blob: d0b420aba7266e23026293fd47156decfec760eb [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>
Naoya Horiguchi7af446a2010-05-28 09:29:17 +090048#include <linux/hugetlb.h>
Andi Kleen6a460792009-09-16 11:50:15 +020049#include "internal.h"
50
51int sysctl_memory_failure_early_kill __read_mostly = 0;
52
53int sysctl_memory_failure_recovery __read_mostly = 1;
54
55atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
56
Andi Kleen27df5062009-12-21 19:56:42 +010057#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
58
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010059u32 hwpoison_filter_enable = 0;
Wu Fengguang7c116f22009-12-16 12:19:59 +010060u32 hwpoison_filter_dev_major = ~0U;
61u32 hwpoison_filter_dev_minor = ~0U;
Wu Fengguang478c5ff2009-12-16 12:19:59 +010062u64 hwpoison_filter_flags_mask;
63u64 hwpoison_filter_flags_value;
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010064EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
Wu Fengguang7c116f22009-12-16 12:19:59 +010065EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
66EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
Wu Fengguang478c5ff2009-12-16 12:19:59 +010067EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
68EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
Wu Fengguang7c116f22009-12-16 12:19:59 +010069
70static int hwpoison_filter_dev(struct page *p)
71{
72 struct address_space *mapping;
73 dev_t dev;
74
75 if (hwpoison_filter_dev_major == ~0U &&
76 hwpoison_filter_dev_minor == ~0U)
77 return 0;
78
79 /*
80 * page_mapping() does not accept slab page
81 */
82 if (PageSlab(p))
83 return -EINVAL;
84
85 mapping = page_mapping(p);
86 if (mapping == NULL || mapping->host == NULL)
87 return -EINVAL;
88
89 dev = mapping->host->i_sb->s_dev;
90 if (hwpoison_filter_dev_major != ~0U &&
91 hwpoison_filter_dev_major != MAJOR(dev))
92 return -EINVAL;
93 if (hwpoison_filter_dev_minor != ~0U &&
94 hwpoison_filter_dev_minor != MINOR(dev))
95 return -EINVAL;
96
97 return 0;
98}
99
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100100static int hwpoison_filter_flags(struct page *p)
101{
102 if (!hwpoison_filter_flags_mask)
103 return 0;
104
105 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
106 hwpoison_filter_flags_value)
107 return 0;
108 else
109 return -EINVAL;
110}
111
Andi Kleen4fd466e2009-12-16 12:19:59 +0100112/*
113 * This allows stress tests to limit test scope to a collection of tasks
114 * by putting them under some memcg. This prevents killing unrelated/important
115 * processes such as /sbin/init. Note that the target task may share clean
116 * pages with init (eg. libc text), which is harmless. If the target task
117 * share _dirty_ pages with another task B, the test scheme must make sure B
118 * is also included in the memcg. At last, due to race conditions this filter
119 * can only guarantee that the page either belongs to the memcg tasks, or is
120 * a freed page.
121 */
122#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
123u64 hwpoison_filter_memcg;
124EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
125static int hwpoison_filter_task(struct page *p)
126{
127 struct mem_cgroup *mem;
128 struct cgroup_subsys_state *css;
129 unsigned long ino;
130
131 if (!hwpoison_filter_memcg)
132 return 0;
133
134 mem = try_get_mem_cgroup_from_page(p);
135 if (!mem)
136 return -EINVAL;
137
138 css = mem_cgroup_css(mem);
139 /* root_mem_cgroup has NULL dentries */
140 if (!css->cgroup->dentry)
141 return -EINVAL;
142
143 ino = css->cgroup->dentry->d_inode->i_ino;
144 css_put(css);
145
146 if (ino != hwpoison_filter_memcg)
147 return -EINVAL;
148
149 return 0;
150}
151#else
152static int hwpoison_filter_task(struct page *p) { return 0; }
153#endif
154
Wu Fengguang7c116f22009-12-16 12:19:59 +0100155int hwpoison_filter(struct page *p)
156{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100157 if (!hwpoison_filter_enable)
158 return 0;
159
Wu Fengguang7c116f22009-12-16 12:19:59 +0100160 if (hwpoison_filter_dev(p))
161 return -EINVAL;
162
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100163 if (hwpoison_filter_flags(p))
164 return -EINVAL;
165
Andi Kleen4fd466e2009-12-16 12:19:59 +0100166 if (hwpoison_filter_task(p))
167 return -EINVAL;
168
Wu Fengguang7c116f22009-12-16 12:19:59 +0100169 return 0;
170}
Andi Kleen27df5062009-12-21 19:56:42 +0100171#else
172int hwpoison_filter(struct page *p)
173{
174 return 0;
175}
176#endif
177
Wu Fengguang7c116f22009-12-16 12:19:59 +0100178EXPORT_SYMBOL_GPL(hwpoison_filter);
179
Andi Kleen6a460792009-09-16 11:50:15 +0200180/*
181 * Send all the processes who have the page mapped an ``action optional''
182 * signal.
183 */
184static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
185 unsigned long pfn)
186{
187 struct siginfo si;
188 int ret;
189
190 printk(KERN_ERR
191 "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
192 pfn, t->comm, t->pid);
193 si.si_signo = SIGBUS;
194 si.si_errno = 0;
195 si.si_code = BUS_MCEERR_AO;
196 si.si_addr = (void *)addr;
197#ifdef __ARCH_SI_TRAPNO
198 si.si_trapno = trapno;
199#endif
200 si.si_addr_lsb = PAGE_SHIFT;
201 /*
202 * Don't use force here, it's convenient if the signal
203 * can be temporarily blocked.
204 * This could cause a loop when the user sets SIGBUS
205 * to SIG_IGN, but hopefully noone will do that?
206 */
207 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
208 if (ret < 0)
209 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
210 t->comm, t->pid, ret);
211 return ret;
212}
213
214/*
Andi Kleen588f9ce2009-12-16 12:19:57 +0100215 * When a unknown page type is encountered drain as many buffers as possible
216 * in the hope to turn the page into a LRU or free page, which we can handle.
217 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100218void shake_page(struct page *p, int access)
Andi Kleen588f9ce2009-12-16 12:19:57 +0100219{
220 if (!PageSlab(p)) {
221 lru_add_drain_all();
222 if (PageLRU(p))
223 return;
224 drain_all_pages();
225 if (PageLRU(p) || is_free_buddy_page(p))
226 return;
227 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100228
Andi Kleen588f9ce2009-12-16 12:19:57 +0100229 /*
Andi Kleenfacb6012009-12-16 12:20:00 +0100230 * Only all shrink_slab here (which would also
231 * shrink other caches) if access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100232 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100233 if (access) {
234 int nr;
235 do {
236 nr = shrink_slab(1000, GFP_KERNEL, 1000);
237 if (page_count(p) == 0)
238 break;
239 } while (nr > 10);
240 }
Andi Kleen588f9ce2009-12-16 12:19:57 +0100241}
242EXPORT_SYMBOL_GPL(shake_page);
243
244/*
Andi Kleen6a460792009-09-16 11:50:15 +0200245 * Kill all processes that have a poisoned page mapped and then isolate
246 * the page.
247 *
248 * General strategy:
249 * Find all processes having the page mapped and kill them.
250 * But we keep a page reference around so that the page is not
251 * actually freed yet.
252 * Then stash the page away
253 *
254 * There's no convenient way to get back to mapped processes
255 * from the VMAs. So do a brute-force search over all
256 * running processes.
257 *
258 * Remember that machine checks are not common (or rather
259 * if they are common you have other problems), so this shouldn't
260 * be a performance issue.
261 *
262 * Also there are some races possible while we get from the
263 * error detection to actually handle it.
264 */
265
266struct to_kill {
267 struct list_head nd;
268 struct task_struct *tsk;
269 unsigned long addr;
270 unsigned addr_valid:1;
271};
272
273/*
274 * Failure handling: if we can't find or can't kill a process there's
275 * not much we can do. We just print a message and ignore otherwise.
276 */
277
278/*
279 * Schedule a process for later kill.
280 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
281 * TBD would GFP_NOIO be enough?
282 */
283static void add_to_kill(struct task_struct *tsk, struct page *p,
284 struct vm_area_struct *vma,
285 struct list_head *to_kill,
286 struct to_kill **tkc)
287{
288 struct to_kill *tk;
289
290 if (*tkc) {
291 tk = *tkc;
292 *tkc = NULL;
293 } else {
294 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
295 if (!tk) {
296 printk(KERN_ERR
297 "MCE: Out of memory while machine check handling\n");
298 return;
299 }
300 }
301 tk->addr = page_address_in_vma(p, vma);
302 tk->addr_valid = 1;
303
304 /*
305 * In theory we don't have to kill when the page was
306 * munmaped. But it could be also a mremap. Since that's
307 * likely very rare kill anyways just out of paranoia, but use
308 * a SIGKILL because the error is not contained anymore.
309 */
310 if (tk->addr == -EFAULT) {
311 pr_debug("MCE: Unable to find user space address %lx in %s\n",
312 page_to_pfn(p), tsk->comm);
313 tk->addr_valid = 0;
314 }
315 get_task_struct(tsk);
316 tk->tsk = tsk;
317 list_add_tail(&tk->nd, to_kill);
318}
319
320/*
321 * Kill the processes that have been collected earlier.
322 *
323 * Only do anything when DOIT is set, otherwise just free the list
324 * (this is used for clean pages which do not need killing)
325 * Also when FAIL is set do a force kill because something went
326 * wrong earlier.
327 */
328static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
329 int fail, unsigned long pfn)
330{
331 struct to_kill *tk, *next;
332
333 list_for_each_entry_safe (tk, next, to_kill, nd) {
334 if (doit) {
335 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200336 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200337 * make sure the process doesn't catch the
338 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200339 */
340 if (fail || tk->addr_valid == 0) {
341 printk(KERN_ERR
342 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
343 pfn, tk->tsk->comm, tk->tsk->pid);
344 force_sig(SIGKILL, tk->tsk);
345 }
346
347 /*
348 * In theory the process could have mapped
349 * something else on the address in-between. We could
350 * check for that, but we need to tell the
351 * process anyways.
352 */
353 else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
354 pfn) < 0)
355 printk(KERN_ERR
356 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
357 pfn, tk->tsk->comm, tk->tsk->pid);
358 }
359 put_task_struct(tk->tsk);
360 kfree(tk);
361 }
362}
363
364static int task_early_kill(struct task_struct *tsk)
365{
366 if (!tsk->mm)
367 return 0;
368 if (tsk->flags & PF_MCE_PROCESS)
369 return !!(tsk->flags & PF_MCE_EARLY);
370 return sysctl_memory_failure_early_kill;
371}
372
373/*
374 * Collect processes when the error hit an anonymous page.
375 */
376static void collect_procs_anon(struct page *page, struct list_head *to_kill,
377 struct to_kill **tkc)
378{
379 struct vm_area_struct *vma;
380 struct task_struct *tsk;
381 struct anon_vma *av;
382
383 read_lock(&tasklist_lock);
384 av = page_lock_anon_vma(page);
385 if (av == NULL) /* Not actually mapped anymore */
386 goto out;
387 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800388 struct anon_vma_chain *vmac;
389
Andi Kleen6a460792009-09-16 11:50:15 +0200390 if (!task_early_kill(tsk))
391 continue;
Rik van Riel5beb4932010-03-05 13:42:07 -0800392 list_for_each_entry(vmac, &av->head, same_anon_vma) {
393 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200394 if (!page_mapped_in_vma(page, vma))
395 continue;
396 if (vma->vm_mm == tsk->mm)
397 add_to_kill(tsk, page, vma, to_kill, tkc);
398 }
399 }
400 page_unlock_anon_vma(av);
401out:
402 read_unlock(&tasklist_lock);
403}
404
405/*
406 * Collect processes when the error hit a file mapped page.
407 */
408static void collect_procs_file(struct page *page, struct list_head *to_kill,
409 struct to_kill **tkc)
410{
411 struct vm_area_struct *vma;
412 struct task_struct *tsk;
413 struct prio_tree_iter iter;
414 struct address_space *mapping = page->mapping;
415
416 /*
417 * A note on the locking order between the two locks.
418 * We don't rely on this particular order.
419 * If you have some other code that needs a different order
420 * feel free to switch them around. Or add a reverse link
421 * from mm_struct to task_struct, then this could be all
422 * done without taking tasklist_lock and looping over all tasks.
423 */
424
425 read_lock(&tasklist_lock);
426 spin_lock(&mapping->i_mmap_lock);
427 for_each_process(tsk) {
428 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
429
430 if (!task_early_kill(tsk))
431 continue;
432
433 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
434 pgoff) {
435 /*
436 * Send early kill signal to tasks where a vma covers
437 * the page but the corrupted page is not necessarily
438 * mapped it in its pte.
439 * Assume applications who requested early kill want
440 * to be informed of all such data corruptions.
441 */
442 if (vma->vm_mm == tsk->mm)
443 add_to_kill(tsk, page, vma, to_kill, tkc);
444 }
445 }
446 spin_unlock(&mapping->i_mmap_lock);
447 read_unlock(&tasklist_lock);
448}
449
450/*
451 * Collect the processes who have the corrupted page mapped to kill.
452 * This is done in two steps for locking reasons.
453 * First preallocate one tokill structure outside the spin locks,
454 * so that we can kill at least one process reasonably reliable.
455 */
456static void collect_procs(struct page *page, struct list_head *tokill)
457{
458 struct to_kill *tk;
459
460 if (!page->mapping)
461 return;
462
463 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
464 if (!tk)
465 return;
466 if (PageAnon(page))
467 collect_procs_anon(page, tokill, &tk);
468 else
469 collect_procs_file(page, tokill, &tk);
470 kfree(tk);
471}
472
473/*
474 * Error handlers for various types of pages.
475 */
476
477enum outcome {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100478 IGNORED, /* Error: cannot be handled */
479 FAILED, /* Error: handling failed */
Andi Kleen6a460792009-09-16 11:50:15 +0200480 DELAYED, /* Will be handled later */
Andi Kleen6a460792009-09-16 11:50:15 +0200481 RECOVERED, /* Successfully recovered */
482};
483
484static const char *action_name[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100485 [IGNORED] = "Ignored",
Andi Kleen6a460792009-09-16 11:50:15 +0200486 [FAILED] = "Failed",
487 [DELAYED] = "Delayed",
Andi Kleen6a460792009-09-16 11:50:15 +0200488 [RECOVERED] = "Recovered",
489};
490
491/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100492 * XXX: It is possible that a page is isolated from LRU cache,
493 * and then kept in swap cache or failed to remove from page cache.
494 * The page count will stop it from being freed by unpoison.
495 * Stress tests should be aware of this memory leak problem.
496 */
497static int delete_from_lru_cache(struct page *p)
498{
499 if (!isolate_lru_page(p)) {
500 /*
501 * Clear sensible page flags, so that the buddy system won't
502 * complain when the page is unpoison-and-freed.
503 */
504 ClearPageActive(p);
505 ClearPageUnevictable(p);
506 /*
507 * drop the page count elevated by isolate_lru_page()
508 */
509 page_cache_release(p);
510 return 0;
511 }
512 return -EIO;
513}
514
515/*
Andi Kleen6a460792009-09-16 11:50:15 +0200516 * Error hit kernel page.
517 * Do nothing, try to be lucky and not touch this instead. For a few cases we
518 * could be more sophisticated.
519 */
520static int me_kernel(struct page *p, unsigned long pfn)
521{
Andi Kleen6a460792009-09-16 11:50:15 +0200522 return IGNORED;
523}
524
525/*
526 * Page in unknown state. Do nothing.
527 */
528static int me_unknown(struct page *p, unsigned long pfn)
529{
530 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
531 return FAILED;
532}
533
534/*
Andi Kleen6a460792009-09-16 11:50:15 +0200535 * Clean (or cleaned) page cache page.
536 */
537static int me_pagecache_clean(struct page *p, unsigned long pfn)
538{
539 int err;
540 int ret = FAILED;
541 struct address_space *mapping;
542
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100543 delete_from_lru_cache(p);
544
Andi Kleen6a460792009-09-16 11:50:15 +0200545 /*
546 * For anonymous pages we're done the only reference left
547 * should be the one m_f() holds.
548 */
549 if (PageAnon(p))
550 return RECOVERED;
551
552 /*
553 * Now truncate the page in the page cache. This is really
554 * more like a "temporary hole punch"
555 * Don't do this for block devices when someone else
556 * has a reference, because it could be file system metadata
557 * and that's not safe to truncate.
558 */
559 mapping = page_mapping(p);
560 if (!mapping) {
561 /*
562 * Page has been teared down in the meanwhile
563 */
564 return FAILED;
565 }
566
567 /*
568 * Truncation is a bit tricky. Enable it per file system for now.
569 *
570 * Open: to take i_mutex or not for this? Right now we don't.
571 */
572 if (mapping->a_ops->error_remove_page) {
573 err = mapping->a_ops->error_remove_page(mapping, p);
574 if (err != 0) {
575 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
576 pfn, err);
577 } else if (page_has_private(p) &&
578 !try_to_release_page(p, GFP_NOIO)) {
579 pr_debug("MCE %#lx: failed to release buffers\n", pfn);
580 } else {
581 ret = RECOVERED;
582 }
583 } else {
584 /*
585 * If the file system doesn't support it just invalidate
586 * This fails on dirty or anything with private pages
587 */
588 if (invalidate_inode_page(p))
589 ret = RECOVERED;
590 else
591 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
592 pfn);
593 }
594 return ret;
595}
596
597/*
598 * Dirty cache page page
599 * Issues: when the error hit a hole page the error is not properly
600 * propagated.
601 */
602static int me_pagecache_dirty(struct page *p, unsigned long pfn)
603{
604 struct address_space *mapping = page_mapping(p);
605
606 SetPageError(p);
607 /* TBD: print more information about the file. */
608 if (mapping) {
609 /*
610 * IO error will be reported by write(), fsync(), etc.
611 * who check the mapping.
612 * This way the application knows that something went
613 * wrong with its dirty file data.
614 *
615 * There's one open issue:
616 *
617 * The EIO will be only reported on the next IO
618 * operation and then cleared through the IO map.
619 * Normally Linux has two mechanisms to pass IO error
620 * first through the AS_EIO flag in the address space
621 * and then through the PageError flag in the page.
622 * Since we drop pages on memory failure handling the
623 * only mechanism open to use is through AS_AIO.
624 *
625 * This has the disadvantage that it gets cleared on
626 * the first operation that returns an error, while
627 * the PageError bit is more sticky and only cleared
628 * when the page is reread or dropped. If an
629 * application assumes it will always get error on
630 * fsync, but does other operations on the fd before
631 * and the page is dropped inbetween then the error
632 * will not be properly reported.
633 *
634 * This can already happen even without hwpoisoned
635 * pages: first on metadata IO errors (which only
636 * report through AS_EIO) or when the page is dropped
637 * at the wrong time.
638 *
639 * So right now we assume that the application DTRT on
640 * the first EIO, but we're not worse than other parts
641 * of the kernel.
642 */
643 mapping_set_error(mapping, EIO);
644 }
645
646 return me_pagecache_clean(p, pfn);
647}
648
649/*
650 * Clean and dirty swap cache.
651 *
652 * Dirty swap cache page is tricky to handle. The page could live both in page
653 * cache and swap cache(ie. page is freshly swapped in). So it could be
654 * referenced concurrently by 2 types of PTEs:
655 * normal PTEs and swap PTEs. We try to handle them consistently by calling
656 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
657 * and then
658 * - clear dirty bit to prevent IO
659 * - remove from LRU
660 * - but keep in the swap cache, so that when we return to it on
661 * a later page fault, we know the application is accessing
662 * corrupted data and shall be killed (we installed simple
663 * interception code in do_swap_page to catch it).
664 *
665 * Clean swap cache pages can be directly isolated. A later page fault will
666 * bring in the known good data from disk.
667 */
668static int me_swapcache_dirty(struct page *p, unsigned long pfn)
669{
Andi Kleen6a460792009-09-16 11:50:15 +0200670 ClearPageDirty(p);
671 /* Trigger EIO in shmem: */
672 ClearPageUptodate(p);
673
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100674 if (!delete_from_lru_cache(p))
675 return DELAYED;
676 else
677 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200678}
679
680static int me_swapcache_clean(struct page *p, unsigned long pfn)
681{
Andi Kleen6a460792009-09-16 11:50:15 +0200682 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800683
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100684 if (!delete_from_lru_cache(p))
685 return RECOVERED;
686 else
687 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200688}
689
690/*
691 * Huge pages. Needs work.
692 * Issues:
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900693 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
694 * To narrow down kill region to one page, we need to break up pmd.
695 * - To support soft-offlining for hugepage, we need to support hugepage
696 * migration.
Andi Kleen6a460792009-09-16 11:50:15 +0200697 */
698static int me_huge_page(struct page *p, unsigned long pfn)
699{
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900700 struct page *hpage = compound_head(p);
701 /*
702 * We can safely recover from error on free or reserved (i.e.
703 * not in-use) hugepage by dequeuing it from freelist.
704 * To check whether a hugepage is in-use or not, we can't use
705 * page->lru because it can be used in other hugepage operations,
706 * such as __unmap_hugepage_range() and gather_surplus_pages().
707 * So instead we use page_mapping() and PageAnon().
708 * We assume that this function is called with page lock held,
709 * so there is no race between isolation and mapping/unmapping.
710 */
711 if (!(page_mapping(hpage) || PageAnon(hpage))) {
712 __isolate_hwpoisoned_huge_page(hpage);
713 return RECOVERED;
714 }
715 return DELAYED;
Andi Kleen6a460792009-09-16 11:50:15 +0200716}
717
718/*
719 * Various page states we can handle.
720 *
721 * A page state is defined by its current page->flags bits.
722 * The table matches them in order and calls the right handler.
723 *
724 * This is quite tricky because we can access page at any time
725 * in its live cycle, so all accesses have to be extremly careful.
726 *
727 * This is not complete. More states could be added.
728 * For any missing state don't attempt recovery.
729 */
730
731#define dirty (1UL << PG_dirty)
732#define sc (1UL << PG_swapcache)
733#define unevict (1UL << PG_unevictable)
734#define mlock (1UL << PG_mlocked)
735#define writeback (1UL << PG_writeback)
736#define lru (1UL << PG_lru)
737#define swapbacked (1UL << PG_swapbacked)
738#define head (1UL << PG_head)
739#define tail (1UL << PG_tail)
740#define compound (1UL << PG_compound)
741#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200742#define reserved (1UL << PG_reserved)
743
744static struct page_state {
745 unsigned long mask;
746 unsigned long res;
747 char *msg;
748 int (*action)(struct page *p, unsigned long pfn);
749} error_states[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100750 { reserved, reserved, "reserved kernel", me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100751 /*
752 * free pages are specially detected outside this table:
753 * PG_buddy pages only make a small fraction of all free pages.
754 */
Andi Kleen6a460792009-09-16 11:50:15 +0200755
756 /*
757 * Could in theory check if slab page is free or if we can drop
758 * currently unused objects without touching them. But just
759 * treat it as standard kernel for now.
760 */
761 { slab, slab, "kernel slab", me_kernel },
762
763#ifdef CONFIG_PAGEFLAGS_EXTENDED
764 { head, head, "huge", me_huge_page },
765 { tail, tail, "huge", me_huge_page },
766#else
767 { compound, compound, "huge", me_huge_page },
768#endif
769
770 { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
771 { sc|dirty, sc, "swapcache", me_swapcache_clean },
772
773 { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
774 { unevict, unevict, "unevictable LRU", me_pagecache_clean},
775
Andi Kleen6a460792009-09-16 11:50:15 +0200776 { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
777 { mlock, mlock, "mlocked LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200778
779 { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
780 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200781
782 /*
783 * Catchall entry: must be at end.
784 */
785 { 0, 0, "unknown page state", me_unknown },
786};
787
Andi Kleen2326c462009-12-16 12:20:00 +0100788#undef dirty
789#undef sc
790#undef unevict
791#undef mlock
792#undef writeback
793#undef lru
794#undef swapbacked
795#undef head
796#undef tail
797#undef compound
798#undef slab
799#undef reserved
800
Andi Kleen6a460792009-09-16 11:50:15 +0200801static void action_result(unsigned long pfn, char *msg, int result)
802{
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100803 struct page *page = pfn_to_page(pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200804
805 printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
806 pfn,
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100807 PageDirty(page) ? "dirty " : "",
Andi Kleen6a460792009-09-16 11:50:15 +0200808 msg, action_name[result]);
809}
810
811static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100812 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200813{
814 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200815 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200816
817 result = ps->action(p, pfn);
818 action_result(pfn, ps->msg, result);
Wu Fengguang7456b042009-10-19 08:15:01 +0200819
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100820 count = page_count(p) - 1;
Wu Fengguang138ce282009-12-16 12:19:58 +0100821 if (ps->action == me_swapcache_dirty && result == DELAYED)
822 count--;
823 if (count != 0) {
Andi Kleen6a460792009-09-16 11:50:15 +0200824 printk(KERN_ERR
825 "MCE %#lx: %s page still referenced by %d users\n",
Wu Fengguang7456b042009-10-19 08:15:01 +0200826 pfn, ps->msg, count);
Wu Fengguang138ce282009-12-16 12:19:58 +0100827 result = FAILED;
828 }
Andi Kleen6a460792009-09-16 11:50:15 +0200829
830 /* Could do more checks here if page looks ok */
831 /*
832 * Could adjust zone counters here to correct for the missing page.
833 */
834
Wu Fengguang138ce282009-12-16 12:19:58 +0100835 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200836}
837
838#define N_UNMAP_TRIES 5
839
840/*
841 * Do all that is necessary to remove user space mappings. Unmap
842 * the pages and send SIGBUS to the processes if the data was dirty.
843 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100844static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Andi Kleen6a460792009-09-16 11:50:15 +0200845 int trapno)
846{
847 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
848 struct address_space *mapping;
849 LIST_HEAD(tokill);
850 int ret;
851 int i;
852 int kill = 1;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900853 struct page *hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +0200854
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100855 if (PageReserved(p) || PageSlab(p))
856 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200857
Andi Kleen6a460792009-09-16 11:50:15 +0200858 /*
859 * This check implies we don't kill processes if their pages
860 * are in the swap cache early. Those are always late kills.
861 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900862 if (!page_mapped(hpage))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100863 return SWAP_SUCCESS;
864
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900865 if (PageKsm(p))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100866 return SWAP_FAIL;
Andi Kleen6a460792009-09-16 11:50:15 +0200867
868 if (PageSwapCache(p)) {
869 printk(KERN_ERR
870 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
871 ttu |= TTU_IGNORE_HWPOISON;
872 }
873
874 /*
875 * Propagate the dirty bit from PTEs to struct page first, because we
876 * need this to decide if we should kill or just drop the page.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100877 * XXX: the dirty test could be racy: set_page_dirty() may not always
878 * be called inside page lock (it's recommended but not enforced).
Andi Kleen6a460792009-09-16 11:50:15 +0200879 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900880 mapping = page_mapping(hpage);
881 if (!PageDirty(hpage) && mapping &&
882 mapping_cap_writeback_dirty(mapping)) {
883 if (page_mkclean(hpage)) {
884 SetPageDirty(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +0200885 } else {
886 kill = 0;
887 ttu |= TTU_IGNORE_HWPOISON;
888 printk(KERN_INFO
889 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
890 pfn);
891 }
892 }
893
894 /*
895 * First collect all the processes that have the page
896 * mapped in dirty form. This has to be done before try_to_unmap,
897 * because ttu takes the rmap data structures down.
898 *
899 * Error handling: We ignore errors here because
900 * there's nothing that can be done.
901 */
902 if (kill)
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900903 collect_procs(hpage, &tokill);
Andi Kleen6a460792009-09-16 11:50:15 +0200904
905 /*
906 * try_to_unmap can fail temporarily due to races.
907 * Try a few times (RED-PEN better strategy?)
908 */
909 for (i = 0; i < N_UNMAP_TRIES; i++) {
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900910 ret = try_to_unmap(hpage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +0200911 if (ret == SWAP_SUCCESS)
912 break;
913 pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn, ret);
914 }
915
916 if (ret != SWAP_SUCCESS)
917 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900918 pfn, page_mapcount(hpage));
Andi Kleen6a460792009-09-16 11:50:15 +0200919
920 /*
921 * Now that the dirty bit has been propagated to the
922 * struct page and all unmaps done we can decide if
923 * killing is needed or not. Only kill when the page
924 * was dirty, otherwise the tokill list is merely
925 * freed. When there was a problem unmapping earlier
926 * use a more force-full uncatchable kill to prevent
927 * any accesses to the poisoned memory.
928 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900929 kill_procs_ao(&tokill, !!PageDirty(hpage), trapno,
Andi Kleen6a460792009-09-16 11:50:15 +0200930 ret != SWAP_SUCCESS, pfn);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100931
932 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200933}
934
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900935static void set_page_hwpoison_huge_page(struct page *hpage)
936{
937 int i;
938 int nr_pages = 1 << compound_order(hpage);
939 for (i = 0; i < nr_pages; i++)
940 SetPageHWPoison(hpage + i);
941}
942
943static void clear_page_hwpoison_huge_page(struct page *hpage)
944{
945 int i;
946 int nr_pages = 1 << compound_order(hpage);
947 for (i = 0; i < nr_pages; i++)
948 ClearPageHWPoison(hpage + i);
949}
950
Andi Kleen82ba0112009-12-16 12:19:57 +0100951int __memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200952{
953 struct page_state *ps;
954 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900955 struct page *hpage;
Andi Kleen6a460792009-09-16 11:50:15 +0200956 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +0900957 unsigned int nr_pages;
Andi Kleen6a460792009-09-16 11:50:15 +0200958
959 if (!sysctl_memory_failure_recovery)
960 panic("Memory failure from trap %d on page %lx", trapno, pfn);
961
962 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100963 printk(KERN_ERR
964 "MCE %#lx: memory outside kernel control\n",
965 pfn);
966 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +0200967 }
968
969 p = pfn_to_page(pfn);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900970 hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +0200971 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100972 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200973 return 0;
974 }
975
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +0900976 nr_pages = 1 << compound_order(hpage);
977 atomic_long_add(nr_pages, &mce_bad_pages);
Andi Kleen6a460792009-09-16 11:50:15 +0200978
979 /*
980 * We need/can do nothing about count=0 pages.
981 * 1) it's a free page, and therefore in safe hand:
982 * prep_new_page() will be the gate keeper.
983 * 2) it's part of a non-compound high order page.
984 * Implies some kernel user: cannot stop them from
985 * R/W the page; let's pray that the page has been
986 * used and will be freed some time later.
987 * In fact it's dangerous to directly bump up page count from 0,
988 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
989 */
Andi Kleen82ba0112009-12-16 12:19:57 +0100990 if (!(flags & MF_COUNT_INCREASED) &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900991 !get_page_unless_zero(hpage)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +0100992 if (is_free_buddy_page(p)) {
993 action_result(pfn, "free buddy", DELAYED);
994 return 0;
995 } else {
996 action_result(pfn, "high order kernel", IGNORED);
997 return -EBUSY;
998 }
Andi Kleen6a460792009-09-16 11:50:15 +0200999 }
1000
1001 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001002 * We ignore non-LRU pages for good reasons.
1003 * - PG_locked is only well defined for LRU pages and a few others
1004 * - to avoid races with __set_page_locked()
1005 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1006 * The check (unnecessarily) ignores LRU pages being isolated and
1007 * walked by the page reclaim code, however that's not a big loss.
1008 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001009 if (!PageLRU(p) && !PageHuge(p))
Andi Kleenfacb6012009-12-16 12:20:00 +01001010 shake_page(p, 0);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001011 if (!PageLRU(p) && !PageHuge(p)) {
Andi Kleen0474a602009-12-16 12:20:00 +01001012 /*
1013 * shake_page could have turned it free.
1014 */
1015 if (is_free_buddy_page(p)) {
1016 action_result(pfn, "free buddy, 2nd try", DELAYED);
1017 return 0;
1018 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001019 action_result(pfn, "non LRU", IGNORED);
1020 put_page(p);
1021 return -EBUSY;
1022 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001023
1024 /*
Andi Kleen6a460792009-09-16 11:50:15 +02001025 * Lock the page and wait for writeback to finish.
1026 * It's very difficult to mess with pages currently under IO
1027 * and in many cases impossible, so we just avoid it here.
1028 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001029 lock_page_nosync(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001030
1031 /*
1032 * unpoison always clear PG_hwpoison inside page lock
1033 */
1034 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001035 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001036 res = 0;
1037 goto out;
1038 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001039 if (hwpoison_filter(p)) {
1040 if (TestClearPageHWPoison(p))
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001041 atomic_long_sub(nr_pages, &mce_bad_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001042 unlock_page(hpage);
1043 put_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001044 return 0;
1045 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001046
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001047 /*
1048 * For error on the tail page, we should set PG_hwpoison
1049 * on the head page to show that the hugepage is hwpoisoned
1050 */
1051 if (PageTail(p) && TestSetPageHWPoison(hpage)) {
1052 action_result(pfn, "hugepage already hardware poisoned",
1053 IGNORED);
1054 unlock_page(hpage);
1055 put_page(hpage);
1056 return 0;
1057 }
1058 /*
1059 * Set PG_hwpoison on all pages in an error hugepage,
1060 * because containment is done in hugepage unit for now.
1061 * Since we have done TestSetPageHWPoison() for the head page with
1062 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1063 */
1064 if (PageHuge(p))
1065 set_page_hwpoison_huge_page(hpage);
1066
Andi Kleen6a460792009-09-16 11:50:15 +02001067 wait_on_page_writeback(p);
1068
1069 /*
1070 * Now take care of user space mappings.
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001071 * Abort on fail: __remove_from_page_cache() assumes unmapped page.
Andi Kleen6a460792009-09-16 11:50:15 +02001072 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001073 if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
1074 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1075 res = -EBUSY;
1076 goto out;
1077 }
Andi Kleen6a460792009-09-16 11:50:15 +02001078
1079 /*
1080 * Torn down by someone else?
1081 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001082 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Andi Kleen6a460792009-09-16 11:50:15 +02001083 action_result(pfn, "already truncated LRU", IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001084 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001085 goto out;
1086 }
1087
1088 res = -EBUSY;
1089 for (ps = error_states;; ps++) {
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001090 if ((p->flags & ps->mask) == ps->res) {
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +01001091 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001092 break;
1093 }
1094 }
1095out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001096 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001097 return res;
1098}
1099EXPORT_SYMBOL_GPL(__memory_failure);
1100
1101/**
1102 * memory_failure - Handle memory failure of a page.
1103 * @pfn: Page Number of the corrupted page
1104 * @trapno: Trap number reported in the signal to user space.
1105 *
1106 * This function is called by the low level machine check code
1107 * of an architecture when it detects hardware memory corruption
1108 * of a page. It tries its best to recover, which includes
1109 * dropping pages, killing processes etc.
1110 *
1111 * The function is primarily of use for corruptions that
1112 * happen outside the current execution context (e.g. when
1113 * detected by a background scrubber)
1114 *
1115 * Must run in process context (e.g. a work queue) with interrupts
1116 * enabled and no spinlocks hold.
1117 */
1118void memory_failure(unsigned long pfn, int trapno)
1119{
1120 __memory_failure(pfn, trapno, 0);
1121}
Wu Fengguang847ce402009-12-16 12:19:58 +01001122
1123/**
1124 * unpoison_memory - Unpoison a previously poisoned page
1125 * @pfn: Page number of the to be unpoisoned page
1126 *
1127 * Software-unpoison a page that has been poisoned by
1128 * memory_failure() earlier.
1129 *
1130 * This is only done on the software-level, so it only works
1131 * for linux injected failures, not real hardware failures
1132 *
1133 * Returns 0 for success, otherwise -errno.
1134 */
1135int unpoison_memory(unsigned long pfn)
1136{
1137 struct page *page;
1138 struct page *p;
1139 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001140 unsigned int nr_pages;
Wu Fengguang847ce402009-12-16 12:19:58 +01001141
1142 if (!pfn_valid(pfn))
1143 return -ENXIO;
1144
1145 p = pfn_to_page(pfn);
1146 page = compound_head(p);
1147
1148 if (!PageHWPoison(p)) {
1149 pr_debug("MCE: Page was already unpoisoned %#lx\n", pfn);
1150 return 0;
1151 }
1152
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001153 nr_pages = 1 << compound_order(page);
1154
Wu Fengguang847ce402009-12-16 12:19:58 +01001155 if (!get_page_unless_zero(page)) {
1156 if (TestClearPageHWPoison(p))
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001157 atomic_long_sub(nr_pages, &mce_bad_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001158 pr_debug("MCE: Software-unpoisoned free page %#lx\n", pfn);
1159 return 0;
1160 }
1161
1162 lock_page_nosync(page);
1163 /*
1164 * This test is racy because PG_hwpoison is set outside of page lock.
1165 * That's acceptable because that won't trigger kernel panic. Instead,
1166 * the PG_hwpoison page will be caught and isolated on the entrance to
1167 * the free buddy page pool.
1168 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001169 if (TestClearPageHWPoison(page)) {
Wu Fengguang847ce402009-12-16 12:19:58 +01001170 pr_debug("MCE: Software-unpoisoned page %#lx\n", pfn);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001171 atomic_long_sub(nr_pages, &mce_bad_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001172 freeit = 1;
1173 }
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001174 if (PageHuge(p))
1175 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001176 unlock_page(page);
1177
1178 put_page(page);
1179 if (freeit)
1180 put_page(page);
1181
1182 return 0;
1183}
1184EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001185
1186static struct page *new_page(struct page *p, unsigned long private, int **x)
1187{
Andi Kleen12686d12009-12-16 12:20:01 +01001188 int nid = page_to_nid(p);
1189 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001190}
1191
1192/*
1193 * Safely get reference count of an arbitrary page.
1194 * Returns 0 for a free page, -EIO for a zero refcount page
1195 * that is not free, and 1 for any other page type.
1196 * For 1 the page is returned with increased page count, otherwise not.
1197 */
1198static int get_any_page(struct page *p, unsigned long pfn, int flags)
1199{
1200 int ret;
1201
1202 if (flags & MF_COUNT_INCREASED)
1203 return 1;
1204
1205 /*
1206 * The lock_system_sleep prevents a race with memory hotplug,
1207 * because the isolation assumes there's only a single user.
1208 * This is a big hammer, a better would be nicer.
1209 */
1210 lock_system_sleep();
1211
1212 /*
1213 * Isolate the page, so that it doesn't get reallocated if it
1214 * was free.
1215 */
1216 set_migratetype_isolate(p);
1217 if (!get_page_unless_zero(compound_head(p))) {
1218 if (is_free_buddy_page(p)) {
1219 pr_debug("get_any_page: %#lx free buddy page\n", pfn);
1220 /* Set hwpoison bit while page is still isolated */
1221 SetPageHWPoison(p);
1222 ret = 0;
1223 } else {
1224 pr_debug("get_any_page: %#lx: unknown zero refcount page type %lx\n",
1225 pfn, p->flags);
1226 ret = -EIO;
1227 }
1228 } else {
1229 /* Not a free page */
1230 ret = 1;
1231 }
1232 unset_migratetype_isolate(p);
1233 unlock_system_sleep();
1234 return ret;
1235}
1236
1237/**
1238 * soft_offline_page - Soft offline a page.
1239 * @page: page to offline
1240 * @flags: flags. Same as memory_failure().
1241 *
1242 * Returns 0 on success, otherwise negated errno.
1243 *
1244 * Soft offline a page, by migration or invalidation,
1245 * without killing anything. This is for the case when
1246 * a page is not corrupted yet (so it's still valid to access),
1247 * but has had a number of corrected errors and is better taken
1248 * out.
1249 *
1250 * The actual policy on when to do that is maintained by
1251 * user space.
1252 *
1253 * This should never impact any application or cause data loss,
1254 * however it might take some time.
1255 *
1256 * This is not a 100% solution for all memory, but tries to be
1257 * ``good enough'' for the majority of memory.
1258 */
1259int soft_offline_page(struct page *page, int flags)
1260{
1261 int ret;
1262 unsigned long pfn = page_to_pfn(page);
1263
1264 ret = get_any_page(page, pfn, flags);
1265 if (ret < 0)
1266 return ret;
1267 if (ret == 0)
1268 goto done;
1269
1270 /*
1271 * Page cache page we can handle?
1272 */
1273 if (!PageLRU(page)) {
1274 /*
1275 * Try to free it.
1276 */
1277 put_page(page);
1278 shake_page(page, 1);
1279
1280 /*
1281 * Did it turn free?
1282 */
1283 ret = get_any_page(page, pfn, 0);
1284 if (ret < 0)
1285 return ret;
1286 if (ret == 0)
1287 goto done;
1288 }
1289 if (!PageLRU(page)) {
1290 pr_debug("soft_offline: %#lx: unknown non LRU page type %lx\n",
1291 pfn, page->flags);
1292 return -EIO;
1293 }
1294
1295 lock_page(page);
1296 wait_on_page_writeback(page);
1297
1298 /*
1299 * Synchronized using the page lock with memory_failure()
1300 */
1301 if (PageHWPoison(page)) {
1302 unlock_page(page);
1303 put_page(page);
1304 pr_debug("soft offline: %#lx page already poisoned\n", pfn);
1305 return -EBUSY;
1306 }
1307
1308 /*
1309 * Try to invalidate first. This should work for
1310 * non dirty unmapped page cache pages.
1311 */
1312 ret = invalidate_inode_page(page);
1313 unlock_page(page);
1314
1315 /*
1316 * Drop count because page migration doesn't like raised
1317 * counts. The page could get re-allocated, but if it becomes
1318 * LRU the isolation will just fail.
1319 * RED-PEN would be better to keep it isolated here, but we
1320 * would need to fix isolation locking first.
1321 */
1322 put_page(page);
1323 if (ret == 1) {
1324 ret = 0;
1325 pr_debug("soft_offline: %#lx: invalidated\n", pfn);
1326 goto done;
1327 }
1328
1329 /*
1330 * Simple invalidation didn't work.
1331 * Try to migrate to a new page instead. migrate.c
1332 * handles a large number of cases for us.
1333 */
1334 ret = isolate_lru_page(page);
1335 if (!ret) {
1336 LIST_HEAD(pagelist);
1337
1338 list_add(&page->lru, &pagelist);
1339 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0);
1340 if (ret) {
1341 pr_debug("soft offline: %#lx: migration failed %d, type %lx\n",
1342 pfn, ret, page->flags);
1343 if (ret > 0)
1344 ret = -EIO;
1345 }
1346 } else {
1347 pr_debug("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1348 pfn, ret, page_count(page), page->flags);
1349 }
1350 if (ret)
1351 return ret;
1352
1353done:
1354 atomic_long_add(1, &mce_bad_pages);
1355 SetPageHWPoison(page);
1356 /* keep elevated page count for bad page */
1357 return ret;
1358}