blob: d9359b770cd96b37ee945553a4230edbfd720e46 [file] [log] [blame]
Andi Kleen6a460792009-09-16 11:50:15 +02001/*
2 * Copyright (C) 2008, 2009 Intel Corporation
3 * Authors: Andi Kleen, Fengguang Wu
4 *
5 * This software may be redistributed and/or modified under the terms of
6 * the GNU General Public License ("GPL") version 2 only as published by the
7 * Free Software Foundation.
8 *
9 * High level machine check handler. Handles pages reported by the
Andi Kleen1c80b992010-09-27 23:09:51 +020010 * hardware as being corrupted usually due to a multi-bit ECC memory or cache
Andi Kleen6a460792009-09-16 11:50:15 +020011 * failure.
Andi Kleen1c80b992010-09-27 23:09:51 +020012 *
13 * In addition there is a "soft offline" entry point that allows stop using
14 * not-yet-corrupted-by-suspicious pages without killing anything.
Andi Kleen6a460792009-09-16 11:50:15 +020015 *
16 * Handles page cache pages in various states. The tricky part
Andi Kleen1c80b992010-09-27 23:09:51 +020017 * here is that we can access any page asynchronously in respect to
18 * other VM users, because memory failures could happen anytime and
19 * anywhere. This could violate some of their assumptions. This is why
20 * this code has to be extremely careful. Generally it tries to use
21 * normal locking rules, as in get the standard locks, even if that means
22 * the error handling takes potentially a long time.
23 *
24 * There are several operations here with exponential complexity because
25 * of unsuitable VM data structures. For example the operation to map back
26 * from RMAP chains to processes has to walk the complete process list and
27 * has non linear complexity with the number. But since memory corruptions
28 * are rare we hope to get away with this. This avoids impacting the core
29 * VM.
Andi Kleen6a460792009-09-16 11:50:15 +020030 */
31
32/*
33 * Notebook:
34 * - hugetlb needs more code
35 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
36 * - pass bad pages to kdump next kernel
37 */
Andi Kleen6a460792009-09-16 11:50:15 +020038#include <linux/kernel.h>
39#include <linux/mm.h>
40#include <linux/page-flags.h>
Wu Fengguang478c5ff2009-12-16 12:19:59 +010041#include <linux/kernel-page-flags.h>
Andi Kleen6a460792009-09-16 11:50:15 +020042#include <linux/sched.h>
Hugh Dickins01e00f82009-10-13 15:02:11 +010043#include <linux/ksm.h>
Andi Kleen6a460792009-09-16 11:50:15 +020044#include <linux/rmap.h>
Paul Gortmakerb9e15ba2011-05-26 16:00:52 -040045#include <linux/export.h>
Andi Kleen6a460792009-09-16 11:50:15 +020046#include <linux/pagemap.h>
47#include <linux/swap.h>
48#include <linux/backing-dev.h>
Andi Kleenfacb6012009-12-16 12:20:00 +010049#include <linux/migrate.h>
50#include <linux/page-isolation.h>
51#include <linux/suspend.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090052#include <linux/slab.h>
Huang Yingbf998152010-05-31 14:28:19 +080053#include <linux/swapops.h>
Naoya Horiguchi7af446a2010-05-28 09:29:17 +090054#include <linux/hugetlb.h>
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -080055#include <linux/memory_hotplug.h>
Minchan Kim5db8a732011-06-15 15:08:48 -070056#include <linux/mm_inline.h>
Huang Yingea8f5fb2011-07-13 13:14:27 +080057#include <linux/kfifo.h>
Andi Kleen6a460792009-09-16 11:50:15 +020058#include "internal.h"
59
60int sysctl_memory_failure_early_kill __read_mostly = 0;
61
62int sysctl_memory_failure_recovery __read_mostly = 1;
63
Xishi Qiu293c07e2013-02-22 16:34:02 -080064atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
Andi Kleen6a460792009-09-16 11:50:15 +020065
Andi Kleen27df5062009-12-21 19:56:42 +010066#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
67
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010068u32 hwpoison_filter_enable = 0;
Wu Fengguang7c116f22009-12-16 12:19:59 +010069u32 hwpoison_filter_dev_major = ~0U;
70u32 hwpoison_filter_dev_minor = ~0U;
Wu Fengguang478c5ff2009-12-16 12:19:59 +010071u64 hwpoison_filter_flags_mask;
72u64 hwpoison_filter_flags_value;
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010073EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
Wu Fengguang7c116f22009-12-16 12:19:59 +010074EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
75EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
Wu Fengguang478c5ff2009-12-16 12:19:59 +010076EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
77EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
Wu Fengguang7c116f22009-12-16 12:19:59 +010078
79static int hwpoison_filter_dev(struct page *p)
80{
81 struct address_space *mapping;
82 dev_t dev;
83
84 if (hwpoison_filter_dev_major == ~0U &&
85 hwpoison_filter_dev_minor == ~0U)
86 return 0;
87
88 /*
Andi Kleen1c80b992010-09-27 23:09:51 +020089 * page_mapping() does not accept slab pages.
Wu Fengguang7c116f22009-12-16 12:19:59 +010090 */
91 if (PageSlab(p))
92 return -EINVAL;
93
94 mapping = page_mapping(p);
95 if (mapping == NULL || mapping->host == NULL)
96 return -EINVAL;
97
98 dev = mapping->host->i_sb->s_dev;
99 if (hwpoison_filter_dev_major != ~0U &&
100 hwpoison_filter_dev_major != MAJOR(dev))
101 return -EINVAL;
102 if (hwpoison_filter_dev_minor != ~0U &&
103 hwpoison_filter_dev_minor != MINOR(dev))
104 return -EINVAL;
105
106 return 0;
107}
108
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100109static int hwpoison_filter_flags(struct page *p)
110{
111 if (!hwpoison_filter_flags_mask)
112 return 0;
113
114 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
115 hwpoison_filter_flags_value)
116 return 0;
117 else
118 return -EINVAL;
119}
120
Andi Kleen4fd466e2009-12-16 12:19:59 +0100121/*
122 * This allows stress tests to limit test scope to a collection of tasks
123 * by putting them under some memcg. This prevents killing unrelated/important
124 * processes such as /sbin/init. Note that the target task may share clean
125 * pages with init (eg. libc text), which is harmless. If the target task
126 * share _dirty_ pages with another task B, the test scheme must make sure B
127 * is also included in the memcg. At last, due to race conditions this filter
128 * can only guarantee that the page either belongs to the memcg tasks, or is
129 * a freed page.
130 */
Andrew Mortonc255a452012-07-31 16:43:02 -0700131#ifdef CONFIG_MEMCG_SWAP
Andi Kleen4fd466e2009-12-16 12:19:59 +0100132u64 hwpoison_filter_memcg;
133EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
134static int hwpoison_filter_task(struct page *p)
135{
136 struct mem_cgroup *mem;
137 struct cgroup_subsys_state *css;
138 unsigned long ino;
139
140 if (!hwpoison_filter_memcg)
141 return 0;
142
143 mem = try_get_mem_cgroup_from_page(p);
144 if (!mem)
145 return -EINVAL;
146
147 css = mem_cgroup_css(mem);
Tejun Heob1664922014-02-11 11:52:49 -0500148 ino = cgroup_ino(css->cgroup);
Andi Kleen4fd466e2009-12-16 12:19:59 +0100149 css_put(css);
150
Zefan Lif29374b2014-09-19 16:29:31 +0800151 if (ino != hwpoison_filter_memcg)
Andi Kleen4fd466e2009-12-16 12:19:59 +0100152 return -EINVAL;
153
154 return 0;
155}
156#else
157static int hwpoison_filter_task(struct page *p) { return 0; }
158#endif
159
Wu Fengguang7c116f22009-12-16 12:19:59 +0100160int hwpoison_filter(struct page *p)
161{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100162 if (!hwpoison_filter_enable)
163 return 0;
164
Wu Fengguang7c116f22009-12-16 12:19:59 +0100165 if (hwpoison_filter_dev(p))
166 return -EINVAL;
167
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100168 if (hwpoison_filter_flags(p))
169 return -EINVAL;
170
Andi Kleen4fd466e2009-12-16 12:19:59 +0100171 if (hwpoison_filter_task(p))
172 return -EINVAL;
173
Wu Fengguang7c116f22009-12-16 12:19:59 +0100174 return 0;
175}
Andi Kleen27df5062009-12-21 19:56:42 +0100176#else
177int hwpoison_filter(struct page *p)
178{
179 return 0;
180}
181#endif
182
Wu Fengguang7c116f22009-12-16 12:19:59 +0100183EXPORT_SYMBOL_GPL(hwpoison_filter);
184
Andi Kleen6a460792009-09-16 11:50:15 +0200185/*
Tony Luck7329bbe2011-12-13 09:27:58 -0800186 * Send all the processes who have the page mapped a signal.
187 * ``action optional'' if they are not immediately affected by the error
188 * ``action required'' if error happened in current execution context
Andi Kleen6a460792009-09-16 11:50:15 +0200189 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800190static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
191 unsigned long pfn, struct page *page, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200192{
193 struct siginfo si;
194 int ret;
195
196 printk(KERN_ERR
Tony Luck7329bbe2011-12-13 09:27:58 -0800197 "MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200198 pfn, t->comm, t->pid);
199 si.si_signo = SIGBUS;
200 si.si_errno = 0;
Andi Kleen6a460792009-09-16 11:50:15 +0200201 si.si_addr = (void *)addr;
202#ifdef __ARCH_SI_TRAPNO
203 si.si_trapno = trapno;
204#endif
Wanpeng Lif9121152013-09-11 14:22:52 -0700205 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
Tony Luck7329bbe2011-12-13 09:27:58 -0800206
Tony Lucka70ffca2014-06-04 16:10:59 -0700207 if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
Tony Luck7329bbe2011-12-13 09:27:58 -0800208 si.si_code = BUS_MCEERR_AR;
Tony Lucka70ffca2014-06-04 16:10:59 -0700209 ret = force_sig_info(SIGBUS, &si, current);
Tony Luck7329bbe2011-12-13 09:27:58 -0800210 } else {
211 /*
212 * Don't use force here, it's convenient if the signal
213 * can be temporarily blocked.
214 * This could cause a loop when the user sets SIGBUS
215 * to SIG_IGN, but hopefully no one will do that?
216 */
217 si.si_code = BUS_MCEERR_AO;
218 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
219 }
Andi Kleen6a460792009-09-16 11:50:15 +0200220 if (ret < 0)
221 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
222 t->comm, t->pid, ret);
223 return ret;
224}
225
226/*
Andi Kleen588f9ce2009-12-16 12:19:57 +0100227 * When a unknown page type is encountered drain as many buffers as possible
228 * in the hope to turn the page into a LRU or free page, which we can handle.
229 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100230void shake_page(struct page *p, int access)
Andi Kleen588f9ce2009-12-16 12:19:57 +0100231{
232 if (!PageSlab(p)) {
233 lru_add_drain_all();
234 if (PageLRU(p))
235 return;
Vlastimil Babkac0554322014-12-10 15:43:10 -0800236 drain_all_pages(page_zone(p));
Andi Kleen588f9ce2009-12-16 12:19:57 +0100237 if (PageLRU(p) || is_free_buddy_page(p))
238 return;
239 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100240
Andi Kleen588f9ce2009-12-16 12:19:57 +0100241 /*
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800242 * Only call shrink_node_slabs here (which would also shrink
243 * other caches) if access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100244 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800245 if (access)
246 drop_slab_node(page_to_nid(p));
Andi Kleen588f9ce2009-12-16 12:19:57 +0100247}
248EXPORT_SYMBOL_GPL(shake_page);
249
250/*
Andi Kleen6a460792009-09-16 11:50:15 +0200251 * Kill all processes that have a poisoned page mapped and then isolate
252 * the page.
253 *
254 * General strategy:
255 * Find all processes having the page mapped and kill them.
256 * But we keep a page reference around so that the page is not
257 * actually freed yet.
258 * Then stash the page away
259 *
260 * There's no convenient way to get back to mapped processes
261 * from the VMAs. So do a brute-force search over all
262 * running processes.
263 *
264 * Remember that machine checks are not common (or rather
265 * if they are common you have other problems), so this shouldn't
266 * be a performance issue.
267 *
268 * Also there are some races possible while we get from the
269 * error detection to actually handle it.
270 */
271
272struct to_kill {
273 struct list_head nd;
274 struct task_struct *tsk;
275 unsigned long addr;
Andi Kleen9033ae12010-09-27 23:36:05 +0200276 char addr_valid;
Andi Kleen6a460792009-09-16 11:50:15 +0200277};
278
279/*
280 * Failure handling: if we can't find or can't kill a process there's
281 * not much we can do. We just print a message and ignore otherwise.
282 */
283
284/*
285 * Schedule a process for later kill.
286 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
287 * TBD would GFP_NOIO be enough?
288 */
289static void add_to_kill(struct task_struct *tsk, struct page *p,
290 struct vm_area_struct *vma,
291 struct list_head *to_kill,
292 struct to_kill **tkc)
293{
294 struct to_kill *tk;
295
296 if (*tkc) {
297 tk = *tkc;
298 *tkc = NULL;
299 } else {
300 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
301 if (!tk) {
302 printk(KERN_ERR
303 "MCE: Out of memory while machine check handling\n");
304 return;
305 }
306 }
307 tk->addr = page_address_in_vma(p, vma);
308 tk->addr_valid = 1;
309
310 /*
311 * In theory we don't have to kill when the page was
312 * munmaped. But it could be also a mremap. Since that's
313 * likely very rare kill anyways just out of paranoia, but use
314 * a SIGKILL because the error is not contained anymore.
315 */
316 if (tk->addr == -EFAULT) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200317 pr_info("MCE: Unable to find user space address %lx in %s\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200318 page_to_pfn(p), tsk->comm);
319 tk->addr_valid = 0;
320 }
321 get_task_struct(tsk);
322 tk->tsk = tsk;
323 list_add_tail(&tk->nd, to_kill);
324}
325
326/*
327 * Kill the processes that have been collected earlier.
328 *
329 * Only do anything when DOIT is set, otherwise just free the list
330 * (this is used for clean pages which do not need killing)
331 * Also when FAIL is set do a force kill because something went
332 * wrong earlier.
333 */
Tony Luck6751ed62012-07-11 10:20:47 -0700334static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -0800335 int fail, struct page *page, unsigned long pfn,
336 int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200337{
338 struct to_kill *tk, *next;
339
340 list_for_each_entry_safe (tk, next, to_kill, nd) {
Tony Luck6751ed62012-07-11 10:20:47 -0700341 if (forcekill) {
Andi Kleen6a460792009-09-16 11:50:15 +0200342 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200343 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200344 * make sure the process doesn't catch the
345 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200346 */
347 if (fail || tk->addr_valid == 0) {
348 printk(KERN_ERR
349 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
350 pfn, tk->tsk->comm, tk->tsk->pid);
351 force_sig(SIGKILL, tk->tsk);
352 }
353
354 /*
355 * In theory the process could have mapped
356 * something else on the address in-between. We could
357 * check for that, but we need to tell the
358 * process anyways.
359 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800360 else if (kill_proc(tk->tsk, tk->addr, trapno,
361 pfn, page, flags) < 0)
Andi Kleen6a460792009-09-16 11:50:15 +0200362 printk(KERN_ERR
363 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
364 pfn, tk->tsk->comm, tk->tsk->pid);
365 }
366 put_task_struct(tk->tsk);
367 kfree(tk);
368 }
369}
370
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700371/*
372 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
373 * on behalf of the thread group. Return task_struct of the (first found)
374 * dedicated thread if found, and return NULL otherwise.
375 *
376 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
377 * have to call rcu_read_lock/unlock() in this function.
378 */
379static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
Andi Kleen6a460792009-09-16 11:50:15 +0200380{
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700381 struct task_struct *t;
382
383 for_each_thread(tsk, t)
384 if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
385 return t;
386 return NULL;
387}
388
389/*
390 * Determine whether a given process is "early kill" process which expects
391 * to be signaled when some page under the process is hwpoisoned.
392 * Return task_struct of the dedicated thread (main thread unless explicitly
393 * specified) if the process is "early kill," and otherwise returns NULL.
394 */
395static struct task_struct *task_early_kill(struct task_struct *tsk,
396 int force_early)
397{
398 struct task_struct *t;
Andi Kleen6a460792009-09-16 11:50:15 +0200399 if (!tsk->mm)
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700400 return NULL;
Tony Luck74614de2014-06-04 16:11:01 -0700401 if (force_early)
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700402 return tsk;
403 t = find_early_kill_thread(tsk);
404 if (t)
405 return t;
406 if (sysctl_memory_failure_early_kill)
407 return tsk;
408 return NULL;
Andi Kleen6a460792009-09-16 11:50:15 +0200409}
410
411/*
412 * Collect processes when the error hit an anonymous page.
413 */
414static void collect_procs_anon(struct page *page, struct list_head *to_kill,
Tony Luck74614de2014-06-04 16:11:01 -0700415 struct to_kill **tkc, int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200416{
417 struct vm_area_struct *vma;
418 struct task_struct *tsk;
419 struct anon_vma *av;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700420 pgoff_t pgoff;
Andi Kleen6a460792009-09-16 11:50:15 +0200421
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000422 av = page_lock_anon_vma_read(page);
Andi Kleen6a460792009-09-16 11:50:15 +0200423 if (av == NULL) /* Not actually mapped anymore */
Peter Zijlstra9b679322011-06-27 16:18:09 -0700424 return;
425
Naoya Horiguchia0f7a752014-07-23 14:00:01 -0700426 pgoff = page_to_pgoff(page);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700427 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200428 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800429 struct anon_vma_chain *vmac;
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700430 struct task_struct *t = task_early_kill(tsk, force_early);
Rik van Riel5beb4932010-03-05 13:42:07 -0800431
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700432 if (!t)
Andi Kleen6a460792009-09-16 11:50:15 +0200433 continue;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700434 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
435 pgoff, pgoff) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800436 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200437 if (!page_mapped_in_vma(page, vma))
438 continue;
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700439 if (vma->vm_mm == t->mm)
440 add_to_kill(t, page, vma, to_kill, tkc);
Andi Kleen6a460792009-09-16 11:50:15 +0200441 }
442 }
Andi Kleen6a460792009-09-16 11:50:15 +0200443 read_unlock(&tasklist_lock);
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000444 page_unlock_anon_vma_read(av);
Andi Kleen6a460792009-09-16 11:50:15 +0200445}
446
447/*
448 * Collect processes when the error hit a file mapped page.
449 */
450static void collect_procs_file(struct page *page, struct list_head *to_kill,
Tony Luck74614de2014-06-04 16:11:01 -0700451 struct to_kill **tkc, int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200452{
453 struct vm_area_struct *vma;
454 struct task_struct *tsk;
Andi Kleen6a460792009-09-16 11:50:15 +0200455 struct address_space *mapping = page->mapping;
456
Davidlohr Buesod28eb9c2014-12-12 16:54:36 -0800457 i_mmap_lock_read(mapping);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700458 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200459 for_each_process(tsk) {
Naoya Horiguchia0f7a752014-07-23 14:00:01 -0700460 pgoff_t pgoff = page_to_pgoff(page);
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700461 struct task_struct *t = task_early_kill(tsk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200462
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700463 if (!t)
Andi Kleen6a460792009-09-16 11:50:15 +0200464 continue;
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -0700465 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
Andi Kleen6a460792009-09-16 11:50:15 +0200466 pgoff) {
467 /*
468 * Send early kill signal to tasks where a vma covers
469 * the page but the corrupted page is not necessarily
470 * mapped it in its pte.
471 * Assume applications who requested early kill want
472 * to be informed of all such data corruptions.
473 */
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700474 if (vma->vm_mm == t->mm)
475 add_to_kill(t, page, vma, to_kill, tkc);
Andi Kleen6a460792009-09-16 11:50:15 +0200476 }
477 }
Andi Kleen6a460792009-09-16 11:50:15 +0200478 read_unlock(&tasklist_lock);
Davidlohr Buesod28eb9c2014-12-12 16:54:36 -0800479 i_mmap_unlock_read(mapping);
Andi Kleen6a460792009-09-16 11:50:15 +0200480}
481
482/*
483 * Collect the processes who have the corrupted page mapped to kill.
484 * This is done in two steps for locking reasons.
485 * First preallocate one tokill structure outside the spin locks,
486 * so that we can kill at least one process reasonably reliable.
487 */
Tony Luck74614de2014-06-04 16:11:01 -0700488static void collect_procs(struct page *page, struct list_head *tokill,
489 int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200490{
491 struct to_kill *tk;
492
493 if (!page->mapping)
494 return;
495
496 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
497 if (!tk)
498 return;
499 if (PageAnon(page))
Tony Luck74614de2014-06-04 16:11:01 -0700500 collect_procs_anon(page, tokill, &tk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200501 else
Tony Luck74614de2014-06-04 16:11:01 -0700502 collect_procs_file(page, tokill, &tk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200503 kfree(tk);
504}
505
506/*
507 * Error handlers for various types of pages.
508 */
509
510enum outcome {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100511 IGNORED, /* Error: cannot be handled */
512 FAILED, /* Error: handling failed */
Andi Kleen6a460792009-09-16 11:50:15 +0200513 DELAYED, /* Will be handled later */
Andi Kleen6a460792009-09-16 11:50:15 +0200514 RECOVERED, /* Successfully recovered */
515};
516
517static const char *action_name[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100518 [IGNORED] = "Ignored",
Andi Kleen6a460792009-09-16 11:50:15 +0200519 [FAILED] = "Failed",
520 [DELAYED] = "Delayed",
Andi Kleen6a460792009-09-16 11:50:15 +0200521 [RECOVERED] = "Recovered",
522};
523
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700524enum action_page_type {
525 MSG_KERNEL,
526 MSG_KERNEL_HIGH_ORDER,
527 MSG_SLAB,
528 MSG_DIFFERENT_COMPOUND,
529 MSG_POISONED_HUGE,
530 MSG_HUGE,
531 MSG_FREE_HUGE,
532 MSG_UNMAP_FAILED,
533 MSG_DIRTY_SWAPCACHE,
534 MSG_CLEAN_SWAPCACHE,
535 MSG_DIRTY_MLOCKED_LRU,
536 MSG_CLEAN_MLOCKED_LRU,
537 MSG_DIRTY_UNEVICTABLE_LRU,
538 MSG_CLEAN_UNEVICTABLE_LRU,
539 MSG_DIRTY_LRU,
540 MSG_CLEAN_LRU,
541 MSG_TRUNCATED_LRU,
542 MSG_BUDDY,
543 MSG_BUDDY_2ND,
544 MSG_UNKNOWN,
545};
546
547static const char * const action_page_types[] = {
548 [MSG_KERNEL] = "reserved kernel page",
549 [MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
550 [MSG_SLAB] = "kernel slab page",
551 [MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
552 [MSG_POISONED_HUGE] = "huge page already hardware poisoned",
553 [MSG_HUGE] = "huge page",
554 [MSG_FREE_HUGE] = "free huge page",
555 [MSG_UNMAP_FAILED] = "unmapping failed page",
556 [MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
557 [MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
558 [MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
559 [MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
560 [MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
561 [MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
562 [MSG_DIRTY_LRU] = "dirty LRU page",
563 [MSG_CLEAN_LRU] = "clean LRU page",
564 [MSG_TRUNCATED_LRU] = "already truncated LRU page",
565 [MSG_BUDDY] = "free buddy page",
566 [MSG_BUDDY_2ND] = "free buddy page (2nd try)",
567 [MSG_UNKNOWN] = "unknown page",
568};
569
Andi Kleen6a460792009-09-16 11:50:15 +0200570/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100571 * XXX: It is possible that a page is isolated from LRU cache,
572 * and then kept in swap cache or failed to remove from page cache.
573 * The page count will stop it from being freed by unpoison.
574 * Stress tests should be aware of this memory leak problem.
575 */
576static int delete_from_lru_cache(struct page *p)
577{
578 if (!isolate_lru_page(p)) {
579 /*
580 * Clear sensible page flags, so that the buddy system won't
581 * complain when the page is unpoison-and-freed.
582 */
583 ClearPageActive(p);
584 ClearPageUnevictable(p);
585 /*
586 * drop the page count elevated by isolate_lru_page()
587 */
588 page_cache_release(p);
589 return 0;
590 }
591 return -EIO;
592}
593
594/*
Andi Kleen6a460792009-09-16 11:50:15 +0200595 * Error hit kernel page.
596 * Do nothing, try to be lucky and not touch this instead. For a few cases we
597 * could be more sophisticated.
598 */
599static int me_kernel(struct page *p, unsigned long pfn)
600{
Andi Kleen6a460792009-09-16 11:50:15 +0200601 return IGNORED;
602}
603
604/*
605 * Page in unknown state. Do nothing.
606 */
607static int me_unknown(struct page *p, unsigned long pfn)
608{
609 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
610 return FAILED;
611}
612
613/*
Andi Kleen6a460792009-09-16 11:50:15 +0200614 * Clean (or cleaned) page cache page.
615 */
616static int me_pagecache_clean(struct page *p, unsigned long pfn)
617{
618 int err;
619 int ret = FAILED;
620 struct address_space *mapping;
621
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100622 delete_from_lru_cache(p);
623
Andi Kleen6a460792009-09-16 11:50:15 +0200624 /*
625 * For anonymous pages we're done the only reference left
626 * should be the one m_f() holds.
627 */
628 if (PageAnon(p))
629 return RECOVERED;
630
631 /*
632 * Now truncate the page in the page cache. This is really
633 * more like a "temporary hole punch"
634 * Don't do this for block devices when someone else
635 * has a reference, because it could be file system metadata
636 * and that's not safe to truncate.
637 */
638 mapping = page_mapping(p);
639 if (!mapping) {
640 /*
641 * Page has been teared down in the meanwhile
642 */
643 return FAILED;
644 }
645
646 /*
647 * Truncation is a bit tricky. Enable it per file system for now.
648 *
649 * Open: to take i_mutex or not for this? Right now we don't.
650 */
651 if (mapping->a_ops->error_remove_page) {
652 err = mapping->a_ops->error_remove_page(mapping, p);
653 if (err != 0) {
654 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
655 pfn, err);
656 } else if (page_has_private(p) &&
657 !try_to_release_page(p, GFP_NOIO)) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200658 pr_info("MCE %#lx: failed to release buffers\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200659 } else {
660 ret = RECOVERED;
661 }
662 } else {
663 /*
664 * If the file system doesn't support it just invalidate
665 * This fails on dirty or anything with private pages
666 */
667 if (invalidate_inode_page(p))
668 ret = RECOVERED;
669 else
670 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
671 pfn);
672 }
673 return ret;
674}
675
676/*
Zhi Yong Wu549543d2014-01-21 15:49:08 -0800677 * Dirty pagecache page
Andi Kleen6a460792009-09-16 11:50:15 +0200678 * Issues: when the error hit a hole page the error is not properly
679 * propagated.
680 */
681static int me_pagecache_dirty(struct page *p, unsigned long pfn)
682{
683 struct address_space *mapping = page_mapping(p);
684
685 SetPageError(p);
686 /* TBD: print more information about the file. */
687 if (mapping) {
688 /*
689 * IO error will be reported by write(), fsync(), etc.
690 * who check the mapping.
691 * This way the application knows that something went
692 * wrong with its dirty file data.
693 *
694 * There's one open issue:
695 *
696 * The EIO will be only reported on the next IO
697 * operation and then cleared through the IO map.
698 * Normally Linux has two mechanisms to pass IO error
699 * first through the AS_EIO flag in the address space
700 * and then through the PageError flag in the page.
701 * Since we drop pages on memory failure handling the
702 * only mechanism open to use is through AS_AIO.
703 *
704 * This has the disadvantage that it gets cleared on
705 * the first operation that returns an error, while
706 * the PageError bit is more sticky and only cleared
707 * when the page is reread or dropped. If an
708 * application assumes it will always get error on
709 * fsync, but does other operations on the fd before
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300710 * and the page is dropped between then the error
Andi Kleen6a460792009-09-16 11:50:15 +0200711 * will not be properly reported.
712 *
713 * This can already happen even without hwpoisoned
714 * pages: first on metadata IO errors (which only
715 * report through AS_EIO) or when the page is dropped
716 * at the wrong time.
717 *
718 * So right now we assume that the application DTRT on
719 * the first EIO, but we're not worse than other parts
720 * of the kernel.
721 */
722 mapping_set_error(mapping, EIO);
723 }
724
725 return me_pagecache_clean(p, pfn);
726}
727
728/*
729 * Clean and dirty swap cache.
730 *
731 * Dirty swap cache page is tricky to handle. The page could live both in page
732 * cache and swap cache(ie. page is freshly swapped in). So it could be
733 * referenced concurrently by 2 types of PTEs:
734 * normal PTEs and swap PTEs. We try to handle them consistently by calling
735 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
736 * and then
737 * - clear dirty bit to prevent IO
738 * - remove from LRU
739 * - but keep in the swap cache, so that when we return to it on
740 * a later page fault, we know the application is accessing
741 * corrupted data and shall be killed (we installed simple
742 * interception code in do_swap_page to catch it).
743 *
744 * Clean swap cache pages can be directly isolated. A later page fault will
745 * bring in the known good data from disk.
746 */
747static int me_swapcache_dirty(struct page *p, unsigned long pfn)
748{
Andi Kleen6a460792009-09-16 11:50:15 +0200749 ClearPageDirty(p);
750 /* Trigger EIO in shmem: */
751 ClearPageUptodate(p);
752
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100753 if (!delete_from_lru_cache(p))
754 return DELAYED;
755 else
756 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200757}
758
759static int me_swapcache_clean(struct page *p, unsigned long pfn)
760{
Andi Kleen6a460792009-09-16 11:50:15 +0200761 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800762
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100763 if (!delete_from_lru_cache(p))
764 return RECOVERED;
765 else
766 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200767}
768
769/*
770 * Huge pages. Needs work.
771 * Issues:
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900772 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
773 * To narrow down kill region to one page, we need to break up pmd.
Andi Kleen6a460792009-09-16 11:50:15 +0200774 */
775static int me_huge_page(struct page *p, unsigned long pfn)
776{
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900777 int res = 0;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900778 struct page *hpage = compound_head(p);
779 /*
780 * We can safely recover from error on free or reserved (i.e.
781 * not in-use) hugepage by dequeuing it from freelist.
782 * To check whether a hugepage is in-use or not, we can't use
783 * page->lru because it can be used in other hugepage operations,
784 * such as __unmap_hugepage_range() and gather_surplus_pages().
785 * So instead we use page_mapping() and PageAnon().
786 * We assume that this function is called with page lock held,
787 * so there is no race between isolation and mapping/unmapping.
788 */
789 if (!(page_mapping(hpage) || PageAnon(hpage))) {
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900790 res = dequeue_hwpoisoned_huge_page(hpage);
791 if (!res)
792 return RECOVERED;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900793 }
794 return DELAYED;
Andi Kleen6a460792009-09-16 11:50:15 +0200795}
796
797/*
798 * Various page states we can handle.
799 *
800 * A page state is defined by its current page->flags bits.
801 * The table matches them in order and calls the right handler.
802 *
803 * This is quite tricky because we can access page at any time
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300804 * in its live cycle, so all accesses have to be extremely careful.
Andi Kleen6a460792009-09-16 11:50:15 +0200805 *
806 * This is not complete. More states could be added.
807 * For any missing state don't attempt recovery.
808 */
809
810#define dirty (1UL << PG_dirty)
811#define sc (1UL << PG_swapcache)
812#define unevict (1UL << PG_unevictable)
813#define mlock (1UL << PG_mlocked)
814#define writeback (1UL << PG_writeback)
815#define lru (1UL << PG_lru)
816#define swapbacked (1UL << PG_swapbacked)
817#define head (1UL << PG_head)
818#define tail (1UL << PG_tail)
819#define compound (1UL << PG_compound)
820#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200821#define reserved (1UL << PG_reserved)
822
823static struct page_state {
824 unsigned long mask;
825 unsigned long res;
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700826 enum action_page_type type;
Andi Kleen6a460792009-09-16 11:50:15 +0200827 int (*action)(struct page *p, unsigned long pfn);
828} error_states[] = {
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700829 { reserved, reserved, MSG_KERNEL, me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100830 /*
831 * free pages are specially detected outside this table:
832 * PG_buddy pages only make a small fraction of all free pages.
833 */
Andi Kleen6a460792009-09-16 11:50:15 +0200834
835 /*
836 * Could in theory check if slab page is free or if we can drop
837 * currently unused objects without touching them. But just
838 * treat it as standard kernel for now.
839 */
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700840 { slab, slab, MSG_SLAB, me_kernel },
Andi Kleen6a460792009-09-16 11:50:15 +0200841
842#ifdef CONFIG_PAGEFLAGS_EXTENDED
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700843 { head, head, MSG_HUGE, me_huge_page },
844 { tail, tail, MSG_HUGE, me_huge_page },
Andi Kleen6a460792009-09-16 11:50:15 +0200845#else
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700846 { compound, compound, MSG_HUGE, me_huge_page },
Andi Kleen6a460792009-09-16 11:50:15 +0200847#endif
848
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700849 { sc|dirty, sc|dirty, MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
850 { sc|dirty, sc, MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200851
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700852 { mlock|dirty, mlock|dirty, MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
853 { mlock|dirty, mlock, MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200854
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700855 { unevict|dirty, unevict|dirty, MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
856 { unevict|dirty, unevict, MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
Naoya Horiguchi5f4b9fc2013-02-22 16:35:53 -0800857
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700858 { lru|dirty, lru|dirty, MSG_DIRTY_LRU, me_pagecache_dirty },
859 { lru|dirty, lru, MSG_CLEAN_LRU, me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200860
861 /*
862 * Catchall entry: must be at end.
863 */
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700864 { 0, 0, MSG_UNKNOWN, me_unknown },
Andi Kleen6a460792009-09-16 11:50:15 +0200865};
866
Andi Kleen2326c462009-12-16 12:20:00 +0100867#undef dirty
868#undef sc
869#undef unevict
870#undef mlock
871#undef writeback
872#undef lru
873#undef swapbacked
874#undef head
875#undef tail
876#undef compound
877#undef slab
878#undef reserved
879
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800880/*
881 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
882 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
883 */
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700884static void action_result(unsigned long pfn, enum action_page_type type, int result)
Andi Kleen6a460792009-09-16 11:50:15 +0200885{
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700886 pr_err("MCE %#lx: recovery action for %s: %s\n",
887 pfn, action_page_types[type], action_name[result]);
Andi Kleen6a460792009-09-16 11:50:15 +0200888}
889
890static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100891 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200892{
893 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200894 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200895
896 result = ps->action(p, pfn);
Wu Fengguang7456b042009-10-19 08:15:01 +0200897
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100898 count = page_count(p) - 1;
Wu Fengguang138ce282009-12-16 12:19:58 +0100899 if (ps->action == me_swapcache_dirty && result == DELAYED)
900 count--;
901 if (count != 0) {
Andi Kleen6a460792009-09-16 11:50:15 +0200902 printk(KERN_ERR
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700903 "MCE %#lx: %s still referenced by %d users\n",
904 pfn, action_page_types[ps->type], count);
Wu Fengguang138ce282009-12-16 12:19:58 +0100905 result = FAILED;
906 }
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700907 action_result(pfn, ps->type, result);
Andi Kleen6a460792009-09-16 11:50:15 +0200908
909 /* Could do more checks here if page looks ok */
910 /*
911 * Could adjust zone counters here to correct for the missing page.
912 */
913
Wu Fengguang138ce282009-12-16 12:19:58 +0100914 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200915}
916
Andi Kleen6a460792009-09-16 11:50:15 +0200917/*
918 * Do all that is necessary to remove user space mappings. Unmap
919 * the pages and send SIGBUS to the processes if the data was dirty.
920 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100921static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800922 int trapno, int flags, struct page **hpagep)
Andi Kleen6a460792009-09-16 11:50:15 +0200923{
924 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
925 struct address_space *mapping;
926 LIST_HEAD(tokill);
927 int ret;
Tony Luck6751ed62012-07-11 10:20:47 -0700928 int kill = 1, forcekill;
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800929 struct page *hpage = *hpagep;
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800930 struct page *ppage;
Andi Kleen6a460792009-09-16 11:50:15 +0200931
Naoya Horiguchi93a9eb32014-07-30 16:08:28 -0700932 /*
933 * Here we are interested only in user-mapped pages, so skip any
934 * other types of pages.
935 */
936 if (PageReserved(p) || PageSlab(p))
937 return SWAP_SUCCESS;
938 if (!(PageLRU(hpage) || PageHuge(p)))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100939 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200940
Andi Kleen6a460792009-09-16 11:50:15 +0200941 /*
942 * This check implies we don't kill processes if their pages
943 * are in the swap cache early. Those are always late kills.
944 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900945 if (!page_mapped(hpage))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100946 return SWAP_SUCCESS;
947
Naoya Horiguchi52089b12014-07-30 16:08:30 -0700948 if (PageKsm(p)) {
949 pr_err("MCE %#lx: can't handle KSM pages.\n", pfn);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100950 return SWAP_FAIL;
Naoya Horiguchi52089b12014-07-30 16:08:30 -0700951 }
Andi Kleen6a460792009-09-16 11:50:15 +0200952
953 if (PageSwapCache(p)) {
954 printk(KERN_ERR
955 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
956 ttu |= TTU_IGNORE_HWPOISON;
957 }
958
959 /*
960 * Propagate the dirty bit from PTEs to struct page first, because we
961 * need this to decide if we should kill or just drop the page.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100962 * XXX: the dirty test could be racy: set_page_dirty() may not always
963 * be called inside page lock (it's recommended but not enforced).
Andi Kleen6a460792009-09-16 11:50:15 +0200964 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900965 mapping = page_mapping(hpage);
Tony Luck6751ed62012-07-11 10:20:47 -0700966 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900967 mapping_cap_writeback_dirty(mapping)) {
968 if (page_mkclean(hpage)) {
969 SetPageDirty(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +0200970 } else {
971 kill = 0;
972 ttu |= TTU_IGNORE_HWPOISON;
973 printk(KERN_INFO
974 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
975 pfn);
976 }
977 }
978
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800979 /*
980 * ppage: poisoned page
981 * if p is regular page(4k page)
982 * ppage == real poisoned page;
983 * else p is hugetlb or THP, ppage == head page.
984 */
985 ppage = hpage;
986
Jin Dongmingefeda7a2011-02-01 15:52:39 -0800987 if (PageTransHuge(hpage)) {
988 /*
989 * Verify that this isn't a hugetlbfs head page, the check for
990 * PageAnon is just for avoid tripping a split_huge_page
991 * internal debug check, as split_huge_page refuses to deal with
992 * anything that isn't an anon page. PageAnon can't go away fro
993 * under us because we hold a refcount on the hpage, without a
994 * refcount on the hpage. split_huge_page can't be safely called
995 * in the first place, having a refcount on the tail isn't
996 * enough * to be safe.
997 */
998 if (!PageHuge(hpage) && PageAnon(hpage)) {
999 if (unlikely(split_huge_page(hpage))) {
1000 /*
1001 * FIXME: if splitting THP is failed, it is
1002 * better to stop the following operation rather
1003 * than causing panic by unmapping. System might
1004 * survive if the page is freed later.
1005 */
1006 printk(KERN_INFO
1007 "MCE %#lx: failed to split THP\n", pfn);
1008
1009 BUG_ON(!PageHWPoison(p));
1010 return SWAP_FAIL;
1011 }
Naoya Horiguchia3e0f9e2014-01-02 12:58:51 -08001012 /*
1013 * We pinned the head page for hwpoison handling,
1014 * now we split the thp and we are interested in
1015 * the hwpoisoned raw page, so move the refcount
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001016 * to it. Similarly, page lock is shifted.
Naoya Horiguchia3e0f9e2014-01-02 12:58:51 -08001017 */
1018 if (hpage != p) {
Naoya Horiguchi8d547ff2014-02-10 14:25:50 -08001019 if (!(flags & MF_COUNT_INCREASED)) {
1020 put_page(hpage);
1021 get_page(p);
1022 }
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001023 lock_page(p);
1024 unlock_page(hpage);
1025 *hpagep = p;
Naoya Horiguchia3e0f9e2014-01-02 12:58:51 -08001026 }
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001027 /* THP is split, so ppage should be the real poisoned page. */
1028 ppage = p;
Jin Dongmingefeda7a2011-02-01 15:52:39 -08001029 }
1030 }
1031
Andi Kleen6a460792009-09-16 11:50:15 +02001032 /*
1033 * First collect all the processes that have the page
1034 * mapped in dirty form. This has to be done before try_to_unmap,
1035 * because ttu takes the rmap data structures down.
1036 *
1037 * Error handling: We ignore errors here because
1038 * there's nothing that can be done.
1039 */
1040 if (kill)
Tony Luck74614de2014-06-04 16:11:01 -07001041 collect_procs(ppage, &tokill, flags & MF_ACTION_REQUIRED);
Andi Kleen6a460792009-09-16 11:50:15 +02001042
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001043 ret = try_to_unmap(ppage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +02001044 if (ret != SWAP_SUCCESS)
1045 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001046 pfn, page_mapcount(ppage));
1047
Andi Kleen6a460792009-09-16 11:50:15 +02001048 /*
1049 * Now that the dirty bit has been propagated to the
1050 * struct page and all unmaps done we can decide if
1051 * killing is needed or not. Only kill when the page
Tony Luck6751ed62012-07-11 10:20:47 -07001052 * was dirty or the process is not restartable,
1053 * otherwise the tokill list is merely
Andi Kleen6a460792009-09-16 11:50:15 +02001054 * freed. When there was a problem unmapping earlier
1055 * use a more force-full uncatchable kill to prevent
1056 * any accesses to the poisoned memory.
1057 */
Tony Luck6751ed62012-07-11 10:20:47 -07001058 forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL);
1059 kill_procs(&tokill, forcekill, trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -08001060 ret != SWAP_SUCCESS, p, pfn, flags);
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001061
1062 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +02001063}
1064
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001065static void set_page_hwpoison_huge_page(struct page *hpage)
1066{
1067 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -07001068 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001069 for (i = 0; i < nr_pages; i++)
1070 SetPageHWPoison(hpage + i);
1071}
1072
1073static void clear_page_hwpoison_huge_page(struct page *hpage)
1074{
1075 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -07001076 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001077 for (i = 0; i < nr_pages; i++)
1078 ClearPageHWPoison(hpage + i);
1079}
1080
Tony Luckcd42f4a2011-12-15 10:48:12 -08001081/**
1082 * memory_failure - Handle memory failure of a page.
1083 * @pfn: Page Number of the corrupted page
1084 * @trapno: Trap number reported in the signal to user space.
1085 * @flags: fine tune action taken
1086 *
1087 * This function is called by the low level machine check code
1088 * of an architecture when it detects hardware memory corruption
1089 * of a page. It tries its best to recover, which includes
1090 * dropping pages, killing processes etc.
1091 *
1092 * The function is primarily of use for corruptions that
1093 * happen outside the current execution context (e.g. when
1094 * detected by a background scrubber)
1095 *
1096 * Must run in process context (e.g. a work queue) with interrupts
1097 * enabled and no spinlocks hold.
1098 */
1099int memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +02001100{
1101 struct page_state *ps;
1102 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001103 struct page *hpage;
Andi Kleen6a460792009-09-16 11:50:15 +02001104 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001105 unsigned int nr_pages;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001106 unsigned long page_flags;
Andi Kleen6a460792009-09-16 11:50:15 +02001107
1108 if (!sysctl_memory_failure_recovery)
1109 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1110
1111 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +01001112 printk(KERN_ERR
1113 "MCE %#lx: memory outside kernel control\n",
1114 pfn);
1115 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +02001116 }
1117
1118 p = pfn_to_page(pfn);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001119 hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +02001120 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001121 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001122 return 0;
1123 }
1124
Naoya Horiguchi4db0e952013-02-22 16:34:05 -08001125 /*
1126 * Currently errors on hugetlbfs pages are measured in hugepage units,
1127 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1128 * transparent hugepages, they are supposed to be split and error
1129 * measurement is done in normal page units. So nr_pages should be one
1130 * in this case.
1131 */
1132 if (PageHuge(p))
1133 nr_pages = 1 << compound_order(hpage);
1134 else /* normal page or thp */
1135 nr_pages = 1;
Xishi Qiu293c07e2013-02-22 16:34:02 -08001136 atomic_long_add(nr_pages, &num_poisoned_pages);
Andi Kleen6a460792009-09-16 11:50:15 +02001137
1138 /*
1139 * We need/can do nothing about count=0 pages.
1140 * 1) it's a free page, and therefore in safe hand:
1141 * prep_new_page() will be the gate keeper.
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001142 * 2) it's a free hugepage, which is also safe:
1143 * an affected hugepage will be dequeued from hugepage freelist,
1144 * so there's no concern about reusing it ever after.
1145 * 3) it's part of a non-compound high order page.
Andi Kleen6a460792009-09-16 11:50:15 +02001146 * Implies some kernel user: cannot stop them from
1147 * R/W the page; let's pray that the page has been
1148 * used and will be freed some time later.
1149 * In fact it's dangerous to directly bump up page count from 0,
1150 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1151 */
Andi Kleen82ba0112009-12-16 12:19:57 +01001152 if (!(flags & MF_COUNT_INCREASED) &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001153 !get_page_unless_zero(hpage)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001154 if (is_free_buddy_page(p)) {
Naoya Horiguchi64d37a22015-04-15 16:13:05 -07001155 action_result(pfn, MSG_BUDDY, DELAYED);
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001156 return 0;
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001157 } else if (PageHuge(hpage)) {
1158 /*
Chen Yucongb9851942014-05-22 11:54:15 -07001159 * Check "filter hit" and "race with other subpage."
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001160 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001161 lock_page(hpage);
Chen Yucongb9851942014-05-22 11:54:15 -07001162 if (PageHWPoison(hpage)) {
1163 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1164 || (p != hpage && TestSetPageHWPoison(hpage))) {
1165 atomic_long_sub(nr_pages, &num_poisoned_pages);
1166 unlock_page(hpage);
1167 return 0;
1168 }
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001169 }
1170 set_page_hwpoison_huge_page(hpage);
1171 res = dequeue_hwpoisoned_huge_page(hpage);
Naoya Horiguchi64d37a22015-04-15 16:13:05 -07001172 action_result(pfn, MSG_FREE_HUGE,
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001173 res ? IGNORED : DELAYED);
1174 unlock_page(hpage);
1175 return res;
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001176 } else {
Naoya Horiguchi64d37a22015-04-15 16:13:05 -07001177 action_result(pfn, MSG_KERNEL_HIGH_ORDER, IGNORED);
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001178 return -EBUSY;
1179 }
Andi Kleen6a460792009-09-16 11:50:15 +02001180 }
1181
1182 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001183 * We ignore non-LRU pages for good reasons.
1184 * - PG_locked is only well defined for LRU pages and a few others
1185 * - to avoid races with __set_page_locked()
1186 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1187 * The check (unnecessarily) ignores LRU pages being isolated and
1188 * walked by the page reclaim code, however that's not a big loss.
1189 */
Dean Nelson385de352012-03-21 16:34:05 -07001190 if (!PageHuge(p) && !PageTransTail(p)) {
Jin Dongmingaf241a02011-02-01 15:52:41 -08001191 if (!PageLRU(p))
1192 shake_page(p, 0);
1193 if (!PageLRU(p)) {
1194 /*
1195 * shake_page could have turned it free.
1196 */
1197 if (is_free_buddy_page(p)) {
Wanpeng Li2d421ac2013-09-30 13:45:23 -07001198 if (flags & MF_COUNT_INCREASED)
Naoya Horiguchi64d37a22015-04-15 16:13:05 -07001199 action_result(pfn, MSG_BUDDY, DELAYED);
Wanpeng Li2d421ac2013-09-30 13:45:23 -07001200 else
Naoya Horiguchi64d37a22015-04-15 16:13:05 -07001201 action_result(pfn, MSG_BUDDY_2ND,
1202 DELAYED);
Jin Dongmingaf241a02011-02-01 15:52:41 -08001203 return 0;
1204 }
Andi Kleen0474a602009-12-16 12:20:00 +01001205 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001206 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001207
Jens Axboe7eaceac2011-03-10 08:52:07 +01001208 lock_page(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001209
1210 /*
Andi Kleenf37d4292014-08-06 16:06:49 -07001211 * The page could have changed compound pages during the locking.
1212 * If this happens just bail out.
1213 */
1214 if (compound_head(p) != hpage) {
Naoya Horiguchi64d37a22015-04-15 16:13:05 -07001215 action_result(pfn, MSG_DIFFERENT_COMPOUND, IGNORED);
Andi Kleenf37d4292014-08-06 16:06:49 -07001216 res = -EBUSY;
1217 goto out;
1218 }
1219
1220 /*
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001221 * We use page flags to determine what action should be taken, but
1222 * the flags can be modified by the error containment action. One
1223 * example is an mlocked page, where PG_mlocked is cleared by
1224 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1225 * correctly, we save a copy of the page flags at this time.
1226 */
1227 page_flags = p->flags;
1228
1229 /*
Wu Fengguang847ce402009-12-16 12:19:58 +01001230 * unpoison always clear PG_hwpoison inside page lock
1231 */
1232 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001233 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Naoya Horiguchi3e030ec2014-05-22 11:54:21 -07001234 atomic_long_sub(nr_pages, &num_poisoned_pages);
1235 put_page(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001236 res = 0;
1237 goto out;
1238 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001239 if (hwpoison_filter(p)) {
1240 if (TestClearPageHWPoison(p))
Xishi Qiu293c07e2013-02-22 16:34:02 -08001241 atomic_long_sub(nr_pages, &num_poisoned_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001242 unlock_page(hpage);
1243 put_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001244 return 0;
1245 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001246
Chen Yucong0bc1f8b2014-07-02 15:22:37 -07001247 if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p))
1248 goto identify_page_state;
1249
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001250 /*
1251 * For error on the tail page, we should set PG_hwpoison
1252 * on the head page to show that the hugepage is hwpoisoned
1253 */
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001254 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
Naoya Horiguchi64d37a22015-04-15 16:13:05 -07001255 action_result(pfn, MSG_POISONED_HUGE, IGNORED);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001256 unlock_page(hpage);
1257 put_page(hpage);
1258 return 0;
1259 }
1260 /*
1261 * Set PG_hwpoison on all pages in an error hugepage,
1262 * because containment is done in hugepage unit for now.
1263 * Since we have done TestSetPageHWPoison() for the head page with
1264 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1265 */
1266 if (PageHuge(p))
1267 set_page_hwpoison_huge_page(hpage);
1268
Naoya Horiguchi6edd6cc2014-06-04 16:10:35 -07001269 /*
1270 * It's very difficult to mess with pages currently under IO
1271 * and in many cases impossible, so we just avoid it here.
1272 */
Andi Kleen6a460792009-09-16 11:50:15 +02001273 wait_on_page_writeback(p);
1274
1275 /*
1276 * Now take care of user space mappings.
Minchan Kime64a7822011-03-22 16:32:44 -07001277 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001278 *
1279 * When the raw error page is thp tail page, hpage points to the raw
1280 * page after thp split.
Andi Kleen6a460792009-09-16 11:50:15 +02001281 */
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001282 if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)
1283 != SWAP_SUCCESS) {
Naoya Horiguchi64d37a22015-04-15 16:13:05 -07001284 action_result(pfn, MSG_UNMAP_FAILED, IGNORED);
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001285 res = -EBUSY;
1286 goto out;
1287 }
Andi Kleen6a460792009-09-16 11:50:15 +02001288
1289 /*
1290 * Torn down by someone else?
1291 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001292 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Naoya Horiguchi64d37a22015-04-15 16:13:05 -07001293 action_result(pfn, MSG_TRUNCATED_LRU, IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001294 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001295 goto out;
1296 }
1297
Chen Yucong0bc1f8b2014-07-02 15:22:37 -07001298identify_page_state:
Andi Kleen6a460792009-09-16 11:50:15 +02001299 res = -EBUSY;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001300 /*
1301 * The first check uses the current page flags which may not have any
1302 * relevant information. The second check with the saved page flagss is
1303 * carried out only if the first check can't determine the page status.
1304 */
1305 for (ps = error_states;; ps++)
1306 if ((p->flags & ps->mask) == ps->res)
Andi Kleen6a460792009-09-16 11:50:15 +02001307 break;
Wanpeng Li841fcc52013-09-11 14:22:50 -07001308
1309 page_flags |= (p->flags & (1UL << PG_dirty));
1310
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001311 if (!ps->mask)
1312 for (ps = error_states;; ps++)
1313 if ((page_flags & ps->mask) == ps->res)
1314 break;
1315 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001316out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001317 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001318 return res;
1319}
Tony Luckcd42f4a2011-12-15 10:48:12 -08001320EXPORT_SYMBOL_GPL(memory_failure);
Wu Fengguang847ce402009-12-16 12:19:58 +01001321
Huang Yingea8f5fb2011-07-13 13:14:27 +08001322#define MEMORY_FAILURE_FIFO_ORDER 4
1323#define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1324
1325struct memory_failure_entry {
1326 unsigned long pfn;
1327 int trapno;
1328 int flags;
1329};
1330
1331struct memory_failure_cpu {
1332 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1333 MEMORY_FAILURE_FIFO_SIZE);
1334 spinlock_t lock;
1335 struct work_struct work;
1336};
1337
1338static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1339
1340/**
1341 * memory_failure_queue - Schedule handling memory failure of a page.
1342 * @pfn: Page Number of the corrupted page
1343 * @trapno: Trap number reported in the signal to user space.
1344 * @flags: Flags for memory failure handling
1345 *
1346 * This function is called by the low level hardware error handler
1347 * when it detects hardware memory corruption of a page. It schedules
1348 * the recovering of error page, including dropping pages, killing
1349 * processes etc.
1350 *
1351 * The function is primarily of use for corruptions that
1352 * happen outside the current execution context (e.g. when
1353 * detected by a background scrubber)
1354 *
1355 * Can run in IRQ context.
1356 */
1357void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1358{
1359 struct memory_failure_cpu *mf_cpu;
1360 unsigned long proc_flags;
1361 struct memory_failure_entry entry = {
1362 .pfn = pfn,
1363 .trapno = trapno,
1364 .flags = flags,
1365 };
1366
1367 mf_cpu = &get_cpu_var(memory_failure_cpu);
1368 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
Stefani Seibold498d3192013-11-14 14:32:17 -08001369 if (kfifo_put(&mf_cpu->fifo, entry))
Huang Yingea8f5fb2011-07-13 13:14:27 +08001370 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1371 else
Joe Perches8e33a522013-07-25 11:53:25 -07001372 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
Huang Yingea8f5fb2011-07-13 13:14:27 +08001373 pfn);
1374 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1375 put_cpu_var(memory_failure_cpu);
1376}
1377EXPORT_SYMBOL_GPL(memory_failure_queue);
1378
1379static void memory_failure_work_func(struct work_struct *work)
1380{
1381 struct memory_failure_cpu *mf_cpu;
1382 struct memory_failure_entry entry = { 0, };
1383 unsigned long proc_flags;
1384 int gotten;
1385
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001386 mf_cpu = this_cpu_ptr(&memory_failure_cpu);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001387 for (;;) {
1388 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1389 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1390 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1391 if (!gotten)
1392 break;
Naveen N. Raocf870c72013-07-10 14:57:01 +05301393 if (entry.flags & MF_SOFT_OFFLINE)
1394 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1395 else
1396 memory_failure(entry.pfn, entry.trapno, entry.flags);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001397 }
1398}
1399
1400static int __init memory_failure_init(void)
1401{
1402 struct memory_failure_cpu *mf_cpu;
1403 int cpu;
1404
1405 for_each_possible_cpu(cpu) {
1406 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1407 spin_lock_init(&mf_cpu->lock);
1408 INIT_KFIFO(mf_cpu->fifo);
1409 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1410 }
1411
1412 return 0;
1413}
1414core_initcall(memory_failure_init);
1415
Wu Fengguang847ce402009-12-16 12:19:58 +01001416/**
1417 * unpoison_memory - Unpoison a previously poisoned page
1418 * @pfn: Page number of the to be unpoisoned page
1419 *
1420 * Software-unpoison a page that has been poisoned by
1421 * memory_failure() earlier.
1422 *
1423 * This is only done on the software-level, so it only works
1424 * for linux injected failures, not real hardware failures
1425 *
1426 * Returns 0 for success, otherwise -errno.
1427 */
1428int unpoison_memory(unsigned long pfn)
1429{
1430 struct page *page;
1431 struct page *p;
1432 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001433 unsigned int nr_pages;
Wu Fengguang847ce402009-12-16 12:19:58 +01001434
1435 if (!pfn_valid(pfn))
1436 return -ENXIO;
1437
1438 p = pfn_to_page(pfn);
1439 page = compound_head(p);
1440
1441 if (!PageHWPoison(p)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001442 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001443 return 0;
1444 }
1445
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001446 /*
1447 * unpoison_memory() can encounter thp only when the thp is being
1448 * worked by memory_failure() and the page lock is not held yet.
1449 * In such case, we yield to memory_failure() and make unpoison fail.
1450 */
Wanpeng Lie76d30e2013-09-30 13:45:22 -07001451 if (!PageHuge(page) && PageTransHuge(page)) {
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001452 pr_info("MCE: Memory failure is now running on %#lx\n", pfn);
1453 return 0;
1454 }
1455
Wanpeng Lif9121152013-09-11 14:22:52 -07001456 nr_pages = 1 << compound_order(page);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001457
Wu Fengguang847ce402009-12-16 12:19:58 +01001458 if (!get_page_unless_zero(page)) {
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001459 /*
1460 * Since HWPoisoned hugepage should have non-zero refcount,
1461 * race between memory failure and unpoison seems to happen.
1462 * In such case unpoison fails and memory failure runs
1463 * to the end.
1464 */
1465 if (PageHuge(page)) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001466 pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001467 return 0;
1468 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001469 if (TestClearPageHWPoison(p))
Wanpeng Lidd9538a2013-09-11 14:22:54 -07001470 atomic_long_dec(&num_poisoned_pages);
Andi Kleenfb46e732010-09-27 23:31:30 +02001471 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001472 return 0;
1473 }
1474
Jens Axboe7eaceac2011-03-10 08:52:07 +01001475 lock_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001476 /*
1477 * This test is racy because PG_hwpoison is set outside of page lock.
1478 * That's acceptable because that won't trigger kernel panic. Instead,
1479 * the PG_hwpoison page will be caught and isolated on the entrance to
1480 * the free buddy page pool.
1481 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001482 if (TestClearPageHWPoison(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001483 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
Xishi Qiu293c07e2013-02-22 16:34:02 -08001484 atomic_long_sub(nr_pages, &num_poisoned_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001485 freeit = 1;
Naoya Horiguchi6a901812010-09-08 10:19:40 +09001486 if (PageHuge(page))
1487 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001488 }
1489 unlock_page(page);
1490
1491 put_page(page);
Wanpeng Li3ba5eeb2013-09-11 14:23:01 -07001492 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
Wu Fengguang847ce402009-12-16 12:19:58 +01001493 put_page(page);
1494
1495 return 0;
1496}
1497EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001498
1499static struct page *new_page(struct page *p, unsigned long private, int **x)
1500{
Andi Kleen12686d12009-12-16 12:20:01 +01001501 int nid = page_to_nid(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001502 if (PageHuge(p))
1503 return alloc_huge_page_node(page_hstate(compound_head(p)),
1504 nid);
1505 else
1506 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001507}
1508
1509/*
1510 * Safely get reference count of an arbitrary page.
1511 * Returns 0 for a free page, -EIO for a zero refcount page
1512 * that is not free, and 1 for any other page type.
1513 * For 1 the page is returned with increased page count, otherwise not.
1514 */
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001515static int __get_any_page(struct page *p, unsigned long pfn, int flags)
Andi Kleenfacb6012009-12-16 12:20:00 +01001516{
1517 int ret;
1518
1519 if (flags & MF_COUNT_INCREASED)
1520 return 1;
1521
1522 /*
Naoya Horiguchid950b952010-09-08 10:19:39 +09001523 * When the target page is a free hugepage, just remove it
1524 * from free hugepage list.
1525 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001526 if (!get_page_unless_zero(compound_head(p))) {
Naoya Horiguchid950b952010-09-08 10:19:39 +09001527 if (PageHuge(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001528 pr_info("%s: %#lx free huge page\n", __func__, pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001529 ret = 0;
Naoya Horiguchid950b952010-09-08 10:19:39 +09001530 } else if (is_free_buddy_page(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001531 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001532 ret = 0;
1533 } else {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001534 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1535 __func__, pfn, p->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001536 ret = -EIO;
1537 }
1538 } else {
1539 /* Not a free page */
1540 ret = 1;
1541 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001542 return ret;
1543}
1544
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001545static int get_any_page(struct page *page, unsigned long pfn, int flags)
1546{
1547 int ret = __get_any_page(page, pfn, flags);
1548
1549 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1550 /*
1551 * Try to free it.
1552 */
1553 put_page(page);
1554 shake_page(page, 1);
1555
1556 /*
1557 * Did it turn free?
1558 */
1559 ret = __get_any_page(page, pfn, 0);
1560 if (!PageLRU(page)) {
1561 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1562 pfn, page->flags);
1563 return -EIO;
1564 }
1565 }
1566 return ret;
1567}
1568
Naoya Horiguchid950b952010-09-08 10:19:39 +09001569static int soft_offline_huge_page(struct page *page, int flags)
1570{
1571 int ret;
1572 unsigned long pfn = page_to_pfn(page);
1573 struct page *hpage = compound_head(page);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001574 LIST_HEAD(pagelist);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001575
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001576 /*
1577 * This double-check of PageHWPoison is to avoid the race with
1578 * memory_failure(). See also comment in __soft_offline_page().
1579 */
1580 lock_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001581 if (PageHWPoison(hpage)) {
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001582 unlock_page(hpage);
1583 put_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001584 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001585 return -EBUSY;
Xishi Qiu0ebff322013-02-22 16:33:59 -08001586 }
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001587 unlock_page(hpage);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001588
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001589 ret = isolate_huge_page(hpage, &pagelist);
1590 if (ret) {
1591 /*
1592 * get_any_page() and isolate_huge_page() takes a refcount each,
1593 * so need to drop one here.
1594 */
1595 put_page(hpage);
1596 } else {
1597 pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn);
1598 return -EBUSY;
1599 }
1600
David Rientjes68711a72014-06-04 16:08:25 -07001601 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001602 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001603 if (ret) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001604 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1605 pfn, ret, page->flags);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001606 /*
1607 * We know that soft_offline_huge_page() tries to migrate
1608 * only one hugepage pointed to by hpage, so we need not
1609 * run through the pagelist here.
1610 */
1611 putback_active_hugepage(hpage);
1612 if (ret > 0)
1613 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001614 } else {
Jianguo Wua49ecbc2013-12-18 17:08:54 -08001615 /* overcommit hugetlb page will be freed to buddy */
1616 if (PageHuge(page)) {
1617 set_page_hwpoison_huge_page(hpage);
1618 dequeue_hwpoisoned_huge_page(hpage);
1619 atomic_long_add(1 << compound_order(hpage),
1620 &num_poisoned_pages);
1621 } else {
1622 SetPageHWPoison(page);
1623 atomic_long_inc(&num_poisoned_pages);
1624 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001625 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001626 return ret;
1627}
1628
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001629static int __soft_offline_page(struct page *page, int flags)
1630{
1631 int ret;
1632 unsigned long pfn = page_to_pfn(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001633
1634 /*
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001635 * Check PageHWPoison again inside page lock because PageHWPoison
1636 * is set by memory_failure() outside page lock. Note that
1637 * memory_failure() also double-checks PageHWPoison inside page lock,
1638 * so there's no race between soft_offline_page() and memory_failure().
Andi Kleenfacb6012009-12-16 12:20:00 +01001639 */
Xishi Qiu0ebff322013-02-22 16:33:59 -08001640 lock_page(page);
1641 wait_on_page_writeback(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001642 if (PageHWPoison(page)) {
1643 unlock_page(page);
1644 put_page(page);
1645 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1646 return -EBUSY;
1647 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001648 /*
1649 * Try to invalidate first. This should work for
1650 * non dirty unmapped page cache pages.
1651 */
1652 ret = invalidate_inode_page(page);
1653 unlock_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001654 /*
Andi Kleenfacb6012009-12-16 12:20:00 +01001655 * RED-PEN would be better to keep it isolated here, but we
1656 * would need to fix isolation locking first.
1657 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001658 if (ret == 1) {
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001659 put_page(page);
Andi Kleenfb46e732010-09-27 23:31:30 +02001660 pr_info("soft_offline: %#lx: invalidated\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001661 SetPageHWPoison(page);
1662 atomic_long_inc(&num_poisoned_pages);
1663 return 0;
Andi Kleenfacb6012009-12-16 12:20:00 +01001664 }
1665
1666 /*
1667 * Simple invalidation didn't work.
1668 * Try to migrate to a new page instead. migrate.c
1669 * handles a large number of cases for us.
1670 */
1671 ret = isolate_lru_page(page);
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001672 /*
1673 * Drop page reference which is came from get_any_page()
1674 * successful isolate_lru_page() already took another one.
1675 */
1676 put_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001677 if (!ret) {
1678 LIST_HEAD(pagelist);
Minchan Kim5db8a732011-06-15 15:08:48 -07001679 inc_zone_page_state(page, NR_ISOLATED_ANON +
Hugh Dickins9c620e22013-02-22 16:35:14 -08001680 page_is_file_cache(page));
Andi Kleenfacb6012009-12-16 12:20:00 +01001681 list_add(&page->lru, &pagelist);
David Rientjes68711a72014-06-04 16:08:25 -07001682 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001683 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Andi Kleenfacb6012009-12-16 12:20:00 +01001684 if (ret) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001685 if (!list_empty(&pagelist)) {
1686 list_del(&page->lru);
1687 dec_zone_page_state(page, NR_ISOLATED_ANON +
1688 page_is_file_cache(page));
1689 putback_lru_page(page);
1690 }
1691
Andi Kleenfb46e732010-09-27 23:31:30 +02001692 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001693 pfn, ret, page->flags);
1694 if (ret > 0)
1695 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001696 } else {
Naoya Horiguchif15bdfa2013-07-03 15:02:37 -07001697 /*
1698 * After page migration succeeds, the source page can
1699 * be trapped in pagevec and actual freeing is delayed.
1700 * Freeing code works differently based on PG_hwpoison,
1701 * so there's a race. We need to make sure that the
1702 * source page should be freed back to buddy before
1703 * setting PG_hwpoison.
1704 */
1705 if (!is_free_buddy_page(page))
Vlastimil Babkac0554322014-12-10 15:43:10 -08001706 drain_all_pages(page_zone(page));
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001707 SetPageHWPoison(page);
Naoya Horiguchif15bdfa2013-07-03 15:02:37 -07001708 if (!is_free_buddy_page(page))
1709 pr_info("soft offline: %#lx: page leaked\n",
1710 pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001711 atomic_long_inc(&num_poisoned_pages);
Andi Kleenfacb6012009-12-16 12:20:00 +01001712 }
1713 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001714 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
Dean Nelsondd73e852011-10-31 17:09:04 -07001715 pfn, ret, page_count(page), page->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001716 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001717 return ret;
1718}
Wanpeng Li86e05772013-09-11 14:22:56 -07001719
1720/**
1721 * soft_offline_page - Soft offline a page.
1722 * @page: page to offline
1723 * @flags: flags. Same as memory_failure().
1724 *
1725 * Returns 0 on success, otherwise negated errno.
1726 *
1727 * Soft offline a page, by migration or invalidation,
1728 * without killing anything. This is for the case when
1729 * a page is not corrupted yet (so it's still valid to access),
1730 * but has had a number of corrected errors and is better taken
1731 * out.
1732 *
1733 * The actual policy on when to do that is maintained by
1734 * user space.
1735 *
1736 * This should never impact any application or cause data loss,
1737 * however it might take some time.
1738 *
1739 * This is not a 100% solution for all memory, but tries to be
1740 * ``good enough'' for the majority of memory.
1741 */
1742int soft_offline_page(struct page *page, int flags)
1743{
1744 int ret;
1745 unsigned long pfn = page_to_pfn(page);
David Rientjes668f9abb2014-03-03 15:38:18 -08001746 struct page *hpage = compound_head(page);
Wanpeng Li86e05772013-09-11 14:22:56 -07001747
1748 if (PageHWPoison(page)) {
1749 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1750 return -EBUSY;
1751 }
1752 if (!PageHuge(page) && PageTransHuge(hpage)) {
1753 if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1754 pr_info("soft offline: %#lx: failed to split THP\n",
1755 pfn);
1756 return -EBUSY;
1757 }
1758 }
1759
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07001760 get_online_mems();
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001761
1762 /*
1763 * Isolate the page, so that it doesn't get reallocated if it
1764 * was free. This flag should be kept set until the source page
1765 * is freed and PG_hwpoison on it is set.
1766 */
1767 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
1768 set_migratetype_isolate(page, true);
1769
Wanpeng Li86e05772013-09-11 14:22:56 -07001770 ret = get_any_page(page, pfn, flags);
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07001771 put_online_mems();
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001772 if (ret > 0) { /* for in-use pages */
Wanpeng Li86e05772013-09-11 14:22:56 -07001773 if (PageHuge(page))
1774 ret = soft_offline_huge_page(page, flags);
1775 else
1776 ret = __soft_offline_page(page, flags);
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001777 } else if (ret == 0) { /* for free pages */
Wanpeng Li86e05772013-09-11 14:22:56 -07001778 if (PageHuge(page)) {
1779 set_page_hwpoison_huge_page(hpage);
1780 dequeue_hwpoisoned_huge_page(hpage);
1781 atomic_long_add(1 << compound_order(hpage),
1782 &num_poisoned_pages);
1783 } else {
1784 SetPageHWPoison(page);
1785 atomic_long_inc(&num_poisoned_pages);
1786 }
1787 }
Wanpeng Li86e05772013-09-11 14:22:56 -07001788 unset_migratetype_isolate(page, MIGRATE_MOVABLE);
1789 return ret;
1790}