]>
Commit | Line | Data |
---|---|---|
6a46079c AK |
1 | /* |
2 | * Copyright (C) 2008, 2009 Intel Corporation | |
3 | * Authors: Andi Kleen, Fengguang Wu | |
4 | * | |
5 | * This software may be redistributed and/or modified under the terms of | |
6 | * the GNU General Public License ("GPL") version 2 only as published by the | |
7 | * Free Software Foundation. | |
8 | * | |
9 | * High level machine check handler. Handles pages reported by the | |
10 | * hardware as being corrupted usually due to a 2bit ECC memory or cache | |
11 | * failure. | |
12 | * | |
13 | * Handles page cache pages in various states. The tricky part | |
14 | * here is that we can access any page asynchronous to other VM | |
15 | * users, because memory failures could happen anytime and anywhere, | |
16 | * possibly violating some of their assumptions. This is why this code | |
17 | * has to be extremely careful. Generally it tries to use normal locking | |
18 | * rules, as in get the standard locks, even if that means the | |
19 | * error handling takes potentially a long time. | |
20 | * | |
21 | * The operation to map back from RMAP chains to processes has to walk | |
22 | * the complete process list and has non linear complexity with the number | |
23 | * mappings. In short it can be quite slow. But since memory corruptions | |
24 | * are rare we hope to get away with this. | |
25 | */ | |
26 | ||
27 | /* | |
28 | * Notebook: | |
29 | * - hugetlb needs more code | |
30 | * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages | |
31 | * - pass bad pages to kdump next kernel | |
32 | */ | |
33 | #define DEBUG 1 /* remove me in 2.6.34 */ | |
34 | #include <linux/kernel.h> | |
35 | #include <linux/mm.h> | |
36 | #include <linux/page-flags.h> | |
478c5ffc | 37 | #include <linux/kernel-page-flags.h> |
6a46079c | 38 | #include <linux/sched.h> |
01e00f88 | 39 | #include <linux/ksm.h> |
6a46079c AK |
40 | #include <linux/rmap.h> |
41 | #include <linux/pagemap.h> | |
42 | #include <linux/swap.h> | |
43 | #include <linux/backing-dev.h> | |
facb6011 AK |
44 | #include <linux/migrate.h> |
45 | #include <linux/page-isolation.h> | |
46 | #include <linux/suspend.h> | |
6a46079c AK |
47 | #include "internal.h" |
48 | ||
49 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
50 | ||
51 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
52 | ||
53 | atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0); | |
54 | ||
1bfe5feb | 55 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
56 | u32 hwpoison_filter_dev_major = ~0U; |
57 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
58 | u64 hwpoison_filter_flags_mask; |
59 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 60 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
61 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
62 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
63 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
64 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
65 | |
66 | static int hwpoison_filter_dev(struct page *p) | |
67 | { | |
68 | struct address_space *mapping; | |
69 | dev_t dev; | |
70 | ||
71 | if (hwpoison_filter_dev_major == ~0U && | |
72 | hwpoison_filter_dev_minor == ~0U) | |
73 | return 0; | |
74 | ||
75 | /* | |
76 | * page_mapping() does not accept slab page | |
77 | */ | |
78 | if (PageSlab(p)) | |
79 | return -EINVAL; | |
80 | ||
81 | mapping = page_mapping(p); | |
82 | if (mapping == NULL || mapping->host == NULL) | |
83 | return -EINVAL; | |
84 | ||
85 | dev = mapping->host->i_sb->s_dev; | |
86 | if (hwpoison_filter_dev_major != ~0U && | |
87 | hwpoison_filter_dev_major != MAJOR(dev)) | |
88 | return -EINVAL; | |
89 | if (hwpoison_filter_dev_minor != ~0U && | |
90 | hwpoison_filter_dev_minor != MINOR(dev)) | |
91 | return -EINVAL; | |
92 | ||
93 | return 0; | |
94 | } | |
95 | ||
478c5ffc WF |
96 | static int hwpoison_filter_flags(struct page *p) |
97 | { | |
98 | if (!hwpoison_filter_flags_mask) | |
99 | return 0; | |
100 | ||
101 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
102 | hwpoison_filter_flags_value) | |
103 | return 0; | |
104 | else | |
105 | return -EINVAL; | |
106 | } | |
107 | ||
4fd466eb AK |
108 | /* |
109 | * This allows stress tests to limit test scope to a collection of tasks | |
110 | * by putting them under some memcg. This prevents killing unrelated/important | |
111 | * processes such as /sbin/init. Note that the target task may share clean | |
112 | * pages with init (eg. libc text), which is harmless. If the target task | |
113 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
114 | * is also included in the memcg. At last, due to race conditions this filter | |
115 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
116 | * a freed page. | |
117 | */ | |
118 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
119 | u64 hwpoison_filter_memcg; | |
120 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
121 | static int hwpoison_filter_task(struct page *p) | |
122 | { | |
123 | struct mem_cgroup *mem; | |
124 | struct cgroup_subsys_state *css; | |
125 | unsigned long ino; | |
126 | ||
127 | if (!hwpoison_filter_memcg) | |
128 | return 0; | |
129 | ||
130 | mem = try_get_mem_cgroup_from_page(p); | |
131 | if (!mem) | |
132 | return -EINVAL; | |
133 | ||
134 | css = mem_cgroup_css(mem); | |
135 | /* root_mem_cgroup has NULL dentries */ | |
136 | if (!css->cgroup->dentry) | |
137 | return -EINVAL; | |
138 | ||
139 | ino = css->cgroup->dentry->d_inode->i_ino; | |
140 | css_put(css); | |
141 | ||
142 | if (ino != hwpoison_filter_memcg) | |
143 | return -EINVAL; | |
144 | ||
145 | return 0; | |
146 | } | |
147 | #else | |
148 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
149 | #endif | |
150 | ||
7c116f2b WF |
151 | int hwpoison_filter(struct page *p) |
152 | { | |
1bfe5feb HL |
153 | if (!hwpoison_filter_enable) |
154 | return 0; | |
155 | ||
7c116f2b WF |
156 | if (hwpoison_filter_dev(p)) |
157 | return -EINVAL; | |
158 | ||
478c5ffc WF |
159 | if (hwpoison_filter_flags(p)) |
160 | return -EINVAL; | |
161 | ||
4fd466eb AK |
162 | if (hwpoison_filter_task(p)) |
163 | return -EINVAL; | |
164 | ||
7c116f2b WF |
165 | return 0; |
166 | } | |
167 | EXPORT_SYMBOL_GPL(hwpoison_filter); | |
168 | ||
6a46079c AK |
169 | /* |
170 | * Send all the processes who have the page mapped an ``action optional'' | |
171 | * signal. | |
172 | */ | |
173 | static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno, | |
174 | unsigned long pfn) | |
175 | { | |
176 | struct siginfo si; | |
177 | int ret; | |
178 | ||
179 | printk(KERN_ERR | |
180 | "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n", | |
181 | pfn, t->comm, t->pid); | |
182 | si.si_signo = SIGBUS; | |
183 | si.si_errno = 0; | |
184 | si.si_code = BUS_MCEERR_AO; | |
185 | si.si_addr = (void *)addr; | |
186 | #ifdef __ARCH_SI_TRAPNO | |
187 | si.si_trapno = trapno; | |
188 | #endif | |
189 | si.si_addr_lsb = PAGE_SHIFT; | |
190 | /* | |
191 | * Don't use force here, it's convenient if the signal | |
192 | * can be temporarily blocked. | |
193 | * This could cause a loop when the user sets SIGBUS | |
194 | * to SIG_IGN, but hopefully noone will do that? | |
195 | */ | |
196 | ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */ | |
197 | if (ret < 0) | |
198 | printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n", | |
199 | t->comm, t->pid, ret); | |
200 | return ret; | |
201 | } | |
202 | ||
588f9ce6 AK |
203 | /* |
204 | * When a unknown page type is encountered drain as many buffers as possible | |
205 | * in the hope to turn the page into a LRU or free page, which we can handle. | |
206 | */ | |
facb6011 | 207 | void shake_page(struct page *p, int access) |
588f9ce6 AK |
208 | { |
209 | if (!PageSlab(p)) { | |
210 | lru_add_drain_all(); | |
211 | if (PageLRU(p)) | |
212 | return; | |
213 | drain_all_pages(); | |
214 | if (PageLRU(p) || is_free_buddy_page(p)) | |
215 | return; | |
216 | } | |
facb6011 | 217 | |
588f9ce6 | 218 | /* |
facb6011 AK |
219 | * Only all shrink_slab here (which would also |
220 | * shrink other caches) if access is not potentially fatal. | |
588f9ce6 | 221 | */ |
facb6011 AK |
222 | if (access) { |
223 | int nr; | |
224 | do { | |
225 | nr = shrink_slab(1000, GFP_KERNEL, 1000); | |
226 | if (page_count(p) == 0) | |
227 | break; | |
228 | } while (nr > 10); | |
229 | } | |
588f9ce6 AK |
230 | } |
231 | EXPORT_SYMBOL_GPL(shake_page); | |
232 | ||
6a46079c AK |
233 | /* |
234 | * Kill all processes that have a poisoned page mapped and then isolate | |
235 | * the page. | |
236 | * | |
237 | * General strategy: | |
238 | * Find all processes having the page mapped and kill them. | |
239 | * But we keep a page reference around so that the page is not | |
240 | * actually freed yet. | |
241 | * Then stash the page away | |
242 | * | |
243 | * There's no convenient way to get back to mapped processes | |
244 | * from the VMAs. So do a brute-force search over all | |
245 | * running processes. | |
246 | * | |
247 | * Remember that machine checks are not common (or rather | |
248 | * if they are common you have other problems), so this shouldn't | |
249 | * be a performance issue. | |
250 | * | |
251 | * Also there are some races possible while we get from the | |
252 | * error detection to actually handle it. | |
253 | */ | |
254 | ||
255 | struct to_kill { | |
256 | struct list_head nd; | |
257 | struct task_struct *tsk; | |
258 | unsigned long addr; | |
259 | unsigned addr_valid:1; | |
260 | }; | |
261 | ||
262 | /* | |
263 | * Failure handling: if we can't find or can't kill a process there's | |
264 | * not much we can do. We just print a message and ignore otherwise. | |
265 | */ | |
266 | ||
267 | /* | |
268 | * Schedule a process for later kill. | |
269 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
270 | * TBD would GFP_NOIO be enough? | |
271 | */ | |
272 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
273 | struct vm_area_struct *vma, | |
274 | struct list_head *to_kill, | |
275 | struct to_kill **tkc) | |
276 | { | |
277 | struct to_kill *tk; | |
278 | ||
279 | if (*tkc) { | |
280 | tk = *tkc; | |
281 | *tkc = NULL; | |
282 | } else { | |
283 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); | |
284 | if (!tk) { | |
285 | printk(KERN_ERR | |
286 | "MCE: Out of memory while machine check handling\n"); | |
287 | return; | |
288 | } | |
289 | } | |
290 | tk->addr = page_address_in_vma(p, vma); | |
291 | tk->addr_valid = 1; | |
292 | ||
293 | /* | |
294 | * In theory we don't have to kill when the page was | |
295 | * munmaped. But it could be also a mremap. Since that's | |
296 | * likely very rare kill anyways just out of paranoia, but use | |
297 | * a SIGKILL because the error is not contained anymore. | |
298 | */ | |
299 | if (tk->addr == -EFAULT) { | |
300 | pr_debug("MCE: Unable to find user space address %lx in %s\n", | |
301 | page_to_pfn(p), tsk->comm); | |
302 | tk->addr_valid = 0; | |
303 | } | |
304 | get_task_struct(tsk); | |
305 | tk->tsk = tsk; | |
306 | list_add_tail(&tk->nd, to_kill); | |
307 | } | |
308 | ||
309 | /* | |
310 | * Kill the processes that have been collected earlier. | |
311 | * | |
312 | * Only do anything when DOIT is set, otherwise just free the list | |
313 | * (this is used for clean pages which do not need killing) | |
314 | * Also when FAIL is set do a force kill because something went | |
315 | * wrong earlier. | |
316 | */ | |
317 | static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno, | |
318 | int fail, unsigned long pfn) | |
319 | { | |
320 | struct to_kill *tk, *next; | |
321 | ||
322 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
323 | if (doit) { | |
324 | /* | |
af901ca1 | 325 | * In case something went wrong with munmapping |
6a46079c AK |
326 | * make sure the process doesn't catch the |
327 | * signal and then access the memory. Just kill it. | |
328 | * the signal handlers | |
329 | */ | |
330 | if (fail || tk->addr_valid == 0) { | |
331 | printk(KERN_ERR | |
332 | "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", | |
333 | pfn, tk->tsk->comm, tk->tsk->pid); | |
334 | force_sig(SIGKILL, tk->tsk); | |
335 | } | |
336 | ||
337 | /* | |
338 | * In theory the process could have mapped | |
339 | * something else on the address in-between. We could | |
340 | * check for that, but we need to tell the | |
341 | * process anyways. | |
342 | */ | |
343 | else if (kill_proc_ao(tk->tsk, tk->addr, trapno, | |
344 | pfn) < 0) | |
345 | printk(KERN_ERR | |
346 | "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n", | |
347 | pfn, tk->tsk->comm, tk->tsk->pid); | |
348 | } | |
349 | put_task_struct(tk->tsk); | |
350 | kfree(tk); | |
351 | } | |
352 | } | |
353 | ||
354 | static int task_early_kill(struct task_struct *tsk) | |
355 | { | |
356 | if (!tsk->mm) | |
357 | return 0; | |
358 | if (tsk->flags & PF_MCE_PROCESS) | |
359 | return !!(tsk->flags & PF_MCE_EARLY); | |
360 | return sysctl_memory_failure_early_kill; | |
361 | } | |
362 | ||
363 | /* | |
364 | * Collect processes when the error hit an anonymous page. | |
365 | */ | |
366 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
367 | struct to_kill **tkc) | |
368 | { | |
369 | struct vm_area_struct *vma; | |
370 | struct task_struct *tsk; | |
371 | struct anon_vma *av; | |
372 | ||
373 | read_lock(&tasklist_lock); | |
374 | av = page_lock_anon_vma(page); | |
375 | if (av == NULL) /* Not actually mapped anymore */ | |
376 | goto out; | |
377 | for_each_process (tsk) { | |
378 | if (!task_early_kill(tsk)) | |
379 | continue; | |
380 | list_for_each_entry (vma, &av->head, anon_vma_node) { | |
381 | if (!page_mapped_in_vma(page, vma)) | |
382 | continue; | |
383 | if (vma->vm_mm == tsk->mm) | |
384 | add_to_kill(tsk, page, vma, to_kill, tkc); | |
385 | } | |
386 | } | |
387 | page_unlock_anon_vma(av); | |
388 | out: | |
389 | read_unlock(&tasklist_lock); | |
390 | } | |
391 | ||
392 | /* | |
393 | * Collect processes when the error hit a file mapped page. | |
394 | */ | |
395 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
396 | struct to_kill **tkc) | |
397 | { | |
398 | struct vm_area_struct *vma; | |
399 | struct task_struct *tsk; | |
400 | struct prio_tree_iter iter; | |
401 | struct address_space *mapping = page->mapping; | |
402 | ||
403 | /* | |
404 | * A note on the locking order between the two locks. | |
405 | * We don't rely on this particular order. | |
406 | * If you have some other code that needs a different order | |
407 | * feel free to switch them around. Or add a reverse link | |
408 | * from mm_struct to task_struct, then this could be all | |
409 | * done without taking tasklist_lock and looping over all tasks. | |
410 | */ | |
411 | ||
412 | read_lock(&tasklist_lock); | |
413 | spin_lock(&mapping->i_mmap_lock); | |
414 | for_each_process(tsk) { | |
415 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
416 | ||
417 | if (!task_early_kill(tsk)) | |
418 | continue; | |
419 | ||
420 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, | |
421 | pgoff) { | |
422 | /* | |
423 | * Send early kill signal to tasks where a vma covers | |
424 | * the page but the corrupted page is not necessarily | |
425 | * mapped it in its pte. | |
426 | * Assume applications who requested early kill want | |
427 | * to be informed of all such data corruptions. | |
428 | */ | |
429 | if (vma->vm_mm == tsk->mm) | |
430 | add_to_kill(tsk, page, vma, to_kill, tkc); | |
431 | } | |
432 | } | |
433 | spin_unlock(&mapping->i_mmap_lock); | |
434 | read_unlock(&tasklist_lock); | |
435 | } | |
436 | ||
437 | /* | |
438 | * Collect the processes who have the corrupted page mapped to kill. | |
439 | * This is done in two steps for locking reasons. | |
440 | * First preallocate one tokill structure outside the spin locks, | |
441 | * so that we can kill at least one process reasonably reliable. | |
442 | */ | |
443 | static void collect_procs(struct page *page, struct list_head *tokill) | |
444 | { | |
445 | struct to_kill *tk; | |
446 | ||
447 | if (!page->mapping) | |
448 | return; | |
449 | ||
450 | tk = kmalloc(sizeof(struct to_kill), GFP_NOIO); | |
451 | if (!tk) | |
452 | return; | |
453 | if (PageAnon(page)) | |
454 | collect_procs_anon(page, tokill, &tk); | |
455 | else | |
456 | collect_procs_file(page, tokill, &tk); | |
457 | kfree(tk); | |
458 | } | |
459 | ||
460 | /* | |
461 | * Error handlers for various types of pages. | |
462 | */ | |
463 | ||
464 | enum outcome { | |
d95ea51e WF |
465 | IGNORED, /* Error: cannot be handled */ |
466 | FAILED, /* Error: handling failed */ | |
6a46079c | 467 | DELAYED, /* Will be handled later */ |
6a46079c AK |
468 | RECOVERED, /* Successfully recovered */ |
469 | }; | |
470 | ||
471 | static const char *action_name[] = { | |
d95ea51e | 472 | [IGNORED] = "Ignored", |
6a46079c AK |
473 | [FAILED] = "Failed", |
474 | [DELAYED] = "Delayed", | |
6a46079c AK |
475 | [RECOVERED] = "Recovered", |
476 | }; | |
477 | ||
dc2a1cbf WF |
478 | /* |
479 | * XXX: It is possible that a page is isolated from LRU cache, | |
480 | * and then kept in swap cache or failed to remove from page cache. | |
481 | * The page count will stop it from being freed by unpoison. | |
482 | * Stress tests should be aware of this memory leak problem. | |
483 | */ | |
484 | static int delete_from_lru_cache(struct page *p) | |
485 | { | |
486 | if (!isolate_lru_page(p)) { | |
487 | /* | |
488 | * Clear sensible page flags, so that the buddy system won't | |
489 | * complain when the page is unpoison-and-freed. | |
490 | */ | |
491 | ClearPageActive(p); | |
492 | ClearPageUnevictable(p); | |
493 | /* | |
494 | * drop the page count elevated by isolate_lru_page() | |
495 | */ | |
496 | page_cache_release(p); | |
497 | return 0; | |
498 | } | |
499 | return -EIO; | |
500 | } | |
501 | ||
6a46079c AK |
502 | /* |
503 | * Error hit kernel page. | |
504 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
505 | * could be more sophisticated. | |
506 | */ | |
507 | static int me_kernel(struct page *p, unsigned long pfn) | |
6a46079c AK |
508 | { |
509 | return IGNORED; | |
510 | } | |
511 | ||
512 | /* | |
513 | * Page in unknown state. Do nothing. | |
514 | */ | |
515 | static int me_unknown(struct page *p, unsigned long pfn) | |
516 | { | |
517 | printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn); | |
518 | return FAILED; | |
519 | } | |
520 | ||
6a46079c AK |
521 | /* |
522 | * Clean (or cleaned) page cache page. | |
523 | */ | |
524 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
525 | { | |
526 | int err; | |
527 | int ret = FAILED; | |
528 | struct address_space *mapping; | |
529 | ||
dc2a1cbf WF |
530 | delete_from_lru_cache(p); |
531 | ||
6a46079c AK |
532 | /* |
533 | * For anonymous pages we're done the only reference left | |
534 | * should be the one m_f() holds. | |
535 | */ | |
536 | if (PageAnon(p)) | |
537 | return RECOVERED; | |
538 | ||
539 | /* | |
540 | * Now truncate the page in the page cache. This is really | |
541 | * more like a "temporary hole punch" | |
542 | * Don't do this for block devices when someone else | |
543 | * has a reference, because it could be file system metadata | |
544 | * and that's not safe to truncate. | |
545 | */ | |
546 | mapping = page_mapping(p); | |
547 | if (!mapping) { | |
548 | /* | |
549 | * Page has been teared down in the meanwhile | |
550 | */ | |
551 | return FAILED; | |
552 | } | |
553 | ||
554 | /* | |
555 | * Truncation is a bit tricky. Enable it per file system for now. | |
556 | * | |
557 | * Open: to take i_mutex or not for this? Right now we don't. | |
558 | */ | |
559 | if (mapping->a_ops->error_remove_page) { | |
560 | err = mapping->a_ops->error_remove_page(mapping, p); | |
561 | if (err != 0) { | |
562 | printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n", | |
563 | pfn, err); | |
564 | } else if (page_has_private(p) && | |
565 | !try_to_release_page(p, GFP_NOIO)) { | |
566 | pr_debug("MCE %#lx: failed to release buffers\n", pfn); | |
567 | } else { | |
568 | ret = RECOVERED; | |
569 | } | |
570 | } else { | |
571 | /* | |
572 | * If the file system doesn't support it just invalidate | |
573 | * This fails on dirty or anything with private pages | |
574 | */ | |
575 | if (invalidate_inode_page(p)) | |
576 | ret = RECOVERED; | |
577 | else | |
578 | printk(KERN_INFO "MCE %#lx: Failed to invalidate\n", | |
579 | pfn); | |
580 | } | |
581 | return ret; | |
582 | } | |
583 | ||
584 | /* | |
585 | * Dirty cache page page | |
586 | * Issues: when the error hit a hole page the error is not properly | |
587 | * propagated. | |
588 | */ | |
589 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
590 | { | |
591 | struct address_space *mapping = page_mapping(p); | |
592 | ||
593 | SetPageError(p); | |
594 | /* TBD: print more information about the file. */ | |
595 | if (mapping) { | |
596 | /* | |
597 | * IO error will be reported by write(), fsync(), etc. | |
598 | * who check the mapping. | |
599 | * This way the application knows that something went | |
600 | * wrong with its dirty file data. | |
601 | * | |
602 | * There's one open issue: | |
603 | * | |
604 | * The EIO will be only reported on the next IO | |
605 | * operation and then cleared through the IO map. | |
606 | * Normally Linux has two mechanisms to pass IO error | |
607 | * first through the AS_EIO flag in the address space | |
608 | * and then through the PageError flag in the page. | |
609 | * Since we drop pages on memory failure handling the | |
610 | * only mechanism open to use is through AS_AIO. | |
611 | * | |
612 | * This has the disadvantage that it gets cleared on | |
613 | * the first operation that returns an error, while | |
614 | * the PageError bit is more sticky and only cleared | |
615 | * when the page is reread or dropped. If an | |
616 | * application assumes it will always get error on | |
617 | * fsync, but does other operations on the fd before | |
618 | * and the page is dropped inbetween then the error | |
619 | * will not be properly reported. | |
620 | * | |
621 | * This can already happen even without hwpoisoned | |
622 | * pages: first on metadata IO errors (which only | |
623 | * report through AS_EIO) or when the page is dropped | |
624 | * at the wrong time. | |
625 | * | |
626 | * So right now we assume that the application DTRT on | |
627 | * the first EIO, but we're not worse than other parts | |
628 | * of the kernel. | |
629 | */ | |
630 | mapping_set_error(mapping, EIO); | |
631 | } | |
632 | ||
633 | return me_pagecache_clean(p, pfn); | |
634 | } | |
635 | ||
636 | /* | |
637 | * Clean and dirty swap cache. | |
638 | * | |
639 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
640 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
641 | * referenced concurrently by 2 types of PTEs: | |
642 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
643 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
644 | * and then | |
645 | * - clear dirty bit to prevent IO | |
646 | * - remove from LRU | |
647 | * - but keep in the swap cache, so that when we return to it on | |
648 | * a later page fault, we know the application is accessing | |
649 | * corrupted data and shall be killed (we installed simple | |
650 | * interception code in do_swap_page to catch it). | |
651 | * | |
652 | * Clean swap cache pages can be directly isolated. A later page fault will | |
653 | * bring in the known good data from disk. | |
654 | */ | |
655 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
656 | { | |
6a46079c AK |
657 | ClearPageDirty(p); |
658 | /* Trigger EIO in shmem: */ | |
659 | ClearPageUptodate(p); | |
660 | ||
dc2a1cbf WF |
661 | if (!delete_from_lru_cache(p)) |
662 | return DELAYED; | |
663 | else | |
664 | return FAILED; | |
6a46079c AK |
665 | } |
666 | ||
667 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
668 | { | |
6a46079c | 669 | delete_from_swap_cache(p); |
e43c3afb | 670 | |
dc2a1cbf WF |
671 | if (!delete_from_lru_cache(p)) |
672 | return RECOVERED; | |
673 | else | |
674 | return FAILED; | |
6a46079c AK |
675 | } |
676 | ||
677 | /* | |
678 | * Huge pages. Needs work. | |
679 | * Issues: | |
680 | * No rmap support so we cannot find the original mapper. In theory could walk | |
681 | * all MMs and look for the mappings, but that would be non atomic and racy. | |
682 | * Need rmap for hugepages for this. Alternatively we could employ a heuristic, | |
683 | * like just walking the current process and hoping it has it mapped (that | |
684 | * should be usually true for the common "shared database cache" case) | |
685 | * Should handle free huge pages and dequeue them too, but this needs to | |
686 | * handle huge page accounting correctly. | |
687 | */ | |
688 | static int me_huge_page(struct page *p, unsigned long pfn) | |
689 | { | |
690 | return FAILED; | |
691 | } | |
692 | ||
693 | /* | |
694 | * Various page states we can handle. | |
695 | * | |
696 | * A page state is defined by its current page->flags bits. | |
697 | * The table matches them in order and calls the right handler. | |
698 | * | |
699 | * This is quite tricky because we can access page at any time | |
700 | * in its live cycle, so all accesses have to be extremly careful. | |
701 | * | |
702 | * This is not complete. More states could be added. | |
703 | * For any missing state don't attempt recovery. | |
704 | */ | |
705 | ||
706 | #define dirty (1UL << PG_dirty) | |
707 | #define sc (1UL << PG_swapcache) | |
708 | #define unevict (1UL << PG_unevictable) | |
709 | #define mlock (1UL << PG_mlocked) | |
710 | #define writeback (1UL << PG_writeback) | |
711 | #define lru (1UL << PG_lru) | |
712 | #define swapbacked (1UL << PG_swapbacked) | |
713 | #define head (1UL << PG_head) | |
714 | #define tail (1UL << PG_tail) | |
715 | #define compound (1UL << PG_compound) | |
716 | #define slab (1UL << PG_slab) | |
6a46079c AK |
717 | #define reserved (1UL << PG_reserved) |
718 | ||
719 | static struct page_state { | |
720 | unsigned long mask; | |
721 | unsigned long res; | |
722 | char *msg; | |
723 | int (*action)(struct page *p, unsigned long pfn); | |
724 | } error_states[] = { | |
d95ea51e | 725 | { reserved, reserved, "reserved kernel", me_kernel }, |
95d01fc6 WF |
726 | /* |
727 | * free pages are specially detected outside this table: | |
728 | * PG_buddy pages only make a small fraction of all free pages. | |
729 | */ | |
6a46079c AK |
730 | |
731 | /* | |
732 | * Could in theory check if slab page is free or if we can drop | |
733 | * currently unused objects without touching them. But just | |
734 | * treat it as standard kernel for now. | |
735 | */ | |
736 | { slab, slab, "kernel slab", me_kernel }, | |
737 | ||
738 | #ifdef CONFIG_PAGEFLAGS_EXTENDED | |
739 | { head, head, "huge", me_huge_page }, | |
740 | { tail, tail, "huge", me_huge_page }, | |
741 | #else | |
742 | { compound, compound, "huge", me_huge_page }, | |
743 | #endif | |
744 | ||
745 | { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty }, | |
746 | { sc|dirty, sc, "swapcache", me_swapcache_clean }, | |
747 | ||
748 | { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty}, | |
749 | { unevict, unevict, "unevictable LRU", me_pagecache_clean}, | |
750 | ||
6a46079c AK |
751 | { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty }, |
752 | { mlock, mlock, "mlocked LRU", me_pagecache_clean }, | |
6a46079c AK |
753 | |
754 | { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty }, | |
755 | { lru|dirty, lru, "clean LRU", me_pagecache_clean }, | |
6a46079c AK |
756 | |
757 | /* | |
758 | * Catchall entry: must be at end. | |
759 | */ | |
760 | { 0, 0, "unknown page state", me_unknown }, | |
761 | }; | |
762 | ||
2326c467 AK |
763 | #undef dirty |
764 | #undef sc | |
765 | #undef unevict | |
766 | #undef mlock | |
767 | #undef writeback | |
768 | #undef lru | |
769 | #undef swapbacked | |
770 | #undef head | |
771 | #undef tail | |
772 | #undef compound | |
773 | #undef slab | |
774 | #undef reserved | |
775 | ||
6a46079c AK |
776 | static void action_result(unsigned long pfn, char *msg, int result) |
777 | { | |
a7560fc8 | 778 | struct page *page = pfn_to_page(pfn); |
6a46079c AK |
779 | |
780 | printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n", | |
781 | pfn, | |
a7560fc8 | 782 | PageDirty(page) ? "dirty " : "", |
6a46079c AK |
783 | msg, action_name[result]); |
784 | } | |
785 | ||
786 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 787 | unsigned long pfn) |
6a46079c AK |
788 | { |
789 | int result; | |
7456b040 | 790 | int count; |
6a46079c AK |
791 | |
792 | result = ps->action(p, pfn); | |
793 | action_result(pfn, ps->msg, result); | |
7456b040 | 794 | |
bd1ce5f9 | 795 | count = page_count(p) - 1; |
138ce286 WF |
796 | if (ps->action == me_swapcache_dirty && result == DELAYED) |
797 | count--; | |
798 | if (count != 0) { | |
6a46079c AK |
799 | printk(KERN_ERR |
800 | "MCE %#lx: %s page still referenced by %d users\n", | |
7456b040 | 801 | pfn, ps->msg, count); |
138ce286 WF |
802 | result = FAILED; |
803 | } | |
6a46079c AK |
804 | |
805 | /* Could do more checks here if page looks ok */ | |
806 | /* | |
807 | * Could adjust zone counters here to correct for the missing page. | |
808 | */ | |
809 | ||
138ce286 | 810 | return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
811 | } |
812 | ||
813 | #define N_UNMAP_TRIES 5 | |
814 | ||
815 | /* | |
816 | * Do all that is necessary to remove user space mappings. Unmap | |
817 | * the pages and send SIGBUS to the processes if the data was dirty. | |
818 | */ | |
1668bfd5 | 819 | static int hwpoison_user_mappings(struct page *p, unsigned long pfn, |
6a46079c AK |
820 | int trapno) |
821 | { | |
822 | enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
823 | struct address_space *mapping; | |
824 | LIST_HEAD(tokill); | |
825 | int ret; | |
826 | int i; | |
827 | int kill = 1; | |
828 | ||
1668bfd5 WF |
829 | if (PageReserved(p) || PageSlab(p)) |
830 | return SWAP_SUCCESS; | |
6a46079c | 831 | |
6a46079c AK |
832 | /* |
833 | * This check implies we don't kill processes if their pages | |
834 | * are in the swap cache early. Those are always late kills. | |
835 | */ | |
836 | if (!page_mapped(p)) | |
1668bfd5 WF |
837 | return SWAP_SUCCESS; |
838 | ||
839 | if (PageCompound(p) || PageKsm(p)) | |
840 | return SWAP_FAIL; | |
6a46079c AK |
841 | |
842 | if (PageSwapCache(p)) { | |
843 | printk(KERN_ERR | |
844 | "MCE %#lx: keeping poisoned page in swap cache\n", pfn); | |
845 | ttu |= TTU_IGNORE_HWPOISON; | |
846 | } | |
847 | ||
848 | /* | |
849 | * Propagate the dirty bit from PTEs to struct page first, because we | |
850 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
851 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
852 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c AK |
853 | */ |
854 | mapping = page_mapping(p); | |
855 | if (!PageDirty(p) && mapping && mapping_cap_writeback_dirty(mapping)) { | |
856 | if (page_mkclean(p)) { | |
857 | SetPageDirty(p); | |
858 | } else { | |
859 | kill = 0; | |
860 | ttu |= TTU_IGNORE_HWPOISON; | |
861 | printk(KERN_INFO | |
862 | "MCE %#lx: corrupted page was clean: dropped without side effects\n", | |
863 | pfn); | |
864 | } | |
865 | } | |
866 | ||
867 | /* | |
868 | * First collect all the processes that have the page | |
869 | * mapped in dirty form. This has to be done before try_to_unmap, | |
870 | * because ttu takes the rmap data structures down. | |
871 | * | |
872 | * Error handling: We ignore errors here because | |
873 | * there's nothing that can be done. | |
874 | */ | |
875 | if (kill) | |
876 | collect_procs(p, &tokill); | |
877 | ||
878 | /* | |
879 | * try_to_unmap can fail temporarily due to races. | |
880 | * Try a few times (RED-PEN better strategy?) | |
881 | */ | |
882 | for (i = 0; i < N_UNMAP_TRIES; i++) { | |
883 | ret = try_to_unmap(p, ttu); | |
884 | if (ret == SWAP_SUCCESS) | |
885 | break; | |
886 | pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn, ret); | |
887 | } | |
888 | ||
889 | if (ret != SWAP_SUCCESS) | |
890 | printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n", | |
891 | pfn, page_mapcount(p)); | |
892 | ||
893 | /* | |
894 | * Now that the dirty bit has been propagated to the | |
895 | * struct page and all unmaps done we can decide if | |
896 | * killing is needed or not. Only kill when the page | |
897 | * was dirty, otherwise the tokill list is merely | |
898 | * freed. When there was a problem unmapping earlier | |
899 | * use a more force-full uncatchable kill to prevent | |
900 | * any accesses to the poisoned memory. | |
901 | */ | |
902 | kill_procs_ao(&tokill, !!PageDirty(p), trapno, | |
903 | ret != SWAP_SUCCESS, pfn); | |
1668bfd5 WF |
904 | |
905 | return ret; | |
6a46079c AK |
906 | } |
907 | ||
82ba011b | 908 | int __memory_failure(unsigned long pfn, int trapno, int flags) |
6a46079c AK |
909 | { |
910 | struct page_state *ps; | |
911 | struct page *p; | |
912 | int res; | |
913 | ||
914 | if (!sysctl_memory_failure_recovery) | |
915 | panic("Memory failure from trap %d on page %lx", trapno, pfn); | |
916 | ||
917 | if (!pfn_valid(pfn)) { | |
a7560fc8 WF |
918 | printk(KERN_ERR |
919 | "MCE %#lx: memory outside kernel control\n", | |
920 | pfn); | |
921 | return -ENXIO; | |
6a46079c AK |
922 | } |
923 | ||
924 | p = pfn_to_page(pfn); | |
925 | if (TestSetPageHWPoison(p)) { | |
d95ea51e | 926 | printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn); |
6a46079c AK |
927 | return 0; |
928 | } | |
929 | ||
930 | atomic_long_add(1, &mce_bad_pages); | |
931 | ||
932 | /* | |
933 | * We need/can do nothing about count=0 pages. | |
934 | * 1) it's a free page, and therefore in safe hand: | |
935 | * prep_new_page() will be the gate keeper. | |
936 | * 2) it's part of a non-compound high order page. | |
937 | * Implies some kernel user: cannot stop them from | |
938 | * R/W the page; let's pray that the page has been | |
939 | * used and will be freed some time later. | |
940 | * In fact it's dangerous to directly bump up page count from 0, | |
941 | * that may make page_freeze_refs()/page_unfreeze_refs() mismatch. | |
942 | */ | |
82ba011b AK |
943 | if (!(flags & MF_COUNT_INCREASED) && |
944 | !get_page_unless_zero(compound_head(p))) { | |
8d22ba1b WF |
945 | if (is_free_buddy_page(p)) { |
946 | action_result(pfn, "free buddy", DELAYED); | |
947 | return 0; | |
948 | } else { | |
949 | action_result(pfn, "high order kernel", IGNORED); | |
950 | return -EBUSY; | |
951 | } | |
6a46079c AK |
952 | } |
953 | ||
e43c3afb WF |
954 | /* |
955 | * We ignore non-LRU pages for good reasons. | |
956 | * - PG_locked is only well defined for LRU pages and a few others | |
957 | * - to avoid races with __set_page_locked() | |
958 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) | |
959 | * The check (unnecessarily) ignores LRU pages being isolated and | |
960 | * walked by the page reclaim code, however that's not a big loss. | |
961 | */ | |
962 | if (!PageLRU(p)) | |
facb6011 | 963 | shake_page(p, 0); |
dc2a1cbf | 964 | if (!PageLRU(p)) { |
0474a60e AK |
965 | /* |
966 | * shake_page could have turned it free. | |
967 | */ | |
968 | if (is_free_buddy_page(p)) { | |
969 | action_result(pfn, "free buddy, 2nd try", DELAYED); | |
970 | return 0; | |
971 | } | |
e43c3afb WF |
972 | action_result(pfn, "non LRU", IGNORED); |
973 | put_page(p); | |
974 | return -EBUSY; | |
975 | } | |
e43c3afb | 976 | |
6a46079c AK |
977 | /* |
978 | * Lock the page and wait for writeback to finish. | |
979 | * It's very difficult to mess with pages currently under IO | |
980 | * and in many cases impossible, so we just avoid it here. | |
981 | */ | |
982 | lock_page_nosync(p); | |
847ce401 WF |
983 | |
984 | /* | |
985 | * unpoison always clear PG_hwpoison inside page lock | |
986 | */ | |
987 | if (!PageHWPoison(p)) { | |
d95ea51e | 988 | printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn); |
847ce401 WF |
989 | res = 0; |
990 | goto out; | |
991 | } | |
7c116f2b WF |
992 | if (hwpoison_filter(p)) { |
993 | if (TestClearPageHWPoison(p)) | |
994 | atomic_long_dec(&mce_bad_pages); | |
995 | unlock_page(p); | |
996 | put_page(p); | |
997 | return 0; | |
998 | } | |
847ce401 | 999 | |
6a46079c AK |
1000 | wait_on_page_writeback(p); |
1001 | ||
1002 | /* | |
1003 | * Now take care of user space mappings. | |
1668bfd5 | 1004 | * Abort on fail: __remove_from_page_cache() assumes unmapped page. |
6a46079c | 1005 | */ |
1668bfd5 WF |
1006 | if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) { |
1007 | printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn); | |
1008 | res = -EBUSY; | |
1009 | goto out; | |
1010 | } | |
6a46079c AK |
1011 | |
1012 | /* | |
1013 | * Torn down by someone else? | |
1014 | */ | |
dc2a1cbf | 1015 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
6a46079c | 1016 | action_result(pfn, "already truncated LRU", IGNORED); |
d95ea51e | 1017 | res = -EBUSY; |
6a46079c AK |
1018 | goto out; |
1019 | } | |
1020 | ||
1021 | res = -EBUSY; | |
1022 | for (ps = error_states;; ps++) { | |
dc2a1cbf | 1023 | if ((p->flags & ps->mask) == ps->res) { |
bd1ce5f9 | 1024 | res = page_action(ps, p, pfn); |
6a46079c AK |
1025 | break; |
1026 | } | |
1027 | } | |
1028 | out: | |
1029 | unlock_page(p); | |
1030 | return res; | |
1031 | } | |
1032 | EXPORT_SYMBOL_GPL(__memory_failure); | |
1033 | ||
1034 | /** | |
1035 | * memory_failure - Handle memory failure of a page. | |
1036 | * @pfn: Page Number of the corrupted page | |
1037 | * @trapno: Trap number reported in the signal to user space. | |
1038 | * | |
1039 | * This function is called by the low level machine check code | |
1040 | * of an architecture when it detects hardware memory corruption | |
1041 | * of a page. It tries its best to recover, which includes | |
1042 | * dropping pages, killing processes etc. | |
1043 | * | |
1044 | * The function is primarily of use for corruptions that | |
1045 | * happen outside the current execution context (e.g. when | |
1046 | * detected by a background scrubber) | |
1047 | * | |
1048 | * Must run in process context (e.g. a work queue) with interrupts | |
1049 | * enabled and no spinlocks hold. | |
1050 | */ | |
1051 | void memory_failure(unsigned long pfn, int trapno) | |
1052 | { | |
1053 | __memory_failure(pfn, trapno, 0); | |
1054 | } | |
847ce401 WF |
1055 | |
1056 | /** | |
1057 | * unpoison_memory - Unpoison a previously poisoned page | |
1058 | * @pfn: Page number of the to be unpoisoned page | |
1059 | * | |
1060 | * Software-unpoison a page that has been poisoned by | |
1061 | * memory_failure() earlier. | |
1062 | * | |
1063 | * This is only done on the software-level, so it only works | |
1064 | * for linux injected failures, not real hardware failures | |
1065 | * | |
1066 | * Returns 0 for success, otherwise -errno. | |
1067 | */ | |
1068 | int unpoison_memory(unsigned long pfn) | |
1069 | { | |
1070 | struct page *page; | |
1071 | struct page *p; | |
1072 | int freeit = 0; | |
1073 | ||
1074 | if (!pfn_valid(pfn)) | |
1075 | return -ENXIO; | |
1076 | ||
1077 | p = pfn_to_page(pfn); | |
1078 | page = compound_head(p); | |
1079 | ||
1080 | if (!PageHWPoison(p)) { | |
1081 | pr_debug("MCE: Page was already unpoisoned %#lx\n", pfn); | |
1082 | return 0; | |
1083 | } | |
1084 | ||
1085 | if (!get_page_unless_zero(page)) { | |
1086 | if (TestClearPageHWPoison(p)) | |
1087 | atomic_long_dec(&mce_bad_pages); | |
1088 | pr_debug("MCE: Software-unpoisoned free page %#lx\n", pfn); | |
1089 | return 0; | |
1090 | } | |
1091 | ||
1092 | lock_page_nosync(page); | |
1093 | /* | |
1094 | * This test is racy because PG_hwpoison is set outside of page lock. | |
1095 | * That's acceptable because that won't trigger kernel panic. Instead, | |
1096 | * the PG_hwpoison page will be caught and isolated on the entrance to | |
1097 | * the free buddy page pool. | |
1098 | */ | |
1099 | if (TestClearPageHWPoison(p)) { | |
1100 | pr_debug("MCE: Software-unpoisoned page %#lx\n", pfn); | |
1101 | atomic_long_dec(&mce_bad_pages); | |
1102 | freeit = 1; | |
1103 | } | |
1104 | unlock_page(page); | |
1105 | ||
1106 | put_page(page); | |
1107 | if (freeit) | |
1108 | put_page(page); | |
1109 | ||
1110 | return 0; | |
1111 | } | |
1112 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 AK |
1113 | |
1114 | static struct page *new_page(struct page *p, unsigned long private, int **x) | |
1115 | { | |
1116 | return alloc_pages(GFP_HIGHUSER_MOVABLE, 0); | |
1117 | } | |
1118 | ||
1119 | /* | |
1120 | * Safely get reference count of an arbitrary page. | |
1121 | * Returns 0 for a free page, -EIO for a zero refcount page | |
1122 | * that is not free, and 1 for any other page type. | |
1123 | * For 1 the page is returned with increased page count, otherwise not. | |
1124 | */ | |
1125 | static int get_any_page(struct page *p, unsigned long pfn, int flags) | |
1126 | { | |
1127 | int ret; | |
1128 | ||
1129 | if (flags & MF_COUNT_INCREASED) | |
1130 | return 1; | |
1131 | ||
1132 | /* | |
1133 | * The lock_system_sleep prevents a race with memory hotplug, | |
1134 | * because the isolation assumes there's only a single user. | |
1135 | * This is a big hammer, a better would be nicer. | |
1136 | */ | |
1137 | lock_system_sleep(); | |
1138 | ||
1139 | /* | |
1140 | * Isolate the page, so that it doesn't get reallocated if it | |
1141 | * was free. | |
1142 | */ | |
1143 | set_migratetype_isolate(p); | |
1144 | if (!get_page_unless_zero(compound_head(p))) { | |
1145 | if (is_free_buddy_page(p)) { | |
1146 | pr_debug("get_any_page: %#lx free buddy page\n", pfn); | |
1147 | /* Set hwpoison bit while page is still isolated */ | |
1148 | SetPageHWPoison(p); | |
1149 | ret = 0; | |
1150 | } else { | |
1151 | pr_debug("get_any_page: %#lx: unknown zero refcount page type %lx\n", | |
1152 | pfn, p->flags); | |
1153 | ret = -EIO; | |
1154 | } | |
1155 | } else { | |
1156 | /* Not a free page */ | |
1157 | ret = 1; | |
1158 | } | |
1159 | unset_migratetype_isolate(p); | |
1160 | unlock_system_sleep(); | |
1161 | return ret; | |
1162 | } | |
1163 | ||
1164 | /** | |
1165 | * soft_offline_page - Soft offline a page. | |
1166 | * @page: page to offline | |
1167 | * @flags: flags. Same as memory_failure(). | |
1168 | * | |
1169 | * Returns 0 on success, otherwise negated errno. | |
1170 | * | |
1171 | * Soft offline a page, by migration or invalidation, | |
1172 | * without killing anything. This is for the case when | |
1173 | * a page is not corrupted yet (so it's still valid to access), | |
1174 | * but has had a number of corrected errors and is better taken | |
1175 | * out. | |
1176 | * | |
1177 | * The actual policy on when to do that is maintained by | |
1178 | * user space. | |
1179 | * | |
1180 | * This should never impact any application or cause data loss, | |
1181 | * however it might take some time. | |
1182 | * | |
1183 | * This is not a 100% solution for all memory, but tries to be | |
1184 | * ``good enough'' for the majority of memory. | |
1185 | */ | |
1186 | int soft_offline_page(struct page *page, int flags) | |
1187 | { | |
1188 | int ret; | |
1189 | unsigned long pfn = page_to_pfn(page); | |
1190 | ||
1191 | ret = get_any_page(page, pfn, flags); | |
1192 | if (ret < 0) | |
1193 | return ret; | |
1194 | if (ret == 0) | |
1195 | goto done; | |
1196 | ||
1197 | /* | |
1198 | * Page cache page we can handle? | |
1199 | */ | |
1200 | if (!PageLRU(page)) { | |
1201 | /* | |
1202 | * Try to free it. | |
1203 | */ | |
1204 | put_page(page); | |
1205 | shake_page(page, 1); | |
1206 | ||
1207 | /* | |
1208 | * Did it turn free? | |
1209 | */ | |
1210 | ret = get_any_page(page, pfn, 0); | |
1211 | if (ret < 0) | |
1212 | return ret; | |
1213 | if (ret == 0) | |
1214 | goto done; | |
1215 | } | |
1216 | if (!PageLRU(page)) { | |
1217 | pr_debug("soft_offline: %#lx: unknown non LRU page type %lx\n", | |
1218 | pfn, page->flags); | |
1219 | return -EIO; | |
1220 | } | |
1221 | ||
1222 | lock_page(page); | |
1223 | wait_on_page_writeback(page); | |
1224 | ||
1225 | /* | |
1226 | * Synchronized using the page lock with memory_failure() | |
1227 | */ | |
1228 | if (PageHWPoison(page)) { | |
1229 | unlock_page(page); | |
1230 | put_page(page); | |
1231 | pr_debug("soft offline: %#lx page already poisoned\n", pfn); | |
1232 | return -EBUSY; | |
1233 | } | |
1234 | ||
1235 | /* | |
1236 | * Try to invalidate first. This should work for | |
1237 | * non dirty unmapped page cache pages. | |
1238 | */ | |
1239 | ret = invalidate_inode_page(page); | |
1240 | unlock_page(page); | |
1241 | ||
1242 | /* | |
1243 | * Drop count because page migration doesn't like raised | |
1244 | * counts. The page could get re-allocated, but if it becomes | |
1245 | * LRU the isolation will just fail. | |
1246 | * RED-PEN would be better to keep it isolated here, but we | |
1247 | * would need to fix isolation locking first. | |
1248 | */ | |
1249 | put_page(page); | |
1250 | if (ret == 1) { | |
1251 | ret = 0; | |
1252 | pr_debug("soft_offline: %#lx: invalidated\n", pfn); | |
1253 | goto done; | |
1254 | } | |
1255 | ||
1256 | /* | |
1257 | * Simple invalidation didn't work. | |
1258 | * Try to migrate to a new page instead. migrate.c | |
1259 | * handles a large number of cases for us. | |
1260 | */ | |
1261 | ret = isolate_lru_page(page); | |
1262 | if (!ret) { | |
1263 | LIST_HEAD(pagelist); | |
1264 | ||
1265 | list_add(&page->lru, &pagelist); | |
1266 | ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0); | |
1267 | if (ret) { | |
1268 | pr_debug("soft offline: %#lx: migration failed %d, type %lx\n", | |
1269 | pfn, ret, page->flags); | |
1270 | if (ret > 0) | |
1271 | ret = -EIO; | |
1272 | } | |
1273 | } else { | |
1274 | pr_debug("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n", | |
1275 | pfn, ret, page_count(page), page->flags); | |
1276 | } | |
1277 | if (ret) | |
1278 | return ret; | |
1279 | ||
1280 | done: | |
1281 | atomic_long_add(1, &mce_bad_pages); | |
1282 | SetPageHWPoison(page); | |
1283 | /* keep elevated page count for bad page */ | |
1284 | return ret; | |
1285 | } |