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/rmap.h>
39 #include <linux/pagemap.h>
40 #include <linux/swap.h>
41 #include <linux/backing-dev.h>
44 int sysctl_memory_failure_early_kill __read_mostly
= 0;
46 int sysctl_memory_failure_recovery __read_mostly
= 1;
48 atomic_long_t mce_bad_pages __read_mostly
= ATOMIC_LONG_INIT(0);
51 * Send all the processes who have the page mapped an ``action optional''
54 static int kill_proc_ao(struct task_struct
*t
, unsigned long addr
, int trapno
,
61 "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
62 pfn
, t
->comm
, t
->pid
);
65 si
.si_code
= BUS_MCEERR_AO
;
66 si
.si_addr
= (void *)addr
;
67 #ifdef __ARCH_SI_TRAPNO
68 si
.si_trapno
= trapno
;
70 si
.si_addr_lsb
= PAGE_SHIFT
;
72 * Don't use force here, it's convenient if the signal
73 * can be temporarily blocked.
74 * This could cause a loop when the user sets SIGBUS
75 * to SIG_IGN, but hopefully noone will do that?
77 ret
= send_sig_info(SIGBUS
, &si
, t
); /* synchronous? */
79 printk(KERN_INFO
"MCE: Error sending signal to %s:%d: %d\n",
80 t
->comm
, t
->pid
, ret
);
85 * Kill all processes that have a poisoned page mapped and then isolate
89 * Find all processes having the page mapped and kill them.
90 * But we keep a page reference around so that the page is not
92 * Then stash the page away
94 * There's no convenient way to get back to mapped processes
95 * from the VMAs. So do a brute-force search over all
98 * Remember that machine checks are not common (or rather
99 * if they are common you have other problems), so this shouldn't
100 * be a performance issue.
102 * Also there are some races possible while we get from the
103 * error detection to actually handle it.
108 struct task_struct
*tsk
;
110 unsigned addr_valid
:1;
114 * Failure handling: if we can't find or can't kill a process there's
115 * not much we can do. We just print a message and ignore otherwise.
119 * Schedule a process for later kill.
120 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
121 * TBD would GFP_NOIO be enough?
123 static void add_to_kill(struct task_struct
*tsk
, struct page
*p
,
124 struct vm_area_struct
*vma
,
125 struct list_head
*to_kill
,
126 struct to_kill
**tkc
)
134 tk
= kmalloc(sizeof(struct to_kill
), GFP_ATOMIC
);
137 "MCE: Out of memory while machine check handling\n");
141 tk
->addr
= page_address_in_vma(p
, vma
);
145 * In theory we don't have to kill when the page was
146 * munmaped. But it could be also a mremap. Since that's
147 * likely very rare kill anyways just out of paranoia, but use
148 * a SIGKILL because the error is not contained anymore.
150 if (tk
->addr
== -EFAULT
) {
151 pr_debug("MCE: Unable to find user space address %lx in %s\n",
152 page_to_pfn(p
), tsk
->comm
);
155 get_task_struct(tsk
);
157 list_add_tail(&tk
->nd
, to_kill
);
161 * Kill the processes that have been collected earlier.
163 * Only do anything when DOIT is set, otherwise just free the list
164 * (this is used for clean pages which do not need killing)
165 * Also when FAIL is set do a force kill because something went
168 static void kill_procs_ao(struct list_head
*to_kill
, int doit
, int trapno
,
169 int fail
, unsigned long pfn
)
171 struct to_kill
*tk
, *next
;
173 list_for_each_entry_safe (tk
, next
, to_kill
, nd
) {
176 * In case something went wrong with munmaping
177 * make sure the process doesn't catch the
178 * signal and then access the memory. Just kill it.
179 * the signal handlers
181 if (fail
|| tk
->addr_valid
== 0) {
183 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
184 pfn
, tk
->tsk
->comm
, tk
->tsk
->pid
);
185 force_sig(SIGKILL
, tk
->tsk
);
189 * In theory the process could have mapped
190 * something else on the address in-between. We could
191 * check for that, but we need to tell the
194 else if (kill_proc_ao(tk
->tsk
, tk
->addr
, trapno
,
197 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
198 pfn
, tk
->tsk
->comm
, tk
->tsk
->pid
);
200 put_task_struct(tk
->tsk
);
205 static int task_early_kill(struct task_struct
*tsk
)
209 if (tsk
->flags
& PF_MCE_PROCESS
)
210 return !!(tsk
->flags
& PF_MCE_EARLY
);
211 return sysctl_memory_failure_early_kill
;
215 * Collect processes when the error hit an anonymous page.
217 static void collect_procs_anon(struct page
*page
, struct list_head
*to_kill
,
218 struct to_kill
**tkc
)
220 struct vm_area_struct
*vma
;
221 struct task_struct
*tsk
;
224 read_lock(&tasklist_lock
);
225 av
= page_lock_anon_vma(page
);
226 if (av
== NULL
) /* Not actually mapped anymore */
228 for_each_process (tsk
) {
229 if (!task_early_kill(tsk
))
231 list_for_each_entry (vma
, &av
->head
, anon_vma_node
) {
232 if (!page_mapped_in_vma(page
, vma
))
234 if (vma
->vm_mm
== tsk
->mm
)
235 add_to_kill(tsk
, page
, vma
, to_kill
, tkc
);
238 page_unlock_anon_vma(av
);
240 read_unlock(&tasklist_lock
);
244 * Collect processes when the error hit a file mapped page.
246 static void collect_procs_file(struct page
*page
, struct list_head
*to_kill
,
247 struct to_kill
**tkc
)
249 struct vm_area_struct
*vma
;
250 struct task_struct
*tsk
;
251 struct prio_tree_iter iter
;
252 struct address_space
*mapping
= page
->mapping
;
255 * A note on the locking order between the two locks.
256 * We don't rely on this particular order.
257 * If you have some other code that needs a different order
258 * feel free to switch them around. Or add a reverse link
259 * from mm_struct to task_struct, then this could be all
260 * done without taking tasklist_lock and looping over all tasks.
263 read_lock(&tasklist_lock
);
264 spin_lock(&mapping
->i_mmap_lock
);
265 for_each_process(tsk
) {
266 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
268 if (!task_early_kill(tsk
))
271 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
,
274 * Send early kill signal to tasks where a vma covers
275 * the page but the corrupted page is not necessarily
276 * mapped it in its pte.
277 * Assume applications who requested early kill want
278 * to be informed of all such data corruptions.
280 if (vma
->vm_mm
== tsk
->mm
)
281 add_to_kill(tsk
, page
, vma
, to_kill
, tkc
);
284 spin_unlock(&mapping
->i_mmap_lock
);
285 read_unlock(&tasklist_lock
);
289 * Collect the processes who have the corrupted page mapped to kill.
290 * This is done in two steps for locking reasons.
291 * First preallocate one tokill structure outside the spin locks,
292 * so that we can kill at least one process reasonably reliable.
294 static void collect_procs(struct page
*page
, struct list_head
*tokill
)
301 tk
= kmalloc(sizeof(struct to_kill
), GFP_NOIO
);
305 collect_procs_anon(page
, tokill
, &tk
);
307 collect_procs_file(page
, tokill
, &tk
);
312 * Error handlers for various types of pages.
316 FAILED
, /* Error handling failed */
317 DELAYED
, /* Will be handled later */
318 IGNORED
, /* Error safely ignored */
319 RECOVERED
, /* Successfully recovered */
322 static const char *action_name
[] = {
324 [DELAYED
] = "Delayed",
325 [IGNORED
] = "Ignored",
326 [RECOVERED
] = "Recovered",
330 * Error hit kernel page.
331 * Do nothing, try to be lucky and not touch this instead. For a few cases we
332 * could be more sophisticated.
334 static int me_kernel(struct page
*p
, unsigned long pfn
)
340 * Already poisoned page.
342 static int me_ignore(struct page
*p
, unsigned long pfn
)
348 * Page in unknown state. Do nothing.
350 static int me_unknown(struct page
*p
, unsigned long pfn
)
352 printk(KERN_ERR
"MCE %#lx: Unknown page state\n", pfn
);
359 static int me_free(struct page
*p
, unsigned long pfn
)
365 * Clean (or cleaned) page cache page.
367 static int me_pagecache_clean(struct page
*p
, unsigned long pfn
)
371 struct address_space
*mapping
;
374 * For anonymous pages we're done the only reference left
375 * should be the one m_f() holds.
381 * Now truncate the page in the page cache. This is really
382 * more like a "temporary hole punch"
383 * Don't do this for block devices when someone else
384 * has a reference, because it could be file system metadata
385 * and that's not safe to truncate.
387 mapping
= page_mapping(p
);
390 * Page has been teared down in the meanwhile
396 * Truncation is a bit tricky. Enable it per file system for now.
398 * Open: to take i_mutex or not for this? Right now we don't.
400 if (mapping
->a_ops
->error_remove_page
) {
401 err
= mapping
->a_ops
->error_remove_page(mapping
, p
);
403 printk(KERN_INFO
"MCE %#lx: Failed to punch page: %d\n",
405 } else if (page_has_private(p
) &&
406 !try_to_release_page(p
, GFP_NOIO
)) {
407 pr_debug("MCE %#lx: failed to release buffers\n", pfn
);
413 * If the file system doesn't support it just invalidate
414 * This fails on dirty or anything with private pages
416 if (invalidate_inode_page(p
))
419 printk(KERN_INFO
"MCE %#lx: Failed to invalidate\n",
426 * Dirty cache page page
427 * Issues: when the error hit a hole page the error is not properly
430 static int me_pagecache_dirty(struct page
*p
, unsigned long pfn
)
432 struct address_space
*mapping
= page_mapping(p
);
435 /* TBD: print more information about the file. */
438 * IO error will be reported by write(), fsync(), etc.
439 * who check the mapping.
440 * This way the application knows that something went
441 * wrong with its dirty file data.
443 * There's one open issue:
445 * The EIO will be only reported on the next IO
446 * operation and then cleared through the IO map.
447 * Normally Linux has two mechanisms to pass IO error
448 * first through the AS_EIO flag in the address space
449 * and then through the PageError flag in the page.
450 * Since we drop pages on memory failure handling the
451 * only mechanism open to use is through AS_AIO.
453 * This has the disadvantage that it gets cleared on
454 * the first operation that returns an error, while
455 * the PageError bit is more sticky and only cleared
456 * when the page is reread or dropped. If an
457 * application assumes it will always get error on
458 * fsync, but does other operations on the fd before
459 * and the page is dropped inbetween then the error
460 * will not be properly reported.
462 * This can already happen even without hwpoisoned
463 * pages: first on metadata IO errors (which only
464 * report through AS_EIO) or when the page is dropped
467 * So right now we assume that the application DTRT on
468 * the first EIO, but we're not worse than other parts
471 mapping_set_error(mapping
, EIO
);
474 return me_pagecache_clean(p
, pfn
);
478 * Clean and dirty swap cache.
480 * Dirty swap cache page is tricky to handle. The page could live both in page
481 * cache and swap cache(ie. page is freshly swapped in). So it could be
482 * referenced concurrently by 2 types of PTEs:
483 * normal PTEs and swap PTEs. We try to handle them consistently by calling
484 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
486 * - clear dirty bit to prevent IO
488 * - but keep in the swap cache, so that when we return to it on
489 * a later page fault, we know the application is accessing
490 * corrupted data and shall be killed (we installed simple
491 * interception code in do_swap_page to catch it).
493 * Clean swap cache pages can be directly isolated. A later page fault will
494 * bring in the known good data from disk.
496 static int me_swapcache_dirty(struct page
*p
, unsigned long pfn
)
499 /* Trigger EIO in shmem: */
500 ClearPageUptodate(p
);
505 static int me_swapcache_clean(struct page
*p
, unsigned long pfn
)
507 delete_from_swap_cache(p
);
513 * Huge pages. Needs work.
515 * No rmap support so we cannot find the original mapper. In theory could walk
516 * all MMs and look for the mappings, but that would be non atomic and racy.
517 * Need rmap for hugepages for this. Alternatively we could employ a heuristic,
518 * like just walking the current process and hoping it has it mapped (that
519 * should be usually true for the common "shared database cache" case)
520 * Should handle free huge pages and dequeue them too, but this needs to
521 * handle huge page accounting correctly.
523 static int me_huge_page(struct page
*p
, unsigned long pfn
)
529 * Various page states we can handle.
531 * A page state is defined by its current page->flags bits.
532 * The table matches them in order and calls the right handler.
534 * This is quite tricky because we can access page at any time
535 * in its live cycle, so all accesses have to be extremly careful.
537 * This is not complete. More states could be added.
538 * For any missing state don't attempt recovery.
541 #define dirty (1UL << PG_dirty)
542 #define sc (1UL << PG_swapcache)
543 #define unevict (1UL << PG_unevictable)
544 #define mlock (1UL << PG_mlocked)
545 #define writeback (1UL << PG_writeback)
546 #define lru (1UL << PG_lru)
547 #define swapbacked (1UL << PG_swapbacked)
548 #define head (1UL << PG_head)
549 #define tail (1UL << PG_tail)
550 #define compound (1UL << PG_compound)
551 #define slab (1UL << PG_slab)
552 #define buddy (1UL << PG_buddy)
553 #define reserved (1UL << PG_reserved)
555 static struct page_state
{
559 int (*action
)(struct page
*p
, unsigned long pfn
);
561 { reserved
, reserved
, "reserved kernel", me_ignore
},
562 { buddy
, buddy
, "free kernel", me_free
},
565 * Could in theory check if slab page is free or if we can drop
566 * currently unused objects without touching them. But just
567 * treat it as standard kernel for now.
569 { slab
, slab
, "kernel slab", me_kernel
},
571 #ifdef CONFIG_PAGEFLAGS_EXTENDED
572 { head
, head
, "huge", me_huge_page
},
573 { tail
, tail
, "huge", me_huge_page
},
575 { compound
, compound
, "huge", me_huge_page
},
578 { sc
|dirty
, sc
|dirty
, "swapcache", me_swapcache_dirty
},
579 { sc
|dirty
, sc
, "swapcache", me_swapcache_clean
},
581 { unevict
|dirty
, unevict
|dirty
, "unevictable LRU", me_pagecache_dirty
},
582 { unevict
, unevict
, "unevictable LRU", me_pagecache_clean
},
584 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
585 { mlock
|dirty
, mlock
|dirty
, "mlocked LRU", me_pagecache_dirty
},
586 { mlock
, mlock
, "mlocked LRU", me_pagecache_clean
},
589 { lru
|dirty
, lru
|dirty
, "LRU", me_pagecache_dirty
},
590 { lru
|dirty
, lru
, "clean LRU", me_pagecache_clean
},
591 { swapbacked
, swapbacked
, "anonymous", me_pagecache_clean
},
594 * Catchall entry: must be at end.
596 { 0, 0, "unknown page state", me_unknown
},
599 static void action_result(unsigned long pfn
, char *msg
, int result
)
601 struct page
*page
= NULL
;
603 page
= pfn_to_page(pfn
);
605 printk(KERN_ERR
"MCE %#lx: %s%s page recovery: %s\n",
607 page
&& PageDirty(page
) ? "dirty " : "",
608 msg
, action_name
[result
]);
611 static int page_action(struct page_state
*ps
, struct page
*p
,
612 unsigned long pfn
, int ref
)
616 result
= ps
->action(p
, pfn
);
617 action_result(pfn
, ps
->msg
, result
);
618 if (page_count(p
) != 1 + ref
)
620 "MCE %#lx: %s page still referenced by %d users\n",
621 pfn
, ps
->msg
, page_count(p
) - 1);
623 /* Could do more checks here if page looks ok */
625 * Could adjust zone counters here to correct for the missing page.
628 return result
== RECOVERED
? 0 : -EBUSY
;
631 #define N_UNMAP_TRIES 5
634 * Do all that is necessary to remove user space mappings. Unmap
635 * the pages and send SIGBUS to the processes if the data was dirty.
637 static void hwpoison_user_mappings(struct page
*p
, unsigned long pfn
,
640 enum ttu_flags ttu
= TTU_UNMAP
| TTU_IGNORE_MLOCK
| TTU_IGNORE_ACCESS
;
641 struct address_space
*mapping
;
647 if (PageReserved(p
) || PageCompound(p
) || PageSlab(p
))
651 * This check implies we don't kill processes if their pages
652 * are in the swap cache early. Those are always late kills.
657 if (PageSwapCache(p
)) {
659 "MCE %#lx: keeping poisoned page in swap cache\n", pfn
);
660 ttu
|= TTU_IGNORE_HWPOISON
;
664 * Propagate the dirty bit from PTEs to struct page first, because we
665 * need this to decide if we should kill or just drop the page.
667 mapping
= page_mapping(p
);
668 if (!PageDirty(p
) && mapping
&& mapping_cap_writeback_dirty(mapping
)) {
669 if (page_mkclean(p
)) {
673 ttu
|= TTU_IGNORE_HWPOISON
;
675 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
681 * First collect all the processes that have the page
682 * mapped in dirty form. This has to be done before try_to_unmap,
683 * because ttu takes the rmap data structures down.
685 * Error handling: We ignore errors here because
686 * there's nothing that can be done.
689 collect_procs(p
, &tokill
);
692 * try_to_unmap can fail temporarily due to races.
693 * Try a few times (RED-PEN better strategy?)
695 for (i
= 0; i
< N_UNMAP_TRIES
; i
++) {
696 ret
= try_to_unmap(p
, ttu
);
697 if (ret
== SWAP_SUCCESS
)
699 pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn
, ret
);
702 if (ret
!= SWAP_SUCCESS
)
703 printk(KERN_ERR
"MCE %#lx: failed to unmap page (mapcount=%d)\n",
704 pfn
, page_mapcount(p
));
707 * Now that the dirty bit has been propagated to the
708 * struct page and all unmaps done we can decide if
709 * killing is needed or not. Only kill when the page
710 * was dirty, otherwise the tokill list is merely
711 * freed. When there was a problem unmapping earlier
712 * use a more force-full uncatchable kill to prevent
713 * any accesses to the poisoned memory.
715 kill_procs_ao(&tokill
, !!PageDirty(p
), trapno
,
716 ret
!= SWAP_SUCCESS
, pfn
);
719 int __memory_failure(unsigned long pfn
, int trapno
, int ref
)
721 unsigned long lru_flag
;
722 struct page_state
*ps
;
726 if (!sysctl_memory_failure_recovery
)
727 panic("Memory failure from trap %d on page %lx", trapno
, pfn
);
729 if (!pfn_valid(pfn
)) {
730 action_result(pfn
, "memory outside kernel control", IGNORED
);
734 p
= pfn_to_page(pfn
);
735 if (TestSetPageHWPoison(p
)) {
736 action_result(pfn
, "already hardware poisoned", IGNORED
);
740 atomic_long_add(1, &mce_bad_pages
);
743 * We need/can do nothing about count=0 pages.
744 * 1) it's a free page, and therefore in safe hand:
745 * prep_new_page() will be the gate keeper.
746 * 2) it's part of a non-compound high order page.
747 * Implies some kernel user: cannot stop them from
748 * R/W the page; let's pray that the page has been
749 * used and will be freed some time later.
750 * In fact it's dangerous to directly bump up page count from 0,
751 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
753 if (!get_page_unless_zero(compound_head(p
))) {
754 action_result(pfn
, "free or high order kernel", IGNORED
);
755 return PageBuddy(compound_head(p
)) ? 0 : -EBUSY
;
759 * We ignore non-LRU pages for good reasons.
760 * - PG_locked is only well defined for LRU pages and a few others
761 * - to avoid races with __set_page_locked()
762 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
763 * The check (unnecessarily) ignores LRU pages being isolated and
764 * walked by the page reclaim code, however that's not a big loss.
768 lru_flag
= p
->flags
& lru
;
769 if (isolate_lru_page(p
)) {
770 action_result(pfn
, "non LRU", IGNORED
);
774 page_cache_release(p
);
777 * Lock the page and wait for writeback to finish.
778 * It's very difficult to mess with pages currently under IO
779 * and in many cases impossible, so we just avoid it here.
782 wait_on_page_writeback(p
);
785 * Now take care of user space mappings.
787 hwpoison_user_mappings(p
, pfn
, trapno
);
790 * Torn down by someone else?
792 if ((lru_flag
& lru
) && !PageSwapCache(p
) && p
->mapping
== NULL
) {
793 action_result(pfn
, "already truncated LRU", IGNORED
);
799 for (ps
= error_states
;; ps
++) {
800 if (((p
->flags
| lru_flag
)& ps
->mask
) == ps
->res
) {
801 res
= page_action(ps
, p
, pfn
, ref
);
809 EXPORT_SYMBOL_GPL(__memory_failure
);
812 * memory_failure - Handle memory failure of a page.
813 * @pfn: Page Number of the corrupted page
814 * @trapno: Trap number reported in the signal to user space.
816 * This function is called by the low level machine check code
817 * of an architecture when it detects hardware memory corruption
818 * of a page. It tries its best to recover, which includes
819 * dropping pages, killing processes etc.
821 * The function is primarily of use for corruptions that
822 * happen outside the current execution context (e.g. when
823 * detected by a background scrubber)
825 * Must run in process context (e.g. a work queue) with interrupts
826 * enabled and no spinlocks hold.
828 void memory_failure(unsigned long pfn
, int trapno
)
830 __memory_failure(pfn
, trapno
, 0);