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