Loading...
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
29#include <nvif/class.h>
30#include <nvif/object.h>
31#include <nvif/if500b.h>
32#include <nvif/if900b.h>
33
34#include <linux/sched/mm.h>
35#include <linux/hmm.h>
36
37/*
38 * FIXME: this is ugly right now we are using TTM to allocate vram and we pin
39 * it in vram while in use. We likely want to overhaul memory management for
40 * nouveau to be more page like (not necessarily with system page size but a
41 * bigger page size) at lowest level and have some shim layer on top that would
42 * provide the same functionality as TTM.
43 */
44#define DMEM_CHUNK_SIZE (2UL << 20)
45#define DMEM_CHUNK_NPAGES (DMEM_CHUNK_SIZE >> PAGE_SHIFT)
46
47enum nouveau_aper {
48 NOUVEAU_APER_VIRT,
49 NOUVEAU_APER_VRAM,
50 NOUVEAU_APER_HOST,
51};
52
53typedef int (*nouveau_migrate_copy_t)(struct nouveau_drm *drm, u64 npages,
54 enum nouveau_aper, u64 dst_addr,
55 enum nouveau_aper, u64 src_addr);
56
57struct nouveau_dmem_chunk {
58 struct list_head list;
59 struct nouveau_bo *bo;
60 struct nouveau_drm *drm;
61 unsigned long pfn_first;
62 unsigned long callocated;
63 unsigned long bitmap[BITS_TO_LONGS(DMEM_CHUNK_NPAGES)];
64 spinlock_t lock;
65};
66
67struct nouveau_dmem_migrate {
68 nouveau_migrate_copy_t copy_func;
69 struct nouveau_channel *chan;
70};
71
72struct nouveau_dmem {
73 struct nouveau_drm *drm;
74 struct dev_pagemap pagemap;
75 struct nouveau_dmem_migrate migrate;
76 struct list_head chunk_free;
77 struct list_head chunk_full;
78 struct list_head chunk_empty;
79 struct mutex mutex;
80};
81
82static inline struct nouveau_dmem *page_to_dmem(struct page *page)
83{
84 return container_of(page->pgmap, struct nouveau_dmem, pagemap);
85}
86
87static unsigned long nouveau_dmem_page_addr(struct page *page)
88{
89 struct nouveau_dmem_chunk *chunk = page->zone_device_data;
90 unsigned long idx = page_to_pfn(page) - chunk->pfn_first;
91
92 return (idx << PAGE_SHIFT) + chunk->bo->bo.offset;
93}
94
95static void nouveau_dmem_page_free(struct page *page)
96{
97 struct nouveau_dmem_chunk *chunk = page->zone_device_data;
98 unsigned long idx = page_to_pfn(page) - chunk->pfn_first;
99
100 /*
101 * FIXME:
102 *
103 * This is really a bad example, we need to overhaul nouveau memory
104 * management to be more page focus and allow lighter locking scheme
105 * to be use in the process.
106 */
107 spin_lock(&chunk->lock);
108 clear_bit(idx, chunk->bitmap);
109 WARN_ON(!chunk->callocated);
110 chunk->callocated--;
111 /*
112 * FIXME when chunk->callocated reach 0 we should add the chunk to
113 * a reclaim list so that it can be freed in case of memory pressure.
114 */
115 spin_unlock(&chunk->lock);
116}
117
118static void nouveau_dmem_fence_done(struct nouveau_fence **fence)
119{
120 if (fence) {
121 nouveau_fence_wait(*fence, true, false);
122 nouveau_fence_unref(fence);
123 } else {
124 /*
125 * FIXME wait for channel to be IDLE before calling finalizing
126 * the hmem object.
127 */
128 }
129}
130
131static vm_fault_t nouveau_dmem_fault_copy_one(struct nouveau_drm *drm,
132 struct vm_fault *vmf, struct migrate_vma *args,
133 dma_addr_t *dma_addr)
134{
135 struct device *dev = drm->dev->dev;
136 struct page *dpage, *spage;
137
138 spage = migrate_pfn_to_page(args->src[0]);
139 if (!spage || !(args->src[0] & MIGRATE_PFN_MIGRATE))
140 return 0;
141
142 dpage = alloc_page_vma(GFP_HIGHUSER, vmf->vma, vmf->address);
143 if (!dpage)
144 return VM_FAULT_SIGBUS;
145 lock_page(dpage);
146
147 *dma_addr = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
148 if (dma_mapping_error(dev, *dma_addr))
149 goto error_free_page;
150
151 if (drm->dmem->migrate.copy_func(drm, 1, NOUVEAU_APER_HOST, *dma_addr,
152 NOUVEAU_APER_VRAM, nouveau_dmem_page_addr(spage)))
153 goto error_dma_unmap;
154
155 args->dst[0] = migrate_pfn(page_to_pfn(dpage)) | MIGRATE_PFN_LOCKED;
156 return 0;
157
158error_dma_unmap:
159 dma_unmap_page(dev, *dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
160error_free_page:
161 __free_page(dpage);
162 return VM_FAULT_SIGBUS;
163}
164
165static vm_fault_t nouveau_dmem_migrate_to_ram(struct vm_fault *vmf)
166{
167 struct nouveau_dmem *dmem = page_to_dmem(vmf->page);
168 struct nouveau_drm *drm = dmem->drm;
169 struct nouveau_fence *fence;
170 unsigned long src = 0, dst = 0;
171 dma_addr_t dma_addr = 0;
172 vm_fault_t ret;
173 struct migrate_vma args = {
174 .vma = vmf->vma,
175 .start = vmf->address,
176 .end = vmf->address + PAGE_SIZE,
177 .src = &src,
178 .dst = &dst,
179 };
180
181 /*
182 * FIXME what we really want is to find some heuristic to migrate more
183 * than just one page on CPU fault. When such fault happens it is very
184 * likely that more surrounding page will CPU fault too.
185 */
186 if (migrate_vma_setup(&args) < 0)
187 return VM_FAULT_SIGBUS;
188 if (!args.cpages)
189 return 0;
190
191 ret = nouveau_dmem_fault_copy_one(drm, vmf, &args, &dma_addr);
192 if (ret || dst == 0)
193 goto done;
194
195 nouveau_fence_new(dmem->migrate.chan, false, &fence);
196 migrate_vma_pages(&args);
197 nouveau_dmem_fence_done(&fence);
198 dma_unmap_page(drm->dev->dev, dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
199done:
200 migrate_vma_finalize(&args);
201 return ret;
202}
203
204static const struct dev_pagemap_ops nouveau_dmem_pagemap_ops = {
205 .page_free = nouveau_dmem_page_free,
206 .migrate_to_ram = nouveau_dmem_migrate_to_ram,
207};
208
209static int
210nouveau_dmem_chunk_alloc(struct nouveau_drm *drm)
211{
212 struct nouveau_dmem_chunk *chunk;
213 int ret;
214
215 if (drm->dmem == NULL)
216 return -EINVAL;
217
218 mutex_lock(&drm->dmem->mutex);
219 chunk = list_first_entry_or_null(&drm->dmem->chunk_empty,
220 struct nouveau_dmem_chunk,
221 list);
222 if (chunk == NULL) {
223 mutex_unlock(&drm->dmem->mutex);
224 return -ENOMEM;
225 }
226
227 list_del(&chunk->list);
228 mutex_unlock(&drm->dmem->mutex);
229
230 ret = nouveau_bo_new(&drm->client, DMEM_CHUNK_SIZE, 0,
231 TTM_PL_FLAG_VRAM, 0, 0, NULL, NULL,
232 &chunk->bo);
233 if (ret)
234 goto out;
235
236 ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false);
237 if (ret) {
238 nouveau_bo_ref(NULL, &chunk->bo);
239 goto out;
240 }
241
242 bitmap_zero(chunk->bitmap, DMEM_CHUNK_NPAGES);
243 spin_lock_init(&chunk->lock);
244
245out:
246 mutex_lock(&drm->dmem->mutex);
247 if (chunk->bo)
248 list_add(&chunk->list, &drm->dmem->chunk_empty);
249 else
250 list_add_tail(&chunk->list, &drm->dmem->chunk_empty);
251 mutex_unlock(&drm->dmem->mutex);
252
253 return ret;
254}
255
256static struct nouveau_dmem_chunk *
257nouveau_dmem_chunk_first_free_locked(struct nouveau_drm *drm)
258{
259 struct nouveau_dmem_chunk *chunk;
260
261 chunk = list_first_entry_or_null(&drm->dmem->chunk_free,
262 struct nouveau_dmem_chunk,
263 list);
264 if (chunk)
265 return chunk;
266
267 chunk = list_first_entry_or_null(&drm->dmem->chunk_empty,
268 struct nouveau_dmem_chunk,
269 list);
270 if (chunk->bo)
271 return chunk;
272
273 return NULL;
274}
275
276static int
277nouveau_dmem_pages_alloc(struct nouveau_drm *drm,
278 unsigned long npages,
279 unsigned long *pages)
280{
281 struct nouveau_dmem_chunk *chunk;
282 unsigned long c;
283 int ret;
284
285 memset(pages, 0xff, npages * sizeof(*pages));
286
287 mutex_lock(&drm->dmem->mutex);
288 for (c = 0; c < npages;) {
289 unsigned long i;
290
291 chunk = nouveau_dmem_chunk_first_free_locked(drm);
292 if (chunk == NULL) {
293 mutex_unlock(&drm->dmem->mutex);
294 ret = nouveau_dmem_chunk_alloc(drm);
295 if (ret) {
296 if (c)
297 return 0;
298 return ret;
299 }
300 mutex_lock(&drm->dmem->mutex);
301 continue;
302 }
303
304 spin_lock(&chunk->lock);
305 i = find_first_zero_bit(chunk->bitmap, DMEM_CHUNK_NPAGES);
306 while (i < DMEM_CHUNK_NPAGES && c < npages) {
307 pages[c] = chunk->pfn_first + i;
308 set_bit(i, chunk->bitmap);
309 chunk->callocated++;
310 c++;
311
312 i = find_next_zero_bit(chunk->bitmap,
313 DMEM_CHUNK_NPAGES, i);
314 }
315 spin_unlock(&chunk->lock);
316 }
317 mutex_unlock(&drm->dmem->mutex);
318
319 return 0;
320}
321
322static struct page *
323nouveau_dmem_page_alloc_locked(struct nouveau_drm *drm)
324{
325 unsigned long pfns[1];
326 struct page *page;
327 int ret;
328
329 /* FIXME stop all the miss-match API ... */
330 ret = nouveau_dmem_pages_alloc(drm, 1, pfns);
331 if (ret)
332 return NULL;
333
334 page = pfn_to_page(pfns[0]);
335 get_page(page);
336 lock_page(page);
337 return page;
338}
339
340static void
341nouveau_dmem_page_free_locked(struct nouveau_drm *drm, struct page *page)
342{
343 unlock_page(page);
344 put_page(page);
345}
346
347void
348nouveau_dmem_resume(struct nouveau_drm *drm)
349{
350 struct nouveau_dmem_chunk *chunk;
351 int ret;
352
353 if (drm->dmem == NULL)
354 return;
355
356 mutex_lock(&drm->dmem->mutex);
357 list_for_each_entry (chunk, &drm->dmem->chunk_free, list) {
358 ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false);
359 /* FIXME handle pin failure */
360 WARN_ON(ret);
361 }
362 list_for_each_entry (chunk, &drm->dmem->chunk_full, list) {
363 ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false);
364 /* FIXME handle pin failure */
365 WARN_ON(ret);
366 }
367 mutex_unlock(&drm->dmem->mutex);
368}
369
370void
371nouveau_dmem_suspend(struct nouveau_drm *drm)
372{
373 struct nouveau_dmem_chunk *chunk;
374
375 if (drm->dmem == NULL)
376 return;
377
378 mutex_lock(&drm->dmem->mutex);
379 list_for_each_entry (chunk, &drm->dmem->chunk_free, list) {
380 nouveau_bo_unpin(chunk->bo);
381 }
382 list_for_each_entry (chunk, &drm->dmem->chunk_full, list) {
383 nouveau_bo_unpin(chunk->bo);
384 }
385 mutex_unlock(&drm->dmem->mutex);
386}
387
388void
389nouveau_dmem_fini(struct nouveau_drm *drm)
390{
391 struct nouveau_dmem_chunk *chunk, *tmp;
392
393 if (drm->dmem == NULL)
394 return;
395
396 mutex_lock(&drm->dmem->mutex);
397
398 WARN_ON(!list_empty(&drm->dmem->chunk_free));
399 WARN_ON(!list_empty(&drm->dmem->chunk_full));
400
401 list_for_each_entry_safe (chunk, tmp, &drm->dmem->chunk_empty, list) {
402 if (chunk->bo) {
403 nouveau_bo_unpin(chunk->bo);
404 nouveau_bo_ref(NULL, &chunk->bo);
405 }
406 list_del(&chunk->list);
407 kfree(chunk);
408 }
409
410 mutex_unlock(&drm->dmem->mutex);
411}
412
413static int
414nvc0b5_migrate_copy(struct nouveau_drm *drm, u64 npages,
415 enum nouveau_aper dst_aper, u64 dst_addr,
416 enum nouveau_aper src_aper, u64 src_addr)
417{
418 struct nouveau_channel *chan = drm->dmem->migrate.chan;
419 u32 launch_dma = (1 << 9) /* MULTI_LINE_ENABLE. */ |
420 (1 << 8) /* DST_MEMORY_LAYOUT_PITCH. */ |
421 (1 << 7) /* SRC_MEMORY_LAYOUT_PITCH. */ |
422 (1 << 2) /* FLUSH_ENABLE_TRUE. */ |
423 (2 << 0) /* DATA_TRANSFER_TYPE_NON_PIPELINED. */;
424 int ret;
425
426 ret = RING_SPACE(chan, 13);
427 if (ret)
428 return ret;
429
430 if (src_aper != NOUVEAU_APER_VIRT) {
431 switch (src_aper) {
432 case NOUVEAU_APER_VRAM:
433 BEGIN_IMC0(chan, NvSubCopy, 0x0260, 0);
434 break;
435 case NOUVEAU_APER_HOST:
436 BEGIN_IMC0(chan, NvSubCopy, 0x0260, 1);
437 break;
438 default:
439 return -EINVAL;
440 }
441 launch_dma |= 0x00001000; /* SRC_TYPE_PHYSICAL. */
442 }
443
444 if (dst_aper != NOUVEAU_APER_VIRT) {
445 switch (dst_aper) {
446 case NOUVEAU_APER_VRAM:
447 BEGIN_IMC0(chan, NvSubCopy, 0x0264, 0);
448 break;
449 case NOUVEAU_APER_HOST:
450 BEGIN_IMC0(chan, NvSubCopy, 0x0264, 1);
451 break;
452 default:
453 return -EINVAL;
454 }
455 launch_dma |= 0x00002000; /* DST_TYPE_PHYSICAL. */
456 }
457
458 BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
459 OUT_RING (chan, upper_32_bits(src_addr));
460 OUT_RING (chan, lower_32_bits(src_addr));
461 OUT_RING (chan, upper_32_bits(dst_addr));
462 OUT_RING (chan, lower_32_bits(dst_addr));
463 OUT_RING (chan, PAGE_SIZE);
464 OUT_RING (chan, PAGE_SIZE);
465 OUT_RING (chan, PAGE_SIZE);
466 OUT_RING (chan, npages);
467 BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
468 OUT_RING (chan, launch_dma);
469 return 0;
470}
471
472static int
473nouveau_dmem_migrate_init(struct nouveau_drm *drm)
474{
475 switch (drm->ttm.copy.oclass) {
476 case PASCAL_DMA_COPY_A:
477 case PASCAL_DMA_COPY_B:
478 case VOLTA_DMA_COPY_A:
479 case TURING_DMA_COPY_A:
480 drm->dmem->migrate.copy_func = nvc0b5_migrate_copy;
481 drm->dmem->migrate.chan = drm->ttm.chan;
482 return 0;
483 default:
484 break;
485 }
486 return -ENODEV;
487}
488
489void
490nouveau_dmem_init(struct nouveau_drm *drm)
491{
492 struct device *device = drm->dev->dev;
493 struct resource *res;
494 unsigned long i, size, pfn_first;
495 int ret;
496
497 /* This only make sense on PASCAL or newer */
498 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_PASCAL)
499 return;
500
501 if (!(drm->dmem = kzalloc(sizeof(*drm->dmem), GFP_KERNEL)))
502 return;
503
504 drm->dmem->drm = drm;
505 mutex_init(&drm->dmem->mutex);
506 INIT_LIST_HEAD(&drm->dmem->chunk_free);
507 INIT_LIST_HEAD(&drm->dmem->chunk_full);
508 INIT_LIST_HEAD(&drm->dmem->chunk_empty);
509
510 size = ALIGN(drm->client.device.info.ram_user, DMEM_CHUNK_SIZE);
511
512 /* Initialize migration dma helpers before registering memory */
513 ret = nouveau_dmem_migrate_init(drm);
514 if (ret)
515 goto out_free;
516
517 /*
518 * FIXME we need some kind of policy to decide how much VRAM we
519 * want to register with HMM. For now just register everything
520 * and latter if we want to do thing like over commit then we
521 * could revisit this.
522 */
523 res = devm_request_free_mem_region(device, &iomem_resource, size);
524 if (IS_ERR(res))
525 goto out_free;
526 drm->dmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
527 drm->dmem->pagemap.res = *res;
528 drm->dmem->pagemap.ops = &nouveau_dmem_pagemap_ops;
529 if (IS_ERR(devm_memremap_pages(device, &drm->dmem->pagemap)))
530 goto out_free;
531
532 pfn_first = res->start >> PAGE_SHIFT;
533 for (i = 0; i < (size / DMEM_CHUNK_SIZE); ++i) {
534 struct nouveau_dmem_chunk *chunk;
535 struct page *page;
536 unsigned long j;
537
538 chunk = kzalloc(sizeof(*chunk), GFP_KERNEL);
539 if (chunk == NULL) {
540 nouveau_dmem_fini(drm);
541 return;
542 }
543
544 chunk->drm = drm;
545 chunk->pfn_first = pfn_first + (i * DMEM_CHUNK_NPAGES);
546 list_add_tail(&chunk->list, &drm->dmem->chunk_empty);
547
548 page = pfn_to_page(chunk->pfn_first);
549 for (j = 0; j < DMEM_CHUNK_NPAGES; ++j, ++page)
550 page->zone_device_data = chunk;
551 }
552
553 NV_INFO(drm, "DMEM: registered %ldMB of device memory\n", size >> 20);
554 return;
555out_free:
556 kfree(drm->dmem);
557 drm->dmem = NULL;
558}
559
560static unsigned long nouveau_dmem_migrate_copy_one(struct nouveau_drm *drm,
561 unsigned long src, dma_addr_t *dma_addr)
562{
563 struct device *dev = drm->dev->dev;
564 struct page *dpage, *spage;
565
566 spage = migrate_pfn_to_page(src);
567 if (!spage || !(src & MIGRATE_PFN_MIGRATE))
568 goto out;
569
570 dpage = nouveau_dmem_page_alloc_locked(drm);
571 if (!dpage)
572 return 0;
573
574 *dma_addr = dma_map_page(dev, spage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
575 if (dma_mapping_error(dev, *dma_addr))
576 goto out_free_page;
577
578 if (drm->dmem->migrate.copy_func(drm, 1, NOUVEAU_APER_VRAM,
579 nouveau_dmem_page_addr(dpage), NOUVEAU_APER_HOST,
580 *dma_addr))
581 goto out_dma_unmap;
582
583 return migrate_pfn(page_to_pfn(dpage)) | MIGRATE_PFN_LOCKED;
584
585out_dma_unmap:
586 dma_unmap_page(dev, *dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
587out_free_page:
588 nouveau_dmem_page_free_locked(drm, dpage);
589out:
590 return 0;
591}
592
593static void nouveau_dmem_migrate_chunk(struct nouveau_drm *drm,
594 struct migrate_vma *args, dma_addr_t *dma_addrs)
595{
596 struct nouveau_fence *fence;
597 unsigned long addr = args->start, nr_dma = 0, i;
598
599 for (i = 0; addr < args->end; i++) {
600 args->dst[i] = nouveau_dmem_migrate_copy_one(drm, args->src[i],
601 dma_addrs + nr_dma);
602 if (args->dst[i])
603 nr_dma++;
604 addr += PAGE_SIZE;
605 }
606
607 nouveau_fence_new(drm->dmem->migrate.chan, false, &fence);
608 migrate_vma_pages(args);
609 nouveau_dmem_fence_done(&fence);
610
611 while (nr_dma--) {
612 dma_unmap_page(drm->dev->dev, dma_addrs[nr_dma], PAGE_SIZE,
613 DMA_BIDIRECTIONAL);
614 }
615 /*
616 * FIXME optimization: update GPU page table to point to newly migrated
617 * memory.
618 */
619 migrate_vma_finalize(args);
620}
621
622int
623nouveau_dmem_migrate_vma(struct nouveau_drm *drm,
624 struct vm_area_struct *vma,
625 unsigned long start,
626 unsigned long end)
627{
628 unsigned long npages = (end - start) >> PAGE_SHIFT;
629 unsigned long max = min(SG_MAX_SINGLE_ALLOC, npages);
630 dma_addr_t *dma_addrs;
631 struct migrate_vma args = {
632 .vma = vma,
633 .start = start,
634 };
635 unsigned long c, i;
636 int ret = -ENOMEM;
637
638 args.src = kcalloc(max, sizeof(args.src), GFP_KERNEL);
639 if (!args.src)
640 goto out;
641 args.dst = kcalloc(max, sizeof(args.dst), GFP_KERNEL);
642 if (!args.dst)
643 goto out_free_src;
644
645 dma_addrs = kmalloc_array(max, sizeof(*dma_addrs), GFP_KERNEL);
646 if (!dma_addrs)
647 goto out_free_dst;
648
649 for (i = 0; i < npages; i += c) {
650 c = min(SG_MAX_SINGLE_ALLOC, npages);
651 args.end = start + (c << PAGE_SHIFT);
652 ret = migrate_vma_setup(&args);
653 if (ret)
654 goto out_free_dma;
655
656 if (args.cpages)
657 nouveau_dmem_migrate_chunk(drm, &args, dma_addrs);
658 args.start = args.end;
659 }
660
661 ret = 0;
662out_free_dma:
663 kfree(dma_addrs);
664out_free_dst:
665 kfree(args.dst);
666out_free_src:
667 kfree(args.src);
668out:
669 return ret;
670}
671
672static inline bool
673nouveau_dmem_page(struct nouveau_drm *drm, struct page *page)
674{
675 return is_device_private_page(page) && drm->dmem == page_to_dmem(page);
676}
677
678void
679nouveau_dmem_convert_pfn(struct nouveau_drm *drm,
680 struct hmm_range *range)
681{
682 unsigned long i, npages;
683
684 npages = (range->end - range->start) >> PAGE_SHIFT;
685 for (i = 0; i < npages; ++i) {
686 struct page *page;
687 uint64_t addr;
688
689 page = hmm_device_entry_to_page(range, range->pfns[i]);
690 if (page == NULL)
691 continue;
692
693 if (!(range->pfns[i] & range->flags[HMM_PFN_DEVICE_PRIVATE])) {
694 continue;
695 }
696
697 if (!nouveau_dmem_page(drm, page)) {
698 WARN(1, "Some unknown device memory !\n");
699 range->pfns[i] = 0;
700 continue;
701 }
702
703 addr = nouveau_dmem_page_addr(page);
704 range->pfns[i] &= ((1UL << range->pfn_shift) - 1);
705 range->pfns[i] |= (addr >> PAGE_SHIFT) << range->pfn_shift;
706 }
707}
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
37#include <nvhw/class/cla0b5.h>
38
39#include <linux/sched/mm.h>
40#include <linux/hmm.h>
41#include <linux/memremap.h>
42#include <linux/migrate.h>
43
44/*
45 * FIXME: this is ugly right now we are using TTM to allocate vram and we pin
46 * it in vram while in use. We likely want to overhaul memory management for
47 * nouveau to be more page like (not necessarily with system page size but a
48 * bigger page size) at lowest level and have some shim layer on top that would
49 * provide the same functionality as TTM.
50 */
51#define DMEM_CHUNK_SIZE (2UL << 20)
52#define DMEM_CHUNK_NPAGES (DMEM_CHUNK_SIZE >> PAGE_SHIFT)
53
54enum nouveau_aper {
55 NOUVEAU_APER_VIRT,
56 NOUVEAU_APER_VRAM,
57 NOUVEAU_APER_HOST,
58};
59
60typedef int (*nouveau_migrate_copy_t)(struct nouveau_drm *drm, u64 npages,
61 enum nouveau_aper, u64 dst_addr,
62 enum nouveau_aper, u64 src_addr);
63typedef int (*nouveau_clear_page_t)(struct nouveau_drm *drm, u32 length,
64 enum nouveau_aper, u64 dst_addr);
65
66struct nouveau_dmem_chunk {
67 struct list_head list;
68 struct nouveau_bo *bo;
69 struct nouveau_drm *drm;
70 unsigned long callocated;
71 struct dev_pagemap pagemap;
72};
73
74struct nouveau_dmem_migrate {
75 nouveau_migrate_copy_t copy_func;
76 nouveau_clear_page_t clear_func;
77 struct nouveau_channel *chan;
78};
79
80struct nouveau_dmem {
81 struct nouveau_drm *drm;
82 struct nouveau_dmem_migrate migrate;
83 struct list_head chunks;
84 struct mutex mutex;
85 struct page *free_pages;
86 spinlock_t lock;
87};
88
89static struct nouveau_dmem_chunk *nouveau_page_to_chunk(struct page *page)
90{
91 return container_of(page->pgmap, struct nouveau_dmem_chunk, pagemap);
92}
93
94static struct nouveau_drm *page_to_drm(struct page *page)
95{
96 struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
97
98 return chunk->drm;
99}
100
101unsigned long nouveau_dmem_page_addr(struct page *page)
102{
103 struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
104 unsigned long off = (page_to_pfn(page) << PAGE_SHIFT) -
105 chunk->pagemap.range.start;
106
107 return chunk->bo->offset + off;
108}
109
110static void nouveau_dmem_page_free(struct page *page)
111{
112 struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
113 struct nouveau_dmem *dmem = chunk->drm->dmem;
114
115 spin_lock(&dmem->lock);
116 page->zone_device_data = dmem->free_pages;
117 dmem->free_pages = page;
118
119 WARN_ON(!chunk->callocated);
120 chunk->callocated--;
121 /*
122 * FIXME when chunk->callocated reach 0 we should add the chunk to
123 * a reclaim list so that it can be freed in case of memory pressure.
124 */
125 spin_unlock(&dmem->lock);
126}
127
128static void nouveau_dmem_fence_done(struct nouveau_fence **fence)
129{
130 if (fence) {
131 nouveau_fence_wait(*fence, true, false);
132 nouveau_fence_unref(fence);
133 } else {
134 /*
135 * FIXME wait for channel to be IDLE before calling finalizing
136 * the hmem object.
137 */
138 }
139}
140
141static int nouveau_dmem_copy_one(struct nouveau_drm *drm, struct page *spage,
142 struct page *dpage, dma_addr_t *dma_addr)
143{
144 struct device *dev = drm->dev->dev;
145
146 lock_page(dpage);
147
148 *dma_addr = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
149 if (dma_mapping_error(dev, *dma_addr))
150 return -EIO;
151
152 if (drm->dmem->migrate.copy_func(drm, 1, NOUVEAU_APER_HOST, *dma_addr,
153 NOUVEAU_APER_VRAM, nouveau_dmem_page_addr(spage))) {
154 dma_unmap_page(dev, *dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
155 return -EIO;
156 }
157
158 return 0;
159}
160
161static vm_fault_t nouveau_dmem_migrate_to_ram(struct vm_fault *vmf)
162{
163 struct nouveau_drm *drm = page_to_drm(vmf->page);
164 struct nouveau_dmem *dmem = drm->dmem;
165 struct nouveau_fence *fence;
166 struct nouveau_svmm *svmm;
167 struct page *spage, *dpage;
168 unsigned long src = 0, dst = 0;
169 dma_addr_t dma_addr = 0;
170 vm_fault_t ret = 0;
171 struct migrate_vma args = {
172 .vma = vmf->vma,
173 .start = vmf->address,
174 .end = vmf->address + PAGE_SIZE,
175 .src = &src,
176 .dst = &dst,
177 .pgmap_owner = drm->dev,
178 .fault_page = vmf->page,
179 .flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE,
180 };
181
182 /*
183 * FIXME what we really want is to find some heuristic to migrate more
184 * than just one page on CPU fault. When such fault happens it is very
185 * likely that more surrounding page will CPU fault too.
186 */
187 if (migrate_vma_setup(&args) < 0)
188 return VM_FAULT_SIGBUS;
189 if (!args.cpages)
190 return 0;
191
192 spage = migrate_pfn_to_page(src);
193 if (!spage || !(src & MIGRATE_PFN_MIGRATE))
194 goto done;
195
196 dpage = alloc_page_vma(GFP_HIGHUSER, vmf->vma, vmf->address);
197 if (!dpage)
198 goto done;
199
200 dst = migrate_pfn(page_to_pfn(dpage));
201
202 svmm = spage->zone_device_data;
203 mutex_lock(&svmm->mutex);
204 nouveau_svmm_invalidate(svmm, args.start, args.end);
205 ret = nouveau_dmem_copy_one(drm, spage, dpage, &dma_addr);
206 mutex_unlock(&svmm->mutex);
207 if (ret) {
208 ret = VM_FAULT_SIGBUS;
209 goto done;
210 }
211
212 nouveau_fence_new(&fence, dmem->migrate.chan);
213 migrate_vma_pages(&args);
214 nouveau_dmem_fence_done(&fence);
215 dma_unmap_page(drm->dev->dev, dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
216done:
217 migrate_vma_finalize(&args);
218 return ret;
219}
220
221static 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
226static int
227nouveau_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
294out_bo_unpin:
295 nouveau_bo_unpin(chunk->bo);
296out_bo_free:
297 nouveau_bo_ref(NULL, &chunk->bo);
298out_release:
299 release_mem_region(chunk->pagemap.range.start, range_len(&chunk->pagemap.range));
300out_free:
301 kfree(chunk);
302out:
303 return ret;
304}
305
306static struct page *
307nouveau_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 zone_device_page_init(page);
328 return page;
329}
330
331static void
332nouveau_dmem_page_free_locked(struct nouveau_drm *drm, struct page *page)
333{
334 unlock_page(page);
335 put_page(page);
336}
337
338void
339nouveau_dmem_resume(struct nouveau_drm *drm)
340{
341 struct nouveau_dmem_chunk *chunk;
342 int ret;
343
344 if (drm->dmem == NULL)
345 return;
346
347 mutex_lock(&drm->dmem->mutex);
348 list_for_each_entry(chunk, &drm->dmem->chunks, list) {
349 ret = nouveau_bo_pin(chunk->bo, NOUVEAU_GEM_DOMAIN_VRAM, false);
350 /* FIXME handle pin failure */
351 WARN_ON(ret);
352 }
353 mutex_unlock(&drm->dmem->mutex);
354}
355
356void
357nouveau_dmem_suspend(struct nouveau_drm *drm)
358{
359 struct nouveau_dmem_chunk *chunk;
360
361 if (drm->dmem == NULL)
362 return;
363
364 mutex_lock(&drm->dmem->mutex);
365 list_for_each_entry(chunk, &drm->dmem->chunks, list)
366 nouveau_bo_unpin(chunk->bo);
367 mutex_unlock(&drm->dmem->mutex);
368}
369
370/*
371 * Evict all pages mapping a chunk.
372 */
373static void
374nouveau_dmem_evict_chunk(struct nouveau_dmem_chunk *chunk)
375{
376 unsigned long i, npages = range_len(&chunk->pagemap.range) >> PAGE_SHIFT;
377 unsigned long *src_pfns, *dst_pfns;
378 dma_addr_t *dma_addrs;
379 struct nouveau_fence *fence;
380
381 src_pfns = kcalloc(npages, sizeof(*src_pfns), GFP_KERNEL);
382 dst_pfns = kcalloc(npages, sizeof(*dst_pfns), GFP_KERNEL);
383 dma_addrs = kcalloc(npages, sizeof(*dma_addrs), GFP_KERNEL);
384
385 migrate_device_range(src_pfns, chunk->pagemap.range.start >> PAGE_SHIFT,
386 npages);
387
388 for (i = 0; i < npages; i++) {
389 if (src_pfns[i] & MIGRATE_PFN_MIGRATE) {
390 struct page *dpage;
391
392 /*
393 * _GFP_NOFAIL because the GPU is going away and there
394 * is nothing sensible we can do if we can't copy the
395 * data back.
396 */
397 dpage = alloc_page(GFP_HIGHUSER | __GFP_NOFAIL);
398 dst_pfns[i] = migrate_pfn(page_to_pfn(dpage));
399 nouveau_dmem_copy_one(chunk->drm,
400 migrate_pfn_to_page(src_pfns[i]), dpage,
401 &dma_addrs[i]);
402 }
403 }
404
405 nouveau_fence_new(&fence, chunk->drm->dmem->migrate.chan);
406 migrate_device_pages(src_pfns, dst_pfns, npages);
407 nouveau_dmem_fence_done(&fence);
408 migrate_device_finalize(src_pfns, dst_pfns, npages);
409 kfree(src_pfns);
410 kfree(dst_pfns);
411 for (i = 0; i < npages; i++)
412 dma_unmap_page(chunk->drm->dev->dev, dma_addrs[i], PAGE_SIZE, DMA_BIDIRECTIONAL);
413 kfree(dma_addrs);
414}
415
416void
417nouveau_dmem_fini(struct nouveau_drm *drm)
418{
419 struct nouveau_dmem_chunk *chunk, *tmp;
420
421 if (drm->dmem == NULL)
422 return;
423
424 mutex_lock(&drm->dmem->mutex);
425
426 list_for_each_entry_safe(chunk, tmp, &drm->dmem->chunks, list) {
427 nouveau_dmem_evict_chunk(chunk);
428 nouveau_bo_unpin(chunk->bo);
429 nouveau_bo_ref(NULL, &chunk->bo);
430 WARN_ON(chunk->callocated);
431 list_del(&chunk->list);
432 memunmap_pages(&chunk->pagemap);
433 release_mem_region(chunk->pagemap.range.start,
434 range_len(&chunk->pagemap.range));
435 kfree(chunk);
436 }
437
438 mutex_unlock(&drm->dmem->mutex);
439}
440
441static int
442nvc0b5_migrate_copy(struct nouveau_drm *drm, u64 npages,
443 enum nouveau_aper dst_aper, u64 dst_addr,
444 enum nouveau_aper src_aper, u64 src_addr)
445{
446 struct nvif_push *push = drm->dmem->migrate.chan->chan.push;
447 u32 launch_dma = 0;
448 int ret;
449
450 ret = PUSH_WAIT(push, 13);
451 if (ret)
452 return ret;
453
454 if (src_aper != NOUVEAU_APER_VIRT) {
455 switch (src_aper) {
456 case NOUVEAU_APER_VRAM:
457 PUSH_IMMD(push, NVA0B5, SET_SRC_PHYS_MODE,
458 NVDEF(NVA0B5, SET_SRC_PHYS_MODE, TARGET, LOCAL_FB));
459 break;
460 case NOUVEAU_APER_HOST:
461 PUSH_IMMD(push, NVA0B5, SET_SRC_PHYS_MODE,
462 NVDEF(NVA0B5, SET_SRC_PHYS_MODE, TARGET, COHERENT_SYSMEM));
463 break;
464 default:
465 return -EINVAL;
466 }
467
468 launch_dma |= NVDEF(NVA0B5, LAUNCH_DMA, SRC_TYPE, PHYSICAL);
469 }
470
471 if (dst_aper != NOUVEAU_APER_VIRT) {
472 switch (dst_aper) {
473 case NOUVEAU_APER_VRAM:
474 PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
475 NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, LOCAL_FB));
476 break;
477 case NOUVEAU_APER_HOST:
478 PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
479 NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, COHERENT_SYSMEM));
480 break;
481 default:
482 return -EINVAL;
483 }
484
485 launch_dma |= NVDEF(NVA0B5, LAUNCH_DMA, DST_TYPE, PHYSICAL);
486 }
487
488 PUSH_MTHD(push, NVA0B5, OFFSET_IN_UPPER,
489 NVVAL(NVA0B5, OFFSET_IN_UPPER, UPPER, upper_32_bits(src_addr)),
490
491 OFFSET_IN_LOWER, lower_32_bits(src_addr),
492
493 OFFSET_OUT_UPPER,
494 NVVAL(NVA0B5, OFFSET_OUT_UPPER, UPPER, upper_32_bits(dst_addr)),
495
496 OFFSET_OUT_LOWER, lower_32_bits(dst_addr),
497 PITCH_IN, PAGE_SIZE,
498 PITCH_OUT, PAGE_SIZE,
499 LINE_LENGTH_IN, PAGE_SIZE,
500 LINE_COUNT, npages);
501
502 PUSH_MTHD(push, NVA0B5, LAUNCH_DMA, launch_dma |
503 NVDEF(NVA0B5, LAUNCH_DMA, DATA_TRANSFER_TYPE, NON_PIPELINED) |
504 NVDEF(NVA0B5, LAUNCH_DMA, FLUSH_ENABLE, TRUE) |
505 NVDEF(NVA0B5, LAUNCH_DMA, SEMAPHORE_TYPE, NONE) |
506 NVDEF(NVA0B5, LAUNCH_DMA, INTERRUPT_TYPE, NONE) |
507 NVDEF(NVA0B5, LAUNCH_DMA, SRC_MEMORY_LAYOUT, PITCH) |
508 NVDEF(NVA0B5, LAUNCH_DMA, DST_MEMORY_LAYOUT, PITCH) |
509 NVDEF(NVA0B5, LAUNCH_DMA, MULTI_LINE_ENABLE, TRUE) |
510 NVDEF(NVA0B5, LAUNCH_DMA, REMAP_ENABLE, FALSE) |
511 NVDEF(NVA0B5, LAUNCH_DMA, BYPASS_L2, USE_PTE_SETTING));
512 return 0;
513}
514
515static int
516nvc0b5_migrate_clear(struct nouveau_drm *drm, u32 length,
517 enum nouveau_aper dst_aper, u64 dst_addr)
518{
519 struct nvif_push *push = drm->dmem->migrate.chan->chan.push;
520 u32 launch_dma = 0;
521 int ret;
522
523 ret = PUSH_WAIT(push, 12);
524 if (ret)
525 return ret;
526
527 switch (dst_aper) {
528 case NOUVEAU_APER_VRAM:
529 PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
530 NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, LOCAL_FB));
531 break;
532 case NOUVEAU_APER_HOST:
533 PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
534 NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, COHERENT_SYSMEM));
535 break;
536 default:
537 return -EINVAL;
538 }
539
540 launch_dma |= NVDEF(NVA0B5, LAUNCH_DMA, DST_TYPE, PHYSICAL);
541
542 PUSH_MTHD(push, NVA0B5, SET_REMAP_CONST_A, 0,
543 SET_REMAP_CONST_B, 0,
544
545 SET_REMAP_COMPONENTS,
546 NVDEF(NVA0B5, SET_REMAP_COMPONENTS, DST_X, CONST_A) |
547 NVDEF(NVA0B5, SET_REMAP_COMPONENTS, DST_Y, CONST_B) |
548 NVDEF(NVA0B5, SET_REMAP_COMPONENTS, COMPONENT_SIZE, FOUR) |
549 NVDEF(NVA0B5, SET_REMAP_COMPONENTS, NUM_DST_COMPONENTS, TWO));
550
551 PUSH_MTHD(push, NVA0B5, OFFSET_OUT_UPPER,
552 NVVAL(NVA0B5, OFFSET_OUT_UPPER, UPPER, upper_32_bits(dst_addr)),
553
554 OFFSET_OUT_LOWER, lower_32_bits(dst_addr));
555
556 PUSH_MTHD(push, NVA0B5, LINE_LENGTH_IN, length >> 3);
557
558 PUSH_MTHD(push, NVA0B5, LAUNCH_DMA, launch_dma |
559 NVDEF(NVA0B5, LAUNCH_DMA, DATA_TRANSFER_TYPE, NON_PIPELINED) |
560 NVDEF(NVA0B5, LAUNCH_DMA, FLUSH_ENABLE, TRUE) |
561 NVDEF(NVA0B5, LAUNCH_DMA, SEMAPHORE_TYPE, NONE) |
562 NVDEF(NVA0B5, LAUNCH_DMA, INTERRUPT_TYPE, NONE) |
563 NVDEF(NVA0B5, LAUNCH_DMA, SRC_MEMORY_LAYOUT, PITCH) |
564 NVDEF(NVA0B5, LAUNCH_DMA, DST_MEMORY_LAYOUT, PITCH) |
565 NVDEF(NVA0B5, LAUNCH_DMA, MULTI_LINE_ENABLE, FALSE) |
566 NVDEF(NVA0B5, LAUNCH_DMA, REMAP_ENABLE, TRUE) |
567 NVDEF(NVA0B5, LAUNCH_DMA, BYPASS_L2, USE_PTE_SETTING));
568 return 0;
569}
570
571static int
572nouveau_dmem_migrate_init(struct nouveau_drm *drm)
573{
574 switch (drm->ttm.copy.oclass) {
575 case PASCAL_DMA_COPY_A:
576 case PASCAL_DMA_COPY_B:
577 case VOLTA_DMA_COPY_A:
578 case TURING_DMA_COPY_A:
579 drm->dmem->migrate.copy_func = nvc0b5_migrate_copy;
580 drm->dmem->migrate.clear_func = nvc0b5_migrate_clear;
581 drm->dmem->migrate.chan = drm->ttm.chan;
582 return 0;
583 default:
584 break;
585 }
586 return -ENODEV;
587}
588
589void
590nouveau_dmem_init(struct nouveau_drm *drm)
591{
592 int ret;
593
594 /* This only make sense on PASCAL or newer */
595 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_PASCAL)
596 return;
597
598 if (!(drm->dmem = kzalloc(sizeof(*drm->dmem), GFP_KERNEL)))
599 return;
600
601 drm->dmem->drm = drm;
602 mutex_init(&drm->dmem->mutex);
603 INIT_LIST_HEAD(&drm->dmem->chunks);
604 mutex_init(&drm->dmem->mutex);
605 spin_lock_init(&drm->dmem->lock);
606
607 /* Initialize migration dma helpers before registering memory */
608 ret = nouveau_dmem_migrate_init(drm);
609 if (ret) {
610 kfree(drm->dmem);
611 drm->dmem = NULL;
612 }
613}
614
615static unsigned long nouveau_dmem_migrate_copy_one(struct nouveau_drm *drm,
616 struct nouveau_svmm *svmm, unsigned long src,
617 dma_addr_t *dma_addr, u64 *pfn)
618{
619 struct device *dev = drm->dev->dev;
620 struct page *dpage, *spage;
621 unsigned long paddr;
622
623 spage = migrate_pfn_to_page(src);
624 if (!(src & MIGRATE_PFN_MIGRATE))
625 goto out;
626
627 dpage = nouveau_dmem_page_alloc_locked(drm);
628 if (!dpage)
629 goto out;
630
631 paddr = nouveau_dmem_page_addr(dpage);
632 if (spage) {
633 *dma_addr = dma_map_page(dev, spage, 0, page_size(spage),
634 DMA_BIDIRECTIONAL);
635 if (dma_mapping_error(dev, *dma_addr))
636 goto out_free_page;
637 if (drm->dmem->migrate.copy_func(drm, 1,
638 NOUVEAU_APER_VRAM, paddr, NOUVEAU_APER_HOST, *dma_addr))
639 goto out_dma_unmap;
640 } else {
641 *dma_addr = DMA_MAPPING_ERROR;
642 if (drm->dmem->migrate.clear_func(drm, page_size(dpage),
643 NOUVEAU_APER_VRAM, paddr))
644 goto out_free_page;
645 }
646
647 dpage->zone_device_data = svmm;
648 *pfn = NVIF_VMM_PFNMAP_V0_V | NVIF_VMM_PFNMAP_V0_VRAM |
649 ((paddr >> PAGE_SHIFT) << NVIF_VMM_PFNMAP_V0_ADDR_SHIFT);
650 if (src & MIGRATE_PFN_WRITE)
651 *pfn |= NVIF_VMM_PFNMAP_V0_W;
652 return migrate_pfn(page_to_pfn(dpage));
653
654out_dma_unmap:
655 dma_unmap_page(dev, *dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
656out_free_page:
657 nouveau_dmem_page_free_locked(drm, dpage);
658out:
659 *pfn = NVIF_VMM_PFNMAP_V0_NONE;
660 return 0;
661}
662
663static void nouveau_dmem_migrate_chunk(struct nouveau_drm *drm,
664 struct nouveau_svmm *svmm, struct migrate_vma *args,
665 dma_addr_t *dma_addrs, u64 *pfns)
666{
667 struct nouveau_fence *fence;
668 unsigned long addr = args->start, nr_dma = 0, i;
669
670 for (i = 0; addr < args->end; i++) {
671 args->dst[i] = nouveau_dmem_migrate_copy_one(drm, svmm,
672 args->src[i], dma_addrs + nr_dma, pfns + i);
673 if (!dma_mapping_error(drm->dev->dev, dma_addrs[nr_dma]))
674 nr_dma++;
675 addr += PAGE_SIZE;
676 }
677
678 nouveau_fence_new(&fence, drm->dmem->migrate.chan);
679 migrate_vma_pages(args);
680 nouveau_dmem_fence_done(&fence);
681 nouveau_pfns_map(svmm, args->vma->vm_mm, args->start, pfns, i);
682
683 while (nr_dma--) {
684 dma_unmap_page(drm->dev->dev, dma_addrs[nr_dma], PAGE_SIZE,
685 DMA_BIDIRECTIONAL);
686 }
687 migrate_vma_finalize(args);
688}
689
690int
691nouveau_dmem_migrate_vma(struct nouveau_drm *drm,
692 struct nouveau_svmm *svmm,
693 struct vm_area_struct *vma,
694 unsigned long start,
695 unsigned long end)
696{
697 unsigned long npages = (end - start) >> PAGE_SHIFT;
698 unsigned long max = min(SG_MAX_SINGLE_ALLOC, npages);
699 dma_addr_t *dma_addrs;
700 struct migrate_vma args = {
701 .vma = vma,
702 .start = start,
703 .pgmap_owner = drm->dev,
704 .flags = MIGRATE_VMA_SELECT_SYSTEM,
705 };
706 unsigned long i;
707 u64 *pfns;
708 int ret = -ENOMEM;
709
710 if (drm->dmem == NULL)
711 return -ENODEV;
712
713 args.src = kcalloc(max, sizeof(*args.src), GFP_KERNEL);
714 if (!args.src)
715 goto out;
716 args.dst = kcalloc(max, sizeof(*args.dst), GFP_KERNEL);
717 if (!args.dst)
718 goto out_free_src;
719
720 dma_addrs = kmalloc_array(max, sizeof(*dma_addrs), GFP_KERNEL);
721 if (!dma_addrs)
722 goto out_free_dst;
723
724 pfns = nouveau_pfns_alloc(max);
725 if (!pfns)
726 goto out_free_dma;
727
728 for (i = 0; i < npages; i += max) {
729 if (args.start + (max << PAGE_SHIFT) > end)
730 args.end = end;
731 else
732 args.end = args.start + (max << PAGE_SHIFT);
733
734 ret = migrate_vma_setup(&args);
735 if (ret)
736 goto out_free_pfns;
737
738 if (args.cpages)
739 nouveau_dmem_migrate_chunk(drm, svmm, &args, dma_addrs,
740 pfns);
741 args.start = args.end;
742 }
743
744 ret = 0;
745out_free_pfns:
746 nouveau_pfns_free(pfns);
747out_free_dma:
748 kfree(dma_addrs);
749out_free_dst:
750 kfree(args.dst);
751out_free_src:
752 kfree(args.src);
753out:
754 return ret;
755}