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
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 * It can be very tempting to add handling for obscure cases here.
25 * In general any code for handling new cases should only be added iff:
26 * - You know how to test it.
27 * - You have a test that can be added to mce-test
28 * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/
29 * - The case actually shows up as a frequent (top 10) page state in
30 * tools/vm/page-types when running a real workload.
32 * There are several operations here with exponential complexity because
33 * of unsuitable VM data structures. For example the operation to map back
34 * from RMAP chains to processes has to walk the complete process list and
35 * has non linear complexity with the number. But since memory corruptions
36 * are rare we hope to get away with this. This avoids impacting the core
39 #include <linux/kernel.h>
41 #include <linux/page-flags.h>
42 #include <linux/kernel-page-flags.h>
43 #include <linux/sched.h>
44 #include <linux/ksm.h>
45 #include <linux/rmap.h>
46 #include <linux/export.h>
47 #include <linux/pagemap.h>
48 #include <linux/swap.h>
49 #include <linux/backing-dev.h>
50 #include <linux/migrate.h>
51 #include <linux/page-isolation.h>
52 #include <linux/suspend.h>
53 #include <linux/slab.h>
54 #include <linux/swapops.h>
55 #include <linux/hugetlb.h>
56 #include <linux/memory_hotplug.h>
57 #include <linux/mm_inline.h>
58 #include <linux/kfifo.h>
59 #include <linux/ratelimit.h>
61 #include "ras/ras_event.h"
63 int sysctl_memory_failure_early_kill __read_mostly
= 0;
65 int sysctl_memory_failure_recovery __read_mostly
= 1;
67 atomic_long_t num_poisoned_pages __read_mostly
= ATOMIC_LONG_INIT(0);
69 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
71 u32 hwpoison_filter_enable
= 0;
72 u32 hwpoison_filter_dev_major
= ~0U;
73 u32 hwpoison_filter_dev_minor
= ~0U;
74 u64 hwpoison_filter_flags_mask
;
75 u64 hwpoison_filter_flags_value
;
76 EXPORT_SYMBOL_GPL(hwpoison_filter_enable
);
77 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major
);
78 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor
);
79 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask
);
80 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value
);
82 static int hwpoison_filter_dev(struct page
*p
)
84 struct address_space
*mapping
;
87 if (hwpoison_filter_dev_major
== ~0U &&
88 hwpoison_filter_dev_minor
== ~0U)
92 * page_mapping() does not accept slab pages.
97 mapping
= page_mapping(p
);
98 if (mapping
== NULL
|| mapping
->host
== NULL
)
101 dev
= mapping
->host
->i_sb
->s_dev
;
102 if (hwpoison_filter_dev_major
!= ~0U &&
103 hwpoison_filter_dev_major
!= MAJOR(dev
))
105 if (hwpoison_filter_dev_minor
!= ~0U &&
106 hwpoison_filter_dev_minor
!= MINOR(dev
))
112 static int hwpoison_filter_flags(struct page
*p
)
114 if (!hwpoison_filter_flags_mask
)
117 if ((stable_page_flags(p
) & hwpoison_filter_flags_mask
) ==
118 hwpoison_filter_flags_value
)
125 * This allows stress tests to limit test scope to a collection of tasks
126 * by putting them under some memcg. This prevents killing unrelated/important
127 * processes such as /sbin/init. Note that the target task may share clean
128 * pages with init (eg. libc text), which is harmless. If the target task
129 * share _dirty_ pages with another task B, the test scheme must make sure B
130 * is also included in the memcg. At last, due to race conditions this filter
131 * can only guarantee that the page either belongs to the memcg tasks, or is
135 u64 hwpoison_filter_memcg
;
136 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg
);
137 static int hwpoison_filter_task(struct page
*p
)
139 if (!hwpoison_filter_memcg
)
142 if (page_cgroup_ino(p
) != hwpoison_filter_memcg
)
148 static int hwpoison_filter_task(struct page
*p
) { return 0; }
151 int hwpoison_filter(struct page
*p
)
153 if (!hwpoison_filter_enable
)
156 if (hwpoison_filter_dev(p
))
159 if (hwpoison_filter_flags(p
))
162 if (hwpoison_filter_task(p
))
168 int hwpoison_filter(struct page
*p
)
174 EXPORT_SYMBOL_GPL(hwpoison_filter
);
177 * Send all the processes who have the page mapped a signal.
178 * ``action optional'' if they are not immediately affected by the error
179 * ``action required'' if error happened in current execution context
181 static int kill_proc(struct task_struct
*t
, unsigned long addr
, int trapno
,
182 unsigned long pfn
, struct page
*page
, int flags
)
187 pr_err("MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
188 pfn
, t
->comm
, t
->pid
);
189 si
.si_signo
= SIGBUS
;
191 si
.si_addr
= (void *)addr
;
192 #ifdef __ARCH_SI_TRAPNO
193 si
.si_trapno
= trapno
;
195 si
.si_addr_lsb
= compound_order(compound_head(page
)) + PAGE_SHIFT
;
197 if ((flags
& MF_ACTION_REQUIRED
) && t
->mm
== current
->mm
) {
198 si
.si_code
= BUS_MCEERR_AR
;
199 ret
= force_sig_info(SIGBUS
, &si
, current
);
202 * Don't use force here, it's convenient if the signal
203 * can be temporarily blocked.
204 * This could cause a loop when the user sets SIGBUS
205 * to SIG_IGN, but hopefully no one will do that?
207 si
.si_code
= BUS_MCEERR_AO
;
208 ret
= send_sig_info(SIGBUS
, &si
, t
); /* synchronous? */
211 pr_info("MCE: Error sending signal to %s:%d: %d\n",
212 t
->comm
, t
->pid
, ret
);
217 * When a unknown page type is encountered drain as many buffers as possible
218 * in the hope to turn the page into a LRU or free page, which we can handle.
220 void shake_page(struct page
*p
, int access
)
226 drain_all_pages(page_zone(p
));
227 if (PageLRU(p
) || is_free_buddy_page(p
))
232 * Only call shrink_node_slabs here (which would also shrink
233 * other caches) if access is not potentially fatal.
236 drop_slab_node(page_to_nid(p
));
238 EXPORT_SYMBOL_GPL(shake_page
);
241 * Kill all processes that have a poisoned page mapped and then isolate
245 * Find all processes having the page mapped and kill them.
246 * But we keep a page reference around so that the page is not
247 * actually freed yet.
248 * Then stash the page away
250 * There's no convenient way to get back to mapped processes
251 * from the VMAs. So do a brute-force search over all
254 * Remember that machine checks are not common (or rather
255 * if they are common you have other problems), so this shouldn't
256 * be a performance issue.
258 * Also there are some races possible while we get from the
259 * error detection to actually handle it.
264 struct task_struct
*tsk
;
270 * Failure handling: if we can't find or can't kill a process there's
271 * not much we can do. We just print a message and ignore otherwise.
275 * Schedule a process for later kill.
276 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
277 * TBD would GFP_NOIO be enough?
279 static void add_to_kill(struct task_struct
*tsk
, struct page
*p
,
280 struct vm_area_struct
*vma
,
281 struct list_head
*to_kill
,
282 struct to_kill
**tkc
)
290 tk
= kmalloc(sizeof(struct to_kill
), GFP_ATOMIC
);
292 pr_err("MCE: Out of memory while machine check handling\n");
296 tk
->addr
= page_address_in_vma(p
, vma
);
300 * In theory we don't have to kill when the page was
301 * munmaped. But it could be also a mremap. Since that's
302 * likely very rare kill anyways just out of paranoia, but use
303 * a SIGKILL because the error is not contained anymore.
305 if (tk
->addr
== -EFAULT
) {
306 pr_info("MCE: Unable to find user space address %lx in %s\n",
307 page_to_pfn(p
), tsk
->comm
);
310 get_task_struct(tsk
);
312 list_add_tail(&tk
->nd
, to_kill
);
316 * Kill the processes that have been collected earlier.
318 * Only do anything when DOIT is set, otherwise just free the list
319 * (this is used for clean pages which do not need killing)
320 * Also when FAIL is set do a force kill because something went
323 static void kill_procs(struct list_head
*to_kill
, int forcekill
, int trapno
,
324 int fail
, struct page
*page
, unsigned long pfn
,
327 struct to_kill
*tk
, *next
;
329 list_for_each_entry_safe (tk
, next
, to_kill
, nd
) {
332 * In case something went wrong with munmapping
333 * make sure the process doesn't catch the
334 * signal and then access the memory. Just kill it.
336 if (fail
|| tk
->addr_valid
== 0) {
337 pr_err("MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
338 pfn
, tk
->tsk
->comm
, tk
->tsk
->pid
);
339 force_sig(SIGKILL
, tk
->tsk
);
343 * In theory the process could have mapped
344 * something else on the address in-between. We could
345 * check for that, but we need to tell the
348 else if (kill_proc(tk
->tsk
, tk
->addr
, trapno
,
349 pfn
, page
, flags
) < 0)
350 pr_err("MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
351 pfn
, tk
->tsk
->comm
, tk
->tsk
->pid
);
353 put_task_struct(tk
->tsk
);
359 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
360 * on behalf of the thread group. Return task_struct of the (first found)
361 * dedicated thread if found, and return NULL otherwise.
363 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
364 * have to call rcu_read_lock/unlock() in this function.
366 static struct task_struct
*find_early_kill_thread(struct task_struct
*tsk
)
368 struct task_struct
*t
;
370 for_each_thread(tsk
, t
)
371 if ((t
->flags
& PF_MCE_PROCESS
) && (t
->flags
& PF_MCE_EARLY
))
377 * Determine whether a given process is "early kill" process which expects
378 * to be signaled when some page under the process is hwpoisoned.
379 * Return task_struct of the dedicated thread (main thread unless explicitly
380 * specified) if the process is "early kill," and otherwise returns NULL.
382 static struct task_struct
*task_early_kill(struct task_struct
*tsk
,
385 struct task_struct
*t
;
390 t
= find_early_kill_thread(tsk
);
393 if (sysctl_memory_failure_early_kill
)
399 * Collect processes when the error hit an anonymous page.
401 static void collect_procs_anon(struct page
*page
, struct list_head
*to_kill
,
402 struct to_kill
**tkc
, int force_early
)
404 struct vm_area_struct
*vma
;
405 struct task_struct
*tsk
;
409 av
= page_lock_anon_vma_read(page
);
410 if (av
== NULL
) /* Not actually mapped anymore */
413 pgoff
= page_to_pgoff(page
);
414 read_lock(&tasklist_lock
);
415 for_each_process (tsk
) {
416 struct anon_vma_chain
*vmac
;
417 struct task_struct
*t
= task_early_kill(tsk
, force_early
);
421 anon_vma_interval_tree_foreach(vmac
, &av
->rb_root
,
424 if (!page_mapped_in_vma(page
, vma
))
426 if (vma
->vm_mm
== t
->mm
)
427 add_to_kill(t
, page
, vma
, to_kill
, tkc
);
430 read_unlock(&tasklist_lock
);
431 page_unlock_anon_vma_read(av
);
435 * Collect processes when the error hit a file mapped page.
437 static void collect_procs_file(struct page
*page
, struct list_head
*to_kill
,
438 struct to_kill
**tkc
, int force_early
)
440 struct vm_area_struct
*vma
;
441 struct task_struct
*tsk
;
442 struct address_space
*mapping
= page
->mapping
;
444 i_mmap_lock_read(mapping
);
445 read_lock(&tasklist_lock
);
446 for_each_process(tsk
) {
447 pgoff_t pgoff
= page_to_pgoff(page
);
448 struct task_struct
*t
= task_early_kill(tsk
, force_early
);
452 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
,
455 * Send early kill signal to tasks where a vma covers
456 * the page but the corrupted page is not necessarily
457 * mapped it in its pte.
458 * Assume applications who requested early kill want
459 * to be informed of all such data corruptions.
461 if (vma
->vm_mm
== t
->mm
)
462 add_to_kill(t
, page
, vma
, to_kill
, tkc
);
465 read_unlock(&tasklist_lock
);
466 i_mmap_unlock_read(mapping
);
470 * Collect the processes who have the corrupted page mapped to kill.
471 * This is done in two steps for locking reasons.
472 * First preallocate one tokill structure outside the spin locks,
473 * so that we can kill at least one process reasonably reliable.
475 static void collect_procs(struct page
*page
, struct list_head
*tokill
,
483 tk
= kmalloc(sizeof(struct to_kill
), GFP_NOIO
);
487 collect_procs_anon(page
, tokill
, &tk
, force_early
);
489 collect_procs_file(page
, tokill
, &tk
, force_early
);
493 static const char *action_name
[] = {
494 [MF_IGNORED
] = "Ignored",
495 [MF_FAILED
] = "Failed",
496 [MF_DELAYED
] = "Delayed",
497 [MF_RECOVERED
] = "Recovered",
500 static const char * const action_page_types
[] = {
501 [MF_MSG_KERNEL
] = "reserved kernel page",
502 [MF_MSG_KERNEL_HIGH_ORDER
] = "high-order kernel page",
503 [MF_MSG_SLAB
] = "kernel slab page",
504 [MF_MSG_DIFFERENT_COMPOUND
] = "different compound page after locking",
505 [MF_MSG_POISONED_HUGE
] = "huge page already hardware poisoned",
506 [MF_MSG_HUGE
] = "huge page",
507 [MF_MSG_FREE_HUGE
] = "free huge page",
508 [MF_MSG_UNMAP_FAILED
] = "unmapping failed page",
509 [MF_MSG_DIRTY_SWAPCACHE
] = "dirty swapcache page",
510 [MF_MSG_CLEAN_SWAPCACHE
] = "clean swapcache page",
511 [MF_MSG_DIRTY_MLOCKED_LRU
] = "dirty mlocked LRU page",
512 [MF_MSG_CLEAN_MLOCKED_LRU
] = "clean mlocked LRU page",
513 [MF_MSG_DIRTY_UNEVICTABLE_LRU
] = "dirty unevictable LRU page",
514 [MF_MSG_CLEAN_UNEVICTABLE_LRU
] = "clean unevictable LRU page",
515 [MF_MSG_DIRTY_LRU
] = "dirty LRU page",
516 [MF_MSG_CLEAN_LRU
] = "clean LRU page",
517 [MF_MSG_TRUNCATED_LRU
] = "already truncated LRU page",
518 [MF_MSG_BUDDY
] = "free buddy page",
519 [MF_MSG_BUDDY_2ND
] = "free buddy page (2nd try)",
520 [MF_MSG_UNKNOWN
] = "unknown page",
524 * XXX: It is possible that a page is isolated from LRU cache,
525 * and then kept in swap cache or failed to remove from page cache.
526 * The page count will stop it from being freed by unpoison.
527 * Stress tests should be aware of this memory leak problem.
529 static int delete_from_lru_cache(struct page
*p
)
531 if (!isolate_lru_page(p
)) {
533 * Clear sensible page flags, so that the buddy system won't
534 * complain when the page is unpoison-and-freed.
537 ClearPageUnevictable(p
);
539 * drop the page count elevated by isolate_lru_page()
548 * Error hit kernel page.
549 * Do nothing, try to be lucky and not touch this instead. For a few cases we
550 * could be more sophisticated.
552 static int me_kernel(struct page
*p
, unsigned long pfn
)
558 * Page in unknown state. Do nothing.
560 static int me_unknown(struct page
*p
, unsigned long pfn
)
562 pr_err("MCE %#lx: Unknown page state\n", pfn
);
567 * Clean (or cleaned) page cache page.
569 static int me_pagecache_clean(struct page
*p
, unsigned long pfn
)
573 struct address_space
*mapping
;
575 delete_from_lru_cache(p
);
578 * For anonymous pages we're done the only reference left
579 * should be the one m_f() holds.
585 * Now truncate the page in the page cache. This is really
586 * more like a "temporary hole punch"
587 * Don't do this for block devices when someone else
588 * has a reference, because it could be file system metadata
589 * and that's not safe to truncate.
591 mapping
= page_mapping(p
);
594 * Page has been teared down in the meanwhile
600 * Truncation is a bit tricky. Enable it per file system for now.
602 * Open: to take i_mutex or not for this? Right now we don't.
604 if (mapping
->a_ops
->error_remove_page
) {
605 err
= mapping
->a_ops
->error_remove_page(mapping
, p
);
607 pr_info("MCE %#lx: Failed to punch page: %d\n",
609 } else if (page_has_private(p
) &&
610 !try_to_release_page(p
, GFP_NOIO
)) {
611 pr_info("MCE %#lx: failed to release buffers\n", pfn
);
617 * If the file system doesn't support it just invalidate
618 * This fails on dirty or anything with private pages
620 if (invalidate_inode_page(p
))
623 pr_info("MCE %#lx: Failed to invalidate\n", pfn
);
629 * Dirty pagecache page
630 * Issues: when the error hit a hole page the error is not properly
633 static int me_pagecache_dirty(struct page
*p
, unsigned long pfn
)
635 struct address_space
*mapping
= page_mapping(p
);
638 /* TBD: print more information about the file. */
641 * IO error will be reported by write(), fsync(), etc.
642 * who check the mapping.
643 * This way the application knows that something went
644 * wrong with its dirty file data.
646 * There's one open issue:
648 * The EIO will be only reported on the next IO
649 * operation and then cleared through the IO map.
650 * Normally Linux has two mechanisms to pass IO error
651 * first through the AS_EIO flag in the address space
652 * and then through the PageError flag in the page.
653 * Since we drop pages on memory failure handling the
654 * only mechanism open to use is through AS_AIO.
656 * This has the disadvantage that it gets cleared on
657 * the first operation that returns an error, while
658 * the PageError bit is more sticky and only cleared
659 * when the page is reread or dropped. If an
660 * application assumes it will always get error on
661 * fsync, but does other operations on the fd before
662 * and the page is dropped between then the error
663 * will not be properly reported.
665 * This can already happen even without hwpoisoned
666 * pages: first on metadata IO errors (which only
667 * report through AS_EIO) or when the page is dropped
670 * So right now we assume that the application DTRT on
671 * the first EIO, but we're not worse than other parts
674 mapping_set_error(mapping
, EIO
);
677 return me_pagecache_clean(p
, pfn
);
681 * Clean and dirty swap cache.
683 * Dirty swap cache page is tricky to handle. The page could live both in page
684 * cache and swap cache(ie. page is freshly swapped in). So it could be
685 * referenced concurrently by 2 types of PTEs:
686 * normal PTEs and swap PTEs. We try to handle them consistently by calling
687 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
689 * - clear dirty bit to prevent IO
691 * - but keep in the swap cache, so that when we return to it on
692 * a later page fault, we know the application is accessing
693 * corrupted data and shall be killed (we installed simple
694 * interception code in do_swap_page to catch it).
696 * Clean swap cache pages can be directly isolated. A later page fault will
697 * bring in the known good data from disk.
699 static int me_swapcache_dirty(struct page
*p
, unsigned long pfn
)
702 /* Trigger EIO in shmem: */
703 ClearPageUptodate(p
);
705 if (!delete_from_lru_cache(p
))
711 static int me_swapcache_clean(struct page
*p
, unsigned long pfn
)
713 delete_from_swap_cache(p
);
715 if (!delete_from_lru_cache(p
))
722 * Huge pages. Needs work.
724 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
725 * To narrow down kill region to one page, we need to break up pmd.
727 static int me_huge_page(struct page
*p
, unsigned long pfn
)
730 struct page
*hpage
= compound_head(p
);
732 if (!PageHuge(hpage
))
736 * We can safely recover from error on free or reserved (i.e.
737 * not in-use) hugepage by dequeuing it from freelist.
738 * To check whether a hugepage is in-use or not, we can't use
739 * page->lru because it can be used in other hugepage operations,
740 * such as __unmap_hugepage_range() and gather_surplus_pages().
741 * So instead we use page_mapping() and PageAnon().
742 * We assume that this function is called with page lock held,
743 * so there is no race between isolation and mapping/unmapping.
745 if (!(page_mapping(hpage
) || PageAnon(hpage
))) {
746 res
= dequeue_hwpoisoned_huge_page(hpage
);
754 * Various page states we can handle.
756 * A page state is defined by its current page->flags bits.
757 * The table matches them in order and calls the right handler.
759 * This is quite tricky because we can access page at any time
760 * in its live cycle, so all accesses have to be extremely careful.
762 * This is not complete. More states could be added.
763 * For any missing state don't attempt recovery.
766 #define dirty (1UL << PG_dirty)
767 #define sc (1UL << PG_swapcache)
768 #define unevict (1UL << PG_unevictable)
769 #define mlock (1UL << PG_mlocked)
770 #define writeback (1UL << PG_writeback)
771 #define lru (1UL << PG_lru)
772 #define swapbacked (1UL << PG_swapbacked)
773 #define head (1UL << PG_head)
774 #define slab (1UL << PG_slab)
775 #define reserved (1UL << PG_reserved)
777 static struct page_state
{
780 enum mf_action_page_type type
;
781 int (*action
)(struct page
*p
, unsigned long pfn
);
783 { reserved
, reserved
, MF_MSG_KERNEL
, me_kernel
},
785 * free pages are specially detected outside this table:
786 * PG_buddy pages only make a small fraction of all free pages.
790 * Could in theory check if slab page is free or if we can drop
791 * currently unused objects without touching them. But just
792 * treat it as standard kernel for now.
794 { slab
, slab
, MF_MSG_SLAB
, me_kernel
},
796 { head
, head
, MF_MSG_HUGE
, me_huge_page
},
798 { sc
|dirty
, sc
|dirty
, MF_MSG_DIRTY_SWAPCACHE
, me_swapcache_dirty
},
799 { sc
|dirty
, sc
, MF_MSG_CLEAN_SWAPCACHE
, me_swapcache_clean
},
801 { mlock
|dirty
, mlock
|dirty
, MF_MSG_DIRTY_MLOCKED_LRU
, me_pagecache_dirty
},
802 { mlock
|dirty
, mlock
, MF_MSG_CLEAN_MLOCKED_LRU
, me_pagecache_clean
},
804 { unevict
|dirty
, unevict
|dirty
, MF_MSG_DIRTY_UNEVICTABLE_LRU
, me_pagecache_dirty
},
805 { unevict
|dirty
, unevict
, MF_MSG_CLEAN_UNEVICTABLE_LRU
, me_pagecache_clean
},
807 { lru
|dirty
, lru
|dirty
, MF_MSG_DIRTY_LRU
, me_pagecache_dirty
},
808 { lru
|dirty
, lru
, MF_MSG_CLEAN_LRU
, me_pagecache_clean
},
811 * Catchall entry: must be at end.
813 { 0, 0, MF_MSG_UNKNOWN
, me_unknown
},
828 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
829 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
831 static void action_result(unsigned long pfn
, enum mf_action_page_type type
,
832 enum mf_result result
)
834 trace_memory_failure_event(pfn
, type
, result
);
836 pr_err("MCE %#lx: recovery action for %s: %s\n",
837 pfn
, action_page_types
[type
], action_name
[result
]);
840 static int page_action(struct page_state
*ps
, struct page
*p
,
846 result
= ps
->action(p
, pfn
);
848 count
= page_count(p
) - 1;
849 if (ps
->action
== me_swapcache_dirty
&& result
== MF_DELAYED
)
852 pr_err("MCE %#lx: %s still referenced by %d users\n",
853 pfn
, action_page_types
[ps
->type
], count
);
856 action_result(pfn
, ps
->type
, result
);
858 /* Could do more checks here if page looks ok */
860 * Could adjust zone counters here to correct for the missing page.
863 return (result
== MF_RECOVERED
|| result
== MF_DELAYED
) ? 0 : -EBUSY
;
867 * get_hwpoison_page() - Get refcount for memory error handling:
868 * @page: raw error page (hit by memory error)
870 * Return: return 0 if failed to grab the refcount, otherwise true (some
873 int get_hwpoison_page(struct page
*page
)
875 struct page
*head
= compound_head(page
);
877 if (!PageHuge(head
) && PageTransHuge(head
)) {
879 * Non anonymous thp exists only in allocation/free time. We
880 * can't handle such a case correctly, so let's give it up.
881 * This should be better than triggering BUG_ON when kernel
882 * tries to touch the "partially handled" page.
884 if (!PageAnon(head
)) {
885 pr_err("MCE: %#lx: non anonymous thp\n",
891 return get_page_unless_zero(head
);
893 EXPORT_SYMBOL_GPL(get_hwpoison_page
);
896 * Do all that is necessary to remove user space mappings. Unmap
897 * the pages and send SIGBUS to the processes if the data was dirty.
899 static int hwpoison_user_mappings(struct page
*p
, unsigned long pfn
,
900 int trapno
, int flags
, struct page
**hpagep
)
902 enum ttu_flags ttu
= TTU_UNMAP
| TTU_IGNORE_MLOCK
| TTU_IGNORE_ACCESS
;
903 struct address_space
*mapping
;
906 int kill
= 1, forcekill
;
907 struct page
*hpage
= *hpagep
;
910 * Here we are interested only in user-mapped pages, so skip any
911 * other types of pages.
913 if (PageReserved(p
) || PageSlab(p
))
915 if (!(PageLRU(hpage
) || PageHuge(p
)))
919 * This check implies we don't kill processes if their pages
920 * are in the swap cache early. Those are always late kills.
922 if (!page_mapped(hpage
))
926 pr_err("MCE %#lx: can't handle KSM pages.\n", pfn
);
930 if (PageSwapCache(p
)) {
931 pr_err("MCE %#lx: keeping poisoned page in swap cache\n", pfn
);
932 ttu
|= TTU_IGNORE_HWPOISON
;
936 * Propagate the dirty bit from PTEs to struct page first, because we
937 * need this to decide if we should kill or just drop the page.
938 * XXX: the dirty test could be racy: set_page_dirty() may not always
939 * be called inside page lock (it's recommended but not enforced).
941 mapping
= page_mapping(hpage
);
942 if (!(flags
& MF_MUST_KILL
) && !PageDirty(hpage
) && mapping
&&
943 mapping_cap_writeback_dirty(mapping
)) {
944 if (page_mkclean(hpage
)) {
948 ttu
|= TTU_IGNORE_HWPOISON
;
949 pr_info("MCE %#lx: corrupted page was clean: dropped without side effects\n",
955 * First collect all the processes that have the page
956 * mapped in dirty form. This has to be done before try_to_unmap,
957 * because ttu takes the rmap data structures down.
959 * Error handling: We ignore errors here because
960 * there's nothing that can be done.
963 collect_procs(hpage
, &tokill
, flags
& MF_ACTION_REQUIRED
);
965 ret
= try_to_unmap(hpage
, ttu
);
966 if (ret
!= SWAP_SUCCESS
)
967 pr_err("MCE %#lx: failed to unmap page (mapcount=%d)\n",
968 pfn
, page_mapcount(hpage
));
971 * Now that the dirty bit has been propagated to the
972 * struct page and all unmaps done we can decide if
973 * killing is needed or not. Only kill when the page
974 * was dirty or the process is not restartable,
975 * otherwise the tokill list is merely
976 * freed. When there was a problem unmapping earlier
977 * use a more force-full uncatchable kill to prevent
978 * any accesses to the poisoned memory.
980 forcekill
= PageDirty(hpage
) || (flags
& MF_MUST_KILL
);
981 kill_procs(&tokill
, forcekill
, trapno
,
982 ret
!= SWAP_SUCCESS
, p
, pfn
, flags
);
987 static void set_page_hwpoison_huge_page(struct page
*hpage
)
990 int nr_pages
= 1 << compound_order(hpage
);
991 for (i
= 0; i
< nr_pages
; i
++)
992 SetPageHWPoison(hpage
+ i
);
995 static void clear_page_hwpoison_huge_page(struct page
*hpage
)
998 int nr_pages
= 1 << compound_order(hpage
);
999 for (i
= 0; i
< nr_pages
; i
++)
1000 ClearPageHWPoison(hpage
+ i
);
1004 * memory_failure - Handle memory failure of a page.
1005 * @pfn: Page Number of the corrupted page
1006 * @trapno: Trap number reported in the signal to user space.
1007 * @flags: fine tune action taken
1009 * This function is called by the low level machine check code
1010 * of an architecture when it detects hardware memory corruption
1011 * of a page. It tries its best to recover, which includes
1012 * dropping pages, killing processes etc.
1014 * The function is primarily of use for corruptions that
1015 * happen outside the current execution context (e.g. when
1016 * detected by a background scrubber)
1018 * Must run in process context (e.g. a work queue) with interrupts
1019 * enabled and no spinlocks hold.
1021 int memory_failure(unsigned long pfn
, int trapno
, int flags
)
1023 struct page_state
*ps
;
1026 struct page
*orig_head
;
1028 unsigned int nr_pages
;
1029 unsigned long page_flags
;
1031 if (!sysctl_memory_failure_recovery
)
1032 panic("Memory failure from trap %d on page %lx", trapno
, pfn
);
1034 if (!pfn_valid(pfn
)) {
1035 pr_err("MCE %#lx: memory outside kernel control\n", pfn
);
1039 p
= pfn_to_page(pfn
);
1040 orig_head
= hpage
= compound_head(p
);
1041 if (TestSetPageHWPoison(p
)) {
1042 pr_err("MCE %#lx: already hardware poisoned\n", pfn
);
1047 * Currently errors on hugetlbfs pages are measured in hugepage units,
1048 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1049 * transparent hugepages, they are supposed to be split and error
1050 * measurement is done in normal page units. So nr_pages should be one
1054 nr_pages
= 1 << compound_order(hpage
);
1055 else /* normal page or thp */
1057 num_poisoned_pages_add(nr_pages
);
1060 * We need/can do nothing about count=0 pages.
1061 * 1) it's a free page, and therefore in safe hand:
1062 * prep_new_page() will be the gate keeper.
1063 * 2) it's a free hugepage, which is also safe:
1064 * an affected hugepage will be dequeued from hugepage freelist,
1065 * so there's no concern about reusing it ever after.
1066 * 3) it's part of a non-compound high order page.
1067 * Implies some kernel user: cannot stop them from
1068 * R/W the page; let's pray that the page has been
1069 * used and will be freed some time later.
1070 * In fact it's dangerous to directly bump up page count from 0,
1071 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1073 if (!(flags
& MF_COUNT_INCREASED
) && !get_hwpoison_page(p
)) {
1074 if (is_free_buddy_page(p
)) {
1075 action_result(pfn
, MF_MSG_BUDDY
, MF_DELAYED
);
1077 } else if (PageHuge(hpage
)) {
1079 * Check "filter hit" and "race with other subpage."
1082 if (PageHWPoison(hpage
)) {
1083 if ((hwpoison_filter(p
) && TestClearPageHWPoison(p
))
1084 || (p
!= hpage
&& TestSetPageHWPoison(hpage
))) {
1085 num_poisoned_pages_sub(nr_pages
);
1090 set_page_hwpoison_huge_page(hpage
);
1091 res
= dequeue_hwpoisoned_huge_page(hpage
);
1092 action_result(pfn
, MF_MSG_FREE_HUGE
,
1093 res
? MF_IGNORED
: MF_DELAYED
);
1097 action_result(pfn
, MF_MSG_KERNEL_HIGH_ORDER
, MF_IGNORED
);
1102 if (!PageHuge(p
) && PageTransHuge(hpage
)) {
1104 if (!PageAnon(hpage
) || unlikely(split_huge_page(hpage
))) {
1106 if (!PageAnon(hpage
))
1107 pr_err("MCE: %#lx: non anonymous thp\n", pfn
);
1109 pr_err("MCE: %#lx: thp split failed\n", pfn
);
1110 if (TestClearPageHWPoison(p
))
1111 num_poisoned_pages_sub(nr_pages
);
1112 put_hwpoison_page(p
);
1116 get_hwpoison_page(p
);
1117 put_hwpoison_page(hpage
);
1118 VM_BUG_ON_PAGE(!page_count(p
), p
);
1119 hpage
= compound_head(p
);
1123 * We ignore non-LRU pages for good reasons.
1124 * - PG_locked is only well defined for LRU pages and a few others
1125 * - to avoid races with __SetPageLocked()
1126 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1127 * The check (unnecessarily) ignores LRU pages being isolated and
1128 * walked by the page reclaim code, however that's not a big loss.
1135 * shake_page could have turned it free.
1137 if (is_free_buddy_page(p
)) {
1138 if (flags
& MF_COUNT_INCREASED
)
1139 action_result(pfn
, MF_MSG_BUDDY
, MF_DELAYED
);
1141 action_result(pfn
, MF_MSG_BUDDY_2ND
,
1151 * The page could have changed compound pages during the locking.
1152 * If this happens just bail out.
1154 if (PageCompound(p
) && compound_head(p
) != orig_head
) {
1155 action_result(pfn
, MF_MSG_DIFFERENT_COMPOUND
, MF_IGNORED
);
1161 * We use page flags to determine what action should be taken, but
1162 * the flags can be modified by the error containment action. One
1163 * example is an mlocked page, where PG_mlocked is cleared by
1164 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1165 * correctly, we save a copy of the page flags at this time.
1167 page_flags
= p
->flags
;
1170 * unpoison always clear PG_hwpoison inside page lock
1172 if (!PageHWPoison(p
)) {
1173 pr_err("MCE %#lx: just unpoisoned\n", pfn
);
1174 num_poisoned_pages_sub(nr_pages
);
1176 put_hwpoison_page(hpage
);
1179 if (hwpoison_filter(p
)) {
1180 if (TestClearPageHWPoison(p
))
1181 num_poisoned_pages_sub(nr_pages
);
1183 put_hwpoison_page(hpage
);
1187 if (!PageHuge(p
) && !PageTransTail(p
) && !PageLRU(p
))
1188 goto identify_page_state
;
1191 * For error on the tail page, we should set PG_hwpoison
1192 * on the head page to show that the hugepage is hwpoisoned
1194 if (PageHuge(p
) && PageTail(p
) && TestSetPageHWPoison(hpage
)) {
1195 action_result(pfn
, MF_MSG_POISONED_HUGE
, MF_IGNORED
);
1197 put_hwpoison_page(hpage
);
1201 * Set PG_hwpoison on all pages in an error hugepage,
1202 * because containment is done in hugepage unit for now.
1203 * Since we have done TestSetPageHWPoison() for the head page with
1204 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1207 set_page_hwpoison_huge_page(hpage
);
1210 * It's very difficult to mess with pages currently under IO
1211 * and in many cases impossible, so we just avoid it here.
1213 wait_on_page_writeback(p
);
1216 * Now take care of user space mappings.
1217 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1219 * When the raw error page is thp tail page, hpage points to the raw
1220 * page after thp split.
1222 if (hwpoison_user_mappings(p
, pfn
, trapno
, flags
, &hpage
)
1224 action_result(pfn
, MF_MSG_UNMAP_FAILED
, MF_IGNORED
);
1230 * Torn down by someone else?
1232 if (PageLRU(p
) && !PageSwapCache(p
) && p
->mapping
== NULL
) {
1233 action_result(pfn
, MF_MSG_TRUNCATED_LRU
, MF_IGNORED
);
1238 identify_page_state
:
1241 * The first check uses the current page flags which may not have any
1242 * relevant information. The second check with the saved page flagss is
1243 * carried out only if the first check can't determine the page status.
1245 for (ps
= error_states
;; ps
++)
1246 if ((p
->flags
& ps
->mask
) == ps
->res
)
1249 page_flags
|= (p
->flags
& (1UL << PG_dirty
));
1252 for (ps
= error_states
;; ps
++)
1253 if ((page_flags
& ps
->mask
) == ps
->res
)
1255 res
= page_action(ps
, p
, pfn
);
1260 EXPORT_SYMBOL_GPL(memory_failure
);
1262 #define MEMORY_FAILURE_FIFO_ORDER 4
1263 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1265 struct memory_failure_entry
{
1271 struct memory_failure_cpu
{
1272 DECLARE_KFIFO(fifo
, struct memory_failure_entry
,
1273 MEMORY_FAILURE_FIFO_SIZE
);
1275 struct work_struct work
;
1278 static DEFINE_PER_CPU(struct memory_failure_cpu
, memory_failure_cpu
);
1281 * memory_failure_queue - Schedule handling memory failure of a page.
1282 * @pfn: Page Number of the corrupted page
1283 * @trapno: Trap number reported in the signal to user space.
1284 * @flags: Flags for memory failure handling
1286 * This function is called by the low level hardware error handler
1287 * when it detects hardware memory corruption of a page. It schedules
1288 * the recovering of error page, including dropping pages, killing
1291 * The function is primarily of use for corruptions that
1292 * happen outside the current execution context (e.g. when
1293 * detected by a background scrubber)
1295 * Can run in IRQ context.
1297 void memory_failure_queue(unsigned long pfn
, int trapno
, int flags
)
1299 struct memory_failure_cpu
*mf_cpu
;
1300 unsigned long proc_flags
;
1301 struct memory_failure_entry entry
= {
1307 mf_cpu
= &get_cpu_var(memory_failure_cpu
);
1308 spin_lock_irqsave(&mf_cpu
->lock
, proc_flags
);
1309 if (kfifo_put(&mf_cpu
->fifo
, entry
))
1310 schedule_work_on(smp_processor_id(), &mf_cpu
->work
);
1312 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1314 spin_unlock_irqrestore(&mf_cpu
->lock
, proc_flags
);
1315 put_cpu_var(memory_failure_cpu
);
1317 EXPORT_SYMBOL_GPL(memory_failure_queue
);
1319 static void memory_failure_work_func(struct work_struct
*work
)
1321 struct memory_failure_cpu
*mf_cpu
;
1322 struct memory_failure_entry entry
= { 0, };
1323 unsigned long proc_flags
;
1326 mf_cpu
= this_cpu_ptr(&memory_failure_cpu
);
1328 spin_lock_irqsave(&mf_cpu
->lock
, proc_flags
);
1329 gotten
= kfifo_get(&mf_cpu
->fifo
, &entry
);
1330 spin_unlock_irqrestore(&mf_cpu
->lock
, proc_flags
);
1333 if (entry
.flags
& MF_SOFT_OFFLINE
)
1334 soft_offline_page(pfn_to_page(entry
.pfn
), entry
.flags
);
1336 memory_failure(entry
.pfn
, entry
.trapno
, entry
.flags
);
1340 static int __init
memory_failure_init(void)
1342 struct memory_failure_cpu
*mf_cpu
;
1345 for_each_possible_cpu(cpu
) {
1346 mf_cpu
= &per_cpu(memory_failure_cpu
, cpu
);
1347 spin_lock_init(&mf_cpu
->lock
);
1348 INIT_KFIFO(mf_cpu
->fifo
);
1349 INIT_WORK(&mf_cpu
->work
, memory_failure_work_func
);
1354 core_initcall(memory_failure_init
);
1356 #define unpoison_pr_info(fmt, pfn, rs) \
1358 if (__ratelimit(rs)) \
1359 pr_info(fmt, pfn); \
1363 * unpoison_memory - Unpoison a previously poisoned page
1364 * @pfn: Page number of the to be unpoisoned page
1366 * Software-unpoison a page that has been poisoned by
1367 * memory_failure() earlier.
1369 * This is only done on the software-level, so it only works
1370 * for linux injected failures, not real hardware failures
1372 * Returns 0 for success, otherwise -errno.
1374 int unpoison_memory(unsigned long pfn
)
1379 unsigned int nr_pages
;
1380 static DEFINE_RATELIMIT_STATE(unpoison_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1381 DEFAULT_RATELIMIT_BURST
);
1383 if (!pfn_valid(pfn
))
1386 p
= pfn_to_page(pfn
);
1387 page
= compound_head(p
);
1389 if (!PageHWPoison(p
)) {
1390 unpoison_pr_info("MCE: Page was already unpoisoned %#lx\n",
1395 if (page_count(page
) > 1) {
1396 unpoison_pr_info("MCE: Someone grabs the hwpoison page %#lx\n",
1401 if (page_mapped(page
)) {
1402 unpoison_pr_info("MCE: Someone maps the hwpoison page %#lx\n",
1407 if (page_mapping(page
)) {
1408 unpoison_pr_info("MCE: the hwpoison page has non-NULL mapping %#lx\n",
1414 * unpoison_memory() can encounter thp only when the thp is being
1415 * worked by memory_failure() and the page lock is not held yet.
1416 * In such case, we yield to memory_failure() and make unpoison fail.
1418 if (!PageHuge(page
) && PageTransHuge(page
)) {
1419 unpoison_pr_info("MCE: Memory failure is now running on %#lx\n",
1424 nr_pages
= 1 << compound_order(page
);
1426 if (!get_hwpoison_page(p
)) {
1428 * Since HWPoisoned hugepage should have non-zero refcount,
1429 * race between memory failure and unpoison seems to happen.
1430 * In such case unpoison fails and memory failure runs
1433 if (PageHuge(page
)) {
1434 unpoison_pr_info("MCE: Memory failure is now running on free hugepage %#lx\n",
1438 if (TestClearPageHWPoison(p
))
1439 num_poisoned_pages_dec();
1440 unpoison_pr_info("MCE: Software-unpoisoned free page %#lx\n",
1447 * This test is racy because PG_hwpoison is set outside of page lock.
1448 * That's acceptable because that won't trigger kernel panic. Instead,
1449 * the PG_hwpoison page will be caught and isolated on the entrance to
1450 * the free buddy page pool.
1452 if (TestClearPageHWPoison(page
)) {
1453 unpoison_pr_info("MCE: Software-unpoisoned page %#lx\n",
1455 num_poisoned_pages_sub(nr_pages
);
1458 clear_page_hwpoison_huge_page(page
);
1462 put_hwpoison_page(page
);
1463 if (freeit
&& !(pfn
== my_zero_pfn(0) && page_count(p
) == 1))
1464 put_hwpoison_page(page
);
1468 EXPORT_SYMBOL(unpoison_memory
);
1470 static struct page
*new_page(struct page
*p
, unsigned long private, int **x
)
1472 int nid
= page_to_nid(p
);
1474 return alloc_huge_page_node(page_hstate(compound_head(p
)),
1477 return __alloc_pages_node(nid
, GFP_HIGHUSER_MOVABLE
, 0);
1481 * Safely get reference count of an arbitrary page.
1482 * Returns 0 for a free page, -EIO for a zero refcount page
1483 * that is not free, and 1 for any other page type.
1484 * For 1 the page is returned with increased page count, otherwise not.
1486 static int __get_any_page(struct page
*p
, unsigned long pfn
, int flags
)
1490 if (flags
& MF_COUNT_INCREASED
)
1494 * When the target page is a free hugepage, just remove it
1495 * from free hugepage list.
1497 if (!get_hwpoison_page(p
)) {
1499 pr_info("%s: %#lx free huge page\n", __func__
, pfn
);
1501 } else if (is_free_buddy_page(p
)) {
1502 pr_info("%s: %#lx free buddy page\n", __func__
, pfn
);
1505 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1506 __func__
, pfn
, p
->flags
);
1510 /* Not a free page */
1516 static int get_any_page(struct page
*page
, unsigned long pfn
, int flags
)
1518 int ret
= __get_any_page(page
, pfn
, flags
);
1520 if (ret
== 1 && !PageHuge(page
) && !PageLRU(page
)) {
1524 put_hwpoison_page(page
);
1525 shake_page(page
, 1);
1530 ret
= __get_any_page(page
, pfn
, 0);
1531 if (ret
== 1 && !PageLRU(page
)) {
1532 /* Drop page reference which is from __get_any_page() */
1533 put_hwpoison_page(page
);
1534 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1542 static int soft_offline_huge_page(struct page
*page
, int flags
)
1545 unsigned long pfn
= page_to_pfn(page
);
1546 struct page
*hpage
= compound_head(page
);
1547 LIST_HEAD(pagelist
);
1550 * This double-check of PageHWPoison is to avoid the race with
1551 * memory_failure(). See also comment in __soft_offline_page().
1554 if (PageHWPoison(hpage
)) {
1556 put_hwpoison_page(hpage
);
1557 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn
);
1562 ret
= isolate_huge_page(hpage
, &pagelist
);
1564 * get_any_page() and isolate_huge_page() takes a refcount each,
1565 * so need to drop one here.
1567 put_hwpoison_page(hpage
);
1569 pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn
);
1573 ret
= migrate_pages(&pagelist
, new_page
, NULL
, MPOL_MF_MOVE_ALL
,
1574 MIGRATE_SYNC
, MR_MEMORY_FAILURE
);
1576 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1577 pfn
, ret
, page
->flags
);
1579 * We know that soft_offline_huge_page() tries to migrate
1580 * only one hugepage pointed to by hpage, so we need not
1581 * run through the pagelist here.
1583 putback_active_hugepage(hpage
);
1587 /* overcommit hugetlb page will be freed to buddy */
1588 if (PageHuge(page
)) {
1589 set_page_hwpoison_huge_page(hpage
);
1590 dequeue_hwpoisoned_huge_page(hpage
);
1591 num_poisoned_pages_add(1 << compound_order(hpage
));
1593 SetPageHWPoison(page
);
1594 num_poisoned_pages_inc();
1600 static int __soft_offline_page(struct page
*page
, int flags
)
1603 unsigned long pfn
= page_to_pfn(page
);
1606 * Check PageHWPoison again inside page lock because PageHWPoison
1607 * is set by memory_failure() outside page lock. Note that
1608 * memory_failure() also double-checks PageHWPoison inside page lock,
1609 * so there's no race between soft_offline_page() and memory_failure().
1612 wait_on_page_writeback(page
);
1613 if (PageHWPoison(page
)) {
1615 put_hwpoison_page(page
);
1616 pr_info("soft offline: %#lx page already poisoned\n", pfn
);
1620 * Try to invalidate first. This should work for
1621 * non dirty unmapped page cache pages.
1623 ret
= invalidate_inode_page(page
);
1626 * RED-PEN would be better to keep it isolated here, but we
1627 * would need to fix isolation locking first.
1630 put_hwpoison_page(page
);
1631 pr_info("soft_offline: %#lx: invalidated\n", pfn
);
1632 SetPageHWPoison(page
);
1633 num_poisoned_pages_inc();
1638 * Simple invalidation didn't work.
1639 * Try to migrate to a new page instead. migrate.c
1640 * handles a large number of cases for us.
1642 ret
= isolate_lru_page(page
);
1644 * Drop page reference which is came from get_any_page()
1645 * successful isolate_lru_page() already took another one.
1647 put_hwpoison_page(page
);
1649 LIST_HEAD(pagelist
);
1650 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
1651 page_is_file_cache(page
));
1652 list_add(&page
->lru
, &pagelist
);
1653 ret
= migrate_pages(&pagelist
, new_page
, NULL
, MPOL_MF_MOVE_ALL
,
1654 MIGRATE_SYNC
, MR_MEMORY_FAILURE
);
1656 if (!list_empty(&pagelist
)) {
1657 list_del(&page
->lru
);
1658 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
1659 page_is_file_cache(page
));
1660 putback_lru_page(page
);
1663 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1664 pfn
, ret
, page
->flags
);
1669 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1670 pfn
, ret
, page_count(page
), page
->flags
);
1675 static int soft_offline_in_use_page(struct page
*page
, int flags
)
1678 struct page
*hpage
= compound_head(page
);
1680 if (!PageHuge(page
) && PageTransHuge(hpage
)) {
1682 if (!PageAnon(hpage
) || unlikely(split_huge_page(hpage
))) {
1684 if (!PageAnon(hpage
))
1685 pr_info("soft offline: %#lx: non anonymous thp\n", page_to_pfn(page
));
1687 pr_info("soft offline: %#lx: thp split failed\n", page_to_pfn(page
));
1688 put_hwpoison_page(hpage
);
1692 get_hwpoison_page(page
);
1693 put_hwpoison_page(hpage
);
1697 ret
= soft_offline_huge_page(page
, flags
);
1699 ret
= __soft_offline_page(page
, flags
);
1704 static void soft_offline_free_page(struct page
*page
)
1706 if (PageHuge(page
)) {
1707 struct page
*hpage
= compound_head(page
);
1709 set_page_hwpoison_huge_page(hpage
);
1710 if (!dequeue_hwpoisoned_huge_page(hpage
))
1711 num_poisoned_pages_add(1 << compound_order(hpage
));
1713 if (!TestSetPageHWPoison(page
))
1714 num_poisoned_pages_inc();
1719 * soft_offline_page - Soft offline a page.
1720 * @page: page to offline
1721 * @flags: flags. Same as memory_failure().
1723 * Returns 0 on success, otherwise negated errno.
1725 * Soft offline a page, by migration or invalidation,
1726 * without killing anything. This is for the case when
1727 * a page is not corrupted yet (so it's still valid to access),
1728 * but has had a number of corrected errors and is better taken
1731 * The actual policy on when to do that is maintained by
1734 * This should never impact any application or cause data loss,
1735 * however it might take some time.
1737 * This is not a 100% solution for all memory, but tries to be
1738 * ``good enough'' for the majority of memory.
1740 int soft_offline_page(struct page
*page
, int flags
)
1743 unsigned long pfn
= page_to_pfn(page
);
1745 if (PageHWPoison(page
)) {
1746 pr_info("soft offline: %#lx page already poisoned\n", pfn
);
1747 if (flags
& MF_COUNT_INCREASED
)
1748 put_hwpoison_page(page
);
1753 ret
= get_any_page(page
, pfn
, flags
);
1757 ret
= soft_offline_in_use_page(page
, flags
);
1759 soft_offline_free_page(page
);