blob: 9c26eeca13425886690cddaf6dd45954fd3f0097 [file] [log] [blame]
Andi Kleen6a460792009-09-16 11:50:15 +02001/*
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 2bit ECC memory or cache
11 * failure.
12 *
13 * Handles page cache pages in various states. The tricky part
14 * here is that we can access any page asynchronous to other VM
15 * users, because memory failures could happen anytime and anywhere,
16 * possibly violating some of their assumptions. This is why this code
17 * has to be extremely careful. Generally it tries to use normal locking
18 * rules, as in get the standard locks, even if that means the
19 * error handling takes potentially a long time.
20 *
21 * The operation to map back from RMAP chains to processes has to walk
22 * the complete process list and has non linear complexity with the number
23 * mappings. In short it can be quite slow. But since memory corruptions
24 * are rare we hope to get away with this.
25 */
26
27/*
28 * Notebook:
29 * - hugetlb needs more code
30 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
31 * - pass bad pages to kdump next kernel
32 */
33#define DEBUG 1 /* remove me in 2.6.34 */
34#include <linux/kernel.h>
35#include <linux/mm.h>
36#include <linux/page-flags.h>
Wu Fengguang478c5ff2009-12-16 12:19:59 +010037#include <linux/kernel-page-flags.h>
Andi Kleen6a460792009-09-16 11:50:15 +020038#include <linux/sched.h>
Hugh Dickins01e00f82009-10-13 15:02:11 +010039#include <linux/ksm.h>
Andi Kleen6a460792009-09-16 11:50:15 +020040#include <linux/rmap.h>
41#include <linux/pagemap.h>
42#include <linux/swap.h>
43#include <linux/backing-dev.h>
Andi Kleenfacb6012009-12-16 12:20:00 +010044#include <linux/migrate.h>
45#include <linux/page-isolation.h>
46#include <linux/suspend.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090047#include <linux/slab.h>
Huang Yingbf998152010-05-31 14:28:19 +080048#include <linux/swapops.h>
Naoya Horiguchi7af446a2010-05-28 09:29:17 +090049#include <linux/hugetlb.h>
Andi Kleen6a460792009-09-16 11:50:15 +020050#include "internal.h"
51
52int sysctl_memory_failure_early_kill __read_mostly = 0;
53
54int sysctl_memory_failure_recovery __read_mostly = 1;
55
56atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
57
Andi Kleen27df5062009-12-21 19:56:42 +010058#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
59
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010060u32 hwpoison_filter_enable = 0;
Wu Fengguang7c116f22009-12-16 12:19:59 +010061u32 hwpoison_filter_dev_major = ~0U;
62u32 hwpoison_filter_dev_minor = ~0U;
Wu Fengguang478c5ff2009-12-16 12:19:59 +010063u64 hwpoison_filter_flags_mask;
64u64 hwpoison_filter_flags_value;
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010065EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
Wu Fengguang7c116f22009-12-16 12:19:59 +010066EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
67EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
Wu Fengguang478c5ff2009-12-16 12:19:59 +010068EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
69EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
Wu Fengguang7c116f22009-12-16 12:19:59 +010070
71static int hwpoison_filter_dev(struct page *p)
72{
73 struct address_space *mapping;
74 dev_t dev;
75
76 if (hwpoison_filter_dev_major == ~0U &&
77 hwpoison_filter_dev_minor == ~0U)
78 return 0;
79
80 /*
81 * page_mapping() does not accept slab page
82 */
83 if (PageSlab(p))
84 return -EINVAL;
85
86 mapping = page_mapping(p);
87 if (mapping == NULL || mapping->host == NULL)
88 return -EINVAL;
89
90 dev = mapping->host->i_sb->s_dev;
91 if (hwpoison_filter_dev_major != ~0U &&
92 hwpoison_filter_dev_major != MAJOR(dev))
93 return -EINVAL;
94 if (hwpoison_filter_dev_minor != ~0U &&
95 hwpoison_filter_dev_minor != MINOR(dev))
96 return -EINVAL;
97
98 return 0;
99}
100
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100101static int hwpoison_filter_flags(struct page *p)
102{
103 if (!hwpoison_filter_flags_mask)
104 return 0;
105
106 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
107 hwpoison_filter_flags_value)
108 return 0;
109 else
110 return -EINVAL;
111}
112
Andi Kleen4fd466e2009-12-16 12:19:59 +0100113/*
114 * This allows stress tests to limit test scope to a collection of tasks
115 * by putting them under some memcg. This prevents killing unrelated/important
116 * processes such as /sbin/init. Note that the target task may share clean
117 * pages with init (eg. libc text), which is harmless. If the target task
118 * share _dirty_ pages with another task B, the test scheme must make sure B
119 * is also included in the memcg. At last, due to race conditions this filter
120 * can only guarantee that the page either belongs to the memcg tasks, or is
121 * a freed page.
122 */
123#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
124u64 hwpoison_filter_memcg;
125EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
126static int hwpoison_filter_task(struct page *p)
127{
128 struct mem_cgroup *mem;
129 struct cgroup_subsys_state *css;
130 unsigned long ino;
131
132 if (!hwpoison_filter_memcg)
133 return 0;
134
135 mem = try_get_mem_cgroup_from_page(p);
136 if (!mem)
137 return -EINVAL;
138
139 css = mem_cgroup_css(mem);
140 /* root_mem_cgroup has NULL dentries */
141 if (!css->cgroup->dentry)
142 return -EINVAL;
143
144 ino = css->cgroup->dentry->d_inode->i_ino;
145 css_put(css);
146
147 if (ino != hwpoison_filter_memcg)
148 return -EINVAL;
149
150 return 0;
151}
152#else
153static int hwpoison_filter_task(struct page *p) { return 0; }
154#endif
155
Wu Fengguang7c116f22009-12-16 12:19:59 +0100156int hwpoison_filter(struct page *p)
157{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100158 if (!hwpoison_filter_enable)
159 return 0;
160
Wu Fengguang7c116f22009-12-16 12:19:59 +0100161 if (hwpoison_filter_dev(p))
162 return -EINVAL;
163
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100164 if (hwpoison_filter_flags(p))
165 return -EINVAL;
166
Andi Kleen4fd466e2009-12-16 12:19:59 +0100167 if (hwpoison_filter_task(p))
168 return -EINVAL;
169
Wu Fengguang7c116f22009-12-16 12:19:59 +0100170 return 0;
171}
Andi Kleen27df5062009-12-21 19:56:42 +0100172#else
173int hwpoison_filter(struct page *p)
174{
175 return 0;
176}
177#endif
178
Wu Fengguang7c116f22009-12-16 12:19:59 +0100179EXPORT_SYMBOL_GPL(hwpoison_filter);
180
Andi Kleen6a460792009-09-16 11:50:15 +0200181/*
182 * Send all the processes who have the page mapped an ``action optional''
183 * signal.
184 */
185static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
186 unsigned long pfn)
187{
188 struct siginfo si;
189 int ret;
190
191 printk(KERN_ERR
192 "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
193 pfn, t->comm, t->pid);
194 si.si_signo = SIGBUS;
195 si.si_errno = 0;
196 si.si_code = BUS_MCEERR_AO;
197 si.si_addr = (void *)addr;
198#ifdef __ARCH_SI_TRAPNO
199 si.si_trapno = trapno;
200#endif
201 si.si_addr_lsb = PAGE_SHIFT;
202 /*
203 * Don't use force here, it's convenient if the signal
204 * can be temporarily blocked.
205 * This could cause a loop when the user sets SIGBUS
206 * to SIG_IGN, but hopefully noone will do that?
207 */
208 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
209 if (ret < 0)
210 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
211 t->comm, t->pid, ret);
212 return ret;
213}
214
215/*
Andi Kleen588f9ce2009-12-16 12:19:57 +0100216 * When a unknown page type is encountered drain as many buffers as possible
217 * in the hope to turn the page into a LRU or free page, which we can handle.
218 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100219void shake_page(struct page *p, int access)
Andi Kleen588f9ce2009-12-16 12:19:57 +0100220{
221 if (!PageSlab(p)) {
222 lru_add_drain_all();
223 if (PageLRU(p))
224 return;
225 drain_all_pages();
226 if (PageLRU(p) || is_free_buddy_page(p))
227 return;
228 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100229
Andi Kleen588f9ce2009-12-16 12:19:57 +0100230 /*
Andi Kleenfacb6012009-12-16 12:20:00 +0100231 * Only all shrink_slab here (which would also
232 * shrink other caches) if access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100233 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100234 if (access) {
235 int nr;
236 do {
237 nr = shrink_slab(1000, GFP_KERNEL, 1000);
238 if (page_count(p) == 0)
239 break;
240 } while (nr > 10);
241 }
Andi Kleen588f9ce2009-12-16 12:19:57 +0100242}
243EXPORT_SYMBOL_GPL(shake_page);
244
245/*
Andi Kleen6a460792009-09-16 11:50:15 +0200246 * Kill all processes that have a poisoned page mapped and then isolate
247 * the page.
248 *
249 * General strategy:
250 * Find all processes having the page mapped and kill them.
251 * But we keep a page reference around so that the page is not
252 * actually freed yet.
253 * Then stash the page away
254 *
255 * There's no convenient way to get back to mapped processes
256 * from the VMAs. So do a brute-force search over all
257 * running processes.
258 *
259 * Remember that machine checks are not common (or rather
260 * if they are common you have other problems), so this shouldn't
261 * be a performance issue.
262 *
263 * Also there are some races possible while we get from the
264 * error detection to actually handle it.
265 */
266
267struct to_kill {
268 struct list_head nd;
269 struct task_struct *tsk;
270 unsigned long addr;
271 unsigned addr_valid:1;
272};
273
274/*
275 * Failure handling: if we can't find or can't kill a process there's
276 * not much we can do. We just print a message and ignore otherwise.
277 */
278
279/*
280 * Schedule a process for later kill.
281 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
282 * TBD would GFP_NOIO be enough?
283 */
284static void add_to_kill(struct task_struct *tsk, struct page *p,
285 struct vm_area_struct *vma,
286 struct list_head *to_kill,
287 struct to_kill **tkc)
288{
289 struct to_kill *tk;
290
291 if (*tkc) {
292 tk = *tkc;
293 *tkc = NULL;
294 } else {
295 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
296 if (!tk) {
297 printk(KERN_ERR
298 "MCE: Out of memory while machine check handling\n");
299 return;
300 }
301 }
302 tk->addr = page_address_in_vma(p, vma);
303 tk->addr_valid = 1;
304
305 /*
306 * In theory we don't have to kill when the page was
307 * munmaped. But it could be also a mremap. Since that's
308 * likely very rare kill anyways just out of paranoia, but use
309 * a SIGKILL because the error is not contained anymore.
310 */
311 if (tk->addr == -EFAULT) {
312 pr_debug("MCE: Unable to find user space address %lx in %s\n",
313 page_to_pfn(p), tsk->comm);
314 tk->addr_valid = 0;
315 }
316 get_task_struct(tsk);
317 tk->tsk = tsk;
318 list_add_tail(&tk->nd, to_kill);
319}
320
321/*
322 * Kill the processes that have been collected earlier.
323 *
324 * Only do anything when DOIT is set, otherwise just free the list
325 * (this is used for clean pages which do not need killing)
326 * Also when FAIL is set do a force kill because something went
327 * wrong earlier.
328 */
329static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
330 int fail, unsigned long pfn)
331{
332 struct to_kill *tk, *next;
333
334 list_for_each_entry_safe (tk, next, to_kill, nd) {
335 if (doit) {
336 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200337 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200338 * make sure the process doesn't catch the
339 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200340 */
341 if (fail || tk->addr_valid == 0) {
342 printk(KERN_ERR
343 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
344 pfn, tk->tsk->comm, tk->tsk->pid);
345 force_sig(SIGKILL, tk->tsk);
346 }
347
348 /*
349 * In theory the process could have mapped
350 * something else on the address in-between. We could
351 * check for that, but we need to tell the
352 * process anyways.
353 */
354 else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
355 pfn) < 0)
356 printk(KERN_ERR
357 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
358 pfn, tk->tsk->comm, tk->tsk->pid);
359 }
360 put_task_struct(tk->tsk);
361 kfree(tk);
362 }
363}
364
365static int task_early_kill(struct task_struct *tsk)
366{
367 if (!tsk->mm)
368 return 0;
369 if (tsk->flags & PF_MCE_PROCESS)
370 return !!(tsk->flags & PF_MCE_EARLY);
371 return sysctl_memory_failure_early_kill;
372}
373
374/*
375 * Collect processes when the error hit an anonymous page.
376 */
377static void collect_procs_anon(struct page *page, struct list_head *to_kill,
378 struct to_kill **tkc)
379{
380 struct vm_area_struct *vma;
381 struct task_struct *tsk;
382 struct anon_vma *av;
383
384 read_lock(&tasklist_lock);
385 av = page_lock_anon_vma(page);
386 if (av == NULL) /* Not actually mapped anymore */
387 goto out;
388 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800389 struct anon_vma_chain *vmac;
390
Andi Kleen6a460792009-09-16 11:50:15 +0200391 if (!task_early_kill(tsk))
392 continue;
Rik van Riel5beb4932010-03-05 13:42:07 -0800393 list_for_each_entry(vmac, &av->head, same_anon_vma) {
394 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200395 if (!page_mapped_in_vma(page, vma))
396 continue;
397 if (vma->vm_mm == tsk->mm)
398 add_to_kill(tsk, page, vma, to_kill, tkc);
399 }
400 }
401 page_unlock_anon_vma(av);
402out:
403 read_unlock(&tasklist_lock);
404}
405
406/*
407 * Collect processes when the error hit a file mapped page.
408 */
409static void collect_procs_file(struct page *page, struct list_head *to_kill,
410 struct to_kill **tkc)
411{
412 struct vm_area_struct *vma;
413 struct task_struct *tsk;
414 struct prio_tree_iter iter;
415 struct address_space *mapping = page->mapping;
416
417 /*
418 * A note on the locking order between the two locks.
419 * We don't rely on this particular order.
420 * If you have some other code that needs a different order
421 * feel free to switch them around. Or add a reverse link
422 * from mm_struct to task_struct, then this could be all
423 * done without taking tasklist_lock and looping over all tasks.
424 */
425
426 read_lock(&tasklist_lock);
427 spin_lock(&mapping->i_mmap_lock);
428 for_each_process(tsk) {
429 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
430
431 if (!task_early_kill(tsk))
432 continue;
433
434 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
435 pgoff) {
436 /*
437 * Send early kill signal to tasks where a vma covers
438 * the page but the corrupted page is not necessarily
439 * mapped it in its pte.
440 * Assume applications who requested early kill want
441 * to be informed of all such data corruptions.
442 */
443 if (vma->vm_mm == tsk->mm)
444 add_to_kill(tsk, page, vma, to_kill, tkc);
445 }
446 }
447 spin_unlock(&mapping->i_mmap_lock);
448 read_unlock(&tasklist_lock);
449}
450
451/*
452 * Collect the processes who have the corrupted page mapped to kill.
453 * This is done in two steps for locking reasons.
454 * First preallocate one tokill structure outside the spin locks,
455 * so that we can kill at least one process reasonably reliable.
456 */
457static void collect_procs(struct page *page, struct list_head *tokill)
458{
459 struct to_kill *tk;
460
461 if (!page->mapping)
462 return;
463
464 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
465 if (!tk)
466 return;
467 if (PageAnon(page))
468 collect_procs_anon(page, tokill, &tk);
469 else
470 collect_procs_file(page, tokill, &tk);
471 kfree(tk);
472}
473
474/*
475 * Error handlers for various types of pages.
476 */
477
478enum outcome {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100479 IGNORED, /* Error: cannot be handled */
480 FAILED, /* Error: handling failed */
Andi Kleen6a460792009-09-16 11:50:15 +0200481 DELAYED, /* Will be handled later */
Andi Kleen6a460792009-09-16 11:50:15 +0200482 RECOVERED, /* Successfully recovered */
483};
484
485static const char *action_name[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100486 [IGNORED] = "Ignored",
Andi Kleen6a460792009-09-16 11:50:15 +0200487 [FAILED] = "Failed",
488 [DELAYED] = "Delayed",
Andi Kleen6a460792009-09-16 11:50:15 +0200489 [RECOVERED] = "Recovered",
490};
491
492/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100493 * XXX: It is possible that a page is isolated from LRU cache,
494 * and then kept in swap cache or failed to remove from page cache.
495 * The page count will stop it from being freed by unpoison.
496 * Stress tests should be aware of this memory leak problem.
497 */
498static int delete_from_lru_cache(struct page *p)
499{
500 if (!isolate_lru_page(p)) {
501 /*
502 * Clear sensible page flags, so that the buddy system won't
503 * complain when the page is unpoison-and-freed.
504 */
505 ClearPageActive(p);
506 ClearPageUnevictable(p);
507 /*
508 * drop the page count elevated by isolate_lru_page()
509 */
510 page_cache_release(p);
511 return 0;
512 }
513 return -EIO;
514}
515
516/*
Andi Kleen6a460792009-09-16 11:50:15 +0200517 * Error hit kernel page.
518 * Do nothing, try to be lucky and not touch this instead. For a few cases we
519 * could be more sophisticated.
520 */
521static int me_kernel(struct page *p, unsigned long pfn)
522{
Andi Kleen6a460792009-09-16 11:50:15 +0200523 return IGNORED;
524}
525
526/*
527 * Page in unknown state. Do nothing.
528 */
529static int me_unknown(struct page *p, unsigned long pfn)
530{
531 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
532 return FAILED;
533}
534
535/*
Andi Kleen6a460792009-09-16 11:50:15 +0200536 * Clean (or cleaned) page cache page.
537 */
538static int me_pagecache_clean(struct page *p, unsigned long pfn)
539{
540 int err;
541 int ret = FAILED;
542 struct address_space *mapping;
543
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100544 delete_from_lru_cache(p);
545
Andi Kleen6a460792009-09-16 11:50:15 +0200546 /*
547 * For anonymous pages we're done the only reference left
548 * should be the one m_f() holds.
549 */
550 if (PageAnon(p))
551 return RECOVERED;
552
553 /*
554 * Now truncate the page in the page cache. This is really
555 * more like a "temporary hole punch"
556 * Don't do this for block devices when someone else
557 * has a reference, because it could be file system metadata
558 * and that's not safe to truncate.
559 */
560 mapping = page_mapping(p);
561 if (!mapping) {
562 /*
563 * Page has been teared down in the meanwhile
564 */
565 return FAILED;
566 }
567
568 /*
569 * Truncation is a bit tricky. Enable it per file system for now.
570 *
571 * Open: to take i_mutex or not for this? Right now we don't.
572 */
573 if (mapping->a_ops->error_remove_page) {
574 err = mapping->a_ops->error_remove_page(mapping, p);
575 if (err != 0) {
576 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
577 pfn, err);
578 } else if (page_has_private(p) &&
579 !try_to_release_page(p, GFP_NOIO)) {
580 pr_debug("MCE %#lx: failed to release buffers\n", pfn);
581 } else {
582 ret = RECOVERED;
583 }
584 } else {
585 /*
586 * If the file system doesn't support it just invalidate
587 * This fails on dirty or anything with private pages
588 */
589 if (invalidate_inode_page(p))
590 ret = RECOVERED;
591 else
592 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
593 pfn);
594 }
595 return ret;
596}
597
598/*
599 * Dirty cache page page
600 * Issues: when the error hit a hole page the error is not properly
601 * propagated.
602 */
603static int me_pagecache_dirty(struct page *p, unsigned long pfn)
604{
605 struct address_space *mapping = page_mapping(p);
606
607 SetPageError(p);
608 /* TBD: print more information about the file. */
609 if (mapping) {
610 /*
611 * IO error will be reported by write(), fsync(), etc.
612 * who check the mapping.
613 * This way the application knows that something went
614 * wrong with its dirty file data.
615 *
616 * There's one open issue:
617 *
618 * The EIO will be only reported on the next IO
619 * operation and then cleared through the IO map.
620 * Normally Linux has two mechanisms to pass IO error
621 * first through the AS_EIO flag in the address space
622 * and then through the PageError flag in the page.
623 * Since we drop pages on memory failure handling the
624 * only mechanism open to use is through AS_AIO.
625 *
626 * This has the disadvantage that it gets cleared on
627 * the first operation that returns an error, while
628 * the PageError bit is more sticky and only cleared
629 * when the page is reread or dropped. If an
630 * application assumes it will always get error on
631 * fsync, but does other operations on the fd before
632 * and the page is dropped inbetween then the error
633 * will not be properly reported.
634 *
635 * This can already happen even without hwpoisoned
636 * pages: first on metadata IO errors (which only
637 * report through AS_EIO) or when the page is dropped
638 * at the wrong time.
639 *
640 * So right now we assume that the application DTRT on
641 * the first EIO, but we're not worse than other parts
642 * of the kernel.
643 */
644 mapping_set_error(mapping, EIO);
645 }
646
647 return me_pagecache_clean(p, pfn);
648}
649
650/*
651 * Clean and dirty swap cache.
652 *
653 * Dirty swap cache page is tricky to handle. The page could live both in page
654 * cache and swap cache(ie. page is freshly swapped in). So it could be
655 * referenced concurrently by 2 types of PTEs:
656 * normal PTEs and swap PTEs. We try to handle them consistently by calling
657 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
658 * and then
659 * - clear dirty bit to prevent IO
660 * - remove from LRU
661 * - but keep in the swap cache, so that when we return to it on
662 * a later page fault, we know the application is accessing
663 * corrupted data and shall be killed (we installed simple
664 * interception code in do_swap_page to catch it).
665 *
666 * Clean swap cache pages can be directly isolated. A later page fault will
667 * bring in the known good data from disk.
668 */
669static int me_swapcache_dirty(struct page *p, unsigned long pfn)
670{
Andi Kleen6a460792009-09-16 11:50:15 +0200671 ClearPageDirty(p);
672 /* Trigger EIO in shmem: */
673 ClearPageUptodate(p);
674
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100675 if (!delete_from_lru_cache(p))
676 return DELAYED;
677 else
678 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200679}
680
681static int me_swapcache_clean(struct page *p, unsigned long pfn)
682{
Andi Kleen6a460792009-09-16 11:50:15 +0200683 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800684
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100685 if (!delete_from_lru_cache(p))
686 return RECOVERED;
687 else
688 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200689}
690
691/*
692 * Huge pages. Needs work.
693 * Issues:
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900694 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
695 * To narrow down kill region to one page, we need to break up pmd.
696 * - To support soft-offlining for hugepage, we need to support hugepage
697 * migration.
Andi Kleen6a460792009-09-16 11:50:15 +0200698 */
699static int me_huge_page(struct page *p, unsigned long pfn)
700{
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900701 struct page *hpage = compound_head(p);
702 /*
703 * We can safely recover from error on free or reserved (i.e.
704 * not in-use) hugepage by dequeuing it from freelist.
705 * To check whether a hugepage is in-use or not, we can't use
706 * page->lru because it can be used in other hugepage operations,
707 * such as __unmap_hugepage_range() and gather_surplus_pages().
708 * So instead we use page_mapping() and PageAnon().
709 * We assume that this function is called with page lock held,
710 * so there is no race between isolation and mapping/unmapping.
711 */
712 if (!(page_mapping(hpage) || PageAnon(hpage))) {
713 __isolate_hwpoisoned_huge_page(hpage);
714 return RECOVERED;
715 }
716 return DELAYED;
Andi Kleen6a460792009-09-16 11:50:15 +0200717}
718
719/*
720 * Various page states we can handle.
721 *
722 * A page state is defined by its current page->flags bits.
723 * The table matches them in order and calls the right handler.
724 *
725 * This is quite tricky because we can access page at any time
726 * in its live cycle, so all accesses have to be extremly careful.
727 *
728 * This is not complete. More states could be added.
729 * For any missing state don't attempt recovery.
730 */
731
732#define dirty (1UL << PG_dirty)
733#define sc (1UL << PG_swapcache)
734#define unevict (1UL << PG_unevictable)
735#define mlock (1UL << PG_mlocked)
736#define writeback (1UL << PG_writeback)
737#define lru (1UL << PG_lru)
738#define swapbacked (1UL << PG_swapbacked)
739#define head (1UL << PG_head)
740#define tail (1UL << PG_tail)
741#define compound (1UL << PG_compound)
742#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200743#define reserved (1UL << PG_reserved)
744
745static struct page_state {
746 unsigned long mask;
747 unsigned long res;
748 char *msg;
749 int (*action)(struct page *p, unsigned long pfn);
750} error_states[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100751 { reserved, reserved, "reserved kernel", me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100752 /*
753 * free pages are specially detected outside this table:
754 * PG_buddy pages only make a small fraction of all free pages.
755 */
Andi Kleen6a460792009-09-16 11:50:15 +0200756
757 /*
758 * Could in theory check if slab page is free or if we can drop
759 * currently unused objects without touching them. But just
760 * treat it as standard kernel for now.
761 */
762 { slab, slab, "kernel slab", me_kernel },
763
764#ifdef CONFIG_PAGEFLAGS_EXTENDED
765 { head, head, "huge", me_huge_page },
766 { tail, tail, "huge", me_huge_page },
767#else
768 { compound, compound, "huge", me_huge_page },
769#endif
770
771 { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
772 { sc|dirty, sc, "swapcache", me_swapcache_clean },
773
774 { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
775 { unevict, unevict, "unevictable LRU", me_pagecache_clean},
776
Andi Kleen6a460792009-09-16 11:50:15 +0200777 { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
778 { mlock, mlock, "mlocked LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200779
780 { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
781 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200782
783 /*
784 * Catchall entry: must be at end.
785 */
786 { 0, 0, "unknown page state", me_unknown },
787};
788
Andi Kleen2326c462009-12-16 12:20:00 +0100789#undef dirty
790#undef sc
791#undef unevict
792#undef mlock
793#undef writeback
794#undef lru
795#undef swapbacked
796#undef head
797#undef tail
798#undef compound
799#undef slab
800#undef reserved
801
Andi Kleen6a460792009-09-16 11:50:15 +0200802static void action_result(unsigned long pfn, char *msg, int result)
803{
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100804 struct page *page = pfn_to_page(pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200805
806 printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
807 pfn,
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100808 PageDirty(page) ? "dirty " : "",
Andi Kleen6a460792009-09-16 11:50:15 +0200809 msg, action_name[result]);
810}
811
812static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100813 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200814{
815 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200816 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200817
818 result = ps->action(p, pfn);
819 action_result(pfn, ps->msg, result);
Wu Fengguang7456b042009-10-19 08:15:01 +0200820
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100821 count = page_count(p) - 1;
Wu Fengguang138ce282009-12-16 12:19:58 +0100822 if (ps->action == me_swapcache_dirty && result == DELAYED)
823 count--;
824 if (count != 0) {
Andi Kleen6a460792009-09-16 11:50:15 +0200825 printk(KERN_ERR
826 "MCE %#lx: %s page still referenced by %d users\n",
Wu Fengguang7456b042009-10-19 08:15:01 +0200827 pfn, ps->msg, count);
Wu Fengguang138ce282009-12-16 12:19:58 +0100828 result = FAILED;
829 }
Andi Kleen6a460792009-09-16 11:50:15 +0200830
831 /* Could do more checks here if page looks ok */
832 /*
833 * Could adjust zone counters here to correct for the missing page.
834 */
835
Wu Fengguang138ce282009-12-16 12:19:58 +0100836 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200837}
838
839#define N_UNMAP_TRIES 5
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 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100845static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Andi Kleen6a460792009-09-16 11:50:15 +0200846 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 i;
853 int kill = 1;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900854 struct page *hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +0200855
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100856 if (PageReserved(p) || PageSlab(p))
857 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200858
Andi Kleen6a460792009-09-16 11:50:15 +0200859 /*
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 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900863 if (!page_mapped(hpage))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100864 return SWAP_SUCCESS;
865
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900866 if (PageKsm(p))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100867 return SWAP_FAIL;
Andi Kleen6a460792009-09-16 11:50:15 +0200868
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.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100878 * 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).
Andi Kleen6a460792009-09-16 11:50:15 +0200880 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900881 mapping = page_mapping(hpage);
882 if (!PageDirty(hpage) && mapping &&
883 mapping_cap_writeback_dirty(mapping)) {
884 if (page_mkclean(hpage)) {
885 SetPageDirty(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +0200886 } 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 * First collect all the processes that have the page
897 * mapped in dirty form. This has to be done before try_to_unmap,
898 * because ttu takes the rmap data structures down.
899 *
900 * Error handling: We ignore errors here because
901 * there's nothing that can be done.
902 */
903 if (kill)
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900904 collect_procs(hpage, &tokill);
Andi Kleen6a460792009-09-16 11:50:15 +0200905
906 /*
907 * try_to_unmap can fail temporarily due to races.
908 * Try a few times (RED-PEN better strategy?)
909 */
910 for (i = 0; i < N_UNMAP_TRIES; i++) {
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900911 ret = try_to_unmap(hpage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +0200912 if (ret == SWAP_SUCCESS)
913 break;
914 pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn, ret);
915 }
916
917 if (ret != SWAP_SUCCESS)
918 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900919 pfn, page_mapcount(hpage));
Andi Kleen6a460792009-09-16 11:50:15 +0200920
921 /*
922 * Now that the dirty bit has been propagated to the
923 * struct page and all unmaps done we can decide if
924 * killing is needed or not. Only kill when the page
925 * was dirty, otherwise the tokill list is merely
926 * freed. When there was a problem unmapping earlier
927 * use a more force-full uncatchable kill to prevent
928 * any accesses to the poisoned memory.
929 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900930 kill_procs_ao(&tokill, !!PageDirty(hpage), trapno,
Andi Kleen6a460792009-09-16 11:50:15 +0200931 ret != SWAP_SUCCESS, pfn);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100932
933 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200934}
935
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900936static void set_page_hwpoison_huge_page(struct page *hpage)
937{
938 int i;
939 int nr_pages = 1 << compound_order(hpage);
940 for (i = 0; i < nr_pages; i++)
941 SetPageHWPoison(hpage + i);
942}
943
944static void clear_page_hwpoison_huge_page(struct page *hpage)
945{
946 int i;
947 int nr_pages = 1 << compound_order(hpage);
948 for (i = 0; i < nr_pages; i++)
949 ClearPageHWPoison(hpage + i);
950}
951
Andi Kleen82ba0112009-12-16 12:19:57 +0100952int __memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200953{
954 struct page_state *ps;
955 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900956 struct page *hpage;
Andi Kleen6a460792009-09-16 11:50:15 +0200957 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +0900958 unsigned int nr_pages;
Andi Kleen6a460792009-09-16 11:50:15 +0200959
960 if (!sysctl_memory_failure_recovery)
961 panic("Memory failure from trap %d on page %lx", trapno, pfn);
962
963 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100964 printk(KERN_ERR
965 "MCE %#lx: memory outside kernel control\n",
966 pfn);
967 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +0200968 }
969
970 p = pfn_to_page(pfn);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900971 hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +0200972 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100973 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200974 return 0;
975 }
976
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +0900977 nr_pages = 1 << compound_order(hpage);
978 atomic_long_add(nr_pages, &mce_bad_pages);
Andi Kleen6a460792009-09-16 11:50:15 +0200979
980 /*
981 * We need/can do nothing about count=0 pages.
982 * 1) it's a free page, and therefore in safe hand:
983 * prep_new_page() will be the gate keeper.
984 * 2) it's part of a non-compound high order page.
985 * Implies some kernel user: cannot stop them from
986 * R/W the page; let's pray that the page has been
987 * used and will be freed some time later.
988 * In fact it's dangerous to directly bump up page count from 0,
989 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
990 */
Andi Kleen82ba0112009-12-16 12:19:57 +0100991 if (!(flags & MF_COUNT_INCREASED) &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900992 !get_page_unless_zero(hpage)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +0100993 if (is_free_buddy_page(p)) {
994 action_result(pfn, "free buddy", DELAYED);
995 return 0;
996 } else {
997 action_result(pfn, "high order kernel", IGNORED);
998 return -EBUSY;
999 }
Andi Kleen6a460792009-09-16 11:50:15 +02001000 }
1001
1002 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001003 * We ignore non-LRU pages for good reasons.
1004 * - PG_locked is only well defined for LRU pages and a few others
1005 * - to avoid races with __set_page_locked()
1006 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1007 * The check (unnecessarily) ignores LRU pages being isolated and
1008 * walked by the page reclaim code, however that's not a big loss.
1009 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001010 if (!PageLRU(p) && !PageHuge(p))
Andi Kleenfacb6012009-12-16 12:20:00 +01001011 shake_page(p, 0);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001012 if (!PageLRU(p) && !PageHuge(p)) {
Andi Kleen0474a602009-12-16 12:20:00 +01001013 /*
1014 * shake_page could have turned it free.
1015 */
1016 if (is_free_buddy_page(p)) {
1017 action_result(pfn, "free buddy, 2nd try", DELAYED);
1018 return 0;
1019 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001020 action_result(pfn, "non LRU", IGNORED);
1021 put_page(p);
1022 return -EBUSY;
1023 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001024
1025 /*
Andi Kleen6a460792009-09-16 11:50:15 +02001026 * Lock the page and wait for writeback to finish.
1027 * It's very difficult to mess with pages currently under IO
1028 * and in many cases impossible, so we just avoid it here.
1029 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001030 lock_page_nosync(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001031
1032 /*
1033 * unpoison always clear PG_hwpoison inside page lock
1034 */
1035 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001036 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001037 res = 0;
1038 goto out;
1039 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001040 if (hwpoison_filter(p)) {
1041 if (TestClearPageHWPoison(p))
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001042 atomic_long_sub(nr_pages, &mce_bad_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001043 unlock_page(hpage);
1044 put_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001045 return 0;
1046 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001047
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001048 /*
1049 * For error on the tail page, we should set PG_hwpoison
1050 * on the head page to show that the hugepage is hwpoisoned
1051 */
1052 if (PageTail(p) && TestSetPageHWPoison(hpage)) {
1053 action_result(pfn, "hugepage already hardware poisoned",
1054 IGNORED);
1055 unlock_page(hpage);
1056 put_page(hpage);
1057 return 0;
1058 }
1059 /*
1060 * Set PG_hwpoison on all pages in an error hugepage,
1061 * because containment is done in hugepage unit for now.
1062 * Since we have done TestSetPageHWPoison() for the head page with
1063 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1064 */
1065 if (PageHuge(p))
1066 set_page_hwpoison_huge_page(hpage);
1067
Andi Kleen6a460792009-09-16 11:50:15 +02001068 wait_on_page_writeback(p);
1069
1070 /*
1071 * Now take care of user space mappings.
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001072 * Abort on fail: __remove_from_page_cache() assumes unmapped page.
Andi Kleen6a460792009-09-16 11:50:15 +02001073 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001074 if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
1075 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1076 res = -EBUSY;
1077 goto out;
1078 }
Andi Kleen6a460792009-09-16 11:50:15 +02001079
1080 /*
1081 * Torn down by someone else?
1082 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001083 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Andi Kleen6a460792009-09-16 11:50:15 +02001084 action_result(pfn, "already truncated LRU", IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001085 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001086 goto out;
1087 }
1088
1089 res = -EBUSY;
1090 for (ps = error_states;; ps++) {
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001091 if ((p->flags & ps->mask) == ps->res) {
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +01001092 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001093 break;
1094 }
1095 }
1096out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001097 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001098 return res;
1099}
1100EXPORT_SYMBOL_GPL(__memory_failure);
1101
1102/**
1103 * memory_failure - Handle memory failure of a page.
1104 * @pfn: Page Number of the corrupted page
1105 * @trapno: Trap number reported in the signal to user space.
1106 *
1107 * This function is called by the low level machine check code
1108 * of an architecture when it detects hardware memory corruption
1109 * of a page. It tries its best to recover, which includes
1110 * dropping pages, killing processes etc.
1111 *
1112 * The function is primarily of use for corruptions that
1113 * happen outside the current execution context (e.g. when
1114 * detected by a background scrubber)
1115 *
1116 * Must run in process context (e.g. a work queue) with interrupts
1117 * enabled and no spinlocks hold.
1118 */
1119void memory_failure(unsigned long pfn, int trapno)
1120{
1121 __memory_failure(pfn, trapno, 0);
1122}
Wu Fengguang847ce402009-12-16 12:19:58 +01001123
1124/**
1125 * unpoison_memory - Unpoison a previously poisoned page
1126 * @pfn: Page number of the to be unpoisoned page
1127 *
1128 * Software-unpoison a page that has been poisoned by
1129 * memory_failure() earlier.
1130 *
1131 * This is only done on the software-level, so it only works
1132 * for linux injected failures, not real hardware failures
1133 *
1134 * Returns 0 for success, otherwise -errno.
1135 */
1136int unpoison_memory(unsigned long pfn)
1137{
1138 struct page *page;
1139 struct page *p;
1140 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001141 unsigned int nr_pages;
Wu Fengguang847ce402009-12-16 12:19:58 +01001142
1143 if (!pfn_valid(pfn))
1144 return -ENXIO;
1145
1146 p = pfn_to_page(pfn);
1147 page = compound_head(p);
1148
1149 if (!PageHWPoison(p)) {
1150 pr_debug("MCE: Page was already unpoisoned %#lx\n", pfn);
1151 return 0;
1152 }
1153
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001154 nr_pages = 1 << compound_order(page);
1155
Wu Fengguang847ce402009-12-16 12:19:58 +01001156 if (!get_page_unless_zero(page)) {
1157 if (TestClearPageHWPoison(p))
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001158 atomic_long_sub(nr_pages, &mce_bad_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001159 pr_debug("MCE: Software-unpoisoned free page %#lx\n", pfn);
1160 return 0;
1161 }
1162
1163 lock_page_nosync(page);
1164 /*
1165 * This test is racy because PG_hwpoison is set outside of page lock.
1166 * That's acceptable because that won't trigger kernel panic. Instead,
1167 * the PG_hwpoison page will be caught and isolated on the entrance to
1168 * the free buddy page pool.
1169 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001170 if (TestClearPageHWPoison(page)) {
Wu Fengguang847ce402009-12-16 12:19:58 +01001171 pr_debug("MCE: Software-unpoisoned page %#lx\n", pfn);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001172 atomic_long_sub(nr_pages, &mce_bad_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001173 freeit = 1;
1174 }
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001175 if (PageHuge(p))
1176 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001177 unlock_page(page);
1178
1179 put_page(page);
1180 if (freeit)
1181 put_page(page);
1182
1183 return 0;
1184}
1185EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001186
1187static struct page *new_page(struct page *p, unsigned long private, int **x)
1188{
Andi Kleen12686d12009-12-16 12:20:01 +01001189 int nid = page_to_nid(p);
1190 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001191}
1192
1193/*
1194 * Safely get reference count of an arbitrary page.
1195 * Returns 0 for a free page, -EIO for a zero refcount page
1196 * that is not free, and 1 for any other page type.
1197 * For 1 the page is returned with increased page count, otherwise not.
1198 */
1199static int get_any_page(struct page *p, unsigned long pfn, int flags)
1200{
1201 int ret;
1202
1203 if (flags & MF_COUNT_INCREASED)
1204 return 1;
1205
1206 /*
1207 * The lock_system_sleep prevents a race with memory hotplug,
1208 * because the isolation assumes there's only a single user.
1209 * This is a big hammer, a better would be nicer.
1210 */
1211 lock_system_sleep();
1212
1213 /*
1214 * Isolate the page, so that it doesn't get reallocated if it
1215 * was free.
1216 */
1217 set_migratetype_isolate(p);
1218 if (!get_page_unless_zero(compound_head(p))) {
1219 if (is_free_buddy_page(p)) {
1220 pr_debug("get_any_page: %#lx free buddy page\n", pfn);
1221 /* Set hwpoison bit while page is still isolated */
1222 SetPageHWPoison(p);
1223 ret = 0;
1224 } else {
1225 pr_debug("get_any_page: %#lx: unknown zero refcount page type %lx\n",
1226 pfn, p->flags);
1227 ret = -EIO;
1228 }
1229 } else {
1230 /* Not a free page */
1231 ret = 1;
1232 }
1233 unset_migratetype_isolate(p);
1234 unlock_system_sleep();
1235 return ret;
1236}
1237
1238/**
1239 * soft_offline_page - Soft offline a page.
1240 * @page: page to offline
1241 * @flags: flags. Same as memory_failure().
1242 *
1243 * Returns 0 on success, otherwise negated errno.
1244 *
1245 * Soft offline a page, by migration or invalidation,
1246 * without killing anything. This is for the case when
1247 * a page is not corrupted yet (so it's still valid to access),
1248 * but has had a number of corrected errors and is better taken
1249 * out.
1250 *
1251 * The actual policy on when to do that is maintained by
1252 * user space.
1253 *
1254 * This should never impact any application or cause data loss,
1255 * however it might take some time.
1256 *
1257 * This is not a 100% solution for all memory, but tries to be
1258 * ``good enough'' for the majority of memory.
1259 */
1260int soft_offline_page(struct page *page, int flags)
1261{
1262 int ret;
1263 unsigned long pfn = page_to_pfn(page);
1264
1265 ret = get_any_page(page, pfn, flags);
1266 if (ret < 0)
1267 return ret;
1268 if (ret == 0)
1269 goto done;
1270
1271 /*
1272 * Page cache page we can handle?
1273 */
1274 if (!PageLRU(page)) {
1275 /*
1276 * Try to free it.
1277 */
1278 put_page(page);
1279 shake_page(page, 1);
1280
1281 /*
1282 * Did it turn free?
1283 */
1284 ret = get_any_page(page, pfn, 0);
1285 if (ret < 0)
1286 return ret;
1287 if (ret == 0)
1288 goto done;
1289 }
1290 if (!PageLRU(page)) {
1291 pr_debug("soft_offline: %#lx: unknown non LRU page type %lx\n",
1292 pfn, page->flags);
1293 return -EIO;
1294 }
1295
1296 lock_page(page);
1297 wait_on_page_writeback(page);
1298
1299 /*
1300 * Synchronized using the page lock with memory_failure()
1301 */
1302 if (PageHWPoison(page)) {
1303 unlock_page(page);
1304 put_page(page);
1305 pr_debug("soft offline: %#lx page already poisoned\n", pfn);
1306 return -EBUSY;
1307 }
1308
1309 /*
1310 * Try to invalidate first. This should work for
1311 * non dirty unmapped page cache pages.
1312 */
1313 ret = invalidate_inode_page(page);
1314 unlock_page(page);
1315
1316 /*
1317 * Drop count because page migration doesn't like raised
1318 * counts. The page could get re-allocated, but if it becomes
1319 * LRU the isolation will just fail.
1320 * RED-PEN would be better to keep it isolated here, but we
1321 * would need to fix isolation locking first.
1322 */
1323 put_page(page);
1324 if (ret == 1) {
1325 ret = 0;
1326 pr_debug("soft_offline: %#lx: invalidated\n", pfn);
1327 goto done;
1328 }
1329
1330 /*
1331 * Simple invalidation didn't work.
1332 * Try to migrate to a new page instead. migrate.c
1333 * handles a large number of cases for us.
1334 */
1335 ret = isolate_lru_page(page);
1336 if (!ret) {
1337 LIST_HEAD(pagelist);
1338
1339 list_add(&page->lru, &pagelist);
1340 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0);
1341 if (ret) {
1342 pr_debug("soft offline: %#lx: migration failed %d, type %lx\n",
1343 pfn, ret, page->flags);
1344 if (ret > 0)
1345 ret = -EIO;
1346 }
1347 } else {
1348 pr_debug("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1349 pfn, ret, page_count(page), page->flags);
1350 }
1351 if (ret)
1352 return ret;
1353
1354done:
1355 atomic_long_add(1, &mce_bad_pages);
1356 SetPageHWPoison(page);
1357 /* keep elevated page count for bad page */
1358 return ret;
1359}
Huang Yingbf998152010-05-31 14:28:19 +08001360
Huang Yingbbeb3402010-06-22 14:23:11 +08001361/*
1362 * The caller must hold current->mm->mmap_sem in read mode.
1363 */
Huang Yingbf998152010-05-31 14:28:19 +08001364int is_hwpoison_address(unsigned long addr)
1365{
1366 pgd_t *pgdp;
1367 pud_t pud, *pudp;
1368 pmd_t pmd, *pmdp;
1369 pte_t pte, *ptep;
1370 swp_entry_t entry;
1371
1372 pgdp = pgd_offset(current->mm, addr);
1373 if (!pgd_present(*pgdp))
1374 return 0;
1375 pudp = pud_offset(pgdp, addr);
1376 pud = *pudp;
1377 if (!pud_present(pud) || pud_large(pud))
1378 return 0;
1379 pmdp = pmd_offset(pudp, addr);
1380 pmd = *pmdp;
1381 if (!pmd_present(pmd) || pmd_large(pmd))
1382 return 0;
1383 ptep = pte_offset_map(pmdp, addr);
1384 pte = *ptep;
1385 pte_unmap(ptep);
1386 if (!is_swap_pte(pte))
1387 return 0;
1388 entry = pte_to_swp_entry(pte);
1389 return is_hwpoison_entry(entry);
1390}
1391EXPORT_SYMBOL_GPL(is_hwpoison_address);