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