blob: 95882692e747c2a534488190287e5954fba35d39 [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.
Andi Kleene0de78d2015-06-24 16:56:02 -070023 *
24 * It can be very tempting to add handling for obscure cases here.
25 * In general any code for handling new cases should only be added iff:
26 * - You know how to test it.
27 * - You have a test that can be added to mce-test
28 * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/
29 * - The case actually shows up as a frequent (top 10) page state in
30 * tools/vm/page-types when running a real workload.
Andi Kleen1c80b992010-09-27 23:09:51 +020031 *
32 * There are several operations here with exponential complexity because
33 * of unsuitable VM data structures. For example the operation to map back
34 * from RMAP chains to processes has to walk the complete process list and
35 * has non linear complexity with the number. But since memory corruptions
36 * are rare we hope to get away with this. This avoids impacting the core
37 * VM.
Andi Kleen6a460792009-09-16 11:50:15 +020038 */
Andi Kleen6a460792009-09-16 11:50:15 +020039#include <linux/kernel.h>
40#include <linux/mm.h>
41#include <linux/page-flags.h>
Wu Fengguang478c5ff2009-12-16 12:19:59 +010042#include <linux/kernel-page-flags.h>
Andi Kleen6a460792009-09-16 11:50:15 +020043#include <linux/sched.h>
Hugh Dickins01e00f82009-10-13 15:02:11 +010044#include <linux/ksm.h>
Andi Kleen6a460792009-09-16 11:50:15 +020045#include <linux/rmap.h>
Paul Gortmakerb9e15ba2011-05-26 16:00:52 -040046#include <linux/export.h>
Andi Kleen6a460792009-09-16 11:50:15 +020047#include <linux/pagemap.h>
48#include <linux/swap.h>
49#include <linux/backing-dev.h>
Andi Kleenfacb6012009-12-16 12:20:00 +010050#include <linux/migrate.h>
51#include <linux/page-isolation.h>
52#include <linux/suspend.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090053#include <linux/slab.h>
Huang Yingbf998152010-05-31 14:28:19 +080054#include <linux/swapops.h>
Naoya Horiguchi7af446a2010-05-28 09:29:17 +090055#include <linux/hugetlb.h>
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -080056#include <linux/memory_hotplug.h>
Minchan Kim5db8a732011-06-15 15:08:48 -070057#include <linux/mm_inline.h>
Huang Yingea8f5fb2011-07-13 13:14:27 +080058#include <linux/kfifo.h>
Andi Kleen6a460792009-09-16 11:50:15 +020059#include "internal.h"
Xie XiuQi97f0b132015-06-24 16:57:36 -070060#include "ras/ras_event.h"
Andi Kleen6a460792009-09-16 11:50:15 +020061
62int sysctl_memory_failure_early_kill __read_mostly = 0;
63
64int sysctl_memory_failure_recovery __read_mostly = 1;
65
Xishi Qiu293c07e2013-02-22 16:34:02 -080066atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
Andi Kleen6a460792009-09-16 11:50:15 +020067
Andi Kleen27df5062009-12-21 19:56:42 +010068#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
69
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010070u32 hwpoison_filter_enable = 0;
Wu Fengguang7c116f22009-12-16 12:19:59 +010071u32 hwpoison_filter_dev_major = ~0U;
72u32 hwpoison_filter_dev_minor = ~0U;
Wu Fengguang478c5ff2009-12-16 12:19:59 +010073u64 hwpoison_filter_flags_mask;
74u64 hwpoison_filter_flags_value;
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010075EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
Wu Fengguang7c116f22009-12-16 12:19:59 +010076EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
77EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
Wu Fengguang478c5ff2009-12-16 12:19:59 +010078EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
79EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
Wu Fengguang7c116f22009-12-16 12:19:59 +010080
81static int hwpoison_filter_dev(struct page *p)
82{
83 struct address_space *mapping;
84 dev_t dev;
85
86 if (hwpoison_filter_dev_major == ~0U &&
87 hwpoison_filter_dev_minor == ~0U)
88 return 0;
89
90 /*
Andi Kleen1c80b992010-09-27 23:09:51 +020091 * page_mapping() does not accept slab pages.
Wu Fengguang7c116f22009-12-16 12:19:59 +010092 */
93 if (PageSlab(p))
94 return -EINVAL;
95
96 mapping = page_mapping(p);
97 if (mapping == NULL || mapping->host == NULL)
98 return -EINVAL;
99
100 dev = mapping->host->i_sb->s_dev;
101 if (hwpoison_filter_dev_major != ~0U &&
102 hwpoison_filter_dev_major != MAJOR(dev))
103 return -EINVAL;
104 if (hwpoison_filter_dev_minor != ~0U &&
105 hwpoison_filter_dev_minor != MINOR(dev))
106 return -EINVAL;
107
108 return 0;
109}
110
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100111static int hwpoison_filter_flags(struct page *p)
112{
113 if (!hwpoison_filter_flags_mask)
114 return 0;
115
116 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
117 hwpoison_filter_flags_value)
118 return 0;
119 else
120 return -EINVAL;
121}
122
Andi Kleen4fd466e2009-12-16 12:19:59 +0100123/*
124 * This allows stress tests to limit test scope to a collection of tasks
125 * by putting them under some memcg. This prevents killing unrelated/important
126 * processes such as /sbin/init. Note that the target task may share clean
127 * pages with init (eg. libc text), which is harmless. If the target task
128 * share _dirty_ pages with another task B, the test scheme must make sure B
129 * is also included in the memcg. At last, due to race conditions this filter
130 * can only guarantee that the page either belongs to the memcg tasks, or is
131 * a freed page.
132 */
Vladimir Davydov94a59fb2015-09-09 15:35:31 -0700133#ifdef CONFIG_MEMCG
Andi Kleen4fd466e2009-12-16 12:19:59 +0100134u64 hwpoison_filter_memcg;
135EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
136static int hwpoison_filter_task(struct page *p)
137{
Andi Kleen4fd466e2009-12-16 12:19:59 +0100138 if (!hwpoison_filter_memcg)
139 return 0;
140
Vladimir Davydov94a59fb2015-09-09 15:35:31 -0700141 if (page_cgroup_ino(p) != hwpoison_filter_memcg)
Andi Kleen4fd466e2009-12-16 12:19:59 +0100142 return -EINVAL;
143
144 return 0;
145}
146#else
147static int hwpoison_filter_task(struct page *p) { return 0; }
148#endif
149
Wu Fengguang7c116f22009-12-16 12:19:59 +0100150int hwpoison_filter(struct page *p)
151{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100152 if (!hwpoison_filter_enable)
153 return 0;
154
Wu Fengguang7c116f22009-12-16 12:19:59 +0100155 if (hwpoison_filter_dev(p))
156 return -EINVAL;
157
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100158 if (hwpoison_filter_flags(p))
159 return -EINVAL;
160
Andi Kleen4fd466e2009-12-16 12:19:59 +0100161 if (hwpoison_filter_task(p))
162 return -EINVAL;
163
Wu Fengguang7c116f22009-12-16 12:19:59 +0100164 return 0;
165}
Andi Kleen27df5062009-12-21 19:56:42 +0100166#else
167int hwpoison_filter(struct page *p)
168{
169 return 0;
170}
171#endif
172
Wu Fengguang7c116f22009-12-16 12:19:59 +0100173EXPORT_SYMBOL_GPL(hwpoison_filter);
174
Andi Kleen6a460792009-09-16 11:50:15 +0200175/*
Tony Luck7329bbe2011-12-13 09:27:58 -0800176 * Send all the processes who have the page mapped a signal.
177 * ``action optional'' if they are not immediately affected by the error
178 * ``action required'' if error happened in current execution context
Andi Kleen6a460792009-09-16 11:50:15 +0200179 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800180static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
181 unsigned long pfn, struct page *page, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200182{
183 struct siginfo si;
184 int ret;
185
186 printk(KERN_ERR
Tony Luck7329bbe2011-12-13 09:27:58 -0800187 "MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200188 pfn, t->comm, t->pid);
189 si.si_signo = SIGBUS;
190 si.si_errno = 0;
Andi Kleen6a460792009-09-16 11:50:15 +0200191 si.si_addr = (void *)addr;
192#ifdef __ARCH_SI_TRAPNO
193 si.si_trapno = trapno;
194#endif
Wanpeng Lif9121152013-09-11 14:22:52 -0700195 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
Tony Luck7329bbe2011-12-13 09:27:58 -0800196
Tony Lucka70ffca2014-06-04 16:10:59 -0700197 if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
Tony Luck7329bbe2011-12-13 09:27:58 -0800198 si.si_code = BUS_MCEERR_AR;
Tony Lucka70ffca2014-06-04 16:10:59 -0700199 ret = force_sig_info(SIGBUS, &si, current);
Tony Luck7329bbe2011-12-13 09:27:58 -0800200 } else {
201 /*
202 * Don't use force here, it's convenient if the signal
203 * can be temporarily blocked.
204 * This could cause a loop when the user sets SIGBUS
205 * to SIG_IGN, but hopefully no one will do that?
206 */
207 si.si_code = BUS_MCEERR_AO;
208 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
209 }
Andi Kleen6a460792009-09-16 11:50:15 +0200210 if (ret < 0)
211 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
212 t->comm, t->pid, ret);
213 return ret;
214}
215
216/*
Andi Kleen588f9ce2009-12-16 12:19:57 +0100217 * When a unknown page type is encountered drain as many buffers as possible
218 * in the hope to turn the page into a LRU or free page, which we can handle.
219 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100220void shake_page(struct page *p, int access)
Andi Kleen588f9ce2009-12-16 12:19:57 +0100221{
222 if (!PageSlab(p)) {
223 lru_add_drain_all();
224 if (PageLRU(p))
225 return;
Vlastimil Babkac0554322014-12-10 15:43:10 -0800226 drain_all_pages(page_zone(p));
Andi Kleen588f9ce2009-12-16 12:19:57 +0100227 if (PageLRU(p) || is_free_buddy_page(p))
228 return;
229 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100230
Andi Kleen588f9ce2009-12-16 12:19:57 +0100231 /*
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800232 * Only call shrink_node_slabs here (which would also shrink
233 * other caches) if access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100234 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800235 if (access)
236 drop_slab_node(page_to_nid(p));
Andi Kleen588f9ce2009-12-16 12:19:57 +0100237}
238EXPORT_SYMBOL_GPL(shake_page);
239
240/*
Andi Kleen6a460792009-09-16 11:50:15 +0200241 * Kill all processes that have a poisoned page mapped and then isolate
242 * the page.
243 *
244 * General strategy:
245 * Find all processes having the page mapped and kill them.
246 * But we keep a page reference around so that the page is not
247 * actually freed yet.
248 * Then stash the page away
249 *
250 * There's no convenient way to get back to mapped processes
251 * from the VMAs. So do a brute-force search over all
252 * running processes.
253 *
254 * Remember that machine checks are not common (or rather
255 * if they are common you have other problems), so this shouldn't
256 * be a performance issue.
257 *
258 * Also there are some races possible while we get from the
259 * error detection to actually handle it.
260 */
261
262struct to_kill {
263 struct list_head nd;
264 struct task_struct *tsk;
265 unsigned long addr;
Andi Kleen9033ae12010-09-27 23:36:05 +0200266 char addr_valid;
Andi Kleen6a460792009-09-16 11:50:15 +0200267};
268
269/*
270 * Failure handling: if we can't find or can't kill a process there's
271 * not much we can do. We just print a message and ignore otherwise.
272 */
273
274/*
275 * Schedule a process for later kill.
276 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
277 * TBD would GFP_NOIO be enough?
278 */
279static void add_to_kill(struct task_struct *tsk, struct page *p,
280 struct vm_area_struct *vma,
281 struct list_head *to_kill,
282 struct to_kill **tkc)
283{
284 struct to_kill *tk;
285
286 if (*tkc) {
287 tk = *tkc;
288 *tkc = NULL;
289 } else {
290 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
291 if (!tk) {
292 printk(KERN_ERR
293 "MCE: Out of memory while machine check handling\n");
294 return;
295 }
296 }
297 tk->addr = page_address_in_vma(p, vma);
298 tk->addr_valid = 1;
299
300 /*
301 * In theory we don't have to kill when the page was
302 * munmaped. But it could be also a mremap. Since that's
303 * likely very rare kill anyways just out of paranoia, but use
304 * a SIGKILL because the error is not contained anymore.
305 */
306 if (tk->addr == -EFAULT) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200307 pr_info("MCE: Unable to find user space address %lx in %s\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200308 page_to_pfn(p), tsk->comm);
309 tk->addr_valid = 0;
310 }
311 get_task_struct(tsk);
312 tk->tsk = tsk;
313 list_add_tail(&tk->nd, to_kill);
314}
315
316/*
317 * Kill the processes that have been collected earlier.
318 *
319 * Only do anything when DOIT is set, otherwise just free the list
320 * (this is used for clean pages which do not need killing)
321 * Also when FAIL is set do a force kill because something went
322 * wrong earlier.
323 */
Tony Luck6751ed62012-07-11 10:20:47 -0700324static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -0800325 int fail, struct page *page, unsigned long pfn,
326 int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200327{
328 struct to_kill *tk, *next;
329
330 list_for_each_entry_safe (tk, next, to_kill, nd) {
Tony Luck6751ed62012-07-11 10:20:47 -0700331 if (forcekill) {
Andi Kleen6a460792009-09-16 11:50:15 +0200332 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200333 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200334 * make sure the process doesn't catch the
335 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200336 */
337 if (fail || tk->addr_valid == 0) {
338 printk(KERN_ERR
339 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
340 pfn, tk->tsk->comm, tk->tsk->pid);
341 force_sig(SIGKILL, tk->tsk);
342 }
343
344 /*
345 * In theory the process could have mapped
346 * something else on the address in-between. We could
347 * check for that, but we need to tell the
348 * process anyways.
349 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800350 else if (kill_proc(tk->tsk, tk->addr, trapno,
351 pfn, page, flags) < 0)
Andi Kleen6a460792009-09-16 11:50:15 +0200352 printk(KERN_ERR
353 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
354 pfn, tk->tsk->comm, tk->tsk->pid);
355 }
356 put_task_struct(tk->tsk);
357 kfree(tk);
358 }
359}
360
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700361/*
362 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
363 * on behalf of the thread group. Return task_struct of the (first found)
364 * dedicated thread if found, and return NULL otherwise.
365 *
366 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
367 * have to call rcu_read_lock/unlock() in this function.
368 */
369static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
Andi Kleen6a460792009-09-16 11:50:15 +0200370{
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700371 struct task_struct *t;
372
373 for_each_thread(tsk, t)
374 if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
375 return t;
376 return NULL;
377}
378
379/*
380 * Determine whether a given process is "early kill" process which expects
381 * to be signaled when some page under the process is hwpoisoned.
382 * Return task_struct of the dedicated thread (main thread unless explicitly
383 * specified) if the process is "early kill," and otherwise returns NULL.
384 */
385static struct task_struct *task_early_kill(struct task_struct *tsk,
386 int force_early)
387{
388 struct task_struct *t;
Andi Kleen6a460792009-09-16 11:50:15 +0200389 if (!tsk->mm)
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700390 return NULL;
Tony Luck74614de2014-06-04 16:11:01 -0700391 if (force_early)
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700392 return tsk;
393 t = find_early_kill_thread(tsk);
394 if (t)
395 return t;
396 if (sysctl_memory_failure_early_kill)
397 return tsk;
398 return NULL;
Andi Kleen6a460792009-09-16 11:50:15 +0200399}
400
401/*
402 * Collect processes when the error hit an anonymous page.
403 */
404static void collect_procs_anon(struct page *page, struct list_head *to_kill,
Tony Luck74614de2014-06-04 16:11:01 -0700405 struct to_kill **tkc, int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200406{
407 struct vm_area_struct *vma;
408 struct task_struct *tsk;
409 struct anon_vma *av;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700410 pgoff_t pgoff;
Andi Kleen6a460792009-09-16 11:50:15 +0200411
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000412 av = page_lock_anon_vma_read(page);
Andi Kleen6a460792009-09-16 11:50:15 +0200413 if (av == NULL) /* Not actually mapped anymore */
Peter Zijlstra9b679322011-06-27 16:18:09 -0700414 return;
415
Naoya Horiguchia0f7a752014-07-23 14:00:01 -0700416 pgoff = page_to_pgoff(page);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700417 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200418 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800419 struct anon_vma_chain *vmac;
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700420 struct task_struct *t = task_early_kill(tsk, force_early);
Rik van Riel5beb4932010-03-05 13:42:07 -0800421
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700422 if (!t)
Andi Kleen6a460792009-09-16 11:50:15 +0200423 continue;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700424 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
425 pgoff, pgoff) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800426 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200427 if (!page_mapped_in_vma(page, vma))
428 continue;
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700429 if (vma->vm_mm == t->mm)
430 add_to_kill(t, page, vma, to_kill, tkc);
Andi Kleen6a460792009-09-16 11:50:15 +0200431 }
432 }
Andi Kleen6a460792009-09-16 11:50:15 +0200433 read_unlock(&tasklist_lock);
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000434 page_unlock_anon_vma_read(av);
Andi Kleen6a460792009-09-16 11:50:15 +0200435}
436
437/*
438 * Collect processes when the error hit a file mapped page.
439 */
440static void collect_procs_file(struct page *page, struct list_head *to_kill,
Tony Luck74614de2014-06-04 16:11:01 -0700441 struct to_kill **tkc, int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200442{
443 struct vm_area_struct *vma;
444 struct task_struct *tsk;
Andi Kleen6a460792009-09-16 11:50:15 +0200445 struct address_space *mapping = page->mapping;
446
Davidlohr Buesod28eb9c2014-12-12 16:54:36 -0800447 i_mmap_lock_read(mapping);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700448 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200449 for_each_process(tsk) {
Naoya Horiguchia0f7a752014-07-23 14:00:01 -0700450 pgoff_t pgoff = page_to_pgoff(page);
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700451 struct task_struct *t = task_early_kill(tsk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200452
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700453 if (!t)
Andi Kleen6a460792009-09-16 11:50:15 +0200454 continue;
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -0700455 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
Andi Kleen6a460792009-09-16 11:50:15 +0200456 pgoff) {
457 /*
458 * Send early kill signal to tasks where a vma covers
459 * the page but the corrupted page is not necessarily
460 * mapped it in its pte.
461 * Assume applications who requested early kill want
462 * to be informed of all such data corruptions.
463 */
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700464 if (vma->vm_mm == t->mm)
465 add_to_kill(t, page, vma, to_kill, tkc);
Andi Kleen6a460792009-09-16 11:50:15 +0200466 }
467 }
Andi Kleen6a460792009-09-16 11:50:15 +0200468 read_unlock(&tasklist_lock);
Davidlohr Buesod28eb9c2014-12-12 16:54:36 -0800469 i_mmap_unlock_read(mapping);
Andi Kleen6a460792009-09-16 11:50:15 +0200470}
471
472/*
473 * Collect the processes who have the corrupted page mapped to kill.
474 * This is done in two steps for locking reasons.
475 * First preallocate one tokill structure outside the spin locks,
476 * so that we can kill at least one process reasonably reliable.
477 */
Tony Luck74614de2014-06-04 16:11:01 -0700478static void collect_procs(struct page *page, struct list_head *tokill,
479 int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200480{
481 struct to_kill *tk;
482
483 if (!page->mapping)
484 return;
485
486 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
487 if (!tk)
488 return;
489 if (PageAnon(page))
Tony Luck74614de2014-06-04 16:11:01 -0700490 collect_procs_anon(page, tokill, &tk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200491 else
Tony Luck74614de2014-06-04 16:11:01 -0700492 collect_procs_file(page, tokill, &tk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200493 kfree(tk);
494}
495
Andi Kleen6a460792009-09-16 11:50:15 +0200496static const char *action_name[] = {
Xie XiuQicc637b12015-06-24 16:57:30 -0700497 [MF_IGNORED] = "Ignored",
498 [MF_FAILED] = "Failed",
499 [MF_DELAYED] = "Delayed",
500 [MF_RECOVERED] = "Recovered",
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700501};
502
503static const char * const action_page_types[] = {
Xie XiuQicc637b12015-06-24 16:57:30 -0700504 [MF_MSG_KERNEL] = "reserved kernel page",
505 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
506 [MF_MSG_SLAB] = "kernel slab page",
507 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
508 [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned",
509 [MF_MSG_HUGE] = "huge page",
510 [MF_MSG_FREE_HUGE] = "free huge page",
511 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
512 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
513 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
514 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
515 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
516 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
517 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
518 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
519 [MF_MSG_CLEAN_LRU] = "clean LRU page",
520 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
521 [MF_MSG_BUDDY] = "free buddy page",
522 [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)",
523 [MF_MSG_UNKNOWN] = "unknown page",
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700524};
525
Andi Kleen6a460792009-09-16 11:50:15 +0200526/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100527 * XXX: It is possible that a page is isolated from LRU cache,
528 * and then kept in swap cache or failed to remove from page cache.
529 * The page count will stop it from being freed by unpoison.
530 * Stress tests should be aware of this memory leak problem.
531 */
532static int delete_from_lru_cache(struct page *p)
533{
534 if (!isolate_lru_page(p)) {
535 /*
536 * Clear sensible page flags, so that the buddy system won't
537 * complain when the page is unpoison-and-freed.
538 */
539 ClearPageActive(p);
540 ClearPageUnevictable(p);
541 /*
542 * drop the page count elevated by isolate_lru_page()
543 */
544 page_cache_release(p);
545 return 0;
546 }
547 return -EIO;
548}
549
550/*
Andi Kleen6a460792009-09-16 11:50:15 +0200551 * Error hit kernel page.
552 * Do nothing, try to be lucky and not touch this instead. For a few cases we
553 * could be more sophisticated.
554 */
555static int me_kernel(struct page *p, unsigned long pfn)
556{
Xie XiuQicc637b12015-06-24 16:57:30 -0700557 return MF_IGNORED;
Andi Kleen6a460792009-09-16 11:50:15 +0200558}
559
560/*
561 * Page in unknown state. Do nothing.
562 */
563static int me_unknown(struct page *p, unsigned long pfn)
564{
565 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
Xie XiuQicc637b12015-06-24 16:57:30 -0700566 return MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200567}
568
569/*
Andi Kleen6a460792009-09-16 11:50:15 +0200570 * Clean (or cleaned) page cache page.
571 */
572static int me_pagecache_clean(struct page *p, unsigned long pfn)
573{
574 int err;
Xie XiuQicc637b12015-06-24 16:57:30 -0700575 int ret = MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200576 struct address_space *mapping;
577
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100578 delete_from_lru_cache(p);
579
Andi Kleen6a460792009-09-16 11:50:15 +0200580 /*
581 * For anonymous pages we're done the only reference left
582 * should be the one m_f() holds.
583 */
584 if (PageAnon(p))
Xie XiuQicc637b12015-06-24 16:57:30 -0700585 return MF_RECOVERED;
Andi Kleen6a460792009-09-16 11:50:15 +0200586
587 /*
588 * Now truncate the page in the page cache. This is really
589 * more like a "temporary hole punch"
590 * Don't do this for block devices when someone else
591 * has a reference, because it could be file system metadata
592 * and that's not safe to truncate.
593 */
594 mapping = page_mapping(p);
595 if (!mapping) {
596 /*
597 * Page has been teared down in the meanwhile
598 */
Xie XiuQicc637b12015-06-24 16:57:30 -0700599 return MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200600 }
601
602 /*
603 * Truncation is a bit tricky. Enable it per file system for now.
604 *
605 * Open: to take i_mutex or not for this? Right now we don't.
606 */
607 if (mapping->a_ops->error_remove_page) {
608 err = mapping->a_ops->error_remove_page(mapping, p);
609 if (err != 0) {
610 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
611 pfn, err);
612 } else if (page_has_private(p) &&
613 !try_to_release_page(p, GFP_NOIO)) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200614 pr_info("MCE %#lx: failed to release buffers\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200615 } else {
Xie XiuQicc637b12015-06-24 16:57:30 -0700616 ret = MF_RECOVERED;
Andi Kleen6a460792009-09-16 11:50:15 +0200617 }
618 } else {
619 /*
620 * If the file system doesn't support it just invalidate
621 * This fails on dirty or anything with private pages
622 */
623 if (invalidate_inode_page(p))
Xie XiuQicc637b12015-06-24 16:57:30 -0700624 ret = MF_RECOVERED;
Andi Kleen6a460792009-09-16 11:50:15 +0200625 else
626 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
627 pfn);
628 }
629 return ret;
630}
631
632/*
Zhi Yong Wu549543d2014-01-21 15:49:08 -0800633 * Dirty pagecache page
Andi Kleen6a460792009-09-16 11:50:15 +0200634 * Issues: when the error hit a hole page the error is not properly
635 * propagated.
636 */
637static int me_pagecache_dirty(struct page *p, unsigned long pfn)
638{
639 struct address_space *mapping = page_mapping(p);
640
641 SetPageError(p);
642 /* TBD: print more information about the file. */
643 if (mapping) {
644 /*
645 * IO error will be reported by write(), fsync(), etc.
646 * who check the mapping.
647 * This way the application knows that something went
648 * wrong with its dirty file data.
649 *
650 * There's one open issue:
651 *
652 * The EIO will be only reported on the next IO
653 * operation and then cleared through the IO map.
654 * Normally Linux has two mechanisms to pass IO error
655 * first through the AS_EIO flag in the address space
656 * and then through the PageError flag in the page.
657 * Since we drop pages on memory failure handling the
658 * only mechanism open to use is through AS_AIO.
659 *
660 * This has the disadvantage that it gets cleared on
661 * the first operation that returns an error, while
662 * the PageError bit is more sticky and only cleared
663 * when the page is reread or dropped. If an
664 * application assumes it will always get error on
665 * fsync, but does other operations on the fd before
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300666 * and the page is dropped between then the error
Andi Kleen6a460792009-09-16 11:50:15 +0200667 * will not be properly reported.
668 *
669 * This can already happen even without hwpoisoned
670 * pages: first on metadata IO errors (which only
671 * report through AS_EIO) or when the page is dropped
672 * at the wrong time.
673 *
674 * So right now we assume that the application DTRT on
675 * the first EIO, but we're not worse than other parts
676 * of the kernel.
677 */
678 mapping_set_error(mapping, EIO);
679 }
680
681 return me_pagecache_clean(p, pfn);
682}
683
684/*
685 * Clean and dirty swap cache.
686 *
687 * Dirty swap cache page is tricky to handle. The page could live both in page
688 * cache and swap cache(ie. page is freshly swapped in). So it could be
689 * referenced concurrently by 2 types of PTEs:
690 * normal PTEs and swap PTEs. We try to handle them consistently by calling
691 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
692 * and then
693 * - clear dirty bit to prevent IO
694 * - remove from LRU
695 * - but keep in the swap cache, so that when we return to it on
696 * a later page fault, we know the application is accessing
697 * corrupted data and shall be killed (we installed simple
698 * interception code in do_swap_page to catch it).
699 *
700 * Clean swap cache pages can be directly isolated. A later page fault will
701 * bring in the known good data from disk.
702 */
703static int me_swapcache_dirty(struct page *p, unsigned long pfn)
704{
Andi Kleen6a460792009-09-16 11:50:15 +0200705 ClearPageDirty(p);
706 /* Trigger EIO in shmem: */
707 ClearPageUptodate(p);
708
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100709 if (!delete_from_lru_cache(p))
Xie XiuQicc637b12015-06-24 16:57:30 -0700710 return MF_DELAYED;
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100711 else
Xie XiuQicc637b12015-06-24 16:57:30 -0700712 return MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200713}
714
715static int me_swapcache_clean(struct page *p, unsigned long pfn)
716{
Andi Kleen6a460792009-09-16 11:50:15 +0200717 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800718
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100719 if (!delete_from_lru_cache(p))
Xie XiuQicc637b12015-06-24 16:57:30 -0700720 return MF_RECOVERED;
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100721 else
Xie XiuQicc637b12015-06-24 16:57:30 -0700722 return MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200723}
724
725/*
726 * Huge pages. Needs work.
727 * Issues:
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900728 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
729 * To narrow down kill region to one page, we need to break up pmd.
Andi Kleen6a460792009-09-16 11:50:15 +0200730 */
731static int me_huge_page(struct page *p, unsigned long pfn)
732{
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900733 int res = 0;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900734 struct page *hpage = compound_head(p);
Naoya Horiguchi2491ffe2015-06-24 16:56:53 -0700735
736 if (!PageHuge(hpage))
737 return MF_DELAYED;
738
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900739 /*
740 * We can safely recover from error on free or reserved (i.e.
741 * not in-use) hugepage by dequeuing it from freelist.
742 * To check whether a hugepage is in-use or not, we can't use
743 * page->lru because it can be used in other hugepage operations,
744 * such as __unmap_hugepage_range() and gather_surplus_pages().
745 * So instead we use page_mapping() and PageAnon().
746 * We assume that this function is called with page lock held,
747 * so there is no race between isolation and mapping/unmapping.
748 */
749 if (!(page_mapping(hpage) || PageAnon(hpage))) {
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900750 res = dequeue_hwpoisoned_huge_page(hpage);
751 if (!res)
Xie XiuQicc637b12015-06-24 16:57:30 -0700752 return MF_RECOVERED;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900753 }
Xie XiuQicc637b12015-06-24 16:57:30 -0700754 return MF_DELAYED;
Andi Kleen6a460792009-09-16 11:50:15 +0200755}
756
757/*
758 * Various page states we can handle.
759 *
760 * A page state is defined by its current page->flags bits.
761 * The table matches them in order and calls the right handler.
762 *
763 * This is quite tricky because we can access page at any time
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300764 * in its live cycle, so all accesses have to be extremely careful.
Andi Kleen6a460792009-09-16 11:50:15 +0200765 *
766 * This is not complete. More states could be added.
767 * For any missing state don't attempt recovery.
768 */
769
770#define dirty (1UL << PG_dirty)
771#define sc (1UL << PG_swapcache)
772#define unevict (1UL << PG_unevictable)
773#define mlock (1UL << PG_mlocked)
774#define writeback (1UL << PG_writeback)
775#define lru (1UL << PG_lru)
776#define swapbacked (1UL << PG_swapbacked)
777#define head (1UL << PG_head)
778#define tail (1UL << PG_tail)
779#define compound (1UL << PG_compound)
780#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200781#define reserved (1UL << PG_reserved)
782
783static struct page_state {
784 unsigned long mask;
785 unsigned long res;
Xie XiuQicc637b12015-06-24 16:57:30 -0700786 enum mf_action_page_type type;
Andi Kleen6a460792009-09-16 11:50:15 +0200787 int (*action)(struct page *p, unsigned long pfn);
788} error_states[] = {
Xie XiuQicc637b12015-06-24 16:57:30 -0700789 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100790 /*
791 * free pages are specially detected outside this table:
792 * PG_buddy pages only make a small fraction of all free pages.
793 */
Andi Kleen6a460792009-09-16 11:50:15 +0200794
795 /*
796 * Could in theory check if slab page is free or if we can drop
797 * currently unused objects without touching them. But just
798 * treat it as standard kernel for now.
799 */
Xie XiuQicc637b12015-06-24 16:57:30 -0700800 { slab, slab, MF_MSG_SLAB, me_kernel },
Andi Kleen6a460792009-09-16 11:50:15 +0200801
802#ifdef CONFIG_PAGEFLAGS_EXTENDED
Xie XiuQicc637b12015-06-24 16:57:30 -0700803 { head, head, MF_MSG_HUGE, me_huge_page },
804 { tail, tail, MF_MSG_HUGE, me_huge_page },
Andi Kleen6a460792009-09-16 11:50:15 +0200805#else
Xie XiuQicc637b12015-06-24 16:57:30 -0700806 { compound, compound, MF_MSG_HUGE, me_huge_page },
Andi Kleen6a460792009-09-16 11:50:15 +0200807#endif
808
Xie XiuQicc637b12015-06-24 16:57:30 -0700809 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
810 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200811
Xie XiuQicc637b12015-06-24 16:57:30 -0700812 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
813 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200814
Xie XiuQicc637b12015-06-24 16:57:30 -0700815 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
816 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
Naoya Horiguchi5f4b9fc2013-02-22 16:35:53 -0800817
Xie XiuQicc637b12015-06-24 16:57:30 -0700818 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
819 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200820
821 /*
822 * Catchall entry: must be at end.
823 */
Xie XiuQicc637b12015-06-24 16:57:30 -0700824 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
Andi Kleen6a460792009-09-16 11:50:15 +0200825};
826
Andi Kleen2326c462009-12-16 12:20:00 +0100827#undef dirty
828#undef sc
829#undef unevict
830#undef mlock
831#undef writeback
832#undef lru
833#undef swapbacked
834#undef head
835#undef tail
836#undef compound
837#undef slab
838#undef reserved
839
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800840/*
841 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
842 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
843 */
Xie XiuQicc3e2af2015-06-24 16:57:33 -0700844static void action_result(unsigned long pfn, enum mf_action_page_type type,
845 enum mf_result result)
Andi Kleen6a460792009-09-16 11:50:15 +0200846{
Xie XiuQi97f0b132015-06-24 16:57:36 -0700847 trace_memory_failure_event(pfn, type, result);
848
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700849 pr_err("MCE %#lx: recovery action for %s: %s\n",
850 pfn, action_page_types[type], action_name[result]);
Andi Kleen6a460792009-09-16 11:50:15 +0200851}
852
853static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100854 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200855{
856 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200857 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200858
859 result = ps->action(p, pfn);
Wu Fengguang7456b042009-10-19 08:15:01 +0200860
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100861 count = page_count(p) - 1;
Xie XiuQicc637b12015-06-24 16:57:30 -0700862 if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
Wu Fengguang138ce282009-12-16 12:19:58 +0100863 count--;
864 if (count != 0) {
Andi Kleen6a460792009-09-16 11:50:15 +0200865 printk(KERN_ERR
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700866 "MCE %#lx: %s still referenced by %d users\n",
867 pfn, action_page_types[ps->type], count);
Xie XiuQicc637b12015-06-24 16:57:30 -0700868 result = MF_FAILED;
Wu Fengguang138ce282009-12-16 12:19:58 +0100869 }
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700870 action_result(pfn, ps->type, result);
Andi Kleen6a460792009-09-16 11:50:15 +0200871
872 /* Could do more checks here if page looks ok */
873 /*
874 * Could adjust zone counters here to correct for the missing page.
875 */
876
Xie XiuQicc637b12015-06-24 16:57:30 -0700877 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200878}
879
Naoya Horiguchiead07f62015-06-24 16:56:48 -0700880/**
881 * get_hwpoison_page() - Get refcount for memory error handling:
882 * @page: raw error page (hit by memory error)
883 *
884 * Return: return 0 if failed to grab the refcount, otherwise true (some
885 * non-zero value.)
886 */
887int get_hwpoison_page(struct page *page)
888{
889 struct page *head = compound_head(page);
890
891 if (PageHuge(head))
892 return get_page_unless_zero(head);
893
894 /*
895 * Thp tail page has special refcounting rule (refcount of tail pages
896 * is stored in ->_mapcount,) so we can't call get_page_unless_zero()
897 * directly for tail pages.
898 */
899 if (PageTransHuge(head)) {
Naoya Horiguchi98ed2b02015-08-06 15:47:04 -0700900 /*
901 * Non anonymous thp exists only in allocation/free time. We
902 * can't handle such a case correctly, so let's give it up.
903 * This should be better than triggering BUG_ON when kernel
904 * tries to touch the "partially handled" page.
905 */
906 if (!PageAnon(head)) {
907 pr_err("MCE: %#lx: non anonymous thp\n",
908 page_to_pfn(page));
909 return 0;
910 }
911
Naoya Horiguchiead07f62015-06-24 16:56:48 -0700912 if (get_page_unless_zero(head)) {
913 if (PageTail(page))
914 get_page(page);
915 return 1;
916 } else {
917 return 0;
918 }
919 }
920
921 return get_page_unless_zero(page);
922}
923EXPORT_SYMBOL_GPL(get_hwpoison_page);
924
Wanpeng Li94bf4ec2015-09-08 15:03:15 -0700925/**
926 * put_hwpoison_page() - Put refcount for memory error handling:
927 * @page: raw error page (hit by memory error)
928 */
929void put_hwpoison_page(struct page *page)
930{
931 struct page *head = compound_head(page);
932
933 if (PageHuge(head)) {
934 put_page(head);
935 return;
936 }
937
938 if (PageTransHuge(head))
939 if (page != head)
940 put_page(head);
941
942 put_page(page);
943}
944EXPORT_SYMBOL_GPL(put_hwpoison_page);
945
Andi Kleen6a460792009-09-16 11:50:15 +0200946/*
947 * Do all that is necessary to remove user space mappings. Unmap
948 * the pages and send SIGBUS to the processes if the data was dirty.
949 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100950static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800951 int trapno, int flags, struct page **hpagep)
Andi Kleen6a460792009-09-16 11:50:15 +0200952{
953 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
954 struct address_space *mapping;
955 LIST_HEAD(tokill);
956 int ret;
Tony Luck6751ed62012-07-11 10:20:47 -0700957 int kill = 1, forcekill;
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800958 struct page *hpage = *hpagep;
Andi Kleen6a460792009-09-16 11:50:15 +0200959
Naoya Horiguchi93a9eb32014-07-30 16:08:28 -0700960 /*
961 * Here we are interested only in user-mapped pages, so skip any
962 * other types of pages.
963 */
964 if (PageReserved(p) || PageSlab(p))
965 return SWAP_SUCCESS;
966 if (!(PageLRU(hpage) || PageHuge(p)))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100967 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200968
Andi Kleen6a460792009-09-16 11:50:15 +0200969 /*
970 * This check implies we don't kill processes if their pages
971 * are in the swap cache early. Those are always late kills.
972 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900973 if (!page_mapped(hpage))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100974 return SWAP_SUCCESS;
975
Naoya Horiguchi52089b12014-07-30 16:08:30 -0700976 if (PageKsm(p)) {
977 pr_err("MCE %#lx: can't handle KSM pages.\n", pfn);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100978 return SWAP_FAIL;
Naoya Horiguchi52089b12014-07-30 16:08:30 -0700979 }
Andi Kleen6a460792009-09-16 11:50:15 +0200980
981 if (PageSwapCache(p)) {
982 printk(KERN_ERR
983 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
984 ttu |= TTU_IGNORE_HWPOISON;
985 }
986
987 /*
988 * Propagate the dirty bit from PTEs to struct page first, because we
989 * need this to decide if we should kill or just drop the page.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100990 * XXX: the dirty test could be racy: set_page_dirty() may not always
991 * be called inside page lock (it's recommended but not enforced).
Andi Kleen6a460792009-09-16 11:50:15 +0200992 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900993 mapping = page_mapping(hpage);
Tony Luck6751ed62012-07-11 10:20:47 -0700994 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900995 mapping_cap_writeback_dirty(mapping)) {
996 if (page_mkclean(hpage)) {
997 SetPageDirty(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +0200998 } else {
999 kill = 0;
1000 ttu |= TTU_IGNORE_HWPOISON;
1001 printk(KERN_INFO
1002 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
1003 pfn);
1004 }
1005 }
1006
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001007 /*
Andi Kleen6a460792009-09-16 11:50:15 +02001008 * First collect all the processes that have the page
1009 * mapped in dirty form. This has to be done before try_to_unmap,
1010 * because ttu takes the rmap data structures down.
1011 *
1012 * Error handling: We ignore errors here because
1013 * there's nothing that can be done.
1014 */
1015 if (kill)
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001016 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
Andi Kleen6a460792009-09-16 11:50:15 +02001017
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001018 ret = try_to_unmap(hpage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +02001019 if (ret != SWAP_SUCCESS)
1020 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001021 pfn, page_mapcount(hpage));
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001022
Andi Kleen6a460792009-09-16 11:50:15 +02001023 /*
1024 * Now that the dirty bit has been propagated to the
1025 * struct page and all unmaps done we can decide if
1026 * killing is needed or not. Only kill when the page
Tony Luck6751ed62012-07-11 10:20:47 -07001027 * was dirty or the process is not restartable,
1028 * otherwise the tokill list is merely
Andi Kleen6a460792009-09-16 11:50:15 +02001029 * freed. When there was a problem unmapping earlier
1030 * use a more force-full uncatchable kill to prevent
1031 * any accesses to the poisoned memory.
1032 */
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001033 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
Tony Luck6751ed62012-07-11 10:20:47 -07001034 kill_procs(&tokill, forcekill, trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -08001035 ret != SWAP_SUCCESS, p, pfn, flags);
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001036
1037 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +02001038}
1039
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001040static void set_page_hwpoison_huge_page(struct page *hpage)
1041{
1042 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -07001043 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001044 for (i = 0; i < nr_pages; i++)
1045 SetPageHWPoison(hpage + i);
1046}
1047
1048static void clear_page_hwpoison_huge_page(struct page *hpage)
1049{
1050 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -07001051 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001052 for (i = 0; i < nr_pages; i++)
1053 ClearPageHWPoison(hpage + i);
1054}
1055
Tony Luckcd42f4a2011-12-15 10:48:12 -08001056/**
1057 * memory_failure - Handle memory failure of a page.
1058 * @pfn: Page Number of the corrupted page
1059 * @trapno: Trap number reported in the signal to user space.
1060 * @flags: fine tune action taken
1061 *
1062 * This function is called by the low level machine check code
1063 * of an architecture when it detects hardware memory corruption
1064 * of a page. It tries its best to recover, which includes
1065 * dropping pages, killing processes etc.
1066 *
1067 * The function is primarily of use for corruptions that
1068 * happen outside the current execution context (e.g. when
1069 * detected by a background scrubber)
1070 *
1071 * Must run in process context (e.g. a work queue) with interrupts
1072 * enabled and no spinlocks hold.
1073 */
1074int memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +02001075{
1076 struct page_state *ps;
1077 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001078 struct page *hpage;
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001079 struct page *orig_head;
Andi Kleen6a460792009-09-16 11:50:15 +02001080 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001081 unsigned int nr_pages;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001082 unsigned long page_flags;
Andi Kleen6a460792009-09-16 11:50:15 +02001083
1084 if (!sysctl_memory_failure_recovery)
1085 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1086
1087 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +01001088 printk(KERN_ERR
1089 "MCE %#lx: memory outside kernel control\n",
1090 pfn);
1091 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +02001092 }
1093
1094 p = pfn_to_page(pfn);
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001095 orig_head = hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +02001096 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001097 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001098 return 0;
1099 }
1100
Naoya Horiguchi4db0e952013-02-22 16:34:05 -08001101 /*
1102 * Currently errors on hugetlbfs pages are measured in hugepage units,
1103 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1104 * transparent hugepages, they are supposed to be split and error
1105 * measurement is done in normal page units. So nr_pages should be one
1106 * in this case.
1107 */
1108 if (PageHuge(p))
1109 nr_pages = 1 << compound_order(hpage);
1110 else /* normal page or thp */
1111 nr_pages = 1;
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001112 num_poisoned_pages_add(nr_pages);
Andi Kleen6a460792009-09-16 11:50:15 +02001113
1114 /*
1115 * We need/can do nothing about count=0 pages.
1116 * 1) it's a free page, and therefore in safe hand:
1117 * prep_new_page() will be the gate keeper.
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001118 * 2) it's a free hugepage, which is also safe:
1119 * an affected hugepage will be dequeued from hugepage freelist,
1120 * so there's no concern about reusing it ever after.
1121 * 3) it's part of a non-compound high order page.
Andi Kleen6a460792009-09-16 11:50:15 +02001122 * Implies some kernel user: cannot stop them from
1123 * R/W the page; let's pray that the page has been
1124 * used and will be freed some time later.
1125 * In fact it's dangerous to directly bump up page count from 0,
1126 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1127 */
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001128 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001129 if (is_free_buddy_page(p)) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001130 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001131 return 0;
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001132 } else if (PageHuge(hpage)) {
1133 /*
Chen Yucongb9851942014-05-22 11:54:15 -07001134 * Check "filter hit" and "race with other subpage."
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001135 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001136 lock_page(hpage);
Chen Yucongb9851942014-05-22 11:54:15 -07001137 if (PageHWPoison(hpage)) {
1138 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1139 || (p != hpage && TestSetPageHWPoison(hpage))) {
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001140 num_poisoned_pages_sub(nr_pages);
Chen Yucongb9851942014-05-22 11:54:15 -07001141 unlock_page(hpage);
1142 return 0;
1143 }
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001144 }
1145 set_page_hwpoison_huge_page(hpage);
1146 res = dequeue_hwpoisoned_huge_page(hpage);
Xie XiuQicc637b12015-06-24 16:57:30 -07001147 action_result(pfn, MF_MSG_FREE_HUGE,
1148 res ? MF_IGNORED : MF_DELAYED);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001149 unlock_page(hpage);
1150 return res;
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001151 } else {
Xie XiuQicc637b12015-06-24 16:57:30 -07001152 action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001153 return -EBUSY;
1154 }
Andi Kleen6a460792009-09-16 11:50:15 +02001155 }
1156
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001157 if (!PageHuge(p) && PageTransHuge(hpage)) {
Wanpeng Li7f6bf392015-08-14 15:35:08 -07001158 if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) {
1159 if (!PageAnon(hpage))
1160 pr_err("MCE: %#lx: non anonymous thp\n", pfn);
1161 else
1162 pr_err("MCE: %#lx: thp split failed\n", pfn);
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001163 if (TestClearPageHWPoison(p))
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001164 num_poisoned_pages_sub(nr_pages);
Wanpeng Li665d9da2015-09-08 15:03:21 -07001165 put_hwpoison_page(p);
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001166 return -EBUSY;
1167 }
1168 VM_BUG_ON_PAGE(!page_count(p), p);
1169 hpage = compound_head(p);
1170 }
1171
Andi Kleen6a460792009-09-16 11:50:15 +02001172 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001173 * We ignore non-LRU pages for good reasons.
1174 * - PG_locked is only well defined for LRU pages and a few others
1175 * - to avoid races with __set_page_locked()
1176 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1177 * The check (unnecessarily) ignores LRU pages being isolated and
1178 * walked by the page reclaim code, however that's not a big loss.
1179 */
Naoya Horiguchi09789e52015-05-05 16:23:35 -07001180 if (!PageHuge(p)) {
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001181 if (!PageLRU(p))
1182 shake_page(p, 0);
1183 if (!PageLRU(p)) {
Jin Dongmingaf241a02011-02-01 15:52:41 -08001184 /*
1185 * shake_page could have turned it free.
1186 */
1187 if (is_free_buddy_page(p)) {
Wanpeng Li2d421ac2013-09-30 13:45:23 -07001188 if (flags & MF_COUNT_INCREASED)
Xie XiuQicc637b12015-06-24 16:57:30 -07001189 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
Wanpeng Li2d421ac2013-09-30 13:45:23 -07001190 else
Xie XiuQicc637b12015-06-24 16:57:30 -07001191 action_result(pfn, MF_MSG_BUDDY_2ND,
1192 MF_DELAYED);
Jin Dongmingaf241a02011-02-01 15:52:41 -08001193 return 0;
1194 }
Andi Kleen0474a602009-12-16 12:20:00 +01001195 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001196 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001197
Jens Axboe7eaceac2011-03-10 08:52:07 +01001198 lock_page(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001199
1200 /*
Andi Kleenf37d4292014-08-06 16:06:49 -07001201 * The page could have changed compound pages during the locking.
1202 * If this happens just bail out.
1203 */
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001204 if (PageCompound(p) && compound_head(p) != orig_head) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001205 action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
Andi Kleenf37d4292014-08-06 16:06:49 -07001206 res = -EBUSY;
1207 goto out;
1208 }
1209
1210 /*
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001211 * We use page flags to determine what action should be taken, but
1212 * the flags can be modified by the error containment action. One
1213 * example is an mlocked page, where PG_mlocked is cleared by
1214 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1215 * correctly, we save a copy of the page flags at this time.
1216 */
1217 page_flags = p->flags;
1218
1219 /*
Wu Fengguang847ce402009-12-16 12:19:58 +01001220 * unpoison always clear PG_hwpoison inside page lock
1221 */
1222 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001223 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001224 num_poisoned_pages_sub(nr_pages);
Naoya Horiguchia09233f2015-08-06 15:46:58 -07001225 unlock_page(hpage);
Wanpeng Li665d9da2015-09-08 15:03:21 -07001226 put_hwpoison_page(hpage);
Naoya Horiguchia09233f2015-08-06 15:46:58 -07001227 return 0;
Wu Fengguang847ce402009-12-16 12:19:58 +01001228 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001229 if (hwpoison_filter(p)) {
1230 if (TestClearPageHWPoison(p))
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001231 num_poisoned_pages_sub(nr_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001232 unlock_page(hpage);
Wanpeng Li665d9da2015-09-08 15:03:21 -07001233 put_hwpoison_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001234 return 0;
1235 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001236
Chen Yucong0bc1f8b2014-07-02 15:22:37 -07001237 if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p))
1238 goto identify_page_state;
1239
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001240 /*
1241 * For error on the tail page, we should set PG_hwpoison
1242 * on the head page to show that the hugepage is hwpoisoned
1243 */
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001244 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001245 action_result(pfn, MF_MSG_POISONED_HUGE, MF_IGNORED);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001246 unlock_page(hpage);
Wanpeng Li665d9da2015-09-08 15:03:21 -07001247 put_hwpoison_page(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001248 return 0;
1249 }
1250 /*
1251 * Set PG_hwpoison on all pages in an error hugepage,
1252 * because containment is done in hugepage unit for now.
1253 * Since we have done TestSetPageHWPoison() for the head page with
1254 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1255 */
1256 if (PageHuge(p))
1257 set_page_hwpoison_huge_page(hpage);
1258
Naoya Horiguchi6edd6cc2014-06-04 16:10:35 -07001259 /*
1260 * It's very difficult to mess with pages currently under IO
1261 * and in many cases impossible, so we just avoid it here.
1262 */
Andi Kleen6a460792009-09-16 11:50:15 +02001263 wait_on_page_writeback(p);
1264
1265 /*
1266 * Now take care of user space mappings.
Minchan Kime64a7822011-03-22 16:32:44 -07001267 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001268 *
1269 * When the raw error page is thp tail page, hpage points to the raw
1270 * page after thp split.
Andi Kleen6a460792009-09-16 11:50:15 +02001271 */
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001272 if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)
1273 != SWAP_SUCCESS) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001274 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001275 res = -EBUSY;
1276 goto out;
1277 }
Andi Kleen6a460792009-09-16 11:50:15 +02001278
1279 /*
1280 * Torn down by someone else?
1281 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001282 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001283 action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001284 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001285 goto out;
1286 }
1287
Chen Yucong0bc1f8b2014-07-02 15:22:37 -07001288identify_page_state:
Andi Kleen6a460792009-09-16 11:50:15 +02001289 res = -EBUSY;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001290 /*
1291 * The first check uses the current page flags which may not have any
1292 * relevant information. The second check with the saved page flagss is
1293 * carried out only if the first check can't determine the page status.
1294 */
1295 for (ps = error_states;; ps++)
1296 if ((p->flags & ps->mask) == ps->res)
Andi Kleen6a460792009-09-16 11:50:15 +02001297 break;
Wanpeng Li841fcc52013-09-11 14:22:50 -07001298
1299 page_flags |= (p->flags & (1UL << PG_dirty));
1300
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001301 if (!ps->mask)
1302 for (ps = error_states;; ps++)
1303 if ((page_flags & ps->mask) == ps->res)
1304 break;
1305 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001306out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001307 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001308 return res;
1309}
Tony Luckcd42f4a2011-12-15 10:48:12 -08001310EXPORT_SYMBOL_GPL(memory_failure);
Wu Fengguang847ce402009-12-16 12:19:58 +01001311
Huang Yingea8f5fb2011-07-13 13:14:27 +08001312#define MEMORY_FAILURE_FIFO_ORDER 4
1313#define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1314
1315struct memory_failure_entry {
1316 unsigned long pfn;
1317 int trapno;
1318 int flags;
1319};
1320
1321struct memory_failure_cpu {
1322 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1323 MEMORY_FAILURE_FIFO_SIZE);
1324 spinlock_t lock;
1325 struct work_struct work;
1326};
1327
1328static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1329
1330/**
1331 * memory_failure_queue - Schedule handling memory failure of a page.
1332 * @pfn: Page Number of the corrupted page
1333 * @trapno: Trap number reported in the signal to user space.
1334 * @flags: Flags for memory failure handling
1335 *
1336 * This function is called by the low level hardware error handler
1337 * when it detects hardware memory corruption of a page. It schedules
1338 * the recovering of error page, including dropping pages, killing
1339 * processes etc.
1340 *
1341 * The function is primarily of use for corruptions that
1342 * happen outside the current execution context (e.g. when
1343 * detected by a background scrubber)
1344 *
1345 * Can run in IRQ context.
1346 */
1347void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1348{
1349 struct memory_failure_cpu *mf_cpu;
1350 unsigned long proc_flags;
1351 struct memory_failure_entry entry = {
1352 .pfn = pfn,
1353 .trapno = trapno,
1354 .flags = flags,
1355 };
1356
1357 mf_cpu = &get_cpu_var(memory_failure_cpu);
1358 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
Stefani Seibold498d3192013-11-14 14:32:17 -08001359 if (kfifo_put(&mf_cpu->fifo, entry))
Huang Yingea8f5fb2011-07-13 13:14:27 +08001360 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1361 else
Joe Perches8e33a522013-07-25 11:53:25 -07001362 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
Huang Yingea8f5fb2011-07-13 13:14:27 +08001363 pfn);
1364 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1365 put_cpu_var(memory_failure_cpu);
1366}
1367EXPORT_SYMBOL_GPL(memory_failure_queue);
1368
1369static void memory_failure_work_func(struct work_struct *work)
1370{
1371 struct memory_failure_cpu *mf_cpu;
1372 struct memory_failure_entry entry = { 0, };
1373 unsigned long proc_flags;
1374 int gotten;
1375
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001376 mf_cpu = this_cpu_ptr(&memory_failure_cpu);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001377 for (;;) {
1378 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1379 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1380 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1381 if (!gotten)
1382 break;
Naveen N. Raocf870c72013-07-10 14:57:01 +05301383 if (entry.flags & MF_SOFT_OFFLINE)
1384 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1385 else
1386 memory_failure(entry.pfn, entry.trapno, entry.flags);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001387 }
1388}
1389
1390static int __init memory_failure_init(void)
1391{
1392 struct memory_failure_cpu *mf_cpu;
1393 int cpu;
1394
1395 for_each_possible_cpu(cpu) {
1396 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1397 spin_lock_init(&mf_cpu->lock);
1398 INIT_KFIFO(mf_cpu->fifo);
1399 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1400 }
1401
1402 return 0;
1403}
1404core_initcall(memory_failure_init);
1405
Wu Fengguang847ce402009-12-16 12:19:58 +01001406/**
1407 * unpoison_memory - Unpoison a previously poisoned page
1408 * @pfn: Page number of the to be unpoisoned page
1409 *
1410 * Software-unpoison a page that has been poisoned by
1411 * memory_failure() earlier.
1412 *
1413 * This is only done on the software-level, so it only works
1414 * for linux injected failures, not real hardware failures
1415 *
1416 * Returns 0 for success, otherwise -errno.
1417 */
1418int unpoison_memory(unsigned long pfn)
1419{
1420 struct page *page;
1421 struct page *p;
1422 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001423 unsigned int nr_pages;
Wu Fengguang847ce402009-12-16 12:19:58 +01001424
1425 if (!pfn_valid(pfn))
1426 return -ENXIO;
1427
1428 p = pfn_to_page(pfn);
1429 page = compound_head(p);
1430
1431 if (!PageHWPoison(p)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001432 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001433 return 0;
1434 }
1435
Naoya Horiguchi230ac712015-09-08 15:03:29 -07001436 if (page_count(page) > 1) {
1437 pr_info("MCE: Someone grabs the hwpoison page %#lx\n", pfn);
1438 return 0;
1439 }
1440
1441 if (page_mapped(page)) {
1442 pr_info("MCE: Someone maps the hwpoison page %#lx\n", pfn);
1443 return 0;
1444 }
1445
1446 if (page_mapping(page)) {
1447 pr_info("MCE: the hwpoison page has non-NULL mapping %#lx\n",
1448 pfn);
1449 return 0;
1450 }
1451
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001452 /*
1453 * unpoison_memory() can encounter thp only when the thp is being
1454 * worked by memory_failure() and the page lock is not held yet.
1455 * In such case, we yield to memory_failure() and make unpoison fail.
1456 */
Wanpeng Lie76d30e2013-09-30 13:45:22 -07001457 if (!PageHuge(page) && PageTransHuge(page)) {
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001458 pr_info("MCE: Memory failure is now running on %#lx\n", pfn);
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001459 return 0;
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001460 }
1461
Wanpeng Lif9121152013-09-11 14:22:52 -07001462 nr_pages = 1 << compound_order(page);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001463
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001464 if (!get_hwpoison_page(p)) {
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001465 /*
1466 * Since HWPoisoned hugepage should have non-zero refcount,
1467 * race between memory failure and unpoison seems to happen.
1468 * In such case unpoison fails and memory failure runs
1469 * to the end.
1470 */
1471 if (PageHuge(page)) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001472 pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001473 return 0;
1474 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001475 if (TestClearPageHWPoison(p))
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001476 num_poisoned_pages_dec();
Andi Kleenfb46e732010-09-27 23:31:30 +02001477 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001478 return 0;
1479 }
1480
Jens Axboe7eaceac2011-03-10 08:52:07 +01001481 lock_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001482 /*
1483 * This test is racy because PG_hwpoison is set outside of page lock.
1484 * That's acceptable because that won't trigger kernel panic. Instead,
1485 * the PG_hwpoison page will be caught and isolated on the entrance to
1486 * the free buddy page pool.
1487 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001488 if (TestClearPageHWPoison(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001489 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001490 num_poisoned_pages_sub(nr_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001491 freeit = 1;
Naoya Horiguchi6a901812010-09-08 10:19:40 +09001492 if (PageHuge(page))
1493 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001494 }
1495 unlock_page(page);
1496
Wanpeng Li665d9da2015-09-08 15:03:21 -07001497 put_hwpoison_page(page);
Wanpeng Li3ba5eeb2013-09-11 14:23:01 -07001498 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
Wanpeng Li665d9da2015-09-08 15:03:21 -07001499 put_hwpoison_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001500
1501 return 0;
1502}
1503EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001504
1505static struct page *new_page(struct page *p, unsigned long private, int **x)
1506{
Andi Kleen12686d12009-12-16 12:20:01 +01001507 int nid = page_to_nid(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001508 if (PageHuge(p))
1509 return alloc_huge_page_node(page_hstate(compound_head(p)),
1510 nid);
1511 else
Vlastimil Babka96db8002015-09-08 15:03:50 -07001512 return __alloc_pages_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001513}
1514
1515/*
1516 * Safely get reference count of an arbitrary page.
1517 * Returns 0 for a free page, -EIO for a zero refcount page
1518 * that is not free, and 1 for any other page type.
1519 * For 1 the page is returned with increased page count, otherwise not.
1520 */
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001521static int __get_any_page(struct page *p, unsigned long pfn, int flags)
Andi Kleenfacb6012009-12-16 12:20:00 +01001522{
1523 int ret;
1524
1525 if (flags & MF_COUNT_INCREASED)
1526 return 1;
1527
1528 /*
Naoya Horiguchid950b952010-09-08 10:19:39 +09001529 * When the target page is a free hugepage, just remove it
1530 * from free hugepage list.
1531 */
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001532 if (!get_hwpoison_page(p)) {
Naoya Horiguchid950b952010-09-08 10:19:39 +09001533 if (PageHuge(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001534 pr_info("%s: %#lx free huge page\n", __func__, pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001535 ret = 0;
Naoya Horiguchid950b952010-09-08 10:19:39 +09001536 } else if (is_free_buddy_page(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001537 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001538 ret = 0;
1539 } else {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001540 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1541 __func__, pfn, p->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001542 ret = -EIO;
1543 }
1544 } else {
1545 /* Not a free page */
1546 ret = 1;
1547 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001548 return ret;
1549}
1550
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001551static int get_any_page(struct page *page, unsigned long pfn, int flags)
1552{
1553 int ret = __get_any_page(page, pfn, flags);
1554
1555 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1556 /*
1557 * Try to free it.
1558 */
Wanpeng Li665d9da2015-09-08 15:03:21 -07001559 put_hwpoison_page(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001560 shake_page(page, 1);
1561
1562 /*
1563 * Did it turn free?
1564 */
1565 ret = __get_any_page(page, pfn, 0);
1566 if (!PageLRU(page)) {
Wanpeng Li4f32be62015-08-14 15:34:56 -07001567 /* Drop page reference which is from __get_any_page() */
Wanpeng Li665d9da2015-09-08 15:03:21 -07001568 put_hwpoison_page(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001569 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1570 pfn, page->flags);
1571 return -EIO;
1572 }
1573 }
1574 return ret;
1575}
1576
Naoya Horiguchid950b952010-09-08 10:19:39 +09001577static int soft_offline_huge_page(struct page *page, int flags)
1578{
1579 int ret;
1580 unsigned long pfn = page_to_pfn(page);
1581 struct page *hpage = compound_head(page);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001582 LIST_HEAD(pagelist);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001583
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001584 /*
1585 * This double-check of PageHWPoison is to avoid the race with
1586 * memory_failure(). See also comment in __soft_offline_page().
1587 */
1588 lock_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001589 if (PageHWPoison(hpage)) {
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001590 unlock_page(hpage);
Wanpeng Li665d9da2015-09-08 15:03:21 -07001591 put_hwpoison_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001592 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001593 return -EBUSY;
Xishi Qiu0ebff322013-02-22 16:33:59 -08001594 }
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001595 unlock_page(hpage);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001596
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001597 ret = isolate_huge_page(hpage, &pagelist);
Wanpeng Li03613802015-08-14 15:34:59 -07001598 /*
1599 * get_any_page() and isolate_huge_page() takes a refcount each,
1600 * so need to drop one here.
1601 */
Wanpeng Li665d9da2015-09-08 15:03:21 -07001602 put_hwpoison_page(hpage);
Wanpeng Li03613802015-08-14 15:34:59 -07001603 if (!ret) {
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001604 pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn);
1605 return -EBUSY;
1606 }
1607
David Rientjes68711a72014-06-04 16:08:25 -07001608 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001609 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001610 if (ret) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001611 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1612 pfn, ret, page->flags);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001613 /*
1614 * We know that soft_offline_huge_page() tries to migrate
1615 * only one hugepage pointed to by hpage, so we need not
1616 * run through the pagelist here.
1617 */
1618 putback_active_hugepage(hpage);
1619 if (ret > 0)
1620 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001621 } else {
Jianguo Wua49ecbc2013-12-18 17:08:54 -08001622 /* overcommit hugetlb page will be freed to buddy */
1623 if (PageHuge(page)) {
1624 set_page_hwpoison_huge_page(hpage);
1625 dequeue_hwpoisoned_huge_page(hpage);
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001626 num_poisoned_pages_add(1 << compound_order(hpage));
Jianguo Wua49ecbc2013-12-18 17:08:54 -08001627 } else {
1628 SetPageHWPoison(page);
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001629 num_poisoned_pages_inc();
Jianguo Wua49ecbc2013-12-18 17:08:54 -08001630 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001631 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001632 return ret;
1633}
1634
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001635static int __soft_offline_page(struct page *page, int flags)
1636{
1637 int ret;
1638 unsigned long pfn = page_to_pfn(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001639
1640 /*
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001641 * Check PageHWPoison again inside page lock because PageHWPoison
1642 * is set by memory_failure() outside page lock. Note that
1643 * memory_failure() also double-checks PageHWPoison inside page lock,
1644 * so there's no race between soft_offline_page() and memory_failure().
Andi Kleenfacb6012009-12-16 12:20:00 +01001645 */
Xishi Qiu0ebff322013-02-22 16:33:59 -08001646 lock_page(page);
1647 wait_on_page_writeback(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001648 if (PageHWPoison(page)) {
1649 unlock_page(page);
Wanpeng Li665d9da2015-09-08 15:03:21 -07001650 put_hwpoison_page(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001651 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1652 return -EBUSY;
1653 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001654 /*
1655 * Try to invalidate first. This should work for
1656 * non dirty unmapped page cache pages.
1657 */
1658 ret = invalidate_inode_page(page);
1659 unlock_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001660 /*
Andi Kleenfacb6012009-12-16 12:20:00 +01001661 * RED-PEN would be better to keep it isolated here, but we
1662 * would need to fix isolation locking first.
1663 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001664 if (ret == 1) {
Wanpeng Li665d9da2015-09-08 15:03:21 -07001665 put_hwpoison_page(page);
Andi Kleenfb46e732010-09-27 23:31:30 +02001666 pr_info("soft_offline: %#lx: invalidated\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001667 SetPageHWPoison(page);
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001668 num_poisoned_pages_inc();
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001669 return 0;
Andi Kleenfacb6012009-12-16 12:20:00 +01001670 }
1671
1672 /*
1673 * Simple invalidation didn't work.
1674 * Try to migrate to a new page instead. migrate.c
1675 * handles a large number of cases for us.
1676 */
1677 ret = isolate_lru_page(page);
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001678 /*
1679 * Drop page reference which is came from get_any_page()
1680 * successful isolate_lru_page() already took another one.
1681 */
Wanpeng Li665d9da2015-09-08 15:03:21 -07001682 put_hwpoison_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001683 if (!ret) {
1684 LIST_HEAD(pagelist);
Minchan Kim5db8a732011-06-15 15:08:48 -07001685 inc_zone_page_state(page, NR_ISOLATED_ANON +
Hugh Dickins9c620e22013-02-22 16:35:14 -08001686 page_is_file_cache(page));
Andi Kleenfacb6012009-12-16 12:20:00 +01001687 list_add(&page->lru, &pagelist);
David Rientjes68711a72014-06-04 16:08:25 -07001688 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001689 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Andi Kleenfacb6012009-12-16 12:20:00 +01001690 if (ret) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001691 if (!list_empty(&pagelist)) {
1692 list_del(&page->lru);
1693 dec_zone_page_state(page, NR_ISOLATED_ANON +
1694 page_is_file_cache(page));
1695 putback_lru_page(page);
1696 }
1697
Andi Kleenfb46e732010-09-27 23:31:30 +02001698 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001699 pfn, ret, page->flags);
1700 if (ret > 0)
1701 ret = -EIO;
1702 }
1703 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001704 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
Dean Nelsondd73e852011-10-31 17:09:04 -07001705 pfn, ret, page_count(page), page->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001706 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001707 return ret;
1708}
Wanpeng Li86e05772013-09-11 14:22:56 -07001709
1710/**
1711 * soft_offline_page - Soft offline a page.
1712 * @page: page to offline
1713 * @flags: flags. Same as memory_failure().
1714 *
1715 * Returns 0 on success, otherwise negated errno.
1716 *
1717 * Soft offline a page, by migration or invalidation,
1718 * without killing anything. This is for the case when
1719 * a page is not corrupted yet (so it's still valid to access),
1720 * but has had a number of corrected errors and is better taken
1721 * out.
1722 *
1723 * The actual policy on when to do that is maintained by
1724 * user space.
1725 *
1726 * This should never impact any application or cause data loss,
1727 * however it might take some time.
1728 *
1729 * This is not a 100% solution for all memory, but tries to be
1730 * ``good enough'' for the majority of memory.
1731 */
1732int soft_offline_page(struct page *page, int flags)
1733{
1734 int ret;
1735 unsigned long pfn = page_to_pfn(page);
David Rientjes668f9abb2014-03-03 15:38:18 -08001736 struct page *hpage = compound_head(page);
Wanpeng Li86e05772013-09-11 14:22:56 -07001737
1738 if (PageHWPoison(page)) {
1739 pr_info("soft offline: %#lx page already poisoned\n", pfn);
Wanpeng Li1e0e6352015-09-08 15:03:13 -07001740 if (flags & MF_COUNT_INCREASED)
Wanpeng Li665d9da2015-09-08 15:03:21 -07001741 put_hwpoison_page(page);
Wanpeng Li86e05772013-09-11 14:22:56 -07001742 return -EBUSY;
1743 }
1744 if (!PageHuge(page) && PageTransHuge(hpage)) {
1745 if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1746 pr_info("soft offline: %#lx: failed to split THP\n",
1747 pfn);
Wanpeng Li7d1900c2015-09-08 15:03:10 -07001748 if (flags & MF_COUNT_INCREASED)
Wanpeng Li665d9da2015-09-08 15:03:21 -07001749 put_hwpoison_page(page);
Wanpeng Li86e05772013-09-11 14:22:56 -07001750 return -EBUSY;
1751 }
1752 }
1753
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07001754 get_online_mems();
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001755
Wanpeng Li86e05772013-09-11 14:22:56 -07001756 ret = get_any_page(page, pfn, flags);
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07001757 put_online_mems();
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001758 if (ret > 0) { /* for in-use pages */
Wanpeng Li86e05772013-09-11 14:22:56 -07001759 if (PageHuge(page))
1760 ret = soft_offline_huge_page(page, flags);
1761 else
1762 ret = __soft_offline_page(page, flags);
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001763 } else if (ret == 0) { /* for free pages */
Wanpeng Li86e05772013-09-11 14:22:56 -07001764 if (PageHuge(page)) {
1765 set_page_hwpoison_huge_page(hpage);
Naoya Horiguchi602498f2015-05-05 16:23:46 -07001766 if (!dequeue_hwpoisoned_huge_page(hpage))
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001767 num_poisoned_pages_add(1 << compound_order(hpage));
Wanpeng Li86e05772013-09-11 14:22:56 -07001768 } else {
Naoya Horiguchi602498f2015-05-05 16:23:46 -07001769 if (!TestSetPageHWPoison(page))
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001770 num_poisoned_pages_inc();
Wanpeng Li86e05772013-09-11 14:22:56 -07001771 }
1772 }
Wanpeng Li86e05772013-09-11 14:22:56 -07001773 return ret;
1774}