tracing: extend sched_pi_setprio
[deliverable/linux.git] / mm / memory-failure.c
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.h>
44 #include <linux/ksm.h>
45 #include <linux/rmap.h>
46 #include <linux/export.h>
47 #include <linux/pagemap.h>
48 #include <linux/swap.h>
49 #include <linux/backing-dev.h>
50 #include <linux/migrate.h>
51 #include <linux/page-isolation.h>
52 #include <linux/suspend.h>
53 #include <linux/slab.h>
54 #include <linux/swapops.h>
55 #include <linux/hugetlb.h>
56 #include <linux/memory_hotplug.h>
57 #include <linux/mm_inline.h>
58 #include <linux/kfifo.h>
59 #include <linux/ratelimit.h>
60 #include "internal.h"
61 #include "ras/ras_event.h"
62
63 int sysctl_memory_failure_early_kill __read_mostly = 0;
64
65 int sysctl_memory_failure_recovery __read_mostly = 1;
66
67 atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
68
69 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
70
71 u32 hwpoison_filter_enable = 0;
72 u32 hwpoison_filter_dev_major = ~0U;
73 u32 hwpoison_filter_dev_minor = ~0U;
74 u64 hwpoison_filter_flags_mask;
75 u64 hwpoison_filter_flags_value;
76 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
77 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
78 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
79 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
80 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
81
82 static int hwpoison_filter_dev(struct page *p)
83 {
84 struct address_space *mapping;
85 dev_t dev;
86
87 if (hwpoison_filter_dev_major == ~0U &&
88 hwpoison_filter_dev_minor == ~0U)
89 return 0;
90
91 /*
92 * page_mapping() does not accept slab pages.
93 */
94 if (PageSlab(p))
95 return -EINVAL;
96
97 mapping = page_mapping(p);
98 if (mapping == NULL || mapping->host == NULL)
99 return -EINVAL;
100
101 dev = mapping->host->i_sb->s_dev;
102 if (hwpoison_filter_dev_major != ~0U &&
103 hwpoison_filter_dev_major != MAJOR(dev))
104 return -EINVAL;
105 if (hwpoison_filter_dev_minor != ~0U &&
106 hwpoison_filter_dev_minor != MINOR(dev))
107 return -EINVAL;
108
109 return 0;
110 }
111
112 static int hwpoison_filter_flags(struct page *p)
113 {
114 if (!hwpoison_filter_flags_mask)
115 return 0;
116
117 if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
118 hwpoison_filter_flags_value)
119 return 0;
120 else
121 return -EINVAL;
122 }
123
124 /*
125 * This allows stress tests to limit test scope to a collection of tasks
126 * by putting them under some memcg. This prevents killing unrelated/important
127 * processes such as /sbin/init. Note that the target task may share clean
128 * pages with init (eg. libc text), which is harmless. If the target task
129 * share _dirty_ pages with another task B, the test scheme must make sure B
130 * is also included in the memcg. At last, due to race conditions this filter
131 * can only guarantee that the page either belongs to the memcg tasks, or is
132 * a freed page.
133 */
134 #ifdef CONFIG_MEMCG
135 u64 hwpoison_filter_memcg;
136 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
137 static int hwpoison_filter_task(struct page *p)
138 {
139 if (!hwpoison_filter_memcg)
140 return 0;
141
142 if (page_cgroup_ino(p) != hwpoison_filter_memcg)
143 return -EINVAL;
144
145 return 0;
146 }
147 #else
148 static int hwpoison_filter_task(struct page *p) { return 0; }
149 #endif
150
151 int hwpoison_filter(struct page *p)
152 {
153 if (!hwpoison_filter_enable)
154 return 0;
155
156 if (hwpoison_filter_dev(p))
157 return -EINVAL;
158
159 if (hwpoison_filter_flags(p))
160 return -EINVAL;
161
162 if (hwpoison_filter_task(p))
163 return -EINVAL;
164
165 return 0;
166 }
167 #else
168 int hwpoison_filter(struct page *p)
169 {
170 return 0;
171 }
172 #endif
173
174 EXPORT_SYMBOL_GPL(hwpoison_filter);
175
176 /*
177 * Send all the processes who have the page mapped a signal.
178 * ``action optional'' if they are not immediately affected by the error
179 * ``action required'' if error happened in current execution context
180 */
181 static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
182 unsigned long pfn, struct page *page, int flags)
183 {
184 struct siginfo si;
185 int ret;
186
187 pr_err("Memory failure: %#lx: Killing %s:%d due to hardware memory corruption\n",
188 pfn, t->comm, t->pid);
189 si.si_signo = SIGBUS;
190 si.si_errno = 0;
191 si.si_addr = (void *)addr;
192 #ifdef __ARCH_SI_TRAPNO
193 si.si_trapno = trapno;
194 #endif
195 si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
196
197 if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
198 si.si_code = BUS_MCEERR_AR;
199 ret = force_sig_info(SIGBUS, &si, current);
200 } else {
201 /*
202 * Don't use force here, it's convenient if the signal
203 * can be temporarily blocked.
204 * This could cause a loop when the user sets SIGBUS
205 * to SIG_IGN, but hopefully no one will do that?
206 */
207 si.si_code = BUS_MCEERR_AO;
208 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
209 }
210 if (ret < 0)
211 pr_info("Memory failure: Error sending signal to %s:%d: %d\n",
212 t->comm, t->pid, ret);
213 return ret;
214 }
215
216 /*
217 * When a unknown page type is encountered drain as many buffers as possible
218 * in the hope to turn the page into a LRU or free page, which we can handle.
219 */
220 void shake_page(struct page *p, int access)
221 {
222 if (!PageSlab(p)) {
223 lru_add_drain_all();
224 if (PageLRU(p))
225 return;
226 drain_all_pages(page_zone(p));
227 if (PageLRU(p) || is_free_buddy_page(p))
228 return;
229 }
230
231 /*
232 * Only call shrink_node_slabs here (which would also shrink
233 * other caches) if access is not potentially fatal.
234 */
235 if (access)
236 drop_slab_node(page_to_nid(p));
237 }
238 EXPORT_SYMBOL_GPL(shake_page);
239
240 /*
241 * Kill all processes that have a poisoned page mapped and then isolate
242 * the page.
243 *
244 * General strategy:
245 * Find all processes having the page mapped and kill them.
246 * But we keep a page reference around so that the page is not
247 * actually freed yet.
248 * Then stash the page away
249 *
250 * There's no convenient way to get back to mapped processes
251 * from the VMAs. So do a brute-force search over all
252 * running processes.
253 *
254 * Remember that machine checks are not common (or rather
255 * if they are common you have other problems), so this shouldn't
256 * be a performance issue.
257 *
258 * Also there are some races possible while we get from the
259 * error detection to actually handle it.
260 */
261
262 struct to_kill {
263 struct list_head nd;
264 struct task_struct *tsk;
265 unsigned long addr;
266 char addr_valid;
267 };
268
269 /*
270 * Failure handling: if we can't find or can't kill a process there's
271 * not much we can do. We just print a message and ignore otherwise.
272 */
273
274 /*
275 * Schedule a process for later kill.
276 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
277 * TBD would GFP_NOIO be enough?
278 */
279 static void add_to_kill(struct task_struct *tsk, struct page *p,
280 struct vm_area_struct *vma,
281 struct list_head *to_kill,
282 struct to_kill **tkc)
283 {
284 struct to_kill *tk;
285
286 if (*tkc) {
287 tk = *tkc;
288 *tkc = NULL;
289 } else {
290 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
291 if (!tk) {
292 pr_err("Memory failure: Out of memory while machine check handling\n");
293 return;
294 }
295 }
296 tk->addr = page_address_in_vma(p, vma);
297 tk->addr_valid = 1;
298
299 /*
300 * In theory we don't have to kill when the page was
301 * munmaped. But it could be also a mremap. Since that's
302 * likely very rare kill anyways just out of paranoia, but use
303 * a SIGKILL because the error is not contained anymore.
304 */
305 if (tk->addr == -EFAULT) {
306 pr_info("Memory failure: Unable to find user space address %lx in %s\n",
307 page_to_pfn(p), tsk->comm);
308 tk->addr_valid = 0;
309 }
310 get_task_struct(tsk);
311 tk->tsk = tsk;
312 list_add_tail(&tk->nd, to_kill);
313 }
314
315 /*
316 * Kill the processes that have been collected earlier.
317 *
318 * Only do anything when DOIT is set, otherwise just free the list
319 * (this is used for clean pages which do not need killing)
320 * Also when FAIL is set do a force kill because something went
321 * wrong earlier.
322 */
323 static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
324 int fail, struct page *page, unsigned long pfn,
325 int flags)
326 {
327 struct to_kill *tk, *next;
328
329 list_for_each_entry_safe (tk, next, to_kill, nd) {
330 if (forcekill) {
331 /*
332 * In case something went wrong with munmapping
333 * make sure the process doesn't catch the
334 * signal and then access the memory. Just kill it.
335 */
336 if (fail || tk->addr_valid == 0) {
337 pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
338 pfn, tk->tsk->comm, tk->tsk->pid);
339 force_sig(SIGKILL, tk->tsk);
340 }
341
342 /*
343 * In theory the process could have mapped
344 * something else on the address in-between. We could
345 * check for that, but we need to tell the
346 * process anyways.
347 */
348 else if (kill_proc(tk->tsk, tk->addr, trapno,
349 pfn, page, flags) < 0)
350 pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n",
351 pfn, tk->tsk->comm, tk->tsk->pid);
352 }
353 put_task_struct(tk->tsk);
354 kfree(tk);
355 }
356 }
357
358 /*
359 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
360 * on behalf of the thread group. Return task_struct of the (first found)
361 * dedicated thread if found, and return NULL otherwise.
362 *
363 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
364 * have to call rcu_read_lock/unlock() in this function.
365 */
366 static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
367 {
368 struct task_struct *t;
369
370 for_each_thread(tsk, t)
371 if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
372 return t;
373 return NULL;
374 }
375
376 /*
377 * Determine whether a given process is "early kill" process which expects
378 * to be signaled when some page under the process is hwpoisoned.
379 * Return task_struct of the dedicated thread (main thread unless explicitly
380 * specified) if the process is "early kill," and otherwise returns NULL.
381 */
382 static struct task_struct *task_early_kill(struct task_struct *tsk,
383 int force_early)
384 {
385 struct task_struct *t;
386 if (!tsk->mm)
387 return NULL;
388 if (force_early)
389 return tsk;
390 t = find_early_kill_thread(tsk);
391 if (t)
392 return t;
393 if (sysctl_memory_failure_early_kill)
394 return tsk;
395 return NULL;
396 }
397
398 /*
399 * Collect processes when the error hit an anonymous page.
400 */
401 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
402 struct to_kill **tkc, int force_early)
403 {
404 struct vm_area_struct *vma;
405 struct task_struct *tsk;
406 struct anon_vma *av;
407 pgoff_t pgoff;
408
409 av = page_lock_anon_vma_read(page);
410 if (av == NULL) /* Not actually mapped anymore */
411 return;
412
413 pgoff = page_to_pgoff(page);
414 read_lock(&tasklist_lock);
415 for_each_process (tsk) {
416 struct anon_vma_chain *vmac;
417 struct task_struct *t = task_early_kill(tsk, force_early);
418
419 if (!t)
420 continue;
421 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
422 pgoff, pgoff) {
423 vma = vmac->vma;
424 if (!page_mapped_in_vma(page, vma))
425 continue;
426 if (vma->vm_mm == t->mm)
427 add_to_kill(t, page, vma, to_kill, tkc);
428 }
429 }
430 read_unlock(&tasklist_lock);
431 page_unlock_anon_vma_read(av);
432 }
433
434 /*
435 * Collect processes when the error hit a file mapped page.
436 */
437 static void collect_procs_file(struct page *page, struct list_head *to_kill,
438 struct to_kill **tkc, int force_early)
439 {
440 struct vm_area_struct *vma;
441 struct task_struct *tsk;
442 struct address_space *mapping = page->mapping;
443
444 i_mmap_lock_read(mapping);
445 read_lock(&tasklist_lock);
446 for_each_process(tsk) {
447 pgoff_t pgoff = page_to_pgoff(page);
448 struct task_struct *t = task_early_kill(tsk, force_early);
449
450 if (!t)
451 continue;
452 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
453 pgoff) {
454 /*
455 * Send early kill signal to tasks where a vma covers
456 * the page but the corrupted page is not necessarily
457 * mapped it in its pte.
458 * Assume applications who requested early kill want
459 * to be informed of all such data corruptions.
460 */
461 if (vma->vm_mm == t->mm)
462 add_to_kill(t, page, vma, to_kill, tkc);
463 }
464 }
465 read_unlock(&tasklist_lock);
466 i_mmap_unlock_read(mapping);
467 }
468
469 /*
470 * Collect the processes who have the corrupted page mapped to kill.
471 * This is done in two steps for locking reasons.
472 * First preallocate one tokill structure outside the spin locks,
473 * so that we can kill at least one process reasonably reliable.
474 */
475 static void collect_procs(struct page *page, struct list_head *tokill,
476 int force_early)
477 {
478 struct to_kill *tk;
479
480 if (!page->mapping)
481 return;
482
483 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
484 if (!tk)
485 return;
486 if (PageAnon(page))
487 collect_procs_anon(page, tokill, &tk, force_early);
488 else
489 collect_procs_file(page, tokill, &tk, force_early);
490 kfree(tk);
491 }
492
493 static const char *action_name[] = {
494 [MF_IGNORED] = "Ignored",
495 [MF_FAILED] = "Failed",
496 [MF_DELAYED] = "Delayed",
497 [MF_RECOVERED] = "Recovered",
498 };
499
500 static const char * const action_page_types[] = {
501 [MF_MSG_KERNEL] = "reserved kernel page",
502 [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page",
503 [MF_MSG_SLAB] = "kernel slab page",
504 [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
505 [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned",
506 [MF_MSG_HUGE] = "huge page",
507 [MF_MSG_FREE_HUGE] = "free huge page",
508 [MF_MSG_UNMAP_FAILED] = "unmapping failed page",
509 [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page",
510 [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page",
511 [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page",
512 [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page",
513 [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page",
514 [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page",
515 [MF_MSG_DIRTY_LRU] = "dirty LRU page",
516 [MF_MSG_CLEAN_LRU] = "clean LRU page",
517 [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page",
518 [MF_MSG_BUDDY] = "free buddy page",
519 [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)",
520 [MF_MSG_UNKNOWN] = "unknown page",
521 };
522
523 /*
524 * XXX: It is possible that a page is isolated from LRU cache,
525 * and then kept in swap cache or failed to remove from page cache.
526 * The page count will stop it from being freed by unpoison.
527 * Stress tests should be aware of this memory leak problem.
528 */
529 static int delete_from_lru_cache(struct page *p)
530 {
531 if (!isolate_lru_page(p)) {
532 /*
533 * Clear sensible page flags, so that the buddy system won't
534 * complain when the page is unpoison-and-freed.
535 */
536 ClearPageActive(p);
537 ClearPageUnevictable(p);
538 /*
539 * drop the page count elevated by isolate_lru_page()
540 */
541 put_page(p);
542 return 0;
543 }
544 return -EIO;
545 }
546
547 /*
548 * Error hit kernel page.
549 * Do nothing, try to be lucky and not touch this instead. For a few cases we
550 * could be more sophisticated.
551 */
552 static int me_kernel(struct page *p, unsigned long pfn)
553 {
554 return MF_IGNORED;
555 }
556
557 /*
558 * Page in unknown state. Do nothing.
559 */
560 static int me_unknown(struct page *p, unsigned long pfn)
561 {
562 pr_err("Memory failure: %#lx: Unknown page state\n", pfn);
563 return MF_FAILED;
564 }
565
566 /*
567 * Clean (or cleaned) page cache page.
568 */
569 static int me_pagecache_clean(struct page *p, unsigned long pfn)
570 {
571 int err;
572 int ret = MF_FAILED;
573 struct address_space *mapping;
574
575 delete_from_lru_cache(p);
576
577 /*
578 * For anonymous pages we're done the only reference left
579 * should be the one m_f() holds.
580 */
581 if (PageAnon(p))
582 return MF_RECOVERED;
583
584 /*
585 * Now truncate the page in the page cache. This is really
586 * more like a "temporary hole punch"
587 * Don't do this for block devices when someone else
588 * has a reference, because it could be file system metadata
589 * and that's not safe to truncate.
590 */
591 mapping = page_mapping(p);
592 if (!mapping) {
593 /*
594 * Page has been teared down in the meanwhile
595 */
596 return MF_FAILED;
597 }
598
599 /*
600 * Truncation is a bit tricky. Enable it per file system for now.
601 *
602 * Open: to take i_mutex or not for this? Right now we don't.
603 */
604 if (mapping->a_ops->error_remove_page) {
605 err = mapping->a_ops->error_remove_page(mapping, p);
606 if (err != 0) {
607 pr_info("Memory failure: %#lx: Failed to punch page: %d\n",
608 pfn, err);
609 } else if (page_has_private(p) &&
610 !try_to_release_page(p, GFP_NOIO)) {
611 pr_info("Memory failure: %#lx: failed to release buffers\n",
612 pfn);
613 } else {
614 ret = MF_RECOVERED;
615 }
616 } else {
617 /*
618 * If the file system doesn't support it just invalidate
619 * This fails on dirty or anything with private pages
620 */
621 if (invalidate_inode_page(p))
622 ret = MF_RECOVERED;
623 else
624 pr_info("Memory failure: %#lx: Failed to invalidate\n",
625 pfn);
626 }
627 return ret;
628 }
629
630 /*
631 * Dirty pagecache page
632 * Issues: when the error hit a hole page the error is not properly
633 * propagated.
634 */
635 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
636 {
637 struct address_space *mapping = page_mapping(p);
638
639 SetPageError(p);
640 /* TBD: print more information about the file. */
641 if (mapping) {
642 /*
643 * IO error will be reported by write(), fsync(), etc.
644 * who check the mapping.
645 * This way the application knows that something went
646 * wrong with its dirty file data.
647 *
648 * There's one open issue:
649 *
650 * The EIO will be only reported on the next IO
651 * operation and then cleared through the IO map.
652 * Normally Linux has two mechanisms to pass IO error
653 * first through the AS_EIO flag in the address space
654 * and then through the PageError flag in the page.
655 * Since we drop pages on memory failure handling the
656 * only mechanism open to use is through AS_AIO.
657 *
658 * This has the disadvantage that it gets cleared on
659 * the first operation that returns an error, while
660 * the PageError bit is more sticky and only cleared
661 * when the page is reread or dropped. If an
662 * application assumes it will always get error on
663 * fsync, but does other operations on the fd before
664 * and the page is dropped between then the error
665 * will not be properly reported.
666 *
667 * This can already happen even without hwpoisoned
668 * pages: first on metadata IO errors (which only
669 * report through AS_EIO) or when the page is dropped
670 * at the wrong time.
671 *
672 * So right now we assume that the application DTRT on
673 * the first EIO, but we're not worse than other parts
674 * of the kernel.
675 */
676 mapping_set_error(mapping, EIO);
677 }
678
679 return me_pagecache_clean(p, pfn);
680 }
681
682 /*
683 * Clean and dirty swap cache.
684 *
685 * Dirty swap cache page is tricky to handle. The page could live both in page
686 * cache and swap cache(ie. page is freshly swapped in). So it could be
687 * referenced concurrently by 2 types of PTEs:
688 * normal PTEs and swap PTEs. We try to handle them consistently by calling
689 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
690 * and then
691 * - clear dirty bit to prevent IO
692 * - remove from LRU
693 * - but keep in the swap cache, so that when we return to it on
694 * a later page fault, we know the application is accessing
695 * corrupted data and shall be killed (we installed simple
696 * interception code in do_swap_page to catch it).
697 *
698 * Clean swap cache pages can be directly isolated. A later page fault will
699 * bring in the known good data from disk.
700 */
701 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
702 {
703 ClearPageDirty(p);
704 /* Trigger EIO in shmem: */
705 ClearPageUptodate(p);
706
707 if (!delete_from_lru_cache(p))
708 return MF_DELAYED;
709 else
710 return MF_FAILED;
711 }
712
713 static int me_swapcache_clean(struct page *p, unsigned long pfn)
714 {
715 delete_from_swap_cache(p);
716
717 if (!delete_from_lru_cache(p))
718 return MF_RECOVERED;
719 else
720 return MF_FAILED;
721 }
722
723 /*
724 * Huge pages. Needs work.
725 * Issues:
726 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
727 * To narrow down kill region to one page, we need to break up pmd.
728 */
729 static int me_huge_page(struct page *p, unsigned long pfn)
730 {
731 int res = 0;
732 struct page *hpage = compound_head(p);
733
734 if (!PageHuge(hpage))
735 return MF_DELAYED;
736
737 /*
738 * We can safely recover from error on free or reserved (i.e.
739 * not in-use) hugepage by dequeuing it from freelist.
740 * To check whether a hugepage is in-use or not, we can't use
741 * page->lru because it can be used in other hugepage operations,
742 * such as __unmap_hugepage_range() and gather_surplus_pages().
743 * So instead we use page_mapping() and PageAnon().
744 */
745 if (!(page_mapping(hpage) || PageAnon(hpage))) {
746 res = dequeue_hwpoisoned_huge_page(hpage);
747 if (!res)
748 return MF_RECOVERED;
749 }
750 return MF_DELAYED;
751 }
752
753 /*
754 * Various page states we can handle.
755 *
756 * A page state is defined by its current page->flags bits.
757 * The table matches them in order and calls the right handler.
758 *
759 * This is quite tricky because we can access page at any time
760 * in its live cycle, so all accesses have to be extremely careful.
761 *
762 * This is not complete. More states could be added.
763 * For any missing state don't attempt recovery.
764 */
765
766 #define dirty (1UL << PG_dirty)
767 #define sc (1UL << PG_swapcache)
768 #define unevict (1UL << PG_unevictable)
769 #define mlock (1UL << PG_mlocked)
770 #define writeback (1UL << PG_writeback)
771 #define lru (1UL << PG_lru)
772 #define swapbacked (1UL << PG_swapbacked)
773 #define head (1UL << PG_head)
774 #define slab (1UL << PG_slab)
775 #define reserved (1UL << PG_reserved)
776
777 static struct page_state {
778 unsigned long mask;
779 unsigned long res;
780 enum mf_action_page_type type;
781 int (*action)(struct page *p, unsigned long pfn);
782 } error_states[] = {
783 { reserved, reserved, MF_MSG_KERNEL, me_kernel },
784 /*
785 * free pages are specially detected outside this table:
786 * PG_buddy pages only make a small fraction of all free pages.
787 */
788
789 /*
790 * Could in theory check if slab page is free or if we can drop
791 * currently unused objects without touching them. But just
792 * treat it as standard kernel for now.
793 */
794 { slab, slab, MF_MSG_SLAB, me_kernel },
795
796 { head, head, MF_MSG_HUGE, me_huge_page },
797
798 { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
799 { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
800
801 { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty },
802 { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean },
803
804 { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty },
805 { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean },
806
807 { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty },
808 { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean },
809
810 /*
811 * Catchall entry: must be at end.
812 */
813 { 0, 0, MF_MSG_UNKNOWN, me_unknown },
814 };
815
816 #undef dirty
817 #undef sc
818 #undef unevict
819 #undef mlock
820 #undef writeback
821 #undef lru
822 #undef swapbacked
823 #undef head
824 #undef slab
825 #undef reserved
826
827 /*
828 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
829 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
830 */
831 static void action_result(unsigned long pfn, enum mf_action_page_type type,
832 enum mf_result result)
833 {
834 trace_memory_failure_event(pfn, type, result);
835
836 pr_err("Memory failure: %#lx: recovery action for %s: %s\n",
837 pfn, action_page_types[type], action_name[result]);
838 }
839
840 static int page_action(struct page_state *ps, struct page *p,
841 unsigned long pfn)
842 {
843 int result;
844 int count;
845
846 result = ps->action(p, pfn);
847
848 count = page_count(p) - 1;
849 if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
850 count--;
851 if (count != 0) {
852 pr_err("Memory failure: %#lx: %s still referenced by %d users\n",
853 pfn, action_page_types[ps->type], count);
854 result = MF_FAILED;
855 }
856 action_result(pfn, ps->type, result);
857
858 /* Could do more checks here if page looks ok */
859 /*
860 * Could adjust zone counters here to correct for the missing page.
861 */
862
863 return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
864 }
865
866 /**
867 * get_hwpoison_page() - Get refcount for memory error handling:
868 * @page: raw error page (hit by memory error)
869 *
870 * Return: return 0 if failed to grab the refcount, otherwise true (some
871 * non-zero value.)
872 */
873 int get_hwpoison_page(struct page *page)
874 {
875 struct page *head = compound_head(page);
876
877 if (!PageHuge(head) && PageTransHuge(head)) {
878 /*
879 * Non anonymous thp exists only in allocation/free time. We
880 * can't handle such a case correctly, so let's give it up.
881 * This should be better than triggering BUG_ON when kernel
882 * tries to touch the "partially handled" page.
883 */
884 if (!PageAnon(head)) {
885 pr_err("Memory failure: %#lx: non anonymous thp\n",
886 page_to_pfn(page));
887 return 0;
888 }
889 }
890
891 if (get_page_unless_zero(head)) {
892 if (head == compound_head(page))
893 return 1;
894
895 pr_info("Memory failure: %#lx cannot catch tail\n",
896 page_to_pfn(page));
897 put_page(head);
898 }
899
900 return 0;
901 }
902 EXPORT_SYMBOL_GPL(get_hwpoison_page);
903
904 /*
905 * Do all that is necessary to remove user space mappings. Unmap
906 * the pages and send SIGBUS to the processes if the data was dirty.
907 */
908 static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
909 int trapno, int flags, struct page **hpagep)
910 {
911 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
912 struct address_space *mapping;
913 LIST_HEAD(tokill);
914 int ret;
915 int kill = 1, forcekill;
916 struct page *hpage = *hpagep;
917
918 /*
919 * Here we are interested only in user-mapped pages, so skip any
920 * other types of pages.
921 */
922 if (PageReserved(p) || PageSlab(p))
923 return SWAP_SUCCESS;
924 if (!(PageLRU(hpage) || PageHuge(p)))
925 return SWAP_SUCCESS;
926
927 /*
928 * This check implies we don't kill processes if their pages
929 * are in the swap cache early. Those are always late kills.
930 */
931 if (!page_mapped(hpage))
932 return SWAP_SUCCESS;
933
934 if (PageKsm(p)) {
935 pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn);
936 return SWAP_FAIL;
937 }
938
939 if (PageSwapCache(p)) {
940 pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n",
941 pfn);
942 ttu |= TTU_IGNORE_HWPOISON;
943 }
944
945 /*
946 * Propagate the dirty bit from PTEs to struct page first, because we
947 * need this to decide if we should kill or just drop the page.
948 * XXX: the dirty test could be racy: set_page_dirty() may not always
949 * be called inside page lock (it's recommended but not enforced).
950 */
951 mapping = page_mapping(hpage);
952 if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
953 mapping_cap_writeback_dirty(mapping)) {
954 if (page_mkclean(hpage)) {
955 SetPageDirty(hpage);
956 } else {
957 kill = 0;
958 ttu |= TTU_IGNORE_HWPOISON;
959 pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n",
960 pfn);
961 }
962 }
963
964 /*
965 * First collect all the processes that have the page
966 * mapped in dirty form. This has to be done before try_to_unmap,
967 * because ttu takes the rmap data structures down.
968 *
969 * Error handling: We ignore errors here because
970 * there's nothing that can be done.
971 */
972 if (kill)
973 collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
974
975 ret = try_to_unmap(hpage, ttu);
976 if (ret != SWAP_SUCCESS)
977 pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n",
978 pfn, page_mapcount(hpage));
979
980 /*
981 * Now that the dirty bit has been propagated to the
982 * struct page and all unmaps done we can decide if
983 * killing is needed or not. Only kill when the page
984 * was dirty or the process is not restartable,
985 * otherwise the tokill list is merely
986 * freed. When there was a problem unmapping earlier
987 * use a more force-full uncatchable kill to prevent
988 * any accesses to the poisoned memory.
989 */
990 forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
991 kill_procs(&tokill, forcekill, trapno,
992 ret != SWAP_SUCCESS, p, pfn, flags);
993
994 return ret;
995 }
996
997 static void set_page_hwpoison_huge_page(struct page *hpage)
998 {
999 int i;
1000 int nr_pages = 1 << compound_order(hpage);
1001 for (i = 0; i < nr_pages; i++)
1002 SetPageHWPoison(hpage + i);
1003 }
1004
1005 static void clear_page_hwpoison_huge_page(struct page *hpage)
1006 {
1007 int i;
1008 int nr_pages = 1 << compound_order(hpage);
1009 for (i = 0; i < nr_pages; i++)
1010 ClearPageHWPoison(hpage + i);
1011 }
1012
1013 /**
1014 * memory_failure - Handle memory failure of a page.
1015 * @pfn: Page Number of the corrupted page
1016 * @trapno: Trap number reported in the signal to user space.
1017 * @flags: fine tune action taken
1018 *
1019 * This function is called by the low level machine check code
1020 * of an architecture when it detects hardware memory corruption
1021 * of a page. It tries its best to recover, which includes
1022 * dropping pages, killing processes etc.
1023 *
1024 * The function is primarily of use for corruptions that
1025 * happen outside the current execution context (e.g. when
1026 * detected by a background scrubber)
1027 *
1028 * Must run in process context (e.g. a work queue) with interrupts
1029 * enabled and no spinlocks hold.
1030 */
1031 int memory_failure(unsigned long pfn, int trapno, int flags)
1032 {
1033 struct page_state *ps;
1034 struct page *p;
1035 struct page *hpage;
1036 struct page *orig_head;
1037 int res;
1038 unsigned int nr_pages;
1039 unsigned long page_flags;
1040
1041 if (!sysctl_memory_failure_recovery)
1042 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1043
1044 if (!pfn_valid(pfn)) {
1045 pr_err("Memory failure: %#lx: memory outside kernel control\n",
1046 pfn);
1047 return -ENXIO;
1048 }
1049
1050 p = pfn_to_page(pfn);
1051 orig_head = hpage = compound_head(p);
1052 if (TestSetPageHWPoison(p)) {
1053 pr_err("Memory failure: %#lx: already hardware poisoned\n",
1054 pfn);
1055 return 0;
1056 }
1057
1058 /*
1059 * Currently errors on hugetlbfs pages are measured in hugepage units,
1060 * so nr_pages should be 1 << compound_order. OTOH when errors are on
1061 * transparent hugepages, they are supposed to be split and error
1062 * measurement is done in normal page units. So nr_pages should be one
1063 * in this case.
1064 */
1065 if (PageHuge(p))
1066 nr_pages = 1 << compound_order(hpage);
1067 else /* normal page or thp */
1068 nr_pages = 1;
1069 num_poisoned_pages_add(nr_pages);
1070
1071 /*
1072 * We need/can do nothing about count=0 pages.
1073 * 1) it's a free page, and therefore in safe hand:
1074 * prep_new_page() will be the gate keeper.
1075 * 2) it's a free hugepage, which is also safe:
1076 * an affected hugepage will be dequeued from hugepage freelist,
1077 * so there's no concern about reusing it ever after.
1078 * 3) it's part of a non-compound high order page.
1079 * Implies some kernel user: cannot stop them from
1080 * R/W the page; let's pray that the page has been
1081 * used and will be freed some time later.
1082 * In fact it's dangerous to directly bump up page count from 0,
1083 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1084 */
1085 if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
1086 if (is_free_buddy_page(p)) {
1087 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1088 return 0;
1089 } else if (PageHuge(hpage)) {
1090 /*
1091 * Check "filter hit" and "race with other subpage."
1092 */
1093 lock_page(hpage);
1094 if (PageHWPoison(hpage)) {
1095 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1096 || (p != hpage && TestSetPageHWPoison(hpage))) {
1097 num_poisoned_pages_sub(nr_pages);
1098 unlock_page(hpage);
1099 return 0;
1100 }
1101 }
1102 set_page_hwpoison_huge_page(hpage);
1103 res = dequeue_hwpoisoned_huge_page(hpage);
1104 action_result(pfn, MF_MSG_FREE_HUGE,
1105 res ? MF_IGNORED : MF_DELAYED);
1106 unlock_page(hpage);
1107 return res;
1108 } else {
1109 action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
1110 return -EBUSY;
1111 }
1112 }
1113
1114 if (!PageHuge(p) && PageTransHuge(hpage)) {
1115 lock_page(hpage);
1116 if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) {
1117 unlock_page(hpage);
1118 if (!PageAnon(hpage))
1119 pr_err("Memory failure: %#lx: non anonymous thp\n",
1120 pfn);
1121 else
1122 pr_err("Memory failure: %#lx: thp split failed\n",
1123 pfn);
1124 if (TestClearPageHWPoison(p))
1125 num_poisoned_pages_sub(nr_pages);
1126 put_hwpoison_page(p);
1127 return -EBUSY;
1128 }
1129 unlock_page(hpage);
1130 get_hwpoison_page(p);
1131 put_hwpoison_page(hpage);
1132 VM_BUG_ON_PAGE(!page_count(p), p);
1133 hpage = compound_head(p);
1134 }
1135
1136 /*
1137 * We ignore non-LRU pages for good reasons.
1138 * - PG_locked is only well defined for LRU pages and a few others
1139 * - to avoid races with __SetPageLocked()
1140 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1141 * The check (unnecessarily) ignores LRU pages being isolated and
1142 * walked by the page reclaim code, however that's not a big loss.
1143 */
1144 if (!PageHuge(p)) {
1145 if (!PageLRU(p))
1146 shake_page(p, 0);
1147 if (!PageLRU(p)) {
1148 /*
1149 * shake_page could have turned it free.
1150 */
1151 if (is_free_buddy_page(p)) {
1152 if (flags & MF_COUNT_INCREASED)
1153 action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1154 else
1155 action_result(pfn, MF_MSG_BUDDY_2ND,
1156 MF_DELAYED);
1157 return 0;
1158 }
1159 }
1160 }
1161
1162 lock_page(hpage);
1163
1164 /*
1165 * The page could have changed compound pages during the locking.
1166 * If this happens just bail out.
1167 */
1168 if (PageCompound(p) && compound_head(p) != orig_head) {
1169 action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
1170 res = -EBUSY;
1171 goto out;
1172 }
1173
1174 /*
1175 * We use page flags to determine what action should be taken, but
1176 * the flags can be modified by the error containment action. One
1177 * example is an mlocked page, where PG_mlocked is cleared by
1178 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1179 * correctly, we save a copy of the page flags at this time.
1180 */
1181 page_flags = p->flags;
1182
1183 /*
1184 * unpoison always clear PG_hwpoison inside page lock
1185 */
1186 if (!PageHWPoison(p)) {
1187 pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1188 num_poisoned_pages_sub(nr_pages);
1189 unlock_page(hpage);
1190 put_hwpoison_page(hpage);
1191 return 0;
1192 }
1193 if (hwpoison_filter(p)) {
1194 if (TestClearPageHWPoison(p))
1195 num_poisoned_pages_sub(nr_pages);
1196 unlock_page(hpage);
1197 put_hwpoison_page(hpage);
1198 return 0;
1199 }
1200
1201 if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p))
1202 goto identify_page_state;
1203
1204 /*
1205 * For error on the tail page, we should set PG_hwpoison
1206 * on the head page to show that the hugepage is hwpoisoned
1207 */
1208 if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1209 action_result(pfn, MF_MSG_POISONED_HUGE, MF_IGNORED);
1210 unlock_page(hpage);
1211 put_hwpoison_page(hpage);
1212 return 0;
1213 }
1214 /*
1215 * Set PG_hwpoison on all pages in an error hugepage,
1216 * because containment is done in hugepage unit for now.
1217 * Since we have done TestSetPageHWPoison() for the head page with
1218 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1219 */
1220 if (PageHuge(p))
1221 set_page_hwpoison_huge_page(hpage);
1222
1223 /*
1224 * It's very difficult to mess with pages currently under IO
1225 * and in many cases impossible, so we just avoid it here.
1226 */
1227 wait_on_page_writeback(p);
1228
1229 /*
1230 * Now take care of user space mappings.
1231 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1232 *
1233 * When the raw error page is thp tail page, hpage points to the raw
1234 * page after thp split.
1235 */
1236 if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)
1237 != SWAP_SUCCESS) {
1238 action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1239 res = -EBUSY;
1240 goto out;
1241 }
1242
1243 /*
1244 * Torn down by someone else?
1245 */
1246 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1247 action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
1248 res = -EBUSY;
1249 goto out;
1250 }
1251
1252 identify_page_state:
1253 res = -EBUSY;
1254 /*
1255 * The first check uses the current page flags which may not have any
1256 * relevant information. The second check with the saved page flagss is
1257 * carried out only if the first check can't determine the page status.
1258 */
1259 for (ps = error_states;; ps++)
1260 if ((p->flags & ps->mask) == ps->res)
1261 break;
1262
1263 page_flags |= (p->flags & (1UL << PG_dirty));
1264
1265 if (!ps->mask)
1266 for (ps = error_states;; ps++)
1267 if ((page_flags & ps->mask) == ps->res)
1268 break;
1269 res = page_action(ps, p, pfn);
1270 out:
1271 unlock_page(hpage);
1272 return res;
1273 }
1274 EXPORT_SYMBOL_GPL(memory_failure);
1275
1276 #define MEMORY_FAILURE_FIFO_ORDER 4
1277 #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER)
1278
1279 struct memory_failure_entry {
1280 unsigned long pfn;
1281 int trapno;
1282 int flags;
1283 };
1284
1285 struct memory_failure_cpu {
1286 DECLARE_KFIFO(fifo, struct memory_failure_entry,
1287 MEMORY_FAILURE_FIFO_SIZE);
1288 spinlock_t lock;
1289 struct work_struct work;
1290 };
1291
1292 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1293
1294 /**
1295 * memory_failure_queue - Schedule handling memory failure of a page.
1296 * @pfn: Page Number of the corrupted page
1297 * @trapno: Trap number reported in the signal to user space.
1298 * @flags: Flags for memory failure handling
1299 *
1300 * This function is called by the low level hardware error handler
1301 * when it detects hardware memory corruption of a page. It schedules
1302 * the recovering of error page, including dropping pages, killing
1303 * processes etc.
1304 *
1305 * The function is primarily of use for corruptions that
1306 * happen outside the current execution context (e.g. when
1307 * detected by a background scrubber)
1308 *
1309 * Can run in IRQ context.
1310 */
1311 void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1312 {
1313 struct memory_failure_cpu *mf_cpu;
1314 unsigned long proc_flags;
1315 struct memory_failure_entry entry = {
1316 .pfn = pfn,
1317 .trapno = trapno,
1318 .flags = flags,
1319 };
1320
1321 mf_cpu = &get_cpu_var(memory_failure_cpu);
1322 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1323 if (kfifo_put(&mf_cpu->fifo, entry))
1324 schedule_work_on(smp_processor_id(), &mf_cpu->work);
1325 else
1326 pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1327 pfn);
1328 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1329 put_cpu_var(memory_failure_cpu);
1330 }
1331 EXPORT_SYMBOL_GPL(memory_failure_queue);
1332
1333 static void memory_failure_work_func(struct work_struct *work)
1334 {
1335 struct memory_failure_cpu *mf_cpu;
1336 struct memory_failure_entry entry = { 0, };
1337 unsigned long proc_flags;
1338 int gotten;
1339
1340 mf_cpu = this_cpu_ptr(&memory_failure_cpu);
1341 for (;;) {
1342 spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1343 gotten = kfifo_get(&mf_cpu->fifo, &entry);
1344 spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1345 if (!gotten)
1346 break;
1347 if (entry.flags & MF_SOFT_OFFLINE)
1348 soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1349 else
1350 memory_failure(entry.pfn, entry.trapno, entry.flags);
1351 }
1352 }
1353
1354 static int __init memory_failure_init(void)
1355 {
1356 struct memory_failure_cpu *mf_cpu;
1357 int cpu;
1358
1359 for_each_possible_cpu(cpu) {
1360 mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1361 spin_lock_init(&mf_cpu->lock);
1362 INIT_KFIFO(mf_cpu->fifo);
1363 INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1364 }
1365
1366 return 0;
1367 }
1368 core_initcall(memory_failure_init);
1369
1370 #define unpoison_pr_info(fmt, pfn, rs) \
1371 ({ \
1372 if (__ratelimit(rs)) \
1373 pr_info(fmt, pfn); \
1374 })
1375
1376 /**
1377 * unpoison_memory - Unpoison a previously poisoned page
1378 * @pfn: Page number of the to be unpoisoned page
1379 *
1380 * Software-unpoison a page that has been poisoned by
1381 * memory_failure() earlier.
1382 *
1383 * This is only done on the software-level, so it only works
1384 * for linux injected failures, not real hardware failures
1385 *
1386 * Returns 0 for success, otherwise -errno.
1387 */
1388 int unpoison_memory(unsigned long pfn)
1389 {
1390 struct page *page;
1391 struct page *p;
1392 int freeit = 0;
1393 unsigned int nr_pages;
1394 static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
1395 DEFAULT_RATELIMIT_BURST);
1396
1397 if (!pfn_valid(pfn))
1398 return -ENXIO;
1399
1400 p = pfn_to_page(pfn);
1401 page = compound_head(p);
1402
1403 if (!PageHWPoison(p)) {
1404 unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
1405 pfn, &unpoison_rs);
1406 return 0;
1407 }
1408
1409 if (page_count(page) > 1) {
1410 unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
1411 pfn, &unpoison_rs);
1412 return 0;
1413 }
1414
1415 if (page_mapped(page)) {
1416 unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
1417 pfn, &unpoison_rs);
1418 return 0;
1419 }
1420
1421 if (page_mapping(page)) {
1422 unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
1423 pfn, &unpoison_rs);
1424 return 0;
1425 }
1426
1427 /*
1428 * unpoison_memory() can encounter thp only when the thp is being
1429 * worked by memory_failure() and the page lock is not held yet.
1430 * In such case, we yield to memory_failure() and make unpoison fail.
1431 */
1432 if (!PageHuge(page) && PageTransHuge(page)) {
1433 unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n",
1434 pfn, &unpoison_rs);
1435 return 0;
1436 }
1437
1438 nr_pages = 1 << compound_order(page);
1439
1440 if (!get_hwpoison_page(p)) {
1441 /*
1442 * Since HWPoisoned hugepage should have non-zero refcount,
1443 * race between memory failure and unpoison seems to happen.
1444 * In such case unpoison fails and memory failure runs
1445 * to the end.
1446 */
1447 if (PageHuge(page)) {
1448 unpoison_pr_info("Unpoison: Memory failure is now running on free hugepage %#lx\n",
1449 pfn, &unpoison_rs);
1450 return 0;
1451 }
1452 if (TestClearPageHWPoison(p))
1453 num_poisoned_pages_dec();
1454 unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n",
1455 pfn, &unpoison_rs);
1456 return 0;
1457 }
1458
1459 lock_page(page);
1460 /*
1461 * This test is racy because PG_hwpoison is set outside of page lock.
1462 * That's acceptable because that won't trigger kernel panic. Instead,
1463 * the PG_hwpoison page will be caught and isolated on the entrance to
1464 * the free buddy page pool.
1465 */
1466 if (TestClearPageHWPoison(page)) {
1467 unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
1468 pfn, &unpoison_rs);
1469 num_poisoned_pages_sub(nr_pages);
1470 freeit = 1;
1471 if (PageHuge(page))
1472 clear_page_hwpoison_huge_page(page);
1473 }
1474 unlock_page(page);
1475
1476 put_hwpoison_page(page);
1477 if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
1478 put_hwpoison_page(page);
1479
1480 return 0;
1481 }
1482 EXPORT_SYMBOL(unpoison_memory);
1483
1484 static struct page *new_page(struct page *p, unsigned long private, int **x)
1485 {
1486 int nid = page_to_nid(p);
1487 if (PageHuge(p))
1488 return alloc_huge_page_node(page_hstate(compound_head(p)),
1489 nid);
1490 else
1491 return __alloc_pages_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1492 }
1493
1494 /*
1495 * Safely get reference count of an arbitrary page.
1496 * Returns 0 for a free page, -EIO for a zero refcount page
1497 * that is not free, and 1 for any other page type.
1498 * For 1 the page is returned with increased page count, otherwise not.
1499 */
1500 static int __get_any_page(struct page *p, unsigned long pfn, int flags)
1501 {
1502 int ret;
1503
1504 if (flags & MF_COUNT_INCREASED)
1505 return 1;
1506
1507 /*
1508 * When the target page is a free hugepage, just remove it
1509 * from free hugepage list.
1510 */
1511 if (!get_hwpoison_page(p)) {
1512 if (PageHuge(p)) {
1513 pr_info("%s: %#lx free huge page\n", __func__, pfn);
1514 ret = 0;
1515 } else if (is_free_buddy_page(p)) {
1516 pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1517 ret = 0;
1518 } else {
1519 pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1520 __func__, pfn, p->flags);
1521 ret = -EIO;
1522 }
1523 } else {
1524 /* Not a free page */
1525 ret = 1;
1526 }
1527 return ret;
1528 }
1529
1530 static int get_any_page(struct page *page, unsigned long pfn, int flags)
1531 {
1532 int ret = __get_any_page(page, pfn, flags);
1533
1534 if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1535 /*
1536 * Try to free it.
1537 */
1538 put_hwpoison_page(page);
1539 shake_page(page, 1);
1540
1541 /*
1542 * Did it turn free?
1543 */
1544 ret = __get_any_page(page, pfn, 0);
1545 if (ret == 1 && !PageLRU(page)) {
1546 /* Drop page reference which is from __get_any_page() */
1547 put_hwpoison_page(page);
1548 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1549 pfn, page->flags);
1550 return -EIO;
1551 }
1552 }
1553 return ret;
1554 }
1555
1556 static int soft_offline_huge_page(struct page *page, int flags)
1557 {
1558 int ret;
1559 unsigned long pfn = page_to_pfn(page);
1560 struct page *hpage = compound_head(page);
1561 LIST_HEAD(pagelist);
1562
1563 /*
1564 * This double-check of PageHWPoison is to avoid the race with
1565 * memory_failure(). See also comment in __soft_offline_page().
1566 */
1567 lock_page(hpage);
1568 if (PageHWPoison(hpage)) {
1569 unlock_page(hpage);
1570 put_hwpoison_page(hpage);
1571 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1572 return -EBUSY;
1573 }
1574 unlock_page(hpage);
1575
1576 ret = isolate_huge_page(hpage, &pagelist);
1577 /*
1578 * get_any_page() and isolate_huge_page() takes a refcount each,
1579 * so need to drop one here.
1580 */
1581 put_hwpoison_page(hpage);
1582 if (!ret) {
1583 pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn);
1584 return -EBUSY;
1585 }
1586
1587 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1588 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1589 if (ret) {
1590 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1591 pfn, ret, page->flags);
1592 /*
1593 * We know that soft_offline_huge_page() tries to migrate
1594 * only one hugepage pointed to by hpage, so we need not
1595 * run through the pagelist here.
1596 */
1597 putback_active_hugepage(hpage);
1598 if (ret > 0)
1599 ret = -EIO;
1600 } else {
1601 /* overcommit hugetlb page will be freed to buddy */
1602 if (PageHuge(page)) {
1603 set_page_hwpoison_huge_page(hpage);
1604 dequeue_hwpoisoned_huge_page(hpage);
1605 num_poisoned_pages_add(1 << compound_order(hpage));
1606 } else {
1607 SetPageHWPoison(page);
1608 num_poisoned_pages_inc();
1609 }
1610 }
1611 return ret;
1612 }
1613
1614 static int __soft_offline_page(struct page *page, int flags)
1615 {
1616 int ret;
1617 unsigned long pfn = page_to_pfn(page);
1618
1619 /*
1620 * Check PageHWPoison again inside page lock because PageHWPoison
1621 * is set by memory_failure() outside page lock. Note that
1622 * memory_failure() also double-checks PageHWPoison inside page lock,
1623 * so there's no race between soft_offline_page() and memory_failure().
1624 */
1625 lock_page(page);
1626 wait_on_page_writeback(page);
1627 if (PageHWPoison(page)) {
1628 unlock_page(page);
1629 put_hwpoison_page(page);
1630 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1631 return -EBUSY;
1632 }
1633 /*
1634 * Try to invalidate first. This should work for
1635 * non dirty unmapped page cache pages.
1636 */
1637 ret = invalidate_inode_page(page);
1638 unlock_page(page);
1639 /*
1640 * RED-PEN would be better to keep it isolated here, but we
1641 * would need to fix isolation locking first.
1642 */
1643 if (ret == 1) {
1644 put_hwpoison_page(page);
1645 pr_info("soft_offline: %#lx: invalidated\n", pfn);
1646 SetPageHWPoison(page);
1647 num_poisoned_pages_inc();
1648 return 0;
1649 }
1650
1651 /*
1652 * Simple invalidation didn't work.
1653 * Try to migrate to a new page instead. migrate.c
1654 * handles a large number of cases for us.
1655 */
1656 ret = isolate_lru_page(page);
1657 /*
1658 * Drop page reference which is came from get_any_page()
1659 * successful isolate_lru_page() already took another one.
1660 */
1661 put_hwpoison_page(page);
1662 if (!ret) {
1663 LIST_HEAD(pagelist);
1664 inc_node_page_state(page, NR_ISOLATED_ANON +
1665 page_is_file_cache(page));
1666 list_add(&page->lru, &pagelist);
1667 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1668 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1669 if (ret) {
1670 if (!list_empty(&pagelist)) {
1671 list_del(&page->lru);
1672 dec_node_page_state(page, NR_ISOLATED_ANON +
1673 page_is_file_cache(page));
1674 putback_lru_page(page);
1675 }
1676
1677 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1678 pfn, ret, page->flags);
1679 if (ret > 0)
1680 ret = -EIO;
1681 }
1682 } else {
1683 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1684 pfn, ret, page_count(page), page->flags);
1685 }
1686 return ret;
1687 }
1688
1689 static int soft_offline_in_use_page(struct page *page, int flags)
1690 {
1691 int ret;
1692 struct page *hpage = compound_head(page);
1693
1694 if (!PageHuge(page) && PageTransHuge(hpage)) {
1695 lock_page(hpage);
1696 if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) {
1697 unlock_page(hpage);
1698 if (!PageAnon(hpage))
1699 pr_info("soft offline: %#lx: non anonymous thp\n", page_to_pfn(page));
1700 else
1701 pr_info("soft offline: %#lx: thp split failed\n", page_to_pfn(page));
1702 put_hwpoison_page(hpage);
1703 return -EBUSY;
1704 }
1705 unlock_page(hpage);
1706 get_hwpoison_page(page);
1707 put_hwpoison_page(hpage);
1708 }
1709
1710 if (PageHuge(page))
1711 ret = soft_offline_huge_page(page, flags);
1712 else
1713 ret = __soft_offline_page(page, flags);
1714
1715 return ret;
1716 }
1717
1718 static void soft_offline_free_page(struct page *page)
1719 {
1720 if (PageHuge(page)) {
1721 struct page *hpage = compound_head(page);
1722
1723 set_page_hwpoison_huge_page(hpage);
1724 if (!dequeue_hwpoisoned_huge_page(hpage))
1725 num_poisoned_pages_add(1 << compound_order(hpage));
1726 } else {
1727 if (!TestSetPageHWPoison(page))
1728 num_poisoned_pages_inc();
1729 }
1730 }
1731
1732 /**
1733 * soft_offline_page - Soft offline a page.
1734 * @page: page to offline
1735 * @flags: flags. Same as memory_failure().
1736 *
1737 * Returns 0 on success, otherwise negated errno.
1738 *
1739 * Soft offline a page, by migration or invalidation,
1740 * without killing anything. This is for the case when
1741 * a page is not corrupted yet (so it's still valid to access),
1742 * but has had a number of corrected errors and is better taken
1743 * out.
1744 *
1745 * The actual policy on when to do that is maintained by
1746 * user space.
1747 *
1748 * This should never impact any application or cause data loss,
1749 * however it might take some time.
1750 *
1751 * This is not a 100% solution for all memory, but tries to be
1752 * ``good enough'' for the majority of memory.
1753 */
1754 int soft_offline_page(struct page *page, int flags)
1755 {
1756 int ret;
1757 unsigned long pfn = page_to_pfn(page);
1758
1759 if (PageHWPoison(page)) {
1760 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1761 if (flags & MF_COUNT_INCREASED)
1762 put_hwpoison_page(page);
1763 return -EBUSY;
1764 }
1765
1766 get_online_mems();
1767 ret = get_any_page(page, pfn, flags);
1768 put_online_mems();
1769
1770 if (ret > 0)
1771 ret = soft_offline_in_use_page(page, flags);
1772 else if (ret == 0)
1773 soft_offline_free_page(page);
1774
1775 return ret;
1776 }
This page took 0.069359 seconds and 5 git commands to generate.