blob: 5a544c6c0717c4908fe23e17a976847fb22981ce [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>
Naoya Horiguchia5f65102015-11-05 18:47:26 -080059#include <linux/ratelimit.h>
Andi Kleen6a460792009-09-16 11:50:15 +020060#include "internal.h"
Xie XiuQi97f0b132015-06-24 16:57:36 -070061#include "ras/ras_event.h"
Andi Kleen6a460792009-09-16 11:50:15 +020062
63int sysctl_memory_failure_early_kill __read_mostly = 0;
64
65int sysctl_memory_failure_recovery __read_mostly = 1;
66
Xishi Qiu293c07e2013-02-22 16:34:02 -080067atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
Andi Kleen6a460792009-09-16 11:50:15 +020068
Andi Kleen27df5062009-12-21 19:56:42 +010069#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
70
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010071u32 hwpoison_filter_enable = 0;
Wu Fengguang7c116f22009-12-16 12:19:59 +010072u32 hwpoison_filter_dev_major = ~0U;
73u32 hwpoison_filter_dev_minor = ~0U;
Wu Fengguang478c5ff2009-12-16 12:19:59 +010074u64 hwpoison_filter_flags_mask;
75u64 hwpoison_filter_flags_value;
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010076EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
Wu Fengguang7c116f22009-12-16 12:19:59 +010077EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
78EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
Wu Fengguang478c5ff2009-12-16 12:19:59 +010079EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
80EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
Wu Fengguang7c116f22009-12-16 12:19:59 +010081
82static int hwpoison_filter_dev(struct page *p)
83{
84 struct address_space *mapping;
85 dev_t dev;
86
87 if (hwpoison_filter_dev_major == ~0U &&
88 hwpoison_filter_dev_minor == ~0U)
89 return 0;
90
91 /*
Andi Kleen1c80b992010-09-27 23:09:51 +020092 * page_mapping() does not accept slab pages.
Wu Fengguang7c116f22009-12-16 12:19:59 +010093 */
94 if (PageSlab(p))
95 return -EINVAL;
96
97 mapping = page_mapping(p);
98 if (mapping == NULL || mapping->host == NULL)
99 return -EINVAL;
100
101 dev = mapping->host->i_sb->s_dev;
102 if (hwpoison_filter_dev_major != ~0U &&
103 hwpoison_filter_dev_major != MAJOR(dev))
104 return -EINVAL;
105 if (hwpoison_filter_dev_minor != ~0U &&
106 hwpoison_filter_dev_minor != MINOR(dev))
107 return -EINVAL;
108
109 return 0;
110}
111
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100112static int hwpoison_filter_flags(struct page *p)
113{
114 if (!hwpoison_filter_flags_mask)
115 return 0;
116
117 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
118 hwpoison_filter_flags_value)
119 return 0;
120 else
121 return -EINVAL;
122}
123
Andi Kleen4fd466e2009-12-16 12:19:59 +0100124/*
125 * This allows stress tests to limit test scope to a collection of tasks
126 * by putting them under some memcg. This prevents killing unrelated/important
127 * processes such as /sbin/init. Note that the target task may share clean
128 * pages with init (eg. libc text), which is harmless. If the target task
129 * share _dirty_ pages with another task B, the test scheme must make sure B
130 * is also included in the memcg. At last, due to race conditions this filter
131 * can only guarantee that the page either belongs to the memcg tasks, or is
132 * a freed page.
133 */
Vladimir Davydov94a59fb2015-09-09 15:35:31 -0700134#ifdef CONFIG_MEMCG
Andi Kleen4fd466e2009-12-16 12:19:59 +0100135u64 hwpoison_filter_memcg;
136EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
137static int hwpoison_filter_task(struct page *p)
138{
Andi Kleen4fd466e2009-12-16 12:19:59 +0100139 if (!hwpoison_filter_memcg)
140 return 0;
141
Vladimir Davydov94a59fb2015-09-09 15:35:31 -0700142 if (page_cgroup_ino(p) != hwpoison_filter_memcg)
Andi Kleen4fd466e2009-12-16 12:19:59 +0100143 return -EINVAL;
144
145 return 0;
146}
147#else
148static int hwpoison_filter_task(struct page *p) { return 0; }
149#endif
150
Wu Fengguang7c116f22009-12-16 12:19:59 +0100151int hwpoison_filter(struct page *p)
152{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100153 if (!hwpoison_filter_enable)
154 return 0;
155
Wu Fengguang7c116f22009-12-16 12:19:59 +0100156 if (hwpoison_filter_dev(p))
157 return -EINVAL;
158
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100159 if (hwpoison_filter_flags(p))
160 return -EINVAL;
161
Andi Kleen4fd466e2009-12-16 12:19:59 +0100162 if (hwpoison_filter_task(p))
163 return -EINVAL;
164
Wu Fengguang7c116f22009-12-16 12:19:59 +0100165 return 0;
166}
Andi Kleen27df5062009-12-21 19:56:42 +0100167#else
168int hwpoison_filter(struct page *p)
169{
170 return 0;
171}
172#endif
173
Wu Fengguang7c116f22009-12-16 12:19:59 +0100174EXPORT_SYMBOL_GPL(hwpoison_filter);
175
Andi Kleen6a460792009-09-16 11:50:15 +0200176/*
Tony Luck7329bbe2011-12-13 09:27:58 -0800177 * Send all the processes who have the page mapped a signal.
178 * ``action optional'' if they are not immediately affected by the error
179 * ``action required'' if error happened in current execution context
Andi Kleen6a460792009-09-16 11:50:15 +0200180 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800181static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
182 unsigned long pfn, struct page *page, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200183{
184 struct siginfo si;
185 int ret;
186
Joe Perches11705322016-03-17 14:19:50 -0700187 pr_err("MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
188 pfn, t->comm, t->pid);
Andi Kleen6a460792009-09-16 11:50:15 +0200189 si.si_signo = SIGBUS;
190 si.si_errno = 0;
Andi Kleen6a460792009-09-16 11:50:15 +0200191 si.si_addr = (void *)addr;
192#ifdef __ARCH_SI_TRAPNO
193 si.si_trapno = trapno;
194#endif
Wanpeng Lif9121152013-09-11 14:22:52 -0700195 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
Tony Luck7329bbe2011-12-13 09:27:58 -0800196
Tony Lucka70ffca2014-06-04 16:10:59 -0700197 if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
Tony Luck7329bbe2011-12-13 09:27:58 -0800198 si.si_code = BUS_MCEERR_AR;
Tony Lucka70ffca2014-06-04 16:10:59 -0700199 ret = force_sig_info(SIGBUS, &si, current);
Tony Luck7329bbe2011-12-13 09:27:58 -0800200 } else {
201 /*
202 * Don't use force here, it's convenient if the signal
203 * can be temporarily blocked.
204 * This could cause a loop when the user sets SIGBUS
205 * to SIG_IGN, but hopefully no one will do that?
206 */
207 si.si_code = BUS_MCEERR_AO;
208 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
209 }
Andi Kleen6a460792009-09-16 11:50:15 +0200210 if (ret < 0)
Joe Perches11705322016-03-17 14:19:50 -0700211 pr_info("MCE: Error sending signal to %s:%d: %d\n",
212 t->comm, t->pid, ret);
Andi Kleen6a460792009-09-16 11:50:15 +0200213 return ret;
214}
215
216/*
Andi Kleen588f9ce2009-12-16 12:19:57 +0100217 * When a unknown page type is encountered drain as many buffers as possible
218 * in the hope to turn the page into a LRU or free page, which we can handle.
219 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100220void shake_page(struct page *p, int access)
Andi Kleen588f9ce2009-12-16 12:19:57 +0100221{
222 if (!PageSlab(p)) {
223 lru_add_drain_all();
224 if (PageLRU(p))
225 return;
Vlastimil Babkac0554322014-12-10 15:43:10 -0800226 drain_all_pages(page_zone(p));
Andi Kleen588f9ce2009-12-16 12:19:57 +0100227 if (PageLRU(p) || is_free_buddy_page(p))
228 return;
229 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100230
Andi Kleen588f9ce2009-12-16 12:19:57 +0100231 /*
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800232 * Only call shrink_node_slabs here (which would also shrink
233 * other caches) if access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100234 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800235 if (access)
236 drop_slab_node(page_to_nid(p));
Andi Kleen588f9ce2009-12-16 12:19:57 +0100237}
238EXPORT_SYMBOL_GPL(shake_page);
239
240/*
Andi Kleen6a460792009-09-16 11:50:15 +0200241 * Kill all processes that have a poisoned page mapped and then isolate
242 * the page.
243 *
244 * General strategy:
245 * Find all processes having the page mapped and kill them.
246 * But we keep a page reference around so that the page is not
247 * actually freed yet.
248 * Then stash the page away
249 *
250 * There's no convenient way to get back to mapped processes
251 * from the VMAs. So do a brute-force search over all
252 * running processes.
253 *
254 * Remember that machine checks are not common (or rather
255 * if they are common you have other problems), so this shouldn't
256 * be a performance issue.
257 *
258 * Also there are some races possible while we get from the
259 * error detection to actually handle it.
260 */
261
262struct to_kill {
263 struct list_head nd;
264 struct task_struct *tsk;
265 unsigned long addr;
Andi Kleen9033ae12010-09-27 23:36:05 +0200266 char addr_valid;
Andi Kleen6a460792009-09-16 11:50:15 +0200267};
268
269/*
270 * Failure handling: if we can't find or can't kill a process there's
271 * not much we can do. We just print a message and ignore otherwise.
272 */
273
274/*
275 * Schedule a process for later kill.
276 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
277 * TBD would GFP_NOIO be enough?
278 */
279static void add_to_kill(struct task_struct *tsk, struct page *p,
280 struct vm_area_struct *vma,
281 struct list_head *to_kill,
282 struct to_kill **tkc)
283{
284 struct to_kill *tk;
285
286 if (*tkc) {
287 tk = *tkc;
288 *tkc = NULL;
289 } else {
290 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
291 if (!tk) {
Joe Perches11705322016-03-17 14:19:50 -0700292 pr_err("MCE: Out of memory while machine check handling\n");
Andi Kleen6a460792009-09-16 11:50:15 +0200293 return;
294 }
295 }
296 tk->addr = page_address_in_vma(p, vma);
297 tk->addr_valid = 1;
298
299 /*
300 * In theory we don't have to kill when the page was
301 * munmaped. But it could be also a mremap. Since that's
302 * likely very rare kill anyways just out of paranoia, but use
303 * a SIGKILL because the error is not contained anymore.
304 */
305 if (tk->addr == -EFAULT) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200306 pr_info("MCE: Unable to find user space address %lx in %s\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200307 page_to_pfn(p), tsk->comm);
308 tk->addr_valid = 0;
309 }
310 get_task_struct(tsk);
311 tk->tsk = tsk;
312 list_add_tail(&tk->nd, to_kill);
313}
314
315/*
316 * Kill the processes that have been collected earlier.
317 *
318 * Only do anything when DOIT is set, otherwise just free the list
319 * (this is used for clean pages which do not need killing)
320 * Also when FAIL is set do a force kill because something went
321 * wrong earlier.
322 */
Tony Luck6751ed62012-07-11 10:20:47 -0700323static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -0800324 int fail, struct page *page, unsigned long pfn,
325 int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200326{
327 struct to_kill *tk, *next;
328
329 list_for_each_entry_safe (tk, next, to_kill, nd) {
Tony Luck6751ed62012-07-11 10:20:47 -0700330 if (forcekill) {
Andi Kleen6a460792009-09-16 11:50:15 +0200331 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200332 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200333 * make sure the process doesn't catch the
334 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200335 */
336 if (fail || tk->addr_valid == 0) {
Joe Perches11705322016-03-17 14:19:50 -0700337 pr_err("MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
338 pfn, tk->tsk->comm, tk->tsk->pid);
Andi Kleen6a460792009-09-16 11:50:15 +0200339 force_sig(SIGKILL, tk->tsk);
340 }
341
342 /*
343 * In theory the process could have mapped
344 * something else on the address in-between. We could
345 * check for that, but we need to tell the
346 * process anyways.
347 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800348 else if (kill_proc(tk->tsk, tk->addr, trapno,
349 pfn, page, flags) < 0)
Joe Perches11705322016-03-17 14:19:50 -0700350 pr_err("MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
351 pfn, tk->tsk->comm, tk->tsk->pid);
Andi Kleen6a460792009-09-16 11:50:15 +0200352 }
353 put_task_struct(tk->tsk);
354 kfree(tk);
355 }
356}
357
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700358/*
359 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
360 * on behalf of the thread group. Return task_struct of the (first found)
361 * dedicated thread if found, and return NULL otherwise.
362 *
363 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
364 * have to call rcu_read_lock/unlock() in this function.
365 */
366static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
Andi Kleen6a460792009-09-16 11:50:15 +0200367{
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700368 struct task_struct *t;
369
370 for_each_thread(tsk, t)
371 if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
372 return t;
373 return NULL;
374}
375
376/*
377 * Determine whether a given process is "early kill" process which expects
378 * to be signaled when some page under the process is hwpoisoned.
379 * Return task_struct of the dedicated thread (main thread unless explicitly
380 * specified) if the process is "early kill," and otherwise returns NULL.
381 */
382static struct task_struct *task_early_kill(struct task_struct *tsk,
383 int force_early)
384{
385 struct task_struct *t;
Andi Kleen6a460792009-09-16 11:50:15 +0200386 if (!tsk->mm)
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700387 return NULL;
Tony Luck74614de2014-06-04 16:11:01 -0700388 if (force_early)
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700389 return tsk;
390 t = find_early_kill_thread(tsk);
391 if (t)
392 return t;
393 if (sysctl_memory_failure_early_kill)
394 return tsk;
395 return NULL;
Andi Kleen6a460792009-09-16 11:50:15 +0200396}
397
398/*
399 * Collect processes when the error hit an anonymous page.
400 */
401static void collect_procs_anon(struct page *page, struct list_head *to_kill,
Tony Luck74614de2014-06-04 16:11:01 -0700402 struct to_kill **tkc, int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200403{
404 struct vm_area_struct *vma;
405 struct task_struct *tsk;
406 struct anon_vma *av;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700407 pgoff_t pgoff;
Andi Kleen6a460792009-09-16 11:50:15 +0200408
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000409 av = page_lock_anon_vma_read(page);
Andi Kleen6a460792009-09-16 11:50:15 +0200410 if (av == NULL) /* Not actually mapped anymore */
Peter Zijlstra9b679322011-06-27 16:18:09 -0700411 return;
412
Naoya Horiguchia0f7a752014-07-23 14:00:01 -0700413 pgoff = page_to_pgoff(page);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700414 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200415 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800416 struct anon_vma_chain *vmac;
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700417 struct task_struct *t = task_early_kill(tsk, force_early);
Rik van Riel5beb4932010-03-05 13:42:07 -0800418
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700419 if (!t)
Andi Kleen6a460792009-09-16 11:50:15 +0200420 continue;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700421 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
422 pgoff, pgoff) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800423 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200424 if (!page_mapped_in_vma(page, vma))
425 continue;
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700426 if (vma->vm_mm == t->mm)
427 add_to_kill(t, page, vma, to_kill, tkc);
Andi Kleen6a460792009-09-16 11:50:15 +0200428 }
429 }
Andi Kleen6a460792009-09-16 11:50:15 +0200430 read_unlock(&tasklist_lock);
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000431 page_unlock_anon_vma_read(av);
Andi Kleen6a460792009-09-16 11:50:15 +0200432}
433
434/*
435 * Collect processes when the error hit a file mapped page.
436 */
437static void collect_procs_file(struct page *page, struct list_head *to_kill,
Tony Luck74614de2014-06-04 16:11:01 -0700438 struct to_kill **tkc, int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200439{
440 struct vm_area_struct *vma;
441 struct task_struct *tsk;
Andi Kleen6a460792009-09-16 11:50:15 +0200442 struct address_space *mapping = page->mapping;
443
Davidlohr Buesod28eb9c2014-12-12 16:54:36 -0800444 i_mmap_lock_read(mapping);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700445 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200446 for_each_process(tsk) {
Naoya Horiguchia0f7a752014-07-23 14:00:01 -0700447 pgoff_t pgoff = page_to_pgoff(page);
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700448 struct task_struct *t = task_early_kill(tsk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200449
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700450 if (!t)
Andi Kleen6a460792009-09-16 11:50:15 +0200451 continue;
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -0700452 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
Andi Kleen6a460792009-09-16 11:50:15 +0200453 pgoff) {
454 /*
455 * Send early kill signal to tasks where a vma covers
456 * the page but the corrupted page is not necessarily
457 * mapped it in its pte.
458 * Assume applications who requested early kill want
459 * to be informed of all such data corruptions.
460 */
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700461 if (vma->vm_mm == t->mm)
462 add_to_kill(t, page, vma, to_kill, tkc);
Andi Kleen6a460792009-09-16 11:50:15 +0200463 }
464 }
Andi Kleen6a460792009-09-16 11:50:15 +0200465 read_unlock(&tasklist_lock);
Davidlohr Buesod28eb9c2014-12-12 16:54:36 -0800466 i_mmap_unlock_read(mapping);
Andi Kleen6a460792009-09-16 11:50:15 +0200467}
468
469/*
470 * Collect the processes who have the corrupted page mapped to kill.
471 * This is done in two steps for locking reasons.
472 * First preallocate one tokill structure outside the spin locks,
473 * so that we can kill at least one process reasonably reliable.
474 */
Tony Luck74614de2014-06-04 16:11:01 -0700475static void collect_procs(struct page *page, struct list_head *tokill,
476 int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200477{
478 struct to_kill *tk;
479
480 if (!page->mapping)
481 return;
482
483 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
484 if (!tk)
485 return;
486 if (PageAnon(page))
Tony Luck74614de2014-06-04 16:11:01 -0700487 collect_procs_anon(page, tokill, &tk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200488 else
Tony Luck74614de2014-06-04 16:11:01 -0700489 collect_procs_file(page, tokill, &tk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200490 kfree(tk);
491}
492
Andi Kleen6a460792009-09-16 11:50:15 +0200493static const char *action_name[] = {
Xie XiuQicc637b12015-06-24 16:57:30 -0700494 [MF_IGNORED] = "Ignored",
495 [MF_FAILED] = "Failed",
496 [MF_DELAYED] = "Delayed",
497 [MF_RECOVERED] = "Recovered",
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700498};
499
500static const char * const action_page_types[] = {
Xie XiuQicc637b12015-06-24 16:57:30 -0700501 [MF_MSG_KERNEL] = "reserved kernel page",
502 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
503 [MF_MSG_SLAB] = "kernel slab page",
504 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
505 [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned",
506 [MF_MSG_HUGE] = "huge page",
507 [MF_MSG_FREE_HUGE] = "free huge page",
508 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
509 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
510 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
511 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
512 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
513 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
514 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
515 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
516 [MF_MSG_CLEAN_LRU] = "clean LRU page",
517 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
518 [MF_MSG_BUDDY] = "free buddy page",
519 [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)",
520 [MF_MSG_UNKNOWN] = "unknown page",
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700521};
522
Andi Kleen6a460792009-09-16 11:50:15 +0200523/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100524 * XXX: It is possible that a page is isolated from LRU cache,
525 * and then kept in swap cache or failed to remove from page cache.
526 * The page count will stop it from being freed by unpoison.
527 * Stress tests should be aware of this memory leak problem.
528 */
529static int delete_from_lru_cache(struct page *p)
530{
531 if (!isolate_lru_page(p)) {
532 /*
533 * Clear sensible page flags, so that the buddy system won't
534 * complain when the page is unpoison-and-freed.
535 */
536 ClearPageActive(p);
537 ClearPageUnevictable(p);
538 /*
539 * drop the page count elevated by isolate_lru_page()
540 */
541 page_cache_release(p);
542 return 0;
543 }
544 return -EIO;
545}
546
547/*
Andi Kleen6a460792009-09-16 11:50:15 +0200548 * Error hit kernel page.
549 * Do nothing, try to be lucky and not touch this instead. For a few cases we
550 * could be more sophisticated.
551 */
552static int me_kernel(struct page *p, unsigned long pfn)
553{
Xie XiuQicc637b12015-06-24 16:57:30 -0700554 return MF_IGNORED;
Andi Kleen6a460792009-09-16 11:50:15 +0200555}
556
557/*
558 * Page in unknown state. Do nothing.
559 */
560static int me_unknown(struct page *p, unsigned long pfn)
561{
Joe Perches11705322016-03-17 14:19:50 -0700562 pr_err("MCE %#lx: Unknown page state\n", pfn);
Xie XiuQicc637b12015-06-24 16:57:30 -0700563 return MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200564}
565
566/*
Andi Kleen6a460792009-09-16 11:50:15 +0200567 * Clean (or cleaned) page cache page.
568 */
569static int me_pagecache_clean(struct page *p, unsigned long pfn)
570{
571 int err;
Xie XiuQicc637b12015-06-24 16:57:30 -0700572 int ret = MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200573 struct address_space *mapping;
574
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100575 delete_from_lru_cache(p);
576
Andi Kleen6a460792009-09-16 11:50:15 +0200577 /*
578 * For anonymous pages we're done the only reference left
579 * should be the one m_f() holds.
580 */
581 if (PageAnon(p))
Xie XiuQicc637b12015-06-24 16:57:30 -0700582 return MF_RECOVERED;
Andi Kleen6a460792009-09-16 11:50:15 +0200583
584 /*
585 * Now truncate the page in the page cache. This is really
586 * more like a "temporary hole punch"
587 * Don't do this for block devices when someone else
588 * has a reference, because it could be file system metadata
589 * and that's not safe to truncate.
590 */
591 mapping = page_mapping(p);
592 if (!mapping) {
593 /*
594 * Page has been teared down in the meanwhile
595 */
Xie XiuQicc637b12015-06-24 16:57:30 -0700596 return MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200597 }
598
599 /*
600 * Truncation is a bit tricky. Enable it per file system for now.
601 *
602 * Open: to take i_mutex or not for this? Right now we don't.
603 */
604 if (mapping->a_ops->error_remove_page) {
605 err = mapping->a_ops->error_remove_page(mapping, p);
606 if (err != 0) {
Joe Perches11705322016-03-17 14:19:50 -0700607 pr_info("MCE %#lx: Failed to punch page: %d\n",
608 pfn, err);
Andi Kleen6a460792009-09-16 11:50:15 +0200609 } else if (page_has_private(p) &&
610 !try_to_release_page(p, GFP_NOIO)) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200611 pr_info("MCE %#lx: failed to release buffers\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200612 } else {
Xie XiuQicc637b12015-06-24 16:57:30 -0700613 ret = MF_RECOVERED;
Andi Kleen6a460792009-09-16 11:50:15 +0200614 }
615 } else {
616 /*
617 * If the file system doesn't support it just invalidate
618 * This fails on dirty or anything with private pages
619 */
620 if (invalidate_inode_page(p))
Xie XiuQicc637b12015-06-24 16:57:30 -0700621 ret = MF_RECOVERED;
Andi Kleen6a460792009-09-16 11:50:15 +0200622 else
Joe Perches11705322016-03-17 14:19:50 -0700623 pr_info("MCE %#lx: Failed to invalidate\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200624 }
625 return ret;
626}
627
628/*
Zhi Yong Wu549543d2014-01-21 15:49:08 -0800629 * Dirty pagecache page
Andi Kleen6a460792009-09-16 11:50:15 +0200630 * Issues: when the error hit a hole page the error is not properly
631 * propagated.
632 */
633static int me_pagecache_dirty(struct page *p, unsigned long pfn)
634{
635 struct address_space *mapping = page_mapping(p);
636
637 SetPageError(p);
638 /* TBD: print more information about the file. */
639 if (mapping) {
640 /*
641 * IO error will be reported by write(), fsync(), etc.
642 * who check the mapping.
643 * This way the application knows that something went
644 * wrong with its dirty file data.
645 *
646 * There's one open issue:
647 *
648 * The EIO will be only reported on the next IO
649 * operation and then cleared through the IO map.
650 * Normally Linux has two mechanisms to pass IO error
651 * first through the AS_EIO flag in the address space
652 * and then through the PageError flag in the page.
653 * Since we drop pages on memory failure handling the
654 * only mechanism open to use is through AS_AIO.
655 *
656 * This has the disadvantage that it gets cleared on
657 * the first operation that returns an error, while
658 * the PageError bit is more sticky and only cleared
659 * when the page is reread or dropped. If an
660 * application assumes it will always get error on
661 * fsync, but does other operations on the fd before
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300662 * and the page is dropped between then the error
Andi Kleen6a460792009-09-16 11:50:15 +0200663 * will not be properly reported.
664 *
665 * This can already happen even without hwpoisoned
666 * pages: first on metadata IO errors (which only
667 * report through AS_EIO) or when the page is dropped
668 * at the wrong time.
669 *
670 * So right now we assume that the application DTRT on
671 * the first EIO, but we're not worse than other parts
672 * of the kernel.
673 */
674 mapping_set_error(mapping, EIO);
675 }
676
677 return me_pagecache_clean(p, pfn);
678}
679
680/*
681 * Clean and dirty swap cache.
682 *
683 * Dirty swap cache page is tricky to handle. The page could live both in page
684 * cache and swap cache(ie. page is freshly swapped in). So it could be
685 * referenced concurrently by 2 types of PTEs:
686 * normal PTEs and swap PTEs. We try to handle them consistently by calling
687 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
688 * and then
689 * - clear dirty bit to prevent IO
690 * - remove from LRU
691 * - but keep in the swap cache, so that when we return to it on
692 * a later page fault, we know the application is accessing
693 * corrupted data and shall be killed (we installed simple
694 * interception code in do_swap_page to catch it).
695 *
696 * Clean swap cache pages can be directly isolated. A later page fault will
697 * bring in the known good data from disk.
698 */
699static int me_swapcache_dirty(struct page *p, unsigned long pfn)
700{
Andi Kleen6a460792009-09-16 11:50:15 +0200701 ClearPageDirty(p);
702 /* Trigger EIO in shmem: */
703 ClearPageUptodate(p);
704
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100705 if (!delete_from_lru_cache(p))
Xie XiuQicc637b12015-06-24 16:57:30 -0700706 return MF_DELAYED;
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100707 else
Xie XiuQicc637b12015-06-24 16:57:30 -0700708 return MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200709}
710
711static int me_swapcache_clean(struct page *p, unsigned long pfn)
712{
Andi Kleen6a460792009-09-16 11:50:15 +0200713 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800714
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100715 if (!delete_from_lru_cache(p))
Xie XiuQicc637b12015-06-24 16:57:30 -0700716 return MF_RECOVERED;
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100717 else
Xie XiuQicc637b12015-06-24 16:57:30 -0700718 return MF_FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200719}
720
721/*
722 * Huge pages. Needs work.
723 * Issues:
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900724 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
725 * To narrow down kill region to one page, we need to break up pmd.
Andi Kleen6a460792009-09-16 11:50:15 +0200726 */
727static int me_huge_page(struct page *p, unsigned long pfn)
728{
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900729 int res = 0;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900730 struct page *hpage = compound_head(p);
Naoya Horiguchi2491ffe2015-06-24 16:56:53 -0700731
732 if (!PageHuge(hpage))
733 return MF_DELAYED;
734
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900735 /*
736 * We can safely recover from error on free or reserved (i.e.
737 * not in-use) hugepage by dequeuing it from freelist.
738 * To check whether a hugepage is in-use or not, we can't use
739 * page->lru because it can be used in other hugepage operations,
740 * such as __unmap_hugepage_range() and gather_surplus_pages().
741 * So instead we use page_mapping() and PageAnon().
742 * We assume that this function is called with page lock held,
743 * so there is no race between isolation and mapping/unmapping.
744 */
745 if (!(page_mapping(hpage) || PageAnon(hpage))) {
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900746 res = dequeue_hwpoisoned_huge_page(hpage);
747 if (!res)
Xie XiuQicc637b12015-06-24 16:57:30 -0700748 return MF_RECOVERED;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900749 }
Xie XiuQicc637b12015-06-24 16:57:30 -0700750 return MF_DELAYED;
Andi Kleen6a460792009-09-16 11:50:15 +0200751}
752
753/*
754 * Various page states we can handle.
755 *
756 * A page state is defined by its current page->flags bits.
757 * The table matches them in order and calls the right handler.
758 *
759 * This is quite tricky because we can access page at any time
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300760 * in its live cycle, so all accesses have to be extremely careful.
Andi Kleen6a460792009-09-16 11:50:15 +0200761 *
762 * This is not complete. More states could be added.
763 * For any missing state don't attempt recovery.
764 */
765
766#define dirty (1UL << PG_dirty)
767#define sc (1UL << PG_swapcache)
768#define unevict (1UL << PG_unevictable)
769#define mlock (1UL << PG_mlocked)
770#define writeback (1UL << PG_writeback)
771#define lru (1UL << PG_lru)
772#define swapbacked (1UL << PG_swapbacked)
773#define head (1UL << PG_head)
Andi Kleen6a460792009-09-16 11:50:15 +0200774#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200775#define reserved (1UL << PG_reserved)
776
777static struct page_state {
778 unsigned long mask;
779 unsigned long res;
Xie XiuQicc637b12015-06-24 16:57:30 -0700780 enum mf_action_page_type type;
Andi Kleen6a460792009-09-16 11:50:15 +0200781 int (*action)(struct page *p, unsigned long pfn);
782} error_states[] = {
Xie XiuQicc637b12015-06-24 16:57:30 -0700783 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100784 /*
785 * free pages are specially detected outside this table:
786 * PG_buddy pages only make a small fraction of all free pages.
787 */
Andi Kleen6a460792009-09-16 11:50:15 +0200788
789 /*
790 * Could in theory check if slab page is free or if we can drop
791 * currently unused objects without touching them. But just
792 * treat it as standard kernel for now.
793 */
Xie XiuQicc637b12015-06-24 16:57:30 -0700794 { slab, slab, MF_MSG_SLAB, me_kernel },
Andi Kleen6a460792009-09-16 11:50:15 +0200795
Xie XiuQicc637b12015-06-24 16:57:30 -0700796 { head, head, MF_MSG_HUGE, me_huge_page },
Andi Kleen6a460792009-09-16 11:50:15 +0200797
Xie XiuQicc637b12015-06-24 16:57:30 -0700798 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
799 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200800
Xie XiuQicc637b12015-06-24 16:57:30 -0700801 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
802 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200803
Xie XiuQicc637b12015-06-24 16:57:30 -0700804 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
805 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
Naoya Horiguchi5f4b9fc2013-02-22 16:35:53 -0800806
Xie XiuQicc637b12015-06-24 16:57:30 -0700807 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
808 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200809
810 /*
811 * Catchall entry: must be at end.
812 */
Xie XiuQicc637b12015-06-24 16:57:30 -0700813 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
Andi Kleen6a460792009-09-16 11:50:15 +0200814};
815
Andi Kleen2326c462009-12-16 12:20:00 +0100816#undef dirty
817#undef sc
818#undef unevict
819#undef mlock
820#undef writeback
821#undef lru
822#undef swapbacked
823#undef head
Andi Kleen2326c462009-12-16 12:20:00 +0100824#undef slab
825#undef reserved
826
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800827/*
828 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
829 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
830 */
Xie XiuQicc3e2af2015-06-24 16:57:33 -0700831static void action_result(unsigned long pfn, enum mf_action_page_type type,
832 enum mf_result result)
Andi Kleen6a460792009-09-16 11:50:15 +0200833{
Xie XiuQi97f0b132015-06-24 16:57:36 -0700834 trace_memory_failure_event(pfn, type, result);
835
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700836 pr_err("MCE %#lx: recovery action for %s: %s\n",
837 pfn, action_page_types[type], action_name[result]);
Andi Kleen6a460792009-09-16 11:50:15 +0200838}
839
840static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100841 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200842{
843 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200844 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200845
846 result = ps->action(p, pfn);
Wu Fengguang7456b042009-10-19 08:15:01 +0200847
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100848 count = page_count(p) - 1;
Xie XiuQicc637b12015-06-24 16:57:30 -0700849 if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
Wu Fengguang138ce282009-12-16 12:19:58 +0100850 count--;
851 if (count != 0) {
Joe Perches11705322016-03-17 14:19:50 -0700852 pr_err("MCE %#lx: %s still referenced by %d users\n",
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700853 pfn, action_page_types[ps->type], count);
Xie XiuQicc637b12015-06-24 16:57:30 -0700854 result = MF_FAILED;
Wu Fengguang138ce282009-12-16 12:19:58 +0100855 }
Naoya Horiguchi64d37a22015-04-15 16:13:05 -0700856 action_result(pfn, ps->type, result);
Andi Kleen6a460792009-09-16 11:50:15 +0200857
858 /* Could do more checks here if page looks ok */
859 /*
860 * Could adjust zone counters here to correct for the missing page.
861 */
862
Xie XiuQicc637b12015-06-24 16:57:30 -0700863 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200864}
865
Naoya Horiguchiead07f62015-06-24 16:56:48 -0700866/**
867 * get_hwpoison_page() - Get refcount for memory error handling:
868 * @page: raw error page (hit by memory error)
869 *
870 * Return: return 0 if failed to grab the refcount, otherwise true (some
871 * non-zero value.)
872 */
873int get_hwpoison_page(struct page *page)
874{
875 struct page *head = compound_head(page);
876
Naoya Horiguchi4e41a302016-01-15 16:54:07 -0800877 if (!PageHuge(head) && PageTransHuge(head)) {
Naoya Horiguchi98ed2b02015-08-06 15:47:04 -0700878 /*
879 * Non anonymous thp exists only in allocation/free time. We
880 * can't handle such a case correctly, so let's give it up.
881 * This should be better than triggering BUG_ON when kernel
882 * tries to touch the "partially handled" page.
883 */
884 if (!PageAnon(head)) {
885 pr_err("MCE: %#lx: non anonymous thp\n",
886 page_to_pfn(page));
887 return 0;
888 }
Naoya Horiguchiead07f62015-06-24 16:56:48 -0700889 }
890
Naoya Horiguchi4e41a302016-01-15 16:54:07 -0800891 return get_page_unless_zero(head);
Naoya Horiguchiead07f62015-06-24 16:56:48 -0700892}
893EXPORT_SYMBOL_GPL(get_hwpoison_page);
894
Andi Kleen6a460792009-09-16 11:50:15 +0200895/*
896 * Do all that is necessary to remove user space mappings. Unmap
897 * the pages and send SIGBUS to the processes if the data was dirty.
898 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100899static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800900 int trapno, int flags, struct page **hpagep)
Andi Kleen6a460792009-09-16 11:50:15 +0200901{
902 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
903 struct address_space *mapping;
904 LIST_HEAD(tokill);
905 int ret;
Tony Luck6751ed62012-07-11 10:20:47 -0700906 int kill = 1, forcekill;
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800907 struct page *hpage = *hpagep;
Andi Kleen6a460792009-09-16 11:50:15 +0200908
Naoya Horiguchi93a9eb32014-07-30 16:08:28 -0700909 /*
910 * Here we are interested only in user-mapped pages, so skip any
911 * other types of pages.
912 */
913 if (PageReserved(p) || PageSlab(p))
914 return SWAP_SUCCESS;
915 if (!(PageLRU(hpage) || PageHuge(p)))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100916 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200917
Andi Kleen6a460792009-09-16 11:50:15 +0200918 /*
919 * This check implies we don't kill processes if their pages
920 * are in the swap cache early. Those are always late kills.
921 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900922 if (!page_mapped(hpage))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100923 return SWAP_SUCCESS;
924
Naoya Horiguchi52089b12014-07-30 16:08:30 -0700925 if (PageKsm(p)) {
926 pr_err("MCE %#lx: can't handle KSM pages.\n", pfn);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100927 return SWAP_FAIL;
Naoya Horiguchi52089b12014-07-30 16:08:30 -0700928 }
Andi Kleen6a460792009-09-16 11:50:15 +0200929
930 if (PageSwapCache(p)) {
Joe Perches11705322016-03-17 14:19:50 -0700931 pr_err("MCE %#lx: keeping poisoned page in swap cache\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200932 ttu |= TTU_IGNORE_HWPOISON;
933 }
934
935 /*
936 * Propagate the dirty bit from PTEs to struct page first, because we
937 * need this to decide if we should kill or just drop the page.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100938 * XXX: the dirty test could be racy: set_page_dirty() may not always
939 * be called inside page lock (it's recommended but not enforced).
Andi Kleen6a460792009-09-16 11:50:15 +0200940 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900941 mapping = page_mapping(hpage);
Tony Luck6751ed62012-07-11 10:20:47 -0700942 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900943 mapping_cap_writeback_dirty(mapping)) {
944 if (page_mkclean(hpage)) {
945 SetPageDirty(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +0200946 } else {
947 kill = 0;
948 ttu |= TTU_IGNORE_HWPOISON;
Joe Perches11705322016-03-17 14:19:50 -0700949 pr_info("MCE %#lx: corrupted page was clean: dropped without side effects\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200950 pfn);
951 }
952 }
953
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800954 /*
Andi Kleen6a460792009-09-16 11:50:15 +0200955 * First collect all the processes that have the page
956 * mapped in dirty form. This has to be done before try_to_unmap,
957 * because ttu takes the rmap data structures down.
958 *
959 * Error handling: We ignore errors here because
960 * there's nothing that can be done.
961 */
962 if (kill)
Naoya Horiguchi415c64c2015-06-24 16:56:45 -0700963 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
Andi Kleen6a460792009-09-16 11:50:15 +0200964
Naoya Horiguchi415c64c2015-06-24 16:56:45 -0700965 ret = try_to_unmap(hpage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +0200966 if (ret != SWAP_SUCCESS)
Joe Perches11705322016-03-17 14:19:50 -0700967 pr_err("MCE %#lx: failed to unmap page (mapcount=%d)\n",
968 pfn, page_mapcount(hpage));
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800969
Andi Kleen6a460792009-09-16 11:50:15 +0200970 /*
971 * Now that the dirty bit has been propagated to the
972 * struct page and all unmaps done we can decide if
973 * killing is needed or not. Only kill when the page
Tony Luck6751ed62012-07-11 10:20:47 -0700974 * was dirty or the process is not restartable,
975 * otherwise the tokill list is merely
Andi Kleen6a460792009-09-16 11:50:15 +0200976 * freed. When there was a problem unmapping earlier
977 * use a more force-full uncatchable kill to prevent
978 * any accesses to the poisoned memory.
979 */
Naoya Horiguchi415c64c2015-06-24 16:56:45 -0700980 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
Tony Luck6751ed62012-07-11 10:20:47 -0700981 kill_procs(&tokill, forcekill, trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -0800982 ret != SWAP_SUCCESS, p, pfn, flags);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100983
984 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200985}
986
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900987static void set_page_hwpoison_huge_page(struct page *hpage)
988{
989 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -0700990 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900991 for (i = 0; i < nr_pages; i++)
992 SetPageHWPoison(hpage + i);
993}
994
995static void clear_page_hwpoison_huge_page(struct page *hpage)
996{
997 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -0700998 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900999 for (i = 0; i < nr_pages; i++)
1000 ClearPageHWPoison(hpage + i);
1001}
1002
Tony Luckcd42f4a2011-12-15 10:48:12 -08001003/**
1004 * memory_failure - Handle memory failure of a page.
1005 * @pfn: Page Number of the corrupted page
1006 * @trapno: Trap number reported in the signal to user space.
1007 * @flags: fine tune action taken
1008 *
1009 * This function is called by the low level machine check code
1010 * of an architecture when it detects hardware memory corruption
1011 * of a page. It tries its best to recover, which includes
1012 * dropping pages, killing processes etc.
1013 *
1014 * The function is primarily of use for corruptions that
1015 * happen outside the current execution context (e.g. when
1016 * detected by a background scrubber)
1017 *
1018 * Must run in process context (e.g. a work queue) with interrupts
1019 * enabled and no spinlocks hold.
1020 */
1021int memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +02001022{
1023 struct page_state *ps;
1024 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001025 struct page *hpage;
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001026 struct page *orig_head;
Andi Kleen6a460792009-09-16 11:50:15 +02001027 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001028 unsigned int nr_pages;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001029 unsigned long page_flags;
Andi Kleen6a460792009-09-16 11:50:15 +02001030
1031 if (!sysctl_memory_failure_recovery)
1032 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1033
1034 if (!pfn_valid(pfn)) {
Joe Perches11705322016-03-17 14:19:50 -07001035 pr_err("MCE %#lx: memory outside kernel control\n", pfn);
Wu Fengguanga7560fc2009-12-16 12:19:57 +01001036 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +02001037 }
1038
1039 p = pfn_to_page(pfn);
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001040 orig_head = hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +02001041 if (TestSetPageHWPoison(p)) {
Joe Perches11705322016-03-17 14:19:50 -07001042 pr_err("MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001043 return 0;
1044 }
1045
Naoya Horiguchi4db0e952013-02-22 16:34:05 -08001046 /*
1047 * Currently errors on hugetlbfs pages are measured in hugepage units,
1048 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1049 * transparent hugepages, they are supposed to be split and error
1050 * measurement is done in normal page units. So nr_pages should be one
1051 * in this case.
1052 */
1053 if (PageHuge(p))
1054 nr_pages = 1 << compound_order(hpage);
1055 else /* normal page or thp */
1056 nr_pages = 1;
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001057 num_poisoned_pages_add(nr_pages);
Andi Kleen6a460792009-09-16 11:50:15 +02001058
1059 /*
1060 * We need/can do nothing about count=0 pages.
1061 * 1) it's a free page, and therefore in safe hand:
1062 * prep_new_page() will be the gate keeper.
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001063 * 2) it's a free hugepage, which is also safe:
1064 * an affected hugepage will be dequeued from hugepage freelist,
1065 * so there's no concern about reusing it ever after.
1066 * 3) it's part of a non-compound high order page.
Andi Kleen6a460792009-09-16 11:50:15 +02001067 * Implies some kernel user: cannot stop them from
1068 * R/W the page; let's pray that the page has been
1069 * used and will be freed some time later.
1070 * In fact it's dangerous to directly bump up page count from 0,
1071 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1072 */
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001073 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001074 if (is_free_buddy_page(p)) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001075 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001076 return 0;
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001077 } else if (PageHuge(hpage)) {
1078 /*
Chen Yucongb9851942014-05-22 11:54:15 -07001079 * Check "filter hit" and "race with other subpage."
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001080 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001081 lock_page(hpage);
Chen Yucongb9851942014-05-22 11:54:15 -07001082 if (PageHWPoison(hpage)) {
1083 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1084 || (p != hpage && TestSetPageHWPoison(hpage))) {
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001085 num_poisoned_pages_sub(nr_pages);
Chen Yucongb9851942014-05-22 11:54:15 -07001086 unlock_page(hpage);
1087 return 0;
1088 }
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001089 }
1090 set_page_hwpoison_huge_page(hpage);
1091 res = dequeue_hwpoisoned_huge_page(hpage);
Xie XiuQicc637b12015-06-24 16:57:30 -07001092 action_result(pfn, MF_MSG_FREE_HUGE,
1093 res ? MF_IGNORED : MF_DELAYED);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001094 unlock_page(hpage);
1095 return res;
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001096 } else {
Xie XiuQicc637b12015-06-24 16:57:30 -07001097 action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001098 return -EBUSY;
1099 }
Andi Kleen6a460792009-09-16 11:50:15 +02001100 }
1101
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001102 if (!PageHuge(p) && PageTransHuge(hpage)) {
Kirill A. Shutemov4d2fa962016-01-15 16:54:00 -08001103 lock_page(hpage);
Wanpeng Li7f6bf392015-08-14 15:35:08 -07001104 if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) {
Kirill A. Shutemov4d2fa962016-01-15 16:54:00 -08001105 unlock_page(hpage);
Wanpeng Li7f6bf392015-08-14 15:35:08 -07001106 if (!PageAnon(hpage))
1107 pr_err("MCE: %#lx: non anonymous thp\n", pfn);
1108 else
1109 pr_err("MCE: %#lx: thp split failed\n", pfn);
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001110 if (TestClearPageHWPoison(p))
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001111 num_poisoned_pages_sub(nr_pages);
Wanpeng Li665d9da2015-09-08 15:03:21 -07001112 put_hwpoison_page(p);
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001113 return -EBUSY;
1114 }
Kirill A. Shutemov4d2fa962016-01-15 16:54:00 -08001115 unlock_page(hpage);
Naoya Horiguchi4e41a302016-01-15 16:54:07 -08001116 get_hwpoison_page(p);
1117 put_hwpoison_page(hpage);
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001118 VM_BUG_ON_PAGE(!page_count(p), p);
1119 hpage = compound_head(p);
1120 }
1121
Andi Kleen6a460792009-09-16 11:50:15 +02001122 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001123 * We ignore non-LRU pages for good reasons.
1124 * - PG_locked is only well defined for LRU pages and a few others
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001125 * - to avoid races with __SetPageLocked()
Wu Fengguange43c3af2009-09-29 13:16:20 +08001126 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1127 * The check (unnecessarily) ignores LRU pages being isolated and
1128 * walked by the page reclaim code, however that's not a big loss.
1129 */
Naoya Horiguchi09789e52015-05-05 16:23:35 -07001130 if (!PageHuge(p)) {
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001131 if (!PageLRU(p))
1132 shake_page(p, 0);
1133 if (!PageLRU(p)) {
Jin Dongmingaf241a02011-02-01 15:52:41 -08001134 /*
1135 * shake_page could have turned it free.
1136 */
1137 if (is_free_buddy_page(p)) {
Wanpeng Li2d421ac2013-09-30 13:45:23 -07001138 if (flags & MF_COUNT_INCREASED)
Xie XiuQicc637b12015-06-24 16:57:30 -07001139 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
Wanpeng Li2d421ac2013-09-30 13:45:23 -07001140 else
Xie XiuQicc637b12015-06-24 16:57:30 -07001141 action_result(pfn, MF_MSG_BUDDY_2ND,
1142 MF_DELAYED);
Jin Dongmingaf241a02011-02-01 15:52:41 -08001143 return 0;
1144 }
Andi Kleen0474a602009-12-16 12:20:00 +01001145 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001146 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001147
Jens Axboe7eaceac2011-03-10 08:52:07 +01001148 lock_page(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001149
1150 /*
Andi Kleenf37d4292014-08-06 16:06:49 -07001151 * The page could have changed compound pages during the locking.
1152 * If this happens just bail out.
1153 */
Naoya Horiguchi415c64c2015-06-24 16:56:45 -07001154 if (PageCompound(p) && compound_head(p) != orig_head) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001155 action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
Andi Kleenf37d4292014-08-06 16:06:49 -07001156 res = -EBUSY;
1157 goto out;
1158 }
1159
1160 /*
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001161 * We use page flags to determine what action should be taken, but
1162 * the flags can be modified by the error containment action. One
1163 * example is an mlocked page, where PG_mlocked is cleared by
1164 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1165 * correctly, we save a copy of the page flags at this time.
1166 */
1167 page_flags = p->flags;
1168
1169 /*
Wu Fengguang847ce402009-12-16 12:19:58 +01001170 * unpoison always clear PG_hwpoison inside page lock
1171 */
1172 if (!PageHWPoison(p)) {
Joe Perches11705322016-03-17 14:19:50 -07001173 pr_err("MCE %#lx: just unpoisoned\n", pfn);
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001174 num_poisoned_pages_sub(nr_pages);
Naoya Horiguchia09233f2015-08-06 15:46:58 -07001175 unlock_page(hpage);
Wanpeng Li665d9da2015-09-08 15:03:21 -07001176 put_hwpoison_page(hpage);
Naoya Horiguchia09233f2015-08-06 15:46:58 -07001177 return 0;
Wu Fengguang847ce402009-12-16 12:19:58 +01001178 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001179 if (hwpoison_filter(p)) {
1180 if (TestClearPageHWPoison(p))
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001181 num_poisoned_pages_sub(nr_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001182 unlock_page(hpage);
Wanpeng Li665d9da2015-09-08 15:03:21 -07001183 put_hwpoison_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001184 return 0;
1185 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001186
Chen Yucong0bc1f8b2014-07-02 15:22:37 -07001187 if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p))
1188 goto identify_page_state;
1189
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001190 /*
1191 * For error on the tail page, we should set PG_hwpoison
1192 * on the head page to show that the hugepage is hwpoisoned
1193 */
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001194 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001195 action_result(pfn, MF_MSG_POISONED_HUGE, MF_IGNORED);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001196 unlock_page(hpage);
Wanpeng Li665d9da2015-09-08 15:03:21 -07001197 put_hwpoison_page(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001198 return 0;
1199 }
1200 /*
1201 * Set PG_hwpoison on all pages in an error hugepage,
1202 * because containment is done in hugepage unit for now.
1203 * Since we have done TestSetPageHWPoison() for the head page with
1204 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1205 */
1206 if (PageHuge(p))
1207 set_page_hwpoison_huge_page(hpage);
1208
Naoya Horiguchi6edd6cc2014-06-04 16:10:35 -07001209 /*
1210 * It's very difficult to mess with pages currently under IO
1211 * and in many cases impossible, so we just avoid it here.
1212 */
Andi Kleen6a460792009-09-16 11:50:15 +02001213 wait_on_page_writeback(p);
1214
1215 /*
1216 * Now take care of user space mappings.
Minchan Kime64a7822011-03-22 16:32:44 -07001217 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001218 *
1219 * When the raw error page is thp tail page, hpage points to the raw
1220 * page after thp split.
Andi Kleen6a460792009-09-16 11:50:15 +02001221 */
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001222 if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)
1223 != SWAP_SUCCESS) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001224 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001225 res = -EBUSY;
1226 goto out;
1227 }
Andi Kleen6a460792009-09-16 11:50:15 +02001228
1229 /*
1230 * Torn down by someone else?
1231 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001232 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Xie XiuQicc637b12015-06-24 16:57:30 -07001233 action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001234 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001235 goto out;
1236 }
1237
Chen Yucong0bc1f8b2014-07-02 15:22:37 -07001238identify_page_state:
Andi Kleen6a460792009-09-16 11:50:15 +02001239 res = -EBUSY;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001240 /*
1241 * The first check uses the current page flags which may not have any
1242 * relevant information. The second check with the saved page flagss is
1243 * carried out only if the first check can't determine the page status.
1244 */
1245 for (ps = error_states;; ps++)
1246 if ((p->flags & ps->mask) == ps->res)
Andi Kleen6a460792009-09-16 11:50:15 +02001247 break;
Wanpeng Li841fcc52013-09-11 14:22:50 -07001248
1249 page_flags |= (p->flags & (1UL << PG_dirty));
1250
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001251 if (!ps->mask)
1252 for (ps = error_states;; ps++)
1253 if ((page_flags & ps->mask) == ps->res)
1254 break;
1255 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001256out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001257 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001258 return res;
1259}
Tony Luckcd42f4a2011-12-15 10:48:12 -08001260EXPORT_SYMBOL_GPL(memory_failure);
Wu Fengguang847ce402009-12-16 12:19:58 +01001261
Huang Yingea8f5fb2011-07-13 13:14:27 +08001262#define MEMORY_FAILURE_FIFO_ORDER 4
1263#define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1264
1265struct memory_failure_entry {
1266 unsigned long pfn;
1267 int trapno;
1268 int flags;
1269};
1270
1271struct memory_failure_cpu {
1272 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1273 MEMORY_FAILURE_FIFO_SIZE);
1274 spinlock_t lock;
1275 struct work_struct work;
1276};
1277
1278static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1279
1280/**
1281 * memory_failure_queue - Schedule handling memory failure of a page.
1282 * @pfn: Page Number of the corrupted page
1283 * @trapno: Trap number reported in the signal to user space.
1284 * @flags: Flags for memory failure handling
1285 *
1286 * This function is called by the low level hardware error handler
1287 * when it detects hardware memory corruption of a page. It schedules
1288 * the recovering of error page, including dropping pages, killing
1289 * processes etc.
1290 *
1291 * The function is primarily of use for corruptions that
1292 * happen outside the current execution context (e.g. when
1293 * detected by a background scrubber)
1294 *
1295 * Can run in IRQ context.
1296 */
1297void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1298{
1299 struct memory_failure_cpu *mf_cpu;
1300 unsigned long proc_flags;
1301 struct memory_failure_entry entry = {
1302 .pfn = pfn,
1303 .trapno = trapno,
1304 .flags = flags,
1305 };
1306
1307 mf_cpu = &get_cpu_var(memory_failure_cpu);
1308 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
Stefani Seibold498d3192013-11-14 14:32:17 -08001309 if (kfifo_put(&mf_cpu->fifo, entry))
Huang Yingea8f5fb2011-07-13 13:14:27 +08001310 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1311 else
Joe Perches8e33a522013-07-25 11:53:25 -07001312 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
Huang Yingea8f5fb2011-07-13 13:14:27 +08001313 pfn);
1314 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1315 put_cpu_var(memory_failure_cpu);
1316}
1317EXPORT_SYMBOL_GPL(memory_failure_queue);
1318
1319static void memory_failure_work_func(struct work_struct *work)
1320{
1321 struct memory_failure_cpu *mf_cpu;
1322 struct memory_failure_entry entry = { 0, };
1323 unsigned long proc_flags;
1324 int gotten;
1325
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001326 mf_cpu = this_cpu_ptr(&memory_failure_cpu);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001327 for (;;) {
1328 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1329 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1330 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1331 if (!gotten)
1332 break;
Naveen N. Raocf870c72013-07-10 14:57:01 +05301333 if (entry.flags & MF_SOFT_OFFLINE)
1334 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1335 else
1336 memory_failure(entry.pfn, entry.trapno, entry.flags);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001337 }
1338}
1339
1340static int __init memory_failure_init(void)
1341{
1342 struct memory_failure_cpu *mf_cpu;
1343 int cpu;
1344
1345 for_each_possible_cpu(cpu) {
1346 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1347 spin_lock_init(&mf_cpu->lock);
1348 INIT_KFIFO(mf_cpu->fifo);
1349 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1350 }
1351
1352 return 0;
1353}
1354core_initcall(memory_failure_init);
1355
Naoya Horiguchia5f65102015-11-05 18:47:26 -08001356#define unpoison_pr_info(fmt, pfn, rs) \
1357({ \
1358 if (__ratelimit(rs)) \
1359 pr_info(fmt, pfn); \
1360})
1361
Wu Fengguang847ce402009-12-16 12:19:58 +01001362/**
1363 * unpoison_memory - Unpoison a previously poisoned page
1364 * @pfn: Page number of the to be unpoisoned page
1365 *
1366 * Software-unpoison a page that has been poisoned by
1367 * memory_failure() earlier.
1368 *
1369 * This is only done on the software-level, so it only works
1370 * for linux injected failures, not real hardware failures
1371 *
1372 * Returns 0 for success, otherwise -errno.
1373 */
1374int unpoison_memory(unsigned long pfn)
1375{
1376 struct page *page;
1377 struct page *p;
1378 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001379 unsigned int nr_pages;
Naoya Horiguchia5f65102015-11-05 18:47:26 -08001380 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
1381 DEFAULT_RATELIMIT_BURST);
Wu Fengguang847ce402009-12-16 12:19:58 +01001382
1383 if (!pfn_valid(pfn))
1384 return -ENXIO;
1385
1386 p = pfn_to_page(pfn);
1387 page = compound_head(p);
1388
1389 if (!PageHWPoison(p)) {
Naoya Horiguchia5f65102015-11-05 18:47:26 -08001390 unpoison_pr_info("MCE: Page was already unpoisoned %#lx\n",
1391 pfn, &unpoison_rs);
Wu Fengguang847ce402009-12-16 12:19:58 +01001392 return 0;
1393 }
1394
Naoya Horiguchi230ac712015-09-08 15:03:29 -07001395 if (page_count(page) > 1) {
Naoya Horiguchia5f65102015-11-05 18:47:26 -08001396 unpoison_pr_info("MCE: Someone grabs the hwpoison page %#lx\n",
1397 pfn, &unpoison_rs);
Naoya Horiguchi230ac712015-09-08 15:03:29 -07001398 return 0;
1399 }
1400
1401 if (page_mapped(page)) {
Naoya Horiguchia5f65102015-11-05 18:47:26 -08001402 unpoison_pr_info("MCE: Someone maps the hwpoison page %#lx\n",
1403 pfn, &unpoison_rs);
Naoya Horiguchi230ac712015-09-08 15:03:29 -07001404 return 0;
1405 }
1406
1407 if (page_mapping(page)) {
Naoya Horiguchia5f65102015-11-05 18:47:26 -08001408 unpoison_pr_info("MCE: the hwpoison page has non-NULL mapping %#lx\n",
1409 pfn, &unpoison_rs);
Naoya Horiguchi230ac712015-09-08 15:03:29 -07001410 return 0;
1411 }
1412
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001413 /*
1414 * unpoison_memory() can encounter thp only when the thp is being
1415 * worked by memory_failure() and the page lock is not held yet.
1416 * In such case, we yield to memory_failure() and make unpoison fail.
1417 */
Wanpeng Lie76d30e2013-09-30 13:45:22 -07001418 if (!PageHuge(page) && PageTransHuge(page)) {
Naoya Horiguchia5f65102015-11-05 18:47:26 -08001419 unpoison_pr_info("MCE: Memory failure is now running on %#lx\n",
1420 pfn, &unpoison_rs);
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001421 return 0;
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001422 }
1423
Wanpeng Lif9121152013-09-11 14:22:52 -07001424 nr_pages = 1 << compound_order(page);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001425
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001426 if (!get_hwpoison_page(p)) {
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001427 /*
1428 * Since HWPoisoned hugepage should have non-zero refcount,
1429 * race between memory failure and unpoison seems to happen.
1430 * In such case unpoison fails and memory failure runs
1431 * to the end.
1432 */
1433 if (PageHuge(page)) {
Naoya Horiguchia5f65102015-11-05 18:47:26 -08001434 unpoison_pr_info("MCE: Memory failure is now running on free hugepage %#lx\n",
1435 pfn, &unpoison_rs);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001436 return 0;
1437 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001438 if (TestClearPageHWPoison(p))
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001439 num_poisoned_pages_dec();
Naoya Horiguchia5f65102015-11-05 18:47:26 -08001440 unpoison_pr_info("MCE: Software-unpoisoned free page %#lx\n",
1441 pfn, &unpoison_rs);
Wu Fengguang847ce402009-12-16 12:19:58 +01001442 return 0;
1443 }
1444
Jens Axboe7eaceac2011-03-10 08:52:07 +01001445 lock_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001446 /*
1447 * This test is racy because PG_hwpoison is set outside of page lock.
1448 * That's acceptable because that won't trigger kernel panic. Instead,
1449 * the PG_hwpoison page will be caught and isolated on the entrance to
1450 * the free buddy page pool.
1451 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001452 if (TestClearPageHWPoison(page)) {
Naoya Horiguchia5f65102015-11-05 18:47:26 -08001453 unpoison_pr_info("MCE: Software-unpoisoned page %#lx\n",
1454 pfn, &unpoison_rs);
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001455 num_poisoned_pages_sub(nr_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001456 freeit = 1;
Naoya Horiguchi6a901812010-09-08 10:19:40 +09001457 if (PageHuge(page))
1458 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001459 }
1460 unlock_page(page);
1461
Wanpeng Li665d9da2015-09-08 15:03:21 -07001462 put_hwpoison_page(page);
Wanpeng Li3ba5eeb2013-09-11 14:23:01 -07001463 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
Wanpeng Li665d9da2015-09-08 15:03:21 -07001464 put_hwpoison_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001465
1466 return 0;
1467}
1468EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001469
1470static struct page *new_page(struct page *p, unsigned long private, int **x)
1471{
Andi Kleen12686d12009-12-16 12:20:01 +01001472 int nid = page_to_nid(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001473 if (PageHuge(p))
1474 return alloc_huge_page_node(page_hstate(compound_head(p)),
1475 nid);
1476 else
Vlastimil Babka96db8002015-09-08 15:03:50 -07001477 return __alloc_pages_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001478}
1479
1480/*
1481 * Safely get reference count of an arbitrary page.
1482 * Returns 0 for a free page, -EIO for a zero refcount page
1483 * that is not free, and 1 for any other page type.
1484 * For 1 the page is returned with increased page count, otherwise not.
1485 */
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001486static int __get_any_page(struct page *p, unsigned long pfn, int flags)
Andi Kleenfacb6012009-12-16 12:20:00 +01001487{
1488 int ret;
1489
1490 if (flags & MF_COUNT_INCREASED)
1491 return 1;
1492
1493 /*
Naoya Horiguchid950b952010-09-08 10:19:39 +09001494 * When the target page is a free hugepage, just remove it
1495 * from free hugepage list.
1496 */
Naoya Horiguchiead07f62015-06-24 16:56:48 -07001497 if (!get_hwpoison_page(p)) {
Naoya Horiguchid950b952010-09-08 10:19:39 +09001498 if (PageHuge(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001499 pr_info("%s: %#lx free huge page\n", __func__, pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001500 ret = 0;
Naoya Horiguchid950b952010-09-08 10:19:39 +09001501 } else if (is_free_buddy_page(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001502 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001503 ret = 0;
1504 } else {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001505 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1506 __func__, pfn, p->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001507 ret = -EIO;
1508 }
1509 } else {
1510 /* Not a free page */
1511 ret = 1;
1512 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001513 return ret;
1514}
1515
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001516static int get_any_page(struct page *page, unsigned long pfn, int flags)
1517{
1518 int ret = __get_any_page(page, pfn, flags);
1519
1520 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1521 /*
1522 * Try to free it.
1523 */
Wanpeng Li665d9da2015-09-08 15:03:21 -07001524 put_hwpoison_page(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001525 shake_page(page, 1);
1526
1527 /*
1528 * Did it turn free?
1529 */
1530 ret = __get_any_page(page, pfn, 0);
Naoya Horiguchid96b3392016-01-15 16:54:03 -08001531 if (ret == 1 && !PageLRU(page)) {
Wanpeng Li4f32be62015-08-14 15:34:56 -07001532 /* Drop page reference which is from __get_any_page() */
Wanpeng Li665d9da2015-09-08 15:03:21 -07001533 put_hwpoison_page(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001534 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1535 pfn, page->flags);
1536 return -EIO;
1537 }
1538 }
1539 return ret;
1540}
1541
Naoya Horiguchid950b952010-09-08 10:19:39 +09001542static int soft_offline_huge_page(struct page *page, int flags)
1543{
1544 int ret;
1545 unsigned long pfn = page_to_pfn(page);
1546 struct page *hpage = compound_head(page);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001547 LIST_HEAD(pagelist);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001548
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001549 /*
1550 * This double-check of PageHWPoison is to avoid the race with
1551 * memory_failure(). See also comment in __soft_offline_page().
1552 */
1553 lock_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001554 if (PageHWPoison(hpage)) {
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001555 unlock_page(hpage);
Wanpeng Li665d9da2015-09-08 15:03:21 -07001556 put_hwpoison_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001557 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001558 return -EBUSY;
Xishi Qiu0ebff322013-02-22 16:33:59 -08001559 }
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001560 unlock_page(hpage);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001561
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001562 ret = isolate_huge_page(hpage, &pagelist);
Wanpeng Li03613802015-08-14 15:34:59 -07001563 /*
1564 * get_any_page() and isolate_huge_page() takes a refcount each,
1565 * so need to drop one here.
1566 */
Wanpeng Li665d9da2015-09-08 15:03:21 -07001567 put_hwpoison_page(hpage);
Wanpeng Li03613802015-08-14 15:34:59 -07001568 if (!ret) {
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001569 pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn);
1570 return -EBUSY;
1571 }
1572
David Rientjes68711a72014-06-04 16:08:25 -07001573 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001574 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001575 if (ret) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001576 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1577 pfn, ret, page->flags);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001578 /*
1579 * We know that soft_offline_huge_page() tries to migrate
1580 * only one hugepage pointed to by hpage, so we need not
1581 * run through the pagelist here.
1582 */
1583 putback_active_hugepage(hpage);
1584 if (ret > 0)
1585 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001586 } else {
Jianguo Wua49ecbc2013-12-18 17:08:54 -08001587 /* overcommit hugetlb page will be freed to buddy */
1588 if (PageHuge(page)) {
1589 set_page_hwpoison_huge_page(hpage);
1590 dequeue_hwpoisoned_huge_page(hpage);
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001591 num_poisoned_pages_add(1 << compound_order(hpage));
Jianguo Wua49ecbc2013-12-18 17:08:54 -08001592 } else {
1593 SetPageHWPoison(page);
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001594 num_poisoned_pages_inc();
Jianguo Wua49ecbc2013-12-18 17:08:54 -08001595 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001596 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001597 return ret;
1598}
1599
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001600static int __soft_offline_page(struct page *page, int flags)
1601{
1602 int ret;
1603 unsigned long pfn = page_to_pfn(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001604
1605 /*
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001606 * Check PageHWPoison again inside page lock because PageHWPoison
1607 * is set by memory_failure() outside page lock. Note that
1608 * memory_failure() also double-checks PageHWPoison inside page lock,
1609 * so there's no race between soft_offline_page() and memory_failure().
Andi Kleenfacb6012009-12-16 12:20:00 +01001610 */
Xishi Qiu0ebff322013-02-22 16:33:59 -08001611 lock_page(page);
1612 wait_on_page_writeback(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001613 if (PageHWPoison(page)) {
1614 unlock_page(page);
Wanpeng Li665d9da2015-09-08 15:03:21 -07001615 put_hwpoison_page(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001616 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1617 return -EBUSY;
1618 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001619 /*
1620 * Try to invalidate first. This should work for
1621 * non dirty unmapped page cache pages.
1622 */
1623 ret = invalidate_inode_page(page);
1624 unlock_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001625 /*
Andi Kleenfacb6012009-12-16 12:20:00 +01001626 * RED-PEN would be better to keep it isolated here, but we
1627 * would need to fix isolation locking first.
1628 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001629 if (ret == 1) {
Wanpeng Li665d9da2015-09-08 15:03:21 -07001630 put_hwpoison_page(page);
Andi Kleenfb46e732010-09-27 23:31:30 +02001631 pr_info("soft_offline: %#lx: invalidated\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001632 SetPageHWPoison(page);
Naoya Horiguchi8e304562015-09-08 15:03:24 -07001633 num_poisoned_pages_inc();
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001634 return 0;
Andi Kleenfacb6012009-12-16 12:20:00 +01001635 }
1636
1637 /*
1638 * Simple invalidation didn't work.
1639 * Try to migrate to a new page instead. migrate.c
1640 * handles a large number of cases for us.
1641 */
1642 ret = isolate_lru_page(page);
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001643 /*
1644 * Drop page reference which is came from get_any_page()
1645 * successful isolate_lru_page() already took another one.
1646 */
Wanpeng Li665d9da2015-09-08 15:03:21 -07001647 put_hwpoison_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001648 if (!ret) {
1649 LIST_HEAD(pagelist);
Minchan Kim5db8a732011-06-15 15:08:48 -07001650 inc_zone_page_state(page, NR_ISOLATED_ANON +
Hugh Dickins9c620e22013-02-22 16:35:14 -08001651 page_is_file_cache(page));
Andi Kleenfacb6012009-12-16 12:20:00 +01001652 list_add(&page->lru, &pagelist);
David Rientjes68711a72014-06-04 16:08:25 -07001653 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001654 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Andi Kleenfacb6012009-12-16 12:20:00 +01001655 if (ret) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001656 if (!list_empty(&pagelist)) {
1657 list_del(&page->lru);
1658 dec_zone_page_state(page, NR_ISOLATED_ANON +
1659 page_is_file_cache(page));
1660 putback_lru_page(page);
1661 }
1662
Andi Kleenfb46e732010-09-27 23:31:30 +02001663 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001664 pfn, ret, page->flags);
1665 if (ret > 0)
1666 ret = -EIO;
1667 }
1668 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001669 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
Dean Nelsondd73e852011-10-31 17:09:04 -07001670 pfn, ret, page_count(page), page->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001671 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001672 return ret;
1673}
Wanpeng Li86e05772013-09-11 14:22:56 -07001674
Naoya Horiguchiacc14dc2016-01-15 16:57:43 -08001675static int soft_offline_in_use_page(struct page *page, int flags)
1676{
1677 int ret;
1678 struct page *hpage = compound_head(page);
1679
1680 if (!PageHuge(page) && PageTransHuge(hpage)) {
1681 lock_page(hpage);
Naoya Horiguchi98fd1ef2016-01-15 16:57:46 -08001682 if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) {
1683 unlock_page(hpage);
1684 if (!PageAnon(hpage))
1685 pr_info("soft offline: %#lx: non anonymous thp\n", page_to_pfn(page));
1686 else
1687 pr_info("soft offline: %#lx: thp split failed\n", page_to_pfn(page));
1688 put_hwpoison_page(hpage);
Naoya Horiguchiacc14dc2016-01-15 16:57:43 -08001689 return -EBUSY;
1690 }
Naoya Horiguchi98fd1ef2016-01-15 16:57:46 -08001691 unlock_page(hpage);
Naoya Horiguchiacc14dc2016-01-15 16:57:43 -08001692 get_hwpoison_page(page);
1693 put_hwpoison_page(hpage);
1694 }
1695
1696 if (PageHuge(page))
1697 ret = soft_offline_huge_page(page, flags);
1698 else
1699 ret = __soft_offline_page(page, flags);
1700
1701 return ret;
1702}
1703
1704static void soft_offline_free_page(struct page *page)
1705{
1706 if (PageHuge(page)) {
1707 struct page *hpage = compound_head(page);
1708
1709 set_page_hwpoison_huge_page(hpage);
1710 if (!dequeue_hwpoisoned_huge_page(hpage))
1711 num_poisoned_pages_add(1 << compound_order(hpage));
1712 } else {
1713 if (!TestSetPageHWPoison(page))
1714 num_poisoned_pages_inc();
1715 }
1716}
1717
Wanpeng Li86e05772013-09-11 14:22:56 -07001718/**
1719 * soft_offline_page - Soft offline a page.
1720 * @page: page to offline
1721 * @flags: flags. Same as memory_failure().
1722 *
1723 * Returns 0 on success, otherwise negated errno.
1724 *
1725 * Soft offline a page, by migration or invalidation,
1726 * without killing anything. This is for the case when
1727 * a page is not corrupted yet (so it's still valid to access),
1728 * but has had a number of corrected errors and is better taken
1729 * out.
1730 *
1731 * The actual policy on when to do that is maintained by
1732 * user space.
1733 *
1734 * This should never impact any application or cause data loss,
1735 * however it might take some time.
1736 *
1737 * This is not a 100% solution for all memory, but tries to be
1738 * ``good enough'' for the majority of memory.
1739 */
1740int soft_offline_page(struct page *page, int flags)
1741{
1742 int ret;
1743 unsigned long pfn = page_to_pfn(page);
Wanpeng Li86e05772013-09-11 14:22:56 -07001744
1745 if (PageHWPoison(page)) {
1746 pr_info("soft offline: %#lx page already poisoned\n", pfn);
Wanpeng Li1e0e6352015-09-08 15:03:13 -07001747 if (flags & MF_COUNT_INCREASED)
Wanpeng Li665d9da2015-09-08 15:03:21 -07001748 put_hwpoison_page(page);
Wanpeng Li86e05772013-09-11 14:22:56 -07001749 return -EBUSY;
1750 }
Wanpeng Li86e05772013-09-11 14:22:56 -07001751
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07001752 get_online_mems();
Wanpeng Li86e05772013-09-11 14:22:56 -07001753 ret = get_any_page(page, pfn, flags);
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07001754 put_online_mems();
Naoya Horiguchi4e41a302016-01-15 16:54:07 -08001755
Naoya Horiguchiacc14dc2016-01-15 16:57:43 -08001756 if (ret > 0)
1757 ret = soft_offline_in_use_page(page, flags);
1758 else if (ret == 0)
1759 soft_offline_free_page(page);
Naoya Horiguchi4e41a302016-01-15 16:54:07 -08001760
Wanpeng Li86e05772013-09-11 14:22:56 -07001761 return ret;
1762}