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

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

Version: ~ [ linux-5.2-rc1 ] ~ [ 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/pagemap.h>
 46 #include <linux/swap.h>
 47 #include <linux/backing-dev.h>
 48 #include <linux/migrate.h>
 49 #include <linux/page-isolation.h>
 50 #include <linux/suspend.h>
 51 #include <linux/slab.h>
 52 #include <linux/swapops.h>
 53 #include <linux/hugetlb.h>
 54 #include <linux/memory_hotplug.h>
 55 #include "internal.h"
 56 
 57 int sysctl_memory_failure_early_kill __read_mostly = 0;
 58 
 59 int sysctl_memory_failure_recovery __read_mostly = 1;
 60 
 61 atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
 62 
 63 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
 64 
 65 u32 hwpoison_filter_enable = 0;
 66 u32 hwpoison_filter_dev_major = ~0U;
 67 u32 hwpoison_filter_dev_minor = ~0U;
 68 u64 hwpoison_filter_flags_mask;
 69 u64 hwpoison_filter_flags_value;
 70 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
 71 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
 72 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
 73 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
 74 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
 75 
 76 static int hwpoison_filter_dev(struct page *p)
 77 {
 78         struct address_space *mapping;
 79         dev_t dev;
 80 
 81         if (hwpoison_filter_dev_major == ~0U &&
 82             hwpoison_filter_dev_minor == ~0U)
 83                 return 0;
 84 
 85         /*
 86          * page_mapping() does not accept slab pages.
 87          */
 88         if (PageSlab(p))
 89                 return -EINVAL;
 90 
 91         mapping = page_mapping(p);
 92         if (mapping == NULL || mapping->host == NULL)
 93                 return -EINVAL;
 94 
 95         dev = mapping->host->i_sb->s_dev;
 96         if (hwpoison_filter_dev_major != ~0U &&
 97             hwpoison_filter_dev_major != MAJOR(dev))
 98                 return -EINVAL;
 99         if (hwpoison_filter_dev_minor != ~0U &&
100             hwpoison_filter_dev_minor != MINOR(dev))
101                 return -EINVAL;
102 
103         return 0;
104 }
105 
106 static int hwpoison_filter_flags(struct page *p)
107 {
108         if (!hwpoison_filter_flags_mask)
109                 return 0;
110 
111         if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
112                                     hwpoison_filter_flags_value)
113                 return 0;
114         else
115                 return -EINVAL;
116 }
117 
118 /*
119  * This allows stress tests to limit test scope to a collection of tasks
120  * by putting them under some memcg. This prevents killing unrelated/important
121  * processes such as /sbin/init. Note that the target task may share clean
122  * pages with init (eg. libc text), which is harmless. If the target task
123  * share _dirty_ pages with another task B, the test scheme must make sure B
124  * is also included in the memcg. At last, due to race conditions this filter
125  * can only guarantee that the page either belongs to the memcg tasks, or is
126  * a freed page.
127  */
128 #ifdef  CONFIG_CGROUP_MEM_RES_CTLR_SWAP
129 u64 hwpoison_filter_memcg;
130 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
131 static int hwpoison_filter_task(struct page *p)
132 {
133         struct mem_cgroup *mem;
134         struct cgroup_subsys_state *css;
135         unsigned long ino;
136 
137         if (!hwpoison_filter_memcg)
138                 return 0;
139 
140         mem = try_get_mem_cgroup_from_page(p);
141         if (!mem)
142                 return -EINVAL;
143 
144         css = mem_cgroup_css(mem);
145         /* root_mem_cgroup has NULL dentries */
146         if (!css->cgroup->dentry)
147                 return -EINVAL;
148 
149         ino = css->cgroup->dentry->d_inode->i_ino;
150         css_put(css);
151 
152         if (ino != hwpoison_filter_memcg)
153                 return -EINVAL;
154 
155         return 0;
156 }
157 #else
158 static int hwpoison_filter_task(struct page *p) { return 0; }
159 #endif
160 
161 int hwpoison_filter(struct page *p)
162 {
163         if (!hwpoison_filter_enable)
164                 return 0;
165 
166         if (hwpoison_filter_dev(p))
167                 return -EINVAL;
168 
169         if (hwpoison_filter_flags(p))
170                 return -EINVAL;
171 
172         if (hwpoison_filter_task(p))
173                 return -EINVAL;
174 
175         return 0;
176 }
177 #else
178 int hwpoison_filter(struct page *p)
179 {
180         return 0;
181 }
182 #endif
183 
184 EXPORT_SYMBOL_GPL(hwpoison_filter);
185 
186 /*
187  * Send all the processes who have the page mapped an ``action optional''
188  * signal.
189  */
190 static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
191                         unsigned long pfn, struct page *page)
192 {
193         struct siginfo si;
194         int ret;
195 
196         printk(KERN_ERR
197                 "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
198                 pfn, t->comm, t->pid);
199         si.si_signo = SIGBUS;
200         si.si_errno = 0;
201         si.si_code = BUS_MCEERR_AO;
202         si.si_addr = (void *)addr;
203 #ifdef __ARCH_SI_TRAPNO
204         si.si_trapno = trapno;
205 #endif
206         si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
207         /*
208          * Don't use force here, it's convenient if the signal
209          * can be temporarily blocked.
210          * This could cause a loop when the user sets SIGBUS
211          * to SIG_IGN, but hopefully noone will do that?
212          */
213         ret = send_sig_info(SIGBUS, &si, t);  /* synchronous? */
214         if (ret < 0)
215                 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
216                        t->comm, t->pid, ret);
217         return ret;
218 }
219 
220 /*
221  * When a unknown page type is encountered drain as many buffers as possible
222  * in the hope to turn the page into a LRU or free page, which we can handle.
223  */
224 void shake_page(struct page *p, int access)
225 {
226         if (!PageSlab(p)) {
227                 lru_add_drain_all();
228                 if (PageLRU(p))
229                         return;
230                 drain_all_pages();
231                 if (PageLRU(p) || is_free_buddy_page(p))
232                         return;
233         }
234 
235         /*
236          * Only all shrink_slab here (which would also
237          * shrink other caches) if access is not potentially fatal.
238          */
239         if (access) {
240                 int nr;
241                 do {
242                         nr = shrink_slab(1000, GFP_KERNEL, 1000);
243                         if (page_count(p) == 1)
244                                 break;
245                 } while (nr > 10);
246         }
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_ao(struct list_head *to_kill, int doit, int trapno,
335                           int fail, struct page *page, unsigned long pfn)
336 {
337         struct to_kill *tk, *next;
338 
339         list_for_each_entry_safe (tk, next, to_kill, nd) {
340                 if (doit) {
341                         /*
342                          * In case something went wrong with munmapping
343                          * make sure the process doesn't catch the
344                          * signal and then access the memory. Just kill it.
345                          */
346                         if (fail || tk->addr_valid == 0) {
347                                 printk(KERN_ERR
348                 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
349                                         pfn, tk->tsk->comm, tk->tsk->pid);
350                                 force_sig(SIGKILL, tk->tsk);
351                         }
352 
353                         /*
354                          * In theory the process could have mapped
355                          * something else on the address in-between. We could
356                          * check for that, but we need to tell the
357                          * process anyways.
358                          */
359                         else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
360                                               pfn, page) < 0)
361                                 printk(KERN_ERR
362                 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
363                                         pfn, tk->tsk->comm, tk->tsk->pid);
364                 }
365                 put_task_struct(tk->tsk);
366                 kfree(tk);
367         }
368 }
369 
370 static int task_early_kill(struct task_struct *tsk)
371 {
372         if (!tsk->mm)
373                 return 0;
374         if (tsk->flags & PF_MCE_PROCESS)
375                 return !!(tsk->flags & PF_MCE_EARLY);
376         return sysctl_memory_failure_early_kill;
377 }
378 
379 /*
380  * Collect processes when the error hit an anonymous page.
381  */
382 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
383                               struct to_kill **tkc)
384 {
385         struct vm_area_struct *vma;
386         struct task_struct *tsk;
387         struct anon_vma *av;
388 
389         read_lock(&tasklist_lock);
390         av = page_lock_anon_vma(page);
391         if (av == NULL) /* Not actually mapped anymore */
392                 goto out;
393         for_each_process (tsk) {
394                 struct anon_vma_chain *vmac;
395 
396                 if (!task_early_kill(tsk))
397                         continue;
398                 list_for_each_entry(vmac, &av->head, same_anon_vma) {
399                         vma = vmac->vma;
400                         if (!page_mapped_in_vma(page, vma))
401                                 continue;
402                         if (vma->vm_mm == tsk->mm)
403                                 add_to_kill(tsk, page, vma, to_kill, tkc);
404                 }
405         }
406         page_unlock_anon_vma(av);
407 out:
408         read_unlock(&tasklist_lock);
409 }
410 
411 /*
412  * Collect processes when the error hit a file mapped page.
413  */
414 static void collect_procs_file(struct page *page, struct list_head *to_kill,
415                               struct to_kill **tkc)
416 {
417         struct vm_area_struct *vma;
418         struct task_struct *tsk;
419         struct prio_tree_iter iter;
420         struct address_space *mapping = page->mapping;
421 
422         /*
423          * A note on the locking order between the two locks.
424          * We don't rely on this particular order.
425          * If you have some other code that needs a different order
426          * feel free to switch them around. Or add a reverse link
427          * from mm_struct to task_struct, then this could be all
428          * done without taking tasklist_lock and looping over all tasks.
429          */
430 
431         read_lock(&tasklist_lock);
432         spin_lock(&mapping->i_mmap_lock);
433         for_each_process(tsk) {
434                 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
435 
436                 if (!task_early_kill(tsk))
437                         continue;
438 
439                 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
440                                       pgoff) {
441                         /*
442                          * Send early kill signal to tasks where a vma covers
443                          * the page but the corrupted page is not necessarily
444                          * mapped it in its pte.
445                          * Assume applications who requested early kill want
446                          * to be informed of all such data corruptions.
447                          */
448                         if (vma->vm_mm == tsk->mm)
449                                 add_to_kill(tsk, page, vma, to_kill, tkc);
450                 }
451         }
452         spin_unlock(&mapping->i_mmap_lock);
453         read_unlock(&tasklist_lock);
454 }
455 
456 /*
457  * Collect the processes who have the corrupted page mapped to kill.
458  * This is done in two steps for locking reasons.
459  * First preallocate one tokill structure outside the spin locks,
460  * so that we can kill at least one process reasonably reliable.
461  */
462 static void collect_procs(struct page *page, struct list_head *tokill)
463 {
464         struct to_kill *tk;
465 
466         if (!page->mapping)
467                 return;
468 
469         tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
470         if (!tk)
471                 return;
472         if (PageAnon(page))
473                 collect_procs_anon(page, tokill, &tk);
474         else
475                 collect_procs_file(page, tokill, &tk);
476         kfree(tk);
477 }
478 
479 /*
480  * Error handlers for various types of pages.
481  */
482 
483 enum outcome {
484         IGNORED,        /* Error: cannot be handled */
485         FAILED,         /* Error: handling failed */
486         DELAYED,        /* Will be handled later */
487         RECOVERED,      /* Successfully recovered */
488 };
489 
490 static const char *action_name[] = {
491         [IGNORED] = "Ignored",
492         [FAILED] = "Failed",
493         [DELAYED] = "Delayed",
494         [RECOVERED] = "Recovered",
495 };
496 
497 /*
498  * XXX: It is possible that a page is isolated from LRU cache,
499  * and then kept in swap cache or failed to remove from page cache.
500  * The page count will stop it from being freed by unpoison.
501  * Stress tests should be aware of this memory leak problem.
502  */
503 static int delete_from_lru_cache(struct page *p)
504 {
505         if (!isolate_lru_page(p)) {
506                 /*
507                  * Clear sensible page flags, so that the buddy system won't
508                  * complain when the page is unpoison-and-freed.
509                  */
510                 ClearPageActive(p);
511                 ClearPageUnevictable(p);
512                 /*
513                  * drop the page count elevated by isolate_lru_page()
514                  */
515                 page_cache_release(p);
516                 return 0;
517         }
518         return -EIO;
519 }
520 
521 /*
522  * Error hit kernel page.
523  * Do nothing, try to be lucky and not touch this instead. For a few cases we
524  * could be more sophisticated.
525  */
526 static int me_kernel(struct page *p, unsigned long pfn)
527 {
528         return IGNORED;
529 }
530 
531 /*
532  * Page in unknown state. Do nothing.
533  */
534 static int me_unknown(struct page *p, unsigned long pfn)
535 {
536         printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
537         return FAILED;
538 }
539 
540 /*
541  * Clean (or cleaned) page cache page.
542  */
543 static int me_pagecache_clean(struct page *p, unsigned long pfn)
544 {
545         int err;
546         int ret = FAILED;
547         struct address_space *mapping;
548 
549         delete_from_lru_cache(p);
550 
551         /*
552          * For anonymous pages we're done the only reference left
553          * should be the one m_f() holds.
554          */
555         if (PageAnon(p))
556                 return RECOVERED;
557 
558         /*
559          * Now truncate the page in the page cache. This is really
560          * more like a "temporary hole punch"
561          * Don't do this for block devices when someone else
562          * has a reference, because it could be file system metadata
563          * and that's not safe to truncate.
564          */
565         mapping = page_mapping(p);
566         if (!mapping) {
567                 /*
568                  * Page has been teared down in the meanwhile
569                  */
570                 return FAILED;
571         }
572 
573         /*
574          * Truncation is a bit tricky. Enable it per file system for now.
575          *
576          * Open: to take i_mutex or not for this? Right now we don't.
577          */
578         if (mapping->a_ops->error_remove_page) {
579                 err = mapping->a_ops->error_remove_page(mapping, p);
580                 if (err != 0) {
581                         printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
582                                         pfn, err);
583                 } else if (page_has_private(p) &&
584                                 !try_to_release_page(p, GFP_NOIO)) {
585                         pr_info("MCE %#lx: failed to release buffers\n", pfn);
586                 } else {
587                         ret = RECOVERED;
588                 }
589         } else {
590                 /*
591                  * If the file system doesn't support it just invalidate
592                  * This fails on dirty or anything with private pages
593                  */
594                 if (invalidate_inode_page(p))
595                         ret = RECOVERED;
596                 else
597                         printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
598                                 pfn);
599         }
600         return ret;
601 }
602 
603 /*
604  * Dirty cache page page
605  * Issues: when the error hit a hole page the error is not properly
606  * propagated.
607  */
608 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
609 {
610         struct address_space *mapping = page_mapping(p);
611 
612         SetPageError(p);
613         /* TBD: print more information about the file. */
614         if (mapping) {
615                 /*
616                  * IO error will be reported by write(), fsync(), etc.
617                  * who check the mapping.
618                  * This way the application knows that something went
619                  * wrong with its dirty file data.
620                  *
621                  * There's one open issue:
622                  *
623                  * The EIO will be only reported on the next IO
624                  * operation and then cleared through the IO map.
625                  * Normally Linux has two mechanisms to pass IO error
626                  * first through the AS_EIO flag in the address space
627                  * and then through the PageError flag in the page.
628                  * Since we drop pages on memory failure handling the
629                  * only mechanism open to use is through AS_AIO.
630                  *
631                  * This has the disadvantage that it gets cleared on
632                  * the first operation that returns an error, while
633                  * the PageError bit is more sticky and only cleared
634                  * when the page is reread or dropped.  If an
635                  * application assumes it will always get error on
636                  * fsync, but does other operations on the fd before
637                  * and the page is dropped inbetween then the error
638                  * will not be properly reported.
639                  *
640                  * This can already happen even without hwpoisoned
641                  * pages: first on metadata IO errors (which only
642                  * report through AS_EIO) or when the page is dropped
643                  * at the wrong time.
644                  *
645                  * So right now we assume that the application DTRT on
646                  * the first EIO, but we're not worse than other parts
647                  * of the kernel.
648                  */
649                 mapping_set_error(mapping, EIO);
650         }
651 
652         return me_pagecache_clean(p, pfn);
653 }
654 
655 /*
656  * Clean and dirty swap cache.
657  *
658  * Dirty swap cache page is tricky to handle. The page could live both in page
659  * cache and swap cache(ie. page is freshly swapped in). So it could be
660  * referenced concurrently by 2 types of PTEs:
661  * normal PTEs and swap PTEs. We try to handle them consistently by calling
662  * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
663  * and then
664  *      - clear dirty bit to prevent IO
665  *      - remove from LRU
666  *      - but keep in the swap cache, so that when we return to it on
667  *        a later page fault, we know the application is accessing
668  *        corrupted data and shall be killed (we installed simple
669  *        interception code in do_swap_page to catch it).
670  *
671  * Clean swap cache pages can be directly isolated. A later page fault will
672  * bring in the known good data from disk.
673  */
674 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
675 {
676         ClearPageDirty(p);
677         /* Trigger EIO in shmem: */
678         ClearPageUptodate(p);
679 
680         if (!delete_from_lru_cache(p))
681                 return DELAYED;
682         else
683                 return FAILED;
684 }
685 
686 static int me_swapcache_clean(struct page *p, unsigned long pfn)
687 {
688         delete_from_swap_cache(p);
689 
690         if (!delete_from_lru_cache(p))
691                 return RECOVERED;
692         else
693                 return FAILED;
694 }
695 
696 /*
697  * Huge pages. Needs work.
698  * Issues:
699  * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
700  *   To narrow down kill region to one page, we need to break up pmd.
701  */
702 static int me_huge_page(struct page *p, unsigned long pfn)
703 {
704         int res = 0;
705         struct page *hpage = compound_head(p);
706         /*
707          * We can safely recover from error on free or reserved (i.e.
708          * not in-use) hugepage by dequeuing it from freelist.
709          * To check whether a hugepage is in-use or not, we can't use
710          * page->lru because it can be used in other hugepage operations,
711          * such as __unmap_hugepage_range() and gather_surplus_pages().
712          * So instead we use page_mapping() and PageAnon().
713          * We assume that this function is called with page lock held,
714          * so there is no race between isolation and mapping/unmapping.
715          */
716         if (!(page_mapping(hpage) || PageAnon(hpage))) {
717                 res = dequeue_hwpoisoned_huge_page(hpage);
718                 if (!res)
719                         return RECOVERED;
720         }
721         return DELAYED;
722 }
723 
724 /*
725  * Various page states we can handle.
726  *
727  * A page state is defined by its current page->flags bits.
728  * The table matches them in order and calls the right handler.
729  *
730  * This is quite tricky because we can access page at any time
731  * in its live cycle, so all accesses have to be extremly careful.
732  *
733  * This is not complete. More states could be added.
734  * For any missing state don't attempt recovery.
735  */
736 
737 #define dirty           (1UL << PG_dirty)
738 #define sc              (1UL << PG_swapcache)
739 #define unevict         (1UL << PG_unevictable)
740 #define mlock           (1UL << PG_mlocked)
741 #define writeback       (1UL << PG_writeback)
742 #define lru             (1UL << PG_lru)
743 #define swapbacked      (1UL << PG_swapbacked)
744 #define head            (1UL << PG_head)
745 #define tail            (1UL << PG_tail)
746 #define compound        (1UL << PG_compound)
747 #define slab            (1UL << PG_slab)
748 #define reserved        (1UL << PG_reserved)
749 
750 static struct page_state {
751         unsigned long mask;
752         unsigned long res;
753         char *msg;
754         int (*action)(struct page *p, unsigned long pfn);
755 } error_states[] = {
756         { reserved,     reserved,       "reserved kernel",      me_kernel },
757         /*
758          * free pages are specially detected outside this table:
759          * PG_buddy pages only make a small fraction of all free pages.
760          */
761 
762         /*
763          * Could in theory check if slab page is free or if we can drop
764          * currently unused objects without touching them. But just
765          * treat it as standard kernel for now.
766          */
767         { slab,         slab,           "kernel slab",  me_kernel },
768 
769 #ifdef CONFIG_PAGEFLAGS_EXTENDED
770         { head,         head,           "huge",         me_huge_page },
771         { tail,         tail,           "huge",         me_huge_page },
772 #else
773         { compound,     compound,       "huge",         me_huge_page },
774 #endif
775 
776         { sc|dirty,     sc|dirty,       "swapcache",    me_swapcache_dirty },
777         { sc|dirty,     sc,             "swapcache",    me_swapcache_clean },
778 
779         { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
780         { unevict,      unevict,        "unevictable LRU", me_pagecache_clean},
781 
782         { mlock|dirty,  mlock|dirty,    "mlocked LRU",  me_pagecache_dirty },
783         { mlock,        mlock,          "mlocked LRU",  me_pagecache_clean },
784 
785         { lru|dirty,    lru|dirty,      "LRU",          me_pagecache_dirty },
786         { lru|dirty,    lru,            "clean LRU",    me_pagecache_clean },
787 
788         /*
789          * Catchall entry: must be at end.
790          */
791         { 0,            0,              "unknown page state",   me_unknown },
792 };
793 
794 #undef dirty
795 #undef sc
796 #undef unevict
797 #undef mlock
798 #undef writeback
799 #undef lru
800 #undef swapbacked
801 #undef head
802 #undef tail
803 #undef compound
804 #undef slab
805 #undef reserved
806 
807 static void action_result(unsigned long pfn, char *msg, int result)
808 {
809         struct page *page = pfn_to_page(pfn);
810 
811         printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
812                 pfn,
813                 PageDirty(page) ? "dirty " : "",
814                 msg, action_name[result]);
815 }
816 
817 static int page_action(struct page_state *ps, struct page *p,
818                         unsigned long pfn)
819 {
820         int result;
821         int count;
822 
823         result = ps->action(p, pfn);
824         action_result(pfn, ps->msg, result);
825 
826         count = page_count(p) - 1;
827         if (ps->action == me_swapcache_dirty && result == DELAYED)
828                 count--;
829         if (count != 0) {
830                 printk(KERN_ERR
831                        "MCE %#lx: %s page still referenced by %d users\n",
832                        pfn, ps->msg, count);
833                 result = FAILED;
834         }
835 
836         /* Could do more checks here if page looks ok */
837         /*
838          * Could adjust zone counters here to correct for the missing page.
839          */
840 
841         return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
842 }
843 
844 /*
845  * Do all that is necessary to remove user space mappings. Unmap
846  * the pages and send SIGBUS to the processes if the data was dirty.
847  */
848 static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
849                                   int trapno)
850 {
851         enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
852         struct address_space *mapping;
853         LIST_HEAD(tokill);
854         int ret;
855         int kill = 1;
856         struct page *hpage = compound_head(p);
857 
858         if (PageReserved(p) || PageSlab(p))
859                 return SWAP_SUCCESS;
860 
861         /*
862          * This check implies we don't kill processes if their pages
863          * are in the swap cache early. Those are always late kills.
864          */
865         if (!page_mapped(hpage))
866                 return SWAP_SUCCESS;
867 
868         if (PageKsm(p))
869                 return SWAP_FAIL;
870 
871         if (PageSwapCache(p)) {
872                 printk(KERN_ERR
873                        "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
874                 ttu |= TTU_IGNORE_HWPOISON;
875         }
876 
877         /*
878          * Propagate the dirty bit from PTEs to struct page first, because we
879          * need this to decide if we should kill or just drop the page.
880          * XXX: the dirty test could be racy: set_page_dirty() may not always
881          * be called inside page lock (it's recommended but not enforced).
882          */
883         mapping = page_mapping(hpage);
884         if (!PageDirty(hpage) && mapping &&
885             mapping_cap_writeback_dirty(mapping)) {
886                 if (page_mkclean(hpage)) {
887                         SetPageDirty(hpage);
888                 } else {
889                         kill = 0;
890                         ttu |= TTU_IGNORE_HWPOISON;
891                         printk(KERN_INFO
892         "MCE %#lx: corrupted page was clean: dropped without side effects\n",
893                                 pfn);
894                 }
895         }
896 
897         /*
898          * First collect all the processes that have the page
899          * mapped in dirty form.  This has to be done before try_to_unmap,
900          * because ttu takes the rmap data structures down.
901          *
902          * Error handling: We ignore errors here because
903          * there's nothing that can be done.
904          */
905         if (kill)
906                 collect_procs(hpage, &tokill);
907 
908         ret = try_to_unmap(hpage, ttu);
909         if (ret != SWAP_SUCCESS)
910                 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
911                                 pfn, page_mapcount(hpage));
912 
913         /*
914          * Now that the dirty bit has been propagated to the
915          * struct page and all unmaps done we can decide if
916          * killing is needed or not.  Only kill when the page
917          * was dirty, otherwise the tokill list is merely
918          * freed.  When there was a problem unmapping earlier
919          * use a more force-full uncatchable kill to prevent
920          * any accesses to the poisoned memory.
921          */
922         kill_procs_ao(&tokill, !!PageDirty(hpage), trapno,
923                       ret != SWAP_SUCCESS, p, pfn);
924 
925         return ret;
926 }
927 
928 static void set_page_hwpoison_huge_page(struct page *hpage)
929 {
930         int i;
931         int nr_pages = 1 << compound_order(hpage);
932         for (i = 0; i < nr_pages; i++)
933                 SetPageHWPoison(hpage + i);
934 }
935 
936 static void clear_page_hwpoison_huge_page(struct page *hpage)
937 {
938         int i;
939         int nr_pages = 1 << compound_order(hpage);
940         for (i = 0; i < nr_pages; i++)
941                 ClearPageHWPoison(hpage + i);
942 }
943 
944 int __memory_failure(unsigned long pfn, int trapno, int flags)
945 {
946         struct page_state *ps;
947         struct page *p;
948         struct page *hpage;
949         int res;
950         unsigned int nr_pages;
951 
952         if (!sysctl_memory_failure_recovery)
953                 panic("Memory failure from trap %d on page %lx", trapno, pfn);
954 
955         if (!pfn_valid(pfn)) {
956                 printk(KERN_ERR
957                        "MCE %#lx: memory outside kernel control\n",
958                        pfn);
959                 return -ENXIO;
960         }
961 
962         p = pfn_to_page(pfn);
963         hpage = compound_head(p);
964         if (TestSetPageHWPoison(p)) {
965                 printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
966                 return 0;
967         }
968 
969         nr_pages = 1 << compound_order(hpage);
970         atomic_long_add(nr_pages, &mce_bad_pages);
971 
972         /*
973          * We need/can do nothing about count=0 pages.
974          * 1) it's a free page, and therefore in safe hand:
975          *    prep_new_page() will be the gate keeper.
976          * 2) it's a free hugepage, which is also safe:
977          *    an affected hugepage will be dequeued from hugepage freelist,
978          *    so there's no concern about reusing it ever after.
979          * 3) it's part of a non-compound high order page.
980          *    Implies some kernel user: cannot stop them from
981          *    R/W the page; let's pray that the page has been
982          *    used and will be freed some time later.
983          * In fact it's dangerous to directly bump up page count from 0,
984          * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
985          */
986         if (!(flags & MF_COUNT_INCREASED) &&
987                 !get_page_unless_zero(hpage)) {
988                 if (is_free_buddy_page(p)) {
989                         action_result(pfn, "free buddy", DELAYED);
990                         return 0;
991                 } else if (PageHuge(hpage)) {
992                         /*
993                          * Check "just unpoisoned", "filter hit", and
994                          * "race with other subpage."
995                          */
996                         lock_page_nosync(hpage);
997                         if (!PageHWPoison(hpage)
998                             || (hwpoison_filter(p) && TestClearPageHWPoison(p))
999                             || (p != hpage && TestSetPageHWPoison(hpage))) {
1000                                 atomic_long_sub(nr_pages, &mce_bad_pages);
1001                                 return 0;
1002                         }
1003                         set_page_hwpoison_huge_page(hpage);
1004                         res = dequeue_hwpoisoned_huge_page(hpage);
1005                         action_result(pfn, "free huge",
1006                                       res ? IGNORED : DELAYED);
1007                         unlock_page(hpage);
1008                         return res;
1009                 } else {
1010                         action_result(pfn, "high order kernel", IGNORED);
1011                         return -EBUSY;
1012                 }
1013         }
1014 
1015         /*
1016          * We ignore non-LRU pages for good reasons.
1017          * - PG_locked is only well defined for LRU pages and a few others
1018          * - to avoid races with __set_page_locked()
1019          * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1020          * The check (unnecessarily) ignores LRU pages being isolated and
1021          * walked by the page reclaim code, however that's not a big loss.
1022          */
1023         if (!PageLRU(p) && !PageHuge(p))
1024                 shake_page(p, 0);
1025         if (!PageLRU(p) && !PageHuge(p)) {
1026                 /*
1027                  * shake_page could have turned it free.
1028                  */
1029                 if (is_free_buddy_page(p)) {
1030                         action_result(pfn, "free buddy, 2nd try", DELAYED);
1031                         return 0;
1032                 }
1033                 action_result(pfn, "non LRU", IGNORED);
1034                 put_page(p);
1035                 return -EBUSY;
1036         }
1037 
1038         /*
1039          * Lock the page and wait for writeback to finish.
1040          * It's very difficult to mess with pages currently under IO
1041          * and in many cases impossible, so we just avoid it here.
1042          */
1043         lock_page_nosync(hpage);
1044 
1045         /*
1046          * unpoison always clear PG_hwpoison inside page lock
1047          */
1048         if (!PageHWPoison(p)) {
1049                 printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1050                 res = 0;
1051                 goto out;
1052         }
1053         if (hwpoison_filter(p)) {
1054                 if (TestClearPageHWPoison(p))
1055                         atomic_long_sub(nr_pages, &mce_bad_pages);
1056                 unlock_page(hpage);
1057                 put_page(hpage);
1058                 return 0;
1059         }
1060 
1061         /*
1062          * For error on the tail page, we should set PG_hwpoison
1063          * on the head page to show that the hugepage is hwpoisoned
1064          */
1065         if (PageTail(p) && TestSetPageHWPoison(hpage)) {
1066                 action_result(pfn, "hugepage already hardware poisoned",
1067                                 IGNORED);
1068                 unlock_page(hpage);
1069                 put_page(hpage);
1070                 return 0;
1071         }
1072         /*
1073          * Set PG_hwpoison on all pages in an error hugepage,
1074          * because containment is done in hugepage unit for now.
1075          * Since we have done TestSetPageHWPoison() for the head page with
1076          * page lock held, we can safely set PG_hwpoison bits on tail pages.
1077          */
1078         if (PageHuge(p))
1079                 set_page_hwpoison_huge_page(hpage);
1080 
1081         wait_on_page_writeback(p);
1082 
1083         /*
1084          * Now take care of user space mappings.
1085          * Abort on fail: __remove_from_page_cache() assumes unmapped page.
1086          */
1087         if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
1088                 printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1089                 res = -EBUSY;
1090                 goto out;
1091         }
1092 
1093         /*
1094          * Torn down by someone else?
1095          */
1096         if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1097                 action_result(pfn, "already truncated LRU", IGNORED);
1098                 res = -EBUSY;
1099                 goto out;
1100         }
1101 
1102         res = -EBUSY;
1103         for (ps = error_states;; ps++) {
1104                 if ((p->flags & ps->mask) == ps->res) {
1105                         res = page_action(ps, p, pfn);
1106                         break;
1107                 }
1108         }
1109 out:
1110         unlock_page(hpage);
1111         return res;
1112 }
1113 EXPORT_SYMBOL_GPL(__memory_failure);
1114 
1115 /**
1116  * memory_failure - Handle memory failure of a page.
1117  * @pfn: Page Number of the corrupted page
1118  * @trapno: Trap number reported in the signal to user space.
1119  *
1120  * This function is called by the low level machine check code
1121  * of an architecture when it detects hardware memory corruption
1122  * of a page. It tries its best to recover, which includes
1123  * dropping pages, killing processes etc.
1124  *
1125  * The function is primarily of use for corruptions that
1126  * happen outside the current execution context (e.g. when
1127  * detected by a background scrubber)
1128  *
1129  * Must run in process context (e.g. a work queue) with interrupts
1130  * enabled and no spinlocks hold.
1131  */
1132 void memory_failure(unsigned long pfn, int trapno)
1133 {
1134         __memory_failure(pfn, trapno, 0);
1135 }
1136 
1137 /**
1138  * unpoison_memory - Unpoison a previously poisoned page
1139  * @pfn: Page number of the to be unpoisoned page
1140  *
1141  * Software-unpoison a page that has been poisoned by
1142  * memory_failure() earlier.
1143  *
1144  * This is only done on the software-level, so it only works
1145  * for linux injected failures, not real hardware failures
1146  *
1147  * Returns 0 for success, otherwise -errno.
1148  */
1149 int unpoison_memory(unsigned long pfn)
1150 {
1151         struct page *page;
1152         struct page *p;
1153         int freeit = 0;
1154         unsigned int nr_pages;
1155 
1156         if (!pfn_valid(pfn))
1157                 return -ENXIO;
1158 
1159         p = pfn_to_page(pfn);
1160         page = compound_head(p);
1161 
1162         if (!PageHWPoison(p)) {
1163                 pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
1164                 return 0;
1165         }
1166 
1167         nr_pages = 1 << compound_order(page);
1168 
1169         if (!get_page_unless_zero(page)) {
1170                 /*
1171                  * Since HWPoisoned hugepage should have non-zero refcount,
1172                  * race between memory failure and unpoison seems to happen.
1173                  * In such case unpoison fails and memory failure runs
1174                  * to the end.
1175                  */
1176                 if (PageHuge(page)) {
1177                         pr_debug("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1178                         return 0;
1179                 }
1180                 if (TestClearPageHWPoison(p))
1181                         atomic_long_sub(nr_pages, &mce_bad_pages);
1182                 pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
1183                 return 0;
1184         }
1185 
1186         lock_page_nosync(page);
1187         /*
1188          * This test is racy because PG_hwpoison is set outside of page lock.
1189          * That's acceptable because that won't trigger kernel panic. Instead,
1190          * the PG_hwpoison page will be caught and isolated on the entrance to
1191          * the free buddy page pool.
1192          */
1193         if (TestClearPageHWPoison(page)) {
1194                 pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
1195                 atomic_long_sub(nr_pages, &mce_bad_pages);
1196                 freeit = 1;
1197                 if (PageHuge(page))
1198                         clear_page_hwpoison_huge_page(page);
1199         }
1200         unlock_page(page);
1201 
1202         put_page(page);
1203         if (freeit)
1204                 put_page(page);
1205 
1206         return 0;
1207 }
1208 EXPORT_SYMBOL(unpoison_memory);
1209 
1210 static struct page *new_page(struct page *p, unsigned long private, int **x)
1211 {
1212         int nid = page_to_nid(p);
1213         if (PageHuge(p))
1214                 return alloc_huge_page_node(page_hstate(compound_head(p)),
1215                                                    nid);
1216         else
1217                 return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1218 }
1219 
1220 /*
1221  * Safely get reference count of an arbitrary page.
1222  * Returns 0 for a free page, -EIO for a zero refcount page
1223  * that is not free, and 1 for any other page type.
1224  * For 1 the page is returned with increased page count, otherwise not.
1225  */
1226 static int get_any_page(struct page *p, unsigned long pfn, int flags)
1227 {
1228         int ret;
1229 
1230         if (flags & MF_COUNT_INCREASED)
1231                 return 1;
1232 
1233         /*
1234          * The lock_memory_hotplug prevents a race with memory hotplug.
1235          * This is a big hammer, a better would be nicer.
1236          */
1237         lock_memory_hotplug();
1238 
1239         /*
1240          * Isolate the page, so that it doesn't get reallocated if it
1241          * was free.
1242          */
1243         set_migratetype_isolate(p);
1244         /*
1245          * When the target page is a free hugepage, just remove it
1246          * from free hugepage list.
1247          */
1248         if (!get_page_unless_zero(compound_head(p))) {
1249                 if (PageHuge(p)) {
1250                         pr_info("get_any_page: %#lx free huge page\n", pfn);
1251                         ret = dequeue_hwpoisoned_huge_page(compound_head(p));
1252                 } else if (is_free_buddy_page(p)) {
1253                         pr_info("get_any_page: %#lx free buddy page\n", pfn);
1254                         /* Set hwpoison bit while page is still isolated */
1255                         SetPageHWPoison(p);
1256                         ret = 0;
1257                 } else {
1258                         pr_info("get_any_page: %#lx: unknown zero refcount page type %lx\n",
1259                                 pfn, p->flags);
1260                         ret = -EIO;
1261                 }
1262         } else {
1263                 /* Not a free page */
1264                 ret = 1;
1265         }
1266         unset_migratetype_isolate(p);
1267         unlock_memory_hotplug();
1268         return ret;
1269 }
1270 
1271 static int soft_offline_huge_page(struct page *page, int flags)
1272 {
1273         int ret;
1274         unsigned long pfn = page_to_pfn(page);
1275         struct page *hpage = compound_head(page);
1276         LIST_HEAD(pagelist);
1277 
1278         ret = get_any_page(page, pfn, flags);
1279         if (ret < 0)
1280                 return ret;
1281         if (ret == 0)
1282                 goto done;
1283 
1284         if (PageHWPoison(hpage)) {
1285                 put_page(hpage);
1286                 pr_debug("soft offline: %#lx hugepage already poisoned\n", pfn);
1287                 return -EBUSY;
1288         }
1289 
1290         /* Keep page count to indicate a given hugepage is isolated. */
1291 
1292         list_add(&hpage->lru, &pagelist);
1293         ret = migrate_huge_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0);
1294         if (ret) {
1295                         putback_lru_pages(&pagelist);
1296                 pr_debug("soft offline: %#lx: migration failed %d, type %lx\n",
1297                          pfn, ret, page->flags);
1298                 if (ret > 0)
1299                         ret = -EIO;
1300                 return ret;
1301         }
1302 done:
1303         if (!PageHWPoison(hpage))
1304                 atomic_long_add(1 << compound_order(hpage), &mce_bad_pages);
1305         set_page_hwpoison_huge_page(hpage);
1306         dequeue_hwpoisoned_huge_page(hpage);
1307         /* keep elevated page count for bad page */
1308         return ret;
1309 }
1310 
1311 /**
1312  * soft_offline_page - Soft offline a page.
1313  * @page: page to offline
1314  * @flags: flags. Same as memory_failure().
1315  *
1316  * Returns 0 on success, otherwise negated errno.
1317  *
1318  * Soft offline a page, by migration or invalidation,
1319  * without killing anything. This is for the case when
1320  * a page is not corrupted yet (so it's still valid to access),
1321  * but has had a number of corrected errors and is better taken
1322  * out.
1323  *
1324  * The actual policy on when to do that is maintained by
1325  * user space.
1326  *
1327  * This should never impact any application or cause data loss,
1328  * however it might take some time.
1329  *
1330  * This is not a 100% solution for all memory, but tries to be
1331  * ``good enough'' for the majority of memory.
1332  */
1333 int soft_offline_page(struct page *page, int flags)
1334 {
1335         int ret;
1336         unsigned long pfn = page_to_pfn(page);
1337 
1338         if (PageHuge(page))
1339                 return soft_offline_huge_page(page, flags);
1340 
1341         ret = get_any_page(page, pfn, flags);
1342         if (ret < 0)
1343                 return ret;
1344         if (ret == 0)
1345                 goto done;
1346 
1347         /*
1348          * Page cache page we can handle?
1349          */
1350         if (!PageLRU(page)) {
1351                 /*
1352                  * Try to free it.
1353                  */
1354                 put_page(page);
1355                 shake_page(page, 1);
1356 
1357                 /*
1358                  * Did it turn free?
1359                  */
1360                 ret = get_any_page(page, pfn, 0);
1361                 if (ret < 0)
1362                         return ret;
1363                 if (ret == 0)
1364                         goto done;
1365         }
1366         if (!PageLRU(page)) {
1367                 pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1368                                 pfn, page->flags);
1369                 return -EIO;
1370         }
1371 
1372         lock_page(page);
1373         wait_on_page_writeback(page);
1374 
1375         /*
1376          * Synchronized using the page lock with memory_failure()
1377          */
1378         if (PageHWPoison(page)) {
1379                 unlock_page(page);
1380                 put_page(page);
1381                 pr_info("soft offline: %#lx page already poisoned\n", pfn);
1382                 return -EBUSY;
1383         }
1384 
1385         /*
1386          * Try to invalidate first. This should work for
1387          * non dirty unmapped page cache pages.
1388          */
1389         ret = invalidate_inode_page(page);
1390         unlock_page(page);
1391 
1392         /*
1393          * Drop count because page migration doesn't like raised
1394          * counts. The page could get re-allocated, but if it becomes
1395          * LRU the isolation will just fail.
1396          * RED-PEN would be better to keep it isolated here, but we
1397          * would need to fix isolation locking first.
1398          */
1399         put_page(page);
1400         if (ret == 1) {
1401                 ret = 0;
1402                 pr_info("soft_offline: %#lx: invalidated\n", pfn);
1403                 goto done;
1404         }
1405 
1406         /*
1407          * Simple invalidation didn't work.
1408          * Try to migrate to a new page instead. migrate.c
1409          * handles a large number of cases for us.
1410          */
1411         ret = isolate_lru_page(page);
1412         if (!ret) {
1413                 LIST_HEAD(pagelist);
1414 
1415                 list_add(&page->lru, &pagelist);
1416                 ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0);
1417                 if (ret) {
1418                         pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1419                                 pfn, ret, page->flags);
1420                         if (ret > 0)
1421                                 ret = -EIO;
1422                 }
1423         } else {
1424                 pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1425                                 pfn, ret, page_count(page), page->flags);
1426         }
1427         if (ret)
1428                 return ret;
1429 
1430 done:
1431         atomic_long_add(1, &mce_bad_pages);
1432         SetPageHWPoison(page);
1433         /* keep elevated page count for bad page */
1434         return ret;
1435 }
1436 
1437 /*
1438  * The caller must hold current->mm->mmap_sem in read mode.
1439  */
1440 int is_hwpoison_address(unsigned long addr)
1441 {
1442         pgd_t *pgdp;
1443         pud_t pud, *pudp;
1444         pmd_t pmd, *pmdp;
1445         pte_t pte, *ptep;
1446         swp_entry_t entry;
1447 
1448         pgdp = pgd_offset(current->mm, addr);
1449         if (!pgd_present(*pgdp))
1450                 return 0;
1451         pudp = pud_offset(pgdp, addr);
1452         pud = *pudp;
1453         if (!pud_present(pud) || pud_large(pud))
1454                 return 0;
1455         pmdp = pmd_offset(pudp, addr);
1456         pmd = *pmdp;
1457         if (!pmd_present(pmd) || pmd_large(pmd))
1458                 return 0;
1459         ptep = pte_offset_map(pmdp, addr);
1460         pte = *ptep;
1461         pte_unmap(ptep);
1462         if (!is_swap_pte(pte))
1463                 return 0;
1464         entry = pte_to_swp_entry(pte);
1465         return is_hwpoison_entry(entry);
1466 }
1467 EXPORT_SYMBOL_GPL(is_hwpoison_address);
1468 

~ [ 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