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/pagemap.h> 19 #include <linux/buffer_head.h> 20 #include <linux/mm_inline.h> 21 #include <linux/pagevec.h> 22 #include <linux/rmap.h> 23 #include <linux/topology.h> 24 #include <linux/cpu.h> 25 #include <linux/cpuset.h> 26 #include <linux/swapops.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 /* Must have swap device for migration */ 74 if (nr_swap_pages <= 0) 75 return -ENODEV; 76 77 /* 78 * Clear the LRU lists so pages can be isolated. 79 * Note that pages may be moved off the LRU after we have 80 * drained them. Those pages will fail to migrate like other 81 * pages that may be busy. 82 */ 83 lru_add_drain_all(); 84 85 return 0; 86 } 87 88 static inline void move_to_lru(struct page *page) 89 { 90 list_del(&page->lru); 91 if (PageActive(page)) { 92 /* 93 * lru_cache_add_active checks that 94 * the PG_active bit is off. 95 */ 96 ClearPageActive(page); 97 lru_cache_add_active(page); 98 } else { 99 lru_cache_add(page); 100 } 101 put_page(page); 102 } 103 104 /* 105 * Add isolated pages on the list back to the LRU. 106 * 107 * returns the number of pages put back. 108 */ 109 int putback_lru_pages(struct list_head *l) 110 { 111 struct page *page; 112 struct page *page2; 113 int count = 0; 114 115 list_for_each_entry_safe(page, page2, l, lru) { 116 move_to_lru(page); 117 count++; 118 } 119 return count; 120 } 121 122 /* 123 * Non migratable page 124 */ 125 int fail_migrate_page(struct page *newpage, struct page *page) 126 { 127 return -EIO; 128 } 129 EXPORT_SYMBOL(fail_migrate_page); 130 131 /* 132 * swapout a single page 133 * page is locked upon entry, unlocked on exit 134 */ 135 static int swap_page(struct page *page) 136 { 137 struct address_space *mapping = page_mapping(page); 138 139 if (page_mapped(page) && mapping) 140 if (try_to_unmap(page, 1) != SWAP_SUCCESS) 141 goto unlock_retry; 142 143 if (PageDirty(page)) { 144 /* Page is dirty, try to write it out here */ 145 switch(pageout(page, mapping)) { 146 case PAGE_KEEP: 147 case PAGE_ACTIVATE: 148 goto unlock_retry; 149 150 case PAGE_SUCCESS: 151 goto retry; 152 153 case PAGE_CLEAN: 154 ; /* try to free the page below */ 155 } 156 } 157 158 if (PagePrivate(page)) { 159 if (!try_to_release_page(page, GFP_KERNEL) || 160 (!mapping && page_count(page) == 1)) 161 goto unlock_retry; 162 } 163 164 if (remove_mapping(mapping, page)) { 165 /* Success */ 166 unlock_page(page); 167 return 0; 168 } 169 170 unlock_retry: 171 unlock_page(page); 172 173 retry: 174 return -EAGAIN; 175 } 176 177 /* 178 * Remove references for a page and establish the new page with the correct 179 * basic settings to be able to stop accesses to the page. 180 */ 181 int migrate_page_remove_references(struct page *newpage, 182 struct page *page, int nr_refs) 183 { 184 struct address_space *mapping = page_mapping(page); 185 struct page **radix_pointer; 186 187 /* 188 * Avoid doing any of the following work if the page count 189 * indicates that the page is in use or truncate has removed 190 * the page. 191 */ 192 if (!mapping || page_mapcount(page) + nr_refs != page_count(page)) 193 return -EAGAIN; 194 195 /* 196 * Establish swap ptes for anonymous pages or destroy pte 197 * maps for files. 198 * 199 * In order to reestablish file backed mappings the fault handlers 200 * will take the radix tree_lock which may then be used to stop 201 * processses from accessing this page until the new page is ready. 202 * 203 * A process accessing via a swap pte (an anonymous page) will take a 204 * page_lock on the old page which will block the process until the 205 * migration attempt is complete. At that time the PageSwapCache bit 206 * will be examined. If the page was migrated then the PageSwapCache 207 * bit will be clear and the operation to retrieve the page will be 208 * retried which will find the new page in the radix tree. Then a new 209 * direct mapping may be generated based on the radix tree contents. 210 * 211 * If the page was not migrated then the PageSwapCache bit 212 * is still set and the operation may continue. 213 */ 214 if (try_to_unmap(page, 1) == SWAP_FAIL) 215 /* A vma has VM_LOCKED set -> permanent failure */ 216 return -EPERM; 217 218 /* 219 * Give up if we were unable to remove all mappings. 220 */ 221 if (page_mapcount(page)) 222 return -EAGAIN; 223 224 write_lock_irq(&mapping->tree_lock); 225 226 radix_pointer = (struct page **)radix_tree_lookup_slot( 227 &mapping->page_tree, 228 page_index(page)); 229 230 if (!page_mapping(page) || page_count(page) != nr_refs || 231 *radix_pointer != page) { 232 write_unlock_irq(&mapping->tree_lock); 233 return -EAGAIN; 234 } 235 236 /* 237 * Now we know that no one else is looking at the page. 238 * 239 * Certain minimal information about a page must be available 240 * in order for other subsystems to properly handle the page if they 241 * find it through the radix tree update before we are finished 242 * copying the page. 243 */ 244 get_page(newpage); 245 newpage->index = page->index; 246 newpage->mapping = page->mapping; 247 if (PageSwapCache(page)) { 248 SetPageSwapCache(newpage); 249 set_page_private(newpage, page_private(page)); 250 } 251 252 *radix_pointer = newpage; 253 __put_page(page); 254 write_unlock_irq(&mapping->tree_lock); 255 256 return 0; 257 } 258 EXPORT_SYMBOL(migrate_page_remove_references); 259 260 /* 261 * Copy the page to its new location 262 */ 263 void migrate_page_copy(struct page *newpage, struct page *page) 264 { 265 copy_highpage(newpage, page); 266 267 if (PageError(page)) 268 SetPageError(newpage); 269 if (PageReferenced(page)) 270 SetPageReferenced(newpage); 271 if (PageUptodate(page)) 272 SetPageUptodate(newpage); 273 if (PageActive(page)) 274 SetPageActive(newpage); 275 if (PageChecked(page)) 276 SetPageChecked(newpage); 277 if (PageMappedToDisk(page)) 278 SetPageMappedToDisk(newpage); 279 280 if (PageDirty(page)) { 281 clear_page_dirty_for_io(page); 282 set_page_dirty(newpage); 283 } 284 285 ClearPageSwapCache(page); 286 ClearPageActive(page); 287 ClearPagePrivate(page); 288 set_page_private(page, 0); 289 page->mapping = NULL; 290 291 /* 292 * If any waiters have accumulated on the new page then 293 * wake them up. 294 */ 295 if (PageWriteback(newpage)) 296 end_page_writeback(newpage); 297 } 298 EXPORT_SYMBOL(migrate_page_copy); 299 300 /* 301 * Common logic to directly migrate a single page suitable for 302 * pages that do not use PagePrivate. 303 * 304 * Pages are locked upon entry and exit. 305 */ 306 int migrate_page(struct page *newpage, struct page *page) 307 { 308 int rc; 309 310 BUG_ON(PageWriteback(page)); /* Writeback must be complete */ 311 312 rc = migrate_page_remove_references(newpage, page, 2); 313 314 if (rc) 315 return rc; 316 317 migrate_page_copy(newpage, page); 318 319 /* 320 * Remove auxiliary swap entries and replace 321 * them with real ptes. 322 * 323 * Note that a real pte entry will allow processes that are not 324 * waiting on the page lock to use the new page via the page tables 325 * before the new page is unlocked. 326 */ 327 remove_from_swap(newpage); 328 return 0; 329 } 330 EXPORT_SYMBOL(migrate_page); 331 332 /* 333 * migrate_pages 334 * 335 * Two lists are passed to this function. The first list 336 * contains the pages isolated from the LRU to be migrated. 337 * The second list contains new pages that the pages isolated 338 * can be moved to. If the second list is NULL then all 339 * pages are swapped out. 340 * 341 * The function returns after 10 attempts or if no pages 342 * are movable anymore because to has become empty 343 * or no retryable pages exist anymore. 344 * 345 * Return: Number of pages not migrated when "to" ran empty. 346 */ 347 int migrate_pages(struct list_head *from, struct list_head *to, 348 struct list_head *moved, struct list_head *failed) 349 { 350 int retry; 351 int nr_failed = 0; 352 int pass = 0; 353 struct page *page; 354 struct page *page2; 355 int swapwrite = current->flags & PF_SWAPWRITE; 356 int rc; 357 358 if (!swapwrite) 359 current->flags |= PF_SWAPWRITE; 360 361 redo: 362 retry = 0; 363 364 list_for_each_entry_safe(page, page2, from, lru) { 365 struct page *newpage = NULL; 366 struct address_space *mapping; 367 368 cond_resched(); 369 370 rc = 0; 371 if (page_count(page) == 1) 372 /* page was freed from under us. So we are done. */ 373 goto next; 374 375 if (to && list_empty(to)) 376 break; 377 378 /* 379 * Skip locked pages during the first two passes to give the 380 * functions holding the lock time to release the page. Later we 381 * use lock_page() to have a higher chance of acquiring the 382 * lock. 383 */ 384 rc = -EAGAIN; 385 if (pass > 2) 386 lock_page(page); 387 else 388 if (TestSetPageLocked(page)) 389 goto next; 390 391 /* 392 * Only wait on writeback if we have already done a pass where 393 * we we may have triggered writeouts for lots of pages. 394 */ 395 if (pass > 0) { 396 wait_on_page_writeback(page); 397 } else { 398 if (PageWriteback(page)) 399 goto unlock_page; 400 } 401 402 /* 403 * Anonymous pages must have swap cache references otherwise 404 * the information contained in the page maps cannot be 405 * preserved. 406 */ 407 if (PageAnon(page) && !PageSwapCache(page)) { 408 if (!add_to_swap(page, GFP_KERNEL)) { 409 rc = -ENOMEM; 410 goto unlock_page; 411 } 412 } 413 414 if (!to) { 415 rc = swap_page(page); 416 goto next; 417 } 418 419 newpage = lru_to_page(to); 420 lock_page(newpage); 421 422 /* 423 * Pages are properly locked and writeback is complete. 424 * Try to migrate the page. 425 */ 426 mapping = page_mapping(page); 427 if (!mapping) 428 goto unlock_both; 429 430 if (mapping->a_ops->migratepage) { 431 /* 432 * Most pages have a mapping and most filesystems 433 * should provide a migration function. Anonymous 434 * pages are part of swap space which also has its 435 * own migration function. This is the most common 436 * path for page migration. 437 */ 438 rc = mapping->a_ops->migratepage(newpage, page); 439 goto unlock_both; 440 } 441 442 /* Make sure the dirty bit is up to date */ 443 if (try_to_unmap(page, 1) == SWAP_FAIL) { 444 rc = -EPERM; 445 goto unlock_both; 446 } 447 448 if (page_mapcount(page)) { 449 rc = -EAGAIN; 450 goto unlock_both; 451 } 452 453 /* 454 * Default handling if a filesystem does not provide 455 * a migration function. We can only migrate clean 456 * pages so try to write out any dirty pages first. 457 */ 458 if (PageDirty(page)) { 459 switch (pageout(page, mapping)) { 460 case PAGE_KEEP: 461 case PAGE_ACTIVATE: 462 goto unlock_both; 463 464 case PAGE_SUCCESS: 465 unlock_page(newpage); 466 goto next; 467 468 case PAGE_CLEAN: 469 ; /* try to migrate the page below */ 470 } 471 } 472 473 /* 474 * Buffers are managed in a filesystem specific way. 475 * We must have no buffers or drop them. 476 */ 477 if (!page_has_buffers(page) || 478 try_to_release_page(page, GFP_KERNEL)) { 479 rc = migrate_page(newpage, page); 480 goto unlock_both; 481 } 482 483 /* 484 * On early passes with mapped pages simply 485 * retry. There may be a lock held for some 486 * buffers that may go away. Later 487 * swap them out. 488 */ 489 if (pass > 4) { 490 /* 491 * Persistently unable to drop buffers..... As a 492 * measure of last resort we fall back to 493 * swap_page(). 494 */ 495 unlock_page(newpage); 496 newpage = NULL; 497 rc = swap_page(page); 498 goto next; 499 } 500 501 unlock_both: 502 unlock_page(newpage); 503 504 unlock_page: 505 unlock_page(page); 506 507 next: 508 if (rc == -EAGAIN) { 509 retry++; 510 } else if (rc) { 511 /* Permanent failure */ 512 list_move(&page->lru, failed); 513 nr_failed++; 514 } else { 515 if (newpage) { 516 /* Successful migration. Return page to LRU */ 517 move_to_lru(newpage); 518 } 519 list_move(&page->lru, moved); 520 } 521 } 522 if (retry && pass++ < 10) 523 goto redo; 524 525 if (!swapwrite) 526 current->flags &= ~PF_SWAPWRITE; 527 528 return nr_failed + retry; 529 } 530 531 /* 532 * Migration function for pages with buffers. This function can only be used 533 * if the underlying filesystem guarantees that no other references to "page" 534 * exist. 535 */ 536 int buffer_migrate_page(struct page *newpage, struct page *page) 537 { 538 struct address_space *mapping = page->mapping; 539 struct buffer_head *bh, *head; 540 int rc; 541 542 if (!mapping) 543 return -EAGAIN; 544 545 if (!page_has_buffers(page)) 546 return migrate_page(newpage, page); 547 548 head = page_buffers(page); 549 550 rc = migrate_page_remove_references(newpage, page, 3); 551 552 if (rc) 553 return rc; 554 555 bh = head; 556 do { 557 get_bh(bh); 558 lock_buffer(bh); 559 bh = bh->b_this_page; 560 561 } while (bh != head); 562 563 ClearPagePrivate(page); 564 set_page_private(newpage, page_private(page)); 565 set_page_private(page, 0); 566 put_page(page); 567 get_page(newpage); 568 569 bh = head; 570 do { 571 set_bh_page(bh, newpage, bh_offset(bh)); 572 bh = bh->b_this_page; 573 574 } while (bh != head); 575 576 SetPagePrivate(newpage); 577 578 migrate_page_copy(newpage, page); 579 580 bh = head; 581 do { 582 unlock_buffer(bh); 583 put_bh(bh); 584 bh = bh->b_this_page; 585 586 } while (bh != head); 587 588 return 0; 589 } 590 EXPORT_SYMBOL(buffer_migrate_page); 591 592 /* 593 * Migrate the list 'pagelist' of pages to a certain destination. 594 * 595 * Specify destination with either non-NULL vma or dest_node >= 0 596 * Return the number of pages not migrated or error code 597 */ 598 int migrate_pages_to(struct list_head *pagelist, 599 struct vm_area_struct *vma, int dest) 600 { 601 LIST_HEAD(newlist); 602 LIST_HEAD(moved); 603 LIST_HEAD(failed); 604 int err = 0; 605 unsigned long offset = 0; 606 int nr_pages; 607 struct page *page; 608 struct list_head *p; 609 610 redo: 611 nr_pages = 0; 612 list_for_each(p, pagelist) { 613 if (vma) { 614 /* 615 * The address passed to alloc_page_vma is used to 616 * generate the proper interleave behavior. We fake 617 * the address here by an increasing offset in order 618 * to get the proper distribution of pages. 619 * 620 * No decision has been made as to which page 621 * a certain old page is moved to so we cannot 622 * specify the correct address. 623 */ 624 page = alloc_page_vma(GFP_HIGHUSER, vma, 625 offset + vma->vm_start); 626 offset += PAGE_SIZE; 627 } 628 else 629 page = alloc_pages_node(dest, GFP_HIGHUSER, 0); 630 631 if (!page) { 632 err = -ENOMEM; 633 goto out; 634 } 635 list_add_tail(&page->lru, &newlist); 636 nr_pages++; 637 if (nr_pages > MIGRATE_CHUNK_SIZE) 638 break; 639 } 640 err = migrate_pages(pagelist, &newlist, &moved, &failed); 641 642 putback_lru_pages(&moved); /* Call release pages instead ?? */ 643 644 if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist)) 645 goto redo; 646 out: 647 /* Return leftover allocated pages */ 648 while (!list_empty(&newlist)) { 649 page = list_entry(newlist.next, struct page, lru); 650 list_del(&page->lru); 651 __free_page(page); 652 } 653 list_splice(&failed, pagelist); 654 if (err < 0) 655 return err; 656 657 /* Calculate number of leftover pages */ 658 nr_pages = 0; 659 list_for_each(p, pagelist) 660 nr_pages++; 661 return nr_pages; 662 } 663