blob: cd8989c1027ea6e2446af6ff148b5a947c8cbc0f [file] [log] [blame]
Andi Kleen6a460792009-09-16 11:50:15 +02001/*
2 * Copyright (C) 2008, 2009 Intel Corporation
3 * Authors: Andi Kleen, Fengguang Wu
4 *
5 * This software may be redistributed and/or modified under the terms of
6 * the GNU General Public License ("GPL") version 2 only as published by the
7 * Free Software Foundation.
8 *
9 * High level machine check handler. Handles pages reported by the
Andi Kleen1c80b992010-09-27 23:09:51 +020010 * hardware as being corrupted usually due to a multi-bit ECC memory or cache
Andi Kleen6a460792009-09-16 11:50:15 +020011 * failure.
Andi Kleen1c80b992010-09-27 23:09:51 +020012 *
13 * In addition there is a "soft offline" entry point that allows stop using
14 * not-yet-corrupted-by-suspicious pages without killing anything.
Andi Kleen6a460792009-09-16 11:50:15 +020015 *
16 * Handles page cache pages in various states. The tricky part
Andi Kleen1c80b992010-09-27 23:09:51 +020017 * here is that we can access any page asynchronously in respect to
18 * other VM users, because memory failures could happen anytime and
19 * anywhere. This could violate some of their assumptions. This is why
20 * this code has to be extremely careful. Generally it tries to use
21 * normal locking rules, as in get the standard locks, even if that means
22 * the error handling takes potentially a long time.
23 *
24 * There are several operations here with exponential complexity because
25 * of unsuitable VM data structures. For example the operation to map back
26 * from RMAP chains to processes has to walk the complete process list and
27 * has non linear complexity with the number. But since memory corruptions
28 * are rare we hope to get away with this. This avoids impacting the core
29 * VM.
Andi Kleen6a460792009-09-16 11:50:15 +020030 */
31
32/*
33 * Notebook:
34 * - hugetlb needs more code
35 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
36 * - pass bad pages to kdump next kernel
37 */
Andi Kleen6a460792009-09-16 11:50:15 +020038#include <linux/kernel.h>
39#include <linux/mm.h>
40#include <linux/page-flags.h>
Wu Fengguang478c5ff2009-12-16 12:19:59 +010041#include <linux/kernel-page-flags.h>
Andi Kleen6a460792009-09-16 11:50:15 +020042#include <linux/sched.h>
Hugh Dickins01e00f82009-10-13 15:02:11 +010043#include <linux/ksm.h>
Andi Kleen6a460792009-09-16 11:50:15 +020044#include <linux/rmap.h>
Paul Gortmakerb9e15ba2011-05-26 16:00:52 -040045#include <linux/export.h>
Andi Kleen6a460792009-09-16 11:50:15 +020046#include <linux/pagemap.h>
47#include <linux/swap.h>
48#include <linux/backing-dev.h>
Andi Kleenfacb6012009-12-16 12:20:00 +010049#include <linux/migrate.h>
50#include <linux/page-isolation.h>
51#include <linux/suspend.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090052#include <linux/slab.h>
Huang Yingbf998152010-05-31 14:28:19 +080053#include <linux/swapops.h>
Naoya Horiguchi7af446a2010-05-28 09:29:17 +090054#include <linux/hugetlb.h>
KOSAKI Motohiro20d6c962010-12-02 14:31:19 -080055#include <linux/memory_hotplug.h>
Minchan Kim5db8a732011-06-15 15:08:48 -070056#include <linux/mm_inline.h>
Huang Yingea8f5fb2011-07-13 13:14:27 +080057#include <linux/kfifo.h>
Andi Kleen6a460792009-09-16 11:50:15 +020058#include "internal.h"
59
60int sysctl_memory_failure_early_kill __read_mostly = 0;
61
62int sysctl_memory_failure_recovery __read_mostly = 1;
63
Xishi Qiu293c07e2013-02-22 16:34:02 -080064atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
Andi Kleen6a460792009-09-16 11:50:15 +020065
Andi Kleen27df5062009-12-21 19:56:42 +010066#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
67
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010068u32 hwpoison_filter_enable = 0;
Wu Fengguang7c116f22009-12-16 12:19:59 +010069u32 hwpoison_filter_dev_major = ~0U;
70u32 hwpoison_filter_dev_minor = ~0U;
Wu Fengguang478c5ff2009-12-16 12:19:59 +010071u64 hwpoison_filter_flags_mask;
72u64 hwpoison_filter_flags_value;
Haicheng Li1bfe5fe2009-12-16 12:19:59 +010073EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
Wu Fengguang7c116f22009-12-16 12:19:59 +010074EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
75EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
Wu Fengguang478c5ff2009-12-16 12:19:59 +010076EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
77EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
Wu Fengguang7c116f22009-12-16 12:19:59 +010078
79static int hwpoison_filter_dev(struct page *p)
80{
81 struct address_space *mapping;
82 dev_t dev;
83
84 if (hwpoison_filter_dev_major == ~0U &&
85 hwpoison_filter_dev_minor == ~0U)
86 return 0;
87
88 /*
Andi Kleen1c80b992010-09-27 23:09:51 +020089 * page_mapping() does not accept slab pages.
Wu Fengguang7c116f22009-12-16 12:19:59 +010090 */
91 if (PageSlab(p))
92 return -EINVAL;
93
94 mapping = page_mapping(p);
95 if (mapping == NULL || mapping->host == NULL)
96 return -EINVAL;
97
98 dev = mapping->host->i_sb->s_dev;
99 if (hwpoison_filter_dev_major != ~0U &&
100 hwpoison_filter_dev_major != MAJOR(dev))
101 return -EINVAL;
102 if (hwpoison_filter_dev_minor != ~0U &&
103 hwpoison_filter_dev_minor != MINOR(dev))
104 return -EINVAL;
105
106 return 0;
107}
108
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100109static int hwpoison_filter_flags(struct page *p)
110{
111 if (!hwpoison_filter_flags_mask)
112 return 0;
113
114 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
115 hwpoison_filter_flags_value)
116 return 0;
117 else
118 return -EINVAL;
119}
120
Andi Kleen4fd466e2009-12-16 12:19:59 +0100121/*
122 * This allows stress tests to limit test scope to a collection of tasks
123 * by putting them under some memcg. This prevents killing unrelated/important
124 * processes such as /sbin/init. Note that the target task may share clean
125 * pages with init (eg. libc text), which is harmless. If the target task
126 * share _dirty_ pages with another task B, the test scheme must make sure B
127 * is also included in the memcg. At last, due to race conditions this filter
128 * can only guarantee that the page either belongs to the memcg tasks, or is
129 * a freed page.
130 */
Andrew Mortonc255a452012-07-31 16:43:02 -0700131#ifdef CONFIG_MEMCG_SWAP
Andi Kleen4fd466e2009-12-16 12:19:59 +0100132u64 hwpoison_filter_memcg;
133EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
134static int hwpoison_filter_task(struct page *p)
135{
136 struct mem_cgroup *mem;
137 struct cgroup_subsys_state *css;
138 unsigned long ino;
139
140 if (!hwpoison_filter_memcg)
141 return 0;
142
143 mem = try_get_mem_cgroup_from_page(p);
144 if (!mem)
145 return -EINVAL;
146
147 css = mem_cgroup_css(mem);
Tejun Heob1664922014-02-11 11:52:49 -0500148 ino = cgroup_ino(css->cgroup);
Andi Kleen4fd466e2009-12-16 12:19:59 +0100149 css_put(css);
150
Tejun Heob1664922014-02-11 11:52:49 -0500151 if (!ino || ino != hwpoison_filter_memcg)
Andi Kleen4fd466e2009-12-16 12:19:59 +0100152 return -EINVAL;
153
154 return 0;
155}
156#else
157static int hwpoison_filter_task(struct page *p) { return 0; }
158#endif
159
Wu Fengguang7c116f22009-12-16 12:19:59 +0100160int hwpoison_filter(struct page *p)
161{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100162 if (!hwpoison_filter_enable)
163 return 0;
164
Wu Fengguang7c116f22009-12-16 12:19:59 +0100165 if (hwpoison_filter_dev(p))
166 return -EINVAL;
167
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100168 if (hwpoison_filter_flags(p))
169 return -EINVAL;
170
Andi Kleen4fd466e2009-12-16 12:19:59 +0100171 if (hwpoison_filter_task(p))
172 return -EINVAL;
173
Wu Fengguang7c116f22009-12-16 12:19:59 +0100174 return 0;
175}
Andi Kleen27df5062009-12-21 19:56:42 +0100176#else
177int hwpoison_filter(struct page *p)
178{
179 return 0;
180}
181#endif
182
Wu Fengguang7c116f22009-12-16 12:19:59 +0100183EXPORT_SYMBOL_GPL(hwpoison_filter);
184
Andi Kleen6a460792009-09-16 11:50:15 +0200185/*
Tony Luck7329bbe2011-12-13 09:27:58 -0800186 * Send all the processes who have the page mapped a signal.
187 * ``action optional'' if they are not immediately affected by the error
188 * ``action required'' if error happened in current execution context
Andi Kleen6a460792009-09-16 11:50:15 +0200189 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800190static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
191 unsigned long pfn, struct page *page, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200192{
193 struct siginfo si;
194 int ret;
195
196 printk(KERN_ERR
Tony Luck7329bbe2011-12-13 09:27:58 -0800197 "MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200198 pfn, t->comm, t->pid);
199 si.si_signo = SIGBUS;
200 si.si_errno = 0;
Andi Kleen6a460792009-09-16 11:50:15 +0200201 si.si_addr = (void *)addr;
202#ifdef __ARCH_SI_TRAPNO
203 si.si_trapno = trapno;
204#endif
Wanpeng Lif9121152013-09-11 14:22:52 -0700205 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
Tony Luck7329bbe2011-12-13 09:27:58 -0800206
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;
236 drain_all_pages();
237 if (PageLRU(p) || is_free_buddy_page(p))
238 return;
239 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100240
Andi Kleen588f9ce2009-12-16 12:19:57 +0100241 /*
Jin Dongmingaf241a02011-02-01 15:52:41 -0800242 * Only call shrink_slab here (which would also shrink other caches) if
243 * access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100244 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100245 if (access) {
246 int nr;
Dave Chinner0ce3d742013-08-28 10:18:03 +1000247 int nid = page_to_nid(p);
Andi Kleenfacb6012009-12-16 12:20:00 +0100248 do {
Ying Hana09ed5e2011-05-24 17:12:26 -0700249 struct shrink_control shrink = {
250 .gfp_mask = GFP_KERNEL,
Ying Hana09ed5e2011-05-24 17:12:26 -0700251 };
Dave Chinner0ce3d742013-08-28 10:18:03 +1000252 node_set(nid, shrink.nodes_to_scan);
Ying Hana09ed5e2011-05-24 17:12:26 -0700253
Ying Han1495f232011-05-24 17:12:27 -0700254 nr = shrink_slab(&shrink, 1000, 1000);
Andi Kleen47f43e72010-09-28 07:37:55 +0200255 if (page_count(p) == 1)
Andi Kleenfacb6012009-12-16 12:20:00 +0100256 break;
257 } while (nr > 10);
258 }
Andi Kleen588f9ce2009-12-16 12:19:57 +0100259}
260EXPORT_SYMBOL_GPL(shake_page);
261
262/*
Andi Kleen6a460792009-09-16 11:50:15 +0200263 * Kill all processes that have a poisoned page mapped and then isolate
264 * the page.
265 *
266 * General strategy:
267 * Find all processes having the page mapped and kill them.
268 * But we keep a page reference around so that the page is not
269 * actually freed yet.
270 * Then stash the page away
271 *
272 * There's no convenient way to get back to mapped processes
273 * from the VMAs. So do a brute-force search over all
274 * running processes.
275 *
276 * Remember that machine checks are not common (or rather
277 * if they are common you have other problems), so this shouldn't
278 * be a performance issue.
279 *
280 * Also there are some races possible while we get from the
281 * error detection to actually handle it.
282 */
283
284struct to_kill {
285 struct list_head nd;
286 struct task_struct *tsk;
287 unsigned long addr;
Andi Kleen9033ae12010-09-27 23:36:05 +0200288 char addr_valid;
Andi Kleen6a460792009-09-16 11:50:15 +0200289};
290
291/*
292 * Failure handling: if we can't find or can't kill a process there's
293 * not much we can do. We just print a message and ignore otherwise.
294 */
295
296/*
297 * Schedule a process for later kill.
298 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
299 * TBD would GFP_NOIO be enough?
300 */
301static void add_to_kill(struct task_struct *tsk, struct page *p,
302 struct vm_area_struct *vma,
303 struct list_head *to_kill,
304 struct to_kill **tkc)
305{
306 struct to_kill *tk;
307
308 if (*tkc) {
309 tk = *tkc;
310 *tkc = NULL;
311 } else {
312 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
313 if (!tk) {
314 printk(KERN_ERR
315 "MCE: Out of memory while machine check handling\n");
316 return;
317 }
318 }
319 tk->addr = page_address_in_vma(p, vma);
320 tk->addr_valid = 1;
321
322 /*
323 * In theory we don't have to kill when the page was
324 * munmaped. But it could be also a mremap. Since that's
325 * likely very rare kill anyways just out of paranoia, but use
326 * a SIGKILL because the error is not contained anymore.
327 */
328 if (tk->addr == -EFAULT) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200329 pr_info("MCE: Unable to find user space address %lx in %s\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200330 page_to_pfn(p), tsk->comm);
331 tk->addr_valid = 0;
332 }
333 get_task_struct(tsk);
334 tk->tsk = tsk;
335 list_add_tail(&tk->nd, to_kill);
336}
337
338/*
339 * Kill the processes that have been collected earlier.
340 *
341 * Only do anything when DOIT is set, otherwise just free the list
342 * (this is used for clean pages which do not need killing)
343 * Also when FAIL is set do a force kill because something went
344 * wrong earlier.
345 */
Tony Luck6751ed62012-07-11 10:20:47 -0700346static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -0800347 int fail, struct page *page, unsigned long pfn,
348 int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200349{
350 struct to_kill *tk, *next;
351
352 list_for_each_entry_safe (tk, next, to_kill, nd) {
Tony Luck6751ed62012-07-11 10:20:47 -0700353 if (forcekill) {
Andi Kleen6a460792009-09-16 11:50:15 +0200354 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200355 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200356 * make sure the process doesn't catch the
357 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200358 */
359 if (fail || tk->addr_valid == 0) {
360 printk(KERN_ERR
361 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
362 pfn, tk->tsk->comm, tk->tsk->pid);
363 force_sig(SIGKILL, tk->tsk);
364 }
365
366 /*
367 * In theory the process could have mapped
368 * something else on the address in-between. We could
369 * check for that, but we need to tell the
370 * process anyways.
371 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800372 else if (kill_proc(tk->tsk, tk->addr, trapno,
373 pfn, page, flags) < 0)
Andi Kleen6a460792009-09-16 11:50:15 +0200374 printk(KERN_ERR
375 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
376 pfn, tk->tsk->comm, tk->tsk->pid);
377 }
378 put_task_struct(tk->tsk);
379 kfree(tk);
380 }
381}
382
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700383/*
384 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
385 * on behalf of the thread group. Return task_struct of the (first found)
386 * dedicated thread if found, and return NULL otherwise.
387 *
388 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
389 * have to call rcu_read_lock/unlock() in this function.
390 */
391static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
Andi Kleen6a460792009-09-16 11:50:15 +0200392{
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700393 struct task_struct *t;
394
395 for_each_thread(tsk, t)
396 if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
397 return t;
398 return NULL;
399}
400
401/*
402 * Determine whether a given process is "early kill" process which expects
403 * to be signaled when some page under the process is hwpoisoned.
404 * Return task_struct of the dedicated thread (main thread unless explicitly
405 * specified) if the process is "early kill," and otherwise returns NULL.
406 */
407static struct task_struct *task_early_kill(struct task_struct *tsk,
408 int force_early)
409{
410 struct task_struct *t;
Andi Kleen6a460792009-09-16 11:50:15 +0200411 if (!tsk->mm)
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700412 return NULL;
Tony Luck74614de2014-06-04 16:11:01 -0700413 if (force_early)
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700414 return tsk;
415 t = find_early_kill_thread(tsk);
416 if (t)
417 return t;
418 if (sysctl_memory_failure_early_kill)
419 return tsk;
420 return NULL;
Andi Kleen6a460792009-09-16 11:50:15 +0200421}
422
423/*
424 * Collect processes when the error hit an anonymous page.
425 */
426static void collect_procs_anon(struct page *page, struct list_head *to_kill,
Tony Luck74614de2014-06-04 16:11:01 -0700427 struct to_kill **tkc, int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200428{
429 struct vm_area_struct *vma;
430 struct task_struct *tsk;
431 struct anon_vma *av;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700432 pgoff_t pgoff;
Andi Kleen6a460792009-09-16 11:50:15 +0200433
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000434 av = page_lock_anon_vma_read(page);
Andi Kleen6a460792009-09-16 11:50:15 +0200435 if (av == NULL) /* Not actually mapped anymore */
Peter Zijlstra9b679322011-06-27 16:18:09 -0700436 return;
437
Michel Lespinassebf181b92012-10-08 16:31:39 -0700438 pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700439 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200440 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800441 struct anon_vma_chain *vmac;
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700442 struct task_struct *t = task_early_kill(tsk, force_early);
Rik van Riel5beb4932010-03-05 13:42:07 -0800443
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700444 if (!t)
Andi Kleen6a460792009-09-16 11:50:15 +0200445 continue;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700446 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
447 pgoff, pgoff) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800448 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200449 if (!page_mapped_in_vma(page, vma))
450 continue;
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700451 if (vma->vm_mm == t->mm)
452 add_to_kill(t, page, vma, to_kill, tkc);
Andi Kleen6a460792009-09-16 11:50:15 +0200453 }
454 }
Andi Kleen6a460792009-09-16 11:50:15 +0200455 read_unlock(&tasklist_lock);
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000456 page_unlock_anon_vma_read(av);
Andi Kleen6a460792009-09-16 11:50:15 +0200457}
458
459/*
460 * Collect processes when the error hit a file mapped page.
461 */
462static void collect_procs_file(struct page *page, struct list_head *to_kill,
Tony Luck74614de2014-06-04 16:11:01 -0700463 struct to_kill **tkc, int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200464{
465 struct vm_area_struct *vma;
466 struct task_struct *tsk;
Andi Kleen6a460792009-09-16 11:50:15 +0200467 struct address_space *mapping = page->mapping;
468
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700469 mutex_lock(&mapping->i_mmap_mutex);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700470 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200471 for_each_process(tsk) {
472 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700473 struct task_struct *t = task_early_kill(tsk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200474
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700475 if (!t)
Andi Kleen6a460792009-09-16 11:50:15 +0200476 continue;
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -0700477 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
Andi Kleen6a460792009-09-16 11:50:15 +0200478 pgoff) {
479 /*
480 * Send early kill signal to tasks where a vma covers
481 * the page but the corrupted page is not necessarily
482 * mapped it in its pte.
483 * Assume applications who requested early kill want
484 * to be informed of all such data corruptions.
485 */
Naoya Horiguchi3ba08122014-06-04 16:11:02 -0700486 if (vma->vm_mm == t->mm)
487 add_to_kill(t, page, vma, to_kill, tkc);
Andi Kleen6a460792009-09-16 11:50:15 +0200488 }
489 }
Andi Kleen6a460792009-09-16 11:50:15 +0200490 read_unlock(&tasklist_lock);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700491 mutex_unlock(&mapping->i_mmap_mutex);
Andi Kleen6a460792009-09-16 11:50:15 +0200492}
493
494/*
495 * Collect the processes who have the corrupted page mapped to kill.
496 * This is done in two steps for locking reasons.
497 * First preallocate one tokill structure outside the spin locks,
498 * so that we can kill at least one process reasonably reliable.
499 */
Tony Luck74614de2014-06-04 16:11:01 -0700500static void collect_procs(struct page *page, struct list_head *tokill,
501 int force_early)
Andi Kleen6a460792009-09-16 11:50:15 +0200502{
503 struct to_kill *tk;
504
505 if (!page->mapping)
506 return;
507
508 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
509 if (!tk)
510 return;
511 if (PageAnon(page))
Tony Luck74614de2014-06-04 16:11:01 -0700512 collect_procs_anon(page, tokill, &tk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200513 else
Tony Luck74614de2014-06-04 16:11:01 -0700514 collect_procs_file(page, tokill, &tk, force_early);
Andi Kleen6a460792009-09-16 11:50:15 +0200515 kfree(tk);
516}
517
518/*
519 * Error handlers for various types of pages.
520 */
521
522enum outcome {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100523 IGNORED, /* Error: cannot be handled */
524 FAILED, /* Error: handling failed */
Andi Kleen6a460792009-09-16 11:50:15 +0200525 DELAYED, /* Will be handled later */
Andi Kleen6a460792009-09-16 11:50:15 +0200526 RECOVERED, /* Successfully recovered */
527};
528
529static const char *action_name[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100530 [IGNORED] = "Ignored",
Andi Kleen6a460792009-09-16 11:50:15 +0200531 [FAILED] = "Failed",
532 [DELAYED] = "Delayed",
Andi Kleen6a460792009-09-16 11:50:15 +0200533 [RECOVERED] = "Recovered",
534};
535
536/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100537 * XXX: It is possible that a page is isolated from LRU cache,
538 * and then kept in swap cache or failed to remove from page cache.
539 * The page count will stop it from being freed by unpoison.
540 * Stress tests should be aware of this memory leak problem.
541 */
542static int delete_from_lru_cache(struct page *p)
543{
544 if (!isolate_lru_page(p)) {
545 /*
546 * Clear sensible page flags, so that the buddy system won't
547 * complain when the page is unpoison-and-freed.
548 */
549 ClearPageActive(p);
550 ClearPageUnevictable(p);
551 /*
552 * drop the page count elevated by isolate_lru_page()
553 */
554 page_cache_release(p);
555 return 0;
556 }
557 return -EIO;
558}
559
560/*
Andi Kleen6a460792009-09-16 11:50:15 +0200561 * Error hit kernel page.
562 * Do nothing, try to be lucky and not touch this instead. For a few cases we
563 * could be more sophisticated.
564 */
565static int me_kernel(struct page *p, unsigned long pfn)
566{
Andi Kleen6a460792009-09-16 11:50:15 +0200567 return IGNORED;
568}
569
570/*
571 * Page in unknown state. Do nothing.
572 */
573static int me_unknown(struct page *p, unsigned long pfn)
574{
575 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
576 return FAILED;
577}
578
579/*
Andi Kleen6a460792009-09-16 11:50:15 +0200580 * Clean (or cleaned) page cache page.
581 */
582static int me_pagecache_clean(struct page *p, unsigned long pfn)
583{
584 int err;
585 int ret = FAILED;
586 struct address_space *mapping;
587
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100588 delete_from_lru_cache(p);
589
Andi Kleen6a460792009-09-16 11:50:15 +0200590 /*
591 * For anonymous pages we're done the only reference left
592 * should be the one m_f() holds.
593 */
594 if (PageAnon(p))
595 return RECOVERED;
596
597 /*
598 * Now truncate the page in the page cache. This is really
599 * more like a "temporary hole punch"
600 * Don't do this for block devices when someone else
601 * has a reference, because it could be file system metadata
602 * and that's not safe to truncate.
603 */
604 mapping = page_mapping(p);
605 if (!mapping) {
606 /*
607 * Page has been teared down in the meanwhile
608 */
609 return FAILED;
610 }
611
612 /*
613 * Truncation is a bit tricky. Enable it per file system for now.
614 *
615 * Open: to take i_mutex or not for this? Right now we don't.
616 */
617 if (mapping->a_ops->error_remove_page) {
618 err = mapping->a_ops->error_remove_page(mapping, p);
619 if (err != 0) {
620 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
621 pfn, err);
622 } else if (page_has_private(p) &&
623 !try_to_release_page(p, GFP_NOIO)) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200624 pr_info("MCE %#lx: failed to release buffers\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200625 } else {
626 ret = RECOVERED;
627 }
628 } else {
629 /*
630 * If the file system doesn't support it just invalidate
631 * This fails on dirty or anything with private pages
632 */
633 if (invalidate_inode_page(p))
634 ret = RECOVERED;
635 else
636 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
637 pfn);
638 }
639 return ret;
640}
641
642/*
Zhi Yong Wu549543d2014-01-21 15:49:08 -0800643 * Dirty pagecache page
Andi Kleen6a460792009-09-16 11:50:15 +0200644 * Issues: when the error hit a hole page the error is not properly
645 * propagated.
646 */
647static int me_pagecache_dirty(struct page *p, unsigned long pfn)
648{
649 struct address_space *mapping = page_mapping(p);
650
651 SetPageError(p);
652 /* TBD: print more information about the file. */
653 if (mapping) {
654 /*
655 * IO error will be reported by write(), fsync(), etc.
656 * who check the mapping.
657 * This way the application knows that something went
658 * wrong with its dirty file data.
659 *
660 * There's one open issue:
661 *
662 * The EIO will be only reported on the next IO
663 * operation and then cleared through the IO map.
664 * Normally Linux has two mechanisms to pass IO error
665 * first through the AS_EIO flag in the address space
666 * and then through the PageError flag in the page.
667 * Since we drop pages on memory failure handling the
668 * only mechanism open to use is through AS_AIO.
669 *
670 * This has the disadvantage that it gets cleared on
671 * the first operation that returns an error, while
672 * the PageError bit is more sticky and only cleared
673 * when the page is reread or dropped. If an
674 * application assumes it will always get error on
675 * fsync, but does other operations on the fd before
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300676 * and the page is dropped between then the error
Andi Kleen6a460792009-09-16 11:50:15 +0200677 * will not be properly reported.
678 *
679 * This can already happen even without hwpoisoned
680 * pages: first on metadata IO errors (which only
681 * report through AS_EIO) or when the page is dropped
682 * at the wrong time.
683 *
684 * So right now we assume that the application DTRT on
685 * the first EIO, but we're not worse than other parts
686 * of the kernel.
687 */
688 mapping_set_error(mapping, EIO);
689 }
690
691 return me_pagecache_clean(p, pfn);
692}
693
694/*
695 * Clean and dirty swap cache.
696 *
697 * Dirty swap cache page is tricky to handle. The page could live both in page
698 * cache and swap cache(ie. page is freshly swapped in). So it could be
699 * referenced concurrently by 2 types of PTEs:
700 * normal PTEs and swap PTEs. We try to handle them consistently by calling
701 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
702 * and then
703 * - clear dirty bit to prevent IO
704 * - remove from LRU
705 * - but keep in the swap cache, so that when we return to it on
706 * a later page fault, we know the application is accessing
707 * corrupted data and shall be killed (we installed simple
708 * interception code in do_swap_page to catch it).
709 *
710 * Clean swap cache pages can be directly isolated. A later page fault will
711 * bring in the known good data from disk.
712 */
713static int me_swapcache_dirty(struct page *p, unsigned long pfn)
714{
Andi Kleen6a460792009-09-16 11:50:15 +0200715 ClearPageDirty(p);
716 /* Trigger EIO in shmem: */
717 ClearPageUptodate(p);
718
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100719 if (!delete_from_lru_cache(p))
720 return DELAYED;
721 else
722 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200723}
724
725static int me_swapcache_clean(struct page *p, unsigned long pfn)
726{
Andi Kleen6a460792009-09-16 11:50:15 +0200727 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800728
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100729 if (!delete_from_lru_cache(p))
730 return RECOVERED;
731 else
732 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200733}
734
735/*
736 * Huge pages. Needs work.
737 * Issues:
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900738 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
739 * To narrow down kill region to one page, we need to break up pmd.
Andi Kleen6a460792009-09-16 11:50:15 +0200740 */
741static int me_huge_page(struct page *p, unsigned long pfn)
742{
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900743 int res = 0;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900744 struct page *hpage = compound_head(p);
745 /*
746 * We can safely recover from error on free or reserved (i.e.
747 * not in-use) hugepage by dequeuing it from freelist.
748 * To check whether a hugepage is in-use or not, we can't use
749 * page->lru because it can be used in other hugepage operations,
750 * such as __unmap_hugepage_range() and gather_surplus_pages().
751 * So instead we use page_mapping() and PageAnon().
752 * We assume that this function is called with page lock held,
753 * so there is no race between isolation and mapping/unmapping.
754 */
755 if (!(page_mapping(hpage) || PageAnon(hpage))) {
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900756 res = dequeue_hwpoisoned_huge_page(hpage);
757 if (!res)
758 return RECOVERED;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900759 }
760 return DELAYED;
Andi Kleen6a460792009-09-16 11:50:15 +0200761}
762
763/*
764 * Various page states we can handle.
765 *
766 * A page state is defined by its current page->flags bits.
767 * The table matches them in order and calls the right handler.
768 *
769 * This is quite tricky because we can access page at any time
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300770 * in its live cycle, so all accesses have to be extremely careful.
Andi Kleen6a460792009-09-16 11:50:15 +0200771 *
772 * This is not complete. More states could be added.
773 * For any missing state don't attempt recovery.
774 */
775
776#define dirty (1UL << PG_dirty)
777#define sc (1UL << PG_swapcache)
778#define unevict (1UL << PG_unevictable)
779#define mlock (1UL << PG_mlocked)
780#define writeback (1UL << PG_writeback)
781#define lru (1UL << PG_lru)
782#define swapbacked (1UL << PG_swapbacked)
783#define head (1UL << PG_head)
784#define tail (1UL << PG_tail)
785#define compound (1UL << PG_compound)
786#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200787#define reserved (1UL << PG_reserved)
788
789static struct page_state {
790 unsigned long mask;
791 unsigned long res;
792 char *msg;
793 int (*action)(struct page *p, unsigned long pfn);
794} error_states[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100795 { reserved, reserved, "reserved kernel", me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100796 /*
797 * free pages are specially detected outside this table:
798 * PG_buddy pages only make a small fraction of all free pages.
799 */
Andi Kleen6a460792009-09-16 11:50:15 +0200800
801 /*
802 * Could in theory check if slab page is free or if we can drop
803 * currently unused objects without touching them. But just
804 * treat it as standard kernel for now.
805 */
806 { slab, slab, "kernel slab", me_kernel },
807
808#ifdef CONFIG_PAGEFLAGS_EXTENDED
809 { head, head, "huge", me_huge_page },
810 { tail, tail, "huge", me_huge_page },
811#else
812 { compound, compound, "huge", me_huge_page },
813#endif
814
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800815 { sc|dirty, sc|dirty, "dirty swapcache", me_swapcache_dirty },
816 { sc|dirty, sc, "clean swapcache", me_swapcache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200817
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800818 { mlock|dirty, mlock|dirty, "dirty mlocked LRU", me_pagecache_dirty },
Naoya Horiguchie3986292013-04-29 15:06:08 -0700819 { mlock|dirty, mlock, "clean mlocked LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200820
Naoya Horiguchi5f4b9fc2013-02-22 16:35:53 -0800821 { unevict|dirty, unevict|dirty, "dirty unevictable LRU", me_pagecache_dirty },
Naoya Horiguchie3986292013-04-29 15:06:08 -0700822 { unevict|dirty, unevict, "clean unevictable LRU", me_pagecache_clean },
Naoya Horiguchi5f4b9fc2013-02-22 16:35:53 -0800823
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800824 { lru|dirty, lru|dirty, "dirty LRU", me_pagecache_dirty },
Andi Kleen6a460792009-09-16 11:50:15 +0200825 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200826
827 /*
828 * Catchall entry: must be at end.
829 */
830 { 0, 0, "unknown page state", me_unknown },
831};
832
Andi Kleen2326c462009-12-16 12:20:00 +0100833#undef dirty
834#undef sc
835#undef unevict
836#undef mlock
837#undef writeback
838#undef lru
839#undef swapbacked
840#undef head
841#undef tail
842#undef compound
843#undef slab
844#undef reserved
845
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800846/*
847 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
848 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
849 */
Andi Kleen6a460792009-09-16 11:50:15 +0200850static void action_result(unsigned long pfn, char *msg, int result)
851{
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800852 pr_err("MCE %#lx: %s page recovery: %s\n",
853 pfn, msg, action_name[result]);
Andi Kleen6a460792009-09-16 11:50:15 +0200854}
855
856static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100857 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200858{
859 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200860 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200861
862 result = ps->action(p, pfn);
863 action_result(pfn, ps->msg, result);
Wu Fengguang7456b042009-10-19 08:15:01 +0200864
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100865 count = page_count(p) - 1;
Wu Fengguang138ce282009-12-16 12:19:58 +0100866 if (ps->action == me_swapcache_dirty && result == DELAYED)
867 count--;
868 if (count != 0) {
Andi Kleen6a460792009-09-16 11:50:15 +0200869 printk(KERN_ERR
870 "MCE %#lx: %s page still referenced by %d users\n",
Wu Fengguang7456b042009-10-19 08:15:01 +0200871 pfn, ps->msg, count);
Wu Fengguang138ce282009-12-16 12:19:58 +0100872 result = FAILED;
873 }
Andi Kleen6a460792009-09-16 11:50:15 +0200874
875 /* Could do more checks here if page looks ok */
876 /*
877 * Could adjust zone counters here to correct for the missing page.
878 */
879
Wu Fengguang138ce282009-12-16 12:19:58 +0100880 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200881}
882
Andi Kleen6a460792009-09-16 11:50:15 +0200883/*
884 * Do all that is necessary to remove user space mappings. Unmap
885 * the pages and send SIGBUS to the processes if the data was dirty.
886 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100887static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800888 int trapno, int flags, struct page **hpagep)
Andi Kleen6a460792009-09-16 11:50:15 +0200889{
890 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
891 struct address_space *mapping;
892 LIST_HEAD(tokill);
893 int ret;
Tony Luck6751ed62012-07-11 10:20:47 -0700894 int kill = 1, forcekill;
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800895 struct page *hpage = *hpagep;
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800896 struct page *ppage;
Andi Kleen6a460792009-09-16 11:50:15 +0200897
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100898 if (PageReserved(p) || PageSlab(p))
899 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200900
Andi Kleen6a460792009-09-16 11:50:15 +0200901 /*
902 * This check implies we don't kill processes if their pages
903 * are in the swap cache early. Those are always late kills.
904 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900905 if (!page_mapped(hpage))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100906 return SWAP_SUCCESS;
907
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900908 if (PageKsm(p))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100909 return SWAP_FAIL;
Andi Kleen6a460792009-09-16 11:50:15 +0200910
911 if (PageSwapCache(p)) {
912 printk(KERN_ERR
913 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
914 ttu |= TTU_IGNORE_HWPOISON;
915 }
916
917 /*
918 * Propagate the dirty bit from PTEs to struct page first, because we
919 * need this to decide if we should kill or just drop the page.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100920 * XXX: the dirty test could be racy: set_page_dirty() may not always
921 * be called inside page lock (it's recommended but not enforced).
Andi Kleen6a460792009-09-16 11:50:15 +0200922 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900923 mapping = page_mapping(hpage);
Tony Luck6751ed62012-07-11 10:20:47 -0700924 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900925 mapping_cap_writeback_dirty(mapping)) {
926 if (page_mkclean(hpage)) {
927 SetPageDirty(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +0200928 } else {
929 kill = 0;
930 ttu |= TTU_IGNORE_HWPOISON;
931 printk(KERN_INFO
932 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
933 pfn);
934 }
935 }
936
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800937 /*
938 * ppage: poisoned page
939 * if p is regular page(4k page)
940 * ppage == real poisoned page;
941 * else p is hugetlb or THP, ppage == head page.
942 */
943 ppage = hpage;
944
Jin Dongmingefeda7a2011-02-01 15:52:39 -0800945 if (PageTransHuge(hpage)) {
946 /*
947 * Verify that this isn't a hugetlbfs head page, the check for
948 * PageAnon is just for avoid tripping a split_huge_page
949 * internal debug check, as split_huge_page refuses to deal with
950 * anything that isn't an anon page. PageAnon can't go away fro
951 * under us because we hold a refcount on the hpage, without a
952 * refcount on the hpage. split_huge_page can't be safely called
953 * in the first place, having a refcount on the tail isn't
954 * enough * to be safe.
955 */
956 if (!PageHuge(hpage) && PageAnon(hpage)) {
957 if (unlikely(split_huge_page(hpage))) {
958 /*
959 * FIXME: if splitting THP is failed, it is
960 * better to stop the following operation rather
961 * than causing panic by unmapping. System might
962 * survive if the page is freed later.
963 */
964 printk(KERN_INFO
965 "MCE %#lx: failed to split THP\n", pfn);
966
967 BUG_ON(!PageHWPoison(p));
968 return SWAP_FAIL;
969 }
Naoya Horiguchia3e0f9e2014-01-02 12:58:51 -0800970 /*
971 * We pinned the head page for hwpoison handling,
972 * now we split the thp and we are interested in
973 * the hwpoisoned raw page, so move the refcount
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800974 * to it. Similarly, page lock is shifted.
Naoya Horiguchia3e0f9e2014-01-02 12:58:51 -0800975 */
976 if (hpage != p) {
Naoya Horiguchi8d547ff2014-02-10 14:25:50 -0800977 if (!(flags & MF_COUNT_INCREASED)) {
978 put_page(hpage);
979 get_page(p);
980 }
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -0800981 lock_page(p);
982 unlock_page(hpage);
983 *hpagep = p;
Naoya Horiguchia3e0f9e2014-01-02 12:58:51 -0800984 }
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800985 /* THP is split, so ppage should be the real poisoned page. */
986 ppage = p;
Jin Dongmingefeda7a2011-02-01 15:52:39 -0800987 }
988 }
989
Andi Kleen6a460792009-09-16 11:50:15 +0200990 /*
991 * First collect all the processes that have the page
992 * mapped in dirty form. This has to be done before try_to_unmap,
993 * because ttu takes the rmap data structures down.
994 *
995 * Error handling: We ignore errors here because
996 * there's nothing that can be done.
997 */
998 if (kill)
Tony Luck74614de2014-06-04 16:11:01 -0700999 collect_procs(ppage, &tokill, flags & MF_ACTION_REQUIRED);
Andi Kleen6a460792009-09-16 11:50:15 +02001000
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001001 ret = try_to_unmap(ppage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +02001002 if (ret != SWAP_SUCCESS)
1003 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001004 pfn, page_mapcount(ppage));
1005
Andi Kleen6a460792009-09-16 11:50:15 +02001006 /*
1007 * Now that the dirty bit has been propagated to the
1008 * struct page and all unmaps done we can decide if
1009 * killing is needed or not. Only kill when the page
Tony Luck6751ed62012-07-11 10:20:47 -07001010 * was dirty or the process is not restartable,
1011 * otherwise the tokill list is merely
Andi Kleen6a460792009-09-16 11:50:15 +02001012 * freed. When there was a problem unmapping earlier
1013 * use a more force-full uncatchable kill to prevent
1014 * any accesses to the poisoned memory.
1015 */
Tony Luck6751ed62012-07-11 10:20:47 -07001016 forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL);
1017 kill_procs(&tokill, forcekill, trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -08001018 ret != SWAP_SUCCESS, p, pfn, flags);
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001019
1020 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +02001021}
1022
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001023static void set_page_hwpoison_huge_page(struct page *hpage)
1024{
1025 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -07001026 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001027 for (i = 0; i < nr_pages; i++)
1028 SetPageHWPoison(hpage + i);
1029}
1030
1031static void clear_page_hwpoison_huge_page(struct page *hpage)
1032{
1033 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -07001034 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001035 for (i = 0; i < nr_pages; i++)
1036 ClearPageHWPoison(hpage + i);
1037}
1038
Tony Luckcd42f4a2011-12-15 10:48:12 -08001039/**
1040 * memory_failure - Handle memory failure of a page.
1041 * @pfn: Page Number of the corrupted page
1042 * @trapno: Trap number reported in the signal to user space.
1043 * @flags: fine tune action taken
1044 *
1045 * This function is called by the low level machine check code
1046 * of an architecture when it detects hardware memory corruption
1047 * of a page. It tries its best to recover, which includes
1048 * dropping pages, killing processes etc.
1049 *
1050 * The function is primarily of use for corruptions that
1051 * happen outside the current execution context (e.g. when
1052 * detected by a background scrubber)
1053 *
1054 * Must run in process context (e.g. a work queue) with interrupts
1055 * enabled and no spinlocks hold.
1056 */
1057int memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +02001058{
1059 struct page_state *ps;
1060 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001061 struct page *hpage;
Andi Kleen6a460792009-09-16 11:50:15 +02001062 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001063 unsigned int nr_pages;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001064 unsigned long page_flags;
Andi Kleen6a460792009-09-16 11:50:15 +02001065
1066 if (!sysctl_memory_failure_recovery)
1067 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1068
1069 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +01001070 printk(KERN_ERR
1071 "MCE %#lx: memory outside kernel control\n",
1072 pfn);
1073 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +02001074 }
1075
1076 p = pfn_to_page(pfn);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001077 hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +02001078 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001079 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001080 return 0;
1081 }
1082
Naoya Horiguchi4db0e952013-02-22 16:34:05 -08001083 /*
1084 * Currently errors on hugetlbfs pages are measured in hugepage units,
1085 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1086 * transparent hugepages, they are supposed to be split and error
1087 * measurement is done in normal page units. So nr_pages should be one
1088 * in this case.
1089 */
1090 if (PageHuge(p))
1091 nr_pages = 1 << compound_order(hpage);
1092 else /* normal page or thp */
1093 nr_pages = 1;
Xishi Qiu293c07e2013-02-22 16:34:02 -08001094 atomic_long_add(nr_pages, &num_poisoned_pages);
Andi Kleen6a460792009-09-16 11:50:15 +02001095
1096 /*
1097 * We need/can do nothing about count=0 pages.
1098 * 1) it's a free page, and therefore in safe hand:
1099 * prep_new_page() will be the gate keeper.
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001100 * 2) it's a free hugepage, which is also safe:
1101 * an affected hugepage will be dequeued from hugepage freelist,
1102 * so there's no concern about reusing it ever after.
1103 * 3) it's part of a non-compound high order page.
Andi Kleen6a460792009-09-16 11:50:15 +02001104 * Implies some kernel user: cannot stop them from
1105 * R/W the page; let's pray that the page has been
1106 * used and will be freed some time later.
1107 * In fact it's dangerous to directly bump up page count from 0,
1108 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1109 */
Andi Kleen82ba0112009-12-16 12:19:57 +01001110 if (!(flags & MF_COUNT_INCREASED) &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001111 !get_page_unless_zero(hpage)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001112 if (is_free_buddy_page(p)) {
1113 action_result(pfn, "free buddy", DELAYED);
1114 return 0;
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001115 } else if (PageHuge(hpage)) {
1116 /*
Chen Yucongb9851942014-05-22 11:54:15 -07001117 * Check "filter hit" and "race with other subpage."
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001118 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001119 lock_page(hpage);
Chen Yucongb9851942014-05-22 11:54:15 -07001120 if (PageHWPoison(hpage)) {
1121 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1122 || (p != hpage && TestSetPageHWPoison(hpage))) {
1123 atomic_long_sub(nr_pages, &num_poisoned_pages);
1124 unlock_page(hpage);
1125 return 0;
1126 }
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001127 }
1128 set_page_hwpoison_huge_page(hpage);
1129 res = dequeue_hwpoisoned_huge_page(hpage);
1130 action_result(pfn, "free huge",
1131 res ? IGNORED : DELAYED);
1132 unlock_page(hpage);
1133 return res;
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001134 } else {
1135 action_result(pfn, "high order kernel", IGNORED);
1136 return -EBUSY;
1137 }
Andi Kleen6a460792009-09-16 11:50:15 +02001138 }
1139
1140 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001141 * We ignore non-LRU pages for good reasons.
1142 * - PG_locked is only well defined for LRU pages and a few others
1143 * - to avoid races with __set_page_locked()
1144 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1145 * The check (unnecessarily) ignores LRU pages being isolated and
1146 * walked by the page reclaim code, however that's not a big loss.
1147 */
Dean Nelson385de352012-03-21 16:34:05 -07001148 if (!PageHuge(p) && !PageTransTail(p)) {
Jin Dongmingaf241a02011-02-01 15:52:41 -08001149 if (!PageLRU(p))
1150 shake_page(p, 0);
1151 if (!PageLRU(p)) {
1152 /*
1153 * shake_page could have turned it free.
1154 */
1155 if (is_free_buddy_page(p)) {
Wanpeng Li2d421ac2013-09-30 13:45:23 -07001156 if (flags & MF_COUNT_INCREASED)
1157 action_result(pfn, "free buddy", DELAYED);
1158 else
1159 action_result(pfn, "free buddy, 2nd try", DELAYED);
Jin Dongmingaf241a02011-02-01 15:52:41 -08001160 return 0;
1161 }
1162 action_result(pfn, "non LRU", IGNORED);
1163 put_page(p);
1164 return -EBUSY;
Andi Kleen0474a602009-12-16 12:20:00 +01001165 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001166 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001167
Jens Axboe7eaceac2011-03-10 08:52:07 +01001168 lock_page(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001169
1170 /*
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001171 * We use page flags to determine what action should be taken, but
1172 * the flags can be modified by the error containment action. One
1173 * example is an mlocked page, where PG_mlocked is cleared by
1174 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1175 * correctly, we save a copy of the page flags at this time.
1176 */
1177 page_flags = p->flags;
1178
1179 /*
Wu Fengguang847ce402009-12-16 12:19:58 +01001180 * unpoison always clear PG_hwpoison inside page lock
1181 */
1182 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001183 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Naoya Horiguchi3e030ec2014-05-22 11:54:21 -07001184 atomic_long_sub(nr_pages, &num_poisoned_pages);
1185 put_page(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001186 res = 0;
1187 goto out;
1188 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001189 if (hwpoison_filter(p)) {
1190 if (TestClearPageHWPoison(p))
Xishi Qiu293c07e2013-02-22 16:34:02 -08001191 atomic_long_sub(nr_pages, &num_poisoned_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001192 unlock_page(hpage);
1193 put_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001194 return 0;
1195 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001196
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001197 /*
1198 * For error on the tail page, we should set PG_hwpoison
1199 * on the head page to show that the hugepage is hwpoisoned
1200 */
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001201 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001202 action_result(pfn, "hugepage already hardware poisoned",
1203 IGNORED);
1204 unlock_page(hpage);
1205 put_page(hpage);
1206 return 0;
1207 }
1208 /*
1209 * Set PG_hwpoison on all pages in an error hugepage,
1210 * because containment is done in hugepage unit for now.
1211 * Since we have done TestSetPageHWPoison() for the head page with
1212 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1213 */
1214 if (PageHuge(p))
1215 set_page_hwpoison_huge_page(hpage);
1216
Naoya Horiguchi6edd6cc2014-06-04 16:10:35 -07001217 /*
1218 * It's very difficult to mess with pages currently under IO
1219 * and in many cases impossible, so we just avoid it here.
1220 */
Andi Kleen6a460792009-09-16 11:50:15 +02001221 wait_on_page_writeback(p);
1222
1223 /*
1224 * Now take care of user space mappings.
Minchan Kime64a7822011-03-22 16:32:44 -07001225 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001226 *
1227 * When the raw error page is thp tail page, hpage points to the raw
1228 * page after thp split.
Andi Kleen6a460792009-09-16 11:50:15 +02001229 */
Naoya Horiguchi54b9dd12014-01-23 15:53:14 -08001230 if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)
1231 != SWAP_SUCCESS) {
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001232 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1233 res = -EBUSY;
1234 goto out;
1235 }
Andi Kleen6a460792009-09-16 11:50:15 +02001236
1237 /*
1238 * Torn down by someone else?
1239 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001240 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Andi Kleen6a460792009-09-16 11:50:15 +02001241 action_result(pfn, "already truncated LRU", IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001242 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001243 goto out;
1244 }
1245
1246 res = -EBUSY;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001247 /*
1248 * The first check uses the current page flags which may not have any
1249 * relevant information. The second check with the saved page flagss is
1250 * carried out only if the first check can't determine the page status.
1251 */
1252 for (ps = error_states;; ps++)
1253 if ((p->flags & ps->mask) == ps->res)
Andi Kleen6a460792009-09-16 11:50:15 +02001254 break;
Wanpeng Li841fcc52013-09-11 14:22:50 -07001255
1256 page_flags |= (p->flags & (1UL << PG_dirty));
1257
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001258 if (!ps->mask)
1259 for (ps = error_states;; ps++)
1260 if ((page_flags & ps->mask) == ps->res)
1261 break;
1262 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001263out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001264 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001265 return res;
1266}
Tony Luckcd42f4a2011-12-15 10:48:12 -08001267EXPORT_SYMBOL_GPL(memory_failure);
Wu Fengguang847ce402009-12-16 12:19:58 +01001268
Huang Yingea8f5fb2011-07-13 13:14:27 +08001269#define MEMORY_FAILURE_FIFO_ORDER 4
1270#define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1271
1272struct memory_failure_entry {
1273 unsigned long pfn;
1274 int trapno;
1275 int flags;
1276};
1277
1278struct memory_failure_cpu {
1279 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1280 MEMORY_FAILURE_FIFO_SIZE);
1281 spinlock_t lock;
1282 struct work_struct work;
1283};
1284
1285static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1286
1287/**
1288 * memory_failure_queue - Schedule handling memory failure of a page.
1289 * @pfn: Page Number of the corrupted page
1290 * @trapno: Trap number reported in the signal to user space.
1291 * @flags: Flags for memory failure handling
1292 *
1293 * This function is called by the low level hardware error handler
1294 * when it detects hardware memory corruption of a page. It schedules
1295 * the recovering of error page, including dropping pages, killing
1296 * processes etc.
1297 *
1298 * The function is primarily of use for corruptions that
1299 * happen outside the current execution context (e.g. when
1300 * detected by a background scrubber)
1301 *
1302 * Can run in IRQ context.
1303 */
1304void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1305{
1306 struct memory_failure_cpu *mf_cpu;
1307 unsigned long proc_flags;
1308 struct memory_failure_entry entry = {
1309 .pfn = pfn,
1310 .trapno = trapno,
1311 .flags = flags,
1312 };
1313
1314 mf_cpu = &get_cpu_var(memory_failure_cpu);
1315 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
Stefani Seibold498d3192013-11-14 14:32:17 -08001316 if (kfifo_put(&mf_cpu->fifo, entry))
Huang Yingea8f5fb2011-07-13 13:14:27 +08001317 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1318 else
Joe Perches8e33a522013-07-25 11:53:25 -07001319 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
Huang Yingea8f5fb2011-07-13 13:14:27 +08001320 pfn);
1321 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1322 put_cpu_var(memory_failure_cpu);
1323}
1324EXPORT_SYMBOL_GPL(memory_failure_queue);
1325
1326static void memory_failure_work_func(struct work_struct *work)
1327{
1328 struct memory_failure_cpu *mf_cpu;
1329 struct memory_failure_entry entry = { 0, };
1330 unsigned long proc_flags;
1331 int gotten;
1332
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001333 mf_cpu = this_cpu_ptr(&memory_failure_cpu);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001334 for (;;) {
1335 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1336 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1337 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1338 if (!gotten)
1339 break;
Naveen N. Raocf870c72013-07-10 14:57:01 +05301340 if (entry.flags & MF_SOFT_OFFLINE)
1341 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1342 else
1343 memory_failure(entry.pfn, entry.trapno, entry.flags);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001344 }
1345}
1346
1347static int __init memory_failure_init(void)
1348{
1349 struct memory_failure_cpu *mf_cpu;
1350 int cpu;
1351
1352 for_each_possible_cpu(cpu) {
1353 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1354 spin_lock_init(&mf_cpu->lock);
1355 INIT_KFIFO(mf_cpu->fifo);
1356 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1357 }
1358
1359 return 0;
1360}
1361core_initcall(memory_failure_init);
1362
Wu Fengguang847ce402009-12-16 12:19:58 +01001363/**
1364 * unpoison_memory - Unpoison a previously poisoned page
1365 * @pfn: Page number of the to be unpoisoned page
1366 *
1367 * Software-unpoison a page that has been poisoned by
1368 * memory_failure() earlier.
1369 *
1370 * This is only done on the software-level, so it only works
1371 * for linux injected failures, not real hardware failures
1372 *
1373 * Returns 0 for success, otherwise -errno.
1374 */
1375int unpoison_memory(unsigned long pfn)
1376{
1377 struct page *page;
1378 struct page *p;
1379 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001380 unsigned int nr_pages;
Wu Fengguang847ce402009-12-16 12:19:58 +01001381
1382 if (!pfn_valid(pfn))
1383 return -ENXIO;
1384
1385 p = pfn_to_page(pfn);
1386 page = compound_head(p);
1387
1388 if (!PageHWPoison(p)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001389 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001390 return 0;
1391 }
1392
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001393 /*
1394 * unpoison_memory() can encounter thp only when the thp is being
1395 * worked by memory_failure() and the page lock is not held yet.
1396 * In such case, we yield to memory_failure() and make unpoison fail.
1397 */
Wanpeng Lie76d30e2013-09-30 13:45:22 -07001398 if (!PageHuge(page) && PageTransHuge(page)) {
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001399 pr_info("MCE: Memory failure is now running on %#lx\n", pfn);
1400 return 0;
1401 }
1402
Wanpeng Lif9121152013-09-11 14:22:52 -07001403 nr_pages = 1 << compound_order(page);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001404
Wu Fengguang847ce402009-12-16 12:19:58 +01001405 if (!get_page_unless_zero(page)) {
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001406 /*
1407 * Since HWPoisoned hugepage should have non-zero refcount,
1408 * race between memory failure and unpoison seems to happen.
1409 * In such case unpoison fails and memory failure runs
1410 * to the end.
1411 */
1412 if (PageHuge(page)) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001413 pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001414 return 0;
1415 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001416 if (TestClearPageHWPoison(p))
Wanpeng Lidd9538a2013-09-11 14:22:54 -07001417 atomic_long_dec(&num_poisoned_pages);
Andi Kleenfb46e732010-09-27 23:31:30 +02001418 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001419 return 0;
1420 }
1421
Jens Axboe7eaceac2011-03-10 08:52:07 +01001422 lock_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001423 /*
1424 * This test is racy because PG_hwpoison is set outside of page lock.
1425 * That's acceptable because that won't trigger kernel panic. Instead,
1426 * the PG_hwpoison page will be caught and isolated on the entrance to
1427 * the free buddy page pool.
1428 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001429 if (TestClearPageHWPoison(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001430 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
Xishi Qiu293c07e2013-02-22 16:34:02 -08001431 atomic_long_sub(nr_pages, &num_poisoned_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001432 freeit = 1;
Naoya Horiguchi6a901812010-09-08 10:19:40 +09001433 if (PageHuge(page))
1434 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001435 }
1436 unlock_page(page);
1437
1438 put_page(page);
Wanpeng Li3ba5eeb2013-09-11 14:23:01 -07001439 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
Wu Fengguang847ce402009-12-16 12:19:58 +01001440 put_page(page);
1441
1442 return 0;
1443}
1444EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001445
1446static struct page *new_page(struct page *p, unsigned long private, int **x)
1447{
Andi Kleen12686d12009-12-16 12:20:01 +01001448 int nid = page_to_nid(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001449 if (PageHuge(p))
1450 return alloc_huge_page_node(page_hstate(compound_head(p)),
1451 nid);
1452 else
1453 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001454}
1455
1456/*
1457 * Safely get reference count of an arbitrary page.
1458 * Returns 0 for a free page, -EIO for a zero refcount page
1459 * that is not free, and 1 for any other page type.
1460 * For 1 the page is returned with increased page count, otherwise not.
1461 */
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001462static int __get_any_page(struct page *p, unsigned long pfn, int flags)
Andi Kleenfacb6012009-12-16 12:20:00 +01001463{
1464 int ret;
1465
1466 if (flags & MF_COUNT_INCREASED)
1467 return 1;
1468
1469 /*
Naoya Horiguchid950b952010-09-08 10:19:39 +09001470 * When the target page is a free hugepage, just remove it
1471 * from free hugepage list.
1472 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001473 if (!get_page_unless_zero(compound_head(p))) {
Naoya Horiguchid950b952010-09-08 10:19:39 +09001474 if (PageHuge(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001475 pr_info("%s: %#lx free huge page\n", __func__, pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001476 ret = 0;
Naoya Horiguchid950b952010-09-08 10:19:39 +09001477 } else if (is_free_buddy_page(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001478 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001479 ret = 0;
1480 } else {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001481 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1482 __func__, pfn, p->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001483 ret = -EIO;
1484 }
1485 } else {
1486 /* Not a free page */
1487 ret = 1;
1488 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001489 return ret;
1490}
1491
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001492static int get_any_page(struct page *page, unsigned long pfn, int flags)
1493{
1494 int ret = __get_any_page(page, pfn, flags);
1495
1496 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1497 /*
1498 * Try to free it.
1499 */
1500 put_page(page);
1501 shake_page(page, 1);
1502
1503 /*
1504 * Did it turn free?
1505 */
1506 ret = __get_any_page(page, pfn, 0);
1507 if (!PageLRU(page)) {
1508 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1509 pfn, page->flags);
1510 return -EIO;
1511 }
1512 }
1513 return ret;
1514}
1515
Naoya Horiguchid950b952010-09-08 10:19:39 +09001516static int soft_offline_huge_page(struct page *page, int flags)
1517{
1518 int ret;
1519 unsigned long pfn = page_to_pfn(page);
1520 struct page *hpage = compound_head(page);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001521 LIST_HEAD(pagelist);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001522
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001523 /*
1524 * This double-check of PageHWPoison is to avoid the race with
1525 * memory_failure(). See also comment in __soft_offline_page().
1526 */
1527 lock_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001528 if (PageHWPoison(hpage)) {
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001529 unlock_page(hpage);
1530 put_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001531 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001532 return -EBUSY;
Xishi Qiu0ebff322013-02-22 16:33:59 -08001533 }
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001534 unlock_page(hpage);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001535
Naoya Horiguchid950b952010-09-08 10:19:39 +09001536 /* Keep page count to indicate a given hugepage is isolated. */
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001537 list_move(&hpage->lru, &pagelist);
David Rientjes68711a72014-06-04 16:08:25 -07001538 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001539 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001540 if (ret) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001541 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1542 pfn, ret, page->flags);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001543 /*
1544 * We know that soft_offline_huge_page() tries to migrate
1545 * only one hugepage pointed to by hpage, so we need not
1546 * run through the pagelist here.
1547 */
1548 putback_active_hugepage(hpage);
1549 if (ret > 0)
1550 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001551 } else {
Jianguo Wua49ecbc2013-12-18 17:08:54 -08001552 /* overcommit hugetlb page will be freed to buddy */
1553 if (PageHuge(page)) {
1554 set_page_hwpoison_huge_page(hpage);
1555 dequeue_hwpoisoned_huge_page(hpage);
1556 atomic_long_add(1 << compound_order(hpage),
1557 &num_poisoned_pages);
1558 } else {
1559 SetPageHWPoison(page);
1560 atomic_long_inc(&num_poisoned_pages);
1561 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001562 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001563 return ret;
1564}
1565
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001566static int __soft_offline_page(struct page *page, int flags)
1567{
1568 int ret;
1569 unsigned long pfn = page_to_pfn(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001570
1571 /*
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001572 * Check PageHWPoison again inside page lock because PageHWPoison
1573 * is set by memory_failure() outside page lock. Note that
1574 * memory_failure() also double-checks PageHWPoison inside page lock,
1575 * so there's no race between soft_offline_page() and memory_failure().
Andi Kleenfacb6012009-12-16 12:20:00 +01001576 */
Xishi Qiu0ebff322013-02-22 16:33:59 -08001577 lock_page(page);
1578 wait_on_page_writeback(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001579 if (PageHWPoison(page)) {
1580 unlock_page(page);
1581 put_page(page);
1582 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1583 return -EBUSY;
1584 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001585 /*
1586 * Try to invalidate first. This should work for
1587 * non dirty unmapped page cache pages.
1588 */
1589 ret = invalidate_inode_page(page);
1590 unlock_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001591 /*
Andi Kleenfacb6012009-12-16 12:20:00 +01001592 * RED-PEN would be better to keep it isolated here, but we
1593 * would need to fix isolation locking first.
1594 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001595 if (ret == 1) {
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001596 put_page(page);
Andi Kleenfb46e732010-09-27 23:31:30 +02001597 pr_info("soft_offline: %#lx: invalidated\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001598 SetPageHWPoison(page);
1599 atomic_long_inc(&num_poisoned_pages);
1600 return 0;
Andi Kleenfacb6012009-12-16 12:20:00 +01001601 }
1602
1603 /*
1604 * Simple invalidation didn't work.
1605 * Try to migrate to a new page instead. migrate.c
1606 * handles a large number of cases for us.
1607 */
1608 ret = isolate_lru_page(page);
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001609 /*
1610 * Drop page reference which is came from get_any_page()
1611 * successful isolate_lru_page() already took another one.
1612 */
1613 put_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001614 if (!ret) {
1615 LIST_HEAD(pagelist);
Minchan Kim5db8a732011-06-15 15:08:48 -07001616 inc_zone_page_state(page, NR_ISOLATED_ANON +
Hugh Dickins9c620e22013-02-22 16:35:14 -08001617 page_is_file_cache(page));
Andi Kleenfacb6012009-12-16 12:20:00 +01001618 list_add(&page->lru, &pagelist);
David Rientjes68711a72014-06-04 16:08:25 -07001619 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001620 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Andi Kleenfacb6012009-12-16 12:20:00 +01001621 if (ret) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001622 if (!list_empty(&pagelist)) {
1623 list_del(&page->lru);
1624 dec_zone_page_state(page, NR_ISOLATED_ANON +
1625 page_is_file_cache(page));
1626 putback_lru_page(page);
1627 }
1628
Andi Kleenfb46e732010-09-27 23:31:30 +02001629 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001630 pfn, ret, page->flags);
1631 if (ret > 0)
1632 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001633 } else {
Naoya Horiguchif15bdfa2013-07-03 15:02:37 -07001634 /*
1635 * After page migration succeeds, the source page can
1636 * be trapped in pagevec and actual freeing is delayed.
1637 * Freeing code works differently based on PG_hwpoison,
1638 * so there's a race. We need to make sure that the
1639 * source page should be freed back to buddy before
1640 * setting PG_hwpoison.
1641 */
1642 if (!is_free_buddy_page(page))
1643 lru_add_drain_all();
1644 if (!is_free_buddy_page(page))
1645 drain_all_pages();
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001646 SetPageHWPoison(page);
Naoya Horiguchif15bdfa2013-07-03 15:02:37 -07001647 if (!is_free_buddy_page(page))
1648 pr_info("soft offline: %#lx: page leaked\n",
1649 pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001650 atomic_long_inc(&num_poisoned_pages);
Andi Kleenfacb6012009-12-16 12:20:00 +01001651 }
1652 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001653 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
Dean Nelsondd73e852011-10-31 17:09:04 -07001654 pfn, ret, page_count(page), page->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001655 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001656 return ret;
1657}
Wanpeng Li86e05772013-09-11 14:22:56 -07001658
1659/**
1660 * soft_offline_page - Soft offline a page.
1661 * @page: page to offline
1662 * @flags: flags. Same as memory_failure().
1663 *
1664 * Returns 0 on success, otherwise negated errno.
1665 *
1666 * Soft offline a page, by migration or invalidation,
1667 * without killing anything. This is for the case when
1668 * a page is not corrupted yet (so it's still valid to access),
1669 * but has had a number of corrected errors and is better taken
1670 * out.
1671 *
1672 * The actual policy on when to do that is maintained by
1673 * user space.
1674 *
1675 * This should never impact any application or cause data loss,
1676 * however it might take some time.
1677 *
1678 * This is not a 100% solution for all memory, but tries to be
1679 * ``good enough'' for the majority of memory.
1680 */
1681int soft_offline_page(struct page *page, int flags)
1682{
1683 int ret;
1684 unsigned long pfn = page_to_pfn(page);
David Rientjes668f9abb2014-03-03 15:38:18 -08001685 struct page *hpage = compound_head(page);
Wanpeng Li86e05772013-09-11 14:22:56 -07001686
1687 if (PageHWPoison(page)) {
1688 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1689 return -EBUSY;
1690 }
1691 if (!PageHuge(page) && PageTransHuge(hpage)) {
1692 if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1693 pr_info("soft offline: %#lx: failed to split THP\n",
1694 pfn);
1695 return -EBUSY;
1696 }
1697 }
1698
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07001699 get_online_mems();
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001700
1701 /*
1702 * Isolate the page, so that it doesn't get reallocated if it
1703 * was free. This flag should be kept set until the source page
1704 * is freed and PG_hwpoison on it is set.
1705 */
1706 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
1707 set_migratetype_isolate(page, true);
1708
Wanpeng Li86e05772013-09-11 14:22:56 -07001709 ret = get_any_page(page, pfn, flags);
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07001710 put_online_mems();
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001711 if (ret > 0) { /* for in-use pages */
Wanpeng Li86e05772013-09-11 14:22:56 -07001712 if (PageHuge(page))
1713 ret = soft_offline_huge_page(page, flags);
1714 else
1715 ret = __soft_offline_page(page, flags);
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001716 } else if (ret == 0) { /* for free pages */
Wanpeng Li86e05772013-09-11 14:22:56 -07001717 if (PageHuge(page)) {
1718 set_page_hwpoison_huge_page(hpage);
1719 dequeue_hwpoisoned_huge_page(hpage);
1720 atomic_long_add(1 << compound_order(hpage),
1721 &num_poisoned_pages);
1722 } else {
1723 SetPageHWPoison(page);
1724 atomic_long_inc(&num_poisoned_pages);
1725 }
1726 }
Wanpeng Li86e05772013-09-11 14:22:56 -07001727 unset_migratetype_isolate(page, MIGRATE_MOVABLE);
1728 return ret;
1729}