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 * swapout a single page 124 * page is locked upon entry, unlocked on exit 125 */ 126 static int swap_page(struct page *page) 127 { 128 struct address_space *mapping = page_mapping(page); 129 130 if (page_mapped(page) && mapping) 131 if (try_to_unmap(page, 1) != SWAP_SUCCESS) 132 goto unlock_retry; 133 134 if (PageDirty(page)) { 135 /* Page is dirty, try to write it out here */ 136 switch(pageout(page, mapping)) { 137 case PAGE_KEEP: 138 case PAGE_ACTIVATE: 139 goto unlock_retry; 140 141 case PAGE_SUCCESS: 142 goto retry; 143 144 case PAGE_CLEAN: 145 ; /* try to free the page below */ 146 } 147 } 148 149 if (PagePrivate(page)) { 150 if (!try_to_release_page(page, GFP_KERNEL) || 151 (!mapping && page_count(page) == 1)) 152 goto unlock_retry; 153 } 154 155 if (remove_mapping(mapping, page)) { 156 /* Success */ 157 unlock_page(page); 158 return 0; 159 } 160 161 unlock_retry: 162 unlock_page(page); 163 164 retry: 165 return -EAGAIN; 166 } 167 168 /* 169 * Remove references for a page and establish the new page with the correct 170 * basic settings to be able to stop accesses to the page. 171 * 172 * The number of remaining references must be: 173 * 1 for anonymous pages without a mapping 174 * 2 for pages with a mapping 175 * 3 for pages with a mapping and PagePrivate set. 176 */ 177 static int migrate_page_remove_references(struct page *newpage, 178 struct page *page) 179 { 180 struct address_space *mapping = page_mapping(page); 181 struct page **radix_pointer; 182 183 if (!mapping) 184 return -EAGAIN; 185 186 /* 187 * Establish swap ptes for anonymous pages or destroy pte 188 * maps for files. 189 * 190 * In order to reestablish file backed mappings the fault handlers 191 * will take the radix tree_lock which may then be used to stop 192 * processses from accessing this page until the new page is ready. 193 * 194 * A process accessing via a swap pte (an anonymous page) will take a 195 * page_lock on the old page which will block the process until the 196 * migration attempt is complete. At that time the PageSwapCache bit 197 * will be examined. If the page was migrated then the PageSwapCache 198 * bit will be clear and the operation to retrieve the page will be 199 * retried which will find the new page in the radix tree. Then a new 200 * direct mapping may be generated based on the radix tree contents. 201 * 202 * If the page was not migrated then the PageSwapCache bit 203 * is still set and the operation may continue. 204 */ 205 if (try_to_unmap(page, 1) == SWAP_FAIL) 206 /* A vma has VM_LOCKED set -> permanent failure */ 207 return -EPERM; 208 209 /* 210 * Give up if we were unable to remove all mappings. 211 */ 212 if (page_mapcount(page)) 213 return -EAGAIN; 214 215 write_lock_irq(&mapping->tree_lock); 216 217 radix_pointer = (struct page **)radix_tree_lookup_slot( 218 &mapping->page_tree, 219 page_index(page)); 220 221 if (!page_mapping(page) || 222 page_count(page) != 2 + !!PagePrivate(page) || 223 *radix_pointer != page) { 224 write_unlock_irq(&mapping->tree_lock); 225 return -EAGAIN; 226 } 227 228 /* 229 * Now we know that no one else is looking at the page. 230 * 231 * Certain minimal information about a page must be available 232 * in order for other subsystems to properly handle the page if they 233 * find it through the radix tree update before we are finished 234 * copying the page. 235 */ 236 get_page(newpage); 237 newpage->index = page->index; 238 newpage->mapping = page->mapping; 239 if (PageSwapCache(page)) { 240 SetPageSwapCache(newpage); 241 set_page_private(newpage, page_private(page)); 242 } 243 244 *radix_pointer = newpage; 245 __put_page(page); 246 write_unlock_irq(&mapping->tree_lock); 247 248 return 0; 249 } 250 251 /* 252 * Copy the page to its new location 253 */ 254 static void migrate_page_copy(struct page *newpage, struct page *page) 255 { 256 copy_highpage(newpage, page); 257 258 if (PageError(page)) 259 SetPageError(newpage); 260 if (PageReferenced(page)) 261 SetPageReferenced(newpage); 262 if (PageUptodate(page)) 263 SetPageUptodate(newpage); 264 if (PageActive(page)) 265 SetPageActive(newpage); 266 if (PageChecked(page)) 267 SetPageChecked(newpage); 268 if (PageMappedToDisk(page)) 269 SetPageMappedToDisk(newpage); 270 271 if (PageDirty(page)) { 272 clear_page_dirty_for_io(page); 273 set_page_dirty(newpage); 274 } 275 276 ClearPageSwapCache(page); 277 ClearPageActive(page); 278 ClearPagePrivate(page); 279 set_page_private(page, 0); 280 page->mapping = NULL; 281 282 /* 283 * If any waiters have accumulated on the new page then 284 * wake them up. 285 */ 286 if (PageWriteback(newpage)) 287 end_page_writeback(newpage); 288 } 289 290 /************************************************************ 291 * Migration functions 292 ***********************************************************/ 293 294 /* Always fail migration. Used for mappings that are not movable */ 295 int fail_migrate_page(struct page *newpage, struct page *page) 296 { 297 return -EIO; 298 } 299 EXPORT_SYMBOL(fail_migrate_page); 300 301 /* 302 * Common logic to directly migrate a single page suitable for 303 * pages that do not use PagePrivate. 304 * 305 * Pages are locked upon entry and exit. 306 */ 307 int migrate_page(struct page *newpage, struct page *page) 308 { 309 int rc; 310 311 BUG_ON(PageWriteback(page)); /* Writeback must be complete */ 312 313 rc = migrate_page_remove_references(newpage, page); 314 315 if (rc) 316 return rc; 317 318 migrate_page_copy(newpage, page); 319 320 /* 321 * Remove auxiliary swap entries and replace 322 * them with real ptes. 323 * 324 * Note that a real pte entry will allow processes that are not 325 * waiting on the page lock to use the new page via the page tables 326 * before the new page is unlocked. 327 */ 328 remove_from_swap(newpage); 329 return 0; 330 } 331 EXPORT_SYMBOL(migrate_page); 332 333 /* 334 * Migration function for pages with buffers. This function can only be used 335 * if the underlying filesystem guarantees that no other references to "page" 336 * exist. 337 */ 338 int buffer_migrate_page(struct page *newpage, struct page *page) 339 { 340 struct address_space *mapping = page->mapping; 341 struct buffer_head *bh, *head; 342 int rc; 343 344 if (!mapping) 345 return -EAGAIN; 346 347 if (!page_has_buffers(page)) 348 return migrate_page(newpage, page); 349 350 head = page_buffers(page); 351 352 rc = migrate_page_remove_references(newpage, page); 353 354 if (rc) 355 return rc; 356 357 bh = head; 358 do { 359 get_bh(bh); 360 lock_buffer(bh); 361 bh = bh->b_this_page; 362 363 } while (bh != head); 364 365 ClearPagePrivate(page); 366 set_page_private(newpage, page_private(page)); 367 set_page_private(page, 0); 368 put_page(page); 369 get_page(newpage); 370 371 bh = head; 372 do { 373 set_bh_page(bh, newpage, bh_offset(bh)); 374 bh = bh->b_this_page; 375 376 } while (bh != head); 377 378 SetPagePrivate(newpage); 379 380 migrate_page_copy(newpage, page); 381 382 bh = head; 383 do { 384 unlock_buffer(bh); 385 put_bh(bh); 386 bh = bh->b_this_page; 387 388 } while (bh != head); 389 390 return 0; 391 } 392 EXPORT_SYMBOL(buffer_migrate_page); 393 394 /* 395 * migrate_pages 396 * 397 * Two lists are passed to this function. The first list 398 * contains the pages isolated from the LRU to be migrated. 399 * The second list contains new pages that the pages isolated 400 * can be moved to. If the second list is NULL then all 401 * pages are swapped out. 402 * 403 * The function returns after 10 attempts or if no pages 404 * are movable anymore because to has become empty 405 * or no retryable pages exist anymore. 406 * 407 * Return: Number of pages not migrated when "to" ran empty. 408 */ 409 int migrate_pages(struct list_head *from, struct list_head *to, 410 struct list_head *moved, struct list_head *failed) 411 { 412 int retry; 413 int nr_failed = 0; 414 int pass = 0; 415 struct page *page; 416 struct page *page2; 417 int swapwrite = current->flags & PF_SWAPWRITE; 418 int rc; 419 420 if (!swapwrite) 421 current->flags |= PF_SWAPWRITE; 422 423 redo: 424 retry = 0; 425 426 list_for_each_entry_safe(page, page2, from, lru) { 427 struct page *newpage = NULL; 428 struct address_space *mapping; 429 430 cond_resched(); 431 432 rc = 0; 433 if (page_count(page) == 1) 434 /* page was freed from under us. So we are done. */ 435 goto next; 436 437 if (to && list_empty(to)) 438 break; 439 440 /* 441 * Skip locked pages during the first two passes to give the 442 * functions holding the lock time to release the page. Later we 443 * use lock_page() to have a higher chance of acquiring the 444 * lock. 445 */ 446 rc = -EAGAIN; 447 if (pass > 2) 448 lock_page(page); 449 else 450 if (TestSetPageLocked(page)) 451 goto next; 452 453 /* 454 * Only wait on writeback if we have already done a pass where 455 * we we may have triggered writeouts for lots of pages. 456 */ 457 if (pass > 0) { 458 wait_on_page_writeback(page); 459 } else { 460 if (PageWriteback(page)) 461 goto unlock_page; 462 } 463 464 /* 465 * Anonymous pages must have swap cache references otherwise 466 * the information contained in the page maps cannot be 467 * preserved. 468 */ 469 if (PageAnon(page) && !PageSwapCache(page)) { 470 if (!add_to_swap(page, GFP_KERNEL)) { 471 rc = -ENOMEM; 472 goto unlock_page; 473 } 474 } 475 476 if (!to) { 477 rc = swap_page(page); 478 goto next; 479 } 480 481 newpage = lru_to_page(to); 482 lock_page(newpage); 483 484 /* 485 * Pages are properly locked and writeback is complete. 486 * Try to migrate the page. 487 */ 488 mapping = page_mapping(page); 489 if (!mapping) 490 goto unlock_both; 491 492 if (mapping->a_ops->migratepage) { 493 /* 494 * Most pages have a mapping and most filesystems 495 * should provide a migration function. Anonymous 496 * pages are part of swap space which also has its 497 * own migration function. This is the most common 498 * path for page migration. 499 */ 500 rc = mapping->a_ops->migratepage(newpage, page); 501 goto unlock_both; 502 } 503 504 /* Make sure the dirty bit is up to date */ 505 if (try_to_unmap(page, 1) == SWAP_FAIL) { 506 rc = -EPERM; 507 goto unlock_both; 508 } 509 510 if (page_mapcount(page)) { 511 rc = -EAGAIN; 512 goto unlock_both; 513 } 514 515 /* 516 * Default handling if a filesystem does not provide 517 * a migration function. We can only migrate clean 518 * pages so try to write out any dirty pages first. 519 */ 520 if (PageDirty(page)) { 521 switch (pageout(page, mapping)) { 522 case PAGE_KEEP: 523 case PAGE_ACTIVATE: 524 goto unlock_both; 525 526 case PAGE_SUCCESS: 527 unlock_page(newpage); 528 goto next; 529 530 case PAGE_CLEAN: 531 ; /* try to migrate the page below */ 532 } 533 } 534 535 /* 536 * Buffers are managed in a filesystem specific way. 537 * We must have no buffers or drop them. 538 */ 539 if (!page_has_buffers(page) || 540 try_to_release_page(page, GFP_KERNEL)) { 541 rc = migrate_page(newpage, page); 542 goto unlock_both; 543 } 544 545 /* 546 * On early passes with mapped pages simply 547 * retry. There may be a lock held for some 548 * buffers that may go away. Later 549 * swap them out. 550 */ 551 if (pass > 4) { 552 /* 553 * Persistently unable to drop buffers..... As a 554 * measure of last resort we fall back to 555 * swap_page(). 556 */ 557 unlock_page(newpage); 558 newpage = NULL; 559 rc = swap_page(page); 560 goto next; 561 } 562 563 unlock_both: 564 unlock_page(newpage); 565 566 unlock_page: 567 unlock_page(page); 568 569 next: 570 if (rc == -EAGAIN) { 571 retry++; 572 } else if (rc) { 573 /* Permanent failure */ 574 list_move(&page->lru, failed); 575 nr_failed++; 576 } else { 577 if (newpage) { 578 /* Successful migration. Return page to LRU */ 579 move_to_lru(newpage); 580 } 581 list_move(&page->lru, moved); 582 } 583 } 584 if (retry && pass++ < 10) 585 goto redo; 586 587 if (!swapwrite) 588 current->flags &= ~PF_SWAPWRITE; 589 590 return nr_failed + retry; 591 } 592 593 /* 594 * Migrate the list 'pagelist' of pages to a certain destination. 595 * 596 * Specify destination with either non-NULL vma or dest_node >= 0 597 * Return the number of pages not migrated or error code 598 */ 599 int migrate_pages_to(struct list_head *pagelist, 600 struct vm_area_struct *vma, int dest) 601 { 602 LIST_HEAD(newlist); 603 LIST_HEAD(moved); 604 LIST_HEAD(failed); 605 int err = 0; 606 unsigned long offset = 0; 607 int nr_pages; 608 struct page *page; 609 struct list_head *p; 610 611 redo: 612 nr_pages = 0; 613 list_for_each(p, pagelist) { 614 if (vma) { 615 /* 616 * The address passed to alloc_page_vma is used to 617 * generate the proper interleave behavior. We fake 618 * the address here by an increasing offset in order 619 * to get the proper distribution of pages. 620 * 621 * No decision has been made as to which page 622 * a certain old page is moved to so we cannot 623 * specify the correct address. 624 */ 625 page = alloc_page_vma(GFP_HIGHUSER, vma, 626 offset + vma->vm_start); 627 offset += PAGE_SIZE; 628 } 629 else 630 page = alloc_pages_node(dest, GFP_HIGHUSER, 0); 631 632 if (!page) { 633 err = -ENOMEM; 634 goto out; 635 } 636 list_add_tail(&page->lru, &newlist); 637 nr_pages++; 638 if (nr_pages > MIGRATE_CHUNK_SIZE) 639 break; 640 } 641 err = migrate_pages(pagelist, &newlist, &moved, &failed); 642 643 putback_lru_pages(&moved); /* Call release pages instead ?? */ 644 645 if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist)) 646 goto redo; 647 out: 648 /* Return leftover allocated pages */ 649 while (!list_empty(&newlist)) { 650 page = list_entry(newlist.next, struct page, lru); 651 list_del(&page->lru); 652 __free_page(page); 653 } 654 list_splice(&failed, pagelist); 655 if (err < 0) 656 return err; 657 658 /* Calculate number of leftover pages */ 659 nr_pages = 0; 660 list_for_each(p, pagelist) 661 nr_pages++; 662 return nr_pages; 663 } 664