1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2020-2021 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 #include <linux/types.h>
24 #include <linux/hmm.h>
25 #include <linux/dma-direction.h>
26 #include <linux/dma-mapping.h>
27 #include <linux/migrate.h>
28 #include "amdgpu_sync.h"
29 #include "amdgpu_object.h"
30 #include "amdgpu_vm.h"
31 #include "amdgpu_mn.h"
32 #include "amdgpu_res_cursor.h"
33 #include "kfd_priv.h"
34 #include "kfd_svm.h"
35 #include "kfd_migrate.h"
36 
37 #ifdef dev_fmt
38 #undef dev_fmt
39 #endif
40 #define dev_fmt(fmt) "kfd_migrate: " fmt
41 
42 static uint64_t
43 svm_migrate_direct_mapping_addr(struct amdgpu_device *adev, uint64_t addr)
44 {
45 	return addr + amdgpu_ttm_domain_start(adev, TTM_PL_VRAM);
46 }
47 
48 static int
49 svm_migrate_gart_map(struct amdgpu_ring *ring, uint64_t npages,
50 		     dma_addr_t *addr, uint64_t *gart_addr, uint64_t flags)
51 {
52 	struct amdgpu_device *adev = ring->adev;
53 	struct amdgpu_job *job;
54 	unsigned int num_dw, num_bytes;
55 	struct dma_fence *fence;
56 	uint64_t src_addr, dst_addr;
57 	uint64_t pte_flags;
58 	void *cpu_addr;
59 	int r;
60 
61 	/* use gart window 0 */
62 	*gart_addr = adev->gmc.gart_start;
63 
64 	num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
65 	num_bytes = npages * 8;
66 
67 	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes,
68 				     AMDGPU_IB_POOL_DELAYED, &job);
69 	if (r)
70 		return r;
71 
72 	src_addr = num_dw * 4;
73 	src_addr += job->ibs[0].gpu_addr;
74 
75 	dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
76 	amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
77 				dst_addr, num_bytes, false);
78 
79 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
80 	WARN_ON(job->ibs[0].length_dw > num_dw);
81 
82 	pte_flags = AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE;
83 	pte_flags |= AMDGPU_PTE_SYSTEM | AMDGPU_PTE_SNOOPED;
84 	if (!(flags & KFD_IOCTL_SVM_FLAG_GPU_RO))
85 		pte_flags |= AMDGPU_PTE_WRITEABLE;
86 	pte_flags |= adev->gart.gart_pte_flags;
87 
88 	cpu_addr = &job->ibs[0].ptr[num_dw];
89 
90 	amdgpu_gart_map(adev, 0, npages, addr, pte_flags, cpu_addr);
91 	r = amdgpu_job_submit(job, &adev->mman.entity,
92 			      AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
93 	if (r)
94 		goto error_free;
95 
96 	dma_fence_put(fence);
97 
98 	return r;
99 
100 error_free:
101 	amdgpu_job_free(job);
102 	return r;
103 }
104 
105 /**
106  * svm_migrate_copy_memory_gart - sdma copy data between ram and vram
107  *
108  * @adev: amdgpu device the sdma ring running
109  * @sys: system DMA pointer to be copied
110  * @vram: vram destination DMA pointer
111  * @npages: number of pages to copy
112  * @direction: enum MIGRATION_COPY_DIR
113  * @mfence: output, sdma fence to signal after sdma is done
114  *
115  * ram address uses GART table continuous entries mapping to ram pages,
116  * vram address uses direct mapping of vram pages, which must have npages
117  * number of continuous pages.
118  * GART update and sdma uses same buf copy function ring, sdma is splited to
119  * multiple GTT_MAX_PAGES transfer, all sdma operations are serialized, wait for
120  * the last sdma finish fence which is returned to check copy memory is done.
121  *
122  * Context: Process context, takes and releases gtt_window_lock
123  *
124  * Return:
125  * 0 - OK, otherwise error code
126  */
127 
128 static int
129 svm_migrate_copy_memory_gart(struct amdgpu_device *adev, dma_addr_t *sys,
130 			     uint64_t *vram, uint64_t npages,
131 			     enum MIGRATION_COPY_DIR direction,
132 			     struct dma_fence **mfence)
133 {
134 	const uint64_t GTT_MAX_PAGES = AMDGPU_GTT_MAX_TRANSFER_SIZE;
135 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
136 	uint64_t gart_s, gart_d;
137 	struct dma_fence *next;
138 	uint64_t size;
139 	int r;
140 
141 	mutex_lock(&adev->mman.gtt_window_lock);
142 
143 	while (npages) {
144 		size = min(GTT_MAX_PAGES, npages);
145 
146 		if (direction == FROM_VRAM_TO_RAM) {
147 			gart_s = svm_migrate_direct_mapping_addr(adev, *vram);
148 			r = svm_migrate_gart_map(ring, size, sys, &gart_d, 0);
149 
150 		} else if (direction == FROM_RAM_TO_VRAM) {
151 			r = svm_migrate_gart_map(ring, size, sys, &gart_s,
152 						 KFD_IOCTL_SVM_FLAG_GPU_RO);
153 			gart_d = svm_migrate_direct_mapping_addr(adev, *vram);
154 		}
155 		if (r) {
156 			dev_err(adev->dev, "fail %d create gart mapping\n", r);
157 			goto out_unlock;
158 		}
159 
160 		r = amdgpu_copy_buffer(ring, gart_s, gart_d, size * PAGE_SIZE,
161 				       NULL, &next, false, true, false);
162 		if (r) {
163 			dev_err(adev->dev, "fail %d to copy memory\n", r);
164 			goto out_unlock;
165 		}
166 
167 		dma_fence_put(*mfence);
168 		*mfence = next;
169 		npages -= size;
170 		if (npages) {
171 			sys += size;
172 			vram += size;
173 		}
174 	}
175 
176 out_unlock:
177 	mutex_unlock(&adev->mman.gtt_window_lock);
178 
179 	return r;
180 }
181 
182 /**
183  * svm_migrate_copy_done - wait for memory copy sdma is done
184  *
185  * @adev: amdgpu device the sdma memory copy is executing on
186  * @mfence: migrate fence
187  *
188  * Wait for dma fence is signaled, if the copy ssplit into multiple sdma
189  * operations, this is the last sdma operation fence.
190  *
191  * Context: called after svm_migrate_copy_memory
192  *
193  * Return:
194  * 0		- success
195  * otherwise	- error code from dma fence signal
196  */
197 static int
198 svm_migrate_copy_done(struct amdgpu_device *adev, struct dma_fence *mfence)
199 {
200 	int r = 0;
201 
202 	if (mfence) {
203 		r = dma_fence_wait(mfence, false);
204 		dma_fence_put(mfence);
205 		pr_debug("sdma copy memory fence done\n");
206 	}
207 
208 	return r;
209 }
210 
211 unsigned long
212 svm_migrate_addr_to_pfn(struct amdgpu_device *adev, unsigned long addr)
213 {
214 	return (addr + adev->kfd.dev->pgmap.range.start) >> PAGE_SHIFT;
215 }
216 
217 static void
218 svm_migrate_get_vram_page(struct svm_range *prange, unsigned long pfn)
219 {
220 	struct page *page;
221 
222 	page = pfn_to_page(pfn);
223 	svm_range_bo_ref(prange->svm_bo);
224 	page->zone_device_data = prange->svm_bo;
225 	lock_page(page);
226 }
227 
228 static void
229 svm_migrate_put_vram_page(struct amdgpu_device *adev, unsigned long addr)
230 {
231 	struct page *page;
232 
233 	page = pfn_to_page(svm_migrate_addr_to_pfn(adev, addr));
234 	unlock_page(page);
235 	put_page(page);
236 }
237 
238 static unsigned long
239 svm_migrate_addr(struct amdgpu_device *adev, struct page *page)
240 {
241 	unsigned long addr;
242 
243 	addr = page_to_pfn(page) << PAGE_SHIFT;
244 	return (addr - adev->kfd.dev->pgmap.range.start);
245 }
246 
247 static struct page *
248 svm_migrate_get_sys_page(struct vm_area_struct *vma, unsigned long addr)
249 {
250 	struct page *page;
251 
252 	page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
253 	if (page)
254 		lock_page(page);
255 
256 	return page;
257 }
258 
259 static void svm_migrate_put_sys_page(unsigned long addr)
260 {
261 	struct page *page;
262 
263 	page = pfn_to_page(addr >> PAGE_SHIFT);
264 	unlock_page(page);
265 	put_page(page);
266 }
267 
268 static unsigned long svm_migrate_successful_pages(struct migrate_vma *migrate)
269 {
270 	unsigned long cpages = 0;
271 	unsigned long i;
272 
273 	for (i = 0; i < migrate->npages; i++) {
274 		if (migrate->src[i] & MIGRATE_PFN_VALID &&
275 		    migrate->src[i] & MIGRATE_PFN_MIGRATE)
276 			cpages++;
277 	}
278 	return cpages;
279 }
280 
281 static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate)
282 {
283 	unsigned long upages = 0;
284 	unsigned long i;
285 
286 	for (i = 0; i < migrate->npages; i++) {
287 		if (migrate->src[i] & MIGRATE_PFN_VALID &&
288 		    !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
289 			upages++;
290 	}
291 	return upages;
292 }
293 
294 static int
295 svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
296 			 struct migrate_vma *migrate, struct dma_fence **mfence,
297 			 dma_addr_t *scratch)
298 {
299 	uint64_t npages = migrate->npages;
300 	struct device *dev = adev->dev;
301 	struct amdgpu_res_cursor cursor;
302 	dma_addr_t *src;
303 	uint64_t *dst;
304 	uint64_t i, j;
305 	int r;
306 
307 	pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start,
308 		 prange->last);
309 
310 	src = scratch;
311 	dst = (uint64_t *)(scratch + npages);
312 
313 	r = svm_range_vram_node_new(adev, prange, true);
314 	if (r) {
315 		dev_dbg(adev->dev, "fail %d to alloc vram\n", r);
316 		goto out;
317 	}
318 
319 	amdgpu_res_first(prange->ttm_res, prange->offset << PAGE_SHIFT,
320 			 npages << PAGE_SHIFT, &cursor);
321 	for (i = j = 0; i < npages; i++) {
322 		struct page *spage;
323 
324 		spage = migrate_pfn_to_page(migrate->src[i]);
325 		if (spage && !is_zone_device_page(spage)) {
326 			dst[i] = cursor.start + (j << PAGE_SHIFT);
327 			migrate->dst[i] = svm_migrate_addr_to_pfn(adev, dst[i]);
328 			svm_migrate_get_vram_page(prange, migrate->dst[i]);
329 			migrate->dst[i] = migrate_pfn(migrate->dst[i]);
330 			src[i] = dma_map_page(dev, spage, 0, PAGE_SIZE,
331 					      DMA_TO_DEVICE);
332 			r = dma_mapping_error(dev, src[i]);
333 			if (r) {
334 				dev_err(adev->dev, "%s: fail %d dma_map_page\n",
335 					__func__, r);
336 				goto out_free_vram_pages;
337 			}
338 		} else {
339 			if (j) {
340 				r = svm_migrate_copy_memory_gart(
341 						adev, src + i - j,
342 						dst + i - j, j,
343 						FROM_RAM_TO_VRAM,
344 						mfence);
345 				if (r)
346 					goto out_free_vram_pages;
347 				amdgpu_res_next(&cursor, (j + 1) << PAGE_SHIFT);
348 				j = 0;
349 			} else {
350 				amdgpu_res_next(&cursor, PAGE_SIZE);
351 			}
352 			continue;
353 		}
354 
355 		pr_debug_ratelimited("dma mapping src to 0x%llx, pfn 0x%lx\n",
356 				     src[i] >> PAGE_SHIFT, page_to_pfn(spage));
357 
358 		if (j >= (cursor.size >> PAGE_SHIFT) - 1 && i < npages - 1) {
359 			r = svm_migrate_copy_memory_gart(adev, src + i - j,
360 							 dst + i - j, j + 1,
361 							 FROM_RAM_TO_VRAM,
362 							 mfence);
363 			if (r)
364 				goto out_free_vram_pages;
365 			amdgpu_res_next(&cursor, (j + 1) * PAGE_SIZE);
366 			j = 0;
367 		} else {
368 			j++;
369 		}
370 	}
371 
372 	r = svm_migrate_copy_memory_gart(adev, src + i - j, dst + i - j, j,
373 					 FROM_RAM_TO_VRAM, mfence);
374 
375 out_free_vram_pages:
376 	if (r) {
377 		pr_debug("failed %d to copy memory to vram\n", r);
378 		while (i--) {
379 			svm_migrate_put_vram_page(adev, dst[i]);
380 			migrate->dst[i] = 0;
381 		}
382 	}
383 
384 #ifdef DEBUG_FORCE_MIXED_DOMAINS
385 	for (i = 0, j = 0; i < npages; i += 4, j++) {
386 		if (j & 1)
387 			continue;
388 		svm_migrate_put_vram_page(adev, dst[i]);
389 		migrate->dst[i] = 0;
390 		svm_migrate_put_vram_page(adev, dst[i + 1]);
391 		migrate->dst[i + 1] = 0;
392 		svm_migrate_put_vram_page(adev, dst[i + 2]);
393 		migrate->dst[i + 2] = 0;
394 		svm_migrate_put_vram_page(adev, dst[i + 3]);
395 		migrate->dst[i + 3] = 0;
396 	}
397 #endif
398 out:
399 	return r;
400 }
401 
402 static long
403 svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
404 			struct vm_area_struct *vma, uint64_t start,
405 			uint64_t end)
406 {
407 	uint64_t npages = (end - start) >> PAGE_SHIFT;
408 	struct kfd_process_device *pdd;
409 	struct dma_fence *mfence = NULL;
410 	struct migrate_vma migrate;
411 	unsigned long cpages = 0;
412 	dma_addr_t *scratch;
413 	void *buf;
414 	int r = -ENOMEM;
415 
416 	memset(&migrate, 0, sizeof(migrate));
417 	migrate.vma = vma;
418 	migrate.start = start;
419 	migrate.end = end;
420 	migrate.flags = MIGRATE_VMA_SELECT_SYSTEM;
421 	migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev);
422 
423 	buf = kvcalloc(npages,
424 		       2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t),
425 		       GFP_KERNEL);
426 	if (!buf)
427 		goto out;
428 
429 	migrate.src = buf;
430 	migrate.dst = migrate.src + npages;
431 	scratch = (dma_addr_t *)(migrate.dst + npages);
432 
433 	r = migrate_vma_setup(&migrate);
434 	if (r) {
435 		dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n",
436 			__func__, r, prange->start, prange->last);
437 		goto out_free;
438 	}
439 
440 	cpages = migrate.cpages;
441 	if (!cpages) {
442 		pr_debug("failed collect migrate sys pages [0x%lx 0x%lx]\n",
443 			 prange->start, prange->last);
444 		goto out_free;
445 	}
446 	if (cpages != npages)
447 		pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n",
448 			 cpages, npages);
449 	else
450 		pr_debug("0x%lx pages migrated\n", cpages);
451 
452 	r = svm_migrate_copy_to_vram(adev, prange, &migrate, &mfence, scratch);
453 	migrate_vma_pages(&migrate);
454 
455 	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
456 		svm_migrate_successful_pages(&migrate), cpages, migrate.npages);
457 
458 	svm_migrate_copy_done(adev, mfence);
459 	migrate_vma_finalize(&migrate);
460 
461 	svm_range_dma_unmap(adev->dev, scratch, 0, npages);
462 	svm_range_free_dma_mappings(prange);
463 
464 out_free:
465 	kvfree(buf);
466 out:
467 	if (!r && cpages) {
468 		pdd = svm_range_get_pdd_by_adev(prange, adev);
469 		if (pdd)
470 			WRITE_ONCE(pdd->page_in, pdd->page_in + cpages);
471 
472 		return cpages;
473 	}
474 	return r;
475 }
476 
477 /**
478  * svm_migrate_ram_to_vram - migrate svm range from system to device
479  * @prange: range structure
480  * @best_loc: the device to migrate to
481  * @mm: the process mm structure
482  *
483  * Context: Process context, caller hold mmap read lock, svms lock, prange lock
484  *
485  * Return:
486  * 0 - OK, otherwise error code
487  */
488 static int
489 svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,
490 			struct mm_struct *mm)
491 {
492 	unsigned long addr, start, end;
493 	struct vm_area_struct *vma;
494 	struct amdgpu_device *adev;
495 	unsigned long cpages = 0;
496 	long r = 0;
497 
498 	if (prange->actual_loc == best_loc) {
499 		pr_debug("svms 0x%p [0x%lx 0x%lx] already on best_loc 0x%x\n",
500 			 prange->svms, prange->start, prange->last, best_loc);
501 		return 0;
502 	}
503 
504 	adev = svm_range_get_adev_by_id(prange, best_loc);
505 	if (!adev) {
506 		pr_debug("failed to get device by id 0x%x\n", best_loc);
507 		return -ENODEV;
508 	}
509 
510 	pr_debug("svms 0x%p [0x%lx 0x%lx] to gpu 0x%x\n", prange->svms,
511 		 prange->start, prange->last, best_loc);
512 
513 	/* FIXME: workaround for page locking bug with invalid pages */
514 	svm_range_prefault(prange, mm, SVM_ADEV_PGMAP_OWNER(adev));
515 
516 	start = prange->start << PAGE_SHIFT;
517 	end = (prange->last + 1) << PAGE_SHIFT;
518 
519 	for (addr = start; addr < end;) {
520 		unsigned long next;
521 
522 		vma = find_vma(mm, addr);
523 		if (!vma || addr < vma->vm_start)
524 			break;
525 
526 		next = min(vma->vm_end, end);
527 		r = svm_migrate_vma_to_vram(adev, prange, vma, addr, next);
528 		if (r < 0) {
529 			pr_debug("failed %ld to migrate\n", r);
530 			break;
531 		} else {
532 			cpages += r;
533 		}
534 		addr = next;
535 	}
536 
537 	if (cpages)
538 		prange->actual_loc = best_loc;
539 
540 	return r < 0 ? r : 0;
541 }
542 
543 static void svm_migrate_page_free(struct page *page)
544 {
545 	struct svm_range_bo *svm_bo = page->zone_device_data;
546 
547 	if (svm_bo) {
548 		pr_debug_ratelimited("ref: %d\n", kref_read(&svm_bo->kref));
549 		svm_range_bo_unref_async(svm_bo);
550 	}
551 }
552 
553 static int
554 svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
555 			struct migrate_vma *migrate, struct dma_fence **mfence,
556 			dma_addr_t *scratch, uint64_t npages)
557 {
558 	struct device *dev = adev->dev;
559 	uint64_t *src;
560 	dma_addr_t *dst;
561 	struct page *dpage;
562 	uint64_t i = 0, j;
563 	uint64_t addr;
564 	int r = 0;
565 
566 	pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start,
567 		 prange->last);
568 
569 	addr = prange->start << PAGE_SHIFT;
570 
571 	src = (uint64_t *)(scratch + npages);
572 	dst = scratch;
573 
574 	for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) {
575 		struct page *spage;
576 
577 		spage = migrate_pfn_to_page(migrate->src[i]);
578 		if (!spage || !is_zone_device_page(spage)) {
579 			pr_debug("invalid page. Could be in CPU already svms 0x%p [0x%lx 0x%lx]\n",
580 				 prange->svms, prange->start, prange->last);
581 			if (j) {
582 				r = svm_migrate_copy_memory_gart(adev, dst + i - j,
583 								 src + i - j, j,
584 								 FROM_VRAM_TO_RAM,
585 								 mfence);
586 				if (r)
587 					goto out_oom;
588 				j = 0;
589 			}
590 			continue;
591 		}
592 		src[i] = svm_migrate_addr(adev, spage);
593 		if (j > 0 && src[i] != src[i - 1] + PAGE_SIZE) {
594 			r = svm_migrate_copy_memory_gart(adev, dst + i - j,
595 							 src + i - j, j,
596 							 FROM_VRAM_TO_RAM,
597 							 mfence);
598 			if (r)
599 				goto out_oom;
600 			j = 0;
601 		}
602 
603 		dpage = svm_migrate_get_sys_page(migrate->vma, addr);
604 		if (!dpage) {
605 			pr_debug("failed get page svms 0x%p [0x%lx 0x%lx]\n",
606 				 prange->svms, prange->start, prange->last);
607 			r = -ENOMEM;
608 			goto out_oom;
609 		}
610 
611 		dst[i] = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_FROM_DEVICE);
612 		r = dma_mapping_error(dev, dst[i]);
613 		if (r) {
614 			dev_err(adev->dev, "%s: fail %d dma_map_page\n", __func__, r);
615 			goto out_oom;
616 		}
617 
618 		pr_debug_ratelimited("dma mapping dst to 0x%llx, pfn 0x%lx\n",
619 				     dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
620 
621 		migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
622 		j++;
623 	}
624 
625 	r = svm_migrate_copy_memory_gart(adev, dst + i - j, src + i - j, j,
626 					 FROM_VRAM_TO_RAM, mfence);
627 
628 out_oom:
629 	if (r) {
630 		pr_debug("failed %d copy to ram\n", r);
631 		while (i--) {
632 			svm_migrate_put_sys_page(dst[i]);
633 			migrate->dst[i] = 0;
634 		}
635 	}
636 
637 	return r;
638 }
639 
640 /**
641  * svm_migrate_vma_to_ram - migrate range inside one vma from device to system
642  *
643  * @adev: amdgpu device to migrate from
644  * @prange: svm range structure
645  * @vma: vm_area_struct that range [start, end] belongs to
646  * @start: range start virtual address in pages
647  * @end: range end virtual address in pages
648  *
649  * Context: Process context, caller hold mmap read lock, prange->migrate_mutex
650  *
651  * Return:
652  *   0 - success with all pages migrated
653  *   negative values - indicate error
654  *   positive values - partial migration, number of pages not migrated
655  */
656 static long
657 svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
658 		       struct vm_area_struct *vma, uint64_t start, uint64_t end)
659 {
660 	uint64_t npages = (end - start) >> PAGE_SHIFT;
661 	unsigned long upages = npages;
662 	unsigned long cpages = 0;
663 	struct kfd_process_device *pdd;
664 	struct dma_fence *mfence = NULL;
665 	struct migrate_vma migrate;
666 	dma_addr_t *scratch;
667 	void *buf;
668 	int r = -ENOMEM;
669 
670 	memset(&migrate, 0, sizeof(migrate));
671 	migrate.vma = vma;
672 	migrate.start = start;
673 	migrate.end = end;
674 	migrate.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE;
675 	migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev);
676 
677 	buf = kvcalloc(npages,
678 		       2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t),
679 		       GFP_KERNEL);
680 
681 	if (!buf)
682 		goto out;
683 
684 	migrate.src = buf;
685 	migrate.dst = migrate.src + npages;
686 	scratch = (dma_addr_t *)(migrate.dst + npages);
687 
688 	r = migrate_vma_setup(&migrate);
689 	if (r) {
690 		dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n",
691 			__func__, r, prange->start, prange->last);
692 		goto out_free;
693 	}
694 
695 	cpages = migrate.cpages;
696 	if (!cpages) {
697 		pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n",
698 			 prange->start, prange->last);
699 		upages = svm_migrate_unsuccessful_pages(&migrate);
700 		goto out_free;
701 	}
702 	if (cpages != npages)
703 		pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n",
704 			 cpages, npages);
705 	else
706 		pr_debug("0x%lx pages migrated\n", cpages);
707 
708 	r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
709 				    scratch, npages);
710 	migrate_vma_pages(&migrate);
711 
712 	upages = svm_migrate_unsuccessful_pages(&migrate);
713 	pr_debug("unsuccessful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
714 		 upages, cpages, migrate.npages);
715 
716 	svm_migrate_copy_done(adev, mfence);
717 	migrate_vma_finalize(&migrate);
718 	svm_range_dma_unmap(adev->dev, scratch, 0, npages);
719 
720 out_free:
721 	kvfree(buf);
722 out:
723 	if (!r && cpages) {
724 		pdd = svm_range_get_pdd_by_adev(prange, adev);
725 		if (pdd)
726 			WRITE_ONCE(pdd->page_out, pdd->page_out + cpages);
727 	}
728 	return r ? r : upages;
729 }
730 
731 /**
732  * svm_migrate_vram_to_ram - migrate svm range from device to system
733  * @prange: range structure
734  * @mm: process mm, use current->mm if NULL
735  *
736  * Context: Process context, caller hold mmap read lock, prange->migrate_mutex
737  *
738  * Return:
739  * 0 - OK, otherwise error code
740  */
741 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)
742 {
743 	struct amdgpu_device *adev;
744 	struct vm_area_struct *vma;
745 	unsigned long addr;
746 	unsigned long start;
747 	unsigned long end;
748 	unsigned long upages = 0;
749 	long r = 0;
750 
751 	if (!prange->actual_loc) {
752 		pr_debug("[0x%lx 0x%lx] already migrated to ram\n",
753 			 prange->start, prange->last);
754 		return 0;
755 	}
756 
757 	adev = svm_range_get_adev_by_id(prange, prange->actual_loc);
758 	if (!adev) {
759 		pr_debug("failed to get device by id 0x%x\n",
760 			 prange->actual_loc);
761 		return -ENODEV;
762 	}
763 
764 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n",
765 		 prange->svms, prange, prange->start, prange->last,
766 		 prange->actual_loc);
767 
768 	start = prange->start << PAGE_SHIFT;
769 	end = (prange->last + 1) << PAGE_SHIFT;
770 
771 	for (addr = start; addr < end;) {
772 		unsigned long next;
773 
774 		vma = find_vma(mm, addr);
775 		if (!vma || addr < vma->vm_start) {
776 			pr_debug("failed to find vma for prange %p\n", prange);
777 			r = -EFAULT;
778 			break;
779 		}
780 
781 		next = min(vma->vm_end, end);
782 		r = svm_migrate_vma_to_ram(adev, prange, vma, addr, next);
783 		if (r < 0) {
784 			pr_debug("failed %ld to migrate prange %p\n", r, prange);
785 			break;
786 		} else {
787 			upages += r;
788 		}
789 		addr = next;
790 	}
791 
792 	if (r >= 0 && !upages) {
793 		svm_range_vram_node_free(prange);
794 		prange->actual_loc = 0;
795 	}
796 
797 	return r < 0 ? r : 0;
798 }
799 
800 /**
801  * svm_migrate_vram_to_vram - migrate svm range from device to device
802  * @prange: range structure
803  * @best_loc: the device to migrate to
804  * @mm: process mm, use current->mm if NULL
805  *
806  * Context: Process context, caller hold mmap read lock, svms lock, prange lock
807  *
808  * Return:
809  * 0 - OK, otherwise error code
810  */
811 static int
812 svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,
813 			 struct mm_struct *mm)
814 {
815 	int r, retries = 3;
816 
817 	/*
818 	 * TODO: for both devices with PCIe large bar or on same xgmi hive, skip
819 	 * system memory as migration bridge
820 	 */
821 
822 	pr_debug("from gpu 0x%x to gpu 0x%x\n", prange->actual_loc, best_loc);
823 
824 	do {
825 		r = svm_migrate_vram_to_ram(prange, mm);
826 		if (r)
827 			return r;
828 	} while (prange->actual_loc && --retries);
829 
830 	if (prange->actual_loc)
831 		return -EDEADLK;
832 
833 	return svm_migrate_ram_to_vram(prange, best_loc, mm);
834 }
835 
836 int
837 svm_migrate_to_vram(struct svm_range *prange, uint32_t best_loc,
838 		    struct mm_struct *mm)
839 {
840 	if  (!prange->actual_loc)
841 		return svm_migrate_ram_to_vram(prange, best_loc, mm);
842 	else
843 		return svm_migrate_vram_to_vram(prange, best_loc, mm);
844 
845 }
846 
847 /**
848  * svm_migrate_to_ram - CPU page fault handler
849  * @vmf: CPU vm fault vma, address
850  *
851  * Context: vm fault handler, caller holds the mmap read lock
852  *
853  * Return:
854  * 0 - OK
855  * VM_FAULT_SIGBUS - notice application to have SIGBUS page fault
856  */
857 static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf)
858 {
859 	unsigned long addr = vmf->address;
860 	struct vm_area_struct *vma;
861 	enum svm_work_list_ops op;
862 	struct svm_range *parent;
863 	struct svm_range *prange;
864 	struct kfd_process *p;
865 	struct mm_struct *mm;
866 	int r = 0;
867 
868 	vma = vmf->vma;
869 	mm = vma->vm_mm;
870 
871 	p = kfd_lookup_process_by_mm(vma->vm_mm);
872 	if (!p) {
873 		pr_debug("failed find process at fault address 0x%lx\n", addr);
874 		return VM_FAULT_SIGBUS;
875 	}
876 	if (READ_ONCE(p->svms.faulting_task) == current) {
877 		pr_debug("skipping ram migration\n");
878 		kfd_unref_process(p);
879 		return 0;
880 	}
881 	addr >>= PAGE_SHIFT;
882 	pr_debug("CPU page fault svms 0x%p address 0x%lx\n", &p->svms, addr);
883 
884 	mutex_lock(&p->svms.lock);
885 
886 	prange = svm_range_from_addr(&p->svms, addr, &parent);
887 	if (!prange) {
888 		pr_debug("cannot find svm range at 0x%lx\n", addr);
889 		r = -EFAULT;
890 		goto out;
891 	}
892 
893 	mutex_lock(&parent->migrate_mutex);
894 	if (prange != parent)
895 		mutex_lock_nested(&prange->migrate_mutex, 1);
896 
897 	if (!prange->actual_loc)
898 		goto out_unlock_prange;
899 
900 	svm_range_lock(parent);
901 	if (prange != parent)
902 		mutex_lock_nested(&prange->lock, 1);
903 	r = svm_range_split_by_granularity(p, mm, addr, parent, prange);
904 	if (prange != parent)
905 		mutex_unlock(&prange->lock);
906 	svm_range_unlock(parent);
907 	if (r) {
908 		pr_debug("failed %d to split range by granularity\n", r);
909 		goto out_unlock_prange;
910 	}
911 
912 	r = svm_migrate_vram_to_ram(prange, mm);
913 	if (r)
914 		pr_debug("failed %d migrate 0x%p [0x%lx 0x%lx] to ram\n", r,
915 			 prange, prange->start, prange->last);
916 
917 	/* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
918 	if (p->xnack_enabled && parent == prange)
919 		op = SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP;
920 	else
921 		op = SVM_OP_UPDATE_RANGE_NOTIFIER;
922 	svm_range_add_list_work(&p->svms, parent, mm, op);
923 	schedule_deferred_list_work(&p->svms);
924 
925 out_unlock_prange:
926 	if (prange != parent)
927 		mutex_unlock(&prange->migrate_mutex);
928 	mutex_unlock(&parent->migrate_mutex);
929 out:
930 	mutex_unlock(&p->svms.lock);
931 	kfd_unref_process(p);
932 
933 	pr_debug("CPU fault svms 0x%p address 0x%lx done\n", &p->svms, addr);
934 
935 	return r ? VM_FAULT_SIGBUS : 0;
936 }
937 
938 static const struct dev_pagemap_ops svm_migrate_pgmap_ops = {
939 	.page_free		= svm_migrate_page_free,
940 	.migrate_to_ram		= svm_migrate_to_ram,
941 };
942 
943 /* Each VRAM page uses sizeof(struct page) on system memory */
944 #define SVM_HMM_PAGE_STRUCT_SIZE(size) ((size)/PAGE_SIZE * sizeof(struct page))
945 
946 int svm_migrate_init(struct amdgpu_device *adev)
947 {
948 	struct kfd_dev *kfddev = adev->kfd.dev;
949 	struct dev_pagemap *pgmap;
950 	struct resource *res;
951 	unsigned long size;
952 	void *r;
953 
954 	/* Page migration works on Vega10 or newer */
955 	if (!KFD_IS_SOC15(kfddev))
956 		return -EINVAL;
957 
958 	pgmap = &kfddev->pgmap;
959 	memset(pgmap, 0, sizeof(*pgmap));
960 
961 	/* TODO: register all vram to HMM for now.
962 	 * should remove reserved size
963 	 */
964 	size = ALIGN(adev->gmc.real_vram_size, 2ULL << 20);
965 	res = devm_request_free_mem_region(adev->dev, &iomem_resource, size);
966 	if (IS_ERR(res))
967 		return -ENOMEM;
968 
969 	pgmap->type = MEMORY_DEVICE_PRIVATE;
970 	pgmap->nr_range = 1;
971 	pgmap->range.start = res->start;
972 	pgmap->range.end = res->end;
973 	pgmap->ops = &svm_migrate_pgmap_ops;
974 	pgmap->owner = SVM_ADEV_PGMAP_OWNER(adev);
975 	pgmap->flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE;
976 
977 	/* Device manager releases device-specific resources, memory region and
978 	 * pgmap when driver disconnects from device.
979 	 */
980 	r = devm_memremap_pages(adev->dev, pgmap);
981 	if (IS_ERR(r)) {
982 		pr_err("failed to register HMM device memory\n");
983 
984 		/* Disable SVM support capability */
985 		pgmap->type = 0;
986 		devm_release_mem_region(adev->dev, res->start, resource_size(res));
987 		return PTR_ERR(r);
988 	}
989 
990 	pr_debug("reserve %ldMB system memory for VRAM pages struct\n",
991 		 SVM_HMM_PAGE_STRUCT_SIZE(size) >> 20);
992 
993 	amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size));
994 
995 	pr_info("HMM registered %ldMB device memory\n", size >> 20);
996 
997 	return 0;
998 }
999