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 | 44 | #include <linux/rmap.h> |
b9e15baf | 45 | #include <linux/export.h> |
6a46079c AK |
46 | #include <linux/pagemap.h> |
47 | #include <linux/swap.h> | |
48 | #include <linux/backing-dev.h> | |
facb6011 AK |
49 | #include <linux/migrate.h> |
50 | #include <linux/page-isolation.h> | |
51 | #include <linux/suspend.h> | |
5a0e3ad6 | 52 | #include <linux/slab.h> |
bf998156 | 53 | #include <linux/swapops.h> |
7af446a8 | 54 | #include <linux/hugetlb.h> |
20d6c96b | 55 | #include <linux/memory_hotplug.h> |
5db8a73a | 56 | #include <linux/mm_inline.h> |
ea8f5fb8 | 57 | #include <linux/kfifo.h> |
6a46079c AK |
58 | #include "internal.h" |
59 | ||
60 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
61 | ||
62 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
63 | ||
64 | atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0); | |
65 | ||
27df5068 AK |
66 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
67 | ||
1bfe5feb | 68 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
69 | u32 hwpoison_filter_dev_major = ~0U; |
70 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
71 | u64 hwpoison_filter_flags_mask; |
72 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 73 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
74 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
75 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
76 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
77 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
78 | |
79 | static int hwpoison_filter_dev(struct page *p) | |
80 | { | |
81 | struct address_space *mapping; | |
82 | dev_t dev; | |
83 | ||
84 | if (hwpoison_filter_dev_major == ~0U && | |
85 | hwpoison_filter_dev_minor == ~0U) | |
86 | return 0; | |
87 | ||
88 | /* | |
1c80b990 | 89 | * page_mapping() does not accept slab pages. |
7c116f2b WF |
90 | */ |
91 | if (PageSlab(p)) | |
92 | return -EINVAL; | |
93 | ||
94 | mapping = page_mapping(p); | |
95 | if (mapping == NULL || mapping->host == NULL) | |
96 | return -EINVAL; | |
97 | ||
98 | dev = mapping->host->i_sb->s_dev; | |
99 | if (hwpoison_filter_dev_major != ~0U && | |
100 | hwpoison_filter_dev_major != MAJOR(dev)) | |
101 | return -EINVAL; | |
102 | if (hwpoison_filter_dev_minor != ~0U && | |
103 | hwpoison_filter_dev_minor != MINOR(dev)) | |
104 | return -EINVAL; | |
105 | ||
106 | return 0; | |
107 | } | |
108 | ||
478c5ffc WF |
109 | static int hwpoison_filter_flags(struct page *p) |
110 | { | |
111 | if (!hwpoison_filter_flags_mask) | |
112 | return 0; | |
113 | ||
114 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
115 | hwpoison_filter_flags_value) | |
116 | return 0; | |
117 | else | |
118 | return -EINVAL; | |
119 | } | |
120 | ||
4fd466eb AK |
121 | /* |
122 | * This allows stress tests to limit test scope to a collection of tasks | |
123 | * by putting them under some memcg. This prevents killing unrelated/important | |
124 | * processes such as /sbin/init. Note that the target task may share clean | |
125 | * pages with init (eg. libc text), which is harmless. If the target task | |
126 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
127 | * is also included in the memcg. At last, due to race conditions this filter | |
128 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
129 | * a freed page. | |
130 | */ | |
c255a458 | 131 | #ifdef CONFIG_MEMCG_SWAP |
4fd466eb AK |
132 | u64 hwpoison_filter_memcg; |
133 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
134 | static int hwpoison_filter_task(struct page *p) | |
135 | { | |
136 | struct mem_cgroup *mem; | |
137 | struct cgroup_subsys_state *css; | |
138 | unsigned long ino; | |
139 | ||
140 | if (!hwpoison_filter_memcg) | |
141 | return 0; | |
142 | ||
143 | mem = try_get_mem_cgroup_from_page(p); | |
144 | if (!mem) | |
145 | return -EINVAL; | |
146 | ||
147 | css = mem_cgroup_css(mem); | |
148 | /* root_mem_cgroup has NULL dentries */ | |
149 | if (!css->cgroup->dentry) | |
150 | return -EINVAL; | |
151 | ||
152 | ino = css->cgroup->dentry->d_inode->i_ino; | |
153 | css_put(css); | |
154 | ||
155 | if (ino != hwpoison_filter_memcg) | |
156 | return -EINVAL; | |
157 | ||
158 | return 0; | |
159 | } | |
160 | #else | |
161 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
162 | #endif | |
163 | ||
7c116f2b WF |
164 | int hwpoison_filter(struct page *p) |
165 | { | |
1bfe5feb HL |
166 | if (!hwpoison_filter_enable) |
167 | return 0; | |
168 | ||
7c116f2b WF |
169 | if (hwpoison_filter_dev(p)) |
170 | return -EINVAL; | |
171 | ||
478c5ffc WF |
172 | if (hwpoison_filter_flags(p)) |
173 | return -EINVAL; | |
174 | ||
4fd466eb AK |
175 | if (hwpoison_filter_task(p)) |
176 | return -EINVAL; | |
177 | ||
7c116f2b WF |
178 | return 0; |
179 | } | |
27df5068 AK |
180 | #else |
181 | int hwpoison_filter(struct page *p) | |
182 | { | |
183 | return 0; | |
184 | } | |
185 | #endif | |
186 | ||
7c116f2b WF |
187 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
188 | ||
6a46079c | 189 | /* |
7329bbeb TL |
190 | * Send all the processes who have the page mapped a signal. |
191 | * ``action optional'' if they are not immediately affected by the error | |
192 | * ``action required'' if error happened in current execution context | |
6a46079c | 193 | */ |
7329bbeb TL |
194 | static int kill_proc(struct task_struct *t, unsigned long addr, int trapno, |
195 | unsigned long pfn, struct page *page, int flags) | |
6a46079c AK |
196 | { |
197 | struct siginfo si; | |
198 | int ret; | |
199 | ||
200 | printk(KERN_ERR | |
7329bbeb | 201 | "MCE %#lx: Killing %s:%d due to hardware memory corruption\n", |
6a46079c AK |
202 | pfn, t->comm, t->pid); |
203 | si.si_signo = SIGBUS; | |
204 | si.si_errno = 0; | |
6a46079c AK |
205 | si.si_addr = (void *)addr; |
206 | #ifdef __ARCH_SI_TRAPNO | |
207 | si.si_trapno = trapno; | |
208 | #endif | |
37c2ac78 | 209 | si.si_addr_lsb = compound_trans_order(compound_head(page)) + PAGE_SHIFT; |
7329bbeb TL |
210 | |
211 | if ((flags & MF_ACTION_REQUIRED) && t == current) { | |
212 | si.si_code = BUS_MCEERR_AR; | |
213 | ret = force_sig_info(SIGBUS, &si, t); | |
214 | } else { | |
215 | /* | |
216 | * Don't use force here, it's convenient if the signal | |
217 | * can be temporarily blocked. | |
218 | * This could cause a loop when the user sets SIGBUS | |
219 | * to SIG_IGN, but hopefully no one will do that? | |
220 | */ | |
221 | si.si_code = BUS_MCEERR_AO; | |
222 | ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */ | |
223 | } | |
6a46079c AK |
224 | if (ret < 0) |
225 | printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n", | |
226 | t->comm, t->pid, ret); | |
227 | return ret; | |
228 | } | |
229 | ||
588f9ce6 AK |
230 | /* |
231 | * When a unknown page type is encountered drain as many buffers as possible | |
232 | * in the hope to turn the page into a LRU or free page, which we can handle. | |
233 | */ | |
facb6011 | 234 | void shake_page(struct page *p, int access) |
588f9ce6 AK |
235 | { |
236 | if (!PageSlab(p)) { | |
237 | lru_add_drain_all(); | |
238 | if (PageLRU(p)) | |
239 | return; | |
240 | drain_all_pages(); | |
241 | if (PageLRU(p) || is_free_buddy_page(p)) | |
242 | return; | |
243 | } | |
facb6011 | 244 | |
588f9ce6 | 245 | /* |
af241a08 JD |
246 | * Only call shrink_slab here (which would also shrink other caches) if |
247 | * access is not potentially fatal. | |
588f9ce6 | 248 | */ |
facb6011 AK |
249 | if (access) { |
250 | int nr; | |
251 | do { | |
a09ed5e0 YH |
252 | struct shrink_control shrink = { |
253 | .gfp_mask = GFP_KERNEL, | |
a09ed5e0 YH |
254 | }; |
255 | ||
1495f230 | 256 | nr = shrink_slab(&shrink, 1000, 1000); |
47f43e7e | 257 | if (page_count(p) == 1) |
facb6011 AK |
258 | break; |
259 | } while (nr > 10); | |
260 | } | |
588f9ce6 AK |
261 | } |
262 | EXPORT_SYMBOL_GPL(shake_page); | |
263 | ||
6a46079c AK |
264 | /* |
265 | * Kill all processes that have a poisoned page mapped and then isolate | |
266 | * the page. | |
267 | * | |
268 | * General strategy: | |
269 | * Find all processes having the page mapped and kill them. | |
270 | * But we keep a page reference around so that the page is not | |
271 | * actually freed yet. | |
272 | * Then stash the page away | |
273 | * | |
274 | * There's no convenient way to get back to mapped processes | |
275 | * from the VMAs. So do a brute-force search over all | |
276 | * running processes. | |
277 | * | |
278 | * Remember that machine checks are not common (or rather | |
279 | * if they are common you have other problems), so this shouldn't | |
280 | * be a performance issue. | |
281 | * | |
282 | * Also there are some races possible while we get from the | |
283 | * error detection to actually handle it. | |
284 | */ | |
285 | ||
286 | struct to_kill { | |
287 | struct list_head nd; | |
288 | struct task_struct *tsk; | |
289 | unsigned long addr; | |
9033ae16 | 290 | char addr_valid; |
6a46079c AK |
291 | }; |
292 | ||
293 | /* | |
294 | * Failure handling: if we can't find or can't kill a process there's | |
295 | * not much we can do. We just print a message and ignore otherwise. | |
296 | */ | |
297 | ||
298 | /* | |
299 | * Schedule a process for later kill. | |
300 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
301 | * TBD would GFP_NOIO be enough? | |
302 | */ | |
303 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
304 | struct vm_area_struct *vma, | |
305 | struct list_head *to_kill, | |
306 | struct to_kill **tkc) | |
307 | { | |
308 | struct to_kill *tk; | |
309 | ||
310 | if (*tkc) { | |
311 | tk = *tkc; | |
312 | *tkc = NULL; | |
313 | } else { | |
314 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); | |
315 | if (!tk) { | |
316 | printk(KERN_ERR | |
317 | "MCE: Out of memory while machine check handling\n"); | |
318 | return; | |
319 | } | |
320 | } | |
321 | tk->addr = page_address_in_vma(p, vma); | |
322 | tk->addr_valid = 1; | |
323 | ||
324 | /* | |
325 | * In theory we don't have to kill when the page was | |
326 | * munmaped. But it could be also a mremap. Since that's | |
327 | * likely very rare kill anyways just out of paranoia, but use | |
328 | * a SIGKILL because the error is not contained anymore. | |
329 | */ | |
330 | if (tk->addr == -EFAULT) { | |
fb46e735 | 331 | pr_info("MCE: Unable to find user space address %lx in %s\n", |
6a46079c AK |
332 | page_to_pfn(p), tsk->comm); |
333 | tk->addr_valid = 0; | |
334 | } | |
335 | get_task_struct(tsk); | |
336 | tk->tsk = tsk; | |
337 | list_add_tail(&tk->nd, to_kill); | |
338 | } | |
339 | ||
340 | /* | |
341 | * Kill the processes that have been collected earlier. | |
342 | * | |
343 | * Only do anything when DOIT is set, otherwise just free the list | |
344 | * (this is used for clean pages which do not need killing) | |
345 | * Also when FAIL is set do a force kill because something went | |
346 | * wrong earlier. | |
347 | */ | |
6751ed65 | 348 | static void kill_procs(struct list_head *to_kill, int forcekill, int trapno, |
7329bbeb TL |
349 | int fail, struct page *page, unsigned long pfn, |
350 | int flags) | |
6a46079c AK |
351 | { |
352 | struct to_kill *tk, *next; | |
353 | ||
354 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
6751ed65 | 355 | if (forcekill) { |
6a46079c | 356 | /* |
af901ca1 | 357 | * In case something went wrong with munmapping |
6a46079c AK |
358 | * make sure the process doesn't catch the |
359 | * signal and then access the memory. Just kill it. | |
6a46079c AK |
360 | */ |
361 | if (fail || tk->addr_valid == 0) { | |
362 | printk(KERN_ERR | |
363 | "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", | |
364 | pfn, tk->tsk->comm, tk->tsk->pid); | |
365 | force_sig(SIGKILL, tk->tsk); | |
366 | } | |
367 | ||
368 | /* | |
369 | * In theory the process could have mapped | |
370 | * something else on the address in-between. We could | |
371 | * check for that, but we need to tell the | |
372 | * process anyways. | |
373 | */ | |
7329bbeb TL |
374 | else if (kill_proc(tk->tsk, tk->addr, trapno, |
375 | pfn, page, flags) < 0) | |
6a46079c AK |
376 | printk(KERN_ERR |
377 | "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n", | |
378 | pfn, tk->tsk->comm, tk->tsk->pid); | |
379 | } | |
380 | put_task_struct(tk->tsk); | |
381 | kfree(tk); | |
382 | } | |
383 | } | |
384 | ||
385 | static int task_early_kill(struct task_struct *tsk) | |
386 | { | |
387 | if (!tsk->mm) | |
388 | return 0; | |
389 | if (tsk->flags & PF_MCE_PROCESS) | |
390 | return !!(tsk->flags & PF_MCE_EARLY); | |
391 | return sysctl_memory_failure_early_kill; | |
392 | } | |
393 | ||
394 | /* | |
395 | * Collect processes when the error hit an anonymous page. | |
396 | */ | |
397 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
398 | struct to_kill **tkc) | |
399 | { | |
400 | struct vm_area_struct *vma; | |
401 | struct task_struct *tsk; | |
402 | struct anon_vma *av; | |
403 | ||
6a46079c AK |
404 | av = page_lock_anon_vma(page); |
405 | if (av == NULL) /* Not actually mapped anymore */ | |
9b679320 PZ |
406 | return; |
407 | ||
408 | read_lock(&tasklist_lock); | |
6a46079c | 409 | for_each_process (tsk) { |
5beb4930 RR |
410 | struct anon_vma_chain *vmac; |
411 | ||
6a46079c AK |
412 | if (!task_early_kill(tsk)) |
413 | continue; | |
5beb4930 RR |
414 | list_for_each_entry(vmac, &av->head, same_anon_vma) { |
415 | vma = vmac->vma; | |
6a46079c AK |
416 | if (!page_mapped_in_vma(page, vma)) |
417 | continue; | |
418 | if (vma->vm_mm == tsk->mm) | |
419 | add_to_kill(tsk, page, vma, to_kill, tkc); | |
420 | } | |
421 | } | |
6a46079c | 422 | read_unlock(&tasklist_lock); |
9b679320 | 423 | page_unlock_anon_vma(av); |
6a46079c AK |
424 | } |
425 | ||
426 | /* | |
427 | * Collect processes when the error hit a file mapped page. | |
428 | */ | |
429 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
430 | struct to_kill **tkc) | |
431 | { | |
432 | struct vm_area_struct *vma; | |
433 | struct task_struct *tsk; | |
6a46079c AK |
434 | struct address_space *mapping = page->mapping; |
435 | ||
3d48ae45 | 436 | mutex_lock(&mapping->i_mmap_mutex); |
9b679320 | 437 | read_lock(&tasklist_lock); |
6a46079c AK |
438 | for_each_process(tsk) { |
439 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
440 | ||
441 | if (!task_early_kill(tsk)) | |
442 | continue; | |
443 | ||
6b2dbba8 | 444 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
6a46079c AK |
445 | pgoff) { |
446 | /* | |
447 | * Send early kill signal to tasks where a vma covers | |
448 | * the page but the corrupted page is not necessarily | |
449 | * mapped it in its pte. | |
450 | * Assume applications who requested early kill want | |
451 | * to be informed of all such data corruptions. | |
452 | */ | |
453 | if (vma->vm_mm == tsk->mm) | |
454 | add_to_kill(tsk, page, vma, to_kill, tkc); | |
455 | } | |
456 | } | |
6a46079c | 457 | read_unlock(&tasklist_lock); |
9b679320 | 458 | mutex_unlock(&mapping->i_mmap_mutex); |
6a46079c AK |
459 | } |
460 | ||
461 | /* | |
462 | * Collect the processes who have the corrupted page mapped to kill. | |
463 | * This is done in two steps for locking reasons. | |
464 | * First preallocate one tokill structure outside the spin locks, | |
465 | * so that we can kill at least one process reasonably reliable. | |
466 | */ | |
467 | static void collect_procs(struct page *page, struct list_head *tokill) | |
468 | { | |
469 | struct to_kill *tk; | |
470 | ||
471 | if (!page->mapping) | |
472 | return; | |
473 | ||
474 | tk = kmalloc(sizeof(struct to_kill), GFP_NOIO); | |
475 | if (!tk) | |
476 | return; | |
477 | if (PageAnon(page)) | |
478 | collect_procs_anon(page, tokill, &tk); | |
479 | else | |
480 | collect_procs_file(page, tokill, &tk); | |
481 | kfree(tk); | |
482 | } | |
483 | ||
484 | /* | |
485 | * Error handlers for various types of pages. | |
486 | */ | |
487 | ||
488 | enum outcome { | |
d95ea51e WF |
489 | IGNORED, /* Error: cannot be handled */ |
490 | FAILED, /* Error: handling failed */ | |
6a46079c | 491 | DELAYED, /* Will be handled later */ |
6a46079c AK |
492 | RECOVERED, /* Successfully recovered */ |
493 | }; | |
494 | ||
495 | static const char *action_name[] = { | |
d95ea51e | 496 | [IGNORED] = "Ignored", |
6a46079c AK |
497 | [FAILED] = "Failed", |
498 | [DELAYED] = "Delayed", | |
6a46079c AK |
499 | [RECOVERED] = "Recovered", |
500 | }; | |
501 | ||
dc2a1cbf WF |
502 | /* |
503 | * XXX: It is possible that a page is isolated from LRU cache, | |
504 | * and then kept in swap cache or failed to remove from page cache. | |
505 | * The page count will stop it from being freed by unpoison. | |
506 | * Stress tests should be aware of this memory leak problem. | |
507 | */ | |
508 | static int delete_from_lru_cache(struct page *p) | |
509 | { | |
510 | if (!isolate_lru_page(p)) { | |
511 | /* | |
512 | * Clear sensible page flags, so that the buddy system won't | |
513 | * complain when the page is unpoison-and-freed. | |
514 | */ | |
515 | ClearPageActive(p); | |
516 | ClearPageUnevictable(p); | |
517 | /* | |
518 | * drop the page count elevated by isolate_lru_page() | |
519 | */ | |
520 | page_cache_release(p); | |
521 | return 0; | |
522 | } | |
523 | return -EIO; | |
524 | } | |
525 | ||
6a46079c AK |
526 | /* |
527 | * Error hit kernel page. | |
528 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
529 | * could be more sophisticated. | |
530 | */ | |
531 | static int me_kernel(struct page *p, unsigned long pfn) | |
6a46079c AK |
532 | { |
533 | return IGNORED; | |
534 | } | |
535 | ||
536 | /* | |
537 | * Page in unknown state. Do nothing. | |
538 | */ | |
539 | static int me_unknown(struct page *p, unsigned long pfn) | |
540 | { | |
541 | printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn); | |
542 | return FAILED; | |
543 | } | |
544 | ||
6a46079c AK |
545 | /* |
546 | * Clean (or cleaned) page cache page. | |
547 | */ | |
548 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
549 | { | |
550 | int err; | |
551 | int ret = FAILED; | |
552 | struct address_space *mapping; | |
553 | ||
dc2a1cbf WF |
554 | delete_from_lru_cache(p); |
555 | ||
6a46079c AK |
556 | /* |
557 | * For anonymous pages we're done the only reference left | |
558 | * should be the one m_f() holds. | |
559 | */ | |
560 | if (PageAnon(p)) | |
561 | return RECOVERED; | |
562 | ||
563 | /* | |
564 | * Now truncate the page in the page cache. This is really | |
565 | * more like a "temporary hole punch" | |
566 | * Don't do this for block devices when someone else | |
567 | * has a reference, because it could be file system metadata | |
568 | * and that's not safe to truncate. | |
569 | */ | |
570 | mapping = page_mapping(p); | |
571 | if (!mapping) { | |
572 | /* | |
573 | * Page has been teared down in the meanwhile | |
574 | */ | |
575 | return FAILED; | |
576 | } | |
577 | ||
578 | /* | |
579 | * Truncation is a bit tricky. Enable it per file system for now. | |
580 | * | |
581 | * Open: to take i_mutex or not for this? Right now we don't. | |
582 | */ | |
583 | if (mapping->a_ops->error_remove_page) { | |
584 | err = mapping->a_ops->error_remove_page(mapping, p); | |
585 | if (err != 0) { | |
586 | printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n", | |
587 | pfn, err); | |
588 | } else if (page_has_private(p) && | |
589 | !try_to_release_page(p, GFP_NOIO)) { | |
fb46e735 | 590 | pr_info("MCE %#lx: failed to release buffers\n", pfn); |
6a46079c AK |
591 | } else { |
592 | ret = RECOVERED; | |
593 | } | |
594 | } else { | |
595 | /* | |
596 | * If the file system doesn't support it just invalidate | |
597 | * This fails on dirty or anything with private pages | |
598 | */ | |
599 | if (invalidate_inode_page(p)) | |
600 | ret = RECOVERED; | |
601 | else | |
602 | printk(KERN_INFO "MCE %#lx: Failed to invalidate\n", | |
603 | pfn); | |
604 | } | |
605 | return ret; | |
606 | } | |
607 | ||
608 | /* | |
609 | * Dirty cache page page | |
610 | * Issues: when the error hit a hole page the error is not properly | |
611 | * propagated. | |
612 | */ | |
613 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
614 | { | |
615 | struct address_space *mapping = page_mapping(p); | |
616 | ||
617 | SetPageError(p); | |
618 | /* TBD: print more information about the file. */ | |
619 | if (mapping) { | |
620 | /* | |
621 | * IO error will be reported by write(), fsync(), etc. | |
622 | * who check the mapping. | |
623 | * This way the application knows that something went | |
624 | * wrong with its dirty file data. | |
625 | * | |
626 | * There's one open issue: | |
627 | * | |
628 | * The EIO will be only reported on the next IO | |
629 | * operation and then cleared through the IO map. | |
630 | * Normally Linux has two mechanisms to pass IO error | |
631 | * first through the AS_EIO flag in the address space | |
632 | * and then through the PageError flag in the page. | |
633 | * Since we drop pages on memory failure handling the | |
634 | * only mechanism open to use is through AS_AIO. | |
635 | * | |
636 | * This has the disadvantage that it gets cleared on | |
637 | * the first operation that returns an error, while | |
638 | * the PageError bit is more sticky and only cleared | |
639 | * when the page is reread or dropped. If an | |
640 | * application assumes it will always get error on | |
641 | * fsync, but does other operations on the fd before | |
25985edc | 642 | * and the page is dropped between then the error |
6a46079c AK |
643 | * will not be properly reported. |
644 | * | |
645 | * This can already happen even without hwpoisoned | |
646 | * pages: first on metadata IO errors (which only | |
647 | * report through AS_EIO) or when the page is dropped | |
648 | * at the wrong time. | |
649 | * | |
650 | * So right now we assume that the application DTRT on | |
651 | * the first EIO, but we're not worse than other parts | |
652 | * of the kernel. | |
653 | */ | |
654 | mapping_set_error(mapping, EIO); | |
655 | } | |
656 | ||
657 | return me_pagecache_clean(p, pfn); | |
658 | } | |
659 | ||
660 | /* | |
661 | * Clean and dirty swap cache. | |
662 | * | |
663 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
664 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
665 | * referenced concurrently by 2 types of PTEs: | |
666 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
667 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
668 | * and then | |
669 | * - clear dirty bit to prevent IO | |
670 | * - remove from LRU | |
671 | * - but keep in the swap cache, so that when we return to it on | |
672 | * a later page fault, we know the application is accessing | |
673 | * corrupted data and shall be killed (we installed simple | |
674 | * interception code in do_swap_page to catch it). | |
675 | * | |
676 | * Clean swap cache pages can be directly isolated. A later page fault will | |
677 | * bring in the known good data from disk. | |
678 | */ | |
679 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
680 | { | |
6a46079c AK |
681 | ClearPageDirty(p); |
682 | /* Trigger EIO in shmem: */ | |
683 | ClearPageUptodate(p); | |
684 | ||
dc2a1cbf WF |
685 | if (!delete_from_lru_cache(p)) |
686 | return DELAYED; | |
687 | else | |
688 | return FAILED; | |
6a46079c AK |
689 | } |
690 | ||
691 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
692 | { | |
6a46079c | 693 | delete_from_swap_cache(p); |
e43c3afb | 694 | |
dc2a1cbf WF |
695 | if (!delete_from_lru_cache(p)) |
696 | return RECOVERED; | |
697 | else | |
698 | return FAILED; | |
6a46079c AK |
699 | } |
700 | ||
701 | /* | |
702 | * Huge pages. Needs work. | |
703 | * Issues: | |
93f70f90 NH |
704 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
705 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c AK |
706 | */ |
707 | static int me_huge_page(struct page *p, unsigned long pfn) | |
708 | { | |
6de2b1aa | 709 | int res = 0; |
93f70f90 NH |
710 | struct page *hpage = compound_head(p); |
711 | /* | |
712 | * We can safely recover from error on free or reserved (i.e. | |
713 | * not in-use) hugepage by dequeuing it from freelist. | |
714 | * To check whether a hugepage is in-use or not, we can't use | |
715 | * page->lru because it can be used in other hugepage operations, | |
716 | * such as __unmap_hugepage_range() and gather_surplus_pages(). | |
717 | * So instead we use page_mapping() and PageAnon(). | |
718 | * We assume that this function is called with page lock held, | |
719 | * so there is no race between isolation and mapping/unmapping. | |
720 | */ | |
721 | if (!(page_mapping(hpage) || PageAnon(hpage))) { | |
6de2b1aa NH |
722 | res = dequeue_hwpoisoned_huge_page(hpage); |
723 | if (!res) | |
724 | return RECOVERED; | |
93f70f90 NH |
725 | } |
726 | return DELAYED; | |
6a46079c AK |
727 | } |
728 | ||
729 | /* | |
730 | * Various page states we can handle. | |
731 | * | |
732 | * A page state is defined by its current page->flags bits. | |
733 | * The table matches them in order and calls the right handler. | |
734 | * | |
735 | * This is quite tricky because we can access page at any time | |
25985edc | 736 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
737 | * |
738 | * This is not complete. More states could be added. | |
739 | * For any missing state don't attempt recovery. | |
740 | */ | |
741 | ||
742 | #define dirty (1UL << PG_dirty) | |
743 | #define sc (1UL << PG_swapcache) | |
744 | #define unevict (1UL << PG_unevictable) | |
745 | #define mlock (1UL << PG_mlocked) | |
746 | #define writeback (1UL << PG_writeback) | |
747 | #define lru (1UL << PG_lru) | |
748 | #define swapbacked (1UL << PG_swapbacked) | |
749 | #define head (1UL << PG_head) | |
750 | #define tail (1UL << PG_tail) | |
751 | #define compound (1UL << PG_compound) | |
752 | #define slab (1UL << PG_slab) | |
6a46079c AK |
753 | #define reserved (1UL << PG_reserved) |
754 | ||
755 | static struct page_state { | |
756 | unsigned long mask; | |
757 | unsigned long res; | |
758 | char *msg; | |
759 | int (*action)(struct page *p, unsigned long pfn); | |
760 | } error_states[] = { | |
d95ea51e | 761 | { reserved, reserved, "reserved kernel", me_kernel }, |
95d01fc6 WF |
762 | /* |
763 | * free pages are specially detected outside this table: | |
764 | * PG_buddy pages only make a small fraction of all free pages. | |
765 | */ | |
6a46079c AK |
766 | |
767 | /* | |
768 | * Could in theory check if slab page is free or if we can drop | |
769 | * currently unused objects without touching them. But just | |
770 | * treat it as standard kernel for now. | |
771 | */ | |
772 | { slab, slab, "kernel slab", me_kernel }, | |
773 | ||
774 | #ifdef CONFIG_PAGEFLAGS_EXTENDED | |
775 | { head, head, "huge", me_huge_page }, | |
776 | { tail, tail, "huge", me_huge_page }, | |
777 | #else | |
778 | { compound, compound, "huge", me_huge_page }, | |
779 | #endif | |
780 | ||
781 | { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty }, | |
782 | { sc|dirty, sc, "swapcache", me_swapcache_clean }, | |
783 | ||
784 | { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty}, | |
785 | { unevict, unevict, "unevictable LRU", me_pagecache_clean}, | |
786 | ||
6a46079c AK |
787 | { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty }, |
788 | { mlock, mlock, "mlocked LRU", me_pagecache_clean }, | |
6a46079c AK |
789 | |
790 | { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty }, | |
791 | { lru|dirty, lru, "clean LRU", me_pagecache_clean }, | |
6a46079c AK |
792 | |
793 | /* | |
794 | * Catchall entry: must be at end. | |
795 | */ | |
796 | { 0, 0, "unknown page state", me_unknown }, | |
797 | }; | |
798 | ||
2326c467 AK |
799 | #undef dirty |
800 | #undef sc | |
801 | #undef unevict | |
802 | #undef mlock | |
803 | #undef writeback | |
804 | #undef lru | |
805 | #undef swapbacked | |
806 | #undef head | |
807 | #undef tail | |
808 | #undef compound | |
809 | #undef slab | |
810 | #undef reserved | |
811 | ||
6a46079c AK |
812 | static void action_result(unsigned long pfn, char *msg, int result) |
813 | { | |
a7560fc8 | 814 | struct page *page = pfn_to_page(pfn); |
6a46079c AK |
815 | |
816 | printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n", | |
817 | pfn, | |
a7560fc8 | 818 | PageDirty(page) ? "dirty " : "", |
6a46079c AK |
819 | msg, action_name[result]); |
820 | } | |
821 | ||
822 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 823 | unsigned long pfn) |
6a46079c AK |
824 | { |
825 | int result; | |
7456b040 | 826 | int count; |
6a46079c AK |
827 | |
828 | result = ps->action(p, pfn); | |
829 | action_result(pfn, ps->msg, result); | |
7456b040 | 830 | |
bd1ce5f9 | 831 | count = page_count(p) - 1; |
138ce286 WF |
832 | if (ps->action == me_swapcache_dirty && result == DELAYED) |
833 | count--; | |
834 | if (count != 0) { | |
6a46079c AK |
835 | printk(KERN_ERR |
836 | "MCE %#lx: %s page still referenced by %d users\n", | |
7456b040 | 837 | pfn, ps->msg, count); |
138ce286 WF |
838 | result = FAILED; |
839 | } | |
6a46079c AK |
840 | |
841 | /* Could do more checks here if page looks ok */ | |
842 | /* | |
843 | * Could adjust zone counters here to correct for the missing page. | |
844 | */ | |
845 | ||
138ce286 | 846 | return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
847 | } |
848 | ||
6a46079c AK |
849 | /* |
850 | * Do all that is necessary to remove user space mappings. Unmap | |
851 | * the pages and send SIGBUS to the processes if the data was dirty. | |
852 | */ | |
1668bfd5 | 853 | static int hwpoison_user_mappings(struct page *p, unsigned long pfn, |
7329bbeb | 854 | int trapno, int flags) |
6a46079c AK |
855 | { |
856 | enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
857 | struct address_space *mapping; | |
858 | LIST_HEAD(tokill); | |
859 | int ret; | |
6751ed65 | 860 | int kill = 1, forcekill; |
7af446a8 | 861 | struct page *hpage = compound_head(p); |
a6d30ddd | 862 | struct page *ppage; |
6a46079c | 863 | |
1668bfd5 WF |
864 | if (PageReserved(p) || PageSlab(p)) |
865 | return SWAP_SUCCESS; | |
6a46079c | 866 | |
6a46079c AK |
867 | /* |
868 | * This check implies we don't kill processes if their pages | |
869 | * are in the swap cache early. Those are always late kills. | |
870 | */ | |
7af446a8 | 871 | if (!page_mapped(hpage)) |
1668bfd5 WF |
872 | return SWAP_SUCCESS; |
873 | ||
7af446a8 | 874 | if (PageKsm(p)) |
1668bfd5 | 875 | return SWAP_FAIL; |
6a46079c AK |
876 | |
877 | if (PageSwapCache(p)) { | |
878 | printk(KERN_ERR | |
879 | "MCE %#lx: keeping poisoned page in swap cache\n", pfn); | |
880 | ttu |= TTU_IGNORE_HWPOISON; | |
881 | } | |
882 | ||
883 | /* | |
884 | * Propagate the dirty bit from PTEs to struct page first, because we | |
885 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
886 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
887 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 888 | */ |
7af446a8 | 889 | mapping = page_mapping(hpage); |
6751ed65 | 890 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && |
7af446a8 NH |
891 | mapping_cap_writeback_dirty(mapping)) { |
892 | if (page_mkclean(hpage)) { | |
893 | SetPageDirty(hpage); | |
6a46079c AK |
894 | } else { |
895 | kill = 0; | |
896 | ttu |= TTU_IGNORE_HWPOISON; | |
897 | printk(KERN_INFO | |
898 | "MCE %#lx: corrupted page was clean: dropped without side effects\n", | |
899 | pfn); | |
900 | } | |
901 | } | |
902 | ||
a6d30ddd JD |
903 | /* |
904 | * ppage: poisoned page | |
905 | * if p is regular page(4k page) | |
906 | * ppage == real poisoned page; | |
907 | * else p is hugetlb or THP, ppage == head page. | |
908 | */ | |
909 | ppage = hpage; | |
910 | ||
efeda7a4 JD |
911 | if (PageTransHuge(hpage)) { |
912 | /* | |
913 | * Verify that this isn't a hugetlbfs head page, the check for | |
914 | * PageAnon is just for avoid tripping a split_huge_page | |
915 | * internal debug check, as split_huge_page refuses to deal with | |
916 | * anything that isn't an anon page. PageAnon can't go away fro | |
917 | * under us because we hold a refcount on the hpage, without a | |
918 | * refcount on the hpage. split_huge_page can't be safely called | |
919 | * in the first place, having a refcount on the tail isn't | |
920 | * enough * to be safe. | |
921 | */ | |
922 | if (!PageHuge(hpage) && PageAnon(hpage)) { | |
923 | if (unlikely(split_huge_page(hpage))) { | |
924 | /* | |
925 | * FIXME: if splitting THP is failed, it is | |
926 | * better to stop the following operation rather | |
927 | * than causing panic by unmapping. System might | |
928 | * survive if the page is freed later. | |
929 | */ | |
930 | printk(KERN_INFO | |
931 | "MCE %#lx: failed to split THP\n", pfn); | |
932 | ||
933 | BUG_ON(!PageHWPoison(p)); | |
934 | return SWAP_FAIL; | |
935 | } | |
a6d30ddd JD |
936 | /* THP is split, so ppage should be the real poisoned page. */ |
937 | ppage = p; | |
efeda7a4 JD |
938 | } |
939 | } | |
940 | ||
6a46079c AK |
941 | /* |
942 | * First collect all the processes that have the page | |
943 | * mapped in dirty form. This has to be done before try_to_unmap, | |
944 | * because ttu takes the rmap data structures down. | |
945 | * | |
946 | * Error handling: We ignore errors here because | |
947 | * there's nothing that can be done. | |
948 | */ | |
949 | if (kill) | |
a6d30ddd | 950 | collect_procs(ppage, &tokill); |
6a46079c | 951 | |
a6d30ddd | 952 | if (hpage != ppage) |
7eaceacc | 953 | lock_page(ppage); |
a6d30ddd JD |
954 | |
955 | ret = try_to_unmap(ppage, ttu); | |
6a46079c AK |
956 | if (ret != SWAP_SUCCESS) |
957 | printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n", | |
a6d30ddd JD |
958 | pfn, page_mapcount(ppage)); |
959 | ||
960 | if (hpage != ppage) | |
961 | unlock_page(ppage); | |
6a46079c AK |
962 | |
963 | /* | |
964 | * Now that the dirty bit has been propagated to the | |
965 | * struct page and all unmaps done we can decide if | |
966 | * killing is needed or not. Only kill when the page | |
6751ed65 TL |
967 | * was dirty or the process is not restartable, |
968 | * otherwise the tokill list is merely | |
6a46079c AK |
969 | * freed. When there was a problem unmapping earlier |
970 | * use a more force-full uncatchable kill to prevent | |
971 | * any accesses to the poisoned memory. | |
972 | */ | |
6751ed65 TL |
973 | forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL); |
974 | kill_procs(&tokill, forcekill, trapno, | |
7329bbeb | 975 | ret != SWAP_SUCCESS, p, pfn, flags); |
1668bfd5 WF |
976 | |
977 | return ret; | |
6a46079c AK |
978 | } |
979 | ||
7013febc NH |
980 | static void set_page_hwpoison_huge_page(struct page *hpage) |
981 | { | |
982 | int i; | |
37c2ac78 | 983 | int nr_pages = 1 << compound_trans_order(hpage); |
7013febc NH |
984 | for (i = 0; i < nr_pages; i++) |
985 | SetPageHWPoison(hpage + i); | |
986 | } | |
987 | ||
988 | static void clear_page_hwpoison_huge_page(struct page *hpage) | |
989 | { | |
990 | int i; | |
37c2ac78 | 991 | int nr_pages = 1 << compound_trans_order(hpage); |
7013febc NH |
992 | for (i = 0; i < nr_pages; i++) |
993 | ClearPageHWPoison(hpage + i); | |
994 | } | |
995 | ||
cd42f4a3 TL |
996 | /** |
997 | * memory_failure - Handle memory failure of a page. | |
998 | * @pfn: Page Number of the corrupted page | |
999 | * @trapno: Trap number reported in the signal to user space. | |
1000 | * @flags: fine tune action taken | |
1001 | * | |
1002 | * This function is called by the low level machine check code | |
1003 | * of an architecture when it detects hardware memory corruption | |
1004 | * of a page. It tries its best to recover, which includes | |
1005 | * dropping pages, killing processes etc. | |
1006 | * | |
1007 | * The function is primarily of use for corruptions that | |
1008 | * happen outside the current execution context (e.g. when | |
1009 | * detected by a background scrubber) | |
1010 | * | |
1011 | * Must run in process context (e.g. a work queue) with interrupts | |
1012 | * enabled and no spinlocks hold. | |
1013 | */ | |
1014 | int memory_failure(unsigned long pfn, int trapno, int flags) | |
6a46079c AK |
1015 | { |
1016 | struct page_state *ps; | |
1017 | struct page *p; | |
7af446a8 | 1018 | struct page *hpage; |
6a46079c | 1019 | int res; |
c9fbdd5f | 1020 | unsigned int nr_pages; |
6a46079c AK |
1021 | |
1022 | if (!sysctl_memory_failure_recovery) | |
1023 | panic("Memory failure from trap %d on page %lx", trapno, pfn); | |
1024 | ||
1025 | if (!pfn_valid(pfn)) { | |
a7560fc8 WF |
1026 | printk(KERN_ERR |
1027 | "MCE %#lx: memory outside kernel control\n", | |
1028 | pfn); | |
1029 | return -ENXIO; | |
6a46079c AK |
1030 | } |
1031 | ||
1032 | p = pfn_to_page(pfn); | |
7af446a8 | 1033 | hpage = compound_head(p); |
6a46079c | 1034 | if (TestSetPageHWPoison(p)) { |
d95ea51e | 1035 | printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn); |
6a46079c AK |
1036 | return 0; |
1037 | } | |
1038 | ||
37c2ac78 | 1039 | nr_pages = 1 << compound_trans_order(hpage); |
c9fbdd5f | 1040 | atomic_long_add(nr_pages, &mce_bad_pages); |
6a46079c AK |
1041 | |
1042 | /* | |
1043 | * We need/can do nothing about count=0 pages. | |
1044 | * 1) it's a free page, and therefore in safe hand: | |
1045 | * prep_new_page() will be the gate keeper. | |
8c6c2ecb NH |
1046 | * 2) it's a free hugepage, which is also safe: |
1047 | * an affected hugepage will be dequeued from hugepage freelist, | |
1048 | * so there's no concern about reusing it ever after. | |
1049 | * 3) it's part of a non-compound high order page. | |
6a46079c AK |
1050 | * Implies some kernel user: cannot stop them from |
1051 | * R/W the page; let's pray that the page has been | |
1052 | * used and will be freed some time later. | |
1053 | * In fact it's dangerous to directly bump up page count from 0, | |
1054 | * that may make page_freeze_refs()/page_unfreeze_refs() mismatch. | |
1055 | */ | |
82ba011b | 1056 | if (!(flags & MF_COUNT_INCREASED) && |
7af446a8 | 1057 | !get_page_unless_zero(hpage)) { |
8d22ba1b WF |
1058 | if (is_free_buddy_page(p)) { |
1059 | action_result(pfn, "free buddy", DELAYED); | |
1060 | return 0; | |
8c6c2ecb NH |
1061 | } else if (PageHuge(hpage)) { |
1062 | /* | |
1063 | * Check "just unpoisoned", "filter hit", and | |
1064 | * "race with other subpage." | |
1065 | */ | |
7eaceacc | 1066 | lock_page(hpage); |
8c6c2ecb NH |
1067 | if (!PageHWPoison(hpage) |
1068 | || (hwpoison_filter(p) && TestClearPageHWPoison(p)) | |
1069 | || (p != hpage && TestSetPageHWPoison(hpage))) { | |
1070 | atomic_long_sub(nr_pages, &mce_bad_pages); | |
1071 | return 0; | |
1072 | } | |
1073 | set_page_hwpoison_huge_page(hpage); | |
1074 | res = dequeue_hwpoisoned_huge_page(hpage); | |
1075 | action_result(pfn, "free huge", | |
1076 | res ? IGNORED : DELAYED); | |
1077 | unlock_page(hpage); | |
1078 | return res; | |
8d22ba1b WF |
1079 | } else { |
1080 | action_result(pfn, "high order kernel", IGNORED); | |
1081 | return -EBUSY; | |
1082 | } | |
6a46079c AK |
1083 | } |
1084 | ||
e43c3afb WF |
1085 | /* |
1086 | * We ignore non-LRU pages for good reasons. | |
1087 | * - PG_locked is only well defined for LRU pages and a few others | |
1088 | * - to avoid races with __set_page_locked() | |
1089 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) | |
1090 | * The check (unnecessarily) ignores LRU pages being isolated and | |
1091 | * walked by the page reclaim code, however that's not a big loss. | |
1092 | */ | |
385de357 | 1093 | if (!PageHuge(p) && !PageTransTail(p)) { |
af241a08 JD |
1094 | if (!PageLRU(p)) |
1095 | shake_page(p, 0); | |
1096 | if (!PageLRU(p)) { | |
1097 | /* | |
1098 | * shake_page could have turned it free. | |
1099 | */ | |
1100 | if (is_free_buddy_page(p)) { | |
1101 | action_result(pfn, "free buddy, 2nd try", | |
1102 | DELAYED); | |
1103 | return 0; | |
1104 | } | |
1105 | action_result(pfn, "non LRU", IGNORED); | |
1106 | put_page(p); | |
1107 | return -EBUSY; | |
0474a60e | 1108 | } |
e43c3afb | 1109 | } |
e43c3afb | 1110 | |
6a46079c AK |
1111 | /* |
1112 | * Lock the page and wait for writeback to finish. | |
1113 | * It's very difficult to mess with pages currently under IO | |
1114 | * and in many cases impossible, so we just avoid it here. | |
1115 | */ | |
7eaceacc | 1116 | lock_page(hpage); |
847ce401 WF |
1117 | |
1118 | /* | |
1119 | * unpoison always clear PG_hwpoison inside page lock | |
1120 | */ | |
1121 | if (!PageHWPoison(p)) { | |
d95ea51e | 1122 | printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn); |
847ce401 WF |
1123 | res = 0; |
1124 | goto out; | |
1125 | } | |
7c116f2b WF |
1126 | if (hwpoison_filter(p)) { |
1127 | if (TestClearPageHWPoison(p)) | |
c9fbdd5f | 1128 | atomic_long_sub(nr_pages, &mce_bad_pages); |
7af446a8 NH |
1129 | unlock_page(hpage); |
1130 | put_page(hpage); | |
7c116f2b WF |
1131 | return 0; |
1132 | } | |
847ce401 | 1133 | |
7013febc NH |
1134 | /* |
1135 | * For error on the tail page, we should set PG_hwpoison | |
1136 | * on the head page to show that the hugepage is hwpoisoned | |
1137 | */ | |
a6d30ddd | 1138 | if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) { |
7013febc NH |
1139 | action_result(pfn, "hugepage already hardware poisoned", |
1140 | IGNORED); | |
1141 | unlock_page(hpage); | |
1142 | put_page(hpage); | |
1143 | return 0; | |
1144 | } | |
1145 | /* | |
1146 | * Set PG_hwpoison on all pages in an error hugepage, | |
1147 | * because containment is done in hugepage unit for now. | |
1148 | * Since we have done TestSetPageHWPoison() for the head page with | |
1149 | * page lock held, we can safely set PG_hwpoison bits on tail pages. | |
1150 | */ | |
1151 | if (PageHuge(p)) | |
1152 | set_page_hwpoison_huge_page(hpage); | |
1153 | ||
6a46079c AK |
1154 | wait_on_page_writeback(p); |
1155 | ||
1156 | /* | |
1157 | * Now take care of user space mappings. | |
e64a782f | 1158 | * Abort on fail: __delete_from_page_cache() assumes unmapped page. |
6a46079c | 1159 | */ |
7329bbeb | 1160 | if (hwpoison_user_mappings(p, pfn, trapno, flags) != SWAP_SUCCESS) { |
1668bfd5 WF |
1161 | printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn); |
1162 | res = -EBUSY; | |
1163 | goto out; | |
1164 | } | |
6a46079c AK |
1165 | |
1166 | /* | |
1167 | * Torn down by someone else? | |
1168 | */ | |
dc2a1cbf | 1169 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
6a46079c | 1170 | action_result(pfn, "already truncated LRU", IGNORED); |
d95ea51e | 1171 | res = -EBUSY; |
6a46079c AK |
1172 | goto out; |
1173 | } | |
1174 | ||
1175 | res = -EBUSY; | |
1176 | for (ps = error_states;; ps++) { | |
dc2a1cbf | 1177 | if ((p->flags & ps->mask) == ps->res) { |
bd1ce5f9 | 1178 | res = page_action(ps, p, pfn); |
6a46079c AK |
1179 | break; |
1180 | } | |
1181 | } | |
1182 | out: | |
7af446a8 | 1183 | unlock_page(hpage); |
6a46079c AK |
1184 | return res; |
1185 | } | |
cd42f4a3 | 1186 | EXPORT_SYMBOL_GPL(memory_failure); |
847ce401 | 1187 | |
ea8f5fb8 HY |
1188 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
1189 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
1190 | ||
1191 | struct memory_failure_entry { | |
1192 | unsigned long pfn; | |
1193 | int trapno; | |
1194 | int flags; | |
1195 | }; | |
1196 | ||
1197 | struct memory_failure_cpu { | |
1198 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
1199 | MEMORY_FAILURE_FIFO_SIZE); | |
1200 | spinlock_t lock; | |
1201 | struct work_struct work; | |
1202 | }; | |
1203 | ||
1204 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
1205 | ||
1206 | /** | |
1207 | * memory_failure_queue - Schedule handling memory failure of a page. | |
1208 | * @pfn: Page Number of the corrupted page | |
1209 | * @trapno: Trap number reported in the signal to user space. | |
1210 | * @flags: Flags for memory failure handling | |
1211 | * | |
1212 | * This function is called by the low level hardware error handler | |
1213 | * when it detects hardware memory corruption of a page. It schedules | |
1214 | * the recovering of error page, including dropping pages, killing | |
1215 | * processes etc. | |
1216 | * | |
1217 | * The function is primarily of use for corruptions that | |
1218 | * happen outside the current execution context (e.g. when | |
1219 | * detected by a background scrubber) | |
1220 | * | |
1221 | * Can run in IRQ context. | |
1222 | */ | |
1223 | void memory_failure_queue(unsigned long pfn, int trapno, int flags) | |
1224 | { | |
1225 | struct memory_failure_cpu *mf_cpu; | |
1226 | unsigned long proc_flags; | |
1227 | struct memory_failure_entry entry = { | |
1228 | .pfn = pfn, | |
1229 | .trapno = trapno, | |
1230 | .flags = flags, | |
1231 | }; | |
1232 | ||
1233 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
1234 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
1235 | if (kfifo_put(&mf_cpu->fifo, &entry)) | |
1236 | schedule_work_on(smp_processor_id(), &mf_cpu->work); | |
1237 | else | |
1238 | pr_err("Memory failure: buffer overflow when queuing memory failure at 0x%#lx\n", | |
1239 | pfn); | |
1240 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1241 | put_cpu_var(memory_failure_cpu); | |
1242 | } | |
1243 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
1244 | ||
1245 | static void memory_failure_work_func(struct work_struct *work) | |
1246 | { | |
1247 | struct memory_failure_cpu *mf_cpu; | |
1248 | struct memory_failure_entry entry = { 0, }; | |
1249 | unsigned long proc_flags; | |
1250 | int gotten; | |
1251 | ||
1252 | mf_cpu = &__get_cpu_var(memory_failure_cpu); | |
1253 | for (;;) { | |
1254 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
1255 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
1256 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1257 | if (!gotten) | |
1258 | break; | |
cd42f4a3 | 1259 | memory_failure(entry.pfn, entry.trapno, entry.flags); |
ea8f5fb8 HY |
1260 | } |
1261 | } | |
1262 | ||
1263 | static int __init memory_failure_init(void) | |
1264 | { | |
1265 | struct memory_failure_cpu *mf_cpu; | |
1266 | int cpu; | |
1267 | ||
1268 | for_each_possible_cpu(cpu) { | |
1269 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1270 | spin_lock_init(&mf_cpu->lock); | |
1271 | INIT_KFIFO(mf_cpu->fifo); | |
1272 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
1273 | } | |
1274 | ||
1275 | return 0; | |
1276 | } | |
1277 | core_initcall(memory_failure_init); | |
1278 | ||
847ce401 WF |
1279 | /** |
1280 | * unpoison_memory - Unpoison a previously poisoned page | |
1281 | * @pfn: Page number of the to be unpoisoned page | |
1282 | * | |
1283 | * Software-unpoison a page that has been poisoned by | |
1284 | * memory_failure() earlier. | |
1285 | * | |
1286 | * This is only done on the software-level, so it only works | |
1287 | * for linux injected failures, not real hardware failures | |
1288 | * | |
1289 | * Returns 0 for success, otherwise -errno. | |
1290 | */ | |
1291 | int unpoison_memory(unsigned long pfn) | |
1292 | { | |
1293 | struct page *page; | |
1294 | struct page *p; | |
1295 | int freeit = 0; | |
c9fbdd5f | 1296 | unsigned int nr_pages; |
847ce401 WF |
1297 | |
1298 | if (!pfn_valid(pfn)) | |
1299 | return -ENXIO; | |
1300 | ||
1301 | p = pfn_to_page(pfn); | |
1302 | page = compound_head(p); | |
1303 | ||
1304 | if (!PageHWPoison(p)) { | |
fb46e735 | 1305 | pr_info("MCE: Page was already unpoisoned %#lx\n", pfn); |
847ce401 WF |
1306 | return 0; |
1307 | } | |
1308 | ||
37c2ac78 | 1309 | nr_pages = 1 << compound_trans_order(page); |
c9fbdd5f | 1310 | |
847ce401 | 1311 | if (!get_page_unless_zero(page)) { |
8c6c2ecb NH |
1312 | /* |
1313 | * Since HWPoisoned hugepage should have non-zero refcount, | |
1314 | * race between memory failure and unpoison seems to happen. | |
1315 | * In such case unpoison fails and memory failure runs | |
1316 | * to the end. | |
1317 | */ | |
1318 | if (PageHuge(page)) { | |
dd73e85f | 1319 | pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn); |
8c6c2ecb NH |
1320 | return 0; |
1321 | } | |
847ce401 | 1322 | if (TestClearPageHWPoison(p)) |
c9fbdd5f | 1323 | atomic_long_sub(nr_pages, &mce_bad_pages); |
fb46e735 | 1324 | pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn); |
847ce401 WF |
1325 | return 0; |
1326 | } | |
1327 | ||
7eaceacc | 1328 | lock_page(page); |
847ce401 WF |
1329 | /* |
1330 | * This test is racy because PG_hwpoison is set outside of page lock. | |
1331 | * That's acceptable because that won't trigger kernel panic. Instead, | |
1332 | * the PG_hwpoison page will be caught and isolated on the entrance to | |
1333 | * the free buddy page pool. | |
1334 | */ | |
c9fbdd5f | 1335 | if (TestClearPageHWPoison(page)) { |
fb46e735 | 1336 | pr_info("MCE: Software-unpoisoned page %#lx\n", pfn); |
c9fbdd5f | 1337 | atomic_long_sub(nr_pages, &mce_bad_pages); |
847ce401 | 1338 | freeit = 1; |
6a90181c NH |
1339 | if (PageHuge(page)) |
1340 | clear_page_hwpoison_huge_page(page); | |
847ce401 WF |
1341 | } |
1342 | unlock_page(page); | |
1343 | ||
1344 | put_page(page); | |
1345 | if (freeit) | |
1346 | put_page(page); | |
1347 | ||
1348 | return 0; | |
1349 | } | |
1350 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 AK |
1351 | |
1352 | static struct page *new_page(struct page *p, unsigned long private, int **x) | |
1353 | { | |
12686d15 | 1354 | int nid = page_to_nid(p); |
d950b958 NH |
1355 | if (PageHuge(p)) |
1356 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1357 | nid); | |
1358 | else | |
1359 | return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0); | |
facb6011 AK |
1360 | } |
1361 | ||
1362 | /* | |
1363 | * Safely get reference count of an arbitrary page. | |
1364 | * Returns 0 for a free page, -EIO for a zero refcount page | |
1365 | * that is not free, and 1 for any other page type. | |
1366 | * For 1 the page is returned with increased page count, otherwise not. | |
1367 | */ | |
1368 | static int get_any_page(struct page *p, unsigned long pfn, int flags) | |
1369 | { | |
1370 | int ret; | |
1371 | ||
1372 | if (flags & MF_COUNT_INCREASED) | |
1373 | return 1; | |
1374 | ||
1375 | /* | |
20d6c96b | 1376 | * The lock_memory_hotplug prevents a race with memory hotplug. |
facb6011 AK |
1377 | * This is a big hammer, a better would be nicer. |
1378 | */ | |
20d6c96b | 1379 | lock_memory_hotplug(); |
facb6011 AK |
1380 | |
1381 | /* | |
1382 | * Isolate the page, so that it doesn't get reallocated if it | |
1383 | * was free. | |
1384 | */ | |
1385 | set_migratetype_isolate(p); | |
d950b958 NH |
1386 | /* |
1387 | * When the target page is a free hugepage, just remove it | |
1388 | * from free hugepage list. | |
1389 | */ | |
facb6011 | 1390 | if (!get_page_unless_zero(compound_head(p))) { |
d950b958 | 1391 | if (PageHuge(p)) { |
71dd0b8a | 1392 | pr_info("%s: %#lx free huge page\n", __func__, pfn); |
d950b958 NH |
1393 | ret = dequeue_hwpoisoned_huge_page(compound_head(p)); |
1394 | } else if (is_free_buddy_page(p)) { | |
71dd0b8a | 1395 | pr_info("%s: %#lx free buddy page\n", __func__, pfn); |
facb6011 AK |
1396 | /* Set hwpoison bit while page is still isolated */ |
1397 | SetPageHWPoison(p); | |
1398 | ret = 0; | |
1399 | } else { | |
71dd0b8a BP |
1400 | pr_info("%s: %#lx: unknown zero refcount page type %lx\n", |
1401 | __func__, pfn, p->flags); | |
facb6011 AK |
1402 | ret = -EIO; |
1403 | } | |
1404 | } else { | |
1405 | /* Not a free page */ | |
1406 | ret = 1; | |
1407 | } | |
0815f3d8 | 1408 | unset_migratetype_isolate(p, MIGRATE_MOVABLE); |
20d6c96b | 1409 | unlock_memory_hotplug(); |
facb6011 AK |
1410 | return ret; |
1411 | } | |
1412 | ||
d950b958 NH |
1413 | static int soft_offline_huge_page(struct page *page, int flags) |
1414 | { | |
1415 | int ret; | |
1416 | unsigned long pfn = page_to_pfn(page); | |
1417 | struct page *hpage = compound_head(page); | |
d950b958 NH |
1418 | |
1419 | ret = get_any_page(page, pfn, flags); | |
1420 | if (ret < 0) | |
1421 | return ret; | |
1422 | if (ret == 0) | |
1423 | goto done; | |
1424 | ||
1425 | if (PageHWPoison(hpage)) { | |
1426 | put_page(hpage); | |
dd73e85f | 1427 | pr_info("soft offline: %#lx hugepage already poisoned\n", pfn); |
d950b958 NH |
1428 | return -EBUSY; |
1429 | } | |
1430 | ||
1431 | /* Keep page count to indicate a given hugepage is isolated. */ | |
189ebff2 | 1432 | ret = migrate_huge_page(hpage, new_page, MPOL_MF_MOVE_ALL, false, |
dc32f634 | 1433 | MIGRATE_SYNC); |
189ebff2 | 1434 | put_page(hpage); |
d950b958 | 1435 | if (ret) { |
dd73e85f DN |
1436 | pr_info("soft offline: %#lx: migration failed %d, type %lx\n", |
1437 | pfn, ret, page->flags); | |
d950b958 NH |
1438 | return ret; |
1439 | } | |
1440 | done: | |
1441 | if (!PageHWPoison(hpage)) | |
189ebff2 AK |
1442 | atomic_long_add(1 << compound_trans_order(hpage), |
1443 | &mce_bad_pages); | |
d950b958 NH |
1444 | set_page_hwpoison_huge_page(hpage); |
1445 | dequeue_hwpoisoned_huge_page(hpage); | |
1446 | /* keep elevated page count for bad page */ | |
1447 | return ret; | |
1448 | } | |
1449 | ||
facb6011 AK |
1450 | /** |
1451 | * soft_offline_page - Soft offline a page. | |
1452 | * @page: page to offline | |
1453 | * @flags: flags. Same as memory_failure(). | |
1454 | * | |
1455 | * Returns 0 on success, otherwise negated errno. | |
1456 | * | |
1457 | * Soft offline a page, by migration or invalidation, | |
1458 | * without killing anything. This is for the case when | |
1459 | * a page is not corrupted yet (so it's still valid to access), | |
1460 | * but has had a number of corrected errors and is better taken | |
1461 | * out. | |
1462 | * | |
1463 | * The actual policy on when to do that is maintained by | |
1464 | * user space. | |
1465 | * | |
1466 | * This should never impact any application or cause data loss, | |
1467 | * however it might take some time. | |
1468 | * | |
1469 | * This is not a 100% solution for all memory, but tries to be | |
1470 | * ``good enough'' for the majority of memory. | |
1471 | */ | |
1472 | int soft_offline_page(struct page *page, int flags) | |
1473 | { | |
1474 | int ret; | |
1475 | unsigned long pfn = page_to_pfn(page); | |
1476 | ||
d950b958 NH |
1477 | if (PageHuge(page)) |
1478 | return soft_offline_huge_page(page, flags); | |
1479 | ||
facb6011 AK |
1480 | ret = get_any_page(page, pfn, flags); |
1481 | if (ret < 0) | |
1482 | return ret; | |
1483 | if (ret == 0) | |
1484 | goto done; | |
1485 | ||
1486 | /* | |
1487 | * Page cache page we can handle? | |
1488 | */ | |
1489 | if (!PageLRU(page)) { | |
1490 | /* | |
1491 | * Try to free it. | |
1492 | */ | |
1493 | put_page(page); | |
1494 | shake_page(page, 1); | |
1495 | ||
1496 | /* | |
1497 | * Did it turn free? | |
1498 | */ | |
1499 | ret = get_any_page(page, pfn, 0); | |
1500 | if (ret < 0) | |
1501 | return ret; | |
1502 | if (ret == 0) | |
1503 | goto done; | |
1504 | } | |
1505 | if (!PageLRU(page)) { | |
fb46e735 | 1506 | pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n", |
dd73e85f | 1507 | pfn, page->flags); |
facb6011 AK |
1508 | return -EIO; |
1509 | } | |
1510 | ||
1511 | lock_page(page); | |
1512 | wait_on_page_writeback(page); | |
1513 | ||
1514 | /* | |
1515 | * Synchronized using the page lock with memory_failure() | |
1516 | */ | |
1517 | if (PageHWPoison(page)) { | |
1518 | unlock_page(page); | |
1519 | put_page(page); | |
fb46e735 | 1520 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
facb6011 AK |
1521 | return -EBUSY; |
1522 | } | |
1523 | ||
1524 | /* | |
1525 | * Try to invalidate first. This should work for | |
1526 | * non dirty unmapped page cache pages. | |
1527 | */ | |
1528 | ret = invalidate_inode_page(page); | |
1529 | unlock_page(page); | |
facb6011 | 1530 | /* |
facb6011 AK |
1531 | * RED-PEN would be better to keep it isolated here, but we |
1532 | * would need to fix isolation locking first. | |
1533 | */ | |
facb6011 | 1534 | if (ret == 1) { |
bd486285 | 1535 | put_page(page); |
facb6011 | 1536 | ret = 0; |
fb46e735 | 1537 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
facb6011 AK |
1538 | goto done; |
1539 | } | |
1540 | ||
1541 | /* | |
1542 | * Simple invalidation didn't work. | |
1543 | * Try to migrate to a new page instead. migrate.c | |
1544 | * handles a large number of cases for us. | |
1545 | */ | |
1546 | ret = isolate_lru_page(page); | |
bd486285 KK |
1547 | /* |
1548 | * Drop page reference which is came from get_any_page() | |
1549 | * successful isolate_lru_page() already took another one. | |
1550 | */ | |
1551 | put_page(page); | |
facb6011 AK |
1552 | if (!ret) { |
1553 | LIST_HEAD(pagelist); | |
5db8a73a MK |
1554 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
1555 | page_is_file_cache(page)); | |
facb6011 | 1556 | list_add(&page->lru, &pagelist); |
77f1fe6b | 1557 | ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, |
dc32f634 | 1558 | false, MIGRATE_SYNC); |
facb6011 | 1559 | if (ret) { |
57fc4a5e | 1560 | putback_lru_pages(&pagelist); |
fb46e735 | 1561 | pr_info("soft offline: %#lx: migration failed %d, type %lx\n", |
facb6011 AK |
1562 | pfn, ret, page->flags); |
1563 | if (ret > 0) | |
1564 | ret = -EIO; | |
1565 | } | |
1566 | } else { | |
fb46e735 | 1567 | pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n", |
dd73e85f | 1568 | pfn, ret, page_count(page), page->flags); |
facb6011 AK |
1569 | } |
1570 | if (ret) | |
1571 | return ret; | |
1572 | ||
1573 | done: | |
1574 | atomic_long_add(1, &mce_bad_pages); | |
1575 | SetPageHWPoison(page); | |
1576 | /* keep elevated page count for bad page */ | |
1577 | return ret; | |
1578 | } |