Root/mm/memory-failure.c

1/*
2 * Copyright (C) 2008, 2009 Intel Corporation
3 * Authors: Andi Kleen, Fengguang Wu
4 *
5 * This software may be redistributed and/or modified under the terms of
6 * the GNU General Public License ("GPL") version 2 only as published by the
7 * Free Software Foundation.
8 *
9 * High level machine check handler. Handles pages reported by the
10 * hardware as being corrupted usually due to a 2bit ECC memory or cache
11 * failure.
12 *
13 * Handles page cache pages in various states. The tricky part
14 * here is that we can access any page asynchronous to other VM
15 * users, because memory failures could happen anytime and anywhere,
16 * possibly violating some of their assumptions. This is why this code
17 * has to be extremely careful. Generally it tries to use normal locking
18 * rules, as in get the standard locks, even if that means the
19 * error handling takes potentially a long time.
20 *
21 * The operation to map back from RMAP chains to processes has to walk
22 * the complete process list and has non linear complexity with the number
23 * mappings. In short it can be quite slow. But since memory corruptions
24 * are rare we hope to get away with this.
25 */
26
27/*
28 * Notebook:
29 * - hugetlb needs more code
30 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
31 * - pass bad pages to kdump next kernel
32 */
33#define DEBUG 1 /* remove me in 2.6.34 */
34#include <linux/kernel.h>
35#include <linux/mm.h>
36#include <linux/page-flags.h>
37#include <linux/kernel-page-flags.h>
38#include <linux/sched.h>
39#include <linux/ksm.h>
40#include <linux/rmap.h>
41#include <linux/pagemap.h>
42#include <linux/swap.h>
43#include <linux/backing-dev.h>
44#include <linux/migrate.h>
45#include <linux/page-isolation.h>
46#include <linux/suspend.h>
47#include <linux/slab.h>
48#include "internal.h"
49
50int sysctl_memory_failure_early_kill __read_mostly = 0;
51
52int sysctl_memory_failure_recovery __read_mostly = 1;
53
54atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
55
56#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
57
58u32 hwpoison_filter_enable = 0;
59u32 hwpoison_filter_dev_major = ~0U;
60u32 hwpoison_filter_dev_minor = ~0U;
61u64 hwpoison_filter_flags_mask;
62u64 hwpoison_filter_flags_value;
63EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
64EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
65EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
66EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
67EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
68
69static int hwpoison_filter_dev(struct page *p)
70{
71    struct address_space *mapping;
72    dev_t dev;
73
74    if (hwpoison_filter_dev_major == ~0U &&
75        hwpoison_filter_dev_minor == ~0U)
76        return 0;
77
78    /*
79     * page_mapping() does not accept slab page
80     */
81    if (PageSlab(p))
82        return -EINVAL;
83
84    mapping = page_mapping(p);
85    if (mapping == NULL || mapping->host == NULL)
86        return -EINVAL;
87
88    dev = mapping->host->i_sb->s_dev;
89    if (hwpoison_filter_dev_major != ~0U &&
90        hwpoison_filter_dev_major != MAJOR(dev))
91        return -EINVAL;
92    if (hwpoison_filter_dev_minor != ~0U &&
93        hwpoison_filter_dev_minor != MINOR(dev))
94        return -EINVAL;
95
96    return 0;
97}
98
99static int hwpoison_filter_flags(struct page *p)
100{
101    if (!hwpoison_filter_flags_mask)
102        return 0;
103
104    if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
105                    hwpoison_filter_flags_value)
106        return 0;
107    else
108        return -EINVAL;
109}
110
111/*
112 * This allows stress tests to limit test scope to a collection of tasks
113 * by putting them under some memcg. This prevents killing unrelated/important
114 * processes such as /sbin/init. Note that the target task may share clean
115 * pages with init (eg. libc text), which is harmless. If the target task
116 * share _dirty_ pages with another task B, the test scheme must make sure B
117 * is also included in the memcg. At last, due to race conditions this filter
118 * can only guarantee that the page either belongs to the memcg tasks, or is
119 * a freed page.
120 */
121#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
122u64 hwpoison_filter_memcg;
123EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
124static int hwpoison_filter_task(struct page *p)
125{
126    struct mem_cgroup *mem;
127    struct cgroup_subsys_state *css;
128    unsigned long ino;
129
130    if (!hwpoison_filter_memcg)
131        return 0;
132
133    mem = try_get_mem_cgroup_from_page(p);
134    if (!mem)
135        return -EINVAL;
136
137    css = mem_cgroup_css(mem);
138    /* root_mem_cgroup has NULL dentries */
139    if (!css->cgroup->dentry)
140        return -EINVAL;
141
142    ino = css->cgroup->dentry->d_inode->i_ino;
143    css_put(css);
144
145    if (ino != hwpoison_filter_memcg)
146        return -EINVAL;
147
148    return 0;
149}
150#else
151static int hwpoison_filter_task(struct page *p) { return 0; }
152#endif
153
154int hwpoison_filter(struct page *p)
155{
156    if (!hwpoison_filter_enable)
157        return 0;
158
159    if (hwpoison_filter_dev(p))
160        return -EINVAL;
161
162    if (hwpoison_filter_flags(p))
163        return -EINVAL;
164
165    if (hwpoison_filter_task(p))
166        return -EINVAL;
167
168    return 0;
169}
170#else
171int hwpoison_filter(struct page *p)
172{
173    return 0;
174}
175#endif
176
177EXPORT_SYMBOL_GPL(hwpoison_filter);
178
179/*
180 * Send all the processes who have the page mapped an ``action optional''
181 * signal.
182 */
183static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
184            unsigned long pfn)
185{
186    struct siginfo si;
187    int ret;
188
189    printk(KERN_ERR
190        "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
191        pfn, t->comm, t->pid);
192    si.si_signo = SIGBUS;
193    si.si_errno = 0;
194    si.si_code = BUS_MCEERR_AO;
195    si.si_addr = (void *)addr;
196#ifdef __ARCH_SI_TRAPNO
197    si.si_trapno = trapno;
198#endif
199    si.si_addr_lsb = PAGE_SHIFT;
200    /*
201     * Don't use force here, it's convenient if the signal
202     * can be temporarily blocked.
203     * This could cause a loop when the user sets SIGBUS
204     * to SIG_IGN, but hopefully noone will do that?
205     */
206    ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
207    if (ret < 0)
208        printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
209               t->comm, t->pid, ret);
210    return ret;
211}
212
213/*
214 * When a unknown page type is encountered drain as many buffers as possible
215 * in the hope to turn the page into a LRU or free page, which we can handle.
216 */
217void shake_page(struct page *p, int access)
218{
219    if (!PageSlab(p)) {
220        lru_add_drain_all();
221        if (PageLRU(p))
222            return;
223        drain_all_pages();
224        if (PageLRU(p) || is_free_buddy_page(p))
225            return;
226    }
227
228    /*
229     * Only all shrink_slab here (which would also
230     * shrink other caches) if access is not potentially fatal.
231     */
232    if (access) {
233        int nr;
234        do {
235            nr = shrink_slab(1000, GFP_KERNEL, 1000);
236            if (page_count(p) == 0)
237                break;
238        } while (nr > 10);
239    }
240}
241EXPORT_SYMBOL_GPL(shake_page);
242
243/*
244 * Kill all processes that have a poisoned page mapped and then isolate
245 * the page.
246 *
247 * General strategy:
248 * Find all processes having the page mapped and kill them.
249 * But we keep a page reference around so that the page is not
250 * actually freed yet.
251 * Then stash the page away
252 *
253 * There's no convenient way to get back to mapped processes
254 * from the VMAs. So do a brute-force search over all
255 * running processes.
256 *
257 * Remember that machine checks are not common (or rather
258 * if they are common you have other problems), so this shouldn't
259 * be a performance issue.
260 *
261 * Also there are some races possible while we get from the
262 * error detection to actually handle it.
263 */
264
265struct to_kill {
266    struct list_head nd;
267    struct task_struct *tsk;
268    unsigned long addr;
269    unsigned addr_valid:1;
270};
271
272/*
273 * Failure handling: if we can't find or can't kill a process there's
274 * not much we can do. We just print a message and ignore otherwise.
275 */
276
277/*
278 * Schedule a process for later kill.
279 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
280 * TBD would GFP_NOIO be enough?
281 */
282static void add_to_kill(struct task_struct *tsk, struct page *p,
283               struct vm_area_struct *vma,
284               struct list_head *to_kill,
285               struct to_kill **tkc)
286{
287    struct to_kill *tk;
288
289    if (*tkc) {
290        tk = *tkc;
291        *tkc = NULL;
292    } else {
293        tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
294        if (!tk) {
295            printk(KERN_ERR
296        "MCE: Out of memory while machine check handling\n");
297            return;
298        }
299    }
300    tk->addr = page_address_in_vma(p, vma);
301    tk->addr_valid = 1;
302
303    /*
304     * In theory we don't have to kill when the page was
305     * munmaped. But it could be also a mremap. Since that's
306     * likely very rare kill anyways just out of paranoia, but use
307     * a SIGKILL because the error is not contained anymore.
308     */
309    if (tk->addr == -EFAULT) {
310        pr_debug("MCE: Unable to find user space address %lx in %s\n",
311            page_to_pfn(p), tsk->comm);
312        tk->addr_valid = 0;
313    }
314    get_task_struct(tsk);
315    tk->tsk = tsk;
316    list_add_tail(&tk->nd, to_kill);
317}
318
319/*
320 * Kill the processes that have been collected earlier.
321 *
322 * Only do anything when DOIT is set, otherwise just free the list
323 * (this is used for clean pages which do not need killing)
324 * Also when FAIL is set do a force kill because something went
325 * wrong earlier.
326 */
327static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
328              int fail, unsigned long pfn)
329{
330    struct to_kill *tk, *next;
331
332    list_for_each_entry_safe (tk, next, to_kill, nd) {
333        if (doit) {
334            /*
335             * In case something went wrong with munmapping
336             * make sure the process doesn't catch the
337             * signal and then access the memory. Just kill it.
338             */
339            if (fail || tk->addr_valid == 0) {
340                printk(KERN_ERR
341        "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
342                    pfn, tk->tsk->comm, tk->tsk->pid);
343                force_sig(SIGKILL, tk->tsk);
344            }
345
346            /*
347             * In theory the process could have mapped
348             * something else on the address in-between. We could
349             * check for that, but we need to tell the
350             * process anyways.
351             */
352            else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
353                          pfn) < 0)
354                printk(KERN_ERR
355        "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
356                    pfn, tk->tsk->comm, tk->tsk->pid);
357        }
358        put_task_struct(tk->tsk);
359        kfree(tk);
360    }
361}
362
363static int task_early_kill(struct task_struct *tsk)
364{
365    if (!tsk->mm)
366        return 0;
367    if (tsk->flags & PF_MCE_PROCESS)
368        return !!(tsk->flags & PF_MCE_EARLY);
369    return sysctl_memory_failure_early_kill;
370}
371
372/*
373 * Collect processes when the error hit an anonymous page.
374 */
375static void collect_procs_anon(struct page *page, struct list_head *to_kill,
376                  struct to_kill **tkc)
377{
378    struct vm_area_struct *vma;
379    struct task_struct *tsk;
380    struct anon_vma *av;
381
382    read_lock(&tasklist_lock);
383    av = page_lock_anon_vma(page);
384    if (av == NULL) /* Not actually mapped anymore */
385        goto out;
386    for_each_process (tsk) {
387        struct anon_vma_chain *vmac;
388
389        if (!task_early_kill(tsk))
390            continue;
391        list_for_each_entry(vmac, &av->head, same_anon_vma) {
392            vma = vmac->vma;
393            if (!page_mapped_in_vma(page, vma))
394                continue;
395            if (vma->vm_mm == tsk->mm)
396                add_to_kill(tsk, page, vma, to_kill, tkc);
397        }
398    }
399    page_unlock_anon_vma(av);
400out:
401    read_unlock(&tasklist_lock);
402}
403
404/*
405 * Collect processes when the error hit a file mapped page.
406 */
407static void collect_procs_file(struct page *page, struct list_head *to_kill,
408                  struct to_kill **tkc)
409{
410    struct vm_area_struct *vma;
411    struct task_struct *tsk;
412    struct prio_tree_iter iter;
413    struct address_space *mapping = page->mapping;
414
415    /*
416     * A note on the locking order between the two locks.
417     * We don't rely on this particular order.
418     * If you have some other code that needs a different order
419     * feel free to switch them around. Or add a reverse link
420     * from mm_struct to task_struct, then this could be all
421     * done without taking tasklist_lock and looping over all tasks.
422     */
423
424    read_lock(&tasklist_lock);
425    spin_lock(&mapping->i_mmap_lock);
426    for_each_process(tsk) {
427        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
428
429        if (!task_early_kill(tsk))
430            continue;
431
432        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
433                      pgoff) {
434            /*
435             * Send early kill signal to tasks where a vma covers
436             * the page but the corrupted page is not necessarily
437             * mapped it in its pte.
438             * Assume applications who requested early kill want
439             * to be informed of all such data corruptions.
440             */
441            if (vma->vm_mm == tsk->mm)
442                add_to_kill(tsk, page, vma, to_kill, tkc);
443        }
444    }
445    spin_unlock(&mapping->i_mmap_lock);
446    read_unlock(&tasklist_lock);
447}
448
449/*
450 * Collect the processes who have the corrupted page mapped to kill.
451 * This is done in two steps for locking reasons.
452 * First preallocate one tokill structure outside the spin locks,
453 * so that we can kill at least one process reasonably reliable.
454 */
455static void collect_procs(struct page *page, struct list_head *tokill)
456{
457    struct to_kill *tk;
458
459    if (!page->mapping)
460        return;
461
462    tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
463    if (!tk)
464        return;
465    if (PageAnon(page))
466        collect_procs_anon(page, tokill, &tk);
467    else
468        collect_procs_file(page, tokill, &tk);
469    kfree(tk);
470}
471
472/*
473 * Error handlers for various types of pages.
474 */
475
476enum outcome {
477    IGNORED, /* Error: cannot be handled */
478    FAILED, /* Error: handling failed */
479    DELAYED, /* Will be handled later */
480    RECOVERED, /* Successfully recovered */
481};
482
483static const char *action_name[] = {
484    [IGNORED] = "Ignored",
485    [FAILED] = "Failed",
486    [DELAYED] = "Delayed",
487    [RECOVERED] = "Recovered",
488};
489
490/*
491 * XXX: It is possible that a page is isolated from LRU cache,
492 * and then kept in swap cache or failed to remove from page cache.
493 * The page count will stop it from being freed by unpoison.
494 * Stress tests should be aware of this memory leak problem.
495 */
496static int delete_from_lru_cache(struct page *p)
497{
498    if (!isolate_lru_page(p)) {
499        /*
500         * Clear sensible page flags, so that the buddy system won't
501         * complain when the page is unpoison-and-freed.
502         */
503        ClearPageActive(p);
504        ClearPageUnevictable(p);
505        /*
506         * drop the page count elevated by isolate_lru_page()
507         */
508        page_cache_release(p);
509        return 0;
510    }
511    return -EIO;
512}
513
514/*
515 * Error hit kernel page.
516 * Do nothing, try to be lucky and not touch this instead. For a few cases we
517 * could be more sophisticated.
518 */
519static int me_kernel(struct page *p, unsigned long pfn)
520{
521    return IGNORED;
522}
523
524/*
525 * Page in unknown state. Do nothing.
526 */
527static int me_unknown(struct page *p, unsigned long pfn)
528{
529    printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
530    return FAILED;
531}
532
533/*
534 * Clean (or cleaned) page cache page.
535 */
536static int me_pagecache_clean(struct page *p, unsigned long pfn)
537{
538    int err;
539    int ret = FAILED;
540    struct address_space *mapping;
541
542    delete_from_lru_cache(p);
543
544    /*
545     * For anonymous pages we're done the only reference left
546     * should be the one m_f() holds.
547     */
548    if (PageAnon(p))
549        return RECOVERED;
550
551    /*
552     * Now truncate the page in the page cache. This is really
553     * more like a "temporary hole punch"
554     * Don't do this for block devices when someone else
555     * has a reference, because it could be file system metadata
556     * and that's not safe to truncate.
557     */
558    mapping = page_mapping(p);
559    if (!mapping) {
560        /*
561         * Page has been teared down in the meanwhile
562         */
563        return FAILED;
564    }
565
566    /*
567     * Truncation is a bit tricky. Enable it per file system for now.
568     *
569     * Open: to take i_mutex or not for this? Right now we don't.
570     */
571    if (mapping->a_ops->error_remove_page) {
572        err = mapping->a_ops->error_remove_page(mapping, p);
573        if (err != 0) {
574            printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
575                    pfn, err);
576        } else if (page_has_private(p) &&
577                !try_to_release_page(p, GFP_NOIO)) {
578            pr_debug("MCE %#lx: failed to release buffers\n", pfn);
579        } else {
580            ret = RECOVERED;
581        }
582    } else {
583        /*
584         * If the file system doesn't support it just invalidate
585         * This fails on dirty or anything with private pages
586         */
587        if (invalidate_inode_page(p))
588            ret = RECOVERED;
589        else
590            printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
591                pfn);
592    }
593    return ret;
594}
595
596/*
597 * Dirty cache page page
598 * Issues: when the error hit a hole page the error is not properly
599 * propagated.
600 */
601static int me_pagecache_dirty(struct page *p, unsigned long pfn)
602{
603    struct address_space *mapping = page_mapping(p);
604
605    SetPageError(p);
606    /* TBD: print more information about the file. */
607    if (mapping) {
608        /*
609         * IO error will be reported by write(), fsync(), etc.
610         * who check the mapping.
611         * This way the application knows that something went
612         * wrong with its dirty file data.
613         *
614         * There's one open issue:
615         *
616         * The EIO will be only reported on the next IO
617         * operation and then cleared through the IO map.
618         * Normally Linux has two mechanisms to pass IO error
619         * first through the AS_EIO flag in the address space
620         * and then through the PageError flag in the page.
621         * Since we drop pages on memory failure handling the
622         * only mechanism open to use is through AS_AIO.
623         *
624         * This has the disadvantage that it gets cleared on
625         * the first operation that returns an error, while
626         * the PageError bit is more sticky and only cleared
627         * when the page is reread or dropped. If an
628         * application assumes it will always get error on
629         * fsync, but does other operations on the fd before
630         * and the page is dropped inbetween then the error
631         * will not be properly reported.
632         *
633         * This can already happen even without hwpoisoned
634         * pages: first on metadata IO errors (which only
635         * report through AS_EIO) or when the page is dropped
636         * at the wrong time.
637         *
638         * So right now we assume that the application DTRT on
639         * the first EIO, but we're not worse than other parts
640         * of the kernel.
641         */
642        mapping_set_error(mapping, EIO);
643    }
644
645    return me_pagecache_clean(p, pfn);
646}
647
648/*
649 * Clean and dirty swap cache.
650 *
651 * Dirty swap cache page is tricky to handle. The page could live both in page
652 * cache and swap cache(ie. page is freshly swapped in). So it could be
653 * referenced concurrently by 2 types of PTEs:
654 * normal PTEs and swap PTEs. We try to handle them consistently by calling
655 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
656 * and then
657 * - clear dirty bit to prevent IO
658 * - remove from LRU
659 * - but keep in the swap cache, so that when we return to it on
660 * a later page fault, we know the application is accessing
661 * corrupted data and shall be killed (we installed simple
662 * interception code in do_swap_page to catch it).
663 *
664 * Clean swap cache pages can be directly isolated. A later page fault will
665 * bring in the known good data from disk.
666 */
667static int me_swapcache_dirty(struct page *p, unsigned long pfn)
668{
669    ClearPageDirty(p);
670    /* Trigger EIO in shmem: */
671    ClearPageUptodate(p);
672
673    if (!delete_from_lru_cache(p))
674        return DELAYED;
675    else
676        return FAILED;
677}
678
679static int me_swapcache_clean(struct page *p, unsigned long pfn)
680{
681    delete_from_swap_cache(p);
682
683    if (!delete_from_lru_cache(p))
684        return RECOVERED;
685    else
686        return FAILED;
687}
688
689/*
690 * Huge pages. Needs work.
691 * Issues:
692 * No rmap support so we cannot find the original mapper. In theory could walk
693 * all MMs and look for the mappings, but that would be non atomic and racy.
694 * Need rmap for hugepages for this. Alternatively we could employ a heuristic,
695 * like just walking the current process and hoping it has it mapped (that
696 * should be usually true for the common "shared database cache" case)
697 * Should handle free huge pages and dequeue them too, but this needs to
698 * handle huge page accounting correctly.
699 */
700static int me_huge_page(struct page *p, unsigned long pfn)
701{
702    return FAILED;
703}
704
705/*
706 * Various page states we can handle.
707 *
708 * A page state is defined by its current page->flags bits.
709 * The table matches them in order and calls the right handler.
710 *
711 * This is quite tricky because we can access page at any time
712 * in its live cycle, so all accesses have to be extremly careful.
713 *
714 * This is not complete. More states could be added.
715 * For any missing state don't attempt recovery.
716 */
717
718#define dirty (1UL << PG_dirty)
719#define sc (1UL << PG_swapcache)
720#define unevict (1UL << PG_unevictable)
721#define mlock (1UL << PG_mlocked)
722#define writeback (1UL << PG_writeback)
723#define lru (1UL << PG_lru)
724#define swapbacked (1UL << PG_swapbacked)
725#define head (1UL << PG_head)
726#define tail (1UL << PG_tail)
727#define compound (1UL << PG_compound)
728#define slab (1UL << PG_slab)
729#define reserved (1UL << PG_reserved)
730
731static struct page_state {
732    unsigned long mask;
733    unsigned long res;
734    char *msg;
735    int (*action)(struct page *p, unsigned long pfn);
736} error_states[] = {
737    { reserved, reserved, "reserved kernel", me_kernel },
738    /*
739     * free pages are specially detected outside this table:
740     * PG_buddy pages only make a small fraction of all free pages.
741     */
742
743    /*
744     * Could in theory check if slab page is free or if we can drop
745     * currently unused objects without touching them. But just
746     * treat it as standard kernel for now.
747     */
748    { slab, slab, "kernel slab", me_kernel },
749
750#ifdef CONFIG_PAGEFLAGS_EXTENDED
751    { head, head, "huge", me_huge_page },
752    { tail, tail, "huge", me_huge_page },
753#else
754    { compound, compound, "huge", me_huge_page },
755#endif
756
757    { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
758    { sc|dirty, sc, "swapcache", me_swapcache_clean },
759
760    { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
761    { unevict, unevict, "unevictable LRU", me_pagecache_clean},
762
763    { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
764    { mlock, mlock, "mlocked LRU", me_pagecache_clean },
765
766    { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
767    { lru|dirty, lru, "clean LRU", me_pagecache_clean },
768
769    /*
770     * Catchall entry: must be at end.
771     */
772    { 0, 0, "unknown page state", me_unknown },
773};
774
775#undef dirty
776#undef sc
777#undef unevict
778#undef mlock
779#undef writeback
780#undef lru
781#undef swapbacked
782#undef head
783#undef tail
784#undef compound
785#undef slab
786#undef reserved
787
788static void action_result(unsigned long pfn, char *msg, int result)
789{
790    struct page *page = pfn_to_page(pfn);
791
792    printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
793        pfn,
794        PageDirty(page) ? "dirty " : "",
795        msg, action_name[result]);
796}
797
798static int page_action(struct page_state *ps, struct page *p,
799            unsigned long pfn)
800{
801    int result;
802    int count;
803
804    result = ps->action(p, pfn);
805    action_result(pfn, ps->msg, result);
806
807    count = page_count(p) - 1;
808    if (ps->action == me_swapcache_dirty && result == DELAYED)
809        count--;
810    if (count != 0) {
811        printk(KERN_ERR
812               "MCE %#lx: %s page still referenced by %d users\n",
813               pfn, ps->msg, count);
814        result = FAILED;
815    }
816
817    /* Could do more checks here if page looks ok */
818    /*
819     * Could adjust zone counters here to correct for the missing page.
820     */
821
822    return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
823}
824
825#define N_UNMAP_TRIES 5
826
827/*
828 * Do all that is necessary to remove user space mappings. Unmap
829 * the pages and send SIGBUS to the processes if the data was dirty.
830 */
831static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
832                  int trapno)
833{
834    enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
835    struct address_space *mapping;
836    LIST_HEAD(tokill);
837    int ret;
838    int i;
839    int kill = 1;
840
841    if (PageReserved(p) || PageSlab(p))
842        return SWAP_SUCCESS;
843
844    /*
845     * This check implies we don't kill processes if their pages
846     * are in the swap cache early. Those are always late kills.
847     */
848    if (!page_mapped(p))
849        return SWAP_SUCCESS;
850
851    if (PageCompound(p) || PageKsm(p))
852        return SWAP_FAIL;
853
854    if (PageSwapCache(p)) {
855        printk(KERN_ERR
856               "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
857        ttu |= TTU_IGNORE_HWPOISON;
858    }
859
860    /*
861     * Propagate the dirty bit from PTEs to struct page first, because we
862     * need this to decide if we should kill or just drop the page.
863     * XXX: the dirty test could be racy: set_page_dirty() may not always
864     * be called inside page lock (it's recommended but not enforced).
865     */
866    mapping = page_mapping(p);
867    if (!PageDirty(p) && mapping && mapping_cap_writeback_dirty(mapping)) {
868        if (page_mkclean(p)) {
869            SetPageDirty(p);
870        } else {
871            kill = 0;
872            ttu |= TTU_IGNORE_HWPOISON;
873            printk(KERN_INFO
874    "MCE %#lx: corrupted page was clean: dropped without side effects\n",
875                pfn);
876        }
877    }
878
879    /*
880     * First collect all the processes that have the page
881     * mapped in dirty form. This has to be done before try_to_unmap,
882     * because ttu takes the rmap data structures down.
883     *
884     * Error handling: We ignore errors here because
885     * there's nothing that can be done.
886     */
887    if (kill)
888        collect_procs(p, &tokill);
889
890    /*
891     * try_to_unmap can fail temporarily due to races.
892     * Try a few times (RED-PEN better strategy?)
893     */
894    for (i = 0; i < N_UNMAP_TRIES; i++) {
895        ret = try_to_unmap(p, ttu);
896        if (ret == SWAP_SUCCESS)
897            break;
898        pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn, ret);
899    }
900
901    if (ret != SWAP_SUCCESS)
902        printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
903                pfn, page_mapcount(p));
904
905    /*
906     * Now that the dirty bit has been propagated to the
907     * struct page and all unmaps done we can decide if
908     * killing is needed or not. Only kill when the page
909     * was dirty, otherwise the tokill list is merely
910     * freed. When there was a problem unmapping earlier
911     * use a more force-full uncatchable kill to prevent
912     * any accesses to the poisoned memory.
913     */
914    kill_procs_ao(&tokill, !!PageDirty(p), trapno,
915              ret != SWAP_SUCCESS, pfn);
916
917    return ret;
918}
919
920int __memory_failure(unsigned long pfn, int trapno, int flags)
921{
922    struct page_state *ps;
923    struct page *p;
924    int res;
925
926    if (!sysctl_memory_failure_recovery)
927        panic("Memory failure from trap %d on page %lx", trapno, pfn);
928
929    if (!pfn_valid(pfn)) {
930        printk(KERN_ERR
931               "MCE %#lx: memory outside kernel control\n",
932               pfn);
933        return -ENXIO;
934    }
935
936    p = pfn_to_page(pfn);
937    if (TestSetPageHWPoison(p)) {
938        printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
939        return 0;
940    }
941
942    atomic_long_add(1, &mce_bad_pages);
943
944    /*
945     * We need/can do nothing about count=0 pages.
946     * 1) it's a free page, and therefore in safe hand:
947     * prep_new_page() will be the gate keeper.
948     * 2) it's part of a non-compound high order page.
949     * Implies some kernel user: cannot stop them from
950     * R/W the page; let's pray that the page has been
951     * used and will be freed some time later.
952     * In fact it's dangerous to directly bump up page count from 0,
953     * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
954     */
955    if (!(flags & MF_COUNT_INCREASED) &&
956        !get_page_unless_zero(compound_head(p))) {
957        if (is_free_buddy_page(p)) {
958            action_result(pfn, "free buddy", DELAYED);
959            return 0;
960        } else {
961            action_result(pfn, "high order kernel", IGNORED);
962            return -EBUSY;
963        }
964    }
965
966    /*
967     * We ignore non-LRU pages for good reasons.
968     * - PG_locked is only well defined for LRU pages and a few others
969     * - to avoid races with __set_page_locked()
970     * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
971     * The check (unnecessarily) ignores LRU pages being isolated and
972     * walked by the page reclaim code, however that's not a big loss.
973     */
974    if (!PageLRU(p))
975        shake_page(p, 0);
976    if (!PageLRU(p)) {
977        /*
978         * shake_page could have turned it free.
979         */
980        if (is_free_buddy_page(p)) {
981            action_result(pfn, "free buddy, 2nd try", DELAYED);
982            return 0;
983        }
984        action_result(pfn, "non LRU", IGNORED);
985        put_page(p);
986        return -EBUSY;
987    }
988
989    /*
990     * Lock the page and wait for writeback to finish.
991     * It's very difficult to mess with pages currently under IO
992     * and in many cases impossible, so we just avoid it here.
993     */
994    lock_page_nosync(p);
995
996    /*
997     * unpoison always clear PG_hwpoison inside page lock
998     */
999    if (!PageHWPoison(p)) {
1000        printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1001        res = 0;
1002        goto out;
1003    }
1004    if (hwpoison_filter(p)) {
1005        if (TestClearPageHWPoison(p))
1006            atomic_long_dec(&mce_bad_pages);
1007        unlock_page(p);
1008        put_page(p);
1009        return 0;
1010    }
1011
1012    wait_on_page_writeback(p);
1013
1014    /*
1015     * Now take care of user space mappings.
1016     * Abort on fail: __remove_from_page_cache() assumes unmapped page.
1017     */
1018    if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
1019        printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1020        res = -EBUSY;
1021        goto out;
1022    }
1023
1024    /*
1025     * Torn down by someone else?
1026     */
1027    if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1028        action_result(pfn, "already truncated LRU", IGNORED);
1029        res = -EBUSY;
1030        goto out;
1031    }
1032
1033    res = -EBUSY;
1034    for (ps = error_states;; ps++) {
1035        if ((p->flags & ps->mask) == ps->res) {
1036            res = page_action(ps, p, pfn);
1037            break;
1038        }
1039    }
1040out:
1041    unlock_page(p);
1042    return res;
1043}
1044EXPORT_SYMBOL_GPL(__memory_failure);
1045
1046/**
1047 * memory_failure - Handle memory failure of a page.
1048 * @pfn: Page Number of the corrupted page
1049 * @trapno: Trap number reported in the signal to user space.
1050 *
1051 * This function is called by the low level machine check code
1052 * of an architecture when it detects hardware memory corruption
1053 * of a page. It tries its best to recover, which includes
1054 * dropping pages, killing processes etc.
1055 *
1056 * The function is primarily of use for corruptions that
1057 * happen outside the current execution context (e.g. when
1058 * detected by a background scrubber)
1059 *
1060 * Must run in process context (e.g. a work queue) with interrupts
1061 * enabled and no spinlocks hold.
1062 */
1063void memory_failure(unsigned long pfn, int trapno)
1064{
1065    __memory_failure(pfn, trapno, 0);
1066}
1067
1068/**
1069 * unpoison_memory - Unpoison a previously poisoned page
1070 * @pfn: Page number of the to be unpoisoned page
1071 *
1072 * Software-unpoison a page that has been poisoned by
1073 * memory_failure() earlier.
1074 *
1075 * This is only done on the software-level, so it only works
1076 * for linux injected failures, not real hardware failures
1077 *
1078 * Returns 0 for success, otherwise -errno.
1079 */
1080int unpoison_memory(unsigned long pfn)
1081{
1082    struct page *page;
1083    struct page *p;
1084    int freeit = 0;
1085
1086    if (!pfn_valid(pfn))
1087        return -ENXIO;
1088
1089    p = pfn_to_page(pfn);
1090    page = compound_head(p);
1091
1092    if (!PageHWPoison(p)) {
1093        pr_debug("MCE: Page was already unpoisoned %#lx\n", pfn);
1094        return 0;
1095    }
1096
1097    if (!get_page_unless_zero(page)) {
1098        if (TestClearPageHWPoison(p))
1099            atomic_long_dec(&mce_bad_pages);
1100        pr_debug("MCE: Software-unpoisoned free page %#lx\n", pfn);
1101        return 0;
1102    }
1103
1104    lock_page_nosync(page);
1105    /*
1106     * This test is racy because PG_hwpoison is set outside of page lock.
1107     * That's acceptable because that won't trigger kernel panic. Instead,
1108     * the PG_hwpoison page will be caught and isolated on the entrance to
1109     * the free buddy page pool.
1110     */
1111    if (TestClearPageHWPoison(p)) {
1112        pr_debug("MCE: Software-unpoisoned page %#lx\n", pfn);
1113        atomic_long_dec(&mce_bad_pages);
1114        freeit = 1;
1115    }
1116    unlock_page(page);
1117
1118    put_page(page);
1119    if (freeit)
1120        put_page(page);
1121
1122    return 0;
1123}
1124EXPORT_SYMBOL(unpoison_memory);
1125
1126static struct page *new_page(struct page *p, unsigned long private, int **x)
1127{
1128    int nid = page_to_nid(p);
1129    return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1130}
1131
1132/*
1133 * Safely get reference count of an arbitrary page.
1134 * Returns 0 for a free page, -EIO for a zero refcount page
1135 * that is not free, and 1 for any other page type.
1136 * For 1 the page is returned with increased page count, otherwise not.
1137 */
1138static int get_any_page(struct page *p, unsigned long pfn, int flags)
1139{
1140    int ret;
1141
1142    if (flags & MF_COUNT_INCREASED)
1143        return 1;
1144
1145    /*
1146     * The lock_system_sleep prevents a race with memory hotplug,
1147     * because the isolation assumes there's only a single user.
1148     * This is a big hammer, a better would be nicer.
1149     */
1150    lock_system_sleep();
1151
1152    /*
1153     * Isolate the page, so that it doesn't get reallocated if it
1154     * was free.
1155     */
1156    set_migratetype_isolate(p);
1157    if (!get_page_unless_zero(compound_head(p))) {
1158        if (is_free_buddy_page(p)) {
1159            pr_debug("get_any_page: %#lx free buddy page\n", pfn);
1160            /* Set hwpoison bit while page is still isolated */
1161            SetPageHWPoison(p);
1162            ret = 0;
1163        } else {
1164            pr_debug("get_any_page: %#lx: unknown zero refcount page type %lx\n",
1165                pfn, p->flags);
1166            ret = -EIO;
1167        }
1168    } else {
1169        /* Not a free page */
1170        ret = 1;
1171    }
1172    unset_migratetype_isolate(p);
1173    unlock_system_sleep();
1174    return ret;
1175}
1176
1177/**
1178 * soft_offline_page - Soft offline a page.
1179 * @page: page to offline
1180 * @flags: flags. Same as memory_failure().
1181 *
1182 * Returns 0 on success, otherwise negated errno.
1183 *
1184 * Soft offline a page, by migration or invalidation,
1185 * without killing anything. This is for the case when
1186 * a page is not corrupted yet (so it's still valid to access),
1187 * but has had a number of corrected errors and is better taken
1188 * out.
1189 *
1190 * The actual policy on when to do that is maintained by
1191 * user space.
1192 *
1193 * This should never impact any application or cause data loss,
1194 * however it might take some time.
1195 *
1196 * This is not a 100% solution for all memory, but tries to be
1197 * ``good enough'' for the majority of memory.
1198 */
1199int soft_offline_page(struct page *page, int flags)
1200{
1201    int ret;
1202    unsigned long pfn = page_to_pfn(page);
1203
1204    ret = get_any_page(page, pfn, flags);
1205    if (ret < 0)
1206        return ret;
1207    if (ret == 0)
1208        goto done;
1209
1210    /*
1211     * Page cache page we can handle?
1212     */
1213    if (!PageLRU(page)) {
1214        /*
1215         * Try to free it.
1216         */
1217        put_page(page);
1218        shake_page(page, 1);
1219
1220        /*
1221         * Did it turn free?
1222         */
1223        ret = get_any_page(page, pfn, 0);
1224        if (ret < 0)
1225            return ret;
1226        if (ret == 0)
1227            goto done;
1228    }
1229    if (!PageLRU(page)) {
1230        pr_debug("soft_offline: %#lx: unknown non LRU page type %lx\n",
1231                pfn, page->flags);
1232        return -EIO;
1233    }
1234
1235    lock_page(page);
1236    wait_on_page_writeback(page);
1237
1238    /*
1239     * Synchronized using the page lock with memory_failure()
1240     */
1241    if (PageHWPoison(page)) {
1242        unlock_page(page);
1243        put_page(page);
1244        pr_debug("soft offline: %#lx page already poisoned\n", pfn);
1245        return -EBUSY;
1246    }
1247
1248    /*
1249     * Try to invalidate first. This should work for
1250     * non dirty unmapped page cache pages.
1251     */
1252    ret = invalidate_inode_page(page);
1253    unlock_page(page);
1254
1255    /*
1256     * Drop count because page migration doesn't like raised
1257     * counts. The page could get re-allocated, but if it becomes
1258     * LRU the isolation will just fail.
1259     * RED-PEN would be better to keep it isolated here, but we
1260     * would need to fix isolation locking first.
1261     */
1262    put_page(page);
1263    if (ret == 1) {
1264        ret = 0;
1265        pr_debug("soft_offline: %#lx: invalidated\n", pfn);
1266        goto done;
1267    }
1268
1269    /*
1270     * Simple invalidation didn't work.
1271     * Try to migrate to a new page instead. migrate.c
1272     * handles a large number of cases for us.
1273     */
1274    ret = isolate_lru_page(page);
1275    if (!ret) {
1276        LIST_HEAD(pagelist);
1277
1278        list_add(&page->lru, &pagelist);
1279        ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0);
1280        if (ret) {
1281            pr_debug("soft offline: %#lx: migration failed %d, type %lx\n",
1282                pfn, ret, page->flags);
1283            if (ret > 0)
1284                ret = -EIO;
1285        }
1286    } else {
1287        pr_debug("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1288                pfn, ret, page_count(page), page->flags);
1289    }
1290    if (ret)
1291        return ret;
1292
1293done:
1294    atomic_long_add(1, &mce_bad_pages);
1295    SetPageHWPoison(page);
1296    /* keep elevated page count for bad page */
1297    return ret;
1298}
1299

Archive Download this file



interactive