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> | |
37 | #include <linux/sched.h> | |
01e00f88 | 38 | #include <linux/ksm.h> |
6a46079c AK |
39 | #include <linux/rmap.h> |
40 | #include <linux/pagemap.h> | |
41 | #include <linux/swap.h> | |
42 | #include <linux/backing-dev.h> | |
43 | #include "internal.h" | |
44 | ||
45 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
46 | ||
47 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
48 | ||
49 | atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0); | |
50 | ||
51 | /* | |
52 | * Send all the processes who have the page mapped an ``action optional'' | |
53 | * signal. | |
54 | */ | |
55 | static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno, | |
56 | unsigned long pfn) | |
57 | { | |
58 | struct siginfo si; | |
59 | int ret; | |
60 | ||
61 | printk(KERN_ERR | |
62 | "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n", | |
63 | pfn, t->comm, t->pid); | |
64 | si.si_signo = SIGBUS; | |
65 | si.si_errno = 0; | |
66 | si.si_code = BUS_MCEERR_AO; | |
67 | si.si_addr = (void *)addr; | |
68 | #ifdef __ARCH_SI_TRAPNO | |
69 | si.si_trapno = trapno; | |
70 | #endif | |
71 | si.si_addr_lsb = PAGE_SHIFT; | |
72 | /* | |
73 | * Don't use force here, it's convenient if the signal | |
74 | * can be temporarily blocked. | |
75 | * This could cause a loop when the user sets SIGBUS | |
76 | * to SIG_IGN, but hopefully noone will do that? | |
77 | */ | |
78 | ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */ | |
79 | if (ret < 0) | |
80 | printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n", | |
81 | t->comm, t->pid, ret); | |
82 | return ret; | |
83 | } | |
84 | ||
85 | /* | |
86 | * Kill all processes that have a poisoned page mapped and then isolate | |
87 | * the page. | |
88 | * | |
89 | * General strategy: | |
90 | * Find all processes having the page mapped and kill them. | |
91 | * But we keep a page reference around so that the page is not | |
92 | * actually freed yet. | |
93 | * Then stash the page away | |
94 | * | |
95 | * There's no convenient way to get back to mapped processes | |
96 | * from the VMAs. So do a brute-force search over all | |
97 | * running processes. | |
98 | * | |
99 | * Remember that machine checks are not common (or rather | |
100 | * if they are common you have other problems), so this shouldn't | |
101 | * be a performance issue. | |
102 | * | |
103 | * Also there are some races possible while we get from the | |
104 | * error detection to actually handle it. | |
105 | */ | |
106 | ||
107 | struct to_kill { | |
108 | struct list_head nd; | |
109 | struct task_struct *tsk; | |
110 | unsigned long addr; | |
111 | unsigned addr_valid:1; | |
112 | }; | |
113 | ||
114 | /* | |
115 | * Failure handling: if we can't find or can't kill a process there's | |
116 | * not much we can do. We just print a message and ignore otherwise. | |
117 | */ | |
118 | ||
119 | /* | |
120 | * Schedule a process for later kill. | |
121 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
122 | * TBD would GFP_NOIO be enough? | |
123 | */ | |
124 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
125 | struct vm_area_struct *vma, | |
126 | struct list_head *to_kill, | |
127 | struct to_kill **tkc) | |
128 | { | |
129 | struct to_kill *tk; | |
130 | ||
131 | if (*tkc) { | |
132 | tk = *tkc; | |
133 | *tkc = NULL; | |
134 | } else { | |
135 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); | |
136 | if (!tk) { | |
137 | printk(KERN_ERR | |
138 | "MCE: Out of memory while machine check handling\n"); | |
139 | return; | |
140 | } | |
141 | } | |
142 | tk->addr = page_address_in_vma(p, vma); | |
143 | tk->addr_valid = 1; | |
144 | ||
145 | /* | |
146 | * In theory we don't have to kill when the page was | |
147 | * munmaped. But it could be also a mremap. Since that's | |
148 | * likely very rare kill anyways just out of paranoia, but use | |
149 | * a SIGKILL because the error is not contained anymore. | |
150 | */ | |
151 | if (tk->addr == -EFAULT) { | |
152 | pr_debug("MCE: Unable to find user space address %lx in %s\n", | |
153 | page_to_pfn(p), tsk->comm); | |
154 | tk->addr_valid = 0; | |
155 | } | |
156 | get_task_struct(tsk); | |
157 | tk->tsk = tsk; | |
158 | list_add_tail(&tk->nd, to_kill); | |
159 | } | |
160 | ||
161 | /* | |
162 | * Kill the processes that have been collected earlier. | |
163 | * | |
164 | * Only do anything when DOIT is set, otherwise just free the list | |
165 | * (this is used for clean pages which do not need killing) | |
166 | * Also when FAIL is set do a force kill because something went | |
167 | * wrong earlier. | |
168 | */ | |
169 | static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno, | |
170 | int fail, unsigned long pfn) | |
171 | { | |
172 | struct to_kill *tk, *next; | |
173 | ||
174 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
175 | if (doit) { | |
176 | /* | |
177 | * In case something went wrong with munmaping | |
178 | * make sure the process doesn't catch the | |
179 | * signal and then access the memory. Just kill it. | |
180 | * the signal handlers | |
181 | */ | |
182 | if (fail || tk->addr_valid == 0) { | |
183 | printk(KERN_ERR | |
184 | "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", | |
185 | pfn, tk->tsk->comm, tk->tsk->pid); | |
186 | force_sig(SIGKILL, tk->tsk); | |
187 | } | |
188 | ||
189 | /* | |
190 | * In theory the process could have mapped | |
191 | * something else on the address in-between. We could | |
192 | * check for that, but we need to tell the | |
193 | * process anyways. | |
194 | */ | |
195 | else if (kill_proc_ao(tk->tsk, tk->addr, trapno, | |
196 | pfn) < 0) | |
197 | printk(KERN_ERR | |
198 | "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n", | |
199 | pfn, tk->tsk->comm, tk->tsk->pid); | |
200 | } | |
201 | put_task_struct(tk->tsk); | |
202 | kfree(tk); | |
203 | } | |
204 | } | |
205 | ||
206 | static int task_early_kill(struct task_struct *tsk) | |
207 | { | |
208 | if (!tsk->mm) | |
209 | return 0; | |
210 | if (tsk->flags & PF_MCE_PROCESS) | |
211 | return !!(tsk->flags & PF_MCE_EARLY); | |
212 | return sysctl_memory_failure_early_kill; | |
213 | } | |
214 | ||
215 | /* | |
216 | * Collect processes when the error hit an anonymous page. | |
217 | */ | |
218 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
219 | struct to_kill **tkc) | |
220 | { | |
221 | struct vm_area_struct *vma; | |
222 | struct task_struct *tsk; | |
223 | struct anon_vma *av; | |
224 | ||
225 | read_lock(&tasklist_lock); | |
226 | av = page_lock_anon_vma(page); | |
227 | if (av == NULL) /* Not actually mapped anymore */ | |
228 | goto out; | |
229 | for_each_process (tsk) { | |
230 | if (!task_early_kill(tsk)) | |
231 | continue; | |
232 | list_for_each_entry (vma, &av->head, anon_vma_node) { | |
233 | if (!page_mapped_in_vma(page, vma)) | |
234 | continue; | |
235 | if (vma->vm_mm == tsk->mm) | |
236 | add_to_kill(tsk, page, vma, to_kill, tkc); | |
237 | } | |
238 | } | |
239 | page_unlock_anon_vma(av); | |
240 | out: | |
241 | read_unlock(&tasklist_lock); | |
242 | } | |
243 | ||
244 | /* | |
245 | * Collect processes when the error hit a file mapped page. | |
246 | */ | |
247 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
248 | struct to_kill **tkc) | |
249 | { | |
250 | struct vm_area_struct *vma; | |
251 | struct task_struct *tsk; | |
252 | struct prio_tree_iter iter; | |
253 | struct address_space *mapping = page->mapping; | |
254 | ||
255 | /* | |
256 | * A note on the locking order between the two locks. | |
257 | * We don't rely on this particular order. | |
258 | * If you have some other code that needs a different order | |
259 | * feel free to switch them around. Or add a reverse link | |
260 | * from mm_struct to task_struct, then this could be all | |
261 | * done without taking tasklist_lock and looping over all tasks. | |
262 | */ | |
263 | ||
264 | read_lock(&tasklist_lock); | |
265 | spin_lock(&mapping->i_mmap_lock); | |
266 | for_each_process(tsk) { | |
267 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
268 | ||
269 | if (!task_early_kill(tsk)) | |
270 | continue; | |
271 | ||
272 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, | |
273 | pgoff) { | |
274 | /* | |
275 | * Send early kill signal to tasks where a vma covers | |
276 | * the page but the corrupted page is not necessarily | |
277 | * mapped it in its pte. | |
278 | * Assume applications who requested early kill want | |
279 | * to be informed of all such data corruptions. | |
280 | */ | |
281 | if (vma->vm_mm == tsk->mm) | |
282 | add_to_kill(tsk, page, vma, to_kill, tkc); | |
283 | } | |
284 | } | |
285 | spin_unlock(&mapping->i_mmap_lock); | |
286 | read_unlock(&tasklist_lock); | |
287 | } | |
288 | ||
289 | /* | |
290 | * Collect the processes who have the corrupted page mapped to kill. | |
291 | * This is done in two steps for locking reasons. | |
292 | * First preallocate one tokill structure outside the spin locks, | |
293 | * so that we can kill at least one process reasonably reliable. | |
294 | */ | |
295 | static void collect_procs(struct page *page, struct list_head *tokill) | |
296 | { | |
297 | struct to_kill *tk; | |
298 | ||
299 | if (!page->mapping) | |
300 | return; | |
301 | ||
302 | tk = kmalloc(sizeof(struct to_kill), GFP_NOIO); | |
303 | if (!tk) | |
304 | return; | |
305 | if (PageAnon(page)) | |
306 | collect_procs_anon(page, tokill, &tk); | |
307 | else | |
308 | collect_procs_file(page, tokill, &tk); | |
309 | kfree(tk); | |
310 | } | |
311 | ||
312 | /* | |
313 | * Error handlers for various types of pages. | |
314 | */ | |
315 | ||
316 | enum outcome { | |
317 | FAILED, /* Error handling failed */ | |
318 | DELAYED, /* Will be handled later */ | |
319 | IGNORED, /* Error safely ignored */ | |
320 | RECOVERED, /* Successfully recovered */ | |
321 | }; | |
322 | ||
323 | static const char *action_name[] = { | |
324 | [FAILED] = "Failed", | |
325 | [DELAYED] = "Delayed", | |
326 | [IGNORED] = "Ignored", | |
327 | [RECOVERED] = "Recovered", | |
328 | }; | |
329 | ||
330 | /* | |
331 | * Error hit kernel page. | |
332 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
333 | * could be more sophisticated. | |
334 | */ | |
335 | static int me_kernel(struct page *p, unsigned long pfn) | |
336 | { | |
337 | return DELAYED; | |
338 | } | |
339 | ||
340 | /* | |
341 | * Already poisoned page. | |
342 | */ | |
343 | static int me_ignore(struct page *p, unsigned long pfn) | |
344 | { | |
345 | return IGNORED; | |
346 | } | |
347 | ||
348 | /* | |
349 | * Page in unknown state. Do nothing. | |
350 | */ | |
351 | static int me_unknown(struct page *p, unsigned long pfn) | |
352 | { | |
353 | printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn); | |
354 | return FAILED; | |
355 | } | |
356 | ||
357 | /* | |
358 | * Free memory | |
359 | */ | |
360 | static int me_free(struct page *p, unsigned long pfn) | |
361 | { | |
362 | return DELAYED; | |
363 | } | |
364 | ||
365 | /* | |
366 | * Clean (or cleaned) page cache page. | |
367 | */ | |
368 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
369 | { | |
370 | int err; | |
371 | int ret = FAILED; | |
372 | struct address_space *mapping; | |
373 | ||
6a46079c AK |
374 | /* |
375 | * For anonymous pages we're done the only reference left | |
376 | * should be the one m_f() holds. | |
377 | */ | |
378 | if (PageAnon(p)) | |
379 | return RECOVERED; | |
380 | ||
381 | /* | |
382 | * Now truncate the page in the page cache. This is really | |
383 | * more like a "temporary hole punch" | |
384 | * Don't do this for block devices when someone else | |
385 | * has a reference, because it could be file system metadata | |
386 | * and that's not safe to truncate. | |
387 | */ | |
388 | mapping = page_mapping(p); | |
389 | if (!mapping) { | |
390 | /* | |
391 | * Page has been teared down in the meanwhile | |
392 | */ | |
393 | return FAILED; | |
394 | } | |
395 | ||
396 | /* | |
397 | * Truncation is a bit tricky. Enable it per file system for now. | |
398 | * | |
399 | * Open: to take i_mutex or not for this? Right now we don't. | |
400 | */ | |
401 | if (mapping->a_ops->error_remove_page) { | |
402 | err = mapping->a_ops->error_remove_page(mapping, p); | |
403 | if (err != 0) { | |
404 | printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n", | |
405 | pfn, err); | |
406 | } else if (page_has_private(p) && | |
407 | !try_to_release_page(p, GFP_NOIO)) { | |
408 | pr_debug("MCE %#lx: failed to release buffers\n", pfn); | |
409 | } else { | |
410 | ret = RECOVERED; | |
411 | } | |
412 | } else { | |
413 | /* | |
414 | * If the file system doesn't support it just invalidate | |
415 | * This fails on dirty or anything with private pages | |
416 | */ | |
417 | if (invalidate_inode_page(p)) | |
418 | ret = RECOVERED; | |
419 | else | |
420 | printk(KERN_INFO "MCE %#lx: Failed to invalidate\n", | |
421 | pfn); | |
422 | } | |
423 | return ret; | |
424 | } | |
425 | ||
426 | /* | |
427 | * Dirty cache page page | |
428 | * Issues: when the error hit a hole page the error is not properly | |
429 | * propagated. | |
430 | */ | |
431 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
432 | { | |
433 | struct address_space *mapping = page_mapping(p); | |
434 | ||
435 | SetPageError(p); | |
436 | /* TBD: print more information about the file. */ | |
437 | if (mapping) { | |
438 | /* | |
439 | * IO error will be reported by write(), fsync(), etc. | |
440 | * who check the mapping. | |
441 | * This way the application knows that something went | |
442 | * wrong with its dirty file data. | |
443 | * | |
444 | * There's one open issue: | |
445 | * | |
446 | * The EIO will be only reported on the next IO | |
447 | * operation and then cleared through the IO map. | |
448 | * Normally Linux has two mechanisms to pass IO error | |
449 | * first through the AS_EIO flag in the address space | |
450 | * and then through the PageError flag in the page. | |
451 | * Since we drop pages on memory failure handling the | |
452 | * only mechanism open to use is through AS_AIO. | |
453 | * | |
454 | * This has the disadvantage that it gets cleared on | |
455 | * the first operation that returns an error, while | |
456 | * the PageError bit is more sticky and only cleared | |
457 | * when the page is reread or dropped. If an | |
458 | * application assumes it will always get error on | |
459 | * fsync, but does other operations on the fd before | |
460 | * and the page is dropped inbetween then the error | |
461 | * will not be properly reported. | |
462 | * | |
463 | * This can already happen even without hwpoisoned | |
464 | * pages: first on metadata IO errors (which only | |
465 | * report through AS_EIO) or when the page is dropped | |
466 | * at the wrong time. | |
467 | * | |
468 | * So right now we assume that the application DTRT on | |
469 | * the first EIO, but we're not worse than other parts | |
470 | * of the kernel. | |
471 | */ | |
472 | mapping_set_error(mapping, EIO); | |
473 | } | |
474 | ||
475 | return me_pagecache_clean(p, pfn); | |
476 | } | |
477 | ||
478 | /* | |
479 | * Clean and dirty swap cache. | |
480 | * | |
481 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
482 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
483 | * referenced concurrently by 2 types of PTEs: | |
484 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
485 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
486 | * and then | |
487 | * - clear dirty bit to prevent IO | |
488 | * - remove from LRU | |
489 | * - but keep in the swap cache, so that when we return to it on | |
490 | * a later page fault, we know the application is accessing | |
491 | * corrupted data and shall be killed (we installed simple | |
492 | * interception code in do_swap_page to catch it). | |
493 | * | |
494 | * Clean swap cache pages can be directly isolated. A later page fault will | |
495 | * bring in the known good data from disk. | |
496 | */ | |
497 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
498 | { | |
6a46079c AK |
499 | ClearPageDirty(p); |
500 | /* Trigger EIO in shmem: */ | |
501 | ClearPageUptodate(p); | |
502 | ||
e43c3afb | 503 | return DELAYED; |
6a46079c AK |
504 | } |
505 | ||
506 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
507 | { | |
6a46079c | 508 | delete_from_swap_cache(p); |
e43c3afb WF |
509 | |
510 | return RECOVERED; | |
6a46079c AK |
511 | } |
512 | ||
513 | /* | |
514 | * Huge pages. Needs work. | |
515 | * Issues: | |
516 | * No rmap support so we cannot find the original mapper. In theory could walk | |
517 | * all MMs and look for the mappings, but that would be non atomic and racy. | |
518 | * Need rmap for hugepages for this. Alternatively we could employ a heuristic, | |
519 | * like just walking the current process and hoping it has it mapped (that | |
520 | * should be usually true for the common "shared database cache" case) | |
521 | * Should handle free huge pages and dequeue them too, but this needs to | |
522 | * handle huge page accounting correctly. | |
523 | */ | |
524 | static int me_huge_page(struct page *p, unsigned long pfn) | |
525 | { | |
526 | return FAILED; | |
527 | } | |
528 | ||
529 | /* | |
530 | * Various page states we can handle. | |
531 | * | |
532 | * A page state is defined by its current page->flags bits. | |
533 | * The table matches them in order and calls the right handler. | |
534 | * | |
535 | * This is quite tricky because we can access page at any time | |
536 | * in its live cycle, so all accesses have to be extremly careful. | |
537 | * | |
538 | * This is not complete. More states could be added. | |
539 | * For any missing state don't attempt recovery. | |
540 | */ | |
541 | ||
542 | #define dirty (1UL << PG_dirty) | |
543 | #define sc (1UL << PG_swapcache) | |
544 | #define unevict (1UL << PG_unevictable) | |
545 | #define mlock (1UL << PG_mlocked) | |
546 | #define writeback (1UL << PG_writeback) | |
547 | #define lru (1UL << PG_lru) | |
548 | #define swapbacked (1UL << PG_swapbacked) | |
549 | #define head (1UL << PG_head) | |
550 | #define tail (1UL << PG_tail) | |
551 | #define compound (1UL << PG_compound) | |
552 | #define slab (1UL << PG_slab) | |
553 | #define buddy (1UL << PG_buddy) | |
554 | #define reserved (1UL << PG_reserved) | |
555 | ||
556 | static struct page_state { | |
557 | unsigned long mask; | |
558 | unsigned long res; | |
559 | char *msg; | |
560 | int (*action)(struct page *p, unsigned long pfn); | |
561 | } error_states[] = { | |
562 | { reserved, reserved, "reserved kernel", me_ignore }, | |
563 | { buddy, buddy, "free kernel", me_free }, | |
564 | ||
565 | /* | |
566 | * Could in theory check if slab page is free or if we can drop | |
567 | * currently unused objects without touching them. But just | |
568 | * treat it as standard kernel for now. | |
569 | */ | |
570 | { slab, slab, "kernel slab", me_kernel }, | |
571 | ||
572 | #ifdef CONFIG_PAGEFLAGS_EXTENDED | |
573 | { head, head, "huge", me_huge_page }, | |
574 | { tail, tail, "huge", me_huge_page }, | |
575 | #else | |
576 | { compound, compound, "huge", me_huge_page }, | |
577 | #endif | |
578 | ||
579 | { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty }, | |
580 | { sc|dirty, sc, "swapcache", me_swapcache_clean }, | |
581 | ||
582 | { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty}, | |
583 | { unevict, unevict, "unevictable LRU", me_pagecache_clean}, | |
584 | ||
585 | #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT | |
586 | { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty }, | |
587 | { mlock, mlock, "mlocked LRU", me_pagecache_clean }, | |
588 | #endif | |
589 | ||
590 | { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty }, | |
591 | { lru|dirty, lru, "clean LRU", me_pagecache_clean }, | |
592 | { swapbacked, swapbacked, "anonymous", me_pagecache_clean }, | |
593 | ||
594 | /* | |
595 | * Catchall entry: must be at end. | |
596 | */ | |
597 | { 0, 0, "unknown page state", me_unknown }, | |
598 | }; | |
599 | ||
6a46079c AK |
600 | static void action_result(unsigned long pfn, char *msg, int result) |
601 | { | |
602 | struct page *page = NULL; | |
603 | if (pfn_valid(pfn)) | |
604 | page = pfn_to_page(pfn); | |
605 | ||
606 | printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n", | |
607 | pfn, | |
608 | page && PageDirty(page) ? "dirty " : "", | |
609 | msg, action_name[result]); | |
610 | } | |
611 | ||
612 | static int page_action(struct page_state *ps, struct page *p, | |
613 | unsigned long pfn, int ref) | |
614 | { | |
615 | int result; | |
7456b040 | 616 | int count; |
6a46079c AK |
617 | |
618 | result = ps->action(p, pfn); | |
619 | action_result(pfn, ps->msg, result); | |
7456b040 WF |
620 | |
621 | count = page_count(p) - 1 - ref; | |
622 | if (count != 0) | |
6a46079c AK |
623 | printk(KERN_ERR |
624 | "MCE %#lx: %s page still referenced by %d users\n", | |
7456b040 | 625 | pfn, ps->msg, count); |
6a46079c AK |
626 | |
627 | /* Could do more checks here if page looks ok */ | |
628 | /* | |
629 | * Could adjust zone counters here to correct for the missing page. | |
630 | */ | |
631 | ||
632 | return result == RECOVERED ? 0 : -EBUSY; | |
633 | } | |
634 | ||
635 | #define N_UNMAP_TRIES 5 | |
636 | ||
637 | /* | |
638 | * Do all that is necessary to remove user space mappings. Unmap | |
639 | * the pages and send SIGBUS to the processes if the data was dirty. | |
640 | */ | |
641 | static void hwpoison_user_mappings(struct page *p, unsigned long pfn, | |
642 | int trapno) | |
643 | { | |
644 | enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
645 | struct address_space *mapping; | |
646 | LIST_HEAD(tokill); | |
647 | int ret; | |
648 | int i; | |
649 | int kill = 1; | |
650 | ||
01e00f88 | 651 | if (PageReserved(p) || PageCompound(p) || PageSlab(p) || PageKsm(p)) |
6a46079c AK |
652 | return; |
653 | ||
6a46079c AK |
654 | /* |
655 | * This check implies we don't kill processes if their pages | |
656 | * are in the swap cache early. Those are always late kills. | |
657 | */ | |
658 | if (!page_mapped(p)) | |
659 | return; | |
660 | ||
661 | if (PageSwapCache(p)) { | |
662 | printk(KERN_ERR | |
663 | "MCE %#lx: keeping poisoned page in swap cache\n", pfn); | |
664 | ttu |= TTU_IGNORE_HWPOISON; | |
665 | } | |
666 | ||
667 | /* | |
668 | * Propagate the dirty bit from PTEs to struct page first, because we | |
669 | * need this to decide if we should kill or just drop the page. | |
670 | */ | |
671 | mapping = page_mapping(p); | |
672 | if (!PageDirty(p) && mapping && mapping_cap_writeback_dirty(mapping)) { | |
673 | if (page_mkclean(p)) { | |
674 | SetPageDirty(p); | |
675 | } else { | |
676 | kill = 0; | |
677 | ttu |= TTU_IGNORE_HWPOISON; | |
678 | printk(KERN_INFO | |
679 | "MCE %#lx: corrupted page was clean: dropped without side effects\n", | |
680 | pfn); | |
681 | } | |
682 | } | |
683 | ||
684 | /* | |
685 | * First collect all the processes that have the page | |
686 | * mapped in dirty form. This has to be done before try_to_unmap, | |
687 | * because ttu takes the rmap data structures down. | |
688 | * | |
689 | * Error handling: We ignore errors here because | |
690 | * there's nothing that can be done. | |
691 | */ | |
692 | if (kill) | |
693 | collect_procs(p, &tokill); | |
694 | ||
695 | /* | |
696 | * try_to_unmap can fail temporarily due to races. | |
697 | * Try a few times (RED-PEN better strategy?) | |
698 | */ | |
699 | for (i = 0; i < N_UNMAP_TRIES; i++) { | |
700 | ret = try_to_unmap(p, ttu); | |
701 | if (ret == SWAP_SUCCESS) | |
702 | break; | |
703 | pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn, ret); | |
704 | } | |
705 | ||
706 | if (ret != SWAP_SUCCESS) | |
707 | printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n", | |
708 | pfn, page_mapcount(p)); | |
709 | ||
710 | /* | |
711 | * Now that the dirty bit has been propagated to the | |
712 | * struct page and all unmaps done we can decide if | |
713 | * killing is needed or not. Only kill when the page | |
714 | * was dirty, otherwise the tokill list is merely | |
715 | * freed. When there was a problem unmapping earlier | |
716 | * use a more force-full uncatchable kill to prevent | |
717 | * any accesses to the poisoned memory. | |
718 | */ | |
719 | kill_procs_ao(&tokill, !!PageDirty(p), trapno, | |
720 | ret != SWAP_SUCCESS, pfn); | |
721 | } | |
722 | ||
723 | int __memory_failure(unsigned long pfn, int trapno, int ref) | |
724 | { | |
e43c3afb | 725 | unsigned long lru_flag; |
6a46079c AK |
726 | struct page_state *ps; |
727 | struct page *p; | |
728 | int res; | |
729 | ||
730 | if (!sysctl_memory_failure_recovery) | |
731 | panic("Memory failure from trap %d on page %lx", trapno, pfn); | |
732 | ||
733 | if (!pfn_valid(pfn)) { | |
734 | action_result(pfn, "memory outside kernel control", IGNORED); | |
735 | return -EIO; | |
736 | } | |
737 | ||
738 | p = pfn_to_page(pfn); | |
739 | if (TestSetPageHWPoison(p)) { | |
740 | action_result(pfn, "already hardware poisoned", IGNORED); | |
741 | return 0; | |
742 | } | |
743 | ||
744 | atomic_long_add(1, &mce_bad_pages); | |
745 | ||
746 | /* | |
747 | * We need/can do nothing about count=0 pages. | |
748 | * 1) it's a free page, and therefore in safe hand: | |
749 | * prep_new_page() will be the gate keeper. | |
750 | * 2) it's part of a non-compound high order page. | |
751 | * Implies some kernel user: cannot stop them from | |
752 | * R/W the page; let's pray that the page has been | |
753 | * used and will be freed some time later. | |
754 | * In fact it's dangerous to directly bump up page count from 0, | |
755 | * that may make page_freeze_refs()/page_unfreeze_refs() mismatch. | |
756 | */ | |
757 | if (!get_page_unless_zero(compound_head(p))) { | |
758 | action_result(pfn, "free or high order kernel", IGNORED); | |
759 | return PageBuddy(compound_head(p)) ? 0 : -EBUSY; | |
760 | } | |
761 | ||
e43c3afb WF |
762 | /* |
763 | * We ignore non-LRU pages for good reasons. | |
764 | * - PG_locked is only well defined for LRU pages and a few others | |
765 | * - to avoid races with __set_page_locked() | |
766 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) | |
767 | * The check (unnecessarily) ignores LRU pages being isolated and | |
768 | * walked by the page reclaim code, however that's not a big loss. | |
769 | */ | |
770 | if (!PageLRU(p)) | |
771 | lru_add_drain_all(); | |
772 | lru_flag = p->flags & lru; | |
773 | if (isolate_lru_page(p)) { | |
774 | action_result(pfn, "non LRU", IGNORED); | |
775 | put_page(p); | |
776 | return -EBUSY; | |
777 | } | |
778 | page_cache_release(p); | |
779 | ||
6a46079c AK |
780 | /* |
781 | * Lock the page and wait for writeback to finish. | |
782 | * It's very difficult to mess with pages currently under IO | |
783 | * and in many cases impossible, so we just avoid it here. | |
784 | */ | |
785 | lock_page_nosync(p); | |
786 | wait_on_page_writeback(p); | |
787 | ||
788 | /* | |
789 | * Now take care of user space mappings. | |
790 | */ | |
791 | hwpoison_user_mappings(p, pfn, trapno); | |
792 | ||
793 | /* | |
794 | * Torn down by someone else? | |
795 | */ | |
e43c3afb | 796 | if ((lru_flag & lru) && !PageSwapCache(p) && p->mapping == NULL) { |
6a46079c AK |
797 | action_result(pfn, "already truncated LRU", IGNORED); |
798 | res = 0; | |
799 | goto out; | |
800 | } | |
801 | ||
802 | res = -EBUSY; | |
803 | for (ps = error_states;; ps++) { | |
e43c3afb | 804 | if (((p->flags | lru_flag)& ps->mask) == ps->res) { |
6a46079c AK |
805 | res = page_action(ps, p, pfn, ref); |
806 | break; | |
807 | } | |
808 | } | |
809 | out: | |
810 | unlock_page(p); | |
811 | return res; | |
812 | } | |
813 | EXPORT_SYMBOL_GPL(__memory_failure); | |
814 | ||
815 | /** | |
816 | * memory_failure - Handle memory failure of a page. | |
817 | * @pfn: Page Number of the corrupted page | |
818 | * @trapno: Trap number reported in the signal to user space. | |
819 | * | |
820 | * This function is called by the low level machine check code | |
821 | * of an architecture when it detects hardware memory corruption | |
822 | * of a page. It tries its best to recover, which includes | |
823 | * dropping pages, killing processes etc. | |
824 | * | |
825 | * The function is primarily of use for corruptions that | |
826 | * happen outside the current execution context (e.g. when | |
827 | * detected by a background scrubber) | |
828 | * | |
829 | * Must run in process context (e.g. a work queue) with interrupts | |
830 | * enabled and no spinlocks hold. | |
831 | */ | |
832 | void memory_failure(unsigned long pfn, int trapno) | |
833 | { | |
834 | __memory_failure(pfn, trapno, 0); | |
835 | } |