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