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/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 <linux/pagewalk.h>
61 #include "ras/ras_event.h"
63 int sysctl_memory_failure_early_kill __read_mostly = 0;
65 int sysctl_memory_failure_recovery __read_mostly = 1;
67 atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
69 static bool __page_handle_poison(struct page *page)
73 zone_pcp_disable(page_zone(page));
74 ret = dissolve_free_huge_page(page);
76 ret = take_page_off_buddy(page);
77 zone_pcp_enable(page_zone(page));
82 static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release)
84 if (hugepage_or_freepage) {
86 * Doing this check for free pages is also fine since dissolve_free_huge_page
87 * returns 0 for non-hugetlb pages as well.
89 if (!__page_handle_poison(page))
91 * We could fail to take off the target page from buddy
92 * for example due to racy page allocation, but that's
93 * acceptable because soft-offlined page is not broken
94 * and if someone really want to use it, they should
100 SetPageHWPoison(page);
104 num_poisoned_pages_inc();
109 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
111 u32 hwpoison_filter_enable = 0;
112 u32 hwpoison_filter_dev_major = ~0U;
113 u32 hwpoison_filter_dev_minor = ~0U;
114 u64 hwpoison_filter_flags_mask;
115 u64 hwpoison_filter_flags_value;
116 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
117 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
118 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
119 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
120 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
122 static int hwpoison_filter_dev(struct page *p)
124 struct address_space *mapping;
127 if (hwpoison_filter_dev_major == ~0U &&
128 hwpoison_filter_dev_minor == ~0U)
132 * page_mapping() does not accept slab pages.
137 mapping = page_mapping(p);
138 if (mapping == NULL || mapping->host == NULL)
141 dev = mapping->host->i_sb->s_dev;
142 if (hwpoison_filter_dev_major != ~0U &&
143 hwpoison_filter_dev_major != MAJOR(dev))
145 if (hwpoison_filter_dev_minor != ~0U &&
146 hwpoison_filter_dev_minor != MINOR(dev))
152 static int hwpoison_filter_flags(struct page *p)
154 if (!hwpoison_filter_flags_mask)
157 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
158 hwpoison_filter_flags_value)
165 * This allows stress tests to limit test scope to a collection of tasks
166 * by putting them under some memcg. This prevents killing unrelated/important
167 * processes such as /sbin/init. Note that the target task may share clean
168 * pages with init (eg. libc text), which is harmless. If the target task
169 * share _dirty_ pages with another task B, the test scheme must make sure B
170 * is also included in the memcg. At last, due to race conditions this filter
171 * can only guarantee that the page either belongs to the memcg tasks, or is
175 u64 hwpoison_filter_memcg;
176 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
177 static int hwpoison_filter_task(struct page *p)
179 if (!hwpoison_filter_memcg)
182 if (page_cgroup_ino(p) != hwpoison_filter_memcg)
188 static int hwpoison_filter_task(struct page *p) { return 0; }
191 int hwpoison_filter(struct page *p)
193 if (!hwpoison_filter_enable)
196 if (hwpoison_filter_dev(p))
199 if (hwpoison_filter_flags(p))
202 if (hwpoison_filter_task(p))
208 int hwpoison_filter(struct page *p)
214 EXPORT_SYMBOL_GPL(hwpoison_filter);
217 * Kill all processes that have a poisoned page mapped and then isolate
221 * Find all processes having the page mapped and kill them.
222 * But we keep a page reference around so that the page is not
223 * actually freed yet.
224 * Then stash the page away
226 * There's no convenient way to get back to mapped processes
227 * from the VMAs. So do a brute-force search over all
230 * Remember that machine checks are not common (or rather
231 * if they are common you have other problems), so this shouldn't
232 * be a performance issue.
234 * Also there are some races possible while we get from the
235 * error detection to actually handle it.
240 struct task_struct *tsk;
246 * Send all the processes who have the page mapped a signal.
247 * ``action optional'' if they are not immediately affected by the error
248 * ``action required'' if error happened in current execution context
250 static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
252 struct task_struct *t = tk->tsk;
253 short addr_lsb = tk->size_shift;
256 pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
257 pfn, t->comm, t->pid);
259 if (flags & MF_ACTION_REQUIRED) {
261 ret = force_sig_mceerr(BUS_MCEERR_AR,
262 (void __user *)tk->addr, addr_lsb);
264 /* Signal other processes sharing the page if they have PF_MCE_EARLY set. */
265 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
269 * Don't use force here, it's convenient if the signal
270 * can be temporarily blocked.
271 * This could cause a loop when the user sets SIGBUS
272 * to SIG_IGN, but hopefully no one will do that?
274 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
275 addr_lsb, t); /* synchronous? */
278 pr_info("Memory failure: Error sending signal to %s:%d: %d\n",
279 t->comm, t->pid, ret);
284 * Unknown page type encountered. Try to check whether it can turn PageLRU by
287 void shake_page(struct page *p)
294 if (PageLRU(p) || is_free_buddy_page(p))
299 * TODO: Could shrink slab caches here if a lightweight range-based
300 * shrinker will be available.
303 EXPORT_SYMBOL_GPL(shake_page);
305 static unsigned long dev_pagemap_mapping_shift(struct page *page,
306 struct vm_area_struct *vma)
308 unsigned long address = vma_address(page, vma);
309 unsigned long ret = 0;
316 pgd = pgd_offset(vma->vm_mm, address);
317 if (!pgd_present(*pgd))
319 p4d = p4d_offset(pgd, address);
320 if (!p4d_present(*p4d))
322 pud = pud_offset(p4d, address);
323 if (!pud_present(*pud))
325 if (pud_devmap(*pud))
327 pmd = pmd_offset(pud, address);
328 if (!pmd_present(*pmd))
330 if (pmd_devmap(*pmd))
332 pte = pte_offset_map(pmd, address);
333 if (pte_present(*pte) && pte_devmap(*pte))
340 * Failure handling: if we can't find or can't kill a process there's
341 * not much we can do. We just print a message and ignore otherwise.
345 * Schedule a process for later kill.
346 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
348 static void add_to_kill(struct task_struct *tsk, struct page *p,
349 struct vm_area_struct *vma,
350 struct list_head *to_kill)
354 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
356 pr_err("Memory failure: Out of memory while machine check handling\n");
360 tk->addr = page_address_in_vma(p, vma);
361 if (is_zone_device_page(p))
362 tk->size_shift = dev_pagemap_mapping_shift(p, vma);
364 tk->size_shift = page_shift(compound_head(p));
367 * Send SIGKILL if "tk->addr == -EFAULT". Also, as
368 * "tk->size_shift" is always non-zero for !is_zone_device_page(),
369 * so "tk->size_shift == 0" effectively checks no mapping on
370 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times
371 * to a process' address space, it's possible not all N VMAs
372 * contain mappings for the page, but at least one VMA does.
373 * Only deliver SIGBUS with payload derived from the VMA that
374 * has a mapping for the page.
376 if (tk->addr == -EFAULT) {
377 pr_info("Memory failure: Unable to find user space address %lx in %s\n",
378 page_to_pfn(p), tsk->comm);
379 } else if (tk->size_shift == 0) {
384 get_task_struct(tsk);
386 list_add_tail(&tk->nd, to_kill);
390 * Kill the processes that have been collected earlier.
392 * Only do anything when FORCEKILL is set, otherwise just free the
393 * list (this is used for clean pages which do not need killing)
394 * Also when FAIL is set do a force kill because something went
397 static void kill_procs(struct list_head *to_kill, int forcekill, bool fail,
398 unsigned long pfn, int flags)
400 struct to_kill *tk, *next;
402 list_for_each_entry_safe (tk, next, to_kill, nd) {
405 * In case something went wrong with munmapping
406 * make sure the process doesn't catch the
407 * signal and then access the memory. Just kill it.
409 if (fail || tk->addr == -EFAULT) {
410 pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
411 pfn, tk->tsk->comm, tk->tsk->pid);
412 do_send_sig_info(SIGKILL, SEND_SIG_PRIV,
413 tk->tsk, PIDTYPE_PID);
417 * In theory the process could have mapped
418 * something else on the address in-between. We could
419 * check for that, but we need to tell the
422 else if (kill_proc(tk, pfn, flags) < 0)
423 pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n",
424 pfn, tk->tsk->comm, tk->tsk->pid);
426 put_task_struct(tk->tsk);
432 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
433 * on behalf of the thread group. Return task_struct of the (first found)
434 * dedicated thread if found, and return NULL otherwise.
436 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
437 * have to call rcu_read_lock/unlock() in this function.
439 static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
441 struct task_struct *t;
443 for_each_thread(tsk, t) {
444 if (t->flags & PF_MCE_PROCESS) {
445 if (t->flags & PF_MCE_EARLY)
448 if (sysctl_memory_failure_early_kill)
456 * Determine whether a given process is "early kill" process which expects
457 * to be signaled when some page under the process is hwpoisoned.
458 * Return task_struct of the dedicated thread (main thread unless explicitly
459 * specified) if the process is "early kill" and otherwise returns NULL.
461 * Note that the above is true for Action Optional case. For Action Required
462 * case, it's only meaningful to the current thread which need to be signaled
463 * with SIGBUS, this error is Action Optional for other non current
464 * processes sharing the same error page,if the process is "early kill", the
465 * task_struct of the dedicated thread will also be returned.
467 static struct task_struct *task_early_kill(struct task_struct *tsk,
473 * Comparing ->mm here because current task might represent
474 * a subthread, while tsk always points to the main thread.
476 if (force_early && tsk->mm == current->mm)
479 return find_early_kill_thread(tsk);
483 * Collect processes when the error hit an anonymous page.
485 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
488 struct vm_area_struct *vma;
489 struct task_struct *tsk;
493 av = page_lock_anon_vma_read(page);
494 if (av == NULL) /* Not actually mapped anymore */
497 pgoff = page_to_pgoff(page);
498 read_lock(&tasklist_lock);
499 for_each_process (tsk) {
500 struct anon_vma_chain *vmac;
501 struct task_struct *t = task_early_kill(tsk, force_early);
505 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
508 if (!page_mapped_in_vma(page, vma))
510 if (vma->vm_mm == t->mm)
511 add_to_kill(t, page, vma, to_kill);
514 read_unlock(&tasklist_lock);
515 page_unlock_anon_vma_read(av);
519 * Collect processes when the error hit a file mapped page.
521 static void collect_procs_file(struct page *page, struct list_head *to_kill,
524 struct vm_area_struct *vma;
525 struct task_struct *tsk;
526 struct address_space *mapping = page->mapping;
529 i_mmap_lock_read(mapping);
530 read_lock(&tasklist_lock);
531 pgoff = page_to_pgoff(page);
532 for_each_process(tsk) {
533 struct task_struct *t = task_early_kill(tsk, force_early);
537 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
540 * Send early kill signal to tasks where a vma covers
541 * the page but the corrupted page is not necessarily
542 * mapped it in its pte.
543 * Assume applications who requested early kill want
544 * to be informed of all such data corruptions.
546 if (vma->vm_mm == t->mm)
547 add_to_kill(t, page, vma, to_kill);
550 read_unlock(&tasklist_lock);
551 i_mmap_unlock_read(mapping);
555 * Collect the processes who have the corrupted page mapped to kill.
557 static void collect_procs(struct page *page, struct list_head *tokill,
564 collect_procs_anon(page, tokill, force_early);
566 collect_procs_file(page, tokill, force_early);
575 static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift)
578 tk->size_shift = shift;
581 static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift,
582 unsigned long poisoned_pfn, struct to_kill *tk)
584 unsigned long pfn = 0;
586 if (pte_present(pte)) {
589 swp_entry_t swp = pte_to_swp_entry(pte);
591 if (is_hwpoison_entry(swp))
592 pfn = hwpoison_entry_to_pfn(swp);
595 if (!pfn || pfn != poisoned_pfn)
598 set_to_kill(tk, addr, shift);
602 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
603 static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
604 struct hwp_walk *hwp)
608 unsigned long hwpoison_vaddr;
610 if (!pmd_present(pmd))
613 if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) {
614 hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT);
615 set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT);
621 static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
622 struct hwp_walk *hwp)
628 static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr,
629 unsigned long end, struct mm_walk *walk)
631 struct hwp_walk *hwp = (struct hwp_walk *)walk->private;
633 pte_t *ptep, *mapped_pte;
636 ptl = pmd_trans_huge_lock(pmdp, walk->vma);
638 ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp);
643 if (pmd_trans_unstable(pmdp))
646 mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp,
648 for (; addr != end; ptep++, addr += PAGE_SIZE) {
649 ret = check_hwpoisoned_entry(*ptep, addr, PAGE_SHIFT,
654 pte_unmap_unlock(mapped_pte, ptl);
660 #ifdef CONFIG_HUGETLB_PAGE
661 static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask,
662 unsigned long addr, unsigned long end,
663 struct mm_walk *walk)
665 struct hwp_walk *hwp = (struct hwp_walk *)walk->private;
666 pte_t pte = huge_ptep_get(ptep);
667 struct hstate *h = hstate_vma(walk->vma);
669 return check_hwpoisoned_entry(pte, addr, huge_page_shift(h),
673 #define hwpoison_hugetlb_range NULL
676 static struct mm_walk_ops hwp_walk_ops = {
677 .pmd_entry = hwpoison_pte_range,
678 .hugetlb_entry = hwpoison_hugetlb_range,
682 * Sends SIGBUS to the current process with error info.
684 * This function is intended to handle "Action Required" MCEs on already
685 * hardware poisoned pages. They could happen, for example, when
686 * memory_failure() failed to unmap the error page at the first call, or
687 * when multiple local machine checks happened on different CPUs.
689 * MCE handler currently has no easy access to the error virtual address,
690 * so this function walks page table to find it. The returned virtual address
691 * is proper in most cases, but it could be wrong when the application
692 * process has multiple entries mapping the error page.
694 static int kill_accessing_process(struct task_struct *p, unsigned long pfn,
698 struct hwp_walk priv = {
706 mmap_read_lock(p->mm);
707 ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwp_walk_ops,
709 if (ret == 1 && priv.tk.addr)
710 kill_proc(&priv.tk, pfn, flags);
713 mmap_read_unlock(p->mm);
714 return ret > 0 ? -EHWPOISON : -EFAULT;
717 static const char *action_name[] = {
718 [MF_IGNORED] = "Ignored",
719 [MF_FAILED] = "Failed",
720 [MF_DELAYED] = "Delayed",
721 [MF_RECOVERED] = "Recovered",
724 static const char * const action_page_types[] = {
725 [MF_MSG_KERNEL] = "reserved kernel page",
726 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
727 [MF_MSG_SLAB] = "kernel slab page",
728 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
729 [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned",
730 [MF_MSG_HUGE] = "huge page",
731 [MF_MSG_FREE_HUGE] = "free huge page",
732 [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page",
733 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
734 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
735 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
736 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
737 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
738 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
739 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
740 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
741 [MF_MSG_CLEAN_LRU] = "clean LRU page",
742 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
743 [MF_MSG_BUDDY] = "free buddy page",
744 [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)",
745 [MF_MSG_DAX] = "dax page",
746 [MF_MSG_UNSPLIT_THP] = "unsplit thp",
747 [MF_MSG_UNKNOWN] = "unknown page",
751 * XXX: It is possible that a page is isolated from LRU cache,
752 * and then kept in swap cache or failed to remove from page cache.
753 * The page count will stop it from being freed by unpoison.
754 * Stress tests should be aware of this memory leak problem.
756 static int delete_from_lru_cache(struct page *p)
758 if (!isolate_lru_page(p)) {
760 * Clear sensible page flags, so that the buddy system won't
761 * complain when the page is unpoison-and-freed.
764 ClearPageUnevictable(p);
767 * Poisoned page might never drop its ref count to 0 so we have
768 * to uncharge it manually from its memcg.
770 mem_cgroup_uncharge(p);
773 * drop the page count elevated by isolate_lru_page()
781 static int truncate_error_page(struct page *p, unsigned long pfn,
782 struct address_space *mapping)
786 if (mapping->a_ops->error_remove_page) {
787 int err = mapping->a_ops->error_remove_page(mapping, p);
790 pr_info("Memory failure: %#lx: Failed to punch page: %d\n",
792 } else if (page_has_private(p) &&
793 !try_to_release_page(p, GFP_NOIO)) {
794 pr_info("Memory failure: %#lx: failed to release buffers\n",
801 * If the file system doesn't support it just invalidate
802 * This fails on dirty or anything with private pages
804 if (invalidate_inode_page(p))
807 pr_info("Memory failure: %#lx: Failed to invalidate\n",
815 * Error hit kernel page.
816 * Do nothing, try to be lucky and not touch this instead. For a few cases we
817 * could be more sophisticated.
819 static int me_kernel(struct page *p, unsigned long pfn)
826 * Page in unknown state. Do nothing.
828 static int me_unknown(struct page *p, unsigned long pfn)
830 pr_err("Memory failure: %#lx: Unknown page state\n", pfn);
836 * Clean (or cleaned) page cache page.
838 static int me_pagecache_clean(struct page *p, unsigned long pfn)
841 struct address_space *mapping;
843 delete_from_lru_cache(p);
846 * For anonymous pages we're done the only reference left
847 * should be the one m_f() holds.
855 * Now truncate the page in the page cache. This is really
856 * more like a "temporary hole punch"
857 * Don't do this for block devices when someone else
858 * has a reference, because it could be file system metadata
859 * and that's not safe to truncate.
861 mapping = page_mapping(p);
864 * Page has been teared down in the meanwhile
871 * Truncation is a bit tricky. Enable it per file system for now.
873 * Open: to take i_rwsem or not for this? Right now we don't.
875 ret = truncate_error_page(p, pfn, mapping);
882 * Dirty pagecache page
883 * Issues: when the error hit a hole page the error is not properly
886 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
888 struct address_space *mapping = page_mapping(p);
891 /* TBD: print more information about the file. */
894 * IO error will be reported by write(), fsync(), etc.
895 * who check the mapping.
896 * This way the application knows that something went
897 * wrong with its dirty file data.
899 * There's one open issue:
901 * The EIO will be only reported on the next IO
902 * operation and then cleared through the IO map.
903 * Normally Linux has two mechanisms to pass IO error
904 * first through the AS_EIO flag in the address space
905 * and then through the PageError flag in the page.
906 * Since we drop pages on memory failure handling the
907 * only mechanism open to use is through AS_AIO.
909 * This has the disadvantage that it gets cleared on
910 * the first operation that returns an error, while
911 * the PageError bit is more sticky and only cleared
912 * when the page is reread or dropped. If an
913 * application assumes it will always get error on
914 * fsync, but does other operations on the fd before
915 * and the page is dropped between then the error
916 * will not be properly reported.
918 * This can already happen even without hwpoisoned
919 * pages: first on metadata IO errors (which only
920 * report through AS_EIO) or when the page is dropped
923 * So right now we assume that the application DTRT on
924 * the first EIO, but we're not worse than other parts
927 mapping_set_error(mapping, -EIO);
930 return me_pagecache_clean(p, pfn);
934 * Clean and dirty swap cache.
936 * Dirty swap cache page is tricky to handle. The page could live both in page
937 * cache and swap cache(ie. page is freshly swapped in). So it could be
938 * referenced concurrently by 2 types of PTEs:
939 * normal PTEs and swap PTEs. We try to handle them consistently by calling
940 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
942 * - clear dirty bit to prevent IO
944 * - but keep in the swap cache, so that when we return to it on
945 * a later page fault, we know the application is accessing
946 * corrupted data and shall be killed (we installed simple
947 * interception code in do_swap_page to catch it).
949 * Clean swap cache pages can be directly isolated. A later page fault will
950 * bring in the known good data from disk.
952 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
957 /* Trigger EIO in shmem: */
958 ClearPageUptodate(p);
960 ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED;
965 static int me_swapcache_clean(struct page *p, unsigned long pfn)
969 delete_from_swap_cache(p);
971 ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED;
977 * Huge pages. Needs work.
979 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
980 * To narrow down kill region to one page, we need to break up pmd.
982 static int me_huge_page(struct page *p, unsigned long pfn)
985 struct page *hpage = compound_head(p);
986 struct address_space *mapping;
988 if (!PageHuge(hpage))
991 mapping = page_mapping(hpage);
993 res = truncate_error_page(hpage, pfn, mapping);
999 * migration entry prevents later access on error anonymous
1000 * hugepage, so we can free and dissolve it into buddy to
1001 * save healthy subpages.
1003 if (PageAnon(hpage))
1005 if (__page_handle_poison(p)) {
1015 * Various page states we can handle.
1017 * A page state is defined by its current page->flags bits.
1018 * The table matches them in order and calls the right handler.
1020 * This is quite tricky because we can access page at any time
1021 * in its live cycle, so all accesses have to be extremely careful.
1023 * This is not complete. More states could be added.
1024 * For any missing state don't attempt recovery.
1027 #define dirty (1UL << PG_dirty)
1028 #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked))
1029 #define unevict (1UL << PG_unevictable)
1030 #define mlock (1UL << PG_mlocked)
1031 #define lru (1UL << PG_lru)
1032 #define head (1UL << PG_head)
1033 #define slab (1UL << PG_slab)
1034 #define reserved (1UL << PG_reserved)
1036 static struct page_state {
1039 enum mf_action_page_type type;
1041 /* Callback ->action() has to unlock the relevant page inside it. */
1042 int (*action)(struct page *p, unsigned long pfn);
1043 } error_states[] = {
1044 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
1046 * free pages are specially detected outside this table:
1047 * PG_buddy pages only make a small fraction of all free pages.
1051 * Could in theory check if slab page is free or if we can drop
1052 * currently unused objects without touching them. But just
1053 * treat it as standard kernel for now.
1055 { slab, slab, MF_MSG_SLAB, me_kernel },
1057 { head, head, MF_MSG_HUGE, me_huge_page },
1059 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
1060 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
1062 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
1063 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
1065 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
1066 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
1068 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
1069 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
1072 * Catchall entry: must be at end.
1074 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
1087 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
1088 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
1090 static void action_result(unsigned long pfn, enum mf_action_page_type type,
1091 enum mf_result result)
1093 trace_memory_failure_event(pfn, type, result);
1095 pr_err("Memory failure: %#lx: recovery action for %s: %s\n",
1096 pfn, action_page_types[type], action_name[result]);
1099 static int page_action(struct page_state *ps, struct page *p,
1105 /* page p should be unlocked after returning from ps->action(). */
1106 result = ps->action(p, pfn);
1108 count = page_count(p) - 1;
1109 if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
1112 pr_err("Memory failure: %#lx: %s still referenced by %d users\n",
1113 pfn, action_page_types[ps->type], count);
1116 action_result(pfn, ps->type, result);
1118 /* Could do more checks here if page looks ok */
1120 * Could adjust zone counters here to correct for the missing page.
1123 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
1127 * Return true if a page type of a given page is supported by hwpoison
1128 * mechanism (while handling could fail), otherwise false. This function
1129 * does not return true for hugetlb or device memory pages, so it's assumed
1130 * to be called only in the context where we never have such pages.
1132 static inline bool HWPoisonHandlable(struct page *page)
1134 return PageLRU(page) || __PageMovable(page) || is_free_buddy_page(page);
1137 static int __get_hwpoison_page(struct page *page)
1139 struct page *head = compound_head(page);
1141 bool hugetlb = false;
1143 ret = get_hwpoison_huge_page(head, &hugetlb);
1148 * This check prevents from calling get_hwpoison_unless_zero()
1149 * for any unsupported type of page in order to reduce the risk of
1150 * unexpected races caused by taking a page refcount.
1152 if (!HWPoisonHandlable(head))
1155 if (get_page_unless_zero(head)) {
1156 if (head == compound_head(page))
1159 pr_info("Memory failure: %#lx cannot catch tail\n",
1167 static int get_any_page(struct page *p, unsigned long flags)
1169 int ret = 0, pass = 0;
1170 bool count_increased = false;
1172 if (flags & MF_COUNT_INCREASED)
1173 count_increased = true;
1176 if (!count_increased) {
1177 ret = __get_hwpoison_page(p);
1179 if (page_count(p)) {
1180 /* We raced with an allocation, retry. */
1184 } else if (!PageHuge(p) && !is_free_buddy_page(p)) {
1185 /* We raced with put_page, retry. */
1191 } else if (ret == -EBUSY) {
1193 * We raced with (possibly temporary) unhandlable
1205 if (PageHuge(p) || HWPoisonHandlable(p)) {
1209 * A page we cannot handle. Check whether we can turn
1210 * it into something we can handle.
1215 count_increased = false;
1223 pr_err("Memory failure: %#lx: unhandlable page.\n", page_to_pfn(p));
1229 * get_hwpoison_page() - Get refcount for memory error handling
1230 * @p: Raw error page (hit by memory error)
1231 * @flags: Flags controlling behavior of error handling
1233 * get_hwpoison_page() takes a page refcount of an error page to handle memory
1234 * error on it, after checking that the error page is in a well-defined state
1235 * (defined as a page-type we can successfully handle the memor error on it,
1236 * such as LRU page and hugetlb page).
1238 * Memory error handling could be triggered at any time on any type of page,
1239 * so it's prone to race with typical memory management lifecycle (like
1240 * allocation and free). So to avoid such races, get_hwpoison_page() takes
1241 * extra care for the error page's state (as done in __get_hwpoison_page()),
1242 * and has some retry logic in get_any_page().
1244 * Return: 0 on failure,
1245 * 1 on success for in-use pages in a well-defined state,
1246 * -EIO for pages on which we can not handle memory errors,
1247 * -EBUSY when get_hwpoison_page() has raced with page lifecycle
1248 * operations like allocation and free.
1250 static int get_hwpoison_page(struct page *p, unsigned long flags)
1254 zone_pcp_disable(page_zone(p));
1255 ret = get_any_page(p, flags);
1256 zone_pcp_enable(page_zone(p));
1262 * Do all that is necessary to remove user space mappings. Unmap
1263 * the pages and send SIGBUS to the processes if the data was dirty.
1265 static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
1266 int flags, struct page *hpage)
1268 enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC;
1269 struct address_space *mapping;
1272 int kill = 1, forcekill;
1273 bool mlocked = PageMlocked(hpage);
1276 * Here we are interested only in user-mapped pages, so skip any
1277 * other types of pages.
1279 if (PageReserved(p) || PageSlab(p))
1281 if (!(PageLRU(hpage) || PageHuge(p)))
1285 * This check implies we don't kill processes if their pages
1286 * are in the swap cache early. Those are always late kills.
1288 if (!page_mapped(hpage))
1292 pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn);
1296 if (PageSwapCache(p)) {
1297 pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n",
1299 ttu |= TTU_IGNORE_HWPOISON;
1303 * Propagate the dirty bit from PTEs to struct page first, because we
1304 * need this to decide if we should kill or just drop the page.
1305 * XXX: the dirty test could be racy: set_page_dirty() may not always
1306 * be called inside page lock (it's recommended but not enforced).
1308 mapping = page_mapping(hpage);
1309 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
1310 mapping_can_writeback(mapping)) {
1311 if (page_mkclean(hpage)) {
1312 SetPageDirty(hpage);
1315 ttu |= TTU_IGNORE_HWPOISON;
1316 pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n",
1322 * First collect all the processes that have the page
1323 * mapped in dirty form. This has to be done before try_to_unmap,
1324 * because ttu takes the rmap data structures down.
1326 * Error handling: We ignore errors here because
1327 * there's nothing that can be done.
1330 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
1332 if (!PageHuge(hpage)) {
1333 try_to_unmap(hpage, ttu);
1335 if (!PageAnon(hpage)) {
1337 * For hugetlb pages in shared mappings, try_to_unmap
1338 * could potentially call huge_pmd_unshare. Because of
1339 * this, take semaphore in write mode here and set
1340 * TTU_RMAP_LOCKED to indicate we have taken the lock
1341 * at this higher level.
1343 mapping = hugetlb_page_mapping_lock_write(hpage);
1345 try_to_unmap(hpage, ttu|TTU_RMAP_LOCKED);
1346 i_mmap_unlock_write(mapping);
1348 pr_info("Memory failure: %#lx: could not lock mapping for mapped huge page\n", pfn);
1350 try_to_unmap(hpage, ttu);
1354 unmap_success = !page_mapped(hpage);
1356 pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n",
1357 pfn, page_mapcount(hpage));
1360 * try_to_unmap() might put mlocked page in lru cache, so call
1361 * shake_page() again to ensure that it's flushed.
1367 * Now that the dirty bit has been propagated to the
1368 * struct page and all unmaps done we can decide if
1369 * killing is needed or not. Only kill when the page
1370 * was dirty or the process is not restartable,
1371 * otherwise the tokill list is merely
1372 * freed. When there was a problem unmapping earlier
1373 * use a more force-full uncatchable kill to prevent
1374 * any accesses to the poisoned memory.
1376 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
1377 kill_procs(&tokill, forcekill, !unmap_success, pfn, flags);
1379 return unmap_success;
1382 static int identify_page_state(unsigned long pfn, struct page *p,
1383 unsigned long page_flags)
1385 struct page_state *ps;
1388 * The first check uses the current page flags which may not have any
1389 * relevant information. The second check with the saved page flags is
1390 * carried out only if the first check can't determine the page status.
1392 for (ps = error_states;; ps++)
1393 if ((p->flags & ps->mask) == ps->res)
1396 page_flags |= (p->flags & (1UL << PG_dirty));
1399 for (ps = error_states;; ps++)
1400 if ((page_flags & ps->mask) == ps->res)
1402 return page_action(ps, p, pfn);
1405 static int try_to_split_thp_page(struct page *page, const char *msg)
1408 if (!PageAnon(page) || unlikely(split_huge_page(page))) {
1409 unsigned long pfn = page_to_pfn(page);
1412 if (!PageAnon(page))
1413 pr_info("%s: %#lx: non anonymous thp\n", msg, pfn);
1415 pr_info("%s: %#lx: thp split failed\n", msg, pfn);
1425 * Called from hugetlb code with hugetlb_lock held.
1429 * 1 - in-use hugepage
1430 * 2 - not a hugepage
1431 * -EBUSY - the hugepage is busy (try to retry)
1432 * -EHWPOISON - the hugepage is already hwpoisoned
1434 int __get_huge_page_for_hwpoison(unsigned long pfn, int flags)
1436 struct page *page = pfn_to_page(pfn);
1437 struct page *head = compound_head(page);
1438 int ret = 2; /* fallback to normal page handling */
1439 bool count_increased = false;
1441 if (!PageHeadHuge(head))
1444 if (flags & MF_COUNT_INCREASED) {
1446 count_increased = true;
1447 } else if (HPageFreed(head) || HPageMigratable(head)) {
1448 ret = get_page_unless_zero(head);
1450 count_increased = true;
1456 if (TestSetPageHWPoison(head)) {
1463 if (count_increased)
1468 #ifdef CONFIG_HUGETLB_PAGE
1470 * Taking refcount of hugetlb pages needs extra care about race conditions
1471 * with basic operations like hugepage allocation/free/demotion.
1472 * So some of prechecks for hwpoison (pinning, and testing/setting
1473 * PageHWPoison) should be done in single hugetlb_lock range.
1475 static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb)
1478 struct page *p = pfn_to_page(pfn);
1480 unsigned long page_flags;
1485 res = get_huge_page_for_hwpoison(pfn, flags);
1486 if (res == 2) { /* fallback to normal page handling */
1489 } else if (res == -EHWPOISON) {
1490 pr_err("Memory failure: %#lx: already hardware poisoned\n", pfn);
1491 if (flags & MF_ACTION_REQUIRED) {
1492 head = compound_head(p);
1493 res = kill_accessing_process(current, page_to_pfn(head), flags);
1496 } else if (res == -EBUSY) {
1501 action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
1505 head = compound_head(p);
1508 if (hwpoison_filter(p)) {
1509 ClearPageHWPoison(head);
1514 num_poisoned_pages_inc();
1517 * Handling free hugepage. The possible race with hugepage allocation
1518 * or demotion can be prevented by PageHWPoison flag.
1523 if (__page_handle_poison(p)) {
1527 action_result(pfn, MF_MSG_FREE_HUGE, res);
1528 return res == MF_RECOVERED ? 0 : -EBUSY;
1531 page_flags = head->flags;
1534 * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so
1535 * simply disable it. In order to make it work properly, we need
1537 * - conversion of a pud that maps an error hugetlb into hwpoison
1538 * entry properly works, and
1539 * - other mm code walking over page table is aware of pud-aligned
1542 if (huge_page_size(page_hstate(head)) > PMD_SIZE) {
1543 action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED);
1548 if (!hwpoison_user_mappings(p, pfn, flags, head)) {
1549 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1554 return identify_page_state(pfn, p, page_flags);
1560 static inline int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb)
1566 static int memory_failure_dev_pagemap(unsigned long pfn, int flags,
1567 struct dev_pagemap *pgmap)
1569 struct page *page = pfn_to_page(pfn);
1570 unsigned long size = 0;
1577 if (flags & MF_COUNT_INCREASED)
1579 * Drop the extra refcount in case we come from madvise().
1583 /* device metadata space is not recoverable */
1584 if (!pgmap_pfn_valid(pgmap, pfn)) {
1590 * Prevent the inode from being freed while we are interrogating
1591 * the address_space, typically this would be handled by
1592 * lock_page(), but dax pages do not use the page lock. This
1593 * also prevents changes to the mapping of this pfn until
1594 * poison signaling is complete.
1596 cookie = dax_lock_page(page);
1600 if (hwpoison_filter(page)) {
1605 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
1607 * TODO: Handle HMM pages which may need coordination
1608 * with device-side memory.
1614 * Use this flag as an indication that the dax page has been
1615 * remapped UC to prevent speculative consumption of poison.
1617 SetPageHWPoison(page);
1620 * Unlike System-RAM there is no possibility to swap in a
1621 * different physical page at a given virtual address, so all
1622 * userspace consumption of ZONE_DEVICE memory necessitates
1623 * SIGBUS (i.e. MF_MUST_KILL)
1625 flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1626 collect_procs(page, &tokill, flags & MF_ACTION_REQUIRED);
1628 list_for_each_entry(tk, &tokill, nd)
1630 size = max(size, 1UL << tk->size_shift);
1633 * Unmap the largest mapping to avoid breaking up
1634 * device-dax mappings which are constant size. The
1635 * actual size of the mapping being torn down is
1636 * communicated in siginfo, see kill_proc()
1638 start = (page->index << PAGE_SHIFT) & ~(size - 1);
1639 unmap_mapping_range(page->mapping, start, size, 0);
1641 kill_procs(&tokill, flags & MF_MUST_KILL, false, pfn, flags);
1644 dax_unlock_page(page, cookie);
1646 /* drop pgmap ref acquired in caller */
1647 put_dev_pagemap(pgmap);
1648 action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED);
1652 static DEFINE_MUTEX(mf_mutex);
1655 * memory_failure - Handle memory failure of a page.
1656 * @pfn: Page Number of the corrupted page
1657 * @flags: fine tune action taken
1659 * This function is called by the low level machine check code
1660 * of an architecture when it detects hardware memory corruption
1661 * of a page. It tries its best to recover, which includes
1662 * dropping pages, killing processes etc.
1664 * The function is primarily of use for corruptions that
1665 * happen outside the current execution context (e.g. when
1666 * detected by a background scrubber)
1668 * Must run in process context (e.g. a work queue) with interrupts
1669 * enabled and no spinlocks hold.
1671 * Return: 0 for successfully handled the memory error,
1672 * -EOPNOTSUPP for memory_filter() filtered the error event,
1673 * < 0(except -EOPNOTSUPP) on failure.
1675 int memory_failure(unsigned long pfn, int flags)
1679 struct page *orig_head;
1680 struct dev_pagemap *pgmap;
1682 unsigned long page_flags;
1686 if (!sysctl_memory_failure_recovery)
1687 panic("Memory failure on page %lx", pfn);
1689 p = pfn_to_online_page(pfn);
1691 if (pfn_valid(pfn)) {
1692 pgmap = get_dev_pagemap(pfn, NULL);
1694 return memory_failure_dev_pagemap(pfn, flags,
1697 pr_err("Memory failure: %#lx: memory outside kernel control\n",
1702 mutex_lock(&mf_mutex);
1705 res = try_memory_failure_hugetlb(pfn, flags, &hugetlb);
1709 if (TestSetPageHWPoison(p)) {
1710 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1713 if (flags & MF_ACTION_REQUIRED)
1714 res = kill_accessing_process(current, pfn, flags);
1718 orig_head = hpage = compound_head(p);
1719 num_poisoned_pages_inc();
1722 * We need/can do nothing about count=0 pages.
1723 * 1) it's a free page, and therefore in safe hand:
1724 * prep_new_page() will be the gate keeper.
1725 * 2) it's part of a non-compound high order page.
1726 * Implies some kernel user: cannot stop them from
1727 * R/W the page; let's pray that the page has been
1728 * used and will be freed some time later.
1729 * In fact it's dangerous to directly bump up page count from 0,
1730 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
1732 if (!(flags & MF_COUNT_INCREASED)) {
1733 res = get_hwpoison_page(p, flags);
1735 if (is_free_buddy_page(p)) {
1736 if (take_page_off_buddy(p)) {
1740 /* We lost the race, try again */
1742 ClearPageHWPoison(p);
1743 num_poisoned_pages_dec();
1749 action_result(pfn, MF_MSG_BUDDY, res);
1750 res = res == MF_RECOVERED ? 0 : -EBUSY;
1752 action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
1756 } else if (res < 0) {
1757 action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
1763 if (PageTransHuge(hpage)) {
1765 * The flag must be set after the refcount is bumped
1766 * otherwise it may race with THP split.
1767 * And the flag can't be set in get_hwpoison_page() since
1768 * it is called by soft offline too and it is just called
1769 * for !MF_COUNT_INCREASE. So here seems to be the best
1772 * Don't need care about the above error handling paths for
1773 * get_hwpoison_page() since they handle either free page
1774 * or unhandlable page. The refcount is bumped iff the
1775 * page is a valid handlable page.
1777 SetPageHasHWPoisoned(hpage);
1778 if (try_to_split_thp_page(p, "Memory Failure") < 0) {
1779 action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED);
1783 VM_BUG_ON_PAGE(!page_count(p), p);
1787 * We ignore non-LRU pages for good reasons.
1788 * - PG_locked is only well defined for LRU pages and a few others
1789 * - to avoid races with __SetPageLocked()
1790 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1791 * The check (unnecessarily) ignores LRU pages being isolated and
1792 * walked by the page reclaim code, however that's not a big loss.
1799 * The page could have changed compound pages during the locking.
1800 * If this happens just bail out.
1802 if (PageCompound(p) && compound_head(p) != orig_head) {
1803 action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
1809 * We use page flags to determine what action should be taken, but
1810 * the flags can be modified by the error containment action. One
1811 * example is an mlocked page, where PG_mlocked is cleared by
1812 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1813 * correctly, we save a copy of the page flags at this time.
1815 page_flags = p->flags;
1817 if (hwpoison_filter(p)) {
1818 if (TestClearPageHWPoison(p))
1819 num_poisoned_pages_dec();
1827 * __munlock_pagevec may clear a writeback page's LRU flag without
1828 * page_lock. We need wait writeback completion for this page or it
1829 * may trigger vfs BUG while evict inode.
1831 if (!PageTransTail(p) && !PageLRU(p) && !PageWriteback(p))
1832 goto identify_page_state;
1835 * It's very difficult to mess with pages currently under IO
1836 * and in many cases impossible, so we just avoid it here.
1838 wait_on_page_writeback(p);
1841 * Now take care of user space mappings.
1842 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1844 if (!hwpoison_user_mappings(p, pfn, flags, p)) {
1845 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1851 * Torn down by someone else?
1853 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1854 action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
1859 identify_page_state:
1860 res = identify_page_state(pfn, p, page_flags);
1861 mutex_unlock(&mf_mutex);
1866 mutex_unlock(&mf_mutex);
1869 EXPORT_SYMBOL_GPL(memory_failure);
1871 #define MEMORY_FAILURE_FIFO_ORDER 4
1872 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1874 struct memory_failure_entry {
1879 struct memory_failure_cpu {
1880 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1881 MEMORY_FAILURE_FIFO_SIZE);
1883 struct work_struct work;
1886 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1889 * memory_failure_queue - Schedule handling memory failure of a page.
1890 * @pfn: Page Number of the corrupted page
1891 * @flags: Flags for memory failure handling
1893 * This function is called by the low level hardware error handler
1894 * when it detects hardware memory corruption of a page. It schedules
1895 * the recovering of error page, including dropping pages, killing
1898 * The function is primarily of use for corruptions that
1899 * happen outside the current execution context (e.g. when
1900 * detected by a background scrubber)
1902 * Can run in IRQ context.
1904 void memory_failure_queue(unsigned long pfn, int flags)
1906 struct memory_failure_cpu *mf_cpu;
1907 unsigned long proc_flags;
1908 struct memory_failure_entry entry = {
1913 mf_cpu = &get_cpu_var(memory_failure_cpu);
1914 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1915 if (kfifo_put(&mf_cpu->fifo, entry))
1916 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1918 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1920 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1921 put_cpu_var(memory_failure_cpu);
1923 EXPORT_SYMBOL_GPL(memory_failure_queue);
1925 static void memory_failure_work_func(struct work_struct *work)
1927 struct memory_failure_cpu *mf_cpu;
1928 struct memory_failure_entry entry = { 0, };
1929 unsigned long proc_flags;
1932 mf_cpu = container_of(work, struct memory_failure_cpu, work);
1934 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1935 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1936 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1939 if (entry.flags & MF_SOFT_OFFLINE)
1940 soft_offline_page(entry.pfn, entry.flags);
1942 memory_failure(entry.pfn, entry.flags);
1947 * Process memory_failure work queued on the specified CPU.
1948 * Used to avoid return-to-userspace racing with the memory_failure workqueue.
1950 void memory_failure_queue_kick(int cpu)
1952 struct memory_failure_cpu *mf_cpu;
1954 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1955 cancel_work_sync(&mf_cpu->work);
1956 memory_failure_work_func(&mf_cpu->work);
1959 static int __init memory_failure_init(void)
1961 struct memory_failure_cpu *mf_cpu;
1964 for_each_possible_cpu(cpu) {
1965 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1966 spin_lock_init(&mf_cpu->lock);
1967 INIT_KFIFO(mf_cpu->fifo);
1968 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1973 core_initcall(memory_failure_init);
1975 #define unpoison_pr_info(fmt, pfn, rs) \
1977 if (__ratelimit(rs)) \
1978 pr_info(fmt, pfn); \
1982 * unpoison_memory - Unpoison a previously poisoned page
1983 * @pfn: Page number of the to be unpoisoned page
1985 * Software-unpoison a page that has been poisoned by
1986 * memory_failure() earlier.
1988 * This is only done on the software-level, so it only works
1989 * for linux injected failures, not real hardware failures
1991 * Returns 0 for success, otherwise -errno.
1993 int unpoison_memory(unsigned long pfn)
1999 unsigned long flags = 0;
2000 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
2001 DEFAULT_RATELIMIT_BURST);
2003 if (!pfn_valid(pfn))
2006 p = pfn_to_page(pfn);
2007 page = compound_head(p);
2009 mutex_lock(&mf_mutex);
2011 if (!PageHWPoison(p)) {
2012 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
2017 if (page_count(page) > 1) {
2018 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
2023 if (page_mapped(page)) {
2024 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
2029 if (page_mapping(page)) {
2030 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
2035 if (!get_hwpoison_page(p, flags)) {
2036 if (TestClearPageHWPoison(p))
2037 num_poisoned_pages_dec();
2038 unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n",
2043 if (TestClearPageHWPoison(page)) {
2044 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
2046 num_poisoned_pages_dec();
2051 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
2055 mutex_unlock(&mf_mutex);
2058 EXPORT_SYMBOL(unpoison_memory);
2060 static bool isolate_page(struct page *page, struct list_head *pagelist)
2062 bool isolated = false;
2063 bool lru = PageLRU(page);
2065 if (PageHuge(page)) {
2066 isolated = isolate_huge_page(page, pagelist);
2069 isolated = !isolate_lru_page(page);
2071 isolated = !isolate_movable_page(page, ISOLATE_UNEVICTABLE);
2074 list_add(&page->lru, pagelist);
2077 if (isolated && lru)
2078 inc_node_page_state(page, NR_ISOLATED_ANON +
2079 page_is_file_lru(page));
2082 * If we succeed to isolate the page, we grabbed another refcount on
2083 * the page, so we can safely drop the one we got from get_any_pages().
2084 * If we failed to isolate the page, it means that we cannot go further
2085 * and we will return an error, so drop the reference we got from
2086 * get_any_pages() as well.
2093 * __soft_offline_page handles hugetlb-pages and non-hugetlb pages.
2094 * If the page is a non-dirty unmapped page-cache page, it simply invalidates.
2095 * If the page is mapped, it migrates the contents over.
2097 static int __soft_offline_page(struct page *page)
2100 unsigned long pfn = page_to_pfn(page);
2101 struct page *hpage = compound_head(page);
2102 char const *msg_page[] = {"page", "hugepage"};
2103 bool huge = PageHuge(page);
2104 LIST_HEAD(pagelist);
2105 struct migration_target_control mtc = {
2106 .nid = NUMA_NO_NODE,
2107 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
2111 * Check PageHWPoison again inside page lock because PageHWPoison
2112 * is set by memory_failure() outside page lock. Note that
2113 * memory_failure() also double-checks PageHWPoison inside page lock,
2114 * so there's no race between soft_offline_page() and memory_failure().
2117 if (!PageHuge(page))
2118 wait_on_page_writeback(page);
2119 if (PageHWPoison(page)) {
2122 pr_info("soft offline: %#lx page already poisoned\n", pfn);
2126 if (!PageHuge(page))
2128 * Try to invalidate first. This should work for
2129 * non dirty unmapped page cache pages.
2131 ret = invalidate_inode_page(page);
2135 * RED-PEN would be better to keep it isolated here, but we
2136 * would need to fix isolation locking first.
2139 pr_info("soft_offline: %#lx: invalidated\n", pfn);
2140 page_handle_poison(page, false, true);
2144 if (isolate_page(hpage, &pagelist)) {
2145 ret = migrate_pages(&pagelist, alloc_migration_target, NULL,
2146 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE, NULL);
2148 bool release = !huge;
2150 if (!page_handle_poison(page, huge, release))
2153 if (!list_empty(&pagelist))
2154 putback_movable_pages(&pagelist);
2156 pr_info("soft offline: %#lx: %s migration failed %d, type %lx (%pGp)\n",
2157 pfn, msg_page[huge], ret, page->flags, &page->flags);
2162 pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %lx (%pGp)\n",
2163 pfn, msg_page[huge], page_count(page), page->flags, &page->flags);
2169 static int soft_offline_in_use_page(struct page *page)
2171 struct page *hpage = compound_head(page);
2173 if (!PageHuge(page) && PageTransHuge(hpage))
2174 if (try_to_split_thp_page(page, "soft offline") < 0)
2176 return __soft_offline_page(page);
2179 static int soft_offline_free_page(struct page *page)
2183 if (!page_handle_poison(page, true, false))
2189 static void put_ref_page(struct page *page)
2196 * soft_offline_page - Soft offline a page.
2197 * @pfn: pfn to soft-offline
2198 * @flags: flags. Same as memory_failure().
2200 * Returns 0 on success, otherwise negated errno.
2202 * Soft offline a page, by migration or invalidation,
2203 * without killing anything. This is for the case when
2204 * a page is not corrupted yet (so it's still valid to access),
2205 * but has had a number of corrected errors and is better taken
2208 * The actual policy on when to do that is maintained by
2211 * This should never impact any application or cause data loss,
2212 * however it might take some time.
2214 * This is not a 100% solution for all memory, but tries to be
2215 * ``good enough'' for the majority of memory.
2217 int soft_offline_page(unsigned long pfn, int flags)
2220 bool try_again = true;
2221 struct page *page, *ref_page = NULL;
2223 WARN_ON_ONCE(!pfn_valid(pfn) && (flags & MF_COUNT_INCREASED));
2225 if (!pfn_valid(pfn))
2227 if (flags & MF_COUNT_INCREASED)
2228 ref_page = pfn_to_page(pfn);
2230 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
2231 page = pfn_to_online_page(pfn);
2233 put_ref_page(ref_page);
2237 mutex_lock(&mf_mutex);
2239 if (PageHWPoison(page)) {
2240 pr_info("%s: %#lx page already poisoned\n", __func__, pfn);
2241 put_ref_page(ref_page);
2242 mutex_unlock(&mf_mutex);
2248 ret = get_hwpoison_page(page, flags);
2252 ret = soft_offline_in_use_page(page);
2253 } else if (ret == 0) {
2254 if (soft_offline_free_page(page) && try_again) {
2256 flags &= ~MF_COUNT_INCREASED;
2261 mutex_unlock(&mf_mutex);