1 /*
2 * Copyright 2018 Red Hat Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 */
22 #include "nouveau_dmem.h"
23 #include "nouveau_drv.h"
24 #include "nouveau_chan.h"
25 #include "nouveau_dma.h"
26 #include "nouveau_mem.h"
27 #include "nouveau_bo.h"
28 #include "nouveau_svm.h"
29
30 #include <nvif/class.h>
31 #include <nvif/object.h>
32 #include <nvif/push906f.h>
33 #include <nvif/if000c.h>
34 #include <nvif/if500b.h>
35 #include <nvif/if900b.h>
36 #include <nvif/if000c.h>
37
38 #include <nvhw/class/cla0b5.h>
39
40 #include <linux/sched/mm.h>
41 #include <linux/hmm.h>
42
43 /*
44 * FIXME: this is ugly right now we are using TTM to allocate vram and we pin
45 * it in vram while in use. We likely want to overhaul memory management for
46 * nouveau to be more page like (not necessarily with system page size but a
47 * bigger page size) at lowest level and have some shim layer on top that would
48 * provide the same functionality as TTM.
49 */
50 #define DMEM_CHUNK_SIZE (2UL << 20)
51 #define DMEM_CHUNK_NPAGES (DMEM_CHUNK_SIZE >> PAGE_SHIFT)
52
53 enum nouveau_aper {
54 NOUVEAU_APER_VIRT,
55 NOUVEAU_APER_VRAM,
56 NOUVEAU_APER_HOST,
57 };
58
59 typedef int (*nouveau_migrate_copy_t)(struct nouveau_drm *drm, u64 npages,
60 enum nouveau_aper, u64 dst_addr,
61 enum nouveau_aper, u64 src_addr);
62 typedef int (*nouveau_clear_page_t)(struct nouveau_drm *drm, u32 length,
63 enum nouveau_aper, u64 dst_addr);
64
65 struct nouveau_dmem_chunk {
66 struct list_head list;
67 struct nouveau_bo *bo;
68 struct nouveau_drm *drm;
69 unsigned long callocated;
70 struct dev_pagemap pagemap;
71 };
72
73 struct nouveau_dmem_migrate {
74 nouveau_migrate_copy_t copy_func;
75 nouveau_clear_page_t clear_func;
76 struct nouveau_channel *chan;
77 };
78
79 struct nouveau_dmem {
80 struct nouveau_drm *drm;
81 struct nouveau_dmem_migrate migrate;
82 struct list_head chunks;
83 struct mutex mutex;
84 struct page *free_pages;
85 spinlock_t lock;
86 };
87
nouveau_page_to_chunk(struct page * page)88 static struct nouveau_dmem_chunk *nouveau_page_to_chunk(struct page *page)
89 {
90 return container_of(page->pgmap, struct nouveau_dmem_chunk, pagemap);
91 }
92
page_to_drm(struct page * page)93 static struct nouveau_drm *page_to_drm(struct page *page)
94 {
95 struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
96
97 return chunk->drm;
98 }
99
nouveau_dmem_page_addr(struct page * page)100 unsigned long nouveau_dmem_page_addr(struct page *page)
101 {
102 struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
103 unsigned long off = (page_to_pfn(page) << PAGE_SHIFT) -
104 chunk->pagemap.range.start;
105
106 return chunk->bo->offset + off;
107 }
108
nouveau_dmem_page_free(struct page * page)109 static void nouveau_dmem_page_free(struct page *page)
110 {
111 struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
112 struct nouveau_dmem *dmem = chunk->drm->dmem;
113
114 spin_lock(&dmem->lock);
115 page->zone_device_data = dmem->free_pages;
116 dmem->free_pages = page;
117
118 WARN_ON(!chunk->callocated);
119 chunk->callocated--;
120 /*
121 * FIXME when chunk->callocated reach 0 we should add the chunk to
122 * a reclaim list so that it can be freed in case of memory pressure.
123 */
124 spin_unlock(&dmem->lock);
125 }
126
nouveau_dmem_fence_done(struct nouveau_fence ** fence)127 static void nouveau_dmem_fence_done(struct nouveau_fence **fence)
128 {
129 if (fence) {
130 nouveau_fence_wait(*fence, true, false);
131 nouveau_fence_unref(fence);
132 } else {
133 /*
134 * FIXME wait for channel to be IDLE before calling finalizing
135 * the hmem object.
136 */
137 }
138 }
139
nouveau_dmem_fault_copy_one(struct nouveau_drm * drm,struct vm_fault * vmf,struct migrate_vma * args,dma_addr_t * dma_addr)140 static vm_fault_t nouveau_dmem_fault_copy_one(struct nouveau_drm *drm,
141 struct vm_fault *vmf, struct migrate_vma *args,
142 dma_addr_t *dma_addr)
143 {
144 struct device *dev = drm->dev->dev;
145 struct page *dpage, *spage;
146 struct nouveau_svmm *svmm;
147
148 spage = migrate_pfn_to_page(args->src[0]);
149 if (!spage || !(args->src[0] & MIGRATE_PFN_MIGRATE))
150 return 0;
151
152 dpage = alloc_page_vma(GFP_HIGHUSER, vmf->vma, vmf->address);
153 if (!dpage)
154 return VM_FAULT_SIGBUS;
155 lock_page(dpage);
156
157 *dma_addr = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
158 if (dma_mapping_error(dev, *dma_addr))
159 goto error_free_page;
160
161 svmm = spage->zone_device_data;
162 mutex_lock(&svmm->mutex);
163 nouveau_svmm_invalidate(svmm, args->start, args->end);
164 if (drm->dmem->migrate.copy_func(drm, 1, NOUVEAU_APER_HOST, *dma_addr,
165 NOUVEAU_APER_VRAM, nouveau_dmem_page_addr(spage)))
166 goto error_dma_unmap;
167 mutex_unlock(&svmm->mutex);
168
169 args->dst[0] = migrate_pfn(page_to_pfn(dpage)) | MIGRATE_PFN_LOCKED;
170 return 0;
171
172 error_dma_unmap:
173 mutex_unlock(&svmm->mutex);
174 dma_unmap_page(dev, *dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
175 error_free_page:
176 __free_page(dpage);
177 return VM_FAULT_SIGBUS;
178 }
179
nouveau_dmem_migrate_to_ram(struct vm_fault * vmf)180 static vm_fault_t nouveau_dmem_migrate_to_ram(struct vm_fault *vmf)
181 {
182 struct nouveau_drm *drm = page_to_drm(vmf->page);
183 struct nouveau_dmem *dmem = drm->dmem;
184 struct nouveau_fence *fence;
185 unsigned long src = 0, dst = 0;
186 dma_addr_t dma_addr = 0;
187 vm_fault_t ret;
188 struct migrate_vma args = {
189 .vma = vmf->vma,
190 .start = vmf->address,
191 .end = vmf->address + PAGE_SIZE,
192 .src = &src,
193 .dst = &dst,
194 .pgmap_owner = drm->dev,
195 .flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE,
196 };
197
198 /*
199 * FIXME what we really want is to find some heuristic to migrate more
200 * than just one page on CPU fault. When such fault happens it is very
201 * likely that more surrounding page will CPU fault too.
202 */
203 if (migrate_vma_setup(&args) < 0)
204 return VM_FAULT_SIGBUS;
205 if (!args.cpages)
206 return 0;
207
208 ret = nouveau_dmem_fault_copy_one(drm, vmf, &args, &dma_addr);
209 if (ret || dst == 0)
210 goto done;
211
212 nouveau_fence_new(dmem->migrate.chan, false, &fence);
213 migrate_vma_pages(&args);
214 nouveau_dmem_fence_done(&fence);
215 dma_unmap_page(drm->dev->dev, dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
216 done:
217 migrate_vma_finalize(&args);
218 return ret;
219 }
220
221 static const struct dev_pagemap_ops nouveau_dmem_pagemap_ops = {
222 .page_free = nouveau_dmem_page_free,
223 .migrate_to_ram = nouveau_dmem_migrate_to_ram,
224 };
225
226 static int
nouveau_dmem_chunk_alloc(struct nouveau_drm * drm,struct page ** ppage)227 nouveau_dmem_chunk_alloc(struct nouveau_drm *drm, struct page **ppage)
228 {
229 struct nouveau_dmem_chunk *chunk;
230 struct resource *res;
231 struct page *page;
232 void *ptr;
233 unsigned long i, pfn_first;
234 int ret;
235
236 chunk = kzalloc(sizeof(*chunk), GFP_KERNEL);
237 if (chunk == NULL) {
238 ret = -ENOMEM;
239 goto out;
240 }
241
242 /* Allocate unused physical address space for device private pages. */
243 res = request_free_mem_region(&iomem_resource, DMEM_CHUNK_SIZE,
244 "nouveau_dmem");
245 if (IS_ERR(res)) {
246 ret = PTR_ERR(res);
247 goto out_free;
248 }
249
250 chunk->drm = drm;
251 chunk->pagemap.type = MEMORY_DEVICE_PRIVATE;
252 chunk->pagemap.range.start = res->start;
253 chunk->pagemap.range.end = res->end;
254 chunk->pagemap.nr_range = 1;
255 chunk->pagemap.ops = &nouveau_dmem_pagemap_ops;
256 chunk->pagemap.owner = drm->dev;
257
258 ret = nouveau_bo_new(&drm->client, DMEM_CHUNK_SIZE, 0,
259 NOUVEAU_GEM_DOMAIN_VRAM, 0, 0, NULL, NULL,
260 &chunk->bo);
261 if (ret)
262 goto out_release;
263
264 ret = nouveau_bo_pin(chunk->bo, NOUVEAU_GEM_DOMAIN_VRAM, false);
265 if (ret)
266 goto out_bo_free;
267
268 ptr = memremap_pages(&chunk->pagemap, numa_node_id());
269 if (IS_ERR(ptr)) {
270 ret = PTR_ERR(ptr);
271 goto out_bo_unpin;
272 }
273
274 mutex_lock(&drm->dmem->mutex);
275 list_add(&chunk->list, &drm->dmem->chunks);
276 mutex_unlock(&drm->dmem->mutex);
277
278 pfn_first = chunk->pagemap.range.start >> PAGE_SHIFT;
279 page = pfn_to_page(pfn_first);
280 spin_lock(&drm->dmem->lock);
281 for (i = 0; i < DMEM_CHUNK_NPAGES - 1; ++i, ++page) {
282 page->zone_device_data = drm->dmem->free_pages;
283 drm->dmem->free_pages = page;
284 }
285 *ppage = page;
286 chunk->callocated++;
287 spin_unlock(&drm->dmem->lock);
288
289 NV_INFO(drm, "DMEM: registered %ldMB of device memory\n",
290 DMEM_CHUNK_SIZE >> 20);
291
292 return 0;
293
294 out_bo_unpin:
295 nouveau_bo_unpin(chunk->bo);
296 out_bo_free:
297 nouveau_bo_ref(NULL, &chunk->bo);
298 out_release:
299 release_mem_region(chunk->pagemap.range.start, range_len(&chunk->pagemap.range));
300 out_free:
301 kfree(chunk);
302 out:
303 return ret;
304 }
305
306 static struct page *
nouveau_dmem_page_alloc_locked(struct nouveau_drm * drm)307 nouveau_dmem_page_alloc_locked(struct nouveau_drm *drm)
308 {
309 struct nouveau_dmem_chunk *chunk;
310 struct page *page = NULL;
311 int ret;
312
313 spin_lock(&drm->dmem->lock);
314 if (drm->dmem->free_pages) {
315 page = drm->dmem->free_pages;
316 drm->dmem->free_pages = page->zone_device_data;
317 chunk = nouveau_page_to_chunk(page);
318 chunk->callocated++;
319 spin_unlock(&drm->dmem->lock);
320 } else {
321 spin_unlock(&drm->dmem->lock);
322 ret = nouveau_dmem_chunk_alloc(drm, &page);
323 if (ret)
324 return NULL;
325 }
326
327 get_page(page);
328 lock_page(page);
329 return page;
330 }
331
332 static void
nouveau_dmem_page_free_locked(struct nouveau_drm * drm,struct page * page)333 nouveau_dmem_page_free_locked(struct nouveau_drm *drm, struct page *page)
334 {
335 unlock_page(page);
336 put_page(page);
337 }
338
339 void
nouveau_dmem_resume(struct nouveau_drm * drm)340 nouveau_dmem_resume(struct nouveau_drm *drm)
341 {
342 struct nouveau_dmem_chunk *chunk;
343 int ret;
344
345 if (drm->dmem == NULL)
346 return;
347
348 mutex_lock(&drm->dmem->mutex);
349 list_for_each_entry(chunk, &drm->dmem->chunks, list) {
350 ret = nouveau_bo_pin(chunk->bo, NOUVEAU_GEM_DOMAIN_VRAM, false);
351 /* FIXME handle pin failure */
352 WARN_ON(ret);
353 }
354 mutex_unlock(&drm->dmem->mutex);
355 }
356
357 void
nouveau_dmem_suspend(struct nouveau_drm * drm)358 nouveau_dmem_suspend(struct nouveau_drm *drm)
359 {
360 struct nouveau_dmem_chunk *chunk;
361
362 if (drm->dmem == NULL)
363 return;
364
365 mutex_lock(&drm->dmem->mutex);
366 list_for_each_entry(chunk, &drm->dmem->chunks, list)
367 nouveau_bo_unpin(chunk->bo);
368 mutex_unlock(&drm->dmem->mutex);
369 }
370
371 void
nouveau_dmem_fini(struct nouveau_drm * drm)372 nouveau_dmem_fini(struct nouveau_drm *drm)
373 {
374 struct nouveau_dmem_chunk *chunk, *tmp;
375
376 if (drm->dmem == NULL)
377 return;
378
379 mutex_lock(&drm->dmem->mutex);
380
381 list_for_each_entry_safe(chunk, tmp, &drm->dmem->chunks, list) {
382 nouveau_bo_unpin(chunk->bo);
383 nouveau_bo_ref(NULL, &chunk->bo);
384 list_del(&chunk->list);
385 memunmap_pages(&chunk->pagemap);
386 release_mem_region(chunk->pagemap.range.start,
387 range_len(&chunk->pagemap.range));
388 kfree(chunk);
389 }
390
391 mutex_unlock(&drm->dmem->mutex);
392 }
393
394 static int
nvc0b5_migrate_copy(struct nouveau_drm * drm,u64 npages,enum nouveau_aper dst_aper,u64 dst_addr,enum nouveau_aper src_aper,u64 src_addr)395 nvc0b5_migrate_copy(struct nouveau_drm *drm, u64 npages,
396 enum nouveau_aper dst_aper, u64 dst_addr,
397 enum nouveau_aper src_aper, u64 src_addr)
398 {
399 struct nvif_push *push = drm->dmem->migrate.chan->chan.push;
400 u32 launch_dma = 0;
401 int ret;
402
403 ret = PUSH_WAIT(push, 13);
404 if (ret)
405 return ret;
406
407 if (src_aper != NOUVEAU_APER_VIRT) {
408 switch (src_aper) {
409 case NOUVEAU_APER_VRAM:
410 PUSH_IMMD(push, NVA0B5, SET_SRC_PHYS_MODE,
411 NVDEF(NVA0B5, SET_SRC_PHYS_MODE, TARGET, LOCAL_FB));
412 break;
413 case NOUVEAU_APER_HOST:
414 PUSH_IMMD(push, NVA0B5, SET_SRC_PHYS_MODE,
415 NVDEF(NVA0B5, SET_SRC_PHYS_MODE, TARGET, COHERENT_SYSMEM));
416 break;
417 default:
418 return -EINVAL;
419 }
420
421 launch_dma |= NVDEF(NVA0B5, LAUNCH_DMA, SRC_TYPE, PHYSICAL);
422 }
423
424 if (dst_aper != NOUVEAU_APER_VIRT) {
425 switch (dst_aper) {
426 case NOUVEAU_APER_VRAM:
427 PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
428 NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, LOCAL_FB));
429 break;
430 case NOUVEAU_APER_HOST:
431 PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
432 NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, COHERENT_SYSMEM));
433 break;
434 default:
435 return -EINVAL;
436 }
437
438 launch_dma |= NVDEF(NVA0B5, LAUNCH_DMA, DST_TYPE, PHYSICAL);
439 }
440
441 PUSH_MTHD(push, NVA0B5, OFFSET_IN_UPPER,
442 NVVAL(NVA0B5, OFFSET_IN_UPPER, UPPER, upper_32_bits(src_addr)),
443
444 OFFSET_IN_LOWER, lower_32_bits(src_addr),
445
446 OFFSET_OUT_UPPER,
447 NVVAL(NVA0B5, OFFSET_OUT_UPPER, UPPER, upper_32_bits(dst_addr)),
448
449 OFFSET_OUT_LOWER, lower_32_bits(dst_addr),
450 PITCH_IN, PAGE_SIZE,
451 PITCH_OUT, PAGE_SIZE,
452 LINE_LENGTH_IN, PAGE_SIZE,
453 LINE_COUNT, npages);
454
455 PUSH_MTHD(push, NVA0B5, LAUNCH_DMA, launch_dma |
456 NVDEF(NVA0B5, LAUNCH_DMA, DATA_TRANSFER_TYPE, NON_PIPELINED) |
457 NVDEF(NVA0B5, LAUNCH_DMA, FLUSH_ENABLE, TRUE) |
458 NVDEF(NVA0B5, LAUNCH_DMA, SEMAPHORE_TYPE, NONE) |
459 NVDEF(NVA0B5, LAUNCH_DMA, INTERRUPT_TYPE, NONE) |
460 NVDEF(NVA0B5, LAUNCH_DMA, SRC_MEMORY_LAYOUT, PITCH) |
461 NVDEF(NVA0B5, LAUNCH_DMA, DST_MEMORY_LAYOUT, PITCH) |
462 NVDEF(NVA0B5, LAUNCH_DMA, MULTI_LINE_ENABLE, TRUE) |
463 NVDEF(NVA0B5, LAUNCH_DMA, REMAP_ENABLE, FALSE) |
464 NVDEF(NVA0B5, LAUNCH_DMA, BYPASS_L2, USE_PTE_SETTING));
465 return 0;
466 }
467
468 static int
nvc0b5_migrate_clear(struct nouveau_drm * drm,u32 length,enum nouveau_aper dst_aper,u64 dst_addr)469 nvc0b5_migrate_clear(struct nouveau_drm *drm, u32 length,
470 enum nouveau_aper dst_aper, u64 dst_addr)
471 {
472 struct nvif_push *push = drm->dmem->migrate.chan->chan.push;
473 u32 launch_dma = 0;
474 int ret;
475
476 ret = PUSH_WAIT(push, 12);
477 if (ret)
478 return ret;
479
480 switch (dst_aper) {
481 case NOUVEAU_APER_VRAM:
482 PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
483 NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, LOCAL_FB));
484 break;
485 case NOUVEAU_APER_HOST:
486 PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
487 NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, COHERENT_SYSMEM));
488 break;
489 default:
490 return -EINVAL;
491 }
492
493 launch_dma |= NVDEF(NVA0B5, LAUNCH_DMA, DST_TYPE, PHYSICAL);
494
495 PUSH_MTHD(push, NVA0B5, SET_REMAP_CONST_A, 0,
496 SET_REMAP_CONST_B, 0,
497
498 SET_REMAP_COMPONENTS,
499 NVDEF(NVA0B5, SET_REMAP_COMPONENTS, DST_X, CONST_A) |
500 NVDEF(NVA0B5, SET_REMAP_COMPONENTS, DST_Y, CONST_B) |
501 NVDEF(NVA0B5, SET_REMAP_COMPONENTS, COMPONENT_SIZE, FOUR) |
502 NVDEF(NVA0B5, SET_REMAP_COMPONENTS, NUM_DST_COMPONENTS, TWO));
503
504 PUSH_MTHD(push, NVA0B5, OFFSET_OUT_UPPER,
505 NVVAL(NVA0B5, OFFSET_OUT_UPPER, UPPER, upper_32_bits(dst_addr)),
506
507 OFFSET_OUT_LOWER, lower_32_bits(dst_addr));
508
509 PUSH_MTHD(push, NVA0B5, LINE_LENGTH_IN, length >> 3);
510
511 PUSH_MTHD(push, NVA0B5, LAUNCH_DMA, launch_dma |
512 NVDEF(NVA0B5, LAUNCH_DMA, DATA_TRANSFER_TYPE, NON_PIPELINED) |
513 NVDEF(NVA0B5, LAUNCH_DMA, FLUSH_ENABLE, TRUE) |
514 NVDEF(NVA0B5, LAUNCH_DMA, SEMAPHORE_TYPE, NONE) |
515 NVDEF(NVA0B5, LAUNCH_DMA, INTERRUPT_TYPE, NONE) |
516 NVDEF(NVA0B5, LAUNCH_DMA, SRC_MEMORY_LAYOUT, PITCH) |
517 NVDEF(NVA0B5, LAUNCH_DMA, DST_MEMORY_LAYOUT, PITCH) |
518 NVDEF(NVA0B5, LAUNCH_DMA, MULTI_LINE_ENABLE, FALSE) |
519 NVDEF(NVA0B5, LAUNCH_DMA, REMAP_ENABLE, TRUE) |
520 NVDEF(NVA0B5, LAUNCH_DMA, BYPASS_L2, USE_PTE_SETTING));
521 return 0;
522 }
523
524 static int
nouveau_dmem_migrate_init(struct nouveau_drm * drm)525 nouveau_dmem_migrate_init(struct nouveau_drm *drm)
526 {
527 switch (drm->ttm.copy.oclass) {
528 case PASCAL_DMA_COPY_A:
529 case PASCAL_DMA_COPY_B:
530 case VOLTA_DMA_COPY_A:
531 case TURING_DMA_COPY_A:
532 drm->dmem->migrate.copy_func = nvc0b5_migrate_copy;
533 drm->dmem->migrate.clear_func = nvc0b5_migrate_clear;
534 drm->dmem->migrate.chan = drm->ttm.chan;
535 return 0;
536 default:
537 break;
538 }
539 return -ENODEV;
540 }
541
542 void
nouveau_dmem_init(struct nouveau_drm * drm)543 nouveau_dmem_init(struct nouveau_drm *drm)
544 {
545 int ret;
546
547 /* This only make sense on PASCAL or newer */
548 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_PASCAL)
549 return;
550
551 if (!(drm->dmem = kzalloc(sizeof(*drm->dmem), GFP_KERNEL)))
552 return;
553
554 drm->dmem->drm = drm;
555 mutex_init(&drm->dmem->mutex);
556 INIT_LIST_HEAD(&drm->dmem->chunks);
557 mutex_init(&drm->dmem->mutex);
558 spin_lock_init(&drm->dmem->lock);
559
560 /* Initialize migration dma helpers before registering memory */
561 ret = nouveau_dmem_migrate_init(drm);
562 if (ret) {
563 kfree(drm->dmem);
564 drm->dmem = NULL;
565 }
566 }
567
nouveau_dmem_migrate_copy_one(struct nouveau_drm * drm,struct nouveau_svmm * svmm,unsigned long src,dma_addr_t * dma_addr,u64 * pfn)568 static unsigned long nouveau_dmem_migrate_copy_one(struct nouveau_drm *drm,
569 struct nouveau_svmm *svmm, unsigned long src,
570 dma_addr_t *dma_addr, u64 *pfn)
571 {
572 struct device *dev = drm->dev->dev;
573 struct page *dpage, *spage;
574 unsigned long paddr;
575
576 spage = migrate_pfn_to_page(src);
577 if (!(src & MIGRATE_PFN_MIGRATE))
578 goto out;
579
580 dpage = nouveau_dmem_page_alloc_locked(drm);
581 if (!dpage)
582 goto out;
583
584 paddr = nouveau_dmem_page_addr(dpage);
585 if (spage) {
586 *dma_addr = dma_map_page(dev, spage, 0, page_size(spage),
587 DMA_BIDIRECTIONAL);
588 if (dma_mapping_error(dev, *dma_addr))
589 goto out_free_page;
590 if (drm->dmem->migrate.copy_func(drm, 1,
591 NOUVEAU_APER_VRAM, paddr, NOUVEAU_APER_HOST, *dma_addr))
592 goto out_dma_unmap;
593 } else {
594 *dma_addr = DMA_MAPPING_ERROR;
595 if (drm->dmem->migrate.clear_func(drm, page_size(dpage),
596 NOUVEAU_APER_VRAM, paddr))
597 goto out_free_page;
598 }
599
600 dpage->zone_device_data = svmm;
601 *pfn = NVIF_VMM_PFNMAP_V0_V | NVIF_VMM_PFNMAP_V0_VRAM |
602 ((paddr >> PAGE_SHIFT) << NVIF_VMM_PFNMAP_V0_ADDR_SHIFT);
603 if (src & MIGRATE_PFN_WRITE)
604 *pfn |= NVIF_VMM_PFNMAP_V0_W;
605 return migrate_pfn(page_to_pfn(dpage)) | MIGRATE_PFN_LOCKED;
606
607 out_dma_unmap:
608 dma_unmap_page(dev, *dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
609 out_free_page:
610 nouveau_dmem_page_free_locked(drm, dpage);
611 out:
612 *pfn = NVIF_VMM_PFNMAP_V0_NONE;
613 return 0;
614 }
615
nouveau_dmem_migrate_chunk(struct nouveau_drm * drm,struct nouveau_svmm * svmm,struct migrate_vma * args,dma_addr_t * dma_addrs,u64 * pfns)616 static void nouveau_dmem_migrate_chunk(struct nouveau_drm *drm,
617 struct nouveau_svmm *svmm, struct migrate_vma *args,
618 dma_addr_t *dma_addrs, u64 *pfns)
619 {
620 struct nouveau_fence *fence;
621 unsigned long addr = args->start, nr_dma = 0, i;
622
623 for (i = 0; addr < args->end; i++) {
624 args->dst[i] = nouveau_dmem_migrate_copy_one(drm, svmm,
625 args->src[i], dma_addrs + nr_dma, pfns + i);
626 if (!dma_mapping_error(drm->dev->dev, dma_addrs[nr_dma]))
627 nr_dma++;
628 addr += PAGE_SIZE;
629 }
630
631 nouveau_fence_new(drm->dmem->migrate.chan, false, &fence);
632 migrate_vma_pages(args);
633 nouveau_dmem_fence_done(&fence);
634 nouveau_pfns_map(svmm, args->vma->vm_mm, args->start, pfns, i);
635
636 while (nr_dma--) {
637 dma_unmap_page(drm->dev->dev, dma_addrs[nr_dma], PAGE_SIZE,
638 DMA_BIDIRECTIONAL);
639 }
640 migrate_vma_finalize(args);
641 }
642
643 int
nouveau_dmem_migrate_vma(struct nouveau_drm * drm,struct nouveau_svmm * svmm,struct vm_area_struct * vma,unsigned long start,unsigned long end)644 nouveau_dmem_migrate_vma(struct nouveau_drm *drm,
645 struct nouveau_svmm *svmm,
646 struct vm_area_struct *vma,
647 unsigned long start,
648 unsigned long end)
649 {
650 unsigned long npages = (end - start) >> PAGE_SHIFT;
651 unsigned long max = min(SG_MAX_SINGLE_ALLOC, npages);
652 dma_addr_t *dma_addrs;
653 struct migrate_vma args = {
654 .vma = vma,
655 .start = start,
656 .pgmap_owner = drm->dev,
657 .flags = MIGRATE_VMA_SELECT_SYSTEM,
658 };
659 unsigned long i;
660 u64 *pfns;
661 int ret = -ENOMEM;
662
663 if (drm->dmem == NULL)
664 return -ENODEV;
665
666 args.src = kcalloc(max, sizeof(*args.src), GFP_KERNEL);
667 if (!args.src)
668 goto out;
669 args.dst = kcalloc(max, sizeof(*args.dst), GFP_KERNEL);
670 if (!args.dst)
671 goto out_free_src;
672
673 dma_addrs = kmalloc_array(max, sizeof(*dma_addrs), GFP_KERNEL);
674 if (!dma_addrs)
675 goto out_free_dst;
676
677 pfns = nouveau_pfns_alloc(max);
678 if (!pfns)
679 goto out_free_dma;
680
681 for (i = 0; i < npages; i += max) {
682 if (args.start + (max << PAGE_SHIFT) > end)
683 args.end = end;
684 else
685 args.end = args.start + (max << PAGE_SHIFT);
686
687 ret = migrate_vma_setup(&args);
688 if (ret)
689 goto out_free_pfns;
690
691 if (args.cpages)
692 nouveau_dmem_migrate_chunk(drm, svmm, &args, dma_addrs,
693 pfns);
694 args.start = args.end;
695 }
696
697 ret = 0;
698 out_free_pfns:
699 nouveau_pfns_free(pfns);
700 out_free_dma:
701 kfree(dma_addrs);
702 out_free_dst:
703 kfree(args.dst);
704 out_free_src:
705 kfree(args.src);
706 out:
707 return ret;
708 }
709