blob: fabe55046c1d7b9638172aa8e26ecb4f3d6ed4b9 [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);
148 /* root_mem_cgroup has NULL dentries */
149 if (!css->cgroup->dentry)
150 return -EINVAL;
151
152 ino = css->cgroup->dentry->d_inode->i_ino;
153 css_put(css);
154
155 if (ino != hwpoison_filter_memcg)
156 return -EINVAL;
157
158 return 0;
159}
160#else
161static int hwpoison_filter_task(struct page *p) { return 0; }
162#endif
163
Wu Fengguang7c116f22009-12-16 12:19:59 +0100164int hwpoison_filter(struct page *p)
165{
Haicheng Li1bfe5fe2009-12-16 12:19:59 +0100166 if (!hwpoison_filter_enable)
167 return 0;
168
Wu Fengguang7c116f22009-12-16 12:19:59 +0100169 if (hwpoison_filter_dev(p))
170 return -EINVAL;
171
Wu Fengguang478c5ff2009-12-16 12:19:59 +0100172 if (hwpoison_filter_flags(p))
173 return -EINVAL;
174
Andi Kleen4fd466e2009-12-16 12:19:59 +0100175 if (hwpoison_filter_task(p))
176 return -EINVAL;
177
Wu Fengguang7c116f22009-12-16 12:19:59 +0100178 return 0;
179}
Andi Kleen27df5062009-12-21 19:56:42 +0100180#else
181int hwpoison_filter(struct page *p)
182{
183 return 0;
184}
185#endif
186
Wu Fengguang7c116f22009-12-16 12:19:59 +0100187EXPORT_SYMBOL_GPL(hwpoison_filter);
188
Andi Kleen6a460792009-09-16 11:50:15 +0200189/*
Tony Luck7329bbe2011-12-13 09:27:58 -0800190 * Send all the processes who have the page mapped a signal.
191 * ``action optional'' if they are not immediately affected by the error
192 * ``action required'' if error happened in current execution context
Andi Kleen6a460792009-09-16 11:50:15 +0200193 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800194static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
195 unsigned long pfn, struct page *page, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200196{
197 struct siginfo si;
198 int ret;
199
200 printk(KERN_ERR
Tony Luck7329bbe2011-12-13 09:27:58 -0800201 "MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200202 pfn, t->comm, t->pid);
203 si.si_signo = SIGBUS;
204 si.si_errno = 0;
Andi Kleen6a460792009-09-16 11:50:15 +0200205 si.si_addr = (void *)addr;
206#ifdef __ARCH_SI_TRAPNO
207 si.si_trapno = trapno;
208#endif
Wanpeng Lif9121152013-09-11 14:22:52 -0700209 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
Tony Luck7329bbe2011-12-13 09:27:58 -0800210
211 if ((flags & MF_ACTION_REQUIRED) && t == current) {
212 si.si_code = BUS_MCEERR_AR;
213 ret = force_sig_info(SIGBUS, &si, t);
214 } else {
215 /*
216 * Don't use force here, it's convenient if the signal
217 * can be temporarily blocked.
218 * This could cause a loop when the user sets SIGBUS
219 * to SIG_IGN, but hopefully no one will do that?
220 */
221 si.si_code = BUS_MCEERR_AO;
222 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
223 }
Andi Kleen6a460792009-09-16 11:50:15 +0200224 if (ret < 0)
225 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
226 t->comm, t->pid, ret);
227 return ret;
228}
229
230/*
Andi Kleen588f9ce2009-12-16 12:19:57 +0100231 * When a unknown page type is encountered drain as many buffers as possible
232 * in the hope to turn the page into a LRU or free page, which we can handle.
233 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100234void shake_page(struct page *p, int access)
Andi Kleen588f9ce2009-12-16 12:19:57 +0100235{
236 if (!PageSlab(p)) {
237 lru_add_drain_all();
238 if (PageLRU(p))
239 return;
240 drain_all_pages();
241 if (PageLRU(p) || is_free_buddy_page(p))
242 return;
243 }
Andi Kleenfacb6012009-12-16 12:20:00 +0100244
Andi Kleen588f9ce2009-12-16 12:19:57 +0100245 /*
Jin Dongmingaf241a02011-02-01 15:52:41 -0800246 * Only call shrink_slab here (which would also shrink other caches) if
247 * access is not potentially fatal.
Andi Kleen588f9ce2009-12-16 12:19:57 +0100248 */
Andi Kleenfacb6012009-12-16 12:20:00 +0100249 if (access) {
250 int nr;
Dave Chinner0ce3d742013-08-28 10:18:03 +1000251 int nid = page_to_nid(p);
Andi Kleenfacb6012009-12-16 12:20:00 +0100252 do {
Ying Hana09ed5e2011-05-24 17:12:26 -0700253 struct shrink_control shrink = {
254 .gfp_mask = GFP_KERNEL,
Ying Hana09ed5e2011-05-24 17:12:26 -0700255 };
Dave Chinner0ce3d742013-08-28 10:18:03 +1000256 node_set(nid, shrink.nodes_to_scan);
Ying Hana09ed5e2011-05-24 17:12:26 -0700257
Ying Han1495f232011-05-24 17:12:27 -0700258 nr = shrink_slab(&shrink, 1000, 1000);
Andi Kleen47f43e72010-09-28 07:37:55 +0200259 if (page_count(p) == 1)
Andi Kleenfacb6012009-12-16 12:20:00 +0100260 break;
261 } while (nr > 10);
262 }
Andi Kleen588f9ce2009-12-16 12:19:57 +0100263}
264EXPORT_SYMBOL_GPL(shake_page);
265
266/*
Andi Kleen6a460792009-09-16 11:50:15 +0200267 * Kill all processes that have a poisoned page mapped and then isolate
268 * the page.
269 *
270 * General strategy:
271 * Find all processes having the page mapped and kill them.
272 * But we keep a page reference around so that the page is not
273 * actually freed yet.
274 * Then stash the page away
275 *
276 * There's no convenient way to get back to mapped processes
277 * from the VMAs. So do a brute-force search over all
278 * running processes.
279 *
280 * Remember that machine checks are not common (or rather
281 * if they are common you have other problems), so this shouldn't
282 * be a performance issue.
283 *
284 * Also there are some races possible while we get from the
285 * error detection to actually handle it.
286 */
287
288struct to_kill {
289 struct list_head nd;
290 struct task_struct *tsk;
291 unsigned long addr;
Andi Kleen9033ae12010-09-27 23:36:05 +0200292 char addr_valid;
Andi Kleen6a460792009-09-16 11:50:15 +0200293};
294
295/*
296 * Failure handling: if we can't find or can't kill a process there's
297 * not much we can do. We just print a message and ignore otherwise.
298 */
299
300/*
301 * Schedule a process for later kill.
302 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
303 * TBD would GFP_NOIO be enough?
304 */
305static void add_to_kill(struct task_struct *tsk, struct page *p,
306 struct vm_area_struct *vma,
307 struct list_head *to_kill,
308 struct to_kill **tkc)
309{
310 struct to_kill *tk;
311
312 if (*tkc) {
313 tk = *tkc;
314 *tkc = NULL;
315 } else {
316 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
317 if (!tk) {
318 printk(KERN_ERR
319 "MCE: Out of memory while machine check handling\n");
320 return;
321 }
322 }
323 tk->addr = page_address_in_vma(p, vma);
324 tk->addr_valid = 1;
325
326 /*
327 * In theory we don't have to kill when the page was
328 * munmaped. But it could be also a mremap. Since that's
329 * likely very rare kill anyways just out of paranoia, but use
330 * a SIGKILL because the error is not contained anymore.
331 */
332 if (tk->addr == -EFAULT) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200333 pr_info("MCE: Unable to find user space address %lx in %s\n",
Andi Kleen6a460792009-09-16 11:50:15 +0200334 page_to_pfn(p), tsk->comm);
335 tk->addr_valid = 0;
336 }
337 get_task_struct(tsk);
338 tk->tsk = tsk;
339 list_add_tail(&tk->nd, to_kill);
340}
341
342/*
343 * Kill the processes that have been collected earlier.
344 *
345 * Only do anything when DOIT is set, otherwise just free the list
346 * (this is used for clean pages which do not need killing)
347 * Also when FAIL is set do a force kill because something went
348 * wrong earlier.
349 */
Tony Luck6751ed62012-07-11 10:20:47 -0700350static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -0800351 int fail, struct page *page, unsigned long pfn,
352 int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200353{
354 struct to_kill *tk, *next;
355
356 list_for_each_entry_safe (tk, next, to_kill, nd) {
Tony Luck6751ed62012-07-11 10:20:47 -0700357 if (forcekill) {
Andi Kleen6a460792009-09-16 11:50:15 +0200358 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200359 * In case something went wrong with munmapping
Andi Kleen6a460792009-09-16 11:50:15 +0200360 * make sure the process doesn't catch the
361 * signal and then access the memory. Just kill it.
Andi Kleen6a460792009-09-16 11:50:15 +0200362 */
363 if (fail || tk->addr_valid == 0) {
364 printk(KERN_ERR
365 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
366 pfn, tk->tsk->comm, tk->tsk->pid);
367 force_sig(SIGKILL, tk->tsk);
368 }
369
370 /*
371 * In theory the process could have mapped
372 * something else on the address in-between. We could
373 * check for that, but we need to tell the
374 * process anyways.
375 */
Tony Luck7329bbe2011-12-13 09:27:58 -0800376 else if (kill_proc(tk->tsk, tk->addr, trapno,
377 pfn, page, flags) < 0)
Andi Kleen6a460792009-09-16 11:50:15 +0200378 printk(KERN_ERR
379 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
380 pfn, tk->tsk->comm, tk->tsk->pid);
381 }
382 put_task_struct(tk->tsk);
383 kfree(tk);
384 }
385}
386
387static int task_early_kill(struct task_struct *tsk)
388{
389 if (!tsk->mm)
390 return 0;
391 if (tsk->flags & PF_MCE_PROCESS)
392 return !!(tsk->flags & PF_MCE_EARLY);
393 return sysctl_memory_failure_early_kill;
394}
395
396/*
397 * Collect processes when the error hit an anonymous page.
398 */
399static void collect_procs_anon(struct page *page, struct list_head *to_kill,
400 struct to_kill **tkc)
401{
402 struct vm_area_struct *vma;
403 struct task_struct *tsk;
404 struct anon_vma *av;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700405 pgoff_t pgoff;
Andi Kleen6a460792009-09-16 11:50:15 +0200406
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000407 av = page_lock_anon_vma_read(page);
Andi Kleen6a460792009-09-16 11:50:15 +0200408 if (av == NULL) /* Not actually mapped anymore */
Peter Zijlstra9b679322011-06-27 16:18:09 -0700409 return;
410
Michel Lespinassebf181b92012-10-08 16:31:39 -0700411 pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700412 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200413 for_each_process (tsk) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800414 struct anon_vma_chain *vmac;
415
Andi Kleen6a460792009-09-16 11:50:15 +0200416 if (!task_early_kill(tsk))
417 continue;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700418 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
419 pgoff, pgoff) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800420 vma = vmac->vma;
Andi Kleen6a460792009-09-16 11:50:15 +0200421 if (!page_mapped_in_vma(page, vma))
422 continue;
423 if (vma->vm_mm == tsk->mm)
424 add_to_kill(tsk, page, vma, to_kill, tkc);
425 }
426 }
Andi Kleen6a460792009-09-16 11:50:15 +0200427 read_unlock(&tasklist_lock);
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000428 page_unlock_anon_vma_read(av);
Andi Kleen6a460792009-09-16 11:50:15 +0200429}
430
431/*
432 * Collect processes when the error hit a file mapped page.
433 */
434static void collect_procs_file(struct page *page, struct list_head *to_kill,
435 struct to_kill **tkc)
436{
437 struct vm_area_struct *vma;
438 struct task_struct *tsk;
Andi Kleen6a460792009-09-16 11:50:15 +0200439 struct address_space *mapping = page->mapping;
440
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700441 mutex_lock(&mapping->i_mmap_mutex);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700442 read_lock(&tasklist_lock);
Andi Kleen6a460792009-09-16 11:50:15 +0200443 for_each_process(tsk) {
444 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
445
446 if (!task_early_kill(tsk))
447 continue;
448
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -0700449 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
Andi Kleen6a460792009-09-16 11:50:15 +0200450 pgoff) {
451 /*
452 * Send early kill signal to tasks where a vma covers
453 * the page but the corrupted page is not necessarily
454 * mapped it in its pte.
455 * Assume applications who requested early kill want
456 * to be informed of all such data corruptions.
457 */
458 if (vma->vm_mm == tsk->mm)
459 add_to_kill(tsk, page, vma, to_kill, tkc);
460 }
461 }
Andi Kleen6a460792009-09-16 11:50:15 +0200462 read_unlock(&tasklist_lock);
Peter Zijlstra9b679322011-06-27 16:18:09 -0700463 mutex_unlock(&mapping->i_mmap_mutex);
Andi Kleen6a460792009-09-16 11:50:15 +0200464}
465
466/*
467 * Collect the processes who have the corrupted page mapped to kill.
468 * This is done in two steps for locking reasons.
469 * First preallocate one tokill structure outside the spin locks,
470 * so that we can kill at least one process reasonably reliable.
471 */
472static void collect_procs(struct page *page, struct list_head *tokill)
473{
474 struct to_kill *tk;
475
476 if (!page->mapping)
477 return;
478
479 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
480 if (!tk)
481 return;
482 if (PageAnon(page))
483 collect_procs_anon(page, tokill, &tk);
484 else
485 collect_procs_file(page, tokill, &tk);
486 kfree(tk);
487}
488
489/*
490 * Error handlers for various types of pages.
491 */
492
493enum outcome {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100494 IGNORED, /* Error: cannot be handled */
495 FAILED, /* Error: handling failed */
Andi Kleen6a460792009-09-16 11:50:15 +0200496 DELAYED, /* Will be handled later */
Andi Kleen6a460792009-09-16 11:50:15 +0200497 RECOVERED, /* Successfully recovered */
498};
499
500static const char *action_name[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100501 [IGNORED] = "Ignored",
Andi Kleen6a460792009-09-16 11:50:15 +0200502 [FAILED] = "Failed",
503 [DELAYED] = "Delayed",
Andi Kleen6a460792009-09-16 11:50:15 +0200504 [RECOVERED] = "Recovered",
505};
506
507/*
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100508 * XXX: It is possible that a page is isolated from LRU cache,
509 * and then kept in swap cache or failed to remove from page cache.
510 * The page count will stop it from being freed by unpoison.
511 * Stress tests should be aware of this memory leak problem.
512 */
513static int delete_from_lru_cache(struct page *p)
514{
515 if (!isolate_lru_page(p)) {
516 /*
517 * Clear sensible page flags, so that the buddy system won't
518 * complain when the page is unpoison-and-freed.
519 */
520 ClearPageActive(p);
521 ClearPageUnevictable(p);
522 /*
523 * drop the page count elevated by isolate_lru_page()
524 */
525 page_cache_release(p);
526 return 0;
527 }
528 return -EIO;
529}
530
531/*
Andi Kleen6a460792009-09-16 11:50:15 +0200532 * Error hit kernel page.
533 * Do nothing, try to be lucky and not touch this instead. For a few cases we
534 * could be more sophisticated.
535 */
536static int me_kernel(struct page *p, unsigned long pfn)
537{
Andi Kleen6a460792009-09-16 11:50:15 +0200538 return IGNORED;
539}
540
541/*
542 * Page in unknown state. Do nothing.
543 */
544static int me_unknown(struct page *p, unsigned long pfn)
545{
546 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
547 return FAILED;
548}
549
550/*
Andi Kleen6a460792009-09-16 11:50:15 +0200551 * Clean (or cleaned) page cache page.
552 */
553static int me_pagecache_clean(struct page *p, unsigned long pfn)
554{
555 int err;
556 int ret = FAILED;
557 struct address_space *mapping;
558
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100559 delete_from_lru_cache(p);
560
Andi Kleen6a460792009-09-16 11:50:15 +0200561 /*
562 * For anonymous pages we're done the only reference left
563 * should be the one m_f() holds.
564 */
565 if (PageAnon(p))
566 return RECOVERED;
567
568 /*
569 * Now truncate the page in the page cache. This is really
570 * more like a "temporary hole punch"
571 * Don't do this for block devices when someone else
572 * has a reference, because it could be file system metadata
573 * and that's not safe to truncate.
574 */
575 mapping = page_mapping(p);
576 if (!mapping) {
577 /*
578 * Page has been teared down in the meanwhile
579 */
580 return FAILED;
581 }
582
583 /*
584 * Truncation is a bit tricky. Enable it per file system for now.
585 *
586 * Open: to take i_mutex or not for this? Right now we don't.
587 */
588 if (mapping->a_ops->error_remove_page) {
589 err = mapping->a_ops->error_remove_page(mapping, p);
590 if (err != 0) {
591 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
592 pfn, err);
593 } else if (page_has_private(p) &&
594 !try_to_release_page(p, GFP_NOIO)) {
Andi Kleenfb46e732010-09-27 23:31:30 +0200595 pr_info("MCE %#lx: failed to release buffers\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +0200596 } else {
597 ret = RECOVERED;
598 }
599 } else {
600 /*
601 * If the file system doesn't support it just invalidate
602 * This fails on dirty or anything with private pages
603 */
604 if (invalidate_inode_page(p))
605 ret = RECOVERED;
606 else
607 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
608 pfn);
609 }
610 return ret;
611}
612
613/*
614 * Dirty cache page page
615 * Issues: when the error hit a hole page the error is not properly
616 * propagated.
617 */
618static int me_pagecache_dirty(struct page *p, unsigned long pfn)
619{
620 struct address_space *mapping = page_mapping(p);
621
622 SetPageError(p);
623 /* TBD: print more information about the file. */
624 if (mapping) {
625 /*
626 * IO error will be reported by write(), fsync(), etc.
627 * who check the mapping.
628 * This way the application knows that something went
629 * wrong with its dirty file data.
630 *
631 * There's one open issue:
632 *
633 * The EIO will be only reported on the next IO
634 * operation and then cleared through the IO map.
635 * Normally Linux has two mechanisms to pass IO error
636 * first through the AS_EIO flag in the address space
637 * and then through the PageError flag in the page.
638 * Since we drop pages on memory failure handling the
639 * only mechanism open to use is through AS_AIO.
640 *
641 * This has the disadvantage that it gets cleared on
642 * the first operation that returns an error, while
643 * the PageError bit is more sticky and only cleared
644 * when the page is reread or dropped. If an
645 * application assumes it will always get error on
646 * fsync, but does other operations on the fd before
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300647 * and the page is dropped between then the error
Andi Kleen6a460792009-09-16 11:50:15 +0200648 * will not be properly reported.
649 *
650 * This can already happen even without hwpoisoned
651 * pages: first on metadata IO errors (which only
652 * report through AS_EIO) or when the page is dropped
653 * at the wrong time.
654 *
655 * So right now we assume that the application DTRT on
656 * the first EIO, but we're not worse than other parts
657 * of the kernel.
658 */
659 mapping_set_error(mapping, EIO);
660 }
661
662 return me_pagecache_clean(p, pfn);
663}
664
665/*
666 * Clean and dirty swap cache.
667 *
668 * Dirty swap cache page is tricky to handle. The page could live both in page
669 * cache and swap cache(ie. page is freshly swapped in). So it could be
670 * referenced concurrently by 2 types of PTEs:
671 * normal PTEs and swap PTEs. We try to handle them consistently by calling
672 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
673 * and then
674 * - clear dirty bit to prevent IO
675 * - remove from LRU
676 * - but keep in the swap cache, so that when we return to it on
677 * a later page fault, we know the application is accessing
678 * corrupted data and shall be killed (we installed simple
679 * interception code in do_swap_page to catch it).
680 *
681 * Clean swap cache pages can be directly isolated. A later page fault will
682 * bring in the known good data from disk.
683 */
684static int me_swapcache_dirty(struct page *p, unsigned long pfn)
685{
Andi Kleen6a460792009-09-16 11:50:15 +0200686 ClearPageDirty(p);
687 /* Trigger EIO in shmem: */
688 ClearPageUptodate(p);
689
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100690 if (!delete_from_lru_cache(p))
691 return DELAYED;
692 else
693 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200694}
695
696static int me_swapcache_clean(struct page *p, unsigned long pfn)
697{
Andi Kleen6a460792009-09-16 11:50:15 +0200698 delete_from_swap_cache(p);
Wu Fengguange43c3af2009-09-29 13:16:20 +0800699
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +0100700 if (!delete_from_lru_cache(p))
701 return RECOVERED;
702 else
703 return FAILED;
Andi Kleen6a460792009-09-16 11:50:15 +0200704}
705
706/*
707 * Huge pages. Needs work.
708 * Issues:
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900709 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
710 * To narrow down kill region to one page, we need to break up pmd.
Andi Kleen6a460792009-09-16 11:50:15 +0200711 */
712static int me_huge_page(struct page *p, unsigned long pfn)
713{
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900714 int res = 0;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900715 struct page *hpage = compound_head(p);
716 /*
717 * We can safely recover from error on free or reserved (i.e.
718 * not in-use) hugepage by dequeuing it from freelist.
719 * To check whether a hugepage is in-use or not, we can't use
720 * page->lru because it can be used in other hugepage operations,
721 * such as __unmap_hugepage_range() and gather_surplus_pages().
722 * So instead we use page_mapping() and PageAnon().
723 * We assume that this function is called with page lock held,
724 * so there is no race between isolation and mapping/unmapping.
725 */
726 if (!(page_mapping(hpage) || PageAnon(hpage))) {
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +0900727 res = dequeue_hwpoisoned_huge_page(hpage);
728 if (!res)
729 return RECOVERED;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +0900730 }
731 return DELAYED;
Andi Kleen6a460792009-09-16 11:50:15 +0200732}
733
734/*
735 * Various page states we can handle.
736 *
737 * A page state is defined by its current page->flags bits.
738 * The table matches them in order and calls the right handler.
739 *
740 * This is quite tricky because we can access page at any time
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300741 * in its live cycle, so all accesses have to be extremely careful.
Andi Kleen6a460792009-09-16 11:50:15 +0200742 *
743 * This is not complete. More states could be added.
744 * For any missing state don't attempt recovery.
745 */
746
747#define dirty (1UL << PG_dirty)
748#define sc (1UL << PG_swapcache)
749#define unevict (1UL << PG_unevictable)
750#define mlock (1UL << PG_mlocked)
751#define writeback (1UL << PG_writeback)
752#define lru (1UL << PG_lru)
753#define swapbacked (1UL << PG_swapbacked)
754#define head (1UL << PG_head)
755#define tail (1UL << PG_tail)
756#define compound (1UL << PG_compound)
757#define slab (1UL << PG_slab)
Andi Kleen6a460792009-09-16 11:50:15 +0200758#define reserved (1UL << PG_reserved)
759
760static struct page_state {
761 unsigned long mask;
762 unsigned long res;
763 char *msg;
764 int (*action)(struct page *p, unsigned long pfn);
765} error_states[] = {
Wu Fengguangd95ea512009-12-16 12:19:58 +0100766 { reserved, reserved, "reserved kernel", me_kernel },
Wu Fengguang95d01fc2009-12-16 12:19:58 +0100767 /*
768 * free pages are specially detected outside this table:
769 * PG_buddy pages only make a small fraction of all free pages.
770 */
Andi Kleen6a460792009-09-16 11:50:15 +0200771
772 /*
773 * Could in theory check if slab page is free or if we can drop
774 * currently unused objects without touching them. But just
775 * treat it as standard kernel for now.
776 */
777 { slab, slab, "kernel slab", me_kernel },
778
779#ifdef CONFIG_PAGEFLAGS_EXTENDED
780 { head, head, "huge", me_huge_page },
781 { tail, tail, "huge", me_huge_page },
782#else
783 { compound, compound, "huge", me_huge_page },
784#endif
785
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800786 { sc|dirty, sc|dirty, "dirty swapcache", me_swapcache_dirty },
787 { sc|dirty, sc, "clean swapcache", me_swapcache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200788
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800789 { mlock|dirty, mlock|dirty, "dirty mlocked LRU", me_pagecache_dirty },
Naoya Horiguchie3986292013-04-29 15:06:08 -0700790 { mlock|dirty, mlock, "clean mlocked LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200791
Naoya Horiguchi5f4b9fc2013-02-22 16:35:53 -0800792 { unevict|dirty, unevict|dirty, "dirty unevictable LRU", me_pagecache_dirty },
Naoya Horiguchie3986292013-04-29 15:06:08 -0700793 { unevict|dirty, unevict, "clean unevictable LRU", me_pagecache_clean },
Naoya Horiguchi5f4b9fc2013-02-22 16:35:53 -0800794
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800795 { lru|dirty, lru|dirty, "dirty LRU", me_pagecache_dirty },
Andi Kleen6a460792009-09-16 11:50:15 +0200796 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
Andi Kleen6a460792009-09-16 11:50:15 +0200797
798 /*
799 * Catchall entry: must be at end.
800 */
801 { 0, 0, "unknown page state", me_unknown },
802};
803
Andi Kleen2326c462009-12-16 12:20:00 +0100804#undef dirty
805#undef sc
806#undef unevict
807#undef mlock
808#undef writeback
809#undef lru
810#undef swapbacked
811#undef head
812#undef tail
813#undef compound
814#undef slab
815#undef reserved
816
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800817/*
818 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
819 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
820 */
Andi Kleen6a460792009-09-16 11:50:15 +0200821static void action_result(unsigned long pfn, char *msg, int result)
822{
Naoya Horiguchiff604cf2012-12-11 16:01:32 -0800823 pr_err("MCE %#lx: %s page recovery: %s\n",
824 pfn, msg, action_name[result]);
Andi Kleen6a460792009-09-16 11:50:15 +0200825}
826
827static int page_action(struct page_state *ps, struct page *p,
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100828 unsigned long pfn)
Andi Kleen6a460792009-09-16 11:50:15 +0200829{
830 int result;
Wu Fengguang7456b042009-10-19 08:15:01 +0200831 int count;
Andi Kleen6a460792009-09-16 11:50:15 +0200832
833 result = ps->action(p, pfn);
834 action_result(pfn, ps->msg, result);
Wu Fengguang7456b042009-10-19 08:15:01 +0200835
Wu Fengguangbd1ce5f2009-12-16 12:19:57 +0100836 count = page_count(p) - 1;
Wu Fengguang138ce282009-12-16 12:19:58 +0100837 if (ps->action == me_swapcache_dirty && result == DELAYED)
838 count--;
839 if (count != 0) {
Andi Kleen6a460792009-09-16 11:50:15 +0200840 printk(KERN_ERR
841 "MCE %#lx: %s page still referenced by %d users\n",
Wu Fengguang7456b042009-10-19 08:15:01 +0200842 pfn, ps->msg, count);
Wu Fengguang138ce282009-12-16 12:19:58 +0100843 result = FAILED;
844 }
Andi Kleen6a460792009-09-16 11:50:15 +0200845
846 /* Could do more checks here if page looks ok */
847 /*
848 * Could adjust zone counters here to correct for the missing page.
849 */
850
Wu Fengguang138ce282009-12-16 12:19:58 +0100851 return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +0200852}
853
Andi Kleen6a460792009-09-16 11:50:15 +0200854/*
855 * Do all that is necessary to remove user space mappings. Unmap
856 * the pages and send SIGBUS to the processes if the data was dirty.
857 */
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100858static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
Tony Luck7329bbe2011-12-13 09:27:58 -0800859 int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +0200860{
861 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
862 struct address_space *mapping;
863 LIST_HEAD(tokill);
864 int ret;
Tony Luck6751ed62012-07-11 10:20:47 -0700865 int kill = 1, forcekill;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900866 struct page *hpage = compound_head(p);
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800867 struct page *ppage;
Andi Kleen6a460792009-09-16 11:50:15 +0200868
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100869 if (PageReserved(p) || PageSlab(p))
870 return SWAP_SUCCESS;
Andi Kleen6a460792009-09-16 11:50:15 +0200871
Andi Kleen6a460792009-09-16 11:50:15 +0200872 /*
873 * This check implies we don't kill processes if their pages
874 * are in the swap cache early. Those are always late kills.
875 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900876 if (!page_mapped(hpage))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100877 return SWAP_SUCCESS;
878
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900879 if (PageKsm(p))
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100880 return SWAP_FAIL;
Andi Kleen6a460792009-09-16 11:50:15 +0200881
882 if (PageSwapCache(p)) {
883 printk(KERN_ERR
884 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
885 ttu |= TTU_IGNORE_HWPOISON;
886 }
887
888 /*
889 * Propagate the dirty bit from PTEs to struct page first, because we
890 * need this to decide if we should kill or just drop the page.
Wu Fengguangdb0480b2009-12-16 12:19:58 +0100891 * XXX: the dirty test could be racy: set_page_dirty() may not always
892 * be called inside page lock (it's recommended but not enforced).
Andi Kleen6a460792009-09-16 11:50:15 +0200893 */
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900894 mapping = page_mapping(hpage);
Tony Luck6751ed62012-07-11 10:20:47 -0700895 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +0900896 mapping_cap_writeback_dirty(mapping)) {
897 if (page_mkclean(hpage)) {
898 SetPageDirty(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +0200899 } else {
900 kill = 0;
901 ttu |= TTU_IGNORE_HWPOISON;
902 printk(KERN_INFO
903 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
904 pfn);
905 }
906 }
907
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800908 /*
909 * ppage: poisoned page
910 * if p is regular page(4k page)
911 * ppage == real poisoned page;
912 * else p is hugetlb or THP, ppage == head page.
913 */
914 ppage = hpage;
915
Jin Dongmingefeda7a2011-02-01 15:52:39 -0800916 if (PageTransHuge(hpage)) {
917 /*
918 * Verify that this isn't a hugetlbfs head page, the check for
919 * PageAnon is just for avoid tripping a split_huge_page
920 * internal debug check, as split_huge_page refuses to deal with
921 * anything that isn't an anon page. PageAnon can't go away fro
922 * under us because we hold a refcount on the hpage, without a
923 * refcount on the hpage. split_huge_page can't be safely called
924 * in the first place, having a refcount on the tail isn't
925 * enough * to be safe.
926 */
927 if (!PageHuge(hpage) && PageAnon(hpage)) {
928 if (unlikely(split_huge_page(hpage))) {
929 /*
930 * FIXME: if splitting THP is failed, it is
931 * better to stop the following operation rather
932 * than causing panic by unmapping. System might
933 * survive if the page is freed later.
934 */
935 printk(KERN_INFO
936 "MCE %#lx: failed to split THP\n", pfn);
937
938 BUG_ON(!PageHWPoison(p));
939 return SWAP_FAIL;
940 }
Naoya Horiguchia3e0f9e2014-01-02 12:58:51 -0800941 /*
942 * We pinned the head page for hwpoison handling,
943 * now we split the thp and we are interested in
944 * the hwpoisoned raw page, so move the refcount
945 * to it.
946 */
947 if (hpage != p) {
948 put_page(hpage);
949 get_page(p);
950 }
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800951 /* THP is split, so ppage should be the real poisoned page. */
952 ppage = p;
Jin Dongmingefeda7a2011-02-01 15:52:39 -0800953 }
954 }
955
Andi Kleen6a460792009-09-16 11:50:15 +0200956 /*
957 * First collect all the processes that have the page
958 * mapped in dirty form. This has to be done before try_to_unmap,
959 * because ttu takes the rmap data structures down.
960 *
961 * Error handling: We ignore errors here because
962 * there's nothing that can be done.
963 */
964 if (kill)
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800965 collect_procs(ppage, &tokill);
Andi Kleen6a460792009-09-16 11:50:15 +0200966
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800967 if (hpage != ppage)
Jens Axboe7eaceac2011-03-10 08:52:07 +0100968 lock_page(ppage);
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800969
970 ret = try_to_unmap(ppage, ttu);
Andi Kleen6a460792009-09-16 11:50:15 +0200971 if (ret != SWAP_SUCCESS)
972 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
Jin Dongminga6d30dd2011-02-01 15:52:40 -0800973 pfn, page_mapcount(ppage));
974
975 if (hpage != ppage)
976 unlock_page(ppage);
Andi Kleen6a460792009-09-16 11:50:15 +0200977
978 /*
979 * Now that the dirty bit has been propagated to the
980 * struct page and all unmaps done we can decide if
981 * killing is needed or not. Only kill when the page
Tony Luck6751ed62012-07-11 10:20:47 -0700982 * was dirty or the process is not restartable,
983 * otherwise the tokill list is merely
Andi Kleen6a460792009-09-16 11:50:15 +0200984 * freed. When there was a problem unmapping earlier
985 * use a more force-full uncatchable kill to prevent
986 * any accesses to the poisoned memory.
987 */
Tony Luck6751ed62012-07-11 10:20:47 -0700988 forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL);
989 kill_procs(&tokill, forcekill, trapno,
Tony Luck7329bbe2011-12-13 09:27:58 -0800990 ret != SWAP_SUCCESS, p, pfn, flags);
Wu Fengguang1668bfd2009-12-16 12:19:58 +0100991
992 return ret;
Andi Kleen6a460792009-09-16 11:50:15 +0200993}
994
Naoya Horiguchi7013feb2010-05-28 09:29:18 +0900995static void set_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 SetPageHWPoison(hpage + i);
1001}
1002
1003static void clear_page_hwpoison_huge_page(struct page *hpage)
1004{
1005 int i;
Wanpeng Lif9121152013-09-11 14:22:52 -07001006 int nr_pages = 1 << compound_order(hpage);
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001007 for (i = 0; i < nr_pages; i++)
1008 ClearPageHWPoison(hpage + i);
1009}
1010
Tony Luckcd42f4a2011-12-15 10:48:12 -08001011/**
1012 * memory_failure - Handle memory failure of a page.
1013 * @pfn: Page Number of the corrupted page
1014 * @trapno: Trap number reported in the signal to user space.
1015 * @flags: fine tune action taken
1016 *
1017 * This function is called by the low level machine check code
1018 * of an architecture when it detects hardware memory corruption
1019 * of a page. It tries its best to recover, which includes
1020 * dropping pages, killing processes etc.
1021 *
1022 * The function is primarily of use for corruptions that
1023 * happen outside the current execution context (e.g. when
1024 * detected by a background scrubber)
1025 *
1026 * Must run in process context (e.g. a work queue) with interrupts
1027 * enabled and no spinlocks hold.
1028 */
1029int memory_failure(unsigned long pfn, int trapno, int flags)
Andi Kleen6a460792009-09-16 11:50:15 +02001030{
1031 struct page_state *ps;
1032 struct page *p;
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001033 struct page *hpage;
Andi Kleen6a460792009-09-16 11:50:15 +02001034 int res;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001035 unsigned int nr_pages;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001036 unsigned long page_flags;
Andi Kleen6a460792009-09-16 11:50:15 +02001037
1038 if (!sysctl_memory_failure_recovery)
1039 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1040
1041 if (!pfn_valid(pfn)) {
Wu Fengguanga7560fc2009-12-16 12:19:57 +01001042 printk(KERN_ERR
1043 "MCE %#lx: memory outside kernel control\n",
1044 pfn);
1045 return -ENXIO;
Andi Kleen6a460792009-09-16 11:50:15 +02001046 }
1047
1048 p = pfn_to_page(pfn);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001049 hpage = compound_head(p);
Andi Kleen6a460792009-09-16 11:50:15 +02001050 if (TestSetPageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001051 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001052 return 0;
1053 }
1054
Naoya Horiguchi4db0e952013-02-22 16:34:05 -08001055 /*
1056 * Currently errors on hugetlbfs pages are measured in hugepage units,
1057 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1058 * transparent hugepages, they are supposed to be split and error
1059 * measurement is done in normal page units. So nr_pages should be one
1060 * in this case.
1061 */
1062 if (PageHuge(p))
1063 nr_pages = 1 << compound_order(hpage);
1064 else /* normal page or thp */
1065 nr_pages = 1;
Xishi Qiu293c07e2013-02-22 16:34:02 -08001066 atomic_long_add(nr_pages, &num_poisoned_pages);
Andi Kleen6a460792009-09-16 11:50:15 +02001067
1068 /*
1069 * We need/can do nothing about count=0 pages.
1070 * 1) it's a free page, and therefore in safe hand:
1071 * prep_new_page() will be the gate keeper.
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001072 * 2) it's a free hugepage, which is also safe:
1073 * an affected hugepage will be dequeued from hugepage freelist,
1074 * so there's no concern about reusing it ever after.
1075 * 3) it's part of a non-compound high order page.
Andi Kleen6a460792009-09-16 11:50:15 +02001076 * Implies some kernel user: cannot stop them from
1077 * R/W the page; let's pray that the page has been
1078 * used and will be freed some time later.
1079 * In fact it's dangerous to directly bump up page count from 0,
1080 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1081 */
Andi Kleen82ba0112009-12-16 12:19:57 +01001082 if (!(flags & MF_COUNT_INCREASED) &&
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001083 !get_page_unless_zero(hpage)) {
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001084 if (is_free_buddy_page(p)) {
1085 action_result(pfn, "free buddy", DELAYED);
1086 return 0;
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001087 } else if (PageHuge(hpage)) {
1088 /*
1089 * Check "just unpoisoned", "filter hit", and
1090 * "race with other subpage."
1091 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001092 lock_page(hpage);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001093 if (!PageHWPoison(hpage)
1094 || (hwpoison_filter(p) && TestClearPageHWPoison(p))
1095 || (p != hpage && TestSetPageHWPoison(hpage))) {
Xishi Qiu293c07e2013-02-22 16:34:02 -08001096 atomic_long_sub(nr_pages, &num_poisoned_pages);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001097 return 0;
1098 }
1099 set_page_hwpoison_huge_page(hpage);
1100 res = dequeue_hwpoisoned_huge_page(hpage);
1101 action_result(pfn, "free huge",
1102 res ? IGNORED : DELAYED);
1103 unlock_page(hpage);
1104 return res;
Wu Fengguang8d22ba12009-12-16 12:19:58 +01001105 } else {
1106 action_result(pfn, "high order kernel", IGNORED);
1107 return -EBUSY;
1108 }
Andi Kleen6a460792009-09-16 11:50:15 +02001109 }
1110
1111 /*
Wu Fengguange43c3af2009-09-29 13:16:20 +08001112 * We ignore non-LRU pages for good reasons.
1113 * - PG_locked is only well defined for LRU pages and a few others
1114 * - to avoid races with __set_page_locked()
1115 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1116 * The check (unnecessarily) ignores LRU pages being isolated and
1117 * walked by the page reclaim code, however that's not a big loss.
1118 */
Dean Nelson385de352012-03-21 16:34:05 -07001119 if (!PageHuge(p) && !PageTransTail(p)) {
Jin Dongmingaf241a02011-02-01 15:52:41 -08001120 if (!PageLRU(p))
1121 shake_page(p, 0);
1122 if (!PageLRU(p)) {
1123 /*
1124 * shake_page could have turned it free.
1125 */
1126 if (is_free_buddy_page(p)) {
Wanpeng Li2d421ac2013-09-30 13:45:23 -07001127 if (flags & MF_COUNT_INCREASED)
1128 action_result(pfn, "free buddy", DELAYED);
1129 else
1130 action_result(pfn, "free buddy, 2nd try", DELAYED);
Jin Dongmingaf241a02011-02-01 15:52:41 -08001131 return 0;
1132 }
1133 action_result(pfn, "non LRU", IGNORED);
1134 put_page(p);
1135 return -EBUSY;
Andi Kleen0474a602009-12-16 12:20:00 +01001136 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001137 }
Wu Fengguange43c3af2009-09-29 13:16:20 +08001138
1139 /*
Andi Kleen6a460792009-09-16 11:50:15 +02001140 * Lock the page and wait for writeback to finish.
1141 * It's very difficult to mess with pages currently under IO
1142 * and in many cases impossible, so we just avoid it here.
1143 */
Jens Axboe7eaceac2011-03-10 08:52:07 +01001144 lock_page(hpage);
Wu Fengguang847ce402009-12-16 12:19:58 +01001145
1146 /*
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001147 * We use page flags to determine what action should be taken, but
1148 * the flags can be modified by the error containment action. One
1149 * example is an mlocked page, where PG_mlocked is cleared by
1150 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1151 * correctly, we save a copy of the page flags at this time.
1152 */
1153 page_flags = p->flags;
1154
1155 /*
Wu Fengguang847ce402009-12-16 12:19:58 +01001156 * unpoison always clear PG_hwpoison inside page lock
1157 */
1158 if (!PageHWPoison(p)) {
Wu Fengguangd95ea512009-12-16 12:19:58 +01001159 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001160 res = 0;
1161 goto out;
1162 }
Wu Fengguang7c116f22009-12-16 12:19:59 +01001163 if (hwpoison_filter(p)) {
1164 if (TestClearPageHWPoison(p))
Xishi Qiu293c07e2013-02-22 16:34:02 -08001165 atomic_long_sub(nr_pages, &num_poisoned_pages);
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001166 unlock_page(hpage);
1167 put_page(hpage);
Wu Fengguang7c116f22009-12-16 12:19:59 +01001168 return 0;
1169 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001170
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001171 /*
1172 * For error on the tail page, we should set PG_hwpoison
1173 * on the head page to show that the hugepage is hwpoisoned
1174 */
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001175 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
Naoya Horiguchi7013feb2010-05-28 09:29:18 +09001176 action_result(pfn, "hugepage already hardware poisoned",
1177 IGNORED);
1178 unlock_page(hpage);
1179 put_page(hpage);
1180 return 0;
1181 }
1182 /*
1183 * Set PG_hwpoison on all pages in an error hugepage,
1184 * because containment is done in hugepage unit for now.
1185 * Since we have done TestSetPageHWPoison() for the head page with
1186 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1187 */
1188 if (PageHuge(p))
1189 set_page_hwpoison_huge_page(hpage);
1190
Andi Kleen6a460792009-09-16 11:50:15 +02001191 wait_on_page_writeback(p);
1192
1193 /*
1194 * Now take care of user space mappings.
Minchan Kime64a7822011-03-22 16:32:44 -07001195 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
Andi Kleen6a460792009-09-16 11:50:15 +02001196 */
Tony Luck7329bbe2011-12-13 09:27:58 -08001197 if (hwpoison_user_mappings(p, pfn, trapno, flags) != SWAP_SUCCESS) {
Wu Fengguang1668bfd2009-12-16 12:19:58 +01001198 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1199 res = -EBUSY;
1200 goto out;
1201 }
Andi Kleen6a460792009-09-16 11:50:15 +02001202
1203 /*
1204 * Torn down by someone else?
1205 */
Wu Fengguangdc2a1cb2009-12-16 12:19:58 +01001206 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
Andi Kleen6a460792009-09-16 11:50:15 +02001207 action_result(pfn, "already truncated LRU", IGNORED);
Wu Fengguangd95ea512009-12-16 12:19:58 +01001208 res = -EBUSY;
Andi Kleen6a460792009-09-16 11:50:15 +02001209 goto out;
1210 }
1211
1212 res = -EBUSY;
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001213 /*
1214 * The first check uses the current page flags which may not have any
1215 * relevant information. The second check with the saved page flagss is
1216 * carried out only if the first check can't determine the page status.
1217 */
1218 for (ps = error_states;; ps++)
1219 if ((p->flags & ps->mask) == ps->res)
Andi Kleen6a460792009-09-16 11:50:15 +02001220 break;
Wanpeng Li841fcc52013-09-11 14:22:50 -07001221
1222 page_flags |= (p->flags & (1UL << PG_dirty));
1223
Naoya Horiguchi524fca12013-02-22 16:35:51 -08001224 if (!ps->mask)
1225 for (ps = error_states;; ps++)
1226 if ((page_flags & ps->mask) == ps->res)
1227 break;
1228 res = page_action(ps, p, pfn);
Andi Kleen6a460792009-09-16 11:50:15 +02001229out:
Naoya Horiguchi7af446a2010-05-28 09:29:17 +09001230 unlock_page(hpage);
Andi Kleen6a460792009-09-16 11:50:15 +02001231 return res;
1232}
Tony Luckcd42f4a2011-12-15 10:48:12 -08001233EXPORT_SYMBOL_GPL(memory_failure);
Wu Fengguang847ce402009-12-16 12:19:58 +01001234
Huang Yingea8f5fb2011-07-13 13:14:27 +08001235#define MEMORY_FAILURE_FIFO_ORDER 4
1236#define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1237
1238struct memory_failure_entry {
1239 unsigned long pfn;
1240 int trapno;
1241 int flags;
1242};
1243
1244struct memory_failure_cpu {
1245 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1246 MEMORY_FAILURE_FIFO_SIZE);
1247 spinlock_t lock;
1248 struct work_struct work;
1249};
1250
1251static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1252
1253/**
1254 * memory_failure_queue - Schedule handling memory failure of a page.
1255 * @pfn: Page Number of the corrupted page
1256 * @trapno: Trap number reported in the signal to user space.
1257 * @flags: Flags for memory failure handling
1258 *
1259 * This function is called by the low level hardware error handler
1260 * when it detects hardware memory corruption of a page. It schedules
1261 * the recovering of error page, including dropping pages, killing
1262 * processes etc.
1263 *
1264 * The function is primarily of use for corruptions that
1265 * happen outside the current execution context (e.g. when
1266 * detected by a background scrubber)
1267 *
1268 * Can run in IRQ context.
1269 */
1270void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1271{
1272 struct memory_failure_cpu *mf_cpu;
1273 unsigned long proc_flags;
1274 struct memory_failure_entry entry = {
1275 .pfn = pfn,
1276 .trapno = trapno,
1277 .flags = flags,
1278 };
1279
1280 mf_cpu = &get_cpu_var(memory_failure_cpu);
1281 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
Stefani Seibold498d3192013-11-14 14:32:17 -08001282 if (kfifo_put(&mf_cpu->fifo, entry))
Huang Yingea8f5fb2011-07-13 13:14:27 +08001283 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1284 else
Joe Perches8e33a522013-07-25 11:53:25 -07001285 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
Huang Yingea8f5fb2011-07-13 13:14:27 +08001286 pfn);
1287 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1288 put_cpu_var(memory_failure_cpu);
1289}
1290EXPORT_SYMBOL_GPL(memory_failure_queue);
1291
1292static void memory_failure_work_func(struct work_struct *work)
1293{
1294 struct memory_failure_cpu *mf_cpu;
1295 struct memory_failure_entry entry = { 0, };
1296 unsigned long proc_flags;
1297 int gotten;
1298
1299 mf_cpu = &__get_cpu_var(memory_failure_cpu);
1300 for (;;) {
1301 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1302 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1303 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1304 if (!gotten)
1305 break;
Naveen N. Raocf870c72013-07-10 14:57:01 +05301306 if (entry.flags & MF_SOFT_OFFLINE)
1307 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1308 else
1309 memory_failure(entry.pfn, entry.trapno, entry.flags);
Huang Yingea8f5fb2011-07-13 13:14:27 +08001310 }
1311}
1312
1313static int __init memory_failure_init(void)
1314{
1315 struct memory_failure_cpu *mf_cpu;
1316 int cpu;
1317
1318 for_each_possible_cpu(cpu) {
1319 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1320 spin_lock_init(&mf_cpu->lock);
1321 INIT_KFIFO(mf_cpu->fifo);
1322 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1323 }
1324
1325 return 0;
1326}
1327core_initcall(memory_failure_init);
1328
Wu Fengguang847ce402009-12-16 12:19:58 +01001329/**
1330 * unpoison_memory - Unpoison a previously poisoned page
1331 * @pfn: Page number of the to be unpoisoned page
1332 *
1333 * Software-unpoison a page that has been poisoned by
1334 * memory_failure() earlier.
1335 *
1336 * This is only done on the software-level, so it only works
1337 * for linux injected failures, not real hardware failures
1338 *
1339 * Returns 0 for success, otherwise -errno.
1340 */
1341int unpoison_memory(unsigned long pfn)
1342{
1343 struct page *page;
1344 struct page *p;
1345 int freeit = 0;
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001346 unsigned int nr_pages;
Wu Fengguang847ce402009-12-16 12:19:58 +01001347
1348 if (!pfn_valid(pfn))
1349 return -ENXIO;
1350
1351 p = pfn_to_page(pfn);
1352 page = compound_head(p);
1353
1354 if (!PageHWPoison(p)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001355 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001356 return 0;
1357 }
1358
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001359 /*
1360 * unpoison_memory() can encounter thp only when the thp is being
1361 * worked by memory_failure() and the page lock is not held yet.
1362 * In such case, we yield to memory_failure() and make unpoison fail.
1363 */
Wanpeng Lie76d30e2013-09-30 13:45:22 -07001364 if (!PageHuge(page) && PageTransHuge(page)) {
Wanpeng Li0cea3fd2013-09-11 14:22:53 -07001365 pr_info("MCE: Memory failure is now running on %#lx\n", pfn);
1366 return 0;
1367 }
1368
Wanpeng Lif9121152013-09-11 14:22:52 -07001369 nr_pages = 1 << compound_order(page);
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001370
Wu Fengguang847ce402009-12-16 12:19:58 +01001371 if (!get_page_unless_zero(page)) {
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001372 /*
1373 * Since HWPoisoned hugepage should have non-zero refcount,
1374 * race between memory failure and unpoison seems to happen.
1375 * In such case unpoison fails and memory failure runs
1376 * to the end.
1377 */
1378 if (PageHuge(page)) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001379 pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09001380 return 0;
1381 }
Wu Fengguang847ce402009-12-16 12:19:58 +01001382 if (TestClearPageHWPoison(p))
Wanpeng Lidd9538a2013-09-11 14:22:54 -07001383 atomic_long_dec(&num_poisoned_pages);
Andi Kleenfb46e732010-09-27 23:31:30 +02001384 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
Wu Fengguang847ce402009-12-16 12:19:58 +01001385 return 0;
1386 }
1387
Jens Axboe7eaceac2011-03-10 08:52:07 +01001388 lock_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001389 /*
1390 * This test is racy because PG_hwpoison is set outside of page lock.
1391 * That's acceptable because that won't trigger kernel panic. Instead,
1392 * the PG_hwpoison page will be caught and isolated on the entrance to
1393 * the free buddy page pool.
1394 */
Naoya Horiguchic9fbdd52010-05-28 09:29:19 +09001395 if (TestClearPageHWPoison(page)) {
Andi Kleenfb46e732010-09-27 23:31:30 +02001396 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
Xishi Qiu293c07e2013-02-22 16:34:02 -08001397 atomic_long_sub(nr_pages, &num_poisoned_pages);
Wu Fengguang847ce402009-12-16 12:19:58 +01001398 freeit = 1;
Naoya Horiguchi6a901812010-09-08 10:19:40 +09001399 if (PageHuge(page))
1400 clear_page_hwpoison_huge_page(page);
Wu Fengguang847ce402009-12-16 12:19:58 +01001401 }
1402 unlock_page(page);
1403
1404 put_page(page);
Wanpeng Li3ba5eeb2013-09-11 14:23:01 -07001405 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
Wu Fengguang847ce402009-12-16 12:19:58 +01001406 put_page(page);
1407
1408 return 0;
1409}
1410EXPORT_SYMBOL(unpoison_memory);
Andi Kleenfacb6012009-12-16 12:20:00 +01001411
1412static struct page *new_page(struct page *p, unsigned long private, int **x)
1413{
Andi Kleen12686d12009-12-16 12:20:01 +01001414 int nid = page_to_nid(p);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001415 if (PageHuge(p))
1416 return alloc_huge_page_node(page_hstate(compound_head(p)),
1417 nid);
1418 else
1419 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +01001420}
1421
1422/*
1423 * Safely get reference count of an arbitrary page.
1424 * Returns 0 for a free page, -EIO for a zero refcount page
1425 * that is not free, and 1 for any other page type.
1426 * For 1 the page is returned with increased page count, otherwise not.
1427 */
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001428static int __get_any_page(struct page *p, unsigned long pfn, int flags)
Andi Kleenfacb6012009-12-16 12:20:00 +01001429{
1430 int ret;
1431
1432 if (flags & MF_COUNT_INCREASED)
1433 return 1;
1434
1435 /*
Naoya Horiguchid950b952010-09-08 10:19:39 +09001436 * When the target page is a free hugepage, just remove it
1437 * from free hugepage list.
1438 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001439 if (!get_page_unless_zero(compound_head(p))) {
Naoya Horiguchid950b952010-09-08 10:19:39 +09001440 if (PageHuge(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001441 pr_info("%s: %#lx free huge page\n", __func__, pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001442 ret = 0;
Naoya Horiguchid950b952010-09-08 10:19:39 +09001443 } else if (is_free_buddy_page(p)) {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001444 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
Andi Kleenfacb6012009-12-16 12:20:00 +01001445 ret = 0;
1446 } else {
Borislav Petkov71dd0b82012-05-29 15:06:16 -07001447 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1448 __func__, pfn, p->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001449 ret = -EIO;
1450 }
1451 } else {
1452 /* Not a free page */
1453 ret = 1;
1454 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001455 return ret;
1456}
1457
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001458static int get_any_page(struct page *page, unsigned long pfn, int flags)
1459{
1460 int ret = __get_any_page(page, pfn, flags);
1461
1462 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1463 /*
1464 * Try to free it.
1465 */
1466 put_page(page);
1467 shake_page(page, 1);
1468
1469 /*
1470 * Did it turn free?
1471 */
1472 ret = __get_any_page(page, pfn, 0);
1473 if (!PageLRU(page)) {
1474 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1475 pfn, page->flags);
1476 return -EIO;
1477 }
1478 }
1479 return ret;
1480}
1481
Naoya Horiguchid950b952010-09-08 10:19:39 +09001482static int soft_offline_huge_page(struct page *page, int flags)
1483{
1484 int ret;
1485 unsigned long pfn = page_to_pfn(page);
1486 struct page *hpage = compound_head(page);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001487 LIST_HEAD(pagelist);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001488
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001489 /*
1490 * This double-check of PageHWPoison is to avoid the race with
1491 * memory_failure(). See also comment in __soft_offline_page().
1492 */
1493 lock_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001494 if (PageHWPoison(hpage)) {
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001495 unlock_page(hpage);
1496 put_page(hpage);
Xishi Qiu0ebff322013-02-22 16:33:59 -08001497 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001498 return -EBUSY;
Xishi Qiu0ebff322013-02-22 16:33:59 -08001499 }
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001500 unlock_page(hpage);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001501
Naoya Horiguchid950b952010-09-08 10:19:39 +09001502 /* Keep page count to indicate a given hugepage is isolated. */
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001503 list_move(&hpage->lru, &pagelist);
1504 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1505 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Naoya Horiguchid950b952010-09-08 10:19:39 +09001506 if (ret) {
Dean Nelsondd73e852011-10-31 17:09:04 -07001507 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1508 pfn, ret, page->flags);
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001509 /*
1510 * We know that soft_offline_huge_page() tries to migrate
1511 * only one hugepage pointed to by hpage, so we need not
1512 * run through the pagelist here.
1513 */
1514 putback_active_hugepage(hpage);
1515 if (ret > 0)
1516 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001517 } else {
Jianguo Wua49ecbc2013-12-18 17:08:54 -08001518 /* overcommit hugetlb page will be freed to buddy */
1519 if (PageHuge(page)) {
1520 set_page_hwpoison_huge_page(hpage);
1521 dequeue_hwpoisoned_huge_page(hpage);
1522 atomic_long_add(1 << compound_order(hpage),
1523 &num_poisoned_pages);
1524 } else {
1525 SetPageHWPoison(page);
1526 atomic_long_inc(&num_poisoned_pages);
1527 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001528 }
Naoya Horiguchid950b952010-09-08 10:19:39 +09001529 return ret;
1530}
1531
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001532static int __soft_offline_page(struct page *page, int flags)
1533{
1534 int ret;
1535 unsigned long pfn = page_to_pfn(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001536
1537 /*
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001538 * Check PageHWPoison again inside page lock because PageHWPoison
1539 * is set by memory_failure() outside page lock. Note that
1540 * memory_failure() also double-checks PageHWPoison inside page lock,
1541 * so there's no race between soft_offline_page() and memory_failure().
Andi Kleenfacb6012009-12-16 12:20:00 +01001542 */
Xishi Qiu0ebff322013-02-22 16:33:59 -08001543 lock_page(page);
1544 wait_on_page_writeback(page);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001545 if (PageHWPoison(page)) {
1546 unlock_page(page);
1547 put_page(page);
1548 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1549 return -EBUSY;
1550 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001551 /*
1552 * Try to invalidate first. This should work for
1553 * non dirty unmapped page cache pages.
1554 */
1555 ret = invalidate_inode_page(page);
1556 unlock_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001557 /*
Andi Kleenfacb6012009-12-16 12:20:00 +01001558 * RED-PEN would be better to keep it isolated here, but we
1559 * would need to fix isolation locking first.
1560 */
Andi Kleenfacb6012009-12-16 12:20:00 +01001561 if (ret == 1) {
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001562 put_page(page);
Andi Kleenfb46e732010-09-27 23:31:30 +02001563 pr_info("soft_offline: %#lx: invalidated\n", pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001564 SetPageHWPoison(page);
1565 atomic_long_inc(&num_poisoned_pages);
1566 return 0;
Andi Kleenfacb6012009-12-16 12:20:00 +01001567 }
1568
1569 /*
1570 * Simple invalidation didn't work.
1571 * Try to migrate to a new page instead. migrate.c
1572 * handles a large number of cases for us.
1573 */
1574 ret = isolate_lru_page(page);
Konstantin Khlebnikovbd486282011-05-24 17:12:20 -07001575 /*
1576 * Drop page reference which is came from get_any_page()
1577 * successful isolate_lru_page() already took another one.
1578 */
1579 put_page(page);
Andi Kleenfacb6012009-12-16 12:20:00 +01001580 if (!ret) {
1581 LIST_HEAD(pagelist);
Minchan Kim5db8a732011-06-15 15:08:48 -07001582 inc_zone_page_state(page, NR_ISOLATED_ANON +
Hugh Dickins9c620e22013-02-22 16:35:14 -08001583 page_is_file_cache(page));
Andi Kleenfacb6012009-12-16 12:20:00 +01001584 list_add(&page->lru, &pagelist);
Mel Gorman77f1fe62011-01-13 15:45:57 -08001585 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001586 MIGRATE_SYNC, MR_MEMORY_FAILURE);
Andi Kleenfacb6012009-12-16 12:20:00 +01001587 if (ret) {
Andrea Arcangeli57fc4a52011-02-01 15:52:32 -08001588 putback_lru_pages(&pagelist);
Andi Kleenfb46e732010-09-27 23:31:30 +02001589 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
Andi Kleenfacb6012009-12-16 12:20:00 +01001590 pfn, ret, page->flags);
1591 if (ret > 0)
1592 ret = -EIO;
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001593 } else {
Naoya Horiguchif15bdfa2013-07-03 15:02:37 -07001594 /*
1595 * After page migration succeeds, the source page can
1596 * be trapped in pagevec and actual freeing is delayed.
1597 * Freeing code works differently based on PG_hwpoison,
1598 * so there's a race. We need to make sure that the
1599 * source page should be freed back to buddy before
1600 * setting PG_hwpoison.
1601 */
1602 if (!is_free_buddy_page(page))
1603 lru_add_drain_all();
1604 if (!is_free_buddy_page(page))
1605 drain_all_pages();
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001606 SetPageHWPoison(page);
Naoya Horiguchif15bdfa2013-07-03 15:02:37 -07001607 if (!is_free_buddy_page(page))
1608 pr_info("soft offline: %#lx: page leaked\n",
1609 pfn);
Naoya Horiguchiaf8fae72013-02-22 16:34:03 -08001610 atomic_long_inc(&num_poisoned_pages);
Andi Kleenfacb6012009-12-16 12:20:00 +01001611 }
1612 } else {
Andi Kleenfb46e732010-09-27 23:31:30 +02001613 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
Dean Nelsondd73e852011-10-31 17:09:04 -07001614 pfn, ret, page_count(page), page->flags);
Andi Kleenfacb6012009-12-16 12:20:00 +01001615 }
Andi Kleenfacb6012009-12-16 12:20:00 +01001616 return ret;
1617}
Wanpeng Li86e05772013-09-11 14:22:56 -07001618
1619/**
1620 * soft_offline_page - Soft offline a page.
1621 * @page: page to offline
1622 * @flags: flags. Same as memory_failure().
1623 *
1624 * Returns 0 on success, otherwise negated errno.
1625 *
1626 * Soft offline a page, by migration or invalidation,
1627 * without killing anything. This is for the case when
1628 * a page is not corrupted yet (so it's still valid to access),
1629 * but has had a number of corrected errors and is better taken
1630 * out.
1631 *
1632 * The actual policy on when to do that is maintained by
1633 * user space.
1634 *
1635 * This should never impact any application or cause data loss,
1636 * however it might take some time.
1637 *
1638 * This is not a 100% solution for all memory, but tries to be
1639 * ``good enough'' for the majority of memory.
1640 */
1641int soft_offline_page(struct page *page, int flags)
1642{
1643 int ret;
1644 unsigned long pfn = page_to_pfn(page);
1645 struct page *hpage = compound_trans_head(page);
1646
1647 if (PageHWPoison(page)) {
1648 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1649 return -EBUSY;
1650 }
1651 if (!PageHuge(page) && PageTransHuge(hpage)) {
1652 if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1653 pr_info("soft offline: %#lx: failed to split THP\n",
1654 pfn);
1655 return -EBUSY;
1656 }
1657 }
1658
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001659 /*
1660 * The lock_memory_hotplug prevents a race with memory hotplug.
1661 * This is a big hammer, a better would be nicer.
1662 */
1663 lock_memory_hotplug();
1664
1665 /*
1666 * Isolate the page, so that it doesn't get reallocated if it
1667 * was free. This flag should be kept set until the source page
1668 * is freed and PG_hwpoison on it is set.
1669 */
1670 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
1671 set_migratetype_isolate(page, true);
1672
Wanpeng Li86e05772013-09-11 14:22:56 -07001673 ret = get_any_page(page, pfn, flags);
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001674 unlock_memory_hotplug();
1675 if (ret > 0) { /* for in-use pages */
Wanpeng Li86e05772013-09-11 14:22:56 -07001676 if (PageHuge(page))
1677 ret = soft_offline_huge_page(page, flags);
1678 else
1679 ret = __soft_offline_page(page, flags);
Naoya Horiguchi03b61ff2013-11-12 15:07:26 -08001680 } else if (ret == 0) { /* for free pages */
Wanpeng Li86e05772013-09-11 14:22:56 -07001681 if (PageHuge(page)) {
1682 set_page_hwpoison_huge_page(hpage);
1683 dequeue_hwpoisoned_huge_page(hpage);
1684 atomic_long_add(1 << compound_order(hpage),
1685 &num_poisoned_pages);
1686 } else {
1687 SetPageHWPoison(page);
1688 atomic_long_inc(&num_poisoned_pages);
1689 }
1690 }
Wanpeng Li86e05772013-09-11 14:22:56 -07001691 unset_migratetype_isolate(page, MIGRATE_MOVABLE);
1692 return ret;
1693}