1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2008, 2009 Intel Corporation
4 * Authors: Andi Kleen, Fengguang Wu
6 * High level machine check handler. Handles pages reported by the
7 * hardware as being corrupted usually due to a multi-bit ECC memory or cache
10 * In addition there is a "soft offline" entry point that allows stop using
11 * not-yet-corrupted-by-suspicious pages without killing anything.
13 * Handles page cache pages in various states. The tricky part
14 * here is that we can access any page asynchronously in respect to
15 * other VM users, because memory failures could happen anytime and
16 * anywhere. This could violate some of their assumptions. This is why
17 * this code has to be extremely careful. Generally it tries to use
18 * normal locking rules, as in get the standard locks, even if that means
19 * the error handling takes potentially a long time.
21 * It can be very tempting to add handling for obscure cases here.
22 * In general any code for handling new cases should only be added iff:
23 * - You know how to test it.
24 * - You have a test that can be added to mce-test
25 * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/
26 * - The case actually shows up as a frequent (top 10) page state in
27 * tools/vm/page-types when running a real workload.
29 * There are several operations here with exponential complexity because
30 * of unsuitable VM data structures. For example the operation to map back
31 * from RMAP chains to processes has to walk the complete process list and
32 * has non linear complexity with the number. But since memory corruptions
33 * are rare we hope to get away with this. This avoids impacting the core
36 #include <linux/kernel.h>
38 #include <linux/page-flags.h>
39 #include <linux/kernel-page-flags.h>
40 #include <linux/sched/signal.h>
41 #include <linux/sched/task.h>
42 #include <linux/ksm.h>
43 #include <linux/rmap.h>
44 #include <linux/export.h>
45 #include <linux/pagemap.h>
46 #include <linux/swap.h>
47 #include <linux/backing-dev.h>
48 #include <linux/migrate.h>
49 #include <linux/suspend.h>
50 #include <linux/slab.h>
51 #include <linux/swapops.h>
52 #include <linux/hugetlb.h>
53 #include <linux/memory_hotplug.h>
54 #include <linux/mm_inline.h>
55 #include <linux/memremap.h>
56 #include <linux/kfifo.h>
57 #include <linux/ratelimit.h>
58 #include <linux/page-isolation.h>
60 #include "ras/ras_event.h"
62 int sysctl_memory_failure_early_kill __read_mostly = 0;
64 int sysctl_memory_failure_recovery __read_mostly = 1;
66 atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
68 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
70 u32 hwpoison_filter_enable = 0;
71 u32 hwpoison_filter_dev_major = ~0U;
72 u32 hwpoison_filter_dev_minor = ~0U;
73 u64 hwpoison_filter_flags_mask;
74 u64 hwpoison_filter_flags_value;
75 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
76 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
77 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
78 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
79 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
81 static int hwpoison_filter_dev(struct page *p)
83 struct address_space *mapping;
86 if (hwpoison_filter_dev_major == ~0U &&
87 hwpoison_filter_dev_minor == ~0U)
91 * page_mapping() does not accept slab pages.
96 mapping = page_mapping(p);
97 if (mapping == NULL || mapping->host == NULL)
100 dev = mapping->host->i_sb->s_dev;
101 if (hwpoison_filter_dev_major != ~0U &&
102 hwpoison_filter_dev_major != MAJOR(dev))
104 if (hwpoison_filter_dev_minor != ~0U &&
105 hwpoison_filter_dev_minor != MINOR(dev))
111 static int hwpoison_filter_flags(struct page *p)
113 if (!hwpoison_filter_flags_mask)
116 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
117 hwpoison_filter_flags_value)
124 * This allows stress tests to limit test scope to a collection of tasks
125 * by putting them under some memcg. This prevents killing unrelated/important
126 * processes such as /sbin/init. Note that the target task may share clean
127 * pages with init (eg. libc text), which is harmless. If the target task
128 * share _dirty_ pages with another task B, the test scheme must make sure B
129 * is also included in the memcg. At last, due to race conditions this filter
130 * can only guarantee that the page either belongs to the memcg tasks, or is
134 u64 hwpoison_filter_memcg;
135 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
136 static int hwpoison_filter_task(struct page *p)
138 if (!hwpoison_filter_memcg)
141 if (page_cgroup_ino(p) != hwpoison_filter_memcg)
147 static int hwpoison_filter_task(struct page *p) { return 0; }
150 int hwpoison_filter(struct page *p)
152 if (!hwpoison_filter_enable)
155 if (hwpoison_filter_dev(p))
158 if (hwpoison_filter_flags(p))
161 if (hwpoison_filter_task(p))
167 int hwpoison_filter(struct page *p)
173 EXPORT_SYMBOL_GPL(hwpoison_filter);
176 * Kill all processes that have a poisoned page mapped and then isolate
180 * Find all processes having the page mapped and kill them.
181 * But we keep a page reference around so that the page is not
182 * actually freed yet.
183 * Then stash the page away
185 * There's no convenient way to get back to mapped processes
186 * from the VMAs. So do a brute-force search over all
189 * Remember that machine checks are not common (or rather
190 * if they are common you have other problems), so this shouldn't
191 * be a performance issue.
193 * Also there are some races possible while we get from the
194 * error detection to actually handle it.
199 struct task_struct *tsk;
205 * Send all the processes who have the page mapped a signal.
206 * ``action optional'' if they are not immediately affected by the error
207 * ``action required'' if error happened in current execution context
209 static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
211 struct task_struct *t = tk->tsk;
212 short addr_lsb = tk->size_shift;
215 pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
216 pfn, t->comm, t->pid);
218 if (flags & MF_ACTION_REQUIRED) {
219 WARN_ON_ONCE(t != current);
220 ret = force_sig_mceerr(BUS_MCEERR_AR,
221 (void __user *)tk->addr, addr_lsb);
224 * Don't use force here, it's convenient if the signal
225 * can be temporarily blocked.
226 * This could cause a loop when the user sets SIGBUS
227 * to SIG_IGN, but hopefully no one will do that?
229 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
230 addr_lsb, t); /* synchronous? */
233 pr_info("Memory failure: Error sending signal to %s:%d: %d\n",
234 t->comm, t->pid, ret);
239 * When a unknown page type is encountered drain as many buffers as possible
240 * in the hope to turn the page into a LRU or free page, which we can handle.
242 void shake_page(struct page *p, int access)
251 drain_all_pages(page_zone(p));
252 if (PageLRU(p) || is_free_buddy_page(p))
257 * Only call shrink_node_slabs here (which would also shrink
258 * other caches) if access is not potentially fatal.
261 drop_slab_node(page_to_nid(p));
263 EXPORT_SYMBOL_GPL(shake_page);
265 static unsigned long dev_pagemap_mapping_shift(struct page *page,
266 struct vm_area_struct *vma)
268 unsigned long address = vma_address(page, vma);
275 pgd = pgd_offset(vma->vm_mm, address);
276 if (!pgd_present(*pgd))
278 p4d = p4d_offset(pgd, address);
279 if (!p4d_present(*p4d))
281 pud = pud_offset(p4d, address);
282 if (!pud_present(*pud))
284 if (pud_devmap(*pud))
286 pmd = pmd_offset(pud, address);
287 if (!pmd_present(*pmd))
289 if (pmd_devmap(*pmd))
291 pte = pte_offset_map(pmd, address);
292 if (!pte_present(*pte))
294 if (pte_devmap(*pte))
300 * Failure handling: if we can't find or can't kill a process there's
301 * not much we can do. We just print a message and ignore otherwise.
305 * Schedule a process for later kill.
306 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
308 static void add_to_kill(struct task_struct *tsk, struct page *p,
309 struct vm_area_struct *vma,
310 struct list_head *to_kill)
314 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
316 pr_err("Memory failure: Out of memory while machine check handling\n");
320 tk->addr = page_address_in_vma(p, vma);
321 if (is_zone_device_page(p))
322 tk->size_shift = dev_pagemap_mapping_shift(p, vma);
324 tk->size_shift = page_shift(compound_head(p));
327 * Send SIGKILL if "tk->addr == -EFAULT". Also, as
328 * "tk->size_shift" is always non-zero for !is_zone_device_page(),
329 * so "tk->size_shift == 0" effectively checks no mapping on
330 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times
331 * to a process' address space, it's possible not all N VMAs
332 * contain mappings for the page, but at least one VMA does.
333 * Only deliver SIGBUS with payload derived from the VMA that
334 * has a mapping for the page.
336 if (tk->addr == -EFAULT) {
337 pr_info("Memory failure: Unable to find user space address %lx in %s\n",
338 page_to_pfn(p), tsk->comm);
339 } else if (tk->size_shift == 0) {
344 get_task_struct(tsk);
346 list_add_tail(&tk->nd, to_kill);
350 * Kill the processes that have been collected earlier.
352 * Only do anything when DOIT is set, otherwise just free the list
353 * (this is used for clean pages which do not need killing)
354 * Also when FAIL is set do a force kill because something went
357 static void kill_procs(struct list_head *to_kill, int forcekill, bool fail,
358 unsigned long pfn, int flags)
360 struct to_kill *tk, *next;
362 list_for_each_entry_safe (tk, next, to_kill, nd) {
365 * In case something went wrong with munmapping
366 * make sure the process doesn't catch the
367 * signal and then access the memory. Just kill it.
369 if (fail || tk->addr == -EFAULT) {
370 pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
371 pfn, tk->tsk->comm, tk->tsk->pid);
372 do_send_sig_info(SIGKILL, SEND_SIG_PRIV,
373 tk->tsk, PIDTYPE_PID);
377 * In theory the process could have mapped
378 * something else on the address in-between. We could
379 * check for that, but we need to tell the
382 else if (kill_proc(tk, pfn, flags) < 0)
383 pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n",
384 pfn, tk->tsk->comm, tk->tsk->pid);
386 put_task_struct(tk->tsk);
392 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
393 * on behalf of the thread group. Return task_struct of the (first found)
394 * dedicated thread if found, and return NULL otherwise.
396 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
397 * have to call rcu_read_lock/unlock() in this function.
399 static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
401 struct task_struct *t;
403 for_each_thread(tsk, t) {
404 if (t->flags & PF_MCE_PROCESS) {
405 if (t->flags & PF_MCE_EARLY)
408 if (sysctl_memory_failure_early_kill)
416 * Determine whether a given process is "early kill" process which expects
417 * to be signaled when some page under the process is hwpoisoned.
418 * Return task_struct of the dedicated thread (main thread unless explicitly
419 * specified) if the process is "early kill," and otherwise returns NULL.
421 * Note that the above is true for Action Optional case, but not for Action
422 * Required case where SIGBUS should sent only to the current thread.
424 static struct task_struct *task_early_kill(struct task_struct *tsk,
431 * Comparing ->mm here because current task might represent
432 * a subthread, while tsk always points to the main thread.
434 if (tsk->mm == current->mm)
439 return find_early_kill_thread(tsk);
443 * Collect processes when the error hit an anonymous page.
445 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
448 struct vm_area_struct *vma;
449 struct task_struct *tsk;
453 av = page_lock_anon_vma_read(page);
454 if (av == NULL) /* Not actually mapped anymore */
457 pgoff = page_to_pgoff(page);
458 read_lock(&tasklist_lock);
459 for_each_process (tsk) {
460 struct anon_vma_chain *vmac;
461 struct task_struct *t = task_early_kill(tsk, force_early);
465 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
468 if (!page_mapped_in_vma(page, vma))
470 if (vma->vm_mm == t->mm)
471 add_to_kill(t, page, vma, to_kill);
474 read_unlock(&tasklist_lock);
475 page_unlock_anon_vma_read(av);
479 * Collect processes when the error hit a file mapped page.
481 static void collect_procs_file(struct page *page, struct list_head *to_kill,
484 struct vm_area_struct *vma;
485 struct task_struct *tsk;
486 struct address_space *mapping = page->mapping;
489 i_mmap_lock_read(mapping);
490 read_lock(&tasklist_lock);
491 pgoff = page_to_pgoff(page);
492 for_each_process(tsk) {
493 struct task_struct *t = task_early_kill(tsk, force_early);
497 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
500 * Send early kill signal to tasks where a vma covers
501 * the page but the corrupted page is not necessarily
502 * mapped it in its pte.
503 * Assume applications who requested early kill want
504 * to be informed of all such data corruptions.
506 if (vma->vm_mm == t->mm)
507 add_to_kill(t, page, vma, to_kill);
510 read_unlock(&tasklist_lock);
511 i_mmap_unlock_read(mapping);
515 * Collect the processes who have the corrupted page mapped to kill.
517 static void collect_procs(struct page *page, struct list_head *tokill,
524 collect_procs_anon(page, tokill, force_early);
526 collect_procs_file(page, tokill, force_early);
529 static const char *action_name[] = {
530 [MF_IGNORED] = "Ignored",
531 [MF_FAILED] = "Failed",
532 [MF_DELAYED] = "Delayed",
533 [MF_RECOVERED] = "Recovered",
536 static const char * const action_page_types[] = {
537 [MF_MSG_KERNEL] = "reserved kernel page",
538 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
539 [MF_MSG_SLAB] = "kernel slab page",
540 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
541 [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned",
542 [MF_MSG_HUGE] = "huge page",
543 [MF_MSG_FREE_HUGE] = "free huge page",
544 [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page",
545 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
546 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
547 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
548 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
549 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
550 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
551 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
552 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
553 [MF_MSG_CLEAN_LRU] = "clean LRU page",
554 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
555 [MF_MSG_BUDDY] = "free buddy page",
556 [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)",
557 [MF_MSG_DAX] = "dax page",
558 [MF_MSG_UNKNOWN] = "unknown page",
562 * XXX: It is possible that a page is isolated from LRU cache,
563 * and then kept in swap cache or failed to remove from page cache.
564 * The page count will stop it from being freed by unpoison.
565 * Stress tests should be aware of this memory leak problem.
567 static int delete_from_lru_cache(struct page *p)
569 if (!isolate_lru_page(p)) {
571 * Clear sensible page flags, so that the buddy system won't
572 * complain when the page is unpoison-and-freed.
575 ClearPageUnevictable(p);
578 * Poisoned page might never drop its ref count to 0 so we have
579 * to uncharge it manually from its memcg.
581 mem_cgroup_uncharge(p);
584 * drop the page count elevated by isolate_lru_page()
592 static int truncate_error_page(struct page *p, unsigned long pfn,
593 struct address_space *mapping)
597 if (mapping->a_ops->error_remove_page) {
598 int err = mapping->a_ops->error_remove_page(mapping, p);
601 pr_info("Memory failure: %#lx: Failed to punch page: %d\n",
603 } else if (page_has_private(p) &&
604 !try_to_release_page(p, GFP_NOIO)) {
605 pr_info("Memory failure: %#lx: failed to release buffers\n",
612 * If the file system doesn't support it just invalidate
613 * This fails on dirty or anything with private pages
615 if (invalidate_inode_page(p))
618 pr_info("Memory failure: %#lx: Failed to invalidate\n",
626 * Error hit kernel page.
627 * Do nothing, try to be lucky and not touch this instead. For a few cases we
628 * could be more sophisticated.
630 static int me_kernel(struct page *p, unsigned long pfn)
636 * Page in unknown state. Do nothing.
638 static int me_unknown(struct page *p, unsigned long pfn)
640 pr_err("Memory failure: %#lx: Unknown page state\n", pfn);
645 * Clean (or cleaned) page cache page.
647 static int me_pagecache_clean(struct page *p, unsigned long pfn)
649 struct address_space *mapping;
651 delete_from_lru_cache(p);
654 * For anonymous pages we're done the only reference left
655 * should be the one m_f() holds.
661 * Now truncate the page in the page cache. This is really
662 * more like a "temporary hole punch"
663 * Don't do this for block devices when someone else
664 * has a reference, because it could be file system metadata
665 * and that's not safe to truncate.
667 mapping = page_mapping(p);
670 * Page has been teared down in the meanwhile
676 * Truncation is a bit tricky. Enable it per file system for now.
678 * Open: to take i_mutex or not for this? Right now we don't.
680 return truncate_error_page(p, pfn, mapping);
684 * Dirty pagecache page
685 * Issues: when the error hit a hole page the error is not properly
688 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
690 struct address_space *mapping = page_mapping(p);
693 /* TBD: print more information about the file. */
696 * IO error will be reported by write(), fsync(), etc.
697 * who check the mapping.
698 * This way the application knows that something went
699 * wrong with its dirty file data.
701 * There's one open issue:
703 * The EIO will be only reported on the next IO
704 * operation and then cleared through the IO map.
705 * Normally Linux has two mechanisms to pass IO error
706 * first through the AS_EIO flag in the address space
707 * and then through the PageError flag in the page.
708 * Since we drop pages on memory failure handling the
709 * only mechanism open to use is through AS_AIO.
711 * This has the disadvantage that it gets cleared on
712 * the first operation that returns an error, while
713 * the PageError bit is more sticky and only cleared
714 * when the page is reread or dropped. If an
715 * application assumes it will always get error on
716 * fsync, but does other operations on the fd before
717 * and the page is dropped between then the error
718 * will not be properly reported.
720 * This can already happen even without hwpoisoned
721 * pages: first on metadata IO errors (which only
722 * report through AS_EIO) or when the page is dropped
725 * So right now we assume that the application DTRT on
726 * the first EIO, but we're not worse than other parts
729 mapping_set_error(mapping, -EIO);
732 return me_pagecache_clean(p, pfn);
736 * Clean and dirty swap cache.
738 * Dirty swap cache page is tricky to handle. The page could live both in page
739 * cache and swap cache(ie. page is freshly swapped in). So it could be
740 * referenced concurrently by 2 types of PTEs:
741 * normal PTEs and swap PTEs. We try to handle them consistently by calling
742 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
744 * - clear dirty bit to prevent IO
746 * - but keep in the swap cache, so that when we return to it on
747 * a later page fault, we know the application is accessing
748 * corrupted data and shall be killed (we installed simple
749 * interception code in do_swap_page to catch it).
751 * Clean swap cache pages can be directly isolated. A later page fault will
752 * bring in the known good data from disk.
754 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
757 /* Trigger EIO in shmem: */
758 ClearPageUptodate(p);
760 if (!delete_from_lru_cache(p))
766 static int me_swapcache_clean(struct page *p, unsigned long pfn)
768 delete_from_swap_cache(p);
770 if (!delete_from_lru_cache(p))
777 * Huge pages. Needs work.
779 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
780 * To narrow down kill region to one page, we need to break up pmd.
782 static int me_huge_page(struct page *p, unsigned long pfn)
785 struct page *hpage = compound_head(p);
786 struct address_space *mapping;
788 if (!PageHuge(hpage))
791 mapping = page_mapping(hpage);
793 res = truncate_error_page(hpage, pfn, mapping);
797 * migration entry prevents later access on error anonymous
798 * hugepage, so we can free and dissolve it into buddy to
799 * save healthy subpages.
803 dissolve_free_huge_page(p);
812 * Various page states we can handle.
814 * A page state is defined by its current page->flags bits.
815 * The table matches them in order and calls the right handler.
817 * This is quite tricky because we can access page at any time
818 * in its live cycle, so all accesses have to be extremely careful.
820 * This is not complete. More states could be added.
821 * For any missing state don't attempt recovery.
824 #define dirty (1UL << PG_dirty)
825 #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked))
826 #define unevict (1UL << PG_unevictable)
827 #define mlock (1UL << PG_mlocked)
828 #define lru (1UL << PG_lru)
829 #define head (1UL << PG_head)
830 #define slab (1UL << PG_slab)
831 #define reserved (1UL << PG_reserved)
833 static struct page_state {
836 enum mf_action_page_type type;
837 int (*action)(struct page *p, unsigned long pfn);
839 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
841 * free pages are specially detected outside this table:
842 * PG_buddy pages only make a small fraction of all free pages.
846 * Could in theory check if slab page is free or if we can drop
847 * currently unused objects without touching them. But just
848 * treat it as standard kernel for now.
850 { slab, slab, MF_MSG_SLAB, me_kernel },
852 { head, head, MF_MSG_HUGE, me_huge_page },
854 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
855 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
857 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
858 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
860 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
861 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
863 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
864 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
867 * Catchall entry: must be at end.
869 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
882 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
883 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
885 static void action_result(unsigned long pfn, enum mf_action_page_type type,
886 enum mf_result result)
888 trace_memory_failure_event(pfn, type, result);
890 pr_err("Memory failure: %#lx: recovery action for %s: %s\n",
891 pfn, action_page_types[type], action_name[result]);
894 static int page_action(struct page_state *ps, struct page *p,
900 result = ps->action(p, pfn);
902 count = page_count(p) - 1;
903 if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
906 pr_err("Memory failure: %#lx: %s still referenced by %d users\n",
907 pfn, action_page_types[ps->type], count);
910 action_result(pfn, ps->type, result);
912 /* Could do more checks here if page looks ok */
914 * Could adjust zone counters here to correct for the missing page.
917 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
921 * get_hwpoison_page() - Get refcount for memory error handling:
922 * @page: raw error page (hit by memory error)
924 * Return: return 0 if failed to grab the refcount, otherwise true (some
927 int get_hwpoison_page(struct page *page)
929 struct page *head = compound_head(page);
931 if (!PageHuge(head) && PageTransHuge(head)) {
933 * Non anonymous thp exists only in allocation/free time. We
934 * can't handle such a case correctly, so let's give it up.
935 * This should be better than triggering BUG_ON when kernel
936 * tries to touch the "partially handled" page.
938 if (!PageAnon(head)) {
939 pr_err("Memory failure: %#lx: non anonymous thp\n",
945 if (get_page_unless_zero(head)) {
946 if (head == compound_head(page))
949 pr_info("Memory failure: %#lx cannot catch tail\n",
956 EXPORT_SYMBOL_GPL(get_hwpoison_page);
959 * Do all that is necessary to remove user space mappings. Unmap
960 * the pages and send SIGBUS to the processes if the data was dirty.
962 static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
963 int flags, struct page **hpagep)
965 enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
966 struct address_space *mapping;
968 bool unmap_success = true;
969 int kill = 1, forcekill;
970 struct page *hpage = *hpagep;
971 bool mlocked = PageMlocked(hpage);
974 * Here we are interested only in user-mapped pages, so skip any
975 * other types of pages.
977 if (PageReserved(p) || PageSlab(p))
979 if (!(PageLRU(hpage) || PageHuge(p)))
983 * This check implies we don't kill processes if their pages
984 * are in the swap cache early. Those are always late kills.
986 if (!page_mapped(hpage))
990 pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn);
994 if (PageSwapCache(p)) {
995 pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n",
997 ttu |= TTU_IGNORE_HWPOISON;
1001 * Propagate the dirty bit from PTEs to struct page first, because we
1002 * need this to decide if we should kill or just drop the page.
1003 * XXX: the dirty test could be racy: set_page_dirty() may not always
1004 * be called inside page lock (it's recommended but not enforced).
1006 mapping = page_mapping(hpage);
1007 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
1008 mapping_can_writeback(mapping)) {
1009 if (page_mkclean(hpage)) {
1010 SetPageDirty(hpage);
1013 ttu |= TTU_IGNORE_HWPOISON;
1014 pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n",
1020 * First collect all the processes that have the page
1021 * mapped in dirty form. This has to be done before try_to_unmap,
1022 * because ttu takes the rmap data structures down.
1024 * Error handling: We ignore errors here because
1025 * there's nothing that can be done.
1028 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
1030 if (!PageHuge(hpage)) {
1031 unmap_success = try_to_unmap(hpage, ttu);
1034 * For hugetlb pages, try_to_unmap could potentially call
1035 * huge_pmd_unshare. Because of this, take semaphore in
1036 * write mode here and set TTU_RMAP_LOCKED to indicate we
1037 * have taken the lock at this higer level.
1039 * Note that the call to hugetlb_page_mapping_lock_write
1040 * is necessary even if mapping is already set. It handles
1041 * ugliness of potentially having to drop page lock to obtain
1044 mapping = hugetlb_page_mapping_lock_write(hpage);
1047 unmap_success = try_to_unmap(hpage,
1048 ttu|TTU_RMAP_LOCKED);
1049 i_mmap_unlock_write(mapping);
1051 pr_info("Memory failure: %#lx: could not find mapping for mapped huge page\n",
1053 unmap_success = false;
1057 pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n",
1058 pfn, page_mapcount(hpage));
1061 * try_to_unmap() might put mlocked page in lru cache, so call
1062 * shake_page() again to ensure that it's flushed.
1065 shake_page(hpage, 0);
1068 * Now that the dirty bit has been propagated to the
1069 * struct page and all unmaps done we can decide if
1070 * killing is needed or not. Only kill when the page
1071 * was dirty or the process is not restartable,
1072 * otherwise the tokill list is merely
1073 * freed. When there was a problem unmapping earlier
1074 * use a more force-full uncatchable kill to prevent
1075 * any accesses to the poisoned memory.
1077 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
1078 kill_procs(&tokill, forcekill, !unmap_success, pfn, flags);
1080 return unmap_success;
1083 static int identify_page_state(unsigned long pfn, struct page *p,
1084 unsigned long page_flags)
1086 struct page_state *ps;
1089 * The first check uses the current page flags which may not have any
1090 * relevant information. The second check with the saved page flags is
1091 * carried out only if the first check can't determine the page status.
1093 for (ps = error_states;; ps++)
1094 if ((p->flags & ps->mask) == ps->res)
1097 page_flags |= (p->flags & (1UL << PG_dirty));
1100 for (ps = error_states;; ps++)
1101 if ((page_flags & ps->mask) == ps->res)
1103 return page_action(ps, p, pfn);
1106 static int memory_failure_hugetlb(unsigned long pfn, int flags)
1108 struct page *p = pfn_to_page(pfn);
1109 struct page *head = compound_head(p);
1111 unsigned long page_flags;
1113 if (TestSetPageHWPoison(head)) {
1114 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1119 num_poisoned_pages_inc();
1121 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
1123 * Check "filter hit" and "race with other subpage."
1126 if (PageHWPoison(head)) {
1127 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1128 || (p != head && TestSetPageHWPoison(head))) {
1129 num_poisoned_pages_dec();
1135 dissolve_free_huge_page(p);
1136 action_result(pfn, MF_MSG_FREE_HUGE, MF_DELAYED);
1141 page_flags = head->flags;
1143 if (!PageHWPoison(head)) {
1144 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1145 num_poisoned_pages_dec();
1147 put_hwpoison_page(head);
1152 * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so
1153 * simply disable it. In order to make it work properly, we need
1155 * - conversion of a pud that maps an error hugetlb into hwpoison
1156 * entry properly works, and
1157 * - other mm code walking over page table is aware of pud-aligned
1160 if (huge_page_size(page_hstate(head)) > PMD_SIZE) {
1161 action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED);
1166 if (!hwpoison_user_mappings(p, pfn, flags, &head)) {
1167 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1172 res = identify_page_state(pfn, p, page_flags);
1178 static int memory_failure_dev_pagemap(unsigned long pfn, int flags,
1179 struct dev_pagemap *pgmap)
1181 struct page *page = pfn_to_page(pfn);
1182 const bool unmap_success = true;
1183 unsigned long size = 0;
1191 * Prevent the inode from being freed while we are interrogating
1192 * the address_space, typically this would be handled by
1193 * lock_page(), but dax pages do not use the page lock. This
1194 * also prevents changes to the mapping of this pfn until
1195 * poison signaling is complete.
1197 cookie = dax_lock_page(page);
1201 if (hwpoison_filter(page)) {
1206 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
1208 * TODO: Handle HMM pages which may need coordination
1209 * with device-side memory.
1215 * Use this flag as an indication that the dax page has been
1216 * remapped UC to prevent speculative consumption of poison.
1218 SetPageHWPoison(page);
1221 * Unlike System-RAM there is no possibility to swap in a
1222 * different physical page at a given virtual address, so all
1223 * userspace consumption of ZONE_DEVICE memory necessitates
1224 * SIGBUS (i.e. MF_MUST_KILL)
1226 flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1227 collect_procs(page, &tokill, flags & MF_ACTION_REQUIRED);
1229 list_for_each_entry(tk, &tokill, nd)
1231 size = max(size, 1UL << tk->size_shift);
1234 * Unmap the largest mapping to avoid breaking up
1235 * device-dax mappings which are constant size. The
1236 * actual size of the mapping being torn down is
1237 * communicated in siginfo, see kill_proc()
1239 start = (page->index << PAGE_SHIFT) & ~(size - 1);
1240 unmap_mapping_range(page->mapping, start, start + size, 0);
1242 kill_procs(&tokill, flags & MF_MUST_KILL, !unmap_success, pfn, flags);
1245 dax_unlock_page(page, cookie);
1247 /* drop pgmap ref acquired in caller */
1248 put_dev_pagemap(pgmap);
1249 action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED);
1254 * memory_failure - Handle memory failure of a page.
1255 * @pfn: Page Number of the corrupted page
1256 * @flags: fine tune action taken
1258 * This function is called by the low level machine check code
1259 * of an architecture when it detects hardware memory corruption
1260 * of a page. It tries its best to recover, which includes
1261 * dropping pages, killing processes etc.
1263 * The function is primarily of use for corruptions that
1264 * happen outside the current execution context (e.g. when
1265 * detected by a background scrubber)
1267 * Must run in process context (e.g. a work queue) with interrupts
1268 * enabled and no spinlocks hold.
1270 int memory_failure(unsigned long pfn, int flags)
1274 struct page *orig_head;
1275 struct dev_pagemap *pgmap;
1277 unsigned long page_flags;
1279 if (!sysctl_memory_failure_recovery)
1280 panic("Memory failure on page %lx", pfn);
1282 p = pfn_to_online_page(pfn);
1284 if (pfn_valid(pfn)) {
1285 pgmap = get_dev_pagemap(pfn, NULL);
1287 return memory_failure_dev_pagemap(pfn, flags,
1290 pr_err("Memory failure: %#lx: memory outside kernel control\n",
1296 return memory_failure_hugetlb(pfn, flags);
1297 if (TestSetPageHWPoison(p)) {
1298 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1303 orig_head = hpage = compound_head(p);
1304 num_poisoned_pages_inc();
1307 * We need/can do nothing about count=0 pages.
1308 * 1) it's a free page, and therefore in safe hand:
1309 * prep_new_page() will be the gate keeper.
1310 * 2) it's part of a non-compound high order page.
1311 * Implies some kernel user: cannot stop them from
1312 * R/W the page; let's pray that the page has been
1313 * used and will be freed some time later.
1314 * In fact it's dangerous to directly bump up page count from 0,
1315 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
1317 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
1318 if (is_free_buddy_page(p)) {
1319 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1322 action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
1327 if (PageTransHuge(hpage)) {
1329 if (!PageAnon(p) || unlikely(split_huge_page(p))) {
1332 pr_err("Memory failure: %#lx: non anonymous thp\n",
1335 pr_err("Memory failure: %#lx: thp split failed\n",
1337 if (TestClearPageHWPoison(p))
1338 num_poisoned_pages_dec();
1339 put_hwpoison_page(p);
1343 VM_BUG_ON_PAGE(!page_count(p), p);
1347 * We ignore non-LRU pages for good reasons.
1348 * - PG_locked is only well defined for LRU pages and a few others
1349 * - to avoid races with __SetPageLocked()
1350 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1351 * The check (unnecessarily) ignores LRU pages being isolated and
1352 * walked by the page reclaim code, however that's not a big loss.
1355 /* shake_page could have turned it free. */
1356 if (!PageLRU(p) && is_free_buddy_page(p)) {
1357 if (flags & MF_COUNT_INCREASED)
1358 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1360 action_result(pfn, MF_MSG_BUDDY_2ND, MF_DELAYED);
1367 * The page could have changed compound pages during the locking.
1368 * If this happens just bail out.
1370 if (PageCompound(p) && compound_head(p) != orig_head) {
1371 action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
1377 * We use page flags to determine what action should be taken, but
1378 * the flags can be modified by the error containment action. One
1379 * example is an mlocked page, where PG_mlocked is cleared by
1380 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1381 * correctly, we save a copy of the page flags at this time.
1383 page_flags = p->flags;
1386 * unpoison always clear PG_hwpoison inside page lock
1388 if (!PageHWPoison(p)) {
1389 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1390 num_poisoned_pages_dec();
1392 put_hwpoison_page(p);
1395 if (hwpoison_filter(p)) {
1396 if (TestClearPageHWPoison(p))
1397 num_poisoned_pages_dec();
1399 put_hwpoison_page(p);
1403 if (!PageTransTail(p) && !PageLRU(p))
1404 goto identify_page_state;
1407 * It's very difficult to mess with pages currently under IO
1408 * and in many cases impossible, so we just avoid it here.
1410 wait_on_page_writeback(p);
1413 * Now take care of user space mappings.
1414 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1416 if (!hwpoison_user_mappings(p, pfn, flags, &p)) {
1417 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1423 * Torn down by someone else?
1425 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1426 action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
1431 identify_page_state:
1432 res = identify_page_state(pfn, p, page_flags);
1437 EXPORT_SYMBOL_GPL(memory_failure);
1439 #define MEMORY_FAILURE_FIFO_ORDER 4
1440 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1442 struct memory_failure_entry {
1447 struct memory_failure_cpu {
1448 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1449 MEMORY_FAILURE_FIFO_SIZE);
1451 struct work_struct work;
1454 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1457 * memory_failure_queue - Schedule handling memory failure of a page.
1458 * @pfn: Page Number of the corrupted page
1459 * @flags: Flags for memory failure handling
1461 * This function is called by the low level hardware error handler
1462 * when it detects hardware memory corruption of a page. It schedules
1463 * the recovering of error page, including dropping pages, killing
1466 * The function is primarily of use for corruptions that
1467 * happen outside the current execution context (e.g. when
1468 * detected by a background scrubber)
1470 * Can run in IRQ context.
1472 void memory_failure_queue(unsigned long pfn, int flags)
1474 struct memory_failure_cpu *mf_cpu;
1475 unsigned long proc_flags;
1476 struct memory_failure_entry entry = {
1481 mf_cpu = &get_cpu_var(memory_failure_cpu);
1482 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1483 if (kfifo_put(&mf_cpu->fifo, entry))
1484 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1486 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1488 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1489 put_cpu_var(memory_failure_cpu);
1491 EXPORT_SYMBOL_GPL(memory_failure_queue);
1493 static void memory_failure_work_func(struct work_struct *work)
1495 struct memory_failure_cpu *mf_cpu;
1496 struct memory_failure_entry entry = { 0, };
1497 unsigned long proc_flags;
1500 mf_cpu = container_of(work, struct memory_failure_cpu, work);
1502 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1503 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1504 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1507 if (entry.flags & MF_SOFT_OFFLINE)
1508 soft_offline_page(entry.pfn, entry.flags);
1510 memory_failure(entry.pfn, entry.flags);
1515 * Process memory_failure work queued on the specified CPU.
1516 * Used to avoid return-to-userspace racing with the memory_failure workqueue.
1518 void memory_failure_queue_kick(int cpu)
1520 struct memory_failure_cpu *mf_cpu;
1522 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1523 cancel_work_sync(&mf_cpu->work);
1524 memory_failure_work_func(&mf_cpu->work);
1527 static int __init memory_failure_init(void)
1529 struct memory_failure_cpu *mf_cpu;
1532 for_each_possible_cpu(cpu) {
1533 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1534 spin_lock_init(&mf_cpu->lock);
1535 INIT_KFIFO(mf_cpu->fifo);
1536 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1541 core_initcall(memory_failure_init);
1543 #define unpoison_pr_info(fmt, pfn, rs) \
1545 if (__ratelimit(rs)) \
1546 pr_info(fmt, pfn); \
1550 * unpoison_memory - Unpoison a previously poisoned page
1551 * @pfn: Page number of the to be unpoisoned page
1553 * Software-unpoison a page that has been poisoned by
1554 * memory_failure() earlier.
1556 * This is only done on the software-level, so it only works
1557 * for linux injected failures, not real hardware failures
1559 * Returns 0 for success, otherwise -errno.
1561 int unpoison_memory(unsigned long pfn)
1566 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
1567 DEFAULT_RATELIMIT_BURST);
1569 if (!pfn_valid(pfn))
1572 p = pfn_to_page(pfn);
1573 page = compound_head(p);
1575 if (!PageHWPoison(p)) {
1576 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
1581 if (page_count(page) > 1) {
1582 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
1587 if (page_mapped(page)) {
1588 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
1593 if (page_mapping(page)) {
1594 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
1600 * unpoison_memory() can encounter thp only when the thp is being
1601 * worked by memory_failure() and the page lock is not held yet.
1602 * In such case, we yield to memory_failure() and make unpoison fail.
1604 if (!PageHuge(page) && PageTransHuge(page)) {
1605 unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n",
1610 if (!get_hwpoison_page(p)) {
1611 if (TestClearPageHWPoison(p))
1612 num_poisoned_pages_dec();
1613 unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n",
1620 * This test is racy because PG_hwpoison is set outside of page lock.
1621 * That's acceptable because that won't trigger kernel panic. Instead,
1622 * the PG_hwpoison page will be caught and isolated on the entrance to
1623 * the free buddy page pool.
1625 if (TestClearPageHWPoison(page)) {
1626 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
1628 num_poisoned_pages_dec();
1633 put_hwpoison_page(page);
1634 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
1635 put_hwpoison_page(page);
1639 EXPORT_SYMBOL(unpoison_memory);
1641 static struct page *new_page(struct page *p, unsigned long private)
1643 struct migration_target_control mtc = {
1644 .nid = page_to_nid(p),
1645 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1648 return alloc_migration_target(p, (unsigned long)&mtc);
1652 * Safely get reference count of an arbitrary page.
1653 * Returns 0 for a free page, -EIO for a zero refcount page
1654 * that is not free, and 1 for any other page type.
1655 * For 1 the page is returned with increased page count, otherwise not.
1657 static int __get_any_page(struct page *p, unsigned long pfn, int flags)
1661 if (flags & MF_COUNT_INCREASED)
1665 * When the target page is a free hugepage, just remove it
1666 * from free hugepage list.
1668 if (!get_hwpoison_page(p)) {
1670 pr_info("%s: %#lx free huge page\n", __func__, pfn);
1672 } else if (is_free_buddy_page(p)) {
1673 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1676 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1677 __func__, pfn, p->flags);
1681 /* Not a free page */
1687 static int get_any_page(struct page *page, unsigned long pfn, int flags)
1689 int ret = __get_any_page(page, pfn, flags);
1691 if (ret == 1 && !PageHuge(page) &&
1692 !PageLRU(page) && !__PageMovable(page)) {
1696 put_hwpoison_page(page);
1697 shake_page(page, 1);
1702 ret = __get_any_page(page, pfn, 0);
1703 if (ret == 1 && !PageLRU(page)) {
1704 /* Drop page reference which is from __get_any_page() */
1705 put_hwpoison_page(page);
1706 pr_info("soft_offline: %#lx: unknown non LRU page type %lx (%pGp)\n",
1707 pfn, page->flags, &page->flags);
1714 static int soft_offline_huge_page(struct page *page, int flags)
1717 unsigned long pfn = page_to_pfn(page);
1718 struct page *hpage = compound_head(page);
1719 LIST_HEAD(pagelist);
1722 * This double-check of PageHWPoison is to avoid the race with
1723 * memory_failure(). See also comment in __soft_offline_page().
1726 if (PageHWPoison(hpage)) {
1728 put_hwpoison_page(hpage);
1729 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1734 ret = isolate_huge_page(hpage, &pagelist);
1736 * get_any_page() and isolate_huge_page() takes a refcount each,
1737 * so need to drop one here.
1739 put_hwpoison_page(hpage);
1741 pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn);
1745 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1746 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1748 pr_info("soft offline: %#lx: hugepage migration failed %d, type %lx (%pGp)\n",
1749 pfn, ret, page->flags, &page->flags);
1750 if (!list_empty(&pagelist))
1751 putback_movable_pages(&pagelist);
1756 * We set PG_hwpoison only when the migration source hugepage
1757 * was successfully dissolved, because otherwise hwpoisoned
1758 * hugepage remains on free hugepage list, then userspace will
1759 * find it as SIGBUS by allocation failure. That's not expected
1760 * in soft-offlining.
1762 ret = dissolve_free_huge_page(page);
1764 if (set_hwpoison_free_buddy_page(page))
1765 num_poisoned_pages_inc();
1773 static int __soft_offline_page(struct page *page, int flags)
1776 unsigned long pfn = page_to_pfn(page);
1779 * Check PageHWPoison again inside page lock because PageHWPoison
1780 * is set by memory_failure() outside page lock. Note that
1781 * memory_failure() also double-checks PageHWPoison inside page lock,
1782 * so there's no race between soft_offline_page() and memory_failure().
1785 wait_on_page_writeback(page);
1786 if (PageHWPoison(page)) {
1788 put_hwpoison_page(page);
1789 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1793 * Try to invalidate first. This should work for
1794 * non dirty unmapped page cache pages.
1796 ret = invalidate_inode_page(page);
1799 * RED-PEN would be better to keep it isolated here, but we
1800 * would need to fix isolation locking first.
1803 put_hwpoison_page(page);
1804 pr_info("soft_offline: %#lx: invalidated\n", pfn);
1805 SetPageHWPoison(page);
1806 num_poisoned_pages_inc();
1811 * Simple invalidation didn't work.
1812 * Try to migrate to a new page instead. migrate.c
1813 * handles a large number of cases for us.
1816 ret = isolate_lru_page(page);
1818 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1820 * Drop page reference which is came from get_any_page()
1821 * successful isolate_lru_page() already took another one.
1823 put_hwpoison_page(page);
1825 LIST_HEAD(pagelist);
1827 * After isolated lru page, the PageLRU will be cleared,
1828 * so use !__PageMovable instead for LRU page's mapping
1829 * cannot have PAGE_MAPPING_MOVABLE.
1831 if (!__PageMovable(page))
1832 inc_node_page_state(page, NR_ISOLATED_ANON +
1833 page_is_file_lru(page));
1834 list_add(&page->lru, &pagelist);
1835 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1836 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1838 if (!list_empty(&pagelist))
1839 putback_movable_pages(&pagelist);
1841 pr_info("soft offline: %#lx: migration failed %d, type %lx (%pGp)\n",
1842 pfn, ret, page->flags, &page->flags);
1847 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx (%pGp)\n",
1848 pfn, ret, page_count(page), page->flags, &page->flags);
1853 static int soft_offline_in_use_page(struct page *page, int flags)
1857 struct page *hpage = compound_head(page);
1859 if (!PageHuge(page) && PageTransHuge(hpage)) {
1861 if (!PageAnon(page) || unlikely(split_huge_page(page))) {
1863 if (!PageAnon(page))
1864 pr_info("soft offline: %#lx: non anonymous thp\n", page_to_pfn(page));
1866 pr_info("soft offline: %#lx: thp split failed\n", page_to_pfn(page));
1867 put_hwpoison_page(page);
1874 * Setting MIGRATE_ISOLATE here ensures that the page will be linked
1875 * to free list immediately (not via pcplist) when released after
1876 * successful page migration. Otherwise we can't guarantee that the
1877 * page is really free after put_page() returns, so
1878 * set_hwpoison_free_buddy_page() highly likely fails.
1880 mt = get_pageblock_migratetype(page);
1881 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
1883 ret = soft_offline_huge_page(page, flags);
1885 ret = __soft_offline_page(page, flags);
1886 set_pageblock_migratetype(page, mt);
1890 static int soft_offline_free_page(struct page *page)
1892 int rc = dissolve_free_huge_page(page);
1895 if (set_hwpoison_free_buddy_page(page))
1896 num_poisoned_pages_inc();
1904 * soft_offline_page - Soft offline a page.
1905 * @pfn: pfn to soft-offline
1906 * @flags: flags. Same as memory_failure().
1908 * Returns 0 on success, otherwise negated errno.
1910 * Soft offline a page, by migration or invalidation,
1911 * without killing anything. This is for the case when
1912 * a page is not corrupted yet (so it's still valid to access),
1913 * but has had a number of corrected errors and is better taken
1916 * The actual policy on when to do that is maintained by
1919 * This should never impact any application or cause data loss,
1920 * however it might take some time.
1922 * This is not a 100% solution for all memory, but tries to be
1923 * ``good enough'' for the majority of memory.
1925 int soft_offline_page(unsigned long pfn, int flags)
1930 if (!pfn_valid(pfn))
1932 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
1933 page = pfn_to_online_page(pfn);
1937 if (PageHWPoison(page)) {
1938 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1939 if (flags & MF_COUNT_INCREASED)
1940 put_hwpoison_page(page);
1945 ret = get_any_page(page, pfn, flags);
1949 ret = soft_offline_in_use_page(page, flags);
1951 ret = soft_offline_free_page(page);