1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2008, 2009 Intel Corporation
4 * Authors: Andi Kleen, Fengguang Wu
6 * High level machine check handler. Handles pages reported by the
7 * hardware as being corrupted usually due to a multi-bit ECC memory or cache
10 * In addition there is a "soft offline" entry point that allows stop using
11 * not-yet-corrupted-by-suspicious pages without killing anything.
13 * Handles page cache pages in various states. The tricky part
14 * here is that we can access any page asynchronously in respect to
15 * other VM users, because memory failures could happen anytime and
16 * anywhere. This could violate some of their assumptions. This is why
17 * this code has to be extremely careful. Generally it tries to use
18 * normal locking rules, as in get the standard locks, even if that means
19 * the error handling takes potentially a long time.
21 * It can be very tempting to add handling for obscure cases here.
22 * In general any code for handling new cases should only be added iff:
23 * - You know how to test it.
24 * - You have a test that can be added to mce-test
25 * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/
26 * - The case actually shows up as a frequent (top 10) page state in
27 * tools/vm/page-types when running a real workload.
29 * There are several operations here with exponential complexity because
30 * of unsuitable VM data structures. For example the operation to map back
31 * from RMAP chains to processes has to walk the complete process list and
32 * has non linear complexity with the number. But since memory corruptions
33 * are rare we hope to get away with this. This avoids impacting the core
36 #include <linux/kernel.h>
38 #include <linux/page-flags.h>
39 #include <linux/kernel-page-flags.h>
40 #include <linux/sched/signal.h>
41 #include <linux/sched/task.h>
42 #include <linux/ksm.h>
43 #include <linux/rmap.h>
44 #include <linux/export.h>
45 #include <linux/pagemap.h>
46 #include <linux/swap.h>
47 #include <linux/backing-dev.h>
48 #include <linux/migrate.h>
49 #include <linux/suspend.h>
50 #include <linux/slab.h>
51 #include <linux/swapops.h>
52 #include <linux/hugetlb.h>
53 #include <linux/memory_hotplug.h>
54 #include <linux/mm_inline.h>
55 #include <linux/memremap.h>
56 #include <linux/kfifo.h>
57 #include <linux/ratelimit.h>
58 #include <linux/page-isolation.h>
60 #include "ras/ras_event.h"
62 int sysctl_memory_failure_early_kill __read_mostly
= 0;
64 int sysctl_memory_failure_recovery __read_mostly
= 1;
66 atomic_long_t num_poisoned_pages __read_mostly
= ATOMIC_LONG_INIT(0);
68 static bool page_handle_poison(struct page
*page
, bool hugepage_or_freepage
, bool release
)
70 if (hugepage_or_freepage
) {
72 * Doing this check for free pages is also fine since dissolve_free_huge_page
73 * returns 0 for non-hugetlb pages as well.
75 if (dissolve_free_huge_page(page
) || !take_page_off_buddy(page
))
77 * We could fail to take off the target page from buddy
78 * for example due to racy page allocation, but that's
79 * acceptable because soft-offlined page is not broken
80 * and if someone really want to use it, they should
86 SetPageHWPoison(page
);
90 num_poisoned_pages_inc();
95 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
97 u32 hwpoison_filter_enable
= 0;
98 u32 hwpoison_filter_dev_major
= ~0U;
99 u32 hwpoison_filter_dev_minor
= ~0U;
100 u64 hwpoison_filter_flags_mask
;
101 u64 hwpoison_filter_flags_value
;
102 EXPORT_SYMBOL_GPL(hwpoison_filter_enable
);
103 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major
);
104 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor
);
105 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask
);
106 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value
);
108 static int hwpoison_filter_dev(struct page
*p
)
110 struct address_space
*mapping
;
113 if (hwpoison_filter_dev_major
== ~0U &&
114 hwpoison_filter_dev_minor
== ~0U)
118 * page_mapping() does not accept slab pages.
123 mapping
= page_mapping(p
);
124 if (mapping
== NULL
|| mapping
->host
== NULL
)
127 dev
= mapping
->host
->i_sb
->s_dev
;
128 if (hwpoison_filter_dev_major
!= ~0U &&
129 hwpoison_filter_dev_major
!= MAJOR(dev
))
131 if (hwpoison_filter_dev_minor
!= ~0U &&
132 hwpoison_filter_dev_minor
!= MINOR(dev
))
138 static int hwpoison_filter_flags(struct page
*p
)
140 if (!hwpoison_filter_flags_mask
)
143 if ((stable_page_flags(p
) & hwpoison_filter_flags_mask
) ==
144 hwpoison_filter_flags_value
)
151 * This allows stress tests to limit test scope to a collection of tasks
152 * by putting them under some memcg. This prevents killing unrelated/important
153 * processes such as /sbin/init. Note that the target task may share clean
154 * pages with init (eg. libc text), which is harmless. If the target task
155 * share _dirty_ pages with another task B, the test scheme must make sure B
156 * is also included in the memcg. At last, due to race conditions this filter
157 * can only guarantee that the page either belongs to the memcg tasks, or is
161 u64 hwpoison_filter_memcg
;
162 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg
);
163 static int hwpoison_filter_task(struct page
*p
)
165 if (!hwpoison_filter_memcg
)
168 if (page_cgroup_ino(p
) != hwpoison_filter_memcg
)
174 static int hwpoison_filter_task(struct page
*p
) { return 0; }
177 int hwpoison_filter(struct page
*p
)
179 if (!hwpoison_filter_enable
)
182 if (hwpoison_filter_dev(p
))
185 if (hwpoison_filter_flags(p
))
188 if (hwpoison_filter_task(p
))
194 int hwpoison_filter(struct page
*p
)
200 EXPORT_SYMBOL_GPL(hwpoison_filter
);
203 * Kill all processes that have a poisoned page mapped and then isolate
207 * Find all processes having the page mapped and kill them.
208 * But we keep a page reference around so that the page is not
209 * actually freed yet.
210 * Then stash the page away
212 * There's no convenient way to get back to mapped processes
213 * from the VMAs. So do a brute-force search over all
216 * Remember that machine checks are not common (or rather
217 * if they are common you have other problems), so this shouldn't
218 * be a performance issue.
220 * Also there are some races possible while we get from the
221 * error detection to actually handle it.
226 struct task_struct
*tsk
;
232 * Send all the processes who have the page mapped a signal.
233 * ``action optional'' if they are not immediately affected by the error
234 * ``action required'' if error happened in current execution context
236 static int kill_proc(struct to_kill
*tk
, unsigned long pfn
, int flags
)
238 struct task_struct
*t
= tk
->tsk
;
239 short addr_lsb
= tk
->size_shift
;
242 pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
243 pfn
, t
->comm
, t
->pid
);
245 if (flags
& MF_ACTION_REQUIRED
) {
247 ret
= force_sig_mceerr(BUS_MCEERR_AR
,
248 (void __user
*)tk
->addr
, addr_lsb
);
250 /* Signal other processes sharing the page if they have PF_MCE_EARLY set. */
251 ret
= send_sig_mceerr(BUS_MCEERR_AO
, (void __user
*)tk
->addr
,
255 * Don't use force here, it's convenient if the signal
256 * can be temporarily blocked.
257 * This could cause a loop when the user sets SIGBUS
258 * to SIG_IGN, but hopefully no one will do that?
260 ret
= send_sig_mceerr(BUS_MCEERR_AO
, (void __user
*)tk
->addr
,
261 addr_lsb
, t
); /* synchronous? */
264 pr_info("Memory failure: Error sending signal to %s:%d: %d\n",
265 t
->comm
, t
->pid
, ret
);
270 * Unknown page type encountered. Try to check whether it can turn PageLRU by
271 * lru_add_drain_all, or a free page by reclaiming slabs when possible.
273 void shake_page(struct page
*p
, int access
)
280 if (PageLRU(p
) || is_free_buddy_page(p
))
285 * Only call shrink_node_slabs here (which would also shrink
286 * other caches) if access is not potentially fatal.
289 drop_slab_node(page_to_nid(p
));
291 EXPORT_SYMBOL_GPL(shake_page
);
293 static unsigned long dev_pagemap_mapping_shift(struct page
*page
,
294 struct vm_area_struct
*vma
)
296 unsigned long address
= vma_address(page
, vma
);
303 pgd
= pgd_offset(vma
->vm_mm
, address
);
304 if (!pgd_present(*pgd
))
306 p4d
= p4d_offset(pgd
, address
);
307 if (!p4d_present(*p4d
))
309 pud
= pud_offset(p4d
, address
);
310 if (!pud_present(*pud
))
312 if (pud_devmap(*pud
))
314 pmd
= pmd_offset(pud
, address
);
315 if (!pmd_present(*pmd
))
317 if (pmd_devmap(*pmd
))
319 pte
= pte_offset_map(pmd
, address
);
320 if (!pte_present(*pte
))
322 if (pte_devmap(*pte
))
328 * Failure handling: if we can't find or can't kill a process there's
329 * not much we can do. We just print a message and ignore otherwise.
333 * Schedule a process for later kill.
334 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
336 static void add_to_kill(struct task_struct
*tsk
, struct page
*p
,
337 struct vm_area_struct
*vma
,
338 struct list_head
*to_kill
)
342 tk
= kmalloc(sizeof(struct to_kill
), GFP_ATOMIC
);
344 pr_err("Memory failure: Out of memory while machine check handling\n");
348 tk
->addr
= page_address_in_vma(p
, vma
);
349 if (is_zone_device_page(p
))
350 tk
->size_shift
= dev_pagemap_mapping_shift(p
, vma
);
352 tk
->size_shift
= page_shift(compound_head(p
));
355 * Send SIGKILL if "tk->addr == -EFAULT". Also, as
356 * "tk->size_shift" is always non-zero for !is_zone_device_page(),
357 * so "tk->size_shift == 0" effectively checks no mapping on
358 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times
359 * to a process' address space, it's possible not all N VMAs
360 * contain mappings for the page, but at least one VMA does.
361 * Only deliver SIGBUS with payload derived from the VMA that
362 * has a mapping for the page.
364 if (tk
->addr
== -EFAULT
) {
365 pr_info("Memory failure: Unable to find user space address %lx in %s\n",
366 page_to_pfn(p
), tsk
->comm
);
367 } else if (tk
->size_shift
== 0) {
372 get_task_struct(tsk
);
374 list_add_tail(&tk
->nd
, to_kill
);
378 * Kill the processes that have been collected earlier.
380 * Only do anything when DOIT is set, otherwise just free the list
381 * (this is used for clean pages which do not need killing)
382 * Also when FAIL is set do a force kill because something went
385 static void kill_procs(struct list_head
*to_kill
, int forcekill
, bool fail
,
386 unsigned long pfn
, int flags
)
388 struct to_kill
*tk
, *next
;
390 list_for_each_entry_safe (tk
, next
, to_kill
, nd
) {
393 * In case something went wrong with munmapping
394 * make sure the process doesn't catch the
395 * signal and then access the memory. Just kill it.
397 if (fail
|| tk
->addr
== -EFAULT
) {
398 pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
399 pfn
, tk
->tsk
->comm
, tk
->tsk
->pid
);
400 do_send_sig_info(SIGKILL
, SEND_SIG_PRIV
,
401 tk
->tsk
, PIDTYPE_PID
);
405 * In theory the process could have mapped
406 * something else on the address in-between. We could
407 * check for that, but we need to tell the
410 else if (kill_proc(tk
, pfn
, flags
) < 0)
411 pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n",
412 pfn
, tk
->tsk
->comm
, tk
->tsk
->pid
);
414 put_task_struct(tk
->tsk
);
420 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
421 * on behalf of the thread group. Return task_struct of the (first found)
422 * dedicated thread if found, and return NULL otherwise.
424 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
425 * have to call rcu_read_lock/unlock() in this function.
427 static struct task_struct
*find_early_kill_thread(struct task_struct
*tsk
)
429 struct task_struct
*t
;
431 for_each_thread(tsk
, t
) {
432 if (t
->flags
& PF_MCE_PROCESS
) {
433 if (t
->flags
& PF_MCE_EARLY
)
436 if (sysctl_memory_failure_early_kill
)
444 * Determine whether a given process is "early kill" process which expects
445 * to be signaled when some page under the process is hwpoisoned.
446 * Return task_struct of the dedicated thread (main thread unless explicitly
447 * specified) if the process is "early kill" and otherwise returns NULL.
449 * Note that the above is true for Action Optional case. For Action Required
450 * case, it's only meaningful to the current thread which need to be signaled
451 * with SIGBUS, this error is Action Optional for other non current
452 * processes sharing the same error page,if the process is "early kill", the
453 * task_struct of the dedicated thread will also be returned.
455 static struct task_struct
*task_early_kill(struct task_struct
*tsk
,
461 * Comparing ->mm here because current task might represent
462 * a subthread, while tsk always points to the main thread.
464 if (force_early
&& tsk
->mm
== current
->mm
)
467 return find_early_kill_thread(tsk
);
471 * Collect processes when the error hit an anonymous page.
473 static void collect_procs_anon(struct page
*page
, struct list_head
*to_kill
,
476 struct vm_area_struct
*vma
;
477 struct task_struct
*tsk
;
481 av
= page_lock_anon_vma_read(page
);
482 if (av
== NULL
) /* Not actually mapped anymore */
485 pgoff
= page_to_pgoff(page
);
486 read_lock(&tasklist_lock
);
487 for_each_process (tsk
) {
488 struct anon_vma_chain
*vmac
;
489 struct task_struct
*t
= task_early_kill(tsk
, force_early
);
493 anon_vma_interval_tree_foreach(vmac
, &av
->rb_root
,
496 if (!page_mapped_in_vma(page
, vma
))
498 if (vma
->vm_mm
== t
->mm
)
499 add_to_kill(t
, page
, vma
, to_kill
);
502 read_unlock(&tasklist_lock
);
503 page_unlock_anon_vma_read(av
);
507 * Collect processes when the error hit a file mapped page.
509 static void collect_procs_file(struct page
*page
, struct list_head
*to_kill
,
512 struct vm_area_struct
*vma
;
513 struct task_struct
*tsk
;
514 struct address_space
*mapping
= page
->mapping
;
517 i_mmap_lock_read(mapping
);
518 read_lock(&tasklist_lock
);
519 pgoff
= page_to_pgoff(page
);
520 for_each_process(tsk
) {
521 struct task_struct
*t
= task_early_kill(tsk
, force_early
);
525 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
,
528 * Send early kill signal to tasks where a vma covers
529 * the page but the corrupted page is not necessarily
530 * mapped it in its pte.
531 * Assume applications who requested early kill want
532 * to be informed of all such data corruptions.
534 if (vma
->vm_mm
== t
->mm
)
535 add_to_kill(t
, page
, vma
, to_kill
);
538 read_unlock(&tasklist_lock
);
539 i_mmap_unlock_read(mapping
);
543 * Collect the processes who have the corrupted page mapped to kill.
545 static void collect_procs(struct page
*page
, struct list_head
*tokill
,
552 collect_procs_anon(page
, tokill
, force_early
);
554 collect_procs_file(page
, tokill
, force_early
);
557 static const char *action_name
[] = {
558 [MF_IGNORED
] = "Ignored",
559 [MF_FAILED
] = "Failed",
560 [MF_DELAYED
] = "Delayed",
561 [MF_RECOVERED
] = "Recovered",
564 static const char * const action_page_types
[] = {
565 [MF_MSG_KERNEL
] = "reserved kernel page",
566 [MF_MSG_KERNEL_HIGH_ORDER
] = "high-order kernel page",
567 [MF_MSG_SLAB
] = "kernel slab page",
568 [MF_MSG_DIFFERENT_COMPOUND
] = "different compound page after locking",
569 [MF_MSG_POISONED_HUGE
] = "huge page already hardware poisoned",
570 [MF_MSG_HUGE
] = "huge page",
571 [MF_MSG_FREE_HUGE
] = "free huge page",
572 [MF_MSG_NON_PMD_HUGE
] = "non-pmd-sized huge page",
573 [MF_MSG_UNMAP_FAILED
] = "unmapping failed page",
574 [MF_MSG_DIRTY_SWAPCACHE
] = "dirty swapcache page",
575 [MF_MSG_CLEAN_SWAPCACHE
] = "clean swapcache page",
576 [MF_MSG_DIRTY_MLOCKED_LRU
] = "dirty mlocked LRU page",
577 [MF_MSG_CLEAN_MLOCKED_LRU
] = "clean mlocked LRU page",
578 [MF_MSG_DIRTY_UNEVICTABLE_LRU
] = "dirty unevictable LRU page",
579 [MF_MSG_CLEAN_UNEVICTABLE_LRU
] = "clean unevictable LRU page",
580 [MF_MSG_DIRTY_LRU
] = "dirty LRU page",
581 [MF_MSG_CLEAN_LRU
] = "clean LRU page",
582 [MF_MSG_TRUNCATED_LRU
] = "already truncated LRU page",
583 [MF_MSG_BUDDY
] = "free buddy page",
584 [MF_MSG_BUDDY_2ND
] = "free buddy page (2nd try)",
585 [MF_MSG_DAX
] = "dax page",
586 [MF_MSG_UNSPLIT_THP
] = "unsplit thp",
587 [MF_MSG_UNKNOWN
] = "unknown page",
591 * XXX: It is possible that a page is isolated from LRU cache,
592 * and then kept in swap cache or failed to remove from page cache.
593 * The page count will stop it from being freed by unpoison.
594 * Stress tests should be aware of this memory leak problem.
596 static int delete_from_lru_cache(struct page
*p
)
598 if (!isolate_lru_page(p
)) {
600 * Clear sensible page flags, so that the buddy system won't
601 * complain when the page is unpoison-and-freed.
604 ClearPageUnevictable(p
);
607 * Poisoned page might never drop its ref count to 0 so we have
608 * to uncharge it manually from its memcg.
610 mem_cgroup_uncharge(p
);
613 * drop the page count elevated by isolate_lru_page()
621 static int truncate_error_page(struct page
*p
, unsigned long pfn
,
622 struct address_space
*mapping
)
626 if (mapping
->a_ops
->error_remove_page
) {
627 int err
= mapping
->a_ops
->error_remove_page(mapping
, p
);
630 pr_info("Memory failure: %#lx: Failed to punch page: %d\n",
632 } else if (page_has_private(p
) &&
633 !try_to_release_page(p
, GFP_NOIO
)) {
634 pr_info("Memory failure: %#lx: failed to release buffers\n",
641 * If the file system doesn't support it just invalidate
642 * This fails on dirty or anything with private pages
644 if (invalidate_inode_page(p
))
647 pr_info("Memory failure: %#lx: Failed to invalidate\n",
655 * Error hit kernel page.
656 * Do nothing, try to be lucky and not touch this instead. For a few cases we
657 * could be more sophisticated.
659 static int me_kernel(struct page
*p
, unsigned long pfn
)
666 * Page in unknown state. Do nothing.
668 static int me_unknown(struct page
*p
, unsigned long pfn
)
670 pr_err("Memory failure: %#lx: Unknown page state\n", pfn
);
676 * Clean (or cleaned) page cache page.
678 static int me_pagecache_clean(struct page
*p
, unsigned long pfn
)
681 struct address_space
*mapping
;
683 delete_from_lru_cache(p
);
686 * For anonymous pages we're done the only reference left
687 * should be the one m_f() holds.
695 * Now truncate the page in the page cache. This is really
696 * more like a "temporary hole punch"
697 * Don't do this for block devices when someone else
698 * has a reference, because it could be file system metadata
699 * and that's not safe to truncate.
701 mapping
= page_mapping(p
);
704 * Page has been teared down in the meanwhile
711 * Truncation is a bit tricky. Enable it per file system for now.
713 * Open: to take i_mutex or not for this? Right now we don't.
715 ret
= truncate_error_page(p
, pfn
, mapping
);
722 * Dirty pagecache page
723 * Issues: when the error hit a hole page the error is not properly
726 static int me_pagecache_dirty(struct page
*p
, unsigned long pfn
)
728 struct address_space
*mapping
= page_mapping(p
);
731 /* TBD: print more information about the file. */
734 * IO error will be reported by write(), fsync(), etc.
735 * who check the mapping.
736 * This way the application knows that something went
737 * wrong with its dirty file data.
739 * There's one open issue:
741 * The EIO will be only reported on the next IO
742 * operation and then cleared through the IO map.
743 * Normally Linux has two mechanisms to pass IO error
744 * first through the AS_EIO flag in the address space
745 * and then through the PageError flag in the page.
746 * Since we drop pages on memory failure handling the
747 * only mechanism open to use is through AS_AIO.
749 * This has the disadvantage that it gets cleared on
750 * the first operation that returns an error, while
751 * the PageError bit is more sticky and only cleared
752 * when the page is reread or dropped. If an
753 * application assumes it will always get error on
754 * fsync, but does other operations on the fd before
755 * and the page is dropped between then the error
756 * will not be properly reported.
758 * This can already happen even without hwpoisoned
759 * pages: first on metadata IO errors (which only
760 * report through AS_EIO) or when the page is dropped
763 * So right now we assume that the application DTRT on
764 * the first EIO, but we're not worse than other parts
767 mapping_set_error(mapping
, -EIO
);
770 return me_pagecache_clean(p
, pfn
);
774 * Clean and dirty swap cache.
776 * Dirty swap cache page is tricky to handle. The page could live both in page
777 * cache and swap cache(ie. page is freshly swapped in). So it could be
778 * referenced concurrently by 2 types of PTEs:
779 * normal PTEs and swap PTEs. We try to handle them consistently by calling
780 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
782 * - clear dirty bit to prevent IO
784 * - but keep in the swap cache, so that when we return to it on
785 * a later page fault, we know the application is accessing
786 * corrupted data and shall be killed (we installed simple
787 * interception code in do_swap_page to catch it).
789 * Clean swap cache pages can be directly isolated. A later page fault will
790 * bring in the known good data from disk.
792 static int me_swapcache_dirty(struct page
*p
, unsigned long pfn
)
797 /* Trigger EIO in shmem: */
798 ClearPageUptodate(p
);
800 ret
= delete_from_lru_cache(p
) ? MF_FAILED
: MF_DELAYED
;
805 static int me_swapcache_clean(struct page
*p
, unsigned long pfn
)
809 delete_from_swap_cache(p
);
811 ret
= delete_from_lru_cache(p
) ? MF_FAILED
: MF_RECOVERED
;
817 * Huge pages. Needs work.
819 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
820 * To narrow down kill region to one page, we need to break up pmd.
822 static int me_huge_page(struct page
*p
, unsigned long pfn
)
825 struct page
*hpage
= compound_head(p
);
826 struct address_space
*mapping
;
828 if (!PageHuge(hpage
))
831 mapping
= page_mapping(hpage
);
833 res
= truncate_error_page(hpage
, pfn
, mapping
);
839 * migration entry prevents later access on error anonymous
840 * hugepage, so we can free and dissolve it into buddy to
841 * save healthy subpages.
845 if (!dissolve_free_huge_page(p
) && take_page_off_buddy(p
)) {
855 * Various page states we can handle.
857 * A page state is defined by its current page->flags bits.
858 * The table matches them in order and calls the right handler.
860 * This is quite tricky because we can access page at any time
861 * in its live cycle, so all accesses have to be extremely careful.
863 * This is not complete. More states could be added.
864 * For any missing state don't attempt recovery.
867 #define dirty (1UL << PG_dirty)
868 #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked))
869 #define unevict (1UL << PG_unevictable)
870 #define mlock (1UL << PG_mlocked)
871 #define lru (1UL << PG_lru)
872 #define head (1UL << PG_head)
873 #define slab (1UL << PG_slab)
874 #define reserved (1UL << PG_reserved)
876 static struct page_state
{
879 enum mf_action_page_type type
;
881 /* Callback ->action() has to unlock the relevant page inside it. */
882 int (*action
)(struct page
*p
, unsigned long pfn
);
884 { reserved
, reserved
, MF_MSG_KERNEL
, me_kernel
},
886 * free pages are specially detected outside this table:
887 * PG_buddy pages only make a small fraction of all free pages.
891 * Could in theory check if slab page is free or if we can drop
892 * currently unused objects without touching them. But just
893 * treat it as standard kernel for now.
895 { slab
, slab
, MF_MSG_SLAB
, me_kernel
},
897 { head
, head
, MF_MSG_HUGE
, me_huge_page
},
899 { sc
|dirty
, sc
|dirty
, MF_MSG_DIRTY_SWAPCACHE
, me_swapcache_dirty
},
900 { sc
|dirty
, sc
, MF_MSG_CLEAN_SWAPCACHE
, me_swapcache_clean
},
902 { mlock
|dirty
, mlock
|dirty
, MF_MSG_DIRTY_MLOCKED_LRU
, me_pagecache_dirty
},
903 { mlock
|dirty
, mlock
, MF_MSG_CLEAN_MLOCKED_LRU
, me_pagecache_clean
},
905 { unevict
|dirty
, unevict
|dirty
, MF_MSG_DIRTY_UNEVICTABLE_LRU
, me_pagecache_dirty
},
906 { unevict
|dirty
, unevict
, MF_MSG_CLEAN_UNEVICTABLE_LRU
, me_pagecache_clean
},
908 { lru
|dirty
, lru
|dirty
, MF_MSG_DIRTY_LRU
, me_pagecache_dirty
},
909 { lru
|dirty
, lru
, MF_MSG_CLEAN_LRU
, me_pagecache_clean
},
912 * Catchall entry: must be at end.
914 { 0, 0, MF_MSG_UNKNOWN
, me_unknown
},
927 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
928 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
930 static void action_result(unsigned long pfn
, enum mf_action_page_type type
,
931 enum mf_result result
)
933 trace_memory_failure_event(pfn
, type
, result
);
935 pr_err("Memory failure: %#lx: recovery action for %s: %s\n",
936 pfn
, action_page_types
[type
], action_name
[result
]);
939 static int page_action(struct page_state
*ps
, struct page
*p
,
945 /* page p should be unlocked after returning from ps->action(). */
946 result
= ps
->action(p
, pfn
);
948 count
= page_count(p
) - 1;
949 if (ps
->action
== me_swapcache_dirty
&& result
== MF_DELAYED
)
952 pr_err("Memory failure: %#lx: %s still referenced by %d users\n",
953 pfn
, action_page_types
[ps
->type
], count
);
956 action_result(pfn
, ps
->type
, result
);
958 /* Could do more checks here if page looks ok */
960 * Could adjust zone counters here to correct for the missing page.
963 return (result
== MF_RECOVERED
|| result
== MF_DELAYED
) ? 0 : -EBUSY
;
967 * Return true if a page type of a given page is supported by hwpoison
968 * mechanism (while handling could fail), otherwise false. This function
969 * does not return true for hugetlb or device memory pages, so it's assumed
970 * to be called only in the context where we never have such pages.
972 static inline bool HWPoisonHandlable(struct page
*page
)
974 return PageLRU(page
) || __PageMovable(page
);
978 * __get_hwpoison_page() - Get refcount for memory error handling:
979 * @page: raw error page (hit by memory error)
981 * Return: return 0 if failed to grab the refcount, otherwise true (some
984 static int __get_hwpoison_page(struct page
*page
)
986 struct page
*head
= compound_head(page
);
988 bool hugetlb
= false;
990 ret
= get_hwpoison_huge_page(head
, &hugetlb
);
995 * This check prevents from calling get_hwpoison_unless_zero()
996 * for any unsupported type of page in order to reduce the risk of
997 * unexpected races caused by taking a page refcount.
999 if (!HWPoisonHandlable(head
))
1002 if (PageTransHuge(head
)) {
1004 * Non anonymous thp exists only in allocation/free time. We
1005 * can't handle such a case correctly, so let's give it up.
1006 * This should be better than triggering BUG_ON when kernel
1007 * tries to touch the "partially handled" page.
1009 if (!PageAnon(head
)) {
1010 pr_err("Memory failure: %#lx: non anonymous thp\n",
1016 if (get_page_unless_zero(head
)) {
1017 if (head
== compound_head(page
))
1020 pr_info("Memory failure: %#lx cannot catch tail\n",
1029 * Safely get reference count of an arbitrary page.
1031 * Returns 0 for a free page, 1 for an in-use page,
1032 * -EIO for a page-type we cannot handle and -EBUSY if we raced with an
1034 * We only incremented refcount in case the page was already in-use and it
1035 * is a known type we can handle.
1037 static int get_any_page(struct page
*p
, unsigned long flags
)
1039 int ret
= 0, pass
= 0;
1040 bool count_increased
= false;
1042 if (flags
& MF_COUNT_INCREASED
)
1043 count_increased
= true;
1046 if (!count_increased
&& !__get_hwpoison_page(p
)) {
1047 if (page_count(p
)) {
1048 /* We raced with an allocation, retry. */
1052 } else if (!PageHuge(p
) && !is_free_buddy_page(p
)) {
1053 /* We raced with put_page, retry. */
1059 if (PageHuge(p
) || HWPoisonHandlable(p
)) {
1063 * A page we cannot handle. Check whether we can turn
1064 * it into something we can handle.
1069 count_increased
= false;
1080 static int get_hwpoison_page(struct page
*p
, unsigned long flags
,
1085 zone_pcp_disable(page_zone(p
));
1086 if (ctxt
== MF_SOFT_OFFLINE
)
1087 ret
= get_any_page(p
, flags
);
1089 ret
= __get_hwpoison_page(p
);
1090 zone_pcp_enable(page_zone(p
));
1096 * Do all that is necessary to remove user space mappings. Unmap
1097 * the pages and send SIGBUS to the processes if the data was dirty.
1099 static bool hwpoison_user_mappings(struct page
*p
, unsigned long pfn
,
1100 int flags
, struct page
**hpagep
)
1102 enum ttu_flags ttu
= TTU_IGNORE_MLOCK
;
1103 struct address_space
*mapping
;
1105 bool unmap_success
= true;
1106 int kill
= 1, forcekill
;
1107 struct page
*hpage
= *hpagep
;
1108 bool mlocked
= PageMlocked(hpage
);
1111 * Here we are interested only in user-mapped pages, so skip any
1112 * other types of pages.
1114 if (PageReserved(p
) || PageSlab(p
))
1116 if (!(PageLRU(hpage
) || PageHuge(p
)))
1120 * This check implies we don't kill processes if their pages
1121 * are in the swap cache early. Those are always late kills.
1123 if (!page_mapped(hpage
))
1127 pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn
);
1131 if (PageSwapCache(p
)) {
1132 pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n",
1134 ttu
|= TTU_IGNORE_HWPOISON
;
1138 * Propagate the dirty bit from PTEs to struct page first, because we
1139 * need this to decide if we should kill or just drop the page.
1140 * XXX: the dirty test could be racy: set_page_dirty() may not always
1141 * be called inside page lock (it's recommended but not enforced).
1143 mapping
= page_mapping(hpage
);
1144 if (!(flags
& MF_MUST_KILL
) && !PageDirty(hpage
) && mapping
&&
1145 mapping_can_writeback(mapping
)) {
1146 if (page_mkclean(hpage
)) {
1147 SetPageDirty(hpage
);
1150 ttu
|= TTU_IGNORE_HWPOISON
;
1151 pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n",
1157 * First collect all the processes that have the page
1158 * mapped in dirty form. This has to be done before try_to_unmap,
1159 * because ttu takes the rmap data structures down.
1161 * Error handling: We ignore errors here because
1162 * there's nothing that can be done.
1165 collect_procs(hpage
, &tokill
, flags
& MF_ACTION_REQUIRED
);
1167 if (!PageHuge(hpage
)) {
1168 unmap_success
= try_to_unmap(hpage
, ttu
);
1170 if (!PageAnon(hpage
)) {
1172 * For hugetlb pages in shared mappings, try_to_unmap
1173 * could potentially call huge_pmd_unshare. Because of
1174 * this, take semaphore in write mode here and set
1175 * TTU_RMAP_LOCKED to indicate we have taken the lock
1176 * at this higer level.
1178 mapping
= hugetlb_page_mapping_lock_write(hpage
);
1180 unmap_success
= try_to_unmap(hpage
,
1181 ttu
|TTU_RMAP_LOCKED
);
1182 i_mmap_unlock_write(mapping
);
1184 pr_info("Memory failure: %#lx: could not lock mapping for mapped huge page\n", pfn
);
1185 unmap_success
= false;
1188 unmap_success
= try_to_unmap(hpage
, ttu
);
1192 pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n",
1193 pfn
, page_mapcount(hpage
));
1196 * try_to_unmap() might put mlocked page in lru cache, so call
1197 * shake_page() again to ensure that it's flushed.
1200 shake_page(hpage
, 0);
1203 * Now that the dirty bit has been propagated to the
1204 * struct page and all unmaps done we can decide if
1205 * killing is needed or not. Only kill when the page
1206 * was dirty or the process is not restartable,
1207 * otherwise the tokill list is merely
1208 * freed. When there was a problem unmapping earlier
1209 * use a more force-full uncatchable kill to prevent
1210 * any accesses to the poisoned memory.
1212 forcekill
= PageDirty(hpage
) || (flags
& MF_MUST_KILL
);
1213 kill_procs(&tokill
, forcekill
, !unmap_success
, pfn
, flags
);
1215 return unmap_success
;
1218 static int identify_page_state(unsigned long pfn
, struct page
*p
,
1219 unsigned long page_flags
)
1221 struct page_state
*ps
;
1224 * The first check uses the current page flags which may not have any
1225 * relevant information. The second check with the saved page flags is
1226 * carried out only if the first check can't determine the page status.
1228 for (ps
= error_states
;; ps
++)
1229 if ((p
->flags
& ps
->mask
) == ps
->res
)
1232 page_flags
|= (p
->flags
& (1UL << PG_dirty
));
1235 for (ps
= error_states
;; ps
++)
1236 if ((page_flags
& ps
->mask
) == ps
->res
)
1238 return page_action(ps
, p
, pfn
);
1241 static int try_to_split_thp_page(struct page
*page
, const char *msg
)
1244 if (!PageAnon(page
) || unlikely(split_huge_page(page
))) {
1245 unsigned long pfn
= page_to_pfn(page
);
1248 if (!PageAnon(page
))
1249 pr_info("%s: %#lx: non anonymous thp\n", msg
, pfn
);
1251 pr_info("%s: %#lx: thp split failed\n", msg
, pfn
);
1260 static int memory_failure_hugetlb(unsigned long pfn
, int flags
)
1262 struct page
*p
= pfn_to_page(pfn
);
1263 struct page
*head
= compound_head(p
);
1265 unsigned long page_flags
;
1267 if (TestSetPageHWPoison(head
)) {
1268 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1273 num_poisoned_pages_inc();
1275 if (!(flags
& MF_COUNT_INCREASED
) && !get_hwpoison_page(p
, flags
, 0)) {
1277 * Check "filter hit" and "race with other subpage."
1280 if (PageHWPoison(head
)) {
1281 if ((hwpoison_filter(p
) && TestClearPageHWPoison(p
))
1282 || (p
!= head
&& TestSetPageHWPoison(head
))) {
1283 num_poisoned_pages_dec();
1290 if (!dissolve_free_huge_page(p
) && take_page_off_buddy(p
)) {
1294 action_result(pfn
, MF_MSG_FREE_HUGE
, res
);
1295 return res
== MF_RECOVERED
? 0 : -EBUSY
;
1299 page_flags
= head
->flags
;
1301 if (!PageHWPoison(head
)) {
1302 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn
);
1303 num_poisoned_pages_dec();
1310 * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so
1311 * simply disable it. In order to make it work properly, we need
1313 * - conversion of a pud that maps an error hugetlb into hwpoison
1314 * entry properly works, and
1315 * - other mm code walking over page table is aware of pud-aligned
1318 if (huge_page_size(page_hstate(head
)) > PMD_SIZE
) {
1319 action_result(pfn
, MF_MSG_NON_PMD_HUGE
, MF_IGNORED
);
1324 if (!hwpoison_user_mappings(p
, pfn
, flags
, &head
)) {
1325 action_result(pfn
, MF_MSG_UNMAP_FAILED
, MF_IGNORED
);
1330 return identify_page_state(pfn
, p
, page_flags
);
1336 static int memory_failure_dev_pagemap(unsigned long pfn
, int flags
,
1337 struct dev_pagemap
*pgmap
)
1339 struct page
*page
= pfn_to_page(pfn
);
1340 const bool unmap_success
= true;
1341 unsigned long size
= 0;
1348 if (flags
& MF_COUNT_INCREASED
)
1350 * Drop the extra refcount in case we come from madvise().
1354 /* device metadata space is not recoverable */
1355 if (!pgmap_pfn_valid(pgmap
, pfn
)) {
1361 * Prevent the inode from being freed while we are interrogating
1362 * the address_space, typically this would be handled by
1363 * lock_page(), but dax pages do not use the page lock. This
1364 * also prevents changes to the mapping of this pfn until
1365 * poison signaling is complete.
1367 cookie
= dax_lock_page(page
);
1371 if (hwpoison_filter(page
)) {
1376 if (pgmap
->type
== MEMORY_DEVICE_PRIVATE
) {
1378 * TODO: Handle HMM pages which may need coordination
1379 * with device-side memory.
1385 * Use this flag as an indication that the dax page has been
1386 * remapped UC to prevent speculative consumption of poison.
1388 SetPageHWPoison(page
);
1391 * Unlike System-RAM there is no possibility to swap in a
1392 * different physical page at a given virtual address, so all
1393 * userspace consumption of ZONE_DEVICE memory necessitates
1394 * SIGBUS (i.e. MF_MUST_KILL)
1396 flags
|= MF_ACTION_REQUIRED
| MF_MUST_KILL
;
1397 collect_procs(page
, &tokill
, flags
& MF_ACTION_REQUIRED
);
1399 list_for_each_entry(tk
, &tokill
, nd
)
1401 size
= max(size
, 1UL << tk
->size_shift
);
1404 * Unmap the largest mapping to avoid breaking up
1405 * device-dax mappings which are constant size. The
1406 * actual size of the mapping being torn down is
1407 * communicated in siginfo, see kill_proc()
1409 start
= (page
->index
<< PAGE_SHIFT
) & ~(size
- 1);
1410 unmap_mapping_range(page
->mapping
, start
, size
, 0);
1412 kill_procs(&tokill
, flags
& MF_MUST_KILL
, !unmap_success
, pfn
, flags
);
1415 dax_unlock_page(page
, cookie
);
1417 /* drop pgmap ref acquired in caller */
1418 put_dev_pagemap(pgmap
);
1419 action_result(pfn
, MF_MSG_DAX
, rc
? MF_FAILED
: MF_RECOVERED
);
1424 * memory_failure - Handle memory failure of a page.
1425 * @pfn: Page Number of the corrupted page
1426 * @flags: fine tune action taken
1428 * This function is called by the low level machine check code
1429 * of an architecture when it detects hardware memory corruption
1430 * of a page. It tries its best to recover, which includes
1431 * dropping pages, killing processes etc.
1433 * The function is primarily of use for corruptions that
1434 * happen outside the current execution context (e.g. when
1435 * detected by a background scrubber)
1437 * Must run in process context (e.g. a work queue) with interrupts
1438 * enabled and no spinlocks hold.
1440 int memory_failure(unsigned long pfn
, int flags
)
1444 struct page
*orig_head
;
1445 struct dev_pagemap
*pgmap
;
1447 unsigned long page_flags
;
1449 static DEFINE_MUTEX(mf_mutex
);
1451 if (!sysctl_memory_failure_recovery
)
1452 panic("Memory failure on page %lx", pfn
);
1454 p
= pfn_to_online_page(pfn
);
1456 if (pfn_valid(pfn
)) {
1457 pgmap
= get_dev_pagemap(pfn
, NULL
);
1459 return memory_failure_dev_pagemap(pfn
, flags
,
1462 pr_err("Memory failure: %#lx: memory outside kernel control\n",
1467 mutex_lock(&mf_mutex
);
1471 res
= memory_failure_hugetlb(pfn
, flags
);
1475 if (TestSetPageHWPoison(p
)) {
1476 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1482 orig_head
= hpage
= compound_head(p
);
1483 num_poisoned_pages_inc();
1486 * We need/can do nothing about count=0 pages.
1487 * 1) it's a free page, and therefore in safe hand:
1488 * prep_new_page() will be the gate keeper.
1489 * 2) it's part of a non-compound high order page.
1490 * Implies some kernel user: cannot stop them from
1491 * R/W the page; let's pray that the page has been
1492 * used and will be freed some time later.
1493 * In fact it's dangerous to directly bump up page count from 0,
1494 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
1496 if (!(flags
& MF_COUNT_INCREASED
) && !get_hwpoison_page(p
, flags
, 0)) {
1497 if (is_free_buddy_page(p
)) {
1498 if (take_page_off_buddy(p
)) {
1502 /* We lost the race, try again */
1504 ClearPageHWPoison(p
);
1505 num_poisoned_pages_dec();
1511 action_result(pfn
, MF_MSG_BUDDY
, res
);
1512 res
= res
== MF_RECOVERED
? 0 : -EBUSY
;
1514 action_result(pfn
, MF_MSG_KERNEL_HIGH_ORDER
, MF_IGNORED
);
1520 if (PageTransHuge(hpage
)) {
1521 if (try_to_split_thp_page(p
, "Memory Failure") < 0) {
1522 action_result(pfn
, MF_MSG_UNSPLIT_THP
, MF_IGNORED
);
1526 VM_BUG_ON_PAGE(!page_count(p
), p
);
1530 * We ignore non-LRU pages for good reasons.
1531 * - PG_locked is only well defined for LRU pages and a few others
1532 * - to avoid races with __SetPageLocked()
1533 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1534 * The check (unnecessarily) ignores LRU pages being isolated and
1535 * walked by the page reclaim code, however that's not a big loss.
1542 * The page could have changed compound pages during the locking.
1543 * If this happens just bail out.
1545 if (PageCompound(p
) && compound_head(p
) != orig_head
) {
1546 action_result(pfn
, MF_MSG_DIFFERENT_COMPOUND
, MF_IGNORED
);
1552 * We use page flags to determine what action should be taken, but
1553 * the flags can be modified by the error containment action. One
1554 * example is an mlocked page, where PG_mlocked is cleared by
1555 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1556 * correctly, we save a copy of the page flags at this time.
1558 page_flags
= p
->flags
;
1561 * unpoison always clear PG_hwpoison inside page lock
1563 if (!PageHWPoison(p
)) {
1564 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn
);
1565 num_poisoned_pages_dec();
1570 if (hwpoison_filter(p
)) {
1571 if (TestClearPageHWPoison(p
))
1572 num_poisoned_pages_dec();
1579 * __munlock_pagevec may clear a writeback page's LRU flag without
1580 * page_lock. We need wait writeback completion for this page or it
1581 * may trigger vfs BUG while evict inode.
1583 if (!PageTransTail(p
) && !PageLRU(p
) && !PageWriteback(p
))
1584 goto identify_page_state
;
1587 * It's very difficult to mess with pages currently under IO
1588 * and in many cases impossible, so we just avoid it here.
1590 wait_on_page_writeback(p
);
1593 * Now take care of user space mappings.
1594 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1596 if (!hwpoison_user_mappings(p
, pfn
, flags
, &p
)) {
1597 action_result(pfn
, MF_MSG_UNMAP_FAILED
, MF_IGNORED
);
1603 * Torn down by someone else?
1605 if (PageLRU(p
) && !PageSwapCache(p
) && p
->mapping
== NULL
) {
1606 action_result(pfn
, MF_MSG_TRUNCATED_LRU
, MF_IGNORED
);
1611 identify_page_state
:
1612 res
= identify_page_state(pfn
, p
, page_flags
);
1613 mutex_unlock(&mf_mutex
);
1618 mutex_unlock(&mf_mutex
);
1621 EXPORT_SYMBOL_GPL(memory_failure
);
1623 #define MEMORY_FAILURE_FIFO_ORDER 4
1624 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1626 struct memory_failure_entry
{
1631 struct memory_failure_cpu
{
1632 DECLARE_KFIFO(fifo
, struct memory_failure_entry
,
1633 MEMORY_FAILURE_FIFO_SIZE
);
1635 struct work_struct work
;
1638 static DEFINE_PER_CPU(struct memory_failure_cpu
, memory_failure_cpu
);
1641 * memory_failure_queue - Schedule handling memory failure of a page.
1642 * @pfn: Page Number of the corrupted page
1643 * @flags: Flags for memory failure handling
1645 * This function is called by the low level hardware error handler
1646 * when it detects hardware memory corruption of a page. It schedules
1647 * the recovering of error page, including dropping pages, killing
1650 * The function is primarily of use for corruptions that
1651 * happen outside the current execution context (e.g. when
1652 * detected by a background scrubber)
1654 * Can run in IRQ context.
1656 void memory_failure_queue(unsigned long pfn
, int flags
)
1658 struct memory_failure_cpu
*mf_cpu
;
1659 unsigned long proc_flags
;
1660 struct memory_failure_entry entry
= {
1665 mf_cpu
= &get_cpu_var(memory_failure_cpu
);
1666 spin_lock_irqsave(&mf_cpu
->lock
, proc_flags
);
1667 if (kfifo_put(&mf_cpu
->fifo
, entry
))
1668 schedule_work_on(smp_processor_id(), &mf_cpu
->work
);
1670 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1672 spin_unlock_irqrestore(&mf_cpu
->lock
, proc_flags
);
1673 put_cpu_var(memory_failure_cpu
);
1675 EXPORT_SYMBOL_GPL(memory_failure_queue
);
1677 static void memory_failure_work_func(struct work_struct
*work
)
1679 struct memory_failure_cpu
*mf_cpu
;
1680 struct memory_failure_entry entry
= { 0, };
1681 unsigned long proc_flags
;
1684 mf_cpu
= container_of(work
, struct memory_failure_cpu
, work
);
1686 spin_lock_irqsave(&mf_cpu
->lock
, proc_flags
);
1687 gotten
= kfifo_get(&mf_cpu
->fifo
, &entry
);
1688 spin_unlock_irqrestore(&mf_cpu
->lock
, proc_flags
);
1691 if (entry
.flags
& MF_SOFT_OFFLINE
)
1692 soft_offline_page(entry
.pfn
, entry
.flags
);
1694 memory_failure(entry
.pfn
, entry
.flags
);
1699 * Process memory_failure work queued on the specified CPU.
1700 * Used to avoid return-to-userspace racing with the memory_failure workqueue.
1702 void memory_failure_queue_kick(int cpu
)
1704 struct memory_failure_cpu
*mf_cpu
;
1706 mf_cpu
= &per_cpu(memory_failure_cpu
, cpu
);
1707 cancel_work_sync(&mf_cpu
->work
);
1708 memory_failure_work_func(&mf_cpu
->work
);
1711 static int __init
memory_failure_init(void)
1713 struct memory_failure_cpu
*mf_cpu
;
1716 for_each_possible_cpu(cpu
) {
1717 mf_cpu
= &per_cpu(memory_failure_cpu
, cpu
);
1718 spin_lock_init(&mf_cpu
->lock
);
1719 INIT_KFIFO(mf_cpu
->fifo
);
1720 INIT_WORK(&mf_cpu
->work
, memory_failure_work_func
);
1725 core_initcall(memory_failure_init
);
1727 #define unpoison_pr_info(fmt, pfn, rs) \
1729 if (__ratelimit(rs)) \
1730 pr_info(fmt, pfn); \
1734 * unpoison_memory - Unpoison a previously poisoned page
1735 * @pfn: Page number of the to be unpoisoned page
1737 * Software-unpoison a page that has been poisoned by
1738 * memory_failure() earlier.
1740 * This is only done on the software-level, so it only works
1741 * for linux injected failures, not real hardware failures
1743 * Returns 0 for success, otherwise -errno.
1745 int unpoison_memory(unsigned long pfn
)
1750 unsigned long flags
= 0;
1751 static DEFINE_RATELIMIT_STATE(unpoison_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1752 DEFAULT_RATELIMIT_BURST
);
1754 if (!pfn_valid(pfn
))
1757 p
= pfn_to_page(pfn
);
1758 page
= compound_head(p
);
1760 if (!PageHWPoison(p
)) {
1761 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
1766 if (page_count(page
) > 1) {
1767 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
1772 if (page_mapped(page
)) {
1773 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
1778 if (page_mapping(page
)) {
1779 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
1785 * unpoison_memory() can encounter thp only when the thp is being
1786 * worked by memory_failure() and the page lock is not held yet.
1787 * In such case, we yield to memory_failure() and make unpoison fail.
1789 if (!PageHuge(page
) && PageTransHuge(page
)) {
1790 unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n",
1795 if (!get_hwpoison_page(p
, flags
, 0)) {
1796 if (TestClearPageHWPoison(p
))
1797 num_poisoned_pages_dec();
1798 unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n",
1805 * This test is racy because PG_hwpoison is set outside of page lock.
1806 * That's acceptable because that won't trigger kernel panic. Instead,
1807 * the PG_hwpoison page will be caught and isolated on the entrance to
1808 * the free buddy page pool.
1810 if (TestClearPageHWPoison(page
)) {
1811 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
1813 num_poisoned_pages_dec();
1819 if (freeit
&& !(pfn
== my_zero_pfn(0) && page_count(p
) == 1))
1824 EXPORT_SYMBOL(unpoison_memory
);
1826 static bool isolate_page(struct page
*page
, struct list_head
*pagelist
)
1828 bool isolated
= false;
1829 bool lru
= PageLRU(page
);
1831 if (PageHuge(page
)) {
1832 isolated
= isolate_huge_page(page
, pagelist
);
1835 isolated
= !isolate_lru_page(page
);
1837 isolated
= !isolate_movable_page(page
, ISOLATE_UNEVICTABLE
);
1840 list_add(&page
->lru
, pagelist
);
1843 if (isolated
&& lru
)
1844 inc_node_page_state(page
, NR_ISOLATED_ANON
+
1845 page_is_file_lru(page
));
1848 * If we succeed to isolate the page, we grabbed another refcount on
1849 * the page, so we can safely drop the one we got from get_any_pages().
1850 * If we failed to isolate the page, it means that we cannot go further
1851 * and we will return an error, so drop the reference we got from
1852 * get_any_pages() as well.
1859 * __soft_offline_page handles hugetlb-pages and non-hugetlb pages.
1860 * If the page is a non-dirty unmapped page-cache page, it simply invalidates.
1861 * If the page is mapped, it migrates the contents over.
1863 static int __soft_offline_page(struct page
*page
)
1866 unsigned long pfn
= page_to_pfn(page
);
1867 struct page
*hpage
= compound_head(page
);
1868 char const *msg_page
[] = {"page", "hugepage"};
1869 bool huge
= PageHuge(page
);
1870 LIST_HEAD(pagelist
);
1871 struct migration_target_control mtc
= {
1872 .nid
= NUMA_NO_NODE
,
1873 .gfp_mask
= GFP_USER
| __GFP_MOVABLE
| __GFP_RETRY_MAYFAIL
,
1877 * Check PageHWPoison again inside page lock because PageHWPoison
1878 * is set by memory_failure() outside page lock. Note that
1879 * memory_failure() also double-checks PageHWPoison inside page lock,
1880 * so there's no race between soft_offline_page() and memory_failure().
1883 if (!PageHuge(page
))
1884 wait_on_page_writeback(page
);
1885 if (PageHWPoison(page
)) {
1888 pr_info("soft offline: %#lx page already poisoned\n", pfn
);
1892 if (!PageHuge(page
))
1894 * Try to invalidate first. This should work for
1895 * non dirty unmapped page cache pages.
1897 ret
= invalidate_inode_page(page
);
1901 * RED-PEN would be better to keep it isolated here, but we
1902 * would need to fix isolation locking first.
1905 pr_info("soft_offline: %#lx: invalidated\n", pfn
);
1906 page_handle_poison(page
, false, true);
1910 if (isolate_page(hpage
, &pagelist
)) {
1911 ret
= migrate_pages(&pagelist
, alloc_migration_target
, NULL
,
1912 (unsigned long)&mtc
, MIGRATE_SYNC
, MR_MEMORY_FAILURE
);
1914 bool release
= !huge
;
1916 if (!page_handle_poison(page
, huge
, release
))
1919 if (!list_empty(&pagelist
))
1920 putback_movable_pages(&pagelist
);
1922 pr_info("soft offline: %#lx: %s migration failed %d, type %lx (%pGp)\n",
1923 pfn
, msg_page
[huge
], ret
, page
->flags
, &page
->flags
);
1928 pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %lx (%pGp)\n",
1929 pfn
, msg_page
[huge
], page_count(page
), page
->flags
, &page
->flags
);
1935 static int soft_offline_in_use_page(struct page
*page
)
1937 struct page
*hpage
= compound_head(page
);
1939 if (!PageHuge(page
) && PageTransHuge(hpage
))
1940 if (try_to_split_thp_page(page
, "soft offline") < 0)
1942 return __soft_offline_page(page
);
1945 static int soft_offline_free_page(struct page
*page
)
1949 if (!page_handle_poison(page
, true, false))
1955 static void put_ref_page(struct page
*page
)
1962 * soft_offline_page - Soft offline a page.
1963 * @pfn: pfn to soft-offline
1964 * @flags: flags. Same as memory_failure().
1966 * Returns 0 on success, otherwise negated errno.
1968 * Soft offline a page, by migration or invalidation,
1969 * without killing anything. This is for the case when
1970 * a page is not corrupted yet (so it's still valid to access),
1971 * but has had a number of corrected errors and is better taken
1974 * The actual policy on when to do that is maintained by
1977 * This should never impact any application or cause data loss,
1978 * however it might take some time.
1980 * This is not a 100% solution for all memory, but tries to be
1981 * ``good enough'' for the majority of memory.
1983 int soft_offline_page(unsigned long pfn
, int flags
)
1986 bool try_again
= true;
1987 struct page
*page
, *ref_page
= NULL
;
1989 WARN_ON_ONCE(!pfn_valid(pfn
) && (flags
& MF_COUNT_INCREASED
));
1991 if (!pfn_valid(pfn
))
1993 if (flags
& MF_COUNT_INCREASED
)
1994 ref_page
= pfn_to_page(pfn
);
1996 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
1997 page
= pfn_to_online_page(pfn
);
1999 put_ref_page(ref_page
);
2003 if (PageHWPoison(page
)) {
2004 pr_info("%s: %#lx page already poisoned\n", __func__
, pfn
);
2005 put_ref_page(ref_page
);
2011 ret
= get_hwpoison_page(page
, flags
, MF_SOFT_OFFLINE
);
2015 ret
= soft_offline_in_use_page(page
);
2016 } else if (ret
== 0) {
2017 if (soft_offline_free_page(page
) && try_again
) {
2021 } else if (ret
== -EIO
) {
2022 pr_info("%s: %#lx: unknown page type: %lx (%pGp)\n",
2023 __func__
, pfn
, page
->flags
, &page
->flags
);