blob: 35ef28acf137c0ab76393ede3dbc1c3d820f5c37 [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);
Tejun Heob1664922014-02-11 11:52:49 -0500148 ino = cgroup_ino(css->cgroup);
Andi Kleen4fd466e2009-12-16 12:19:59 +0100149 css_put(css);
150
Tejun Heob1664922014-02-11 11:52:49 -0500151 if (!ino || ino != hwpoison_filter_memcg)
Andi Kleen4fd466e2009-12-16 12:19:59 +0100152 return -EINVAL;
153
154 return 0;
155}
156#else
157static int hwpoison_filter_task(struct page *p) { return 0; }
158#endif
159
Wu Fengguang7c116f22009-12-16 12:19:59 +0100160int hwpoison_filter(struct page *p)
161{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100162 if (!hwpoison_filter_enable)
163 return 0;
164
Wu Fengguang7c116f22009-12-16 12:19:59 +0100165 if (hwpoison_filter_dev(p))
166 return -EINVAL;
167
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100168 if (hwpoison_filter_flags(p))
169 return -EINVAL;
170
Andi Kleen4fd466e2009-12-16 12:19:59 +0100171 if (hwpoison_filter_task(p))
172 return -EINVAL;
173
Wu Fengguang7c116f22009-12-16 12:19:59 +0100174 return 0;
175}
Andi Kleen27df5062009-12-21 19:56:42 +0100176#else
177int hwpoison_filter(struct page *p)
178{
179 return 0;
180}
181#endif
182
Wu Fengguang7c116f22009-12-16 12:19:59 +0100183EXPORT_SYMBOL_GPL(hwpoison_filter);
184
Andi Kleen6a460792009-09-16 11:50:15 +0200185/*
Tony Luck7329bbe2011-12-13 09:27:58 -0800186 * Send all the processes who have the page mapped a signal.
187 * ``action optional'' if they are not immediately affected by the error
188 * ``action required'' if error happened in current execution context
Andi Kleen6a460792009-09-16 11:50:15 +0200189 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800190static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
191 unsigned long pfn, struct page *page, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200192{
193 struct siginfo si;
194 int ret;
195
196 printk(KERN_ERR
Tony Luck7329bbe2011-12-13 09:27:58 -0800197 "MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200198 pfn, t->comm, t->pid);
199 si.si_signo = SIGBUS;
200 si.si_errno = 0;
Andi Kleen6a460792009-09-16 11:50:15 +0200201 si.si_addr = (void *)addr;
202#ifdef __ARCH_SI_TRAPNO
203 si.si_trapno = trapno;
204#endif
Wanpeng Lif9121152013-09-11 14:22:52 -0700205 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
Tony Luck7329bbe2011-12-13 09:27:58 -0800206
207 if ((flags & MF_ACTION_REQUIRED) && t == current) {
208 si.si_code = BUS_MCEERR_AR;
209 ret = force_sig_info(SIGBUS, &si, t);
210 } else {
211 /*
212 * Don't use force here, it's convenient if the signal
213 * can be temporarily blocked.
214 * This could cause a loop when the user sets SIGBUS
215 * to SIG_IGN, but hopefully no one will do that?
216 */
217 si.si_code = BUS_MCEERR_AO;
218 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
219 }
Andi Kleen6a460792009-09-16 11:50:15 +0200220 if (ret < 0)
221 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
222 t->comm, t->pid, ret);
223 return ret;
224}
225
226/*
Andi Kleen588f9ce2009-12-16 12:19:57 +0100227 * When a unknown page type is encountered drain as many buffers as possible
228 * in the hope to turn the page into a LRU or free page, which we can handle.
229 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100230void shake_page(struct page *p, int access)
Andi Kleen588f9ce2009-12-16 12:19:57 +0100231{
232 if (!PageSlab(p)) {
233 lru_add_drain_all();
234 if (PageLRU(p))
235 return;
236 drain_all_pages();
237 if (PageLRU(p) || is_free_buddy_page(p))
238 return;
239 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100240
Andi Kleen588f9ce2009-12-16 12:19:57 +0100241 /*
Jin Dongmingaf241a02011-02-01 15:52:41 -0800242 * Only call shrink_slab here (which would also shrink other caches) if
243 * access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100244 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100245 if (access) {
246 int nr;
Dave Chinner0ce3d742013-08-28 10:18:03 +1000247 int nid = page_to_nid(p);
Andi Kleenfacb6012009-12-16 12:20:00 +0100248 do {
Ying Hana09ed5e2011-05-24 17:12:26 -0700249 struct shrink_control shrink = {
250 .gfp_mask = GFP_KERNEL,
Ying Hana09ed5e2011-05-24 17:12:26 -0700251 };
Dave Chinner0ce3d742013-08-28 10:18:03 +1000252 node_set(nid, shrink.nodes_to_scan);
Ying Hana09ed5e2011-05-24 17:12:26 -0700253
Ying Han1495f232011-05-24 17:12:27 -0700254 nr = shrink_slab(&shrink, 1000, 1000);
Andi Kleen47f43e72010-09-28 07:37:55 +0200255 if (page_count(p) == 1)
Andi Kleenfacb6012009-12-16 12:20:00 +0100256 break;
257 } while (nr > 10);
258 }
Andi Kleen588f9ce2009-12-16 12:19:57 +0100259}
260EXPORT_SYMBOL_GPL(shake_page);
261
262/*
Andi Kleen6a460792009-09-16 11:50:15 +0200263 * Kill all processes that have a poisoned page mapped and then isolate
264 * the page.
265 *
266 * General strategy:
267 * Find all processes having the page mapped and kill them.
268 * But we keep a page reference around so that the page is not
269 * actually freed yet.
270 * Then stash the page away
271 *
272 * There's no convenient way to get back to mapped processes
273 * from the VMAs. So do a brute-force search over all
274 * running processes.
275 *
276 * Remember that machine checks are not common (or rather
277 * if they are common you have other problems), so this shouldn't
278 * be a performance issue.
279 *
280 * Also there are some races possible while we get from the
281 * error detection to actually handle it.
282 */
283
284struct to_kill {
285 struct list_head nd;
286 struct task_struct *tsk;
287 unsigned long addr;
Andi Kleen9033ae12010-09-27 23:36:05 +0200288 char addr_valid;
Andi Kleen6a460792009-09-16 11:50:15 +0200289};
290
291/*
292 * Failure handling: if we can't find or can't kill a process there's
293 * not much we can do. We just print a message and ignore otherwise.
294 */
295
296/*
297 * Schedule a process for later kill.
298 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
299 * TBD would GFP_NOIO be enough?
300 */
301static void add_to_kill(struct task_struct *tsk, struct page *p,
302 struct vm_area_struct *vma,
303 struct list_head *to_kill,
304 struct to_kill **tkc)
305{
306 struct to_kill *tk;
307
308 if (*tkc) {
309 tk = *tkc;
310 *tkc = NULL;
311 } else {
312 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
313 if (!tk) {
314 printk(KERN_ERR
315 "MCE: Out of memory while machine check handling\n");
316 return;
317 }
318 }
319 tk->addr = page_address_in_vma(p, vma);
320 tk->addr_valid = 1;
321
322 /*
323 * In theory we don't have to kill when the page was
324 * munmaped. But it could be also a mremap. Since that's
325 * likely very rare kill anyways just out of paranoia, but use
326 * a SIGKILL because the error is not contained anymore.
327 */
328 if (tk->addr == -EFAULT) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200329 pr_info("MCE: Unable to find user space address %lx in %s\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200330 page_to_pfn(p), tsk->comm);
331 tk->addr_valid = 0;
332 }
333 get_task_struct(tsk);
334 tk->tsk = tsk;
335 list_add_tail(&tk->nd, to_kill);
336}
337
338/*
339 * Kill the processes that have been collected earlier.
340 *
341 * Only do anything when DOIT is set, otherwise just free the list
342 * (this is used for clean pages which do not need killing)
343 * Also when FAIL is set do a force kill because something went
344 * wrong earlier.
345 */
Tony Luck6751ed62012-07-11 10:20:47 -0700346static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -0800347 int fail, struct page *page, unsigned long pfn,
348 int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200349{
350 struct to_kill *tk, *next;
351
352 list_for_each_entry_safe (tk, next, to_kill, nd) {
Tony Luck6751ed62012-07-11 10:20:47 -0700353 if (forcekill) {
Andi Kleen6a460792009-09-16 11:50:15 +0200354 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200355 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200356 * make sure the process doesn't catch the
357 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200358 */
359 if (fail || tk->addr_valid == 0) {
360 printk(KERN_ERR
361 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
362 pfn, tk->tsk->comm, tk->tsk->pid);
363 force_sig(SIGKILL, tk->tsk);
364 }
365
366 /*
367 * In theory the process could have mapped
368 * something else on the address in-between. We could
369 * check for that, but we need to tell the
370 * process anyways.
371 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800372 else if (kill_proc(tk->tsk, tk->addr, trapno,
373 pfn, page, flags) < 0)
Andi Kleen6a460792009-09-16 11:50:15 +0200374 printk(KERN_ERR
375 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
376 pfn, tk->tsk->comm, tk->tsk->pid);
377 }
378 put_task_struct(tk->tsk);
379 kfree(tk);
380 }
381}
382
383static int task_early_kill(struct task_struct *tsk)
384{
385 if (!tsk->mm)
386 return 0;
387 if (tsk->flags & PF_MCE_PROCESS)
388 return !!(tsk->flags & PF_MCE_EARLY);
389 return sysctl_memory_failure_early_kill;
390}
391
392/*
393 * Collect processes when the error hit an anonymous page.
394 */
395static void collect_procs_anon(struct page *page, struct list_head *to_kill,
396 struct to_kill **tkc)
397{
398 struct vm_area_struct *vma;
399 struct task_struct *tsk;
400 struct anon_vma *av;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700401 pgoff_t pgoff;
Andi Kleen6a460792009-09-16 11:50:15 +0200402
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000403 av = page_lock_anon_vma_read(page);
Andi Kleen6a460792009-09-16 11:50:15 +0200404 if (av == NULL) /* Not actually mapped anymore */
Peter Zijlstra9b679322011-06-27 16:18:09 -0700405 return;
406
Michel Lespinassebf181b92012-10-08 16:31:39 -0700407 pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700408 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;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700414 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
415 pgoff, pgoff) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800416 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200417 if (!page_mapped_in_vma(page, vma))
418 continue;
419 if (vma->vm_mm == tsk->mm)
420 add_to_kill(tsk, page, vma, to_kill, tkc);
421 }
422 }
Andi Kleen6a460792009-09-16 11:50:15 +0200423 read_unlock(&tasklist_lock);
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000424 page_unlock_anon_vma_read(av);
Andi Kleen6a460792009-09-16 11:50:15 +0200425}
426
427/*
428 * Collect processes when the error hit a file mapped page.
429 */
430static void collect_procs_file(struct page *page, struct list_head *to_kill,
431 struct to_kill **tkc)
432{
433 struct vm_area_struct *vma;
434 struct task_struct *tsk;
Andi Kleen6a460792009-09-16 11:50:15 +0200435 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
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -0700445 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
Andi Kleen6a460792009-09-16 11:50:15 +0200446 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/*
Zhi Yong Wu549543d2014-01-21 15:49:08 -0800610 * Dirty pagecache page
Andi Kleen6a460792009-09-16 11:50:15 +0200611 * 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
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800782 { sc|dirty, sc|dirty, "dirty swapcache", me_swapcache_dirty },
783 { sc|dirty, sc, "clean swapcache", me_swapcache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200784
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800785 { mlock|dirty, mlock|dirty, "dirty mlocked LRU", me_pagecache_dirty },
Naoya Horiguchie3986292013-04-29 15:06:08 -0700786 { mlock|dirty, mlock, "clean mlocked LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200787
Naoya Horiguchi5f4b9fc2013-02-22 16:35:53 -0800788 { unevict|dirty, unevict|dirty, "dirty unevictable LRU", me_pagecache_dirty },
Naoya Horiguchie3986292013-04-29 15:06:08 -0700789 { unevict|dirty, unevict, "clean unevictable LRU", me_pagecache_clean },
Naoya Horiguchi5f4b9fc2013-02-22 16:35:53 -0800790
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800791 { lru|dirty, lru|dirty, "dirty LRU", me_pagecache_dirty },
Andi Kleen6a460792009-09-16 11:50:15 +0200792 { 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
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800813/*
814 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
815 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
816 */
Andi Kleen6a460792009-09-16 11:50:15 +0200817static void action_result(unsigned long pfn, char *msg, int result)
818{
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800819 pr_err("MCE %#lx: %s page recovery: %s\n",
820 pfn, msg, action_name[result]);
Andi Kleen6a460792009-09-16 11:50:15 +0200821}
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,
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800855 int trapno, int flags, struct page **hpagep)
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;
Tony Luck6751ed62012-07-11 10:20:47 -0700861 int kill = 1, forcekill;
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800862 struct page *hpage = *hpagep;
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);
Tony Luck6751ed62012-07-11 10:20:47 -0700891 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900892 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 }
Naoya Horiguchia3e0f9e2014-01-02 12:58:51 -0800937 /*
938 * We pinned the head page for hwpoison handling,
939 * now we split the thp and we are interested in
940 * the hwpoisoned raw page, so move the refcount
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800941 * to it. Similarly, page lock is shifted.
Naoya Horiguchia3e0f9e2014-01-02 12:58:51 -0800942 */
943 if (hpage != p) {
Naoya Horiguchi8d547ff2014-02-10 14:25:50 -0800944 if (!(flags & MF_COUNT_INCREASED)) {
945 put_page(hpage);
946 get_page(p);
947 }
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800948 lock_page(p);
949 unlock_page(hpage);
950 *hpagep = p;
Naoya Horiguchia3e0f9e2014-01-02 12:58:51 -0800951 }
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800952 /* THP is split, so ppage should be the real poisoned page. */
953 ppage = p;
Jin Dongmingefeda7a2011-02-01 15:52:39 -0800954 }
955 }
956
Andi Kleen6a460792009-09-16 11:50:15 +0200957 /*
958 * First collect all the processes that have the page
959 * mapped in dirty form. This has to be done before try_to_unmap,
960 * because ttu takes the rmap data structures down.
961 *
962 * Error handling: We ignore errors here because
963 * there's nothing that can be done.
964 */
965 if (kill)
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800966 collect_procs(ppage, &tokill);
Andi Kleen6a460792009-09-16 11:50:15 +0200967
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800968 ret = try_to_unmap(ppage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +0200969 if (ret != SWAP_SUCCESS)
970 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800971 pfn, page_mapcount(ppage));
972
Andi Kleen6a460792009-09-16 11:50:15 +0200973 /*
974 * Now that the dirty bit has been propagated to the
975 * struct page and all unmaps done we can decide if
976 * killing is needed or not. Only kill when the page
Tony Luck6751ed62012-07-11 10:20:47 -0700977 * was dirty or the process is not restartable,
978 * otherwise the tokill list is merely
Andi Kleen6a460792009-09-16 11:50:15 +0200979 * freed. When there was a problem unmapping earlier
980 * use a more force-full uncatchable kill to prevent
981 * any accesses to the poisoned memory.
982 */
Tony Luck6751ed62012-07-11 10:20:47 -0700983 forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL);
984 kill_procs(&tokill, forcekill, trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -0800985 ret != SWAP_SUCCESS, p, pfn, flags);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100986
987 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200988}
989
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900990static void set_page_hwpoison_huge_page(struct page *hpage)
991{
992 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -0700993 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900994 for (i = 0; i < nr_pages; i++)
995 SetPageHWPoison(hpage + i);
996}
997
998static void clear_page_hwpoison_huge_page(struct page *hpage)
999{
1000 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -07001001 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001002 for (i = 0; i < nr_pages; i++)
1003 ClearPageHWPoison(hpage + i);
1004}
1005
Tony Luckcd42f4a2011-12-15 10:48:12 -08001006/**
1007 * memory_failure - Handle memory failure of a page.
1008 * @pfn: Page Number of the corrupted page
1009 * @trapno: Trap number reported in the signal to user space.
1010 * @flags: fine tune action taken
1011 *
1012 * This function is called by the low level machine check code
1013 * of an architecture when it detects hardware memory corruption
1014 * of a page. It tries its best to recover, which includes
1015 * dropping pages, killing processes etc.
1016 *
1017 * The function is primarily of use for corruptions that
1018 * happen outside the current execution context (e.g. when
1019 * detected by a background scrubber)
1020 *
1021 * Must run in process context (e.g. a work queue) with interrupts
1022 * enabled and no spinlocks hold.
1023 */
1024int memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +02001025{
1026 struct page_state *ps;
1027 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001028 struct page *hpage;
Andi Kleen6a460792009-09-16 11:50:15 +02001029 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001030 unsigned int nr_pages;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001031 unsigned long page_flags;
Andi Kleen6a460792009-09-16 11:50:15 +02001032
1033 if (!sysctl_memory_failure_recovery)
1034 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1035
1036 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +01001037 printk(KERN_ERR
1038 "MCE %#lx: memory outside kernel control\n",
1039 pfn);
1040 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +02001041 }
1042
1043 p = pfn_to_page(pfn);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001044 hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +02001045 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001046 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001047 return 0;
1048 }
1049
Naoya Horiguchi4db0e952013-02-22 16:34:05 -08001050 /*
1051 * Currently errors on hugetlbfs pages are measured in hugepage units,
1052 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1053 * transparent hugepages, they are supposed to be split and error
1054 * measurement is done in normal page units. So nr_pages should be one
1055 * in this case.
1056 */
1057 if (PageHuge(p))
1058 nr_pages = 1 << compound_order(hpage);
1059 else /* normal page or thp */
1060 nr_pages = 1;
Xishi Qiu293c07e2013-02-22 16:34:02 -08001061 atomic_long_add(nr_pages, &num_poisoned_pages);
Andi Kleen6a460792009-09-16 11:50:15 +02001062
1063 /*
1064 * We need/can do nothing about count=0 pages.
1065 * 1) it's a free page, and therefore in safe hand:
1066 * prep_new_page() will be the gate keeper.
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001067 * 2) it's a free hugepage, which is also safe:
1068 * an affected hugepage will be dequeued from hugepage freelist,
1069 * so there's no concern about reusing it ever after.
1070 * 3) it's part of a non-compound high order page.
Andi Kleen6a460792009-09-16 11:50:15 +02001071 * Implies some kernel user: cannot stop them from
1072 * R/W the page; let's pray that the page has been
1073 * used and will be freed some time later.
1074 * In fact it's dangerous to directly bump up page count from 0,
1075 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1076 */
Andi Kleen82ba0112009-12-16 12:19:57 +01001077 if (!(flags & MF_COUNT_INCREASED) &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001078 !get_page_unless_zero(hpage)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001079 if (is_free_buddy_page(p)) {
1080 action_result(pfn, "free buddy", DELAYED);
1081 return 0;
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001082 } else if (PageHuge(hpage)) {
1083 /*
1084 * Check "just unpoisoned", "filter hit", and
1085 * "race with other subpage."
1086 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001087 lock_page(hpage);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001088 if (!PageHWPoison(hpage)
1089 || (hwpoison_filter(p) && TestClearPageHWPoison(p))
1090 || (p != hpage && TestSetPageHWPoison(hpage))) {
Xishi Qiu293c07e2013-02-22 16:34:02 -08001091 atomic_long_sub(nr_pages, &num_poisoned_pages);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001092 return 0;
1093 }
1094 set_page_hwpoison_huge_page(hpage);
1095 res = dequeue_hwpoisoned_huge_page(hpage);
1096 action_result(pfn, "free huge",
1097 res ? IGNORED : DELAYED);
1098 unlock_page(hpage);
1099 return res;
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001100 } else {
1101 action_result(pfn, "high order kernel", IGNORED);
1102 return -EBUSY;
1103 }
Andi Kleen6a460792009-09-16 11:50:15 +02001104 }
1105
1106 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001107 * We ignore non-LRU pages for good reasons.
1108 * - PG_locked is only well defined for LRU pages and a few others
1109 * - to avoid races with __set_page_locked()
1110 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1111 * The check (unnecessarily) ignores LRU pages being isolated and
1112 * walked by the page reclaim code, however that's not a big loss.
1113 */
Dean Nelson385de352012-03-21 16:34:05 -07001114 if (!PageHuge(p) && !PageTransTail(p)) {
Jin Dongmingaf241a02011-02-01 15:52:41 -08001115 if (!PageLRU(p))
1116 shake_page(p, 0);
1117 if (!PageLRU(p)) {
1118 /*
1119 * shake_page could have turned it free.
1120 */
1121 if (is_free_buddy_page(p)) {
Wanpeng Li2d421ac2013-09-30 13:45:23 -07001122 if (flags & MF_COUNT_INCREASED)
1123 action_result(pfn, "free buddy", DELAYED);
1124 else
1125 action_result(pfn, "free buddy, 2nd try", DELAYED);
Jin Dongmingaf241a02011-02-01 15:52:41 -08001126 return 0;
1127 }
1128 action_result(pfn, "non LRU", IGNORED);
1129 put_page(p);
1130 return -EBUSY;
Andi Kleen0474a602009-12-16 12:20:00 +01001131 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001132 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001133
1134 /*
Andi Kleen6a460792009-09-16 11:50:15 +02001135 * Lock the page and wait for writeback to finish.
1136 * It's very difficult to mess with pages currently under IO
1137 * and in many cases impossible, so we just avoid it here.
1138 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001139 lock_page(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001140
1141 /*
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001142 * We use page flags to determine what action should be taken, but
1143 * the flags can be modified by the error containment action. One
1144 * example is an mlocked page, where PG_mlocked is cleared by
1145 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1146 * correctly, we save a copy of the page flags at this time.
1147 */
1148 page_flags = p->flags;
1149
1150 /*
Wu Fengguang847ce402009-12-16 12:19:58 +01001151 * unpoison always clear PG_hwpoison inside page lock
1152 */
1153 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001154 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001155 res = 0;
1156 goto out;
1157 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001158 if (hwpoison_filter(p)) {
1159 if (TestClearPageHWPoison(p))
Xishi Qiu293c07e2013-02-22 16:34:02 -08001160 atomic_long_sub(nr_pages, &num_poisoned_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001161 unlock_page(hpage);
1162 put_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001163 return 0;
1164 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001165
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001166 /*
1167 * For error on the tail page, we should set PG_hwpoison
1168 * on the head page to show that the hugepage is hwpoisoned
1169 */
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001170 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001171 action_result(pfn, "hugepage already hardware poisoned",
1172 IGNORED);
1173 unlock_page(hpage);
1174 put_page(hpage);
1175 return 0;
1176 }
1177 /*
1178 * Set PG_hwpoison on all pages in an error hugepage,
1179 * because containment is done in hugepage unit for now.
1180 * Since we have done TestSetPageHWPoison() for the head page with
1181 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1182 */
1183 if (PageHuge(p))
1184 set_page_hwpoison_huge_page(hpage);
1185
Andi Kleen6a460792009-09-16 11:50:15 +02001186 wait_on_page_writeback(p);
1187
1188 /*
1189 * Now take care of user space mappings.
Minchan Kime64a7822011-03-22 16:32:44 -07001190 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001191 *
1192 * When the raw error page is thp tail page, hpage points to the raw
1193 * page after thp split.
Andi Kleen6a460792009-09-16 11:50:15 +02001194 */
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001195 if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)
1196 != SWAP_SUCCESS) {
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001197 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1198 res = -EBUSY;
1199 goto out;
1200 }
Andi Kleen6a460792009-09-16 11:50:15 +02001201
1202 /*
1203 * Torn down by someone else?
1204 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001205 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Andi Kleen6a460792009-09-16 11:50:15 +02001206 action_result(pfn, "already truncated LRU", IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001207 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001208 goto out;
1209 }
1210
1211 res = -EBUSY;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001212 /*
1213 * The first check uses the current page flags which may not have any
1214 * relevant information. The second check with the saved page flagss is
1215 * carried out only if the first check can't determine the page status.
1216 */
1217 for (ps = error_states;; ps++)
1218 if ((p->flags & ps->mask) == ps->res)
Andi Kleen6a460792009-09-16 11:50:15 +02001219 break;
Wanpeng Li841fcc52013-09-11 14:22:50 -07001220
1221 page_flags |= (p->flags & (1UL << PG_dirty));
1222
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001223 if (!ps->mask)
1224 for (ps = error_states;; ps++)
1225 if ((page_flags & ps->mask) == ps->res)
1226 break;
1227 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001228out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001229 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001230 return res;
1231}
Tony Luckcd42f4a2011-12-15 10:48:12 -08001232EXPORT_SYMBOL_GPL(memory_failure);
Wu Fengguang847ce402009-12-16 12:19:58 +01001233
Huang Yingea8f5fb2011-07-13 13:14:27 +08001234#define MEMORY_FAILURE_FIFO_ORDER 4
1235#define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1236
1237struct memory_failure_entry {
1238 unsigned long pfn;
1239 int trapno;
1240 int flags;
1241};
1242
1243struct memory_failure_cpu {
1244 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1245 MEMORY_FAILURE_FIFO_SIZE);
1246 spinlock_t lock;
1247 struct work_struct work;
1248};
1249
1250static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1251
1252/**
1253 * memory_failure_queue - Schedule handling memory failure of a page.
1254 * @pfn: Page Number of the corrupted page
1255 * @trapno: Trap number reported in the signal to user space.
1256 * @flags: Flags for memory failure handling
1257 *
1258 * This function is called by the low level hardware error handler
1259 * when it detects hardware memory corruption of a page. It schedules
1260 * the recovering of error page, including dropping pages, killing
1261 * processes etc.
1262 *
1263 * The function is primarily of use for corruptions that
1264 * happen outside the current execution context (e.g. when
1265 * detected by a background scrubber)
1266 *
1267 * Can run in IRQ context.
1268 */
1269void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1270{
1271 struct memory_failure_cpu *mf_cpu;
1272 unsigned long proc_flags;
1273 struct memory_failure_entry entry = {
1274 .pfn = pfn,
1275 .trapno = trapno,
1276 .flags = flags,
1277 };
1278
1279 mf_cpu = &get_cpu_var(memory_failure_cpu);
1280 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
Stefani Seibold498d3192013-11-14 14:32:17 -08001281 if (kfifo_put(&mf_cpu->fifo, entry))
Huang Yingea8f5fb2011-07-13 13:14:27 +08001282 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1283 else
Joe Perches8e33a522013-07-25 11:53:25 -07001284 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
Huang Yingea8f5fb2011-07-13 13:14:27 +08001285 pfn);
1286 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1287 put_cpu_var(memory_failure_cpu);
1288}
1289EXPORT_SYMBOL_GPL(memory_failure_queue);
1290
1291static void memory_failure_work_func(struct work_struct *work)
1292{
1293 struct memory_failure_cpu *mf_cpu;
1294 struct memory_failure_entry entry = { 0, };
1295 unsigned long proc_flags;
1296 int gotten;
1297
1298 mf_cpu = &__get_cpu_var(memory_failure_cpu);
1299 for (;;) {
1300 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1301 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1302 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1303 if (!gotten)
1304 break;
Naveen N. Raocf870c72013-07-10 14:57:01 +05301305 if (entry.flags & MF_SOFT_OFFLINE)
1306 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1307 else
1308 memory_failure(entry.pfn, entry.trapno, entry.flags);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001309 }
1310}
1311
1312static int __init memory_failure_init(void)
1313{
1314 struct memory_failure_cpu *mf_cpu;
1315 int cpu;
1316
1317 for_each_possible_cpu(cpu) {
1318 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1319 spin_lock_init(&mf_cpu->lock);
1320 INIT_KFIFO(mf_cpu->fifo);
1321 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1322 }
1323
1324 return 0;
1325}
1326core_initcall(memory_failure_init);
1327
Wu Fengguang847ce402009-12-16 12:19:58 +01001328/**
1329 * unpoison_memory - Unpoison a previously poisoned page
1330 * @pfn: Page number of the to be unpoisoned page
1331 *
1332 * Software-unpoison a page that has been poisoned by
1333 * memory_failure() earlier.
1334 *
1335 * This is only done on the software-level, so it only works
1336 * for linux injected failures, not real hardware failures
1337 *
1338 * Returns 0 for success, otherwise -errno.
1339 */
1340int unpoison_memory(unsigned long pfn)
1341{
1342 struct page *page;
1343 struct page *p;
1344 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001345 unsigned int nr_pages;
Wu Fengguang847ce402009-12-16 12:19:58 +01001346
1347 if (!pfn_valid(pfn))
1348 return -ENXIO;
1349
1350 p = pfn_to_page(pfn);
1351 page = compound_head(p);
1352
1353 if (!PageHWPoison(p)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001354 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001355 return 0;
1356 }
1357
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001358 /*
1359 * unpoison_memory() can encounter thp only when the thp is being
1360 * worked by memory_failure() and the page lock is not held yet.
1361 * In such case, we yield to memory_failure() and make unpoison fail.
1362 */
Wanpeng Lie76d30e2013-09-30 13:45:22 -07001363 if (!PageHuge(page) && PageTransHuge(page)) {
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001364 pr_info("MCE: Memory failure is now running on %#lx\n", pfn);
1365 return 0;
1366 }
1367
Wanpeng Lif9121152013-09-11 14:22:52 -07001368 nr_pages = 1 << compound_order(page);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001369
Wu Fengguang847ce402009-12-16 12:19:58 +01001370 if (!get_page_unless_zero(page)) {
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001371 /*
1372 * Since HWPoisoned hugepage should have non-zero refcount,
1373 * race between memory failure and unpoison seems to happen.
1374 * In such case unpoison fails and memory failure runs
1375 * to the end.
1376 */
1377 if (PageHuge(page)) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001378 pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001379 return 0;
1380 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001381 if (TestClearPageHWPoison(p))
Wanpeng Lidd9538a2013-09-11 14:22:54 -07001382 atomic_long_dec(&num_poisoned_pages);
Andi Kleenfb46e732010-09-27 23:31:30 +02001383 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001384 return 0;
1385 }
1386
Jens Axboe7eaceac2011-03-10 08:52:07 +01001387 lock_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001388 /*
1389 * This test is racy because PG_hwpoison is set outside of page lock.
1390 * That's acceptable because that won't trigger kernel panic. Instead,
1391 * the PG_hwpoison page will be caught and isolated on the entrance to
1392 * the free buddy page pool.
1393 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001394 if (TestClearPageHWPoison(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001395 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
Xishi Qiu293c07e2013-02-22 16:34:02 -08001396 atomic_long_sub(nr_pages, &num_poisoned_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001397 freeit = 1;
Naoya Horiguchi6a901812010-09-08 10:19:40 +09001398 if (PageHuge(page))
1399 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001400 }
1401 unlock_page(page);
1402
1403 put_page(page);
Wanpeng Li3ba5eeb2013-09-11 14:23:01 -07001404 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
Wu Fengguang847ce402009-12-16 12:19:58 +01001405 put_page(page);
1406
1407 return 0;
1408}
1409EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001410
1411static struct page *new_page(struct page *p, unsigned long private, int **x)
1412{
Andi Kleen12686d12009-12-16 12:20:01 +01001413 int nid = page_to_nid(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001414 if (PageHuge(p))
1415 return alloc_huge_page_node(page_hstate(compound_head(p)),
1416 nid);
1417 else
1418 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001419}
1420
1421/*
1422 * Safely get reference count of an arbitrary page.
1423 * Returns 0 for a free page, -EIO for a zero refcount page
1424 * that is not free, and 1 for any other page type.
1425 * For 1 the page is returned with increased page count, otherwise not.
1426 */
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001427static int __get_any_page(struct page *p, unsigned long pfn, int flags)
Andi Kleenfacb6012009-12-16 12:20:00 +01001428{
1429 int ret;
1430
1431 if (flags & MF_COUNT_INCREASED)
1432 return 1;
1433
1434 /*
Naoya Horiguchid950b952010-09-08 10:19:39 +09001435 * When the target page is a free hugepage, just remove it
1436 * from free hugepage list.
1437 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001438 if (!get_page_unless_zero(compound_head(p))) {
Naoya Horiguchid950b952010-09-08 10:19:39 +09001439 if (PageHuge(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001440 pr_info("%s: %#lx free huge page\n", __func__, pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001441 ret = 0;
Naoya Horiguchid950b952010-09-08 10:19:39 +09001442 } else if (is_free_buddy_page(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001443 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001444 ret = 0;
1445 } else {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001446 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1447 __func__, pfn, p->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001448 ret = -EIO;
1449 }
1450 } else {
1451 /* Not a free page */
1452 ret = 1;
1453 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001454 return ret;
1455}
1456
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001457static int get_any_page(struct page *page, unsigned long pfn, int flags)
1458{
1459 int ret = __get_any_page(page, pfn, flags);
1460
1461 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1462 /*
1463 * Try to free it.
1464 */
1465 put_page(page);
1466 shake_page(page, 1);
1467
1468 /*
1469 * Did it turn free?
1470 */
1471 ret = __get_any_page(page, pfn, 0);
1472 if (!PageLRU(page)) {
1473 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1474 pfn, page->flags);
1475 return -EIO;
1476 }
1477 }
1478 return ret;
1479}
1480
Naoya Horiguchid950b952010-09-08 10:19:39 +09001481static int soft_offline_huge_page(struct page *page, int flags)
1482{
1483 int ret;
1484 unsigned long pfn = page_to_pfn(page);
1485 struct page *hpage = compound_head(page);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001486 LIST_HEAD(pagelist);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001487
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001488 /*
1489 * This double-check of PageHWPoison is to avoid the race with
1490 * memory_failure(). See also comment in __soft_offline_page().
1491 */
1492 lock_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001493 if (PageHWPoison(hpage)) {
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001494 unlock_page(hpage);
1495 put_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001496 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001497 return -EBUSY;
Xishi Qiu0ebff322013-02-22 16:33:59 -08001498 }
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001499 unlock_page(hpage);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001500
Naoya Horiguchid950b952010-09-08 10:19:39 +09001501 /* Keep page count to indicate a given hugepage is isolated. */
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001502 list_move(&hpage->lru, &pagelist);
1503 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1504 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001505 if (ret) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001506 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1507 pfn, ret, page->flags);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001508 /*
1509 * We know that soft_offline_huge_page() tries to migrate
1510 * only one hugepage pointed to by hpage, so we need not
1511 * run through the pagelist here.
1512 */
1513 putback_active_hugepage(hpage);
1514 if (ret > 0)
1515 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001516 } else {
Jianguo Wua49ecbc2013-12-18 17:08:54 -08001517 /* overcommit hugetlb page will be freed to buddy */
1518 if (PageHuge(page)) {
1519 set_page_hwpoison_huge_page(hpage);
1520 dequeue_hwpoisoned_huge_page(hpage);
1521 atomic_long_add(1 << compound_order(hpage),
1522 &num_poisoned_pages);
1523 } else {
1524 SetPageHWPoison(page);
1525 atomic_long_inc(&num_poisoned_pages);
1526 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001527 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001528 return ret;
1529}
1530
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001531static int __soft_offline_page(struct page *page, int flags)
1532{
1533 int ret;
1534 unsigned long pfn = page_to_pfn(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001535
1536 /*
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001537 * Check PageHWPoison again inside page lock because PageHWPoison
1538 * is set by memory_failure() outside page lock. Note that
1539 * memory_failure() also double-checks PageHWPoison inside page lock,
1540 * so there's no race between soft_offline_page() and memory_failure().
Andi Kleenfacb6012009-12-16 12:20:00 +01001541 */
Xishi Qiu0ebff322013-02-22 16:33:59 -08001542 lock_page(page);
1543 wait_on_page_writeback(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001544 if (PageHWPoison(page)) {
1545 unlock_page(page);
1546 put_page(page);
1547 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1548 return -EBUSY;
1549 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001550 /*
1551 * Try to invalidate first. This should work for
1552 * non dirty unmapped page cache pages.
1553 */
1554 ret = invalidate_inode_page(page);
1555 unlock_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001556 /*
Andi Kleenfacb6012009-12-16 12:20:00 +01001557 * RED-PEN would be better to keep it isolated here, but we
1558 * would need to fix isolation locking first.
1559 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001560 if (ret == 1) {
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001561 put_page(page);
Andi Kleenfb46e732010-09-27 23:31:30 +02001562 pr_info("soft_offline: %#lx: invalidated\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001563 SetPageHWPoison(page);
1564 atomic_long_inc(&num_poisoned_pages);
1565 return 0;
Andi Kleenfacb6012009-12-16 12:20:00 +01001566 }
1567
1568 /*
1569 * Simple invalidation didn't work.
1570 * Try to migrate to a new page instead. migrate.c
1571 * handles a large number of cases for us.
1572 */
1573 ret = isolate_lru_page(page);
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001574 /*
1575 * Drop page reference which is came from get_any_page()
1576 * successful isolate_lru_page() already took another one.
1577 */
1578 put_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001579 if (!ret) {
1580 LIST_HEAD(pagelist);
Minchan Kim5db8a732011-06-15 15:08:48 -07001581 inc_zone_page_state(page, NR_ISOLATED_ANON +
Hugh Dickins9c620e22013-02-22 16:35:14 -08001582 page_is_file_cache(page));
Andi Kleenfacb6012009-12-16 12:20:00 +01001583 list_add(&page->lru, &pagelist);
Mel Gorman77f1fe62011-01-13 15:45:57 -08001584 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001585 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Andi Kleenfacb6012009-12-16 12:20:00 +01001586 if (ret) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001587 if (!list_empty(&pagelist)) {
1588 list_del(&page->lru);
1589 dec_zone_page_state(page, NR_ISOLATED_ANON +
1590 page_is_file_cache(page));
1591 putback_lru_page(page);
1592 }
1593
Andi Kleenfb46e732010-09-27 23:31:30 +02001594 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001595 pfn, ret, page->flags);
1596 if (ret > 0)
1597 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001598 } else {
Naoya Horiguchif15bdfa2013-07-03 15:02:37 -07001599 /*
1600 * After page migration succeeds, the source page can
1601 * be trapped in pagevec and actual freeing is delayed.
1602 * Freeing code works differently based on PG_hwpoison,
1603 * so there's a race. We need to make sure that the
1604 * source page should be freed back to buddy before
1605 * setting PG_hwpoison.
1606 */
1607 if (!is_free_buddy_page(page))
1608 lru_add_drain_all();
1609 if (!is_free_buddy_page(page))
1610 drain_all_pages();
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001611 SetPageHWPoison(page);
Naoya Horiguchif15bdfa2013-07-03 15:02:37 -07001612 if (!is_free_buddy_page(page))
1613 pr_info("soft offline: %#lx: page leaked\n",
1614 pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001615 atomic_long_inc(&num_poisoned_pages);
Andi Kleenfacb6012009-12-16 12:20:00 +01001616 }
1617 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001618 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
Dean Nelsondd73e852011-10-31 17:09:04 -07001619 pfn, ret, page_count(page), page->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001620 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001621 return ret;
1622}
Wanpeng Li86e05772013-09-11 14:22:56 -07001623
1624/**
1625 * soft_offline_page - Soft offline a page.
1626 * @page: page to offline
1627 * @flags: flags. Same as memory_failure().
1628 *
1629 * Returns 0 on success, otherwise negated errno.
1630 *
1631 * Soft offline a page, by migration or invalidation,
1632 * without killing anything. This is for the case when
1633 * a page is not corrupted yet (so it's still valid to access),
1634 * but has had a number of corrected errors and is better taken
1635 * out.
1636 *
1637 * The actual policy on when to do that is maintained by
1638 * user space.
1639 *
1640 * This should never impact any application or cause data loss,
1641 * however it might take some time.
1642 *
1643 * This is not a 100% solution for all memory, but tries to be
1644 * ``good enough'' for the majority of memory.
1645 */
1646int soft_offline_page(struct page *page, int flags)
1647{
1648 int ret;
1649 unsigned long pfn = page_to_pfn(page);
David Rientjes668f9abb2014-03-03 15:38:18 -08001650 struct page *hpage = compound_head(page);
Wanpeng Li86e05772013-09-11 14:22:56 -07001651
1652 if (PageHWPoison(page)) {
1653 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1654 return -EBUSY;
1655 }
1656 if (!PageHuge(page) && PageTransHuge(hpage)) {
1657 if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1658 pr_info("soft offline: %#lx: failed to split THP\n",
1659 pfn);
1660 return -EBUSY;
1661 }
1662 }
1663
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001664 /*
1665 * The lock_memory_hotplug prevents a race with memory hotplug.
1666 * This is a big hammer, a better would be nicer.
1667 */
1668 lock_memory_hotplug();
1669
1670 /*
1671 * Isolate the page, so that it doesn't get reallocated if it
1672 * was free. This flag should be kept set until the source page
1673 * is freed and PG_hwpoison on it is set.
1674 */
1675 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
1676 set_migratetype_isolate(page, true);
1677
Wanpeng Li86e05772013-09-11 14:22:56 -07001678 ret = get_any_page(page, pfn, flags);
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001679 unlock_memory_hotplug();
1680 if (ret > 0) { /* for in-use pages */
Wanpeng Li86e05772013-09-11 14:22:56 -07001681 if (PageHuge(page))
1682 ret = soft_offline_huge_page(page, flags);
1683 else
1684 ret = __soft_offline_page(page, flags);
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001685 } else if (ret == 0) { /* for free pages */
Wanpeng Li86e05772013-09-11 14:22:56 -07001686 if (PageHuge(page)) {
1687 set_page_hwpoison_huge_page(hpage);
1688 dequeue_hwpoisoned_huge_page(hpage);
1689 atomic_long_add(1 << compound_order(hpage),
1690 &num_poisoned_pages);
1691 } else {
1692 SetPageHWPoison(page);
1693 atomic_long_inc(&num_poisoned_pages);
1694 }
1695 }
Wanpeng Li86e05772013-09-11 14:22:56 -07001696 unset_migratetype_isolate(page, MIGRATE_MOVABLE);
1697 return ret;
1698}