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 * Error hit kernel page.
354 * Do nothing, try to be lucky and not touch this instead. For a few cases we
355 * could be more sophisticated.
357 static int me_kernel(struct page
*p
, unsigned long pfn
)
363 * Already poisoned page.
365 static int me_ignore(struct page
*p
, unsigned long pfn
)
371 * Page in unknown state. Do nothing.
373 static int me_unknown(struct page
*p
, unsigned long pfn
)
375 printk(KERN_ERR
"MCE %#lx: Unknown page state\n", pfn
);
382 static int me_free(struct page
*p
, unsigned long pfn
)
388 * Clean (or cleaned) page cache page.
390 static int me_pagecache_clean(struct page
*p
, unsigned long pfn
)
394 struct address_space
*mapping
;
397 * For anonymous pages we're done the only reference left
398 * should be the one m_f() holds.
404 * Now truncate the page in the page cache. This is really
405 * more like a "temporary hole punch"
406 * Don't do this for block devices when someone else
407 * has a reference, because it could be file system metadata
408 * and that's not safe to truncate.
410 mapping
= page_mapping(p
);
413 * Page has been teared down in the meanwhile
419 * Truncation is a bit tricky. Enable it per file system for now.
421 * Open: to take i_mutex or not for this? Right now we don't.
423 if (mapping
->a_ops
->error_remove_page
) {
424 err
= mapping
->a_ops
->error_remove_page(mapping
, p
);
426 printk(KERN_INFO
"MCE %#lx: Failed to punch page: %d\n",
428 } else if (page_has_private(p
) &&
429 !try_to_release_page(p
, GFP_NOIO
)) {
430 pr_debug("MCE %#lx: failed to release buffers\n", pfn
);
436 * If the file system doesn't support it just invalidate
437 * This fails on dirty or anything with private pages
439 if (invalidate_inode_page(p
))
442 printk(KERN_INFO
"MCE %#lx: Failed to invalidate\n",
449 * Dirty cache page page
450 * Issues: when the error hit a hole page the error is not properly
453 static int me_pagecache_dirty(struct page
*p
, unsigned long pfn
)
455 struct address_space
*mapping
= page_mapping(p
);
458 /* TBD: print more information about the file. */
461 * IO error will be reported by write(), fsync(), etc.
462 * who check the mapping.
463 * This way the application knows that something went
464 * wrong with its dirty file data.
466 * There's one open issue:
468 * The EIO will be only reported on the next IO
469 * operation and then cleared through the IO map.
470 * Normally Linux has two mechanisms to pass IO error
471 * first through the AS_EIO flag in the address space
472 * and then through the PageError flag in the page.
473 * Since we drop pages on memory failure handling the
474 * only mechanism open to use is through AS_AIO.
476 * This has the disadvantage that it gets cleared on
477 * the first operation that returns an error, while
478 * the PageError bit is more sticky and only cleared
479 * when the page is reread or dropped. If an
480 * application assumes it will always get error on
481 * fsync, but does other operations on the fd before
482 * and the page is dropped inbetween then the error
483 * will not be properly reported.
485 * This can already happen even without hwpoisoned
486 * pages: first on metadata IO errors (which only
487 * report through AS_EIO) or when the page is dropped
490 * So right now we assume that the application DTRT on
491 * the first EIO, but we're not worse than other parts
494 mapping_set_error(mapping
, EIO
);
497 return me_pagecache_clean(p
, pfn
);
501 * Clean and dirty swap cache.
503 * Dirty swap cache page is tricky to handle. The page could live both in page
504 * cache and swap cache(ie. page is freshly swapped in). So it could be
505 * referenced concurrently by 2 types of PTEs:
506 * normal PTEs and swap PTEs. We try to handle them consistently by calling
507 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
509 * - clear dirty bit to prevent IO
511 * - but keep in the swap cache, so that when we return to it on
512 * a later page fault, we know the application is accessing
513 * corrupted data and shall be killed (we installed simple
514 * interception code in do_swap_page to catch it).
516 * Clean swap cache pages can be directly isolated. A later page fault will
517 * bring in the known good data from disk.
519 static int me_swapcache_dirty(struct page
*p
, unsigned long pfn
)
522 /* Trigger EIO in shmem: */
523 ClearPageUptodate(p
);
528 static int me_swapcache_clean(struct page
*p
, unsigned long pfn
)
530 delete_from_swap_cache(p
);
536 * Huge pages. Needs work.
538 * No rmap support so we cannot find the original mapper. In theory could walk
539 * all MMs and look for the mappings, but that would be non atomic and racy.
540 * Need rmap for hugepages for this. Alternatively we could employ a heuristic,
541 * like just walking the current process and hoping it has it mapped (that
542 * should be usually true for the common "shared database cache" case)
543 * Should handle free huge pages and dequeue them too, but this needs to
544 * handle huge page accounting correctly.
546 static int me_huge_page(struct page
*p
, unsigned long pfn
)
552 * Various page states we can handle.
554 * A page state is defined by its current page->flags bits.
555 * The table matches them in order and calls the right handler.
557 * This is quite tricky because we can access page at any time
558 * in its live cycle, so all accesses have to be extremly careful.
560 * This is not complete. More states could be added.
561 * For any missing state don't attempt recovery.
564 #define dirty (1UL << PG_dirty)
565 #define sc (1UL << PG_swapcache)
566 #define unevict (1UL << PG_unevictable)
567 #define mlock (1UL << PG_mlocked)
568 #define writeback (1UL << PG_writeback)
569 #define lru (1UL << PG_lru)
570 #define swapbacked (1UL << PG_swapbacked)
571 #define head (1UL << PG_head)
572 #define tail (1UL << PG_tail)
573 #define compound (1UL << PG_compound)
574 #define slab (1UL << PG_slab)
575 #define buddy (1UL << PG_buddy)
576 #define reserved (1UL << PG_reserved)
578 static struct page_state
{
582 int (*action
)(struct page
*p
, unsigned long pfn
);
584 { reserved
, reserved
, "reserved kernel", me_ignore
},
585 { buddy
, buddy
, "free kernel", me_free
},
588 * Could in theory check if slab page is free or if we can drop
589 * currently unused objects without touching them. But just
590 * treat it as standard kernel for now.
592 { slab
, slab
, "kernel slab", me_kernel
},
594 #ifdef CONFIG_PAGEFLAGS_EXTENDED
595 { head
, head
, "huge", me_huge_page
},
596 { tail
, tail
, "huge", me_huge_page
},
598 { compound
, compound
, "huge", me_huge_page
},
601 { sc
|dirty
, sc
|dirty
, "swapcache", me_swapcache_dirty
},
602 { sc
|dirty
, sc
, "swapcache", me_swapcache_clean
},
604 { unevict
|dirty
, unevict
|dirty
, "unevictable LRU", me_pagecache_dirty
},
605 { unevict
, unevict
, "unevictable LRU", me_pagecache_clean
},
607 { mlock
|dirty
, mlock
|dirty
, "mlocked LRU", me_pagecache_dirty
},
608 { mlock
, mlock
, "mlocked LRU", me_pagecache_clean
},
610 { lru
|dirty
, lru
|dirty
, "LRU", me_pagecache_dirty
},
611 { lru
|dirty
, lru
, "clean LRU", me_pagecache_clean
},
614 * Catchall entry: must be at end.
616 { 0, 0, "unknown page state", me_unknown
},
619 static void action_result(unsigned long pfn
, char *msg
, int result
)
621 struct page
*page
= pfn_to_page(pfn
);
623 printk(KERN_ERR
"MCE %#lx: %s%s page recovery: %s\n",
625 PageDirty(page
) ? "dirty " : "",
626 msg
, action_name
[result
]);
629 static int page_action(struct page_state
*ps
, struct page
*p
,
635 result
= ps
->action(p
, pfn
);
636 action_result(pfn
, ps
->msg
, result
);
638 count
= page_count(p
) - 1;
641 "MCE %#lx: %s page still referenced by %d users\n",
642 pfn
, ps
->msg
, count
);
644 /* Could do more checks here if page looks ok */
646 * Could adjust zone counters here to correct for the missing page.
649 return result
== RECOVERED
? 0 : -EBUSY
;
652 #define N_UNMAP_TRIES 5
655 * Do all that is necessary to remove user space mappings. Unmap
656 * the pages and send SIGBUS to the processes if the data was dirty.
658 static void hwpoison_user_mappings(struct page
*p
, unsigned long pfn
,
661 enum ttu_flags ttu
= TTU_UNMAP
| TTU_IGNORE_MLOCK
| TTU_IGNORE_ACCESS
;
662 struct address_space
*mapping
;
668 if (PageReserved(p
) || PageCompound(p
) || PageSlab(p
) || PageKsm(p
))
672 * This check implies we don't kill processes if their pages
673 * are in the swap cache early. Those are always late kills.
678 if (PageSwapCache(p
)) {
680 "MCE %#lx: keeping poisoned page in swap cache\n", pfn
);
681 ttu
|= TTU_IGNORE_HWPOISON
;
685 * Propagate the dirty bit from PTEs to struct page first, because we
686 * need this to decide if we should kill or just drop the page.
688 mapping
= page_mapping(p
);
689 if (!PageDirty(p
) && mapping
&& mapping_cap_writeback_dirty(mapping
)) {
690 if (page_mkclean(p
)) {
694 ttu
|= TTU_IGNORE_HWPOISON
;
696 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
702 * First collect all the processes that have the page
703 * mapped in dirty form. This has to be done before try_to_unmap,
704 * because ttu takes the rmap data structures down.
706 * Error handling: We ignore errors here because
707 * there's nothing that can be done.
710 collect_procs(p
, &tokill
);
713 * try_to_unmap can fail temporarily due to races.
714 * Try a few times (RED-PEN better strategy?)
716 for (i
= 0; i
< N_UNMAP_TRIES
; i
++) {
717 ret
= try_to_unmap(p
, ttu
);
718 if (ret
== SWAP_SUCCESS
)
720 pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn
, ret
);
723 if (ret
!= SWAP_SUCCESS
)
724 printk(KERN_ERR
"MCE %#lx: failed to unmap page (mapcount=%d)\n",
725 pfn
, page_mapcount(p
));
728 * Now that the dirty bit has been propagated to the
729 * struct page and all unmaps done we can decide if
730 * killing is needed or not. Only kill when the page
731 * was dirty, otherwise the tokill list is merely
732 * freed. When there was a problem unmapping earlier
733 * use a more force-full uncatchable kill to prevent
734 * any accesses to the poisoned memory.
736 kill_procs_ao(&tokill
, !!PageDirty(p
), trapno
,
737 ret
!= SWAP_SUCCESS
, pfn
);
740 int __memory_failure(unsigned long pfn
, int trapno
, int ref
)
742 unsigned long lru_flag
;
743 struct page_state
*ps
;
747 if (!sysctl_memory_failure_recovery
)
748 panic("Memory failure from trap %d on page %lx", trapno
, pfn
);
750 if (!pfn_valid(pfn
)) {
752 "MCE %#lx: memory outside kernel control\n",
757 p
= pfn_to_page(pfn
);
758 if (TestSetPageHWPoison(p
)) {
759 action_result(pfn
, "already hardware poisoned", IGNORED
);
763 atomic_long_add(1, &mce_bad_pages
);
766 * We need/can do nothing about count=0 pages.
767 * 1) it's a free page, and therefore in safe hand:
768 * prep_new_page() will be the gate keeper.
769 * 2) it's part of a non-compound high order page.
770 * Implies some kernel user: cannot stop them from
771 * R/W the page; let's pray that the page has been
772 * used and will be freed some time later.
773 * In fact it's dangerous to directly bump up page count from 0,
774 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
776 if (!ref
&& !get_page_unless_zero(compound_head(p
))) {
777 action_result(pfn
, "free or high order kernel", IGNORED
);
778 return PageBuddy(compound_head(p
)) ? 0 : -EBUSY
;
782 * We ignore non-LRU pages for good reasons.
783 * - PG_locked is only well defined for LRU pages and a few others
784 * - to avoid races with __set_page_locked()
785 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
786 * The check (unnecessarily) ignores LRU pages being isolated and
787 * walked by the page reclaim code, however that's not a big loss.
791 lru_flag
= p
->flags
& lru
;
792 if (isolate_lru_page(p
)) {
793 action_result(pfn
, "non LRU", IGNORED
);
797 page_cache_release(p
);
800 * Lock the page and wait for writeback to finish.
801 * It's very difficult to mess with pages currently under IO
802 * and in many cases impossible, so we just avoid it here.
805 wait_on_page_writeback(p
);
808 * Now take care of user space mappings.
810 hwpoison_user_mappings(p
, pfn
, trapno
);
813 * Torn down by someone else?
815 if ((lru_flag
& lru
) && !PageSwapCache(p
) && p
->mapping
== NULL
) {
816 action_result(pfn
, "already truncated LRU", IGNORED
);
822 for (ps
= error_states
;; ps
++) {
823 if (((p
->flags
| lru_flag
)& ps
->mask
) == ps
->res
) {
824 res
= page_action(ps
, p
, pfn
);
832 EXPORT_SYMBOL_GPL(__memory_failure
);
835 * memory_failure - Handle memory failure of a page.
836 * @pfn: Page Number of the corrupted page
837 * @trapno: Trap number reported in the signal to user space.
839 * This function is called by the low level machine check code
840 * of an architecture when it detects hardware memory corruption
841 * of a page. It tries its best to recover, which includes
842 * dropping pages, killing processes etc.
844 * The function is primarily of use for corruptions that
845 * happen outside the current execution context (e.g. when
846 * detected by a background scrubber)
848 * Must run in process context (e.g. a work queue) with interrupts
849 * enabled and no spinlocks hold.
851 void memory_failure(unsigned long pfn
, int trapno
)
853 __memory_failure(pfn
, trapno
, 0);