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