2 * Copyright (C) 2008, 2009 Intel Corporation
3 * Authors: Andi Kleen, Fengguang Wu
5 * This software may be redistributed and/or modified under the terms of
6 * the GNU General Public License ("GPL") version 2 only as published by the
7 * Free Software Foundation.
9 * High level machine check handler. Handles pages reported by the
10 * hardware as being corrupted usually due to a multi-bit ECC memory or cache
13 * In addition there is a "soft offline" entry point that allows stop using
14 * not-yet-corrupted-by-suspicious pages without killing anything.
16 * Handles page cache pages in various states. The tricky part
17 * here is that we can access any page asynchronously in respect to
18 * other VM users, because memory failures could happen anytime and
19 * anywhere. This could violate some of their assumptions. This is why
20 * this code has to be extremely careful. Generally it tries to use
21 * normal locking rules, as in get the standard locks, even if that means
22 * the error handling takes potentially a long time.
24 * There are several operations here with exponential complexity because
25 * of unsuitable VM data structures. For example the operation to map back
26 * from RMAP chains to processes has to walk the complete process list and
27 * has non linear complexity with the number. But since memory corruptions
28 * are rare we hope to get away with this. This avoids impacting the core
34 * - hugetlb needs more code
35 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
36 * - pass bad pages to kdump next kernel
38 #include <linux/kernel.h>
40 #include <linux/page-flags.h>
41 #include <linux/kernel-page-flags.h>
42 #include <linux/sched.h>
43 #include <linux/ksm.h>
44 #include <linux/rmap.h>
45 #include <linux/export.h>
46 #include <linux/pagemap.h>
47 #include <linux/swap.h>
48 #include <linux/backing-dev.h>
49 #include <linux/migrate.h>
50 #include <linux/page-isolation.h>
51 #include <linux/suspend.h>
52 #include <linux/slab.h>
53 #include <linux/swapops.h>
54 #include <linux/hugetlb.h>
55 #include <linux/memory_hotplug.h>
56 #include <linux/mm_inline.h>
57 #include <linux/kfifo.h>
60 int sysctl_memory_failure_early_kill __read_mostly = 0;
62 int sysctl_memory_failure_recovery __read_mostly = 1;
64 atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
66 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
68 u32 hwpoison_filter_enable = 0;
69 u32 hwpoison_filter_dev_major = ~0U;
70 u32 hwpoison_filter_dev_minor = ~0U;
71 u64 hwpoison_filter_flags_mask;
72 u64 hwpoison_filter_flags_value;
73 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
74 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
75 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
76 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
77 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
79 static int hwpoison_filter_dev(struct page *p)
81 struct address_space *mapping;
84 if (hwpoison_filter_dev_major == ~0U &&
85 hwpoison_filter_dev_minor == ~0U)
89 * page_mapping() does not accept slab pages.
94 mapping = page_mapping(p);
95 if (mapping == NULL || mapping->host == NULL)
98 dev = mapping->host->i_sb->s_dev;
99 if (hwpoison_filter_dev_major != ~0U &&
100 hwpoison_filter_dev_major != MAJOR(dev))
102 if (hwpoison_filter_dev_minor != ~0U &&
103 hwpoison_filter_dev_minor != MINOR(dev))
109 static int hwpoison_filter_flags(struct page *p)
111 if (!hwpoison_filter_flags_mask)
114 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
115 hwpoison_filter_flags_value)
122 * This allows stress tests to limit test scope to a collection of tasks
123 * by putting them under some memcg. This prevents killing unrelated/important
124 * processes such as /sbin/init. Note that the target task may share clean
125 * pages with init (eg. libc text), which is harmless. If the target task
126 * share _dirty_ pages with another task B, the test scheme must make sure B
127 * is also included in the memcg. At last, due to race conditions this filter
128 * can only guarantee that the page either belongs to the memcg tasks, or is
131 #ifdef CONFIG_MEMCG_SWAP
132 u64 hwpoison_filter_memcg;
133 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
134 static int hwpoison_filter_task(struct page *p)
136 struct mem_cgroup *mem;
137 struct cgroup_subsys_state *css;
140 if (!hwpoison_filter_memcg)
143 mem = try_get_mem_cgroup_from_page(p);
147 css = mem_cgroup_css(mem);
148 /* root_mem_cgroup has NULL dentries */
149 if (!css->cgroup->dentry)
152 ino = css->cgroup->dentry->d_inode->i_ino;
155 if (ino != hwpoison_filter_memcg)
161 static int hwpoison_filter_task(struct page *p) { return 0; }
164 int hwpoison_filter(struct page *p)
166 if (!hwpoison_filter_enable)
169 if (hwpoison_filter_dev(p))
172 if (hwpoison_filter_flags(p))
175 if (hwpoison_filter_task(p))
181 int hwpoison_filter(struct page *p)
187 EXPORT_SYMBOL_GPL(hwpoison_filter);
190 * Send all the processes who have the page mapped a signal.
191 * ``action optional'' if they are not immediately affected by the error
192 * ``action required'' if error happened in current execution context
194 static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
195 unsigned long pfn, struct page *page, int flags)
201 "MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
202 pfn, t->comm, t->pid);
203 si.si_signo = SIGBUS;
205 si.si_addr = (void *)addr;
206 #ifdef __ARCH_SI_TRAPNO
207 si.si_trapno = trapno;
209 si.si_addr_lsb = compound_trans_order(compound_head(page)) + PAGE_SHIFT;
211 if ((flags & MF_ACTION_REQUIRED) && t == current) {
212 si.si_code = BUS_MCEERR_AR;
213 ret = force_sig_info(SIGBUS, &si, t);
216 * Don't use force here, it's convenient if the signal
217 * can be temporarily blocked.
218 * This could cause a loop when the user sets SIGBUS
219 * to SIG_IGN, but hopefully no one will do that?
221 si.si_code = BUS_MCEERR_AO;
222 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
225 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
226 t->comm, t->pid, ret);
231 * When a unknown page type is encountered drain as many buffers as possible
232 * in the hope to turn the page into a LRU or free page, which we can handle.
234 void shake_page(struct page *p, int access)
241 if (PageLRU(p) || is_free_buddy_page(p))
246 * Only call shrink_slab here (which would also shrink other caches) if
247 * access is not potentially fatal.
252 struct shrink_control shrink = {
253 .gfp_mask = GFP_KERNEL,
256 nr = shrink_slab(&shrink, 1000, 1000);
257 if (page_count(p) == 1)
262 EXPORT_SYMBOL_GPL(shake_page);
265 * Kill all processes that have a poisoned page mapped and then isolate
269 * Find all processes having the page mapped and kill them.
270 * But we keep a page reference around so that the page is not
271 * actually freed yet.
272 * Then stash the page away
274 * There's no convenient way to get back to mapped processes
275 * from the VMAs. So do a brute-force search over all
278 * Remember that machine checks are not common (or rather
279 * if they are common you have other problems), so this shouldn't
280 * be a performance issue.
282 * Also there are some races possible while we get from the
283 * error detection to actually handle it.
288 struct task_struct *tsk;
294 * Failure handling: if we can't find or can't kill a process there's
295 * not much we can do. We just print a message and ignore otherwise.
299 * Schedule a process for later kill.
300 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
301 * TBD would GFP_NOIO be enough?
303 static void add_to_kill(struct task_struct *tsk, struct page *p,
304 struct vm_area_struct *vma,
305 struct list_head *to_kill,
306 struct to_kill **tkc)
314 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
317 "MCE: Out of memory while machine check handling\n");
321 tk->addr = page_address_in_vma(p, vma);
325 * In theory we don't have to kill when the page was
326 * munmaped. But it could be also a mremap. Since that's
327 * likely very rare kill anyways just out of paranoia, but use
328 * a SIGKILL because the error is not contained anymore.
330 if (tk->addr == -EFAULT) {
331 pr_info("MCE: Unable to find user space address %lx in %s\n",
332 page_to_pfn(p), tsk->comm);
335 get_task_struct(tsk);
337 list_add_tail(&tk->nd, to_kill);
341 * Kill the processes that have been collected earlier.
343 * Only do anything when DOIT is set, otherwise just free the list
344 * (this is used for clean pages which do not need killing)
345 * Also when FAIL is set do a force kill because something went
348 static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
349 int fail, struct page *page, unsigned long pfn,
352 struct to_kill *tk, *next;
354 list_for_each_entry_safe (tk, next, to_kill, nd) {
357 * In case something went wrong with munmapping
358 * make sure the process doesn't catch the
359 * signal and then access the memory. Just kill it.
361 if (fail || tk->addr_valid == 0) {
363 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
364 pfn, tk->tsk->comm, tk->tsk->pid);
365 force_sig(SIGKILL, tk->tsk);
369 * In theory the process could have mapped
370 * something else on the address in-between. We could
371 * check for that, but we need to tell the
374 else if (kill_proc(tk->tsk, tk->addr, trapno,
375 pfn, page, flags) < 0)
377 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
378 pfn, tk->tsk->comm, tk->tsk->pid);
380 put_task_struct(tk->tsk);
385 static int task_early_kill(struct task_struct *tsk)
389 if (tsk->flags & PF_MCE_PROCESS)
390 return !!(tsk->flags & PF_MCE_EARLY);
391 return sysctl_memory_failure_early_kill;
395 * Collect processes when the error hit an anonymous page.
397 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
398 struct to_kill **tkc)
400 struct vm_area_struct *vma;
401 struct task_struct *tsk;
404 av = page_lock_anon_vma(page);
405 if (av == NULL) /* Not actually mapped anymore */
408 read_lock(&tasklist_lock);
409 for_each_process (tsk) {
410 struct anon_vma_chain *vmac;
412 if (!task_early_kill(tsk))
414 list_for_each_entry(vmac, &av->head, same_anon_vma) {
416 if (!page_mapped_in_vma(page, vma))
418 if (vma->vm_mm == tsk->mm)
419 add_to_kill(tsk, page, vma, to_kill, tkc);
422 read_unlock(&tasklist_lock);
423 page_unlock_anon_vma(av);
427 * Collect processes when the error hit a file mapped page.
429 static void collect_procs_file(struct page *page, struct list_head *to_kill,
430 struct to_kill **tkc)
432 struct vm_area_struct *vma;
433 struct task_struct *tsk;
434 struct address_space *mapping = page->mapping;
436 mutex_lock(&mapping->i_mmap_mutex);
437 read_lock(&tasklist_lock);
438 for_each_process(tsk) {
439 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
441 if (!task_early_kill(tsk))
444 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
447 * Send early kill signal to tasks where a vma covers
448 * the page but the corrupted page is not necessarily
449 * mapped it in its pte.
450 * Assume applications who requested early kill want
451 * to be informed of all such data corruptions.
453 if (vma->vm_mm == tsk->mm)
454 add_to_kill(tsk, page, vma, to_kill, tkc);
457 read_unlock(&tasklist_lock);
458 mutex_unlock(&mapping->i_mmap_mutex);
462 * Collect the processes who have the corrupted page mapped to kill.
463 * This is done in two steps for locking reasons.
464 * First preallocate one tokill structure outside the spin locks,
465 * so that we can kill at least one process reasonably reliable.
467 static void collect_procs(struct page *page, struct list_head *tokill)
474 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
478 collect_procs_anon(page, tokill, &tk);
480 collect_procs_file(page, tokill, &tk);
485 * Error handlers for various types of pages.
489 IGNORED, /* Error: cannot be handled */
490 FAILED, /* Error: handling failed */
491 DELAYED, /* Will be handled later */
492 RECOVERED, /* Successfully recovered */
495 static const char *action_name[] = {
496 [IGNORED] = "Ignored",
498 [DELAYED] = "Delayed",
499 [RECOVERED] = "Recovered",
503 * XXX: It is possible that a page is isolated from LRU cache,
504 * and then kept in swap cache or failed to remove from page cache.
505 * The page count will stop it from being freed by unpoison.
506 * Stress tests should be aware of this memory leak problem.
508 static int delete_from_lru_cache(struct page *p)
510 if (!isolate_lru_page(p)) {
512 * Clear sensible page flags, so that the buddy system won't
513 * complain when the page is unpoison-and-freed.
516 ClearPageUnevictable(p);
518 * drop the page count elevated by isolate_lru_page()
520 page_cache_release(p);
527 * Error hit kernel page.
528 * Do nothing, try to be lucky and not touch this instead. For a few cases we
529 * could be more sophisticated.
531 static int me_kernel(struct page *p, unsigned long pfn)
537 * Page in unknown state. Do nothing.
539 static int me_unknown(struct page *p, unsigned long pfn)
541 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
546 * Clean (or cleaned) page cache page.
548 static int me_pagecache_clean(struct page *p, unsigned long pfn)
552 struct address_space *mapping;
554 delete_from_lru_cache(p);
557 * For anonymous pages we're done the only reference left
558 * should be the one m_f() holds.
564 * Now truncate the page in the page cache. This is really
565 * more like a "temporary hole punch"
566 * Don't do this for block devices when someone else
567 * has a reference, because it could be file system metadata
568 * and that's not safe to truncate.
570 mapping = page_mapping(p);
573 * Page has been teared down in the meanwhile
579 * Truncation is a bit tricky. Enable it per file system for now.
581 * Open: to take i_mutex or not for this? Right now we don't.
583 if (mapping->a_ops->error_remove_page) {
584 err = mapping->a_ops->error_remove_page(mapping, p);
586 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
588 } else if (page_has_private(p) &&
589 !try_to_release_page(p, GFP_NOIO)) {
590 pr_info("MCE %#lx: failed to release buffers\n", pfn);
596 * If the file system doesn't support it just invalidate
597 * This fails on dirty or anything with private pages
599 if (invalidate_inode_page(p))
602 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
609 * Dirty cache page page
610 * Issues: when the error hit a hole page the error is not properly
613 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
615 struct address_space *mapping = page_mapping(p);
618 /* TBD: print more information about the file. */
621 * IO error will be reported by write(), fsync(), etc.
622 * who check the mapping.
623 * This way the application knows that something went
624 * wrong with its dirty file data.
626 * There's one open issue:
628 * The EIO will be only reported on the next IO
629 * operation and then cleared through the IO map.
630 * Normally Linux has two mechanisms to pass IO error
631 * first through the AS_EIO flag in the address space
632 * and then through the PageError flag in the page.
633 * Since we drop pages on memory failure handling the
634 * only mechanism open to use is through AS_AIO.
636 * This has the disadvantage that it gets cleared on
637 * the first operation that returns an error, while
638 * the PageError bit is more sticky and only cleared
639 * when the page is reread or dropped. If an
640 * application assumes it will always get error on
641 * fsync, but does other operations on the fd before
642 * and the page is dropped between then the error
643 * will not be properly reported.
645 * This can already happen even without hwpoisoned
646 * pages: first on metadata IO errors (which only
647 * report through AS_EIO) or when the page is dropped
650 * So right now we assume that the application DTRT on
651 * the first EIO, but we're not worse than other parts
654 mapping_set_error(mapping, EIO);
657 return me_pagecache_clean(p, pfn);
661 * Clean and dirty swap cache.
663 * Dirty swap cache page is tricky to handle. The page could live both in page
664 * cache and swap cache(ie. page is freshly swapped in). So it could be
665 * referenced concurrently by 2 types of PTEs:
666 * normal PTEs and swap PTEs. We try to handle them consistently by calling
667 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
669 * - clear dirty bit to prevent IO
671 * - but keep in the swap cache, so that when we return to it on
672 * a later page fault, we know the application is accessing
673 * corrupted data and shall be killed (we installed simple
674 * interception code in do_swap_page to catch it).
676 * Clean swap cache pages can be directly isolated. A later page fault will
677 * bring in the known good data from disk.
679 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
682 /* Trigger EIO in shmem: */
683 ClearPageUptodate(p);
685 if (!delete_from_lru_cache(p))
691 static int me_swapcache_clean(struct page *p, unsigned long pfn)
693 delete_from_swap_cache(p);
695 if (!delete_from_lru_cache(p))
702 * Huge pages. Needs work.
704 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
705 * To narrow down kill region to one page, we need to break up pmd.
707 static int me_huge_page(struct page *p, unsigned long pfn)
710 struct page *hpage = compound_head(p);
712 * We can safely recover from error on free or reserved (i.e.
713 * not in-use) hugepage by dequeuing it from freelist.
714 * To check whether a hugepage is in-use or not, we can't use
715 * page->lru because it can be used in other hugepage operations,
716 * such as __unmap_hugepage_range() and gather_surplus_pages().
717 * So instead we use page_mapping() and PageAnon().
718 * We assume that this function is called with page lock held,
719 * so there is no race between isolation and mapping/unmapping.
721 if (!(page_mapping(hpage) || PageAnon(hpage))) {
722 res = dequeue_hwpoisoned_huge_page(hpage);
730 * Various page states we can handle.
732 * A page state is defined by its current page->flags bits.
733 * The table matches them in order and calls the right handler.
735 * This is quite tricky because we can access page at any time
736 * in its live cycle, so all accesses have to be extremely careful.
738 * This is not complete. More states could be added.
739 * For any missing state don't attempt recovery.
742 #define dirty (1UL << PG_dirty)
743 #define sc (1UL << PG_swapcache)
744 #define unevict (1UL << PG_unevictable)
745 #define mlock (1UL << PG_mlocked)
746 #define writeback (1UL << PG_writeback)
747 #define lru (1UL << PG_lru)
748 #define swapbacked (1UL << PG_swapbacked)
749 #define head (1UL << PG_head)
750 #define tail (1UL << PG_tail)
751 #define compound (1UL << PG_compound)
752 #define slab (1UL << PG_slab)
753 #define reserved (1UL << PG_reserved)
755 static struct page_state {
759 int (*action)(struct page *p, unsigned long pfn);
761 { reserved, reserved, "reserved kernel", me_kernel },
763 * free pages are specially detected outside this table:
764 * PG_buddy pages only make a small fraction of all free pages.
768 * Could in theory check if slab page is free or if we can drop
769 * currently unused objects without touching them. But just
770 * treat it as standard kernel for now.
772 { slab, slab, "kernel slab", me_kernel },
774 #ifdef CONFIG_PAGEFLAGS_EXTENDED
775 { head, head, "huge", me_huge_page },
776 { tail, tail, "huge", me_huge_page },
778 { compound, compound, "huge", me_huge_page },
781 { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
782 { sc|dirty, sc, "swapcache", me_swapcache_clean },
784 { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
785 { unevict, unevict, "unevictable LRU", me_pagecache_clean},
787 { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
788 { mlock, mlock, "mlocked LRU", me_pagecache_clean },
790 { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
791 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
794 * Catchall entry: must be at end.
796 { 0, 0, "unknown page state", me_unknown },
812 static void action_result(unsigned long pfn, char *msg, int result)
814 struct page *page = pfn_to_page(pfn);
816 printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
818 PageDirty(page) ? "dirty " : "",
819 msg, action_name[result]);
822 static int page_action(struct page_state *ps, struct page *p,
828 result = ps->action(p, pfn);
829 action_result(pfn, ps->msg, result);
831 count = page_count(p) - 1;
832 if (ps->action == me_swapcache_dirty && result == DELAYED)
836 "MCE %#lx: %s page still referenced by %d users\n",
837 pfn, ps->msg, count);
841 /* Could do more checks here if page looks ok */
843 * Could adjust zone counters here to correct for the missing page.
846 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
850 * Do all that is necessary to remove user space mappings. Unmap
851 * the pages and send SIGBUS to the processes if the data was dirty.
853 static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
854 int trapno, int flags)
856 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
857 struct address_space *mapping;
860 int kill = 1, forcekill;
861 struct page *hpage = compound_head(p);
864 if (PageReserved(p) || PageSlab(p))
868 * This check implies we don't kill processes if their pages
869 * are in the swap cache early. Those are always late kills.
871 if (!page_mapped(hpage))
877 if (PageSwapCache(p)) {
879 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
880 ttu |= TTU_IGNORE_HWPOISON;
884 * Propagate the dirty bit from PTEs to struct page first, because we
885 * need this to decide if we should kill or just drop the page.
886 * XXX: the dirty test could be racy: set_page_dirty() may not always
887 * be called inside page lock (it's recommended but not enforced).
889 mapping = page_mapping(hpage);
890 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
891 mapping_cap_writeback_dirty(mapping)) {
892 if (page_mkclean(hpage)) {
896 ttu |= TTU_IGNORE_HWPOISON;
898 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
904 * ppage: poisoned page
905 * if p is regular page(4k page)
906 * ppage == real poisoned page;
907 * else p is hugetlb or THP, ppage == head page.
911 if (PageTransHuge(hpage)) {
913 * Verify that this isn't a hugetlbfs head page, the check for
914 * PageAnon is just for avoid tripping a split_huge_page
915 * internal debug check, as split_huge_page refuses to deal with
916 * anything that isn't an anon page. PageAnon can't go away fro
917 * under us because we hold a refcount on the hpage, without a
918 * refcount on the hpage. split_huge_page can't be safely called
919 * in the first place, having a refcount on the tail isn't
920 * enough * to be safe.
922 if (!PageHuge(hpage) && PageAnon(hpage)) {
923 if (unlikely(split_huge_page(hpage))) {
925 * FIXME: if splitting THP is failed, it is
926 * better to stop the following operation rather
927 * than causing panic by unmapping. System might
928 * survive if the page is freed later.
931 "MCE %#lx: failed to split THP\n", pfn);
933 BUG_ON(!PageHWPoison(p));
936 /* THP is split, so ppage should be the real poisoned page. */
942 * First collect all the processes that have the page
943 * mapped in dirty form. This has to be done before try_to_unmap,
944 * because ttu takes the rmap data structures down.
946 * Error handling: We ignore errors here because
947 * there's nothing that can be done.
950 collect_procs(ppage, &tokill);
955 ret = try_to_unmap(ppage, ttu);
956 if (ret != SWAP_SUCCESS)
957 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
958 pfn, page_mapcount(ppage));
964 * Now that the dirty bit has been propagated to the
965 * struct page and all unmaps done we can decide if
966 * killing is needed or not. Only kill when the page
967 * was dirty or the process is not restartable,
968 * otherwise the tokill list is merely
969 * freed. When there was a problem unmapping earlier
970 * use a more force-full uncatchable kill to prevent
971 * any accesses to the poisoned memory.
973 forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL);
974 kill_procs(&tokill, forcekill, trapno,
975 ret != SWAP_SUCCESS, p, pfn, flags);
980 static void set_page_hwpoison_huge_page(struct page *hpage)
983 int nr_pages = 1 << compound_trans_order(hpage);
984 for (i = 0; i < nr_pages; i++)
985 SetPageHWPoison(hpage + i);
988 static void clear_page_hwpoison_huge_page(struct page *hpage)
991 int nr_pages = 1 << compound_trans_order(hpage);
992 for (i = 0; i < nr_pages; i++)
993 ClearPageHWPoison(hpage + i);
997 * memory_failure - Handle memory failure of a page.
998 * @pfn: Page Number of the corrupted page
999 * @trapno: Trap number reported in the signal to user space.
1000 * @flags: fine tune action taken
1002 * This function is called by the low level machine check code
1003 * of an architecture when it detects hardware memory corruption
1004 * of a page. It tries its best to recover, which includes
1005 * dropping pages, killing processes etc.
1007 * The function is primarily of use for corruptions that
1008 * happen outside the current execution context (e.g. when
1009 * detected by a background scrubber)
1011 * Must run in process context (e.g. a work queue) with interrupts
1012 * enabled and no spinlocks hold.
1014 int memory_failure(unsigned long pfn, int trapno, int flags)
1016 struct page_state *ps;
1020 unsigned int nr_pages;
1022 if (!sysctl_memory_failure_recovery)
1023 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1025 if (!pfn_valid(pfn)) {
1027 "MCE %#lx: memory outside kernel control\n",
1032 p = pfn_to_page(pfn);
1033 hpage = compound_head(p);
1034 if (TestSetPageHWPoison(p)) {
1035 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
1039 nr_pages = 1 << compound_trans_order(hpage);
1040 atomic_long_add(nr_pages, &mce_bad_pages);
1043 * We need/can do nothing about count=0 pages.
1044 * 1) it's a free page, and therefore in safe hand:
1045 * prep_new_page() will be the gate keeper.
1046 * 2) it's a free hugepage, which is also safe:
1047 * an affected hugepage will be dequeued from hugepage freelist,
1048 * so there's no concern about reusing it ever after.
1049 * 3) it's part of a non-compound high order page.
1050 * Implies some kernel user: cannot stop them from
1051 * R/W the page; let's pray that the page has been
1052 * used and will be freed some time later.
1053 * In fact it's dangerous to directly bump up page count from 0,
1054 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1056 if (!(flags & MF_COUNT_INCREASED) &&
1057 !get_page_unless_zero(hpage)) {
1058 if (is_free_buddy_page(p)) {
1059 action_result(pfn, "free buddy", DELAYED);
1061 } else if (PageHuge(hpage)) {
1063 * Check "just unpoisoned", "filter hit", and
1064 * "race with other subpage."
1067 if (!PageHWPoison(hpage)
1068 || (hwpoison_filter(p) && TestClearPageHWPoison(p))
1069 || (p != hpage && TestSetPageHWPoison(hpage))) {
1070 atomic_long_sub(nr_pages, &mce_bad_pages);
1073 set_page_hwpoison_huge_page(hpage);
1074 res = dequeue_hwpoisoned_huge_page(hpage);
1075 action_result(pfn, "free huge",
1076 res ? IGNORED : DELAYED);
1080 action_result(pfn, "high order kernel", IGNORED);
1086 * We ignore non-LRU pages for good reasons.
1087 * - PG_locked is only well defined for LRU pages and a few others
1088 * - to avoid races with __set_page_locked()
1089 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1090 * The check (unnecessarily) ignores LRU pages being isolated and
1091 * walked by the page reclaim code, however that's not a big loss.
1093 if (!PageHuge(p) && !PageTransTail(p)) {
1098 * shake_page could have turned it free.
1100 if (is_free_buddy_page(p)) {
1101 action_result(pfn, "free buddy, 2nd try",
1105 action_result(pfn, "non LRU", IGNORED);
1112 * Lock the page and wait for writeback to finish.
1113 * It's very difficult to mess with pages currently under IO
1114 * and in many cases impossible, so we just avoid it here.
1119 * unpoison always clear PG_hwpoison inside page lock
1121 if (!PageHWPoison(p)) {
1122 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1126 if (hwpoison_filter(p)) {
1127 if (TestClearPageHWPoison(p))
1128 atomic_long_sub(nr_pages, &mce_bad_pages);
1135 * For error on the tail page, we should set PG_hwpoison
1136 * on the head page to show that the hugepage is hwpoisoned
1138 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1139 action_result(pfn, "hugepage already hardware poisoned",
1146 * Set PG_hwpoison on all pages in an error hugepage,
1147 * because containment is done in hugepage unit for now.
1148 * Since we have done TestSetPageHWPoison() for the head page with
1149 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1152 set_page_hwpoison_huge_page(hpage);
1154 wait_on_page_writeback(p);
1157 * Now take care of user space mappings.
1158 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1160 if (hwpoison_user_mappings(p, pfn, trapno, flags) != SWAP_SUCCESS) {
1161 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1167 * Torn down by someone else?
1169 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1170 action_result(pfn, "already truncated LRU", IGNORED);
1176 for (ps = error_states;; ps++) {
1177 if ((p->flags & ps->mask) == ps->res) {
1178 res = page_action(ps, p, pfn);
1186 EXPORT_SYMBOL_GPL(memory_failure);
1188 #define MEMORY_FAILURE_FIFO_ORDER 4
1189 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1191 struct memory_failure_entry {
1197 struct memory_failure_cpu {
1198 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1199 MEMORY_FAILURE_FIFO_SIZE);
1201 struct work_struct work;
1204 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1207 * memory_failure_queue - Schedule handling memory failure of a page.
1208 * @pfn: Page Number of the corrupted page
1209 * @trapno: Trap number reported in the signal to user space.
1210 * @flags: Flags for memory failure handling
1212 * This function is called by the low level hardware error handler
1213 * when it detects hardware memory corruption of a page. It schedules
1214 * the recovering of error page, including dropping pages, killing
1217 * The function is primarily of use for corruptions that
1218 * happen outside the current execution context (e.g. when
1219 * detected by a background scrubber)
1221 * Can run in IRQ context.
1223 void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1225 struct memory_failure_cpu *mf_cpu;
1226 unsigned long proc_flags;
1227 struct memory_failure_entry entry = {
1233 mf_cpu = &get_cpu_var(memory_failure_cpu);
1234 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1235 if (kfifo_put(&mf_cpu->fifo, &entry))
1236 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1238 pr_err("Memory failure: buffer overflow when queuing memory failure at 0x%#lx\n",
1240 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1241 put_cpu_var(memory_failure_cpu);
1243 EXPORT_SYMBOL_GPL(memory_failure_queue);
1245 static void memory_failure_work_func(struct work_struct *work)
1247 struct memory_failure_cpu *mf_cpu;
1248 struct memory_failure_entry entry = { 0, };
1249 unsigned long proc_flags;
1252 mf_cpu = &__get_cpu_var(memory_failure_cpu);
1254 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1255 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1256 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1259 memory_failure(entry.pfn, entry.trapno, entry.flags);
1263 static int __init memory_failure_init(void)
1265 struct memory_failure_cpu *mf_cpu;
1268 for_each_possible_cpu(cpu) {
1269 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1270 spin_lock_init(&mf_cpu->lock);
1271 INIT_KFIFO(mf_cpu->fifo);
1272 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1277 core_initcall(memory_failure_init);
1280 * unpoison_memory - Unpoison a previously poisoned page
1281 * @pfn: Page number of the to be unpoisoned page
1283 * Software-unpoison a page that has been poisoned by
1284 * memory_failure() earlier.
1286 * This is only done on the software-level, so it only works
1287 * for linux injected failures, not real hardware failures
1289 * Returns 0 for success, otherwise -errno.
1291 int unpoison_memory(unsigned long pfn)
1296 unsigned int nr_pages;
1298 if (!pfn_valid(pfn))
1301 p = pfn_to_page(pfn);
1302 page = compound_head(p);
1304 if (!PageHWPoison(p)) {
1305 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
1309 nr_pages = 1 << compound_trans_order(page);
1311 if (!get_page_unless_zero(page)) {
1313 * Since HWPoisoned hugepage should have non-zero refcount,
1314 * race between memory failure and unpoison seems to happen.
1315 * In such case unpoison fails and memory failure runs
1318 if (PageHuge(page)) {
1319 pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1322 if (TestClearPageHWPoison(p))
1323 atomic_long_sub(nr_pages, &mce_bad_pages);
1324 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
1330 * This test is racy because PG_hwpoison is set outside of page lock.
1331 * That's acceptable because that won't trigger kernel panic. Instead,
1332 * the PG_hwpoison page will be caught and isolated on the entrance to
1333 * the free buddy page pool.
1335 if (TestClearPageHWPoison(page)) {
1336 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
1337 atomic_long_sub(nr_pages, &mce_bad_pages);
1340 clear_page_hwpoison_huge_page(page);
1350 EXPORT_SYMBOL(unpoison_memory);
1352 static struct page *new_page(struct page *p, unsigned long private, int **x)
1354 int nid = page_to_nid(p);
1356 return alloc_huge_page_node(page_hstate(compound_head(p)),
1359 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1363 * Safely get reference count of an arbitrary page.
1364 * Returns 0 for a free page, -EIO for a zero refcount page
1365 * that is not free, and 1 for any other page type.
1366 * For 1 the page is returned with increased page count, otherwise not.
1368 static int get_any_page(struct page *p, unsigned long pfn, int flags)
1372 if (flags & MF_COUNT_INCREASED)
1376 * The lock_memory_hotplug prevents a race with memory hotplug.
1377 * This is a big hammer, a better would be nicer.
1379 lock_memory_hotplug();
1382 * Isolate the page, so that it doesn't get reallocated if it
1385 set_migratetype_isolate(p);
1387 * When the target page is a free hugepage, just remove it
1388 * from free hugepage list.
1390 if (!get_page_unless_zero(compound_head(p))) {
1392 pr_info("%s: %#lx free huge page\n", __func__, pfn);
1393 ret = dequeue_hwpoisoned_huge_page(compound_head(p));
1394 } else if (is_free_buddy_page(p)) {
1395 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1396 /* Set hwpoison bit while page is still isolated */
1400 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1401 __func__, pfn, p->flags);
1405 /* Not a free page */
1408 unset_migratetype_isolate(p, MIGRATE_MOVABLE);
1409 unlock_memory_hotplug();
1413 static int soft_offline_huge_page(struct page *page, int flags)
1416 unsigned long pfn = page_to_pfn(page);
1417 struct page *hpage = compound_head(page);
1419 ret = get_any_page(page, pfn, flags);
1425 if (PageHWPoison(hpage)) {
1427 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1431 /* Keep page count to indicate a given hugepage is isolated. */
1432 ret = migrate_huge_page(hpage, new_page, MPOL_MF_MOVE_ALL, false,
1436 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1437 pfn, ret, page->flags);
1441 if (!PageHWPoison(hpage))
1442 atomic_long_add(1 << compound_trans_order(hpage),
1444 set_page_hwpoison_huge_page(hpage);
1445 dequeue_hwpoisoned_huge_page(hpage);
1446 /* keep elevated page count for bad page */
1451 * soft_offline_page - Soft offline a page.
1452 * @page: page to offline
1453 * @flags: flags. Same as memory_failure().
1455 * Returns 0 on success, otherwise negated errno.
1457 * Soft offline a page, by migration or invalidation,
1458 * without killing anything. This is for the case when
1459 * a page is not corrupted yet (so it's still valid to access),
1460 * but has had a number of corrected errors and is better taken
1463 * The actual policy on when to do that is maintained by
1466 * This should never impact any application or cause data loss,
1467 * however it might take some time.
1469 * This is not a 100% solution for all memory, but tries to be
1470 * ``good enough'' for the majority of memory.
1472 int soft_offline_page(struct page *page, int flags)
1475 unsigned long pfn = page_to_pfn(page);
1478 return soft_offline_huge_page(page, flags);
1480 ret = get_any_page(page, pfn, flags);
1487 * Page cache page we can handle?
1489 if (!PageLRU(page)) {
1494 shake_page(page, 1);
1499 ret = get_any_page(page, pfn, 0);
1505 if (!PageLRU(page)) {
1506 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1512 wait_on_page_writeback(page);
1515 * Synchronized using the page lock with memory_failure()
1517 if (PageHWPoison(page)) {
1520 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1525 * Try to invalidate first. This should work for
1526 * non dirty unmapped page cache pages.
1528 ret = invalidate_inode_page(page);
1531 * RED-PEN would be better to keep it isolated here, but we
1532 * would need to fix isolation locking first.
1537 pr_info("soft_offline: %#lx: invalidated\n", pfn);
1542 * Simple invalidation didn't work.
1543 * Try to migrate to a new page instead. migrate.c
1544 * handles a large number of cases for us.
1546 ret = isolate_lru_page(page);
1548 * Drop page reference which is came from get_any_page()
1549 * successful isolate_lru_page() already took another one.
1553 LIST_HEAD(pagelist);
1554 inc_zone_page_state(page, NR_ISOLATED_ANON +
1555 page_is_file_cache(page));
1556 list_add(&page->lru, &pagelist);
1557 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1558 false, MIGRATE_SYNC);
1560 putback_lru_pages(&pagelist);
1561 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1562 pfn, ret, page->flags);
1567 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1568 pfn, ret, page_count(page), page->flags);
1574 atomic_long_add(1, &mce_bad_pages);
1575 SetPageHWPoison(page);
1576 /* keep elevated page count for bad page */