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/mm/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
37 #define pr_fmt(fmt) "Memory failure: " fmt
39 #include <linux/kernel.h>
41 #include <linux/page-flags.h>
42 #include <linux/kernel-page-flags.h>
43 #include <linux/sched/signal.h>
44 #include <linux/sched/task.h>
45 #include <linux/dax.h>
46 #include <linux/ksm.h>
47 #include <linux/rmap.h>
48 #include <linux/export.h>
49 #include <linux/pagemap.h>
50 #include <linux/swap.h>
51 #include <linux/backing-dev.h>
52 #include <linux/migrate.h>
53 #include <linux/suspend.h>
54 #include <linux/slab.h>
55 #include <linux/swapops.h>
56 #include <linux/hugetlb.h>
57 #include <linux/memory_hotplug.h>
58 #include <linux/mm_inline.h>
59 #include <linux/memremap.h>
60 #include <linux/kfifo.h>
61 #include <linux/ratelimit.h>
62 #include <linux/page-isolation.h>
63 #include <linux/pagewalk.h>
64 #include <linux/shmem_fs.h>
65 #include <linux/sysctl.h>
68 #include "ras/ras_event.h"
70 static int sysctl_memory_failure_early_kill __read_mostly;
72 static int sysctl_memory_failure_recovery __read_mostly = 1;
74 atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
76 static bool hw_memory_failure __read_mostly = false;
78 inline void num_poisoned_pages_inc(unsigned long pfn)
80 atomic_long_inc(&num_poisoned_pages);
81 memblk_nr_poison_inc(pfn);
84 inline void num_poisoned_pages_sub(unsigned long pfn, long i)
86 atomic_long_sub(i, &num_poisoned_pages);
88 memblk_nr_poison_sub(pfn, i);
92 * MF_ATTR_RO - Create sysfs entry for each memory failure statistics.
93 * @_name: name of the file in the per NUMA sysfs directory.
95 #define MF_ATTR_RO(_name) \
96 static ssize_t _name##_show(struct device *dev, \
97 struct device_attribute *attr, \
100 struct memory_failure_stats *mf_stats = \
101 &NODE_DATA(dev->id)->mf_stats; \
102 return sprintf(buf, "%lu\n", mf_stats->_name); \
104 static DEVICE_ATTR_RO(_name)
110 MF_ATTR_RO(recovered);
112 static struct attribute *memory_failure_attr[] = {
113 &dev_attr_total.attr,
114 &dev_attr_ignored.attr,
115 &dev_attr_failed.attr,
116 &dev_attr_delayed.attr,
117 &dev_attr_recovered.attr,
121 const struct attribute_group memory_failure_attr_group = {
122 .name = "memory_failure",
123 .attrs = memory_failure_attr,
126 static struct ctl_table memory_failure_table[] = {
128 .procname = "memory_failure_early_kill",
129 .data = &sysctl_memory_failure_early_kill,
130 .maxlen = sizeof(sysctl_memory_failure_early_kill),
132 .proc_handler = proc_dointvec_minmax,
133 .extra1 = SYSCTL_ZERO,
134 .extra2 = SYSCTL_ONE,
137 .procname = "memory_failure_recovery",
138 .data = &sysctl_memory_failure_recovery,
139 .maxlen = sizeof(sysctl_memory_failure_recovery),
141 .proc_handler = proc_dointvec_minmax,
142 .extra1 = SYSCTL_ZERO,
143 .extra2 = SYSCTL_ONE,
150 * 1: the page is dissolved (if needed) and taken off from buddy,
151 * 0: the page is dissolved (if needed) and not taken off from buddy,
152 * < 0: failed to dissolve.
154 static int __page_handle_poison(struct page *page)
158 zone_pcp_disable(page_zone(page));
159 ret = dissolve_free_huge_page(page);
161 ret = take_page_off_buddy(page);
162 zone_pcp_enable(page_zone(page));
167 static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release)
169 if (hugepage_or_freepage) {
171 * Doing this check for free pages is also fine since dissolve_free_huge_page
172 * returns 0 for non-hugetlb pages as well.
174 if (__page_handle_poison(page) <= 0)
176 * We could fail to take off the target page from buddy
177 * for example due to racy page allocation, but that's
178 * acceptable because soft-offlined page is not broken
179 * and if someone really want to use it, they should
185 SetPageHWPoison(page);
189 num_poisoned_pages_inc(page_to_pfn(page));
194 #if IS_ENABLED(CONFIG_HWPOISON_INJECT)
196 u32 hwpoison_filter_enable = 0;
197 u32 hwpoison_filter_dev_major = ~0U;
198 u32 hwpoison_filter_dev_minor = ~0U;
199 u64 hwpoison_filter_flags_mask;
200 u64 hwpoison_filter_flags_value;
201 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
202 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
203 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
204 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
205 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
207 static int hwpoison_filter_dev(struct page *p)
209 struct address_space *mapping;
212 if (hwpoison_filter_dev_major == ~0U &&
213 hwpoison_filter_dev_minor == ~0U)
216 mapping = page_mapping(p);
217 if (mapping == NULL || mapping->host == NULL)
220 dev = mapping->host->i_sb->s_dev;
221 if (hwpoison_filter_dev_major != ~0U &&
222 hwpoison_filter_dev_major != MAJOR(dev))
224 if (hwpoison_filter_dev_minor != ~0U &&
225 hwpoison_filter_dev_minor != MINOR(dev))
231 static int hwpoison_filter_flags(struct page *p)
233 if (!hwpoison_filter_flags_mask)
236 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
237 hwpoison_filter_flags_value)
244 * This allows stress tests to limit test scope to a collection of tasks
245 * by putting them under some memcg. This prevents killing unrelated/important
246 * processes such as /sbin/init. Note that the target task may share clean
247 * pages with init (eg. libc text), which is harmless. If the target task
248 * share _dirty_ pages with another task B, the test scheme must make sure B
249 * is also included in the memcg. At last, due to race conditions this filter
250 * can only guarantee that the page either belongs to the memcg tasks, or is
254 u64 hwpoison_filter_memcg;
255 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
256 static int hwpoison_filter_task(struct page *p)
258 if (!hwpoison_filter_memcg)
261 if (page_cgroup_ino(p) != hwpoison_filter_memcg)
267 static int hwpoison_filter_task(struct page *p) { return 0; }
270 int hwpoison_filter(struct page *p)
272 if (!hwpoison_filter_enable)
275 if (hwpoison_filter_dev(p))
278 if (hwpoison_filter_flags(p))
281 if (hwpoison_filter_task(p))
287 int hwpoison_filter(struct page *p)
293 EXPORT_SYMBOL_GPL(hwpoison_filter);
296 * Kill all processes that have a poisoned page mapped and then isolate
300 * Find all processes having the page mapped and kill them.
301 * But we keep a page reference around so that the page is not
302 * actually freed yet.
303 * Then stash the page away
305 * There's no convenient way to get back to mapped processes
306 * from the VMAs. So do a brute-force search over all
309 * Remember that machine checks are not common (or rather
310 * if they are common you have other problems), so this shouldn't
311 * be a performance issue.
313 * Also there are some races possible while we get from the
314 * error detection to actually handle it.
319 struct task_struct *tsk;
325 * Send all the processes who have the page mapped a signal.
326 * ``action optional'' if they are not immediately affected by the error
327 * ``action required'' if error happened in current execution context
329 static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
331 struct task_struct *t = tk->tsk;
332 short addr_lsb = tk->size_shift;
335 pr_err("%#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
336 pfn, t->comm, t->pid);
338 if ((flags & MF_ACTION_REQUIRED) && (t == current))
339 ret = force_sig_mceerr(BUS_MCEERR_AR,
340 (void __user *)tk->addr, addr_lsb);
343 * Signal other processes sharing the page if they have
345 * Don't use force here, it's convenient if the signal
346 * can be temporarily blocked.
347 * This could cause a loop when the user sets SIGBUS
348 * to SIG_IGN, but hopefully no one will do that?
350 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
353 pr_info("Error sending signal to %s:%d: %d\n",
354 t->comm, t->pid, ret);
359 * Unknown page type encountered. Try to check whether it can turn PageLRU by
362 void shake_page(struct page *p)
369 if (PageLRU(p) || is_free_buddy_page(p))
374 * TODO: Could shrink slab caches here if a lightweight range-based
375 * shrinker will be available.
378 EXPORT_SYMBOL_GPL(shake_page);
380 static unsigned long dev_pagemap_mapping_shift(struct vm_area_struct *vma,
381 unsigned long address)
383 unsigned long ret = 0;
390 VM_BUG_ON_VMA(address == -EFAULT, vma);
391 pgd = pgd_offset(vma->vm_mm, address);
392 if (!pgd_present(*pgd))
394 p4d = p4d_offset(pgd, address);
395 if (!p4d_present(*p4d))
397 pud = pud_offset(p4d, address);
398 if (!pud_present(*pud))
400 if (pud_devmap(*pud))
402 pmd = pmd_offset(pud, address);
403 if (!pmd_present(*pmd))
405 if (pmd_devmap(*pmd))
407 pte = pte_offset_map(pmd, address);
410 if (pte_present(*pte) && pte_devmap(*pte))
417 * Failure handling: if we can't find or can't kill a process there's
418 * not much we can do. We just print a message and ignore otherwise.
421 #define FSDAX_INVALID_PGOFF ULONG_MAX
424 * Schedule a process for later kill.
425 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
427 * Note: @fsdax_pgoff is used only when @p is a fsdax page and a
428 * filesystem with a memory failure handler has claimed the
429 * memory_failure event. In all other cases, page->index and
430 * page->mapping are sufficient for mapping the page back to its
431 * corresponding user virtual address.
433 static void __add_to_kill(struct task_struct *tsk, struct page *p,
434 struct vm_area_struct *vma, struct list_head *to_kill,
435 unsigned long ksm_addr, pgoff_t fsdax_pgoff)
439 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
441 pr_err("Out of memory while machine check handling\n");
445 tk->addr = ksm_addr ? ksm_addr : page_address_in_vma(p, vma);
446 if (is_zone_device_page(p)) {
447 if (fsdax_pgoff != FSDAX_INVALID_PGOFF)
448 tk->addr = vma_pgoff_address(fsdax_pgoff, 1, vma);
449 tk->size_shift = dev_pagemap_mapping_shift(vma, tk->addr);
451 tk->size_shift = page_shift(compound_head(p));
454 * Send SIGKILL if "tk->addr == -EFAULT". Also, as
455 * "tk->size_shift" is always non-zero for !is_zone_device_page(),
456 * so "tk->size_shift == 0" effectively checks no mapping on
457 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times
458 * to a process' address space, it's possible not all N VMAs
459 * contain mappings for the page, but at least one VMA does.
460 * Only deliver SIGBUS with payload derived from the VMA that
461 * has a mapping for the page.
463 if (tk->addr == -EFAULT) {
464 pr_info("Unable to find user space address %lx in %s\n",
465 page_to_pfn(p), tsk->comm);
466 } else if (tk->size_shift == 0) {
471 get_task_struct(tsk);
473 list_add_tail(&tk->nd, to_kill);
476 static void add_to_kill_anon_file(struct task_struct *tsk, struct page *p,
477 struct vm_area_struct *vma,
478 struct list_head *to_kill)
480 __add_to_kill(tsk, p, vma, to_kill, 0, FSDAX_INVALID_PGOFF);
484 static bool task_in_to_kill_list(struct list_head *to_kill,
485 struct task_struct *tsk)
487 struct to_kill *tk, *next;
489 list_for_each_entry_safe(tk, next, to_kill, nd) {
496 void add_to_kill_ksm(struct task_struct *tsk, struct page *p,
497 struct vm_area_struct *vma, struct list_head *to_kill,
498 unsigned long ksm_addr)
500 if (!task_in_to_kill_list(to_kill, tsk))
501 __add_to_kill(tsk, p, vma, to_kill, ksm_addr, FSDAX_INVALID_PGOFF);
505 * Kill the processes that have been collected earlier.
507 * Only do anything when FORCEKILL is set, otherwise just free the
508 * list (this is used for clean pages which do not need killing)
509 * Also when FAIL is set do a force kill because something went
512 static void kill_procs(struct list_head *to_kill, int forcekill, bool fail,
513 unsigned long pfn, int flags)
515 struct to_kill *tk, *next;
517 list_for_each_entry_safe(tk, next, to_kill, nd) {
520 * In case something went wrong with munmapping
521 * make sure the process doesn't catch the
522 * signal and then access the memory. Just kill it.
524 if (fail || tk->addr == -EFAULT) {
525 pr_err("%#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
526 pfn, tk->tsk->comm, tk->tsk->pid);
527 do_send_sig_info(SIGKILL, SEND_SIG_PRIV,
528 tk->tsk, PIDTYPE_PID);
532 * In theory the process could have mapped
533 * something else on the address in-between. We could
534 * check for that, but we need to tell the
537 else if (kill_proc(tk, pfn, flags) < 0)
538 pr_err("%#lx: Cannot send advisory machine check signal to %s:%d\n",
539 pfn, tk->tsk->comm, tk->tsk->pid);
542 put_task_struct(tk->tsk);
548 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
549 * on behalf of the thread group. Return task_struct of the (first found)
550 * dedicated thread if found, and return NULL otherwise.
552 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
553 * have to call rcu_read_lock/unlock() in this function.
555 static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
557 struct task_struct *t;
559 for_each_thread(tsk, t) {
560 if (t->flags & PF_MCE_PROCESS) {
561 if (t->flags & PF_MCE_EARLY)
564 if (sysctl_memory_failure_early_kill)
572 * Determine whether a given process is "early kill" process which expects
573 * to be signaled when some page under the process is hwpoisoned.
574 * Return task_struct of the dedicated thread (main thread unless explicitly
575 * specified) if the process is "early kill" and otherwise returns NULL.
577 * Note that the above is true for Action Optional case. For Action Required
578 * case, it's only meaningful to the current thread which need to be signaled
579 * with SIGBUS, this error is Action Optional for other non current
580 * processes sharing the same error page,if the process is "early kill", the
581 * task_struct of the dedicated thread will also be returned.
583 struct task_struct *task_early_kill(struct task_struct *tsk, int force_early)
588 * Comparing ->mm here because current task might represent
589 * a subthread, while tsk always points to the main thread.
591 if (force_early && tsk->mm == current->mm)
594 return find_early_kill_thread(tsk);
598 * Collect processes when the error hit an anonymous page.
600 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
603 struct folio *folio = page_folio(page);
604 struct vm_area_struct *vma;
605 struct task_struct *tsk;
609 av = folio_lock_anon_vma_read(folio, NULL);
610 if (av == NULL) /* Not actually mapped anymore */
613 pgoff = page_to_pgoff(page);
614 read_lock(&tasklist_lock);
615 for_each_process (tsk) {
616 struct anon_vma_chain *vmac;
617 struct task_struct *t = task_early_kill(tsk, force_early);
621 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
624 if (vma->vm_mm != t->mm)
626 if (!page_mapped_in_vma(page, vma))
628 add_to_kill_anon_file(t, page, vma, to_kill);
631 read_unlock(&tasklist_lock);
632 anon_vma_unlock_read(av);
636 * Collect processes when the error hit a file mapped page.
638 static void collect_procs_file(struct page *page, struct list_head *to_kill,
641 struct vm_area_struct *vma;
642 struct task_struct *tsk;
643 struct address_space *mapping = page->mapping;
646 i_mmap_lock_read(mapping);
647 read_lock(&tasklist_lock);
648 pgoff = page_to_pgoff(page);
649 for_each_process(tsk) {
650 struct task_struct *t = task_early_kill(tsk, force_early);
654 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
657 * Send early kill signal to tasks where a vma covers
658 * the page but the corrupted page is not necessarily
659 * mapped it in its pte.
660 * Assume applications who requested early kill want
661 * to be informed of all such data corruptions.
663 if (vma->vm_mm == t->mm)
664 add_to_kill_anon_file(t, page, vma, to_kill);
667 read_unlock(&tasklist_lock);
668 i_mmap_unlock_read(mapping);
672 static void add_to_kill_fsdax(struct task_struct *tsk, struct page *p,
673 struct vm_area_struct *vma,
674 struct list_head *to_kill, pgoff_t pgoff)
676 __add_to_kill(tsk, p, vma, to_kill, 0, pgoff);
680 * Collect processes when the error hit a fsdax page.
682 static void collect_procs_fsdax(struct page *page,
683 struct address_space *mapping, pgoff_t pgoff,
684 struct list_head *to_kill)
686 struct vm_area_struct *vma;
687 struct task_struct *tsk;
689 i_mmap_lock_read(mapping);
690 read_lock(&tasklist_lock);
691 for_each_process(tsk) {
692 struct task_struct *t = task_early_kill(tsk, true);
696 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
697 if (vma->vm_mm == t->mm)
698 add_to_kill_fsdax(t, page, vma, to_kill, pgoff);
701 read_unlock(&tasklist_lock);
702 i_mmap_unlock_read(mapping);
704 #endif /* CONFIG_FS_DAX */
707 * Collect the processes who have the corrupted page mapped to kill.
709 static void collect_procs(struct page *page, struct list_head *tokill,
714 if (unlikely(PageKsm(page)))
715 collect_procs_ksm(page, tokill, force_early);
716 else if (PageAnon(page))
717 collect_procs_anon(page, tokill, force_early);
719 collect_procs_file(page, tokill, force_early);
728 static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift)
731 tk->size_shift = shift;
734 static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift,
735 unsigned long poisoned_pfn, struct to_kill *tk)
737 unsigned long pfn = 0;
739 if (pte_present(pte)) {
742 swp_entry_t swp = pte_to_swp_entry(pte);
744 if (is_hwpoison_entry(swp))
745 pfn = swp_offset_pfn(swp);
748 if (!pfn || pfn != poisoned_pfn)
751 set_to_kill(tk, addr, shift);
755 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
756 static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
757 struct hwp_walk *hwp)
761 unsigned long hwpoison_vaddr;
763 if (!pmd_present(pmd))
766 if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) {
767 hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT);
768 set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT);
774 static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
775 struct hwp_walk *hwp)
781 static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr,
782 unsigned long end, struct mm_walk *walk)
784 struct hwp_walk *hwp = walk->private;
786 pte_t *ptep, *mapped_pte;
789 ptl = pmd_trans_huge_lock(pmdp, walk->vma);
791 ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp);
796 mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp,
801 for (; addr != end; ptep++, addr += PAGE_SIZE) {
802 ret = check_hwpoisoned_entry(*ptep, addr, PAGE_SHIFT,
807 pte_unmap_unlock(mapped_pte, ptl);
813 #ifdef CONFIG_HUGETLB_PAGE
814 static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask,
815 unsigned long addr, unsigned long end,
816 struct mm_walk *walk)
818 struct hwp_walk *hwp = walk->private;
819 pte_t pte = huge_ptep_get(ptep);
820 struct hstate *h = hstate_vma(walk->vma);
822 return check_hwpoisoned_entry(pte, addr, huge_page_shift(h),
826 #define hwpoison_hugetlb_range NULL
829 static const struct mm_walk_ops hwp_walk_ops = {
830 .pmd_entry = hwpoison_pte_range,
831 .hugetlb_entry = hwpoison_hugetlb_range,
835 * Sends SIGBUS to the current process with error info.
837 * This function is intended to handle "Action Required" MCEs on already
838 * hardware poisoned pages. They could happen, for example, when
839 * memory_failure() failed to unmap the error page at the first call, or
840 * when multiple local machine checks happened on different CPUs.
842 * MCE handler currently has no easy access to the error virtual address,
843 * so this function walks page table to find it. The returned virtual address
844 * is proper in most cases, but it could be wrong when the application
845 * process has multiple entries mapping the error page.
847 static int kill_accessing_process(struct task_struct *p, unsigned long pfn,
851 struct hwp_walk priv = {
859 mmap_read_lock(p->mm);
860 ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwp_walk_ops,
862 if (ret == 1 && priv.tk.addr)
863 kill_proc(&priv.tk, pfn, flags);
866 mmap_read_unlock(p->mm);
867 return ret > 0 ? -EHWPOISON : -EFAULT;
870 static const char *action_name[] = {
871 [MF_IGNORED] = "Ignored",
872 [MF_FAILED] = "Failed",
873 [MF_DELAYED] = "Delayed",
874 [MF_RECOVERED] = "Recovered",
877 static const char * const action_page_types[] = {
878 [MF_MSG_KERNEL] = "reserved kernel page",
879 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
880 [MF_MSG_SLAB] = "kernel slab page",
881 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
882 [MF_MSG_HUGE] = "huge page",
883 [MF_MSG_FREE_HUGE] = "free huge page",
884 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
885 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
886 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
887 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
888 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
889 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
890 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
891 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
892 [MF_MSG_CLEAN_LRU] = "clean LRU page",
893 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
894 [MF_MSG_BUDDY] = "free buddy page",
895 [MF_MSG_DAX] = "dax page",
896 [MF_MSG_UNSPLIT_THP] = "unsplit thp",
897 [MF_MSG_UNKNOWN] = "unknown page",
901 * XXX: It is possible that a page is isolated from LRU cache,
902 * and then kept in swap cache or failed to remove from page cache.
903 * The page count will stop it from being freed by unpoison.
904 * Stress tests should be aware of this memory leak problem.
906 static int delete_from_lru_cache(struct page *p)
908 if (isolate_lru_page(p)) {
910 * Clear sensible page flags, so that the buddy system won't
911 * complain when the page is unpoison-and-freed.
914 ClearPageUnevictable(p);
917 * Poisoned page might never drop its ref count to 0 so we have
918 * to uncharge it manually from its memcg.
920 mem_cgroup_uncharge(page_folio(p));
923 * drop the page count elevated by isolate_lru_page()
931 static int truncate_error_page(struct page *p, unsigned long pfn,
932 struct address_space *mapping)
936 if (mapping->a_ops->error_remove_page) {
937 struct folio *folio = page_folio(p);
938 int err = mapping->a_ops->error_remove_page(mapping, p);
941 pr_info("%#lx: Failed to punch page: %d\n", pfn, err);
942 } else if (folio_has_private(folio) &&
943 !filemap_release_folio(folio, GFP_NOIO)) {
944 pr_info("%#lx: failed to release buffers\n", pfn);
950 * If the file system doesn't support it just invalidate
951 * This fails on dirty or anything with private pages
953 if (invalidate_inode_page(p))
956 pr_info("%#lx: Failed to invalidate\n", pfn);
965 enum mf_action_page_type type;
967 /* Callback ->action() has to unlock the relevant page inside it. */
968 int (*action)(struct page_state *ps, struct page *p);
972 * Return true if page is still referenced by others, otherwise return
975 * The extra_pins is true when one extra refcount is expected.
977 static bool has_extra_refcount(struct page_state *ps, struct page *p,
980 int count = page_count(p) - 1;
986 pr_err("%#lx: %s still referenced by %d users\n",
987 page_to_pfn(p), action_page_types[ps->type], count);
995 * Error hit kernel page.
996 * Do nothing, try to be lucky and not touch this instead. For a few cases we
997 * could be more sophisticated.
999 static int me_kernel(struct page_state *ps, struct page *p)
1006 * Page in unknown state. Do nothing.
1008 static int me_unknown(struct page_state *ps, struct page *p)
1010 pr_err("%#lx: Unknown page state\n", page_to_pfn(p));
1016 * Clean (or cleaned) page cache page.
1018 static int me_pagecache_clean(struct page_state *ps, struct page *p)
1021 struct address_space *mapping;
1024 delete_from_lru_cache(p);
1027 * For anonymous pages we're done the only reference left
1028 * should be the one m_f() holds.
1036 * Now truncate the page in the page cache. This is really
1037 * more like a "temporary hole punch"
1038 * Don't do this for block devices when someone else
1039 * has a reference, because it could be file system metadata
1040 * and that's not safe to truncate.
1042 mapping = page_mapping(p);
1045 * Page has been teared down in the meanwhile
1052 * The shmem page is kept in page cache instead of truncating
1053 * so is expected to have an extra refcount after error-handling.
1055 extra_pins = shmem_mapping(mapping);
1058 * Truncation is a bit tricky. Enable it per file system for now.
1060 * Open: to take i_rwsem or not for this? Right now we don't.
1062 ret = truncate_error_page(p, page_to_pfn(p), mapping);
1063 if (has_extra_refcount(ps, p, extra_pins))
1073 * Dirty pagecache page
1074 * Issues: when the error hit a hole page the error is not properly
1077 static int me_pagecache_dirty(struct page_state *ps, struct page *p)
1079 struct address_space *mapping = page_mapping(p);
1082 /* TBD: print more information about the file. */
1085 * IO error will be reported by write(), fsync(), etc.
1086 * who check the mapping.
1087 * This way the application knows that something went
1088 * wrong with its dirty file data.
1090 * There's one open issue:
1092 * The EIO will be only reported on the next IO
1093 * operation and then cleared through the IO map.
1094 * Normally Linux has two mechanisms to pass IO error
1095 * first through the AS_EIO flag in the address space
1096 * and then through the PageError flag in the page.
1097 * Since we drop pages on memory failure handling the
1098 * only mechanism open to use is through AS_AIO.
1100 * This has the disadvantage that it gets cleared on
1101 * the first operation that returns an error, while
1102 * the PageError bit is more sticky and only cleared
1103 * when the page is reread or dropped. If an
1104 * application assumes it will always get error on
1105 * fsync, but does other operations on the fd before
1106 * and the page is dropped between then the error
1107 * will not be properly reported.
1109 * This can already happen even without hwpoisoned
1110 * pages: first on metadata IO errors (which only
1111 * report through AS_EIO) or when the page is dropped
1112 * at the wrong time.
1114 * So right now we assume that the application DTRT on
1115 * the first EIO, but we're not worse than other parts
1118 mapping_set_error(mapping, -EIO);
1121 return me_pagecache_clean(ps, p);
1125 * Clean and dirty swap cache.
1127 * Dirty swap cache page is tricky to handle. The page could live both in page
1128 * cache and swap cache(ie. page is freshly swapped in). So it could be
1129 * referenced concurrently by 2 types of PTEs:
1130 * normal PTEs and swap PTEs. We try to handle them consistently by calling
1131 * try_to_unmap(!TTU_HWPOISON) to convert the normal PTEs to swap PTEs,
1133 * - clear dirty bit to prevent IO
1135 * - but keep in the swap cache, so that when we return to it on
1136 * a later page fault, we know the application is accessing
1137 * corrupted data and shall be killed (we installed simple
1138 * interception code in do_swap_page to catch it).
1140 * Clean swap cache pages can be directly isolated. A later page fault will
1141 * bring in the known good data from disk.
1143 static int me_swapcache_dirty(struct page_state *ps, struct page *p)
1146 bool extra_pins = false;
1149 /* Trigger EIO in shmem: */
1150 ClearPageUptodate(p);
1152 ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED;
1155 if (ret == MF_DELAYED)
1158 if (has_extra_refcount(ps, p, extra_pins))
1164 static int me_swapcache_clean(struct page_state *ps, struct page *p)
1166 struct folio *folio = page_folio(p);
1169 delete_from_swap_cache(folio);
1171 ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED;
1172 folio_unlock(folio);
1174 if (has_extra_refcount(ps, p, false))
1181 * Huge pages. Needs work.
1183 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
1184 * To narrow down kill region to one page, we need to break up pmd.
1186 static int me_huge_page(struct page_state *ps, struct page *p)
1189 struct page *hpage = compound_head(p);
1190 struct address_space *mapping;
1191 bool extra_pins = false;
1193 if (!PageHuge(hpage))
1196 mapping = page_mapping(hpage);
1198 res = truncate_error_page(hpage, page_to_pfn(p), mapping);
1199 /* The page is kept in page cache. */
1205 * migration entry prevents later access on error hugepage,
1206 * so we can free and dissolve it into buddy to save healthy
1210 if (__page_handle_poison(p) >= 0) {
1218 if (has_extra_refcount(ps, p, extra_pins))
1225 * Various page states we can handle.
1227 * A page state is defined by its current page->flags bits.
1228 * The table matches them in order and calls the right handler.
1230 * This is quite tricky because we can access page at any time
1231 * in its live cycle, so all accesses have to be extremely careful.
1233 * This is not complete. More states could be added.
1234 * For any missing state don't attempt recovery.
1237 #define dirty (1UL << PG_dirty)
1238 #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked))
1239 #define unevict (1UL << PG_unevictable)
1240 #define mlock (1UL << PG_mlocked)
1241 #define lru (1UL << PG_lru)
1242 #define head (1UL << PG_head)
1243 #define slab (1UL << PG_slab)
1244 #define reserved (1UL << PG_reserved)
1246 static struct page_state error_states[] = {
1247 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
1249 * free pages are specially detected outside this table:
1250 * PG_buddy pages only make a small fraction of all free pages.
1254 * Could in theory check if slab page is free or if we can drop
1255 * currently unused objects without touching them. But just
1256 * treat it as standard kernel for now.
1258 { slab, slab, MF_MSG_SLAB, me_kernel },
1260 { head, head, MF_MSG_HUGE, me_huge_page },
1262 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
1263 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
1265 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
1266 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
1268 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
1269 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
1271 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
1272 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
1275 * Catchall entry: must be at end.
1277 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
1289 static void update_per_node_mf_stats(unsigned long pfn,
1290 enum mf_result result)
1292 int nid = MAX_NUMNODES;
1293 struct memory_failure_stats *mf_stats = NULL;
1295 nid = pfn_to_nid(pfn);
1296 if (unlikely(nid < 0 || nid >= MAX_NUMNODES)) {
1297 WARN_ONCE(1, "Memory failure: pfn=%#lx, invalid nid=%d", pfn, nid);
1301 mf_stats = &NODE_DATA(nid)->mf_stats;
1304 ++mf_stats->ignored;
1310 ++mf_stats->delayed;
1313 ++mf_stats->recovered;
1316 WARN_ONCE(1, "Memory failure: mf_result=%d is not properly handled", result);
1323 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
1324 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
1326 static int action_result(unsigned long pfn, enum mf_action_page_type type,
1327 enum mf_result result)
1329 trace_memory_failure_event(pfn, type, result);
1331 num_poisoned_pages_inc(pfn);
1333 update_per_node_mf_stats(pfn, result);
1335 pr_err("%#lx: recovery action for %s: %s\n",
1336 pfn, action_page_types[type], action_name[result]);
1338 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
1341 static int page_action(struct page_state *ps, struct page *p,
1346 /* page p should be unlocked after returning from ps->action(). */
1347 result = ps->action(ps, p);
1349 /* Could do more checks here if page looks ok */
1351 * Could adjust zone counters here to correct for the missing page.
1354 return action_result(pfn, ps->type, result);
1357 static inline bool PageHWPoisonTakenOff(struct page *page)
1359 return PageHWPoison(page) && page_private(page) == MAGIC_HWPOISON;
1362 void SetPageHWPoisonTakenOff(struct page *page)
1364 set_page_private(page, MAGIC_HWPOISON);
1367 void ClearPageHWPoisonTakenOff(struct page *page)
1369 if (PageHWPoison(page))
1370 set_page_private(page, 0);
1374 * Return true if a page type of a given page is supported by hwpoison
1375 * mechanism (while handling could fail), otherwise false. This function
1376 * does not return true for hugetlb or device memory pages, so it's assumed
1377 * to be called only in the context where we never have such pages.
1379 static inline bool HWPoisonHandlable(struct page *page, unsigned long flags)
1381 /* Soft offline could migrate non-LRU movable pages */
1382 if ((flags & MF_SOFT_OFFLINE) && __PageMovable(page))
1385 return PageLRU(page) || is_free_buddy_page(page);
1388 static int __get_hwpoison_page(struct page *page, unsigned long flags)
1390 struct folio *folio = page_folio(page);
1392 bool hugetlb = false;
1394 ret = get_hwpoison_hugetlb_folio(folio, &hugetlb, false);
1399 * This check prevents from calling folio_try_get() for any
1400 * unsupported type of folio in order to reduce the risk of unexpected
1401 * races caused by taking a folio refcount.
1403 if (!HWPoisonHandlable(&folio->page, flags))
1406 if (folio_try_get(folio)) {
1407 if (folio == page_folio(page))
1410 pr_info("%#lx cannot catch tail\n", page_to_pfn(page));
1417 static int get_any_page(struct page *p, unsigned long flags)
1419 int ret = 0, pass = 0;
1420 bool count_increased = false;
1422 if (flags & MF_COUNT_INCREASED)
1423 count_increased = true;
1426 if (!count_increased) {
1427 ret = __get_hwpoison_page(p, flags);
1429 if (page_count(p)) {
1430 /* We raced with an allocation, retry. */
1434 } else if (!PageHuge(p) && !is_free_buddy_page(p)) {
1435 /* We raced with put_page, retry. */
1441 } else if (ret == -EBUSY) {
1443 * We raced with (possibly temporary) unhandlable
1455 if (PageHuge(p) || HWPoisonHandlable(p, flags)) {
1459 * A page we cannot handle. Check whether we can turn
1460 * it into something we can handle.
1465 count_increased = false;
1473 pr_err("%#lx: unhandlable page.\n", page_to_pfn(p));
1478 static int __get_unpoison_page(struct page *page)
1480 struct folio *folio = page_folio(page);
1482 bool hugetlb = false;
1484 ret = get_hwpoison_hugetlb_folio(folio, &hugetlb, true);
1489 * PageHWPoisonTakenOff pages are not only marked as PG_hwpoison,
1490 * but also isolated from buddy freelist, so need to identify the
1491 * state and have to cancel both operations to unpoison.
1493 if (PageHWPoisonTakenOff(page))
1496 return get_page_unless_zero(page) ? 1 : 0;
1500 * get_hwpoison_page() - Get refcount for memory error handling
1501 * @p: Raw error page (hit by memory error)
1502 * @flags: Flags controlling behavior of error handling
1504 * get_hwpoison_page() takes a page refcount of an error page to handle memory
1505 * error on it, after checking that the error page is in a well-defined state
1506 * (defined as a page-type we can successfully handle the memory error on it,
1507 * such as LRU page and hugetlb page).
1509 * Memory error handling could be triggered at any time on any type of page,
1510 * so it's prone to race with typical memory management lifecycle (like
1511 * allocation and free). So to avoid such races, get_hwpoison_page() takes
1512 * extra care for the error page's state (as done in __get_hwpoison_page()),
1513 * and has some retry logic in get_any_page().
1515 * When called from unpoison_memory(), the caller should already ensure that
1516 * the given page has PG_hwpoison. So it's never reused for other page
1517 * allocations, and __get_unpoison_page() never races with them.
1519 * Return: 0 on failure,
1520 * 1 on success for in-use pages in a well-defined state,
1521 * -EIO for pages on which we can not handle memory errors,
1522 * -EBUSY when get_hwpoison_page() has raced with page lifecycle
1523 * operations like allocation and free,
1524 * -EHWPOISON when the page is hwpoisoned and taken off from buddy.
1526 static int get_hwpoison_page(struct page *p, unsigned long flags)
1530 zone_pcp_disable(page_zone(p));
1531 if (flags & MF_UNPOISON)
1532 ret = __get_unpoison_page(p);
1534 ret = get_any_page(p, flags);
1535 zone_pcp_enable(page_zone(p));
1541 * Do all that is necessary to remove user space mappings. Unmap
1542 * the pages and send SIGBUS to the processes if the data was dirty.
1544 static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
1545 int flags, struct page *hpage)
1547 struct folio *folio = page_folio(hpage);
1548 enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC | TTU_HWPOISON;
1549 struct address_space *mapping;
1553 bool mlocked = PageMlocked(hpage);
1556 * Here we are interested only in user-mapped pages, so skip any
1557 * other types of pages.
1559 if (PageReserved(p) || PageSlab(p) || PageTable(p))
1561 if (!(PageLRU(hpage) || PageHuge(p)))
1565 * This check implies we don't kill processes if their pages
1566 * are in the swap cache early. Those are always late kills.
1568 if (!page_mapped(hpage))
1571 if (PageSwapCache(p)) {
1572 pr_err("%#lx: keeping poisoned page in swap cache\n", pfn);
1573 ttu &= ~TTU_HWPOISON;
1577 * Propagate the dirty bit from PTEs to struct page first, because we
1578 * need this to decide if we should kill or just drop the page.
1579 * XXX: the dirty test could be racy: set_page_dirty() may not always
1580 * be called inside page lock (it's recommended but not enforced).
1582 mapping = page_mapping(hpage);
1583 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
1584 mapping_can_writeback(mapping)) {
1585 if (page_mkclean(hpage)) {
1586 SetPageDirty(hpage);
1588 ttu &= ~TTU_HWPOISON;
1589 pr_info("%#lx: corrupted page was clean: dropped without side effects\n",
1595 * First collect all the processes that have the page
1596 * mapped in dirty form. This has to be done before try_to_unmap,
1597 * because ttu takes the rmap data structures down.
1599 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
1601 if (PageHuge(hpage) && !PageAnon(hpage)) {
1603 * For hugetlb pages in shared mappings, try_to_unmap
1604 * could potentially call huge_pmd_unshare. Because of
1605 * this, take semaphore in write mode here and set
1606 * TTU_RMAP_LOCKED to indicate we have taken the lock
1607 * at this higher level.
1609 mapping = hugetlb_page_mapping_lock_write(hpage);
1611 try_to_unmap(folio, ttu|TTU_RMAP_LOCKED);
1612 i_mmap_unlock_write(mapping);
1614 pr_info("%#lx: could not lock mapping for mapped huge page\n", pfn);
1616 try_to_unmap(folio, ttu);
1619 unmap_success = !page_mapped(hpage);
1621 pr_err("%#lx: failed to unmap page (mapcount=%d)\n",
1622 pfn, page_mapcount(hpage));
1625 * try_to_unmap() might put mlocked page in lru cache, so call
1626 * shake_page() again to ensure that it's flushed.
1632 * Now that the dirty bit has been propagated to the
1633 * struct page and all unmaps done we can decide if
1634 * killing is needed or not. Only kill when the page
1635 * was dirty or the process is not restartable,
1636 * otherwise the tokill list is merely
1637 * freed. When there was a problem unmapping earlier
1638 * use a more force-full uncatchable kill to prevent
1639 * any accesses to the poisoned memory.
1641 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL) ||
1643 kill_procs(&tokill, forcekill, !unmap_success, pfn, flags);
1645 return unmap_success;
1648 static int identify_page_state(unsigned long pfn, struct page *p,
1649 unsigned long page_flags)
1651 struct page_state *ps;
1654 * The first check uses the current page flags which may not have any
1655 * relevant information. The second check with the saved page flags is
1656 * carried out only if the first check can't determine the page status.
1658 for (ps = error_states;; ps++)
1659 if ((p->flags & ps->mask) == ps->res)
1662 page_flags |= (p->flags & (1UL << PG_dirty));
1665 for (ps = error_states;; ps++)
1666 if ((page_flags & ps->mask) == ps->res)
1668 return page_action(ps, p, pfn);
1671 static int try_to_split_thp_page(struct page *page)
1676 ret = split_huge_page(page);
1685 static void unmap_and_kill(struct list_head *to_kill, unsigned long pfn,
1686 struct address_space *mapping, pgoff_t index, int flags)
1689 unsigned long size = 0;
1691 list_for_each_entry(tk, to_kill, nd)
1693 size = max(size, 1UL << tk->size_shift);
1697 * Unmap the largest mapping to avoid breaking up device-dax
1698 * mappings which are constant size. The actual size of the
1699 * mapping being torn down is communicated in siginfo, see
1702 loff_t start = (index << PAGE_SHIFT) & ~(size - 1);
1704 unmap_mapping_range(mapping, start, size, 0);
1707 kill_procs(to_kill, flags & MF_MUST_KILL, false, pfn, flags);
1710 static int mf_generic_kill_procs(unsigned long long pfn, int flags,
1711 struct dev_pagemap *pgmap)
1713 struct page *page = pfn_to_page(pfn);
1719 * Pages instantiated by device-dax (not filesystem-dax)
1720 * may be compound pages.
1722 page = compound_head(page);
1725 * Prevent the inode from being freed while we are interrogating
1726 * the address_space, typically this would be handled by
1727 * lock_page(), but dax pages do not use the page lock. This
1728 * also prevents changes to the mapping of this pfn until
1729 * poison signaling is complete.
1731 cookie = dax_lock_page(page);
1735 if (hwpoison_filter(page)) {
1740 switch (pgmap->type) {
1741 case MEMORY_DEVICE_PRIVATE:
1742 case MEMORY_DEVICE_COHERENT:
1744 * TODO: Handle device pages which may need coordination
1745 * with device-side memory.
1754 * Use this flag as an indication that the dax page has been
1755 * remapped UC to prevent speculative consumption of poison.
1757 SetPageHWPoison(page);
1760 * Unlike System-RAM there is no possibility to swap in a
1761 * different physical page at a given virtual address, so all
1762 * userspace consumption of ZONE_DEVICE memory necessitates
1763 * SIGBUS (i.e. MF_MUST_KILL)
1765 flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1766 collect_procs(page, &to_kill, true);
1768 unmap_and_kill(&to_kill, pfn, page->mapping, page->index, flags);
1770 dax_unlock_page(page, cookie);
1774 #ifdef CONFIG_FS_DAX
1776 * mf_dax_kill_procs - Collect and kill processes who are using this file range
1777 * @mapping: address_space of the file in use
1778 * @index: start pgoff of the range within the file
1779 * @count: length of the range, in unit of PAGE_SIZE
1780 * @mf_flags: memory failure flags
1782 int mf_dax_kill_procs(struct address_space *mapping, pgoff_t index,
1783 unsigned long count, int mf_flags)
1788 size_t end = index + count;
1790 mf_flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1792 for (; index < end; index++) {
1794 cookie = dax_lock_mapping_entry(mapping, index, &page);
1800 SetPageHWPoison(page);
1802 collect_procs_fsdax(page, mapping, index, &to_kill);
1803 unmap_and_kill(&to_kill, page_to_pfn(page), mapping,
1806 dax_unlock_mapping_entry(mapping, index, cookie);
1810 EXPORT_SYMBOL_GPL(mf_dax_kill_procs);
1811 #endif /* CONFIG_FS_DAX */
1813 #ifdef CONFIG_HUGETLB_PAGE
1815 * Struct raw_hwp_page represents information about "raw error page",
1816 * constructing singly linked list from ->_hugetlb_hwpoison field of folio.
1818 struct raw_hwp_page {
1819 struct llist_node node;
1823 static inline struct llist_head *raw_hwp_list_head(struct folio *folio)
1825 return (struct llist_head *)&folio->_hugetlb_hwpoison;
1828 static unsigned long __folio_free_raw_hwp(struct folio *folio, bool move_flag)
1830 struct llist_head *head;
1831 struct llist_node *t, *tnode;
1832 unsigned long count = 0;
1834 head = raw_hwp_list_head(folio);
1835 llist_for_each_safe(tnode, t, head->first) {
1836 struct raw_hwp_page *p = container_of(tnode, struct raw_hwp_page, node);
1839 SetPageHWPoison(p->page);
1841 num_poisoned_pages_sub(page_to_pfn(p->page), 1);
1845 llist_del_all(head);
1849 static int folio_set_hugetlb_hwpoison(struct folio *folio, struct page *page)
1851 struct llist_head *head;
1852 struct raw_hwp_page *raw_hwp;
1853 struct llist_node *t, *tnode;
1854 int ret = folio_test_set_hwpoison(folio) ? -EHWPOISON : 0;
1857 * Once the hwpoison hugepage has lost reliable raw error info,
1858 * there is little meaning to keep additional error info precisely,
1859 * so skip to add additional raw error info.
1861 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1863 head = raw_hwp_list_head(folio);
1864 llist_for_each_safe(tnode, t, head->first) {
1865 struct raw_hwp_page *p = container_of(tnode, struct raw_hwp_page, node);
1867 if (p->page == page)
1871 raw_hwp = kmalloc(sizeof(struct raw_hwp_page), GFP_ATOMIC);
1873 raw_hwp->page = page;
1874 llist_add(&raw_hwp->node, head);
1875 /* the first error event will be counted in action_result(). */
1877 num_poisoned_pages_inc(page_to_pfn(page));
1880 * Failed to save raw error info. We no longer trace all
1881 * hwpoisoned subpages, and we need refuse to free/dissolve
1882 * this hwpoisoned hugepage.
1884 folio_set_hugetlb_raw_hwp_unreliable(folio);
1886 * Once hugetlb_raw_hwp_unreliable is set, raw_hwp_page is not
1887 * used any more, so free it.
1889 __folio_free_raw_hwp(folio, false);
1894 static unsigned long folio_free_raw_hwp(struct folio *folio, bool move_flag)
1897 * hugetlb_vmemmap_optimized hugepages can't be freed because struct
1898 * pages for tail pages are required but they don't exist.
1900 if (move_flag && folio_test_hugetlb_vmemmap_optimized(folio))
1904 * hugetlb_raw_hwp_unreliable hugepages shouldn't be unpoisoned by
1907 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1910 return __folio_free_raw_hwp(folio, move_flag);
1913 void folio_clear_hugetlb_hwpoison(struct folio *folio)
1915 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1917 folio_clear_hwpoison(folio);
1918 folio_free_raw_hwp(folio, true);
1922 * Called from hugetlb code with hugetlb_lock held.
1926 * 1 - in-use hugepage
1927 * 2 - not a hugepage
1928 * -EBUSY - the hugepage is busy (try to retry)
1929 * -EHWPOISON - the hugepage is already hwpoisoned
1931 int __get_huge_page_for_hwpoison(unsigned long pfn, int flags,
1932 bool *migratable_cleared)
1934 struct page *page = pfn_to_page(pfn);
1935 struct folio *folio = page_folio(page);
1936 int ret = 2; /* fallback to normal page handling */
1937 bool count_increased = false;
1939 if (!folio_test_hugetlb(folio))
1942 if (flags & MF_COUNT_INCREASED) {
1944 count_increased = true;
1945 } else if (folio_test_hugetlb_freed(folio)) {
1947 } else if (folio_test_hugetlb_migratable(folio)) {
1948 ret = folio_try_get(folio);
1950 count_increased = true;
1953 if (!(flags & MF_NO_RETRY))
1957 if (folio_set_hugetlb_hwpoison(folio, page)) {
1963 * Clearing hugetlb_migratable for hwpoisoned hugepages to prevent them
1964 * from being migrated by memory hotremove.
1966 if (count_increased && folio_test_hugetlb_migratable(folio)) {
1967 folio_clear_hugetlb_migratable(folio);
1968 *migratable_cleared = true;
1973 if (count_increased)
1979 * Taking refcount of hugetlb pages needs extra care about race conditions
1980 * with basic operations like hugepage allocation/free/demotion.
1981 * So some of prechecks for hwpoison (pinning, and testing/setting
1982 * PageHWPoison) should be done in single hugetlb_lock range.
1984 static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb)
1987 struct page *p = pfn_to_page(pfn);
1988 struct folio *folio;
1989 unsigned long page_flags;
1990 bool migratable_cleared = false;
1994 res = get_huge_page_for_hwpoison(pfn, flags, &migratable_cleared);
1995 if (res == 2) { /* fallback to normal page handling */
1998 } else if (res == -EHWPOISON) {
1999 pr_err("%#lx: already hardware poisoned\n", pfn);
2000 if (flags & MF_ACTION_REQUIRED) {
2001 folio = page_folio(p);
2002 res = kill_accessing_process(current, folio_pfn(folio), flags);
2005 } else if (res == -EBUSY) {
2006 if (!(flags & MF_NO_RETRY)) {
2007 flags |= MF_NO_RETRY;
2010 return action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
2013 folio = page_folio(p);
2016 if (hwpoison_filter(p)) {
2017 folio_clear_hugetlb_hwpoison(folio);
2018 if (migratable_cleared)
2019 folio_set_hugetlb_migratable(folio);
2020 folio_unlock(folio);
2027 * Handling free hugepage. The possible race with hugepage allocation
2028 * or demotion can be prevented by PageHWPoison flag.
2031 folio_unlock(folio);
2032 if (__page_handle_poison(p) >= 0) {
2038 return action_result(pfn, MF_MSG_FREE_HUGE, res);
2041 page_flags = folio->flags;
2043 if (!hwpoison_user_mappings(p, pfn, flags, &folio->page)) {
2044 folio_unlock(folio);
2045 return action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
2048 return identify_page_state(pfn, p, page_flags);
2052 static inline int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb)
2057 static inline unsigned long folio_free_raw_hwp(struct folio *folio, bool flag)
2061 #endif /* CONFIG_HUGETLB_PAGE */
2063 /* Drop the extra refcount in case we come from madvise() */
2064 static void put_ref_page(unsigned long pfn, int flags)
2068 if (!(flags & MF_COUNT_INCREASED))
2071 page = pfn_to_page(pfn);
2076 static int memory_failure_dev_pagemap(unsigned long pfn, int flags,
2077 struct dev_pagemap *pgmap)
2081 put_ref_page(pfn, flags);
2083 /* device metadata space is not recoverable */
2084 if (!pgmap_pfn_valid(pgmap, pfn))
2088 * Call driver's implementation to handle the memory failure, otherwise
2089 * fall back to generic handler.
2091 if (pgmap_has_memory_failure(pgmap)) {
2092 rc = pgmap->ops->memory_failure(pgmap, pfn, 1, flags);
2094 * Fall back to generic handler too if operation is not
2095 * supported inside the driver/device/filesystem.
2097 if (rc != -EOPNOTSUPP)
2101 rc = mf_generic_kill_procs(pfn, flags, pgmap);
2103 /* drop pgmap ref acquired in caller */
2104 put_dev_pagemap(pgmap);
2105 action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED);
2109 static DEFINE_MUTEX(mf_mutex);
2112 * memory_failure - Handle memory failure of a page.
2113 * @pfn: Page Number of the corrupted page
2114 * @flags: fine tune action taken
2116 * This function is called by the low level machine check code
2117 * of an architecture when it detects hardware memory corruption
2118 * of a page. It tries its best to recover, which includes
2119 * dropping pages, killing processes etc.
2121 * The function is primarily of use for corruptions that
2122 * happen outside the current execution context (e.g. when
2123 * detected by a background scrubber)
2125 * Must run in process context (e.g. a work queue) with interrupts
2126 * enabled and no spinlocks hold.
2128 * Return: 0 for successfully handled the memory error,
2129 * -EOPNOTSUPP for hwpoison_filter() filtered the error event,
2130 * < 0(except -EOPNOTSUPP) on failure.
2132 int memory_failure(unsigned long pfn, int flags)
2136 struct dev_pagemap *pgmap;
2138 unsigned long page_flags;
2142 if (!sysctl_memory_failure_recovery)
2143 panic("Memory failure on page %lx", pfn);
2145 mutex_lock(&mf_mutex);
2147 if (!(flags & MF_SW_SIMULATED))
2148 hw_memory_failure = true;
2150 p = pfn_to_online_page(pfn);
2152 res = arch_memory_failure(pfn, flags);
2156 if (pfn_valid(pfn)) {
2157 pgmap = get_dev_pagemap(pfn, NULL);
2159 res = memory_failure_dev_pagemap(pfn, flags,
2164 pr_err("%#lx: memory outside kernel control\n", pfn);
2170 res = try_memory_failure_hugetlb(pfn, flags, &hugetlb);
2174 if (TestSetPageHWPoison(p)) {
2175 pr_err("%#lx: already hardware poisoned\n", pfn);
2177 if (flags & MF_ACTION_REQUIRED)
2178 res = kill_accessing_process(current, pfn, flags);
2179 if (flags & MF_COUNT_INCREASED)
2184 hpage = compound_head(p);
2187 * We need/can do nothing about count=0 pages.
2188 * 1) it's a free page, and therefore in safe hand:
2189 * check_new_page() will be the gate keeper.
2190 * 2) it's part of a non-compound high order page.
2191 * Implies some kernel user: cannot stop them from
2192 * R/W the page; let's pray that the page has been
2193 * used and will be freed some time later.
2194 * In fact it's dangerous to directly bump up page count from 0,
2195 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
2197 if (!(flags & MF_COUNT_INCREASED)) {
2198 res = get_hwpoison_page(p, flags);
2200 if (is_free_buddy_page(p)) {
2201 if (take_page_off_buddy(p)) {
2205 /* We lost the race, try again */
2207 ClearPageHWPoison(p);
2213 res = action_result(pfn, MF_MSG_BUDDY, res);
2215 res = action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
2218 } else if (res < 0) {
2219 res = action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
2224 if (PageTransHuge(hpage)) {
2226 * The flag must be set after the refcount is bumped
2227 * otherwise it may race with THP split.
2228 * And the flag can't be set in get_hwpoison_page() since
2229 * it is called by soft offline too and it is just called
2230 * for !MF_COUNT_INCREASE. So here seems to be the best
2233 * Don't need care about the above error handling paths for
2234 * get_hwpoison_page() since they handle either free page
2235 * or unhandlable page. The refcount is bumped iff the
2236 * page is a valid handlable page.
2238 SetPageHasHWPoisoned(hpage);
2239 if (try_to_split_thp_page(p) < 0) {
2240 res = action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED);
2243 VM_BUG_ON_PAGE(!page_count(p), p);
2247 * We ignore non-LRU pages for good reasons.
2248 * - PG_locked is only well defined for LRU pages and a few others
2249 * - to avoid races with __SetPageLocked()
2250 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
2251 * The check (unnecessarily) ignores LRU pages being isolated and
2252 * walked by the page reclaim code, however that's not a big loss.
2259 * We're only intended to deal with the non-Compound page here.
2260 * However, the page could have changed compound pages due to
2261 * race window. If this happens, we could try again to hopefully
2262 * handle the page next round.
2264 if (PageCompound(p)) {
2266 ClearPageHWPoison(p);
2269 flags &= ~MF_COUNT_INCREASED;
2273 res = action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
2278 * We use page flags to determine what action should be taken, but
2279 * the flags can be modified by the error containment action. One
2280 * example is an mlocked page, where PG_mlocked is cleared by
2281 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
2282 * correctly, we save a copy of the page flags at this time.
2284 page_flags = p->flags;
2286 if (hwpoison_filter(p)) {
2287 ClearPageHWPoison(p);
2295 * __munlock_folio() may clear a writeback page's LRU flag without
2296 * page_lock. We need wait writeback completion for this page or it
2297 * may trigger vfs BUG while evict inode.
2299 if (!PageLRU(p) && !PageWriteback(p))
2300 goto identify_page_state;
2303 * It's very difficult to mess with pages currently under IO
2304 * and in many cases impossible, so we just avoid it here.
2306 wait_on_page_writeback(p);
2309 * Now take care of user space mappings.
2310 * Abort on fail: __filemap_remove_folio() assumes unmapped page.
2312 if (!hwpoison_user_mappings(p, pfn, flags, p)) {
2313 res = action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
2318 * Torn down by someone else?
2320 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
2321 res = action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
2325 identify_page_state:
2326 res = identify_page_state(pfn, p, page_flags);
2327 mutex_unlock(&mf_mutex);
2332 mutex_unlock(&mf_mutex);
2335 EXPORT_SYMBOL_GPL(memory_failure);
2337 #define MEMORY_FAILURE_FIFO_ORDER 4
2338 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
2340 struct memory_failure_entry {
2345 struct memory_failure_cpu {
2346 DECLARE_KFIFO(fifo, struct memory_failure_entry,
2347 MEMORY_FAILURE_FIFO_SIZE);
2349 struct work_struct work;
2352 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
2355 * memory_failure_queue - Schedule handling memory failure of a page.
2356 * @pfn: Page Number of the corrupted page
2357 * @flags: Flags for memory failure handling
2359 * This function is called by the low level hardware error handler
2360 * when it detects hardware memory corruption of a page. It schedules
2361 * the recovering of error page, including dropping pages, killing
2364 * The function is primarily of use for corruptions that
2365 * happen outside the current execution context (e.g. when
2366 * detected by a background scrubber)
2368 * Can run in IRQ context.
2370 void memory_failure_queue(unsigned long pfn, int flags)
2372 struct memory_failure_cpu *mf_cpu;
2373 unsigned long proc_flags;
2374 struct memory_failure_entry entry = {
2379 mf_cpu = &get_cpu_var(memory_failure_cpu);
2380 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
2381 if (kfifo_put(&mf_cpu->fifo, entry))
2382 schedule_work_on(smp_processor_id(), &mf_cpu->work);
2384 pr_err("buffer overflow when queuing memory failure at %#lx\n",
2386 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
2387 put_cpu_var(memory_failure_cpu);
2389 EXPORT_SYMBOL_GPL(memory_failure_queue);
2391 static void memory_failure_work_func(struct work_struct *work)
2393 struct memory_failure_cpu *mf_cpu;
2394 struct memory_failure_entry entry = { 0, };
2395 unsigned long proc_flags;
2398 mf_cpu = container_of(work, struct memory_failure_cpu, work);
2400 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
2401 gotten = kfifo_get(&mf_cpu->fifo, &entry);
2402 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
2405 if (entry.flags & MF_SOFT_OFFLINE)
2406 soft_offline_page(entry.pfn, entry.flags);
2408 memory_failure(entry.pfn, entry.flags);
2413 * Process memory_failure work queued on the specified CPU.
2414 * Used to avoid return-to-userspace racing with the memory_failure workqueue.
2416 void memory_failure_queue_kick(int cpu)
2418 struct memory_failure_cpu *mf_cpu;
2420 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
2421 cancel_work_sync(&mf_cpu->work);
2422 memory_failure_work_func(&mf_cpu->work);
2425 static int __init memory_failure_init(void)
2427 struct memory_failure_cpu *mf_cpu;
2430 for_each_possible_cpu(cpu) {
2431 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
2432 spin_lock_init(&mf_cpu->lock);
2433 INIT_KFIFO(mf_cpu->fifo);
2434 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
2437 register_sysctl_init("vm", memory_failure_table);
2441 core_initcall(memory_failure_init);
2444 #define pr_fmt(fmt) "" fmt
2445 #define unpoison_pr_info(fmt, pfn, rs) \
2447 if (__ratelimit(rs)) \
2448 pr_info(fmt, pfn); \
2452 * unpoison_memory - Unpoison a previously poisoned page
2453 * @pfn: Page number of the to be unpoisoned page
2455 * Software-unpoison a page that has been poisoned by
2456 * memory_failure() earlier.
2458 * This is only done on the software-level, so it only works
2459 * for linux injected failures, not real hardware failures
2461 * Returns 0 for success, otherwise -errno.
2463 int unpoison_memory(unsigned long pfn)
2465 struct folio *folio;
2468 unsigned long count = 1;
2470 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
2471 DEFAULT_RATELIMIT_BURST);
2473 if (!pfn_valid(pfn))
2476 p = pfn_to_page(pfn);
2477 folio = page_folio(p);
2479 mutex_lock(&mf_mutex);
2481 if (hw_memory_failure) {
2482 unpoison_pr_info("Unpoison: Disabled after HW memory failure %#lx\n",
2488 if (!folio_test_hwpoison(folio)) {
2489 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
2494 if (folio_ref_count(folio) > 1) {
2495 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
2500 if (folio_mapped(folio)) {
2501 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
2506 if (folio_mapping(folio)) {
2507 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
2512 if (folio_test_slab(folio) || PageTable(&folio->page) || folio_test_reserved(folio))
2515 ret = get_hwpoison_page(p, MF_UNPOISON);
2519 count = folio_free_raw_hwp(folio, false);
2525 ret = folio_test_clear_hwpoison(folio) ? 0 : -EBUSY;
2526 } else if (ret < 0) {
2527 if (ret == -EHWPOISON) {
2528 ret = put_page_back_buddy(p) ? 0 : -EBUSY;
2530 unpoison_pr_info("Unpoison: failed to grab page %#lx\n",
2535 count = folio_free_raw_hwp(folio, false);
2544 if (TestClearPageHWPoison(p)) {
2551 mutex_unlock(&mf_mutex);
2554 num_poisoned_pages_sub(pfn, 1);
2555 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
2556 page_to_pfn(p), &unpoison_rs);
2560 EXPORT_SYMBOL(unpoison_memory);
2562 static bool isolate_page(struct page *page, struct list_head *pagelist)
2564 bool isolated = false;
2566 if (PageHuge(page)) {
2567 isolated = isolate_hugetlb(page_folio(page), pagelist);
2569 bool lru = !__PageMovable(page);
2572 isolated = isolate_lru_page(page);
2574 isolated = isolate_movable_page(page,
2575 ISOLATE_UNEVICTABLE);
2578 list_add(&page->lru, pagelist);
2580 inc_node_page_state(page, NR_ISOLATED_ANON +
2581 page_is_file_lru(page));
2586 * If we succeed to isolate the page, we grabbed another refcount on
2587 * the page, so we can safely drop the one we got from get_any_pages().
2588 * If we failed to isolate the page, it means that we cannot go further
2589 * and we will return an error, so drop the reference we got from
2590 * get_any_pages() as well.
2597 * soft_offline_in_use_page handles hugetlb-pages and non-hugetlb pages.
2598 * If the page is a non-dirty unmapped page-cache page, it simply invalidates.
2599 * If the page is mapped, it migrates the contents over.
2601 static int soft_offline_in_use_page(struct page *page)
2604 unsigned long pfn = page_to_pfn(page);
2605 struct page *hpage = compound_head(page);
2606 char const *msg_page[] = {"page", "hugepage"};
2607 bool huge = PageHuge(page);
2608 LIST_HEAD(pagelist);
2609 struct migration_target_control mtc = {
2610 .nid = NUMA_NO_NODE,
2611 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
2614 if (!huge && PageTransHuge(hpage)) {
2615 if (try_to_split_thp_page(page)) {
2616 pr_info("soft offline: %#lx: thp split failed\n", pfn);
2623 if (!PageHuge(page))
2624 wait_on_page_writeback(page);
2625 if (PageHWPoison(page)) {
2628 pr_info("soft offline: %#lx page already poisoned\n", pfn);
2632 if (!PageHuge(page) && PageLRU(page) && !PageSwapCache(page))
2634 * Try to invalidate first. This should work for
2635 * non dirty unmapped page cache pages.
2637 ret = invalidate_inode_page(page);
2641 pr_info("soft_offline: %#lx: invalidated\n", pfn);
2642 page_handle_poison(page, false, true);
2646 if (isolate_page(hpage, &pagelist)) {
2647 ret = migrate_pages(&pagelist, alloc_migration_target, NULL,
2648 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE, NULL);
2650 bool release = !huge;
2652 if (!page_handle_poison(page, huge, release))
2655 if (!list_empty(&pagelist))
2656 putback_movable_pages(&pagelist);
2658 pr_info("soft offline: %#lx: %s migration failed %ld, type %pGp\n",
2659 pfn, msg_page[huge], ret, &page->flags);
2664 pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %pGp\n",
2665 pfn, msg_page[huge], page_count(page), &page->flags);
2672 * soft_offline_page - Soft offline a page.
2673 * @pfn: pfn to soft-offline
2674 * @flags: flags. Same as memory_failure().
2676 * Returns 0 on success
2677 * -EOPNOTSUPP for hwpoison_filter() filtered the error event
2678 * < 0 otherwise negated errno.
2680 * Soft offline a page, by migration or invalidation,
2681 * without killing anything. This is for the case when
2682 * a page is not corrupted yet (so it's still valid to access),
2683 * but has had a number of corrected errors and is better taken
2686 * The actual policy on when to do that is maintained by
2689 * This should never impact any application or cause data loss,
2690 * however it might take some time.
2692 * This is not a 100% solution for all memory, but tries to be
2693 * ``good enough'' for the majority of memory.
2695 int soft_offline_page(unsigned long pfn, int flags)
2698 bool try_again = true;
2701 if (!pfn_valid(pfn)) {
2702 WARN_ON_ONCE(flags & MF_COUNT_INCREASED);
2706 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
2707 page = pfn_to_online_page(pfn);
2709 put_ref_page(pfn, flags);
2713 mutex_lock(&mf_mutex);
2715 if (PageHWPoison(page)) {
2716 pr_info("%s: %#lx page already poisoned\n", __func__, pfn);
2717 put_ref_page(pfn, flags);
2718 mutex_unlock(&mf_mutex);
2724 ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE);
2727 if (hwpoison_filter(page)) {
2731 mutex_unlock(&mf_mutex);
2736 ret = soft_offline_in_use_page(page);
2737 } else if (ret == 0) {
2738 if (!page_handle_poison(page, true, false) && try_again) {
2740 flags &= ~MF_COUNT_INCREASED;
2745 mutex_unlock(&mf_mutex);