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); | |
76cbbead CH |
86 | } else { |
87 | int ret; | |
88 | ||
89 | get_page(page); | |
90 | spin_unlock(ptl); | |
91 | if (unlikely(!trylock_page(page))) | |
92 | return migrate_vma_collect_skip(start, end, | |
93 | walk); | |
94 | ret = split_huge_page(page); | |
95 | unlock_page(page); | |
96 | put_page(page); | |
97 | if (ret) | |
98 | return migrate_vma_collect_skip(start, end, | |
99 | walk); | |
76cbbead CH |
100 | } |
101 | } | |
102 | ||
76cbbead | 103 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); |
4b56069c HD |
104 | if (!ptep) |
105 | goto again; | |
76cbbead CH |
106 | arch_enter_lazy_mmu_mode(); |
107 | ||
108 | for (; addr < end; addr += PAGE_SIZE, ptep++) { | |
109 | unsigned long mpfn = 0, pfn; | |
5b205c7f | 110 | struct folio *folio; |
76cbbead CH |
111 | struct page *page; |
112 | swp_entry_t entry; | |
113 | pte_t pte; | |
114 | ||
c33c7948 | 115 | pte = ptep_get(ptep); |
76cbbead CH |
116 | |
117 | if (pte_none(pte)) { | |
118 | if (vma_is_anonymous(vma)) { | |
119 | mpfn = MIGRATE_PFN_MIGRATE; | |
120 | migrate->cpages++; | |
121 | } | |
122 | goto next; | |
123 | } | |
124 | ||
125 | if (!pte_present(pte)) { | |
126 | /* | |
127 | * Only care about unaddressable device page special | |
128 | * page table entry. Other special swap entries are not | |
129 | * migratable, and we ignore regular swapped page. | |
130 | */ | |
131 | entry = pte_to_swp_entry(pte); | |
132 | if (!is_device_private_entry(entry)) | |
133 | goto next; | |
134 | ||
135 | page = pfn_swap_entry_to_page(entry); | |
136 | if (!(migrate->flags & | |
137 | MIGRATE_VMA_SELECT_DEVICE_PRIVATE) || | |
138 | page->pgmap->owner != migrate->pgmap_owner) | |
139 | goto next; | |
140 | ||
141 | mpfn = migrate_pfn(page_to_pfn(page)) | | |
142 | MIGRATE_PFN_MIGRATE; | |
143 | if (is_writable_device_private_entry(entry)) | |
144 | mpfn |= MIGRATE_PFN_WRITE; | |
145 | } else { | |
76cbbead | 146 | pfn = pte_pfn(pte); |
dd19e6d8 AS |
147 | if (is_zero_pfn(pfn) && |
148 | (migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) { | |
76cbbead CH |
149 | mpfn = MIGRATE_PFN_MIGRATE; |
150 | migrate->cpages++; | |
151 | goto next; | |
152 | } | |
153 | page = vm_normal_page(migrate->vma, addr, pte); | |
dd19e6d8 AS |
154 | if (page && !is_zone_device_page(page) && |
155 | !(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) | |
156 | goto next; | |
157 | else if (page && is_device_coherent_page(page) && | |
158 | (!(migrate->flags & MIGRATE_VMA_SELECT_DEVICE_COHERENT) || | |
159 | page->pgmap->owner != migrate->pgmap_owner)) | |
160 | goto next; | |
76cbbead CH |
161 | mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; |
162 | mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; | |
163 | } | |
164 | ||
165 | /* FIXME support THP */ | |
166 | if (!page || !page->mapping || PageTransCompound(page)) { | |
167 | mpfn = 0; | |
168 | goto next; | |
169 | } | |
170 | ||
171 | /* | |
5b205c7f | 172 | * By getting a reference on the folio we pin it and that blocks |
76cbbead CH |
173 | * any kind of migration. Side effect is that it "freezes" the |
174 | * pte. | |
175 | * | |
5b205c7f DH |
176 | * We drop this reference after isolating the folio from the lru |
177 | * for non device folio (device folio are not on the lru and thus | |
76cbbead CH |
178 | * can't be dropped from it). |
179 | */ | |
5b205c7f DH |
180 | folio = page_folio(page); |
181 | folio_get(folio); | |
76cbbead CH |
182 | |
183 | /* | |
5b205c7f | 184 | * We rely on folio_trylock() to avoid deadlock between |
0742e490 | 185 | * concurrent migrations where each is waiting on the others |
5b205c7f | 186 | * folio lock. If we can't immediately lock the folio we fail this |
0742e490 AP |
187 | * migration as it is only best effort anyway. |
188 | * | |
5b205c7f DH |
189 | * If we can lock the folio it's safe to set up a migration entry |
190 | * now. In the common case where the folio is mapped once in a | |
0742e490 AP |
191 | * single process setting up the migration entry now is an |
192 | * optimisation to avoid walking the rmap later with | |
193 | * try_to_migrate(). | |
76cbbead | 194 | */ |
5b205c7f | 195 | if (folio_trylock(folio)) { |
6c287605 | 196 | bool anon_exclusive; |
76cbbead CH |
197 | pte_t swp_pte; |
198 | ||
c33c7948 | 199 | flush_cache_page(vma, addr, pte_pfn(pte)); |
5b205c7f DH |
200 | anon_exclusive = folio_test_anon(folio) && |
201 | PageAnonExclusive(page); | |
6c287605 | 202 | if (anon_exclusive) { |
fd35ca3d | 203 | pte = ptep_clear_flush(vma, addr, ptep); |
6c287605 | 204 | |
e3b4b137 | 205 | if (folio_try_share_anon_rmap_pte(folio, page)) { |
6c287605 | 206 | set_pte_at(mm, addr, ptep, pte); |
5b205c7f DH |
207 | folio_unlock(folio); |
208 | folio_put(folio); | |
6c287605 DH |
209 | mpfn = 0; |
210 | goto next; | |
211 | } | |
212 | } else { | |
fd35ca3d | 213 | pte = ptep_get_and_clear(mm, addr, ptep); |
6c287605 DH |
214 | } |
215 | ||
76cbbead | 216 | migrate->cpages++; |
76cbbead | 217 | |
fd35ca3d AP |
218 | /* Set the dirty flag on the folio now the pte is gone. */ |
219 | if (pte_dirty(pte)) | |
5b205c7f | 220 | folio_mark_dirty(folio); |
fd35ca3d | 221 | |
76cbbead CH |
222 | /* Setup special migration page table entry */ |
223 | if (mpfn & MIGRATE_PFN_WRITE) | |
224 | entry = make_writable_migration_entry( | |
225 | page_to_pfn(page)); | |
6c287605 DH |
226 | else if (anon_exclusive) |
227 | entry = make_readable_exclusive_migration_entry( | |
228 | page_to_pfn(page)); | |
76cbbead CH |
229 | else |
230 | entry = make_readable_migration_entry( | |
231 | page_to_pfn(page)); | |
2e346877 PX |
232 | if (pte_present(pte)) { |
233 | if (pte_young(pte)) | |
234 | entry = make_migration_entry_young(entry); | |
235 | if (pte_dirty(pte)) | |
236 | entry = make_migration_entry_dirty(entry); | |
237 | } | |
76cbbead CH |
238 | swp_pte = swp_entry_to_pte(entry); |
239 | if (pte_present(pte)) { | |
240 | if (pte_soft_dirty(pte)) | |
241 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
242 | if (pte_uffd_wp(pte)) | |
243 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
244 | } else { | |
245 | if (pte_swp_soft_dirty(pte)) | |
246 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
247 | if (pte_swp_uffd_wp(pte)) | |
248 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
249 | } | |
250 | set_pte_at(mm, addr, ptep, swp_pte); | |
251 | ||
252 | /* | |
253 | * This is like regular unmap: we remove the rmap and | |
5b205c7f DH |
254 | * drop the folio refcount. The folio won't be freed, as |
255 | * we took a reference just above. | |
76cbbead | 256 | */ |
5b205c7f DH |
257 | folio_remove_rmap_pte(folio, page, vma); |
258 | folio_put(folio); | |
76cbbead CH |
259 | |
260 | if (pte_present(pte)) | |
261 | unmapped++; | |
262 | } else { | |
5b205c7f | 263 | folio_put(folio); |
76cbbead CH |
264 | mpfn = 0; |
265 | } | |
266 | ||
267 | next: | |
268 | migrate->dst[migrate->npages] = 0; | |
269 | migrate->src[migrate->npages++] = mpfn; | |
270 | } | |
76cbbead CH |
271 | |
272 | /* Only flush the TLB if we actually modified any entries */ | |
273 | if (unmapped) | |
274 | flush_tlb_range(walk->vma, start, end); | |
275 | ||
60bae737 AP |
276 | arch_leave_lazy_mmu_mode(); |
277 | pte_unmap_unlock(ptep - 1, ptl); | |
278 | ||
76cbbead CH |
279 | return 0; |
280 | } | |
281 | ||
282 | static const struct mm_walk_ops migrate_vma_walk_ops = { | |
283 | .pmd_entry = migrate_vma_collect_pmd, | |
284 | .pte_hole = migrate_vma_collect_hole, | |
49b06385 | 285 | .walk_lock = PGWALK_RDLOCK, |
76cbbead CH |
286 | }; |
287 | ||
288 | /* | |
289 | * migrate_vma_collect() - collect pages over a range of virtual addresses | |
290 | * @migrate: migrate struct containing all migration information | |
291 | * | |
292 | * This will walk the CPU page table. For each virtual address backed by a | |
293 | * valid page, it updates the src array and takes a reference on the page, in | |
294 | * order to pin the page until we lock it and unmap it. | |
295 | */ | |
296 | static void migrate_vma_collect(struct migrate_vma *migrate) | |
297 | { | |
298 | struct mmu_notifier_range range; | |
299 | ||
300 | /* | |
301 | * Note that the pgmap_owner is passed to the mmu notifier callback so | |
302 | * that the registered device driver can skip invalidating device | |
303 | * private page mappings that won't be migrated. | |
304 | */ | |
305 | mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0, | |
7d4a8be0 | 306 | migrate->vma->vm_mm, migrate->start, migrate->end, |
76cbbead CH |
307 | migrate->pgmap_owner); |
308 | mmu_notifier_invalidate_range_start(&range); | |
309 | ||
310 | walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end, | |
311 | &migrate_vma_walk_ops, migrate); | |
312 | ||
313 | mmu_notifier_invalidate_range_end(&range); | |
314 | migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); | |
315 | } | |
316 | ||
317 | /* | |
318 | * migrate_vma_check_page() - check if page is pinned or not | |
319 | * @page: struct page to check | |
320 | * | |
321 | * Pinned pages cannot be migrated. This is the same test as in | |
322 | * folio_migrate_mapping(), except that here we allow migration of a | |
323 | * ZONE_DEVICE page. | |
324 | */ | |
16ce101d | 325 | static bool migrate_vma_check_page(struct page *page, struct page *fault_page) |
76cbbead CH |
326 | { |
327 | /* | |
328 | * One extra ref because caller holds an extra reference, either from | |
329 | * isolate_lru_page() for a regular page, or migrate_vma_collect() for | |
330 | * a device page. | |
331 | */ | |
16ce101d | 332 | int extra = 1 + (page == fault_page); |
76cbbead CH |
333 | |
334 | /* | |
335 | * FIXME support THP (transparent huge page), it is bit more complex to | |
336 | * check them than regular pages, because they can be mapped with a pmd | |
337 | * or with a pte (split pte mapping). | |
338 | */ | |
339 | if (PageCompound(page)) | |
340 | return false; | |
341 | ||
342 | /* Page from ZONE_DEVICE have one extra reference */ | |
343 | if (is_zone_device_page(page)) | |
344 | extra++; | |
345 | ||
346 | /* For file back page */ | |
347 | if (page_mapping(page)) | |
348 | extra += 1 + page_has_private(page); | |
349 | ||
350 | if ((page_count(page) - extra) > page_mapcount(page)) | |
351 | return false; | |
352 | ||
353 | return true; | |
354 | } | |
355 | ||
356 | /* | |
44af0b45 AP |
357 | * Unmaps pages for migration. Returns number of source pfns marked as |
358 | * migrating. | |
76cbbead | 359 | */ |
241f6885 AP |
360 | static unsigned long migrate_device_unmap(unsigned long *src_pfns, |
361 | unsigned long npages, | |
362 | struct page *fault_page) | |
76cbbead | 363 | { |
76cbbead CH |
364 | unsigned long i, restore = 0; |
365 | bool allow_drain = true; | |
241f6885 | 366 | unsigned long unmapped = 0; |
76cbbead CH |
367 | |
368 | lru_add_drain(); | |
369 | ||
370 | for (i = 0; i < npages; i++) { | |
241f6885 | 371 | struct page *page = migrate_pfn_to_page(src_pfns[i]); |
4b8554c5 | 372 | struct folio *folio; |
76cbbead | 373 | |
44af0b45 AP |
374 | if (!page) { |
375 | if (src_pfns[i] & MIGRATE_PFN_MIGRATE) | |
376 | unmapped++; | |
76cbbead | 377 | continue; |
44af0b45 | 378 | } |
76cbbead CH |
379 | |
380 | /* ZONE_DEVICE pages are not on LRU */ | |
381 | if (!is_zone_device_page(page)) { | |
382 | if (!PageLRU(page) && allow_drain) { | |
1fec6890 | 383 | /* Drain CPU's lru cache */ |
76cbbead CH |
384 | lru_add_drain_all(); |
385 | allow_drain = false; | |
386 | } | |
387 | ||
f7f9c00d | 388 | if (!isolate_lru_page(page)) { |
241f6885 | 389 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
390 | restore++; |
391 | continue; | |
392 | } | |
393 | ||
394 | /* Drop the reference we took in collect */ | |
395 | put_page(page); | |
396 | } | |
397 | ||
4b8554c5 MWO |
398 | folio = page_folio(page); |
399 | if (folio_mapped(folio)) | |
400 | try_to_migrate(folio, 0); | |
76cbbead | 401 | |
16ce101d | 402 | if (page_mapped(page) || |
241f6885 | 403 | !migrate_vma_check_page(page, fault_page)) { |
76cbbead CH |
404 | if (!is_zone_device_page(page)) { |
405 | get_page(page); | |
406 | putback_lru_page(page); | |
407 | } | |
408 | ||
241f6885 | 409 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
410 | restore++; |
411 | continue; | |
412 | } | |
241f6885 AP |
413 | |
414 | unmapped++; | |
76cbbead CH |
415 | } |
416 | ||
417 | for (i = 0; i < npages && restore; i++) { | |
241f6885 | 418 | struct page *page = migrate_pfn_to_page(src_pfns[i]); |
4eecb8b9 | 419 | struct folio *folio; |
76cbbead | 420 | |
241f6885 | 421 | if (!page || (src_pfns[i] & MIGRATE_PFN_MIGRATE)) |
76cbbead CH |
422 | continue; |
423 | ||
4eecb8b9 MWO |
424 | folio = page_folio(page); |
425 | remove_migration_ptes(folio, folio, false); | |
76cbbead | 426 | |
241f6885 | 427 | src_pfns[i] = 0; |
4eecb8b9 MWO |
428 | folio_unlock(folio); |
429 | folio_put(folio); | |
76cbbead CH |
430 | restore--; |
431 | } | |
241f6885 AP |
432 | |
433 | return unmapped; | |
434 | } | |
435 | ||
436 | /* | |
437 | * migrate_vma_unmap() - replace page mapping with special migration pte entry | |
438 | * @migrate: migrate struct containing all migration information | |
439 | * | |
440 | * Isolate pages from the LRU and replace mappings (CPU page table pte) with a | |
441 | * special migration pte entry and check if it has been pinned. Pinned pages are | |
442 | * restored because we cannot migrate them. | |
443 | * | |
444 | * This is the last step before we call the device driver callback to allocate | |
445 | * destination memory and copy contents of original page over to new page. | |
446 | */ | |
447 | static void migrate_vma_unmap(struct migrate_vma *migrate) | |
448 | { | |
449 | migrate->cpages = migrate_device_unmap(migrate->src, migrate->npages, | |
450 | migrate->fault_page); | |
76cbbead CH |
451 | } |
452 | ||
453 | /** | |
454 | * migrate_vma_setup() - prepare to migrate a range of memory | |
455 | * @args: contains the vma, start, and pfns arrays for the migration | |
456 | * | |
457 | * Returns: negative errno on failures, 0 when 0 or more pages were migrated | |
458 | * without an error. | |
459 | * | |
460 | * Prepare to migrate a range of memory virtual address range by collecting all | |
461 | * the pages backing each virtual address in the range, saving them inside the | |
462 | * src array. Then lock those pages and unmap them. Once the pages are locked | |
463 | * and unmapped, check whether each page is pinned or not. Pages that aren't | |
464 | * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the | |
465 | * corresponding src array entry. Then restores any pages that are pinned, by | |
466 | * remapping and unlocking those pages. | |
467 | * | |
468 | * The caller should then allocate destination memory and copy source memory to | |
469 | * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE | |
470 | * flag set). Once these are allocated and copied, the caller must update each | |
471 | * corresponding entry in the dst array with the pfn value of the destination | |
472 | * page and with MIGRATE_PFN_VALID. Destination pages must be locked via | |
473 | * lock_page(). | |
474 | * | |
475 | * Note that the caller does not have to migrate all the pages that are marked | |
476 | * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from | |
477 | * device memory to system memory. If the caller cannot migrate a device page | |
478 | * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe | |
479 | * consequences for the userspace process, so it must be avoided if at all | |
480 | * possible. | |
481 | * | |
482 | * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we | |
483 | * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus | |
484 | * allowing the caller to allocate device memory for those unbacked virtual | |
485 | * addresses. For this the caller simply has to allocate device memory and | |
486 | * properly set the destination entry like for regular migration. Note that | |
487 | * this can still fail, and thus inside the device driver you must check if the | |
488 | * migration was successful for those entries after calling migrate_vma_pages(), | |
489 | * just like for regular migration. | |
490 | * | |
491 | * After that, the callers must call migrate_vma_pages() to go over each entry | |
492 | * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag | |
493 | * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, | |
494 | * then migrate_vma_pages() to migrate struct page information from the source | |
495 | * struct page to the destination struct page. If it fails to migrate the | |
496 | * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the | |
497 | * src array. | |
498 | * | |
499 | * At this point all successfully migrated pages have an entry in the src | |
500 | * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst | |
501 | * array entry with MIGRATE_PFN_VALID flag set. | |
502 | * | |
503 | * Once migrate_vma_pages() returns the caller may inspect which pages were | |
504 | * successfully migrated, and which were not. Successfully migrated pages will | |
505 | * have the MIGRATE_PFN_MIGRATE flag set for their src array entry. | |
506 | * | |
507 | * It is safe to update device page table after migrate_vma_pages() because | |
508 | * both destination and source page are still locked, and the mmap_lock is held | |
509 | * in read mode (hence no one can unmap the range being migrated). | |
510 | * | |
511 | * Once the caller is done cleaning up things and updating its page table (if it | |
512 | * chose to do so, this is not an obligation) it finally calls | |
513 | * migrate_vma_finalize() to update the CPU page table to point to new pages | |
514 | * for successfully migrated pages or otherwise restore the CPU page table to | |
515 | * point to the original source pages. | |
516 | */ | |
517 | int migrate_vma_setup(struct migrate_vma *args) | |
518 | { | |
519 | long nr_pages = (args->end - args->start) >> PAGE_SHIFT; | |
520 | ||
521 | args->start &= PAGE_MASK; | |
522 | args->end &= PAGE_MASK; | |
523 | if (!args->vma || is_vm_hugetlb_page(args->vma) || | |
524 | (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma)) | |
525 | return -EINVAL; | |
526 | if (nr_pages <= 0) | |
527 | return -EINVAL; | |
528 | if (args->start < args->vma->vm_start || | |
529 | args->start >= args->vma->vm_end) | |
530 | return -EINVAL; | |
531 | if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end) | |
532 | return -EINVAL; | |
533 | if (!args->src || !args->dst) | |
534 | return -EINVAL; | |
16ce101d AP |
535 | if (args->fault_page && !is_device_private_page(args->fault_page)) |
536 | return -EINVAL; | |
76cbbead CH |
537 | |
538 | memset(args->src, 0, sizeof(*args->src) * nr_pages); | |
539 | args->cpages = 0; | |
540 | args->npages = 0; | |
541 | ||
542 | migrate_vma_collect(args); | |
543 | ||
544 | if (args->cpages) | |
545 | migrate_vma_unmap(args); | |
546 | ||
547 | /* | |
548 | * At this point pages are locked and unmapped, and thus they have | |
549 | * stable content and can safely be copied to destination memory that | |
550 | * is allocated by the drivers. | |
551 | */ | |
552 | return 0; | |
553 | ||
554 | } | |
555 | EXPORT_SYMBOL(migrate_vma_setup); | |
556 | ||
557 | /* | |
558 | * This code closely matches the code in: | |
559 | * __handle_mm_fault() | |
560 | * handle_pte_fault() | |
561 | * do_anonymous_page() | |
562 | * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE | |
f25cbb7a | 563 | * private or coherent page. |
76cbbead CH |
564 | */ |
565 | static void migrate_vma_insert_page(struct migrate_vma *migrate, | |
566 | unsigned long addr, | |
567 | struct page *page, | |
568 | unsigned long *src) | |
569 | { | |
d3b08273 | 570 | struct folio *folio = page_folio(page); |
76cbbead CH |
571 | struct vm_area_struct *vma = migrate->vma; |
572 | struct mm_struct *mm = vma->vm_mm; | |
573 | bool flush = false; | |
574 | spinlock_t *ptl; | |
575 | pte_t entry; | |
576 | pgd_t *pgdp; | |
577 | p4d_t *p4dp; | |
578 | pud_t *pudp; | |
579 | pmd_t *pmdp; | |
580 | pte_t *ptep; | |
c33c7948 | 581 | pte_t orig_pte; |
76cbbead CH |
582 | |
583 | /* Only allow populating anonymous memory */ | |
584 | if (!vma_is_anonymous(vma)) | |
585 | goto abort; | |
586 | ||
587 | pgdp = pgd_offset(mm, addr); | |
588 | p4dp = p4d_alloc(mm, pgdp, addr); | |
589 | if (!p4dp) | |
590 | goto abort; | |
591 | pudp = pud_alloc(mm, p4dp, addr); | |
592 | if (!pudp) | |
593 | goto abort; | |
594 | pmdp = pmd_alloc(mm, pudp, addr); | |
595 | if (!pmdp) | |
596 | goto abort; | |
76cbbead CH |
597 | if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) |
598 | goto abort; | |
76cbbead CH |
599 | if (pte_alloc(mm, pmdp)) |
600 | goto abort; | |
76cbbead CH |
601 | if (unlikely(anon_vma_prepare(vma))) |
602 | goto abort; | |
d3b08273 | 603 | if (mem_cgroup_charge(folio, vma->vm_mm, GFP_KERNEL)) |
76cbbead CH |
604 | goto abort; |
605 | ||
606 | /* | |
d3b08273 MWO |
607 | * The memory barrier inside __folio_mark_uptodate makes sure that |
608 | * preceding stores to the folio contents become visible before | |
76cbbead CH |
609 | * the set_pte_at() write. |
610 | */ | |
d3b08273 | 611 | __folio_mark_uptodate(folio); |
76cbbead | 612 | |
d3b08273 | 613 | if (folio_is_device_private(folio)) { |
76cbbead CH |
614 | swp_entry_t swp_entry; |
615 | ||
616 | if (vma->vm_flags & VM_WRITE) | |
617 | swp_entry = make_writable_device_private_entry( | |
618 | page_to_pfn(page)); | |
619 | else | |
620 | swp_entry = make_readable_device_private_entry( | |
621 | page_to_pfn(page)); | |
622 | entry = swp_entry_to_pte(swp_entry); | |
623 | } else { | |
d3b08273 MWO |
624 | if (folio_is_zone_device(folio) && |
625 | !folio_is_device_coherent(folio)) { | |
76cbbead CH |
626 | pr_warn_once("Unsupported ZONE_DEVICE page type.\n"); |
627 | goto abort; | |
628 | } | |
629 | entry = mk_pte(page, vma->vm_page_prot); | |
630 | if (vma->vm_flags & VM_WRITE) | |
161e393c | 631 | entry = pte_mkwrite(pte_mkdirty(entry), vma); |
76cbbead CH |
632 | } |
633 | ||
634 | ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); | |
4b56069c HD |
635 | if (!ptep) |
636 | goto abort; | |
c33c7948 RR |
637 | orig_pte = ptep_get(ptep); |
638 | ||
76cbbead CH |
639 | if (check_stable_address_space(mm)) |
640 | goto unlock_abort; | |
641 | ||
c33c7948 RR |
642 | if (pte_present(orig_pte)) { |
643 | unsigned long pfn = pte_pfn(orig_pte); | |
76cbbead CH |
644 | |
645 | if (!is_zero_pfn(pfn)) | |
646 | goto unlock_abort; | |
647 | flush = true; | |
c33c7948 | 648 | } else if (!pte_none(orig_pte)) |
76cbbead CH |
649 | goto unlock_abort; |
650 | ||
651 | /* | |
652 | * Check for userfaultfd but do not deliver the fault. Instead, | |
653 | * just back off. | |
654 | */ | |
655 | if (userfaultfd_missing(vma)) | |
656 | goto unlock_abort; | |
657 | ||
658 | inc_mm_counter(mm, MM_ANONPAGES); | |
d3b08273 MWO |
659 | folio_add_new_anon_rmap(folio, vma, addr); |
660 | if (!folio_is_zone_device(folio)) | |
661 | folio_add_lru_vma(folio, vma); | |
662 | folio_get(folio); | |
76cbbead CH |
663 | |
664 | if (flush) { | |
c33c7948 | 665 | flush_cache_page(vma, addr, pte_pfn(orig_pte)); |
ec8832d0 | 666 | ptep_clear_flush(vma, addr, ptep); |
76cbbead CH |
667 | set_pte_at_notify(mm, addr, ptep, entry); |
668 | update_mmu_cache(vma, addr, ptep); | |
669 | } else { | |
670 | /* No need to invalidate - it was non-present before */ | |
671 | set_pte_at(mm, addr, ptep, entry); | |
672 | update_mmu_cache(vma, addr, ptep); | |
673 | } | |
674 | ||
675 | pte_unmap_unlock(ptep, ptl); | |
676 | *src = MIGRATE_PFN_MIGRATE; | |
677 | return; | |
678 | ||
679 | unlock_abort: | |
680 | pte_unmap_unlock(ptep, ptl); | |
681 | abort: | |
682 | *src &= ~MIGRATE_PFN_MIGRATE; | |
683 | } | |
684 | ||
e778406b | 685 | static void __migrate_device_pages(unsigned long *src_pfns, |
241f6885 AP |
686 | unsigned long *dst_pfns, unsigned long npages, |
687 | struct migrate_vma *migrate) | |
76cbbead | 688 | { |
76cbbead | 689 | struct mmu_notifier_range range; |
241f6885 | 690 | unsigned long i; |
76cbbead CH |
691 | bool notified = false; |
692 | ||
241f6885 AP |
693 | for (i = 0; i < npages; i++) { |
694 | struct page *newpage = migrate_pfn_to_page(dst_pfns[i]); | |
695 | struct page *page = migrate_pfn_to_page(src_pfns[i]); | |
76cbbead CH |
696 | struct address_space *mapping; |
697 | int r; | |
698 | ||
699 | if (!newpage) { | |
241f6885 | 700 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
701 | continue; |
702 | } | |
703 | ||
704 | if (!page) { | |
241f6885 AP |
705 | unsigned long addr; |
706 | ||
e778406b AP |
707 | if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE)) |
708 | continue; | |
709 | ||
b05a79d4 AP |
710 | /* |
711 | * The only time there is no vma is when called from | |
712 | * migrate_device_coherent_page(). However this isn't | |
713 | * called if the page could not be unmapped. | |
714 | */ | |
241f6885 AP |
715 | VM_BUG_ON(!migrate); |
716 | addr = migrate->start + i*PAGE_SIZE; | |
76cbbead CH |
717 | if (!notified) { |
718 | notified = true; | |
719 | ||
720 | mmu_notifier_range_init_owner(&range, | |
7d4a8be0 | 721 | MMU_NOTIFY_MIGRATE, 0, |
76cbbead CH |
722 | migrate->vma->vm_mm, addr, migrate->end, |
723 | migrate->pgmap_owner); | |
724 | mmu_notifier_invalidate_range_start(&range); | |
725 | } | |
726 | migrate_vma_insert_page(migrate, addr, newpage, | |
241f6885 | 727 | &src_pfns[i]); |
76cbbead CH |
728 | continue; |
729 | } | |
730 | ||
731 | mapping = page_mapping(page); | |
732 | ||
f25cbb7a AS |
733 | if (is_device_private_page(newpage) || |
734 | is_device_coherent_page(newpage)) { | |
76cbbead | 735 | if (mapping) { |
df263d9a MP |
736 | struct folio *folio; |
737 | ||
738 | folio = page_folio(page); | |
739 | ||
740 | /* | |
741 | * For now only support anonymous memory migrating to | |
742 | * device private or coherent memory. | |
743 | * | |
744 | * Try to get rid of swap cache if possible. | |
745 | */ | |
746 | if (!folio_test_anon(folio) || | |
747 | !folio_free_swap(folio)) { | |
748 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; | |
749 | continue; | |
750 | } | |
76cbbead CH |
751 | } |
752 | } else if (is_zone_device_page(newpage)) { | |
753 | /* | |
754 | * Other types of ZONE_DEVICE page are not supported. | |
755 | */ | |
241f6885 | 756 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
757 | continue; |
758 | } | |
759 | ||
241f6885 | 760 | if (migrate && migrate->fault_page == page) |
16ce101d AP |
761 | r = migrate_folio_extra(mapping, page_folio(newpage), |
762 | page_folio(page), | |
763 | MIGRATE_SYNC_NO_COPY, 1); | |
764 | else | |
765 | r = migrate_folio(mapping, page_folio(newpage), | |
766 | page_folio(page), MIGRATE_SYNC_NO_COPY); | |
76cbbead | 767 | if (r != MIGRATEPAGE_SUCCESS) |
241f6885 | 768 | src_pfns[i] &= ~MIGRATE_PFN_MIGRATE; |
76cbbead CH |
769 | } |
770 | ||
76cbbead | 771 | if (notified) |
ec8832d0 | 772 | mmu_notifier_invalidate_range_end(&range); |
76cbbead | 773 | } |
76cbbead | 774 | |
e778406b AP |
775 | /** |
776 | * migrate_device_pages() - migrate meta-data from src page to dst page | |
777 | * @src_pfns: src_pfns returned from migrate_device_range() | |
778 | * @dst_pfns: array of pfns allocated by the driver to migrate memory to | |
779 | * @npages: number of pages in the range | |
780 | * | |
781 | * Equivalent to migrate_vma_pages(). This is called to migrate struct page | |
782 | * meta-data from source struct page to destination. | |
783 | */ | |
784 | void migrate_device_pages(unsigned long *src_pfns, unsigned long *dst_pfns, | |
785 | unsigned long npages) | |
786 | { | |
787 | __migrate_device_pages(src_pfns, dst_pfns, npages, NULL); | |
788 | } | |
789 | EXPORT_SYMBOL(migrate_device_pages); | |
790 | ||
76cbbead | 791 | /** |
241f6885 | 792 | * migrate_vma_pages() - migrate meta-data from src page to dst page |
76cbbead CH |
793 | * @migrate: migrate struct containing all migration information |
794 | * | |
241f6885 AP |
795 | * This migrates struct page meta-data from source struct page to destination |
796 | * struct page. This effectively finishes the migration from source page to the | |
797 | * destination page. | |
76cbbead | 798 | */ |
241f6885 AP |
799 | void migrate_vma_pages(struct migrate_vma *migrate) |
800 | { | |
e778406b | 801 | __migrate_device_pages(migrate->src, migrate->dst, migrate->npages, migrate); |
241f6885 AP |
802 | } |
803 | EXPORT_SYMBOL(migrate_vma_pages); | |
804 | ||
e778406b AP |
805 | /* |
806 | * migrate_device_finalize() - complete page migration | |
807 | * @src_pfns: src_pfns returned from migrate_device_range() | |
808 | * @dst_pfns: array of pfns allocated by the driver to migrate memory to | |
809 | * @npages: number of pages in the range | |
810 | * | |
811 | * Completes migration of the page by removing special migration entries. | |
812 | * Drivers must ensure copying of page data is complete and visible to the CPU | |
813 | * before calling this. | |
814 | */ | |
815 | void migrate_device_finalize(unsigned long *src_pfns, | |
816 | unsigned long *dst_pfns, unsigned long npages) | |
76cbbead | 817 | { |
76cbbead CH |
818 | unsigned long i; |
819 | ||
820 | for (i = 0; i < npages; i++) { | |
4eecb8b9 | 821 | struct folio *dst, *src; |
241f6885 AP |
822 | struct page *newpage = migrate_pfn_to_page(dst_pfns[i]); |
823 | struct page *page = migrate_pfn_to_page(src_pfns[i]); | |
76cbbead CH |
824 | |
825 | if (!page) { | |
826 | if (newpage) { | |
827 | unlock_page(newpage); | |
828 | put_page(newpage); | |
829 | } | |
830 | continue; | |
831 | } | |
832 | ||
241f6885 | 833 | if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE) || !newpage) { |
76cbbead CH |
834 | if (newpage) { |
835 | unlock_page(newpage); | |
836 | put_page(newpage); | |
837 | } | |
838 | newpage = page; | |
839 | } | |
840 | ||
4eecb8b9 MWO |
841 | src = page_folio(page); |
842 | dst = page_folio(newpage); | |
843 | remove_migration_ptes(src, dst, false); | |
844 | folio_unlock(src); | |
76cbbead CH |
845 | |
846 | if (is_zone_device_page(page)) | |
847 | put_page(page); | |
848 | else | |
849 | putback_lru_page(page); | |
850 | ||
851 | if (newpage != page) { | |
852 | unlock_page(newpage); | |
853 | if (is_zone_device_page(newpage)) | |
854 | put_page(newpage); | |
855 | else | |
856 | putback_lru_page(newpage); | |
857 | } | |
858 | } | |
859 | } | |
e778406b | 860 | EXPORT_SYMBOL(migrate_device_finalize); |
241f6885 AP |
861 | |
862 | /** | |
863 | * migrate_vma_finalize() - restore CPU page table entry | |
864 | * @migrate: migrate struct containing all migration information | |
865 | * | |
866 | * This replaces the special migration pte entry with either a mapping to the | |
867 | * new page if migration was successful for that page, or to the original page | |
868 | * otherwise. | |
869 | * | |
870 | * This also unlocks the pages and puts them back on the lru, or drops the extra | |
871 | * refcount, for device pages. | |
872 | */ | |
873 | void migrate_vma_finalize(struct migrate_vma *migrate) | |
874 | { | |
875 | migrate_device_finalize(migrate->src, migrate->dst, migrate->npages); | |
876 | } | |
76cbbead | 877 | EXPORT_SYMBOL(migrate_vma_finalize); |
b05a79d4 | 878 | |
e778406b AP |
879 | /** |
880 | * migrate_device_range() - migrate device private pfns to normal memory. | |
881 | * @src_pfns: array large enough to hold migrating source device private pfns. | |
882 | * @start: starting pfn in the range to migrate. | |
883 | * @npages: number of pages to migrate. | |
884 | * | |
885 | * migrate_vma_setup() is similar in concept to migrate_vma_setup() except that | |
886 | * instead of looking up pages based on virtual address mappings a range of | |
887 | * device pfns that should be migrated to system memory is used instead. | |
888 | * | |
889 | * This is useful when a driver needs to free device memory but doesn't know the | |
890 | * virtual mappings of every page that may be in device memory. For example this | |
891 | * is often the case when a driver is being unloaded or unbound from a device. | |
892 | * | |
893 | * Like migrate_vma_setup() this function will take a reference and lock any | |
894 | * migrating pages that aren't free before unmapping them. Drivers may then | |
895 | * allocate destination pages and start copying data from the device to CPU | |
896 | * memory before calling migrate_device_pages(). | |
897 | */ | |
898 | int migrate_device_range(unsigned long *src_pfns, unsigned long start, | |
899 | unsigned long npages) | |
900 | { | |
901 | unsigned long i, pfn; | |
902 | ||
903 | for (pfn = start, i = 0; i < npages; pfn++, i++) { | |
904 | struct page *page = pfn_to_page(pfn); | |
905 | ||
906 | if (!get_page_unless_zero(page)) { | |
907 | src_pfns[i] = 0; | |
908 | continue; | |
909 | } | |
910 | ||
911 | if (!trylock_page(page)) { | |
912 | src_pfns[i] = 0; | |
913 | put_page(page); | |
914 | continue; | |
915 | } | |
916 | ||
917 | src_pfns[i] = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; | |
918 | } | |
919 | ||
920 | migrate_device_unmap(src_pfns, npages, NULL); | |
921 | ||
922 | return 0; | |
923 | } | |
924 | EXPORT_SYMBOL(migrate_device_range); | |
925 | ||
b05a79d4 AP |
926 | /* |
927 | * Migrate a device coherent page back to normal memory. The caller should have | |
928 | * a reference on page which will be copied to the new page if migration is | |
929 | * successful or dropped on failure. | |
930 | */ | |
931 | int migrate_device_coherent_page(struct page *page) | |
932 | { | |
933 | unsigned long src_pfn, dst_pfn = 0; | |
b05a79d4 AP |
934 | struct page *dpage; |
935 | ||
936 | WARN_ON_ONCE(PageCompound(page)); | |
937 | ||
938 | lock_page(page); | |
939 | src_pfn = migrate_pfn(page_to_pfn(page)) | MIGRATE_PFN_MIGRATE; | |
b05a79d4 AP |
940 | |
941 | /* | |
942 | * We don't have a VMA and don't need to walk the page tables to find | |
943 | * the source page. So call migrate_vma_unmap() directly to unmap the | |
944 | * page as migrate_vma_setup() will fail if args.vma == NULL. | |
945 | */ | |
241f6885 | 946 | migrate_device_unmap(&src_pfn, 1, NULL); |
b05a79d4 AP |
947 | if (!(src_pfn & MIGRATE_PFN_MIGRATE)) |
948 | return -EBUSY; | |
949 | ||
950 | dpage = alloc_page(GFP_USER | __GFP_NOWARN); | |
951 | if (dpage) { | |
952 | lock_page(dpage); | |
953 | dst_pfn = migrate_pfn(page_to_pfn(dpage)); | |
954 | } | |
955 | ||
e778406b | 956 | migrate_device_pages(&src_pfn, &dst_pfn, 1); |
b05a79d4 AP |
957 | if (src_pfn & MIGRATE_PFN_MIGRATE) |
958 | copy_highpage(dpage, page); | |
241f6885 | 959 | migrate_device_finalize(&src_pfn, &dst_pfn, 1); |
b05a79d4 AP |
960 | |
961 | if (src_pfn & MIGRATE_PFN_MIGRATE) | |
962 | return 0; | |
963 | return -EBUSY; | |
964 | } |