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, bool hugepage_or_freepage, bool release)
71 if (hugepage_or_freepage) {
73 * Doing this check for free pages is also fine since dissolve_free_huge_page
74 * returns 0 for non-hugetlb pages as well.
76 if (dissolve_free_huge_page(page) || !take_page_off_buddy(page))
78 * We could fail to take off the target page from buddy
79 * for example due to racy page allocation, but that's
80 * acceptable because soft-offlined page is not broken
81 * and if someone really want to use it, they should
87 SetPageHWPoison(page);
91 num_poisoned_pages_inc();
96 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
98 u32 hwpoison_filter_enable = 0;
99 u32 hwpoison_filter_dev_major = ~0U;
100 u32 hwpoison_filter_dev_minor = ~0U;
101 u64 hwpoison_filter_flags_mask;
102 u64 hwpoison_filter_flags_value;
103 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
104 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
105 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
106 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
107 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
109 static int hwpoison_filter_dev(struct page *p)
111 struct address_space *mapping;
114 if (hwpoison_filter_dev_major == ~0U &&
115 hwpoison_filter_dev_minor == ~0U)
119 * page_mapping() does not accept slab pages.
124 mapping = page_mapping(p);
125 if (mapping == NULL || mapping->host == NULL)
128 dev = mapping->host->i_sb->s_dev;
129 if (hwpoison_filter_dev_major != ~0U &&
130 hwpoison_filter_dev_major != MAJOR(dev))
132 if (hwpoison_filter_dev_minor != ~0U &&
133 hwpoison_filter_dev_minor != MINOR(dev))
139 static int hwpoison_filter_flags(struct page *p)
141 if (!hwpoison_filter_flags_mask)
144 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
145 hwpoison_filter_flags_value)
152 * This allows stress tests to limit test scope to a collection of tasks
153 * by putting them under some memcg. This prevents killing unrelated/important
154 * processes such as /sbin/init. Note that the target task may share clean
155 * pages with init (eg. libc text), which is harmless. If the target task
156 * share _dirty_ pages with another task B, the test scheme must make sure B
157 * is also included in the memcg. At last, due to race conditions this filter
158 * can only guarantee that the page either belongs to the memcg tasks, or is
162 u64 hwpoison_filter_memcg;
163 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
164 static int hwpoison_filter_task(struct page *p)
166 if (!hwpoison_filter_memcg)
169 if (page_cgroup_ino(p) != hwpoison_filter_memcg)
175 static int hwpoison_filter_task(struct page *p) { return 0; }
178 int hwpoison_filter(struct page *p)
180 if (!hwpoison_filter_enable)
183 if (hwpoison_filter_dev(p))
186 if (hwpoison_filter_flags(p))
189 if (hwpoison_filter_task(p))
195 int hwpoison_filter(struct page *p)
201 EXPORT_SYMBOL_GPL(hwpoison_filter);
204 * Kill all processes that have a poisoned page mapped and then isolate
208 * Find all processes having the page mapped and kill them.
209 * But we keep a page reference around so that the page is not
210 * actually freed yet.
211 * Then stash the page away
213 * There's no convenient way to get back to mapped processes
214 * from the VMAs. So do a brute-force search over all
217 * Remember that machine checks are not common (or rather
218 * if they are common you have other problems), so this shouldn't
219 * be a performance issue.
221 * Also there are some races possible while we get from the
222 * error detection to actually handle it.
227 struct task_struct *tsk;
233 * Send all the processes who have the page mapped a signal.
234 * ``action optional'' if they are not immediately affected by the error
235 * ``action required'' if error happened in current execution context
237 static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
239 struct task_struct *t = tk->tsk;
240 short addr_lsb = tk->size_shift;
243 pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
244 pfn, t->comm, t->pid);
246 if (flags & MF_ACTION_REQUIRED) {
248 ret = force_sig_mceerr(BUS_MCEERR_AR,
249 (void __user *)tk->addr, addr_lsb);
251 /* Signal other processes sharing the page if they have PF_MCE_EARLY set. */
252 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
256 * Don't use force here, it's convenient if the signal
257 * can be temporarily blocked.
258 * This could cause a loop when the user sets SIGBUS
259 * to SIG_IGN, but hopefully no one will do that?
261 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
262 addr_lsb, t); /* synchronous? */
265 pr_info("Memory failure: Error sending signal to %s:%d: %d\n",
266 t->comm, t->pid, ret);
271 * Unknown page type encountered. Try to check whether it can turn PageLRU by
272 * lru_add_drain_all, or a free page by reclaiming slabs when possible.
274 void shake_page(struct page *p, int access)
281 if (PageLRU(p) || is_free_buddy_page(p))
286 * Only call shrink_node_slabs here (which would also shrink
287 * other caches) if access is not potentially fatal.
290 drop_slab_node(page_to_nid(p));
292 EXPORT_SYMBOL_GPL(shake_page);
294 static unsigned long dev_pagemap_mapping_shift(struct page *page,
295 struct vm_area_struct *vma)
297 unsigned long address = vma_address(page, vma);
304 pgd = pgd_offset(vma->vm_mm, address);
305 if (!pgd_present(*pgd))
307 p4d = p4d_offset(pgd, address);
308 if (!p4d_present(*p4d))
310 pud = pud_offset(p4d, address);
311 if (!pud_present(*pud))
313 if (pud_devmap(*pud))
315 pmd = pmd_offset(pud, address);
316 if (!pmd_present(*pmd))
318 if (pmd_devmap(*pmd))
320 pte = pte_offset_map(pmd, address);
321 if (!pte_present(*pte))
323 if (pte_devmap(*pte))
329 * Failure handling: if we can't find or can't kill a process there's
330 * not much we can do. We just print a message and ignore otherwise.
334 * Schedule a process for later kill.
335 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
337 static void add_to_kill(struct task_struct *tsk, struct page *p,
338 struct vm_area_struct *vma,
339 struct list_head *to_kill)
343 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
345 pr_err("Memory failure: Out of memory while machine check handling\n");
349 tk->addr = page_address_in_vma(p, vma);
350 if (is_zone_device_page(p))
351 tk->size_shift = dev_pagemap_mapping_shift(p, vma);
353 tk->size_shift = page_shift(compound_head(p));
356 * Send SIGKILL if "tk->addr == -EFAULT". Also, as
357 * "tk->size_shift" is always non-zero for !is_zone_device_page(),
358 * so "tk->size_shift == 0" effectively checks no mapping on
359 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times
360 * to a process' address space, it's possible not all N VMAs
361 * contain mappings for the page, but at least one VMA does.
362 * Only deliver SIGBUS with payload derived from the VMA that
363 * has a mapping for the page.
365 if (tk->addr == -EFAULT) {
366 pr_info("Memory failure: Unable to find user space address %lx in %s\n",
367 page_to_pfn(p), tsk->comm);
368 } else if (tk->size_shift == 0) {
373 get_task_struct(tsk);
375 list_add_tail(&tk->nd, to_kill);
379 * Kill the processes that have been collected earlier.
381 * Only do anything when DOIT is set, otherwise just free the list
382 * (this is used for clean pages which do not need killing)
383 * Also when FAIL is set do a force kill because something went
386 static void kill_procs(struct list_head *to_kill, int forcekill, bool fail,
387 unsigned long pfn, int flags)
389 struct to_kill *tk, *next;
391 list_for_each_entry_safe (tk, next, to_kill, nd) {
394 * In case something went wrong with munmapping
395 * make sure the process doesn't catch the
396 * signal and then access the memory. Just kill it.
398 if (fail || tk->addr == -EFAULT) {
399 pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
400 pfn, tk->tsk->comm, tk->tsk->pid);
401 do_send_sig_info(SIGKILL, SEND_SIG_PRIV,
402 tk->tsk, PIDTYPE_PID);
406 * In theory the process could have mapped
407 * something else on the address in-between. We could
408 * check for that, but we need to tell the
411 else if (kill_proc(tk, pfn, flags) < 0)
412 pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n",
413 pfn, tk->tsk->comm, tk->tsk->pid);
415 put_task_struct(tk->tsk);
421 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
422 * on behalf of the thread group. Return task_struct of the (first found)
423 * dedicated thread if found, and return NULL otherwise.
425 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
426 * have to call rcu_read_lock/unlock() in this function.
428 static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
430 struct task_struct *t;
432 for_each_thread(tsk, t) {
433 if (t->flags & PF_MCE_PROCESS) {
434 if (t->flags & PF_MCE_EARLY)
437 if (sysctl_memory_failure_early_kill)
445 * Determine whether a given process is "early kill" process which expects
446 * to be signaled when some page under the process is hwpoisoned.
447 * Return task_struct of the dedicated thread (main thread unless explicitly
448 * specified) if the process is "early kill" and otherwise returns NULL.
450 * Note that the above is true for Action Optional case. For Action Required
451 * case, it's only meaningful to the current thread which need to be signaled
452 * with SIGBUS, this error is Action Optional for other non current
453 * processes sharing the same error page,if the process is "early kill", the
454 * task_struct of the dedicated thread will also be returned.
456 static struct task_struct *task_early_kill(struct task_struct *tsk,
462 * Comparing ->mm here because current task might represent
463 * a subthread, while tsk always points to the main thread.
465 if (force_early && tsk->mm == current->mm)
468 return find_early_kill_thread(tsk);
472 * Collect processes when the error hit an anonymous page.
474 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
477 struct vm_area_struct *vma;
478 struct task_struct *tsk;
482 av = page_lock_anon_vma_read(page);
483 if (av == NULL) /* Not actually mapped anymore */
486 pgoff = page_to_pgoff(page);
487 read_lock(&tasklist_lock);
488 for_each_process (tsk) {
489 struct anon_vma_chain *vmac;
490 struct task_struct *t = task_early_kill(tsk, force_early);
494 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
497 if (!page_mapped_in_vma(page, vma))
499 if (vma->vm_mm == t->mm)
500 add_to_kill(t, page, vma, to_kill);
503 read_unlock(&tasklist_lock);
504 page_unlock_anon_vma_read(av);
508 * Collect processes when the error hit a file mapped page.
510 static void collect_procs_file(struct page *page, struct list_head *to_kill,
513 struct vm_area_struct *vma;
514 struct task_struct *tsk;
515 struct address_space *mapping = page->mapping;
518 i_mmap_lock_read(mapping);
519 read_lock(&tasklist_lock);
520 pgoff = page_to_pgoff(page);
521 for_each_process(tsk) {
522 struct task_struct *t = task_early_kill(tsk, force_early);
526 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
529 * Send early kill signal to tasks where a vma covers
530 * the page but the corrupted page is not necessarily
531 * mapped it in its pte.
532 * Assume applications who requested early kill want
533 * to be informed of all such data corruptions.
535 if (vma->vm_mm == t->mm)
536 add_to_kill(t, page, vma, to_kill);
539 read_unlock(&tasklist_lock);
540 i_mmap_unlock_read(mapping);
544 * Collect the processes who have the corrupted page mapped to kill.
546 static void collect_procs(struct page *page, struct list_head *tokill,
553 collect_procs_anon(page, tokill, force_early);
555 collect_procs_file(page, tokill, force_early);
564 static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift)
567 tk->size_shift = shift;
570 static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift,
571 unsigned long poisoned_pfn, struct to_kill *tk)
573 unsigned long pfn = 0;
575 if (pte_present(pte)) {
578 swp_entry_t swp = pte_to_swp_entry(pte);
580 if (is_hwpoison_entry(swp))
581 pfn = hwpoison_entry_to_pfn(swp);
584 if (!pfn || pfn != poisoned_pfn)
587 set_to_kill(tk, addr, shift);
591 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
592 static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
593 struct hwp_walk *hwp)
597 unsigned long hwpoison_vaddr;
599 if (!pmd_present(pmd))
602 if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) {
603 hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT);
604 set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT);
610 static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
611 struct hwp_walk *hwp)
617 static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr,
618 unsigned long end, struct mm_walk *walk)
620 struct hwp_walk *hwp = (struct hwp_walk *)walk->private;
625 ptl = pmd_trans_huge_lock(pmdp, walk->vma);
627 ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp);
632 if (pmd_trans_unstable(pmdp))
635 ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp, addr, &ptl);
636 for (; addr != end; ptep++, addr += PAGE_SIZE) {
637 ret = check_hwpoisoned_entry(*ptep, addr, PAGE_SHIFT,
642 pte_unmap_unlock(ptep - 1, ptl);
648 #ifdef CONFIG_HUGETLB_PAGE
649 static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask,
650 unsigned long addr, unsigned long end,
651 struct mm_walk *walk)
653 struct hwp_walk *hwp = (struct hwp_walk *)walk->private;
654 pte_t pte = huge_ptep_get(ptep);
655 struct hstate *h = hstate_vma(walk->vma);
657 return check_hwpoisoned_entry(pte, addr, huge_page_shift(h),
661 #define hwpoison_hugetlb_range NULL
664 static struct mm_walk_ops hwp_walk_ops = {
665 .pmd_entry = hwpoison_pte_range,
666 .hugetlb_entry = hwpoison_hugetlb_range,
670 * Sends SIGBUS to the current process with error info.
672 * This function is intended to handle "Action Required" MCEs on already
673 * hardware poisoned pages. They could happen, for example, when
674 * memory_failure() failed to unmap the error page at the first call, or
675 * when multiple local machine checks happened on different CPUs.
677 * MCE handler currently has no easy access to the error virtual address,
678 * so this function walks page table to find it. The returned virtual address
679 * is proper in most cases, but it could be wrong when the application
680 * process has multiple entries mapping the error page.
682 static int kill_accessing_process(struct task_struct *p, unsigned long pfn,
686 struct hwp_walk priv = {
691 mmap_read_lock(p->mm);
692 ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwp_walk_ops,
694 if (ret == 1 && priv.tk.addr)
695 kill_proc(&priv.tk, pfn, flags);
696 mmap_read_unlock(p->mm);
697 return ret ? -EFAULT : -EHWPOISON;
700 static const char *action_name[] = {
701 [MF_IGNORED] = "Ignored",
702 [MF_FAILED] = "Failed",
703 [MF_DELAYED] = "Delayed",
704 [MF_RECOVERED] = "Recovered",
707 static const char * const action_page_types[] = {
708 [MF_MSG_KERNEL] = "reserved kernel page",
709 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
710 [MF_MSG_SLAB] = "kernel slab page",
711 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
712 [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned",
713 [MF_MSG_HUGE] = "huge page",
714 [MF_MSG_FREE_HUGE] = "free huge page",
715 [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page",
716 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
717 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
718 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
719 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
720 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
721 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
722 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
723 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
724 [MF_MSG_CLEAN_LRU] = "clean LRU page",
725 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
726 [MF_MSG_BUDDY] = "free buddy page",
727 [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)",
728 [MF_MSG_DAX] = "dax page",
729 [MF_MSG_UNSPLIT_THP] = "unsplit thp",
730 [MF_MSG_UNKNOWN] = "unknown page",
734 * XXX: It is possible that a page is isolated from LRU cache,
735 * and then kept in swap cache or failed to remove from page cache.
736 * The page count will stop it from being freed by unpoison.
737 * Stress tests should be aware of this memory leak problem.
739 static int delete_from_lru_cache(struct page *p)
741 if (!isolate_lru_page(p)) {
743 * Clear sensible page flags, so that the buddy system won't
744 * complain when the page is unpoison-and-freed.
747 ClearPageUnevictable(p);
750 * Poisoned page might never drop its ref count to 0 so we have
751 * to uncharge it manually from its memcg.
753 mem_cgroup_uncharge(p);
756 * drop the page count elevated by isolate_lru_page()
764 static int truncate_error_page(struct page *p, unsigned long pfn,
765 struct address_space *mapping)
769 if (mapping->a_ops->error_remove_page) {
770 int err = mapping->a_ops->error_remove_page(mapping, p);
773 pr_info("Memory failure: %#lx: Failed to punch page: %d\n",
775 } else if (page_has_private(p) &&
776 !try_to_release_page(p, GFP_NOIO)) {
777 pr_info("Memory failure: %#lx: failed to release buffers\n",
784 * If the file system doesn't support it just invalidate
785 * This fails on dirty or anything with private pages
787 if (invalidate_inode_page(p))
790 pr_info("Memory failure: %#lx: Failed to invalidate\n",
798 * Error hit kernel page.
799 * Do nothing, try to be lucky and not touch this instead. For a few cases we
800 * could be more sophisticated.
802 static int me_kernel(struct page *p, unsigned long pfn)
809 * Page in unknown state. Do nothing.
811 static int me_unknown(struct page *p, unsigned long pfn)
813 pr_err("Memory failure: %#lx: Unknown page state\n", pfn);
819 * Clean (or cleaned) page cache page.
821 static int me_pagecache_clean(struct page *p, unsigned long pfn)
824 struct address_space *mapping;
826 delete_from_lru_cache(p);
829 * For anonymous pages we're done the only reference left
830 * should be the one m_f() holds.
838 * Now truncate the page in the page cache. This is really
839 * more like a "temporary hole punch"
840 * Don't do this for block devices when someone else
841 * has a reference, because it could be file system metadata
842 * and that's not safe to truncate.
844 mapping = page_mapping(p);
847 * Page has been teared down in the meanwhile
854 * Truncation is a bit tricky. Enable it per file system for now.
856 * Open: to take i_mutex or not for this? Right now we don't.
858 ret = truncate_error_page(p, pfn, mapping);
865 * Dirty pagecache page
866 * Issues: when the error hit a hole page the error is not properly
869 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
871 struct address_space *mapping = page_mapping(p);
874 /* TBD: print more information about the file. */
877 * IO error will be reported by write(), fsync(), etc.
878 * who check the mapping.
879 * This way the application knows that something went
880 * wrong with its dirty file data.
882 * There's one open issue:
884 * The EIO will be only reported on the next IO
885 * operation and then cleared through the IO map.
886 * Normally Linux has two mechanisms to pass IO error
887 * first through the AS_EIO flag in the address space
888 * and then through the PageError flag in the page.
889 * Since we drop pages on memory failure handling the
890 * only mechanism open to use is through AS_AIO.
892 * This has the disadvantage that it gets cleared on
893 * the first operation that returns an error, while
894 * the PageError bit is more sticky and only cleared
895 * when the page is reread or dropped. If an
896 * application assumes it will always get error on
897 * fsync, but does other operations on the fd before
898 * and the page is dropped between then the error
899 * will not be properly reported.
901 * This can already happen even without hwpoisoned
902 * pages: first on metadata IO errors (which only
903 * report through AS_EIO) or when the page is dropped
906 * So right now we assume that the application DTRT on
907 * the first EIO, but we're not worse than other parts
910 mapping_set_error(mapping, -EIO);
913 return me_pagecache_clean(p, pfn);
917 * Clean and dirty swap cache.
919 * Dirty swap cache page is tricky to handle. The page could live both in page
920 * cache and swap cache(ie. page is freshly swapped in). So it could be
921 * referenced concurrently by 2 types of PTEs:
922 * normal PTEs and swap PTEs. We try to handle them consistently by calling
923 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
925 * - clear dirty bit to prevent IO
927 * - but keep in the swap cache, so that when we return to it on
928 * a later page fault, we know the application is accessing
929 * corrupted data and shall be killed (we installed simple
930 * interception code in do_swap_page to catch it).
932 * Clean swap cache pages can be directly isolated. A later page fault will
933 * bring in the known good data from disk.
935 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
940 /* Trigger EIO in shmem: */
941 ClearPageUptodate(p);
943 ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED;
948 static int me_swapcache_clean(struct page *p, unsigned long pfn)
952 delete_from_swap_cache(p);
954 ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED;
960 * Huge pages. Needs work.
962 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
963 * To narrow down kill region to one page, we need to break up pmd.
965 static int me_huge_page(struct page *p, unsigned long pfn)
968 struct page *hpage = compound_head(p);
969 struct address_space *mapping;
971 if (!PageHuge(hpage))
974 mapping = page_mapping(hpage);
976 res = truncate_error_page(hpage, pfn, mapping);
982 * migration entry prevents later access on error anonymous
983 * hugepage, so we can free and dissolve it into buddy to
984 * save healthy subpages.
988 if (!dissolve_free_huge_page(p) && take_page_off_buddy(p)) {
998 * Various page states we can handle.
1000 * A page state is defined by its current page->flags bits.
1001 * The table matches them in order and calls the right handler.
1003 * This is quite tricky because we can access page at any time
1004 * in its live cycle, so all accesses have to be extremely careful.
1006 * This is not complete. More states could be added.
1007 * For any missing state don't attempt recovery.
1010 #define dirty (1UL << PG_dirty)
1011 #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked))
1012 #define unevict (1UL << PG_unevictable)
1013 #define mlock (1UL << PG_mlocked)
1014 #define lru (1UL << PG_lru)
1015 #define head (1UL << PG_head)
1016 #define slab (1UL << PG_slab)
1017 #define reserved (1UL << PG_reserved)
1019 static struct page_state {
1022 enum mf_action_page_type type;
1024 /* Callback ->action() has to unlock the relevant page inside it. */
1025 int (*action)(struct page *p, unsigned long pfn);
1026 } error_states[] = {
1027 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
1029 * free pages are specially detected outside this table:
1030 * PG_buddy pages only make a small fraction of all free pages.
1034 * Could in theory check if slab page is free or if we can drop
1035 * currently unused objects without touching them. But just
1036 * treat it as standard kernel for now.
1038 { slab, slab, MF_MSG_SLAB, me_kernel },
1040 { head, head, MF_MSG_HUGE, me_huge_page },
1042 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
1043 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
1045 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
1046 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
1048 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
1049 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
1051 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
1052 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
1055 * Catchall entry: must be at end.
1057 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
1070 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
1071 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
1073 static void action_result(unsigned long pfn, enum mf_action_page_type type,
1074 enum mf_result result)
1076 trace_memory_failure_event(pfn, type, result);
1078 pr_err("Memory failure: %#lx: recovery action for %s: %s\n",
1079 pfn, action_page_types[type], action_name[result]);
1082 static int page_action(struct page_state *ps, struct page *p,
1088 /* page p should be unlocked after returning from ps->action(). */
1089 result = ps->action(p, pfn);
1091 count = page_count(p) - 1;
1092 if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
1095 pr_err("Memory failure: %#lx: %s still referenced by %d users\n",
1096 pfn, action_page_types[ps->type], count);
1099 action_result(pfn, ps->type, result);
1101 /* Could do more checks here if page looks ok */
1103 * Could adjust zone counters here to correct for the missing page.
1106 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
1110 * Return true if a page type of a given page is supported by hwpoison
1111 * mechanism (while handling could fail), otherwise false. This function
1112 * does not return true for hugetlb or device memory pages, so it's assumed
1113 * to be called only in the context where we never have such pages.
1115 static inline bool HWPoisonHandlable(struct page *page)
1117 return PageLRU(page) || __PageMovable(page);
1121 * __get_hwpoison_page() - Get refcount for memory error handling:
1122 * @page: raw error page (hit by memory error)
1124 * Return: return 0 if failed to grab the refcount, otherwise true (some
1127 static int __get_hwpoison_page(struct page *page)
1129 struct page *head = compound_head(page);
1131 bool hugetlb = false;
1133 ret = get_hwpoison_huge_page(head, &hugetlb);
1138 * This check prevents from calling get_hwpoison_unless_zero()
1139 * for any unsupported type of page in order to reduce the risk of
1140 * unexpected races caused by taking a page refcount.
1142 if (!HWPoisonHandlable(head))
1145 if (PageTransHuge(head)) {
1147 * Non anonymous thp exists only in allocation/free time. We
1148 * can't handle such a case correctly, so let's give it up.
1149 * This should be better than triggering BUG_ON when kernel
1150 * tries to touch the "partially handled" page.
1152 if (!PageAnon(head)) {
1153 pr_err("Memory failure: %#lx: non anonymous thp\n",
1159 if (get_page_unless_zero(head)) {
1160 if (head == compound_head(page))
1163 pr_info("Memory failure: %#lx cannot catch tail\n",
1172 * Safely get reference count of an arbitrary page.
1174 * Returns 0 for a free page, 1 for an in-use page,
1175 * -EIO for a page-type we cannot handle and -EBUSY if we raced with an
1177 * We only incremented refcount in case the page was already in-use and it
1178 * is a known type we can handle.
1180 static int get_any_page(struct page *p, unsigned long flags)
1182 int ret = 0, pass = 0;
1183 bool count_increased = false;
1185 if (flags & MF_COUNT_INCREASED)
1186 count_increased = true;
1189 if (!count_increased && !__get_hwpoison_page(p)) {
1190 if (page_count(p)) {
1191 /* We raced with an allocation, retry. */
1195 } else if (!PageHuge(p) && !is_free_buddy_page(p)) {
1196 /* We raced with put_page, retry. */
1202 if (PageHuge(p) || HWPoisonHandlable(p)) {
1206 * A page we cannot handle. Check whether we can turn
1207 * it into something we can handle.
1212 count_increased = false;
1223 static int get_hwpoison_page(struct page *p, unsigned long flags,
1228 zone_pcp_disable(page_zone(p));
1229 if (ctxt == MF_SOFT_OFFLINE)
1230 ret = get_any_page(p, flags);
1232 ret = __get_hwpoison_page(p);
1233 zone_pcp_enable(page_zone(p));
1239 * Do all that is necessary to remove user space mappings. Unmap
1240 * the pages and send SIGBUS to the processes if the data was dirty.
1242 static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
1243 int flags, struct page **hpagep)
1245 enum ttu_flags ttu = TTU_IGNORE_MLOCK;
1246 struct address_space *mapping;
1248 bool unmap_success = true;
1249 int kill = 1, forcekill;
1250 struct page *hpage = *hpagep;
1251 bool mlocked = PageMlocked(hpage);
1254 * Here we are interested only in user-mapped pages, so skip any
1255 * other types of pages.
1257 if (PageReserved(p) || PageSlab(p))
1259 if (!(PageLRU(hpage) || PageHuge(p)))
1263 * This check implies we don't kill processes if their pages
1264 * are in the swap cache early. Those are always late kills.
1266 if (!page_mapped(hpage))
1270 pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn);
1274 if (PageSwapCache(p)) {
1275 pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n",
1277 ttu |= TTU_IGNORE_HWPOISON;
1281 * Propagate the dirty bit from PTEs to struct page first, because we
1282 * need this to decide if we should kill or just drop the page.
1283 * XXX: the dirty test could be racy: set_page_dirty() may not always
1284 * be called inside page lock (it's recommended but not enforced).
1286 mapping = page_mapping(hpage);
1287 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
1288 mapping_can_writeback(mapping)) {
1289 if (page_mkclean(hpage)) {
1290 SetPageDirty(hpage);
1293 ttu |= TTU_IGNORE_HWPOISON;
1294 pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n",
1300 * First collect all the processes that have the page
1301 * mapped in dirty form. This has to be done before try_to_unmap,
1302 * because ttu takes the rmap data structures down.
1304 * Error handling: We ignore errors here because
1305 * there's nothing that can be done.
1308 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
1310 if (!PageHuge(hpage)) {
1311 unmap_success = try_to_unmap(hpage, ttu);
1313 if (!PageAnon(hpage)) {
1315 * For hugetlb pages in shared mappings, try_to_unmap
1316 * could potentially call huge_pmd_unshare. Because of
1317 * this, take semaphore in write mode here and set
1318 * TTU_RMAP_LOCKED to indicate we have taken the lock
1319 * at this higer level.
1321 mapping = hugetlb_page_mapping_lock_write(hpage);
1323 unmap_success = try_to_unmap(hpage,
1324 ttu|TTU_RMAP_LOCKED);
1325 i_mmap_unlock_write(mapping);
1327 pr_info("Memory failure: %#lx: could not lock mapping for mapped huge page\n", pfn);
1328 unmap_success = false;
1331 unmap_success = try_to_unmap(hpage, ttu);
1335 pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n",
1336 pfn, page_mapcount(hpage));
1339 * try_to_unmap() might put mlocked page in lru cache, so call
1340 * shake_page() again to ensure that it's flushed.
1343 shake_page(hpage, 0);
1346 * Now that the dirty bit has been propagated to the
1347 * struct page and all unmaps done we can decide if
1348 * killing is needed or not. Only kill when the page
1349 * was dirty or the process is not restartable,
1350 * otherwise the tokill list is merely
1351 * freed. When there was a problem unmapping earlier
1352 * use a more force-full uncatchable kill to prevent
1353 * any accesses to the poisoned memory.
1355 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
1356 kill_procs(&tokill, forcekill, !unmap_success, pfn, flags);
1358 return unmap_success;
1361 static int identify_page_state(unsigned long pfn, struct page *p,
1362 unsigned long page_flags)
1364 struct page_state *ps;
1367 * The first check uses the current page flags which may not have any
1368 * relevant information. The second check with the saved page flags is
1369 * carried out only if the first check can't determine the page status.
1371 for (ps = error_states;; ps++)
1372 if ((p->flags & ps->mask) == ps->res)
1375 page_flags |= (p->flags & (1UL << PG_dirty));
1378 for (ps = error_states;; ps++)
1379 if ((page_flags & ps->mask) == ps->res)
1381 return page_action(ps, p, pfn);
1384 static int try_to_split_thp_page(struct page *page, const char *msg)
1387 if (!PageAnon(page) || unlikely(split_huge_page(page))) {
1388 unsigned long pfn = page_to_pfn(page);
1391 if (!PageAnon(page))
1392 pr_info("%s: %#lx: non anonymous thp\n", msg, pfn);
1394 pr_info("%s: %#lx: thp split failed\n", msg, pfn);
1403 static int memory_failure_hugetlb(unsigned long pfn, int flags)
1405 struct page *p = pfn_to_page(pfn);
1406 struct page *head = compound_head(p);
1408 unsigned long page_flags;
1410 if (TestSetPageHWPoison(head)) {
1411 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1414 if (flags & MF_ACTION_REQUIRED)
1415 res = kill_accessing_process(current, page_to_pfn(head), flags);
1419 num_poisoned_pages_inc();
1421 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p, flags, 0)) {
1423 * Check "filter hit" and "race with other subpage."
1426 if (PageHWPoison(head)) {
1427 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1428 || (p != head && TestSetPageHWPoison(head))) {
1429 num_poisoned_pages_dec();
1436 if (!dissolve_free_huge_page(p) && take_page_off_buddy(p)) {
1440 action_result(pfn, MF_MSG_FREE_HUGE, res);
1441 return res == MF_RECOVERED ? 0 : -EBUSY;
1445 page_flags = head->flags;
1447 if (!PageHWPoison(head)) {
1448 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1449 num_poisoned_pages_dec();
1456 * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so
1457 * simply disable it. In order to make it work properly, we need
1459 * - conversion of a pud that maps an error hugetlb into hwpoison
1460 * entry properly works, and
1461 * - other mm code walking over page table is aware of pud-aligned
1464 if (huge_page_size(page_hstate(head)) > PMD_SIZE) {
1465 action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED);
1470 if (!hwpoison_user_mappings(p, pfn, flags, &head)) {
1471 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1476 return identify_page_state(pfn, p, page_flags);
1482 static int memory_failure_dev_pagemap(unsigned long pfn, int flags,
1483 struct dev_pagemap *pgmap)
1485 struct page *page = pfn_to_page(pfn);
1486 const bool unmap_success = true;
1487 unsigned long size = 0;
1494 if (flags & MF_COUNT_INCREASED)
1496 * Drop the extra refcount in case we come from madvise().
1500 /* device metadata space is not recoverable */
1501 if (!pgmap_pfn_valid(pgmap, pfn)) {
1507 * Prevent the inode from being freed while we are interrogating
1508 * the address_space, typically this would be handled by
1509 * lock_page(), but dax pages do not use the page lock. This
1510 * also prevents changes to the mapping of this pfn until
1511 * poison signaling is complete.
1513 cookie = dax_lock_page(page);
1517 if (hwpoison_filter(page)) {
1522 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
1524 * TODO: Handle HMM pages which may need coordination
1525 * with device-side memory.
1531 * Use this flag as an indication that the dax page has been
1532 * remapped UC to prevent speculative consumption of poison.
1534 SetPageHWPoison(page);
1537 * Unlike System-RAM there is no possibility to swap in a
1538 * different physical page at a given virtual address, so all
1539 * userspace consumption of ZONE_DEVICE memory necessitates
1540 * SIGBUS (i.e. MF_MUST_KILL)
1542 flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1543 collect_procs(page, &tokill, flags & MF_ACTION_REQUIRED);
1545 list_for_each_entry(tk, &tokill, nd)
1547 size = max(size, 1UL << tk->size_shift);
1550 * Unmap the largest mapping to avoid breaking up
1551 * device-dax mappings which are constant size. The
1552 * actual size of the mapping being torn down is
1553 * communicated in siginfo, see kill_proc()
1555 start = (page->index << PAGE_SHIFT) & ~(size - 1);
1556 unmap_mapping_range(page->mapping, start, size, 0);
1558 kill_procs(&tokill, flags & MF_MUST_KILL, !unmap_success, pfn, flags);
1561 dax_unlock_page(page, cookie);
1563 /* drop pgmap ref acquired in caller */
1564 put_dev_pagemap(pgmap);
1565 action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED);
1570 * memory_failure - Handle memory failure of a page.
1571 * @pfn: Page Number of the corrupted page
1572 * @flags: fine tune action taken
1574 * This function is called by the low level machine check code
1575 * of an architecture when it detects hardware memory corruption
1576 * of a page. It tries its best to recover, which includes
1577 * dropping pages, killing processes etc.
1579 * The function is primarily of use for corruptions that
1580 * happen outside the current execution context (e.g. when
1581 * detected by a background scrubber)
1583 * Must run in process context (e.g. a work queue) with interrupts
1584 * enabled and no spinlocks hold.
1586 int memory_failure(unsigned long pfn, int flags)
1590 struct page *orig_head;
1591 struct dev_pagemap *pgmap;
1593 unsigned long page_flags;
1595 static DEFINE_MUTEX(mf_mutex);
1597 if (!sysctl_memory_failure_recovery)
1598 panic("Memory failure on page %lx", pfn);
1600 p = pfn_to_online_page(pfn);
1602 if (pfn_valid(pfn)) {
1603 pgmap = get_dev_pagemap(pfn, NULL);
1605 return memory_failure_dev_pagemap(pfn, flags,
1608 pr_err("Memory failure: %#lx: memory outside kernel control\n",
1613 mutex_lock(&mf_mutex);
1617 res = memory_failure_hugetlb(pfn, flags);
1621 if (TestSetPageHWPoison(p)) {
1622 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1625 if (flags & MF_ACTION_REQUIRED)
1626 res = kill_accessing_process(current, pfn, flags);
1630 orig_head = hpage = compound_head(p);
1631 num_poisoned_pages_inc();
1634 * We need/can do nothing about count=0 pages.
1635 * 1) it's a free page, and therefore in safe hand:
1636 * prep_new_page() will be the gate keeper.
1637 * 2) it's part of a non-compound high order page.
1638 * Implies some kernel user: cannot stop them from
1639 * R/W the page; let's pray that the page has been
1640 * used and will be freed some time later.
1641 * In fact it's dangerous to directly bump up page count from 0,
1642 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
1644 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p, flags, 0)) {
1645 if (is_free_buddy_page(p)) {
1646 if (take_page_off_buddy(p)) {
1650 /* We lost the race, try again */
1652 ClearPageHWPoison(p);
1653 num_poisoned_pages_dec();
1659 action_result(pfn, MF_MSG_BUDDY, res);
1660 res = res == MF_RECOVERED ? 0 : -EBUSY;
1662 action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
1668 if (PageTransHuge(hpage)) {
1669 if (try_to_split_thp_page(p, "Memory Failure") < 0) {
1670 action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED);
1674 VM_BUG_ON_PAGE(!page_count(p), p);
1678 * We ignore non-LRU pages for good reasons.
1679 * - PG_locked is only well defined for LRU pages and a few others
1680 * - to avoid races with __SetPageLocked()
1681 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1682 * The check (unnecessarily) ignores LRU pages being isolated and
1683 * walked by the page reclaim code, however that's not a big loss.
1690 * The page could have changed compound pages during the locking.
1691 * If this happens just bail out.
1693 if (PageCompound(p) && compound_head(p) != orig_head) {
1694 action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
1700 * We use page flags to determine what action should be taken, but
1701 * the flags can be modified by the error containment action. One
1702 * example is an mlocked page, where PG_mlocked is cleared by
1703 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1704 * correctly, we save a copy of the page flags at this time.
1706 page_flags = p->flags;
1709 * unpoison always clear PG_hwpoison inside page lock
1711 if (!PageHWPoison(p)) {
1712 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1713 num_poisoned_pages_dec();
1718 if (hwpoison_filter(p)) {
1719 if (TestClearPageHWPoison(p))
1720 num_poisoned_pages_dec();
1727 * __munlock_pagevec may clear a writeback page's LRU flag without
1728 * page_lock. We need wait writeback completion for this page or it
1729 * may trigger vfs BUG while evict inode.
1731 if (!PageTransTail(p) && !PageLRU(p) && !PageWriteback(p))
1732 goto identify_page_state;
1735 * It's very difficult to mess with pages currently under IO
1736 * and in many cases impossible, so we just avoid it here.
1738 wait_on_page_writeback(p);
1741 * Now take care of user space mappings.
1742 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1744 if (!hwpoison_user_mappings(p, pfn, flags, &p)) {
1745 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1751 * Torn down by someone else?
1753 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1754 action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
1759 identify_page_state:
1760 res = identify_page_state(pfn, p, page_flags);
1761 mutex_unlock(&mf_mutex);
1766 mutex_unlock(&mf_mutex);
1769 EXPORT_SYMBOL_GPL(memory_failure);
1771 #define MEMORY_FAILURE_FIFO_ORDER 4
1772 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1774 struct memory_failure_entry {
1779 struct memory_failure_cpu {
1780 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1781 MEMORY_FAILURE_FIFO_SIZE);
1783 struct work_struct work;
1786 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1789 * memory_failure_queue - Schedule handling memory failure of a page.
1790 * @pfn: Page Number of the corrupted page
1791 * @flags: Flags for memory failure handling
1793 * This function is called by the low level hardware error handler
1794 * when it detects hardware memory corruption of a page. It schedules
1795 * the recovering of error page, including dropping pages, killing
1798 * The function is primarily of use for corruptions that
1799 * happen outside the current execution context (e.g. when
1800 * detected by a background scrubber)
1802 * Can run in IRQ context.
1804 void memory_failure_queue(unsigned long pfn, int flags)
1806 struct memory_failure_cpu *mf_cpu;
1807 unsigned long proc_flags;
1808 struct memory_failure_entry entry = {
1813 mf_cpu = &get_cpu_var(memory_failure_cpu);
1814 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1815 if (kfifo_put(&mf_cpu->fifo, entry))
1816 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1818 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1820 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1821 put_cpu_var(memory_failure_cpu);
1823 EXPORT_SYMBOL_GPL(memory_failure_queue);
1825 static void memory_failure_work_func(struct work_struct *work)
1827 struct memory_failure_cpu *mf_cpu;
1828 struct memory_failure_entry entry = { 0, };
1829 unsigned long proc_flags;
1832 mf_cpu = container_of(work, struct memory_failure_cpu, work);
1834 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1835 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1836 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1839 if (entry.flags & MF_SOFT_OFFLINE)
1840 soft_offline_page(entry.pfn, entry.flags);
1842 memory_failure(entry.pfn, entry.flags);
1847 * Process memory_failure work queued on the specified CPU.
1848 * Used to avoid return-to-userspace racing with the memory_failure workqueue.
1850 void memory_failure_queue_kick(int cpu)
1852 struct memory_failure_cpu *mf_cpu;
1854 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1855 cancel_work_sync(&mf_cpu->work);
1856 memory_failure_work_func(&mf_cpu->work);
1859 static int __init memory_failure_init(void)
1861 struct memory_failure_cpu *mf_cpu;
1864 for_each_possible_cpu(cpu) {
1865 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1866 spin_lock_init(&mf_cpu->lock);
1867 INIT_KFIFO(mf_cpu->fifo);
1868 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1873 core_initcall(memory_failure_init);
1875 #define unpoison_pr_info(fmt, pfn, rs) \
1877 if (__ratelimit(rs)) \
1878 pr_info(fmt, pfn); \
1882 * unpoison_memory - Unpoison a previously poisoned page
1883 * @pfn: Page number of the to be unpoisoned page
1885 * Software-unpoison a page that has been poisoned by
1886 * memory_failure() earlier.
1888 * This is only done on the software-level, so it only works
1889 * for linux injected failures, not real hardware failures
1891 * Returns 0 for success, otherwise -errno.
1893 int unpoison_memory(unsigned long pfn)
1898 unsigned long flags = 0;
1899 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
1900 DEFAULT_RATELIMIT_BURST);
1902 if (!pfn_valid(pfn))
1905 p = pfn_to_page(pfn);
1906 page = compound_head(p);
1908 if (!PageHWPoison(p)) {
1909 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
1914 if (page_count(page) > 1) {
1915 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
1920 if (page_mapped(page)) {
1921 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
1926 if (page_mapping(page)) {
1927 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
1933 * unpoison_memory() can encounter thp only when the thp is being
1934 * worked by memory_failure() and the page lock is not held yet.
1935 * In such case, we yield to memory_failure() and make unpoison fail.
1937 if (!PageHuge(page) && PageTransHuge(page)) {
1938 unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n",
1943 if (!get_hwpoison_page(p, flags, 0)) {
1944 if (TestClearPageHWPoison(p))
1945 num_poisoned_pages_dec();
1946 unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n",
1953 * This test is racy because PG_hwpoison is set outside of page lock.
1954 * That's acceptable because that won't trigger kernel panic. Instead,
1955 * the PG_hwpoison page will be caught and isolated on the entrance to
1956 * the free buddy page pool.
1958 if (TestClearPageHWPoison(page)) {
1959 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
1961 num_poisoned_pages_dec();
1967 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
1972 EXPORT_SYMBOL(unpoison_memory);
1974 static bool isolate_page(struct page *page, struct list_head *pagelist)
1976 bool isolated = false;
1977 bool lru = PageLRU(page);
1979 if (PageHuge(page)) {
1980 isolated = isolate_huge_page(page, pagelist);
1983 isolated = !isolate_lru_page(page);
1985 isolated = !isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1988 list_add(&page->lru, pagelist);
1991 if (isolated && lru)
1992 inc_node_page_state(page, NR_ISOLATED_ANON +
1993 page_is_file_lru(page));
1996 * If we succeed to isolate the page, we grabbed another refcount on
1997 * the page, so we can safely drop the one we got from get_any_pages().
1998 * If we failed to isolate the page, it means that we cannot go further
1999 * and we will return an error, so drop the reference we got from
2000 * get_any_pages() as well.
2007 * __soft_offline_page handles hugetlb-pages and non-hugetlb pages.
2008 * If the page is a non-dirty unmapped page-cache page, it simply invalidates.
2009 * If the page is mapped, it migrates the contents over.
2011 static int __soft_offline_page(struct page *page)
2014 unsigned long pfn = page_to_pfn(page);
2015 struct page *hpage = compound_head(page);
2016 char const *msg_page[] = {"page", "hugepage"};
2017 bool huge = PageHuge(page);
2018 LIST_HEAD(pagelist);
2019 struct migration_target_control mtc = {
2020 .nid = NUMA_NO_NODE,
2021 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
2025 * Check PageHWPoison again inside page lock because PageHWPoison
2026 * is set by memory_failure() outside page lock. Note that
2027 * memory_failure() also double-checks PageHWPoison inside page lock,
2028 * so there's no race between soft_offline_page() and memory_failure().
2031 if (!PageHuge(page))
2032 wait_on_page_writeback(page);
2033 if (PageHWPoison(page)) {
2036 pr_info("soft offline: %#lx page already poisoned\n", pfn);
2040 if (!PageHuge(page))
2042 * Try to invalidate first. This should work for
2043 * non dirty unmapped page cache pages.
2045 ret = invalidate_inode_page(page);
2049 * RED-PEN would be better to keep it isolated here, but we
2050 * would need to fix isolation locking first.
2053 pr_info("soft_offline: %#lx: invalidated\n", pfn);
2054 page_handle_poison(page, false, true);
2058 if (isolate_page(hpage, &pagelist)) {
2059 ret = migrate_pages(&pagelist, alloc_migration_target, NULL,
2060 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE);
2062 bool release = !huge;
2064 if (!page_handle_poison(page, huge, release))
2067 if (!list_empty(&pagelist))
2068 putback_movable_pages(&pagelist);
2070 pr_info("soft offline: %#lx: %s migration failed %d, type %lx (%pGp)\n",
2071 pfn, msg_page[huge], ret, page->flags, &page->flags);
2076 pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %lx (%pGp)\n",
2077 pfn, msg_page[huge], page_count(page), page->flags, &page->flags);
2083 static int soft_offline_in_use_page(struct page *page)
2085 struct page *hpage = compound_head(page);
2087 if (!PageHuge(page) && PageTransHuge(hpage))
2088 if (try_to_split_thp_page(page, "soft offline") < 0)
2090 return __soft_offline_page(page);
2093 static int soft_offline_free_page(struct page *page)
2097 if (!page_handle_poison(page, true, false))
2103 static void put_ref_page(struct page *page)
2110 * soft_offline_page - Soft offline a page.
2111 * @pfn: pfn to soft-offline
2112 * @flags: flags. Same as memory_failure().
2114 * Returns 0 on success, otherwise negated errno.
2116 * Soft offline a page, by migration or invalidation,
2117 * without killing anything. This is for the case when
2118 * a page is not corrupted yet (so it's still valid to access),
2119 * but has had a number of corrected errors and is better taken
2122 * The actual policy on when to do that is maintained by
2125 * This should never impact any application or cause data loss,
2126 * however it might take some time.
2128 * This is not a 100% solution for all memory, but tries to be
2129 * ``good enough'' for the majority of memory.
2131 int soft_offline_page(unsigned long pfn, int flags)
2134 bool try_again = true;
2135 struct page *page, *ref_page = NULL;
2137 WARN_ON_ONCE(!pfn_valid(pfn) && (flags & MF_COUNT_INCREASED));
2139 if (!pfn_valid(pfn))
2141 if (flags & MF_COUNT_INCREASED)
2142 ref_page = pfn_to_page(pfn);
2144 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
2145 page = pfn_to_online_page(pfn);
2147 put_ref_page(ref_page);
2151 if (PageHWPoison(page)) {
2152 pr_info("%s: %#lx page already poisoned\n", __func__, pfn);
2153 put_ref_page(ref_page);
2159 ret = get_hwpoison_page(page, flags, MF_SOFT_OFFLINE);
2163 ret = soft_offline_in_use_page(page);
2164 } else if (ret == 0) {
2165 if (soft_offline_free_page(page) && try_again) {
2169 } else if (ret == -EIO) {
2170 pr_info("%s: %#lx: unknown page type: %lx (%pGp)\n",
2171 __func__, pfn, page->flags, &page->flags);