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 #include "kfd_smi_events.h"
37 
38 #ifdef dev_fmt
39 #undef dev_fmt
40 #endif
41 #define dev_fmt(fmt) "kfd_migrate: " fmt
42 
43 static uint64_t
44 svm_migrate_direct_mapping_addr(struct amdgpu_device *adev, uint64_t addr)
45 {
46 	return addr + amdgpu_ttm_domain_start(adev, TTM_PL_VRAM);
47 }
48 
49 static int
50 svm_migrate_gart_map(struct amdgpu_ring *ring, uint64_t npages,
51 		     dma_addr_t *addr, uint64_t *gart_addr, uint64_t flags)
52 {
53 	struct amdgpu_device *adev = ring->adev;
54 	struct amdgpu_job *job;
55 	unsigned int num_dw, num_bytes;
56 	struct dma_fence *fence;
57 	uint64_t src_addr, dst_addr;
58 	uint64_t pte_flags;
59 	void *cpu_addr;
60 	int r;
61 
62 	/* use gart window 0 */
63 	*gart_addr = adev->gmc.gart_start;
64 
65 	num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
66 	num_bytes = npages * 8;
67 
68 	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes,
69 				     AMDGPU_IB_POOL_DELAYED, &job);
70 	if (r)
71 		return r;
72 
73 	src_addr = num_dw * 4;
74 	src_addr += job->ibs[0].gpu_addr;
75 
76 	dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
77 	amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
78 				dst_addr, num_bytes, false);
79 
80 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
81 	WARN_ON(job->ibs[0].length_dw > num_dw);
82 
83 	pte_flags = AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE;
84 	pte_flags |= AMDGPU_PTE_SYSTEM | AMDGPU_PTE_SNOOPED;
85 	if (!(flags & KFD_IOCTL_SVM_FLAG_GPU_RO))
86 		pte_flags |= AMDGPU_PTE_WRITEABLE;
87 	pte_flags |= adev->gart.gart_pte_flags;
88 
89 	cpu_addr = &job->ibs[0].ptr[num_dw];
90 
91 	amdgpu_gart_map(adev, 0, npages, addr, pte_flags, cpu_addr);
92 	r = amdgpu_job_submit(job, &adev->mman.entity,
93 			      AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
94 	if (r)
95 		goto error_free;
96 
97 	dma_fence_put(fence);
98 
99 	return r;
100 
101 error_free:
102 	amdgpu_job_free(job);
103 	return r;
104 }
105 
106 /**
107  * svm_migrate_copy_memory_gart - sdma copy data between ram and vram
108  *
109  * @adev: amdgpu device the sdma ring running
110  * @sys: system DMA pointer to be copied
111  * @vram: vram destination DMA pointer
112  * @npages: number of pages to copy
113  * @direction: enum MIGRATION_COPY_DIR
114  * @mfence: output, sdma fence to signal after sdma is done
115  *
116  * ram address uses GART table continuous entries mapping to ram pages,
117  * vram address uses direct mapping of vram pages, which must have npages
118  * number of continuous pages.
119  * GART update and sdma uses same buf copy function ring, sdma is splited to
120  * multiple GTT_MAX_PAGES transfer, all sdma operations are serialized, wait for
121  * the last sdma finish fence which is returned to check copy memory is done.
122  *
123  * Context: Process context, takes and releases gtt_window_lock
124  *
125  * Return:
126  * 0 - OK, otherwise error code
127  */
128 
129 static int
130 svm_migrate_copy_memory_gart(struct amdgpu_device *adev, dma_addr_t *sys,
131 			     uint64_t *vram, uint64_t npages,
132 			     enum MIGRATION_COPY_DIR direction,
133 			     struct dma_fence **mfence)
134 {
135 	const uint64_t GTT_MAX_PAGES = AMDGPU_GTT_MAX_TRANSFER_SIZE;
136 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
137 	uint64_t gart_s, gart_d;
138 	struct dma_fence *next;
139 	uint64_t size;
140 	int r;
141 
142 	mutex_lock(&adev->mman.gtt_window_lock);
143 
144 	while (npages) {
145 		size = min(GTT_MAX_PAGES, npages);
146 
147 		if (direction == FROM_VRAM_TO_RAM) {
148 			gart_s = svm_migrate_direct_mapping_addr(adev, *vram);
149 			r = svm_migrate_gart_map(ring, size, sys, &gart_d, 0);
150 
151 		} else if (direction == FROM_RAM_TO_VRAM) {
152 			r = svm_migrate_gart_map(ring, size, sys, &gart_s,
153 						 KFD_IOCTL_SVM_FLAG_GPU_RO);
154 			gart_d = svm_migrate_direct_mapping_addr(adev, *vram);
155 		}
156 		if (r) {
157 			dev_err(adev->dev, "fail %d create gart mapping\n", r);
158 			goto out_unlock;
159 		}
160 
161 		r = amdgpu_copy_buffer(ring, gart_s, gart_d, size * PAGE_SIZE,
162 				       NULL, &next, false, true, false);
163 		if (r) {
164 			dev_err(adev->dev, "fail %d to copy memory\n", r);
165 			goto out_unlock;
166 		}
167 
168 		dma_fence_put(*mfence);
169 		*mfence = next;
170 		npages -= size;
171 		if (npages) {
172 			sys += size;
173 			vram += size;
174 		}
175 	}
176 
177 out_unlock:
178 	mutex_unlock(&adev->mman.gtt_window_lock);
179 
180 	return r;
181 }
182 
183 /**
184  * svm_migrate_copy_done - wait for memory copy sdma is done
185  *
186  * @adev: amdgpu device the sdma memory copy is executing on
187  * @mfence: migrate fence
188  *
189  * Wait for dma fence is signaled, if the copy ssplit into multiple sdma
190  * operations, this is the last sdma operation fence.
191  *
192  * Context: called after svm_migrate_copy_memory
193  *
194  * Return:
195  * 0		- success
196  * otherwise	- error code from dma fence signal
197  */
198 static int
199 svm_migrate_copy_done(struct amdgpu_device *adev, struct dma_fence *mfence)
200 {
201 	int r = 0;
202 
203 	if (mfence) {
204 		r = dma_fence_wait(mfence, false);
205 		dma_fence_put(mfence);
206 		pr_debug("sdma copy memory fence done\n");
207 	}
208 
209 	return r;
210 }
211 
212 unsigned long
213 svm_migrate_addr_to_pfn(struct amdgpu_device *adev, unsigned long addr)
214 {
215 	return (addr + adev->kfd.dev->pgmap.range.start) >> PAGE_SHIFT;
216 }
217 
218 static void
219 svm_migrate_get_vram_page(struct svm_range *prange, unsigned long pfn)
220 {
221 	struct page *page;
222 
223 	page = pfn_to_page(pfn);
224 	svm_range_bo_ref(prange->svm_bo);
225 	page->zone_device_data = prange->svm_bo;
226 	zone_device_page_init(page);
227 }
228 
229 static void
230 svm_migrate_put_vram_page(struct amdgpu_device *adev, unsigned long addr)
231 {
232 	struct page *page;
233 
234 	page = pfn_to_page(svm_migrate_addr_to_pfn(adev, addr));
235 	unlock_page(page);
236 	put_page(page);
237 }
238 
239 static unsigned long
240 svm_migrate_addr(struct amdgpu_device *adev, struct page *page)
241 {
242 	unsigned long addr;
243 
244 	addr = page_to_pfn(page) << PAGE_SHIFT;
245 	return (addr - adev->kfd.dev->pgmap.range.start);
246 }
247 
248 static struct page *
249 svm_migrate_get_sys_page(struct vm_area_struct *vma, unsigned long addr)
250 {
251 	struct page *page;
252 
253 	page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
254 	if (page)
255 		lock_page(page);
256 
257 	return page;
258 }
259 
260 static void svm_migrate_put_sys_page(unsigned long addr)
261 {
262 	struct page *page;
263 
264 	page = pfn_to_page(addr >> PAGE_SHIFT);
265 	unlock_page(page);
266 	put_page(page);
267 }
268 
269 static unsigned long svm_migrate_successful_pages(struct migrate_vma *migrate)
270 {
271 	unsigned long cpages = 0;
272 	unsigned long i;
273 
274 	for (i = 0; i < migrate->npages; i++) {
275 		if (migrate->src[i] & MIGRATE_PFN_VALID &&
276 		    migrate->src[i] & MIGRATE_PFN_MIGRATE)
277 			cpages++;
278 	}
279 	return cpages;
280 }
281 
282 static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate)
283 {
284 	unsigned long upages = 0;
285 	unsigned long i;
286 
287 	for (i = 0; i < migrate->npages; i++) {
288 		if (migrate->src[i] & MIGRATE_PFN_VALID &&
289 		    !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
290 			upages++;
291 	}
292 	return upages;
293 }
294 
295 static int
296 svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
297 			 struct migrate_vma *migrate, struct dma_fence **mfence,
298 			 dma_addr_t *scratch)
299 {
300 	uint64_t npages = migrate->npages;
301 	struct device *dev = adev->dev;
302 	struct amdgpu_res_cursor cursor;
303 	dma_addr_t *src;
304 	uint64_t *dst;
305 	uint64_t i, j;
306 	int r;
307 
308 	pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start,
309 		 prange->last);
310 
311 	src = scratch;
312 	dst = (uint64_t *)(scratch + npages);
313 
314 	r = svm_range_vram_node_new(adev, prange, true);
315 	if (r) {
316 		dev_dbg(adev->dev, "fail %d to alloc vram\n", r);
317 		goto out;
318 	}
319 
320 	amdgpu_res_first(prange->ttm_res, prange->offset << PAGE_SHIFT,
321 			 npages << PAGE_SHIFT, &cursor);
322 	for (i = j = 0; i < npages; i++) {
323 		struct page *spage;
324 
325 		dst[i] = cursor.start + (j << PAGE_SHIFT);
326 		migrate->dst[i] = svm_migrate_addr_to_pfn(adev, dst[i]);
327 		svm_migrate_get_vram_page(prange, migrate->dst[i]);
328 		migrate->dst[i] = migrate_pfn(migrate->dst[i]);
329 
330 		spage = migrate_pfn_to_page(migrate->src[i]);
331 		if (spage && !is_zone_device_page(spage)) {
332 			src[i] = dma_map_page(dev, spage, 0, PAGE_SIZE,
333 					      DMA_TO_DEVICE);
334 			r = dma_mapping_error(dev, src[i]);
335 			if (r) {
336 				dev_err(adev->dev, "%s: fail %d dma_map_page\n",
337 					__func__, r);
338 				goto out_free_vram_pages;
339 			}
340 		} else {
341 			if (j) {
342 				r = svm_migrate_copy_memory_gart(
343 						adev, src + i - j,
344 						dst + i - j, j,
345 						FROM_RAM_TO_VRAM,
346 						mfence);
347 				if (r)
348 					goto out_free_vram_pages;
349 				amdgpu_res_next(&cursor, (j + 1) << PAGE_SHIFT);
350 				j = 0;
351 			} else {
352 				amdgpu_res_next(&cursor, PAGE_SIZE);
353 			}
354 			continue;
355 		}
356 
357 		pr_debug_ratelimited("dma mapping src to 0x%llx, pfn 0x%lx\n",
358 				     src[i] >> PAGE_SHIFT, page_to_pfn(spage));
359 
360 		if (j >= (cursor.size >> PAGE_SHIFT) - 1 && i < npages - 1) {
361 			r = svm_migrate_copy_memory_gart(adev, src + i - j,
362 							 dst + i - j, j + 1,
363 							 FROM_RAM_TO_VRAM,
364 							 mfence);
365 			if (r)
366 				goto out_free_vram_pages;
367 			amdgpu_res_next(&cursor, (j + 1) * PAGE_SIZE);
368 			j = 0;
369 		} else {
370 			j++;
371 		}
372 	}
373 
374 	r = svm_migrate_copy_memory_gart(adev, src + i - j, dst + i - j, j,
375 					 FROM_RAM_TO_VRAM, mfence);
376 
377 out_free_vram_pages:
378 	if (r) {
379 		pr_debug("failed %d to copy memory to vram\n", r);
380 		while (i--) {
381 			svm_migrate_put_vram_page(adev, dst[i]);
382 			migrate->dst[i] = 0;
383 		}
384 	}
385 
386 #ifdef DEBUG_FORCE_MIXED_DOMAINS
387 	for (i = 0, j = 0; i < npages; i += 4, j++) {
388 		if (j & 1)
389 			continue;
390 		svm_migrate_put_vram_page(adev, dst[i]);
391 		migrate->dst[i] = 0;
392 		svm_migrate_put_vram_page(adev, dst[i + 1]);
393 		migrate->dst[i + 1] = 0;
394 		svm_migrate_put_vram_page(adev, dst[i + 2]);
395 		migrate->dst[i + 2] = 0;
396 		svm_migrate_put_vram_page(adev, dst[i + 3]);
397 		migrate->dst[i + 3] = 0;
398 	}
399 #endif
400 out:
401 	return r;
402 }
403 
404 static long
405 svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
406 			struct vm_area_struct *vma, uint64_t start,
407 			uint64_t end, uint32_t trigger)
408 {
409 	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
410 	uint64_t npages = (end - start) >> PAGE_SHIFT;
411 	struct kfd_process_device *pdd;
412 	struct dma_fence *mfence = NULL;
413 	struct migrate_vma migrate = { 0 };
414 	unsigned long cpages = 0;
415 	dma_addr_t *scratch;
416 	void *buf;
417 	int r = -ENOMEM;
418 
419 	memset(&migrate, 0, sizeof(migrate));
420 	migrate.vma = vma;
421 	migrate.start = start;
422 	migrate.end = end;
423 	migrate.flags = MIGRATE_VMA_SELECT_SYSTEM;
424 	migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev);
425 
426 	buf = kvcalloc(npages,
427 		       2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t),
428 		       GFP_KERNEL);
429 	if (!buf)
430 		goto out;
431 
432 	migrate.src = buf;
433 	migrate.dst = migrate.src + npages;
434 	scratch = (dma_addr_t *)(migrate.dst + npages);
435 
436 	kfd_smi_event_migration_start(adev->kfd.dev, p->lead_thread->pid,
437 				      start >> PAGE_SHIFT, end >> PAGE_SHIFT,
438 				      0, adev->kfd.dev->id, prange->prefetch_loc,
439 				      prange->preferred_loc, trigger);
440 
441 	r = migrate_vma_setup(&migrate);
442 	if (r) {
443 		dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n",
444 			__func__, r, prange->start, prange->last);
445 		goto out_free;
446 	}
447 
448 	cpages = migrate.cpages;
449 	if (!cpages) {
450 		pr_debug("failed collect migrate sys pages [0x%lx 0x%lx]\n",
451 			 prange->start, prange->last);
452 		goto out_free;
453 	}
454 	if (cpages != npages)
455 		pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n",
456 			 cpages, npages);
457 	else
458 		pr_debug("0x%lx pages migrated\n", cpages);
459 
460 	r = svm_migrate_copy_to_vram(adev, prange, &migrate, &mfence, scratch);
461 	migrate_vma_pages(&migrate);
462 
463 	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
464 		svm_migrate_successful_pages(&migrate), cpages, migrate.npages);
465 
466 	svm_migrate_copy_done(adev, mfence);
467 	migrate_vma_finalize(&migrate);
468 
469 	kfd_smi_event_migration_end(adev->kfd.dev, p->lead_thread->pid,
470 				    start >> PAGE_SHIFT, end >> PAGE_SHIFT,
471 				    0, adev->kfd.dev->id, trigger);
472 
473 	svm_range_dma_unmap(adev->dev, scratch, 0, npages);
474 	svm_range_free_dma_mappings(prange);
475 
476 out_free:
477 	kvfree(buf);
478 out:
479 	if (!r && cpages) {
480 		pdd = svm_range_get_pdd_by_adev(prange, adev);
481 		if (pdd)
482 			WRITE_ONCE(pdd->page_in, pdd->page_in + cpages);
483 
484 		return cpages;
485 	}
486 	return r;
487 }
488 
489 /**
490  * svm_migrate_ram_to_vram - migrate svm range from system to device
491  * @prange: range structure
492  * @best_loc: the device to migrate to
493  * @mm: the process mm structure
494  * @trigger: reason of migration
495  *
496  * Context: Process context, caller hold mmap read lock, svms lock, prange lock
497  *
498  * Return:
499  * 0 - OK, otherwise error code
500  */
501 static int
502 svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,
503 			struct mm_struct *mm, uint32_t trigger)
504 {
505 	unsigned long addr, start, end;
506 	struct vm_area_struct *vma;
507 	struct amdgpu_device *adev;
508 	unsigned long cpages = 0;
509 	long r = 0;
510 
511 	if (prange->actual_loc == best_loc) {
512 		pr_debug("svms 0x%p [0x%lx 0x%lx] already on best_loc 0x%x\n",
513 			 prange->svms, prange->start, prange->last, best_loc);
514 		return 0;
515 	}
516 
517 	adev = svm_range_get_adev_by_id(prange, best_loc);
518 	if (!adev) {
519 		pr_debug("failed to get device by id 0x%x\n", best_loc);
520 		return -ENODEV;
521 	}
522 
523 	pr_debug("svms 0x%p [0x%lx 0x%lx] to gpu 0x%x\n", prange->svms,
524 		 prange->start, prange->last, best_loc);
525 
526 	start = prange->start << PAGE_SHIFT;
527 	end = (prange->last + 1) << PAGE_SHIFT;
528 
529 	for (addr = start; addr < end;) {
530 		unsigned long next;
531 
532 		vma = find_vma(mm, addr);
533 		if (!vma || addr < vma->vm_start)
534 			break;
535 
536 		next = min(vma->vm_end, end);
537 		r = svm_migrate_vma_to_vram(adev, prange, vma, addr, next, trigger);
538 		if (r < 0) {
539 			pr_debug("failed %ld to migrate\n", r);
540 			break;
541 		} else {
542 			cpages += r;
543 		}
544 		addr = next;
545 	}
546 
547 	if (cpages)
548 		prange->actual_loc = best_loc;
549 
550 	return r < 0 ? r : 0;
551 }
552 
553 static void svm_migrate_page_free(struct page *page)
554 {
555 	struct svm_range_bo *svm_bo = page->zone_device_data;
556 
557 	if (svm_bo) {
558 		pr_debug_ratelimited("ref: %d\n", kref_read(&svm_bo->kref));
559 		svm_range_bo_unref_async(svm_bo);
560 	}
561 }
562 
563 static int
564 svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
565 			struct migrate_vma *migrate, struct dma_fence **mfence,
566 			dma_addr_t *scratch, uint64_t npages)
567 {
568 	struct device *dev = adev->dev;
569 	uint64_t *src;
570 	dma_addr_t *dst;
571 	struct page *dpage;
572 	uint64_t i = 0, j;
573 	uint64_t addr;
574 	int r = 0;
575 
576 	pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start,
577 		 prange->last);
578 
579 	addr = prange->start << PAGE_SHIFT;
580 
581 	src = (uint64_t *)(scratch + npages);
582 	dst = scratch;
583 
584 	for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) {
585 		struct page *spage;
586 
587 		spage = migrate_pfn_to_page(migrate->src[i]);
588 		if (!spage || !is_zone_device_page(spage)) {
589 			pr_debug("invalid page. Could be in CPU already svms 0x%p [0x%lx 0x%lx]\n",
590 				 prange->svms, prange->start, prange->last);
591 			if (j) {
592 				r = svm_migrate_copy_memory_gart(adev, dst + i - j,
593 								 src + i - j, j,
594 								 FROM_VRAM_TO_RAM,
595 								 mfence);
596 				if (r)
597 					goto out_oom;
598 				j = 0;
599 			}
600 			continue;
601 		}
602 		src[i] = svm_migrate_addr(adev, spage);
603 		if (j > 0 && src[i] != src[i - 1] + PAGE_SIZE) {
604 			r = svm_migrate_copy_memory_gart(adev, dst + i - j,
605 							 src + i - j, j,
606 							 FROM_VRAM_TO_RAM,
607 							 mfence);
608 			if (r)
609 				goto out_oom;
610 			j = 0;
611 		}
612 
613 		dpage = svm_migrate_get_sys_page(migrate->vma, addr);
614 		if (!dpage) {
615 			pr_debug("failed get page svms 0x%p [0x%lx 0x%lx]\n",
616 				 prange->svms, prange->start, prange->last);
617 			r = -ENOMEM;
618 			goto out_oom;
619 		}
620 
621 		dst[i] = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_FROM_DEVICE);
622 		r = dma_mapping_error(dev, dst[i]);
623 		if (r) {
624 			dev_err(adev->dev, "%s: fail %d dma_map_page\n", __func__, r);
625 			goto out_oom;
626 		}
627 
628 		pr_debug_ratelimited("dma mapping dst to 0x%llx, pfn 0x%lx\n",
629 				     dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
630 
631 		migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
632 		j++;
633 	}
634 
635 	r = svm_migrate_copy_memory_gart(adev, dst + i - j, src + i - j, j,
636 					 FROM_VRAM_TO_RAM, mfence);
637 
638 out_oom:
639 	if (r) {
640 		pr_debug("failed %d copy to ram\n", r);
641 		while (i--) {
642 			svm_migrate_put_sys_page(dst[i]);
643 			migrate->dst[i] = 0;
644 		}
645 	}
646 
647 	return r;
648 }
649 
650 /**
651  * svm_migrate_vma_to_ram - migrate range inside one vma from device to system
652  *
653  * @adev: amdgpu device to migrate from
654  * @prange: svm range structure
655  * @vma: vm_area_struct that range [start, end] belongs to
656  * @start: range start virtual address in pages
657  * @end: range end virtual address in pages
658  *
659  * Context: Process context, caller hold mmap read lock, prange->migrate_mutex
660  *
661  * Return:
662  *   0 - success with all pages migrated
663  *   negative values - indicate error
664  *   positive values - partial migration, number of pages not migrated
665  */
666 static long
667 svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
668 		       struct vm_area_struct *vma, uint64_t start, uint64_t end,
669 		       uint32_t trigger, struct page *fault_page)
670 {
671 	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
672 	uint64_t npages = (end - start) >> PAGE_SHIFT;
673 	unsigned long upages = npages;
674 	unsigned long cpages = 0;
675 	struct kfd_process_device *pdd;
676 	struct dma_fence *mfence = NULL;
677 	struct migrate_vma migrate = { 0 };
678 	dma_addr_t *scratch;
679 	void *buf;
680 	int r = -ENOMEM;
681 
682 	memset(&migrate, 0, sizeof(migrate));
683 	migrate.vma = vma;
684 	migrate.start = start;
685 	migrate.end = end;
686 	migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev);
687 	if (adev->gmc.xgmi.connected_to_cpu)
688 		migrate.flags = MIGRATE_VMA_SELECT_DEVICE_COHERENT;
689 	else
690 		migrate.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE;
691 
692 	buf = kvcalloc(npages,
693 		       2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t),
694 		       GFP_KERNEL);
695 	if (!buf)
696 		goto out;
697 
698 	migrate.src = buf;
699 	migrate.dst = migrate.src + npages;
700 	migrate.fault_page = fault_page;
701 	scratch = (dma_addr_t *)(migrate.dst + npages);
702 
703 	kfd_smi_event_migration_start(adev->kfd.dev, p->lead_thread->pid,
704 				      start >> PAGE_SHIFT, end >> PAGE_SHIFT,
705 				      adev->kfd.dev->id, 0, prange->prefetch_loc,
706 				      prange->preferred_loc, trigger);
707 
708 	r = migrate_vma_setup(&migrate);
709 	if (r) {
710 		dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n",
711 			__func__, r, prange->start, prange->last);
712 		goto out_free;
713 	}
714 
715 	cpages = migrate.cpages;
716 	if (!cpages) {
717 		pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n",
718 			 prange->start, prange->last);
719 		upages = svm_migrate_unsuccessful_pages(&migrate);
720 		goto out_free;
721 	}
722 	if (cpages != npages)
723 		pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n",
724 			 cpages, npages);
725 	else
726 		pr_debug("0x%lx pages migrated\n", cpages);
727 
728 	r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
729 				    scratch, npages);
730 	migrate_vma_pages(&migrate);
731 
732 	upages = svm_migrate_unsuccessful_pages(&migrate);
733 	pr_debug("unsuccessful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
734 		 upages, cpages, migrate.npages);
735 
736 	svm_migrate_copy_done(adev, mfence);
737 	migrate_vma_finalize(&migrate);
738 
739 	kfd_smi_event_migration_end(adev->kfd.dev, p->lead_thread->pid,
740 				    start >> PAGE_SHIFT, end >> PAGE_SHIFT,
741 				    adev->kfd.dev->id, 0, trigger);
742 
743 	svm_range_dma_unmap(adev->dev, scratch, 0, npages);
744 
745 out_free:
746 	kvfree(buf);
747 out:
748 	if (!r && cpages) {
749 		pdd = svm_range_get_pdd_by_adev(prange, adev);
750 		if (pdd)
751 			WRITE_ONCE(pdd->page_out, pdd->page_out + cpages);
752 	}
753 	return r ? r : upages;
754 }
755 
756 /**
757  * svm_migrate_vram_to_ram - migrate svm range from device to system
758  * @prange: range structure
759  * @mm: process mm, use current->mm if NULL
760  * @trigger: reason of migration
761  *
762  * Context: Process context, caller hold mmap read lock, prange->migrate_mutex
763  *
764  * Return:
765  * 0 - OK, otherwise error code
766  */
767 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
768 			    uint32_t trigger, struct page *fault_page)
769 {
770 	struct amdgpu_device *adev;
771 	struct vm_area_struct *vma;
772 	unsigned long addr;
773 	unsigned long start;
774 	unsigned long end;
775 	unsigned long upages = 0;
776 	long r = 0;
777 
778 	if (!prange->actual_loc) {
779 		pr_debug("[0x%lx 0x%lx] already migrated to ram\n",
780 			 prange->start, prange->last);
781 		return 0;
782 	}
783 
784 	adev = svm_range_get_adev_by_id(prange, prange->actual_loc);
785 	if (!adev) {
786 		pr_debug("failed to get device by id 0x%x\n",
787 			 prange->actual_loc);
788 		return -ENODEV;
789 	}
790 
791 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n",
792 		 prange->svms, prange, prange->start, prange->last,
793 		 prange->actual_loc);
794 
795 	start = prange->start << PAGE_SHIFT;
796 	end = (prange->last + 1) << PAGE_SHIFT;
797 
798 	for (addr = start; addr < end;) {
799 		unsigned long next;
800 
801 		vma = find_vma(mm, addr);
802 		if (!vma || addr < vma->vm_start) {
803 			pr_debug("failed to find vma for prange %p\n", prange);
804 			r = -EFAULT;
805 			break;
806 		}
807 
808 		next = min(vma->vm_end, end);
809 		r = svm_migrate_vma_to_ram(adev, prange, vma, addr, next, trigger,
810 			fault_page);
811 		if (r < 0) {
812 			pr_debug("failed %ld to migrate prange %p\n", r, prange);
813 			break;
814 		} else {
815 			upages += r;
816 		}
817 		addr = next;
818 	}
819 
820 	if (r >= 0 && !upages) {
821 		svm_range_vram_node_free(prange);
822 		prange->actual_loc = 0;
823 	}
824 
825 	return r < 0 ? r : 0;
826 }
827 
828 /**
829  * svm_migrate_vram_to_vram - migrate svm range from device to device
830  * @prange: range structure
831  * @best_loc: the device to migrate to
832  * @mm: process mm, use current->mm if NULL
833  * @trigger: reason of migration
834  *
835  * Context: Process context, caller hold mmap read lock, svms lock, prange lock
836  *
837  * Return:
838  * 0 - OK, otherwise error code
839  */
840 static int
841 svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,
842 			 struct mm_struct *mm, uint32_t trigger)
843 {
844 	int r, retries = 3;
845 
846 	/*
847 	 * TODO: for both devices with PCIe large bar or on same xgmi hive, skip
848 	 * system memory as migration bridge
849 	 */
850 
851 	pr_debug("from gpu 0x%x to gpu 0x%x\n", prange->actual_loc, best_loc);
852 
853 	do {
854 		r = svm_migrate_vram_to_ram(prange, mm, trigger, NULL);
855 		if (r)
856 			return r;
857 	} while (prange->actual_loc && --retries);
858 
859 	if (prange->actual_loc)
860 		return -EDEADLK;
861 
862 	return svm_migrate_ram_to_vram(prange, best_loc, mm, trigger);
863 }
864 
865 int
866 svm_migrate_to_vram(struct svm_range *prange, uint32_t best_loc,
867 		    struct mm_struct *mm, uint32_t trigger)
868 {
869 	if  (!prange->actual_loc)
870 		return svm_migrate_ram_to_vram(prange, best_loc, mm, trigger);
871 	else
872 		return svm_migrate_vram_to_vram(prange, best_loc, mm, trigger);
873 
874 }
875 
876 /**
877  * svm_migrate_to_ram - CPU page fault handler
878  * @vmf: CPU vm fault vma, address
879  *
880  * Context: vm fault handler, caller holds the mmap read lock
881  *
882  * Return:
883  * 0 - OK
884  * VM_FAULT_SIGBUS - notice application to have SIGBUS page fault
885  */
886 static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf)
887 {
888 	unsigned long addr = vmf->address;
889 	struct svm_range_bo *svm_bo;
890 	enum svm_work_list_ops op;
891 	struct svm_range *parent;
892 	struct svm_range *prange;
893 	struct kfd_process *p;
894 	struct mm_struct *mm;
895 	int r = 0;
896 
897 	svm_bo = vmf->page->zone_device_data;
898 	if (!svm_bo) {
899 		pr_debug("failed get device page at addr 0x%lx\n", addr);
900 		return VM_FAULT_SIGBUS;
901 	}
902 	if (!mmget_not_zero(svm_bo->eviction_fence->mm)) {
903 		pr_debug("addr 0x%lx of process mm is destroyed\n", addr);
904 		return VM_FAULT_SIGBUS;
905 	}
906 
907 	mm = svm_bo->eviction_fence->mm;
908 	if (mm != vmf->vma->vm_mm)
909 		pr_debug("addr 0x%lx is COW mapping in child process\n", addr);
910 
911 	p = kfd_lookup_process_by_mm(mm);
912 	if (!p) {
913 		pr_debug("failed find process at fault address 0x%lx\n", addr);
914 		r = VM_FAULT_SIGBUS;
915 		goto out_mmput;
916 	}
917 	if (READ_ONCE(p->svms.faulting_task) == current) {
918 		pr_debug("skipping ram migration\n");
919 		r = 0;
920 		goto out_unref_process;
921 	}
922 
923 	pr_debug("CPU page fault svms 0x%p address 0x%lx\n", &p->svms, addr);
924 	addr >>= PAGE_SHIFT;
925 
926 	mutex_lock(&p->svms.lock);
927 
928 	prange = svm_range_from_addr(&p->svms, addr, &parent);
929 	if (!prange) {
930 		pr_debug("failed get range svms 0x%p addr 0x%lx\n", &p->svms, addr);
931 		r = -EFAULT;
932 		goto out_unlock_svms;
933 	}
934 
935 	mutex_lock(&parent->migrate_mutex);
936 	if (prange != parent)
937 		mutex_lock_nested(&prange->migrate_mutex, 1);
938 
939 	if (!prange->actual_loc)
940 		goto out_unlock_prange;
941 
942 	svm_range_lock(parent);
943 	if (prange != parent)
944 		mutex_lock_nested(&prange->lock, 1);
945 	r = svm_range_split_by_granularity(p, mm, addr, parent, prange);
946 	if (prange != parent)
947 		mutex_unlock(&prange->lock);
948 	svm_range_unlock(parent);
949 	if (r) {
950 		pr_debug("failed %d to split range by granularity\n", r);
951 		goto out_unlock_prange;
952 	}
953 
954 	r = svm_migrate_vram_to_ram(prange, vmf->vma->vm_mm,
955 				    KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU,
956 				    vmf->page);
957 	if (r)
958 		pr_debug("failed %d migrate svms 0x%p range 0x%p [0x%lx 0x%lx]\n",
959 			 r, prange->svms, prange, prange->start, prange->last);
960 
961 	/* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
962 	if (p->xnack_enabled && parent == prange)
963 		op = SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP;
964 	else
965 		op = SVM_OP_UPDATE_RANGE_NOTIFIER;
966 	svm_range_add_list_work(&p->svms, parent, mm, op);
967 	schedule_deferred_list_work(&p->svms);
968 
969 out_unlock_prange:
970 	if (prange != parent)
971 		mutex_unlock(&prange->migrate_mutex);
972 	mutex_unlock(&parent->migrate_mutex);
973 out_unlock_svms:
974 	mutex_unlock(&p->svms.lock);
975 out_unref_process:
976 	kfd_unref_process(p);
977 out_mmput:
978 	mmput(mm);
979 
980 	pr_debug("CPU fault svms 0x%p address 0x%lx done\n", &p->svms, addr);
981 
982 	return r ? VM_FAULT_SIGBUS : 0;
983 }
984 
985 static const struct dev_pagemap_ops svm_migrate_pgmap_ops = {
986 	.page_free		= svm_migrate_page_free,
987 	.migrate_to_ram		= svm_migrate_to_ram,
988 };
989 
990 /* Each VRAM page uses sizeof(struct page) on system memory */
991 #define SVM_HMM_PAGE_STRUCT_SIZE(size) ((size)/PAGE_SIZE * sizeof(struct page))
992 
993 int svm_migrate_init(struct amdgpu_device *adev)
994 {
995 	struct kfd_dev *kfddev = adev->kfd.dev;
996 	struct dev_pagemap *pgmap;
997 	struct resource *res = NULL;
998 	unsigned long size;
999 	void *r;
1000 
1001 	/* Page migration works on Vega10 or newer */
1002 	if (!KFD_IS_SOC15(kfddev))
1003 		return -EINVAL;
1004 
1005 	pgmap = &kfddev->pgmap;
1006 	memset(pgmap, 0, sizeof(*pgmap));
1007 
1008 	/* TODO: register all vram to HMM for now.
1009 	 * should remove reserved size
1010 	 */
1011 	size = ALIGN(adev->gmc.real_vram_size, 2ULL << 20);
1012 	if (adev->gmc.xgmi.connected_to_cpu) {
1013 		pgmap->range.start = adev->gmc.aper_base;
1014 		pgmap->range.end = adev->gmc.aper_base + adev->gmc.aper_size - 1;
1015 		pgmap->type = MEMORY_DEVICE_COHERENT;
1016 	} else {
1017 		res = devm_request_free_mem_region(adev->dev, &iomem_resource, size);
1018 		if (IS_ERR(res))
1019 			return -ENOMEM;
1020 		pgmap->range.start = res->start;
1021 		pgmap->range.end = res->end;
1022 		pgmap->type = MEMORY_DEVICE_PRIVATE;
1023 	}
1024 
1025 	pgmap->nr_range = 1;
1026 	pgmap->ops = &svm_migrate_pgmap_ops;
1027 	pgmap->owner = SVM_ADEV_PGMAP_OWNER(adev);
1028 	pgmap->flags = 0;
1029 	/* Device manager releases device-specific resources, memory region and
1030 	 * pgmap when driver disconnects from device.
1031 	 */
1032 	r = devm_memremap_pages(adev->dev, pgmap);
1033 	if (IS_ERR(r)) {
1034 		pr_err("failed to register HMM device memory\n");
1035 		/* Disable SVM support capability */
1036 		pgmap->type = 0;
1037 		if (pgmap->type == MEMORY_DEVICE_PRIVATE)
1038 			devm_release_mem_region(adev->dev, res->start,
1039 						res->end - res->start + 1);
1040 		return PTR_ERR(r);
1041 	}
1042 
1043 	pr_debug("reserve %ldMB system memory for VRAM pages struct\n",
1044 		 SVM_HMM_PAGE_STRUCT_SIZE(size) >> 20);
1045 
1046 	amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size));
1047 
1048 	svm_range_set_max_pages(adev);
1049 
1050 	pr_info("HMM registered %ldMB device memory\n", size >> 20);
1051 
1052 	return 0;
1053 }
1054