Commit | Line | Data |
---|---|---|
6a46079c AK |
1 | /* |
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 | |
10 | * hardware as being corrupted usually due to a 2bit ECC memory or cache | |
11 | * failure. | |
12 | * | |
13 | * Handles page cache pages in various states. The tricky part | |
14 | * here is that we can access any page asynchronous to other VM | |
15 | * users, because memory failures could happen anytime and anywhere, | |
16 | * possibly violating some of their assumptions. This is why this code | |
17 | * has to be extremely careful. Generally it tries to use normal locking | |
18 | * rules, as in get the standard locks, even if that means the | |
19 | * error handling takes potentially a long time. | |
20 | * | |
21 | * The operation to map back from RMAP chains to processes has to walk | |
22 | * the complete process list and has non linear complexity with the number | |
23 | * mappings. In short it can be quite slow. But since memory corruptions | |
24 | * are rare we hope to get away with this. | |
25 | */ | |
26 | ||
27 | /* | |
28 | * Notebook: | |
29 | * - hugetlb needs more code | |
30 | * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages | |
31 | * - pass bad pages to kdump next kernel | |
32 | */ | |
33 | #define DEBUG 1 /* remove me in 2.6.34 */ | |
34 | #include <linux/kernel.h> | |
35 | #include <linux/mm.h> | |
36 | #include <linux/page-flags.h> | |
478c5ffc | 37 | #include <linux/kernel-page-flags.h> |
6a46079c | 38 | #include <linux/sched.h> |
01e00f88 | 39 | #include <linux/ksm.h> |
6a46079c AK |
40 | #include <linux/rmap.h> |
41 | #include <linux/pagemap.h> | |
42 | #include <linux/swap.h> | |
43 | #include <linux/backing-dev.h> | |
facb6011 AK |
44 | #include <linux/migrate.h> |
45 | #include <linux/page-isolation.h> | |
46 | #include <linux/suspend.h> | |
6a46079c AK |
47 | #include "internal.h" |
48 | ||
49 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
50 | ||
51 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
52 | ||
53 | atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0); | |
54 | ||
27df5068 AK |
55 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
56 | ||
1bfe5feb | 57 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
58 | u32 hwpoison_filter_dev_major = ~0U; |
59 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
60 | u64 hwpoison_filter_flags_mask; |
61 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 62 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
63 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
64 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
65 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
66 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
67 | |
68 | static int hwpoison_filter_dev(struct page *p) | |
69 | { | |
70 | struct address_space *mapping; | |
71 | dev_t dev; | |
72 | ||
73 | if (hwpoison_filter_dev_major == ~0U && | |
74 | hwpoison_filter_dev_minor == ~0U) | |
75 | return 0; | |
76 | ||
77 | /* | |
78 | * page_mapping() does not accept slab page | |
79 | */ | |
80 | if (PageSlab(p)) | |
81 | return -EINVAL; | |
82 | ||
83 | mapping = page_mapping(p); | |
84 | if (mapping == NULL || mapping->host == NULL) | |
85 | return -EINVAL; | |
86 | ||
87 | dev = mapping->host->i_sb->s_dev; | |
88 | if (hwpoison_filter_dev_major != ~0U && | |
89 | hwpoison_filter_dev_major != MAJOR(dev)) | |
90 | return -EINVAL; | |
91 | if (hwpoison_filter_dev_minor != ~0U && | |
92 | hwpoison_filter_dev_minor != MINOR(dev)) | |
93 | return -EINVAL; | |
94 | ||
95 | return 0; | |
96 | } | |
97 | ||
478c5ffc WF |
98 | static int hwpoison_filter_flags(struct page *p) |
99 | { | |
100 | if (!hwpoison_filter_flags_mask) | |
101 | return 0; | |
102 | ||
103 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
104 | hwpoison_filter_flags_value) | |
105 | return 0; | |
106 | else | |
107 | return -EINVAL; | |
108 | } | |
109 | ||
4fd466eb AK |
110 | /* |
111 | * This allows stress tests to limit test scope to a collection of tasks | |
112 | * by putting them under some memcg. This prevents killing unrelated/important | |
113 | * processes such as /sbin/init. Note that the target task may share clean | |
114 | * pages with init (eg. libc text), which is harmless. If the target task | |
115 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
116 | * is also included in the memcg. At last, due to race conditions this filter | |
117 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
118 | * a freed page. | |
119 | */ | |
120 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
121 | u64 hwpoison_filter_memcg; | |
122 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
123 | static int hwpoison_filter_task(struct page *p) | |
124 | { | |
125 | struct mem_cgroup *mem; | |
126 | struct cgroup_subsys_state *css; | |
127 | unsigned long ino; | |
128 | ||
129 | if (!hwpoison_filter_memcg) | |
130 | return 0; | |
131 | ||
132 | mem = try_get_mem_cgroup_from_page(p); | |
133 | if (!mem) | |
134 | return -EINVAL; | |
135 | ||
136 | css = mem_cgroup_css(mem); | |
137 | /* root_mem_cgroup has NULL dentries */ | |
138 | if (!css->cgroup->dentry) | |
139 | return -EINVAL; | |
140 | ||
141 | ino = css->cgroup->dentry->d_inode->i_ino; | |
142 | css_put(css); | |
143 | ||
144 | if (ino != hwpoison_filter_memcg) | |
145 | return -EINVAL; | |
146 | ||
147 | return 0; | |
148 | } | |
149 | #else | |
150 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
151 | #endif | |
152 | ||
7c116f2b WF |
153 | int hwpoison_filter(struct page *p) |
154 | { | |
1bfe5feb HL |
155 | if (!hwpoison_filter_enable) |
156 | return 0; | |
157 | ||
7c116f2b WF |
158 | if (hwpoison_filter_dev(p)) |
159 | return -EINVAL; | |
160 | ||
478c5ffc WF |
161 | if (hwpoison_filter_flags(p)) |
162 | return -EINVAL; | |
163 | ||
4fd466eb AK |
164 | if (hwpoison_filter_task(p)) |
165 | return -EINVAL; | |
166 | ||
7c116f2b WF |
167 | return 0; |
168 | } | |
27df5068 AK |
169 | #else |
170 | int hwpoison_filter(struct page *p) | |
171 | { | |
172 | return 0; | |
173 | } | |
174 | #endif | |
175 | ||
7c116f2b WF |
176 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
177 | ||
6a46079c AK |
178 | /* |
179 | * Send all the processes who have the page mapped an ``action optional'' | |
180 | * signal. | |
181 | */ | |
182 | static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno, | |
183 | unsigned long pfn) | |
184 | { | |
185 | struct siginfo si; | |
186 | int ret; | |
187 | ||
188 | printk(KERN_ERR | |
189 | "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n", | |
190 | pfn, t->comm, t->pid); | |
191 | si.si_signo = SIGBUS; | |
192 | si.si_errno = 0; | |
193 | si.si_code = BUS_MCEERR_AO; | |
194 | si.si_addr = (void *)addr; | |
195 | #ifdef __ARCH_SI_TRAPNO | |
196 | si.si_trapno = trapno; | |
197 | #endif | |
198 | si.si_addr_lsb = PAGE_SHIFT; | |
199 | /* | |
200 | * Don't use force here, it's convenient if the signal | |
201 | * can be temporarily blocked. | |
202 | * This could cause a loop when the user sets SIGBUS | |
203 | * to SIG_IGN, but hopefully noone will do that? | |
204 | */ | |
205 | ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */ | |
206 | if (ret < 0) | |
207 | printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n", | |
208 | t->comm, t->pid, ret); | |
209 | return ret; | |
210 | } | |
211 | ||
588f9ce6 AK |
212 | /* |
213 | * When a unknown page type is encountered drain as many buffers as possible | |
214 | * in the hope to turn the page into a LRU or free page, which we can handle. | |
215 | */ | |
facb6011 | 216 | void shake_page(struct page *p, int access) |
588f9ce6 AK |
217 | { |
218 | if (!PageSlab(p)) { | |
219 | lru_add_drain_all(); | |
220 | if (PageLRU(p)) | |
221 | return; | |
222 | drain_all_pages(); | |
223 | if (PageLRU(p) || is_free_buddy_page(p)) | |
224 | return; | |
225 | } | |
facb6011 | 226 | |
588f9ce6 | 227 | /* |
facb6011 AK |
228 | * Only all shrink_slab here (which would also |
229 | * shrink other caches) if access is not potentially fatal. | |
588f9ce6 | 230 | */ |
facb6011 AK |
231 | if (access) { |
232 | int nr; | |
233 | do { | |
234 | nr = shrink_slab(1000, GFP_KERNEL, 1000); | |
235 | if (page_count(p) == 0) | |
236 | break; | |
237 | } while (nr > 10); | |
238 | } | |
588f9ce6 AK |
239 | } |
240 | EXPORT_SYMBOL_GPL(shake_page); | |
241 | ||
6a46079c AK |
242 | /* |
243 | * Kill all processes that have a poisoned page mapped and then isolate | |
244 | * the page. | |
245 | * | |
246 | * General strategy: | |
247 | * Find all processes having the page mapped and kill them. | |
248 | * But we keep a page reference around so that the page is not | |
249 | * actually freed yet. | |
250 | * Then stash the page away | |
251 | * | |
252 | * There's no convenient way to get back to mapped processes | |
253 | * from the VMAs. So do a brute-force search over all | |
254 | * running processes. | |
255 | * | |
256 | * Remember that machine checks are not common (or rather | |
257 | * if they are common you have other problems), so this shouldn't | |
258 | * be a performance issue. | |
259 | * | |
260 | * Also there are some races possible while we get from the | |
261 | * error detection to actually handle it. | |
262 | */ | |
263 | ||
264 | struct to_kill { | |
265 | struct list_head nd; | |
266 | struct task_struct *tsk; | |
267 | unsigned long addr; | |
268 | unsigned addr_valid:1; | |
269 | }; | |
270 | ||
271 | /* | |
272 | * Failure handling: if we can't find or can't kill a process there's | |
273 | * not much we can do. We just print a message and ignore otherwise. | |
274 | */ | |
275 | ||
276 | /* | |
277 | * Schedule a process for later kill. | |
278 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
279 | * TBD would GFP_NOIO be enough? | |
280 | */ | |
281 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
282 | struct vm_area_struct *vma, | |
283 | struct list_head *to_kill, | |
284 | struct to_kill **tkc) | |
285 | { | |
286 | struct to_kill *tk; | |
287 | ||
288 | if (*tkc) { | |
289 | tk = *tkc; | |
290 | *tkc = NULL; | |
291 | } else { | |
292 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); | |
293 | if (!tk) { | |
294 | printk(KERN_ERR | |
295 | "MCE: Out of memory while machine check handling\n"); | |
296 | return; | |
297 | } | |
298 | } | |
299 | tk->addr = page_address_in_vma(p, vma); | |
300 | tk->addr_valid = 1; | |
301 | ||
302 | /* | |
303 | * In theory we don't have to kill when the page was | |
304 | * munmaped. But it could be also a mremap. Since that's | |
305 | * likely very rare kill anyways just out of paranoia, but use | |
306 | * a SIGKILL because the error is not contained anymore. | |
307 | */ | |
308 | if (tk->addr == -EFAULT) { | |
309 | pr_debug("MCE: Unable to find user space address %lx in %s\n", | |
310 | page_to_pfn(p), tsk->comm); | |
311 | tk->addr_valid = 0; | |
312 | } | |
313 | get_task_struct(tsk); | |
314 | tk->tsk = tsk; | |
315 | list_add_tail(&tk->nd, to_kill); | |
316 | } | |
317 | ||
318 | /* | |
319 | * Kill the processes that have been collected earlier. | |
320 | * | |
321 | * Only do anything when DOIT is set, otherwise just free the list | |
322 | * (this is used for clean pages which do not need killing) | |
323 | * Also when FAIL is set do a force kill because something went | |
324 | * wrong earlier. | |
325 | */ | |
326 | static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno, | |
327 | int fail, unsigned long pfn) | |
328 | { | |
329 | struct to_kill *tk, *next; | |
330 | ||
331 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
332 | if (doit) { | |
333 | /* | |
af901ca1 | 334 | * In case something went wrong with munmapping |
6a46079c AK |
335 | * make sure the process doesn't catch the |
336 | * signal and then access the memory. Just kill it. | |
6a46079c AK |
337 | */ |
338 | if (fail || tk->addr_valid == 0) { | |
339 | printk(KERN_ERR | |
340 | "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", | |
341 | pfn, tk->tsk->comm, tk->tsk->pid); | |
342 | force_sig(SIGKILL, tk->tsk); | |
343 | } | |
344 | ||
345 | /* | |
346 | * In theory the process could have mapped | |
347 | * something else on the address in-between. We could | |
348 | * check for that, but we need to tell the | |
349 | * process anyways. | |
350 | */ | |
351 | else if (kill_proc_ao(tk->tsk, tk->addr, trapno, | |
352 | pfn) < 0) | |
353 | printk(KERN_ERR | |
354 | "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n", | |
355 | pfn, tk->tsk->comm, tk->tsk->pid); | |
356 | } | |
357 | put_task_struct(tk->tsk); | |
358 | kfree(tk); | |
359 | } | |
360 | } | |
361 | ||
362 | static int task_early_kill(struct task_struct *tsk) | |
363 | { | |
364 | if (!tsk->mm) | |
365 | return 0; | |
366 | if (tsk->flags & PF_MCE_PROCESS) | |
367 | return !!(tsk->flags & PF_MCE_EARLY); | |
368 | return sysctl_memory_failure_early_kill; | |
369 | } | |
370 | ||
371 | /* | |
372 | * Collect processes when the error hit an anonymous page. | |
373 | */ | |
374 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
375 | struct to_kill **tkc) | |
376 | { | |
377 | struct vm_area_struct *vma; | |
378 | struct task_struct *tsk; | |
379 | struct anon_vma *av; | |
380 | ||
381 | read_lock(&tasklist_lock); | |
382 | av = page_lock_anon_vma(page); | |
383 | if (av == NULL) /* Not actually mapped anymore */ | |
384 | goto out; | |
385 | for_each_process (tsk) { | |
5beb4930 RR |
386 | struct anon_vma_chain *vmac; |
387 | ||
6a46079c AK |
388 | if (!task_early_kill(tsk)) |
389 | continue; | |
5beb4930 RR |
390 | list_for_each_entry(vmac, &av->head, same_anon_vma) { |
391 | vma = vmac->vma; | |
6a46079c AK |
392 | if (!page_mapped_in_vma(page, vma)) |
393 | continue; | |
394 | if (vma->vm_mm == tsk->mm) | |
395 | add_to_kill(tsk, page, vma, to_kill, tkc); | |
396 | } | |
397 | } | |
398 | page_unlock_anon_vma(av); | |
399 | out: | |
400 | read_unlock(&tasklist_lock); | |
401 | } | |
402 | ||
403 | /* | |
404 | * Collect processes when the error hit a file mapped page. | |
405 | */ | |
406 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
407 | struct to_kill **tkc) | |
408 | { | |
409 | struct vm_area_struct *vma; | |
410 | struct task_struct *tsk; | |
411 | struct prio_tree_iter iter; | |
412 | struct address_space *mapping = page->mapping; | |
413 | ||
414 | /* | |
415 | * A note on the locking order between the two locks. | |
416 | * We don't rely on this particular order. | |
417 | * If you have some other code that needs a different order | |
418 | * feel free to switch them around. Or add a reverse link | |
419 | * from mm_struct to task_struct, then this could be all | |
420 | * done without taking tasklist_lock and looping over all tasks. | |
421 | */ | |
422 | ||
423 | read_lock(&tasklist_lock); | |
424 | spin_lock(&mapping->i_mmap_lock); | |
425 | for_each_process(tsk) { | |
426 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
427 | ||
428 | if (!task_early_kill(tsk)) | |
429 | continue; | |
430 | ||
431 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, | |
432 | pgoff) { | |
433 | /* | |
434 | * Send early kill signal to tasks where a vma covers | |
435 | * the page but the corrupted page is not necessarily | |
436 | * mapped it in its pte. | |
437 | * Assume applications who requested early kill want | |
438 | * to be informed of all such data corruptions. | |
439 | */ | |
440 | if (vma->vm_mm == tsk->mm) | |
441 | add_to_kill(tsk, page, vma, to_kill, tkc); | |
442 | } | |
443 | } | |
444 | spin_unlock(&mapping->i_mmap_lock); | |
445 | read_unlock(&tasklist_lock); | |
446 | } | |
447 | ||
448 | /* | |
449 | * Collect the processes who have the corrupted page mapped to kill. | |
450 | * This is done in two steps for locking reasons. | |
451 | * First preallocate one tokill structure outside the spin locks, | |
452 | * so that we can kill at least one process reasonably reliable. | |
453 | */ | |
454 | static void collect_procs(struct page *page, struct list_head *tokill) | |
455 | { | |
456 | struct to_kill *tk; | |
457 | ||
458 | if (!page->mapping) | |
459 | return; | |
460 | ||
461 | tk = kmalloc(sizeof(struct to_kill), GFP_NOIO); | |
462 | if (!tk) | |
463 | return; | |
464 | if (PageAnon(page)) | |
465 | collect_procs_anon(page, tokill, &tk); | |
466 | else | |
467 | collect_procs_file(page, tokill, &tk); | |
468 | kfree(tk); | |
469 | } | |
470 | ||
471 | /* | |
472 | * Error handlers for various types of pages. | |
473 | */ | |
474 | ||
475 | enum outcome { | |
d95ea51e WF |
476 | IGNORED, /* Error: cannot be handled */ |
477 | FAILED, /* Error: handling failed */ | |
6a46079c | 478 | DELAYED, /* Will be handled later */ |
6a46079c AK |
479 | RECOVERED, /* Successfully recovered */ |
480 | }; | |
481 | ||
482 | static const char *action_name[] = { | |
d95ea51e | 483 | [IGNORED] = "Ignored", |
6a46079c AK |
484 | [FAILED] = "Failed", |
485 | [DELAYED] = "Delayed", | |
6a46079c AK |
486 | [RECOVERED] = "Recovered", |
487 | }; | |
488 | ||
dc2a1cbf WF |
489 | /* |
490 | * XXX: It is possible that a page is isolated from LRU cache, | |
491 | * and then kept in swap cache or failed to remove from page cache. | |
492 | * The page count will stop it from being freed by unpoison. | |
493 | * Stress tests should be aware of this memory leak problem. | |
494 | */ | |
495 | static int delete_from_lru_cache(struct page *p) | |
496 | { | |
497 | if (!isolate_lru_page(p)) { | |
498 | /* | |
499 | * Clear sensible page flags, so that the buddy system won't | |
500 | * complain when the page is unpoison-and-freed. | |
501 | */ | |
502 | ClearPageActive(p); | |
503 | ClearPageUnevictable(p); | |
504 | /* | |
505 | * drop the page count elevated by isolate_lru_page() | |
506 | */ | |
507 | page_cache_release(p); | |
508 | return 0; | |
509 | } | |
510 | return -EIO; | |
511 | } | |
512 | ||
6a46079c AK |
513 | /* |
514 | * Error hit kernel page. | |
515 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
516 | * could be more sophisticated. | |
517 | */ | |
518 | static int me_kernel(struct page *p, unsigned long pfn) | |
6a46079c AK |
519 | { |
520 | return IGNORED; | |
521 | } | |
522 | ||
523 | /* | |
524 | * Page in unknown state. Do nothing. | |
525 | */ | |
526 | static int me_unknown(struct page *p, unsigned long pfn) | |
527 | { | |
528 | printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn); | |
529 | return FAILED; | |
530 | } | |
531 | ||
6a46079c AK |
532 | /* |
533 | * Clean (or cleaned) page cache page. | |
534 | */ | |
535 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
536 | { | |
537 | int err; | |
538 | int ret = FAILED; | |
539 | struct address_space *mapping; | |
540 | ||
dc2a1cbf WF |
541 | delete_from_lru_cache(p); |
542 | ||
6a46079c AK |
543 | /* |
544 | * For anonymous pages we're done the only reference left | |
545 | * should be the one m_f() holds. | |
546 | */ | |
547 | if (PageAnon(p)) | |
548 | return RECOVERED; | |
549 | ||
550 | /* | |
551 | * Now truncate the page in the page cache. This is really | |
552 | * more like a "temporary hole punch" | |
553 | * Don't do this for block devices when someone else | |
554 | * has a reference, because it could be file system metadata | |
555 | * and that's not safe to truncate. | |
556 | */ | |
557 | mapping = page_mapping(p); | |
558 | if (!mapping) { | |
559 | /* | |
560 | * Page has been teared down in the meanwhile | |
561 | */ | |
562 | return FAILED; | |
563 | } | |
564 | ||
565 | /* | |
566 | * Truncation is a bit tricky. Enable it per file system for now. | |
567 | * | |
568 | * Open: to take i_mutex or not for this? Right now we don't. | |
569 | */ | |
570 | if (mapping->a_ops->error_remove_page) { | |
571 | err = mapping->a_ops->error_remove_page(mapping, p); | |
572 | if (err != 0) { | |
573 | printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n", | |
574 | pfn, err); | |
575 | } else if (page_has_private(p) && | |
576 | !try_to_release_page(p, GFP_NOIO)) { | |
577 | pr_debug("MCE %#lx: failed to release buffers\n", pfn); | |
578 | } else { | |
579 | ret = RECOVERED; | |
580 | } | |
581 | } else { | |
582 | /* | |
583 | * If the file system doesn't support it just invalidate | |
584 | * This fails on dirty or anything with private pages | |
585 | */ | |
586 | if (invalidate_inode_page(p)) | |
587 | ret = RECOVERED; | |
588 | else | |
589 | printk(KERN_INFO "MCE %#lx: Failed to invalidate\n", | |
590 | pfn); | |
591 | } | |
592 | return ret; | |
593 | } | |
594 | ||
595 | /* | |
596 | * Dirty cache page page | |
597 | * Issues: when the error hit a hole page the error is not properly | |
598 | * propagated. | |
599 | */ | |
600 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
601 | { | |
602 | struct address_space *mapping = page_mapping(p); | |
603 | ||
604 | SetPageError(p); | |
605 | /* TBD: print more information about the file. */ | |
606 | if (mapping) { | |
607 | /* | |
608 | * IO error will be reported by write(), fsync(), etc. | |
609 | * who check the mapping. | |
610 | * This way the application knows that something went | |
611 | * wrong with its dirty file data. | |
612 | * | |
613 | * There's one open issue: | |
614 | * | |
615 | * The EIO will be only reported on the next IO | |
616 | * operation and then cleared through the IO map. | |
617 | * Normally Linux has two mechanisms to pass IO error | |
618 | * first through the AS_EIO flag in the address space | |
619 | * and then through the PageError flag in the page. | |
620 | * Since we drop pages on memory failure handling the | |
621 | * only mechanism open to use is through AS_AIO. | |
622 | * | |
623 | * This has the disadvantage that it gets cleared on | |
624 | * the first operation that returns an error, while | |
625 | * the PageError bit is more sticky and only cleared | |
626 | * when the page is reread or dropped. If an | |
627 | * application assumes it will always get error on | |
628 | * fsync, but does other operations on the fd before | |
629 | * and the page is dropped inbetween then the error | |
630 | * will not be properly reported. | |
631 | * | |
632 | * This can already happen even without hwpoisoned | |
633 | * pages: first on metadata IO errors (which only | |
634 | * report through AS_EIO) or when the page is dropped | |
635 | * at the wrong time. | |
636 | * | |
637 | * So right now we assume that the application DTRT on | |
638 | * the first EIO, but we're not worse than other parts | |
639 | * of the kernel. | |
640 | */ | |
641 | mapping_set_error(mapping, EIO); | |
642 | } | |
643 | ||
644 | return me_pagecache_clean(p, pfn); | |
645 | } | |
646 | ||
647 | /* | |
648 | * Clean and dirty swap cache. | |
649 | * | |
650 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
651 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
652 | * referenced concurrently by 2 types of PTEs: | |
653 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
654 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
655 | * and then | |
656 | * - clear dirty bit to prevent IO | |
657 | * - remove from LRU | |
658 | * - but keep in the swap cache, so that when we return to it on | |
659 | * a later page fault, we know the application is accessing | |
660 | * corrupted data and shall be killed (we installed simple | |
661 | * interception code in do_swap_page to catch it). | |
662 | * | |
663 | * Clean swap cache pages can be directly isolated. A later page fault will | |
664 | * bring in the known good data from disk. | |
665 | */ | |
666 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
667 | { | |
6a46079c AK |
668 | ClearPageDirty(p); |
669 | /* Trigger EIO in shmem: */ | |
670 | ClearPageUptodate(p); | |
671 | ||
dc2a1cbf WF |
672 | if (!delete_from_lru_cache(p)) |
673 | return DELAYED; | |
674 | else | |
675 | return FAILED; | |
6a46079c AK |
676 | } |
677 | ||
678 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
679 | { | |
6a46079c | 680 | delete_from_swap_cache(p); |
e43c3afb | 681 | |
dc2a1cbf WF |
682 | if (!delete_from_lru_cache(p)) |
683 | return RECOVERED; | |
684 | else | |
685 | return FAILED; | |
6a46079c AK |
686 | } |
687 | ||
688 | /* | |
689 | * Huge pages. Needs work. | |
690 | * Issues: | |
691 | * No rmap support so we cannot find the original mapper. In theory could walk | |
692 | * all MMs and look for the mappings, but that would be non atomic and racy. | |
693 | * Need rmap for hugepages for this. Alternatively we could employ a heuristic, | |
694 | * like just walking the current process and hoping it has it mapped (that | |
695 | * should be usually true for the common "shared database cache" case) | |
696 | * Should handle free huge pages and dequeue them too, but this needs to | |
697 | * handle huge page accounting correctly. | |
698 | */ | |
699 | static int me_huge_page(struct page *p, unsigned long pfn) | |
700 | { | |
701 | return FAILED; | |
702 | } | |
703 | ||
704 | /* | |
705 | * Various page states we can handle. | |
706 | * | |
707 | * A page state is defined by its current page->flags bits. | |
708 | * The table matches them in order and calls the right handler. | |
709 | * | |
710 | * This is quite tricky because we can access page at any time | |
711 | * in its live cycle, so all accesses have to be extremly careful. | |
712 | * | |
713 | * This is not complete. More states could be added. | |
714 | * For any missing state don't attempt recovery. | |
715 | */ | |
716 | ||
717 | #define dirty (1UL << PG_dirty) | |
718 | #define sc (1UL << PG_swapcache) | |
719 | #define unevict (1UL << PG_unevictable) | |
720 | #define mlock (1UL << PG_mlocked) | |
721 | #define writeback (1UL << PG_writeback) | |
722 | #define lru (1UL << PG_lru) | |
723 | #define swapbacked (1UL << PG_swapbacked) | |
724 | #define head (1UL << PG_head) | |
725 | #define tail (1UL << PG_tail) | |
726 | #define compound (1UL << PG_compound) | |
727 | #define slab (1UL << PG_slab) | |
6a46079c AK |
728 | #define reserved (1UL << PG_reserved) |
729 | ||
730 | static struct page_state { | |
731 | unsigned long mask; | |
732 | unsigned long res; | |
733 | char *msg; | |
734 | int (*action)(struct page *p, unsigned long pfn); | |
735 | } error_states[] = { | |
d95ea51e | 736 | { reserved, reserved, "reserved kernel", me_kernel }, |
95d01fc6 WF |
737 | /* |
738 | * free pages are specially detected outside this table: | |
739 | * PG_buddy pages only make a small fraction of all free pages. | |
740 | */ | |
6a46079c AK |
741 | |
742 | /* | |
743 | * Could in theory check if slab page is free or if we can drop | |
744 | * currently unused objects without touching them. But just | |
745 | * treat it as standard kernel for now. | |
746 | */ | |
747 | { slab, slab, "kernel slab", me_kernel }, | |
748 | ||
749 | #ifdef CONFIG_PAGEFLAGS_EXTENDED | |
750 | { head, head, "huge", me_huge_page }, | |
751 | { tail, tail, "huge", me_huge_page }, | |
752 | #else | |
753 | { compound, compound, "huge", me_huge_page }, | |
754 | #endif | |
755 | ||
756 | { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty }, | |
757 | { sc|dirty, sc, "swapcache", me_swapcache_clean }, | |
758 | ||
759 | { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty}, | |
760 | { unevict, unevict, "unevictable LRU", me_pagecache_clean}, | |
761 | ||
6a46079c AK |
762 | { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty }, |
763 | { mlock, mlock, "mlocked LRU", me_pagecache_clean }, | |
6a46079c AK |
764 | |
765 | { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty }, | |
766 | { lru|dirty, lru, "clean LRU", me_pagecache_clean }, | |
6a46079c AK |
767 | |
768 | /* | |
769 | * Catchall entry: must be at end. | |
770 | */ | |
771 | { 0, 0, "unknown page state", me_unknown }, | |
772 | }; | |
773 | ||
2326c467 AK |
774 | #undef dirty |
775 | #undef sc | |
776 | #undef unevict | |
777 | #undef mlock | |
778 | #undef writeback | |
779 | #undef lru | |
780 | #undef swapbacked | |
781 | #undef head | |
782 | #undef tail | |
783 | #undef compound | |
784 | #undef slab | |
785 | #undef reserved | |
786 | ||
6a46079c AK |
787 | static void action_result(unsigned long pfn, char *msg, int result) |
788 | { | |
a7560fc8 | 789 | struct page *page = pfn_to_page(pfn); |
6a46079c AK |
790 | |
791 | printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n", | |
792 | pfn, | |
a7560fc8 | 793 | PageDirty(page) ? "dirty " : "", |
6a46079c AK |
794 | msg, action_name[result]); |
795 | } | |
796 | ||
797 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 798 | unsigned long pfn) |
6a46079c AK |
799 | { |
800 | int result; | |
7456b040 | 801 | int count; |
6a46079c AK |
802 | |
803 | result = ps->action(p, pfn); | |
804 | action_result(pfn, ps->msg, result); | |
7456b040 | 805 | |
bd1ce5f9 | 806 | count = page_count(p) - 1; |
138ce286 WF |
807 | if (ps->action == me_swapcache_dirty && result == DELAYED) |
808 | count--; | |
809 | if (count != 0) { | |
6a46079c AK |
810 | printk(KERN_ERR |
811 | "MCE %#lx: %s page still referenced by %d users\n", | |
7456b040 | 812 | pfn, ps->msg, count); |
138ce286 WF |
813 | result = FAILED; |
814 | } | |
6a46079c AK |
815 | |
816 | /* Could do more checks here if page looks ok */ | |
817 | /* | |
818 | * Could adjust zone counters here to correct for the missing page. | |
819 | */ | |
820 | ||
138ce286 | 821 | return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
822 | } |
823 | ||
824 | #define N_UNMAP_TRIES 5 | |
825 | ||
826 | /* | |
827 | * Do all that is necessary to remove user space mappings. Unmap | |
828 | * the pages and send SIGBUS to the processes if the data was dirty. | |
829 | */ | |
1668bfd5 | 830 | static int hwpoison_user_mappings(struct page *p, unsigned long pfn, |
6a46079c AK |
831 | int trapno) |
832 | { | |
833 | enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
834 | struct address_space *mapping; | |
835 | LIST_HEAD(tokill); | |
836 | int ret; | |
837 | int i; | |
838 | int kill = 1; | |
839 | ||
1668bfd5 WF |
840 | if (PageReserved(p) || PageSlab(p)) |
841 | return SWAP_SUCCESS; | |
6a46079c | 842 | |
6a46079c AK |
843 | /* |
844 | * This check implies we don't kill processes if their pages | |
845 | * are in the swap cache early. Those are always late kills. | |
846 | */ | |
847 | if (!page_mapped(p)) | |
1668bfd5 WF |
848 | return SWAP_SUCCESS; |
849 | ||
850 | if (PageCompound(p) || PageKsm(p)) | |
851 | return SWAP_FAIL; | |
6a46079c AK |
852 | |
853 | if (PageSwapCache(p)) { | |
854 | printk(KERN_ERR | |
855 | "MCE %#lx: keeping poisoned page in swap cache\n", pfn); | |
856 | ttu |= TTU_IGNORE_HWPOISON; | |
857 | } | |
858 | ||
859 | /* | |
860 | * Propagate the dirty bit from PTEs to struct page first, because we | |
861 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
862 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
863 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c AK |
864 | */ |
865 | mapping = page_mapping(p); | |
866 | if (!PageDirty(p) && mapping && mapping_cap_writeback_dirty(mapping)) { | |
867 | if (page_mkclean(p)) { | |
868 | SetPageDirty(p); | |
869 | } else { | |
870 | kill = 0; | |
871 | ttu |= TTU_IGNORE_HWPOISON; | |
872 | printk(KERN_INFO | |
873 | "MCE %#lx: corrupted page was clean: dropped without side effects\n", | |
874 | pfn); | |
875 | } | |
876 | } | |
877 | ||
878 | /* | |
879 | * First collect all the processes that have the page | |
880 | * mapped in dirty form. This has to be done before try_to_unmap, | |
881 | * because ttu takes the rmap data structures down. | |
882 | * | |
883 | * Error handling: We ignore errors here because | |
884 | * there's nothing that can be done. | |
885 | */ | |
886 | if (kill) | |
887 | collect_procs(p, &tokill); | |
888 | ||
889 | /* | |
890 | * try_to_unmap can fail temporarily due to races. | |
891 | * Try a few times (RED-PEN better strategy?) | |
892 | */ | |
893 | for (i = 0; i < N_UNMAP_TRIES; i++) { | |
894 | ret = try_to_unmap(p, ttu); | |
895 | if (ret == SWAP_SUCCESS) | |
896 | break; | |
897 | pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn, ret); | |
898 | } | |
899 | ||
900 | if (ret != SWAP_SUCCESS) | |
901 | printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n", | |
902 | pfn, page_mapcount(p)); | |
903 | ||
904 | /* | |
905 | * Now that the dirty bit has been propagated to the | |
906 | * struct page and all unmaps done we can decide if | |
907 | * killing is needed or not. Only kill when the page | |
908 | * was dirty, otherwise the tokill list is merely | |
909 | * freed. When there was a problem unmapping earlier | |
910 | * use a more force-full uncatchable kill to prevent | |
911 | * any accesses to the poisoned memory. | |
912 | */ | |
913 | kill_procs_ao(&tokill, !!PageDirty(p), trapno, | |
914 | ret != SWAP_SUCCESS, pfn); | |
1668bfd5 WF |
915 | |
916 | return ret; | |
6a46079c AK |
917 | } |
918 | ||
82ba011b | 919 | int __memory_failure(unsigned long pfn, int trapno, int flags) |
6a46079c AK |
920 | { |
921 | struct page_state *ps; | |
922 | struct page *p; | |
923 | int res; | |
924 | ||
925 | if (!sysctl_memory_failure_recovery) | |
926 | panic("Memory failure from trap %d on page %lx", trapno, pfn); | |
927 | ||
928 | if (!pfn_valid(pfn)) { | |
a7560fc8 WF |
929 | printk(KERN_ERR |
930 | "MCE %#lx: memory outside kernel control\n", | |
931 | pfn); | |
932 | return -ENXIO; | |
6a46079c AK |
933 | } |
934 | ||
935 | p = pfn_to_page(pfn); | |
936 | if (TestSetPageHWPoison(p)) { | |
d95ea51e | 937 | printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn); |
6a46079c AK |
938 | return 0; |
939 | } | |
940 | ||
941 | atomic_long_add(1, &mce_bad_pages); | |
942 | ||
943 | /* | |
944 | * We need/can do nothing about count=0 pages. | |
945 | * 1) it's a free page, and therefore in safe hand: | |
946 | * prep_new_page() will be the gate keeper. | |
947 | * 2) it's part of a non-compound high order page. | |
948 | * Implies some kernel user: cannot stop them from | |
949 | * R/W the page; let's pray that the page has been | |
950 | * used and will be freed some time later. | |
951 | * In fact it's dangerous to directly bump up page count from 0, | |
952 | * that may make page_freeze_refs()/page_unfreeze_refs() mismatch. | |
953 | */ | |
82ba011b AK |
954 | if (!(flags & MF_COUNT_INCREASED) && |
955 | !get_page_unless_zero(compound_head(p))) { | |
8d22ba1b WF |
956 | if (is_free_buddy_page(p)) { |
957 | action_result(pfn, "free buddy", DELAYED); | |
958 | return 0; | |
959 | } else { | |
960 | action_result(pfn, "high order kernel", IGNORED); | |
961 | return -EBUSY; | |
962 | } | |
6a46079c AK |
963 | } |
964 | ||
e43c3afb WF |
965 | /* |
966 | * We ignore non-LRU pages for good reasons. | |
967 | * - PG_locked is only well defined for LRU pages and a few others | |
968 | * - to avoid races with __set_page_locked() | |
969 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) | |
970 | * The check (unnecessarily) ignores LRU pages being isolated and | |
971 | * walked by the page reclaim code, however that's not a big loss. | |
972 | */ | |
973 | if (!PageLRU(p)) | |
facb6011 | 974 | shake_page(p, 0); |
dc2a1cbf | 975 | if (!PageLRU(p)) { |
0474a60e AK |
976 | /* |
977 | * shake_page could have turned it free. | |
978 | */ | |
979 | if (is_free_buddy_page(p)) { | |
980 | action_result(pfn, "free buddy, 2nd try", DELAYED); | |
981 | return 0; | |
982 | } | |
e43c3afb WF |
983 | action_result(pfn, "non LRU", IGNORED); |
984 | put_page(p); | |
985 | return -EBUSY; | |
986 | } | |
e43c3afb | 987 | |
6a46079c AK |
988 | /* |
989 | * Lock the page and wait for writeback to finish. | |
990 | * It's very difficult to mess with pages currently under IO | |
991 | * and in many cases impossible, so we just avoid it here. | |
992 | */ | |
993 | lock_page_nosync(p); | |
847ce401 WF |
994 | |
995 | /* | |
996 | * unpoison always clear PG_hwpoison inside page lock | |
997 | */ | |
998 | if (!PageHWPoison(p)) { | |
d95ea51e | 999 | printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn); |
847ce401 WF |
1000 | res = 0; |
1001 | goto out; | |
1002 | } | |
7c116f2b WF |
1003 | if (hwpoison_filter(p)) { |
1004 | if (TestClearPageHWPoison(p)) | |
1005 | atomic_long_dec(&mce_bad_pages); | |
1006 | unlock_page(p); | |
1007 | put_page(p); | |
1008 | return 0; | |
1009 | } | |
847ce401 | 1010 | |
6a46079c AK |
1011 | wait_on_page_writeback(p); |
1012 | ||
1013 | /* | |
1014 | * Now take care of user space mappings. | |
1668bfd5 | 1015 | * Abort on fail: __remove_from_page_cache() assumes unmapped page. |
6a46079c | 1016 | */ |
1668bfd5 WF |
1017 | if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) { |
1018 | printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn); | |
1019 | res = -EBUSY; | |
1020 | goto out; | |
1021 | } | |
6a46079c AK |
1022 | |
1023 | /* | |
1024 | * Torn down by someone else? | |
1025 | */ | |
dc2a1cbf | 1026 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
6a46079c | 1027 | action_result(pfn, "already truncated LRU", IGNORED); |
d95ea51e | 1028 | res = -EBUSY; |
6a46079c AK |
1029 | goto out; |
1030 | } | |
1031 | ||
1032 | res = -EBUSY; | |
1033 | for (ps = error_states;; ps++) { | |
dc2a1cbf | 1034 | if ((p->flags & ps->mask) == ps->res) { |
bd1ce5f9 | 1035 | res = page_action(ps, p, pfn); |
6a46079c AK |
1036 | break; |
1037 | } | |
1038 | } | |
1039 | out: | |
1040 | unlock_page(p); | |
1041 | return res; | |
1042 | } | |
1043 | EXPORT_SYMBOL_GPL(__memory_failure); | |
1044 | ||
1045 | /** | |
1046 | * memory_failure - Handle memory failure of a page. | |
1047 | * @pfn: Page Number of the corrupted page | |
1048 | * @trapno: Trap number reported in the signal to user space. | |
1049 | * | |
1050 | * This function is called by the low level machine check code | |
1051 | * of an architecture when it detects hardware memory corruption | |
1052 | * of a page. It tries its best to recover, which includes | |
1053 | * dropping pages, killing processes etc. | |
1054 | * | |
1055 | * The function is primarily of use for corruptions that | |
1056 | * happen outside the current execution context (e.g. when | |
1057 | * detected by a background scrubber) | |
1058 | * | |
1059 | * Must run in process context (e.g. a work queue) with interrupts | |
1060 | * enabled and no spinlocks hold. | |
1061 | */ | |
1062 | void memory_failure(unsigned long pfn, int trapno) | |
1063 | { | |
1064 | __memory_failure(pfn, trapno, 0); | |
1065 | } | |
847ce401 WF |
1066 | |
1067 | /** | |
1068 | * unpoison_memory - Unpoison a previously poisoned page | |
1069 | * @pfn: Page number of the to be unpoisoned page | |
1070 | * | |
1071 | * Software-unpoison a page that has been poisoned by | |
1072 | * memory_failure() earlier. | |
1073 | * | |
1074 | * This is only done on the software-level, so it only works | |
1075 | * for linux injected failures, not real hardware failures | |
1076 | * | |
1077 | * Returns 0 for success, otherwise -errno. | |
1078 | */ | |
1079 | int unpoison_memory(unsigned long pfn) | |
1080 | { | |
1081 | struct page *page; | |
1082 | struct page *p; | |
1083 | int freeit = 0; | |
1084 | ||
1085 | if (!pfn_valid(pfn)) | |
1086 | return -ENXIO; | |
1087 | ||
1088 | p = pfn_to_page(pfn); | |
1089 | page = compound_head(p); | |
1090 | ||
1091 | if (!PageHWPoison(p)) { | |
1092 | pr_debug("MCE: Page was already unpoisoned %#lx\n", pfn); | |
1093 | return 0; | |
1094 | } | |
1095 | ||
1096 | if (!get_page_unless_zero(page)) { | |
1097 | if (TestClearPageHWPoison(p)) | |
1098 | atomic_long_dec(&mce_bad_pages); | |
1099 | pr_debug("MCE: Software-unpoisoned free page %#lx\n", pfn); | |
1100 | return 0; | |
1101 | } | |
1102 | ||
1103 | lock_page_nosync(page); | |
1104 | /* | |
1105 | * This test is racy because PG_hwpoison is set outside of page lock. | |
1106 | * That's acceptable because that won't trigger kernel panic. Instead, | |
1107 | * the PG_hwpoison page will be caught and isolated on the entrance to | |
1108 | * the free buddy page pool. | |
1109 | */ | |
1110 | if (TestClearPageHWPoison(p)) { | |
1111 | pr_debug("MCE: Software-unpoisoned page %#lx\n", pfn); | |
1112 | atomic_long_dec(&mce_bad_pages); | |
1113 | freeit = 1; | |
1114 | } | |
1115 | unlock_page(page); | |
1116 | ||
1117 | put_page(page); | |
1118 | if (freeit) | |
1119 | put_page(page); | |
1120 | ||
1121 | return 0; | |
1122 | } | |
1123 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 AK |
1124 | |
1125 | static struct page *new_page(struct page *p, unsigned long private, int **x) | |
1126 | { | |
12686d15 AK |
1127 | int nid = page_to_nid(p); |
1128 | return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0); | |
facb6011 AK |
1129 | } |
1130 | ||
1131 | /* | |
1132 | * Safely get reference count of an arbitrary page. | |
1133 | * Returns 0 for a free page, -EIO for a zero refcount page | |
1134 | * that is not free, and 1 for any other page type. | |
1135 | * For 1 the page is returned with increased page count, otherwise not. | |
1136 | */ | |
1137 | static int get_any_page(struct page *p, unsigned long pfn, int flags) | |
1138 | { | |
1139 | int ret; | |
1140 | ||
1141 | if (flags & MF_COUNT_INCREASED) | |
1142 | return 1; | |
1143 | ||
1144 | /* | |
1145 | * The lock_system_sleep prevents a race with memory hotplug, | |
1146 | * because the isolation assumes there's only a single user. | |
1147 | * This is a big hammer, a better would be nicer. | |
1148 | */ | |
1149 | lock_system_sleep(); | |
1150 | ||
1151 | /* | |
1152 | * Isolate the page, so that it doesn't get reallocated if it | |
1153 | * was free. | |
1154 | */ | |
1155 | set_migratetype_isolate(p); | |
1156 | if (!get_page_unless_zero(compound_head(p))) { | |
1157 | if (is_free_buddy_page(p)) { | |
1158 | pr_debug("get_any_page: %#lx free buddy page\n", pfn); | |
1159 | /* Set hwpoison bit while page is still isolated */ | |
1160 | SetPageHWPoison(p); | |
1161 | ret = 0; | |
1162 | } else { | |
1163 | pr_debug("get_any_page: %#lx: unknown zero refcount page type %lx\n", | |
1164 | pfn, p->flags); | |
1165 | ret = -EIO; | |
1166 | } | |
1167 | } else { | |
1168 | /* Not a free page */ | |
1169 | ret = 1; | |
1170 | } | |
1171 | unset_migratetype_isolate(p); | |
1172 | unlock_system_sleep(); | |
1173 | return ret; | |
1174 | } | |
1175 | ||
1176 | /** | |
1177 | * soft_offline_page - Soft offline a page. | |
1178 | * @page: page to offline | |
1179 | * @flags: flags. Same as memory_failure(). | |
1180 | * | |
1181 | * Returns 0 on success, otherwise negated errno. | |
1182 | * | |
1183 | * Soft offline a page, by migration or invalidation, | |
1184 | * without killing anything. This is for the case when | |
1185 | * a page is not corrupted yet (so it's still valid to access), | |
1186 | * but has had a number of corrected errors and is better taken | |
1187 | * out. | |
1188 | * | |
1189 | * The actual policy on when to do that is maintained by | |
1190 | * user space. | |
1191 | * | |
1192 | * This should never impact any application or cause data loss, | |
1193 | * however it might take some time. | |
1194 | * | |
1195 | * This is not a 100% solution for all memory, but tries to be | |
1196 | * ``good enough'' for the majority of memory. | |
1197 | */ | |
1198 | int soft_offline_page(struct page *page, int flags) | |
1199 | { | |
1200 | int ret; | |
1201 | unsigned long pfn = page_to_pfn(page); | |
1202 | ||
1203 | ret = get_any_page(page, pfn, flags); | |
1204 | if (ret < 0) | |
1205 | return ret; | |
1206 | if (ret == 0) | |
1207 | goto done; | |
1208 | ||
1209 | /* | |
1210 | * Page cache page we can handle? | |
1211 | */ | |
1212 | if (!PageLRU(page)) { | |
1213 | /* | |
1214 | * Try to free it. | |
1215 | */ | |
1216 | put_page(page); | |
1217 | shake_page(page, 1); | |
1218 | ||
1219 | /* | |
1220 | * Did it turn free? | |
1221 | */ | |
1222 | ret = get_any_page(page, pfn, 0); | |
1223 | if (ret < 0) | |
1224 | return ret; | |
1225 | if (ret == 0) | |
1226 | goto done; | |
1227 | } | |
1228 | if (!PageLRU(page)) { | |
1229 | pr_debug("soft_offline: %#lx: unknown non LRU page type %lx\n", | |
1230 | pfn, page->flags); | |
1231 | return -EIO; | |
1232 | } | |
1233 | ||
1234 | lock_page(page); | |
1235 | wait_on_page_writeback(page); | |
1236 | ||
1237 | /* | |
1238 | * Synchronized using the page lock with memory_failure() | |
1239 | */ | |
1240 | if (PageHWPoison(page)) { | |
1241 | unlock_page(page); | |
1242 | put_page(page); | |
1243 | pr_debug("soft offline: %#lx page already poisoned\n", pfn); | |
1244 | return -EBUSY; | |
1245 | } | |
1246 | ||
1247 | /* | |
1248 | * Try to invalidate first. This should work for | |
1249 | * non dirty unmapped page cache pages. | |
1250 | */ | |
1251 | ret = invalidate_inode_page(page); | |
1252 | unlock_page(page); | |
1253 | ||
1254 | /* | |
1255 | * Drop count because page migration doesn't like raised | |
1256 | * counts. The page could get re-allocated, but if it becomes | |
1257 | * LRU the isolation will just fail. | |
1258 | * RED-PEN would be better to keep it isolated here, but we | |
1259 | * would need to fix isolation locking first. | |
1260 | */ | |
1261 | put_page(page); | |
1262 | if (ret == 1) { | |
1263 | ret = 0; | |
1264 | pr_debug("soft_offline: %#lx: invalidated\n", pfn); | |
1265 | goto done; | |
1266 | } | |
1267 | ||
1268 | /* | |
1269 | * Simple invalidation didn't work. | |
1270 | * Try to migrate to a new page instead. migrate.c | |
1271 | * handles a large number of cases for us. | |
1272 | */ | |
1273 | ret = isolate_lru_page(page); | |
1274 | if (!ret) { | |
1275 | LIST_HEAD(pagelist); | |
1276 | ||
1277 | list_add(&page->lru, &pagelist); | |
1278 | ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0); | |
1279 | if (ret) { | |
1280 | pr_debug("soft offline: %#lx: migration failed %d, type %lx\n", | |
1281 | pfn, ret, page->flags); | |
1282 | if (ret > 0) | |
1283 | ret = -EIO; | |
1284 | } | |
1285 | } else { | |
1286 | pr_debug("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n", | |
1287 | pfn, ret, page_count(page), page->flags); | |
1288 | } | |
1289 | if (ret) | |
1290 | return ret; | |
1291 | ||
1292 | done: | |
1293 | atomic_long_add(1, &mce_bad_pages); | |
1294 | SetPageHWPoison(page); | |
1295 | /* keep elevated page count for bad page */ | |
1296 | return ret; | |
1297 | } |