mm/migrate.c

   1 /*
   2  * Memory Migration functionality - linux/mm/migration.c
   3  *
   4  * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
   5  *
   6  * Page migration was first developed in the context of the memory hotplug
   7  * project. The main authors of the migration code are:
   8  *
   9  * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
  10  * Hirokazu Takahashi <taka@valinux.co.jp>
  11  * Dave Hansen <haveblue@us.ibm.com>
  12  * Christoph Lameter <clameter@sgi.com>
  13  */
  14
  15 #include <linux/migrate.h>
  16 #include <linux/module.h>
  17 #include <linux/swap.h>
  18 #include <linux/swapops.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/buffer_head.h>
  21 #include <linux/mm_inline.h>
  22 #include <linux/pagevec.h>
  23 #include <linux/rmap.h>
  24 #include <linux/topology.h>
  25 #include <linux/cpu.h>
  26 #include <linux/cpuset.h>
  27
  28 #include "internal.h"
  29
  30 /* The maximum number of pages to take off the LRU for migration */
  31 #define MIGRATE_CHUNK_SIZE 256
  32
  33 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  34
  35 /*
  36  * Isolate one page from the LRU lists. If successful put it onto
  37  * the indicated list with elevated page count.
  38  *
  39  * Result:
  40  *  -EBUSY: page not on LRU list
  41  *  0: page removed from LRU list and added to the specified list.
  42  */
  43 int isolate_lru_page(struct page *page, struct list_head *pagelist)
  44 {
  45         int ret = -EBUSY;
  46
  47         if (PageLRU(page)) {
  48                 struct zone *zone = page_zone(page);
  49
  50                 spin_lock_irq(&zone->lru_lock);
  51                 if (PageLRU(page)) {
  52                         ret = 0;
  53                         get_page(page);
  54                         ClearPageLRU(page);
  55                         if (PageActive(page))
  56                                 del_page_from_active_list(zone, page);
  57                         else
  58                                 del_page_from_inactive_list(zone, page);
  59                         list_add_tail(&page->lru, pagelist);
  60                 }
  61                 spin_unlock_irq(&zone->lru_lock);
  62         }
  63         return ret;
  64 }
  65
  66 /*
  67  * migrate_prep() needs to be called after we have compiled the list of pages
  68  * to be migrated using isolate_lru_page() but before we begin a series of calls
  69  * to migrate_pages().
  70  */
  71 int migrate_prep(void)
  72 {
  73         /*
  74          * Clear the LRU lists so pages can be isolated.
  75          * Note that pages may be moved off the LRU after we have
  76          * drained them. Those pages will fail to migrate like other
  77          * pages that may be busy.
  78          */
  79         lru_add_drain_all();
  80
  81         return 0;
  82 }
  83
  84 static inline void move_to_lru(struct page *page)
  85 {
  86         list_del(&page->lru);
  87         if (PageActive(page)) {
  88                 /*
  89                  * lru_cache_add_active checks that
  90                  * the PG_active bit is off.
  91                  */
  92                 ClearPageActive(page);
  93                 lru_cache_add_active(page);
  94         } else {
  95                 lru_cache_add(page);
  96         }
  97         put_page(page);
  98 }
  99
 100 /*
 101  * Add isolated pages on the list back to the LRU.
 102  *
 103  * returns the number of pages put back.
 104  */
 105 int putback_lru_pages(struct list_head *l)
 106 {
 107         struct page *page;
 108         struct page *page2;
 109         int count = 0;
 110
 111         list_for_each_entry_safe(page, page2, l, lru) {
 112                 move_to_lru(page);
 113                 count++;
 114         }
 115         return count;
 116 }
 117
 118 static inline int is_swap_pte(pte_t pte)
 119 {
 120         return !pte_none(pte) && !pte_present(pte) && !pte_file(pte);
 121 }
 122
 123 /*
 124  * Restore a potential migration pte to a working pte entry
 125  */
 126 static void remove_migration_pte(struct vm_area_struct *vma, unsigned long addr,
 127                 struct page *old, struct page *new)
 128 {
 129         struct mm_struct *mm = vma->vm_mm;
 130         swp_entry_t entry;
 131         pgd_t *pgd;
 132         pud_t *pud;
 133         pmd_t *pmd;
 134         pte_t *ptep, pte;
 135         spinlock_t *ptl;
 136
 137         pgd = pgd_offset(mm, addr);
 138         if (!pgd_present(*pgd))
 139                 return;
 140
 141         pud = pud_offset(pgd, addr);
 142         if (!pud_present(*pud))
 143                 return;
 144
 145         pmd = pmd_offset(pud, addr);
 146         if (!pmd_present(*pmd))
 147                 return;
 148
 149         ptep = pte_offset_map(pmd, addr);
 150
 151         if (!is_swap_pte(*ptep)) {
 152                 pte_unmap(ptep);
 153                 return;
 154         }
 155
 156         ptl = pte_lockptr(mm, pmd);
 157         spin_lock(ptl);
 158         pte = *ptep;
 159         if (!is_swap_pte(pte))
 160                 goto out;
 161
 162         entry = pte_to_swp_entry(pte);
 163
 164         if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
 165                 goto out;
 166
 167         inc_mm_counter(mm, anon_rss);
 168         get_page(new);
 169         pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
 170         if (is_write_migration_entry(entry))
 171                 pte = pte_mkwrite(pte);
 172         set_pte_at(mm, addr, ptep, pte);
 173         page_add_anon_rmap(new, vma, addr);
 174 out:
 175         pte_unmap_unlock(ptep, ptl);
 176 }
 177
 178 /*
 179  * Get rid of all migration entries and replace them by
 180  * references to the indicated page.
 181  *
 182  * Must hold mmap_sem lock on at least one of the vmas containing
 183  * the page so that the anon_vma cannot vanish.
 184  */
 185 static void remove_migration_ptes(struct page *old, struct page *new)
 186 {
 187         struct anon_vma *anon_vma;
 188         struct vm_area_struct *vma;
 189         unsigned long mapping;
 190
 191         mapping = (unsigned long)new->mapping;
 192
 193         if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
 194                 return;
 195
 196         /*
 197          * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
 198          */
 199         anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
 200         spin_lock(&anon_vma->lock);
 201
 202         list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
 203                 remove_migration_pte(vma, page_address_in_vma(new, vma),
 204                                         old, new);
 205
 206         spin_unlock(&anon_vma->lock);
 207 }
 208
 209 /*
 210  * Something used the pte of a page under migration. We need to
 211  * get to the page and wait until migration is finished.
 212  * When we return from this function the fault will be retried.
 213  *
 214  * This function is called from do_swap_page().
 215  */
 216 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 217                                 unsigned long address)
 218 {
 219         pte_t *ptep, pte;
 220         spinlock_t *ptl;
 221         swp_entry_t entry;
 222         struct page *page;
 223
 224         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 225         pte = *ptep;
 226         if (!is_swap_pte(pte))
 227                 goto out;
 228
 229         entry = pte_to_swp_entry(pte);
 230         if (!is_migration_entry(entry))
 231                 goto out;
 232
 233         page = migration_entry_to_page(entry);
 234
 235         get_page(page);
 236         pte_unmap_unlock(ptep, ptl);
 237         wait_on_page_locked(page);
 238         put_page(page);
 239         return;
 240 out:
 241         pte_unmap_unlock(ptep, ptl);
 242 }
 243
 244 /*
 245  * Replace the page in the mapping.
 246  *
 247  * The number of remaining references must be:
 248  * 1 for anonymous pages without a mapping
 249  * 2 for pages with a mapping
 250  * 3 for pages with a mapping and PagePrivate set.
 251  */
 252 static int migrate_page_move_mapping(struct address_space *mapping,
 253                 struct page *newpage, struct page *page)
 254 {
 255         struct page **radix_pointer;
 256
 257         if (!mapping) {
 258                 /* Anonymous page */
 259                 if (page_count(page) != 1)
 260                         return -EAGAIN;
 261                 return 0;
 262         }
 263
 264         write_lock_irq(&mapping->tree_lock);
 265
 266         radix_pointer = (struct page **)radix_tree_lookup_slot(
 267                                                 &mapping->page_tree,
 268                                                 page_index(page));
 269
 270         if (page_count(page) != 2 + !!PagePrivate(page) ||
 271                         *radix_pointer != page) {
 272                 write_unlock_irq(&mapping->tree_lock);
 273                 return -EAGAIN;
 274         }
 275
 276         /*
 277          * Now we know that no one else is looking at the page.
 278          */
 279         get_page(newpage);
 280 #ifdef CONFIG_SWAP
 281         if (PageSwapCache(page)) {
 282                 SetPageSwapCache(newpage);
 283                 set_page_private(newpage, page_private(page));
 284         }
 285 #endif
 286
 287         *radix_pointer = newpage;
 288         __put_page(page);
 289         write_unlock_irq(&mapping->tree_lock);
 290
 291         return 0;
 292 }
 293
 294 /*
 295  * Copy the page to its new location
 296  */
 297 static void migrate_page_copy(struct page *newpage, struct page *page)
 298 {
 299         copy_highpage(newpage, page);
 300
 301         if (PageError(page))
 302                 SetPageError(newpage);
 303         if (PageReferenced(page))
 304                 SetPageReferenced(newpage);
 305         if (PageUptodate(page))
 306                 SetPageUptodate(newpage);
 307         if (PageActive(page))
 308                 SetPageActive(newpage);
 309         if (PageChecked(page))
 310                 SetPageChecked(newpage);
 311         if (PageMappedToDisk(page))
 312                 SetPageMappedToDisk(newpage);
 313
 314         if (PageDirty(page)) {
 315                 clear_page_dirty_for_io(page);
 316                 set_page_dirty(newpage);
 317         }
 318
 319 #ifdef CONFIG_SWAP
 320         ClearPageSwapCache(page);
 321 #endif
 322         ClearPageActive(page);
 323         ClearPagePrivate(page);
 324         set_page_private(page, 0);
 325         page->mapping = NULL;
 326
 327         /*
 328          * If any waiters have accumulated on the new page then
 329          * wake them up.
 330          */
 331         if (PageWriteback(newpage))
 332                 end_page_writeback(newpage);
 333 }
 334
 335 /************************************************************
 336  *                    Migration functions
 337  ***********************************************************/
 338
 339 /* Always fail migration. Used for mappings that are not movable */
 340 int fail_migrate_page(struct address_space *mapping,
 341                         struct page *newpage, struct page *page)
 342 {
 343         return -EIO;
 344 }
 345 EXPORT_SYMBOL(fail_migrate_page);
 346
 347 /*
 348  * Common logic to directly migrate a single page suitable for
 349  * pages that do not use PagePrivate.
 350  *
 351  * Pages are locked upon entry and exit.
 352  */
 353 int migrate_page(struct address_space *mapping,
 354                 struct page *newpage, struct page *page)
 355 {
 356         int rc;
 357
 358         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 359
 360         rc = migrate_page_move_mapping(mapping, newpage, page);
 361
 362         if (rc)
 363                 return rc;
 364
 365         migrate_page_copy(newpage, page);
 366         return 0;
 367 }
 368 EXPORT_SYMBOL(migrate_page);
 369
 370 /*
 371  * Migration function for pages with buffers. This function can only be used
 372  * if the underlying filesystem guarantees that no other references to "page"
 373  * exist.
 374  */
 375 int buffer_migrate_page(struct address_space *mapping,
 376                 struct page *newpage, struct page *page)
 377 {
 378         struct buffer_head *bh, *head;
 379         int rc;
 380
 381         if (!page_has_buffers(page))
 382                 return migrate_page(mapping, newpage, page);
 383
 384         head = page_buffers(page);
 385
 386         rc = migrate_page_move_mapping(mapping, newpage, page);
 387
 388         if (rc)
 389                 return rc;
 390
 391         bh = head;
 392         do {
 393                 get_bh(bh);
 394                 lock_buffer(bh);
 395                 bh = bh->b_this_page;
 396
 397         } while (bh != head);
 398
 399         ClearPagePrivate(page);
 400         set_page_private(newpage, page_private(page));
 401         set_page_private(page, 0);
 402         put_page(page);
 403         get_page(newpage);
 404
 405         bh = head;
 406         do {
 407                 set_bh_page(bh, newpage, bh_offset(bh));
 408                 bh = bh->b_this_page;
 409
 410         } while (bh != head);
 411
 412         SetPagePrivate(newpage);
 413
 414         migrate_page_copy(newpage, page);
 415
 416         bh = head;
 417         do {
 418                 unlock_buffer(bh);
 419                 put_bh(bh);
 420                 bh = bh->b_this_page;
 421
 422         } while (bh != head);
 423
 424         return 0;
 425 }
 426 EXPORT_SYMBOL(buffer_migrate_page);
 427
 428 static int fallback_migrate_page(struct address_space *mapping,
 429         struct page *newpage, struct page *page)
 430 {
 431         /*
 432          * Default handling if a filesystem does not provide
 433          * a migration function. We can only migrate clean
 434          * pages so try to write out any dirty pages first.
 435          */
 436         if (PageDirty(page)) {
 437                 switch (pageout(page, mapping)) {
 438                 case PAGE_KEEP:
 439                 case PAGE_ACTIVATE:
 440                         return -EAGAIN;
 441
 442                 case PAGE_SUCCESS:
 443                         /* Relock since we lost the lock */
 444                         lock_page(page);
 445                         /* Must retry since page state may have changed */
 446                         return -EAGAIN;
 447
 448                 case PAGE_CLEAN:
 449                         ; /* try to migrate the page below */
 450                 }
 451         }
 452
 453         /*
 454          * Buffers may be managed in a filesystem specific way.
 455          * We must have no buffers or drop them.
 456          */
 457         if (page_has_buffers(page) &&
 458             !try_to_release_page(page, GFP_KERNEL))
 459                 return -EAGAIN;
 460
 461         return migrate_page(mapping, newpage, page);
 462 }
 463
 464 /*
 465  * migrate_pages
 466  *
 467  * Two lists are passed to this function. The first list
 468  * contains the pages isolated from the LRU to be migrated.
 469  * The second list contains new pages that the pages isolated
 470  * can be moved to.
 471  *
 472  * The function returns after 10 attempts or if no pages
 473  * are movable anymore because to has become empty
 474  * or no retryable pages exist anymore.
 475  *
 476  * Return: Number of pages not migrated when "to" ran empty.
 477  */
 478 int migrate_pages(struct list_head *from, struct list_head *to,
 479                   struct list_head *moved, struct list_head *failed)
 480 {
 481         int retry;
 482         int nr_failed = 0;
 483         int pass = 0;
 484         struct page *page;
 485         struct page *page2;
 486         int swapwrite = current->flags & PF_SWAPWRITE;
 487         int rc;
 488
 489         if (!swapwrite)
 490                 current->flags |= PF_SWAPWRITE;
 491
 492 redo:
 493         retry = 0;
 494
 495         list_for_each_entry_safe(page, page2, from, lru) {
 496                 struct page *newpage = NULL;
 497                 struct address_space *mapping;
 498
 499                 cond_resched();
 500
 501                 rc = 0;
 502                 if (page_count(page) == 1)
 503                         /* page was freed from under us. So we are done. */
 504                         goto next;
 505
 506                 if (to && list_empty(to))
 507                         break;
 508
 509                 /*
 510                  * Skip locked pages during the first two passes to give the
 511                  * functions holding the lock time to release the page. Later we
 512                  * use lock_page() to have a higher chance of acquiring the
 513                  * lock.
 514                  */
 515                 rc = -EAGAIN;
 516                 if (pass > 2)
 517                         lock_page(page);
 518                 else
 519                         if (TestSetPageLocked(page))
 520                                 goto next;
 521
 522                 /*
 523                  * Only wait on writeback if we have already done a pass where
 524                  * we we may have triggered writeouts for lots of pages.
 525                  */
 526                 if (pass > 0)
 527                         wait_on_page_writeback(page);
 528                 else
 529                         if (PageWriteback(page))
 530                                 goto unlock_page;
 531
 532                 /*
 533                  * Establish migration ptes or remove ptes
 534                  */
 535                 rc = -EPERM;
 536                 if (try_to_unmap(page, 1) == SWAP_FAIL)
 537                         /* A vma has VM_LOCKED set -> permanent failure */
 538                         goto unlock_page;
 539
 540                 rc = -EAGAIN;
 541                 if (page_mapped(page))
 542                         goto unlock_page;
 543
 544                 newpage = lru_to_page(to);
 545                 lock_page(newpage);
 546                 /* Prepare mapping for the new page.*/
 547                 newpage->index = page->index;
 548                 newpage->mapping = page->mapping;
 549
 550                 /*
 551                  * Pages are properly locked and writeback is complete.
 552                  * Try to migrate the page.
 553                  */
 554                 mapping = page_mapping(page);
 555                 if (!mapping)
 556                         rc = migrate_page(mapping, newpage, page);
 557
 558                 else if (mapping->a_ops->migratepage)
 559                         /*
 560                          * Most pages have a mapping and most filesystems
 561                          * should provide a migration function. Anonymous
 562                          * pages are part of swap space which also has its
 563                          * own migration function. This is the most common
 564                          * path for page migration.
 565                          */
 566                         rc = mapping->a_ops->migratepage(mapping,
 567                                                         newpage, page);
 568                 else
 569                         rc = fallback_migrate_page(mapping, newpage, page);
 570
 571                 if (!rc)
 572                         remove_migration_ptes(page, newpage);
 573
 574                 unlock_page(newpage);
 575
 576 unlock_page:
 577                 if (rc)
 578                         remove_migration_ptes(page, page);
 579
 580                 unlock_page(page);
 581
 582 next:
 583                 if (rc) {
 584                         if (newpage)
 585                                 newpage->mapping = NULL;
 586
 587                         if (rc == -EAGAIN)
 588                                 retry++;
 589                         else {
 590                                 /* Permanent failure */
 591                                 list_move(&page->lru, failed);
 592                                 nr_failed++;
 593                         }
 594                 } else {
 595                         if (newpage) {
 596                                 /* Successful migration. Return page to LRU */
 597                                 move_to_lru(newpage);
 598                         }
 599                         list_move(&page->lru, moved);
 600                 }
 601         }
 602         if (retry && pass++ < 10)
 603                 goto redo;
 604
 605         if (!swapwrite)
 606                 current->flags &= ~PF_SWAPWRITE;
 607
 608         return nr_failed + retry;
 609 }
 610
 611 /*
 612  * Migrate the list 'pagelist' of pages to a certain destination.
 613  *
 614  * Specify destination with either non-NULL vma or dest_node >= 0
 615  * Return the number of pages not migrated or error code
 616  */
 617 int migrate_pages_to(struct list_head *pagelist,
 618                         struct vm_area_struct *vma, int dest)
 619 {
 620         LIST_HEAD(newlist);
 621         LIST_HEAD(moved);
 622         LIST_HEAD(failed);
 623         int err = 0;
 624         unsigned long offset = 0;
 625         int nr_pages;
 626         struct page *page;
 627         struct list_head *p;
 628
 629 redo:
 630         nr_pages = 0;
 631         list_for_each(p, pagelist) {
 632                 if (vma) {
 633                         /*
 634                          * The address passed to alloc_page_vma is used to
 635                          * generate the proper interleave behavior. We fake
 636                          * the address here by an increasing offset in order
 637                          * to get the proper distribution of pages.
 638                          *
 639                          * No decision has been made as to which page
 640                          * a certain old page is moved to so we cannot
 641                          * specify the correct address.
 642                          */
 643                         page = alloc_page_vma(GFP_HIGHUSER, vma,
 644                                         offset + vma->vm_start);
 645                         offset += PAGE_SIZE;
 646                 }
 647                 else
 648                         page = alloc_pages_node(dest, GFP_HIGHUSER, 0);
 649
 650                 if (!page) {
 651                         err = -ENOMEM;
 652                         goto out;
 653                 }
 654                 list_add_tail(&page->lru, &newlist);
 655                 nr_pages++;
 656                 if (nr_pages > MIGRATE_CHUNK_SIZE)
 657                         break;
 658         }
 659         err = migrate_pages(pagelist, &newlist, &moved, &failed);
 660
 661         putback_lru_pages(&moved);      /* Call release pages instead ?? */
 662
 663         if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist))
 664                 goto redo;
 665 out:
 666         /* Return leftover allocated pages */
 667         while (!list_empty(&newlist)) {
 668                 page = list_entry(newlist.next, struct page, lru);
 669                 list_del(&page->lru);
 670                 __free_page(page);
 671         }
 672         list_splice(&failed, pagelist);
 673         if (err < 0)
 674                 return err;
 675
 676         /* Calculate number of leftover pages */
 677         nr_pages = 0;
 678         list_for_each(p, pagelist)
 679                 nr_pages++;
 680         return nr_pages;
 681 }