Merge tag 'v3.2-rc6' of /home/airlied/devel/kernel/linux-2.6 into drm-core-next
[pandora-kernel.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 #include "ttm/ttm_page_alloc.h"
32
33 #include "nouveau_drm.h"
34 #include "nouveau_drv.h"
35 #include "nouveau_dma.h"
36 #include "nouveau_mm.h"
37 #include "nouveau_vm.h"
38
39 #include <linux/log2.h>
40 #include <linux/slab.h>
41
42 static void
43 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
44 {
45         struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
46         struct drm_device *dev = dev_priv->dev;
47         struct nouveau_bo *nvbo = nouveau_bo(bo);
48
49         if (unlikely(nvbo->gem))
50                 DRM_ERROR("bo %p still attached to GEM object\n", bo);
51
52         nv10_mem_put_tile_region(dev, nvbo->tile, NULL);
53         kfree(nvbo);
54 }
55
56 static void
57 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
58                        int *align, int *size)
59 {
60         struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
61
62         if (dev_priv->card_type < NV_50) {
63                 if (nvbo->tile_mode) {
64                         if (dev_priv->chipset >= 0x40) {
65                                 *align = 65536;
66                                 *size = roundup(*size, 64 * nvbo->tile_mode);
67
68                         } else if (dev_priv->chipset >= 0x30) {
69                                 *align = 32768;
70                                 *size = roundup(*size, 64 * nvbo->tile_mode);
71
72                         } else if (dev_priv->chipset >= 0x20) {
73                                 *align = 16384;
74                                 *size = roundup(*size, 64 * nvbo->tile_mode);
75
76                         } else if (dev_priv->chipset >= 0x10) {
77                                 *align = 16384;
78                                 *size = roundup(*size, 32 * nvbo->tile_mode);
79                         }
80                 }
81         } else {
82                 *size = roundup(*size, (1 << nvbo->page_shift));
83                 *align = max((1 <<  nvbo->page_shift), *align);
84         }
85
86         *size = roundup(*size, PAGE_SIZE);
87 }
88
89 int
90 nouveau_bo_new(struct drm_device *dev, int size, int align,
91                uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
92                struct nouveau_bo **pnvbo)
93 {
94         struct drm_nouveau_private *dev_priv = dev->dev_private;
95         struct nouveau_bo *nvbo;
96         size_t acc_size;
97         int ret;
98
99         nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
100         if (!nvbo)
101                 return -ENOMEM;
102         INIT_LIST_HEAD(&nvbo->head);
103         INIT_LIST_HEAD(&nvbo->entry);
104         INIT_LIST_HEAD(&nvbo->vma_list);
105         nvbo->tile_mode = tile_mode;
106         nvbo->tile_flags = tile_flags;
107         nvbo->bo.bdev = &dev_priv->ttm.bdev;
108
109         nvbo->page_shift = 12;
110         if (dev_priv->bar1_vm) {
111                 if (!(flags & TTM_PL_FLAG_TT) && size > 256 * 1024)
112                         nvbo->page_shift = dev_priv->bar1_vm->lpg_shift;
113         }
114
115         nouveau_bo_fixup_align(nvbo, flags, &align, &size);
116         nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
117         nouveau_bo_placement_set(nvbo, flags, 0);
118
119         acc_size = ttm_bo_dma_acc_size(&dev_priv->ttm.bdev, size,
120                                        sizeof(struct nouveau_bo));
121
122         ret = ttm_bo_init(&dev_priv->ttm.bdev, &nvbo->bo, size,
123                           ttm_bo_type_device, &nvbo->placement,
124                           align >> PAGE_SHIFT, 0, false, NULL, acc_size,
125                           nouveau_bo_del_ttm);
126         if (ret) {
127                 /* ttm will call nouveau_bo_del_ttm if it fails.. */
128                 return ret;
129         }
130
131         *pnvbo = nvbo;
132         return 0;
133 }
134
135 static void
136 set_placement_list(uint32_t *pl, unsigned *n, uint32_t type, uint32_t flags)
137 {
138         *n = 0;
139
140         if (type & TTM_PL_FLAG_VRAM)
141                 pl[(*n)++] = TTM_PL_FLAG_VRAM | flags;
142         if (type & TTM_PL_FLAG_TT)
143                 pl[(*n)++] = TTM_PL_FLAG_TT | flags;
144         if (type & TTM_PL_FLAG_SYSTEM)
145                 pl[(*n)++] = TTM_PL_FLAG_SYSTEM | flags;
146 }
147
148 static void
149 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
150 {
151         struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
152         int vram_pages = dev_priv->vram_size >> PAGE_SHIFT;
153
154         if (dev_priv->card_type == NV_10 &&
155             nvbo->tile_mode && (type & TTM_PL_FLAG_VRAM) &&
156             nvbo->bo.mem.num_pages < vram_pages / 4) {
157                 /*
158                  * Make sure that the color and depth buffers are handled
159                  * by independent memory controller units. Up to a 9x
160                  * speed up when alpha-blending and depth-test are enabled
161                  * at the same time.
162                  */
163                 if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
164                         nvbo->placement.fpfn = vram_pages / 2;
165                         nvbo->placement.lpfn = ~0;
166                 } else {
167                         nvbo->placement.fpfn = 0;
168                         nvbo->placement.lpfn = vram_pages / 2;
169                 }
170         }
171 }
172
173 void
174 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
175 {
176         struct ttm_placement *pl = &nvbo->placement;
177         uint32_t flags = TTM_PL_MASK_CACHING |
178                 (nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
179
180         pl->placement = nvbo->placements;
181         set_placement_list(nvbo->placements, &pl->num_placement,
182                            type, flags);
183
184         pl->busy_placement = nvbo->busy_placements;
185         set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
186                            type | busy, flags);
187
188         set_placement_range(nvbo, type);
189 }
190
191 int
192 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype)
193 {
194         struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
195         struct ttm_buffer_object *bo = &nvbo->bo;
196         int ret;
197
198         if (nvbo->pin_refcnt && !(memtype & (1 << bo->mem.mem_type))) {
199                 NV_ERROR(nouveau_bdev(bo->bdev)->dev,
200                          "bo %p pinned elsewhere: 0x%08x vs 0x%08x\n", bo,
201                          1 << bo->mem.mem_type, memtype);
202                 return -EINVAL;
203         }
204
205         if (nvbo->pin_refcnt++)
206                 return 0;
207
208         ret = ttm_bo_reserve(bo, false, false, false, 0);
209         if (ret)
210                 goto out;
211
212         nouveau_bo_placement_set(nvbo, memtype, 0);
213
214         ret = nouveau_bo_validate(nvbo, false, false, false);
215         if (ret == 0) {
216                 switch (bo->mem.mem_type) {
217                 case TTM_PL_VRAM:
218                         dev_priv->fb_aper_free -= bo->mem.size;
219                         break;
220                 case TTM_PL_TT:
221                         dev_priv->gart_info.aper_free -= bo->mem.size;
222                         break;
223                 default:
224                         break;
225                 }
226         }
227         ttm_bo_unreserve(bo);
228 out:
229         if (unlikely(ret))
230                 nvbo->pin_refcnt--;
231         return ret;
232 }
233
234 int
235 nouveau_bo_unpin(struct nouveau_bo *nvbo)
236 {
237         struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
238         struct ttm_buffer_object *bo = &nvbo->bo;
239         int ret;
240
241         if (--nvbo->pin_refcnt)
242                 return 0;
243
244         ret = ttm_bo_reserve(bo, false, false, false, 0);
245         if (ret)
246                 return ret;
247
248         nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
249
250         ret = nouveau_bo_validate(nvbo, false, false, false);
251         if (ret == 0) {
252                 switch (bo->mem.mem_type) {
253                 case TTM_PL_VRAM:
254                         dev_priv->fb_aper_free += bo->mem.size;
255                         break;
256                 case TTM_PL_TT:
257                         dev_priv->gart_info.aper_free += bo->mem.size;
258                         break;
259                 default:
260                         break;
261                 }
262         }
263
264         ttm_bo_unreserve(bo);
265         return ret;
266 }
267
268 int
269 nouveau_bo_map(struct nouveau_bo *nvbo)
270 {
271         int ret;
272
273         ret = ttm_bo_reserve(&nvbo->bo, false, false, false, 0);
274         if (ret)
275                 return ret;
276
277         ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
278         ttm_bo_unreserve(&nvbo->bo);
279         return ret;
280 }
281
282 void
283 nouveau_bo_unmap(struct nouveau_bo *nvbo)
284 {
285         if (nvbo)
286                 ttm_bo_kunmap(&nvbo->kmap);
287 }
288
289 int
290 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
291                     bool no_wait_reserve, bool no_wait_gpu)
292 {
293         int ret;
294
295         ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, interruptible,
296                               no_wait_reserve, no_wait_gpu);
297         if (ret)
298                 return ret;
299
300         return 0;
301 }
302
303 u16
304 nouveau_bo_rd16(struct nouveau_bo *nvbo, unsigned index)
305 {
306         bool is_iomem;
307         u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
308         mem = &mem[index];
309         if (is_iomem)
310                 return ioread16_native((void __force __iomem *)mem);
311         else
312                 return *mem;
313 }
314
315 void
316 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
317 {
318         bool is_iomem;
319         u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
320         mem = &mem[index];
321         if (is_iomem)
322                 iowrite16_native(val, (void __force __iomem *)mem);
323         else
324                 *mem = val;
325 }
326
327 u32
328 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
329 {
330         bool is_iomem;
331         u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
332         mem = &mem[index];
333         if (is_iomem)
334                 return ioread32_native((void __force __iomem *)mem);
335         else
336                 return *mem;
337 }
338
339 void
340 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
341 {
342         bool is_iomem;
343         u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
344         mem = &mem[index];
345         if (is_iomem)
346                 iowrite32_native(val, (void __force __iomem *)mem);
347         else
348                 *mem = val;
349 }
350
351 static struct ttm_tt *
352 nouveau_ttm_tt_create(struct ttm_bo_device *bdev,
353                       unsigned long size, uint32_t page_flags,
354                       struct page *dummy_read_page)
355 {
356         struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
357         struct drm_device *dev = dev_priv->dev;
358
359         switch (dev_priv->gart_info.type) {
360 #if __OS_HAS_AGP
361         case NOUVEAU_GART_AGP:
362                 return ttm_agp_tt_create(bdev, dev->agp->bridge,
363                                          size, page_flags, dummy_read_page);
364 #endif
365         case NOUVEAU_GART_PDMA:
366         case NOUVEAU_GART_HW:
367                 return nouveau_sgdma_create_ttm(bdev, size, page_flags,
368                                                 dummy_read_page);
369         default:
370                 NV_ERROR(dev, "Unknown GART type %d\n",
371                          dev_priv->gart_info.type);
372                 break;
373         }
374
375         return NULL;
376 }
377
378 static int
379 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
380 {
381         /* We'll do this from user space. */
382         return 0;
383 }
384
385 static int
386 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
387                          struct ttm_mem_type_manager *man)
388 {
389         struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
390         struct drm_device *dev = dev_priv->dev;
391
392         switch (type) {
393         case TTM_PL_SYSTEM:
394                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
395                 man->available_caching = TTM_PL_MASK_CACHING;
396                 man->default_caching = TTM_PL_FLAG_CACHED;
397                 break;
398         case TTM_PL_VRAM:
399                 if (dev_priv->card_type >= NV_50) {
400                         man->func = &nouveau_vram_manager;
401                         man->io_reserve_fastpath = false;
402                         man->use_io_reserve_lru = true;
403                 } else {
404                         man->func = &ttm_bo_manager_func;
405                 }
406                 man->flags = TTM_MEMTYPE_FLAG_FIXED |
407                              TTM_MEMTYPE_FLAG_MAPPABLE;
408                 man->available_caching = TTM_PL_FLAG_UNCACHED |
409                                          TTM_PL_FLAG_WC;
410                 man->default_caching = TTM_PL_FLAG_WC;
411                 break;
412         case TTM_PL_TT:
413                 if (dev_priv->card_type >= NV_50)
414                         man->func = &nouveau_gart_manager;
415                 else
416                         man->func = &ttm_bo_manager_func;
417                 switch (dev_priv->gart_info.type) {
418                 case NOUVEAU_GART_AGP:
419                         man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
420                         man->available_caching = TTM_PL_FLAG_UNCACHED |
421                                 TTM_PL_FLAG_WC;
422                         man->default_caching = TTM_PL_FLAG_WC;
423                         break;
424                 case NOUVEAU_GART_PDMA:
425                 case NOUVEAU_GART_HW:
426                         man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
427                                      TTM_MEMTYPE_FLAG_CMA;
428                         man->available_caching = TTM_PL_MASK_CACHING;
429                         man->default_caching = TTM_PL_FLAG_CACHED;
430                         break;
431                 default:
432                         NV_ERROR(dev, "Unknown GART type: %d\n",
433                                  dev_priv->gart_info.type);
434                         return -EINVAL;
435                 }
436                 break;
437         default:
438                 NV_ERROR(dev, "Unsupported memory type %u\n", (unsigned)type);
439                 return -EINVAL;
440         }
441         return 0;
442 }
443
444 static void
445 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
446 {
447         struct nouveau_bo *nvbo = nouveau_bo(bo);
448
449         switch (bo->mem.mem_type) {
450         case TTM_PL_VRAM:
451                 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
452                                          TTM_PL_FLAG_SYSTEM);
453                 break;
454         default:
455                 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
456                 break;
457         }
458
459         *pl = nvbo->placement;
460 }
461
462
463 /* GPU-assisted copy using NV_MEMORY_TO_MEMORY_FORMAT, can access
464  * TTM_PL_{VRAM,TT} directly.
465  */
466
467 static int
468 nouveau_bo_move_accel_cleanup(struct nouveau_channel *chan,
469                               struct nouveau_bo *nvbo, bool evict,
470                               bool no_wait_reserve, bool no_wait_gpu,
471                               struct ttm_mem_reg *new_mem)
472 {
473         struct nouveau_fence *fence = NULL;
474         int ret;
475
476         ret = nouveau_fence_new(chan, &fence, true);
477         if (ret)
478                 return ret;
479
480         ret = ttm_bo_move_accel_cleanup(&nvbo->bo, fence, NULL, evict,
481                                         no_wait_reserve, no_wait_gpu, new_mem);
482         nouveau_fence_unref(&fence);
483         return ret;
484 }
485
486 static int
487 nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
488                   struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
489 {
490         struct nouveau_mem *node = old_mem->mm_node;
491         u64 src_offset = node->vma[0].offset;
492         u64 dst_offset = node->vma[1].offset;
493         u32 page_count = new_mem->num_pages;
494         int ret;
495
496         page_count = new_mem->num_pages;
497         while (page_count) {
498                 int line_count = (page_count > 2047) ? 2047 : page_count;
499
500                 ret = RING_SPACE(chan, 12);
501                 if (ret)
502                         return ret;
503
504                 BEGIN_NVC0(chan, 2, NvSubM2MF, 0x0238, 2);
505                 OUT_RING  (chan, upper_32_bits(dst_offset));
506                 OUT_RING  (chan, lower_32_bits(dst_offset));
507                 BEGIN_NVC0(chan, 2, NvSubM2MF, 0x030c, 6);
508                 OUT_RING  (chan, upper_32_bits(src_offset));
509                 OUT_RING  (chan, lower_32_bits(src_offset));
510                 OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
511                 OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
512                 OUT_RING  (chan, PAGE_SIZE); /* line_length */
513                 OUT_RING  (chan, line_count);
514                 BEGIN_NVC0(chan, 2, NvSubM2MF, 0x0300, 1);
515                 OUT_RING  (chan, 0x00100110);
516
517                 page_count -= line_count;
518                 src_offset += (PAGE_SIZE * line_count);
519                 dst_offset += (PAGE_SIZE * line_count);
520         }
521
522         return 0;
523 }
524
525 static int
526 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
527                   struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
528 {
529         struct nouveau_mem *node = old_mem->mm_node;
530         struct nouveau_bo *nvbo = nouveau_bo(bo);
531         u64 length = (new_mem->num_pages << PAGE_SHIFT);
532         u64 src_offset = node->vma[0].offset;
533         u64 dst_offset = node->vma[1].offset;
534         int ret;
535
536         while (length) {
537                 u32 amount, stride, height;
538
539                 amount  = min(length, (u64)(4 * 1024 * 1024));
540                 stride  = 16 * 4;
541                 height  = amount / stride;
542
543                 if (new_mem->mem_type == TTM_PL_VRAM &&
544                     nouveau_bo_tile_layout(nvbo)) {
545                         ret = RING_SPACE(chan, 8);
546                         if (ret)
547                                 return ret;
548
549                         BEGIN_RING(chan, NvSubM2MF, 0x0200, 7);
550                         OUT_RING  (chan, 0);
551                         OUT_RING  (chan, 0);
552                         OUT_RING  (chan, stride);
553                         OUT_RING  (chan, height);
554                         OUT_RING  (chan, 1);
555                         OUT_RING  (chan, 0);
556                         OUT_RING  (chan, 0);
557                 } else {
558                         ret = RING_SPACE(chan, 2);
559                         if (ret)
560                                 return ret;
561
562                         BEGIN_RING(chan, NvSubM2MF, 0x0200, 1);
563                         OUT_RING  (chan, 1);
564                 }
565                 if (old_mem->mem_type == TTM_PL_VRAM &&
566                     nouveau_bo_tile_layout(nvbo)) {
567                         ret = RING_SPACE(chan, 8);
568                         if (ret)
569                                 return ret;
570
571                         BEGIN_RING(chan, NvSubM2MF, 0x021c, 7);
572                         OUT_RING  (chan, 0);
573                         OUT_RING  (chan, 0);
574                         OUT_RING  (chan, stride);
575                         OUT_RING  (chan, height);
576                         OUT_RING  (chan, 1);
577                         OUT_RING  (chan, 0);
578                         OUT_RING  (chan, 0);
579                 } else {
580                         ret = RING_SPACE(chan, 2);
581                         if (ret)
582                                 return ret;
583
584                         BEGIN_RING(chan, NvSubM2MF, 0x021c, 1);
585                         OUT_RING  (chan, 1);
586                 }
587
588                 ret = RING_SPACE(chan, 14);
589                 if (ret)
590                         return ret;
591
592                 BEGIN_RING(chan, NvSubM2MF, 0x0238, 2);
593                 OUT_RING  (chan, upper_32_bits(src_offset));
594                 OUT_RING  (chan, upper_32_bits(dst_offset));
595                 BEGIN_RING(chan, NvSubM2MF, 0x030c, 8);
596                 OUT_RING  (chan, lower_32_bits(src_offset));
597                 OUT_RING  (chan, lower_32_bits(dst_offset));
598                 OUT_RING  (chan, stride);
599                 OUT_RING  (chan, stride);
600                 OUT_RING  (chan, stride);
601                 OUT_RING  (chan, height);
602                 OUT_RING  (chan, 0x00000101);
603                 OUT_RING  (chan, 0x00000000);
604                 BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
605                 OUT_RING  (chan, 0);
606
607                 length -= amount;
608                 src_offset += amount;
609                 dst_offset += amount;
610         }
611
612         return 0;
613 }
614
615 static inline uint32_t
616 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
617                       struct nouveau_channel *chan, struct ttm_mem_reg *mem)
618 {
619         if (mem->mem_type == TTM_PL_TT)
620                 return chan->gart_handle;
621         return chan->vram_handle;
622 }
623
624 static int
625 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
626                   struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
627 {
628         u32 src_offset = old_mem->start << PAGE_SHIFT;
629         u32 dst_offset = new_mem->start << PAGE_SHIFT;
630         u32 page_count = new_mem->num_pages;
631         int ret;
632
633         ret = RING_SPACE(chan, 3);
634         if (ret)
635                 return ret;
636
637         BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
638         OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
639         OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
640
641         page_count = new_mem->num_pages;
642         while (page_count) {
643                 int line_count = (page_count > 2047) ? 2047 : page_count;
644
645                 ret = RING_SPACE(chan, 11);
646                 if (ret)
647                         return ret;
648
649                 BEGIN_RING(chan, NvSubM2MF,
650                                  NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
651                 OUT_RING  (chan, src_offset);
652                 OUT_RING  (chan, dst_offset);
653                 OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
654                 OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
655                 OUT_RING  (chan, PAGE_SIZE); /* line_length */
656                 OUT_RING  (chan, line_count);
657                 OUT_RING  (chan, 0x00000101);
658                 OUT_RING  (chan, 0x00000000);
659                 BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
660                 OUT_RING  (chan, 0);
661
662                 page_count -= line_count;
663                 src_offset += (PAGE_SIZE * line_count);
664                 dst_offset += (PAGE_SIZE * line_count);
665         }
666
667         return 0;
668 }
669
670 static int
671 nouveau_vma_getmap(struct nouveau_channel *chan, struct nouveau_bo *nvbo,
672                    struct ttm_mem_reg *mem, struct nouveau_vma *vma)
673 {
674         struct nouveau_mem *node = mem->mm_node;
675         int ret;
676
677         ret = nouveau_vm_get(chan->vm, mem->num_pages << PAGE_SHIFT,
678                              node->page_shift, NV_MEM_ACCESS_RO, vma);
679         if (ret)
680                 return ret;
681
682         if (mem->mem_type == TTM_PL_VRAM)
683                 nouveau_vm_map(vma, node);
684         else
685                 nouveau_vm_map_sg(vma, 0, mem->num_pages << PAGE_SHIFT,
686                                   node, node->pages);
687
688         return 0;
689 }
690
691 static int
692 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
693                      bool no_wait_reserve, bool no_wait_gpu,
694                      struct ttm_mem_reg *new_mem)
695 {
696         struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
697         struct nouveau_bo *nvbo = nouveau_bo(bo);
698         struct ttm_mem_reg *old_mem = &bo->mem;
699         struct nouveau_channel *chan;
700         int ret;
701
702         chan = nvbo->channel;
703         if (!chan) {
704                 chan = dev_priv->channel;
705                 mutex_lock_nested(&chan->mutex, NOUVEAU_KCHANNEL_MUTEX);
706         }
707
708         /* create temporary vmas for the transfer and attach them to the
709          * old nouveau_mem node, these will get cleaned up after ttm has
710          * destroyed the ttm_mem_reg
711          */
712         if (dev_priv->card_type >= NV_50) {
713                 struct nouveau_mem *node = old_mem->mm_node;
714
715                 ret = nouveau_vma_getmap(chan, nvbo, old_mem, &node->vma[0]);
716                 if (ret)
717                         goto out;
718
719                 ret = nouveau_vma_getmap(chan, nvbo, new_mem, &node->vma[1]);
720                 if (ret)
721                         goto out;
722         }
723
724         if (dev_priv->card_type < NV_50)
725                 ret = nv04_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
726         else
727         if (dev_priv->card_type < NV_C0)
728                 ret = nv50_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
729         else
730                 ret = nvc0_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
731         if (ret == 0) {
732                 ret = nouveau_bo_move_accel_cleanup(chan, nvbo, evict,
733                                                     no_wait_reserve,
734                                                     no_wait_gpu, new_mem);
735         }
736
737 out:
738         if (chan == dev_priv->channel)
739                 mutex_unlock(&chan->mutex);
740         return ret;
741 }
742
743 static int
744 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
745                       bool no_wait_reserve, bool no_wait_gpu,
746                       struct ttm_mem_reg *new_mem)
747 {
748         u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
749         struct ttm_placement placement;
750         struct ttm_mem_reg tmp_mem;
751         int ret;
752
753         placement.fpfn = placement.lpfn = 0;
754         placement.num_placement = placement.num_busy_placement = 1;
755         placement.placement = placement.busy_placement = &placement_memtype;
756
757         tmp_mem = *new_mem;
758         tmp_mem.mm_node = NULL;
759         ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_reserve, no_wait_gpu);
760         if (ret)
761                 return ret;
762
763         ret = ttm_tt_bind(bo->ttm, &tmp_mem);
764         if (ret)
765                 goto out;
766
767         ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_reserve, no_wait_gpu, &tmp_mem);
768         if (ret)
769                 goto out;
770
771         ret = ttm_bo_move_ttm(bo, true, no_wait_reserve, no_wait_gpu, new_mem);
772 out:
773         ttm_bo_mem_put(bo, &tmp_mem);
774         return ret;
775 }
776
777 static int
778 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
779                       bool no_wait_reserve, bool no_wait_gpu,
780                       struct ttm_mem_reg *new_mem)
781 {
782         u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
783         struct ttm_placement placement;
784         struct ttm_mem_reg tmp_mem;
785         int ret;
786
787         placement.fpfn = placement.lpfn = 0;
788         placement.num_placement = placement.num_busy_placement = 1;
789         placement.placement = placement.busy_placement = &placement_memtype;
790
791         tmp_mem = *new_mem;
792         tmp_mem.mm_node = NULL;
793         ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_reserve, no_wait_gpu);
794         if (ret)
795                 return ret;
796
797         ret = ttm_bo_move_ttm(bo, true, no_wait_reserve, no_wait_gpu, &tmp_mem);
798         if (ret)
799                 goto out;
800
801         ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_reserve, no_wait_gpu, new_mem);
802         if (ret)
803                 goto out;
804
805 out:
806         ttm_bo_mem_put(bo, &tmp_mem);
807         return ret;
808 }
809
810 static void
811 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem)
812 {
813         struct nouveau_mem *node = new_mem->mm_node;
814         struct nouveau_bo *nvbo = nouveau_bo(bo);
815         struct nouveau_vma *vma;
816
817         list_for_each_entry(vma, &nvbo->vma_list, head) {
818                 if (new_mem && new_mem->mem_type == TTM_PL_VRAM) {
819                         nouveau_vm_map(vma, new_mem->mm_node);
820                 } else
821                 if (new_mem && new_mem->mem_type == TTM_PL_TT &&
822                     nvbo->page_shift == vma->vm->spg_shift) {
823                         nouveau_vm_map_sg(vma, 0, new_mem->
824                                           num_pages << PAGE_SHIFT,
825                                           node, node->pages);
826                 } else {
827                         nouveau_vm_unmap(vma);
828                 }
829         }
830 }
831
832 static int
833 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
834                    struct nouveau_tile_reg **new_tile)
835 {
836         struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
837         struct drm_device *dev = dev_priv->dev;
838         struct nouveau_bo *nvbo = nouveau_bo(bo);
839         u64 offset = new_mem->start << PAGE_SHIFT;
840
841         *new_tile = NULL;
842         if (new_mem->mem_type != TTM_PL_VRAM)
843                 return 0;
844
845         if (dev_priv->card_type >= NV_10) {
846                 *new_tile = nv10_mem_set_tiling(dev, offset, new_mem->size,
847                                                 nvbo->tile_mode,
848                                                 nvbo->tile_flags);
849         }
850
851         return 0;
852 }
853
854 static void
855 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
856                       struct nouveau_tile_reg *new_tile,
857                       struct nouveau_tile_reg **old_tile)
858 {
859         struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
860         struct drm_device *dev = dev_priv->dev;
861
862         nv10_mem_put_tile_region(dev, *old_tile, bo->sync_obj);
863         *old_tile = new_tile;
864 }
865
866 static int
867 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
868                 bool no_wait_reserve, bool no_wait_gpu,
869                 struct ttm_mem_reg *new_mem)
870 {
871         struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
872         struct nouveau_bo *nvbo = nouveau_bo(bo);
873         struct ttm_mem_reg *old_mem = &bo->mem;
874         struct nouveau_tile_reg *new_tile = NULL;
875         int ret = 0;
876
877         if (dev_priv->card_type < NV_50) {
878                 ret = nouveau_bo_vm_bind(bo, new_mem, &new_tile);
879                 if (ret)
880                         return ret;
881         }
882
883         /* Fake bo copy. */
884         if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
885                 BUG_ON(bo->mem.mm_node != NULL);
886                 bo->mem = *new_mem;
887                 new_mem->mm_node = NULL;
888                 goto out;
889         }
890
891         /* Software copy if the card isn't up and running yet. */
892         if (!dev_priv->channel) {
893                 ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
894                 goto out;
895         }
896
897         /* Hardware assisted copy. */
898         if (new_mem->mem_type == TTM_PL_SYSTEM)
899                 ret = nouveau_bo_move_flipd(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
900         else if (old_mem->mem_type == TTM_PL_SYSTEM)
901                 ret = nouveau_bo_move_flips(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
902         else
903                 ret = nouveau_bo_move_m2mf(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
904
905         if (!ret)
906                 goto out;
907
908         /* Fallback to software copy. */
909         ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
910
911 out:
912         if (dev_priv->card_type < NV_50) {
913                 if (ret)
914                         nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
915                 else
916                         nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
917         }
918
919         return ret;
920 }
921
922 static int
923 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
924 {
925         return 0;
926 }
927
928 static int
929 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
930 {
931         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
932         struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
933         struct drm_device *dev = dev_priv->dev;
934         int ret;
935
936         mem->bus.addr = NULL;
937         mem->bus.offset = 0;
938         mem->bus.size = mem->num_pages << PAGE_SHIFT;
939         mem->bus.base = 0;
940         mem->bus.is_iomem = false;
941         if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
942                 return -EINVAL;
943         switch (mem->mem_type) {
944         case TTM_PL_SYSTEM:
945                 /* System memory */
946                 return 0;
947         case TTM_PL_TT:
948 #if __OS_HAS_AGP
949                 if (dev_priv->gart_info.type == NOUVEAU_GART_AGP) {
950                         mem->bus.offset = mem->start << PAGE_SHIFT;
951                         mem->bus.base = dev_priv->gart_info.aper_base;
952                         mem->bus.is_iomem = true;
953                 }
954 #endif
955                 break;
956         case TTM_PL_VRAM:
957         {
958                 struct nouveau_mem *node = mem->mm_node;
959                 u8 page_shift;
960
961                 if (!dev_priv->bar1_vm) {
962                         mem->bus.offset = mem->start << PAGE_SHIFT;
963                         mem->bus.base = pci_resource_start(dev->pdev, 1);
964                         mem->bus.is_iomem = true;
965                         break;
966                 }
967
968                 if (dev_priv->card_type >= NV_C0)
969                         page_shift = node->page_shift;
970                 else
971                         page_shift = 12;
972
973                 ret = nouveau_vm_get(dev_priv->bar1_vm, mem->bus.size,
974                                      page_shift, NV_MEM_ACCESS_RW,
975                                      &node->bar_vma);
976                 if (ret)
977                         return ret;
978
979                 nouveau_vm_map(&node->bar_vma, node);
980                 if (ret) {
981                         nouveau_vm_put(&node->bar_vma);
982                         return ret;
983                 }
984
985                 mem->bus.offset = node->bar_vma.offset;
986                 if (dev_priv->card_type == NV_50) /*XXX*/
987                         mem->bus.offset -= 0x0020000000ULL;
988                 mem->bus.base = pci_resource_start(dev->pdev, 1);
989                 mem->bus.is_iomem = true;
990         }
991                 break;
992         default:
993                 return -EINVAL;
994         }
995         return 0;
996 }
997
998 static void
999 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1000 {
1001         struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
1002         struct nouveau_mem *node = mem->mm_node;
1003
1004         if (!dev_priv->bar1_vm || mem->mem_type != TTM_PL_VRAM)
1005                 return;
1006
1007         if (!node->bar_vma.node)
1008                 return;
1009
1010         nouveau_vm_unmap(&node->bar_vma);
1011         nouveau_vm_put(&node->bar_vma);
1012 }
1013
1014 static int
1015 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1016 {
1017         struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
1018         struct nouveau_bo *nvbo = nouveau_bo(bo);
1019
1020         /* as long as the bo isn't in vram, and isn't tiled, we've got
1021          * nothing to do here.
1022          */
1023         if (bo->mem.mem_type != TTM_PL_VRAM) {
1024                 if (dev_priv->card_type < NV_50 ||
1025                     !nouveau_bo_tile_layout(nvbo))
1026                         return 0;
1027         }
1028
1029         /* make sure bo is in mappable vram */
1030         if (bo->mem.start + bo->mem.num_pages < dev_priv->fb_mappable_pages)
1031                 return 0;
1032
1033
1034         nvbo->placement.fpfn = 0;
1035         nvbo->placement.lpfn = dev_priv->fb_mappable_pages;
1036         nouveau_bo_placement_set(nvbo, TTM_PL_VRAM, 0);
1037         return nouveau_bo_validate(nvbo, false, true, false);
1038 }
1039
1040 void
1041 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
1042 {
1043         struct nouveau_fence *old_fence;
1044
1045         if (likely(fence))
1046                 nouveau_fence_ref(fence);
1047
1048         spin_lock(&nvbo->bo.bdev->fence_lock);
1049         old_fence = nvbo->bo.sync_obj;
1050         nvbo->bo.sync_obj = fence;
1051         spin_unlock(&nvbo->bo.bdev->fence_lock);
1052
1053         nouveau_fence_unref(&old_fence);
1054 }
1055
1056 static int
1057 nouveau_ttm_tt_populate(struct ttm_tt *ttm)
1058 {
1059         struct ttm_dma_tt *ttm_dma = (void *)ttm;
1060         struct drm_nouveau_private *dev_priv;
1061         struct drm_device *dev;
1062         unsigned i;
1063         int r;
1064
1065         if (ttm->state != tt_unpopulated)
1066                 return 0;
1067
1068         dev_priv = nouveau_bdev(ttm->bdev);
1069         dev = dev_priv->dev;
1070
1071 #ifdef CONFIG_SWIOTLB
1072         if (swiotlb_nr_tbl()) {
1073                 return ttm_dma_populate((void *)ttm, dev->dev);
1074         }
1075 #endif
1076
1077         r = ttm_pool_populate(ttm);
1078         if (r) {
1079                 return r;
1080         }
1081
1082         for (i = 0; i < ttm->num_pages; i++) {
1083                 ttm_dma->dma_address[i] = pci_map_page(dev->pdev, ttm->pages[i],
1084                                                    0, PAGE_SIZE,
1085                                                    PCI_DMA_BIDIRECTIONAL);
1086                 if (pci_dma_mapping_error(dev->pdev, ttm_dma->dma_address[i])) {
1087                         while (--i) {
1088                                 pci_unmap_page(dev->pdev, ttm_dma->dma_address[i],
1089                                                PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
1090                                 ttm_dma->dma_address[i] = 0;
1091                         }
1092                         ttm_pool_unpopulate(ttm);
1093                         return -EFAULT;
1094                 }
1095         }
1096         return 0;
1097 }
1098
1099 static void
1100 nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
1101 {
1102         struct ttm_dma_tt *ttm_dma = (void *)ttm;
1103         struct drm_nouveau_private *dev_priv;
1104         struct drm_device *dev;
1105         unsigned i;
1106
1107         dev_priv = nouveau_bdev(ttm->bdev);
1108         dev = dev_priv->dev;
1109
1110 #ifdef CONFIG_SWIOTLB
1111         if (swiotlb_nr_tbl()) {
1112                 ttm_dma_unpopulate((void *)ttm, dev->dev);
1113                 return;
1114         }
1115 #endif
1116
1117         for (i = 0; i < ttm->num_pages; i++) {
1118                 if (ttm_dma->dma_address[i]) {
1119                         pci_unmap_page(dev->pdev, ttm_dma->dma_address[i],
1120                                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
1121                 }
1122         }
1123
1124         ttm_pool_unpopulate(ttm);
1125 }
1126
1127 struct ttm_bo_driver nouveau_bo_driver = {
1128         .ttm_tt_create = &nouveau_ttm_tt_create,
1129         .ttm_tt_populate = &nouveau_ttm_tt_populate,
1130         .ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1131         .invalidate_caches = nouveau_bo_invalidate_caches,
1132         .init_mem_type = nouveau_bo_init_mem_type,
1133         .evict_flags = nouveau_bo_evict_flags,
1134         .move_notify = nouveau_bo_move_ntfy,
1135         .move = nouveau_bo_move,
1136         .verify_access = nouveau_bo_verify_access,
1137         .sync_obj_signaled = __nouveau_fence_signalled,
1138         .sync_obj_wait = __nouveau_fence_wait,
1139         .sync_obj_flush = __nouveau_fence_flush,
1140         .sync_obj_unref = __nouveau_fence_unref,
1141         .sync_obj_ref = __nouveau_fence_ref,
1142         .fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1143         .io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1144         .io_mem_free = &nouveau_ttm_io_mem_free,
1145 };
1146
1147 struct nouveau_vma *
1148 nouveau_bo_vma_find(struct nouveau_bo *nvbo, struct nouveau_vm *vm)
1149 {
1150         struct nouveau_vma *vma;
1151         list_for_each_entry(vma, &nvbo->vma_list, head) {
1152                 if (vma->vm == vm)
1153                         return vma;
1154         }
1155
1156         return NULL;
1157 }
1158
1159 int
1160 nouveau_bo_vma_add(struct nouveau_bo *nvbo, struct nouveau_vm *vm,
1161                    struct nouveau_vma *vma)
1162 {
1163         const u32 size = nvbo->bo.mem.num_pages << PAGE_SHIFT;
1164         struct nouveau_mem *node = nvbo->bo.mem.mm_node;
1165         int ret;
1166
1167         ret = nouveau_vm_get(vm, size, nvbo->page_shift,
1168                              NV_MEM_ACCESS_RW, vma);
1169         if (ret)
1170                 return ret;
1171
1172         if (nvbo->bo.mem.mem_type == TTM_PL_VRAM)
1173                 nouveau_vm_map(vma, nvbo->bo.mem.mm_node);
1174         else
1175         if (nvbo->bo.mem.mem_type == TTM_PL_TT)
1176                 nouveau_vm_map_sg(vma, 0, size, node, node->pages);
1177
1178         list_add_tail(&vma->head, &nvbo->vma_list);
1179         vma->refcount = 1;
1180         return 0;
1181 }
1182
1183 void
1184 nouveau_bo_vma_del(struct nouveau_bo *nvbo, struct nouveau_vma *vma)
1185 {
1186         if (vma->node) {
1187                 if (nvbo->bo.mem.mem_type != TTM_PL_SYSTEM) {
1188                         spin_lock(&nvbo->bo.bdev->fence_lock);
1189                         ttm_bo_wait(&nvbo->bo, false, false, false);
1190                         spin_unlock(&nvbo->bo.bdev->fence_lock);
1191                         nouveau_vm_unmap(vma);
1192                 }
1193
1194                 nouveau_vm_put(vma);
1195                 list_del(&vma->head);
1196         }
1197 }