ARM: hw_breakpoint: ensure OS lock is clear before writing to debug registers
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / memory-failure.c
blob124324134ff67b3c4a0bc06661b4f70f2406840f
1 /*
2 * Copyright (C) 2008, 2009 Intel Corporation
3 * Authors: Andi Kleen, Fengguang Wu
5 * This software may be redistributed and/or modified under the terms of
6 * the GNU General Public License ("GPL") version 2 only as published by the
7 * Free Software Foundation.
9 * High level machine check handler. Handles pages reported by the
10 * hardware as being corrupted usually due to a multi-bit ECC memory or cache
11 * failure.
13 * In addition there is a "soft offline" entry point that allows stop using
14 * not-yet-corrupted-by-suspicious pages without killing anything.
16 * Handles page cache pages in various states. The tricky part
17 * here is that we can access any page asynchronously in respect to
18 * other VM users, because memory failures could happen anytime and
19 * anywhere. This could violate some of their assumptions. This is why
20 * this code has to be extremely careful. Generally it tries to use
21 * normal locking rules, as in get the standard locks, even if that means
22 * the error handling takes potentially a long time.
24 * There are several operations here with exponential complexity because
25 * of unsuitable VM data structures. For example the operation to map back
26 * from RMAP chains to processes has to walk the complete process list and
27 * has non linear complexity with the number. But since memory corruptions
28 * are rare we hope to get away with this. This avoids impacting the core
29 * VM.
33 * Notebook:
34 * - hugetlb needs more code
35 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
36 * - pass bad pages to kdump next kernel
38 #include <linux/kernel.h>
39 #include <linux/mm.h>
40 #include <linux/page-flags.h>
41 #include <linux/kernel-page-flags.h>
42 #include <linux/sched.h>
43 #include <linux/ksm.h>
44 #include <linux/rmap.h>
45 #include <linux/pagemap.h>
46 #include <linux/swap.h>
47 #include <linux/backing-dev.h>
48 #include <linux/migrate.h>
49 #include <linux/page-isolation.h>
50 #include <linux/suspend.h>
51 #include <linux/slab.h>
52 #include <linux/swapops.h>
53 #include <linux/hugetlb.h>
54 #include "internal.h"
56 int sysctl_memory_failure_early_kill __read_mostly = 0;
58 int sysctl_memory_failure_recovery __read_mostly = 1;
60 atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
62 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
64 u32 hwpoison_filter_enable = 0;
65 u32 hwpoison_filter_dev_major = ~0U;
66 u32 hwpoison_filter_dev_minor = ~0U;
67 u64 hwpoison_filter_flags_mask;
68 u64 hwpoison_filter_flags_value;
69 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
70 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
71 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
72 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
73 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
75 static int hwpoison_filter_dev(struct page *p)
77 struct address_space *mapping;
78 dev_t dev;
80 if (hwpoison_filter_dev_major == ~0U &&
81 hwpoison_filter_dev_minor == ~0U)
82 return 0;
85 * page_mapping() does not accept slab pages.
87 if (PageSlab(p))
88 return -EINVAL;
90 mapping = page_mapping(p);
91 if (mapping == NULL || mapping->host == NULL)
92 return -EINVAL;
94 dev = mapping->host->i_sb->s_dev;
95 if (hwpoison_filter_dev_major != ~0U &&
96 hwpoison_filter_dev_major != MAJOR(dev))
97 return -EINVAL;
98 if (hwpoison_filter_dev_minor != ~0U &&
99 hwpoison_filter_dev_minor != MINOR(dev))
100 return -EINVAL;
102 return 0;
105 static int hwpoison_filter_flags(struct page *p)
107 if (!hwpoison_filter_flags_mask)
108 return 0;
110 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
111 hwpoison_filter_flags_value)
112 return 0;
113 else
114 return -EINVAL;
118 * This allows stress tests to limit test scope to a collection of tasks
119 * by putting them under some memcg. This prevents killing unrelated/important
120 * processes such as /sbin/init. Note that the target task may share clean
121 * pages with init (eg. libc text), which is harmless. If the target task
122 * share _dirty_ pages with another task B, the test scheme must make sure B
123 * is also included in the memcg. At last, due to race conditions this filter
124 * can only guarantee that the page either belongs to the memcg tasks, or is
125 * a freed page.
127 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
128 u64 hwpoison_filter_memcg;
129 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
130 static int hwpoison_filter_task(struct page *p)
132 struct mem_cgroup *mem;
133 struct cgroup_subsys_state *css;
134 unsigned long ino;
136 if (!hwpoison_filter_memcg)
137 return 0;
139 mem = try_get_mem_cgroup_from_page(p);
140 if (!mem)
141 return -EINVAL;
143 css = mem_cgroup_css(mem);
144 /* root_mem_cgroup has NULL dentries */
145 if (!css->cgroup->dentry)
146 return -EINVAL;
148 ino = css->cgroup->dentry->d_inode->i_ino;
149 css_put(css);
151 if (ino != hwpoison_filter_memcg)
152 return -EINVAL;
154 return 0;
156 #else
157 static int hwpoison_filter_task(struct page *p) { return 0; }
158 #endif
160 int hwpoison_filter(struct page *p)
162 if (!hwpoison_filter_enable)
163 return 0;
165 if (hwpoison_filter_dev(p))
166 return -EINVAL;
168 if (hwpoison_filter_flags(p))
169 return -EINVAL;
171 if (hwpoison_filter_task(p))
172 return -EINVAL;
174 return 0;
176 #else
177 int hwpoison_filter(struct page *p)
179 return 0;
181 #endif
183 EXPORT_SYMBOL_GPL(hwpoison_filter);
186 * Send all the processes who have the page mapped an ``action optional''
187 * signal.
189 static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
190 unsigned long pfn, struct page *page)
192 struct siginfo si;
193 int ret;
195 printk(KERN_ERR
196 "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
197 pfn, t->comm, t->pid);
198 si.si_signo = SIGBUS;
199 si.si_errno = 0;
200 si.si_code = BUS_MCEERR_AO;
201 si.si_addr = (void *)addr;
202 #ifdef __ARCH_SI_TRAPNO
203 si.si_trapno = trapno;
204 #endif
205 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
207 * Don't use force here, it's convenient if the signal
208 * can be temporarily blocked.
209 * This could cause a loop when the user sets SIGBUS
210 * to SIG_IGN, but hopefully noone will do that?
212 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
213 if (ret < 0)
214 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
215 t->comm, t->pid, ret);
216 return ret;
220 * When a unknown page type is encountered drain as many buffers as possible
221 * in the hope to turn the page into a LRU or free page, which we can handle.
223 void shake_page(struct page *p, int access)
225 if (!PageSlab(p)) {
226 lru_add_drain_all();
227 if (PageLRU(p))
228 return;
229 drain_all_pages();
230 if (PageLRU(p) || is_free_buddy_page(p))
231 return;
235 * Only all shrink_slab here (which would also
236 * shrink other caches) if access is not potentially fatal.
238 if (access) {
239 int nr;
240 do {
241 nr = shrink_slab(1000, GFP_KERNEL, 1000);
242 if (page_count(p) == 1)
243 break;
244 } while (nr > 10);
247 EXPORT_SYMBOL_GPL(shake_page);
250 * Kill all processes that have a poisoned page mapped and then isolate
251 * the page.
253 * General strategy:
254 * Find all processes having the page mapped and kill them.
255 * But we keep a page reference around so that the page is not
256 * actually freed yet.
257 * Then stash the page away
259 * There's no convenient way to get back to mapped processes
260 * from the VMAs. So do a brute-force search over all
261 * running processes.
263 * Remember that machine checks are not common (or rather
264 * if they are common you have other problems), so this shouldn't
265 * be a performance issue.
267 * Also there are some races possible while we get from the
268 * error detection to actually handle it.
271 struct to_kill {
272 struct list_head nd;
273 struct task_struct *tsk;
274 unsigned long addr;
275 char addr_valid;
279 * Failure handling: if we can't find or can't kill a process there's
280 * not much we can do. We just print a message and ignore otherwise.
284 * Schedule a process for later kill.
285 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
286 * TBD would GFP_NOIO be enough?
288 static void add_to_kill(struct task_struct *tsk, struct page *p,
289 struct vm_area_struct *vma,
290 struct list_head *to_kill,
291 struct to_kill **tkc)
293 struct to_kill *tk;
295 if (*tkc) {
296 tk = *tkc;
297 *tkc = NULL;
298 } else {
299 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
300 if (!tk) {
301 printk(KERN_ERR
302 "MCE: Out of memory while machine check handling\n");
303 return;
306 tk->addr = page_address_in_vma(p, vma);
307 tk->addr_valid = 1;
310 * In theory we don't have to kill when the page was
311 * munmaped. But it could be also a mremap. Since that's
312 * likely very rare kill anyways just out of paranoia, but use
313 * a SIGKILL because the error is not contained anymore.
315 if (tk->addr == -EFAULT) {
316 pr_info("MCE: Unable to find user space address %lx in %s\n",
317 page_to_pfn(p), tsk->comm);
318 tk->addr_valid = 0;
320 get_task_struct(tsk);
321 tk->tsk = tsk;
322 list_add_tail(&tk->nd, to_kill);
326 * Kill the processes that have been collected earlier.
328 * Only do anything when DOIT is set, otherwise just free the list
329 * (this is used for clean pages which do not need killing)
330 * Also when FAIL is set do a force kill because something went
331 * wrong earlier.
333 static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
334 int fail, struct page *page, unsigned long pfn)
336 struct to_kill *tk, *next;
338 list_for_each_entry_safe (tk, next, to_kill, nd) {
339 if (doit) {
341 * In case something went wrong with munmapping
342 * make sure the process doesn't catch the
343 * signal and then access the memory. Just kill it.
345 if (fail || tk->addr_valid == 0) {
346 printk(KERN_ERR
347 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
348 pfn, tk->tsk->comm, tk->tsk->pid);
349 force_sig(SIGKILL, tk->tsk);
353 * In theory the process could have mapped
354 * something else on the address in-between. We could
355 * check for that, but we need to tell the
356 * process anyways.
358 else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
359 pfn, page) < 0)
360 printk(KERN_ERR
361 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
362 pfn, tk->tsk->comm, tk->tsk->pid);
364 put_task_struct(tk->tsk);
365 kfree(tk);
369 static int task_early_kill(struct task_struct *tsk)
371 if (!tsk->mm)
372 return 0;
373 if (tsk->flags & PF_MCE_PROCESS)
374 return !!(tsk->flags & PF_MCE_EARLY);
375 return sysctl_memory_failure_early_kill;
379 * Collect processes when the error hit an anonymous page.
381 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
382 struct to_kill **tkc)
384 struct vm_area_struct *vma;
385 struct task_struct *tsk;
386 struct anon_vma *av;
388 read_lock(&tasklist_lock);
389 av = page_lock_anon_vma(page);
390 if (av == NULL) /* Not actually mapped anymore */
391 goto out;
392 for_each_process (tsk) {
393 struct anon_vma_chain *vmac;
395 if (!task_early_kill(tsk))
396 continue;
397 list_for_each_entry(vmac, &av->head, same_anon_vma) {
398 vma = vmac->vma;
399 if (!page_mapped_in_vma(page, vma))
400 continue;
401 if (vma->vm_mm == tsk->mm)
402 add_to_kill(tsk, page, vma, to_kill, tkc);
405 page_unlock_anon_vma(av);
406 out:
407 read_unlock(&tasklist_lock);
411 * Collect processes when the error hit a file mapped page.
413 static void collect_procs_file(struct page *page, struct list_head *to_kill,
414 struct to_kill **tkc)
416 struct vm_area_struct *vma;
417 struct task_struct *tsk;
418 struct prio_tree_iter iter;
419 struct address_space *mapping = page->mapping;
422 * A note on the locking order between the two locks.
423 * We don't rely on this particular order.
424 * If you have some other code that needs a different order
425 * feel free to switch them around. Or add a reverse link
426 * from mm_struct to task_struct, then this could be all
427 * done without taking tasklist_lock and looping over all tasks.
430 read_lock(&tasklist_lock);
431 spin_lock(&mapping->i_mmap_lock);
432 for_each_process(tsk) {
433 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
435 if (!task_early_kill(tsk))
436 continue;
438 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
439 pgoff) {
441 * Send early kill signal to tasks where a vma covers
442 * the page but the corrupted page is not necessarily
443 * mapped it in its pte.
444 * Assume applications who requested early kill want
445 * to be informed of all such data corruptions.
447 if (vma->vm_mm == tsk->mm)
448 add_to_kill(tsk, page, vma, to_kill, tkc);
451 spin_unlock(&mapping->i_mmap_lock);
452 read_unlock(&tasklist_lock);
456 * Collect the processes who have the corrupted page mapped to kill.
457 * This is done in two steps for locking reasons.
458 * First preallocate one tokill structure outside the spin locks,
459 * so that we can kill at least one process reasonably reliable.
461 static void collect_procs(struct page *page, struct list_head *tokill)
463 struct to_kill *tk;
465 if (!page->mapping)
466 return;
468 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
469 if (!tk)
470 return;
471 if (PageAnon(page))
472 collect_procs_anon(page, tokill, &tk);
473 else
474 collect_procs_file(page, tokill, &tk);
475 kfree(tk);
479 * Error handlers for various types of pages.
482 enum outcome {
483 IGNORED, /* Error: cannot be handled */
484 FAILED, /* Error: handling failed */
485 DELAYED, /* Will be handled later */
486 RECOVERED, /* Successfully recovered */
489 static const char *action_name[] = {
490 [IGNORED] = "Ignored",
491 [FAILED] = "Failed",
492 [DELAYED] = "Delayed",
493 [RECOVERED] = "Recovered",
497 * XXX: It is possible that a page is isolated from LRU cache,
498 * and then kept in swap cache or failed to remove from page cache.
499 * The page count will stop it from being freed by unpoison.
500 * Stress tests should be aware of this memory leak problem.
502 static int delete_from_lru_cache(struct page *p)
504 if (!isolate_lru_page(p)) {
506 * Clear sensible page flags, so that the buddy system won't
507 * complain when the page is unpoison-and-freed.
509 ClearPageActive(p);
510 ClearPageUnevictable(p);
512 * drop the page count elevated by isolate_lru_page()
514 page_cache_release(p);
515 return 0;
517 return -EIO;
521 * Error hit kernel page.
522 * Do nothing, try to be lucky and not touch this instead. For a few cases we
523 * could be more sophisticated.
525 static int me_kernel(struct page *p, unsigned long pfn)
527 return IGNORED;
531 * Page in unknown state. Do nothing.
533 static int me_unknown(struct page *p, unsigned long pfn)
535 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
536 return FAILED;
540 * Clean (or cleaned) page cache page.
542 static int me_pagecache_clean(struct page *p, unsigned long pfn)
544 int err;
545 int ret = FAILED;
546 struct address_space *mapping;
548 delete_from_lru_cache(p);
551 * For anonymous pages we're done the only reference left
552 * should be the one m_f() holds.
554 if (PageAnon(p))
555 return RECOVERED;
558 * Now truncate the page in the page cache. This is really
559 * more like a "temporary hole punch"
560 * Don't do this for block devices when someone else
561 * has a reference, because it could be file system metadata
562 * and that's not safe to truncate.
564 mapping = page_mapping(p);
565 if (!mapping) {
567 * Page has been teared down in the meanwhile
569 return FAILED;
573 * Truncation is a bit tricky. Enable it per file system for now.
575 * Open: to take i_mutex or not for this? Right now we don't.
577 if (mapping->a_ops->error_remove_page) {
578 err = mapping->a_ops->error_remove_page(mapping, p);
579 if (err != 0) {
580 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
581 pfn, err);
582 } else if (page_has_private(p) &&
583 !try_to_release_page(p, GFP_NOIO)) {
584 pr_info("MCE %#lx: failed to release buffers\n", pfn);
585 } else {
586 ret = RECOVERED;
588 } else {
590 * If the file system doesn't support it just invalidate
591 * This fails on dirty or anything with private pages
593 if (invalidate_inode_page(p))
594 ret = RECOVERED;
595 else
596 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
597 pfn);
599 return ret;
603 * Dirty cache page page
604 * Issues: when the error hit a hole page the error is not properly
605 * propagated.
607 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
609 struct address_space *mapping = page_mapping(p);
611 SetPageError(p);
612 /* TBD: print more information about the file. */
613 if (mapping) {
615 * IO error will be reported by write(), fsync(), etc.
616 * who check the mapping.
617 * This way the application knows that something went
618 * wrong with its dirty file data.
620 * There's one open issue:
622 * The EIO will be only reported on the next IO
623 * operation and then cleared through the IO map.
624 * Normally Linux has two mechanisms to pass IO error
625 * first through the AS_EIO flag in the address space
626 * and then through the PageError flag in the page.
627 * Since we drop pages on memory failure handling the
628 * only mechanism open to use is through AS_AIO.
630 * This has the disadvantage that it gets cleared on
631 * the first operation that returns an error, while
632 * the PageError bit is more sticky and only cleared
633 * when the page is reread or dropped. If an
634 * application assumes it will always get error on
635 * fsync, but does other operations on the fd before
636 * and the page is dropped inbetween then the error
637 * will not be properly reported.
639 * This can already happen even without hwpoisoned
640 * pages: first on metadata IO errors (which only
641 * report through AS_EIO) or when the page is dropped
642 * at the wrong time.
644 * So right now we assume that the application DTRT on
645 * the first EIO, but we're not worse than other parts
646 * of the kernel.
648 mapping_set_error(mapping, EIO);
651 return me_pagecache_clean(p, pfn);
655 * Clean and dirty swap cache.
657 * Dirty swap cache page is tricky to handle. The page could live both in page
658 * cache and swap cache(ie. page is freshly swapped in). So it could be
659 * referenced concurrently by 2 types of PTEs:
660 * normal PTEs and swap PTEs. We try to handle them consistently by calling
661 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
662 * and then
663 * - clear dirty bit to prevent IO
664 * - remove from LRU
665 * - but keep in the swap cache, so that when we return to it on
666 * a later page fault, we know the application is accessing
667 * corrupted data and shall be killed (we installed simple
668 * interception code in do_swap_page to catch it).
670 * Clean swap cache pages can be directly isolated. A later page fault will
671 * bring in the known good data from disk.
673 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
675 ClearPageDirty(p);
676 /* Trigger EIO in shmem: */
677 ClearPageUptodate(p);
679 if (!delete_from_lru_cache(p))
680 return DELAYED;
681 else
682 return FAILED;
685 static int me_swapcache_clean(struct page *p, unsigned long pfn)
687 delete_from_swap_cache(p);
689 if (!delete_from_lru_cache(p))
690 return RECOVERED;
691 else
692 return FAILED;
696 * Huge pages. Needs work.
697 * Issues:
698 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
699 * To narrow down kill region to one page, we need to break up pmd.
701 static int me_huge_page(struct page *p, unsigned long pfn)
703 int res = 0;
704 struct page *hpage = compound_head(p);
706 * We can safely recover from error on free or reserved (i.e.
707 * not in-use) hugepage by dequeuing it from freelist.
708 * To check whether a hugepage is in-use or not, we can't use
709 * page->lru because it can be used in other hugepage operations,
710 * such as __unmap_hugepage_range() and gather_surplus_pages().
711 * So instead we use page_mapping() and PageAnon().
712 * We assume that this function is called with page lock held,
713 * so there is no race between isolation and mapping/unmapping.
715 if (!(page_mapping(hpage) || PageAnon(hpage))) {
716 res = dequeue_hwpoisoned_huge_page(hpage);
717 if (!res)
718 return RECOVERED;
720 return DELAYED;
724 * Various page states we can handle.
726 * A page state is defined by its current page->flags bits.
727 * The table matches them in order and calls the right handler.
729 * This is quite tricky because we can access page at any time
730 * in its live cycle, so all accesses have to be extremly careful.
732 * This is not complete. More states could be added.
733 * For any missing state don't attempt recovery.
736 #define dirty (1UL << PG_dirty)
737 #define sc (1UL << PG_swapcache)
738 #define unevict (1UL << PG_unevictable)
739 #define mlock (1UL << PG_mlocked)
740 #define writeback (1UL << PG_writeback)
741 #define lru (1UL << PG_lru)
742 #define swapbacked (1UL << PG_swapbacked)
743 #define head (1UL << PG_head)
744 #define tail (1UL << PG_tail)
745 #define compound (1UL << PG_compound)
746 #define slab (1UL << PG_slab)
747 #define reserved (1UL << PG_reserved)
749 static struct page_state {
750 unsigned long mask;
751 unsigned long res;
752 char *msg;
753 int (*action)(struct page *p, unsigned long pfn);
754 } error_states[] = {
755 { reserved, reserved, "reserved kernel", me_kernel },
757 * free pages are specially detected outside this table:
758 * PG_buddy pages only make a small fraction of all free pages.
762 * Could in theory check if slab page is free or if we can drop
763 * currently unused objects without touching them. But just
764 * treat it as standard kernel for now.
766 { slab, slab, "kernel slab", me_kernel },
768 #ifdef CONFIG_PAGEFLAGS_EXTENDED
769 { head, head, "huge", me_huge_page },
770 { tail, tail, "huge", me_huge_page },
771 #else
772 { compound, compound, "huge", me_huge_page },
773 #endif
775 { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
776 { sc|dirty, sc, "swapcache", me_swapcache_clean },
778 { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
779 { unevict, unevict, "unevictable LRU", me_pagecache_clean},
781 { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
782 { mlock, mlock, "mlocked LRU", me_pagecache_clean },
784 { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
785 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
788 * Catchall entry: must be at end.
790 { 0, 0, "unknown page state", me_unknown },
793 #undef dirty
794 #undef sc
795 #undef unevict
796 #undef mlock
797 #undef writeback
798 #undef lru
799 #undef swapbacked
800 #undef head
801 #undef tail
802 #undef compound
803 #undef slab
804 #undef reserved
806 static void action_result(unsigned long pfn, char *msg, int result)
808 struct page *page = pfn_to_page(pfn);
810 printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
811 pfn,
812 PageDirty(page) ? "dirty " : "",
813 msg, action_name[result]);
816 static int page_action(struct page_state *ps, struct page *p,
817 unsigned long pfn)
819 int result;
820 int count;
822 result = ps->action(p, pfn);
823 action_result(pfn, ps->msg, result);
825 count = page_count(p) - 1;
826 if (ps->action == me_swapcache_dirty && result == DELAYED)
827 count--;
828 if (count != 0) {
829 printk(KERN_ERR
830 "MCE %#lx: %s page still referenced by %d users\n",
831 pfn, ps->msg, count);
832 result = FAILED;
835 /* Could do more checks here if page looks ok */
837 * Could adjust zone counters here to correct for the missing page.
840 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
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,
848 int trapno)
850 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
851 struct address_space *mapping;
852 LIST_HEAD(tokill);
853 int ret;
854 int kill = 1;
855 struct page *hpage = compound_head(p);
857 if (PageReserved(p) || PageSlab(p))
858 return SWAP_SUCCESS;
861 * This check implies we don't kill processes if their pages
862 * are in the swap cache early. Those are always late kills.
864 if (!page_mapped(hpage))
865 return SWAP_SUCCESS;
867 if (PageKsm(p))
868 return SWAP_FAIL;
870 if (PageSwapCache(p)) {
871 printk(KERN_ERR
872 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
873 ttu |= TTU_IGNORE_HWPOISON;
877 * Propagate the dirty bit from PTEs to struct page first, because we
878 * need this to decide if we should kill or just drop the page.
879 * XXX: the dirty test could be racy: set_page_dirty() may not always
880 * be called inside page lock (it's recommended but not enforced).
882 mapping = page_mapping(hpage);
883 if (!PageDirty(hpage) && mapping &&
884 mapping_cap_writeback_dirty(mapping)) {
885 if (page_mkclean(hpage)) {
886 SetPageDirty(hpage);
887 } else {
888 kill = 0;
889 ttu |= TTU_IGNORE_HWPOISON;
890 printk(KERN_INFO
891 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
892 pfn);
897 * First collect all the processes that have the page
898 * mapped in dirty form. This has to be done before try_to_unmap,
899 * because ttu takes the rmap data structures down.
901 * Error handling: We ignore errors here because
902 * there's nothing that can be done.
904 if (kill)
905 collect_procs(hpage, &tokill);
907 ret = try_to_unmap(hpage, ttu);
908 if (ret != SWAP_SUCCESS)
909 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
910 pfn, page_mapcount(hpage));
913 * Now that the dirty bit has been propagated to the
914 * struct page and all unmaps done we can decide if
915 * killing is needed or not. Only kill when the page
916 * was dirty, otherwise the tokill list is merely
917 * freed. When there was a problem unmapping earlier
918 * use a more force-full uncatchable kill to prevent
919 * any accesses to the poisoned memory.
921 kill_procs_ao(&tokill, !!PageDirty(hpage), trapno,
922 ret != SWAP_SUCCESS, p, pfn);
924 return ret;
927 static void set_page_hwpoison_huge_page(struct page *hpage)
929 int i;
930 int nr_pages = 1 << compound_order(hpage);
931 for (i = 0; i < nr_pages; i++)
932 SetPageHWPoison(hpage + i);
935 static void clear_page_hwpoison_huge_page(struct page *hpage)
937 int i;
938 int nr_pages = 1 << compound_order(hpage);
939 for (i = 0; i < nr_pages; i++)
940 ClearPageHWPoison(hpage + i);
943 int __memory_failure(unsigned long pfn, int trapno, int flags)
945 struct page_state *ps;
946 struct page *p;
947 struct page *hpage;
948 int res;
949 unsigned int nr_pages;
951 if (!sysctl_memory_failure_recovery)
952 panic("Memory failure from trap %d on page %lx", trapno, pfn);
954 if (!pfn_valid(pfn)) {
955 printk(KERN_ERR
956 "MCE %#lx: memory outside kernel control\n",
957 pfn);
958 return -ENXIO;
961 p = pfn_to_page(pfn);
962 hpage = compound_head(p);
963 if (TestSetPageHWPoison(p)) {
964 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
965 return 0;
968 nr_pages = 1 << compound_order(hpage);
969 atomic_long_add(nr_pages, &mce_bad_pages);
972 * We need/can do nothing about count=0 pages.
973 * 1) it's a free page, and therefore in safe hand:
974 * prep_new_page() will be the gate keeper.
975 * 2) it's a free hugepage, which is also safe:
976 * an affected hugepage will be dequeued from hugepage freelist,
977 * so there's no concern about reusing it ever after.
978 * 3) it's part of a non-compound high order page.
979 * Implies some kernel user: cannot stop them from
980 * R/W the page; let's pray that the page has been
981 * used and will be freed some time later.
982 * In fact it's dangerous to directly bump up page count from 0,
983 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
985 if (!(flags & MF_COUNT_INCREASED) &&
986 !get_page_unless_zero(hpage)) {
987 if (is_free_buddy_page(p)) {
988 action_result(pfn, "free buddy", DELAYED);
989 return 0;
990 } else if (PageHuge(hpage)) {
992 * Check "just unpoisoned", "filter hit", and
993 * "race with other subpage."
995 lock_page_nosync(hpage);
996 if (!PageHWPoison(hpage)
997 || (hwpoison_filter(p) && TestClearPageHWPoison(p))
998 || (p != hpage && TestSetPageHWPoison(hpage))) {
999 atomic_long_sub(nr_pages, &mce_bad_pages);
1000 return 0;
1002 set_page_hwpoison_huge_page(hpage);
1003 res = dequeue_hwpoisoned_huge_page(hpage);
1004 action_result(pfn, "free huge",
1005 res ? IGNORED : DELAYED);
1006 unlock_page(hpage);
1007 return res;
1008 } else {
1009 action_result(pfn, "high order kernel", IGNORED);
1010 return -EBUSY;
1015 * We ignore non-LRU pages for good reasons.
1016 * - PG_locked is only well defined for LRU pages and a few others
1017 * - to avoid races with __set_page_locked()
1018 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1019 * The check (unnecessarily) ignores LRU pages being isolated and
1020 * walked by the page reclaim code, however that's not a big loss.
1022 if (!PageLRU(p) && !PageHuge(p))
1023 shake_page(p, 0);
1024 if (!PageLRU(p) && !PageHuge(p)) {
1026 * shake_page could have turned it free.
1028 if (is_free_buddy_page(p)) {
1029 action_result(pfn, "free buddy, 2nd try", DELAYED);
1030 return 0;
1032 action_result(pfn, "non LRU", IGNORED);
1033 put_page(p);
1034 return -EBUSY;
1038 * Lock the page and wait for writeback to finish.
1039 * It's very difficult to mess with pages currently under IO
1040 * and in many cases impossible, so we just avoid it here.
1042 lock_page_nosync(hpage);
1045 * unpoison always clear PG_hwpoison inside page lock
1047 if (!PageHWPoison(p)) {
1048 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1049 res = 0;
1050 goto out;
1052 if (hwpoison_filter(p)) {
1053 if (TestClearPageHWPoison(p))
1054 atomic_long_sub(nr_pages, &mce_bad_pages);
1055 unlock_page(hpage);
1056 put_page(hpage);
1057 return 0;
1061 * For error on the tail page, we should set PG_hwpoison
1062 * on the head page to show that the hugepage is hwpoisoned
1064 if (PageTail(p) && TestSetPageHWPoison(hpage)) {
1065 action_result(pfn, "hugepage already hardware poisoned",
1066 IGNORED);
1067 unlock_page(hpage);
1068 put_page(hpage);
1069 return 0;
1072 * Set PG_hwpoison on all pages in an error hugepage,
1073 * because containment is done in hugepage unit for now.
1074 * Since we have done TestSetPageHWPoison() for the head page with
1075 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1077 if (PageHuge(p))
1078 set_page_hwpoison_huge_page(hpage);
1080 wait_on_page_writeback(p);
1083 * Now take care of user space mappings.
1084 * Abort on fail: __remove_from_page_cache() assumes unmapped page.
1086 if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
1087 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1088 res = -EBUSY;
1089 goto out;
1093 * Torn down by someone else?
1095 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1096 action_result(pfn, "already truncated LRU", IGNORED);
1097 res = -EBUSY;
1098 goto out;
1101 res = -EBUSY;
1102 for (ps = error_states;; ps++) {
1103 if ((p->flags & ps->mask) == ps->res) {
1104 res = page_action(ps, p, pfn);
1105 break;
1108 out:
1109 unlock_page(hpage);
1110 return res;
1112 EXPORT_SYMBOL_GPL(__memory_failure);
1115 * memory_failure - Handle memory failure of a page.
1116 * @pfn: Page Number of the corrupted page
1117 * @trapno: Trap number reported in the signal to user space.
1119 * This function is called by the low level machine check code
1120 * of an architecture when it detects hardware memory corruption
1121 * of a page. It tries its best to recover, which includes
1122 * dropping pages, killing processes etc.
1124 * The function is primarily of use for corruptions that
1125 * happen outside the current execution context (e.g. when
1126 * detected by a background scrubber)
1128 * Must run in process context (e.g. a work queue) with interrupts
1129 * enabled and no spinlocks hold.
1131 void memory_failure(unsigned long pfn, int trapno)
1133 __memory_failure(pfn, trapno, 0);
1137 * unpoison_memory - Unpoison a previously poisoned page
1138 * @pfn: Page number of the to be unpoisoned page
1140 * Software-unpoison a page that has been poisoned by
1141 * memory_failure() earlier.
1143 * This is only done on the software-level, so it only works
1144 * for linux injected failures, not real hardware failures
1146 * Returns 0 for success, otherwise -errno.
1148 int unpoison_memory(unsigned long pfn)
1150 struct page *page;
1151 struct page *p;
1152 int freeit = 0;
1153 unsigned int nr_pages;
1155 if (!pfn_valid(pfn))
1156 return -ENXIO;
1158 p = pfn_to_page(pfn);
1159 page = compound_head(p);
1161 if (!PageHWPoison(p)) {
1162 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
1163 return 0;
1166 nr_pages = 1 << compound_order(page);
1168 if (!get_page_unless_zero(page)) {
1170 * Since HWPoisoned hugepage should have non-zero refcount,
1171 * race between memory failure and unpoison seems to happen.
1172 * In such case unpoison fails and memory failure runs
1173 * to the end.
1175 if (PageHuge(page)) {
1176 pr_debug("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1177 return 0;
1179 if (TestClearPageHWPoison(p))
1180 atomic_long_sub(nr_pages, &mce_bad_pages);
1181 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
1182 return 0;
1185 lock_page_nosync(page);
1187 * This test is racy because PG_hwpoison is set outside of page lock.
1188 * That's acceptable because that won't trigger kernel panic. Instead,
1189 * the PG_hwpoison page will be caught and isolated on the entrance to
1190 * the free buddy page pool.
1192 if (TestClearPageHWPoison(page)) {
1193 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
1194 atomic_long_sub(nr_pages, &mce_bad_pages);
1195 freeit = 1;
1196 if (PageHuge(page))
1197 clear_page_hwpoison_huge_page(page);
1199 unlock_page(page);
1201 put_page(page);
1202 if (freeit)
1203 put_page(page);
1205 return 0;
1207 EXPORT_SYMBOL(unpoison_memory);
1209 static struct page *new_page(struct page *p, unsigned long private, int **x)
1211 int nid = page_to_nid(p);
1212 if (PageHuge(p))
1213 return alloc_huge_page_node(page_hstate(compound_head(p)),
1214 nid);
1215 else
1216 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1220 * Safely get reference count of an arbitrary page.
1221 * Returns 0 for a free page, -EIO for a zero refcount page
1222 * that is not free, and 1 for any other page type.
1223 * For 1 the page is returned with increased page count, otherwise not.
1225 static int get_any_page(struct page *p, unsigned long pfn, int flags)
1227 int ret;
1229 if (flags & MF_COUNT_INCREASED)
1230 return 1;
1233 * The lock_system_sleep prevents a race with memory hotplug,
1234 * because the isolation assumes there's only a single user.
1235 * This is a big hammer, a better would be nicer.
1237 lock_system_sleep();
1240 * Isolate the page, so that it doesn't get reallocated if it
1241 * was free.
1243 set_migratetype_isolate(p);
1245 * When the target page is a free hugepage, just remove it
1246 * from free hugepage list.
1248 if (!get_page_unless_zero(compound_head(p))) {
1249 if (PageHuge(p)) {
1250 pr_info("get_any_page: %#lx free huge page\n", pfn);
1251 ret = dequeue_hwpoisoned_huge_page(compound_head(p));
1252 } else if (is_free_buddy_page(p)) {
1253 pr_info("get_any_page: %#lx free buddy page\n", pfn);
1254 /* Set hwpoison bit while page is still isolated */
1255 SetPageHWPoison(p);
1256 ret = 0;
1257 } else {
1258 pr_info("get_any_page: %#lx: unknown zero refcount page type %lx\n",
1259 pfn, p->flags);
1260 ret = -EIO;
1262 } else {
1263 /* Not a free page */
1264 ret = 1;
1266 unset_migratetype_isolate(p);
1267 unlock_system_sleep();
1268 return ret;
1271 static int soft_offline_huge_page(struct page *page, int flags)
1273 int ret;
1274 unsigned long pfn = page_to_pfn(page);
1275 struct page *hpage = compound_head(page);
1276 LIST_HEAD(pagelist);
1278 ret = get_any_page(page, pfn, flags);
1279 if (ret < 0)
1280 return ret;
1281 if (ret == 0)
1282 goto done;
1284 if (PageHWPoison(hpage)) {
1285 put_page(hpage);
1286 pr_debug("soft offline: %#lx hugepage already poisoned\n", pfn);
1287 return -EBUSY;
1290 /* Keep page count to indicate a given hugepage is isolated. */
1292 list_add(&hpage->lru, &pagelist);
1293 ret = migrate_huge_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0);
1294 if (ret) {
1295 putback_lru_pages(&pagelist);
1296 pr_debug("soft offline: %#lx: migration failed %d, type %lx\n",
1297 pfn, ret, page->flags);
1298 if (ret > 0)
1299 ret = -EIO;
1300 return ret;
1302 done:
1303 if (!PageHWPoison(hpage))
1304 atomic_long_add(1 << compound_order(hpage), &mce_bad_pages);
1305 set_page_hwpoison_huge_page(hpage);
1306 dequeue_hwpoisoned_huge_page(hpage);
1307 /* keep elevated page count for bad page */
1308 return ret;
1312 * soft_offline_page - Soft offline a page.
1313 * @page: page to offline
1314 * @flags: flags. Same as memory_failure().
1316 * Returns 0 on success, otherwise negated errno.
1318 * Soft offline a page, by migration or invalidation,
1319 * without killing anything. This is for the case when
1320 * a page is not corrupted yet (so it's still valid to access),
1321 * but has had a number of corrected errors and is better taken
1322 * out.
1324 * The actual policy on when to do that is maintained by
1325 * user space.
1327 * This should never impact any application or cause data loss,
1328 * however it might take some time.
1330 * This is not a 100% solution for all memory, but tries to be
1331 * ``good enough'' for the majority of memory.
1333 int soft_offline_page(struct page *page, int flags)
1335 int ret;
1336 unsigned long pfn = page_to_pfn(page);
1338 if (PageHuge(page))
1339 return soft_offline_huge_page(page, flags);
1341 ret = get_any_page(page, pfn, flags);
1342 if (ret < 0)
1343 return ret;
1344 if (ret == 0)
1345 goto done;
1348 * Page cache page we can handle?
1350 if (!PageLRU(page)) {
1352 * Try to free it.
1354 put_page(page);
1355 shake_page(page, 1);
1358 * Did it turn free?
1360 ret = get_any_page(page, pfn, 0);
1361 if (ret < 0)
1362 return ret;
1363 if (ret == 0)
1364 goto done;
1366 if (!PageLRU(page)) {
1367 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1368 pfn, page->flags);
1369 return -EIO;
1372 lock_page(page);
1373 wait_on_page_writeback(page);
1376 * Synchronized using the page lock with memory_failure()
1378 if (PageHWPoison(page)) {
1379 unlock_page(page);
1380 put_page(page);
1381 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1382 return -EBUSY;
1386 * Try to invalidate first. This should work for
1387 * non dirty unmapped page cache pages.
1389 ret = invalidate_inode_page(page);
1390 unlock_page(page);
1393 * Drop count because page migration doesn't like raised
1394 * counts. The page could get re-allocated, but if it becomes
1395 * LRU the isolation will just fail.
1396 * RED-PEN would be better to keep it isolated here, but we
1397 * would need to fix isolation locking first.
1399 put_page(page);
1400 if (ret == 1) {
1401 ret = 0;
1402 pr_info("soft_offline: %#lx: invalidated\n", pfn);
1403 goto done;
1407 * Simple invalidation didn't work.
1408 * Try to migrate to a new page instead. migrate.c
1409 * handles a large number of cases for us.
1411 ret = isolate_lru_page(page);
1412 if (!ret) {
1413 LIST_HEAD(pagelist);
1415 list_add(&page->lru, &pagelist);
1416 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0);
1417 if (ret) {
1418 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1419 pfn, ret, page->flags);
1420 if (ret > 0)
1421 ret = -EIO;
1423 } else {
1424 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1425 pfn, ret, page_count(page), page->flags);
1427 if (ret)
1428 return ret;
1430 done:
1431 atomic_long_add(1, &mce_bad_pages);
1432 SetPageHWPoison(page);
1433 /* keep elevated page count for bad page */
1434 return ret;
1438 * The caller must hold current->mm->mmap_sem in read mode.
1440 int is_hwpoison_address(unsigned long addr)
1442 pgd_t *pgdp;
1443 pud_t pud, *pudp;
1444 pmd_t pmd, *pmdp;
1445 pte_t pte, *ptep;
1446 swp_entry_t entry;
1448 pgdp = pgd_offset(current->mm, addr);
1449 if (!pgd_present(*pgdp))
1450 return 0;
1451 pudp = pud_offset(pgdp, addr);
1452 pud = *pudp;
1453 if (!pud_present(pud) || pud_large(pud))
1454 return 0;
1455 pmdp = pmd_offset(pudp, addr);
1456 pmd = *pmdp;
1457 if (!pmd_present(pmd) || pmd_large(pmd))
1458 return 0;
1459 ptep = pte_offset_map(pmdp, addr);
1460 pte = *ptep;
1461 pte_unmap(ptep);
1462 if (!is_swap_pte(pte))
1463 return 0;
1464 entry = pte_to_swp_entry(pte);
1465 return is_hwpoison_entry(entry);
1467 EXPORT_SYMBOL_GPL(is_hwpoison_address);