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 * It can be very tempting to add handling for obscure cases here.
25 * In general any code for handling new cases should only be added iff:
26 * - You know how to test it.
27 * - You have a test that can be added to mce-test
28 * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/
29 * - The case actually shows up as a frequent (top 10) page state in
30 * tools/vm/page-types when running a real workload.
31 *
32 * There are several operations here with exponential complexity because
33 * of unsuitable VM data structures. For example the operation to map back
34 * from RMAP chains to processes has to walk the complete process list and
35 * has non linear complexity with the number. But since memory corruptions
36 * are rare we hope to get away with this. This avoids impacting the core
37 * VM.
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/ksm.h>
46#include <linux/rmap.h>
47#include <linux/export.h>
48#include <linux/pagemap.h>
49#include <linux/swap.h>
50#include <linux/backing-dev.h>
51#include <linux/migrate.h>
52#include <linux/suspend.h>
53#include <linux/slab.h>
54#include <linux/swapops.h>
55#include <linux/hugetlb.h>
56#include <linux/memory_hotplug.h>
57#include <linux/mm_inline.h>
58#include <linux/memremap.h>
59#include <linux/kfifo.h>
60#include <linux/ratelimit.h>
61#include <linux/page-isolation.h>
62#include "internal.h"
63#include "ras/ras_event.h"
64
65int sysctl_memory_failure_early_kill __read_mostly = 0;
66
67int sysctl_memory_failure_recovery __read_mostly = 1;
68
69atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
70
71#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
72
73u32 hwpoison_filter_enable = 0;
74u32 hwpoison_filter_dev_major = ~0U;
75u32 hwpoison_filter_dev_minor = ~0U;
76u64 hwpoison_filter_flags_mask;
77u64 hwpoison_filter_flags_value;
78EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
79EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
80EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
81EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
82EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
83
84static int hwpoison_filter_dev(struct page *p)
85{
86 struct address_space *mapping;
87 dev_t dev;
88
89 if (hwpoison_filter_dev_major == ~0U &&
90 hwpoison_filter_dev_minor == ~0U)
91 return 0;
92
93 /*
94 * page_mapping() does not accept slab pages.
95 */
96 if (PageSlab(p))
97 return -EINVAL;
98
99 mapping = page_mapping(p);
100 if (mapping == NULL || mapping->host == NULL)
101 return -EINVAL;
102
103 dev = mapping->host->i_sb->s_dev;
104 if (hwpoison_filter_dev_major != ~0U &&
105 hwpoison_filter_dev_major != MAJOR(dev))
106 return -EINVAL;
107 if (hwpoison_filter_dev_minor != ~0U &&
108 hwpoison_filter_dev_minor != MINOR(dev))
109 return -EINVAL;
110
111 return 0;
112}
113
114static int hwpoison_filter_flags(struct page *p)
115{
116 if (!hwpoison_filter_flags_mask)
117 return 0;
118
119 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
120 hwpoison_filter_flags_value)
121 return 0;
122 else
123 return -EINVAL;
124}
125
126/*
127 * This allows stress tests to limit test scope to a collection of tasks
128 * by putting them under some memcg. This prevents killing unrelated/important
129 * processes such as /sbin/init. Note that the target task may share clean
130 * pages with init (eg. libc text), which is harmless. If the target task
131 * share _dirty_ pages with another task B, the test scheme must make sure B
132 * is also included in the memcg. At last, due to race conditions this filter
133 * can only guarantee that the page either belongs to the memcg tasks, or is
134 * a freed page.
135 */
136#ifdef CONFIG_MEMCG
137u64 hwpoison_filter_memcg;
138EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
139static int hwpoison_filter_task(struct page *p)
140{
141 if (!hwpoison_filter_memcg)
142 return 0;
143
144 if (page_cgroup_ino(p) != hwpoison_filter_memcg)
145 return -EINVAL;
146
147 return 0;
148}
149#else
150static int hwpoison_filter_task(struct page *p) { return 0; }
151#endif
152
153int hwpoison_filter(struct page *p)
154{
155 if (!hwpoison_filter_enable)
156 return 0;
157
158 if (hwpoison_filter_dev(p))
159 return -EINVAL;
160
161 if (hwpoison_filter_flags(p))
162 return -EINVAL;
163
164 if (hwpoison_filter_task(p))
165 return -EINVAL;
166
167 return 0;
168}
169#else
170int hwpoison_filter(struct page *p)
171{
172 return 0;
173}
174#endif
175
176EXPORT_SYMBOL_GPL(hwpoison_filter);
177
178/*
179 * Kill all processes that have a poisoned page mapped and then isolate
180 * the page.
181 *
182 * General strategy:
183 * Find all processes having the page mapped and kill them.
184 * But we keep a page reference around so that the page is not
185 * actually freed yet.
186 * Then stash the page away
187 *
188 * There's no convenient way to get back to mapped processes
189 * from the VMAs. So do a brute-force search over all
190 * running processes.
191 *
192 * Remember that machine checks are not common (or rather
193 * if they are common you have other problems), so this shouldn't
194 * be a performance issue.
195 *
196 * Also there are some races possible while we get from the
197 * error detection to actually handle it.
198 */
199
200struct to_kill {
201 struct list_head nd;
202 struct task_struct *tsk;
203 unsigned long addr;
204 short size_shift;
205 char addr_valid;
206};
207
208/*
209 * Send all the processes who have the page mapped a signal.
210 * ``action optional'' if they are not immediately affected by the error
211 * ``action required'' if error happened in current execution context
212 */
213static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
214{
215 struct task_struct *t = tk->tsk;
216 short addr_lsb = tk->size_shift;
217 int ret;
218
219 pr_err("Memory failure: %#lx: Killing %s:%d due to hardware memory corruption\n",
220 pfn, t->comm, t->pid);
221
222 if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
223 ret = force_sig_mceerr(BUS_MCEERR_AR, (void __user *)tk->addr,
224 addr_lsb, current);
225 } else {
226 /*
227 * Don't use force here, it's convenient if the signal
228 * can be temporarily blocked.
229 * This could cause a loop when the user sets SIGBUS
230 * to SIG_IGN, but hopefully no one will do that?
231 */
232 ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
233 addr_lsb, t); /* synchronous? */
234 }
235 if (ret < 0)
236 pr_info("Memory failure: Error sending signal to %s:%d: %d\n",
237 t->comm, t->pid, ret);
238 return ret;
239}
240
241/*
242 * When a unknown page type is encountered drain as many buffers as possible
243 * in the hope to turn the page into a LRU or free page, which we can handle.
244 */
245void shake_page(struct page *p, int access)
246{
247 if (PageHuge(p))
248 return;
249
250 if (!PageSlab(p)) {
251 lru_add_drain_all();
252 if (PageLRU(p))
253 return;
254 drain_all_pages(page_zone(p));
255 if (PageLRU(p) || is_free_buddy_page(p))
256 return;
257 }
258
259 /*
260 * Only call shrink_node_slabs here (which would also shrink
261 * other caches) if access is not potentially fatal.
262 */
263 if (access)
264 drop_slab_node(page_to_nid(p));
265}
266EXPORT_SYMBOL_GPL(shake_page);
267
268static unsigned long dev_pagemap_mapping_shift(struct page *page,
269 struct vm_area_struct *vma)
270{
271 unsigned long address = vma_address(page, vma);
272 pgd_t *pgd;
273 p4d_t *p4d;
274 pud_t *pud;
275 pmd_t *pmd;
276 pte_t *pte;
277
278 pgd = pgd_offset(vma->vm_mm, address);
279 if (!pgd_present(*pgd))
280 return 0;
281 p4d = p4d_offset(pgd, address);
282 if (!p4d_present(*p4d))
283 return 0;
284 pud = pud_offset(p4d, address);
285 if (!pud_present(*pud))
286 return 0;
287 if (pud_devmap(*pud))
288 return PUD_SHIFT;
289 pmd = pmd_offset(pud, address);
290 if (!pmd_present(*pmd))
291 return 0;
292 if (pmd_devmap(*pmd))
293 return PMD_SHIFT;
294 pte = pte_offset_map(pmd, address);
295 if (!pte_present(*pte))
296 return 0;
297 if (pte_devmap(*pte))
298 return PAGE_SHIFT;
299 return 0;
300}
301
302/*
303 * Failure handling: if we can't find or can't kill a process there's
304 * not much we can do. We just print a message and ignore otherwise.
305 */
306
307/*
308 * Schedule a process for later kill.
309 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
310 * TBD would GFP_NOIO be enough?
311 */
312static void add_to_kill(struct task_struct *tsk, struct page *p,
313 struct vm_area_struct *vma,
314 struct list_head *to_kill,
315 struct to_kill **tkc)
316{
317 struct to_kill *tk;
318
319 if (*tkc) {
320 tk = *tkc;
321 *tkc = NULL;
322 } else {
323 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
324 if (!tk) {
325 pr_err("Memory failure: Out of memory while machine check handling\n");
326 return;
327 }
328 }
329 tk->addr = page_address_in_vma(p, vma);
330 tk->addr_valid = 1;
331 if (is_zone_device_page(p))
332 tk->size_shift = dev_pagemap_mapping_shift(p, vma);
333 else
334 tk->size_shift = compound_order(compound_head(p)) + PAGE_SHIFT;
335
336 /*
337 * In theory we don't have to kill when the page was
338 * munmaped. But it could be also a mremap. Since that's
339 * likely very rare kill anyways just out of paranoia, but use
340 * a SIGKILL because the error is not contained anymore.
341 */
342 if (tk->addr == -EFAULT || tk->size_shift == 0) {
343 pr_info("Memory failure: Unable to find user space address %lx in %s\n",
344 page_to_pfn(p), tsk->comm);
345 tk->addr_valid = 0;
346 }
347 get_task_struct(tsk);
348 tk->tsk = tsk;
349 list_add_tail(&tk->nd, to_kill);
350}
351
352/*
353 * Kill the processes that have been collected earlier.
354 *
355 * Only do anything when DOIT is set, otherwise just free the list
356 * (this is used for clean pages which do not need killing)
357 * Also when FAIL is set do a force kill because something went
358 * wrong earlier.
359 */
360static void kill_procs(struct list_head *to_kill, int forcekill, bool fail,
361 unsigned long pfn, int flags)
362{
363 struct to_kill *tk, *next;
364
365 list_for_each_entry_safe (tk, next, to_kill, nd) {
366 if (forcekill) {
367 /*
368 * In case something went wrong with munmapping
369 * make sure the process doesn't catch the
370 * signal and then access the memory. Just kill it.
371 */
372 if (fail || tk->addr_valid == 0) {
373 pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
374 pfn, tk->tsk->comm, tk->tsk->pid);
375 do_send_sig_info(SIGKILL, SEND_SIG_PRIV,
376 tk->tsk, PIDTYPE_PID);
377 }
378
379 /*
380 * In theory the process could have mapped
381 * something else on the address in-between. We could
382 * check for that, but we need to tell the
383 * process anyways.
384 */
385 else if (kill_proc(tk, pfn, flags) < 0)
386 pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n",
387 pfn, tk->tsk->comm, tk->tsk->pid);
388 }
389 put_task_struct(tk->tsk);
390 kfree(tk);
391 }
392}
393
394/*
395 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
396 * on behalf of the thread group. Return task_struct of the (first found)
397 * dedicated thread if found, and return NULL otherwise.
398 *
399 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
400 * have to call rcu_read_lock/unlock() in this function.
401 */
402static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
403{
404 struct task_struct *t;
405
406 for_each_thread(tsk, t)
407 if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
408 return t;
409 return NULL;
410}
411
412/*
413 * Determine whether a given process is "early kill" process which expects
414 * to be signaled when some page under the process is hwpoisoned.
415 * Return task_struct of the dedicated thread (main thread unless explicitly
416 * specified) if the process is "early kill," and otherwise returns NULL.
417 */
418static struct task_struct *task_early_kill(struct task_struct *tsk,
419 int force_early)
420{
421 struct task_struct *t;
422 if (!tsk->mm)
423 return NULL;
424 if (force_early)
425 return tsk;
426 t = find_early_kill_thread(tsk);
427 if (t)
428 return t;
429 if (sysctl_memory_failure_early_kill)
430 return tsk;
431 return NULL;
432}
433
434/*
435 * Collect processes when the error hit an anonymous page.
436 */
437static void collect_procs_anon(struct page *page, struct list_head *to_kill,
438 struct to_kill **tkc, int force_early)
439{
440 struct vm_area_struct *vma;
441 struct task_struct *tsk;
442 struct anon_vma *av;
443 pgoff_t pgoff;
444
445 av = page_lock_anon_vma_read(page);
446 if (av == NULL) /* Not actually mapped anymore */
447 return;
448
449 pgoff = page_to_pgoff(page);
450 read_lock(&tasklist_lock);
451 for_each_process (tsk) {
452 struct anon_vma_chain *vmac;
453 struct task_struct *t = task_early_kill(tsk, force_early);
454
455 if (!t)
456 continue;
457 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
458 pgoff, pgoff) {
459 vma = vmac->vma;
460 if (!page_mapped_in_vma(page, vma))
461 continue;
462 if (vma->vm_mm == t->mm)
463 add_to_kill(t, page, vma, to_kill, tkc);
464 }
465 }
466 read_unlock(&tasklist_lock);
467 page_unlock_anon_vma_read(av);
468}
469
470/*
471 * Collect processes when the error hit a file mapped page.
472 */
473static void collect_procs_file(struct page *page, struct list_head *to_kill,
474 struct to_kill **tkc, int force_early)
475{
476 struct vm_area_struct *vma;
477 struct task_struct *tsk;
478 struct address_space *mapping = page->mapping;
479
480 i_mmap_lock_read(mapping);
481 read_lock(&tasklist_lock);
482 for_each_process(tsk) {
483 pgoff_t pgoff = page_to_pgoff(page);
484 struct task_struct *t = task_early_kill(tsk, force_early);
485
486 if (!t)
487 continue;
488 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
489 pgoff) {
490 /*
491 * Send early kill signal to tasks where a vma covers
492 * the page but the corrupted page is not necessarily
493 * mapped it in its pte.
494 * Assume applications who requested early kill want
495 * to be informed of all such data corruptions.
496 */
497 if (vma->vm_mm == t->mm)
498 add_to_kill(t, page, vma, to_kill, tkc);
499 }
500 }
501 read_unlock(&tasklist_lock);
502 i_mmap_unlock_read(mapping);
503}
504
505/*
506 * Collect the processes who have the corrupted page mapped to kill.
507 * This is done in two steps for locking reasons.
508 * First preallocate one tokill structure outside the spin locks,
509 * so that we can kill at least one process reasonably reliable.
510 */
511static void collect_procs(struct page *page, struct list_head *tokill,
512 int force_early)
513{
514 struct to_kill *tk;
515
516 if (!page->mapping)
517 return;
518
519 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
520 if (!tk)
521 return;
522 if (PageAnon(page))
523 collect_procs_anon(page, tokill, &tk, force_early);
524 else
525 collect_procs_file(page, tokill, &tk, force_early);
526 kfree(tk);
527}
528
529static const char *action_name[] = {
530 [MF_IGNORED] = "Ignored",
531 [MF_FAILED] = "Failed",
532 [MF_DELAYED] = "Delayed",
533 [MF_RECOVERED] = "Recovered",
534};
535
536static const char * const action_page_types[] = {
537 [MF_MSG_KERNEL] = "reserved kernel page",
538 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
539 [MF_MSG_SLAB] = "kernel slab page",
540 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
541 [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned",
542 [MF_MSG_HUGE] = "huge page",
543 [MF_MSG_FREE_HUGE] = "free huge page",
544 [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page",
545 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
546 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
547 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
548 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
549 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
550 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
551 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
552 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
553 [MF_MSG_CLEAN_LRU] = "clean LRU page",
554 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
555 [MF_MSG_BUDDY] = "free buddy page",
556 [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)",
557 [MF_MSG_DAX] = "dax page",
558 [MF_MSG_UNKNOWN] = "unknown page",
559};
560
561/*
562 * XXX: It is possible that a page is isolated from LRU cache,
563 * and then kept in swap cache or failed to remove from page cache.
564 * The page count will stop it from being freed by unpoison.
565 * Stress tests should be aware of this memory leak problem.
566 */
567static int delete_from_lru_cache(struct page *p)
568{
569 if (!isolate_lru_page(p)) {
570 /*
571 * Clear sensible page flags, so that the buddy system won't
572 * complain when the page is unpoison-and-freed.
573 */
574 ClearPageActive(p);
575 ClearPageUnevictable(p);
576
577 /*
578 * Poisoned page might never drop its ref count to 0 so we have
579 * to uncharge it manually from its memcg.
580 */
581 mem_cgroup_uncharge(p);
582
583 /*
584 * drop the page count elevated by isolate_lru_page()
585 */
586 put_page(p);
587 return 0;
588 }
589 return -EIO;
590}
591
592static int truncate_error_page(struct page *p, unsigned long pfn,
593 struct address_space *mapping)
594{
595 int ret = MF_FAILED;
596
597 if (mapping->a_ops->error_remove_page) {
598 int err = mapping->a_ops->error_remove_page(mapping, p);
599
600 if (err != 0) {
601 pr_info("Memory failure: %#lx: Failed to punch page: %d\n",
602 pfn, err);
603 } else if (page_has_private(p) &&
604 !try_to_release_page(p, GFP_NOIO)) {
605 pr_info("Memory failure: %#lx: failed to release buffers\n",
606 pfn);
607 } else {
608 ret = MF_RECOVERED;
609 }
610 } else {
611 /*
612 * If the file system doesn't support it just invalidate
613 * This fails on dirty or anything with private pages
614 */
615 if (invalidate_inode_page(p))
616 ret = MF_RECOVERED;
617 else
618 pr_info("Memory failure: %#lx: Failed to invalidate\n",
619 pfn);
620 }
621
622 return ret;
623}
624
625/*
626 * Error hit kernel page.
627 * Do nothing, try to be lucky and not touch this instead. For a few cases we
628 * could be more sophisticated.
629 */
630static int me_kernel(struct page *p, unsigned long pfn)
631{
632 return MF_IGNORED;
633}
634
635/*
636 * Page in unknown state. Do nothing.
637 */
638static int me_unknown(struct page *p, unsigned long pfn)
639{
640 pr_err("Memory failure: %#lx: Unknown page state\n", pfn);
641 return MF_FAILED;
642}
643
644/*
645 * Clean (or cleaned) page cache page.
646 */
647static int me_pagecache_clean(struct page *p, unsigned long pfn)
648{
649 struct address_space *mapping;
650
651 delete_from_lru_cache(p);
652
653 /*
654 * For anonymous pages we're done the only reference left
655 * should be the one m_f() holds.
656 */
657 if (PageAnon(p))
658 return MF_RECOVERED;
659
660 /*
661 * Now truncate the page in the page cache. This is really
662 * more like a "temporary hole punch"
663 * Don't do this for block devices when someone else
664 * has a reference, because it could be file system metadata
665 * and that's not safe to truncate.
666 */
667 mapping = page_mapping(p);
668 if (!mapping) {
669 /*
670 * Page has been teared down in the meanwhile
671 */
672 return MF_FAILED;
673 }
674
675 /*
676 * Truncation is a bit tricky. Enable it per file system for now.
677 *
678 * Open: to take i_mutex or not for this? Right now we don't.
679 */
680 return truncate_error_page(p, pfn, mapping);
681}
682
683/*
684 * Dirty pagecache page
685 * Issues: when the error hit a hole page the error is not properly
686 * propagated.
687 */
688static int me_pagecache_dirty(struct page *p, unsigned long pfn)
689{
690 struct address_space *mapping = page_mapping(p);
691
692 SetPageError(p);
693 /* TBD: print more information about the file. */
694 if (mapping) {
695 /*
696 * IO error will be reported by write(), fsync(), etc.
697 * who check the mapping.
698 * This way the application knows that something went
699 * wrong with its dirty file data.
700 *
701 * There's one open issue:
702 *
703 * The EIO will be only reported on the next IO
704 * operation and then cleared through the IO map.
705 * Normally Linux has two mechanisms to pass IO error
706 * first through the AS_EIO flag in the address space
707 * and then through the PageError flag in the page.
708 * Since we drop pages on memory failure handling the
709 * only mechanism open to use is through AS_AIO.
710 *
711 * This has the disadvantage that it gets cleared on
712 * the first operation that returns an error, while
713 * the PageError bit is more sticky and only cleared
714 * when the page is reread or dropped. If an
715 * application assumes it will always get error on
716 * fsync, but does other operations on the fd before
717 * and the page is dropped between then the error
718 * will not be properly reported.
719 *
720 * This can already happen even without hwpoisoned
721 * pages: first on metadata IO errors (which only
722 * report through AS_EIO) or when the page is dropped
723 * at the wrong time.
724 *
725 * So right now we assume that the application DTRT on
726 * the first EIO, but we're not worse than other parts
727 * of the kernel.
728 */
729 mapping_set_error(mapping, -EIO);
730 }
731
732 return me_pagecache_clean(p, pfn);
733}
734
735/*
736 * Clean and dirty swap cache.
737 *
738 * Dirty swap cache page is tricky to handle. The page could live both in page
739 * cache and swap cache(ie. page is freshly swapped in). So it could be
740 * referenced concurrently by 2 types of PTEs:
741 * normal PTEs and swap PTEs. We try to handle them consistently by calling
742 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
743 * and then
744 * - clear dirty bit to prevent IO
745 * - remove from LRU
746 * - but keep in the swap cache, so that when we return to it on
747 * a later page fault, we know the application is accessing
748 * corrupted data and shall be killed (we installed simple
749 * interception code in do_swap_page to catch it).
750 *
751 * Clean swap cache pages can be directly isolated. A later page fault will
752 * bring in the known good data from disk.
753 */
754static int me_swapcache_dirty(struct page *p, unsigned long pfn)
755{
756 ClearPageDirty(p);
757 /* Trigger EIO in shmem: */
758 ClearPageUptodate(p);
759
760 if (!delete_from_lru_cache(p))
761 return MF_DELAYED;
762 else
763 return MF_FAILED;
764}
765
766static int me_swapcache_clean(struct page *p, unsigned long pfn)
767{
768 delete_from_swap_cache(p);
769
770 if (!delete_from_lru_cache(p))
771 return MF_RECOVERED;
772 else
773 return MF_FAILED;
774}
775
776/*
777 * Huge pages. Needs work.
778 * Issues:
779 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
780 * To narrow down kill region to one page, we need to break up pmd.
781 */
782static int me_huge_page(struct page *p, unsigned long pfn)
783{
784 int res = 0;
785 struct page *hpage = compound_head(p);
786 struct address_space *mapping;
787
788 if (!PageHuge(hpage))
789 return MF_DELAYED;
790
791 mapping = page_mapping(hpage);
792 if (mapping) {
793 res = truncate_error_page(hpage, pfn, mapping);
794 } else {
795 unlock_page(hpage);
796 /*
797 * migration entry prevents later access on error anonymous
798 * hugepage, so we can free and dissolve it into buddy to
799 * save healthy subpages.
800 */
801 if (PageAnon(hpage))
802 put_page(hpage);
803 dissolve_free_huge_page(p);
804 res = MF_RECOVERED;
805 lock_page(hpage);
806 }
807
808 return res;
809}
810
811/*
812 * Various page states we can handle.
813 *
814 * A page state is defined by its current page->flags bits.
815 * The table matches them in order and calls the right handler.
816 *
817 * This is quite tricky because we can access page at any time
818 * in its live cycle, so all accesses have to be extremely careful.
819 *
820 * This is not complete. More states could be added.
821 * For any missing state don't attempt recovery.
822 */
823
824#define dirty (1UL << PG_dirty)
825#define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked))
826#define unevict (1UL << PG_unevictable)
827#define mlock (1UL << PG_mlocked)
828#define writeback (1UL << PG_writeback)
829#define lru (1UL << PG_lru)
830#define head (1UL << PG_head)
831#define slab (1UL << PG_slab)
832#define reserved (1UL << PG_reserved)
833
834static struct page_state {
835 unsigned long mask;
836 unsigned long res;
837 enum mf_action_page_type type;
838 int (*action)(struct page *p, unsigned long pfn);
839} error_states[] = {
840 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
841 /*
842 * free pages are specially detected outside this table:
843 * PG_buddy pages only make a small fraction of all free pages.
844 */
845
846 /*
847 * Could in theory check if slab page is free or if we can drop
848 * currently unused objects without touching them. But just
849 * treat it as standard kernel for now.
850 */
851 { slab, slab, MF_MSG_SLAB, me_kernel },
852
853 { head, head, MF_MSG_HUGE, me_huge_page },
854
855 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
856 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
857
858 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
859 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
860
861 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
862 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
863
864 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
865 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
866
867 /*
868 * Catchall entry: must be at end.
869 */
870 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
871};
872
873#undef dirty
874#undef sc
875#undef unevict
876#undef mlock
877#undef writeback
878#undef lru
879#undef head
880#undef slab
881#undef reserved
882
883/*
884 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
885 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
886 */
887static void action_result(unsigned long pfn, enum mf_action_page_type type,
888 enum mf_result result)
889{
890 trace_memory_failure_event(pfn, type, result);
891
892 pr_err("Memory failure: %#lx: recovery action for %s: %s\n",
893 pfn, action_page_types[type], action_name[result]);
894}
895
896static int page_action(struct page_state *ps, struct page *p,
897 unsigned long pfn)
898{
899 int result;
900 int count;
901
902 result = ps->action(p, pfn);
903
904 count = page_count(p) - 1;
905 if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
906 count--;
907 if (count > 0) {
908 pr_err("Memory failure: %#lx: %s still referenced by %d users\n",
909 pfn, action_page_types[ps->type], count);
910 result = MF_FAILED;
911 }
912 action_result(pfn, ps->type, result);
913
914 /* Could do more checks here if page looks ok */
915 /*
916 * Could adjust zone counters here to correct for the missing page.
917 */
918
919 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
920}
921
922/**
923 * get_hwpoison_page() - Get refcount for memory error handling:
924 * @page: raw error page (hit by memory error)
925 *
926 * Return: return 0 if failed to grab the refcount, otherwise true (some
927 * non-zero value.)
928 */
929int get_hwpoison_page(struct page *page)
930{
931 struct page *head = compound_head(page);
932
933 if (!PageHuge(head) && PageTransHuge(head)) {
934 /*
935 * Non anonymous thp exists only in allocation/free time. We
936 * can't handle such a case correctly, so let's give it up.
937 * This should be better than triggering BUG_ON when kernel
938 * tries to touch the "partially handled" page.
939 */
940 if (!PageAnon(head)) {
941 pr_err("Memory failure: %#lx: non anonymous thp\n",
942 page_to_pfn(page));
943 return 0;
944 }
945 }
946
947 if (get_page_unless_zero(head)) {
948 if (head == compound_head(page))
949 return 1;
950
951 pr_info("Memory failure: %#lx cannot catch tail\n",
952 page_to_pfn(page));
953 put_page(head);
954 }
955
956 return 0;
957}
958EXPORT_SYMBOL_GPL(get_hwpoison_page);
959
960/*
961 * Do all that is necessary to remove user space mappings. Unmap
962 * the pages and send SIGBUS to the processes if the data was dirty.
963 */
964static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
965 int flags, struct page **hpagep)
966{
967 enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
968 struct address_space *mapping;
969 LIST_HEAD(tokill);
970 bool unmap_success;
971 int kill = 1, forcekill;
972 struct page *hpage = *hpagep;
973 bool mlocked = PageMlocked(hpage);
974
975 /*
976 * Here we are interested only in user-mapped pages, so skip any
977 * other types of pages.
978 */
979 if (PageReserved(p) || PageSlab(p))
980 return true;
981 if (!(PageLRU(hpage) || PageHuge(p)))
982 return true;
983
984 /*
985 * This check implies we don't kill processes if their pages
986 * are in the swap cache early. Those are always late kills.
987 */
988 if (!page_mapped(hpage))
989 return true;
990
991 if (PageKsm(p)) {
992 pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn);
993 return false;
994 }
995
996 if (PageSwapCache(p)) {
997 pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n",
998 pfn);
999 ttu |= TTU_IGNORE_HWPOISON;
1000 }
1001
1002 /*
1003 * Propagate the dirty bit from PTEs to struct page first, because we
1004 * need this to decide if we should kill or just drop the page.
1005 * XXX: the dirty test could be racy: set_page_dirty() may not always
1006 * be called inside page lock (it's recommended but not enforced).
1007 */
1008 mapping = page_mapping(hpage);
1009 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
1010 mapping_cap_writeback_dirty(mapping)) {
1011 if (page_mkclean(hpage)) {
1012 SetPageDirty(hpage);
1013 } else {
1014 kill = 0;
1015 ttu |= TTU_IGNORE_HWPOISON;
1016 pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n",
1017 pfn);
1018 }
1019 }
1020
1021 /*
1022 * First collect all the processes that have the page
1023 * mapped in dirty form. This has to be done before try_to_unmap,
1024 * because ttu takes the rmap data structures down.
1025 *
1026 * Error handling: We ignore errors here because
1027 * there's nothing that can be done.
1028 */
1029 if (kill)
1030 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
1031
1032 unmap_success = try_to_unmap(hpage, ttu);
1033 if (!unmap_success)
1034 pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n",
1035 pfn, page_mapcount(hpage));
1036
1037 /*
1038 * try_to_unmap() might put mlocked page in lru cache, so call
1039 * shake_page() again to ensure that it's flushed.
1040 */
1041 if (mlocked)
1042 shake_page(hpage, 0);
1043
1044 /*
1045 * Now that the dirty bit has been propagated to the
1046 * struct page and all unmaps done we can decide if
1047 * killing is needed or not. Only kill when the page
1048 * was dirty or the process is not restartable,
1049 * otherwise the tokill list is merely
1050 * freed. When there was a problem unmapping earlier
1051 * use a more force-full uncatchable kill to prevent
1052 * any accesses to the poisoned memory.
1053 */
1054 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
1055 kill_procs(&tokill, forcekill, !unmap_success, pfn, flags);
1056
1057 return unmap_success;
1058}
1059
1060static int identify_page_state(unsigned long pfn, struct page *p,
1061 unsigned long page_flags)
1062{
1063 struct page_state *ps;
1064
1065 /*
1066 * The first check uses the current page flags which may not have any
1067 * relevant information. The second check with the saved page flags is
1068 * carried out only if the first check can't determine the page status.
1069 */
1070 for (ps = error_states;; ps++)
1071 if ((p->flags & ps->mask) == ps->res)
1072 break;
1073
1074 page_flags |= (p->flags & (1UL << PG_dirty));
1075
1076 if (!ps->mask)
1077 for (ps = error_states;; ps++)
1078 if ((page_flags & ps->mask) == ps->res)
1079 break;
1080 return page_action(ps, p, pfn);
1081}
1082
1083static int memory_failure_hugetlb(unsigned long pfn, int flags)
1084{
1085 struct page *p = pfn_to_page(pfn);
1086 struct page *head = compound_head(p);
1087 int res;
1088 unsigned long page_flags;
1089
1090 if (TestSetPageHWPoison(head)) {
1091 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1092 pfn);
1093 return 0;
1094 }
1095
1096 num_poisoned_pages_inc();
1097
1098 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
1099 /*
1100 * Check "filter hit" and "race with other subpage."
1101 */
1102 lock_page(head);
1103 if (PageHWPoison(head)) {
1104 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1105 || (p != head && TestSetPageHWPoison(head))) {
1106 num_poisoned_pages_dec();
1107 unlock_page(head);
1108 return 0;
1109 }
1110 }
1111 unlock_page(head);
1112 dissolve_free_huge_page(p);
1113 action_result(pfn, MF_MSG_FREE_HUGE, MF_DELAYED);
1114 return 0;
1115 }
1116
1117 lock_page(head);
1118 page_flags = head->flags;
1119
1120 if (!PageHWPoison(head)) {
1121 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1122 num_poisoned_pages_dec();
1123 unlock_page(head);
1124 put_hwpoison_page(head);
1125 return 0;
1126 }
1127
1128 /*
1129 * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so
1130 * simply disable it. In order to make it work properly, we need
1131 * make sure that:
1132 * - conversion of a pud that maps an error hugetlb into hwpoison
1133 * entry properly works, and
1134 * - other mm code walking over page table is aware of pud-aligned
1135 * hwpoison entries.
1136 */
1137 if (huge_page_size(page_hstate(head)) > PMD_SIZE) {
1138 action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED);
1139 res = -EBUSY;
1140 goto out;
1141 }
1142
1143 if (!hwpoison_user_mappings(p, pfn, flags, &head)) {
1144 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1145 res = -EBUSY;
1146 goto out;
1147 }
1148
1149 res = identify_page_state(pfn, p, page_flags);
1150out:
1151 unlock_page(head);
1152 return res;
1153}
1154
1155static int memory_failure_dev_pagemap(unsigned long pfn, int flags,
1156 struct dev_pagemap *pgmap)
1157{
1158 struct page *page = pfn_to_page(pfn);
1159 const bool unmap_success = true;
1160 unsigned long size = 0;
1161 struct to_kill *tk;
1162 LIST_HEAD(tokill);
1163 int rc = -EBUSY;
1164 loff_t start;
1165 dax_entry_t cookie;
1166
1167 /*
1168 * Prevent the inode from being freed while we are interrogating
1169 * the address_space, typically this would be handled by
1170 * lock_page(), but dax pages do not use the page lock. This
1171 * also prevents changes to the mapping of this pfn until
1172 * poison signaling is complete.
1173 */
1174 cookie = dax_lock_page(page);
1175 if (!cookie)
1176 goto out;
1177
1178 if (hwpoison_filter(page)) {
1179 rc = 0;
1180 goto unlock;
1181 }
1182
1183 switch (pgmap->type) {
1184 case MEMORY_DEVICE_PRIVATE:
1185 case MEMORY_DEVICE_PUBLIC:
1186 /*
1187 * TODO: Handle HMM pages which may need coordination
1188 * with device-side memory.
1189 */
1190 goto unlock;
1191 default:
1192 break;
1193 }
1194
1195 /*
1196 * Use this flag as an indication that the dax page has been
1197 * remapped UC to prevent speculative consumption of poison.
1198 */
1199 SetPageHWPoison(page);
1200
1201 /*
1202 * Unlike System-RAM there is no possibility to swap in a
1203 * different physical page at a given virtual address, so all
1204 * userspace consumption of ZONE_DEVICE memory necessitates
1205 * SIGBUS (i.e. MF_MUST_KILL)
1206 */
1207 flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1208 collect_procs(page, &tokill, flags & MF_ACTION_REQUIRED);
1209
1210 list_for_each_entry(tk, &tokill, nd)
1211 if (tk->size_shift)
1212 size = max(size, 1UL << tk->size_shift);
1213 if (size) {
1214 /*
1215 * Unmap the largest mapping to avoid breaking up
1216 * device-dax mappings which are constant size. The
1217 * actual size of the mapping being torn down is
1218 * communicated in siginfo, see kill_proc()
1219 */
1220 start = (page->index << PAGE_SHIFT) & ~(size - 1);
1221 unmap_mapping_range(page->mapping, start, start + size, 0);
1222 }
1223 kill_procs(&tokill, flags & MF_MUST_KILL, !unmap_success, pfn, flags);
1224 rc = 0;
1225unlock:
1226 dax_unlock_page(page, cookie);
1227out:
1228 /* drop pgmap ref acquired in caller */
1229 put_dev_pagemap(pgmap);
1230 action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED);
1231 return rc;
1232}
1233
1234/**
1235 * memory_failure - Handle memory failure of a page.
1236 * @pfn: Page Number of the corrupted page
1237 * @flags: fine tune action taken
1238 *
1239 * This function is called by the low level machine check code
1240 * of an architecture when it detects hardware memory corruption
1241 * of a page. It tries its best to recover, which includes
1242 * dropping pages, killing processes etc.
1243 *
1244 * The function is primarily of use for corruptions that
1245 * happen outside the current execution context (e.g. when
1246 * detected by a background scrubber)
1247 *
1248 * Must run in process context (e.g. a work queue) with interrupts
1249 * enabled and no spinlocks hold.
1250 */
1251int memory_failure(unsigned long pfn, int flags)
1252{
1253 struct page *p;
1254 struct page *hpage;
1255 struct page *orig_head;
1256 struct dev_pagemap *pgmap;
1257 int res;
1258 unsigned long page_flags;
1259
1260 if (!sysctl_memory_failure_recovery)
1261 panic("Memory failure on page %lx", pfn);
1262
1263 if (!pfn_valid(pfn)) {
1264 pr_err("Memory failure: %#lx: memory outside kernel control\n",
1265 pfn);
1266 return -ENXIO;
1267 }
1268
1269 pgmap = get_dev_pagemap(pfn, NULL);
1270 if (pgmap)
1271 return memory_failure_dev_pagemap(pfn, flags, pgmap);
1272
1273 p = pfn_to_page(pfn);
1274 if (PageHuge(p))
1275 return memory_failure_hugetlb(pfn, flags);
1276 if (TestSetPageHWPoison(p)) {
1277 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1278 pfn);
1279 return 0;
1280 }
1281
1282 orig_head = hpage = compound_head(p);
1283 num_poisoned_pages_inc();
1284
1285 /*
1286 * We need/can do nothing about count=0 pages.
1287 * 1) it's a free page, and therefore in safe hand:
1288 * prep_new_page() will be the gate keeper.
1289 * 2) it's part of a non-compound high order page.
1290 * Implies some kernel user: cannot stop them from
1291 * R/W the page; let's pray that the page has been
1292 * used and will be freed some time later.
1293 * In fact it's dangerous to directly bump up page count from 0,
1294 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
1295 */
1296 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
1297 if (is_free_buddy_page(p)) {
1298 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1299 return 0;
1300 } else {
1301 action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
1302 return -EBUSY;
1303 }
1304 }
1305
1306 if (PageTransHuge(hpage)) {
1307 lock_page(p);
1308 if (!PageAnon(p) || unlikely(split_huge_page(p))) {
1309 unlock_page(p);
1310 if (!PageAnon(p))
1311 pr_err("Memory failure: %#lx: non anonymous thp\n",
1312 pfn);
1313 else
1314 pr_err("Memory failure: %#lx: thp split failed\n",
1315 pfn);
1316 if (TestClearPageHWPoison(p))
1317 num_poisoned_pages_dec();
1318 put_hwpoison_page(p);
1319 return -EBUSY;
1320 }
1321 unlock_page(p);
1322 VM_BUG_ON_PAGE(!page_count(p), p);
1323 hpage = compound_head(p);
1324 }
1325
1326 /*
1327 * We ignore non-LRU pages for good reasons.
1328 * - PG_locked is only well defined for LRU pages and a few others
1329 * - to avoid races with __SetPageLocked()
1330 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1331 * The check (unnecessarily) ignores LRU pages being isolated and
1332 * walked by the page reclaim code, however that's not a big loss.
1333 */
1334 shake_page(p, 0);
1335 /* shake_page could have turned it free. */
1336 if (!PageLRU(p) && is_free_buddy_page(p)) {
1337 if (flags & MF_COUNT_INCREASED)
1338 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1339 else
1340 action_result(pfn, MF_MSG_BUDDY_2ND, MF_DELAYED);
1341 return 0;
1342 }
1343
1344 lock_page(p);
1345
1346 /*
1347 * The page could have changed compound pages during the locking.
1348 * If this happens just bail out.
1349 */
1350 if (PageCompound(p) && compound_head(p) != orig_head) {
1351 action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
1352 res = -EBUSY;
1353 goto out;
1354 }
1355
1356 /*
1357 * We use page flags to determine what action should be taken, but
1358 * the flags can be modified by the error containment action. One
1359 * example is an mlocked page, where PG_mlocked is cleared by
1360 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1361 * correctly, we save a copy of the page flags at this time.
1362 */
1363 if (PageHuge(p))
1364 page_flags = hpage->flags;
1365 else
1366 page_flags = p->flags;
1367
1368 /*
1369 * unpoison always clear PG_hwpoison inside page lock
1370 */
1371 if (!PageHWPoison(p)) {
1372 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1373 num_poisoned_pages_dec();
1374 unlock_page(p);
1375 put_hwpoison_page(p);
1376 return 0;
1377 }
1378 if (hwpoison_filter(p)) {
1379 if (TestClearPageHWPoison(p))
1380 num_poisoned_pages_dec();
1381 unlock_page(p);
1382 put_hwpoison_page(p);
1383 return 0;
1384 }
1385
1386 if (!PageTransTail(p) && !PageLRU(p))
1387 goto identify_page_state;
1388
1389 /*
1390 * It's very difficult to mess with pages currently under IO
1391 * and in many cases impossible, so we just avoid it here.
1392 */
1393 wait_on_page_writeback(p);
1394
1395 /*
1396 * Now take care of user space mappings.
1397 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1398 *
1399 * When the raw error page is thp tail page, hpage points to the raw
1400 * page after thp split.
1401 */
1402 if (!hwpoison_user_mappings(p, pfn, flags, &hpage)) {
1403 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1404 res = -EBUSY;
1405 goto out;
1406 }
1407
1408 /*
1409 * Torn down by someone else?
1410 */
1411 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1412 action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
1413 res = -EBUSY;
1414 goto out;
1415 }
1416
1417identify_page_state:
1418 res = identify_page_state(pfn, p, page_flags);
1419out:
1420 unlock_page(p);
1421 return res;
1422}
1423EXPORT_SYMBOL_GPL(memory_failure);
1424
1425#define MEMORY_FAILURE_FIFO_ORDER 4
1426#define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1427
1428struct memory_failure_entry {
1429 unsigned long pfn;
1430 int flags;
1431};
1432
1433struct memory_failure_cpu {
1434 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1435 MEMORY_FAILURE_FIFO_SIZE);
1436 spinlock_t lock;
1437 struct work_struct work;
1438};
1439
1440static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1441
1442/**
1443 * memory_failure_queue - Schedule handling memory failure of a page.
1444 * @pfn: Page Number of the corrupted page
1445 * @flags: Flags for memory failure handling
1446 *
1447 * This function is called by the low level hardware error handler
1448 * when it detects hardware memory corruption of a page. It schedules
1449 * the recovering of error page, including dropping pages, killing
1450 * processes etc.
1451 *
1452 * The function is primarily of use for corruptions that
1453 * happen outside the current execution context (e.g. when
1454 * detected by a background scrubber)
1455 *
1456 * Can run in IRQ context.
1457 */
1458void memory_failure_queue(unsigned long pfn, int flags)
1459{
1460 struct memory_failure_cpu *mf_cpu;
1461 unsigned long proc_flags;
1462 struct memory_failure_entry entry = {
1463 .pfn = pfn,
1464 .flags = flags,
1465 };
1466
1467 mf_cpu = &get_cpu_var(memory_failure_cpu);
1468 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1469 if (kfifo_put(&mf_cpu->fifo, entry))
1470 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1471 else
1472 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1473 pfn);
1474 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1475 put_cpu_var(memory_failure_cpu);
1476}
1477EXPORT_SYMBOL_GPL(memory_failure_queue);
1478
1479static void memory_failure_work_func(struct work_struct *work)
1480{
1481 struct memory_failure_cpu *mf_cpu;
1482 struct memory_failure_entry entry = { 0, };
1483 unsigned long proc_flags;
1484 int gotten;
1485
1486 mf_cpu = this_cpu_ptr(&memory_failure_cpu);
1487 for (;;) {
1488 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1489 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1490 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1491 if (!gotten)
1492 break;
1493 if (entry.flags & MF_SOFT_OFFLINE)
1494 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1495 else
1496 memory_failure(entry.pfn, entry.flags);
1497 }
1498}
1499
1500static int __init memory_failure_init(void)
1501{
1502 struct memory_failure_cpu *mf_cpu;
1503 int cpu;
1504
1505 for_each_possible_cpu(cpu) {
1506 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1507 spin_lock_init(&mf_cpu->lock);
1508 INIT_KFIFO(mf_cpu->fifo);
1509 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1510 }
1511
1512 return 0;
1513}
1514core_initcall(memory_failure_init);
1515
1516#define unpoison_pr_info(fmt, pfn, rs) \
1517({ \
1518 if (__ratelimit(rs)) \
1519 pr_info(fmt, pfn); \
1520})
1521
1522/**
1523 * unpoison_memory - Unpoison a previously poisoned page
1524 * @pfn: Page number of the to be unpoisoned page
1525 *
1526 * Software-unpoison a page that has been poisoned by
1527 * memory_failure() earlier.
1528 *
1529 * This is only done on the software-level, so it only works
1530 * for linux injected failures, not real hardware failures
1531 *
1532 * Returns 0 for success, otherwise -errno.
1533 */
1534int unpoison_memory(unsigned long pfn)
1535{
1536 struct page *page;
1537 struct page *p;
1538 int freeit = 0;
1539 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
1540 DEFAULT_RATELIMIT_BURST);
1541
1542 if (!pfn_valid(pfn))
1543 return -ENXIO;
1544
1545 p = pfn_to_page(pfn);
1546 page = compound_head(p);
1547
1548 if (!PageHWPoison(p)) {
1549 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
1550 pfn, &unpoison_rs);
1551 return 0;
1552 }
1553
1554 if (page_count(page) > 1) {
1555 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
1556 pfn, &unpoison_rs);
1557 return 0;
1558 }
1559
1560 if (page_mapped(page)) {
1561 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
1562 pfn, &unpoison_rs);
1563 return 0;
1564 }
1565
1566 if (page_mapping(page)) {
1567 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
1568 pfn, &unpoison_rs);
1569 return 0;
1570 }
1571
1572 /*
1573 * unpoison_memory() can encounter thp only when the thp is being
1574 * worked by memory_failure() and the page lock is not held yet.
1575 * In such case, we yield to memory_failure() and make unpoison fail.
1576 */
1577 if (!PageHuge(page) && PageTransHuge(page)) {
1578 unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n",
1579 pfn, &unpoison_rs);
1580 return 0;
1581 }
1582
1583 if (!get_hwpoison_page(p)) {
1584 if (TestClearPageHWPoison(p))
1585 num_poisoned_pages_dec();
1586 unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n",
1587 pfn, &unpoison_rs);
1588 return 0;
1589 }
1590
1591 lock_page(page);
1592 /*
1593 * This test is racy because PG_hwpoison is set outside of page lock.
1594 * That's acceptable because that won't trigger kernel panic. Instead,
1595 * the PG_hwpoison page will be caught and isolated on the entrance to
1596 * the free buddy page pool.
1597 */
1598 if (TestClearPageHWPoison(page)) {
1599 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
1600 pfn, &unpoison_rs);
1601 num_poisoned_pages_dec();
1602 freeit = 1;
1603 }
1604 unlock_page(page);
1605
1606 put_hwpoison_page(page);
1607 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
1608 put_hwpoison_page(page);
1609
1610 return 0;
1611}
1612EXPORT_SYMBOL(unpoison_memory);
1613
1614static struct page *new_page(struct page *p, unsigned long private)
1615{
1616 int nid = page_to_nid(p);
1617
1618 return new_page_nodemask(p, nid, &node_states[N_MEMORY]);
1619}
1620
1621/*
1622 * Safely get reference count of an arbitrary page.
1623 * Returns 0 for a free page, -EIO for a zero refcount page
1624 * that is not free, and 1 for any other page type.
1625 * For 1 the page is returned with increased page count, otherwise not.
1626 */
1627static int __get_any_page(struct page *p, unsigned long pfn, int flags)
1628{
1629 int ret;
1630
1631 if (flags & MF_COUNT_INCREASED)
1632 return 1;
1633
1634 /*
1635 * When the target page is a free hugepage, just remove it
1636 * from free hugepage list.
1637 */
1638 if (!get_hwpoison_page(p)) {
1639 if (PageHuge(p)) {
1640 pr_info("%s: %#lx free huge page\n", __func__, pfn);
1641 ret = 0;
1642 } else if (is_free_buddy_page(p)) {
1643 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1644 ret = 0;
1645 } else {
1646 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1647 __func__, pfn, p->flags);
1648 ret = -EIO;
1649 }
1650 } else {
1651 /* Not a free page */
1652 ret = 1;
1653 }
1654 return ret;
1655}
1656
1657static int get_any_page(struct page *page, unsigned long pfn, int flags)
1658{
1659 int ret = __get_any_page(page, pfn, flags);
1660
1661 if (ret == 1 && !PageHuge(page) &&
1662 !PageLRU(page) && !__PageMovable(page)) {
1663 /*
1664 * Try to free it.
1665 */
1666 put_hwpoison_page(page);
1667 shake_page(page, 1);
1668
1669 /*
1670 * Did it turn free?
1671 */
1672 ret = __get_any_page(page, pfn, 0);
1673 if (ret == 1 && !PageLRU(page)) {
1674 /* Drop page reference which is from __get_any_page() */
1675 put_hwpoison_page(page);
1676 pr_info("soft_offline: %#lx: unknown non LRU page type %lx (%pGp)\n",
1677 pfn, page->flags, &page->flags);
1678 return -EIO;
1679 }
1680 }
1681 return ret;
1682}
1683
1684static int soft_offline_huge_page(struct page *page, int flags)
1685{
1686 int ret;
1687 unsigned long pfn = page_to_pfn(page);
1688 struct page *hpage = compound_head(page);
1689 LIST_HEAD(pagelist);
1690
1691 /*
1692 * This double-check of PageHWPoison is to avoid the race with
1693 * memory_failure(). See also comment in __soft_offline_page().
1694 */
1695 lock_page(hpage);
1696 if (PageHWPoison(hpage)) {
1697 unlock_page(hpage);
1698 put_hwpoison_page(hpage);
1699 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1700 return -EBUSY;
1701 }
1702 unlock_page(hpage);
1703
1704 ret = isolate_huge_page(hpage, &pagelist);
1705 /*
1706 * get_any_page() and isolate_huge_page() takes a refcount each,
1707 * so need to drop one here.
1708 */
1709 put_hwpoison_page(hpage);
1710 if (!ret) {
1711 pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn);
1712 return -EBUSY;
1713 }
1714
1715 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1716 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1717 if (ret) {
1718 pr_info("soft offline: %#lx: hugepage migration failed %d, type %lx (%pGp)\n",
1719 pfn, ret, page->flags, &page->flags);
1720 if (!list_empty(&pagelist))
1721 putback_movable_pages(&pagelist);
1722 if (ret > 0)
1723 ret = -EIO;
1724 } else {
1725 /*
1726 * We set PG_hwpoison only when the migration source hugepage
1727 * was successfully dissolved, because otherwise hwpoisoned
1728 * hugepage remains on free hugepage list, then userspace will
1729 * find it as SIGBUS by allocation failure. That's not expected
1730 * in soft-offlining.
1731 */
1732 ret = dissolve_free_huge_page(page);
1733 if (!ret) {
1734 if (set_hwpoison_free_buddy_page(page))
1735 num_poisoned_pages_inc();
1736 }
1737 }
1738 return ret;
1739}
1740
1741static int __soft_offline_page(struct page *page, int flags)
1742{
1743 int ret;
1744 unsigned long pfn = page_to_pfn(page);
1745
1746 /*
1747 * Check PageHWPoison again inside page lock because PageHWPoison
1748 * is set by memory_failure() outside page lock. Note that
1749 * memory_failure() also double-checks PageHWPoison inside page lock,
1750 * so there's no race between soft_offline_page() and memory_failure().
1751 */
1752 lock_page(page);
1753 wait_on_page_writeback(page);
1754 if (PageHWPoison(page)) {
1755 unlock_page(page);
1756 put_hwpoison_page(page);
1757 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1758 return -EBUSY;
1759 }
1760 /*
1761 * Try to invalidate first. This should work for
1762 * non dirty unmapped page cache pages.
1763 */
1764 ret = invalidate_inode_page(page);
1765 unlock_page(page);
1766 /*
1767 * RED-PEN would be better to keep it isolated here, but we
1768 * would need to fix isolation locking first.
1769 */
1770 if (ret == 1) {
1771 put_hwpoison_page(page);
1772 pr_info("soft_offline: %#lx: invalidated\n", pfn);
1773 SetPageHWPoison(page);
1774 num_poisoned_pages_inc();
1775 return 0;
1776 }
1777
1778 /*
1779 * Simple invalidation didn't work.
1780 * Try to migrate to a new page instead. migrate.c
1781 * handles a large number of cases for us.
1782 */
1783 if (PageLRU(page))
1784 ret = isolate_lru_page(page);
1785 else
1786 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1787 /*
1788 * Drop page reference which is came from get_any_page()
1789 * successful isolate_lru_page() already took another one.
1790 */
1791 put_hwpoison_page(page);
1792 if (!ret) {
1793 LIST_HEAD(pagelist);
1794 /*
1795 * After isolated lru page, the PageLRU will be cleared,
1796 * so use !__PageMovable instead for LRU page's mapping
1797 * cannot have PAGE_MAPPING_MOVABLE.
1798 */
1799 if (!__PageMovable(page))
1800 inc_node_page_state(page, NR_ISOLATED_ANON +
1801 page_is_file_cache(page));
1802 list_add(&page->lru, &pagelist);
1803 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1804 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1805 if (ret) {
1806 if (!list_empty(&pagelist))
1807 putback_movable_pages(&pagelist);
1808
1809 pr_info("soft offline: %#lx: migration failed %d, type %lx (%pGp)\n",
1810 pfn, ret, page->flags, &page->flags);
1811 if (ret > 0)
1812 ret = -EIO;
1813 }
1814 } else {
1815 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx (%pGp)\n",
1816 pfn, ret, page_count(page), page->flags, &page->flags);
1817 }
1818 return ret;
1819}
1820
1821static int soft_offline_in_use_page(struct page *page, int flags)
1822{
1823 int ret;
1824 int mt;
1825 struct page *hpage = compound_head(page);
1826
1827 if (!PageHuge(page) && PageTransHuge(hpage)) {
1828 lock_page(page);
1829 if (!PageAnon(page) || unlikely(split_huge_page(page))) {
1830 unlock_page(page);
1831 if (!PageAnon(page))
1832 pr_info("soft offline: %#lx: non anonymous thp\n", page_to_pfn(page));
1833 else
1834 pr_info("soft offline: %#lx: thp split failed\n", page_to_pfn(page));
1835 put_hwpoison_page(page);
1836 return -EBUSY;
1837 }
1838 unlock_page(page);
1839 }
1840
1841 /*
1842 * Setting MIGRATE_ISOLATE here ensures that the page will be linked
1843 * to free list immediately (not via pcplist) when released after
1844 * successful page migration. Otherwise we can't guarantee that the
1845 * page is really free after put_page() returns, so
1846 * set_hwpoison_free_buddy_page() highly likely fails.
1847 */
1848 mt = get_pageblock_migratetype(page);
1849 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
1850 if (PageHuge(page))
1851 ret = soft_offline_huge_page(page, flags);
1852 else
1853 ret = __soft_offline_page(page, flags);
1854 set_pageblock_migratetype(page, mt);
1855 return ret;
1856}
1857
1858static int soft_offline_free_page(struct page *page)
1859{
1860 int rc = 0;
1861 struct page *head = compound_head(page);
1862
1863 if (PageHuge(head))
1864 rc = dissolve_free_huge_page(page);
1865 if (!rc) {
1866 if (set_hwpoison_free_buddy_page(page))
1867 num_poisoned_pages_inc();
1868 else
1869 rc = -EBUSY;
1870 }
1871 return rc;
1872}
1873
1874/**
1875 * soft_offline_page - Soft offline a page.
1876 * @page: page to offline
1877 * @flags: flags. Same as memory_failure().
1878 *
1879 * Returns 0 on success, otherwise negated errno.
1880 *
1881 * Soft offline a page, by migration or invalidation,
1882 * without killing anything. This is for the case when
1883 * a page is not corrupted yet (so it's still valid to access),
1884 * but has had a number of corrected errors and is better taken
1885 * out.
1886 *
1887 * The actual policy on when to do that is maintained by
1888 * user space.
1889 *
1890 * This should never impact any application or cause data loss,
1891 * however it might take some time.
1892 *
1893 * This is not a 100% solution for all memory, but tries to be
1894 * ``good enough'' for the majority of memory.
1895 */
1896int soft_offline_page(struct page *page, int flags)
1897{
1898 int ret;
1899 unsigned long pfn = page_to_pfn(page);
1900
1901 if (is_zone_device_page(page)) {
1902 pr_debug_ratelimited("soft_offline: %#lx page is device page\n",
1903 pfn);
1904 if (flags & MF_COUNT_INCREASED)
1905 put_page(page);
1906 return -EIO;
1907 }
1908
1909 if (PageHWPoison(page)) {
1910 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1911 if (flags & MF_COUNT_INCREASED)
1912 put_hwpoison_page(page);
1913 return -EBUSY;
1914 }
1915
1916 get_online_mems();
1917 ret = get_any_page(page, pfn, flags);
1918 put_online_mems();
1919
1920 if (ret > 0)
1921 ret = soft_offline_in_use_page(page, flags);
1922 else if (ret == 0)
1923 ret = soft_offline_free_page(page);
1924
1925 return ret;
1926}
1927