• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include <linux/memremap.h>
43 #include <linux/migrate.h>
44 
45 /*
46  * FIXME: this is ugly right now we are using TTM to allocate vram and we pin
47  * it in vram while in use. We likely want to overhaul memory management for
48  * nouveau to be more page like (not necessarily with system page size but a
49  * bigger page size) at lowest level and have some shim layer on top that would
50  * provide the same functionality as TTM.
51  */
52 #define DMEM_CHUNK_SIZE (2UL << 20)
53 #define DMEM_CHUNK_NPAGES (DMEM_CHUNK_SIZE >> PAGE_SHIFT)
54 
55 enum nouveau_aper {
56 	NOUVEAU_APER_VIRT,
57 	NOUVEAU_APER_VRAM,
58 	NOUVEAU_APER_HOST,
59 };
60 
61 typedef int (*nouveau_migrate_copy_t)(struct nouveau_drm *drm, u64 npages,
62 				      enum nouveau_aper, u64 dst_addr,
63 				      enum nouveau_aper, u64 src_addr);
64 typedef int (*nouveau_clear_page_t)(struct nouveau_drm *drm, u32 length,
65 				      enum nouveau_aper, u64 dst_addr);
66 
67 struct nouveau_dmem_chunk {
68 	struct list_head list;
69 	struct nouveau_bo *bo;
70 	struct nouveau_drm *drm;
71 	unsigned long callocated;
72 	struct dev_pagemap pagemap;
73 };
74 
75 struct nouveau_dmem_migrate {
76 	nouveau_migrate_copy_t copy_func;
77 	nouveau_clear_page_t clear_func;
78 	struct nouveau_channel *chan;
79 };
80 
81 struct nouveau_dmem {
82 	struct nouveau_drm *drm;
83 	struct nouveau_dmem_migrate migrate;
84 	struct list_head chunks;
85 	struct mutex mutex;
86 	struct page *free_pages;
87 	spinlock_t lock;
88 };
89 
nouveau_page_to_chunk(struct page * page)90 static struct nouveau_dmem_chunk *nouveau_page_to_chunk(struct page *page)
91 {
92 	return container_of(page->pgmap, struct nouveau_dmem_chunk, pagemap);
93 }
94 
page_to_drm(struct page * page)95 static struct nouveau_drm *page_to_drm(struct page *page)
96 {
97 	struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
98 
99 	return chunk->drm;
100 }
101 
nouveau_dmem_page_addr(struct page * page)102 unsigned long nouveau_dmem_page_addr(struct page *page)
103 {
104 	struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
105 	unsigned long off = (page_to_pfn(page) << PAGE_SHIFT) -
106 				chunk->pagemap.range.start;
107 
108 	return chunk->bo->offset + off;
109 }
110 
nouveau_dmem_page_free(struct page * page)111 static void nouveau_dmem_page_free(struct page *page)
112 {
113 	struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
114 	struct nouveau_dmem *dmem = chunk->drm->dmem;
115 
116 	spin_lock(&dmem->lock);
117 	page->zone_device_data = dmem->free_pages;
118 	dmem->free_pages = page;
119 
120 	WARN_ON(!chunk->callocated);
121 	chunk->callocated--;
122 	/*
123 	 * FIXME when chunk->callocated reach 0 we should add the chunk to
124 	 * a reclaim list so that it can be freed in case of memory pressure.
125 	 */
126 	spin_unlock(&dmem->lock);
127 }
128 
nouveau_dmem_fence_done(struct nouveau_fence ** fence)129 static void nouveau_dmem_fence_done(struct nouveau_fence **fence)
130 {
131 	if (fence) {
132 		nouveau_fence_wait(*fence, true, false);
133 		nouveau_fence_unref(fence);
134 	} else {
135 		/*
136 		 * FIXME wait for channel to be IDLE before calling finalizing
137 		 * the hmem object.
138 		 */
139 	}
140 }
141 
nouveau_dmem_copy_one(struct nouveau_drm * drm,struct page * spage,struct page * dpage,dma_addr_t * dma_addr)142 static int nouveau_dmem_copy_one(struct nouveau_drm *drm, struct page *spage,
143 				struct page *dpage, dma_addr_t *dma_addr)
144 {
145 	struct device *dev = drm->dev->dev;
146 
147 	lock_page(dpage);
148 
149 	*dma_addr = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
150 	if (dma_mapping_error(dev, *dma_addr))
151 		return -EIO;
152 
153 	if (drm->dmem->migrate.copy_func(drm, 1, NOUVEAU_APER_HOST, *dma_addr,
154 					 NOUVEAU_APER_VRAM, nouveau_dmem_page_addr(spage))) {
155 		dma_unmap_page(dev, *dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
156 		return -EIO;
157 	}
158 
159 	return 0;
160 }
161 
nouveau_dmem_migrate_to_ram(struct vm_fault * vmf)162 static vm_fault_t nouveau_dmem_migrate_to_ram(struct vm_fault *vmf)
163 {
164 	struct nouveau_drm *drm = page_to_drm(vmf->page);
165 	struct nouveau_dmem *dmem = drm->dmem;
166 	struct nouveau_fence *fence;
167 	struct nouveau_svmm *svmm;
168 	struct page *spage, *dpage;
169 	unsigned long src = 0, dst = 0;
170 	dma_addr_t dma_addr = 0;
171 	vm_fault_t ret = 0;
172 	struct migrate_vma args = {
173 		.vma		= vmf->vma,
174 		.start		= vmf->address,
175 		.end		= vmf->address + PAGE_SIZE,
176 		.src		= &src,
177 		.dst		= &dst,
178 		.pgmap_owner	= drm->dev,
179 		.fault_page	= vmf->page,
180 		.flags		= MIGRATE_VMA_SELECT_DEVICE_PRIVATE,
181 	};
182 
183 	/*
184 	 * FIXME what we really want is to find some heuristic to migrate more
185 	 * than just one page on CPU fault. When such fault happens it is very
186 	 * likely that more surrounding page will CPU fault too.
187 	 */
188 	if (migrate_vma_setup(&args) < 0)
189 		return VM_FAULT_SIGBUS;
190 	if (!args.cpages)
191 		return 0;
192 
193 	spage = migrate_pfn_to_page(src);
194 	if (!spage || !(src & MIGRATE_PFN_MIGRATE))
195 		goto done;
196 
197 	dpage = alloc_page_vma(GFP_HIGHUSER, vmf->vma, vmf->address);
198 	if (!dpage)
199 		goto done;
200 
201 	dst = migrate_pfn(page_to_pfn(dpage));
202 
203 	svmm = spage->zone_device_data;
204 	mutex_lock(&svmm->mutex);
205 	nouveau_svmm_invalidate(svmm, args.start, args.end);
206 	ret = nouveau_dmem_copy_one(drm, spage, dpage, &dma_addr);
207 	mutex_unlock(&svmm->mutex);
208 	if (ret) {
209 		ret = VM_FAULT_SIGBUS;
210 		goto done;
211 	}
212 
213 	nouveau_fence_new(dmem->migrate.chan, false, &fence);
214 	migrate_vma_pages(&args);
215 	nouveau_dmem_fence_done(&fence);
216 	dma_unmap_page(drm->dev->dev, dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
217 done:
218 	migrate_vma_finalize(&args);
219 	return ret;
220 }
221 
222 static const struct dev_pagemap_ops nouveau_dmem_pagemap_ops = {
223 	.page_free		= nouveau_dmem_page_free,
224 	.migrate_to_ram		= nouveau_dmem_migrate_to_ram,
225 };
226 
227 static int
nouveau_dmem_chunk_alloc(struct nouveau_drm * drm,struct page ** ppage)228 nouveau_dmem_chunk_alloc(struct nouveau_drm *drm, struct page **ppage)
229 {
230 	struct nouveau_dmem_chunk *chunk;
231 	struct resource *res;
232 	struct page *page;
233 	void *ptr;
234 	unsigned long i, pfn_first;
235 	int ret;
236 
237 	chunk = kzalloc(sizeof(*chunk), GFP_KERNEL);
238 	if (chunk == NULL) {
239 		ret = -ENOMEM;
240 		goto out;
241 	}
242 
243 	/* Allocate unused physical address space for device private pages. */
244 	res = request_free_mem_region(&iomem_resource, DMEM_CHUNK_SIZE,
245 				      "nouveau_dmem");
246 	if (IS_ERR(res)) {
247 		ret = PTR_ERR(res);
248 		goto out_free;
249 	}
250 
251 	chunk->drm = drm;
252 	chunk->pagemap.type = MEMORY_DEVICE_PRIVATE;
253 	chunk->pagemap.range.start = res->start;
254 	chunk->pagemap.range.end = res->end;
255 	chunk->pagemap.nr_range = 1;
256 	chunk->pagemap.ops = &nouveau_dmem_pagemap_ops;
257 	chunk->pagemap.owner = drm->dev;
258 
259 	ret = nouveau_bo_new(&drm->client, DMEM_CHUNK_SIZE, 0,
260 			     NOUVEAU_GEM_DOMAIN_VRAM, 0, 0, NULL, NULL,
261 			     &chunk->bo);
262 	if (ret)
263 		goto out_release;
264 
265 	ret = nouveau_bo_pin(chunk->bo, NOUVEAU_GEM_DOMAIN_VRAM, false);
266 	if (ret)
267 		goto out_bo_free;
268 
269 	ptr = memremap_pages(&chunk->pagemap, numa_node_id());
270 	if (IS_ERR(ptr)) {
271 		ret = PTR_ERR(ptr);
272 		goto out_bo_unpin;
273 	}
274 
275 	mutex_lock(&drm->dmem->mutex);
276 	list_add(&chunk->list, &drm->dmem->chunks);
277 	mutex_unlock(&drm->dmem->mutex);
278 
279 	pfn_first = chunk->pagemap.range.start >> PAGE_SHIFT;
280 	page = pfn_to_page(pfn_first);
281 	spin_lock(&drm->dmem->lock);
282 	for (i = 0; i < DMEM_CHUNK_NPAGES - 1; ++i, ++page) {
283 		page->zone_device_data = drm->dmem->free_pages;
284 		drm->dmem->free_pages = page;
285 	}
286 	*ppage = page;
287 	chunk->callocated++;
288 	spin_unlock(&drm->dmem->lock);
289 
290 	NV_INFO(drm, "DMEM: registered %ldMB of device memory\n",
291 		DMEM_CHUNK_SIZE >> 20);
292 
293 	return 0;
294 
295 out_bo_unpin:
296 	nouveau_bo_unpin(chunk->bo);
297 out_bo_free:
298 	nouveau_bo_ref(NULL, &chunk->bo);
299 out_release:
300 	release_mem_region(chunk->pagemap.range.start, range_len(&chunk->pagemap.range));
301 out_free:
302 	kfree(chunk);
303 out:
304 	return ret;
305 }
306 
307 static struct page *
nouveau_dmem_page_alloc_locked(struct nouveau_drm * drm)308 nouveau_dmem_page_alloc_locked(struct nouveau_drm *drm)
309 {
310 	struct nouveau_dmem_chunk *chunk;
311 	struct page *page = NULL;
312 	int ret;
313 
314 	spin_lock(&drm->dmem->lock);
315 	if (drm->dmem->free_pages) {
316 		page = drm->dmem->free_pages;
317 		drm->dmem->free_pages = page->zone_device_data;
318 		chunk = nouveau_page_to_chunk(page);
319 		chunk->callocated++;
320 		spin_unlock(&drm->dmem->lock);
321 	} else {
322 		spin_unlock(&drm->dmem->lock);
323 		ret = nouveau_dmem_chunk_alloc(drm, &page);
324 		if (ret)
325 			return NULL;
326 	}
327 
328 	zone_device_page_init(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 /*
372  * Evict all pages mapping a chunk.
373  */
374 static void
nouveau_dmem_evict_chunk(struct nouveau_dmem_chunk * chunk)375 nouveau_dmem_evict_chunk(struct nouveau_dmem_chunk *chunk)
376 {
377 	unsigned long i, npages = range_len(&chunk->pagemap.range) >> PAGE_SHIFT;
378 	unsigned long *src_pfns, *dst_pfns;
379 	dma_addr_t *dma_addrs;
380 	struct nouveau_fence *fence;
381 
382 	src_pfns = kcalloc(npages, sizeof(*src_pfns), GFP_KERNEL);
383 	dst_pfns = kcalloc(npages, sizeof(*dst_pfns), GFP_KERNEL);
384 	dma_addrs = kcalloc(npages, sizeof(*dma_addrs), GFP_KERNEL);
385 
386 	migrate_device_range(src_pfns, chunk->pagemap.range.start >> PAGE_SHIFT,
387 			npages);
388 
389 	for (i = 0; i < npages; i++) {
390 		if (src_pfns[i] & MIGRATE_PFN_MIGRATE) {
391 			struct page *dpage;
392 
393 			/*
394 			 * _GFP_NOFAIL because the GPU is going away and there
395 			 * is nothing sensible we can do if we can't copy the
396 			 * data back.
397 			 */
398 			dpage = alloc_page(GFP_HIGHUSER | __GFP_NOFAIL);
399 			dst_pfns[i] = migrate_pfn(page_to_pfn(dpage));
400 			nouveau_dmem_copy_one(chunk->drm,
401 					migrate_pfn_to_page(src_pfns[i]), dpage,
402 					&dma_addrs[i]);
403 		}
404 	}
405 
406 	nouveau_fence_new(chunk->drm->dmem->migrate.chan, false, &fence);
407 	migrate_device_pages(src_pfns, dst_pfns, npages);
408 	nouveau_dmem_fence_done(&fence);
409 	migrate_device_finalize(src_pfns, dst_pfns, npages);
410 	kfree(src_pfns);
411 	kfree(dst_pfns);
412 	for (i = 0; i < npages; i++)
413 		dma_unmap_page(chunk->drm->dev->dev, dma_addrs[i], PAGE_SIZE, DMA_BIDIRECTIONAL);
414 	kfree(dma_addrs);
415 }
416 
417 void
nouveau_dmem_fini(struct nouveau_drm * drm)418 nouveau_dmem_fini(struct nouveau_drm *drm)
419 {
420 	struct nouveau_dmem_chunk *chunk, *tmp;
421 
422 	if (drm->dmem == NULL)
423 		return;
424 
425 	mutex_lock(&drm->dmem->mutex);
426 
427 	list_for_each_entry_safe(chunk, tmp, &drm->dmem->chunks, list) {
428 		nouveau_dmem_evict_chunk(chunk);
429 		nouveau_bo_unpin(chunk->bo);
430 		nouveau_bo_ref(NULL, &chunk->bo);
431 		WARN_ON(chunk->callocated);
432 		list_del(&chunk->list);
433 		memunmap_pages(&chunk->pagemap);
434 		release_mem_region(chunk->pagemap.range.start,
435 				   range_len(&chunk->pagemap.range));
436 		kfree(chunk);
437 	}
438 
439 	mutex_unlock(&drm->dmem->mutex);
440 }
441 
442 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)443 nvc0b5_migrate_copy(struct nouveau_drm *drm, u64 npages,
444 		    enum nouveau_aper dst_aper, u64 dst_addr,
445 		    enum nouveau_aper src_aper, u64 src_addr)
446 {
447 	struct nvif_push *push = drm->dmem->migrate.chan->chan.push;
448 	u32 launch_dma = 0;
449 	int ret;
450 
451 	ret = PUSH_WAIT(push, 13);
452 	if (ret)
453 		return ret;
454 
455 	if (src_aper != NOUVEAU_APER_VIRT) {
456 		switch (src_aper) {
457 		case NOUVEAU_APER_VRAM:
458 			PUSH_IMMD(push, NVA0B5, SET_SRC_PHYS_MODE,
459 				  NVDEF(NVA0B5, SET_SRC_PHYS_MODE, TARGET, LOCAL_FB));
460 			break;
461 		case NOUVEAU_APER_HOST:
462 			PUSH_IMMD(push, NVA0B5, SET_SRC_PHYS_MODE,
463 				  NVDEF(NVA0B5, SET_SRC_PHYS_MODE, TARGET, COHERENT_SYSMEM));
464 			break;
465 		default:
466 			return -EINVAL;
467 		}
468 
469 		launch_dma |= NVDEF(NVA0B5, LAUNCH_DMA, SRC_TYPE, PHYSICAL);
470 	}
471 
472 	if (dst_aper != NOUVEAU_APER_VIRT) {
473 		switch (dst_aper) {
474 		case NOUVEAU_APER_VRAM:
475 			PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
476 				  NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, LOCAL_FB));
477 			break;
478 		case NOUVEAU_APER_HOST:
479 			PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
480 				  NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, COHERENT_SYSMEM));
481 			break;
482 		default:
483 			return -EINVAL;
484 		}
485 
486 		launch_dma |= NVDEF(NVA0B5, LAUNCH_DMA, DST_TYPE, PHYSICAL);
487 	}
488 
489 	PUSH_MTHD(push, NVA0B5, OFFSET_IN_UPPER,
490 		  NVVAL(NVA0B5, OFFSET_IN_UPPER, UPPER, upper_32_bits(src_addr)),
491 
492 				OFFSET_IN_LOWER, lower_32_bits(src_addr),
493 
494 				OFFSET_OUT_UPPER,
495 		  NVVAL(NVA0B5, OFFSET_OUT_UPPER, UPPER, upper_32_bits(dst_addr)),
496 
497 				OFFSET_OUT_LOWER, lower_32_bits(dst_addr),
498 				PITCH_IN, PAGE_SIZE,
499 				PITCH_OUT, PAGE_SIZE,
500 				LINE_LENGTH_IN, PAGE_SIZE,
501 				LINE_COUNT, npages);
502 
503 	PUSH_MTHD(push, NVA0B5, LAUNCH_DMA, launch_dma |
504 		  NVDEF(NVA0B5, LAUNCH_DMA, DATA_TRANSFER_TYPE, NON_PIPELINED) |
505 		  NVDEF(NVA0B5, LAUNCH_DMA, FLUSH_ENABLE, TRUE) |
506 		  NVDEF(NVA0B5, LAUNCH_DMA, SEMAPHORE_TYPE, NONE) |
507 		  NVDEF(NVA0B5, LAUNCH_DMA, INTERRUPT_TYPE, NONE) |
508 		  NVDEF(NVA0B5, LAUNCH_DMA, SRC_MEMORY_LAYOUT, PITCH) |
509 		  NVDEF(NVA0B5, LAUNCH_DMA, DST_MEMORY_LAYOUT, PITCH) |
510 		  NVDEF(NVA0B5, LAUNCH_DMA, MULTI_LINE_ENABLE, TRUE) |
511 		  NVDEF(NVA0B5, LAUNCH_DMA, REMAP_ENABLE, FALSE) |
512 		  NVDEF(NVA0B5, LAUNCH_DMA, BYPASS_L2, USE_PTE_SETTING));
513 	return 0;
514 }
515 
516 static int
nvc0b5_migrate_clear(struct nouveau_drm * drm,u32 length,enum nouveau_aper dst_aper,u64 dst_addr)517 nvc0b5_migrate_clear(struct nouveau_drm *drm, u32 length,
518 		     enum nouveau_aper dst_aper, u64 dst_addr)
519 {
520 	struct nvif_push *push = drm->dmem->migrate.chan->chan.push;
521 	u32 launch_dma = 0;
522 	int ret;
523 
524 	ret = PUSH_WAIT(push, 12);
525 	if (ret)
526 		return ret;
527 
528 	switch (dst_aper) {
529 	case NOUVEAU_APER_VRAM:
530 		PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
531 			  NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, LOCAL_FB));
532 		break;
533 	case NOUVEAU_APER_HOST:
534 		PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
535 			  NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, COHERENT_SYSMEM));
536 		break;
537 	default:
538 		return -EINVAL;
539 	}
540 
541 	launch_dma |= NVDEF(NVA0B5, LAUNCH_DMA, DST_TYPE, PHYSICAL);
542 
543 	PUSH_MTHD(push, NVA0B5, SET_REMAP_CONST_A, 0,
544 				SET_REMAP_CONST_B, 0,
545 
546 				SET_REMAP_COMPONENTS,
547 		  NVDEF(NVA0B5, SET_REMAP_COMPONENTS, DST_X, CONST_A) |
548 		  NVDEF(NVA0B5, SET_REMAP_COMPONENTS, DST_Y, CONST_B) |
549 		  NVDEF(NVA0B5, SET_REMAP_COMPONENTS, COMPONENT_SIZE, FOUR) |
550 		  NVDEF(NVA0B5, SET_REMAP_COMPONENTS, NUM_DST_COMPONENTS, TWO));
551 
552 	PUSH_MTHD(push, NVA0B5, OFFSET_OUT_UPPER,
553 		  NVVAL(NVA0B5, OFFSET_OUT_UPPER, UPPER, upper_32_bits(dst_addr)),
554 
555 				OFFSET_OUT_LOWER, lower_32_bits(dst_addr));
556 
557 	PUSH_MTHD(push, NVA0B5, LINE_LENGTH_IN, length >> 3);
558 
559 	PUSH_MTHD(push, NVA0B5, LAUNCH_DMA, launch_dma |
560 		  NVDEF(NVA0B5, LAUNCH_DMA, DATA_TRANSFER_TYPE, NON_PIPELINED) |
561 		  NVDEF(NVA0B5, LAUNCH_DMA, FLUSH_ENABLE, TRUE) |
562 		  NVDEF(NVA0B5, LAUNCH_DMA, SEMAPHORE_TYPE, NONE) |
563 		  NVDEF(NVA0B5, LAUNCH_DMA, INTERRUPT_TYPE, NONE) |
564 		  NVDEF(NVA0B5, LAUNCH_DMA, SRC_MEMORY_LAYOUT, PITCH) |
565 		  NVDEF(NVA0B5, LAUNCH_DMA, DST_MEMORY_LAYOUT, PITCH) |
566 		  NVDEF(NVA0B5, LAUNCH_DMA, MULTI_LINE_ENABLE, FALSE) |
567 		  NVDEF(NVA0B5, LAUNCH_DMA, REMAP_ENABLE, TRUE) |
568 		  NVDEF(NVA0B5, LAUNCH_DMA, BYPASS_L2, USE_PTE_SETTING));
569 	return 0;
570 }
571 
572 static int
nouveau_dmem_migrate_init(struct nouveau_drm * drm)573 nouveau_dmem_migrate_init(struct nouveau_drm *drm)
574 {
575 	switch (drm->ttm.copy.oclass) {
576 	case PASCAL_DMA_COPY_A:
577 	case PASCAL_DMA_COPY_B:
578 	case  VOLTA_DMA_COPY_A:
579 	case TURING_DMA_COPY_A:
580 		drm->dmem->migrate.copy_func = nvc0b5_migrate_copy;
581 		drm->dmem->migrate.clear_func = nvc0b5_migrate_clear;
582 		drm->dmem->migrate.chan = drm->ttm.chan;
583 		return 0;
584 	default:
585 		break;
586 	}
587 	return -ENODEV;
588 }
589 
590 void
nouveau_dmem_init(struct nouveau_drm * drm)591 nouveau_dmem_init(struct nouveau_drm *drm)
592 {
593 	int ret;
594 
595 	/* This only make sense on PASCAL or newer */
596 	if (drm->client.device.info.family < NV_DEVICE_INFO_V0_PASCAL)
597 		return;
598 
599 	if (!(drm->dmem = kzalloc(sizeof(*drm->dmem), GFP_KERNEL)))
600 		return;
601 
602 	drm->dmem->drm = drm;
603 	mutex_init(&drm->dmem->mutex);
604 	INIT_LIST_HEAD(&drm->dmem->chunks);
605 	mutex_init(&drm->dmem->mutex);
606 	spin_lock_init(&drm->dmem->lock);
607 
608 	/* Initialize migration dma helpers before registering memory */
609 	ret = nouveau_dmem_migrate_init(drm);
610 	if (ret) {
611 		kfree(drm->dmem);
612 		drm->dmem = NULL;
613 	}
614 }
615 
nouveau_dmem_migrate_copy_one(struct nouveau_drm * drm,struct nouveau_svmm * svmm,unsigned long src,dma_addr_t * dma_addr,u64 * pfn)616 static unsigned long nouveau_dmem_migrate_copy_one(struct nouveau_drm *drm,
617 		struct nouveau_svmm *svmm, unsigned long src,
618 		dma_addr_t *dma_addr, u64 *pfn)
619 {
620 	struct device *dev = drm->dev->dev;
621 	struct page *dpage, *spage;
622 	unsigned long paddr;
623 
624 	spage = migrate_pfn_to_page(src);
625 	if (!(src & MIGRATE_PFN_MIGRATE))
626 		goto out;
627 
628 	dpage = nouveau_dmem_page_alloc_locked(drm);
629 	if (!dpage)
630 		goto out;
631 
632 	paddr = nouveau_dmem_page_addr(dpage);
633 	if (spage) {
634 		*dma_addr = dma_map_page(dev, spage, 0, page_size(spage),
635 					 DMA_BIDIRECTIONAL);
636 		if (dma_mapping_error(dev, *dma_addr))
637 			goto out_free_page;
638 		if (drm->dmem->migrate.copy_func(drm, 1,
639 			NOUVEAU_APER_VRAM, paddr, NOUVEAU_APER_HOST, *dma_addr))
640 			goto out_dma_unmap;
641 	} else {
642 		*dma_addr = DMA_MAPPING_ERROR;
643 		if (drm->dmem->migrate.clear_func(drm, page_size(dpage),
644 			NOUVEAU_APER_VRAM, paddr))
645 			goto out_free_page;
646 	}
647 
648 	dpage->zone_device_data = svmm;
649 	*pfn = NVIF_VMM_PFNMAP_V0_V | NVIF_VMM_PFNMAP_V0_VRAM |
650 		((paddr >> PAGE_SHIFT) << NVIF_VMM_PFNMAP_V0_ADDR_SHIFT);
651 	if (src & MIGRATE_PFN_WRITE)
652 		*pfn |= NVIF_VMM_PFNMAP_V0_W;
653 	return migrate_pfn(page_to_pfn(dpage));
654 
655 out_dma_unmap:
656 	dma_unmap_page(dev, *dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
657 out_free_page:
658 	nouveau_dmem_page_free_locked(drm, dpage);
659 out:
660 	*pfn = NVIF_VMM_PFNMAP_V0_NONE;
661 	return 0;
662 }
663 
nouveau_dmem_migrate_chunk(struct nouveau_drm * drm,struct nouveau_svmm * svmm,struct migrate_vma * args,dma_addr_t * dma_addrs,u64 * pfns)664 static void nouveau_dmem_migrate_chunk(struct nouveau_drm *drm,
665 		struct nouveau_svmm *svmm, struct migrate_vma *args,
666 		dma_addr_t *dma_addrs, u64 *pfns)
667 {
668 	struct nouveau_fence *fence;
669 	unsigned long addr = args->start, nr_dma = 0, i;
670 
671 	for (i = 0; addr < args->end; i++) {
672 		args->dst[i] = nouveau_dmem_migrate_copy_one(drm, svmm,
673 				args->src[i], dma_addrs + nr_dma, pfns + i);
674 		if (!dma_mapping_error(drm->dev->dev, dma_addrs[nr_dma]))
675 			nr_dma++;
676 		addr += PAGE_SIZE;
677 	}
678 
679 	nouveau_fence_new(drm->dmem->migrate.chan, false, &fence);
680 	migrate_vma_pages(args);
681 	nouveau_dmem_fence_done(&fence);
682 	nouveau_pfns_map(svmm, args->vma->vm_mm, args->start, pfns, i);
683 
684 	while (nr_dma--) {
685 		dma_unmap_page(drm->dev->dev, dma_addrs[nr_dma], PAGE_SIZE,
686 				DMA_BIDIRECTIONAL);
687 	}
688 	migrate_vma_finalize(args);
689 }
690 
691 int
nouveau_dmem_migrate_vma(struct nouveau_drm * drm,struct nouveau_svmm * svmm,struct vm_area_struct * vma,unsigned long start,unsigned long end)692 nouveau_dmem_migrate_vma(struct nouveau_drm *drm,
693 			 struct nouveau_svmm *svmm,
694 			 struct vm_area_struct *vma,
695 			 unsigned long start,
696 			 unsigned long end)
697 {
698 	unsigned long npages = (end - start) >> PAGE_SHIFT;
699 	unsigned long max = min(SG_MAX_SINGLE_ALLOC, npages);
700 	dma_addr_t *dma_addrs;
701 	struct migrate_vma args = {
702 		.vma		= vma,
703 		.start		= start,
704 		.pgmap_owner	= drm->dev,
705 		.flags		= MIGRATE_VMA_SELECT_SYSTEM,
706 	};
707 	unsigned long i;
708 	u64 *pfns;
709 	int ret = -ENOMEM;
710 
711 	if (drm->dmem == NULL)
712 		return -ENODEV;
713 
714 	args.src = kcalloc(max, sizeof(*args.src), GFP_KERNEL);
715 	if (!args.src)
716 		goto out;
717 	args.dst = kcalloc(max, sizeof(*args.dst), GFP_KERNEL);
718 	if (!args.dst)
719 		goto out_free_src;
720 
721 	dma_addrs = kmalloc_array(max, sizeof(*dma_addrs), GFP_KERNEL);
722 	if (!dma_addrs)
723 		goto out_free_dst;
724 
725 	pfns = nouveau_pfns_alloc(max);
726 	if (!pfns)
727 		goto out_free_dma;
728 
729 	for (i = 0; i < npages; i += max) {
730 		if (args.start + (max << PAGE_SHIFT) > end)
731 			args.end = end;
732 		else
733 			args.end = args.start + (max << PAGE_SHIFT);
734 
735 		ret = migrate_vma_setup(&args);
736 		if (ret)
737 			goto out_free_pfns;
738 
739 		if (args.cpages)
740 			nouveau_dmem_migrate_chunk(drm, svmm, &args, dma_addrs,
741 						   pfns);
742 		args.start = args.end;
743 	}
744 
745 	ret = 0;
746 out_free_pfns:
747 	nouveau_pfns_free(pfns);
748 out_free_dma:
749 	kfree(dma_addrs);
750 out_free_dst:
751 	kfree(args.dst);
752 out_free_src:
753 	kfree(args.src);
754 out:
755 	return ret;
756 }
757