1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Memory Migration functionality - linux/mm/migrate.c 4 * 5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter 6 * 7 * Page migration was first developed in the context of the memory hotplug 8 * project. The main authors of the migration code are: 9 * 10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> 11 * Hirokazu Takahashi <taka@valinux.co.jp> 12 * Dave Hansen <haveblue@us.ibm.com> 13 * Christoph Lameter 14 */ 15 16 #include <linux/migrate.h> 17 #include <linux/export.h> 18 #include <linux/swap.h> 19 #include <linux/swapops.h> 20 #include <linux/pagemap.h> 21 #include <linux/buffer_head.h> 22 #include <linux/mm_inline.h> 23 #include <linux/nsproxy.h> 24 #include <linux/pagevec.h> 25 #include <linux/ksm.h> 26 #include <linux/rmap.h> 27 #include <linux/topology.h> 28 #include <linux/cpu.h> 29 #include <linux/cpuset.h> 30 #include <linux/writeback.h> 31 #include <linux/mempolicy.h> 32 #include <linux/vmalloc.h> 33 #include <linux/security.h> 34 #include <linux/backing-dev.h> 35 #include <linux/compaction.h> 36 #include <linux/syscalls.h> 37 #include <linux/compat.h> 38 #include <linux/hugetlb.h> 39 #include <linux/hugetlb_cgroup.h> 40 #include <linux/gfp.h> 41 #include <linux/pfn_t.h> 42 #include <linux/memremap.h> 43 #include <linux/userfaultfd_k.h> 44 #include <linux/balloon_compaction.h> 45 #include <linux/page_idle.h> 46 #include <linux/page_owner.h> 47 #include <linux/sched/mm.h> 48 #include <linux/ptrace.h> 49 #include <linux/oom.h> 50 #include <linux/memory.h> 51 #include <linux/random.h> 52 #include <linux/sched/sysctl.h> 53 #include <linux/memory-tiers.h> 54 55 #include <asm/tlbflush.h> 56 57 #include <trace/events/migrate.h> 58 59 #include "internal.h" 60 61 bool isolate_movable_page(struct page *page, isolate_mode_t mode) 62 { 63 struct folio *folio = folio_get_nontail_page(page); 64 const struct movable_operations *mops; 65 66 /* 67 * Avoid burning cycles with pages that are yet under __free_pages(), 68 * or just got freed under us. 69 * 70 * In case we 'win' a race for a movable page being freed under us and 71 * raise its refcount preventing __free_pages() from doing its job 72 * the put_page() at the end of this block will take care of 73 * release this page, thus avoiding a nasty leakage. 74 */ 75 if (!folio) 76 goto out; 77 78 if (unlikely(folio_test_slab(folio))) 79 goto out_putfolio; 80 /* Pairs with smp_wmb() in slab freeing, e.g. SLUB's __free_slab() */ 81 smp_rmb(); 82 /* 83 * Check movable flag before taking the page lock because 84 * we use non-atomic bitops on newly allocated page flags so 85 * unconditionally grabbing the lock ruins page's owner side. 86 */ 87 if (unlikely(!__folio_test_movable(folio))) 88 goto out_putfolio; 89 /* Pairs with smp_wmb() in slab allocation, e.g. SLUB's alloc_slab_page() */ 90 smp_rmb(); 91 if (unlikely(folio_test_slab(folio))) 92 goto out_putfolio; 93 94 /* 95 * As movable pages are not isolated from LRU lists, concurrent 96 * compaction threads can race against page migration functions 97 * as well as race against the releasing a page. 98 * 99 * In order to avoid having an already isolated movable page 100 * being (wrongly) re-isolated while it is under migration, 101 * or to avoid attempting to isolate pages being released, 102 * lets be sure we have the page lock 103 * before proceeding with the movable page isolation steps. 104 */ 105 if (unlikely(!folio_trylock(folio))) 106 goto out_putfolio; 107 108 if (!folio_test_movable(folio) || folio_test_isolated(folio)) 109 goto out_no_isolated; 110 111 mops = folio_movable_ops(folio); 112 VM_BUG_ON_FOLIO(!mops, folio); 113 114 if (!mops->isolate_page(&folio->page, mode)) 115 goto out_no_isolated; 116 117 /* Driver shouldn't use PG_isolated bit of page->flags */ 118 WARN_ON_ONCE(folio_test_isolated(folio)); 119 folio_set_isolated(folio); 120 folio_unlock(folio); 121 122 return true; 123 124 out_no_isolated: 125 folio_unlock(folio); 126 out_putfolio: 127 folio_put(folio); 128 out: 129 return false; 130 } 131 132 static void putback_movable_folio(struct folio *folio) 133 { 134 const struct movable_operations *mops = folio_movable_ops(folio); 135 136 mops->putback_page(&folio->page); 137 folio_clear_isolated(folio); 138 } 139 140 /* 141 * Put previously isolated pages back onto the appropriate lists 142 * from where they were once taken off for compaction/migration. 143 * 144 * This function shall be used whenever the isolated pageset has been 145 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() 146 * and isolate_hugetlb(). 147 */ 148 void putback_movable_pages(struct list_head *l) 149 { 150 struct folio *folio; 151 struct folio *folio2; 152 153 list_for_each_entry_safe(folio, folio2, l, lru) { 154 if (unlikely(folio_test_hugetlb(folio))) { 155 folio_putback_active_hugetlb(folio); 156 continue; 157 } 158 list_del(&folio->lru); 159 /* 160 * We isolated non-lru movable folio so here we can use 161 * __PageMovable because LRU folio's mapping cannot have 162 * PAGE_MAPPING_MOVABLE. 163 */ 164 if (unlikely(__folio_test_movable(folio))) { 165 VM_BUG_ON_FOLIO(!folio_test_isolated(folio), folio); 166 folio_lock(folio); 167 if (folio_test_movable(folio)) 168 putback_movable_folio(folio); 169 else 170 folio_clear_isolated(folio); 171 folio_unlock(folio); 172 folio_put(folio); 173 } else { 174 node_stat_mod_folio(folio, NR_ISOLATED_ANON + 175 folio_is_file_lru(folio), -folio_nr_pages(folio)); 176 folio_putback_lru(folio); 177 } 178 } 179 } 180 181 /* 182 * Restore a potential migration pte to a working pte entry 183 */ 184 static bool remove_migration_pte(struct folio *folio, 185 struct vm_area_struct *vma, unsigned long addr, void *old) 186 { 187 DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION); 188 189 while (page_vma_mapped_walk(&pvmw)) { 190 rmap_t rmap_flags = RMAP_NONE; 191 pte_t pte; 192 swp_entry_t entry; 193 struct page *new; 194 unsigned long idx = 0; 195 196 /* pgoff is invalid for ksm pages, but they are never large */ 197 if (folio_test_large(folio) && !folio_test_hugetlb(folio)) 198 idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff; 199 new = folio_page(folio, idx); 200 201 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 202 /* PMD-mapped THP migration entry */ 203 if (!pvmw.pte) { 204 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) || 205 !folio_test_pmd_mappable(folio), folio); 206 remove_migration_pmd(&pvmw, new); 207 continue; 208 } 209 #endif 210 211 folio_get(folio); 212 pte = mk_pte(new, READ_ONCE(vma->vm_page_prot)); 213 if (pte_swp_soft_dirty(*pvmw.pte)) 214 pte = pte_mksoft_dirty(pte); 215 216 entry = pte_to_swp_entry(*pvmw.pte); 217 if (!is_migration_entry_young(entry)) 218 pte = pte_mkold(pte); 219 if (folio_test_dirty(folio) && is_migration_entry_dirty(entry)) 220 pte = pte_mkdirty(pte); 221 if (is_writable_migration_entry(entry)) 222 pte = pte_mkwrite(pte); 223 else if (pte_swp_uffd_wp(*pvmw.pte)) 224 pte = pte_mkuffd_wp(pte); 225 226 if (folio_test_anon(folio) && !is_readable_migration_entry(entry)) 227 rmap_flags |= RMAP_EXCLUSIVE; 228 229 if (unlikely(is_device_private_page(new))) { 230 if (pte_write(pte)) 231 entry = make_writable_device_private_entry( 232 page_to_pfn(new)); 233 else 234 entry = make_readable_device_private_entry( 235 page_to_pfn(new)); 236 pte = swp_entry_to_pte(entry); 237 if (pte_swp_soft_dirty(*pvmw.pte)) 238 pte = pte_swp_mksoft_dirty(pte); 239 if (pte_swp_uffd_wp(*pvmw.pte)) 240 pte = pte_swp_mkuffd_wp(pte); 241 } 242 243 #ifdef CONFIG_HUGETLB_PAGE 244 if (folio_test_hugetlb(folio)) { 245 unsigned int shift = huge_page_shift(hstate_vma(vma)); 246 247 pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 248 if (folio_test_anon(folio)) 249 hugepage_add_anon_rmap(new, vma, pvmw.address, 250 rmap_flags); 251 else 252 page_dup_file_rmap(new, true); 253 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); 254 } else 255 #endif 256 { 257 if (folio_test_anon(folio)) 258 page_add_anon_rmap(new, vma, pvmw.address, 259 rmap_flags); 260 else 261 page_add_file_rmap(new, vma, false); 262 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); 263 } 264 if (vma->vm_flags & VM_LOCKED) 265 mlock_drain_local(); 266 267 trace_remove_migration_pte(pvmw.address, pte_val(pte), 268 compound_order(new)); 269 270 /* No need to invalidate - it was non-present before */ 271 update_mmu_cache(vma, pvmw.address, pvmw.pte); 272 } 273 274 return true; 275 } 276 277 /* 278 * Get rid of all migration entries and replace them by 279 * references to the indicated page. 280 */ 281 void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked) 282 { 283 struct rmap_walk_control rwc = { 284 .rmap_one = remove_migration_pte, 285 .arg = src, 286 }; 287 288 if (locked) 289 rmap_walk_locked(dst, &rwc); 290 else 291 rmap_walk(dst, &rwc); 292 } 293 294 /* 295 * Something used the pte of a page under migration. We need to 296 * get to the page and wait until migration is finished. 297 * When we return from this function the fault will be retried. 298 */ 299 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, 300 spinlock_t *ptl) 301 { 302 pte_t pte; 303 swp_entry_t entry; 304 305 spin_lock(ptl); 306 pte = *ptep; 307 if (!is_swap_pte(pte)) 308 goto out; 309 310 entry = pte_to_swp_entry(pte); 311 if (!is_migration_entry(entry)) 312 goto out; 313 314 migration_entry_wait_on_locked(entry, ptep, ptl); 315 return; 316 out: 317 pte_unmap_unlock(ptep, ptl); 318 } 319 320 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, 321 unsigned long address) 322 { 323 spinlock_t *ptl = pte_lockptr(mm, pmd); 324 pte_t *ptep = pte_offset_map(pmd, address); 325 __migration_entry_wait(mm, ptep, ptl); 326 } 327 328 #ifdef CONFIG_HUGETLB_PAGE 329 /* 330 * The vma read lock must be held upon entry. Holding that lock prevents either 331 * the pte or the ptl from being freed. 332 * 333 * This function will release the vma lock before returning. 334 */ 335 void __migration_entry_wait_huge(struct vm_area_struct *vma, 336 pte_t *ptep, spinlock_t *ptl) 337 { 338 pte_t pte; 339 340 hugetlb_vma_assert_locked(vma); 341 spin_lock(ptl); 342 pte = huge_ptep_get(ptep); 343 344 if (unlikely(!is_hugetlb_entry_migration(pte))) { 345 spin_unlock(ptl); 346 hugetlb_vma_unlock_read(vma); 347 } else { 348 /* 349 * If migration entry existed, safe to release vma lock 350 * here because the pgtable page won't be freed without the 351 * pgtable lock released. See comment right above pgtable 352 * lock release in migration_entry_wait_on_locked(). 353 */ 354 hugetlb_vma_unlock_read(vma); 355 migration_entry_wait_on_locked(pte_to_swp_entry(pte), NULL, ptl); 356 } 357 } 358 359 void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *pte) 360 { 361 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, pte); 362 363 __migration_entry_wait_huge(vma, pte, ptl); 364 } 365 #endif 366 367 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 368 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) 369 { 370 spinlock_t *ptl; 371 372 ptl = pmd_lock(mm, pmd); 373 if (!is_pmd_migration_entry(*pmd)) 374 goto unlock; 375 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl); 376 return; 377 unlock: 378 spin_unlock(ptl); 379 } 380 #endif 381 382 static int folio_expected_refs(struct address_space *mapping, 383 struct folio *folio) 384 { 385 int refs = 1; 386 if (!mapping) 387 return refs; 388 389 refs += folio_nr_pages(folio); 390 if (folio_test_private(folio)) 391 refs++; 392 393 return refs; 394 } 395 396 /* 397 * Replace the page in the mapping. 398 * 399 * The number of remaining references must be: 400 * 1 for anonymous pages without a mapping 401 * 2 for pages with a mapping 402 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. 403 */ 404 int folio_migrate_mapping(struct address_space *mapping, 405 struct folio *newfolio, struct folio *folio, int extra_count) 406 { 407 XA_STATE(xas, &mapping->i_pages, folio_index(folio)); 408 struct zone *oldzone, *newzone; 409 int dirty; 410 int expected_count = folio_expected_refs(mapping, folio) + extra_count; 411 long nr = folio_nr_pages(folio); 412 413 if (!mapping) { 414 /* Anonymous page without mapping */ 415 if (folio_ref_count(folio) != expected_count) 416 return -EAGAIN; 417 418 /* No turning back from here */ 419 newfolio->index = folio->index; 420 newfolio->mapping = folio->mapping; 421 if (folio_test_swapbacked(folio)) 422 __folio_set_swapbacked(newfolio); 423 424 return MIGRATEPAGE_SUCCESS; 425 } 426 427 oldzone = folio_zone(folio); 428 newzone = folio_zone(newfolio); 429 430 xas_lock_irq(&xas); 431 if (!folio_ref_freeze(folio, expected_count)) { 432 xas_unlock_irq(&xas); 433 return -EAGAIN; 434 } 435 436 /* 437 * Now we know that no one else is looking at the folio: 438 * no turning back from here. 439 */ 440 newfolio->index = folio->index; 441 newfolio->mapping = folio->mapping; 442 folio_ref_add(newfolio, nr); /* add cache reference */ 443 if (folio_test_swapbacked(folio)) { 444 __folio_set_swapbacked(newfolio); 445 if (folio_test_swapcache(folio)) { 446 folio_set_swapcache(newfolio); 447 newfolio->private = folio_get_private(folio); 448 } 449 } else { 450 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio); 451 } 452 453 /* Move dirty while page refs frozen and newpage not yet exposed */ 454 dirty = folio_test_dirty(folio); 455 if (dirty) { 456 folio_clear_dirty(folio); 457 folio_set_dirty(newfolio); 458 } 459 460 xas_store(&xas, newfolio); 461 462 /* 463 * Drop cache reference from old page by unfreezing 464 * to one less reference. 465 * We know this isn't the last reference. 466 */ 467 folio_ref_unfreeze(folio, expected_count - nr); 468 469 xas_unlock(&xas); 470 /* Leave irq disabled to prevent preemption while updating stats */ 471 472 /* 473 * If moved to a different zone then also account 474 * the page for that zone. Other VM counters will be 475 * taken care of when we establish references to the 476 * new page and drop references to the old page. 477 * 478 * Note that anonymous pages are accounted for 479 * via NR_FILE_PAGES and NR_ANON_MAPPED if they 480 * are mapped to swap space. 481 */ 482 if (newzone != oldzone) { 483 struct lruvec *old_lruvec, *new_lruvec; 484 struct mem_cgroup *memcg; 485 486 memcg = folio_memcg(folio); 487 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); 488 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); 489 490 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr); 491 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr); 492 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) { 493 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr); 494 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr); 495 } 496 #ifdef CONFIG_SWAP 497 if (folio_test_swapcache(folio)) { 498 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr); 499 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr); 500 } 501 #endif 502 if (dirty && mapping_can_writeback(mapping)) { 503 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr); 504 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr); 505 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr); 506 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr); 507 } 508 } 509 local_irq_enable(); 510 511 return MIGRATEPAGE_SUCCESS; 512 } 513 EXPORT_SYMBOL(folio_migrate_mapping); 514 515 /* 516 * The expected number of remaining references is the same as that 517 * of folio_migrate_mapping(). 518 */ 519 int migrate_huge_page_move_mapping(struct address_space *mapping, 520 struct folio *dst, struct folio *src) 521 { 522 XA_STATE(xas, &mapping->i_pages, folio_index(src)); 523 int expected_count; 524 525 xas_lock_irq(&xas); 526 expected_count = 2 + folio_has_private(src); 527 if (!folio_ref_freeze(src, expected_count)) { 528 xas_unlock_irq(&xas); 529 return -EAGAIN; 530 } 531 532 dst->index = src->index; 533 dst->mapping = src->mapping; 534 535 folio_get(dst); 536 537 xas_store(&xas, dst); 538 539 folio_ref_unfreeze(src, expected_count - 1); 540 541 xas_unlock_irq(&xas); 542 543 return MIGRATEPAGE_SUCCESS; 544 } 545 546 /* 547 * Copy the flags and some other ancillary information 548 */ 549 void folio_migrate_flags(struct folio *newfolio, struct folio *folio) 550 { 551 int cpupid; 552 553 if (folio_test_error(folio)) 554 folio_set_error(newfolio); 555 if (folio_test_referenced(folio)) 556 folio_set_referenced(newfolio); 557 if (folio_test_uptodate(folio)) 558 folio_mark_uptodate(newfolio); 559 if (folio_test_clear_active(folio)) { 560 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio); 561 folio_set_active(newfolio); 562 } else if (folio_test_clear_unevictable(folio)) 563 folio_set_unevictable(newfolio); 564 if (folio_test_workingset(folio)) 565 folio_set_workingset(newfolio); 566 if (folio_test_checked(folio)) 567 folio_set_checked(newfolio); 568 /* 569 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via 570 * migration entries. We can still have PG_anon_exclusive set on an 571 * effectively unmapped and unreferenced first sub-pages of an 572 * anonymous THP: we can simply copy it here via PG_mappedtodisk. 573 */ 574 if (folio_test_mappedtodisk(folio)) 575 folio_set_mappedtodisk(newfolio); 576 577 /* Move dirty on pages not done by folio_migrate_mapping() */ 578 if (folio_test_dirty(folio)) 579 folio_set_dirty(newfolio); 580 581 if (folio_test_young(folio)) 582 folio_set_young(newfolio); 583 if (folio_test_idle(folio)) 584 folio_set_idle(newfolio); 585 586 /* 587 * Copy NUMA information to the new page, to prevent over-eager 588 * future migrations of this same page. 589 */ 590 cpupid = page_cpupid_xchg_last(&folio->page, -1); 591 /* 592 * For memory tiering mode, when migrate between slow and fast 593 * memory node, reset cpupid, because that is used to record 594 * page access time in slow memory node. 595 */ 596 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) { 597 bool f_toptier = node_is_toptier(page_to_nid(&folio->page)); 598 bool t_toptier = node_is_toptier(page_to_nid(&newfolio->page)); 599 600 if (f_toptier != t_toptier) 601 cpupid = -1; 602 } 603 page_cpupid_xchg_last(&newfolio->page, cpupid); 604 605 folio_migrate_ksm(newfolio, folio); 606 /* 607 * Please do not reorder this without considering how mm/ksm.c's 608 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). 609 */ 610 if (folio_test_swapcache(folio)) 611 folio_clear_swapcache(folio); 612 folio_clear_private(folio); 613 614 /* page->private contains hugetlb specific flags */ 615 if (!folio_test_hugetlb(folio)) 616 folio->private = NULL; 617 618 /* 619 * If any waiters have accumulated on the new page then 620 * wake them up. 621 */ 622 if (folio_test_writeback(newfolio)) 623 folio_end_writeback(newfolio); 624 625 /* 626 * PG_readahead shares the same bit with PG_reclaim. The above 627 * end_page_writeback() may clear PG_readahead mistakenly, so set the 628 * bit after that. 629 */ 630 if (folio_test_readahead(folio)) 631 folio_set_readahead(newfolio); 632 633 folio_copy_owner(newfolio, folio); 634 635 if (!folio_test_hugetlb(folio)) 636 mem_cgroup_migrate(folio, newfolio); 637 } 638 EXPORT_SYMBOL(folio_migrate_flags); 639 640 void folio_migrate_copy(struct folio *newfolio, struct folio *folio) 641 { 642 folio_copy(newfolio, folio); 643 folio_migrate_flags(newfolio, folio); 644 } 645 EXPORT_SYMBOL(folio_migrate_copy); 646 647 /************************************************************ 648 * Migration functions 649 ***********************************************************/ 650 651 int migrate_folio_extra(struct address_space *mapping, struct folio *dst, 652 struct folio *src, enum migrate_mode mode, int extra_count) 653 { 654 int rc; 655 656 BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */ 657 658 rc = folio_migrate_mapping(mapping, dst, src, extra_count); 659 660 if (rc != MIGRATEPAGE_SUCCESS) 661 return rc; 662 663 if (mode != MIGRATE_SYNC_NO_COPY) 664 folio_migrate_copy(dst, src); 665 else 666 folio_migrate_flags(dst, src); 667 return MIGRATEPAGE_SUCCESS; 668 } 669 670 /** 671 * migrate_folio() - Simple folio migration. 672 * @mapping: The address_space containing the folio. 673 * @dst: The folio to migrate the data to. 674 * @src: The folio containing the current data. 675 * @mode: How to migrate the page. 676 * 677 * Common logic to directly migrate a single LRU folio suitable for 678 * folios that do not use PagePrivate/PagePrivate2. 679 * 680 * Folios are locked upon entry and exit. 681 */ 682 int migrate_folio(struct address_space *mapping, struct folio *dst, 683 struct folio *src, enum migrate_mode mode) 684 { 685 return migrate_folio_extra(mapping, dst, src, mode, 0); 686 } 687 EXPORT_SYMBOL(migrate_folio); 688 689 #ifdef CONFIG_BLOCK 690 /* Returns true if all buffers are successfully locked */ 691 static bool buffer_migrate_lock_buffers(struct buffer_head *head, 692 enum migrate_mode mode) 693 { 694 struct buffer_head *bh = head; 695 696 /* Simple case, sync compaction */ 697 if (mode != MIGRATE_ASYNC) { 698 do { 699 lock_buffer(bh); 700 bh = bh->b_this_page; 701 702 } while (bh != head); 703 704 return true; 705 } 706 707 /* async case, we cannot block on lock_buffer so use trylock_buffer */ 708 do { 709 if (!trylock_buffer(bh)) { 710 /* 711 * We failed to lock the buffer and cannot stall in 712 * async migration. Release the taken locks 713 */ 714 struct buffer_head *failed_bh = bh; 715 bh = head; 716 while (bh != failed_bh) { 717 unlock_buffer(bh); 718 bh = bh->b_this_page; 719 } 720 return false; 721 } 722 723 bh = bh->b_this_page; 724 } while (bh != head); 725 return true; 726 } 727 728 static int __buffer_migrate_folio(struct address_space *mapping, 729 struct folio *dst, struct folio *src, enum migrate_mode mode, 730 bool check_refs) 731 { 732 struct buffer_head *bh, *head; 733 int rc; 734 int expected_count; 735 736 head = folio_buffers(src); 737 if (!head) 738 return migrate_folio(mapping, dst, src, mode); 739 740 /* Check whether page does not have extra refs before we do more work */ 741 expected_count = folio_expected_refs(mapping, src); 742 if (folio_ref_count(src) != expected_count) 743 return -EAGAIN; 744 745 if (!buffer_migrate_lock_buffers(head, mode)) 746 return -EAGAIN; 747 748 if (check_refs) { 749 bool busy; 750 bool invalidated = false; 751 752 recheck_buffers: 753 busy = false; 754 spin_lock(&mapping->private_lock); 755 bh = head; 756 do { 757 if (atomic_read(&bh->b_count)) { 758 busy = true; 759 break; 760 } 761 bh = bh->b_this_page; 762 } while (bh != head); 763 if (busy) { 764 if (invalidated) { 765 rc = -EAGAIN; 766 goto unlock_buffers; 767 } 768 spin_unlock(&mapping->private_lock); 769 invalidate_bh_lrus(); 770 invalidated = true; 771 goto recheck_buffers; 772 } 773 } 774 775 rc = folio_migrate_mapping(mapping, dst, src, 0); 776 if (rc != MIGRATEPAGE_SUCCESS) 777 goto unlock_buffers; 778 779 folio_attach_private(dst, folio_detach_private(src)); 780 781 bh = head; 782 do { 783 set_bh_page(bh, &dst->page, bh_offset(bh)); 784 bh = bh->b_this_page; 785 } while (bh != head); 786 787 if (mode != MIGRATE_SYNC_NO_COPY) 788 folio_migrate_copy(dst, src); 789 else 790 folio_migrate_flags(dst, src); 791 792 rc = MIGRATEPAGE_SUCCESS; 793 unlock_buffers: 794 if (check_refs) 795 spin_unlock(&mapping->private_lock); 796 bh = head; 797 do { 798 unlock_buffer(bh); 799 bh = bh->b_this_page; 800 } while (bh != head); 801 802 return rc; 803 } 804 805 /** 806 * buffer_migrate_folio() - Migration function for folios with buffers. 807 * @mapping: The address space containing @src. 808 * @dst: The folio to migrate to. 809 * @src: The folio to migrate from. 810 * @mode: How to migrate the folio. 811 * 812 * This function can only be used if the underlying filesystem guarantees 813 * that no other references to @src exist. For example attached buffer 814 * heads are accessed only under the folio lock. If your filesystem cannot 815 * provide this guarantee, buffer_migrate_folio_norefs() may be more 816 * appropriate. 817 * 818 * Return: 0 on success or a negative errno on failure. 819 */ 820 int buffer_migrate_folio(struct address_space *mapping, 821 struct folio *dst, struct folio *src, enum migrate_mode mode) 822 { 823 return __buffer_migrate_folio(mapping, dst, src, mode, false); 824 } 825 EXPORT_SYMBOL(buffer_migrate_folio); 826 827 /** 828 * buffer_migrate_folio_norefs() - Migration function for folios with buffers. 829 * @mapping: The address space containing @src. 830 * @dst: The folio to migrate to. 831 * @src: The folio to migrate from. 832 * @mode: How to migrate the folio. 833 * 834 * Like buffer_migrate_folio() except that this variant is more careful 835 * and checks that there are also no buffer head references. This function 836 * is the right one for mappings where buffer heads are directly looked 837 * up and referenced (such as block device mappings). 838 * 839 * Return: 0 on success or a negative errno on failure. 840 */ 841 int buffer_migrate_folio_norefs(struct address_space *mapping, 842 struct folio *dst, struct folio *src, enum migrate_mode mode) 843 { 844 return __buffer_migrate_folio(mapping, dst, src, mode, true); 845 } 846 EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs); 847 #endif 848 849 int filemap_migrate_folio(struct address_space *mapping, 850 struct folio *dst, struct folio *src, enum migrate_mode mode) 851 { 852 int ret; 853 854 ret = folio_migrate_mapping(mapping, dst, src, 0); 855 if (ret != MIGRATEPAGE_SUCCESS) 856 return ret; 857 858 if (folio_get_private(src)) 859 folio_attach_private(dst, folio_detach_private(src)); 860 861 if (mode != MIGRATE_SYNC_NO_COPY) 862 folio_migrate_copy(dst, src); 863 else 864 folio_migrate_flags(dst, src); 865 return MIGRATEPAGE_SUCCESS; 866 } 867 EXPORT_SYMBOL_GPL(filemap_migrate_folio); 868 869 /* 870 * Writeback a folio to clean the dirty state 871 */ 872 static int writeout(struct address_space *mapping, struct folio *folio) 873 { 874 struct writeback_control wbc = { 875 .sync_mode = WB_SYNC_NONE, 876 .nr_to_write = 1, 877 .range_start = 0, 878 .range_end = LLONG_MAX, 879 .for_reclaim = 1 880 }; 881 int rc; 882 883 if (!mapping->a_ops->writepage) 884 /* No write method for the address space */ 885 return -EINVAL; 886 887 if (!folio_clear_dirty_for_io(folio)) 888 /* Someone else already triggered a write */ 889 return -EAGAIN; 890 891 /* 892 * A dirty folio may imply that the underlying filesystem has 893 * the folio on some queue. So the folio must be clean for 894 * migration. Writeout may mean we lose the lock and the 895 * folio state is no longer what we checked for earlier. 896 * At this point we know that the migration attempt cannot 897 * be successful. 898 */ 899 remove_migration_ptes(folio, folio, false); 900 901 rc = mapping->a_ops->writepage(&folio->page, &wbc); 902 903 if (rc != AOP_WRITEPAGE_ACTIVATE) 904 /* unlocked. Relock */ 905 folio_lock(folio); 906 907 return (rc < 0) ? -EIO : -EAGAIN; 908 } 909 910 /* 911 * Default handling if a filesystem does not provide a migration function. 912 */ 913 static int fallback_migrate_folio(struct address_space *mapping, 914 struct folio *dst, struct folio *src, enum migrate_mode mode) 915 { 916 if (folio_test_dirty(src)) { 917 /* Only writeback folios in full synchronous migration */ 918 switch (mode) { 919 case MIGRATE_SYNC: 920 case MIGRATE_SYNC_NO_COPY: 921 break; 922 default: 923 return -EBUSY; 924 } 925 return writeout(mapping, src); 926 } 927 928 /* 929 * Buffers may be managed in a filesystem specific way. 930 * We must have no buffers or drop them. 931 */ 932 if (folio_test_private(src) && 933 !filemap_release_folio(src, GFP_KERNEL)) 934 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; 935 936 return migrate_folio(mapping, dst, src, mode); 937 } 938 939 /* 940 * Move a page to a newly allocated page 941 * The page is locked and all ptes have been successfully removed. 942 * 943 * The new page will have replaced the old page if this function 944 * is successful. 945 * 946 * Return value: 947 * < 0 - error code 948 * MIGRATEPAGE_SUCCESS - success 949 */ 950 static int move_to_new_folio(struct folio *dst, struct folio *src, 951 enum migrate_mode mode) 952 { 953 int rc = -EAGAIN; 954 bool is_lru = !__PageMovable(&src->page); 955 956 VM_BUG_ON_FOLIO(!folio_test_locked(src), src); 957 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst); 958 959 if (likely(is_lru)) { 960 struct address_space *mapping = folio_mapping(src); 961 962 if (!mapping) 963 rc = migrate_folio(mapping, dst, src, mode); 964 else if (mapping->a_ops->migrate_folio) 965 /* 966 * Most folios have a mapping and most filesystems 967 * provide a migrate_folio callback. Anonymous folios 968 * are part of swap space which also has its own 969 * migrate_folio callback. This is the most common path 970 * for page migration. 971 */ 972 rc = mapping->a_ops->migrate_folio(mapping, dst, src, 973 mode); 974 else 975 rc = fallback_migrate_folio(mapping, dst, src, mode); 976 } else { 977 const struct movable_operations *mops; 978 979 /* 980 * In case of non-lru page, it could be released after 981 * isolation step. In that case, we shouldn't try migration. 982 */ 983 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); 984 if (!folio_test_movable(src)) { 985 rc = MIGRATEPAGE_SUCCESS; 986 folio_clear_isolated(src); 987 goto out; 988 } 989 990 mops = folio_movable_ops(src); 991 rc = mops->migrate_page(&dst->page, &src->page, mode); 992 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && 993 !folio_test_isolated(src)); 994 } 995 996 /* 997 * When successful, old pagecache src->mapping must be cleared before 998 * src is freed; but stats require that PageAnon be left as PageAnon. 999 */ 1000 if (rc == MIGRATEPAGE_SUCCESS) { 1001 if (__PageMovable(&src->page)) { 1002 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); 1003 1004 /* 1005 * We clear PG_movable under page_lock so any compactor 1006 * cannot try to migrate this page. 1007 */ 1008 folio_clear_isolated(src); 1009 } 1010 1011 /* 1012 * Anonymous and movable src->mapping will be cleared by 1013 * free_pages_prepare so don't reset it here for keeping 1014 * the type to work PageAnon, for example. 1015 */ 1016 if (!folio_mapping_flags(src)) 1017 src->mapping = NULL; 1018 1019 if (likely(!folio_is_zone_device(dst))) 1020 flush_dcache_folio(dst); 1021 } 1022 out: 1023 return rc; 1024 } 1025 1026 /* 1027 * To record some information during migration, we use some unused 1028 * fields (mapping and private) of struct folio of the newly allocated 1029 * destination folio. This is safe because nobody is using them 1030 * except us. 1031 */ 1032 union migration_ptr { 1033 struct anon_vma *anon_vma; 1034 struct address_space *mapping; 1035 }; 1036 static void __migrate_folio_record(struct folio *dst, 1037 unsigned long page_was_mapped, 1038 struct anon_vma *anon_vma) 1039 { 1040 union migration_ptr ptr = { .anon_vma = anon_vma }; 1041 dst->mapping = ptr.mapping; 1042 dst->private = (void *)page_was_mapped; 1043 } 1044 1045 static void __migrate_folio_extract(struct folio *dst, 1046 int *page_was_mappedp, 1047 struct anon_vma **anon_vmap) 1048 { 1049 union migration_ptr ptr = { .mapping = dst->mapping }; 1050 *anon_vmap = ptr.anon_vma; 1051 *page_was_mappedp = (unsigned long)dst->private; 1052 dst->mapping = NULL; 1053 dst->private = NULL; 1054 } 1055 1056 /* Restore the source folio to the original state upon failure */ 1057 static void migrate_folio_undo_src(struct folio *src, 1058 int page_was_mapped, 1059 struct anon_vma *anon_vma, 1060 bool locked, 1061 struct list_head *ret) 1062 { 1063 if (page_was_mapped) 1064 remove_migration_ptes(src, src, false); 1065 /* Drop an anon_vma reference if we took one */ 1066 if (anon_vma) 1067 put_anon_vma(anon_vma); 1068 if (locked) 1069 folio_unlock(src); 1070 if (ret) 1071 list_move_tail(&src->lru, ret); 1072 } 1073 1074 /* Restore the destination folio to the original state upon failure */ 1075 static void migrate_folio_undo_dst(struct folio *dst, 1076 bool locked, 1077 free_page_t put_new_page, 1078 unsigned long private) 1079 { 1080 if (locked) 1081 folio_unlock(dst); 1082 if (put_new_page) 1083 put_new_page(&dst->page, private); 1084 else 1085 folio_put(dst); 1086 } 1087 1088 /* Cleanup src folio upon migration success */ 1089 static void migrate_folio_done(struct folio *src, 1090 enum migrate_reason reason) 1091 { 1092 /* 1093 * Compaction can migrate also non-LRU pages which are 1094 * not accounted to NR_ISOLATED_*. They can be recognized 1095 * as __PageMovable 1096 */ 1097 if (likely(!__folio_test_movable(src))) 1098 mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON + 1099 folio_is_file_lru(src), -folio_nr_pages(src)); 1100 1101 if (reason != MR_MEMORY_FAILURE) 1102 /* We release the page in page_handle_poison. */ 1103 folio_put(src); 1104 } 1105 1106 /* Obtain the lock on page, remove all ptes. */ 1107 static int migrate_folio_unmap(new_page_t get_new_page, free_page_t put_new_page, 1108 unsigned long private, struct folio *src, 1109 struct folio **dstp, enum migrate_mode mode, 1110 enum migrate_reason reason, struct list_head *ret) 1111 { 1112 struct folio *dst; 1113 int rc = -EAGAIN; 1114 struct page *newpage = NULL; 1115 int page_was_mapped = 0; 1116 struct anon_vma *anon_vma = NULL; 1117 bool is_lru = !__PageMovable(&src->page); 1118 bool locked = false; 1119 bool dst_locked = false; 1120 1121 if (folio_ref_count(src) == 1) { 1122 /* Folio was freed from under us. So we are done. */ 1123 folio_clear_active(src); 1124 folio_clear_unevictable(src); 1125 /* free_pages_prepare() will clear PG_isolated. */ 1126 list_del(&src->lru); 1127 migrate_folio_done(src, reason); 1128 return MIGRATEPAGE_SUCCESS; 1129 } 1130 1131 newpage = get_new_page(&src->page, private); 1132 if (!newpage) 1133 return -ENOMEM; 1134 dst = page_folio(newpage); 1135 *dstp = dst; 1136 1137 dst->private = NULL; 1138 1139 if (!folio_trylock(src)) { 1140 if (mode == MIGRATE_ASYNC) 1141 goto out; 1142 1143 /* 1144 * It's not safe for direct compaction to call lock_page. 1145 * For example, during page readahead pages are added locked 1146 * to the LRU. Later, when the IO completes the pages are 1147 * marked uptodate and unlocked. However, the queueing 1148 * could be merging multiple pages for one bio (e.g. 1149 * mpage_readahead). If an allocation happens for the 1150 * second or third page, the process can end up locking 1151 * the same page twice and deadlocking. Rather than 1152 * trying to be clever about what pages can be locked, 1153 * avoid the use of lock_page for direct compaction 1154 * altogether. 1155 */ 1156 if (current->flags & PF_MEMALLOC) 1157 goto out; 1158 1159 folio_lock(src); 1160 } 1161 locked = true; 1162 1163 if (folio_test_writeback(src)) { 1164 /* 1165 * Only in the case of a full synchronous migration is it 1166 * necessary to wait for PageWriteback. In the async case, 1167 * the retry loop is too short and in the sync-light case, 1168 * the overhead of stalling is too much 1169 */ 1170 switch (mode) { 1171 case MIGRATE_SYNC: 1172 case MIGRATE_SYNC_NO_COPY: 1173 break; 1174 default: 1175 rc = -EBUSY; 1176 goto out; 1177 } 1178 folio_wait_writeback(src); 1179 } 1180 1181 /* 1182 * By try_to_migrate(), src->mapcount goes down to 0 here. In this case, 1183 * we cannot notice that anon_vma is freed while we migrate a page. 1184 * This get_anon_vma() delays freeing anon_vma pointer until the end 1185 * of migration. File cache pages are no problem because of page_lock() 1186 * File Caches may use write_page() or lock_page() in migration, then, 1187 * just care Anon page here. 1188 * 1189 * Only folio_get_anon_vma() understands the subtleties of 1190 * getting a hold on an anon_vma from outside one of its mms. 1191 * But if we cannot get anon_vma, then we won't need it anyway, 1192 * because that implies that the anon page is no longer mapped 1193 * (and cannot be remapped so long as we hold the page lock). 1194 */ 1195 if (folio_test_anon(src) && !folio_test_ksm(src)) 1196 anon_vma = folio_get_anon_vma(src); 1197 1198 /* 1199 * Block others from accessing the new page when we get around to 1200 * establishing additional references. We are usually the only one 1201 * holding a reference to dst at this point. We used to have a BUG 1202 * here if folio_trylock(dst) fails, but would like to allow for 1203 * cases where there might be a race with the previous use of dst. 1204 * This is much like races on refcount of oldpage: just don't BUG(). 1205 */ 1206 if (unlikely(!folio_trylock(dst))) 1207 goto out; 1208 dst_locked = true; 1209 1210 if (unlikely(!is_lru)) { 1211 __migrate_folio_record(dst, page_was_mapped, anon_vma); 1212 return MIGRATEPAGE_UNMAP; 1213 } 1214 1215 /* 1216 * Corner case handling: 1217 * 1. When a new swap-cache page is read into, it is added to the LRU 1218 * and treated as swapcache but it has no rmap yet. 1219 * Calling try_to_unmap() against a src->mapping==NULL page will 1220 * trigger a BUG. So handle it here. 1221 * 2. An orphaned page (see truncate_cleanup_page) might have 1222 * fs-private metadata. The page can be picked up due to memory 1223 * offlining. Everywhere else except page reclaim, the page is 1224 * invisible to the vm, so the page can not be migrated. So try to 1225 * free the metadata, so the page can be freed. 1226 */ 1227 if (!src->mapping) { 1228 if (folio_test_private(src)) { 1229 try_to_free_buffers(src); 1230 goto out; 1231 } 1232 } else if (folio_mapped(src)) { 1233 /* Establish migration ptes */ 1234 VM_BUG_ON_FOLIO(folio_test_anon(src) && 1235 !folio_test_ksm(src) && !anon_vma, src); 1236 try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0); 1237 page_was_mapped = 1; 1238 } 1239 1240 if (!folio_mapped(src)) { 1241 __migrate_folio_record(dst, page_was_mapped, anon_vma); 1242 return MIGRATEPAGE_UNMAP; 1243 } 1244 1245 out: 1246 /* 1247 * A folio that has not been unmapped will be restored to 1248 * right list unless we want to retry. 1249 */ 1250 if (rc == -EAGAIN) 1251 ret = NULL; 1252 1253 migrate_folio_undo_src(src, page_was_mapped, anon_vma, locked, ret); 1254 migrate_folio_undo_dst(dst, dst_locked, put_new_page, private); 1255 1256 return rc; 1257 } 1258 1259 /* Migrate the folio to the newly allocated folio in dst. */ 1260 static int migrate_folio_move(free_page_t put_new_page, unsigned long private, 1261 struct folio *src, struct folio *dst, 1262 enum migrate_mode mode, enum migrate_reason reason, 1263 struct list_head *ret) 1264 { 1265 int rc; 1266 int page_was_mapped = 0; 1267 struct anon_vma *anon_vma = NULL; 1268 bool is_lru = !__PageMovable(&src->page); 1269 struct list_head *prev; 1270 1271 __migrate_folio_extract(dst, &page_was_mapped, &anon_vma); 1272 prev = dst->lru.prev; 1273 list_del(&dst->lru); 1274 1275 rc = move_to_new_folio(dst, src, mode); 1276 if (rc) 1277 goto out; 1278 1279 if (unlikely(!is_lru)) 1280 goto out_unlock_both; 1281 1282 /* 1283 * When successful, push dst to LRU immediately: so that if it 1284 * turns out to be an mlocked page, remove_migration_ptes() will 1285 * automatically build up the correct dst->mlock_count for it. 1286 * 1287 * We would like to do something similar for the old page, when 1288 * unsuccessful, and other cases when a page has been temporarily 1289 * isolated from the unevictable LRU: but this case is the easiest. 1290 */ 1291 folio_add_lru(dst); 1292 if (page_was_mapped) 1293 lru_add_drain(); 1294 1295 if (page_was_mapped) 1296 remove_migration_ptes(src, dst, false); 1297 1298 out_unlock_both: 1299 folio_unlock(dst); 1300 set_page_owner_migrate_reason(&dst->page, reason); 1301 /* 1302 * If migration is successful, decrease refcount of dst, 1303 * which will not free the page because new page owner increased 1304 * refcounter. 1305 */ 1306 folio_put(dst); 1307 1308 /* 1309 * A folio that has been migrated has all references removed 1310 * and will be freed. 1311 */ 1312 list_del(&src->lru); 1313 /* Drop an anon_vma reference if we took one */ 1314 if (anon_vma) 1315 put_anon_vma(anon_vma); 1316 folio_unlock(src); 1317 migrate_folio_done(src, reason); 1318 1319 return rc; 1320 out: 1321 /* 1322 * A folio that has not been migrated will be restored to 1323 * right list unless we want to retry. 1324 */ 1325 if (rc == -EAGAIN) { 1326 list_add(&dst->lru, prev); 1327 __migrate_folio_record(dst, page_was_mapped, anon_vma); 1328 return rc; 1329 } 1330 1331 migrate_folio_undo_src(src, page_was_mapped, anon_vma, true, ret); 1332 migrate_folio_undo_dst(dst, true, put_new_page, private); 1333 1334 return rc; 1335 } 1336 1337 /* 1338 * Counterpart of unmap_and_move_page() for hugepage migration. 1339 * 1340 * This function doesn't wait the completion of hugepage I/O 1341 * because there is no race between I/O and migration for hugepage. 1342 * Note that currently hugepage I/O occurs only in direct I/O 1343 * where no lock is held and PG_writeback is irrelevant, 1344 * and writeback status of all subpages are counted in the reference 1345 * count of the head page (i.e. if all subpages of a 2MB hugepage are 1346 * under direct I/O, the reference of the head page is 512 and a bit more.) 1347 * This means that when we try to migrate hugepage whose subpages are 1348 * doing direct I/O, some references remain after try_to_unmap() and 1349 * hugepage migration fails without data corruption. 1350 * 1351 * There is also no race when direct I/O is issued on the page under migration, 1352 * because then pte is replaced with migration swap entry and direct I/O code 1353 * will wait in the page fault for migration to complete. 1354 */ 1355 static int unmap_and_move_huge_page(new_page_t get_new_page, 1356 free_page_t put_new_page, unsigned long private, 1357 struct page *hpage, int force, 1358 enum migrate_mode mode, int reason, 1359 struct list_head *ret) 1360 { 1361 struct folio *dst, *src = page_folio(hpage); 1362 int rc = -EAGAIN; 1363 int page_was_mapped = 0; 1364 struct page *new_hpage; 1365 struct anon_vma *anon_vma = NULL; 1366 struct address_space *mapping = NULL; 1367 1368 if (folio_ref_count(src) == 1) { 1369 /* page was freed from under us. So we are done. */ 1370 folio_putback_active_hugetlb(src); 1371 return MIGRATEPAGE_SUCCESS; 1372 } 1373 1374 new_hpage = get_new_page(hpage, private); 1375 if (!new_hpage) 1376 return -ENOMEM; 1377 dst = page_folio(new_hpage); 1378 1379 if (!folio_trylock(src)) { 1380 if (!force) 1381 goto out; 1382 switch (mode) { 1383 case MIGRATE_SYNC: 1384 case MIGRATE_SYNC_NO_COPY: 1385 break; 1386 default: 1387 goto out; 1388 } 1389 folio_lock(src); 1390 } 1391 1392 /* 1393 * Check for pages which are in the process of being freed. Without 1394 * folio_mapping() set, hugetlbfs specific move page routine will not 1395 * be called and we could leak usage counts for subpools. 1396 */ 1397 if (hugetlb_folio_subpool(src) && !folio_mapping(src)) { 1398 rc = -EBUSY; 1399 goto out_unlock; 1400 } 1401 1402 if (folio_test_anon(src)) 1403 anon_vma = folio_get_anon_vma(src); 1404 1405 if (unlikely(!folio_trylock(dst))) 1406 goto put_anon; 1407 1408 if (folio_mapped(src)) { 1409 enum ttu_flags ttu = 0; 1410 1411 if (!folio_test_anon(src)) { 1412 /* 1413 * In shared mappings, try_to_unmap could potentially 1414 * call huge_pmd_unshare. Because of this, take 1415 * semaphore in write mode here and set TTU_RMAP_LOCKED 1416 * to let lower levels know we have taken the lock. 1417 */ 1418 mapping = hugetlb_page_mapping_lock_write(hpage); 1419 if (unlikely(!mapping)) 1420 goto unlock_put_anon; 1421 1422 ttu = TTU_RMAP_LOCKED; 1423 } 1424 1425 try_to_migrate(src, ttu); 1426 page_was_mapped = 1; 1427 1428 if (ttu & TTU_RMAP_LOCKED) 1429 i_mmap_unlock_write(mapping); 1430 } 1431 1432 if (!folio_mapped(src)) 1433 rc = move_to_new_folio(dst, src, mode); 1434 1435 if (page_was_mapped) 1436 remove_migration_ptes(src, 1437 rc == MIGRATEPAGE_SUCCESS ? dst : src, false); 1438 1439 unlock_put_anon: 1440 folio_unlock(dst); 1441 1442 put_anon: 1443 if (anon_vma) 1444 put_anon_vma(anon_vma); 1445 1446 if (rc == MIGRATEPAGE_SUCCESS) { 1447 move_hugetlb_state(src, dst, reason); 1448 put_new_page = NULL; 1449 } 1450 1451 out_unlock: 1452 folio_unlock(src); 1453 out: 1454 if (rc == MIGRATEPAGE_SUCCESS) 1455 folio_putback_active_hugetlb(src); 1456 else if (rc != -EAGAIN) 1457 list_move_tail(&src->lru, ret); 1458 1459 /* 1460 * If migration was not successful and there's a freeing callback, use 1461 * it. Otherwise, put_page() will drop the reference grabbed during 1462 * isolation. 1463 */ 1464 if (put_new_page) 1465 put_new_page(new_hpage, private); 1466 else 1467 folio_putback_active_hugetlb(dst); 1468 1469 return rc; 1470 } 1471 1472 static inline int try_split_folio(struct folio *folio, struct list_head *split_folios) 1473 { 1474 int rc; 1475 1476 folio_lock(folio); 1477 rc = split_folio_to_list(folio, split_folios); 1478 folio_unlock(folio); 1479 if (!rc) 1480 list_move_tail(&folio->lru, split_folios); 1481 1482 return rc; 1483 } 1484 1485 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 1486 #define NR_MAX_BATCHED_MIGRATION HPAGE_PMD_NR 1487 #else 1488 #define NR_MAX_BATCHED_MIGRATION 512 1489 #endif 1490 #define NR_MAX_MIGRATE_PAGES_RETRY 10 1491 #define NR_MAX_MIGRATE_ASYNC_RETRY 3 1492 #define NR_MAX_MIGRATE_SYNC_RETRY \ 1493 (NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY) 1494 1495 struct migrate_pages_stats { 1496 int nr_succeeded; /* Normal and large folios migrated successfully, in 1497 units of base pages */ 1498 int nr_failed_pages; /* Normal and large folios failed to be migrated, in 1499 units of base pages. Untried folios aren't counted */ 1500 int nr_thp_succeeded; /* THP migrated successfully */ 1501 int nr_thp_failed; /* THP failed to be migrated */ 1502 int nr_thp_split; /* THP split before migrating */ 1503 }; 1504 1505 /* 1506 * Returns the number of hugetlb folios that were not migrated, or an error code 1507 * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable 1508 * any more because the list has become empty or no retryable hugetlb folios 1509 * exist any more. It is caller's responsibility to call putback_movable_pages() 1510 * only if ret != 0. 1511 */ 1512 static int migrate_hugetlbs(struct list_head *from, new_page_t get_new_page, 1513 free_page_t put_new_page, unsigned long private, 1514 enum migrate_mode mode, int reason, 1515 struct migrate_pages_stats *stats, 1516 struct list_head *ret_folios) 1517 { 1518 int retry = 1; 1519 int nr_failed = 0; 1520 int nr_retry_pages = 0; 1521 int pass = 0; 1522 struct folio *folio, *folio2; 1523 int rc, nr_pages; 1524 1525 for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) { 1526 retry = 0; 1527 nr_retry_pages = 0; 1528 1529 list_for_each_entry_safe(folio, folio2, from, lru) { 1530 if (!folio_test_hugetlb(folio)) 1531 continue; 1532 1533 nr_pages = folio_nr_pages(folio); 1534 1535 cond_resched(); 1536 1537 /* 1538 * Migratability of hugepages depends on architectures and 1539 * their size. This check is necessary because some callers 1540 * of hugepage migration like soft offline and memory 1541 * hotremove don't walk through page tables or check whether 1542 * the hugepage is pmd-based or not before kicking migration. 1543 */ 1544 if (!hugepage_migration_supported(folio_hstate(folio))) { 1545 nr_failed++; 1546 stats->nr_failed_pages += nr_pages; 1547 list_move_tail(&folio->lru, ret_folios); 1548 continue; 1549 } 1550 1551 rc = unmap_and_move_huge_page(get_new_page, 1552 put_new_page, private, 1553 &folio->page, pass > 2, mode, 1554 reason, ret_folios); 1555 /* 1556 * The rules are: 1557 * Success: hugetlb folio will be put back 1558 * -EAGAIN: stay on the from list 1559 * -ENOMEM: stay on the from list 1560 * Other errno: put on ret_folios list 1561 */ 1562 switch(rc) { 1563 case -ENOMEM: 1564 /* 1565 * When memory is low, don't bother to try to migrate 1566 * other folios, just exit. 1567 */ 1568 stats->nr_failed_pages += nr_pages + nr_retry_pages; 1569 return -ENOMEM; 1570 case -EAGAIN: 1571 retry++; 1572 nr_retry_pages += nr_pages; 1573 break; 1574 case MIGRATEPAGE_SUCCESS: 1575 stats->nr_succeeded += nr_pages; 1576 break; 1577 default: 1578 /* 1579 * Permanent failure (-EBUSY, etc.): 1580 * unlike -EAGAIN case, the failed folio is 1581 * removed from migration folio list and not 1582 * retried in the next outer loop. 1583 */ 1584 nr_failed++; 1585 stats->nr_failed_pages += nr_pages; 1586 break; 1587 } 1588 } 1589 } 1590 /* 1591 * nr_failed is number of hugetlb folios failed to be migrated. After 1592 * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb 1593 * folios as failed. 1594 */ 1595 nr_failed += retry; 1596 stats->nr_failed_pages += nr_retry_pages; 1597 1598 return nr_failed; 1599 } 1600 1601 /* 1602 * migrate_pages_batch() first unmaps folios in the from list as many as 1603 * possible, then move the unmapped folios. 1604 * 1605 * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a 1606 * lock or bit when we have locked more than one folio. Which may cause 1607 * deadlock (e.g., for loop device). So, if mode != MIGRATE_ASYNC, the 1608 * length of the from list must be <= 1. 1609 */ 1610 static int migrate_pages_batch(struct list_head *from, new_page_t get_new_page, 1611 free_page_t put_new_page, unsigned long private, 1612 enum migrate_mode mode, int reason, struct list_head *ret_folios, 1613 struct list_head *split_folios, struct migrate_pages_stats *stats, 1614 int nr_pass) 1615 { 1616 int retry = 1; 1617 int large_retry = 1; 1618 int thp_retry = 1; 1619 int nr_failed = 0; 1620 int nr_retry_pages = 0; 1621 int nr_large_failed = 0; 1622 int pass = 0; 1623 bool is_large = false; 1624 bool is_thp = false; 1625 struct folio *folio, *folio2, *dst = NULL, *dst2; 1626 int rc, rc_saved = 0, nr_pages; 1627 LIST_HEAD(unmap_folios); 1628 LIST_HEAD(dst_folios); 1629 bool nosplit = (reason == MR_NUMA_MISPLACED); 1630 1631 VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC && 1632 !list_empty(from) && !list_is_singular(from)); 1633 1634 for (pass = 0; pass < nr_pass && (retry || large_retry); pass++) { 1635 retry = 0; 1636 large_retry = 0; 1637 thp_retry = 0; 1638 nr_retry_pages = 0; 1639 1640 list_for_each_entry_safe(folio, folio2, from, lru) { 1641 /* 1642 * Large folio statistics is based on the source large 1643 * folio. Capture required information that might get 1644 * lost during migration. 1645 */ 1646 is_large = folio_test_large(folio); 1647 is_thp = is_large && folio_test_pmd_mappable(folio); 1648 nr_pages = folio_nr_pages(folio); 1649 1650 cond_resched(); 1651 1652 /* 1653 * Large folio migration might be unsupported or 1654 * the allocation might be failed so we should retry 1655 * on the same folio with the large folio split 1656 * to normal folios. 1657 * 1658 * Split folios are put in split_folios, and 1659 * we will migrate them after the rest of the 1660 * list is processed. 1661 */ 1662 if (!thp_migration_supported() && is_thp) { 1663 nr_large_failed++; 1664 stats->nr_thp_failed++; 1665 if (!try_split_folio(folio, split_folios)) { 1666 stats->nr_thp_split++; 1667 continue; 1668 } 1669 stats->nr_failed_pages += nr_pages; 1670 list_move_tail(&folio->lru, ret_folios); 1671 continue; 1672 } 1673 1674 rc = migrate_folio_unmap(get_new_page, put_new_page, private, 1675 folio, &dst, mode, reason, ret_folios); 1676 /* 1677 * The rules are: 1678 * Success: folio will be freed 1679 * Unmap: folio will be put on unmap_folios list, 1680 * dst folio put on dst_folios list 1681 * -EAGAIN: stay on the from list 1682 * -ENOMEM: stay on the from list 1683 * Other errno: put on ret_folios list 1684 */ 1685 switch(rc) { 1686 case -ENOMEM: 1687 /* 1688 * When memory is low, don't bother to try to migrate 1689 * other folios, move unmapped folios, then exit. 1690 */ 1691 if (is_large) { 1692 nr_large_failed++; 1693 stats->nr_thp_failed += is_thp; 1694 /* Large folio NUMA faulting doesn't split to retry. */ 1695 if (!nosplit) { 1696 int ret = try_split_folio(folio, split_folios); 1697 1698 if (!ret) { 1699 stats->nr_thp_split += is_thp; 1700 break; 1701 } else if (reason == MR_LONGTERM_PIN && 1702 ret == -EAGAIN) { 1703 /* 1704 * Try again to split large folio to 1705 * mitigate the failure of longterm pinning. 1706 */ 1707 large_retry++; 1708 thp_retry += is_thp; 1709 nr_retry_pages += nr_pages; 1710 /* Undo duplicated failure counting. */ 1711 nr_large_failed--; 1712 stats->nr_thp_failed -= is_thp; 1713 break; 1714 } 1715 } 1716 } else { 1717 nr_failed++; 1718 } 1719 1720 stats->nr_failed_pages += nr_pages + nr_retry_pages; 1721 /* nr_failed isn't updated for not used */ 1722 nr_large_failed += large_retry; 1723 stats->nr_thp_failed += thp_retry; 1724 rc_saved = rc; 1725 if (list_empty(&unmap_folios)) 1726 goto out; 1727 else 1728 goto move; 1729 case -EAGAIN: 1730 if (is_large) { 1731 large_retry++; 1732 thp_retry += is_thp; 1733 } else { 1734 retry++; 1735 } 1736 nr_retry_pages += nr_pages; 1737 break; 1738 case MIGRATEPAGE_SUCCESS: 1739 stats->nr_succeeded += nr_pages; 1740 stats->nr_thp_succeeded += is_thp; 1741 break; 1742 case MIGRATEPAGE_UNMAP: 1743 list_move_tail(&folio->lru, &unmap_folios); 1744 list_add_tail(&dst->lru, &dst_folios); 1745 break; 1746 default: 1747 /* 1748 * Permanent failure (-EBUSY, etc.): 1749 * unlike -EAGAIN case, the failed folio is 1750 * removed from migration folio list and not 1751 * retried in the next outer loop. 1752 */ 1753 if (is_large) { 1754 nr_large_failed++; 1755 stats->nr_thp_failed += is_thp; 1756 } else { 1757 nr_failed++; 1758 } 1759 1760 stats->nr_failed_pages += nr_pages; 1761 break; 1762 } 1763 } 1764 } 1765 nr_failed += retry; 1766 nr_large_failed += large_retry; 1767 stats->nr_thp_failed += thp_retry; 1768 stats->nr_failed_pages += nr_retry_pages; 1769 move: 1770 /* Flush TLBs for all unmapped folios */ 1771 try_to_unmap_flush(); 1772 1773 retry = 1; 1774 for (pass = 0; pass < nr_pass && (retry || large_retry); pass++) { 1775 retry = 0; 1776 large_retry = 0; 1777 thp_retry = 0; 1778 nr_retry_pages = 0; 1779 1780 dst = list_first_entry(&dst_folios, struct folio, lru); 1781 dst2 = list_next_entry(dst, lru); 1782 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) { 1783 is_large = folio_test_large(folio); 1784 is_thp = is_large && folio_test_pmd_mappable(folio); 1785 nr_pages = folio_nr_pages(folio); 1786 1787 cond_resched(); 1788 1789 rc = migrate_folio_move(put_new_page, private, 1790 folio, dst, mode, 1791 reason, ret_folios); 1792 /* 1793 * The rules are: 1794 * Success: folio will be freed 1795 * -EAGAIN: stay on the unmap_folios list 1796 * Other errno: put on ret_folios list 1797 */ 1798 switch(rc) { 1799 case -EAGAIN: 1800 if (is_large) { 1801 large_retry++; 1802 thp_retry += is_thp; 1803 } else { 1804 retry++; 1805 } 1806 nr_retry_pages += nr_pages; 1807 break; 1808 case MIGRATEPAGE_SUCCESS: 1809 stats->nr_succeeded += nr_pages; 1810 stats->nr_thp_succeeded += is_thp; 1811 break; 1812 default: 1813 if (is_large) { 1814 nr_large_failed++; 1815 stats->nr_thp_failed += is_thp; 1816 } else { 1817 nr_failed++; 1818 } 1819 1820 stats->nr_failed_pages += nr_pages; 1821 break; 1822 } 1823 dst = dst2; 1824 dst2 = list_next_entry(dst, lru); 1825 } 1826 } 1827 nr_failed += retry; 1828 nr_large_failed += large_retry; 1829 stats->nr_thp_failed += thp_retry; 1830 stats->nr_failed_pages += nr_retry_pages; 1831 1832 if (rc_saved) 1833 rc = rc_saved; 1834 else 1835 rc = nr_failed + nr_large_failed; 1836 out: 1837 /* Cleanup remaining folios */ 1838 dst = list_first_entry(&dst_folios, struct folio, lru); 1839 dst2 = list_next_entry(dst, lru); 1840 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) { 1841 int page_was_mapped = 0; 1842 struct anon_vma *anon_vma = NULL; 1843 1844 __migrate_folio_extract(dst, &page_was_mapped, &anon_vma); 1845 migrate_folio_undo_src(folio, page_was_mapped, anon_vma, 1846 true, ret_folios); 1847 list_del(&dst->lru); 1848 migrate_folio_undo_dst(dst, true, put_new_page, private); 1849 dst = dst2; 1850 dst2 = list_next_entry(dst, lru); 1851 } 1852 1853 return rc; 1854 } 1855 1856 static int migrate_pages_sync(struct list_head *from, new_page_t get_new_page, 1857 free_page_t put_new_page, unsigned long private, 1858 enum migrate_mode mode, int reason, struct list_head *ret_folios, 1859 struct list_head *split_folios, struct migrate_pages_stats *stats) 1860 { 1861 int rc, nr_failed = 0; 1862 LIST_HEAD(folios); 1863 struct migrate_pages_stats astats; 1864 1865 memset(&astats, 0, sizeof(astats)); 1866 /* Try to migrate in batch with MIGRATE_ASYNC mode firstly */ 1867 rc = migrate_pages_batch(from, get_new_page, put_new_page, private, MIGRATE_ASYNC, 1868 reason, &folios, split_folios, &astats, 1869 NR_MAX_MIGRATE_ASYNC_RETRY); 1870 stats->nr_succeeded += astats.nr_succeeded; 1871 stats->nr_thp_succeeded += astats.nr_thp_succeeded; 1872 stats->nr_thp_split += astats.nr_thp_split; 1873 if (rc < 0) { 1874 stats->nr_failed_pages += astats.nr_failed_pages; 1875 stats->nr_thp_failed += astats.nr_thp_failed; 1876 list_splice_tail(&folios, ret_folios); 1877 return rc; 1878 } 1879 stats->nr_thp_failed += astats.nr_thp_split; 1880 nr_failed += astats.nr_thp_split; 1881 /* 1882 * Fall back to migrate all failed folios one by one synchronously. All 1883 * failed folios except split THPs will be retried, so their failure 1884 * isn't counted 1885 */ 1886 list_splice_tail_init(&folios, from); 1887 while (!list_empty(from)) { 1888 list_move(from->next, &folios); 1889 rc = migrate_pages_batch(&folios, get_new_page, put_new_page, 1890 private, mode, reason, ret_folios, 1891 split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY); 1892 list_splice_tail_init(&folios, ret_folios); 1893 if (rc < 0) 1894 return rc; 1895 nr_failed += rc; 1896 } 1897 1898 return nr_failed; 1899 } 1900 1901 /* 1902 * migrate_pages - migrate the folios specified in a list, to the free folios 1903 * supplied as the target for the page migration 1904 * 1905 * @from: The list of folios to be migrated. 1906 * @get_new_page: The function used to allocate free folios to be used 1907 * as the target of the folio migration. 1908 * @put_new_page: The function used to free target folios if migration 1909 * fails, or NULL if no special handling is necessary. 1910 * @private: Private data to be passed on to get_new_page() 1911 * @mode: The migration mode that specifies the constraints for 1912 * folio migration, if any. 1913 * @reason: The reason for folio migration. 1914 * @ret_succeeded: Set to the number of folios migrated successfully if 1915 * the caller passes a non-NULL pointer. 1916 * 1917 * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios 1918 * are movable any more because the list has become empty or no retryable folios 1919 * exist any more. It is caller's responsibility to call putback_movable_pages() 1920 * only if ret != 0. 1921 * 1922 * Returns the number of {normal folio, large folio, hugetlb} that were not 1923 * migrated, or an error code. The number of large folio splits will be 1924 * considered as the number of non-migrated large folio, no matter how many 1925 * split folios of the large folio are migrated successfully. 1926 */ 1927 int migrate_pages(struct list_head *from, new_page_t get_new_page, 1928 free_page_t put_new_page, unsigned long private, 1929 enum migrate_mode mode, int reason, unsigned int *ret_succeeded) 1930 { 1931 int rc, rc_gather; 1932 int nr_pages; 1933 struct folio *folio, *folio2; 1934 LIST_HEAD(folios); 1935 LIST_HEAD(ret_folios); 1936 LIST_HEAD(split_folios); 1937 struct migrate_pages_stats stats; 1938 1939 trace_mm_migrate_pages_start(mode, reason); 1940 1941 memset(&stats, 0, sizeof(stats)); 1942 1943 rc_gather = migrate_hugetlbs(from, get_new_page, put_new_page, private, 1944 mode, reason, &stats, &ret_folios); 1945 if (rc_gather < 0) 1946 goto out; 1947 1948 again: 1949 nr_pages = 0; 1950 list_for_each_entry_safe(folio, folio2, from, lru) { 1951 /* Retried hugetlb folios will be kept in list */ 1952 if (folio_test_hugetlb(folio)) { 1953 list_move_tail(&folio->lru, &ret_folios); 1954 continue; 1955 } 1956 1957 nr_pages += folio_nr_pages(folio); 1958 if (nr_pages >= NR_MAX_BATCHED_MIGRATION) 1959 break; 1960 } 1961 if (nr_pages >= NR_MAX_BATCHED_MIGRATION) 1962 list_cut_before(&folios, from, &folio2->lru); 1963 else 1964 list_splice_init(from, &folios); 1965 if (mode == MIGRATE_ASYNC) 1966 rc = migrate_pages_batch(&folios, get_new_page, put_new_page, private, 1967 mode, reason, &ret_folios, &split_folios, &stats, 1968 NR_MAX_MIGRATE_PAGES_RETRY); 1969 else 1970 rc = migrate_pages_sync(&folios, get_new_page, put_new_page, private, 1971 mode, reason, &ret_folios, &split_folios, &stats); 1972 list_splice_tail_init(&folios, &ret_folios); 1973 if (rc < 0) { 1974 rc_gather = rc; 1975 list_splice_tail(&split_folios, &ret_folios); 1976 goto out; 1977 } 1978 if (!list_empty(&split_folios)) { 1979 /* 1980 * Failure isn't counted since all split folios of a large folio 1981 * is counted as 1 failure already. And, we only try to migrate 1982 * with minimal effort, force MIGRATE_ASYNC mode and retry once. 1983 */ 1984 migrate_pages_batch(&split_folios, get_new_page, put_new_page, private, 1985 MIGRATE_ASYNC, reason, &ret_folios, NULL, &stats, 1); 1986 list_splice_tail_init(&split_folios, &ret_folios); 1987 } 1988 rc_gather += rc; 1989 if (!list_empty(from)) 1990 goto again; 1991 out: 1992 /* 1993 * Put the permanent failure folio back to migration list, they 1994 * will be put back to the right list by the caller. 1995 */ 1996 list_splice(&ret_folios, from); 1997 1998 /* 1999 * Return 0 in case all split folios of fail-to-migrate large folios 2000 * are migrated successfully. 2001 */ 2002 if (list_empty(from)) 2003 rc_gather = 0; 2004 2005 count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded); 2006 count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages); 2007 count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded); 2008 count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed); 2009 count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split); 2010 trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages, 2011 stats.nr_thp_succeeded, stats.nr_thp_failed, 2012 stats.nr_thp_split, mode, reason); 2013 2014 if (ret_succeeded) 2015 *ret_succeeded = stats.nr_succeeded; 2016 2017 return rc_gather; 2018 } 2019 2020 struct page *alloc_migration_target(struct page *page, unsigned long private) 2021 { 2022 struct folio *folio = page_folio(page); 2023 struct migration_target_control *mtc; 2024 gfp_t gfp_mask; 2025 unsigned int order = 0; 2026 struct folio *hugetlb_folio = NULL; 2027 struct folio *new_folio = NULL; 2028 int nid; 2029 int zidx; 2030 2031 mtc = (struct migration_target_control *)private; 2032 gfp_mask = mtc->gfp_mask; 2033 nid = mtc->nid; 2034 if (nid == NUMA_NO_NODE) 2035 nid = folio_nid(folio); 2036 2037 if (folio_test_hugetlb(folio)) { 2038 struct hstate *h = folio_hstate(folio); 2039 2040 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); 2041 hugetlb_folio = alloc_hugetlb_folio_nodemask(h, nid, 2042 mtc->nmask, gfp_mask); 2043 return &hugetlb_folio->page; 2044 } 2045 2046 if (folio_test_large(folio)) { 2047 /* 2048 * clear __GFP_RECLAIM to make the migration callback 2049 * consistent with regular THP allocations. 2050 */ 2051 gfp_mask &= ~__GFP_RECLAIM; 2052 gfp_mask |= GFP_TRANSHUGE; 2053 order = folio_order(folio); 2054 } 2055 zidx = zone_idx(folio_zone(folio)); 2056 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) 2057 gfp_mask |= __GFP_HIGHMEM; 2058 2059 new_folio = __folio_alloc(gfp_mask, order, nid, mtc->nmask); 2060 2061 return &new_folio->page; 2062 } 2063 2064 #ifdef CONFIG_NUMA 2065 2066 static int store_status(int __user *status, int start, int value, int nr) 2067 { 2068 while (nr-- > 0) { 2069 if (put_user(value, status + start)) 2070 return -EFAULT; 2071 start++; 2072 } 2073 2074 return 0; 2075 } 2076 2077 static int do_move_pages_to_node(struct mm_struct *mm, 2078 struct list_head *pagelist, int node) 2079 { 2080 int err; 2081 struct migration_target_control mtc = { 2082 .nid = node, 2083 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 2084 }; 2085 2086 err = migrate_pages(pagelist, alloc_migration_target, NULL, 2087 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL); 2088 if (err) 2089 putback_movable_pages(pagelist); 2090 return err; 2091 } 2092 2093 /* 2094 * Resolves the given address to a struct page, isolates it from the LRU and 2095 * puts it to the given pagelist. 2096 * Returns: 2097 * errno - if the page cannot be found/isolated 2098 * 0 - when it doesn't have to be migrated because it is already on the 2099 * target node 2100 * 1 - when it has been queued 2101 */ 2102 static int add_page_for_migration(struct mm_struct *mm, const void __user *p, 2103 int node, struct list_head *pagelist, bool migrate_all) 2104 { 2105 struct vm_area_struct *vma; 2106 unsigned long addr; 2107 struct page *page; 2108 int err; 2109 bool isolated; 2110 2111 mmap_read_lock(mm); 2112 addr = (unsigned long)untagged_addr_remote(mm, p); 2113 2114 err = -EFAULT; 2115 vma = vma_lookup(mm, addr); 2116 if (!vma || !vma_migratable(vma)) 2117 goto out; 2118 2119 /* FOLL_DUMP to ignore special (like zero) pages */ 2120 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 2121 2122 err = PTR_ERR(page); 2123 if (IS_ERR(page)) 2124 goto out; 2125 2126 err = -ENOENT; 2127 if (!page) 2128 goto out; 2129 2130 if (is_zone_device_page(page)) 2131 goto out_putpage; 2132 2133 err = 0; 2134 if (page_to_nid(page) == node) 2135 goto out_putpage; 2136 2137 err = -EACCES; 2138 if (page_mapcount(page) > 1 && !migrate_all) 2139 goto out_putpage; 2140 2141 if (PageHuge(page)) { 2142 if (PageHead(page)) { 2143 isolated = isolate_hugetlb(page_folio(page), pagelist); 2144 err = isolated ? 1 : -EBUSY; 2145 } 2146 } else { 2147 struct page *head; 2148 2149 head = compound_head(page); 2150 isolated = isolate_lru_page(head); 2151 if (!isolated) { 2152 err = -EBUSY; 2153 goto out_putpage; 2154 } 2155 2156 err = 1; 2157 list_add_tail(&head->lru, pagelist); 2158 mod_node_page_state(page_pgdat(head), 2159 NR_ISOLATED_ANON + page_is_file_lru(head), 2160 thp_nr_pages(head)); 2161 } 2162 out_putpage: 2163 /* 2164 * Either remove the duplicate refcount from 2165 * isolate_lru_page() or drop the page ref if it was 2166 * not isolated. 2167 */ 2168 put_page(page); 2169 out: 2170 mmap_read_unlock(mm); 2171 return err; 2172 } 2173 2174 static int move_pages_and_store_status(struct mm_struct *mm, int node, 2175 struct list_head *pagelist, int __user *status, 2176 int start, int i, unsigned long nr_pages) 2177 { 2178 int err; 2179 2180 if (list_empty(pagelist)) 2181 return 0; 2182 2183 err = do_move_pages_to_node(mm, pagelist, node); 2184 if (err) { 2185 /* 2186 * Positive err means the number of failed 2187 * pages to migrate. Since we are going to 2188 * abort and return the number of non-migrated 2189 * pages, so need to include the rest of the 2190 * nr_pages that have not been attempted as 2191 * well. 2192 */ 2193 if (err > 0) 2194 err += nr_pages - i; 2195 return err; 2196 } 2197 return store_status(status, start, node, i - start); 2198 } 2199 2200 /* 2201 * Migrate an array of page address onto an array of nodes and fill 2202 * the corresponding array of status. 2203 */ 2204 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, 2205 unsigned long nr_pages, 2206 const void __user * __user *pages, 2207 const int __user *nodes, 2208 int __user *status, int flags) 2209 { 2210 int current_node = NUMA_NO_NODE; 2211 LIST_HEAD(pagelist); 2212 int start, i; 2213 int err = 0, err1; 2214 2215 lru_cache_disable(); 2216 2217 for (i = start = 0; i < nr_pages; i++) { 2218 const void __user *p; 2219 int node; 2220 2221 err = -EFAULT; 2222 if (get_user(p, pages + i)) 2223 goto out_flush; 2224 if (get_user(node, nodes + i)) 2225 goto out_flush; 2226 2227 err = -ENODEV; 2228 if (node < 0 || node >= MAX_NUMNODES) 2229 goto out_flush; 2230 if (!node_state(node, N_MEMORY)) 2231 goto out_flush; 2232 2233 err = -EACCES; 2234 if (!node_isset(node, task_nodes)) 2235 goto out_flush; 2236 2237 if (current_node == NUMA_NO_NODE) { 2238 current_node = node; 2239 start = i; 2240 } else if (node != current_node) { 2241 err = move_pages_and_store_status(mm, current_node, 2242 &pagelist, status, start, i, nr_pages); 2243 if (err) 2244 goto out; 2245 start = i; 2246 current_node = node; 2247 } 2248 2249 /* 2250 * Errors in the page lookup or isolation are not fatal and we simply 2251 * report them via status 2252 */ 2253 err = add_page_for_migration(mm, p, current_node, &pagelist, 2254 flags & MPOL_MF_MOVE_ALL); 2255 2256 if (err > 0) { 2257 /* The page is successfully queued for migration */ 2258 continue; 2259 } 2260 2261 /* 2262 * The move_pages() man page does not have an -EEXIST choice, so 2263 * use -EFAULT instead. 2264 */ 2265 if (err == -EEXIST) 2266 err = -EFAULT; 2267 2268 /* 2269 * If the page is already on the target node (!err), store the 2270 * node, otherwise, store the err. 2271 */ 2272 err = store_status(status, i, err ? : current_node, 1); 2273 if (err) 2274 goto out_flush; 2275 2276 err = move_pages_and_store_status(mm, current_node, &pagelist, 2277 status, start, i, nr_pages); 2278 if (err) { 2279 /* We have accounted for page i */ 2280 if (err > 0) 2281 err--; 2282 goto out; 2283 } 2284 current_node = NUMA_NO_NODE; 2285 } 2286 out_flush: 2287 /* Make sure we do not overwrite the existing error */ 2288 err1 = move_pages_and_store_status(mm, current_node, &pagelist, 2289 status, start, i, nr_pages); 2290 if (err >= 0) 2291 err = err1; 2292 out: 2293 lru_cache_enable(); 2294 return err; 2295 } 2296 2297 /* 2298 * Determine the nodes of an array of pages and store it in an array of status. 2299 */ 2300 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, 2301 const void __user **pages, int *status) 2302 { 2303 unsigned long i; 2304 2305 mmap_read_lock(mm); 2306 2307 for (i = 0; i < nr_pages; i++) { 2308 unsigned long addr = (unsigned long)(*pages); 2309 struct vm_area_struct *vma; 2310 struct page *page; 2311 int err = -EFAULT; 2312 2313 vma = vma_lookup(mm, addr); 2314 if (!vma) 2315 goto set_status; 2316 2317 /* FOLL_DUMP to ignore special (like zero) pages */ 2318 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 2319 2320 err = PTR_ERR(page); 2321 if (IS_ERR(page)) 2322 goto set_status; 2323 2324 err = -ENOENT; 2325 if (!page) 2326 goto set_status; 2327 2328 if (!is_zone_device_page(page)) 2329 err = page_to_nid(page); 2330 2331 put_page(page); 2332 set_status: 2333 *status = err; 2334 2335 pages++; 2336 status++; 2337 } 2338 2339 mmap_read_unlock(mm); 2340 } 2341 2342 static int get_compat_pages_array(const void __user *chunk_pages[], 2343 const void __user * __user *pages, 2344 unsigned long chunk_nr) 2345 { 2346 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages; 2347 compat_uptr_t p; 2348 int i; 2349 2350 for (i = 0; i < chunk_nr; i++) { 2351 if (get_user(p, pages32 + i)) 2352 return -EFAULT; 2353 chunk_pages[i] = compat_ptr(p); 2354 } 2355 2356 return 0; 2357 } 2358 2359 /* 2360 * Determine the nodes of a user array of pages and store it in 2361 * a user array of status. 2362 */ 2363 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, 2364 const void __user * __user *pages, 2365 int __user *status) 2366 { 2367 #define DO_PAGES_STAT_CHUNK_NR 16UL 2368 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; 2369 int chunk_status[DO_PAGES_STAT_CHUNK_NR]; 2370 2371 while (nr_pages) { 2372 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR); 2373 2374 if (in_compat_syscall()) { 2375 if (get_compat_pages_array(chunk_pages, pages, 2376 chunk_nr)) 2377 break; 2378 } else { 2379 if (copy_from_user(chunk_pages, pages, 2380 chunk_nr * sizeof(*chunk_pages))) 2381 break; 2382 } 2383 2384 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); 2385 2386 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) 2387 break; 2388 2389 pages += chunk_nr; 2390 status += chunk_nr; 2391 nr_pages -= chunk_nr; 2392 } 2393 return nr_pages ? -EFAULT : 0; 2394 } 2395 2396 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) 2397 { 2398 struct task_struct *task; 2399 struct mm_struct *mm; 2400 2401 /* 2402 * There is no need to check if current process has the right to modify 2403 * the specified process when they are same. 2404 */ 2405 if (!pid) { 2406 mmget(current->mm); 2407 *mem_nodes = cpuset_mems_allowed(current); 2408 return current->mm; 2409 } 2410 2411 /* Find the mm_struct */ 2412 rcu_read_lock(); 2413 task = find_task_by_vpid(pid); 2414 if (!task) { 2415 rcu_read_unlock(); 2416 return ERR_PTR(-ESRCH); 2417 } 2418 get_task_struct(task); 2419 2420 /* 2421 * Check if this process has the right to modify the specified 2422 * process. Use the regular "ptrace_may_access()" checks. 2423 */ 2424 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { 2425 rcu_read_unlock(); 2426 mm = ERR_PTR(-EPERM); 2427 goto out; 2428 } 2429 rcu_read_unlock(); 2430 2431 mm = ERR_PTR(security_task_movememory(task)); 2432 if (IS_ERR(mm)) 2433 goto out; 2434 *mem_nodes = cpuset_mems_allowed(task); 2435 mm = get_task_mm(task); 2436 out: 2437 put_task_struct(task); 2438 if (!mm) 2439 mm = ERR_PTR(-EINVAL); 2440 return mm; 2441 } 2442 2443 /* 2444 * Move a list of pages in the address space of the currently executing 2445 * process. 2446 */ 2447 static int kernel_move_pages(pid_t pid, unsigned long nr_pages, 2448 const void __user * __user *pages, 2449 const int __user *nodes, 2450 int __user *status, int flags) 2451 { 2452 struct mm_struct *mm; 2453 int err; 2454 nodemask_t task_nodes; 2455 2456 /* Check flags */ 2457 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) 2458 return -EINVAL; 2459 2460 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) 2461 return -EPERM; 2462 2463 mm = find_mm_struct(pid, &task_nodes); 2464 if (IS_ERR(mm)) 2465 return PTR_ERR(mm); 2466 2467 if (nodes) 2468 err = do_pages_move(mm, task_nodes, nr_pages, pages, 2469 nodes, status, flags); 2470 else 2471 err = do_pages_stat(mm, nr_pages, pages, status); 2472 2473 mmput(mm); 2474 return err; 2475 } 2476 2477 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, 2478 const void __user * __user *, pages, 2479 const int __user *, nodes, 2480 int __user *, status, int, flags) 2481 { 2482 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); 2483 } 2484 2485 #ifdef CONFIG_NUMA_BALANCING 2486 /* 2487 * Returns true if this is a safe migration target node for misplaced NUMA 2488 * pages. Currently it only checks the watermarks which is crude. 2489 */ 2490 static bool migrate_balanced_pgdat(struct pglist_data *pgdat, 2491 unsigned long nr_migrate_pages) 2492 { 2493 int z; 2494 2495 for (z = pgdat->nr_zones - 1; z >= 0; z--) { 2496 struct zone *zone = pgdat->node_zones + z; 2497 2498 if (!managed_zone(zone)) 2499 continue; 2500 2501 /* Avoid waking kswapd by allocating pages_to_migrate pages. */ 2502 if (!zone_watermark_ok(zone, 0, 2503 high_wmark_pages(zone) + 2504 nr_migrate_pages, 2505 ZONE_MOVABLE, 0)) 2506 continue; 2507 return true; 2508 } 2509 return false; 2510 } 2511 2512 static struct page *alloc_misplaced_dst_page(struct page *page, 2513 unsigned long data) 2514 { 2515 int nid = (int) data; 2516 int order = compound_order(page); 2517 gfp_t gfp = __GFP_THISNODE; 2518 struct folio *new; 2519 2520 if (order > 0) 2521 gfp |= GFP_TRANSHUGE_LIGHT; 2522 else { 2523 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY | 2524 __GFP_NOWARN; 2525 gfp &= ~__GFP_RECLAIM; 2526 } 2527 new = __folio_alloc_node(gfp, order, nid); 2528 2529 return &new->page; 2530 } 2531 2532 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) 2533 { 2534 int nr_pages = thp_nr_pages(page); 2535 int order = compound_order(page); 2536 2537 VM_BUG_ON_PAGE(order && !PageTransHuge(page), page); 2538 2539 /* Do not migrate THP mapped by multiple processes */ 2540 if (PageTransHuge(page) && total_mapcount(page) > 1) 2541 return 0; 2542 2543 /* Avoid migrating to a node that is nearly full */ 2544 if (!migrate_balanced_pgdat(pgdat, nr_pages)) { 2545 int z; 2546 2547 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)) 2548 return 0; 2549 for (z = pgdat->nr_zones - 1; z >= 0; z--) { 2550 if (managed_zone(pgdat->node_zones + z)) 2551 break; 2552 } 2553 wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE); 2554 return 0; 2555 } 2556 2557 if (!isolate_lru_page(page)) 2558 return 0; 2559 2560 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page), 2561 nr_pages); 2562 2563 /* 2564 * Isolating the page has taken another reference, so the 2565 * caller's reference can be safely dropped without the page 2566 * disappearing underneath us during migration. 2567 */ 2568 put_page(page); 2569 return 1; 2570 } 2571 2572 /* 2573 * Attempt to migrate a misplaced page to the specified destination 2574 * node. Caller is expected to have an elevated reference count on 2575 * the page that will be dropped by this function before returning. 2576 */ 2577 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, 2578 int node) 2579 { 2580 pg_data_t *pgdat = NODE_DATA(node); 2581 int isolated; 2582 int nr_remaining; 2583 unsigned int nr_succeeded; 2584 LIST_HEAD(migratepages); 2585 int nr_pages = thp_nr_pages(page); 2586 2587 /* 2588 * Don't migrate file pages that are mapped in multiple processes 2589 * with execute permissions as they are probably shared libraries. 2590 */ 2591 if (page_mapcount(page) != 1 && page_is_file_lru(page) && 2592 (vma->vm_flags & VM_EXEC)) 2593 goto out; 2594 2595 /* 2596 * Also do not migrate dirty pages as not all filesystems can move 2597 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. 2598 */ 2599 if (page_is_file_lru(page) && PageDirty(page)) 2600 goto out; 2601 2602 isolated = numamigrate_isolate_page(pgdat, page); 2603 if (!isolated) 2604 goto out; 2605 2606 list_add(&page->lru, &migratepages); 2607 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, 2608 NULL, node, MIGRATE_ASYNC, 2609 MR_NUMA_MISPLACED, &nr_succeeded); 2610 if (nr_remaining) { 2611 if (!list_empty(&migratepages)) { 2612 list_del(&page->lru); 2613 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + 2614 page_is_file_lru(page), -nr_pages); 2615 putback_lru_page(page); 2616 } 2617 isolated = 0; 2618 } 2619 if (nr_succeeded) { 2620 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded); 2621 if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node)) 2622 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS, 2623 nr_succeeded); 2624 } 2625 BUG_ON(!list_empty(&migratepages)); 2626 return isolated; 2627 2628 out: 2629 put_page(page); 2630 return 0; 2631 } 2632 #endif /* CONFIG_NUMA_BALANCING */ 2633 #endif /* CONFIG_NUMA */ 2634