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
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>
67 #include "ras/ras_event.h"
69 int sysctl_memory_failure_early_kill __read_mostly
= 0;
71 int sysctl_memory_failure_recovery __read_mostly
= 1;
73 atomic_long_t num_poisoned_pages __read_mostly
= ATOMIC_LONG_INIT(0);
75 static bool hw_memory_failure __read_mostly
= false;
79 * 1: the page is dissolved (if needed) and taken off from buddy,
80 * 0: the page is dissolved (if needed) and not taken off from buddy,
81 * < 0: failed to dissolve.
83 static int __page_handle_poison(struct page
*page
)
87 zone_pcp_disable(page_zone(page
));
88 ret
= dissolve_free_huge_page(page
);
90 ret
= take_page_off_buddy(page
);
91 zone_pcp_enable(page_zone(page
));
96 static bool page_handle_poison(struct page
*page
, bool hugepage_or_freepage
, bool release
)
98 if (hugepage_or_freepage
) {
100 * Doing this check for free pages is also fine since dissolve_free_huge_page
101 * returns 0 for non-hugetlb pages as well.
103 if (__page_handle_poison(page
) <= 0)
105 * We could fail to take off the target page from buddy
106 * for example due to racy page allocation, but that's
107 * acceptable because soft-offlined page is not broken
108 * and if someone really want to use it, they should
114 SetPageHWPoison(page
);
118 num_poisoned_pages_inc();
123 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
125 u32 hwpoison_filter_enable
= 0;
126 u32 hwpoison_filter_dev_major
= ~0U;
127 u32 hwpoison_filter_dev_minor
= ~0U;
128 u64 hwpoison_filter_flags_mask
;
129 u64 hwpoison_filter_flags_value
;
130 EXPORT_SYMBOL_GPL(hwpoison_filter_enable
);
131 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major
);
132 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor
);
133 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask
);
134 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value
);
136 static int hwpoison_filter_dev(struct page
*p
)
138 struct address_space
*mapping
;
141 if (hwpoison_filter_dev_major
== ~0U &&
142 hwpoison_filter_dev_minor
== ~0U)
145 mapping
= page_mapping(p
);
146 if (mapping
== NULL
|| mapping
->host
== NULL
)
149 dev
= mapping
->host
->i_sb
->s_dev
;
150 if (hwpoison_filter_dev_major
!= ~0U &&
151 hwpoison_filter_dev_major
!= MAJOR(dev
))
153 if (hwpoison_filter_dev_minor
!= ~0U &&
154 hwpoison_filter_dev_minor
!= MINOR(dev
))
160 static int hwpoison_filter_flags(struct page
*p
)
162 if (!hwpoison_filter_flags_mask
)
165 if ((stable_page_flags(p
) & hwpoison_filter_flags_mask
) ==
166 hwpoison_filter_flags_value
)
173 * This allows stress tests to limit test scope to a collection of tasks
174 * by putting them under some memcg. This prevents killing unrelated/important
175 * processes such as /sbin/init. Note that the target task may share clean
176 * pages with init (eg. libc text), which is harmless. If the target task
177 * share _dirty_ pages with another task B, the test scheme must make sure B
178 * is also included in the memcg. At last, due to race conditions this filter
179 * can only guarantee that the page either belongs to the memcg tasks, or is
183 u64 hwpoison_filter_memcg
;
184 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg
);
185 static int hwpoison_filter_task(struct page
*p
)
187 if (!hwpoison_filter_memcg
)
190 if (page_cgroup_ino(p
) != hwpoison_filter_memcg
)
196 static int hwpoison_filter_task(struct page
*p
) { return 0; }
199 int hwpoison_filter(struct page
*p
)
201 if (!hwpoison_filter_enable
)
204 if (hwpoison_filter_dev(p
))
207 if (hwpoison_filter_flags(p
))
210 if (hwpoison_filter_task(p
))
216 int hwpoison_filter(struct page
*p
)
222 EXPORT_SYMBOL_GPL(hwpoison_filter
);
225 * Kill all processes that have a poisoned page mapped and then isolate
229 * Find all processes having the page mapped and kill them.
230 * But we keep a page reference around so that the page is not
231 * actually freed yet.
232 * Then stash the page away
234 * There's no convenient way to get back to mapped processes
235 * from the VMAs. So do a brute-force search over all
238 * Remember that machine checks are not common (or rather
239 * if they are common you have other problems), so this shouldn't
240 * be a performance issue.
242 * Also there are some races possible while we get from the
243 * error detection to actually handle it.
248 struct task_struct
*tsk
;
254 * Send all the processes who have the page mapped a signal.
255 * ``action optional'' if they are not immediately affected by the error
256 * ``action required'' if error happened in current execution context
258 static int kill_proc(struct to_kill
*tk
, unsigned long pfn
, int flags
)
260 struct task_struct
*t
= tk
->tsk
;
261 short addr_lsb
= tk
->size_shift
;
264 pr_err("%#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
265 pfn
, t
->comm
, t
->pid
);
267 if ((flags
& MF_ACTION_REQUIRED
) && (t
== current
))
268 ret
= force_sig_mceerr(BUS_MCEERR_AR
,
269 (void __user
*)tk
->addr
, addr_lsb
);
272 * Signal other processes sharing the page if they have
274 * Don't use force here, it's convenient if the signal
275 * can be temporarily blocked.
276 * This could cause a loop when the user sets SIGBUS
277 * to SIG_IGN, but hopefully no one will do that?
279 ret
= send_sig_mceerr(BUS_MCEERR_AO
, (void __user
*)tk
->addr
,
280 addr_lsb
, t
); /* synchronous? */
282 pr_info("Error sending signal to %s:%d: %d\n",
283 t
->comm
, t
->pid
, ret
);
288 * Unknown page type encountered. Try to check whether it can turn PageLRU by
291 void shake_page(struct page
*p
)
298 if (PageLRU(p
) || is_free_buddy_page(p
))
303 * TODO: Could shrink slab caches here if a lightweight range-based
304 * shrinker will be available.
307 EXPORT_SYMBOL_GPL(shake_page
);
309 static unsigned long dev_pagemap_mapping_shift(struct vm_area_struct
*vma
,
310 unsigned long address
)
312 unsigned long ret
= 0;
319 VM_BUG_ON_VMA(address
== -EFAULT
, vma
);
320 pgd
= pgd_offset(vma
->vm_mm
, address
);
321 if (!pgd_present(*pgd
))
323 p4d
= p4d_offset(pgd
, address
);
324 if (!p4d_present(*p4d
))
326 pud
= pud_offset(p4d
, address
);
327 if (!pud_present(*pud
))
329 if (pud_devmap(*pud
))
331 pmd
= pmd_offset(pud
, address
);
332 if (!pmd_present(*pmd
))
334 if (pmd_devmap(*pmd
))
336 pte
= pte_offset_map(pmd
, address
);
337 if (pte_present(*pte
) && pte_devmap(*pte
))
344 * Failure handling: if we can't find or can't kill a process there's
345 * not much we can do. We just print a message and ignore otherwise.
349 * Schedule a process for later kill.
350 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
352 * Notice: @fsdax_pgoff is used only when @p is a fsdax page.
353 * In other cases, such as anonymous and file-backend page, the address to be
354 * killed can be caculated by @p itself.
356 static void add_to_kill(struct task_struct
*tsk
, struct page
*p
,
357 pgoff_t fsdax_pgoff
, struct vm_area_struct
*vma
,
358 struct list_head
*to_kill
)
362 tk
= kmalloc(sizeof(struct to_kill
), GFP_ATOMIC
);
364 pr_err("Out of memory while machine check handling\n");
368 tk
->addr
= page_address_in_vma(p
, vma
);
369 if (is_zone_device_page(p
)) {
371 * Since page->mapping is not used for fsdax, we need
372 * calculate the address based on the vma.
374 if (p
->pgmap
->type
== MEMORY_DEVICE_FS_DAX
)
375 tk
->addr
= vma_pgoff_address(fsdax_pgoff
, 1, vma
);
376 tk
->size_shift
= dev_pagemap_mapping_shift(vma
, tk
->addr
);
378 tk
->size_shift
= page_shift(compound_head(p
));
381 * Send SIGKILL if "tk->addr == -EFAULT". Also, as
382 * "tk->size_shift" is always non-zero for !is_zone_device_page(),
383 * so "tk->size_shift == 0" effectively checks no mapping on
384 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times
385 * to a process' address space, it's possible not all N VMAs
386 * contain mappings for the page, but at least one VMA does.
387 * Only deliver SIGBUS with payload derived from the VMA that
388 * has a mapping for the page.
390 if (tk
->addr
== -EFAULT
) {
391 pr_info("Unable to find user space address %lx in %s\n",
392 page_to_pfn(p
), tsk
->comm
);
393 } else if (tk
->size_shift
== 0) {
398 get_task_struct(tsk
);
400 list_add_tail(&tk
->nd
, to_kill
);
404 * Kill the processes that have been collected earlier.
406 * Only do anything when FORCEKILL is set, otherwise just free the
407 * list (this is used for clean pages which do not need killing)
408 * Also when FAIL is set do a force kill because something went
411 static void kill_procs(struct list_head
*to_kill
, int forcekill
, bool fail
,
412 unsigned long pfn
, int flags
)
414 struct to_kill
*tk
, *next
;
416 list_for_each_entry_safe (tk
, next
, to_kill
, nd
) {
419 * In case something went wrong with munmapping
420 * make sure the process doesn't catch the
421 * signal and then access the memory. Just kill it.
423 if (fail
|| tk
->addr
== -EFAULT
) {
424 pr_err("%#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
425 pfn
, tk
->tsk
->comm
, tk
->tsk
->pid
);
426 do_send_sig_info(SIGKILL
, SEND_SIG_PRIV
,
427 tk
->tsk
, PIDTYPE_PID
);
431 * In theory the process could have mapped
432 * something else on the address in-between. We could
433 * check for that, but we need to tell the
436 else if (kill_proc(tk
, pfn
, flags
) < 0)
437 pr_err("%#lx: Cannot send advisory machine check signal to %s:%d\n",
438 pfn
, tk
->tsk
->comm
, tk
->tsk
->pid
);
440 put_task_struct(tk
->tsk
);
446 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
447 * on behalf of the thread group. Return task_struct of the (first found)
448 * dedicated thread if found, and return NULL otherwise.
450 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
451 * have to call rcu_read_lock/unlock() in this function.
453 static struct task_struct
*find_early_kill_thread(struct task_struct
*tsk
)
455 struct task_struct
*t
;
457 for_each_thread(tsk
, t
) {
458 if (t
->flags
& PF_MCE_PROCESS
) {
459 if (t
->flags
& PF_MCE_EARLY
)
462 if (sysctl_memory_failure_early_kill
)
470 * Determine whether a given process is "early kill" process which expects
471 * to be signaled when some page under the process is hwpoisoned.
472 * Return task_struct of the dedicated thread (main thread unless explicitly
473 * specified) if the process is "early kill" and otherwise returns NULL.
475 * Note that the above is true for Action Optional case. For Action Required
476 * case, it's only meaningful to the current thread which need to be signaled
477 * with SIGBUS, this error is Action Optional for other non current
478 * processes sharing the same error page,if the process is "early kill", the
479 * task_struct of the dedicated thread will also be returned.
481 static struct task_struct
*task_early_kill(struct task_struct
*tsk
,
487 * Comparing ->mm here because current task might represent
488 * a subthread, while tsk always points to the main thread.
490 if (force_early
&& tsk
->mm
== current
->mm
)
493 return find_early_kill_thread(tsk
);
497 * Collect processes when the error hit an anonymous page.
499 static void collect_procs_anon(struct page
*page
, struct list_head
*to_kill
,
502 struct folio
*folio
= page_folio(page
);
503 struct vm_area_struct
*vma
;
504 struct task_struct
*tsk
;
508 av
= folio_lock_anon_vma_read(folio
, NULL
);
509 if (av
== NULL
) /* Not actually mapped anymore */
512 pgoff
= page_to_pgoff(page
);
513 read_lock(&tasklist_lock
);
514 for_each_process (tsk
) {
515 struct anon_vma_chain
*vmac
;
516 struct task_struct
*t
= task_early_kill(tsk
, force_early
);
520 anon_vma_interval_tree_foreach(vmac
, &av
->rb_root
,
523 if (!page_mapped_in_vma(page
, vma
))
525 if (vma
->vm_mm
== t
->mm
)
526 add_to_kill(t
, page
, 0, vma
, to_kill
);
529 read_unlock(&tasklist_lock
);
530 page_unlock_anon_vma_read(av
);
534 * Collect processes when the error hit a file mapped page.
536 static void collect_procs_file(struct page
*page
, struct list_head
*to_kill
,
539 struct vm_area_struct
*vma
;
540 struct task_struct
*tsk
;
541 struct address_space
*mapping
= page
->mapping
;
544 i_mmap_lock_read(mapping
);
545 read_lock(&tasklist_lock
);
546 pgoff
= page_to_pgoff(page
);
547 for_each_process(tsk
) {
548 struct task_struct
*t
= task_early_kill(tsk
, force_early
);
552 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
,
555 * Send early kill signal to tasks where a vma covers
556 * the page but the corrupted page is not necessarily
557 * mapped it in its pte.
558 * Assume applications who requested early kill want
559 * to be informed of all such data corruptions.
561 if (vma
->vm_mm
== t
->mm
)
562 add_to_kill(t
, page
, 0, vma
, to_kill
);
565 read_unlock(&tasklist_lock
);
566 i_mmap_unlock_read(mapping
);
571 * Collect processes when the error hit a fsdax page.
573 static void collect_procs_fsdax(struct page
*page
,
574 struct address_space
*mapping
, pgoff_t pgoff
,
575 struct list_head
*to_kill
)
577 struct vm_area_struct
*vma
;
578 struct task_struct
*tsk
;
580 i_mmap_lock_read(mapping
);
581 read_lock(&tasklist_lock
);
582 for_each_process(tsk
) {
583 struct task_struct
*t
= task_early_kill(tsk
, true);
587 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
588 if (vma
->vm_mm
== t
->mm
)
589 add_to_kill(t
, page
, pgoff
, vma
, to_kill
);
592 read_unlock(&tasklist_lock
);
593 i_mmap_unlock_read(mapping
);
595 #endif /* CONFIG_FS_DAX */
598 * Collect the processes who have the corrupted page mapped to kill.
600 static void collect_procs(struct page
*page
, struct list_head
*tokill
,
607 collect_procs_anon(page
, tokill
, force_early
);
609 collect_procs_file(page
, tokill
, force_early
);
618 static void set_to_kill(struct to_kill
*tk
, unsigned long addr
, short shift
)
621 tk
->size_shift
= shift
;
624 static int check_hwpoisoned_entry(pte_t pte
, unsigned long addr
, short shift
,
625 unsigned long poisoned_pfn
, struct to_kill
*tk
)
627 unsigned long pfn
= 0;
629 if (pte_present(pte
)) {
632 swp_entry_t swp
= pte_to_swp_entry(pte
);
634 if (is_hwpoison_entry(swp
))
635 pfn
= hwpoison_entry_to_pfn(swp
);
638 if (!pfn
|| pfn
!= poisoned_pfn
)
641 set_to_kill(tk
, addr
, shift
);
645 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
646 static int check_hwpoisoned_pmd_entry(pmd_t
*pmdp
, unsigned long addr
,
647 struct hwp_walk
*hwp
)
651 unsigned long hwpoison_vaddr
;
653 if (!pmd_present(pmd
))
656 if (pfn
<= hwp
->pfn
&& hwp
->pfn
< pfn
+ HPAGE_PMD_NR
) {
657 hwpoison_vaddr
= addr
+ ((hwp
->pfn
- pfn
) << PAGE_SHIFT
);
658 set_to_kill(&hwp
->tk
, hwpoison_vaddr
, PAGE_SHIFT
);
664 static int check_hwpoisoned_pmd_entry(pmd_t
*pmdp
, unsigned long addr
,
665 struct hwp_walk
*hwp
)
671 static int hwpoison_pte_range(pmd_t
*pmdp
, unsigned long addr
,
672 unsigned long end
, struct mm_walk
*walk
)
674 struct hwp_walk
*hwp
= walk
->private;
676 pte_t
*ptep
, *mapped_pte
;
679 ptl
= pmd_trans_huge_lock(pmdp
, walk
->vma
);
681 ret
= check_hwpoisoned_pmd_entry(pmdp
, addr
, hwp
);
686 if (pmd_trans_unstable(pmdp
))
689 mapped_pte
= ptep
= pte_offset_map_lock(walk
->vma
->vm_mm
, pmdp
,
691 for (; addr
!= end
; ptep
++, addr
+= PAGE_SIZE
) {
692 ret
= check_hwpoisoned_entry(*ptep
, addr
, PAGE_SHIFT
,
697 pte_unmap_unlock(mapped_pte
, ptl
);
703 #ifdef CONFIG_HUGETLB_PAGE
704 static int hwpoison_hugetlb_range(pte_t
*ptep
, unsigned long hmask
,
705 unsigned long addr
, unsigned long end
,
706 struct mm_walk
*walk
)
708 struct hwp_walk
*hwp
= walk
->private;
709 pte_t pte
= huge_ptep_get(ptep
);
710 struct hstate
*h
= hstate_vma(walk
->vma
);
712 return check_hwpoisoned_entry(pte
, addr
, huge_page_shift(h
),
716 #define hwpoison_hugetlb_range NULL
719 static const struct mm_walk_ops hwp_walk_ops
= {
720 .pmd_entry
= hwpoison_pte_range
,
721 .hugetlb_entry
= hwpoison_hugetlb_range
,
725 * Sends SIGBUS to the current process with error info.
727 * This function is intended to handle "Action Required" MCEs on already
728 * hardware poisoned pages. They could happen, for example, when
729 * memory_failure() failed to unmap the error page at the first call, or
730 * when multiple local machine checks happened on different CPUs.
732 * MCE handler currently has no easy access to the error virtual address,
733 * so this function walks page table to find it. The returned virtual address
734 * is proper in most cases, but it could be wrong when the application
735 * process has multiple entries mapping the error page.
737 static int kill_accessing_process(struct task_struct
*p
, unsigned long pfn
,
741 struct hwp_walk priv
= {
746 mmap_read_lock(p
->mm
);
747 ret
= walk_page_range(p
->mm
, 0, TASK_SIZE
, &hwp_walk_ops
,
749 if (ret
== 1 && priv
.tk
.addr
)
750 kill_proc(&priv
.tk
, pfn
, flags
);
753 mmap_read_unlock(p
->mm
);
754 return ret
> 0 ? -EHWPOISON
: -EFAULT
;
757 static const char *action_name
[] = {
758 [MF_IGNORED
] = "Ignored",
759 [MF_FAILED
] = "Failed",
760 [MF_DELAYED
] = "Delayed",
761 [MF_RECOVERED
] = "Recovered",
764 static const char * const action_page_types
[] = {
765 [MF_MSG_KERNEL
] = "reserved kernel page",
766 [MF_MSG_KERNEL_HIGH_ORDER
] = "high-order kernel page",
767 [MF_MSG_SLAB
] = "kernel slab page",
768 [MF_MSG_DIFFERENT_COMPOUND
] = "different compound page after locking",
769 [MF_MSG_HUGE
] = "huge page",
770 [MF_MSG_FREE_HUGE
] = "free huge page",
771 [MF_MSG_UNMAP_FAILED
] = "unmapping failed page",
772 [MF_MSG_DIRTY_SWAPCACHE
] = "dirty swapcache page",
773 [MF_MSG_CLEAN_SWAPCACHE
] = "clean swapcache page",
774 [MF_MSG_DIRTY_MLOCKED_LRU
] = "dirty mlocked LRU page",
775 [MF_MSG_CLEAN_MLOCKED_LRU
] = "clean mlocked LRU page",
776 [MF_MSG_DIRTY_UNEVICTABLE_LRU
] = "dirty unevictable LRU page",
777 [MF_MSG_CLEAN_UNEVICTABLE_LRU
] = "clean unevictable LRU page",
778 [MF_MSG_DIRTY_LRU
] = "dirty LRU page",
779 [MF_MSG_CLEAN_LRU
] = "clean LRU page",
780 [MF_MSG_TRUNCATED_LRU
] = "already truncated LRU page",
781 [MF_MSG_BUDDY
] = "free buddy page",
782 [MF_MSG_DAX
] = "dax page",
783 [MF_MSG_UNSPLIT_THP
] = "unsplit thp",
784 [MF_MSG_UNKNOWN
] = "unknown page",
788 * XXX: It is possible that a page is isolated from LRU cache,
789 * and then kept in swap cache or failed to remove from page cache.
790 * The page count will stop it from being freed by unpoison.
791 * Stress tests should be aware of this memory leak problem.
793 static int delete_from_lru_cache(struct page
*p
)
795 if (!isolate_lru_page(p
)) {
797 * Clear sensible page flags, so that the buddy system won't
798 * complain when the page is unpoison-and-freed.
801 ClearPageUnevictable(p
);
804 * Poisoned page might never drop its ref count to 0 so we have
805 * to uncharge it manually from its memcg.
807 mem_cgroup_uncharge(page_folio(p
));
810 * drop the page count elevated by isolate_lru_page()
818 static int truncate_error_page(struct page
*p
, unsigned long pfn
,
819 struct address_space
*mapping
)
823 if (mapping
->a_ops
->error_remove_page
) {
824 int err
= mapping
->a_ops
->error_remove_page(mapping
, p
);
827 pr_info("%#lx: Failed to punch page: %d\n", pfn
, err
);
828 } else if (page_has_private(p
) &&
829 !try_to_release_page(p
, GFP_NOIO
)) {
830 pr_info("%#lx: failed to release buffers\n", pfn
);
836 * If the file system doesn't support it just invalidate
837 * This fails on dirty or anything with private pages
839 if (invalidate_inode_page(p
))
842 pr_info("%#lx: Failed to invalidate\n", pfn
);
851 enum mf_action_page_type type
;
853 /* Callback ->action() has to unlock the relevant page inside it. */
854 int (*action
)(struct page_state
*ps
, struct page
*p
);
858 * Return true if page is still referenced by others, otherwise return
861 * The extra_pins is true when one extra refcount is expected.
863 static bool has_extra_refcount(struct page_state
*ps
, struct page
*p
,
866 int count
= page_count(p
) - 1;
872 pr_err("%#lx: %s still referenced by %d users\n",
873 page_to_pfn(p
), action_page_types
[ps
->type
], count
);
881 * Error hit kernel page.
882 * Do nothing, try to be lucky and not touch this instead. For a few cases we
883 * could be more sophisticated.
885 static int me_kernel(struct page_state
*ps
, struct page
*p
)
892 * Page in unknown state. Do nothing.
894 static int me_unknown(struct page_state
*ps
, struct page
*p
)
896 pr_err("%#lx: Unknown page state\n", page_to_pfn(p
));
902 * Clean (or cleaned) page cache page.
904 static int me_pagecache_clean(struct page_state
*ps
, struct page
*p
)
907 struct address_space
*mapping
;
910 delete_from_lru_cache(p
);
913 * For anonymous pages we're done the only reference left
914 * should be the one m_f() holds.
922 * Now truncate the page in the page cache. This is really
923 * more like a "temporary hole punch"
924 * Don't do this for block devices when someone else
925 * has a reference, because it could be file system metadata
926 * and that's not safe to truncate.
928 mapping
= page_mapping(p
);
931 * Page has been teared down in the meanwhile
938 * The shmem page is kept in page cache instead of truncating
939 * so is expected to have an extra refcount after error-handling.
941 extra_pins
= shmem_mapping(mapping
);
944 * Truncation is a bit tricky. Enable it per file system for now.
946 * Open: to take i_rwsem or not for this? Right now we don't.
948 ret
= truncate_error_page(p
, page_to_pfn(p
), mapping
);
949 if (has_extra_refcount(ps
, p
, extra_pins
))
959 * Dirty pagecache page
960 * Issues: when the error hit a hole page the error is not properly
963 static int me_pagecache_dirty(struct page_state
*ps
, struct page
*p
)
965 struct address_space
*mapping
= page_mapping(p
);
968 /* TBD: print more information about the file. */
971 * IO error will be reported by write(), fsync(), etc.
972 * who check the mapping.
973 * This way the application knows that something went
974 * wrong with its dirty file data.
976 * There's one open issue:
978 * The EIO will be only reported on the next IO
979 * operation and then cleared through the IO map.
980 * Normally Linux has two mechanisms to pass IO error
981 * first through the AS_EIO flag in the address space
982 * and then through the PageError flag in the page.
983 * Since we drop pages on memory failure handling the
984 * only mechanism open to use is through AS_AIO.
986 * This has the disadvantage that it gets cleared on
987 * the first operation that returns an error, while
988 * the PageError bit is more sticky and only cleared
989 * when the page is reread or dropped. If an
990 * application assumes it will always get error on
991 * fsync, but does other operations on the fd before
992 * and the page is dropped between then the error
993 * will not be properly reported.
995 * This can already happen even without hwpoisoned
996 * pages: first on metadata IO errors (which only
997 * report through AS_EIO) or when the page is dropped
1000 * So right now we assume that the application DTRT on
1001 * the first EIO, but we're not worse than other parts
1004 mapping_set_error(mapping
, -EIO
);
1007 return me_pagecache_clean(ps
, p
);
1011 * Clean and dirty swap cache.
1013 * Dirty swap cache page is tricky to handle. The page could live both in page
1014 * cache and swap cache(ie. page is freshly swapped in). So it could be
1015 * referenced concurrently by 2 types of PTEs:
1016 * normal PTEs and swap PTEs. We try to handle them consistently by calling
1017 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
1019 * - clear dirty bit to prevent IO
1021 * - but keep in the swap cache, so that when we return to it on
1022 * a later page fault, we know the application is accessing
1023 * corrupted data and shall be killed (we installed simple
1024 * interception code in do_swap_page to catch it).
1026 * Clean swap cache pages can be directly isolated. A later page fault will
1027 * bring in the known good data from disk.
1029 static int me_swapcache_dirty(struct page_state
*ps
, struct page
*p
)
1032 bool extra_pins
= false;
1035 /* Trigger EIO in shmem: */
1036 ClearPageUptodate(p
);
1038 ret
= delete_from_lru_cache(p
) ? MF_FAILED
: MF_DELAYED
;
1041 if (ret
== MF_DELAYED
)
1044 if (has_extra_refcount(ps
, p
, extra_pins
))
1050 static int me_swapcache_clean(struct page_state
*ps
, struct page
*p
)
1052 struct folio
*folio
= page_folio(p
);
1055 delete_from_swap_cache(folio
);
1057 ret
= delete_from_lru_cache(p
) ? MF_FAILED
: MF_RECOVERED
;
1058 folio_unlock(folio
);
1060 if (has_extra_refcount(ps
, p
, false))
1067 * Huge pages. Needs work.
1069 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
1070 * To narrow down kill region to one page, we need to break up pmd.
1072 static int me_huge_page(struct page_state
*ps
, struct page
*p
)
1075 struct page
*hpage
= compound_head(p
);
1076 struct address_space
*mapping
;
1078 if (!PageHuge(hpage
))
1081 mapping
= page_mapping(hpage
);
1083 res
= truncate_error_page(hpage
, page_to_pfn(p
), mapping
);
1088 * migration entry prevents later access on error hugepage,
1089 * so we can free and dissolve it into buddy to save healthy
1093 if (__page_handle_poison(p
) >= 0) {
1101 if (has_extra_refcount(ps
, p
, false))
1108 * Various page states we can handle.
1110 * A page state is defined by its current page->flags bits.
1111 * The table matches them in order and calls the right handler.
1113 * This is quite tricky because we can access page at any time
1114 * in its live cycle, so all accesses have to be extremely careful.
1116 * This is not complete. More states could be added.
1117 * For any missing state don't attempt recovery.
1120 #define dirty (1UL << PG_dirty)
1121 #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked))
1122 #define unevict (1UL << PG_unevictable)
1123 #define mlock (1UL << PG_mlocked)
1124 #define lru (1UL << PG_lru)
1125 #define head (1UL << PG_head)
1126 #define slab (1UL << PG_slab)
1127 #define reserved (1UL << PG_reserved)
1129 static struct page_state error_states
[] = {
1130 { reserved
, reserved
, MF_MSG_KERNEL
, me_kernel
},
1132 * free pages are specially detected outside this table:
1133 * PG_buddy pages only make a small fraction of all free pages.
1137 * Could in theory check if slab page is free or if we can drop
1138 * currently unused objects without touching them. But just
1139 * treat it as standard kernel for now.
1141 { slab
, slab
, MF_MSG_SLAB
, me_kernel
},
1143 { head
, head
, MF_MSG_HUGE
, me_huge_page
},
1145 { sc
|dirty
, sc
|dirty
, MF_MSG_DIRTY_SWAPCACHE
, me_swapcache_dirty
},
1146 { sc
|dirty
, sc
, MF_MSG_CLEAN_SWAPCACHE
, me_swapcache_clean
},
1148 { mlock
|dirty
, mlock
|dirty
, MF_MSG_DIRTY_MLOCKED_LRU
, me_pagecache_dirty
},
1149 { mlock
|dirty
, mlock
, MF_MSG_CLEAN_MLOCKED_LRU
, me_pagecache_clean
},
1151 { unevict
|dirty
, unevict
|dirty
, MF_MSG_DIRTY_UNEVICTABLE_LRU
, me_pagecache_dirty
},
1152 { unevict
|dirty
, unevict
, MF_MSG_CLEAN_UNEVICTABLE_LRU
, me_pagecache_clean
},
1154 { lru
|dirty
, lru
|dirty
, MF_MSG_DIRTY_LRU
, me_pagecache_dirty
},
1155 { lru
|dirty
, lru
, MF_MSG_CLEAN_LRU
, me_pagecache_clean
},
1158 * Catchall entry: must be at end.
1160 { 0, 0, MF_MSG_UNKNOWN
, me_unknown
},
1173 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
1174 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
1176 static void action_result(unsigned long pfn
, enum mf_action_page_type type
,
1177 enum mf_result result
)
1179 trace_memory_failure_event(pfn
, type
, result
);
1181 num_poisoned_pages_inc();
1182 pr_err("%#lx: recovery action for %s: %s\n",
1183 pfn
, action_page_types
[type
], action_name
[result
]);
1186 static int page_action(struct page_state
*ps
, struct page
*p
,
1191 /* page p should be unlocked after returning from ps->action(). */
1192 result
= ps
->action(ps
, p
);
1194 action_result(pfn
, ps
->type
, result
);
1196 /* Could do more checks here if page looks ok */
1198 * Could adjust zone counters here to correct for the missing page.
1201 return (result
== MF_RECOVERED
|| result
== MF_DELAYED
) ? 0 : -EBUSY
;
1204 static inline bool PageHWPoisonTakenOff(struct page
*page
)
1206 return PageHWPoison(page
) && page_private(page
) == MAGIC_HWPOISON
;
1209 void SetPageHWPoisonTakenOff(struct page
*page
)
1211 set_page_private(page
, MAGIC_HWPOISON
);
1214 void ClearPageHWPoisonTakenOff(struct page
*page
)
1216 if (PageHWPoison(page
))
1217 set_page_private(page
, 0);
1221 * Return true if a page type of a given page is supported by hwpoison
1222 * mechanism (while handling could fail), otherwise false. This function
1223 * does not return true for hugetlb or device memory pages, so it's assumed
1224 * to be called only in the context where we never have such pages.
1226 static inline bool HWPoisonHandlable(struct page
*page
, unsigned long flags
)
1228 /* Soft offline could migrate non-LRU movable pages */
1229 if ((flags
& MF_SOFT_OFFLINE
) && __PageMovable(page
))
1232 return PageLRU(page
) || is_free_buddy_page(page
);
1235 static int __get_hwpoison_page(struct page
*page
, unsigned long flags
)
1237 struct page
*head
= compound_head(page
);
1239 bool hugetlb
= false;
1241 ret
= get_hwpoison_huge_page(head
, &hugetlb
);
1246 * This check prevents from calling get_hwpoison_unless_zero()
1247 * for any unsupported type of page in order to reduce the risk of
1248 * unexpected races caused by taking a page refcount.
1250 if (!HWPoisonHandlable(head
, flags
))
1253 if (get_page_unless_zero(head
)) {
1254 if (head
== compound_head(page
))
1257 pr_info("%#lx cannot catch tail\n", page_to_pfn(page
));
1264 static int get_any_page(struct page
*p
, unsigned long flags
)
1266 int ret
= 0, pass
= 0;
1267 bool count_increased
= false;
1269 if (flags
& MF_COUNT_INCREASED
)
1270 count_increased
= true;
1273 if (!count_increased
) {
1274 ret
= __get_hwpoison_page(p
, flags
);
1276 if (page_count(p
)) {
1277 /* We raced with an allocation, retry. */
1281 } else if (!PageHuge(p
) && !is_free_buddy_page(p
)) {
1282 /* We raced with put_page, retry. */
1288 } else if (ret
== -EBUSY
) {
1290 * We raced with (possibly temporary) unhandlable
1302 if (PageHuge(p
) || HWPoisonHandlable(p
, flags
)) {
1306 * A page we cannot handle. Check whether we can turn
1307 * it into something we can handle.
1312 count_increased
= false;
1320 pr_err("%#lx: unhandlable page.\n", page_to_pfn(p
));
1325 static int __get_unpoison_page(struct page
*page
)
1327 struct page
*head
= compound_head(page
);
1329 bool hugetlb
= false;
1331 ret
= get_hwpoison_huge_page(head
, &hugetlb
);
1336 * PageHWPoisonTakenOff pages are not only marked as PG_hwpoison,
1337 * but also isolated from buddy freelist, so need to identify the
1338 * state and have to cancel both operations to unpoison.
1340 if (PageHWPoisonTakenOff(page
))
1343 return get_page_unless_zero(page
) ? 1 : 0;
1347 * get_hwpoison_page() - Get refcount for memory error handling
1348 * @p: Raw error page (hit by memory error)
1349 * @flags: Flags controlling behavior of error handling
1351 * get_hwpoison_page() takes a page refcount of an error page to handle memory
1352 * error on it, after checking that the error page is in a well-defined state
1353 * (defined as a page-type we can successfully handle the memory error on it,
1354 * such as LRU page and hugetlb page).
1356 * Memory error handling could be triggered at any time on any type of page,
1357 * so it's prone to race with typical memory management lifecycle (like
1358 * allocation and free). So to avoid such races, get_hwpoison_page() takes
1359 * extra care for the error page's state (as done in __get_hwpoison_page()),
1360 * and has some retry logic in get_any_page().
1362 * When called from unpoison_memory(), the caller should already ensure that
1363 * the given page has PG_hwpoison. So it's never reused for other page
1364 * allocations, and __get_unpoison_page() never races with them.
1366 * Return: 0 on failure,
1367 * 1 on success for in-use pages in a well-defined state,
1368 * -EIO for pages on which we can not handle memory errors,
1369 * -EBUSY when get_hwpoison_page() has raced with page lifecycle
1370 * operations like allocation and free,
1371 * -EHWPOISON when the page is hwpoisoned and taken off from buddy.
1373 static int get_hwpoison_page(struct page
*p
, unsigned long flags
)
1377 zone_pcp_disable(page_zone(p
));
1378 if (flags
& MF_UNPOISON
)
1379 ret
= __get_unpoison_page(p
);
1381 ret
= get_any_page(p
, flags
);
1382 zone_pcp_enable(page_zone(p
));
1388 * Do all that is necessary to remove user space mappings. Unmap
1389 * the pages and send SIGBUS to the processes if the data was dirty.
1391 static bool hwpoison_user_mappings(struct page
*p
, unsigned long pfn
,
1392 int flags
, struct page
*hpage
)
1394 struct folio
*folio
= page_folio(hpage
);
1395 enum ttu_flags ttu
= TTU_IGNORE_MLOCK
| TTU_SYNC
;
1396 struct address_space
*mapping
;
1399 int kill
= 1, forcekill
;
1400 bool mlocked
= PageMlocked(hpage
);
1403 * Here we are interested only in user-mapped pages, so skip any
1404 * other types of pages.
1406 if (PageReserved(p
) || PageSlab(p
))
1408 if (!(PageLRU(hpage
) || PageHuge(p
)))
1412 * This check implies we don't kill processes if their pages
1413 * are in the swap cache early. Those are always late kills.
1415 if (!page_mapped(hpage
))
1419 pr_err("%#lx: can't handle KSM pages.\n", pfn
);
1423 if (PageSwapCache(p
)) {
1424 pr_err("%#lx: keeping poisoned page in swap cache\n", pfn
);
1425 ttu
|= TTU_IGNORE_HWPOISON
;
1429 * Propagate the dirty bit from PTEs to struct page first, because we
1430 * need this to decide if we should kill or just drop the page.
1431 * XXX: the dirty test could be racy: set_page_dirty() may not always
1432 * be called inside page lock (it's recommended but not enforced).
1434 mapping
= page_mapping(hpage
);
1435 if (!(flags
& MF_MUST_KILL
) && !PageDirty(hpage
) && mapping
&&
1436 mapping_can_writeback(mapping
)) {
1437 if (page_mkclean(hpage
)) {
1438 SetPageDirty(hpage
);
1441 ttu
|= TTU_IGNORE_HWPOISON
;
1442 pr_info("%#lx: corrupted page was clean: dropped without side effects\n",
1448 * First collect all the processes that have the page
1449 * mapped in dirty form. This has to be done before try_to_unmap,
1450 * because ttu takes the rmap data structures down.
1452 * Error handling: We ignore errors here because
1453 * there's nothing that can be done.
1456 collect_procs(hpage
, &tokill
, flags
& MF_ACTION_REQUIRED
);
1458 if (PageHuge(hpage
) && !PageAnon(hpage
)) {
1460 * For hugetlb pages in shared mappings, try_to_unmap
1461 * could potentially call huge_pmd_unshare. Because of
1462 * this, take semaphore in write mode here and set
1463 * TTU_RMAP_LOCKED to indicate we have taken the lock
1464 * at this higher level.
1466 mapping
= hugetlb_page_mapping_lock_write(hpage
);
1468 try_to_unmap(folio
, ttu
|TTU_RMAP_LOCKED
);
1469 i_mmap_unlock_write(mapping
);
1471 pr_info("%#lx: could not lock mapping for mapped huge page\n", pfn
);
1473 try_to_unmap(folio
, ttu
);
1476 unmap_success
= !page_mapped(hpage
);
1478 pr_err("%#lx: failed to unmap page (mapcount=%d)\n",
1479 pfn
, page_mapcount(hpage
));
1482 * try_to_unmap() might put mlocked page in lru cache, so call
1483 * shake_page() again to ensure that it's flushed.
1489 * Now that the dirty bit has been propagated to the
1490 * struct page and all unmaps done we can decide if
1491 * killing is needed or not. Only kill when the page
1492 * was dirty or the process is not restartable,
1493 * otherwise the tokill list is merely
1494 * freed. When there was a problem unmapping earlier
1495 * use a more force-full uncatchable kill to prevent
1496 * any accesses to the poisoned memory.
1498 forcekill
= PageDirty(hpage
) || (flags
& MF_MUST_KILL
);
1499 kill_procs(&tokill
, forcekill
, !unmap_success
, pfn
, flags
);
1501 return unmap_success
;
1504 static int identify_page_state(unsigned long pfn
, struct page
*p
,
1505 unsigned long page_flags
)
1507 struct page_state
*ps
;
1510 * The first check uses the current page flags which may not have any
1511 * relevant information. The second check with the saved page flags is
1512 * carried out only if the first check can't determine the page status.
1514 for (ps
= error_states
;; ps
++)
1515 if ((p
->flags
& ps
->mask
) == ps
->res
)
1518 page_flags
|= (p
->flags
& (1UL << PG_dirty
));
1521 for (ps
= error_states
;; ps
++)
1522 if ((page_flags
& ps
->mask
) == ps
->res
)
1524 return page_action(ps
, p
, pfn
);
1527 static int try_to_split_thp_page(struct page
*page
, const char *msg
)
1530 if (unlikely(split_huge_page(page
))) {
1531 unsigned long pfn
= page_to_pfn(page
);
1534 pr_info("%s: %#lx: thp split failed\n", msg
, pfn
);
1543 static void unmap_and_kill(struct list_head
*to_kill
, unsigned long pfn
,
1544 struct address_space
*mapping
, pgoff_t index
, int flags
)
1547 unsigned long size
= 0;
1549 list_for_each_entry(tk
, to_kill
, nd
)
1551 size
= max(size
, 1UL << tk
->size_shift
);
1555 * Unmap the largest mapping to avoid breaking up device-dax
1556 * mappings which are constant size. The actual size of the
1557 * mapping being torn down is communicated in siginfo, see
1560 loff_t start
= (index
<< PAGE_SHIFT
) & ~(size
- 1);
1562 unmap_mapping_range(mapping
, start
, size
, 0);
1565 kill_procs(to_kill
, flags
& MF_MUST_KILL
, false, pfn
, flags
);
1568 static int mf_generic_kill_procs(unsigned long long pfn
, int flags
,
1569 struct dev_pagemap
*pgmap
)
1571 struct page
*page
= pfn_to_page(pfn
);
1577 * Pages instantiated by device-dax (not filesystem-dax)
1578 * may be compound pages.
1580 page
= compound_head(page
);
1583 * Prevent the inode from being freed while we are interrogating
1584 * the address_space, typically this would be handled by
1585 * lock_page(), but dax pages do not use the page lock. This
1586 * also prevents changes to the mapping of this pfn until
1587 * poison signaling is complete.
1589 cookie
= dax_lock_page(page
);
1593 if (hwpoison_filter(page
)) {
1598 switch (pgmap
->type
) {
1599 case MEMORY_DEVICE_PRIVATE
:
1600 case MEMORY_DEVICE_COHERENT
:
1602 * TODO: Handle device pages which may need coordination
1603 * with device-side memory.
1612 * Use this flag as an indication that the dax page has been
1613 * remapped UC to prevent speculative consumption of poison.
1615 SetPageHWPoison(page
);
1618 * Unlike System-RAM there is no possibility to swap in a
1619 * different physical page at a given virtual address, so all
1620 * userspace consumption of ZONE_DEVICE memory necessitates
1621 * SIGBUS (i.e. MF_MUST_KILL)
1623 flags
|= MF_ACTION_REQUIRED
| MF_MUST_KILL
;
1624 collect_procs(page
, &to_kill
, true);
1626 unmap_and_kill(&to_kill
, pfn
, page
->mapping
, page
->index
, flags
);
1628 dax_unlock_page(page
, cookie
);
1632 #ifdef CONFIG_FS_DAX
1634 * mf_dax_kill_procs - Collect and kill processes who are using this file range
1635 * @mapping: address_space of the file in use
1636 * @index: start pgoff of the range within the file
1637 * @count: length of the range, in unit of PAGE_SIZE
1638 * @mf_flags: memory failure flags
1640 int mf_dax_kill_procs(struct address_space
*mapping
, pgoff_t index
,
1641 unsigned long count
, int mf_flags
)
1646 size_t end
= index
+ count
;
1648 mf_flags
|= MF_ACTION_REQUIRED
| MF_MUST_KILL
;
1650 for (; index
< end
; index
++) {
1652 cookie
= dax_lock_mapping_entry(mapping
, index
, &page
);
1658 SetPageHWPoison(page
);
1660 collect_procs_fsdax(page
, mapping
, index
, &to_kill
);
1661 unmap_and_kill(&to_kill
, page_to_pfn(page
), mapping
,
1664 dax_unlock_mapping_entry(mapping
, index
, cookie
);
1668 EXPORT_SYMBOL_GPL(mf_dax_kill_procs
);
1669 #endif /* CONFIG_FS_DAX */
1671 #ifdef CONFIG_HUGETLB_PAGE
1673 * Struct raw_hwp_page represents information about "raw error page",
1674 * constructing singly linked list originated from ->private field of
1675 * SUBPAGE_INDEX_HWPOISON-th tail page.
1677 struct raw_hwp_page
{
1678 struct llist_node node
;
1682 static inline struct llist_head
*raw_hwp_list_head(struct page
*hpage
)
1684 return (struct llist_head
*)&page_private(hpage
+ SUBPAGE_INDEX_HWPOISON
);
1687 static unsigned long __free_raw_hwp_pages(struct page
*hpage
, bool move_flag
)
1689 struct llist_head
*head
;
1690 struct llist_node
*t
, *tnode
;
1691 unsigned long count
= 0;
1693 head
= raw_hwp_list_head(hpage
);
1694 llist_for_each_safe(tnode
, t
, head
->first
) {
1695 struct raw_hwp_page
*p
= container_of(tnode
, struct raw_hwp_page
, node
);
1698 SetPageHWPoison(p
->page
);
1702 llist_del_all(head
);
1706 static int hugetlb_set_page_hwpoison(struct page
*hpage
, struct page
*page
)
1708 struct llist_head
*head
;
1709 struct raw_hwp_page
*raw_hwp
;
1710 struct llist_node
*t
, *tnode
;
1711 int ret
= TestSetPageHWPoison(hpage
) ? -EHWPOISON
: 0;
1714 * Once the hwpoison hugepage has lost reliable raw error info,
1715 * there is little meaning to keep additional error info precisely,
1716 * so skip to add additional raw error info.
1718 if (HPageRawHwpUnreliable(hpage
))
1720 head
= raw_hwp_list_head(hpage
);
1721 llist_for_each_safe(tnode
, t
, head
->first
) {
1722 struct raw_hwp_page
*p
= container_of(tnode
, struct raw_hwp_page
, node
);
1724 if (p
->page
== page
)
1728 raw_hwp
= kmalloc(sizeof(struct raw_hwp_page
), GFP_ATOMIC
);
1730 raw_hwp
->page
= page
;
1731 llist_add(&raw_hwp
->node
, head
);
1732 /* the first error event will be counted in action_result(). */
1734 num_poisoned_pages_inc();
1737 * Failed to save raw error info. We no longer trace all
1738 * hwpoisoned subpages, and we need refuse to free/dissolve
1739 * this hwpoisoned hugepage.
1741 SetHPageRawHwpUnreliable(hpage
);
1743 * Once HPageRawHwpUnreliable is set, raw_hwp_page is not
1744 * used any more, so free it.
1746 __free_raw_hwp_pages(hpage
, false);
1751 static unsigned long free_raw_hwp_pages(struct page
*hpage
, bool move_flag
)
1754 * HPageVmemmapOptimized hugepages can't be freed because struct
1755 * pages for tail pages are required but they don't exist.
1757 if (move_flag
&& HPageVmemmapOptimized(hpage
))
1761 * HPageRawHwpUnreliable hugepages shouldn't be unpoisoned by
1764 if (HPageRawHwpUnreliable(hpage
))
1767 return __free_raw_hwp_pages(hpage
, move_flag
);
1770 void hugetlb_clear_page_hwpoison(struct page
*hpage
)
1772 if (HPageRawHwpUnreliable(hpage
))
1774 ClearPageHWPoison(hpage
);
1775 free_raw_hwp_pages(hpage
, true);
1779 * Called from hugetlb code with hugetlb_lock held.
1783 * 1 - in-use hugepage
1784 * 2 - not a hugepage
1785 * -EBUSY - the hugepage is busy (try to retry)
1786 * -EHWPOISON - the hugepage is already hwpoisoned
1788 int __get_huge_page_for_hwpoison(unsigned long pfn
, int flags
)
1790 struct page
*page
= pfn_to_page(pfn
);
1791 struct page
*head
= compound_head(page
);
1792 int ret
= 2; /* fallback to normal page handling */
1793 bool count_increased
= false;
1795 if (!PageHeadHuge(head
))
1798 if (flags
& MF_COUNT_INCREASED
) {
1800 count_increased
= true;
1801 } else if (HPageFreed(head
)) {
1803 } else if (HPageMigratable(head
)) {
1804 ret
= get_page_unless_zero(head
);
1806 count_increased
= true;
1809 if (!(flags
& MF_NO_RETRY
))
1813 if (hugetlb_set_page_hwpoison(head
, page
)) {
1820 if (count_increased
)
1826 * Taking refcount of hugetlb pages needs extra care about race conditions
1827 * with basic operations like hugepage allocation/free/demotion.
1828 * So some of prechecks for hwpoison (pinning, and testing/setting
1829 * PageHWPoison) should be done in single hugetlb_lock range.
1831 static int try_memory_failure_hugetlb(unsigned long pfn
, int flags
, int *hugetlb
)
1834 struct page
*p
= pfn_to_page(pfn
);
1836 unsigned long page_flags
;
1840 res
= get_huge_page_for_hwpoison(pfn
, flags
);
1841 if (res
== 2) { /* fallback to normal page handling */
1844 } else if (res
== -EHWPOISON
) {
1845 pr_err("%#lx: already hardware poisoned\n", pfn
);
1846 if (flags
& MF_ACTION_REQUIRED
) {
1847 head
= compound_head(p
);
1848 res
= kill_accessing_process(current
, page_to_pfn(head
), flags
);
1851 } else if (res
== -EBUSY
) {
1852 if (!(flags
& MF_NO_RETRY
)) {
1853 flags
|= MF_NO_RETRY
;
1856 action_result(pfn
, MF_MSG_UNKNOWN
, MF_IGNORED
);
1860 head
= compound_head(p
);
1863 if (hwpoison_filter(p
)) {
1864 hugetlb_clear_page_hwpoison(head
);
1870 * Handling free hugepage. The possible race with hugepage allocation
1871 * or demotion can be prevented by PageHWPoison flag.
1875 if (__page_handle_poison(p
) >= 0) {
1881 action_result(pfn
, MF_MSG_FREE_HUGE
, res
);
1882 return res
== MF_RECOVERED
? 0 : -EBUSY
;
1885 page_flags
= head
->flags
;
1887 if (!hwpoison_user_mappings(p
, pfn
, flags
, head
)) {
1888 action_result(pfn
, MF_MSG_UNMAP_FAILED
, MF_IGNORED
);
1893 return identify_page_state(pfn
, p
, page_flags
);
1900 static inline int try_memory_failure_hugetlb(unsigned long pfn
, int flags
, int *hugetlb
)
1905 static inline unsigned long free_raw_hwp_pages(struct page
*hpage
, bool flag
)
1909 #endif /* CONFIG_HUGETLB_PAGE */
1911 static int memory_failure_dev_pagemap(unsigned long pfn
, int flags
,
1912 struct dev_pagemap
*pgmap
)
1914 struct page
*page
= pfn_to_page(pfn
);
1917 if (flags
& MF_COUNT_INCREASED
)
1919 * Drop the extra refcount in case we come from madvise().
1923 /* device metadata space is not recoverable */
1924 if (!pgmap_pfn_valid(pgmap
, pfn
))
1928 * Call driver's implementation to handle the memory failure, otherwise
1929 * fall back to generic handler.
1931 if (pgmap
->ops
->memory_failure
) {
1932 rc
= pgmap
->ops
->memory_failure(pgmap
, pfn
, 1, flags
);
1934 * Fall back to generic handler too if operation is not
1935 * supported inside the driver/device/filesystem.
1937 if (rc
!= -EOPNOTSUPP
)
1941 rc
= mf_generic_kill_procs(pfn
, flags
, pgmap
);
1943 /* drop pgmap ref acquired in caller */
1944 put_dev_pagemap(pgmap
);
1945 action_result(pfn
, MF_MSG_DAX
, rc
? MF_FAILED
: MF_RECOVERED
);
1949 static DEFINE_MUTEX(mf_mutex
);
1952 * memory_failure - Handle memory failure of a page.
1953 * @pfn: Page Number of the corrupted page
1954 * @flags: fine tune action taken
1956 * This function is called by the low level machine check code
1957 * of an architecture when it detects hardware memory corruption
1958 * of a page. It tries its best to recover, which includes
1959 * dropping pages, killing processes etc.
1961 * The function is primarily of use for corruptions that
1962 * happen outside the current execution context (e.g. when
1963 * detected by a background scrubber)
1965 * Must run in process context (e.g. a work queue) with interrupts
1966 * enabled and no spinlocks hold.
1968 * Return: 0 for successfully handled the memory error,
1969 * -EOPNOTSUPP for hwpoison_filter() filtered the error event,
1970 * < 0(except -EOPNOTSUPP) on failure.
1972 int memory_failure(unsigned long pfn
, int flags
)
1976 struct dev_pagemap
*pgmap
;
1978 unsigned long page_flags
;
1982 if (!sysctl_memory_failure_recovery
)
1983 panic("Memory failure on page %lx", pfn
);
1985 mutex_lock(&mf_mutex
);
1987 if (!(flags
& MF_SW_SIMULATED
))
1988 hw_memory_failure
= true;
1990 p
= pfn_to_online_page(pfn
);
1992 res
= arch_memory_failure(pfn
, flags
);
1996 if (pfn_valid(pfn
)) {
1997 pgmap
= get_dev_pagemap(pfn
, NULL
);
1999 res
= memory_failure_dev_pagemap(pfn
, flags
,
2004 pr_err("%#lx: memory outside kernel control\n", pfn
);
2010 res
= try_memory_failure_hugetlb(pfn
, flags
, &hugetlb
);
2014 if (TestSetPageHWPoison(p
)) {
2015 pr_err("%#lx: already hardware poisoned\n", pfn
);
2017 if (flags
& MF_ACTION_REQUIRED
)
2018 res
= kill_accessing_process(current
, pfn
, flags
);
2019 if (flags
& MF_COUNT_INCREASED
)
2024 hpage
= compound_head(p
);
2027 * We need/can do nothing about count=0 pages.
2028 * 1) it's a free page, and therefore in safe hand:
2029 * prep_new_page() will be the gate keeper.
2030 * 2) it's part of a non-compound high order page.
2031 * Implies some kernel user: cannot stop them from
2032 * R/W the page; let's pray that the page has been
2033 * used and will be freed some time later.
2034 * In fact it's dangerous to directly bump up page count from 0,
2035 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
2037 if (!(flags
& MF_COUNT_INCREASED
)) {
2038 res
= get_hwpoison_page(p
, flags
);
2040 if (is_free_buddy_page(p
)) {
2041 if (take_page_off_buddy(p
)) {
2045 /* We lost the race, try again */
2047 ClearPageHWPoison(p
);
2053 action_result(pfn
, MF_MSG_BUDDY
, res
);
2054 res
= res
== MF_RECOVERED
? 0 : -EBUSY
;
2056 action_result(pfn
, MF_MSG_KERNEL_HIGH_ORDER
, MF_IGNORED
);
2060 } else if (res
< 0) {
2061 action_result(pfn
, MF_MSG_UNKNOWN
, MF_IGNORED
);
2067 if (PageTransHuge(hpage
)) {
2069 * The flag must be set after the refcount is bumped
2070 * otherwise it may race with THP split.
2071 * And the flag can't be set in get_hwpoison_page() since
2072 * it is called by soft offline too and it is just called
2073 * for !MF_COUNT_INCREASE. So here seems to be the best
2076 * Don't need care about the above error handling paths for
2077 * get_hwpoison_page() since they handle either free page
2078 * or unhandlable page. The refcount is bumped iff the
2079 * page is a valid handlable page.
2081 SetPageHasHWPoisoned(hpage
);
2082 if (try_to_split_thp_page(p
, "Memory Failure") < 0) {
2083 action_result(pfn
, MF_MSG_UNSPLIT_THP
, MF_IGNORED
);
2087 VM_BUG_ON_PAGE(!page_count(p
), p
);
2091 * We ignore non-LRU pages for good reasons.
2092 * - PG_locked is only well defined for LRU pages and a few others
2093 * - to avoid races with __SetPageLocked()
2094 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
2095 * The check (unnecessarily) ignores LRU pages being isolated and
2096 * walked by the page reclaim code, however that's not a big loss.
2103 * We're only intended to deal with the non-Compound page here.
2104 * However, the page could have changed compound pages due to
2105 * race window. If this happens, we could try again to hopefully
2106 * handle the page next round.
2108 if (PageCompound(p
)) {
2110 ClearPageHWPoison(p
);
2113 flags
&= ~MF_COUNT_INCREASED
;
2117 action_result(pfn
, MF_MSG_DIFFERENT_COMPOUND
, MF_IGNORED
);
2123 * We use page flags to determine what action should be taken, but
2124 * the flags can be modified by the error containment action. One
2125 * example is an mlocked page, where PG_mlocked is cleared by
2126 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
2127 * correctly, we save a copy of the page flags at this time.
2129 page_flags
= p
->flags
;
2131 if (hwpoison_filter(p
)) {
2132 TestClearPageHWPoison(p
);
2140 * __munlock_pagevec may clear a writeback page's LRU flag without
2141 * page_lock. We need wait writeback completion for this page or it
2142 * may trigger vfs BUG while evict inode.
2144 if (!PageLRU(p
) && !PageWriteback(p
))
2145 goto identify_page_state
;
2148 * It's very difficult to mess with pages currently under IO
2149 * and in many cases impossible, so we just avoid it here.
2151 wait_on_page_writeback(p
);
2154 * Now take care of user space mappings.
2155 * Abort on fail: __filemap_remove_folio() assumes unmapped page.
2157 if (!hwpoison_user_mappings(p
, pfn
, flags
, p
)) {
2158 action_result(pfn
, MF_MSG_UNMAP_FAILED
, MF_IGNORED
);
2164 * Torn down by someone else?
2166 if (PageLRU(p
) && !PageSwapCache(p
) && p
->mapping
== NULL
) {
2167 action_result(pfn
, MF_MSG_TRUNCATED_LRU
, MF_IGNORED
);
2172 identify_page_state
:
2173 res
= identify_page_state(pfn
, p
, page_flags
);
2174 mutex_unlock(&mf_mutex
);
2179 mutex_unlock(&mf_mutex
);
2182 EXPORT_SYMBOL_GPL(memory_failure
);
2184 #define MEMORY_FAILURE_FIFO_ORDER 4
2185 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
2187 struct memory_failure_entry
{
2192 struct memory_failure_cpu
{
2193 DECLARE_KFIFO(fifo
, struct memory_failure_entry
,
2194 MEMORY_FAILURE_FIFO_SIZE
);
2196 struct work_struct work
;
2199 static DEFINE_PER_CPU(struct memory_failure_cpu
, memory_failure_cpu
);
2202 * memory_failure_queue - Schedule handling memory failure of a page.
2203 * @pfn: Page Number of the corrupted page
2204 * @flags: Flags for memory failure handling
2206 * This function is called by the low level hardware error handler
2207 * when it detects hardware memory corruption of a page. It schedules
2208 * the recovering of error page, including dropping pages, killing
2211 * The function is primarily of use for corruptions that
2212 * happen outside the current execution context (e.g. when
2213 * detected by a background scrubber)
2215 * Can run in IRQ context.
2217 void memory_failure_queue(unsigned long pfn
, int flags
)
2219 struct memory_failure_cpu
*mf_cpu
;
2220 unsigned long proc_flags
;
2221 struct memory_failure_entry entry
= {
2226 mf_cpu
= &get_cpu_var(memory_failure_cpu
);
2227 spin_lock_irqsave(&mf_cpu
->lock
, proc_flags
);
2228 if (kfifo_put(&mf_cpu
->fifo
, entry
))
2229 schedule_work_on(smp_processor_id(), &mf_cpu
->work
);
2231 pr_err("buffer overflow when queuing memory failure at %#lx\n",
2233 spin_unlock_irqrestore(&mf_cpu
->lock
, proc_flags
);
2234 put_cpu_var(memory_failure_cpu
);
2236 EXPORT_SYMBOL_GPL(memory_failure_queue
);
2238 static void memory_failure_work_func(struct work_struct
*work
)
2240 struct memory_failure_cpu
*mf_cpu
;
2241 struct memory_failure_entry entry
= { 0, };
2242 unsigned long proc_flags
;
2245 mf_cpu
= container_of(work
, struct memory_failure_cpu
, work
);
2247 spin_lock_irqsave(&mf_cpu
->lock
, proc_flags
);
2248 gotten
= kfifo_get(&mf_cpu
->fifo
, &entry
);
2249 spin_unlock_irqrestore(&mf_cpu
->lock
, proc_flags
);
2252 if (entry
.flags
& MF_SOFT_OFFLINE
)
2253 soft_offline_page(entry
.pfn
, entry
.flags
);
2255 memory_failure(entry
.pfn
, entry
.flags
);
2260 * Process memory_failure work queued on the specified CPU.
2261 * Used to avoid return-to-userspace racing with the memory_failure workqueue.
2263 void memory_failure_queue_kick(int cpu
)
2265 struct memory_failure_cpu
*mf_cpu
;
2267 mf_cpu
= &per_cpu(memory_failure_cpu
, cpu
);
2268 cancel_work_sync(&mf_cpu
->work
);
2269 memory_failure_work_func(&mf_cpu
->work
);
2272 static int __init
memory_failure_init(void)
2274 struct memory_failure_cpu
*mf_cpu
;
2277 for_each_possible_cpu(cpu
) {
2278 mf_cpu
= &per_cpu(memory_failure_cpu
, cpu
);
2279 spin_lock_init(&mf_cpu
->lock
);
2280 INIT_KFIFO(mf_cpu
->fifo
);
2281 INIT_WORK(&mf_cpu
->work
, memory_failure_work_func
);
2286 core_initcall(memory_failure_init
);
2289 #define pr_fmt(fmt) "" fmt
2290 #define unpoison_pr_info(fmt, pfn, rs) \
2292 if (__ratelimit(rs)) \
2293 pr_info(fmt, pfn); \
2297 * unpoison_memory - Unpoison a previously poisoned page
2298 * @pfn: Page number of the to be unpoisoned page
2300 * Software-unpoison a page that has been poisoned by
2301 * memory_failure() earlier.
2303 * This is only done on the software-level, so it only works
2304 * for linux injected failures, not real hardware failures
2306 * Returns 0 for success, otherwise -errno.
2308 int unpoison_memory(unsigned long pfn
)
2314 unsigned long count
= 1;
2315 static DEFINE_RATELIMIT_STATE(unpoison_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2316 DEFAULT_RATELIMIT_BURST
);
2318 if (!pfn_valid(pfn
))
2321 p
= pfn_to_page(pfn
);
2322 page
= compound_head(p
);
2324 mutex_lock(&mf_mutex
);
2326 if (hw_memory_failure
) {
2327 unpoison_pr_info("Unpoison: Disabled after HW memory failure %#lx\n",
2333 if (!PageHWPoison(p
)) {
2334 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
2339 if (page_count(page
) > 1) {
2340 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
2345 if (page_mapped(page
)) {
2346 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
2351 if (page_mapping(page
)) {
2352 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
2357 if (PageSlab(page
) || PageTable(page
))
2360 ret
= get_hwpoison_page(p
, MF_UNPOISON
);
2363 count
= free_raw_hwp_pages(page
, false);
2369 ret
= TestClearPageHWPoison(page
) ? 0 : -EBUSY
;
2370 } else if (ret
< 0) {
2371 if (ret
== -EHWPOISON
) {
2372 ret
= put_page_back_buddy(p
) ? 0 : -EBUSY
;
2374 unpoison_pr_info("Unpoison: failed to grab page %#lx\n",
2378 count
= free_raw_hwp_pages(page
, false);
2384 freeit
= !!TestClearPageHWPoison(p
);
2387 if (freeit
&& !(pfn
== my_zero_pfn(0) && page_count(p
) == 1)) {
2394 mutex_unlock(&mf_mutex
);
2395 if (!ret
|| freeit
) {
2396 num_poisoned_pages_sub(count
);
2397 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
2398 page_to_pfn(p
), &unpoison_rs
);
2402 EXPORT_SYMBOL(unpoison_memory
);
2404 static bool isolate_page(struct page
*page
, struct list_head
*pagelist
)
2406 bool isolated
= false;
2407 bool lru
= PageLRU(page
);
2409 if (PageHuge(page
)) {
2410 isolated
= !isolate_hugetlb(page
, pagelist
);
2413 isolated
= !isolate_lru_page(page
);
2415 isolated
= !isolate_movable_page(page
, ISOLATE_UNEVICTABLE
);
2418 list_add(&page
->lru
, pagelist
);
2421 if (isolated
&& lru
)
2422 inc_node_page_state(page
, NR_ISOLATED_ANON
+
2423 page_is_file_lru(page
));
2426 * If we succeed to isolate the page, we grabbed another refcount on
2427 * the page, so we can safely drop the one we got from get_any_pages().
2428 * If we failed to isolate the page, it means that we cannot go further
2429 * and we will return an error, so drop the reference we got from
2430 * get_any_pages() as well.
2437 * __soft_offline_page handles hugetlb-pages and non-hugetlb pages.
2438 * If the page is a non-dirty unmapped page-cache page, it simply invalidates.
2439 * If the page is mapped, it migrates the contents over.
2441 static int __soft_offline_page(struct page
*page
)
2444 unsigned long pfn
= page_to_pfn(page
);
2445 struct page
*hpage
= compound_head(page
);
2446 char const *msg_page
[] = {"page", "hugepage"};
2447 bool huge
= PageHuge(page
);
2448 LIST_HEAD(pagelist
);
2449 struct migration_target_control mtc
= {
2450 .nid
= NUMA_NO_NODE
,
2451 .gfp_mask
= GFP_USER
| __GFP_MOVABLE
| __GFP_RETRY_MAYFAIL
,
2455 if (!PageHuge(page
))
2456 wait_on_page_writeback(page
);
2457 if (PageHWPoison(page
)) {
2460 pr_info("soft offline: %#lx page already poisoned\n", pfn
);
2464 if (!PageHuge(page
) && PageLRU(page
) && !PageSwapCache(page
))
2466 * Try to invalidate first. This should work for
2467 * non dirty unmapped page cache pages.
2469 ret
= invalidate_inode_page(page
);
2473 pr_info("soft_offline: %#lx: invalidated\n", pfn
);
2474 page_handle_poison(page
, false, true);
2478 if (isolate_page(hpage
, &pagelist
)) {
2479 ret
= migrate_pages(&pagelist
, alloc_migration_target
, NULL
,
2480 (unsigned long)&mtc
, MIGRATE_SYNC
, MR_MEMORY_FAILURE
, NULL
);
2482 bool release
= !huge
;
2484 if (!page_handle_poison(page
, huge
, release
))
2487 if (!list_empty(&pagelist
))
2488 putback_movable_pages(&pagelist
);
2490 pr_info("soft offline: %#lx: %s migration failed %ld, type %pGp\n",
2491 pfn
, msg_page
[huge
], ret
, &page
->flags
);
2496 pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %pGp\n",
2497 pfn
, msg_page
[huge
], page_count(page
), &page
->flags
);
2503 static int soft_offline_in_use_page(struct page
*page
)
2505 struct page
*hpage
= compound_head(page
);
2507 if (!PageHuge(page
) && PageTransHuge(hpage
))
2508 if (try_to_split_thp_page(page
, "soft offline") < 0)
2510 return __soft_offline_page(page
);
2513 static int soft_offline_free_page(struct page
*page
)
2517 if (!page_handle_poison(page
, true, false))
2523 static void put_ref_page(struct page
*page
)
2530 * soft_offline_page - Soft offline a page.
2531 * @pfn: pfn to soft-offline
2532 * @flags: flags. Same as memory_failure().
2534 * Returns 0 on success
2535 * -EOPNOTSUPP for hwpoison_filter() filtered the error event
2536 * < 0 otherwise negated errno.
2538 * Soft offline a page, by migration or invalidation,
2539 * without killing anything. This is for the case when
2540 * a page is not corrupted yet (so it's still valid to access),
2541 * but has had a number of corrected errors and is better taken
2544 * The actual policy on when to do that is maintained by
2547 * This should never impact any application or cause data loss,
2548 * however it might take some time.
2550 * This is not a 100% solution for all memory, but tries to be
2551 * ``good enough'' for the majority of memory.
2553 int soft_offline_page(unsigned long pfn
, int flags
)
2556 bool try_again
= true;
2557 struct page
*page
, *ref_page
= NULL
;
2559 WARN_ON_ONCE(!pfn_valid(pfn
) && (flags
& MF_COUNT_INCREASED
));
2561 if (!pfn_valid(pfn
))
2563 if (flags
& MF_COUNT_INCREASED
)
2564 ref_page
= pfn_to_page(pfn
);
2566 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
2567 page
= pfn_to_online_page(pfn
);
2569 put_ref_page(ref_page
);
2573 mutex_lock(&mf_mutex
);
2575 if (PageHWPoison(page
)) {
2576 pr_info("%s: %#lx page already poisoned\n", __func__
, pfn
);
2577 put_ref_page(ref_page
);
2578 mutex_unlock(&mf_mutex
);
2584 ret
= get_hwpoison_page(page
, flags
| MF_SOFT_OFFLINE
);
2587 if (hwpoison_filter(page
)) {
2591 put_ref_page(ref_page
);
2593 mutex_unlock(&mf_mutex
);
2598 ret
= soft_offline_in_use_page(page
);
2599 } else if (ret
== 0) {
2600 if (soft_offline_free_page(page
) && try_again
) {
2602 flags
&= ~MF_COUNT_INCREASED
;
2607 mutex_unlock(&mf_mutex
);
2612 void clear_hwpoisoned_pages(struct page
*memmap
, int nr_pages
)
2617 * A further optimization is to have per section refcounted
2618 * num_poisoned_pages. But that would need more space per memmap, so
2619 * for now just do a quick global check to speed up this routine in the
2620 * absence of bad pages.
2622 if (atomic_long_read(&num_poisoned_pages
) == 0)
2625 for (i
= 0; i
< nr_pages
; i
++) {
2626 if (PageHWPoison(&memmap
[i
])) {
2627 num_poisoned_pages_dec();
2628 ClearPageHWPoison(&memmap
[i
]);