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 == 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.
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)
391 if (tsk->flags & PF_MCE_PROCESS)
392 return !!(tsk->flags & PF_MCE_EARLY);
393 return sysctl_memory_failure_early_kill;
397 * Collect processes when the error hit an anonymous page.
399 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
400 struct to_kill **tkc)
402 struct vm_area_struct *vma;
403 struct task_struct *tsk;
407 av = page_lock_anon_vma_read(page);
408 if (av == NULL) /* Not actually mapped anymore */
411 pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
412 read_lock(&tasklist_lock);
413 for_each_process (tsk) {
414 struct anon_vma_chain *vmac;
416 if (!task_early_kill(tsk))
418 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
421 if (!page_mapped_in_vma(page, vma))
423 if (vma->vm_mm == tsk->mm)
424 add_to_kill(tsk, page, vma, to_kill, tkc);
427 read_unlock(&tasklist_lock);
428 page_unlock_anon_vma_read(av);
432 * Collect processes when the error hit a file mapped page.
434 static void collect_procs_file(struct page *page, struct list_head *to_kill,
435 struct to_kill **tkc)
437 struct vm_area_struct *vma;
438 struct task_struct *tsk;
439 struct address_space *mapping = page->mapping;
441 mutex_lock(&mapping->i_mmap_mutex);
442 read_lock(&tasklist_lock);
443 for_each_process(tsk) {
444 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
446 if (!task_early_kill(tsk))
449 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
452 * Send early kill signal to tasks where a vma covers
453 * the page but the corrupted page is not necessarily
454 * mapped it in its pte.
455 * Assume applications who requested early kill want
456 * to be informed of all such data corruptions.
458 if (vma->vm_mm == tsk->mm)
459 add_to_kill(tsk, page, vma, to_kill, tkc);
462 read_unlock(&tasklist_lock);
463 mutex_unlock(&mapping->i_mmap_mutex);
467 * Collect the processes who have the corrupted page mapped to kill.
468 * This is done in two steps for locking reasons.
469 * First preallocate one tokill structure outside the spin locks,
470 * so that we can kill at least one process reasonably reliable.
472 static void collect_procs(struct page *page, struct list_head *tokill)
479 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
483 collect_procs_anon(page, tokill, &tk);
485 collect_procs_file(page, tokill, &tk);
490 * Error handlers for various types of pages.
494 IGNORED, /* Error: cannot be handled */
495 FAILED, /* Error: handling failed */
496 DELAYED, /* Will be handled later */
497 RECOVERED, /* Successfully recovered */
500 static const char *action_name[] = {
501 [IGNORED] = "Ignored",
503 [DELAYED] = "Delayed",
504 [RECOVERED] = "Recovered",
508 * XXX: It is possible that a page is isolated from LRU cache,
509 * and then kept in swap cache or failed to remove from page cache.
510 * The page count will stop it from being freed by unpoison.
511 * Stress tests should be aware of this memory leak problem.
513 static int delete_from_lru_cache(struct page *p)
515 if (!isolate_lru_page(p)) {
517 * Clear sensible page flags, so that the buddy system won't
518 * complain when the page is unpoison-and-freed.
521 ClearPageUnevictable(p);
523 * drop the page count elevated by isolate_lru_page()
525 page_cache_release(p);
532 * Error hit kernel page.
533 * Do nothing, try to be lucky and not touch this instead. For a few cases we
534 * could be more sophisticated.
536 static int me_kernel(struct page *p, unsigned long pfn)
542 * Page in unknown state. Do nothing.
544 static int me_unknown(struct page *p, unsigned long pfn)
546 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
551 * Clean (or cleaned) page cache page.
553 static int me_pagecache_clean(struct page *p, unsigned long pfn)
557 struct address_space *mapping;
559 delete_from_lru_cache(p);
562 * For anonymous pages we're done the only reference left
563 * should be the one m_f() holds.
569 * Now truncate the page in the page cache. This is really
570 * more like a "temporary hole punch"
571 * Don't do this for block devices when someone else
572 * has a reference, because it could be file system metadata
573 * and that's not safe to truncate.
575 mapping = page_mapping(p);
578 * Page has been teared down in the meanwhile
584 * Truncation is a bit tricky. Enable it per file system for now.
586 * Open: to take i_mutex or not for this? Right now we don't.
588 if (mapping->a_ops->error_remove_page) {
589 err = mapping->a_ops->error_remove_page(mapping, p);
591 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
593 } else if (page_has_private(p) &&
594 !try_to_release_page(p, GFP_NOIO)) {
595 pr_info("MCE %#lx: failed to release buffers\n", pfn);
601 * If the file system doesn't support it just invalidate
602 * This fails on dirty or anything with private pages
604 if (invalidate_inode_page(p))
607 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
614 * Dirty cache page page
615 * Issues: when the error hit a hole page the error is not properly
618 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
620 struct address_space *mapping = page_mapping(p);
623 /* TBD: print more information about the file. */
626 * IO error will be reported by write(), fsync(), etc.
627 * who check the mapping.
628 * This way the application knows that something went
629 * wrong with its dirty file data.
631 * There's one open issue:
633 * The EIO will be only reported on the next IO
634 * operation and then cleared through the IO map.
635 * Normally Linux has two mechanisms to pass IO error
636 * first through the AS_EIO flag in the address space
637 * and then through the PageError flag in the page.
638 * Since we drop pages on memory failure handling the
639 * only mechanism open to use is through AS_AIO.
641 * This has the disadvantage that it gets cleared on
642 * the first operation that returns an error, while
643 * the PageError bit is more sticky and only cleared
644 * when the page is reread or dropped. If an
645 * application assumes it will always get error on
646 * fsync, but does other operations on the fd before
647 * and the page is dropped between then the error
648 * will not be properly reported.
650 * This can already happen even without hwpoisoned
651 * pages: first on metadata IO errors (which only
652 * report through AS_EIO) or when the page is dropped
655 * So right now we assume that the application DTRT on
656 * the first EIO, but we're not worse than other parts
659 mapping_set_error(mapping, EIO);
662 return me_pagecache_clean(p, pfn);
666 * Clean and dirty swap cache.
668 * Dirty swap cache page is tricky to handle. The page could live both in page
669 * cache and swap cache(ie. page is freshly swapped in). So it could be
670 * referenced concurrently by 2 types of PTEs:
671 * normal PTEs and swap PTEs. We try to handle them consistently by calling
672 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
674 * - clear dirty bit to prevent IO
676 * - but keep in the swap cache, so that when we return to it on
677 * a later page fault, we know the application is accessing
678 * corrupted data and shall be killed (we installed simple
679 * interception code in do_swap_page to catch it).
681 * Clean swap cache pages can be directly isolated. A later page fault will
682 * bring in the known good data from disk.
684 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
687 /* Trigger EIO in shmem: */
688 ClearPageUptodate(p);
690 if (!delete_from_lru_cache(p))
696 static int me_swapcache_clean(struct page *p, unsigned long pfn)
698 delete_from_swap_cache(p);
700 if (!delete_from_lru_cache(p))
707 * Huge pages. Needs work.
709 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
710 * To narrow down kill region to one page, we need to break up pmd.
712 static int me_huge_page(struct page *p, unsigned long pfn)
715 struct page *hpage = compound_head(p);
717 * We can safely recover from error on free or reserved (i.e.
718 * not in-use) hugepage by dequeuing it from freelist.
719 * To check whether a hugepage is in-use or not, we can't use
720 * page->lru because it can be used in other hugepage operations,
721 * such as __unmap_hugepage_range() and gather_surplus_pages().
722 * So instead we use page_mapping() and PageAnon().
723 * We assume that this function is called with page lock held,
724 * so there is no race between isolation and mapping/unmapping.
726 if (!(page_mapping(hpage) || PageAnon(hpage))) {
727 res = dequeue_hwpoisoned_huge_page(hpage);
735 * Various page states we can handle.
737 * A page state is defined by its current page->flags bits.
738 * The table matches them in order and calls the right handler.
740 * This is quite tricky because we can access page at any time
741 * in its live cycle, so all accesses have to be extremely careful.
743 * This is not complete. More states could be added.
744 * For any missing state don't attempt recovery.
747 #define dirty (1UL << PG_dirty)
748 #define sc (1UL << PG_swapcache)
749 #define unevict (1UL << PG_unevictable)
750 #define mlock (1UL << PG_mlocked)
751 #define writeback (1UL << PG_writeback)
752 #define lru (1UL << PG_lru)
753 #define swapbacked (1UL << PG_swapbacked)
754 #define head (1UL << PG_head)
755 #define tail (1UL << PG_tail)
756 #define compound (1UL << PG_compound)
757 #define slab (1UL << PG_slab)
758 #define reserved (1UL << PG_reserved)
760 static struct page_state {
764 int (*action)(struct page *p, unsigned long pfn);
766 { reserved, reserved, "reserved kernel", me_kernel },
768 * free pages are specially detected outside this table:
769 * PG_buddy pages only make a small fraction of all free pages.
773 * Could in theory check if slab page is free or if we can drop
774 * currently unused objects without touching them. But just
775 * treat it as standard kernel for now.
777 { slab, slab, "kernel slab", me_kernel },
779 #ifdef CONFIG_PAGEFLAGS_EXTENDED
780 { head, head, "huge", me_huge_page },
781 { tail, tail, "huge", me_huge_page },
783 { compound, compound, "huge", me_huge_page },
786 { sc|dirty, sc|dirty, "dirty swapcache", me_swapcache_dirty },
787 { sc|dirty, sc, "clean swapcache", me_swapcache_clean },
789 { mlock|dirty, mlock|dirty, "dirty mlocked LRU", me_pagecache_dirty },
790 { mlock|dirty, mlock, "clean mlocked LRU", me_pagecache_clean },
792 { unevict|dirty, unevict|dirty, "dirty unevictable LRU", me_pagecache_dirty },
793 { unevict|dirty, unevict, "clean unevictable LRU", me_pagecache_clean },
795 { lru|dirty, lru|dirty, "dirty LRU", me_pagecache_dirty },
796 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
799 * Catchall entry: must be at end.
801 { 0, 0, "unknown page state", me_unknown },
818 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
819 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
821 static void action_result(unsigned long pfn, char *msg, int result)
823 pr_err("MCE %#lx: %s page recovery: %s\n",
824 pfn, msg, action_name[result]);
827 static int page_action(struct page_state *ps, struct page *p,
833 result = ps->action(p, pfn);
834 action_result(pfn, ps->msg, result);
836 count = page_count(p) - 1;
837 if (ps->action == me_swapcache_dirty && result == DELAYED)
841 "MCE %#lx: %s page still referenced by %d users\n",
842 pfn, ps->msg, count);
846 /* Could do more checks here if page looks ok */
848 * Could adjust zone counters here to correct for the missing page.
851 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
855 * Do all that is necessary to remove user space mappings. Unmap
856 * the pages and send SIGBUS to the processes if the data was dirty.
858 static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
859 int trapno, int flags)
861 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
862 struct address_space *mapping;
865 int kill = 1, forcekill;
866 struct page *hpage = compound_head(p);
869 if (PageReserved(p) || PageSlab(p))
873 * This check implies we don't kill processes if their pages
874 * are in the swap cache early. Those are always late kills.
876 if (!page_mapped(hpage))
882 if (PageSwapCache(p)) {
884 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
885 ttu |= TTU_IGNORE_HWPOISON;
889 * Propagate the dirty bit from PTEs to struct page first, because we
890 * need this to decide if we should kill or just drop the page.
891 * XXX: the dirty test could be racy: set_page_dirty() may not always
892 * be called inside page lock (it's recommended but not enforced).
894 mapping = page_mapping(hpage);
895 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
896 mapping_cap_writeback_dirty(mapping)) {
897 if (page_mkclean(hpage)) {
901 ttu |= TTU_IGNORE_HWPOISON;
903 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
909 * ppage: poisoned page
910 * if p is regular page(4k page)
911 * ppage == real poisoned page;
912 * else p is hugetlb or THP, ppage == head page.
916 if (PageTransHuge(hpage)) {
918 * Verify that this isn't a hugetlbfs head page, the check for
919 * PageAnon is just for avoid tripping a split_huge_page
920 * internal debug check, as split_huge_page refuses to deal with
921 * anything that isn't an anon page. PageAnon can't go away fro
922 * under us because we hold a refcount on the hpage, without a
923 * refcount on the hpage. split_huge_page can't be safely called
924 * in the first place, having a refcount on the tail isn't
925 * enough * to be safe.
927 if (!PageHuge(hpage) && PageAnon(hpage)) {
928 if (unlikely(split_huge_page(hpage))) {
930 * FIXME: if splitting THP is failed, it is
931 * better to stop the following operation rather
932 * than causing panic by unmapping. System might
933 * survive if the page is freed later.
936 "MCE %#lx: failed to split THP\n", pfn);
938 BUG_ON(!PageHWPoison(p));
941 /* THP is split, so ppage should be the real poisoned page. */
947 * First collect all the processes that have the page
948 * mapped in dirty form. This has to be done before try_to_unmap,
949 * because ttu takes the rmap data structures down.
951 * Error handling: We ignore errors here because
952 * there's nothing that can be done.
955 collect_procs(ppage, &tokill);
960 ret = try_to_unmap(ppage, ttu);
961 if (ret != SWAP_SUCCESS)
962 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
963 pfn, page_mapcount(ppage));
969 * Now that the dirty bit has been propagated to the
970 * struct page and all unmaps done we can decide if
971 * killing is needed or not. Only kill when the page
972 * was dirty or the process is not restartable,
973 * otherwise the tokill list is merely
974 * freed. When there was a problem unmapping earlier
975 * use a more force-full uncatchable kill to prevent
976 * any accesses to the poisoned memory.
978 forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL);
979 kill_procs(&tokill, forcekill, trapno,
980 ret != SWAP_SUCCESS, p, pfn, flags);
985 static void set_page_hwpoison_huge_page(struct page *hpage)
988 int nr_pages = 1 << compound_order(hpage);
989 for (i = 0; i < nr_pages; i++)
990 SetPageHWPoison(hpage + i);
993 static void clear_page_hwpoison_huge_page(struct page *hpage)
996 int nr_pages = 1 << compound_order(hpage);
997 for (i = 0; i < nr_pages; i++)
998 ClearPageHWPoison(hpage + i);
1002 * memory_failure - Handle memory failure of a page.
1003 * @pfn: Page Number of the corrupted page
1004 * @trapno: Trap number reported in the signal to user space.
1005 * @flags: fine tune action taken
1007 * This function is called by the low level machine check code
1008 * of an architecture when it detects hardware memory corruption
1009 * of a page. It tries its best to recover, which includes
1010 * dropping pages, killing processes etc.
1012 * The function is primarily of use for corruptions that
1013 * happen outside the current execution context (e.g. when
1014 * detected by a background scrubber)
1016 * Must run in process context (e.g. a work queue) with interrupts
1017 * enabled and no spinlocks hold.
1019 int memory_failure(unsigned long pfn, int trapno, int flags)
1021 struct page_state *ps;
1025 unsigned int nr_pages;
1026 unsigned long page_flags;
1028 if (!sysctl_memory_failure_recovery)
1029 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1031 if (!pfn_valid(pfn)) {
1033 "MCE %#lx: memory outside kernel control\n",
1038 p = pfn_to_page(pfn);
1039 hpage = compound_head(p);
1040 if (TestSetPageHWPoison(p)) {
1041 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
1046 * Currently errors on hugetlbfs pages are measured in hugepage units,
1047 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1048 * transparent hugepages, they are supposed to be split and error
1049 * measurement is done in normal page units. So nr_pages should be one
1053 nr_pages = 1 << compound_order(hpage);
1054 else /* normal page or thp */
1056 atomic_long_add(nr_pages, &num_poisoned_pages);
1059 * We need/can do nothing about count=0 pages.
1060 * 1) it's a free page, and therefore in safe hand:
1061 * prep_new_page() will be the gate keeper.
1062 * 2) it's a free hugepage, which is also safe:
1063 * an affected hugepage will be dequeued from hugepage freelist,
1064 * so there's no concern about reusing it ever after.
1065 * 3) it's part of a non-compound high order page.
1066 * Implies some kernel user: cannot stop them from
1067 * R/W the page; let's pray that the page has been
1068 * used and will be freed some time later.
1069 * In fact it's dangerous to directly bump up page count from 0,
1070 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1072 if (!(flags & MF_COUNT_INCREASED) &&
1073 !get_page_unless_zero(hpage)) {
1074 if (is_free_buddy_page(p)) {
1075 action_result(pfn, "free buddy", DELAYED);
1077 } else if (PageHuge(hpage)) {
1079 * Check "just unpoisoned", "filter hit", and
1080 * "race with other subpage."
1083 if (!PageHWPoison(hpage)
1084 || (hwpoison_filter(p) && TestClearPageHWPoison(p))
1085 || (p != hpage && TestSetPageHWPoison(hpage))) {
1086 atomic_long_sub(nr_pages, &num_poisoned_pages);
1089 set_page_hwpoison_huge_page(hpage);
1090 res = dequeue_hwpoisoned_huge_page(hpage);
1091 action_result(pfn, "free huge",
1092 res ? IGNORED : DELAYED);
1096 action_result(pfn, "high order kernel", IGNORED);
1102 * We ignore non-LRU pages for good reasons.
1103 * - PG_locked is only well defined for LRU pages and a few others
1104 * - to avoid races with __set_page_locked()
1105 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1106 * The check (unnecessarily) ignores LRU pages being isolated and
1107 * walked by the page reclaim code, however that's not a big loss.
1109 if (!PageHuge(p) && !PageTransTail(p)) {
1114 * shake_page could have turned it free.
1116 if (is_free_buddy_page(p)) {
1117 action_result(pfn, "free buddy, 2nd try",
1121 action_result(pfn, "non LRU", IGNORED);
1128 * Lock the page and wait for writeback to finish.
1129 * It's very difficult to mess with pages currently under IO
1130 * and in many cases impossible, so we just avoid it here.
1135 * We use page flags to determine what action should be taken, but
1136 * the flags can be modified by the error containment action. One
1137 * example is an mlocked page, where PG_mlocked is cleared by
1138 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1139 * correctly, we save a copy of the page flags at this time.
1141 page_flags = p->flags;
1144 * unpoison always clear PG_hwpoison inside page lock
1146 if (!PageHWPoison(p)) {
1147 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1151 if (hwpoison_filter(p)) {
1152 if (TestClearPageHWPoison(p))
1153 atomic_long_sub(nr_pages, &num_poisoned_pages);
1160 * For error on the tail page, we should set PG_hwpoison
1161 * on the head page to show that the hugepage is hwpoisoned
1163 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1164 action_result(pfn, "hugepage already hardware poisoned",
1171 * Set PG_hwpoison on all pages in an error hugepage,
1172 * because containment is done in hugepage unit for now.
1173 * Since we have done TestSetPageHWPoison() for the head page with
1174 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1177 set_page_hwpoison_huge_page(hpage);
1179 wait_on_page_writeback(p);
1182 * Now take care of user space mappings.
1183 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1185 if (hwpoison_user_mappings(p, pfn, trapno, flags) != SWAP_SUCCESS) {
1186 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1192 * Torn down by someone else?
1194 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1195 action_result(pfn, "already truncated LRU", IGNORED);
1202 * The first check uses the current page flags which may not have any
1203 * relevant information. The second check with the saved page flagss is
1204 * carried out only if the first check can't determine the page status.
1206 for (ps = error_states;; ps++)
1207 if ((p->flags & ps->mask) == ps->res)
1210 page_flags |= (p->flags & (1UL << PG_dirty));
1213 for (ps = error_states;; ps++)
1214 if ((page_flags & ps->mask) == ps->res)
1216 res = page_action(ps, p, pfn);
1221 EXPORT_SYMBOL_GPL(memory_failure);
1223 #define MEMORY_FAILURE_FIFO_ORDER 4
1224 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1226 struct memory_failure_entry {
1232 struct memory_failure_cpu {
1233 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1234 MEMORY_FAILURE_FIFO_SIZE);
1236 struct work_struct work;
1239 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1242 * memory_failure_queue - Schedule handling memory failure of a page.
1243 * @pfn: Page Number of the corrupted page
1244 * @trapno: Trap number reported in the signal to user space.
1245 * @flags: Flags for memory failure handling
1247 * This function is called by the low level hardware error handler
1248 * when it detects hardware memory corruption of a page. It schedules
1249 * the recovering of error page, including dropping pages, killing
1252 * The function is primarily of use for corruptions that
1253 * happen outside the current execution context (e.g. when
1254 * detected by a background scrubber)
1256 * Can run in IRQ context.
1258 void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1260 struct memory_failure_cpu *mf_cpu;
1261 unsigned long proc_flags;
1262 struct memory_failure_entry entry = {
1268 mf_cpu = &get_cpu_var(memory_failure_cpu);
1269 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1270 if (kfifo_put(&mf_cpu->fifo, &entry))
1271 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1273 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1275 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1276 put_cpu_var(memory_failure_cpu);
1278 EXPORT_SYMBOL_GPL(memory_failure_queue);
1280 static void memory_failure_work_func(struct work_struct *work)
1282 struct memory_failure_cpu *mf_cpu;
1283 struct memory_failure_entry entry = { 0, };
1284 unsigned long proc_flags;
1287 mf_cpu = &__get_cpu_var(memory_failure_cpu);
1289 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1290 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1291 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1294 if (entry.flags & MF_SOFT_OFFLINE)
1295 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1297 memory_failure(entry.pfn, entry.trapno, entry.flags);
1301 static int __init memory_failure_init(void)
1303 struct memory_failure_cpu *mf_cpu;
1306 for_each_possible_cpu(cpu) {
1307 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1308 spin_lock_init(&mf_cpu->lock);
1309 INIT_KFIFO(mf_cpu->fifo);
1310 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1315 core_initcall(memory_failure_init);
1318 * unpoison_memory - Unpoison a previously poisoned page
1319 * @pfn: Page number of the to be unpoisoned page
1321 * Software-unpoison a page that has been poisoned by
1322 * memory_failure() earlier.
1324 * This is only done on the software-level, so it only works
1325 * for linux injected failures, not real hardware failures
1327 * Returns 0 for success, otherwise -errno.
1329 int unpoison_memory(unsigned long pfn)
1334 unsigned int nr_pages;
1336 if (!pfn_valid(pfn))
1339 p = pfn_to_page(pfn);
1340 page = compound_head(p);
1342 if (!PageHWPoison(p)) {
1343 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
1348 * unpoison_memory() can encounter thp only when the thp is being
1349 * worked by memory_failure() and the page lock is not held yet.
1350 * In such case, we yield to memory_failure() and make unpoison fail.
1352 if (!PageHuge(page) && PageTransHuge(page)) {
1353 pr_info("MCE: Memory failure is now running on %#lx\n", pfn);
1357 nr_pages = 1 << compound_order(page);
1359 if (!get_page_unless_zero(page)) {
1361 * Since HWPoisoned hugepage should have non-zero refcount,
1362 * race between memory failure and unpoison seems to happen.
1363 * In such case unpoison fails and memory failure runs
1366 if (PageHuge(page)) {
1367 pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1370 if (TestClearPageHWPoison(p))
1371 atomic_long_dec(&num_poisoned_pages);
1372 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
1378 * This test is racy because PG_hwpoison is set outside of page lock.
1379 * That's acceptable because that won't trigger kernel panic. Instead,
1380 * the PG_hwpoison page will be caught and isolated on the entrance to
1381 * the free buddy page pool.
1383 if (TestClearPageHWPoison(page)) {
1384 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
1385 atomic_long_sub(nr_pages, &num_poisoned_pages);
1388 clear_page_hwpoison_huge_page(page);
1393 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
1398 EXPORT_SYMBOL(unpoison_memory);
1400 static struct page *new_page(struct page *p, unsigned long private, int **x)
1402 int nid = page_to_nid(p);
1404 return alloc_huge_page_node(page_hstate(compound_head(p)),
1407 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1411 * Safely get reference count of an arbitrary page.
1412 * Returns 0 for a free page, -EIO for a zero refcount page
1413 * that is not free, and 1 for any other page type.
1414 * For 1 the page is returned with increased page count, otherwise not.
1416 static int __get_any_page(struct page *p, unsigned long pfn, int flags)
1420 if (flags & MF_COUNT_INCREASED)
1424 * The lock_memory_hotplug prevents a race with memory hotplug.
1425 * This is a big hammer, a better would be nicer.
1427 lock_memory_hotplug();
1430 * Isolate the page, so that it doesn't get reallocated if it
1431 * was free. This flag should be kept set until the source page
1432 * is freed and PG_hwpoison on it is set.
1434 if (get_pageblock_migratetype(p) != MIGRATE_ISOLATE)
1435 set_migratetype_isolate(p, true);
1437 * When the target page is a free hugepage, just remove it
1438 * from free hugepage list.
1440 if (!get_page_unless_zero(compound_head(p))) {
1442 pr_info("%s: %#lx free huge page\n", __func__, pfn);
1444 } else if (is_free_buddy_page(p)) {
1445 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1448 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1449 __func__, pfn, p->flags);
1453 /* Not a free page */
1456 unlock_memory_hotplug();
1460 static int get_any_page(struct page *page, unsigned long pfn, int flags)
1462 int ret = __get_any_page(page, pfn, flags);
1464 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1469 shake_page(page, 1);
1474 ret = __get_any_page(page, pfn, 0);
1475 if (!PageLRU(page)) {
1476 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1484 static int soft_offline_huge_page(struct page *page, int flags)
1487 unsigned long pfn = page_to_pfn(page);
1488 struct page *hpage = compound_head(page);
1489 LIST_HEAD(pagelist);
1492 * This double-check of PageHWPoison is to avoid the race with
1493 * memory_failure(). See also comment in __soft_offline_page().
1496 if (PageHWPoison(hpage)) {
1499 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1504 /* Keep page count to indicate a given hugepage is isolated. */
1505 list_move(&hpage->lru, &pagelist);
1506 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1507 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1509 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1510 pfn, ret, page->flags);
1512 * We know that soft_offline_huge_page() tries to migrate
1513 * only one hugepage pointed to by hpage, so we need not
1514 * run through the pagelist here.
1516 putback_active_hugepage(hpage);
1520 set_page_hwpoison_huge_page(hpage);
1521 dequeue_hwpoisoned_huge_page(hpage);
1522 atomic_long_add(1 << compound_order(hpage),
1523 &num_poisoned_pages);
1528 static int __soft_offline_page(struct page *page, int flags)
1531 unsigned long pfn = page_to_pfn(page);
1534 * Check PageHWPoison again inside page lock because PageHWPoison
1535 * is set by memory_failure() outside page lock. Note that
1536 * memory_failure() also double-checks PageHWPoison inside page lock,
1537 * so there's no race between soft_offline_page() and memory_failure().
1540 wait_on_page_writeback(page);
1541 if (PageHWPoison(page)) {
1544 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1548 * Try to invalidate first. This should work for
1549 * non dirty unmapped page cache pages.
1551 ret = invalidate_inode_page(page);
1554 * RED-PEN would be better to keep it isolated here, but we
1555 * would need to fix isolation locking first.
1559 pr_info("soft_offline: %#lx: invalidated\n", pfn);
1560 SetPageHWPoison(page);
1561 atomic_long_inc(&num_poisoned_pages);
1566 * Simple invalidation didn't work.
1567 * Try to migrate to a new page instead. migrate.c
1568 * handles a large number of cases for us.
1570 ret = isolate_lru_page(page);
1572 * Drop page reference which is came from get_any_page()
1573 * successful isolate_lru_page() already took another one.
1577 LIST_HEAD(pagelist);
1578 inc_zone_page_state(page, NR_ISOLATED_ANON +
1579 page_is_file_cache(page));
1580 list_add(&page->lru, &pagelist);
1581 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1582 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1584 putback_lru_pages(&pagelist);
1585 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1586 pfn, ret, page->flags);
1591 * After page migration succeeds, the source page can
1592 * be trapped in pagevec and actual freeing is delayed.
1593 * Freeing code works differently based on PG_hwpoison,
1594 * so there's a race. We need to make sure that the
1595 * source page should be freed back to buddy before
1596 * setting PG_hwpoison.
1598 if (!is_free_buddy_page(page))
1599 lru_add_drain_all();
1600 if (!is_free_buddy_page(page))
1602 SetPageHWPoison(page);
1603 if (!is_free_buddy_page(page))
1604 pr_info("soft offline: %#lx: page leaked\n",
1606 atomic_long_inc(&num_poisoned_pages);
1609 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1610 pfn, ret, page_count(page), page->flags);
1616 * soft_offline_page - Soft offline a page.
1617 * @page: page to offline
1618 * @flags: flags. Same as memory_failure().
1620 * Returns 0 on success, otherwise negated errno.
1622 * Soft offline a page, by migration or invalidation,
1623 * without killing anything. This is for the case when
1624 * a page is not corrupted yet (so it's still valid to access),
1625 * but has had a number of corrected errors and is better taken
1628 * The actual policy on when to do that is maintained by
1631 * This should never impact any application or cause data loss,
1632 * however it might take some time.
1634 * This is not a 100% solution for all memory, but tries to be
1635 * ``good enough'' for the majority of memory.
1637 int soft_offline_page(struct page *page, int flags)
1640 unsigned long pfn = page_to_pfn(page);
1641 struct page *hpage = compound_trans_head(page);
1643 if (PageHWPoison(page)) {
1644 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1647 if (!PageHuge(page) && PageTransHuge(hpage)) {
1648 if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1649 pr_info("soft offline: %#lx: failed to split THP\n",
1655 ret = get_any_page(page, pfn, flags);
1658 if (ret) { /* for in-use pages */
1660 ret = soft_offline_huge_page(page, flags);
1662 ret = __soft_offline_page(page, flags);
1663 } else { /* for free pages */
1664 if (PageHuge(page)) {
1665 set_page_hwpoison_huge_page(hpage);
1666 dequeue_hwpoisoned_huge_page(hpage);
1667 atomic_long_add(1 << compound_order(hpage),
1668 &num_poisoned_pages);
1670 SetPageHWPoison(page);
1671 atomic_long_inc(&num_poisoned_pages);
1675 unset_migratetype_isolate(page, MIGRATE_MOVABLE);