blob: 8ad923a93539b4ac99e65cf507cab377388b618c [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 */
Andrew Mortonc255a452012-07-31 16:43:02 -0700133#ifdef CONFIG_MEMCG_SWAP
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{
138 struct mem_cgroup *mem;
139 struct cgroup_subsys_state *css;
140 unsigned long ino;
141
142 if (!hwpoison_filter_memcg)
143 return 0;
144
145 mem = try_get_mem_cgroup_from_page(p);
146 if (!mem)
147 return -EINVAL;
148
Michal Hocko33398cf2015-09-08 15:01:02 -0700149 css = &mem->css;
Tejun Heob1664922014-02-11 11:52:49 -0500150 ino = cgroup_ino(css->cgroup);
Andi Kleen4fd466e2009-12-16 12:19:59 +0100151 css_put(css);
152
Zefan Lif29374b2014-09-19 16:29:31 +0800153 if (ino != hwpoison_filter_memcg)
Andi Kleen4fd466e2009-12-16 12:19:59 +0100154 return -EINVAL;
155
156 return 0;
157}
158#else
159static int hwpoison_filter_task(struct page *p) { return 0; }
160#endif
161
Wu Fengguang7c116f22009-12-16 12:19:59 +0100162int hwpoison_filter(struct page *p)
163{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100164 if (!hwpoison_filter_enable)
165 return 0;
166
Wu Fengguang7c116f22009-12-16 12:19:59 +0100167 if (hwpoison_filter_dev(p))
168 return -EINVAL;
169
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100170 if (hwpoison_filter_flags(p))
171 return -EINVAL;
172
Andi Kleen4fd466e2009-12-16 12:19:59 +0100173 if (hwpoison_filter_task(p))
174 return -EINVAL;
175
Wu Fengguang7c116f22009-12-16 12:19:59 +0100176 return 0;
177}
Andi Kleen27df5062009-12-21 19:56:42 +0100178#else
179int hwpoison_filter(struct page *p)
180{
181 return 0;
182}
183#endif
184
Wu Fengguang7c116f22009-12-16 12:19:59 +0100185EXPORT_SYMBOL_GPL(hwpoison_filter);
186
Andi Kleen6a460792009-09-16 11:50:15 +0200187/*
Tony Luck7329bbe2011-12-13 09:27:58 -0800188 * Send all the processes who have the page mapped a signal.
189 * ``action optional'' if they are not immediately affected by the error
190 * ``action required'' if error happened in current execution context
Andi Kleen6a460792009-09-16 11:50:15 +0200191 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800192static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
193 unsigned long pfn, struct page *page, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200194{
195 struct siginfo si;
196 int ret;
197
198 printk(KERN_ERR
Tony Luck7329bbe2011-12-13 09:27:58 -0800199 "MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200200 pfn, t->comm, t->pid);
201 si.si_signo = SIGBUS;
202 si.si_errno = 0;
Andi Kleen6a460792009-09-16 11:50:15 +0200203 si.si_addr = (void *)addr;
204#ifdef __ARCH_SI_TRAPNO
205 si.si_trapno = trapno;
206#endif
Wanpeng Lif9121152013-09-11 14:22:52 -0700207 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
Tony Luck7329bbe2011-12-13 09:27:58 -0800208
Tony Lucka70ffca2014-06-04 16:10:59 -0700209 if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
Tony Luck7329bbe2011-12-13 09:27:58 -0800210 si.si_code = BUS_MCEERR_AR;
Tony Lucka70ffca2014-06-04 16:10:59 -0700211 ret = force_sig_info(SIGBUS, &si, current);
Tony Luck7329bbe2011-12-13 09:27:58 -0800212 } else {
213 /*
214 * Don't use force here, it's convenient if the signal
215 * can be temporarily blocked.
216 * This could cause a loop when the user sets SIGBUS
217 * to SIG_IGN, but hopefully no one will do that?
218 */
219 si.si_code = BUS_MCEERR_AO;
220 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
221 }
Andi Kleen6a460792009-09-16 11:50:15 +0200222 if (ret < 0)
223 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
224 t->comm, t->pid, ret);
225 return ret;
226}
227
228/*
Andi Kleen588f9ce2009-12-16 12:19:57 +0100229 * When a unknown page type is encountered drain as many buffers as possible
230 * in the hope to turn the page into a LRU or free page, which we can handle.
231 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100232void shake_page(struct page *p, int access)
Andi Kleen588f9ce2009-12-16 12:19:57 +0100233{
234 if (!PageSlab(p)) {
235 lru_add_drain_all();
236 if (PageLRU(p))
237 return;
Vlastimil Babkac0554322014-12-10 15:43:10 -0800238 drain_all_pages(page_zone(p));
Andi Kleen588f9ce2009-12-16 12:19:57 +0100239 if (PageLRU(p) || is_free_buddy_page(p))
240 return;
241 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100242
Andi Kleen588f9ce2009-12-16 12:19:57 +0100243 /*
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800244 * Only call shrink_node_slabs here (which would also shrink
245 * other caches) if access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100246 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800247 if (access)
248 drop_slab_node(page_to_nid(p));
Andi Kleen588f9ce2009-12-16 12:19:57 +0100249}
250EXPORT_SYMBOL_GPL(shake_page);
251
252/*
Andi Kleen6a460792009-09-16 11:50:15 +0200253 * Kill all processes that have a poisoned page mapped and then isolate
254 * the page.
255 *
256 * General strategy:
257 * Find all processes having the page mapped and kill them.
258 * But we keep a page reference around so that the page is not
259 * actually freed yet.
260 * Then stash the page away
261 *
262 * There's no convenient way to get back to mapped processes
263 * from the VMAs. So do a brute-force search over all
264 * running processes.
265 *
266 * Remember that machine checks are not common (or rather
267 * if they are common you have other problems), so this shouldn't
268 * be a performance issue.
269 *
270 * Also there are some races possible while we get from the
271 * error detection to actually handle it.
272 */
273
274struct to_kill {
275 struct list_head nd;
276 struct task_struct *tsk;
277 unsigned long addr;
Andi Kleen9033ae12010-09-27 23:36:05 +0200278 char addr_valid;
Andi Kleen6a460792009-09-16 11:50:15 +0200279};
280
281/*
282 * Failure handling: if we can't find or can't kill a process there's
283 * not much we can do. We just print a message and ignore otherwise.
284 */
285
286/*
287 * Schedule a process for later kill.
288 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
289 * TBD would GFP_NOIO be enough?
290 */
291static void add_to_kill(struct task_struct *tsk, struct page *p,
292 struct vm_area_struct *vma,
293 struct list_head *to_kill,
294 struct to_kill **tkc)
295{
296 struct to_kill *tk;
297
298 if (*tkc) {
299 tk = *tkc;
300 *tkc = NULL;
301 } else {
302 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
303 if (!tk) {
304 printk(KERN_ERR
305 "MCE: Out of memory while machine check handling\n");
306 return;
307 }
308 }
309 tk->addr = page_address_in_vma(p, vma);
310 tk->addr_valid = 1;
311
312 /*
313 * In theory we don't have to kill when the page was
314 * munmaped. But it could be also a mremap. Since that's
315 * likely very rare kill anyways just out of paranoia, but use
316 * a SIGKILL because the error is not contained anymore.
317 */
318 if (tk->addr == -EFAULT) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200319 pr_info("MCE: Unable to find user space address %lx in %s\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200320 page_to_pfn(p), tsk->comm);
321 tk->addr_valid = 0;
322 }
323 get_task_struct(tsk);
324 tk->tsk = tsk;
325 list_add_tail(&tk->nd, to_kill);
326}
327
328/*
329 * Kill the processes that have been collected earlier.
330 *
331 * Only do anything when DOIT is set, otherwise just free the list
332 * (this is used for clean pages which do not need killing)
333 * Also when FAIL is set do a force kill because something went
334 * wrong earlier.
335 */
Tony Luck6751ed62012-07-11 10:20:47 -0700336static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -0800337 int fail, struct page *page, unsigned long pfn,
338 int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200339{
340 struct to_kill *tk, *next;
341
342 list_for_each_entry_safe (tk, next, to_kill, nd) {
Tony Luck6751ed62012-07-11 10:20:47 -0700343 if (forcekill) {
Andi Kleen6a460792009-09-16 11:50:15 +0200344 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200345 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200346 * make sure the process doesn't catch the
347 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200348 */
349 if (fail || tk->addr_valid == 0) {
350 printk(KERN_ERR
351 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
352 pfn, tk->tsk->comm, tk->tsk->pid);
353 force_sig(SIGKILL, tk->tsk);
354 }
355
356 /*
357 * In theory the process could have mapped
358 * something else on the address in-between. We could
359 * check for that, but we need to tell the
360 * process anyways.
361 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800362 else if (kill_proc(tk->tsk, tk->addr, trapno,
363 pfn, page, flags) < 0)
Andi Kleen6a460792009-09-16 11:50:15 +0200364 printk(KERN_ERR
365 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
366 pfn, tk->tsk->comm, tk->tsk->pid);
367 }
368 put_task_struct(tk->tsk);
369 kfree(tk);
370 }
371}
372
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700373/*
374 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
375 * on behalf of the thread group. Return task_struct of the (first found)
376 * dedicated thread if found, and return NULL otherwise.
377 *
378 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
379 * have to call rcu_read_lock/unlock() in this function.
380 */
381static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
Andi Kleen6a460792009-09-16 11:50:15 +0200382{
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700383 struct task_struct *t;
384
385 for_each_thread(tsk, t)
386 if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
387 return t;
388 return NULL;
389}
390
391/*
392 * Determine whether a given process is "early kill" process which expects
393 * to be signaled when some page under the process is hwpoisoned.
394 * Return task_struct of the dedicated thread (main thread unless explicitly
395 * specified) if the process is "early kill," and otherwise returns NULL.
396 */
397static struct task_struct *task_early_kill(struct task_struct *tsk,
398 int force_early)
399{
400 struct task_struct *t;
Andi Kleen6a460792009-09-16 11:50:15 +0200401 if (!tsk->mm)
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700402 return NULL;
Tony Luck74614de2014-06-04 16:11:01 -0700403 if (force_early)
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700404 return tsk;
405 t = find_early_kill_thread(tsk);
406 if (t)
407 return t;
408 if (sysctl_memory_failure_early_kill)
409 return tsk;
410 return NULL;
Andi Kleen6a460792009-09-16 11:50:15 +0200411}
412
413/*
414 * Collect processes when the error hit an anonymous page.
415 */
416static void collect_procs_anon(struct page *page, struct list_head *to_kill,
Tony Luck74614de2014-06-04 16:11:01 -0700417 struct to_kill **tkc, int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200418{
419 struct vm_area_struct *vma;
420 struct task_struct *tsk;
421 struct anon_vma *av;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700422 pgoff_t pgoff;
Andi Kleen6a460792009-09-16 11:50:15 +0200423
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000424 av = page_lock_anon_vma_read(page);
Andi Kleen6a460792009-09-16 11:50:15 +0200425 if (av == NULL) /* Not actually mapped anymore */
Peter Zijlstra9b679322011-06-27 16:18:09 -0700426 return;
427
Naoya Horiguchia0f7a752014-07-23 14:00:01 -0700428 pgoff = page_to_pgoff(page);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700429 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200430 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800431 struct anon_vma_chain *vmac;
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700432 struct task_struct *t = task_early_kill(tsk, force_early);
Rik van Riel5beb4932010-03-05 13:42:07 -0800433
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700434 if (!t)
Andi Kleen6a460792009-09-16 11:50:15 +0200435 continue;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700436 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
437 pgoff, pgoff) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800438 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200439 if (!page_mapped_in_vma(page, vma))
440 continue;
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700441 if (vma->vm_mm == t->mm)
442 add_to_kill(t, page, vma, to_kill, tkc);
Andi Kleen6a460792009-09-16 11:50:15 +0200443 }
444 }
Andi Kleen6a460792009-09-16 11:50:15 +0200445 read_unlock(&tasklist_lock);
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000446 page_unlock_anon_vma_read(av);
Andi Kleen6a460792009-09-16 11:50:15 +0200447}
448
449/*
450 * Collect processes when the error hit a file mapped page.
451 */
452static void collect_procs_file(struct page *page, struct list_head *to_kill,
Tony Luck74614de2014-06-04 16:11:01 -0700453 struct to_kill **tkc, int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200454{
455 struct vm_area_struct *vma;
456 struct task_struct *tsk;
Andi Kleen6a460792009-09-16 11:50:15 +0200457 struct address_space *mapping = page->mapping;
458
Davidlohr Buesod28eb9c2014-12-12 16:54:36 -0800459 i_mmap_lock_read(mapping);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700460 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200461 for_each_process(tsk) {
Naoya Horiguchia0f7a752014-07-23 14:00:01 -0700462 pgoff_t pgoff = page_to_pgoff(page);
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700463 struct task_struct *t = task_early_kill(tsk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200464
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700465 if (!t)
Andi Kleen6a460792009-09-16 11:50:15 +0200466 continue;
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -0700467 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
Andi Kleen6a460792009-09-16 11:50:15 +0200468 pgoff) {
469 /*
470 * Send early kill signal to tasks where a vma covers
471 * the page but the corrupted page is not necessarily
472 * mapped it in its pte.
473 * Assume applications who requested early kill want
474 * to be informed of all such data corruptions.
475 */
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700476 if (vma->vm_mm == t->mm)
477 add_to_kill(t, page, vma, to_kill, tkc);
Andi Kleen6a460792009-09-16 11:50:15 +0200478 }
479 }
Andi Kleen6a460792009-09-16 11:50:15 +0200480 read_unlock(&tasklist_lock);
Davidlohr Buesod28eb9c2014-12-12 16:54:36 -0800481 i_mmap_unlock_read(mapping);
Andi Kleen6a460792009-09-16 11:50:15 +0200482}
483
484/*
485 * Collect the processes who have the corrupted page mapped to kill.
486 * This is done in two steps for locking reasons.
487 * First preallocate one tokill structure outside the spin locks,
488 * so that we can kill at least one process reasonably reliable.
489 */
Tony Luck74614de2014-06-04 16:11:01 -0700490static void collect_procs(struct page *page, struct list_head *tokill,
491 int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200492{
493 struct to_kill *tk;
494
495 if (!page->mapping)
496 return;
497
498 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
499 if (!tk)
500 return;
501 if (PageAnon(page))
Tony Luck74614de2014-06-04 16:11:01 -0700502 collect_procs_anon(page, tokill, &tk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200503 else
Tony Luck74614de2014-06-04 16:11:01 -0700504 collect_procs_file(page, tokill, &tk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200505 kfree(tk);
506}
507
Andi Kleen6a460792009-09-16 11:50:15 +0200508static const char *action_name[] = {
Xie XiuQicc637b12015-06-24 16:57:30 -0700509 [MF_IGNORED] = "Ignored",
510 [MF_FAILED] = "Failed",
511 [MF_DELAYED] = "Delayed",
512 [MF_RECOVERED] = "Recovered",
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700513};
514
515static const char * const action_page_types[] = {
Xie XiuQicc637b12015-06-24 16:57:30 -0700516 [MF_MSG_KERNEL] = "reserved kernel page",
517 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
518 [MF_MSG_SLAB] = "kernel slab page",
519 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
520 [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned",
521 [MF_MSG_HUGE] = "huge page",
522 [MF_MSG_FREE_HUGE] = "free huge page",
523 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
524 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
525 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
526 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
527 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
528 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
529 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
530 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
531 [MF_MSG_CLEAN_LRU] = "clean LRU page",
532 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
533 [MF_MSG_BUDDY] = "free buddy page",
534 [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)",
535 [MF_MSG_UNKNOWN] = "unknown page",
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700536};
537
Andi Kleen6a460792009-09-16 11:50:15 +0200538/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100539 * XXX: It is possible that a page is isolated from LRU cache,
540 * and then kept in swap cache or failed to remove from page cache.
541 * The page count will stop it from being freed by unpoison.
542 * Stress tests should be aware of this memory leak problem.
543 */
544static int delete_from_lru_cache(struct page *p)
545{
546 if (!isolate_lru_page(p)) {
547 /*
548 * Clear sensible page flags, so that the buddy system won't
549 * complain when the page is unpoison-and-freed.
550 */
551 ClearPageActive(p);
552 ClearPageUnevictable(p);
553 /*
554 * drop the page count elevated by isolate_lru_page()
555 */
556 page_cache_release(p);
557 return 0;
558 }
559 return -EIO;
560}
561
562/*
Andi Kleen6a460792009-09-16 11:50:15 +0200563 * Error hit kernel page.
564 * Do nothing, try to be lucky and not touch this instead. For a few cases we
565 * could be more sophisticated.
566 */
567static int me_kernel(struct page *p, unsigned long pfn)
568{
Xie XiuQicc637b12015-06-24 16:57:30 -0700569 return MF_IGNORED;
Andi Kleen6a460792009-09-16 11:50:15 +0200570}
571
572/*
573 * Page in unknown state. Do nothing.
574 */
575static int me_unknown(struct page *p, unsigned long pfn)
576{
577 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
Xie XiuQicc637b12015-06-24 16:57:30 -0700578 return MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200579}
580
581/*
Andi Kleen6a460792009-09-16 11:50:15 +0200582 * Clean (or cleaned) page cache page.
583 */
584static int me_pagecache_clean(struct page *p, unsigned long pfn)
585{
586 int err;
Xie XiuQicc637b12015-06-24 16:57:30 -0700587 int ret = MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200588 struct address_space *mapping;
589
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100590 delete_from_lru_cache(p);
591
Andi Kleen6a460792009-09-16 11:50:15 +0200592 /*
593 * For anonymous pages we're done the only reference left
594 * should be the one m_f() holds.
595 */
596 if (PageAnon(p))
Xie XiuQicc637b12015-06-24 16:57:30 -0700597 return MF_RECOVERED;
Andi Kleen6a460792009-09-16 11:50:15 +0200598
599 /*
600 * Now truncate the page in the page cache. This is really
601 * more like a "temporary hole punch"
602 * Don't do this for block devices when someone else
603 * has a reference, because it could be file system metadata
604 * and that's not safe to truncate.
605 */
606 mapping = page_mapping(p);
607 if (!mapping) {
608 /*
609 * Page has been teared down in the meanwhile
610 */
Xie XiuQicc637b12015-06-24 16:57:30 -0700611 return MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200612 }
613
614 /*
615 * Truncation is a bit tricky. Enable it per file system for now.
616 *
617 * Open: to take i_mutex or not for this? Right now we don't.
618 */
619 if (mapping->a_ops->error_remove_page) {
620 err = mapping->a_ops->error_remove_page(mapping, p);
621 if (err != 0) {
622 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
623 pfn, err);
624 } else if (page_has_private(p) &&
625 !try_to_release_page(p, GFP_NOIO)) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200626 pr_info("MCE %#lx: failed to release buffers\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200627 } else {
Xie XiuQicc637b12015-06-24 16:57:30 -0700628 ret = MF_RECOVERED;
Andi Kleen6a460792009-09-16 11:50:15 +0200629 }
630 } else {
631 /*
632 * If the file system doesn't support it just invalidate
633 * This fails on dirty or anything with private pages
634 */
635 if (invalidate_inode_page(p))
Xie XiuQicc637b12015-06-24 16:57:30 -0700636 ret = MF_RECOVERED;
Andi Kleen6a460792009-09-16 11:50:15 +0200637 else
638 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
639 pfn);
640 }
641 return ret;
642}
643
644/*
Zhi Yong Wu549543d2014-01-21 15:49:08 -0800645 * Dirty pagecache page
Andi Kleen6a460792009-09-16 11:50:15 +0200646 * Issues: when the error hit a hole page the error is not properly
647 * propagated.
648 */
649static int me_pagecache_dirty(struct page *p, unsigned long pfn)
650{
651 struct address_space *mapping = page_mapping(p);
652
653 SetPageError(p);
654 /* TBD: print more information about the file. */
655 if (mapping) {
656 /*
657 * IO error will be reported by write(), fsync(), etc.
658 * who check the mapping.
659 * This way the application knows that something went
660 * wrong with its dirty file data.
661 *
662 * There's one open issue:
663 *
664 * The EIO will be only reported on the next IO
665 * operation and then cleared through the IO map.
666 * Normally Linux has two mechanisms to pass IO error
667 * first through the AS_EIO flag in the address space
668 * and then through the PageError flag in the page.
669 * Since we drop pages on memory failure handling the
670 * only mechanism open to use is through AS_AIO.
671 *
672 * This has the disadvantage that it gets cleared on
673 * the first operation that returns an error, while
674 * the PageError bit is more sticky and only cleared
675 * when the page is reread or dropped. If an
676 * application assumes it will always get error on
677 * fsync, but does other operations on the fd before
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300678 * and the page is dropped between then the error
Andi Kleen6a460792009-09-16 11:50:15 +0200679 * will not be properly reported.
680 *
681 * This can already happen even without hwpoisoned
682 * pages: first on metadata IO errors (which only
683 * report through AS_EIO) or when the page is dropped
684 * at the wrong time.
685 *
686 * So right now we assume that the application DTRT on
687 * the first EIO, but we're not worse than other parts
688 * of the kernel.
689 */
690 mapping_set_error(mapping, EIO);
691 }
692
693 return me_pagecache_clean(p, pfn);
694}
695
696/*
697 * Clean and dirty swap cache.
698 *
699 * Dirty swap cache page is tricky to handle. The page could live both in page
700 * cache and swap cache(ie. page is freshly swapped in). So it could be
701 * referenced concurrently by 2 types of PTEs:
702 * normal PTEs and swap PTEs. We try to handle them consistently by calling
703 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
704 * and then
705 * - clear dirty bit to prevent IO
706 * - remove from LRU
707 * - but keep in the swap cache, so that when we return to it on
708 * a later page fault, we know the application is accessing
709 * corrupted data and shall be killed (we installed simple
710 * interception code in do_swap_page to catch it).
711 *
712 * Clean swap cache pages can be directly isolated. A later page fault will
713 * bring in the known good data from disk.
714 */
715static int me_swapcache_dirty(struct page *p, unsigned long pfn)
716{
Andi Kleen6a460792009-09-16 11:50:15 +0200717 ClearPageDirty(p);
718 /* Trigger EIO in shmem: */
719 ClearPageUptodate(p);
720
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100721 if (!delete_from_lru_cache(p))
Xie XiuQicc637b12015-06-24 16:57:30 -0700722 return MF_DELAYED;
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100723 else
Xie XiuQicc637b12015-06-24 16:57:30 -0700724 return MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200725}
726
727static int me_swapcache_clean(struct page *p, unsigned long pfn)
728{
Andi Kleen6a460792009-09-16 11:50:15 +0200729 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800730
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100731 if (!delete_from_lru_cache(p))
Xie XiuQicc637b12015-06-24 16:57:30 -0700732 return MF_RECOVERED;
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100733 else
Xie XiuQicc637b12015-06-24 16:57:30 -0700734 return MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200735}
736
737/*
738 * Huge pages. Needs work.
739 * Issues:
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900740 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
741 * To narrow down kill region to one page, we need to break up pmd.
Andi Kleen6a460792009-09-16 11:50:15 +0200742 */
743static int me_huge_page(struct page *p, unsigned long pfn)
744{
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900745 int res = 0;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900746 struct page *hpage = compound_head(p);
Naoya Horiguchi2491ffe2015-06-24 16:56:53 -0700747
748 if (!PageHuge(hpage))
749 return MF_DELAYED;
750
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900751 /*
752 * We can safely recover from error on free or reserved (i.e.
753 * not in-use) hugepage by dequeuing it from freelist.
754 * To check whether a hugepage is in-use or not, we can't use
755 * page->lru because it can be used in other hugepage operations,
756 * such as __unmap_hugepage_range() and gather_surplus_pages().
757 * So instead we use page_mapping() and PageAnon().
758 * We assume that this function is called with page lock held,
759 * so there is no race between isolation and mapping/unmapping.
760 */
761 if (!(page_mapping(hpage) || PageAnon(hpage))) {
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900762 res = dequeue_hwpoisoned_huge_page(hpage);
763 if (!res)
Xie XiuQicc637b12015-06-24 16:57:30 -0700764 return MF_RECOVERED;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900765 }
Xie XiuQicc637b12015-06-24 16:57:30 -0700766 return MF_DELAYED;
Andi Kleen6a460792009-09-16 11:50:15 +0200767}
768
769/*
770 * Various page states we can handle.
771 *
772 * A page state is defined by its current page->flags bits.
773 * The table matches them in order and calls the right handler.
774 *
775 * This is quite tricky because we can access page at any time
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300776 * in its live cycle, so all accesses have to be extremely careful.
Andi Kleen6a460792009-09-16 11:50:15 +0200777 *
778 * This is not complete. More states could be added.
779 * For any missing state don't attempt recovery.
780 */
781
782#define dirty (1UL << PG_dirty)
783#define sc (1UL << PG_swapcache)
784#define unevict (1UL << PG_unevictable)
785#define mlock (1UL << PG_mlocked)
786#define writeback (1UL << PG_writeback)
787#define lru (1UL << PG_lru)
788#define swapbacked (1UL << PG_swapbacked)
789#define head (1UL << PG_head)
790#define tail (1UL << PG_tail)
791#define compound (1UL << PG_compound)
792#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200793#define reserved (1UL << PG_reserved)
794
795static struct page_state {
796 unsigned long mask;
797 unsigned long res;
Xie XiuQicc637b12015-06-24 16:57:30 -0700798 enum mf_action_page_type type;
Andi Kleen6a460792009-09-16 11:50:15 +0200799 int (*action)(struct page *p, unsigned long pfn);
800} error_states[] = {
Xie XiuQicc637b12015-06-24 16:57:30 -0700801 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100802 /*
803 * free pages are specially detected outside this table:
804 * PG_buddy pages only make a small fraction of all free pages.
805 */
Andi Kleen6a460792009-09-16 11:50:15 +0200806
807 /*
808 * Could in theory check if slab page is free or if we can drop
809 * currently unused objects without touching them. But just
810 * treat it as standard kernel for now.
811 */
Xie XiuQicc637b12015-06-24 16:57:30 -0700812 { slab, slab, MF_MSG_SLAB, me_kernel },
Andi Kleen6a460792009-09-16 11:50:15 +0200813
814#ifdef CONFIG_PAGEFLAGS_EXTENDED
Xie XiuQicc637b12015-06-24 16:57:30 -0700815 { head, head, MF_MSG_HUGE, me_huge_page },
816 { tail, tail, MF_MSG_HUGE, me_huge_page },
Andi Kleen6a460792009-09-16 11:50:15 +0200817#else
Xie XiuQicc637b12015-06-24 16:57:30 -0700818 { compound, compound, MF_MSG_HUGE, me_huge_page },
Andi Kleen6a460792009-09-16 11:50:15 +0200819#endif
820
Xie XiuQicc637b12015-06-24 16:57:30 -0700821 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
822 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200823
Xie XiuQicc637b12015-06-24 16:57:30 -0700824 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
825 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200826
Xie XiuQicc637b12015-06-24 16:57:30 -0700827 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
828 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
Naoya Horiguchi5f4b9fc2013-02-22 16:35:53 -0800829
Xie XiuQicc637b12015-06-24 16:57:30 -0700830 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
831 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200832
833 /*
834 * Catchall entry: must be at end.
835 */
Xie XiuQicc637b12015-06-24 16:57:30 -0700836 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
Andi Kleen6a460792009-09-16 11:50:15 +0200837};
838
Andi Kleen2326c462009-12-16 12:20:00 +0100839#undef dirty
840#undef sc
841#undef unevict
842#undef mlock
843#undef writeback
844#undef lru
845#undef swapbacked
846#undef head
847#undef tail
848#undef compound
849#undef slab
850#undef reserved
851
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800852/*
853 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
854 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
855 */
Xie XiuQicc3e2af2015-06-24 16:57:33 -0700856static void action_result(unsigned long pfn, enum mf_action_page_type type,
857 enum mf_result result)
Andi Kleen6a460792009-09-16 11:50:15 +0200858{
Xie XiuQi97f0b132015-06-24 16:57:36 -0700859 trace_memory_failure_event(pfn, type, result);
860
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700861 pr_err("MCE %#lx: recovery action for %s: %s\n",
862 pfn, action_page_types[type], action_name[result]);
Andi Kleen6a460792009-09-16 11:50:15 +0200863}
864
865static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100866 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200867{
868 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200869 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200870
871 result = ps->action(p, pfn);
Wu Fengguang7456b042009-10-19 08:15:01 +0200872
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100873 count = page_count(p) - 1;
Xie XiuQicc637b12015-06-24 16:57:30 -0700874 if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
Wu Fengguang138ce282009-12-16 12:19:58 +0100875 count--;
876 if (count != 0) {
Andi Kleen6a460792009-09-16 11:50:15 +0200877 printk(KERN_ERR
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700878 "MCE %#lx: %s still referenced by %d users\n",
879 pfn, action_page_types[ps->type], count);
Xie XiuQicc637b12015-06-24 16:57:30 -0700880 result = MF_FAILED;
Wu Fengguang138ce282009-12-16 12:19:58 +0100881 }
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700882 action_result(pfn, ps->type, result);
Andi Kleen6a460792009-09-16 11:50:15 +0200883
884 /* Could do more checks here if page looks ok */
885 /*
886 * Could adjust zone counters here to correct for the missing page.
887 */
888
Xie XiuQicc637b12015-06-24 16:57:30 -0700889 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200890}
891
Naoya Horiguchiead07f62015-06-24 16:56:48 -0700892/**
893 * get_hwpoison_page() - Get refcount for memory error handling:
894 * @page: raw error page (hit by memory error)
895 *
896 * Return: return 0 if failed to grab the refcount, otherwise true (some
897 * non-zero value.)
898 */
899int get_hwpoison_page(struct page *page)
900{
901 struct page *head = compound_head(page);
902
903 if (PageHuge(head))
904 return get_page_unless_zero(head);
905
906 /*
907 * Thp tail page has special refcounting rule (refcount of tail pages
908 * is stored in ->_mapcount,) so we can't call get_page_unless_zero()
909 * directly for tail pages.
910 */
911 if (PageTransHuge(head)) {
Naoya Horiguchi98ed2b02015-08-06 15:47:04 -0700912 /*
913 * Non anonymous thp exists only in allocation/free time. We
914 * can't handle such a case correctly, so let's give it up.
915 * This should be better than triggering BUG_ON when kernel
916 * tries to touch the "partially handled" page.
917 */
918 if (!PageAnon(head)) {
919 pr_err("MCE: %#lx: non anonymous thp\n",
920 page_to_pfn(page));
921 return 0;
922 }
923
Naoya Horiguchiead07f62015-06-24 16:56:48 -0700924 if (get_page_unless_zero(head)) {
925 if (PageTail(page))
926 get_page(page);
927 return 1;
928 } else {
929 return 0;
930 }
931 }
932
933 return get_page_unless_zero(page);
934}
935EXPORT_SYMBOL_GPL(get_hwpoison_page);
936
Andi Kleen6a460792009-09-16 11:50:15 +0200937/*
938 * Do all that is necessary to remove user space mappings. Unmap
939 * the pages and send SIGBUS to the processes if the data was dirty.
940 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100941static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800942 int trapno, int flags, struct page **hpagep)
Andi Kleen6a460792009-09-16 11:50:15 +0200943{
944 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
945 struct address_space *mapping;
946 LIST_HEAD(tokill);
947 int ret;
Tony Luck6751ed62012-07-11 10:20:47 -0700948 int kill = 1, forcekill;
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800949 struct page *hpage = *hpagep;
Andi Kleen6a460792009-09-16 11:50:15 +0200950
Naoya Horiguchi93a9eb32014-07-30 16:08:28 -0700951 /*
952 * Here we are interested only in user-mapped pages, so skip any
953 * other types of pages.
954 */
955 if (PageReserved(p) || PageSlab(p))
956 return SWAP_SUCCESS;
957 if (!(PageLRU(hpage) || PageHuge(p)))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100958 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200959
Andi Kleen6a460792009-09-16 11:50:15 +0200960 /*
961 * This check implies we don't kill processes if their pages
962 * are in the swap cache early. Those are always late kills.
963 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900964 if (!page_mapped(hpage))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100965 return SWAP_SUCCESS;
966
Naoya Horiguchi52089b12014-07-30 16:08:30 -0700967 if (PageKsm(p)) {
968 pr_err("MCE %#lx: can't handle KSM pages.\n", pfn);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100969 return SWAP_FAIL;
Naoya Horiguchi52089b12014-07-30 16:08:30 -0700970 }
Andi Kleen6a460792009-09-16 11:50:15 +0200971
972 if (PageSwapCache(p)) {
973 printk(KERN_ERR
974 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
975 ttu |= TTU_IGNORE_HWPOISON;
976 }
977
978 /*
979 * Propagate the dirty bit from PTEs to struct page first, because we
980 * need this to decide if we should kill or just drop the page.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100981 * XXX: the dirty test could be racy: set_page_dirty() may not always
982 * be called inside page lock (it's recommended but not enforced).
Andi Kleen6a460792009-09-16 11:50:15 +0200983 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900984 mapping = page_mapping(hpage);
Tony Luck6751ed62012-07-11 10:20:47 -0700985 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900986 mapping_cap_writeback_dirty(mapping)) {
987 if (page_mkclean(hpage)) {
988 SetPageDirty(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +0200989 } else {
990 kill = 0;
991 ttu |= TTU_IGNORE_HWPOISON;
992 printk(KERN_INFO
993 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
994 pfn);
995 }
996 }
997
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800998 /*
Andi Kleen6a460792009-09-16 11:50:15 +0200999 * First collect all the processes that have the page
1000 * mapped in dirty form. This has to be done before try_to_unmap,
1001 * because ttu takes the rmap data structures down.
1002 *
1003 * Error handling: We ignore errors here because
1004 * there's nothing that can be done.
1005 */
1006 if (kill)
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001007 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
Andi Kleen6a460792009-09-16 11:50:15 +02001008
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001009 ret = try_to_unmap(hpage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +02001010 if (ret != SWAP_SUCCESS)
1011 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001012 pfn, page_mapcount(hpage));
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001013
Andi Kleen6a460792009-09-16 11:50:15 +02001014 /*
1015 * Now that the dirty bit has been propagated to the
1016 * struct page and all unmaps done we can decide if
1017 * killing is needed or not. Only kill when the page
Tony Luck6751ed62012-07-11 10:20:47 -07001018 * was dirty or the process is not restartable,
1019 * otherwise the tokill list is merely
Andi Kleen6a460792009-09-16 11:50:15 +02001020 * freed. When there was a problem unmapping earlier
1021 * use a more force-full uncatchable kill to prevent
1022 * any accesses to the poisoned memory.
1023 */
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001024 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
Tony Luck6751ed62012-07-11 10:20:47 -07001025 kill_procs(&tokill, forcekill, trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -08001026 ret != SWAP_SUCCESS, p, pfn, flags);
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001027
1028 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +02001029}
1030
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001031static void set_page_hwpoison_huge_page(struct page *hpage)
1032{
1033 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -07001034 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001035 for (i = 0; i < nr_pages; i++)
1036 SetPageHWPoison(hpage + i);
1037}
1038
1039static void clear_page_hwpoison_huge_page(struct page *hpage)
1040{
1041 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -07001042 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001043 for (i = 0; i < nr_pages; i++)
1044 ClearPageHWPoison(hpage + i);
1045}
1046
Tony Luckcd42f4a2011-12-15 10:48:12 -08001047/**
1048 * memory_failure - Handle memory failure of a page.
1049 * @pfn: Page Number of the corrupted page
1050 * @trapno: Trap number reported in the signal to user space.
1051 * @flags: fine tune action taken
1052 *
1053 * This function is called by the low level machine check code
1054 * of an architecture when it detects hardware memory corruption
1055 * of a page. It tries its best to recover, which includes
1056 * dropping pages, killing processes etc.
1057 *
1058 * The function is primarily of use for corruptions that
1059 * happen outside the current execution context (e.g. when
1060 * detected by a background scrubber)
1061 *
1062 * Must run in process context (e.g. a work queue) with interrupts
1063 * enabled and no spinlocks hold.
1064 */
1065int memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +02001066{
1067 struct page_state *ps;
1068 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001069 struct page *hpage;
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001070 struct page *orig_head;
Andi Kleen6a460792009-09-16 11:50:15 +02001071 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001072 unsigned int nr_pages;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001073 unsigned long page_flags;
Andi Kleen6a460792009-09-16 11:50:15 +02001074
1075 if (!sysctl_memory_failure_recovery)
1076 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1077
1078 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +01001079 printk(KERN_ERR
1080 "MCE %#lx: memory outside kernel control\n",
1081 pfn);
1082 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +02001083 }
1084
1085 p = pfn_to_page(pfn);
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001086 orig_head = hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +02001087 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001088 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001089 return 0;
1090 }
1091
Naoya Horiguchi4db0e952013-02-22 16:34:05 -08001092 /*
1093 * Currently errors on hugetlbfs pages are measured in hugepage units,
1094 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1095 * transparent hugepages, they are supposed to be split and error
1096 * measurement is done in normal page units. So nr_pages should be one
1097 * in this case.
1098 */
1099 if (PageHuge(p))
1100 nr_pages = 1 << compound_order(hpage);
1101 else /* normal page or thp */
1102 nr_pages = 1;
Xishi Qiu293c07e2013-02-22 16:34:02 -08001103 atomic_long_add(nr_pages, &num_poisoned_pages);
Andi Kleen6a460792009-09-16 11:50:15 +02001104
1105 /*
1106 * We need/can do nothing about count=0 pages.
1107 * 1) it's a free page, and therefore in safe hand:
1108 * prep_new_page() will be the gate keeper.
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001109 * 2) it's a free hugepage, which is also safe:
1110 * an affected hugepage will be dequeued from hugepage freelist,
1111 * so there's no concern about reusing it ever after.
1112 * 3) it's part of a non-compound high order page.
Andi Kleen6a460792009-09-16 11:50:15 +02001113 * Implies some kernel user: cannot stop them from
1114 * R/W the page; let's pray that the page has been
1115 * used and will be freed some time later.
1116 * In fact it's dangerous to directly bump up page count from 0,
1117 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1118 */
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001119 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001120 if (is_free_buddy_page(p)) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001121 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001122 return 0;
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001123 } else if (PageHuge(hpage)) {
1124 /*
Chen Yucongb9851942014-05-22 11:54:15 -07001125 * Check "filter hit" and "race with other subpage."
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001126 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001127 lock_page(hpage);
Chen Yucongb9851942014-05-22 11:54:15 -07001128 if (PageHWPoison(hpage)) {
1129 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1130 || (p != hpage && TestSetPageHWPoison(hpage))) {
1131 atomic_long_sub(nr_pages, &num_poisoned_pages);
1132 unlock_page(hpage);
1133 return 0;
1134 }
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001135 }
1136 set_page_hwpoison_huge_page(hpage);
1137 res = dequeue_hwpoisoned_huge_page(hpage);
Xie XiuQicc637b12015-06-24 16:57:30 -07001138 action_result(pfn, MF_MSG_FREE_HUGE,
1139 res ? MF_IGNORED : MF_DELAYED);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001140 unlock_page(hpage);
1141 return res;
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001142 } else {
Xie XiuQicc637b12015-06-24 16:57:30 -07001143 action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001144 return -EBUSY;
1145 }
Andi Kleen6a460792009-09-16 11:50:15 +02001146 }
1147
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001148 if (!PageHuge(p) && PageTransHuge(hpage)) {
Wanpeng Li7f6bf392015-08-14 15:35:08 -07001149 if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) {
1150 if (!PageAnon(hpage))
1151 pr_err("MCE: %#lx: non anonymous thp\n", pfn);
1152 else
1153 pr_err("MCE: %#lx: thp split failed\n", pfn);
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001154 if (TestClearPageHWPoison(p))
1155 atomic_long_sub(nr_pages, &num_poisoned_pages);
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001156 put_page(p);
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001157 if (p != hpage)
1158 put_page(hpage);
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001159 return -EBUSY;
1160 }
1161 VM_BUG_ON_PAGE(!page_count(p), p);
1162 hpage = compound_head(p);
1163 }
1164
Andi Kleen6a460792009-09-16 11:50:15 +02001165 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001166 * We ignore non-LRU pages for good reasons.
1167 * - PG_locked is only well defined for LRU pages and a few others
1168 * - to avoid races with __set_page_locked()
1169 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1170 * The check (unnecessarily) ignores LRU pages being isolated and
1171 * walked by the page reclaim code, however that's not a big loss.
1172 */
Naoya Horiguchi09789e52015-05-05 16:23:35 -07001173 if (!PageHuge(p)) {
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001174 if (!PageLRU(p))
1175 shake_page(p, 0);
1176 if (!PageLRU(p)) {
Jin Dongmingaf241a02011-02-01 15:52:41 -08001177 /*
1178 * shake_page could have turned it free.
1179 */
1180 if (is_free_buddy_page(p)) {
Wanpeng Li2d421ac2013-09-30 13:45:23 -07001181 if (flags & MF_COUNT_INCREASED)
Xie XiuQicc637b12015-06-24 16:57:30 -07001182 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
Wanpeng Li2d421ac2013-09-30 13:45:23 -07001183 else
Xie XiuQicc637b12015-06-24 16:57:30 -07001184 action_result(pfn, MF_MSG_BUDDY_2ND,
1185 MF_DELAYED);
Jin Dongmingaf241a02011-02-01 15:52:41 -08001186 return 0;
1187 }
Andi Kleen0474a602009-12-16 12:20:00 +01001188 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001189 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001190
Jens Axboe7eaceac2011-03-10 08:52:07 +01001191 lock_page(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001192
1193 /*
Andi Kleenf37d4292014-08-06 16:06:49 -07001194 * The page could have changed compound pages during the locking.
1195 * If this happens just bail out.
1196 */
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001197 if (PageCompound(p) && compound_head(p) != orig_head) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001198 action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
Andi Kleenf37d4292014-08-06 16:06:49 -07001199 res = -EBUSY;
1200 goto out;
1201 }
1202
1203 /*
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001204 * We use page flags to determine what action should be taken, but
1205 * the flags can be modified by the error containment action. One
1206 * example is an mlocked page, where PG_mlocked is cleared by
1207 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1208 * correctly, we save a copy of the page flags at this time.
1209 */
1210 page_flags = p->flags;
1211
1212 /*
Wu Fengguang847ce402009-12-16 12:19:58 +01001213 * unpoison always clear PG_hwpoison inside page lock
1214 */
1215 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001216 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Naoya Horiguchi3e030ec2014-05-22 11:54:21 -07001217 atomic_long_sub(nr_pages, &num_poisoned_pages);
Naoya Horiguchia09233f2015-08-06 15:46:58 -07001218 unlock_page(hpage);
Naoya Horiguchi3e030ec2014-05-22 11:54:21 -07001219 put_page(hpage);
Naoya Horiguchia09233f2015-08-06 15:46:58 -07001220 return 0;
Wu Fengguang847ce402009-12-16 12:19:58 +01001221 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001222 if (hwpoison_filter(p)) {
1223 if (TestClearPageHWPoison(p))
Xishi Qiu293c07e2013-02-22 16:34:02 -08001224 atomic_long_sub(nr_pages, &num_poisoned_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001225 unlock_page(hpage);
1226 put_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001227 return 0;
1228 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001229
Chen Yucong0bc1f8b2014-07-02 15:22:37 -07001230 if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p))
1231 goto identify_page_state;
1232
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001233 /*
1234 * For error on the tail page, we should set PG_hwpoison
1235 * on the head page to show that the hugepage is hwpoisoned
1236 */
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001237 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001238 action_result(pfn, MF_MSG_POISONED_HUGE, MF_IGNORED);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001239 unlock_page(hpage);
1240 put_page(hpage);
1241 return 0;
1242 }
1243 /*
1244 * Set PG_hwpoison on all pages in an error hugepage,
1245 * because containment is done in hugepage unit for now.
1246 * Since we have done TestSetPageHWPoison() for the head page with
1247 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1248 */
1249 if (PageHuge(p))
1250 set_page_hwpoison_huge_page(hpage);
1251
Naoya Horiguchi6edd6cc2014-06-04 16:10:35 -07001252 /*
1253 * It's very difficult to mess with pages currently under IO
1254 * and in many cases impossible, so we just avoid it here.
1255 */
Andi Kleen6a460792009-09-16 11:50:15 +02001256 wait_on_page_writeback(p);
1257
1258 /*
1259 * Now take care of user space mappings.
Minchan Kime64a7822011-03-22 16:32:44 -07001260 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001261 *
1262 * When the raw error page is thp tail page, hpage points to the raw
1263 * page after thp split.
Andi Kleen6a460792009-09-16 11:50:15 +02001264 */
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001265 if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)
1266 != SWAP_SUCCESS) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001267 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001268 res = -EBUSY;
1269 goto out;
1270 }
Andi Kleen6a460792009-09-16 11:50:15 +02001271
1272 /*
1273 * Torn down by someone else?
1274 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001275 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001276 action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001277 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001278 goto out;
1279 }
1280
Chen Yucong0bc1f8b2014-07-02 15:22:37 -07001281identify_page_state:
Andi Kleen6a460792009-09-16 11:50:15 +02001282 res = -EBUSY;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001283 /*
1284 * The first check uses the current page flags which may not have any
1285 * relevant information. The second check with the saved page flagss is
1286 * carried out only if the first check can't determine the page status.
1287 */
1288 for (ps = error_states;; ps++)
1289 if ((p->flags & ps->mask) == ps->res)
Andi Kleen6a460792009-09-16 11:50:15 +02001290 break;
Wanpeng Li841fcc52013-09-11 14:22:50 -07001291
1292 page_flags |= (p->flags & (1UL << PG_dirty));
1293
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001294 if (!ps->mask)
1295 for (ps = error_states;; ps++)
1296 if ((page_flags & ps->mask) == ps->res)
1297 break;
1298 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001299out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001300 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001301 return res;
1302}
Tony Luckcd42f4a2011-12-15 10:48:12 -08001303EXPORT_SYMBOL_GPL(memory_failure);
Wu Fengguang847ce402009-12-16 12:19:58 +01001304
Huang Yingea8f5fb2011-07-13 13:14:27 +08001305#define MEMORY_FAILURE_FIFO_ORDER 4
1306#define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1307
1308struct memory_failure_entry {
1309 unsigned long pfn;
1310 int trapno;
1311 int flags;
1312};
1313
1314struct memory_failure_cpu {
1315 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1316 MEMORY_FAILURE_FIFO_SIZE);
1317 spinlock_t lock;
1318 struct work_struct work;
1319};
1320
1321static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1322
1323/**
1324 * memory_failure_queue - Schedule handling memory failure of a page.
1325 * @pfn: Page Number of the corrupted page
1326 * @trapno: Trap number reported in the signal to user space.
1327 * @flags: Flags for memory failure handling
1328 *
1329 * This function is called by the low level hardware error handler
1330 * when it detects hardware memory corruption of a page. It schedules
1331 * the recovering of error page, including dropping pages, killing
1332 * processes etc.
1333 *
1334 * The function is primarily of use for corruptions that
1335 * happen outside the current execution context (e.g. when
1336 * detected by a background scrubber)
1337 *
1338 * Can run in IRQ context.
1339 */
1340void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1341{
1342 struct memory_failure_cpu *mf_cpu;
1343 unsigned long proc_flags;
1344 struct memory_failure_entry entry = {
1345 .pfn = pfn,
1346 .trapno = trapno,
1347 .flags = flags,
1348 };
1349
1350 mf_cpu = &get_cpu_var(memory_failure_cpu);
1351 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
Stefani Seibold498d3192013-11-14 14:32:17 -08001352 if (kfifo_put(&mf_cpu->fifo, entry))
Huang Yingea8f5fb2011-07-13 13:14:27 +08001353 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1354 else
Joe Perches8e33a522013-07-25 11:53:25 -07001355 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
Huang Yingea8f5fb2011-07-13 13:14:27 +08001356 pfn);
1357 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1358 put_cpu_var(memory_failure_cpu);
1359}
1360EXPORT_SYMBOL_GPL(memory_failure_queue);
1361
1362static void memory_failure_work_func(struct work_struct *work)
1363{
1364 struct memory_failure_cpu *mf_cpu;
1365 struct memory_failure_entry entry = { 0, };
1366 unsigned long proc_flags;
1367 int gotten;
1368
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001369 mf_cpu = this_cpu_ptr(&memory_failure_cpu);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001370 for (;;) {
1371 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1372 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1373 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1374 if (!gotten)
1375 break;
Naveen N. Raocf870c72013-07-10 14:57:01 +05301376 if (entry.flags & MF_SOFT_OFFLINE)
1377 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1378 else
1379 memory_failure(entry.pfn, entry.trapno, entry.flags);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001380 }
1381}
1382
1383static int __init memory_failure_init(void)
1384{
1385 struct memory_failure_cpu *mf_cpu;
1386 int cpu;
1387
1388 for_each_possible_cpu(cpu) {
1389 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1390 spin_lock_init(&mf_cpu->lock);
1391 INIT_KFIFO(mf_cpu->fifo);
1392 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1393 }
1394
1395 return 0;
1396}
1397core_initcall(memory_failure_init);
1398
Wu Fengguang847ce402009-12-16 12:19:58 +01001399/**
1400 * unpoison_memory - Unpoison a previously poisoned page
1401 * @pfn: Page number of the to be unpoisoned page
1402 *
1403 * Software-unpoison a page that has been poisoned by
1404 * memory_failure() earlier.
1405 *
1406 * This is only done on the software-level, so it only works
1407 * for linux injected failures, not real hardware failures
1408 *
1409 * Returns 0 for success, otherwise -errno.
1410 */
1411int unpoison_memory(unsigned long pfn)
1412{
1413 struct page *page;
1414 struct page *p;
1415 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001416 unsigned int nr_pages;
Wu Fengguang847ce402009-12-16 12:19:58 +01001417
1418 if (!pfn_valid(pfn))
1419 return -ENXIO;
1420
1421 p = pfn_to_page(pfn);
1422 page = compound_head(p);
1423
1424 if (!PageHWPoison(p)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001425 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001426 return 0;
1427 }
1428
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001429 /*
1430 * unpoison_memory() can encounter thp only when the thp is being
1431 * worked by memory_failure() and the page lock is not held yet.
1432 * In such case, we yield to memory_failure() and make unpoison fail.
1433 */
Wanpeng Lie76d30e2013-09-30 13:45:22 -07001434 if (!PageHuge(page) && PageTransHuge(page)) {
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001435 pr_info("MCE: Memory failure is now running on %#lx\n", pfn);
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001436 return 0;
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001437 }
1438
Wanpeng Lif9121152013-09-11 14:22:52 -07001439 nr_pages = 1 << compound_order(page);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001440
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001441 if (!get_hwpoison_page(p)) {
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001442 /*
1443 * Since HWPoisoned hugepage should have non-zero refcount,
1444 * race between memory failure and unpoison seems to happen.
1445 * In such case unpoison fails and memory failure runs
1446 * to the end.
1447 */
1448 if (PageHuge(page)) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001449 pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001450 return 0;
1451 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001452 if (TestClearPageHWPoison(p))
Wanpeng Lidd9538a2013-09-11 14:22:54 -07001453 atomic_long_dec(&num_poisoned_pages);
Andi Kleenfb46e732010-09-27 23:31:30 +02001454 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001455 return 0;
1456 }
1457
Jens Axboe7eaceac2011-03-10 08:52:07 +01001458 lock_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001459 /*
1460 * This test is racy because PG_hwpoison is set outside of page lock.
1461 * That's acceptable because that won't trigger kernel panic. Instead,
1462 * the PG_hwpoison page will be caught and isolated on the entrance to
1463 * the free buddy page pool.
1464 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001465 if (TestClearPageHWPoison(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001466 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
Xishi Qiu293c07e2013-02-22 16:34:02 -08001467 atomic_long_sub(nr_pages, &num_poisoned_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001468 freeit = 1;
Naoya Horiguchi6a901812010-09-08 10:19:40 +09001469 if (PageHuge(page))
1470 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001471 }
1472 unlock_page(page);
1473
1474 put_page(page);
Wanpeng Li3ba5eeb2013-09-11 14:23:01 -07001475 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
Wu Fengguang847ce402009-12-16 12:19:58 +01001476 put_page(page);
1477
1478 return 0;
1479}
1480EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001481
1482static struct page *new_page(struct page *p, unsigned long private, int **x)
1483{
Andi Kleen12686d12009-12-16 12:20:01 +01001484 int nid = page_to_nid(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001485 if (PageHuge(p))
1486 return alloc_huge_page_node(page_hstate(compound_head(p)),
1487 nid);
1488 else
1489 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001490}
1491
1492/*
1493 * Safely get reference count of an arbitrary page.
1494 * Returns 0 for a free page, -EIO for a zero refcount page
1495 * that is not free, and 1 for any other page type.
1496 * For 1 the page is returned with increased page count, otherwise not.
1497 */
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001498static int __get_any_page(struct page *p, unsigned long pfn, int flags)
Andi Kleenfacb6012009-12-16 12:20:00 +01001499{
1500 int ret;
1501
1502 if (flags & MF_COUNT_INCREASED)
1503 return 1;
1504
1505 /*
Naoya Horiguchid950b952010-09-08 10:19:39 +09001506 * When the target page is a free hugepage, just remove it
1507 * from free hugepage list.
1508 */
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001509 if (!get_hwpoison_page(p)) {
Naoya Horiguchid950b952010-09-08 10:19:39 +09001510 if (PageHuge(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001511 pr_info("%s: %#lx free huge page\n", __func__, pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001512 ret = 0;
Naoya Horiguchid950b952010-09-08 10:19:39 +09001513 } else if (is_free_buddy_page(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001514 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001515 ret = 0;
1516 } else {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001517 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1518 __func__, pfn, p->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001519 ret = -EIO;
1520 }
1521 } else {
1522 /* Not a free page */
1523 ret = 1;
1524 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001525 return ret;
1526}
1527
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001528static int get_any_page(struct page *page, unsigned long pfn, int flags)
1529{
1530 int ret = __get_any_page(page, pfn, flags);
1531
1532 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1533 /*
1534 * Try to free it.
1535 */
1536 put_page(page);
1537 shake_page(page, 1);
1538
1539 /*
1540 * Did it turn free?
1541 */
1542 ret = __get_any_page(page, pfn, 0);
1543 if (!PageLRU(page)) {
Wanpeng Li4f32be62015-08-14 15:34:56 -07001544 /* Drop page reference which is from __get_any_page() */
1545 put_page(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001546 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1547 pfn, page->flags);
1548 return -EIO;
1549 }
1550 }
1551 return ret;
1552}
1553
Naoya Horiguchid950b952010-09-08 10:19:39 +09001554static int soft_offline_huge_page(struct page *page, int flags)
1555{
1556 int ret;
1557 unsigned long pfn = page_to_pfn(page);
1558 struct page *hpage = compound_head(page);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001559 LIST_HEAD(pagelist);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001560
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001561 /*
1562 * This double-check of PageHWPoison is to avoid the race with
1563 * memory_failure(). See also comment in __soft_offline_page().
1564 */
1565 lock_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001566 if (PageHWPoison(hpage)) {
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001567 unlock_page(hpage);
1568 put_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001569 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001570 return -EBUSY;
Xishi Qiu0ebff322013-02-22 16:33:59 -08001571 }
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001572 unlock_page(hpage);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001573
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001574 ret = isolate_huge_page(hpage, &pagelist);
Wanpeng Li03613802015-08-14 15:34:59 -07001575 /*
1576 * get_any_page() and isolate_huge_page() takes a refcount each,
1577 * so need to drop one here.
1578 */
1579 put_page(hpage);
1580 if (!ret) {
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001581 pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn);
1582 return -EBUSY;
1583 }
1584
David Rientjes68711a72014-06-04 16:08:25 -07001585 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001586 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001587 if (ret) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001588 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1589 pfn, ret, page->flags);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001590 /*
1591 * We know that soft_offline_huge_page() tries to migrate
1592 * only one hugepage pointed to by hpage, so we need not
1593 * run through the pagelist here.
1594 */
1595 putback_active_hugepage(hpage);
1596 if (ret > 0)
1597 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001598 } else {
Jianguo Wua49ecbc2013-12-18 17:08:54 -08001599 /* overcommit hugetlb page will be freed to buddy */
1600 if (PageHuge(page)) {
1601 set_page_hwpoison_huge_page(hpage);
1602 dequeue_hwpoisoned_huge_page(hpage);
1603 atomic_long_add(1 << compound_order(hpage),
1604 &num_poisoned_pages);
1605 } else {
1606 SetPageHWPoison(page);
1607 atomic_long_inc(&num_poisoned_pages);
1608 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001609 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001610 return ret;
1611}
1612
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001613static int __soft_offline_page(struct page *page, int flags)
1614{
1615 int ret;
1616 unsigned long pfn = page_to_pfn(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001617
1618 /*
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001619 * Check PageHWPoison again inside page lock because PageHWPoison
1620 * is set by memory_failure() outside page lock. Note that
1621 * memory_failure() also double-checks PageHWPoison inside page lock,
1622 * so there's no race between soft_offline_page() and memory_failure().
Andi Kleenfacb6012009-12-16 12:20:00 +01001623 */
Xishi Qiu0ebff322013-02-22 16:33:59 -08001624 lock_page(page);
1625 wait_on_page_writeback(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001626 if (PageHWPoison(page)) {
1627 unlock_page(page);
1628 put_page(page);
1629 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1630 return -EBUSY;
1631 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001632 /*
1633 * Try to invalidate first. This should work for
1634 * non dirty unmapped page cache pages.
1635 */
1636 ret = invalidate_inode_page(page);
1637 unlock_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001638 /*
Andi Kleenfacb6012009-12-16 12:20:00 +01001639 * RED-PEN would be better to keep it isolated here, but we
1640 * would need to fix isolation locking first.
1641 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001642 if (ret == 1) {
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001643 put_page(page);
Andi Kleenfb46e732010-09-27 23:31:30 +02001644 pr_info("soft_offline: %#lx: invalidated\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001645 SetPageHWPoison(page);
1646 atomic_long_inc(&num_poisoned_pages);
1647 return 0;
Andi Kleenfacb6012009-12-16 12:20:00 +01001648 }
1649
1650 /*
1651 * Simple invalidation didn't work.
1652 * Try to migrate to a new page instead. migrate.c
1653 * handles a large number of cases for us.
1654 */
1655 ret = isolate_lru_page(page);
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001656 /*
1657 * Drop page reference which is came from get_any_page()
1658 * successful isolate_lru_page() already took another one.
1659 */
1660 put_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001661 if (!ret) {
1662 LIST_HEAD(pagelist);
Minchan Kim5db8a732011-06-15 15:08:48 -07001663 inc_zone_page_state(page, NR_ISOLATED_ANON +
Hugh Dickins9c620e22013-02-22 16:35:14 -08001664 page_is_file_cache(page));
Andi Kleenfacb6012009-12-16 12:20:00 +01001665 list_add(&page->lru, &pagelist);
Naoya Horiguchi4491f712015-08-06 15:47:11 -07001666 if (!TestSetPageHWPoison(page))
1667 atomic_long_inc(&num_poisoned_pages);
David Rientjes68711a72014-06-04 16:08:25 -07001668 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001669 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Andi Kleenfacb6012009-12-16 12:20:00 +01001670 if (ret) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001671 if (!list_empty(&pagelist)) {
1672 list_del(&page->lru);
1673 dec_zone_page_state(page, NR_ISOLATED_ANON +
1674 page_is_file_cache(page));
1675 putback_lru_page(page);
1676 }
1677
Andi Kleenfb46e732010-09-27 23:31:30 +02001678 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001679 pfn, ret, page->flags);
1680 if (ret > 0)
1681 ret = -EIO;
Naoya Horiguchi4491f712015-08-06 15:47:11 -07001682 if (TestClearPageHWPoison(page))
1683 atomic_long_dec(&num_poisoned_pages);
Andi Kleenfacb6012009-12-16 12:20:00 +01001684 }
1685 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001686 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
Dean Nelsondd73e852011-10-31 17:09:04 -07001687 pfn, ret, page_count(page), page->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001688 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001689 return ret;
1690}
Wanpeng Li86e05772013-09-11 14:22:56 -07001691
1692/**
1693 * soft_offline_page - Soft offline a page.
1694 * @page: page to offline
1695 * @flags: flags. Same as memory_failure().
1696 *
1697 * Returns 0 on success, otherwise negated errno.
1698 *
1699 * Soft offline a page, by migration or invalidation,
1700 * without killing anything. This is for the case when
1701 * a page is not corrupted yet (so it's still valid to access),
1702 * but has had a number of corrected errors and is better taken
1703 * out.
1704 *
1705 * The actual policy on when to do that is maintained by
1706 * user space.
1707 *
1708 * This should never impact any application or cause data loss,
1709 * however it might take some time.
1710 *
1711 * This is not a 100% solution for all memory, but tries to be
1712 * ``good enough'' for the majority of memory.
1713 */
1714int soft_offline_page(struct page *page, int flags)
1715{
1716 int ret;
1717 unsigned long pfn = page_to_pfn(page);
David Rientjes668f9abb2014-03-03 15:38:18 -08001718 struct page *hpage = compound_head(page);
Wanpeng Li86e05772013-09-11 14:22:56 -07001719
1720 if (PageHWPoison(page)) {
1721 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1722 return -EBUSY;
1723 }
1724 if (!PageHuge(page) && PageTransHuge(hpage)) {
1725 if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1726 pr_info("soft offline: %#lx: failed to split THP\n",
1727 pfn);
Wanpeng Li7d1900c2015-09-08 15:03:10 -07001728 if (flags & MF_COUNT_INCREASED)
1729 put_page(page);
Wanpeng Li86e05772013-09-11 14:22:56 -07001730 return -EBUSY;
1731 }
1732 }
1733
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07001734 get_online_mems();
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001735
Wanpeng Li86e05772013-09-11 14:22:56 -07001736 ret = get_any_page(page, pfn, flags);
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07001737 put_online_mems();
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001738 if (ret > 0) { /* for in-use pages */
Wanpeng Li86e05772013-09-11 14:22:56 -07001739 if (PageHuge(page))
1740 ret = soft_offline_huge_page(page, flags);
1741 else
1742 ret = __soft_offline_page(page, flags);
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001743 } else if (ret == 0) { /* for free pages */
Wanpeng Li86e05772013-09-11 14:22:56 -07001744 if (PageHuge(page)) {
1745 set_page_hwpoison_huge_page(hpage);
Naoya Horiguchi602498f2015-05-05 16:23:46 -07001746 if (!dequeue_hwpoisoned_huge_page(hpage))
1747 atomic_long_add(1 << compound_order(hpage),
Wanpeng Li86e05772013-09-11 14:22:56 -07001748 &num_poisoned_pages);
1749 } else {
Naoya Horiguchi602498f2015-05-05 16:23:46 -07001750 if (!TestSetPageHWPoison(page))
1751 atomic_long_inc(&num_poisoned_pages);
Wanpeng Li86e05772013-09-11 14:22:56 -07001752 }
1753 }
Wanpeng Li86e05772013-09-11 14:22:56 -07001754 return ret;
1755}