Commit | Line | Data |
---|---|---|
76cbbead CH |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Device Memory Migration functionality. | |
4 | * | |
5 | * Originally written by Jérôme Glisse. | |
6 | */ | |
7 | #include <linux/export.h> | |
8 | #include <linux/memremap.h> | |
9 | #include <linux/migrate.h> | |
fd35ca3d | 10 | #include <linux/mm.h> |
76cbbead CH |
11 | #include <linux/mm_inline.h> |
12 | #include <linux/mmu_notifier.h> | |
13 | #include <linux/oom.h> | |
14 | #include <linux/pagewalk.h> | |
15 | #include <linux/rmap.h> | |
16 | #include <linux/swapops.h> | |
17 | #include <asm/tlbflush.h> | |
18 | #include "internal.h" | |
19 | ||
20 | static int migrate_vma_collect_skip(unsigned long start, | |
21 | unsigned long end, | |
22 | struct mm_walk *walk) | |
23 | { | |
24 | struct migrate_vma *migrate = walk->private; | |
25 | unsigned long addr; | |
26 | ||
27 | for (addr = start; addr < end; addr += PAGE_SIZE) { | |
28 | migrate->dst[migrate->npages] = 0; | |
29 | migrate->src[migrate->npages++] = 0; | |
30 | } | |
31 | ||
32 | return 0; | |
33 | } | |
34 | ||
35 | static int migrate_vma_collect_hole(unsigned long start, | |
36 | unsigned long end, | |
37 | __always_unused int depth, | |
38 | struct mm_walk *walk) | |
39 | { | |
40 | struct migrate_vma *migrate = walk->private; | |
41 | unsigned long addr; | |
42 | ||
43 | /* Only allow populating anonymous memory. */ | |
44 | if (!vma_is_anonymous(walk->vma)) | |
45 | return migrate_vma_collect_skip(start, end, walk); | |
46 | ||
47 | for (addr = start; addr < end; addr += PAGE_SIZE) { | |
48 | migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; | |
49 | migrate->dst[migrate->npages] = 0; | |
50 | migrate->npages++; | |
51 | migrate->cpages++; | |
52 | } | |
53 | ||
54 | return 0; | |
55 | } | |
56 | ||
57 | static int migrate_vma_collect_pmd(pmd_t *pmdp, | |
58 | unsigned long start, | |
59 | unsigned long end, | |
60 | struct mm_walk *walk) | |
61 | { | |
62 | struct migrate_vma *migrate = walk->private; | |
63 | struct vm_area_struct *vma = walk->vma; | |
64 | struct mm_struct *mm = vma->vm_mm; | |
65 | unsigned long addr = start, unmapped = 0; | |
66 | spinlock_t *ptl; | |
67 | pte_t *ptep; | |
68 | ||
69 | again: | |
70 | if (pmd_none(*pmdp)) | |
71 | return migrate_vma_collect_hole(start, end, -1, walk); | |
72 | ||
73 | if (pmd_trans_huge(*pmdp)) { | |
74 | struct page *page; | |
75 | ||
76 | ptl = pmd_lock(mm, pmdp); | |
77 | if (unlikely(!pmd_trans_huge(*pmdp))) { | |
78 | spin_unlock(ptl); | |
79 | goto again; | |
80 | } | |
81 | ||
82 | page = pmd_page(*pmdp); | |
83 | if (is_huge_zero_page(page)) { | |
84 | spin_unlock(ptl); | |
85 | split_huge_pmd(vma, pmdp, addr); | |
86 | if (pmd_trans_unstable(pmdp)) | |
87 | return migrate_vma_collect_skip(start, end, | |
88 | walk); | |
89 | } else { | |
90 | int ret; | |
91 | ||
92 | get_page(page); | |
93 | spin_unlock(ptl); | |
94 | if (unlikely(!trylock_page(page))) | |
95 | return migrate_vma_collect_skip(start, end, | |
96 | walk); | |
97 | ret = split_huge_page(page); | |
98 | unlock_page(page); | |
99 | put_page(page); | |
100 | if (ret) | |
101 | return migrate_vma_collect_skip(start, end, | |
102 | walk); | |
103 | if (pmd_none(*pmdp)) | |
104 | return migrate_vma_collect_hole(start, end, -1, | |
105 | walk); | |
106 | } | |
107 | } | |
108 | ||
109 | if (unlikely(pmd_bad(*pmdp))) | |
110 | return migrate_vma_collect_skip(start, end, walk); | |
111 | ||
112 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
113 | arch_enter_lazy_mmu_mode(); | |
114 | ||
115 | for (; addr < end; addr += PAGE_SIZE, ptep++) { | |
116 | unsigned long mpfn = 0, pfn; | |
117 | struct page *page; | |
118 | swp_entry_t entry; | |
119 | pte_t pte; | |
120 | ||
121 | pte = *ptep; | |
122 | ||
123 | if (pte_none(pte)) { | |
124 | if (vma_is_anonymous(vma)) { | |
125 | mpfn = MIGRATE_PFN_MIGRATE; | |
126 | migrate->cpages++; | |
127 | } | |
128 | goto next; | |
129 | } | |
130 | ||
131 | if (!pte_present(pte)) { | |
132 | /* | |
133 | * Only care about unaddressable device page special | |
134 | * page table entry. Other special swap entries are not | |
135 | * migratable, and we ignore regular swapped page. | |
136 | */ | |
137 | entry = pte_to_swp_entry(pte); | |
138 | if (!is_device_private_entry(entry)) | |
139 | goto next; | |
140 | ||
141 | page = pfn_swap_entry_to_page(entry); | |
142 | if (!(migrate->flags & | |
143 | MIGRATE_VMA_SELECT_DEVICE_PRIVATE) || | |
144 | page->pgmap->owner != migrate->pgmap_owner) | |
145 | goto next; | |
146 | ||
147 | mpfn = migrate_pfn(page_to_pfn(page)) | | |
148 | MIGRATE_PFN_MIGRATE; | |
149 | if (is_writable_device_private_entry(entry)) | |
150 | mpfn |= MIGRATE_PFN_WRITE; | |
151 | } else { | |
76cbbead | 152 | pfn = pte_pfn(pte); |
dd19e6d8 AS |
153 | if (is_zero_pfn(pfn) && |
154 | (migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) { | |
76cbbead CH |
155 | mpfn = MIGRATE_PFN_MIGRATE; |
156 | migrate->cpages++; | |
157 | goto next; | |
158 | } | |
159 | page = vm_normal_page(migrate->vma, addr, pte); | |
dd19e6d8 AS |
160 | if (page && !is_zone_device_page(page) && |
161 | !(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) | |
162 | goto next; | |
163 | else if (page && is_device_coherent_page(page) && | |
164 | (!(migrate->flags & MIGRATE_VMA_SELECT_DEVICE_COHERENT) || | |
165 | page->pgmap->owner != migrate->pgmap_owner)) | |
166 | goto next; | |
76cbbead CH |
167 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; |
168 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; | |
169 | } | |
170 | ||
171 | /* FIXME support THP */ | |
172 | if (!page || !page->mapping || PageTransCompound(page)) { | |
173 | mpfn = 0; | |
174 | goto next; | |
175 | } | |
176 | ||
177 | /* | |
178 | * By getting a reference on the page we pin it and that blocks | |
179 | * any kind of migration. Side effect is that it "freezes" the | |
180 | * pte. | |
181 | * | |
182 | * We drop this reference after isolating the page from the lru | |
183 | * for non device page (device page are not on the lru and thus | |
184 | * can't be dropped from it). | |
185 | */ | |
186 | get_page(page); | |
187 | ||
188 | /* | |
0742e490 AP |
189 | * We rely on trylock_page() to avoid deadlock between |
190 | * concurrent migrations where each is waiting on the others | |
191 | * page lock. If we can't immediately lock the page we fail this | |
192 | * migration as it is only best effort anyway. | |
193 | * | |
194 | * If we can lock the page it's safe to set up a migration entry | |
195 | * now. In the common case where the page is mapped once in a | |
196 | * single process setting up the migration entry now is an | |
197 | * optimisation to avoid walking the rmap later with | |
198 | * try_to_migrate(). | |
76cbbead CH |
199 | */ |
200 | if (trylock_page(page)) { | |
6c287605 | 201 | bool anon_exclusive; |
76cbbead CH |
202 | pte_t swp_pte; |
203 | ||
a3589e1d | 204 | flush_cache_page(vma, addr, pte_pfn(*ptep)); |
6c287605 DH |
205 | anon_exclusive = PageAnon(page) && PageAnonExclusive(page); |
206 | if (anon_exclusive) { | |
fd35ca3d | 207 | pte = ptep_clear_flush(vma, addr, ptep); |
6c287605 DH |
208 | |
209 | if (page_try_share_anon_rmap(page)) { | |
210 | set_pte_at(mm, addr, ptep, pte); | |
211 | unlock_page(page); | |
212 | put_page(page); | |
213 | mpfn = 0; | |
214 | goto next; | |
215 | } | |
216 | } else { | |
fd35ca3d | 217 | pte = ptep_get_and_clear(mm, addr, ptep); |
6c287605 DH |
218 | } |
219 | ||
76cbbead | 220 | migrate->cpages++; |
76cbbead | 221 | |
fd35ca3d AP |
222 | /* Set the dirty flag on the folio now the pte is gone. */ |
223 | if (pte_dirty(pte)) | |
224 | folio_mark_dirty(page_folio(page)); | |
225 | ||
76cbbead CH |
226 | /* Setup special migration page table entry */ |
227 | if (mpfn & MIGRATE_PFN_WRITE) | |
228 | entry = make_writable_migration_entry( | |
229 | page_to_pfn(page)); | |
6c287605 DH |
230 | else if (anon_exclusive) |
231 | entry = make_readable_exclusive_migration_entry( | |
232 | page_to_pfn(page)); | |
76cbbead CH |
233 | else |
234 | entry = make_readable_migration_entry( | |
235 | page_to_pfn(page)); | |
2e346877 PX |
236 | if (pte_present(pte)) { |
237 | if (pte_young(pte)) | |
238 | entry = make_migration_entry_young(entry); | |
239 | if (pte_dirty(pte)) | |
240 | entry = make_migration_entry_dirty(entry); | |
241 | } | |
76cbbead CH |
242 | swp_pte = swp_entry_to_pte(entry); |
243 | if (pte_present(pte)) { | |
244 | if (pte_soft_dirty(pte)) | |
245 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
246 | if (pte_uffd_wp(pte)) | |
247 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
248 | } else { | |
249 | if (pte_swp_soft_dirty(pte)) | |
250 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
251 | if (pte_swp_uffd_wp(pte)) | |
252 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
253 | } | |
254 | set_pte_at(mm, addr, ptep, swp_pte); | |
255 | ||
256 | /* | |
257 | * This is like regular unmap: we remove the rmap and | |
258 | * drop page refcount. Page won't be freed, as we took | |
259 | * a reference just above. | |
260 | */ | |
261 | page_remove_rmap(page, vma, false); | |
262 | put_page(page); | |
263 | ||
264 | if (pte_present(pte)) | |
265 | unmapped++; | |
266 | } else { | |
267 | put_page(page); | |
268 | mpfn = 0; | |
269 | } | |
270 | ||
271 | next: | |
272 | migrate->dst[migrate->npages] = 0; | |
273 | migrate->src[migrate->npages++] = mpfn; | |
274 | } | |
76cbbead CH |
275 | |
276 | /* Only flush the TLB if we actually modified any entries */ | |
277 | if (unmapped) | |
278 | flush_tlb_range(walk->vma, start, end); | |
279 | ||
60bae737 AP |
280 | arch_leave_lazy_mmu_mode(); |
281 | pte_unmap_unlock(ptep - 1, ptl); | |
282 | ||
76cbbead CH |
283 | return 0; |
284 | } | |
285 | ||
286 | static const struct mm_walk_ops migrate_vma_walk_ops = { | |
287 | .pmd_entry = migrate_vma_collect_pmd, | |
288 | .pte_hole = migrate_vma_collect_hole, | |
289 | }; | |
290 | ||
291 | /* | |
292 | * migrate_vma_collect() - collect pages over a range of virtual addresses | |
293 | * @migrate: migrate struct containing all migration information | |
294 | * | |
295 | * This will walk the CPU page table. For each virtual address backed by a | |
296 | * valid page, it updates the src array and takes a reference on the page, in | |
297 | * order to pin the page until we lock it and unmap it. | |
298 | */ | |
299 | static void migrate_vma_collect(struct migrate_vma *migrate) | |
300 | { | |
301 | struct mmu_notifier_range range; | |
302 | ||
303 | /* | |
304 | * Note that the pgmap_owner is passed to the mmu notifier callback so | |
305 | * that the registered device driver can skip invalidating device | |
306 | * private page mappings that won't be migrated. | |
307 | */ | |
308 | mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0, | |
7d4a8be0 | 309 | migrate->vma->vm_mm, migrate->start, migrate->end, |
76cbbead CH |
310 | migrate->pgmap_owner); |
311 | mmu_notifier_invalidate_range_start(&range); | |
312 | ||
313 | walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end, | |
314 | &migrate_vma_walk_ops, migrate); | |
315 | ||
316 | mmu_notifier_invalidate_range_end(&range); | |
317 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); | |
318 | } | |
319 | ||
320 | /* | |
321 | * migrate_vma_check_page() - check if page is pinned or not | |
322 | * @page: struct page to check | |
323 | * | |
324 | * Pinned pages cannot be migrated. This is the same test as in | |
325 | * folio_migrate_mapping(), except that here we allow migration of a | |
326 | * ZONE_DEVICE page. | |
327 | */ | |
16ce101d | 328 | static bool migrate_vma_check_page(struct page *page, struct page *fault_page) |
76cbbead CH |
329 | { |
330 | /* | |
331 | * One extra ref because caller holds an extra reference, either from | |
332 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for | |
333 | * a device page. | |
334 | */ | |
16ce101d | 335 | int extra = 1 + (page == fault_page); |
76cbbead CH |
336 | |
337 | /* | |
338 | * FIXME support THP (transparent huge page), it is bit more complex to | |
339 | * check them than regular pages, because they can be mapped with a pmd | |
340 | * or with a pte (split pte mapping). | |
341 | */ | |
342 | if (PageCompound(page)) | |
343 | return false; | |
344 | ||
345 | /* Page from ZONE_DEVICE have one extra reference */ | |
346 | if (is_zone_device_page(page)) | |
347 | extra++; | |
348 | ||
349 | /* For file back page */ | |
350 | if (page_mapping(page)) | |
351 | extra += 1 + page_has_private(page); | |
352 | ||
353 | if ((page_count(page) - extra) > page_mapcount(page)) | |
354 | return false; | |
355 | ||
356 | return true; | |
357 | } | |
358 | ||
359 | /* | |
44af0b45 AP |
360 | * Unmaps pages for migration. Returns number of source pfns marked as |
361 | * migrating. | |
76cbbead | 362 | */ |
241f6885 AP |
363 | static unsigned long migrate_device_unmap(unsigned long *src_pfns, |
364 | unsigned long npages, | |
365 | struct page *fault_page) | |
76cbbead | 366 | { |
76cbbead CH |
367 | unsigned long i, restore = 0; |
368 | bool allow_drain = true; | |
241f6885 | 369 | unsigned long unmapped = 0; |
76cbbead CH |
370 | |
371 | lru_add_drain(); | |
372 | ||
373 | for (i = 0; i < npages; i++) { | |
241f6885 | 374 | struct page *page = migrate_pfn_to_page(src_pfns[i]); |
4b8554c5 | 375 | struct folio *folio; |
76cbbead | 376 | |
44af0b45 AP |
377 | if (!page) { |
378 | if (src_pfns[i] & MIGRATE_PFN_MIGRATE) | |
379 | unmapped++; | |
76cbbead | 380 | continue; |
44af0b45 | 381 | } |
76cbbead CH |
382 | |
383 | /* ZONE_DEVICE pages are not on LRU */ | |
384 | if (!is_zone_device_page(page)) { | |
385 | if (!PageLRU(page) && allow_drain) { | |
386 | /* Drain CPU's pagevec */ | |
387 | lru_add_drain_all(); | |
388 | allow_drain = false; | |
389 | } | |
390 | ||
f7f9c00d | 391 | if (!isolate_lru_page(page)) { |
241f6885 | 392 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
393 | restore++; |
394 | continue; | |
395 | } | |
396 | ||
397 | /* Drop the reference we took in collect */ | |
398 | put_page(page); | |
399 | } | |
400 | ||
4b8554c5 MWO |
401 | folio = page_folio(page); |
402 | if (folio_mapped(folio)) | |
403 | try_to_migrate(folio, 0); | |
76cbbead | 404 | |
16ce101d | 405 | if (page_mapped(page) || |
241f6885 | 406 | !migrate_vma_check_page(page, fault_page)) { |
76cbbead CH |
407 | if (!is_zone_device_page(page)) { |
408 | get_page(page); | |
409 | putback_lru_page(page); | |
410 | } | |
411 | ||
241f6885 | 412 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
413 | restore++; |
414 | continue; | |
415 | } | |
241f6885 AP |
416 | |
417 | unmapped++; | |
76cbbead CH |
418 | } |
419 | ||
420 | for (i = 0; i < npages && restore; i++) { | |
241f6885 | 421 | struct page *page = migrate_pfn_to_page(src_pfns[i]); |
4eecb8b9 | 422 | struct folio *folio; |
76cbbead | 423 | |
241f6885 | 424 | if (!page || (src_pfns[i] & MIGRATE_PFN_MIGRATE)) |
76cbbead CH |
425 | continue; |
426 | ||
4eecb8b9 MWO |
427 | folio = page_folio(page); |
428 | remove_migration_ptes(folio, folio, false); | |
76cbbead | 429 | |
241f6885 | 430 | src_pfns[i] = 0; |
4eecb8b9 MWO |
431 | folio_unlock(folio); |
432 | folio_put(folio); | |
76cbbead CH |
433 | restore--; |
434 | } | |
241f6885 AP |
435 | |
436 | return unmapped; | |
437 | } | |
438 | ||
439 | /* | |
440 | * migrate_vma_unmap() - replace page mapping with special migration pte entry | |
441 | * @migrate: migrate struct containing all migration information | |
442 | * | |
443 | * Isolate pages from the LRU and replace mappings (CPU page table pte) with a | |
444 | * special migration pte entry and check if it has been pinned. Pinned pages are | |
445 | * restored because we cannot migrate them. | |
446 | * | |
447 | * This is the last step before we call the device driver callback to allocate | |
448 | * destination memory and copy contents of original page over to new page. | |
449 | */ | |
450 | static void migrate_vma_unmap(struct migrate_vma *migrate) | |
451 | { | |
452 | migrate->cpages = migrate_device_unmap(migrate->src, migrate->npages, | |
453 | migrate->fault_page); | |
76cbbead CH |
454 | } |
455 | ||
456 | /** | |
457 | * migrate_vma_setup() - prepare to migrate a range of memory | |
458 | * @args: contains the vma, start, and pfns arrays for the migration | |
459 | * | |
460 | * Returns: negative errno on failures, 0 when 0 or more pages were migrated | |
461 | * without an error. | |
462 | * | |
463 | * Prepare to migrate a range of memory virtual address range by collecting all | |
464 | * the pages backing each virtual address in the range, saving them inside the | |
465 | * src array. Then lock those pages and unmap them. Once the pages are locked | |
466 | * and unmapped, check whether each page is pinned or not. Pages that aren't | |
467 | * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the | |
468 | * corresponding src array entry. Then restores any pages that are pinned, by | |
469 | * remapping and unlocking those pages. | |
470 | * | |
471 | * The caller should then allocate destination memory and copy source memory to | |
472 | * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE | |
473 | * flag set). Once these are allocated and copied, the caller must update each | |
474 | * corresponding entry in the dst array with the pfn value of the destination | |
475 | * page and with MIGRATE_PFN_VALID. Destination pages must be locked via | |
476 | * lock_page(). | |
477 | * | |
478 | * Note that the caller does not have to migrate all the pages that are marked | |
479 | * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from | |
480 | * device memory to system memory. If the caller cannot migrate a device page | |
481 | * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe | |
482 | * consequences for the userspace process, so it must be avoided if at all | |
483 | * possible. | |
484 | * | |
485 | * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we | |
486 | * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus | |
487 | * allowing the caller to allocate device memory for those unbacked virtual | |
488 | * addresses. For this the caller simply has to allocate device memory and | |
489 | * properly set the destination entry like for regular migration. Note that | |
490 | * this can still fail, and thus inside the device driver you must check if the | |
491 | * migration was successful for those entries after calling migrate_vma_pages(), | |
492 | * just like for regular migration. | |
493 | * | |
494 | * After that, the callers must call migrate_vma_pages() to go over each entry | |
495 | * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag | |
496 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, | |
497 | * then migrate_vma_pages() to migrate struct page information from the source | |
498 | * struct page to the destination struct page. If it fails to migrate the | |
499 | * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the | |
500 | * src array. | |
501 | * | |
502 | * At this point all successfully migrated pages have an entry in the src | |
503 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst | |
504 | * array entry with MIGRATE_PFN_VALID flag set. | |
505 | * | |
506 | * Once migrate_vma_pages() returns the caller may inspect which pages were | |
507 | * successfully migrated, and which were not. Successfully migrated pages will | |
508 | * have the MIGRATE_PFN_MIGRATE flag set for their src array entry. | |
509 | * | |
510 | * It is safe to update device page table after migrate_vma_pages() because | |
511 | * both destination and source page are still locked, and the mmap_lock is held | |
512 | * in read mode (hence no one can unmap the range being migrated). | |
513 | * | |
514 | * Once the caller is done cleaning up things and updating its page table (if it | |
515 | * chose to do so, this is not an obligation) it finally calls | |
516 | * migrate_vma_finalize() to update the CPU page table to point to new pages | |
517 | * for successfully migrated pages or otherwise restore the CPU page table to | |
518 | * point to the original source pages. | |
519 | */ | |
520 | int migrate_vma_setup(struct migrate_vma *args) | |
521 | { | |
522 | long nr_pages = (args->end - args->start) >> PAGE_SHIFT; | |
523 | ||
524 | args->start &= PAGE_MASK; | |
525 | args->end &= PAGE_MASK; | |
526 | if (!args->vma || is_vm_hugetlb_page(args->vma) || | |
527 | (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma)) | |
528 | return -EINVAL; | |
529 | if (nr_pages <= 0) | |
530 | return -EINVAL; | |
531 | if (args->start < args->vma->vm_start || | |
532 | args->start >= args->vma->vm_end) | |
533 | return -EINVAL; | |
534 | if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end) | |
535 | return -EINVAL; | |
536 | if (!args->src || !args->dst) | |
537 | return -EINVAL; | |
16ce101d AP |
538 | if (args->fault_page && !is_device_private_page(args->fault_page)) |
539 | return -EINVAL; | |
76cbbead CH |
540 | |
541 | memset(args->src, 0, sizeof(*args->src) * nr_pages); | |
542 | args->cpages = 0; | |
543 | args->npages = 0; | |
544 | ||
545 | migrate_vma_collect(args); | |
546 | ||
547 | if (args->cpages) | |
548 | migrate_vma_unmap(args); | |
549 | ||
550 | /* | |
551 | * At this point pages are locked and unmapped, and thus they have | |
552 | * stable content and can safely be copied to destination memory that | |
553 | * is allocated by the drivers. | |
554 | */ | |
555 | return 0; | |
556 | ||
557 | } | |
558 | EXPORT_SYMBOL(migrate_vma_setup); | |
559 | ||
560 | /* | |
561 | * This code closely matches the code in: | |
562 | * __handle_mm_fault() | |
563 | * handle_pte_fault() | |
564 | * do_anonymous_page() | |
565 | * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE | |
f25cbb7a | 566 | * private or coherent page. |
76cbbead CH |
567 | */ |
568 | static void migrate_vma_insert_page(struct migrate_vma *migrate, | |
569 | unsigned long addr, | |
570 | struct page *page, | |
571 | unsigned long *src) | |
572 | { | |
573 | struct vm_area_struct *vma = migrate->vma; | |
574 | struct mm_struct *mm = vma->vm_mm; | |
575 | bool flush = false; | |
576 | spinlock_t *ptl; | |
577 | pte_t entry; | |
578 | pgd_t *pgdp; | |
579 | p4d_t *p4dp; | |
580 | pud_t *pudp; | |
581 | pmd_t *pmdp; | |
582 | pte_t *ptep; | |
583 | ||
584 | /* Only allow populating anonymous memory */ | |
585 | if (!vma_is_anonymous(vma)) | |
586 | goto abort; | |
587 | ||
588 | pgdp = pgd_offset(mm, addr); | |
589 | p4dp = p4d_alloc(mm, pgdp, addr); | |
590 | if (!p4dp) | |
591 | goto abort; | |
592 | pudp = pud_alloc(mm, p4dp, addr); | |
593 | if (!pudp) | |
594 | goto abort; | |
595 | pmdp = pmd_alloc(mm, pudp, addr); | |
596 | if (!pmdp) | |
597 | goto abort; | |
598 | ||
599 | if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) | |
600 | goto abort; | |
601 | ||
602 | /* | |
603 | * Use pte_alloc() instead of pte_alloc_map(). We can't run | |
604 | * pte_offset_map() on pmds where a huge pmd might be created | |
605 | * from a different thread. | |
606 | * | |
607 | * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when | |
608 | * parallel threads are excluded by other means. | |
609 | * | |
610 | * Here we only have mmap_read_lock(mm). | |
611 | */ | |
612 | if (pte_alloc(mm, pmdp)) | |
613 | goto abort; | |
614 | ||
615 | /* See the comment in pte_alloc_one_map() */ | |
616 | if (unlikely(pmd_trans_unstable(pmdp))) | |
617 | goto abort; | |
618 | ||
619 | if (unlikely(anon_vma_prepare(vma))) | |
620 | goto abort; | |
621 | if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL)) | |
622 | goto abort; | |
623 | ||
624 | /* | |
625 | * The memory barrier inside __SetPageUptodate makes sure that | |
626 | * preceding stores to the page contents become visible before | |
627 | * the set_pte_at() write. | |
628 | */ | |
629 | __SetPageUptodate(page); | |
630 | ||
631 | if (is_device_private_page(page)) { | |
632 | swp_entry_t swp_entry; | |
633 | ||
634 | if (vma->vm_flags & VM_WRITE) | |
635 | swp_entry = make_writable_device_private_entry( | |
636 | page_to_pfn(page)); | |
637 | else | |
638 | swp_entry = make_readable_device_private_entry( | |
639 | page_to_pfn(page)); | |
640 | entry = swp_entry_to_pte(swp_entry); | |
641 | } else { | |
f25cbb7a AS |
642 | if (is_zone_device_page(page) && |
643 | !is_device_coherent_page(page)) { | |
76cbbead CH |
644 | pr_warn_once("Unsupported ZONE_DEVICE page type.\n"); |
645 | goto abort; | |
646 | } | |
647 | entry = mk_pte(page, vma->vm_page_prot); | |
648 | if (vma->vm_flags & VM_WRITE) | |
649 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
650 | } | |
651 | ||
652 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
653 | ||
654 | if (check_stable_address_space(mm)) | |
655 | goto unlock_abort; | |
656 | ||
657 | if (pte_present(*ptep)) { | |
658 | unsigned long pfn = pte_pfn(*ptep); | |
659 | ||
660 | if (!is_zero_pfn(pfn)) | |
661 | goto unlock_abort; | |
662 | flush = true; | |
663 | } else if (!pte_none(*ptep)) | |
664 | goto unlock_abort; | |
665 | ||
666 | /* | |
667 | * Check for userfaultfd but do not deliver the fault. Instead, | |
668 | * just back off. | |
669 | */ | |
670 | if (userfaultfd_missing(vma)) | |
671 | goto unlock_abort; | |
672 | ||
673 | inc_mm_counter(mm, MM_ANONPAGES); | |
40f2bbf7 | 674 | page_add_new_anon_rmap(page, vma, addr); |
76cbbead CH |
675 | if (!is_zone_device_page(page)) |
676 | lru_cache_add_inactive_or_unevictable(page, vma); | |
677 | get_page(page); | |
678 | ||
679 | if (flush) { | |
680 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
681 | ptep_clear_flush_notify(vma, addr, ptep); | |
682 | set_pte_at_notify(mm, addr, ptep, entry); | |
683 | update_mmu_cache(vma, addr, ptep); | |
684 | } else { | |
685 | /* No need to invalidate - it was non-present before */ | |
686 | set_pte_at(mm, addr, ptep, entry); | |
687 | update_mmu_cache(vma, addr, ptep); | |
688 | } | |
689 | ||
690 | pte_unmap_unlock(ptep, ptl); | |
691 | *src = MIGRATE_PFN_MIGRATE; | |
692 | return; | |
693 | ||
694 | unlock_abort: | |
695 | pte_unmap_unlock(ptep, ptl); | |
696 | abort: | |
697 | *src &= ~MIGRATE_PFN_MIGRATE; | |
698 | } | |
699 | ||
e778406b | 700 | static void __migrate_device_pages(unsigned long *src_pfns, |
241f6885 AP |
701 | unsigned long *dst_pfns, unsigned long npages, |
702 | struct migrate_vma *migrate) | |
76cbbead | 703 | { |
76cbbead | 704 | struct mmu_notifier_range range; |
241f6885 | 705 | unsigned long i; |
76cbbead CH |
706 | bool notified = false; |
707 | ||
241f6885 AP |
708 | for (i = 0; i < npages; i++) { |
709 | struct page *newpage = migrate_pfn_to_page(dst_pfns[i]); | |
710 | struct page *page = migrate_pfn_to_page(src_pfns[i]); | |
76cbbead CH |
711 | struct address_space *mapping; |
712 | int r; | |
713 | ||
714 | if (!newpage) { | |
241f6885 | 715 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
716 | continue; |
717 | } | |
718 | ||
719 | if (!page) { | |
241f6885 AP |
720 | unsigned long addr; |
721 | ||
e778406b AP |
722 | if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE)) |
723 | continue; | |
724 | ||
b05a79d4 AP |
725 | /* |
726 | * The only time there is no vma is when called from | |
727 | * migrate_device_coherent_page(). However this isn't | |
728 | * called if the page could not be unmapped. | |
729 | */ | |
241f6885 AP |
730 | VM_BUG_ON(!migrate); |
731 | addr = migrate->start + i*PAGE_SIZE; | |
76cbbead CH |
732 | if (!notified) { |
733 | notified = true; | |
734 | ||
735 | mmu_notifier_range_init_owner(&range, | |
7d4a8be0 | 736 | MMU_NOTIFY_MIGRATE, 0, |
76cbbead CH |
737 | migrate->vma->vm_mm, addr, migrate->end, |
738 | migrate->pgmap_owner); | |
739 | mmu_notifier_invalidate_range_start(&range); | |
740 | } | |
741 | migrate_vma_insert_page(migrate, addr, newpage, | |
241f6885 | 742 | &src_pfns[i]); |
76cbbead CH |
743 | continue; |
744 | } | |
745 | ||
746 | mapping = page_mapping(page); | |
747 | ||
f25cbb7a AS |
748 | if (is_device_private_page(newpage) || |
749 | is_device_coherent_page(newpage)) { | |
76cbbead | 750 | /* |
f25cbb7a AS |
751 | * For now only support anonymous memory migrating to |
752 | * device private or coherent memory. | |
76cbbead CH |
753 | */ |
754 | if (mapping) { | |
241f6885 | 755 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
756 | continue; |
757 | } | |
758 | } else if (is_zone_device_page(newpage)) { | |
759 | /* | |
760 | * Other types of ZONE_DEVICE page are not supported. | |
761 | */ | |
241f6885 | 762 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
763 | continue; |
764 | } | |
765 | ||
241f6885 | 766 | if (migrate && migrate->fault_page == page) |
16ce101d AP |
767 | r = migrate_folio_extra(mapping, page_folio(newpage), |
768 | page_folio(page), | |
769 | MIGRATE_SYNC_NO_COPY, 1); | |
770 | else | |
771 | r = migrate_folio(mapping, page_folio(newpage), | |
772 | page_folio(page), MIGRATE_SYNC_NO_COPY); | |
76cbbead | 773 | if (r != MIGRATEPAGE_SUCCESS) |
241f6885 | 774 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
775 | } |
776 | ||
777 | /* | |
778 | * No need to double call mmu_notifier->invalidate_range() callback as | |
779 | * the above ptep_clear_flush_notify() inside migrate_vma_insert_page() | |
780 | * did already call it. | |
781 | */ | |
782 | if (notified) | |
783 | mmu_notifier_invalidate_range_only_end(&range); | |
784 | } | |
76cbbead | 785 | |
e778406b AP |
786 | /** |
787 | * migrate_device_pages() - migrate meta-data from src page to dst page | |
788 | * @src_pfns: src_pfns returned from migrate_device_range() | |
789 | * @dst_pfns: array of pfns allocated by the driver to migrate memory to | |
790 | * @npages: number of pages in the range | |
791 | * | |
792 | * Equivalent to migrate_vma_pages(). This is called to migrate struct page | |
793 | * meta-data from source struct page to destination. | |
794 | */ | |
795 | void migrate_device_pages(unsigned long *src_pfns, unsigned long *dst_pfns, | |
796 | unsigned long npages) | |
797 | { | |
798 | __migrate_device_pages(src_pfns, dst_pfns, npages, NULL); | |
799 | } | |
800 | EXPORT_SYMBOL(migrate_device_pages); | |
801 | ||
76cbbead | 802 | /** |
241f6885 | 803 | * migrate_vma_pages() - migrate meta-data from src page to dst page |
76cbbead CH |
804 | * @migrate: migrate struct containing all migration information |
805 | * | |
241f6885 AP |
806 | * This migrates struct page meta-data from source struct page to destination |
807 | * struct page. This effectively finishes the migration from source page to the | |
808 | * destination page. | |
76cbbead | 809 | */ |
241f6885 AP |
810 | void migrate_vma_pages(struct migrate_vma *migrate) |
811 | { | |
e778406b | 812 | __migrate_device_pages(migrate->src, migrate->dst, migrate->npages, migrate); |
241f6885 AP |
813 | } |
814 | EXPORT_SYMBOL(migrate_vma_pages); | |
815 | ||
e778406b AP |
816 | /* |
817 | * migrate_device_finalize() - complete page migration | |
818 | * @src_pfns: src_pfns returned from migrate_device_range() | |
819 | * @dst_pfns: array of pfns allocated by the driver to migrate memory to | |
820 | * @npages: number of pages in the range | |
821 | * | |
822 | * Completes migration of the page by removing special migration entries. | |
823 | * Drivers must ensure copying of page data is complete and visible to the CPU | |
824 | * before calling this. | |
825 | */ | |
826 | void migrate_device_finalize(unsigned long *src_pfns, | |
827 | unsigned long *dst_pfns, unsigned long npages) | |
76cbbead | 828 | { |
76cbbead CH |
829 | unsigned long i; |
830 | ||
831 | for (i = 0; i < npages; i++) { | |
4eecb8b9 | 832 | struct folio *dst, *src; |
241f6885 AP |
833 | struct page *newpage = migrate_pfn_to_page(dst_pfns[i]); |
834 | struct page *page = migrate_pfn_to_page(src_pfns[i]); | |
76cbbead CH |
835 | |
836 | if (!page) { | |
837 | if (newpage) { | |
838 | unlock_page(newpage); | |
839 | put_page(newpage); | |
840 | } | |
841 | continue; | |
842 | } | |
843 | ||
241f6885 | 844 | if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE) || !newpage) { |
76cbbead CH |
845 | if (newpage) { |
846 | unlock_page(newpage); | |
847 | put_page(newpage); | |
848 | } | |
849 | newpage = page; | |
850 | } | |
851 | ||
4eecb8b9 MWO |
852 | src = page_folio(page); |
853 | dst = page_folio(newpage); | |
854 | remove_migration_ptes(src, dst, false); | |
855 | folio_unlock(src); | |
76cbbead CH |
856 | |
857 | if (is_zone_device_page(page)) | |
858 | put_page(page); | |
859 | else | |
860 | putback_lru_page(page); | |
861 | ||
862 | if (newpage != page) { | |
863 | unlock_page(newpage); | |
864 | if (is_zone_device_page(newpage)) | |
865 | put_page(newpage); | |
866 | else | |
867 | putback_lru_page(newpage); | |
868 | } | |
869 | } | |
870 | } | |
e778406b | 871 | EXPORT_SYMBOL(migrate_device_finalize); |
241f6885 AP |
872 | |
873 | /** | |
874 | * migrate_vma_finalize() - restore CPU page table entry | |
875 | * @migrate: migrate struct containing all migration information | |
876 | * | |
877 | * This replaces the special migration pte entry with either a mapping to the | |
878 | * new page if migration was successful for that page, or to the original page | |
879 | * otherwise. | |
880 | * | |
881 | * This also unlocks the pages and puts them back on the lru, or drops the extra | |
882 | * refcount, for device pages. | |
883 | */ | |
884 | void migrate_vma_finalize(struct migrate_vma *migrate) | |
885 | { | |
886 | migrate_device_finalize(migrate->src, migrate->dst, migrate->npages); | |
887 | } | |
76cbbead | 888 | EXPORT_SYMBOL(migrate_vma_finalize); |
b05a79d4 | 889 | |
e778406b AP |
890 | /** |
891 | * migrate_device_range() - migrate device private pfns to normal memory. | |
892 | * @src_pfns: array large enough to hold migrating source device private pfns. | |
893 | * @start: starting pfn in the range to migrate. | |
894 | * @npages: number of pages to migrate. | |
895 | * | |
896 | * migrate_vma_setup() is similar in concept to migrate_vma_setup() except that | |
897 | * instead of looking up pages based on virtual address mappings a range of | |
898 | * device pfns that should be migrated to system memory is used instead. | |
899 | * | |
900 | * This is useful when a driver needs to free device memory but doesn't know the | |
901 | * virtual mappings of every page that may be in device memory. For example this | |
902 | * is often the case when a driver is being unloaded or unbound from a device. | |
903 | * | |
904 | * Like migrate_vma_setup() this function will take a reference and lock any | |
905 | * migrating pages that aren't free before unmapping them. Drivers may then | |
906 | * allocate destination pages and start copying data from the device to CPU | |
907 | * memory before calling migrate_device_pages(). | |
908 | */ | |
909 | int migrate_device_range(unsigned long *src_pfns, unsigned long start, | |
910 | unsigned long npages) | |
911 | { | |
912 | unsigned long i, pfn; | |
913 | ||
914 | for (pfn = start, i = 0; i < npages; pfn++, i++) { | |
915 | struct page *page = pfn_to_page(pfn); | |
916 | ||
917 | if (!get_page_unless_zero(page)) { | |
918 | src_pfns[i] = 0; | |
919 | continue; | |
920 | } | |
921 | ||
922 | if (!trylock_page(page)) { | |
923 | src_pfns[i] = 0; | |
924 | put_page(page); | |
925 | continue; | |
926 | } | |
927 | ||
928 | src_pfns[i] = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; | |
929 | } | |
930 | ||
931 | migrate_device_unmap(src_pfns, npages, NULL); | |
932 | ||
933 | return 0; | |
934 | } | |
935 | EXPORT_SYMBOL(migrate_device_range); | |
936 | ||
b05a79d4 AP |
937 | /* |
938 | * Migrate a device coherent page back to normal memory. The caller should have | |
939 | * a reference on page which will be copied to the new page if migration is | |
940 | * successful or dropped on failure. | |
941 | */ | |
942 | int migrate_device_coherent_page(struct page *page) | |
943 | { | |
944 | unsigned long src_pfn, dst_pfn = 0; | |
b05a79d4 AP |
945 | struct page *dpage; |
946 | ||
947 | WARN_ON_ONCE(PageCompound(page)); | |
948 | ||
949 | lock_page(page); | |
950 | src_pfn = migrate_pfn(page_to_pfn(page)) | MIGRATE_PFN_MIGRATE; | |
b05a79d4 AP |
951 | |
952 | /* | |
953 | * We don't have a VMA and don't need to walk the page tables to find | |
954 | * the source page. So call migrate_vma_unmap() directly to unmap the | |
955 | * page as migrate_vma_setup() will fail if args.vma == NULL. | |
956 | */ | |
241f6885 | 957 | migrate_device_unmap(&src_pfn, 1, NULL); |
b05a79d4 AP |
958 | if (!(src_pfn & MIGRATE_PFN_MIGRATE)) |
959 | return -EBUSY; | |
960 | ||
961 | dpage = alloc_page(GFP_USER | __GFP_NOWARN); | |
962 | if (dpage) { | |
963 | lock_page(dpage); | |
964 | dst_pfn = migrate_pfn(page_to_pfn(dpage)); | |
965 | } | |
966 | ||
e778406b | 967 | migrate_device_pages(&src_pfn, &dst_pfn, 1); |
b05a79d4 AP |
968 | if (src_pfn & MIGRATE_PFN_MIGRATE) |
969 | copy_highpage(dpage, page); | |
241f6885 | 970 | migrate_device_finalize(&src_pfn, &dst_pfn, 1); |
b05a79d4 AP |
971 | |
972 | if (src_pfn & MIGRATE_PFN_MIGRATE) | |
973 | return 0; | |
974 | return -EBUSY; | |
975 | } |