1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2008, 2009 Intel Corporation 4 * Authors: Andi Kleen, Fengguang Wu 5 * 6 * High level machine check handler. Handles pages reported by the 7 * hardware as being corrupted usually due to a multi-bit ECC memory or cache 8 * failure. 9 * 10 * In addition there is a "soft offline" entry point that allows stop using 11 * not-yet-corrupted-by-suspicious pages without killing anything. 12 * 13 * Handles page cache pages in various states. The tricky part 14 * here is that we can access any page asynchronously in respect to 15 * other VM users, because memory failures could happen anytime and 16 * anywhere. This could violate some of their assumptions. This is why 17 * this code has to be extremely careful. Generally it tries to use 18 * normal locking rules, as in get the standard locks, even if that means 19 * the error handling takes potentially a long time. 20 * 21 * It can be very tempting to add handling for obscure cases here. 22 * In general any code for handling new cases should only be added iff: 23 * - You know how to test it. 24 * - You have a test that can be added to mce-test 25 * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/ 26 * - The case actually shows up as a frequent (top 10) page state in 27 * tools/vm/page-types when running a real workload. 28 * 29 * There are several operations here with exponential complexity because 30 * of unsuitable VM data structures. For example the operation to map back 31 * from RMAP chains to processes has to walk the complete process list and 32 * has non linear complexity with the number. But since memory corruptions 33 * are rare we hope to get away with this. This avoids impacting the core 34 * VM. 35 */ 36 #include <linux/kernel.h> 37 #include <linux/mm.h> 38 #include <linux/page-flags.h> 39 #include <linux/kernel-page-flags.h> 40 #include <linux/sched/signal.h> 41 #include <linux/sched/task.h> 42 #include <linux/ksm.h> 43 #include <linux/rmap.h> 44 #include <linux/export.h> 45 #include <linux/pagemap.h> 46 #include <linux/swap.h> 47 #include <linux/backing-dev.h> 48 #include <linux/migrate.h> 49 #include <linux/suspend.h> 50 #include <linux/slab.h> 51 #include <linux/swapops.h> 52 #include <linux/hugetlb.h> 53 #include <linux/memory_hotplug.h> 54 #include <linux/mm_inline.h> 55 #include <linux/memremap.h> 56 #include <linux/kfifo.h> 57 #include <linux/ratelimit.h> 58 #include <linux/page-isolation.h> 59 #include "internal.h" 60 #include "ras/ras_event.h" 61 62 int sysctl_memory_failure_early_kill __read_mostly = 0; 63 64 int sysctl_memory_failure_recovery __read_mostly = 1; 65 66 atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); 67 68 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) 69 70 u32 hwpoison_filter_enable = 0; 71 u32 hwpoison_filter_dev_major = ~0U; 72 u32 hwpoison_filter_dev_minor = ~0U; 73 u64 hwpoison_filter_flags_mask; 74 u64 hwpoison_filter_flags_value; 75 EXPORT_SYMBOL_GPL(hwpoison_filter_enable); 76 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); 77 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); 78 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); 79 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); 80 81 static int hwpoison_filter_dev(struct page *p) 82 { 83 struct address_space *mapping; 84 dev_t dev; 85 86 if (hwpoison_filter_dev_major == ~0U && 87 hwpoison_filter_dev_minor == ~0U) 88 return 0; 89 90 /* 91 * page_mapping() does not accept slab pages. 92 */ 93 if (PageSlab(p)) 94 return -EINVAL; 95 96 mapping = page_mapping(p); 97 if (mapping == NULL || mapping->host == NULL) 98 return -EINVAL; 99 100 dev = mapping->host->i_sb->s_dev; 101 if (hwpoison_filter_dev_major != ~0U && 102 hwpoison_filter_dev_major != MAJOR(dev)) 103 return -EINVAL; 104 if (hwpoison_filter_dev_minor != ~0U && 105 hwpoison_filter_dev_minor != MINOR(dev)) 106 return -EINVAL; 107 108 return 0; 109 } 110 111 static int hwpoison_filter_flags(struct page *p) 112 { 113 if (!hwpoison_filter_flags_mask) 114 return 0; 115 116 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == 117 hwpoison_filter_flags_value) 118 return 0; 119 else 120 return -EINVAL; 121 } 122 123 /* 124 * This allows stress tests to limit test scope to a collection of tasks 125 * by putting them under some memcg. This prevents killing unrelated/important 126 * processes such as /sbin/init. Note that the target task may share clean 127 * pages with init (eg. libc text), which is harmless. If the target task 128 * share _dirty_ pages with another task B, the test scheme must make sure B 129 * is also included in the memcg. At last, due to race conditions this filter 130 * can only guarantee that the page either belongs to the memcg tasks, or is 131 * a freed page. 132 */ 133 #ifdef CONFIG_MEMCG 134 u64 hwpoison_filter_memcg; 135 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); 136 static int hwpoison_filter_task(struct page *p) 137 { 138 if (!hwpoison_filter_memcg) 139 return 0; 140 141 if (page_cgroup_ino(p) != hwpoison_filter_memcg) 142 return -EINVAL; 143 144 return 0; 145 } 146 #else 147 static int hwpoison_filter_task(struct page *p) { return 0; } 148 #endif 149 150 int hwpoison_filter(struct page *p) 151 { 152 if (!hwpoison_filter_enable) 153 return 0; 154 155 if (hwpoison_filter_dev(p)) 156 return -EINVAL; 157 158 if (hwpoison_filter_flags(p)) 159 return -EINVAL; 160 161 if (hwpoison_filter_task(p)) 162 return -EINVAL; 163 164 return 0; 165 } 166 #else 167 int hwpoison_filter(struct page *p) 168 { 169 return 0; 170 } 171 #endif 172 173 EXPORT_SYMBOL_GPL(hwpoison_filter); 174 175 /* 176 * Kill all processes that have a poisoned page mapped and then isolate 177 * the page. 178 * 179 * General strategy: 180 * Find all processes having the page mapped and kill them. 181 * But we keep a page reference around so that the page is not 182 * actually freed yet. 183 * Then stash the page away 184 * 185 * There's no convenient way to get back to mapped processes 186 * from the VMAs. So do a brute-force search over all 187 * running processes. 188 * 189 * Remember that machine checks are not common (or rather 190 * if they are common you have other problems), so this shouldn't 191 * be a performance issue. 192 * 193 * Also there are some races possible while we get from the 194 * error detection to actually handle it. 195 */ 196 197 struct to_kill { 198 struct list_head nd; 199 struct task_struct *tsk; 200 unsigned long addr; 201 short size_shift; 202 }; 203 204 /* 205 * Send all the processes who have the page mapped a signal. 206 * ``action optional'' if they are not immediately affected by the error 207 * ``action required'' if error happened in current execution context 208 */ 209 static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags) 210 { 211 struct task_struct *t = tk->tsk; 212 short addr_lsb = tk->size_shift; 213 int ret = 0; 214 215 if ((t->mm == current->mm) || !(flags & MF_ACTION_REQUIRED)) 216 pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n", 217 pfn, t->comm, t->pid); 218 219 if (flags & MF_ACTION_REQUIRED) { 220 if (t->mm == current->mm) 221 ret = force_sig_mceerr(BUS_MCEERR_AR, 222 (void __user *)tk->addr, addr_lsb); 223 /* send no signal to non-current processes */ 224 } else { 225 /* 226 * Don't use force here, it's convenient if the signal 227 * can be temporarily blocked. 228 * This could cause a loop when the user sets SIGBUS 229 * to SIG_IGN, but hopefully no one will do that? 230 */ 231 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, 232 addr_lsb, t); /* synchronous? */ 233 } 234 if (ret < 0) 235 pr_info("Memory failure: Error sending signal to %s:%d: %d\n", 236 t->comm, t->pid, ret); 237 return ret; 238 } 239 240 /* 241 * When a unknown page type is encountered drain as many buffers as possible 242 * in the hope to turn the page into a LRU or free page, which we can handle. 243 */ 244 void shake_page(struct page *p, int access) 245 { 246 if (PageHuge(p)) 247 return; 248 249 if (!PageSlab(p)) { 250 lru_add_drain_all(); 251 if (PageLRU(p)) 252 return; 253 drain_all_pages(page_zone(p)); 254 if (PageLRU(p) || is_free_buddy_page(p)) 255 return; 256 } 257 258 /* 259 * Only call shrink_node_slabs here (which would also shrink 260 * other caches) if access is not potentially fatal. 261 */ 262 if (access) 263 drop_slab_node(page_to_nid(p)); 264 } 265 EXPORT_SYMBOL_GPL(shake_page); 266 267 static unsigned long dev_pagemap_mapping_shift(struct page *page, 268 struct vm_area_struct *vma) 269 { 270 unsigned long address = vma_address(page, vma); 271 pgd_t *pgd; 272 p4d_t *p4d; 273 pud_t *pud; 274 pmd_t *pmd; 275 pte_t *pte; 276 277 pgd = pgd_offset(vma->vm_mm, address); 278 if (!pgd_present(*pgd)) 279 return 0; 280 p4d = p4d_offset(pgd, address); 281 if (!p4d_present(*p4d)) 282 return 0; 283 pud = pud_offset(p4d, address); 284 if (!pud_present(*pud)) 285 return 0; 286 if (pud_devmap(*pud)) 287 return PUD_SHIFT; 288 pmd = pmd_offset(pud, address); 289 if (!pmd_present(*pmd)) 290 return 0; 291 if (pmd_devmap(*pmd)) 292 return PMD_SHIFT; 293 pte = pte_offset_map(pmd, address); 294 if (!pte_present(*pte)) 295 return 0; 296 if (pte_devmap(*pte)) 297 return PAGE_SHIFT; 298 return 0; 299 } 300 301 /* 302 * Failure handling: if we can't find or can't kill a process there's 303 * not much we can do. We just print a message and ignore otherwise. 304 */ 305 306 /* 307 * Schedule a process for later kill. 308 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. 309 */ 310 static void add_to_kill(struct task_struct *tsk, struct page *p, 311 struct vm_area_struct *vma, 312 struct list_head *to_kill) 313 { 314 struct to_kill *tk; 315 316 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); 317 if (!tk) { 318 pr_err("Memory failure: Out of memory while machine check handling\n"); 319 return; 320 } 321 322 tk->addr = page_address_in_vma(p, vma); 323 if (is_zone_device_page(p)) 324 tk->size_shift = dev_pagemap_mapping_shift(p, vma); 325 else 326 tk->size_shift = page_shift(compound_head(p)); 327 328 /* 329 * Send SIGKILL if "tk->addr == -EFAULT". Also, as 330 * "tk->size_shift" is always non-zero for !is_zone_device_page(), 331 * so "tk->size_shift == 0" effectively checks no mapping on 332 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times 333 * to a process' address space, it's possible not all N VMAs 334 * contain mappings for the page, but at least one VMA does. 335 * Only deliver SIGBUS with payload derived from the VMA that 336 * has a mapping for the page. 337 */ 338 if (tk->addr == -EFAULT) { 339 pr_info("Memory failure: Unable to find user space address %lx in %s\n", 340 page_to_pfn(p), tsk->comm); 341 } else if (tk->size_shift == 0) { 342 kfree(tk); 343 return; 344 } 345 346 get_task_struct(tsk); 347 tk->tsk = tsk; 348 list_add_tail(&tk->nd, to_kill); 349 } 350 351 /* 352 * Kill the processes that have been collected earlier. 353 * 354 * Only do anything when DOIT is set, otherwise just free the list 355 * (this is used for clean pages which do not need killing) 356 * Also when FAIL is set do a force kill because something went 357 * wrong earlier. 358 */ 359 static void kill_procs(struct list_head *to_kill, int forcekill, bool fail, 360 unsigned long pfn, int flags) 361 { 362 struct to_kill *tk, *next; 363 364 list_for_each_entry_safe (tk, next, to_kill, nd) { 365 if (forcekill) { 366 /* 367 * In case something went wrong with munmapping 368 * make sure the process doesn't catch the 369 * signal and then access the memory. Just kill it. 370 */ 371 if (fail || tk->addr == -EFAULT) { 372 pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", 373 pfn, tk->tsk->comm, tk->tsk->pid); 374 do_send_sig_info(SIGKILL, SEND_SIG_PRIV, 375 tk->tsk, PIDTYPE_PID); 376 } 377 378 /* 379 * In theory the process could have mapped 380 * something else on the address in-between. We could 381 * check for that, but we need to tell the 382 * process anyways. 383 */ 384 else if (kill_proc(tk, pfn, flags) < 0) 385 pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n", 386 pfn, tk->tsk->comm, tk->tsk->pid); 387 } 388 put_task_struct(tk->tsk); 389 kfree(tk); 390 } 391 } 392 393 /* 394 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) 395 * on behalf of the thread group. Return task_struct of the (first found) 396 * dedicated thread if found, and return NULL otherwise. 397 * 398 * We already hold read_lock(&tasklist_lock) in the caller, so we don't 399 * have to call rcu_read_lock/unlock() in this function. 400 */ 401 static struct task_struct *find_early_kill_thread(struct task_struct *tsk) 402 { 403 struct task_struct *t; 404 405 for_each_thread(tsk, t) 406 if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY)) 407 return t; 408 return NULL; 409 } 410 411 /* 412 * Determine whether a given process is "early kill" process which expects 413 * to be signaled when some page under the process is hwpoisoned. 414 * Return task_struct of the dedicated thread (main thread unless explicitly 415 * specified) if the process is "early kill," and otherwise returns NULL. 416 */ 417 static struct task_struct *task_early_kill(struct task_struct *tsk, 418 int force_early) 419 { 420 struct task_struct *t; 421 if (!tsk->mm) 422 return NULL; 423 if (force_early) 424 return tsk; 425 t = find_early_kill_thread(tsk); 426 if (t) 427 return t; 428 if (sysctl_memory_failure_early_kill) 429 return tsk; 430 return NULL; 431 } 432 433 /* 434 * Collect processes when the error hit an anonymous page. 435 */ 436 static void collect_procs_anon(struct page *page, struct list_head *to_kill, 437 int force_early) 438 { 439 struct vm_area_struct *vma; 440 struct task_struct *tsk; 441 struct anon_vma *av; 442 pgoff_t pgoff; 443 444 av = page_lock_anon_vma_read(page); 445 if (av == NULL) /* Not actually mapped anymore */ 446 return; 447 448 pgoff = page_to_pgoff(page); 449 read_lock(&tasklist_lock); 450 for_each_process (tsk) { 451 struct anon_vma_chain *vmac; 452 struct task_struct *t = task_early_kill(tsk, force_early); 453 454 if (!t) 455 continue; 456 anon_vma_interval_tree_foreach(vmac, &av->rb_root, 457 pgoff, pgoff) { 458 vma = vmac->vma; 459 if (!page_mapped_in_vma(page, vma)) 460 continue; 461 if (vma->vm_mm == t->mm) 462 add_to_kill(t, page, vma, to_kill); 463 } 464 } 465 read_unlock(&tasklist_lock); 466 page_unlock_anon_vma_read(av); 467 } 468 469 /* 470 * Collect processes when the error hit a file mapped page. 471 */ 472 static void collect_procs_file(struct page *page, struct list_head *to_kill, 473 int force_early) 474 { 475 struct vm_area_struct *vma; 476 struct task_struct *tsk; 477 struct address_space *mapping = page->mapping; 478 479 i_mmap_lock_read(mapping); 480 read_lock(&tasklist_lock); 481 for_each_process(tsk) { 482 pgoff_t pgoff = page_to_pgoff(page); 483 struct task_struct *t = task_early_kill(tsk, force_early); 484 485 if (!t) 486 continue; 487 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, 488 pgoff) { 489 /* 490 * Send early kill signal to tasks where a vma covers 491 * the page but the corrupted page is not necessarily 492 * mapped it in its pte. 493 * Assume applications who requested early kill want 494 * to be informed of all such data corruptions. 495 */ 496 if (vma->vm_mm == t->mm) 497 add_to_kill(t, page, vma, to_kill); 498 } 499 } 500 read_unlock(&tasklist_lock); 501 i_mmap_unlock_read(mapping); 502 } 503 504 /* 505 * Collect the processes who have the corrupted page mapped to kill. 506 */ 507 static void collect_procs(struct page *page, struct list_head *tokill, 508 int force_early) 509 { 510 if (!page->mapping) 511 return; 512 513 if (PageAnon(page)) 514 collect_procs_anon(page, tokill, force_early); 515 else 516 collect_procs_file(page, tokill, force_early); 517 } 518 519 static const char *action_name[] = { 520 [MF_IGNORED] = "Ignored", 521 [MF_FAILED] = "Failed", 522 [MF_DELAYED] = "Delayed", 523 [MF_RECOVERED] = "Recovered", 524 }; 525 526 static const char * const action_page_types[] = { 527 [MF_MSG_KERNEL] = "reserved kernel page", 528 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", 529 [MF_MSG_SLAB] = "kernel slab page", 530 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", 531 [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned", 532 [MF_MSG_HUGE] = "huge page", 533 [MF_MSG_FREE_HUGE] = "free huge page", 534 [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page", 535 [MF_MSG_UNMAP_FAILED] = "unmapping failed page", 536 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", 537 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", 538 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", 539 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", 540 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", 541 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", 542 [MF_MSG_DIRTY_LRU] = "dirty LRU page", 543 [MF_MSG_CLEAN_LRU] = "clean LRU page", 544 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", 545 [MF_MSG_BUDDY] = "free buddy page", 546 [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)", 547 [MF_MSG_DAX] = "dax page", 548 [MF_MSG_UNKNOWN] = "unknown page", 549 }; 550 551 /* 552 * XXX: It is possible that a page is isolated from LRU cache, 553 * and then kept in swap cache or failed to remove from page cache. 554 * The page count will stop it from being freed by unpoison. 555 * Stress tests should be aware of this memory leak problem. 556 */ 557 static int delete_from_lru_cache(struct page *p) 558 { 559 if (!isolate_lru_page(p)) { 560 /* 561 * Clear sensible page flags, so that the buddy system won't 562 * complain when the page is unpoison-and-freed. 563 */ 564 ClearPageActive(p); 565 ClearPageUnevictable(p); 566 567 /* 568 * Poisoned page might never drop its ref count to 0 so we have 569 * to uncharge it manually from its memcg. 570 */ 571 mem_cgroup_uncharge(p); 572 573 /* 574 * drop the page count elevated by isolate_lru_page() 575 */ 576 put_page(p); 577 return 0; 578 } 579 return -EIO; 580 } 581 582 static int truncate_error_page(struct page *p, unsigned long pfn, 583 struct address_space *mapping) 584 { 585 int ret = MF_FAILED; 586 587 if (mapping->a_ops->error_remove_page) { 588 int err = mapping->a_ops->error_remove_page(mapping, p); 589 590 if (err != 0) { 591 pr_info("Memory failure: %#lx: Failed to punch page: %d\n", 592 pfn, err); 593 } else if (page_has_private(p) && 594 !try_to_release_page(p, GFP_NOIO)) { 595 pr_info("Memory failure: %#lx: failed to release buffers\n", 596 pfn); 597 } else { 598 ret = MF_RECOVERED; 599 } 600 } else { 601 /* 602 * If the file system doesn't support it just invalidate 603 * This fails on dirty or anything with private pages 604 */ 605 if (invalidate_inode_page(p)) 606 ret = MF_RECOVERED; 607 else 608 pr_info("Memory failure: %#lx: Failed to invalidate\n", 609 pfn); 610 } 611 612 return ret; 613 } 614 615 /* 616 * Error hit kernel page. 617 * Do nothing, try to be lucky and not touch this instead. For a few cases we 618 * could be more sophisticated. 619 */ 620 static int me_kernel(struct page *p, unsigned long pfn) 621 { 622 return MF_IGNORED; 623 } 624 625 /* 626 * Page in unknown state. Do nothing. 627 */ 628 static int me_unknown(struct page *p, unsigned long pfn) 629 { 630 pr_err("Memory failure: %#lx: Unknown page state\n", pfn); 631 return MF_FAILED; 632 } 633 634 /* 635 * Clean (or cleaned) page cache page. 636 */ 637 static int me_pagecache_clean(struct page *p, unsigned long pfn) 638 { 639 struct address_space *mapping; 640 641 delete_from_lru_cache(p); 642 643 /* 644 * For anonymous pages we're done the only reference left 645 * should be the one m_f() holds. 646 */ 647 if (PageAnon(p)) 648 return MF_RECOVERED; 649 650 /* 651 * Now truncate the page in the page cache. This is really 652 * more like a "temporary hole punch" 653 * Don't do this for block devices when someone else 654 * has a reference, because it could be file system metadata 655 * and that's not safe to truncate. 656 */ 657 mapping = page_mapping(p); 658 if (!mapping) { 659 /* 660 * Page has been teared down in the meanwhile 661 */ 662 return MF_FAILED; 663 } 664 665 /* 666 * Truncation is a bit tricky. Enable it per file system for now. 667 * 668 * Open: to take i_mutex or not for this? Right now we don't. 669 */ 670 return truncate_error_page(p, pfn, mapping); 671 } 672 673 /* 674 * Dirty pagecache page 675 * Issues: when the error hit a hole page the error is not properly 676 * propagated. 677 */ 678 static int me_pagecache_dirty(struct page *p, unsigned long pfn) 679 { 680 struct address_space *mapping = page_mapping(p); 681 682 SetPageError(p); 683 /* TBD: print more information about the file. */ 684 if (mapping) { 685 /* 686 * IO error will be reported by write(), fsync(), etc. 687 * who check the mapping. 688 * This way the application knows that something went 689 * wrong with its dirty file data. 690 * 691 * There's one open issue: 692 * 693 * The EIO will be only reported on the next IO 694 * operation and then cleared through the IO map. 695 * Normally Linux has two mechanisms to pass IO error 696 * first through the AS_EIO flag in the address space 697 * and then through the PageError flag in the page. 698 * Since we drop pages on memory failure handling the 699 * only mechanism open to use is through AS_AIO. 700 * 701 * This has the disadvantage that it gets cleared on 702 * the first operation that returns an error, while 703 * the PageError bit is more sticky and only cleared 704 * when the page is reread or dropped. If an 705 * application assumes it will always get error on 706 * fsync, but does other operations on the fd before 707 * and the page is dropped between then the error 708 * will not be properly reported. 709 * 710 * This can already happen even without hwpoisoned 711 * pages: first on metadata IO errors (which only 712 * report through AS_EIO) or when the page is dropped 713 * at the wrong time. 714 * 715 * So right now we assume that the application DTRT on 716 * the first EIO, but we're not worse than other parts 717 * of the kernel. 718 */ 719 mapping_set_error(mapping, -EIO); 720 } 721 722 return me_pagecache_clean(p, pfn); 723 } 724 725 /* 726 * Clean and dirty swap cache. 727 * 728 * Dirty swap cache page is tricky to handle. The page could live both in page 729 * cache and swap cache(ie. page is freshly swapped in). So it could be 730 * referenced concurrently by 2 types of PTEs: 731 * normal PTEs and swap PTEs. We try to handle them consistently by calling 732 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, 733 * and then 734 * - clear dirty bit to prevent IO 735 * - remove from LRU 736 * - but keep in the swap cache, so that when we return to it on 737 * a later page fault, we know the application is accessing 738 * corrupted data and shall be killed (we installed simple 739 * interception code in do_swap_page to catch it). 740 * 741 * Clean swap cache pages can be directly isolated. A later page fault will 742 * bring in the known good data from disk. 743 */ 744 static int me_swapcache_dirty(struct page *p, unsigned long pfn) 745 { 746 ClearPageDirty(p); 747 /* Trigger EIO in shmem: */ 748 ClearPageUptodate(p); 749 750 if (!delete_from_lru_cache(p)) 751 return MF_DELAYED; 752 else 753 return MF_FAILED; 754 } 755 756 static int me_swapcache_clean(struct page *p, unsigned long pfn) 757 { 758 delete_from_swap_cache(p); 759 760 if (!delete_from_lru_cache(p)) 761 return MF_RECOVERED; 762 else 763 return MF_FAILED; 764 } 765 766 /* 767 * Huge pages. Needs work. 768 * Issues: 769 * - Error on hugepage is contained in hugepage unit (not in raw page unit.) 770 * To narrow down kill region to one page, we need to break up pmd. 771 */ 772 static int me_huge_page(struct page *p, unsigned long pfn) 773 { 774 int res = 0; 775 struct page *hpage = compound_head(p); 776 struct address_space *mapping; 777 778 if (!PageHuge(hpage)) 779 return MF_DELAYED; 780 781 mapping = page_mapping(hpage); 782 if (mapping) { 783 res = truncate_error_page(hpage, pfn, mapping); 784 } else { 785 unlock_page(hpage); 786 /* 787 * migration entry prevents later access on error anonymous 788 * hugepage, so we can free and dissolve it into buddy to 789 * save healthy subpages. 790 */ 791 if (PageAnon(hpage)) 792 put_page(hpage); 793 dissolve_free_huge_page(p); 794 res = MF_RECOVERED; 795 lock_page(hpage); 796 } 797 798 return res; 799 } 800 801 /* 802 * Various page states we can handle. 803 * 804 * A page state is defined by its current page->flags bits. 805 * The table matches them in order and calls the right handler. 806 * 807 * This is quite tricky because we can access page at any time 808 * in its live cycle, so all accesses have to be extremely careful. 809 * 810 * This is not complete. More states could be added. 811 * For any missing state don't attempt recovery. 812 */ 813 814 #define dirty (1UL << PG_dirty) 815 #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) 816 #define unevict (1UL << PG_unevictable) 817 #define mlock (1UL << PG_mlocked) 818 #define writeback (1UL << PG_writeback) 819 #define lru (1UL << PG_lru) 820 #define head (1UL << PG_head) 821 #define slab (1UL << PG_slab) 822 #define reserved (1UL << PG_reserved) 823 824 static struct page_state { 825 unsigned long mask; 826 unsigned long res; 827 enum mf_action_page_type type; 828 int (*action)(struct page *p, unsigned long pfn); 829 } error_states[] = { 830 { reserved, reserved, MF_MSG_KERNEL, me_kernel }, 831 /* 832 * free pages are specially detected outside this table: 833 * PG_buddy pages only make a small fraction of all free pages. 834 */ 835 836 /* 837 * Could in theory check if slab page is free or if we can drop 838 * currently unused objects without touching them. But just 839 * treat it as standard kernel for now. 840 */ 841 { slab, slab, MF_MSG_SLAB, me_kernel }, 842 843 { head, head, MF_MSG_HUGE, me_huge_page }, 844 845 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, 846 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, 847 848 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, 849 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, 850 851 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, 852 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, 853 854 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, 855 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, 856 857 /* 858 * Catchall entry: must be at end. 859 */ 860 { 0, 0, MF_MSG_UNKNOWN, me_unknown }, 861 }; 862 863 #undef dirty 864 #undef sc 865 #undef unevict 866 #undef mlock 867 #undef writeback 868 #undef lru 869 #undef head 870 #undef slab 871 #undef reserved 872 873 /* 874 * "Dirty/Clean" indication is not 100% accurate due to the possibility of 875 * setting PG_dirty outside page lock. See also comment above set_page_dirty(). 876 */ 877 static void action_result(unsigned long pfn, enum mf_action_page_type type, 878 enum mf_result result) 879 { 880 trace_memory_failure_event(pfn, type, result); 881 882 pr_err("Memory failure: %#lx: recovery action for %s: %s\n", 883 pfn, action_page_types[type], action_name[result]); 884 } 885 886 static int page_action(struct page_state *ps, struct page *p, 887 unsigned long pfn) 888 { 889 int result; 890 int count; 891 892 result = ps->action(p, pfn); 893 894 count = page_count(p) - 1; 895 if (ps->action == me_swapcache_dirty && result == MF_DELAYED) 896 count--; 897 if (count > 0) { 898 pr_err("Memory failure: %#lx: %s still referenced by %d users\n", 899 pfn, action_page_types[ps->type], count); 900 result = MF_FAILED; 901 } 902 action_result(pfn, ps->type, result); 903 904 /* Could do more checks here if page looks ok */ 905 /* 906 * Could adjust zone counters here to correct for the missing page. 907 */ 908 909 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; 910 } 911 912 /** 913 * get_hwpoison_page() - Get refcount for memory error handling: 914 * @page: raw error page (hit by memory error) 915 * 916 * Return: return 0 if failed to grab the refcount, otherwise true (some 917 * non-zero value.) 918 */ 919 int get_hwpoison_page(struct page *page) 920 { 921 struct page *head = compound_head(page); 922 923 if (!PageHuge(head) && PageTransHuge(head)) { 924 /* 925 * Non anonymous thp exists only in allocation/free time. We 926 * can't handle such a case correctly, so let's give it up. 927 * This should be better than triggering BUG_ON when kernel 928 * tries to touch the "partially handled" page. 929 */ 930 if (!PageAnon(head)) { 931 pr_err("Memory failure: %#lx: non anonymous thp\n", 932 page_to_pfn(page)); 933 return 0; 934 } 935 } 936 937 if (get_page_unless_zero(head)) { 938 if (head == compound_head(page)) 939 return 1; 940 941 pr_info("Memory failure: %#lx cannot catch tail\n", 942 page_to_pfn(page)); 943 put_page(head); 944 } 945 946 return 0; 947 } 948 EXPORT_SYMBOL_GPL(get_hwpoison_page); 949 950 /* 951 * Do all that is necessary to remove user space mappings. Unmap 952 * the pages and send SIGBUS to the processes if the data was dirty. 953 */ 954 static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, 955 int flags, struct page **hpagep) 956 { 957 enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; 958 struct address_space *mapping; 959 LIST_HEAD(tokill); 960 bool unmap_success = true; 961 int kill = 1, forcekill; 962 struct page *hpage = *hpagep; 963 bool mlocked = PageMlocked(hpage); 964 965 /* 966 * Here we are interested only in user-mapped pages, so skip any 967 * other types of pages. 968 */ 969 if (PageReserved(p) || PageSlab(p)) 970 return true; 971 if (!(PageLRU(hpage) || PageHuge(p))) 972 return true; 973 974 /* 975 * This check implies we don't kill processes if their pages 976 * are in the swap cache early. Those are always late kills. 977 */ 978 if (!page_mapped(hpage)) 979 return true; 980 981 if (PageKsm(p)) { 982 pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn); 983 return false; 984 } 985 986 if (PageSwapCache(p)) { 987 pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n", 988 pfn); 989 ttu |= TTU_IGNORE_HWPOISON; 990 } 991 992 /* 993 * Propagate the dirty bit from PTEs to struct page first, because we 994 * need this to decide if we should kill or just drop the page. 995 * XXX: the dirty test could be racy: set_page_dirty() may not always 996 * be called inside page lock (it's recommended but not enforced). 997 */ 998 mapping = page_mapping(hpage); 999 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && 1000 mapping_cap_writeback_dirty(mapping)) { 1001 if (page_mkclean(hpage)) { 1002 SetPageDirty(hpage); 1003 } else { 1004 kill = 0; 1005 ttu |= TTU_IGNORE_HWPOISON; 1006 pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n", 1007 pfn); 1008 } 1009 } 1010 1011 /* 1012 * First collect all the processes that have the page 1013 * mapped in dirty form. This has to be done before try_to_unmap, 1014 * because ttu takes the rmap data structures down. 1015 * 1016 * Error handling: We ignore errors here because 1017 * there's nothing that can be done. 1018 */ 1019 if (kill) 1020 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); 1021 1022 if (!PageHuge(hpage)) { 1023 unmap_success = try_to_unmap(hpage, ttu); 1024 } else { 1025 /* 1026 * For hugetlb pages, try_to_unmap could potentially call 1027 * huge_pmd_unshare. Because of this, take semaphore in 1028 * write mode here and set TTU_RMAP_LOCKED to indicate we 1029 * have taken the lock at this higer level. 1030 * 1031 * Note that the call to hugetlb_page_mapping_lock_write 1032 * is necessary even if mapping is already set. It handles 1033 * ugliness of potentially having to drop page lock to obtain 1034 * i_mmap_rwsem. 1035 */ 1036 mapping = hugetlb_page_mapping_lock_write(hpage); 1037 1038 if (mapping) { 1039 unmap_success = try_to_unmap(hpage, 1040 ttu|TTU_RMAP_LOCKED); 1041 i_mmap_unlock_write(mapping); 1042 } else { 1043 pr_info("Memory failure: %#lx: could not find mapping for mapped huge page\n", 1044 pfn); 1045 unmap_success = false; 1046 } 1047 } 1048 if (!unmap_success) 1049 pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n", 1050 pfn, page_mapcount(hpage)); 1051 1052 /* 1053 * try_to_unmap() might put mlocked page in lru cache, so call 1054 * shake_page() again to ensure that it's flushed. 1055 */ 1056 if (mlocked) 1057 shake_page(hpage, 0); 1058 1059 /* 1060 * Now that the dirty bit has been propagated to the 1061 * struct page and all unmaps done we can decide if 1062 * killing is needed or not. Only kill when the page 1063 * was dirty or the process is not restartable, 1064 * otherwise the tokill list is merely 1065 * freed. When there was a problem unmapping earlier 1066 * use a more force-full uncatchable kill to prevent 1067 * any accesses to the poisoned memory. 1068 */ 1069 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL); 1070 kill_procs(&tokill, forcekill, !unmap_success, pfn, flags); 1071 1072 return unmap_success; 1073 } 1074 1075 static int identify_page_state(unsigned long pfn, struct page *p, 1076 unsigned long page_flags) 1077 { 1078 struct page_state *ps; 1079 1080 /* 1081 * The first check uses the current page flags which may not have any 1082 * relevant information. The second check with the saved page flags is 1083 * carried out only if the first check can't determine the page status. 1084 */ 1085 for (ps = error_states;; ps++) 1086 if ((p->flags & ps->mask) == ps->res) 1087 break; 1088 1089 page_flags |= (p->flags & (1UL << PG_dirty)); 1090 1091 if (!ps->mask) 1092 for (ps = error_states;; ps++) 1093 if ((page_flags & ps->mask) == ps->res) 1094 break; 1095 return page_action(ps, p, pfn); 1096 } 1097 1098 static int memory_failure_hugetlb(unsigned long pfn, int flags) 1099 { 1100 struct page *p = pfn_to_page(pfn); 1101 struct page *head = compound_head(p); 1102 int res; 1103 unsigned long page_flags; 1104 1105 if (TestSetPageHWPoison(head)) { 1106 pr_err("Memory failure: %#lx: already hardware poisoned\n", 1107 pfn); 1108 return 0; 1109 } 1110 1111 num_poisoned_pages_inc(); 1112 1113 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) { 1114 /* 1115 * Check "filter hit" and "race with other subpage." 1116 */ 1117 lock_page(head); 1118 if (PageHWPoison(head)) { 1119 if ((hwpoison_filter(p) && TestClearPageHWPoison(p)) 1120 || (p != head && TestSetPageHWPoison(head))) { 1121 num_poisoned_pages_dec(); 1122 unlock_page(head); 1123 return 0; 1124 } 1125 } 1126 unlock_page(head); 1127 dissolve_free_huge_page(p); 1128 action_result(pfn, MF_MSG_FREE_HUGE, MF_DELAYED); 1129 return 0; 1130 } 1131 1132 lock_page(head); 1133 page_flags = head->flags; 1134 1135 if (!PageHWPoison(head)) { 1136 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); 1137 num_poisoned_pages_dec(); 1138 unlock_page(head); 1139 put_hwpoison_page(head); 1140 return 0; 1141 } 1142 1143 /* 1144 * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so 1145 * simply disable it. In order to make it work properly, we need 1146 * make sure that: 1147 * - conversion of a pud that maps an error hugetlb into hwpoison 1148 * entry properly works, and 1149 * - other mm code walking over page table is aware of pud-aligned 1150 * hwpoison entries. 1151 */ 1152 if (huge_page_size(page_hstate(head)) > PMD_SIZE) { 1153 action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED); 1154 res = -EBUSY; 1155 goto out; 1156 } 1157 1158 if (!hwpoison_user_mappings(p, pfn, flags, &head)) { 1159 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); 1160 res = -EBUSY; 1161 goto out; 1162 } 1163 1164 res = identify_page_state(pfn, p, page_flags); 1165 out: 1166 unlock_page(head); 1167 return res; 1168 } 1169 1170 static int memory_failure_dev_pagemap(unsigned long pfn, int flags, 1171 struct dev_pagemap *pgmap) 1172 { 1173 struct page *page = pfn_to_page(pfn); 1174 const bool unmap_success = true; 1175 unsigned long size = 0; 1176 struct to_kill *tk; 1177 LIST_HEAD(tokill); 1178 int rc = -EBUSY; 1179 loff_t start; 1180 dax_entry_t cookie; 1181 1182 /* 1183 * Prevent the inode from being freed while we are interrogating 1184 * the address_space, typically this would be handled by 1185 * lock_page(), but dax pages do not use the page lock. This 1186 * also prevents changes to the mapping of this pfn until 1187 * poison signaling is complete. 1188 */ 1189 cookie = dax_lock_page(page); 1190 if (!cookie) 1191 goto out; 1192 1193 if (hwpoison_filter(page)) { 1194 rc = 0; 1195 goto unlock; 1196 } 1197 1198 if (pgmap->type == MEMORY_DEVICE_PRIVATE) { 1199 /* 1200 * TODO: Handle HMM pages which may need coordination 1201 * with device-side memory. 1202 */ 1203 goto unlock; 1204 } 1205 1206 /* 1207 * Use this flag as an indication that the dax page has been 1208 * remapped UC to prevent speculative consumption of poison. 1209 */ 1210 SetPageHWPoison(page); 1211 1212 /* 1213 * Unlike System-RAM there is no possibility to swap in a 1214 * different physical page at a given virtual address, so all 1215 * userspace consumption of ZONE_DEVICE memory necessitates 1216 * SIGBUS (i.e. MF_MUST_KILL) 1217 */ 1218 flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; 1219 collect_procs(page, &tokill, flags & MF_ACTION_REQUIRED); 1220 1221 list_for_each_entry(tk, &tokill, nd) 1222 if (tk->size_shift) 1223 size = max(size, 1UL << tk->size_shift); 1224 if (size) { 1225 /* 1226 * Unmap the largest mapping to avoid breaking up 1227 * device-dax mappings which are constant size. The 1228 * actual size of the mapping being torn down is 1229 * communicated in siginfo, see kill_proc() 1230 */ 1231 start = (page->index << PAGE_SHIFT) & ~(size - 1); 1232 unmap_mapping_range(page->mapping, start, start + size, 0); 1233 } 1234 kill_procs(&tokill, flags & MF_MUST_KILL, !unmap_success, pfn, flags); 1235 rc = 0; 1236 unlock: 1237 dax_unlock_page(page, cookie); 1238 out: 1239 /* drop pgmap ref acquired in caller */ 1240 put_dev_pagemap(pgmap); 1241 action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED); 1242 return rc; 1243 } 1244 1245 /** 1246 * memory_failure - Handle memory failure of a page. 1247 * @pfn: Page Number of the corrupted page 1248 * @flags: fine tune action taken 1249 * 1250 * This function is called by the low level machine check code 1251 * of an architecture when it detects hardware memory corruption 1252 * of a page. It tries its best to recover, which includes 1253 * dropping pages, killing processes etc. 1254 * 1255 * The function is primarily of use for corruptions that 1256 * happen outside the current execution context (e.g. when 1257 * detected by a background scrubber) 1258 * 1259 * Must run in process context (e.g. a work queue) with interrupts 1260 * enabled and no spinlocks hold. 1261 */ 1262 int memory_failure(unsigned long pfn, int flags) 1263 { 1264 struct page *p; 1265 struct page *hpage; 1266 struct page *orig_head; 1267 struct dev_pagemap *pgmap; 1268 int res; 1269 unsigned long page_flags; 1270 1271 if (!sysctl_memory_failure_recovery) 1272 panic("Memory failure on page %lx", pfn); 1273 1274 p = pfn_to_online_page(pfn); 1275 if (!p) { 1276 if (pfn_valid(pfn)) { 1277 pgmap = get_dev_pagemap(pfn, NULL); 1278 if (pgmap) 1279 return memory_failure_dev_pagemap(pfn, flags, 1280 pgmap); 1281 } 1282 pr_err("Memory failure: %#lx: memory outside kernel control\n", 1283 pfn); 1284 return -ENXIO; 1285 } 1286 1287 if (PageHuge(p)) 1288 return memory_failure_hugetlb(pfn, flags); 1289 if (TestSetPageHWPoison(p)) { 1290 pr_err("Memory failure: %#lx: already hardware poisoned\n", 1291 pfn); 1292 return 0; 1293 } 1294 1295 orig_head = hpage = compound_head(p); 1296 num_poisoned_pages_inc(); 1297 1298 /* 1299 * We need/can do nothing about count=0 pages. 1300 * 1) it's a free page, and therefore in safe hand: 1301 * prep_new_page() will be the gate keeper. 1302 * 2) it's part of a non-compound high order page. 1303 * Implies some kernel user: cannot stop them from 1304 * R/W the page; let's pray that the page has been 1305 * used and will be freed some time later. 1306 * In fact it's dangerous to directly bump up page count from 0, 1307 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. 1308 */ 1309 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) { 1310 if (is_free_buddy_page(p)) { 1311 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED); 1312 return 0; 1313 } else { 1314 action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); 1315 return -EBUSY; 1316 } 1317 } 1318 1319 if (PageTransHuge(hpage)) { 1320 lock_page(p); 1321 if (!PageAnon(p) || unlikely(split_huge_page(p))) { 1322 unlock_page(p); 1323 if (!PageAnon(p)) 1324 pr_err("Memory failure: %#lx: non anonymous thp\n", 1325 pfn); 1326 else 1327 pr_err("Memory failure: %#lx: thp split failed\n", 1328 pfn); 1329 if (TestClearPageHWPoison(p)) 1330 num_poisoned_pages_dec(); 1331 put_hwpoison_page(p); 1332 return -EBUSY; 1333 } 1334 unlock_page(p); 1335 VM_BUG_ON_PAGE(!page_count(p), p); 1336 hpage = compound_head(p); 1337 } 1338 1339 /* 1340 * We ignore non-LRU pages for good reasons. 1341 * - PG_locked is only well defined for LRU pages and a few others 1342 * - to avoid races with __SetPageLocked() 1343 * - to avoid races with __SetPageSlab*() (and more non-atomic ops) 1344 * The check (unnecessarily) ignores LRU pages being isolated and 1345 * walked by the page reclaim code, however that's not a big loss. 1346 */ 1347 shake_page(p, 0); 1348 /* shake_page could have turned it free. */ 1349 if (!PageLRU(p) && is_free_buddy_page(p)) { 1350 if (flags & MF_COUNT_INCREASED) 1351 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED); 1352 else 1353 action_result(pfn, MF_MSG_BUDDY_2ND, MF_DELAYED); 1354 return 0; 1355 } 1356 1357 lock_page(p); 1358 1359 /* 1360 * The page could have changed compound pages during the locking. 1361 * If this happens just bail out. 1362 */ 1363 if (PageCompound(p) && compound_head(p) != orig_head) { 1364 action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); 1365 res = -EBUSY; 1366 goto out; 1367 } 1368 1369 /* 1370 * We use page flags to determine what action should be taken, but 1371 * the flags can be modified by the error containment action. One 1372 * example is an mlocked page, where PG_mlocked is cleared by 1373 * page_remove_rmap() in try_to_unmap_one(). So to determine page status 1374 * correctly, we save a copy of the page flags at this time. 1375 */ 1376 if (PageHuge(p)) 1377 page_flags = hpage->flags; 1378 else 1379 page_flags = p->flags; 1380 1381 /* 1382 * unpoison always clear PG_hwpoison inside page lock 1383 */ 1384 if (!PageHWPoison(p)) { 1385 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); 1386 num_poisoned_pages_dec(); 1387 unlock_page(p); 1388 put_hwpoison_page(p); 1389 return 0; 1390 } 1391 if (hwpoison_filter(p)) { 1392 if (TestClearPageHWPoison(p)) 1393 num_poisoned_pages_dec(); 1394 unlock_page(p); 1395 put_hwpoison_page(p); 1396 return 0; 1397 } 1398 1399 if (!PageTransTail(p) && !PageLRU(p)) 1400 goto identify_page_state; 1401 1402 /* 1403 * It's very difficult to mess with pages currently under IO 1404 * and in many cases impossible, so we just avoid it here. 1405 */ 1406 wait_on_page_writeback(p); 1407 1408 /* 1409 * Now take care of user space mappings. 1410 * Abort on fail: __delete_from_page_cache() assumes unmapped page. 1411 * 1412 * When the raw error page is thp tail page, hpage points to the raw 1413 * page after thp split. 1414 */ 1415 if (!hwpoison_user_mappings(p, pfn, flags, &hpage)) { 1416 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); 1417 res = -EBUSY; 1418 goto out; 1419 } 1420 1421 /* 1422 * Torn down by someone else? 1423 */ 1424 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { 1425 action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); 1426 res = -EBUSY; 1427 goto out; 1428 } 1429 1430 identify_page_state: 1431 res = identify_page_state(pfn, p, page_flags); 1432 out: 1433 unlock_page(p); 1434 return res; 1435 } 1436 EXPORT_SYMBOL_GPL(memory_failure); 1437 1438 #define MEMORY_FAILURE_FIFO_ORDER 4 1439 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) 1440 1441 struct memory_failure_entry { 1442 unsigned long pfn; 1443 int flags; 1444 }; 1445 1446 struct memory_failure_cpu { 1447 DECLARE_KFIFO(fifo, struct memory_failure_entry, 1448 MEMORY_FAILURE_FIFO_SIZE); 1449 spinlock_t lock; 1450 struct work_struct work; 1451 }; 1452 1453 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); 1454 1455 /** 1456 * memory_failure_queue - Schedule handling memory failure of a page. 1457 * @pfn: Page Number of the corrupted page 1458 * @flags: Flags for memory failure handling 1459 * 1460 * This function is called by the low level hardware error handler 1461 * when it detects hardware memory corruption of a page. It schedules 1462 * the recovering of error page, including dropping pages, killing 1463 * processes etc. 1464 * 1465 * The function is primarily of use for corruptions that 1466 * happen outside the current execution context (e.g. when 1467 * detected by a background scrubber) 1468 * 1469 * Can run in IRQ context. 1470 */ 1471 void memory_failure_queue(unsigned long pfn, int flags) 1472 { 1473 struct memory_failure_cpu *mf_cpu; 1474 unsigned long proc_flags; 1475 struct memory_failure_entry entry = { 1476 .pfn = pfn, 1477 .flags = flags, 1478 }; 1479 1480 mf_cpu = &get_cpu_var(memory_failure_cpu); 1481 spin_lock_irqsave(&mf_cpu->lock, proc_flags); 1482 if (kfifo_put(&mf_cpu->fifo, entry)) 1483 schedule_work_on(smp_processor_id(), &mf_cpu->work); 1484 else 1485 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n", 1486 pfn); 1487 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); 1488 put_cpu_var(memory_failure_cpu); 1489 } 1490 EXPORT_SYMBOL_GPL(memory_failure_queue); 1491 1492 static void memory_failure_work_func(struct work_struct *work) 1493 { 1494 struct memory_failure_cpu *mf_cpu; 1495 struct memory_failure_entry entry = { 0, }; 1496 unsigned long proc_flags; 1497 int gotten; 1498 1499 mf_cpu = container_of(work, struct memory_failure_cpu, work); 1500 for (;;) { 1501 spin_lock_irqsave(&mf_cpu->lock, proc_flags); 1502 gotten = kfifo_get(&mf_cpu->fifo, &entry); 1503 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); 1504 if (!gotten) 1505 break; 1506 if (entry.flags & MF_SOFT_OFFLINE) 1507 soft_offline_page(entry.pfn, entry.flags); 1508 else 1509 memory_failure(entry.pfn, entry.flags); 1510 } 1511 } 1512 1513 /* 1514 * Process memory_failure work queued on the specified CPU. 1515 * Used to avoid return-to-userspace racing with the memory_failure workqueue. 1516 */ 1517 void memory_failure_queue_kick(int cpu) 1518 { 1519 struct memory_failure_cpu *mf_cpu; 1520 1521 mf_cpu = &per_cpu(memory_failure_cpu, cpu); 1522 cancel_work_sync(&mf_cpu->work); 1523 memory_failure_work_func(&mf_cpu->work); 1524 } 1525 1526 static int __init memory_failure_init(void) 1527 { 1528 struct memory_failure_cpu *mf_cpu; 1529 int cpu; 1530 1531 for_each_possible_cpu(cpu) { 1532 mf_cpu = &per_cpu(memory_failure_cpu, cpu); 1533 spin_lock_init(&mf_cpu->lock); 1534 INIT_KFIFO(mf_cpu->fifo); 1535 INIT_WORK(&mf_cpu->work, memory_failure_work_func); 1536 } 1537 1538 return 0; 1539 } 1540 core_initcall(memory_failure_init); 1541 1542 #define unpoison_pr_info(fmt, pfn, rs) \ 1543 ({ \ 1544 if (__ratelimit(rs)) \ 1545 pr_info(fmt, pfn); \ 1546 }) 1547 1548 /** 1549 * unpoison_memory - Unpoison a previously poisoned page 1550 * @pfn: Page number of the to be unpoisoned page 1551 * 1552 * Software-unpoison a page that has been poisoned by 1553 * memory_failure() earlier. 1554 * 1555 * This is only done on the software-level, so it only works 1556 * for linux injected failures, not real hardware failures 1557 * 1558 * Returns 0 for success, otherwise -errno. 1559 */ 1560 int unpoison_memory(unsigned long pfn) 1561 { 1562 struct page *page; 1563 struct page *p; 1564 int freeit = 0; 1565 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, 1566 DEFAULT_RATELIMIT_BURST); 1567 1568 if (!pfn_valid(pfn)) 1569 return -ENXIO; 1570 1571 p = pfn_to_page(pfn); 1572 page = compound_head(p); 1573 1574 if (!PageHWPoison(p)) { 1575 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n", 1576 pfn, &unpoison_rs); 1577 return 0; 1578 } 1579 1580 if (page_count(page) > 1) { 1581 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n", 1582 pfn, &unpoison_rs); 1583 return 0; 1584 } 1585 1586 if (page_mapped(page)) { 1587 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n", 1588 pfn, &unpoison_rs); 1589 return 0; 1590 } 1591 1592 if (page_mapping(page)) { 1593 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n", 1594 pfn, &unpoison_rs); 1595 return 0; 1596 } 1597 1598 /* 1599 * unpoison_memory() can encounter thp only when the thp is being 1600 * worked by memory_failure() and the page lock is not held yet. 1601 * In such case, we yield to memory_failure() and make unpoison fail. 1602 */ 1603 if (!PageHuge(page) && PageTransHuge(page)) { 1604 unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n", 1605 pfn, &unpoison_rs); 1606 return 0; 1607 } 1608 1609 if (!get_hwpoison_page(p)) { 1610 if (TestClearPageHWPoison(p)) 1611 num_poisoned_pages_dec(); 1612 unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n", 1613 pfn, &unpoison_rs); 1614 return 0; 1615 } 1616 1617 lock_page(page); 1618 /* 1619 * This test is racy because PG_hwpoison is set outside of page lock. 1620 * That's acceptable because that won't trigger kernel panic. Instead, 1621 * the PG_hwpoison page will be caught and isolated on the entrance to 1622 * the free buddy page pool. 1623 */ 1624 if (TestClearPageHWPoison(page)) { 1625 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n", 1626 pfn, &unpoison_rs); 1627 num_poisoned_pages_dec(); 1628 freeit = 1; 1629 } 1630 unlock_page(page); 1631 1632 put_hwpoison_page(page); 1633 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1)) 1634 put_hwpoison_page(page); 1635 1636 return 0; 1637 } 1638 EXPORT_SYMBOL(unpoison_memory); 1639 1640 static struct page *new_page(struct page *p, unsigned long private) 1641 { 1642 int nid = page_to_nid(p); 1643 1644 return new_page_nodemask(p, nid, &node_states[N_MEMORY]); 1645 } 1646 1647 /* 1648 * Safely get reference count of an arbitrary page. 1649 * Returns 0 for a free page, -EIO for a zero refcount page 1650 * that is not free, and 1 for any other page type. 1651 * For 1 the page is returned with increased page count, otherwise not. 1652 */ 1653 static int __get_any_page(struct page *p, unsigned long pfn, int flags) 1654 { 1655 int ret; 1656 1657 if (flags & MF_COUNT_INCREASED) 1658 return 1; 1659 1660 /* 1661 * When the target page is a free hugepage, just remove it 1662 * from free hugepage list. 1663 */ 1664 if (!get_hwpoison_page(p)) { 1665 if (PageHuge(p)) { 1666 pr_info("%s: %#lx free huge page\n", __func__, pfn); 1667 ret = 0; 1668 } else if (is_free_buddy_page(p)) { 1669 pr_info("%s: %#lx free buddy page\n", __func__, pfn); 1670 ret = 0; 1671 } else { 1672 pr_info("%s: %#lx: unknown zero refcount page type %lx\n", 1673 __func__, pfn, p->flags); 1674 ret = -EIO; 1675 } 1676 } else { 1677 /* Not a free page */ 1678 ret = 1; 1679 } 1680 return ret; 1681 } 1682 1683 static int get_any_page(struct page *page, unsigned long pfn, int flags) 1684 { 1685 int ret = __get_any_page(page, pfn, flags); 1686 1687 if (ret == 1 && !PageHuge(page) && 1688 !PageLRU(page) && !__PageMovable(page)) { 1689 /* 1690 * Try to free it. 1691 */ 1692 put_hwpoison_page(page); 1693 shake_page(page, 1); 1694 1695 /* 1696 * Did it turn free? 1697 */ 1698 ret = __get_any_page(page, pfn, 0); 1699 if (ret == 1 && !PageLRU(page)) { 1700 /* Drop page reference which is from __get_any_page() */ 1701 put_hwpoison_page(page); 1702 pr_info("soft_offline: %#lx: unknown non LRU page type %lx (%pGp)\n", 1703 pfn, page->flags, &page->flags); 1704 return -EIO; 1705 } 1706 } 1707 return ret; 1708 } 1709 1710 static int soft_offline_huge_page(struct page *page, int flags) 1711 { 1712 int ret; 1713 unsigned long pfn = page_to_pfn(page); 1714 struct page *hpage = compound_head(page); 1715 LIST_HEAD(pagelist); 1716 1717 /* 1718 * This double-check of PageHWPoison is to avoid the race with 1719 * memory_failure(). See also comment in __soft_offline_page(). 1720 */ 1721 lock_page(hpage); 1722 if (PageHWPoison(hpage)) { 1723 unlock_page(hpage); 1724 put_hwpoison_page(hpage); 1725 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn); 1726 return -EBUSY; 1727 } 1728 unlock_page(hpage); 1729 1730 ret = isolate_huge_page(hpage, &pagelist); 1731 /* 1732 * get_any_page() and isolate_huge_page() takes a refcount each, 1733 * so need to drop one here. 1734 */ 1735 put_hwpoison_page(hpage); 1736 if (!ret) { 1737 pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn); 1738 return -EBUSY; 1739 } 1740 1741 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL, 1742 MIGRATE_SYNC, MR_MEMORY_FAILURE); 1743 if (ret) { 1744 pr_info("soft offline: %#lx: hugepage migration failed %d, type %lx (%pGp)\n", 1745 pfn, ret, page->flags, &page->flags); 1746 if (!list_empty(&pagelist)) 1747 putback_movable_pages(&pagelist); 1748 if (ret > 0) 1749 ret = -EIO; 1750 } else { 1751 /* 1752 * We set PG_hwpoison only when the migration source hugepage 1753 * was successfully dissolved, because otherwise hwpoisoned 1754 * hugepage remains on free hugepage list, then userspace will 1755 * find it as SIGBUS by allocation failure. That's not expected 1756 * in soft-offlining. 1757 */ 1758 ret = dissolve_free_huge_page(page); 1759 if (!ret) { 1760 if (set_hwpoison_free_buddy_page(page)) 1761 num_poisoned_pages_inc(); 1762 else 1763 ret = -EBUSY; 1764 } 1765 } 1766 return ret; 1767 } 1768 1769 static int __soft_offline_page(struct page *page, int flags) 1770 { 1771 int ret; 1772 unsigned long pfn = page_to_pfn(page); 1773 1774 /* 1775 * Check PageHWPoison again inside page lock because PageHWPoison 1776 * is set by memory_failure() outside page lock. Note that 1777 * memory_failure() also double-checks PageHWPoison inside page lock, 1778 * so there's no race between soft_offline_page() and memory_failure(). 1779 */ 1780 lock_page(page); 1781 wait_on_page_writeback(page); 1782 if (PageHWPoison(page)) { 1783 unlock_page(page); 1784 put_hwpoison_page(page); 1785 pr_info("soft offline: %#lx page already poisoned\n", pfn); 1786 return -EBUSY; 1787 } 1788 /* 1789 * Try to invalidate first. This should work for 1790 * non dirty unmapped page cache pages. 1791 */ 1792 ret = invalidate_inode_page(page); 1793 unlock_page(page); 1794 /* 1795 * RED-PEN would be better to keep it isolated here, but we 1796 * would need to fix isolation locking first. 1797 */ 1798 if (ret == 1) { 1799 put_hwpoison_page(page); 1800 pr_info("soft_offline: %#lx: invalidated\n", pfn); 1801 SetPageHWPoison(page); 1802 num_poisoned_pages_inc(); 1803 return 0; 1804 } 1805 1806 /* 1807 * Simple invalidation didn't work. 1808 * Try to migrate to a new page instead. migrate.c 1809 * handles a large number of cases for us. 1810 */ 1811 if (PageLRU(page)) 1812 ret = isolate_lru_page(page); 1813 else 1814 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE); 1815 /* 1816 * Drop page reference which is came from get_any_page() 1817 * successful isolate_lru_page() already took another one. 1818 */ 1819 put_hwpoison_page(page); 1820 if (!ret) { 1821 LIST_HEAD(pagelist); 1822 /* 1823 * After isolated lru page, the PageLRU will be cleared, 1824 * so use !__PageMovable instead for LRU page's mapping 1825 * cannot have PAGE_MAPPING_MOVABLE. 1826 */ 1827 if (!__PageMovable(page)) 1828 inc_node_page_state(page, NR_ISOLATED_ANON + 1829 page_is_file_lru(page)); 1830 list_add(&page->lru, &pagelist); 1831 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL, 1832 MIGRATE_SYNC, MR_MEMORY_FAILURE); 1833 if (ret) { 1834 if (!list_empty(&pagelist)) 1835 putback_movable_pages(&pagelist); 1836 1837 pr_info("soft offline: %#lx: migration failed %d, type %lx (%pGp)\n", 1838 pfn, ret, page->flags, &page->flags); 1839 if (ret > 0) 1840 ret = -EIO; 1841 } 1842 } else { 1843 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx (%pGp)\n", 1844 pfn, ret, page_count(page), page->flags, &page->flags); 1845 } 1846 return ret; 1847 } 1848 1849 static int soft_offline_in_use_page(struct page *page, int flags) 1850 { 1851 int ret; 1852 int mt; 1853 struct page *hpage = compound_head(page); 1854 1855 if (!PageHuge(page) && PageTransHuge(hpage)) { 1856 lock_page(page); 1857 if (!PageAnon(page) || unlikely(split_huge_page(page))) { 1858 unlock_page(page); 1859 if (!PageAnon(page)) 1860 pr_info("soft offline: %#lx: non anonymous thp\n", page_to_pfn(page)); 1861 else 1862 pr_info("soft offline: %#lx: thp split failed\n", page_to_pfn(page)); 1863 put_hwpoison_page(page); 1864 return -EBUSY; 1865 } 1866 unlock_page(page); 1867 } 1868 1869 /* 1870 * Setting MIGRATE_ISOLATE here ensures that the page will be linked 1871 * to free list immediately (not via pcplist) when released after 1872 * successful page migration. Otherwise we can't guarantee that the 1873 * page is really free after put_page() returns, so 1874 * set_hwpoison_free_buddy_page() highly likely fails. 1875 */ 1876 mt = get_pageblock_migratetype(page); 1877 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 1878 if (PageHuge(page)) 1879 ret = soft_offline_huge_page(page, flags); 1880 else 1881 ret = __soft_offline_page(page, flags); 1882 set_pageblock_migratetype(page, mt); 1883 return ret; 1884 } 1885 1886 static int soft_offline_free_page(struct page *page) 1887 { 1888 int rc = dissolve_free_huge_page(page); 1889 1890 if (!rc) { 1891 if (set_hwpoison_free_buddy_page(page)) 1892 num_poisoned_pages_inc(); 1893 else 1894 rc = -EBUSY; 1895 } 1896 return rc; 1897 } 1898 1899 /** 1900 * soft_offline_page - Soft offline a page. 1901 * @pfn: pfn to soft-offline 1902 * @flags: flags. Same as memory_failure(). 1903 * 1904 * Returns 0 on success, otherwise negated errno. 1905 * 1906 * Soft offline a page, by migration or invalidation, 1907 * without killing anything. This is for the case when 1908 * a page is not corrupted yet (so it's still valid to access), 1909 * but has had a number of corrected errors and is better taken 1910 * out. 1911 * 1912 * The actual policy on when to do that is maintained by 1913 * user space. 1914 * 1915 * This should never impact any application or cause data loss, 1916 * however it might take some time. 1917 * 1918 * This is not a 100% solution for all memory, but tries to be 1919 * ``good enough'' for the majority of memory. 1920 */ 1921 int soft_offline_page(unsigned long pfn, int flags) 1922 { 1923 int ret; 1924 struct page *page; 1925 1926 if (!pfn_valid(pfn)) 1927 return -ENXIO; 1928 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */ 1929 page = pfn_to_online_page(pfn); 1930 if (!page) 1931 return -EIO; 1932 1933 if (PageHWPoison(page)) { 1934 pr_info("soft offline: %#lx page already poisoned\n", pfn); 1935 if (flags & MF_COUNT_INCREASED) 1936 put_hwpoison_page(page); 1937 return -EBUSY; 1938 } 1939 1940 get_online_mems(); 1941 ret = get_any_page(page, pfn, flags); 1942 put_online_mems(); 1943 1944 if (ret > 0) 1945 ret = soft_offline_in_use_page(page, flags); 1946 else if (ret == 0) 1947 ret = soft_offline_free_page(page); 1948 1949 return ret; 1950 } 1951