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
svm_migrate_direct_mapping_addr(struct amdgpu_device * adev,uint64_t addr)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
svm_migrate_gart_map(struct amdgpu_ring * ring,uint64_t npages,dma_addr_t * addr,uint64_t * gart_addr,uint64_t flags)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
svm_migrate_copy_memory_gart(struct amdgpu_device * adev,dma_addr_t * sys,uint64_t * vram,uint64_t npages,enum MIGRATION_COPY_DIR direction,struct dma_fence ** mfence)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
svm_migrate_copy_done(struct amdgpu_device * adev,struct dma_fence * mfence)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
svm_migrate_addr_to_pfn(struct amdgpu_device * adev,unsigned long addr)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
svm_migrate_get_vram_page(struct svm_range * prange,unsigned long pfn)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
svm_migrate_put_vram_page(struct amdgpu_device * adev,unsigned long addr)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
svm_migrate_addr(struct amdgpu_device * adev,struct page * page)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 *
svm_migrate_get_sys_page(struct vm_area_struct * vma,unsigned long addr)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
svm_migrate_put_sys_page(unsigned long addr)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
svm_migrate_successful_pages(struct migrate_vma * migrate)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
svm_migrate_unsuccessful_pages(struct migrate_vma * migrate)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
svm_migrate_copy_to_vram(struct kfd_node * node,struct svm_range * prange,struct migrate_vma * migrate,struct dma_fence ** mfence,dma_addr_t * scratch,uint64_t ttm_res_offset)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
svm_migrate_vma_to_vram(struct kfd_node * node,struct svm_range * prange,struct vm_area_struct * vma,uint64_t start,uint64_t end,uint32_t trigger,uint64_t ttm_res_offset)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
svm_migrate_ram_to_vram(struct svm_range * prange,uint32_t best_loc,struct mm_struct * mm,uint32_t trigger)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 = amdgpu_amdkfd_reserve_mem_limit(node->adev,
520 prange->npages * PAGE_SIZE,
521 KFD_IOC_ALLOC_MEM_FLAGS_VRAM,
522 node->xcp ? node->xcp->id : 0);
523 if (r) {
524 dev_dbg(node->adev->dev, "failed to reserve VRAM, r: %ld\n", r);
525 return -ENOSPC;
526 }
527
528 r = svm_range_vram_node_new(node, prange, true);
529 if (r) {
530 dev_dbg(node->adev->dev, "fail %ld to alloc vram\n", r);
531 goto out;
532 }
533 ttm_res_offset = prange->offset << PAGE_SHIFT;
534
535 for (addr = start; addr < end;) {
536 unsigned long next;
537
538 vma = vma_lookup(mm, addr);
539 if (!vma)
540 break;
541
542 next = min(vma->vm_end, end);
543 r = svm_migrate_vma_to_vram(node, prange, vma, addr, next, trigger, ttm_res_offset);
544 if (r < 0) {
545 pr_debug("failed %ld to migrate\n", r);
546 break;
547 } else {
548 cpages += r;
549 }
550 ttm_res_offset += next - addr;
551 addr = next;
552 }
553
554 if (cpages) {
555 prange->actual_loc = best_loc;
556 svm_range_free_dma_mappings(prange, true);
557 } else {
558 svm_range_vram_node_free(prange);
559 }
560
561 out:
562 amdgpu_amdkfd_unreserve_mem_limit(node->adev,
563 prange->npages * PAGE_SIZE,
564 KFD_IOC_ALLOC_MEM_FLAGS_VRAM,
565 node->xcp ? node->xcp->id : 0);
566 return r < 0 ? r : 0;
567 }
568
svm_migrate_page_free(struct page * page)569 static void svm_migrate_page_free(struct page *page)
570 {
571 struct svm_range_bo *svm_bo = page->zone_device_data;
572
573 if (svm_bo) {
574 pr_debug_ratelimited("ref: %d\n", kref_read(&svm_bo->kref));
575 svm_range_bo_unref_async(svm_bo);
576 }
577 }
578
579 static int
svm_migrate_copy_to_ram(struct amdgpu_device * adev,struct svm_range * prange,struct migrate_vma * migrate,struct dma_fence ** mfence,dma_addr_t * scratch,uint64_t npages)580 svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
581 struct migrate_vma *migrate, struct dma_fence **mfence,
582 dma_addr_t *scratch, uint64_t npages)
583 {
584 struct device *dev = adev->dev;
585 uint64_t *src;
586 dma_addr_t *dst;
587 struct page *dpage;
588 uint64_t i = 0, j;
589 uint64_t addr;
590 int r = 0;
591
592 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start,
593 prange->last);
594
595 addr = prange->start << PAGE_SHIFT;
596
597 src = (uint64_t *)(scratch + npages);
598 dst = scratch;
599
600 for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) {
601 struct page *spage;
602
603 spage = migrate_pfn_to_page(migrate->src[i]);
604 if (!spage || !is_zone_device_page(spage)) {
605 pr_debug("invalid page. Could be in CPU already svms 0x%p [0x%lx 0x%lx]\n",
606 prange->svms, prange->start, prange->last);
607 if (j) {
608 r = svm_migrate_copy_memory_gart(adev, dst + i - j,
609 src + i - j, j,
610 FROM_VRAM_TO_RAM,
611 mfence);
612 if (r)
613 goto out_oom;
614 j = 0;
615 }
616 continue;
617 }
618 src[i] = svm_migrate_addr(adev, spage);
619 if (j > 0 && src[i] != src[i - 1] + PAGE_SIZE) {
620 r = svm_migrate_copy_memory_gart(adev, dst + i - j,
621 src + i - j, j,
622 FROM_VRAM_TO_RAM,
623 mfence);
624 if (r)
625 goto out_oom;
626 j = 0;
627 }
628
629 dpage = svm_migrate_get_sys_page(migrate->vma, addr);
630 if (!dpage) {
631 pr_debug("failed get page svms 0x%p [0x%lx 0x%lx]\n",
632 prange->svms, prange->start, prange->last);
633 r = -ENOMEM;
634 goto out_oom;
635 }
636
637 dst[i] = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_FROM_DEVICE);
638 r = dma_mapping_error(dev, dst[i]);
639 if (r) {
640 dev_err(adev->dev, "%s: fail %d dma_map_page\n", __func__, r);
641 goto out_oom;
642 }
643
644 pr_debug_ratelimited("dma mapping dst to 0x%llx, pfn 0x%lx\n",
645 dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
646
647 migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
648 j++;
649 }
650
651 r = svm_migrate_copy_memory_gart(adev, dst + i - j, src + i - j, j,
652 FROM_VRAM_TO_RAM, mfence);
653
654 out_oom:
655 if (r) {
656 pr_debug("failed %d copy to ram\n", r);
657 while (i--) {
658 svm_migrate_put_sys_page(dst[i]);
659 migrate->dst[i] = 0;
660 }
661 }
662
663 return r;
664 }
665
666 /**
667 * svm_migrate_vma_to_ram - migrate range inside one vma from device to system
668 *
669 * @prange: svm range structure
670 * @vma: vm_area_struct that range [start, end] belongs to
671 * @start: range start virtual address in pages
672 * @end: range end virtual address in pages
673 * @node: kfd node device to migrate from
674 * @trigger: reason of migration
675 * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback
676 *
677 * Context: Process context, caller hold mmap read lock, prange->migrate_mutex
678 *
679 * Return:
680 * 0 - success with all pages migrated
681 * negative values - indicate error
682 * positive values - partial migration, number of pages not migrated
683 */
684 static long
svm_migrate_vma_to_ram(struct kfd_node * node,struct svm_range * prange,struct vm_area_struct * vma,uint64_t start,uint64_t end,uint32_t trigger,struct page * fault_page)685 svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
686 struct vm_area_struct *vma, uint64_t start, uint64_t end,
687 uint32_t trigger, struct page *fault_page)
688 {
689 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
690 uint64_t npages = (end - start) >> PAGE_SHIFT;
691 unsigned long upages = npages;
692 unsigned long cpages = 0;
693 struct amdgpu_device *adev = node->adev;
694 struct kfd_process_device *pdd;
695 struct dma_fence *mfence = NULL;
696 struct migrate_vma migrate = { 0 };
697 dma_addr_t *scratch;
698 void *buf;
699 int r = -ENOMEM;
700
701 memset(&migrate, 0, sizeof(migrate));
702 migrate.vma = vma;
703 migrate.start = start;
704 migrate.end = end;
705 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev);
706 if (adev->gmc.xgmi.connected_to_cpu)
707 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_COHERENT;
708 else
709 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE;
710
711 buf = kvcalloc(npages,
712 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t),
713 GFP_KERNEL);
714 if (!buf)
715 goto out;
716
717 migrate.src = buf;
718 migrate.dst = migrate.src + npages;
719 migrate.fault_page = fault_page;
720 scratch = (dma_addr_t *)(migrate.dst + npages);
721
722 kfd_smi_event_migration_start(node, p->lead_thread->pid,
723 start >> PAGE_SHIFT, end >> PAGE_SHIFT,
724 node->id, 0, prange->prefetch_loc,
725 prange->preferred_loc, trigger);
726
727 r = migrate_vma_setup(&migrate);
728 if (r) {
729 dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n",
730 __func__, r, prange->start, prange->last);
731 goto out_free;
732 }
733
734 cpages = migrate.cpages;
735 if (!cpages) {
736 pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n",
737 prange->start, prange->last);
738 upages = svm_migrate_unsuccessful_pages(&migrate);
739 goto out_free;
740 }
741 if (cpages != npages)
742 pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n",
743 cpages, npages);
744 else
745 pr_debug("0x%lx pages migrated\n", cpages);
746
747 r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
748 scratch, npages);
749 migrate_vma_pages(&migrate);
750
751 upages = svm_migrate_unsuccessful_pages(&migrate);
752 pr_debug("unsuccessful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
753 upages, cpages, migrate.npages);
754
755 svm_migrate_copy_done(adev, mfence);
756 migrate_vma_finalize(&migrate);
757
758 kfd_smi_event_migration_end(node, p->lead_thread->pid,
759 start >> PAGE_SHIFT, end >> PAGE_SHIFT,
760 node->id, 0, trigger);
761
762 svm_range_dma_unmap(adev->dev, scratch, 0, npages);
763
764 out_free:
765 kvfree(buf);
766 out:
767 if (!r && cpages) {
768 pdd = svm_range_get_pdd_by_node(prange, node);
769 if (pdd)
770 WRITE_ONCE(pdd->page_out, pdd->page_out + cpages);
771 }
772 return r ? r : upages;
773 }
774
775 /**
776 * svm_migrate_vram_to_ram - migrate svm range from device to system
777 * @prange: range structure
778 * @mm: process mm, use current->mm if NULL
779 * @trigger: reason of migration
780 * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback
781 *
782 * Context: Process context, caller hold mmap read lock, prange->migrate_mutex
783 *
784 * Return:
785 * 0 - OK, otherwise error code
786 */
svm_migrate_vram_to_ram(struct svm_range * prange,struct mm_struct * mm,uint32_t trigger,struct page * fault_page)787 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
788 uint32_t trigger, struct page *fault_page)
789 {
790 struct kfd_node *node;
791 struct vm_area_struct *vma;
792 unsigned long addr;
793 unsigned long start;
794 unsigned long end;
795 unsigned long upages = 0;
796 long r = 0;
797
798 if (!prange->actual_loc) {
799 pr_debug("[0x%lx 0x%lx] already migrated to ram\n",
800 prange->start, prange->last);
801 return 0;
802 }
803
804 node = svm_range_get_node_by_id(prange, prange->actual_loc);
805 if (!node) {
806 pr_debug("failed to get kfd node by id 0x%x\n", prange->actual_loc);
807 return -ENODEV;
808 }
809 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n",
810 prange->svms, prange, prange->start, prange->last,
811 prange->actual_loc);
812
813 start = prange->start << PAGE_SHIFT;
814 end = (prange->last + 1) << PAGE_SHIFT;
815
816 for (addr = start; addr < end;) {
817 unsigned long next;
818
819 vma = vma_lookup(mm, addr);
820 if (!vma) {
821 pr_debug("failed to find vma for prange %p\n", prange);
822 r = -EFAULT;
823 break;
824 }
825
826 next = min(vma->vm_end, end);
827 r = svm_migrate_vma_to_ram(node, prange, vma, addr, next, trigger,
828 fault_page);
829 if (r < 0) {
830 pr_debug("failed %ld to migrate prange %p\n", r, prange);
831 break;
832 } else {
833 upages += r;
834 }
835 addr = next;
836 }
837
838 if (r >= 0 && !upages) {
839 svm_range_vram_node_free(prange);
840 prange->actual_loc = 0;
841 }
842
843 return r < 0 ? r : 0;
844 }
845
846 /**
847 * svm_migrate_vram_to_vram - migrate svm range from device to device
848 * @prange: range structure
849 * @best_loc: the device to migrate to
850 * @mm: process mm, use current->mm if NULL
851 * @trigger: reason of migration
852 *
853 * Context: Process context, caller hold mmap read lock, svms lock, prange lock
854 *
855 * Return:
856 * 0 - OK, otherwise error code
857 */
858 static int
svm_migrate_vram_to_vram(struct svm_range * prange,uint32_t best_loc,struct mm_struct * mm,uint32_t trigger)859 svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,
860 struct mm_struct *mm, uint32_t trigger)
861 {
862 int r, retries = 3;
863
864 /*
865 * TODO: for both devices with PCIe large bar or on same xgmi hive, skip
866 * system memory as migration bridge
867 */
868
869 pr_debug("from gpu 0x%x to gpu 0x%x\n", prange->actual_loc, best_loc);
870
871 do {
872 r = svm_migrate_vram_to_ram(prange, mm, trigger, NULL);
873 if (r)
874 return r;
875 } while (prange->actual_loc && --retries);
876
877 if (prange->actual_loc)
878 return -EDEADLK;
879
880 return svm_migrate_ram_to_vram(prange, best_loc, mm, trigger);
881 }
882
883 int
svm_migrate_to_vram(struct svm_range * prange,uint32_t best_loc,struct mm_struct * mm,uint32_t trigger)884 svm_migrate_to_vram(struct svm_range *prange, uint32_t best_loc,
885 struct mm_struct *mm, uint32_t trigger)
886 {
887 if (!prange->actual_loc)
888 return svm_migrate_ram_to_vram(prange, best_loc, mm, trigger);
889 else
890 return svm_migrate_vram_to_vram(prange, best_loc, mm, trigger);
891
892 }
893
894 /**
895 * svm_migrate_to_ram - CPU page fault handler
896 * @vmf: CPU vm fault vma, address
897 *
898 * Context: vm fault handler, caller holds the mmap read lock
899 *
900 * Return:
901 * 0 - OK
902 * VM_FAULT_SIGBUS - notice application to have SIGBUS page fault
903 */
svm_migrate_to_ram(struct vm_fault * vmf)904 static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf)
905 {
906 unsigned long addr = vmf->address;
907 struct svm_range_bo *svm_bo;
908 enum svm_work_list_ops op;
909 struct svm_range *parent;
910 struct svm_range *prange;
911 struct kfd_process *p;
912 struct mm_struct *mm;
913 int r = 0;
914
915 svm_bo = vmf->page->zone_device_data;
916 if (!svm_bo) {
917 pr_debug("failed get device page at addr 0x%lx\n", addr);
918 return VM_FAULT_SIGBUS;
919 }
920 if (!mmget_not_zero(svm_bo->eviction_fence->mm)) {
921 pr_debug("addr 0x%lx of process mm is destroyed\n", addr);
922 return VM_FAULT_SIGBUS;
923 }
924
925 mm = svm_bo->eviction_fence->mm;
926 if (mm != vmf->vma->vm_mm)
927 pr_debug("addr 0x%lx is COW mapping in child process\n", addr);
928
929 p = kfd_lookup_process_by_mm(mm);
930 if (!p) {
931 pr_debug("failed find process at fault address 0x%lx\n", addr);
932 r = VM_FAULT_SIGBUS;
933 goto out_mmput;
934 }
935 if (READ_ONCE(p->svms.faulting_task) == current) {
936 pr_debug("skipping ram migration\n");
937 r = 0;
938 goto out_unref_process;
939 }
940
941 pr_debug("CPU page fault svms 0x%p address 0x%lx\n", &p->svms, addr);
942 addr >>= PAGE_SHIFT;
943
944 mutex_lock(&p->svms.lock);
945
946 prange = svm_range_from_addr(&p->svms, addr, &parent);
947 if (!prange) {
948 pr_debug("failed get range svms 0x%p addr 0x%lx\n", &p->svms, addr);
949 r = -EFAULT;
950 goto out_unlock_svms;
951 }
952
953 mutex_lock(&parent->migrate_mutex);
954 if (prange != parent)
955 mutex_lock_nested(&prange->migrate_mutex, 1);
956
957 if (!prange->actual_loc)
958 goto out_unlock_prange;
959
960 svm_range_lock(parent);
961 if (prange != parent)
962 mutex_lock_nested(&prange->lock, 1);
963 r = svm_range_split_by_granularity(p, mm, addr, parent, prange);
964 if (prange != parent)
965 mutex_unlock(&prange->lock);
966 svm_range_unlock(parent);
967 if (r) {
968 pr_debug("failed %d to split range by granularity\n", r);
969 goto out_unlock_prange;
970 }
971
972 r = svm_migrate_vram_to_ram(prange, vmf->vma->vm_mm,
973 KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU,
974 vmf->page);
975 if (r)
976 pr_debug("failed %d migrate svms 0x%p range 0x%p [0x%lx 0x%lx]\n",
977 r, prange->svms, prange, prange->start, prange->last);
978
979 /* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
980 if (p->xnack_enabled && parent == prange)
981 op = SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP;
982 else
983 op = SVM_OP_UPDATE_RANGE_NOTIFIER;
984 svm_range_add_list_work(&p->svms, parent, mm, op);
985 schedule_deferred_list_work(&p->svms);
986
987 out_unlock_prange:
988 if (prange != parent)
989 mutex_unlock(&prange->migrate_mutex);
990 mutex_unlock(&parent->migrate_mutex);
991 out_unlock_svms:
992 mutex_unlock(&p->svms.lock);
993 out_unref_process:
994 pr_debug("CPU fault svms 0x%p address 0x%lx done\n", &p->svms, addr);
995 kfd_unref_process(p);
996 out_mmput:
997 mmput(mm);
998 return r ? VM_FAULT_SIGBUS : 0;
999 }
1000
1001 static const struct dev_pagemap_ops svm_migrate_pgmap_ops = {
1002 .page_free = svm_migrate_page_free,
1003 .migrate_to_ram = svm_migrate_to_ram,
1004 };
1005
1006 /* Each VRAM page uses sizeof(struct page) on system memory */
1007 #define SVM_HMM_PAGE_STRUCT_SIZE(size) ((size)/PAGE_SIZE * sizeof(struct page))
1008
kgd2kfd_init_zone_device(struct amdgpu_device * adev)1009 int kgd2kfd_init_zone_device(struct amdgpu_device *adev)
1010 {
1011 struct amdgpu_kfd_dev *kfddev = &adev->kfd;
1012 struct dev_pagemap *pgmap;
1013 struct resource *res = NULL;
1014 unsigned long size;
1015 void *r;
1016
1017 /* Page migration works on gfx9 or newer */
1018 if (adev->ip_versions[GC_HWIP][0] < IP_VERSION(9, 0, 1))
1019 return -EINVAL;
1020
1021 if (adev->gmc.is_app_apu)
1022 return 0;
1023
1024 pgmap = &kfddev->pgmap;
1025 memset(pgmap, 0, sizeof(*pgmap));
1026
1027 /* TODO: register all vram to HMM for now.
1028 * should remove reserved size
1029 */
1030 size = ALIGN(adev->gmc.real_vram_size, 2ULL << 20);
1031 if (adev->gmc.xgmi.connected_to_cpu) {
1032 pgmap->range.start = adev->gmc.aper_base;
1033 pgmap->range.end = adev->gmc.aper_base + adev->gmc.aper_size - 1;
1034 pgmap->type = MEMORY_DEVICE_COHERENT;
1035 } else {
1036 res = devm_request_free_mem_region(adev->dev, &iomem_resource, size);
1037 if (IS_ERR(res))
1038 return PTR_ERR(res);
1039 pgmap->range.start = res->start;
1040 pgmap->range.end = res->end;
1041 pgmap->type = MEMORY_DEVICE_PRIVATE;
1042 }
1043
1044 pgmap->nr_range = 1;
1045 pgmap->ops = &svm_migrate_pgmap_ops;
1046 pgmap->owner = SVM_ADEV_PGMAP_OWNER(adev);
1047 pgmap->flags = 0;
1048 /* Device manager releases device-specific resources, memory region and
1049 * pgmap when driver disconnects from device.
1050 */
1051 r = devm_memremap_pages(adev->dev, pgmap);
1052 if (IS_ERR(r)) {
1053 pr_err("failed to register HMM device memory\n");
1054 if (pgmap->type == MEMORY_DEVICE_PRIVATE)
1055 devm_release_mem_region(adev->dev, res->start, resource_size(res));
1056 /* Disable SVM support capability */
1057 pgmap->type = 0;
1058 return PTR_ERR(r);
1059 }
1060
1061 pr_debug("reserve %ldMB system memory for VRAM pages struct\n",
1062 SVM_HMM_PAGE_STRUCT_SIZE(size) >> 20);
1063
1064 amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size));
1065
1066 pr_info("HMM registered %ldMB device memory\n", size >> 20);
1067
1068 return 0;
1069 }
1070