xref: /linux/drivers/gpu/drm/nouveau/nouveau_dmem.c (revision 95298d63c67673c654c08952672d016212b26054)
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/if000c.h>
33 #include <nvif/if500b.h>
34 #include <nvif/if900b.h>
35 #include <nvif/if000c.h>
36 
37 #include <linux/sched/mm.h>
38 #include <linux/hmm.h>
39 
40 /*
41  * FIXME: this is ugly right now we are using TTM to allocate vram and we pin
42  * it in vram while in use. We likely want to overhaul memory management for
43  * nouveau to be more page like (not necessarily with system page size but a
44  * bigger page size) at lowest level and have some shim layer on top that would
45  * provide the same functionality as TTM.
46  */
47 #define DMEM_CHUNK_SIZE (2UL << 20)
48 #define DMEM_CHUNK_NPAGES (DMEM_CHUNK_SIZE >> PAGE_SHIFT)
49 
50 enum nouveau_aper {
51 	NOUVEAU_APER_VIRT,
52 	NOUVEAU_APER_VRAM,
53 	NOUVEAU_APER_HOST,
54 };
55 
56 typedef int (*nouveau_migrate_copy_t)(struct nouveau_drm *drm, u64 npages,
57 				      enum nouveau_aper, u64 dst_addr,
58 				      enum nouveau_aper, u64 src_addr);
59 typedef int (*nouveau_clear_page_t)(struct nouveau_drm *drm, u32 length,
60 				      enum nouveau_aper, u64 dst_addr);
61 
62 struct nouveau_dmem_chunk {
63 	struct list_head list;
64 	struct nouveau_bo *bo;
65 	struct nouveau_drm *drm;
66 	unsigned long callocated;
67 	struct dev_pagemap pagemap;
68 };
69 
70 struct nouveau_dmem_migrate {
71 	nouveau_migrate_copy_t copy_func;
72 	nouveau_clear_page_t clear_func;
73 	struct nouveau_channel *chan;
74 };
75 
76 struct nouveau_dmem {
77 	struct nouveau_drm *drm;
78 	struct nouveau_dmem_migrate migrate;
79 	struct list_head chunks;
80 	struct mutex mutex;
81 	struct page *free_pages;
82 	spinlock_t lock;
83 };
84 
85 static struct nouveau_dmem_chunk *nouveau_page_to_chunk(struct page *page)
86 {
87 	return container_of(page->pgmap, struct nouveau_dmem_chunk, pagemap);
88 }
89 
90 static struct nouveau_drm *page_to_drm(struct page *page)
91 {
92 	struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
93 
94 	return chunk->drm;
95 }
96 
97 unsigned long nouveau_dmem_page_addr(struct page *page)
98 {
99 	struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
100 	unsigned long off = (page_to_pfn(page) << PAGE_SHIFT) -
101 				chunk->pagemap.res.start;
102 
103 	return chunk->bo->bo.offset + off;
104 }
105 
106 static void nouveau_dmem_page_free(struct page *page)
107 {
108 	struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
109 	struct nouveau_dmem *dmem = chunk->drm->dmem;
110 
111 	spin_lock(&dmem->lock);
112 	page->zone_device_data = dmem->free_pages;
113 	dmem->free_pages = page;
114 
115 	WARN_ON(!chunk->callocated);
116 	chunk->callocated--;
117 	/*
118 	 * FIXME when chunk->callocated reach 0 we should add the chunk to
119 	 * a reclaim list so that it can be freed in case of memory pressure.
120 	 */
121 	spin_unlock(&dmem->lock);
122 }
123 
124 static void nouveau_dmem_fence_done(struct nouveau_fence **fence)
125 {
126 	if (fence) {
127 		nouveau_fence_wait(*fence, true, false);
128 		nouveau_fence_unref(fence);
129 	} else {
130 		/*
131 		 * FIXME wait for channel to be IDLE before calling finalizing
132 		 * the hmem object.
133 		 */
134 	}
135 }
136 
137 static vm_fault_t nouveau_dmem_fault_copy_one(struct nouveau_drm *drm,
138 		struct vm_fault *vmf, struct migrate_vma *args,
139 		dma_addr_t *dma_addr)
140 {
141 	struct device *dev = drm->dev->dev;
142 	struct page *dpage, *spage;
143 
144 	spage = migrate_pfn_to_page(args->src[0]);
145 	if (!spage || !(args->src[0] & MIGRATE_PFN_MIGRATE))
146 		return 0;
147 
148 	dpage = alloc_page_vma(GFP_HIGHUSER, vmf->vma, vmf->address);
149 	if (!dpage)
150 		return VM_FAULT_SIGBUS;
151 	lock_page(dpage);
152 
153 	*dma_addr = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
154 	if (dma_mapping_error(dev, *dma_addr))
155 		goto error_free_page;
156 
157 	if (drm->dmem->migrate.copy_func(drm, 1, NOUVEAU_APER_HOST, *dma_addr,
158 			NOUVEAU_APER_VRAM, nouveau_dmem_page_addr(spage)))
159 		goto error_dma_unmap;
160 
161 	args->dst[0] = migrate_pfn(page_to_pfn(dpage)) | MIGRATE_PFN_LOCKED;
162 	return 0;
163 
164 error_dma_unmap:
165 	dma_unmap_page(dev, *dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
166 error_free_page:
167 	__free_page(dpage);
168 	return VM_FAULT_SIGBUS;
169 }
170 
171 static vm_fault_t nouveau_dmem_migrate_to_ram(struct vm_fault *vmf)
172 {
173 	struct nouveau_drm *drm = page_to_drm(vmf->page);
174 	struct nouveau_dmem *dmem = drm->dmem;
175 	struct nouveau_fence *fence;
176 	unsigned long src = 0, dst = 0;
177 	dma_addr_t dma_addr = 0;
178 	vm_fault_t ret;
179 	struct migrate_vma args = {
180 		.vma		= vmf->vma,
181 		.start		= vmf->address,
182 		.end		= vmf->address + PAGE_SIZE,
183 		.src		= &src,
184 		.dst		= &dst,
185 		.src_owner	= drm->dev,
186 	};
187 
188 	/*
189 	 * FIXME what we really want is to find some heuristic to migrate more
190 	 * than just one page on CPU fault. When such fault happens it is very
191 	 * likely that more surrounding page will CPU fault too.
192 	 */
193 	if (migrate_vma_setup(&args) < 0)
194 		return VM_FAULT_SIGBUS;
195 	if (!args.cpages)
196 		return 0;
197 
198 	ret = nouveau_dmem_fault_copy_one(drm, vmf, &args, &dma_addr);
199 	if (ret || dst == 0)
200 		goto done;
201 
202 	nouveau_fence_new(dmem->migrate.chan, false, &fence);
203 	migrate_vma_pages(&args);
204 	nouveau_dmem_fence_done(&fence);
205 	dma_unmap_page(drm->dev->dev, dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
206 done:
207 	migrate_vma_finalize(&args);
208 	return ret;
209 }
210 
211 static const struct dev_pagemap_ops nouveau_dmem_pagemap_ops = {
212 	.page_free		= nouveau_dmem_page_free,
213 	.migrate_to_ram		= nouveau_dmem_migrate_to_ram,
214 };
215 
216 static int
217 nouveau_dmem_chunk_alloc(struct nouveau_drm *drm, struct page **ppage)
218 {
219 	struct nouveau_dmem_chunk *chunk;
220 	struct resource *res;
221 	struct page *page;
222 	void *ptr;
223 	unsigned long i, pfn_first;
224 	int ret;
225 
226 	chunk = kzalloc(sizeof(*chunk), GFP_KERNEL);
227 	if (chunk == NULL) {
228 		ret = -ENOMEM;
229 		goto out;
230 	}
231 
232 	/* Allocate unused physical address space for device private pages. */
233 	res = request_free_mem_region(&iomem_resource, DMEM_CHUNK_SIZE,
234 				      "nouveau_dmem");
235 	if (IS_ERR(res)) {
236 		ret = PTR_ERR(res);
237 		goto out_free;
238 	}
239 
240 	chunk->drm = drm;
241 	chunk->pagemap.type = MEMORY_DEVICE_PRIVATE;
242 	chunk->pagemap.res = *res;
243 	chunk->pagemap.ops = &nouveau_dmem_pagemap_ops;
244 	chunk->pagemap.owner = drm->dev;
245 
246 	ret = nouveau_bo_new(&drm->client, DMEM_CHUNK_SIZE, 0,
247 			     TTM_PL_FLAG_VRAM, 0, 0, NULL, NULL,
248 			     &chunk->bo);
249 	if (ret)
250 		goto out_release;
251 
252 	ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false);
253 	if (ret)
254 		goto out_bo_free;
255 
256 	ptr = memremap_pages(&chunk->pagemap, numa_node_id());
257 	if (IS_ERR(ptr)) {
258 		ret = PTR_ERR(ptr);
259 		goto out_bo_unpin;
260 	}
261 
262 	mutex_lock(&drm->dmem->mutex);
263 	list_add(&chunk->list, &drm->dmem->chunks);
264 	mutex_unlock(&drm->dmem->mutex);
265 
266 	pfn_first = chunk->pagemap.res.start >> PAGE_SHIFT;
267 	page = pfn_to_page(pfn_first);
268 	spin_lock(&drm->dmem->lock);
269 	for (i = 0; i < DMEM_CHUNK_NPAGES - 1; ++i, ++page) {
270 		page->zone_device_data = drm->dmem->free_pages;
271 		drm->dmem->free_pages = page;
272 	}
273 	*ppage = page;
274 	chunk->callocated++;
275 	spin_unlock(&drm->dmem->lock);
276 
277 	NV_INFO(drm, "DMEM: registered %ldMB of device memory\n",
278 		DMEM_CHUNK_SIZE >> 20);
279 
280 	return 0;
281 
282 out_bo_unpin:
283 	nouveau_bo_unpin(chunk->bo);
284 out_bo_free:
285 	nouveau_bo_ref(NULL, &chunk->bo);
286 out_release:
287 	release_mem_region(chunk->pagemap.res.start,
288 			   resource_size(&chunk->pagemap.res));
289 out_free:
290 	kfree(chunk);
291 out:
292 	return ret;
293 }
294 
295 static struct page *
296 nouveau_dmem_page_alloc_locked(struct nouveau_drm *drm)
297 {
298 	struct nouveau_dmem_chunk *chunk;
299 	struct page *page = NULL;
300 	int ret;
301 
302 	spin_lock(&drm->dmem->lock);
303 	if (drm->dmem->free_pages) {
304 		page = drm->dmem->free_pages;
305 		drm->dmem->free_pages = page->zone_device_data;
306 		chunk = nouveau_page_to_chunk(page);
307 		chunk->callocated++;
308 		spin_unlock(&drm->dmem->lock);
309 	} else {
310 		spin_unlock(&drm->dmem->lock);
311 		ret = nouveau_dmem_chunk_alloc(drm, &page);
312 		if (ret)
313 			return NULL;
314 	}
315 
316 	get_page(page);
317 	lock_page(page);
318 	return page;
319 }
320 
321 static void
322 nouveau_dmem_page_free_locked(struct nouveau_drm *drm, struct page *page)
323 {
324 	unlock_page(page);
325 	put_page(page);
326 }
327 
328 void
329 nouveau_dmem_resume(struct nouveau_drm *drm)
330 {
331 	struct nouveau_dmem_chunk *chunk;
332 	int ret;
333 
334 	if (drm->dmem == NULL)
335 		return;
336 
337 	mutex_lock(&drm->dmem->mutex);
338 	list_for_each_entry(chunk, &drm->dmem->chunks, list) {
339 		ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false);
340 		/* FIXME handle pin failure */
341 		WARN_ON(ret);
342 	}
343 	mutex_unlock(&drm->dmem->mutex);
344 }
345 
346 void
347 nouveau_dmem_suspend(struct nouveau_drm *drm)
348 {
349 	struct nouveau_dmem_chunk *chunk;
350 
351 	if (drm->dmem == NULL)
352 		return;
353 
354 	mutex_lock(&drm->dmem->mutex);
355 	list_for_each_entry(chunk, &drm->dmem->chunks, list)
356 		nouveau_bo_unpin(chunk->bo);
357 	mutex_unlock(&drm->dmem->mutex);
358 }
359 
360 void
361 nouveau_dmem_fini(struct nouveau_drm *drm)
362 {
363 	struct nouveau_dmem_chunk *chunk, *tmp;
364 
365 	if (drm->dmem == NULL)
366 		return;
367 
368 	mutex_lock(&drm->dmem->mutex);
369 
370 	list_for_each_entry_safe(chunk, tmp, &drm->dmem->chunks, list) {
371 		nouveau_bo_unpin(chunk->bo);
372 		nouveau_bo_ref(NULL, &chunk->bo);
373 		list_del(&chunk->list);
374 		memunmap_pages(&chunk->pagemap);
375 		release_mem_region(chunk->pagemap.res.start,
376 				   resource_size(&chunk->pagemap.res));
377 		kfree(chunk);
378 	}
379 
380 	mutex_unlock(&drm->dmem->mutex);
381 }
382 
383 static int
384 nvc0b5_migrate_copy(struct nouveau_drm *drm, u64 npages,
385 		    enum nouveau_aper dst_aper, u64 dst_addr,
386 		    enum nouveau_aper src_aper, u64 src_addr)
387 {
388 	struct nouveau_channel *chan = drm->dmem->migrate.chan;
389 	u32 launch_dma = (1 << 9) /* MULTI_LINE_ENABLE. */ |
390 			 (1 << 8) /* DST_MEMORY_LAYOUT_PITCH. */ |
391 			 (1 << 7) /* SRC_MEMORY_LAYOUT_PITCH. */ |
392 			 (1 << 2) /* FLUSH_ENABLE_TRUE. */ |
393 			 (2 << 0) /* DATA_TRANSFER_TYPE_NON_PIPELINED. */;
394 	int ret;
395 
396 	ret = RING_SPACE(chan, 13);
397 	if (ret)
398 		return ret;
399 
400 	if (src_aper != NOUVEAU_APER_VIRT) {
401 		switch (src_aper) {
402 		case NOUVEAU_APER_VRAM:
403 			BEGIN_IMC0(chan, NvSubCopy, 0x0260, 0);
404 			break;
405 		case NOUVEAU_APER_HOST:
406 			BEGIN_IMC0(chan, NvSubCopy, 0x0260, 1);
407 			break;
408 		default:
409 			return -EINVAL;
410 		}
411 		launch_dma |= 0x00001000; /* SRC_TYPE_PHYSICAL. */
412 	}
413 
414 	if (dst_aper != NOUVEAU_APER_VIRT) {
415 		switch (dst_aper) {
416 		case NOUVEAU_APER_VRAM:
417 			BEGIN_IMC0(chan, NvSubCopy, 0x0264, 0);
418 			break;
419 		case NOUVEAU_APER_HOST:
420 			BEGIN_IMC0(chan, NvSubCopy, 0x0264, 1);
421 			break;
422 		default:
423 			return -EINVAL;
424 		}
425 		launch_dma |= 0x00002000; /* DST_TYPE_PHYSICAL. */
426 	}
427 
428 	BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
429 	OUT_RING  (chan, upper_32_bits(src_addr));
430 	OUT_RING  (chan, lower_32_bits(src_addr));
431 	OUT_RING  (chan, upper_32_bits(dst_addr));
432 	OUT_RING  (chan, lower_32_bits(dst_addr));
433 	OUT_RING  (chan, PAGE_SIZE);
434 	OUT_RING  (chan, PAGE_SIZE);
435 	OUT_RING  (chan, PAGE_SIZE);
436 	OUT_RING  (chan, npages);
437 	BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
438 	OUT_RING  (chan, launch_dma);
439 	return 0;
440 }
441 
442 static int
443 nvc0b5_migrate_clear(struct nouveau_drm *drm, u32 length,
444 		     enum nouveau_aper dst_aper, u64 dst_addr)
445 {
446 	struct nouveau_channel *chan = drm->dmem->migrate.chan;
447 	u32 launch_dma = (1 << 10) /* REMAP_ENABLE_TRUE */ |
448 			 (1 << 8) /* DST_MEMORY_LAYOUT_PITCH. */ |
449 			 (1 << 7) /* SRC_MEMORY_LAYOUT_PITCH. */ |
450 			 (1 << 2) /* FLUSH_ENABLE_TRUE. */ |
451 			 (2 << 0) /* DATA_TRANSFER_TYPE_NON_PIPELINED. */;
452 	u32 remap = (4 <<  0) /* DST_X_CONST_A */ |
453 		    (5 <<  4) /* DST_Y_CONST_B */ |
454 		    (3 << 16) /* COMPONENT_SIZE_FOUR */ |
455 		    (1 << 24) /* NUM_DST_COMPONENTS_TWO */;
456 	int ret;
457 
458 	ret = RING_SPACE(chan, 12);
459 	if (ret)
460 		return ret;
461 
462 	switch (dst_aper) {
463 	case NOUVEAU_APER_VRAM:
464 		BEGIN_IMC0(chan, NvSubCopy, 0x0264, 0);
465 			break;
466 	case NOUVEAU_APER_HOST:
467 		BEGIN_IMC0(chan, NvSubCopy, 0x0264, 1);
468 		break;
469 	default:
470 		return -EINVAL;
471 	}
472 	launch_dma |= 0x00002000; /* DST_TYPE_PHYSICAL. */
473 
474 	BEGIN_NVC0(chan, NvSubCopy, 0x0700, 3);
475 	OUT_RING(chan, 0);
476 	OUT_RING(chan, 0);
477 	OUT_RING(chan, remap);
478 	BEGIN_NVC0(chan, NvSubCopy, 0x0408, 2);
479 	OUT_RING(chan, upper_32_bits(dst_addr));
480 	OUT_RING(chan, lower_32_bits(dst_addr));
481 	BEGIN_NVC0(chan, NvSubCopy, 0x0418, 1);
482 	OUT_RING(chan, length >> 3);
483 	BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
484 	OUT_RING(chan, launch_dma);
485 	return 0;
486 }
487 
488 static int
489 nouveau_dmem_migrate_init(struct nouveau_drm *drm)
490 {
491 	switch (drm->ttm.copy.oclass) {
492 	case PASCAL_DMA_COPY_A:
493 	case PASCAL_DMA_COPY_B:
494 	case  VOLTA_DMA_COPY_A:
495 	case TURING_DMA_COPY_A:
496 		drm->dmem->migrate.copy_func = nvc0b5_migrate_copy;
497 		drm->dmem->migrate.clear_func = nvc0b5_migrate_clear;
498 		drm->dmem->migrate.chan = drm->ttm.chan;
499 		return 0;
500 	default:
501 		break;
502 	}
503 	return -ENODEV;
504 }
505 
506 void
507 nouveau_dmem_init(struct nouveau_drm *drm)
508 {
509 	int ret;
510 
511 	/* This only make sense on PASCAL or newer */
512 	if (drm->client.device.info.family < NV_DEVICE_INFO_V0_PASCAL)
513 		return;
514 
515 	if (!(drm->dmem = kzalloc(sizeof(*drm->dmem), GFP_KERNEL)))
516 		return;
517 
518 	drm->dmem->drm = drm;
519 	mutex_init(&drm->dmem->mutex);
520 	INIT_LIST_HEAD(&drm->dmem->chunks);
521 	mutex_init(&drm->dmem->mutex);
522 	spin_lock_init(&drm->dmem->lock);
523 
524 	/* Initialize migration dma helpers before registering memory */
525 	ret = nouveau_dmem_migrate_init(drm);
526 	if (ret) {
527 		kfree(drm->dmem);
528 		drm->dmem = NULL;
529 	}
530 }
531 
532 static unsigned long nouveau_dmem_migrate_copy_one(struct nouveau_drm *drm,
533 		unsigned long src, dma_addr_t *dma_addr, u64 *pfn)
534 {
535 	struct device *dev = drm->dev->dev;
536 	struct page *dpage, *spage;
537 	unsigned long paddr;
538 
539 	spage = migrate_pfn_to_page(src);
540 	if (!(src & MIGRATE_PFN_MIGRATE))
541 		goto out;
542 
543 	dpage = nouveau_dmem_page_alloc_locked(drm);
544 	if (!dpage)
545 		goto out;
546 
547 	paddr = nouveau_dmem_page_addr(dpage);
548 	if (spage) {
549 		*dma_addr = dma_map_page(dev, spage, 0, page_size(spage),
550 					 DMA_BIDIRECTIONAL);
551 		if (dma_mapping_error(dev, *dma_addr))
552 			goto out_free_page;
553 		if (drm->dmem->migrate.copy_func(drm, page_size(spage),
554 			NOUVEAU_APER_VRAM, paddr, NOUVEAU_APER_HOST, *dma_addr))
555 			goto out_dma_unmap;
556 	} else {
557 		*dma_addr = DMA_MAPPING_ERROR;
558 		if (drm->dmem->migrate.clear_func(drm, page_size(dpage),
559 			NOUVEAU_APER_VRAM, paddr))
560 			goto out_free_page;
561 	}
562 
563 	*pfn = NVIF_VMM_PFNMAP_V0_V | NVIF_VMM_PFNMAP_V0_VRAM |
564 		((paddr >> PAGE_SHIFT) << NVIF_VMM_PFNMAP_V0_ADDR_SHIFT);
565 	if (src & MIGRATE_PFN_WRITE)
566 		*pfn |= NVIF_VMM_PFNMAP_V0_W;
567 	return migrate_pfn(page_to_pfn(dpage)) | MIGRATE_PFN_LOCKED;
568 
569 out_dma_unmap:
570 	dma_unmap_page(dev, *dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
571 out_free_page:
572 	nouveau_dmem_page_free_locked(drm, dpage);
573 out:
574 	*pfn = NVIF_VMM_PFNMAP_V0_NONE;
575 	return 0;
576 }
577 
578 static void nouveau_dmem_migrate_chunk(struct nouveau_drm *drm,
579 		struct nouveau_svmm *svmm, struct migrate_vma *args,
580 		dma_addr_t *dma_addrs, u64 *pfns)
581 {
582 	struct nouveau_fence *fence;
583 	unsigned long addr = args->start, nr_dma = 0, i;
584 
585 	for (i = 0; addr < args->end; i++) {
586 		args->dst[i] = nouveau_dmem_migrate_copy_one(drm, args->src[i],
587 				dma_addrs + nr_dma, pfns + i);
588 		if (!dma_mapping_error(drm->dev->dev, dma_addrs[nr_dma]))
589 			nr_dma++;
590 		addr += PAGE_SIZE;
591 	}
592 
593 	nouveau_fence_new(drm->dmem->migrate.chan, false, &fence);
594 	migrate_vma_pages(args);
595 	nouveau_dmem_fence_done(&fence);
596 	nouveau_pfns_map(svmm, args->vma->vm_mm, args->start, pfns, i);
597 
598 	while (nr_dma--) {
599 		dma_unmap_page(drm->dev->dev, dma_addrs[nr_dma], PAGE_SIZE,
600 				DMA_BIDIRECTIONAL);
601 	}
602 	migrate_vma_finalize(args);
603 }
604 
605 int
606 nouveau_dmem_migrate_vma(struct nouveau_drm *drm,
607 			 struct nouveau_svmm *svmm,
608 			 struct vm_area_struct *vma,
609 			 unsigned long start,
610 			 unsigned long end)
611 {
612 	unsigned long npages = (end - start) >> PAGE_SHIFT;
613 	unsigned long max = min(SG_MAX_SINGLE_ALLOC, npages);
614 	dma_addr_t *dma_addrs;
615 	struct migrate_vma args = {
616 		.vma		= vma,
617 		.start		= start,
618 	};
619 	unsigned long i;
620 	u64 *pfns;
621 	int ret = -ENOMEM;
622 
623 	if (drm->dmem == NULL)
624 		return -ENODEV;
625 
626 	args.src = kcalloc(max, sizeof(*args.src), GFP_KERNEL);
627 	if (!args.src)
628 		goto out;
629 	args.dst = kcalloc(max, sizeof(*args.dst), GFP_KERNEL);
630 	if (!args.dst)
631 		goto out_free_src;
632 
633 	dma_addrs = kmalloc_array(max, sizeof(*dma_addrs), GFP_KERNEL);
634 	if (!dma_addrs)
635 		goto out_free_dst;
636 
637 	pfns = nouveau_pfns_alloc(max);
638 	if (!pfns)
639 		goto out_free_dma;
640 
641 	for (i = 0; i < npages; i += max) {
642 		args.end = start + (max << PAGE_SHIFT);
643 		ret = migrate_vma_setup(&args);
644 		if (ret)
645 			goto out_free_pfns;
646 
647 		if (args.cpages)
648 			nouveau_dmem_migrate_chunk(drm, svmm, &args, dma_addrs,
649 						   pfns);
650 		args.start = args.end;
651 	}
652 
653 	ret = 0;
654 out_free_pfns:
655 	nouveau_pfns_free(pfns);
656 out_free_dma:
657 	kfree(dma_addrs);
658 out_free_dst:
659 	kfree(args.dst);
660 out_free_src:
661 	kfree(args.src);
662 out:
663 	return ret;
664 }
665