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
42 * - hugetlb needs more code
43 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
44 * - pass bad pages to kdump next kernel
46 #include <linux/kernel.h>
48 #include <linux/page-flags.h>
49 #include <linux/kernel-page-flags.h>
50 #include <linux/sched.h>
51 #include <linux/ksm.h>
52 #include <linux/rmap.h>
53 #include <linux/export.h>
54 #include <linux/pagemap.h>
55 #include <linux/swap.h>
56 #include <linux/backing-dev.h>
57 #include <linux/migrate.h>
58 #include <linux/page-isolation.h>
59 #include <linux/suspend.h>
60 #include <linux/slab.h>
61 #include <linux/swapops.h>
62 #include <linux/hugetlb.h>
63 #include <linux/memory_hotplug.h>
64 #include <linux/mm_inline.h>
65 #include <linux/kfifo.h>
68 int sysctl_memory_failure_early_kill __read_mostly = 0;
70 int sysctl_memory_failure_recovery __read_mostly = 1;
72 atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
74 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
76 u32 hwpoison_filter_enable = 0;
77 u32 hwpoison_filter_dev_major = ~0U;
78 u32 hwpoison_filter_dev_minor = ~0U;
79 u64 hwpoison_filter_flags_mask;
80 u64 hwpoison_filter_flags_value;
81 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
82 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
83 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
84 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
85 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
87 static int hwpoison_filter_dev(struct page *p)
89 struct address_space *mapping;
92 if (hwpoison_filter_dev_major == ~0U &&
93 hwpoison_filter_dev_minor == ~0U)
97 * page_mapping() does not accept slab pages.
102 mapping = page_mapping(p);
103 if (mapping == NULL || mapping->host == NULL)
106 dev = mapping->host->i_sb->s_dev;
107 if (hwpoison_filter_dev_major != ~0U &&
108 hwpoison_filter_dev_major != MAJOR(dev))
110 if (hwpoison_filter_dev_minor != ~0U &&
111 hwpoison_filter_dev_minor != MINOR(dev))
117 static int hwpoison_filter_flags(struct page *p)
119 if (!hwpoison_filter_flags_mask)
122 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
123 hwpoison_filter_flags_value)
130 * This allows stress tests to limit test scope to a collection of tasks
131 * by putting them under some memcg. This prevents killing unrelated/important
132 * processes such as /sbin/init. Note that the target task may share clean
133 * pages with init (eg. libc text), which is harmless. If the target task
134 * share _dirty_ pages with another task B, the test scheme must make sure B
135 * is also included in the memcg. At last, due to race conditions this filter
136 * can only guarantee that the page either belongs to the memcg tasks, or is
139 #ifdef CONFIG_MEMCG_SWAP
140 u64 hwpoison_filter_memcg;
141 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
142 static int hwpoison_filter_task(struct page *p)
144 struct mem_cgroup *mem;
145 struct cgroup_subsys_state *css;
148 if (!hwpoison_filter_memcg)
151 mem = try_get_mem_cgroup_from_page(p);
155 css = mem_cgroup_css(mem);
156 ino = cgroup_ino(css->cgroup);
159 if (ino != hwpoison_filter_memcg)
165 static int hwpoison_filter_task(struct page *p) { return 0; }
168 int hwpoison_filter(struct page *p)
170 if (!hwpoison_filter_enable)
173 if (hwpoison_filter_dev(p))
176 if (hwpoison_filter_flags(p))
179 if (hwpoison_filter_task(p))
185 int hwpoison_filter(struct page *p)
191 EXPORT_SYMBOL_GPL(hwpoison_filter);
194 * Send all the processes who have the page mapped a signal.
195 * ``action optional'' if they are not immediately affected by the error
196 * ``action required'' if error happened in current execution context
198 static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
199 unsigned long pfn, struct page *page, int flags)
205 "MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
206 pfn, t->comm, t->pid);
207 si.si_signo = SIGBUS;
209 si.si_addr = (void *)addr;
210 #ifdef __ARCH_SI_TRAPNO
211 si.si_trapno = trapno;
213 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
215 if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
216 si.si_code = BUS_MCEERR_AR;
217 ret = force_sig_info(SIGBUS, &si, current);
220 * Don't use force here, it's convenient if the signal
221 * can be temporarily blocked.
222 * This could cause a loop when the user sets SIGBUS
223 * to SIG_IGN, but hopefully no one will do that?
225 si.si_code = BUS_MCEERR_AO;
226 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
229 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
230 t->comm, t->pid, ret);
235 * When a unknown page type is encountered drain as many buffers as possible
236 * in the hope to turn the page into a LRU or free page, which we can handle.
238 void shake_page(struct page *p, int access)
244 drain_all_pages(page_zone(p));
245 if (PageLRU(p) || is_free_buddy_page(p))
250 * Only call shrink_node_slabs here (which would also shrink
251 * other caches) if access is not potentially fatal.
254 drop_slab_node(page_to_nid(p));
256 EXPORT_SYMBOL_GPL(shake_page);
259 * Kill all processes that have a poisoned page mapped and then isolate
263 * Find all processes having the page mapped and kill them.
264 * But we keep a page reference around so that the page is not
265 * actually freed yet.
266 * Then stash the page away
268 * There's no convenient way to get back to mapped processes
269 * from the VMAs. So do a brute-force search over all
272 * Remember that machine checks are not common (or rather
273 * if they are common you have other problems), so this shouldn't
274 * be a performance issue.
276 * Also there are some races possible while we get from the
277 * error detection to actually handle it.
282 struct task_struct *tsk;
288 * Failure handling: if we can't find or can't kill a process there's
289 * not much we can do. We just print a message and ignore otherwise.
293 * Schedule a process for later kill.
294 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
295 * TBD would GFP_NOIO be enough?
297 static void add_to_kill(struct task_struct *tsk, struct page *p,
298 struct vm_area_struct *vma,
299 struct list_head *to_kill,
300 struct to_kill **tkc)
308 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
311 "MCE: Out of memory while machine check handling\n");
315 tk->addr = page_address_in_vma(p, vma);
319 * In theory we don't have to kill when the page was
320 * munmaped. But it could be also a mremap. Since that's
321 * likely very rare kill anyways just out of paranoia, but use
322 * a SIGKILL because the error is not contained anymore.
324 if (tk->addr == -EFAULT) {
325 pr_info("MCE: Unable to find user space address %lx in %s\n",
326 page_to_pfn(p), tsk->comm);
329 get_task_struct(tsk);
331 list_add_tail(&tk->nd, to_kill);
335 * Kill the processes that have been collected earlier.
337 * Only do anything when DOIT is set, otherwise just free the list
338 * (this is used for clean pages which do not need killing)
339 * Also when FAIL is set do a force kill because something went
342 static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
343 int fail, struct page *page, unsigned long pfn,
346 struct to_kill *tk, *next;
348 list_for_each_entry_safe (tk, next, to_kill, nd) {
351 * In case something went wrong with munmapping
352 * make sure the process doesn't catch the
353 * signal and then access the memory. Just kill it.
355 if (fail || tk->addr_valid == 0) {
357 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
358 pfn, tk->tsk->comm, tk->tsk->pid);
359 force_sig(SIGKILL, tk->tsk);
363 * In theory the process could have mapped
364 * something else on the address in-between. We could
365 * check for that, but we need to tell the
368 else if (kill_proc(tk->tsk, tk->addr, trapno,
369 pfn, page, flags) < 0)
371 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
372 pfn, tk->tsk->comm, tk->tsk->pid);
374 put_task_struct(tk->tsk);
380 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
381 * on behalf of the thread group. Return task_struct of the (first found)
382 * dedicated thread if found, and return NULL otherwise.
384 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
385 * have to call rcu_read_lock/unlock() in this function.
387 static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
389 struct task_struct *t;
391 for_each_thread(tsk, t)
392 if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
398 * Determine whether a given process is "early kill" process which expects
399 * to be signaled when some page under the process is hwpoisoned.
400 * Return task_struct of the dedicated thread (main thread unless explicitly
401 * specified) if the process is "early kill," and otherwise returns NULL.
403 static struct task_struct *task_early_kill(struct task_struct *tsk,
406 struct task_struct *t;
411 t = find_early_kill_thread(tsk);
414 if (sysctl_memory_failure_early_kill)
420 * Collect processes when the error hit an anonymous page.
422 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
423 struct to_kill **tkc, int force_early)
425 struct vm_area_struct *vma;
426 struct task_struct *tsk;
430 av = page_lock_anon_vma_read(page);
431 if (av == NULL) /* Not actually mapped anymore */
434 pgoff = page_to_pgoff(page);
435 read_lock(&tasklist_lock);
436 for_each_process (tsk) {
437 struct anon_vma_chain *vmac;
438 struct task_struct *t = task_early_kill(tsk, force_early);
442 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
445 if (!page_mapped_in_vma(page, vma))
447 if (vma->vm_mm == t->mm)
448 add_to_kill(t, page, vma, to_kill, tkc);
451 read_unlock(&tasklist_lock);
452 page_unlock_anon_vma_read(av);
456 * Collect processes when the error hit a file mapped page.
458 static void collect_procs_file(struct page *page, struct list_head *to_kill,
459 struct to_kill **tkc, int force_early)
461 struct vm_area_struct *vma;
462 struct task_struct *tsk;
463 struct address_space *mapping = page->mapping;
465 i_mmap_lock_read(mapping);
466 read_lock(&tasklist_lock);
467 for_each_process(tsk) {
468 pgoff_t pgoff = page_to_pgoff(page);
469 struct task_struct *t = task_early_kill(tsk, force_early);
473 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
476 * Send early kill signal to tasks where a vma covers
477 * the page but the corrupted page is not necessarily
478 * mapped it in its pte.
479 * Assume applications who requested early kill want
480 * to be informed of all such data corruptions.
482 if (vma->vm_mm == t->mm)
483 add_to_kill(t, page, vma, to_kill, tkc);
486 read_unlock(&tasklist_lock);
487 i_mmap_unlock_read(mapping);
491 * Collect the processes who have the corrupted page mapped to kill.
492 * This is done in two steps for locking reasons.
493 * First preallocate one tokill structure outside the spin locks,
494 * so that we can kill at least one process reasonably reliable.
496 static void collect_procs(struct page *page, struct list_head *tokill,
504 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
508 collect_procs_anon(page, tokill, &tk, force_early);
510 collect_procs_file(page, tokill, &tk, force_early);
515 * Error handlers for various types of pages.
519 IGNORED, /* Error: cannot be handled */
520 FAILED, /* Error: handling failed */
521 DELAYED, /* Will be handled later */
522 RECOVERED, /* Successfully recovered */
525 static const char *action_name[] = {
526 [IGNORED] = "Ignored",
528 [DELAYED] = "Delayed",
529 [RECOVERED] = "Recovered",
532 enum action_page_type {
534 MSG_KERNEL_HIGH_ORDER,
536 MSG_DIFFERENT_COMPOUND,
543 MSG_DIRTY_MLOCKED_LRU,
544 MSG_CLEAN_MLOCKED_LRU,
545 MSG_DIRTY_UNEVICTABLE_LRU,
546 MSG_CLEAN_UNEVICTABLE_LRU,
555 static const char * const action_page_types[] = {
556 [MSG_KERNEL] = "reserved kernel page",
557 [MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
558 [MSG_SLAB] = "kernel slab page",
559 [MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
560 [MSG_POISONED_HUGE] = "huge page already hardware poisoned",
561 [MSG_HUGE] = "huge page",
562 [MSG_FREE_HUGE] = "free huge page",
563 [MSG_UNMAP_FAILED] = "unmapping failed page",
564 [MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
565 [MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
566 [MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
567 [MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
568 [MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
569 [MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
570 [MSG_DIRTY_LRU] = "dirty LRU page",
571 [MSG_CLEAN_LRU] = "clean LRU page",
572 [MSG_TRUNCATED_LRU] = "already truncated LRU page",
573 [MSG_BUDDY] = "free buddy page",
574 [MSG_BUDDY_2ND] = "free buddy page (2nd try)",
575 [MSG_UNKNOWN] = "unknown page",
579 * XXX: It is possible that a page is isolated from LRU cache,
580 * and then kept in swap cache or failed to remove from page cache.
581 * The page count will stop it from being freed by unpoison.
582 * Stress tests should be aware of this memory leak problem.
584 static int delete_from_lru_cache(struct page *p)
586 if (!isolate_lru_page(p)) {
588 * Clear sensible page flags, so that the buddy system won't
589 * complain when the page is unpoison-and-freed.
592 ClearPageUnevictable(p);
594 * drop the page count elevated by isolate_lru_page()
596 page_cache_release(p);
603 * Error hit kernel page.
604 * Do nothing, try to be lucky and not touch this instead. For a few cases we
605 * could be more sophisticated.
607 static int me_kernel(struct page *p, unsigned long pfn)
613 * Page in unknown state. Do nothing.
615 static int me_unknown(struct page *p, unsigned long pfn)
617 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
622 * Clean (or cleaned) page cache page.
624 static int me_pagecache_clean(struct page *p, unsigned long pfn)
628 struct address_space *mapping;
630 delete_from_lru_cache(p);
633 * For anonymous pages we're done the only reference left
634 * should be the one m_f() holds.
640 * Now truncate the page in the page cache. This is really
641 * more like a "temporary hole punch"
642 * Don't do this for block devices when someone else
643 * has a reference, because it could be file system metadata
644 * and that's not safe to truncate.
646 mapping = page_mapping(p);
649 * Page has been teared down in the meanwhile
655 * Truncation is a bit tricky. Enable it per file system for now.
657 * Open: to take i_mutex or not for this? Right now we don't.
659 if (mapping->a_ops->error_remove_page) {
660 err = mapping->a_ops->error_remove_page(mapping, p);
662 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
664 } else if (page_has_private(p) &&
665 !try_to_release_page(p, GFP_NOIO)) {
666 pr_info("MCE %#lx: failed to release buffers\n", pfn);
672 * If the file system doesn't support it just invalidate
673 * This fails on dirty or anything with private pages
675 if (invalidate_inode_page(p))
678 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
685 * Dirty pagecache page
686 * Issues: when the error hit a hole page the error is not properly
689 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
691 struct address_space *mapping = page_mapping(p);
694 /* TBD: print more information about the file. */
697 * IO error will be reported by write(), fsync(), etc.
698 * who check the mapping.
699 * This way the application knows that something went
700 * wrong with its dirty file data.
702 * There's one open issue:
704 * The EIO will be only reported on the next IO
705 * operation and then cleared through the IO map.
706 * Normally Linux has two mechanisms to pass IO error
707 * first through the AS_EIO flag in the address space
708 * and then through the PageError flag in the page.
709 * Since we drop pages on memory failure handling the
710 * only mechanism open to use is through AS_AIO.
712 * This has the disadvantage that it gets cleared on
713 * the first operation that returns an error, while
714 * the PageError bit is more sticky and only cleared
715 * when the page is reread or dropped. If an
716 * application assumes it will always get error on
717 * fsync, but does other operations on the fd before
718 * and the page is dropped between then the error
719 * will not be properly reported.
721 * This can already happen even without hwpoisoned
722 * pages: first on metadata IO errors (which only
723 * report through AS_EIO) or when the page is dropped
726 * So right now we assume that the application DTRT on
727 * the first EIO, but we're not worse than other parts
730 mapping_set_error(mapping, EIO);
733 return me_pagecache_clean(p, pfn);
737 * Clean and dirty swap cache.
739 * Dirty swap cache page is tricky to handle. The page could live both in page
740 * cache and swap cache(ie. page is freshly swapped in). So it could be
741 * referenced concurrently by 2 types of PTEs:
742 * normal PTEs and swap PTEs. We try to handle them consistently by calling
743 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
745 * - clear dirty bit to prevent IO
747 * - but keep in the swap cache, so that when we return to it on
748 * a later page fault, we know the application is accessing
749 * corrupted data and shall be killed (we installed simple
750 * interception code in do_swap_page to catch it).
752 * Clean swap cache pages can be directly isolated. A later page fault will
753 * bring in the known good data from disk.
755 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
758 /* Trigger EIO in shmem: */
759 ClearPageUptodate(p);
761 if (!delete_from_lru_cache(p))
767 static int me_swapcache_clean(struct page *p, unsigned long pfn)
769 delete_from_swap_cache(p);
771 if (!delete_from_lru_cache(p))
778 * Huge pages. Needs work.
780 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
781 * To narrow down kill region to one page, we need to break up pmd.
783 static int me_huge_page(struct page *p, unsigned long pfn)
786 struct page *hpage = compound_head(p);
788 * We can safely recover from error on free or reserved (i.e.
789 * not in-use) hugepage by dequeuing it from freelist.
790 * To check whether a hugepage is in-use or not, we can't use
791 * page->lru because it can be used in other hugepage operations,
792 * such as __unmap_hugepage_range() and gather_surplus_pages().
793 * So instead we use page_mapping() and PageAnon().
794 * We assume that this function is called with page lock held,
795 * so there is no race between isolation and mapping/unmapping.
797 if (!(page_mapping(hpage) || PageAnon(hpage))) {
798 res = dequeue_hwpoisoned_huge_page(hpage);
806 * Various page states we can handle.
808 * A page state is defined by its current page->flags bits.
809 * The table matches them in order and calls the right handler.
811 * This is quite tricky because we can access page at any time
812 * in its live cycle, so all accesses have to be extremely careful.
814 * This is not complete. More states could be added.
815 * For any missing state don't attempt recovery.
818 #define dirty (1UL << PG_dirty)
819 #define sc (1UL << PG_swapcache)
820 #define unevict (1UL << PG_unevictable)
821 #define mlock (1UL << PG_mlocked)
822 #define writeback (1UL << PG_writeback)
823 #define lru (1UL << PG_lru)
824 #define swapbacked (1UL << PG_swapbacked)
825 #define head (1UL << PG_head)
826 #define tail (1UL << PG_tail)
827 #define compound (1UL << PG_compound)
828 #define slab (1UL << PG_slab)
829 #define reserved (1UL << PG_reserved)
831 static struct page_state {
834 enum action_page_type type;
835 int (*action)(struct page *p, unsigned long pfn);
837 { reserved, reserved, MSG_KERNEL, me_kernel },
839 * free pages are specially detected outside this table:
840 * PG_buddy pages only make a small fraction of all free pages.
844 * Could in theory check if slab page is free or if we can drop
845 * currently unused objects without touching them. But just
846 * treat it as standard kernel for now.
848 { slab, slab, MSG_SLAB, me_kernel },
850 #ifdef CONFIG_PAGEFLAGS_EXTENDED
851 { head, head, MSG_HUGE, me_huge_page },
852 { tail, tail, MSG_HUGE, me_huge_page },
854 { compound, compound, MSG_HUGE, me_huge_page },
857 { sc|dirty, sc|dirty, MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
858 { sc|dirty, sc, MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
860 { mlock|dirty, mlock|dirty, MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
861 { mlock|dirty, mlock, MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
863 { unevict|dirty, unevict|dirty, MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
864 { unevict|dirty, unevict, MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
866 { lru|dirty, lru|dirty, MSG_DIRTY_LRU, me_pagecache_dirty },
867 { lru|dirty, lru, MSG_CLEAN_LRU, me_pagecache_clean },
870 * Catchall entry: must be at end.
872 { 0, 0, MSG_UNKNOWN, me_unknown },
889 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
890 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
892 static void action_result(unsigned long pfn, enum action_page_type type, int result)
894 pr_err("MCE %#lx: recovery action for %s: %s\n",
895 pfn, action_page_types[type], action_name[result]);
898 static int page_action(struct page_state *ps, struct page *p,
904 result = ps->action(p, pfn);
906 count = page_count(p) - 1;
907 if (ps->action == me_swapcache_dirty && result == DELAYED)
911 "MCE %#lx: %s still referenced by %d users\n",
912 pfn, action_page_types[ps->type], count);
915 action_result(pfn, ps->type, result);
917 /* Could do more checks here if page looks ok */
919 * Could adjust zone counters here to correct for the missing page.
922 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
926 * Do all that is necessary to remove user space mappings. Unmap
927 * the pages and send SIGBUS to the processes if the data was dirty.
929 static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
930 int trapno, int flags, struct page **hpagep)
932 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
933 struct address_space *mapping;
936 int kill = 1, forcekill;
937 struct page *hpage = *hpagep;
941 * Here we are interested only in user-mapped pages, so skip any
942 * other types of pages.
944 if (PageReserved(p) || PageSlab(p))
946 if (!(PageLRU(hpage) || PageHuge(p)))
950 * This check implies we don't kill processes if their pages
951 * are in the swap cache early. Those are always late kills.
953 if (!page_mapped(hpage))
957 pr_err("MCE %#lx: can't handle KSM pages.\n", pfn);
961 if (PageSwapCache(p)) {
963 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
964 ttu |= TTU_IGNORE_HWPOISON;
968 * Propagate the dirty bit from PTEs to struct page first, because we
969 * need this to decide if we should kill or just drop the page.
970 * XXX: the dirty test could be racy: set_page_dirty() may not always
971 * be called inside page lock (it's recommended but not enforced).
973 mapping = page_mapping(hpage);
974 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
975 mapping_cap_writeback_dirty(mapping)) {
976 if (page_mkclean(hpage)) {
980 ttu |= TTU_IGNORE_HWPOISON;
982 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
988 * ppage: poisoned page
989 * if p is regular page(4k page)
990 * ppage == real poisoned page;
991 * else p is hugetlb or THP, ppage == head page.
995 if (PageTransHuge(hpage)) {
997 * Verify that this isn't a hugetlbfs head page, the check for
998 * PageAnon is just for avoid tripping a split_huge_page
999 * internal debug check, as split_huge_page refuses to deal with
1000 * anything that isn't an anon page. PageAnon can't go away fro
1001 * under us because we hold a refcount on the hpage, without a
1002 * refcount on the hpage. split_huge_page can't be safely called
1003 * in the first place, having a refcount on the tail isn't
1004 * enough * to be safe.
1006 if (!PageHuge(hpage) && PageAnon(hpage)) {
1007 if (unlikely(split_huge_page(hpage))) {
1009 * FIXME: if splitting THP is failed, it is
1010 * better to stop the following operation rather
1011 * than causing panic by unmapping. System might
1012 * survive if the page is freed later.
1015 "MCE %#lx: failed to split THP\n", pfn);
1017 BUG_ON(!PageHWPoison(p));
1021 * We pinned the head page for hwpoison handling,
1022 * now we split the thp and we are interested in
1023 * the hwpoisoned raw page, so move the refcount
1024 * to it. Similarly, page lock is shifted.
1027 if (!(flags & MF_COUNT_INCREASED)) {
1035 /* THP is split, so ppage should be the real poisoned page. */
1041 * First collect all the processes that have the page
1042 * mapped in dirty form. This has to be done before try_to_unmap,
1043 * because ttu takes the rmap data structures down.
1045 * Error handling: We ignore errors here because
1046 * there's nothing that can be done.
1049 collect_procs(ppage, &tokill, flags & MF_ACTION_REQUIRED);
1051 ret = try_to_unmap(ppage, ttu);
1052 if (ret != SWAP_SUCCESS)
1053 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
1054 pfn, page_mapcount(ppage));
1057 * Now that the dirty bit has been propagated to the
1058 * struct page and all unmaps done we can decide if
1059 * killing is needed or not. Only kill when the page
1060 * was dirty or the process is not restartable,
1061 * otherwise the tokill list is merely
1062 * freed. When there was a problem unmapping earlier
1063 * use a more force-full uncatchable kill to prevent
1064 * any accesses to the poisoned memory.
1066 forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL);
1067 kill_procs(&tokill, forcekill, trapno,
1068 ret != SWAP_SUCCESS, p, pfn, flags);
1073 static void set_page_hwpoison_huge_page(struct page *hpage)
1076 int nr_pages = 1 << compound_order(hpage);
1077 for (i = 0; i < nr_pages; i++)
1078 SetPageHWPoison(hpage + i);
1081 static void clear_page_hwpoison_huge_page(struct page *hpage)
1084 int nr_pages = 1 << compound_order(hpage);
1085 for (i = 0; i < nr_pages; i++)
1086 ClearPageHWPoison(hpage + i);
1090 * memory_failure - Handle memory failure of a page.
1091 * @pfn: Page Number of the corrupted page
1092 * @trapno: Trap number reported in the signal to user space.
1093 * @flags: fine tune action taken
1095 * This function is called by the low level machine check code
1096 * of an architecture when it detects hardware memory corruption
1097 * of a page. It tries its best to recover, which includes
1098 * dropping pages, killing processes etc.
1100 * The function is primarily of use for corruptions that
1101 * happen outside the current execution context (e.g. when
1102 * detected by a background scrubber)
1104 * Must run in process context (e.g. a work queue) with interrupts
1105 * enabled and no spinlocks hold.
1107 int memory_failure(unsigned long pfn, int trapno, int flags)
1109 struct page_state *ps;
1113 unsigned int nr_pages;
1114 unsigned long page_flags;
1116 if (!sysctl_memory_failure_recovery)
1117 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1119 if (!pfn_valid(pfn)) {
1121 "MCE %#lx: memory outside kernel control\n",
1126 p = pfn_to_page(pfn);
1127 hpage = compound_head(p);
1128 if (TestSetPageHWPoison(p)) {
1129 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
1134 * Currently errors on hugetlbfs pages are measured in hugepage units,
1135 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1136 * transparent hugepages, they are supposed to be split and error
1137 * measurement is done in normal page units. So nr_pages should be one
1141 nr_pages = 1 << compound_order(hpage);
1142 else /* normal page or thp */
1144 atomic_long_add(nr_pages, &num_poisoned_pages);
1147 * We need/can do nothing about count=0 pages.
1148 * 1) it's a free page, and therefore in safe hand:
1149 * prep_new_page() will be the gate keeper.
1150 * 2) it's a free hugepage, which is also safe:
1151 * an affected hugepage will be dequeued from hugepage freelist,
1152 * so there's no concern about reusing it ever after.
1153 * 3) it's part of a non-compound high order page.
1154 * Implies some kernel user: cannot stop them from
1155 * R/W the page; let's pray that the page has been
1156 * used and will be freed some time later.
1157 * In fact it's dangerous to directly bump up page count from 0,
1158 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1160 if (!(flags & MF_COUNT_INCREASED) &&
1161 !get_page_unless_zero(hpage)) {
1162 if (is_free_buddy_page(p)) {
1163 action_result(pfn, MSG_BUDDY, DELAYED);
1165 } else if (PageHuge(hpage)) {
1167 * Check "filter hit" and "race with other subpage."
1170 if (PageHWPoison(hpage)) {
1171 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1172 || (p != hpage && TestSetPageHWPoison(hpage))) {
1173 atomic_long_sub(nr_pages, &num_poisoned_pages);
1178 set_page_hwpoison_huge_page(hpage);
1179 res = dequeue_hwpoisoned_huge_page(hpage);
1180 action_result(pfn, MSG_FREE_HUGE,
1181 res ? IGNORED : DELAYED);
1185 action_result(pfn, MSG_KERNEL_HIGH_ORDER, IGNORED);
1191 * We ignore non-LRU pages for good reasons.
1192 * - PG_locked is only well defined for LRU pages and a few others
1193 * - to avoid races with __set_page_locked()
1194 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1195 * The check (unnecessarily) ignores LRU pages being isolated and
1196 * walked by the page reclaim code, however that's not a big loss.
1199 if (!PageLRU(hpage))
1200 shake_page(hpage, 0);
1201 if (!PageLRU(hpage)) {
1203 * shake_page could have turned it free.
1205 if (is_free_buddy_page(p)) {
1206 if (flags & MF_COUNT_INCREASED)
1207 action_result(pfn, MSG_BUDDY, DELAYED);
1209 action_result(pfn, MSG_BUDDY_2ND,
1219 * The page could have changed compound pages during the locking.
1220 * If this happens just bail out.
1222 if (compound_head(p) != hpage) {
1223 action_result(pfn, MSG_DIFFERENT_COMPOUND, IGNORED);
1229 * We use page flags to determine what action should be taken, but
1230 * the flags can be modified by the error containment action. One
1231 * example is an mlocked page, where PG_mlocked is cleared by
1232 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1233 * correctly, we save a copy of the page flags at this time.
1235 page_flags = p->flags;
1238 * unpoison always clear PG_hwpoison inside page lock
1240 if (!PageHWPoison(p)) {
1241 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1242 atomic_long_sub(nr_pages, &num_poisoned_pages);
1247 if (hwpoison_filter(p)) {
1248 if (TestClearPageHWPoison(p))
1249 atomic_long_sub(nr_pages, &num_poisoned_pages);
1255 if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p))
1256 goto identify_page_state;
1259 * For error on the tail page, we should set PG_hwpoison
1260 * on the head page to show that the hugepage is hwpoisoned
1262 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1263 action_result(pfn, MSG_POISONED_HUGE, IGNORED);
1269 * Set PG_hwpoison on all pages in an error hugepage,
1270 * because containment is done in hugepage unit for now.
1271 * Since we have done TestSetPageHWPoison() for the head page with
1272 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1275 set_page_hwpoison_huge_page(hpage);
1278 * It's very difficult to mess with pages currently under IO
1279 * and in many cases impossible, so we just avoid it here.
1281 wait_on_page_writeback(p);
1284 * Now take care of user space mappings.
1285 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1287 * When the raw error page is thp tail page, hpage points to the raw
1288 * page after thp split.
1290 if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)
1292 action_result(pfn, MSG_UNMAP_FAILED, IGNORED);
1298 * Torn down by someone else?
1300 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1301 action_result(pfn, MSG_TRUNCATED_LRU, IGNORED);
1306 identify_page_state:
1309 * The first check uses the current page flags which may not have any
1310 * relevant information. The second check with the saved page flagss is
1311 * carried out only if the first check can't determine the page status.
1313 for (ps = error_states;; ps++)
1314 if ((p->flags & ps->mask) == ps->res)
1317 page_flags |= (p->flags & (1UL << PG_dirty));
1320 for (ps = error_states;; ps++)
1321 if ((page_flags & ps->mask) == ps->res)
1323 res = page_action(ps, p, pfn);
1328 EXPORT_SYMBOL_GPL(memory_failure);
1330 #define MEMORY_FAILURE_FIFO_ORDER 4
1331 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1333 struct memory_failure_entry {
1339 struct memory_failure_cpu {
1340 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1341 MEMORY_FAILURE_FIFO_SIZE);
1343 struct work_struct work;
1346 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1349 * memory_failure_queue - Schedule handling memory failure of a page.
1350 * @pfn: Page Number of the corrupted page
1351 * @trapno: Trap number reported in the signal to user space.
1352 * @flags: Flags for memory failure handling
1354 * This function is called by the low level hardware error handler
1355 * when it detects hardware memory corruption of a page. It schedules
1356 * the recovering of error page, including dropping pages, killing
1359 * The function is primarily of use for corruptions that
1360 * happen outside the current execution context (e.g. when
1361 * detected by a background scrubber)
1363 * Can run in IRQ context.
1365 void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1367 struct memory_failure_cpu *mf_cpu;
1368 unsigned long proc_flags;
1369 struct memory_failure_entry entry = {
1375 mf_cpu = &get_cpu_var(memory_failure_cpu);
1376 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1377 if (kfifo_put(&mf_cpu->fifo, entry))
1378 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1380 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1382 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1383 put_cpu_var(memory_failure_cpu);
1385 EXPORT_SYMBOL_GPL(memory_failure_queue);
1387 static void memory_failure_work_func(struct work_struct *work)
1389 struct memory_failure_cpu *mf_cpu;
1390 struct memory_failure_entry entry = { 0, };
1391 unsigned long proc_flags;
1394 mf_cpu = this_cpu_ptr(&memory_failure_cpu);
1396 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1397 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1398 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1401 if (entry.flags & MF_SOFT_OFFLINE)
1402 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1404 memory_failure(entry.pfn, entry.trapno, entry.flags);
1408 static int __init memory_failure_init(void)
1410 struct memory_failure_cpu *mf_cpu;
1413 for_each_possible_cpu(cpu) {
1414 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1415 spin_lock_init(&mf_cpu->lock);
1416 INIT_KFIFO(mf_cpu->fifo);
1417 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1422 core_initcall(memory_failure_init);
1425 * unpoison_memory - Unpoison a previously poisoned page
1426 * @pfn: Page number of the to be unpoisoned page
1428 * Software-unpoison a page that has been poisoned by
1429 * memory_failure() earlier.
1431 * This is only done on the software-level, so it only works
1432 * for linux injected failures, not real hardware failures
1434 * Returns 0 for success, otherwise -errno.
1436 int unpoison_memory(unsigned long pfn)
1441 unsigned int nr_pages;
1443 if (!pfn_valid(pfn))
1446 p = pfn_to_page(pfn);
1447 page = compound_head(p);
1449 if (!PageHWPoison(p)) {
1450 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
1455 * unpoison_memory() can encounter thp only when the thp is being
1456 * worked by memory_failure() and the page lock is not held yet.
1457 * In such case, we yield to memory_failure() and make unpoison fail.
1459 if (!PageHuge(page) && PageTransHuge(page)) {
1460 pr_info("MCE: Memory failure is now running on %#lx\n", pfn);
1464 nr_pages = 1 << compound_order(page);
1466 if (!get_page_unless_zero(page)) {
1468 * Since HWPoisoned hugepage should have non-zero refcount,
1469 * race between memory failure and unpoison seems to happen.
1470 * In such case unpoison fails and memory failure runs
1473 if (PageHuge(page)) {
1474 pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1477 if (TestClearPageHWPoison(p))
1478 atomic_long_dec(&num_poisoned_pages);
1479 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
1485 * This test is racy because PG_hwpoison is set outside of page lock.
1486 * That's acceptable because that won't trigger kernel panic. Instead,
1487 * the PG_hwpoison page will be caught and isolated on the entrance to
1488 * the free buddy page pool.
1490 if (TestClearPageHWPoison(page)) {
1491 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
1492 atomic_long_sub(nr_pages, &num_poisoned_pages);
1495 clear_page_hwpoison_huge_page(page);
1500 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
1505 EXPORT_SYMBOL(unpoison_memory);
1507 static struct page *new_page(struct page *p, unsigned long private, int **x)
1509 int nid = page_to_nid(p);
1511 return alloc_huge_page_node(page_hstate(compound_head(p)),
1514 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1518 * Safely get reference count of an arbitrary page.
1519 * Returns 0 for a free page, -EIO for a zero refcount page
1520 * that is not free, and 1 for any other page type.
1521 * For 1 the page is returned with increased page count, otherwise not.
1523 static int __get_any_page(struct page *p, unsigned long pfn, int flags)
1527 if (flags & MF_COUNT_INCREASED)
1531 * When the target page is a free hugepage, just remove it
1532 * from free hugepage list.
1534 if (!get_page_unless_zero(compound_head(p))) {
1536 pr_info("%s: %#lx free huge page\n", __func__, pfn);
1538 } else if (is_free_buddy_page(p)) {
1539 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1542 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1543 __func__, pfn, p->flags);
1547 /* Not a free page */
1553 static int get_any_page(struct page *page, unsigned long pfn, int flags)
1555 int ret = __get_any_page(page, pfn, flags);
1557 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1562 shake_page(page, 1);
1567 ret = __get_any_page(page, pfn, 0);
1568 if (!PageLRU(page)) {
1569 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1577 static int soft_offline_huge_page(struct page *page, int flags)
1580 unsigned long pfn = page_to_pfn(page);
1581 struct page *hpage = compound_head(page);
1582 LIST_HEAD(pagelist);
1585 * This double-check of PageHWPoison is to avoid the race with
1586 * memory_failure(). See also comment in __soft_offline_page().
1589 if (PageHWPoison(hpage)) {
1592 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1597 ret = isolate_huge_page(hpage, &pagelist);
1600 * get_any_page() and isolate_huge_page() takes a refcount each,
1601 * so need to drop one here.
1605 pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn);
1609 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1610 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1612 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1613 pfn, ret, page->flags);
1615 * We know that soft_offline_huge_page() tries to migrate
1616 * only one hugepage pointed to by hpage, so we need not
1617 * run through the pagelist here.
1619 putback_active_hugepage(hpage);
1623 /* overcommit hugetlb page will be freed to buddy */
1624 if (PageHuge(page)) {
1625 set_page_hwpoison_huge_page(hpage);
1626 dequeue_hwpoisoned_huge_page(hpage);
1627 atomic_long_add(1 << compound_order(hpage),
1628 &num_poisoned_pages);
1630 SetPageHWPoison(page);
1631 atomic_long_inc(&num_poisoned_pages);
1637 static int __soft_offline_page(struct page *page, int flags)
1640 unsigned long pfn = page_to_pfn(page);
1643 * Check PageHWPoison again inside page lock because PageHWPoison
1644 * is set by memory_failure() outside page lock. Note that
1645 * memory_failure() also double-checks PageHWPoison inside page lock,
1646 * so there's no race between soft_offline_page() and memory_failure().
1649 wait_on_page_writeback(page);
1650 if (PageHWPoison(page)) {
1653 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1657 * Try to invalidate first. This should work for
1658 * non dirty unmapped page cache pages.
1660 ret = invalidate_inode_page(page);
1663 * RED-PEN would be better to keep it isolated here, but we
1664 * would need to fix isolation locking first.
1668 pr_info("soft_offline: %#lx: invalidated\n", pfn);
1669 SetPageHWPoison(page);
1670 atomic_long_inc(&num_poisoned_pages);
1675 * Simple invalidation didn't work.
1676 * Try to migrate to a new page instead. migrate.c
1677 * handles a large number of cases for us.
1679 ret = isolate_lru_page(page);
1681 * Drop page reference which is came from get_any_page()
1682 * successful isolate_lru_page() already took another one.
1686 LIST_HEAD(pagelist);
1687 inc_zone_page_state(page, NR_ISOLATED_ANON +
1688 page_is_file_cache(page));
1689 list_add(&page->lru, &pagelist);
1690 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1691 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1693 if (!list_empty(&pagelist)) {
1694 list_del(&page->lru);
1695 dec_zone_page_state(page, NR_ISOLATED_ANON +
1696 page_is_file_cache(page));
1697 putback_lru_page(page);
1700 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1701 pfn, ret, page->flags);
1706 * After page migration succeeds, the source page can
1707 * be trapped in pagevec and actual freeing is delayed.
1708 * Freeing code works differently based on PG_hwpoison,
1709 * so there's a race. We need to make sure that the
1710 * source page should be freed back to buddy before
1711 * setting PG_hwpoison.
1713 if (!is_free_buddy_page(page))
1714 drain_all_pages(page_zone(page));
1715 SetPageHWPoison(page);
1716 if (!is_free_buddy_page(page))
1717 pr_info("soft offline: %#lx: page leaked\n",
1719 atomic_long_inc(&num_poisoned_pages);
1722 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1723 pfn, ret, page_count(page), page->flags);
1729 * soft_offline_page - Soft offline a page.
1730 * @page: page to offline
1731 * @flags: flags. Same as memory_failure().
1733 * Returns 0 on success, otherwise negated errno.
1735 * Soft offline a page, by migration or invalidation,
1736 * without killing anything. This is for the case when
1737 * a page is not corrupted yet (so it's still valid to access),
1738 * but has had a number of corrected errors and is better taken
1741 * The actual policy on when to do that is maintained by
1744 * This should never impact any application or cause data loss,
1745 * however it might take some time.
1747 * This is not a 100% solution for all memory, but tries to be
1748 * ``good enough'' for the majority of memory.
1750 int soft_offline_page(struct page *page, int flags)
1753 unsigned long pfn = page_to_pfn(page);
1754 struct page *hpage = compound_head(page);
1756 if (PageHWPoison(page)) {
1757 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1760 if (!PageHuge(page) && PageTransHuge(hpage)) {
1761 if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1762 pr_info("soft offline: %#lx: failed to split THP\n",
1771 * Isolate the page, so that it doesn't get reallocated if it
1772 * was free. This flag should be kept set until the source page
1773 * is freed and PG_hwpoison on it is set.
1775 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
1776 set_migratetype_isolate(page, true);
1778 ret = get_any_page(page, pfn, flags);
1780 if (ret > 0) { /* for in-use pages */
1782 ret = soft_offline_huge_page(page, flags);
1784 ret = __soft_offline_page(page, flags);
1785 } else if (ret == 0) { /* for free pages */
1786 if (PageHuge(page)) {
1787 set_page_hwpoison_huge_page(hpage);
1788 if (!dequeue_hwpoisoned_huge_page(hpage))
1789 atomic_long_add(1 << compound_order(hpage),
1790 &num_poisoned_pages);
1792 if (!TestSetPageHWPoison(page))
1793 atomic_long_inc(&num_poisoned_pages);
1796 unset_migratetype_isolate(page, MIGRATE_MOVABLE);