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 2bit ECC memory or cache
13 * Handles page cache pages in various states. The tricky part
14 * here is that we can access any page asynchronous to other VM
15 * users, because memory failures could happen anytime and anywhere,
16 * possibly violating some of their assumptions. This is why this code
17 * has to be extremely careful. Generally it tries to use normal locking
18 * rules, as in get the standard locks, even if that means the
19 * error handling takes potentially a long time.
21 * The operation to map back from RMAP chains to processes has to walk
22 * the complete process list and has non linear complexity with the number
23 * mappings. In short it can be quite slow. But since memory corruptions
24 * are rare we hope to get away with this.
29 * - hugetlb needs more code
30 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
31 * - pass bad pages to kdump next kernel
33 #define DEBUG 1 /* remove me in 2.6.34 */
34 #include <linux/kernel.h>
36 #include <linux/page-flags.h>
37 #include <linux/kernel-page-flags.h>
38 #include <linux/sched.h>
39 #include <linux/ksm.h>
40 #include <linux/rmap.h>
41 #include <linux/pagemap.h>
42 #include <linux/swap.h>
43 #include <linux/backing-dev.h>
44 #include <linux/migrate.h>
45 #include <linux/page-isolation.h>
46 #include <linux/suspend.h>
47 #include <linux/slab.h>
48 #include <linux/swapops.h>
49 #include <linux/hugetlb.h>
52 int sysctl_memory_failure_early_kill __read_mostly = 0;
54 int sysctl_memory_failure_recovery __read_mostly = 1;
56 atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
58 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
60 u32 hwpoison_filter_enable = 0;
61 u32 hwpoison_filter_dev_major = ~0U;
62 u32 hwpoison_filter_dev_minor = ~0U;
63 u64 hwpoison_filter_flags_mask;
64 u64 hwpoison_filter_flags_value;
65 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
66 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
67 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
68 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
69 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
71 static int hwpoison_filter_dev(struct page *p)
73 struct address_space *mapping;
76 if (hwpoison_filter_dev_major == ~0U &&
77 hwpoison_filter_dev_minor == ~0U)
81 * page_mapping() does not accept slab page
86 mapping = page_mapping(p);
87 if (mapping == NULL || mapping->host == NULL)
90 dev = mapping->host->i_sb->s_dev;
91 if (hwpoison_filter_dev_major != ~0U &&
92 hwpoison_filter_dev_major != MAJOR(dev))
94 if (hwpoison_filter_dev_minor != ~0U &&
95 hwpoison_filter_dev_minor != MINOR(dev))
101 static int hwpoison_filter_flags(struct page *p)
103 if (!hwpoison_filter_flags_mask)
106 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
107 hwpoison_filter_flags_value)
114 * This allows stress tests to limit test scope to a collection of tasks
115 * by putting them under some memcg. This prevents killing unrelated/important
116 * processes such as /sbin/init. Note that the target task may share clean
117 * pages with init (eg. libc text), which is harmless. If the target task
118 * share _dirty_ pages with another task B, the test scheme must make sure B
119 * is also included in the memcg. At last, due to race conditions this filter
120 * can only guarantee that the page either belongs to the memcg tasks, or is
123 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
124 u64 hwpoison_filter_memcg;
125 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
126 static int hwpoison_filter_task(struct page *p)
128 struct mem_cgroup *mem;
129 struct cgroup_subsys_state *css;
132 if (!hwpoison_filter_memcg)
135 mem = try_get_mem_cgroup_from_page(p);
139 css = mem_cgroup_css(mem);
140 /* root_mem_cgroup has NULL dentries */
141 if (!css->cgroup->dentry)
144 ino = css->cgroup->dentry->d_inode->i_ino;
147 if (ino != hwpoison_filter_memcg)
153 static int hwpoison_filter_task(struct page *p) { return 0; }
156 int hwpoison_filter(struct page *p)
158 if (!hwpoison_filter_enable)
161 if (hwpoison_filter_dev(p))
164 if (hwpoison_filter_flags(p))
167 if (hwpoison_filter_task(p))
173 int hwpoison_filter(struct page *p)
179 EXPORT_SYMBOL_GPL(hwpoison_filter);
182 * Send all the processes who have the page mapped an ``action optional''
185 static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
186 unsigned long pfn, struct page *page)
192 "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
193 pfn, t->comm, t->pid);
194 si.si_signo = SIGBUS;
196 si.si_code = BUS_MCEERR_AO;
197 si.si_addr = (void *)addr;
198 #ifdef __ARCH_SI_TRAPNO
199 si.si_trapno = trapno;
201 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
203 * Don't use force here, it's convenient if the signal
204 * can be temporarily blocked.
205 * This could cause a loop when the user sets SIGBUS
206 * to SIG_IGN, but hopefully noone will do that?
208 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
210 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
211 t->comm, t->pid, ret);
216 * When a unknown page type is encountered drain as many buffers as possible
217 * in the hope to turn the page into a LRU or free page, which we can handle.
219 void shake_page(struct page *p, int access)
226 if (PageLRU(p) || is_free_buddy_page(p))
231 * Only all shrink_slab here (which would also
232 * shrink other caches) if access is not potentially fatal.
237 nr = shrink_slab(1000, GFP_KERNEL, 1000);
238 if (page_count(p) == 1)
243 EXPORT_SYMBOL_GPL(shake_page);
246 * Kill all processes that have a poisoned page mapped and then isolate
250 * Find all processes having the page mapped and kill them.
251 * But we keep a page reference around so that the page is not
252 * actually freed yet.
253 * Then stash the page away
255 * There's no convenient way to get back to mapped processes
256 * from the VMAs. So do a brute-force search over all
259 * Remember that machine checks are not common (or rather
260 * if they are common you have other problems), so this shouldn't
261 * be a performance issue.
263 * Also there are some races possible while we get from the
264 * error detection to actually handle it.
269 struct task_struct *tsk;
271 unsigned addr_valid:1;
275 * Failure handling: if we can't find or can't kill a process there's
276 * not much we can do. We just print a message and ignore otherwise.
280 * Schedule a process for later kill.
281 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
282 * TBD would GFP_NOIO be enough?
284 static void add_to_kill(struct task_struct *tsk, struct page *p,
285 struct vm_area_struct *vma,
286 struct list_head *to_kill,
287 struct to_kill **tkc)
295 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
298 "MCE: Out of memory while machine check handling\n");
302 tk->addr = page_address_in_vma(p, vma);
306 * In theory we don't have to kill when the page was
307 * munmaped. But it could be also a mremap. Since that's
308 * likely very rare kill anyways just out of paranoia, but use
309 * a SIGKILL because the error is not contained anymore.
311 if (tk->addr == -EFAULT) {
312 pr_debug("MCE: Unable to find user space address %lx in %s\n",
313 page_to_pfn(p), tsk->comm);
316 get_task_struct(tsk);
318 list_add_tail(&tk->nd, to_kill);
322 * Kill the processes that have been collected earlier.
324 * Only do anything when DOIT is set, otherwise just free the list
325 * (this is used for clean pages which do not need killing)
326 * Also when FAIL is set do a force kill because something went
329 static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
330 int fail, struct page *page, unsigned long pfn)
332 struct to_kill *tk, *next;
334 list_for_each_entry_safe (tk, next, to_kill, nd) {
337 * In case something went wrong with munmapping
338 * make sure the process doesn't catch the
339 * signal and then access the memory. Just kill it.
341 if (fail || tk->addr_valid == 0) {
343 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
344 pfn, tk->tsk->comm, tk->tsk->pid);
345 force_sig(SIGKILL, tk->tsk);
349 * In theory the process could have mapped
350 * something else on the address in-between. We could
351 * check for that, but we need to tell the
354 else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
357 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
358 pfn, tk->tsk->comm, tk->tsk->pid);
360 put_task_struct(tk->tsk);
365 static int task_early_kill(struct task_struct *tsk)
369 if (tsk->flags & PF_MCE_PROCESS)
370 return !!(tsk->flags & PF_MCE_EARLY);
371 return sysctl_memory_failure_early_kill;
375 * Collect processes when the error hit an anonymous page.
377 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
378 struct to_kill **tkc)
380 struct vm_area_struct *vma;
381 struct task_struct *tsk;
384 read_lock(&tasklist_lock);
385 av = page_lock_anon_vma(page);
386 if (av == NULL) /* Not actually mapped anymore */
388 for_each_process (tsk) {
389 struct anon_vma_chain *vmac;
391 if (!task_early_kill(tsk))
393 list_for_each_entry(vmac, &av->head, same_anon_vma) {
395 if (!page_mapped_in_vma(page, vma))
397 if (vma->vm_mm == tsk->mm)
398 add_to_kill(tsk, page, vma, to_kill, tkc);
401 page_unlock_anon_vma(av);
403 read_unlock(&tasklist_lock);
407 * Collect processes when the error hit a file mapped page.
409 static void collect_procs_file(struct page *page, struct list_head *to_kill,
410 struct to_kill **tkc)
412 struct vm_area_struct *vma;
413 struct task_struct *tsk;
414 struct prio_tree_iter iter;
415 struct address_space *mapping = page->mapping;
418 * A note on the locking order between the two locks.
419 * We don't rely on this particular order.
420 * If you have some other code that needs a different order
421 * feel free to switch them around. Or add a reverse link
422 * from mm_struct to task_struct, then this could be all
423 * done without taking tasklist_lock and looping over all tasks.
426 read_lock(&tasklist_lock);
427 spin_lock(&mapping->i_mmap_lock);
428 for_each_process(tsk) {
429 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
431 if (!task_early_kill(tsk))
434 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
437 * Send early kill signal to tasks where a vma covers
438 * the page but the corrupted page is not necessarily
439 * mapped it in its pte.
440 * Assume applications who requested early kill want
441 * to be informed of all such data corruptions.
443 if (vma->vm_mm == tsk->mm)
444 add_to_kill(tsk, page, vma, to_kill, tkc);
447 spin_unlock(&mapping->i_mmap_lock);
448 read_unlock(&tasklist_lock);
452 * Collect the processes who have the corrupted page mapped to kill.
453 * This is done in two steps for locking reasons.
454 * First preallocate one tokill structure outside the spin locks,
455 * so that we can kill at least one process reasonably reliable.
457 static void collect_procs(struct page *page, struct list_head *tokill)
464 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
468 collect_procs_anon(page, tokill, &tk);
470 collect_procs_file(page, tokill, &tk);
475 * Error handlers for various types of pages.
479 IGNORED, /* Error: cannot be handled */
480 FAILED, /* Error: handling failed */
481 DELAYED, /* Will be handled later */
482 RECOVERED, /* Successfully recovered */
485 static const char *action_name[] = {
486 [IGNORED] = "Ignored",
488 [DELAYED] = "Delayed",
489 [RECOVERED] = "Recovered",
493 * XXX: It is possible that a page is isolated from LRU cache,
494 * and then kept in swap cache or failed to remove from page cache.
495 * The page count will stop it from being freed by unpoison.
496 * Stress tests should be aware of this memory leak problem.
498 static int delete_from_lru_cache(struct page *p)
500 if (!isolate_lru_page(p)) {
502 * Clear sensible page flags, so that the buddy system won't
503 * complain when the page is unpoison-and-freed.
506 ClearPageUnevictable(p);
508 * drop the page count elevated by isolate_lru_page()
510 page_cache_release(p);
517 * Error hit kernel page.
518 * Do nothing, try to be lucky and not touch this instead. For a few cases we
519 * could be more sophisticated.
521 static int me_kernel(struct page *p, unsigned long pfn)
527 * Page in unknown state. Do nothing.
529 static int me_unknown(struct page *p, unsigned long pfn)
531 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
536 * Clean (or cleaned) page cache page.
538 static int me_pagecache_clean(struct page *p, unsigned long pfn)
542 struct address_space *mapping;
544 delete_from_lru_cache(p);
547 * For anonymous pages we're done the only reference left
548 * should be the one m_f() holds.
554 * Now truncate the page in the page cache. This is really
555 * more like a "temporary hole punch"
556 * Don't do this for block devices when someone else
557 * has a reference, because it could be file system metadata
558 * and that's not safe to truncate.
560 mapping = page_mapping(p);
563 * Page has been teared down in the meanwhile
569 * Truncation is a bit tricky. Enable it per file system for now.
571 * Open: to take i_mutex or not for this? Right now we don't.
573 if (mapping->a_ops->error_remove_page) {
574 err = mapping->a_ops->error_remove_page(mapping, p);
576 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
578 } else if (page_has_private(p) &&
579 !try_to_release_page(p, GFP_NOIO)) {
580 pr_debug("MCE %#lx: failed to release buffers\n", pfn);
586 * If the file system doesn't support it just invalidate
587 * This fails on dirty or anything with private pages
589 if (invalidate_inode_page(p))
592 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
599 * Dirty cache page page
600 * Issues: when the error hit a hole page the error is not properly
603 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
605 struct address_space *mapping = page_mapping(p);
608 /* TBD: print more information about the file. */
611 * IO error will be reported by write(), fsync(), etc.
612 * who check the mapping.
613 * This way the application knows that something went
614 * wrong with its dirty file data.
616 * There's one open issue:
618 * The EIO will be only reported on the next IO
619 * operation and then cleared through the IO map.
620 * Normally Linux has two mechanisms to pass IO error
621 * first through the AS_EIO flag in the address space
622 * and then through the PageError flag in the page.
623 * Since we drop pages on memory failure handling the
624 * only mechanism open to use is through AS_AIO.
626 * This has the disadvantage that it gets cleared on
627 * the first operation that returns an error, while
628 * the PageError bit is more sticky and only cleared
629 * when the page is reread or dropped. If an
630 * application assumes it will always get error on
631 * fsync, but does other operations on the fd before
632 * and the page is dropped inbetween then the error
633 * will not be properly reported.
635 * This can already happen even without hwpoisoned
636 * pages: first on metadata IO errors (which only
637 * report through AS_EIO) or when the page is dropped
640 * So right now we assume that the application DTRT on
641 * the first EIO, but we're not worse than other parts
644 mapping_set_error(mapping, EIO);
647 return me_pagecache_clean(p, pfn);
651 * Clean and dirty swap cache.
653 * Dirty swap cache page is tricky to handle. The page could live both in page
654 * cache and swap cache(ie. page is freshly swapped in). So it could be
655 * referenced concurrently by 2 types of PTEs:
656 * normal PTEs and swap PTEs. We try to handle them consistently by calling
657 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
659 * - clear dirty bit to prevent IO
661 * - but keep in the swap cache, so that when we return to it on
662 * a later page fault, we know the application is accessing
663 * corrupted data and shall be killed (we installed simple
664 * interception code in do_swap_page to catch it).
666 * Clean swap cache pages can be directly isolated. A later page fault will
667 * bring in the known good data from disk.
669 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
672 /* Trigger EIO in shmem: */
673 ClearPageUptodate(p);
675 if (!delete_from_lru_cache(p))
681 static int me_swapcache_clean(struct page *p, unsigned long pfn)
683 delete_from_swap_cache(p);
685 if (!delete_from_lru_cache(p))
692 * Huge pages. Needs work.
694 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
695 * To narrow down kill region to one page, we need to break up pmd.
696 * - To support soft-offlining for hugepage, we need to support hugepage
699 static int me_huge_page(struct page *p, unsigned long pfn)
702 struct page *hpage = compound_head(p);
704 * We can safely recover from error on free or reserved (i.e.
705 * not in-use) hugepage by dequeuing it from freelist.
706 * To check whether a hugepage is in-use or not, we can't use
707 * page->lru because it can be used in other hugepage operations,
708 * such as __unmap_hugepage_range() and gather_surplus_pages().
709 * So instead we use page_mapping() and PageAnon().
710 * We assume that this function is called with page lock held,
711 * so there is no race between isolation and mapping/unmapping.
713 if (!(page_mapping(hpage) || PageAnon(hpage))) {
714 res = dequeue_hwpoisoned_huge_page(hpage);
722 * Various page states we can handle.
724 * A page state is defined by its current page->flags bits.
725 * The table matches them in order and calls the right handler.
727 * This is quite tricky because we can access page at any time
728 * in its live cycle, so all accesses have to be extremly careful.
730 * This is not complete. More states could be added.
731 * For any missing state don't attempt recovery.
734 #define dirty (1UL << PG_dirty)
735 #define sc (1UL << PG_swapcache)
736 #define unevict (1UL << PG_unevictable)
737 #define mlock (1UL << PG_mlocked)
738 #define writeback (1UL << PG_writeback)
739 #define lru (1UL << PG_lru)
740 #define swapbacked (1UL << PG_swapbacked)
741 #define head (1UL << PG_head)
742 #define tail (1UL << PG_tail)
743 #define compound (1UL << PG_compound)
744 #define slab (1UL << PG_slab)
745 #define reserved (1UL << PG_reserved)
747 static struct page_state {
751 int (*action)(struct page *p, unsigned long pfn);
753 { reserved, reserved, "reserved kernel", me_kernel },
755 * free pages are specially detected outside this table:
756 * PG_buddy pages only make a small fraction of all free pages.
760 * Could in theory check if slab page is free or if we can drop
761 * currently unused objects without touching them. But just
762 * treat it as standard kernel for now.
764 { slab, slab, "kernel slab", me_kernel },
766 #ifdef CONFIG_PAGEFLAGS_EXTENDED
767 { head, head, "huge", me_huge_page },
768 { tail, tail, "huge", me_huge_page },
770 { compound, compound, "huge", me_huge_page },
773 { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
774 { sc|dirty, sc, "swapcache", me_swapcache_clean },
776 { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
777 { unevict, unevict, "unevictable LRU", me_pagecache_clean},
779 { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
780 { mlock, mlock, "mlocked LRU", me_pagecache_clean },
782 { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
783 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
786 * Catchall entry: must be at end.
788 { 0, 0, "unknown page state", me_unknown },
804 static void action_result(unsigned long pfn, char *msg, int result)
806 struct page *page = pfn_to_page(pfn);
808 printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
810 PageDirty(page) ? "dirty " : "",
811 msg, action_name[result]);
814 static int page_action(struct page_state *ps, struct page *p,
820 result = ps->action(p, pfn);
821 action_result(pfn, ps->msg, result);
823 count = page_count(p) - 1;
824 if (ps->action == me_swapcache_dirty && result == DELAYED)
828 "MCE %#lx: %s page still referenced by %d users\n",
829 pfn, ps->msg, count);
833 /* Could do more checks here if page looks ok */
835 * Could adjust zone counters here to correct for the missing page.
838 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
841 #define N_UNMAP_TRIES 5
844 * Do all that is necessary to remove user space mappings. Unmap
845 * the pages and send SIGBUS to the processes if the data was dirty.
847 static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
850 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
851 struct address_space *mapping;
856 struct page *hpage = compound_head(p);
858 if (PageReserved(p) || PageSlab(p))
862 * This check implies we don't kill processes if their pages
863 * are in the swap cache early. Those are always late kills.
865 if (!page_mapped(hpage))
871 if (PageSwapCache(p)) {
873 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
874 ttu |= TTU_IGNORE_HWPOISON;
878 * Propagate the dirty bit from PTEs to struct page first, because we
879 * need this to decide if we should kill or just drop the page.
880 * XXX: the dirty test could be racy: set_page_dirty() may not always
881 * be called inside page lock (it's recommended but not enforced).
883 mapping = page_mapping(hpage);
884 if (!PageDirty(hpage) && mapping &&
885 mapping_cap_writeback_dirty(mapping)) {
886 if (page_mkclean(hpage)) {
890 ttu |= TTU_IGNORE_HWPOISON;
892 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
898 * First collect all the processes that have the page
899 * mapped in dirty form. This has to be done before try_to_unmap,
900 * because ttu takes the rmap data structures down.
902 * Error handling: We ignore errors here because
903 * there's nothing that can be done.
906 collect_procs(hpage, &tokill);
909 * try_to_unmap can fail temporarily due to races.
910 * Try a few times (RED-PEN better strategy?)
912 for (i = 0; i < N_UNMAP_TRIES; i++) {
913 ret = try_to_unmap(hpage, ttu);
914 if (ret == SWAP_SUCCESS)
916 pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn, ret);
919 if (ret != SWAP_SUCCESS)
920 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
921 pfn, page_mapcount(hpage));
924 * Now that the dirty bit has been propagated to the
925 * struct page and all unmaps done we can decide if
926 * killing is needed or not. Only kill when the page
927 * was dirty, otherwise the tokill list is merely
928 * freed. When there was a problem unmapping earlier
929 * use a more force-full uncatchable kill to prevent
930 * any accesses to the poisoned memory.
932 kill_procs_ao(&tokill, !!PageDirty(hpage), trapno,
933 ret != SWAP_SUCCESS, p, pfn);
938 static void set_page_hwpoison_huge_page(struct page *hpage)
941 int nr_pages = 1 << compound_order(hpage);
942 for (i = 0; i < nr_pages; i++)
943 SetPageHWPoison(hpage + i);
946 static void clear_page_hwpoison_huge_page(struct page *hpage)
949 int nr_pages = 1 << compound_order(hpage);
950 for (i = 0; i < nr_pages; i++)
951 ClearPageHWPoison(hpage + i);
954 int __memory_failure(unsigned long pfn, int trapno, int flags)
956 struct page_state *ps;
960 unsigned int nr_pages;
962 if (!sysctl_memory_failure_recovery)
963 panic("Memory failure from trap %d on page %lx", trapno, pfn);
965 if (!pfn_valid(pfn)) {
967 "MCE %#lx: memory outside kernel control\n",
972 p = pfn_to_page(pfn);
973 hpage = compound_head(p);
974 if (TestSetPageHWPoison(p)) {
975 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
979 nr_pages = 1 << compound_order(hpage);
980 atomic_long_add(nr_pages, &mce_bad_pages);
983 * We need/can do nothing about count=0 pages.
984 * 1) it's a free page, and therefore in safe hand:
985 * prep_new_page() will be the gate keeper.
986 * 2) it's part of a non-compound high order page.
987 * Implies some kernel user: cannot stop them from
988 * R/W the page; let's pray that the page has been
989 * used and will be freed some time later.
990 * In fact it's dangerous to directly bump up page count from 0,
991 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
993 if (!(flags & MF_COUNT_INCREASED) &&
994 !get_page_unless_zero(hpage)) {
995 if (is_free_buddy_page(p)) {
996 action_result(pfn, "free buddy", DELAYED);
999 action_result(pfn, "high order kernel", IGNORED);
1005 * We ignore non-LRU pages for good reasons.
1006 * - PG_locked is only well defined for LRU pages and a few others
1007 * - to avoid races with __set_page_locked()
1008 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1009 * The check (unnecessarily) ignores LRU pages being isolated and
1010 * walked by the page reclaim code, however that's not a big loss.
1012 if (!PageLRU(p) && !PageHuge(p))
1014 if (!PageLRU(p) && !PageHuge(p)) {
1016 * shake_page could have turned it free.
1018 if (is_free_buddy_page(p)) {
1019 action_result(pfn, "free buddy, 2nd try", DELAYED);
1022 action_result(pfn, "non LRU", IGNORED);
1028 * Lock the page and wait for writeback to finish.
1029 * It's very difficult to mess with pages currently under IO
1030 * and in many cases impossible, so we just avoid it here.
1032 lock_page_nosync(hpage);
1035 * unpoison always clear PG_hwpoison inside page lock
1037 if (!PageHWPoison(p)) {
1038 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1042 if (hwpoison_filter(p)) {
1043 if (TestClearPageHWPoison(p))
1044 atomic_long_sub(nr_pages, &mce_bad_pages);
1051 * For error on the tail page, we should set PG_hwpoison
1052 * on the head page to show that the hugepage is hwpoisoned
1054 if (PageTail(p) && TestSetPageHWPoison(hpage)) {
1055 action_result(pfn, "hugepage already hardware poisoned",
1062 * Set PG_hwpoison on all pages in an error hugepage,
1063 * because containment is done in hugepage unit for now.
1064 * Since we have done TestSetPageHWPoison() for the head page with
1065 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1068 set_page_hwpoison_huge_page(hpage);
1070 wait_on_page_writeback(p);
1073 * Now take care of user space mappings.
1074 * Abort on fail: __remove_from_page_cache() assumes unmapped page.
1076 if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
1077 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1083 * Torn down by someone else?
1085 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1086 action_result(pfn, "already truncated LRU", IGNORED);
1092 for (ps = error_states;; ps++) {
1093 if ((p->flags & ps->mask) == ps->res) {
1094 res = page_action(ps, p, pfn);
1102 EXPORT_SYMBOL_GPL(__memory_failure);
1105 * memory_failure - Handle memory failure of a page.
1106 * @pfn: Page Number of the corrupted page
1107 * @trapno: Trap number reported in the signal to user space.
1109 * This function is called by the low level machine check code
1110 * of an architecture when it detects hardware memory corruption
1111 * of a page. It tries its best to recover, which includes
1112 * dropping pages, killing processes etc.
1114 * The function is primarily of use for corruptions that
1115 * happen outside the current execution context (e.g. when
1116 * detected by a background scrubber)
1118 * Must run in process context (e.g. a work queue) with interrupts
1119 * enabled and no spinlocks hold.
1121 void memory_failure(unsigned long pfn, int trapno)
1123 __memory_failure(pfn, trapno, 0);
1127 * unpoison_memory - Unpoison a previously poisoned page
1128 * @pfn: Page number of the to be unpoisoned page
1130 * Software-unpoison a page that has been poisoned by
1131 * memory_failure() earlier.
1133 * This is only done on the software-level, so it only works
1134 * for linux injected failures, not real hardware failures
1136 * Returns 0 for success, otherwise -errno.
1138 int unpoison_memory(unsigned long pfn)
1143 unsigned int nr_pages;
1145 if (!pfn_valid(pfn))
1148 p = pfn_to_page(pfn);
1149 page = compound_head(p);
1151 if (!PageHWPoison(p)) {
1152 pr_debug("MCE: Page was already unpoisoned %#lx\n", pfn);
1156 nr_pages = 1 << compound_order(page);
1158 if (!get_page_unless_zero(page)) {
1159 if (TestClearPageHWPoison(p))
1160 atomic_long_sub(nr_pages, &mce_bad_pages);
1161 pr_debug("MCE: Software-unpoisoned free page %#lx\n", pfn);
1165 lock_page_nosync(page);
1167 * This test is racy because PG_hwpoison is set outside of page lock.
1168 * That's acceptable because that won't trigger kernel panic. Instead,
1169 * the PG_hwpoison page will be caught and isolated on the entrance to
1170 * the free buddy page pool.
1172 if (TestClearPageHWPoison(page)) {
1173 pr_debug("MCE: Software-unpoisoned page %#lx\n", pfn);
1174 atomic_long_sub(nr_pages, &mce_bad_pages);
1178 clear_page_hwpoison_huge_page(page);
1187 EXPORT_SYMBOL(unpoison_memory);
1189 static struct page *new_page(struct page *p, unsigned long private, int **x)
1191 int nid = page_to_nid(p);
1192 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1196 * Safely get reference count of an arbitrary page.
1197 * Returns 0 for a free page, -EIO for a zero refcount page
1198 * that is not free, and 1 for any other page type.
1199 * For 1 the page is returned with increased page count, otherwise not.
1201 static int get_any_page(struct page *p, unsigned long pfn, int flags)
1205 if (flags & MF_COUNT_INCREASED)
1209 * The lock_system_sleep prevents a race with memory hotplug,
1210 * because the isolation assumes there's only a single user.
1211 * This is a big hammer, a better would be nicer.
1213 lock_system_sleep();
1216 * Isolate the page, so that it doesn't get reallocated if it
1219 set_migratetype_isolate(p);
1220 if (!get_page_unless_zero(compound_head(p))) {
1221 if (is_free_buddy_page(p)) {
1222 pr_debug("get_any_page: %#lx free buddy page\n", pfn);
1223 /* Set hwpoison bit while page is still isolated */
1227 pr_debug("get_any_page: %#lx: unknown zero refcount page type %lx\n",
1232 /* Not a free page */
1235 unset_migratetype_isolate(p);
1236 unlock_system_sleep();
1241 * soft_offline_page - Soft offline a page.
1242 * @page: page to offline
1243 * @flags: flags. Same as memory_failure().
1245 * Returns 0 on success, otherwise negated errno.
1247 * Soft offline a page, by migration or invalidation,
1248 * without killing anything. This is for the case when
1249 * a page is not corrupted yet (so it's still valid to access),
1250 * but has had a number of corrected errors and is better taken
1253 * The actual policy on when to do that is maintained by
1256 * This should never impact any application or cause data loss,
1257 * however it might take some time.
1259 * This is not a 100% solution for all memory, but tries to be
1260 * ``good enough'' for the majority of memory.
1262 int soft_offline_page(struct page *page, int flags)
1265 unsigned long pfn = page_to_pfn(page);
1267 ret = get_any_page(page, pfn, flags);
1274 * Page cache page we can handle?
1276 if (!PageLRU(page)) {
1281 shake_page(page, 1);
1286 ret = get_any_page(page, pfn, 0);
1292 if (!PageLRU(page)) {
1293 pr_debug("soft_offline: %#lx: unknown non LRU page type %lx\n",
1299 wait_on_page_writeback(page);
1302 * Synchronized using the page lock with memory_failure()
1304 if (PageHWPoison(page)) {
1307 pr_debug("soft offline: %#lx page already poisoned\n", pfn);
1312 * Try to invalidate first. This should work for
1313 * non dirty unmapped page cache pages.
1315 ret = invalidate_inode_page(page);
1319 * Drop count because page migration doesn't like raised
1320 * counts. The page could get re-allocated, but if it becomes
1321 * LRU the isolation will just fail.
1322 * RED-PEN would be better to keep it isolated here, but we
1323 * would need to fix isolation locking first.
1328 pr_debug("soft_offline: %#lx: invalidated\n", pfn);
1333 * Simple invalidation didn't work.
1334 * Try to migrate to a new page instead. migrate.c
1335 * handles a large number of cases for us.
1337 ret = isolate_lru_page(page);
1339 LIST_HEAD(pagelist);
1341 list_add(&page->lru, &pagelist);
1342 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0);
1344 pr_debug("soft offline: %#lx: migration failed %d, type %lx\n",
1345 pfn, ret, page->flags);
1350 pr_debug("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1351 pfn, ret, page_count(page), page->flags);
1357 atomic_long_add(1, &mce_bad_pages);
1358 SetPageHWPoison(page);
1359 /* keep elevated page count for bad page */
1364 * The caller must hold current->mm->mmap_sem in read mode.
1366 int is_hwpoison_address(unsigned long addr)
1374 pgdp = pgd_offset(current->mm, addr);
1375 if (!pgd_present(*pgdp))
1377 pudp = pud_offset(pgdp, addr);
1379 if (!pud_present(pud) || pud_large(pud))
1381 pmdp = pmd_offset(pudp, addr);
1383 if (!pmd_present(pmd) || pmd_large(pmd))
1385 ptep = pte_offset_map(pmdp, addr);
1388 if (!is_swap_pte(pte))
1390 entry = pte_to_swp_entry(pte);
1391 return is_hwpoison_entry(entry);
1393 EXPORT_SYMBOL_GPL(is_hwpoison_address);