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 <core/engine.h>
31 #include <linux/swiotlb.h>
32
33 #include <subdev/fb.h>
34 #include <subdev/vm.h>
35 #include <subdev/bar.h>
36
37 #include "nouveau_drm.h"
38 #include "nouveau_dma.h"
39 #include "nouveau_fence.h"
40
41 #include "nouveau_bo.h"
42 #include "nouveau_ttm.h"
43 #include "nouveau_gem.h"
44
45 /*
46 * NV10-NV40 tiling helpers
47 */
48
49 static void
nv10_bo_update_tile_region(struct drm_device * dev,struct nouveau_drm_tile * reg,u32 addr,u32 size,u32 pitch,u32 flags)50 nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
51 u32 addr, u32 size, u32 pitch, u32 flags)
52 {
53 struct nouveau_drm *drm = nouveau_drm(dev);
54 int i = reg - drm->tile.reg;
55 struct nouveau_fb *pfb = nouveau_fb(drm->device);
56 struct nouveau_fb_tile *tile = &pfb->tile.region[i];
57 struct nouveau_engine *engine;
58
59 nouveau_fence_unref(®->fence);
60
61 if (tile->pitch)
62 pfb->tile.fini(pfb, i, tile);
63
64 if (pitch)
65 pfb->tile.init(pfb, i, addr, size, pitch, flags, tile);
66
67 pfb->tile.prog(pfb, i, tile);
68
69 if ((engine = nouveau_engine(pfb, NVDEV_ENGINE_GR)))
70 engine->tile_prog(engine, i);
71 if ((engine = nouveau_engine(pfb, NVDEV_ENGINE_MPEG)))
72 engine->tile_prog(engine, i);
73 }
74
75 static struct nouveau_drm_tile *
nv10_bo_get_tile_region(struct drm_device * dev,int i)76 nv10_bo_get_tile_region(struct drm_device *dev, int i)
77 {
78 struct nouveau_drm *drm = nouveau_drm(dev);
79 struct nouveau_drm_tile *tile = &drm->tile.reg[i];
80
81 spin_lock(&drm->tile.lock);
82
83 if (!tile->used &&
84 (!tile->fence || nouveau_fence_done(tile->fence)))
85 tile->used = true;
86 else
87 tile = NULL;
88
89 spin_unlock(&drm->tile.lock);
90 return tile;
91 }
92
93 static void
nv10_bo_put_tile_region(struct drm_device * dev,struct nouveau_drm_tile * tile,struct nouveau_fence * fence)94 nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
95 struct nouveau_fence *fence)
96 {
97 struct nouveau_drm *drm = nouveau_drm(dev);
98
99 if (tile) {
100 spin_lock(&drm->tile.lock);
101 if (fence) {
102 /* Mark it as pending. */
103 tile->fence = fence;
104 nouveau_fence_ref(fence);
105 }
106
107 tile->used = false;
108 spin_unlock(&drm->tile.lock);
109 }
110 }
111
112 static struct nouveau_drm_tile *
nv10_bo_set_tiling(struct drm_device * dev,u32 addr,u32 size,u32 pitch,u32 flags)113 nv10_bo_set_tiling(struct drm_device *dev, u32 addr,
114 u32 size, u32 pitch, u32 flags)
115 {
116 struct nouveau_drm *drm = nouveau_drm(dev);
117 struct nouveau_fb *pfb = nouveau_fb(drm->device);
118 struct nouveau_drm_tile *tile, *found = NULL;
119 int i;
120
121 for (i = 0; i < pfb->tile.regions; i++) {
122 tile = nv10_bo_get_tile_region(dev, i);
123
124 if (pitch && !found) {
125 found = tile;
126 continue;
127
128 } else if (tile && pfb->tile.region[i].pitch) {
129 /* Kill an unused tile region. */
130 nv10_bo_update_tile_region(dev, tile, 0, 0, 0, 0);
131 }
132
133 nv10_bo_put_tile_region(dev, tile, NULL);
134 }
135
136 if (found)
137 nv10_bo_update_tile_region(dev, found, addr, size,
138 pitch, flags);
139 return found;
140 }
141
142 static void
nouveau_bo_del_ttm(struct ttm_buffer_object * bo)143 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
144 {
145 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
146 struct drm_device *dev = drm->dev;
147 struct nouveau_bo *nvbo = nouveau_bo(bo);
148
149 if (unlikely(nvbo->gem))
150 DRM_ERROR("bo %p still attached to GEM object\n", bo);
151 nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
152 kfree(nvbo);
153 }
154
155 static void
nouveau_bo_fixup_align(struct nouveau_bo * nvbo,u32 flags,int * align,int * size)156 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
157 int *align, int *size)
158 {
159 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
160 struct nouveau_device *device = nv_device(drm->device);
161
162 if (device->card_type < NV_50) {
163 if (nvbo->tile_mode) {
164 if (device->chipset >= 0x40) {
165 *align = 65536;
166 *size = roundup(*size, 64 * nvbo->tile_mode);
167
168 } else if (device->chipset >= 0x30) {
169 *align = 32768;
170 *size = roundup(*size, 64 * nvbo->tile_mode);
171
172 } else if (device->chipset >= 0x20) {
173 *align = 16384;
174 *size = roundup(*size, 64 * nvbo->tile_mode);
175
176 } else if (device->chipset >= 0x10) {
177 *align = 16384;
178 *size = roundup(*size, 32 * nvbo->tile_mode);
179 }
180 }
181 } else {
182 *size = roundup(*size, (1 << nvbo->page_shift));
183 *align = max((1 << nvbo->page_shift), *align);
184 }
185
186 *size = roundup(*size, PAGE_SIZE);
187 }
188
189 int
nouveau_bo_new(struct drm_device * dev,int size,int align,uint32_t flags,uint32_t tile_mode,uint32_t tile_flags,struct sg_table * sg,struct nouveau_bo ** pnvbo)190 nouveau_bo_new(struct drm_device *dev, int size, int align,
191 uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
192 struct sg_table *sg,
193 struct nouveau_bo **pnvbo)
194 {
195 struct nouveau_drm *drm = nouveau_drm(dev);
196 struct nouveau_bo *nvbo;
197 size_t acc_size;
198 int ret;
199 int type = ttm_bo_type_device;
200
201 if (sg)
202 type = ttm_bo_type_sg;
203
204 nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
205 if (!nvbo)
206 return -ENOMEM;
207 INIT_LIST_HEAD(&nvbo->head);
208 INIT_LIST_HEAD(&nvbo->entry);
209 INIT_LIST_HEAD(&nvbo->vma_list);
210 nvbo->tile_mode = tile_mode;
211 nvbo->tile_flags = tile_flags;
212 nvbo->bo.bdev = &drm->ttm.bdev;
213
214 nvbo->page_shift = 12;
215 if (drm->client.base.vm) {
216 if (!(flags & TTM_PL_FLAG_TT) && size > 256 * 1024)
217 nvbo->page_shift = drm->client.base.vm->vmm->lpg_shift;
218 }
219
220 nouveau_bo_fixup_align(nvbo, flags, &align, &size);
221 nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
222 nouveau_bo_placement_set(nvbo, flags, 0);
223
224 acc_size = ttm_bo_dma_acc_size(&drm->ttm.bdev, size,
225 sizeof(struct nouveau_bo));
226
227 ret = ttm_bo_init(&drm->ttm.bdev, &nvbo->bo, size,
228 type, &nvbo->placement,
229 align >> PAGE_SHIFT, false, NULL, acc_size, sg,
230 nouveau_bo_del_ttm);
231 if (ret) {
232 /* ttm will call nouveau_bo_del_ttm if it fails.. */
233 return ret;
234 }
235
236 *pnvbo = nvbo;
237 return 0;
238 }
239
240 static void
set_placement_list(uint32_t * pl,unsigned * n,uint32_t type,uint32_t flags)241 set_placement_list(uint32_t *pl, unsigned *n, uint32_t type, uint32_t flags)
242 {
243 *n = 0;
244
245 if (type & TTM_PL_FLAG_VRAM)
246 pl[(*n)++] = TTM_PL_FLAG_VRAM | flags;
247 if (type & TTM_PL_FLAG_TT)
248 pl[(*n)++] = TTM_PL_FLAG_TT | flags;
249 if (type & TTM_PL_FLAG_SYSTEM)
250 pl[(*n)++] = TTM_PL_FLAG_SYSTEM | flags;
251 }
252
253 static void
set_placement_range(struct nouveau_bo * nvbo,uint32_t type)254 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
255 {
256 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
257 struct nouveau_fb *pfb = nouveau_fb(drm->device);
258 u32 vram_pages = pfb->ram.size >> PAGE_SHIFT;
259
260 if (nv_device(drm->device)->card_type == NV_10 &&
261 nvbo->tile_mode && (type & TTM_PL_FLAG_VRAM) &&
262 nvbo->bo.mem.num_pages < vram_pages / 4) {
263 /*
264 * Make sure that the color and depth buffers are handled
265 * by independent memory controller units. Up to a 9x
266 * speed up when alpha-blending and depth-test are enabled
267 * at the same time.
268 */
269 if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
270 nvbo->placement.fpfn = vram_pages / 2;
271 nvbo->placement.lpfn = ~0;
272 } else {
273 nvbo->placement.fpfn = 0;
274 nvbo->placement.lpfn = vram_pages / 2;
275 }
276 }
277 }
278
279 void
nouveau_bo_placement_set(struct nouveau_bo * nvbo,uint32_t type,uint32_t busy)280 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
281 {
282 struct ttm_placement *pl = &nvbo->placement;
283 uint32_t flags = TTM_PL_MASK_CACHING |
284 (nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
285
286 pl->placement = nvbo->placements;
287 set_placement_list(nvbo->placements, &pl->num_placement,
288 type, flags);
289
290 pl->busy_placement = nvbo->busy_placements;
291 set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
292 type | busy, flags);
293
294 set_placement_range(nvbo, type);
295 }
296
297 int
nouveau_bo_pin(struct nouveau_bo * nvbo,uint32_t memtype)298 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype)
299 {
300 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
301 struct ttm_buffer_object *bo = &nvbo->bo;
302 int ret;
303
304 ret = ttm_bo_reserve(bo, false, false, false, 0);
305 if (ret)
306 goto out;
307
308 if (nvbo->pin_refcnt && !(memtype & (1 << bo->mem.mem_type))) {
309 NV_ERROR(drm, "bo %p pinned elsewhere: 0x%08x vs 0x%08x\n", bo,
310 1 << bo->mem.mem_type, memtype);
311 ret = -EINVAL;
312 goto out;
313 }
314
315 if (nvbo->pin_refcnt++)
316 goto out;
317
318 nouveau_bo_placement_set(nvbo, memtype, 0);
319
320 ret = nouveau_bo_validate(nvbo, false, false);
321 if (ret == 0) {
322 switch (bo->mem.mem_type) {
323 case TTM_PL_VRAM:
324 drm->gem.vram_available -= bo->mem.size;
325 break;
326 case TTM_PL_TT:
327 drm->gem.gart_available -= bo->mem.size;
328 break;
329 default:
330 break;
331 }
332 }
333 out:
334 ttm_bo_unreserve(bo);
335 return ret;
336 }
337
338 int
nouveau_bo_unpin(struct nouveau_bo * nvbo)339 nouveau_bo_unpin(struct nouveau_bo *nvbo)
340 {
341 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
342 struct ttm_buffer_object *bo = &nvbo->bo;
343 int ret;
344
345 ret = ttm_bo_reserve(bo, false, false, false, 0);
346 if (ret)
347 return ret;
348
349 if (--nvbo->pin_refcnt)
350 goto out;
351
352 nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
353
354 ret = nouveau_bo_validate(nvbo, false, false);
355 if (ret == 0) {
356 switch (bo->mem.mem_type) {
357 case TTM_PL_VRAM:
358 drm->gem.vram_available += bo->mem.size;
359 break;
360 case TTM_PL_TT:
361 drm->gem.gart_available += bo->mem.size;
362 break;
363 default:
364 break;
365 }
366 }
367
368 out:
369 ttm_bo_unreserve(bo);
370 return ret;
371 }
372
373 int
nouveau_bo_map(struct nouveau_bo * nvbo)374 nouveau_bo_map(struct nouveau_bo *nvbo)
375 {
376 int ret;
377
378 ret = ttm_bo_reserve(&nvbo->bo, false, false, false, 0);
379 if (ret)
380 return ret;
381
382 ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
383 ttm_bo_unreserve(&nvbo->bo);
384 return ret;
385 }
386
387 void
nouveau_bo_unmap(struct nouveau_bo * nvbo)388 nouveau_bo_unmap(struct nouveau_bo *nvbo)
389 {
390 if (nvbo)
391 ttm_bo_kunmap(&nvbo->kmap);
392 }
393
394 int
nouveau_bo_validate(struct nouveau_bo * nvbo,bool interruptible,bool no_wait_gpu)395 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
396 bool no_wait_gpu)
397 {
398 int ret;
399
400 ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement,
401 interruptible, no_wait_gpu);
402 if (ret)
403 return ret;
404
405 return 0;
406 }
407
408 u16
nouveau_bo_rd16(struct nouveau_bo * nvbo,unsigned index)409 nouveau_bo_rd16(struct nouveau_bo *nvbo, unsigned index)
410 {
411 bool is_iomem;
412 u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
413 mem = &mem[index];
414 if (is_iomem)
415 return ioread16_native((void __force __iomem *)mem);
416 else
417 return *mem;
418 }
419
420 void
nouveau_bo_wr16(struct nouveau_bo * nvbo,unsigned index,u16 val)421 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
422 {
423 bool is_iomem;
424 u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
425 mem = &mem[index];
426 if (is_iomem)
427 iowrite16_native(val, (void __force __iomem *)mem);
428 else
429 *mem = val;
430 }
431
432 u32
nouveau_bo_rd32(struct nouveau_bo * nvbo,unsigned index)433 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
434 {
435 bool is_iomem;
436 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
437 mem = &mem[index];
438 if (is_iomem)
439 return ioread32_native((void __force __iomem *)mem);
440 else
441 return *mem;
442 }
443
444 void
nouveau_bo_wr32(struct nouveau_bo * nvbo,unsigned index,u32 val)445 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
446 {
447 bool is_iomem;
448 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
449 mem = &mem[index];
450 if (is_iomem)
451 iowrite32_native(val, (void __force __iomem *)mem);
452 else
453 *mem = val;
454 }
455
456 static struct ttm_tt *
nouveau_ttm_tt_create(struct ttm_bo_device * bdev,unsigned long size,uint32_t page_flags,struct page * dummy_read)457 nouveau_ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
458 uint32_t page_flags, struct page *dummy_read)
459 {
460 #if __OS_HAS_AGP
461 struct nouveau_drm *drm = nouveau_bdev(bdev);
462 struct drm_device *dev = drm->dev;
463
464 if (drm->agp.stat == ENABLED) {
465 return ttm_agp_tt_create(bdev, dev->agp->bridge, size,
466 page_flags, dummy_read);
467 }
468 #endif
469
470 return nouveau_sgdma_create_ttm(bdev, size, page_flags, dummy_read);
471 }
472
473 static int
nouveau_bo_invalidate_caches(struct ttm_bo_device * bdev,uint32_t flags)474 nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
475 {
476 /* We'll do this from user space. */
477 return 0;
478 }
479
480 static int
nouveau_bo_init_mem_type(struct ttm_bo_device * bdev,uint32_t type,struct ttm_mem_type_manager * man)481 nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
482 struct ttm_mem_type_manager *man)
483 {
484 struct nouveau_drm *drm = nouveau_bdev(bdev);
485
486 switch (type) {
487 case TTM_PL_SYSTEM:
488 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
489 man->available_caching = TTM_PL_MASK_CACHING;
490 man->default_caching = TTM_PL_FLAG_CACHED;
491 break;
492 case TTM_PL_VRAM:
493 if (nv_device(drm->device)->card_type >= NV_50) {
494 man->func = &nouveau_vram_manager;
495 man->io_reserve_fastpath = false;
496 man->use_io_reserve_lru = true;
497 } else {
498 man->func = &ttm_bo_manager_func;
499 }
500 man->flags = TTM_MEMTYPE_FLAG_FIXED |
501 TTM_MEMTYPE_FLAG_MAPPABLE;
502 man->available_caching = TTM_PL_FLAG_UNCACHED |
503 TTM_PL_FLAG_WC;
504 man->default_caching = TTM_PL_FLAG_WC;
505 break;
506 case TTM_PL_TT:
507 if (nv_device(drm->device)->card_type >= NV_50)
508 man->func = &nouveau_gart_manager;
509 else
510 if (drm->agp.stat != ENABLED)
511 man->func = &nv04_gart_manager;
512 else
513 man->func = &ttm_bo_manager_func;
514
515 if (drm->agp.stat == ENABLED) {
516 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
517 man->available_caching = TTM_PL_FLAG_UNCACHED |
518 TTM_PL_FLAG_WC;
519 man->default_caching = TTM_PL_FLAG_WC;
520 } else {
521 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
522 TTM_MEMTYPE_FLAG_CMA;
523 man->available_caching = TTM_PL_MASK_CACHING;
524 man->default_caching = TTM_PL_FLAG_CACHED;
525 }
526
527 break;
528 default:
529 return -EINVAL;
530 }
531 return 0;
532 }
533
534 static void
nouveau_bo_evict_flags(struct ttm_buffer_object * bo,struct ttm_placement * pl)535 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
536 {
537 struct nouveau_bo *nvbo = nouveau_bo(bo);
538
539 switch (bo->mem.mem_type) {
540 case TTM_PL_VRAM:
541 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
542 TTM_PL_FLAG_SYSTEM);
543 break;
544 default:
545 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
546 break;
547 }
548
549 *pl = nvbo->placement;
550 }
551
552
553 /* GPU-assisted copy using NV_MEMORY_TO_MEMORY_FORMAT, can access
554 * TTM_PL_{VRAM,TT} directly.
555 */
556
557 static int
nouveau_bo_move_accel_cleanup(struct nouveau_channel * chan,struct nouveau_bo * nvbo,bool evict,bool no_wait_gpu,struct ttm_mem_reg * new_mem)558 nouveau_bo_move_accel_cleanup(struct nouveau_channel *chan,
559 struct nouveau_bo *nvbo, bool evict,
560 bool no_wait_gpu, struct ttm_mem_reg *new_mem)
561 {
562 struct nouveau_fence *fence = NULL;
563 int ret;
564
565 ret = nouveau_fence_new(chan, false, &fence);
566 if (ret)
567 return ret;
568
569 ret = ttm_bo_move_accel_cleanup(&nvbo->bo, fence, evict,
570 no_wait_gpu, new_mem);
571 nouveau_fence_unref(&fence);
572 return ret;
573 }
574
575 static int
nve0_bo_move_init(struct nouveau_channel * chan,u32 handle)576 nve0_bo_move_init(struct nouveau_channel *chan, u32 handle)
577 {
578 int ret = RING_SPACE(chan, 2);
579 if (ret == 0) {
580 BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
581 OUT_RING (chan, handle);
582 FIRE_RING (chan);
583 }
584 return ret;
585 }
586
587 static int
nve0_bo_move_copy(struct nouveau_channel * chan,struct ttm_buffer_object * bo,struct ttm_mem_reg * old_mem,struct ttm_mem_reg * new_mem)588 nve0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
589 struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
590 {
591 struct nouveau_mem *node = old_mem->mm_node;
592 int ret = RING_SPACE(chan, 10);
593 if (ret == 0) {
594 BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
595 OUT_RING (chan, upper_32_bits(node->vma[0].offset));
596 OUT_RING (chan, lower_32_bits(node->vma[0].offset));
597 OUT_RING (chan, upper_32_bits(node->vma[1].offset));
598 OUT_RING (chan, lower_32_bits(node->vma[1].offset));
599 OUT_RING (chan, PAGE_SIZE);
600 OUT_RING (chan, PAGE_SIZE);
601 OUT_RING (chan, PAGE_SIZE);
602 OUT_RING (chan, new_mem->num_pages);
603 BEGIN_IMC0(chan, NvSubCopy, 0x0300, 0x0386);
604 }
605 return ret;
606 }
607
608 static int
nvc0_bo_move_init(struct nouveau_channel * chan,u32 handle)609 nvc0_bo_move_init(struct nouveau_channel *chan, u32 handle)
610 {
611 int ret = RING_SPACE(chan, 2);
612 if (ret == 0) {
613 BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
614 OUT_RING (chan, handle);
615 }
616 return ret;
617 }
618
619 static int
nvc0_bo_move_copy(struct nouveau_channel * chan,struct ttm_buffer_object * bo,struct ttm_mem_reg * old_mem,struct ttm_mem_reg * new_mem)620 nvc0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
621 struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
622 {
623 struct nouveau_mem *node = old_mem->mm_node;
624 u64 src_offset = node->vma[0].offset;
625 u64 dst_offset = node->vma[1].offset;
626 u32 page_count = new_mem->num_pages;
627 int ret;
628
629 page_count = new_mem->num_pages;
630 while (page_count) {
631 int line_count = (page_count > 8191) ? 8191 : page_count;
632
633 ret = RING_SPACE(chan, 11);
634 if (ret)
635 return ret;
636
637 BEGIN_NVC0(chan, NvSubCopy, 0x030c, 8);
638 OUT_RING (chan, upper_32_bits(src_offset));
639 OUT_RING (chan, lower_32_bits(src_offset));
640 OUT_RING (chan, upper_32_bits(dst_offset));
641 OUT_RING (chan, lower_32_bits(dst_offset));
642 OUT_RING (chan, PAGE_SIZE);
643 OUT_RING (chan, PAGE_SIZE);
644 OUT_RING (chan, PAGE_SIZE);
645 OUT_RING (chan, line_count);
646 BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
647 OUT_RING (chan, 0x00000110);
648
649 page_count -= line_count;
650 src_offset += (PAGE_SIZE * line_count);
651 dst_offset += (PAGE_SIZE * line_count);
652 }
653
654 return 0;
655 }
656
657 static int
nvc0_bo_move_m2mf(struct nouveau_channel * chan,struct ttm_buffer_object * bo,struct ttm_mem_reg * old_mem,struct ttm_mem_reg * new_mem)658 nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
659 struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
660 {
661 struct nouveau_mem *node = old_mem->mm_node;
662 u64 src_offset = node->vma[0].offset;
663 u64 dst_offset = node->vma[1].offset;
664 u32 page_count = new_mem->num_pages;
665 int ret;
666
667 page_count = new_mem->num_pages;
668 while (page_count) {
669 int line_count = (page_count > 2047) ? 2047 : page_count;
670
671 ret = RING_SPACE(chan, 12);
672 if (ret)
673 return ret;
674
675 BEGIN_NVC0(chan, NvSubCopy, 0x0238, 2);
676 OUT_RING (chan, upper_32_bits(dst_offset));
677 OUT_RING (chan, lower_32_bits(dst_offset));
678 BEGIN_NVC0(chan, NvSubCopy, 0x030c, 6);
679 OUT_RING (chan, upper_32_bits(src_offset));
680 OUT_RING (chan, lower_32_bits(src_offset));
681 OUT_RING (chan, PAGE_SIZE); /* src_pitch */
682 OUT_RING (chan, PAGE_SIZE); /* dst_pitch */
683 OUT_RING (chan, PAGE_SIZE); /* line_length */
684 OUT_RING (chan, line_count);
685 BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
686 OUT_RING (chan, 0x00100110);
687
688 page_count -= line_count;
689 src_offset += (PAGE_SIZE * line_count);
690 dst_offset += (PAGE_SIZE * line_count);
691 }
692
693 return 0;
694 }
695
696 static int
nva3_bo_move_copy(struct nouveau_channel * chan,struct ttm_buffer_object * bo,struct ttm_mem_reg * old_mem,struct ttm_mem_reg * new_mem)697 nva3_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
698 struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
699 {
700 struct nouveau_mem *node = old_mem->mm_node;
701 u64 src_offset = node->vma[0].offset;
702 u64 dst_offset = node->vma[1].offset;
703 u32 page_count = new_mem->num_pages;
704 int ret;
705
706 page_count = new_mem->num_pages;
707 while (page_count) {
708 int line_count = (page_count > 8191) ? 8191 : page_count;
709
710 ret = RING_SPACE(chan, 11);
711 if (ret)
712 return ret;
713
714 BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
715 OUT_RING (chan, upper_32_bits(src_offset));
716 OUT_RING (chan, lower_32_bits(src_offset));
717 OUT_RING (chan, upper_32_bits(dst_offset));
718 OUT_RING (chan, lower_32_bits(dst_offset));
719 OUT_RING (chan, PAGE_SIZE);
720 OUT_RING (chan, PAGE_SIZE);
721 OUT_RING (chan, PAGE_SIZE);
722 OUT_RING (chan, line_count);
723 BEGIN_NV04(chan, NvSubCopy, 0x0300, 1);
724 OUT_RING (chan, 0x00000110);
725
726 page_count -= line_count;
727 src_offset += (PAGE_SIZE * line_count);
728 dst_offset += (PAGE_SIZE * line_count);
729 }
730
731 return 0;
732 }
733
734 static int
nv98_bo_move_exec(struct nouveau_channel * chan,struct ttm_buffer_object * bo,struct ttm_mem_reg * old_mem,struct ttm_mem_reg * new_mem)735 nv98_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
736 struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
737 {
738 struct nouveau_mem *node = old_mem->mm_node;
739 int ret = RING_SPACE(chan, 7);
740 if (ret == 0) {
741 BEGIN_NV04(chan, NvSubCopy, 0x0320, 6);
742 OUT_RING (chan, upper_32_bits(node->vma[0].offset));
743 OUT_RING (chan, lower_32_bits(node->vma[0].offset));
744 OUT_RING (chan, upper_32_bits(node->vma[1].offset));
745 OUT_RING (chan, lower_32_bits(node->vma[1].offset));
746 OUT_RING (chan, 0x00000000 /* COPY */);
747 OUT_RING (chan, new_mem->num_pages << PAGE_SHIFT);
748 }
749 return ret;
750 }
751
752 static int
nv84_bo_move_exec(struct nouveau_channel * chan,struct ttm_buffer_object * bo,struct ttm_mem_reg * old_mem,struct ttm_mem_reg * new_mem)753 nv84_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
754 struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
755 {
756 struct nouveau_mem *node = old_mem->mm_node;
757 int ret = RING_SPACE(chan, 7);
758 if (ret == 0) {
759 BEGIN_NV04(chan, NvSubCopy, 0x0304, 6);
760 OUT_RING (chan, new_mem->num_pages << PAGE_SHIFT);
761 OUT_RING (chan, upper_32_bits(node->vma[0].offset));
762 OUT_RING (chan, lower_32_bits(node->vma[0].offset));
763 OUT_RING (chan, upper_32_bits(node->vma[1].offset));
764 OUT_RING (chan, lower_32_bits(node->vma[1].offset));
765 OUT_RING (chan, 0x00000000 /* MODE_COPY, QUERY_NONE */);
766 }
767 return ret;
768 }
769
770 static int
nv50_bo_move_init(struct nouveau_channel * chan,u32 handle)771 nv50_bo_move_init(struct nouveau_channel *chan, u32 handle)
772 {
773 int ret = RING_SPACE(chan, 6);
774 if (ret == 0) {
775 BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
776 OUT_RING (chan, handle);
777 BEGIN_NV04(chan, NvSubCopy, 0x0180, 3);
778 OUT_RING (chan, NvNotify0);
779 OUT_RING (chan, NvDmaFB);
780 OUT_RING (chan, NvDmaFB);
781 }
782
783 return ret;
784 }
785
786 static int
nv50_bo_move_m2mf(struct nouveau_channel * chan,struct ttm_buffer_object * bo,struct ttm_mem_reg * old_mem,struct ttm_mem_reg * new_mem)787 nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
788 struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
789 {
790 struct nouveau_mem *node = old_mem->mm_node;
791 struct nouveau_bo *nvbo = nouveau_bo(bo);
792 u64 length = (new_mem->num_pages << PAGE_SHIFT);
793 u64 src_offset = node->vma[0].offset;
794 u64 dst_offset = node->vma[1].offset;
795 int ret;
796
797 while (length) {
798 u32 amount, stride, height;
799
800 amount = min(length, (u64)(4 * 1024 * 1024));
801 stride = 16 * 4;
802 height = amount / stride;
803
804 if (old_mem->mem_type == TTM_PL_VRAM &&
805 nouveau_bo_tile_layout(nvbo)) {
806 ret = RING_SPACE(chan, 8);
807 if (ret)
808 return ret;
809
810 BEGIN_NV04(chan, NvSubCopy, 0x0200, 7);
811 OUT_RING (chan, 0);
812 OUT_RING (chan, 0);
813 OUT_RING (chan, stride);
814 OUT_RING (chan, height);
815 OUT_RING (chan, 1);
816 OUT_RING (chan, 0);
817 OUT_RING (chan, 0);
818 } else {
819 ret = RING_SPACE(chan, 2);
820 if (ret)
821 return ret;
822
823 BEGIN_NV04(chan, NvSubCopy, 0x0200, 1);
824 OUT_RING (chan, 1);
825 }
826 if (new_mem->mem_type == TTM_PL_VRAM &&
827 nouveau_bo_tile_layout(nvbo)) {
828 ret = RING_SPACE(chan, 8);
829 if (ret)
830 return ret;
831
832 BEGIN_NV04(chan, NvSubCopy, 0x021c, 7);
833 OUT_RING (chan, 0);
834 OUT_RING (chan, 0);
835 OUT_RING (chan, stride);
836 OUT_RING (chan, height);
837 OUT_RING (chan, 1);
838 OUT_RING (chan, 0);
839 OUT_RING (chan, 0);
840 } else {
841 ret = RING_SPACE(chan, 2);
842 if (ret)
843 return ret;
844
845 BEGIN_NV04(chan, NvSubCopy, 0x021c, 1);
846 OUT_RING (chan, 1);
847 }
848
849 ret = RING_SPACE(chan, 14);
850 if (ret)
851 return ret;
852
853 BEGIN_NV04(chan, NvSubCopy, 0x0238, 2);
854 OUT_RING (chan, upper_32_bits(src_offset));
855 OUT_RING (chan, upper_32_bits(dst_offset));
856 BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
857 OUT_RING (chan, lower_32_bits(src_offset));
858 OUT_RING (chan, lower_32_bits(dst_offset));
859 OUT_RING (chan, stride);
860 OUT_RING (chan, stride);
861 OUT_RING (chan, stride);
862 OUT_RING (chan, height);
863 OUT_RING (chan, 0x00000101);
864 OUT_RING (chan, 0x00000000);
865 BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
866 OUT_RING (chan, 0);
867
868 length -= amount;
869 src_offset += amount;
870 dst_offset += amount;
871 }
872
873 return 0;
874 }
875
876 static int
nv04_bo_move_init(struct nouveau_channel * chan,u32 handle)877 nv04_bo_move_init(struct nouveau_channel *chan, u32 handle)
878 {
879 int ret = RING_SPACE(chan, 4);
880 if (ret == 0) {
881 BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
882 OUT_RING (chan, handle);
883 BEGIN_NV04(chan, NvSubCopy, 0x0180, 1);
884 OUT_RING (chan, NvNotify0);
885 }
886
887 return ret;
888 }
889
890 static inline uint32_t
nouveau_bo_mem_ctxdma(struct ttm_buffer_object * bo,struct nouveau_channel * chan,struct ttm_mem_reg * mem)891 nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
892 struct nouveau_channel *chan, struct ttm_mem_reg *mem)
893 {
894 if (mem->mem_type == TTM_PL_TT)
895 return NvDmaTT;
896 return NvDmaFB;
897 }
898
899 static int
nv04_bo_move_m2mf(struct nouveau_channel * chan,struct ttm_buffer_object * bo,struct ttm_mem_reg * old_mem,struct ttm_mem_reg * new_mem)900 nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
901 struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
902 {
903 u32 src_offset = old_mem->start << PAGE_SHIFT;
904 u32 dst_offset = new_mem->start << PAGE_SHIFT;
905 u32 page_count = new_mem->num_pages;
906 int ret;
907
908 ret = RING_SPACE(chan, 3);
909 if (ret)
910 return ret;
911
912 BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
913 OUT_RING (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
914 OUT_RING (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
915
916 page_count = new_mem->num_pages;
917 while (page_count) {
918 int line_count = (page_count > 2047) ? 2047 : page_count;
919
920 ret = RING_SPACE(chan, 11);
921 if (ret)
922 return ret;
923
924 BEGIN_NV04(chan, NvSubCopy,
925 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
926 OUT_RING (chan, src_offset);
927 OUT_RING (chan, dst_offset);
928 OUT_RING (chan, PAGE_SIZE); /* src_pitch */
929 OUT_RING (chan, PAGE_SIZE); /* dst_pitch */
930 OUT_RING (chan, PAGE_SIZE); /* line_length */
931 OUT_RING (chan, line_count);
932 OUT_RING (chan, 0x00000101);
933 OUT_RING (chan, 0x00000000);
934 BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
935 OUT_RING (chan, 0);
936
937 page_count -= line_count;
938 src_offset += (PAGE_SIZE * line_count);
939 dst_offset += (PAGE_SIZE * line_count);
940 }
941
942 return 0;
943 }
944
945 static int
nouveau_vma_getmap(struct nouveau_channel * chan,struct nouveau_bo * nvbo,struct ttm_mem_reg * mem,struct nouveau_vma * vma)946 nouveau_vma_getmap(struct nouveau_channel *chan, struct nouveau_bo *nvbo,
947 struct ttm_mem_reg *mem, struct nouveau_vma *vma)
948 {
949 struct nouveau_mem *node = mem->mm_node;
950 int ret;
951
952 ret = nouveau_vm_get(nv_client(chan->cli)->vm, mem->num_pages <<
953 PAGE_SHIFT, node->page_shift,
954 NV_MEM_ACCESS_RW, vma);
955 if (ret)
956 return ret;
957
958 if (mem->mem_type == TTM_PL_VRAM)
959 nouveau_vm_map(vma, node);
960 else
961 nouveau_vm_map_sg(vma, 0, mem->num_pages << PAGE_SHIFT, node);
962
963 return 0;
964 }
965
966 static int
nouveau_bo_move_m2mf(struct ttm_buffer_object * bo,int evict,bool intr,bool no_wait_gpu,struct ttm_mem_reg * new_mem)967 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
968 bool no_wait_gpu, struct ttm_mem_reg *new_mem)
969 {
970 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
971 struct nouveau_channel *chan = chan = drm->channel;
972 struct nouveau_bo *nvbo = nouveau_bo(bo);
973 struct ttm_mem_reg *old_mem = &bo->mem;
974 int ret;
975
976 mutex_lock(&chan->cli->mutex);
977
978 /* create temporary vmas for the transfer and attach them to the
979 * old nouveau_mem node, these will get cleaned up after ttm has
980 * destroyed the ttm_mem_reg
981 */
982 if (nv_device(drm->device)->card_type >= NV_50) {
983 struct nouveau_mem *node = old_mem->mm_node;
984
985 ret = nouveau_vma_getmap(chan, nvbo, old_mem, &node->vma[0]);
986 if (ret)
987 goto out;
988
989 ret = nouveau_vma_getmap(chan, nvbo, new_mem, &node->vma[1]);
990 if (ret)
991 goto out;
992 }
993
994 ret = drm->ttm.move(chan, bo, &bo->mem, new_mem);
995 if (ret == 0) {
996 ret = nouveau_bo_move_accel_cleanup(chan, nvbo, evict,
997 no_wait_gpu, new_mem);
998 }
999
1000 out:
1001 mutex_unlock(&chan->cli->mutex);
1002 return ret;
1003 }
1004
1005 void
nouveau_bo_move_init(struct nouveau_drm * drm)1006 nouveau_bo_move_init(struct nouveau_drm *drm)
1007 {
1008 static const struct {
1009 const char *name;
1010 int engine;
1011 u32 oclass;
1012 int (*exec)(struct nouveau_channel *,
1013 struct ttm_buffer_object *,
1014 struct ttm_mem_reg *, struct ttm_mem_reg *);
1015 int (*init)(struct nouveau_channel *, u32 handle);
1016 } _methods[] = {
1017 { "COPY", 0, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
1018 { "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1019 { "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
1020 { "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
1021 { "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
1022 { "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
1023 { "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
1024 { "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
1025 { "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
1026 {},
1027 { "CRYPT", 0, 0x88b4, nv98_bo_move_exec, nv50_bo_move_init },
1028 }, *mthd = _methods;
1029 const char *name = "CPU";
1030 int ret;
1031
1032 do {
1033 struct nouveau_object *object;
1034 struct nouveau_channel *chan;
1035 u32 handle = (mthd->engine << 16) | mthd->oclass;
1036
1037 if (mthd->init == nve0_bo_move_init)
1038 chan = drm->cechan;
1039 else
1040 chan = drm->channel;
1041 if (chan == NULL)
1042 continue;
1043
1044 ret = nouveau_object_new(nv_object(drm), chan->handle, handle,
1045 mthd->oclass, NULL, 0, &object);
1046 if (ret == 0) {
1047 ret = mthd->init(chan, handle);
1048 if (ret) {
1049 nouveau_object_del(nv_object(drm),
1050 chan->handle, handle);
1051 continue;
1052 }
1053
1054 drm->ttm.move = mthd->exec;
1055 name = mthd->name;
1056 break;
1057 }
1058 } while ((++mthd)->exec);
1059
1060 NV_INFO(drm, "MM: using %s for buffer copies\n", name);
1061 }
1062
1063 static int
nouveau_bo_move_flipd(struct ttm_buffer_object * bo,bool evict,bool intr,bool no_wait_gpu,struct ttm_mem_reg * new_mem)1064 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
1065 bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1066 {
1067 u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
1068 struct ttm_placement placement;
1069 struct ttm_mem_reg tmp_mem;
1070 int ret;
1071
1072 placement.fpfn = placement.lpfn = 0;
1073 placement.num_placement = placement.num_busy_placement = 1;
1074 placement.placement = placement.busy_placement = &placement_memtype;
1075
1076 tmp_mem = *new_mem;
1077 tmp_mem.mm_node = NULL;
1078 ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1079 if (ret)
1080 return ret;
1081
1082 ret = ttm_tt_bind(bo->ttm, &tmp_mem);
1083 if (ret)
1084 goto out;
1085
1086 ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, &tmp_mem);
1087 if (ret)
1088 goto out;
1089
1090 ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
1091 out:
1092 ttm_bo_mem_put(bo, &tmp_mem);
1093 return ret;
1094 }
1095
1096 static int
nouveau_bo_move_flips(struct ttm_buffer_object * bo,bool evict,bool intr,bool no_wait_gpu,struct ttm_mem_reg * new_mem)1097 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
1098 bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1099 {
1100 u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
1101 struct ttm_placement placement;
1102 struct ttm_mem_reg tmp_mem;
1103 int ret;
1104
1105 placement.fpfn = placement.lpfn = 0;
1106 placement.num_placement = placement.num_busy_placement = 1;
1107 placement.placement = placement.busy_placement = &placement_memtype;
1108
1109 tmp_mem = *new_mem;
1110 tmp_mem.mm_node = NULL;
1111 ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1112 if (ret)
1113 return ret;
1114
1115 ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
1116 if (ret)
1117 goto out;
1118
1119 ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_mem);
1120 if (ret)
1121 goto out;
1122
1123 out:
1124 ttm_bo_mem_put(bo, &tmp_mem);
1125 return ret;
1126 }
1127
1128 static void
nouveau_bo_move_ntfy(struct ttm_buffer_object * bo,struct ttm_mem_reg * new_mem)1129 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem)
1130 {
1131 struct nouveau_bo *nvbo = nouveau_bo(bo);
1132 struct nouveau_vma *vma;
1133
1134 /* ttm can now (stupidly) pass the driver bos it didn't create... */
1135 if (bo->destroy != nouveau_bo_del_ttm)
1136 return;
1137
1138 list_for_each_entry(vma, &nvbo->vma_list, head) {
1139 if (new_mem && new_mem->mem_type == TTM_PL_VRAM) {
1140 nouveau_vm_map(vma, new_mem->mm_node);
1141 } else
1142 if (new_mem && new_mem->mem_type == TTM_PL_TT &&
1143 nvbo->page_shift == vma->vm->vmm->spg_shift) {
1144 if (((struct nouveau_mem *)new_mem->mm_node)->sg)
1145 nouveau_vm_map_sg_table(vma, 0, new_mem->
1146 num_pages << PAGE_SHIFT,
1147 new_mem->mm_node);
1148 else
1149 nouveau_vm_map_sg(vma, 0, new_mem->
1150 num_pages << PAGE_SHIFT,
1151 new_mem->mm_node);
1152 } else {
1153 nouveau_vm_unmap(vma);
1154 }
1155 }
1156 }
1157
1158 static int
nouveau_bo_vm_bind(struct ttm_buffer_object * bo,struct ttm_mem_reg * new_mem,struct nouveau_drm_tile ** new_tile)1159 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
1160 struct nouveau_drm_tile **new_tile)
1161 {
1162 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1163 struct drm_device *dev = drm->dev;
1164 struct nouveau_bo *nvbo = nouveau_bo(bo);
1165 u64 offset = new_mem->start << PAGE_SHIFT;
1166
1167 *new_tile = NULL;
1168 if (new_mem->mem_type != TTM_PL_VRAM)
1169 return 0;
1170
1171 if (nv_device(drm->device)->card_type >= NV_10) {
1172 *new_tile = nv10_bo_set_tiling(dev, offset, new_mem->size,
1173 nvbo->tile_mode,
1174 nvbo->tile_flags);
1175 }
1176
1177 return 0;
1178 }
1179
1180 static void
nouveau_bo_vm_cleanup(struct ttm_buffer_object * bo,struct nouveau_drm_tile * new_tile,struct nouveau_drm_tile ** old_tile)1181 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1182 struct nouveau_drm_tile *new_tile,
1183 struct nouveau_drm_tile **old_tile)
1184 {
1185 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1186 struct drm_device *dev = drm->dev;
1187
1188 nv10_bo_put_tile_region(dev, *old_tile, bo->sync_obj);
1189 *old_tile = new_tile;
1190 }
1191
1192 static int
nouveau_bo_move(struct ttm_buffer_object * bo,bool evict,bool intr,bool no_wait_gpu,struct ttm_mem_reg * new_mem)1193 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
1194 bool no_wait_gpu, struct ttm_mem_reg *new_mem)
1195 {
1196 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1197 struct nouveau_bo *nvbo = nouveau_bo(bo);
1198 struct ttm_mem_reg *old_mem = &bo->mem;
1199 struct nouveau_drm_tile *new_tile = NULL;
1200 int ret = 0;
1201
1202 if (nv_device(drm->device)->card_type < NV_50) {
1203 ret = nouveau_bo_vm_bind(bo, new_mem, &new_tile);
1204 if (ret)
1205 return ret;
1206 }
1207
1208 /* Fake bo copy. */
1209 if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
1210 BUG_ON(bo->mem.mm_node != NULL);
1211 bo->mem = *new_mem;
1212 new_mem->mm_node = NULL;
1213 goto out;
1214 }
1215
1216 /* CPU copy if we have no accelerated method available */
1217 if (!drm->ttm.move) {
1218 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
1219 goto out;
1220 }
1221
1222 /* Hardware assisted copy. */
1223 if (new_mem->mem_type == TTM_PL_SYSTEM)
1224 ret = nouveau_bo_move_flipd(bo, evict, intr,
1225 no_wait_gpu, new_mem);
1226 else if (old_mem->mem_type == TTM_PL_SYSTEM)
1227 ret = nouveau_bo_move_flips(bo, evict, intr,
1228 no_wait_gpu, new_mem);
1229 else
1230 ret = nouveau_bo_move_m2mf(bo, evict, intr,
1231 no_wait_gpu, new_mem);
1232
1233 if (!ret)
1234 goto out;
1235
1236 /* Fallback to software copy. */
1237 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
1238
1239 out:
1240 if (nv_device(drm->device)->card_type < NV_50) {
1241 if (ret)
1242 nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1243 else
1244 nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1245 }
1246
1247 return ret;
1248 }
1249
1250 static int
nouveau_bo_verify_access(struct ttm_buffer_object * bo,struct file * filp)1251 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1252 {
1253 return 0;
1254 }
1255
1256 static int
nouveau_ttm_io_mem_reserve(struct ttm_bo_device * bdev,struct ttm_mem_reg * mem)1257 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1258 {
1259 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1260 struct nouveau_drm *drm = nouveau_bdev(bdev);
1261 struct drm_device *dev = drm->dev;
1262 int ret;
1263
1264 mem->bus.addr = NULL;
1265 mem->bus.offset = 0;
1266 mem->bus.size = mem->num_pages << PAGE_SHIFT;
1267 mem->bus.base = 0;
1268 mem->bus.is_iomem = false;
1269 if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
1270 return -EINVAL;
1271 switch (mem->mem_type) {
1272 case TTM_PL_SYSTEM:
1273 /* System memory */
1274 return 0;
1275 case TTM_PL_TT:
1276 #if __OS_HAS_AGP
1277 if (drm->agp.stat == ENABLED) {
1278 mem->bus.offset = mem->start << PAGE_SHIFT;
1279 mem->bus.base = drm->agp.base;
1280 mem->bus.is_iomem = !dev->agp->cant_use_aperture;
1281 }
1282 #endif
1283 break;
1284 case TTM_PL_VRAM:
1285 mem->bus.offset = mem->start << PAGE_SHIFT;
1286 mem->bus.base = pci_resource_start(dev->pdev, 1);
1287 mem->bus.is_iomem = true;
1288 if (nv_device(drm->device)->card_type >= NV_50) {
1289 struct nouveau_bar *bar = nouveau_bar(drm->device);
1290 struct nouveau_mem *node = mem->mm_node;
1291
1292 ret = bar->umap(bar, node, NV_MEM_ACCESS_RW,
1293 &node->bar_vma);
1294 if (ret)
1295 return ret;
1296
1297 mem->bus.offset = node->bar_vma.offset;
1298 }
1299 break;
1300 default:
1301 return -EINVAL;
1302 }
1303 return 0;
1304 }
1305
1306 static void
nouveau_ttm_io_mem_free(struct ttm_bo_device * bdev,struct ttm_mem_reg * mem)1307 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1308 {
1309 struct nouveau_drm *drm = nouveau_bdev(bdev);
1310 struct nouveau_bar *bar = nouveau_bar(drm->device);
1311 struct nouveau_mem *node = mem->mm_node;
1312
1313 if (!node->bar_vma.node)
1314 return;
1315
1316 bar->unmap(bar, &node->bar_vma);
1317 }
1318
1319 static int
nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object * bo)1320 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1321 {
1322 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1323 struct nouveau_bo *nvbo = nouveau_bo(bo);
1324 struct nouveau_device *device = nv_device(drm->device);
1325 u32 mappable = pci_resource_len(device->pdev, 1) >> PAGE_SHIFT;
1326
1327 /* as long as the bo isn't in vram, and isn't tiled, we've got
1328 * nothing to do here.
1329 */
1330 if (bo->mem.mem_type != TTM_PL_VRAM) {
1331 if (nv_device(drm->device)->card_type < NV_50 ||
1332 !nouveau_bo_tile_layout(nvbo))
1333 return 0;
1334 }
1335
1336 /* make sure bo is in mappable vram */
1337 if (bo->mem.start + bo->mem.num_pages < mappable)
1338 return 0;
1339
1340
1341 nvbo->placement.fpfn = 0;
1342 nvbo->placement.lpfn = mappable;
1343 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_VRAM, 0);
1344 return nouveau_bo_validate(nvbo, false, false);
1345 }
1346
1347 static int
nouveau_ttm_tt_populate(struct ttm_tt * ttm)1348 nouveau_ttm_tt_populate(struct ttm_tt *ttm)
1349 {
1350 struct ttm_dma_tt *ttm_dma = (void *)ttm;
1351 struct nouveau_drm *drm;
1352 struct drm_device *dev;
1353 unsigned i;
1354 int r;
1355 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1356
1357 if (ttm->state != tt_unpopulated)
1358 return 0;
1359
1360 if (slave && ttm->sg) {
1361 /* make userspace faulting work */
1362 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1363 ttm_dma->dma_address, ttm->num_pages);
1364 ttm->state = tt_unbound;
1365 return 0;
1366 }
1367
1368 drm = nouveau_bdev(ttm->bdev);
1369 dev = drm->dev;
1370
1371 #if __OS_HAS_AGP
1372 if (drm->agp.stat == ENABLED) {
1373 return ttm_agp_tt_populate(ttm);
1374 }
1375 #endif
1376
1377 #ifdef CONFIG_SWIOTLB
1378 if (swiotlb_nr_tbl()) {
1379 return ttm_dma_populate((void *)ttm, dev->dev);
1380 }
1381 #endif
1382
1383 r = ttm_pool_populate(ttm);
1384 if (r) {
1385 return r;
1386 }
1387
1388 for (i = 0; i < ttm->num_pages; i++) {
1389 ttm_dma->dma_address[i] = pci_map_page(dev->pdev, ttm->pages[i],
1390 0, PAGE_SIZE,
1391 PCI_DMA_BIDIRECTIONAL);
1392 if (pci_dma_mapping_error(dev->pdev, ttm_dma->dma_address[i])) {
1393 while (--i) {
1394 pci_unmap_page(dev->pdev, ttm_dma->dma_address[i],
1395 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
1396 ttm_dma->dma_address[i] = 0;
1397 }
1398 ttm_pool_unpopulate(ttm);
1399 return -EFAULT;
1400 }
1401 }
1402 return 0;
1403 }
1404
1405 static void
nouveau_ttm_tt_unpopulate(struct ttm_tt * ttm)1406 nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
1407 {
1408 struct ttm_dma_tt *ttm_dma = (void *)ttm;
1409 struct nouveau_drm *drm;
1410 struct drm_device *dev;
1411 unsigned i;
1412 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1413
1414 if (slave)
1415 return;
1416
1417 drm = nouveau_bdev(ttm->bdev);
1418 dev = drm->dev;
1419
1420 #if __OS_HAS_AGP
1421 if (drm->agp.stat == ENABLED) {
1422 ttm_agp_tt_unpopulate(ttm);
1423 return;
1424 }
1425 #endif
1426
1427 #ifdef CONFIG_SWIOTLB
1428 if (swiotlb_nr_tbl()) {
1429 ttm_dma_unpopulate((void *)ttm, dev->dev);
1430 return;
1431 }
1432 #endif
1433
1434 for (i = 0; i < ttm->num_pages; i++) {
1435 if (ttm_dma->dma_address[i]) {
1436 pci_unmap_page(dev->pdev, ttm_dma->dma_address[i],
1437 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
1438 }
1439 }
1440
1441 ttm_pool_unpopulate(ttm);
1442 }
1443
1444 void
nouveau_bo_fence(struct nouveau_bo * nvbo,struct nouveau_fence * fence)1445 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
1446 {
1447 struct nouveau_fence *old_fence = NULL;
1448
1449 if (likely(fence))
1450 nouveau_fence_ref(fence);
1451
1452 spin_lock(&nvbo->bo.bdev->fence_lock);
1453 old_fence = nvbo->bo.sync_obj;
1454 nvbo->bo.sync_obj = fence;
1455 spin_unlock(&nvbo->bo.bdev->fence_lock);
1456
1457 nouveau_fence_unref(&old_fence);
1458 }
1459
1460 static void
nouveau_bo_fence_unref(void ** sync_obj)1461 nouveau_bo_fence_unref(void **sync_obj)
1462 {
1463 nouveau_fence_unref((struct nouveau_fence **)sync_obj);
1464 }
1465
1466 static void *
nouveau_bo_fence_ref(void * sync_obj)1467 nouveau_bo_fence_ref(void *sync_obj)
1468 {
1469 return nouveau_fence_ref(sync_obj);
1470 }
1471
1472 static bool
nouveau_bo_fence_signalled(void * sync_obj)1473 nouveau_bo_fence_signalled(void *sync_obj)
1474 {
1475 return nouveau_fence_done(sync_obj);
1476 }
1477
1478 static int
nouveau_bo_fence_wait(void * sync_obj,bool lazy,bool intr)1479 nouveau_bo_fence_wait(void *sync_obj, bool lazy, bool intr)
1480 {
1481 return nouveau_fence_wait(sync_obj, lazy, intr);
1482 }
1483
1484 static int
nouveau_bo_fence_flush(void * sync_obj)1485 nouveau_bo_fence_flush(void *sync_obj)
1486 {
1487 return 0;
1488 }
1489
1490 struct ttm_bo_driver nouveau_bo_driver = {
1491 .ttm_tt_create = &nouveau_ttm_tt_create,
1492 .ttm_tt_populate = &nouveau_ttm_tt_populate,
1493 .ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1494 .invalidate_caches = nouveau_bo_invalidate_caches,
1495 .init_mem_type = nouveau_bo_init_mem_type,
1496 .evict_flags = nouveau_bo_evict_flags,
1497 .move_notify = nouveau_bo_move_ntfy,
1498 .move = nouveau_bo_move,
1499 .verify_access = nouveau_bo_verify_access,
1500 .sync_obj_signaled = nouveau_bo_fence_signalled,
1501 .sync_obj_wait = nouveau_bo_fence_wait,
1502 .sync_obj_flush = nouveau_bo_fence_flush,
1503 .sync_obj_unref = nouveau_bo_fence_unref,
1504 .sync_obj_ref = nouveau_bo_fence_ref,
1505 .fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1506 .io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1507 .io_mem_free = &nouveau_ttm_io_mem_free,
1508 };
1509
1510 struct nouveau_vma *
nouveau_bo_vma_find(struct nouveau_bo * nvbo,struct nouveau_vm * vm)1511 nouveau_bo_vma_find(struct nouveau_bo *nvbo, struct nouveau_vm *vm)
1512 {
1513 struct nouveau_vma *vma;
1514 list_for_each_entry(vma, &nvbo->vma_list, head) {
1515 if (vma->vm == vm)
1516 return vma;
1517 }
1518
1519 return NULL;
1520 }
1521
1522 int
nouveau_bo_vma_add(struct nouveau_bo * nvbo,struct nouveau_vm * vm,struct nouveau_vma * vma)1523 nouveau_bo_vma_add(struct nouveau_bo *nvbo, struct nouveau_vm *vm,
1524 struct nouveau_vma *vma)
1525 {
1526 const u32 size = nvbo->bo.mem.num_pages << PAGE_SHIFT;
1527 struct nouveau_mem *node = nvbo->bo.mem.mm_node;
1528 int ret;
1529
1530 ret = nouveau_vm_get(vm, size, nvbo->page_shift,
1531 NV_MEM_ACCESS_RW, vma);
1532 if (ret)
1533 return ret;
1534
1535 if (nvbo->bo.mem.mem_type == TTM_PL_VRAM)
1536 nouveau_vm_map(vma, nvbo->bo.mem.mm_node);
1537 else if (nvbo->bo.mem.mem_type == TTM_PL_TT) {
1538 if (node->sg)
1539 nouveau_vm_map_sg_table(vma, 0, size, node);
1540 else
1541 nouveau_vm_map_sg(vma, 0, size, node);
1542 }
1543
1544 list_add_tail(&vma->head, &nvbo->vma_list);
1545 vma->refcount = 1;
1546 return 0;
1547 }
1548
1549 void
nouveau_bo_vma_del(struct nouveau_bo * nvbo,struct nouveau_vma * vma)1550 nouveau_bo_vma_del(struct nouveau_bo *nvbo, struct nouveau_vma *vma)
1551 {
1552 if (vma->node) {
1553 if (nvbo->bo.mem.mem_type != TTM_PL_SYSTEM) {
1554 spin_lock(&nvbo->bo.bdev->fence_lock);
1555 ttm_bo_wait(&nvbo->bo, false, false, false);
1556 spin_unlock(&nvbo->bo.bdev->fence_lock);
1557 nouveau_vm_unmap(vma);
1558 }
1559
1560 nouveau_vm_put(vma);
1561 list_del(&vma->head);
1562 }
1563 }
1564