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/sched.h>
38 #include <linux/ksm.h>
39 #include <linux/rmap.h>
40 #include <linux/pagemap.h>
41 #include <linux/swap.h>
42 #include <linux/backing-dev.h>
45 int sysctl_memory_failure_early_kill __read_mostly = 0;
47 int sysctl_memory_failure_recovery __read_mostly = 1;
49 atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
52 * Send all the processes who have the page mapped an ``action optional''
55 static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
62 "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
63 pfn, t->comm, t->pid);
66 si.si_code = BUS_MCEERR_AO;
67 si.si_addr = (void *)addr;
68 #ifdef __ARCH_SI_TRAPNO
69 si.si_trapno = trapno;
71 si.si_addr_lsb = PAGE_SHIFT;
73 * Don't use force here, it's convenient if the signal
74 * can be temporarily blocked.
75 * This could cause a loop when the user sets SIGBUS
76 * to SIG_IGN, but hopefully noone will do that?
78 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
80 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
81 t->comm, t->pid, ret);
86 * When a unknown page type is encountered drain as many buffers as possible
87 * in the hope to turn the page into a LRU or free page, which we can handle.
89 void shake_page(struct page *p)
96 if (PageLRU(p) || is_free_buddy_page(p))
100 * Could call shrink_slab here (which would also
101 * shrink other caches). Unfortunately that might
102 * also access the corrupted page, which could be fatal.
105 EXPORT_SYMBOL_GPL(shake_page);
108 * Kill all processes that have a poisoned page mapped and then isolate
112 * Find all processes having the page mapped and kill them.
113 * But we keep a page reference around so that the page is not
114 * actually freed yet.
115 * Then stash the page away
117 * There's no convenient way to get back to mapped processes
118 * from the VMAs. So do a brute-force search over all
121 * Remember that machine checks are not common (or rather
122 * if they are common you have other problems), so this shouldn't
123 * be a performance issue.
125 * Also there are some races possible while we get from the
126 * error detection to actually handle it.
131 struct task_struct *tsk;
133 unsigned addr_valid:1;
137 * Failure handling: if we can't find or can't kill a process there's
138 * not much we can do. We just print a message and ignore otherwise.
142 * Schedule a process for later kill.
143 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
144 * TBD would GFP_NOIO be enough?
146 static void add_to_kill(struct task_struct *tsk, struct page *p,
147 struct vm_area_struct *vma,
148 struct list_head *to_kill,
149 struct to_kill **tkc)
157 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
160 "MCE: Out of memory while machine check handling\n");
164 tk->addr = page_address_in_vma(p, vma);
168 * In theory we don't have to kill when the page was
169 * munmaped. But it could be also a mremap. Since that's
170 * likely very rare kill anyways just out of paranoia, but use
171 * a SIGKILL because the error is not contained anymore.
173 if (tk->addr == -EFAULT) {
174 pr_debug("MCE: Unable to find user space address %lx in %s\n",
175 page_to_pfn(p), tsk->comm);
178 get_task_struct(tsk);
180 list_add_tail(&tk->nd, to_kill);
184 * Kill the processes that have been collected earlier.
186 * Only do anything when DOIT is set, otherwise just free the list
187 * (this is used for clean pages which do not need killing)
188 * Also when FAIL is set do a force kill because something went
191 static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
192 int fail, unsigned long pfn)
194 struct to_kill *tk, *next;
196 list_for_each_entry_safe (tk, next, to_kill, nd) {
199 * In case something went wrong with munmapping
200 * make sure the process doesn't catch the
201 * signal and then access the memory. Just kill it.
202 * the signal handlers
204 if (fail || tk->addr_valid == 0) {
206 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
207 pfn, tk->tsk->comm, tk->tsk->pid);
208 force_sig(SIGKILL, tk->tsk);
212 * In theory the process could have mapped
213 * something else on the address in-between. We could
214 * check for that, but we need to tell the
217 else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
220 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
221 pfn, tk->tsk->comm, tk->tsk->pid);
223 put_task_struct(tk->tsk);
228 static int task_early_kill(struct task_struct *tsk)
232 if (tsk->flags & PF_MCE_PROCESS)
233 return !!(tsk->flags & PF_MCE_EARLY);
234 return sysctl_memory_failure_early_kill;
238 * Collect processes when the error hit an anonymous page.
240 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
241 struct to_kill **tkc)
243 struct vm_area_struct *vma;
244 struct task_struct *tsk;
247 read_lock(&tasklist_lock);
248 av = page_lock_anon_vma(page);
249 if (av == NULL) /* Not actually mapped anymore */
251 for_each_process (tsk) {
252 if (!task_early_kill(tsk))
254 list_for_each_entry (vma, &av->head, anon_vma_node) {
255 if (!page_mapped_in_vma(page, vma))
257 if (vma->vm_mm == tsk->mm)
258 add_to_kill(tsk, page, vma, to_kill, tkc);
261 page_unlock_anon_vma(av);
263 read_unlock(&tasklist_lock);
267 * Collect processes when the error hit a file mapped page.
269 static void collect_procs_file(struct page *page, struct list_head *to_kill,
270 struct to_kill **tkc)
272 struct vm_area_struct *vma;
273 struct task_struct *tsk;
274 struct prio_tree_iter iter;
275 struct address_space *mapping = page->mapping;
278 * A note on the locking order between the two locks.
279 * We don't rely on this particular order.
280 * If you have some other code that needs a different order
281 * feel free to switch them around. Or add a reverse link
282 * from mm_struct to task_struct, then this could be all
283 * done without taking tasklist_lock and looping over all tasks.
286 read_lock(&tasklist_lock);
287 spin_lock(&mapping->i_mmap_lock);
288 for_each_process(tsk) {
289 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
291 if (!task_early_kill(tsk))
294 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
297 * Send early kill signal to tasks where a vma covers
298 * the page but the corrupted page is not necessarily
299 * mapped it in its pte.
300 * Assume applications who requested early kill want
301 * to be informed of all such data corruptions.
303 if (vma->vm_mm == tsk->mm)
304 add_to_kill(tsk, page, vma, to_kill, tkc);
307 spin_unlock(&mapping->i_mmap_lock);
308 read_unlock(&tasklist_lock);
312 * Collect the processes who have the corrupted page mapped to kill.
313 * This is done in two steps for locking reasons.
314 * First preallocate one tokill structure outside the spin locks,
315 * so that we can kill at least one process reasonably reliable.
317 static void collect_procs(struct page *page, struct list_head *tokill)
324 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
328 collect_procs_anon(page, tokill, &tk);
330 collect_procs_file(page, tokill, &tk);
335 * Error handlers for various types of pages.
339 FAILED, /* Error handling failed */
340 DELAYED, /* Will be handled later */
341 IGNORED, /* Error safely ignored */
342 RECOVERED, /* Successfully recovered */
345 static const char *action_name[] = {
347 [DELAYED] = "Delayed",
348 [IGNORED] = "Ignored",
349 [RECOVERED] = "Recovered",
353 * XXX: It is possible that a page is isolated from LRU cache,
354 * and then kept in swap cache or failed to remove from page cache.
355 * The page count will stop it from being freed by unpoison.
356 * Stress tests should be aware of this memory leak problem.
358 static int delete_from_lru_cache(struct page *p)
360 if (!isolate_lru_page(p)) {
362 * Clear sensible page flags, so that the buddy system won't
363 * complain when the page is unpoison-and-freed.
366 ClearPageUnevictable(p);
368 * drop the page count elevated by isolate_lru_page()
370 page_cache_release(p);
377 * Error hit kernel page.
378 * Do nothing, try to be lucky and not touch this instead. For a few cases we
379 * could be more sophisticated.
381 static int me_kernel(struct page *p, unsigned long pfn)
387 * Already poisoned page.
389 static int me_ignore(struct page *p, unsigned long pfn)
395 * Page in unknown state. Do nothing.
397 static int me_unknown(struct page *p, unsigned long pfn)
399 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
404 * Clean (or cleaned) page cache page.
406 static int me_pagecache_clean(struct page *p, unsigned long pfn)
410 struct address_space *mapping;
412 delete_from_lru_cache(p);
415 * For anonymous pages we're done the only reference left
416 * should be the one m_f() holds.
422 * Now truncate the page in the page cache. This is really
423 * more like a "temporary hole punch"
424 * Don't do this for block devices when someone else
425 * has a reference, because it could be file system metadata
426 * and that's not safe to truncate.
428 mapping = page_mapping(p);
431 * Page has been teared down in the meanwhile
437 * Truncation is a bit tricky. Enable it per file system for now.
439 * Open: to take i_mutex or not for this? Right now we don't.
441 if (mapping->a_ops->error_remove_page) {
442 err = mapping->a_ops->error_remove_page(mapping, p);
444 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
446 } else if (page_has_private(p) &&
447 !try_to_release_page(p, GFP_NOIO)) {
448 pr_debug("MCE %#lx: failed to release buffers\n", pfn);
454 * If the file system doesn't support it just invalidate
455 * This fails on dirty or anything with private pages
457 if (invalidate_inode_page(p))
460 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
467 * Dirty cache page page
468 * Issues: when the error hit a hole page the error is not properly
471 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
473 struct address_space *mapping = page_mapping(p);
476 /* TBD: print more information about the file. */
479 * IO error will be reported by write(), fsync(), etc.
480 * who check the mapping.
481 * This way the application knows that something went
482 * wrong with its dirty file data.
484 * There's one open issue:
486 * The EIO will be only reported on the next IO
487 * operation and then cleared through the IO map.
488 * Normally Linux has two mechanisms to pass IO error
489 * first through the AS_EIO flag in the address space
490 * and then through the PageError flag in the page.
491 * Since we drop pages on memory failure handling the
492 * only mechanism open to use is through AS_AIO.
494 * This has the disadvantage that it gets cleared on
495 * the first operation that returns an error, while
496 * the PageError bit is more sticky and only cleared
497 * when the page is reread or dropped. If an
498 * application assumes it will always get error on
499 * fsync, but does other operations on the fd before
500 * and the page is dropped inbetween then the error
501 * will not be properly reported.
503 * This can already happen even without hwpoisoned
504 * pages: first on metadata IO errors (which only
505 * report through AS_EIO) or when the page is dropped
508 * So right now we assume that the application DTRT on
509 * the first EIO, but we're not worse than other parts
512 mapping_set_error(mapping, EIO);
515 return me_pagecache_clean(p, pfn);
519 * Clean and dirty swap cache.
521 * Dirty swap cache page is tricky to handle. The page could live both in page
522 * cache and swap cache(ie. page is freshly swapped in). So it could be
523 * referenced concurrently by 2 types of PTEs:
524 * normal PTEs and swap PTEs. We try to handle them consistently by calling
525 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
527 * - clear dirty bit to prevent IO
529 * - but keep in the swap cache, so that when we return to it on
530 * a later page fault, we know the application is accessing
531 * corrupted data and shall be killed (we installed simple
532 * interception code in do_swap_page to catch it).
534 * Clean swap cache pages can be directly isolated. A later page fault will
535 * bring in the known good data from disk.
537 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
540 /* Trigger EIO in shmem: */
541 ClearPageUptodate(p);
543 if (!delete_from_lru_cache(p))
549 static int me_swapcache_clean(struct page *p, unsigned long pfn)
551 delete_from_swap_cache(p);
553 if (!delete_from_lru_cache(p))
560 * Huge pages. Needs work.
562 * No rmap support so we cannot find the original mapper. In theory could walk
563 * all MMs and look for the mappings, but that would be non atomic and racy.
564 * Need rmap for hugepages for this. Alternatively we could employ a heuristic,
565 * like just walking the current process and hoping it has it mapped (that
566 * should be usually true for the common "shared database cache" case)
567 * Should handle free huge pages and dequeue them too, but this needs to
568 * handle huge page accounting correctly.
570 static int me_huge_page(struct page *p, unsigned long pfn)
576 * Various page states we can handle.
578 * A page state is defined by its current page->flags bits.
579 * The table matches them in order and calls the right handler.
581 * This is quite tricky because we can access page at any time
582 * in its live cycle, so all accesses have to be extremly careful.
584 * This is not complete. More states could be added.
585 * For any missing state don't attempt recovery.
588 #define dirty (1UL << PG_dirty)
589 #define sc (1UL << PG_swapcache)
590 #define unevict (1UL << PG_unevictable)
591 #define mlock (1UL << PG_mlocked)
592 #define writeback (1UL << PG_writeback)
593 #define lru (1UL << PG_lru)
594 #define swapbacked (1UL << PG_swapbacked)
595 #define head (1UL << PG_head)
596 #define tail (1UL << PG_tail)
597 #define compound (1UL << PG_compound)
598 #define slab (1UL << PG_slab)
599 #define reserved (1UL << PG_reserved)
601 static struct page_state {
605 int (*action)(struct page *p, unsigned long pfn);
607 { reserved, reserved, "reserved kernel", me_ignore },
609 * free pages are specially detected outside this table:
610 * PG_buddy pages only make a small fraction of all free pages.
614 * Could in theory check if slab page is free or if we can drop
615 * currently unused objects without touching them. But just
616 * treat it as standard kernel for now.
618 { slab, slab, "kernel slab", me_kernel },
620 #ifdef CONFIG_PAGEFLAGS_EXTENDED
621 { head, head, "huge", me_huge_page },
622 { tail, tail, "huge", me_huge_page },
624 { compound, compound, "huge", me_huge_page },
627 { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
628 { sc|dirty, sc, "swapcache", me_swapcache_clean },
630 { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
631 { unevict, unevict, "unevictable LRU", me_pagecache_clean},
633 { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
634 { mlock, mlock, "mlocked LRU", me_pagecache_clean },
636 { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
637 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
640 * Catchall entry: must be at end.
642 { 0, 0, "unknown page state", me_unknown },
645 static void action_result(unsigned long pfn, char *msg, int result)
647 struct page *page = pfn_to_page(pfn);
649 printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
651 PageDirty(page) ? "dirty " : "",
652 msg, action_name[result]);
655 static int page_action(struct page_state *ps, struct page *p,
661 result = ps->action(p, pfn);
662 action_result(pfn, ps->msg, result);
664 count = page_count(p) - 1;
667 "MCE %#lx: %s page still referenced by %d users\n",
668 pfn, ps->msg, count);
670 /* Could do more checks here if page looks ok */
672 * Could adjust zone counters here to correct for the missing page.
675 return result == RECOVERED ? 0 : -EBUSY;
678 #define N_UNMAP_TRIES 5
681 * Do all that is necessary to remove user space mappings. Unmap
682 * the pages and send SIGBUS to the processes if the data was dirty.
684 static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
687 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
688 struct address_space *mapping;
694 if (PageReserved(p) || PageSlab(p))
698 * This check implies we don't kill processes if their pages
699 * are in the swap cache early. Those are always late kills.
704 if (PageCompound(p) || PageKsm(p))
707 if (PageSwapCache(p)) {
709 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
710 ttu |= TTU_IGNORE_HWPOISON;
714 * Propagate the dirty bit from PTEs to struct page first, because we
715 * need this to decide if we should kill or just drop the page.
716 * XXX: the dirty test could be racy: set_page_dirty() may not always
717 * be called inside page lock (it's recommended but not enforced).
719 mapping = page_mapping(p);
720 if (!PageDirty(p) && mapping && mapping_cap_writeback_dirty(mapping)) {
721 if (page_mkclean(p)) {
725 ttu |= TTU_IGNORE_HWPOISON;
727 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
733 * First collect all the processes that have the page
734 * mapped in dirty form. This has to be done before try_to_unmap,
735 * because ttu takes the rmap data structures down.
737 * Error handling: We ignore errors here because
738 * there's nothing that can be done.
741 collect_procs(p, &tokill);
744 * try_to_unmap can fail temporarily due to races.
745 * Try a few times (RED-PEN better strategy?)
747 for (i = 0; i < N_UNMAP_TRIES; i++) {
748 ret = try_to_unmap(p, ttu);
749 if (ret == SWAP_SUCCESS)
751 pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn, ret);
754 if (ret != SWAP_SUCCESS)
755 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
756 pfn, page_mapcount(p));
759 * Now that the dirty bit has been propagated to the
760 * struct page and all unmaps done we can decide if
761 * killing is needed or not. Only kill when the page
762 * was dirty, otherwise the tokill list is merely
763 * freed. When there was a problem unmapping earlier
764 * use a more force-full uncatchable kill to prevent
765 * any accesses to the poisoned memory.
767 kill_procs_ao(&tokill, !!PageDirty(p), trapno,
768 ret != SWAP_SUCCESS, pfn);
773 int __memory_failure(unsigned long pfn, int trapno, int flags)
775 struct page_state *ps;
779 if (!sysctl_memory_failure_recovery)
780 panic("Memory failure from trap %d on page %lx", trapno, pfn);
782 if (!pfn_valid(pfn)) {
784 "MCE %#lx: memory outside kernel control\n",
789 p = pfn_to_page(pfn);
790 if (TestSetPageHWPoison(p)) {
791 action_result(pfn, "already hardware poisoned", IGNORED);
795 atomic_long_add(1, &mce_bad_pages);
798 * We need/can do nothing about count=0 pages.
799 * 1) it's a free page, and therefore in safe hand:
800 * prep_new_page() will be the gate keeper.
801 * 2) it's part of a non-compound high order page.
802 * Implies some kernel user: cannot stop them from
803 * R/W the page; let's pray that the page has been
804 * used and will be freed some time later.
805 * In fact it's dangerous to directly bump up page count from 0,
806 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
808 if (!(flags & MF_COUNT_INCREASED) &&
809 !get_page_unless_zero(compound_head(p))) {
810 if (is_free_buddy_page(p)) {
811 action_result(pfn, "free buddy", DELAYED);
814 action_result(pfn, "high order kernel", IGNORED);
820 * We ignore non-LRU pages for good reasons.
821 * - PG_locked is only well defined for LRU pages and a few others
822 * - to avoid races with __set_page_locked()
823 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
824 * The check (unnecessarily) ignores LRU pages being isolated and
825 * walked by the page reclaim code, however that's not a big loss.
830 action_result(pfn, "non LRU", IGNORED);
836 * Lock the page and wait for writeback to finish.
837 * It's very difficult to mess with pages currently under IO
838 * and in many cases impossible, so we just avoid it here.
841 wait_on_page_writeback(p);
844 * Now take care of user space mappings.
845 * Abort on fail: __remove_from_page_cache() assumes unmapped page.
847 if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
848 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
854 * Torn down by someone else?
856 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
857 action_result(pfn, "already truncated LRU", IGNORED);
863 for (ps = error_states;; ps++) {
864 if ((p->flags & ps->mask) == ps->res) {
865 res = page_action(ps, p, pfn);
873 EXPORT_SYMBOL_GPL(__memory_failure);
876 * memory_failure - Handle memory failure of a page.
877 * @pfn: Page Number of the corrupted page
878 * @trapno: Trap number reported in the signal to user space.
880 * This function is called by the low level machine check code
881 * of an architecture when it detects hardware memory corruption
882 * of a page. It tries its best to recover, which includes
883 * dropping pages, killing processes etc.
885 * The function is primarily of use for corruptions that
886 * happen outside the current execution context (e.g. when
887 * detected by a background scrubber)
889 * Must run in process context (e.g. a work queue) with interrupts
890 * enabled and no spinlocks hold.
892 void memory_failure(unsigned long pfn, int trapno)
894 __memory_failure(pfn, trapno, 0);