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)
188 "MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
189 pfn, t->comm, t->pid);
190 si.si_signo = SIGBUS;
192 si.si_addr = (void *)addr;
193 #ifdef __ARCH_SI_TRAPNO
194 si.si_trapno = trapno;
196 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
198 if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
199 si.si_code = BUS_MCEERR_AR;
200 ret = force_sig_info(SIGBUS, &si, current);
203 * Don't use force here, it's convenient if the signal
204 * can be temporarily blocked.
205 * This could cause a loop when the user sets SIGBUS
206 * to SIG_IGN, but hopefully no one will do that?
208 si.si_code = BUS_MCEERR_AO;
209 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
212 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
213 t->comm, t->pid, ret);
218 * When a unknown page type is encountered drain as many buffers as possible
219 * in the hope to turn the page into a LRU or free page, which we can handle.
221 void shake_page(struct page *p, int access)
227 drain_all_pages(page_zone(p));
228 if (PageLRU(p) || is_free_buddy_page(p))
233 * Only call shrink_node_slabs here (which would also shrink
234 * other caches) if access is not potentially fatal.
237 drop_slab_node(page_to_nid(p));
239 EXPORT_SYMBOL_GPL(shake_page);
242 * Kill all processes that have a poisoned page mapped and then isolate
246 * Find all processes having the page mapped and kill them.
247 * But we keep a page reference around so that the page is not
248 * actually freed yet.
249 * Then stash the page away
251 * There's no convenient way to get back to mapped processes
252 * from the VMAs. So do a brute-force search over all
255 * Remember that machine checks are not common (or rather
256 * if they are common you have other problems), so this shouldn't
257 * be a performance issue.
259 * Also there are some races possible while we get from the
260 * error detection to actually handle it.
265 struct task_struct *tsk;
271 * Failure handling: if we can't find or can't kill a process there's
272 * not much we can do. We just print a message and ignore otherwise.
276 * Schedule a process for later kill.
277 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
278 * TBD would GFP_NOIO be enough?
280 static void add_to_kill(struct task_struct *tsk, struct page *p,
281 struct vm_area_struct *vma,
282 struct list_head *to_kill,
283 struct to_kill **tkc)
291 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
294 "MCE: Out of memory while machine check handling\n");
298 tk->addr = page_address_in_vma(p, vma);
302 * In theory we don't have to kill when the page was
303 * munmaped. But it could be also a mremap. Since that's
304 * likely very rare kill anyways just out of paranoia, but use
305 * a SIGKILL because the error is not contained anymore.
307 if (tk->addr == -EFAULT) {
308 pr_info("MCE: Unable to find user space address %lx in %s\n",
309 page_to_pfn(p), tsk->comm);
312 get_task_struct(tsk);
314 list_add_tail(&tk->nd, to_kill);
318 * Kill the processes that have been collected earlier.
320 * Only do anything when DOIT is set, otherwise just free the list
321 * (this is used for clean pages which do not need killing)
322 * Also when FAIL is set do a force kill because something went
325 static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
326 int fail, struct page *page, unsigned long pfn,
329 struct to_kill *tk, *next;
331 list_for_each_entry_safe (tk, next, to_kill, nd) {
334 * In case something went wrong with munmapping
335 * make sure the process doesn't catch the
336 * signal and then access the memory. Just kill it.
338 if (fail || tk->addr_valid == 0) {
340 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
341 pfn, tk->tsk->comm, tk->tsk->pid);
342 force_sig(SIGKILL, tk->tsk);
346 * In theory the process could have mapped
347 * something else on the address in-between. We could
348 * check for that, but we need to tell the
351 else if (kill_proc(tk->tsk, tk->addr, trapno,
352 pfn, page, flags) < 0)
354 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
355 pfn, tk->tsk->comm, tk->tsk->pid);
357 put_task_struct(tk->tsk);
363 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
364 * on behalf of the thread group. Return task_struct of the (first found)
365 * dedicated thread if found, and return NULL otherwise.
367 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
368 * have to call rcu_read_lock/unlock() in this function.
370 static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
372 struct task_struct *t;
374 for_each_thread(tsk, t)
375 if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
381 * Determine whether a given process is "early kill" process which expects
382 * to be signaled when some page under the process is hwpoisoned.
383 * Return task_struct of the dedicated thread (main thread unless explicitly
384 * specified) if the process is "early kill," and otherwise returns NULL.
386 static struct task_struct *task_early_kill(struct task_struct *tsk,
389 struct task_struct *t;
394 t = find_early_kill_thread(tsk);
397 if (sysctl_memory_failure_early_kill)
403 * Collect processes when the error hit an anonymous page.
405 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
406 struct to_kill **tkc, int force_early)
408 struct vm_area_struct *vma;
409 struct task_struct *tsk;
413 av = page_lock_anon_vma_read(page);
414 if (av == NULL) /* Not actually mapped anymore */
417 pgoff = page_to_pgoff(page);
418 read_lock(&tasklist_lock);
419 for_each_process (tsk) {
420 struct anon_vma_chain *vmac;
421 struct task_struct *t = task_early_kill(tsk, force_early);
425 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
428 if (!page_mapped_in_vma(page, vma))
430 if (vma->vm_mm == t->mm)
431 add_to_kill(t, page, vma, to_kill, tkc);
434 read_unlock(&tasklist_lock);
435 page_unlock_anon_vma_read(av);
439 * Collect processes when the error hit a file mapped page.
441 static void collect_procs_file(struct page *page, struct list_head *to_kill,
442 struct to_kill **tkc, int force_early)
444 struct vm_area_struct *vma;
445 struct task_struct *tsk;
446 struct address_space *mapping = page->mapping;
448 i_mmap_lock_read(mapping);
449 read_lock(&tasklist_lock);
450 for_each_process(tsk) {
451 pgoff_t pgoff = page_to_pgoff(page);
452 struct task_struct *t = task_early_kill(tsk, force_early);
456 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
459 * Send early kill signal to tasks where a vma covers
460 * the page but the corrupted page is not necessarily
461 * mapped it in its pte.
462 * Assume applications who requested early kill want
463 * to be informed of all such data corruptions.
465 if (vma->vm_mm == t->mm)
466 add_to_kill(t, page, vma, to_kill, tkc);
469 read_unlock(&tasklist_lock);
470 i_mmap_unlock_read(mapping);
474 * Collect the processes who have the corrupted page mapped to kill.
475 * This is done in two steps for locking reasons.
476 * First preallocate one tokill structure outside the spin locks,
477 * so that we can kill at least one process reasonably reliable.
479 static void collect_procs(struct page *page, struct list_head *tokill,
487 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
491 collect_procs_anon(page, tokill, &tk, force_early);
493 collect_procs_file(page, tokill, &tk, force_early);
497 static const char *action_name[] = {
498 [MF_IGNORED] = "Ignored",
499 [MF_FAILED] = "Failed",
500 [MF_DELAYED] = "Delayed",
501 [MF_RECOVERED] = "Recovered",
504 static const char * const action_page_types[] = {
505 [MF_MSG_KERNEL] = "reserved kernel page",
506 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
507 [MF_MSG_SLAB] = "kernel slab page",
508 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
509 [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned",
510 [MF_MSG_HUGE] = "huge page",
511 [MF_MSG_FREE_HUGE] = "free huge page",
512 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
513 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
514 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
515 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
516 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
517 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
518 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
519 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
520 [MF_MSG_CLEAN_LRU] = "clean LRU page",
521 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
522 [MF_MSG_BUDDY] = "free buddy page",
523 [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)",
524 [MF_MSG_UNKNOWN] = "unknown page",
528 * XXX: It is possible that a page is isolated from LRU cache,
529 * and then kept in swap cache or failed to remove from page cache.
530 * The page count will stop it from being freed by unpoison.
531 * Stress tests should be aware of this memory leak problem.
533 static int delete_from_lru_cache(struct page *p)
535 if (!isolate_lru_page(p)) {
537 * Clear sensible page flags, so that the buddy system won't
538 * complain when the page is unpoison-and-freed.
541 ClearPageUnevictable(p);
543 * drop the page count elevated by isolate_lru_page()
545 page_cache_release(p);
552 * Error hit kernel page.
553 * Do nothing, try to be lucky and not touch this instead. For a few cases we
554 * could be more sophisticated.
556 static int me_kernel(struct page *p, unsigned long pfn)
562 * Page in unknown state. Do nothing.
564 static int me_unknown(struct page *p, unsigned long pfn)
566 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
571 * Clean (or cleaned) page cache page.
573 static int me_pagecache_clean(struct page *p, unsigned long pfn)
577 struct address_space *mapping;
579 delete_from_lru_cache(p);
582 * For anonymous pages we're done the only reference left
583 * should be the one m_f() holds.
589 * Now truncate the page in the page cache. This is really
590 * more like a "temporary hole punch"
591 * Don't do this for block devices when someone else
592 * has a reference, because it could be file system metadata
593 * and that's not safe to truncate.
595 mapping = page_mapping(p);
598 * Page has been teared down in the meanwhile
604 * Truncation is a bit tricky. Enable it per file system for now.
606 * Open: to take i_mutex or not for this? Right now we don't.
608 if (mapping->a_ops->error_remove_page) {
609 err = mapping->a_ops->error_remove_page(mapping, p);
611 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
613 } else if (page_has_private(p) &&
614 !try_to_release_page(p, GFP_NOIO)) {
615 pr_info("MCE %#lx: failed to release buffers\n", pfn);
621 * If the file system doesn't support it just invalidate
622 * This fails on dirty or anything with private pages
624 if (invalidate_inode_page(p))
627 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
634 * Dirty pagecache page
635 * Issues: when the error hit a hole page the error is not properly
638 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
640 struct address_space *mapping = page_mapping(p);
643 /* TBD: print more information about the file. */
646 * IO error will be reported by write(), fsync(), etc.
647 * who check the mapping.
648 * This way the application knows that something went
649 * wrong with its dirty file data.
651 * There's one open issue:
653 * The EIO will be only reported on the next IO
654 * operation and then cleared through the IO map.
655 * Normally Linux has two mechanisms to pass IO error
656 * first through the AS_EIO flag in the address space
657 * and then through the PageError flag in the page.
658 * Since we drop pages on memory failure handling the
659 * only mechanism open to use is through AS_AIO.
661 * This has the disadvantage that it gets cleared on
662 * the first operation that returns an error, while
663 * the PageError bit is more sticky and only cleared
664 * when the page is reread or dropped. If an
665 * application assumes it will always get error on
666 * fsync, but does other operations on the fd before
667 * and the page is dropped between then the error
668 * will not be properly reported.
670 * This can already happen even without hwpoisoned
671 * pages: first on metadata IO errors (which only
672 * report through AS_EIO) or when the page is dropped
675 * So right now we assume that the application DTRT on
676 * the first EIO, but we're not worse than other parts
679 mapping_set_error(mapping, EIO);
682 return me_pagecache_clean(p, pfn);
686 * Clean and dirty swap cache.
688 * Dirty swap cache page is tricky to handle. The page could live both in page
689 * cache and swap cache(ie. page is freshly swapped in). So it could be
690 * referenced concurrently by 2 types of PTEs:
691 * normal PTEs and swap PTEs. We try to handle them consistently by calling
692 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
694 * - clear dirty bit to prevent IO
696 * - but keep in the swap cache, so that when we return to it on
697 * a later page fault, we know the application is accessing
698 * corrupted data and shall be killed (we installed simple
699 * interception code in do_swap_page to catch it).
701 * Clean swap cache pages can be directly isolated. A later page fault will
702 * bring in the known good data from disk.
704 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
707 /* Trigger EIO in shmem: */
708 ClearPageUptodate(p);
710 if (!delete_from_lru_cache(p))
716 static int me_swapcache_clean(struct page *p, unsigned long pfn)
718 delete_from_swap_cache(p);
720 if (!delete_from_lru_cache(p))
727 * Huge pages. Needs work.
729 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
730 * To narrow down kill region to one page, we need to break up pmd.
732 static int me_huge_page(struct page *p, unsigned long pfn)
735 struct page *hpage = compound_head(p);
737 if (!PageHuge(hpage))
741 * We can safely recover from error on free or reserved (i.e.
742 * not in-use) hugepage by dequeuing it from freelist.
743 * To check whether a hugepage is in-use or not, we can't use
744 * page->lru because it can be used in other hugepage operations,
745 * such as __unmap_hugepage_range() and gather_surplus_pages().
746 * So instead we use page_mapping() and PageAnon().
747 * We assume that this function is called with page lock held,
748 * so there is no race between isolation and mapping/unmapping.
750 if (!(page_mapping(hpage) || PageAnon(hpage))) {
751 res = dequeue_hwpoisoned_huge_page(hpage);
759 * Various page states we can handle.
761 * A page state is defined by its current page->flags bits.
762 * The table matches them in order and calls the right handler.
764 * This is quite tricky because we can access page at any time
765 * in its live cycle, so all accesses have to be extremely careful.
767 * This is not complete. More states could be added.
768 * For any missing state don't attempt recovery.
771 #define dirty (1UL << PG_dirty)
772 #define sc (1UL << PG_swapcache)
773 #define unevict (1UL << PG_unevictable)
774 #define mlock (1UL << PG_mlocked)
775 #define writeback (1UL << PG_writeback)
776 #define lru (1UL << PG_lru)
777 #define swapbacked (1UL << PG_swapbacked)
778 #define head (1UL << PG_head)
779 #define slab (1UL << PG_slab)
780 #define reserved (1UL << PG_reserved)
782 static struct page_state {
785 enum mf_action_page_type type;
786 int (*action)(struct page *p, unsigned long pfn);
788 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
790 * free pages are specially detected outside this table:
791 * PG_buddy pages only make a small fraction of all free pages.
795 * Could in theory check if slab page is free or if we can drop
796 * currently unused objects without touching them. But just
797 * treat it as standard kernel for now.
799 { slab, slab, MF_MSG_SLAB, me_kernel },
801 { head, head, MF_MSG_HUGE, me_huge_page },
803 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
804 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
806 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
807 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
809 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
810 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
812 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
813 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
816 * Catchall entry: must be at end.
818 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
835 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
836 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
838 static void action_result(unsigned long pfn, enum mf_action_page_type type,
839 enum mf_result result)
841 trace_memory_failure_event(pfn, type, result);
843 pr_err("MCE %#lx: recovery action for %s: %s\n",
844 pfn, action_page_types[type], action_name[result]);
847 static int page_action(struct page_state *ps, struct page *p,
853 result = ps->action(p, pfn);
855 count = page_count(p) - 1;
856 if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
860 "MCE %#lx: %s still referenced by %d users\n",
861 pfn, action_page_types[ps->type], count);
864 action_result(pfn, ps->type, result);
866 /* Could do more checks here if page looks ok */
868 * Could adjust zone counters here to correct for the missing page.
871 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
875 * get_hwpoison_page() - Get refcount for memory error handling:
876 * @page: raw error page (hit by memory error)
878 * Return: return 0 if failed to grab the refcount, otherwise true (some
881 int get_hwpoison_page(struct page *page)
883 struct page *head = compound_head(page);
886 return get_page_unless_zero(head);
889 * Thp tail page has special refcounting rule (refcount of tail pages
890 * is stored in ->_mapcount,) so we can't call get_page_unless_zero()
891 * directly for tail pages.
893 if (PageTransHuge(head)) {
895 * Non anonymous thp exists only in allocation/free time. We
896 * can't handle such a case correctly, so let's give it up.
897 * This should be better than triggering BUG_ON when kernel
898 * tries to touch the "partially handled" page.
900 if (!PageAnon(head)) {
901 pr_err("MCE: %#lx: non anonymous thp\n",
906 if (get_page_unless_zero(head)) {
915 return get_page_unless_zero(page);
917 EXPORT_SYMBOL_GPL(get_hwpoison_page);
920 * put_hwpoison_page() - Put refcount for memory error handling:
921 * @page: raw error page (hit by memory error)
923 void put_hwpoison_page(struct page *page)
925 struct page *head = compound_head(page);
927 if (PageHuge(head)) {
932 if (PageTransHuge(head))
938 EXPORT_SYMBOL_GPL(put_hwpoison_page);
941 * Do all that is necessary to remove user space mappings. Unmap
942 * the pages and send SIGBUS to the processes if the data was dirty.
944 static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
945 int trapno, int flags, struct page **hpagep)
947 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
948 struct address_space *mapping;
951 int kill = 1, forcekill;
952 struct page *hpage = *hpagep;
955 * Here we are interested only in user-mapped pages, so skip any
956 * other types of pages.
958 if (PageReserved(p) || PageSlab(p))
960 if (!(PageLRU(hpage) || PageHuge(p)))
964 * This check implies we don't kill processes if their pages
965 * are in the swap cache early. Those are always late kills.
967 if (!page_mapped(hpage))
971 pr_err("MCE %#lx: can't handle KSM pages.\n", pfn);
975 if (PageSwapCache(p)) {
977 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
978 ttu |= TTU_IGNORE_HWPOISON;
982 * Propagate the dirty bit from PTEs to struct page first, because we
983 * need this to decide if we should kill or just drop the page.
984 * XXX: the dirty test could be racy: set_page_dirty() may not always
985 * be called inside page lock (it's recommended but not enforced).
987 mapping = page_mapping(hpage);
988 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
989 mapping_cap_writeback_dirty(mapping)) {
990 if (page_mkclean(hpage)) {
994 ttu |= TTU_IGNORE_HWPOISON;
996 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
1002 * First collect all the processes that have the page
1003 * mapped in dirty form. This has to be done before try_to_unmap,
1004 * because ttu takes the rmap data structures down.
1006 * Error handling: We ignore errors here because
1007 * there's nothing that can be done.
1010 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
1012 ret = try_to_unmap(hpage, ttu);
1013 if (ret != SWAP_SUCCESS)
1014 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
1015 pfn, page_mapcount(hpage));
1018 * Now that the dirty bit has been propagated to the
1019 * struct page and all unmaps done we can decide if
1020 * killing is needed or not. Only kill when the page
1021 * was dirty or the process is not restartable,
1022 * otherwise the tokill list is merely
1023 * freed. When there was a problem unmapping earlier
1024 * use a more force-full uncatchable kill to prevent
1025 * any accesses to the poisoned memory.
1027 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
1028 kill_procs(&tokill, forcekill, trapno,
1029 ret != SWAP_SUCCESS, p, pfn, flags);
1034 static void set_page_hwpoison_huge_page(struct page *hpage)
1037 int nr_pages = 1 << compound_order(hpage);
1038 for (i = 0; i < nr_pages; i++)
1039 SetPageHWPoison(hpage + i);
1042 static void clear_page_hwpoison_huge_page(struct page *hpage)
1045 int nr_pages = 1 << compound_order(hpage);
1046 for (i = 0; i < nr_pages; i++)
1047 ClearPageHWPoison(hpage + i);
1051 * memory_failure - Handle memory failure of a page.
1052 * @pfn: Page Number of the corrupted page
1053 * @trapno: Trap number reported in the signal to user space.
1054 * @flags: fine tune action taken
1056 * This function is called by the low level machine check code
1057 * of an architecture when it detects hardware memory corruption
1058 * of a page. It tries its best to recover, which includes
1059 * dropping pages, killing processes etc.
1061 * The function is primarily of use for corruptions that
1062 * happen outside the current execution context (e.g. when
1063 * detected by a background scrubber)
1065 * Must run in process context (e.g. a work queue) with interrupts
1066 * enabled and no spinlocks hold.
1068 int memory_failure(unsigned long pfn, int trapno, int flags)
1070 struct page_state *ps;
1073 struct page *orig_head;
1075 unsigned int nr_pages;
1076 unsigned long page_flags;
1078 if (!sysctl_memory_failure_recovery)
1079 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1081 if (!pfn_valid(pfn)) {
1083 "MCE %#lx: memory outside kernel control\n",
1088 p = pfn_to_page(pfn);
1089 orig_head = hpage = compound_head(p);
1090 if (TestSetPageHWPoison(p)) {
1091 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
1096 * Currently errors on hugetlbfs pages are measured in hugepage units,
1097 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1098 * transparent hugepages, they are supposed to be split and error
1099 * measurement is done in normal page units. So nr_pages should be one
1103 nr_pages = 1 << compound_order(hpage);
1104 else /* normal page or thp */
1106 num_poisoned_pages_add(nr_pages);
1109 * We need/can do nothing about count=0 pages.
1110 * 1) it's a free page, and therefore in safe hand:
1111 * prep_new_page() will be the gate keeper.
1112 * 2) it's a free hugepage, which is also safe:
1113 * an affected hugepage will be dequeued from hugepage freelist,
1114 * so there's no concern about reusing it ever after.
1115 * 3) it's part of a non-compound high order page.
1116 * Implies some kernel user: cannot stop them from
1117 * R/W the page; let's pray that the page has been
1118 * used and will be freed some time later.
1119 * In fact it's dangerous to directly bump up page count from 0,
1120 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1122 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
1123 if (is_free_buddy_page(p)) {
1124 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1126 } else if (PageHuge(hpage)) {
1128 * Check "filter hit" and "race with other subpage."
1131 if (PageHWPoison(hpage)) {
1132 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1133 || (p != hpage && TestSetPageHWPoison(hpage))) {
1134 num_poisoned_pages_sub(nr_pages);
1139 set_page_hwpoison_huge_page(hpage);
1140 res = dequeue_hwpoisoned_huge_page(hpage);
1141 action_result(pfn, MF_MSG_FREE_HUGE,
1142 res ? MF_IGNORED : MF_DELAYED);
1146 action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
1151 if (!PageHuge(p) && PageTransHuge(hpage)) {
1152 if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) {
1153 if (!PageAnon(hpage))
1154 pr_err("MCE: %#lx: non anonymous thp\n", pfn);
1156 pr_err("MCE: %#lx: thp split failed\n", pfn);
1157 if (TestClearPageHWPoison(p))
1158 num_poisoned_pages_sub(nr_pages);
1159 put_hwpoison_page(p);
1162 VM_BUG_ON_PAGE(!page_count(p), p);
1163 hpage = compound_head(p);
1167 * We ignore non-LRU pages for good reasons.
1168 * - PG_locked is only well defined for LRU pages and a few others
1169 * - to avoid races with __set_page_locked()
1170 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1171 * The check (unnecessarily) ignores LRU pages being isolated and
1172 * walked by the page reclaim code, however that's not a big loss.
1179 * shake_page could have turned it free.
1181 if (is_free_buddy_page(p)) {
1182 if (flags & MF_COUNT_INCREASED)
1183 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1185 action_result(pfn, MF_MSG_BUDDY_2ND,
1195 * The page could have changed compound pages during the locking.
1196 * If this happens just bail out.
1198 if (PageCompound(p) && compound_head(p) != orig_head) {
1199 action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
1205 * We use page flags to determine what action should be taken, but
1206 * the flags can be modified by the error containment action. One
1207 * example is an mlocked page, where PG_mlocked is cleared by
1208 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1209 * correctly, we save a copy of the page flags at this time.
1211 page_flags = p->flags;
1214 * unpoison always clear PG_hwpoison inside page lock
1216 if (!PageHWPoison(p)) {
1217 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1218 num_poisoned_pages_sub(nr_pages);
1220 put_hwpoison_page(hpage);
1223 if (hwpoison_filter(p)) {
1224 if (TestClearPageHWPoison(p))
1225 num_poisoned_pages_sub(nr_pages);
1227 put_hwpoison_page(hpage);
1231 if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p))
1232 goto identify_page_state;
1235 * For error on the tail page, we should set PG_hwpoison
1236 * on the head page to show that the hugepage is hwpoisoned
1238 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1239 action_result(pfn, MF_MSG_POISONED_HUGE, MF_IGNORED);
1241 put_hwpoison_page(hpage);
1245 * Set PG_hwpoison on all pages in an error hugepage,
1246 * because containment is done in hugepage unit for now.
1247 * Since we have done TestSetPageHWPoison() for the head page with
1248 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1251 set_page_hwpoison_huge_page(hpage);
1254 * It's very difficult to mess with pages currently under IO
1255 * and in many cases impossible, so we just avoid it here.
1257 wait_on_page_writeback(p);
1260 * Now take care of user space mappings.
1261 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1263 * When the raw error page is thp tail page, hpage points to the raw
1264 * page after thp split.
1266 if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)
1268 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1274 * Torn down by someone else?
1276 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1277 action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
1282 identify_page_state:
1285 * The first check uses the current page flags which may not have any
1286 * relevant information. The second check with the saved page flagss is
1287 * carried out only if the first check can't determine the page status.
1289 for (ps = error_states;; ps++)
1290 if ((p->flags & ps->mask) == ps->res)
1293 page_flags |= (p->flags & (1UL << PG_dirty));
1296 for (ps = error_states;; ps++)
1297 if ((page_flags & ps->mask) == ps->res)
1299 res = page_action(ps, p, pfn);
1304 EXPORT_SYMBOL_GPL(memory_failure);
1306 #define MEMORY_FAILURE_FIFO_ORDER 4
1307 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1309 struct memory_failure_entry {
1315 struct memory_failure_cpu {
1316 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1317 MEMORY_FAILURE_FIFO_SIZE);
1319 struct work_struct work;
1322 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1325 * memory_failure_queue - Schedule handling memory failure of a page.
1326 * @pfn: Page Number of the corrupted page
1327 * @trapno: Trap number reported in the signal to user space.
1328 * @flags: Flags for memory failure handling
1330 * This function is called by the low level hardware error handler
1331 * when it detects hardware memory corruption of a page. It schedules
1332 * the recovering of error page, including dropping pages, killing
1335 * The function is primarily of use for corruptions that
1336 * happen outside the current execution context (e.g. when
1337 * detected by a background scrubber)
1339 * Can run in IRQ context.
1341 void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1343 struct memory_failure_cpu *mf_cpu;
1344 unsigned long proc_flags;
1345 struct memory_failure_entry entry = {
1351 mf_cpu = &get_cpu_var(memory_failure_cpu);
1352 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1353 if (kfifo_put(&mf_cpu->fifo, entry))
1354 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1356 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1358 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1359 put_cpu_var(memory_failure_cpu);
1361 EXPORT_SYMBOL_GPL(memory_failure_queue);
1363 static void memory_failure_work_func(struct work_struct *work)
1365 struct memory_failure_cpu *mf_cpu;
1366 struct memory_failure_entry entry = { 0, };
1367 unsigned long proc_flags;
1370 mf_cpu = this_cpu_ptr(&memory_failure_cpu);
1372 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1373 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1374 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1377 if (entry.flags & MF_SOFT_OFFLINE)
1378 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1380 memory_failure(entry.pfn, entry.trapno, entry.flags);
1384 static int __init memory_failure_init(void)
1386 struct memory_failure_cpu *mf_cpu;
1389 for_each_possible_cpu(cpu) {
1390 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1391 spin_lock_init(&mf_cpu->lock);
1392 INIT_KFIFO(mf_cpu->fifo);
1393 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1398 core_initcall(memory_failure_init);
1400 #define unpoison_pr_info(fmt, pfn, rs) \
1402 if (__ratelimit(rs)) \
1403 pr_info(fmt, pfn); \
1407 * unpoison_memory - Unpoison a previously poisoned page
1408 * @pfn: Page number of the to be unpoisoned page
1410 * Software-unpoison a page that has been poisoned by
1411 * memory_failure() earlier.
1413 * This is only done on the software-level, so it only works
1414 * for linux injected failures, not real hardware failures
1416 * Returns 0 for success, otherwise -errno.
1418 int unpoison_memory(unsigned long pfn)
1423 unsigned int nr_pages;
1424 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
1425 DEFAULT_RATELIMIT_BURST);
1427 if (!pfn_valid(pfn))
1430 p = pfn_to_page(pfn);
1431 page = compound_head(p);
1433 if (!PageHWPoison(p)) {
1434 unpoison_pr_info("MCE: Page was already unpoisoned %#lx\n",
1439 if (page_count(page) > 1) {
1440 unpoison_pr_info("MCE: Someone grabs the hwpoison page %#lx\n",
1445 if (page_mapped(page)) {
1446 unpoison_pr_info("MCE: Someone maps the hwpoison page %#lx\n",
1451 if (page_mapping(page)) {
1452 unpoison_pr_info("MCE: the hwpoison page has non-NULL mapping %#lx\n",
1458 * unpoison_memory() can encounter thp only when the thp is being
1459 * worked by memory_failure() and the page lock is not held yet.
1460 * In such case, we yield to memory_failure() and make unpoison fail.
1462 if (!PageHuge(page) && PageTransHuge(page)) {
1463 unpoison_pr_info("MCE: Memory failure is now running on %#lx\n",
1468 nr_pages = 1 << compound_order(page);
1470 if (!get_hwpoison_page(p)) {
1472 * Since HWPoisoned hugepage should have non-zero refcount,
1473 * race between memory failure and unpoison seems to happen.
1474 * In such case unpoison fails and memory failure runs
1477 if (PageHuge(page)) {
1478 unpoison_pr_info("MCE: Memory failure is now running on free hugepage %#lx\n",
1482 if (TestClearPageHWPoison(p))
1483 num_poisoned_pages_dec();
1484 unpoison_pr_info("MCE: Software-unpoisoned free page %#lx\n",
1491 * This test is racy because PG_hwpoison is set outside of page lock.
1492 * That's acceptable because that won't trigger kernel panic. Instead,
1493 * the PG_hwpoison page will be caught and isolated on the entrance to
1494 * the free buddy page pool.
1496 if (TestClearPageHWPoison(page)) {
1497 unpoison_pr_info("MCE: Software-unpoisoned page %#lx\n",
1499 num_poisoned_pages_sub(nr_pages);
1502 clear_page_hwpoison_huge_page(page);
1506 put_hwpoison_page(page);
1507 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
1508 put_hwpoison_page(page);
1512 EXPORT_SYMBOL(unpoison_memory);
1514 static struct page *new_page(struct page *p, unsigned long private, int **x)
1516 int nid = page_to_nid(p);
1518 return alloc_huge_page_node(page_hstate(compound_head(p)),
1521 return __alloc_pages_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1525 * Safely get reference count of an arbitrary page.
1526 * Returns 0 for a free page, -EIO for a zero refcount page
1527 * that is not free, and 1 for any other page type.
1528 * For 1 the page is returned with increased page count, otherwise not.
1530 static int __get_any_page(struct page *p, unsigned long pfn, int flags)
1534 if (flags & MF_COUNT_INCREASED)
1538 * When the target page is a free hugepage, just remove it
1539 * from free hugepage list.
1541 if (!get_hwpoison_page(p)) {
1543 pr_info("%s: %#lx free huge page\n", __func__, pfn);
1545 } else if (is_free_buddy_page(p)) {
1546 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1549 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1550 __func__, pfn, p->flags);
1554 /* Not a free page */
1560 static int get_any_page(struct page *page, unsigned long pfn, int flags)
1562 int ret = __get_any_page(page, pfn, flags);
1564 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1568 put_hwpoison_page(page);
1569 shake_page(page, 1);
1574 ret = __get_any_page(page, pfn, 0);
1575 if (ret == 1 && !PageLRU(page)) {
1576 /* Drop page reference which is from __get_any_page() */
1577 put_hwpoison_page(page);
1578 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1586 static int soft_offline_huge_page(struct page *page, int flags)
1589 unsigned long pfn = page_to_pfn(page);
1590 struct page *hpage = compound_head(page);
1591 LIST_HEAD(pagelist);
1594 * This double-check of PageHWPoison is to avoid the race with
1595 * memory_failure(). See also comment in __soft_offline_page().
1598 if (PageHWPoison(hpage)) {
1600 put_hwpoison_page(hpage);
1601 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1606 ret = isolate_huge_page(hpage, &pagelist);
1608 * get_any_page() and isolate_huge_page() takes a refcount each,
1609 * so need to drop one here.
1611 put_hwpoison_page(hpage);
1613 pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn);
1617 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1618 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1620 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1621 pfn, ret, page->flags);
1623 * We know that soft_offline_huge_page() tries to migrate
1624 * only one hugepage pointed to by hpage, so we need not
1625 * run through the pagelist here.
1627 putback_active_hugepage(hpage);
1631 /* overcommit hugetlb page will be freed to buddy */
1632 if (PageHuge(page)) {
1633 set_page_hwpoison_huge_page(hpage);
1634 dequeue_hwpoisoned_huge_page(hpage);
1635 num_poisoned_pages_add(1 << compound_order(hpage));
1637 SetPageHWPoison(page);
1638 num_poisoned_pages_inc();
1644 static int __soft_offline_page(struct page *page, int flags)
1647 unsigned long pfn = page_to_pfn(page);
1650 * Check PageHWPoison again inside page lock because PageHWPoison
1651 * is set by memory_failure() outside page lock. Note that
1652 * memory_failure() also double-checks PageHWPoison inside page lock,
1653 * so there's no race between soft_offline_page() and memory_failure().
1656 wait_on_page_writeback(page);
1657 if (PageHWPoison(page)) {
1659 put_hwpoison_page(page);
1660 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1664 * Try to invalidate first. This should work for
1665 * non dirty unmapped page cache pages.
1667 ret = invalidate_inode_page(page);
1670 * RED-PEN would be better to keep it isolated here, but we
1671 * would need to fix isolation locking first.
1674 put_hwpoison_page(page);
1675 pr_info("soft_offline: %#lx: invalidated\n", pfn);
1676 SetPageHWPoison(page);
1677 num_poisoned_pages_inc();
1682 * Simple invalidation didn't work.
1683 * Try to migrate to a new page instead. migrate.c
1684 * handles a large number of cases for us.
1686 ret = isolate_lru_page(page);
1688 * Drop page reference which is came from get_any_page()
1689 * successful isolate_lru_page() already took another one.
1691 put_hwpoison_page(page);
1693 LIST_HEAD(pagelist);
1694 inc_zone_page_state(page, NR_ISOLATED_ANON +
1695 page_is_file_cache(page));
1696 list_add(&page->lru, &pagelist);
1697 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1698 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1700 if (!list_empty(&pagelist)) {
1701 list_del(&page->lru);
1702 dec_zone_page_state(page, NR_ISOLATED_ANON +
1703 page_is_file_cache(page));
1704 putback_lru_page(page);
1707 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1708 pfn, ret, page->flags);
1713 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1714 pfn, ret, page_count(page), page->flags);
1720 * soft_offline_page - Soft offline a page.
1721 * @page: page to offline
1722 * @flags: flags. Same as memory_failure().
1724 * Returns 0 on success, otherwise negated errno.
1726 * Soft offline a page, by migration or invalidation,
1727 * without killing anything. This is for the case when
1728 * a page is not corrupted yet (so it's still valid to access),
1729 * but has had a number of corrected errors and is better taken
1732 * The actual policy on when to do that is maintained by
1735 * This should never impact any application or cause data loss,
1736 * however it might take some time.
1738 * This is not a 100% solution for all memory, but tries to be
1739 * ``good enough'' for the majority of memory.
1741 int soft_offline_page(struct page *page, int flags)
1744 unsigned long pfn = page_to_pfn(page);
1745 struct page *hpage = compound_head(page);
1747 if (PageHWPoison(page)) {
1748 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1749 if (flags & MF_COUNT_INCREASED)
1750 put_hwpoison_page(page);
1753 if (!PageHuge(page) && PageTransHuge(hpage)) {
1754 if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1755 pr_info("soft offline: %#lx: failed to split THP\n",
1757 if (flags & MF_COUNT_INCREASED)
1758 put_hwpoison_page(page);
1765 ret = get_any_page(page, pfn, flags);
1767 if (ret > 0) { /* for in-use pages */
1769 ret = soft_offline_huge_page(page, flags);
1771 ret = __soft_offline_page(page, flags);
1772 } else if (ret == 0) { /* for free pages */
1773 if (PageHuge(page)) {
1774 set_page_hwpoison_huge_page(hpage);
1775 if (!dequeue_hwpoisoned_huge_page(hpage))
1776 num_poisoned_pages_add(1 << compound_order(hpage));
1778 if (!TestSetPageHWPoison(page))
1779 num_poisoned_pages_inc();