blob: 124324134ff67b3c4a0bc06661b4f70f2406840f [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>
Andi Kleen6a460792009-09-16 11:50:15 +020054#include "internal.h"
55
56int sysctl_memory_failure_early_kill __read_mostly = 0;
57
58int sysctl_memory_failure_recovery __read_mostly = 1;
59
60atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
61
Andi Kleen27df5062009-12-21 19:56:42 +010062#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
63
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010064u32 hwpoison_filter_enable = 0;
Wu Fengguang7c116f22009-12-16 12:19:59 +010065u32 hwpoison_filter_dev_major = ~0U;
66u32 hwpoison_filter_dev_minor = ~0U;
Wu Fengguang478c5ff2009-12-16 12:19:59 +010067u64 hwpoison_filter_flags_mask;
68u64 hwpoison_filter_flags_value;
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010069EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
Wu Fengguang7c116f22009-12-16 12:19:59 +010070EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
71EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
Wu Fengguang478c5ff2009-12-16 12:19:59 +010072EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
73EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
Wu Fengguang7c116f22009-12-16 12:19:59 +010074
75static int hwpoison_filter_dev(struct page *p)
76{
77 struct address_space *mapping;
78 dev_t dev;
79
80 if (hwpoison_filter_dev_major == ~0U &&
81 hwpoison_filter_dev_minor == ~0U)
82 return 0;
83
84 /*
Andi Kleen1c80b992010-09-27 23:09:51 +020085 * page_mapping() does not accept slab pages.
Wu Fengguang7c116f22009-12-16 12:19:59 +010086 */
87 if (PageSlab(p))
88 return -EINVAL;
89
90 mapping = page_mapping(p);
91 if (mapping == NULL || mapping->host == NULL)
92 return -EINVAL;
93
94 dev = mapping->host->i_sb->s_dev;
95 if (hwpoison_filter_dev_major != ~0U &&
96 hwpoison_filter_dev_major != MAJOR(dev))
97 return -EINVAL;
98 if (hwpoison_filter_dev_minor != ~0U &&
99 hwpoison_filter_dev_minor != MINOR(dev))
100 return -EINVAL;
101
102 return 0;
103}
104
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100105static int hwpoison_filter_flags(struct page *p)
106{
107 if (!hwpoison_filter_flags_mask)
108 return 0;
109
110 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
111 hwpoison_filter_flags_value)
112 return 0;
113 else
114 return -EINVAL;
115}
116
Andi Kleen4fd466e2009-12-16 12:19:59 +0100117/*
118 * This allows stress tests to limit test scope to a collection of tasks
119 * by putting them under some memcg. This prevents killing unrelated/important
120 * processes such as /sbin/init. Note that the target task may share clean
121 * pages with init (eg. libc text), which is harmless. If the target task
122 * share _dirty_ pages with another task B, the test scheme must make sure B
123 * is also included in the memcg. At last, due to race conditions this filter
124 * can only guarantee that the page either belongs to the memcg tasks, or is
125 * a freed page.
126 */
127#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
128u64 hwpoison_filter_memcg;
129EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
130static int hwpoison_filter_task(struct page *p)
131{
132 struct mem_cgroup *mem;
133 struct cgroup_subsys_state *css;
134 unsigned long ino;
135
136 if (!hwpoison_filter_memcg)
137 return 0;
138
139 mem = try_get_mem_cgroup_from_page(p);
140 if (!mem)
141 return -EINVAL;
142
143 css = mem_cgroup_css(mem);
144 /* root_mem_cgroup has NULL dentries */
145 if (!css->cgroup->dentry)
146 return -EINVAL;
147
148 ino = css->cgroup->dentry->d_inode->i_ino;
149 css_put(css);
150
151 if (ino != hwpoison_filter_memcg)
152 return -EINVAL;
153
154 return 0;
155}
156#else
157static int hwpoison_filter_task(struct page *p) { return 0; }
158#endif
159
Wu Fengguang7c116f22009-12-16 12:19:59 +0100160int hwpoison_filter(struct page *p)
161{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100162 if (!hwpoison_filter_enable)
163 return 0;
164
Wu Fengguang7c116f22009-12-16 12:19:59 +0100165 if (hwpoison_filter_dev(p))
166 return -EINVAL;
167
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100168 if (hwpoison_filter_flags(p))
169 return -EINVAL;
170
Andi Kleen4fd466e2009-12-16 12:19:59 +0100171 if (hwpoison_filter_task(p))
172 return -EINVAL;
173
Wu Fengguang7c116f22009-12-16 12:19:59 +0100174 return 0;
175}
Andi Kleen27df5062009-12-21 19:56:42 +0100176#else
177int hwpoison_filter(struct page *p)
178{
179 return 0;
180}
181#endif
182
Wu Fengguang7c116f22009-12-16 12:19:59 +0100183EXPORT_SYMBOL_GPL(hwpoison_filter);
184
Andi Kleen6a460792009-09-16 11:50:15 +0200185/*
186 * Send all the processes who have the page mapped an ``action optional''
187 * signal.
188 */
189static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
Andi Kleen0d9ee6a2010-09-27 22:03:33 +0200190 unsigned long pfn, struct page *page)
Andi Kleen6a460792009-09-16 11:50:15 +0200191{
192 struct siginfo si;
193 int ret;
194
195 printk(KERN_ERR
196 "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
197 pfn, t->comm, t->pid);
198 si.si_signo = SIGBUS;
199 si.si_errno = 0;
200 si.si_code = BUS_MCEERR_AO;
201 si.si_addr = (void *)addr;
202#ifdef __ARCH_SI_TRAPNO
203 si.si_trapno = trapno;
204#endif
Andi Kleen0d9ee6a2010-09-27 22:03:33 +0200205 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
Andi Kleen6a460792009-09-16 11:50:15 +0200206 /*
207 * Don't use force here, it's convenient if the signal
208 * can be temporarily blocked.
209 * This could cause a loop when the user sets SIGBUS
210 * to SIG_IGN, but hopefully noone will do that?
211 */
212 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
213 if (ret < 0)
214 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
215 t->comm, t->pid, ret);
216 return ret;
217}
218
219/*
Andi Kleen588f9ce2009-12-16 12:19:57 +0100220 * When a unknown page type is encountered drain as many buffers as possible
221 * in the hope to turn the page into a LRU or free page, which we can handle.
222 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100223void shake_page(struct page *p, int access)
Andi Kleen588f9ce2009-12-16 12:19:57 +0100224{
225 if (!PageSlab(p)) {
226 lru_add_drain_all();
227 if (PageLRU(p))
228 return;
229 drain_all_pages();
230 if (PageLRU(p) || is_free_buddy_page(p))
231 return;
232 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100233
Andi Kleen588f9ce2009-12-16 12:19:57 +0100234 /*
Andi Kleenfacb6012009-12-16 12:20:00 +0100235 * Only all shrink_slab here (which would also
236 * shrink other caches) if access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100237 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100238 if (access) {
239 int nr;
240 do {
241 nr = shrink_slab(1000, GFP_KERNEL, 1000);
Andi Kleen47f43e72010-09-28 07:37:55 +0200242 if (page_count(p) == 1)
Andi Kleenfacb6012009-12-16 12:20:00 +0100243 break;
244 } while (nr > 10);
245 }
Andi Kleen588f9ce2009-12-16 12:19:57 +0100246}
247EXPORT_SYMBOL_GPL(shake_page);
248
249/*
Andi Kleen6a460792009-09-16 11:50:15 +0200250 * Kill all processes that have a poisoned page mapped and then isolate
251 * the page.
252 *
253 * General strategy:
254 * Find all processes having the page mapped and kill them.
255 * But we keep a page reference around so that the page is not
256 * actually freed yet.
257 * Then stash the page away
258 *
259 * There's no convenient way to get back to mapped processes
260 * from the VMAs. So do a brute-force search over all
261 * running processes.
262 *
263 * Remember that machine checks are not common (or rather
264 * if they are common you have other problems), so this shouldn't
265 * be a performance issue.
266 *
267 * Also there are some races possible while we get from the
268 * error detection to actually handle it.
269 */
270
271struct to_kill {
272 struct list_head nd;
273 struct task_struct *tsk;
274 unsigned long addr;
Andi Kleen9033ae12010-09-27 23:36:05 +0200275 char addr_valid;
Andi Kleen6a460792009-09-16 11:50:15 +0200276};
277
278/*
279 * Failure handling: if we can't find or can't kill a process there's
280 * not much we can do. We just print a message and ignore otherwise.
281 */
282
283/*
284 * Schedule a process for later kill.
285 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
286 * TBD would GFP_NOIO be enough?
287 */
288static void add_to_kill(struct task_struct *tsk, struct page *p,
289 struct vm_area_struct *vma,
290 struct list_head *to_kill,
291 struct to_kill **tkc)
292{
293 struct to_kill *tk;
294
295 if (*tkc) {
296 tk = *tkc;
297 *tkc = NULL;
298 } else {
299 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
300 if (!tk) {
301 printk(KERN_ERR
302 "MCE: Out of memory while machine check handling\n");
303 return;
304 }
305 }
306 tk->addr = page_address_in_vma(p, vma);
307 tk->addr_valid = 1;
308
309 /*
310 * In theory we don't have to kill when the page was
311 * munmaped. But it could be also a mremap. Since that's
312 * likely very rare kill anyways just out of paranoia, but use
313 * a SIGKILL because the error is not contained anymore.
314 */
315 if (tk->addr == -EFAULT) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200316 pr_info("MCE: Unable to find user space address %lx in %s\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200317 page_to_pfn(p), tsk->comm);
318 tk->addr_valid = 0;
319 }
320 get_task_struct(tsk);
321 tk->tsk = tsk;
322 list_add_tail(&tk->nd, to_kill);
323}
324
325/*
326 * Kill the processes that have been collected earlier.
327 *
328 * Only do anything when DOIT is set, otherwise just free the list
329 * (this is used for clean pages which do not need killing)
330 * Also when FAIL is set do a force kill because something went
331 * wrong earlier.
332 */
333static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
Andi Kleen0d9ee6a2010-09-27 22:03:33 +0200334 int fail, struct page *page, unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200335{
336 struct to_kill *tk, *next;
337
338 list_for_each_entry_safe (tk, next, to_kill, nd) {
339 if (doit) {
340 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200341 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200342 * make sure the process doesn't catch the
343 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200344 */
345 if (fail || tk->addr_valid == 0) {
346 printk(KERN_ERR
347 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
348 pfn, tk->tsk->comm, tk->tsk->pid);
349 force_sig(SIGKILL, tk->tsk);
350 }
351
352 /*
353 * In theory the process could have mapped
354 * something else on the address in-between. We could
355 * check for that, but we need to tell the
356 * process anyways.
357 */
358 else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
Andi Kleen0d9ee6a2010-09-27 22:03:33 +0200359 pfn, page) < 0)
Andi Kleen6a460792009-09-16 11:50:15 +0200360 printk(KERN_ERR
361 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
362 pfn, tk->tsk->comm, tk->tsk->pid);
363 }
364 put_task_struct(tk->tsk);
365 kfree(tk);
366 }
367}
368
369static int task_early_kill(struct task_struct *tsk)
370{
371 if (!tsk->mm)
372 return 0;
373 if (tsk->flags & PF_MCE_PROCESS)
374 return !!(tsk->flags & PF_MCE_EARLY);
375 return sysctl_memory_failure_early_kill;
376}
377
378/*
379 * Collect processes when the error hit an anonymous page.
380 */
381static void collect_procs_anon(struct page *page, struct list_head *to_kill,
382 struct to_kill **tkc)
383{
384 struct vm_area_struct *vma;
385 struct task_struct *tsk;
386 struct anon_vma *av;
387
388 read_lock(&tasklist_lock);
389 av = page_lock_anon_vma(page);
390 if (av == NULL) /* Not actually mapped anymore */
391 goto out;
392 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800393 struct anon_vma_chain *vmac;
394
Andi Kleen6a460792009-09-16 11:50:15 +0200395 if (!task_early_kill(tsk))
396 continue;
Rik van Riel5beb4932010-03-05 13:42:07 -0800397 list_for_each_entry(vmac, &av->head, same_anon_vma) {
398 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200399 if (!page_mapped_in_vma(page, vma))
400 continue;
401 if (vma->vm_mm == tsk->mm)
402 add_to_kill(tsk, page, vma, to_kill, tkc);
403 }
404 }
405 page_unlock_anon_vma(av);
406out:
407 read_unlock(&tasklist_lock);
408}
409
410/*
411 * Collect processes when the error hit a file mapped page.
412 */
413static void collect_procs_file(struct page *page, struct list_head *to_kill,
414 struct to_kill **tkc)
415{
416 struct vm_area_struct *vma;
417 struct task_struct *tsk;
418 struct prio_tree_iter iter;
419 struct address_space *mapping = page->mapping;
420
421 /*
422 * A note on the locking order between the two locks.
423 * We don't rely on this particular order.
424 * If you have some other code that needs a different order
425 * feel free to switch them around. Or add a reverse link
426 * from mm_struct to task_struct, then this could be all
427 * done without taking tasklist_lock and looping over all tasks.
428 */
429
430 read_lock(&tasklist_lock);
431 spin_lock(&mapping->i_mmap_lock);
432 for_each_process(tsk) {
433 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
434
435 if (!task_early_kill(tsk))
436 continue;
437
438 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
439 pgoff) {
440 /*
441 * Send early kill signal to tasks where a vma covers
442 * the page but the corrupted page is not necessarily
443 * mapped it in its pte.
444 * Assume applications who requested early kill want
445 * to be informed of all such data corruptions.
446 */
447 if (vma->vm_mm == tsk->mm)
448 add_to_kill(tsk, page, vma, to_kill, tkc);
449 }
450 }
451 spin_unlock(&mapping->i_mmap_lock);
452 read_unlock(&tasklist_lock);
453}
454
455/*
456 * Collect the processes who have the corrupted page mapped to kill.
457 * This is done in two steps for locking reasons.
458 * First preallocate one tokill structure outside the spin locks,
459 * so that we can kill at least one process reasonably reliable.
460 */
461static void collect_procs(struct page *page, struct list_head *tokill)
462{
463 struct to_kill *tk;
464
465 if (!page->mapping)
466 return;
467
468 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
469 if (!tk)
470 return;
471 if (PageAnon(page))
472 collect_procs_anon(page, tokill, &tk);
473 else
474 collect_procs_file(page, tokill, &tk);
475 kfree(tk);
476}
477
478/*
479 * Error handlers for various types of pages.
480 */
481
482enum outcome {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100483 IGNORED, /* Error: cannot be handled */
484 FAILED, /* Error: handling failed */
Andi Kleen6a460792009-09-16 11:50:15 +0200485 DELAYED, /* Will be handled later */
Andi Kleen6a460792009-09-16 11:50:15 +0200486 RECOVERED, /* Successfully recovered */
487};
488
489static const char *action_name[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100490 [IGNORED] = "Ignored",
Andi Kleen6a460792009-09-16 11:50:15 +0200491 [FAILED] = "Failed",
492 [DELAYED] = "Delayed",
Andi Kleen6a460792009-09-16 11:50:15 +0200493 [RECOVERED] = "Recovered",
494};
495
496/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100497 * XXX: It is possible that a page is isolated from LRU cache,
498 * and then kept in swap cache or failed to remove from page cache.
499 * The page count will stop it from being freed by unpoison.
500 * Stress tests should be aware of this memory leak problem.
501 */
502static int delete_from_lru_cache(struct page *p)
503{
504 if (!isolate_lru_page(p)) {
505 /*
506 * Clear sensible page flags, so that the buddy system won't
507 * complain when the page is unpoison-and-freed.
508 */
509 ClearPageActive(p);
510 ClearPageUnevictable(p);
511 /*
512 * drop the page count elevated by isolate_lru_page()
513 */
514 page_cache_release(p);
515 return 0;
516 }
517 return -EIO;
518}
519
520/*
Andi Kleen6a460792009-09-16 11:50:15 +0200521 * Error hit kernel page.
522 * Do nothing, try to be lucky and not touch this instead. For a few cases we
523 * could be more sophisticated.
524 */
525static int me_kernel(struct page *p, unsigned long pfn)
526{
Andi Kleen6a460792009-09-16 11:50:15 +0200527 return IGNORED;
528}
529
530/*
531 * Page in unknown state. Do nothing.
532 */
533static int me_unknown(struct page *p, unsigned long pfn)
534{
535 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
536 return FAILED;
537}
538
539/*
Andi Kleen6a460792009-09-16 11:50:15 +0200540 * Clean (or cleaned) page cache page.
541 */
542static int me_pagecache_clean(struct page *p, unsigned long pfn)
543{
544 int err;
545 int ret = FAILED;
546 struct address_space *mapping;
547
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100548 delete_from_lru_cache(p);
549
Andi Kleen6a460792009-09-16 11:50:15 +0200550 /*
551 * For anonymous pages we're done the only reference left
552 * should be the one m_f() holds.
553 */
554 if (PageAnon(p))
555 return RECOVERED;
556
557 /*
558 * Now truncate the page in the page cache. This is really
559 * more like a "temporary hole punch"
560 * Don't do this for block devices when someone else
561 * has a reference, because it could be file system metadata
562 * and that's not safe to truncate.
563 */
564 mapping = page_mapping(p);
565 if (!mapping) {
566 /*
567 * Page has been teared down in the meanwhile
568 */
569 return FAILED;
570 }
571
572 /*
573 * Truncation is a bit tricky. Enable it per file system for now.
574 *
575 * Open: to take i_mutex or not for this? Right now we don't.
576 */
577 if (mapping->a_ops->error_remove_page) {
578 err = mapping->a_ops->error_remove_page(mapping, p);
579 if (err != 0) {
580 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
581 pfn, err);
582 } else if (page_has_private(p) &&
583 !try_to_release_page(p, GFP_NOIO)) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200584 pr_info("MCE %#lx: failed to release buffers\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200585 } else {
586 ret = RECOVERED;
587 }
588 } else {
589 /*
590 * If the file system doesn't support it just invalidate
591 * This fails on dirty or anything with private pages
592 */
593 if (invalidate_inode_page(p))
594 ret = RECOVERED;
595 else
596 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
597 pfn);
598 }
599 return ret;
600}
601
602/*
603 * Dirty cache page page
604 * Issues: when the error hit a hole page the error is not properly
605 * propagated.
606 */
607static int me_pagecache_dirty(struct page *p, unsigned long pfn)
608{
609 struct address_space *mapping = page_mapping(p);
610
611 SetPageError(p);
612 /* TBD: print more information about the file. */
613 if (mapping) {
614 /*
615 * IO error will be reported by write(), fsync(), etc.
616 * who check the mapping.
617 * This way the application knows that something went
618 * wrong with its dirty file data.
619 *
620 * There's one open issue:
621 *
622 * The EIO will be only reported on the next IO
623 * operation and then cleared through the IO map.
624 * Normally Linux has two mechanisms to pass IO error
625 * first through the AS_EIO flag in the address space
626 * and then through the PageError flag in the page.
627 * Since we drop pages on memory failure handling the
628 * only mechanism open to use is through AS_AIO.
629 *
630 * This has the disadvantage that it gets cleared on
631 * the first operation that returns an error, while
632 * the PageError bit is more sticky and only cleared
633 * when the page is reread or dropped. If an
634 * application assumes it will always get error on
635 * fsync, but does other operations on the fd before
636 * and the page is dropped inbetween then the error
637 * will not be properly reported.
638 *
639 * This can already happen even without hwpoisoned
640 * pages: first on metadata IO errors (which only
641 * report through AS_EIO) or when the page is dropped
642 * at the wrong time.
643 *
644 * So right now we assume that the application DTRT on
645 * the first EIO, but we're not worse than other parts
646 * of the kernel.
647 */
648 mapping_set_error(mapping, EIO);
649 }
650
651 return me_pagecache_clean(p, pfn);
652}
653
654/*
655 * Clean and dirty swap cache.
656 *
657 * Dirty swap cache page is tricky to handle. The page could live both in page
658 * cache and swap cache(ie. page is freshly swapped in). So it could be
659 * referenced concurrently by 2 types of PTEs:
660 * normal PTEs and swap PTEs. We try to handle them consistently by calling
661 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
662 * and then
663 * - clear dirty bit to prevent IO
664 * - remove from LRU
665 * - but keep in the swap cache, so that when we return to it on
666 * a later page fault, we know the application is accessing
667 * corrupted data and shall be killed (we installed simple
668 * interception code in do_swap_page to catch it).
669 *
670 * Clean swap cache pages can be directly isolated. A later page fault will
671 * bring in the known good data from disk.
672 */
673static int me_swapcache_dirty(struct page *p, unsigned long pfn)
674{
Andi Kleen6a460792009-09-16 11:50:15 +0200675 ClearPageDirty(p);
676 /* Trigger EIO in shmem: */
677 ClearPageUptodate(p);
678
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100679 if (!delete_from_lru_cache(p))
680 return DELAYED;
681 else
682 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200683}
684
685static int me_swapcache_clean(struct page *p, unsigned long pfn)
686{
Andi Kleen6a460792009-09-16 11:50:15 +0200687 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800688
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100689 if (!delete_from_lru_cache(p))
690 return RECOVERED;
691 else
692 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200693}
694
695/*
696 * Huge pages. Needs work.
697 * Issues:
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900698 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
699 * To narrow down kill region to one page, we need to break up pmd.
Andi Kleen6a460792009-09-16 11:50:15 +0200700 */
701static int me_huge_page(struct page *p, unsigned long pfn)
702{
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900703 int res = 0;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900704 struct page *hpage = compound_head(p);
705 /*
706 * We can safely recover from error on free or reserved (i.e.
707 * not in-use) hugepage by dequeuing it from freelist.
708 * To check whether a hugepage is in-use or not, we can't use
709 * page->lru because it can be used in other hugepage operations,
710 * such as __unmap_hugepage_range() and gather_surplus_pages().
711 * So instead we use page_mapping() and PageAnon().
712 * We assume that this function is called with page lock held,
713 * so there is no race between isolation and mapping/unmapping.
714 */
715 if (!(page_mapping(hpage) || PageAnon(hpage))) {
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900716 res = dequeue_hwpoisoned_huge_page(hpage);
717 if (!res)
718 return RECOVERED;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900719 }
720 return DELAYED;
Andi Kleen6a460792009-09-16 11:50:15 +0200721}
722
723/*
724 * Various page states we can handle.
725 *
726 * A page state is defined by its current page->flags bits.
727 * The table matches them in order and calls the right handler.
728 *
729 * This is quite tricky because we can access page at any time
730 * in its live cycle, so all accesses have to be extremly careful.
731 *
732 * This is not complete. More states could be added.
733 * For any missing state don't attempt recovery.
734 */
735
736#define dirty (1UL << PG_dirty)
737#define sc (1UL << PG_swapcache)
738#define unevict (1UL << PG_unevictable)
739#define mlock (1UL << PG_mlocked)
740#define writeback (1UL << PG_writeback)
741#define lru (1UL << PG_lru)
742#define swapbacked (1UL << PG_swapbacked)
743#define head (1UL << PG_head)
744#define tail (1UL << PG_tail)
745#define compound (1UL << PG_compound)
746#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200747#define reserved (1UL << PG_reserved)
748
749static struct page_state {
750 unsigned long mask;
751 unsigned long res;
752 char *msg;
753 int (*action)(struct page *p, unsigned long pfn);
754} error_states[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100755 { reserved, reserved, "reserved kernel", me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100756 /*
757 * free pages are specially detected outside this table:
758 * PG_buddy pages only make a small fraction of all free pages.
759 */
Andi Kleen6a460792009-09-16 11:50:15 +0200760
761 /*
762 * Could in theory check if slab page is free or if we can drop
763 * currently unused objects without touching them. But just
764 * treat it as standard kernel for now.
765 */
766 { slab, slab, "kernel slab", me_kernel },
767
768#ifdef CONFIG_PAGEFLAGS_EXTENDED
769 { head, head, "huge", me_huge_page },
770 { tail, tail, "huge", me_huge_page },
771#else
772 { compound, compound, "huge", me_huge_page },
773#endif
774
775 { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
776 { sc|dirty, sc, "swapcache", me_swapcache_clean },
777
778 { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
779 { unevict, unevict, "unevictable LRU", me_pagecache_clean},
780
Andi Kleen6a460792009-09-16 11:50:15 +0200781 { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
782 { mlock, mlock, "mlocked LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200783
784 { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
785 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200786
787 /*
788 * Catchall entry: must be at end.
789 */
790 { 0, 0, "unknown page state", me_unknown },
791};
792
Andi Kleen2326c462009-12-16 12:20:00 +0100793#undef dirty
794#undef sc
795#undef unevict
796#undef mlock
797#undef writeback
798#undef lru
799#undef swapbacked
800#undef head
801#undef tail
802#undef compound
803#undef slab
804#undef reserved
805
Andi Kleen6a460792009-09-16 11:50:15 +0200806static void action_result(unsigned long pfn, char *msg, int result)
807{
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100808 struct page *page = pfn_to_page(pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200809
810 printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
811 pfn,
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100812 PageDirty(page) ? "dirty " : "",
Andi Kleen6a460792009-09-16 11:50:15 +0200813 msg, action_name[result]);
814}
815
816static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100817 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200818{
819 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200820 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200821
822 result = ps->action(p, pfn);
823 action_result(pfn, ps->msg, result);
Wu Fengguang7456b042009-10-19 08:15:01 +0200824
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100825 count = page_count(p) - 1;
Wu Fengguang138ce282009-12-16 12:19:58 +0100826 if (ps->action == me_swapcache_dirty && result == DELAYED)
827 count--;
828 if (count != 0) {
Andi Kleen6a460792009-09-16 11:50:15 +0200829 printk(KERN_ERR
830 "MCE %#lx: %s page still referenced by %d users\n",
Wu Fengguang7456b042009-10-19 08:15:01 +0200831 pfn, ps->msg, count);
Wu Fengguang138ce282009-12-16 12:19:58 +0100832 result = FAILED;
833 }
Andi Kleen6a460792009-09-16 11:50:15 +0200834
835 /* Could do more checks here if page looks ok */
836 /*
837 * Could adjust zone counters here to correct for the missing page.
838 */
839
Wu Fengguang138ce282009-12-16 12:19:58 +0100840 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200841}
842
Andi Kleen6a460792009-09-16 11:50:15 +0200843/*
844 * Do all that is necessary to remove user space mappings. Unmap
845 * the pages and send SIGBUS to the processes if the data was dirty.
846 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100847static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Andi Kleen6a460792009-09-16 11:50:15 +0200848 int trapno)
849{
850 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
851 struct address_space *mapping;
852 LIST_HEAD(tokill);
853 int ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200854 int kill = 1;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900855 struct page *hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +0200856
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100857 if (PageReserved(p) || PageSlab(p))
858 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200859
Andi Kleen6a460792009-09-16 11:50:15 +0200860 /*
861 * This check implies we don't kill processes if their pages
862 * are in the swap cache early. Those are always late kills.
863 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900864 if (!page_mapped(hpage))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100865 return SWAP_SUCCESS;
866
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900867 if (PageKsm(p))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100868 return SWAP_FAIL;
Andi Kleen6a460792009-09-16 11:50:15 +0200869
870 if (PageSwapCache(p)) {
871 printk(KERN_ERR
872 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
873 ttu |= TTU_IGNORE_HWPOISON;
874 }
875
876 /*
877 * Propagate the dirty bit from PTEs to struct page first, because we
878 * need this to decide if we should kill or just drop the page.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100879 * XXX: the dirty test could be racy: set_page_dirty() may not always
880 * be called inside page lock (it's recommended but not enforced).
Andi Kleen6a460792009-09-16 11:50:15 +0200881 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900882 mapping = page_mapping(hpage);
883 if (!PageDirty(hpage) && mapping &&
884 mapping_cap_writeback_dirty(mapping)) {
885 if (page_mkclean(hpage)) {
886 SetPageDirty(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +0200887 } else {
888 kill = 0;
889 ttu |= TTU_IGNORE_HWPOISON;
890 printk(KERN_INFO
891 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
892 pfn);
893 }
894 }
895
896 /*
897 * First collect all the processes that have the page
898 * mapped in dirty form. This has to be done before try_to_unmap,
899 * because ttu takes the rmap data structures down.
900 *
901 * Error handling: We ignore errors here because
902 * there's nothing that can be done.
903 */
904 if (kill)
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900905 collect_procs(hpage, &tokill);
Andi Kleen6a460792009-09-16 11:50:15 +0200906
Andi Kleena08c80e2010-09-27 23:39:30 +0200907 ret = try_to_unmap(hpage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +0200908 if (ret != SWAP_SUCCESS)
909 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900910 pfn, page_mapcount(hpage));
Andi Kleen6a460792009-09-16 11:50:15 +0200911
912 /*
913 * Now that the dirty bit has been propagated to the
914 * struct page and all unmaps done we can decide if
915 * killing is needed or not. Only kill when the page
916 * was dirty, otherwise the tokill list is merely
917 * freed. When there was a problem unmapping earlier
918 * use a more force-full uncatchable kill to prevent
919 * any accesses to the poisoned memory.
920 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900921 kill_procs_ao(&tokill, !!PageDirty(hpage), trapno,
Andi Kleen0d9ee6a2010-09-27 22:03:33 +0200922 ret != SWAP_SUCCESS, p, pfn);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100923
924 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200925}
926
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900927static void set_page_hwpoison_huge_page(struct page *hpage)
928{
929 int i;
930 int nr_pages = 1 << compound_order(hpage);
931 for (i = 0; i < nr_pages; i++)
932 SetPageHWPoison(hpage + i);
933}
934
935static void clear_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 ClearPageHWPoison(hpage + i);
941}
942
Andi Kleen82ba0112009-12-16 12:19:57 +0100943int __memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200944{
945 struct page_state *ps;
946 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900947 struct page *hpage;
Andi Kleen6a460792009-09-16 11:50:15 +0200948 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +0900949 unsigned int nr_pages;
Andi Kleen6a460792009-09-16 11:50:15 +0200950
951 if (!sysctl_memory_failure_recovery)
952 panic("Memory failure from trap %d on page %lx", trapno, pfn);
953
954 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100955 printk(KERN_ERR
956 "MCE %#lx: memory outside kernel control\n",
957 pfn);
958 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +0200959 }
960
961 p = pfn_to_page(pfn);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900962 hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +0200963 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100964 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200965 return 0;
966 }
967
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +0900968 nr_pages = 1 << compound_order(hpage);
969 atomic_long_add(nr_pages, &mce_bad_pages);
Andi Kleen6a460792009-09-16 11:50:15 +0200970
971 /*
972 * We need/can do nothing about count=0 pages.
973 * 1) it's a free page, and therefore in safe hand:
974 * prep_new_page() will be the gate keeper.
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +0900975 * 2) it's a free hugepage, which is also safe:
976 * an affected hugepage will be dequeued from hugepage freelist,
977 * so there's no concern about reusing it ever after.
978 * 3) it's part of a non-compound high order page.
Andi Kleen6a460792009-09-16 11:50:15 +0200979 * Implies some kernel user: cannot stop them from
980 * R/W the page; let's pray that the page has been
981 * used and will be freed some time later.
982 * In fact it's dangerous to directly bump up page count from 0,
983 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
984 */
Andi Kleen82ba0112009-12-16 12:19:57 +0100985 if (!(flags & MF_COUNT_INCREASED) &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900986 !get_page_unless_zero(hpage)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +0100987 if (is_free_buddy_page(p)) {
988 action_result(pfn, "free buddy", DELAYED);
989 return 0;
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +0900990 } else if (PageHuge(hpage)) {
991 /*
992 * Check "just unpoisoned", "filter hit", and
993 * "race with other subpage."
994 */
995 lock_page_nosync(hpage);
996 if (!PageHWPoison(hpage)
997 || (hwpoison_filter(p) && TestClearPageHWPoison(p))
998 || (p != hpage && TestSetPageHWPoison(hpage))) {
999 atomic_long_sub(nr_pages, &mce_bad_pages);
1000 return 0;
1001 }
1002 set_page_hwpoison_huge_page(hpage);
1003 res = dequeue_hwpoisoned_huge_page(hpage);
1004 action_result(pfn, "free huge",
1005 res ? IGNORED : DELAYED);
1006 unlock_page(hpage);
1007 return res;
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001008 } else {
1009 action_result(pfn, "high order kernel", IGNORED);
1010 return -EBUSY;
1011 }
Andi Kleen6a460792009-09-16 11:50:15 +02001012 }
1013
1014 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001015 * We ignore non-LRU pages for good reasons.
1016 * - PG_locked is only well defined for LRU pages and a few others
1017 * - to avoid races with __set_page_locked()
1018 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1019 * The check (unnecessarily) ignores LRU pages being isolated and
1020 * walked by the page reclaim code, however that's not a big loss.
1021 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001022 if (!PageLRU(p) && !PageHuge(p))
Andi Kleenfacb6012009-12-16 12:20:00 +01001023 shake_page(p, 0);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001024 if (!PageLRU(p) && !PageHuge(p)) {
Andi Kleen0474a602009-12-16 12:20:00 +01001025 /*
1026 * shake_page could have turned it free.
1027 */
1028 if (is_free_buddy_page(p)) {
1029 action_result(pfn, "free buddy, 2nd try", DELAYED);
1030 return 0;
1031 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001032 action_result(pfn, "non LRU", IGNORED);
1033 put_page(p);
1034 return -EBUSY;
1035 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001036
1037 /*
Andi Kleen6a460792009-09-16 11:50:15 +02001038 * Lock the page and wait for writeback to finish.
1039 * It's very difficult to mess with pages currently under IO
1040 * and in many cases impossible, so we just avoid it here.
1041 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001042 lock_page_nosync(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001043
1044 /*
1045 * unpoison always clear PG_hwpoison inside page lock
1046 */
1047 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001048 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001049 res = 0;
1050 goto out;
1051 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001052 if (hwpoison_filter(p)) {
1053 if (TestClearPageHWPoison(p))
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001054 atomic_long_sub(nr_pages, &mce_bad_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001055 unlock_page(hpage);
1056 put_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001057 return 0;
1058 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001059
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001060 /*
1061 * For error on the tail page, we should set PG_hwpoison
1062 * on the head page to show that the hugepage is hwpoisoned
1063 */
1064 if (PageTail(p) && TestSetPageHWPoison(hpage)) {
1065 action_result(pfn, "hugepage already hardware poisoned",
1066 IGNORED);
1067 unlock_page(hpage);
1068 put_page(hpage);
1069 return 0;
1070 }
1071 /*
1072 * Set PG_hwpoison on all pages in an error hugepage,
1073 * because containment is done in hugepage unit for now.
1074 * Since we have done TestSetPageHWPoison() for the head page with
1075 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1076 */
1077 if (PageHuge(p))
1078 set_page_hwpoison_huge_page(hpage);
1079
Andi Kleen6a460792009-09-16 11:50:15 +02001080 wait_on_page_writeback(p);
1081
1082 /*
1083 * Now take care of user space mappings.
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001084 * Abort on fail: __remove_from_page_cache() assumes unmapped page.
Andi Kleen6a460792009-09-16 11:50:15 +02001085 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001086 if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
1087 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1088 res = -EBUSY;
1089 goto out;
1090 }
Andi Kleen6a460792009-09-16 11:50:15 +02001091
1092 /*
1093 * Torn down by someone else?
1094 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001095 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Andi Kleen6a460792009-09-16 11:50:15 +02001096 action_result(pfn, "already truncated LRU", IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001097 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001098 goto out;
1099 }
1100
1101 res = -EBUSY;
1102 for (ps = error_states;; ps++) {
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001103 if ((p->flags & ps->mask) == ps->res) {
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +01001104 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001105 break;
1106 }
1107 }
1108out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001109 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001110 return res;
1111}
1112EXPORT_SYMBOL_GPL(__memory_failure);
1113
1114/**
1115 * memory_failure - Handle memory failure of a page.
1116 * @pfn: Page Number of the corrupted page
1117 * @trapno: Trap number reported in the signal to user space.
1118 *
1119 * This function is called by the low level machine check code
1120 * of an architecture when it detects hardware memory corruption
1121 * of a page. It tries its best to recover, which includes
1122 * dropping pages, killing processes etc.
1123 *
1124 * The function is primarily of use for corruptions that
1125 * happen outside the current execution context (e.g. when
1126 * detected by a background scrubber)
1127 *
1128 * Must run in process context (e.g. a work queue) with interrupts
1129 * enabled and no spinlocks hold.
1130 */
1131void memory_failure(unsigned long pfn, int trapno)
1132{
1133 __memory_failure(pfn, trapno, 0);
1134}
Wu Fengguang847ce402009-12-16 12:19:58 +01001135
1136/**
1137 * unpoison_memory - Unpoison a previously poisoned page
1138 * @pfn: Page number of the to be unpoisoned page
1139 *
1140 * Software-unpoison a page that has been poisoned by
1141 * memory_failure() earlier.
1142 *
1143 * This is only done on the software-level, so it only works
1144 * for linux injected failures, not real hardware failures
1145 *
1146 * Returns 0 for success, otherwise -errno.
1147 */
1148int unpoison_memory(unsigned long pfn)
1149{
1150 struct page *page;
1151 struct page *p;
1152 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001153 unsigned int nr_pages;
Wu Fengguang847ce402009-12-16 12:19:58 +01001154
1155 if (!pfn_valid(pfn))
1156 return -ENXIO;
1157
1158 p = pfn_to_page(pfn);
1159 page = compound_head(p);
1160
1161 if (!PageHWPoison(p)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001162 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001163 return 0;
1164 }
1165
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001166 nr_pages = 1 << compound_order(page);
1167
Wu Fengguang847ce402009-12-16 12:19:58 +01001168 if (!get_page_unless_zero(page)) {
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001169 /*
1170 * Since HWPoisoned hugepage should have non-zero refcount,
1171 * race between memory failure and unpoison seems to happen.
1172 * In such case unpoison fails and memory failure runs
1173 * to the end.
1174 */
1175 if (PageHuge(page)) {
1176 pr_debug("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1177 return 0;
1178 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001179 if (TestClearPageHWPoison(p))
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001180 atomic_long_sub(nr_pages, &mce_bad_pages);
Andi Kleenfb46e732010-09-27 23:31:30 +02001181 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001182 return 0;
1183 }
1184
1185 lock_page_nosync(page);
1186 /*
1187 * This test is racy because PG_hwpoison is set outside of page lock.
1188 * That's acceptable because that won't trigger kernel panic. Instead,
1189 * the PG_hwpoison page will be caught and isolated on the entrance to
1190 * the free buddy page pool.
1191 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001192 if (TestClearPageHWPoison(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001193 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001194 atomic_long_sub(nr_pages, &mce_bad_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001195 freeit = 1;
Naoya Horiguchi6a901812010-09-08 10:19:40 +09001196 if (PageHuge(page))
1197 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001198 }
1199 unlock_page(page);
1200
1201 put_page(page);
1202 if (freeit)
1203 put_page(page);
1204
1205 return 0;
1206}
1207EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001208
1209static struct page *new_page(struct page *p, unsigned long private, int **x)
1210{
Andi Kleen12686d12009-12-16 12:20:01 +01001211 int nid = page_to_nid(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001212 if (PageHuge(p))
1213 return alloc_huge_page_node(page_hstate(compound_head(p)),
1214 nid);
1215 else
1216 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001217}
1218
1219/*
1220 * Safely get reference count of an arbitrary page.
1221 * Returns 0 for a free page, -EIO for a zero refcount page
1222 * that is not free, and 1 for any other page type.
1223 * For 1 the page is returned with increased page count, otherwise not.
1224 */
1225static int get_any_page(struct page *p, unsigned long pfn, int flags)
1226{
1227 int ret;
1228
1229 if (flags & MF_COUNT_INCREASED)
1230 return 1;
1231
1232 /*
1233 * The lock_system_sleep prevents a race with memory hotplug,
1234 * because the isolation assumes there's only a single user.
1235 * This is a big hammer, a better would be nicer.
1236 */
1237 lock_system_sleep();
1238
1239 /*
1240 * Isolate the page, so that it doesn't get reallocated if it
1241 * was free.
1242 */
1243 set_migratetype_isolate(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001244 /*
1245 * When the target page is a free hugepage, just remove it
1246 * from free hugepage list.
1247 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001248 if (!get_page_unless_zero(compound_head(p))) {
Naoya Horiguchid950b952010-09-08 10:19:39 +09001249 if (PageHuge(p)) {
Andi Kleen46e387b2010-10-22 17:40:48 +02001250 pr_info("get_any_page: %#lx free huge page\n", pfn);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001251 ret = dequeue_hwpoisoned_huge_page(compound_head(p));
1252 } else if (is_free_buddy_page(p)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001253 pr_info("get_any_page: %#lx free buddy page\n", pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001254 /* Set hwpoison bit while page is still isolated */
1255 SetPageHWPoison(p);
1256 ret = 0;
1257 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001258 pr_info("get_any_page: %#lx: unknown zero refcount page type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001259 pfn, p->flags);
1260 ret = -EIO;
1261 }
1262 } else {
1263 /* Not a free page */
1264 ret = 1;
1265 }
1266 unset_migratetype_isolate(p);
1267 unlock_system_sleep();
1268 return ret;
1269}
1270
Naoya Horiguchid950b952010-09-08 10:19:39 +09001271static int soft_offline_huge_page(struct page *page, int flags)
1272{
1273 int ret;
1274 unsigned long pfn = page_to_pfn(page);
1275 struct page *hpage = compound_head(page);
1276 LIST_HEAD(pagelist);
1277
1278 ret = get_any_page(page, pfn, flags);
1279 if (ret < 0)
1280 return ret;
1281 if (ret == 0)
1282 goto done;
1283
1284 if (PageHWPoison(hpage)) {
1285 put_page(hpage);
1286 pr_debug("soft offline: %#lx hugepage already poisoned\n", pfn);
1287 return -EBUSY;
1288 }
1289
1290 /* Keep page count to indicate a given hugepage is isolated. */
1291
1292 list_add(&hpage->lru, &pagelist);
1293 ret = migrate_huge_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0);
1294 if (ret) {
Minchan Kimcf608ac2010-10-26 14:21:29 -07001295 putback_lru_pages(&pagelist);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001296 pr_debug("soft offline: %#lx: migration failed %d, type %lx\n",
1297 pfn, ret, page->flags);
1298 if (ret > 0)
1299 ret = -EIO;
1300 return ret;
1301 }
1302done:
1303 if (!PageHWPoison(hpage))
1304 atomic_long_add(1 << compound_order(hpage), &mce_bad_pages);
1305 set_page_hwpoison_huge_page(hpage);
1306 dequeue_hwpoisoned_huge_page(hpage);
1307 /* keep elevated page count for bad page */
1308 return ret;
1309}
1310
Andi Kleenfacb6012009-12-16 12:20:00 +01001311/**
1312 * soft_offline_page - Soft offline a page.
1313 * @page: page to offline
1314 * @flags: flags. Same as memory_failure().
1315 *
1316 * Returns 0 on success, otherwise negated errno.
1317 *
1318 * Soft offline a page, by migration or invalidation,
1319 * without killing anything. This is for the case when
1320 * a page is not corrupted yet (so it's still valid to access),
1321 * but has had a number of corrected errors and is better taken
1322 * out.
1323 *
1324 * The actual policy on when to do that is maintained by
1325 * user space.
1326 *
1327 * This should never impact any application or cause data loss,
1328 * however it might take some time.
1329 *
1330 * This is not a 100% solution for all memory, but tries to be
1331 * ``good enough'' for the majority of memory.
1332 */
1333int soft_offline_page(struct page *page, int flags)
1334{
1335 int ret;
1336 unsigned long pfn = page_to_pfn(page);
1337
Naoya Horiguchid950b952010-09-08 10:19:39 +09001338 if (PageHuge(page))
1339 return soft_offline_huge_page(page, flags);
1340
Andi Kleenfacb6012009-12-16 12:20:00 +01001341 ret = get_any_page(page, pfn, flags);
1342 if (ret < 0)
1343 return ret;
1344 if (ret == 0)
1345 goto done;
1346
1347 /*
1348 * Page cache page we can handle?
1349 */
1350 if (!PageLRU(page)) {
1351 /*
1352 * Try to free it.
1353 */
1354 put_page(page);
1355 shake_page(page, 1);
1356
1357 /*
1358 * Did it turn free?
1359 */
1360 ret = get_any_page(page, pfn, 0);
1361 if (ret < 0)
1362 return ret;
1363 if (ret == 0)
1364 goto done;
1365 }
1366 if (!PageLRU(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001367 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001368 pfn, page->flags);
1369 return -EIO;
1370 }
1371
1372 lock_page(page);
1373 wait_on_page_writeback(page);
1374
1375 /*
1376 * Synchronized using the page lock with memory_failure()
1377 */
1378 if (PageHWPoison(page)) {
1379 unlock_page(page);
1380 put_page(page);
Andi Kleenfb46e732010-09-27 23:31:30 +02001381 pr_info("soft offline: %#lx page already poisoned\n", pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001382 return -EBUSY;
1383 }
1384
1385 /*
1386 * Try to invalidate first. This should work for
1387 * non dirty unmapped page cache pages.
1388 */
1389 ret = invalidate_inode_page(page);
1390 unlock_page(page);
1391
1392 /*
1393 * Drop count because page migration doesn't like raised
1394 * counts. The page could get re-allocated, but if it becomes
1395 * LRU the isolation will just fail.
1396 * RED-PEN would be better to keep it isolated here, but we
1397 * would need to fix isolation locking first.
1398 */
1399 put_page(page);
1400 if (ret == 1) {
1401 ret = 0;
Andi Kleenfb46e732010-09-27 23:31:30 +02001402 pr_info("soft_offline: %#lx: invalidated\n", pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001403 goto done;
1404 }
1405
1406 /*
1407 * Simple invalidation didn't work.
1408 * Try to migrate to a new page instead. migrate.c
1409 * handles a large number of cases for us.
1410 */
1411 ret = isolate_lru_page(page);
1412 if (!ret) {
1413 LIST_HEAD(pagelist);
1414
1415 list_add(&page->lru, &pagelist);
1416 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0);
1417 if (ret) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001418 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001419 pfn, ret, page->flags);
1420 if (ret > 0)
1421 ret = -EIO;
1422 }
1423 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001424 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001425 pfn, ret, page_count(page), page->flags);
1426 }
1427 if (ret)
1428 return ret;
1429
1430done:
1431 atomic_long_add(1, &mce_bad_pages);
1432 SetPageHWPoison(page);
1433 /* keep elevated page count for bad page */
1434 return ret;
1435}
Huang Yingbf998152010-05-31 14:28:19 +08001436
Huang Yingbbeb3402010-06-22 14:23:11 +08001437/*
1438 * The caller must hold current->mm->mmap_sem in read mode.
1439 */
Huang Yingbf998152010-05-31 14:28:19 +08001440int is_hwpoison_address(unsigned long addr)
1441{
1442 pgd_t *pgdp;
1443 pud_t pud, *pudp;
1444 pmd_t pmd, *pmdp;
1445 pte_t pte, *ptep;
1446 swp_entry_t entry;
1447
1448 pgdp = pgd_offset(current->mm, addr);
1449 if (!pgd_present(*pgdp))
1450 return 0;
1451 pudp = pud_offset(pgdp, addr);
1452 pud = *pudp;
1453 if (!pud_present(pud) || pud_large(pud))
1454 return 0;
1455 pmdp = pmd_offset(pudp, addr);
1456 pmd = *pmdp;
1457 if (!pmd_present(pmd) || pmd_large(pmd))
1458 return 0;
1459 ptep = pte_offset_map(pmdp, addr);
1460 pte = *ptep;
1461 pte_unmap(ptep);
1462 if (!is_swap_pte(pte))
1463 return 0;
1464 entry = pte_to_swp_entry(pte);
1465 return is_hwpoison_entry(entry);
1466}
1467EXPORT_SYMBOL_GPL(is_hwpoison_address);