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 #include "nouveau_mm.h"
36 #include "nouveau_vm.h"
37 
38 #include <linux/log2.h>
39 #include <linux/slab.h>
40 
41 static void
42 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
43 {
44 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
45 	struct drm_device *dev = dev_priv->dev;
46 	struct nouveau_bo *nvbo = nouveau_bo(bo);
47 
48 	if (unlikely(nvbo->gem))
49 		DRM_ERROR("bo %p still attached to GEM object\n", bo);
50 
51 	nv10_mem_put_tile_region(dev, nvbo->tile, NULL);
52 	if (nvbo->vma.node) {
53 		nouveau_vm_unmap(&nvbo->vma);
54 		nouveau_vm_put(&nvbo->vma);
55 	}
56 	kfree(nvbo);
57 }
58 
59 static void
60 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, int *align, int *size,
61 		       int *page_shift)
62 {
63 	struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
64 
65 	if (dev_priv->card_type < NV_50) {
66 		if (nvbo->tile_mode) {
67 			if (dev_priv->chipset >= 0x40) {
68 				*align = 65536;
69 				*size = roundup(*size, 64 * nvbo->tile_mode);
70 
71 			} else if (dev_priv->chipset >= 0x30) {
72 				*align = 32768;
73 				*size = roundup(*size, 64 * nvbo->tile_mode);
74 
75 			} else if (dev_priv->chipset >= 0x20) {
76 				*align = 16384;
77 				*size = roundup(*size, 64 * nvbo->tile_mode);
78 
79 			} else if (dev_priv->chipset >= 0x10) {
80 				*align = 16384;
81 				*size = roundup(*size, 32 * nvbo->tile_mode);
82 			}
83 		}
84 	} else {
85 		if (likely(dev_priv->chan_vm)) {
86 			if (*size > 256 * 1024)
87 				*page_shift = dev_priv->chan_vm->lpg_shift;
88 			else
89 				*page_shift = dev_priv->chan_vm->spg_shift;
90 		} else {
91 			*page_shift = 12;
92 		}
93 
94 		*size = roundup(*size, (1 << *page_shift));
95 		*align = max((1 << *page_shift), *align);
96 	}
97 
98 	*size = roundup(*size, PAGE_SIZE);
99 }
100 
101 int
102 nouveau_bo_new(struct drm_device *dev, struct nouveau_channel *chan,
103 	       int size, int align, uint32_t flags, uint32_t tile_mode,
104 	       uint32_t tile_flags, bool no_vm, bool mappable,
105 	       struct nouveau_bo **pnvbo)
106 {
107 	struct drm_nouveau_private *dev_priv = dev->dev_private;
108 	struct nouveau_bo *nvbo;
109 	int ret = 0, page_shift = 0;
110 
111 	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
112 	if (!nvbo)
113 		return -ENOMEM;
114 	INIT_LIST_HEAD(&nvbo->head);
115 	INIT_LIST_HEAD(&nvbo->entry);
116 	nvbo->mappable = mappable;
117 	nvbo->no_vm = no_vm;
118 	nvbo->tile_mode = tile_mode;
119 	nvbo->tile_flags = tile_flags;
120 	nvbo->bo.bdev = &dev_priv->ttm.bdev;
121 
122 	nouveau_bo_fixup_align(nvbo, &align, &size, &page_shift);
123 	align >>= PAGE_SHIFT;
124 
125 	if (!nvbo->no_vm && dev_priv->chan_vm) {
126 		ret = nouveau_vm_get(dev_priv->chan_vm, size, page_shift,
127 				     NV_MEM_ACCESS_RW, &nvbo->vma);
128 		if (ret) {
129 			kfree(nvbo);
130 			return ret;
131 		}
132 	}
133 
134 	nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
135 	nouveau_bo_placement_set(nvbo, flags, 0);
136 
137 	nvbo->channel = chan;
138 	ret = ttm_bo_init(&dev_priv->ttm.bdev, &nvbo->bo, size,
139 			  ttm_bo_type_device, &nvbo->placement, align, 0,
140 			  false, NULL, size, nouveau_bo_del_ttm);
141 	if (ret) {
142 		/* ttm will call nouveau_bo_del_ttm if it fails.. */
143 		return ret;
144 	}
145 	nvbo->channel = NULL;
146 
147 	if (nvbo->vma.node) {
148 		if (nvbo->bo.mem.mem_type == TTM_PL_VRAM)
149 			nvbo->bo.offset = nvbo->vma.offset;
150 	}
151 
152 	*pnvbo = nvbo;
153 	return 0;
154 }
155 
156 static void
157 set_placement_list(uint32_t *pl, unsigned *n, uint32_t type, uint32_t flags)
158 {
159 	*n = 0;
160 
161 	if (type & TTM_PL_FLAG_VRAM)
162 		pl[(*n)++] = TTM_PL_FLAG_VRAM | flags;
163 	if (type & TTM_PL_FLAG_TT)
164 		pl[(*n)++] = TTM_PL_FLAG_TT | flags;
165 	if (type & TTM_PL_FLAG_SYSTEM)
166 		pl[(*n)++] = TTM_PL_FLAG_SYSTEM | flags;
167 }
168 
169 static void
170 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
171 {
172 	struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
173 	int vram_pages = dev_priv->vram_size >> PAGE_SHIFT;
174 
175 	if (dev_priv->card_type == NV_10 &&
176 	    nvbo->tile_mode && (type & TTM_PL_FLAG_VRAM) &&
177 	    nvbo->bo.mem.num_pages < vram_pages / 2) {
178 		/*
179 		 * Make sure that the color and depth buffers are handled
180 		 * by independent memory controller units. Up to a 9x
181 		 * speed up when alpha-blending and depth-test are enabled
182 		 * at the same time.
183 		 */
184 		if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
185 			nvbo->placement.fpfn = vram_pages / 2;
186 			nvbo->placement.lpfn = ~0;
187 		} else {
188 			nvbo->placement.fpfn = 0;
189 			nvbo->placement.lpfn = vram_pages / 2;
190 		}
191 	}
192 }
193 
194 void
195 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
196 {
197 	struct ttm_placement *pl = &nvbo->placement;
198 	uint32_t flags = TTM_PL_MASK_CACHING |
199 		(nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
200 
201 	pl->placement = nvbo->placements;
202 	set_placement_list(nvbo->placements, &pl->num_placement,
203 			   type, flags);
204 
205 	pl->busy_placement = nvbo->busy_placements;
206 	set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
207 			   type | busy, flags);
208 
209 	set_placement_range(nvbo, type);
210 }
211 
212 int
213 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype)
214 {
215 	struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
216 	struct ttm_buffer_object *bo = &nvbo->bo;
217 	int ret;
218 
219 	if (nvbo->pin_refcnt && !(memtype & (1 << bo->mem.mem_type))) {
220 		NV_ERROR(nouveau_bdev(bo->bdev)->dev,
221 			 "bo %p pinned elsewhere: 0x%08x vs 0x%08x\n", bo,
222 			 1 << bo->mem.mem_type, memtype);
223 		return -EINVAL;
224 	}
225 
226 	if (nvbo->pin_refcnt++)
227 		return 0;
228 
229 	ret = ttm_bo_reserve(bo, false, false, false, 0);
230 	if (ret)
231 		goto out;
232 
233 	nouveau_bo_placement_set(nvbo, memtype, 0);
234 
235 	ret = nouveau_bo_validate(nvbo, false, false, false);
236 	if (ret == 0) {
237 		switch (bo->mem.mem_type) {
238 		case TTM_PL_VRAM:
239 			dev_priv->fb_aper_free -= bo->mem.size;
240 			break;
241 		case TTM_PL_TT:
242 			dev_priv->gart_info.aper_free -= bo->mem.size;
243 			break;
244 		default:
245 			break;
246 		}
247 	}
248 	ttm_bo_unreserve(bo);
249 out:
250 	if (unlikely(ret))
251 		nvbo->pin_refcnt--;
252 	return ret;
253 }
254 
255 int
256 nouveau_bo_unpin(struct nouveau_bo *nvbo)
257 {
258 	struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
259 	struct ttm_buffer_object *bo = &nvbo->bo;
260 	int ret;
261 
262 	if (--nvbo->pin_refcnt)
263 		return 0;
264 
265 	ret = ttm_bo_reserve(bo, false, false, false, 0);
266 	if (ret)
267 		return ret;
268 
269 	nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
270 
271 	ret = nouveau_bo_validate(nvbo, false, false, false);
272 	if (ret == 0) {
273 		switch (bo->mem.mem_type) {
274 		case TTM_PL_VRAM:
275 			dev_priv->fb_aper_free += bo->mem.size;
276 			break;
277 		case TTM_PL_TT:
278 			dev_priv->gart_info.aper_free += bo->mem.size;
279 			break;
280 		default:
281 			break;
282 		}
283 	}
284 
285 	ttm_bo_unreserve(bo);
286 	return ret;
287 }
288 
289 int
290 nouveau_bo_map(struct nouveau_bo *nvbo)
291 {
292 	int ret;
293 
294 	ret = ttm_bo_reserve(&nvbo->bo, false, false, false, 0);
295 	if (ret)
296 		return ret;
297 
298 	ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
299 	ttm_bo_unreserve(&nvbo->bo);
300 	return ret;
301 }
302 
303 void
304 nouveau_bo_unmap(struct nouveau_bo *nvbo)
305 {
306 	if (nvbo)
307 		ttm_bo_kunmap(&nvbo->kmap);
308 }
309 
310 int
311 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
312 		    bool no_wait_reserve, bool no_wait_gpu)
313 {
314 	int ret;
315 
316 	ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, interruptible,
317 			      no_wait_reserve, no_wait_gpu);
318 	if (ret)
319 		return ret;
320 
321 	if (nvbo->vma.node) {
322 		if (nvbo->bo.mem.mem_type == TTM_PL_VRAM)
323 			nvbo->bo.offset = nvbo->vma.offset;
324 	}
325 
326 	return 0;
327 }
328 
329 u16
330 nouveau_bo_rd16(struct nouveau_bo *nvbo, unsigned index)
331 {
332 	bool is_iomem;
333 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
334 	mem = &mem[index];
335 	if (is_iomem)
336 		return ioread16_native((void __force __iomem *)mem);
337 	else
338 		return *mem;
339 }
340 
341 void
342 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
343 {
344 	bool is_iomem;
345 	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
346 	mem = &mem[index];
347 	if (is_iomem)
348 		iowrite16_native(val, (void __force __iomem *)mem);
349 	else
350 		*mem = val;
351 }
352 
353 u32
354 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
355 {
356 	bool is_iomem;
357 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
358 	mem = &mem[index];
359 	if (is_iomem)
360 		return ioread32_native((void __force __iomem *)mem);
361 	else
362 		return *mem;
363 }
364 
365 void
366 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
367 {
368 	bool is_iomem;
369 	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
370 	mem = &mem[index];
371 	if (is_iomem)
372 		iowrite32_native(val, (void __force __iomem *)mem);
373 	else
374 		*mem = val;
375 }
376 
377 static struct ttm_backend *
378 nouveau_bo_create_ttm_backend_entry(struct ttm_bo_device *bdev)
379 {
380 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
381 	struct drm_device *dev = dev_priv->dev;
382 
383 	switch (dev_priv->gart_info.type) {
384 #if __OS_HAS_AGP
385 	case NOUVEAU_GART_AGP:
386 		return ttm_agp_backend_init(bdev, dev->agp->bridge);
387 #endif
388 	case NOUVEAU_GART_SGDMA:
389 		return nouveau_sgdma_init_ttm(dev);
390 	default:
391 		NV_ERROR(dev, "Unknown GART type %d\n",
392 			 dev_priv->gart_info.type);
393 		break;
394 	}
395 
396 	return NULL;
397 }
398 
399 static int
400 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
401 {
402 	/* We'll do this from user space. */
403 	return 0;
404 }
405 
406 static int
407 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
408 			 struct ttm_mem_type_manager *man)
409 {
410 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
411 	struct drm_device *dev = dev_priv->dev;
412 
413 	switch (type) {
414 	case TTM_PL_SYSTEM:
415 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
416 		man->available_caching = TTM_PL_MASK_CACHING;
417 		man->default_caching = TTM_PL_FLAG_CACHED;
418 		break;
419 	case TTM_PL_VRAM:
420 		if (dev_priv->card_type >= NV_50) {
421 			man->func = &nouveau_vram_manager;
422 			man->io_reserve_fastpath = false;
423 			man->use_io_reserve_lru = true;
424 		} else {
425 			man->func = &ttm_bo_manager_func;
426 		}
427 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
428 			     TTM_MEMTYPE_FLAG_MAPPABLE;
429 		man->available_caching = TTM_PL_FLAG_UNCACHED |
430 					 TTM_PL_FLAG_WC;
431 		man->default_caching = TTM_PL_FLAG_WC;
432 		break;
433 	case TTM_PL_TT:
434 		man->func = &ttm_bo_manager_func;
435 		switch (dev_priv->gart_info.type) {
436 		case NOUVEAU_GART_AGP:
437 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
438 			man->available_caching = TTM_PL_FLAG_UNCACHED |
439 				TTM_PL_FLAG_WC;
440 			man->default_caching = TTM_PL_FLAG_WC;
441 			break;
442 		case NOUVEAU_GART_SGDMA:
443 			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
444 				     TTM_MEMTYPE_FLAG_CMA;
445 			man->available_caching = TTM_PL_MASK_CACHING;
446 			man->default_caching = TTM_PL_FLAG_CACHED;
447 			man->gpu_offset = dev_priv->gart_info.aper_base;
448 			break;
449 		default:
450 			NV_ERROR(dev, "Unknown GART type: %d\n",
451 				 dev_priv->gart_info.type);
452 			return -EINVAL;
453 		}
454 		break;
455 	default:
456 		NV_ERROR(dev, "Unsupported memory type %u\n", (unsigned)type);
457 		return -EINVAL;
458 	}
459 	return 0;
460 }
461 
462 static void
463 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
464 {
465 	struct nouveau_bo *nvbo = nouveau_bo(bo);
466 
467 	switch (bo->mem.mem_type) {
468 	case TTM_PL_VRAM:
469 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
470 					 TTM_PL_FLAG_SYSTEM);
471 		break;
472 	default:
473 		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
474 		break;
475 	}
476 
477 	*pl = nvbo->placement;
478 }
479 
480 
481 /* GPU-assisted copy using NV_MEMORY_TO_MEMORY_FORMAT, can access
482  * TTM_PL_{VRAM,TT} directly.
483  */
484 
485 static int
486 nouveau_bo_move_accel_cleanup(struct nouveau_channel *chan,
487 			      struct nouveau_bo *nvbo, bool evict,
488 			      bool no_wait_reserve, bool no_wait_gpu,
489 			      struct ttm_mem_reg *new_mem)
490 {
491 	struct nouveau_fence *fence = NULL;
492 	int ret;
493 
494 	ret = nouveau_fence_new(chan, &fence, true);
495 	if (ret)
496 		return ret;
497 
498 	ret = ttm_bo_move_accel_cleanup(&nvbo->bo, fence, NULL, evict,
499 					no_wait_reserve, no_wait_gpu, new_mem);
500 	nouveau_fence_unref(&fence);
501 	return ret;
502 }
503 
504 static inline uint32_t
505 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
506 		      struct nouveau_channel *chan, struct ttm_mem_reg *mem)
507 {
508 	struct nouveau_bo *nvbo = nouveau_bo(bo);
509 
510 	if (nvbo->no_vm) {
511 		if (mem->mem_type == TTM_PL_TT)
512 			return NvDmaGART;
513 		return NvDmaVRAM;
514 	}
515 
516 	if (mem->mem_type == TTM_PL_TT)
517 		return chan->gart_handle;
518 	return chan->vram_handle;
519 }
520 
521 static int
522 nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
523 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
524 {
525 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
526 	struct nouveau_bo *nvbo = nouveau_bo(bo);
527 	u64 src_offset = old_mem->start << PAGE_SHIFT;
528 	u64 dst_offset = new_mem->start << PAGE_SHIFT;
529 	u32 page_count = new_mem->num_pages;
530 	int ret;
531 
532 	if (!nvbo->no_vm) {
533 		if (old_mem->mem_type == TTM_PL_VRAM)
534 			src_offset  = nvbo->vma.offset;
535 		else
536 			src_offset += dev_priv->gart_info.aper_base;
537 
538 		if (new_mem->mem_type == TTM_PL_VRAM)
539 			dst_offset  = nvbo->vma.offset;
540 		else
541 			dst_offset += dev_priv->gart_info.aper_base;
542 	}
543 
544 	page_count = new_mem->num_pages;
545 	while (page_count) {
546 		int line_count = (page_count > 2047) ? 2047 : page_count;
547 
548 		ret = RING_SPACE(chan, 12);
549 		if (ret)
550 			return ret;
551 
552 		BEGIN_NVC0(chan, 2, NvSubM2MF, 0x0238, 2);
553 		OUT_RING  (chan, upper_32_bits(dst_offset));
554 		OUT_RING  (chan, lower_32_bits(dst_offset));
555 		BEGIN_NVC0(chan, 2, NvSubM2MF, 0x030c, 6);
556 		OUT_RING  (chan, upper_32_bits(src_offset));
557 		OUT_RING  (chan, lower_32_bits(src_offset));
558 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
559 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
560 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
561 		OUT_RING  (chan, line_count);
562 		BEGIN_NVC0(chan, 2, NvSubM2MF, 0x0300, 1);
563 		OUT_RING  (chan, 0x00100110);
564 
565 		page_count -= line_count;
566 		src_offset += (PAGE_SIZE * line_count);
567 		dst_offset += (PAGE_SIZE * line_count);
568 	}
569 
570 	return 0;
571 }
572 
573 static int
574 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
575 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
576 {
577 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
578 	struct nouveau_bo *nvbo = nouveau_bo(bo);
579 	u64 length = (new_mem->num_pages << PAGE_SHIFT);
580 	u64 src_offset, dst_offset;
581 	int ret;
582 
583 	src_offset = old_mem->start << PAGE_SHIFT;
584 	dst_offset = new_mem->start << PAGE_SHIFT;
585 	if (!nvbo->no_vm) {
586 		if (old_mem->mem_type == TTM_PL_VRAM)
587 			src_offset  = nvbo->vma.offset;
588 		else
589 			src_offset += dev_priv->gart_info.aper_base;
590 
591 		if (new_mem->mem_type == TTM_PL_VRAM)
592 			dst_offset  = nvbo->vma.offset;
593 		else
594 			dst_offset += dev_priv->gart_info.aper_base;
595 	}
596 
597 	ret = RING_SPACE(chan, 3);
598 	if (ret)
599 		return ret;
600 
601 	BEGIN_RING(chan, NvSubM2MF, 0x0184, 2);
602 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
603 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
604 
605 	while (length) {
606 		u32 amount, stride, height;
607 
608 		amount  = min(length, (u64)(4 * 1024 * 1024));
609 		stride  = 16 * 4;
610 		height  = amount / stride;
611 
612 		if (new_mem->mem_type == TTM_PL_VRAM &&
613 		    nouveau_bo_tile_layout(nvbo)) {
614 			ret = RING_SPACE(chan, 8);
615 			if (ret)
616 				return ret;
617 
618 			BEGIN_RING(chan, NvSubM2MF, 0x0200, 7);
619 			OUT_RING  (chan, 0);
620 			OUT_RING  (chan, 0);
621 			OUT_RING  (chan, stride);
622 			OUT_RING  (chan, height);
623 			OUT_RING  (chan, 1);
624 			OUT_RING  (chan, 0);
625 			OUT_RING  (chan, 0);
626 		} else {
627 			ret = RING_SPACE(chan, 2);
628 			if (ret)
629 				return ret;
630 
631 			BEGIN_RING(chan, NvSubM2MF, 0x0200, 1);
632 			OUT_RING  (chan, 1);
633 		}
634 		if (old_mem->mem_type == TTM_PL_VRAM &&
635 		    nouveau_bo_tile_layout(nvbo)) {
636 			ret = RING_SPACE(chan, 8);
637 			if (ret)
638 				return ret;
639 
640 			BEGIN_RING(chan, NvSubM2MF, 0x021c, 7);
641 			OUT_RING  (chan, 0);
642 			OUT_RING  (chan, 0);
643 			OUT_RING  (chan, stride);
644 			OUT_RING  (chan, height);
645 			OUT_RING  (chan, 1);
646 			OUT_RING  (chan, 0);
647 			OUT_RING  (chan, 0);
648 		} else {
649 			ret = RING_SPACE(chan, 2);
650 			if (ret)
651 				return ret;
652 
653 			BEGIN_RING(chan, NvSubM2MF, 0x021c, 1);
654 			OUT_RING  (chan, 1);
655 		}
656 
657 		ret = RING_SPACE(chan, 14);
658 		if (ret)
659 			return ret;
660 
661 		BEGIN_RING(chan, NvSubM2MF, 0x0238, 2);
662 		OUT_RING  (chan, upper_32_bits(src_offset));
663 		OUT_RING  (chan, upper_32_bits(dst_offset));
664 		BEGIN_RING(chan, NvSubM2MF, 0x030c, 8);
665 		OUT_RING  (chan, lower_32_bits(src_offset));
666 		OUT_RING  (chan, lower_32_bits(dst_offset));
667 		OUT_RING  (chan, stride);
668 		OUT_RING  (chan, stride);
669 		OUT_RING  (chan, stride);
670 		OUT_RING  (chan, height);
671 		OUT_RING  (chan, 0x00000101);
672 		OUT_RING  (chan, 0x00000000);
673 		BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
674 		OUT_RING  (chan, 0);
675 
676 		length -= amount;
677 		src_offset += amount;
678 		dst_offset += amount;
679 	}
680 
681 	return 0;
682 }
683 
684 static int
685 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
686 		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
687 {
688 	u32 src_offset = old_mem->start << PAGE_SHIFT;
689 	u32 dst_offset = new_mem->start << PAGE_SHIFT;
690 	u32 page_count = new_mem->num_pages;
691 	int ret;
692 
693 	ret = RING_SPACE(chan, 3);
694 	if (ret)
695 		return ret;
696 
697 	BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
698 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
699 	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
700 
701 	page_count = new_mem->num_pages;
702 	while (page_count) {
703 		int line_count = (page_count > 2047) ? 2047 : page_count;
704 
705 		ret = RING_SPACE(chan, 11);
706 		if (ret)
707 			return ret;
708 
709 		BEGIN_RING(chan, NvSubM2MF,
710 				 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
711 		OUT_RING  (chan, src_offset);
712 		OUT_RING  (chan, dst_offset);
713 		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
714 		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
715 		OUT_RING  (chan, PAGE_SIZE); /* line_length */
716 		OUT_RING  (chan, line_count);
717 		OUT_RING  (chan, 0x00000101);
718 		OUT_RING  (chan, 0x00000000);
719 		BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
720 		OUT_RING  (chan, 0);
721 
722 		page_count -= line_count;
723 		src_offset += (PAGE_SIZE * line_count);
724 		dst_offset += (PAGE_SIZE * line_count);
725 	}
726 
727 	return 0;
728 }
729 
730 static int
731 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
732 		     bool no_wait_reserve, bool no_wait_gpu,
733 		     struct ttm_mem_reg *new_mem)
734 {
735 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
736 	struct nouveau_bo *nvbo = nouveau_bo(bo);
737 	struct nouveau_channel *chan;
738 	int ret;
739 
740 	chan = nvbo->channel;
741 	if (!chan || nvbo->no_vm) {
742 		chan = dev_priv->channel;
743 		mutex_lock_nested(&chan->mutex, NOUVEAU_KCHANNEL_MUTEX);
744 	}
745 
746 	if (dev_priv->card_type < NV_50)
747 		ret = nv04_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
748 	else
749 	if (dev_priv->card_type < NV_C0)
750 		ret = nv50_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
751 	else
752 		ret = nvc0_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
753 	if (ret == 0) {
754 		ret = nouveau_bo_move_accel_cleanup(chan, nvbo, evict,
755 						    no_wait_reserve,
756 						    no_wait_gpu, new_mem);
757 	}
758 
759 	if (chan == dev_priv->channel)
760 		mutex_unlock(&chan->mutex);
761 	return ret;
762 }
763 
764 static int
765 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
766 		      bool no_wait_reserve, bool no_wait_gpu,
767 		      struct ttm_mem_reg *new_mem)
768 {
769 	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
770 	struct ttm_placement placement;
771 	struct ttm_mem_reg tmp_mem;
772 	int ret;
773 
774 	placement.fpfn = placement.lpfn = 0;
775 	placement.num_placement = placement.num_busy_placement = 1;
776 	placement.placement = placement.busy_placement = &placement_memtype;
777 
778 	tmp_mem = *new_mem;
779 	tmp_mem.mm_node = NULL;
780 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_reserve, no_wait_gpu);
781 	if (ret)
782 		return ret;
783 
784 	ret = ttm_tt_bind(bo->ttm, &tmp_mem);
785 	if (ret)
786 		goto out;
787 
788 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_reserve, no_wait_gpu, &tmp_mem);
789 	if (ret)
790 		goto out;
791 
792 	ret = ttm_bo_move_ttm(bo, true, no_wait_reserve, no_wait_gpu, new_mem);
793 out:
794 	ttm_bo_mem_put(bo, &tmp_mem);
795 	return ret;
796 }
797 
798 static int
799 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
800 		      bool no_wait_reserve, bool no_wait_gpu,
801 		      struct ttm_mem_reg *new_mem)
802 {
803 	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
804 	struct ttm_placement placement;
805 	struct ttm_mem_reg tmp_mem;
806 	int ret;
807 
808 	placement.fpfn = placement.lpfn = 0;
809 	placement.num_placement = placement.num_busy_placement = 1;
810 	placement.placement = placement.busy_placement = &placement_memtype;
811 
812 	tmp_mem = *new_mem;
813 	tmp_mem.mm_node = NULL;
814 	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_reserve, no_wait_gpu);
815 	if (ret)
816 		return ret;
817 
818 	ret = ttm_bo_move_ttm(bo, true, no_wait_reserve, no_wait_gpu, &tmp_mem);
819 	if (ret)
820 		goto out;
821 
822 	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_reserve, no_wait_gpu, new_mem);
823 	if (ret)
824 		goto out;
825 
826 out:
827 	ttm_bo_mem_put(bo, &tmp_mem);
828 	return ret;
829 }
830 
831 static int
832 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
833 		   struct nouveau_tile_reg **new_tile)
834 {
835 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
836 	struct drm_device *dev = dev_priv->dev;
837 	struct nouveau_bo *nvbo = nouveau_bo(bo);
838 	uint64_t offset;
839 
840 	if (nvbo->no_vm || new_mem->mem_type != TTM_PL_VRAM) {
841 		/* Nothing to do. */
842 		*new_tile = NULL;
843 		return 0;
844 	}
845 
846 	offset = new_mem->start << PAGE_SHIFT;
847 
848 	if (dev_priv->chan_vm) {
849 		nouveau_vm_map(&nvbo->vma, new_mem->mm_node);
850 	} else if (dev_priv->card_type >= NV_10) {
851 		*new_tile = nv10_mem_set_tiling(dev, offset, new_mem->size,
852 						nvbo->tile_mode,
853 						nvbo->tile_flags);
854 	}
855 
856 	return 0;
857 }
858 
859 static void
860 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
861 		      struct nouveau_tile_reg *new_tile,
862 		      struct nouveau_tile_reg **old_tile)
863 {
864 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
865 	struct drm_device *dev = dev_priv->dev;
866 
867 	if (dev_priv->card_type >= NV_10 &&
868 	    dev_priv->card_type < NV_50) {
869 		nv10_mem_put_tile_region(dev, *old_tile, bo->sync_obj);
870 		*old_tile = new_tile;
871 	}
872 }
873 
874 static int
875 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
876 		bool no_wait_reserve, bool no_wait_gpu,
877 		struct ttm_mem_reg *new_mem)
878 {
879 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
880 	struct nouveau_bo *nvbo = nouveau_bo(bo);
881 	struct ttm_mem_reg *old_mem = &bo->mem;
882 	struct nouveau_tile_reg *new_tile = NULL;
883 	int ret = 0;
884 
885 	ret = nouveau_bo_vm_bind(bo, new_mem, &new_tile);
886 	if (ret)
887 		return ret;
888 
889 	/* Fake bo copy. */
890 	if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
891 		BUG_ON(bo->mem.mm_node != NULL);
892 		bo->mem = *new_mem;
893 		new_mem->mm_node = NULL;
894 		goto out;
895 	}
896 
897 	/* Software copy if the card isn't up and running yet. */
898 	if (!dev_priv->channel) {
899 		ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
900 		goto out;
901 	}
902 
903 	/* Hardware assisted copy. */
904 	if (new_mem->mem_type == TTM_PL_SYSTEM)
905 		ret = nouveau_bo_move_flipd(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
906 	else if (old_mem->mem_type == TTM_PL_SYSTEM)
907 		ret = nouveau_bo_move_flips(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
908 	else
909 		ret = nouveau_bo_move_m2mf(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
910 
911 	if (!ret)
912 		goto out;
913 
914 	/* Fallback to software copy. */
915 	ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
916 
917 out:
918 	if (ret)
919 		nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
920 	else
921 		nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
922 
923 	return ret;
924 }
925 
926 static int
927 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
928 {
929 	return 0;
930 }
931 
932 static int
933 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
934 {
935 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
936 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
937 	struct drm_device *dev = dev_priv->dev;
938 	int ret;
939 
940 	mem->bus.addr = NULL;
941 	mem->bus.offset = 0;
942 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
943 	mem->bus.base = 0;
944 	mem->bus.is_iomem = false;
945 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
946 		return -EINVAL;
947 	switch (mem->mem_type) {
948 	case TTM_PL_SYSTEM:
949 		/* System memory */
950 		return 0;
951 	case TTM_PL_TT:
952 #if __OS_HAS_AGP
953 		if (dev_priv->gart_info.type == NOUVEAU_GART_AGP) {
954 			mem->bus.offset = mem->start << PAGE_SHIFT;
955 			mem->bus.base = dev_priv->gart_info.aper_base;
956 			mem->bus.is_iomem = true;
957 		}
958 #endif
959 		break;
960 	case TTM_PL_VRAM:
961 	{
962 		struct nouveau_vram *vram = mem->mm_node;
963 		u8 page_shift;
964 
965 		if (!dev_priv->bar1_vm) {
966 			mem->bus.offset = mem->start << PAGE_SHIFT;
967 			mem->bus.base = pci_resource_start(dev->pdev, 1);
968 			mem->bus.is_iomem = true;
969 			break;
970 		}
971 
972 		if (dev_priv->card_type == NV_C0)
973 			page_shift = vram->page_shift;
974 		else
975 			page_shift = 12;
976 
977 		ret = nouveau_vm_get(dev_priv->bar1_vm, mem->bus.size,
978 				     page_shift, NV_MEM_ACCESS_RW,
979 				     &vram->bar_vma);
980 		if (ret)
981 			return ret;
982 
983 		nouveau_vm_map(&vram->bar_vma, vram);
984 		if (ret) {
985 			nouveau_vm_put(&vram->bar_vma);
986 			return ret;
987 		}
988 
989 		mem->bus.offset = vram->bar_vma.offset;
990 		if (dev_priv->card_type == NV_50) /*XXX*/
991 			mem->bus.offset -= 0x0020000000ULL;
992 		mem->bus.base = pci_resource_start(dev->pdev, 1);
993 		mem->bus.is_iomem = true;
994 	}
995 		break;
996 	default:
997 		return -EINVAL;
998 	}
999 	return 0;
1000 }
1001 
1002 static void
1003 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1004 {
1005 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
1006 	struct nouveau_vram *vram = mem->mm_node;
1007 
1008 	if (!dev_priv->bar1_vm || mem->mem_type != TTM_PL_VRAM)
1009 		return;
1010 
1011 	if (!vram->bar_vma.node)
1012 		return;
1013 
1014 	nouveau_vm_unmap(&vram->bar_vma);
1015 	nouveau_vm_put(&vram->bar_vma);
1016 }
1017 
1018 static int
1019 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1020 {
1021 	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
1022 	struct nouveau_bo *nvbo = nouveau_bo(bo);
1023 
1024 	/* as long as the bo isn't in vram, and isn't tiled, we've got
1025 	 * nothing to do here.
1026 	 */
1027 	if (bo->mem.mem_type != TTM_PL_VRAM) {
1028 		if (dev_priv->card_type < NV_50 ||
1029 		    !nouveau_bo_tile_layout(nvbo))
1030 			return 0;
1031 	}
1032 
1033 	/* make sure bo is in mappable vram */
1034 	if (bo->mem.start + bo->mem.num_pages < dev_priv->fb_mappable_pages)
1035 		return 0;
1036 
1037 
1038 	nvbo->placement.fpfn = 0;
1039 	nvbo->placement.lpfn = dev_priv->fb_mappable_pages;
1040 	nouveau_bo_placement_set(nvbo, TTM_PL_VRAM, 0);
1041 	return nouveau_bo_validate(nvbo, false, true, false);
1042 }
1043 
1044 void
1045 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
1046 {
1047 	struct nouveau_fence *old_fence;
1048 
1049 	if (likely(fence))
1050 		nouveau_fence_ref(fence);
1051 
1052 	spin_lock(&nvbo->bo.bdev->fence_lock);
1053 	old_fence = nvbo->bo.sync_obj;
1054 	nvbo->bo.sync_obj = fence;
1055 	spin_unlock(&nvbo->bo.bdev->fence_lock);
1056 
1057 	nouveau_fence_unref(&old_fence);
1058 }
1059 
1060 struct ttm_bo_driver nouveau_bo_driver = {
1061 	.create_ttm_backend_entry = nouveau_bo_create_ttm_backend_entry,
1062 	.invalidate_caches = nouveau_bo_invalidate_caches,
1063 	.init_mem_type = nouveau_bo_init_mem_type,
1064 	.evict_flags = nouveau_bo_evict_flags,
1065 	.move = nouveau_bo_move,
1066 	.verify_access = nouveau_bo_verify_access,
1067 	.sync_obj_signaled = __nouveau_fence_signalled,
1068 	.sync_obj_wait = __nouveau_fence_wait,
1069 	.sync_obj_flush = __nouveau_fence_flush,
1070 	.sync_obj_unref = __nouveau_fence_unref,
1071 	.sync_obj_ref = __nouveau_fence_ref,
1072 	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1073 	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1074 	.io_mem_free = &nouveau_ttm_io_mem_free,
1075 };
1076 
1077