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 | |
1c80b990 | 10 | * hardware as being corrupted usually due to a multi-bit ECC memory or cache |
6a46079c | 11 | * failure. |
1c80b990 AK |
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. | |
6a46079c AK |
15 | * |
16 | * Handles page cache pages in various states. The tricky part | |
1c80b990 AK |
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. | |
6a46079c AK |
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 | */ | |
6a46079c AK |
38 | #include <linux/kernel.h> |
39 | #include <linux/mm.h> | |
40 | #include <linux/page-flags.h> | |
478c5ffc | 41 | #include <linux/kernel-page-flags.h> |
6a46079c | 42 | #include <linux/sched.h> |
01e00f88 | 43 | #include <linux/ksm.h> |
6a46079c AK |
44 | #include <linux/rmap.h> |
45 | #include <linux/pagemap.h> | |
46 | #include <linux/swap.h> | |
47 | #include <linux/backing-dev.h> | |
facb6011 AK |
48 | #include <linux/migrate.h> |
49 | #include <linux/page-isolation.h> | |
50 | #include <linux/suspend.h> | |
5a0e3ad6 | 51 | #include <linux/slab.h> |
bf998156 | 52 | #include <linux/swapops.h> |
7af446a8 | 53 | #include <linux/hugetlb.h> |
20d6c96b | 54 | #include <linux/memory_hotplug.h> |
6a46079c AK |
55 | #include "internal.h" |
56 | ||
57 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
58 | ||
59 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
60 | ||
61 | atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0); | |
62 | ||
27df5068 AK |
63 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
64 | ||
1bfe5feb | 65 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
66 | u32 hwpoison_filter_dev_major = ~0U; |
67 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
68 | u64 hwpoison_filter_flags_mask; |
69 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 70 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
71 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
72 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
73 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
74 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
75 | |
76 | static int hwpoison_filter_dev(struct page *p) | |
77 | { | |
78 | struct address_space *mapping; | |
79 | dev_t dev; | |
80 | ||
81 | if (hwpoison_filter_dev_major == ~0U && | |
82 | hwpoison_filter_dev_minor == ~0U) | |
83 | return 0; | |
84 | ||
85 | /* | |
1c80b990 | 86 | * page_mapping() does not accept slab pages. |
7c116f2b WF |
87 | */ |
88 | if (PageSlab(p)) | |
89 | return -EINVAL; | |
90 | ||
91 | mapping = page_mapping(p); | |
92 | if (mapping == NULL || mapping->host == NULL) | |
93 | return -EINVAL; | |
94 | ||
95 | dev = mapping->host->i_sb->s_dev; | |
96 | if (hwpoison_filter_dev_major != ~0U && | |
97 | hwpoison_filter_dev_major != MAJOR(dev)) | |
98 | return -EINVAL; | |
99 | if (hwpoison_filter_dev_minor != ~0U && | |
100 | hwpoison_filter_dev_minor != MINOR(dev)) | |
101 | return -EINVAL; | |
102 | ||
103 | return 0; | |
104 | } | |
105 | ||
478c5ffc WF |
106 | static int hwpoison_filter_flags(struct page *p) |
107 | { | |
108 | if (!hwpoison_filter_flags_mask) | |
109 | return 0; | |
110 | ||
111 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
112 | hwpoison_filter_flags_value) | |
113 | return 0; | |
114 | else | |
115 | return -EINVAL; | |
116 | } | |
117 | ||
4fd466eb AK |
118 | /* |
119 | * This allows stress tests to limit test scope to a collection of tasks | |
120 | * by putting them under some memcg. This prevents killing unrelated/important | |
121 | * processes such as /sbin/init. Note that the target task may share clean | |
122 | * pages with init (eg. libc text), which is harmless. If the target task | |
123 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
124 | * is also included in the memcg. At last, due to race conditions this filter | |
125 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
126 | * a freed page. | |
127 | */ | |
128 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
129 | u64 hwpoison_filter_memcg; | |
130 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
131 | static int hwpoison_filter_task(struct page *p) | |
132 | { | |
133 | struct mem_cgroup *mem; | |
134 | struct cgroup_subsys_state *css; | |
135 | unsigned long ino; | |
136 | ||
137 | if (!hwpoison_filter_memcg) | |
138 | return 0; | |
139 | ||
140 | mem = try_get_mem_cgroup_from_page(p); | |
141 | if (!mem) | |
142 | return -EINVAL; | |
143 | ||
144 | css = mem_cgroup_css(mem); | |
145 | /* root_mem_cgroup has NULL dentries */ | |
146 | if (!css->cgroup->dentry) | |
147 | return -EINVAL; | |
148 | ||
149 | ino = css->cgroup->dentry->d_inode->i_ino; | |
150 | css_put(css); | |
151 | ||
152 | if (ino != hwpoison_filter_memcg) | |
153 | return -EINVAL; | |
154 | ||
155 | return 0; | |
156 | } | |
157 | #else | |
158 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
159 | #endif | |
160 | ||
7c116f2b WF |
161 | int hwpoison_filter(struct page *p) |
162 | { | |
1bfe5feb HL |
163 | if (!hwpoison_filter_enable) |
164 | return 0; | |
165 | ||
7c116f2b WF |
166 | if (hwpoison_filter_dev(p)) |
167 | return -EINVAL; | |
168 | ||
478c5ffc WF |
169 | if (hwpoison_filter_flags(p)) |
170 | return -EINVAL; | |
171 | ||
4fd466eb AK |
172 | if (hwpoison_filter_task(p)) |
173 | return -EINVAL; | |
174 | ||
7c116f2b WF |
175 | return 0; |
176 | } | |
27df5068 AK |
177 | #else |
178 | int hwpoison_filter(struct page *p) | |
179 | { | |
180 | return 0; | |
181 | } | |
182 | #endif | |
183 | ||
7c116f2b WF |
184 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
185 | ||
6a46079c AK |
186 | /* |
187 | * Send all the processes who have the page mapped an ``action optional'' | |
188 | * signal. | |
189 | */ | |
190 | static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno, | |
0d9ee6a2 | 191 | unsigned long pfn, struct page *page) |
6a46079c AK |
192 | { |
193 | struct siginfo si; | |
194 | int ret; | |
195 | ||
196 | printk(KERN_ERR | |
197 | "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n", | |
198 | pfn, t->comm, t->pid); | |
199 | si.si_signo = SIGBUS; | |
200 | si.si_errno = 0; | |
201 | si.si_code = BUS_MCEERR_AO; | |
202 | si.si_addr = (void *)addr; | |
203 | #ifdef __ARCH_SI_TRAPNO | |
204 | si.si_trapno = trapno; | |
205 | #endif | |
37c2ac78 | 206 | si.si_addr_lsb = compound_trans_order(compound_head(page)) + PAGE_SHIFT; |
6a46079c AK |
207 | /* |
208 | * Don't use force here, it's convenient if the signal | |
209 | * can be temporarily blocked. | |
210 | * This could cause a loop when the user sets SIGBUS | |
25985edc | 211 | * to SIG_IGN, but hopefully no one will do that? |
6a46079c AK |
212 | */ |
213 | ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */ | |
214 | if (ret < 0) | |
215 | printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n", | |
216 | t->comm, t->pid, ret); | |
217 | return ret; | |
218 | } | |
219 | ||
588f9ce6 AK |
220 | /* |
221 | * When a unknown page type is encountered drain as many buffers as possible | |
222 | * in the hope to turn the page into a LRU or free page, which we can handle. | |
223 | */ | |
facb6011 | 224 | void shake_page(struct page *p, int access) |
588f9ce6 AK |
225 | { |
226 | if (!PageSlab(p)) { | |
227 | lru_add_drain_all(); | |
228 | if (PageLRU(p)) | |
229 | return; | |
230 | drain_all_pages(); | |
231 | if (PageLRU(p) || is_free_buddy_page(p)) | |
232 | return; | |
233 | } | |
facb6011 | 234 | |
588f9ce6 | 235 | /* |
af241a08 JD |
236 | * Only call shrink_slab here (which would also shrink other caches) if |
237 | * access is not potentially fatal. | |
588f9ce6 | 238 | */ |
facb6011 AK |
239 | if (access) { |
240 | int nr; | |
241 | do { | |
a09ed5e0 YH |
242 | struct shrink_control shrink = { |
243 | .gfp_mask = GFP_KERNEL, | |
a09ed5e0 YH |
244 | }; |
245 | ||
1495f230 | 246 | nr = shrink_slab(&shrink, 1000, 1000); |
47f43e7e | 247 | if (page_count(p) == 1) |
facb6011 AK |
248 | break; |
249 | } while (nr > 10); | |
250 | } | |
588f9ce6 AK |
251 | } |
252 | EXPORT_SYMBOL_GPL(shake_page); | |
253 | ||
6a46079c AK |
254 | /* |
255 | * Kill all processes that have a poisoned page mapped and then isolate | |
256 | * the page. | |
257 | * | |
258 | * General strategy: | |
259 | * Find all processes having the page mapped and kill them. | |
260 | * But we keep a page reference around so that the page is not | |
261 | * actually freed yet. | |
262 | * Then stash the page away | |
263 | * | |
264 | * There's no convenient way to get back to mapped processes | |
265 | * from the VMAs. So do a brute-force search over all | |
266 | * running processes. | |
267 | * | |
268 | * Remember that machine checks are not common (or rather | |
269 | * if they are common you have other problems), so this shouldn't | |
270 | * be a performance issue. | |
271 | * | |
272 | * Also there are some races possible while we get from the | |
273 | * error detection to actually handle it. | |
274 | */ | |
275 | ||
276 | struct to_kill { | |
277 | struct list_head nd; | |
278 | struct task_struct *tsk; | |
279 | unsigned long addr; | |
9033ae16 | 280 | char addr_valid; |
6a46079c AK |
281 | }; |
282 | ||
283 | /* | |
284 | * Failure handling: if we can't find or can't kill a process there's | |
285 | * not much we can do. We just print a message and ignore otherwise. | |
286 | */ | |
287 | ||
288 | /* | |
289 | * Schedule a process for later kill. | |
290 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
291 | * TBD would GFP_NOIO be enough? | |
292 | */ | |
293 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
294 | struct vm_area_struct *vma, | |
295 | struct list_head *to_kill, | |
296 | struct to_kill **tkc) | |
297 | { | |
298 | struct to_kill *tk; | |
299 | ||
300 | if (*tkc) { | |
301 | tk = *tkc; | |
302 | *tkc = NULL; | |
303 | } else { | |
304 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); | |
305 | if (!tk) { | |
306 | printk(KERN_ERR | |
307 | "MCE: Out of memory while machine check handling\n"); | |
308 | return; | |
309 | } | |
310 | } | |
311 | tk->addr = page_address_in_vma(p, vma); | |
312 | tk->addr_valid = 1; | |
313 | ||
314 | /* | |
315 | * In theory we don't have to kill when the page was | |
316 | * munmaped. But it could be also a mremap. Since that's | |
317 | * likely very rare kill anyways just out of paranoia, but use | |
318 | * a SIGKILL because the error is not contained anymore. | |
319 | */ | |
320 | if (tk->addr == -EFAULT) { | |
fb46e735 | 321 | pr_info("MCE: Unable to find user space address %lx in %s\n", |
6a46079c AK |
322 | page_to_pfn(p), tsk->comm); |
323 | tk->addr_valid = 0; | |
324 | } | |
325 | get_task_struct(tsk); | |
326 | tk->tsk = tsk; | |
327 | list_add_tail(&tk->nd, to_kill); | |
328 | } | |
329 | ||
330 | /* | |
331 | * Kill the processes that have been collected earlier. | |
332 | * | |
333 | * Only do anything when DOIT is set, otherwise just free the list | |
334 | * (this is used for clean pages which do not need killing) | |
335 | * Also when FAIL is set do a force kill because something went | |
336 | * wrong earlier. | |
337 | */ | |
338 | static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno, | |
0d9ee6a2 | 339 | int fail, struct page *page, unsigned long pfn) |
6a46079c AK |
340 | { |
341 | struct to_kill *tk, *next; | |
342 | ||
343 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
344 | if (doit) { | |
345 | /* | |
af901ca1 | 346 | * In case something went wrong with munmapping |
6a46079c AK |
347 | * make sure the process doesn't catch the |
348 | * signal and then access the memory. Just kill it. | |
6a46079c AK |
349 | */ |
350 | if (fail || tk->addr_valid == 0) { | |
351 | printk(KERN_ERR | |
352 | "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", | |
353 | pfn, tk->tsk->comm, tk->tsk->pid); | |
354 | force_sig(SIGKILL, tk->tsk); | |
355 | } | |
356 | ||
357 | /* | |
358 | * In theory the process could have mapped | |
359 | * something else on the address in-between. We could | |
360 | * check for that, but we need to tell the | |
361 | * process anyways. | |
362 | */ | |
363 | else if (kill_proc_ao(tk->tsk, tk->addr, trapno, | |
0d9ee6a2 | 364 | pfn, page) < 0) |
6a46079c AK |
365 | printk(KERN_ERR |
366 | "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n", | |
367 | pfn, tk->tsk->comm, tk->tsk->pid); | |
368 | } | |
369 | put_task_struct(tk->tsk); | |
370 | kfree(tk); | |
371 | } | |
372 | } | |
373 | ||
374 | static int task_early_kill(struct task_struct *tsk) | |
375 | { | |
376 | if (!tsk->mm) | |
377 | return 0; | |
378 | if (tsk->flags & PF_MCE_PROCESS) | |
379 | return !!(tsk->flags & PF_MCE_EARLY); | |
380 | return sysctl_memory_failure_early_kill; | |
381 | } | |
382 | ||
383 | /* | |
384 | * Collect processes when the error hit an anonymous page. | |
385 | */ | |
386 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
387 | struct to_kill **tkc) | |
388 | { | |
389 | struct vm_area_struct *vma; | |
390 | struct task_struct *tsk; | |
391 | struct anon_vma *av; | |
392 | ||
393 | read_lock(&tasklist_lock); | |
394 | av = page_lock_anon_vma(page); | |
395 | if (av == NULL) /* Not actually mapped anymore */ | |
396 | goto out; | |
397 | for_each_process (tsk) { | |
5beb4930 RR |
398 | struct anon_vma_chain *vmac; |
399 | ||
6a46079c AK |
400 | if (!task_early_kill(tsk)) |
401 | continue; | |
5beb4930 RR |
402 | list_for_each_entry(vmac, &av->head, same_anon_vma) { |
403 | vma = vmac->vma; | |
6a46079c AK |
404 | if (!page_mapped_in_vma(page, vma)) |
405 | continue; | |
406 | if (vma->vm_mm == tsk->mm) | |
407 | add_to_kill(tsk, page, vma, to_kill, tkc); | |
408 | } | |
409 | } | |
410 | page_unlock_anon_vma(av); | |
411 | out: | |
412 | read_unlock(&tasklist_lock); | |
413 | } | |
414 | ||
415 | /* | |
416 | * Collect processes when the error hit a file mapped page. | |
417 | */ | |
418 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
419 | struct to_kill **tkc) | |
420 | { | |
421 | struct vm_area_struct *vma; | |
422 | struct task_struct *tsk; | |
423 | struct prio_tree_iter iter; | |
424 | struct address_space *mapping = page->mapping; | |
425 | ||
426 | /* | |
427 | * A note on the locking order between the two locks. | |
428 | * We don't rely on this particular order. | |
429 | * If you have some other code that needs a different order | |
430 | * feel free to switch them around. Or add a reverse link | |
431 | * from mm_struct to task_struct, then this could be all | |
432 | * done without taking tasklist_lock and looping over all tasks. | |
433 | */ | |
434 | ||
435 | read_lock(&tasklist_lock); | |
3d48ae45 | 436 | mutex_lock(&mapping->i_mmap_mutex); |
6a46079c AK |
437 | for_each_process(tsk) { |
438 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
439 | ||
440 | if (!task_early_kill(tsk)) | |
441 | continue; | |
442 | ||
443 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, | |
444 | pgoff) { | |
445 | /* | |
446 | * Send early kill signal to tasks where a vma covers | |
447 | * the page but the corrupted page is not necessarily | |
448 | * mapped it in its pte. | |
449 | * Assume applications who requested early kill want | |
450 | * to be informed of all such data corruptions. | |
451 | */ | |
452 | if (vma->vm_mm == tsk->mm) | |
453 | add_to_kill(tsk, page, vma, to_kill, tkc); | |
454 | } | |
455 | } | |
3d48ae45 | 456 | mutex_unlock(&mapping->i_mmap_mutex); |
6a46079c AK |
457 | read_unlock(&tasklist_lock); |
458 | } | |
459 | ||
460 | /* | |
461 | * Collect the processes who have the corrupted page mapped to kill. | |
462 | * This is done in two steps for locking reasons. | |
463 | * First preallocate one tokill structure outside the spin locks, | |
464 | * so that we can kill at least one process reasonably reliable. | |
465 | */ | |
466 | static void collect_procs(struct page *page, struct list_head *tokill) | |
467 | { | |
468 | struct to_kill *tk; | |
469 | ||
470 | if (!page->mapping) | |
471 | return; | |
472 | ||
473 | tk = kmalloc(sizeof(struct to_kill), GFP_NOIO); | |
474 | if (!tk) | |
475 | return; | |
476 | if (PageAnon(page)) | |
477 | collect_procs_anon(page, tokill, &tk); | |
478 | else | |
479 | collect_procs_file(page, tokill, &tk); | |
480 | kfree(tk); | |
481 | } | |
482 | ||
483 | /* | |
484 | * Error handlers for various types of pages. | |
485 | */ | |
486 | ||
487 | enum outcome { | |
d95ea51e WF |
488 | IGNORED, /* Error: cannot be handled */ |
489 | FAILED, /* Error: handling failed */ | |
6a46079c | 490 | DELAYED, /* Will be handled later */ |
6a46079c AK |
491 | RECOVERED, /* Successfully recovered */ |
492 | }; | |
493 | ||
494 | static const char *action_name[] = { | |
d95ea51e | 495 | [IGNORED] = "Ignored", |
6a46079c AK |
496 | [FAILED] = "Failed", |
497 | [DELAYED] = "Delayed", | |
6a46079c AK |
498 | [RECOVERED] = "Recovered", |
499 | }; | |
500 | ||
dc2a1cbf WF |
501 | /* |
502 | * XXX: It is possible that a page is isolated from LRU cache, | |
503 | * and then kept in swap cache or failed to remove from page cache. | |
504 | * The page count will stop it from being freed by unpoison. | |
505 | * Stress tests should be aware of this memory leak problem. | |
506 | */ | |
507 | static int delete_from_lru_cache(struct page *p) | |
508 | { | |
509 | if (!isolate_lru_page(p)) { | |
510 | /* | |
511 | * Clear sensible page flags, so that the buddy system won't | |
512 | * complain when the page is unpoison-and-freed. | |
513 | */ | |
514 | ClearPageActive(p); | |
515 | ClearPageUnevictable(p); | |
516 | /* | |
517 | * drop the page count elevated by isolate_lru_page() | |
518 | */ | |
519 | page_cache_release(p); | |
520 | return 0; | |
521 | } | |
522 | return -EIO; | |
523 | } | |
524 | ||
6a46079c AK |
525 | /* |
526 | * Error hit kernel page. | |
527 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
528 | * could be more sophisticated. | |
529 | */ | |
530 | static int me_kernel(struct page *p, unsigned long pfn) | |
6a46079c AK |
531 | { |
532 | return IGNORED; | |
533 | } | |
534 | ||
535 | /* | |
536 | * Page in unknown state. Do nothing. | |
537 | */ | |
538 | static int me_unknown(struct page *p, unsigned long pfn) | |
539 | { | |
540 | printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn); | |
541 | return FAILED; | |
542 | } | |
543 | ||
6a46079c AK |
544 | /* |
545 | * Clean (or cleaned) page cache page. | |
546 | */ | |
547 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
548 | { | |
549 | int err; | |
550 | int ret = FAILED; | |
551 | struct address_space *mapping; | |
552 | ||
dc2a1cbf WF |
553 | delete_from_lru_cache(p); |
554 | ||
6a46079c AK |
555 | /* |
556 | * For anonymous pages we're done the only reference left | |
557 | * should be the one m_f() holds. | |
558 | */ | |
559 | if (PageAnon(p)) | |
560 | return RECOVERED; | |
561 | ||
562 | /* | |
563 | * Now truncate the page in the page cache. This is really | |
564 | * more like a "temporary hole punch" | |
565 | * Don't do this for block devices when someone else | |
566 | * has a reference, because it could be file system metadata | |
567 | * and that's not safe to truncate. | |
568 | */ | |
569 | mapping = page_mapping(p); | |
570 | if (!mapping) { | |
571 | /* | |
572 | * Page has been teared down in the meanwhile | |
573 | */ | |
574 | return FAILED; | |
575 | } | |
576 | ||
577 | /* | |
578 | * Truncation is a bit tricky. Enable it per file system for now. | |
579 | * | |
580 | * Open: to take i_mutex or not for this? Right now we don't. | |
581 | */ | |
582 | if (mapping->a_ops->error_remove_page) { | |
583 | err = mapping->a_ops->error_remove_page(mapping, p); | |
584 | if (err != 0) { | |
585 | printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n", | |
586 | pfn, err); | |
587 | } else if (page_has_private(p) && | |
588 | !try_to_release_page(p, GFP_NOIO)) { | |
fb46e735 | 589 | pr_info("MCE %#lx: failed to release buffers\n", pfn); |
6a46079c AK |
590 | } else { |
591 | ret = RECOVERED; | |
592 | } | |
593 | } else { | |
594 | /* | |
595 | * If the file system doesn't support it just invalidate | |
596 | * This fails on dirty or anything with private pages | |
597 | */ | |
598 | if (invalidate_inode_page(p)) | |
599 | ret = RECOVERED; | |
600 | else | |
601 | printk(KERN_INFO "MCE %#lx: Failed to invalidate\n", | |
602 | pfn); | |
603 | } | |
604 | return ret; | |
605 | } | |
606 | ||
607 | /* | |
608 | * Dirty cache page page | |
609 | * Issues: when the error hit a hole page the error is not properly | |
610 | * propagated. | |
611 | */ | |
612 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
613 | { | |
614 | struct address_space *mapping = page_mapping(p); | |
615 | ||
616 | SetPageError(p); | |
617 | /* TBD: print more information about the file. */ | |
618 | if (mapping) { | |
619 | /* | |
620 | * IO error will be reported by write(), fsync(), etc. | |
621 | * who check the mapping. | |
622 | * This way the application knows that something went | |
623 | * wrong with its dirty file data. | |
624 | * | |
625 | * There's one open issue: | |
626 | * | |
627 | * The EIO will be only reported on the next IO | |
628 | * operation and then cleared through the IO map. | |
629 | * Normally Linux has two mechanisms to pass IO error | |
630 | * first through the AS_EIO flag in the address space | |
631 | * and then through the PageError flag in the page. | |
632 | * Since we drop pages on memory failure handling the | |
633 | * only mechanism open to use is through AS_AIO. | |
634 | * | |
635 | * This has the disadvantage that it gets cleared on | |
636 | * the first operation that returns an error, while | |
637 | * the PageError bit is more sticky and only cleared | |
638 | * when the page is reread or dropped. If an | |
639 | * application assumes it will always get error on | |
640 | * fsync, but does other operations on the fd before | |
25985edc | 641 | * and the page is dropped between then the error |
6a46079c AK |
642 | * will not be properly reported. |
643 | * | |
644 | * This can already happen even without hwpoisoned | |
645 | * pages: first on metadata IO errors (which only | |
646 | * report through AS_EIO) or when the page is dropped | |
647 | * at the wrong time. | |
648 | * | |
649 | * So right now we assume that the application DTRT on | |
650 | * the first EIO, but we're not worse than other parts | |
651 | * of the kernel. | |
652 | */ | |
653 | mapping_set_error(mapping, EIO); | |
654 | } | |
655 | ||
656 | return me_pagecache_clean(p, pfn); | |
657 | } | |
658 | ||
659 | /* | |
660 | * Clean and dirty swap cache. | |
661 | * | |
662 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
663 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
664 | * referenced concurrently by 2 types of PTEs: | |
665 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
666 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
667 | * and then | |
668 | * - clear dirty bit to prevent IO | |
669 | * - remove from LRU | |
670 | * - but keep in the swap cache, so that when we return to it on | |
671 | * a later page fault, we know the application is accessing | |
672 | * corrupted data and shall be killed (we installed simple | |
673 | * interception code in do_swap_page to catch it). | |
674 | * | |
675 | * Clean swap cache pages can be directly isolated. A later page fault will | |
676 | * bring in the known good data from disk. | |
677 | */ | |
678 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
679 | { | |
6a46079c AK |
680 | ClearPageDirty(p); |
681 | /* Trigger EIO in shmem: */ | |
682 | ClearPageUptodate(p); | |
683 | ||
dc2a1cbf WF |
684 | if (!delete_from_lru_cache(p)) |
685 | return DELAYED; | |
686 | else | |
687 | return FAILED; | |
6a46079c AK |
688 | } |
689 | ||
690 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
691 | { | |
6a46079c | 692 | delete_from_swap_cache(p); |
e43c3afb | 693 | |
dc2a1cbf WF |
694 | if (!delete_from_lru_cache(p)) |
695 | return RECOVERED; | |
696 | else | |
697 | return FAILED; | |
6a46079c AK |
698 | } |
699 | ||
700 | /* | |
701 | * Huge pages. Needs work. | |
702 | * Issues: | |
93f70f90 NH |
703 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
704 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c AK |
705 | */ |
706 | static int me_huge_page(struct page *p, unsigned long pfn) | |
707 | { | |
6de2b1aa | 708 | int res = 0; |
93f70f90 NH |
709 | struct page *hpage = compound_head(p); |
710 | /* | |
711 | * We can safely recover from error on free or reserved (i.e. | |
712 | * not in-use) hugepage by dequeuing it from freelist. | |
713 | * To check whether a hugepage is in-use or not, we can't use | |
714 | * page->lru because it can be used in other hugepage operations, | |
715 | * such as __unmap_hugepage_range() and gather_surplus_pages(). | |
716 | * So instead we use page_mapping() and PageAnon(). | |
717 | * We assume that this function is called with page lock held, | |
718 | * so there is no race between isolation and mapping/unmapping. | |
719 | */ | |
720 | if (!(page_mapping(hpage) || PageAnon(hpage))) { | |
6de2b1aa NH |
721 | res = dequeue_hwpoisoned_huge_page(hpage); |
722 | if (!res) | |
723 | return RECOVERED; | |
93f70f90 NH |
724 | } |
725 | return DELAYED; | |
6a46079c AK |
726 | } |
727 | ||
728 | /* | |
729 | * Various page states we can handle. | |
730 | * | |
731 | * A page state is defined by its current page->flags bits. | |
732 | * The table matches them in order and calls the right handler. | |
733 | * | |
734 | * This is quite tricky because we can access page at any time | |
25985edc | 735 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
736 | * |
737 | * This is not complete. More states could be added. | |
738 | * For any missing state don't attempt recovery. | |
739 | */ | |
740 | ||
741 | #define dirty (1UL << PG_dirty) | |
742 | #define sc (1UL << PG_swapcache) | |
743 | #define unevict (1UL << PG_unevictable) | |
744 | #define mlock (1UL << PG_mlocked) | |
745 | #define writeback (1UL << PG_writeback) | |
746 | #define lru (1UL << PG_lru) | |
747 | #define swapbacked (1UL << PG_swapbacked) | |
748 | #define head (1UL << PG_head) | |
749 | #define tail (1UL << PG_tail) | |
750 | #define compound (1UL << PG_compound) | |
751 | #define slab (1UL << PG_slab) | |
6a46079c AK |
752 | #define reserved (1UL << PG_reserved) |
753 | ||
754 | static struct page_state { | |
755 | unsigned long mask; | |
756 | unsigned long res; | |
757 | char *msg; | |
758 | int (*action)(struct page *p, unsigned long pfn); | |
759 | } error_states[] = { | |
d95ea51e | 760 | { reserved, reserved, "reserved kernel", me_kernel }, |
95d01fc6 WF |
761 | /* |
762 | * free pages are specially detected outside this table: | |
763 | * PG_buddy pages only make a small fraction of all free pages. | |
764 | */ | |
6a46079c AK |
765 | |
766 | /* | |
767 | * Could in theory check if slab page is free or if we can drop | |
768 | * currently unused objects without touching them. But just | |
769 | * treat it as standard kernel for now. | |
770 | */ | |
771 | { slab, slab, "kernel slab", me_kernel }, | |
772 | ||
773 | #ifdef CONFIG_PAGEFLAGS_EXTENDED | |
774 | { head, head, "huge", me_huge_page }, | |
775 | { tail, tail, "huge", me_huge_page }, | |
776 | #else | |
777 | { compound, compound, "huge", me_huge_page }, | |
778 | #endif | |
779 | ||
780 | { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty }, | |
781 | { sc|dirty, sc, "swapcache", me_swapcache_clean }, | |
782 | ||
783 | { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty}, | |
784 | { unevict, unevict, "unevictable LRU", me_pagecache_clean}, | |
785 | ||
6a46079c AK |
786 | { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty }, |
787 | { mlock, mlock, "mlocked LRU", me_pagecache_clean }, | |
6a46079c AK |
788 | |
789 | { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty }, | |
790 | { lru|dirty, lru, "clean LRU", me_pagecache_clean }, | |
6a46079c AK |
791 | |
792 | /* | |
793 | * Catchall entry: must be at end. | |
794 | */ | |
795 | { 0, 0, "unknown page state", me_unknown }, | |
796 | }; | |
797 | ||
2326c467 AK |
798 | #undef dirty |
799 | #undef sc | |
800 | #undef unevict | |
801 | #undef mlock | |
802 | #undef writeback | |
803 | #undef lru | |
804 | #undef swapbacked | |
805 | #undef head | |
806 | #undef tail | |
807 | #undef compound | |
808 | #undef slab | |
809 | #undef reserved | |
810 | ||
6a46079c AK |
811 | static void action_result(unsigned long pfn, char *msg, int result) |
812 | { | |
a7560fc8 | 813 | struct page *page = pfn_to_page(pfn); |
6a46079c AK |
814 | |
815 | printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n", | |
816 | pfn, | |
a7560fc8 | 817 | PageDirty(page) ? "dirty " : "", |
6a46079c AK |
818 | msg, action_name[result]); |
819 | } | |
820 | ||
821 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 822 | unsigned long pfn) |
6a46079c AK |
823 | { |
824 | int result; | |
7456b040 | 825 | int count; |
6a46079c AK |
826 | |
827 | result = ps->action(p, pfn); | |
828 | action_result(pfn, ps->msg, result); | |
7456b040 | 829 | |
bd1ce5f9 | 830 | count = page_count(p) - 1; |
138ce286 WF |
831 | if (ps->action == me_swapcache_dirty && result == DELAYED) |
832 | count--; | |
833 | if (count != 0) { | |
6a46079c AK |
834 | printk(KERN_ERR |
835 | "MCE %#lx: %s page still referenced by %d users\n", | |
7456b040 | 836 | pfn, ps->msg, count); |
138ce286 WF |
837 | result = FAILED; |
838 | } | |
6a46079c AK |
839 | |
840 | /* Could do more checks here if page looks ok */ | |
841 | /* | |
842 | * Could adjust zone counters here to correct for the missing page. | |
843 | */ | |
844 | ||
138ce286 | 845 | return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
846 | } |
847 | ||
6a46079c AK |
848 | /* |
849 | * Do all that is necessary to remove user space mappings. Unmap | |
850 | * the pages and send SIGBUS to the processes if the data was dirty. | |
851 | */ | |
1668bfd5 | 852 | static int hwpoison_user_mappings(struct page *p, unsigned long pfn, |
6a46079c AK |
853 | int trapno) |
854 | { | |
855 | enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
856 | struct address_space *mapping; | |
857 | LIST_HEAD(tokill); | |
858 | int ret; | |
6a46079c | 859 | int kill = 1; |
7af446a8 | 860 | struct page *hpage = compound_head(p); |
a6d30ddd | 861 | struct page *ppage; |
6a46079c | 862 | |
1668bfd5 WF |
863 | if (PageReserved(p) || PageSlab(p)) |
864 | return SWAP_SUCCESS; | |
6a46079c | 865 | |
6a46079c AK |
866 | /* |
867 | * This check implies we don't kill processes if their pages | |
868 | * are in the swap cache early. Those are always late kills. | |
869 | */ | |
7af446a8 | 870 | if (!page_mapped(hpage)) |
1668bfd5 WF |
871 | return SWAP_SUCCESS; |
872 | ||
7af446a8 | 873 | if (PageKsm(p)) |
1668bfd5 | 874 | return SWAP_FAIL; |
6a46079c AK |
875 | |
876 | if (PageSwapCache(p)) { | |
877 | printk(KERN_ERR | |
878 | "MCE %#lx: keeping poisoned page in swap cache\n", pfn); | |
879 | ttu |= TTU_IGNORE_HWPOISON; | |
880 | } | |
881 | ||
882 | /* | |
883 | * Propagate the dirty bit from PTEs to struct page first, because we | |
884 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
885 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
886 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 887 | */ |
7af446a8 NH |
888 | mapping = page_mapping(hpage); |
889 | if (!PageDirty(hpage) && mapping && | |
890 | mapping_cap_writeback_dirty(mapping)) { | |
891 | if (page_mkclean(hpage)) { | |
892 | SetPageDirty(hpage); | |
6a46079c AK |
893 | } else { |
894 | kill = 0; | |
895 | ttu |= TTU_IGNORE_HWPOISON; | |
896 | printk(KERN_INFO | |
897 | "MCE %#lx: corrupted page was clean: dropped without side effects\n", | |
898 | pfn); | |
899 | } | |
900 | } | |
901 | ||
a6d30ddd JD |
902 | /* |
903 | * ppage: poisoned page | |
904 | * if p is regular page(4k page) | |
905 | * ppage == real poisoned page; | |
906 | * else p is hugetlb or THP, ppage == head page. | |
907 | */ | |
908 | ppage = hpage; | |
909 | ||
efeda7a4 JD |
910 | if (PageTransHuge(hpage)) { |
911 | /* | |
912 | * Verify that this isn't a hugetlbfs head page, the check for | |
913 | * PageAnon is just for avoid tripping a split_huge_page | |
914 | * internal debug check, as split_huge_page refuses to deal with | |
915 | * anything that isn't an anon page. PageAnon can't go away fro | |
916 | * under us because we hold a refcount on the hpage, without a | |
917 | * refcount on the hpage. split_huge_page can't be safely called | |
918 | * in the first place, having a refcount on the tail isn't | |
919 | * enough * to be safe. | |
920 | */ | |
921 | if (!PageHuge(hpage) && PageAnon(hpage)) { | |
922 | if (unlikely(split_huge_page(hpage))) { | |
923 | /* | |
924 | * FIXME: if splitting THP is failed, it is | |
925 | * better to stop the following operation rather | |
926 | * than causing panic by unmapping. System might | |
927 | * survive if the page is freed later. | |
928 | */ | |
929 | printk(KERN_INFO | |
930 | "MCE %#lx: failed to split THP\n", pfn); | |
931 | ||
932 | BUG_ON(!PageHWPoison(p)); | |
933 | return SWAP_FAIL; | |
934 | } | |
a6d30ddd JD |
935 | /* THP is split, so ppage should be the real poisoned page. */ |
936 | ppage = p; | |
efeda7a4 JD |
937 | } |
938 | } | |
939 | ||
6a46079c AK |
940 | /* |
941 | * First collect all the processes that have the page | |
942 | * mapped in dirty form. This has to be done before try_to_unmap, | |
943 | * because ttu takes the rmap data structures down. | |
944 | * | |
945 | * Error handling: We ignore errors here because | |
946 | * there's nothing that can be done. | |
947 | */ | |
948 | if (kill) | |
a6d30ddd | 949 | collect_procs(ppage, &tokill); |
6a46079c | 950 | |
a6d30ddd | 951 | if (hpage != ppage) |
7eaceacc | 952 | lock_page(ppage); |
a6d30ddd JD |
953 | |
954 | ret = try_to_unmap(ppage, ttu); | |
6a46079c AK |
955 | if (ret != SWAP_SUCCESS) |
956 | printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n", | |
a6d30ddd JD |
957 | pfn, page_mapcount(ppage)); |
958 | ||
959 | if (hpage != ppage) | |
960 | unlock_page(ppage); | |
6a46079c AK |
961 | |
962 | /* | |
963 | * Now that the dirty bit has been propagated to the | |
964 | * struct page and all unmaps done we can decide if | |
965 | * killing is needed or not. Only kill when the page | |
966 | * was dirty, otherwise the tokill list is merely | |
967 | * freed. When there was a problem unmapping earlier | |
968 | * use a more force-full uncatchable kill to prevent | |
969 | * any accesses to the poisoned memory. | |
970 | */ | |
a6d30ddd | 971 | kill_procs_ao(&tokill, !!PageDirty(ppage), trapno, |
0d9ee6a2 | 972 | ret != SWAP_SUCCESS, p, pfn); |
1668bfd5 WF |
973 | |
974 | return ret; | |
6a46079c AK |
975 | } |
976 | ||
7013febc NH |
977 | static void set_page_hwpoison_huge_page(struct page *hpage) |
978 | { | |
979 | int i; | |
37c2ac78 | 980 | int nr_pages = 1 << compound_trans_order(hpage); |
7013febc NH |
981 | for (i = 0; i < nr_pages; i++) |
982 | SetPageHWPoison(hpage + i); | |
983 | } | |
984 | ||
985 | static void clear_page_hwpoison_huge_page(struct page *hpage) | |
986 | { | |
987 | int i; | |
37c2ac78 | 988 | int nr_pages = 1 << compound_trans_order(hpage); |
7013febc NH |
989 | for (i = 0; i < nr_pages; i++) |
990 | ClearPageHWPoison(hpage + i); | |
991 | } | |
992 | ||
82ba011b | 993 | int __memory_failure(unsigned long pfn, int trapno, int flags) |
6a46079c AK |
994 | { |
995 | struct page_state *ps; | |
996 | struct page *p; | |
7af446a8 | 997 | struct page *hpage; |
6a46079c | 998 | int res; |
c9fbdd5f | 999 | unsigned int nr_pages; |
6a46079c AK |
1000 | |
1001 | if (!sysctl_memory_failure_recovery) | |
1002 | panic("Memory failure from trap %d on page %lx", trapno, pfn); | |
1003 | ||
1004 | if (!pfn_valid(pfn)) { | |
a7560fc8 WF |
1005 | printk(KERN_ERR |
1006 | "MCE %#lx: memory outside kernel control\n", | |
1007 | pfn); | |
1008 | return -ENXIO; | |
6a46079c AK |
1009 | } |
1010 | ||
1011 | p = pfn_to_page(pfn); | |
7af446a8 | 1012 | hpage = compound_head(p); |
6a46079c | 1013 | if (TestSetPageHWPoison(p)) { |
d95ea51e | 1014 | printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn); |
6a46079c AK |
1015 | return 0; |
1016 | } | |
1017 | ||
37c2ac78 | 1018 | nr_pages = 1 << compound_trans_order(hpage); |
c9fbdd5f | 1019 | atomic_long_add(nr_pages, &mce_bad_pages); |
6a46079c AK |
1020 | |
1021 | /* | |
1022 | * We need/can do nothing about count=0 pages. | |
1023 | * 1) it's a free page, and therefore in safe hand: | |
1024 | * prep_new_page() will be the gate keeper. | |
8c6c2ecb NH |
1025 | * 2) it's a free hugepage, which is also safe: |
1026 | * an affected hugepage will be dequeued from hugepage freelist, | |
1027 | * so there's no concern about reusing it ever after. | |
1028 | * 3) it's part of a non-compound high order page. | |
6a46079c AK |
1029 | * Implies some kernel user: cannot stop them from |
1030 | * R/W the page; let's pray that the page has been | |
1031 | * used and will be freed some time later. | |
1032 | * In fact it's dangerous to directly bump up page count from 0, | |
1033 | * that may make page_freeze_refs()/page_unfreeze_refs() mismatch. | |
1034 | */ | |
82ba011b | 1035 | if (!(flags & MF_COUNT_INCREASED) && |
7af446a8 | 1036 | !get_page_unless_zero(hpage)) { |
8d22ba1b WF |
1037 | if (is_free_buddy_page(p)) { |
1038 | action_result(pfn, "free buddy", DELAYED); | |
1039 | return 0; | |
8c6c2ecb NH |
1040 | } else if (PageHuge(hpage)) { |
1041 | /* | |
1042 | * Check "just unpoisoned", "filter hit", and | |
1043 | * "race with other subpage." | |
1044 | */ | |
7eaceacc | 1045 | lock_page(hpage); |
8c6c2ecb NH |
1046 | if (!PageHWPoison(hpage) |
1047 | || (hwpoison_filter(p) && TestClearPageHWPoison(p)) | |
1048 | || (p != hpage && TestSetPageHWPoison(hpage))) { | |
1049 | atomic_long_sub(nr_pages, &mce_bad_pages); | |
1050 | return 0; | |
1051 | } | |
1052 | set_page_hwpoison_huge_page(hpage); | |
1053 | res = dequeue_hwpoisoned_huge_page(hpage); | |
1054 | action_result(pfn, "free huge", | |
1055 | res ? IGNORED : DELAYED); | |
1056 | unlock_page(hpage); | |
1057 | return res; | |
8d22ba1b WF |
1058 | } else { |
1059 | action_result(pfn, "high order kernel", IGNORED); | |
1060 | return -EBUSY; | |
1061 | } | |
6a46079c AK |
1062 | } |
1063 | ||
e43c3afb WF |
1064 | /* |
1065 | * We ignore non-LRU pages for good reasons. | |
1066 | * - PG_locked is only well defined for LRU pages and a few others | |
1067 | * - to avoid races with __set_page_locked() | |
1068 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) | |
1069 | * The check (unnecessarily) ignores LRU pages being isolated and | |
1070 | * walked by the page reclaim code, however that's not a big loss. | |
1071 | */ | |
af241a08 JD |
1072 | if (!PageHuge(p) && !PageTransCompound(p)) { |
1073 | if (!PageLRU(p)) | |
1074 | shake_page(p, 0); | |
1075 | if (!PageLRU(p)) { | |
1076 | /* | |
1077 | * shake_page could have turned it free. | |
1078 | */ | |
1079 | if (is_free_buddy_page(p)) { | |
1080 | action_result(pfn, "free buddy, 2nd try", | |
1081 | DELAYED); | |
1082 | return 0; | |
1083 | } | |
1084 | action_result(pfn, "non LRU", IGNORED); | |
1085 | put_page(p); | |
1086 | return -EBUSY; | |
0474a60e | 1087 | } |
e43c3afb | 1088 | } |
e43c3afb | 1089 | |
6a46079c AK |
1090 | /* |
1091 | * Lock the page and wait for writeback to finish. | |
1092 | * It's very difficult to mess with pages currently under IO | |
1093 | * and in many cases impossible, so we just avoid it here. | |
1094 | */ | |
7eaceacc | 1095 | lock_page(hpage); |
847ce401 WF |
1096 | |
1097 | /* | |
1098 | * unpoison always clear PG_hwpoison inside page lock | |
1099 | */ | |
1100 | if (!PageHWPoison(p)) { | |
d95ea51e | 1101 | printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn); |
847ce401 WF |
1102 | res = 0; |
1103 | goto out; | |
1104 | } | |
7c116f2b WF |
1105 | if (hwpoison_filter(p)) { |
1106 | if (TestClearPageHWPoison(p)) | |
c9fbdd5f | 1107 | atomic_long_sub(nr_pages, &mce_bad_pages); |
7af446a8 NH |
1108 | unlock_page(hpage); |
1109 | put_page(hpage); | |
7c116f2b WF |
1110 | return 0; |
1111 | } | |
847ce401 | 1112 | |
7013febc NH |
1113 | /* |
1114 | * For error on the tail page, we should set PG_hwpoison | |
1115 | * on the head page to show that the hugepage is hwpoisoned | |
1116 | */ | |
a6d30ddd | 1117 | if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) { |
7013febc NH |
1118 | action_result(pfn, "hugepage already hardware poisoned", |
1119 | IGNORED); | |
1120 | unlock_page(hpage); | |
1121 | put_page(hpage); | |
1122 | return 0; | |
1123 | } | |
1124 | /* | |
1125 | * Set PG_hwpoison on all pages in an error hugepage, | |
1126 | * because containment is done in hugepage unit for now. | |
1127 | * Since we have done TestSetPageHWPoison() for the head page with | |
1128 | * page lock held, we can safely set PG_hwpoison bits on tail pages. | |
1129 | */ | |
1130 | if (PageHuge(p)) | |
1131 | set_page_hwpoison_huge_page(hpage); | |
1132 | ||
6a46079c AK |
1133 | wait_on_page_writeback(p); |
1134 | ||
1135 | /* | |
1136 | * Now take care of user space mappings. | |
e64a782f | 1137 | * Abort on fail: __delete_from_page_cache() assumes unmapped page. |
6a46079c | 1138 | */ |
1668bfd5 WF |
1139 | if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) { |
1140 | printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn); | |
1141 | res = -EBUSY; | |
1142 | goto out; | |
1143 | } | |
6a46079c AK |
1144 | |
1145 | /* | |
1146 | * Torn down by someone else? | |
1147 | */ | |
dc2a1cbf | 1148 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
6a46079c | 1149 | action_result(pfn, "already truncated LRU", IGNORED); |
d95ea51e | 1150 | res = -EBUSY; |
6a46079c AK |
1151 | goto out; |
1152 | } | |
1153 | ||
1154 | res = -EBUSY; | |
1155 | for (ps = error_states;; ps++) { | |
dc2a1cbf | 1156 | if ((p->flags & ps->mask) == ps->res) { |
bd1ce5f9 | 1157 | res = page_action(ps, p, pfn); |
6a46079c AK |
1158 | break; |
1159 | } | |
1160 | } | |
1161 | out: | |
7af446a8 | 1162 | unlock_page(hpage); |
6a46079c AK |
1163 | return res; |
1164 | } | |
1165 | EXPORT_SYMBOL_GPL(__memory_failure); | |
1166 | ||
1167 | /** | |
1168 | * memory_failure - Handle memory failure of a page. | |
1169 | * @pfn: Page Number of the corrupted page | |
1170 | * @trapno: Trap number reported in the signal to user space. | |
1171 | * | |
1172 | * This function is called by the low level machine check code | |
1173 | * of an architecture when it detects hardware memory corruption | |
1174 | * of a page. It tries its best to recover, which includes | |
1175 | * dropping pages, killing processes etc. | |
1176 | * | |
1177 | * The function is primarily of use for corruptions that | |
1178 | * happen outside the current execution context (e.g. when | |
1179 | * detected by a background scrubber) | |
1180 | * | |
1181 | * Must run in process context (e.g. a work queue) with interrupts | |
1182 | * enabled and no spinlocks hold. | |
1183 | */ | |
1184 | void memory_failure(unsigned long pfn, int trapno) | |
1185 | { | |
1186 | __memory_failure(pfn, trapno, 0); | |
1187 | } | |
847ce401 WF |
1188 | |
1189 | /** | |
1190 | * unpoison_memory - Unpoison a previously poisoned page | |
1191 | * @pfn: Page number of the to be unpoisoned page | |
1192 | * | |
1193 | * Software-unpoison a page that has been poisoned by | |
1194 | * memory_failure() earlier. | |
1195 | * | |
1196 | * This is only done on the software-level, so it only works | |
1197 | * for linux injected failures, not real hardware failures | |
1198 | * | |
1199 | * Returns 0 for success, otherwise -errno. | |
1200 | */ | |
1201 | int unpoison_memory(unsigned long pfn) | |
1202 | { | |
1203 | struct page *page; | |
1204 | struct page *p; | |
1205 | int freeit = 0; | |
c9fbdd5f | 1206 | unsigned int nr_pages; |
847ce401 WF |
1207 | |
1208 | if (!pfn_valid(pfn)) | |
1209 | return -ENXIO; | |
1210 | ||
1211 | p = pfn_to_page(pfn); | |
1212 | page = compound_head(p); | |
1213 | ||
1214 | if (!PageHWPoison(p)) { | |
fb46e735 | 1215 | pr_info("MCE: Page was already unpoisoned %#lx\n", pfn); |
847ce401 WF |
1216 | return 0; |
1217 | } | |
1218 | ||
37c2ac78 | 1219 | nr_pages = 1 << compound_trans_order(page); |
c9fbdd5f | 1220 | |
847ce401 | 1221 | if (!get_page_unless_zero(page)) { |
8c6c2ecb NH |
1222 | /* |
1223 | * Since HWPoisoned hugepage should have non-zero refcount, | |
1224 | * race between memory failure and unpoison seems to happen. | |
1225 | * In such case unpoison fails and memory failure runs | |
1226 | * to the end. | |
1227 | */ | |
1228 | if (PageHuge(page)) { | |
1229 | pr_debug("MCE: Memory failure is now running on free hugepage %#lx\n", pfn); | |
1230 | return 0; | |
1231 | } | |
847ce401 | 1232 | if (TestClearPageHWPoison(p)) |
c9fbdd5f | 1233 | atomic_long_sub(nr_pages, &mce_bad_pages); |
fb46e735 | 1234 | pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn); |
847ce401 WF |
1235 | return 0; |
1236 | } | |
1237 | ||
7eaceacc | 1238 | lock_page(page); |
847ce401 WF |
1239 | /* |
1240 | * This test is racy because PG_hwpoison is set outside of page lock. | |
1241 | * That's acceptable because that won't trigger kernel panic. Instead, | |
1242 | * the PG_hwpoison page will be caught and isolated on the entrance to | |
1243 | * the free buddy page pool. | |
1244 | */ | |
c9fbdd5f | 1245 | if (TestClearPageHWPoison(page)) { |
fb46e735 | 1246 | pr_info("MCE: Software-unpoisoned page %#lx\n", pfn); |
c9fbdd5f | 1247 | atomic_long_sub(nr_pages, &mce_bad_pages); |
847ce401 | 1248 | freeit = 1; |
6a90181c NH |
1249 | if (PageHuge(page)) |
1250 | clear_page_hwpoison_huge_page(page); | |
847ce401 WF |
1251 | } |
1252 | unlock_page(page); | |
1253 | ||
1254 | put_page(page); | |
1255 | if (freeit) | |
1256 | put_page(page); | |
1257 | ||
1258 | return 0; | |
1259 | } | |
1260 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 AK |
1261 | |
1262 | static struct page *new_page(struct page *p, unsigned long private, int **x) | |
1263 | { | |
12686d15 | 1264 | int nid = page_to_nid(p); |
d950b958 NH |
1265 | if (PageHuge(p)) |
1266 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1267 | nid); | |
1268 | else | |
1269 | return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0); | |
facb6011 AK |
1270 | } |
1271 | ||
1272 | /* | |
1273 | * Safely get reference count of an arbitrary page. | |
1274 | * Returns 0 for a free page, -EIO for a zero refcount page | |
1275 | * that is not free, and 1 for any other page type. | |
1276 | * For 1 the page is returned with increased page count, otherwise not. | |
1277 | */ | |
1278 | static int get_any_page(struct page *p, unsigned long pfn, int flags) | |
1279 | { | |
1280 | int ret; | |
1281 | ||
1282 | if (flags & MF_COUNT_INCREASED) | |
1283 | return 1; | |
1284 | ||
1285 | /* | |
20d6c96b | 1286 | * The lock_memory_hotplug prevents a race with memory hotplug. |
facb6011 AK |
1287 | * This is a big hammer, a better would be nicer. |
1288 | */ | |
20d6c96b | 1289 | lock_memory_hotplug(); |
facb6011 AK |
1290 | |
1291 | /* | |
1292 | * Isolate the page, so that it doesn't get reallocated if it | |
1293 | * was free. | |
1294 | */ | |
1295 | set_migratetype_isolate(p); | |
d950b958 NH |
1296 | /* |
1297 | * When the target page is a free hugepage, just remove it | |
1298 | * from free hugepage list. | |
1299 | */ | |
facb6011 | 1300 | if (!get_page_unless_zero(compound_head(p))) { |
d950b958 | 1301 | if (PageHuge(p)) { |
46e387bb | 1302 | pr_info("get_any_page: %#lx free huge page\n", pfn); |
d950b958 NH |
1303 | ret = dequeue_hwpoisoned_huge_page(compound_head(p)); |
1304 | } else if (is_free_buddy_page(p)) { | |
fb46e735 | 1305 | pr_info("get_any_page: %#lx free buddy page\n", pfn); |
facb6011 AK |
1306 | /* Set hwpoison bit while page is still isolated */ |
1307 | SetPageHWPoison(p); | |
1308 | ret = 0; | |
1309 | } else { | |
fb46e735 | 1310 | pr_info("get_any_page: %#lx: unknown zero refcount page type %lx\n", |
facb6011 AK |
1311 | pfn, p->flags); |
1312 | ret = -EIO; | |
1313 | } | |
1314 | } else { | |
1315 | /* Not a free page */ | |
1316 | ret = 1; | |
1317 | } | |
1318 | unset_migratetype_isolate(p); | |
20d6c96b | 1319 | unlock_memory_hotplug(); |
facb6011 AK |
1320 | return ret; |
1321 | } | |
1322 | ||
d950b958 NH |
1323 | static int soft_offline_huge_page(struct page *page, int flags) |
1324 | { | |
1325 | int ret; | |
1326 | unsigned long pfn = page_to_pfn(page); | |
1327 | struct page *hpage = compound_head(page); | |
1328 | LIST_HEAD(pagelist); | |
1329 | ||
1330 | ret = get_any_page(page, pfn, flags); | |
1331 | if (ret < 0) | |
1332 | return ret; | |
1333 | if (ret == 0) | |
1334 | goto done; | |
1335 | ||
1336 | if (PageHWPoison(hpage)) { | |
1337 | put_page(hpage); | |
1338 | pr_debug("soft offline: %#lx hugepage already poisoned\n", pfn); | |
1339 | return -EBUSY; | |
1340 | } | |
1341 | ||
1342 | /* Keep page count to indicate a given hugepage is isolated. */ | |
1343 | ||
1344 | list_add(&hpage->lru, &pagelist); | |
77f1fe6b MG |
1345 | ret = migrate_huge_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0, |
1346 | true); | |
d950b958 | 1347 | if (ret) { |
48db54ee MK |
1348 | struct page *page1, *page2; |
1349 | list_for_each_entry_safe(page1, page2, &pagelist, lru) | |
1350 | put_page(page1); | |
1351 | ||
d950b958 NH |
1352 | pr_debug("soft offline: %#lx: migration failed %d, type %lx\n", |
1353 | pfn, ret, page->flags); | |
1354 | if (ret > 0) | |
1355 | ret = -EIO; | |
1356 | return ret; | |
1357 | } | |
1358 | done: | |
1359 | if (!PageHWPoison(hpage)) | |
37c2ac78 | 1360 | atomic_long_add(1 << compound_trans_order(hpage), &mce_bad_pages); |
d950b958 NH |
1361 | set_page_hwpoison_huge_page(hpage); |
1362 | dequeue_hwpoisoned_huge_page(hpage); | |
1363 | /* keep elevated page count for bad page */ | |
1364 | return ret; | |
1365 | } | |
1366 | ||
facb6011 AK |
1367 | /** |
1368 | * soft_offline_page - Soft offline a page. | |
1369 | * @page: page to offline | |
1370 | * @flags: flags. Same as memory_failure(). | |
1371 | * | |
1372 | * Returns 0 on success, otherwise negated errno. | |
1373 | * | |
1374 | * Soft offline a page, by migration or invalidation, | |
1375 | * without killing anything. This is for the case when | |
1376 | * a page is not corrupted yet (so it's still valid to access), | |
1377 | * but has had a number of corrected errors and is better taken | |
1378 | * out. | |
1379 | * | |
1380 | * The actual policy on when to do that is maintained by | |
1381 | * user space. | |
1382 | * | |
1383 | * This should never impact any application or cause data loss, | |
1384 | * however it might take some time. | |
1385 | * | |
1386 | * This is not a 100% solution for all memory, but tries to be | |
1387 | * ``good enough'' for the majority of memory. | |
1388 | */ | |
1389 | int soft_offline_page(struct page *page, int flags) | |
1390 | { | |
1391 | int ret; | |
1392 | unsigned long pfn = page_to_pfn(page); | |
1393 | ||
d950b958 NH |
1394 | if (PageHuge(page)) |
1395 | return soft_offline_huge_page(page, flags); | |
1396 | ||
facb6011 AK |
1397 | ret = get_any_page(page, pfn, flags); |
1398 | if (ret < 0) | |
1399 | return ret; | |
1400 | if (ret == 0) | |
1401 | goto done; | |
1402 | ||
1403 | /* | |
1404 | * Page cache page we can handle? | |
1405 | */ | |
1406 | if (!PageLRU(page)) { | |
1407 | /* | |
1408 | * Try to free it. | |
1409 | */ | |
1410 | put_page(page); | |
1411 | shake_page(page, 1); | |
1412 | ||
1413 | /* | |
1414 | * Did it turn free? | |
1415 | */ | |
1416 | ret = get_any_page(page, pfn, 0); | |
1417 | if (ret < 0) | |
1418 | return ret; | |
1419 | if (ret == 0) | |
1420 | goto done; | |
1421 | } | |
1422 | if (!PageLRU(page)) { | |
fb46e735 | 1423 | pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n", |
facb6011 AK |
1424 | pfn, page->flags); |
1425 | return -EIO; | |
1426 | } | |
1427 | ||
1428 | lock_page(page); | |
1429 | wait_on_page_writeback(page); | |
1430 | ||
1431 | /* | |
1432 | * Synchronized using the page lock with memory_failure() | |
1433 | */ | |
1434 | if (PageHWPoison(page)) { | |
1435 | unlock_page(page); | |
1436 | put_page(page); | |
fb46e735 | 1437 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
facb6011 AK |
1438 | return -EBUSY; |
1439 | } | |
1440 | ||
1441 | /* | |
1442 | * Try to invalidate first. This should work for | |
1443 | * non dirty unmapped page cache pages. | |
1444 | */ | |
1445 | ret = invalidate_inode_page(page); | |
1446 | unlock_page(page); | |
facb6011 | 1447 | /* |
facb6011 AK |
1448 | * RED-PEN would be better to keep it isolated here, but we |
1449 | * would need to fix isolation locking first. | |
1450 | */ | |
facb6011 | 1451 | if (ret == 1) { |
bd486285 | 1452 | put_page(page); |
facb6011 | 1453 | ret = 0; |
fb46e735 | 1454 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
facb6011 AK |
1455 | goto done; |
1456 | } | |
1457 | ||
1458 | /* | |
1459 | * Simple invalidation didn't work. | |
1460 | * Try to migrate to a new page instead. migrate.c | |
1461 | * handles a large number of cases for us. | |
1462 | */ | |
1463 | ret = isolate_lru_page(page); | |
bd486285 KK |
1464 | /* |
1465 | * Drop page reference which is came from get_any_page() | |
1466 | * successful isolate_lru_page() already took another one. | |
1467 | */ | |
1468 | put_page(page); | |
facb6011 AK |
1469 | if (!ret) { |
1470 | LIST_HEAD(pagelist); | |
1471 | ||
1472 | list_add(&page->lru, &pagelist); | |
77f1fe6b MG |
1473 | ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, |
1474 | 0, true); | |
facb6011 | 1475 | if (ret) { |
57fc4a5e | 1476 | putback_lru_pages(&pagelist); |
fb46e735 | 1477 | pr_info("soft offline: %#lx: migration failed %d, type %lx\n", |
facb6011 AK |
1478 | pfn, ret, page->flags); |
1479 | if (ret > 0) | |
1480 | ret = -EIO; | |
1481 | } | |
1482 | } else { | |
fb46e735 | 1483 | pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n", |
facb6011 AK |
1484 | pfn, ret, page_count(page), page->flags); |
1485 | } | |
1486 | if (ret) | |
1487 | return ret; | |
1488 | ||
1489 | done: | |
1490 | atomic_long_add(1, &mce_bad_pages); | |
1491 | SetPageHWPoison(page); | |
1492 | /* keep elevated page count for bad page */ | |
1493 | return ret; | |
1494 | } |