Root/
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 multi-bit ECC memory or cache |
11 | * failure. |
12 | * |
13 | * In addition there is a "soft offline" entry point that allows stop using |
14 | * not-yet-corrupted-by-suspicious pages without killing anything. |
15 | * |
16 | * Handles page cache pages in various states. The tricky part |
17 | * here is that we can access any page asynchronously in respect to |
18 | * other VM users, because memory failures could happen anytime and |
19 | * anywhere. This could violate some of their assumptions. This is why |
20 | * this code has to be extremely careful. Generally it tries to use |
21 | * normal locking rules, as in get the standard locks, even if that means |
22 | * the error handling takes potentially a long time. |
23 | * |
24 | * There are several operations here with exponential complexity because |
25 | * of unsuitable VM data structures. For example the operation to map back |
26 | * from RMAP chains to processes has to walk the complete process list and |
27 | * has non linear complexity with the number. But since memory corruptions |
28 | * are rare we hope to get away with this. This avoids impacting the core |
29 | * VM. |
30 | */ |
31 | |
32 | /* |
33 | * Notebook: |
34 | * - hugetlb needs more code |
35 | * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages |
36 | * - pass bad pages to kdump next kernel |
37 | */ |
38 | #include <linux/kernel.h> |
39 | #include <linux/mm.h> |
40 | #include <linux/page-flags.h> |
41 | #include <linux/kernel-page-flags.h> |
42 | #include <linux/sched.h> |
43 | #include <linux/ksm.h> |
44 | #include <linux/rmap.h> |
45 | #include <linux/export.h> |
46 | #include <linux/pagemap.h> |
47 | #include <linux/swap.h> |
48 | #include <linux/backing-dev.h> |
49 | #include <linux/migrate.h> |
50 | #include <linux/page-isolation.h> |
51 | #include <linux/suspend.h> |
52 | #include <linux/slab.h> |
53 | #include <linux/swapops.h> |
54 | #include <linux/hugetlb.h> |
55 | #include <linux/memory_hotplug.h> |
56 | #include <linux/mm_inline.h> |
57 | #include <linux/kfifo.h> |
58 | #include "internal.h" |
59 | |
60 | int sysctl_memory_failure_early_kill __read_mostly = 0; |
61 | |
62 | int sysctl_memory_failure_recovery __read_mostly = 1; |
63 | |
64 | atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0); |
65 | |
66 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
67 | |
68 | u32 hwpoison_filter_enable = 0; |
69 | u32 hwpoison_filter_dev_major = ~0U; |
70 | u32 hwpoison_filter_dev_minor = ~0U; |
71 | u64 hwpoison_filter_flags_mask; |
72 | u64 hwpoison_filter_flags_value; |
73 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
74 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
75 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); |
76 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
77 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); |
78 | |
79 | static int hwpoison_filter_dev(struct page *p) |
80 | { |
81 | struct address_space *mapping; |
82 | dev_t dev; |
83 | |
84 | if (hwpoison_filter_dev_major == ~0U && |
85 | hwpoison_filter_dev_minor == ~0U) |
86 | return 0; |
87 | |
88 | /* |
89 | * page_mapping() does not accept slab pages. |
90 | */ |
91 | if (PageSlab(p)) |
92 | return -EINVAL; |
93 | |
94 | mapping = page_mapping(p); |
95 | if (mapping == NULL || mapping->host == NULL) |
96 | return -EINVAL; |
97 | |
98 | dev = mapping->host->i_sb->s_dev; |
99 | if (hwpoison_filter_dev_major != ~0U && |
100 | hwpoison_filter_dev_major != MAJOR(dev)) |
101 | return -EINVAL; |
102 | if (hwpoison_filter_dev_minor != ~0U && |
103 | hwpoison_filter_dev_minor != MINOR(dev)) |
104 | return -EINVAL; |
105 | |
106 | return 0; |
107 | } |
108 | |
109 | static int hwpoison_filter_flags(struct page *p) |
110 | { |
111 | if (!hwpoison_filter_flags_mask) |
112 | return 0; |
113 | |
114 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == |
115 | hwpoison_filter_flags_value) |
116 | return 0; |
117 | else |
118 | return -EINVAL; |
119 | } |
120 | |
121 | /* |
122 | * This allows stress tests to limit test scope to a collection of tasks |
123 | * by putting them under some memcg. This prevents killing unrelated/important |
124 | * processes such as /sbin/init. Note that the target task may share clean |
125 | * pages with init (eg. libc text), which is harmless. If the target task |
126 | * share _dirty_ pages with another task B, the test scheme must make sure B |
127 | * is also included in the memcg. At last, due to race conditions this filter |
128 | * can only guarantee that the page either belongs to the memcg tasks, or is |
129 | * a freed page. |
130 | */ |
131 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
132 | u64 hwpoison_filter_memcg; |
133 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); |
134 | static int hwpoison_filter_task(struct page *p) |
135 | { |
136 | struct mem_cgroup *mem; |
137 | struct cgroup_subsys_state *css; |
138 | unsigned long ino; |
139 | |
140 | if (!hwpoison_filter_memcg) |
141 | return 0; |
142 | |
143 | mem = try_get_mem_cgroup_from_page(p); |
144 | if (!mem) |
145 | return -EINVAL; |
146 | |
147 | css = mem_cgroup_css(mem); |
148 | /* root_mem_cgroup has NULL dentries */ |
149 | if (!css->cgroup->dentry) |
150 | return -EINVAL; |
151 | |
152 | ino = css->cgroup->dentry->d_inode->i_ino; |
153 | css_put(css); |
154 | |
155 | if (ino != hwpoison_filter_memcg) |
156 | return -EINVAL; |
157 | |
158 | return 0; |
159 | } |
160 | #else |
161 | static int hwpoison_filter_task(struct page *p) { return 0; } |
162 | #endif |
163 | |
164 | int hwpoison_filter(struct page *p) |
165 | { |
166 | if (!hwpoison_filter_enable) |
167 | return 0; |
168 | |
169 | if (hwpoison_filter_dev(p)) |
170 | return -EINVAL; |
171 | |
172 | if (hwpoison_filter_flags(p)) |
173 | return -EINVAL; |
174 | |
175 | if (hwpoison_filter_task(p)) |
176 | return -EINVAL; |
177 | |
178 | return 0; |
179 | } |
180 | #else |
181 | int hwpoison_filter(struct page *p) |
182 | { |
183 | return 0; |
184 | } |
185 | #endif |
186 | |
187 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
188 | |
189 | /* |
190 | * Send all the processes who have the page mapped an ``action optional'' |
191 | * signal. |
192 | */ |
193 | static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno, |
194 | unsigned long pfn, struct page *page) |
195 | { |
196 | struct siginfo si; |
197 | int ret; |
198 | |
199 | printk(KERN_ERR |
200 | "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n", |
201 | pfn, t->comm, t->pid); |
202 | si.si_signo = SIGBUS; |
203 | si.si_errno = 0; |
204 | si.si_code = BUS_MCEERR_AO; |
205 | si.si_addr = (void *)addr; |
206 | #ifdef __ARCH_SI_TRAPNO |
207 | si.si_trapno = trapno; |
208 | #endif |
209 | si.si_addr_lsb = compound_trans_order(compound_head(page)) + PAGE_SHIFT; |
210 | /* |
211 | * Don't use force here, it's convenient if the signal |
212 | * can be temporarily blocked. |
213 | * This could cause a loop when the user sets SIGBUS |
214 | * to SIG_IGN, but hopefully no one will do that? |
215 | */ |
216 | ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */ |
217 | if (ret < 0) |
218 | printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n", |
219 | t->comm, t->pid, ret); |
220 | return ret; |
221 | } |
222 | |
223 | /* |
224 | * When a unknown page type is encountered drain as many buffers as possible |
225 | * in the hope to turn the page into a LRU or free page, which we can handle. |
226 | */ |
227 | void shake_page(struct page *p, int access) |
228 | { |
229 | if (!PageSlab(p)) { |
230 | lru_add_drain_all(); |
231 | if (PageLRU(p)) |
232 | return; |
233 | drain_all_pages(); |
234 | if (PageLRU(p) || is_free_buddy_page(p)) |
235 | return; |
236 | } |
237 | |
238 | /* |
239 | * Only call shrink_slab here (which would also shrink other caches) if |
240 | * access is not potentially fatal. |
241 | */ |
242 | if (access) { |
243 | int nr; |
244 | do { |
245 | struct shrink_control shrink = { |
246 | .gfp_mask = GFP_KERNEL, |
247 | }; |
248 | |
249 | nr = shrink_slab(&shrink, 1000, 1000); |
250 | if (page_count(p) == 1) |
251 | break; |
252 | } while (nr > 10); |
253 | } |
254 | } |
255 | EXPORT_SYMBOL_GPL(shake_page); |
256 | |
257 | /* |
258 | * Kill all processes that have a poisoned page mapped and then isolate |
259 | * the page. |
260 | * |
261 | * General strategy: |
262 | * Find all processes having the page mapped and kill them. |
263 | * But we keep a page reference around so that the page is not |
264 | * actually freed yet. |
265 | * Then stash the page away |
266 | * |
267 | * There's no convenient way to get back to mapped processes |
268 | * from the VMAs. So do a brute-force search over all |
269 | * running processes. |
270 | * |
271 | * Remember that machine checks are not common (or rather |
272 | * if they are common you have other problems), so this shouldn't |
273 | * be a performance issue. |
274 | * |
275 | * Also there are some races possible while we get from the |
276 | * error detection to actually handle it. |
277 | */ |
278 | |
279 | struct to_kill { |
280 | struct list_head nd; |
281 | struct task_struct *tsk; |
282 | unsigned long addr; |
283 | char addr_valid; |
284 | }; |
285 | |
286 | /* |
287 | * Failure handling: if we can't find or can't kill a process there's |
288 | * not much we can do. We just print a message and ignore otherwise. |
289 | */ |
290 | |
291 | /* |
292 | * Schedule a process for later kill. |
293 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. |
294 | * TBD would GFP_NOIO be enough? |
295 | */ |
296 | static void add_to_kill(struct task_struct *tsk, struct page *p, |
297 | struct vm_area_struct *vma, |
298 | struct list_head *to_kill, |
299 | struct to_kill **tkc) |
300 | { |
301 | struct to_kill *tk; |
302 | |
303 | if (*tkc) { |
304 | tk = *tkc; |
305 | *tkc = NULL; |
306 | } else { |
307 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); |
308 | if (!tk) { |
309 | printk(KERN_ERR |
310 | "MCE: Out of memory while machine check handling\n"); |
311 | return; |
312 | } |
313 | } |
314 | tk->addr = page_address_in_vma(p, vma); |
315 | tk->addr_valid = 1; |
316 | |
317 | /* |
318 | * In theory we don't have to kill when the page was |
319 | * munmaped. But it could be also a mremap. Since that's |
320 | * likely very rare kill anyways just out of paranoia, but use |
321 | * a SIGKILL because the error is not contained anymore. |
322 | */ |
323 | if (tk->addr == -EFAULT) { |
324 | pr_info("MCE: Unable to find user space address %lx in %s\n", |
325 | page_to_pfn(p), tsk->comm); |
326 | tk->addr_valid = 0; |
327 | } |
328 | get_task_struct(tsk); |
329 | tk->tsk = tsk; |
330 | list_add_tail(&tk->nd, to_kill); |
331 | } |
332 | |
333 | /* |
334 | * Kill the processes that have been collected earlier. |
335 | * |
336 | * Only do anything when DOIT is set, otherwise just free the list |
337 | * (this is used for clean pages which do not need killing) |
338 | * Also when FAIL is set do a force kill because something went |
339 | * wrong earlier. |
340 | */ |
341 | static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno, |
342 | int fail, struct page *page, unsigned long pfn) |
343 | { |
344 | struct to_kill *tk, *next; |
345 | |
346 | list_for_each_entry_safe (tk, next, to_kill, nd) { |
347 | if (doit) { |
348 | /* |
349 | * In case something went wrong with munmapping |
350 | * make sure the process doesn't catch the |
351 | * signal and then access the memory. Just kill it. |
352 | */ |
353 | if (fail || tk->addr_valid == 0) { |
354 | printk(KERN_ERR |
355 | "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", |
356 | pfn, tk->tsk->comm, tk->tsk->pid); |
357 | force_sig(SIGKILL, tk->tsk); |
358 | } |
359 | |
360 | /* |
361 | * In theory the process could have mapped |
362 | * something else on the address in-between. We could |
363 | * check for that, but we need to tell the |
364 | * process anyways. |
365 | */ |
366 | else if (kill_proc_ao(tk->tsk, tk->addr, trapno, |
367 | pfn, page) < 0) |
368 | printk(KERN_ERR |
369 | "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n", |
370 | pfn, tk->tsk->comm, tk->tsk->pid); |
371 | } |
372 | put_task_struct(tk->tsk); |
373 | kfree(tk); |
374 | } |
375 | } |
376 | |
377 | static int task_early_kill(struct task_struct *tsk) |
378 | { |
379 | if (!tsk->mm) |
380 | return 0; |
381 | if (tsk->flags & PF_MCE_PROCESS) |
382 | return !!(tsk->flags & PF_MCE_EARLY); |
383 | return sysctl_memory_failure_early_kill; |
384 | } |
385 | |
386 | /* |
387 | * Collect processes when the error hit an anonymous page. |
388 | */ |
389 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, |
390 | struct to_kill **tkc) |
391 | { |
392 | struct vm_area_struct *vma; |
393 | struct task_struct *tsk; |
394 | struct anon_vma *av; |
395 | |
396 | av = page_lock_anon_vma(page); |
397 | if (av == NULL) /* Not actually mapped anymore */ |
398 | return; |
399 | |
400 | read_lock(&tasklist_lock); |
401 | for_each_process (tsk) { |
402 | struct anon_vma_chain *vmac; |
403 | |
404 | if (!task_early_kill(tsk)) |
405 | continue; |
406 | list_for_each_entry(vmac, &av->head, same_anon_vma) { |
407 | vma = vmac->vma; |
408 | if (!page_mapped_in_vma(page, vma)) |
409 | continue; |
410 | if (vma->vm_mm == tsk->mm) |
411 | add_to_kill(tsk, page, vma, to_kill, tkc); |
412 | } |
413 | } |
414 | read_unlock(&tasklist_lock); |
415 | page_unlock_anon_vma(av); |
416 | } |
417 | |
418 | /* |
419 | * Collect processes when the error hit a file mapped page. |
420 | */ |
421 | static void collect_procs_file(struct page *page, struct list_head *to_kill, |
422 | struct to_kill **tkc) |
423 | { |
424 | struct vm_area_struct *vma; |
425 | struct task_struct *tsk; |
426 | struct prio_tree_iter iter; |
427 | struct address_space *mapping = page->mapping; |
428 | |
429 | mutex_lock(&mapping->i_mmap_mutex); |
430 | read_lock(&tasklist_lock); |
431 | for_each_process(tsk) { |
432 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
433 | |
434 | if (!task_early_kill(tsk)) |
435 | continue; |
436 | |
437 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, |
438 | pgoff) { |
439 | /* |
440 | * Send early kill signal to tasks where a vma covers |
441 | * the page but the corrupted page is not necessarily |
442 | * mapped it in its pte. |
443 | * Assume applications who requested early kill want |
444 | * to be informed of all such data corruptions. |
445 | */ |
446 | if (vma->vm_mm == tsk->mm) |
447 | add_to_kill(tsk, page, vma, to_kill, tkc); |
448 | } |
449 | } |
450 | read_unlock(&tasklist_lock); |
451 | mutex_unlock(&mapping->i_mmap_mutex); |
452 | } |
453 | |
454 | /* |
455 | * Collect the processes who have the corrupted page mapped to kill. |
456 | * This is done in two steps for locking reasons. |
457 | * First preallocate one tokill structure outside the spin locks, |
458 | * so that we can kill at least one process reasonably reliable. |
459 | */ |
460 | static void collect_procs(struct page *page, struct list_head *tokill) |
461 | { |
462 | struct to_kill *tk; |
463 | |
464 | if (!page->mapping) |
465 | return; |
466 | |
467 | tk = kmalloc(sizeof(struct to_kill), GFP_NOIO); |
468 | if (!tk) |
469 | return; |
470 | if (PageAnon(page)) |
471 | collect_procs_anon(page, tokill, &tk); |
472 | else |
473 | collect_procs_file(page, tokill, &tk); |
474 | kfree(tk); |
475 | } |
476 | |
477 | /* |
478 | * Error handlers for various types of pages. |
479 | */ |
480 | |
481 | enum outcome { |
482 | IGNORED, /* Error: cannot be handled */ |
483 | FAILED, /* Error: handling failed */ |
484 | DELAYED, /* Will be handled later */ |
485 | RECOVERED, /* Successfully recovered */ |
486 | }; |
487 | |
488 | static const char *action_name[] = { |
489 | [IGNORED] = "Ignored", |
490 | [FAILED] = "Failed", |
491 | [DELAYED] = "Delayed", |
492 | [RECOVERED] = "Recovered", |
493 | }; |
494 | |
495 | /* |
496 | * XXX: It is possible that a page is isolated from LRU cache, |
497 | * and then kept in swap cache or failed to remove from page cache. |
498 | * The page count will stop it from being freed by unpoison. |
499 | * Stress tests should be aware of this memory leak problem. |
500 | */ |
501 | static int delete_from_lru_cache(struct page *p) |
502 | { |
503 | if (!isolate_lru_page(p)) { |
504 | /* |
505 | * Clear sensible page flags, so that the buddy system won't |
506 | * complain when the page is unpoison-and-freed. |
507 | */ |
508 | ClearPageActive(p); |
509 | ClearPageUnevictable(p); |
510 | /* |
511 | * drop the page count elevated by isolate_lru_page() |
512 | */ |
513 | page_cache_release(p); |
514 | return 0; |
515 | } |
516 | return -EIO; |
517 | } |
518 | |
519 | /* |
520 | * Error hit kernel page. |
521 | * Do nothing, try to be lucky and not touch this instead. For a few cases we |
522 | * could be more sophisticated. |
523 | */ |
524 | static int me_kernel(struct page *p, unsigned long pfn) |
525 | { |
526 | return IGNORED; |
527 | } |
528 | |
529 | /* |
530 | * Page in unknown state. Do nothing. |
531 | */ |
532 | static int me_unknown(struct page *p, unsigned long pfn) |
533 | { |
534 | printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn); |
535 | return FAILED; |
536 | } |
537 | |
538 | /* |
539 | * Clean (or cleaned) page cache page. |
540 | */ |
541 | static int me_pagecache_clean(struct page *p, unsigned long pfn) |
542 | { |
543 | int err; |
544 | int ret = FAILED; |
545 | struct address_space *mapping; |
546 | |
547 | delete_from_lru_cache(p); |
548 | |
549 | /* |
550 | * For anonymous pages we're done the only reference left |
551 | * should be the one m_f() holds. |
552 | */ |
553 | if (PageAnon(p)) |
554 | return RECOVERED; |
555 | |
556 | /* |
557 | * Now truncate the page in the page cache. This is really |
558 | * more like a "temporary hole punch" |
559 | * Don't do this for block devices when someone else |
560 | * has a reference, because it could be file system metadata |
561 | * and that's not safe to truncate. |
562 | */ |
563 | mapping = page_mapping(p); |
564 | if (!mapping) { |
565 | /* |
566 | * Page has been teared down in the meanwhile |
567 | */ |
568 | return FAILED; |
569 | } |
570 | |
571 | /* |
572 | * Truncation is a bit tricky. Enable it per file system for now. |
573 | * |
574 | * Open: to take i_mutex or not for this? Right now we don't. |
575 | */ |
576 | if (mapping->a_ops->error_remove_page) { |
577 | err = mapping->a_ops->error_remove_page(mapping, p); |
578 | if (err != 0) { |
579 | printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n", |
580 | pfn, err); |
581 | } else if (page_has_private(p) && |
582 | !try_to_release_page(p, GFP_NOIO)) { |
583 | pr_info("MCE %#lx: failed to release buffers\n", pfn); |
584 | } else { |
585 | ret = RECOVERED; |
586 | } |
587 | } else { |
588 | /* |
589 | * If the file system doesn't support it just invalidate |
590 | * This fails on dirty or anything with private pages |
591 | */ |
592 | if (invalidate_inode_page(p)) |
593 | ret = RECOVERED; |
594 | else |
595 | printk(KERN_INFO "MCE %#lx: Failed to invalidate\n", |
596 | pfn); |
597 | } |
598 | return ret; |
599 | } |
600 | |
601 | /* |
602 | * Dirty cache page page |
603 | * Issues: when the error hit a hole page the error is not properly |
604 | * propagated. |
605 | */ |
606 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) |
607 | { |
608 | struct address_space *mapping = page_mapping(p); |
609 | |
610 | SetPageError(p); |
611 | /* TBD: print more information about the file. */ |
612 | if (mapping) { |
613 | /* |
614 | * IO error will be reported by write(), fsync(), etc. |
615 | * who check the mapping. |
616 | * This way the application knows that something went |
617 | * wrong with its dirty file data. |
618 | * |
619 | * There's one open issue: |
620 | * |
621 | * The EIO will be only reported on the next IO |
622 | * operation and then cleared through the IO map. |
623 | * Normally Linux has two mechanisms to pass IO error |
624 | * first through the AS_EIO flag in the address space |
625 | * and then through the PageError flag in the page. |
626 | * Since we drop pages on memory failure handling the |
627 | * only mechanism open to use is through AS_AIO. |
628 | * |
629 | * This has the disadvantage that it gets cleared on |
630 | * the first operation that returns an error, while |
631 | * the PageError bit is more sticky and only cleared |
632 | * when the page is reread or dropped. If an |
633 | * application assumes it will always get error on |
634 | * fsync, but does other operations on the fd before |
635 | * and the page is dropped between then the error |
636 | * will not be properly reported. |
637 | * |
638 | * This can already happen even without hwpoisoned |
639 | * pages: first on metadata IO errors (which only |
640 | * report through AS_EIO) or when the page is dropped |
641 | * at the wrong time. |
642 | * |
643 | * So right now we assume that the application DTRT on |
644 | * the first EIO, but we're not worse than other parts |
645 | * of the kernel. |
646 | */ |
647 | mapping_set_error(mapping, EIO); |
648 | } |
649 | |
650 | return me_pagecache_clean(p, pfn); |
651 | } |
652 | |
653 | /* |
654 | * Clean and dirty swap cache. |
655 | * |
656 | * Dirty swap cache page is tricky to handle. The page could live both in page |
657 | * cache and swap cache(ie. page is freshly swapped in). So it could be |
658 | * referenced concurrently by 2 types of PTEs: |
659 | * normal PTEs and swap PTEs. We try to handle them consistently by calling |
660 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, |
661 | * and then |
662 | * - clear dirty bit to prevent IO |
663 | * - remove from LRU |
664 | * - but keep in the swap cache, so that when we return to it on |
665 | * a later page fault, we know the application is accessing |
666 | * corrupted data and shall be killed (we installed simple |
667 | * interception code in do_swap_page to catch it). |
668 | * |
669 | * Clean swap cache pages can be directly isolated. A later page fault will |
670 | * bring in the known good data from disk. |
671 | */ |
672 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) |
673 | { |
674 | ClearPageDirty(p); |
675 | /* Trigger EIO in shmem: */ |
676 | ClearPageUptodate(p); |
677 | |
678 | if (!delete_from_lru_cache(p)) |
679 | return DELAYED; |
680 | else |
681 | return FAILED; |
682 | } |
683 | |
684 | static int me_swapcache_clean(struct page *p, unsigned long pfn) |
685 | { |
686 | delete_from_swap_cache(p); |
687 | |
688 | if (!delete_from_lru_cache(p)) |
689 | return RECOVERED; |
690 | else |
691 | return FAILED; |
692 | } |
693 | |
694 | /* |
695 | * Huge pages. Needs work. |
696 | * Issues: |
697 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
698 | * To narrow down kill region to one page, we need to break up pmd. |
699 | */ |
700 | static int me_huge_page(struct page *p, unsigned long pfn) |
701 | { |
702 | int res = 0; |
703 | struct page *hpage = compound_head(p); |
704 | /* |
705 | * We can safely recover from error on free or reserved (i.e. |
706 | * not in-use) hugepage by dequeuing it from freelist. |
707 | * To check whether a hugepage is in-use or not, we can't use |
708 | * page->lru because it can be used in other hugepage operations, |
709 | * such as __unmap_hugepage_range() and gather_surplus_pages(). |
710 | * So instead we use page_mapping() and PageAnon(). |
711 | * We assume that this function is called with page lock held, |
712 | * so there is no race between isolation and mapping/unmapping. |
713 | */ |
714 | if (!(page_mapping(hpage) || PageAnon(hpage))) { |
715 | res = dequeue_hwpoisoned_huge_page(hpage); |
716 | if (!res) |
717 | return RECOVERED; |
718 | } |
719 | return DELAYED; |
720 | } |
721 | |
722 | /* |
723 | * Various page states we can handle. |
724 | * |
725 | * A page state is defined by its current page->flags bits. |
726 | * The table matches them in order and calls the right handler. |
727 | * |
728 | * This is quite tricky because we can access page at any time |
729 | * in its live cycle, so all accesses have to be extremely careful. |
730 | * |
731 | * This is not complete. More states could be added. |
732 | * For any missing state don't attempt recovery. |
733 | */ |
734 | |
735 | #define dirty (1UL << PG_dirty) |
736 | #define sc (1UL << PG_swapcache) |
737 | #define unevict (1UL << PG_unevictable) |
738 | #define mlock (1UL << PG_mlocked) |
739 | #define writeback (1UL << PG_writeback) |
740 | #define lru (1UL << PG_lru) |
741 | #define swapbacked (1UL << PG_swapbacked) |
742 | #define head (1UL << PG_head) |
743 | #define tail (1UL << PG_tail) |
744 | #define compound (1UL << PG_compound) |
745 | #define slab (1UL << PG_slab) |
746 | #define reserved (1UL << PG_reserved) |
747 | |
748 | static struct page_state { |
749 | unsigned long mask; |
750 | unsigned long res; |
751 | char *msg; |
752 | int (*action)(struct page *p, unsigned long pfn); |
753 | } error_states[] = { |
754 | { reserved, reserved, "reserved kernel", me_kernel }, |
755 | /* |
756 | * free pages are specially detected outside this table: |
757 | * PG_buddy pages only make a small fraction of all free pages. |
758 | */ |
759 | |
760 | /* |
761 | * Could in theory check if slab page is free or if we can drop |
762 | * currently unused objects without touching them. But just |
763 | * treat it as standard kernel for now. |
764 | */ |
765 | { slab, slab, "kernel slab", me_kernel }, |
766 | |
767 | #ifdef CONFIG_PAGEFLAGS_EXTENDED |
768 | { head, head, "huge", me_huge_page }, |
769 | { tail, tail, "huge", me_huge_page }, |
770 | #else |
771 | { compound, compound, "huge", me_huge_page }, |
772 | #endif |
773 | |
774 | { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty }, |
775 | { sc|dirty, sc, "swapcache", me_swapcache_clean }, |
776 | |
777 | { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty}, |
778 | { unevict, unevict, "unevictable LRU", me_pagecache_clean}, |
779 | |
780 | { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty }, |
781 | { mlock, mlock, "mlocked LRU", me_pagecache_clean }, |
782 | |
783 | { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty }, |
784 | { lru|dirty, lru, "clean LRU", me_pagecache_clean }, |
785 | |
786 | /* |
787 | * Catchall entry: must be at end. |
788 | */ |
789 | { 0, 0, "unknown page state", me_unknown }, |
790 | }; |
791 | |
792 | #undef dirty |
793 | #undef sc |
794 | #undef unevict |
795 | #undef mlock |
796 | #undef writeback |
797 | #undef lru |
798 | #undef swapbacked |
799 | #undef head |
800 | #undef tail |
801 | #undef compound |
802 | #undef slab |
803 | #undef reserved |
804 | |
805 | static void action_result(unsigned long pfn, char *msg, int result) |
806 | { |
807 | struct page *page = pfn_to_page(pfn); |
808 | |
809 | printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n", |
810 | pfn, |
811 | PageDirty(page) ? "dirty " : "", |
812 | msg, action_name[result]); |
813 | } |
814 | |
815 | static int page_action(struct page_state *ps, struct page *p, |
816 | unsigned long pfn) |
817 | { |
818 | int result; |
819 | int count; |
820 | |
821 | result = ps->action(p, pfn); |
822 | action_result(pfn, ps->msg, result); |
823 | |
824 | count = page_count(p) - 1; |
825 | if (ps->action == me_swapcache_dirty && result == DELAYED) |
826 | count--; |
827 | if (count != 0) { |
828 | printk(KERN_ERR |
829 | "MCE %#lx: %s page still referenced by %d users\n", |
830 | pfn, ps->msg, count); |
831 | result = FAILED; |
832 | } |
833 | |
834 | /* Could do more checks here if page looks ok */ |
835 | /* |
836 | * Could adjust zone counters here to correct for the missing page. |
837 | */ |
838 | |
839 | return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY; |
840 | } |
841 | |
842 | /* |
843 | * Do all that is necessary to remove user space mappings. Unmap |
844 | * the pages and send SIGBUS to the processes if the data was dirty. |
845 | */ |
846 | static int hwpoison_user_mappings(struct page *p, unsigned long pfn, |
847 | int trapno) |
848 | { |
849 | enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; |
850 | struct address_space *mapping; |
851 | LIST_HEAD(tokill); |
852 | int ret; |
853 | int kill = 1; |
854 | struct page *hpage = compound_head(p); |
855 | struct page *ppage; |
856 | |
857 | if (PageReserved(p) || PageSlab(p)) |
858 | return SWAP_SUCCESS; |
859 | |
860 | /* |
861 | * This check implies we don't kill processes if their pages |
862 | * are in the swap cache early. Those are always late kills. |
863 | */ |
864 | if (!page_mapped(hpage)) |
865 | return SWAP_SUCCESS; |
866 | |
867 | if (PageKsm(p)) |
868 | return SWAP_FAIL; |
869 | |
870 | if (PageSwapCache(p)) { |
871 | printk(KERN_ERR |
872 | "MCE %#lx: keeping poisoned page in swap cache\n", pfn); |
873 | ttu |= TTU_IGNORE_HWPOISON; |
874 | } |
875 | |
876 | /* |
877 | * Propagate the dirty bit from PTEs to struct page first, because we |
878 | * need this to decide if we should kill or just drop the page. |
879 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
880 | * be called inside page lock (it's recommended but not enforced). |
881 | */ |
882 | mapping = page_mapping(hpage); |
883 | if (!PageDirty(hpage) && mapping && |
884 | mapping_cap_writeback_dirty(mapping)) { |
885 | if (page_mkclean(hpage)) { |
886 | SetPageDirty(hpage); |
887 | } else { |
888 | kill = 0; |
889 | ttu |= TTU_IGNORE_HWPOISON; |
890 | printk(KERN_INFO |
891 | "MCE %#lx: corrupted page was clean: dropped without side effects\n", |
892 | pfn); |
893 | } |
894 | } |
895 | |
896 | /* |
897 | * ppage: poisoned page |
898 | * if p is regular page(4k page) |
899 | * ppage == real poisoned page; |
900 | * else p is hugetlb or THP, ppage == head page. |
901 | */ |
902 | ppage = hpage; |
903 | |
904 | if (PageTransHuge(hpage)) { |
905 | /* |
906 | * Verify that this isn't a hugetlbfs head page, the check for |
907 | * PageAnon is just for avoid tripping a split_huge_page |
908 | * internal debug check, as split_huge_page refuses to deal with |
909 | * anything that isn't an anon page. PageAnon can't go away fro |
910 | * under us because we hold a refcount on the hpage, without a |
911 | * refcount on the hpage. split_huge_page can't be safely called |
912 | * in the first place, having a refcount on the tail isn't |
913 | * enough * to be safe. |
914 | */ |
915 | if (!PageHuge(hpage) && PageAnon(hpage)) { |
916 | if (unlikely(split_huge_page(hpage))) { |
917 | /* |
918 | * FIXME: if splitting THP is failed, it is |
919 | * better to stop the following operation rather |
920 | * than causing panic by unmapping. System might |
921 | * survive if the page is freed later. |
922 | */ |
923 | printk(KERN_INFO |
924 | "MCE %#lx: failed to split THP\n", pfn); |
925 | |
926 | BUG_ON(!PageHWPoison(p)); |
927 | return SWAP_FAIL; |
928 | } |
929 | /* THP is split, so ppage should be the real poisoned page. */ |
930 | ppage = p; |
931 | } |
932 | } |
933 | |
934 | /* |
935 | * First collect all the processes that have the page |
936 | * mapped in dirty form. This has to be done before try_to_unmap, |
937 | * because ttu takes the rmap data structures down. |
938 | * |
939 | * Error handling: We ignore errors here because |
940 | * there's nothing that can be done. |
941 | */ |
942 | if (kill) |
943 | collect_procs(ppage, &tokill); |
944 | |
945 | if (hpage != ppage) |
946 | lock_page(ppage); |
947 | |
948 | ret = try_to_unmap(ppage, ttu); |
949 | if (ret != SWAP_SUCCESS) |
950 | printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n", |
951 | pfn, page_mapcount(ppage)); |
952 | |
953 | if (hpage != ppage) |
954 | unlock_page(ppage); |
955 | |
956 | /* |
957 | * Now that the dirty bit has been propagated to the |
958 | * struct page and all unmaps done we can decide if |
959 | * killing is needed or not. Only kill when the page |
960 | * was dirty, otherwise the tokill list is merely |
961 | * freed. When there was a problem unmapping earlier |
962 | * use a more force-full uncatchable kill to prevent |
963 | * any accesses to the poisoned memory. |
964 | */ |
965 | kill_procs_ao(&tokill, !!PageDirty(ppage), trapno, |
966 | ret != SWAP_SUCCESS, p, pfn); |
967 | |
968 | return ret; |
969 | } |
970 | |
971 | static void set_page_hwpoison_huge_page(struct page *hpage) |
972 | { |
973 | int i; |
974 | int nr_pages = 1 << compound_trans_order(hpage); |
975 | for (i = 0; i < nr_pages; i++) |
976 | SetPageHWPoison(hpage + i); |
977 | } |
978 | |
979 | static void clear_page_hwpoison_huge_page(struct page *hpage) |
980 | { |
981 | int i; |
982 | int nr_pages = 1 << compound_trans_order(hpage); |
983 | for (i = 0; i < nr_pages; i++) |
984 | ClearPageHWPoison(hpage + i); |
985 | } |
986 | |
987 | int __memory_failure(unsigned long pfn, int trapno, int flags) |
988 | { |
989 | struct page_state *ps; |
990 | struct page *p; |
991 | struct page *hpage; |
992 | int res; |
993 | unsigned int nr_pages; |
994 | |
995 | if (!sysctl_memory_failure_recovery) |
996 | panic("Memory failure from trap %d on page %lx", trapno, pfn); |
997 | |
998 | if (!pfn_valid(pfn)) { |
999 | printk(KERN_ERR |
1000 | "MCE %#lx: memory outside kernel control\n", |
1001 | pfn); |
1002 | return -ENXIO; |
1003 | } |
1004 | |
1005 | p = pfn_to_page(pfn); |
1006 | hpage = compound_head(p); |
1007 | if (TestSetPageHWPoison(p)) { |
1008 | printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn); |
1009 | return 0; |
1010 | } |
1011 | |
1012 | nr_pages = 1 << compound_trans_order(hpage); |
1013 | atomic_long_add(nr_pages, &mce_bad_pages); |
1014 | |
1015 | /* |
1016 | * We need/can do nothing about count=0 pages. |
1017 | * 1) it's a free page, and therefore in safe hand: |
1018 | * prep_new_page() will be the gate keeper. |
1019 | * 2) it's a free hugepage, which is also safe: |
1020 | * an affected hugepage will be dequeued from hugepage freelist, |
1021 | * so there's no concern about reusing it ever after. |
1022 | * 3) it's part of a non-compound high order page. |
1023 | * Implies some kernel user: cannot stop them from |
1024 | * R/W the page; let's pray that the page has been |
1025 | * used and will be freed some time later. |
1026 | * In fact it's dangerous to directly bump up page count from 0, |
1027 | * that may make page_freeze_refs()/page_unfreeze_refs() mismatch. |
1028 | */ |
1029 | if (!(flags & MF_COUNT_INCREASED) && |
1030 | !get_page_unless_zero(hpage)) { |
1031 | if (is_free_buddy_page(p)) { |
1032 | action_result(pfn, "free buddy", DELAYED); |
1033 | return 0; |
1034 | } else if (PageHuge(hpage)) { |
1035 | /* |
1036 | * Check "just unpoisoned", "filter hit", and |
1037 | * "race with other subpage." |
1038 | */ |
1039 | lock_page(hpage); |
1040 | if (!PageHWPoison(hpage) |
1041 | || (hwpoison_filter(p) && TestClearPageHWPoison(p)) |
1042 | || (p != hpage && TestSetPageHWPoison(hpage))) { |
1043 | atomic_long_sub(nr_pages, &mce_bad_pages); |
1044 | return 0; |
1045 | } |
1046 | set_page_hwpoison_huge_page(hpage); |
1047 | res = dequeue_hwpoisoned_huge_page(hpage); |
1048 | action_result(pfn, "free huge", |
1049 | res ? IGNORED : DELAYED); |
1050 | unlock_page(hpage); |
1051 | return res; |
1052 | } else { |
1053 | action_result(pfn, "high order kernel", IGNORED); |
1054 | return -EBUSY; |
1055 | } |
1056 | } |
1057 | |
1058 | /* |
1059 | * We ignore non-LRU pages for good reasons. |
1060 | * - PG_locked is only well defined for LRU pages and a few others |
1061 | * - to avoid races with __set_page_locked() |
1062 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) |
1063 | * The check (unnecessarily) ignores LRU pages being isolated and |
1064 | * walked by the page reclaim code, however that's not a big loss. |
1065 | */ |
1066 | if (!PageHuge(p) && !PageTransCompound(p)) { |
1067 | if (!PageLRU(p)) |
1068 | shake_page(p, 0); |
1069 | if (!PageLRU(p)) { |
1070 | /* |
1071 | * shake_page could have turned it free. |
1072 | */ |
1073 | if (is_free_buddy_page(p)) { |
1074 | action_result(pfn, "free buddy, 2nd try", |
1075 | DELAYED); |
1076 | return 0; |
1077 | } |
1078 | action_result(pfn, "non LRU", IGNORED); |
1079 | put_page(p); |
1080 | return -EBUSY; |
1081 | } |
1082 | } |
1083 | |
1084 | /* |
1085 | * Lock the page and wait for writeback to finish. |
1086 | * It's very difficult to mess with pages currently under IO |
1087 | * and in many cases impossible, so we just avoid it here. |
1088 | */ |
1089 | lock_page(hpage); |
1090 | |
1091 | /* |
1092 | * unpoison always clear PG_hwpoison inside page lock |
1093 | */ |
1094 | if (!PageHWPoison(p)) { |
1095 | printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn); |
1096 | res = 0; |
1097 | goto out; |
1098 | } |
1099 | if (hwpoison_filter(p)) { |
1100 | if (TestClearPageHWPoison(p)) |
1101 | atomic_long_sub(nr_pages, &mce_bad_pages); |
1102 | unlock_page(hpage); |
1103 | put_page(hpage); |
1104 | return 0; |
1105 | } |
1106 | |
1107 | /* |
1108 | * For error on the tail page, we should set PG_hwpoison |
1109 | * on the head page to show that the hugepage is hwpoisoned |
1110 | */ |
1111 | if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) { |
1112 | action_result(pfn, "hugepage already hardware poisoned", |
1113 | IGNORED); |
1114 | unlock_page(hpage); |
1115 | put_page(hpage); |
1116 | return 0; |
1117 | } |
1118 | /* |
1119 | * Set PG_hwpoison on all pages in an error hugepage, |
1120 | * because containment is done in hugepage unit for now. |
1121 | * Since we have done TestSetPageHWPoison() for the head page with |
1122 | * page lock held, we can safely set PG_hwpoison bits on tail pages. |
1123 | */ |
1124 | if (PageHuge(p)) |
1125 | set_page_hwpoison_huge_page(hpage); |
1126 | |
1127 | wait_on_page_writeback(p); |
1128 | |
1129 | /* |
1130 | * Now take care of user space mappings. |
1131 | * Abort on fail: __delete_from_page_cache() assumes unmapped page. |
1132 | */ |
1133 | if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) { |
1134 | printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn); |
1135 | res = -EBUSY; |
1136 | goto out; |
1137 | } |
1138 | |
1139 | /* |
1140 | * Torn down by someone else? |
1141 | */ |
1142 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
1143 | action_result(pfn, "already truncated LRU", IGNORED); |
1144 | res = -EBUSY; |
1145 | goto out; |
1146 | } |
1147 | |
1148 | res = -EBUSY; |
1149 | for (ps = error_states;; ps++) { |
1150 | if ((p->flags & ps->mask) == ps->res) { |
1151 | res = page_action(ps, p, pfn); |
1152 | break; |
1153 | } |
1154 | } |
1155 | out: |
1156 | unlock_page(hpage); |
1157 | return res; |
1158 | } |
1159 | EXPORT_SYMBOL_GPL(__memory_failure); |
1160 | |
1161 | /** |
1162 | * memory_failure - Handle memory failure of a page. |
1163 | * @pfn: Page Number of the corrupted page |
1164 | * @trapno: Trap number reported in the signal to user space. |
1165 | * |
1166 | * This function is called by the low level machine check code |
1167 | * of an architecture when it detects hardware memory corruption |
1168 | * of a page. It tries its best to recover, which includes |
1169 | * dropping pages, killing processes etc. |
1170 | * |
1171 | * The function is primarily of use for corruptions that |
1172 | * happen outside the current execution context (e.g. when |
1173 | * detected by a background scrubber) |
1174 | * |
1175 | * Must run in process context (e.g. a work queue) with interrupts |
1176 | * enabled and no spinlocks hold. |
1177 | */ |
1178 | void memory_failure(unsigned long pfn, int trapno) |
1179 | { |
1180 | __memory_failure(pfn, trapno, 0); |
1181 | } |
1182 | |
1183 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
1184 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) |
1185 | |
1186 | struct memory_failure_entry { |
1187 | unsigned long pfn; |
1188 | int trapno; |
1189 | int flags; |
1190 | }; |
1191 | |
1192 | struct memory_failure_cpu { |
1193 | DECLARE_KFIFO(fifo, struct memory_failure_entry, |
1194 | MEMORY_FAILURE_FIFO_SIZE); |
1195 | spinlock_t lock; |
1196 | struct work_struct work; |
1197 | }; |
1198 | |
1199 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); |
1200 | |
1201 | /** |
1202 | * memory_failure_queue - Schedule handling memory failure of a page. |
1203 | * @pfn: Page Number of the corrupted page |
1204 | * @trapno: Trap number reported in the signal to user space. |
1205 | * @flags: Flags for memory failure handling |
1206 | * |
1207 | * This function is called by the low level hardware error handler |
1208 | * when it detects hardware memory corruption of a page. It schedules |
1209 | * the recovering of error page, including dropping pages, killing |
1210 | * processes etc. |
1211 | * |
1212 | * The function is primarily of use for corruptions that |
1213 | * happen outside the current execution context (e.g. when |
1214 | * detected by a background scrubber) |
1215 | * |
1216 | * Can run in IRQ context. |
1217 | */ |
1218 | void memory_failure_queue(unsigned long pfn, int trapno, int flags) |
1219 | { |
1220 | struct memory_failure_cpu *mf_cpu; |
1221 | unsigned long proc_flags; |
1222 | struct memory_failure_entry entry = { |
1223 | .pfn = pfn, |
1224 | .trapno = trapno, |
1225 | .flags = flags, |
1226 | }; |
1227 | |
1228 | mf_cpu = &get_cpu_var(memory_failure_cpu); |
1229 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); |
1230 | if (kfifo_put(&mf_cpu->fifo, &entry)) |
1231 | schedule_work_on(smp_processor_id(), &mf_cpu->work); |
1232 | else |
1233 | pr_err("Memory failure: buffer overflow when queuing memory failure at 0x%#lx\n", |
1234 | pfn); |
1235 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); |
1236 | put_cpu_var(memory_failure_cpu); |
1237 | } |
1238 | EXPORT_SYMBOL_GPL(memory_failure_queue); |
1239 | |
1240 | static void memory_failure_work_func(struct work_struct *work) |
1241 | { |
1242 | struct memory_failure_cpu *mf_cpu; |
1243 | struct memory_failure_entry entry = { 0, }; |
1244 | unsigned long proc_flags; |
1245 | int gotten; |
1246 | |
1247 | mf_cpu = &__get_cpu_var(memory_failure_cpu); |
1248 | for (;;) { |
1249 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); |
1250 | gotten = kfifo_get(&mf_cpu->fifo, &entry); |
1251 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); |
1252 | if (!gotten) |
1253 | break; |
1254 | __memory_failure(entry.pfn, entry.trapno, entry.flags); |
1255 | } |
1256 | } |
1257 | |
1258 | static int __init memory_failure_init(void) |
1259 | { |
1260 | struct memory_failure_cpu *mf_cpu; |
1261 | int cpu; |
1262 | |
1263 | for_each_possible_cpu(cpu) { |
1264 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); |
1265 | spin_lock_init(&mf_cpu->lock); |
1266 | INIT_KFIFO(mf_cpu->fifo); |
1267 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); |
1268 | } |
1269 | |
1270 | return 0; |
1271 | } |
1272 | core_initcall(memory_failure_init); |
1273 | |
1274 | /** |
1275 | * unpoison_memory - Unpoison a previously poisoned page |
1276 | * @pfn: Page number of the to be unpoisoned page |
1277 | * |
1278 | * Software-unpoison a page that has been poisoned by |
1279 | * memory_failure() earlier. |
1280 | * |
1281 | * This is only done on the software-level, so it only works |
1282 | * for linux injected failures, not real hardware failures |
1283 | * |
1284 | * Returns 0 for success, otherwise -errno. |
1285 | */ |
1286 | int unpoison_memory(unsigned long pfn) |
1287 | { |
1288 | struct page *page; |
1289 | struct page *p; |
1290 | int freeit = 0; |
1291 | unsigned int nr_pages; |
1292 | |
1293 | if (!pfn_valid(pfn)) |
1294 | return -ENXIO; |
1295 | |
1296 | p = pfn_to_page(pfn); |
1297 | page = compound_head(p); |
1298 | |
1299 | if (!PageHWPoison(p)) { |
1300 | pr_info("MCE: Page was already unpoisoned %#lx\n", pfn); |
1301 | return 0; |
1302 | } |
1303 | |
1304 | nr_pages = 1 << compound_trans_order(page); |
1305 | |
1306 | if (!get_page_unless_zero(page)) { |
1307 | /* |
1308 | * Since HWPoisoned hugepage should have non-zero refcount, |
1309 | * race between memory failure and unpoison seems to happen. |
1310 | * In such case unpoison fails and memory failure runs |
1311 | * to the end. |
1312 | */ |
1313 | if (PageHuge(page)) { |
1314 | pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn); |
1315 | return 0; |
1316 | } |
1317 | if (TestClearPageHWPoison(p)) |
1318 | atomic_long_sub(nr_pages, &mce_bad_pages); |
1319 | pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn); |
1320 | return 0; |
1321 | } |
1322 | |
1323 | lock_page(page); |
1324 | /* |
1325 | * This test is racy because PG_hwpoison is set outside of page lock. |
1326 | * That's acceptable because that won't trigger kernel panic. Instead, |
1327 | * the PG_hwpoison page will be caught and isolated on the entrance to |
1328 | * the free buddy page pool. |
1329 | */ |
1330 | if (TestClearPageHWPoison(page)) { |
1331 | pr_info("MCE: Software-unpoisoned page %#lx\n", pfn); |
1332 | atomic_long_sub(nr_pages, &mce_bad_pages); |
1333 | freeit = 1; |
1334 | if (PageHuge(page)) |
1335 | clear_page_hwpoison_huge_page(page); |
1336 | } |
1337 | unlock_page(page); |
1338 | |
1339 | put_page(page); |
1340 | if (freeit) |
1341 | put_page(page); |
1342 | |
1343 | return 0; |
1344 | } |
1345 | EXPORT_SYMBOL(unpoison_memory); |
1346 | |
1347 | static struct page *new_page(struct page *p, unsigned long private, int **x) |
1348 | { |
1349 | int nid = page_to_nid(p); |
1350 | if (PageHuge(p)) |
1351 | return alloc_huge_page_node(page_hstate(compound_head(p)), |
1352 | nid); |
1353 | else |
1354 | return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0); |
1355 | } |
1356 | |
1357 | /* |
1358 | * Safely get reference count of an arbitrary page. |
1359 | * Returns 0 for a free page, -EIO for a zero refcount page |
1360 | * that is not free, and 1 for any other page type. |
1361 | * For 1 the page is returned with increased page count, otherwise not. |
1362 | */ |
1363 | static int get_any_page(struct page *p, unsigned long pfn, int flags) |
1364 | { |
1365 | int ret; |
1366 | |
1367 | if (flags & MF_COUNT_INCREASED) |
1368 | return 1; |
1369 | |
1370 | /* |
1371 | * The lock_memory_hotplug prevents a race with memory hotplug. |
1372 | * This is a big hammer, a better would be nicer. |
1373 | */ |
1374 | lock_memory_hotplug(); |
1375 | |
1376 | /* |
1377 | * Isolate the page, so that it doesn't get reallocated if it |
1378 | * was free. |
1379 | */ |
1380 | set_migratetype_isolate(p); |
1381 | /* |
1382 | * When the target page is a free hugepage, just remove it |
1383 | * from free hugepage list. |
1384 | */ |
1385 | if (!get_page_unless_zero(compound_head(p))) { |
1386 | if (PageHuge(p)) { |
1387 | pr_info("get_any_page: %#lx free huge page\n", pfn); |
1388 | ret = dequeue_hwpoisoned_huge_page(compound_head(p)); |
1389 | } else if (is_free_buddy_page(p)) { |
1390 | pr_info("get_any_page: %#lx free buddy page\n", pfn); |
1391 | /* Set hwpoison bit while page is still isolated */ |
1392 | SetPageHWPoison(p); |
1393 | ret = 0; |
1394 | } else { |
1395 | pr_info("get_any_page: %#lx: unknown zero refcount page type %lx\n", |
1396 | pfn, p->flags); |
1397 | ret = -EIO; |
1398 | } |
1399 | } else { |
1400 | /* Not a free page */ |
1401 | ret = 1; |
1402 | } |
1403 | unset_migratetype_isolate(p); |
1404 | unlock_memory_hotplug(); |
1405 | return ret; |
1406 | } |
1407 | |
1408 | static int soft_offline_huge_page(struct page *page, int flags) |
1409 | { |
1410 | int ret; |
1411 | unsigned long pfn = page_to_pfn(page); |
1412 | struct page *hpage = compound_head(page); |
1413 | LIST_HEAD(pagelist); |
1414 | |
1415 | ret = get_any_page(page, pfn, flags); |
1416 | if (ret < 0) |
1417 | return ret; |
1418 | if (ret == 0) |
1419 | goto done; |
1420 | |
1421 | if (PageHWPoison(hpage)) { |
1422 | put_page(hpage); |
1423 | pr_info("soft offline: %#lx hugepage already poisoned\n", pfn); |
1424 | return -EBUSY; |
1425 | } |
1426 | |
1427 | /* Keep page count to indicate a given hugepage is isolated. */ |
1428 | |
1429 | list_add(&hpage->lru, &pagelist); |
1430 | ret = migrate_huge_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0, |
1431 | true); |
1432 | if (ret) { |
1433 | struct page *page1, *page2; |
1434 | list_for_each_entry_safe(page1, page2, &pagelist, lru) |
1435 | put_page(page1); |
1436 | |
1437 | pr_info("soft offline: %#lx: migration failed %d, type %lx\n", |
1438 | pfn, ret, page->flags); |
1439 | if (ret > 0) |
1440 | ret = -EIO; |
1441 | return ret; |
1442 | } |
1443 | done: |
1444 | if (!PageHWPoison(hpage)) |
1445 | atomic_long_add(1 << compound_trans_order(hpage), &mce_bad_pages); |
1446 | set_page_hwpoison_huge_page(hpage); |
1447 | dequeue_hwpoisoned_huge_page(hpage); |
1448 | /* keep elevated page count for bad page */ |
1449 | return ret; |
1450 | } |
1451 | |
1452 | /** |
1453 | * soft_offline_page - Soft offline a page. |
1454 | * @page: page to offline |
1455 | * @flags: flags. Same as memory_failure(). |
1456 | * |
1457 | * Returns 0 on success, otherwise negated errno. |
1458 | * |
1459 | * Soft offline a page, by migration or invalidation, |
1460 | * without killing anything. This is for the case when |
1461 | * a page is not corrupted yet (so it's still valid to access), |
1462 | * but has had a number of corrected errors and is better taken |
1463 | * out. |
1464 | * |
1465 | * The actual policy on when to do that is maintained by |
1466 | * user space. |
1467 | * |
1468 | * This should never impact any application or cause data loss, |
1469 | * however it might take some time. |
1470 | * |
1471 | * This is not a 100% solution for all memory, but tries to be |
1472 | * ``good enough'' for the majority of memory. |
1473 | */ |
1474 | int soft_offline_page(struct page *page, int flags) |
1475 | { |
1476 | int ret; |
1477 | unsigned long pfn = page_to_pfn(page); |
1478 | |
1479 | if (PageHuge(page)) |
1480 | return soft_offline_huge_page(page, flags); |
1481 | |
1482 | ret = get_any_page(page, pfn, flags); |
1483 | if (ret < 0) |
1484 | return ret; |
1485 | if (ret == 0) |
1486 | goto done; |
1487 | |
1488 | /* |
1489 | * Page cache page we can handle? |
1490 | */ |
1491 | if (!PageLRU(page)) { |
1492 | /* |
1493 | * Try to free it. |
1494 | */ |
1495 | put_page(page); |
1496 | shake_page(page, 1); |
1497 | |
1498 | /* |
1499 | * Did it turn free? |
1500 | */ |
1501 | ret = get_any_page(page, pfn, 0); |
1502 | if (ret < 0) |
1503 | return ret; |
1504 | if (ret == 0) |
1505 | goto done; |
1506 | } |
1507 | if (!PageLRU(page)) { |
1508 | pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n", |
1509 | pfn, page->flags); |
1510 | return -EIO; |
1511 | } |
1512 | |
1513 | lock_page(page); |
1514 | wait_on_page_writeback(page); |
1515 | |
1516 | /* |
1517 | * Synchronized using the page lock with memory_failure() |
1518 | */ |
1519 | if (PageHWPoison(page)) { |
1520 | unlock_page(page); |
1521 | put_page(page); |
1522 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
1523 | return -EBUSY; |
1524 | } |
1525 | |
1526 | /* |
1527 | * Try to invalidate first. This should work for |
1528 | * non dirty unmapped page cache pages. |
1529 | */ |
1530 | ret = invalidate_inode_page(page); |
1531 | unlock_page(page); |
1532 | /* |
1533 | * RED-PEN would be better to keep it isolated here, but we |
1534 | * would need to fix isolation locking first. |
1535 | */ |
1536 | if (ret == 1) { |
1537 | put_page(page); |
1538 | ret = 0; |
1539 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
1540 | goto done; |
1541 | } |
1542 | |
1543 | /* |
1544 | * Simple invalidation didn't work. |
1545 | * Try to migrate to a new page instead. migrate.c |
1546 | * handles a large number of cases for us. |
1547 | */ |
1548 | ret = isolate_lru_page(page); |
1549 | /* |
1550 | * Drop page reference which is came from get_any_page() |
1551 | * successful isolate_lru_page() already took another one. |
1552 | */ |
1553 | put_page(page); |
1554 | if (!ret) { |
1555 | LIST_HEAD(pagelist); |
1556 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
1557 | page_is_file_cache(page)); |
1558 | list_add(&page->lru, &pagelist); |
1559 | ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, |
1560 | 0, true); |
1561 | if (ret) { |
1562 | putback_lru_pages(&pagelist); |
1563 | pr_info("soft offline: %#lx: migration failed %d, type %lx\n", |
1564 | pfn, ret, page->flags); |
1565 | if (ret > 0) |
1566 | ret = -EIO; |
1567 | } |
1568 | } else { |
1569 | pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n", |
1570 | pfn, ret, page_count(page), page->flags); |
1571 | } |
1572 | if (ret) |
1573 | return ret; |
1574 | |
1575 | done: |
1576 | atomic_long_add(1, &mce_bad_pages); |
1577 | SetPageHWPoison(page); |
1578 | /* keep elevated page count for bad page */ |
1579 | return ret; |
1580 | } |
1581 |
Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master
Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9