blob: 6a283cc9317c52053e16a2bcf8ebaabb23634ad8 [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
Andi Kleen1c80b992010-09-27 23:09:51 +020010 * hardware as being corrupted usually due to a multi-bit ECC memory or cache
Andi Kleen6a460792009-09-16 11:50:15 +020011 * failure.
Andi Kleen1c80b992010-09-27 23:09:51 +020012 *
13 * In addition there is a "soft offline" entry point that allows stop using
14 * not-yet-corrupted-by-suspicious pages without killing anything.
Andi Kleen6a460792009-09-16 11:50:15 +020015 *
16 * Handles page cache pages in various states. The tricky part
Andi Kleen1c80b992010-09-27 23:09:51 +020017 * here is that we can access any page asynchronously in respect to
18 * other VM users, because memory failures could happen anytime and
19 * anywhere. This could violate some of their assumptions. This is why
20 * this code has to be extremely careful. Generally it tries to use
21 * normal locking rules, as in get the standard locks, even if that means
22 * the error handling takes potentially a long time.
23 *
24 * There are several operations here with exponential complexity because
25 * of unsuitable VM data structures. For example the operation to map back
26 * from RMAP chains to processes has to walk the complete process list and
27 * has non linear complexity with the number. But since memory corruptions
28 * are rare we hope to get away with this. This avoids impacting the core
29 * VM.
Andi Kleen6a460792009-09-16 11:50:15 +020030 */
31
32/*
33 * Notebook:
34 * - hugetlb needs more code
35 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
36 * - pass bad pages to kdump next kernel
37 */
Andi Kleen6a460792009-09-16 11:50:15 +020038#include <linux/kernel.h>
39#include <linux/mm.h>
40#include <linux/page-flags.h>
Wu Fengguang478c5ff2009-12-16 12:19:59 +010041#include <linux/kernel-page-flags.h>
Andi Kleen6a460792009-09-16 11:50:15 +020042#include <linux/sched.h>
Hugh Dickins01e00f82009-10-13 15:02:11 +010043#include <linux/ksm.h>
Andi Kleen6a460792009-09-16 11:50:15 +020044#include <linux/rmap.h>
45#include <linux/pagemap.h>
46#include <linux/swap.h>
47#include <linux/backing-dev.h>
Andi Kleenfacb6012009-12-16 12:20:00 +010048#include <linux/migrate.h>
49#include <linux/page-isolation.h>
50#include <linux/suspend.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090051#include <linux/slab.h>
Huang Yingbf998152010-05-31 14:28:19 +080052#include <linux/swapops.h>
Naoya Horiguchi7af446a2010-05-28 09:29:17 +090053#include <linux/hugetlb.h>
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -080054#include <linux/memory_hotplug.h>
Andi Kleen6a460792009-09-16 11:50:15 +020055#include "internal.h"
56
57int sysctl_memory_failure_early_kill __read_mostly = 0;
58
59int sysctl_memory_failure_recovery __read_mostly = 1;
60
61atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
62
Andi Kleen27df5062009-12-21 19:56:42 +010063#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
64
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010065u32 hwpoison_filter_enable = 0;
Wu Fengguang7c116f22009-12-16 12:19:59 +010066u32 hwpoison_filter_dev_major = ~0U;
67u32 hwpoison_filter_dev_minor = ~0U;
Wu Fengguang478c5ff2009-12-16 12:19:59 +010068u64 hwpoison_filter_flags_mask;
69u64 hwpoison_filter_flags_value;
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010070EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
Wu Fengguang7c116f22009-12-16 12:19:59 +010071EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
72EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
Wu Fengguang478c5ff2009-12-16 12:19:59 +010073EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
74EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
Wu Fengguang7c116f22009-12-16 12:19:59 +010075
76static int hwpoison_filter_dev(struct page *p)
77{
78 struct address_space *mapping;
79 dev_t dev;
80
81 if (hwpoison_filter_dev_major == ~0U &&
82 hwpoison_filter_dev_minor == ~0U)
83 return 0;
84
85 /*
Andi Kleen1c80b992010-09-27 23:09:51 +020086 * page_mapping() does not accept slab pages.
Wu Fengguang7c116f22009-12-16 12:19:59 +010087 */
88 if (PageSlab(p))
89 return -EINVAL;
90
91 mapping = page_mapping(p);
92 if (mapping == NULL || mapping->host == NULL)
93 return -EINVAL;
94
95 dev = mapping->host->i_sb->s_dev;
96 if (hwpoison_filter_dev_major != ~0U &&
97 hwpoison_filter_dev_major != MAJOR(dev))
98 return -EINVAL;
99 if (hwpoison_filter_dev_minor != ~0U &&
100 hwpoison_filter_dev_minor != MINOR(dev))
101 return -EINVAL;
102
103 return 0;
104}
105
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100106static int hwpoison_filter_flags(struct page *p)
107{
108 if (!hwpoison_filter_flags_mask)
109 return 0;
110
111 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
112 hwpoison_filter_flags_value)
113 return 0;
114 else
115 return -EINVAL;
116}
117
Andi Kleen4fd466e2009-12-16 12:19:59 +0100118/*
119 * This allows stress tests to limit test scope to a collection of tasks
120 * by putting them under some memcg. This prevents killing unrelated/important
121 * processes such as /sbin/init. Note that the target task may share clean
122 * pages with init (eg. libc text), which is harmless. If the target task
123 * share _dirty_ pages with another task B, the test scheme must make sure B
124 * is also included in the memcg. At last, due to race conditions this filter
125 * can only guarantee that the page either belongs to the memcg tasks, or is
126 * a freed page.
127 */
128#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
129u64 hwpoison_filter_memcg;
130EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
131static int hwpoison_filter_task(struct page *p)
132{
133 struct mem_cgroup *mem;
134 struct cgroup_subsys_state *css;
135 unsigned long ino;
136
137 if (!hwpoison_filter_memcg)
138 return 0;
139
140 mem = try_get_mem_cgroup_from_page(p);
141 if (!mem)
142 return -EINVAL;
143
144 css = mem_cgroup_css(mem);
145 /* root_mem_cgroup has NULL dentries */
146 if (!css->cgroup->dentry)
147 return -EINVAL;
148
149 ino = css->cgroup->dentry->d_inode->i_ino;
150 css_put(css);
151
152 if (ino != hwpoison_filter_memcg)
153 return -EINVAL;
154
155 return 0;
156}
157#else
158static int hwpoison_filter_task(struct page *p) { return 0; }
159#endif
160
Wu Fengguang7c116f22009-12-16 12:19:59 +0100161int hwpoison_filter(struct page *p)
162{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100163 if (!hwpoison_filter_enable)
164 return 0;
165
Wu Fengguang7c116f22009-12-16 12:19:59 +0100166 if (hwpoison_filter_dev(p))
167 return -EINVAL;
168
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100169 if (hwpoison_filter_flags(p))
170 return -EINVAL;
171
Andi Kleen4fd466e2009-12-16 12:19:59 +0100172 if (hwpoison_filter_task(p))
173 return -EINVAL;
174
Wu Fengguang7c116f22009-12-16 12:19:59 +0100175 return 0;
176}
Andi Kleen27df5062009-12-21 19:56:42 +0100177#else
178int hwpoison_filter(struct page *p)
179{
180 return 0;
181}
182#endif
183
Wu Fengguang7c116f22009-12-16 12:19:59 +0100184EXPORT_SYMBOL_GPL(hwpoison_filter);
185
Andi Kleen6a460792009-09-16 11:50:15 +0200186/*
187 * Send all the processes who have the page mapped an ``action optional''
188 * signal.
189 */
190static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
Andi Kleen0d9ee6a2010-09-27 22:03:33 +0200191 unsigned long pfn, struct page *page)
Andi Kleen6a460792009-09-16 11:50:15 +0200192{
193 struct siginfo si;
194 int ret;
195
196 printk(KERN_ERR
197 "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
198 pfn, t->comm, t->pid);
199 si.si_signo = SIGBUS;
200 si.si_errno = 0;
201 si.si_code = BUS_MCEERR_AO;
202 si.si_addr = (void *)addr;
203#ifdef __ARCH_SI_TRAPNO
204 si.si_trapno = trapno;
205#endif
Andi Kleen0d9ee6a2010-09-27 22:03:33 +0200206 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
Andi Kleen6a460792009-09-16 11:50:15 +0200207 /*
208 * Don't use force here, it's convenient if the signal
209 * can be temporarily blocked.
210 * This could cause a loop when the user sets SIGBUS
211 * to SIG_IGN, but hopefully noone will do that?
212 */
213 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
214 if (ret < 0)
215 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
216 t->comm, t->pid, ret);
217 return ret;
218}
219
220/*
Andi Kleen588f9ce2009-12-16 12:19:57 +0100221 * When a unknown page type is encountered drain as many buffers as possible
222 * in the hope to turn the page into a LRU or free page, which we can handle.
223 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100224void shake_page(struct page *p, int access)
Andi Kleen588f9ce2009-12-16 12:19:57 +0100225{
226 if (!PageSlab(p)) {
227 lru_add_drain_all();
228 if (PageLRU(p))
229 return;
230 drain_all_pages();
231 if (PageLRU(p) || is_free_buddy_page(p))
232 return;
233 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100234
Andi Kleen588f9ce2009-12-16 12:19:57 +0100235 /*
Andi Kleenfacb6012009-12-16 12:20:00 +0100236 * Only all shrink_slab here (which would also
237 * shrink other caches) if access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100238 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100239 if (access) {
240 int nr;
241 do {
242 nr = shrink_slab(1000, GFP_KERNEL, 1000);
Andi Kleen47f43e72010-09-28 07:37:55 +0200243 if (page_count(p) == 1)
Andi Kleenfacb6012009-12-16 12:20:00 +0100244 break;
245 } while (nr > 10);
246 }
Andi Kleen588f9ce2009-12-16 12:19:57 +0100247}
248EXPORT_SYMBOL_GPL(shake_page);
249
250/*
Andi Kleen6a460792009-09-16 11:50:15 +0200251 * Kill all processes that have a poisoned page mapped and then isolate
252 * the page.
253 *
254 * General strategy:
255 * Find all processes having the page mapped and kill them.
256 * But we keep a page reference around so that the page is not
257 * actually freed yet.
258 * Then stash the page away
259 *
260 * There's no convenient way to get back to mapped processes
261 * from the VMAs. So do a brute-force search over all
262 * running processes.
263 *
264 * Remember that machine checks are not common (or rather
265 * if they are common you have other problems), so this shouldn't
266 * be a performance issue.
267 *
268 * Also there are some races possible while we get from the
269 * error detection to actually handle it.
270 */
271
272struct to_kill {
273 struct list_head nd;
274 struct task_struct *tsk;
275 unsigned long addr;
Andi Kleen9033ae12010-09-27 23:36:05 +0200276 char addr_valid;
Andi Kleen6a460792009-09-16 11:50:15 +0200277};
278
279/*
280 * Failure handling: if we can't find or can't kill a process there's
281 * not much we can do. We just print a message and ignore otherwise.
282 */
283
284/*
285 * Schedule a process for later kill.
286 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
287 * TBD would GFP_NOIO be enough?
288 */
289static void add_to_kill(struct task_struct *tsk, struct page *p,
290 struct vm_area_struct *vma,
291 struct list_head *to_kill,
292 struct to_kill **tkc)
293{
294 struct to_kill *tk;
295
296 if (*tkc) {
297 tk = *tkc;
298 *tkc = NULL;
299 } else {
300 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
301 if (!tk) {
302 printk(KERN_ERR
303 "MCE: Out of memory while machine check handling\n");
304 return;
305 }
306 }
307 tk->addr = page_address_in_vma(p, vma);
308 tk->addr_valid = 1;
309
310 /*
311 * In theory we don't have to kill when the page was
312 * munmaped. But it could be also a mremap. Since that's
313 * likely very rare kill anyways just out of paranoia, but use
314 * a SIGKILL because the error is not contained anymore.
315 */
316 if (tk->addr == -EFAULT) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200317 pr_info("MCE: Unable to find user space address %lx in %s\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200318 page_to_pfn(p), tsk->comm);
319 tk->addr_valid = 0;
320 }
321 get_task_struct(tsk);
322 tk->tsk = tsk;
323 list_add_tail(&tk->nd, to_kill);
324}
325
326/*
327 * Kill the processes that have been collected earlier.
328 *
329 * Only do anything when DOIT is set, otherwise just free the list
330 * (this is used for clean pages which do not need killing)
331 * Also when FAIL is set do a force kill because something went
332 * wrong earlier.
333 */
334static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
Andi Kleen0d9ee6a2010-09-27 22:03:33 +0200335 int fail, struct page *page, unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200336{
337 struct to_kill *tk, *next;
338
339 list_for_each_entry_safe (tk, next, to_kill, nd) {
340 if (doit) {
341 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200342 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200343 * make sure the process doesn't catch the
344 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200345 */
346 if (fail || tk->addr_valid == 0) {
347 printk(KERN_ERR
348 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
349 pfn, tk->tsk->comm, tk->tsk->pid);
350 force_sig(SIGKILL, tk->tsk);
351 }
352
353 /*
354 * In theory the process could have mapped
355 * something else on the address in-between. We could
356 * check for that, but we need to tell the
357 * process anyways.
358 */
359 else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
Andi Kleen0d9ee6a2010-09-27 22:03:33 +0200360 pfn, page) < 0)
Andi Kleen6a460792009-09-16 11:50:15 +0200361 printk(KERN_ERR
362 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
363 pfn, tk->tsk->comm, tk->tsk->pid);
364 }
365 put_task_struct(tk->tsk);
366 kfree(tk);
367 }
368}
369
370static int task_early_kill(struct task_struct *tsk)
371{
372 if (!tsk->mm)
373 return 0;
374 if (tsk->flags & PF_MCE_PROCESS)
375 return !!(tsk->flags & PF_MCE_EARLY);
376 return sysctl_memory_failure_early_kill;
377}
378
379/*
380 * Collect processes when the error hit an anonymous page.
381 */
382static void collect_procs_anon(struct page *page, struct list_head *to_kill,
383 struct to_kill **tkc)
384{
385 struct vm_area_struct *vma;
386 struct task_struct *tsk;
387 struct anon_vma *av;
388
Andrea Arcangeli3f04f622011-01-13 15:46:47 -0800389 if (unlikely(split_huge_page(page)))
390 return;
Andi Kleen6a460792009-09-16 11:50:15 +0200391 read_lock(&tasklist_lock);
392 av = page_lock_anon_vma(page);
393 if (av == NULL) /* Not actually mapped anymore */
394 goto out;
395 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800396 struct anon_vma_chain *vmac;
397
Andi Kleen6a460792009-09-16 11:50:15 +0200398 if (!task_early_kill(tsk))
399 continue;
Rik van Riel5beb4932010-03-05 13:42:07 -0800400 list_for_each_entry(vmac, &av->head, same_anon_vma) {
401 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200402 if (!page_mapped_in_vma(page, vma))
403 continue;
404 if (vma->vm_mm == tsk->mm)
405 add_to_kill(tsk, page, vma, to_kill, tkc);
406 }
407 }
408 page_unlock_anon_vma(av);
409out:
410 read_unlock(&tasklist_lock);
411}
412
413/*
414 * Collect processes when the error hit a file mapped page.
415 */
416static void collect_procs_file(struct page *page, struct list_head *to_kill,
417 struct to_kill **tkc)
418{
419 struct vm_area_struct *vma;
420 struct task_struct *tsk;
421 struct prio_tree_iter iter;
422 struct address_space *mapping = page->mapping;
423
424 /*
425 * A note on the locking order between the two locks.
426 * We don't rely on this particular order.
427 * If you have some other code that needs a different order
428 * feel free to switch them around. Or add a reverse link
429 * from mm_struct to task_struct, then this could be all
430 * done without taking tasklist_lock and looping over all tasks.
431 */
432
433 read_lock(&tasklist_lock);
434 spin_lock(&mapping->i_mmap_lock);
435 for_each_process(tsk) {
436 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
437
438 if (!task_early_kill(tsk))
439 continue;
440
441 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
442 pgoff) {
443 /*
444 * Send early kill signal to tasks where a vma covers
445 * the page but the corrupted page is not necessarily
446 * mapped it in its pte.
447 * Assume applications who requested early kill want
448 * to be informed of all such data corruptions.
449 */
450 if (vma->vm_mm == tsk->mm)
451 add_to_kill(tsk, page, vma, to_kill, tkc);
452 }
453 }
454 spin_unlock(&mapping->i_mmap_lock);
455 read_unlock(&tasklist_lock);
456}
457
458/*
459 * Collect the processes who have the corrupted page mapped to kill.
460 * This is done in two steps for locking reasons.
461 * First preallocate one tokill structure outside the spin locks,
462 * so that we can kill at least one process reasonably reliable.
463 */
464static void collect_procs(struct page *page, struct list_head *tokill)
465{
466 struct to_kill *tk;
467
468 if (!page->mapping)
469 return;
470
471 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
472 if (!tk)
473 return;
474 if (PageAnon(page))
475 collect_procs_anon(page, tokill, &tk);
476 else
477 collect_procs_file(page, tokill, &tk);
478 kfree(tk);
479}
480
481/*
482 * Error handlers for various types of pages.
483 */
484
485enum outcome {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100486 IGNORED, /* Error: cannot be handled */
487 FAILED, /* Error: handling failed */
Andi Kleen6a460792009-09-16 11:50:15 +0200488 DELAYED, /* Will be handled later */
Andi Kleen6a460792009-09-16 11:50:15 +0200489 RECOVERED, /* Successfully recovered */
490};
491
492static const char *action_name[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100493 [IGNORED] = "Ignored",
Andi Kleen6a460792009-09-16 11:50:15 +0200494 [FAILED] = "Failed",
495 [DELAYED] = "Delayed",
Andi Kleen6a460792009-09-16 11:50:15 +0200496 [RECOVERED] = "Recovered",
497};
498
499/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100500 * XXX: It is possible that a page is isolated from LRU cache,
501 * and then kept in swap cache or failed to remove from page cache.
502 * The page count will stop it from being freed by unpoison.
503 * Stress tests should be aware of this memory leak problem.
504 */
505static int delete_from_lru_cache(struct page *p)
506{
507 if (!isolate_lru_page(p)) {
508 /*
509 * Clear sensible page flags, so that the buddy system won't
510 * complain when the page is unpoison-and-freed.
511 */
512 ClearPageActive(p);
513 ClearPageUnevictable(p);
514 /*
515 * drop the page count elevated by isolate_lru_page()
516 */
517 page_cache_release(p);
518 return 0;
519 }
520 return -EIO;
521}
522
523/*
Andi Kleen6a460792009-09-16 11:50:15 +0200524 * Error hit kernel page.
525 * Do nothing, try to be lucky and not touch this instead. For a few cases we
526 * could be more sophisticated.
527 */
528static int me_kernel(struct page *p, unsigned long pfn)
529{
Andi Kleen6a460792009-09-16 11:50:15 +0200530 return IGNORED;
531}
532
533/*
534 * Page in unknown state. Do nothing.
535 */
536static int me_unknown(struct page *p, unsigned long pfn)
537{
538 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
539 return FAILED;
540}
541
542/*
Andi Kleen6a460792009-09-16 11:50:15 +0200543 * Clean (or cleaned) page cache page.
544 */
545static int me_pagecache_clean(struct page *p, unsigned long pfn)
546{
547 int err;
548 int ret = FAILED;
549 struct address_space *mapping;
550
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100551 delete_from_lru_cache(p);
552
Andi Kleen6a460792009-09-16 11:50:15 +0200553 /*
554 * For anonymous pages we're done the only reference left
555 * should be the one m_f() holds.
556 */
557 if (PageAnon(p))
558 return RECOVERED;
559
560 /*
561 * Now truncate the page in the page cache. This is really
562 * more like a "temporary hole punch"
563 * Don't do this for block devices when someone else
564 * has a reference, because it could be file system metadata
565 * and that's not safe to truncate.
566 */
567 mapping = page_mapping(p);
568 if (!mapping) {
569 /*
570 * Page has been teared down in the meanwhile
571 */
572 return FAILED;
573 }
574
575 /*
576 * Truncation is a bit tricky. Enable it per file system for now.
577 *
578 * Open: to take i_mutex or not for this? Right now we don't.
579 */
580 if (mapping->a_ops->error_remove_page) {
581 err = mapping->a_ops->error_remove_page(mapping, p);
582 if (err != 0) {
583 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
584 pfn, err);
585 } else if (page_has_private(p) &&
586 !try_to_release_page(p, GFP_NOIO)) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200587 pr_info("MCE %#lx: failed to release buffers\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200588 } else {
589 ret = RECOVERED;
590 }
591 } else {
592 /*
593 * If the file system doesn't support it just invalidate
594 * This fails on dirty or anything with private pages
595 */
596 if (invalidate_inode_page(p))
597 ret = RECOVERED;
598 else
599 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
600 pfn);
601 }
602 return ret;
603}
604
605/*
606 * Dirty cache page page
607 * Issues: when the error hit a hole page the error is not properly
608 * propagated.
609 */
610static int me_pagecache_dirty(struct page *p, unsigned long pfn)
611{
612 struct address_space *mapping = page_mapping(p);
613
614 SetPageError(p);
615 /* TBD: print more information about the file. */
616 if (mapping) {
617 /*
618 * IO error will be reported by write(), fsync(), etc.
619 * who check the mapping.
620 * This way the application knows that something went
621 * wrong with its dirty file data.
622 *
623 * There's one open issue:
624 *
625 * The EIO will be only reported on the next IO
626 * operation and then cleared through the IO map.
627 * Normally Linux has two mechanisms to pass IO error
628 * first through the AS_EIO flag in the address space
629 * and then through the PageError flag in the page.
630 * Since we drop pages on memory failure handling the
631 * only mechanism open to use is through AS_AIO.
632 *
633 * This has the disadvantage that it gets cleared on
634 * the first operation that returns an error, while
635 * the PageError bit is more sticky and only cleared
636 * when the page is reread or dropped. If an
637 * application assumes it will always get error on
638 * fsync, but does other operations on the fd before
639 * and the page is dropped inbetween then the error
640 * will not be properly reported.
641 *
642 * This can already happen even without hwpoisoned
643 * pages: first on metadata IO errors (which only
644 * report through AS_EIO) or when the page is dropped
645 * at the wrong time.
646 *
647 * So right now we assume that the application DTRT on
648 * the first EIO, but we're not worse than other parts
649 * of the kernel.
650 */
651 mapping_set_error(mapping, EIO);
652 }
653
654 return me_pagecache_clean(p, pfn);
655}
656
657/*
658 * Clean and dirty swap cache.
659 *
660 * Dirty swap cache page is tricky to handle. The page could live both in page
661 * cache and swap cache(ie. page is freshly swapped in). So it could be
662 * referenced concurrently by 2 types of PTEs:
663 * normal PTEs and swap PTEs. We try to handle them consistently by calling
664 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
665 * and then
666 * - clear dirty bit to prevent IO
667 * - remove from LRU
668 * - but keep in the swap cache, so that when we return to it on
669 * a later page fault, we know the application is accessing
670 * corrupted data and shall be killed (we installed simple
671 * interception code in do_swap_page to catch it).
672 *
673 * Clean swap cache pages can be directly isolated. A later page fault will
674 * bring in the known good data from disk.
675 */
676static int me_swapcache_dirty(struct page *p, unsigned long pfn)
677{
Andi Kleen6a460792009-09-16 11:50:15 +0200678 ClearPageDirty(p);
679 /* Trigger EIO in shmem: */
680 ClearPageUptodate(p);
681
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100682 if (!delete_from_lru_cache(p))
683 return DELAYED;
684 else
685 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200686}
687
688static int me_swapcache_clean(struct page *p, unsigned long pfn)
689{
Andi Kleen6a460792009-09-16 11:50:15 +0200690 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800691
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100692 if (!delete_from_lru_cache(p))
693 return RECOVERED;
694 else
695 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200696}
697
698/*
699 * Huge pages. Needs work.
700 * Issues:
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900701 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
702 * To narrow down kill region to one page, we need to break up pmd.
Andi Kleen6a460792009-09-16 11:50:15 +0200703 */
704static int me_huge_page(struct page *p, unsigned long pfn)
705{
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900706 int res = 0;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900707 struct page *hpage = compound_head(p);
708 /*
709 * We can safely recover from error on free or reserved (i.e.
710 * not in-use) hugepage by dequeuing it from freelist.
711 * To check whether a hugepage is in-use or not, we can't use
712 * page->lru because it can be used in other hugepage operations,
713 * such as __unmap_hugepage_range() and gather_surplus_pages().
714 * So instead we use page_mapping() and PageAnon().
715 * We assume that this function is called with page lock held,
716 * so there is no race between isolation and mapping/unmapping.
717 */
718 if (!(page_mapping(hpage) || PageAnon(hpage))) {
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900719 res = dequeue_hwpoisoned_huge_page(hpage);
720 if (!res)
721 return RECOVERED;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900722 }
723 return DELAYED;
Andi Kleen6a460792009-09-16 11:50:15 +0200724}
725
726/*
727 * Various page states we can handle.
728 *
729 * A page state is defined by its current page->flags bits.
730 * The table matches them in order and calls the right handler.
731 *
732 * This is quite tricky because we can access page at any time
733 * in its live cycle, so all accesses have to be extremly careful.
734 *
735 * This is not complete. More states could be added.
736 * For any missing state don't attempt recovery.
737 */
738
739#define dirty (1UL << PG_dirty)
740#define sc (1UL << PG_swapcache)
741#define unevict (1UL << PG_unevictable)
742#define mlock (1UL << PG_mlocked)
743#define writeback (1UL << PG_writeback)
744#define lru (1UL << PG_lru)
745#define swapbacked (1UL << PG_swapbacked)
746#define head (1UL << PG_head)
747#define tail (1UL << PG_tail)
748#define compound (1UL << PG_compound)
749#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200750#define reserved (1UL << PG_reserved)
751
752static struct page_state {
753 unsigned long mask;
754 unsigned long res;
755 char *msg;
756 int (*action)(struct page *p, unsigned long pfn);
757} error_states[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100758 { reserved, reserved, "reserved kernel", me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100759 /*
760 * free pages are specially detected outside this table:
761 * PG_buddy pages only make a small fraction of all free pages.
762 */
Andi Kleen6a460792009-09-16 11:50:15 +0200763
764 /*
765 * Could in theory check if slab page is free or if we can drop
766 * currently unused objects without touching them. But just
767 * treat it as standard kernel for now.
768 */
769 { slab, slab, "kernel slab", me_kernel },
770
771#ifdef CONFIG_PAGEFLAGS_EXTENDED
772 { head, head, "huge", me_huge_page },
773 { tail, tail, "huge", me_huge_page },
774#else
775 { compound, compound, "huge", me_huge_page },
776#endif
777
778 { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
779 { sc|dirty, sc, "swapcache", me_swapcache_clean },
780
781 { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
782 { unevict, unevict, "unevictable LRU", me_pagecache_clean},
783
Andi Kleen6a460792009-09-16 11:50:15 +0200784 { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
785 { mlock, mlock, "mlocked LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200786
787 { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
788 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200789
790 /*
791 * Catchall entry: must be at end.
792 */
793 { 0, 0, "unknown page state", me_unknown },
794};
795
Andi Kleen2326c462009-12-16 12:20:00 +0100796#undef dirty
797#undef sc
798#undef unevict
799#undef mlock
800#undef writeback
801#undef lru
802#undef swapbacked
803#undef head
804#undef tail
805#undef compound
806#undef slab
807#undef reserved
808
Andi Kleen6a460792009-09-16 11:50:15 +0200809static void action_result(unsigned long pfn, char *msg, int result)
810{
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100811 struct page *page = pfn_to_page(pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200812
813 printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
814 pfn,
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100815 PageDirty(page) ? "dirty " : "",
Andi Kleen6a460792009-09-16 11:50:15 +0200816 msg, action_name[result]);
817}
818
819static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100820 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200821{
822 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200823 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200824
825 result = ps->action(p, pfn);
826 action_result(pfn, ps->msg, result);
Wu Fengguang7456b042009-10-19 08:15:01 +0200827
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100828 count = page_count(p) - 1;
Wu Fengguang138ce282009-12-16 12:19:58 +0100829 if (ps->action == me_swapcache_dirty && result == DELAYED)
830 count--;
831 if (count != 0) {
Andi Kleen6a460792009-09-16 11:50:15 +0200832 printk(KERN_ERR
833 "MCE %#lx: %s page still referenced by %d users\n",
Wu Fengguang7456b042009-10-19 08:15:01 +0200834 pfn, ps->msg, count);
Wu Fengguang138ce282009-12-16 12:19:58 +0100835 result = FAILED;
836 }
Andi Kleen6a460792009-09-16 11:50:15 +0200837
838 /* Could do more checks here if page looks ok */
839 /*
840 * Could adjust zone counters here to correct for the missing page.
841 */
842
Wu Fengguang138ce282009-12-16 12:19:58 +0100843 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200844}
845
Andi Kleen6a460792009-09-16 11:50:15 +0200846/*
847 * Do all that is necessary to remove user space mappings. Unmap
848 * the pages and send SIGBUS to the processes if the data was dirty.
849 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100850static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Andi Kleen6a460792009-09-16 11:50:15 +0200851 int trapno)
852{
853 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
854 struct address_space *mapping;
855 LIST_HEAD(tokill);
856 int ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200857 int kill = 1;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900858 struct page *hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +0200859
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100860 if (PageReserved(p) || PageSlab(p))
861 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200862
Andi Kleen6a460792009-09-16 11:50:15 +0200863 /*
864 * This check implies we don't kill processes if their pages
865 * are in the swap cache early. Those are always late kills.
866 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900867 if (!page_mapped(hpage))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100868 return SWAP_SUCCESS;
869
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900870 if (PageKsm(p))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100871 return SWAP_FAIL;
Andi Kleen6a460792009-09-16 11:50:15 +0200872
873 if (PageSwapCache(p)) {
874 printk(KERN_ERR
875 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
876 ttu |= TTU_IGNORE_HWPOISON;
877 }
878
879 /*
880 * Propagate the dirty bit from PTEs to struct page first, because we
881 * need this to decide if we should kill or just drop the page.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100882 * XXX: the dirty test could be racy: set_page_dirty() may not always
883 * be called inside page lock (it's recommended but not enforced).
Andi Kleen6a460792009-09-16 11:50:15 +0200884 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900885 mapping = page_mapping(hpage);
886 if (!PageDirty(hpage) && mapping &&
887 mapping_cap_writeback_dirty(mapping)) {
888 if (page_mkclean(hpage)) {
889 SetPageDirty(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +0200890 } else {
891 kill = 0;
892 ttu |= TTU_IGNORE_HWPOISON;
893 printk(KERN_INFO
894 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
895 pfn);
896 }
897 }
898
899 /*
900 * First collect all the processes that have the page
901 * mapped in dirty form. This has to be done before try_to_unmap,
902 * because ttu takes the rmap data structures down.
903 *
904 * Error handling: We ignore errors here because
905 * there's nothing that can be done.
906 */
907 if (kill)
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900908 collect_procs(hpage, &tokill);
Andi Kleen6a460792009-09-16 11:50:15 +0200909
Andi Kleena08c80e2010-09-27 23:39:30 +0200910 ret = try_to_unmap(hpage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +0200911 if (ret != SWAP_SUCCESS)
912 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900913 pfn, page_mapcount(hpage));
Andi Kleen6a460792009-09-16 11:50:15 +0200914
915 /*
916 * Now that the dirty bit has been propagated to the
917 * struct page and all unmaps done we can decide if
918 * killing is needed or not. Only kill when the page
919 * was dirty, otherwise the tokill list is merely
920 * freed. When there was a problem unmapping earlier
921 * use a more force-full uncatchable kill to prevent
922 * any accesses to the poisoned memory.
923 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900924 kill_procs_ao(&tokill, !!PageDirty(hpage), trapno,
Andi Kleen0d9ee6a2010-09-27 22:03:33 +0200925 ret != SWAP_SUCCESS, p, pfn);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100926
927 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200928}
929
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900930static void set_page_hwpoison_huge_page(struct page *hpage)
931{
932 int i;
933 int nr_pages = 1 << compound_order(hpage);
934 for (i = 0; i < nr_pages; i++)
935 SetPageHWPoison(hpage + i);
936}
937
938static void clear_page_hwpoison_huge_page(struct page *hpage)
939{
940 int i;
941 int nr_pages = 1 << compound_order(hpage);
942 for (i = 0; i < nr_pages; i++)
943 ClearPageHWPoison(hpage + i);
944}
945
Andi Kleen82ba0112009-12-16 12:19:57 +0100946int __memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200947{
948 struct page_state *ps;
949 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900950 struct page *hpage;
Andi Kleen6a460792009-09-16 11:50:15 +0200951 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +0900952 unsigned int nr_pages;
Andi Kleen6a460792009-09-16 11:50:15 +0200953
954 if (!sysctl_memory_failure_recovery)
955 panic("Memory failure from trap %d on page %lx", trapno, pfn);
956
957 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100958 printk(KERN_ERR
959 "MCE %#lx: memory outside kernel control\n",
960 pfn);
961 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +0200962 }
963
964 p = pfn_to_page(pfn);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900965 hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +0200966 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100967 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200968 return 0;
969 }
970
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +0900971 nr_pages = 1 << compound_order(hpage);
972 atomic_long_add(nr_pages, &mce_bad_pages);
Andi Kleen6a460792009-09-16 11:50:15 +0200973
974 /*
975 * We need/can do nothing about count=0 pages.
976 * 1) it's a free page, and therefore in safe hand:
977 * prep_new_page() will be the gate keeper.
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +0900978 * 2) it's a free hugepage, which is also safe:
979 * an affected hugepage will be dequeued from hugepage freelist,
980 * so there's no concern about reusing it ever after.
981 * 3) it's part of a non-compound high order page.
Andi Kleen6a460792009-09-16 11:50:15 +0200982 * Implies some kernel user: cannot stop them from
983 * R/W the page; let's pray that the page has been
984 * used and will be freed some time later.
985 * In fact it's dangerous to directly bump up page count from 0,
986 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
987 */
Andi Kleen82ba0112009-12-16 12:19:57 +0100988 if (!(flags & MF_COUNT_INCREASED) &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900989 !get_page_unless_zero(hpage)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +0100990 if (is_free_buddy_page(p)) {
991 action_result(pfn, "free buddy", DELAYED);
992 return 0;
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +0900993 } else if (PageHuge(hpage)) {
994 /*
995 * Check "just unpoisoned", "filter hit", and
996 * "race with other subpage."
997 */
998 lock_page_nosync(hpage);
999 if (!PageHWPoison(hpage)
1000 || (hwpoison_filter(p) && TestClearPageHWPoison(p))
1001 || (p != hpage && TestSetPageHWPoison(hpage))) {
1002 atomic_long_sub(nr_pages, &mce_bad_pages);
1003 return 0;
1004 }
1005 set_page_hwpoison_huge_page(hpage);
1006 res = dequeue_hwpoisoned_huge_page(hpage);
1007 action_result(pfn, "free huge",
1008 res ? IGNORED : DELAYED);
1009 unlock_page(hpage);
1010 return res;
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001011 } else {
1012 action_result(pfn, "high order kernel", IGNORED);
1013 return -EBUSY;
1014 }
Andi Kleen6a460792009-09-16 11:50:15 +02001015 }
1016
1017 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001018 * We ignore non-LRU pages for good reasons.
1019 * - PG_locked is only well defined for LRU pages and a few others
1020 * - to avoid races with __set_page_locked()
1021 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1022 * The check (unnecessarily) ignores LRU pages being isolated and
1023 * walked by the page reclaim code, however that's not a big loss.
1024 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001025 if (!PageLRU(p) && !PageHuge(p))
Andi Kleenfacb6012009-12-16 12:20:00 +01001026 shake_page(p, 0);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001027 if (!PageLRU(p) && !PageHuge(p)) {
Andi Kleen0474a602009-12-16 12:20:00 +01001028 /*
1029 * shake_page could have turned it free.
1030 */
1031 if (is_free_buddy_page(p)) {
1032 action_result(pfn, "free buddy, 2nd try", DELAYED);
1033 return 0;
1034 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001035 action_result(pfn, "non LRU", IGNORED);
1036 put_page(p);
1037 return -EBUSY;
1038 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001039
1040 /*
Andi Kleen6a460792009-09-16 11:50:15 +02001041 * Lock the page and wait for writeback to finish.
1042 * It's very difficult to mess with pages currently under IO
1043 * and in many cases impossible, so we just avoid it here.
1044 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001045 lock_page_nosync(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001046
1047 /*
1048 * unpoison always clear PG_hwpoison inside page lock
1049 */
1050 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001051 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001052 res = 0;
1053 goto out;
1054 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001055 if (hwpoison_filter(p)) {
1056 if (TestClearPageHWPoison(p))
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001057 atomic_long_sub(nr_pages, &mce_bad_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001058 unlock_page(hpage);
1059 put_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001060 return 0;
1061 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001062
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001063 /*
1064 * For error on the tail page, we should set PG_hwpoison
1065 * on the head page to show that the hugepage is hwpoisoned
1066 */
1067 if (PageTail(p) && TestSetPageHWPoison(hpage)) {
1068 action_result(pfn, "hugepage already hardware poisoned",
1069 IGNORED);
1070 unlock_page(hpage);
1071 put_page(hpage);
1072 return 0;
1073 }
1074 /*
1075 * Set PG_hwpoison on all pages in an error hugepage,
1076 * because containment is done in hugepage unit for now.
1077 * Since we have done TestSetPageHWPoison() for the head page with
1078 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1079 */
1080 if (PageHuge(p))
1081 set_page_hwpoison_huge_page(hpage);
1082
Andi Kleen6a460792009-09-16 11:50:15 +02001083 wait_on_page_writeback(p);
1084
1085 /*
1086 * Now take care of user space mappings.
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001087 * Abort on fail: __remove_from_page_cache() assumes unmapped page.
Andi Kleen6a460792009-09-16 11:50:15 +02001088 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001089 if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
1090 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1091 res = -EBUSY;
1092 goto out;
1093 }
Andi Kleen6a460792009-09-16 11:50:15 +02001094
1095 /*
1096 * Torn down by someone else?
1097 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001098 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Andi Kleen6a460792009-09-16 11:50:15 +02001099 action_result(pfn, "already truncated LRU", IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001100 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001101 goto out;
1102 }
1103
1104 res = -EBUSY;
1105 for (ps = error_states;; ps++) {
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001106 if ((p->flags & ps->mask) == ps->res) {
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +01001107 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001108 break;
1109 }
1110 }
1111out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001112 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001113 return res;
1114}
1115EXPORT_SYMBOL_GPL(__memory_failure);
1116
1117/**
1118 * memory_failure - Handle memory failure of a page.
1119 * @pfn: Page Number of the corrupted page
1120 * @trapno: Trap number reported in the signal to user space.
1121 *
1122 * This function is called by the low level machine check code
1123 * of an architecture when it detects hardware memory corruption
1124 * of a page. It tries its best to recover, which includes
1125 * dropping pages, killing processes etc.
1126 *
1127 * The function is primarily of use for corruptions that
1128 * happen outside the current execution context (e.g. when
1129 * detected by a background scrubber)
1130 *
1131 * Must run in process context (e.g. a work queue) with interrupts
1132 * enabled and no spinlocks hold.
1133 */
1134void memory_failure(unsigned long pfn, int trapno)
1135{
1136 __memory_failure(pfn, trapno, 0);
1137}
Wu Fengguang847ce402009-12-16 12:19:58 +01001138
1139/**
1140 * unpoison_memory - Unpoison a previously poisoned page
1141 * @pfn: Page number of the to be unpoisoned page
1142 *
1143 * Software-unpoison a page that has been poisoned by
1144 * memory_failure() earlier.
1145 *
1146 * This is only done on the software-level, so it only works
1147 * for linux injected failures, not real hardware failures
1148 *
1149 * Returns 0 for success, otherwise -errno.
1150 */
1151int unpoison_memory(unsigned long pfn)
1152{
1153 struct page *page;
1154 struct page *p;
1155 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001156 unsigned int nr_pages;
Wu Fengguang847ce402009-12-16 12:19:58 +01001157
1158 if (!pfn_valid(pfn))
1159 return -ENXIO;
1160
1161 p = pfn_to_page(pfn);
1162 page = compound_head(p);
1163
1164 if (!PageHWPoison(p)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001165 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001166 return 0;
1167 }
1168
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001169 nr_pages = 1 << compound_order(page);
1170
Wu Fengguang847ce402009-12-16 12:19:58 +01001171 if (!get_page_unless_zero(page)) {
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001172 /*
1173 * Since HWPoisoned hugepage should have non-zero refcount,
1174 * race between memory failure and unpoison seems to happen.
1175 * In such case unpoison fails and memory failure runs
1176 * to the end.
1177 */
1178 if (PageHuge(page)) {
1179 pr_debug("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1180 return 0;
1181 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001182 if (TestClearPageHWPoison(p))
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001183 atomic_long_sub(nr_pages, &mce_bad_pages);
Andi Kleenfb46e732010-09-27 23:31:30 +02001184 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001185 return 0;
1186 }
1187
1188 lock_page_nosync(page);
1189 /*
1190 * This test is racy because PG_hwpoison is set outside of page lock.
1191 * That's acceptable because that won't trigger kernel panic. Instead,
1192 * the PG_hwpoison page will be caught and isolated on the entrance to
1193 * the free buddy page pool.
1194 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001195 if (TestClearPageHWPoison(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001196 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001197 atomic_long_sub(nr_pages, &mce_bad_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001198 freeit = 1;
Naoya Horiguchi6a901812010-09-08 10:19:40 +09001199 if (PageHuge(page))
1200 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001201 }
1202 unlock_page(page);
1203
1204 put_page(page);
1205 if (freeit)
1206 put_page(page);
1207
1208 return 0;
1209}
1210EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001211
1212static struct page *new_page(struct page *p, unsigned long private, int **x)
1213{
Andi Kleen12686d12009-12-16 12:20:01 +01001214 int nid = page_to_nid(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001215 if (PageHuge(p))
1216 return alloc_huge_page_node(page_hstate(compound_head(p)),
1217 nid);
1218 else
1219 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001220}
1221
1222/*
1223 * Safely get reference count of an arbitrary page.
1224 * Returns 0 for a free page, -EIO for a zero refcount page
1225 * that is not free, and 1 for any other page type.
1226 * For 1 the page is returned with increased page count, otherwise not.
1227 */
1228static int get_any_page(struct page *p, unsigned long pfn, int flags)
1229{
1230 int ret;
1231
1232 if (flags & MF_COUNT_INCREASED)
1233 return 1;
1234
1235 /*
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -08001236 * The lock_memory_hotplug prevents a race with memory hotplug.
Andi Kleenfacb6012009-12-16 12:20:00 +01001237 * This is a big hammer, a better would be nicer.
1238 */
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -08001239 lock_memory_hotplug();
Andi Kleenfacb6012009-12-16 12:20:00 +01001240
1241 /*
1242 * Isolate the page, so that it doesn't get reallocated if it
1243 * was free.
1244 */
1245 set_migratetype_isolate(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001246 /*
1247 * When the target page is a free hugepage, just remove it
1248 * from free hugepage list.
1249 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001250 if (!get_page_unless_zero(compound_head(p))) {
Naoya Horiguchid950b952010-09-08 10:19:39 +09001251 if (PageHuge(p)) {
Andi Kleen46e387b2010-10-22 17:40:48 +02001252 pr_info("get_any_page: %#lx free huge page\n", pfn);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001253 ret = dequeue_hwpoisoned_huge_page(compound_head(p));
1254 } else if (is_free_buddy_page(p)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001255 pr_info("get_any_page: %#lx free buddy page\n", pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001256 /* Set hwpoison bit while page is still isolated */
1257 SetPageHWPoison(p);
1258 ret = 0;
1259 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001260 pr_info("get_any_page: %#lx: unknown zero refcount page type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001261 pfn, p->flags);
1262 ret = -EIO;
1263 }
1264 } else {
1265 /* Not a free page */
1266 ret = 1;
1267 }
1268 unset_migratetype_isolate(p);
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -08001269 unlock_memory_hotplug();
Andi Kleenfacb6012009-12-16 12:20:00 +01001270 return ret;
1271}
1272
Naoya Horiguchid950b952010-09-08 10:19:39 +09001273static int soft_offline_huge_page(struct page *page, int flags)
1274{
1275 int ret;
1276 unsigned long pfn = page_to_pfn(page);
1277 struct page *hpage = compound_head(page);
1278 LIST_HEAD(pagelist);
1279
1280 ret = get_any_page(page, pfn, flags);
1281 if (ret < 0)
1282 return ret;
1283 if (ret == 0)
1284 goto done;
1285
1286 if (PageHWPoison(hpage)) {
1287 put_page(hpage);
1288 pr_debug("soft offline: %#lx hugepage already poisoned\n", pfn);
1289 return -EBUSY;
1290 }
1291
1292 /* Keep page count to indicate a given hugepage is isolated. */
1293
1294 list_add(&hpage->lru, &pagelist);
Mel Gorman77f1fe62011-01-13 15:45:57 -08001295 ret = migrate_huge_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0,
1296 true);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001297 if (ret) {
Mel Gorman77f1fe62011-01-13 15:45:57 -08001298 putback_lru_pages(&pagelist);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001299 pr_debug("soft offline: %#lx: migration failed %d, type %lx\n",
1300 pfn, ret, page->flags);
1301 if (ret > 0)
1302 ret = -EIO;
1303 return ret;
1304 }
1305done:
1306 if (!PageHWPoison(hpage))
1307 atomic_long_add(1 << compound_order(hpage), &mce_bad_pages);
1308 set_page_hwpoison_huge_page(hpage);
1309 dequeue_hwpoisoned_huge_page(hpage);
1310 /* keep elevated page count for bad page */
1311 return ret;
1312}
1313
Andi Kleenfacb6012009-12-16 12:20:00 +01001314/**
1315 * soft_offline_page - Soft offline a page.
1316 * @page: page to offline
1317 * @flags: flags. Same as memory_failure().
1318 *
1319 * Returns 0 on success, otherwise negated errno.
1320 *
1321 * Soft offline a page, by migration or invalidation,
1322 * without killing anything. This is for the case when
1323 * a page is not corrupted yet (so it's still valid to access),
1324 * but has had a number of corrected errors and is better taken
1325 * out.
1326 *
1327 * The actual policy on when to do that is maintained by
1328 * user space.
1329 *
1330 * This should never impact any application or cause data loss,
1331 * however it might take some time.
1332 *
1333 * This is not a 100% solution for all memory, but tries to be
1334 * ``good enough'' for the majority of memory.
1335 */
1336int soft_offline_page(struct page *page, int flags)
1337{
1338 int ret;
1339 unsigned long pfn = page_to_pfn(page);
1340
Naoya Horiguchid950b952010-09-08 10:19:39 +09001341 if (PageHuge(page))
1342 return soft_offline_huge_page(page, flags);
1343
Andi Kleenfacb6012009-12-16 12:20:00 +01001344 ret = get_any_page(page, pfn, flags);
1345 if (ret < 0)
1346 return ret;
1347 if (ret == 0)
1348 goto done;
1349
1350 /*
1351 * Page cache page we can handle?
1352 */
1353 if (!PageLRU(page)) {
1354 /*
1355 * Try to free it.
1356 */
1357 put_page(page);
1358 shake_page(page, 1);
1359
1360 /*
1361 * Did it turn free?
1362 */
1363 ret = get_any_page(page, pfn, 0);
1364 if (ret < 0)
1365 return ret;
1366 if (ret == 0)
1367 goto done;
1368 }
1369 if (!PageLRU(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001370 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001371 pfn, page->flags);
1372 return -EIO;
1373 }
1374
1375 lock_page(page);
1376 wait_on_page_writeback(page);
1377
1378 /*
1379 * Synchronized using the page lock with memory_failure()
1380 */
1381 if (PageHWPoison(page)) {
1382 unlock_page(page);
1383 put_page(page);
Andi Kleenfb46e732010-09-27 23:31:30 +02001384 pr_info("soft offline: %#lx page already poisoned\n", pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001385 return -EBUSY;
1386 }
1387
1388 /*
1389 * Try to invalidate first. This should work for
1390 * non dirty unmapped page cache pages.
1391 */
1392 ret = invalidate_inode_page(page);
1393 unlock_page(page);
1394
1395 /*
1396 * Drop count because page migration doesn't like raised
1397 * counts. The page could get re-allocated, but if it becomes
1398 * LRU the isolation will just fail.
1399 * RED-PEN would be better to keep it isolated here, but we
1400 * would need to fix isolation locking first.
1401 */
1402 put_page(page);
1403 if (ret == 1) {
1404 ret = 0;
Andi Kleenfb46e732010-09-27 23:31:30 +02001405 pr_info("soft_offline: %#lx: invalidated\n", pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001406 goto done;
1407 }
1408
1409 /*
1410 * Simple invalidation didn't work.
1411 * Try to migrate to a new page instead. migrate.c
1412 * handles a large number of cases for us.
1413 */
1414 ret = isolate_lru_page(page);
1415 if (!ret) {
1416 LIST_HEAD(pagelist);
1417
1418 list_add(&page->lru, &pagelist);
Mel Gorman77f1fe62011-01-13 15:45:57 -08001419 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1420 0, true);
Andi Kleenfacb6012009-12-16 12:20:00 +01001421 if (ret) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001422 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001423 pfn, ret, page->flags);
1424 if (ret > 0)
1425 ret = -EIO;
1426 }
1427 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001428 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001429 pfn, ret, page_count(page), page->flags);
1430 }
1431 if (ret)
1432 return ret;
1433
1434done:
1435 atomic_long_add(1, &mce_bad_pages);
1436 SetPageHWPoison(page);
1437 /* keep elevated page count for bad page */
1438 return ret;
1439}
Huang Yingbf998152010-05-31 14:28:19 +08001440
Huang Yingbbeb3402010-06-22 14:23:11 +08001441/*
1442 * The caller must hold current->mm->mmap_sem in read mode.
1443 */
Huang Yingbf998152010-05-31 14:28:19 +08001444int is_hwpoison_address(unsigned long addr)
1445{
1446 pgd_t *pgdp;
1447 pud_t pud, *pudp;
1448 pmd_t pmd, *pmdp;
1449 pte_t pte, *ptep;
1450 swp_entry_t entry;
1451
1452 pgdp = pgd_offset(current->mm, addr);
1453 if (!pgd_present(*pgdp))
1454 return 0;
1455 pudp = pud_offset(pgdp, addr);
1456 pud = *pudp;
1457 if (!pud_present(pud) || pud_large(pud))
1458 return 0;
1459 pmdp = pmd_offset(pudp, addr);
1460 pmd = *pmdp;
1461 if (!pmd_present(pmd) || pmd_large(pmd))
1462 return 0;
1463 ptep = pte_offset_map(pmdp, addr);
1464 pte = *ptep;
1465 pte_unmap(ptep);
1466 if (!is_swap_pte(pte))
1467 return 0;
1468 entry = pte_to_swp_entry(pte);
1469 return is_hwpoison_entry(entry);
1470}
1471EXPORT_SYMBOL_GPL(is_hwpoison_address);