]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Copyright (C) 2008, 2009 Intel Corporation | |
3 | * Authors: Andi Kleen, Fengguang Wu | |
4 | * | |
5 | * This software may be redistributed and/or modified under the terms of | |
6 | * the GNU General Public License ("GPL") version 2 only as published by the | |
7 | * Free Software Foundation. | |
8 | * | |
9 | * High level machine check handler. Handles pages reported by the | |
10 | * hardware as being corrupted usually due to a multi-bit ECC memory or cache | |
11 | * failure. | |
12 | * | |
13 | * In addition there is a "soft offline" entry point that allows stop using | |
14 | * not-yet-corrupted-by-suspicious pages without killing anything. | |
15 | * | |
16 | * Handles page cache pages in various states. The tricky part | |
17 | * here is that we can access any page asynchronously in respect to | |
18 | * other VM users, because memory failures could happen anytime and | |
19 | * anywhere. This could violate some of their assumptions. This is why | |
20 | * this code has to be extremely careful. Generally it tries to use | |
21 | * normal locking rules, as in get the standard locks, even if that means | |
22 | * the error handling takes potentially a long time. | |
23 | * | |
24 | * It can be very tempting to add handling for obscure cases here. | |
25 | * In general any code for handling new cases should only be added iff: | |
26 | * - You know how to test it. | |
27 | * - You have a test that can be added to mce-test | |
28 | * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/ | |
29 | * - The case actually shows up as a frequent (top 10) page state in | |
30 | * tools/vm/page-types when running a real workload. | |
31 | * | |
32 | * There are several operations here with exponential complexity because | |
33 | * of unsuitable VM data structures. For example the operation to map back | |
34 | * from RMAP chains to processes has to walk the complete process list and | |
35 | * has non linear complexity with the number. But since memory corruptions | |
36 | * are rare we hope to get away with this. This avoids impacting the core | |
37 | * VM. | |
38 | */ | |
39 | #include <linux/kernel.h> | |
40 | #include <linux/mm.h> | |
41 | #include <linux/page-flags.h> | |
42 | #include <linux/kernel-page-flags.h> | |
43 | #include <linux/sched/signal.h> | |
44 | #include <linux/sched/task.h> | |
45 | #include <linux/ksm.h> | |
46 | #include <linux/rmap.h> | |
47 | #include <linux/export.h> | |
48 | #include <linux/pagemap.h> | |
49 | #include <linux/swap.h> | |
50 | #include <linux/backing-dev.h> | |
51 | #include <linux/migrate.h> | |
52 | #include <linux/page-isolation.h> | |
53 | #include <linux/suspend.h> | |
54 | #include <linux/slab.h> | |
55 | #include <linux/swapops.h> | |
56 | #include <linux/hugetlb.h> | |
57 | #include <linux/memory_hotplug.h> | |
58 | #include <linux/mm_inline.h> | |
59 | #include <linux/kfifo.h> | |
60 | #include <linux/ratelimit.h> | |
61 | #include "internal.h" | |
62 | #include "ras/ras_event.h" | |
63 | ||
64 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
65 | ||
66 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
67 | ||
68 | atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); | |
69 | ||
70 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) | |
71 | ||
72 | u32 hwpoison_filter_enable = 0; | |
73 | u32 hwpoison_filter_dev_major = ~0U; | |
74 | u32 hwpoison_filter_dev_minor = ~0U; | |
75 | u64 hwpoison_filter_flags_mask; | |
76 | u64 hwpoison_filter_flags_value; | |
77 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); | |
78 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); | |
79 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
80 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); | |
81 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
82 | ||
83 | static int hwpoison_filter_dev(struct page *p) | |
84 | { | |
85 | struct address_space *mapping; | |
86 | dev_t dev; | |
87 | ||
88 | if (hwpoison_filter_dev_major == ~0U && | |
89 | hwpoison_filter_dev_minor == ~0U) | |
90 | return 0; | |
91 | ||
92 | /* | |
93 | * page_mapping() does not accept slab pages. | |
94 | */ | |
95 | if (PageSlab(p)) | |
96 | return -EINVAL; | |
97 | ||
98 | mapping = page_mapping(p); | |
99 | if (mapping == NULL || mapping->host == NULL) | |
100 | return -EINVAL; | |
101 | ||
102 | dev = mapping->host->i_sb->s_dev; | |
103 | if (hwpoison_filter_dev_major != ~0U && | |
104 | hwpoison_filter_dev_major != MAJOR(dev)) | |
105 | return -EINVAL; | |
106 | if (hwpoison_filter_dev_minor != ~0U && | |
107 | hwpoison_filter_dev_minor != MINOR(dev)) | |
108 | return -EINVAL; | |
109 | ||
110 | return 0; | |
111 | } | |
112 | ||
113 | static int hwpoison_filter_flags(struct page *p) | |
114 | { | |
115 | if (!hwpoison_filter_flags_mask) | |
116 | return 0; | |
117 | ||
118 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
119 | hwpoison_filter_flags_value) | |
120 | return 0; | |
121 | else | |
122 | return -EINVAL; | |
123 | } | |
124 | ||
125 | /* | |
126 | * This allows stress tests to limit test scope to a collection of tasks | |
127 | * by putting them under some memcg. This prevents killing unrelated/important | |
128 | * processes such as /sbin/init. Note that the target task may share clean | |
129 | * pages with init (eg. libc text), which is harmless. If the target task | |
130 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
131 | * is also included in the memcg. At last, due to race conditions this filter | |
132 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
133 | * a freed page. | |
134 | */ | |
135 | #ifdef CONFIG_MEMCG | |
136 | u64 hwpoison_filter_memcg; | |
137 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
138 | static int hwpoison_filter_task(struct page *p) | |
139 | { | |
140 | if (!hwpoison_filter_memcg) | |
141 | return 0; | |
142 | ||
143 | if (page_cgroup_ino(p) != hwpoison_filter_memcg) | |
144 | return -EINVAL; | |
145 | ||
146 | return 0; | |
147 | } | |
148 | #else | |
149 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
150 | #endif | |
151 | ||
152 | int hwpoison_filter(struct page *p) | |
153 | { | |
154 | if (!hwpoison_filter_enable) | |
155 | return 0; | |
156 | ||
157 | if (hwpoison_filter_dev(p)) | |
158 | return -EINVAL; | |
159 | ||
160 | if (hwpoison_filter_flags(p)) | |
161 | return -EINVAL; | |
162 | ||
163 | if (hwpoison_filter_task(p)) | |
164 | return -EINVAL; | |
165 | ||
166 | return 0; | |
167 | } | |
168 | #else | |
169 | int hwpoison_filter(struct page *p) | |
170 | { | |
171 | return 0; | |
172 | } | |
173 | #endif | |
174 | ||
175 | EXPORT_SYMBOL_GPL(hwpoison_filter); | |
176 | ||
177 | /* | |
178 | * Send all the processes who have the page mapped a signal. | |
179 | * ``action optional'' if they are not immediately affected by the error | |
180 | * ``action required'' if error happened in current execution context | |
181 | */ | |
182 | static int kill_proc(struct task_struct *t, unsigned long addr, int trapno, | |
183 | unsigned long pfn, struct page *page, int flags) | |
184 | { | |
185 | struct siginfo si; | |
186 | int ret; | |
187 | ||
188 | pr_err("Memory failure: %#lx: Killing %s:%d due to hardware memory corruption\n", | |
189 | pfn, t->comm, t->pid); | |
190 | si.si_signo = SIGBUS; | |
191 | si.si_errno = 0; | |
192 | si.si_addr = (void *)addr; | |
193 | #ifdef __ARCH_SI_TRAPNO | |
194 | si.si_trapno = trapno; | |
195 | #endif | |
196 | si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT; | |
197 | ||
198 | if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) { | |
199 | si.si_code = BUS_MCEERR_AR; | |
200 | ret = force_sig_info(SIGBUS, &si, current); | |
201 | } else { | |
202 | /* | |
203 | * Don't use force here, it's convenient if the signal | |
204 | * can be temporarily blocked. | |
205 | * This could cause a loop when the user sets SIGBUS | |
206 | * to SIG_IGN, but hopefully no one will do that? | |
207 | */ | |
208 | si.si_code = BUS_MCEERR_AO; | |
209 | ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */ | |
210 | } | |
211 | if (ret < 0) | |
212 | pr_info("Memory failure: Error sending signal to %s:%d: %d\n", | |
213 | t->comm, t->pid, ret); | |
214 | return ret; | |
215 | } | |
216 | ||
217 | /* | |
218 | * When a unknown page type is encountered drain as many buffers as possible | |
219 | * in the hope to turn the page into a LRU or free page, which we can handle. | |
220 | */ | |
221 | void shake_page(struct page *p, int access) | |
222 | { | |
223 | if (PageHuge(p)) | |
224 | return; | |
225 | ||
226 | if (!PageSlab(p)) { | |
227 | lru_add_drain_all(); | |
228 | if (PageLRU(p)) | |
229 | return; | |
230 | drain_all_pages(page_zone(p)); | |
231 | if (PageLRU(p) || is_free_buddy_page(p)) | |
232 | return; | |
233 | } | |
234 | ||
235 | /* | |
236 | * Only call shrink_node_slabs here (which would also shrink | |
237 | * other caches) if access is not potentially fatal. | |
238 | */ | |
239 | if (access) | |
240 | drop_slab_node(page_to_nid(p)); | |
241 | } | |
242 | EXPORT_SYMBOL_GPL(shake_page); | |
243 | ||
244 | /* | |
245 | * Kill all processes that have a poisoned page mapped and then isolate | |
246 | * the page. | |
247 | * | |
248 | * General strategy: | |
249 | * Find all processes having the page mapped and kill them. | |
250 | * But we keep a page reference around so that the page is not | |
251 | * actually freed yet. | |
252 | * Then stash the page away | |
253 | * | |
254 | * There's no convenient way to get back to mapped processes | |
255 | * from the VMAs. So do a brute-force search over all | |
256 | * running processes. | |
257 | * | |
258 | * Remember that machine checks are not common (or rather | |
259 | * if they are common you have other problems), so this shouldn't | |
260 | * be a performance issue. | |
261 | * | |
262 | * Also there are some races possible while we get from the | |
263 | * error detection to actually handle it. | |
264 | */ | |
265 | ||
266 | struct to_kill { | |
267 | struct list_head nd; | |
268 | struct task_struct *tsk; | |
269 | unsigned long addr; | |
270 | char addr_valid; | |
271 | }; | |
272 | ||
273 | /* | |
274 | * Failure handling: if we can't find or can't kill a process there's | |
275 | * not much we can do. We just print a message and ignore otherwise. | |
276 | */ | |
277 | ||
278 | /* | |
279 | * Schedule a process for later kill. | |
280 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
281 | * TBD would GFP_NOIO be enough? | |
282 | */ | |
283 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
284 | struct vm_area_struct *vma, | |
285 | struct list_head *to_kill, | |
286 | struct to_kill **tkc) | |
287 | { | |
288 | struct to_kill *tk; | |
289 | ||
290 | if (*tkc) { | |
291 | tk = *tkc; | |
292 | *tkc = NULL; | |
293 | } else { | |
294 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); | |
295 | if (!tk) { | |
296 | pr_err("Memory failure: Out of memory while machine check handling\n"); | |
297 | return; | |
298 | } | |
299 | } | |
300 | tk->addr = page_address_in_vma(p, vma); | |
301 | tk->addr_valid = 1; | |
302 | ||
303 | /* | |
304 | * In theory we don't have to kill when the page was | |
305 | * munmaped. But it could be also a mremap. Since that's | |
306 | * likely very rare kill anyways just out of paranoia, but use | |
307 | * a SIGKILL because the error is not contained anymore. | |
308 | */ | |
309 | if (tk->addr == -EFAULT) { | |
310 | pr_info("Memory failure: Unable to find user space address %lx in %s\n", | |
311 | page_to_pfn(p), tsk->comm); | |
312 | tk->addr_valid = 0; | |
313 | } | |
314 | get_task_struct(tsk); | |
315 | tk->tsk = tsk; | |
316 | list_add_tail(&tk->nd, to_kill); | |
317 | } | |
318 | ||
319 | /* | |
320 | * Kill the processes that have been collected earlier. | |
321 | * | |
322 | * Only do anything when DOIT is set, otherwise just free the list | |
323 | * (this is used for clean pages which do not need killing) | |
324 | * Also when FAIL is set do a force kill because something went | |
325 | * wrong earlier. | |
326 | */ | |
327 | static void kill_procs(struct list_head *to_kill, int forcekill, int trapno, | |
328 | bool fail, struct page *page, unsigned long pfn, | |
329 | int flags) | |
330 | { | |
331 | struct to_kill *tk, *next; | |
332 | ||
333 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
334 | if (forcekill) { | |
335 | /* | |
336 | * In case something went wrong with munmapping | |
337 | * make sure the process doesn't catch the | |
338 | * signal and then access the memory. Just kill it. | |
339 | */ | |
340 | if (fail || tk->addr_valid == 0) { | |
341 | pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", | |
342 | pfn, tk->tsk->comm, tk->tsk->pid); | |
343 | force_sig(SIGKILL, tk->tsk); | |
344 | } | |
345 | ||
346 | /* | |
347 | * In theory the process could have mapped | |
348 | * something else on the address in-between. We could | |
349 | * check for that, but we need to tell the | |
350 | * process anyways. | |
351 | */ | |
352 | else if (kill_proc(tk->tsk, tk->addr, trapno, | |
353 | pfn, page, flags) < 0) | |
354 | pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n", | |
355 | pfn, tk->tsk->comm, tk->tsk->pid); | |
356 | } | |
357 | put_task_struct(tk->tsk); | |
358 | kfree(tk); | |
359 | } | |
360 | } | |
361 | ||
362 | /* | |
363 | * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) | |
364 | * on behalf of the thread group. Return task_struct of the (first found) | |
365 | * dedicated thread if found, and return NULL otherwise. | |
366 | * | |
367 | * We already hold read_lock(&tasklist_lock) in the caller, so we don't | |
368 | * have to call rcu_read_lock/unlock() in this function. | |
369 | */ | |
370 | static struct task_struct *find_early_kill_thread(struct task_struct *tsk) | |
371 | { | |
372 | struct task_struct *t; | |
373 | ||
374 | for_each_thread(tsk, t) | |
375 | if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY)) | |
376 | return t; | |
377 | return NULL; | |
378 | } | |
379 | ||
380 | /* | |
381 | * Determine whether a given process is "early kill" process which expects | |
382 | * to be signaled when some page under the process is hwpoisoned. | |
383 | * Return task_struct of the dedicated thread (main thread unless explicitly | |
384 | * specified) if the process is "early kill," and otherwise returns NULL. | |
385 | */ | |
386 | static struct task_struct *task_early_kill(struct task_struct *tsk, | |
387 | int force_early) | |
388 | { | |
389 | struct task_struct *t; | |
390 | if (!tsk->mm) | |
391 | return NULL; | |
392 | if (force_early) | |
393 | return tsk; | |
394 | t = find_early_kill_thread(tsk); | |
395 | if (t) | |
396 | return t; | |
397 | if (sysctl_memory_failure_early_kill) | |
398 | return tsk; | |
399 | return NULL; | |
400 | } | |
401 | ||
402 | /* | |
403 | * Collect processes when the error hit an anonymous page. | |
404 | */ | |
405 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
406 | struct to_kill **tkc, int force_early) | |
407 | { | |
408 | struct vm_area_struct *vma; | |
409 | struct task_struct *tsk; | |
410 | struct anon_vma *av; | |
411 | pgoff_t pgoff; | |
412 | ||
413 | av = page_lock_anon_vma_read(page); | |
414 | if (av == NULL) /* Not actually mapped anymore */ | |
415 | return; | |
416 | ||
417 | pgoff = page_to_pgoff(page); | |
418 | read_lock(&tasklist_lock); | |
419 | for_each_process (tsk) { | |
420 | struct anon_vma_chain *vmac; | |
421 | struct task_struct *t = task_early_kill(tsk, force_early); | |
422 | ||
423 | if (!t) | |
424 | continue; | |
425 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, | |
426 | pgoff, pgoff) { | |
427 | vma = vmac->vma; | |
428 | if (!page_mapped_in_vma(page, vma)) | |
429 | continue; | |
430 | if (vma->vm_mm == t->mm) | |
431 | add_to_kill(t, page, vma, to_kill, tkc); | |
432 | } | |
433 | } | |
434 | read_unlock(&tasklist_lock); | |
435 | page_unlock_anon_vma_read(av); | |
436 | } | |
437 | ||
438 | /* | |
439 | * Collect processes when the error hit a file mapped page. | |
440 | */ | |
441 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
442 | struct to_kill **tkc, int force_early) | |
443 | { | |
444 | struct vm_area_struct *vma; | |
445 | struct task_struct *tsk; | |
446 | struct address_space *mapping = page->mapping; | |
447 | ||
448 | i_mmap_lock_read(mapping); | |
449 | read_lock(&tasklist_lock); | |
450 | for_each_process(tsk) { | |
451 | pgoff_t pgoff = page_to_pgoff(page); | |
452 | struct task_struct *t = task_early_kill(tsk, force_early); | |
453 | ||
454 | if (!t) | |
455 | continue; | |
456 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, | |
457 | pgoff) { | |
458 | /* | |
459 | * Send early kill signal to tasks where a vma covers | |
460 | * the page but the corrupted page is not necessarily | |
461 | * mapped it in its pte. | |
462 | * Assume applications who requested early kill want | |
463 | * to be informed of all such data corruptions. | |
464 | */ | |
465 | if (vma->vm_mm == t->mm) | |
466 | add_to_kill(t, page, vma, to_kill, tkc); | |
467 | } | |
468 | } | |
469 | read_unlock(&tasklist_lock); | |
470 | i_mmap_unlock_read(mapping); | |
471 | } | |
472 | ||
473 | /* | |
474 | * Collect the processes who have the corrupted page mapped to kill. | |
475 | * This is done in two steps for locking reasons. | |
476 | * First preallocate one tokill structure outside the spin locks, | |
477 | * so that we can kill at least one process reasonably reliable. | |
478 | */ | |
479 | static void collect_procs(struct page *page, struct list_head *tokill, | |
480 | int force_early) | |
481 | { | |
482 | struct to_kill *tk; | |
483 | ||
484 | if (!page->mapping) | |
485 | return; | |
486 | ||
487 | tk = kmalloc(sizeof(struct to_kill), GFP_NOIO); | |
488 | if (!tk) | |
489 | return; | |
490 | if (PageAnon(page)) | |
491 | collect_procs_anon(page, tokill, &tk, force_early); | |
492 | else | |
493 | collect_procs_file(page, tokill, &tk, force_early); | |
494 | kfree(tk); | |
495 | } | |
496 | ||
497 | static const char *action_name[] = { | |
498 | [MF_IGNORED] = "Ignored", | |
499 | [MF_FAILED] = "Failed", | |
500 | [MF_DELAYED] = "Delayed", | |
501 | [MF_RECOVERED] = "Recovered", | |
502 | }; | |
503 | ||
504 | static const char * const action_page_types[] = { | |
505 | [MF_MSG_KERNEL] = "reserved kernel page", | |
506 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", | |
507 | [MF_MSG_SLAB] = "kernel slab page", | |
508 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", | |
509 | [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned", | |
510 | [MF_MSG_HUGE] = "huge page", | |
511 | [MF_MSG_FREE_HUGE] = "free huge page", | |
512 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page", | |
513 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", | |
514 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", | |
515 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", | |
516 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", | |
517 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", | |
518 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", | |
519 | [MF_MSG_DIRTY_LRU] = "dirty LRU page", | |
520 | [MF_MSG_CLEAN_LRU] = "clean LRU page", | |
521 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", | |
522 | [MF_MSG_BUDDY] = "free buddy page", | |
523 | [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)", | |
524 | [MF_MSG_UNKNOWN] = "unknown page", | |
525 | }; | |
526 | ||
527 | /* | |
528 | * XXX: It is possible that a page is isolated from LRU cache, | |
529 | * and then kept in swap cache or failed to remove from page cache. | |
530 | * The page count will stop it from being freed by unpoison. | |
531 | * Stress tests should be aware of this memory leak problem. | |
532 | */ | |
533 | static int delete_from_lru_cache(struct page *p) | |
534 | { | |
535 | if (!isolate_lru_page(p)) { | |
536 | /* | |
537 | * Clear sensible page flags, so that the buddy system won't | |
538 | * complain when the page is unpoison-and-freed. | |
539 | */ | |
540 | ClearPageActive(p); | |
541 | ClearPageUnevictable(p); | |
542 | ||
543 | /* | |
544 | * Poisoned page might never drop its ref count to 0 so we have | |
545 | * to uncharge it manually from its memcg. | |
546 | */ | |
547 | mem_cgroup_uncharge(p); | |
548 | ||
549 | /* | |
550 | * drop the page count elevated by isolate_lru_page() | |
551 | */ | |
552 | put_page(p); | |
553 | return 0; | |
554 | } | |
555 | return -EIO; | |
556 | } | |
557 | ||
558 | /* | |
559 | * Error hit kernel page. | |
560 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
561 | * could be more sophisticated. | |
562 | */ | |
563 | static int me_kernel(struct page *p, unsigned long pfn) | |
564 | { | |
565 | return MF_IGNORED; | |
566 | } | |
567 | ||
568 | /* | |
569 | * Page in unknown state. Do nothing. | |
570 | */ | |
571 | static int me_unknown(struct page *p, unsigned long pfn) | |
572 | { | |
573 | pr_err("Memory failure: %#lx: Unknown page state\n", pfn); | |
574 | return MF_FAILED; | |
575 | } | |
576 | ||
577 | /* | |
578 | * Clean (or cleaned) page cache page. | |
579 | */ | |
580 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
581 | { | |
582 | int err; | |
583 | int ret = MF_FAILED; | |
584 | struct address_space *mapping; | |
585 | ||
586 | delete_from_lru_cache(p); | |
587 | ||
588 | /* | |
589 | * For anonymous pages we're done the only reference left | |
590 | * should be the one m_f() holds. | |
591 | */ | |
592 | if (PageAnon(p)) | |
593 | return MF_RECOVERED; | |
594 | ||
595 | /* | |
596 | * Now truncate the page in the page cache. This is really | |
597 | * more like a "temporary hole punch" | |
598 | * Don't do this for block devices when someone else | |
599 | * has a reference, because it could be file system metadata | |
600 | * and that's not safe to truncate. | |
601 | */ | |
602 | mapping = page_mapping(p); | |
603 | if (!mapping) { | |
604 | /* | |
605 | * Page has been teared down in the meanwhile | |
606 | */ | |
607 | return MF_FAILED; | |
608 | } | |
609 | ||
610 | /* | |
611 | * Truncation is a bit tricky. Enable it per file system for now. | |
612 | * | |
613 | * Open: to take i_mutex or not for this? Right now we don't. | |
614 | */ | |
615 | if (mapping->a_ops->error_remove_page) { | |
616 | err = mapping->a_ops->error_remove_page(mapping, p); | |
617 | if (err != 0) { | |
618 | pr_info("Memory failure: %#lx: Failed to punch page: %d\n", | |
619 | pfn, err); | |
620 | } else if (page_has_private(p) && | |
621 | !try_to_release_page(p, GFP_NOIO)) { | |
622 | pr_info("Memory failure: %#lx: failed to release buffers\n", | |
623 | pfn); | |
624 | } else { | |
625 | ret = MF_RECOVERED; | |
626 | } | |
627 | } else { | |
628 | /* | |
629 | * If the file system doesn't support it just invalidate | |
630 | * This fails on dirty or anything with private pages | |
631 | */ | |
632 | if (invalidate_inode_page(p)) | |
633 | ret = MF_RECOVERED; | |
634 | else | |
635 | pr_info("Memory failure: %#lx: Failed to invalidate\n", | |
636 | pfn); | |
637 | } | |
638 | return ret; | |
639 | } | |
640 | ||
641 | /* | |
642 | * Dirty pagecache page | |
643 | * Issues: when the error hit a hole page the error is not properly | |
644 | * propagated. | |
645 | */ | |
646 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
647 | { | |
648 | struct address_space *mapping = page_mapping(p); | |
649 | ||
650 | SetPageError(p); | |
651 | /* TBD: print more information about the file. */ | |
652 | if (mapping) { | |
653 | /* | |
654 | * IO error will be reported by write(), fsync(), etc. | |
655 | * who check the mapping. | |
656 | * This way the application knows that something went | |
657 | * wrong with its dirty file data. | |
658 | * | |
659 | * There's one open issue: | |
660 | * | |
661 | * The EIO will be only reported on the next IO | |
662 | * operation and then cleared through the IO map. | |
663 | * Normally Linux has two mechanisms to pass IO error | |
664 | * first through the AS_EIO flag in the address space | |
665 | * and then through the PageError flag in the page. | |
666 | * Since we drop pages on memory failure handling the | |
667 | * only mechanism open to use is through AS_AIO. | |
668 | * | |
669 | * This has the disadvantage that it gets cleared on | |
670 | * the first operation that returns an error, while | |
671 | * the PageError bit is more sticky and only cleared | |
672 | * when the page is reread or dropped. If an | |
673 | * application assumes it will always get error on | |
674 | * fsync, but does other operations on the fd before | |
675 | * and the page is dropped between then the error | |
676 | * will not be properly reported. | |
677 | * | |
678 | * This can already happen even without hwpoisoned | |
679 | * pages: first on metadata IO errors (which only | |
680 | * report through AS_EIO) or when the page is dropped | |
681 | * at the wrong time. | |
682 | * | |
683 | * So right now we assume that the application DTRT on | |
684 | * the first EIO, but we're not worse than other parts | |
685 | * of the kernel. | |
686 | */ | |
687 | mapping_set_error(mapping, EIO); | |
688 | } | |
689 | ||
690 | return me_pagecache_clean(p, pfn); | |
691 | } | |
692 | ||
693 | /* | |
694 | * Clean and dirty swap cache. | |
695 | * | |
696 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
697 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
698 | * referenced concurrently by 2 types of PTEs: | |
699 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
700 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
701 | * and then | |
702 | * - clear dirty bit to prevent IO | |
703 | * - remove from LRU | |
704 | * - but keep in the swap cache, so that when we return to it on | |
705 | * a later page fault, we know the application is accessing | |
706 | * corrupted data and shall be killed (we installed simple | |
707 | * interception code in do_swap_page to catch it). | |
708 | * | |
709 | * Clean swap cache pages can be directly isolated. A later page fault will | |
710 | * bring in the known good data from disk. | |
711 | */ | |
712 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
713 | { | |
714 | ClearPageDirty(p); | |
715 | /* Trigger EIO in shmem: */ | |
716 | ClearPageUptodate(p); | |
717 | ||
718 | if (!delete_from_lru_cache(p)) | |
719 | return MF_DELAYED; | |
720 | else | |
721 | return MF_FAILED; | |
722 | } | |
723 | ||
724 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
725 | { | |
726 | delete_from_swap_cache(p); | |
727 | ||
728 | if (!delete_from_lru_cache(p)) | |
729 | return MF_RECOVERED; | |
730 | else | |
731 | return MF_FAILED; | |
732 | } | |
733 | ||
734 | /* | |
735 | * Huge pages. Needs work. | |
736 | * Issues: | |
737 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) | |
738 | * To narrow down kill region to one page, we need to break up pmd. | |
739 | */ | |
740 | static int me_huge_page(struct page *p, unsigned long pfn) | |
741 | { | |
742 | int res = 0; | |
743 | struct page *hpage = compound_head(p); | |
744 | ||
745 | if (!PageHuge(hpage)) | |
746 | return MF_DELAYED; | |
747 | ||
748 | /* | |
749 | * We can safely recover from error on free or reserved (i.e. | |
750 | * not in-use) hugepage by dequeuing it from freelist. | |
751 | * To check whether a hugepage is in-use or not, we can't use | |
752 | * page->lru because it can be used in other hugepage operations, | |
753 | * such as __unmap_hugepage_range() and gather_surplus_pages(). | |
754 | * So instead we use page_mapping() and PageAnon(). | |
755 | */ | |
756 | if (!(page_mapping(hpage) || PageAnon(hpage))) { | |
757 | res = dequeue_hwpoisoned_huge_page(hpage); | |
758 | if (!res) | |
759 | return MF_RECOVERED; | |
760 | } | |
761 | return MF_DELAYED; | |
762 | } | |
763 | ||
764 | /* | |
765 | * Various page states we can handle. | |
766 | * | |
767 | * A page state is defined by its current page->flags bits. | |
768 | * The table matches them in order and calls the right handler. | |
769 | * | |
770 | * This is quite tricky because we can access page at any time | |
771 | * in its live cycle, so all accesses have to be extremely careful. | |
772 | * | |
773 | * This is not complete. More states could be added. | |
774 | * For any missing state don't attempt recovery. | |
775 | */ | |
776 | ||
777 | #define dirty (1UL << PG_dirty) | |
778 | #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) | |
779 | #define unevict (1UL << PG_unevictable) | |
780 | #define mlock (1UL << PG_mlocked) | |
781 | #define writeback (1UL << PG_writeback) | |
782 | #define lru (1UL << PG_lru) | |
783 | #define head (1UL << PG_head) | |
784 | #define slab (1UL << PG_slab) | |
785 | #define reserved (1UL << PG_reserved) | |
786 | ||
787 | static struct page_state { | |
788 | unsigned long mask; | |
789 | unsigned long res; | |
790 | enum mf_action_page_type type; | |
791 | int (*action)(struct page *p, unsigned long pfn); | |
792 | } error_states[] = { | |
793 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, | |
794 | /* | |
795 | * free pages are specially detected outside this table: | |
796 | * PG_buddy pages only make a small fraction of all free pages. | |
797 | */ | |
798 | ||
799 | /* | |
800 | * Could in theory check if slab page is free or if we can drop | |
801 | * currently unused objects without touching them. But just | |
802 | * treat it as standard kernel for now. | |
803 | */ | |
804 | { slab, slab, MF_MSG_SLAB, me_kernel }, | |
805 | ||
806 | { head, head, MF_MSG_HUGE, me_huge_page }, | |
807 | ||
808 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, | |
809 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, | |
810 | ||
811 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, | |
812 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, | |
813 | ||
814 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, | |
815 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, | |
816 | ||
817 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, | |
818 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, | |
819 | ||
820 | /* | |
821 | * Catchall entry: must be at end. | |
822 | */ | |
823 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, | |
824 | }; | |
825 | ||
826 | #undef dirty | |
827 | #undef sc | |
828 | #undef unevict | |
829 | #undef mlock | |
830 | #undef writeback | |
831 | #undef lru | |
832 | #undef head | |
833 | #undef slab | |
834 | #undef reserved | |
835 | ||
836 | /* | |
837 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of | |
838 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). | |
839 | */ | |
840 | static void action_result(unsigned long pfn, enum mf_action_page_type type, | |
841 | enum mf_result result) | |
842 | { | |
843 | trace_memory_failure_event(pfn, type, result); | |
844 | ||
845 | pr_err("Memory failure: %#lx: recovery action for %s: %s\n", | |
846 | pfn, action_page_types[type], action_name[result]); | |
847 | } | |
848 | ||
849 | static int page_action(struct page_state *ps, struct page *p, | |
850 | unsigned long pfn) | |
851 | { | |
852 | int result; | |
853 | int count; | |
854 | ||
855 | result = ps->action(p, pfn); | |
856 | ||
857 | count = page_count(p) - 1; | |
858 | if (ps->action == me_swapcache_dirty && result == MF_DELAYED) | |
859 | count--; | |
860 | if (count != 0) { | |
861 | pr_err("Memory failure: %#lx: %s still referenced by %d users\n", | |
862 | pfn, action_page_types[ps->type], count); | |
863 | result = MF_FAILED; | |
864 | } | |
865 | action_result(pfn, ps->type, result); | |
866 | ||
867 | /* Could do more checks here if page looks ok */ | |
868 | /* | |
869 | * Could adjust zone counters here to correct for the missing page. | |
870 | */ | |
871 | ||
872 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; | |
873 | } | |
874 | ||
875 | /** | |
876 | * get_hwpoison_page() - Get refcount for memory error handling: | |
877 | * @page: raw error page (hit by memory error) | |
878 | * | |
879 | * Return: return 0 if failed to grab the refcount, otherwise true (some | |
880 | * non-zero value.) | |
881 | */ | |
882 | int get_hwpoison_page(struct page *page) | |
883 | { | |
884 | struct page *head = compound_head(page); | |
885 | ||
886 | if (!PageHuge(head) && PageTransHuge(head)) { | |
887 | /* | |
888 | * Non anonymous thp exists only in allocation/free time. We | |
889 | * can't handle such a case correctly, so let's give it up. | |
890 | * This should be better than triggering BUG_ON when kernel | |
891 | * tries to touch the "partially handled" page. | |
892 | */ | |
893 | if (!PageAnon(head)) { | |
894 | pr_err("Memory failure: %#lx: non anonymous thp\n", | |
895 | page_to_pfn(page)); | |
896 | return 0; | |
897 | } | |
898 | } | |
899 | ||
900 | if (get_page_unless_zero(head)) { | |
901 | if (head == compound_head(page)) | |
902 | return 1; | |
903 | ||
904 | pr_info("Memory failure: %#lx cannot catch tail\n", | |
905 | page_to_pfn(page)); | |
906 | put_page(head); | |
907 | } | |
908 | ||
909 | return 0; | |
910 | } | |
911 | EXPORT_SYMBOL_GPL(get_hwpoison_page); | |
912 | ||
913 | /* | |
914 | * Do all that is necessary to remove user space mappings. Unmap | |
915 | * the pages and send SIGBUS to the processes if the data was dirty. | |
916 | */ | |
917 | static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, | |
918 | int trapno, int flags, struct page **hpagep) | |
919 | { | |
920 | enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
921 | struct address_space *mapping; | |
922 | LIST_HEAD(tokill); | |
923 | bool unmap_success; | |
924 | int kill = 1, forcekill; | |
925 | struct page *hpage = *hpagep; | |
926 | bool mlocked = PageMlocked(hpage); | |
927 | ||
928 | /* | |
929 | * Here we are interested only in user-mapped pages, so skip any | |
930 | * other types of pages. | |
931 | */ | |
932 | if (PageReserved(p) || PageSlab(p)) | |
933 | return true; | |
934 | if (!(PageLRU(hpage) || PageHuge(p))) | |
935 | return true; | |
936 | ||
937 | /* | |
938 | * This check implies we don't kill processes if their pages | |
939 | * are in the swap cache early. Those are always late kills. | |
940 | */ | |
941 | if (!page_mapped(hpage)) | |
942 | return true; | |
943 | ||
944 | if (PageKsm(p)) { | |
945 | pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn); | |
946 | return false; | |
947 | } | |
948 | ||
949 | if (PageSwapCache(p)) { | |
950 | pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n", | |
951 | pfn); | |
952 | ttu |= TTU_IGNORE_HWPOISON; | |
953 | } | |
954 | ||
955 | /* | |
956 | * Propagate the dirty bit from PTEs to struct page first, because we | |
957 | * need this to decide if we should kill or just drop the page. | |
958 | * XXX: the dirty test could be racy: set_page_dirty() may not always | |
959 | * be called inside page lock (it's recommended but not enforced). | |
960 | */ | |
961 | mapping = page_mapping(hpage); | |
962 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && | |
963 | mapping_cap_writeback_dirty(mapping)) { | |
964 | if (page_mkclean(hpage)) { | |
965 | SetPageDirty(hpage); | |
966 | } else { | |
967 | kill = 0; | |
968 | ttu |= TTU_IGNORE_HWPOISON; | |
969 | pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n", | |
970 | pfn); | |
971 | } | |
972 | } | |
973 | ||
974 | /* | |
975 | * First collect all the processes that have the page | |
976 | * mapped in dirty form. This has to be done before try_to_unmap, | |
977 | * because ttu takes the rmap data structures down. | |
978 | * | |
979 | * Error handling: We ignore errors here because | |
980 | * there's nothing that can be done. | |
981 | */ | |
982 | if (kill) | |
983 | collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); | |
984 | ||
985 | unmap_success = try_to_unmap(hpage, ttu); | |
986 | if (!unmap_success) | |
987 | pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n", | |
988 | pfn, page_mapcount(hpage)); | |
989 | ||
990 | /* | |
991 | * try_to_unmap() might put mlocked page in lru cache, so call | |
992 | * shake_page() again to ensure that it's flushed. | |
993 | */ | |
994 | if (mlocked) | |
995 | shake_page(hpage, 0); | |
996 | ||
997 | /* | |
998 | * Now that the dirty bit has been propagated to the | |
999 | * struct page and all unmaps done we can decide if | |
1000 | * killing is needed or not. Only kill when the page | |
1001 | * was dirty or the process is not restartable, | |
1002 | * otherwise the tokill list is merely | |
1003 | * freed. When there was a problem unmapping earlier | |
1004 | * use a more force-full uncatchable kill to prevent | |
1005 | * any accesses to the poisoned memory. | |
1006 | */ | |
1007 | forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL); | |
1008 | kill_procs(&tokill, forcekill, trapno, !unmap_success, p, pfn, flags); | |
1009 | ||
1010 | return unmap_success; | |
1011 | } | |
1012 | ||
1013 | static void set_page_hwpoison_huge_page(struct page *hpage) | |
1014 | { | |
1015 | int i; | |
1016 | int nr_pages = 1 << compound_order(hpage); | |
1017 | for (i = 0; i < nr_pages; i++) | |
1018 | SetPageHWPoison(hpage + i); | |
1019 | } | |
1020 | ||
1021 | static void clear_page_hwpoison_huge_page(struct page *hpage) | |
1022 | { | |
1023 | int i; | |
1024 | int nr_pages = 1 << compound_order(hpage); | |
1025 | for (i = 0; i < nr_pages; i++) | |
1026 | ClearPageHWPoison(hpage + i); | |
1027 | } | |
1028 | ||
1029 | /** | |
1030 | * memory_failure - Handle memory failure of a page. | |
1031 | * @pfn: Page Number of the corrupted page | |
1032 | * @trapno: Trap number reported in the signal to user space. | |
1033 | * @flags: fine tune action taken | |
1034 | * | |
1035 | * This function is called by the low level machine check code | |
1036 | * of an architecture when it detects hardware memory corruption | |
1037 | * of a page. It tries its best to recover, which includes | |
1038 | * dropping pages, killing processes etc. | |
1039 | * | |
1040 | * The function is primarily of use for corruptions that | |
1041 | * happen outside the current execution context (e.g. when | |
1042 | * detected by a background scrubber) | |
1043 | * | |
1044 | * Must run in process context (e.g. a work queue) with interrupts | |
1045 | * enabled and no spinlocks hold. | |
1046 | */ | |
1047 | int memory_failure(unsigned long pfn, int trapno, int flags) | |
1048 | { | |
1049 | struct page_state *ps; | |
1050 | struct page *p; | |
1051 | struct page *hpage; | |
1052 | struct page *orig_head; | |
1053 | int res; | |
1054 | unsigned int nr_pages; | |
1055 | unsigned long page_flags; | |
1056 | ||
1057 | if (!sysctl_memory_failure_recovery) | |
1058 | panic("Memory failure from trap %d on page %lx", trapno, pfn); | |
1059 | ||
1060 | if (!pfn_valid(pfn)) { | |
1061 | pr_err("Memory failure: %#lx: memory outside kernel control\n", | |
1062 | pfn); | |
1063 | return -ENXIO; | |
1064 | } | |
1065 | ||
1066 | p = pfn_to_page(pfn); | |
1067 | orig_head = hpage = compound_head(p); | |
1068 | if (TestSetPageHWPoison(p)) { | |
1069 | pr_err("Memory failure: %#lx: already hardware poisoned\n", | |
1070 | pfn); | |
1071 | return 0; | |
1072 | } | |
1073 | ||
1074 | /* | |
1075 | * Currently errors on hugetlbfs pages are measured in hugepage units, | |
1076 | * so nr_pages should be 1 << compound_order. OTOH when errors are on | |
1077 | * transparent hugepages, they are supposed to be split and error | |
1078 | * measurement is done in normal page units. So nr_pages should be one | |
1079 | * in this case. | |
1080 | */ | |
1081 | if (PageHuge(p)) | |
1082 | nr_pages = 1 << compound_order(hpage); | |
1083 | else /* normal page or thp */ | |
1084 | nr_pages = 1; | |
1085 | num_poisoned_pages_add(nr_pages); | |
1086 | ||
1087 | /* | |
1088 | * We need/can do nothing about count=0 pages. | |
1089 | * 1) it's a free page, and therefore in safe hand: | |
1090 | * prep_new_page() will be the gate keeper. | |
1091 | * 2) it's a free hugepage, which is also safe: | |
1092 | * an affected hugepage will be dequeued from hugepage freelist, | |
1093 | * so there's no concern about reusing it ever after. | |
1094 | * 3) it's part of a non-compound high order page. | |
1095 | * Implies some kernel user: cannot stop them from | |
1096 | * R/W the page; let's pray that the page has been | |
1097 | * used and will be freed some time later. | |
1098 | * In fact it's dangerous to directly bump up page count from 0, | |
1099 | * that may make page_freeze_refs()/page_unfreeze_refs() mismatch. | |
1100 | */ | |
1101 | if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) { | |
1102 | if (is_free_buddy_page(p)) { | |
1103 | action_result(pfn, MF_MSG_BUDDY, MF_DELAYED); | |
1104 | return 0; | |
1105 | } else if (PageHuge(hpage)) { | |
1106 | /* | |
1107 | * Check "filter hit" and "race with other subpage." | |
1108 | */ | |
1109 | lock_page(hpage); | |
1110 | if (PageHWPoison(hpage)) { | |
1111 | if ((hwpoison_filter(p) && TestClearPageHWPoison(p)) | |
1112 | || (p != hpage && TestSetPageHWPoison(hpage))) { | |
1113 | num_poisoned_pages_sub(nr_pages); | |
1114 | unlock_page(hpage); | |
1115 | return 0; | |
1116 | } | |
1117 | } | |
1118 | set_page_hwpoison_huge_page(hpage); | |
1119 | res = dequeue_hwpoisoned_huge_page(hpage); | |
1120 | action_result(pfn, MF_MSG_FREE_HUGE, | |
1121 | res ? MF_IGNORED : MF_DELAYED); | |
1122 | unlock_page(hpage); | |
1123 | return res; | |
1124 | } else { | |
1125 | action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); | |
1126 | return -EBUSY; | |
1127 | } | |
1128 | } | |
1129 | ||
1130 | if (!PageHuge(p) && PageTransHuge(hpage)) { | |
1131 | lock_page(p); | |
1132 | if (!PageAnon(p) || unlikely(split_huge_page(p))) { | |
1133 | unlock_page(p); | |
1134 | if (!PageAnon(p)) | |
1135 | pr_err("Memory failure: %#lx: non anonymous thp\n", | |
1136 | pfn); | |
1137 | else | |
1138 | pr_err("Memory failure: %#lx: thp split failed\n", | |
1139 | pfn); | |
1140 | if (TestClearPageHWPoison(p)) | |
1141 | num_poisoned_pages_sub(nr_pages); | |
1142 | put_hwpoison_page(p); | |
1143 | return -EBUSY; | |
1144 | } | |
1145 | unlock_page(p); | |
1146 | VM_BUG_ON_PAGE(!page_count(p), p); | |
1147 | hpage = compound_head(p); | |
1148 | } | |
1149 | ||
1150 | /* | |
1151 | * We ignore non-LRU pages for good reasons. | |
1152 | * - PG_locked is only well defined for LRU pages and a few others | |
1153 | * - to avoid races with __SetPageLocked() | |
1154 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) | |
1155 | * The check (unnecessarily) ignores LRU pages being isolated and | |
1156 | * walked by the page reclaim code, however that's not a big loss. | |
1157 | */ | |
1158 | shake_page(p, 0); | |
1159 | /* shake_page could have turned it free. */ | |
1160 | if (!PageLRU(p) && is_free_buddy_page(p)) { | |
1161 | if (flags & MF_COUNT_INCREASED) | |
1162 | action_result(pfn, MF_MSG_BUDDY, MF_DELAYED); | |
1163 | else | |
1164 | action_result(pfn, MF_MSG_BUDDY_2ND, MF_DELAYED); | |
1165 | return 0; | |
1166 | } | |
1167 | ||
1168 | lock_page(hpage); | |
1169 | ||
1170 | /* | |
1171 | * The page could have changed compound pages during the locking. | |
1172 | * If this happens just bail out. | |
1173 | */ | |
1174 | if (PageCompound(p) && compound_head(p) != orig_head) { | |
1175 | action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); | |
1176 | res = -EBUSY; | |
1177 | goto out; | |
1178 | } | |
1179 | ||
1180 | /* | |
1181 | * We use page flags to determine what action should be taken, but | |
1182 | * the flags can be modified by the error containment action. One | |
1183 | * example is an mlocked page, where PG_mlocked is cleared by | |
1184 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status | |
1185 | * correctly, we save a copy of the page flags at this time. | |
1186 | */ | |
1187 | page_flags = p->flags; | |
1188 | ||
1189 | /* | |
1190 | * unpoison always clear PG_hwpoison inside page lock | |
1191 | */ | |
1192 | if (!PageHWPoison(p)) { | |
1193 | pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); | |
1194 | num_poisoned_pages_sub(nr_pages); | |
1195 | unlock_page(hpage); | |
1196 | put_hwpoison_page(hpage); | |
1197 | return 0; | |
1198 | } | |
1199 | if (hwpoison_filter(p)) { | |
1200 | if (TestClearPageHWPoison(p)) | |
1201 | num_poisoned_pages_sub(nr_pages); | |
1202 | unlock_page(hpage); | |
1203 | put_hwpoison_page(hpage); | |
1204 | return 0; | |
1205 | } | |
1206 | ||
1207 | if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p)) | |
1208 | goto identify_page_state; | |
1209 | ||
1210 | /* | |
1211 | * For error on the tail page, we should set PG_hwpoison | |
1212 | * on the head page to show that the hugepage is hwpoisoned | |
1213 | */ | |
1214 | if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) { | |
1215 | action_result(pfn, MF_MSG_POISONED_HUGE, MF_IGNORED); | |
1216 | unlock_page(hpage); | |
1217 | put_hwpoison_page(hpage); | |
1218 | return 0; | |
1219 | } | |
1220 | /* | |
1221 | * Set PG_hwpoison on all pages in an error hugepage, | |
1222 | * because containment is done in hugepage unit for now. | |
1223 | * Since we have done TestSetPageHWPoison() for the head page with | |
1224 | * page lock held, we can safely set PG_hwpoison bits on tail pages. | |
1225 | */ | |
1226 | if (PageHuge(p)) | |
1227 | set_page_hwpoison_huge_page(hpage); | |
1228 | ||
1229 | /* | |
1230 | * It's very difficult to mess with pages currently under IO | |
1231 | * and in many cases impossible, so we just avoid it here. | |
1232 | */ | |
1233 | wait_on_page_writeback(p); | |
1234 | ||
1235 | /* | |
1236 | * Now take care of user space mappings. | |
1237 | * Abort on fail: __delete_from_page_cache() assumes unmapped page. | |
1238 | * | |
1239 | * When the raw error page is thp tail page, hpage points to the raw | |
1240 | * page after thp split. | |
1241 | */ | |
1242 | if (!hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)) { | |
1243 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); | |
1244 | res = -EBUSY; | |
1245 | goto out; | |
1246 | } | |
1247 | ||
1248 | /* | |
1249 | * Torn down by someone else? | |
1250 | */ | |
1251 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { | |
1252 | action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); | |
1253 | res = -EBUSY; | |
1254 | goto out; | |
1255 | } | |
1256 | ||
1257 | identify_page_state: | |
1258 | res = -EBUSY; | |
1259 | /* | |
1260 | * The first check uses the current page flags which may not have any | |
1261 | * relevant information. The second check with the saved page flagss is | |
1262 | * carried out only if the first check can't determine the page status. | |
1263 | */ | |
1264 | for (ps = error_states;; ps++) | |
1265 | if ((p->flags & ps->mask) == ps->res) | |
1266 | break; | |
1267 | ||
1268 | page_flags |= (p->flags & (1UL << PG_dirty)); | |
1269 | ||
1270 | if (!ps->mask) | |
1271 | for (ps = error_states;; ps++) | |
1272 | if ((page_flags & ps->mask) == ps->res) | |
1273 | break; | |
1274 | res = page_action(ps, p, pfn); | |
1275 | out: | |
1276 | unlock_page(hpage); | |
1277 | return res; | |
1278 | } | |
1279 | EXPORT_SYMBOL_GPL(memory_failure); | |
1280 | ||
1281 | #define MEMORY_FAILURE_FIFO_ORDER 4 | |
1282 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
1283 | ||
1284 | struct memory_failure_entry { | |
1285 | unsigned long pfn; | |
1286 | int trapno; | |
1287 | int flags; | |
1288 | }; | |
1289 | ||
1290 | struct memory_failure_cpu { | |
1291 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
1292 | MEMORY_FAILURE_FIFO_SIZE); | |
1293 | spinlock_t lock; | |
1294 | struct work_struct work; | |
1295 | }; | |
1296 | ||
1297 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
1298 | ||
1299 | /** | |
1300 | * memory_failure_queue - Schedule handling memory failure of a page. | |
1301 | * @pfn: Page Number of the corrupted page | |
1302 | * @trapno: Trap number reported in the signal to user space. | |
1303 | * @flags: Flags for memory failure handling | |
1304 | * | |
1305 | * This function is called by the low level hardware error handler | |
1306 | * when it detects hardware memory corruption of a page. It schedules | |
1307 | * the recovering of error page, including dropping pages, killing | |
1308 | * processes etc. | |
1309 | * | |
1310 | * The function is primarily of use for corruptions that | |
1311 | * happen outside the current execution context (e.g. when | |
1312 | * detected by a background scrubber) | |
1313 | * | |
1314 | * Can run in IRQ context. | |
1315 | */ | |
1316 | void memory_failure_queue(unsigned long pfn, int trapno, int flags) | |
1317 | { | |
1318 | struct memory_failure_cpu *mf_cpu; | |
1319 | unsigned long proc_flags; | |
1320 | struct memory_failure_entry entry = { | |
1321 | .pfn = pfn, | |
1322 | .trapno = trapno, | |
1323 | .flags = flags, | |
1324 | }; | |
1325 | ||
1326 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
1327 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
1328 | if (kfifo_put(&mf_cpu->fifo, entry)) | |
1329 | schedule_work_on(smp_processor_id(), &mf_cpu->work); | |
1330 | else | |
1331 | pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n", | |
1332 | pfn); | |
1333 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1334 | put_cpu_var(memory_failure_cpu); | |
1335 | } | |
1336 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
1337 | ||
1338 | static void memory_failure_work_func(struct work_struct *work) | |
1339 | { | |
1340 | struct memory_failure_cpu *mf_cpu; | |
1341 | struct memory_failure_entry entry = { 0, }; | |
1342 | unsigned long proc_flags; | |
1343 | int gotten; | |
1344 | ||
1345 | mf_cpu = this_cpu_ptr(&memory_failure_cpu); | |
1346 | for (;;) { | |
1347 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
1348 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
1349 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1350 | if (!gotten) | |
1351 | break; | |
1352 | if (entry.flags & MF_SOFT_OFFLINE) | |
1353 | soft_offline_page(pfn_to_page(entry.pfn), entry.flags); | |
1354 | else | |
1355 | memory_failure(entry.pfn, entry.trapno, entry.flags); | |
1356 | } | |
1357 | } | |
1358 | ||
1359 | static int __init memory_failure_init(void) | |
1360 | { | |
1361 | struct memory_failure_cpu *mf_cpu; | |
1362 | int cpu; | |
1363 | ||
1364 | for_each_possible_cpu(cpu) { | |
1365 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1366 | spin_lock_init(&mf_cpu->lock); | |
1367 | INIT_KFIFO(mf_cpu->fifo); | |
1368 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
1369 | } | |
1370 | ||
1371 | return 0; | |
1372 | } | |
1373 | core_initcall(memory_failure_init); | |
1374 | ||
1375 | #define unpoison_pr_info(fmt, pfn, rs) \ | |
1376 | ({ \ | |
1377 | if (__ratelimit(rs)) \ | |
1378 | pr_info(fmt, pfn); \ | |
1379 | }) | |
1380 | ||
1381 | /** | |
1382 | * unpoison_memory - Unpoison a previously poisoned page | |
1383 | * @pfn: Page number of the to be unpoisoned page | |
1384 | * | |
1385 | * Software-unpoison a page that has been poisoned by | |
1386 | * memory_failure() earlier. | |
1387 | * | |
1388 | * This is only done on the software-level, so it only works | |
1389 | * for linux injected failures, not real hardware failures | |
1390 | * | |
1391 | * Returns 0 for success, otherwise -errno. | |
1392 | */ | |
1393 | int unpoison_memory(unsigned long pfn) | |
1394 | { | |
1395 | struct page *page; | |
1396 | struct page *p; | |
1397 | int freeit = 0; | |
1398 | unsigned int nr_pages; | |
1399 | static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, | |
1400 | DEFAULT_RATELIMIT_BURST); | |
1401 | ||
1402 | if (!pfn_valid(pfn)) | |
1403 | return -ENXIO; | |
1404 | ||
1405 | p = pfn_to_page(pfn); | |
1406 | page = compound_head(p); | |
1407 | ||
1408 | if (!PageHWPoison(p)) { | |
1409 | unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n", | |
1410 | pfn, &unpoison_rs); | |
1411 | return 0; | |
1412 | } | |
1413 | ||
1414 | if (page_count(page) > 1) { | |
1415 | unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n", | |
1416 | pfn, &unpoison_rs); | |
1417 | return 0; | |
1418 | } | |
1419 | ||
1420 | if (page_mapped(page)) { | |
1421 | unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n", | |
1422 | pfn, &unpoison_rs); | |
1423 | return 0; | |
1424 | } | |
1425 | ||
1426 | if (page_mapping(page)) { | |
1427 | unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n", | |
1428 | pfn, &unpoison_rs); | |
1429 | return 0; | |
1430 | } | |
1431 | ||
1432 | /* | |
1433 | * unpoison_memory() can encounter thp only when the thp is being | |
1434 | * worked by memory_failure() and the page lock is not held yet. | |
1435 | * In such case, we yield to memory_failure() and make unpoison fail. | |
1436 | */ | |
1437 | if (!PageHuge(page) && PageTransHuge(page)) { | |
1438 | unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n", | |
1439 | pfn, &unpoison_rs); | |
1440 | return 0; | |
1441 | } | |
1442 | ||
1443 | nr_pages = 1 << compound_order(page); | |
1444 | ||
1445 | if (!get_hwpoison_page(p)) { | |
1446 | /* | |
1447 | * Since HWPoisoned hugepage should have non-zero refcount, | |
1448 | * race between memory failure and unpoison seems to happen. | |
1449 | * In such case unpoison fails and memory failure runs | |
1450 | * to the end. | |
1451 | */ | |
1452 | if (PageHuge(page)) { | |
1453 | unpoison_pr_info("Unpoison: Memory failure is now running on free hugepage %#lx\n", | |
1454 | pfn, &unpoison_rs); | |
1455 | return 0; | |
1456 | } | |
1457 | if (TestClearPageHWPoison(p)) | |
1458 | num_poisoned_pages_dec(); | |
1459 | unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n", | |
1460 | pfn, &unpoison_rs); | |
1461 | return 0; | |
1462 | } | |
1463 | ||
1464 | lock_page(page); | |
1465 | /* | |
1466 | * This test is racy because PG_hwpoison is set outside of page lock. | |
1467 | * That's acceptable because that won't trigger kernel panic. Instead, | |
1468 | * the PG_hwpoison page will be caught and isolated on the entrance to | |
1469 | * the free buddy page pool. | |
1470 | */ | |
1471 | if (TestClearPageHWPoison(page)) { | |
1472 | unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n", | |
1473 | pfn, &unpoison_rs); | |
1474 | num_poisoned_pages_sub(nr_pages); | |
1475 | freeit = 1; | |
1476 | if (PageHuge(page)) | |
1477 | clear_page_hwpoison_huge_page(page); | |
1478 | } | |
1479 | unlock_page(page); | |
1480 | ||
1481 | put_hwpoison_page(page); | |
1482 | if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1)) | |
1483 | put_hwpoison_page(page); | |
1484 | ||
1485 | return 0; | |
1486 | } | |
1487 | EXPORT_SYMBOL(unpoison_memory); | |
1488 | ||
1489 | static struct page *new_page(struct page *p, unsigned long private, int **x) | |
1490 | { | |
1491 | int nid = page_to_nid(p); | |
1492 | if (PageHuge(p)) | |
1493 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1494 | nid); | |
1495 | else | |
1496 | return __alloc_pages_node(nid, GFP_HIGHUSER_MOVABLE, 0); | |
1497 | } | |
1498 | ||
1499 | /* | |
1500 | * Safely get reference count of an arbitrary page. | |
1501 | * Returns 0 for a free page, -EIO for a zero refcount page | |
1502 | * that is not free, and 1 for any other page type. | |
1503 | * For 1 the page is returned with increased page count, otherwise not. | |
1504 | */ | |
1505 | static int __get_any_page(struct page *p, unsigned long pfn, int flags) | |
1506 | { | |
1507 | int ret; | |
1508 | ||
1509 | if (flags & MF_COUNT_INCREASED) | |
1510 | return 1; | |
1511 | ||
1512 | /* | |
1513 | * When the target page is a free hugepage, just remove it | |
1514 | * from free hugepage list. | |
1515 | */ | |
1516 | if (!get_hwpoison_page(p)) { | |
1517 | if (PageHuge(p)) { | |
1518 | pr_info("%s: %#lx free huge page\n", __func__, pfn); | |
1519 | ret = 0; | |
1520 | } else if (is_free_buddy_page(p)) { | |
1521 | pr_info("%s: %#lx free buddy page\n", __func__, pfn); | |
1522 | ret = 0; | |
1523 | } else { | |
1524 | pr_info("%s: %#lx: unknown zero refcount page type %lx\n", | |
1525 | __func__, pfn, p->flags); | |
1526 | ret = -EIO; | |
1527 | } | |
1528 | } else { | |
1529 | /* Not a free page */ | |
1530 | ret = 1; | |
1531 | } | |
1532 | return ret; | |
1533 | } | |
1534 | ||
1535 | static int get_any_page(struct page *page, unsigned long pfn, int flags) | |
1536 | { | |
1537 | int ret = __get_any_page(page, pfn, flags); | |
1538 | ||
1539 | if (ret == 1 && !PageHuge(page) && | |
1540 | !PageLRU(page) && !__PageMovable(page)) { | |
1541 | /* | |
1542 | * Try to free it. | |
1543 | */ | |
1544 | put_hwpoison_page(page); | |
1545 | shake_page(page, 1); | |
1546 | ||
1547 | /* | |
1548 | * Did it turn free? | |
1549 | */ | |
1550 | ret = __get_any_page(page, pfn, 0); | |
1551 | if (ret == 1 && !PageLRU(page)) { | |
1552 | /* Drop page reference which is from __get_any_page() */ | |
1553 | put_hwpoison_page(page); | |
1554 | pr_info("soft_offline: %#lx: unknown non LRU page type %lx (%pGp)\n", | |
1555 | pfn, page->flags, &page->flags); | |
1556 | return -EIO; | |
1557 | } | |
1558 | } | |
1559 | return ret; | |
1560 | } | |
1561 | ||
1562 | static int soft_offline_huge_page(struct page *page, int flags) | |
1563 | { | |
1564 | int ret; | |
1565 | unsigned long pfn = page_to_pfn(page); | |
1566 | struct page *hpage = compound_head(page); | |
1567 | LIST_HEAD(pagelist); | |
1568 | ||
1569 | /* | |
1570 | * This double-check of PageHWPoison is to avoid the race with | |
1571 | * memory_failure(). See also comment in __soft_offline_page(). | |
1572 | */ | |
1573 | lock_page(hpage); | |
1574 | if (PageHWPoison(hpage)) { | |
1575 | unlock_page(hpage); | |
1576 | put_hwpoison_page(hpage); | |
1577 | pr_info("soft offline: %#lx hugepage already poisoned\n", pfn); | |
1578 | return -EBUSY; | |
1579 | } | |
1580 | unlock_page(hpage); | |
1581 | ||
1582 | ret = isolate_huge_page(hpage, &pagelist); | |
1583 | /* | |
1584 | * get_any_page() and isolate_huge_page() takes a refcount each, | |
1585 | * so need to drop one here. | |
1586 | */ | |
1587 | put_hwpoison_page(hpage); | |
1588 | if (!ret) { | |
1589 | pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn); | |
1590 | return -EBUSY; | |
1591 | } | |
1592 | ||
1593 | ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL, | |
1594 | MIGRATE_SYNC, MR_MEMORY_FAILURE); | |
1595 | if (ret) { | |
1596 | pr_info("soft offline: %#lx: migration failed %d, type %lx (%pGp)\n", | |
1597 | pfn, ret, page->flags, &page->flags); | |
1598 | /* | |
1599 | * We know that soft_offline_huge_page() tries to migrate | |
1600 | * only one hugepage pointed to by hpage, so we need not | |
1601 | * run through the pagelist here. | |
1602 | */ | |
1603 | putback_active_hugepage(hpage); | |
1604 | if (ret > 0) | |
1605 | ret = -EIO; | |
1606 | } else { | |
1607 | /* overcommit hugetlb page will be freed to buddy */ | |
1608 | if (PageHuge(page)) { | |
1609 | set_page_hwpoison_huge_page(hpage); | |
1610 | dequeue_hwpoisoned_huge_page(hpage); | |
1611 | num_poisoned_pages_add(1 << compound_order(hpage)); | |
1612 | } else { | |
1613 | SetPageHWPoison(page); | |
1614 | num_poisoned_pages_inc(); | |
1615 | } | |
1616 | } | |
1617 | return ret; | |
1618 | } | |
1619 | ||
1620 | static int __soft_offline_page(struct page *page, int flags) | |
1621 | { | |
1622 | int ret; | |
1623 | unsigned long pfn = page_to_pfn(page); | |
1624 | ||
1625 | /* | |
1626 | * Check PageHWPoison again inside page lock because PageHWPoison | |
1627 | * is set by memory_failure() outside page lock. Note that | |
1628 | * memory_failure() also double-checks PageHWPoison inside page lock, | |
1629 | * so there's no race between soft_offline_page() and memory_failure(). | |
1630 | */ | |
1631 | lock_page(page); | |
1632 | wait_on_page_writeback(page); | |
1633 | if (PageHWPoison(page)) { | |
1634 | unlock_page(page); | |
1635 | put_hwpoison_page(page); | |
1636 | pr_info("soft offline: %#lx page already poisoned\n", pfn); | |
1637 | return -EBUSY; | |
1638 | } | |
1639 | /* | |
1640 | * Try to invalidate first. This should work for | |
1641 | * non dirty unmapped page cache pages. | |
1642 | */ | |
1643 | ret = invalidate_inode_page(page); | |
1644 | unlock_page(page); | |
1645 | /* | |
1646 | * RED-PEN would be better to keep it isolated here, but we | |
1647 | * would need to fix isolation locking first. | |
1648 | */ | |
1649 | if (ret == 1) { | |
1650 | put_hwpoison_page(page); | |
1651 | pr_info("soft_offline: %#lx: invalidated\n", pfn); | |
1652 | SetPageHWPoison(page); | |
1653 | num_poisoned_pages_inc(); | |
1654 | return 0; | |
1655 | } | |
1656 | ||
1657 | /* | |
1658 | * Simple invalidation didn't work. | |
1659 | * Try to migrate to a new page instead. migrate.c | |
1660 | * handles a large number of cases for us. | |
1661 | */ | |
1662 | if (PageLRU(page)) | |
1663 | ret = isolate_lru_page(page); | |
1664 | else | |
1665 | ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE); | |
1666 | /* | |
1667 | * Drop page reference which is came from get_any_page() | |
1668 | * successful isolate_lru_page() already took another one. | |
1669 | */ | |
1670 | put_hwpoison_page(page); | |
1671 | if (!ret) { | |
1672 | LIST_HEAD(pagelist); | |
1673 | /* | |
1674 | * After isolated lru page, the PageLRU will be cleared, | |
1675 | * so use !__PageMovable instead for LRU page's mapping | |
1676 | * cannot have PAGE_MAPPING_MOVABLE. | |
1677 | */ | |
1678 | if (!__PageMovable(page)) | |
1679 | inc_node_page_state(page, NR_ISOLATED_ANON + | |
1680 | page_is_file_cache(page)); | |
1681 | list_add(&page->lru, &pagelist); | |
1682 | ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL, | |
1683 | MIGRATE_SYNC, MR_MEMORY_FAILURE); | |
1684 | if (ret) { | |
1685 | if (!list_empty(&pagelist)) | |
1686 | putback_movable_pages(&pagelist); | |
1687 | ||
1688 | pr_info("soft offline: %#lx: migration failed %d, type %lx (%pGp)\n", | |
1689 | pfn, ret, page->flags, &page->flags); | |
1690 | if (ret > 0) | |
1691 | ret = -EIO; | |
1692 | } | |
1693 | } else { | |
1694 | pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx (%pGp)\n", | |
1695 | pfn, ret, page_count(page), page->flags, &page->flags); | |
1696 | } | |
1697 | return ret; | |
1698 | } | |
1699 | ||
1700 | static int soft_offline_in_use_page(struct page *page, int flags) | |
1701 | { | |
1702 | int ret; | |
1703 | struct page *hpage = compound_head(page); | |
1704 | ||
1705 | if (!PageHuge(page) && PageTransHuge(hpage)) { | |
1706 | lock_page(hpage); | |
1707 | if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) { | |
1708 | unlock_page(hpage); | |
1709 | if (!PageAnon(hpage)) | |
1710 | pr_info("soft offline: %#lx: non anonymous thp\n", page_to_pfn(page)); | |
1711 | else | |
1712 | pr_info("soft offline: %#lx: thp split failed\n", page_to_pfn(page)); | |
1713 | put_hwpoison_page(hpage); | |
1714 | return -EBUSY; | |
1715 | } | |
1716 | unlock_page(hpage); | |
1717 | get_hwpoison_page(page); | |
1718 | put_hwpoison_page(hpage); | |
1719 | } | |
1720 | ||
1721 | if (PageHuge(page)) | |
1722 | ret = soft_offline_huge_page(page, flags); | |
1723 | else | |
1724 | ret = __soft_offline_page(page, flags); | |
1725 | ||
1726 | return ret; | |
1727 | } | |
1728 | ||
1729 | static void soft_offline_free_page(struct page *page) | |
1730 | { | |
1731 | if (PageHuge(page)) { | |
1732 | struct page *hpage = compound_head(page); | |
1733 | ||
1734 | set_page_hwpoison_huge_page(hpage); | |
1735 | if (!dequeue_hwpoisoned_huge_page(hpage)) | |
1736 | num_poisoned_pages_add(1 << compound_order(hpage)); | |
1737 | } else { | |
1738 | if (!TestSetPageHWPoison(page)) | |
1739 | num_poisoned_pages_inc(); | |
1740 | } | |
1741 | } | |
1742 | ||
1743 | /** | |
1744 | * soft_offline_page - Soft offline a page. | |
1745 | * @page: page to offline | |
1746 | * @flags: flags. Same as memory_failure(). | |
1747 | * | |
1748 | * Returns 0 on success, otherwise negated errno. | |
1749 | * | |
1750 | * Soft offline a page, by migration or invalidation, | |
1751 | * without killing anything. This is for the case when | |
1752 | * a page is not corrupted yet (so it's still valid to access), | |
1753 | * but has had a number of corrected errors and is better taken | |
1754 | * out. | |
1755 | * | |
1756 | * The actual policy on when to do that is maintained by | |
1757 | * user space. | |
1758 | * | |
1759 | * This should never impact any application or cause data loss, | |
1760 | * however it might take some time. | |
1761 | * | |
1762 | * This is not a 100% solution for all memory, but tries to be | |
1763 | * ``good enough'' for the majority of memory. | |
1764 | */ | |
1765 | int soft_offline_page(struct page *page, int flags) | |
1766 | { | |
1767 | int ret; | |
1768 | unsigned long pfn = page_to_pfn(page); | |
1769 | ||
1770 | if (PageHWPoison(page)) { | |
1771 | pr_info("soft offline: %#lx page already poisoned\n", pfn); | |
1772 | if (flags & MF_COUNT_INCREASED) | |
1773 | put_hwpoison_page(page); | |
1774 | return -EBUSY; | |
1775 | } | |
1776 | ||
1777 | get_online_mems(); | |
1778 | ret = get_any_page(page, pfn, flags); | |
1779 | put_online_mems(); | |
1780 | ||
1781 | if (ret > 0) | |
1782 | ret = soft_offline_in_use_page(page, flags); | |
1783 | else if (ret == 0) | |
1784 | soft_offline_free_page(page); | |
1785 | ||
1786 | return ret; | |
1787 | } |