drm/ttm: merge ttm_backend and ttm_tt V5
[linux-2.6-block.git] / drivers / gpu / drm / nouveau / nouveau_bo.c
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         kfree(nvbo);
53 }
54
55 static void
56 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
57                        int *align, int *size)
58 {
59         struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
60
61         if (dev_priv->card_type < NV_50) {
62                 if (nvbo->tile_mode) {
63                         if (dev_priv->chipset >= 0x40) {
64                                 *align = 65536;
65                                 *size = roundup(*size, 64 * nvbo->tile_mode);
66
67                         } else if (dev_priv->chipset >= 0x30) {
68                                 *align = 32768;
69                                 *size = roundup(*size, 64 * nvbo->tile_mode);
70
71                         } else if (dev_priv->chipset >= 0x20) {
72                                 *align = 16384;
73                                 *size = roundup(*size, 64 * nvbo->tile_mode);
74
75                         } else if (dev_priv->chipset >= 0x10) {
76                                 *align = 16384;
77                                 *size = roundup(*size, 32 * nvbo->tile_mode);
78                         }
79                 }
80         } else {
81                 *size = roundup(*size, (1 << nvbo->page_shift));
82                 *align = max((1 <<  nvbo->page_shift), *align);
83         }
84
85         *size = roundup(*size, PAGE_SIZE);
86 }
87
88 int
89 nouveau_bo_new(struct drm_device *dev, int size, int align,
90                uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
91                struct nouveau_bo **pnvbo)
92 {
93         struct drm_nouveau_private *dev_priv = dev->dev_private;
94         struct nouveau_bo *nvbo;
95         int ret;
96
97         nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
98         if (!nvbo)
99                 return -ENOMEM;
100         INIT_LIST_HEAD(&nvbo->head);
101         INIT_LIST_HEAD(&nvbo->entry);
102         INIT_LIST_HEAD(&nvbo->vma_list);
103         nvbo->tile_mode = tile_mode;
104         nvbo->tile_flags = tile_flags;
105         nvbo->bo.bdev = &dev_priv->ttm.bdev;
106
107         nvbo->page_shift = 12;
108         if (dev_priv->bar1_vm) {
109                 if (!(flags & TTM_PL_FLAG_TT) && size > 256 * 1024)
110                         nvbo->page_shift = dev_priv->bar1_vm->lpg_shift;
111         }
112
113         nouveau_bo_fixup_align(nvbo, flags, &align, &size);
114         nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
115         nouveau_bo_placement_set(nvbo, flags, 0);
116
117         ret = ttm_bo_init(&dev_priv->ttm.bdev, &nvbo->bo, size,
118                           ttm_bo_type_device, &nvbo->placement,
119                           align >> PAGE_SHIFT, 0, false, NULL, size,
120                           nouveau_bo_del_ttm);
121         if (ret) {
122                 /* ttm will call nouveau_bo_del_ttm if it fails.. */
123                 return ret;
124         }
125
126         *pnvbo = nvbo;
127         return 0;
128 }
129
130 static void
131 set_placement_list(uint32_t *pl, unsigned *n, uint32_t type, uint32_t flags)
132 {
133         *n = 0;
134
135         if (type & TTM_PL_FLAG_VRAM)
136                 pl[(*n)++] = TTM_PL_FLAG_VRAM | flags;
137         if (type & TTM_PL_FLAG_TT)
138                 pl[(*n)++] = TTM_PL_FLAG_TT | flags;
139         if (type & TTM_PL_FLAG_SYSTEM)
140                 pl[(*n)++] = TTM_PL_FLAG_SYSTEM | flags;
141 }
142
143 static void
144 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
145 {
146         struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
147         int vram_pages = dev_priv->vram_size >> PAGE_SHIFT;
148
149         if (dev_priv->card_type == NV_10 &&
150             nvbo->tile_mode && (type & TTM_PL_FLAG_VRAM) &&
151             nvbo->bo.mem.num_pages < vram_pages / 2) {
152                 /*
153                  * Make sure that the color and depth buffers are handled
154                  * by independent memory controller units. Up to a 9x
155                  * speed up when alpha-blending and depth-test are enabled
156                  * at the same time.
157                  */
158                 if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
159                         nvbo->placement.fpfn = vram_pages / 2;
160                         nvbo->placement.lpfn = ~0;
161                 } else {
162                         nvbo->placement.fpfn = 0;
163                         nvbo->placement.lpfn = vram_pages / 2;
164                 }
165         }
166 }
167
168 void
169 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
170 {
171         struct ttm_placement *pl = &nvbo->placement;
172         uint32_t flags = TTM_PL_MASK_CACHING |
173                 (nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
174
175         pl->placement = nvbo->placements;
176         set_placement_list(nvbo->placements, &pl->num_placement,
177                            type, flags);
178
179         pl->busy_placement = nvbo->busy_placements;
180         set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
181                            type | busy, flags);
182
183         set_placement_range(nvbo, type);
184 }
185
186 int
187 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype)
188 {
189         struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
190         struct ttm_buffer_object *bo = &nvbo->bo;
191         int ret;
192
193         if (nvbo->pin_refcnt && !(memtype & (1 << bo->mem.mem_type))) {
194                 NV_ERROR(nouveau_bdev(bo->bdev)->dev,
195                          "bo %p pinned elsewhere: 0x%08x vs 0x%08x\n", bo,
196                          1 << bo->mem.mem_type, memtype);
197                 return -EINVAL;
198         }
199
200         if (nvbo->pin_refcnt++)
201                 return 0;
202
203         ret = ttm_bo_reserve(bo, false, false, false, 0);
204         if (ret)
205                 goto out;
206
207         nouveau_bo_placement_set(nvbo, memtype, 0);
208
209         ret = nouveau_bo_validate(nvbo, false, false, false);
210         if (ret == 0) {
211                 switch (bo->mem.mem_type) {
212                 case TTM_PL_VRAM:
213                         dev_priv->fb_aper_free -= bo->mem.size;
214                         break;
215                 case TTM_PL_TT:
216                         dev_priv->gart_info.aper_free -= bo->mem.size;
217                         break;
218                 default:
219                         break;
220                 }
221         }
222         ttm_bo_unreserve(bo);
223 out:
224         if (unlikely(ret))
225                 nvbo->pin_refcnt--;
226         return ret;
227 }
228
229 int
230 nouveau_bo_unpin(struct nouveau_bo *nvbo)
231 {
232         struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
233         struct ttm_buffer_object *bo = &nvbo->bo;
234         int ret;
235
236         if (--nvbo->pin_refcnt)
237                 return 0;
238
239         ret = ttm_bo_reserve(bo, false, false, false, 0);
240         if (ret)
241                 return ret;
242
243         nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
244
245         ret = nouveau_bo_validate(nvbo, false, false, false);
246         if (ret == 0) {
247                 switch (bo->mem.mem_type) {
248                 case TTM_PL_VRAM:
249                         dev_priv->fb_aper_free += bo->mem.size;
250                         break;
251                 case TTM_PL_TT:
252                         dev_priv->gart_info.aper_free += bo->mem.size;
253                         break;
254                 default:
255                         break;
256                 }
257         }
258
259         ttm_bo_unreserve(bo);
260         return ret;
261 }
262
263 int
264 nouveau_bo_map(struct nouveau_bo *nvbo)
265 {
266         int ret;
267
268         ret = ttm_bo_reserve(&nvbo->bo, false, false, false, 0);
269         if (ret)
270                 return ret;
271
272         ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
273         ttm_bo_unreserve(&nvbo->bo);
274         return ret;
275 }
276
277 void
278 nouveau_bo_unmap(struct nouveau_bo *nvbo)
279 {
280         if (nvbo)
281                 ttm_bo_kunmap(&nvbo->kmap);
282 }
283
284 int
285 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
286                     bool no_wait_reserve, bool no_wait_gpu)
287 {
288         int ret;
289
290         ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, interruptible,
291                               no_wait_reserve, no_wait_gpu);
292         if (ret)
293                 return ret;
294
295         return 0;
296 }
297
298 u16
299 nouveau_bo_rd16(struct nouveau_bo *nvbo, unsigned index)
300 {
301         bool is_iomem;
302         u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
303         mem = &mem[index];
304         if (is_iomem)
305                 return ioread16_native((void __force __iomem *)mem);
306         else
307                 return *mem;
308 }
309
310 void
311 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
312 {
313         bool is_iomem;
314         u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
315         mem = &mem[index];
316         if (is_iomem)
317                 iowrite16_native(val, (void __force __iomem *)mem);
318         else
319                 *mem = val;
320 }
321
322 u32
323 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
324 {
325         bool is_iomem;
326         u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
327         mem = &mem[index];
328         if (is_iomem)
329                 return ioread32_native((void __force __iomem *)mem);
330         else
331                 return *mem;
332 }
333
334 void
335 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
336 {
337         bool is_iomem;
338         u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
339         mem = &mem[index];
340         if (is_iomem)
341                 iowrite32_native(val, (void __force __iomem *)mem);
342         else
343                 *mem = val;
344 }
345
346 static struct ttm_tt *
347 nouveau_ttm_tt_create(struct ttm_bo_device *bdev,
348                       unsigned long size, uint32_t page_flags,
349                       struct page *dummy_read_page)
350 {
351         struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
352         struct drm_device *dev = dev_priv->dev;
353
354         switch (dev_priv->gart_info.type) {
355 #if __OS_HAS_AGP
356         case NOUVEAU_GART_AGP:
357                 return ttm_agp_tt_create(bdev, dev->agp->bridge,
358                                          size, page_flags, dummy_read_page);
359 #endif
360         case NOUVEAU_GART_PDMA:
361         case NOUVEAU_GART_HW:
362                 return nouveau_sgdma_create_ttm(bdev, size, page_flags,
363                                                 dummy_read_page);
364         default:
365                 NV_ERROR(dev, "Unknown GART type %d\n",
366                          dev_priv->gart_info.type);
367                 break;
368         }
369
370         return NULL;
371 }
372
373 static int
374 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
375 {
376         /* We'll do this from user space. */
377         return 0;
378 }
379
380 static int
381 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
382                          struct ttm_mem_type_manager *man)
383 {
384         struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
385         struct drm_device *dev = dev_priv->dev;
386
387         switch (type) {
388         case TTM_PL_SYSTEM:
389                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
390                 man->available_caching = TTM_PL_MASK_CACHING;
391                 man->default_caching = TTM_PL_FLAG_CACHED;
392                 break;
393         case TTM_PL_VRAM:
394                 if (dev_priv->card_type >= NV_50) {
395                         man->func = &nouveau_vram_manager;
396                         man->io_reserve_fastpath = false;
397                         man->use_io_reserve_lru = true;
398                 } else {
399                         man->func = &ttm_bo_manager_func;
400                 }
401                 man->flags = TTM_MEMTYPE_FLAG_FIXED |
402                              TTM_MEMTYPE_FLAG_MAPPABLE;
403                 man->available_caching = TTM_PL_FLAG_UNCACHED |
404                                          TTM_PL_FLAG_WC;
405                 man->default_caching = TTM_PL_FLAG_WC;
406                 break;
407         case TTM_PL_TT:
408                 if (dev_priv->card_type >= NV_50)
409                         man->func = &nouveau_gart_manager;
410                 else
411                         man->func = &ttm_bo_manager_func;
412                 switch (dev_priv->gart_info.type) {
413                 case NOUVEAU_GART_AGP:
414                         man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
415                         man->available_caching = TTM_PL_FLAG_UNCACHED |
416                                 TTM_PL_FLAG_WC;
417                         man->default_caching = TTM_PL_FLAG_WC;
418                         break;
419                 case NOUVEAU_GART_PDMA:
420                 case NOUVEAU_GART_HW:
421                         man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
422                                      TTM_MEMTYPE_FLAG_CMA;
423                         man->available_caching = TTM_PL_MASK_CACHING;
424                         man->default_caching = TTM_PL_FLAG_CACHED;
425                         break;
426                 default:
427                         NV_ERROR(dev, "Unknown GART type: %d\n",
428                                  dev_priv->gart_info.type);
429                         return -EINVAL;
430                 }
431                 break;
432         default:
433                 NV_ERROR(dev, "Unsupported memory type %u\n", (unsigned)type);
434                 return -EINVAL;
435         }
436         return 0;
437 }
438
439 static void
440 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
441 {
442         struct nouveau_bo *nvbo = nouveau_bo(bo);
443
444         switch (bo->mem.mem_type) {
445         case TTM_PL_VRAM:
446                 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
447                                          TTM_PL_FLAG_SYSTEM);
448                 break;
449         default:
450                 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
451                 break;
452         }
453
454         *pl = nvbo->placement;
455 }
456
457
458 /* GPU-assisted copy using NV_MEMORY_TO_MEMORY_FORMAT, can access
459  * TTM_PL_{VRAM,TT} directly.
460  */
461
462 static int
463 nouveau_bo_move_accel_cleanup(struct nouveau_channel *chan,
464                               struct nouveau_bo *nvbo, bool evict,
465                               bool no_wait_reserve, bool no_wait_gpu,
466                               struct ttm_mem_reg *new_mem)
467 {
468         struct nouveau_fence *fence = NULL;
469         int ret;
470
471         ret = nouveau_fence_new(chan, &fence, true);
472         if (ret)
473                 return ret;
474
475         ret = ttm_bo_move_accel_cleanup(&nvbo->bo, fence, NULL, evict,
476                                         no_wait_reserve, no_wait_gpu, new_mem);
477         nouveau_fence_unref(&fence);
478         return ret;
479 }
480
481 static int
482 nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
483                   struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
484 {
485         struct nouveau_mem *node = old_mem->mm_node;
486         u64 src_offset = node->vma[0].offset;
487         u64 dst_offset = node->vma[1].offset;
488         u32 page_count = new_mem->num_pages;
489         int ret;
490
491         page_count = new_mem->num_pages;
492         while (page_count) {
493                 int line_count = (page_count > 2047) ? 2047 : page_count;
494
495                 ret = RING_SPACE(chan, 12);
496                 if (ret)
497                         return ret;
498
499                 BEGIN_NVC0(chan, 2, NvSubM2MF, 0x0238, 2);
500                 OUT_RING  (chan, upper_32_bits(dst_offset));
501                 OUT_RING  (chan, lower_32_bits(dst_offset));
502                 BEGIN_NVC0(chan, 2, NvSubM2MF, 0x030c, 6);
503                 OUT_RING  (chan, upper_32_bits(src_offset));
504                 OUT_RING  (chan, lower_32_bits(src_offset));
505                 OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
506                 OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
507                 OUT_RING  (chan, PAGE_SIZE); /* line_length */
508                 OUT_RING  (chan, line_count);
509                 BEGIN_NVC0(chan, 2, NvSubM2MF, 0x0300, 1);
510                 OUT_RING  (chan, 0x00100110);
511
512                 page_count -= line_count;
513                 src_offset += (PAGE_SIZE * line_count);
514                 dst_offset += (PAGE_SIZE * line_count);
515         }
516
517         return 0;
518 }
519
520 static int
521 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
522                   struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
523 {
524         struct nouveau_mem *node = old_mem->mm_node;
525         struct nouveau_bo *nvbo = nouveau_bo(bo);
526         u64 length = (new_mem->num_pages << PAGE_SHIFT);
527         u64 src_offset = node->vma[0].offset;
528         u64 dst_offset = node->vma[1].offset;
529         int ret;
530
531         while (length) {
532                 u32 amount, stride, height;
533
534                 amount  = min(length, (u64)(4 * 1024 * 1024));
535                 stride  = 16 * 4;
536                 height  = amount / stride;
537
538                 if (new_mem->mem_type == TTM_PL_VRAM &&
539                     nouveau_bo_tile_layout(nvbo)) {
540                         ret = RING_SPACE(chan, 8);
541                         if (ret)
542                                 return ret;
543
544                         BEGIN_RING(chan, NvSubM2MF, 0x0200, 7);
545                         OUT_RING  (chan, 0);
546                         OUT_RING  (chan, 0);
547                         OUT_RING  (chan, stride);
548                         OUT_RING  (chan, height);
549                         OUT_RING  (chan, 1);
550                         OUT_RING  (chan, 0);
551                         OUT_RING  (chan, 0);
552                 } else {
553                         ret = RING_SPACE(chan, 2);
554                         if (ret)
555                                 return ret;
556
557                         BEGIN_RING(chan, NvSubM2MF, 0x0200, 1);
558                         OUT_RING  (chan, 1);
559                 }
560                 if (old_mem->mem_type == TTM_PL_VRAM &&
561                     nouveau_bo_tile_layout(nvbo)) {
562                         ret = RING_SPACE(chan, 8);
563                         if (ret)
564                                 return ret;
565
566                         BEGIN_RING(chan, NvSubM2MF, 0x021c, 7);
567                         OUT_RING  (chan, 0);
568                         OUT_RING  (chan, 0);
569                         OUT_RING  (chan, stride);
570                         OUT_RING  (chan, height);
571                         OUT_RING  (chan, 1);
572                         OUT_RING  (chan, 0);
573                         OUT_RING  (chan, 0);
574                 } else {
575                         ret = RING_SPACE(chan, 2);
576                         if (ret)
577                                 return ret;
578
579                         BEGIN_RING(chan, NvSubM2MF, 0x021c, 1);
580                         OUT_RING  (chan, 1);
581                 }
582
583                 ret = RING_SPACE(chan, 14);
584                 if (ret)
585                         return ret;
586
587                 BEGIN_RING(chan, NvSubM2MF, 0x0238, 2);
588                 OUT_RING  (chan, upper_32_bits(src_offset));
589                 OUT_RING  (chan, upper_32_bits(dst_offset));
590                 BEGIN_RING(chan, NvSubM2MF, 0x030c, 8);
591                 OUT_RING  (chan, lower_32_bits(src_offset));
592                 OUT_RING  (chan, lower_32_bits(dst_offset));
593                 OUT_RING  (chan, stride);
594                 OUT_RING  (chan, stride);
595                 OUT_RING  (chan, stride);
596                 OUT_RING  (chan, height);
597                 OUT_RING  (chan, 0x00000101);
598                 OUT_RING  (chan, 0x00000000);
599                 BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
600                 OUT_RING  (chan, 0);
601
602                 length -= amount;
603                 src_offset += amount;
604                 dst_offset += amount;
605         }
606
607         return 0;
608 }
609
610 static inline uint32_t
611 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
612                       struct nouveau_channel *chan, struct ttm_mem_reg *mem)
613 {
614         if (mem->mem_type == TTM_PL_TT)
615                 return chan->gart_handle;
616         return chan->vram_handle;
617 }
618
619 static int
620 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
621                   struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
622 {
623         u32 src_offset = old_mem->start << PAGE_SHIFT;
624         u32 dst_offset = new_mem->start << PAGE_SHIFT;
625         u32 page_count = new_mem->num_pages;
626         int ret;
627
628         ret = RING_SPACE(chan, 3);
629         if (ret)
630                 return ret;
631
632         BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
633         OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
634         OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
635
636         page_count = new_mem->num_pages;
637         while (page_count) {
638                 int line_count = (page_count > 2047) ? 2047 : page_count;
639
640                 ret = RING_SPACE(chan, 11);
641                 if (ret)
642                         return ret;
643
644                 BEGIN_RING(chan, NvSubM2MF,
645                                  NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
646                 OUT_RING  (chan, src_offset);
647                 OUT_RING  (chan, dst_offset);
648                 OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
649                 OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
650                 OUT_RING  (chan, PAGE_SIZE); /* line_length */
651                 OUT_RING  (chan, line_count);
652                 OUT_RING  (chan, 0x00000101);
653                 OUT_RING  (chan, 0x00000000);
654                 BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
655                 OUT_RING  (chan, 0);
656
657                 page_count -= line_count;
658                 src_offset += (PAGE_SIZE * line_count);
659                 dst_offset += (PAGE_SIZE * line_count);
660         }
661
662         return 0;
663 }
664
665 static int
666 nouveau_vma_getmap(struct nouveau_channel *chan, struct nouveau_bo *nvbo,
667                    struct ttm_mem_reg *mem, struct nouveau_vma *vma)
668 {
669         struct nouveau_mem *node = mem->mm_node;
670         int ret;
671
672         ret = nouveau_vm_get(chan->vm, mem->num_pages << PAGE_SHIFT,
673                              node->page_shift, NV_MEM_ACCESS_RO, vma);
674         if (ret)
675                 return ret;
676
677         if (mem->mem_type == TTM_PL_VRAM)
678                 nouveau_vm_map(vma, node);
679         else
680                 nouveau_vm_map_sg(vma, 0, mem->num_pages << PAGE_SHIFT,
681                                   node, node->pages);
682
683         return 0;
684 }
685
686 static int
687 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
688                      bool no_wait_reserve, bool no_wait_gpu,
689                      struct ttm_mem_reg *new_mem)
690 {
691         struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
692         struct nouveau_bo *nvbo = nouveau_bo(bo);
693         struct ttm_mem_reg *old_mem = &bo->mem;
694         struct nouveau_channel *chan;
695         int ret;
696
697         chan = nvbo->channel;
698         if (!chan) {
699                 chan = dev_priv->channel;
700                 mutex_lock_nested(&chan->mutex, NOUVEAU_KCHANNEL_MUTEX);
701         }
702
703         /* create temporary vmas for the transfer and attach them to the
704          * old nouveau_mem node, these will get cleaned up after ttm has
705          * destroyed the ttm_mem_reg
706          */
707         if (dev_priv->card_type >= NV_50) {
708                 struct nouveau_mem *node = old_mem->mm_node;
709
710                 ret = nouveau_vma_getmap(chan, nvbo, old_mem, &node->vma[0]);
711                 if (ret)
712                         goto out;
713
714                 ret = nouveau_vma_getmap(chan, nvbo, new_mem, &node->vma[1]);
715                 if (ret)
716                         goto out;
717         }
718
719         if (dev_priv->card_type < NV_50)
720                 ret = nv04_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
721         else
722         if (dev_priv->card_type < NV_C0)
723                 ret = nv50_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
724         else
725                 ret = nvc0_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
726         if (ret == 0) {
727                 ret = nouveau_bo_move_accel_cleanup(chan, nvbo, evict,
728                                                     no_wait_reserve,
729                                                     no_wait_gpu, new_mem);
730         }
731
732 out:
733         if (chan == dev_priv->channel)
734                 mutex_unlock(&chan->mutex);
735         return ret;
736 }
737
738 static int
739 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
740                       bool no_wait_reserve, bool no_wait_gpu,
741                       struct ttm_mem_reg *new_mem)
742 {
743         u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
744         struct ttm_placement placement;
745         struct ttm_mem_reg tmp_mem;
746         int ret;
747
748         placement.fpfn = placement.lpfn = 0;
749         placement.num_placement = placement.num_busy_placement = 1;
750         placement.placement = placement.busy_placement = &placement_memtype;
751
752         tmp_mem = *new_mem;
753         tmp_mem.mm_node = NULL;
754         ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_reserve, no_wait_gpu);
755         if (ret)
756                 return ret;
757
758         ret = ttm_tt_bind(bo->ttm, &tmp_mem);
759         if (ret)
760                 goto out;
761
762         ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_reserve, no_wait_gpu, &tmp_mem);
763         if (ret)
764                 goto out;
765
766         ret = ttm_bo_move_ttm(bo, true, no_wait_reserve, no_wait_gpu, new_mem);
767 out:
768         ttm_bo_mem_put(bo, &tmp_mem);
769         return ret;
770 }
771
772 static int
773 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
774                       bool no_wait_reserve, bool no_wait_gpu,
775                       struct ttm_mem_reg *new_mem)
776 {
777         u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
778         struct ttm_placement placement;
779         struct ttm_mem_reg tmp_mem;
780         int ret;
781
782         placement.fpfn = placement.lpfn = 0;
783         placement.num_placement = placement.num_busy_placement = 1;
784         placement.placement = placement.busy_placement = &placement_memtype;
785
786         tmp_mem = *new_mem;
787         tmp_mem.mm_node = NULL;
788         ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_reserve, no_wait_gpu);
789         if (ret)
790                 return ret;
791
792         ret = ttm_bo_move_ttm(bo, true, no_wait_reserve, no_wait_gpu, &tmp_mem);
793         if (ret)
794                 goto out;
795
796         ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_reserve, no_wait_gpu, new_mem);
797         if (ret)
798                 goto out;
799
800 out:
801         ttm_bo_mem_put(bo, &tmp_mem);
802         return ret;
803 }
804
805 static void
806 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem)
807 {
808         struct nouveau_mem *node = new_mem->mm_node;
809         struct nouveau_bo *nvbo = nouveau_bo(bo);
810         struct nouveau_vma *vma;
811
812         list_for_each_entry(vma, &nvbo->vma_list, head) {
813                 if (new_mem->mem_type == TTM_PL_VRAM) {
814                         nouveau_vm_map(vma, new_mem->mm_node);
815                 } else
816                 if (new_mem->mem_type == TTM_PL_TT &&
817                     nvbo->page_shift == vma->vm->spg_shift) {
818                         nouveau_vm_map_sg(vma, 0, new_mem->
819                                           num_pages << PAGE_SHIFT,
820                                           node, node->pages);
821                 } else {
822                         nouveau_vm_unmap(vma);
823                 }
824         }
825 }
826
827 static int
828 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
829                    struct nouveau_tile_reg **new_tile)
830 {
831         struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
832         struct drm_device *dev = dev_priv->dev;
833         struct nouveau_bo *nvbo = nouveau_bo(bo);
834         u64 offset = new_mem->start << PAGE_SHIFT;
835
836         *new_tile = NULL;
837         if (new_mem->mem_type != TTM_PL_VRAM)
838                 return 0;
839
840         if (dev_priv->card_type >= NV_10) {
841                 *new_tile = nv10_mem_set_tiling(dev, offset, new_mem->size,
842                                                 nvbo->tile_mode,
843                                                 nvbo->tile_flags);
844         }
845
846         return 0;
847 }
848
849 static void
850 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
851                       struct nouveau_tile_reg *new_tile,
852                       struct nouveau_tile_reg **old_tile)
853 {
854         struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
855         struct drm_device *dev = dev_priv->dev;
856
857         nv10_mem_put_tile_region(dev, *old_tile, bo->sync_obj);
858         *old_tile = new_tile;
859 }
860
861 static int
862 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
863                 bool no_wait_reserve, bool no_wait_gpu,
864                 struct ttm_mem_reg *new_mem)
865 {
866         struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
867         struct nouveau_bo *nvbo = nouveau_bo(bo);
868         struct ttm_mem_reg *old_mem = &bo->mem;
869         struct nouveau_tile_reg *new_tile = NULL;
870         int ret = 0;
871
872         if (dev_priv->card_type < NV_50) {
873                 ret = nouveau_bo_vm_bind(bo, new_mem, &new_tile);
874                 if (ret)
875                         return ret;
876         }
877
878         /* Fake bo copy. */
879         if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
880                 BUG_ON(bo->mem.mm_node != NULL);
881                 bo->mem = *new_mem;
882                 new_mem->mm_node = NULL;
883                 goto out;
884         }
885
886         /* Software copy if the card isn't up and running yet. */
887         if (!dev_priv->channel) {
888                 ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
889                 goto out;
890         }
891
892         /* Hardware assisted copy. */
893         if (new_mem->mem_type == TTM_PL_SYSTEM)
894                 ret = nouveau_bo_move_flipd(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
895         else if (old_mem->mem_type == TTM_PL_SYSTEM)
896                 ret = nouveau_bo_move_flips(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
897         else
898                 ret = nouveau_bo_move_m2mf(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
899
900         if (!ret)
901                 goto out;
902
903         /* Fallback to software copy. */
904         ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
905
906 out:
907         if (dev_priv->card_type < NV_50) {
908                 if (ret)
909                         nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
910                 else
911                         nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
912         }
913
914         return ret;
915 }
916
917 static int
918 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
919 {
920         return 0;
921 }
922
923 static int
924 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
925 {
926         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
927         struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
928         struct drm_device *dev = dev_priv->dev;
929         int ret;
930
931         mem->bus.addr = NULL;
932         mem->bus.offset = 0;
933         mem->bus.size = mem->num_pages << PAGE_SHIFT;
934         mem->bus.base = 0;
935         mem->bus.is_iomem = false;
936         if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
937                 return -EINVAL;
938         switch (mem->mem_type) {
939         case TTM_PL_SYSTEM:
940                 /* System memory */
941                 return 0;
942         case TTM_PL_TT:
943 #if __OS_HAS_AGP
944                 if (dev_priv->gart_info.type == NOUVEAU_GART_AGP) {
945                         mem->bus.offset = mem->start << PAGE_SHIFT;
946                         mem->bus.base = dev_priv->gart_info.aper_base;
947                         mem->bus.is_iomem = true;
948                 }
949 #endif
950                 break;
951         case TTM_PL_VRAM:
952         {
953                 struct nouveau_mem *node = mem->mm_node;
954                 u8 page_shift;
955
956                 if (!dev_priv->bar1_vm) {
957                         mem->bus.offset = mem->start << PAGE_SHIFT;
958                         mem->bus.base = pci_resource_start(dev->pdev, 1);
959                         mem->bus.is_iomem = true;
960                         break;
961                 }
962
963                 if (dev_priv->card_type >= NV_C0)
964                         page_shift = node->page_shift;
965                 else
966                         page_shift = 12;
967
968                 ret = nouveau_vm_get(dev_priv->bar1_vm, mem->bus.size,
969                                      page_shift, NV_MEM_ACCESS_RW,
970                                      &node->bar_vma);
971                 if (ret)
972                         return ret;
973
974                 nouveau_vm_map(&node->bar_vma, node);
975                 if (ret) {
976                         nouveau_vm_put(&node->bar_vma);
977                         return ret;
978                 }
979
980                 mem->bus.offset = node->bar_vma.offset;
981                 if (dev_priv->card_type == NV_50) /*XXX*/
982                         mem->bus.offset -= 0x0020000000ULL;
983                 mem->bus.base = pci_resource_start(dev->pdev, 1);
984                 mem->bus.is_iomem = true;
985         }
986                 break;
987         default:
988                 return -EINVAL;
989         }
990         return 0;
991 }
992
993 static void
994 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
995 {
996         struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
997         struct nouveau_mem *node = mem->mm_node;
998
999         if (!dev_priv->bar1_vm || mem->mem_type != TTM_PL_VRAM)
1000                 return;
1001
1002         if (!node->bar_vma.node)
1003                 return;
1004
1005         nouveau_vm_unmap(&node->bar_vma);
1006         nouveau_vm_put(&node->bar_vma);
1007 }
1008
1009 static int
1010 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1011 {
1012         struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
1013         struct nouveau_bo *nvbo = nouveau_bo(bo);
1014
1015         /* as long as the bo isn't in vram, and isn't tiled, we've got
1016          * nothing to do here.
1017          */
1018         if (bo->mem.mem_type != TTM_PL_VRAM) {
1019                 if (dev_priv->card_type < NV_50 ||
1020                     !nouveau_bo_tile_layout(nvbo))
1021                         return 0;
1022         }
1023
1024         /* make sure bo is in mappable vram */
1025         if (bo->mem.start + bo->mem.num_pages < dev_priv->fb_mappable_pages)
1026                 return 0;
1027
1028
1029         nvbo->placement.fpfn = 0;
1030         nvbo->placement.lpfn = dev_priv->fb_mappable_pages;
1031         nouveau_bo_placement_set(nvbo, TTM_PL_VRAM, 0);
1032         return nouveau_bo_validate(nvbo, false, true, false);
1033 }
1034
1035 void
1036 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
1037 {
1038         struct nouveau_fence *old_fence;
1039
1040         if (likely(fence))
1041                 nouveau_fence_ref(fence);
1042
1043         spin_lock(&nvbo->bo.bdev->fence_lock);
1044         old_fence = nvbo->bo.sync_obj;
1045         nvbo->bo.sync_obj = fence;
1046         spin_unlock(&nvbo->bo.bdev->fence_lock);
1047
1048         nouveau_fence_unref(&old_fence);
1049 }
1050
1051 struct ttm_bo_driver nouveau_bo_driver = {
1052         .ttm_tt_create = &nouveau_ttm_tt_create,
1053         .invalidate_caches = nouveau_bo_invalidate_caches,
1054         .init_mem_type = nouveau_bo_init_mem_type,
1055         .evict_flags = nouveau_bo_evict_flags,
1056         .move_notify = nouveau_bo_move_ntfy,
1057         .move = nouveau_bo_move,
1058         .verify_access = nouveau_bo_verify_access,
1059         .sync_obj_signaled = __nouveau_fence_signalled,
1060         .sync_obj_wait = __nouveau_fence_wait,
1061         .sync_obj_flush = __nouveau_fence_flush,
1062         .sync_obj_unref = __nouveau_fence_unref,
1063         .sync_obj_ref = __nouveau_fence_ref,
1064         .fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1065         .io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1066         .io_mem_free = &nouveau_ttm_io_mem_free,
1067 };
1068
1069 struct nouveau_vma *
1070 nouveau_bo_vma_find(struct nouveau_bo *nvbo, struct nouveau_vm *vm)
1071 {
1072         struct nouveau_vma *vma;
1073         list_for_each_entry(vma, &nvbo->vma_list, head) {
1074                 if (vma->vm == vm)
1075                         return vma;
1076         }
1077
1078         return NULL;
1079 }
1080
1081 int
1082 nouveau_bo_vma_add(struct nouveau_bo *nvbo, struct nouveau_vm *vm,
1083                    struct nouveau_vma *vma)
1084 {
1085         const u32 size = nvbo->bo.mem.num_pages << PAGE_SHIFT;
1086         struct nouveau_mem *node = nvbo->bo.mem.mm_node;
1087         int ret;
1088
1089         ret = nouveau_vm_get(vm, size, nvbo->page_shift,
1090                              NV_MEM_ACCESS_RW, vma);
1091         if (ret)
1092                 return ret;
1093
1094         if (nvbo->bo.mem.mem_type == TTM_PL_VRAM)
1095                 nouveau_vm_map(vma, nvbo->bo.mem.mm_node);
1096         else
1097         if (nvbo->bo.mem.mem_type == TTM_PL_TT)
1098                 nouveau_vm_map_sg(vma, 0, size, node, node->pages);
1099
1100         list_add_tail(&vma->head, &nvbo->vma_list);
1101         vma->refcount = 1;
1102         return 0;
1103 }
1104
1105 void
1106 nouveau_bo_vma_del(struct nouveau_bo *nvbo, struct nouveau_vma *vma)
1107 {
1108         if (vma->node) {
1109                 if (nvbo->bo.mem.mem_type != TTM_PL_SYSTEM) {
1110                         spin_lock(&nvbo->bo.bdev->fence_lock);
1111                         ttm_bo_wait(&nvbo->bo, false, false, false);
1112                         spin_unlock(&nvbo->bo.bdev->fence_lock);
1113                         nouveau_vm_unmap(vma);
1114                 }
1115
1116                 nouveau_vm_put(vma);
1117                 list_del(&vma->head);
1118         }
1119 }