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 struct buffer_head *failed_bh; 696 697 do { 698 if (!trylock_buffer(bh)) { 699 if (mode == MIGRATE_ASYNC) 700 goto unlock; 701 if (mode == MIGRATE_SYNC_LIGHT && !buffer_uptodate(bh)) 702 goto unlock; 703 lock_buffer(bh); 704 } 705 706 bh = bh->b_this_page; 707 } while (bh != head); 708 709 return true; 710 711 unlock: 712 /* We failed to lock the buffer and cannot stall. */ 713 failed_bh = bh; 714 bh = head; 715 while (bh != failed_bh) { 716 unlock_buffer(bh); 717 bh = bh->b_this_page; 718 } 719 720 return false; 721 } 722 723 static int __buffer_migrate_folio(struct address_space *mapping, 724 struct folio *dst, struct folio *src, enum migrate_mode mode, 725 bool check_refs) 726 { 727 struct buffer_head *bh, *head; 728 int rc; 729 int expected_count; 730 731 head = folio_buffers(src); 732 if (!head) 733 return migrate_folio(mapping, dst, src, mode); 734 735 /* Check whether page does not have extra refs before we do more work */ 736 expected_count = folio_expected_refs(mapping, src); 737 if (folio_ref_count(src) != expected_count) 738 return -EAGAIN; 739 740 if (!buffer_migrate_lock_buffers(head, mode)) 741 return -EAGAIN; 742 743 if (check_refs) { 744 bool busy; 745 bool invalidated = false; 746 747 recheck_buffers: 748 busy = false; 749 spin_lock(&mapping->private_lock); 750 bh = head; 751 do { 752 if (atomic_read(&bh->b_count)) { 753 busy = true; 754 break; 755 } 756 bh = bh->b_this_page; 757 } while (bh != head); 758 if (busy) { 759 if (invalidated) { 760 rc = -EAGAIN; 761 goto unlock_buffers; 762 } 763 spin_unlock(&mapping->private_lock); 764 invalidate_bh_lrus(); 765 invalidated = true; 766 goto recheck_buffers; 767 } 768 } 769 770 rc = folio_migrate_mapping(mapping, dst, src, 0); 771 if (rc != MIGRATEPAGE_SUCCESS) 772 goto unlock_buffers; 773 774 folio_attach_private(dst, folio_detach_private(src)); 775 776 bh = head; 777 do { 778 set_bh_page(bh, &dst->page, bh_offset(bh)); 779 bh = bh->b_this_page; 780 } while (bh != head); 781 782 if (mode != MIGRATE_SYNC_NO_COPY) 783 folio_migrate_copy(dst, src); 784 else 785 folio_migrate_flags(dst, src); 786 787 rc = MIGRATEPAGE_SUCCESS; 788 unlock_buffers: 789 if (check_refs) 790 spin_unlock(&mapping->private_lock); 791 bh = head; 792 do { 793 unlock_buffer(bh); 794 bh = bh->b_this_page; 795 } while (bh != head); 796 797 return rc; 798 } 799 800 /** 801 * buffer_migrate_folio() - Migration function for folios with buffers. 802 * @mapping: The address space containing @src. 803 * @dst: The folio to migrate to. 804 * @src: The folio to migrate from. 805 * @mode: How to migrate the folio. 806 * 807 * This function can only be used if the underlying filesystem guarantees 808 * that no other references to @src exist. For example attached buffer 809 * heads are accessed only under the folio lock. If your filesystem cannot 810 * provide this guarantee, buffer_migrate_folio_norefs() may be more 811 * appropriate. 812 * 813 * Return: 0 on success or a negative errno on failure. 814 */ 815 int buffer_migrate_folio(struct address_space *mapping, 816 struct folio *dst, struct folio *src, enum migrate_mode mode) 817 { 818 return __buffer_migrate_folio(mapping, dst, src, mode, false); 819 } 820 EXPORT_SYMBOL(buffer_migrate_folio); 821 822 /** 823 * buffer_migrate_folio_norefs() - Migration function for folios with buffers. 824 * @mapping: The address space containing @src. 825 * @dst: The folio to migrate to. 826 * @src: The folio to migrate from. 827 * @mode: How to migrate the folio. 828 * 829 * Like buffer_migrate_folio() except that this variant is more careful 830 * and checks that there are also no buffer head references. This function 831 * is the right one for mappings where buffer heads are directly looked 832 * up and referenced (such as block device mappings). 833 * 834 * Return: 0 on success or a negative errno on failure. 835 */ 836 int buffer_migrate_folio_norefs(struct address_space *mapping, 837 struct folio *dst, struct folio *src, enum migrate_mode mode) 838 { 839 return __buffer_migrate_folio(mapping, dst, src, mode, true); 840 } 841 EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs); 842 #endif 843 844 int filemap_migrate_folio(struct address_space *mapping, 845 struct folio *dst, struct folio *src, enum migrate_mode mode) 846 { 847 int ret; 848 849 ret = folio_migrate_mapping(mapping, dst, src, 0); 850 if (ret != MIGRATEPAGE_SUCCESS) 851 return ret; 852 853 if (folio_get_private(src)) 854 folio_attach_private(dst, folio_detach_private(src)); 855 856 if (mode != MIGRATE_SYNC_NO_COPY) 857 folio_migrate_copy(dst, src); 858 else 859 folio_migrate_flags(dst, src); 860 return MIGRATEPAGE_SUCCESS; 861 } 862 EXPORT_SYMBOL_GPL(filemap_migrate_folio); 863 864 /* 865 * Writeback a folio to clean the dirty state 866 */ 867 static int writeout(struct address_space *mapping, struct folio *folio) 868 { 869 struct writeback_control wbc = { 870 .sync_mode = WB_SYNC_NONE, 871 .nr_to_write = 1, 872 .range_start = 0, 873 .range_end = LLONG_MAX, 874 .for_reclaim = 1 875 }; 876 int rc; 877 878 if (!mapping->a_ops->writepage) 879 /* No write method for the address space */ 880 return -EINVAL; 881 882 if (!folio_clear_dirty_for_io(folio)) 883 /* Someone else already triggered a write */ 884 return -EAGAIN; 885 886 /* 887 * A dirty folio may imply that the underlying filesystem has 888 * the folio on some queue. So the folio must be clean for 889 * migration. Writeout may mean we lose the lock and the 890 * folio state is no longer what we checked for earlier. 891 * At this point we know that the migration attempt cannot 892 * be successful. 893 */ 894 remove_migration_ptes(folio, folio, false); 895 896 rc = mapping->a_ops->writepage(&folio->page, &wbc); 897 898 if (rc != AOP_WRITEPAGE_ACTIVATE) 899 /* unlocked. Relock */ 900 folio_lock(folio); 901 902 return (rc < 0) ? -EIO : -EAGAIN; 903 } 904 905 /* 906 * Default handling if a filesystem does not provide a migration function. 907 */ 908 static int fallback_migrate_folio(struct address_space *mapping, 909 struct folio *dst, struct folio *src, enum migrate_mode mode) 910 { 911 if (folio_test_dirty(src)) { 912 /* Only writeback folios in full synchronous migration */ 913 switch (mode) { 914 case MIGRATE_SYNC: 915 case MIGRATE_SYNC_NO_COPY: 916 break; 917 default: 918 return -EBUSY; 919 } 920 return writeout(mapping, src); 921 } 922 923 /* 924 * Buffers may be managed in a filesystem specific way. 925 * We must have no buffers or drop them. 926 */ 927 if (folio_test_private(src) && 928 !filemap_release_folio(src, GFP_KERNEL)) 929 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; 930 931 return migrate_folio(mapping, dst, src, mode); 932 } 933 934 /* 935 * Move a page to a newly allocated page 936 * The page is locked and all ptes have been successfully removed. 937 * 938 * The new page will have replaced the old page if this function 939 * is successful. 940 * 941 * Return value: 942 * < 0 - error code 943 * MIGRATEPAGE_SUCCESS - success 944 */ 945 static int move_to_new_folio(struct folio *dst, struct folio *src, 946 enum migrate_mode mode) 947 { 948 int rc = -EAGAIN; 949 bool is_lru = !__PageMovable(&src->page); 950 951 VM_BUG_ON_FOLIO(!folio_test_locked(src), src); 952 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst); 953 954 if (likely(is_lru)) { 955 struct address_space *mapping = folio_mapping(src); 956 957 if (!mapping) 958 rc = migrate_folio(mapping, dst, src, mode); 959 else if (mapping->a_ops->migrate_folio) 960 /* 961 * Most folios have a mapping and most filesystems 962 * provide a migrate_folio callback. Anonymous folios 963 * are part of swap space which also has its own 964 * migrate_folio callback. This is the most common path 965 * for page migration. 966 */ 967 rc = mapping->a_ops->migrate_folio(mapping, dst, src, 968 mode); 969 else 970 rc = fallback_migrate_folio(mapping, dst, src, mode); 971 } else { 972 const struct movable_operations *mops; 973 974 /* 975 * In case of non-lru page, it could be released after 976 * isolation step. In that case, we shouldn't try migration. 977 */ 978 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); 979 if (!folio_test_movable(src)) { 980 rc = MIGRATEPAGE_SUCCESS; 981 folio_clear_isolated(src); 982 goto out; 983 } 984 985 mops = folio_movable_ops(src); 986 rc = mops->migrate_page(&dst->page, &src->page, mode); 987 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && 988 !folio_test_isolated(src)); 989 } 990 991 /* 992 * When successful, old pagecache src->mapping must be cleared before 993 * src is freed; but stats require that PageAnon be left as PageAnon. 994 */ 995 if (rc == MIGRATEPAGE_SUCCESS) { 996 if (__PageMovable(&src->page)) { 997 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); 998 999 /* 1000 * We clear PG_movable under page_lock so any compactor 1001 * cannot try to migrate this page. 1002 */ 1003 folio_clear_isolated(src); 1004 } 1005 1006 /* 1007 * Anonymous and movable src->mapping will be cleared by 1008 * free_pages_prepare so don't reset it here for keeping 1009 * the type to work PageAnon, for example. 1010 */ 1011 if (!folio_mapping_flags(src)) 1012 src->mapping = NULL; 1013 1014 if (likely(!folio_is_zone_device(dst))) 1015 flush_dcache_folio(dst); 1016 } 1017 out: 1018 return rc; 1019 } 1020 1021 /* 1022 * To record some information during migration, we use some unused 1023 * fields (mapping and private) of struct folio of the newly allocated 1024 * destination folio. This is safe because nobody is using them 1025 * except us. 1026 */ 1027 union migration_ptr { 1028 struct anon_vma *anon_vma; 1029 struct address_space *mapping; 1030 }; 1031 static void __migrate_folio_record(struct folio *dst, 1032 unsigned long page_was_mapped, 1033 struct anon_vma *anon_vma) 1034 { 1035 union migration_ptr ptr = { .anon_vma = anon_vma }; 1036 dst->mapping = ptr.mapping; 1037 dst->private = (void *)page_was_mapped; 1038 } 1039 1040 static void __migrate_folio_extract(struct folio *dst, 1041 int *page_was_mappedp, 1042 struct anon_vma **anon_vmap) 1043 { 1044 union migration_ptr ptr = { .mapping = dst->mapping }; 1045 *anon_vmap = ptr.anon_vma; 1046 *page_was_mappedp = (unsigned long)dst->private; 1047 dst->mapping = NULL; 1048 dst->private = NULL; 1049 } 1050 1051 /* Restore the source folio to the original state upon failure */ 1052 static void migrate_folio_undo_src(struct folio *src, 1053 int page_was_mapped, 1054 struct anon_vma *anon_vma, 1055 bool locked, 1056 struct list_head *ret) 1057 { 1058 if (page_was_mapped) 1059 remove_migration_ptes(src, src, false); 1060 /* Drop an anon_vma reference if we took one */ 1061 if (anon_vma) 1062 put_anon_vma(anon_vma); 1063 if (locked) 1064 folio_unlock(src); 1065 if (ret) 1066 list_move_tail(&src->lru, ret); 1067 } 1068 1069 /* Restore the destination folio to the original state upon failure */ 1070 static void migrate_folio_undo_dst(struct folio *dst, bool locked, 1071 free_folio_t put_new_folio, unsigned long private) 1072 { 1073 if (locked) 1074 folio_unlock(dst); 1075 if (put_new_folio) 1076 put_new_folio(dst, private); 1077 else 1078 folio_put(dst); 1079 } 1080 1081 /* Cleanup src folio upon migration success */ 1082 static void migrate_folio_done(struct folio *src, 1083 enum migrate_reason reason) 1084 { 1085 /* 1086 * Compaction can migrate also non-LRU pages which are 1087 * not accounted to NR_ISOLATED_*. They can be recognized 1088 * as __PageMovable 1089 */ 1090 if (likely(!__folio_test_movable(src))) 1091 mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON + 1092 folio_is_file_lru(src), -folio_nr_pages(src)); 1093 1094 if (reason != MR_MEMORY_FAILURE) 1095 /* We release the page in page_handle_poison. */ 1096 folio_put(src); 1097 } 1098 1099 /* Obtain the lock on page, remove all ptes. */ 1100 static int migrate_folio_unmap(new_folio_t get_new_folio, 1101 free_folio_t put_new_folio, unsigned long private, 1102 struct folio *src, struct folio **dstp, enum migrate_mode mode, 1103 enum migrate_reason reason, struct list_head *ret) 1104 { 1105 struct folio *dst; 1106 int rc = -EAGAIN; 1107 int page_was_mapped = 0; 1108 struct anon_vma *anon_vma = NULL; 1109 bool is_lru = !__PageMovable(&src->page); 1110 bool locked = false; 1111 bool dst_locked = false; 1112 1113 if (folio_ref_count(src) == 1) { 1114 /* Folio was freed from under us. So we are done. */ 1115 folio_clear_active(src); 1116 folio_clear_unevictable(src); 1117 /* free_pages_prepare() will clear PG_isolated. */ 1118 list_del(&src->lru); 1119 migrate_folio_done(src, reason); 1120 return MIGRATEPAGE_SUCCESS; 1121 } 1122 1123 dst = get_new_folio(src, private); 1124 if (!dst) 1125 return -ENOMEM; 1126 *dstp = dst; 1127 1128 dst->private = NULL; 1129 1130 if (!folio_trylock(src)) { 1131 if (mode == MIGRATE_ASYNC) 1132 goto out; 1133 1134 /* 1135 * It's not safe for direct compaction to call lock_page. 1136 * For example, during page readahead pages are added locked 1137 * to the LRU. Later, when the IO completes the pages are 1138 * marked uptodate and unlocked. However, the queueing 1139 * could be merging multiple pages for one bio (e.g. 1140 * mpage_readahead). If an allocation happens for the 1141 * second or third page, the process can end up locking 1142 * the same page twice and deadlocking. Rather than 1143 * trying to be clever about what pages can be locked, 1144 * avoid the use of lock_page for direct compaction 1145 * altogether. 1146 */ 1147 if (current->flags & PF_MEMALLOC) 1148 goto out; 1149 1150 /* 1151 * In "light" mode, we can wait for transient locks (eg 1152 * inserting a page into the page table), but it's not 1153 * worth waiting for I/O. 1154 */ 1155 if (mode == MIGRATE_SYNC_LIGHT && !folio_test_uptodate(src)) 1156 goto out; 1157 1158 folio_lock(src); 1159 } 1160 locked = true; 1161 1162 if (folio_test_writeback(src)) { 1163 /* 1164 * Only in the case of a full synchronous migration is it 1165 * necessary to wait for PageWriteback. In the async case, 1166 * the retry loop is too short and in the sync-light case, 1167 * the overhead of stalling is too much 1168 */ 1169 switch (mode) { 1170 case MIGRATE_SYNC: 1171 case MIGRATE_SYNC_NO_COPY: 1172 break; 1173 default: 1174 rc = -EBUSY; 1175 goto out; 1176 } 1177 folio_wait_writeback(src); 1178 } 1179 1180 /* 1181 * By try_to_migrate(), src->mapcount goes down to 0 here. In this case, 1182 * we cannot notice that anon_vma is freed while we migrate a page. 1183 * This get_anon_vma() delays freeing anon_vma pointer until the end 1184 * of migration. File cache pages are no problem because of page_lock() 1185 * File Caches may use write_page() or lock_page() in migration, then, 1186 * just care Anon page here. 1187 * 1188 * Only folio_get_anon_vma() understands the subtleties of 1189 * getting a hold on an anon_vma from outside one of its mms. 1190 * But if we cannot get anon_vma, then we won't need it anyway, 1191 * because that implies that the anon page is no longer mapped 1192 * (and cannot be remapped so long as we hold the page lock). 1193 */ 1194 if (folio_test_anon(src) && !folio_test_ksm(src)) 1195 anon_vma = folio_get_anon_vma(src); 1196 1197 /* 1198 * Block others from accessing the new page when we get around to 1199 * establishing additional references. We are usually the only one 1200 * holding a reference to dst at this point. We used to have a BUG 1201 * here if folio_trylock(dst) fails, but would like to allow for 1202 * cases where there might be a race with the previous use of dst. 1203 * This is much like races on refcount of oldpage: just don't BUG(). 1204 */ 1205 if (unlikely(!folio_trylock(dst))) 1206 goto out; 1207 dst_locked = true; 1208 1209 if (unlikely(!is_lru)) { 1210 __migrate_folio_record(dst, page_was_mapped, anon_vma); 1211 return MIGRATEPAGE_UNMAP; 1212 } 1213 1214 /* 1215 * Corner case handling: 1216 * 1. When a new swap-cache page is read into, it is added to the LRU 1217 * and treated as swapcache but it has no rmap yet. 1218 * Calling try_to_unmap() against a src->mapping==NULL page will 1219 * trigger a BUG. So handle it here. 1220 * 2. An orphaned page (see truncate_cleanup_page) might have 1221 * fs-private metadata. The page can be picked up due to memory 1222 * offlining. Everywhere else except page reclaim, the page is 1223 * invisible to the vm, so the page can not be migrated. So try to 1224 * free the metadata, so the page can be freed. 1225 */ 1226 if (!src->mapping) { 1227 if (folio_test_private(src)) { 1228 try_to_free_buffers(src); 1229 goto out; 1230 } 1231 } else if (folio_mapped(src)) { 1232 /* Establish migration ptes */ 1233 VM_BUG_ON_FOLIO(folio_test_anon(src) && 1234 !folio_test_ksm(src) && !anon_vma, src); 1235 try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0); 1236 page_was_mapped = 1; 1237 } 1238 1239 if (!folio_mapped(src)) { 1240 __migrate_folio_record(dst, page_was_mapped, anon_vma); 1241 return MIGRATEPAGE_UNMAP; 1242 } 1243 1244 out: 1245 /* 1246 * A folio that has not been unmapped will be restored to 1247 * right list unless we want to retry. 1248 */ 1249 if (rc == -EAGAIN) 1250 ret = NULL; 1251 1252 migrate_folio_undo_src(src, page_was_mapped, anon_vma, locked, ret); 1253 migrate_folio_undo_dst(dst, dst_locked, put_new_folio, private); 1254 1255 return rc; 1256 } 1257 1258 /* Migrate the folio to the newly allocated folio in dst. */ 1259 static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private, 1260 struct folio *src, struct folio *dst, 1261 enum migrate_mode mode, enum migrate_reason reason, 1262 struct list_head *ret) 1263 { 1264 int rc; 1265 int page_was_mapped = 0; 1266 struct anon_vma *anon_vma = NULL; 1267 bool is_lru = !__PageMovable(&src->page); 1268 struct list_head *prev; 1269 1270 __migrate_folio_extract(dst, &page_was_mapped, &anon_vma); 1271 prev = dst->lru.prev; 1272 list_del(&dst->lru); 1273 1274 rc = move_to_new_folio(dst, src, mode); 1275 if (rc) 1276 goto out; 1277 1278 if (unlikely(!is_lru)) 1279 goto out_unlock_both; 1280 1281 /* 1282 * When successful, push dst to LRU immediately: so that if it 1283 * turns out to be an mlocked page, remove_migration_ptes() will 1284 * automatically build up the correct dst->mlock_count for it. 1285 * 1286 * We would like to do something similar for the old page, when 1287 * unsuccessful, and other cases when a page has been temporarily 1288 * isolated from the unevictable LRU: but this case is the easiest. 1289 */ 1290 folio_add_lru(dst); 1291 if (page_was_mapped) 1292 lru_add_drain(); 1293 1294 if (page_was_mapped) 1295 remove_migration_ptes(src, dst, false); 1296 1297 out_unlock_both: 1298 folio_unlock(dst); 1299 set_page_owner_migrate_reason(&dst->page, reason); 1300 /* 1301 * If migration is successful, decrease refcount of dst, 1302 * which will not free the page because new page owner increased 1303 * refcounter. 1304 */ 1305 folio_put(dst); 1306 1307 /* 1308 * A folio that has been migrated has all references removed 1309 * and will be freed. 1310 */ 1311 list_del(&src->lru); 1312 /* Drop an anon_vma reference if we took one */ 1313 if (anon_vma) 1314 put_anon_vma(anon_vma); 1315 folio_unlock(src); 1316 migrate_folio_done(src, reason); 1317 1318 return rc; 1319 out: 1320 /* 1321 * A folio that has not been migrated will be restored to 1322 * right list unless we want to retry. 1323 */ 1324 if (rc == -EAGAIN) { 1325 list_add(&dst->lru, prev); 1326 __migrate_folio_record(dst, page_was_mapped, anon_vma); 1327 return rc; 1328 } 1329 1330 migrate_folio_undo_src(src, page_was_mapped, anon_vma, true, ret); 1331 migrate_folio_undo_dst(dst, true, put_new_folio, private); 1332 1333 return rc; 1334 } 1335 1336 /* 1337 * Counterpart of unmap_and_move_page() for hugepage migration. 1338 * 1339 * This function doesn't wait the completion of hugepage I/O 1340 * because there is no race between I/O and migration for hugepage. 1341 * Note that currently hugepage I/O occurs only in direct I/O 1342 * where no lock is held and PG_writeback is irrelevant, 1343 * and writeback status of all subpages are counted in the reference 1344 * count of the head page (i.e. if all subpages of a 2MB hugepage are 1345 * under direct I/O, the reference of the head page is 512 and a bit more.) 1346 * This means that when we try to migrate hugepage whose subpages are 1347 * doing direct I/O, some references remain after try_to_unmap() and 1348 * hugepage migration fails without data corruption. 1349 * 1350 * There is also no race when direct I/O is issued on the page under migration, 1351 * because then pte is replaced with migration swap entry and direct I/O code 1352 * will wait in the page fault for migration to complete. 1353 */ 1354 static int unmap_and_move_huge_page(new_folio_t get_new_folio, 1355 free_folio_t put_new_folio, unsigned long private, 1356 struct folio *src, int force, enum migrate_mode mode, 1357 int reason, struct list_head *ret) 1358 { 1359 struct folio *dst; 1360 int rc = -EAGAIN; 1361 int page_was_mapped = 0; 1362 struct anon_vma *anon_vma = NULL; 1363 struct address_space *mapping = NULL; 1364 1365 if (folio_ref_count(src) == 1) { 1366 /* page was freed from under us. So we are done. */ 1367 folio_putback_active_hugetlb(src); 1368 return MIGRATEPAGE_SUCCESS; 1369 } 1370 1371 dst = get_new_folio(src, private); 1372 if (!dst) 1373 return -ENOMEM; 1374 1375 if (!folio_trylock(src)) { 1376 if (!force) 1377 goto out; 1378 switch (mode) { 1379 case MIGRATE_SYNC: 1380 case MIGRATE_SYNC_NO_COPY: 1381 break; 1382 default: 1383 goto out; 1384 } 1385 folio_lock(src); 1386 } 1387 1388 /* 1389 * Check for pages which are in the process of being freed. Without 1390 * folio_mapping() set, hugetlbfs specific move page routine will not 1391 * be called and we could leak usage counts for subpools. 1392 */ 1393 if (hugetlb_folio_subpool(src) && !folio_mapping(src)) { 1394 rc = -EBUSY; 1395 goto out_unlock; 1396 } 1397 1398 if (folio_test_anon(src)) 1399 anon_vma = folio_get_anon_vma(src); 1400 1401 if (unlikely(!folio_trylock(dst))) 1402 goto put_anon; 1403 1404 if (folio_mapped(src)) { 1405 enum ttu_flags ttu = 0; 1406 1407 if (!folio_test_anon(src)) { 1408 /* 1409 * In shared mappings, try_to_unmap could potentially 1410 * call huge_pmd_unshare. Because of this, take 1411 * semaphore in write mode here and set TTU_RMAP_LOCKED 1412 * to let lower levels know we have taken the lock. 1413 */ 1414 mapping = hugetlb_page_mapping_lock_write(&src->page); 1415 if (unlikely(!mapping)) 1416 goto unlock_put_anon; 1417 1418 ttu = TTU_RMAP_LOCKED; 1419 } 1420 1421 try_to_migrate(src, ttu); 1422 page_was_mapped = 1; 1423 1424 if (ttu & TTU_RMAP_LOCKED) 1425 i_mmap_unlock_write(mapping); 1426 } 1427 1428 if (!folio_mapped(src)) 1429 rc = move_to_new_folio(dst, src, mode); 1430 1431 if (page_was_mapped) 1432 remove_migration_ptes(src, 1433 rc == MIGRATEPAGE_SUCCESS ? dst : src, false); 1434 1435 unlock_put_anon: 1436 folio_unlock(dst); 1437 1438 put_anon: 1439 if (anon_vma) 1440 put_anon_vma(anon_vma); 1441 1442 if (rc == MIGRATEPAGE_SUCCESS) { 1443 move_hugetlb_state(src, dst, reason); 1444 put_new_folio = NULL; 1445 } 1446 1447 out_unlock: 1448 folio_unlock(src); 1449 out: 1450 if (rc == MIGRATEPAGE_SUCCESS) 1451 folio_putback_active_hugetlb(src); 1452 else if (rc != -EAGAIN) 1453 list_move_tail(&src->lru, ret); 1454 1455 /* 1456 * If migration was not successful and there's a freeing callback, use 1457 * it. Otherwise, put_page() will drop the reference grabbed during 1458 * isolation. 1459 */ 1460 if (put_new_folio) 1461 put_new_folio(dst, private); 1462 else 1463 folio_putback_active_hugetlb(dst); 1464 1465 return rc; 1466 } 1467 1468 static inline int try_split_folio(struct folio *folio, struct list_head *split_folios) 1469 { 1470 int rc; 1471 1472 folio_lock(folio); 1473 rc = split_folio_to_list(folio, split_folios); 1474 folio_unlock(folio); 1475 if (!rc) 1476 list_move_tail(&folio->lru, split_folios); 1477 1478 return rc; 1479 } 1480 1481 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 1482 #define NR_MAX_BATCHED_MIGRATION HPAGE_PMD_NR 1483 #else 1484 #define NR_MAX_BATCHED_MIGRATION 512 1485 #endif 1486 #define NR_MAX_MIGRATE_PAGES_RETRY 10 1487 #define NR_MAX_MIGRATE_ASYNC_RETRY 3 1488 #define NR_MAX_MIGRATE_SYNC_RETRY \ 1489 (NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY) 1490 1491 struct migrate_pages_stats { 1492 int nr_succeeded; /* Normal and large folios migrated successfully, in 1493 units of base pages */ 1494 int nr_failed_pages; /* Normal and large folios failed to be migrated, in 1495 units of base pages. Untried folios aren't counted */ 1496 int nr_thp_succeeded; /* THP migrated successfully */ 1497 int nr_thp_failed; /* THP failed to be migrated */ 1498 int nr_thp_split; /* THP split before migrating */ 1499 }; 1500 1501 /* 1502 * Returns the number of hugetlb folios that were not migrated, or an error code 1503 * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable 1504 * any more because the list has become empty or no retryable hugetlb folios 1505 * exist any more. It is caller's responsibility to call putback_movable_pages() 1506 * only if ret != 0. 1507 */ 1508 static int migrate_hugetlbs(struct list_head *from, new_folio_t get_new_folio, 1509 free_folio_t put_new_folio, unsigned long private, 1510 enum migrate_mode mode, int reason, 1511 struct migrate_pages_stats *stats, 1512 struct list_head *ret_folios) 1513 { 1514 int retry = 1; 1515 int nr_failed = 0; 1516 int nr_retry_pages = 0; 1517 int pass = 0; 1518 struct folio *folio, *folio2; 1519 int rc, nr_pages; 1520 1521 for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) { 1522 retry = 0; 1523 nr_retry_pages = 0; 1524 1525 list_for_each_entry_safe(folio, folio2, from, lru) { 1526 if (!folio_test_hugetlb(folio)) 1527 continue; 1528 1529 nr_pages = folio_nr_pages(folio); 1530 1531 cond_resched(); 1532 1533 /* 1534 * Migratability of hugepages depends on architectures and 1535 * their size. This check is necessary because some callers 1536 * of hugepage migration like soft offline and memory 1537 * hotremove don't walk through page tables or check whether 1538 * the hugepage is pmd-based or not before kicking migration. 1539 */ 1540 if (!hugepage_migration_supported(folio_hstate(folio))) { 1541 nr_failed++; 1542 stats->nr_failed_pages += nr_pages; 1543 list_move_tail(&folio->lru, ret_folios); 1544 continue; 1545 } 1546 1547 rc = unmap_and_move_huge_page(get_new_folio, 1548 put_new_folio, private, 1549 folio, pass > 2, mode, 1550 reason, ret_folios); 1551 /* 1552 * The rules are: 1553 * Success: hugetlb folio will be put back 1554 * -EAGAIN: stay on the from list 1555 * -ENOMEM: stay on the from list 1556 * Other errno: put on ret_folios list 1557 */ 1558 switch(rc) { 1559 case -ENOMEM: 1560 /* 1561 * When memory is low, don't bother to try to migrate 1562 * other folios, just exit. 1563 */ 1564 stats->nr_failed_pages += nr_pages + nr_retry_pages; 1565 return -ENOMEM; 1566 case -EAGAIN: 1567 retry++; 1568 nr_retry_pages += nr_pages; 1569 break; 1570 case MIGRATEPAGE_SUCCESS: 1571 stats->nr_succeeded += nr_pages; 1572 break; 1573 default: 1574 /* 1575 * Permanent failure (-EBUSY, etc.): 1576 * unlike -EAGAIN case, the failed folio is 1577 * removed from migration folio list and not 1578 * retried in the next outer loop. 1579 */ 1580 nr_failed++; 1581 stats->nr_failed_pages += nr_pages; 1582 break; 1583 } 1584 } 1585 } 1586 /* 1587 * nr_failed is number of hugetlb folios failed to be migrated. After 1588 * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb 1589 * folios as failed. 1590 */ 1591 nr_failed += retry; 1592 stats->nr_failed_pages += nr_retry_pages; 1593 1594 return nr_failed; 1595 } 1596 1597 /* 1598 * migrate_pages_batch() first unmaps folios in the from list as many as 1599 * possible, then move the unmapped folios. 1600 * 1601 * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a 1602 * lock or bit when we have locked more than one folio. Which may cause 1603 * deadlock (e.g., for loop device). So, if mode != MIGRATE_ASYNC, the 1604 * length of the from list must be <= 1. 1605 */ 1606 static int migrate_pages_batch(struct list_head *from, 1607 new_folio_t get_new_folio, free_folio_t put_new_folio, 1608 unsigned long private, enum migrate_mode mode, int reason, 1609 struct list_head *ret_folios, struct list_head *split_folios, 1610 struct migrate_pages_stats *stats, int nr_pass) 1611 { 1612 int retry = 1; 1613 int thp_retry = 1; 1614 int nr_failed = 0; 1615 int nr_retry_pages = 0; 1616 int pass = 0; 1617 bool is_thp = false; 1618 struct folio *folio, *folio2, *dst = NULL, *dst2; 1619 int rc, rc_saved = 0, nr_pages; 1620 LIST_HEAD(unmap_folios); 1621 LIST_HEAD(dst_folios); 1622 bool nosplit = (reason == MR_NUMA_MISPLACED); 1623 1624 VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC && 1625 !list_empty(from) && !list_is_singular(from)); 1626 1627 for (pass = 0; pass < nr_pass && retry; pass++) { 1628 retry = 0; 1629 thp_retry = 0; 1630 nr_retry_pages = 0; 1631 1632 list_for_each_entry_safe(folio, folio2, from, lru) { 1633 is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio); 1634 nr_pages = folio_nr_pages(folio); 1635 1636 cond_resched(); 1637 1638 /* 1639 * Large folio migration might be unsupported or 1640 * the allocation might be failed so we should retry 1641 * on the same folio with the large folio split 1642 * to normal folios. 1643 * 1644 * Split folios are put in split_folios, and 1645 * we will migrate them after the rest of the 1646 * list is processed. 1647 */ 1648 if (!thp_migration_supported() && is_thp) { 1649 nr_failed++; 1650 stats->nr_thp_failed++; 1651 if (!try_split_folio(folio, split_folios)) { 1652 stats->nr_thp_split++; 1653 continue; 1654 } 1655 stats->nr_failed_pages += nr_pages; 1656 list_move_tail(&folio->lru, ret_folios); 1657 continue; 1658 } 1659 1660 rc = migrate_folio_unmap(get_new_folio, put_new_folio, 1661 private, folio, &dst, mode, reason, 1662 ret_folios); 1663 /* 1664 * The rules are: 1665 * Success: folio will be freed 1666 * Unmap: folio will be put on unmap_folios list, 1667 * dst folio put on dst_folios list 1668 * -EAGAIN: stay on the from list 1669 * -ENOMEM: stay on the from list 1670 * Other errno: put on ret_folios list 1671 */ 1672 switch(rc) { 1673 case -ENOMEM: 1674 /* 1675 * When memory is low, don't bother to try to migrate 1676 * other folios, move unmapped folios, then exit. 1677 */ 1678 nr_failed++; 1679 stats->nr_thp_failed += is_thp; 1680 /* Large folio NUMA faulting doesn't split to retry. */ 1681 if (folio_test_large(folio) && !nosplit) { 1682 int ret = try_split_folio(folio, split_folios); 1683 1684 if (!ret) { 1685 stats->nr_thp_split += is_thp; 1686 break; 1687 } else if (reason == MR_LONGTERM_PIN && 1688 ret == -EAGAIN) { 1689 /* 1690 * Try again to split large folio to 1691 * mitigate the failure of longterm pinning. 1692 */ 1693 retry++; 1694 thp_retry += is_thp; 1695 nr_retry_pages += nr_pages; 1696 /* Undo duplicated failure counting. */ 1697 nr_failed--; 1698 stats->nr_thp_failed -= is_thp; 1699 break; 1700 } 1701 } 1702 1703 stats->nr_failed_pages += nr_pages + nr_retry_pages; 1704 /* nr_failed isn't updated for not used */ 1705 stats->nr_thp_failed += thp_retry; 1706 rc_saved = rc; 1707 if (list_empty(&unmap_folios)) 1708 goto out; 1709 else 1710 goto move; 1711 case -EAGAIN: 1712 retry++; 1713 thp_retry += is_thp; 1714 nr_retry_pages += nr_pages; 1715 break; 1716 case MIGRATEPAGE_SUCCESS: 1717 stats->nr_succeeded += nr_pages; 1718 stats->nr_thp_succeeded += is_thp; 1719 break; 1720 case MIGRATEPAGE_UNMAP: 1721 list_move_tail(&folio->lru, &unmap_folios); 1722 list_add_tail(&dst->lru, &dst_folios); 1723 break; 1724 default: 1725 /* 1726 * Permanent failure (-EBUSY, etc.): 1727 * unlike -EAGAIN case, the failed folio is 1728 * removed from migration folio list and not 1729 * retried in the next outer loop. 1730 */ 1731 nr_failed++; 1732 stats->nr_thp_failed += is_thp; 1733 stats->nr_failed_pages += nr_pages; 1734 break; 1735 } 1736 } 1737 } 1738 nr_failed += retry; 1739 stats->nr_thp_failed += thp_retry; 1740 stats->nr_failed_pages += nr_retry_pages; 1741 move: 1742 /* Flush TLBs for all unmapped folios */ 1743 try_to_unmap_flush(); 1744 1745 retry = 1; 1746 for (pass = 0; pass < nr_pass && retry; pass++) { 1747 retry = 0; 1748 thp_retry = 0; 1749 nr_retry_pages = 0; 1750 1751 dst = list_first_entry(&dst_folios, struct folio, lru); 1752 dst2 = list_next_entry(dst, lru); 1753 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) { 1754 is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio); 1755 nr_pages = folio_nr_pages(folio); 1756 1757 cond_resched(); 1758 1759 rc = migrate_folio_move(put_new_folio, private, 1760 folio, dst, mode, 1761 reason, ret_folios); 1762 /* 1763 * The rules are: 1764 * Success: folio will be freed 1765 * -EAGAIN: stay on the unmap_folios list 1766 * Other errno: put on ret_folios list 1767 */ 1768 switch(rc) { 1769 case -EAGAIN: 1770 retry++; 1771 thp_retry += is_thp; 1772 nr_retry_pages += nr_pages; 1773 break; 1774 case MIGRATEPAGE_SUCCESS: 1775 stats->nr_succeeded += nr_pages; 1776 stats->nr_thp_succeeded += is_thp; 1777 break; 1778 default: 1779 nr_failed++; 1780 stats->nr_thp_failed += is_thp; 1781 stats->nr_failed_pages += nr_pages; 1782 break; 1783 } 1784 dst = dst2; 1785 dst2 = list_next_entry(dst, lru); 1786 } 1787 } 1788 nr_failed += retry; 1789 stats->nr_thp_failed += thp_retry; 1790 stats->nr_failed_pages += nr_retry_pages; 1791 1792 rc = rc_saved ? : nr_failed; 1793 out: 1794 /* Cleanup remaining folios */ 1795 dst = list_first_entry(&dst_folios, struct folio, lru); 1796 dst2 = list_next_entry(dst, lru); 1797 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) { 1798 int page_was_mapped = 0; 1799 struct anon_vma *anon_vma = NULL; 1800 1801 __migrate_folio_extract(dst, &page_was_mapped, &anon_vma); 1802 migrate_folio_undo_src(folio, page_was_mapped, anon_vma, 1803 true, ret_folios); 1804 list_del(&dst->lru); 1805 migrate_folio_undo_dst(dst, true, put_new_folio, private); 1806 dst = dst2; 1807 dst2 = list_next_entry(dst, lru); 1808 } 1809 1810 return rc; 1811 } 1812 1813 static int migrate_pages_sync(struct list_head *from, new_folio_t get_new_folio, 1814 free_folio_t put_new_folio, unsigned long private, 1815 enum migrate_mode mode, int reason, 1816 struct list_head *ret_folios, struct list_head *split_folios, 1817 struct migrate_pages_stats *stats) 1818 { 1819 int rc, nr_failed = 0; 1820 LIST_HEAD(folios); 1821 struct migrate_pages_stats astats; 1822 1823 memset(&astats, 0, sizeof(astats)); 1824 /* Try to migrate in batch with MIGRATE_ASYNC mode firstly */ 1825 rc = migrate_pages_batch(from, get_new_folio, put_new_folio, private, MIGRATE_ASYNC, 1826 reason, &folios, split_folios, &astats, 1827 NR_MAX_MIGRATE_ASYNC_RETRY); 1828 stats->nr_succeeded += astats.nr_succeeded; 1829 stats->nr_thp_succeeded += astats.nr_thp_succeeded; 1830 stats->nr_thp_split += astats.nr_thp_split; 1831 if (rc < 0) { 1832 stats->nr_failed_pages += astats.nr_failed_pages; 1833 stats->nr_thp_failed += astats.nr_thp_failed; 1834 list_splice_tail(&folios, ret_folios); 1835 return rc; 1836 } 1837 stats->nr_thp_failed += astats.nr_thp_split; 1838 nr_failed += astats.nr_thp_split; 1839 /* 1840 * Fall back to migrate all failed folios one by one synchronously. All 1841 * failed folios except split THPs will be retried, so their failure 1842 * isn't counted 1843 */ 1844 list_splice_tail_init(&folios, from); 1845 while (!list_empty(from)) { 1846 list_move(from->next, &folios); 1847 rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio, 1848 private, mode, reason, ret_folios, 1849 split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY); 1850 list_splice_tail_init(&folios, ret_folios); 1851 if (rc < 0) 1852 return rc; 1853 nr_failed += rc; 1854 } 1855 1856 return nr_failed; 1857 } 1858 1859 /* 1860 * migrate_pages - migrate the folios specified in a list, to the free folios 1861 * supplied as the target for the page migration 1862 * 1863 * @from: The list of folios to be migrated. 1864 * @get_new_folio: The function used to allocate free folios to be used 1865 * as the target of the folio migration. 1866 * @put_new_folio: The function used to free target folios if migration 1867 * fails, or NULL if no special handling is necessary. 1868 * @private: Private data to be passed on to get_new_folio() 1869 * @mode: The migration mode that specifies the constraints for 1870 * folio migration, if any. 1871 * @reason: The reason for folio migration. 1872 * @ret_succeeded: Set to the number of folios migrated successfully if 1873 * the caller passes a non-NULL pointer. 1874 * 1875 * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios 1876 * are movable any more because the list has become empty or no retryable folios 1877 * exist any more. It is caller's responsibility to call putback_movable_pages() 1878 * only if ret != 0. 1879 * 1880 * Returns the number of {normal folio, large folio, hugetlb} that were not 1881 * migrated, or an error code. The number of large folio splits will be 1882 * considered as the number of non-migrated large folio, no matter how many 1883 * split folios of the large folio are migrated successfully. 1884 */ 1885 int migrate_pages(struct list_head *from, new_folio_t get_new_folio, 1886 free_folio_t put_new_folio, unsigned long private, 1887 enum migrate_mode mode, int reason, unsigned int *ret_succeeded) 1888 { 1889 int rc, rc_gather; 1890 int nr_pages; 1891 struct folio *folio, *folio2; 1892 LIST_HEAD(folios); 1893 LIST_HEAD(ret_folios); 1894 LIST_HEAD(split_folios); 1895 struct migrate_pages_stats stats; 1896 1897 trace_mm_migrate_pages_start(mode, reason); 1898 1899 memset(&stats, 0, sizeof(stats)); 1900 1901 rc_gather = migrate_hugetlbs(from, get_new_folio, put_new_folio, private, 1902 mode, reason, &stats, &ret_folios); 1903 if (rc_gather < 0) 1904 goto out; 1905 1906 again: 1907 nr_pages = 0; 1908 list_for_each_entry_safe(folio, folio2, from, lru) { 1909 /* Retried hugetlb folios will be kept in list */ 1910 if (folio_test_hugetlb(folio)) { 1911 list_move_tail(&folio->lru, &ret_folios); 1912 continue; 1913 } 1914 1915 nr_pages += folio_nr_pages(folio); 1916 if (nr_pages >= NR_MAX_BATCHED_MIGRATION) 1917 break; 1918 } 1919 if (nr_pages >= NR_MAX_BATCHED_MIGRATION) 1920 list_cut_before(&folios, from, &folio2->lru); 1921 else 1922 list_splice_init(from, &folios); 1923 if (mode == MIGRATE_ASYNC) 1924 rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio, 1925 private, mode, reason, &ret_folios, 1926 &split_folios, &stats, 1927 NR_MAX_MIGRATE_PAGES_RETRY); 1928 else 1929 rc = migrate_pages_sync(&folios, get_new_folio, put_new_folio, 1930 private, mode, reason, &ret_folios, 1931 &split_folios, &stats); 1932 list_splice_tail_init(&folios, &ret_folios); 1933 if (rc < 0) { 1934 rc_gather = rc; 1935 list_splice_tail(&split_folios, &ret_folios); 1936 goto out; 1937 } 1938 if (!list_empty(&split_folios)) { 1939 /* 1940 * Failure isn't counted since all split folios of a large folio 1941 * is counted as 1 failure already. And, we only try to migrate 1942 * with minimal effort, force MIGRATE_ASYNC mode and retry once. 1943 */ 1944 migrate_pages_batch(&split_folios, get_new_folio, 1945 put_new_folio, private, MIGRATE_ASYNC, reason, 1946 &ret_folios, NULL, &stats, 1); 1947 list_splice_tail_init(&split_folios, &ret_folios); 1948 } 1949 rc_gather += rc; 1950 if (!list_empty(from)) 1951 goto again; 1952 out: 1953 /* 1954 * Put the permanent failure folio back to migration list, they 1955 * will be put back to the right list by the caller. 1956 */ 1957 list_splice(&ret_folios, from); 1958 1959 /* 1960 * Return 0 in case all split folios of fail-to-migrate large folios 1961 * are migrated successfully. 1962 */ 1963 if (list_empty(from)) 1964 rc_gather = 0; 1965 1966 count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded); 1967 count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages); 1968 count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded); 1969 count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed); 1970 count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split); 1971 trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages, 1972 stats.nr_thp_succeeded, stats.nr_thp_failed, 1973 stats.nr_thp_split, mode, reason); 1974 1975 if (ret_succeeded) 1976 *ret_succeeded = stats.nr_succeeded; 1977 1978 return rc_gather; 1979 } 1980 1981 struct folio *alloc_migration_target(struct folio *src, unsigned long private) 1982 { 1983 struct migration_target_control *mtc; 1984 gfp_t gfp_mask; 1985 unsigned int order = 0; 1986 int nid; 1987 int zidx; 1988 1989 mtc = (struct migration_target_control *)private; 1990 gfp_mask = mtc->gfp_mask; 1991 nid = mtc->nid; 1992 if (nid == NUMA_NO_NODE) 1993 nid = folio_nid(src); 1994 1995 if (folio_test_hugetlb(src)) { 1996 struct hstate *h = folio_hstate(src); 1997 1998 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); 1999 return alloc_hugetlb_folio_nodemask(h, nid, 2000 mtc->nmask, gfp_mask); 2001 } 2002 2003 if (folio_test_large(src)) { 2004 /* 2005 * clear __GFP_RECLAIM to make the migration callback 2006 * consistent with regular THP allocations. 2007 */ 2008 gfp_mask &= ~__GFP_RECLAIM; 2009 gfp_mask |= GFP_TRANSHUGE; 2010 order = folio_order(src); 2011 } 2012 zidx = zone_idx(folio_zone(src)); 2013 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) 2014 gfp_mask |= __GFP_HIGHMEM; 2015 2016 return __folio_alloc(gfp_mask, order, nid, mtc->nmask); 2017 } 2018 2019 #ifdef CONFIG_NUMA 2020 2021 static int store_status(int __user *status, int start, int value, int nr) 2022 { 2023 while (nr-- > 0) { 2024 if (put_user(value, status + start)) 2025 return -EFAULT; 2026 start++; 2027 } 2028 2029 return 0; 2030 } 2031 2032 static int do_move_pages_to_node(struct mm_struct *mm, 2033 struct list_head *pagelist, int node) 2034 { 2035 int err; 2036 struct migration_target_control mtc = { 2037 .nid = node, 2038 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 2039 }; 2040 2041 err = migrate_pages(pagelist, alloc_migration_target, NULL, 2042 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL); 2043 if (err) 2044 putback_movable_pages(pagelist); 2045 return err; 2046 } 2047 2048 /* 2049 * Resolves the given address to a struct page, isolates it from the LRU and 2050 * puts it to the given pagelist. 2051 * Returns: 2052 * errno - if the page cannot be found/isolated 2053 * 0 - when it doesn't have to be migrated because it is already on the 2054 * target node 2055 * 1 - when it has been queued 2056 */ 2057 static int add_page_for_migration(struct mm_struct *mm, const void __user *p, 2058 int node, struct list_head *pagelist, bool migrate_all) 2059 { 2060 struct vm_area_struct *vma; 2061 unsigned long addr; 2062 struct page *page; 2063 int err; 2064 bool isolated; 2065 2066 mmap_read_lock(mm); 2067 addr = (unsigned long)untagged_addr_remote(mm, p); 2068 2069 err = -EFAULT; 2070 vma = vma_lookup(mm, addr); 2071 if (!vma || !vma_migratable(vma)) 2072 goto out; 2073 2074 /* FOLL_DUMP to ignore special (like zero) pages */ 2075 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 2076 2077 err = PTR_ERR(page); 2078 if (IS_ERR(page)) 2079 goto out; 2080 2081 err = -ENOENT; 2082 if (!page) 2083 goto out; 2084 2085 if (is_zone_device_page(page)) 2086 goto out_putpage; 2087 2088 err = 0; 2089 if (page_to_nid(page) == node) 2090 goto out_putpage; 2091 2092 err = -EACCES; 2093 if (page_mapcount(page) > 1 && !migrate_all) 2094 goto out_putpage; 2095 2096 if (PageHuge(page)) { 2097 if (PageHead(page)) { 2098 isolated = isolate_hugetlb(page_folio(page), pagelist); 2099 err = isolated ? 1 : -EBUSY; 2100 } 2101 } else { 2102 struct page *head; 2103 2104 head = compound_head(page); 2105 isolated = isolate_lru_page(head); 2106 if (!isolated) { 2107 err = -EBUSY; 2108 goto out_putpage; 2109 } 2110 2111 err = 1; 2112 list_add_tail(&head->lru, pagelist); 2113 mod_node_page_state(page_pgdat(head), 2114 NR_ISOLATED_ANON + page_is_file_lru(head), 2115 thp_nr_pages(head)); 2116 } 2117 out_putpage: 2118 /* 2119 * Either remove the duplicate refcount from 2120 * isolate_lru_page() or drop the page ref if it was 2121 * not isolated. 2122 */ 2123 put_page(page); 2124 out: 2125 mmap_read_unlock(mm); 2126 return err; 2127 } 2128 2129 static int move_pages_and_store_status(struct mm_struct *mm, int node, 2130 struct list_head *pagelist, int __user *status, 2131 int start, int i, unsigned long nr_pages) 2132 { 2133 int err; 2134 2135 if (list_empty(pagelist)) 2136 return 0; 2137 2138 err = do_move_pages_to_node(mm, pagelist, node); 2139 if (err) { 2140 /* 2141 * Positive err means the number of failed 2142 * pages to migrate. Since we are going to 2143 * abort and return the number of non-migrated 2144 * pages, so need to include the rest of the 2145 * nr_pages that have not been attempted as 2146 * well. 2147 */ 2148 if (err > 0) 2149 err += nr_pages - i; 2150 return err; 2151 } 2152 return store_status(status, start, node, i - start); 2153 } 2154 2155 /* 2156 * Migrate an array of page address onto an array of nodes and fill 2157 * the corresponding array of status. 2158 */ 2159 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, 2160 unsigned long nr_pages, 2161 const void __user * __user *pages, 2162 const int __user *nodes, 2163 int __user *status, int flags) 2164 { 2165 int current_node = NUMA_NO_NODE; 2166 LIST_HEAD(pagelist); 2167 int start, i; 2168 int err = 0, err1; 2169 2170 lru_cache_disable(); 2171 2172 for (i = start = 0; i < nr_pages; i++) { 2173 const void __user *p; 2174 int node; 2175 2176 err = -EFAULT; 2177 if (get_user(p, pages + i)) 2178 goto out_flush; 2179 if (get_user(node, nodes + i)) 2180 goto out_flush; 2181 2182 err = -ENODEV; 2183 if (node < 0 || node >= MAX_NUMNODES) 2184 goto out_flush; 2185 if (!node_state(node, N_MEMORY)) 2186 goto out_flush; 2187 2188 err = -EACCES; 2189 if (!node_isset(node, task_nodes)) 2190 goto out_flush; 2191 2192 if (current_node == NUMA_NO_NODE) { 2193 current_node = node; 2194 start = i; 2195 } else if (node != current_node) { 2196 err = move_pages_and_store_status(mm, current_node, 2197 &pagelist, status, start, i, nr_pages); 2198 if (err) 2199 goto out; 2200 start = i; 2201 current_node = node; 2202 } 2203 2204 /* 2205 * Errors in the page lookup or isolation are not fatal and we simply 2206 * report them via status 2207 */ 2208 err = add_page_for_migration(mm, p, current_node, &pagelist, 2209 flags & MPOL_MF_MOVE_ALL); 2210 2211 if (err > 0) { 2212 /* The page is successfully queued for migration */ 2213 continue; 2214 } 2215 2216 /* 2217 * The move_pages() man page does not have an -EEXIST choice, so 2218 * use -EFAULT instead. 2219 */ 2220 if (err == -EEXIST) 2221 err = -EFAULT; 2222 2223 /* 2224 * If the page is already on the target node (!err), store the 2225 * node, otherwise, store the err. 2226 */ 2227 err = store_status(status, i, err ? : current_node, 1); 2228 if (err) 2229 goto out_flush; 2230 2231 err = move_pages_and_store_status(mm, current_node, &pagelist, 2232 status, start, i, nr_pages); 2233 if (err) { 2234 /* We have accounted for page i */ 2235 if (err > 0) 2236 err--; 2237 goto out; 2238 } 2239 current_node = NUMA_NO_NODE; 2240 } 2241 out_flush: 2242 /* Make sure we do not overwrite the existing error */ 2243 err1 = move_pages_and_store_status(mm, current_node, &pagelist, 2244 status, start, i, nr_pages); 2245 if (err >= 0) 2246 err = err1; 2247 out: 2248 lru_cache_enable(); 2249 return err; 2250 } 2251 2252 /* 2253 * Determine the nodes of an array of pages and store it in an array of status. 2254 */ 2255 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, 2256 const void __user **pages, int *status) 2257 { 2258 unsigned long i; 2259 2260 mmap_read_lock(mm); 2261 2262 for (i = 0; i < nr_pages; i++) { 2263 unsigned long addr = (unsigned long)(*pages); 2264 struct vm_area_struct *vma; 2265 struct page *page; 2266 int err = -EFAULT; 2267 2268 vma = vma_lookup(mm, addr); 2269 if (!vma) 2270 goto set_status; 2271 2272 /* FOLL_DUMP to ignore special (like zero) pages */ 2273 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 2274 2275 err = PTR_ERR(page); 2276 if (IS_ERR(page)) 2277 goto set_status; 2278 2279 err = -ENOENT; 2280 if (!page) 2281 goto set_status; 2282 2283 if (!is_zone_device_page(page)) 2284 err = page_to_nid(page); 2285 2286 put_page(page); 2287 set_status: 2288 *status = err; 2289 2290 pages++; 2291 status++; 2292 } 2293 2294 mmap_read_unlock(mm); 2295 } 2296 2297 static int get_compat_pages_array(const void __user *chunk_pages[], 2298 const void __user * __user *pages, 2299 unsigned long chunk_nr) 2300 { 2301 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages; 2302 compat_uptr_t p; 2303 int i; 2304 2305 for (i = 0; i < chunk_nr; i++) { 2306 if (get_user(p, pages32 + i)) 2307 return -EFAULT; 2308 chunk_pages[i] = compat_ptr(p); 2309 } 2310 2311 return 0; 2312 } 2313 2314 /* 2315 * Determine the nodes of a user array of pages and store it in 2316 * a user array of status. 2317 */ 2318 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, 2319 const void __user * __user *pages, 2320 int __user *status) 2321 { 2322 #define DO_PAGES_STAT_CHUNK_NR 16UL 2323 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; 2324 int chunk_status[DO_PAGES_STAT_CHUNK_NR]; 2325 2326 while (nr_pages) { 2327 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR); 2328 2329 if (in_compat_syscall()) { 2330 if (get_compat_pages_array(chunk_pages, pages, 2331 chunk_nr)) 2332 break; 2333 } else { 2334 if (copy_from_user(chunk_pages, pages, 2335 chunk_nr * sizeof(*chunk_pages))) 2336 break; 2337 } 2338 2339 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); 2340 2341 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) 2342 break; 2343 2344 pages += chunk_nr; 2345 status += chunk_nr; 2346 nr_pages -= chunk_nr; 2347 } 2348 return nr_pages ? -EFAULT : 0; 2349 } 2350 2351 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) 2352 { 2353 struct task_struct *task; 2354 struct mm_struct *mm; 2355 2356 /* 2357 * There is no need to check if current process has the right to modify 2358 * the specified process when they are same. 2359 */ 2360 if (!pid) { 2361 mmget(current->mm); 2362 *mem_nodes = cpuset_mems_allowed(current); 2363 return current->mm; 2364 } 2365 2366 /* Find the mm_struct */ 2367 rcu_read_lock(); 2368 task = find_task_by_vpid(pid); 2369 if (!task) { 2370 rcu_read_unlock(); 2371 return ERR_PTR(-ESRCH); 2372 } 2373 get_task_struct(task); 2374 2375 /* 2376 * Check if this process has the right to modify the specified 2377 * process. Use the regular "ptrace_may_access()" checks. 2378 */ 2379 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { 2380 rcu_read_unlock(); 2381 mm = ERR_PTR(-EPERM); 2382 goto out; 2383 } 2384 rcu_read_unlock(); 2385 2386 mm = ERR_PTR(security_task_movememory(task)); 2387 if (IS_ERR(mm)) 2388 goto out; 2389 *mem_nodes = cpuset_mems_allowed(task); 2390 mm = get_task_mm(task); 2391 out: 2392 put_task_struct(task); 2393 if (!mm) 2394 mm = ERR_PTR(-EINVAL); 2395 return mm; 2396 } 2397 2398 /* 2399 * Move a list of pages in the address space of the currently executing 2400 * process. 2401 */ 2402 static int kernel_move_pages(pid_t pid, unsigned long nr_pages, 2403 const void __user * __user *pages, 2404 const int __user *nodes, 2405 int __user *status, int flags) 2406 { 2407 struct mm_struct *mm; 2408 int err; 2409 nodemask_t task_nodes; 2410 2411 /* Check flags */ 2412 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) 2413 return -EINVAL; 2414 2415 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) 2416 return -EPERM; 2417 2418 mm = find_mm_struct(pid, &task_nodes); 2419 if (IS_ERR(mm)) 2420 return PTR_ERR(mm); 2421 2422 if (nodes) 2423 err = do_pages_move(mm, task_nodes, nr_pages, pages, 2424 nodes, status, flags); 2425 else 2426 err = do_pages_stat(mm, nr_pages, pages, status); 2427 2428 mmput(mm); 2429 return err; 2430 } 2431 2432 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, 2433 const void __user * __user *, pages, 2434 const int __user *, nodes, 2435 int __user *, status, int, flags) 2436 { 2437 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); 2438 } 2439 2440 #ifdef CONFIG_NUMA_BALANCING 2441 /* 2442 * Returns true if this is a safe migration target node for misplaced NUMA 2443 * pages. Currently it only checks the watermarks which is crude. 2444 */ 2445 static bool migrate_balanced_pgdat(struct pglist_data *pgdat, 2446 unsigned long nr_migrate_pages) 2447 { 2448 int z; 2449 2450 for (z = pgdat->nr_zones - 1; z >= 0; z--) { 2451 struct zone *zone = pgdat->node_zones + z; 2452 2453 if (!managed_zone(zone)) 2454 continue; 2455 2456 /* Avoid waking kswapd by allocating pages_to_migrate pages. */ 2457 if (!zone_watermark_ok(zone, 0, 2458 high_wmark_pages(zone) + 2459 nr_migrate_pages, 2460 ZONE_MOVABLE, 0)) 2461 continue; 2462 return true; 2463 } 2464 return false; 2465 } 2466 2467 static struct folio *alloc_misplaced_dst_folio(struct folio *src, 2468 unsigned long data) 2469 { 2470 int nid = (int) data; 2471 int order = folio_order(src); 2472 gfp_t gfp = __GFP_THISNODE; 2473 2474 if (order > 0) 2475 gfp |= GFP_TRANSHUGE_LIGHT; 2476 else { 2477 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY | 2478 __GFP_NOWARN; 2479 gfp &= ~__GFP_RECLAIM; 2480 } 2481 return __folio_alloc_node(gfp, order, nid); 2482 } 2483 2484 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) 2485 { 2486 int nr_pages = thp_nr_pages(page); 2487 int order = compound_order(page); 2488 2489 VM_BUG_ON_PAGE(order && !PageTransHuge(page), page); 2490 2491 /* Do not migrate THP mapped by multiple processes */ 2492 if (PageTransHuge(page) && total_mapcount(page) > 1) 2493 return 0; 2494 2495 /* Avoid migrating to a node that is nearly full */ 2496 if (!migrate_balanced_pgdat(pgdat, nr_pages)) { 2497 int z; 2498 2499 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)) 2500 return 0; 2501 for (z = pgdat->nr_zones - 1; z >= 0; z--) { 2502 if (managed_zone(pgdat->node_zones + z)) 2503 break; 2504 } 2505 wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE); 2506 return 0; 2507 } 2508 2509 if (!isolate_lru_page(page)) 2510 return 0; 2511 2512 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page), 2513 nr_pages); 2514 2515 /* 2516 * Isolating the page has taken another reference, so the 2517 * caller's reference can be safely dropped without the page 2518 * disappearing underneath us during migration. 2519 */ 2520 put_page(page); 2521 return 1; 2522 } 2523 2524 /* 2525 * Attempt to migrate a misplaced page to the specified destination 2526 * node. Caller is expected to have an elevated reference count on 2527 * the page that will be dropped by this function before returning. 2528 */ 2529 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, 2530 int node) 2531 { 2532 pg_data_t *pgdat = NODE_DATA(node); 2533 int isolated; 2534 int nr_remaining; 2535 unsigned int nr_succeeded; 2536 LIST_HEAD(migratepages); 2537 int nr_pages = thp_nr_pages(page); 2538 2539 /* 2540 * Don't migrate file pages that are mapped in multiple processes 2541 * with execute permissions as they are probably shared libraries. 2542 */ 2543 if (page_mapcount(page) != 1 && page_is_file_lru(page) && 2544 (vma->vm_flags & VM_EXEC)) 2545 goto out; 2546 2547 /* 2548 * Also do not migrate dirty pages as not all filesystems can move 2549 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. 2550 */ 2551 if (page_is_file_lru(page) && PageDirty(page)) 2552 goto out; 2553 2554 isolated = numamigrate_isolate_page(pgdat, page); 2555 if (!isolated) 2556 goto out; 2557 2558 list_add(&page->lru, &migratepages); 2559 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio, 2560 NULL, node, MIGRATE_ASYNC, 2561 MR_NUMA_MISPLACED, &nr_succeeded); 2562 if (nr_remaining) { 2563 if (!list_empty(&migratepages)) { 2564 list_del(&page->lru); 2565 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + 2566 page_is_file_lru(page), -nr_pages); 2567 putback_lru_page(page); 2568 } 2569 isolated = 0; 2570 } 2571 if (nr_succeeded) { 2572 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded); 2573 if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node)) 2574 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS, 2575 nr_succeeded); 2576 } 2577 BUG_ON(!list_empty(&migratepages)); 2578 return isolated; 2579 2580 out: 2581 put_page(page); 2582 return 0; 2583 } 2584 #endif /* CONFIG_NUMA_BALANCING */ 2585 #endif /* CONFIG_NUMA */ 2586