 static void
1484 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *reg)
1485 {
1486 	struct nouveau_drm *drm = nouveau_bdev(bdev);
1487 	struct nouveau_mem *mem = nouveau_mem(reg);
1488 
1489 	if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1490 		switch (reg->mem_type) {
1491 		case TTM_PL_TT:
1492 			if (mem->kind)
1493 				nvif_object_unmap_handle(&mem->mem.object);
1494 			break;
1495 		case TTM_PL_VRAM:
1496 			nvif_object_unmap_handle(&mem->mem.object);
1497 			break;
1498 		default:
1499 			break;
1500 		}
1501 	}
1502 }
1503 
1504 static int
1505 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1506 {
1507 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1508 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1509 	struct nvkm_device *device = nvxx_device(&drm->client.device);
1510 	u32 mappable = device->func->resource_size(device, 1) >> PAGE_SHIFT;
1511 	int i, ret;
1512 
1513 	/* as long as the bo isn't in vram, and isn't tiled, we've got
1514 	 * nothing to do here.
1515 	 */
1516 	if (bo->mem.mem_type != TTM_PL_VRAM) {
1517 		if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1518 		    !nvbo->kind)
1519 			return 0;
1520 
1521 		if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1522 			nouveau_bo_placement_set(nvbo, TTM_PL_TT, 0);
1523 
1524 			ret = nouveau_bo_validate(nvbo, false, false);
1525 			if (ret)
1526 				return ret;
1527 		}
1528 		return 0;
1529 	}
1530 
1531 	/* make sure bo is in mappable vram */
1532 	if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1533 	    bo->mem.start + bo->mem.num_pages < mappable)
1534 		return 0;
1535 
1536 	for (i = 0; i < nvbo->placement.num_placement; ++i) {
1537 		nvbo->placements[i].fpfn = 0;
1538 		nvbo->placements[i].lpfn = mappable;
1539 	}
1540 
1541 	for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
1542 		nvbo->busy_placements[i].fpfn = 0;
1543 		nvbo->busy_placements[i].lpfn = mappable;
1544 	}
1545 
1546 	nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_VRAM, 0);
1547 	return nouveau_bo_validate(nvbo, false, false);
1548 }
1549 
1550 static int
1551 nouveau_ttm_tt_populate(struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1552 {
1553 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1554 	struct nouveau_drm *drm;
1555 	struct device *dev;
1556 	unsigned i;
1557 	int r;
1558 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1559 
1560 	if (ttm->state != tt_unpopulated)
1561 		return 0;
1562 
1563 	if (slave && ttm->sg) {
1564 		/* make userspace faulting work */
1565 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1566 						 ttm_dma->dma_address, ttm->num_pages);
1567 		ttm->state = tt_unbound;
1568 		return 0;
1569 	}
1570 
1571 	drm = nouveau_bdev(ttm->bdev);
1572 	dev = drm->dev->dev;
1573 
1574 #if IS_ENABLED(CONFIG_AGP)
1575 	if (drm->agp.bridge) {
1576 		return ttm_agp_tt_populate(ttm, ctx);
1577 	}
1578 #endif
1579 
1580 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1581 	if (swiotlb_nr_tbl()) {
1582 		return ttm_dma_populate((void *)ttm, dev, ctx);
1583 	}
1584 #endif
1585 
1586 	r = ttm_pool_populate(ttm, ctx);
1587 	if (r) {
1588 		return r;
1589 	}
1590 
1591 	for (i = 0; i < ttm->num_pages; i++) {
1592 		dma_addr_t addr;
1593 
1594 		addr = dma_map_page(dev, ttm->pages[i], 0, PAGE_SIZE,
1595 				    DMA_BIDIRECTIONAL);
1596 
1597 		if (dma_mapping_error(dev, addr)) {
1598 			while (i--) {
1599 				dma_unmap_page(dev, ttm_dma->dma_address[i],
1600 					       PAGE_SIZE, DMA_BIDIRECTIONAL);
1601 				ttm_dma->dma_address[i] = 0;
1602 			}
1603 			ttm_pool_unpopulate(ttm);
1604 			return -EFAULT;
1605 		}
1606 
1607 		ttm_dma->dma_address[i] = addr;
1608 	}
1609 	return 0;
1610 }
1611 
1612 static void
1613 nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
1614 {
1615 	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1616 	struct nouveau_drm *drm;
1617 	struct device *dev;
1618 	unsigned i;
1619 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1620 
1621 	if (slave)
1622 		return;
1623 
1624 	drm = nouveau_bdev(ttm->bdev);
1625 	dev = drm->dev->dev;
1626 
1627 #if IS_ENABLED(CONFIG_AGP)
1628 	if (drm->agp.bridge) {
1629 		ttm_agp_tt_unpopulate(ttm);
1630 		return;
1631 	}
1632 #endif
1633 
1634 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1635 	if (swiotlb_nr_tbl()) {
1636 		ttm_dma_unpopulate((void *)ttm, dev);
1637 		return;
1638 	}
1639 #endif
1640 
1641 	for (i = 0; i < ttm->num_pages; i++) {
1642 		if (ttm_dma->dma_address[i]) {
1643 			dma_unmap_page(dev, ttm_dma->dma_address[i], PAGE_SIZE,
1644 				       DMA_BIDIRECTIONAL);
1645 		}
1646 	}
1647 
1648 	ttm_pool_unpopulate(ttm);
1649 }
1650 
1651 void
1652 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1653 {
1654 	struct reservation_object *resv = nvbo->bo.resv;
1655 
1656 	if (exclusive)
1657 		reservation_object_add_excl_fence(resv, &fence->base);
1658 	else if (fence)
1659 		reservation_object_add_shared_fence(resv, &fence->base);
1660 }
1661 
1662 struct ttm_bo_driver nouveau_bo_driver = {
1663 	.ttm_tt_create = &nouveau_ttm_tt_create,
1664 	.ttm_tt_populate = &nouveau_ttm_tt_populate,
1665 	.ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1666 	.invalidate_caches = nouveau_bo_invalidate_caches,
1667 	.init_mem_type = nouveau_bo_init_mem_type,
1668 	.eviction_valuable = ttm_bo_eviction_valuable,
1669 	.evict_flags = nouveau_bo_evict_flags,
1670 	.move_notify = nouveau_bo_move_ntfy,
1671 	.move = nouveau_bo_move,
1672 	.verify_access = nouveau_bo_verify_access,
1673 	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1674 	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1675 	.io_mem_free = &nouveau_ttm_io_mem_free,
1676 };
1677