| 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 | |
| 54 | #include <asm/tlbflush.h> |
| 55 | |
| 56 | #include <trace/events/migrate.h> |
| 57 | |
| 58 | #include "internal.h" |
| 59 | |
| 60 | int isolate_movable_page(struct page *page, isolate_mode_t mode) |
| 61 | { |
| 62 | struct address_space *mapping; |
| 63 | |
| 64 | /* |
| 65 | * Avoid burning cycles with pages that are yet under __free_pages(), |
| 66 | * or just got freed under us. |
| 67 | * |
| 68 | * In case we 'win' a race for a movable page being freed under us and |
| 69 | * raise its refcount preventing __free_pages() from doing its job |
| 70 | * the put_page() at the end of this block will take care of |
| 71 | * release this page, thus avoiding a nasty leakage. |
| 72 | */ |
| 73 | if (unlikely(!get_page_unless_zero(page))) |
| 74 | goto out; |
| 75 | |
| 76 | /* |
| 77 | * Check PageMovable before holding a PG_lock because page's owner |
| 78 | * assumes anybody doesn't touch PG_lock of newly allocated page |
| 79 | * so unconditionally grabbing the lock ruins page's owner side. |
| 80 | */ |
| 81 | if (unlikely(!__PageMovable(page))) |
| 82 | goto out_putpage; |
| 83 | /* |
| 84 | * As movable pages are not isolated from LRU lists, concurrent |
| 85 | * compaction threads can race against page migration functions |
| 86 | * as well as race against the releasing a page. |
| 87 | * |
| 88 | * In order to avoid having an already isolated movable page |
| 89 | * being (wrongly) re-isolated while it is under migration, |
| 90 | * or to avoid attempting to isolate pages being released, |
| 91 | * lets be sure we have the page lock |
| 92 | * before proceeding with the movable page isolation steps. |
| 93 | */ |
| 94 | if (unlikely(!trylock_page(page))) |
| 95 | goto out_putpage; |
| 96 | |
| 97 | if (!PageMovable(page) || PageIsolated(page)) |
| 98 | goto out_no_isolated; |
| 99 | |
| 100 | mapping = page_mapping(page); |
| 101 | VM_BUG_ON_PAGE(!mapping, page); |
| 102 | |
| 103 | if (!mapping->a_ops->isolate_page(page, mode)) |
| 104 | goto out_no_isolated; |
| 105 | |
| 106 | /* Driver shouldn't use PG_isolated bit of page->flags */ |
| 107 | WARN_ON_ONCE(PageIsolated(page)); |
| 108 | SetPageIsolated(page); |
| 109 | unlock_page(page); |
| 110 | |
| 111 | return 0; |
| 112 | |
| 113 | out_no_isolated: |
| 114 | unlock_page(page); |
| 115 | out_putpage: |
| 116 | put_page(page); |
| 117 | out: |
| 118 | return -EBUSY; |
| 119 | } |
| 120 | |
| 121 | static void putback_movable_page(struct page *page) |
| 122 | { |
| 123 | struct address_space *mapping; |
| 124 | |
| 125 | mapping = page_mapping(page); |
| 126 | mapping->a_ops->putback_page(page); |
| 127 | ClearPageIsolated(page); |
| 128 | } |
| 129 | |
| 130 | /* |
| 131 | * Put previously isolated pages back onto the appropriate lists |
| 132 | * from where they were once taken off for compaction/migration. |
| 133 | * |
| 134 | * This function shall be used whenever the isolated pageset has been |
| 135 | * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() |
| 136 | * and isolate_huge_page(). |
| 137 | */ |
| 138 | void putback_movable_pages(struct list_head *l) |
| 139 | { |
| 140 | struct page *page; |
| 141 | struct page *page2; |
| 142 | |
| 143 | list_for_each_entry_safe(page, page2, l, lru) { |
| 144 | if (unlikely(PageHuge(page))) { |
| 145 | putback_active_hugepage(page); |
| 146 | continue; |
| 147 | } |
| 148 | list_del(&page->lru); |
| 149 | /* |
| 150 | * We isolated non-lru movable page so here we can use |
| 151 | * __PageMovable because LRU page's mapping cannot have |
| 152 | * PAGE_MAPPING_MOVABLE. |
| 153 | */ |
| 154 | if (unlikely(__PageMovable(page))) { |
| 155 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
| 156 | lock_page(page); |
| 157 | if (PageMovable(page)) |
| 158 | putback_movable_page(page); |
| 159 | else |
| 160 | ClearPageIsolated(page); |
| 161 | unlock_page(page); |
| 162 | put_page(page); |
| 163 | } else { |
| 164 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
| 165 | page_is_file_lru(page), -thp_nr_pages(page)); |
| 166 | putback_lru_page(page); |
| 167 | } |
| 168 | } |
| 169 | } |
| 170 | |
| 171 | /* |
| 172 | * Restore a potential migration pte to a working pte entry |
| 173 | */ |
| 174 | static bool remove_migration_pte(struct folio *folio, |
| 175 | struct vm_area_struct *vma, unsigned long addr, void *old) |
| 176 | { |
| 177 | DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION); |
| 178 | |
| 179 | while (page_vma_mapped_walk(&pvmw)) { |
| 180 | pte_t pte; |
| 181 | swp_entry_t entry; |
| 182 | struct page *new; |
| 183 | unsigned long idx = 0; |
| 184 | |
| 185 | /* pgoff is invalid for ksm pages, but they are never large */ |
| 186 | if (folio_test_large(folio) && !folio_test_hugetlb(folio)) |
| 187 | idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff; |
| 188 | new = folio_page(folio, idx); |
| 189 | |
| 190 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
| 191 | /* PMD-mapped THP migration entry */ |
| 192 | if (!pvmw.pte) { |
| 193 | VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) || |
| 194 | !folio_test_pmd_mappable(folio), folio); |
| 195 | remove_migration_pmd(&pvmw, new); |
| 196 | continue; |
| 197 | } |
| 198 | #endif |
| 199 | |
| 200 | folio_get(folio); |
| 201 | pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); |
| 202 | if (pte_swp_soft_dirty(*pvmw.pte)) |
| 203 | pte = pte_mksoft_dirty(pte); |
| 204 | |
| 205 | /* |
| 206 | * Recheck VMA as permissions can change since migration started |
| 207 | */ |
| 208 | entry = pte_to_swp_entry(*pvmw.pte); |
| 209 | if (is_writable_migration_entry(entry)) |
| 210 | pte = maybe_mkwrite(pte, vma); |
| 211 | else if (pte_swp_uffd_wp(*pvmw.pte)) |
| 212 | pte = pte_mkuffd_wp(pte); |
| 213 | |
| 214 | if (unlikely(is_device_private_page(new))) { |
| 215 | if (pte_write(pte)) |
| 216 | entry = make_writable_device_private_entry( |
| 217 | page_to_pfn(new)); |
| 218 | else |
| 219 | entry = make_readable_device_private_entry( |
| 220 | page_to_pfn(new)); |
| 221 | pte = swp_entry_to_pte(entry); |
| 222 | if (pte_swp_soft_dirty(*pvmw.pte)) |
| 223 | pte = pte_swp_mksoft_dirty(pte); |
| 224 | if (pte_swp_uffd_wp(*pvmw.pte)) |
| 225 | pte = pte_swp_mkuffd_wp(pte); |
| 226 | } |
| 227 | |
| 228 | #ifdef CONFIG_HUGETLB_PAGE |
| 229 | if (folio_test_hugetlb(folio)) { |
| 230 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
| 231 | |
| 232 | pte = pte_mkhuge(pte); |
| 233 | pte = arch_make_huge_pte(pte, shift, vma->vm_flags); |
| 234 | if (folio_test_anon(folio)) |
| 235 | hugepage_add_anon_rmap(new, vma, pvmw.address); |
| 236 | else |
| 237 | page_dup_rmap(new, true); |
| 238 | set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
| 239 | } else |
| 240 | #endif |
| 241 | { |
| 242 | if (folio_test_anon(folio)) |
| 243 | page_add_anon_rmap(new, vma, pvmw.address, false); |
| 244 | else |
| 245 | page_add_file_rmap(new, vma, false); |
| 246 | set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); |
| 247 | } |
| 248 | if (vma->vm_flags & VM_LOCKED) |
| 249 | mlock_page_drain_local(); |
| 250 | |
| 251 | trace_remove_migration_pte(pvmw.address, pte_val(pte), |
| 252 | compound_order(new)); |
| 253 | |
| 254 | /* No need to invalidate - it was non-present before */ |
| 255 | update_mmu_cache(vma, pvmw.address, pvmw.pte); |
| 256 | } |
| 257 | |
| 258 | return true; |
| 259 | } |
| 260 | |
| 261 | /* |
| 262 | * Get rid of all migration entries and replace them by |
| 263 | * references to the indicated page. |
| 264 | */ |
| 265 | void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked) |
| 266 | { |
| 267 | struct rmap_walk_control rwc = { |
| 268 | .rmap_one = remove_migration_pte, |
| 269 | .arg = src, |
| 270 | }; |
| 271 | |
| 272 | if (locked) |
| 273 | rmap_walk_locked(dst, &rwc); |
| 274 | else |
| 275 | rmap_walk(dst, &rwc); |
| 276 | } |
| 277 | |
| 278 | /* |
| 279 | * Something used the pte of a page under migration. We need to |
| 280 | * get to the page and wait until migration is finished. |
| 281 | * When we return from this function the fault will be retried. |
| 282 | */ |
| 283 | void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
| 284 | spinlock_t *ptl) |
| 285 | { |
| 286 | pte_t pte; |
| 287 | swp_entry_t entry; |
| 288 | |
| 289 | spin_lock(ptl); |
| 290 | pte = *ptep; |
| 291 | if (!is_swap_pte(pte)) |
| 292 | goto out; |
| 293 | |
| 294 | entry = pte_to_swp_entry(pte); |
| 295 | if (!is_migration_entry(entry)) |
| 296 | goto out; |
| 297 | |
| 298 | migration_entry_wait_on_locked(entry, ptep, ptl); |
| 299 | return; |
| 300 | out: |
| 301 | pte_unmap_unlock(ptep, ptl); |
| 302 | } |
| 303 | |
| 304 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
| 305 | unsigned long address) |
| 306 | { |
| 307 | spinlock_t *ptl = pte_lockptr(mm, pmd); |
| 308 | pte_t *ptep = pte_offset_map(pmd, address); |
| 309 | __migration_entry_wait(mm, ptep, ptl); |
| 310 | } |
| 311 | |
| 312 | void migration_entry_wait_huge(struct vm_area_struct *vma, |
| 313 | struct mm_struct *mm, pte_t *pte) |
| 314 | { |
| 315 | spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); |
| 316 | __migration_entry_wait(mm, pte, ptl); |
| 317 | } |
| 318 | |
| 319 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
| 320 | void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) |
| 321 | { |
| 322 | spinlock_t *ptl; |
| 323 | |
| 324 | ptl = pmd_lock(mm, pmd); |
| 325 | if (!is_pmd_migration_entry(*pmd)) |
| 326 | goto unlock; |
| 327 | migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl); |
| 328 | return; |
| 329 | unlock: |
| 330 | spin_unlock(ptl); |
| 331 | } |
| 332 | #endif |
| 333 | |
| 334 | static int expected_page_refs(struct address_space *mapping, struct page *page) |
| 335 | { |
| 336 | int expected_count = 1; |
| 337 | |
| 338 | if (mapping) |
| 339 | expected_count += compound_nr(page) + page_has_private(page); |
| 340 | return expected_count; |
| 341 | } |
| 342 | |
| 343 | /* |
| 344 | * Replace the page in the mapping. |
| 345 | * |
| 346 | * The number of remaining references must be: |
| 347 | * 1 for anonymous pages without a mapping |
| 348 | * 2 for pages with a mapping |
| 349 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
| 350 | */ |
| 351 | int folio_migrate_mapping(struct address_space *mapping, |
| 352 | struct folio *newfolio, struct folio *folio, int extra_count) |
| 353 | { |
| 354 | XA_STATE(xas, &mapping->i_pages, folio_index(folio)); |
| 355 | struct zone *oldzone, *newzone; |
| 356 | int dirty; |
| 357 | int expected_count = expected_page_refs(mapping, &folio->page) + extra_count; |
| 358 | long nr = folio_nr_pages(folio); |
| 359 | |
| 360 | if (!mapping) { |
| 361 | /* Anonymous page without mapping */ |
| 362 | if (folio_ref_count(folio) != expected_count) |
| 363 | return -EAGAIN; |
| 364 | |
| 365 | /* No turning back from here */ |
| 366 | newfolio->index = folio->index; |
| 367 | newfolio->mapping = folio->mapping; |
| 368 | if (folio_test_swapbacked(folio)) |
| 369 | __folio_set_swapbacked(newfolio); |
| 370 | |
| 371 | return MIGRATEPAGE_SUCCESS; |
| 372 | } |
| 373 | |
| 374 | oldzone = folio_zone(folio); |
| 375 | newzone = folio_zone(newfolio); |
| 376 | |
| 377 | xas_lock_irq(&xas); |
| 378 | if (!folio_ref_freeze(folio, expected_count)) { |
| 379 | xas_unlock_irq(&xas); |
| 380 | return -EAGAIN; |
| 381 | } |
| 382 | |
| 383 | /* |
| 384 | * Now we know that no one else is looking at the folio: |
| 385 | * no turning back from here. |
| 386 | */ |
| 387 | newfolio->index = folio->index; |
| 388 | newfolio->mapping = folio->mapping; |
| 389 | folio_ref_add(newfolio, nr); /* add cache reference */ |
| 390 | if (folio_test_swapbacked(folio)) { |
| 391 | __folio_set_swapbacked(newfolio); |
| 392 | if (folio_test_swapcache(folio)) { |
| 393 | folio_set_swapcache(newfolio); |
| 394 | newfolio->private = folio_get_private(folio); |
| 395 | } |
| 396 | } else { |
| 397 | VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio); |
| 398 | } |
| 399 | |
| 400 | /* Move dirty while page refs frozen and newpage not yet exposed */ |
| 401 | dirty = folio_test_dirty(folio); |
| 402 | if (dirty) { |
| 403 | folio_clear_dirty(folio); |
| 404 | folio_set_dirty(newfolio); |
| 405 | } |
| 406 | |
| 407 | xas_store(&xas, newfolio); |
| 408 | |
| 409 | /* |
| 410 | * Drop cache reference from old page by unfreezing |
| 411 | * to one less reference. |
| 412 | * We know this isn't the last reference. |
| 413 | */ |
| 414 | folio_ref_unfreeze(folio, expected_count - nr); |
| 415 | |
| 416 | xas_unlock(&xas); |
| 417 | /* Leave irq disabled to prevent preemption while updating stats */ |
| 418 | |
| 419 | /* |
| 420 | * If moved to a different zone then also account |
| 421 | * the page for that zone. Other VM counters will be |
| 422 | * taken care of when we establish references to the |
| 423 | * new page and drop references to the old page. |
| 424 | * |
| 425 | * Note that anonymous pages are accounted for |
| 426 | * via NR_FILE_PAGES and NR_ANON_MAPPED if they |
| 427 | * are mapped to swap space. |
| 428 | */ |
| 429 | if (newzone != oldzone) { |
| 430 | struct lruvec *old_lruvec, *new_lruvec; |
| 431 | struct mem_cgroup *memcg; |
| 432 | |
| 433 | memcg = folio_memcg(folio); |
| 434 | old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); |
| 435 | new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); |
| 436 | |
| 437 | __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr); |
| 438 | __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr); |
| 439 | if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) { |
| 440 | __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr); |
| 441 | __mod_lruvec_state(new_lruvec, NR_SHMEM, nr); |
| 442 | } |
| 443 | #ifdef CONFIG_SWAP |
| 444 | if (folio_test_swapcache(folio)) { |
| 445 | __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr); |
| 446 | __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr); |
| 447 | } |
| 448 | #endif |
| 449 | if (dirty && mapping_can_writeback(mapping)) { |
| 450 | __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr); |
| 451 | __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr); |
| 452 | __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr); |
| 453 | __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr); |
| 454 | } |
| 455 | } |
| 456 | local_irq_enable(); |
| 457 | |
| 458 | return MIGRATEPAGE_SUCCESS; |
| 459 | } |
| 460 | EXPORT_SYMBOL(folio_migrate_mapping); |
| 461 | |
| 462 | /* |
| 463 | * The expected number of remaining references is the same as that |
| 464 | * of folio_migrate_mapping(). |
| 465 | */ |
| 466 | int migrate_huge_page_move_mapping(struct address_space *mapping, |
| 467 | struct page *newpage, struct page *page) |
| 468 | { |
| 469 | XA_STATE(xas, &mapping->i_pages, page_index(page)); |
| 470 | int expected_count; |
| 471 | |
| 472 | xas_lock_irq(&xas); |
| 473 | expected_count = 2 + page_has_private(page); |
| 474 | if (page_count(page) != expected_count || xas_load(&xas) != page) { |
| 475 | xas_unlock_irq(&xas); |
| 476 | return -EAGAIN; |
| 477 | } |
| 478 | |
| 479 | if (!page_ref_freeze(page, expected_count)) { |
| 480 | xas_unlock_irq(&xas); |
| 481 | return -EAGAIN; |
| 482 | } |
| 483 | |
| 484 | newpage->index = page->index; |
| 485 | newpage->mapping = page->mapping; |
| 486 | |
| 487 | get_page(newpage); |
| 488 | |
| 489 | xas_store(&xas, newpage); |
| 490 | |
| 491 | page_ref_unfreeze(page, expected_count - 1); |
| 492 | |
| 493 | xas_unlock_irq(&xas); |
| 494 | |
| 495 | return MIGRATEPAGE_SUCCESS; |
| 496 | } |
| 497 | |
| 498 | /* |
| 499 | * Copy the flags and some other ancillary information |
| 500 | */ |
| 501 | void folio_migrate_flags(struct folio *newfolio, struct folio *folio) |
| 502 | { |
| 503 | int cpupid; |
| 504 | |
| 505 | if (folio_test_error(folio)) |
| 506 | folio_set_error(newfolio); |
| 507 | if (folio_test_referenced(folio)) |
| 508 | folio_set_referenced(newfolio); |
| 509 | if (folio_test_uptodate(folio)) |
| 510 | folio_mark_uptodate(newfolio); |
| 511 | if (folio_test_clear_active(folio)) { |
| 512 | VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio); |
| 513 | folio_set_active(newfolio); |
| 514 | } else if (folio_test_clear_unevictable(folio)) |
| 515 | folio_set_unevictable(newfolio); |
| 516 | if (folio_test_workingset(folio)) |
| 517 | folio_set_workingset(newfolio); |
| 518 | if (folio_test_checked(folio)) |
| 519 | folio_set_checked(newfolio); |
| 520 | if (folio_test_mappedtodisk(folio)) |
| 521 | folio_set_mappedtodisk(newfolio); |
| 522 | |
| 523 | /* Move dirty on pages not done by folio_migrate_mapping() */ |
| 524 | if (folio_test_dirty(folio)) |
| 525 | folio_set_dirty(newfolio); |
| 526 | |
| 527 | if (folio_test_young(folio)) |
| 528 | folio_set_young(newfolio); |
| 529 | if (folio_test_idle(folio)) |
| 530 | folio_set_idle(newfolio); |
| 531 | |
| 532 | /* |
| 533 | * Copy NUMA information to the new page, to prevent over-eager |
| 534 | * future migrations of this same page. |
| 535 | */ |
| 536 | cpupid = page_cpupid_xchg_last(&folio->page, -1); |
| 537 | page_cpupid_xchg_last(&newfolio->page, cpupid); |
| 538 | |
| 539 | folio_migrate_ksm(newfolio, folio); |
| 540 | /* |
| 541 | * Please do not reorder this without considering how mm/ksm.c's |
| 542 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). |
| 543 | */ |
| 544 | if (folio_test_swapcache(folio)) |
| 545 | folio_clear_swapcache(folio); |
| 546 | folio_clear_private(folio); |
| 547 | |
| 548 | /* page->private contains hugetlb specific flags */ |
| 549 | if (!folio_test_hugetlb(folio)) |
| 550 | folio->private = NULL; |
| 551 | |
| 552 | /* |
| 553 | * If any waiters have accumulated on the new page then |
| 554 | * wake them up. |
| 555 | */ |
| 556 | if (folio_test_writeback(newfolio)) |
| 557 | folio_end_writeback(newfolio); |
| 558 | |
| 559 | /* |
| 560 | * PG_readahead shares the same bit with PG_reclaim. The above |
| 561 | * end_page_writeback() may clear PG_readahead mistakenly, so set the |
| 562 | * bit after that. |
| 563 | */ |
| 564 | if (folio_test_readahead(folio)) |
| 565 | folio_set_readahead(newfolio); |
| 566 | |
| 567 | folio_copy_owner(newfolio, folio); |
| 568 | |
| 569 | if (!folio_test_hugetlb(folio)) |
| 570 | mem_cgroup_migrate(folio, newfolio); |
| 571 | } |
| 572 | EXPORT_SYMBOL(folio_migrate_flags); |
| 573 | |
| 574 | void folio_migrate_copy(struct folio *newfolio, struct folio *folio) |
| 575 | { |
| 576 | folio_copy(newfolio, folio); |
| 577 | folio_migrate_flags(newfolio, folio); |
| 578 | } |
| 579 | EXPORT_SYMBOL(folio_migrate_copy); |
| 580 | |
| 581 | /************************************************************ |
| 582 | * Migration functions |
| 583 | ***********************************************************/ |
| 584 | |
| 585 | /* |
| 586 | * Common logic to directly migrate a single LRU page suitable for |
| 587 | * pages that do not use PagePrivate/PagePrivate2. |
| 588 | * |
| 589 | * Pages are locked upon entry and exit. |
| 590 | */ |
| 591 | int migrate_page(struct address_space *mapping, |
| 592 | struct page *newpage, struct page *page, |
| 593 | enum migrate_mode mode) |
| 594 | { |
| 595 | struct folio *newfolio = page_folio(newpage); |
| 596 | struct folio *folio = page_folio(page); |
| 597 | int rc; |
| 598 | |
| 599 | BUG_ON(folio_test_writeback(folio)); /* Writeback must be complete */ |
| 600 | |
| 601 | rc = folio_migrate_mapping(mapping, newfolio, folio, 0); |
| 602 | |
| 603 | if (rc != MIGRATEPAGE_SUCCESS) |
| 604 | return rc; |
| 605 | |
| 606 | if (mode != MIGRATE_SYNC_NO_COPY) |
| 607 | folio_migrate_copy(newfolio, folio); |
| 608 | else |
| 609 | folio_migrate_flags(newfolio, folio); |
| 610 | return MIGRATEPAGE_SUCCESS; |
| 611 | } |
| 612 | EXPORT_SYMBOL(migrate_page); |
| 613 | |
| 614 | #ifdef CONFIG_BLOCK |
| 615 | /* Returns true if all buffers are successfully locked */ |
| 616 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, |
| 617 | enum migrate_mode mode) |
| 618 | { |
| 619 | struct buffer_head *bh = head; |
| 620 | |
| 621 | /* Simple case, sync compaction */ |
| 622 | if (mode != MIGRATE_ASYNC) { |
| 623 | do { |
| 624 | lock_buffer(bh); |
| 625 | bh = bh->b_this_page; |
| 626 | |
| 627 | } while (bh != head); |
| 628 | |
| 629 | return true; |
| 630 | } |
| 631 | |
| 632 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ |
| 633 | do { |
| 634 | if (!trylock_buffer(bh)) { |
| 635 | /* |
| 636 | * We failed to lock the buffer and cannot stall in |
| 637 | * async migration. Release the taken locks |
| 638 | */ |
| 639 | struct buffer_head *failed_bh = bh; |
| 640 | bh = head; |
| 641 | while (bh != failed_bh) { |
| 642 | unlock_buffer(bh); |
| 643 | bh = bh->b_this_page; |
| 644 | } |
| 645 | return false; |
| 646 | } |
| 647 | |
| 648 | bh = bh->b_this_page; |
| 649 | } while (bh != head); |
| 650 | return true; |
| 651 | } |
| 652 | |
| 653 | static int __buffer_migrate_page(struct address_space *mapping, |
| 654 | struct page *newpage, struct page *page, enum migrate_mode mode, |
| 655 | bool check_refs) |
| 656 | { |
| 657 | struct buffer_head *bh, *head; |
| 658 | int rc; |
| 659 | int expected_count; |
| 660 | |
| 661 | if (!page_has_buffers(page)) |
| 662 | return migrate_page(mapping, newpage, page, mode); |
| 663 | |
| 664 | /* Check whether page does not have extra refs before we do more work */ |
| 665 | expected_count = expected_page_refs(mapping, page); |
| 666 | if (page_count(page) != expected_count) |
| 667 | return -EAGAIN; |
| 668 | |
| 669 | head = page_buffers(page); |
| 670 | if (!buffer_migrate_lock_buffers(head, mode)) |
| 671 | return -EAGAIN; |
| 672 | |
| 673 | if (check_refs) { |
| 674 | bool busy; |
| 675 | bool invalidated = false; |
| 676 | |
| 677 | recheck_buffers: |
| 678 | busy = false; |
| 679 | spin_lock(&mapping->private_lock); |
| 680 | bh = head; |
| 681 | do { |
| 682 | if (atomic_read(&bh->b_count)) { |
| 683 | busy = true; |
| 684 | break; |
| 685 | } |
| 686 | bh = bh->b_this_page; |
| 687 | } while (bh != head); |
| 688 | if (busy) { |
| 689 | if (invalidated) { |
| 690 | rc = -EAGAIN; |
| 691 | goto unlock_buffers; |
| 692 | } |
| 693 | spin_unlock(&mapping->private_lock); |
| 694 | invalidate_bh_lrus(); |
| 695 | invalidated = true; |
| 696 | goto recheck_buffers; |
| 697 | } |
| 698 | } |
| 699 | |
| 700 | rc = migrate_page_move_mapping(mapping, newpage, page, 0); |
| 701 | if (rc != MIGRATEPAGE_SUCCESS) |
| 702 | goto unlock_buffers; |
| 703 | |
| 704 | attach_page_private(newpage, detach_page_private(page)); |
| 705 | |
| 706 | bh = head; |
| 707 | do { |
| 708 | set_bh_page(bh, newpage, bh_offset(bh)); |
| 709 | bh = bh->b_this_page; |
| 710 | |
| 711 | } while (bh != head); |
| 712 | |
| 713 | if (mode != MIGRATE_SYNC_NO_COPY) |
| 714 | migrate_page_copy(newpage, page); |
| 715 | else |
| 716 | migrate_page_states(newpage, page); |
| 717 | |
| 718 | rc = MIGRATEPAGE_SUCCESS; |
| 719 | unlock_buffers: |
| 720 | if (check_refs) |
| 721 | spin_unlock(&mapping->private_lock); |
| 722 | bh = head; |
| 723 | do { |
| 724 | unlock_buffer(bh); |
| 725 | bh = bh->b_this_page; |
| 726 | |
| 727 | } while (bh != head); |
| 728 | |
| 729 | return rc; |
| 730 | } |
| 731 | |
| 732 | /* |
| 733 | * Migration function for pages with buffers. This function can only be used |
| 734 | * if the underlying filesystem guarantees that no other references to "page" |
| 735 | * exist. For example attached buffer heads are accessed only under page lock. |
| 736 | */ |
| 737 | int buffer_migrate_page(struct address_space *mapping, |
| 738 | struct page *newpage, struct page *page, enum migrate_mode mode) |
| 739 | { |
| 740 | return __buffer_migrate_page(mapping, newpage, page, mode, false); |
| 741 | } |
| 742 | EXPORT_SYMBOL(buffer_migrate_page); |
| 743 | |
| 744 | /* |
| 745 | * Same as above except that this variant is more careful and checks that there |
| 746 | * are also no buffer head references. This function is the right one for |
| 747 | * mappings where buffer heads are directly looked up and referenced (such as |
| 748 | * block device mappings). |
| 749 | */ |
| 750 | int buffer_migrate_page_norefs(struct address_space *mapping, |
| 751 | struct page *newpage, struct page *page, enum migrate_mode mode) |
| 752 | { |
| 753 | return __buffer_migrate_page(mapping, newpage, page, mode, true); |
| 754 | } |
| 755 | #endif |
| 756 | |
| 757 | /* |
| 758 | * Writeback a page to clean the dirty state |
| 759 | */ |
| 760 | static int writeout(struct address_space *mapping, struct page *page) |
| 761 | { |
| 762 | struct folio *folio = page_folio(page); |
| 763 | struct writeback_control wbc = { |
| 764 | .sync_mode = WB_SYNC_NONE, |
| 765 | .nr_to_write = 1, |
| 766 | .range_start = 0, |
| 767 | .range_end = LLONG_MAX, |
| 768 | .for_reclaim = 1 |
| 769 | }; |
| 770 | int rc; |
| 771 | |
| 772 | if (!mapping->a_ops->writepage) |
| 773 | /* No write method for the address space */ |
| 774 | return -EINVAL; |
| 775 | |
| 776 | if (!clear_page_dirty_for_io(page)) |
| 777 | /* Someone else already triggered a write */ |
| 778 | return -EAGAIN; |
| 779 | |
| 780 | /* |
| 781 | * A dirty page may imply that the underlying filesystem has |
| 782 | * the page on some queue. So the page must be clean for |
| 783 | * migration. Writeout may mean we loose the lock and the |
| 784 | * page state is no longer what we checked for earlier. |
| 785 | * At this point we know that the migration attempt cannot |
| 786 | * be successful. |
| 787 | */ |
| 788 | remove_migration_ptes(folio, folio, false); |
| 789 | |
| 790 | rc = mapping->a_ops->writepage(page, &wbc); |
| 791 | |
| 792 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
| 793 | /* unlocked. Relock */ |
| 794 | lock_page(page); |
| 795 | |
| 796 | return (rc < 0) ? -EIO : -EAGAIN; |
| 797 | } |
| 798 | |
| 799 | /* |
| 800 | * Default handling if a filesystem does not provide a migration function. |
| 801 | */ |
| 802 | static int fallback_migrate_page(struct address_space *mapping, |
| 803 | struct page *newpage, struct page *page, enum migrate_mode mode) |
| 804 | { |
| 805 | if (PageDirty(page)) { |
| 806 | /* Only writeback pages in full synchronous migration */ |
| 807 | switch (mode) { |
| 808 | case MIGRATE_SYNC: |
| 809 | case MIGRATE_SYNC_NO_COPY: |
| 810 | break; |
| 811 | default: |
| 812 | return -EBUSY; |
| 813 | } |
| 814 | return writeout(mapping, page); |
| 815 | } |
| 816 | |
| 817 | /* |
| 818 | * Buffers may be managed in a filesystem specific way. |
| 819 | * We must have no buffers or drop them. |
| 820 | */ |
| 821 | if (page_has_private(page) && |
| 822 | !try_to_release_page(page, GFP_KERNEL)) |
| 823 | return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; |
| 824 | |
| 825 | return migrate_page(mapping, newpage, page, mode); |
| 826 | } |
| 827 | |
| 828 | /* |
| 829 | * Move a page to a newly allocated page |
| 830 | * The page is locked and all ptes have been successfully removed. |
| 831 | * |
| 832 | * The new page will have replaced the old page if this function |
| 833 | * is successful. |
| 834 | * |
| 835 | * Return value: |
| 836 | * < 0 - error code |
| 837 | * MIGRATEPAGE_SUCCESS - success |
| 838 | */ |
| 839 | static int move_to_new_page(struct page *newpage, struct page *page, |
| 840 | enum migrate_mode mode) |
| 841 | { |
| 842 | struct address_space *mapping; |
| 843 | int rc = -EAGAIN; |
| 844 | bool is_lru = !__PageMovable(page); |
| 845 | |
| 846 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
| 847 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); |
| 848 | |
| 849 | mapping = page_mapping(page); |
| 850 | |
| 851 | if (likely(is_lru)) { |
| 852 | if (!mapping) |
| 853 | rc = migrate_page(mapping, newpage, page, mode); |
| 854 | else if (mapping->a_ops->migratepage) |
| 855 | /* |
| 856 | * Most pages have a mapping and most filesystems |
| 857 | * provide a migratepage callback. Anonymous pages |
| 858 | * are part of swap space which also has its own |
| 859 | * migratepage callback. This is the most common path |
| 860 | * for page migration. |
| 861 | */ |
| 862 | rc = mapping->a_ops->migratepage(mapping, newpage, |
| 863 | page, mode); |
| 864 | else |
| 865 | rc = fallback_migrate_page(mapping, newpage, |
| 866 | page, mode); |
| 867 | } else { |
| 868 | /* |
| 869 | * In case of non-lru page, it could be released after |
| 870 | * isolation step. In that case, we shouldn't try migration. |
| 871 | */ |
| 872 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
| 873 | if (!PageMovable(page)) { |
| 874 | rc = MIGRATEPAGE_SUCCESS; |
| 875 | ClearPageIsolated(page); |
| 876 | goto out; |
| 877 | } |
| 878 | |
| 879 | rc = mapping->a_ops->migratepage(mapping, newpage, |
| 880 | page, mode); |
| 881 | WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && |
| 882 | !PageIsolated(page)); |
| 883 | } |
| 884 | |
| 885 | /* |
| 886 | * When successful, old pagecache page->mapping must be cleared before |
| 887 | * page is freed; but stats require that PageAnon be left as PageAnon. |
| 888 | */ |
| 889 | if (rc == MIGRATEPAGE_SUCCESS) { |
| 890 | if (__PageMovable(page)) { |
| 891 | VM_BUG_ON_PAGE(!PageIsolated(page), page); |
| 892 | |
| 893 | /* |
| 894 | * We clear PG_movable under page_lock so any compactor |
| 895 | * cannot try to migrate this page. |
| 896 | */ |
| 897 | ClearPageIsolated(page); |
| 898 | } |
| 899 | |
| 900 | /* |
| 901 | * Anonymous and movable page->mapping will be cleared by |
| 902 | * free_pages_prepare so don't reset it here for keeping |
| 903 | * the type to work PageAnon, for example. |
| 904 | */ |
| 905 | if (!PageMappingFlags(page)) |
| 906 | page->mapping = NULL; |
| 907 | |
| 908 | if (likely(!is_zone_device_page(newpage))) |
| 909 | flush_dcache_folio(page_folio(newpage)); |
| 910 | } |
| 911 | out: |
| 912 | return rc; |
| 913 | } |
| 914 | |
| 915 | static int __unmap_and_move(struct page *page, struct page *newpage, |
| 916 | int force, enum migrate_mode mode) |
| 917 | { |
| 918 | struct folio *folio = page_folio(page); |
| 919 | struct folio *dst = page_folio(newpage); |
| 920 | int rc = -EAGAIN; |
| 921 | bool page_was_mapped = false; |
| 922 | struct anon_vma *anon_vma = NULL; |
| 923 | bool is_lru = !__PageMovable(page); |
| 924 | |
| 925 | if (!trylock_page(page)) { |
| 926 | if (!force || mode == MIGRATE_ASYNC) |
| 927 | goto out; |
| 928 | |
| 929 | /* |
| 930 | * It's not safe for direct compaction to call lock_page. |
| 931 | * For example, during page readahead pages are added locked |
| 932 | * to the LRU. Later, when the IO completes the pages are |
| 933 | * marked uptodate and unlocked. However, the queueing |
| 934 | * could be merging multiple pages for one bio (e.g. |
| 935 | * mpage_readahead). If an allocation happens for the |
| 936 | * second or third page, the process can end up locking |
| 937 | * the same page twice and deadlocking. Rather than |
| 938 | * trying to be clever about what pages can be locked, |
| 939 | * avoid the use of lock_page for direct compaction |
| 940 | * altogether. |
| 941 | */ |
| 942 | if (current->flags & PF_MEMALLOC) |
| 943 | goto out; |
| 944 | |
| 945 | lock_page(page); |
| 946 | } |
| 947 | |
| 948 | if (PageWriteback(page)) { |
| 949 | /* |
| 950 | * Only in the case of a full synchronous migration is it |
| 951 | * necessary to wait for PageWriteback. In the async case, |
| 952 | * the retry loop is too short and in the sync-light case, |
| 953 | * the overhead of stalling is too much |
| 954 | */ |
| 955 | switch (mode) { |
| 956 | case MIGRATE_SYNC: |
| 957 | case MIGRATE_SYNC_NO_COPY: |
| 958 | break; |
| 959 | default: |
| 960 | rc = -EBUSY; |
| 961 | goto out_unlock; |
| 962 | } |
| 963 | if (!force) |
| 964 | goto out_unlock; |
| 965 | wait_on_page_writeback(page); |
| 966 | } |
| 967 | |
| 968 | /* |
| 969 | * By try_to_migrate(), page->mapcount goes down to 0 here. In this case, |
| 970 | * we cannot notice that anon_vma is freed while we migrates a page. |
| 971 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
| 972 | * of migration. File cache pages are no problem because of page_lock() |
| 973 | * File Caches may use write_page() or lock_page() in migration, then, |
| 974 | * just care Anon page here. |
| 975 | * |
| 976 | * Only page_get_anon_vma() understands the subtleties of |
| 977 | * getting a hold on an anon_vma from outside one of its mms. |
| 978 | * But if we cannot get anon_vma, then we won't need it anyway, |
| 979 | * because that implies that the anon page is no longer mapped |
| 980 | * (and cannot be remapped so long as we hold the page lock). |
| 981 | */ |
| 982 | if (PageAnon(page) && !PageKsm(page)) |
| 983 | anon_vma = page_get_anon_vma(page); |
| 984 | |
| 985 | /* |
| 986 | * Block others from accessing the new page when we get around to |
| 987 | * establishing additional references. We are usually the only one |
| 988 | * holding a reference to newpage at this point. We used to have a BUG |
| 989 | * here if trylock_page(newpage) fails, but would like to allow for |
| 990 | * cases where there might be a race with the previous use of newpage. |
| 991 | * This is much like races on refcount of oldpage: just don't BUG(). |
| 992 | */ |
| 993 | if (unlikely(!trylock_page(newpage))) |
| 994 | goto out_unlock; |
| 995 | |
| 996 | if (unlikely(!is_lru)) { |
| 997 | rc = move_to_new_page(newpage, page, mode); |
| 998 | goto out_unlock_both; |
| 999 | } |
| 1000 | |
| 1001 | /* |
| 1002 | * Corner case handling: |
| 1003 | * 1. When a new swap-cache page is read into, it is added to the LRU |
| 1004 | * and treated as swapcache but it has no rmap yet. |
| 1005 | * Calling try_to_unmap() against a page->mapping==NULL page will |
| 1006 | * trigger a BUG. So handle it here. |
| 1007 | * 2. An orphaned page (see truncate_cleanup_page) might have |
| 1008 | * fs-private metadata. The page can be picked up due to memory |
| 1009 | * offlining. Everywhere else except page reclaim, the page is |
| 1010 | * invisible to the vm, so the page can not be migrated. So try to |
| 1011 | * free the metadata, so the page can be freed. |
| 1012 | */ |
| 1013 | if (!page->mapping) { |
| 1014 | VM_BUG_ON_PAGE(PageAnon(page), page); |
| 1015 | if (page_has_private(page)) { |
| 1016 | try_to_free_buffers(page); |
| 1017 | goto out_unlock_both; |
| 1018 | } |
| 1019 | } else if (page_mapped(page)) { |
| 1020 | /* Establish migration ptes */ |
| 1021 | VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, |
| 1022 | page); |
| 1023 | try_to_migrate(folio, 0); |
| 1024 | page_was_mapped = true; |
| 1025 | } |
| 1026 | |
| 1027 | if (!page_mapped(page)) |
| 1028 | rc = move_to_new_page(newpage, page, mode); |
| 1029 | |
| 1030 | /* |
| 1031 | * When successful, push newpage to LRU immediately: so that if it |
| 1032 | * turns out to be an mlocked page, remove_migration_ptes() will |
| 1033 | * automatically build up the correct newpage->mlock_count for it. |
| 1034 | * |
| 1035 | * We would like to do something similar for the old page, when |
| 1036 | * unsuccessful, and other cases when a page has been temporarily |
| 1037 | * isolated from the unevictable LRU: but this case is the easiest. |
| 1038 | */ |
| 1039 | if (rc == MIGRATEPAGE_SUCCESS) { |
| 1040 | lru_cache_add(newpage); |
| 1041 | if (page_was_mapped) |
| 1042 | lru_add_drain(); |
| 1043 | } |
| 1044 | |
| 1045 | if (page_was_mapped) |
| 1046 | remove_migration_ptes(folio, |
| 1047 | rc == MIGRATEPAGE_SUCCESS ? dst : folio, false); |
| 1048 | |
| 1049 | out_unlock_both: |
| 1050 | unlock_page(newpage); |
| 1051 | out_unlock: |
| 1052 | /* Drop an anon_vma reference if we took one */ |
| 1053 | if (anon_vma) |
| 1054 | put_anon_vma(anon_vma); |
| 1055 | unlock_page(page); |
| 1056 | out: |
| 1057 | /* |
| 1058 | * If migration is successful, decrease refcount of the newpage, |
| 1059 | * which will not free the page because new page owner increased |
| 1060 | * refcounter. |
| 1061 | */ |
| 1062 | if (rc == MIGRATEPAGE_SUCCESS) |
| 1063 | put_page(newpage); |
| 1064 | |
| 1065 | return rc; |
| 1066 | } |
| 1067 | |
| 1068 | /* |
| 1069 | * Obtain the lock on page, remove all ptes and migrate the page |
| 1070 | * to the newly allocated page in newpage. |
| 1071 | */ |
| 1072 | static int unmap_and_move(new_page_t get_new_page, |
| 1073 | free_page_t put_new_page, |
| 1074 | unsigned long private, struct page *page, |
| 1075 | int force, enum migrate_mode mode, |
| 1076 | enum migrate_reason reason, |
| 1077 | struct list_head *ret) |
| 1078 | { |
| 1079 | int rc = MIGRATEPAGE_SUCCESS; |
| 1080 | struct page *newpage = NULL; |
| 1081 | |
| 1082 | if (!thp_migration_supported() && PageTransHuge(page)) |
| 1083 | return -ENOSYS; |
| 1084 | |
| 1085 | if (page_count(page) == 1) { |
| 1086 | /* page was freed from under us. So we are done. */ |
| 1087 | ClearPageActive(page); |
| 1088 | ClearPageUnevictable(page); |
| 1089 | if (unlikely(__PageMovable(page))) { |
| 1090 | lock_page(page); |
| 1091 | if (!PageMovable(page)) |
| 1092 | ClearPageIsolated(page); |
| 1093 | unlock_page(page); |
| 1094 | } |
| 1095 | goto out; |
| 1096 | } |
| 1097 | |
| 1098 | newpage = get_new_page(page, private); |
| 1099 | if (!newpage) |
| 1100 | return -ENOMEM; |
| 1101 | |
| 1102 | rc = __unmap_and_move(page, newpage, force, mode); |
| 1103 | if (rc == MIGRATEPAGE_SUCCESS) |
| 1104 | set_page_owner_migrate_reason(newpage, reason); |
| 1105 | |
| 1106 | out: |
| 1107 | if (rc != -EAGAIN) { |
| 1108 | /* |
| 1109 | * A page that has been migrated has all references |
| 1110 | * removed and will be freed. A page that has not been |
| 1111 | * migrated will have kept its references and be restored. |
| 1112 | */ |
| 1113 | list_del(&page->lru); |
| 1114 | } |
| 1115 | |
| 1116 | /* |
| 1117 | * If migration is successful, releases reference grabbed during |
| 1118 | * isolation. Otherwise, restore the page to right list unless |
| 1119 | * we want to retry. |
| 1120 | */ |
| 1121 | if (rc == MIGRATEPAGE_SUCCESS) { |
| 1122 | /* |
| 1123 | * Compaction can migrate also non-LRU pages which are |
| 1124 | * not accounted to NR_ISOLATED_*. They can be recognized |
| 1125 | * as __PageMovable |
| 1126 | */ |
| 1127 | if (likely(!__PageMovable(page))) |
| 1128 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
| 1129 | page_is_file_lru(page), -thp_nr_pages(page)); |
| 1130 | |
| 1131 | if (reason != MR_MEMORY_FAILURE) |
| 1132 | /* |
| 1133 | * We release the page in page_handle_poison. |
| 1134 | */ |
| 1135 | put_page(page); |
| 1136 | } else { |
| 1137 | if (rc != -EAGAIN) |
| 1138 | list_add_tail(&page->lru, ret); |
| 1139 | |
| 1140 | if (put_new_page) |
| 1141 | put_new_page(newpage, private); |
| 1142 | else |
| 1143 | put_page(newpage); |
| 1144 | } |
| 1145 | |
| 1146 | return rc; |
| 1147 | } |
| 1148 | |
| 1149 | /* |
| 1150 | * Counterpart of unmap_and_move_page() for hugepage migration. |
| 1151 | * |
| 1152 | * This function doesn't wait the completion of hugepage I/O |
| 1153 | * because there is no race between I/O and migration for hugepage. |
| 1154 | * Note that currently hugepage I/O occurs only in direct I/O |
| 1155 | * where no lock is held and PG_writeback is irrelevant, |
| 1156 | * and writeback status of all subpages are counted in the reference |
| 1157 | * count of the head page (i.e. if all subpages of a 2MB hugepage are |
| 1158 | * under direct I/O, the reference of the head page is 512 and a bit more.) |
| 1159 | * This means that when we try to migrate hugepage whose subpages are |
| 1160 | * doing direct I/O, some references remain after try_to_unmap() and |
| 1161 | * hugepage migration fails without data corruption. |
| 1162 | * |
| 1163 | * There is also no race when direct I/O is issued on the page under migration, |
| 1164 | * because then pte is replaced with migration swap entry and direct I/O code |
| 1165 | * will wait in the page fault for migration to complete. |
| 1166 | */ |
| 1167 | static int unmap_and_move_huge_page(new_page_t get_new_page, |
| 1168 | free_page_t put_new_page, unsigned long private, |
| 1169 | struct page *hpage, int force, |
| 1170 | enum migrate_mode mode, int reason, |
| 1171 | struct list_head *ret) |
| 1172 | { |
| 1173 | struct folio *dst, *src = page_folio(hpage); |
| 1174 | int rc = -EAGAIN; |
| 1175 | int page_was_mapped = 0; |
| 1176 | struct page *new_hpage; |
| 1177 | struct anon_vma *anon_vma = NULL; |
| 1178 | struct address_space *mapping = NULL; |
| 1179 | |
| 1180 | /* |
| 1181 | * Migratability of hugepages depends on architectures and their size. |
| 1182 | * This check is necessary because some callers of hugepage migration |
| 1183 | * like soft offline and memory hotremove don't walk through page |
| 1184 | * tables or check whether the hugepage is pmd-based or not before |
| 1185 | * kicking migration. |
| 1186 | */ |
| 1187 | if (!hugepage_migration_supported(page_hstate(hpage))) { |
| 1188 | list_move_tail(&hpage->lru, ret); |
| 1189 | return -ENOSYS; |
| 1190 | } |
| 1191 | |
| 1192 | if (page_count(hpage) == 1) { |
| 1193 | /* page was freed from under us. So we are done. */ |
| 1194 | putback_active_hugepage(hpage); |
| 1195 | return MIGRATEPAGE_SUCCESS; |
| 1196 | } |
| 1197 | |
| 1198 | new_hpage = get_new_page(hpage, private); |
| 1199 | if (!new_hpage) |
| 1200 | return -ENOMEM; |
| 1201 | dst = page_folio(new_hpage); |
| 1202 | |
| 1203 | if (!trylock_page(hpage)) { |
| 1204 | if (!force) |
| 1205 | goto out; |
| 1206 | switch (mode) { |
| 1207 | case MIGRATE_SYNC: |
| 1208 | case MIGRATE_SYNC_NO_COPY: |
| 1209 | break; |
| 1210 | default: |
| 1211 | goto out; |
| 1212 | } |
| 1213 | lock_page(hpage); |
| 1214 | } |
| 1215 | |
| 1216 | /* |
| 1217 | * Check for pages which are in the process of being freed. Without |
| 1218 | * page_mapping() set, hugetlbfs specific move page routine will not |
| 1219 | * be called and we could leak usage counts for subpools. |
| 1220 | */ |
| 1221 | if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) { |
| 1222 | rc = -EBUSY; |
| 1223 | goto out_unlock; |
| 1224 | } |
| 1225 | |
| 1226 | if (PageAnon(hpage)) |
| 1227 | anon_vma = page_get_anon_vma(hpage); |
| 1228 | |
| 1229 | if (unlikely(!trylock_page(new_hpage))) |
| 1230 | goto put_anon; |
| 1231 | |
| 1232 | if (page_mapped(hpage)) { |
| 1233 | bool mapping_locked = false; |
| 1234 | enum ttu_flags ttu = 0; |
| 1235 | |
| 1236 | if (!PageAnon(hpage)) { |
| 1237 | /* |
| 1238 | * In shared mappings, try_to_unmap could potentially |
| 1239 | * call huge_pmd_unshare. Because of this, take |
| 1240 | * semaphore in write mode here and set TTU_RMAP_LOCKED |
| 1241 | * to let lower levels know we have taken the lock. |
| 1242 | */ |
| 1243 | mapping = hugetlb_page_mapping_lock_write(hpage); |
| 1244 | if (unlikely(!mapping)) |
| 1245 | goto unlock_put_anon; |
| 1246 | |
| 1247 | mapping_locked = true; |
| 1248 | ttu |= TTU_RMAP_LOCKED; |
| 1249 | } |
| 1250 | |
| 1251 | try_to_migrate(src, ttu); |
| 1252 | page_was_mapped = 1; |
| 1253 | |
| 1254 | if (mapping_locked) |
| 1255 | i_mmap_unlock_write(mapping); |
| 1256 | } |
| 1257 | |
| 1258 | if (!page_mapped(hpage)) |
| 1259 | rc = move_to_new_page(new_hpage, hpage, mode); |
| 1260 | |
| 1261 | if (page_was_mapped) |
| 1262 | remove_migration_ptes(src, |
| 1263 | rc == MIGRATEPAGE_SUCCESS ? dst : src, false); |
| 1264 | |
| 1265 | unlock_put_anon: |
| 1266 | unlock_page(new_hpage); |
| 1267 | |
| 1268 | put_anon: |
| 1269 | if (anon_vma) |
| 1270 | put_anon_vma(anon_vma); |
| 1271 | |
| 1272 | if (rc == MIGRATEPAGE_SUCCESS) { |
| 1273 | move_hugetlb_state(hpage, new_hpage, reason); |
| 1274 | put_new_page = NULL; |
| 1275 | } |
| 1276 | |
| 1277 | out_unlock: |
| 1278 | unlock_page(hpage); |
| 1279 | out: |
| 1280 | if (rc == MIGRATEPAGE_SUCCESS) |
| 1281 | putback_active_hugepage(hpage); |
| 1282 | else if (rc != -EAGAIN) |
| 1283 | list_move_tail(&hpage->lru, ret); |
| 1284 | |
| 1285 | /* |
| 1286 | * If migration was not successful and there's a freeing callback, use |
| 1287 | * it. Otherwise, put_page() will drop the reference grabbed during |
| 1288 | * isolation. |
| 1289 | */ |
| 1290 | if (put_new_page) |
| 1291 | put_new_page(new_hpage, private); |
| 1292 | else |
| 1293 | putback_active_hugepage(new_hpage); |
| 1294 | |
| 1295 | return rc; |
| 1296 | } |
| 1297 | |
| 1298 | static inline int try_split_thp(struct page *page, struct page **page2, |
| 1299 | struct list_head *from) |
| 1300 | { |
| 1301 | int rc = 0; |
| 1302 | |
| 1303 | lock_page(page); |
| 1304 | rc = split_huge_page_to_list(page, from); |
| 1305 | unlock_page(page); |
| 1306 | if (!rc) |
| 1307 | list_safe_reset_next(page, *page2, lru); |
| 1308 | |
| 1309 | return rc; |
| 1310 | } |
| 1311 | |
| 1312 | /* |
| 1313 | * migrate_pages - migrate the pages specified in a list, to the free pages |
| 1314 | * supplied as the target for the page migration |
| 1315 | * |
| 1316 | * @from: The list of pages to be migrated. |
| 1317 | * @get_new_page: The function used to allocate free pages to be used |
| 1318 | * as the target of the page migration. |
| 1319 | * @put_new_page: The function used to free target pages if migration |
| 1320 | * fails, or NULL if no special handling is necessary. |
| 1321 | * @private: Private data to be passed on to get_new_page() |
| 1322 | * @mode: The migration mode that specifies the constraints for |
| 1323 | * page migration, if any. |
| 1324 | * @reason: The reason for page migration. |
| 1325 | * @ret_succeeded: Set to the number of normal pages migrated successfully if |
| 1326 | * the caller passes a non-NULL pointer. |
| 1327 | * |
| 1328 | * The function returns after 10 attempts or if no pages are movable any more |
| 1329 | * because the list has become empty or no retryable pages exist any more. |
| 1330 | * It is caller's responsibility to call putback_movable_pages() to return pages |
| 1331 | * to the LRU or free list only if ret != 0. |
| 1332 | * |
| 1333 | * Returns the number of {normal page, THP, hugetlb} that were not migrated, or |
| 1334 | * an error code. The number of THP splits will be considered as the number of |
| 1335 | * non-migrated THP, no matter how many subpages of the THP are migrated successfully. |
| 1336 | */ |
| 1337 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
| 1338 | free_page_t put_new_page, unsigned long private, |
| 1339 | enum migrate_mode mode, int reason, unsigned int *ret_succeeded) |
| 1340 | { |
| 1341 | int retry = 1; |
| 1342 | int thp_retry = 1; |
| 1343 | int nr_failed = 0; |
| 1344 | int nr_failed_pages = 0; |
| 1345 | int nr_succeeded = 0; |
| 1346 | int nr_thp_succeeded = 0; |
| 1347 | int nr_thp_failed = 0; |
| 1348 | int nr_thp_split = 0; |
| 1349 | int pass = 0; |
| 1350 | bool is_thp = false; |
| 1351 | struct page *page; |
| 1352 | struct page *page2; |
| 1353 | int rc, nr_subpages; |
| 1354 | LIST_HEAD(ret_pages); |
| 1355 | LIST_HEAD(thp_split_pages); |
| 1356 | bool nosplit = (reason == MR_NUMA_MISPLACED); |
| 1357 | bool no_subpage_counting = false; |
| 1358 | |
| 1359 | trace_mm_migrate_pages_start(mode, reason); |
| 1360 | |
| 1361 | thp_subpage_migration: |
| 1362 | for (pass = 0; pass < 10 && (retry || thp_retry); pass++) { |
| 1363 | retry = 0; |
| 1364 | thp_retry = 0; |
| 1365 | |
| 1366 | list_for_each_entry_safe(page, page2, from, lru) { |
| 1367 | retry: |
| 1368 | /* |
| 1369 | * THP statistics is based on the source huge page. |
| 1370 | * Capture required information that might get lost |
| 1371 | * during migration. |
| 1372 | */ |
| 1373 | is_thp = PageTransHuge(page) && !PageHuge(page); |
| 1374 | nr_subpages = compound_nr(page); |
| 1375 | cond_resched(); |
| 1376 | |
| 1377 | if (PageHuge(page)) |
| 1378 | rc = unmap_and_move_huge_page(get_new_page, |
| 1379 | put_new_page, private, page, |
| 1380 | pass > 2, mode, reason, |
| 1381 | &ret_pages); |
| 1382 | else |
| 1383 | rc = unmap_and_move(get_new_page, put_new_page, |
| 1384 | private, page, pass > 2, mode, |
| 1385 | reason, &ret_pages); |
| 1386 | /* |
| 1387 | * The rules are: |
| 1388 | * Success: non hugetlb page will be freed, hugetlb |
| 1389 | * page will be put back |
| 1390 | * -EAGAIN: stay on the from list |
| 1391 | * -ENOMEM: stay on the from list |
| 1392 | * Other errno: put on ret_pages list then splice to |
| 1393 | * from list |
| 1394 | */ |
| 1395 | switch(rc) { |
| 1396 | /* |
| 1397 | * THP migration might be unsupported or the |
| 1398 | * allocation could've failed so we should |
| 1399 | * retry on the same page with the THP split |
| 1400 | * to base pages. |
| 1401 | * |
| 1402 | * Head page is retried immediately and tail |
| 1403 | * pages are added to the tail of the list so |
| 1404 | * we encounter them after the rest of the list |
| 1405 | * is processed. |
| 1406 | */ |
| 1407 | case -ENOSYS: |
| 1408 | /* THP migration is unsupported */ |
| 1409 | if (is_thp) { |
| 1410 | nr_thp_failed++; |
| 1411 | if (!try_split_thp(page, &page2, &thp_split_pages)) { |
| 1412 | nr_thp_split++; |
| 1413 | goto retry; |
| 1414 | } |
| 1415 | |
| 1416 | nr_failed_pages += nr_subpages; |
| 1417 | break; |
| 1418 | } |
| 1419 | |
| 1420 | /* Hugetlb migration is unsupported */ |
| 1421 | if (!no_subpage_counting) |
| 1422 | nr_failed++; |
| 1423 | nr_failed_pages += nr_subpages; |
| 1424 | break; |
| 1425 | case -ENOMEM: |
| 1426 | /* |
| 1427 | * When memory is low, don't bother to try to migrate |
| 1428 | * other pages, just exit. |
| 1429 | * THP NUMA faulting doesn't split THP to retry. |
| 1430 | */ |
| 1431 | if (is_thp && !nosplit) { |
| 1432 | nr_thp_failed++; |
| 1433 | if (!try_split_thp(page, &page2, &thp_split_pages)) { |
| 1434 | nr_thp_split++; |
| 1435 | goto retry; |
| 1436 | } |
| 1437 | |
| 1438 | nr_failed_pages += nr_subpages; |
| 1439 | goto out; |
| 1440 | } |
| 1441 | |
| 1442 | if (!no_subpage_counting) |
| 1443 | nr_failed++; |
| 1444 | nr_failed_pages += nr_subpages; |
| 1445 | goto out; |
| 1446 | case -EAGAIN: |
| 1447 | if (is_thp) { |
| 1448 | thp_retry++; |
| 1449 | break; |
| 1450 | } |
| 1451 | retry++; |
| 1452 | break; |
| 1453 | case MIGRATEPAGE_SUCCESS: |
| 1454 | nr_succeeded += nr_subpages; |
| 1455 | if (is_thp) { |
| 1456 | nr_thp_succeeded++; |
| 1457 | break; |
| 1458 | } |
| 1459 | break; |
| 1460 | default: |
| 1461 | /* |
| 1462 | * Permanent failure (-EBUSY, etc.): |
| 1463 | * unlike -EAGAIN case, the failed page is |
| 1464 | * removed from migration page list and not |
| 1465 | * retried in the next outer loop. |
| 1466 | */ |
| 1467 | if (is_thp) { |
| 1468 | nr_thp_failed++; |
| 1469 | nr_failed_pages += nr_subpages; |
| 1470 | break; |
| 1471 | } |
| 1472 | |
| 1473 | if (!no_subpage_counting) |
| 1474 | nr_failed++; |
| 1475 | nr_failed_pages += nr_subpages; |
| 1476 | break; |
| 1477 | } |
| 1478 | } |
| 1479 | } |
| 1480 | nr_failed += retry; |
| 1481 | nr_thp_failed += thp_retry; |
| 1482 | /* |
| 1483 | * Try to migrate subpages of fail-to-migrate THPs, no nr_failed |
| 1484 | * counting in this round, since all subpages of a THP is counted |
| 1485 | * as 1 failure in the first round. |
| 1486 | */ |
| 1487 | if (!list_empty(&thp_split_pages)) { |
| 1488 | /* |
| 1489 | * Move non-migrated pages (after 10 retries) to ret_pages |
| 1490 | * to avoid migrating them again. |
| 1491 | */ |
| 1492 | list_splice_init(from, &ret_pages); |
| 1493 | list_splice_init(&thp_split_pages, from); |
| 1494 | no_subpage_counting = true; |
| 1495 | retry = 1; |
| 1496 | goto thp_subpage_migration; |
| 1497 | } |
| 1498 | |
| 1499 | rc = nr_failed + nr_thp_failed; |
| 1500 | out: |
| 1501 | /* |
| 1502 | * Put the permanent failure page back to migration list, they |
| 1503 | * will be put back to the right list by the caller. |
| 1504 | */ |
| 1505 | list_splice(&ret_pages, from); |
| 1506 | |
| 1507 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); |
| 1508 | count_vm_events(PGMIGRATE_FAIL, nr_failed_pages); |
| 1509 | count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded); |
| 1510 | count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed); |
| 1511 | count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split); |
| 1512 | trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded, |
| 1513 | nr_thp_failed, nr_thp_split, mode, reason); |
| 1514 | |
| 1515 | if (ret_succeeded) |
| 1516 | *ret_succeeded = nr_succeeded; |
| 1517 | |
| 1518 | return rc; |
| 1519 | } |
| 1520 | |
| 1521 | struct page *alloc_migration_target(struct page *page, unsigned long private) |
| 1522 | { |
| 1523 | struct migration_target_control *mtc; |
| 1524 | gfp_t gfp_mask; |
| 1525 | unsigned int order = 0; |
| 1526 | struct page *new_page = NULL; |
| 1527 | int nid; |
| 1528 | int zidx; |
| 1529 | |
| 1530 | mtc = (struct migration_target_control *)private; |
| 1531 | gfp_mask = mtc->gfp_mask; |
| 1532 | nid = mtc->nid; |
| 1533 | if (nid == NUMA_NO_NODE) |
| 1534 | nid = page_to_nid(page); |
| 1535 | |
| 1536 | if (PageHuge(page)) { |
| 1537 | struct hstate *h = page_hstate(compound_head(page)); |
| 1538 | |
| 1539 | gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); |
| 1540 | return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask); |
| 1541 | } |
| 1542 | |
| 1543 | if (PageTransHuge(page)) { |
| 1544 | /* |
| 1545 | * clear __GFP_RECLAIM to make the migration callback |
| 1546 | * consistent with regular THP allocations. |
| 1547 | */ |
| 1548 | gfp_mask &= ~__GFP_RECLAIM; |
| 1549 | gfp_mask |= GFP_TRANSHUGE; |
| 1550 | order = HPAGE_PMD_ORDER; |
| 1551 | } |
| 1552 | zidx = zone_idx(page_zone(page)); |
| 1553 | if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) |
| 1554 | gfp_mask |= __GFP_HIGHMEM; |
| 1555 | |
| 1556 | new_page = __alloc_pages(gfp_mask, order, nid, mtc->nmask); |
| 1557 | |
| 1558 | if (new_page && PageTransHuge(new_page)) |
| 1559 | prep_transhuge_page(new_page); |
| 1560 | |
| 1561 | return new_page; |
| 1562 | } |
| 1563 | |
| 1564 | #ifdef CONFIG_NUMA |
| 1565 | |
| 1566 | static int store_status(int __user *status, int start, int value, int nr) |
| 1567 | { |
| 1568 | while (nr-- > 0) { |
| 1569 | if (put_user(value, status + start)) |
| 1570 | return -EFAULT; |
| 1571 | start++; |
| 1572 | } |
| 1573 | |
| 1574 | return 0; |
| 1575 | } |
| 1576 | |
| 1577 | static int do_move_pages_to_node(struct mm_struct *mm, |
| 1578 | struct list_head *pagelist, int node) |
| 1579 | { |
| 1580 | int err; |
| 1581 | struct migration_target_control mtc = { |
| 1582 | .nid = node, |
| 1583 | .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, |
| 1584 | }; |
| 1585 | |
| 1586 | err = migrate_pages(pagelist, alloc_migration_target, NULL, |
| 1587 | (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL); |
| 1588 | if (err) |
| 1589 | putback_movable_pages(pagelist); |
| 1590 | return err; |
| 1591 | } |
| 1592 | |
| 1593 | /* |
| 1594 | * Resolves the given address to a struct page, isolates it from the LRU and |
| 1595 | * puts it to the given pagelist. |
| 1596 | * Returns: |
| 1597 | * errno - if the page cannot be found/isolated |
| 1598 | * 0 - when it doesn't have to be migrated because it is already on the |
| 1599 | * target node |
| 1600 | * 1 - when it has been queued |
| 1601 | */ |
| 1602 | static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, |
| 1603 | int node, struct list_head *pagelist, bool migrate_all) |
| 1604 | { |
| 1605 | struct vm_area_struct *vma; |
| 1606 | struct page *page; |
| 1607 | int err; |
| 1608 | |
| 1609 | mmap_read_lock(mm); |
| 1610 | err = -EFAULT; |
| 1611 | vma = find_vma(mm, addr); |
| 1612 | if (!vma || addr < vma->vm_start || !vma_migratable(vma)) |
| 1613 | goto out; |
| 1614 | |
| 1615 | /* FOLL_DUMP to ignore special (like zero) pages */ |
| 1616 | page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); |
| 1617 | |
| 1618 | err = PTR_ERR(page); |
| 1619 | if (IS_ERR(page)) |
| 1620 | goto out; |
| 1621 | |
| 1622 | err = -ENOENT; |
| 1623 | if (!page) |
| 1624 | goto out; |
| 1625 | |
| 1626 | err = 0; |
| 1627 | if (page_to_nid(page) == node) |
| 1628 | goto out_putpage; |
| 1629 | |
| 1630 | err = -EACCES; |
| 1631 | if (page_mapcount(page) > 1 && !migrate_all) |
| 1632 | goto out_putpage; |
| 1633 | |
| 1634 | if (PageHuge(page)) { |
| 1635 | if (PageHead(page)) { |
| 1636 | isolate_huge_page(page, pagelist); |
| 1637 | err = 1; |
| 1638 | } |
| 1639 | } else { |
| 1640 | struct page *head; |
| 1641 | |
| 1642 | head = compound_head(page); |
| 1643 | err = isolate_lru_page(head); |
| 1644 | if (err) |
| 1645 | goto out_putpage; |
| 1646 | |
| 1647 | err = 1; |
| 1648 | list_add_tail(&head->lru, pagelist); |
| 1649 | mod_node_page_state(page_pgdat(head), |
| 1650 | NR_ISOLATED_ANON + page_is_file_lru(head), |
| 1651 | thp_nr_pages(head)); |
| 1652 | } |
| 1653 | out_putpage: |
| 1654 | /* |
| 1655 | * Either remove the duplicate refcount from |
| 1656 | * isolate_lru_page() or drop the page ref if it was |
| 1657 | * not isolated. |
| 1658 | */ |
| 1659 | put_page(page); |
| 1660 | out: |
| 1661 | mmap_read_unlock(mm); |
| 1662 | return err; |
| 1663 | } |
| 1664 | |
| 1665 | static int move_pages_and_store_status(struct mm_struct *mm, int node, |
| 1666 | struct list_head *pagelist, int __user *status, |
| 1667 | int start, int i, unsigned long nr_pages) |
| 1668 | { |
| 1669 | int err; |
| 1670 | |
| 1671 | if (list_empty(pagelist)) |
| 1672 | return 0; |
| 1673 | |
| 1674 | err = do_move_pages_to_node(mm, pagelist, node); |
| 1675 | if (err) { |
| 1676 | /* |
| 1677 | * Positive err means the number of failed |
| 1678 | * pages to migrate. Since we are going to |
| 1679 | * abort and return the number of non-migrated |
| 1680 | * pages, so need to include the rest of the |
| 1681 | * nr_pages that have not been attempted as |
| 1682 | * well. |
| 1683 | */ |
| 1684 | if (err > 0) |
| 1685 | err += nr_pages - i - 1; |
| 1686 | return err; |
| 1687 | } |
| 1688 | return store_status(status, start, node, i - start); |
| 1689 | } |
| 1690 | |
| 1691 | /* |
| 1692 | * Migrate an array of page address onto an array of nodes and fill |
| 1693 | * the corresponding array of status. |
| 1694 | */ |
| 1695 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
| 1696 | unsigned long nr_pages, |
| 1697 | const void __user * __user *pages, |
| 1698 | const int __user *nodes, |
| 1699 | int __user *status, int flags) |
| 1700 | { |
| 1701 | int current_node = NUMA_NO_NODE; |
| 1702 | LIST_HEAD(pagelist); |
| 1703 | int start, i; |
| 1704 | int err = 0, err1; |
| 1705 | |
| 1706 | lru_cache_disable(); |
| 1707 | |
| 1708 | for (i = start = 0; i < nr_pages; i++) { |
| 1709 | const void __user *p; |
| 1710 | unsigned long addr; |
| 1711 | int node; |
| 1712 | |
| 1713 | err = -EFAULT; |
| 1714 | if (get_user(p, pages + i)) |
| 1715 | goto out_flush; |
| 1716 | if (get_user(node, nodes + i)) |
| 1717 | goto out_flush; |
| 1718 | addr = (unsigned long)untagged_addr(p); |
| 1719 | |
| 1720 | err = -ENODEV; |
| 1721 | if (node < 0 || node >= MAX_NUMNODES) |
| 1722 | goto out_flush; |
| 1723 | if (!node_state(node, N_MEMORY)) |
| 1724 | goto out_flush; |
| 1725 | |
| 1726 | err = -EACCES; |
| 1727 | if (!node_isset(node, task_nodes)) |
| 1728 | goto out_flush; |
| 1729 | |
| 1730 | if (current_node == NUMA_NO_NODE) { |
| 1731 | current_node = node; |
| 1732 | start = i; |
| 1733 | } else if (node != current_node) { |
| 1734 | err = move_pages_and_store_status(mm, current_node, |
| 1735 | &pagelist, status, start, i, nr_pages); |
| 1736 | if (err) |
| 1737 | goto out; |
| 1738 | start = i; |
| 1739 | current_node = node; |
| 1740 | } |
| 1741 | |
| 1742 | /* |
| 1743 | * Errors in the page lookup or isolation are not fatal and we simply |
| 1744 | * report them via status |
| 1745 | */ |
| 1746 | err = add_page_for_migration(mm, addr, current_node, |
| 1747 | &pagelist, flags & MPOL_MF_MOVE_ALL); |
| 1748 | |
| 1749 | if (err > 0) { |
| 1750 | /* The page is successfully queued for migration */ |
| 1751 | continue; |
| 1752 | } |
| 1753 | |
| 1754 | /* |
| 1755 | * The move_pages() man page does not have an -EEXIST choice, so |
| 1756 | * use -EFAULT instead. |
| 1757 | */ |
| 1758 | if (err == -EEXIST) |
| 1759 | err = -EFAULT; |
| 1760 | |
| 1761 | /* |
| 1762 | * If the page is already on the target node (!err), store the |
| 1763 | * node, otherwise, store the err. |
| 1764 | */ |
| 1765 | err = store_status(status, i, err ? : current_node, 1); |
| 1766 | if (err) |
| 1767 | goto out_flush; |
| 1768 | |
| 1769 | err = move_pages_and_store_status(mm, current_node, &pagelist, |
| 1770 | status, start, i, nr_pages); |
| 1771 | if (err) |
| 1772 | goto out; |
| 1773 | current_node = NUMA_NO_NODE; |
| 1774 | } |
| 1775 | out_flush: |
| 1776 | /* Make sure we do not overwrite the existing error */ |
| 1777 | err1 = move_pages_and_store_status(mm, current_node, &pagelist, |
| 1778 | status, start, i, nr_pages); |
| 1779 | if (err >= 0) |
| 1780 | err = err1; |
| 1781 | out: |
| 1782 | lru_cache_enable(); |
| 1783 | return err; |
| 1784 | } |
| 1785 | |
| 1786 | /* |
| 1787 | * Determine the nodes of an array of pages and store it in an array of status. |
| 1788 | */ |
| 1789 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
| 1790 | const void __user **pages, int *status) |
| 1791 | { |
| 1792 | unsigned long i; |
| 1793 | |
| 1794 | mmap_read_lock(mm); |
| 1795 | |
| 1796 | for (i = 0; i < nr_pages; i++) { |
| 1797 | unsigned long addr = (unsigned long)(*pages); |
| 1798 | struct vm_area_struct *vma; |
| 1799 | struct page *page; |
| 1800 | int err = -EFAULT; |
| 1801 | |
| 1802 | vma = vma_lookup(mm, addr); |
| 1803 | if (!vma) |
| 1804 | goto set_status; |
| 1805 | |
| 1806 | /* FOLL_DUMP to ignore special (like zero) pages */ |
| 1807 | page = follow_page(vma, addr, FOLL_DUMP); |
| 1808 | |
| 1809 | err = PTR_ERR(page); |
| 1810 | if (IS_ERR(page)) |
| 1811 | goto set_status; |
| 1812 | |
| 1813 | err = page ? page_to_nid(page) : -ENOENT; |
| 1814 | set_status: |
| 1815 | *status = err; |
| 1816 | |
| 1817 | pages++; |
| 1818 | status++; |
| 1819 | } |
| 1820 | |
| 1821 | mmap_read_unlock(mm); |
| 1822 | } |
| 1823 | |
| 1824 | static int get_compat_pages_array(const void __user *chunk_pages[], |
| 1825 | const void __user * __user *pages, |
| 1826 | unsigned long chunk_nr) |
| 1827 | { |
| 1828 | compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages; |
| 1829 | compat_uptr_t p; |
| 1830 | int i; |
| 1831 | |
| 1832 | for (i = 0; i < chunk_nr; i++) { |
| 1833 | if (get_user(p, pages32 + i)) |
| 1834 | return -EFAULT; |
| 1835 | chunk_pages[i] = compat_ptr(p); |
| 1836 | } |
| 1837 | |
| 1838 | return 0; |
| 1839 | } |
| 1840 | |
| 1841 | /* |
| 1842 | * Determine the nodes of a user array of pages and store it in |
| 1843 | * a user array of status. |
| 1844 | */ |
| 1845 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, |
| 1846 | const void __user * __user *pages, |
| 1847 | int __user *status) |
| 1848 | { |
| 1849 | #define DO_PAGES_STAT_CHUNK_NR 16 |
| 1850 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; |
| 1851 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; |
| 1852 | |
| 1853 | while (nr_pages) { |
| 1854 | unsigned long chunk_nr; |
| 1855 | |
| 1856 | chunk_nr = nr_pages; |
| 1857 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) |
| 1858 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; |
| 1859 | |
| 1860 | if (in_compat_syscall()) { |
| 1861 | if (get_compat_pages_array(chunk_pages, pages, |
| 1862 | chunk_nr)) |
| 1863 | break; |
| 1864 | } else { |
| 1865 | if (copy_from_user(chunk_pages, pages, |
| 1866 | chunk_nr * sizeof(*chunk_pages))) |
| 1867 | break; |
| 1868 | } |
| 1869 | |
| 1870 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); |
| 1871 | |
| 1872 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
| 1873 | break; |
| 1874 | |
| 1875 | pages += chunk_nr; |
| 1876 | status += chunk_nr; |
| 1877 | nr_pages -= chunk_nr; |
| 1878 | } |
| 1879 | return nr_pages ? -EFAULT : 0; |
| 1880 | } |
| 1881 | |
| 1882 | static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) |
| 1883 | { |
| 1884 | struct task_struct *task; |
| 1885 | struct mm_struct *mm; |
| 1886 | |
| 1887 | /* |
| 1888 | * There is no need to check if current process has the right to modify |
| 1889 | * the specified process when they are same. |
| 1890 | */ |
| 1891 | if (!pid) { |
| 1892 | mmget(current->mm); |
| 1893 | *mem_nodes = cpuset_mems_allowed(current); |
| 1894 | return current->mm; |
| 1895 | } |
| 1896 | |
| 1897 | /* Find the mm_struct */ |
| 1898 | rcu_read_lock(); |
| 1899 | task = find_task_by_vpid(pid); |
| 1900 | if (!task) { |
| 1901 | rcu_read_unlock(); |
| 1902 | return ERR_PTR(-ESRCH); |
| 1903 | } |
| 1904 | get_task_struct(task); |
| 1905 | |
| 1906 | /* |
| 1907 | * Check if this process has the right to modify the specified |
| 1908 | * process. Use the regular "ptrace_may_access()" checks. |
| 1909 | */ |
| 1910 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { |
| 1911 | rcu_read_unlock(); |
| 1912 | mm = ERR_PTR(-EPERM); |
| 1913 | goto out; |
| 1914 | } |
| 1915 | rcu_read_unlock(); |
| 1916 | |
| 1917 | mm = ERR_PTR(security_task_movememory(task)); |
| 1918 | if (IS_ERR(mm)) |
| 1919 | goto out; |
| 1920 | *mem_nodes = cpuset_mems_allowed(task); |
| 1921 | mm = get_task_mm(task); |
| 1922 | out: |
| 1923 | put_task_struct(task); |
| 1924 | if (!mm) |
| 1925 | mm = ERR_PTR(-EINVAL); |
| 1926 | return mm; |
| 1927 | } |
| 1928 | |
| 1929 | /* |
| 1930 | * Move a list of pages in the address space of the currently executing |
| 1931 | * process. |
| 1932 | */ |
| 1933 | static int kernel_move_pages(pid_t pid, unsigned long nr_pages, |
| 1934 | const void __user * __user *pages, |
| 1935 | const int __user *nodes, |
| 1936 | int __user *status, int flags) |
| 1937 | { |
| 1938 | struct mm_struct *mm; |
| 1939 | int err; |
| 1940 | nodemask_t task_nodes; |
| 1941 | |
| 1942 | /* Check flags */ |
| 1943 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) |
| 1944 | return -EINVAL; |
| 1945 | |
| 1946 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
| 1947 | return -EPERM; |
| 1948 | |
| 1949 | mm = find_mm_struct(pid, &task_nodes); |
| 1950 | if (IS_ERR(mm)) |
| 1951 | return PTR_ERR(mm); |
| 1952 | |
| 1953 | if (nodes) |
| 1954 | err = do_pages_move(mm, task_nodes, nr_pages, pages, |
| 1955 | nodes, status, flags); |
| 1956 | else |
| 1957 | err = do_pages_stat(mm, nr_pages, pages, status); |
| 1958 | |
| 1959 | mmput(mm); |
| 1960 | return err; |
| 1961 | } |
| 1962 | |
| 1963 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
| 1964 | const void __user * __user *, pages, |
| 1965 | const int __user *, nodes, |
| 1966 | int __user *, status, int, flags) |
| 1967 | { |
| 1968 | return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); |
| 1969 | } |
| 1970 | |
| 1971 | #ifdef CONFIG_NUMA_BALANCING |
| 1972 | /* |
| 1973 | * Returns true if this is a safe migration target node for misplaced NUMA |
| 1974 | * pages. Currently it only checks the watermarks which crude |
| 1975 | */ |
| 1976 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, |
| 1977 | unsigned long nr_migrate_pages) |
| 1978 | { |
| 1979 | int z; |
| 1980 | |
| 1981 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { |
| 1982 | struct zone *zone = pgdat->node_zones + z; |
| 1983 | |
| 1984 | if (!populated_zone(zone)) |
| 1985 | continue; |
| 1986 | |
| 1987 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ |
| 1988 | if (!zone_watermark_ok(zone, 0, |
| 1989 | high_wmark_pages(zone) + |
| 1990 | nr_migrate_pages, |
| 1991 | ZONE_MOVABLE, 0)) |
| 1992 | continue; |
| 1993 | return true; |
| 1994 | } |
| 1995 | return false; |
| 1996 | } |
| 1997 | |
| 1998 | static struct page *alloc_misplaced_dst_page(struct page *page, |
| 1999 | unsigned long data) |
| 2000 | { |
| 2001 | int nid = (int) data; |
| 2002 | struct page *newpage; |
| 2003 | |
| 2004 | newpage = __alloc_pages_node(nid, |
| 2005 | (GFP_HIGHUSER_MOVABLE | |
| 2006 | __GFP_THISNODE | __GFP_NOMEMALLOC | |
| 2007 | __GFP_NORETRY | __GFP_NOWARN) & |
| 2008 | ~__GFP_RECLAIM, 0); |
| 2009 | |
| 2010 | return newpage; |
| 2011 | } |
| 2012 | |
| 2013 | static struct page *alloc_misplaced_dst_page_thp(struct page *page, |
| 2014 | unsigned long data) |
| 2015 | { |
| 2016 | int nid = (int) data; |
| 2017 | struct page *newpage; |
| 2018 | |
| 2019 | newpage = alloc_pages_node(nid, (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), |
| 2020 | HPAGE_PMD_ORDER); |
| 2021 | if (!newpage) |
| 2022 | goto out; |
| 2023 | |
| 2024 | prep_transhuge_page(newpage); |
| 2025 | |
| 2026 | out: |
| 2027 | return newpage; |
| 2028 | } |
| 2029 | |
| 2030 | static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) |
| 2031 | { |
| 2032 | int page_lru; |
| 2033 | int nr_pages = thp_nr_pages(page); |
| 2034 | int order = compound_order(page); |
| 2035 | |
| 2036 | VM_BUG_ON_PAGE(order && !PageTransHuge(page), page); |
| 2037 | |
| 2038 | /* Do not migrate THP mapped by multiple processes */ |
| 2039 | if (PageTransHuge(page) && total_mapcount(page) > 1) |
| 2040 | return 0; |
| 2041 | |
| 2042 | /* Avoid migrating to a node that is nearly full */ |
| 2043 | if (!migrate_balanced_pgdat(pgdat, nr_pages)) { |
| 2044 | int z; |
| 2045 | |
| 2046 | if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)) |
| 2047 | return 0; |
| 2048 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { |
| 2049 | if (populated_zone(pgdat->node_zones + z)) |
| 2050 | break; |
| 2051 | } |
| 2052 | wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE); |
| 2053 | return 0; |
| 2054 | } |
| 2055 | |
| 2056 | if (isolate_lru_page(page)) |
| 2057 | return 0; |
| 2058 | |
| 2059 | page_lru = page_is_file_lru(page); |
| 2060 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, |
| 2061 | nr_pages); |
| 2062 | |
| 2063 | /* |
| 2064 | * Isolating the page has taken another reference, so the |
| 2065 | * caller's reference can be safely dropped without the page |
| 2066 | * disappearing underneath us during migration. |
| 2067 | */ |
| 2068 | put_page(page); |
| 2069 | return 1; |
| 2070 | } |
| 2071 | |
| 2072 | /* |
| 2073 | * Attempt to migrate a misplaced page to the specified destination |
| 2074 | * node. Caller is expected to have an elevated reference count on |
| 2075 | * the page that will be dropped by this function before returning. |
| 2076 | */ |
| 2077 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
| 2078 | int node) |
| 2079 | { |
| 2080 | pg_data_t *pgdat = NODE_DATA(node); |
| 2081 | int isolated; |
| 2082 | int nr_remaining; |
| 2083 | unsigned int nr_succeeded; |
| 2084 | LIST_HEAD(migratepages); |
| 2085 | new_page_t *new; |
| 2086 | bool compound; |
| 2087 | int nr_pages = thp_nr_pages(page); |
| 2088 | |
| 2089 | /* |
| 2090 | * PTE mapped THP or HugeTLB page can't reach here so the page could |
| 2091 | * be either base page or THP. And it must be head page if it is |
| 2092 | * THP. |
| 2093 | */ |
| 2094 | compound = PageTransHuge(page); |
| 2095 | |
| 2096 | if (compound) |
| 2097 | new = alloc_misplaced_dst_page_thp; |
| 2098 | else |
| 2099 | new = alloc_misplaced_dst_page; |
| 2100 | |
| 2101 | /* |
| 2102 | * Don't migrate file pages that are mapped in multiple processes |
| 2103 | * with execute permissions as they are probably shared libraries. |
| 2104 | */ |
| 2105 | if (page_mapcount(page) != 1 && page_is_file_lru(page) && |
| 2106 | (vma->vm_flags & VM_EXEC)) |
| 2107 | goto out; |
| 2108 | |
| 2109 | /* |
| 2110 | * Also do not migrate dirty pages as not all filesystems can move |
| 2111 | * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. |
| 2112 | */ |
| 2113 | if (page_is_file_lru(page) && PageDirty(page)) |
| 2114 | goto out; |
| 2115 | |
| 2116 | isolated = numamigrate_isolate_page(pgdat, page); |
| 2117 | if (!isolated) |
| 2118 | goto out; |
| 2119 | |
| 2120 | list_add(&page->lru, &migratepages); |
| 2121 | nr_remaining = migrate_pages(&migratepages, *new, NULL, node, |
| 2122 | MIGRATE_ASYNC, MR_NUMA_MISPLACED, |
| 2123 | &nr_succeeded); |
| 2124 | if (nr_remaining) { |
| 2125 | if (!list_empty(&migratepages)) { |
| 2126 | list_del(&page->lru); |
| 2127 | mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + |
| 2128 | page_is_file_lru(page), -nr_pages); |
| 2129 | putback_lru_page(page); |
| 2130 | } |
| 2131 | isolated = 0; |
| 2132 | } |
| 2133 | if (nr_succeeded) { |
| 2134 | count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded); |
| 2135 | if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node)) |
| 2136 | mod_node_page_state(pgdat, PGPROMOTE_SUCCESS, |
| 2137 | nr_succeeded); |
| 2138 | } |
| 2139 | BUG_ON(!list_empty(&migratepages)); |
| 2140 | return isolated; |
| 2141 | |
| 2142 | out: |
| 2143 | put_page(page); |
| 2144 | return 0; |
| 2145 | } |
| 2146 | #endif /* CONFIG_NUMA_BALANCING */ |
| 2147 | #endif /* CONFIG_NUMA */ |
| 2148 | |
| 2149 | /* |
| 2150 | * node_demotion[] example: |
| 2151 | * |
| 2152 | * Consider a system with two sockets. Each socket has |
| 2153 | * three classes of memory attached: fast, medium and slow. |
| 2154 | * Each memory class is placed in its own NUMA node. The |
| 2155 | * CPUs are placed in the node with the "fast" memory. The |
| 2156 | * 6 NUMA nodes (0-5) might be split among the sockets like |
| 2157 | * this: |
| 2158 | * |
| 2159 | * Socket A: 0, 1, 2 |
| 2160 | * Socket B: 3, 4, 5 |
| 2161 | * |
| 2162 | * When Node 0 fills up, its memory should be migrated to |
| 2163 | * Node 1. When Node 1 fills up, it should be migrated to |
| 2164 | * Node 2. The migration path start on the nodes with the |
| 2165 | * processors (since allocations default to this node) and |
| 2166 | * fast memory, progress through medium and end with the |
| 2167 | * slow memory: |
| 2168 | * |
| 2169 | * 0 -> 1 -> 2 -> stop |
| 2170 | * 3 -> 4 -> 5 -> stop |
| 2171 | * |
| 2172 | * This is represented in the node_demotion[] like this: |
| 2173 | * |
| 2174 | * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1 |
| 2175 | * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2 |
| 2176 | * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate |
| 2177 | * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4 |
| 2178 | * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5 |
| 2179 | * { nr=0, nodes[0]=-1 }, // Node 5 does not migrate |
| 2180 | * |
| 2181 | * Moreover some systems may have multiple slow memory nodes. |
| 2182 | * Suppose a system has one socket with 3 memory nodes, node 0 |
| 2183 | * is fast memory type, and node 1/2 both are slow memory |
| 2184 | * type, and the distance between fast memory node and slow |
| 2185 | * memory node is same. So the migration path should be: |
| 2186 | * |
| 2187 | * 0 -> 1/2 -> stop |
| 2188 | * |
| 2189 | * This is represented in the node_demotion[] like this: |
| 2190 | * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2 |
| 2191 | * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate |
| 2192 | * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate |
| 2193 | */ |
| 2194 | |
| 2195 | /* |
| 2196 | * Writes to this array occur without locking. Cycles are |
| 2197 | * not allowed: Node X demotes to Y which demotes to X... |
| 2198 | * |
| 2199 | * If multiple reads are performed, a single rcu_read_lock() |
| 2200 | * must be held over all reads to ensure that no cycles are |
| 2201 | * observed. |
| 2202 | */ |
| 2203 | #define DEFAULT_DEMOTION_TARGET_NODES 15 |
| 2204 | |
| 2205 | #if MAX_NUMNODES < DEFAULT_DEMOTION_TARGET_NODES |
| 2206 | #define DEMOTION_TARGET_NODES (MAX_NUMNODES - 1) |
| 2207 | #else |
| 2208 | #define DEMOTION_TARGET_NODES DEFAULT_DEMOTION_TARGET_NODES |
| 2209 | #endif |
| 2210 | |
| 2211 | struct demotion_nodes { |
| 2212 | unsigned short nr; |
| 2213 | short nodes[DEMOTION_TARGET_NODES]; |
| 2214 | }; |
| 2215 | |
| 2216 | static struct demotion_nodes *node_demotion __read_mostly; |
| 2217 | |
| 2218 | /** |
| 2219 | * next_demotion_node() - Get the next node in the demotion path |
| 2220 | * @node: The starting node to lookup the next node |
| 2221 | * |
| 2222 | * Return: node id for next memory node in the demotion path hierarchy |
| 2223 | * from @node; NUMA_NO_NODE if @node is terminal. This does not keep |
| 2224 | * @node online or guarantee that it *continues* to be the next demotion |
| 2225 | * target. |
| 2226 | */ |
| 2227 | int next_demotion_node(int node) |
| 2228 | { |
| 2229 | struct demotion_nodes *nd; |
| 2230 | unsigned short target_nr, index; |
| 2231 | int target; |
| 2232 | |
| 2233 | if (!node_demotion) |
| 2234 | return NUMA_NO_NODE; |
| 2235 | |
| 2236 | nd = &node_demotion[node]; |
| 2237 | |
| 2238 | /* |
| 2239 | * node_demotion[] is updated without excluding this |
| 2240 | * function from running. RCU doesn't provide any |
| 2241 | * compiler barriers, so the READ_ONCE() is required |
| 2242 | * to avoid compiler reordering or read merging. |
| 2243 | * |
| 2244 | * Make sure to use RCU over entire code blocks if |
| 2245 | * node_demotion[] reads need to be consistent. |
| 2246 | */ |
| 2247 | rcu_read_lock(); |
| 2248 | target_nr = READ_ONCE(nd->nr); |
| 2249 | |
| 2250 | switch (target_nr) { |
| 2251 | case 0: |
| 2252 | target = NUMA_NO_NODE; |
| 2253 | goto out; |
| 2254 | case 1: |
| 2255 | index = 0; |
| 2256 | break; |
| 2257 | default: |
| 2258 | /* |
| 2259 | * If there are multiple target nodes, just select one |
| 2260 | * target node randomly. |
| 2261 | * |
| 2262 | * In addition, we can also use round-robin to select |
| 2263 | * target node, but we should introduce another variable |
| 2264 | * for node_demotion[] to record last selected target node, |
| 2265 | * that may cause cache ping-pong due to the changing of |
| 2266 | * last target node. Or introducing per-cpu data to avoid |
| 2267 | * caching issue, which seems more complicated. So selecting |
| 2268 | * target node randomly seems better until now. |
| 2269 | */ |
| 2270 | index = get_random_int() % target_nr; |
| 2271 | break; |
| 2272 | } |
| 2273 | |
| 2274 | target = READ_ONCE(nd->nodes[index]); |
| 2275 | |
| 2276 | out: |
| 2277 | rcu_read_unlock(); |
| 2278 | return target; |
| 2279 | } |
| 2280 | |
| 2281 | #if defined(CONFIG_HOTPLUG_CPU) |
| 2282 | /* Disable reclaim-based migration. */ |
| 2283 | static void __disable_all_migrate_targets(void) |
| 2284 | { |
| 2285 | int node, i; |
| 2286 | |
| 2287 | if (!node_demotion) |
| 2288 | return; |
| 2289 | |
| 2290 | for_each_online_node(node) { |
| 2291 | node_demotion[node].nr = 0; |
| 2292 | for (i = 0; i < DEMOTION_TARGET_NODES; i++) |
| 2293 | node_demotion[node].nodes[i] = NUMA_NO_NODE; |
| 2294 | } |
| 2295 | } |
| 2296 | |
| 2297 | static void disable_all_migrate_targets(void) |
| 2298 | { |
| 2299 | __disable_all_migrate_targets(); |
| 2300 | |
| 2301 | /* |
| 2302 | * Ensure that the "disable" is visible across the system. |
| 2303 | * Readers will see either a combination of before+disable |
| 2304 | * state or disable+after. They will never see before and |
| 2305 | * after state together. |
| 2306 | * |
| 2307 | * The before+after state together might have cycles and |
| 2308 | * could cause readers to do things like loop until this |
| 2309 | * function finishes. This ensures they can only see a |
| 2310 | * single "bad" read and would, for instance, only loop |
| 2311 | * once. |
| 2312 | */ |
| 2313 | synchronize_rcu(); |
| 2314 | } |
| 2315 | |
| 2316 | /* |
| 2317 | * Find an automatic demotion target for 'node'. |
| 2318 | * Failing here is OK. It might just indicate |
| 2319 | * being at the end of a chain. |
| 2320 | */ |
| 2321 | static int establish_migrate_target(int node, nodemask_t *used, |
| 2322 | int best_distance) |
| 2323 | { |
| 2324 | int migration_target, index, val; |
| 2325 | struct demotion_nodes *nd; |
| 2326 | |
| 2327 | if (!node_demotion) |
| 2328 | return NUMA_NO_NODE; |
| 2329 | |
| 2330 | nd = &node_demotion[node]; |
| 2331 | |
| 2332 | migration_target = find_next_best_node(node, used); |
| 2333 | if (migration_target == NUMA_NO_NODE) |
| 2334 | return NUMA_NO_NODE; |
| 2335 | |
| 2336 | /* |
| 2337 | * If the node has been set a migration target node before, |
| 2338 | * which means it's the best distance between them. Still |
| 2339 | * check if this node can be demoted to other target nodes |
| 2340 | * if they have a same best distance. |
| 2341 | */ |
| 2342 | if (best_distance != -1) { |
| 2343 | val = node_distance(node, migration_target); |
| 2344 | if (val > best_distance) |
| 2345 | goto out_clear; |
| 2346 | } |
| 2347 | |
| 2348 | index = nd->nr; |
| 2349 | if (WARN_ONCE(index >= DEMOTION_TARGET_NODES, |
| 2350 | "Exceeds maximum demotion target nodes\n")) |
| 2351 | goto out_clear; |
| 2352 | |
| 2353 | nd->nodes[index] = migration_target; |
| 2354 | nd->nr++; |
| 2355 | |
| 2356 | return migration_target; |
| 2357 | out_clear: |
| 2358 | node_clear(migration_target, *used); |
| 2359 | return NUMA_NO_NODE; |
| 2360 | } |
| 2361 | |
| 2362 | /* |
| 2363 | * When memory fills up on a node, memory contents can be |
| 2364 | * automatically migrated to another node instead of |
| 2365 | * discarded at reclaim. |
| 2366 | * |
| 2367 | * Establish a "migration path" which will start at nodes |
| 2368 | * with CPUs and will follow the priorities used to build the |
| 2369 | * page allocator zonelists. |
| 2370 | * |
| 2371 | * The difference here is that cycles must be avoided. If |
| 2372 | * node0 migrates to node1, then neither node1, nor anything |
| 2373 | * node1 migrates to can migrate to node0. Also one node can |
| 2374 | * be migrated to multiple nodes if the target nodes all have |
| 2375 | * a same best-distance against the source node. |
| 2376 | * |
| 2377 | * This function can run simultaneously with readers of |
| 2378 | * node_demotion[]. However, it can not run simultaneously |
| 2379 | * with itself. Exclusion is provided by memory hotplug events |
| 2380 | * being single-threaded. |
| 2381 | */ |
| 2382 | static void __set_migration_target_nodes(void) |
| 2383 | { |
| 2384 | nodemask_t next_pass = NODE_MASK_NONE; |
| 2385 | nodemask_t this_pass = NODE_MASK_NONE; |
| 2386 | nodemask_t used_targets = NODE_MASK_NONE; |
| 2387 | int node, best_distance; |
| 2388 | |
| 2389 | /* |
| 2390 | * Avoid any oddities like cycles that could occur |
| 2391 | * from changes in the topology. This will leave |
| 2392 | * a momentary gap when migration is disabled. |
| 2393 | */ |
| 2394 | disable_all_migrate_targets(); |
| 2395 | |
| 2396 | /* |
| 2397 | * Allocations go close to CPUs, first. Assume that |
| 2398 | * the migration path starts at the nodes with CPUs. |
| 2399 | */ |
| 2400 | next_pass = node_states[N_CPU]; |
| 2401 | again: |
| 2402 | this_pass = next_pass; |
| 2403 | next_pass = NODE_MASK_NONE; |
| 2404 | /* |
| 2405 | * To avoid cycles in the migration "graph", ensure |
| 2406 | * that migration sources are not future targets by |
| 2407 | * setting them in 'used_targets'. Do this only |
| 2408 | * once per pass so that multiple source nodes can |
| 2409 | * share a target node. |
| 2410 | * |
| 2411 | * 'used_targets' will become unavailable in future |
| 2412 | * passes. This limits some opportunities for |
| 2413 | * multiple source nodes to share a destination. |
| 2414 | */ |
| 2415 | nodes_or(used_targets, used_targets, this_pass); |
| 2416 | |
| 2417 | for_each_node_mask(node, this_pass) { |
| 2418 | best_distance = -1; |
| 2419 | |
| 2420 | /* |
| 2421 | * Try to set up the migration path for the node, and the target |
| 2422 | * migration nodes can be multiple, so doing a loop to find all |
| 2423 | * the target nodes if they all have a best node distance. |
| 2424 | */ |
| 2425 | do { |
| 2426 | int target_node = |
| 2427 | establish_migrate_target(node, &used_targets, |
| 2428 | best_distance); |
| 2429 | |
| 2430 | if (target_node == NUMA_NO_NODE) |
| 2431 | break; |
| 2432 | |
| 2433 | if (best_distance == -1) |
| 2434 | best_distance = node_distance(node, target_node); |
| 2435 | |
| 2436 | /* |
| 2437 | * Visit targets from this pass in the next pass. |
| 2438 | * Eventually, every node will have been part of |
| 2439 | * a pass, and will become set in 'used_targets'. |
| 2440 | */ |
| 2441 | node_set(target_node, next_pass); |
| 2442 | } while (1); |
| 2443 | } |
| 2444 | /* |
| 2445 | * 'next_pass' contains nodes which became migration |
| 2446 | * targets in this pass. Make additional passes until |
| 2447 | * no more migrations targets are available. |
| 2448 | */ |
| 2449 | if (!nodes_empty(next_pass)) |
| 2450 | goto again; |
| 2451 | } |
| 2452 | |
| 2453 | /* |
| 2454 | * For callers that do not hold get_online_mems() already. |
| 2455 | */ |
| 2456 | void set_migration_target_nodes(void) |
| 2457 | { |
| 2458 | get_online_mems(); |
| 2459 | __set_migration_target_nodes(); |
| 2460 | put_online_mems(); |
| 2461 | } |
| 2462 | |
| 2463 | /* |
| 2464 | * This leaves migrate-on-reclaim transiently disabled between |
| 2465 | * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs |
| 2466 | * whether reclaim-based migration is enabled or not, which |
| 2467 | * ensures that the user can turn reclaim-based migration at |
| 2468 | * any time without needing to recalculate migration targets. |
| 2469 | * |
| 2470 | * These callbacks already hold get_online_mems(). That is why |
| 2471 | * __set_migration_target_nodes() can be used as opposed to |
| 2472 | * set_migration_target_nodes(). |
| 2473 | */ |
| 2474 | static int __meminit migrate_on_reclaim_callback(struct notifier_block *self, |
| 2475 | unsigned long action, void *_arg) |
| 2476 | { |
| 2477 | struct memory_notify *arg = _arg; |
| 2478 | |
| 2479 | /* |
| 2480 | * Only update the node migration order when a node is |
| 2481 | * changing status, like online->offline. This avoids |
| 2482 | * the overhead of synchronize_rcu() in most cases. |
| 2483 | */ |
| 2484 | if (arg->status_change_nid < 0) |
| 2485 | return notifier_from_errno(0); |
| 2486 | |
| 2487 | switch (action) { |
| 2488 | case MEM_GOING_OFFLINE: |
| 2489 | /* |
| 2490 | * Make sure there are not transient states where |
| 2491 | * an offline node is a migration target. This |
| 2492 | * will leave migration disabled until the offline |
| 2493 | * completes and the MEM_OFFLINE case below runs. |
| 2494 | */ |
| 2495 | disable_all_migrate_targets(); |
| 2496 | break; |
| 2497 | case MEM_OFFLINE: |
| 2498 | case MEM_ONLINE: |
| 2499 | /* |
| 2500 | * Recalculate the target nodes once the node |
| 2501 | * reaches its final state (online or offline). |
| 2502 | */ |
| 2503 | __set_migration_target_nodes(); |
| 2504 | break; |
| 2505 | case MEM_CANCEL_OFFLINE: |
| 2506 | /* |
| 2507 | * MEM_GOING_OFFLINE disabled all the migration |
| 2508 | * targets. Reenable them. |
| 2509 | */ |
| 2510 | __set_migration_target_nodes(); |
| 2511 | break; |
| 2512 | case MEM_GOING_ONLINE: |
| 2513 | case MEM_CANCEL_ONLINE: |
| 2514 | break; |
| 2515 | } |
| 2516 | |
| 2517 | return notifier_from_errno(0); |
| 2518 | } |
| 2519 | |
| 2520 | void __init migrate_on_reclaim_init(void) |
| 2521 | { |
| 2522 | node_demotion = kmalloc_array(nr_node_ids, |
| 2523 | sizeof(struct demotion_nodes), |
| 2524 | GFP_KERNEL); |
| 2525 | WARN_ON(!node_demotion); |
| 2526 | |
| 2527 | hotplug_memory_notifier(migrate_on_reclaim_callback, 100); |
| 2528 | /* |
| 2529 | * At this point, all numa nodes with memory/CPus have their state |
| 2530 | * properly set, so we can build the demotion order now. |
| 2531 | * Let us hold the cpu_hotplug lock just, as we could possibily have |
| 2532 | * CPU hotplug events during boot. |
| 2533 | */ |
| 2534 | cpus_read_lock(); |
| 2535 | set_migration_target_nodes(); |
| 2536 | cpus_read_unlock(); |
| 2537 | } |
| 2538 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 2539 | |
| 2540 | bool numa_demotion_enabled = false; |
| 2541 | |
| 2542 | #ifdef CONFIG_SYSFS |
| 2543 | static ssize_t numa_demotion_enabled_show(struct kobject *kobj, |
| 2544 | struct kobj_attribute *attr, char *buf) |
| 2545 | { |
| 2546 | return sysfs_emit(buf, "%s\n", |
| 2547 | numa_demotion_enabled ? "true" : "false"); |
| 2548 | } |
| 2549 | |
| 2550 | static ssize_t numa_demotion_enabled_store(struct kobject *kobj, |
| 2551 | struct kobj_attribute *attr, |
| 2552 | const char *buf, size_t count) |
| 2553 | { |
| 2554 | if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1)) |
| 2555 | numa_demotion_enabled = true; |
| 2556 | else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1)) |
| 2557 | numa_demotion_enabled = false; |
| 2558 | else |
| 2559 | return -EINVAL; |
| 2560 | |
| 2561 | return count; |
| 2562 | } |
| 2563 | |
| 2564 | static struct kobj_attribute numa_demotion_enabled_attr = |
| 2565 | __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show, |
| 2566 | numa_demotion_enabled_store); |
| 2567 | |
| 2568 | static struct attribute *numa_attrs[] = { |
| 2569 | &numa_demotion_enabled_attr.attr, |
| 2570 | NULL, |
| 2571 | }; |
| 2572 | |
| 2573 | static const struct attribute_group numa_attr_group = { |
| 2574 | .attrs = numa_attrs, |
| 2575 | }; |
| 2576 | |
| 2577 | static int __init numa_init_sysfs(void) |
| 2578 | { |
| 2579 | int err; |
| 2580 | struct kobject *numa_kobj; |
| 2581 | |
| 2582 | numa_kobj = kobject_create_and_add("numa", mm_kobj); |
| 2583 | if (!numa_kobj) { |
| 2584 | pr_err("failed to create numa kobject\n"); |
| 2585 | return -ENOMEM; |
| 2586 | } |
| 2587 | err = sysfs_create_group(numa_kobj, &numa_attr_group); |
| 2588 | if (err) { |
| 2589 | pr_err("failed to register numa group\n"); |
| 2590 | goto delete_obj; |
| 2591 | } |
| 2592 | return 0; |
| 2593 | |
| 2594 | delete_obj: |
| 2595 | kobject_put(numa_kobj); |
| 2596 | return err; |
| 2597 | } |
| 2598 | subsys_initcall(numa_init_sysfs); |
| 2599 | #endif |