blob: ab1e7145e2909c8e1a02359fb1e148118c42e1b2 [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>
Paul Gortmakerb9e15ba2011-05-26 16:00:52 -040045#include <linux/export.h>
Andi Kleen6a460792009-09-16 11:50:15 +020046#include <linux/pagemap.h>
47#include <linux/swap.h>
48#include <linux/backing-dev.h>
Andi Kleenfacb6012009-12-16 12:20:00 +010049#include <linux/migrate.h>
50#include <linux/page-isolation.h>
51#include <linux/suspend.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090052#include <linux/slab.h>
Huang Yingbf998152010-05-31 14:28:19 +080053#include <linux/swapops.h>
Naoya Horiguchi7af446a2010-05-28 09:29:17 +090054#include <linux/hugetlb.h>
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -080055#include <linux/memory_hotplug.h>
Minchan Kim5db8a732011-06-15 15:08:48 -070056#include <linux/mm_inline.h>
Huang Yingea8f5fb2011-07-13 13:14:27 +080057#include <linux/kfifo.h>
Andi Kleen6a460792009-09-16 11:50:15 +020058#include "internal.h"
59
60int sysctl_memory_failure_early_kill __read_mostly = 0;
61
62int sysctl_memory_failure_recovery __read_mostly = 1;
63
64atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
65
Andi Kleen27df5062009-12-21 19:56:42 +010066#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
67
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010068u32 hwpoison_filter_enable = 0;
Wu Fengguang7c116f22009-12-16 12:19:59 +010069u32 hwpoison_filter_dev_major = ~0U;
70u32 hwpoison_filter_dev_minor = ~0U;
Wu Fengguang478c5ff2009-12-16 12:19:59 +010071u64 hwpoison_filter_flags_mask;
72u64 hwpoison_filter_flags_value;
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010073EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
Wu Fengguang7c116f22009-12-16 12:19:59 +010074EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
75EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
Wu Fengguang478c5ff2009-12-16 12:19:59 +010076EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
77EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
Wu Fengguang7c116f22009-12-16 12:19:59 +010078
79static int hwpoison_filter_dev(struct page *p)
80{
81 struct address_space *mapping;
82 dev_t dev;
83
84 if (hwpoison_filter_dev_major == ~0U &&
85 hwpoison_filter_dev_minor == ~0U)
86 return 0;
87
88 /*
Andi Kleen1c80b992010-09-27 23:09:51 +020089 * page_mapping() does not accept slab pages.
Wu Fengguang7c116f22009-12-16 12:19:59 +010090 */
91 if (PageSlab(p))
92 return -EINVAL;
93
94 mapping = page_mapping(p);
95 if (mapping == NULL || mapping->host == NULL)
96 return -EINVAL;
97
98 dev = mapping->host->i_sb->s_dev;
99 if (hwpoison_filter_dev_major != ~0U &&
100 hwpoison_filter_dev_major != MAJOR(dev))
101 return -EINVAL;
102 if (hwpoison_filter_dev_minor != ~0U &&
103 hwpoison_filter_dev_minor != MINOR(dev))
104 return -EINVAL;
105
106 return 0;
107}
108
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100109static int hwpoison_filter_flags(struct page *p)
110{
111 if (!hwpoison_filter_flags_mask)
112 return 0;
113
114 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
115 hwpoison_filter_flags_value)
116 return 0;
117 else
118 return -EINVAL;
119}
120
Andi Kleen4fd466e2009-12-16 12:19:59 +0100121/*
122 * This allows stress tests to limit test scope to a collection of tasks
123 * by putting them under some memcg. This prevents killing unrelated/important
124 * processes such as /sbin/init. Note that the target task may share clean
125 * pages with init (eg. libc text), which is harmless. If the target task
126 * share _dirty_ pages with another task B, the test scheme must make sure B
127 * is also included in the memcg. At last, due to race conditions this filter
128 * can only guarantee that the page either belongs to the memcg tasks, or is
129 * a freed page.
130 */
131#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
132u64 hwpoison_filter_memcg;
133EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
134static int hwpoison_filter_task(struct page *p)
135{
136 struct mem_cgroup *mem;
137 struct cgroup_subsys_state *css;
138 unsigned long ino;
139
140 if (!hwpoison_filter_memcg)
141 return 0;
142
143 mem = try_get_mem_cgroup_from_page(p);
144 if (!mem)
145 return -EINVAL;
146
147 css = mem_cgroup_css(mem);
148 /* root_mem_cgroup has NULL dentries */
149 if (!css->cgroup->dentry)
150 return -EINVAL;
151
152 ino = css->cgroup->dentry->d_inode->i_ino;
153 css_put(css);
154
155 if (ino != hwpoison_filter_memcg)
156 return -EINVAL;
157
158 return 0;
159}
160#else
161static int hwpoison_filter_task(struct page *p) { return 0; }
162#endif
163
Wu Fengguang7c116f22009-12-16 12:19:59 +0100164int hwpoison_filter(struct page *p)
165{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100166 if (!hwpoison_filter_enable)
167 return 0;
168
Wu Fengguang7c116f22009-12-16 12:19:59 +0100169 if (hwpoison_filter_dev(p))
170 return -EINVAL;
171
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100172 if (hwpoison_filter_flags(p))
173 return -EINVAL;
174
Andi Kleen4fd466e2009-12-16 12:19:59 +0100175 if (hwpoison_filter_task(p))
176 return -EINVAL;
177
Wu Fengguang7c116f22009-12-16 12:19:59 +0100178 return 0;
179}
Andi Kleen27df5062009-12-21 19:56:42 +0100180#else
181int hwpoison_filter(struct page *p)
182{
183 return 0;
184}
185#endif
186
Wu Fengguang7c116f22009-12-16 12:19:59 +0100187EXPORT_SYMBOL_GPL(hwpoison_filter);
188
Andi Kleen6a460792009-09-16 11:50:15 +0200189/*
Tony Luck7329bbe2011-12-13 09:27:58 -0800190 * Send all the processes who have the page mapped a signal.
191 * ``action optional'' if they are not immediately affected by the error
192 * ``action required'' if error happened in current execution context
Andi Kleen6a460792009-09-16 11:50:15 +0200193 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800194static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
195 unsigned long pfn, struct page *page, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200196{
197 struct siginfo si;
198 int ret;
199
200 printk(KERN_ERR
Tony Luck7329bbe2011-12-13 09:27:58 -0800201 "MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200202 pfn, t->comm, t->pid);
203 si.si_signo = SIGBUS;
204 si.si_errno = 0;
Andi Kleen6a460792009-09-16 11:50:15 +0200205 si.si_addr = (void *)addr;
206#ifdef __ARCH_SI_TRAPNO
207 si.si_trapno = trapno;
208#endif
Andrea Arcangeli37c2ac72011-01-13 15:47:16 -0800209 si.si_addr_lsb = compound_trans_order(compound_head(page)) + PAGE_SHIFT;
Tony Luck7329bbe2011-12-13 09:27:58 -0800210
211 if ((flags & MF_ACTION_REQUIRED) && t == current) {
212 si.si_code = BUS_MCEERR_AR;
213 ret = force_sig_info(SIGBUS, &si, t);
214 } else {
215 /*
216 * Don't use force here, it's convenient if the signal
217 * can be temporarily blocked.
218 * This could cause a loop when the user sets SIGBUS
219 * to SIG_IGN, but hopefully no one will do that?
220 */
221 si.si_code = BUS_MCEERR_AO;
222 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
223 }
Andi Kleen6a460792009-09-16 11:50:15 +0200224 if (ret < 0)
225 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
226 t->comm, t->pid, ret);
227 return ret;
228}
229
230/*
Andi Kleen588f9ce2009-12-16 12:19:57 +0100231 * When a unknown page type is encountered drain as many buffers as possible
232 * in the hope to turn the page into a LRU or free page, which we can handle.
233 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100234void shake_page(struct page *p, int access)
Andi Kleen588f9ce2009-12-16 12:19:57 +0100235{
236 if (!PageSlab(p)) {
237 lru_add_drain_all();
238 if (PageLRU(p))
239 return;
240 drain_all_pages();
241 if (PageLRU(p) || is_free_buddy_page(p))
242 return;
243 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100244
Andi Kleen588f9ce2009-12-16 12:19:57 +0100245 /*
Jin Dongmingaf241a02011-02-01 15:52:41 -0800246 * Only call shrink_slab here (which would also shrink other caches) if
247 * access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100248 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100249 if (access) {
250 int nr;
251 do {
Ying Hana09ed5e2011-05-24 17:12:26 -0700252 struct shrink_control shrink = {
253 .gfp_mask = GFP_KERNEL,
Ying Hana09ed5e2011-05-24 17:12:26 -0700254 };
255
Ying Han1495f232011-05-24 17:12:27 -0700256 nr = shrink_slab(&shrink, 1000, 1000);
Andi Kleen47f43e72010-09-28 07:37:55 +0200257 if (page_count(p) == 1)
Andi Kleenfacb6012009-12-16 12:20:00 +0100258 break;
259 } while (nr > 10);
260 }
Andi Kleen588f9ce2009-12-16 12:19:57 +0100261}
262EXPORT_SYMBOL_GPL(shake_page);
263
264/*
Andi Kleen6a460792009-09-16 11:50:15 +0200265 * Kill all processes that have a poisoned page mapped and then isolate
266 * the page.
267 *
268 * General strategy:
269 * Find all processes having the page mapped and kill them.
270 * But we keep a page reference around so that the page is not
271 * actually freed yet.
272 * Then stash the page away
273 *
274 * There's no convenient way to get back to mapped processes
275 * from the VMAs. So do a brute-force search over all
276 * running processes.
277 *
278 * Remember that machine checks are not common (or rather
279 * if they are common you have other problems), so this shouldn't
280 * be a performance issue.
281 *
282 * Also there are some races possible while we get from the
283 * error detection to actually handle it.
284 */
285
286struct to_kill {
287 struct list_head nd;
288 struct task_struct *tsk;
289 unsigned long addr;
Andi Kleen9033ae12010-09-27 23:36:05 +0200290 char addr_valid;
Andi Kleen6a460792009-09-16 11:50:15 +0200291};
292
293/*
294 * Failure handling: if we can't find or can't kill a process there's
295 * not much we can do. We just print a message and ignore otherwise.
296 */
297
298/*
299 * Schedule a process for later kill.
300 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
301 * TBD would GFP_NOIO be enough?
302 */
303static void add_to_kill(struct task_struct *tsk, struct page *p,
304 struct vm_area_struct *vma,
305 struct list_head *to_kill,
306 struct to_kill **tkc)
307{
308 struct to_kill *tk;
309
310 if (*tkc) {
311 tk = *tkc;
312 *tkc = NULL;
313 } else {
314 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
315 if (!tk) {
316 printk(KERN_ERR
317 "MCE: Out of memory while machine check handling\n");
318 return;
319 }
320 }
321 tk->addr = page_address_in_vma(p, vma);
322 tk->addr_valid = 1;
323
324 /*
325 * In theory we don't have to kill when the page was
326 * munmaped. But it could be also a mremap. Since that's
327 * likely very rare kill anyways just out of paranoia, but use
328 * a SIGKILL because the error is not contained anymore.
329 */
330 if (tk->addr == -EFAULT) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200331 pr_info("MCE: Unable to find user space address %lx in %s\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200332 page_to_pfn(p), tsk->comm);
333 tk->addr_valid = 0;
334 }
335 get_task_struct(tsk);
336 tk->tsk = tsk;
337 list_add_tail(&tk->nd, to_kill);
338}
339
340/*
341 * Kill the processes that have been collected earlier.
342 *
343 * Only do anything when DOIT is set, otherwise just free the list
344 * (this is used for clean pages which do not need killing)
345 * Also when FAIL is set do a force kill because something went
346 * wrong earlier.
347 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800348static void kill_procs(struct list_head *to_kill, int doit, int trapno,
349 int fail, struct page *page, unsigned long pfn,
350 int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200351{
352 struct to_kill *tk, *next;
353
354 list_for_each_entry_safe (tk, next, to_kill, nd) {
355 if (doit) {
356 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200357 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200358 * make sure the process doesn't catch the
359 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200360 */
361 if (fail || tk->addr_valid == 0) {
362 printk(KERN_ERR
363 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
364 pfn, tk->tsk->comm, tk->tsk->pid);
365 force_sig(SIGKILL, tk->tsk);
366 }
367
368 /*
369 * In theory the process could have mapped
370 * something else on the address in-between. We could
371 * check for that, but we need to tell the
372 * process anyways.
373 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800374 else if (kill_proc(tk->tsk, tk->addr, trapno,
375 pfn, page, flags) < 0)
Andi Kleen6a460792009-09-16 11:50:15 +0200376 printk(KERN_ERR
377 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
378 pfn, tk->tsk->comm, tk->tsk->pid);
379 }
380 put_task_struct(tk->tsk);
381 kfree(tk);
382 }
383}
384
385static int task_early_kill(struct task_struct *tsk)
386{
387 if (!tsk->mm)
388 return 0;
389 if (tsk->flags & PF_MCE_PROCESS)
390 return !!(tsk->flags & PF_MCE_EARLY);
391 return sysctl_memory_failure_early_kill;
392}
393
394/*
395 * Collect processes when the error hit an anonymous page.
396 */
397static void collect_procs_anon(struct page *page, struct list_head *to_kill,
398 struct to_kill **tkc)
399{
400 struct vm_area_struct *vma;
401 struct task_struct *tsk;
402 struct anon_vma *av;
403
Andi Kleen6a460792009-09-16 11:50:15 +0200404 av = page_lock_anon_vma(page);
405 if (av == NULL) /* Not actually mapped anymore */
Peter Zijlstra9b679322011-06-27 16:18:09 -0700406 return;
407
408 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200409 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800410 struct anon_vma_chain *vmac;
411
Andi Kleen6a460792009-09-16 11:50:15 +0200412 if (!task_early_kill(tsk))
413 continue;
Rik van Riel5beb4932010-03-05 13:42:07 -0800414 list_for_each_entry(vmac, &av->head, same_anon_vma) {
415 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200416 if (!page_mapped_in_vma(page, vma))
417 continue;
418 if (vma->vm_mm == tsk->mm)
419 add_to_kill(tsk, page, vma, to_kill, tkc);
420 }
421 }
Andi Kleen6a460792009-09-16 11:50:15 +0200422 read_unlock(&tasklist_lock);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700423 page_unlock_anon_vma(av);
Andi Kleen6a460792009-09-16 11:50:15 +0200424}
425
426/*
427 * Collect processes when the error hit a file mapped page.
428 */
429static void collect_procs_file(struct page *page, struct list_head *to_kill,
430 struct to_kill **tkc)
431{
432 struct vm_area_struct *vma;
433 struct task_struct *tsk;
434 struct prio_tree_iter iter;
435 struct address_space *mapping = page->mapping;
436
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700437 mutex_lock(&mapping->i_mmap_mutex);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700438 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200439 for_each_process(tsk) {
440 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
441
442 if (!task_early_kill(tsk))
443 continue;
444
445 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
446 pgoff) {
447 /*
448 * Send early kill signal to tasks where a vma covers
449 * the page but the corrupted page is not necessarily
450 * mapped it in its pte.
451 * Assume applications who requested early kill want
452 * to be informed of all such data corruptions.
453 */
454 if (vma->vm_mm == tsk->mm)
455 add_to_kill(tsk, page, vma, to_kill, tkc);
456 }
457 }
Andi Kleen6a460792009-09-16 11:50:15 +0200458 read_unlock(&tasklist_lock);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700459 mutex_unlock(&mapping->i_mmap_mutex);
Andi Kleen6a460792009-09-16 11:50:15 +0200460}
461
462/*
463 * Collect the processes who have the corrupted page mapped to kill.
464 * This is done in two steps for locking reasons.
465 * First preallocate one tokill structure outside the spin locks,
466 * so that we can kill at least one process reasonably reliable.
467 */
468static void collect_procs(struct page *page, struct list_head *tokill)
469{
470 struct to_kill *tk;
471
472 if (!page->mapping)
473 return;
474
475 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
476 if (!tk)
477 return;
478 if (PageAnon(page))
479 collect_procs_anon(page, tokill, &tk);
480 else
481 collect_procs_file(page, tokill, &tk);
482 kfree(tk);
483}
484
485/*
486 * Error handlers for various types of pages.
487 */
488
489enum outcome {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100490 IGNORED, /* Error: cannot be handled */
491 FAILED, /* Error: handling failed */
Andi Kleen6a460792009-09-16 11:50:15 +0200492 DELAYED, /* Will be handled later */
Andi Kleen6a460792009-09-16 11:50:15 +0200493 RECOVERED, /* Successfully recovered */
494};
495
496static const char *action_name[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100497 [IGNORED] = "Ignored",
Andi Kleen6a460792009-09-16 11:50:15 +0200498 [FAILED] = "Failed",
499 [DELAYED] = "Delayed",
Andi Kleen6a460792009-09-16 11:50:15 +0200500 [RECOVERED] = "Recovered",
501};
502
503/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100504 * XXX: It is possible that a page is isolated from LRU cache,
505 * and then kept in swap cache or failed to remove from page cache.
506 * The page count will stop it from being freed by unpoison.
507 * Stress tests should be aware of this memory leak problem.
508 */
509static int delete_from_lru_cache(struct page *p)
510{
511 if (!isolate_lru_page(p)) {
512 /*
513 * Clear sensible page flags, so that the buddy system won't
514 * complain when the page is unpoison-and-freed.
515 */
516 ClearPageActive(p);
517 ClearPageUnevictable(p);
518 /*
519 * drop the page count elevated by isolate_lru_page()
520 */
521 page_cache_release(p);
522 return 0;
523 }
524 return -EIO;
525}
526
527/*
Andi Kleen6a460792009-09-16 11:50:15 +0200528 * Error hit kernel page.
529 * Do nothing, try to be lucky and not touch this instead. For a few cases we
530 * could be more sophisticated.
531 */
532static int me_kernel(struct page *p, unsigned long pfn)
533{
Andi Kleen6a460792009-09-16 11:50:15 +0200534 return IGNORED;
535}
536
537/*
538 * Page in unknown state. Do nothing.
539 */
540static int me_unknown(struct page *p, unsigned long pfn)
541{
542 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
543 return FAILED;
544}
545
546/*
Andi Kleen6a460792009-09-16 11:50:15 +0200547 * Clean (or cleaned) page cache page.
548 */
549static int me_pagecache_clean(struct page *p, unsigned long pfn)
550{
551 int err;
552 int ret = FAILED;
553 struct address_space *mapping;
554
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100555 delete_from_lru_cache(p);
556
Andi Kleen6a460792009-09-16 11:50:15 +0200557 /*
558 * For anonymous pages we're done the only reference left
559 * should be the one m_f() holds.
560 */
561 if (PageAnon(p))
562 return RECOVERED;
563
564 /*
565 * Now truncate the page in the page cache. This is really
566 * more like a "temporary hole punch"
567 * Don't do this for block devices when someone else
568 * has a reference, because it could be file system metadata
569 * and that's not safe to truncate.
570 */
571 mapping = page_mapping(p);
572 if (!mapping) {
573 /*
574 * Page has been teared down in the meanwhile
575 */
576 return FAILED;
577 }
578
579 /*
580 * Truncation is a bit tricky. Enable it per file system for now.
581 *
582 * Open: to take i_mutex or not for this? Right now we don't.
583 */
584 if (mapping->a_ops->error_remove_page) {
585 err = mapping->a_ops->error_remove_page(mapping, p);
586 if (err != 0) {
587 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
588 pfn, err);
589 } else if (page_has_private(p) &&
590 !try_to_release_page(p, GFP_NOIO)) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200591 pr_info("MCE %#lx: failed to release buffers\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200592 } else {
593 ret = RECOVERED;
594 }
595 } else {
596 /*
597 * If the file system doesn't support it just invalidate
598 * This fails on dirty or anything with private pages
599 */
600 if (invalidate_inode_page(p))
601 ret = RECOVERED;
602 else
603 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
604 pfn);
605 }
606 return ret;
607}
608
609/*
610 * Dirty cache page page
611 * Issues: when the error hit a hole page the error is not properly
612 * propagated.
613 */
614static int me_pagecache_dirty(struct page *p, unsigned long pfn)
615{
616 struct address_space *mapping = page_mapping(p);
617
618 SetPageError(p);
619 /* TBD: print more information about the file. */
620 if (mapping) {
621 /*
622 * IO error will be reported by write(), fsync(), etc.
623 * who check the mapping.
624 * This way the application knows that something went
625 * wrong with its dirty file data.
626 *
627 * There's one open issue:
628 *
629 * The EIO will be only reported on the next IO
630 * operation and then cleared through the IO map.
631 * Normally Linux has two mechanisms to pass IO error
632 * first through the AS_EIO flag in the address space
633 * and then through the PageError flag in the page.
634 * Since we drop pages on memory failure handling the
635 * only mechanism open to use is through AS_AIO.
636 *
637 * This has the disadvantage that it gets cleared on
638 * the first operation that returns an error, while
639 * the PageError bit is more sticky and only cleared
640 * when the page is reread or dropped. If an
641 * application assumes it will always get error on
642 * fsync, but does other operations on the fd before
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300643 * and the page is dropped between then the error
Andi Kleen6a460792009-09-16 11:50:15 +0200644 * will not be properly reported.
645 *
646 * This can already happen even without hwpoisoned
647 * pages: first on metadata IO errors (which only
648 * report through AS_EIO) or when the page is dropped
649 * at the wrong time.
650 *
651 * So right now we assume that the application DTRT on
652 * the first EIO, but we're not worse than other parts
653 * of the kernel.
654 */
655 mapping_set_error(mapping, EIO);
656 }
657
658 return me_pagecache_clean(p, pfn);
659}
660
661/*
662 * Clean and dirty swap cache.
663 *
664 * Dirty swap cache page is tricky to handle. The page could live both in page
665 * cache and swap cache(ie. page is freshly swapped in). So it could be
666 * referenced concurrently by 2 types of PTEs:
667 * normal PTEs and swap PTEs. We try to handle them consistently by calling
668 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
669 * and then
670 * - clear dirty bit to prevent IO
671 * - remove from LRU
672 * - but keep in the swap cache, so that when we return to it on
673 * a later page fault, we know the application is accessing
674 * corrupted data and shall be killed (we installed simple
675 * interception code in do_swap_page to catch it).
676 *
677 * Clean swap cache pages can be directly isolated. A later page fault will
678 * bring in the known good data from disk.
679 */
680static int me_swapcache_dirty(struct page *p, unsigned long pfn)
681{
Andi Kleen6a460792009-09-16 11:50:15 +0200682 ClearPageDirty(p);
683 /* Trigger EIO in shmem: */
684 ClearPageUptodate(p);
685
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100686 if (!delete_from_lru_cache(p))
687 return DELAYED;
688 else
689 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200690}
691
692static int me_swapcache_clean(struct page *p, unsigned long pfn)
693{
Andi Kleen6a460792009-09-16 11:50:15 +0200694 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800695
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100696 if (!delete_from_lru_cache(p))
697 return RECOVERED;
698 else
699 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200700}
701
702/*
703 * Huge pages. Needs work.
704 * Issues:
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900705 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
706 * To narrow down kill region to one page, we need to break up pmd.
Andi Kleen6a460792009-09-16 11:50:15 +0200707 */
708static int me_huge_page(struct page *p, unsigned long pfn)
709{
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900710 int res = 0;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900711 struct page *hpage = compound_head(p);
712 /*
713 * We can safely recover from error on free or reserved (i.e.
714 * not in-use) hugepage by dequeuing it from freelist.
715 * To check whether a hugepage is in-use or not, we can't use
716 * page->lru because it can be used in other hugepage operations,
717 * such as __unmap_hugepage_range() and gather_surplus_pages().
718 * So instead we use page_mapping() and PageAnon().
719 * We assume that this function is called with page lock held,
720 * so there is no race between isolation and mapping/unmapping.
721 */
722 if (!(page_mapping(hpage) || PageAnon(hpage))) {
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900723 res = dequeue_hwpoisoned_huge_page(hpage);
724 if (!res)
725 return RECOVERED;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900726 }
727 return DELAYED;
Andi Kleen6a460792009-09-16 11:50:15 +0200728}
729
730/*
731 * Various page states we can handle.
732 *
733 * A page state is defined by its current page->flags bits.
734 * The table matches them in order and calls the right handler.
735 *
736 * This is quite tricky because we can access page at any time
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300737 * in its live cycle, so all accesses have to be extremely careful.
Andi Kleen6a460792009-09-16 11:50:15 +0200738 *
739 * This is not complete. More states could be added.
740 * For any missing state don't attempt recovery.
741 */
742
743#define dirty (1UL << PG_dirty)
744#define sc (1UL << PG_swapcache)
745#define unevict (1UL << PG_unevictable)
746#define mlock (1UL << PG_mlocked)
747#define writeback (1UL << PG_writeback)
748#define lru (1UL << PG_lru)
749#define swapbacked (1UL << PG_swapbacked)
750#define head (1UL << PG_head)
751#define tail (1UL << PG_tail)
752#define compound (1UL << PG_compound)
753#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200754#define reserved (1UL << PG_reserved)
755
756static struct page_state {
757 unsigned long mask;
758 unsigned long res;
759 char *msg;
760 int (*action)(struct page *p, unsigned long pfn);
761} error_states[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100762 { reserved, reserved, "reserved kernel", me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100763 /*
764 * free pages are specially detected outside this table:
765 * PG_buddy pages only make a small fraction of all free pages.
766 */
Andi Kleen6a460792009-09-16 11:50:15 +0200767
768 /*
769 * Could in theory check if slab page is free or if we can drop
770 * currently unused objects without touching them. But just
771 * treat it as standard kernel for now.
772 */
773 { slab, slab, "kernel slab", me_kernel },
774
775#ifdef CONFIG_PAGEFLAGS_EXTENDED
776 { head, head, "huge", me_huge_page },
777 { tail, tail, "huge", me_huge_page },
778#else
779 { compound, compound, "huge", me_huge_page },
780#endif
781
782 { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
783 { sc|dirty, sc, "swapcache", me_swapcache_clean },
784
785 { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
786 { unevict, unevict, "unevictable LRU", me_pagecache_clean},
787
Andi Kleen6a460792009-09-16 11:50:15 +0200788 { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
789 { mlock, mlock, "mlocked LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200790
791 { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
792 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200793
794 /*
795 * Catchall entry: must be at end.
796 */
797 { 0, 0, "unknown page state", me_unknown },
798};
799
Andi Kleen2326c462009-12-16 12:20:00 +0100800#undef dirty
801#undef sc
802#undef unevict
803#undef mlock
804#undef writeback
805#undef lru
806#undef swapbacked
807#undef head
808#undef tail
809#undef compound
810#undef slab
811#undef reserved
812
Andi Kleen6a460792009-09-16 11:50:15 +0200813static void action_result(unsigned long pfn, char *msg, int result)
814{
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100815 struct page *page = pfn_to_page(pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200816
817 printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
818 pfn,
Wu Fengguanga7560fc2009-12-16 12:19:57 +0100819 PageDirty(page) ? "dirty " : "",
Andi Kleen6a460792009-09-16 11:50:15 +0200820 msg, action_name[result]);
821}
822
823static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100824 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200825{
826 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200827 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200828
829 result = ps->action(p, pfn);
830 action_result(pfn, ps->msg, result);
Wu Fengguang7456b042009-10-19 08:15:01 +0200831
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100832 count = page_count(p) - 1;
Wu Fengguang138ce282009-12-16 12:19:58 +0100833 if (ps->action == me_swapcache_dirty && result == DELAYED)
834 count--;
835 if (count != 0) {
Andi Kleen6a460792009-09-16 11:50:15 +0200836 printk(KERN_ERR
837 "MCE %#lx: %s page still referenced by %d users\n",
Wu Fengguang7456b042009-10-19 08:15:01 +0200838 pfn, ps->msg, count);
Wu Fengguang138ce282009-12-16 12:19:58 +0100839 result = FAILED;
840 }
Andi Kleen6a460792009-09-16 11:50:15 +0200841
842 /* Could do more checks here if page looks ok */
843 /*
844 * Could adjust zone counters here to correct for the missing page.
845 */
846
Wu Fengguang138ce282009-12-16 12:19:58 +0100847 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200848}
849
Andi Kleen6a460792009-09-16 11:50:15 +0200850/*
851 * Do all that is necessary to remove user space mappings. Unmap
852 * the pages and send SIGBUS to the processes if the data was dirty.
853 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100854static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Tony Luck7329bbe2011-12-13 09:27:58 -0800855 int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200856{
857 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
858 struct address_space *mapping;
859 LIST_HEAD(tokill);
860 int ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200861 int kill = 1;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900862 struct page *hpage = compound_head(p);
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800863 struct page *ppage;
Andi Kleen6a460792009-09-16 11:50:15 +0200864
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100865 if (PageReserved(p) || PageSlab(p))
866 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200867
Andi Kleen6a460792009-09-16 11:50:15 +0200868 /*
869 * This check implies we don't kill processes if their pages
870 * are in the swap cache early. Those are always late kills.
871 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900872 if (!page_mapped(hpage))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100873 return SWAP_SUCCESS;
874
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900875 if (PageKsm(p))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100876 return SWAP_FAIL;
Andi Kleen6a460792009-09-16 11:50:15 +0200877
878 if (PageSwapCache(p)) {
879 printk(KERN_ERR
880 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
881 ttu |= TTU_IGNORE_HWPOISON;
882 }
883
884 /*
885 * Propagate the dirty bit from PTEs to struct page first, because we
886 * need this to decide if we should kill or just drop the page.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100887 * XXX: the dirty test could be racy: set_page_dirty() may not always
888 * be called inside page lock (it's recommended but not enforced).
Andi Kleen6a460792009-09-16 11:50:15 +0200889 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900890 mapping = page_mapping(hpage);
891 if (!PageDirty(hpage) && mapping &&
892 mapping_cap_writeback_dirty(mapping)) {
893 if (page_mkclean(hpage)) {
894 SetPageDirty(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +0200895 } else {
896 kill = 0;
897 ttu |= TTU_IGNORE_HWPOISON;
898 printk(KERN_INFO
899 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
900 pfn);
901 }
902 }
903
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800904 /*
905 * ppage: poisoned page
906 * if p is regular page(4k page)
907 * ppage == real poisoned page;
908 * else p is hugetlb or THP, ppage == head page.
909 */
910 ppage = hpage;
911
Jin Dongmingefeda7a2011-02-01 15:52:39 -0800912 if (PageTransHuge(hpage)) {
913 /*
914 * Verify that this isn't a hugetlbfs head page, the check for
915 * PageAnon is just for avoid tripping a split_huge_page
916 * internal debug check, as split_huge_page refuses to deal with
917 * anything that isn't an anon page. PageAnon can't go away fro
918 * under us because we hold a refcount on the hpage, without a
919 * refcount on the hpage. split_huge_page can't be safely called
920 * in the first place, having a refcount on the tail isn't
921 * enough * to be safe.
922 */
923 if (!PageHuge(hpage) && PageAnon(hpage)) {
924 if (unlikely(split_huge_page(hpage))) {
925 /*
926 * FIXME: if splitting THP is failed, it is
927 * better to stop the following operation rather
928 * than causing panic by unmapping. System might
929 * survive if the page is freed later.
930 */
931 printk(KERN_INFO
932 "MCE %#lx: failed to split THP\n", pfn);
933
934 BUG_ON(!PageHWPoison(p));
935 return SWAP_FAIL;
936 }
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800937 /* THP is split, so ppage should be the real poisoned page. */
938 ppage = p;
Jin Dongmingefeda7a2011-02-01 15:52:39 -0800939 }
940 }
941
Andi Kleen6a460792009-09-16 11:50:15 +0200942 /*
943 * First collect all the processes that have the page
944 * mapped in dirty form. This has to be done before try_to_unmap,
945 * because ttu takes the rmap data structures down.
946 *
947 * Error handling: We ignore errors here because
948 * there's nothing that can be done.
949 */
950 if (kill)
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800951 collect_procs(ppage, &tokill);
Andi Kleen6a460792009-09-16 11:50:15 +0200952
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800953 if (hpage != ppage)
Jens Axboe7eaceac2011-03-10 08:52:07 +0100954 lock_page(ppage);
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800955
956 ret = try_to_unmap(ppage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +0200957 if (ret != SWAP_SUCCESS)
958 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800959 pfn, page_mapcount(ppage));
960
961 if (hpage != ppage)
962 unlock_page(ppage);
Andi Kleen6a460792009-09-16 11:50:15 +0200963
964 /*
965 * Now that the dirty bit has been propagated to the
966 * struct page and all unmaps done we can decide if
967 * killing is needed or not. Only kill when the page
968 * was dirty, otherwise the tokill list is merely
969 * freed. When there was a problem unmapping earlier
970 * use a more force-full uncatchable kill to prevent
971 * any accesses to the poisoned memory.
972 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800973 kill_procs(&tokill, !!PageDirty(ppage), trapno,
974 ret != SWAP_SUCCESS, p, pfn, flags);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100975
976 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200977}
978
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900979static void set_page_hwpoison_huge_page(struct page *hpage)
980{
981 int i;
Andrea Arcangeli37c2ac72011-01-13 15:47:16 -0800982 int nr_pages = 1 << compound_trans_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900983 for (i = 0; i < nr_pages; i++)
984 SetPageHWPoison(hpage + i);
985}
986
987static void clear_page_hwpoison_huge_page(struct page *hpage)
988{
989 int i;
Andrea Arcangeli37c2ac72011-01-13 15:47:16 -0800990 int nr_pages = 1 << compound_trans_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900991 for (i = 0; i < nr_pages; i++)
992 ClearPageHWPoison(hpage + i);
993}
994
Tony Luckcd42f4a2011-12-15 10:48:12 -0800995/**
996 * memory_failure - Handle memory failure of a page.
997 * @pfn: Page Number of the corrupted page
998 * @trapno: Trap number reported in the signal to user space.
999 * @flags: fine tune action taken
1000 *
1001 * This function is called by the low level machine check code
1002 * of an architecture when it detects hardware memory corruption
1003 * of a page. It tries its best to recover, which includes
1004 * dropping pages, killing processes etc.
1005 *
1006 * The function is primarily of use for corruptions that
1007 * happen outside the current execution context (e.g. when
1008 * detected by a background scrubber)
1009 *
1010 * Must run in process context (e.g. a work queue) with interrupts
1011 * enabled and no spinlocks hold.
1012 */
1013int memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +02001014{
1015 struct page_state *ps;
1016 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001017 struct page *hpage;
Andi Kleen6a460792009-09-16 11:50:15 +02001018 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001019 unsigned int nr_pages;
Andi Kleen6a460792009-09-16 11:50:15 +02001020
1021 if (!sysctl_memory_failure_recovery)
1022 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1023
1024 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +01001025 printk(KERN_ERR
1026 "MCE %#lx: memory outside kernel control\n",
1027 pfn);
1028 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +02001029 }
1030
1031 p = pfn_to_page(pfn);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001032 hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +02001033 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001034 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001035 return 0;
1036 }
1037
Andrea Arcangeli37c2ac72011-01-13 15:47:16 -08001038 nr_pages = 1 << compound_trans_order(hpage);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001039 atomic_long_add(nr_pages, &mce_bad_pages);
Andi Kleen6a460792009-09-16 11:50:15 +02001040
1041 /*
1042 * We need/can do nothing about count=0 pages.
1043 * 1) it's a free page, and therefore in safe hand:
1044 * prep_new_page() will be the gate keeper.
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001045 * 2) it's a free hugepage, which is also safe:
1046 * an affected hugepage will be dequeued from hugepage freelist,
1047 * so there's no concern about reusing it ever after.
1048 * 3) it's part of a non-compound high order page.
Andi Kleen6a460792009-09-16 11:50:15 +02001049 * Implies some kernel user: cannot stop them from
1050 * R/W the page; let's pray that the page has been
1051 * used and will be freed some time later.
1052 * In fact it's dangerous to directly bump up page count from 0,
1053 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1054 */
Andi Kleen82ba0112009-12-16 12:19:57 +01001055 if (!(flags & MF_COUNT_INCREASED) &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001056 !get_page_unless_zero(hpage)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001057 if (is_free_buddy_page(p)) {
1058 action_result(pfn, "free buddy", DELAYED);
1059 return 0;
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001060 } else if (PageHuge(hpage)) {
1061 /*
1062 * Check "just unpoisoned", "filter hit", and
1063 * "race with other subpage."
1064 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001065 lock_page(hpage);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001066 if (!PageHWPoison(hpage)
1067 || (hwpoison_filter(p) && TestClearPageHWPoison(p))
1068 || (p != hpage && TestSetPageHWPoison(hpage))) {
1069 atomic_long_sub(nr_pages, &mce_bad_pages);
1070 return 0;
1071 }
1072 set_page_hwpoison_huge_page(hpage);
1073 res = dequeue_hwpoisoned_huge_page(hpage);
1074 action_result(pfn, "free huge",
1075 res ? IGNORED : DELAYED);
1076 unlock_page(hpage);
1077 return res;
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001078 } else {
1079 action_result(pfn, "high order kernel", IGNORED);
1080 return -EBUSY;
1081 }
Andi Kleen6a460792009-09-16 11:50:15 +02001082 }
1083
1084 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001085 * We ignore non-LRU pages for good reasons.
1086 * - PG_locked is only well defined for LRU pages and a few others
1087 * - to avoid races with __set_page_locked()
1088 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1089 * The check (unnecessarily) ignores LRU pages being isolated and
1090 * walked by the page reclaim code, however that's not a big loss.
1091 */
Dean Nelson385de352012-03-21 16:34:05 -07001092 if (!PageHuge(p) && !PageTransTail(p)) {
Jin Dongmingaf241a02011-02-01 15:52:41 -08001093 if (!PageLRU(p))
1094 shake_page(p, 0);
1095 if (!PageLRU(p)) {
1096 /*
1097 * shake_page could have turned it free.
1098 */
1099 if (is_free_buddy_page(p)) {
1100 action_result(pfn, "free buddy, 2nd try",
1101 DELAYED);
1102 return 0;
1103 }
1104 action_result(pfn, "non LRU", IGNORED);
1105 put_page(p);
1106 return -EBUSY;
Andi Kleen0474a602009-12-16 12:20:00 +01001107 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001108 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001109
1110 /*
Andi Kleen6a460792009-09-16 11:50:15 +02001111 * Lock the page and wait for writeback to finish.
1112 * It's very difficult to mess with pages currently under IO
1113 * and in many cases impossible, so we just avoid it here.
1114 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001115 lock_page(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001116
1117 /*
1118 * unpoison always clear PG_hwpoison inside page lock
1119 */
1120 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001121 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001122 res = 0;
1123 goto out;
1124 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001125 if (hwpoison_filter(p)) {
1126 if (TestClearPageHWPoison(p))
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001127 atomic_long_sub(nr_pages, &mce_bad_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001128 unlock_page(hpage);
1129 put_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001130 return 0;
1131 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001132
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001133 /*
1134 * For error on the tail page, we should set PG_hwpoison
1135 * on the head page to show that the hugepage is hwpoisoned
1136 */
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001137 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001138 action_result(pfn, "hugepage already hardware poisoned",
1139 IGNORED);
1140 unlock_page(hpage);
1141 put_page(hpage);
1142 return 0;
1143 }
1144 /*
1145 * Set PG_hwpoison on all pages in an error hugepage,
1146 * because containment is done in hugepage unit for now.
1147 * Since we have done TestSetPageHWPoison() for the head page with
1148 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1149 */
1150 if (PageHuge(p))
1151 set_page_hwpoison_huge_page(hpage);
1152
Andi Kleen6a460792009-09-16 11:50:15 +02001153 wait_on_page_writeback(p);
1154
1155 /*
1156 * Now take care of user space mappings.
Minchan Kime64a7822011-03-22 16:32:44 -07001157 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
Andi Kleen6a460792009-09-16 11:50:15 +02001158 */
Tony Luck7329bbe2011-12-13 09:27:58 -08001159 if (hwpoison_user_mappings(p, pfn, trapno, flags) != SWAP_SUCCESS) {
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001160 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1161 res = -EBUSY;
1162 goto out;
1163 }
Andi Kleen6a460792009-09-16 11:50:15 +02001164
1165 /*
1166 * Torn down by someone else?
1167 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001168 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Andi Kleen6a460792009-09-16 11:50:15 +02001169 action_result(pfn, "already truncated LRU", IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001170 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001171 goto out;
1172 }
1173
1174 res = -EBUSY;
1175 for (ps = error_states;; ps++) {
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001176 if ((p->flags & ps->mask) == ps->res) {
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +01001177 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001178 break;
1179 }
1180 }
1181out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001182 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001183 return res;
1184}
Tony Luckcd42f4a2011-12-15 10:48:12 -08001185EXPORT_SYMBOL_GPL(memory_failure);
Wu Fengguang847ce402009-12-16 12:19:58 +01001186
Huang Yingea8f5fb2011-07-13 13:14:27 +08001187#define MEMORY_FAILURE_FIFO_ORDER 4
1188#define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1189
1190struct memory_failure_entry {
1191 unsigned long pfn;
1192 int trapno;
1193 int flags;
1194};
1195
1196struct memory_failure_cpu {
1197 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1198 MEMORY_FAILURE_FIFO_SIZE);
1199 spinlock_t lock;
1200 struct work_struct work;
1201};
1202
1203static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1204
1205/**
1206 * memory_failure_queue - Schedule handling memory failure of a page.
1207 * @pfn: Page Number of the corrupted page
1208 * @trapno: Trap number reported in the signal to user space.
1209 * @flags: Flags for memory failure handling
1210 *
1211 * This function is called by the low level hardware error handler
1212 * when it detects hardware memory corruption of a page. It schedules
1213 * the recovering of error page, including dropping pages, killing
1214 * processes etc.
1215 *
1216 * The function is primarily of use for corruptions that
1217 * happen outside the current execution context (e.g. when
1218 * detected by a background scrubber)
1219 *
1220 * Can run in IRQ context.
1221 */
1222void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1223{
1224 struct memory_failure_cpu *mf_cpu;
1225 unsigned long proc_flags;
1226 struct memory_failure_entry entry = {
1227 .pfn = pfn,
1228 .trapno = trapno,
1229 .flags = flags,
1230 };
1231
1232 mf_cpu = &get_cpu_var(memory_failure_cpu);
1233 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1234 if (kfifo_put(&mf_cpu->fifo, &entry))
1235 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1236 else
1237 pr_err("Memory failure: buffer overflow when queuing memory failure at 0x%#lx\n",
1238 pfn);
1239 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1240 put_cpu_var(memory_failure_cpu);
1241}
1242EXPORT_SYMBOL_GPL(memory_failure_queue);
1243
1244static void memory_failure_work_func(struct work_struct *work)
1245{
1246 struct memory_failure_cpu *mf_cpu;
1247 struct memory_failure_entry entry = { 0, };
1248 unsigned long proc_flags;
1249 int gotten;
1250
1251 mf_cpu = &__get_cpu_var(memory_failure_cpu);
1252 for (;;) {
1253 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1254 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1255 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1256 if (!gotten)
1257 break;
Tony Luckcd42f4a2011-12-15 10:48:12 -08001258 memory_failure(entry.pfn, entry.trapno, entry.flags);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001259 }
1260}
1261
1262static int __init memory_failure_init(void)
1263{
1264 struct memory_failure_cpu *mf_cpu;
1265 int cpu;
1266
1267 for_each_possible_cpu(cpu) {
1268 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1269 spin_lock_init(&mf_cpu->lock);
1270 INIT_KFIFO(mf_cpu->fifo);
1271 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1272 }
1273
1274 return 0;
1275}
1276core_initcall(memory_failure_init);
1277
Wu Fengguang847ce402009-12-16 12:19:58 +01001278/**
1279 * unpoison_memory - Unpoison a previously poisoned page
1280 * @pfn: Page number of the to be unpoisoned page
1281 *
1282 * Software-unpoison a page that has been poisoned by
1283 * memory_failure() earlier.
1284 *
1285 * This is only done on the software-level, so it only works
1286 * for linux injected failures, not real hardware failures
1287 *
1288 * Returns 0 for success, otherwise -errno.
1289 */
1290int unpoison_memory(unsigned long pfn)
1291{
1292 struct page *page;
1293 struct page *p;
1294 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001295 unsigned int nr_pages;
Wu Fengguang847ce402009-12-16 12:19:58 +01001296
1297 if (!pfn_valid(pfn))
1298 return -ENXIO;
1299
1300 p = pfn_to_page(pfn);
1301 page = compound_head(p);
1302
1303 if (!PageHWPoison(p)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001304 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001305 return 0;
1306 }
1307
Andrea Arcangeli37c2ac72011-01-13 15:47:16 -08001308 nr_pages = 1 << compound_trans_order(page);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001309
Wu Fengguang847ce402009-12-16 12:19:58 +01001310 if (!get_page_unless_zero(page)) {
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001311 /*
1312 * Since HWPoisoned hugepage should have non-zero refcount,
1313 * race between memory failure and unpoison seems to happen.
1314 * In such case unpoison fails and memory failure runs
1315 * to the end.
1316 */
1317 if (PageHuge(page)) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001318 pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001319 return 0;
1320 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001321 if (TestClearPageHWPoison(p))
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001322 atomic_long_sub(nr_pages, &mce_bad_pages);
Andi Kleenfb46e732010-09-27 23:31:30 +02001323 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001324 return 0;
1325 }
1326
Jens Axboe7eaceac2011-03-10 08:52:07 +01001327 lock_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001328 /*
1329 * This test is racy because PG_hwpoison is set outside of page lock.
1330 * That's acceptable because that won't trigger kernel panic. Instead,
1331 * the PG_hwpoison page will be caught and isolated on the entrance to
1332 * the free buddy page pool.
1333 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001334 if (TestClearPageHWPoison(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001335 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001336 atomic_long_sub(nr_pages, &mce_bad_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001337 freeit = 1;
Naoya Horiguchi6a901812010-09-08 10:19:40 +09001338 if (PageHuge(page))
1339 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001340 }
1341 unlock_page(page);
1342
1343 put_page(page);
1344 if (freeit)
1345 put_page(page);
1346
1347 return 0;
1348}
1349EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001350
1351static struct page *new_page(struct page *p, unsigned long private, int **x)
1352{
Andi Kleen12686d12009-12-16 12:20:01 +01001353 int nid = page_to_nid(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001354 if (PageHuge(p))
1355 return alloc_huge_page_node(page_hstate(compound_head(p)),
1356 nid);
1357 else
1358 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001359}
1360
1361/*
1362 * Safely get reference count of an arbitrary page.
1363 * Returns 0 for a free page, -EIO for a zero refcount page
1364 * that is not free, and 1 for any other page type.
1365 * For 1 the page is returned with increased page count, otherwise not.
1366 */
1367static int get_any_page(struct page *p, unsigned long pfn, int flags)
1368{
1369 int ret;
1370
1371 if (flags & MF_COUNT_INCREASED)
1372 return 1;
1373
1374 /*
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -08001375 * The lock_memory_hotplug prevents a race with memory hotplug.
Andi Kleenfacb6012009-12-16 12:20:00 +01001376 * This is a big hammer, a better would be nicer.
1377 */
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -08001378 lock_memory_hotplug();
Andi Kleenfacb6012009-12-16 12:20:00 +01001379
1380 /*
1381 * Isolate the page, so that it doesn't get reallocated if it
1382 * was free.
1383 */
1384 set_migratetype_isolate(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001385 /*
1386 * When the target page is a free hugepage, just remove it
1387 * from free hugepage list.
1388 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001389 if (!get_page_unless_zero(compound_head(p))) {
Naoya Horiguchid950b952010-09-08 10:19:39 +09001390 if (PageHuge(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001391 pr_info("%s: %#lx free huge page\n", __func__, pfn);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001392 ret = dequeue_hwpoisoned_huge_page(compound_head(p));
1393 } else if (is_free_buddy_page(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001394 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001395 /* Set hwpoison bit while page is still isolated */
1396 SetPageHWPoison(p);
1397 ret = 0;
1398 } else {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001399 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1400 __func__, pfn, p->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001401 ret = -EIO;
1402 }
1403 } else {
1404 /* Not a free page */
1405 ret = 1;
1406 }
Michal Nazarewicz0815f3d2012-04-03 15:06:15 +02001407 unset_migratetype_isolate(p, MIGRATE_MOVABLE);
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -08001408 unlock_memory_hotplug();
Andi Kleenfacb6012009-12-16 12:20:00 +01001409 return ret;
1410}
1411
Naoya Horiguchid950b952010-09-08 10:19:39 +09001412static int soft_offline_huge_page(struct page *page, int flags)
1413{
1414 int ret;
1415 unsigned long pfn = page_to_pfn(page);
1416 struct page *hpage = compound_head(page);
1417 LIST_HEAD(pagelist);
1418
1419 ret = get_any_page(page, pfn, flags);
1420 if (ret < 0)
1421 return ret;
1422 if (ret == 0)
1423 goto done;
1424
1425 if (PageHWPoison(hpage)) {
1426 put_page(hpage);
Dean Nelsondd73e852011-10-31 17:09:04 -07001427 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001428 return -EBUSY;
1429 }
1430
1431 /* Keep page count to indicate a given hugepage is isolated. */
1432
1433 list_add(&hpage->lru, &pagelist);
Mel Gorman77f1fe62011-01-13 15:45:57 -08001434 ret = migrate_huge_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0,
1435 true);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001436 if (ret) {
Minchan Kim48db54e2011-02-01 15:52:33 -08001437 struct page *page1, *page2;
1438 list_for_each_entry_safe(page1, page2, &pagelist, lru)
1439 put_page(page1);
1440
Dean Nelsondd73e852011-10-31 17:09:04 -07001441 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1442 pfn, ret, page->flags);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001443 if (ret > 0)
1444 ret = -EIO;
1445 return ret;
1446 }
1447done:
1448 if (!PageHWPoison(hpage))
Andrea Arcangeli37c2ac72011-01-13 15:47:16 -08001449 atomic_long_add(1 << compound_trans_order(hpage), &mce_bad_pages);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001450 set_page_hwpoison_huge_page(hpage);
1451 dequeue_hwpoisoned_huge_page(hpage);
1452 /* keep elevated page count for bad page */
1453 return ret;
1454}
1455
Andi Kleenfacb6012009-12-16 12:20:00 +01001456/**
1457 * soft_offline_page - Soft offline a page.
1458 * @page: page to offline
1459 * @flags: flags. Same as memory_failure().
1460 *
1461 * Returns 0 on success, otherwise negated errno.
1462 *
1463 * Soft offline a page, by migration or invalidation,
1464 * without killing anything. This is for the case when
1465 * a page is not corrupted yet (so it's still valid to access),
1466 * but has had a number of corrected errors and is better taken
1467 * out.
1468 *
1469 * The actual policy on when to do that is maintained by
1470 * user space.
1471 *
1472 * This should never impact any application or cause data loss,
1473 * however it might take some time.
1474 *
1475 * This is not a 100% solution for all memory, but tries to be
1476 * ``good enough'' for the majority of memory.
1477 */
1478int soft_offline_page(struct page *page, int flags)
1479{
1480 int ret;
1481 unsigned long pfn = page_to_pfn(page);
1482
Naoya Horiguchid950b952010-09-08 10:19:39 +09001483 if (PageHuge(page))
1484 return soft_offline_huge_page(page, flags);
1485
Andi Kleenfacb6012009-12-16 12:20:00 +01001486 ret = get_any_page(page, pfn, flags);
1487 if (ret < 0)
1488 return ret;
1489 if (ret == 0)
1490 goto done;
1491
1492 /*
1493 * Page cache page we can handle?
1494 */
1495 if (!PageLRU(page)) {
1496 /*
1497 * Try to free it.
1498 */
1499 put_page(page);
1500 shake_page(page, 1);
1501
1502 /*
1503 * Did it turn free?
1504 */
1505 ret = get_any_page(page, pfn, 0);
1506 if (ret < 0)
1507 return ret;
1508 if (ret == 0)
1509 goto done;
1510 }
1511 if (!PageLRU(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001512 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
Dean Nelsondd73e852011-10-31 17:09:04 -07001513 pfn, page->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001514 return -EIO;
1515 }
1516
1517 lock_page(page);
1518 wait_on_page_writeback(page);
1519
1520 /*
1521 * Synchronized using the page lock with memory_failure()
1522 */
1523 if (PageHWPoison(page)) {
1524 unlock_page(page);
1525 put_page(page);
Andi Kleenfb46e732010-09-27 23:31:30 +02001526 pr_info("soft offline: %#lx page already poisoned\n", pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001527 return -EBUSY;
1528 }
1529
1530 /*
1531 * Try to invalidate first. This should work for
1532 * non dirty unmapped page cache pages.
1533 */
1534 ret = invalidate_inode_page(page);
1535 unlock_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001536 /*
Andi Kleenfacb6012009-12-16 12:20:00 +01001537 * RED-PEN would be better to keep it isolated here, but we
1538 * would need to fix isolation locking first.
1539 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001540 if (ret == 1) {
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001541 put_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001542 ret = 0;
Andi Kleenfb46e732010-09-27 23:31:30 +02001543 pr_info("soft_offline: %#lx: invalidated\n", pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001544 goto done;
1545 }
1546
1547 /*
1548 * Simple invalidation didn't work.
1549 * Try to migrate to a new page instead. migrate.c
1550 * handles a large number of cases for us.
1551 */
1552 ret = isolate_lru_page(page);
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001553 /*
1554 * Drop page reference which is came from get_any_page()
1555 * successful isolate_lru_page() already took another one.
1556 */
1557 put_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001558 if (!ret) {
1559 LIST_HEAD(pagelist);
Minchan Kim5db8a732011-06-15 15:08:48 -07001560 inc_zone_page_state(page, NR_ISOLATED_ANON +
1561 page_is_file_cache(page));
Andi Kleenfacb6012009-12-16 12:20:00 +01001562 list_add(&page->lru, &pagelist);
Mel Gorman77f1fe62011-01-13 15:45:57 -08001563 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
Mel Gormana6bc32b2012-01-12 17:19:43 -08001564 0, MIGRATE_SYNC);
Andi Kleenfacb6012009-12-16 12:20:00 +01001565 if (ret) {
Andrea Arcangeli57fc4a52011-02-01 15:52:32 -08001566 putback_lru_pages(&pagelist);
Andi Kleenfb46e732010-09-27 23:31:30 +02001567 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001568 pfn, ret, page->flags);
1569 if (ret > 0)
1570 ret = -EIO;
1571 }
1572 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001573 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
Dean Nelsondd73e852011-10-31 17:09:04 -07001574 pfn, ret, page_count(page), page->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001575 }
1576 if (ret)
1577 return ret;
1578
1579done:
1580 atomic_long_add(1, &mce_bad_pages);
1581 SetPageHWPoison(page);
1582 /* keep elevated page count for bad page */
1583 return ret;
1584}