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