xref: /openbmc/linux/drivers/gpu/drm/nouveau/nouveau_bo.c (revision baa7eb025ab14f3cba2e35c0a8648f9c9f01d24f)
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 "drmP.h"
31 
32 #include "nouveau_drm.h"
33 #include "nouveau_drv.h"
34 #include "nouveau_dma.h"
35 
36 #include <linux/log2.h>
37 #include <linux/slab.h>
38 
39 static void
40 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
41 {
42 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
43 	struct drm_device *dev = dev_priv->dev;
44 	struct nouveau_bo *nvbo = nouveau_bo(bo);
45 
46 	if (unlikely(nvbo->gem))
47 		DRM_ERROR("bo %p still attached to GEM object\n", bo);
48 
49 	if (nvbo->tile)
50 		nv10_mem_expire_tiling(dev, nvbo->tile, NULL);
51 
52 	kfree(nvbo);
53 }
54 
55 static void
56 nouveau_bo_fixup_align(struct drm_device *dev,
57 		       uint32_t tile_mode, uint32_t tile_flags,
58 		       int *align, int *size)
59 {
60 	struct drm_nouveau_private *dev_priv = dev->dev_private;
61 
62 	/*
63 	 * Some of the tile_flags have a periodic structure of N*4096 bytes,
64 	 * align to to that as well as the page size. Align the size to the
65 	 * appropriate boundaries. This does imply that sizes are rounded up
66 	 * 3-7 pages, so be aware of this and do not waste memory by allocating
67 	 * many small buffers.
68 	 */
69 	if (dev_priv->card_type == NV_50) {
70 		uint32_t block_size = dev_priv->vram_size >> 15;
71 		int i;
72 
73 		switch (tile_flags) {
74 		case 0x1800:
75 		case 0x2800:
76 		case 0x4800:
77 		case 0x7a00:
78 			if (is_power_of_2(block_size)) {
79 				for (i = 1; i < 10; i++) {
80 					*align = 12 * i * block_size;
81 					if (!(*align % 65536))
82 						break;
83 				}
84 			} else {
85 				for (i = 1; i < 10; i++) {
86 					*align = 8 * i * block_size;
87 					if (!(*align % 65536))
88 						break;
89 				}
90 			}
91 			*size = roundup(*size, *align);
92 			break;
93 		default:
94 			break;
95 		}
96 
97 	} else {
98 		if (tile_mode) {
99 			if (dev_priv->chipset >= 0x40) {
100 				*align = 65536;
101 				*size = roundup(*size, 64 * tile_mode);
102 
103 			} else if (dev_priv->chipset >= 0x30) {
104 				*align = 32768;
105 				*size = roundup(*size, 64 * tile_mode);
106 
107 			} else if (dev_priv->chipset >= 0x20) {
108 				*align = 16384;
109 				*size = roundup(*size, 64 * tile_mode);
110 
111 			} else if (dev_priv->chipset >= 0x10) {
112 				*align = 16384;
113 				*size = roundup(*size, 32 * tile_mode);
114 			}
115 		}
116 	}
117 
118 	/* ALIGN works only on powers of two. */
119 	*size = roundup(*size, PAGE_SIZE);
120 
121 	if (dev_priv->card_type == NV_50) {
122 		*size = roundup(*size, 65536);
123 		*align = max(65536, *align);
124 	}
125 }
126 
127 int
128 nouveau_bo_new(struct drm_device *dev, struct nouveau_channel *chan,
129 	       int size, int align, uint32_t flags, uint32_t tile_mode,
130 	       uint32_t tile_flags, bool no_vm, bool mappable,
131 	       struct nouveau_bo **pnvbo)
132 {
133 	struct drm_nouveau_private *dev_priv = dev->dev_private;
134 	struct nouveau_bo *nvbo;
135 	int ret = 0;
136 
137 	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
138 	if (!nvbo)
139 		return -ENOMEM;
140 	INIT_LIST_HEAD(&nvbo->head);
141 	INIT_LIST_HEAD(&nvbo->entry);
142 	nvbo->mappable = mappable;
143 	nvbo->no_vm = no_vm;
144 	nvbo->tile_mode = tile_mode;
145 	nvbo->tile_flags = tile_flags;
146 	nvbo->bo.bdev = &dev_priv->ttm.bdev;
147 
148 	nouveau_bo_fixup_align(dev, tile_mode, nouveau_bo_tile_layout(nvbo),
149 			       &align, &size);
150 	align >>= PAGE_SHIFT;
151 
152 	nouveau_bo_placement_set(nvbo, flags, 0);
153 
154 	nvbo->channel = chan;
155 	ret = ttm_bo_init(&dev_priv->ttm.bdev, &nvbo->bo, size,
156 			  ttm_bo_type_device, &nvbo->placement, align, 0,
157 			  false, NULL, size, nouveau_bo_del_ttm);
158 	if (ret) {
159 		/* ttm will call nouveau_bo_del_ttm if it fails.. */
160 		return ret;
161 	}
162 	nvbo->channel = NULL;
163 
164 	*pnvbo = nvbo;
165 	return 0;
166 }
167 
168 static void
169 set_placement_list(uint32_t *pl, unsigned *n, uint32_t type, uint32_t flags)
170 {
171 	*n = 0;
172 
173 	if (type & TTM_PL_FLAG_VRAM)
174 		pl[(*n)++] = TTM_PL_FLAG_VRAM | flags;
175 	if (type & TTM_PL_FLAG_TT)
176 		pl[(*n)++] = TTM_PL_FLAG_TT | flags;
177 	if (type & TTM_PL_FLAG_SYSTEM)
178 		pl[(*n)++] = TTM_PL_FLAG_SYSTEM | flags;
179 }
180 
181 static void
182 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
183 {
184 	struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
185 
186 	if (dev_priv->card_type == NV_10 &&
187 	    nvbo->tile_mode && (type & TTM_PL_FLAG_VRAM)) {
188 		/*
189 		 * Make sure that the color and depth buffers are handled
190 		 * by independent memory controller units. Up to a 9x
191 		 * speed up when alpha-blending and depth-test are enabled
192 		 * at the same time.
193 		 */
194 		int vram_pages = dev_priv->vram_size >> PAGE_SHIFT;
195 
196 		if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
197 			nvbo->placement.fpfn = vram_pages / 2;
198 			nvbo->placement.lpfn = ~0;
199 		} else {
200 			nvbo->placement.fpfn = 0;
201 			nvbo->placement.lpfn = vram_pages / 2;
202 		}
203 	}
204 }
205 
206 void
207 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
208 {
209 	struct ttm_placement *pl = &nvbo->placement;
210 	uint32_t flags = TTM_PL_MASK_CACHING |
211 		(nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
212 
213 	pl->placement = nvbo->placements;
214 	set_placement_list(nvbo->placements, &pl->num_placement,
215 			   type, flags);
216 
217 	pl->busy_placement = nvbo->busy_placements;
218 	set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
219 			   type | busy, flags);
220 
221 	set_placement_range(nvbo, type);
222 }
223 
224 int
225 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype)
226 {
227 	struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
228 	struct ttm_buffer_object *bo = &nvbo->bo;
229 	int ret;
230 
231 	if (nvbo->pin_refcnt && !(memtype & (1 << bo->mem.mem_type))) {
232 		NV_ERROR(nouveau_bdev(bo->bdev)->dev,
233 			 "bo %p pinned elsewhere: 0x%08x vs 0x%08x\n", bo,
234 			 1 << bo->mem.mem_type, memtype);
235 		return -EINVAL;
236 	}
237 
238 	if (nvbo->pin_refcnt++)
239 		return 0;
240 
241 	ret = ttm_bo_reserve(bo, false, false, false, 0);
242 	if (ret)
243 		goto out;
244 
245 	nouveau_bo_placement_set(nvbo, memtype, 0);
246 
247 	ret = ttm_bo_validate(bo, &nvbo->placement, false, false, false);
248 	if (ret == 0) {
249 		switch (bo->mem.mem_type) {
250 		case TTM_PL_VRAM:
251 			dev_priv->fb_aper_free -= bo->mem.size;
252 			break;
253 		case TTM_PL_TT:
254 			dev_priv->gart_info.aper_free -= bo->mem.size;
255 			break;
256 		default:
257 			break;
258 		}
259 	}
260 	ttm_bo_unreserve(bo);
261 out:
262 	if (unlikely(ret))
263 		nvbo->pin_refcnt--;
264 	return ret;
265 }
266 
267 int
268 nouveau_bo_unpin(struct nouveau_bo *nvbo)
269 {
270 	struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
271 	struct ttm_buffer_object *bo = &nvbo->bo;
272 	int ret;
273 
274 	if (--nvbo->pin_refcnt)
275 		return 0;
276 
277 	ret = ttm_bo_reserve(bo, false, false, false, 0);
278 	if (ret)
279 		return ret;
280 
281 	nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
282 
283 	ret = ttm_bo_validate(bo, &nvbo->placement, false, false, false);
284 	if (ret == 0) {
285 		switch (bo->mem.mem_type) {
286 		case TTM_PL_VRAM:
287 			dev_priv->fb_aper_free += bo->mem.size;
288 			break;
289 		case TTM_PL_TT:
290 			dev_priv->gart_info.aper_free += bo->mem.size;
291 			break;
292 		default:
293 			break;
294 		}
295 	}
296 
297 	ttm_bo_unreserve(bo);
298 	return ret;
299 }
300 
301 int
302 nouveau_bo_map(struct nouveau_bo *nvbo)
303 {
304 	int ret;
305 
306 	ret = ttm_bo_reserve(&nvbo->bo, false, false, false, 0);
307 	if (ret)
308 		return ret;
309 
310 	ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
311 	ttm_bo_unreserve(&nvbo->bo);
312 	return ret;
313 }
314 
315 void
316 nouveau_bo_unmap(struct nouveau_bo *nvbo)
317 {
318 	if (nvbo)
319 		ttm_bo_kunmap(&nvbo->kmap);
320 }
321 
322 u16
323 nouveau_bo_rd16(struct nouveau_bo *nvbo, unsigned index)
324 {
325 	bool is_iomem;
326 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
327 	mem = &mem[index];
328 	if (is_iomem)
329 		return ioread16_native((void __force __iomem *)mem);
330 	else
331 		return *mem;
332 }
333 
334 void
335 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
336 {
337 	bool is_iomem;
338 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
339 	mem = &mem[index];
340 	if (is_iomem)
341 		iowrite16_native(val, (void __force __iomem *)mem);
342 	else
343 		*mem = val;
344 }
345 
346 u32
347 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
348 {
349 	bool is_iomem;
350 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
351 	mem = &mem[index];
352 	if (is_iomem)
353 		return ioread32_native((void __force __iomem *)mem);
354 	else
355 		return *mem;
356 }
357 
358 void
359 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
360 {
361 	bool is_iomem;
362 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
363 	mem = &mem[index];
364 	if (is_iomem)
365 		iowrite32_native(val, (void __force __iomem *)mem);
366 	else
367 		*mem = val;
368 }
369 
370 static struct ttm_backend *
371 nouveau_bo_create_ttm_backend_entry(struct ttm_bo_device *bdev)
372 {
373 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
374 	struct drm_device *dev = dev_priv->dev;
375 
376 	switch (dev_priv->gart_info.type) {
377 #if __OS_HAS_AGP
378 	case NOUVEAU_GART_AGP:
379 		return ttm_agp_backend_init(bdev, dev->agp->bridge);
380 #endif
381 	case NOUVEAU_GART_SGDMA:
382 		return nouveau_sgdma_init_ttm(dev);
383 	default:
384 		NV_ERROR(dev, "Unknown GART type %d\n",
385 			 dev_priv->gart_info.type);
386 		break;
387 	}
388 
389 	return NULL;
390 }
391 
392 static int
393 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
394 {
395 	/* We'll do this from user space. */
396 	return 0;
397 }
398 
399 static int
400 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
401 			 struct ttm_mem_type_manager *man)
402 {
403 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
404 	struct drm_device *dev = dev_priv->dev;
405 
406 	switch (type) {
407 	case TTM_PL_SYSTEM:
408 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
409 		man->available_caching = TTM_PL_MASK_CACHING;
410 		man->default_caching = TTM_PL_FLAG_CACHED;
411 		break;
412 	case TTM_PL_VRAM:
413 		man->func = &ttm_bo_manager_func;
414 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
415 			     TTM_MEMTYPE_FLAG_MAPPABLE;
416 		man->available_caching = TTM_PL_FLAG_UNCACHED |
417 					 TTM_PL_FLAG_WC;
418 		man->default_caching = TTM_PL_FLAG_WC;
419 		if (dev_priv->card_type == NV_50)
420 			man->gpu_offset = 0x40000000;
421 		else
422 			man->gpu_offset = 0;
423 		break;
424 	case TTM_PL_TT:
425 		man->func = &ttm_bo_manager_func;
426 		switch (dev_priv->gart_info.type) {
427 		case NOUVEAU_GART_AGP:
428 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
429 			man->available_caching = TTM_PL_FLAG_UNCACHED;
430 			man->default_caching = TTM_PL_FLAG_UNCACHED;
431 			break;
432 		case NOUVEAU_GART_SGDMA:
433 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
434 				     TTM_MEMTYPE_FLAG_CMA;
435 			man->available_caching = TTM_PL_MASK_CACHING;
436 			man->default_caching = TTM_PL_FLAG_CACHED;
437 			break;
438 		default:
439 			NV_ERROR(dev, "Unknown GART type: %d\n",
440 				 dev_priv->gart_info.type);
441 			return -EINVAL;
442 		}
443 		man->gpu_offset = dev_priv->vm_gart_base;
444 		break;
445 	default:
446 		NV_ERROR(dev, "Unsupported memory type %u\n", (unsigned)type);
447 		return -EINVAL;
448 	}
449 	return 0;
450 }
451 
452 static void
453 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
454 {
455 	struct nouveau_bo *nvbo = nouveau_bo(bo);
456 
457 	switch (bo->mem.mem_type) {
458 	case TTM_PL_VRAM:
459 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
460 					 TTM_PL_FLAG_SYSTEM);
461 		break;
462 	default:
463 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
464 		break;
465 	}
466 
467 	*pl = nvbo->placement;
468 }
469 
470 
471 /* GPU-assisted copy using NV_MEMORY_TO_MEMORY_FORMAT, can access
472  * TTM_PL_{VRAM,TT} directly.
473  */
474 
475 static int
476 nouveau_bo_move_accel_cleanup(struct nouveau_channel *chan,
477 			      struct nouveau_bo *nvbo, bool evict,
478 			      bool no_wait_reserve, bool no_wait_gpu,
479 			      struct ttm_mem_reg *new_mem)
480 {
481 	struct nouveau_fence *fence = NULL;
482 	int ret;
483 
484 	ret = nouveau_fence_new(chan, &fence, true);
485 	if (ret)
486 		return ret;
487 
488 	if (nvbo->channel) {
489 		ret = nouveau_fence_sync(fence, nvbo->channel);
490 		if (ret)
491 			goto out;
492 	}
493 
494 	ret = ttm_bo_move_accel_cleanup(&nvbo->bo, fence, NULL, evict,
495 					no_wait_reserve, no_wait_gpu, new_mem);
496 out:
497 	nouveau_fence_unref((void *)&fence);
498 	return ret;
499 }
500 
501 static inline uint32_t
502 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
503 		      struct nouveau_channel *chan, struct ttm_mem_reg *mem)
504 {
505 	struct nouveau_bo *nvbo = nouveau_bo(bo);
506 
507 	if (nvbo->no_vm) {
508 		if (mem->mem_type == TTM_PL_TT)
509 			return NvDmaGART;
510 		return NvDmaVRAM;
511 	}
512 
513 	if (mem->mem_type == TTM_PL_TT)
514 		return chan->gart_handle;
515 	return chan->vram_handle;
516 }
517 
518 static int
519 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
520 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
521 {
522 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
523 	struct nouveau_bo *nvbo = nouveau_bo(bo);
524 	u64 length = (new_mem->num_pages << PAGE_SHIFT);
525 	u64 src_offset, dst_offset;
526 	int ret;
527 
528 	src_offset = old_mem->start << PAGE_SHIFT;
529 	dst_offset = new_mem->start << PAGE_SHIFT;
530 	if (!nvbo->no_vm) {
531 		if (old_mem->mem_type == TTM_PL_VRAM)
532 			src_offset += dev_priv->vm_vram_base;
533 		else
534 			src_offset += dev_priv->vm_gart_base;
535 
536 		if (new_mem->mem_type == TTM_PL_VRAM)
537 			dst_offset += dev_priv->vm_vram_base;
538 		else
539 			dst_offset += dev_priv->vm_gart_base;
540 	}
541 
542 	ret = RING_SPACE(chan, 3);
543 	if (ret)
544 		return ret;
545 
546 	BEGIN_RING(chan, NvSubM2MF, 0x0184, 2);
547 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
548 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
549 
550 	while (length) {
551 		u32 amount, stride, height;
552 
553 		amount  = min(length, (u64)(4 * 1024 * 1024));
554 		stride  = 16 * 4;
555 		height  = amount / stride;
556 
557 		if (new_mem->mem_type == TTM_PL_VRAM &&
558 		    nouveau_bo_tile_layout(nvbo)) {
559 			ret = RING_SPACE(chan, 8);
560 			if (ret)
561 				return ret;
562 
563 			BEGIN_RING(chan, NvSubM2MF, 0x0200, 7);
564 			OUT_RING  (chan, 0);
565 			OUT_RING  (chan, 0);
566 			OUT_RING  (chan, stride);
567 			OUT_RING  (chan, height);
568 			OUT_RING  (chan, 1);
569 			OUT_RING  (chan, 0);
570 			OUT_RING  (chan, 0);
571 		} else {
572 			ret = RING_SPACE(chan, 2);
573 			if (ret)
574 				return ret;
575 
576 			BEGIN_RING(chan, NvSubM2MF, 0x0200, 1);
577 			OUT_RING  (chan, 1);
578 		}
579 		if (old_mem->mem_type == TTM_PL_VRAM &&
580 		    nouveau_bo_tile_layout(nvbo)) {
581 			ret = RING_SPACE(chan, 8);
582 			if (ret)
583 				return ret;
584 
585 			BEGIN_RING(chan, NvSubM2MF, 0x021c, 7);
586 			OUT_RING  (chan, 0);
587 			OUT_RING  (chan, 0);
588 			OUT_RING  (chan, stride);
589 			OUT_RING  (chan, height);
590 			OUT_RING  (chan, 1);
591 			OUT_RING  (chan, 0);
592 			OUT_RING  (chan, 0);
593 		} else {
594 			ret = RING_SPACE(chan, 2);
595 			if (ret)
596 				return ret;
597 
598 			BEGIN_RING(chan, NvSubM2MF, 0x021c, 1);
599 			OUT_RING  (chan, 1);
600 		}
601 
602 		ret = RING_SPACE(chan, 14);
603 		if (ret)
604 			return ret;
605 
606 		BEGIN_RING(chan, NvSubM2MF, 0x0238, 2);
607 		OUT_RING  (chan, upper_32_bits(src_offset));
608 		OUT_RING  (chan, upper_32_bits(dst_offset));
609 		BEGIN_RING(chan, NvSubM2MF, 0x030c, 8);
610 		OUT_RING  (chan, lower_32_bits(src_offset));
611 		OUT_RING  (chan, lower_32_bits(dst_offset));
612 		OUT_RING  (chan, stride);
613 		OUT_RING  (chan, stride);
614 		OUT_RING  (chan, stride);
615 		OUT_RING  (chan, height);
616 		OUT_RING  (chan, 0x00000101);
617 		OUT_RING  (chan, 0x00000000);
618 		BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
619 		OUT_RING  (chan, 0);
620 
621 		length -= amount;
622 		src_offset += amount;
623 		dst_offset += amount;
624 	}
625 
626 	return 0;
627 }
628 
629 static int
630 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
631 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
632 {
633 	u32 src_offset = old_mem->start << PAGE_SHIFT;
634 	u32 dst_offset = new_mem->start << PAGE_SHIFT;
635 	u32 page_count = new_mem->num_pages;
636 	int ret;
637 
638 	ret = RING_SPACE(chan, 3);
639 	if (ret)
640 		return ret;
641 
642 	BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
643 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
644 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
645 
646 	page_count = new_mem->num_pages;
647 	while (page_count) {
648 		int line_count = (page_count > 2047) ? 2047 : page_count;
649 
650 		ret = RING_SPACE(chan, 11);
651 		if (ret)
652 			return ret;
653 
654 		BEGIN_RING(chan, NvSubM2MF,
655 				 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
656 		OUT_RING  (chan, src_offset);
657 		OUT_RING  (chan, dst_offset);
658 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
659 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
660 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
661 		OUT_RING  (chan, line_count);
662 		OUT_RING  (chan, 0x00000101);
663 		OUT_RING  (chan, 0x00000000);
664 		BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
665 		OUT_RING  (chan, 0);
666 
667 		page_count -= line_count;
668 		src_offset += (PAGE_SIZE * line_count);
669 		dst_offset += (PAGE_SIZE * line_count);
670 	}
671 
672 	return 0;
673 }
674 
675 static int
676 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
677 		     bool no_wait_reserve, bool no_wait_gpu,
678 		     struct ttm_mem_reg *new_mem)
679 {
680 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
681 	struct nouveau_bo *nvbo = nouveau_bo(bo);
682 	struct nouveau_channel *chan;
683 	int ret;
684 
685 	chan = nvbo->channel;
686 	if (!chan || nvbo->no_vm)
687 		chan = dev_priv->channel;
688 
689 	if (dev_priv->card_type < NV_50)
690 		ret = nv04_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
691 	else
692 		ret = nv50_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
693 	if (ret)
694 		return ret;
695 
696 	return nouveau_bo_move_accel_cleanup(chan, nvbo, evict, no_wait_reserve, no_wait_gpu, new_mem);
697 }
698 
699 static int
700 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
701 		      bool no_wait_reserve, bool no_wait_gpu,
702 		      struct ttm_mem_reg *new_mem)
703 {
704 	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
705 	struct ttm_placement placement;
706 	struct ttm_mem_reg tmp_mem;
707 	int ret;
708 
709 	placement.fpfn = placement.lpfn = 0;
710 	placement.num_placement = placement.num_busy_placement = 1;
711 	placement.placement = placement.busy_placement = &placement_memtype;
712 
713 	tmp_mem = *new_mem;
714 	tmp_mem.mm_node = NULL;
715 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_reserve, no_wait_gpu);
716 	if (ret)
717 		return ret;
718 
719 	ret = ttm_tt_bind(bo->ttm, &tmp_mem);
720 	if (ret)
721 		goto out;
722 
723 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_reserve, no_wait_gpu, &tmp_mem);
724 	if (ret)
725 		goto out;
726 
727 	ret = ttm_bo_move_ttm(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
728 out:
729 	ttm_bo_mem_put(bo, &tmp_mem);
730 	return ret;
731 }
732 
733 static int
734 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
735 		      bool no_wait_reserve, bool no_wait_gpu,
736 		      struct ttm_mem_reg *new_mem)
737 {
738 	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
739 	struct ttm_placement placement;
740 	struct ttm_mem_reg tmp_mem;
741 	int ret;
742 
743 	placement.fpfn = placement.lpfn = 0;
744 	placement.num_placement = placement.num_busy_placement = 1;
745 	placement.placement = placement.busy_placement = &placement_memtype;
746 
747 	tmp_mem = *new_mem;
748 	tmp_mem.mm_node = NULL;
749 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_reserve, no_wait_gpu);
750 	if (ret)
751 		return ret;
752 
753 	ret = ttm_bo_move_ttm(bo, evict, no_wait_reserve, no_wait_gpu, &tmp_mem);
754 	if (ret)
755 		goto out;
756 
757 	ret = nouveau_bo_move_m2mf(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
758 	if (ret)
759 		goto out;
760 
761 out:
762 	ttm_bo_mem_put(bo, &tmp_mem);
763 	return ret;
764 }
765 
766 static int
767 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
768 		   struct nouveau_tile_reg **new_tile)
769 {
770 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
771 	struct drm_device *dev = dev_priv->dev;
772 	struct nouveau_bo *nvbo = nouveau_bo(bo);
773 	uint64_t offset;
774 	int ret;
775 
776 	if (nvbo->no_vm || new_mem->mem_type != TTM_PL_VRAM) {
777 		/* Nothing to do. */
778 		*new_tile = NULL;
779 		return 0;
780 	}
781 
782 	offset = new_mem->start << PAGE_SHIFT;
783 
784 	if (dev_priv->card_type == NV_50) {
785 		ret = nv50_mem_vm_bind_linear(dev,
786 					      offset + dev_priv->vm_vram_base,
787 					      new_mem->size,
788 					      nouveau_bo_tile_layout(nvbo),
789 					      offset);
790 		if (ret)
791 			return ret;
792 
793 	} else if (dev_priv->card_type >= NV_10) {
794 		*new_tile = nv10_mem_set_tiling(dev, offset, new_mem->size,
795 						nvbo->tile_mode);
796 	}
797 
798 	return 0;
799 }
800 
801 static void
802 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
803 		      struct nouveau_tile_reg *new_tile,
804 		      struct nouveau_tile_reg **old_tile)
805 {
806 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
807 	struct drm_device *dev = dev_priv->dev;
808 
809 	if (dev_priv->card_type >= NV_10 &&
810 	    dev_priv->card_type < NV_50) {
811 		if (*old_tile)
812 			nv10_mem_expire_tiling(dev, *old_tile, bo->sync_obj);
813 
814 		*old_tile = new_tile;
815 	}
816 }
817 
818 static int
819 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
820 		bool no_wait_reserve, bool no_wait_gpu,
821 		struct ttm_mem_reg *new_mem)
822 {
823 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
824 	struct nouveau_bo *nvbo = nouveau_bo(bo);
825 	struct ttm_mem_reg *old_mem = &bo->mem;
826 	struct nouveau_tile_reg *new_tile = NULL;
827 	int ret = 0;
828 
829 	ret = nouveau_bo_vm_bind(bo, new_mem, &new_tile);
830 	if (ret)
831 		return ret;
832 
833 	/* Fake bo copy. */
834 	if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
835 		BUG_ON(bo->mem.mm_node != NULL);
836 		bo->mem = *new_mem;
837 		new_mem->mm_node = NULL;
838 		goto out;
839 	}
840 
841 	/* Software copy if the card isn't up and running yet. */
842 	if (!dev_priv->channel) {
843 		ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
844 		goto out;
845 	}
846 
847 	/* Hardware assisted copy. */
848 	if (new_mem->mem_type == TTM_PL_SYSTEM)
849 		ret = nouveau_bo_move_flipd(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
850 	else if (old_mem->mem_type == TTM_PL_SYSTEM)
851 		ret = nouveau_bo_move_flips(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
852 	else
853 		ret = nouveau_bo_move_m2mf(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
854 
855 	if (!ret)
856 		goto out;
857 
858 	/* Fallback to software copy. */
859 	ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
860 
861 out:
862 	if (ret)
863 		nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
864 	else
865 		nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
866 
867 	return ret;
868 }
869 
870 static int
871 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
872 {
873 	return 0;
874 }
875 
876 static int
877 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
878 {
879 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
880 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
881 	struct drm_device *dev = dev_priv->dev;
882 
883 	mem->bus.addr = NULL;
884 	mem->bus.offset = 0;
885 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
886 	mem->bus.base = 0;
887 	mem->bus.is_iomem = false;
888 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
889 		return -EINVAL;
890 	switch (mem->mem_type) {
891 	case TTM_PL_SYSTEM:
892 		/* System memory */
893 		return 0;
894 	case TTM_PL_TT:
895 #if __OS_HAS_AGP
896 		if (dev_priv->gart_info.type == NOUVEAU_GART_AGP) {
897 			mem->bus.offset = mem->start << PAGE_SHIFT;
898 			mem->bus.base = dev_priv->gart_info.aper_base;
899 			mem->bus.is_iomem = true;
900 		}
901 #endif
902 		break;
903 	case TTM_PL_VRAM:
904 		mem->bus.offset = mem->start << PAGE_SHIFT;
905 		mem->bus.base = pci_resource_start(dev->pdev, 1);
906 		mem->bus.is_iomem = true;
907 		break;
908 	default:
909 		return -EINVAL;
910 	}
911 	return 0;
912 }
913 
914 static void
915 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
916 {
917 }
918 
919 static int
920 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
921 {
922 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
923 	struct nouveau_bo *nvbo = nouveau_bo(bo);
924 
925 	/* as long as the bo isn't in vram, and isn't tiled, we've got
926 	 * nothing to do here.
927 	 */
928 	if (bo->mem.mem_type != TTM_PL_VRAM) {
929 		if (dev_priv->card_type < NV_50 ||
930 		    !nouveau_bo_tile_layout(nvbo))
931 			return 0;
932 	}
933 
934 	/* make sure bo is in mappable vram */
935 	if (bo->mem.start + bo->mem.num_pages < dev_priv->fb_mappable_pages)
936 		return 0;
937 
938 
939 	nvbo->placement.fpfn = 0;
940 	nvbo->placement.lpfn = dev_priv->fb_mappable_pages;
941 	nouveau_bo_placement_set(nvbo, TTM_PL_VRAM, 0);
942 	return ttm_bo_validate(bo, &nvbo->placement, false, true, false);
943 }
944 
945 struct ttm_bo_driver nouveau_bo_driver = {
946 	.create_ttm_backend_entry = nouveau_bo_create_ttm_backend_entry,
947 	.invalidate_caches = nouveau_bo_invalidate_caches,
948 	.init_mem_type = nouveau_bo_init_mem_type,
949 	.evict_flags = nouveau_bo_evict_flags,
950 	.move = nouveau_bo_move,
951 	.verify_access = nouveau_bo_verify_access,
952 	.sync_obj_signaled = nouveau_fence_signalled,
953 	.sync_obj_wait = nouveau_fence_wait,
954 	.sync_obj_flush = nouveau_fence_flush,
955 	.sync_obj_unref = nouveau_fence_unref,
956 	.sync_obj_ref = nouveau_fence_ref,
957 	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
958 	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
959 	.io_mem_free = &nouveau_ttm_io_mem_free,
960 };
961 
962