1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2008, 2009 Intel Corporation
4 * Authors: Andi Kleen, Fengguang Wu
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
10 * In addition there is a "soft offline" entry point that allows stop using
11 * not-yet-corrupted-by-suspicious pages without killing anything.
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.
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.
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
36 #include <linux/kernel.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/dax.h>
43 #include <linux/ksm.h>
44 #include <linux/rmap.h>
45 #include <linux/export.h>
46 #include <linux/pagemap.h>
47 #include <linux/swap.h>
48 #include <linux/backing-dev.h>
49 #include <linux/migrate.h>
50 #include <linux/suspend.h>
51 #include <linux/slab.h>
52 #include <linux/swapops.h>
53 #include <linux/hugetlb.h>
54 #include <linux/memory_hotplug.h>
55 #include <linux/mm_inline.h>
56 #include <linux/memremap.h>
57 #include <linux/kfifo.h>
58 #include <linux/ratelimit.h>
59 #include <linux/page-isolation.h>
60 #include <linux/pagewalk.h>
62 #include "ras/ras_event.h"
64 int sysctl_memory_failure_early_kill __read_mostly = 0;
66 int sysctl_memory_failure_recovery __read_mostly = 1;
68 atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
70 static bool __page_handle_poison(struct page *page)
74 zone_pcp_disable(page_zone(page));
75 ret = dissolve_free_huge_page(page);
77 ret = take_page_off_buddy(page);
78 zone_pcp_enable(page_zone(page));
83 static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release)
85 if (hugepage_or_freepage) {
87 * Doing this check for free pages is also fine since dissolve_free_huge_page
88 * returns 0 for non-hugetlb pages as well.
90 if (!__page_handle_poison(page))
92 * We could fail to take off the target page from buddy
93 * for example due to racy page allocation, but that's
94 * acceptable because soft-offlined page is not broken
95 * and if someone really want to use it, they should
101 SetPageHWPoison(page);
105 num_poisoned_pages_inc();
110 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
112 u32 hwpoison_filter_enable = 0;
113 u32 hwpoison_filter_dev_major = ~0U;
114 u32 hwpoison_filter_dev_minor = ~0U;
115 u64 hwpoison_filter_flags_mask;
116 u64 hwpoison_filter_flags_value;
117 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
118 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
119 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
120 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
121 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
123 static int hwpoison_filter_dev(struct page *p)
125 struct address_space *mapping;
128 if (hwpoison_filter_dev_major == ~0U &&
129 hwpoison_filter_dev_minor == ~0U)
133 * page_mapping() does not accept slab pages.
138 mapping = page_mapping(p);
139 if (mapping == NULL || mapping->host == NULL)
142 dev = mapping->host->i_sb->s_dev;
143 if (hwpoison_filter_dev_major != ~0U &&
144 hwpoison_filter_dev_major != MAJOR(dev))
146 if (hwpoison_filter_dev_minor != ~0U &&
147 hwpoison_filter_dev_minor != MINOR(dev))
153 static int hwpoison_filter_flags(struct page *p)
155 if (!hwpoison_filter_flags_mask)
158 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
159 hwpoison_filter_flags_value)
166 * This allows stress tests to limit test scope to a collection of tasks
167 * by putting them under some memcg. This prevents killing unrelated/important
168 * processes such as /sbin/init. Note that the target task may share clean
169 * pages with init (eg. libc text), which is harmless. If the target task
170 * share _dirty_ pages with another task B, the test scheme must make sure B
171 * is also included in the memcg. At last, due to race conditions this filter
172 * can only guarantee that the page either belongs to the memcg tasks, or is
176 u64 hwpoison_filter_memcg;
177 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
178 static int hwpoison_filter_task(struct page *p)
180 if (!hwpoison_filter_memcg)
183 if (page_cgroup_ino(p) != hwpoison_filter_memcg)
189 static int hwpoison_filter_task(struct page *p) { return 0; }
192 int hwpoison_filter(struct page *p)
194 if (!hwpoison_filter_enable)
197 if (hwpoison_filter_dev(p))
200 if (hwpoison_filter_flags(p))
203 if (hwpoison_filter_task(p))
209 int hwpoison_filter(struct page *p)
215 EXPORT_SYMBOL_GPL(hwpoison_filter);
218 * Kill all processes that have a poisoned page mapped and then isolate
222 * Find all processes having the page mapped and kill them.
223 * But we keep a page reference around so that the page is not
224 * actually freed yet.
225 * Then stash the page away
227 * There's no convenient way to get back to mapped processes
228 * from the VMAs. So do a brute-force search over all
231 * Remember that machine checks are not common (or rather
232 * if they are common you have other problems), so this shouldn't
233 * be a performance issue.
235 * Also there are some races possible while we get from the
236 * error detection to actually handle it.
241 struct task_struct *tsk;
247 * Send all the processes who have the page mapped a signal.
248 * ``action optional'' if they are not immediately affected by the error
249 * ``action required'' if error happened in current execution context
251 static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
253 struct task_struct *t = tk->tsk;
254 short addr_lsb = tk->size_shift;
257 pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
258 pfn, t->comm, t->pid);
260 if (flags & MF_ACTION_REQUIRED) {
262 ret = force_sig_mceerr(BUS_MCEERR_AR,
263 (void __user *)tk->addr, addr_lsb);
265 /* Signal other processes sharing the page if they have PF_MCE_EARLY set. */
266 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
270 * Don't use force here, it's convenient if the signal
271 * can be temporarily blocked.
272 * This could cause a loop when the user sets SIGBUS
273 * to SIG_IGN, but hopefully no one will do that?
275 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
276 addr_lsb, t); /* synchronous? */
279 pr_info("Memory failure: Error sending signal to %s:%d: %d\n",
280 t->comm, t->pid, ret);
285 * Unknown page type encountered. Try to check whether it can turn PageLRU by
288 void shake_page(struct page *p)
295 if (PageLRU(p) || is_free_buddy_page(p))
300 * TODO: Could shrink slab caches here if a lightweight range-based
301 * shrinker will be available.
304 EXPORT_SYMBOL_GPL(shake_page);
306 static unsigned long dev_pagemap_mapping_shift(struct page *page,
307 struct vm_area_struct *vma)
309 unsigned long address = vma_address(page, vma);
310 unsigned long ret = 0;
317 pgd = pgd_offset(vma->vm_mm, address);
318 if (!pgd_present(*pgd))
320 p4d = p4d_offset(pgd, address);
321 if (!p4d_present(*p4d))
323 pud = pud_offset(p4d, address);
324 if (!pud_present(*pud))
326 if (pud_devmap(*pud))
328 pmd = pmd_offset(pud, address);
329 if (!pmd_present(*pmd))
331 if (pmd_devmap(*pmd))
333 pte = pte_offset_map(pmd, address);
334 if (pte_present(*pte) && pte_devmap(*pte))
341 * Failure handling: if we can't find or can't kill a process there's
342 * not much we can do. We just print a message and ignore otherwise.
346 * Schedule a process for later kill.
347 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
349 static void add_to_kill(struct task_struct *tsk, struct page *p,
350 struct vm_area_struct *vma,
351 struct list_head *to_kill)
355 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
357 pr_err("Memory failure: Out of memory while machine check handling\n");
361 tk->addr = page_address_in_vma(p, vma);
362 if (is_zone_device_page(p))
363 tk->size_shift = dev_pagemap_mapping_shift(p, vma);
365 tk->size_shift = page_shift(compound_head(p));
368 * Send SIGKILL if "tk->addr == -EFAULT". Also, as
369 * "tk->size_shift" is always non-zero for !is_zone_device_page(),
370 * so "tk->size_shift == 0" effectively checks no mapping on
371 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times
372 * to a process' address space, it's possible not all N VMAs
373 * contain mappings for the page, but at least one VMA does.
374 * Only deliver SIGBUS with payload derived from the VMA that
375 * has a mapping for the page.
377 if (tk->addr == -EFAULT) {
378 pr_info("Memory failure: Unable to find user space address %lx in %s\n",
379 page_to_pfn(p), tsk->comm);
380 } else if (tk->size_shift == 0) {
385 get_task_struct(tsk);
387 list_add_tail(&tk->nd, to_kill);
391 * Kill the processes that have been collected earlier.
393 * Only do anything when FORCEKILL is set, otherwise just free the
394 * list (this is used for clean pages which do not need killing)
395 * Also when FAIL is set do a force kill because something went
398 static void kill_procs(struct list_head *to_kill, int forcekill, bool fail,
399 unsigned long pfn, int flags)
401 struct to_kill *tk, *next;
403 list_for_each_entry_safe (tk, next, to_kill, nd) {
406 * In case something went wrong with munmapping
407 * make sure the process doesn't catch the
408 * signal and then access the memory. Just kill it.
410 if (fail || tk->addr == -EFAULT) {
411 pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
412 pfn, tk->tsk->comm, tk->tsk->pid);
413 do_send_sig_info(SIGKILL, SEND_SIG_PRIV,
414 tk->tsk, PIDTYPE_PID);
418 * In theory the process could have mapped
419 * something else on the address in-between. We could
420 * check for that, but we need to tell the
423 else if (kill_proc(tk, pfn, flags) < 0)
424 pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n",
425 pfn, tk->tsk->comm, tk->tsk->pid);
427 put_task_struct(tk->tsk);
433 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
434 * on behalf of the thread group. Return task_struct of the (first found)
435 * dedicated thread if found, and return NULL otherwise.
437 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
438 * have to call rcu_read_lock/unlock() in this function.
440 static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
442 struct task_struct *t;
444 for_each_thread(tsk, t) {
445 if (t->flags & PF_MCE_PROCESS) {
446 if (t->flags & PF_MCE_EARLY)
449 if (sysctl_memory_failure_early_kill)
457 * Determine whether a given process is "early kill" process which expects
458 * to be signaled when some page under the process is hwpoisoned.
459 * Return task_struct of the dedicated thread (main thread unless explicitly
460 * specified) if the process is "early kill" and otherwise returns NULL.
462 * Note that the above is true for Action Optional case. For Action Required
463 * case, it's only meaningful to the current thread which need to be signaled
464 * with SIGBUS, this error is Action Optional for other non current
465 * processes sharing the same error page,if the process is "early kill", the
466 * task_struct of the dedicated thread will also be returned.
468 static struct task_struct *task_early_kill(struct task_struct *tsk,
474 * Comparing ->mm here because current task might represent
475 * a subthread, while tsk always points to the main thread.
477 if (force_early && tsk->mm == current->mm)
480 return find_early_kill_thread(tsk);
484 * Collect processes when the error hit an anonymous page.
486 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
489 struct vm_area_struct *vma;
490 struct task_struct *tsk;
494 av = page_lock_anon_vma_read(page);
495 if (av == NULL) /* Not actually mapped anymore */
498 pgoff = page_to_pgoff(page);
499 read_lock(&tasklist_lock);
500 for_each_process (tsk) {
501 struct anon_vma_chain *vmac;
502 struct task_struct *t = task_early_kill(tsk, force_early);
506 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
509 if (!page_mapped_in_vma(page, vma))
511 if (vma->vm_mm == t->mm)
512 add_to_kill(t, page, vma, to_kill);
515 read_unlock(&tasklist_lock);
516 page_unlock_anon_vma_read(av);
520 * Collect processes when the error hit a file mapped page.
522 static void collect_procs_file(struct page *page, struct list_head *to_kill,
525 struct vm_area_struct *vma;
526 struct task_struct *tsk;
527 struct address_space *mapping = page->mapping;
530 i_mmap_lock_read(mapping);
531 read_lock(&tasklist_lock);
532 pgoff = page_to_pgoff(page);
533 for_each_process(tsk) {
534 struct task_struct *t = task_early_kill(tsk, force_early);
538 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
541 * Send early kill signal to tasks where a vma covers
542 * the page but the corrupted page is not necessarily
543 * mapped it in its pte.
544 * Assume applications who requested early kill want
545 * to be informed of all such data corruptions.
547 if (vma->vm_mm == t->mm)
548 add_to_kill(t, page, vma, to_kill);
551 read_unlock(&tasklist_lock);
552 i_mmap_unlock_read(mapping);
556 * Collect the processes who have the corrupted page mapped to kill.
558 static void collect_procs(struct page *page, struct list_head *tokill,
565 collect_procs_anon(page, tokill, force_early);
567 collect_procs_file(page, tokill, force_early);
576 static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift)
579 tk->size_shift = shift;
582 static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift,
583 unsigned long poisoned_pfn, struct to_kill *tk)
585 unsigned long pfn = 0;
587 if (pte_present(pte)) {
590 swp_entry_t swp = pte_to_swp_entry(pte);
592 if (is_hwpoison_entry(swp))
593 pfn = hwpoison_entry_to_pfn(swp);
596 if (!pfn || pfn != poisoned_pfn)
599 set_to_kill(tk, addr, shift);
603 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
604 static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
605 struct hwp_walk *hwp)
609 unsigned long hwpoison_vaddr;
611 if (!pmd_present(pmd))
614 if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) {
615 hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT);
616 set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT);
622 static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
623 struct hwp_walk *hwp)
629 static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr,
630 unsigned long end, struct mm_walk *walk)
632 struct hwp_walk *hwp = (struct hwp_walk *)walk->private;
634 pte_t *ptep, *mapped_pte;
637 ptl = pmd_trans_huge_lock(pmdp, walk->vma);
639 ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp);
644 if (pmd_trans_unstable(pmdp))
647 mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp,
649 for (; addr != end; ptep++, addr += PAGE_SIZE) {
650 ret = check_hwpoisoned_entry(*ptep, addr, PAGE_SHIFT,
655 pte_unmap_unlock(mapped_pte, ptl);
661 #ifdef CONFIG_HUGETLB_PAGE
662 static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask,
663 unsigned long addr, unsigned long end,
664 struct mm_walk *walk)
666 struct hwp_walk *hwp = (struct hwp_walk *)walk->private;
667 pte_t pte = huge_ptep_get(ptep);
668 struct hstate *h = hstate_vma(walk->vma);
670 return check_hwpoisoned_entry(pte, addr, huge_page_shift(h),
674 #define hwpoison_hugetlb_range NULL
677 static const struct mm_walk_ops hwp_walk_ops = {
678 .pmd_entry = hwpoison_pte_range,
679 .hugetlb_entry = hwpoison_hugetlb_range,
683 * Sends SIGBUS to the current process with error info.
685 * This function is intended to handle "Action Required" MCEs on already
686 * hardware poisoned pages. They could happen, for example, when
687 * memory_failure() failed to unmap the error page at the first call, or
688 * when multiple local machine checks happened on different CPUs.
690 * MCE handler currently has no easy access to the error virtual address,
691 * so this function walks page table to find it. The returned virtual address
692 * is proper in most cases, but it could be wrong when the application
693 * process has multiple entries mapping the error page.
695 static int kill_accessing_process(struct task_struct *p, unsigned long pfn,
699 struct hwp_walk priv = {
704 mmap_read_lock(p->mm);
705 ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwp_walk_ops,
707 if (ret == 1 && priv.tk.addr)
708 kill_proc(&priv.tk, pfn, flags);
709 mmap_read_unlock(p->mm);
710 return ret ? -EFAULT : -EHWPOISON;
713 static const char *action_name[] = {
714 [MF_IGNORED] = "Ignored",
715 [MF_FAILED] = "Failed",
716 [MF_DELAYED] = "Delayed",
717 [MF_RECOVERED] = "Recovered",
720 static const char * const action_page_types[] = {
721 [MF_MSG_KERNEL] = "reserved kernel page",
722 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
723 [MF_MSG_SLAB] = "kernel slab page",
724 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
725 [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned",
726 [MF_MSG_HUGE] = "huge page",
727 [MF_MSG_FREE_HUGE] = "free huge page",
728 [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page",
729 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
730 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
731 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
732 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
733 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
734 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
735 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
736 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
737 [MF_MSG_CLEAN_LRU] = "clean LRU page",
738 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
739 [MF_MSG_BUDDY] = "free buddy page",
740 [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)",
741 [MF_MSG_DAX] = "dax page",
742 [MF_MSG_UNSPLIT_THP] = "unsplit thp",
743 [MF_MSG_UNKNOWN] = "unknown page",
747 * XXX: It is possible that a page is isolated from LRU cache,
748 * and then kept in swap cache or failed to remove from page cache.
749 * The page count will stop it from being freed by unpoison.
750 * Stress tests should be aware of this memory leak problem.
752 static int delete_from_lru_cache(struct page *p)
754 if (!isolate_lru_page(p)) {
756 * Clear sensible page flags, so that the buddy system won't
757 * complain when the page is unpoison-and-freed.
760 ClearPageUnevictable(p);
763 * Poisoned page might never drop its ref count to 0 so we have
764 * to uncharge it manually from its memcg.
766 mem_cgroup_uncharge(page_folio(p));
769 * drop the page count elevated by isolate_lru_page()
777 static int truncate_error_page(struct page *p, unsigned long pfn,
778 struct address_space *mapping)
782 if (mapping->a_ops->error_remove_page) {
783 int err = mapping->a_ops->error_remove_page(mapping, p);
786 pr_info("Memory failure: %#lx: Failed to punch page: %d\n",
788 } else if (page_has_private(p) &&
789 !try_to_release_page(p, GFP_NOIO)) {
790 pr_info("Memory failure: %#lx: failed to release buffers\n",
797 * If the file system doesn't support it just invalidate
798 * This fails on dirty or anything with private pages
800 if (invalidate_inode_page(p))
803 pr_info("Memory failure: %#lx: Failed to invalidate\n",
813 enum mf_action_page_type type;
815 /* Callback ->action() has to unlock the relevant page inside it. */
816 int (*action)(struct page_state *ps, struct page *p);
820 * Return true if page is still referenced by others, otherwise return
823 * The extra_pins is true when one extra refcount is expected.
825 static bool has_extra_refcount(struct page_state *ps, struct page *p,
828 int count = page_count(p) - 1;
834 pr_err("Memory failure: %#lx: %s still referenced by %d users\n",
835 page_to_pfn(p), action_page_types[ps->type], count);
843 * Error hit kernel page.
844 * Do nothing, try to be lucky and not touch this instead. For a few cases we
845 * could be more sophisticated.
847 static int me_kernel(struct page_state *ps, struct page *p)
854 * Page in unknown state. Do nothing.
856 static int me_unknown(struct page_state *ps, struct page *p)
858 pr_err("Memory failure: %#lx: Unknown page state\n", page_to_pfn(p));
864 * Clean (or cleaned) page cache page.
866 static int me_pagecache_clean(struct page_state *ps, struct page *p)
869 struct address_space *mapping;
871 delete_from_lru_cache(p);
874 * For anonymous pages we're done the only reference left
875 * should be the one m_f() holds.
883 * Now truncate the page in the page cache. This is really
884 * more like a "temporary hole punch"
885 * Don't do this for block devices when someone else
886 * has a reference, because it could be file system metadata
887 * and that's not safe to truncate.
889 mapping = page_mapping(p);
892 * Page has been teared down in the meanwhile
899 * Truncation is a bit tricky. Enable it per file system for now.
901 * Open: to take i_rwsem or not for this? Right now we don't.
903 ret = truncate_error_page(p, page_to_pfn(p), mapping);
907 if (has_extra_refcount(ps, p, false))
914 * Dirty pagecache page
915 * Issues: when the error hit a hole page the error is not properly
918 static int me_pagecache_dirty(struct page_state *ps, struct page *p)
920 struct address_space *mapping = page_mapping(p);
923 /* TBD: print more information about the file. */
926 * IO error will be reported by write(), fsync(), etc.
927 * who check the mapping.
928 * This way the application knows that something went
929 * wrong with its dirty file data.
931 * There's one open issue:
933 * The EIO will be only reported on the next IO
934 * operation and then cleared through the IO map.
935 * Normally Linux has two mechanisms to pass IO error
936 * first through the AS_EIO flag in the address space
937 * and then through the PageError flag in the page.
938 * Since we drop pages on memory failure handling the
939 * only mechanism open to use is through AS_AIO.
941 * This has the disadvantage that it gets cleared on
942 * the first operation that returns an error, while
943 * the PageError bit is more sticky and only cleared
944 * when the page is reread or dropped. If an
945 * application assumes it will always get error on
946 * fsync, but does other operations on the fd before
947 * and the page is dropped between then the error
948 * will not be properly reported.
950 * This can already happen even without hwpoisoned
951 * pages: first on metadata IO errors (which only
952 * report through AS_EIO) or when the page is dropped
955 * So right now we assume that the application DTRT on
956 * the first EIO, but we're not worse than other parts
959 mapping_set_error(mapping, -EIO);
962 return me_pagecache_clean(ps, p);
966 * Clean and dirty swap cache.
968 * Dirty swap cache page is tricky to handle. The page could live both in page
969 * cache and swap cache(ie. page is freshly swapped in). So it could be
970 * referenced concurrently by 2 types of PTEs:
971 * normal PTEs and swap PTEs. We try to handle them consistently by calling
972 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
974 * - clear dirty bit to prevent IO
976 * - but keep in the swap cache, so that when we return to it on
977 * a later page fault, we know the application is accessing
978 * corrupted data and shall be killed (we installed simple
979 * interception code in do_swap_page to catch it).
981 * Clean swap cache pages can be directly isolated. A later page fault will
982 * bring in the known good data from disk.
984 static int me_swapcache_dirty(struct page_state *ps, struct page *p)
987 bool extra_pins = false;
990 /* Trigger EIO in shmem: */
991 ClearPageUptodate(p);
993 ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED;
996 if (ret == MF_DELAYED)
999 if (has_extra_refcount(ps, p, extra_pins))
1005 static int me_swapcache_clean(struct page_state *ps, struct page *p)
1009 delete_from_swap_cache(p);
1011 ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED;
1014 if (has_extra_refcount(ps, p, false))
1021 * Huge pages. Needs work.
1023 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
1024 * To narrow down kill region to one page, we need to break up pmd.
1026 static int me_huge_page(struct page_state *ps, struct page *p)
1029 struct page *hpage = compound_head(p);
1030 struct address_space *mapping;
1032 if (!PageHuge(hpage))
1035 mapping = page_mapping(hpage);
1037 res = truncate_error_page(hpage, page_to_pfn(p), mapping);
1043 * migration entry prevents later access on error anonymous
1044 * hugepage, so we can free and dissolve it into buddy to
1045 * save healthy subpages.
1047 if (PageAnon(hpage))
1049 if (__page_handle_poison(p)) {
1055 if (has_extra_refcount(ps, p, false))
1062 * Various page states we can handle.
1064 * A page state is defined by its current page->flags bits.
1065 * The table matches them in order and calls the right handler.
1067 * This is quite tricky because we can access page at any time
1068 * in its live cycle, so all accesses have to be extremely careful.
1070 * This is not complete. More states could be added.
1071 * For any missing state don't attempt recovery.
1074 #define dirty (1UL << PG_dirty)
1075 #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked))
1076 #define unevict (1UL << PG_unevictable)
1077 #define mlock (1UL << PG_mlocked)
1078 #define lru (1UL << PG_lru)
1079 #define head (1UL << PG_head)
1080 #define slab (1UL << PG_slab)
1081 #define reserved (1UL << PG_reserved)
1083 static struct page_state error_states[] = {
1084 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
1086 * free pages are specially detected outside this table:
1087 * PG_buddy pages only make a small fraction of all free pages.
1091 * Could in theory check if slab page is free or if we can drop
1092 * currently unused objects without touching them. But just
1093 * treat it as standard kernel for now.
1095 { slab, slab, MF_MSG_SLAB, me_kernel },
1097 { head, head, MF_MSG_HUGE, me_huge_page },
1099 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
1100 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
1102 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
1103 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
1105 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
1106 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
1108 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
1109 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
1112 * Catchall entry: must be at end.
1114 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
1127 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
1128 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
1130 static void action_result(unsigned long pfn, enum mf_action_page_type type,
1131 enum mf_result result)
1133 trace_memory_failure_event(pfn, type, result);
1135 pr_err("Memory failure: %#lx: recovery action for %s: %s\n",
1136 pfn, action_page_types[type], action_name[result]);
1139 static int page_action(struct page_state *ps, struct page *p,
1144 /* page p should be unlocked after returning from ps->action(). */
1145 result = ps->action(ps, p);
1147 action_result(pfn, ps->type, result);
1149 /* Could do more checks here if page looks ok */
1151 * Could adjust zone counters here to correct for the missing page.
1154 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
1158 * Return true if a page type of a given page is supported by hwpoison
1159 * mechanism (while handling could fail), otherwise false. This function
1160 * does not return true for hugetlb or device memory pages, so it's assumed
1161 * to be called only in the context where we never have such pages.
1163 static inline bool HWPoisonHandlable(struct page *page)
1165 return PageLRU(page) || __PageMovable(page) || is_free_buddy_page(page);
1168 static int __get_hwpoison_page(struct page *page)
1170 struct page *head = compound_head(page);
1172 bool hugetlb = false;
1174 ret = get_hwpoison_huge_page(head, &hugetlb);
1179 * This check prevents from calling get_hwpoison_unless_zero()
1180 * for any unsupported type of page in order to reduce the risk of
1181 * unexpected races caused by taking a page refcount.
1183 if (!HWPoisonHandlable(head))
1186 if (get_page_unless_zero(head)) {
1187 if (head == compound_head(page))
1190 pr_info("Memory failure: %#lx cannot catch tail\n",
1198 static int get_any_page(struct page *p, unsigned long flags)
1200 int ret = 0, pass = 0;
1201 bool count_increased = false;
1203 if (flags & MF_COUNT_INCREASED)
1204 count_increased = true;
1207 if (!count_increased) {
1208 ret = __get_hwpoison_page(p);
1210 if (page_count(p)) {
1211 /* We raced with an allocation, retry. */
1215 } else if (!PageHuge(p) && !is_free_buddy_page(p)) {
1216 /* We raced with put_page, retry. */
1222 } else if (ret == -EBUSY) {
1224 * We raced with (possibly temporary) unhandlable
1236 if (PageHuge(p) || HWPoisonHandlable(p)) {
1240 * A page we cannot handle. Check whether we can turn
1241 * it into something we can handle.
1246 count_increased = false;
1254 dump_page(p, "hwpoison: unhandlable page");
1260 * get_hwpoison_page() - Get refcount for memory error handling
1261 * @p: Raw error page (hit by memory error)
1262 * @flags: Flags controlling behavior of error handling
1264 * get_hwpoison_page() takes a page refcount of an error page to handle memory
1265 * error on it, after checking that the error page is in a well-defined state
1266 * (defined as a page-type we can successfully handle the memor error on it,
1267 * such as LRU page and hugetlb page).
1269 * Memory error handling could be triggered at any time on any type of page,
1270 * so it's prone to race with typical memory management lifecycle (like
1271 * allocation and free). So to avoid such races, get_hwpoison_page() takes
1272 * extra care for the error page's state (as done in __get_hwpoison_page()),
1273 * and has some retry logic in get_any_page().
1275 * Return: 0 on failure,
1276 * 1 on success for in-use pages in a well-defined state,
1277 * -EIO for pages on which we can not handle memory errors,
1278 * -EBUSY when get_hwpoison_page() has raced with page lifecycle
1279 * operations like allocation and free.
1281 static int get_hwpoison_page(struct page *p, unsigned long flags)
1285 zone_pcp_disable(page_zone(p));
1286 ret = get_any_page(p, flags);
1287 zone_pcp_enable(page_zone(p));
1293 * Do all that is necessary to remove user space mappings. Unmap
1294 * the pages and send SIGBUS to the processes if the data was dirty.
1296 static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
1297 int flags, struct page *hpage)
1299 enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC;
1300 struct address_space *mapping;
1303 int kill = 1, forcekill;
1304 bool mlocked = PageMlocked(hpage);
1307 * Here we are interested only in user-mapped pages, so skip any
1308 * other types of pages.
1310 if (PageReserved(p) || PageSlab(p))
1312 if (!(PageLRU(hpage) || PageHuge(p)))
1316 * This check implies we don't kill processes if their pages
1317 * are in the swap cache early. Those are always late kills.
1319 if (!page_mapped(hpage))
1323 pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn);
1327 if (PageSwapCache(p)) {
1328 pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n",
1330 ttu |= TTU_IGNORE_HWPOISON;
1334 * Propagate the dirty bit from PTEs to struct page first, because we
1335 * need this to decide if we should kill or just drop the page.
1336 * XXX: the dirty test could be racy: set_page_dirty() may not always
1337 * be called inside page lock (it's recommended but not enforced).
1339 mapping = page_mapping(hpage);
1340 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
1341 mapping_can_writeback(mapping)) {
1342 if (page_mkclean(hpage)) {
1343 SetPageDirty(hpage);
1346 ttu |= TTU_IGNORE_HWPOISON;
1347 pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n",
1353 * First collect all the processes that have the page
1354 * mapped in dirty form. This has to be done before try_to_unmap,
1355 * because ttu takes the rmap data structures down.
1357 * Error handling: We ignore errors here because
1358 * there's nothing that can be done.
1361 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
1363 if (!PageHuge(hpage)) {
1364 try_to_unmap(hpage, ttu);
1366 if (!PageAnon(hpage)) {
1368 * For hugetlb pages in shared mappings, try_to_unmap
1369 * could potentially call huge_pmd_unshare. Because of
1370 * this, take semaphore in write mode here and set
1371 * TTU_RMAP_LOCKED to indicate we have taken the lock
1372 * at this higher level.
1374 mapping = hugetlb_page_mapping_lock_write(hpage);
1376 try_to_unmap(hpage, ttu|TTU_RMAP_LOCKED);
1377 i_mmap_unlock_write(mapping);
1379 pr_info("Memory failure: %#lx: could not lock mapping for mapped huge page\n", pfn);
1381 try_to_unmap(hpage, ttu);
1385 unmap_success = !page_mapped(hpage);
1387 pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n",
1388 pfn, page_mapcount(hpage));
1391 * try_to_unmap() might put mlocked page in lru cache, so call
1392 * shake_page() again to ensure that it's flushed.
1398 * Now that the dirty bit has been propagated to the
1399 * struct page and all unmaps done we can decide if
1400 * killing is needed or not. Only kill when the page
1401 * was dirty or the process is not restartable,
1402 * otherwise the tokill list is merely
1403 * freed. When there was a problem unmapping earlier
1404 * use a more force-full uncatchable kill to prevent
1405 * any accesses to the poisoned memory.
1407 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
1408 kill_procs(&tokill, forcekill, !unmap_success, pfn, flags);
1410 return unmap_success;
1413 static int identify_page_state(unsigned long pfn, struct page *p,
1414 unsigned long page_flags)
1416 struct page_state *ps;
1419 * The first check uses the current page flags which may not have any
1420 * relevant information. The second check with the saved page flags is
1421 * carried out only if the first check can't determine the page status.
1423 for (ps = error_states;; ps++)
1424 if ((p->flags & ps->mask) == ps->res)
1427 page_flags |= (p->flags & (1UL << PG_dirty));
1430 for (ps = error_states;; ps++)
1431 if ((page_flags & ps->mask) == ps->res)
1433 return page_action(ps, p, pfn);
1436 static int try_to_split_thp_page(struct page *page, const char *msg)
1439 if (unlikely(split_huge_page(page))) {
1440 unsigned long pfn = page_to_pfn(page);
1443 pr_info("%s: %#lx: thp split failed\n", msg, pfn);
1452 static int memory_failure_hugetlb(unsigned long pfn, int flags)
1454 struct page *p = pfn_to_page(pfn);
1455 struct page *head = compound_head(p);
1457 unsigned long page_flags;
1459 if (TestSetPageHWPoison(head)) {
1460 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1463 if (flags & MF_ACTION_REQUIRED)
1464 res = kill_accessing_process(current, page_to_pfn(head), flags);
1468 num_poisoned_pages_inc();
1470 if (!(flags & MF_COUNT_INCREASED)) {
1471 res = get_hwpoison_page(p, flags);
1474 * Check "filter hit" and "race with other subpage."
1477 if (PageHWPoison(head)) {
1478 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1479 || (p != head && TestSetPageHWPoison(head))) {
1480 num_poisoned_pages_dec();
1487 if (__page_handle_poison(p)) {
1491 action_result(pfn, MF_MSG_FREE_HUGE, res);
1492 return res == MF_RECOVERED ? 0 : -EBUSY;
1493 } else if (res < 0) {
1494 action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
1500 page_flags = head->flags;
1502 if (!PageHWPoison(head)) {
1503 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1504 num_poisoned_pages_dec();
1511 * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so
1512 * simply disable it. In order to make it work properly, we need
1514 * - conversion of a pud that maps an error hugetlb into hwpoison
1515 * entry properly works, and
1516 * - other mm code walking over page table is aware of pud-aligned
1519 if (huge_page_size(page_hstate(head)) > PMD_SIZE) {
1520 action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED);
1525 if (!hwpoison_user_mappings(p, pfn, flags, head)) {
1526 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1531 return identify_page_state(pfn, p, page_flags);
1537 static int memory_failure_dev_pagemap(unsigned long pfn, int flags,
1538 struct dev_pagemap *pgmap)
1540 struct page *page = pfn_to_page(pfn);
1541 unsigned long size = 0;
1548 if (flags & MF_COUNT_INCREASED)
1550 * Drop the extra refcount in case we come from madvise().
1554 /* device metadata space is not recoverable */
1555 if (!pgmap_pfn_valid(pgmap, pfn)) {
1561 * Prevent the inode from being freed while we are interrogating
1562 * the address_space, typically this would be handled by
1563 * lock_page(), but dax pages do not use the page lock. This
1564 * also prevents changes to the mapping of this pfn until
1565 * poison signaling is complete.
1567 cookie = dax_lock_page(page);
1571 if (hwpoison_filter(page)) {
1576 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
1578 * TODO: Handle HMM pages which may need coordination
1579 * with device-side memory.
1585 * Use this flag as an indication that the dax page has been
1586 * remapped UC to prevent speculative consumption of poison.
1588 SetPageHWPoison(page);
1591 * Unlike System-RAM there is no possibility to swap in a
1592 * different physical page at a given virtual address, so all
1593 * userspace consumption of ZONE_DEVICE memory necessitates
1594 * SIGBUS (i.e. MF_MUST_KILL)
1596 flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1597 collect_procs(page, &tokill, flags & MF_ACTION_REQUIRED);
1599 list_for_each_entry(tk, &tokill, nd)
1601 size = max(size, 1UL << tk->size_shift);
1604 * Unmap the largest mapping to avoid breaking up
1605 * device-dax mappings which are constant size. The
1606 * actual size of the mapping being torn down is
1607 * communicated in siginfo, see kill_proc()
1609 start = (page->index << PAGE_SHIFT) & ~(size - 1);
1610 unmap_mapping_range(page->mapping, start, size, 0);
1612 kill_procs(&tokill, flags & MF_MUST_KILL, false, pfn, flags);
1615 dax_unlock_page(page, cookie);
1617 /* drop pgmap ref acquired in caller */
1618 put_dev_pagemap(pgmap);
1619 action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED);
1624 * memory_failure - Handle memory failure of a page.
1625 * @pfn: Page Number of the corrupted page
1626 * @flags: fine tune action taken
1628 * This function is called by the low level machine check code
1629 * of an architecture when it detects hardware memory corruption
1630 * of a page. It tries its best to recover, which includes
1631 * dropping pages, killing processes etc.
1633 * The function is primarily of use for corruptions that
1634 * happen outside the current execution context (e.g. when
1635 * detected by a background scrubber)
1637 * Must run in process context (e.g. a work queue) with interrupts
1638 * enabled and no spinlocks hold.
1640 int memory_failure(unsigned long pfn, int flags)
1644 struct page *orig_head;
1645 struct dev_pagemap *pgmap;
1647 unsigned long page_flags;
1649 static DEFINE_MUTEX(mf_mutex);
1651 if (!sysctl_memory_failure_recovery)
1652 panic("Memory failure on page %lx", pfn);
1654 p = pfn_to_online_page(pfn);
1656 if (pfn_valid(pfn)) {
1657 pgmap = get_dev_pagemap(pfn, NULL);
1659 return memory_failure_dev_pagemap(pfn, flags,
1662 pr_err("Memory failure: %#lx: memory outside kernel control\n",
1667 mutex_lock(&mf_mutex);
1671 res = memory_failure_hugetlb(pfn, flags);
1675 if (TestSetPageHWPoison(p)) {
1676 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1679 if (flags & MF_ACTION_REQUIRED)
1680 res = kill_accessing_process(current, pfn, flags);
1684 orig_head = hpage = compound_head(p);
1685 num_poisoned_pages_inc();
1688 * We need/can do nothing about count=0 pages.
1689 * 1) it's a free page, and therefore in safe hand:
1690 * prep_new_page() will be the gate keeper.
1691 * 2) it's part of a non-compound high order page.
1692 * Implies some kernel user: cannot stop them from
1693 * R/W the page; let's pray that the page has been
1694 * used and will be freed some time later.
1695 * In fact it's dangerous to directly bump up page count from 0,
1696 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
1698 if (!(flags & MF_COUNT_INCREASED)) {
1699 res = get_hwpoison_page(p, flags);
1701 if (is_free_buddy_page(p)) {
1702 if (take_page_off_buddy(p)) {
1706 /* We lost the race, try again */
1708 ClearPageHWPoison(p);
1709 num_poisoned_pages_dec();
1715 action_result(pfn, MF_MSG_BUDDY, res);
1716 res = res == MF_RECOVERED ? 0 : -EBUSY;
1718 action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
1722 } else if (res < 0) {
1723 action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
1729 if (PageTransHuge(hpage)) {
1731 * The flag must be set after the refcount is bumped
1732 * otherwise it may race with THP split.
1733 * And the flag can't be set in get_hwpoison_page() since
1734 * it is called by soft offline too and it is just called
1735 * for !MF_COUNT_INCREASE. So here seems to be the best
1738 * Don't need care about the above error handling paths for
1739 * get_hwpoison_page() since they handle either free page
1740 * or unhandlable page. The refcount is bumped iff the
1741 * page is a valid handlable page.
1743 SetPageHasHWPoisoned(hpage);
1744 if (try_to_split_thp_page(p, "Memory Failure") < 0) {
1745 action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED);
1749 VM_BUG_ON_PAGE(!page_count(p), p);
1753 * We ignore non-LRU pages for good reasons.
1754 * - PG_locked is only well defined for LRU pages and a few others
1755 * - to avoid races with __SetPageLocked()
1756 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1757 * The check (unnecessarily) ignores LRU pages being isolated and
1758 * walked by the page reclaim code, however that's not a big loss.
1765 * The page could have changed compound pages during the locking.
1766 * If this happens just bail out.
1768 if (PageCompound(p) && compound_head(p) != orig_head) {
1769 action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
1775 * We use page flags to determine what action should be taken, but
1776 * the flags can be modified by the error containment action. One
1777 * example is an mlocked page, where PG_mlocked is cleared by
1778 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1779 * correctly, we save a copy of the page flags at this time.
1781 page_flags = p->flags;
1784 * unpoison always clear PG_hwpoison inside page lock
1786 if (!PageHWPoison(p)) {
1787 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1788 num_poisoned_pages_dec();
1793 if (hwpoison_filter(p)) {
1794 if (TestClearPageHWPoison(p))
1795 num_poisoned_pages_dec();
1802 * __munlock_pagevec may clear a writeback page's LRU flag without
1803 * page_lock. We need wait writeback completion for this page or it
1804 * may trigger vfs BUG while evict inode.
1806 if (!PageTransTail(p) && !PageLRU(p) && !PageWriteback(p))
1807 goto identify_page_state;
1810 * It's very difficult to mess with pages currently under IO
1811 * and in many cases impossible, so we just avoid it here.
1813 wait_on_page_writeback(p);
1816 * Now take care of user space mappings.
1817 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1819 if (!hwpoison_user_mappings(p, pfn, flags, p)) {
1820 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1826 * Torn down by someone else?
1828 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1829 action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
1834 identify_page_state:
1835 res = identify_page_state(pfn, p, page_flags);
1836 mutex_unlock(&mf_mutex);
1841 mutex_unlock(&mf_mutex);
1844 EXPORT_SYMBOL_GPL(memory_failure);
1846 #define MEMORY_FAILURE_FIFO_ORDER 4
1847 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1849 struct memory_failure_entry {
1854 struct memory_failure_cpu {
1855 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1856 MEMORY_FAILURE_FIFO_SIZE);
1858 struct work_struct work;
1861 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1864 * memory_failure_queue - Schedule handling memory failure of a page.
1865 * @pfn: Page Number of the corrupted page
1866 * @flags: Flags for memory failure handling
1868 * This function is called by the low level hardware error handler
1869 * when it detects hardware memory corruption of a page. It schedules
1870 * the recovering of error page, including dropping pages, killing
1873 * The function is primarily of use for corruptions that
1874 * happen outside the current execution context (e.g. when
1875 * detected by a background scrubber)
1877 * Can run in IRQ context.
1879 void memory_failure_queue(unsigned long pfn, int flags)
1881 struct memory_failure_cpu *mf_cpu;
1882 unsigned long proc_flags;
1883 struct memory_failure_entry entry = {
1888 mf_cpu = &get_cpu_var(memory_failure_cpu);
1889 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1890 if (kfifo_put(&mf_cpu->fifo, entry))
1891 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1893 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1895 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1896 put_cpu_var(memory_failure_cpu);
1898 EXPORT_SYMBOL_GPL(memory_failure_queue);
1900 static void memory_failure_work_func(struct work_struct *work)
1902 struct memory_failure_cpu *mf_cpu;
1903 struct memory_failure_entry entry = { 0, };
1904 unsigned long proc_flags;
1907 mf_cpu = container_of(work, struct memory_failure_cpu, work);
1909 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1910 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1911 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1914 if (entry.flags & MF_SOFT_OFFLINE)
1915 soft_offline_page(entry.pfn, entry.flags);
1917 memory_failure(entry.pfn, entry.flags);
1922 * Process memory_failure work queued on the specified CPU.
1923 * Used to avoid return-to-userspace racing with the memory_failure workqueue.
1925 void memory_failure_queue_kick(int cpu)
1927 struct memory_failure_cpu *mf_cpu;
1929 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1930 cancel_work_sync(&mf_cpu->work);
1931 memory_failure_work_func(&mf_cpu->work);
1934 static int __init memory_failure_init(void)
1936 struct memory_failure_cpu *mf_cpu;
1939 for_each_possible_cpu(cpu) {
1940 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1941 spin_lock_init(&mf_cpu->lock);
1942 INIT_KFIFO(mf_cpu->fifo);
1943 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1948 core_initcall(memory_failure_init);
1950 #define unpoison_pr_info(fmt, pfn, rs) \
1952 if (__ratelimit(rs)) \
1953 pr_info(fmt, pfn); \
1957 * unpoison_memory - Unpoison a previously poisoned page
1958 * @pfn: Page number of the to be unpoisoned page
1960 * Software-unpoison a page that has been poisoned by
1961 * memory_failure() earlier.
1963 * This is only done on the software-level, so it only works
1964 * for linux injected failures, not real hardware failures
1966 * Returns 0 for success, otherwise -errno.
1968 int unpoison_memory(unsigned long pfn)
1973 unsigned long flags = 0;
1974 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
1975 DEFAULT_RATELIMIT_BURST);
1977 if (!pfn_valid(pfn))
1980 p = pfn_to_page(pfn);
1981 page = compound_head(p);
1983 if (!PageHWPoison(p)) {
1984 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
1989 if (page_count(page) > 1) {
1990 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
1995 if (page_mapped(page)) {
1996 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
2001 if (page_mapping(page)) {
2002 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
2008 * unpoison_memory() can encounter thp only when the thp is being
2009 * worked by memory_failure() and the page lock is not held yet.
2010 * In such case, we yield to memory_failure() and make unpoison fail.
2012 if (!PageHuge(page) && PageTransHuge(page)) {
2013 unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n",
2018 if (!get_hwpoison_page(p, flags)) {
2019 if (TestClearPageHWPoison(p))
2020 num_poisoned_pages_dec();
2021 unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n",
2028 * This test is racy because PG_hwpoison is set outside of page lock.
2029 * That's acceptable because that won't trigger kernel panic. Instead,
2030 * the PG_hwpoison page will be caught and isolated on the entrance to
2031 * the free buddy page pool.
2033 if (TestClearPageHWPoison(page)) {
2034 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
2036 num_poisoned_pages_dec();
2042 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
2047 EXPORT_SYMBOL(unpoison_memory);
2049 static bool isolate_page(struct page *page, struct list_head *pagelist)
2051 bool isolated = false;
2052 bool lru = PageLRU(page);
2054 if (PageHuge(page)) {
2055 isolated = isolate_huge_page(page, pagelist);
2058 isolated = !isolate_lru_page(page);
2060 isolated = !isolate_movable_page(page, ISOLATE_UNEVICTABLE);
2063 list_add(&page->lru, pagelist);
2066 if (isolated && lru)
2067 inc_node_page_state(page, NR_ISOLATED_ANON +
2068 page_is_file_lru(page));
2071 * If we succeed to isolate the page, we grabbed another refcount on
2072 * the page, so we can safely drop the one we got from get_any_pages().
2073 * If we failed to isolate the page, it means that we cannot go further
2074 * and we will return an error, so drop the reference we got from
2075 * get_any_pages() as well.
2082 * __soft_offline_page handles hugetlb-pages and non-hugetlb pages.
2083 * If the page is a non-dirty unmapped page-cache page, it simply invalidates.
2084 * If the page is mapped, it migrates the contents over.
2086 static int __soft_offline_page(struct page *page)
2089 unsigned long pfn = page_to_pfn(page);
2090 struct page *hpage = compound_head(page);
2091 char const *msg_page[] = {"page", "hugepage"};
2092 bool huge = PageHuge(page);
2093 LIST_HEAD(pagelist);
2094 struct migration_target_control mtc = {
2095 .nid = NUMA_NO_NODE,
2096 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
2100 * Check PageHWPoison again inside page lock because PageHWPoison
2101 * is set by memory_failure() outside page lock. Note that
2102 * memory_failure() also double-checks PageHWPoison inside page lock,
2103 * so there's no race between soft_offline_page() and memory_failure().
2106 if (!PageHuge(page))
2107 wait_on_page_writeback(page);
2108 if (PageHWPoison(page)) {
2111 pr_info("soft offline: %#lx page already poisoned\n", pfn);
2115 if (!PageHuge(page))
2117 * Try to invalidate first. This should work for
2118 * non dirty unmapped page cache pages.
2120 ret = invalidate_inode_page(page);
2124 * RED-PEN would be better to keep it isolated here, but we
2125 * would need to fix isolation locking first.
2128 pr_info("soft_offline: %#lx: invalidated\n", pfn);
2129 page_handle_poison(page, false, true);
2133 if (isolate_page(hpage, &pagelist)) {
2134 ret = migrate_pages(&pagelist, alloc_migration_target, NULL,
2135 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE, NULL);
2137 bool release = !huge;
2139 if (!page_handle_poison(page, huge, release))
2142 if (!list_empty(&pagelist))
2143 putback_movable_pages(&pagelist);
2145 pr_info("soft offline: %#lx: %s migration failed %d, type %pGp\n",
2146 pfn, msg_page[huge], ret, &page->flags);
2151 pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %pGp\n",
2152 pfn, msg_page[huge], page_count(page), &page->flags);
2158 static int soft_offline_in_use_page(struct page *page)
2160 struct page *hpage = compound_head(page);
2162 if (!PageHuge(page) && PageTransHuge(hpage))
2163 if (try_to_split_thp_page(page, "soft offline") < 0)
2165 return __soft_offline_page(page);
2168 static int soft_offline_free_page(struct page *page)
2172 if (!page_handle_poison(page, true, false))
2178 static void put_ref_page(struct page *page)
2185 * soft_offline_page - Soft offline a page.
2186 * @pfn: pfn to soft-offline
2187 * @flags: flags. Same as memory_failure().
2189 * Returns 0 on success, otherwise negated errno.
2191 * Soft offline a page, by migration or invalidation,
2192 * without killing anything. This is for the case when
2193 * a page is not corrupted yet (so it's still valid to access),
2194 * but has had a number of corrected errors and is better taken
2197 * The actual policy on when to do that is maintained by
2200 * This should never impact any application or cause data loss,
2201 * however it might take some time.
2203 * This is not a 100% solution for all memory, but tries to be
2204 * ``good enough'' for the majority of memory.
2206 int soft_offline_page(unsigned long pfn, int flags)
2209 bool try_again = true;
2210 struct page *page, *ref_page = NULL;
2212 WARN_ON_ONCE(!pfn_valid(pfn) && (flags & MF_COUNT_INCREASED));
2214 if (!pfn_valid(pfn))
2216 if (flags & MF_COUNT_INCREASED)
2217 ref_page = pfn_to_page(pfn);
2219 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
2220 page = pfn_to_online_page(pfn);
2222 put_ref_page(ref_page);
2226 if (PageHWPoison(page)) {
2227 pr_info("%s: %#lx page already poisoned\n", __func__, pfn);
2228 put_ref_page(ref_page);
2234 ret = get_hwpoison_page(page, flags);
2238 ret = soft_offline_in_use_page(page);
2239 } else if (ret == 0) {
2240 if (soft_offline_free_page(page) && try_again) {