~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/mm/memory-failure.c

Version: ~ [ linux-5.1.2 ] ~ [ linux-5.0.16 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.43 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.119 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.176 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.179 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.139 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.67 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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  * There are several operations here with exponential complexity because
 25  * of unsuitable VM data structures. For example the operation to map back 
 26  * from RMAP chains to processes has to walk the complete process list and 
 27  * has non linear complexity with the number. But since memory corruptions
 28  * are rare we hope to get away with this. This avoids impacting the core 
 29  * VM.
 30  */
 31 
 32 /*
 33  * Notebook:
 34  * - hugetlb needs more code
 35  * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
 36  * - pass bad pages to kdump next kernel
 37  */
 38 #include <linux/kernel.h>
 39 #include <linux/mm.h>
 40 #include <linux/page-flags.h>
 41 #include <linux/kernel-page-flags.h>
 42 #include <linux/sched.h>
 43 #include <linux/ksm.h>
 44 #include <linux/rmap.h>
 45 #include <linux/export.h>
 46 #include <linux/pagemap.h>
 47 #include <linux/swap.h>
 48 #include <linux/backing-dev.h>
 49 #include <linux/migrate.h>
 50 #include <linux/page-isolation.h>
 51 #include <linux/suspend.h>
 52 #include <linux/slab.h>
 53 #include <linux/swapops.h>
 54 #include <linux/hugetlb.h>
 55 #include <linux/memory_hotplug.h>
 56 #include <linux/mm_inline.h>
 57 #include <linux/kfifo.h>
 58 #include "internal.h"
 59 
 60 int sysctl_memory_failure_early_kill __read_mostly = 0;
 61 
 62 int sysctl_memory_failure_recovery __read_mostly = 1;
 63 
 64 atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
 65 
 66 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
 67 
 68 u32 hwpoison_filter_enable = 0;
 69 u32 hwpoison_filter_dev_major = ~0U;
 70 u32 hwpoison_filter_dev_minor = ~0U;
 71 u64 hwpoison_filter_flags_mask;
 72 u64 hwpoison_filter_flags_value;
 73 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
 74 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
 75 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
 76 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
 77 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
 78 
 79 static int hwpoison_filter_dev(struct page *p)
 80 {
 81         struct address_space *mapping;
 82         dev_t dev;
 83 
 84         if (hwpoison_filter_dev_major == ~0U &&
 85             hwpoison_filter_dev_minor == ~0U)
 86                 return 0;
 87 
 88         /*
 89          * page_mapping() does not accept slab pages.
 90          */
 91         if (PageSlab(p))
 92                 return -EINVAL;
 93 
 94         mapping = page_mapping(p);
 95         if (mapping == NULL || mapping->host == NULL)
 96                 return -EINVAL;
 97 
 98         dev = mapping->host->i_sb->s_dev;
 99         if (hwpoison_filter_dev_major != ~0U &&
100             hwpoison_filter_dev_major != MAJOR(dev))
101                 return -EINVAL;
102         if (hwpoison_filter_dev_minor != ~0U &&
103             hwpoison_filter_dev_minor != MINOR(dev))
104                 return -EINVAL;
105 
106         return 0;
107 }
108 
109 static int hwpoison_filter_flags(struct page *p)
110 {
111         if (!hwpoison_filter_flags_mask)
112                 return 0;
113 
114         if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
115                                     hwpoison_filter_flags_value)
116                 return 0;
117         else
118                 return -EINVAL;
119 }
120 
121 /*
122  * This allows stress tests to limit test scope to a collection of tasks
123  * by putting them under some memcg. This prevents killing unrelated/important
124  * processes such as /sbin/init. Note that the target task may share clean
125  * pages with init (eg. libc text), which is harmless. If the target task
126  * share _dirty_ pages with another task B, the test scheme must make sure B
127  * is also included in the memcg. At last, due to race conditions this filter
128  * can only guarantee that the page either belongs to the memcg tasks, or is
129  * a freed page.
130  */
131 #ifdef  CONFIG_MEMCG_SWAP
132 u64 hwpoison_filter_memcg;
133 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
134 static int hwpoison_filter_task(struct page *p)
135 {
136         struct mem_cgroup *mem;
137         struct cgroup_subsys_state *css;
138         unsigned long ino;
139 
140         if (!hwpoison_filter_memcg)
141                 return 0;
142 
143         mem = try_get_mem_cgroup_from_page(p);
144         if (!mem)
145                 return -EINVAL;
146 
147         css = mem_cgroup_css(mem);
148         ino = cgroup_ino(css->cgroup);
149         css_put(css);
150 
151         if (ino != hwpoison_filter_memcg)
152                 return -EINVAL;
153 
154         return 0;
155 }
156 #else
157 static int hwpoison_filter_task(struct page *p) { return 0; }
158 #endif
159 
160 int hwpoison_filter(struct page *p)
161 {
162         if (!hwpoison_filter_enable)
163                 return 0;
164 
165         if (hwpoison_filter_dev(p))
166                 return -EINVAL;
167 
168         if (hwpoison_filter_flags(p))
169                 return -EINVAL;
170 
171         if (hwpoison_filter_task(p))
172                 return -EINVAL;
173 
174         return 0;
175 }
176 #else
177 int hwpoison_filter(struct page *p)
178 {
179         return 0;
180 }
181 #endif
182 
183 EXPORT_SYMBOL_GPL(hwpoison_filter);
184 
185 /*
186  * Send all the processes who have the page mapped a signal.
187  * ``action optional'' if they are not immediately affected by the error
188  * ``action required'' if error happened in current execution context
189  */
190 static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
191                         unsigned long pfn, struct page *page, int flags)
192 {
193         struct siginfo si;
194         int ret;
195 
196         printk(KERN_ERR
197                 "MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
198                 pfn, t->comm, t->pid);
199         si.si_signo = SIGBUS;
200         si.si_errno = 0;
201         si.si_addr = (void *)addr;
202 #ifdef __ARCH_SI_TRAPNO
203         si.si_trapno = trapno;
204 #endif
205         si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
206 
207         if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
208                 si.si_code = BUS_MCEERR_AR;
209                 ret = force_sig_info(SIGBUS, &si, current);
210         } else {
211                 /*
212                  * Don't use force here, it's convenient if the signal
213                  * can be temporarily blocked.
214                  * This could cause a loop when the user sets SIGBUS
215                  * to SIG_IGN, but hopefully no one will do that?
216                  */
217                 si.si_code = BUS_MCEERR_AO;
218                 ret = send_sig_info(SIGBUS, &si, t);  /* synchronous? */
219         }
220         if (ret < 0)
221                 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
222                        t->comm, t->pid, ret);
223         return ret;
224 }
225 
226 /*
227  * When a unknown page type is encountered drain as many buffers as possible
228  * in the hope to turn the page into a LRU or free page, which we can handle.
229  */
230 void shake_page(struct page *p, int access)
231 {
232         if (!PageSlab(p)) {
233                 lru_add_drain_all();
234                 if (PageLRU(p))
235                         return;
236                 drain_all_pages(page_zone(p));
237                 if (PageLRU(p) || is_free_buddy_page(p))
238                         return;
239         }
240 
241         /*
242          * Only call shrink_node_slabs here (which would also shrink
243          * other caches) if access is not potentially fatal.
244          */
245         if (access)
246                 drop_slab_node(page_to_nid(p));
247 }
248 EXPORT_SYMBOL_GPL(shake_page);
249 
250 /*
251  * Kill all processes that have a poisoned page mapped and then isolate
252  * the page.
253  *
254  * General strategy:
255  * Find all processes having the page mapped and kill them.
256  * But we keep a page reference around so that the page is not
257  * actually freed yet.
258  * Then stash the page away
259  *
260  * There's no convenient way to get back to mapped processes
261  * from the VMAs. So do a brute-force search over all
262  * running processes.
263  *
264  * Remember that machine checks are not common (or rather
265  * if they are common you have other problems), so this shouldn't
266  * be a performance issue.
267  *
268  * Also there are some races possible while we get from the
269  * error detection to actually handle it.
270  */
271 
272 struct to_kill {
273         struct list_head nd;
274         struct task_struct *tsk;
275         unsigned long addr;
276         char addr_valid;
277 };
278 
279 /*
280  * Failure handling: if we can't find or can't kill a process there's
281  * not much we can do.  We just print a message and ignore otherwise.
282  */
283 
284 /*
285  * Schedule a process for later kill.
286  * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
287  * TBD would GFP_NOIO be enough?
288  */
289 static void add_to_kill(struct task_struct *tsk, struct page *p,
290                        struct vm_area_struct *vma,
291                        struct list_head *to_kill,
292                        struct to_kill **tkc)
293 {
294         struct to_kill *tk;
295 
296         if (*tkc) {
297                 tk = *tkc;
298                 *tkc = NULL;
299         } else {
300                 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
301                 if (!tk) {
302                         printk(KERN_ERR
303                 "MCE: Out of memory while machine check handling\n");
304                         return;
305                 }
306         }
307         tk->addr = page_address_in_vma(p, vma);
308         tk->addr_valid = 1;
309 
310         /*
311          * In theory we don't have to kill when the page was
312          * munmaped. But it could be also a mremap. Since that's
313          * likely very rare kill anyways just out of paranoia, but use
314          * a SIGKILL because the error is not contained anymore.
315          */
316         if (tk->addr == -EFAULT) {
317                 pr_info("MCE: Unable to find user space address %lx in %s\n",
318                         page_to_pfn(p), tsk->comm);
319                 tk->addr_valid = 0;
320         }
321         get_task_struct(tsk);
322         tk->tsk = tsk;
323         list_add_tail(&tk->nd, to_kill);
324 }
325 
326 /*
327  * Kill the processes that have been collected earlier.
328  *
329  * Only do anything when DOIT is set, otherwise just free the list
330  * (this is used for clean pages which do not need killing)
331  * Also when FAIL is set do a force kill because something went
332  * wrong earlier.
333  */
334 static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
335                           int fail, struct page *page, unsigned long pfn,
336                           int flags)
337 {
338         struct to_kill *tk, *next;
339 
340         list_for_each_entry_safe (tk, next, to_kill, nd) {
341                 if (forcekill) {
342                         /*
343                          * In case something went wrong with munmapping
344                          * make sure the process doesn't catch the
345                          * signal and then access the memory. Just kill it.
346                          */
347                         if (fail || tk->addr_valid == 0) {
348                                 printk(KERN_ERR
349                 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
350                                         pfn, tk->tsk->comm, tk->tsk->pid);
351                                 force_sig(SIGKILL, tk->tsk);
352                         }
353 
354                         /*
355                          * In theory the process could have mapped
356                          * something else on the address in-between. We could
357                          * check for that, but we need to tell the
358                          * process anyways.
359                          */
360                         else if (kill_proc(tk->tsk, tk->addr, trapno,
361                                               pfn, page, flags) < 0)
362                                 printk(KERN_ERR
363                 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
364                                         pfn, tk->tsk->comm, tk->tsk->pid);
365                 }
366                 put_task_struct(tk->tsk);
367                 kfree(tk);
368         }
369 }
370 
371 /*
372  * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
373  * on behalf of the thread group. Return task_struct of the (first found)
374  * dedicated thread if found, and return NULL otherwise.
375  *
376  * We already hold read_lock(&tasklist_lock) in the caller, so we don't
377  * have to call rcu_read_lock/unlock() in this function.
378  */
379 static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
380 {
381         struct task_struct *t;
382 
383         for_each_thread(tsk, t)
384                 if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
385                         return t;
386         return NULL;
387 }
388 
389 /*
390  * Determine whether a given process is "early kill" process which expects
391  * to be signaled when some page under the process is hwpoisoned.
392  * Return task_struct of the dedicated thread (main thread unless explicitly
393  * specified) if the process is "early kill," and otherwise returns NULL.
394  */
395 static struct task_struct *task_early_kill(struct task_struct *tsk,
396                                            int force_early)
397 {
398         struct task_struct *t;
399         if (!tsk->mm)
400                 return NULL;
401         if (force_early)
402                 return tsk;
403         t = find_early_kill_thread(tsk);
404         if (t)
405                 return t;
406         if (sysctl_memory_failure_early_kill)
407                 return tsk;
408         return NULL;
409 }
410 
411 /*
412  * Collect processes when the error hit an anonymous page.
413  */
414 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
415                               struct to_kill **tkc, int force_early)
416 {
417         struct vm_area_struct *vma;
418         struct task_struct *tsk;
419         struct anon_vma *av;
420         pgoff_t pgoff;
421 
422         av = page_lock_anon_vma_read(page);
423         if (av == NULL) /* Not actually mapped anymore */
424                 return;
425 
426         pgoff = page_to_pgoff(page);
427         read_lock(&tasklist_lock);
428         for_each_process (tsk) {
429                 struct anon_vma_chain *vmac;
430                 struct task_struct *t = task_early_kill(tsk, force_early);
431 
432                 if (!t)
433                         continue;
434                 anon_vma_interval_tree_foreach(vmac, &av->rb_root,
435                                                pgoff, pgoff) {
436                         vma = vmac->vma;
437                         if (!page_mapped_in_vma(page, vma))
438                                 continue;
439                         if (vma->vm_mm == t->mm)
440                                 add_to_kill(t, page, vma, to_kill, tkc);
441                 }
442         }
443         read_unlock(&tasklist_lock);
444         page_unlock_anon_vma_read(av);
445 }
446 
447 /*
448  * Collect processes when the error hit a file mapped page.
449  */
450 static void collect_procs_file(struct page *page, struct list_head *to_kill,
451                               struct to_kill **tkc, int force_early)
452 {
453         struct vm_area_struct *vma;
454         struct task_struct *tsk;
455         struct address_space *mapping = page->mapping;
456 
457         i_mmap_lock_read(mapping);
458         read_lock(&tasklist_lock);
459         for_each_process(tsk) {
460                 pgoff_t pgoff = page_to_pgoff(page);
461                 struct task_struct *t = task_early_kill(tsk, force_early);
462 
463                 if (!t)
464                         continue;
465                 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
466                                       pgoff) {
467                         /*
468                          * Send early kill signal to tasks where a vma covers
469                          * the page but the corrupted page is not necessarily
470                          * mapped it in its pte.
471                          * Assume applications who requested early kill want
472                          * to be informed of all such data corruptions.
473                          */
474                         if (vma->vm_mm == t->mm)
475                                 add_to_kill(t, page, vma, to_kill, tkc);
476                 }
477         }
478         read_unlock(&tasklist_lock);
479         i_mmap_unlock_read(mapping);
480 }
481 
482 /*
483  * Collect the processes who have the corrupted page mapped to kill.
484  * This is done in two steps for locking reasons.
485  * First preallocate one tokill structure outside the spin locks,
486  * so that we can kill at least one process reasonably reliable.
487  */
488 static void collect_procs(struct page *page, struct list_head *tokill,
489                                 int force_early)
490 {
491         struct to_kill *tk;
492 
493         if (!page->mapping)
494                 return;
495 
496         tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
497         if (!tk)
498                 return;
499         if (PageAnon(page))
500                 collect_procs_anon(page, tokill, &tk, force_early);
501         else
502                 collect_procs_file(page, tokill, &tk, force_early);
503         kfree(tk);
504 }
505 
506 /*
507  * Error handlers for various types of pages.
508  */
509 
510 enum outcome {
511         IGNORED,        /* Error: cannot be handled */
512         FAILED,         /* Error: handling failed */
513         DELAYED,        /* Will be handled later */
514         RECOVERED,      /* Successfully recovered */
515 };
516 
517 static const char *action_name[] = {
518         [IGNORED] = "Ignored",
519         [FAILED] = "Failed",
520         [DELAYED] = "Delayed",
521         [RECOVERED] = "Recovered",
522 };
523 
524 /*
525  * XXX: It is possible that a page is isolated from LRU cache,
526  * and then kept in swap cache or failed to remove from page cache.
527  * The page count will stop it from being freed by unpoison.
528  * Stress tests should be aware of this memory leak problem.
529  */
530 static int delete_from_lru_cache(struct page *p)
531 {
532         if (!isolate_lru_page(p)) {
533                 /*
534                  * Clear sensible page flags, so that the buddy system won't
535                  * complain when the page is unpoison-and-freed.
536                  */
537                 ClearPageActive(p);
538                 ClearPageUnevictable(p);
539                 /*
540                  * drop the page count elevated by isolate_lru_page()
541                  */
542                 page_cache_release(p);
543                 return 0;
544         }
545         return -EIO;
546 }
547 
548 /*
549  * Error hit kernel page.
550  * Do nothing, try to be lucky and not touch this instead. For a few cases we
551  * could be more sophisticated.
552  */
553 static int me_kernel(struct page *p, unsigned long pfn)
554 {
555         return IGNORED;
556 }
557 
558 /*
559  * Page in unknown state. Do nothing.
560  */
561 static int me_unknown(struct page *p, unsigned long pfn)
562 {
563         printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
564         return FAILED;
565 }
566 
567 /*
568  * Clean (or cleaned) page cache page.
569  */
570 static int me_pagecache_clean(struct page *p, unsigned long pfn)
571 {
572         int err;
573         int ret = FAILED;
574         struct address_space *mapping;
575 
576         delete_from_lru_cache(p);
577 
578         /*
579          * For anonymous pages we're done the only reference left
580          * should be the one m_f() holds.
581          */
582         if (PageAnon(p))
583                 return RECOVERED;
584 
585         /*
586          * Now truncate the page in the page cache. This is really
587          * more like a "temporary hole punch"
588          * Don't do this for block devices when someone else
589          * has a reference, because it could be file system metadata
590          * and that's not safe to truncate.
591          */
592         mapping = page_mapping(p);
593         if (!mapping) {
594                 /*
595                  * Page has been teared down in the meanwhile
596                  */
597                 return FAILED;
598         }
599 
600         /*
601          * Truncation is a bit tricky. Enable it per file system for now.
602          *
603          * Open: to take i_mutex or not for this? Right now we don't.
604          */
605         if (mapping->a_ops->error_remove_page) {
606                 err = mapping->a_ops->error_remove_page(mapping, p);
607                 if (err != 0) {
608                         printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
609                                         pfn, err);
610                 } else if (page_has_private(p) &&
611                                 !try_to_release_page(p, GFP_NOIO)) {
612                         pr_info("MCE %#lx: failed to release buffers\n", pfn);
613                 } else {
614                         ret = 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 = RECOVERED;
623                 else
624                         printk(KERN_INFO "MCE %#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 DELAYED;
709         else
710                 return 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 RECOVERED;
719         else
720                 return 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          * We can safely recover from error on free or reserved (i.e.
735          * not in-use) hugepage by dequeuing it from freelist.
736          * To check whether a hugepage is in-use or not, we can't use
737          * page->lru because it can be used in other hugepage operations,
738          * such as __unmap_hugepage_range() and gather_surplus_pages().
739          * So instead we use page_mapping() and PageAnon().
740          * We assume that this function is called with page lock held,
741          * so there is no race between isolation and mapping/unmapping.
742          */
743         if (!(page_mapping(hpage) || PageAnon(hpage))) {
744                 res = dequeue_hwpoisoned_huge_page(hpage);
745                 if (!res)
746                         return RECOVERED;
747         }
748         return DELAYED;
749 }
750 
751 /*
752  * Various page states we can handle.
753  *
754  * A page state is defined by its current page->flags bits.
755  * The table matches them in order and calls the right handler.
756  *
757  * This is quite tricky because we can access page at any time
758  * in its live cycle, so all accesses have to be extremely careful.
759  *
760  * This is not complete. More states could be added.
761  * For any missing state don't attempt recovery.
762  */
763 
764 #define dirty           (1UL << PG_dirty)
765 #define sc              (1UL << PG_swapcache)
766 #define unevict         (1UL << PG_unevictable)
767 #define mlock           (1UL << PG_mlocked)
768 #define writeback       (1UL << PG_writeback)
769 #define lru             (1UL << PG_lru)
770 #define swapbacked      (1UL << PG_swapbacked)
771 #define head            (1UL << PG_head)
772 #define tail            (1UL << PG_tail)
773 #define compound        (1UL << PG_compound)
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         char *msg;
781         int (*action)(struct page *p, unsigned long pfn);
782 } error_states[] = {
783         { reserved,     reserved,       "reserved 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,           "kernel slab",  me_kernel },
795 
796 #ifdef CONFIG_PAGEFLAGS_EXTENDED
797         { head,         head,           "huge",         me_huge_page },
798         { tail,         tail,           "huge",         me_huge_page },
799 #else
800         { compound,     compound,       "huge",         me_huge_page },
801 #endif
802 
803         { sc|dirty,     sc|dirty,       "dirty swapcache",      me_swapcache_dirty },
804         { sc|dirty,     sc,             "clean swapcache",      me_swapcache_clean },
805 
806         { mlock|dirty,  mlock|dirty,    "dirty mlocked LRU",    me_pagecache_dirty },
807         { mlock|dirty,  mlock,          "clean mlocked LRU",    me_pagecache_clean },
808 
809         { unevict|dirty, unevict|dirty, "dirty unevictable LRU", me_pagecache_dirty },
810         { unevict|dirty, unevict,       "clean unevictable LRU", me_pagecache_clean },
811 
812         { lru|dirty,    lru|dirty,      "dirty LRU",    me_pagecache_dirty },
813         { lru|dirty,    lru,            "clean LRU",    me_pagecache_clean },
814 
815         /*
816          * Catchall entry: must be at end.
817          */
818         { 0,            0,              "unknown page state",   me_unknown },
819 };
820 
821 #undef dirty
822 #undef sc
823 #undef unevict
824 #undef mlock
825 #undef writeback
826 #undef lru
827 #undef swapbacked
828 #undef head
829 #undef tail
830 #undef compound
831 #undef slab
832 #undef reserved
833 
834 /*
835  * "Dirty/Clean" indication is not 100% accurate due to the possibility of
836  * setting PG_dirty outside page lock. See also comment above set_page_dirty().
837  */
838 static void action_result(unsigned long pfn, char *msg, int result)
839 {
840         pr_err("MCE %#lx: %s page recovery: %s\n",
841                 pfn, msg, action_name[result]);
842 }
843 
844 static int page_action(struct page_state *ps, struct page *p,
845                         unsigned long pfn)
846 {
847         int result;
848         int count;
849 
850         result = ps->action(p, pfn);
851 
852         count = page_count(p) - 1;
853         if (ps->action == me_swapcache_dirty && result == DELAYED)
854                 count--;
855         if (count != 0) {
856                 printk(KERN_ERR
857                        "MCE %#lx: %s page still referenced by %d users\n",
858                        pfn, ps->msg, count);
859                 result = FAILED;
860         }
861         action_result(pfn, ps->msg, result);
862 
863         /* Could do more checks here if page looks ok */
864         /*
865          * Could adjust zone counters here to correct for the missing page.
866          */
867 
868         return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
869 }
870 
871 /*
872  * Do all that is necessary to remove user space mappings. Unmap
873  * the pages and send SIGBUS to the processes if the data was dirty.
874  */
875 static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
876                                   int trapno, int flags, struct page **hpagep)
877 {
878         enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
879         struct address_space *mapping;
880         LIST_HEAD(tokill);
881         int ret;
882         int kill = 1, forcekill;
883         struct page *hpage = *hpagep;
884         struct page *ppage;
885 
886         /*
887          * Here we are interested only in user-mapped pages, so skip any
888          * other types of pages.
889          */
890         if (PageReserved(p) || PageSlab(p))
891                 return SWAP_SUCCESS;
892         if (!(PageLRU(hpage) || PageHuge(p)))
893                 return SWAP_SUCCESS;
894 
895         /*
896          * This check implies we don't kill processes if their pages
897          * are in the swap cache early. Those are always late kills.
898          */
899         if (!page_mapped(hpage))
900                 return SWAP_SUCCESS;
901 
902         if (PageKsm(p)) {
903                 pr_err("MCE %#lx: can't handle KSM pages.\n", pfn);
904                 return SWAP_FAIL;
905         }
906 
907         if (PageSwapCache(p)) {
908                 printk(KERN_ERR
909                        "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
910                 ttu |= TTU_IGNORE_HWPOISON;
911         }
912 
913         /*
914          * Propagate the dirty bit from PTEs to struct page first, because we
915          * need this to decide if we should kill or just drop the page.
916          * XXX: the dirty test could be racy: set_page_dirty() may not always
917          * be called inside page lock (it's recommended but not enforced).
918          */
919         mapping = page_mapping(hpage);
920         if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
921             mapping_cap_writeback_dirty(mapping)) {
922                 if (page_mkclean(hpage)) {
923                         SetPageDirty(hpage);
924                 } else {
925                         kill = 0;
926                         ttu |= TTU_IGNORE_HWPOISON;
927                         printk(KERN_INFO
928         "MCE %#lx: corrupted page was clean: dropped without side effects\n",
929                                 pfn);
930                 }
931         }
932 
933         /*
934          * ppage: poisoned page
935          *   if p is regular page(4k page)
936          *        ppage == real poisoned page;
937          *   else p is hugetlb or THP, ppage == head page.
938          */
939         ppage = hpage;
940 
941         if (PageTransHuge(hpage)) {
942                 /*
943                  * Verify that this isn't a hugetlbfs head page, the check for
944                  * PageAnon is just for avoid tripping a split_huge_page
945                  * internal debug check, as split_huge_page refuses to deal with
946                  * anything that isn't an anon page. PageAnon can't go away fro
947                  * under us because we hold a refcount on the hpage, without a
948                  * refcount on the hpage. split_huge_page can't be safely called
949                  * in the first place, having a refcount on the tail isn't
950                  * enough * to be safe.
951                  */
952                 if (!PageHuge(hpage) && PageAnon(hpage)) {
953                         if (unlikely(split_huge_page(hpage))) {
954                                 /*
955                                  * FIXME: if splitting THP is failed, it is
956                                  * better to stop the following operation rather
957                                  * than causing panic by unmapping. System might
958                                  * survive if the page is freed later.
959                                  */
960                                 printk(KERN_INFO
961                                         "MCE %#lx: failed to split THP\n", pfn);
962 
963                                 BUG_ON(!PageHWPoison(p));
964                                 return SWAP_FAIL;
965                         }
966                         /*
967                          * We pinned the head page for hwpoison handling,
968                          * now we split the thp and we are interested in
969                          * the hwpoisoned raw page, so move the refcount
970                          * to it. Similarly, page lock is shifted.
971                          */
972                         if (hpage != p) {
973                                 if (!(flags & MF_COUNT_INCREASED)) {
974                                         put_page(hpage);
975                                         get_page(p);
976                                 }
977                                 lock_page(p);
978                                 unlock_page(hpage);
979                                 *hpagep = p;
980                         }
981                         /* THP is split, so ppage should be the real poisoned page. */
982                         ppage = p;
983                 }
984         }
985 
986         /*
987          * First collect all the processes that have the page
988          * mapped in dirty form.  This has to be done before try_to_unmap,
989          * because ttu takes the rmap data structures down.
990          *
991          * Error handling: We ignore errors here because
992          * there's nothing that can be done.
993          */
994         if (kill)
995                 collect_procs(ppage, &tokill, flags & MF_ACTION_REQUIRED);
996 
997         ret = try_to_unmap(ppage, ttu);
998         if (ret != SWAP_SUCCESS)
999                 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
1000                                 pfn, page_mapcount(ppage));
1001 
1002         /*
1003          * Now that the dirty bit has been propagated to the
1004          * struct page and all unmaps done we can decide if
1005          * killing is needed or not.  Only kill when the page
1006          * was dirty or the process is not restartable,
1007          * otherwise the tokill list is merely
1008          * freed.  When there was a problem unmapping earlier
1009          * use a more force-full uncatchable kill to prevent
1010          * any accesses to the poisoned memory.
1011          */
1012         forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL);
1013         kill_procs(&tokill, forcekill, trapno,
1014                       ret != SWAP_SUCCESS, p, pfn, flags);
1015 
1016         return ret;
1017 }
1018 
1019 static void set_page_hwpoison_huge_page(struct page *hpage)
1020 {
1021         int i;
1022         int nr_pages = 1 << compound_order(hpage);
1023         for (i = 0; i < nr_pages; i++)
1024                 SetPageHWPoison(hpage + i);
1025 }
1026 
1027 static void clear_page_hwpoison_huge_page(struct page *hpage)
1028 {
1029         int i;
1030         int nr_pages = 1 << compound_order(hpage);
1031         for (i = 0; i < nr_pages; i++)
1032                 ClearPageHWPoison(hpage + i);
1033 }
1034 
1035 /**
1036  * memory_failure - Handle memory failure of a page.
1037  * @pfn: Page Number of the corrupted page
1038  * @trapno: Trap number reported in the signal to user space.
1039  * @flags: fine tune action taken
1040  *
1041  * This function is called by the low level machine check code
1042  * of an architecture when it detects hardware memory corruption
1043  * of a page. It tries its best to recover, which includes
1044  * dropping pages, killing processes etc.
1045  *
1046  * The function is primarily of use for corruptions that
1047  * happen outside the current execution context (e.g. when
1048  * detected by a background scrubber)
1049  *
1050  * Must run in process context (e.g. a work queue) with interrupts
1051  * enabled and no spinlocks hold.
1052  */
1053 int memory_failure(unsigned long pfn, int trapno, int flags)
1054 {
1055         struct page_state *ps;
1056         struct page *p;
1057         struct page *hpage;
1058         int res;
1059         unsigned int nr_pages;
1060         unsigned long page_flags;
1061 
1062         if (!sysctl_memory_failure_recovery)
1063                 panic("Memory failure from trap %d on page %lx", trapno, pfn);
1064 
1065         if (!pfn_valid(pfn)) {
1066                 printk(KERN_ERR
1067                        "MCE %#lx: memory outside kernel control\n",
1068                        pfn);
1069                 return -ENXIO;
1070         }
1071 
1072         p = pfn_to_page(pfn);
1073         hpage = compound_head(p);
1074         if (TestSetPageHWPoison(p)) {
1075                 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
1076                 return 0;
1077         }
1078 
1079         /*
1080          * Currently errors on hugetlbfs pages are measured in hugepage units,
1081          * so nr_pages should be 1 << compound_order.  OTOH when errors are on
1082          * transparent hugepages, they are supposed to be split and error
1083          * measurement is done in normal page units.  So nr_pages should be one
1084          * in this case.
1085          */
1086         if (PageHuge(p))
1087                 nr_pages = 1 << compound_order(hpage);
1088         else /* normal page or thp */
1089                 nr_pages = 1;
1090         atomic_long_add(nr_pages, &num_poisoned_pages);
1091 
1092         /*
1093          * We need/can do nothing about count=0 pages.
1094          * 1) it's a free page, and therefore in safe hand:
1095          *    prep_new_page() will be the gate keeper.
1096          * 2) it's a free hugepage, which is also safe:
1097          *    an affected hugepage will be dequeued from hugepage freelist,
1098          *    so there's no concern about reusing it ever after.
1099          * 3) it's part of a non-compound high order page.
1100          *    Implies some kernel user: cannot stop them from
1101          *    R/W the page; let's pray that the page has been
1102          *    used and will be freed some time later.
1103          * In fact it's dangerous to directly bump up page count from 0,
1104          * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1105          */
1106         if (!(flags & MF_COUNT_INCREASED) &&
1107                 !get_page_unless_zero(hpage)) {
1108                 if (is_free_buddy_page(p)) {
1109                         action_result(pfn, "free buddy", DELAYED);
1110                         return 0;
1111                 } else if (PageHuge(hpage)) {
1112                         /*
1113                          * Check "filter hit" and "race with other subpage."
1114                          */
1115                         lock_page(hpage);
1116                         if (PageHWPoison(hpage)) {
1117                                 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1118                                     || (p != hpage && TestSetPageHWPoison(hpage))) {
1119                                         atomic_long_sub(nr_pages, &num_poisoned_pages);
1120                                         unlock_page(hpage);
1121                                         return 0;
1122                                 }
1123                         }
1124                         set_page_hwpoison_huge_page(hpage);
1125                         res = dequeue_hwpoisoned_huge_page(hpage);
1126                         action_result(pfn, "free huge",
1127                                       res ? IGNORED : DELAYED);
1128                         unlock_page(hpage);
1129                         return res;
1130                 } else {
1131                         action_result(pfn, "high order kernel", IGNORED);
1132                         return -EBUSY;
1133                 }
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 __set_page_locked()
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(hpage))
1146                         shake_page(hpage, 0);
1147                 if (!PageLRU(hpage)) {
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, "free buddy", DELAYED);
1154                                 else
1155                                         action_result(pfn, "free buddy, 2nd try", DELAYED);
1156                                 return 0;
1157                         }
1158                 }
1159         }
1160 
1161         lock_page(hpage);
1162 
1163         /*
1164          * The page could have changed compound pages during the locking.
1165          * If this happens just bail out.
1166          */
1167         if (compound_head(p) != hpage) {
1168                 action_result(pfn, "different compound page after locking", IGNORED);
1169                 res = -EBUSY;
1170                 goto out;
1171         }
1172 
1173         /*
1174          * We use page flags to determine what action should be taken, but
1175          * the flags can be modified by the error containment action.  One
1176          * example is an mlocked page, where PG_mlocked is cleared by
1177          * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1178          * correctly, we save a copy of the page flags at this time.
1179          */
1180         page_flags = p->flags;
1181 
1182         /*
1183          * unpoison always clear PG_hwpoison inside page lock
1184          */
1185         if (!PageHWPoison(p)) {
1186                 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1187                 atomic_long_sub(nr_pages, &num_poisoned_pages);
1188                 put_page(hpage);
1189                 res = 0;
1190                 goto out;
1191         }
1192         if (hwpoison_filter(p)) {
1193                 if (TestClearPageHWPoison(p))
1194                         atomic_long_sub(nr_pages, &num_poisoned_pages);
1195                 unlock_page(hpage);
1196                 put_page(hpage);
1197                 return 0;
1198         }
1199 
1200         if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p))
1201                 goto identify_page_state;
1202 
1203         /*
1204          * For error on the tail page, we should set PG_hwpoison
1205          * on the head page to show that the hugepage is hwpoisoned
1206          */
1207         if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1208                 action_result(pfn, "hugepage already hardware poisoned",
1209                                 IGNORED);
1210                 unlock_page(hpage);
1211                 put_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, "unmapping failed", 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, "already truncated LRU", 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 /**
1371  * unpoison_memory - Unpoison a previously poisoned page
1372  * @pfn: Page number of the to be unpoisoned page
1373  *
1374  * Software-unpoison a page that has been poisoned by
1375  * memory_failure() earlier.
1376  *
1377  * This is only done on the software-level, so it only works
1378  * for linux injected failures, not real hardware failures
1379  *
1380  * Returns 0 for success, otherwise -errno.
1381  */
1382 int unpoison_memory(unsigned long pfn)
1383 {
1384         struct page *page;
1385         struct page *p;
1386         int freeit = 0;
1387         unsigned int nr_pages;
1388 
1389         if (!pfn_valid(pfn))
1390                 return -ENXIO;
1391 
1392         p = pfn_to_page(pfn);
1393         page = compound_head(p);
1394 
1395         if (!PageHWPoison(p)) {
1396                 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
1397                 return 0;
1398         }
1399 
1400         /*
1401          * unpoison_memory() can encounter thp only when the thp is being
1402          * worked by memory_failure() and the page lock is not held yet.
1403          * In such case, we yield to memory_failure() and make unpoison fail.
1404          */
1405         if (!PageHuge(page) && PageTransHuge(page)) {
1406                 pr_info("MCE: Memory failure is now running on %#lx\n", pfn);
1407                         return 0;
1408         }
1409 
1410         nr_pages = 1 << compound_order(page);
1411 
1412         if (!get_page_unless_zero(page)) {
1413                 /*
1414                  * Since HWPoisoned hugepage should have non-zero refcount,
1415                  * race between memory failure and unpoison seems to happen.
1416                  * In such case unpoison fails and memory failure runs
1417                  * to the end.
1418                  */
1419                 if (PageHuge(page)) {
1420                         pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1421                         return 0;
1422                 }
1423                 if (TestClearPageHWPoison(p))
1424                         atomic_long_dec(&num_poisoned_pages);
1425                 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
1426                 return 0;
1427         }
1428 
1429         lock_page(page);
1430         /*
1431          * This test is racy because PG_hwpoison is set outside of page lock.
1432          * That's acceptable because that won't trigger kernel panic. Instead,
1433          * the PG_hwpoison page will be caught and isolated on the entrance to
1434          * the free buddy page pool.
1435          */
1436         if (TestClearPageHWPoison(page)) {
1437                 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
1438                 atomic_long_sub(nr_pages, &num_poisoned_pages);
1439                 freeit = 1;
1440                 if (PageHuge(page))
1441                         clear_page_hwpoison_huge_page(page);
1442         }
1443         unlock_page(page);
1444 
1445         put_page(page);
1446         if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
1447                 put_page(page);
1448 
1449         return 0;
1450 }
1451 EXPORT_SYMBOL(unpoison_memory);
1452 
1453 static struct page *new_page(struct page *p, unsigned long private, int **x)
1454 {
1455         int nid = page_to_nid(p);
1456         if (PageHuge(p))
1457                 return alloc_huge_page_node(page_hstate(compound_head(p)),
1458                                                    nid);
1459         else
1460                 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1461 }
1462 
1463 /*
1464  * Safely get reference count of an arbitrary page.
1465  * Returns 0 for a free page, -EIO for a zero refcount page
1466  * that is not free, and 1 for any other page type.
1467  * For 1 the page is returned with increased page count, otherwise not.
1468  */
1469 static int __get_any_page(struct page *p, unsigned long pfn, int flags)
1470 {
1471         int ret;
1472 
1473         if (flags & MF_COUNT_INCREASED)
1474                 return 1;
1475 
1476         /*
1477          * When the target page is a free hugepage, just remove it
1478          * from free hugepage list.
1479          */
1480         if (!get_page_unless_zero(compound_head(p))) {
1481                 if (PageHuge(p)) {
1482                         pr_info("%s: %#lx free huge page\n", __func__, pfn);
1483                         ret = 0;
1484                 } else if (is_free_buddy_page(p)) {
1485                         pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1486                         ret = 0;
1487                 } else {
1488                         pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1489                                 __func__, pfn, p->flags);
1490                         ret = -EIO;
1491                 }
1492         } else {
1493                 /* Not a free page */
1494                 ret = 1;
1495         }
1496         return ret;
1497 }
1498 
1499 static int get_any_page(struct page *page, unsigned long pfn, int flags)
1500 {
1501         int ret = __get_any_page(page, pfn, flags);
1502 
1503         if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1504                 /*
1505                  * Try to free it.
1506                  */
1507                 put_page(page);
1508                 shake_page(page, 1);
1509 
1510                 /*
1511                  * Did it turn free?
1512                  */
1513                 ret = __get_any_page(page, pfn, 0);
1514                 if (!PageLRU(page)) {
1515                         pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1516                                 pfn, page->flags);
1517                         return -EIO;
1518                 }
1519         }
1520         return ret;
1521 }
1522 
1523 static int soft_offline_huge_page(struct page *page, int flags)
1524 {
1525         int ret;
1526         unsigned long pfn = page_to_pfn(page);
1527         struct page *hpage = compound_head(page);
1528         LIST_HEAD(pagelist);
1529 
1530         /*
1531          * This double-check of PageHWPoison is to avoid the race with
1532          * memory_failure(). See also comment in __soft_offline_page().
1533          */
1534         lock_page(hpage);
1535         if (PageHWPoison(hpage)) {
1536                 unlock_page(hpage);
1537                 put_page(hpage);
1538                 pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1539                 return -EBUSY;
1540         }
1541         unlock_page(hpage);
1542 
1543         /* Keep page count to indicate a given hugepage is isolated. */
1544         list_move(&hpage->lru, &pagelist);
1545         ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1546                                 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1547         if (ret) {
1548                 pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1549                         pfn, ret, page->flags);
1550                 /*
1551                  * We know that soft_offline_huge_page() tries to migrate
1552                  * only one hugepage pointed to by hpage, so we need not
1553                  * run through the pagelist here.
1554                  */
1555                 putback_active_hugepage(hpage);
1556                 if (ret > 0)
1557                         ret = -EIO;
1558         } else {
1559                 /* overcommit hugetlb page will be freed to buddy */
1560                 if (PageHuge(page)) {
1561                         set_page_hwpoison_huge_page(hpage);
1562                         dequeue_hwpoisoned_huge_page(hpage);
1563                         atomic_long_add(1 << compound_order(hpage),
1564                                         &num_poisoned_pages);
1565                 } else {
1566                         SetPageHWPoison(page);
1567                         atomic_long_inc(&num_poisoned_pages);
1568                 }
1569         }
1570         return ret;
1571 }
1572 
1573 static int __soft_offline_page(struct page *page, int flags)
1574 {
1575         int ret;
1576         unsigned long pfn = page_to_pfn(page);
1577 
1578         /*
1579          * Check PageHWPoison again inside page lock because PageHWPoison
1580          * is set by memory_failure() outside page lock. Note that
1581          * memory_failure() also double-checks PageHWPoison inside page lock,
1582          * so there's no race between soft_offline_page() and memory_failure().
1583          */
1584         lock_page(page);
1585         wait_on_page_writeback(page);
1586         if (PageHWPoison(page)) {
1587                 unlock_page(page);
1588                 put_page(page);
1589                 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1590                 return -EBUSY;
1591         }
1592         /*
1593          * Try to invalidate first. This should work for
1594          * non dirty unmapped page cache pages.
1595          */
1596         ret = invalidate_inode_page(page);
1597         unlock_page(page);
1598         /*
1599          * RED-PEN would be better to keep it isolated here, but we
1600          * would need to fix isolation locking first.
1601          */
1602         if (ret == 1) {
1603                 put_page(page);
1604                 pr_info("soft_offline: %#lx: invalidated\n", pfn);
1605                 SetPageHWPoison(page);
1606                 atomic_long_inc(&num_poisoned_pages);
1607                 return 0;
1608         }
1609 
1610         /*
1611          * Simple invalidation didn't work.
1612          * Try to migrate to a new page instead. migrate.c
1613          * handles a large number of cases for us.
1614          */
1615         ret = isolate_lru_page(page);
1616         /*
1617          * Drop page reference which is came from get_any_page()
1618          * successful isolate_lru_page() already took another one.
1619          */
1620         put_page(page);
1621         if (!ret) {
1622                 LIST_HEAD(pagelist);
1623                 inc_zone_page_state(page, NR_ISOLATED_ANON +
1624                                         page_is_file_cache(page));
1625                 list_add(&page->lru, &pagelist);
1626                 ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1627                                         MIGRATE_SYNC, MR_MEMORY_FAILURE);
1628                 if (ret) {
1629                         if (!list_empty(&pagelist)) {
1630                                 list_del(&page->lru);
1631                                 dec_zone_page_state(page, NR_ISOLATED_ANON +
1632                                                 page_is_file_cache(page));
1633                                 putback_lru_page(page);
1634                         }
1635 
1636                         pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1637                                 pfn, ret, page->flags);
1638                         if (ret > 0)
1639                                 ret = -EIO;
1640                 } else {
1641                         /*
1642                          * After page migration succeeds, the source page can
1643                          * be trapped in pagevec and actual freeing is delayed.
1644                          * Freeing code works differently based on PG_hwpoison,
1645                          * so there's a race. We need to make sure that the
1646                          * source page should be freed back to buddy before
1647                          * setting PG_hwpoison.
1648                          */
1649                         if (!is_free_buddy_page(page))
1650                                 drain_all_pages(page_zone(page));
1651                         SetPageHWPoison(page);
1652                         if (!is_free_buddy_page(page))
1653                                 pr_info("soft offline: %#lx: page leaked\n",
1654                                         pfn);
1655                         atomic_long_inc(&num_poisoned_pages);
1656                 }
1657         } else {
1658                 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1659                         pfn, ret, page_count(page), page->flags);
1660         }
1661         return ret;
1662 }
1663 
1664 /**
1665  * soft_offline_page - Soft offline a page.
1666  * @page: page to offline
1667  * @flags: flags. Same as memory_failure().
1668  *
1669  * Returns 0 on success, otherwise negated errno.
1670  *
1671  * Soft offline a page, by migration or invalidation,
1672  * without killing anything. This is for the case when
1673  * a page is not corrupted yet (so it's still valid to access),
1674  * but has had a number of corrected errors and is better taken
1675  * out.
1676  *
1677  * The actual policy on when to do that is maintained by
1678  * user space.
1679  *
1680  * This should never impact any application or cause data loss,
1681  * however it might take some time.
1682  *
1683  * This is not a 100% solution for all memory, but tries to be
1684  * ``good enough'' for the majority of memory.
1685  */
1686 int soft_offline_page(struct page *page, int flags)
1687 {
1688         int ret;
1689         unsigned long pfn = page_to_pfn(page);
1690         struct page *hpage = compound_head(page);
1691 
1692         if (PageHWPoison(page)) {
1693                 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1694                 return -EBUSY;
1695         }
1696         if (!PageHuge(page) && PageTransHuge(hpage)) {
1697                 if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1698                         pr_info("soft offline: %#lx: failed to split THP\n",
1699                                 pfn);
1700                         return -EBUSY;
1701                 }
1702         }
1703 
1704         get_online_mems();
1705 
1706         /*
1707          * Isolate the page, so that it doesn't get reallocated if it
1708          * was free. This flag should be kept set until the source page
1709          * is freed and PG_hwpoison on it is set.
1710          */
1711         if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
1712                 set_migratetype_isolate(page, true);
1713 
1714         ret = get_any_page(page, pfn, flags);
1715         put_online_mems();
1716         if (ret > 0) { /* for in-use pages */
1717                 if (PageHuge(page))
1718                         ret = soft_offline_huge_page(page, flags);
1719                 else
1720                         ret = __soft_offline_page(page, flags);
1721         } else if (ret == 0) { /* for free pages */
1722                 if (PageHuge(page)) {
1723                         set_page_hwpoison_huge_page(hpage);
1724                         if (!dequeue_hwpoisoned_huge_page(hpage))
1725                                 atomic_long_add(1 << compound_order(hpage),
1726                                         &num_poisoned_pages);
1727                 } else {
1728                         if (!TestSetPageHWPoison(page))
1729                                 atomic_long_inc(&num_poisoned_pages);
1730                 }
1731         }
1732         unset_migratetype_isolate(page, MIGRATE_MOVABLE);
1733         return ret;
1734 }
1735 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp