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.high_pr,
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.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.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 kfd_node *node, struct svm_range *prange,
291 			 struct migrate_vma *migrate, struct dma_fence **mfence,
292 			 dma_addr_t *scratch, uint64_t ttm_res_offset)
293 {
294 	uint64_t npages = migrate->cpages;
295 	struct amdgpu_device *adev = node->adev;
296 	struct device *dev = adev->dev;
297 	struct amdgpu_res_cursor cursor;
298 	dma_addr_t *src;
299 	uint64_t *dst;
300 	uint64_t i, j;
301 	int r;
302 
303 	pr_debug("svms 0x%p [0x%lx 0x%lx 0x%llx]\n", prange->svms, prange->start,
304 		 prange->last, ttm_res_offset);
305 
306 	src = scratch;
307 	dst = (uint64_t *)(scratch + npages);
308 
309 	amdgpu_res_first(prange->ttm_res, ttm_res_offset,
310 			 npages << PAGE_SHIFT, &cursor);
311 	for (i = j = 0; i < npages; i++) {
312 		struct page *spage;
313 
314 		dst[i] = cursor.start + (j << PAGE_SHIFT);
315 		migrate->dst[i] = svm_migrate_addr_to_pfn(adev, dst[i]);
316 		svm_migrate_get_vram_page(prange, migrate->dst[i]);
317 		migrate->dst[i] = migrate_pfn(migrate->dst[i]);
318 
319 		spage = migrate_pfn_to_page(migrate->src[i]);
320 		if (spage && !is_zone_device_page(spage)) {
321 			src[i] = dma_map_page(dev, spage, 0, PAGE_SIZE,
322 					      DMA_TO_DEVICE);
323 			r = dma_mapping_error(dev, src[i]);
324 			if (r) {
325 				dev_err(dev, "%s: fail %d dma_map_page\n",
326 					__func__, r);
327 				goto out_free_vram_pages;
328 			}
329 		} else {
330 			if (j) {
331 				r = svm_migrate_copy_memory_gart(
332 						adev, src + i - j,
333 						dst + i - j, j,
334 						FROM_RAM_TO_VRAM,
335 						mfence);
336 				if (r)
337 					goto out_free_vram_pages;
338 				amdgpu_res_next(&cursor, (j + 1) << PAGE_SHIFT);
339 				j = 0;
340 			} else {
341 				amdgpu_res_next(&cursor, PAGE_SIZE);
342 			}
343 			continue;
344 		}
345 
346 		pr_debug_ratelimited("dma mapping src to 0x%llx, pfn 0x%lx\n",
347 				     src[i] >> PAGE_SHIFT, page_to_pfn(spage));
348 
349 		if (j >= (cursor.size >> PAGE_SHIFT) - 1 && i < npages - 1) {
350 			r = svm_migrate_copy_memory_gart(adev, src + i - j,
351 							 dst + i - j, j + 1,
352 							 FROM_RAM_TO_VRAM,
353 							 mfence);
354 			if (r)
355 				goto out_free_vram_pages;
356 			amdgpu_res_next(&cursor, (j + 1) * PAGE_SIZE);
357 			j = 0;
358 		} else {
359 			j++;
360 		}
361 	}
362 
363 	r = svm_migrate_copy_memory_gart(adev, src + i - j, dst + i - j, j,
364 					 FROM_RAM_TO_VRAM, mfence);
365 
366 out_free_vram_pages:
367 	if (r) {
368 		pr_debug("failed %d to copy memory to vram\n", r);
369 		while (i--) {
370 			svm_migrate_put_vram_page(adev, dst[i]);
371 			migrate->dst[i] = 0;
372 		}
373 	}
374 
375 #ifdef DEBUG_FORCE_MIXED_DOMAINS
376 	for (i = 0, j = 0; i < npages; i += 4, j++) {
377 		if (j & 1)
378 			continue;
379 		svm_migrate_put_vram_page(adev, dst[i]);
380 		migrate->dst[i] = 0;
381 		svm_migrate_put_vram_page(adev, dst[i + 1]);
382 		migrate->dst[i + 1] = 0;
383 		svm_migrate_put_vram_page(adev, dst[i + 2]);
384 		migrate->dst[i + 2] = 0;
385 		svm_migrate_put_vram_page(adev, dst[i + 3]);
386 		migrate->dst[i + 3] = 0;
387 	}
388 #endif
389 
390 	return r;
391 }
392 
393 static long
394 svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
395 			struct vm_area_struct *vma, uint64_t start,
396 			uint64_t end, uint32_t trigger, uint64_t ttm_res_offset)
397 {
398 	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
399 	uint64_t npages = (end - start) >> PAGE_SHIFT;
400 	struct amdgpu_device *adev = node->adev;
401 	struct kfd_process_device *pdd;
402 	struct dma_fence *mfence = NULL;
403 	struct migrate_vma migrate = { 0 };
404 	unsigned long cpages = 0;
405 	dma_addr_t *scratch;
406 	void *buf;
407 	int r = -ENOMEM;
408 
409 	memset(&migrate, 0, sizeof(migrate));
410 	migrate.vma = vma;
411 	migrate.start = start;
412 	migrate.end = end;
413 	migrate.flags = MIGRATE_VMA_SELECT_SYSTEM;
414 	migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev);
415 
416 	buf = kvcalloc(npages,
417 		       2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t),
418 		       GFP_KERNEL);
419 	if (!buf)
420 		goto out;
421 
422 	migrate.src = buf;
423 	migrate.dst = migrate.src + npages;
424 	scratch = (dma_addr_t *)(migrate.dst + npages);
425 
426 	kfd_smi_event_migration_start(node, p->lead_thread->pid,
427 				      start >> PAGE_SHIFT, end >> PAGE_SHIFT,
428 				      0, node->id, prange->prefetch_loc,
429 				      prange->preferred_loc, trigger);
430 
431 	r = migrate_vma_setup(&migrate);
432 	if (r) {
433 		dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n",
434 			__func__, r, prange->start, prange->last);
435 		goto out_free;
436 	}
437 
438 	cpages = migrate.cpages;
439 	if (!cpages) {
440 		pr_debug("failed collect migrate sys pages [0x%lx 0x%lx]\n",
441 			 prange->start, prange->last);
442 		goto out_free;
443 	}
444 	if (cpages != npages)
445 		pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n",
446 			 cpages, npages);
447 	else
448 		pr_debug("0x%lx pages migrated\n", cpages);
449 
450 	r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset);
451 	migrate_vma_pages(&migrate);
452 
453 	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
454 		svm_migrate_successful_pages(&migrate), cpages, migrate.npages);
455 
456 	svm_migrate_copy_done(adev, mfence);
457 	migrate_vma_finalize(&migrate);
458 
459 	kfd_smi_event_migration_end(node, p->lead_thread->pid,
460 				    start >> PAGE_SHIFT, end >> PAGE_SHIFT,
461 				    0, node->id, trigger);
462 
463 	svm_range_dma_unmap(adev->dev, scratch, 0, npages);
464 
465 out_free:
466 	kvfree(buf);
467 out:
468 	if (!r && cpages) {
469 		pdd = svm_range_get_pdd_by_node(prange, node);
470 		if (pdd)
471 			WRITE_ONCE(pdd->page_in, pdd->page_in + cpages);
472 
473 		return cpages;
474 	}
475 	return r;
476 }
477 
478 /**
479  * svm_migrate_ram_to_vram - migrate svm range from system to device
480  * @prange: range structure
481  * @best_loc: the device to migrate to
482  * @mm: the process mm structure
483  * @trigger: reason of migration
484  *
485  * Context: Process context, caller hold mmap read lock, svms lock, prange lock
486  *
487  * Return:
488  * 0 - OK, otherwise error code
489  */
490 static int
491 svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,
492 			struct mm_struct *mm, uint32_t trigger)
493 {
494 	unsigned long addr, start, end;
495 	struct vm_area_struct *vma;
496 	uint64_t ttm_res_offset;
497 	struct kfd_node *node;
498 	unsigned long cpages = 0;
499 	long r = 0;
500 
501 	if (prange->actual_loc == best_loc) {
502 		pr_debug("svms 0x%p [0x%lx 0x%lx] already on best_loc 0x%x\n",
503 			 prange->svms, prange->start, prange->last, best_loc);
504 		return 0;
505 	}
506 
507 	node = svm_range_get_node_by_id(prange, best_loc);
508 	if (!node) {
509 		pr_debug("failed to get kfd node by id 0x%x\n", best_loc);
510 		return -ENODEV;
511 	}
512 
513 	pr_debug("svms 0x%p [0x%lx 0x%lx] to gpu 0x%x\n", prange->svms,
514 		 prange->start, prange->last, best_loc);
515 
516 	start = prange->start << PAGE_SHIFT;
517 	end = (prange->last + 1) << PAGE_SHIFT;
518 
519 	r = svm_range_vram_node_new(node, prange, true);
520 	if (r) {
521 		dev_dbg(node->adev->dev, "fail %ld to alloc vram\n", r);
522 		return r;
523 	}
524 	ttm_res_offset = prange->offset << PAGE_SHIFT;
525 
526 	for (addr = start; addr < end;) {
527 		unsigned long next;
528 
529 		vma = vma_lookup(mm, addr);
530 		if (!vma)
531 			break;
532 
533 		next = min(vma->vm_end, end);
534 		r = svm_migrate_vma_to_vram(node, prange, vma, addr, next, trigger, ttm_res_offset);
535 		if (r < 0) {
536 			pr_debug("failed %ld to migrate\n", r);
537 			break;
538 		} else {
539 			cpages += r;
540 		}
541 		ttm_res_offset += next - addr;
542 		addr = next;
543 	}
544 
545 	if (cpages) {
546 		prange->actual_loc = best_loc;
547 		svm_range_free_dma_mappings(prange, true);
548 	} else {
549 		svm_range_vram_node_free(prange);
550 	}
551 
552 	return r < 0 ? r : 0;
553 }
554 
555 static void svm_migrate_page_free(struct page *page)
556 {
557 	struct svm_range_bo *svm_bo = page->zone_device_data;
558 
559 	if (svm_bo) {
560 		pr_debug_ratelimited("ref: %d\n", kref_read(&svm_bo->kref));
561 		svm_range_bo_unref_async(svm_bo);
562 	}
563 }
564 
565 static int
566 svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
567 			struct migrate_vma *migrate, struct dma_fence **mfence,
568 			dma_addr_t *scratch, uint64_t npages)
569 {
570 	struct device *dev = adev->dev;
571 	uint64_t *src;
572 	dma_addr_t *dst;
573 	struct page *dpage;
574 	uint64_t i = 0, j;
575 	uint64_t addr;
576 	int r = 0;
577 
578 	pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start,
579 		 prange->last);
580 
581 	addr = prange->start << PAGE_SHIFT;
582 
583 	src = (uint64_t *)(scratch + npages);
584 	dst = scratch;
585 
586 	for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) {
587 		struct page *spage;
588 
589 		spage = migrate_pfn_to_page(migrate->src[i]);
590 		if (!spage || !is_zone_device_page(spage)) {
591 			pr_debug("invalid page. Could be in CPU already svms 0x%p [0x%lx 0x%lx]\n",
592 				 prange->svms, prange->start, prange->last);
593 			if (j) {
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 			continue;
603 		}
604 		src[i] = svm_migrate_addr(adev, spage);
605 		if (j > 0 && src[i] != src[i - 1] + PAGE_SIZE) {
606 			r = svm_migrate_copy_memory_gart(adev, dst + i - j,
607 							 src + i - j, j,
608 							 FROM_VRAM_TO_RAM,
609 							 mfence);
610 			if (r)
611 				goto out_oom;
612 			j = 0;
613 		}
614 
615 		dpage = svm_migrate_get_sys_page(migrate->vma, addr);
616 		if (!dpage) {
617 			pr_debug("failed get page svms 0x%p [0x%lx 0x%lx]\n",
618 				 prange->svms, prange->start, prange->last);
619 			r = -ENOMEM;
620 			goto out_oom;
621 		}
622 
623 		dst[i] = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_FROM_DEVICE);
624 		r = dma_mapping_error(dev, dst[i]);
625 		if (r) {
626 			dev_err(adev->dev, "%s: fail %d dma_map_page\n", __func__, r);
627 			goto out_oom;
628 		}
629 
630 		pr_debug_ratelimited("dma mapping dst to 0x%llx, pfn 0x%lx\n",
631 				     dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
632 
633 		migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
634 		j++;
635 	}
636 
637 	r = svm_migrate_copy_memory_gart(adev, dst + i - j, src + i - j, j,
638 					 FROM_VRAM_TO_RAM, mfence);
639 
640 out_oom:
641 	if (r) {
642 		pr_debug("failed %d copy to ram\n", r);
643 		while (i--) {
644 			svm_migrate_put_sys_page(dst[i]);
645 			migrate->dst[i] = 0;
646 		}
647 	}
648 
649 	return r;
650 }
651 
652 /**
653  * svm_migrate_vma_to_ram - migrate range inside one vma from device to system
654  *
655  * @prange: svm range structure
656  * @vma: vm_area_struct that range [start, end] belongs to
657  * @start: range start virtual address in pages
658  * @end: range end virtual address in pages
659  * @node: kfd node device to migrate from
660  * @trigger: reason of migration
661  * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback
662  *
663  * Context: Process context, caller hold mmap read lock, prange->migrate_mutex
664  *
665  * Return:
666  *   0 - success with all pages migrated
667  *   negative values - indicate error
668  *   positive values - partial migration, number of pages not migrated
669  */
670 static long
671 svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
672 		       struct vm_area_struct *vma, uint64_t start, uint64_t end,
673 		       uint32_t trigger, struct page *fault_page)
674 {
675 	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
676 	uint64_t npages = (end - start) >> PAGE_SHIFT;
677 	unsigned long upages = npages;
678 	unsigned long cpages = 0;
679 	struct amdgpu_device *adev = node->adev;
680 	struct kfd_process_device *pdd;
681 	struct dma_fence *mfence = NULL;
682 	struct migrate_vma migrate = { 0 };
683 	dma_addr_t *scratch;
684 	void *buf;
685 	int r = -ENOMEM;
686 
687 	memset(&migrate, 0, sizeof(migrate));
688 	migrate.vma = vma;
689 	migrate.start = start;
690 	migrate.end = end;
691 	migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev);
692 	if (adev->gmc.xgmi.connected_to_cpu)
693 		migrate.flags = MIGRATE_VMA_SELECT_DEVICE_COHERENT;
694 	else
695 		migrate.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE;
696 
697 	buf = kvcalloc(npages,
698 		       2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t),
699 		       GFP_KERNEL);
700 	if (!buf)
701 		goto out;
702 
703 	migrate.src = buf;
704 	migrate.dst = migrate.src + npages;
705 	migrate.fault_page = fault_page;
706 	scratch = (dma_addr_t *)(migrate.dst + npages);
707 
708 	kfd_smi_event_migration_start(node, p->lead_thread->pid,
709 				      start >> PAGE_SHIFT, end >> PAGE_SHIFT,
710 				      node->id, 0, prange->prefetch_loc,
711 				      prange->preferred_loc, trigger);
712 
713 	r = migrate_vma_setup(&migrate);
714 	if (r) {
715 		dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n",
716 			__func__, r, prange->start, prange->last);
717 		goto out_free;
718 	}
719 
720 	cpages = migrate.cpages;
721 	if (!cpages) {
722 		pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n",
723 			 prange->start, prange->last);
724 		upages = svm_migrate_unsuccessful_pages(&migrate);
725 		goto out_free;
726 	}
727 	if (cpages != npages)
728 		pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n",
729 			 cpages, npages);
730 	else
731 		pr_debug("0x%lx pages migrated\n", cpages);
732 
733 	r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
734 				    scratch, npages);
735 	migrate_vma_pages(&migrate);
736 
737 	upages = svm_migrate_unsuccessful_pages(&migrate);
738 	pr_debug("unsuccessful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
739 		 upages, cpages, migrate.npages);
740 
741 	svm_migrate_copy_done(adev, mfence);
742 	migrate_vma_finalize(&migrate);
743 
744 	kfd_smi_event_migration_end(node, p->lead_thread->pid,
745 				    start >> PAGE_SHIFT, end >> PAGE_SHIFT,
746 				    node->id, 0, trigger);
747 
748 	svm_range_dma_unmap(adev->dev, scratch, 0, npages);
749 
750 out_free:
751 	kvfree(buf);
752 out:
753 	if (!r && cpages) {
754 		pdd = svm_range_get_pdd_by_node(prange, node);
755 		if (pdd)
756 			WRITE_ONCE(pdd->page_out, pdd->page_out + cpages);
757 	}
758 	return r ? r : upages;
759 }
760 
761 /**
762  * svm_migrate_vram_to_ram - migrate svm range from device to system
763  * @prange: range structure
764  * @mm: process mm, use current->mm if NULL
765  * @trigger: reason of migration
766  * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback
767  *
768  * Context: Process context, caller hold mmap read lock, prange->migrate_mutex
769  *
770  * Return:
771  * 0 - OK, otherwise error code
772  */
773 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
774 			    uint32_t trigger, struct page *fault_page)
775 {
776 	struct kfd_node *node;
777 	struct vm_area_struct *vma;
778 	unsigned long addr;
779 	unsigned long start;
780 	unsigned long end;
781 	unsigned long upages = 0;
782 	long r = 0;
783 
784 	if (!prange->actual_loc) {
785 		pr_debug("[0x%lx 0x%lx] already migrated to ram\n",
786 			 prange->start, prange->last);
787 		return 0;
788 	}
789 
790 	node = svm_range_get_node_by_id(prange, prange->actual_loc);
791 	if (!node) {
792 		pr_debug("failed to get kfd node by id 0x%x\n", prange->actual_loc);
793 		return -ENODEV;
794 	}
795 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n",
796 		 prange->svms, prange, prange->start, prange->last,
797 		 prange->actual_loc);
798 
799 	start = prange->start << PAGE_SHIFT;
800 	end = (prange->last + 1) << PAGE_SHIFT;
801 
802 	for (addr = start; addr < end;) {
803 		unsigned long next;
804 
805 		vma = vma_lookup(mm, addr);
806 		if (!vma) {
807 			pr_debug("failed to find vma for prange %p\n", prange);
808 			r = -EFAULT;
809 			break;
810 		}
811 
812 		next = min(vma->vm_end, end);
813 		r = svm_migrate_vma_to_ram(node, prange, vma, addr, next, trigger,
814 			fault_page);
815 		if (r < 0) {
816 			pr_debug("failed %ld to migrate prange %p\n", r, prange);
817 			break;
818 		} else {
819 			upages += r;
820 		}
821 		addr = next;
822 	}
823 
824 	if (r >= 0 && !upages) {
825 		svm_range_vram_node_free(prange);
826 		prange->actual_loc = 0;
827 	}
828 
829 	return r < 0 ? r : 0;
830 }
831 
832 /**
833  * svm_migrate_vram_to_vram - migrate svm range from device to device
834  * @prange: range structure
835  * @best_loc: the device to migrate to
836  * @mm: process mm, use current->mm if NULL
837  * @trigger: reason of migration
838  *
839  * Context: Process context, caller hold mmap read lock, svms lock, prange lock
840  *
841  * Return:
842  * 0 - OK, otherwise error code
843  */
844 static int
845 svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,
846 			 struct mm_struct *mm, uint32_t trigger)
847 {
848 	int r, retries = 3;
849 
850 	/*
851 	 * TODO: for both devices with PCIe large bar or on same xgmi hive, skip
852 	 * system memory as migration bridge
853 	 */
854 
855 	pr_debug("from gpu 0x%x to gpu 0x%x\n", prange->actual_loc, best_loc);
856 
857 	do {
858 		r = svm_migrate_vram_to_ram(prange, mm, trigger, NULL);
859 		if (r)
860 			return r;
861 	} while (prange->actual_loc && --retries);
862 
863 	if (prange->actual_loc)
864 		return -EDEADLK;
865 
866 	return svm_migrate_ram_to_vram(prange, best_loc, mm, trigger);
867 }
868 
869 int
870 svm_migrate_to_vram(struct svm_range *prange, uint32_t best_loc,
871 		    struct mm_struct *mm, uint32_t trigger)
872 {
873 	if  (!prange->actual_loc)
874 		return svm_migrate_ram_to_vram(prange, best_loc, mm, trigger);
875 	else
876 		return svm_migrate_vram_to_vram(prange, best_loc, mm, trigger);
877 
878 }
879 
880 /**
881  * svm_migrate_to_ram - CPU page fault handler
882  * @vmf: CPU vm fault vma, address
883  *
884  * Context: vm fault handler, caller holds the mmap read lock
885  *
886  * Return:
887  * 0 - OK
888  * VM_FAULT_SIGBUS - notice application to have SIGBUS page fault
889  */
890 static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf)
891 {
892 	unsigned long addr = vmf->address;
893 	struct svm_range_bo *svm_bo;
894 	enum svm_work_list_ops op;
895 	struct svm_range *parent;
896 	struct svm_range *prange;
897 	struct kfd_process *p;
898 	struct mm_struct *mm;
899 	int r = 0;
900 
901 	svm_bo = vmf->page->zone_device_data;
902 	if (!svm_bo) {
903 		pr_debug("failed get device page at addr 0x%lx\n", addr);
904 		return VM_FAULT_SIGBUS;
905 	}
906 	if (!mmget_not_zero(svm_bo->eviction_fence->mm)) {
907 		pr_debug("addr 0x%lx of process mm is destroyed\n", addr);
908 		return VM_FAULT_SIGBUS;
909 	}
910 
911 	mm = svm_bo->eviction_fence->mm;
912 	if (mm != vmf->vma->vm_mm)
913 		pr_debug("addr 0x%lx is COW mapping in child process\n", addr);
914 
915 	p = kfd_lookup_process_by_mm(mm);
916 	if (!p) {
917 		pr_debug("failed find process at fault address 0x%lx\n", addr);
918 		r = VM_FAULT_SIGBUS;
919 		goto out_mmput;
920 	}
921 	if (READ_ONCE(p->svms.faulting_task) == current) {
922 		pr_debug("skipping ram migration\n");
923 		r = 0;
924 		goto out_unref_process;
925 	}
926 
927 	pr_debug("CPU page fault svms 0x%p address 0x%lx\n", &p->svms, addr);
928 	addr >>= PAGE_SHIFT;
929 
930 	mutex_lock(&p->svms.lock);
931 
932 	prange = svm_range_from_addr(&p->svms, addr, &parent);
933 	if (!prange) {
934 		pr_debug("failed get range svms 0x%p addr 0x%lx\n", &p->svms, addr);
935 		r = -EFAULT;
936 		goto out_unlock_svms;
937 	}
938 
939 	mutex_lock(&parent->migrate_mutex);
940 	if (prange != parent)
941 		mutex_lock_nested(&prange->migrate_mutex, 1);
942 
943 	if (!prange->actual_loc)
944 		goto out_unlock_prange;
945 
946 	svm_range_lock(parent);
947 	if (prange != parent)
948 		mutex_lock_nested(&prange->lock, 1);
949 	r = svm_range_split_by_granularity(p, mm, addr, parent, prange);
950 	if (prange != parent)
951 		mutex_unlock(&prange->lock);
952 	svm_range_unlock(parent);
953 	if (r) {
954 		pr_debug("failed %d to split range by granularity\n", r);
955 		goto out_unlock_prange;
956 	}
957 
958 	r = svm_migrate_vram_to_ram(prange, vmf->vma->vm_mm,
959 				    KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU,
960 				    vmf->page);
961 	if (r)
962 		pr_debug("failed %d migrate svms 0x%p range 0x%p [0x%lx 0x%lx]\n",
963 			 r, prange->svms, prange, prange->start, prange->last);
964 
965 	/* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
966 	if (p->xnack_enabled && parent == prange)
967 		op = SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP;
968 	else
969 		op = SVM_OP_UPDATE_RANGE_NOTIFIER;
970 	svm_range_add_list_work(&p->svms, parent, mm, op);
971 	schedule_deferred_list_work(&p->svms);
972 
973 out_unlock_prange:
974 	if (prange != parent)
975 		mutex_unlock(&prange->migrate_mutex);
976 	mutex_unlock(&parent->migrate_mutex);
977 out_unlock_svms:
978 	mutex_unlock(&p->svms.lock);
979 out_unref_process:
980 	pr_debug("CPU fault svms 0x%p address 0x%lx done\n", &p->svms, addr);
981 	kfd_unref_process(p);
982 out_mmput:
983 	mmput(mm);
984 	return r ? VM_FAULT_SIGBUS : 0;
985 }
986 
987 static const struct dev_pagemap_ops svm_migrate_pgmap_ops = {
988 	.page_free		= svm_migrate_page_free,
989 	.migrate_to_ram		= svm_migrate_to_ram,
990 };
991 
992 /* Each VRAM page uses sizeof(struct page) on system memory */
993 #define SVM_HMM_PAGE_STRUCT_SIZE(size) ((size)/PAGE_SIZE * sizeof(struct page))
994 
995 int kgd2kfd_init_zone_device(struct amdgpu_device *adev)
996 {
997 	struct amdgpu_kfd_dev *kfddev = &adev->kfd;
998 	struct dev_pagemap *pgmap;
999 	struct resource *res = NULL;
1000 	unsigned long size;
1001 	void *r;
1002 
1003 	/* Page migration works on gfx9 or newer */
1004 	if (adev->ip_versions[GC_HWIP][0] < IP_VERSION(9, 0, 1))
1005 		return -EINVAL;
1006 
1007 	if (adev->gmc.is_app_apu)
1008 		return 0;
1009 
1010 	pgmap = &kfddev->pgmap;
1011 	memset(pgmap, 0, sizeof(*pgmap));
1012 
1013 	/* TODO: register all vram to HMM for now.
1014 	 * should remove reserved size
1015 	 */
1016 	size = ALIGN(adev->gmc.real_vram_size, 2ULL << 20);
1017 	if (adev->gmc.xgmi.connected_to_cpu) {
1018 		pgmap->range.start = adev->gmc.aper_base;
1019 		pgmap->range.end = adev->gmc.aper_base + adev->gmc.aper_size - 1;
1020 		pgmap->type = MEMORY_DEVICE_COHERENT;
1021 	} else {
1022 		res = devm_request_free_mem_region(adev->dev, &iomem_resource, size);
1023 		if (IS_ERR(res))
1024 			return -ENOMEM;
1025 		pgmap->range.start = res->start;
1026 		pgmap->range.end = res->end;
1027 		pgmap->type = MEMORY_DEVICE_PRIVATE;
1028 	}
1029 
1030 	pgmap->nr_range = 1;
1031 	pgmap->ops = &svm_migrate_pgmap_ops;
1032 	pgmap->owner = SVM_ADEV_PGMAP_OWNER(adev);
1033 	pgmap->flags = 0;
1034 	/* Device manager releases device-specific resources, memory region and
1035 	 * pgmap when driver disconnects from device.
1036 	 */
1037 	r = devm_memremap_pages(adev->dev, pgmap);
1038 	if (IS_ERR(r)) {
1039 		pr_err("failed to register HMM device memory\n");
1040 		/* Disable SVM support capability */
1041 		pgmap->type = 0;
1042 		if (pgmap->type == MEMORY_DEVICE_PRIVATE)
1043 			devm_release_mem_region(adev->dev, res->start, resource_size(res));
1044 		return PTR_ERR(r);
1045 	}
1046 
1047 	pr_debug("reserve %ldMB system memory for VRAM pages struct\n",
1048 		 SVM_HMM_PAGE_STRUCT_SIZE(size) >> 20);
1049 
1050 	amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size));
1051 
1052 	pr_info("HMM registered %ldMB device memory\n", size >> 20);
1053 
1054 	return 0;
1055 }
1056