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 num_poisoned_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_order(compound_head(page)) + PAGE_SHIFT;
211 if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
212 si.si_code = BUS_MCEERR_AR;
213 ret = force_sig_info(SIGBUS, &si, current);
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.
251 int nid = page_to_nid(p);
253 struct shrink_control shrink = {
254 .gfp_mask = GFP_KERNEL,
256 node_set(nid, shrink.nodes_to_scan);
258 nr = shrink_slab(&shrink, 1000, 1000);
259 if (page_count(p) == 1)
264 EXPORT_SYMBOL_GPL(shake_page);
267 * Kill all processes that have a poisoned page mapped and then isolate
271 * Find all processes having the page mapped and kill them.
272 * But we keep a page reference around so that the page is not
273 * actually freed yet.
274 * Then stash the page away
276 * There's no convenient way to get back to mapped processes
277 * from the VMAs. So do a brute-force search over all
280 * Remember that machine checks are not common (or rather
281 * if they are common you have other problems), so this shouldn't
282 * be a performance issue.
284 * Also there are some races possible while we get from the
285 * error detection to actually handle it.
290 struct task_struct *tsk;
296 * Failure handling: if we can't find or can't kill a process there's
297 * not much we can do. We just print a message and ignore otherwise.
301 * Schedule a process for later kill.
302 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
303 * TBD would GFP_NOIO be enough?
305 static void add_to_kill(struct task_struct *tsk, struct page *p,
306 struct vm_area_struct *vma,
307 struct list_head *to_kill,
308 struct to_kill **tkc)
316 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
319 "MCE: Out of memory while machine check handling\n");
323 tk->addr = page_address_in_vma(p, vma);
327 * In theory we don't have to kill when the page was
328 * munmaped. But it could be also a mremap. Since that's
329 * likely very rare kill anyways just out of paranoia, but use
330 * a SIGKILL because the error is not contained anymore.
332 if (tk->addr == -EFAULT) {
333 pr_info("MCE: Unable to find user space address %lx in %s\n",
334 page_to_pfn(p), tsk->comm);
337 get_task_struct(tsk);
339 list_add_tail(&tk->nd, to_kill);
343 * Kill the processes that have been collected earlier.
345 * Only do anything when DOIT is set, otherwise just free the list
346 * (this is used for clean pages which do not need killing)
347 * Also when FAIL is set do a force kill because something went
350 static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
351 int fail, struct page *page, unsigned long pfn,
354 struct to_kill *tk, *next;
356 list_for_each_entry_safe (tk, next, to_kill, nd) {
359 * In case something went wrong with munmapping
360 * make sure the process doesn't catch the
361 * signal and then access the memory. Just kill it.
363 if (fail || tk->addr_valid == 0) {
365 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
366 pfn, tk->tsk->comm, tk->tsk->pid);
367 force_sig(SIGKILL, tk->tsk);
371 * In theory the process could have mapped
372 * something else on the address in-between. We could
373 * check for that, but we need to tell the
376 else if (kill_proc(tk->tsk, tk->addr, trapno,
377 pfn, page, flags) < 0)
379 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
380 pfn, tk->tsk->comm, tk->tsk->pid);
382 put_task_struct(tk->tsk);
387 static int task_early_kill(struct task_struct *tsk, int force_early)
393 if (tsk->flags & PF_MCE_PROCESS)
394 return !!(tsk->flags & PF_MCE_EARLY);
395 return sysctl_memory_failure_early_kill;
399 * Collect processes when the error hit an anonymous page.
401 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
402 struct to_kill **tkc, int force_early)
404 struct vm_area_struct *vma;
405 struct task_struct *tsk;
409 av = page_lock_anon_vma_read(page);
410 if (av == NULL) /* Not actually mapped anymore */
413 pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
414 read_lock(&tasklist_lock);
415 for_each_process (tsk) {
416 struct anon_vma_chain *vmac;
418 if (!task_early_kill(tsk, force_early))
420 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
423 if (!page_mapped_in_vma(page, vma))
425 if (vma->vm_mm == tsk->mm)
426 add_to_kill(tsk, page, vma, to_kill, tkc);
429 read_unlock(&tasklist_lock);
430 page_unlock_anon_vma_read(av);
434 * Collect processes when the error hit a file mapped page.
436 static void collect_procs_file(struct page *page, struct list_head *to_kill,
437 struct to_kill **tkc, int force_early)
439 struct vm_area_struct *vma;
440 struct task_struct *tsk;
441 struct address_space *mapping = page->mapping;
443 mutex_lock(&mapping->i_mmap_mutex);
444 read_lock(&tasklist_lock);
445 for_each_process(tsk) {
446 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
448 if (!task_early_kill(tsk, force_early))
451 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
454 * Send early kill signal to tasks where a vma covers
455 * the page but the corrupted page is not necessarily
456 * mapped it in its pte.
457 * Assume applications who requested early kill want
458 * to be informed of all such data corruptions.
460 if (vma->vm_mm == tsk->mm)
461 add_to_kill(tsk, page, vma, to_kill, tkc);
464 read_unlock(&tasklist_lock);
465 mutex_unlock(&mapping->i_mmap_mutex);
469 * Collect the processes who have the corrupted page mapped to kill.
470 * This is done in two steps for locking reasons.
471 * First preallocate one tokill structure outside the spin locks,
472 * so that we can kill at least one process reasonably reliable.
474 static void collect_procs(struct page *page, struct list_head *tokill,
482 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
486 collect_procs_anon(page, tokill, &tk, force_early);
488 collect_procs_file(page, tokill, &tk, force_early);
493 * Error handlers for various types of pages.
497 IGNORED, /* Error: cannot be handled */
498 FAILED, /* Error: handling failed */
499 DELAYED, /* Will be handled later */
500 RECOVERED, /* Successfully recovered */
503 static const char *action_name[] = {
504 [IGNORED] = "Ignored",
506 [DELAYED] = "Delayed",
507 [RECOVERED] = "Recovered",
511 * XXX: It is possible that a page is isolated from LRU cache,
512 * and then kept in swap cache or failed to remove from page cache.
513 * The page count will stop it from being freed by unpoison.
514 * Stress tests should be aware of this memory leak problem.
516 static int delete_from_lru_cache(struct page *p)
518 if (!isolate_lru_page(p)) {
520 * Clear sensible page flags, so that the buddy system won't
521 * complain when the page is unpoison-and-freed.
524 ClearPageUnevictable(p);
526 * drop the page count elevated by isolate_lru_page()
528 page_cache_release(p);
535 * Error hit kernel page.
536 * Do nothing, try to be lucky and not touch this instead. For a few cases we
537 * could be more sophisticated.
539 static int me_kernel(struct page *p, unsigned long pfn)
545 * Page in unknown state. Do nothing.
547 static int me_unknown(struct page *p, unsigned long pfn)
549 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
554 * Clean (or cleaned) page cache page.
556 static int me_pagecache_clean(struct page *p, unsigned long pfn)
560 struct address_space *mapping;
562 delete_from_lru_cache(p);
565 * For anonymous pages we're done the only reference left
566 * should be the one m_f() holds.
572 * Now truncate the page in the page cache. This is really
573 * more like a "temporary hole punch"
574 * Don't do this for block devices when someone else
575 * has a reference, because it could be file system metadata
576 * and that's not safe to truncate.
578 mapping = page_mapping(p);
581 * Page has been teared down in the meanwhile
587 * Truncation is a bit tricky. Enable it per file system for now.
589 * Open: to take i_mutex or not for this? Right now we don't.
591 if (mapping->a_ops->error_remove_page) {
592 err = mapping->a_ops->error_remove_page(mapping, p);
594 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
596 } else if (page_has_private(p) &&
597 !try_to_release_page(p, GFP_NOIO)) {
598 pr_info("MCE %#lx: failed to release buffers\n", pfn);
604 * If the file system doesn't support it just invalidate
605 * This fails on dirty or anything with private pages
607 if (invalidate_inode_page(p))
610 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
617 * Dirty pagecache page
618 * Issues: when the error hit a hole page the error is not properly
621 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
623 struct address_space *mapping = page_mapping(p);
626 /* TBD: print more information about the file. */
629 * IO error will be reported by write(), fsync(), etc.
630 * who check the mapping.
631 * This way the application knows that something went
632 * wrong with its dirty file data.
634 * There's one open issue:
636 * The EIO will be only reported on the next IO
637 * operation and then cleared through the IO map.
638 * Normally Linux has two mechanisms to pass IO error
639 * first through the AS_EIO flag in the address space
640 * and then through the PageError flag in the page.
641 * Since we drop pages on memory failure handling the
642 * only mechanism open to use is through AS_AIO.
644 * This has the disadvantage that it gets cleared on
645 * the first operation that returns an error, while
646 * the PageError bit is more sticky and only cleared
647 * when the page is reread or dropped. If an
648 * application assumes it will always get error on
649 * fsync, but does other operations on the fd before
650 * and the page is dropped between then the error
651 * will not be properly reported.
653 * This can already happen even without hwpoisoned
654 * pages: first on metadata IO errors (which only
655 * report through AS_EIO) or when the page is dropped
658 * So right now we assume that the application DTRT on
659 * the first EIO, but we're not worse than other parts
662 mapping_set_error(mapping, EIO);
665 return me_pagecache_clean(p, pfn);
669 * Clean and dirty swap cache.
671 * Dirty swap cache page is tricky to handle. The page could live both in page
672 * cache and swap cache(ie. page is freshly swapped in). So it could be
673 * referenced concurrently by 2 types of PTEs:
674 * normal PTEs and swap PTEs. We try to handle them consistently by calling
675 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
677 * - clear dirty bit to prevent IO
679 * - but keep in the swap cache, so that when we return to it on
680 * a later page fault, we know the application is accessing
681 * corrupted data and shall be killed (we installed simple
682 * interception code in do_swap_page to catch it).
684 * Clean swap cache pages can be directly isolated. A later page fault will
685 * bring in the known good data from disk.
687 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
690 /* Trigger EIO in shmem: */
691 ClearPageUptodate(p);
693 if (!delete_from_lru_cache(p))
699 static int me_swapcache_clean(struct page *p, unsigned long pfn)
701 delete_from_swap_cache(p);
703 if (!delete_from_lru_cache(p))
710 * Huge pages. Needs work.
712 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
713 * To narrow down kill region to one page, we need to break up pmd.
715 static int me_huge_page(struct page *p, unsigned long pfn)
718 struct page *hpage = compound_head(p);
720 * We can safely recover from error on free or reserved (i.e.
721 * not in-use) hugepage by dequeuing it from freelist.
722 * To check whether a hugepage is in-use or not, we can't use
723 * page->lru because it can be used in other hugepage operations,
724 * such as __unmap_hugepage_range() and gather_surplus_pages().
725 * So instead we use page_mapping() and PageAnon().
726 * We assume that this function is called with page lock held,
727 * so there is no race between isolation and mapping/unmapping.
729 if (!(page_mapping(hpage) || PageAnon(hpage))) {
730 res = dequeue_hwpoisoned_huge_page(hpage);
738 * Various page states we can handle.
740 * A page state is defined by its current page->flags bits.
741 * The table matches them in order and calls the right handler.
743 * This is quite tricky because we can access page at any time
744 * in its live cycle, so all accesses have to be extremely careful.
746 * This is not complete. More states could be added.
747 * For any missing state don't attempt recovery.
750 #define dirty (1UL << PG_dirty)
751 #define sc (1UL << PG_swapcache)
752 #define unevict (1UL << PG_unevictable)
753 #define mlock (1UL << PG_mlocked)
754 #define writeback (1UL << PG_writeback)
755 #define lru (1UL << PG_lru)
756 #define swapbacked (1UL << PG_swapbacked)
757 #define head (1UL << PG_head)
758 #define tail (1UL << PG_tail)
759 #define compound (1UL << PG_compound)
760 #define slab (1UL << PG_slab)
761 #define reserved (1UL << PG_reserved)
763 static struct page_state {
767 int (*action)(struct page *p, unsigned long pfn);
769 { reserved, reserved, "reserved kernel", me_kernel },
771 * free pages are specially detected outside this table:
772 * PG_buddy pages only make a small fraction of all free pages.
776 * Could in theory check if slab page is free or if we can drop
777 * currently unused objects without touching them. But just
778 * treat it as standard kernel for now.
780 { slab, slab, "kernel slab", me_kernel },
782 #ifdef CONFIG_PAGEFLAGS_EXTENDED
783 { head, head, "huge", me_huge_page },
784 { tail, tail, "huge", me_huge_page },
786 { compound, compound, "huge", me_huge_page },
789 { sc|dirty, sc|dirty, "dirty swapcache", me_swapcache_dirty },
790 { sc|dirty, sc, "clean swapcache", me_swapcache_clean },
792 { mlock|dirty, mlock|dirty, "dirty mlocked LRU", me_pagecache_dirty },
793 { mlock|dirty, mlock, "clean mlocked LRU", me_pagecache_clean },
795 { unevict|dirty, unevict|dirty, "dirty unevictable LRU", me_pagecache_dirty },
796 { unevict|dirty, unevict, "clean unevictable LRU", me_pagecache_clean },
798 { lru|dirty, lru|dirty, "dirty LRU", me_pagecache_dirty },
799 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
802 * Catchall entry: must be at end.
804 { 0, 0, "unknown page state", me_unknown },
821 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
822 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
824 static void action_result(unsigned long pfn, char *msg, int result)
826 pr_err("MCE %#lx: %s page recovery: %s\n",
827 pfn, msg, action_name[result]);
830 static int page_action(struct page_state *ps, struct page *p,
836 result = ps->action(p, pfn);
837 action_result(pfn, ps->msg, result);
839 count = page_count(p) - 1;
840 if (ps->action == me_swapcache_dirty && result == DELAYED)
844 "MCE %#lx: %s page still referenced by %d users\n",
845 pfn, ps->msg, count);
849 /* Could do more checks here if page looks ok */
851 * Could adjust zone counters here to correct for the missing page.
854 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
858 * Do all that is necessary to remove user space mappings. Unmap
859 * the pages and send SIGBUS to the processes if the data was dirty.
861 static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
862 int trapno, int flags, struct page **hpagep)
864 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
865 struct address_space *mapping;
868 int kill = 1, forcekill;
869 struct page *hpage = *hpagep;
872 if (PageReserved(p) || PageSlab(p))
876 * This check implies we don't kill processes if their pages
877 * are in the swap cache early. Those are always late kills.
879 if (!page_mapped(hpage))
885 if (PageSwapCache(p)) {
887 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
888 ttu |= TTU_IGNORE_HWPOISON;
892 * Propagate the dirty bit from PTEs to struct page first, because we
893 * need this to decide if we should kill or just drop the page.
894 * XXX: the dirty test could be racy: set_page_dirty() may not always
895 * be called inside page lock (it's recommended but not enforced).
897 mapping = page_mapping(hpage);
898 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
899 mapping_cap_writeback_dirty(mapping)) {
900 if (page_mkclean(hpage)) {
904 ttu |= TTU_IGNORE_HWPOISON;
906 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
912 * ppage: poisoned page
913 * if p is regular page(4k page)
914 * ppage == real poisoned page;
915 * else p is hugetlb or THP, ppage == head page.
919 if (PageTransHuge(hpage)) {
921 * Verify that this isn't a hugetlbfs head page, the check for
922 * PageAnon is just for avoid tripping a split_huge_page
923 * internal debug check, as split_huge_page refuses to deal with
924 * anything that isn't an anon page. PageAnon can't go away fro
925 * under us because we hold a refcount on the hpage, without a
926 * refcount on the hpage. split_huge_page can't be safely called
927 * in the first place, having a refcount on the tail isn't
928 * enough * to be safe.
930 if (!PageHuge(hpage) && PageAnon(hpage)) {
931 if (unlikely(split_huge_page(hpage))) {
933 * FIXME: if splitting THP is failed, it is
934 * better to stop the following operation rather
935 * than causing panic by unmapping. System might
936 * survive if the page is freed later.
939 "MCE %#lx: failed to split THP\n", pfn);
941 BUG_ON(!PageHWPoison(p));
945 * We pinned the head page for hwpoison handling,
946 * now we split the thp and we are interested in
947 * the hwpoisoned raw page, so move the refcount
948 * to it. Similarly, page lock is shifted.
951 if (!(flags & MF_COUNT_INCREASED)) {
959 /* THP is split, so ppage should be the real poisoned page. */
965 * First collect all the processes that have the page
966 * mapped in dirty form. This has to be done before try_to_unmap,
967 * because ttu takes the rmap data structures down.
969 * Error handling: We ignore errors here because
970 * there's nothing that can be done.
973 collect_procs(ppage, &tokill, flags & MF_ACTION_REQUIRED);
975 ret = try_to_unmap(ppage, ttu);
976 if (ret != SWAP_SUCCESS)
977 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
978 pfn, page_mapcount(ppage));
981 * Now that the dirty bit has been propagated to the
982 * struct page and all unmaps done we can decide if
983 * killing is needed or not. Only kill when the page
984 * was dirty or the process is not restartable,
985 * otherwise the tokill list is merely
986 * freed. When there was a problem unmapping earlier
987 * use a more force-full uncatchable kill to prevent
988 * any accesses to the poisoned memory.
990 forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL);
991 kill_procs(&tokill, forcekill, trapno,
992 ret != SWAP_SUCCESS, p, pfn, flags);
997 static void set_page_hwpoison_huge_page(struct page *hpage)
1000 int nr_pages = 1 << compound_order(hpage);
1001 for (i = 0; i < nr_pages; i++)
1002 SetPageHWPoison(hpage + i);
1005 static void clear_page_hwpoison_huge_page(struct page *hpage)
1008 int nr_pages = 1 << compound_order(hpage);
1009 for (i = 0; i < nr_pages; i++)
1010 ClearPageHWPoison(hpage + i);
1014 * memory_failure - Handle memory failure of a page.
1015 * @pfn: Page Number of the corrupted page
1016 * @trapno: Trap number reported in the signal to user space.
1017 * @flags: fine tune action taken
1019 * This function is called by the low level machine check code
1020 * of an architecture when it detects hardware memory corruption
1021 * of a page. It tries its best to recover, which includes
1022 * dropping pages, killing processes etc.
1024 * The function is primarily of use for corruptions that
1025 * happen outside the current execution context (e.g. when
1026 * detected by a background scrubber)
1028 * Must run in process context (e.g. a work queue) with interrupts
1029 * enabled and no spinlocks hold.
1031 int memory_failure(unsigned long pfn, int trapno, int flags)
1033 struct page_state *ps;
1037 unsigned int nr_pages;
1038 unsigned long page_flags;
1040 if (!sysctl_memory_failure_recovery)
1041 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1043 if (!pfn_valid(pfn)) {
1045 "MCE %#lx: memory outside kernel control\n",
1050 p = pfn_to_page(pfn);
1051 hpage = compound_head(p);
1052 if (TestSetPageHWPoison(p)) {
1053 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
1058 * Currently errors on hugetlbfs pages are measured in hugepage units,
1059 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1060 * transparent hugepages, they are supposed to be split and error
1061 * measurement is done in normal page units. So nr_pages should be one
1065 nr_pages = 1 << compound_order(hpage);
1066 else /* normal page or thp */
1068 atomic_long_add(nr_pages, &num_poisoned_pages);
1071 * We need/can do nothing about count=0 pages.
1072 * 1) it's a free page, and therefore in safe hand:
1073 * prep_new_page() will be the gate keeper.
1074 * 2) it's a free hugepage, which is also safe:
1075 * an affected hugepage will be dequeued from hugepage freelist,
1076 * so there's no concern about reusing it ever after.
1077 * 3) it's part of a non-compound high order page.
1078 * Implies some kernel user: cannot stop them from
1079 * R/W the page; let's pray that the page has been
1080 * used and will be freed some time later.
1081 * In fact it's dangerous to directly bump up page count from 0,
1082 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1084 if (!(flags & MF_COUNT_INCREASED) &&
1085 !get_page_unless_zero(hpage)) {
1086 if (is_free_buddy_page(p)) {
1087 action_result(pfn, "free buddy", DELAYED);
1089 } else if (PageHuge(hpage)) {
1091 * Check "filter hit" and "race with other subpage."
1094 if (PageHWPoison(hpage)) {
1095 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1096 || (p != hpage && TestSetPageHWPoison(hpage))) {
1097 atomic_long_sub(nr_pages, &num_poisoned_pages);
1102 set_page_hwpoison_huge_page(hpage);
1103 res = dequeue_hwpoisoned_huge_page(hpage);
1104 action_result(pfn, "free huge",
1105 res ? IGNORED : DELAYED);
1109 action_result(pfn, "high order kernel", IGNORED);
1115 * We ignore non-LRU pages for good reasons.
1116 * - PG_locked is only well defined for LRU pages and a few others
1117 * - to avoid races with __set_page_locked()
1118 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1119 * The check (unnecessarily) ignores LRU pages being isolated and
1120 * walked by the page reclaim code, however that's not a big loss.
1122 if (!PageHuge(p) && !PageTransTail(p)) {
1127 * shake_page could have turned it free.
1129 if (is_free_buddy_page(p)) {
1130 if (flags & MF_COUNT_INCREASED)
1131 action_result(pfn, "free buddy", DELAYED);
1133 action_result(pfn, "free buddy, 2nd try", DELAYED);
1136 action_result(pfn, "non LRU", IGNORED);
1143 * Lock the page and wait for writeback to finish.
1144 * It's very difficult to mess with pages currently under IO
1145 * and in many cases impossible, so we just avoid it here.
1150 * We use page flags to determine what action should be taken, but
1151 * the flags can be modified by the error containment action. One
1152 * example is an mlocked page, where PG_mlocked is cleared by
1153 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1154 * correctly, we save a copy of the page flags at this time.
1156 page_flags = p->flags;
1159 * unpoison always clear PG_hwpoison inside page lock
1161 if (!PageHWPoison(p)) {
1162 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1163 atomic_long_sub(nr_pages, &num_poisoned_pages);
1168 if (hwpoison_filter(p)) {
1169 if (TestClearPageHWPoison(p))
1170 atomic_long_sub(nr_pages, &num_poisoned_pages);
1177 * For error on the tail page, we should set PG_hwpoison
1178 * on the head page to show that the hugepage is hwpoisoned
1180 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1181 action_result(pfn, "hugepage already hardware poisoned",
1188 * Set PG_hwpoison on all pages in an error hugepage,
1189 * because containment is done in hugepage unit for now.
1190 * Since we have done TestSetPageHWPoison() for the head page with
1191 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1194 set_page_hwpoison_huge_page(hpage);
1196 wait_on_page_writeback(p);
1199 * Now take care of user space mappings.
1200 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1202 * When the raw error page is thp tail page, hpage points to the raw
1203 * page after thp split.
1205 if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)
1207 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1213 * Torn down by someone else?
1215 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1216 action_result(pfn, "already truncated LRU", IGNORED);
1223 * The first check uses the current page flags which may not have any
1224 * relevant information. The second check with the saved page flagss is
1225 * carried out only if the first check can't determine the page status.
1227 for (ps = error_states;; ps++)
1228 if ((p->flags & ps->mask) == ps->res)
1231 page_flags |= (p->flags & (1UL << PG_dirty));
1234 for (ps = error_states;; ps++)
1235 if ((page_flags & ps->mask) == ps->res)
1237 res = page_action(ps, p, pfn);
1242 EXPORT_SYMBOL_GPL(memory_failure);
1244 #define MEMORY_FAILURE_FIFO_ORDER 4
1245 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1247 struct memory_failure_entry {
1253 struct memory_failure_cpu {
1254 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1255 MEMORY_FAILURE_FIFO_SIZE);
1257 struct work_struct work;
1260 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1263 * memory_failure_queue - Schedule handling memory failure of a page.
1264 * @pfn: Page Number of the corrupted page
1265 * @trapno: Trap number reported in the signal to user space.
1266 * @flags: Flags for memory failure handling
1268 * This function is called by the low level hardware error handler
1269 * when it detects hardware memory corruption of a page. It schedules
1270 * the recovering of error page, including dropping pages, killing
1273 * The function is primarily of use for corruptions that
1274 * happen outside the current execution context (e.g. when
1275 * detected by a background scrubber)
1277 * Can run in IRQ context.
1279 void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1281 struct memory_failure_cpu *mf_cpu;
1282 unsigned long proc_flags;
1283 struct memory_failure_entry entry = {
1289 mf_cpu = &get_cpu_var(memory_failure_cpu);
1290 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1291 if (kfifo_put(&mf_cpu->fifo, entry))
1292 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1294 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1296 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1297 put_cpu_var(memory_failure_cpu);
1299 EXPORT_SYMBOL_GPL(memory_failure_queue);
1301 static void memory_failure_work_func(struct work_struct *work)
1303 struct memory_failure_cpu *mf_cpu;
1304 struct memory_failure_entry entry = { 0, };
1305 unsigned long proc_flags;
1308 mf_cpu = &__get_cpu_var(memory_failure_cpu);
1310 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1311 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1312 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1315 if (entry.flags & MF_SOFT_OFFLINE)
1316 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1318 memory_failure(entry.pfn, entry.trapno, entry.flags);
1322 static int __init memory_failure_init(void)
1324 struct memory_failure_cpu *mf_cpu;
1327 for_each_possible_cpu(cpu) {
1328 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1329 spin_lock_init(&mf_cpu->lock);
1330 INIT_KFIFO(mf_cpu->fifo);
1331 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1336 core_initcall(memory_failure_init);
1339 * unpoison_memory - Unpoison a previously poisoned page
1340 * @pfn: Page number of the to be unpoisoned page
1342 * Software-unpoison a page that has been poisoned by
1343 * memory_failure() earlier.
1345 * This is only done on the software-level, so it only works
1346 * for linux injected failures, not real hardware failures
1348 * Returns 0 for success, otherwise -errno.
1350 int unpoison_memory(unsigned long pfn)
1355 unsigned int nr_pages;
1357 if (!pfn_valid(pfn))
1360 p = pfn_to_page(pfn);
1361 page = compound_head(p);
1363 if (!PageHWPoison(p)) {
1364 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
1369 * unpoison_memory() can encounter thp only when the thp is being
1370 * worked by memory_failure() and the page lock is not held yet.
1371 * In such case, we yield to memory_failure() and make unpoison fail.
1373 if (!PageHuge(page) && PageTransHuge(page)) {
1374 pr_info("MCE: Memory failure is now running on %#lx\n", pfn);
1378 nr_pages = 1 << compound_order(page);
1380 if (!get_page_unless_zero(page)) {
1382 * Since HWPoisoned hugepage should have non-zero refcount,
1383 * race between memory failure and unpoison seems to happen.
1384 * In such case unpoison fails and memory failure runs
1387 if (PageHuge(page)) {
1388 pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1391 if (TestClearPageHWPoison(p))
1392 atomic_long_dec(&num_poisoned_pages);
1393 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
1399 * This test is racy because PG_hwpoison is set outside of page lock.
1400 * That's acceptable because that won't trigger kernel panic. Instead,
1401 * the PG_hwpoison page will be caught and isolated on the entrance to
1402 * the free buddy page pool.
1404 if (TestClearPageHWPoison(page)) {
1405 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
1406 atomic_long_sub(nr_pages, &num_poisoned_pages);
1409 clear_page_hwpoison_huge_page(page);
1414 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
1419 EXPORT_SYMBOL(unpoison_memory);
1421 static struct page *new_page(struct page *p, unsigned long private, int **x)
1423 int nid = page_to_nid(p);
1425 return alloc_huge_page_node(page_hstate(compound_head(p)),
1428 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1432 * Safely get reference count of an arbitrary page.
1433 * Returns 0 for a free page, -EIO for a zero refcount page
1434 * that is not free, and 1 for any other page type.
1435 * For 1 the page is returned with increased page count, otherwise not.
1437 static int __get_any_page(struct page *p, unsigned long pfn, int flags)
1441 if (flags & MF_COUNT_INCREASED)
1445 * When the target page is a free hugepage, just remove it
1446 * from free hugepage list.
1448 if (!get_page_unless_zero(compound_head(p))) {
1450 pr_info("%s: %#lx free huge page\n", __func__, pfn);
1452 } else if (is_free_buddy_page(p)) {
1453 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1456 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1457 __func__, pfn, p->flags);
1461 /* Not a free page */
1467 static int get_any_page(struct page *page, unsigned long pfn, int flags)
1469 int ret = __get_any_page(page, pfn, flags);
1471 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1476 shake_page(page, 1);
1481 ret = __get_any_page(page, pfn, 0);
1482 if (!PageLRU(page)) {
1483 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1491 static int soft_offline_huge_page(struct page *page, int flags)
1494 unsigned long pfn = page_to_pfn(page);
1495 struct page *hpage = compound_head(page);
1496 LIST_HEAD(pagelist);
1499 * This double-check of PageHWPoison is to avoid the race with
1500 * memory_failure(). See also comment in __soft_offline_page().
1503 if (PageHWPoison(hpage)) {
1506 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1511 /* Keep page count to indicate a given hugepage is isolated. */
1512 list_move(&hpage->lru, &pagelist);
1513 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1514 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1516 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1517 pfn, ret, page->flags);
1519 * We know that soft_offline_huge_page() tries to migrate
1520 * only one hugepage pointed to by hpage, so we need not
1521 * run through the pagelist here.
1523 putback_active_hugepage(hpage);
1527 /* overcommit hugetlb page will be freed to buddy */
1528 if (PageHuge(page)) {
1529 set_page_hwpoison_huge_page(hpage);
1530 dequeue_hwpoisoned_huge_page(hpage);
1531 atomic_long_add(1 << compound_order(hpage),
1532 &num_poisoned_pages);
1534 SetPageHWPoison(page);
1535 atomic_long_inc(&num_poisoned_pages);
1541 static int __soft_offline_page(struct page *page, int flags)
1544 unsigned long pfn = page_to_pfn(page);
1547 * Check PageHWPoison again inside page lock because PageHWPoison
1548 * is set by memory_failure() outside page lock. Note that
1549 * memory_failure() also double-checks PageHWPoison inside page lock,
1550 * so there's no race between soft_offline_page() and memory_failure().
1553 wait_on_page_writeback(page);
1554 if (PageHWPoison(page)) {
1557 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1561 * Try to invalidate first. This should work for
1562 * non dirty unmapped page cache pages.
1564 ret = invalidate_inode_page(page);
1567 * RED-PEN would be better to keep it isolated here, but we
1568 * would need to fix isolation locking first.
1572 pr_info("soft_offline: %#lx: invalidated\n", pfn);
1573 SetPageHWPoison(page);
1574 atomic_long_inc(&num_poisoned_pages);
1579 * Simple invalidation didn't work.
1580 * Try to migrate to a new page instead. migrate.c
1581 * handles a large number of cases for us.
1583 ret = isolate_lru_page(page);
1585 * Drop page reference which is came from get_any_page()
1586 * successful isolate_lru_page() already took another one.
1590 LIST_HEAD(pagelist);
1591 inc_zone_page_state(page, NR_ISOLATED_ANON +
1592 page_is_file_cache(page));
1593 list_add(&page->lru, &pagelist);
1594 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1595 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1597 if (!list_empty(&pagelist)) {
1598 list_del(&page->lru);
1599 dec_zone_page_state(page, NR_ISOLATED_ANON +
1600 page_is_file_cache(page));
1601 putback_lru_page(page);
1604 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1605 pfn, ret, page->flags);
1610 * After page migration succeeds, the source page can
1611 * be trapped in pagevec and actual freeing is delayed.
1612 * Freeing code works differently based on PG_hwpoison,
1613 * so there's a race. We need to make sure that the
1614 * source page should be freed back to buddy before
1615 * setting PG_hwpoison.
1617 if (!is_free_buddy_page(page))
1618 lru_add_drain_all();
1619 if (!is_free_buddy_page(page))
1621 SetPageHWPoison(page);
1622 if (!is_free_buddy_page(page))
1623 pr_info("soft offline: %#lx: page leaked\n",
1625 atomic_long_inc(&num_poisoned_pages);
1628 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1629 pfn, ret, page_count(page), page->flags);
1635 * soft_offline_page - Soft offline a page.
1636 * @page: page to offline
1637 * @flags: flags. Same as memory_failure().
1639 * Returns 0 on success, otherwise negated errno.
1641 * Soft offline a page, by migration or invalidation,
1642 * without killing anything. This is for the case when
1643 * a page is not corrupted yet (so it's still valid to access),
1644 * but has had a number of corrected errors and is better taken
1647 * The actual policy on when to do that is maintained by
1650 * This should never impact any application or cause data loss,
1651 * however it might take some time.
1653 * This is not a 100% solution for all memory, but tries to be
1654 * ``good enough'' for the majority of memory.
1656 int soft_offline_page(struct page *page, int flags)
1659 unsigned long pfn = page_to_pfn(page);
1660 struct page *hpage = compound_head(page);
1662 if (PageHWPoison(page)) {
1663 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1666 if (!PageHuge(page) && PageTransHuge(hpage)) {
1667 if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1668 pr_info("soft offline: %#lx: failed to split THP\n",
1675 * The lock_memory_hotplug prevents a race with memory hotplug.
1676 * This is a big hammer, a better would be nicer.
1678 lock_memory_hotplug();
1681 * Isolate the page, so that it doesn't get reallocated if it
1682 * was free. This flag should be kept set until the source page
1683 * is freed and PG_hwpoison on it is set.
1685 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
1686 set_migratetype_isolate(page, true);
1688 ret = get_any_page(page, pfn, flags);
1689 unlock_memory_hotplug();
1690 if (ret > 0) { /* for in-use pages */
1692 ret = soft_offline_huge_page(page, flags);
1694 ret = __soft_offline_page(page, flags);
1695 } else if (ret == 0) { /* for free pages */
1696 if (PageHuge(page)) {
1697 set_page_hwpoison_huge_page(hpage);
1698 dequeue_hwpoisoned_huge_page(hpage);
1699 atomic_long_add(1 << compound_order(hpage),
1700 &num_poisoned_pages);
1702 SetPageHWPoison(page);
1703 atomic_long_inc(&num_poisoned_pages);
1706 unset_migratetype_isolate(page, MIGRATE_MOVABLE);