Loading...
1/*
2 * Copyright (C) 2008, 2009 Intel Corporation
3 * Authors: Andi Kleen, Fengguang Wu
4 *
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.
8 *
9 * High level machine check handler. Handles pages reported by the
10 * hardware as being corrupted usually due to a multi-bit ECC memory or cache
11 * failure.
12 *
13 * In addition there is a "soft offline" entry point that allows stop using
14 * not-yet-corrupted-by-suspicious pages without killing anything.
15 *
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.
23 *
24 * There are several operations here with exponential complexity because
25 * of unsuitable VM data structures. For example the operation to map back
26 * from RMAP chains to processes has to walk the complete process list and
27 * has non linear complexity with the number. But since memory corruptions
28 * are rare we hope to get away with this. This avoids impacting the core
29 * VM.
30 */
31
32/*
33 * Notebook:
34 * - hugetlb needs more code
35 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
36 * - pass bad pages to kdump next kernel
37 */
38#include <linux/kernel.h>
39#include <linux/mm.h>
40#include <linux/page-flags.h>
41#include <linux/kernel-page-flags.h>
42#include <linux/sched.h>
43#include <linux/ksm.h>
44#include <linux/rmap.h>
45#include <linux/pagemap.h>
46#include <linux/swap.h>
47#include <linux/backing-dev.h>
48#include <linux/migrate.h>
49#include <linux/page-isolation.h>
50#include <linux/suspend.h>
51#include <linux/slab.h>
52#include <linux/swapops.h>
53#include <linux/hugetlb.h>
54#include <linux/memory_hotplug.h>
55#include <linux/mm_inline.h>
56#include <linux/kfifo.h>
57#include "internal.h"
58
59int sysctl_memory_failure_early_kill __read_mostly = 0;
60
61int sysctl_memory_failure_recovery __read_mostly = 1;
62
63atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
64
65#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
66
67u32 hwpoison_filter_enable = 0;
68u32 hwpoison_filter_dev_major = ~0U;
69u32 hwpoison_filter_dev_minor = ~0U;
70u64 hwpoison_filter_flags_mask;
71u64 hwpoison_filter_flags_value;
72EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
73EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
74EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
75EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
76EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
77
78static int hwpoison_filter_dev(struct page *p)
79{
80 struct address_space *mapping;
81 dev_t dev;
82
83 if (hwpoison_filter_dev_major == ~0U &&
84 hwpoison_filter_dev_minor == ~0U)
85 return 0;
86
87 /*
88 * page_mapping() does not accept slab pages.
89 */
90 if (PageSlab(p))
91 return -EINVAL;
92
93 mapping = page_mapping(p);
94 if (mapping == NULL || mapping->host == NULL)
95 return -EINVAL;
96
97 dev = mapping->host->i_sb->s_dev;
98 if (hwpoison_filter_dev_major != ~0U &&
99 hwpoison_filter_dev_major != MAJOR(dev))
100 return -EINVAL;
101 if (hwpoison_filter_dev_minor != ~0U &&
102 hwpoison_filter_dev_minor != MINOR(dev))
103 return -EINVAL;
104
105 return 0;
106}
107
108static int hwpoison_filter_flags(struct page *p)
109{
110 if (!hwpoison_filter_flags_mask)
111 return 0;
112
113 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
114 hwpoison_filter_flags_value)
115 return 0;
116 else
117 return -EINVAL;
118}
119
120/*
121 * This allows stress tests to limit test scope to a collection of tasks
122 * by putting them under some memcg. This prevents killing unrelated/important
123 * processes such as /sbin/init. Note that the target task may share clean
124 * pages with init (eg. libc text), which is harmless. If the target task
125 * share _dirty_ pages with another task B, the test scheme must make sure B
126 * is also included in the memcg. At last, due to race conditions this filter
127 * can only guarantee that the page either belongs to the memcg tasks, or is
128 * a freed page.
129 */
130#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
131u64 hwpoison_filter_memcg;
132EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
133static int hwpoison_filter_task(struct page *p)
134{
135 struct mem_cgroup *mem;
136 struct cgroup_subsys_state *css;
137 unsigned long ino;
138
139 if (!hwpoison_filter_memcg)
140 return 0;
141
142 mem = try_get_mem_cgroup_from_page(p);
143 if (!mem)
144 return -EINVAL;
145
146 css = mem_cgroup_css(mem);
147 /* root_mem_cgroup has NULL dentries */
148 if (!css->cgroup->dentry)
149 return -EINVAL;
150
151 ino = css->cgroup->dentry->d_inode->i_ino;
152 css_put(css);
153
154 if (ino != hwpoison_filter_memcg)
155 return -EINVAL;
156
157 return 0;
158}
159#else
160static int hwpoison_filter_task(struct page *p) { return 0; }
161#endif
162
163int hwpoison_filter(struct page *p)
164{
165 if (!hwpoison_filter_enable)
166 return 0;
167
168 if (hwpoison_filter_dev(p))
169 return -EINVAL;
170
171 if (hwpoison_filter_flags(p))
172 return -EINVAL;
173
174 if (hwpoison_filter_task(p))
175 return -EINVAL;
176
177 return 0;
178}
179#else
180int hwpoison_filter(struct page *p)
181{
182 return 0;
183}
184#endif
185
186EXPORT_SYMBOL_GPL(hwpoison_filter);
187
188/*
189 * Send all the processes who have the page mapped an ``action optional''
190 * signal.
191 */
192static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
193 unsigned long pfn, struct page *page)
194{
195 struct siginfo si;
196 int ret;
197
198 printk(KERN_ERR
199 "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
200 pfn, t->comm, t->pid);
201 si.si_signo = SIGBUS;
202 si.si_errno = 0;
203 si.si_code = BUS_MCEERR_AO;
204 si.si_addr = (void *)addr;
205#ifdef __ARCH_SI_TRAPNO
206 si.si_trapno = trapno;
207#endif
208 si.si_addr_lsb = compound_trans_order(compound_head(page)) + PAGE_SHIFT;
209 /*
210 * Don't use force here, it's convenient if the signal
211 * can be temporarily blocked.
212 * This could cause a loop when the user sets SIGBUS
213 * to SIG_IGN, but hopefully no one will do that?
214 */
215 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
216 if (ret < 0)
217 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
218 t->comm, t->pid, ret);
219 return ret;
220}
221
222/*
223 * When a unknown page type is encountered drain as many buffers as possible
224 * in the hope to turn the page into a LRU or free page, which we can handle.
225 */
226void shake_page(struct page *p, int access)
227{
228 if (!PageSlab(p)) {
229 lru_add_drain_all();
230 if (PageLRU(p))
231 return;
232 drain_all_pages();
233 if (PageLRU(p) || is_free_buddy_page(p))
234 return;
235 }
236
237 /*
238 * Only call shrink_slab here (which would also shrink other caches) if
239 * access is not potentially fatal.
240 */
241 if (access) {
242 int nr;
243 do {
244 struct shrink_control shrink = {
245 .gfp_mask = GFP_KERNEL,
246 };
247
248 nr = shrink_slab(&shrink, 1000, 1000);
249 if (page_count(p) == 1)
250 break;
251 } while (nr > 10);
252 }
253}
254EXPORT_SYMBOL_GPL(shake_page);
255
256/*
257 * Kill all processes that have a poisoned page mapped and then isolate
258 * the page.
259 *
260 * General strategy:
261 * Find all processes having the page mapped and kill them.
262 * But we keep a page reference around so that the page is not
263 * actually freed yet.
264 * Then stash the page away
265 *
266 * There's no convenient way to get back to mapped processes
267 * from the VMAs. So do a brute-force search over all
268 * running processes.
269 *
270 * Remember that machine checks are not common (or rather
271 * if they are common you have other problems), so this shouldn't
272 * be a performance issue.
273 *
274 * Also there are some races possible while we get from the
275 * error detection to actually handle it.
276 */
277
278struct to_kill {
279 struct list_head nd;
280 struct task_struct *tsk;
281 unsigned long addr;
282 char addr_valid;
283};
284
285/*
286 * Failure handling: if we can't find or can't kill a process there's
287 * not much we can do. We just print a message and ignore otherwise.
288 */
289
290/*
291 * Schedule a process for later kill.
292 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
293 * TBD would GFP_NOIO be enough?
294 */
295static void add_to_kill(struct task_struct *tsk, struct page *p,
296 struct vm_area_struct *vma,
297 struct list_head *to_kill,
298 struct to_kill **tkc)
299{
300 struct to_kill *tk;
301
302 if (*tkc) {
303 tk = *tkc;
304 *tkc = NULL;
305 } else {
306 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
307 if (!tk) {
308 printk(KERN_ERR
309 "MCE: Out of memory while machine check handling\n");
310 return;
311 }
312 }
313 tk->addr = page_address_in_vma(p, vma);
314 tk->addr_valid = 1;
315
316 /*
317 * In theory we don't have to kill when the page was
318 * munmaped. But it could be also a mremap. Since that's
319 * likely very rare kill anyways just out of paranoia, but use
320 * a SIGKILL because the error is not contained anymore.
321 */
322 if (tk->addr == -EFAULT) {
323 pr_info("MCE: Unable to find user space address %lx in %s\n",
324 page_to_pfn(p), tsk->comm);
325 tk->addr_valid = 0;
326 }
327 get_task_struct(tsk);
328 tk->tsk = tsk;
329 list_add_tail(&tk->nd, to_kill);
330}
331
332/*
333 * Kill the processes that have been collected earlier.
334 *
335 * Only do anything when DOIT is set, otherwise just free the list
336 * (this is used for clean pages which do not need killing)
337 * Also when FAIL is set do a force kill because something went
338 * wrong earlier.
339 */
340static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
341 int fail, struct page *page, unsigned long pfn)
342{
343 struct to_kill *tk, *next;
344
345 list_for_each_entry_safe (tk, next, to_kill, nd) {
346 if (doit) {
347 /*
348 * In case something went wrong with munmapping
349 * make sure the process doesn't catch the
350 * signal and then access the memory. Just kill it.
351 */
352 if (fail || tk->addr_valid == 0) {
353 printk(KERN_ERR
354 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
355 pfn, tk->tsk->comm, tk->tsk->pid);
356 force_sig(SIGKILL, tk->tsk);
357 }
358
359 /*
360 * In theory the process could have mapped
361 * something else on the address in-between. We could
362 * check for that, but we need to tell the
363 * process anyways.
364 */
365 else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
366 pfn, page) < 0)
367 printk(KERN_ERR
368 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
369 pfn, tk->tsk->comm, tk->tsk->pid);
370 }
371 put_task_struct(tk->tsk);
372 kfree(tk);
373 }
374}
375
376static int task_early_kill(struct task_struct *tsk)
377{
378 if (!tsk->mm)
379 return 0;
380 if (tsk->flags & PF_MCE_PROCESS)
381 return !!(tsk->flags & PF_MCE_EARLY);
382 return sysctl_memory_failure_early_kill;
383}
384
385/*
386 * Collect processes when the error hit an anonymous page.
387 */
388static void collect_procs_anon(struct page *page, struct list_head *to_kill,
389 struct to_kill **tkc)
390{
391 struct vm_area_struct *vma;
392 struct task_struct *tsk;
393 struct anon_vma *av;
394
395 av = page_lock_anon_vma(page);
396 if (av == NULL) /* Not actually mapped anymore */
397 return;
398
399 read_lock(&tasklist_lock);
400 for_each_process (tsk) {
401 struct anon_vma_chain *vmac;
402
403 if (!task_early_kill(tsk))
404 continue;
405 list_for_each_entry(vmac, &av->head, same_anon_vma) {
406 vma = vmac->vma;
407 if (!page_mapped_in_vma(page, vma))
408 continue;
409 if (vma->vm_mm == tsk->mm)
410 add_to_kill(tsk, page, vma, to_kill, tkc);
411 }
412 }
413 read_unlock(&tasklist_lock);
414 page_unlock_anon_vma(av);
415}
416
417/*
418 * Collect processes when the error hit a file mapped page.
419 */
420static void collect_procs_file(struct page *page, struct list_head *to_kill,
421 struct to_kill **tkc)
422{
423 struct vm_area_struct *vma;
424 struct task_struct *tsk;
425 struct prio_tree_iter iter;
426 struct address_space *mapping = page->mapping;
427
428 mutex_lock(&mapping->i_mmap_mutex);
429 read_lock(&tasklist_lock);
430 for_each_process(tsk) {
431 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
432
433 if (!task_early_kill(tsk))
434 continue;
435
436 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
437 pgoff) {
438 /*
439 * Send early kill signal to tasks where a vma covers
440 * the page but the corrupted page is not necessarily
441 * mapped it in its pte.
442 * Assume applications who requested early kill want
443 * to be informed of all such data corruptions.
444 */
445 if (vma->vm_mm == tsk->mm)
446 add_to_kill(tsk, page, vma, to_kill, tkc);
447 }
448 }
449 read_unlock(&tasklist_lock);
450 mutex_unlock(&mapping->i_mmap_mutex);
451}
452
453/*
454 * Collect the processes who have the corrupted page mapped to kill.
455 * This is done in two steps for locking reasons.
456 * First preallocate one tokill structure outside the spin locks,
457 * so that we can kill at least one process reasonably reliable.
458 */
459static void collect_procs(struct page *page, struct list_head *tokill)
460{
461 struct to_kill *tk;
462
463 if (!page->mapping)
464 return;
465
466 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
467 if (!tk)
468 return;
469 if (PageAnon(page))
470 collect_procs_anon(page, tokill, &tk);
471 else
472 collect_procs_file(page, tokill, &tk);
473 kfree(tk);
474}
475
476/*
477 * Error handlers for various types of pages.
478 */
479
480enum outcome {
481 IGNORED, /* Error: cannot be handled */
482 FAILED, /* Error: handling failed */
483 DELAYED, /* Will be handled later */
484 RECOVERED, /* Successfully recovered */
485};
486
487static const char *action_name[] = {
488 [IGNORED] = "Ignored",
489 [FAILED] = "Failed",
490 [DELAYED] = "Delayed",
491 [RECOVERED] = "Recovered",
492};
493
494/*
495 * XXX: It is possible that a page is isolated from LRU cache,
496 * and then kept in swap cache or failed to remove from page cache.
497 * The page count will stop it from being freed by unpoison.
498 * Stress tests should be aware of this memory leak problem.
499 */
500static int delete_from_lru_cache(struct page *p)
501{
502 if (!isolate_lru_page(p)) {
503 /*
504 * Clear sensible page flags, so that the buddy system won't
505 * complain when the page is unpoison-and-freed.
506 */
507 ClearPageActive(p);
508 ClearPageUnevictable(p);
509 /*
510 * drop the page count elevated by isolate_lru_page()
511 */
512 page_cache_release(p);
513 return 0;
514 }
515 return -EIO;
516}
517
518/*
519 * Error hit kernel page.
520 * Do nothing, try to be lucky and not touch this instead. For a few cases we
521 * could be more sophisticated.
522 */
523static int me_kernel(struct page *p, unsigned long pfn)
524{
525 return IGNORED;
526}
527
528/*
529 * Page in unknown state. Do nothing.
530 */
531static int me_unknown(struct page *p, unsigned long pfn)
532{
533 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
534 return FAILED;
535}
536
537/*
538 * Clean (or cleaned) page cache page.
539 */
540static int me_pagecache_clean(struct page *p, unsigned long pfn)
541{
542 int err;
543 int ret = FAILED;
544 struct address_space *mapping;
545
546 delete_from_lru_cache(p);
547
548 /*
549 * For anonymous pages we're done the only reference left
550 * should be the one m_f() holds.
551 */
552 if (PageAnon(p))
553 return RECOVERED;
554
555 /*
556 * Now truncate the page in the page cache. This is really
557 * more like a "temporary hole punch"
558 * Don't do this for block devices when someone else
559 * has a reference, because it could be file system metadata
560 * and that's not safe to truncate.
561 */
562 mapping = page_mapping(p);
563 if (!mapping) {
564 /*
565 * Page has been teared down in the meanwhile
566 */
567 return FAILED;
568 }
569
570 /*
571 * Truncation is a bit tricky. Enable it per file system for now.
572 *
573 * Open: to take i_mutex or not for this? Right now we don't.
574 */
575 if (mapping->a_ops->error_remove_page) {
576 err = mapping->a_ops->error_remove_page(mapping, p);
577 if (err != 0) {
578 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
579 pfn, err);
580 } else if (page_has_private(p) &&
581 !try_to_release_page(p, GFP_NOIO)) {
582 pr_info("MCE %#lx: failed to release buffers\n", pfn);
583 } else {
584 ret = RECOVERED;
585 }
586 } else {
587 /*
588 * If the file system doesn't support it just invalidate
589 * This fails on dirty or anything with private pages
590 */
591 if (invalidate_inode_page(p))
592 ret = RECOVERED;
593 else
594 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
595 pfn);
596 }
597 return ret;
598}
599
600/*
601 * Dirty cache page page
602 * Issues: when the error hit a hole page the error is not properly
603 * propagated.
604 */
605static int me_pagecache_dirty(struct page *p, unsigned long pfn)
606{
607 struct address_space *mapping = page_mapping(p);
608
609 SetPageError(p);
610 /* TBD: print more information about the file. */
611 if (mapping) {
612 /*
613 * IO error will be reported by write(), fsync(), etc.
614 * who check the mapping.
615 * This way the application knows that something went
616 * wrong with its dirty file data.
617 *
618 * There's one open issue:
619 *
620 * The EIO will be only reported on the next IO
621 * operation and then cleared through the IO map.
622 * Normally Linux has two mechanisms to pass IO error
623 * first through the AS_EIO flag in the address space
624 * and then through the PageError flag in the page.
625 * Since we drop pages on memory failure handling the
626 * only mechanism open to use is through AS_AIO.
627 *
628 * This has the disadvantage that it gets cleared on
629 * the first operation that returns an error, while
630 * the PageError bit is more sticky and only cleared
631 * when the page is reread or dropped. If an
632 * application assumes it will always get error on
633 * fsync, but does other operations on the fd before
634 * and the page is dropped between then the error
635 * will not be properly reported.
636 *
637 * This can already happen even without hwpoisoned
638 * pages: first on metadata IO errors (which only
639 * report through AS_EIO) or when the page is dropped
640 * at the wrong time.
641 *
642 * So right now we assume that the application DTRT on
643 * the first EIO, but we're not worse than other parts
644 * of the kernel.
645 */
646 mapping_set_error(mapping, EIO);
647 }
648
649 return me_pagecache_clean(p, pfn);
650}
651
652/*
653 * Clean and dirty swap cache.
654 *
655 * Dirty swap cache page is tricky to handle. The page could live both in page
656 * cache and swap cache(ie. page is freshly swapped in). So it could be
657 * referenced concurrently by 2 types of PTEs:
658 * normal PTEs and swap PTEs. We try to handle them consistently by calling
659 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
660 * and then
661 * - clear dirty bit to prevent IO
662 * - remove from LRU
663 * - but keep in the swap cache, so that when we return to it on
664 * a later page fault, we know the application is accessing
665 * corrupted data and shall be killed (we installed simple
666 * interception code in do_swap_page to catch it).
667 *
668 * Clean swap cache pages can be directly isolated. A later page fault will
669 * bring in the known good data from disk.
670 */
671static int me_swapcache_dirty(struct page *p, unsigned long pfn)
672{
673 ClearPageDirty(p);
674 /* Trigger EIO in shmem: */
675 ClearPageUptodate(p);
676
677 if (!delete_from_lru_cache(p))
678 return DELAYED;
679 else
680 return FAILED;
681}
682
683static int me_swapcache_clean(struct page *p, unsigned long pfn)
684{
685 delete_from_swap_cache(p);
686
687 if (!delete_from_lru_cache(p))
688 return RECOVERED;
689 else
690 return FAILED;
691}
692
693/*
694 * Huge pages. Needs work.
695 * Issues:
696 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
697 * To narrow down kill region to one page, we need to break up pmd.
698 */
699static int me_huge_page(struct page *p, unsigned long pfn)
700{
701 int res = 0;
702 struct page *hpage = compound_head(p);
703 /*
704 * We can safely recover from error on free or reserved (i.e.
705 * not in-use) hugepage by dequeuing it from freelist.
706 * To check whether a hugepage is in-use or not, we can't use
707 * page->lru because it can be used in other hugepage operations,
708 * such as __unmap_hugepage_range() and gather_surplus_pages().
709 * So instead we use page_mapping() and PageAnon().
710 * We assume that this function is called with page lock held,
711 * so there is no race between isolation and mapping/unmapping.
712 */
713 if (!(page_mapping(hpage) || PageAnon(hpage))) {
714 res = dequeue_hwpoisoned_huge_page(hpage);
715 if (!res)
716 return RECOVERED;
717 }
718 return DELAYED;
719}
720
721/*
722 * Various page states we can handle.
723 *
724 * A page state is defined by its current page->flags bits.
725 * The table matches them in order and calls the right handler.
726 *
727 * This is quite tricky because we can access page at any time
728 * in its live cycle, so all accesses have to be extremely careful.
729 *
730 * This is not complete. More states could be added.
731 * For any missing state don't attempt recovery.
732 */
733
734#define dirty (1UL << PG_dirty)
735#define sc (1UL << PG_swapcache)
736#define unevict (1UL << PG_unevictable)
737#define mlock (1UL << PG_mlocked)
738#define writeback (1UL << PG_writeback)
739#define lru (1UL << PG_lru)
740#define swapbacked (1UL << PG_swapbacked)
741#define head (1UL << PG_head)
742#define tail (1UL << PG_tail)
743#define compound (1UL << PG_compound)
744#define slab (1UL << PG_slab)
745#define reserved (1UL << PG_reserved)
746
747static struct page_state {
748 unsigned long mask;
749 unsigned long res;
750 char *msg;
751 int (*action)(struct page *p, unsigned long pfn);
752} error_states[] = {
753 { reserved, reserved, "reserved kernel", me_kernel },
754 /*
755 * free pages are specially detected outside this table:
756 * PG_buddy pages only make a small fraction of all free pages.
757 */
758
759 /*
760 * Could in theory check if slab page is free or if we can drop
761 * currently unused objects without touching them. But just
762 * treat it as standard kernel for now.
763 */
764 { slab, slab, "kernel slab", me_kernel },
765
766#ifdef CONFIG_PAGEFLAGS_EXTENDED
767 { head, head, "huge", me_huge_page },
768 { tail, tail, "huge", me_huge_page },
769#else
770 { compound, compound, "huge", me_huge_page },
771#endif
772
773 { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
774 { sc|dirty, sc, "swapcache", me_swapcache_clean },
775
776 { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
777 { unevict, unevict, "unevictable LRU", me_pagecache_clean},
778
779 { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
780 { mlock, mlock, "mlocked LRU", me_pagecache_clean },
781
782 { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
783 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
784
785 /*
786 * Catchall entry: must be at end.
787 */
788 { 0, 0, "unknown page state", me_unknown },
789};
790
791#undef dirty
792#undef sc
793#undef unevict
794#undef mlock
795#undef writeback
796#undef lru
797#undef swapbacked
798#undef head
799#undef tail
800#undef compound
801#undef slab
802#undef reserved
803
804static void action_result(unsigned long pfn, char *msg, int result)
805{
806 struct page *page = pfn_to_page(pfn);
807
808 printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
809 pfn,
810 PageDirty(page) ? "dirty " : "",
811 msg, action_name[result]);
812}
813
814static int page_action(struct page_state *ps, struct page *p,
815 unsigned long pfn)
816{
817 int result;
818 int count;
819
820 result = ps->action(p, pfn);
821 action_result(pfn, ps->msg, result);
822
823 count = page_count(p) - 1;
824 if (ps->action == me_swapcache_dirty && result == DELAYED)
825 count--;
826 if (count != 0) {
827 printk(KERN_ERR
828 "MCE %#lx: %s page still referenced by %d users\n",
829 pfn, ps->msg, count);
830 result = FAILED;
831 }
832
833 /* Could do more checks here if page looks ok */
834 /*
835 * Could adjust zone counters here to correct for the missing page.
836 */
837
838 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
839}
840
841/*
842 * Do all that is necessary to remove user space mappings. Unmap
843 * the pages and send SIGBUS to the processes if the data was dirty.
844 */
845static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
846 int trapno)
847{
848 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
849 struct address_space *mapping;
850 LIST_HEAD(tokill);
851 int ret;
852 int kill = 1;
853 struct page *hpage = compound_head(p);
854 struct page *ppage;
855
856 if (PageReserved(p) || PageSlab(p))
857 return SWAP_SUCCESS;
858
859 /*
860 * This check implies we don't kill processes if their pages
861 * are in the swap cache early. Those are always late kills.
862 */
863 if (!page_mapped(hpage))
864 return SWAP_SUCCESS;
865
866 if (PageKsm(p))
867 return SWAP_FAIL;
868
869 if (PageSwapCache(p)) {
870 printk(KERN_ERR
871 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
872 ttu |= TTU_IGNORE_HWPOISON;
873 }
874
875 /*
876 * Propagate the dirty bit from PTEs to struct page first, because we
877 * need this to decide if we should kill or just drop the page.
878 * XXX: the dirty test could be racy: set_page_dirty() may not always
879 * be called inside page lock (it's recommended but not enforced).
880 */
881 mapping = page_mapping(hpage);
882 if (!PageDirty(hpage) && mapping &&
883 mapping_cap_writeback_dirty(mapping)) {
884 if (page_mkclean(hpage)) {
885 SetPageDirty(hpage);
886 } else {
887 kill = 0;
888 ttu |= TTU_IGNORE_HWPOISON;
889 printk(KERN_INFO
890 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
891 pfn);
892 }
893 }
894
895 /*
896 * ppage: poisoned page
897 * if p is regular page(4k page)
898 * ppage == real poisoned page;
899 * else p is hugetlb or THP, ppage == head page.
900 */
901 ppage = hpage;
902
903 if (PageTransHuge(hpage)) {
904 /*
905 * Verify that this isn't a hugetlbfs head page, the check for
906 * PageAnon is just for avoid tripping a split_huge_page
907 * internal debug check, as split_huge_page refuses to deal with
908 * anything that isn't an anon page. PageAnon can't go away fro
909 * under us because we hold a refcount on the hpage, without a
910 * refcount on the hpage. split_huge_page can't be safely called
911 * in the first place, having a refcount on the tail isn't
912 * enough * to be safe.
913 */
914 if (!PageHuge(hpage) && PageAnon(hpage)) {
915 if (unlikely(split_huge_page(hpage))) {
916 /*
917 * FIXME: if splitting THP is failed, it is
918 * better to stop the following operation rather
919 * than causing panic by unmapping. System might
920 * survive if the page is freed later.
921 */
922 printk(KERN_INFO
923 "MCE %#lx: failed to split THP\n", pfn);
924
925 BUG_ON(!PageHWPoison(p));
926 return SWAP_FAIL;
927 }
928 /* THP is split, so ppage should be the real poisoned page. */
929 ppage = p;
930 }
931 }
932
933 /*
934 * First collect all the processes that have the page
935 * mapped in dirty form. This has to be done before try_to_unmap,
936 * because ttu takes the rmap data structures down.
937 *
938 * Error handling: We ignore errors here because
939 * there's nothing that can be done.
940 */
941 if (kill)
942 collect_procs(ppage, &tokill);
943
944 if (hpage != ppage)
945 lock_page(ppage);
946
947 ret = try_to_unmap(ppage, ttu);
948 if (ret != SWAP_SUCCESS)
949 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
950 pfn, page_mapcount(ppage));
951
952 if (hpage != ppage)
953 unlock_page(ppage);
954
955 /*
956 * Now that the dirty bit has been propagated to the
957 * struct page and all unmaps done we can decide if
958 * killing is needed or not. Only kill when the page
959 * was dirty, otherwise the tokill list is merely
960 * freed. When there was a problem unmapping earlier
961 * use a more force-full uncatchable kill to prevent
962 * any accesses to the poisoned memory.
963 */
964 kill_procs_ao(&tokill, !!PageDirty(ppage), trapno,
965 ret != SWAP_SUCCESS, p, pfn);
966
967 return ret;
968}
969
970static void set_page_hwpoison_huge_page(struct page *hpage)
971{
972 int i;
973 int nr_pages = 1 << compound_trans_order(hpage);
974 for (i = 0; i < nr_pages; i++)
975 SetPageHWPoison(hpage + i);
976}
977
978static void clear_page_hwpoison_huge_page(struct page *hpage)
979{
980 int i;
981 int nr_pages = 1 << compound_trans_order(hpage);
982 for (i = 0; i < nr_pages; i++)
983 ClearPageHWPoison(hpage + i);
984}
985
986int __memory_failure(unsigned long pfn, int trapno, int flags)
987{
988 struct page_state *ps;
989 struct page *p;
990 struct page *hpage;
991 int res;
992 unsigned int nr_pages;
993
994 if (!sysctl_memory_failure_recovery)
995 panic("Memory failure from trap %d on page %lx", trapno, pfn);
996
997 if (!pfn_valid(pfn)) {
998 printk(KERN_ERR
999 "MCE %#lx: memory outside kernel control\n",
1000 pfn);
1001 return -ENXIO;
1002 }
1003
1004 p = pfn_to_page(pfn);
1005 hpage = compound_head(p);
1006 if (TestSetPageHWPoison(p)) {
1007 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
1008 return 0;
1009 }
1010
1011 nr_pages = 1 << compound_trans_order(hpage);
1012 atomic_long_add(nr_pages, &mce_bad_pages);
1013
1014 /*
1015 * We need/can do nothing about count=0 pages.
1016 * 1) it's a free page, and therefore in safe hand:
1017 * prep_new_page() will be the gate keeper.
1018 * 2) it's a free hugepage, which is also safe:
1019 * an affected hugepage will be dequeued from hugepage freelist,
1020 * so there's no concern about reusing it ever after.
1021 * 3) it's part of a non-compound high order page.
1022 * Implies some kernel user: cannot stop them from
1023 * R/W the page; let's pray that the page has been
1024 * used and will be freed some time later.
1025 * In fact it's dangerous to directly bump up page count from 0,
1026 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1027 */
1028 if (!(flags & MF_COUNT_INCREASED) &&
1029 !get_page_unless_zero(hpage)) {
1030 if (is_free_buddy_page(p)) {
1031 action_result(pfn, "free buddy", DELAYED);
1032 return 0;
1033 } else if (PageHuge(hpage)) {
1034 /*
1035 * Check "just unpoisoned", "filter hit", and
1036 * "race with other subpage."
1037 */
1038 lock_page(hpage);
1039 if (!PageHWPoison(hpage)
1040 || (hwpoison_filter(p) && TestClearPageHWPoison(p))
1041 || (p != hpage && TestSetPageHWPoison(hpage))) {
1042 atomic_long_sub(nr_pages, &mce_bad_pages);
1043 return 0;
1044 }
1045 set_page_hwpoison_huge_page(hpage);
1046 res = dequeue_hwpoisoned_huge_page(hpage);
1047 action_result(pfn, "free huge",
1048 res ? IGNORED : DELAYED);
1049 unlock_page(hpage);
1050 return res;
1051 } else {
1052 action_result(pfn, "high order kernel", IGNORED);
1053 return -EBUSY;
1054 }
1055 }
1056
1057 /*
1058 * We ignore non-LRU pages for good reasons.
1059 * - PG_locked is only well defined for LRU pages and a few others
1060 * - to avoid races with __set_page_locked()
1061 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1062 * The check (unnecessarily) ignores LRU pages being isolated and
1063 * walked by the page reclaim code, however that's not a big loss.
1064 */
1065 if (!PageHuge(p) && !PageTransCompound(p)) {
1066 if (!PageLRU(p))
1067 shake_page(p, 0);
1068 if (!PageLRU(p)) {
1069 /*
1070 * shake_page could have turned it free.
1071 */
1072 if (is_free_buddy_page(p)) {
1073 action_result(pfn, "free buddy, 2nd try",
1074 DELAYED);
1075 return 0;
1076 }
1077 action_result(pfn, "non LRU", IGNORED);
1078 put_page(p);
1079 return -EBUSY;
1080 }
1081 }
1082
1083 /*
1084 * Lock the page and wait for writeback to finish.
1085 * It's very difficult to mess with pages currently under IO
1086 * and in many cases impossible, so we just avoid it here.
1087 */
1088 lock_page(hpage);
1089
1090 /*
1091 * unpoison always clear PG_hwpoison inside page lock
1092 */
1093 if (!PageHWPoison(p)) {
1094 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1095 res = 0;
1096 goto out;
1097 }
1098 if (hwpoison_filter(p)) {
1099 if (TestClearPageHWPoison(p))
1100 atomic_long_sub(nr_pages, &mce_bad_pages);
1101 unlock_page(hpage);
1102 put_page(hpage);
1103 return 0;
1104 }
1105
1106 /*
1107 * For error on the tail page, we should set PG_hwpoison
1108 * on the head page to show that the hugepage is hwpoisoned
1109 */
1110 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1111 action_result(pfn, "hugepage already hardware poisoned",
1112 IGNORED);
1113 unlock_page(hpage);
1114 put_page(hpage);
1115 return 0;
1116 }
1117 /*
1118 * Set PG_hwpoison on all pages in an error hugepage,
1119 * because containment is done in hugepage unit for now.
1120 * Since we have done TestSetPageHWPoison() for the head page with
1121 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1122 */
1123 if (PageHuge(p))
1124 set_page_hwpoison_huge_page(hpage);
1125
1126 wait_on_page_writeback(p);
1127
1128 /*
1129 * Now take care of user space mappings.
1130 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1131 */
1132 if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
1133 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1134 res = -EBUSY;
1135 goto out;
1136 }
1137
1138 /*
1139 * Torn down by someone else?
1140 */
1141 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1142 action_result(pfn, "already truncated LRU", IGNORED);
1143 res = -EBUSY;
1144 goto out;
1145 }
1146
1147 res = -EBUSY;
1148 for (ps = error_states;; ps++) {
1149 if ((p->flags & ps->mask) == ps->res) {
1150 res = page_action(ps, p, pfn);
1151 break;
1152 }
1153 }
1154out:
1155 unlock_page(hpage);
1156 return res;
1157}
1158EXPORT_SYMBOL_GPL(__memory_failure);
1159
1160/**
1161 * memory_failure - Handle memory failure of a page.
1162 * @pfn: Page Number of the corrupted page
1163 * @trapno: Trap number reported in the signal to user space.
1164 *
1165 * This function is called by the low level machine check code
1166 * of an architecture when it detects hardware memory corruption
1167 * of a page. It tries its best to recover, which includes
1168 * dropping pages, killing processes etc.
1169 *
1170 * The function is primarily of use for corruptions that
1171 * happen outside the current execution context (e.g. when
1172 * detected by a background scrubber)
1173 *
1174 * Must run in process context (e.g. a work queue) with interrupts
1175 * enabled and no spinlocks hold.
1176 */
1177void memory_failure(unsigned long pfn, int trapno)
1178{
1179 __memory_failure(pfn, trapno, 0);
1180}
1181
1182#define MEMORY_FAILURE_FIFO_ORDER 4
1183#define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1184
1185struct memory_failure_entry {
1186 unsigned long pfn;
1187 int trapno;
1188 int flags;
1189};
1190
1191struct memory_failure_cpu {
1192 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1193 MEMORY_FAILURE_FIFO_SIZE);
1194 spinlock_t lock;
1195 struct work_struct work;
1196};
1197
1198static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1199
1200/**
1201 * memory_failure_queue - Schedule handling memory failure of a page.
1202 * @pfn: Page Number of the corrupted page
1203 * @trapno: Trap number reported in the signal to user space.
1204 * @flags: Flags for memory failure handling
1205 *
1206 * This function is called by the low level hardware error handler
1207 * when it detects hardware memory corruption of a page. It schedules
1208 * the recovering of error page, including dropping pages, killing
1209 * processes etc.
1210 *
1211 * The function is primarily of use for corruptions that
1212 * happen outside the current execution context (e.g. when
1213 * detected by a background scrubber)
1214 *
1215 * Can run in IRQ context.
1216 */
1217void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1218{
1219 struct memory_failure_cpu *mf_cpu;
1220 unsigned long proc_flags;
1221 struct memory_failure_entry entry = {
1222 .pfn = pfn,
1223 .trapno = trapno,
1224 .flags = flags,
1225 };
1226
1227 mf_cpu = &get_cpu_var(memory_failure_cpu);
1228 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1229 if (kfifo_put(&mf_cpu->fifo, &entry))
1230 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1231 else
1232 pr_err("Memory failure: buffer overflow when queuing memory failure at 0x%#lx\n",
1233 pfn);
1234 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1235 put_cpu_var(memory_failure_cpu);
1236}
1237EXPORT_SYMBOL_GPL(memory_failure_queue);
1238
1239static void memory_failure_work_func(struct work_struct *work)
1240{
1241 struct memory_failure_cpu *mf_cpu;
1242 struct memory_failure_entry entry = { 0, };
1243 unsigned long proc_flags;
1244 int gotten;
1245
1246 mf_cpu = &__get_cpu_var(memory_failure_cpu);
1247 for (;;) {
1248 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1249 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1250 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1251 if (!gotten)
1252 break;
1253 __memory_failure(entry.pfn, entry.trapno, entry.flags);
1254 }
1255}
1256
1257static int __init memory_failure_init(void)
1258{
1259 struct memory_failure_cpu *mf_cpu;
1260 int cpu;
1261
1262 for_each_possible_cpu(cpu) {
1263 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1264 spin_lock_init(&mf_cpu->lock);
1265 INIT_KFIFO(mf_cpu->fifo);
1266 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1267 }
1268
1269 return 0;
1270}
1271core_initcall(memory_failure_init);
1272
1273/**
1274 * unpoison_memory - Unpoison a previously poisoned page
1275 * @pfn: Page number of the to be unpoisoned page
1276 *
1277 * Software-unpoison a page that has been poisoned by
1278 * memory_failure() earlier.
1279 *
1280 * This is only done on the software-level, so it only works
1281 * for linux injected failures, not real hardware failures
1282 *
1283 * Returns 0 for success, otherwise -errno.
1284 */
1285int unpoison_memory(unsigned long pfn)
1286{
1287 struct page *page;
1288 struct page *p;
1289 int freeit = 0;
1290 unsigned int nr_pages;
1291
1292 if (!pfn_valid(pfn))
1293 return -ENXIO;
1294
1295 p = pfn_to_page(pfn);
1296 page = compound_head(p);
1297
1298 if (!PageHWPoison(p)) {
1299 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
1300 return 0;
1301 }
1302
1303 nr_pages = 1 << compound_trans_order(page);
1304
1305 if (!get_page_unless_zero(page)) {
1306 /*
1307 * Since HWPoisoned hugepage should have non-zero refcount,
1308 * race between memory failure and unpoison seems to happen.
1309 * In such case unpoison fails and memory failure runs
1310 * to the end.
1311 */
1312 if (PageHuge(page)) {
1313 pr_debug("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1314 return 0;
1315 }
1316 if (TestClearPageHWPoison(p))
1317 atomic_long_sub(nr_pages, &mce_bad_pages);
1318 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
1319 return 0;
1320 }
1321
1322 lock_page(page);
1323 /*
1324 * This test is racy because PG_hwpoison is set outside of page lock.
1325 * That's acceptable because that won't trigger kernel panic. Instead,
1326 * the PG_hwpoison page will be caught and isolated on the entrance to
1327 * the free buddy page pool.
1328 */
1329 if (TestClearPageHWPoison(page)) {
1330 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
1331 atomic_long_sub(nr_pages, &mce_bad_pages);
1332 freeit = 1;
1333 if (PageHuge(page))
1334 clear_page_hwpoison_huge_page(page);
1335 }
1336 unlock_page(page);
1337
1338 put_page(page);
1339 if (freeit)
1340 put_page(page);
1341
1342 return 0;
1343}
1344EXPORT_SYMBOL(unpoison_memory);
1345
1346static struct page *new_page(struct page *p, unsigned long private, int **x)
1347{
1348 int nid = page_to_nid(p);
1349 if (PageHuge(p))
1350 return alloc_huge_page_node(page_hstate(compound_head(p)),
1351 nid);
1352 else
1353 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1354}
1355
1356/*
1357 * Safely get reference count of an arbitrary page.
1358 * Returns 0 for a free page, -EIO for a zero refcount page
1359 * that is not free, and 1 for any other page type.
1360 * For 1 the page is returned with increased page count, otherwise not.
1361 */
1362static int get_any_page(struct page *p, unsigned long pfn, int flags)
1363{
1364 int ret;
1365
1366 if (flags & MF_COUNT_INCREASED)
1367 return 1;
1368
1369 /*
1370 * The lock_memory_hotplug prevents a race with memory hotplug.
1371 * This is a big hammer, a better would be nicer.
1372 */
1373 lock_memory_hotplug();
1374
1375 /*
1376 * Isolate the page, so that it doesn't get reallocated if it
1377 * was free.
1378 */
1379 set_migratetype_isolate(p);
1380 /*
1381 * When the target page is a free hugepage, just remove it
1382 * from free hugepage list.
1383 */
1384 if (!get_page_unless_zero(compound_head(p))) {
1385 if (PageHuge(p)) {
1386 pr_info("get_any_page: %#lx free huge page\n", pfn);
1387 ret = dequeue_hwpoisoned_huge_page(compound_head(p));
1388 } else if (is_free_buddy_page(p)) {
1389 pr_info("get_any_page: %#lx free buddy page\n", pfn);
1390 /* Set hwpoison bit while page is still isolated */
1391 SetPageHWPoison(p);
1392 ret = 0;
1393 } else {
1394 pr_info("get_any_page: %#lx: unknown zero refcount page type %lx\n",
1395 pfn, p->flags);
1396 ret = -EIO;
1397 }
1398 } else {
1399 /* Not a free page */
1400 ret = 1;
1401 }
1402 unset_migratetype_isolate(p);
1403 unlock_memory_hotplug();
1404 return ret;
1405}
1406
1407static int soft_offline_huge_page(struct page *page, int flags)
1408{
1409 int ret;
1410 unsigned long pfn = page_to_pfn(page);
1411 struct page *hpage = compound_head(page);
1412 LIST_HEAD(pagelist);
1413
1414 ret = get_any_page(page, pfn, flags);
1415 if (ret < 0)
1416 return ret;
1417 if (ret == 0)
1418 goto done;
1419
1420 if (PageHWPoison(hpage)) {
1421 put_page(hpage);
1422 pr_debug("soft offline: %#lx hugepage already poisoned\n", pfn);
1423 return -EBUSY;
1424 }
1425
1426 /* Keep page count to indicate a given hugepage is isolated. */
1427
1428 list_add(&hpage->lru, &pagelist);
1429 ret = migrate_huge_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0,
1430 true);
1431 if (ret) {
1432 struct page *page1, *page2;
1433 list_for_each_entry_safe(page1, page2, &pagelist, lru)
1434 put_page(page1);
1435
1436 pr_debug("soft offline: %#lx: migration failed %d, type %lx\n",
1437 pfn, ret, page->flags);
1438 if (ret > 0)
1439 ret = -EIO;
1440 return ret;
1441 }
1442done:
1443 if (!PageHWPoison(hpage))
1444 atomic_long_add(1 << compound_trans_order(hpage), &mce_bad_pages);
1445 set_page_hwpoison_huge_page(hpage);
1446 dequeue_hwpoisoned_huge_page(hpage);
1447 /* keep elevated page count for bad page */
1448 return ret;
1449}
1450
1451/**
1452 * soft_offline_page - Soft offline a page.
1453 * @page: page to offline
1454 * @flags: flags. Same as memory_failure().
1455 *
1456 * Returns 0 on success, otherwise negated errno.
1457 *
1458 * Soft offline a page, by migration or invalidation,
1459 * without killing anything. This is for the case when
1460 * a page is not corrupted yet (so it's still valid to access),
1461 * but has had a number of corrected errors and is better taken
1462 * out.
1463 *
1464 * The actual policy on when to do that is maintained by
1465 * user space.
1466 *
1467 * This should never impact any application or cause data loss,
1468 * however it might take some time.
1469 *
1470 * This is not a 100% solution for all memory, but tries to be
1471 * ``good enough'' for the majority of memory.
1472 */
1473int soft_offline_page(struct page *page, int flags)
1474{
1475 int ret;
1476 unsigned long pfn = page_to_pfn(page);
1477
1478 if (PageHuge(page))
1479 return soft_offline_huge_page(page, flags);
1480
1481 ret = get_any_page(page, pfn, flags);
1482 if (ret < 0)
1483 return ret;
1484 if (ret == 0)
1485 goto done;
1486
1487 /*
1488 * Page cache page we can handle?
1489 */
1490 if (!PageLRU(page)) {
1491 /*
1492 * Try to free it.
1493 */
1494 put_page(page);
1495 shake_page(page, 1);
1496
1497 /*
1498 * Did it turn free?
1499 */
1500 ret = get_any_page(page, pfn, 0);
1501 if (ret < 0)
1502 return ret;
1503 if (ret == 0)
1504 goto done;
1505 }
1506 if (!PageLRU(page)) {
1507 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1508 pfn, page->flags);
1509 return -EIO;
1510 }
1511
1512 lock_page(page);
1513 wait_on_page_writeback(page);
1514
1515 /*
1516 * Synchronized using the page lock with memory_failure()
1517 */
1518 if (PageHWPoison(page)) {
1519 unlock_page(page);
1520 put_page(page);
1521 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1522 return -EBUSY;
1523 }
1524
1525 /*
1526 * Try to invalidate first. This should work for
1527 * non dirty unmapped page cache pages.
1528 */
1529 ret = invalidate_inode_page(page);
1530 unlock_page(page);
1531 /*
1532 * RED-PEN would be better to keep it isolated here, but we
1533 * would need to fix isolation locking first.
1534 */
1535 if (ret == 1) {
1536 put_page(page);
1537 ret = 0;
1538 pr_info("soft_offline: %#lx: invalidated\n", pfn);
1539 goto done;
1540 }
1541
1542 /*
1543 * Simple invalidation didn't work.
1544 * Try to migrate to a new page instead. migrate.c
1545 * handles a large number of cases for us.
1546 */
1547 ret = isolate_lru_page(page);
1548 /*
1549 * Drop page reference which is came from get_any_page()
1550 * successful isolate_lru_page() already took another one.
1551 */
1552 put_page(page);
1553 if (!ret) {
1554 LIST_HEAD(pagelist);
1555 inc_zone_page_state(page, NR_ISOLATED_ANON +
1556 page_is_file_cache(page));
1557 list_add(&page->lru, &pagelist);
1558 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1559 0, true);
1560 if (ret) {
1561 putback_lru_pages(&pagelist);
1562 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1563 pfn, ret, page->flags);
1564 if (ret > 0)
1565 ret = -EIO;
1566 }
1567 } else {
1568 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1569 pfn, ret, page_count(page), page->flags);
1570 }
1571 if (ret)
1572 return ret;
1573
1574done:
1575 atomic_long_add(1, &mce_bad_pages);
1576 SetPageHWPoison(page);
1577 /* keep elevated page count for bad page */
1578 return ret;
1579}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2008, 2009 Intel Corporation
4 * Authors: Andi Kleen, Fengguang Wu
5 *
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
8 * failure.
9 *
10 * In addition there is a "soft offline" entry point that allows stop using
11 * not-yet-corrupted-by-suspicious pages without killing anything.
12 *
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.
20 *
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.
28 *
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
34 * VM.
35 */
36
37#define pr_fmt(fmt) "Memory failure: " fmt
38
39#include <linux/kernel.h>
40#include <linux/mm.h>
41#include <linux/page-flags.h>
42#include <linux/kernel-page-flags.h>
43#include <linux/sched/signal.h>
44#include <linux/sched/task.h>
45#include <linux/dax.h>
46#include <linux/ksm.h>
47#include <linux/rmap.h>
48#include <linux/export.h>
49#include <linux/pagemap.h>
50#include <linux/swap.h>
51#include <linux/backing-dev.h>
52#include <linux/migrate.h>
53#include <linux/suspend.h>
54#include <linux/slab.h>
55#include <linux/swapops.h>
56#include <linux/hugetlb.h>
57#include <linux/memory_hotplug.h>
58#include <linux/mm_inline.h>
59#include <linux/memremap.h>
60#include <linux/kfifo.h>
61#include <linux/ratelimit.h>
62#include <linux/page-isolation.h>
63#include <linux/pagewalk.h>
64#include <linux/shmem_fs.h>
65#include "swap.h"
66#include "internal.h"
67#include "ras/ras_event.h"
68
69int sysctl_memory_failure_early_kill __read_mostly = 0;
70
71int sysctl_memory_failure_recovery __read_mostly = 1;
72
73atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
74
75static bool hw_memory_failure __read_mostly = false;
76
77inline void num_poisoned_pages_inc(unsigned long pfn)
78{
79 atomic_long_inc(&num_poisoned_pages);
80 memblk_nr_poison_inc(pfn);
81}
82
83inline void num_poisoned_pages_sub(unsigned long pfn, long i)
84{
85 atomic_long_sub(i, &num_poisoned_pages);
86 if (pfn != -1UL)
87 memblk_nr_poison_sub(pfn, i);
88}
89
90/*
91 * Return values:
92 * 1: the page is dissolved (if needed) and taken off from buddy,
93 * 0: the page is dissolved (if needed) and not taken off from buddy,
94 * < 0: failed to dissolve.
95 */
96static int __page_handle_poison(struct page *page)
97{
98 int ret;
99
100 zone_pcp_disable(page_zone(page));
101 ret = dissolve_free_huge_page(page);
102 if (!ret)
103 ret = take_page_off_buddy(page);
104 zone_pcp_enable(page_zone(page));
105
106 return ret;
107}
108
109static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release)
110{
111 if (hugepage_or_freepage) {
112 /*
113 * Doing this check for free pages is also fine since dissolve_free_huge_page
114 * returns 0 for non-hugetlb pages as well.
115 */
116 if (__page_handle_poison(page) <= 0)
117 /*
118 * We could fail to take off the target page from buddy
119 * for example due to racy page allocation, but that's
120 * acceptable because soft-offlined page is not broken
121 * and if someone really want to use it, they should
122 * take it.
123 */
124 return false;
125 }
126
127 SetPageHWPoison(page);
128 if (release)
129 put_page(page);
130 page_ref_inc(page);
131 num_poisoned_pages_inc(page_to_pfn(page));
132
133 return true;
134}
135
136#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
137
138u32 hwpoison_filter_enable = 0;
139u32 hwpoison_filter_dev_major = ~0U;
140u32 hwpoison_filter_dev_minor = ~0U;
141u64 hwpoison_filter_flags_mask;
142u64 hwpoison_filter_flags_value;
143EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
144EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
145EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
146EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
147EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
148
149static int hwpoison_filter_dev(struct page *p)
150{
151 struct address_space *mapping;
152 dev_t dev;
153
154 if (hwpoison_filter_dev_major == ~0U &&
155 hwpoison_filter_dev_minor == ~0U)
156 return 0;
157
158 mapping = page_mapping(p);
159 if (mapping == NULL || mapping->host == NULL)
160 return -EINVAL;
161
162 dev = mapping->host->i_sb->s_dev;
163 if (hwpoison_filter_dev_major != ~0U &&
164 hwpoison_filter_dev_major != MAJOR(dev))
165 return -EINVAL;
166 if (hwpoison_filter_dev_minor != ~0U &&
167 hwpoison_filter_dev_minor != MINOR(dev))
168 return -EINVAL;
169
170 return 0;
171}
172
173static int hwpoison_filter_flags(struct page *p)
174{
175 if (!hwpoison_filter_flags_mask)
176 return 0;
177
178 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
179 hwpoison_filter_flags_value)
180 return 0;
181 else
182 return -EINVAL;
183}
184
185/*
186 * This allows stress tests to limit test scope to a collection of tasks
187 * by putting them under some memcg. This prevents killing unrelated/important
188 * processes such as /sbin/init. Note that the target task may share clean
189 * pages with init (eg. libc text), which is harmless. If the target task
190 * share _dirty_ pages with another task B, the test scheme must make sure B
191 * is also included in the memcg. At last, due to race conditions this filter
192 * can only guarantee that the page either belongs to the memcg tasks, or is
193 * a freed page.
194 */
195#ifdef CONFIG_MEMCG
196u64 hwpoison_filter_memcg;
197EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
198static int hwpoison_filter_task(struct page *p)
199{
200 if (!hwpoison_filter_memcg)
201 return 0;
202
203 if (page_cgroup_ino(p) != hwpoison_filter_memcg)
204 return -EINVAL;
205
206 return 0;
207}
208#else
209static int hwpoison_filter_task(struct page *p) { return 0; }
210#endif
211
212int hwpoison_filter(struct page *p)
213{
214 if (!hwpoison_filter_enable)
215 return 0;
216
217 if (hwpoison_filter_dev(p))
218 return -EINVAL;
219
220 if (hwpoison_filter_flags(p))
221 return -EINVAL;
222
223 if (hwpoison_filter_task(p))
224 return -EINVAL;
225
226 return 0;
227}
228#else
229int hwpoison_filter(struct page *p)
230{
231 return 0;
232}
233#endif
234
235EXPORT_SYMBOL_GPL(hwpoison_filter);
236
237/*
238 * Kill all processes that have a poisoned page mapped and then isolate
239 * the page.
240 *
241 * General strategy:
242 * Find all processes having the page mapped and kill them.
243 * But we keep a page reference around so that the page is not
244 * actually freed yet.
245 * Then stash the page away
246 *
247 * There's no convenient way to get back to mapped processes
248 * from the VMAs. So do a brute-force search over all
249 * running processes.
250 *
251 * Remember that machine checks are not common (or rather
252 * if they are common you have other problems), so this shouldn't
253 * be a performance issue.
254 *
255 * Also there are some races possible while we get from the
256 * error detection to actually handle it.
257 */
258
259struct to_kill {
260 struct list_head nd;
261 struct task_struct *tsk;
262 unsigned long addr;
263 short size_shift;
264};
265
266/*
267 * Send all the processes who have the page mapped a signal.
268 * ``action optional'' if they are not immediately affected by the error
269 * ``action required'' if error happened in current execution context
270 */
271static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
272{
273 struct task_struct *t = tk->tsk;
274 short addr_lsb = tk->size_shift;
275 int ret = 0;
276
277 pr_err("%#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
278 pfn, t->comm, t->pid);
279
280 if ((flags & MF_ACTION_REQUIRED) && (t == current))
281 ret = force_sig_mceerr(BUS_MCEERR_AR,
282 (void __user *)tk->addr, addr_lsb);
283 else
284 /*
285 * Signal other processes sharing the page if they have
286 * PF_MCE_EARLY set.
287 * Don't use force here, it's convenient if the signal
288 * can be temporarily blocked.
289 * This could cause a loop when the user sets SIGBUS
290 * to SIG_IGN, but hopefully no one will do that?
291 */
292 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
293 addr_lsb, t);
294 if (ret < 0)
295 pr_info("Error sending signal to %s:%d: %d\n",
296 t->comm, t->pid, ret);
297 return ret;
298}
299
300/*
301 * Unknown page type encountered. Try to check whether it can turn PageLRU by
302 * lru_add_drain_all.
303 */
304void shake_page(struct page *p)
305{
306 if (PageHuge(p))
307 return;
308
309 if (!PageSlab(p)) {
310 lru_add_drain_all();
311 if (PageLRU(p) || is_free_buddy_page(p))
312 return;
313 }
314
315 /*
316 * TODO: Could shrink slab caches here if a lightweight range-based
317 * shrinker will be available.
318 */
319}
320EXPORT_SYMBOL_GPL(shake_page);
321
322static unsigned long dev_pagemap_mapping_shift(struct vm_area_struct *vma,
323 unsigned long address)
324{
325 unsigned long ret = 0;
326 pgd_t *pgd;
327 p4d_t *p4d;
328 pud_t *pud;
329 pmd_t *pmd;
330 pte_t *pte;
331
332 VM_BUG_ON_VMA(address == -EFAULT, vma);
333 pgd = pgd_offset(vma->vm_mm, address);
334 if (!pgd_present(*pgd))
335 return 0;
336 p4d = p4d_offset(pgd, address);
337 if (!p4d_present(*p4d))
338 return 0;
339 pud = pud_offset(p4d, address);
340 if (!pud_present(*pud))
341 return 0;
342 if (pud_devmap(*pud))
343 return PUD_SHIFT;
344 pmd = pmd_offset(pud, address);
345 if (!pmd_present(*pmd))
346 return 0;
347 if (pmd_devmap(*pmd))
348 return PMD_SHIFT;
349 pte = pte_offset_map(pmd, address);
350 if (pte_present(*pte) && pte_devmap(*pte))
351 ret = PAGE_SHIFT;
352 pte_unmap(pte);
353 return ret;
354}
355
356/*
357 * Failure handling: if we can't find or can't kill a process there's
358 * not much we can do. We just print a message and ignore otherwise.
359 */
360
361#define FSDAX_INVALID_PGOFF ULONG_MAX
362
363/*
364 * Schedule a process for later kill.
365 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
366 *
367 * Note: @fsdax_pgoff is used only when @p is a fsdax page and a
368 * filesystem with a memory failure handler has claimed the
369 * memory_failure event. In all other cases, page->index and
370 * page->mapping are sufficient for mapping the page back to its
371 * corresponding user virtual address.
372 */
373static void add_to_kill(struct task_struct *tsk, struct page *p,
374 pgoff_t fsdax_pgoff, struct vm_area_struct *vma,
375 struct list_head *to_kill)
376{
377 struct to_kill *tk;
378
379 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
380 if (!tk) {
381 pr_err("Out of memory while machine check handling\n");
382 return;
383 }
384
385 tk->addr = page_address_in_vma(p, vma);
386 if (is_zone_device_page(p)) {
387 if (fsdax_pgoff != FSDAX_INVALID_PGOFF)
388 tk->addr = vma_pgoff_address(fsdax_pgoff, 1, vma);
389 tk->size_shift = dev_pagemap_mapping_shift(vma, tk->addr);
390 } else
391 tk->size_shift = page_shift(compound_head(p));
392
393 /*
394 * Send SIGKILL if "tk->addr == -EFAULT". Also, as
395 * "tk->size_shift" is always non-zero for !is_zone_device_page(),
396 * so "tk->size_shift == 0" effectively checks no mapping on
397 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times
398 * to a process' address space, it's possible not all N VMAs
399 * contain mappings for the page, but at least one VMA does.
400 * Only deliver SIGBUS with payload derived from the VMA that
401 * has a mapping for the page.
402 */
403 if (tk->addr == -EFAULT) {
404 pr_info("Unable to find user space address %lx in %s\n",
405 page_to_pfn(p), tsk->comm);
406 } else if (tk->size_shift == 0) {
407 kfree(tk);
408 return;
409 }
410
411 get_task_struct(tsk);
412 tk->tsk = tsk;
413 list_add_tail(&tk->nd, to_kill);
414}
415
416/*
417 * Kill the processes that have been collected earlier.
418 *
419 * Only do anything when FORCEKILL is set, otherwise just free the
420 * list (this is used for clean pages which do not need killing)
421 * Also when FAIL is set do a force kill because something went
422 * wrong earlier.
423 */
424static void kill_procs(struct list_head *to_kill, int forcekill, bool fail,
425 unsigned long pfn, int flags)
426{
427 struct to_kill *tk, *next;
428
429 list_for_each_entry_safe(tk, next, to_kill, nd) {
430 if (forcekill) {
431 /*
432 * In case something went wrong with munmapping
433 * make sure the process doesn't catch the
434 * signal and then access the memory. Just kill it.
435 */
436 if (fail || tk->addr == -EFAULT) {
437 pr_err("%#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
438 pfn, tk->tsk->comm, tk->tsk->pid);
439 do_send_sig_info(SIGKILL, SEND_SIG_PRIV,
440 tk->tsk, PIDTYPE_PID);
441 }
442
443 /*
444 * In theory the process could have mapped
445 * something else on the address in-between. We could
446 * check for that, but we need to tell the
447 * process anyways.
448 */
449 else if (kill_proc(tk, pfn, flags) < 0)
450 pr_err("%#lx: Cannot send advisory machine check signal to %s:%d\n",
451 pfn, tk->tsk->comm, tk->tsk->pid);
452 }
453 list_del(&tk->nd);
454 put_task_struct(tk->tsk);
455 kfree(tk);
456 }
457}
458
459/*
460 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
461 * on behalf of the thread group. Return task_struct of the (first found)
462 * dedicated thread if found, and return NULL otherwise.
463 *
464 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
465 * have to call rcu_read_lock/unlock() in this function.
466 */
467static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
468{
469 struct task_struct *t;
470
471 for_each_thread(tsk, t) {
472 if (t->flags & PF_MCE_PROCESS) {
473 if (t->flags & PF_MCE_EARLY)
474 return t;
475 } else {
476 if (sysctl_memory_failure_early_kill)
477 return t;
478 }
479 }
480 return NULL;
481}
482
483/*
484 * Determine whether a given process is "early kill" process which expects
485 * to be signaled when some page under the process is hwpoisoned.
486 * Return task_struct of the dedicated thread (main thread unless explicitly
487 * specified) if the process is "early kill" and otherwise returns NULL.
488 *
489 * Note that the above is true for Action Optional case. For Action Required
490 * case, it's only meaningful to the current thread which need to be signaled
491 * with SIGBUS, this error is Action Optional for other non current
492 * processes sharing the same error page,if the process is "early kill", the
493 * task_struct of the dedicated thread will also be returned.
494 */
495static struct task_struct *task_early_kill(struct task_struct *tsk,
496 int force_early)
497{
498 if (!tsk->mm)
499 return NULL;
500 /*
501 * Comparing ->mm here because current task might represent
502 * a subthread, while tsk always points to the main thread.
503 */
504 if (force_early && tsk->mm == current->mm)
505 return current;
506
507 return find_early_kill_thread(tsk);
508}
509
510/*
511 * Collect processes when the error hit an anonymous page.
512 */
513static void collect_procs_anon(struct page *page, struct list_head *to_kill,
514 int force_early)
515{
516 struct folio *folio = page_folio(page);
517 struct vm_area_struct *vma;
518 struct task_struct *tsk;
519 struct anon_vma *av;
520 pgoff_t pgoff;
521
522 av = folio_lock_anon_vma_read(folio, NULL);
523 if (av == NULL) /* Not actually mapped anymore */
524 return;
525
526 pgoff = page_to_pgoff(page);
527 read_lock(&tasklist_lock);
528 for_each_process (tsk) {
529 struct anon_vma_chain *vmac;
530 struct task_struct *t = task_early_kill(tsk, force_early);
531
532 if (!t)
533 continue;
534 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
535 pgoff, pgoff) {
536 vma = vmac->vma;
537 if (vma->vm_mm != t->mm)
538 continue;
539 if (!page_mapped_in_vma(page, vma))
540 continue;
541 add_to_kill(t, page, FSDAX_INVALID_PGOFF, vma, to_kill);
542 }
543 }
544 read_unlock(&tasklist_lock);
545 anon_vma_unlock_read(av);
546}
547
548/*
549 * Collect processes when the error hit a file mapped page.
550 */
551static void collect_procs_file(struct page *page, struct list_head *to_kill,
552 int force_early)
553{
554 struct vm_area_struct *vma;
555 struct task_struct *tsk;
556 struct address_space *mapping = page->mapping;
557 pgoff_t pgoff;
558
559 i_mmap_lock_read(mapping);
560 read_lock(&tasklist_lock);
561 pgoff = page_to_pgoff(page);
562 for_each_process(tsk) {
563 struct task_struct *t = task_early_kill(tsk, force_early);
564
565 if (!t)
566 continue;
567 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
568 pgoff) {
569 /*
570 * Send early kill signal to tasks where a vma covers
571 * the page but the corrupted page is not necessarily
572 * mapped it in its pte.
573 * Assume applications who requested early kill want
574 * to be informed of all such data corruptions.
575 */
576 if (vma->vm_mm == t->mm)
577 add_to_kill(t, page, FSDAX_INVALID_PGOFF, vma,
578 to_kill);
579 }
580 }
581 read_unlock(&tasklist_lock);
582 i_mmap_unlock_read(mapping);
583}
584
585#ifdef CONFIG_FS_DAX
586/*
587 * Collect processes when the error hit a fsdax page.
588 */
589static void collect_procs_fsdax(struct page *page,
590 struct address_space *mapping, pgoff_t pgoff,
591 struct list_head *to_kill)
592{
593 struct vm_area_struct *vma;
594 struct task_struct *tsk;
595
596 i_mmap_lock_read(mapping);
597 read_lock(&tasklist_lock);
598 for_each_process(tsk) {
599 struct task_struct *t = task_early_kill(tsk, true);
600
601 if (!t)
602 continue;
603 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
604 if (vma->vm_mm == t->mm)
605 add_to_kill(t, page, pgoff, vma, to_kill);
606 }
607 }
608 read_unlock(&tasklist_lock);
609 i_mmap_unlock_read(mapping);
610}
611#endif /* CONFIG_FS_DAX */
612
613/*
614 * Collect the processes who have the corrupted page mapped to kill.
615 */
616static void collect_procs(struct page *page, struct list_head *tokill,
617 int force_early)
618{
619 if (!page->mapping)
620 return;
621
622 if (PageAnon(page))
623 collect_procs_anon(page, tokill, force_early);
624 else
625 collect_procs_file(page, tokill, force_early);
626}
627
628struct hwp_walk {
629 struct to_kill tk;
630 unsigned long pfn;
631 int flags;
632};
633
634static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift)
635{
636 tk->addr = addr;
637 tk->size_shift = shift;
638}
639
640static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift,
641 unsigned long poisoned_pfn, struct to_kill *tk)
642{
643 unsigned long pfn = 0;
644
645 if (pte_present(pte)) {
646 pfn = pte_pfn(pte);
647 } else {
648 swp_entry_t swp = pte_to_swp_entry(pte);
649
650 if (is_hwpoison_entry(swp))
651 pfn = swp_offset_pfn(swp);
652 }
653
654 if (!pfn || pfn != poisoned_pfn)
655 return 0;
656
657 set_to_kill(tk, addr, shift);
658 return 1;
659}
660
661#ifdef CONFIG_TRANSPARENT_HUGEPAGE
662static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
663 struct hwp_walk *hwp)
664{
665 pmd_t pmd = *pmdp;
666 unsigned long pfn;
667 unsigned long hwpoison_vaddr;
668
669 if (!pmd_present(pmd))
670 return 0;
671 pfn = pmd_pfn(pmd);
672 if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) {
673 hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT);
674 set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT);
675 return 1;
676 }
677 return 0;
678}
679#else
680static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
681 struct hwp_walk *hwp)
682{
683 return 0;
684}
685#endif
686
687static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr,
688 unsigned long end, struct mm_walk *walk)
689{
690 struct hwp_walk *hwp = walk->private;
691 int ret = 0;
692 pte_t *ptep, *mapped_pte;
693 spinlock_t *ptl;
694
695 ptl = pmd_trans_huge_lock(pmdp, walk->vma);
696 if (ptl) {
697 ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp);
698 spin_unlock(ptl);
699 goto out;
700 }
701
702 if (pmd_trans_unstable(pmdp))
703 goto out;
704
705 mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp,
706 addr, &ptl);
707 for (; addr != end; ptep++, addr += PAGE_SIZE) {
708 ret = check_hwpoisoned_entry(*ptep, addr, PAGE_SHIFT,
709 hwp->pfn, &hwp->tk);
710 if (ret == 1)
711 break;
712 }
713 pte_unmap_unlock(mapped_pte, ptl);
714out:
715 cond_resched();
716 return ret;
717}
718
719#ifdef CONFIG_HUGETLB_PAGE
720static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask,
721 unsigned long addr, unsigned long end,
722 struct mm_walk *walk)
723{
724 struct hwp_walk *hwp = walk->private;
725 pte_t pte = huge_ptep_get(ptep);
726 struct hstate *h = hstate_vma(walk->vma);
727
728 return check_hwpoisoned_entry(pte, addr, huge_page_shift(h),
729 hwp->pfn, &hwp->tk);
730}
731#else
732#define hwpoison_hugetlb_range NULL
733#endif
734
735static const struct mm_walk_ops hwp_walk_ops = {
736 .pmd_entry = hwpoison_pte_range,
737 .hugetlb_entry = hwpoison_hugetlb_range,
738};
739
740/*
741 * Sends SIGBUS to the current process with error info.
742 *
743 * This function is intended to handle "Action Required" MCEs on already
744 * hardware poisoned pages. They could happen, for example, when
745 * memory_failure() failed to unmap the error page at the first call, or
746 * when multiple local machine checks happened on different CPUs.
747 *
748 * MCE handler currently has no easy access to the error virtual address,
749 * so this function walks page table to find it. The returned virtual address
750 * is proper in most cases, but it could be wrong when the application
751 * process has multiple entries mapping the error page.
752 */
753static int kill_accessing_process(struct task_struct *p, unsigned long pfn,
754 int flags)
755{
756 int ret;
757 struct hwp_walk priv = {
758 .pfn = pfn,
759 };
760 priv.tk.tsk = p;
761
762 if (!p->mm)
763 return -EFAULT;
764
765 mmap_read_lock(p->mm);
766 ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwp_walk_ops,
767 (void *)&priv);
768 if (ret == 1 && priv.tk.addr)
769 kill_proc(&priv.tk, pfn, flags);
770 else
771 ret = 0;
772 mmap_read_unlock(p->mm);
773 return ret > 0 ? -EHWPOISON : -EFAULT;
774}
775
776static const char *action_name[] = {
777 [MF_IGNORED] = "Ignored",
778 [MF_FAILED] = "Failed",
779 [MF_DELAYED] = "Delayed",
780 [MF_RECOVERED] = "Recovered",
781};
782
783static const char * const action_page_types[] = {
784 [MF_MSG_KERNEL] = "reserved kernel page",
785 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
786 [MF_MSG_SLAB] = "kernel slab page",
787 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
788 [MF_MSG_HUGE] = "huge page",
789 [MF_MSG_FREE_HUGE] = "free huge page",
790 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
791 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
792 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
793 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
794 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
795 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
796 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
797 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
798 [MF_MSG_CLEAN_LRU] = "clean LRU page",
799 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
800 [MF_MSG_BUDDY] = "free buddy page",
801 [MF_MSG_DAX] = "dax page",
802 [MF_MSG_UNSPLIT_THP] = "unsplit thp",
803 [MF_MSG_UNKNOWN] = "unknown page",
804};
805
806/*
807 * XXX: It is possible that a page is isolated from LRU cache,
808 * and then kept in swap cache or failed to remove from page cache.
809 * The page count will stop it from being freed by unpoison.
810 * Stress tests should be aware of this memory leak problem.
811 */
812static int delete_from_lru_cache(struct page *p)
813{
814 if (!isolate_lru_page(p)) {
815 /*
816 * Clear sensible page flags, so that the buddy system won't
817 * complain when the page is unpoison-and-freed.
818 */
819 ClearPageActive(p);
820 ClearPageUnevictable(p);
821
822 /*
823 * Poisoned page might never drop its ref count to 0 so we have
824 * to uncharge it manually from its memcg.
825 */
826 mem_cgroup_uncharge(page_folio(p));
827
828 /*
829 * drop the page count elevated by isolate_lru_page()
830 */
831 put_page(p);
832 return 0;
833 }
834 return -EIO;
835}
836
837static int truncate_error_page(struct page *p, unsigned long pfn,
838 struct address_space *mapping)
839{
840 int ret = MF_FAILED;
841
842 if (mapping->a_ops->error_remove_page) {
843 struct folio *folio = page_folio(p);
844 int err = mapping->a_ops->error_remove_page(mapping, p);
845
846 if (err != 0) {
847 pr_info("%#lx: Failed to punch page: %d\n", pfn, err);
848 } else if (folio_has_private(folio) &&
849 !filemap_release_folio(folio, GFP_NOIO)) {
850 pr_info("%#lx: failed to release buffers\n", pfn);
851 } else {
852 ret = MF_RECOVERED;
853 }
854 } else {
855 /*
856 * If the file system doesn't support it just invalidate
857 * This fails on dirty or anything with private pages
858 */
859 if (invalidate_inode_page(p))
860 ret = MF_RECOVERED;
861 else
862 pr_info("%#lx: Failed to invalidate\n", pfn);
863 }
864
865 return ret;
866}
867
868struct page_state {
869 unsigned long mask;
870 unsigned long res;
871 enum mf_action_page_type type;
872
873 /* Callback ->action() has to unlock the relevant page inside it. */
874 int (*action)(struct page_state *ps, struct page *p);
875};
876
877/*
878 * Return true if page is still referenced by others, otherwise return
879 * false.
880 *
881 * The extra_pins is true when one extra refcount is expected.
882 */
883static bool has_extra_refcount(struct page_state *ps, struct page *p,
884 bool extra_pins)
885{
886 int count = page_count(p) - 1;
887
888 if (extra_pins)
889 count -= 1;
890
891 if (count > 0) {
892 pr_err("%#lx: %s still referenced by %d users\n",
893 page_to_pfn(p), action_page_types[ps->type], count);
894 return true;
895 }
896
897 return false;
898}
899
900/*
901 * Error hit kernel page.
902 * Do nothing, try to be lucky and not touch this instead. For a few cases we
903 * could be more sophisticated.
904 */
905static int me_kernel(struct page_state *ps, struct page *p)
906{
907 unlock_page(p);
908 return MF_IGNORED;
909}
910
911/*
912 * Page in unknown state. Do nothing.
913 */
914static int me_unknown(struct page_state *ps, struct page *p)
915{
916 pr_err("%#lx: Unknown page state\n", page_to_pfn(p));
917 unlock_page(p);
918 return MF_FAILED;
919}
920
921/*
922 * Clean (or cleaned) page cache page.
923 */
924static int me_pagecache_clean(struct page_state *ps, struct page *p)
925{
926 int ret;
927 struct address_space *mapping;
928 bool extra_pins;
929
930 delete_from_lru_cache(p);
931
932 /*
933 * For anonymous pages we're done the only reference left
934 * should be the one m_f() holds.
935 */
936 if (PageAnon(p)) {
937 ret = MF_RECOVERED;
938 goto out;
939 }
940
941 /*
942 * Now truncate the page in the page cache. This is really
943 * more like a "temporary hole punch"
944 * Don't do this for block devices when someone else
945 * has a reference, because it could be file system metadata
946 * and that's not safe to truncate.
947 */
948 mapping = page_mapping(p);
949 if (!mapping) {
950 /*
951 * Page has been teared down in the meanwhile
952 */
953 ret = MF_FAILED;
954 goto out;
955 }
956
957 /*
958 * The shmem page is kept in page cache instead of truncating
959 * so is expected to have an extra refcount after error-handling.
960 */
961 extra_pins = shmem_mapping(mapping);
962
963 /*
964 * Truncation is a bit tricky. Enable it per file system for now.
965 *
966 * Open: to take i_rwsem or not for this? Right now we don't.
967 */
968 ret = truncate_error_page(p, page_to_pfn(p), mapping);
969 if (has_extra_refcount(ps, p, extra_pins))
970 ret = MF_FAILED;
971
972out:
973 unlock_page(p);
974
975 return ret;
976}
977
978/*
979 * Dirty pagecache page
980 * Issues: when the error hit a hole page the error is not properly
981 * propagated.
982 */
983static int me_pagecache_dirty(struct page_state *ps, struct page *p)
984{
985 struct address_space *mapping = page_mapping(p);
986
987 SetPageError(p);
988 /* TBD: print more information about the file. */
989 if (mapping) {
990 /*
991 * IO error will be reported by write(), fsync(), etc.
992 * who check the mapping.
993 * This way the application knows that something went
994 * wrong with its dirty file data.
995 *
996 * There's one open issue:
997 *
998 * The EIO will be only reported on the next IO
999 * operation and then cleared through the IO map.
1000 * Normally Linux has two mechanisms to pass IO error
1001 * first through the AS_EIO flag in the address space
1002 * and then through the PageError flag in the page.
1003 * Since we drop pages on memory failure handling the
1004 * only mechanism open to use is through AS_AIO.
1005 *
1006 * This has the disadvantage that it gets cleared on
1007 * the first operation that returns an error, while
1008 * the PageError bit is more sticky and only cleared
1009 * when the page is reread or dropped. If an
1010 * application assumes it will always get error on
1011 * fsync, but does other operations on the fd before
1012 * and the page is dropped between then the error
1013 * will not be properly reported.
1014 *
1015 * This can already happen even without hwpoisoned
1016 * pages: first on metadata IO errors (which only
1017 * report through AS_EIO) or when the page is dropped
1018 * at the wrong time.
1019 *
1020 * So right now we assume that the application DTRT on
1021 * the first EIO, but we're not worse than other parts
1022 * of the kernel.
1023 */
1024 mapping_set_error(mapping, -EIO);
1025 }
1026
1027 return me_pagecache_clean(ps, p);
1028}
1029
1030/*
1031 * Clean and dirty swap cache.
1032 *
1033 * Dirty swap cache page is tricky to handle. The page could live both in page
1034 * cache and swap cache(ie. page is freshly swapped in). So it could be
1035 * referenced concurrently by 2 types of PTEs:
1036 * normal PTEs and swap PTEs. We try to handle them consistently by calling
1037 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
1038 * and then
1039 * - clear dirty bit to prevent IO
1040 * - remove from LRU
1041 * - but keep in the swap cache, so that when we return to it on
1042 * a later page fault, we know the application is accessing
1043 * corrupted data and shall be killed (we installed simple
1044 * interception code in do_swap_page to catch it).
1045 *
1046 * Clean swap cache pages can be directly isolated. A later page fault will
1047 * bring in the known good data from disk.
1048 */
1049static int me_swapcache_dirty(struct page_state *ps, struct page *p)
1050{
1051 int ret;
1052 bool extra_pins = false;
1053
1054 ClearPageDirty(p);
1055 /* Trigger EIO in shmem: */
1056 ClearPageUptodate(p);
1057
1058 ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED;
1059 unlock_page(p);
1060
1061 if (ret == MF_DELAYED)
1062 extra_pins = true;
1063
1064 if (has_extra_refcount(ps, p, extra_pins))
1065 ret = MF_FAILED;
1066
1067 return ret;
1068}
1069
1070static int me_swapcache_clean(struct page_state *ps, struct page *p)
1071{
1072 struct folio *folio = page_folio(p);
1073 int ret;
1074
1075 delete_from_swap_cache(folio);
1076
1077 ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED;
1078 folio_unlock(folio);
1079
1080 if (has_extra_refcount(ps, p, false))
1081 ret = MF_FAILED;
1082
1083 return ret;
1084}
1085
1086/*
1087 * Huge pages. Needs work.
1088 * Issues:
1089 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
1090 * To narrow down kill region to one page, we need to break up pmd.
1091 */
1092static int me_huge_page(struct page_state *ps, struct page *p)
1093{
1094 int res;
1095 struct page *hpage = compound_head(p);
1096 struct address_space *mapping;
1097 bool extra_pins = false;
1098
1099 if (!PageHuge(hpage))
1100 return MF_DELAYED;
1101
1102 mapping = page_mapping(hpage);
1103 if (mapping) {
1104 res = truncate_error_page(hpage, page_to_pfn(p), mapping);
1105 /* The page is kept in page cache. */
1106 extra_pins = true;
1107 unlock_page(hpage);
1108 } else {
1109 unlock_page(hpage);
1110 /*
1111 * migration entry prevents later access on error hugepage,
1112 * so we can free and dissolve it into buddy to save healthy
1113 * subpages.
1114 */
1115 put_page(hpage);
1116 if (__page_handle_poison(p) >= 0) {
1117 page_ref_inc(p);
1118 res = MF_RECOVERED;
1119 } else {
1120 res = MF_FAILED;
1121 }
1122 }
1123
1124 if (has_extra_refcount(ps, p, extra_pins))
1125 res = MF_FAILED;
1126
1127 return res;
1128}
1129
1130/*
1131 * Various page states we can handle.
1132 *
1133 * A page state is defined by its current page->flags bits.
1134 * The table matches them in order and calls the right handler.
1135 *
1136 * This is quite tricky because we can access page at any time
1137 * in its live cycle, so all accesses have to be extremely careful.
1138 *
1139 * This is not complete. More states could be added.
1140 * For any missing state don't attempt recovery.
1141 */
1142
1143#define dirty (1UL << PG_dirty)
1144#define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked))
1145#define unevict (1UL << PG_unevictable)
1146#define mlock (1UL << PG_mlocked)
1147#define lru (1UL << PG_lru)
1148#define head (1UL << PG_head)
1149#define slab (1UL << PG_slab)
1150#define reserved (1UL << PG_reserved)
1151
1152static struct page_state error_states[] = {
1153 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
1154 /*
1155 * free pages are specially detected outside this table:
1156 * PG_buddy pages only make a small fraction of all free pages.
1157 */
1158
1159 /*
1160 * Could in theory check if slab page is free or if we can drop
1161 * currently unused objects without touching them. But just
1162 * treat it as standard kernel for now.
1163 */
1164 { slab, slab, MF_MSG_SLAB, me_kernel },
1165
1166 { head, head, MF_MSG_HUGE, me_huge_page },
1167
1168 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
1169 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
1170
1171 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
1172 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
1173
1174 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
1175 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
1176
1177 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
1178 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
1179
1180 /*
1181 * Catchall entry: must be at end.
1182 */
1183 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
1184};
1185
1186#undef dirty
1187#undef sc
1188#undef unevict
1189#undef mlock
1190#undef lru
1191#undef head
1192#undef slab
1193#undef reserved
1194
1195/*
1196 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
1197 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
1198 */
1199static int action_result(unsigned long pfn, enum mf_action_page_type type,
1200 enum mf_result result)
1201{
1202 trace_memory_failure_event(pfn, type, result);
1203
1204 num_poisoned_pages_inc(pfn);
1205 pr_err("%#lx: recovery action for %s: %s\n",
1206 pfn, action_page_types[type], action_name[result]);
1207
1208 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
1209}
1210
1211static int page_action(struct page_state *ps, struct page *p,
1212 unsigned long pfn)
1213{
1214 int result;
1215
1216 /* page p should be unlocked after returning from ps->action(). */
1217 result = ps->action(ps, p);
1218
1219 /* Could do more checks here if page looks ok */
1220 /*
1221 * Could adjust zone counters here to correct for the missing page.
1222 */
1223
1224 return action_result(pfn, ps->type, result);
1225}
1226
1227static inline bool PageHWPoisonTakenOff(struct page *page)
1228{
1229 return PageHWPoison(page) && page_private(page) == MAGIC_HWPOISON;
1230}
1231
1232void SetPageHWPoisonTakenOff(struct page *page)
1233{
1234 set_page_private(page, MAGIC_HWPOISON);
1235}
1236
1237void ClearPageHWPoisonTakenOff(struct page *page)
1238{
1239 if (PageHWPoison(page))
1240 set_page_private(page, 0);
1241}
1242
1243/*
1244 * Return true if a page type of a given page is supported by hwpoison
1245 * mechanism (while handling could fail), otherwise false. This function
1246 * does not return true for hugetlb or device memory pages, so it's assumed
1247 * to be called only in the context where we never have such pages.
1248 */
1249static inline bool HWPoisonHandlable(struct page *page, unsigned long flags)
1250{
1251 /* Soft offline could migrate non-LRU movable pages */
1252 if ((flags & MF_SOFT_OFFLINE) && __PageMovable(page))
1253 return true;
1254
1255 return PageLRU(page) || is_free_buddy_page(page);
1256}
1257
1258static int __get_hwpoison_page(struct page *page, unsigned long flags)
1259{
1260 struct page *head = compound_head(page);
1261 int ret = 0;
1262 bool hugetlb = false;
1263
1264 ret = get_hwpoison_huge_page(head, &hugetlb, false);
1265 if (hugetlb)
1266 return ret;
1267
1268 /*
1269 * This check prevents from calling get_page_unless_zero() for any
1270 * unsupported type of page in order to reduce the risk of unexpected
1271 * races caused by taking a page refcount.
1272 */
1273 if (!HWPoisonHandlable(head, flags))
1274 return -EBUSY;
1275
1276 if (get_page_unless_zero(head)) {
1277 if (head == compound_head(page))
1278 return 1;
1279
1280 pr_info("%#lx cannot catch tail\n", page_to_pfn(page));
1281 put_page(head);
1282 }
1283
1284 return 0;
1285}
1286
1287static int get_any_page(struct page *p, unsigned long flags)
1288{
1289 int ret = 0, pass = 0;
1290 bool count_increased = false;
1291
1292 if (flags & MF_COUNT_INCREASED)
1293 count_increased = true;
1294
1295try_again:
1296 if (!count_increased) {
1297 ret = __get_hwpoison_page(p, flags);
1298 if (!ret) {
1299 if (page_count(p)) {
1300 /* We raced with an allocation, retry. */
1301 if (pass++ < 3)
1302 goto try_again;
1303 ret = -EBUSY;
1304 } else if (!PageHuge(p) && !is_free_buddy_page(p)) {
1305 /* We raced with put_page, retry. */
1306 if (pass++ < 3)
1307 goto try_again;
1308 ret = -EIO;
1309 }
1310 goto out;
1311 } else if (ret == -EBUSY) {
1312 /*
1313 * We raced with (possibly temporary) unhandlable
1314 * page, retry.
1315 */
1316 if (pass++ < 3) {
1317 shake_page(p);
1318 goto try_again;
1319 }
1320 ret = -EIO;
1321 goto out;
1322 }
1323 }
1324
1325 if (PageHuge(p) || HWPoisonHandlable(p, flags)) {
1326 ret = 1;
1327 } else {
1328 /*
1329 * A page we cannot handle. Check whether we can turn
1330 * it into something we can handle.
1331 */
1332 if (pass++ < 3) {
1333 put_page(p);
1334 shake_page(p);
1335 count_increased = false;
1336 goto try_again;
1337 }
1338 put_page(p);
1339 ret = -EIO;
1340 }
1341out:
1342 if (ret == -EIO)
1343 pr_err("%#lx: unhandlable page.\n", page_to_pfn(p));
1344
1345 return ret;
1346}
1347
1348static int __get_unpoison_page(struct page *page)
1349{
1350 struct page *head = compound_head(page);
1351 int ret = 0;
1352 bool hugetlb = false;
1353
1354 ret = get_hwpoison_huge_page(head, &hugetlb, true);
1355 if (hugetlb)
1356 return ret;
1357
1358 /*
1359 * PageHWPoisonTakenOff pages are not only marked as PG_hwpoison,
1360 * but also isolated from buddy freelist, so need to identify the
1361 * state and have to cancel both operations to unpoison.
1362 */
1363 if (PageHWPoisonTakenOff(page))
1364 return -EHWPOISON;
1365
1366 return get_page_unless_zero(page) ? 1 : 0;
1367}
1368
1369/**
1370 * get_hwpoison_page() - Get refcount for memory error handling
1371 * @p: Raw error page (hit by memory error)
1372 * @flags: Flags controlling behavior of error handling
1373 *
1374 * get_hwpoison_page() takes a page refcount of an error page to handle memory
1375 * error on it, after checking that the error page is in a well-defined state
1376 * (defined as a page-type we can successfully handle the memory error on it,
1377 * such as LRU page and hugetlb page).
1378 *
1379 * Memory error handling could be triggered at any time on any type of page,
1380 * so it's prone to race with typical memory management lifecycle (like
1381 * allocation and free). So to avoid such races, get_hwpoison_page() takes
1382 * extra care for the error page's state (as done in __get_hwpoison_page()),
1383 * and has some retry logic in get_any_page().
1384 *
1385 * When called from unpoison_memory(), the caller should already ensure that
1386 * the given page has PG_hwpoison. So it's never reused for other page
1387 * allocations, and __get_unpoison_page() never races with them.
1388 *
1389 * Return: 0 on failure,
1390 * 1 on success for in-use pages in a well-defined state,
1391 * -EIO for pages on which we can not handle memory errors,
1392 * -EBUSY when get_hwpoison_page() has raced with page lifecycle
1393 * operations like allocation and free,
1394 * -EHWPOISON when the page is hwpoisoned and taken off from buddy.
1395 */
1396static int get_hwpoison_page(struct page *p, unsigned long flags)
1397{
1398 int ret;
1399
1400 zone_pcp_disable(page_zone(p));
1401 if (flags & MF_UNPOISON)
1402 ret = __get_unpoison_page(p);
1403 else
1404 ret = get_any_page(p, flags);
1405 zone_pcp_enable(page_zone(p));
1406
1407 return ret;
1408}
1409
1410/*
1411 * Do all that is necessary to remove user space mappings. Unmap
1412 * the pages and send SIGBUS to the processes if the data was dirty.
1413 */
1414static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
1415 int flags, struct page *hpage)
1416{
1417 struct folio *folio = page_folio(hpage);
1418 enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC;
1419 struct address_space *mapping;
1420 LIST_HEAD(tokill);
1421 bool unmap_success;
1422 int forcekill;
1423 bool mlocked = PageMlocked(hpage);
1424
1425 /*
1426 * Here we are interested only in user-mapped pages, so skip any
1427 * other types of pages.
1428 */
1429 if (PageReserved(p) || PageSlab(p) || PageTable(p))
1430 return true;
1431 if (!(PageLRU(hpage) || PageHuge(p)))
1432 return true;
1433
1434 /*
1435 * This check implies we don't kill processes if their pages
1436 * are in the swap cache early. Those are always late kills.
1437 */
1438 if (!page_mapped(hpage))
1439 return true;
1440
1441 if (PageKsm(p)) {
1442 pr_err("%#lx: can't handle KSM pages.\n", pfn);
1443 return false;
1444 }
1445
1446 if (PageSwapCache(p)) {
1447 pr_err("%#lx: keeping poisoned page in swap cache\n", pfn);
1448 ttu |= TTU_IGNORE_HWPOISON;
1449 }
1450
1451 /*
1452 * Propagate the dirty bit from PTEs to struct page first, because we
1453 * need this to decide if we should kill or just drop the page.
1454 * XXX: the dirty test could be racy: set_page_dirty() may not always
1455 * be called inside page lock (it's recommended but not enforced).
1456 */
1457 mapping = page_mapping(hpage);
1458 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
1459 mapping_can_writeback(mapping)) {
1460 if (page_mkclean(hpage)) {
1461 SetPageDirty(hpage);
1462 } else {
1463 ttu |= TTU_IGNORE_HWPOISON;
1464 pr_info("%#lx: corrupted page was clean: dropped without side effects\n",
1465 pfn);
1466 }
1467 }
1468
1469 /*
1470 * First collect all the processes that have the page
1471 * mapped in dirty form. This has to be done before try_to_unmap,
1472 * because ttu takes the rmap data structures down.
1473 */
1474 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
1475
1476 if (PageHuge(hpage) && !PageAnon(hpage)) {
1477 /*
1478 * For hugetlb pages in shared mappings, try_to_unmap
1479 * could potentially call huge_pmd_unshare. Because of
1480 * this, take semaphore in write mode here and set
1481 * TTU_RMAP_LOCKED to indicate we have taken the lock
1482 * at this higher level.
1483 */
1484 mapping = hugetlb_page_mapping_lock_write(hpage);
1485 if (mapping) {
1486 try_to_unmap(folio, ttu|TTU_RMAP_LOCKED);
1487 i_mmap_unlock_write(mapping);
1488 } else
1489 pr_info("%#lx: could not lock mapping for mapped huge page\n", pfn);
1490 } else {
1491 try_to_unmap(folio, ttu);
1492 }
1493
1494 unmap_success = !page_mapped(hpage);
1495 if (!unmap_success)
1496 pr_err("%#lx: failed to unmap page (mapcount=%d)\n",
1497 pfn, page_mapcount(hpage));
1498
1499 /*
1500 * try_to_unmap() might put mlocked page in lru cache, so call
1501 * shake_page() again to ensure that it's flushed.
1502 */
1503 if (mlocked)
1504 shake_page(hpage);
1505
1506 /*
1507 * Now that the dirty bit has been propagated to the
1508 * struct page and all unmaps done we can decide if
1509 * killing is needed or not. Only kill when the page
1510 * was dirty or the process is not restartable,
1511 * otherwise the tokill list is merely
1512 * freed. When there was a problem unmapping earlier
1513 * use a more force-full uncatchable kill to prevent
1514 * any accesses to the poisoned memory.
1515 */
1516 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL) ||
1517 !unmap_success;
1518 kill_procs(&tokill, forcekill, !unmap_success, pfn, flags);
1519
1520 return unmap_success;
1521}
1522
1523static int identify_page_state(unsigned long pfn, struct page *p,
1524 unsigned long page_flags)
1525{
1526 struct page_state *ps;
1527
1528 /*
1529 * The first check uses the current page flags which may not have any
1530 * relevant information. The second check with the saved page flags is
1531 * carried out only if the first check can't determine the page status.
1532 */
1533 for (ps = error_states;; ps++)
1534 if ((p->flags & ps->mask) == ps->res)
1535 break;
1536
1537 page_flags |= (p->flags & (1UL << PG_dirty));
1538
1539 if (!ps->mask)
1540 for (ps = error_states;; ps++)
1541 if ((page_flags & ps->mask) == ps->res)
1542 break;
1543 return page_action(ps, p, pfn);
1544}
1545
1546static int try_to_split_thp_page(struct page *page)
1547{
1548 int ret;
1549
1550 lock_page(page);
1551 ret = split_huge_page(page);
1552 unlock_page(page);
1553
1554 if (unlikely(ret))
1555 put_page(page);
1556
1557 return ret;
1558}
1559
1560static void unmap_and_kill(struct list_head *to_kill, unsigned long pfn,
1561 struct address_space *mapping, pgoff_t index, int flags)
1562{
1563 struct to_kill *tk;
1564 unsigned long size = 0;
1565
1566 list_for_each_entry(tk, to_kill, nd)
1567 if (tk->size_shift)
1568 size = max(size, 1UL << tk->size_shift);
1569
1570 if (size) {
1571 /*
1572 * Unmap the largest mapping to avoid breaking up device-dax
1573 * mappings which are constant size. The actual size of the
1574 * mapping being torn down is communicated in siginfo, see
1575 * kill_proc()
1576 */
1577 loff_t start = (index << PAGE_SHIFT) & ~(size - 1);
1578
1579 unmap_mapping_range(mapping, start, size, 0);
1580 }
1581
1582 kill_procs(to_kill, flags & MF_MUST_KILL, false, pfn, flags);
1583}
1584
1585static int mf_generic_kill_procs(unsigned long long pfn, int flags,
1586 struct dev_pagemap *pgmap)
1587{
1588 struct page *page = pfn_to_page(pfn);
1589 LIST_HEAD(to_kill);
1590 dax_entry_t cookie;
1591 int rc = 0;
1592
1593 /*
1594 * Pages instantiated by device-dax (not filesystem-dax)
1595 * may be compound pages.
1596 */
1597 page = compound_head(page);
1598
1599 /*
1600 * Prevent the inode from being freed while we are interrogating
1601 * the address_space, typically this would be handled by
1602 * lock_page(), but dax pages do not use the page lock. This
1603 * also prevents changes to the mapping of this pfn until
1604 * poison signaling is complete.
1605 */
1606 cookie = dax_lock_page(page);
1607 if (!cookie)
1608 return -EBUSY;
1609
1610 if (hwpoison_filter(page)) {
1611 rc = -EOPNOTSUPP;
1612 goto unlock;
1613 }
1614
1615 switch (pgmap->type) {
1616 case MEMORY_DEVICE_PRIVATE:
1617 case MEMORY_DEVICE_COHERENT:
1618 /*
1619 * TODO: Handle device pages which may need coordination
1620 * with device-side memory.
1621 */
1622 rc = -ENXIO;
1623 goto unlock;
1624 default:
1625 break;
1626 }
1627
1628 /*
1629 * Use this flag as an indication that the dax page has been
1630 * remapped UC to prevent speculative consumption of poison.
1631 */
1632 SetPageHWPoison(page);
1633
1634 /*
1635 * Unlike System-RAM there is no possibility to swap in a
1636 * different physical page at a given virtual address, so all
1637 * userspace consumption of ZONE_DEVICE memory necessitates
1638 * SIGBUS (i.e. MF_MUST_KILL)
1639 */
1640 flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1641 collect_procs(page, &to_kill, true);
1642
1643 unmap_and_kill(&to_kill, pfn, page->mapping, page->index, flags);
1644unlock:
1645 dax_unlock_page(page, cookie);
1646 return rc;
1647}
1648
1649#ifdef CONFIG_FS_DAX
1650/**
1651 * mf_dax_kill_procs - Collect and kill processes who are using this file range
1652 * @mapping: address_space of the file in use
1653 * @index: start pgoff of the range within the file
1654 * @count: length of the range, in unit of PAGE_SIZE
1655 * @mf_flags: memory failure flags
1656 */
1657int mf_dax_kill_procs(struct address_space *mapping, pgoff_t index,
1658 unsigned long count, int mf_flags)
1659{
1660 LIST_HEAD(to_kill);
1661 dax_entry_t cookie;
1662 struct page *page;
1663 size_t end = index + count;
1664
1665 mf_flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1666
1667 for (; index < end; index++) {
1668 page = NULL;
1669 cookie = dax_lock_mapping_entry(mapping, index, &page);
1670 if (!cookie)
1671 return -EBUSY;
1672 if (!page)
1673 goto unlock;
1674
1675 SetPageHWPoison(page);
1676
1677 collect_procs_fsdax(page, mapping, index, &to_kill);
1678 unmap_and_kill(&to_kill, page_to_pfn(page), mapping,
1679 index, mf_flags);
1680unlock:
1681 dax_unlock_mapping_entry(mapping, index, cookie);
1682 }
1683 return 0;
1684}
1685EXPORT_SYMBOL_GPL(mf_dax_kill_procs);
1686#endif /* CONFIG_FS_DAX */
1687
1688#ifdef CONFIG_HUGETLB_PAGE
1689/*
1690 * Struct raw_hwp_page represents information about "raw error page",
1691 * constructing singly linked list from ->_hugetlb_hwpoison field of folio.
1692 */
1693struct raw_hwp_page {
1694 struct llist_node node;
1695 struct page *page;
1696};
1697
1698static inline struct llist_head *raw_hwp_list_head(struct page *hpage)
1699{
1700 return (struct llist_head *)&page_folio(hpage)->_hugetlb_hwpoison;
1701}
1702
1703static unsigned long __free_raw_hwp_pages(struct page *hpage, bool move_flag)
1704{
1705 struct llist_head *head;
1706 struct llist_node *t, *tnode;
1707 unsigned long count = 0;
1708
1709 head = raw_hwp_list_head(hpage);
1710 llist_for_each_safe(tnode, t, head->first) {
1711 struct raw_hwp_page *p = container_of(tnode, struct raw_hwp_page, node);
1712
1713 if (move_flag)
1714 SetPageHWPoison(p->page);
1715 else
1716 num_poisoned_pages_sub(page_to_pfn(p->page), 1);
1717 kfree(p);
1718 count++;
1719 }
1720 llist_del_all(head);
1721 return count;
1722}
1723
1724static int hugetlb_set_page_hwpoison(struct page *hpage, struct page *page)
1725{
1726 struct llist_head *head;
1727 struct raw_hwp_page *raw_hwp;
1728 struct llist_node *t, *tnode;
1729 int ret = TestSetPageHWPoison(hpage) ? -EHWPOISON : 0;
1730
1731 /*
1732 * Once the hwpoison hugepage has lost reliable raw error info,
1733 * there is little meaning to keep additional error info precisely,
1734 * so skip to add additional raw error info.
1735 */
1736 if (HPageRawHwpUnreliable(hpage))
1737 return -EHWPOISON;
1738 head = raw_hwp_list_head(hpage);
1739 llist_for_each_safe(tnode, t, head->first) {
1740 struct raw_hwp_page *p = container_of(tnode, struct raw_hwp_page, node);
1741
1742 if (p->page == page)
1743 return -EHWPOISON;
1744 }
1745
1746 raw_hwp = kmalloc(sizeof(struct raw_hwp_page), GFP_ATOMIC);
1747 if (raw_hwp) {
1748 raw_hwp->page = page;
1749 llist_add(&raw_hwp->node, head);
1750 /* the first error event will be counted in action_result(). */
1751 if (ret)
1752 num_poisoned_pages_inc(page_to_pfn(page));
1753 } else {
1754 /*
1755 * Failed to save raw error info. We no longer trace all
1756 * hwpoisoned subpages, and we need refuse to free/dissolve
1757 * this hwpoisoned hugepage.
1758 */
1759 SetHPageRawHwpUnreliable(hpage);
1760 /*
1761 * Once HPageRawHwpUnreliable is set, raw_hwp_page is not
1762 * used any more, so free it.
1763 */
1764 __free_raw_hwp_pages(hpage, false);
1765 }
1766 return ret;
1767}
1768
1769static unsigned long free_raw_hwp_pages(struct page *hpage, bool move_flag)
1770{
1771 /*
1772 * HPageVmemmapOptimized hugepages can't be freed because struct
1773 * pages for tail pages are required but they don't exist.
1774 */
1775 if (move_flag && HPageVmemmapOptimized(hpage))
1776 return 0;
1777
1778 /*
1779 * HPageRawHwpUnreliable hugepages shouldn't be unpoisoned by
1780 * definition.
1781 */
1782 if (HPageRawHwpUnreliable(hpage))
1783 return 0;
1784
1785 return __free_raw_hwp_pages(hpage, move_flag);
1786}
1787
1788void hugetlb_clear_page_hwpoison(struct page *hpage)
1789{
1790 if (HPageRawHwpUnreliable(hpage))
1791 return;
1792 ClearPageHWPoison(hpage);
1793 free_raw_hwp_pages(hpage, true);
1794}
1795
1796/*
1797 * Called from hugetlb code with hugetlb_lock held.
1798 *
1799 * Return values:
1800 * 0 - free hugepage
1801 * 1 - in-use hugepage
1802 * 2 - not a hugepage
1803 * -EBUSY - the hugepage is busy (try to retry)
1804 * -EHWPOISON - the hugepage is already hwpoisoned
1805 */
1806int __get_huge_page_for_hwpoison(unsigned long pfn, int flags,
1807 bool *migratable_cleared)
1808{
1809 struct page *page = pfn_to_page(pfn);
1810 struct page *head = compound_head(page);
1811 int ret = 2; /* fallback to normal page handling */
1812 bool count_increased = false;
1813
1814 if (!PageHeadHuge(head))
1815 goto out;
1816
1817 if (flags & MF_COUNT_INCREASED) {
1818 ret = 1;
1819 count_increased = true;
1820 } else if (HPageFreed(head)) {
1821 ret = 0;
1822 } else if (HPageMigratable(head)) {
1823 ret = get_page_unless_zero(head);
1824 if (ret)
1825 count_increased = true;
1826 } else {
1827 ret = -EBUSY;
1828 if (!(flags & MF_NO_RETRY))
1829 goto out;
1830 }
1831
1832 if (hugetlb_set_page_hwpoison(head, page)) {
1833 ret = -EHWPOISON;
1834 goto out;
1835 }
1836
1837 /*
1838 * Clearing HPageMigratable for hwpoisoned hugepages to prevent them
1839 * from being migrated by memory hotremove.
1840 */
1841 if (count_increased && HPageMigratable(head)) {
1842 ClearHPageMigratable(head);
1843 *migratable_cleared = true;
1844 }
1845
1846 return ret;
1847out:
1848 if (count_increased)
1849 put_page(head);
1850 return ret;
1851}
1852
1853/*
1854 * Taking refcount of hugetlb pages needs extra care about race conditions
1855 * with basic operations like hugepage allocation/free/demotion.
1856 * So some of prechecks for hwpoison (pinning, and testing/setting
1857 * PageHWPoison) should be done in single hugetlb_lock range.
1858 */
1859static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb)
1860{
1861 int res;
1862 struct page *p = pfn_to_page(pfn);
1863 struct page *head;
1864 unsigned long page_flags;
1865 bool migratable_cleared = false;
1866
1867 *hugetlb = 1;
1868retry:
1869 res = get_huge_page_for_hwpoison(pfn, flags, &migratable_cleared);
1870 if (res == 2) { /* fallback to normal page handling */
1871 *hugetlb = 0;
1872 return 0;
1873 } else if (res == -EHWPOISON) {
1874 pr_err("%#lx: already hardware poisoned\n", pfn);
1875 if (flags & MF_ACTION_REQUIRED) {
1876 head = compound_head(p);
1877 res = kill_accessing_process(current, page_to_pfn(head), flags);
1878 }
1879 return res;
1880 } else if (res == -EBUSY) {
1881 if (!(flags & MF_NO_RETRY)) {
1882 flags |= MF_NO_RETRY;
1883 goto retry;
1884 }
1885 return action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
1886 }
1887
1888 head = compound_head(p);
1889 lock_page(head);
1890
1891 if (hwpoison_filter(p)) {
1892 hugetlb_clear_page_hwpoison(head);
1893 if (migratable_cleared)
1894 SetHPageMigratable(head);
1895 unlock_page(head);
1896 if (res == 1)
1897 put_page(head);
1898 return -EOPNOTSUPP;
1899 }
1900
1901 /*
1902 * Handling free hugepage. The possible race with hugepage allocation
1903 * or demotion can be prevented by PageHWPoison flag.
1904 */
1905 if (res == 0) {
1906 unlock_page(head);
1907 if (__page_handle_poison(p) >= 0) {
1908 page_ref_inc(p);
1909 res = MF_RECOVERED;
1910 } else {
1911 res = MF_FAILED;
1912 }
1913 return action_result(pfn, MF_MSG_FREE_HUGE, res);
1914 }
1915
1916 page_flags = head->flags;
1917
1918 if (!hwpoison_user_mappings(p, pfn, flags, head)) {
1919 unlock_page(head);
1920 return action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1921 }
1922
1923 return identify_page_state(pfn, p, page_flags);
1924}
1925
1926#else
1927static inline int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb)
1928{
1929 return 0;
1930}
1931
1932static inline unsigned long free_raw_hwp_pages(struct page *hpage, bool flag)
1933{
1934 return 0;
1935}
1936#endif /* CONFIG_HUGETLB_PAGE */
1937
1938/* Drop the extra refcount in case we come from madvise() */
1939static void put_ref_page(unsigned long pfn, int flags)
1940{
1941 struct page *page;
1942
1943 if (!(flags & MF_COUNT_INCREASED))
1944 return;
1945
1946 page = pfn_to_page(pfn);
1947 if (page)
1948 put_page(page);
1949}
1950
1951static int memory_failure_dev_pagemap(unsigned long pfn, int flags,
1952 struct dev_pagemap *pgmap)
1953{
1954 int rc = -ENXIO;
1955
1956 put_ref_page(pfn, flags);
1957
1958 /* device metadata space is not recoverable */
1959 if (!pgmap_pfn_valid(pgmap, pfn))
1960 goto out;
1961
1962 /*
1963 * Call driver's implementation to handle the memory failure, otherwise
1964 * fall back to generic handler.
1965 */
1966 if (pgmap_has_memory_failure(pgmap)) {
1967 rc = pgmap->ops->memory_failure(pgmap, pfn, 1, flags);
1968 /*
1969 * Fall back to generic handler too if operation is not
1970 * supported inside the driver/device/filesystem.
1971 */
1972 if (rc != -EOPNOTSUPP)
1973 goto out;
1974 }
1975
1976 rc = mf_generic_kill_procs(pfn, flags, pgmap);
1977out:
1978 /* drop pgmap ref acquired in caller */
1979 put_dev_pagemap(pgmap);
1980 action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED);
1981 return rc;
1982}
1983
1984static DEFINE_MUTEX(mf_mutex);
1985
1986/**
1987 * memory_failure - Handle memory failure of a page.
1988 * @pfn: Page Number of the corrupted page
1989 * @flags: fine tune action taken
1990 *
1991 * This function is called by the low level machine check code
1992 * of an architecture when it detects hardware memory corruption
1993 * of a page. It tries its best to recover, which includes
1994 * dropping pages, killing processes etc.
1995 *
1996 * The function is primarily of use for corruptions that
1997 * happen outside the current execution context (e.g. when
1998 * detected by a background scrubber)
1999 *
2000 * Must run in process context (e.g. a work queue) with interrupts
2001 * enabled and no spinlocks hold.
2002 *
2003 * Return: 0 for successfully handled the memory error,
2004 * -EOPNOTSUPP for hwpoison_filter() filtered the error event,
2005 * < 0(except -EOPNOTSUPP) on failure.
2006 */
2007int memory_failure(unsigned long pfn, int flags)
2008{
2009 struct page *p;
2010 struct page *hpage;
2011 struct dev_pagemap *pgmap;
2012 int res = 0;
2013 unsigned long page_flags;
2014 bool retry = true;
2015 int hugetlb = 0;
2016
2017 if (!sysctl_memory_failure_recovery)
2018 panic("Memory failure on page %lx", pfn);
2019
2020 mutex_lock(&mf_mutex);
2021
2022 if (!(flags & MF_SW_SIMULATED))
2023 hw_memory_failure = true;
2024
2025 p = pfn_to_online_page(pfn);
2026 if (!p) {
2027 res = arch_memory_failure(pfn, flags);
2028 if (res == 0)
2029 goto unlock_mutex;
2030
2031 if (pfn_valid(pfn)) {
2032 pgmap = get_dev_pagemap(pfn, NULL);
2033 if (pgmap) {
2034 res = memory_failure_dev_pagemap(pfn, flags,
2035 pgmap);
2036 goto unlock_mutex;
2037 }
2038 }
2039 pr_err("%#lx: memory outside kernel control\n", pfn);
2040 res = -ENXIO;
2041 goto unlock_mutex;
2042 }
2043
2044try_again:
2045 res = try_memory_failure_hugetlb(pfn, flags, &hugetlb);
2046 if (hugetlb)
2047 goto unlock_mutex;
2048
2049 if (TestSetPageHWPoison(p)) {
2050 pr_err("%#lx: already hardware poisoned\n", pfn);
2051 res = -EHWPOISON;
2052 if (flags & MF_ACTION_REQUIRED)
2053 res = kill_accessing_process(current, pfn, flags);
2054 if (flags & MF_COUNT_INCREASED)
2055 put_page(p);
2056 goto unlock_mutex;
2057 }
2058
2059 hpage = compound_head(p);
2060
2061 /*
2062 * We need/can do nothing about count=0 pages.
2063 * 1) it's a free page, and therefore in safe hand:
2064 * check_new_page() will be the gate keeper.
2065 * 2) it's part of a non-compound high order page.
2066 * Implies some kernel user: cannot stop them from
2067 * R/W the page; let's pray that the page has been
2068 * used and will be freed some time later.
2069 * In fact it's dangerous to directly bump up page count from 0,
2070 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
2071 */
2072 if (!(flags & MF_COUNT_INCREASED)) {
2073 res = get_hwpoison_page(p, flags);
2074 if (!res) {
2075 if (is_free_buddy_page(p)) {
2076 if (take_page_off_buddy(p)) {
2077 page_ref_inc(p);
2078 res = MF_RECOVERED;
2079 } else {
2080 /* We lost the race, try again */
2081 if (retry) {
2082 ClearPageHWPoison(p);
2083 retry = false;
2084 goto try_again;
2085 }
2086 res = MF_FAILED;
2087 }
2088 res = action_result(pfn, MF_MSG_BUDDY, res);
2089 } else {
2090 res = action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
2091 }
2092 goto unlock_mutex;
2093 } else if (res < 0) {
2094 res = action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
2095 goto unlock_mutex;
2096 }
2097 }
2098
2099 if (PageTransHuge(hpage)) {
2100 /*
2101 * The flag must be set after the refcount is bumped
2102 * otherwise it may race with THP split.
2103 * And the flag can't be set in get_hwpoison_page() since
2104 * it is called by soft offline too and it is just called
2105 * for !MF_COUNT_INCREASE. So here seems to be the best
2106 * place.
2107 *
2108 * Don't need care about the above error handling paths for
2109 * get_hwpoison_page() since they handle either free page
2110 * or unhandlable page. The refcount is bumped iff the
2111 * page is a valid handlable page.
2112 */
2113 SetPageHasHWPoisoned(hpage);
2114 if (try_to_split_thp_page(p) < 0) {
2115 res = action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED);
2116 goto unlock_mutex;
2117 }
2118 VM_BUG_ON_PAGE(!page_count(p), p);
2119 }
2120
2121 /*
2122 * We ignore non-LRU pages for good reasons.
2123 * - PG_locked is only well defined for LRU pages and a few others
2124 * - to avoid races with __SetPageLocked()
2125 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
2126 * The check (unnecessarily) ignores LRU pages being isolated and
2127 * walked by the page reclaim code, however that's not a big loss.
2128 */
2129 shake_page(p);
2130
2131 lock_page(p);
2132
2133 /*
2134 * We're only intended to deal with the non-Compound page here.
2135 * However, the page could have changed compound pages due to
2136 * race window. If this happens, we could try again to hopefully
2137 * handle the page next round.
2138 */
2139 if (PageCompound(p)) {
2140 if (retry) {
2141 ClearPageHWPoison(p);
2142 unlock_page(p);
2143 put_page(p);
2144 flags &= ~MF_COUNT_INCREASED;
2145 retry = false;
2146 goto try_again;
2147 }
2148 res = action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
2149 goto unlock_page;
2150 }
2151
2152 /*
2153 * We use page flags to determine what action should be taken, but
2154 * the flags can be modified by the error containment action. One
2155 * example is an mlocked page, where PG_mlocked is cleared by
2156 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
2157 * correctly, we save a copy of the page flags at this time.
2158 */
2159 page_flags = p->flags;
2160
2161 if (hwpoison_filter(p)) {
2162 ClearPageHWPoison(p);
2163 unlock_page(p);
2164 put_page(p);
2165 res = -EOPNOTSUPP;
2166 goto unlock_mutex;
2167 }
2168
2169 /*
2170 * __munlock_pagevec may clear a writeback page's LRU flag without
2171 * page_lock. We need wait writeback completion for this page or it
2172 * may trigger vfs BUG while evict inode.
2173 */
2174 if (!PageLRU(p) && !PageWriteback(p))
2175 goto identify_page_state;
2176
2177 /*
2178 * It's very difficult to mess with pages currently under IO
2179 * and in many cases impossible, so we just avoid it here.
2180 */
2181 wait_on_page_writeback(p);
2182
2183 /*
2184 * Now take care of user space mappings.
2185 * Abort on fail: __filemap_remove_folio() assumes unmapped page.
2186 */
2187 if (!hwpoison_user_mappings(p, pfn, flags, p)) {
2188 res = action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
2189 goto unlock_page;
2190 }
2191
2192 /*
2193 * Torn down by someone else?
2194 */
2195 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
2196 res = action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
2197 goto unlock_page;
2198 }
2199
2200identify_page_state:
2201 res = identify_page_state(pfn, p, page_flags);
2202 mutex_unlock(&mf_mutex);
2203 return res;
2204unlock_page:
2205 unlock_page(p);
2206unlock_mutex:
2207 mutex_unlock(&mf_mutex);
2208 return res;
2209}
2210EXPORT_SYMBOL_GPL(memory_failure);
2211
2212#define MEMORY_FAILURE_FIFO_ORDER 4
2213#define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
2214
2215struct memory_failure_entry {
2216 unsigned long pfn;
2217 int flags;
2218};
2219
2220struct memory_failure_cpu {
2221 DECLARE_KFIFO(fifo, struct memory_failure_entry,
2222 MEMORY_FAILURE_FIFO_SIZE);
2223 spinlock_t lock;
2224 struct work_struct work;
2225};
2226
2227static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
2228
2229/**
2230 * memory_failure_queue - Schedule handling memory failure of a page.
2231 * @pfn: Page Number of the corrupted page
2232 * @flags: Flags for memory failure handling
2233 *
2234 * This function is called by the low level hardware error handler
2235 * when it detects hardware memory corruption of a page. It schedules
2236 * the recovering of error page, including dropping pages, killing
2237 * processes etc.
2238 *
2239 * The function is primarily of use for corruptions that
2240 * happen outside the current execution context (e.g. when
2241 * detected by a background scrubber)
2242 *
2243 * Can run in IRQ context.
2244 */
2245void memory_failure_queue(unsigned long pfn, int flags)
2246{
2247 struct memory_failure_cpu *mf_cpu;
2248 unsigned long proc_flags;
2249 struct memory_failure_entry entry = {
2250 .pfn = pfn,
2251 .flags = flags,
2252 };
2253
2254 mf_cpu = &get_cpu_var(memory_failure_cpu);
2255 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
2256 if (kfifo_put(&mf_cpu->fifo, entry))
2257 schedule_work_on(smp_processor_id(), &mf_cpu->work);
2258 else
2259 pr_err("buffer overflow when queuing memory failure at %#lx\n",
2260 pfn);
2261 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
2262 put_cpu_var(memory_failure_cpu);
2263}
2264EXPORT_SYMBOL_GPL(memory_failure_queue);
2265
2266static void memory_failure_work_func(struct work_struct *work)
2267{
2268 struct memory_failure_cpu *mf_cpu;
2269 struct memory_failure_entry entry = { 0, };
2270 unsigned long proc_flags;
2271 int gotten;
2272
2273 mf_cpu = container_of(work, struct memory_failure_cpu, work);
2274 for (;;) {
2275 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
2276 gotten = kfifo_get(&mf_cpu->fifo, &entry);
2277 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
2278 if (!gotten)
2279 break;
2280 if (entry.flags & MF_SOFT_OFFLINE)
2281 soft_offline_page(entry.pfn, entry.flags);
2282 else
2283 memory_failure(entry.pfn, entry.flags);
2284 }
2285}
2286
2287/*
2288 * Process memory_failure work queued on the specified CPU.
2289 * Used to avoid return-to-userspace racing with the memory_failure workqueue.
2290 */
2291void memory_failure_queue_kick(int cpu)
2292{
2293 struct memory_failure_cpu *mf_cpu;
2294
2295 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
2296 cancel_work_sync(&mf_cpu->work);
2297 memory_failure_work_func(&mf_cpu->work);
2298}
2299
2300static int __init memory_failure_init(void)
2301{
2302 struct memory_failure_cpu *mf_cpu;
2303 int cpu;
2304
2305 for_each_possible_cpu(cpu) {
2306 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
2307 spin_lock_init(&mf_cpu->lock);
2308 INIT_KFIFO(mf_cpu->fifo);
2309 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
2310 }
2311
2312 return 0;
2313}
2314core_initcall(memory_failure_init);
2315
2316#undef pr_fmt
2317#define pr_fmt(fmt) "" fmt
2318#define unpoison_pr_info(fmt, pfn, rs) \
2319({ \
2320 if (__ratelimit(rs)) \
2321 pr_info(fmt, pfn); \
2322})
2323
2324/**
2325 * unpoison_memory - Unpoison a previously poisoned page
2326 * @pfn: Page number of the to be unpoisoned page
2327 *
2328 * Software-unpoison a page that has been poisoned by
2329 * memory_failure() earlier.
2330 *
2331 * This is only done on the software-level, so it only works
2332 * for linux injected failures, not real hardware failures
2333 *
2334 * Returns 0 for success, otherwise -errno.
2335 */
2336int unpoison_memory(unsigned long pfn)
2337{
2338 struct page *page;
2339 struct page *p;
2340 int ret = -EBUSY;
2341 unsigned long count = 1;
2342 bool huge = false;
2343 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
2344 DEFAULT_RATELIMIT_BURST);
2345
2346 if (!pfn_valid(pfn))
2347 return -ENXIO;
2348
2349 p = pfn_to_page(pfn);
2350 page = compound_head(p);
2351
2352 mutex_lock(&mf_mutex);
2353
2354 if (hw_memory_failure) {
2355 unpoison_pr_info("Unpoison: Disabled after HW memory failure %#lx\n",
2356 pfn, &unpoison_rs);
2357 ret = -EOPNOTSUPP;
2358 goto unlock_mutex;
2359 }
2360
2361 if (!PageHWPoison(p)) {
2362 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
2363 pfn, &unpoison_rs);
2364 goto unlock_mutex;
2365 }
2366
2367 if (page_count(page) > 1) {
2368 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
2369 pfn, &unpoison_rs);
2370 goto unlock_mutex;
2371 }
2372
2373 if (page_mapped(page)) {
2374 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
2375 pfn, &unpoison_rs);
2376 goto unlock_mutex;
2377 }
2378
2379 if (page_mapping(page)) {
2380 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
2381 pfn, &unpoison_rs);
2382 goto unlock_mutex;
2383 }
2384
2385 if (PageSlab(page) || PageTable(page) || PageReserved(page))
2386 goto unlock_mutex;
2387
2388 ret = get_hwpoison_page(p, MF_UNPOISON);
2389 if (!ret) {
2390 if (PageHuge(p)) {
2391 huge = true;
2392 count = free_raw_hwp_pages(page, false);
2393 if (count == 0) {
2394 ret = -EBUSY;
2395 goto unlock_mutex;
2396 }
2397 }
2398 ret = TestClearPageHWPoison(page) ? 0 : -EBUSY;
2399 } else if (ret < 0) {
2400 if (ret == -EHWPOISON) {
2401 ret = put_page_back_buddy(p) ? 0 : -EBUSY;
2402 } else
2403 unpoison_pr_info("Unpoison: failed to grab page %#lx\n",
2404 pfn, &unpoison_rs);
2405 } else {
2406 if (PageHuge(p)) {
2407 huge = true;
2408 count = free_raw_hwp_pages(page, false);
2409 if (count == 0) {
2410 ret = -EBUSY;
2411 put_page(page);
2412 goto unlock_mutex;
2413 }
2414 }
2415
2416 put_page(page);
2417 if (TestClearPageHWPoison(p)) {
2418 put_page(page);
2419 ret = 0;
2420 }
2421 }
2422
2423unlock_mutex:
2424 mutex_unlock(&mf_mutex);
2425 if (!ret) {
2426 if (!huge)
2427 num_poisoned_pages_sub(pfn, 1);
2428 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
2429 page_to_pfn(p), &unpoison_rs);
2430 }
2431 return ret;
2432}
2433EXPORT_SYMBOL(unpoison_memory);
2434
2435static bool isolate_page(struct page *page, struct list_head *pagelist)
2436{
2437 bool isolated = false;
2438
2439 if (PageHuge(page)) {
2440 isolated = !isolate_hugetlb(page, pagelist);
2441 } else {
2442 bool lru = !__PageMovable(page);
2443
2444 if (lru)
2445 isolated = !isolate_lru_page(page);
2446 else
2447 isolated = !isolate_movable_page(page,
2448 ISOLATE_UNEVICTABLE);
2449
2450 if (isolated) {
2451 list_add(&page->lru, pagelist);
2452 if (lru)
2453 inc_node_page_state(page, NR_ISOLATED_ANON +
2454 page_is_file_lru(page));
2455 }
2456 }
2457
2458 /*
2459 * If we succeed to isolate the page, we grabbed another refcount on
2460 * the page, so we can safely drop the one we got from get_any_pages().
2461 * If we failed to isolate the page, it means that we cannot go further
2462 * and we will return an error, so drop the reference we got from
2463 * get_any_pages() as well.
2464 */
2465 put_page(page);
2466 return isolated;
2467}
2468
2469/*
2470 * soft_offline_in_use_page handles hugetlb-pages and non-hugetlb pages.
2471 * If the page is a non-dirty unmapped page-cache page, it simply invalidates.
2472 * If the page is mapped, it migrates the contents over.
2473 */
2474static int soft_offline_in_use_page(struct page *page)
2475{
2476 long ret = 0;
2477 unsigned long pfn = page_to_pfn(page);
2478 struct page *hpage = compound_head(page);
2479 char const *msg_page[] = {"page", "hugepage"};
2480 bool huge = PageHuge(page);
2481 LIST_HEAD(pagelist);
2482 struct migration_target_control mtc = {
2483 .nid = NUMA_NO_NODE,
2484 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
2485 };
2486
2487 if (!huge && PageTransHuge(hpage)) {
2488 if (try_to_split_thp_page(page)) {
2489 pr_info("soft offline: %#lx: thp split failed\n", pfn);
2490 return -EBUSY;
2491 }
2492 hpage = page;
2493 }
2494
2495 lock_page(page);
2496 if (!PageHuge(page))
2497 wait_on_page_writeback(page);
2498 if (PageHWPoison(page)) {
2499 unlock_page(page);
2500 put_page(page);
2501 pr_info("soft offline: %#lx page already poisoned\n", pfn);
2502 return 0;
2503 }
2504
2505 if (!PageHuge(page) && PageLRU(page) && !PageSwapCache(page))
2506 /*
2507 * Try to invalidate first. This should work for
2508 * non dirty unmapped page cache pages.
2509 */
2510 ret = invalidate_inode_page(page);
2511 unlock_page(page);
2512
2513 if (ret) {
2514 pr_info("soft_offline: %#lx: invalidated\n", pfn);
2515 page_handle_poison(page, false, true);
2516 return 0;
2517 }
2518
2519 if (isolate_page(hpage, &pagelist)) {
2520 ret = migrate_pages(&pagelist, alloc_migration_target, NULL,
2521 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE, NULL);
2522 if (!ret) {
2523 bool release = !huge;
2524
2525 if (!page_handle_poison(page, huge, release))
2526 ret = -EBUSY;
2527 } else {
2528 if (!list_empty(&pagelist))
2529 putback_movable_pages(&pagelist);
2530
2531 pr_info("soft offline: %#lx: %s migration failed %ld, type %pGp\n",
2532 pfn, msg_page[huge], ret, &page->flags);
2533 if (ret > 0)
2534 ret = -EBUSY;
2535 }
2536 } else {
2537 pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %pGp\n",
2538 pfn, msg_page[huge], page_count(page), &page->flags);
2539 ret = -EBUSY;
2540 }
2541 return ret;
2542}
2543
2544/**
2545 * soft_offline_page - Soft offline a page.
2546 * @pfn: pfn to soft-offline
2547 * @flags: flags. Same as memory_failure().
2548 *
2549 * Returns 0 on success
2550 * -EOPNOTSUPP for hwpoison_filter() filtered the error event
2551 * < 0 otherwise negated errno.
2552 *
2553 * Soft offline a page, by migration or invalidation,
2554 * without killing anything. This is for the case when
2555 * a page is not corrupted yet (so it's still valid to access),
2556 * but has had a number of corrected errors and is better taken
2557 * out.
2558 *
2559 * The actual policy on when to do that is maintained by
2560 * user space.
2561 *
2562 * This should never impact any application or cause data loss,
2563 * however it might take some time.
2564 *
2565 * This is not a 100% solution for all memory, but tries to be
2566 * ``good enough'' for the majority of memory.
2567 */
2568int soft_offline_page(unsigned long pfn, int flags)
2569{
2570 int ret;
2571 bool try_again = true;
2572 struct page *page;
2573
2574 if (!pfn_valid(pfn)) {
2575 WARN_ON_ONCE(flags & MF_COUNT_INCREASED);
2576 return -ENXIO;
2577 }
2578
2579 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
2580 page = pfn_to_online_page(pfn);
2581 if (!page) {
2582 put_ref_page(pfn, flags);
2583 return -EIO;
2584 }
2585
2586 mutex_lock(&mf_mutex);
2587
2588 if (PageHWPoison(page)) {
2589 pr_info("%s: %#lx page already poisoned\n", __func__, pfn);
2590 put_ref_page(pfn, flags);
2591 mutex_unlock(&mf_mutex);
2592 return 0;
2593 }
2594
2595retry:
2596 get_online_mems();
2597 ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE);
2598 put_online_mems();
2599
2600 if (hwpoison_filter(page)) {
2601 if (ret > 0)
2602 put_page(page);
2603
2604 mutex_unlock(&mf_mutex);
2605 return -EOPNOTSUPP;
2606 }
2607
2608 if (ret > 0) {
2609 ret = soft_offline_in_use_page(page);
2610 } else if (ret == 0) {
2611 if (!page_handle_poison(page, true, false) && try_again) {
2612 try_again = false;
2613 flags &= ~MF_COUNT_INCREASED;
2614 goto retry;
2615 }
2616 }
2617
2618 mutex_unlock(&mf_mutex);
2619
2620 return ret;
2621}