blob: e05ed31c0f61b9bf14a4cc8ce2baa747025fc4ba [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
Xishi Qiu293c07e2013-02-22 16:34:02 -080064atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
Andi Kleen6a460792009-09-16 11:50:15 +020065
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 */
Andrew Mortonc255a452012-07-31 16:43:02 -0700131#ifdef CONFIG_MEMCG_SWAP
Andi Kleen4fd466e2009-12-16 12:19:59 +0100132u64 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 Luck6751ed62012-07-11 10:20:47 -0700348static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -0800349 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) {
Tony Luck6751ed62012-07-11 10:20:47 -0700355 if (forcekill) {
Andi Kleen6a460792009-09-16 11:50:15 +0200356 /*
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;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700403 pgoff_t pgoff;
Andi Kleen6a460792009-09-16 11:50:15 +0200404
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000405 av = page_lock_anon_vma_read(page);
Andi Kleen6a460792009-09-16 11:50:15 +0200406 if (av == NULL) /* Not actually mapped anymore */
Peter Zijlstra9b679322011-06-27 16:18:09 -0700407 return;
408
Michel Lespinassebf181b92012-10-08 16:31:39 -0700409 pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700410 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200411 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800412 struct anon_vma_chain *vmac;
413
Andi Kleen6a460792009-09-16 11:50:15 +0200414 if (!task_early_kill(tsk))
415 continue;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700416 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
417 pgoff, pgoff) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800418 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200419 if (!page_mapped_in_vma(page, vma))
420 continue;
421 if (vma->vm_mm == tsk->mm)
422 add_to_kill(tsk, page, vma, to_kill, tkc);
423 }
424 }
Andi Kleen6a460792009-09-16 11:50:15 +0200425 read_unlock(&tasklist_lock);
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000426 page_unlock_anon_vma_read(av);
Andi Kleen6a460792009-09-16 11:50:15 +0200427}
428
429/*
430 * Collect processes when the error hit a file mapped page.
431 */
432static void collect_procs_file(struct page *page, struct list_head *to_kill,
433 struct to_kill **tkc)
434{
435 struct vm_area_struct *vma;
436 struct task_struct *tsk;
Andi Kleen6a460792009-09-16 11:50:15 +0200437 struct address_space *mapping = page->mapping;
438
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700439 mutex_lock(&mapping->i_mmap_mutex);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700440 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200441 for_each_process(tsk) {
442 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
443
444 if (!task_early_kill(tsk))
445 continue;
446
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -0700447 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
Andi Kleen6a460792009-09-16 11:50:15 +0200448 pgoff) {
449 /*
450 * Send early kill signal to tasks where a vma covers
451 * the page but the corrupted page is not necessarily
452 * mapped it in its pte.
453 * Assume applications who requested early kill want
454 * to be informed of all such data corruptions.
455 */
456 if (vma->vm_mm == tsk->mm)
457 add_to_kill(tsk, page, vma, to_kill, tkc);
458 }
459 }
Andi Kleen6a460792009-09-16 11:50:15 +0200460 read_unlock(&tasklist_lock);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700461 mutex_unlock(&mapping->i_mmap_mutex);
Andi Kleen6a460792009-09-16 11:50:15 +0200462}
463
464/*
465 * Collect the processes who have the corrupted page mapped to kill.
466 * This is done in two steps for locking reasons.
467 * First preallocate one tokill structure outside the spin locks,
468 * so that we can kill at least one process reasonably reliable.
469 */
470static void collect_procs(struct page *page, struct list_head *tokill)
471{
472 struct to_kill *tk;
473
474 if (!page->mapping)
475 return;
476
477 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
478 if (!tk)
479 return;
480 if (PageAnon(page))
481 collect_procs_anon(page, tokill, &tk);
482 else
483 collect_procs_file(page, tokill, &tk);
484 kfree(tk);
485}
486
487/*
488 * Error handlers for various types of pages.
489 */
490
491enum outcome {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100492 IGNORED, /* Error: cannot be handled */
493 FAILED, /* Error: handling failed */
Andi Kleen6a460792009-09-16 11:50:15 +0200494 DELAYED, /* Will be handled later */
Andi Kleen6a460792009-09-16 11:50:15 +0200495 RECOVERED, /* Successfully recovered */
496};
497
498static const char *action_name[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100499 [IGNORED] = "Ignored",
Andi Kleen6a460792009-09-16 11:50:15 +0200500 [FAILED] = "Failed",
501 [DELAYED] = "Delayed",
Andi Kleen6a460792009-09-16 11:50:15 +0200502 [RECOVERED] = "Recovered",
503};
504
505/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100506 * XXX: It is possible that a page is isolated from LRU cache,
507 * and then kept in swap cache or failed to remove from page cache.
508 * The page count will stop it from being freed by unpoison.
509 * Stress tests should be aware of this memory leak problem.
510 */
511static int delete_from_lru_cache(struct page *p)
512{
513 if (!isolate_lru_page(p)) {
514 /*
515 * Clear sensible page flags, so that the buddy system won't
516 * complain when the page is unpoison-and-freed.
517 */
518 ClearPageActive(p);
519 ClearPageUnevictable(p);
520 /*
521 * drop the page count elevated by isolate_lru_page()
522 */
523 page_cache_release(p);
524 return 0;
525 }
526 return -EIO;
527}
528
529/*
Andi Kleen6a460792009-09-16 11:50:15 +0200530 * Error hit kernel page.
531 * Do nothing, try to be lucky and not touch this instead. For a few cases we
532 * could be more sophisticated.
533 */
534static int me_kernel(struct page *p, unsigned long pfn)
535{
Andi Kleen6a460792009-09-16 11:50:15 +0200536 return IGNORED;
537}
538
539/*
540 * Page in unknown state. Do nothing.
541 */
542static int me_unknown(struct page *p, unsigned long pfn)
543{
544 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
545 return FAILED;
546}
547
548/*
Andi Kleen6a460792009-09-16 11:50:15 +0200549 * Clean (or cleaned) page cache page.
550 */
551static int me_pagecache_clean(struct page *p, unsigned long pfn)
552{
553 int err;
554 int ret = FAILED;
555 struct address_space *mapping;
556
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100557 delete_from_lru_cache(p);
558
Andi Kleen6a460792009-09-16 11:50:15 +0200559 /*
560 * For anonymous pages we're done the only reference left
561 * should be the one m_f() holds.
562 */
563 if (PageAnon(p))
564 return RECOVERED;
565
566 /*
567 * Now truncate the page in the page cache. This is really
568 * more like a "temporary hole punch"
569 * Don't do this for block devices when someone else
570 * has a reference, because it could be file system metadata
571 * and that's not safe to truncate.
572 */
573 mapping = page_mapping(p);
574 if (!mapping) {
575 /*
576 * Page has been teared down in the meanwhile
577 */
578 return FAILED;
579 }
580
581 /*
582 * Truncation is a bit tricky. Enable it per file system for now.
583 *
584 * Open: to take i_mutex or not for this? Right now we don't.
585 */
586 if (mapping->a_ops->error_remove_page) {
587 err = mapping->a_ops->error_remove_page(mapping, p);
588 if (err != 0) {
589 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
590 pfn, err);
591 } else if (page_has_private(p) &&
592 !try_to_release_page(p, GFP_NOIO)) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200593 pr_info("MCE %#lx: failed to release buffers\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200594 } else {
595 ret = RECOVERED;
596 }
597 } else {
598 /*
599 * If the file system doesn't support it just invalidate
600 * This fails on dirty or anything with private pages
601 */
602 if (invalidate_inode_page(p))
603 ret = RECOVERED;
604 else
605 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
606 pfn);
607 }
608 return ret;
609}
610
611/*
612 * Dirty cache page page
613 * Issues: when the error hit a hole page the error is not properly
614 * propagated.
615 */
616static int me_pagecache_dirty(struct page *p, unsigned long pfn)
617{
618 struct address_space *mapping = page_mapping(p);
619
620 SetPageError(p);
621 /* TBD: print more information about the file. */
622 if (mapping) {
623 /*
624 * IO error will be reported by write(), fsync(), etc.
625 * who check the mapping.
626 * This way the application knows that something went
627 * wrong with its dirty file data.
628 *
629 * There's one open issue:
630 *
631 * The EIO will be only reported on the next IO
632 * operation and then cleared through the IO map.
633 * Normally Linux has two mechanisms to pass IO error
634 * first through the AS_EIO flag in the address space
635 * and then through the PageError flag in the page.
636 * Since we drop pages on memory failure handling the
637 * only mechanism open to use is through AS_AIO.
638 *
639 * This has the disadvantage that it gets cleared on
640 * the first operation that returns an error, while
641 * the PageError bit is more sticky and only cleared
642 * when the page is reread or dropped. If an
643 * application assumes it will always get error on
644 * fsync, but does other operations on the fd before
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300645 * and the page is dropped between then the error
Andi Kleen6a460792009-09-16 11:50:15 +0200646 * will not be properly reported.
647 *
648 * This can already happen even without hwpoisoned
649 * pages: first on metadata IO errors (which only
650 * report through AS_EIO) or when the page is dropped
651 * at the wrong time.
652 *
653 * So right now we assume that the application DTRT on
654 * the first EIO, but we're not worse than other parts
655 * of the kernel.
656 */
657 mapping_set_error(mapping, EIO);
658 }
659
660 return me_pagecache_clean(p, pfn);
661}
662
663/*
664 * Clean and dirty swap cache.
665 *
666 * Dirty swap cache page is tricky to handle. The page could live both in page
667 * cache and swap cache(ie. page is freshly swapped in). So it could be
668 * referenced concurrently by 2 types of PTEs:
669 * normal PTEs and swap PTEs. We try to handle them consistently by calling
670 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
671 * and then
672 * - clear dirty bit to prevent IO
673 * - remove from LRU
674 * - but keep in the swap cache, so that when we return to it on
675 * a later page fault, we know the application is accessing
676 * corrupted data and shall be killed (we installed simple
677 * interception code in do_swap_page to catch it).
678 *
679 * Clean swap cache pages can be directly isolated. A later page fault will
680 * bring in the known good data from disk.
681 */
682static int me_swapcache_dirty(struct page *p, unsigned long pfn)
683{
Andi Kleen6a460792009-09-16 11:50:15 +0200684 ClearPageDirty(p);
685 /* Trigger EIO in shmem: */
686 ClearPageUptodate(p);
687
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100688 if (!delete_from_lru_cache(p))
689 return DELAYED;
690 else
691 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200692}
693
694static int me_swapcache_clean(struct page *p, unsigned long pfn)
695{
Andi Kleen6a460792009-09-16 11:50:15 +0200696 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800697
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100698 if (!delete_from_lru_cache(p))
699 return RECOVERED;
700 else
701 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200702}
703
704/*
705 * Huge pages. Needs work.
706 * Issues:
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900707 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
708 * To narrow down kill region to one page, we need to break up pmd.
Andi Kleen6a460792009-09-16 11:50:15 +0200709 */
710static int me_huge_page(struct page *p, unsigned long pfn)
711{
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900712 int res = 0;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900713 struct page *hpage = compound_head(p);
714 /*
715 * We can safely recover from error on free or reserved (i.e.
716 * not in-use) hugepage by dequeuing it from freelist.
717 * To check whether a hugepage is in-use or not, we can't use
718 * page->lru because it can be used in other hugepage operations,
719 * such as __unmap_hugepage_range() and gather_surplus_pages().
720 * So instead we use page_mapping() and PageAnon().
721 * We assume that this function is called with page lock held,
722 * so there is no race between isolation and mapping/unmapping.
723 */
724 if (!(page_mapping(hpage) || PageAnon(hpage))) {
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900725 res = dequeue_hwpoisoned_huge_page(hpage);
726 if (!res)
727 return RECOVERED;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900728 }
729 return DELAYED;
Andi Kleen6a460792009-09-16 11:50:15 +0200730}
731
732/*
733 * Various page states we can handle.
734 *
735 * A page state is defined by its current page->flags bits.
736 * The table matches them in order and calls the right handler.
737 *
738 * This is quite tricky because we can access page at any time
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300739 * in its live cycle, so all accesses have to be extremely careful.
Andi Kleen6a460792009-09-16 11:50:15 +0200740 *
741 * This is not complete. More states could be added.
742 * For any missing state don't attempt recovery.
743 */
744
745#define dirty (1UL << PG_dirty)
746#define sc (1UL << PG_swapcache)
747#define unevict (1UL << PG_unevictable)
748#define mlock (1UL << PG_mlocked)
749#define writeback (1UL << PG_writeback)
750#define lru (1UL << PG_lru)
751#define swapbacked (1UL << PG_swapbacked)
752#define head (1UL << PG_head)
753#define tail (1UL << PG_tail)
754#define compound (1UL << PG_compound)
755#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200756#define reserved (1UL << PG_reserved)
757
758static struct page_state {
759 unsigned long mask;
760 unsigned long res;
761 char *msg;
762 int (*action)(struct page *p, unsigned long pfn);
763} error_states[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100764 { reserved, reserved, "reserved kernel", me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100765 /*
766 * free pages are specially detected outside this table:
767 * PG_buddy pages only make a small fraction of all free pages.
768 */
Andi Kleen6a460792009-09-16 11:50:15 +0200769
770 /*
771 * Could in theory check if slab page is free or if we can drop
772 * currently unused objects without touching them. But just
773 * treat it as standard kernel for now.
774 */
775 { slab, slab, "kernel slab", me_kernel },
776
777#ifdef CONFIG_PAGEFLAGS_EXTENDED
778 { head, head, "huge", me_huge_page },
779 { tail, tail, "huge", me_huge_page },
780#else
781 { compound, compound, "huge", me_huge_page },
782#endif
783
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800784 { sc|dirty, sc|dirty, "dirty swapcache", me_swapcache_dirty },
785 { sc|dirty, sc, "clean swapcache", me_swapcache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200786
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800787 { mlock|dirty, mlock|dirty, "dirty mlocked LRU", me_pagecache_dirty },
Naoya Horiguchie3986292013-04-29 15:06:08 -0700788 { mlock|dirty, mlock, "clean mlocked LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200789
Naoya Horiguchi5f4b9fc2013-02-22 16:35:53 -0800790 { unevict|dirty, unevict|dirty, "dirty unevictable LRU", me_pagecache_dirty },
Naoya Horiguchie3986292013-04-29 15:06:08 -0700791 { unevict|dirty, unevict, "clean unevictable LRU", me_pagecache_clean },
Naoya Horiguchi5f4b9fc2013-02-22 16:35:53 -0800792
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800793 { lru|dirty, lru|dirty, "dirty LRU", me_pagecache_dirty },
Andi Kleen6a460792009-09-16 11:50:15 +0200794 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200795
796 /*
797 * Catchall entry: must be at end.
798 */
799 { 0, 0, "unknown page state", me_unknown },
800};
801
Andi Kleen2326c462009-12-16 12:20:00 +0100802#undef dirty
803#undef sc
804#undef unevict
805#undef mlock
806#undef writeback
807#undef lru
808#undef swapbacked
809#undef head
810#undef tail
811#undef compound
812#undef slab
813#undef reserved
814
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800815/*
816 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
817 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
818 */
Andi Kleen6a460792009-09-16 11:50:15 +0200819static void action_result(unsigned long pfn, char *msg, int result)
820{
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800821 pr_err("MCE %#lx: %s page recovery: %s\n",
822 pfn, msg, action_name[result]);
Andi Kleen6a460792009-09-16 11:50:15 +0200823}
824
825static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100826 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200827{
828 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200829 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200830
831 result = ps->action(p, pfn);
832 action_result(pfn, ps->msg, result);
Wu Fengguang7456b042009-10-19 08:15:01 +0200833
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100834 count = page_count(p) - 1;
Wu Fengguang138ce282009-12-16 12:19:58 +0100835 if (ps->action == me_swapcache_dirty && result == DELAYED)
836 count--;
837 if (count != 0) {
Andi Kleen6a460792009-09-16 11:50:15 +0200838 printk(KERN_ERR
839 "MCE %#lx: %s page still referenced by %d users\n",
Wu Fengguang7456b042009-10-19 08:15:01 +0200840 pfn, ps->msg, count);
Wu Fengguang138ce282009-12-16 12:19:58 +0100841 result = FAILED;
842 }
Andi Kleen6a460792009-09-16 11:50:15 +0200843
844 /* Could do more checks here if page looks ok */
845 /*
846 * Could adjust zone counters here to correct for the missing page.
847 */
848
Wu Fengguang138ce282009-12-16 12:19:58 +0100849 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200850}
851
Andi Kleen6a460792009-09-16 11:50:15 +0200852/*
853 * Do all that is necessary to remove user space mappings. Unmap
854 * the pages and send SIGBUS to the processes if the data was dirty.
855 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100856static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Tony Luck7329bbe2011-12-13 09:27:58 -0800857 int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200858{
859 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
860 struct address_space *mapping;
861 LIST_HEAD(tokill);
862 int ret;
Tony Luck6751ed62012-07-11 10:20:47 -0700863 int kill = 1, forcekill;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900864 struct page *hpage = compound_head(p);
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800865 struct page *ppage;
Andi Kleen6a460792009-09-16 11:50:15 +0200866
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100867 if (PageReserved(p) || PageSlab(p))
868 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200869
Andi Kleen6a460792009-09-16 11:50:15 +0200870 /*
871 * This check implies we don't kill processes if their pages
872 * are in the swap cache early. Those are always late kills.
873 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900874 if (!page_mapped(hpage))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100875 return SWAP_SUCCESS;
876
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900877 if (PageKsm(p))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100878 return SWAP_FAIL;
Andi Kleen6a460792009-09-16 11:50:15 +0200879
880 if (PageSwapCache(p)) {
881 printk(KERN_ERR
882 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
883 ttu |= TTU_IGNORE_HWPOISON;
884 }
885
886 /*
887 * Propagate the dirty bit from PTEs to struct page first, because we
888 * need this to decide if we should kill or just drop the page.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100889 * XXX: the dirty test could be racy: set_page_dirty() may not always
890 * be called inside page lock (it's recommended but not enforced).
Andi Kleen6a460792009-09-16 11:50:15 +0200891 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900892 mapping = page_mapping(hpage);
Tony Luck6751ed62012-07-11 10:20:47 -0700893 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900894 mapping_cap_writeback_dirty(mapping)) {
895 if (page_mkclean(hpage)) {
896 SetPageDirty(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +0200897 } else {
898 kill = 0;
899 ttu |= TTU_IGNORE_HWPOISON;
900 printk(KERN_INFO
901 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
902 pfn);
903 }
904 }
905
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800906 /*
907 * ppage: poisoned page
908 * if p is regular page(4k page)
909 * ppage == real poisoned page;
910 * else p is hugetlb or THP, ppage == head page.
911 */
912 ppage = hpage;
913
Jin Dongmingefeda7a2011-02-01 15:52:39 -0800914 if (PageTransHuge(hpage)) {
915 /*
916 * Verify that this isn't a hugetlbfs head page, the check for
917 * PageAnon is just for avoid tripping a split_huge_page
918 * internal debug check, as split_huge_page refuses to deal with
919 * anything that isn't an anon page. PageAnon can't go away fro
920 * under us because we hold a refcount on the hpage, without a
921 * refcount on the hpage. split_huge_page can't be safely called
922 * in the first place, having a refcount on the tail isn't
923 * enough * to be safe.
924 */
925 if (!PageHuge(hpage) && PageAnon(hpage)) {
926 if (unlikely(split_huge_page(hpage))) {
927 /*
928 * FIXME: if splitting THP is failed, it is
929 * better to stop the following operation rather
930 * than causing panic by unmapping. System might
931 * survive if the page is freed later.
932 */
933 printk(KERN_INFO
934 "MCE %#lx: failed to split THP\n", pfn);
935
936 BUG_ON(!PageHWPoison(p));
937 return SWAP_FAIL;
938 }
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800939 /* THP is split, so ppage should be the real poisoned page. */
940 ppage = p;
Jin Dongmingefeda7a2011-02-01 15:52:39 -0800941 }
942 }
943
Andi Kleen6a460792009-09-16 11:50:15 +0200944 /*
945 * First collect all the processes that have the page
946 * mapped in dirty form. This has to be done before try_to_unmap,
947 * because ttu takes the rmap data structures down.
948 *
949 * Error handling: We ignore errors here because
950 * there's nothing that can be done.
951 */
952 if (kill)
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800953 collect_procs(ppage, &tokill);
Andi Kleen6a460792009-09-16 11:50:15 +0200954
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800955 if (hpage != ppage)
Jens Axboe7eaceac2011-03-10 08:52:07 +0100956 lock_page(ppage);
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800957
958 ret = try_to_unmap(ppage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +0200959 if (ret != SWAP_SUCCESS)
960 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800961 pfn, page_mapcount(ppage));
962
963 if (hpage != ppage)
964 unlock_page(ppage);
Andi Kleen6a460792009-09-16 11:50:15 +0200965
966 /*
967 * Now that the dirty bit has been propagated to the
968 * struct page and all unmaps done we can decide if
969 * killing is needed or not. Only kill when the page
Tony Luck6751ed62012-07-11 10:20:47 -0700970 * was dirty or the process is not restartable,
971 * otherwise the tokill list is merely
Andi Kleen6a460792009-09-16 11:50:15 +0200972 * freed. When there was a problem unmapping earlier
973 * use a more force-full uncatchable kill to prevent
974 * any accesses to the poisoned memory.
975 */
Tony Luck6751ed62012-07-11 10:20:47 -0700976 forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL);
977 kill_procs(&tokill, forcekill, trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -0800978 ret != SWAP_SUCCESS, p, pfn, flags);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100979
980 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200981}
982
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900983static void set_page_hwpoison_huge_page(struct page *hpage)
984{
985 int i;
Andrea Arcangeli37c2ac72011-01-13 15:47:16 -0800986 int nr_pages = 1 << compound_trans_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900987 for (i = 0; i < nr_pages; i++)
988 SetPageHWPoison(hpage + i);
989}
990
991static void clear_page_hwpoison_huge_page(struct page *hpage)
992{
993 int i;
Andrea Arcangeli37c2ac72011-01-13 15:47:16 -0800994 int nr_pages = 1 << compound_trans_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900995 for (i = 0; i < nr_pages; i++)
996 ClearPageHWPoison(hpage + i);
997}
998
Tony Luckcd42f4a2011-12-15 10:48:12 -0800999/**
1000 * memory_failure - Handle memory failure of a page.
1001 * @pfn: Page Number of the corrupted page
1002 * @trapno: Trap number reported in the signal to user space.
1003 * @flags: fine tune action taken
1004 *
1005 * This function is called by the low level machine check code
1006 * of an architecture when it detects hardware memory corruption
1007 * of a page. It tries its best to recover, which includes
1008 * dropping pages, killing processes etc.
1009 *
1010 * The function is primarily of use for corruptions that
1011 * happen outside the current execution context (e.g. when
1012 * detected by a background scrubber)
1013 *
1014 * Must run in process context (e.g. a work queue) with interrupts
1015 * enabled and no spinlocks hold.
1016 */
1017int memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +02001018{
1019 struct page_state *ps;
1020 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001021 struct page *hpage;
Andi Kleen6a460792009-09-16 11:50:15 +02001022 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001023 unsigned int nr_pages;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001024 unsigned long page_flags;
Andi Kleen6a460792009-09-16 11:50:15 +02001025
1026 if (!sysctl_memory_failure_recovery)
1027 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1028
1029 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +01001030 printk(KERN_ERR
1031 "MCE %#lx: memory outside kernel control\n",
1032 pfn);
1033 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +02001034 }
1035
1036 p = pfn_to_page(pfn);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001037 hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +02001038 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001039 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001040 return 0;
1041 }
1042
Naoya Horiguchi4db0e952013-02-22 16:34:05 -08001043 /*
1044 * Currently errors on hugetlbfs pages are measured in hugepage units,
1045 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1046 * transparent hugepages, they are supposed to be split and error
1047 * measurement is done in normal page units. So nr_pages should be one
1048 * in this case.
1049 */
1050 if (PageHuge(p))
1051 nr_pages = 1 << compound_order(hpage);
1052 else /* normal page or thp */
1053 nr_pages = 1;
Xishi Qiu293c07e2013-02-22 16:34:02 -08001054 atomic_long_add(nr_pages, &num_poisoned_pages);
Andi Kleen6a460792009-09-16 11:50:15 +02001055
1056 /*
1057 * We need/can do nothing about count=0 pages.
1058 * 1) it's a free page, and therefore in safe hand:
1059 * prep_new_page() will be the gate keeper.
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001060 * 2) it's a free hugepage, which is also safe:
1061 * an affected hugepage will be dequeued from hugepage freelist,
1062 * so there's no concern about reusing it ever after.
1063 * 3) it's part of a non-compound high order page.
Andi Kleen6a460792009-09-16 11:50:15 +02001064 * Implies some kernel user: cannot stop them from
1065 * R/W the page; let's pray that the page has been
1066 * used and will be freed some time later.
1067 * In fact it's dangerous to directly bump up page count from 0,
1068 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1069 */
Andi Kleen82ba0112009-12-16 12:19:57 +01001070 if (!(flags & MF_COUNT_INCREASED) &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001071 !get_page_unless_zero(hpage)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001072 if (is_free_buddy_page(p)) {
1073 action_result(pfn, "free buddy", DELAYED);
1074 return 0;
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001075 } else if (PageHuge(hpage)) {
1076 /*
1077 * Check "just unpoisoned", "filter hit", and
1078 * "race with other subpage."
1079 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001080 lock_page(hpage);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001081 if (!PageHWPoison(hpage)
1082 || (hwpoison_filter(p) && TestClearPageHWPoison(p))
1083 || (p != hpage && TestSetPageHWPoison(hpage))) {
Xishi Qiu293c07e2013-02-22 16:34:02 -08001084 atomic_long_sub(nr_pages, &num_poisoned_pages);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001085 return 0;
1086 }
1087 set_page_hwpoison_huge_page(hpage);
1088 res = dequeue_hwpoisoned_huge_page(hpage);
1089 action_result(pfn, "free huge",
1090 res ? IGNORED : DELAYED);
1091 unlock_page(hpage);
1092 return res;
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001093 } else {
1094 action_result(pfn, "high order kernel", IGNORED);
1095 return -EBUSY;
1096 }
Andi Kleen6a460792009-09-16 11:50:15 +02001097 }
1098
1099 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001100 * We ignore non-LRU pages for good reasons.
1101 * - PG_locked is only well defined for LRU pages and a few others
1102 * - to avoid races with __set_page_locked()
1103 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1104 * The check (unnecessarily) ignores LRU pages being isolated and
1105 * walked by the page reclaim code, however that's not a big loss.
1106 */
Dean Nelson385de352012-03-21 16:34:05 -07001107 if (!PageHuge(p) && !PageTransTail(p)) {
Jin Dongmingaf241a02011-02-01 15:52:41 -08001108 if (!PageLRU(p))
1109 shake_page(p, 0);
1110 if (!PageLRU(p)) {
1111 /*
1112 * shake_page could have turned it free.
1113 */
1114 if (is_free_buddy_page(p)) {
1115 action_result(pfn, "free buddy, 2nd try",
1116 DELAYED);
1117 return 0;
1118 }
1119 action_result(pfn, "non LRU", IGNORED);
1120 put_page(p);
1121 return -EBUSY;
Andi Kleen0474a602009-12-16 12:20:00 +01001122 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001123 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001124
1125 /*
Andi Kleen6a460792009-09-16 11:50:15 +02001126 * Lock the page and wait for writeback to finish.
1127 * It's very difficult to mess with pages currently under IO
1128 * and in many cases impossible, so we just avoid it here.
1129 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001130 lock_page(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001131
1132 /*
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001133 * We use page flags to determine what action should be taken, but
1134 * the flags can be modified by the error containment action. One
1135 * example is an mlocked page, where PG_mlocked is cleared by
1136 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1137 * correctly, we save a copy of the page flags at this time.
1138 */
1139 page_flags = p->flags;
1140
1141 /*
Wu Fengguang847ce402009-12-16 12:19:58 +01001142 * unpoison always clear PG_hwpoison inside page lock
1143 */
1144 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001145 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001146 res = 0;
1147 goto out;
1148 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001149 if (hwpoison_filter(p)) {
1150 if (TestClearPageHWPoison(p))
Xishi Qiu293c07e2013-02-22 16:34:02 -08001151 atomic_long_sub(nr_pages, &num_poisoned_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001152 unlock_page(hpage);
1153 put_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001154 return 0;
1155 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001156
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001157 /*
1158 * For error on the tail page, we should set PG_hwpoison
1159 * on the head page to show that the hugepage is hwpoisoned
1160 */
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001161 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001162 action_result(pfn, "hugepage already hardware poisoned",
1163 IGNORED);
1164 unlock_page(hpage);
1165 put_page(hpage);
1166 return 0;
1167 }
1168 /*
1169 * Set PG_hwpoison on all pages in an error hugepage,
1170 * because containment is done in hugepage unit for now.
1171 * Since we have done TestSetPageHWPoison() for the head page with
1172 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1173 */
1174 if (PageHuge(p))
1175 set_page_hwpoison_huge_page(hpage);
1176
Andi Kleen6a460792009-09-16 11:50:15 +02001177 wait_on_page_writeback(p);
1178
1179 /*
1180 * Now take care of user space mappings.
Minchan Kime64a7822011-03-22 16:32:44 -07001181 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
Andi Kleen6a460792009-09-16 11:50:15 +02001182 */
Tony Luck7329bbe2011-12-13 09:27:58 -08001183 if (hwpoison_user_mappings(p, pfn, trapno, flags) != SWAP_SUCCESS) {
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001184 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1185 res = -EBUSY;
1186 goto out;
1187 }
Andi Kleen6a460792009-09-16 11:50:15 +02001188
1189 /*
1190 * Torn down by someone else?
1191 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001192 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Andi Kleen6a460792009-09-16 11:50:15 +02001193 action_result(pfn, "already truncated LRU", IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001194 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001195 goto out;
1196 }
1197
1198 res = -EBUSY;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001199 /*
1200 * The first check uses the current page flags which may not have any
1201 * relevant information. The second check with the saved page flagss is
1202 * carried out only if the first check can't determine the page status.
1203 */
1204 for (ps = error_states;; ps++)
1205 if ((p->flags & ps->mask) == ps->res)
Andi Kleen6a460792009-09-16 11:50:15 +02001206 break;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001207 if (!ps->mask)
1208 for (ps = error_states;; ps++)
1209 if ((page_flags & ps->mask) == ps->res)
1210 break;
1211 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001212out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001213 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001214 return res;
1215}
Tony Luckcd42f4a2011-12-15 10:48:12 -08001216EXPORT_SYMBOL_GPL(memory_failure);
Wu Fengguang847ce402009-12-16 12:19:58 +01001217
Huang Yingea8f5fb2011-07-13 13:14:27 +08001218#define MEMORY_FAILURE_FIFO_ORDER 4
1219#define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1220
1221struct memory_failure_entry {
1222 unsigned long pfn;
1223 int trapno;
1224 int flags;
1225};
1226
1227struct memory_failure_cpu {
1228 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1229 MEMORY_FAILURE_FIFO_SIZE);
1230 spinlock_t lock;
1231 struct work_struct work;
1232};
1233
1234static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1235
1236/**
1237 * memory_failure_queue - Schedule handling memory failure of a page.
1238 * @pfn: Page Number of the corrupted page
1239 * @trapno: Trap number reported in the signal to user space.
1240 * @flags: Flags for memory failure handling
1241 *
1242 * This function is called by the low level hardware error handler
1243 * when it detects hardware memory corruption of a page. It schedules
1244 * the recovering of error page, including dropping pages, killing
1245 * processes etc.
1246 *
1247 * The function is primarily of use for corruptions that
1248 * happen outside the current execution context (e.g. when
1249 * detected by a background scrubber)
1250 *
1251 * Can run in IRQ context.
1252 */
1253void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1254{
1255 struct memory_failure_cpu *mf_cpu;
1256 unsigned long proc_flags;
1257 struct memory_failure_entry entry = {
1258 .pfn = pfn,
1259 .trapno = trapno,
1260 .flags = flags,
1261 };
1262
1263 mf_cpu = &get_cpu_var(memory_failure_cpu);
1264 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1265 if (kfifo_put(&mf_cpu->fifo, &entry))
1266 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1267 else
Joe Perches8e33a522013-07-25 11:53:25 -07001268 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
Huang Yingea8f5fb2011-07-13 13:14:27 +08001269 pfn);
1270 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1271 put_cpu_var(memory_failure_cpu);
1272}
1273EXPORT_SYMBOL_GPL(memory_failure_queue);
1274
1275static void memory_failure_work_func(struct work_struct *work)
1276{
1277 struct memory_failure_cpu *mf_cpu;
1278 struct memory_failure_entry entry = { 0, };
1279 unsigned long proc_flags;
1280 int gotten;
1281
1282 mf_cpu = &__get_cpu_var(memory_failure_cpu);
1283 for (;;) {
1284 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1285 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1286 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1287 if (!gotten)
1288 break;
Naveen N. Raocf870c72013-07-10 14:57:01 +05301289 if (entry.flags & MF_SOFT_OFFLINE)
1290 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1291 else
1292 memory_failure(entry.pfn, entry.trapno, entry.flags);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001293 }
1294}
1295
1296static int __init memory_failure_init(void)
1297{
1298 struct memory_failure_cpu *mf_cpu;
1299 int cpu;
1300
1301 for_each_possible_cpu(cpu) {
1302 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1303 spin_lock_init(&mf_cpu->lock);
1304 INIT_KFIFO(mf_cpu->fifo);
1305 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1306 }
1307
1308 return 0;
1309}
1310core_initcall(memory_failure_init);
1311
Wu Fengguang847ce402009-12-16 12:19:58 +01001312/**
1313 * unpoison_memory - Unpoison a previously poisoned page
1314 * @pfn: Page number of the to be unpoisoned page
1315 *
1316 * Software-unpoison a page that has been poisoned by
1317 * memory_failure() earlier.
1318 *
1319 * This is only done on the software-level, so it only works
1320 * for linux injected failures, not real hardware failures
1321 *
1322 * Returns 0 for success, otherwise -errno.
1323 */
1324int unpoison_memory(unsigned long pfn)
1325{
1326 struct page *page;
1327 struct page *p;
1328 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001329 unsigned int nr_pages;
Wu Fengguang847ce402009-12-16 12:19:58 +01001330
1331 if (!pfn_valid(pfn))
1332 return -ENXIO;
1333
1334 p = pfn_to_page(pfn);
1335 page = compound_head(p);
1336
1337 if (!PageHWPoison(p)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001338 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001339 return 0;
1340 }
1341
Andrea Arcangeli37c2ac72011-01-13 15:47:16 -08001342 nr_pages = 1 << compound_trans_order(page);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001343
Wu Fengguang847ce402009-12-16 12:19:58 +01001344 if (!get_page_unless_zero(page)) {
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001345 /*
1346 * Since HWPoisoned hugepage should have non-zero refcount,
1347 * race between memory failure and unpoison seems to happen.
1348 * In such case unpoison fails and memory failure runs
1349 * to the end.
1350 */
1351 if (PageHuge(page)) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001352 pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001353 return 0;
1354 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001355 if (TestClearPageHWPoison(p))
Xishi Qiu293c07e2013-02-22 16:34:02 -08001356 atomic_long_sub(nr_pages, &num_poisoned_pages);
Andi Kleenfb46e732010-09-27 23:31:30 +02001357 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001358 return 0;
1359 }
1360
Jens Axboe7eaceac2011-03-10 08:52:07 +01001361 lock_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001362 /*
1363 * This test is racy because PG_hwpoison is set outside of page lock.
1364 * That's acceptable because that won't trigger kernel panic. Instead,
1365 * the PG_hwpoison page will be caught and isolated on the entrance to
1366 * the free buddy page pool.
1367 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001368 if (TestClearPageHWPoison(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001369 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
Xishi Qiu293c07e2013-02-22 16:34:02 -08001370 atomic_long_sub(nr_pages, &num_poisoned_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001371 freeit = 1;
Naoya Horiguchi6a901812010-09-08 10:19:40 +09001372 if (PageHuge(page))
1373 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001374 }
1375 unlock_page(page);
1376
1377 put_page(page);
1378 if (freeit)
1379 put_page(page);
1380
1381 return 0;
1382}
1383EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001384
1385static struct page *new_page(struct page *p, unsigned long private, int **x)
1386{
Andi Kleen12686d12009-12-16 12:20:01 +01001387 int nid = page_to_nid(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001388 if (PageHuge(p))
1389 return alloc_huge_page_node(page_hstate(compound_head(p)),
1390 nid);
1391 else
1392 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001393}
1394
1395/*
1396 * Safely get reference count of an arbitrary page.
1397 * Returns 0 for a free page, -EIO for a zero refcount page
1398 * that is not free, and 1 for any other page type.
1399 * For 1 the page is returned with increased page count, otherwise not.
1400 */
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001401static int __get_any_page(struct page *p, unsigned long pfn, int flags)
Andi Kleenfacb6012009-12-16 12:20:00 +01001402{
1403 int ret;
1404
1405 if (flags & MF_COUNT_INCREASED)
1406 return 1;
1407
1408 /*
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -08001409 * The lock_memory_hotplug prevents a race with memory hotplug.
Andi Kleenfacb6012009-12-16 12:20:00 +01001410 * This is a big hammer, a better would be nicer.
1411 */
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -08001412 lock_memory_hotplug();
Andi Kleenfacb6012009-12-16 12:20:00 +01001413
1414 /*
1415 * Isolate the page, so that it doesn't get reallocated if it
Naoya Horiguchif15bdfa2013-07-03 15:02:37 -07001416 * was free. This flag should be kept set until the source page
1417 * is freed and PG_hwpoison on it is set.
Andi Kleenfacb6012009-12-16 12:20:00 +01001418 */
Wen Congyangb023f462012-12-11 16:00:45 -08001419 set_migratetype_isolate(p, true);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001420 /*
1421 * When the target page is a free hugepage, just remove it
1422 * from free hugepage list.
1423 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001424 if (!get_page_unless_zero(compound_head(p))) {
Naoya Horiguchid950b952010-09-08 10:19:39 +09001425 if (PageHuge(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001426 pr_info("%s: %#lx free huge page\n", __func__, pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001427 ret = 0;
Naoya Horiguchid950b952010-09-08 10:19:39 +09001428 } else if (is_free_buddy_page(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001429 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001430 ret = 0;
1431 } else {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001432 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1433 __func__, pfn, p->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001434 ret = -EIO;
1435 }
1436 } else {
1437 /* Not a free page */
1438 ret = 1;
1439 }
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -08001440 unlock_memory_hotplug();
Andi Kleenfacb6012009-12-16 12:20:00 +01001441 return ret;
1442}
1443
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001444static int get_any_page(struct page *page, unsigned long pfn, int flags)
1445{
1446 int ret = __get_any_page(page, pfn, flags);
1447
1448 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1449 /*
1450 * Try to free it.
1451 */
1452 put_page(page);
1453 shake_page(page, 1);
1454
1455 /*
1456 * Did it turn free?
1457 */
1458 ret = __get_any_page(page, pfn, 0);
1459 if (!PageLRU(page)) {
1460 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1461 pfn, page->flags);
1462 return -EIO;
1463 }
1464 }
1465 return ret;
1466}
1467
Naoya Horiguchid950b952010-09-08 10:19:39 +09001468static int soft_offline_huge_page(struct page *page, int flags)
1469{
1470 int ret;
1471 unsigned long pfn = page_to_pfn(page);
1472 struct page *hpage = compound_head(page);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001473 LIST_HEAD(pagelist);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001474
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001475 /*
1476 * This double-check of PageHWPoison is to avoid the race with
1477 * memory_failure(). See also comment in __soft_offline_page().
1478 */
1479 lock_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001480 if (PageHWPoison(hpage)) {
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001481 unlock_page(hpage);
1482 put_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001483 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001484 return -EBUSY;
Xishi Qiu0ebff322013-02-22 16:33:59 -08001485 }
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001486 unlock_page(hpage);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001487
Naoya Horiguchid950b952010-09-08 10:19:39 +09001488 /* Keep page count to indicate a given hugepage is isolated. */
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001489 list_move(&hpage->lru, &pagelist);
1490 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1491 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001492 if (ret) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001493 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1494 pfn, ret, page->flags);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001495 /*
1496 * We know that soft_offline_huge_page() tries to migrate
1497 * only one hugepage pointed to by hpage, so we need not
1498 * run through the pagelist here.
1499 */
1500 putback_active_hugepage(hpage);
1501 if (ret > 0)
1502 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001503 } else {
1504 set_page_hwpoison_huge_page(hpage);
1505 dequeue_hwpoisoned_huge_page(hpage);
1506 atomic_long_add(1 << compound_trans_order(hpage),
1507 &num_poisoned_pages);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001508 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001509 return ret;
1510}
1511
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001512static int __soft_offline_page(struct page *page, int flags);
1513
Andi Kleenfacb6012009-12-16 12:20:00 +01001514/**
1515 * soft_offline_page - Soft offline a page.
1516 * @page: page to offline
1517 * @flags: flags. Same as memory_failure().
1518 *
1519 * Returns 0 on success, otherwise negated errno.
1520 *
1521 * Soft offline a page, by migration or invalidation,
1522 * without killing anything. This is for the case when
1523 * a page is not corrupted yet (so it's still valid to access),
1524 * but has had a number of corrected errors and is better taken
1525 * out.
1526 *
1527 * The actual policy on when to do that is maintained by
1528 * user space.
1529 *
1530 * This should never impact any application or cause data loss,
1531 * however it might take some time.
1532 *
1533 * This is not a 100% solution for all memory, but tries to be
1534 * ``good enough'' for the majority of memory.
1535 */
1536int soft_offline_page(struct page *page, int flags)
1537{
1538 int ret;
1539 unsigned long pfn = page_to_pfn(page);
Naoya Horiguchi783657a2012-11-29 13:54:34 -08001540 struct page *hpage = compound_trans_head(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001541
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001542 if (PageHWPoison(page)) {
1543 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1544 return -EBUSY;
Xishi Qiufa8dd8a2013-02-22 16:34:00 -08001545 }
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001546 if (!PageHuge(page) && PageTransHuge(hpage)) {
Naoya Horiguchi783657a2012-11-29 13:54:34 -08001547 if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1548 pr_info("soft offline: %#lx: failed to split THP\n",
1549 pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001550 return -EBUSY;
Naoya Horiguchi783657a2012-11-29 13:54:34 -08001551 }
1552 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001553
Andi Kleenfacb6012009-12-16 12:20:00 +01001554 ret = get_any_page(page, pfn, flags);
1555 if (ret < 0)
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001556 return ret;
1557 if (ret) { /* for in-use pages */
1558 if (PageHuge(page))
1559 ret = soft_offline_huge_page(page, flags);
1560 else
1561 ret = __soft_offline_page(page, flags);
1562 } else { /* for free pages */
1563 if (PageHuge(page)) {
1564 set_page_hwpoison_huge_page(hpage);
1565 dequeue_hwpoisoned_huge_page(hpage);
1566 atomic_long_add(1 << compound_trans_order(hpage),
1567 &num_poisoned_pages);
1568 } else {
1569 SetPageHWPoison(page);
1570 atomic_long_inc(&num_poisoned_pages);
1571 }
1572 }
Naoya Horiguchif15bdfa2013-07-03 15:02:37 -07001573 unset_migratetype_isolate(page, MIGRATE_MOVABLE);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001574 return ret;
1575}
1576
1577static int __soft_offline_page(struct page *page, int flags)
1578{
1579 int ret;
1580 unsigned long pfn = page_to_pfn(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001581
1582 /*
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001583 * Check PageHWPoison again inside page lock because PageHWPoison
1584 * is set by memory_failure() outside page lock. Note that
1585 * memory_failure() also double-checks PageHWPoison inside page lock,
1586 * so there's no race between soft_offline_page() and memory_failure().
Andi Kleenfacb6012009-12-16 12:20:00 +01001587 */
Xishi Qiu0ebff322013-02-22 16:33:59 -08001588 lock_page(page);
1589 wait_on_page_writeback(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001590 if (PageHWPoison(page)) {
1591 unlock_page(page);
1592 put_page(page);
1593 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1594 return -EBUSY;
1595 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001596 /*
1597 * Try to invalidate first. This should work for
1598 * non dirty unmapped page cache pages.
1599 */
1600 ret = invalidate_inode_page(page);
1601 unlock_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001602 /*
Andi Kleenfacb6012009-12-16 12:20:00 +01001603 * RED-PEN would be better to keep it isolated here, but we
1604 * would need to fix isolation locking first.
1605 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001606 if (ret == 1) {
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001607 put_page(page);
Andi Kleenfb46e732010-09-27 23:31:30 +02001608 pr_info("soft_offline: %#lx: invalidated\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001609 SetPageHWPoison(page);
1610 atomic_long_inc(&num_poisoned_pages);
1611 return 0;
Andi Kleenfacb6012009-12-16 12:20:00 +01001612 }
1613
1614 /*
1615 * Simple invalidation didn't work.
1616 * Try to migrate to a new page instead. migrate.c
1617 * handles a large number of cases for us.
1618 */
1619 ret = isolate_lru_page(page);
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001620 /*
1621 * Drop page reference which is came from get_any_page()
1622 * successful isolate_lru_page() already took another one.
1623 */
1624 put_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001625 if (!ret) {
1626 LIST_HEAD(pagelist);
Minchan Kim5db8a732011-06-15 15:08:48 -07001627 inc_zone_page_state(page, NR_ISOLATED_ANON +
Hugh Dickins9c620e22013-02-22 16:35:14 -08001628 page_is_file_cache(page));
Andi Kleenfacb6012009-12-16 12:20:00 +01001629 list_add(&page->lru, &pagelist);
Mel Gorman77f1fe62011-01-13 15:45:57 -08001630 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001631 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Andi Kleenfacb6012009-12-16 12:20:00 +01001632 if (ret) {
Andrea Arcangeli57fc4a52011-02-01 15:52:32 -08001633 putback_lru_pages(&pagelist);
Andi Kleenfb46e732010-09-27 23:31:30 +02001634 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001635 pfn, ret, page->flags);
1636 if (ret > 0)
1637 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001638 } else {
Naoya Horiguchif15bdfa2013-07-03 15:02:37 -07001639 /*
1640 * After page migration succeeds, the source page can
1641 * be trapped in pagevec and actual freeing is delayed.
1642 * Freeing code works differently based on PG_hwpoison,
1643 * so there's a race. We need to make sure that the
1644 * source page should be freed back to buddy before
1645 * setting PG_hwpoison.
1646 */
1647 if (!is_free_buddy_page(page))
1648 lru_add_drain_all();
1649 if (!is_free_buddy_page(page))
1650 drain_all_pages();
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001651 SetPageHWPoison(page);
Naoya Horiguchif15bdfa2013-07-03 15:02:37 -07001652 if (!is_free_buddy_page(page))
1653 pr_info("soft offline: %#lx: page leaked\n",
1654 pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001655 atomic_long_inc(&num_poisoned_pages);
Andi Kleenfacb6012009-12-16 12:20:00 +01001656 }
1657 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001658 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
Dean Nelsondd73e852011-10-31 17:09:04 -07001659 pfn, ret, page_count(page), page->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001660 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001661 return ret;
1662}