blob: 6b44e52cacaa19857de7ba08af0469492dbeb2db [file] [log] [blame]
Andi Kleen6a460792009-09-16 11:50:15 +02001/*
2 * Copyright (C) 2008, 2009 Intel Corporation
3 * Authors: Andi Kleen, Fengguang Wu
4 *
5 * This software may be redistributed and/or modified under the terms of
6 * the GNU General Public License ("GPL") version 2 only as published by the
7 * Free Software Foundation.
8 *
9 * High level machine check handler. Handles pages reported by the
10 * hardware as being corrupted usually due to a 2bit ECC memory or cache
11 * failure.
12 *
13 * Handles page cache pages in various states. The tricky part
14 * here is that we can access any page asynchronous to other VM
15 * users, because memory failures could happen anytime and anywhere,
16 * possibly violating some of their assumptions. This is why this code
17 * has to be extremely careful. Generally it tries to use normal locking
18 * rules, as in get the standard locks, even if that means the
19 * error handling takes potentially a long time.
20 *
21 * The operation to map back from RMAP chains to processes has to walk
22 * the complete process list and has non linear complexity with the number
23 * mappings. In short it can be quite slow. But since memory corruptions
24 * are rare we hope to get away with this.
25 */
26
27/*
28 * Notebook:
29 * - hugetlb needs more code
30 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
31 * - pass bad pages to kdump next kernel
32 */
33#define DEBUG 1 /* remove me in 2.6.34 */
34#include <linux/kernel.h>
35#include <linux/mm.h>
36#include <linux/page-flags.h>
Wu Fengguang478c5ff2009-12-16 12:19:59 +010037#include <linux/kernel-page-flags.h>
Andi Kleen6a460792009-09-16 11:50:15 +020038#include <linux/sched.h>
Hugh Dickins01e00f82009-10-13 15:02:11 +010039#include <linux/ksm.h>
Andi Kleen6a460792009-09-16 11:50:15 +020040#include <linux/rmap.h>
41#include <linux/pagemap.h>
42#include <linux/swap.h>
43#include <linux/backing-dev.h>
Andi Kleenfacb6012009-12-16 12:20:00 +010044#include <linux/migrate.h>
45#include <linux/page-isolation.h>
46#include <linux/suspend.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090047#include <linux/slab.h>
Huang Yingbf998152010-05-31 14:28:19 +080048#include <linux/swapops.h>
Andi Kleen6a460792009-09-16 11:50:15 +020049#include "internal.h"
50
51int sysctl_memory_failure_early_kill __read_mostly = 0;
52
53int sysctl_memory_failure_recovery __read_mostly = 1;
54
55atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
56
Andi Kleen27df5062009-12-21 19:56:42 +010057#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
58
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010059u32 hwpoison_filter_enable = 0;
Wu Fengguang7c116f22009-12-16 12:19:59 +010060u32 hwpoison_filter_dev_major = ~0U;
61u32 hwpoison_filter_dev_minor = ~0U;
Wu Fengguang478c5ff2009-12-16 12:19:59 +010062u64 hwpoison_filter_flags_mask;
63u64 hwpoison_filter_flags_value;
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010064EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
Wu Fengguang7c116f22009-12-16 12:19:59 +010065EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
66EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
Wu Fengguang478c5ff2009-12-16 12:19:59 +010067EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
68EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
Wu Fengguang7c116f22009-12-16 12:19:59 +010069
70static int hwpoison_filter_dev(struct page *p)
71{
72 struct address_space *mapping;
73 dev_t dev;
74
75 if (hwpoison_filter_dev_major == ~0U &&
76 hwpoison_filter_dev_minor == ~0U)
77 return 0;
78
79 /*
80 * page_mapping() does not accept slab page
81 */
82 if (PageSlab(p))
83 return -EINVAL;
84
85 mapping = page_mapping(p);
86 if (mapping == NULL || mapping->host == NULL)
87 return -EINVAL;
88
89 dev = mapping->host->i_sb->s_dev;
90 if (hwpoison_filter_dev_major != ~0U &&
91 hwpoison_filter_dev_major != MAJOR(dev))
92 return -EINVAL;
93 if (hwpoison_filter_dev_minor != ~0U &&
94 hwpoison_filter_dev_minor != MINOR(dev))
95 return -EINVAL;
96
97 return 0;
98}
99
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100100static int hwpoison_filter_flags(struct page *p)
101{
102 if (!hwpoison_filter_flags_mask)
103 return 0;
104
105 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
106 hwpoison_filter_flags_value)
107 return 0;
108 else
109 return -EINVAL;
110}
111
Andi Kleen4fd466e2009-12-16 12:19:59 +0100112/*
113 * This allows stress tests to limit test scope to a collection of tasks
114 * by putting them under some memcg. This prevents killing unrelated/important
115 * processes such as /sbin/init. Note that the target task may share clean
116 * pages with init (eg. libc text), which is harmless. If the target task
117 * share _dirty_ pages with another task B, the test scheme must make sure B
118 * is also included in the memcg. At last, due to race conditions this filter
119 * can only guarantee that the page either belongs to the memcg tasks, or is
120 * a freed page.
121 */
122#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
123u64 hwpoison_filter_memcg;
124EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
125static int hwpoison_filter_task(struct page *p)
126{
127 struct mem_cgroup *mem;
128 struct cgroup_subsys_state *css;
129 unsigned long ino;
130
131 if (!hwpoison_filter_memcg)
132 return 0;
133
134 mem = try_get_mem_cgroup_from_page(p);
135 if (!mem)
136 return -EINVAL;
137
138 css = mem_cgroup_css(mem);
139 /* root_mem_cgroup has NULL dentries */
140 if (!css->cgroup->dentry)
141 return -EINVAL;
142
143 ino = css->cgroup->dentry->d_inode->i_ino;
144 css_put(css);
145
146 if (ino != hwpoison_filter_memcg)
147 return -EINVAL;
148
149 return 0;
150}
151#else
152static int hwpoison_filter_task(struct page *p) { return 0; }
153#endif
154
Wu Fengguang7c116f22009-12-16 12:19:59 +0100155int hwpoison_filter(struct page *p)
156{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100157 if (!hwpoison_filter_enable)
158 return 0;
159
Wu Fengguang7c116f22009-12-16 12:19:59 +0100160 if (hwpoison_filter_dev(p))
161 return -EINVAL;
162
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100163 if (hwpoison_filter_flags(p))
164 return -EINVAL;
165
Andi Kleen4fd466e2009-12-16 12:19:59 +0100166 if (hwpoison_filter_task(p))
167 return -EINVAL;
168
Wu Fengguang7c116f22009-12-16 12:19:59 +0100169 return 0;
170}
Andi Kleen27df5062009-12-21 19:56:42 +0100171#else
172int hwpoison_filter(struct page *p)
173{
174 return 0;
175}
176#endif
177
Wu Fengguang7c116f22009-12-16 12:19:59 +0100178EXPORT_SYMBOL_GPL(hwpoison_filter);
179
Andi Kleen6a460792009-09-16 11:50:15 +0200180/*
181 * Send all the processes who have the page mapped an ``action optional''
182 * signal.
183 */
184static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
185 unsigned long pfn)
186{
187 struct siginfo si;
188 int ret;
189
190 printk(KERN_ERR
191 "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
192 pfn, t->comm, t->pid);
193 si.si_signo = SIGBUS;
194 si.si_errno = 0;
195 si.si_code = BUS_MCEERR_AO;
196 si.si_addr = (void *)addr;
197#ifdef __ARCH_SI_TRAPNO
198 si.si_trapno = trapno;
199#endif
200 si.si_addr_lsb = PAGE_SHIFT;
201 /*
202 * Don't use force here, it's convenient if the signal
203 * can be temporarily blocked.
204 * This could cause a loop when the user sets SIGBUS
205 * to SIG_IGN, but hopefully noone will do that?
206 */
207 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
208 if (ret < 0)
209 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
210 t->comm, t->pid, ret);
211 return ret;
212}
213
214/*
Andi Kleen588f9ce2009-12-16 12:19:57 +0100215 * When a unknown page type is encountered drain as many buffers as possible
216 * in the hope to turn the page into a LRU or free page, which we can handle.
217 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100218void shake_page(struct page *p, int access)
Andi Kleen588f9ce2009-12-16 12:19:57 +0100219{
220 if (!PageSlab(p)) {
221 lru_add_drain_all();
222 if (PageLRU(p))
223 return;
224 drain_all_pages();
225 if (PageLRU(p) || is_free_buddy_page(p))
226 return;
227 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100228
Andi Kleen588f9ce2009-12-16 12:19:57 +0100229 /*
Andi Kleenfacb6012009-12-16 12:20:00 +0100230 * Only all shrink_slab here (which would also
231 * shrink other caches) if access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100232 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100233 if (access) {
234 int nr;
235 do {
236 nr = shrink_slab(1000, GFP_KERNEL, 1000);
237 if (page_count(p) == 0)
238 break;
239 } while (nr > 10);
240 }
Andi Kleen588f9ce2009-12-16 12:19:57 +0100241}
242EXPORT_SYMBOL_GPL(shake_page);
243
244/*
Andi Kleen6a460792009-09-16 11:50:15 +0200245 * Kill all processes that have a poisoned page mapped and then isolate
246 * the page.
247 *
248 * General strategy:
249 * Find all processes having the page mapped and kill them.
250 * But we keep a page reference around so that the page is not
251 * actually freed yet.
252 * Then stash the page away
253 *
254 * There's no convenient way to get back to mapped processes
255 * from the VMAs. So do a brute-force search over all
256 * running processes.
257 *
258 * Remember that machine checks are not common (or rather
259 * if they are common you have other problems), so this shouldn't
260 * be a performance issue.
261 *
262 * Also there are some races possible while we get from the
263 * error detection to actually handle it.
264 */
265
266struct to_kill {
267 struct list_head nd;
268 struct task_struct *tsk;
269 unsigned long addr;
270 unsigned addr_valid:1;
271};
272
273/*
274 * Failure handling: if we can't find or can't kill a process there's
275 * not much we can do. We just print a message and ignore otherwise.
276 */
277
278/*
279 * Schedule a process for later kill.
280 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
281 * TBD would GFP_NOIO be enough?
282 */
283static void add_to_kill(struct task_struct *tsk, struct page *p,
284 struct vm_area_struct *vma,
285 struct list_head *to_kill,
286 struct to_kill **tkc)
287{
288 struct to_kill *tk;
289
290 if (*tkc) {
291 tk = *tkc;
292 *tkc = NULL;
293 } else {
294 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
295 if (!tk) {
296 printk(KERN_ERR
297 "MCE: Out of memory while machine check handling\n");
298 return;
299 }
300 }
301 tk->addr = page_address_in_vma(p, vma);
302 tk->addr_valid = 1;
303
304 /*
305 * In theory we don't have to kill when the page was
306 * munmaped. But it could be also a mremap. Since that's
307 * likely very rare kill anyways just out of paranoia, but use
308 * a SIGKILL because the error is not contained anymore.
309 */
310 if (tk->addr == -EFAULT) {
311 pr_debug("MCE: Unable to find user space address %lx in %s\n",
312 page_to_pfn(p), tsk->comm);
313 tk->addr_valid = 0;
314 }
315 get_task_struct(tsk);
316 tk->tsk = tsk;
317 list_add_tail(&tk->nd, to_kill);
318}
319
320/*
321 * Kill the processes that have been collected earlier.
322 *
323 * Only do anything when DOIT is set, otherwise just free the list
324 * (this is used for clean pages which do not need killing)
325 * Also when FAIL is set do a force kill because something went
326 * wrong earlier.
327 */
328static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
329 int fail, unsigned long pfn)
330{
331 struct to_kill *tk, *next;
332
333 list_for_each_entry_safe (tk, next, to_kill, nd) {
334 if (doit) {
335 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200336 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200337 * make sure the process doesn't catch the
338 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200339 */
340 if (fail || tk->addr_valid == 0) {
341 printk(KERN_ERR
342 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
343 pfn, tk->tsk->comm, tk->tsk->pid);
344 force_sig(SIGKILL, tk->tsk);
345 }
346
347 /*
348 * In theory the process could have mapped
349 * something else on the address in-between. We could
350 * check for that, but we need to tell the
351 * process anyways.
352 */
353 else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
354 pfn) < 0)
355 printk(KERN_ERR
356 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
357 pfn, tk->tsk->comm, tk->tsk->pid);
358 }
359 put_task_struct(tk->tsk);
360 kfree(tk);
361 }
362}
363
364static int task_early_kill(struct task_struct *tsk)
365{
366 if (!tsk->mm)
367 return 0;
368 if (tsk->flags & PF_MCE_PROCESS)
369 return !!(tsk->flags & PF_MCE_EARLY);
370 return sysctl_memory_failure_early_kill;
371}
372
373/*
374 * Collect processes when the error hit an anonymous page.
375 */
376static void collect_procs_anon(struct page *page, struct list_head *to_kill,
377 struct to_kill **tkc)
378{
379 struct vm_area_struct *vma;
380 struct task_struct *tsk;
381 struct anon_vma *av;
382
383 read_lock(&tasklist_lock);
384 av = page_lock_anon_vma(page);
385 if (av == NULL) /* Not actually mapped anymore */
386 goto out;
387 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800388 struct anon_vma_chain *vmac;
389
Andi Kleen6a460792009-09-16 11:50:15 +0200390 if (!task_early_kill(tsk))
391 continue;
Rik van Riel5beb4932010-03-05 13:42:07 -0800392 list_for_each_entry(vmac, &av->head, same_anon_vma) {
393 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200394 if (!page_mapped_in_vma(page, vma))
395 continue;
396 if (vma->vm_mm == tsk->mm)
397 add_to_kill(tsk, page, vma, to_kill, tkc);
398 }
399 }
400 page_unlock_anon_vma(av);
401out:
402 read_unlock(&tasklist_lock);
403}
404
405/*
406 * Collect processes when the error hit a file mapped page.
407 */
408static void collect_procs_file(struct page *page, struct list_head *to_kill,
409 struct to_kill **tkc)
410{
411 struct vm_area_struct *vma;
412 struct task_struct *tsk;
413 struct prio_tree_iter iter;
414 struct address_space *mapping = page->mapping;
415
416 /*
417 * A note on the locking order between the two locks.
418 * We don't rely on this particular order.
419 * If you have some other code that needs a different order
420 * feel free to switch them around. Or add a reverse link
421 * from mm_struct to task_struct, then this could be all
422 * done without taking tasklist_lock and looping over all tasks.
423 */
424
425 read_lock(&tasklist_lock);
426 spin_lock(&mapping->i_mmap_lock);
427 for_each_process(tsk) {
428 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
429
430 if (!task_early_kill(tsk))
431 continue;
432
433 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
434 pgoff) {
435 /*
436 * Send early kill signal to tasks where a vma covers
437 * the page but the corrupted page is not necessarily
438 * mapped it in its pte.
439 * Assume applications who requested early kill want
440 * to be informed of all such data corruptions.
441 */
442 if (vma->vm_mm == tsk->mm)
443 add_to_kill(tsk, page, vma, to_kill, tkc);
444 }
445 }
446 spin_unlock(&mapping->i_mmap_lock);
447 read_unlock(&tasklist_lock);
448}
449
450/*
451 * Collect the processes who have the corrupted page mapped to kill.
452 * This is done in two steps for locking reasons.
453 * First preallocate one tokill structure outside the spin locks,
454 * so that we can kill at least one process reasonably reliable.
455 */
456static void collect_procs(struct page *page, struct list_head *tokill)
457{
458 struct to_kill *tk;
459
460 if (!page->mapping)
461 return;
462
463 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
464 if (!tk)
465 return;
466 if (PageAnon(page))
467 collect_procs_anon(page, tokill, &tk);
468 else
469 collect_procs_file(page, tokill, &tk);
470 kfree(tk);
471}
472
473/*
474 * Error handlers for various types of pages.
475 */
476
477enum outcome {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100478 IGNORED, /* Error: cannot be handled */
479 FAILED, /* Error: handling failed */
Andi Kleen6a460792009-09-16 11:50:15 +0200480 DELAYED, /* Will be handled later */
Andi Kleen6a460792009-09-16 11:50:15 +0200481 RECOVERED, /* Successfully recovered */
482};
483
484static const char *action_name[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100485 [IGNORED] = "Ignored",
Andi Kleen6a460792009-09-16 11:50:15 +0200486 [FAILED] = "Failed",
487 [DELAYED] = "Delayed",
Andi Kleen6a460792009-09-16 11:50:15 +0200488 [RECOVERED] = "Recovered",
489};
490
491/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100492 * XXX: It is possible that a page is isolated from LRU cache,
493 * and then kept in swap cache or failed to remove from page cache.
494 * The page count will stop it from being freed by unpoison.
495 * Stress tests should be aware of this memory leak problem.
496 */
497static int delete_from_lru_cache(struct page *p)
498{
499 if (!isolate_lru_page(p)) {
500 /*
501 * Clear sensible page flags, so that the buddy system won't
502 * complain when the page is unpoison-and-freed.
503 */
504 ClearPageActive(p);
505 ClearPageUnevictable(p);
506 /*
507 * drop the page count elevated by isolate_lru_page()
508 */
509 page_cache_release(p);
510 return 0;
511 }
512 return -EIO;
513}
514
515/*
Andi Kleen6a460792009-09-16 11:50:15 +0200516 * Error hit kernel page.
517 * Do nothing, try to be lucky and not touch this instead. For a few cases we
518 * could be more sophisticated.
519 */
520static int me_kernel(struct page *p, unsigned long pfn)
521{
Andi Kleen6a460792009-09-16 11:50:15 +0200522 return IGNORED;
523}
524
525/*
526 * Page in unknown state. Do nothing.
527 */
528static int me_unknown(struct page *p, unsigned long pfn)
529{
530 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
531 return FAILED;
532}
533
534/*
Andi Kleen6a460792009-09-16 11:50:15 +0200535 * Clean (or cleaned) page cache page.
536 */
537static int me_pagecache_clean(struct page *p, unsigned long pfn)
538{
539 int err;
540 int ret = FAILED;
541 struct address_space *mapping;
542
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100543 delete_from_lru_cache(p);
544
Andi Kleen6a460792009-09-16 11:50:15 +0200545 /*
546 * For anonymous pages we're done the only reference left
547 * should be the one m_f() holds.
548 */
549 if (PageAnon(p))
550 return RECOVERED;
551
552 /*
553 * Now truncate the page in the page cache. This is really
554 * more like a "temporary hole punch"
555 * Don't do this for block devices when someone else
556 * has a reference, because it could be file system metadata
557 * and that's not safe to truncate.
558 */
559 mapping = page_mapping(p);
560 if (!mapping) {
561 /*
562 * Page has been teared down in the meanwhile
563 */
564 return FAILED;
565 }
566
567 /*
568 * Truncation is a bit tricky. Enable it per file system for now.
569 *
570 * Open: to take i_mutex or not for this? Right now we don't.
571 */
572 if (mapping->a_ops->error_remove_page) {
573 err = mapping->a_ops->error_remove_page(mapping, p);
574 if (err != 0) {
575 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
576 pfn, err);
577 } else if (page_has_private(p) &&
578 !try_to_release_page(p, GFP_NOIO)) {
579 pr_debug("MCE %#lx: failed to release buffers\n", pfn);
580 } else {
581 ret = RECOVERED;
582 }
583 } else {
584 /*
585 * If the file system doesn't support it just invalidate
586 * This fails on dirty or anything with private pages
587 */
588 if (invalidate_inode_page(p))
589 ret = RECOVERED;
590 else
591 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
592 pfn);
593 }
594 return ret;
595}
596
597/*
598 * Dirty cache page page
599 * Issues: when the error hit a hole page the error is not properly
600 * propagated.
601 */
602static int me_pagecache_dirty(struct page *p, unsigned long pfn)
603{
604 struct address_space *mapping = page_mapping(p);
605
606 SetPageError(p);
607 /* TBD: print more information about the file. */
608 if (mapping) {
609 /*
610 * IO error will be reported by write(), fsync(), etc.
611 * who check the mapping.
612 * This way the application knows that something went
613 * wrong with its dirty file data.
614 *
615 * There's one open issue:
616 *
617 * The EIO will be only reported on the next IO
618 * operation and then cleared through the IO map.
619 * Normally Linux has two mechanisms to pass IO error
620 * first through the AS_EIO flag in the address space
621 * and then through the PageError flag in the page.
622 * Since we drop pages on memory failure handling the
623 * only mechanism open to use is through AS_AIO.
624 *
625 * This has the disadvantage that it gets cleared on
626 * the first operation that returns an error, while
627 * the PageError bit is more sticky and only cleared
628 * when the page is reread or dropped. If an
629 * application assumes it will always get error on
630 * fsync, but does other operations on the fd before
631 * and the page is dropped inbetween then the error
632 * will not be properly reported.
633 *
634 * This can already happen even without hwpoisoned
635 * pages: first on metadata IO errors (which only
636 * report through AS_EIO) or when the page is dropped
637 * at the wrong time.
638 *
639 * So right now we assume that the application DTRT on
640 * the first EIO, but we're not worse than other parts
641 * of the kernel.
642 */
643 mapping_set_error(mapping, EIO);
644 }
645
646 return me_pagecache_clean(p, pfn);
647}
648
649/*
650 * Clean and dirty swap cache.
651 *
652 * Dirty swap cache page is tricky to handle. The page could live both in page
653 * cache and swap cache(ie. page is freshly swapped in). So it could be
654 * referenced concurrently by 2 types of PTEs:
655 * normal PTEs and swap PTEs. We try to handle them consistently by calling
656 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
657 * and then
658 * - clear dirty bit to prevent IO
659 * - remove from LRU
660 * - but keep in the swap cache, so that when we return to it on
661 * a later page fault, we know the application is accessing
662 * corrupted data and shall be killed (we installed simple
663 * interception code in do_swap_page to catch it).
664 *
665 * Clean swap cache pages can be directly isolated. A later page fault will
666 * bring in the known good data from disk.
667 */
668static int me_swapcache_dirty(struct page *p, unsigned long pfn)
669{
Andi Kleen6a460792009-09-16 11:50:15 +0200670 ClearPageDirty(p);
671 /* Trigger EIO in shmem: */
672 ClearPageUptodate(p);
673
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100674 if (!delete_from_lru_cache(p))
675 return DELAYED;
676 else
677 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200678}
679
680static int me_swapcache_clean(struct page *p, unsigned long pfn)
681{
Andi Kleen6a460792009-09-16 11:50:15 +0200682 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800683
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100684 if (!delete_from_lru_cache(p))
685 return RECOVERED;
686 else
687 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200688}
689
690/*
691 * Huge pages. Needs work.
692 * Issues:
693 * No rmap support so we cannot find the original mapper. In theory could walk
694 * all MMs and look for the mappings, but that would be non atomic and racy.
695 * Need rmap for hugepages for this. Alternatively we could employ a heuristic,
696 * like just walking the current process and hoping it has it mapped (that
697 * should be usually true for the common "shared database cache" case)
698 * Should handle free huge pages and dequeue them too, but this needs to
699 * handle huge page accounting correctly.
700 */
701static int me_huge_page(struct page *p, unsigned long pfn)
702{
703 return FAILED;
704}
705
706/*
707 * Various page states we can handle.
708 *
709 * A page state is defined by its current page->flags bits.
710 * The table matches them in order and calls the right handler.
711 *
712 * This is quite tricky because we can access page at any time
713 * in its live cycle, so all accesses have to be extremly careful.
714 *
715 * This is not complete. More states could be added.
716 * For any missing state don't attempt recovery.
717 */
718
719#define dirty (1UL << PG_dirty)
720#define sc (1UL << PG_swapcache)
721#define unevict (1UL << PG_unevictable)
722#define mlock (1UL << PG_mlocked)
723#define writeback (1UL << PG_writeback)
724#define lru (1UL << PG_lru)
725#define swapbacked (1UL << PG_swapbacked)
726#define head (1UL << PG_head)
727#define tail (1UL << PG_tail)
728#define compound (1UL << PG_compound)
729#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200730#define reserved (1UL << PG_reserved)
731
732static struct page_state {
733 unsigned long mask;
734 unsigned long res;
735 char *msg;
736 int (*action)(struct page *p, unsigned long pfn);
737} error_states[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100738 { reserved, reserved, "reserved kernel", me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100739 /*
740 * free pages are specially detected outside this table:
741 * PG_buddy pages only make a small fraction of all free pages.
742 */
Andi Kleen6a460792009-09-16 11:50:15 +0200743
744 /*
745 * Could in theory check if slab page is free or if we can drop
746 * currently unused objects without touching them. But just
747 * treat it as standard kernel for now.
748 */
749 { slab, slab, "kernel slab", me_kernel },
750
751#ifdef CONFIG_PAGEFLAGS_EXTENDED
752 { head, head, "huge", me_huge_page },
753 { tail, tail, "huge", me_huge_page },
754#else
755 { compound, compound, "huge", me_huge_page },
756#endif
757
758 { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
759 { sc|dirty, sc, "swapcache", me_swapcache_clean },
760
761 { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
762 { unevict, unevict, "unevictable LRU", me_pagecache_clean},
763
Andi Kleen6a460792009-09-16 11:50:15 +0200764 { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
765 { mlock, mlock, "mlocked LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200766
767 { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
768 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200769
770 /*
771 * Catchall entry: must be at end.
772 */
773 { 0, 0, "unknown page state", me_unknown },
774};
775
Andi Kleen2326c462009-12-16 12:20:00 +0100776#undef dirty
777#undef sc
778#undef unevict
779#undef mlock
780#undef writeback
781#undef lru
782#undef swapbacked
783#undef head
784#undef tail
785#undef compound
786#undef slab
787#undef reserved
788
Andi Kleen6a460792009-09-16 11:50:15 +0200789static void action_result(unsigned long pfn, char *msg, int result)
790{
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100791 struct page *page = pfn_to_page(pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200792
793 printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
794 pfn,
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100795 PageDirty(page) ? "dirty " : "",
Andi Kleen6a460792009-09-16 11:50:15 +0200796 msg, action_name[result]);
797}
798
799static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100800 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200801{
802 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200803 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200804
805 result = ps->action(p, pfn);
806 action_result(pfn, ps->msg, result);
Wu Fengguang7456b042009-10-19 08:15:01 +0200807
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100808 count = page_count(p) - 1;
Wu Fengguang138ce282009-12-16 12:19:58 +0100809 if (ps->action == me_swapcache_dirty && result == DELAYED)
810 count--;
811 if (count != 0) {
Andi Kleen6a460792009-09-16 11:50:15 +0200812 printk(KERN_ERR
813 "MCE %#lx: %s page still referenced by %d users\n",
Wu Fengguang7456b042009-10-19 08:15:01 +0200814 pfn, ps->msg, count);
Wu Fengguang138ce282009-12-16 12:19:58 +0100815 result = FAILED;
816 }
Andi Kleen6a460792009-09-16 11:50:15 +0200817
818 /* Could do more checks here if page looks ok */
819 /*
820 * Could adjust zone counters here to correct for the missing page.
821 */
822
Wu Fengguang138ce282009-12-16 12:19:58 +0100823 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200824}
825
826#define N_UNMAP_TRIES 5
827
828/*
829 * Do all that is necessary to remove user space mappings. Unmap
830 * the pages and send SIGBUS to the processes if the data was dirty.
831 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100832static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Andi Kleen6a460792009-09-16 11:50:15 +0200833 int trapno)
834{
835 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
836 struct address_space *mapping;
837 LIST_HEAD(tokill);
838 int ret;
839 int i;
840 int kill = 1;
841
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100842 if (PageReserved(p) || PageSlab(p))
843 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200844
Andi Kleen6a460792009-09-16 11:50:15 +0200845 /*
846 * This check implies we don't kill processes if their pages
847 * are in the swap cache early. Those are always late kills.
848 */
849 if (!page_mapped(p))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100850 return SWAP_SUCCESS;
851
852 if (PageCompound(p) || PageKsm(p))
853 return SWAP_FAIL;
Andi Kleen6a460792009-09-16 11:50:15 +0200854
855 if (PageSwapCache(p)) {
856 printk(KERN_ERR
857 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
858 ttu |= TTU_IGNORE_HWPOISON;
859 }
860
861 /*
862 * Propagate the dirty bit from PTEs to struct page first, because we
863 * need this to decide if we should kill or just drop the page.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100864 * XXX: the dirty test could be racy: set_page_dirty() may not always
865 * be called inside page lock (it's recommended but not enforced).
Andi Kleen6a460792009-09-16 11:50:15 +0200866 */
867 mapping = page_mapping(p);
868 if (!PageDirty(p) && mapping && mapping_cap_writeback_dirty(mapping)) {
869 if (page_mkclean(p)) {
870 SetPageDirty(p);
871 } else {
872 kill = 0;
873 ttu |= TTU_IGNORE_HWPOISON;
874 printk(KERN_INFO
875 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
876 pfn);
877 }
878 }
879
880 /*
881 * First collect all the processes that have the page
882 * mapped in dirty form. This has to be done before try_to_unmap,
883 * because ttu takes the rmap data structures down.
884 *
885 * Error handling: We ignore errors here because
886 * there's nothing that can be done.
887 */
888 if (kill)
889 collect_procs(p, &tokill);
890
891 /*
892 * try_to_unmap can fail temporarily due to races.
893 * Try a few times (RED-PEN better strategy?)
894 */
895 for (i = 0; i < N_UNMAP_TRIES; i++) {
896 ret = try_to_unmap(p, ttu);
897 if (ret == SWAP_SUCCESS)
898 break;
899 pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn, ret);
900 }
901
902 if (ret != SWAP_SUCCESS)
903 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
904 pfn, page_mapcount(p));
905
906 /*
907 * Now that the dirty bit has been propagated to the
908 * struct page and all unmaps done we can decide if
909 * killing is needed or not. Only kill when the page
910 * was dirty, otherwise the tokill list is merely
911 * freed. When there was a problem unmapping earlier
912 * use a more force-full uncatchable kill to prevent
913 * any accesses to the poisoned memory.
914 */
915 kill_procs_ao(&tokill, !!PageDirty(p), trapno,
916 ret != SWAP_SUCCESS, pfn);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100917
918 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200919}
920
Andi Kleen82ba0112009-12-16 12:19:57 +0100921int __memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200922{
923 struct page_state *ps;
924 struct page *p;
925 int res;
926
927 if (!sysctl_memory_failure_recovery)
928 panic("Memory failure from trap %d on page %lx", trapno, pfn);
929
930 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100931 printk(KERN_ERR
932 "MCE %#lx: memory outside kernel control\n",
933 pfn);
934 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +0200935 }
936
937 p = pfn_to_page(pfn);
938 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100939 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200940 return 0;
941 }
942
943 atomic_long_add(1, &mce_bad_pages);
944
945 /*
946 * We need/can do nothing about count=0 pages.
947 * 1) it's a free page, and therefore in safe hand:
948 * prep_new_page() will be the gate keeper.
949 * 2) it's part of a non-compound high order page.
950 * Implies some kernel user: cannot stop them from
951 * R/W the page; let's pray that the page has been
952 * used and will be freed some time later.
953 * In fact it's dangerous to directly bump up page count from 0,
954 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
955 */
Andi Kleen82ba0112009-12-16 12:19:57 +0100956 if (!(flags & MF_COUNT_INCREASED) &&
957 !get_page_unless_zero(compound_head(p))) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +0100958 if (is_free_buddy_page(p)) {
959 action_result(pfn, "free buddy", DELAYED);
960 return 0;
961 } else {
962 action_result(pfn, "high order kernel", IGNORED);
963 return -EBUSY;
964 }
Andi Kleen6a460792009-09-16 11:50:15 +0200965 }
966
967 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +0800968 * We ignore non-LRU pages for good reasons.
969 * - PG_locked is only well defined for LRU pages and a few others
970 * - to avoid races with __set_page_locked()
971 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
972 * The check (unnecessarily) ignores LRU pages being isolated and
973 * walked by the page reclaim code, however that's not a big loss.
974 */
975 if (!PageLRU(p))
Andi Kleenfacb6012009-12-16 12:20:00 +0100976 shake_page(p, 0);
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100977 if (!PageLRU(p)) {
Andi Kleen0474a602009-12-16 12:20:00 +0100978 /*
979 * shake_page could have turned it free.
980 */
981 if (is_free_buddy_page(p)) {
982 action_result(pfn, "free buddy, 2nd try", DELAYED);
983 return 0;
984 }
Wu Fengguange43c3af2009-09-29 13:16:20 +0800985 action_result(pfn, "non LRU", IGNORED);
986 put_page(p);
987 return -EBUSY;
988 }
Wu Fengguange43c3af2009-09-29 13:16:20 +0800989
990 /*
Andi Kleen6a460792009-09-16 11:50:15 +0200991 * Lock the page and wait for writeback to finish.
992 * It's very difficult to mess with pages currently under IO
993 * and in many cases impossible, so we just avoid it here.
994 */
995 lock_page_nosync(p);
Wu Fengguang847ce402009-12-16 12:19:58 +0100996
997 /*
998 * unpoison always clear PG_hwpoison inside page lock
999 */
1000 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001001 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001002 res = 0;
1003 goto out;
1004 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001005 if (hwpoison_filter(p)) {
1006 if (TestClearPageHWPoison(p))
1007 atomic_long_dec(&mce_bad_pages);
1008 unlock_page(p);
1009 put_page(p);
1010 return 0;
1011 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001012
Andi Kleen6a460792009-09-16 11:50:15 +02001013 wait_on_page_writeback(p);
1014
1015 /*
1016 * Now take care of user space mappings.
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001017 * Abort on fail: __remove_from_page_cache() assumes unmapped page.
Andi Kleen6a460792009-09-16 11:50:15 +02001018 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001019 if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
1020 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1021 res = -EBUSY;
1022 goto out;
1023 }
Andi Kleen6a460792009-09-16 11:50:15 +02001024
1025 /*
1026 * Torn down by someone else?
1027 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001028 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Andi Kleen6a460792009-09-16 11:50:15 +02001029 action_result(pfn, "already truncated LRU", IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001030 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001031 goto out;
1032 }
1033
1034 res = -EBUSY;
1035 for (ps = error_states;; ps++) {
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001036 if ((p->flags & ps->mask) == ps->res) {
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +01001037 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001038 break;
1039 }
1040 }
1041out:
1042 unlock_page(p);
1043 return res;
1044}
1045EXPORT_SYMBOL_GPL(__memory_failure);
1046
1047/**
1048 * memory_failure - Handle memory failure of a page.
1049 * @pfn: Page Number of the corrupted page
1050 * @trapno: Trap number reported in the signal to user space.
1051 *
1052 * This function is called by the low level machine check code
1053 * of an architecture when it detects hardware memory corruption
1054 * of a page. It tries its best to recover, which includes
1055 * dropping pages, killing processes etc.
1056 *
1057 * The function is primarily of use for corruptions that
1058 * happen outside the current execution context (e.g. when
1059 * detected by a background scrubber)
1060 *
1061 * Must run in process context (e.g. a work queue) with interrupts
1062 * enabled and no spinlocks hold.
1063 */
1064void memory_failure(unsigned long pfn, int trapno)
1065{
1066 __memory_failure(pfn, trapno, 0);
1067}
Wu Fengguang847ce402009-12-16 12:19:58 +01001068
1069/**
1070 * unpoison_memory - Unpoison a previously poisoned page
1071 * @pfn: Page number of the to be unpoisoned page
1072 *
1073 * Software-unpoison a page that has been poisoned by
1074 * memory_failure() earlier.
1075 *
1076 * This is only done on the software-level, so it only works
1077 * for linux injected failures, not real hardware failures
1078 *
1079 * Returns 0 for success, otherwise -errno.
1080 */
1081int unpoison_memory(unsigned long pfn)
1082{
1083 struct page *page;
1084 struct page *p;
1085 int freeit = 0;
1086
1087 if (!pfn_valid(pfn))
1088 return -ENXIO;
1089
1090 p = pfn_to_page(pfn);
1091 page = compound_head(p);
1092
1093 if (!PageHWPoison(p)) {
1094 pr_debug("MCE: Page was already unpoisoned %#lx\n", pfn);
1095 return 0;
1096 }
1097
1098 if (!get_page_unless_zero(page)) {
1099 if (TestClearPageHWPoison(p))
1100 atomic_long_dec(&mce_bad_pages);
1101 pr_debug("MCE: Software-unpoisoned free page %#lx\n", pfn);
1102 return 0;
1103 }
1104
1105 lock_page_nosync(page);
1106 /*
1107 * This test is racy because PG_hwpoison is set outside of page lock.
1108 * That's acceptable because that won't trigger kernel panic. Instead,
1109 * the PG_hwpoison page will be caught and isolated on the entrance to
1110 * the free buddy page pool.
1111 */
1112 if (TestClearPageHWPoison(p)) {
1113 pr_debug("MCE: Software-unpoisoned page %#lx\n", pfn);
1114 atomic_long_dec(&mce_bad_pages);
1115 freeit = 1;
1116 }
1117 unlock_page(page);
1118
1119 put_page(page);
1120 if (freeit)
1121 put_page(page);
1122
1123 return 0;
1124}
1125EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001126
1127static struct page *new_page(struct page *p, unsigned long private, int **x)
1128{
Andi Kleen12686d12009-12-16 12:20:01 +01001129 int nid = page_to_nid(p);
1130 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001131}
1132
1133/*
1134 * Safely get reference count of an arbitrary page.
1135 * Returns 0 for a free page, -EIO for a zero refcount page
1136 * that is not free, and 1 for any other page type.
1137 * For 1 the page is returned with increased page count, otherwise not.
1138 */
1139static int get_any_page(struct page *p, unsigned long pfn, int flags)
1140{
1141 int ret;
1142
1143 if (flags & MF_COUNT_INCREASED)
1144 return 1;
1145
1146 /*
1147 * The lock_system_sleep prevents a race with memory hotplug,
1148 * because the isolation assumes there's only a single user.
1149 * This is a big hammer, a better would be nicer.
1150 */
1151 lock_system_sleep();
1152
1153 /*
1154 * Isolate the page, so that it doesn't get reallocated if it
1155 * was free.
1156 */
1157 set_migratetype_isolate(p);
1158 if (!get_page_unless_zero(compound_head(p))) {
1159 if (is_free_buddy_page(p)) {
1160 pr_debug("get_any_page: %#lx free buddy page\n", pfn);
1161 /* Set hwpoison bit while page is still isolated */
1162 SetPageHWPoison(p);
1163 ret = 0;
1164 } else {
1165 pr_debug("get_any_page: %#lx: unknown zero refcount page type %lx\n",
1166 pfn, p->flags);
1167 ret = -EIO;
1168 }
1169 } else {
1170 /* Not a free page */
1171 ret = 1;
1172 }
1173 unset_migratetype_isolate(p);
1174 unlock_system_sleep();
1175 return ret;
1176}
1177
1178/**
1179 * soft_offline_page - Soft offline a page.
1180 * @page: page to offline
1181 * @flags: flags. Same as memory_failure().
1182 *
1183 * Returns 0 on success, otherwise negated errno.
1184 *
1185 * Soft offline a page, by migration or invalidation,
1186 * without killing anything. This is for the case when
1187 * a page is not corrupted yet (so it's still valid to access),
1188 * but has had a number of corrected errors and is better taken
1189 * out.
1190 *
1191 * The actual policy on when to do that is maintained by
1192 * user space.
1193 *
1194 * This should never impact any application or cause data loss,
1195 * however it might take some time.
1196 *
1197 * This is not a 100% solution for all memory, but tries to be
1198 * ``good enough'' for the majority of memory.
1199 */
1200int soft_offline_page(struct page *page, int flags)
1201{
1202 int ret;
1203 unsigned long pfn = page_to_pfn(page);
1204
1205 ret = get_any_page(page, pfn, flags);
1206 if (ret < 0)
1207 return ret;
1208 if (ret == 0)
1209 goto done;
1210
1211 /*
1212 * Page cache page we can handle?
1213 */
1214 if (!PageLRU(page)) {
1215 /*
1216 * Try to free it.
1217 */
1218 put_page(page);
1219 shake_page(page, 1);
1220
1221 /*
1222 * Did it turn free?
1223 */
1224 ret = get_any_page(page, pfn, 0);
1225 if (ret < 0)
1226 return ret;
1227 if (ret == 0)
1228 goto done;
1229 }
1230 if (!PageLRU(page)) {
1231 pr_debug("soft_offline: %#lx: unknown non LRU page type %lx\n",
1232 pfn, page->flags);
1233 return -EIO;
1234 }
1235
1236 lock_page(page);
1237 wait_on_page_writeback(page);
1238
1239 /*
1240 * Synchronized using the page lock with memory_failure()
1241 */
1242 if (PageHWPoison(page)) {
1243 unlock_page(page);
1244 put_page(page);
1245 pr_debug("soft offline: %#lx page already poisoned\n", pfn);
1246 return -EBUSY;
1247 }
1248
1249 /*
1250 * Try to invalidate first. This should work for
1251 * non dirty unmapped page cache pages.
1252 */
1253 ret = invalidate_inode_page(page);
1254 unlock_page(page);
1255
1256 /*
1257 * Drop count because page migration doesn't like raised
1258 * counts. The page could get re-allocated, but if it becomes
1259 * LRU the isolation will just fail.
1260 * RED-PEN would be better to keep it isolated here, but we
1261 * would need to fix isolation locking first.
1262 */
1263 put_page(page);
1264 if (ret == 1) {
1265 ret = 0;
1266 pr_debug("soft_offline: %#lx: invalidated\n", pfn);
1267 goto done;
1268 }
1269
1270 /*
1271 * Simple invalidation didn't work.
1272 * Try to migrate to a new page instead. migrate.c
1273 * handles a large number of cases for us.
1274 */
1275 ret = isolate_lru_page(page);
1276 if (!ret) {
1277 LIST_HEAD(pagelist);
1278
1279 list_add(&page->lru, &pagelist);
1280 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0);
1281 if (ret) {
1282 pr_debug("soft offline: %#lx: migration failed %d, type %lx\n",
1283 pfn, ret, page->flags);
1284 if (ret > 0)
1285 ret = -EIO;
1286 }
1287 } else {
1288 pr_debug("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1289 pfn, ret, page_count(page), page->flags);
1290 }
1291 if (ret)
1292 return ret;
1293
1294done:
1295 atomic_long_add(1, &mce_bad_pages);
1296 SetPageHWPoison(page);
1297 /* keep elevated page count for bad page */
1298 return ret;
1299}
Huang Yingbf998152010-05-31 14:28:19 +08001300
Huang Yingbbeb3402010-06-22 14:23:11 +08001301/*
1302 * The caller must hold current->mm->mmap_sem in read mode.
1303 */
Huang Yingbf998152010-05-31 14:28:19 +08001304int is_hwpoison_address(unsigned long addr)
1305{
1306 pgd_t *pgdp;
1307 pud_t pud, *pudp;
1308 pmd_t pmd, *pmdp;
1309 pte_t pte, *ptep;
1310 swp_entry_t entry;
1311
1312 pgdp = pgd_offset(current->mm, addr);
1313 if (!pgd_present(*pgdp))
1314 return 0;
1315 pudp = pud_offset(pgdp, addr);
1316 pud = *pudp;
1317 if (!pud_present(pud) || pud_large(pud))
1318 return 0;
1319 pmdp = pmd_offset(pudp, addr);
1320 pmd = *pmdp;
1321 if (!pmd_present(pmd) || pmd_large(pmd))
1322 return 0;
1323 ptep = pte_offset_map(pmdp, addr);
1324 pte = *ptep;
1325 pte_unmap(ptep);
1326 if (!is_swap_pte(pte))
1327 return 0;
1328 entry = pte_to_swp_entry(pte);
1329 return is_hwpoison_entry(entry);
1330}
1331EXPORT_SYMBOL_GPL(is_hwpoison_address);