1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2008, 2009 Intel Corporation
4 * Authors: Andi Kleen, Fengguang Wu
6 * High level machine check handler. Handles pages reported by the
7 * hardware as being corrupted usually due to a multi-bit ECC memory or cache
10 * In addition there is a "soft offline" entry point that allows stop using
11 * not-yet-corrupted-by-suspicious pages without killing anything.
13 * Handles page cache pages in various states. The tricky part
14 * here is that we can access any page asynchronously in respect to
15 * other VM users, because memory failures could happen anytime and
16 * anywhere. This could violate some of their assumptions. This is why
17 * this code has to be extremely careful. Generally it tries to use
18 * normal locking rules, as in get the standard locks, even if that means
19 * the error handling takes potentially a long time.
21 * It can be very tempting to add handling for obscure cases here.
22 * In general any code for handling new cases should only be added iff:
23 * - You know how to test it.
24 * - You have a test that can be added to mce-test
25 * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/
26 * - The case actually shows up as a frequent (top 10) page state in
27 * tools/vm/page-types when running a real workload.
29 * There are several operations here with exponential complexity because
30 * of unsuitable VM data structures. For example the operation to map back
31 * from RMAP chains to processes has to walk the complete process list and
32 * has non linear complexity with the number. But since memory corruptions
33 * are rare we hope to get away with this. This avoids impacting the core
36 #include <linux/kernel.h>
38 #include <linux/page-flags.h>
39 #include <linux/kernel-page-flags.h>
40 #include <linux/sched/signal.h>
41 #include <linux/sched/task.h>
42 #include <linux/dax.h>
43 #include <linux/ksm.h>
44 #include <linux/rmap.h>
45 #include <linux/export.h>
46 #include <linux/pagemap.h>
47 #include <linux/swap.h>
48 #include <linux/backing-dev.h>
49 #include <linux/migrate.h>
50 #include <linux/suspend.h>
51 #include <linux/slab.h>
52 #include <linux/swapops.h>
53 #include <linux/hugetlb.h>
54 #include <linux/memory_hotplug.h>
55 #include <linux/mm_inline.h>
56 #include <linux/memremap.h>
57 #include <linux/kfifo.h>
58 #include <linux/ratelimit.h>
59 #include <linux/page-isolation.h>
60 #include <linux/pagewalk.h>
61 #include <linux/shmem_fs.h>
63 #include "ras/ras_event.h"
65 int sysctl_memory_failure_early_kill __read_mostly = 0;
67 int sysctl_memory_failure_recovery __read_mostly = 1;
69 atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
71 static bool __page_handle_poison(struct page *page)
75 zone_pcp_disable(page_zone(page));
76 ret = dissolve_free_huge_page(page);
78 ret = take_page_off_buddy(page);
79 zone_pcp_enable(page_zone(page));
84 static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release)
86 if (hugepage_or_freepage) {
88 * Doing this check for free pages is also fine since dissolve_free_huge_page
89 * returns 0 for non-hugetlb pages as well.
91 if (!__page_handle_poison(page))
93 * We could fail to take off the target page from buddy
94 * for example due to racy page allocation, but that's
95 * acceptable because soft-offlined page is not broken
96 * and if someone really want to use it, they should
102 SetPageHWPoison(page);
106 num_poisoned_pages_inc();
111 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
113 u32 hwpoison_filter_enable = 0;
114 u32 hwpoison_filter_dev_major = ~0U;
115 u32 hwpoison_filter_dev_minor = ~0U;
116 u64 hwpoison_filter_flags_mask;
117 u64 hwpoison_filter_flags_value;
118 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
119 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
120 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
121 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
122 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
124 static int hwpoison_filter_dev(struct page *p)
126 struct address_space *mapping;
129 if (hwpoison_filter_dev_major == ~0U &&
130 hwpoison_filter_dev_minor == ~0U)
133 mapping = page_mapping(p);
134 if (mapping == NULL || mapping->host == NULL)
137 dev = mapping->host->i_sb->s_dev;
138 if (hwpoison_filter_dev_major != ~0U &&
139 hwpoison_filter_dev_major != MAJOR(dev))
141 if (hwpoison_filter_dev_minor != ~0U &&
142 hwpoison_filter_dev_minor != MINOR(dev))
148 static int hwpoison_filter_flags(struct page *p)
150 if (!hwpoison_filter_flags_mask)
153 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
154 hwpoison_filter_flags_value)
161 * This allows stress tests to limit test scope to a collection of tasks
162 * by putting them under some memcg. This prevents killing unrelated/important
163 * processes such as /sbin/init. Note that the target task may share clean
164 * pages with init (eg. libc text), which is harmless. If the target task
165 * share _dirty_ pages with another task B, the test scheme must make sure B
166 * is also included in the memcg. At last, due to race conditions this filter
167 * can only guarantee that the page either belongs to the memcg tasks, or is
171 u64 hwpoison_filter_memcg;
172 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
173 static int hwpoison_filter_task(struct page *p)
175 if (!hwpoison_filter_memcg)
178 if (page_cgroup_ino(p) != hwpoison_filter_memcg)
184 static int hwpoison_filter_task(struct page *p) { return 0; }
187 int hwpoison_filter(struct page *p)
189 if (!hwpoison_filter_enable)
192 if (hwpoison_filter_dev(p))
195 if (hwpoison_filter_flags(p))
198 if (hwpoison_filter_task(p))
204 int hwpoison_filter(struct page *p)
210 EXPORT_SYMBOL_GPL(hwpoison_filter);
213 * Kill all processes that have a poisoned page mapped and then isolate
217 * Find all processes having the page mapped and kill them.
218 * But we keep a page reference around so that the page is not
219 * actually freed yet.
220 * Then stash the page away
222 * There's no convenient way to get back to mapped processes
223 * from the VMAs. So do a brute-force search over all
226 * Remember that machine checks are not common (or rather
227 * if they are common you have other problems), so this shouldn't
228 * be a performance issue.
230 * Also there are some races possible while we get from the
231 * error detection to actually handle it.
236 struct task_struct *tsk;
242 * Send all the processes who have the page mapped a signal.
243 * ``action optional'' if they are not immediately affected by the error
244 * ``action required'' if error happened in current execution context
246 static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
248 struct task_struct *t = tk->tsk;
249 short addr_lsb = tk->size_shift;
252 pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
253 pfn, t->comm, t->pid);
255 if ((flags & MF_ACTION_REQUIRED) && (t == current))
256 ret = force_sig_mceerr(BUS_MCEERR_AR,
257 (void __user *)tk->addr, addr_lsb);
260 * Signal other processes sharing the page if they have
262 * Don't use force here, it's convenient if the signal
263 * can be temporarily blocked.
264 * This could cause a loop when the user sets SIGBUS
265 * to SIG_IGN, but hopefully no one will do that?
267 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
268 addr_lsb, t); /* synchronous? */
270 pr_info("Memory failure: Error sending signal to %s:%d: %d\n",
271 t->comm, t->pid, ret);
276 * Unknown page type encountered. Try to check whether it can turn PageLRU by
279 void shake_page(struct page *p)
286 if (PageLRU(p) || is_free_buddy_page(p))
291 * TODO: Could shrink slab caches here if a lightweight range-based
292 * shrinker will be available.
295 EXPORT_SYMBOL_GPL(shake_page);
297 static unsigned long dev_pagemap_mapping_shift(struct page *page,
298 struct vm_area_struct *vma)
300 unsigned long address = vma_address(page, vma);
301 unsigned long ret = 0;
308 VM_BUG_ON_VMA(address == -EFAULT, vma);
309 pgd = pgd_offset(vma->vm_mm, address);
310 if (!pgd_present(*pgd))
312 p4d = p4d_offset(pgd, address);
313 if (!p4d_present(*p4d))
315 pud = pud_offset(p4d, address);
316 if (!pud_present(*pud))
318 if (pud_devmap(*pud))
320 pmd = pmd_offset(pud, address);
321 if (!pmd_present(*pmd))
323 if (pmd_devmap(*pmd))
325 pte = pte_offset_map(pmd, address);
326 if (pte_present(*pte) && pte_devmap(*pte))
333 * Failure handling: if we can't find or can't kill a process there's
334 * not much we can do. We just print a message and ignore otherwise.
338 * Schedule a process for later kill.
339 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
341 static void add_to_kill(struct task_struct *tsk, struct page *p,
342 struct vm_area_struct *vma,
343 struct list_head *to_kill)
347 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
349 pr_err("Memory failure: Out of memory while machine check handling\n");
353 tk->addr = page_address_in_vma(p, vma);
354 if (is_zone_device_page(p))
355 tk->size_shift = dev_pagemap_mapping_shift(p, vma);
357 tk->size_shift = page_shift(compound_head(p));
360 * Send SIGKILL if "tk->addr == -EFAULT". Also, as
361 * "tk->size_shift" is always non-zero for !is_zone_device_page(),
362 * so "tk->size_shift == 0" effectively checks no mapping on
363 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times
364 * to a process' address space, it's possible not all N VMAs
365 * contain mappings for the page, but at least one VMA does.
366 * Only deliver SIGBUS with payload derived from the VMA that
367 * has a mapping for the page.
369 if (tk->addr == -EFAULT) {
370 pr_info("Memory failure: Unable to find user space address %lx in %s\n",
371 page_to_pfn(p), tsk->comm);
372 } else if (tk->size_shift == 0) {
377 get_task_struct(tsk);
379 list_add_tail(&tk->nd, to_kill);
383 * Kill the processes that have been collected earlier.
385 * Only do anything when FORCEKILL is set, otherwise just free the
386 * list (this is used for clean pages which do not need killing)
387 * Also when FAIL is set do a force kill because something went
390 static void kill_procs(struct list_head *to_kill, int forcekill, bool fail,
391 unsigned long pfn, int flags)
393 struct to_kill *tk, *next;
395 list_for_each_entry_safe (tk, next, to_kill, nd) {
398 * In case something went wrong with munmapping
399 * make sure the process doesn't catch the
400 * signal and then access the memory. Just kill it.
402 if (fail || tk->addr == -EFAULT) {
403 pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
404 pfn, tk->tsk->comm, tk->tsk->pid);
405 do_send_sig_info(SIGKILL, SEND_SIG_PRIV,
406 tk->tsk, PIDTYPE_PID);
410 * In theory the process could have mapped
411 * something else on the address in-between. We could
412 * check for that, but we need to tell the
415 else if (kill_proc(tk, pfn, flags) < 0)
416 pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n",
417 pfn, tk->tsk->comm, tk->tsk->pid);
419 put_task_struct(tk->tsk);
425 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
426 * on behalf of the thread group. Return task_struct of the (first found)
427 * dedicated thread if found, and return NULL otherwise.
429 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
430 * have to call rcu_read_lock/unlock() in this function.
432 static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
434 struct task_struct *t;
436 for_each_thread(tsk, t) {
437 if (t->flags & PF_MCE_PROCESS) {
438 if (t->flags & PF_MCE_EARLY)
441 if (sysctl_memory_failure_early_kill)
449 * Determine whether a given process is "early kill" process which expects
450 * to be signaled when some page under the process is hwpoisoned.
451 * Return task_struct of the dedicated thread (main thread unless explicitly
452 * specified) if the process is "early kill" and otherwise returns NULL.
454 * Note that the above is true for Action Optional case. For Action Required
455 * case, it's only meaningful to the current thread which need to be signaled
456 * with SIGBUS, this error is Action Optional for other non current
457 * processes sharing the same error page,if the process is "early kill", the
458 * task_struct of the dedicated thread will also be returned.
460 static struct task_struct *task_early_kill(struct task_struct *tsk,
466 * Comparing ->mm here because current task might represent
467 * a subthread, while tsk always points to the main thread.
469 if (force_early && tsk->mm == current->mm)
472 return find_early_kill_thread(tsk);
476 * Collect processes when the error hit an anonymous page.
478 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
481 struct folio *folio = page_folio(page);
482 struct vm_area_struct *vma;
483 struct task_struct *tsk;
487 av = folio_lock_anon_vma_read(folio);
488 if (av == NULL) /* Not actually mapped anymore */
491 pgoff = page_to_pgoff(page);
492 read_lock(&tasklist_lock);
493 for_each_process (tsk) {
494 struct anon_vma_chain *vmac;
495 struct task_struct *t = task_early_kill(tsk, force_early);
499 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
502 if (!page_mapped_in_vma(page, vma))
504 if (vma->vm_mm == t->mm)
505 add_to_kill(t, page, vma, to_kill);
508 read_unlock(&tasklist_lock);
509 page_unlock_anon_vma_read(av);
513 * Collect processes when the error hit a file mapped page.
515 static void collect_procs_file(struct page *page, struct list_head *to_kill,
518 struct vm_area_struct *vma;
519 struct task_struct *tsk;
520 struct address_space *mapping = page->mapping;
523 i_mmap_lock_read(mapping);
524 read_lock(&tasklist_lock);
525 pgoff = page_to_pgoff(page);
526 for_each_process(tsk) {
527 struct task_struct *t = task_early_kill(tsk, force_early);
531 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
534 * Send early kill signal to tasks where a vma covers
535 * the page but the corrupted page is not necessarily
536 * mapped it in its pte.
537 * Assume applications who requested early kill want
538 * to be informed of all such data corruptions.
540 if (vma->vm_mm == t->mm)
541 add_to_kill(t, page, vma, to_kill);
544 read_unlock(&tasklist_lock);
545 i_mmap_unlock_read(mapping);
549 * Collect the processes who have the corrupted page mapped to kill.
551 static void collect_procs(struct page *page, struct list_head *tokill,
558 collect_procs_anon(page, tokill, force_early);
560 collect_procs_file(page, tokill, force_early);
569 static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift)
572 tk->size_shift = shift;
575 static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift,
576 unsigned long poisoned_pfn, struct to_kill *tk)
578 unsigned long pfn = 0;
580 if (pte_present(pte)) {
583 swp_entry_t swp = pte_to_swp_entry(pte);
585 if (is_hwpoison_entry(swp))
586 pfn = hwpoison_entry_to_pfn(swp);
589 if (!pfn || pfn != poisoned_pfn)
592 set_to_kill(tk, addr, shift);
596 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
597 static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
598 struct hwp_walk *hwp)
602 unsigned long hwpoison_vaddr;
604 if (!pmd_present(pmd))
607 if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) {
608 hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT);
609 set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT);
615 static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
616 struct hwp_walk *hwp)
622 static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr,
623 unsigned long end, struct mm_walk *walk)
625 struct hwp_walk *hwp = walk->private;
627 pte_t *ptep, *mapped_pte;
630 ptl = pmd_trans_huge_lock(pmdp, walk->vma);
632 ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp);
637 if (pmd_trans_unstable(pmdp))
640 mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp,
642 for (; addr != end; ptep++, addr += PAGE_SIZE) {
643 ret = check_hwpoisoned_entry(*ptep, addr, PAGE_SHIFT,
648 pte_unmap_unlock(mapped_pte, ptl);
654 #ifdef CONFIG_HUGETLB_PAGE
655 static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask,
656 unsigned long addr, unsigned long end,
657 struct mm_walk *walk)
659 struct hwp_walk *hwp = walk->private;
660 pte_t pte = huge_ptep_get(ptep);
661 struct hstate *h = hstate_vma(walk->vma);
663 return check_hwpoisoned_entry(pte, addr, huge_page_shift(h),
667 #define hwpoison_hugetlb_range NULL
670 static const struct mm_walk_ops hwp_walk_ops = {
671 .pmd_entry = hwpoison_pte_range,
672 .hugetlb_entry = hwpoison_hugetlb_range,
676 * Sends SIGBUS to the current process with error info.
678 * This function is intended to handle "Action Required" MCEs on already
679 * hardware poisoned pages. They could happen, for example, when
680 * memory_failure() failed to unmap the error page at the first call, or
681 * when multiple local machine checks happened on different CPUs.
683 * MCE handler currently has no easy access to the error virtual address,
684 * so this function walks page table to find it. The returned virtual address
685 * is proper in most cases, but it could be wrong when the application
686 * process has multiple entries mapping the error page.
688 static int kill_accessing_process(struct task_struct *p, unsigned long pfn,
692 struct hwp_walk priv = {
697 mmap_read_lock(p->mm);
698 ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwp_walk_ops,
700 if (ret == 1 && priv.tk.addr)
701 kill_proc(&priv.tk, pfn, flags);
704 mmap_read_unlock(p->mm);
705 return ret > 0 ? -EHWPOISON : -EFAULT;
708 static const char *action_name[] = {
709 [MF_IGNORED] = "Ignored",
710 [MF_FAILED] = "Failed",
711 [MF_DELAYED] = "Delayed",
712 [MF_RECOVERED] = "Recovered",
715 static const char * const action_page_types[] = {
716 [MF_MSG_KERNEL] = "reserved kernel page",
717 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
718 [MF_MSG_SLAB] = "kernel slab page",
719 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
720 [MF_MSG_HUGE] = "huge page",
721 [MF_MSG_FREE_HUGE] = "free huge page",
722 [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page",
723 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
724 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
725 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
726 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
727 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
728 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
729 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
730 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
731 [MF_MSG_CLEAN_LRU] = "clean LRU page",
732 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
733 [MF_MSG_BUDDY] = "free buddy page",
734 [MF_MSG_DAX] = "dax page",
735 [MF_MSG_UNSPLIT_THP] = "unsplit thp",
736 [MF_MSG_UNKNOWN] = "unknown page",
740 * XXX: It is possible that a page is isolated from LRU cache,
741 * and then kept in swap cache or failed to remove from page cache.
742 * The page count will stop it from being freed by unpoison.
743 * Stress tests should be aware of this memory leak problem.
745 static int delete_from_lru_cache(struct page *p)
747 if (!isolate_lru_page(p)) {
749 * Clear sensible page flags, so that the buddy system won't
750 * complain when the page is unpoison-and-freed.
753 ClearPageUnevictable(p);
756 * Poisoned page might never drop its ref count to 0 so we have
757 * to uncharge it manually from its memcg.
759 mem_cgroup_uncharge(page_folio(p));
762 * drop the page count elevated by isolate_lru_page()
770 static int truncate_error_page(struct page *p, unsigned long pfn,
771 struct address_space *mapping)
775 if (mapping->a_ops->error_remove_page) {
776 int err = mapping->a_ops->error_remove_page(mapping, p);
779 pr_info("Memory failure: %#lx: Failed to punch page: %d\n",
781 } else if (page_has_private(p) &&
782 !try_to_release_page(p, GFP_NOIO)) {
783 pr_info("Memory failure: %#lx: failed to release buffers\n",
790 * If the file system doesn't support it just invalidate
791 * This fails on dirty or anything with private pages
793 if (invalidate_inode_page(p))
796 pr_info("Memory failure: %#lx: Failed to invalidate\n",
806 enum mf_action_page_type type;
808 /* Callback ->action() has to unlock the relevant page inside it. */
809 int (*action)(struct page_state *ps, struct page *p);
813 * Return true if page is still referenced by others, otherwise return
816 * The extra_pins is true when one extra refcount is expected.
818 static bool has_extra_refcount(struct page_state *ps, struct page *p,
821 int count = page_count(p) - 1;
827 pr_err("Memory failure: %#lx: %s still referenced by %d users\n",
828 page_to_pfn(p), action_page_types[ps->type], count);
836 * Error hit kernel page.
837 * Do nothing, try to be lucky and not touch this instead. For a few cases we
838 * could be more sophisticated.
840 static int me_kernel(struct page_state *ps, struct page *p)
847 * Page in unknown state. Do nothing.
849 static int me_unknown(struct page_state *ps, struct page *p)
851 pr_err("Memory failure: %#lx: Unknown page state\n", page_to_pfn(p));
857 * Clean (or cleaned) page cache page.
859 static int me_pagecache_clean(struct page_state *ps, struct page *p)
862 struct address_space *mapping;
865 delete_from_lru_cache(p);
868 * For anonymous pages we're done the only reference left
869 * should be the one m_f() holds.
877 * Now truncate the page in the page cache. This is really
878 * more like a "temporary hole punch"
879 * Don't do this for block devices when someone else
880 * has a reference, because it could be file system metadata
881 * and that's not safe to truncate.
883 mapping = page_mapping(p);
886 * Page has been teared down in the meanwhile
893 * The shmem page is kept in page cache instead of truncating
894 * so is expected to have an extra refcount after error-handling.
896 extra_pins = shmem_mapping(mapping);
899 * Truncation is a bit tricky. Enable it per file system for now.
901 * Open: to take i_rwsem or not for this? Right now we don't.
903 ret = truncate_error_page(p, page_to_pfn(p), mapping);
904 if (has_extra_refcount(ps, p, extra_pins))
914 * Dirty pagecache page
915 * Issues: when the error hit a hole page the error is not properly
918 static int me_pagecache_dirty(struct page_state *ps, struct page *p)
920 struct address_space *mapping = page_mapping(p);
923 /* TBD: print more information about the file. */
926 * IO error will be reported by write(), fsync(), etc.
927 * who check the mapping.
928 * This way the application knows that something went
929 * wrong with its dirty file data.
931 * There's one open issue:
933 * The EIO will be only reported on the next IO
934 * operation and then cleared through the IO map.
935 * Normally Linux has two mechanisms to pass IO error
936 * first through the AS_EIO flag in the address space
937 * and then through the PageError flag in the page.
938 * Since we drop pages on memory failure handling the
939 * only mechanism open to use is through AS_AIO.
941 * This has the disadvantage that it gets cleared on
942 * the first operation that returns an error, while
943 * the PageError bit is more sticky and only cleared
944 * when the page is reread or dropped. If an
945 * application assumes it will always get error on
946 * fsync, but does other operations on the fd before
947 * and the page is dropped between then the error
948 * will not be properly reported.
950 * This can already happen even without hwpoisoned
951 * pages: first on metadata IO errors (which only
952 * report through AS_EIO) or when the page is dropped
955 * So right now we assume that the application DTRT on
956 * the first EIO, but we're not worse than other parts
959 mapping_set_error(mapping, -EIO);
962 return me_pagecache_clean(ps, p);
966 * Clean and dirty swap cache.
968 * Dirty swap cache page is tricky to handle. The page could live both in page
969 * cache and swap cache(ie. page is freshly swapped in). So it could be
970 * referenced concurrently by 2 types of PTEs:
971 * normal PTEs and swap PTEs. We try to handle them consistently by calling
972 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
974 * - clear dirty bit to prevent IO
976 * - but keep in the swap cache, so that when we return to it on
977 * a later page fault, we know the application is accessing
978 * corrupted data and shall be killed (we installed simple
979 * interception code in do_swap_page to catch it).
981 * Clean swap cache pages can be directly isolated. A later page fault will
982 * bring in the known good data from disk.
984 static int me_swapcache_dirty(struct page_state *ps, struct page *p)
987 bool extra_pins = false;
990 /* Trigger EIO in shmem: */
991 ClearPageUptodate(p);
993 ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED;
996 if (ret == MF_DELAYED)
999 if (has_extra_refcount(ps, p, extra_pins))
1005 static int me_swapcache_clean(struct page_state *ps, struct page *p)
1009 delete_from_swap_cache(p);
1011 ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED;
1014 if (has_extra_refcount(ps, p, false))
1021 * Huge pages. Needs work.
1023 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
1024 * To narrow down kill region to one page, we need to break up pmd.
1026 static int me_huge_page(struct page_state *ps, struct page *p)
1029 struct page *hpage = compound_head(p);
1030 struct address_space *mapping;
1032 if (!PageHuge(hpage))
1035 mapping = page_mapping(hpage);
1037 res = truncate_error_page(hpage, page_to_pfn(p), mapping);
1043 * migration entry prevents later access on error hugepage,
1044 * so we can free and dissolve it into buddy to save healthy
1048 if (__page_handle_poison(p)) {
1054 if (has_extra_refcount(ps, p, false))
1061 * Various page states we can handle.
1063 * A page state is defined by its current page->flags bits.
1064 * The table matches them in order and calls the right handler.
1066 * This is quite tricky because we can access page at any time
1067 * in its live cycle, so all accesses have to be extremely careful.
1069 * This is not complete. More states could be added.
1070 * For any missing state don't attempt recovery.
1073 #define dirty (1UL << PG_dirty)
1074 #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked))
1075 #define unevict (1UL << PG_unevictable)
1076 #define mlock (1UL << PG_mlocked)
1077 #define lru (1UL << PG_lru)
1078 #define head (1UL << PG_head)
1079 #define slab (1UL << PG_slab)
1080 #define reserved (1UL << PG_reserved)
1082 static struct page_state error_states[] = {
1083 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
1085 * free pages are specially detected outside this table:
1086 * PG_buddy pages only make a small fraction of all free pages.
1090 * Could in theory check if slab page is free or if we can drop
1091 * currently unused objects without touching them. But just
1092 * treat it as standard kernel for now.
1094 { slab, slab, MF_MSG_SLAB, me_kernel },
1096 { head, head, MF_MSG_HUGE, me_huge_page },
1098 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
1099 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
1101 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
1102 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
1104 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
1105 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
1107 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
1108 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
1111 * Catchall entry: must be at end.
1113 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
1126 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
1127 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
1129 static void action_result(unsigned long pfn, enum mf_action_page_type type,
1130 enum mf_result result)
1132 trace_memory_failure_event(pfn, type, result);
1134 pr_err("Memory failure: %#lx: recovery action for %s: %s\n",
1135 pfn, action_page_types[type], action_name[result]);
1138 static int page_action(struct page_state *ps, struct page *p,
1143 /* page p should be unlocked after returning from ps->action(). */
1144 result = ps->action(ps, p);
1146 action_result(pfn, ps->type, result);
1148 /* Could do more checks here if page looks ok */
1150 * Could adjust zone counters here to correct for the missing page.
1153 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
1156 static inline bool PageHWPoisonTakenOff(struct page *page)
1158 return PageHWPoison(page) && page_private(page) == MAGIC_HWPOISON;
1161 void SetPageHWPoisonTakenOff(struct page *page)
1163 set_page_private(page, MAGIC_HWPOISON);
1166 void ClearPageHWPoisonTakenOff(struct page *page)
1168 if (PageHWPoison(page))
1169 set_page_private(page, 0);
1173 * Return true if a page type of a given page is supported by hwpoison
1174 * mechanism (while handling could fail), otherwise false. This function
1175 * does not return true for hugetlb or device memory pages, so it's assumed
1176 * to be called only in the context where we never have such pages.
1178 static inline bool HWPoisonHandlable(struct page *page, unsigned long flags)
1180 /* Soft offline could migrate non-LRU movable pages */
1181 if ((flags & MF_SOFT_OFFLINE) && __PageMovable(page))
1184 return PageLRU(page) || is_free_buddy_page(page);
1187 static int __get_hwpoison_page(struct page *page, unsigned long flags)
1189 struct page *head = compound_head(page);
1191 bool hugetlb = false;
1193 ret = get_hwpoison_huge_page(head, &hugetlb);
1198 * This check prevents from calling get_hwpoison_unless_zero()
1199 * for any unsupported type of page in order to reduce the risk of
1200 * unexpected races caused by taking a page refcount.
1202 if (!HWPoisonHandlable(head, flags))
1205 if (get_page_unless_zero(head)) {
1206 if (head == compound_head(page))
1209 pr_info("Memory failure: %#lx cannot catch tail\n",
1217 static int get_any_page(struct page *p, unsigned long flags)
1219 int ret = 0, pass = 0;
1220 bool count_increased = false;
1222 if (flags & MF_COUNT_INCREASED)
1223 count_increased = true;
1226 if (!count_increased) {
1227 ret = __get_hwpoison_page(p, flags);
1229 if (page_count(p)) {
1230 /* We raced with an allocation, retry. */
1234 } else if (!PageHuge(p) && !is_free_buddy_page(p)) {
1235 /* We raced with put_page, retry. */
1241 } else if (ret == -EBUSY) {
1243 * We raced with (possibly temporary) unhandlable
1255 if (PageHuge(p) || HWPoisonHandlable(p, flags)) {
1259 * A page we cannot handle. Check whether we can turn
1260 * it into something we can handle.
1265 count_increased = false;
1273 dump_page(p, "hwpoison: unhandlable page");
1278 static int __get_unpoison_page(struct page *page)
1280 struct page *head = compound_head(page);
1282 bool hugetlb = false;
1284 ret = get_hwpoison_huge_page(head, &hugetlb);
1289 * PageHWPoisonTakenOff pages are not only marked as PG_hwpoison,
1290 * but also isolated from buddy freelist, so need to identify the
1291 * state and have to cancel both operations to unpoison.
1293 if (PageHWPoisonTakenOff(page))
1296 return get_page_unless_zero(page) ? 1 : 0;
1300 * get_hwpoison_page() - Get refcount for memory error handling
1301 * @p: Raw error page (hit by memory error)
1302 * @flags: Flags controlling behavior of error handling
1304 * get_hwpoison_page() takes a page refcount of an error page to handle memory
1305 * error on it, after checking that the error page is in a well-defined state
1306 * (defined as a page-type we can successfully handle the memory error on it,
1307 * such as LRU page and hugetlb page).
1309 * Memory error handling could be triggered at any time on any type of page,
1310 * so it's prone to race with typical memory management lifecycle (like
1311 * allocation and free). So to avoid such races, get_hwpoison_page() takes
1312 * extra care for the error page's state (as done in __get_hwpoison_page()),
1313 * and has some retry logic in get_any_page().
1315 * When called from unpoison_memory(), the caller should already ensure that
1316 * the given page has PG_hwpoison. So it's never reused for other page
1317 * allocations, and __get_unpoison_page() never races with them.
1319 * Return: 0 on failure,
1320 * 1 on success for in-use pages in a well-defined state,
1321 * -EIO for pages on which we can not handle memory errors,
1322 * -EBUSY when get_hwpoison_page() has raced with page lifecycle
1323 * operations like allocation and free,
1324 * -EHWPOISON when the page is hwpoisoned and taken off from buddy.
1326 static int get_hwpoison_page(struct page *p, unsigned long flags)
1330 zone_pcp_disable(page_zone(p));
1331 if (flags & MF_UNPOISON)
1332 ret = __get_unpoison_page(p);
1334 ret = get_any_page(p, flags);
1335 zone_pcp_enable(page_zone(p));
1341 * Do all that is necessary to remove user space mappings. Unmap
1342 * the pages and send SIGBUS to the processes if the data was dirty.
1344 static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
1345 int flags, struct page *hpage)
1347 struct folio *folio = page_folio(hpage);
1348 enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC;
1349 struct address_space *mapping;
1352 int kill = 1, forcekill;
1353 bool mlocked = PageMlocked(hpage);
1356 * Here we are interested only in user-mapped pages, so skip any
1357 * other types of pages.
1359 if (PageReserved(p) || PageSlab(p))
1361 if (!(PageLRU(hpage) || PageHuge(p)))
1365 * This check implies we don't kill processes if their pages
1366 * are in the swap cache early. Those are always late kills.
1368 if (!page_mapped(hpage))
1372 pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn);
1376 if (PageSwapCache(p)) {
1377 pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n",
1379 ttu |= TTU_IGNORE_HWPOISON;
1383 * Propagate the dirty bit from PTEs to struct page first, because we
1384 * need this to decide if we should kill or just drop the page.
1385 * XXX: the dirty test could be racy: set_page_dirty() may not always
1386 * be called inside page lock (it's recommended but not enforced).
1388 mapping = page_mapping(hpage);
1389 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
1390 mapping_can_writeback(mapping)) {
1391 if (page_mkclean(hpage)) {
1392 SetPageDirty(hpage);
1395 ttu |= TTU_IGNORE_HWPOISON;
1396 pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n",
1402 * First collect all the processes that have the page
1403 * mapped in dirty form. This has to be done before try_to_unmap,
1404 * because ttu takes the rmap data structures down.
1406 * Error handling: We ignore errors here because
1407 * there's nothing that can be done.
1410 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
1412 if (PageHuge(hpage) && !PageAnon(hpage)) {
1414 * For hugetlb pages in shared mappings, try_to_unmap
1415 * could potentially call huge_pmd_unshare. Because of
1416 * this, take semaphore in write mode here and set
1417 * TTU_RMAP_LOCKED to indicate we have taken the lock
1418 * at this higher level.
1420 mapping = hugetlb_page_mapping_lock_write(hpage);
1422 try_to_unmap(folio, ttu|TTU_RMAP_LOCKED);
1423 i_mmap_unlock_write(mapping);
1425 pr_info("Memory failure: %#lx: could not lock mapping for mapped huge page\n", pfn);
1427 try_to_unmap(folio, ttu);
1430 unmap_success = !page_mapped(hpage);
1432 pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n",
1433 pfn, page_mapcount(hpage));
1436 * try_to_unmap() might put mlocked page in lru cache, so call
1437 * shake_page() again to ensure that it's flushed.
1443 * Now that the dirty bit has been propagated to the
1444 * struct page and all unmaps done we can decide if
1445 * killing is needed or not. Only kill when the page
1446 * was dirty or the process is not restartable,
1447 * otherwise the tokill list is merely
1448 * freed. When there was a problem unmapping earlier
1449 * use a more force-full uncatchable kill to prevent
1450 * any accesses to the poisoned memory.
1452 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
1453 kill_procs(&tokill, forcekill, !unmap_success, pfn, flags);
1455 return unmap_success;
1458 static int identify_page_state(unsigned long pfn, struct page *p,
1459 unsigned long page_flags)
1461 struct page_state *ps;
1464 * The first check uses the current page flags which may not have any
1465 * relevant information. The second check with the saved page flags is
1466 * carried out only if the first check can't determine the page status.
1468 for (ps = error_states;; ps++)
1469 if ((p->flags & ps->mask) == ps->res)
1472 page_flags |= (p->flags & (1UL << PG_dirty));
1475 for (ps = error_states;; ps++)
1476 if ((page_flags & ps->mask) == ps->res)
1478 return page_action(ps, p, pfn);
1481 static int try_to_split_thp_page(struct page *page, const char *msg)
1484 if (unlikely(split_huge_page(page))) {
1485 unsigned long pfn = page_to_pfn(page);
1488 pr_info("%s: %#lx: thp split failed\n", msg, pfn);
1498 * Called from hugetlb code with hugetlb_lock held.
1502 * 1 - in-use hugepage
1503 * 2 - not a hugepage
1504 * -EBUSY - the hugepage is busy (try to retry)
1505 * -EHWPOISON - the hugepage is already hwpoisoned
1507 int __get_huge_page_for_hwpoison(unsigned long pfn, int flags)
1509 struct page *page = pfn_to_page(pfn);
1510 struct page *head = compound_head(page);
1511 int ret = 2; /* fallback to normal page handling */
1512 bool count_increased = false;
1514 if (!PageHeadHuge(head))
1517 if (flags & MF_COUNT_INCREASED) {
1519 count_increased = true;
1520 } else if (HPageFreed(head) || HPageMigratable(head)) {
1521 ret = get_page_unless_zero(head);
1523 count_increased = true;
1529 if (TestSetPageHWPoison(head)) {
1536 if (count_increased)
1541 #ifdef CONFIG_HUGETLB_PAGE
1543 * Taking refcount of hugetlb pages needs extra care about race conditions
1544 * with basic operations like hugepage allocation/free/demotion.
1545 * So some of prechecks for hwpoison (pinning, and testing/setting
1546 * PageHWPoison) should be done in single hugetlb_lock range.
1548 static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb)
1551 struct page *p = pfn_to_page(pfn);
1553 unsigned long page_flags;
1558 res = get_huge_page_for_hwpoison(pfn, flags);
1559 if (res == 2) { /* fallback to normal page handling */
1562 } else if (res == -EHWPOISON) {
1563 pr_err("Memory failure: %#lx: already hardware poisoned\n", pfn);
1564 if (flags & MF_ACTION_REQUIRED) {
1565 head = compound_head(p);
1566 res = kill_accessing_process(current, page_to_pfn(head), flags);
1569 } else if (res == -EBUSY) {
1574 action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
1578 head = compound_head(p);
1581 if (hwpoison_filter(p)) {
1582 ClearPageHWPoison(head);
1587 num_poisoned_pages_inc();
1590 * Handling free hugepage. The possible race with hugepage allocation
1591 * or demotion can be prevented by PageHWPoison flag.
1596 if (__page_handle_poison(p)) {
1600 action_result(pfn, MF_MSG_FREE_HUGE, res);
1601 return res == MF_RECOVERED ? 0 : -EBUSY;
1604 page_flags = head->flags;
1607 * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so
1608 * simply disable it. In order to make it work properly, we need
1610 * - conversion of a pud that maps an error hugetlb into hwpoison
1611 * entry properly works, and
1612 * - other mm code walking over page table is aware of pud-aligned
1615 if (huge_page_size(page_hstate(head)) > PMD_SIZE) {
1616 action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED);
1621 if (!hwpoison_user_mappings(p, pfn, flags, head)) {
1622 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1627 return identify_page_state(pfn, p, page_flags);
1633 static inline int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb)
1639 static int memory_failure_dev_pagemap(unsigned long pfn, int flags,
1640 struct dev_pagemap *pgmap)
1642 struct page *page = pfn_to_page(pfn);
1643 unsigned long size = 0;
1650 if (flags & MF_COUNT_INCREASED)
1652 * Drop the extra refcount in case we come from madvise().
1656 /* device metadata space is not recoverable */
1657 if (!pgmap_pfn_valid(pgmap, pfn)) {
1663 * Pages instantiated by device-dax (not filesystem-dax)
1664 * may be compound pages.
1666 page = compound_head(page);
1669 * Prevent the inode from being freed while we are interrogating
1670 * the address_space, typically this would be handled by
1671 * lock_page(), but dax pages do not use the page lock. This
1672 * also prevents changes to the mapping of this pfn until
1673 * poison signaling is complete.
1675 cookie = dax_lock_page(page);
1679 if (hwpoison_filter(page)) {
1684 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
1686 * TODO: Handle HMM pages which may need coordination
1687 * with device-side memory.
1693 * Use this flag as an indication that the dax page has been
1694 * remapped UC to prevent speculative consumption of poison.
1696 SetPageHWPoison(page);
1699 * Unlike System-RAM there is no possibility to swap in a
1700 * different physical page at a given virtual address, so all
1701 * userspace consumption of ZONE_DEVICE memory necessitates
1702 * SIGBUS (i.e. MF_MUST_KILL)
1704 flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1705 collect_procs(page, &tokill, true);
1707 list_for_each_entry(tk, &tokill, nd)
1709 size = max(size, 1UL << tk->size_shift);
1712 * Unmap the largest mapping to avoid breaking up
1713 * device-dax mappings which are constant size. The
1714 * actual size of the mapping being torn down is
1715 * communicated in siginfo, see kill_proc()
1717 start = (page->index << PAGE_SHIFT) & ~(size - 1);
1718 unmap_mapping_range(page->mapping, start, size, 0);
1720 kill_procs(&tokill, true, false, pfn, flags);
1723 dax_unlock_page(page, cookie);
1725 /* drop pgmap ref acquired in caller */
1726 put_dev_pagemap(pgmap);
1727 action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED);
1731 static DEFINE_MUTEX(mf_mutex);
1734 * memory_failure - Handle memory failure of a page.
1735 * @pfn: Page Number of the corrupted page
1736 * @flags: fine tune action taken
1738 * This function is called by the low level machine check code
1739 * of an architecture when it detects hardware memory corruption
1740 * of a page. It tries its best to recover, which includes
1741 * dropping pages, killing processes etc.
1743 * The function is primarily of use for corruptions that
1744 * happen outside the current execution context (e.g. when
1745 * detected by a background scrubber)
1747 * Must run in process context (e.g. a work queue) with interrupts
1748 * enabled and no spinlocks hold.
1750 * Return: 0 for successfully handled the memory error,
1751 * -EOPNOTSUPP for memory_filter() filtered the error event,
1752 * < 0(except -EOPNOTSUPP) on failure.
1754 int memory_failure(unsigned long pfn, int flags)
1758 struct dev_pagemap *pgmap;
1760 unsigned long page_flags;
1764 if (!sysctl_memory_failure_recovery)
1765 panic("Memory failure on page %lx", pfn);
1767 mutex_lock(&mf_mutex);
1769 p = pfn_to_online_page(pfn);
1771 res = arch_memory_failure(pfn, flags);
1775 if (pfn_valid(pfn)) {
1776 pgmap = get_dev_pagemap(pfn, NULL);
1778 res = memory_failure_dev_pagemap(pfn, flags,
1783 pr_err("Memory failure: %#lx: memory outside kernel control\n",
1790 res = try_memory_failure_hugetlb(pfn, flags, &hugetlb);
1794 if (TestSetPageHWPoison(p)) {
1795 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1798 if (flags & MF_ACTION_REQUIRED)
1799 res = kill_accessing_process(current, pfn, flags);
1800 if (flags & MF_COUNT_INCREASED)
1805 hpage = compound_head(p);
1806 num_poisoned_pages_inc();
1809 * We need/can do nothing about count=0 pages.
1810 * 1) it's a free page, and therefore in safe hand:
1811 * prep_new_page() will be the gate keeper.
1812 * 2) it's part of a non-compound high order page.
1813 * Implies some kernel user: cannot stop them from
1814 * R/W the page; let's pray that the page has been
1815 * used and will be freed some time later.
1816 * In fact it's dangerous to directly bump up page count from 0,
1817 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
1819 if (!(flags & MF_COUNT_INCREASED)) {
1820 res = get_hwpoison_page(p, flags);
1822 if (is_free_buddy_page(p)) {
1823 if (take_page_off_buddy(p)) {
1827 /* We lost the race, try again */
1829 ClearPageHWPoison(p);
1830 num_poisoned_pages_dec();
1836 action_result(pfn, MF_MSG_BUDDY, res);
1837 res = res == MF_RECOVERED ? 0 : -EBUSY;
1839 action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
1843 } else if (res < 0) {
1844 action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
1850 if (PageTransHuge(hpage)) {
1852 * Bail out before SetPageHasHWPoisoned() if hpage is
1853 * huge_zero_page, although PG_has_hwpoisoned is not
1854 * checked in set_huge_zero_page().
1856 * TODO: Handle memory failure of huge_zero_page thoroughly.
1858 if (is_huge_zero_page(hpage)) {
1859 action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED);
1865 * The flag must be set after the refcount is bumped
1866 * otherwise it may race with THP split.
1867 * And the flag can't be set in get_hwpoison_page() since
1868 * it is called by soft offline too and it is just called
1869 * for !MF_COUNT_INCREASE. So here seems to be the best
1872 * Don't need care about the above error handling paths for
1873 * get_hwpoison_page() since they handle either free page
1874 * or unhandlable page. The refcount is bumped iff the
1875 * page is a valid handlable page.
1877 SetPageHasHWPoisoned(hpage);
1878 if (try_to_split_thp_page(p, "Memory Failure") < 0) {
1879 action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED);
1883 VM_BUG_ON_PAGE(!page_count(p), p);
1887 * We ignore non-LRU pages for good reasons.
1888 * - PG_locked is only well defined for LRU pages and a few others
1889 * - to avoid races with __SetPageLocked()
1890 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1891 * The check (unnecessarily) ignores LRU pages being isolated and
1892 * walked by the page reclaim code, however that's not a big loss.
1899 * We're only intended to deal with the non-Compound page here.
1900 * However, the page could have changed compound pages due to
1901 * race window. If this happens, we could try again to hopefully
1902 * handle the page next round.
1904 if (PageCompound(p)) {
1906 if (TestClearPageHWPoison(p))
1907 num_poisoned_pages_dec();
1910 flags &= ~MF_COUNT_INCREASED;
1914 action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
1920 * We use page flags to determine what action should be taken, but
1921 * the flags can be modified by the error containment action. One
1922 * example is an mlocked page, where PG_mlocked is cleared by
1923 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1924 * correctly, we save a copy of the page flags at this time.
1926 page_flags = p->flags;
1928 if (hwpoison_filter(p)) {
1929 if (TestClearPageHWPoison(p))
1930 num_poisoned_pages_dec();
1938 * __munlock_pagevec may clear a writeback page's LRU flag without
1939 * page_lock. We need wait writeback completion for this page or it
1940 * may trigger vfs BUG while evict inode.
1942 if (!PageLRU(p) && !PageWriteback(p))
1943 goto identify_page_state;
1946 * It's very difficult to mess with pages currently under IO
1947 * and in many cases impossible, so we just avoid it here.
1949 wait_on_page_writeback(p);
1952 * Now take care of user space mappings.
1953 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1955 if (!hwpoison_user_mappings(p, pfn, flags, p)) {
1956 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1962 * Torn down by someone else?
1964 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1965 action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
1970 identify_page_state:
1971 res = identify_page_state(pfn, p, page_flags);
1972 mutex_unlock(&mf_mutex);
1977 mutex_unlock(&mf_mutex);
1980 EXPORT_SYMBOL_GPL(memory_failure);
1982 #define MEMORY_FAILURE_FIFO_ORDER 4
1983 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1985 struct memory_failure_entry {
1990 struct memory_failure_cpu {
1991 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1992 MEMORY_FAILURE_FIFO_SIZE);
1994 struct work_struct work;
1997 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
2000 * memory_failure_queue - Schedule handling memory failure of a page.
2001 * @pfn: Page Number of the corrupted page
2002 * @flags: Flags for memory failure handling
2004 * This function is called by the low level hardware error handler
2005 * when it detects hardware memory corruption of a page. It schedules
2006 * the recovering of error page, including dropping pages, killing
2009 * The function is primarily of use for corruptions that
2010 * happen outside the current execution context (e.g. when
2011 * detected by a background scrubber)
2013 * Can run in IRQ context.
2015 void memory_failure_queue(unsigned long pfn, int flags)
2017 struct memory_failure_cpu *mf_cpu;
2018 unsigned long proc_flags;
2019 struct memory_failure_entry entry = {
2024 mf_cpu = &get_cpu_var(memory_failure_cpu);
2025 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
2026 if (kfifo_put(&mf_cpu->fifo, entry))
2027 schedule_work_on(smp_processor_id(), &mf_cpu->work);
2029 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
2031 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
2032 put_cpu_var(memory_failure_cpu);
2034 EXPORT_SYMBOL_GPL(memory_failure_queue);
2036 static void memory_failure_work_func(struct work_struct *work)
2038 struct memory_failure_cpu *mf_cpu;
2039 struct memory_failure_entry entry = { 0, };
2040 unsigned long proc_flags;
2043 mf_cpu = container_of(work, struct memory_failure_cpu, work);
2045 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
2046 gotten = kfifo_get(&mf_cpu->fifo, &entry);
2047 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
2050 if (entry.flags & MF_SOFT_OFFLINE)
2051 soft_offline_page(entry.pfn, entry.flags);
2053 memory_failure(entry.pfn, entry.flags);
2058 * Process memory_failure work queued on the specified CPU.
2059 * Used to avoid return-to-userspace racing with the memory_failure workqueue.
2061 void memory_failure_queue_kick(int cpu)
2063 struct memory_failure_cpu *mf_cpu;
2065 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
2066 cancel_work_sync(&mf_cpu->work);
2067 memory_failure_work_func(&mf_cpu->work);
2070 static int __init memory_failure_init(void)
2072 struct memory_failure_cpu *mf_cpu;
2075 for_each_possible_cpu(cpu) {
2076 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
2077 spin_lock_init(&mf_cpu->lock);
2078 INIT_KFIFO(mf_cpu->fifo);
2079 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
2084 core_initcall(memory_failure_init);
2086 #define unpoison_pr_info(fmt, pfn, rs) \
2088 if (__ratelimit(rs)) \
2089 pr_info(fmt, pfn); \
2092 static inline int clear_page_hwpoison(struct ratelimit_state *rs, struct page *p)
2094 if (TestClearPageHWPoison(p)) {
2095 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
2096 page_to_pfn(p), rs);
2097 num_poisoned_pages_dec();
2103 static inline int unpoison_taken_off_page(struct ratelimit_state *rs,
2106 if (put_page_back_buddy(p)) {
2107 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
2108 page_to_pfn(p), rs);
2115 * unpoison_memory - Unpoison a previously poisoned page
2116 * @pfn: Page number of the to be unpoisoned page
2118 * Software-unpoison a page that has been poisoned by
2119 * memory_failure() earlier.
2121 * This is only done on the software-level, so it only works
2122 * for linux injected failures, not real hardware failures
2124 * Returns 0 for success, otherwise -errno.
2126 int unpoison_memory(unsigned long pfn)
2131 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
2132 DEFAULT_RATELIMIT_BURST);
2134 if (!pfn_valid(pfn))
2137 p = pfn_to_page(pfn);
2138 page = compound_head(p);
2140 mutex_lock(&mf_mutex);
2142 if (!PageHWPoison(p)) {
2143 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
2148 if (page_count(page) > 1) {
2149 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
2154 if (page_mapped(page)) {
2155 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
2160 if (page_mapping(page)) {
2161 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
2166 if (PageSlab(page) || PageTable(page))
2169 ret = get_hwpoison_page(p, MF_UNPOISON);
2171 if (clear_page_hwpoison(&unpoison_rs, page))
2175 } else if (ret < 0) {
2176 if (ret == -EHWPOISON) {
2177 ret = unpoison_taken_off_page(&unpoison_rs, p);
2179 unpoison_pr_info("Unpoison: failed to grab page %#lx\n",
2182 int freeit = clear_page_hwpoison(&unpoison_rs, p);
2185 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1)) {
2192 mutex_unlock(&mf_mutex);
2195 EXPORT_SYMBOL(unpoison_memory);
2197 static bool isolate_page(struct page *page, struct list_head *pagelist)
2199 bool isolated = false;
2200 bool lru = PageLRU(page);
2202 if (PageHuge(page)) {
2203 isolated = isolate_huge_page(page, pagelist);
2206 isolated = !isolate_lru_page(page);
2208 isolated = !isolate_movable_page(page, ISOLATE_UNEVICTABLE);
2211 list_add(&page->lru, pagelist);
2214 if (isolated && lru)
2215 inc_node_page_state(page, NR_ISOLATED_ANON +
2216 page_is_file_lru(page));
2219 * If we succeed to isolate the page, we grabbed another refcount on
2220 * the page, so we can safely drop the one we got from get_any_pages().
2221 * If we failed to isolate the page, it means that we cannot go further
2222 * and we will return an error, so drop the reference we got from
2223 * get_any_pages() as well.
2230 * __soft_offline_page handles hugetlb-pages and non-hugetlb pages.
2231 * If the page is a non-dirty unmapped page-cache page, it simply invalidates.
2232 * If the page is mapped, it migrates the contents over.
2234 static int __soft_offline_page(struct page *page)
2237 unsigned long pfn = page_to_pfn(page);
2238 struct page *hpage = compound_head(page);
2239 char const *msg_page[] = {"page", "hugepage"};
2240 bool huge = PageHuge(page);
2241 LIST_HEAD(pagelist);
2242 struct migration_target_control mtc = {
2243 .nid = NUMA_NO_NODE,
2244 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
2248 if (!PageHuge(page))
2249 wait_on_page_writeback(page);
2250 if (PageHWPoison(page)) {
2253 pr_info("soft offline: %#lx page already poisoned\n", pfn);
2257 if (!PageHuge(page) && PageLRU(page) && !PageSwapCache(page))
2259 * Try to invalidate first. This should work for
2260 * non dirty unmapped page cache pages.
2262 ret = invalidate_inode_page(page);
2266 pr_info("soft_offline: %#lx: invalidated\n", pfn);
2267 page_handle_poison(page, false, true);
2271 if (isolate_page(hpage, &pagelist)) {
2272 ret = migrate_pages(&pagelist, alloc_migration_target, NULL,
2273 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE, NULL);
2275 bool release = !huge;
2277 if (!page_handle_poison(page, huge, release))
2280 if (!list_empty(&pagelist))
2281 putback_movable_pages(&pagelist);
2283 pr_info("soft offline: %#lx: %s migration failed %ld, type %pGp\n",
2284 pfn, msg_page[huge], ret, &page->flags);
2289 pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %pGp\n",
2290 pfn, msg_page[huge], page_count(page), &page->flags);
2296 static int soft_offline_in_use_page(struct page *page)
2298 struct page *hpage = compound_head(page);
2300 if (!PageHuge(page) && PageTransHuge(hpage))
2301 if (try_to_split_thp_page(page, "soft offline") < 0)
2303 return __soft_offline_page(page);
2306 static int soft_offline_free_page(struct page *page)
2310 if (!page_handle_poison(page, true, false))
2316 static void put_ref_page(struct page *page)
2323 * soft_offline_page - Soft offline a page.
2324 * @pfn: pfn to soft-offline
2325 * @flags: flags. Same as memory_failure().
2327 * Returns 0 on success, otherwise negated errno.
2329 * Soft offline a page, by migration or invalidation,
2330 * without killing anything. This is for the case when
2331 * a page is not corrupted yet (so it's still valid to access),
2332 * but has had a number of corrected errors and is better taken
2335 * The actual policy on when to do that is maintained by
2338 * This should never impact any application or cause data loss,
2339 * however it might take some time.
2341 * This is not a 100% solution for all memory, but tries to be
2342 * ``good enough'' for the majority of memory.
2344 int soft_offline_page(unsigned long pfn, int flags)
2347 bool try_again = true;
2348 struct page *page, *ref_page = NULL;
2350 WARN_ON_ONCE(!pfn_valid(pfn) && (flags & MF_COUNT_INCREASED));
2352 if (!pfn_valid(pfn))
2354 if (flags & MF_COUNT_INCREASED)
2355 ref_page = pfn_to_page(pfn);
2357 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
2358 page = pfn_to_online_page(pfn);
2360 put_ref_page(ref_page);
2364 mutex_lock(&mf_mutex);
2366 if (PageHWPoison(page)) {
2367 pr_info("%s: %#lx page already poisoned\n", __func__, pfn);
2368 put_ref_page(ref_page);
2369 mutex_unlock(&mf_mutex);
2375 ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE);
2379 ret = soft_offline_in_use_page(page);
2380 } else if (ret == 0) {
2381 if (soft_offline_free_page(page) && try_again) {
2383 flags &= ~MF_COUNT_INCREASED;
2388 mutex_unlock(&mf_mutex);