Back to home page

LXR

 
 

    


0001 /*
0002  * Copyright (C) 2008, 2009 Intel Corporation
0003  * Authors: Andi Kleen, Fengguang Wu
0004  *
0005  * This software may be redistributed and/or modified under the terms of
0006  * the GNU General Public License ("GPL") version 2 only as published by the
0007  * Free Software Foundation.
0008  *
0009  * High level machine check handler. Handles pages reported by the
0010  * hardware as being corrupted usually due to a multi-bit ECC memory or cache
0011  * failure.
0012  * 
0013  * In addition there is a "soft offline" entry point that allows stop using
0014  * not-yet-corrupted-by-suspicious pages without killing anything.
0015  *
0016  * Handles page cache pages in various states.  The tricky part
0017  * here is that we can access any page asynchronously in respect to 
0018  * other VM users, because memory failures could happen anytime and 
0019  * anywhere. This could violate some of their assumptions. This is why 
0020  * this code has to be extremely careful. Generally it tries to use 
0021  * normal locking rules, as in get the standard locks, even if that means 
0022  * the error handling takes potentially a long time.
0023  *
0024  * It can be very tempting to add handling for obscure cases here.
0025  * In general any code for handling new cases should only be added iff:
0026  * - You know how to test it.
0027  * - You have a test that can be added to mce-test
0028  *   https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/
0029  * - The case actually shows up as a frequent (top 10) page state in
0030  *   tools/vm/page-types when running a real workload.
0031  * 
0032  * There are several operations here with exponential complexity because
0033  * of unsuitable VM data structures. For example the operation to map back 
0034  * from RMAP chains to processes has to walk the complete process list and 
0035  * has non linear complexity with the number. But since memory corruptions
0036  * are rare we hope to get away with this. This avoids impacting the core 
0037  * VM.
0038  */
0039 #include <linux/kernel.h>
0040 #include <linux/mm.h>
0041 #include <linux/page-flags.h>
0042 #include <linux/kernel-page-flags.h>
0043 #include <linux/sched.h>
0044 #include <linux/ksm.h>
0045 #include <linux/rmap.h>
0046 #include <linux/export.h>
0047 #include <linux/pagemap.h>
0048 #include <linux/swap.h>
0049 #include <linux/backing-dev.h>
0050 #include <linux/migrate.h>
0051 #include <linux/page-isolation.h>
0052 #include <linux/suspend.h>
0053 #include <linux/slab.h>
0054 #include <linux/swapops.h>
0055 #include <linux/hugetlb.h>
0056 #include <linux/memory_hotplug.h>
0057 #include <linux/mm_inline.h>
0058 #include <linux/kfifo.h>
0059 #include <linux/ratelimit.h>
0060 #include "internal.h"
0061 #include "ras/ras_event.h"
0062 
0063 int sysctl_memory_failure_early_kill __read_mostly = 0;
0064 
0065 int sysctl_memory_failure_recovery __read_mostly = 1;
0066 
0067 atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
0068 
0069 #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
0070 
0071 u32 hwpoison_filter_enable = 0;
0072 u32 hwpoison_filter_dev_major = ~0U;
0073 u32 hwpoison_filter_dev_minor = ~0U;
0074 u64 hwpoison_filter_flags_mask;
0075 u64 hwpoison_filter_flags_value;
0076 EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
0077 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
0078 EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
0079 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
0080 EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
0081 
0082 static int hwpoison_filter_dev(struct page *p)
0083 {
0084     struct address_space *mapping;
0085     dev_t dev;
0086 
0087     if (hwpoison_filter_dev_major == ~0U &&
0088         hwpoison_filter_dev_minor == ~0U)
0089         return 0;
0090 
0091     /*
0092      * page_mapping() does not accept slab pages.
0093      */
0094     if (PageSlab(p))
0095         return -EINVAL;
0096 
0097     mapping = page_mapping(p);
0098     if (mapping == NULL || mapping->host == NULL)
0099         return -EINVAL;
0100 
0101     dev = mapping->host->i_sb->s_dev;
0102     if (hwpoison_filter_dev_major != ~0U &&
0103         hwpoison_filter_dev_major != MAJOR(dev))
0104         return -EINVAL;
0105     if (hwpoison_filter_dev_minor != ~0U &&
0106         hwpoison_filter_dev_minor != MINOR(dev))
0107         return -EINVAL;
0108 
0109     return 0;
0110 }
0111 
0112 static int hwpoison_filter_flags(struct page *p)
0113 {
0114     if (!hwpoison_filter_flags_mask)
0115         return 0;
0116 
0117     if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
0118                     hwpoison_filter_flags_value)
0119         return 0;
0120     else
0121         return -EINVAL;
0122 }
0123 
0124 /*
0125  * This allows stress tests to limit test scope to a collection of tasks
0126  * by putting them under some memcg. This prevents killing unrelated/important
0127  * processes such as /sbin/init. Note that the target task may share clean
0128  * pages with init (eg. libc text), which is harmless. If the target task
0129  * share _dirty_ pages with another task B, the test scheme must make sure B
0130  * is also included in the memcg. At last, due to race conditions this filter
0131  * can only guarantee that the page either belongs to the memcg tasks, or is
0132  * a freed page.
0133  */
0134 #ifdef CONFIG_MEMCG
0135 u64 hwpoison_filter_memcg;
0136 EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
0137 static int hwpoison_filter_task(struct page *p)
0138 {
0139     if (!hwpoison_filter_memcg)
0140         return 0;
0141 
0142     if (page_cgroup_ino(p) != hwpoison_filter_memcg)
0143         return -EINVAL;
0144 
0145     return 0;
0146 }
0147 #else
0148 static int hwpoison_filter_task(struct page *p) { return 0; }
0149 #endif
0150 
0151 int hwpoison_filter(struct page *p)
0152 {
0153     if (!hwpoison_filter_enable)
0154         return 0;
0155 
0156     if (hwpoison_filter_dev(p))
0157         return -EINVAL;
0158 
0159     if (hwpoison_filter_flags(p))
0160         return -EINVAL;
0161 
0162     if (hwpoison_filter_task(p))
0163         return -EINVAL;
0164 
0165     return 0;
0166 }
0167 #else
0168 int hwpoison_filter(struct page *p)
0169 {
0170     return 0;
0171 }
0172 #endif
0173 
0174 EXPORT_SYMBOL_GPL(hwpoison_filter);
0175 
0176 /*
0177  * Send all the processes who have the page mapped a signal.
0178  * ``action optional'' if they are not immediately affected by the error
0179  * ``action required'' if error happened in current execution context
0180  */
0181 static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
0182             unsigned long pfn, struct page *page, int flags)
0183 {
0184     struct siginfo si;
0185     int ret;
0186 
0187     pr_err("Memory failure: %#lx: Killing %s:%d due to hardware memory corruption\n",
0188         pfn, t->comm, t->pid);
0189     si.si_signo = SIGBUS;
0190     si.si_errno = 0;
0191     si.si_addr = (void *)addr;
0192 #ifdef __ARCH_SI_TRAPNO
0193     si.si_trapno = trapno;
0194 #endif
0195     si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
0196 
0197     if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
0198         si.si_code = BUS_MCEERR_AR;
0199         ret = force_sig_info(SIGBUS, &si, current);
0200     } else {
0201         /*
0202          * Don't use force here, it's convenient if the signal
0203          * can be temporarily blocked.
0204          * This could cause a loop when the user sets SIGBUS
0205          * to SIG_IGN, but hopefully no one will do that?
0206          */
0207         si.si_code = BUS_MCEERR_AO;
0208         ret = send_sig_info(SIGBUS, &si, t);  /* synchronous? */
0209     }
0210     if (ret < 0)
0211         pr_info("Memory failure: Error sending signal to %s:%d: %d\n",
0212             t->comm, t->pid, ret);
0213     return ret;
0214 }
0215 
0216 /*
0217  * When a unknown page type is encountered drain as many buffers as possible
0218  * in the hope to turn the page into a LRU or free page, which we can handle.
0219  */
0220 void shake_page(struct page *p, int access)
0221 {
0222     if (!PageSlab(p)) {
0223         lru_add_drain_all();
0224         if (PageLRU(p))
0225             return;
0226         drain_all_pages(page_zone(p));
0227         if (PageLRU(p) || is_free_buddy_page(p))
0228             return;
0229     }
0230 
0231     /*
0232      * Only call shrink_node_slabs here (which would also shrink
0233      * other caches) if access is not potentially fatal.
0234      */
0235     if (access)
0236         drop_slab_node(page_to_nid(p));
0237 }
0238 EXPORT_SYMBOL_GPL(shake_page);
0239 
0240 /*
0241  * Kill all processes that have a poisoned page mapped and then isolate
0242  * the page.
0243  *
0244  * General strategy:
0245  * Find all processes having the page mapped and kill them.
0246  * But we keep a page reference around so that the page is not
0247  * actually freed yet.
0248  * Then stash the page away
0249  *
0250  * There's no convenient way to get back to mapped processes
0251  * from the VMAs. So do a brute-force search over all
0252  * running processes.
0253  *
0254  * Remember that machine checks are not common (or rather
0255  * if they are common you have other problems), so this shouldn't
0256  * be a performance issue.
0257  *
0258  * Also there are some races possible while we get from the
0259  * error detection to actually handle it.
0260  */
0261 
0262 struct to_kill {
0263     struct list_head nd;
0264     struct task_struct *tsk;
0265     unsigned long addr;
0266     char addr_valid;
0267 };
0268 
0269 /*
0270  * Failure handling: if we can't find or can't kill a process there's
0271  * not much we can do.  We just print a message and ignore otherwise.
0272  */
0273 
0274 /*
0275  * Schedule a process for later kill.
0276  * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
0277  * TBD would GFP_NOIO be enough?
0278  */
0279 static void add_to_kill(struct task_struct *tsk, struct page *p,
0280                struct vm_area_struct *vma,
0281                struct list_head *to_kill,
0282                struct to_kill **tkc)
0283 {
0284     struct to_kill *tk;
0285 
0286     if (*tkc) {
0287         tk = *tkc;
0288         *tkc = NULL;
0289     } else {
0290         tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
0291         if (!tk) {
0292             pr_err("Memory failure: Out of memory while machine check handling\n");
0293             return;
0294         }
0295     }
0296     tk->addr = page_address_in_vma(p, vma);
0297     tk->addr_valid = 1;
0298 
0299     /*
0300      * In theory we don't have to kill when the page was
0301      * munmaped. But it could be also a mremap. Since that's
0302      * likely very rare kill anyways just out of paranoia, but use
0303      * a SIGKILL because the error is not contained anymore.
0304      */
0305     if (tk->addr == -EFAULT) {
0306         pr_info("Memory failure: Unable to find user space address %lx in %s\n",
0307             page_to_pfn(p), tsk->comm);
0308         tk->addr_valid = 0;
0309     }
0310     get_task_struct(tsk);
0311     tk->tsk = tsk;
0312     list_add_tail(&tk->nd, to_kill);
0313 }
0314 
0315 /*
0316  * Kill the processes that have been collected earlier.
0317  *
0318  * Only do anything when DOIT is set, otherwise just free the list
0319  * (this is used for clean pages which do not need killing)
0320  * Also when FAIL is set do a force kill because something went
0321  * wrong earlier.
0322  */
0323 static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
0324               int fail, struct page *page, unsigned long pfn,
0325               int flags)
0326 {
0327     struct to_kill *tk, *next;
0328 
0329     list_for_each_entry_safe (tk, next, to_kill, nd) {
0330         if (forcekill) {
0331             /*
0332              * In case something went wrong with munmapping
0333              * make sure the process doesn't catch the
0334              * signal and then access the memory. Just kill it.
0335              */
0336             if (fail || tk->addr_valid == 0) {
0337                 pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
0338                        pfn, tk->tsk->comm, tk->tsk->pid);
0339                 force_sig(SIGKILL, tk->tsk);
0340             }
0341 
0342             /*
0343              * In theory the process could have mapped
0344              * something else on the address in-between. We could
0345              * check for that, but we need to tell the
0346              * process anyways.
0347              */
0348             else if (kill_proc(tk->tsk, tk->addr, trapno,
0349                           pfn, page, flags) < 0)
0350                 pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n",
0351                        pfn, tk->tsk->comm, tk->tsk->pid);
0352         }
0353         put_task_struct(tk->tsk);
0354         kfree(tk);
0355     }
0356 }
0357 
0358 /*
0359  * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
0360  * on behalf of the thread group. Return task_struct of the (first found)
0361  * dedicated thread if found, and return NULL otherwise.
0362  *
0363  * We already hold read_lock(&tasklist_lock) in the caller, so we don't
0364  * have to call rcu_read_lock/unlock() in this function.
0365  */
0366 static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
0367 {
0368     struct task_struct *t;
0369 
0370     for_each_thread(tsk, t)
0371         if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
0372             return t;
0373     return NULL;
0374 }
0375 
0376 /*
0377  * Determine whether a given process is "early kill" process which expects
0378  * to be signaled when some page under the process is hwpoisoned.
0379  * Return task_struct of the dedicated thread (main thread unless explicitly
0380  * specified) if the process is "early kill," and otherwise returns NULL.
0381  */
0382 static struct task_struct *task_early_kill(struct task_struct *tsk,
0383                        int force_early)
0384 {
0385     struct task_struct *t;
0386     if (!tsk->mm)
0387         return NULL;
0388     if (force_early)
0389         return tsk;
0390     t = find_early_kill_thread(tsk);
0391     if (t)
0392         return t;
0393     if (sysctl_memory_failure_early_kill)
0394         return tsk;
0395     return NULL;
0396 }
0397 
0398 /*
0399  * Collect processes when the error hit an anonymous page.
0400  */
0401 static void collect_procs_anon(struct page *page, struct list_head *to_kill,
0402                   struct to_kill **tkc, int force_early)
0403 {
0404     struct vm_area_struct *vma;
0405     struct task_struct *tsk;
0406     struct anon_vma *av;
0407     pgoff_t pgoff;
0408 
0409     av = page_lock_anon_vma_read(page);
0410     if (av == NULL) /* Not actually mapped anymore */
0411         return;
0412 
0413     pgoff = page_to_pgoff(page);
0414     read_lock(&tasklist_lock);
0415     for_each_process (tsk) {
0416         struct anon_vma_chain *vmac;
0417         struct task_struct *t = task_early_kill(tsk, force_early);
0418 
0419         if (!t)
0420             continue;
0421         anon_vma_interval_tree_foreach(vmac, &av->rb_root,
0422                            pgoff, pgoff) {
0423             vma = vmac->vma;
0424             if (!page_mapped_in_vma(page, vma))
0425                 continue;
0426             if (vma->vm_mm == t->mm)
0427                 add_to_kill(t, page, vma, to_kill, tkc);
0428         }
0429     }
0430     read_unlock(&tasklist_lock);
0431     page_unlock_anon_vma_read(av);
0432 }
0433 
0434 /*
0435  * Collect processes when the error hit a file mapped page.
0436  */
0437 static void collect_procs_file(struct page *page, struct list_head *to_kill,
0438                   struct to_kill **tkc, int force_early)
0439 {
0440     struct vm_area_struct *vma;
0441     struct task_struct *tsk;
0442     struct address_space *mapping = page->mapping;
0443 
0444     i_mmap_lock_read(mapping);
0445     read_lock(&tasklist_lock);
0446     for_each_process(tsk) {
0447         pgoff_t pgoff = page_to_pgoff(page);
0448         struct task_struct *t = task_early_kill(tsk, force_early);
0449 
0450         if (!t)
0451             continue;
0452         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
0453                       pgoff) {
0454             /*
0455              * Send early kill signal to tasks where a vma covers
0456              * the page but the corrupted page is not necessarily
0457              * mapped it in its pte.
0458              * Assume applications who requested early kill want
0459              * to be informed of all such data corruptions.
0460              */
0461             if (vma->vm_mm == t->mm)
0462                 add_to_kill(t, page, vma, to_kill, tkc);
0463         }
0464     }
0465     read_unlock(&tasklist_lock);
0466     i_mmap_unlock_read(mapping);
0467 }
0468 
0469 /*
0470  * Collect the processes who have the corrupted page mapped to kill.
0471  * This is done in two steps for locking reasons.
0472  * First preallocate one tokill structure outside the spin locks,
0473  * so that we can kill at least one process reasonably reliable.
0474  */
0475 static void collect_procs(struct page *page, struct list_head *tokill,
0476                 int force_early)
0477 {
0478     struct to_kill *tk;
0479 
0480     if (!page->mapping)
0481         return;
0482 
0483     tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
0484     if (!tk)
0485         return;
0486     if (PageAnon(page))
0487         collect_procs_anon(page, tokill, &tk, force_early);
0488     else
0489         collect_procs_file(page, tokill, &tk, force_early);
0490     kfree(tk);
0491 }
0492 
0493 static const char *action_name[] = {
0494     [MF_IGNORED] = "Ignored",
0495     [MF_FAILED] = "Failed",
0496     [MF_DELAYED] = "Delayed",
0497     [MF_RECOVERED] = "Recovered",
0498 };
0499 
0500 static const char * const action_page_types[] = {
0501     [MF_MSG_KERNEL]         = "reserved kernel page",
0502     [MF_MSG_KERNEL_HIGH_ORDER]  = "high-order kernel page",
0503     [MF_MSG_SLAB]           = "kernel slab page",
0504     [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking",
0505     [MF_MSG_POISONED_HUGE]      = "huge page already hardware poisoned",
0506     [MF_MSG_HUGE]           = "huge page",
0507     [MF_MSG_FREE_HUGE]      = "free huge page",
0508     [MF_MSG_UNMAP_FAILED]       = "unmapping failed page",
0509     [MF_MSG_DIRTY_SWAPCACHE]    = "dirty swapcache page",
0510     [MF_MSG_CLEAN_SWAPCACHE]    = "clean swapcache page",
0511     [MF_MSG_DIRTY_MLOCKED_LRU]  = "dirty mlocked LRU page",
0512     [MF_MSG_CLEAN_MLOCKED_LRU]  = "clean mlocked LRU page",
0513     [MF_MSG_DIRTY_UNEVICTABLE_LRU]  = "dirty unevictable LRU page",
0514     [MF_MSG_CLEAN_UNEVICTABLE_LRU]  = "clean unevictable LRU page",
0515     [MF_MSG_DIRTY_LRU]      = "dirty LRU page",
0516     [MF_MSG_CLEAN_LRU]      = "clean LRU page",
0517     [MF_MSG_TRUNCATED_LRU]      = "already truncated LRU page",
0518     [MF_MSG_BUDDY]          = "free buddy page",
0519     [MF_MSG_BUDDY_2ND]      = "free buddy page (2nd try)",
0520     [MF_MSG_UNKNOWN]        = "unknown page",
0521 };
0522 
0523 /*
0524  * XXX: It is possible that a page is isolated from LRU cache,
0525  * and then kept in swap cache or failed to remove from page cache.
0526  * The page count will stop it from being freed by unpoison.
0527  * Stress tests should be aware of this memory leak problem.
0528  */
0529 static int delete_from_lru_cache(struct page *p)
0530 {
0531     if (!isolate_lru_page(p)) {
0532         /*
0533          * Clear sensible page flags, so that the buddy system won't
0534          * complain when the page is unpoison-and-freed.
0535          */
0536         ClearPageActive(p);
0537         ClearPageUnevictable(p);
0538         /*
0539          * drop the page count elevated by isolate_lru_page()
0540          */
0541         put_page(p);
0542         return 0;
0543     }
0544     return -EIO;
0545 }
0546 
0547 /*
0548  * Error hit kernel page.
0549  * Do nothing, try to be lucky and not touch this instead. For a few cases we
0550  * could be more sophisticated.
0551  */
0552 static int me_kernel(struct page *p, unsigned long pfn)
0553 {
0554     return MF_IGNORED;
0555 }
0556 
0557 /*
0558  * Page in unknown state. Do nothing.
0559  */
0560 static int me_unknown(struct page *p, unsigned long pfn)
0561 {
0562     pr_err("Memory failure: %#lx: Unknown page state\n", pfn);
0563     return MF_FAILED;
0564 }
0565 
0566 /*
0567  * Clean (or cleaned) page cache page.
0568  */
0569 static int me_pagecache_clean(struct page *p, unsigned long pfn)
0570 {
0571     int err;
0572     int ret = MF_FAILED;
0573     struct address_space *mapping;
0574 
0575     delete_from_lru_cache(p);
0576 
0577     /*
0578      * For anonymous pages we're done the only reference left
0579      * should be the one m_f() holds.
0580      */
0581     if (PageAnon(p))
0582         return MF_RECOVERED;
0583 
0584     /*
0585      * Now truncate the page in the page cache. This is really
0586      * more like a "temporary hole punch"
0587      * Don't do this for block devices when someone else
0588      * has a reference, because it could be file system metadata
0589      * and that's not safe to truncate.
0590      */
0591     mapping = page_mapping(p);
0592     if (!mapping) {
0593         /*
0594          * Page has been teared down in the meanwhile
0595          */
0596         return MF_FAILED;
0597     }
0598 
0599     /*
0600      * Truncation is a bit tricky. Enable it per file system for now.
0601      *
0602      * Open: to take i_mutex or not for this? Right now we don't.
0603      */
0604     if (mapping->a_ops->error_remove_page) {
0605         err = mapping->a_ops->error_remove_page(mapping, p);
0606         if (err != 0) {
0607             pr_info("Memory failure: %#lx: Failed to punch page: %d\n",
0608                 pfn, err);
0609         } else if (page_has_private(p) &&
0610                 !try_to_release_page(p, GFP_NOIO)) {
0611             pr_info("Memory failure: %#lx: failed to release buffers\n",
0612                 pfn);
0613         } else {
0614             ret = MF_RECOVERED;
0615         }
0616     } else {
0617         /*
0618          * If the file system doesn't support it just invalidate
0619          * This fails on dirty or anything with private pages
0620          */
0621         if (invalidate_inode_page(p))
0622             ret = MF_RECOVERED;
0623         else
0624             pr_info("Memory failure: %#lx: Failed to invalidate\n",
0625                 pfn);
0626     }
0627     return ret;
0628 }
0629 
0630 /*
0631  * Dirty pagecache page
0632  * Issues: when the error hit a hole page the error is not properly
0633  * propagated.
0634  */
0635 static int me_pagecache_dirty(struct page *p, unsigned long pfn)
0636 {
0637     struct address_space *mapping = page_mapping(p);
0638 
0639     SetPageError(p);
0640     /* TBD: print more information about the file. */
0641     if (mapping) {
0642         /*
0643          * IO error will be reported by write(), fsync(), etc.
0644          * who check the mapping.
0645          * This way the application knows that something went
0646          * wrong with its dirty file data.
0647          *
0648          * There's one open issue:
0649          *
0650          * The EIO will be only reported on the next IO
0651          * operation and then cleared through the IO map.
0652          * Normally Linux has two mechanisms to pass IO error
0653          * first through the AS_EIO flag in the address space
0654          * and then through the PageError flag in the page.
0655          * Since we drop pages on memory failure handling the
0656          * only mechanism open to use is through AS_AIO.
0657          *
0658          * This has the disadvantage that it gets cleared on
0659          * the first operation that returns an error, while
0660          * the PageError bit is more sticky and only cleared
0661          * when the page is reread or dropped.  If an
0662          * application assumes it will always get error on
0663          * fsync, but does other operations on the fd before
0664          * and the page is dropped between then the error
0665          * will not be properly reported.
0666          *
0667          * This can already happen even without hwpoisoned
0668          * pages: first on metadata IO errors (which only
0669          * report through AS_EIO) or when the page is dropped
0670          * at the wrong time.
0671          *
0672          * So right now we assume that the application DTRT on
0673          * the first EIO, but we're not worse than other parts
0674          * of the kernel.
0675          */
0676         mapping_set_error(mapping, EIO);
0677     }
0678 
0679     return me_pagecache_clean(p, pfn);
0680 }
0681 
0682 /*
0683  * Clean and dirty swap cache.
0684  *
0685  * Dirty swap cache page is tricky to handle. The page could live both in page
0686  * cache and swap cache(ie. page is freshly swapped in). So it could be
0687  * referenced concurrently by 2 types of PTEs:
0688  * normal PTEs and swap PTEs. We try to handle them consistently by calling
0689  * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
0690  * and then
0691  *      - clear dirty bit to prevent IO
0692  *      - remove from LRU
0693  *      - but keep in the swap cache, so that when we return to it on
0694  *        a later page fault, we know the application is accessing
0695  *        corrupted data and shall be killed (we installed simple
0696  *        interception code in do_swap_page to catch it).
0697  *
0698  * Clean swap cache pages can be directly isolated. A later page fault will
0699  * bring in the known good data from disk.
0700  */
0701 static int me_swapcache_dirty(struct page *p, unsigned long pfn)
0702 {
0703     ClearPageDirty(p);
0704     /* Trigger EIO in shmem: */
0705     ClearPageUptodate(p);
0706 
0707     if (!delete_from_lru_cache(p))
0708         return MF_DELAYED;
0709     else
0710         return MF_FAILED;
0711 }
0712 
0713 static int me_swapcache_clean(struct page *p, unsigned long pfn)
0714 {
0715     delete_from_swap_cache(p);
0716 
0717     if (!delete_from_lru_cache(p))
0718         return MF_RECOVERED;
0719     else
0720         return MF_FAILED;
0721 }
0722 
0723 /*
0724  * Huge pages. Needs work.
0725  * Issues:
0726  * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
0727  *   To narrow down kill region to one page, we need to break up pmd.
0728  */
0729 static int me_huge_page(struct page *p, unsigned long pfn)
0730 {
0731     int res = 0;
0732     struct page *hpage = compound_head(p);
0733 
0734     if (!PageHuge(hpage))
0735         return MF_DELAYED;
0736 
0737     /*
0738      * We can safely recover from error on free or reserved (i.e.
0739      * not in-use) hugepage by dequeuing it from freelist.
0740      * To check whether a hugepage is in-use or not, we can't use
0741      * page->lru because it can be used in other hugepage operations,
0742      * such as __unmap_hugepage_range() and gather_surplus_pages().
0743      * So instead we use page_mapping() and PageAnon().
0744      */
0745     if (!(page_mapping(hpage) || PageAnon(hpage))) {
0746         res = dequeue_hwpoisoned_huge_page(hpage);
0747         if (!res)
0748             return MF_RECOVERED;
0749     }
0750     return MF_DELAYED;
0751 }
0752 
0753 /*
0754  * Various page states we can handle.
0755  *
0756  * A page state is defined by its current page->flags bits.
0757  * The table matches them in order and calls the right handler.
0758  *
0759  * This is quite tricky because we can access page at any time
0760  * in its live cycle, so all accesses have to be extremely careful.
0761  *
0762  * This is not complete. More states could be added.
0763  * For any missing state don't attempt recovery.
0764  */
0765 
0766 #define dirty       (1UL << PG_dirty)
0767 #define sc      ((1UL << PG_swapcache) | (1UL << PG_swapbacked))
0768 #define unevict     (1UL << PG_unevictable)
0769 #define mlock       (1UL << PG_mlocked)
0770 #define writeback   (1UL << PG_writeback)
0771 #define lru     (1UL << PG_lru)
0772 #define head        (1UL << PG_head)
0773 #define slab        (1UL << PG_slab)
0774 #define reserved    (1UL << PG_reserved)
0775 
0776 static struct page_state {
0777     unsigned long mask;
0778     unsigned long res;
0779     enum mf_action_page_type type;
0780     int (*action)(struct page *p, unsigned long pfn);
0781 } error_states[] = {
0782     { reserved, reserved,   MF_MSG_KERNEL,  me_kernel },
0783     /*
0784      * free pages are specially detected outside this table:
0785      * PG_buddy pages only make a small fraction of all free pages.
0786      */
0787 
0788     /*
0789      * Could in theory check if slab page is free or if we can drop
0790      * currently unused objects without touching them. But just
0791      * treat it as standard kernel for now.
0792      */
0793     { slab,     slab,       MF_MSG_SLAB,    me_kernel },
0794 
0795     { head,     head,       MF_MSG_HUGE,        me_huge_page },
0796 
0797     { sc|dirty, sc|dirty,   MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
0798     { sc|dirty, sc,     MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
0799 
0800     { mlock|dirty,  mlock|dirty,    MF_MSG_DIRTY_MLOCKED_LRU,   me_pagecache_dirty },
0801     { mlock|dirty,  mlock,      MF_MSG_CLEAN_MLOCKED_LRU,   me_pagecache_clean },
0802 
0803     { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU,   me_pagecache_dirty },
0804     { unevict|dirty, unevict,   MF_MSG_CLEAN_UNEVICTABLE_LRU,   me_pagecache_clean },
0805 
0806     { lru|dirty,    lru|dirty,  MF_MSG_DIRTY_LRU,   me_pagecache_dirty },
0807     { lru|dirty,    lru,        MF_MSG_CLEAN_LRU,   me_pagecache_clean },
0808 
0809     /*
0810      * Catchall entry: must be at end.
0811      */
0812     { 0,        0,      MF_MSG_UNKNOWN, me_unknown },
0813 };
0814 
0815 #undef dirty
0816 #undef sc
0817 #undef unevict
0818 #undef mlock
0819 #undef writeback
0820 #undef lru
0821 #undef head
0822 #undef slab
0823 #undef reserved
0824 
0825 /*
0826  * "Dirty/Clean" indication is not 100% accurate due to the possibility of
0827  * setting PG_dirty outside page lock. See also comment above set_page_dirty().
0828  */
0829 static void action_result(unsigned long pfn, enum mf_action_page_type type,
0830               enum mf_result result)
0831 {
0832     trace_memory_failure_event(pfn, type, result);
0833 
0834     pr_err("Memory failure: %#lx: recovery action for %s: %s\n",
0835         pfn, action_page_types[type], action_name[result]);
0836 }
0837 
0838 static int page_action(struct page_state *ps, struct page *p,
0839             unsigned long pfn)
0840 {
0841     int result;
0842     int count;
0843 
0844     result = ps->action(p, pfn);
0845 
0846     count = page_count(p) - 1;
0847     if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
0848         count--;
0849     if (count != 0) {
0850         pr_err("Memory failure: %#lx: %s still referenced by %d users\n",
0851                pfn, action_page_types[ps->type], count);
0852         result = MF_FAILED;
0853     }
0854     action_result(pfn, ps->type, result);
0855 
0856     /* Could do more checks here if page looks ok */
0857     /*
0858      * Could adjust zone counters here to correct for the missing page.
0859      */
0860 
0861     return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
0862 }
0863 
0864 /**
0865  * get_hwpoison_page() - Get refcount for memory error handling:
0866  * @page:   raw error page (hit by memory error)
0867  *
0868  * Return: return 0 if failed to grab the refcount, otherwise true (some
0869  * non-zero value.)
0870  */
0871 int get_hwpoison_page(struct page *page)
0872 {
0873     struct page *head = compound_head(page);
0874 
0875     if (!PageHuge(head) && PageTransHuge(head)) {
0876         /*
0877          * Non anonymous thp exists only in allocation/free time. We
0878          * can't handle such a case correctly, so let's give it up.
0879          * This should be better than triggering BUG_ON when kernel
0880          * tries to touch the "partially handled" page.
0881          */
0882         if (!PageAnon(head)) {
0883             pr_err("Memory failure: %#lx: non anonymous thp\n",
0884                 page_to_pfn(page));
0885             return 0;
0886         }
0887     }
0888 
0889     if (get_page_unless_zero(head)) {
0890         if (head == compound_head(page))
0891             return 1;
0892 
0893         pr_info("Memory failure: %#lx cannot catch tail\n",
0894             page_to_pfn(page));
0895         put_page(head);
0896     }
0897 
0898     return 0;
0899 }
0900 EXPORT_SYMBOL_GPL(get_hwpoison_page);
0901 
0902 /*
0903  * Do all that is necessary to remove user space mappings. Unmap
0904  * the pages and send SIGBUS to the processes if the data was dirty.
0905  */
0906 static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
0907                   int trapno, int flags, struct page **hpagep)
0908 {
0909     enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
0910     struct address_space *mapping;
0911     LIST_HEAD(tokill);
0912     int ret;
0913     int kill = 1, forcekill;
0914     struct page *hpage = *hpagep;
0915 
0916     /*
0917      * Here we are interested only in user-mapped pages, so skip any
0918      * other types of pages.
0919      */
0920     if (PageReserved(p) || PageSlab(p))
0921         return SWAP_SUCCESS;
0922     if (!(PageLRU(hpage) || PageHuge(p)))
0923         return SWAP_SUCCESS;
0924 
0925     /*
0926      * This check implies we don't kill processes if their pages
0927      * are in the swap cache early. Those are always late kills.
0928      */
0929     if (!page_mapped(hpage))
0930         return SWAP_SUCCESS;
0931 
0932     if (PageKsm(p)) {
0933         pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn);
0934         return SWAP_FAIL;
0935     }
0936 
0937     if (PageSwapCache(p)) {
0938         pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n",
0939             pfn);
0940         ttu |= TTU_IGNORE_HWPOISON;
0941     }
0942 
0943     /*
0944      * Propagate the dirty bit from PTEs to struct page first, because we
0945      * need this to decide if we should kill or just drop the page.
0946      * XXX: the dirty test could be racy: set_page_dirty() may not always
0947      * be called inside page lock (it's recommended but not enforced).
0948      */
0949     mapping = page_mapping(hpage);
0950     if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
0951         mapping_cap_writeback_dirty(mapping)) {
0952         if (page_mkclean(hpage)) {
0953             SetPageDirty(hpage);
0954         } else {
0955             kill = 0;
0956             ttu |= TTU_IGNORE_HWPOISON;
0957             pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n",
0958                 pfn);
0959         }
0960     }
0961 
0962     /*
0963      * First collect all the processes that have the page
0964      * mapped in dirty form.  This has to be done before try_to_unmap,
0965      * because ttu takes the rmap data structures down.
0966      *
0967      * Error handling: We ignore errors here because
0968      * there's nothing that can be done.
0969      */
0970     if (kill)
0971         collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
0972 
0973     ret = try_to_unmap(hpage, ttu);
0974     if (ret != SWAP_SUCCESS)
0975         pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n",
0976                pfn, page_mapcount(hpage));
0977 
0978     /*
0979      * Now that the dirty bit has been propagated to the
0980      * struct page and all unmaps done we can decide if
0981      * killing is needed or not.  Only kill when the page
0982      * was dirty or the process is not restartable,
0983      * otherwise the tokill list is merely
0984      * freed.  When there was a problem unmapping earlier
0985      * use a more force-full uncatchable kill to prevent
0986      * any accesses to the poisoned memory.
0987      */
0988     forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
0989     kill_procs(&tokill, forcekill, trapno,
0990               ret != SWAP_SUCCESS, p, pfn, flags);
0991 
0992     return ret;
0993 }
0994 
0995 static void set_page_hwpoison_huge_page(struct page *hpage)
0996 {
0997     int i;
0998     int nr_pages = 1 << compound_order(hpage);
0999     for (i = 0; i < nr_pages; i++)
1000         SetPageHWPoison(hpage + i);
1001 }
1002 
1003 static void clear_page_hwpoison_huge_page(struct page *hpage)
1004 {
1005     int i;
1006     int nr_pages = 1 << compound_order(hpage);
1007     for (i = 0; i < nr_pages; i++)
1008         ClearPageHWPoison(hpage + i);
1009 }
1010 
1011 /**
1012  * memory_failure - Handle memory failure of a page.
1013  * @pfn: Page Number of the corrupted page
1014  * @trapno: Trap number reported in the signal to user space.
1015  * @flags: fine tune action taken
1016  *
1017  * This function is called by the low level machine check code
1018  * of an architecture when it detects hardware memory corruption
1019  * of a page. It tries its best to recover, which includes
1020  * dropping pages, killing processes etc.
1021  *
1022  * The function is primarily of use for corruptions that
1023  * happen outside the current execution context (e.g. when
1024  * detected by a background scrubber)
1025  *
1026  * Must run in process context (e.g. a work queue) with interrupts
1027  * enabled and no spinlocks hold.
1028  */
1029 int memory_failure(unsigned long pfn, int trapno, int flags)
1030 {
1031     struct page_state *ps;
1032     struct page *p;
1033     struct page *hpage;
1034     struct page *orig_head;
1035     int res;
1036     unsigned int nr_pages;
1037     unsigned long page_flags;
1038 
1039     if (!sysctl_memory_failure_recovery)
1040         panic("Memory failure from trap %d on page %lx", trapno, pfn);
1041 
1042     if (!pfn_valid(pfn)) {
1043         pr_err("Memory failure: %#lx: memory outside kernel control\n",
1044             pfn);
1045         return -ENXIO;
1046     }
1047 
1048     p = pfn_to_page(pfn);
1049     orig_head = hpage = compound_head(p);
1050     if (TestSetPageHWPoison(p)) {
1051         pr_err("Memory failure: %#lx: already hardware poisoned\n",
1052             pfn);
1053         return 0;
1054     }
1055 
1056     /*
1057      * Currently errors on hugetlbfs pages are measured in hugepage units,
1058      * so nr_pages should be 1 << compound_order.  OTOH when errors are on
1059      * transparent hugepages, they are supposed to be split and error
1060      * measurement is done in normal page units.  So nr_pages should be one
1061      * in this case.
1062      */
1063     if (PageHuge(p))
1064         nr_pages = 1 << compound_order(hpage);
1065     else /* normal page or thp */
1066         nr_pages = 1;
1067     num_poisoned_pages_add(nr_pages);
1068 
1069     /*
1070      * We need/can do nothing about count=0 pages.
1071      * 1) it's a free page, and therefore in safe hand:
1072      *    prep_new_page() will be the gate keeper.
1073      * 2) it's a free hugepage, which is also safe:
1074      *    an affected hugepage will be dequeued from hugepage freelist,
1075      *    so there's no concern about reusing it ever after.
1076      * 3) it's part of a non-compound high order page.
1077      *    Implies some kernel user: cannot stop them from
1078      *    R/W the page; let's pray that the page has been
1079      *    used and will be freed some time later.
1080      * In fact it's dangerous to directly bump up page count from 0,
1081      * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1082      */
1083     if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
1084         if (is_free_buddy_page(p)) {
1085             action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1086             return 0;
1087         } else if (PageHuge(hpage)) {
1088             /*
1089              * Check "filter hit" and "race with other subpage."
1090              */
1091             lock_page(hpage);
1092             if (PageHWPoison(hpage)) {
1093                 if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1094                     || (p != hpage && TestSetPageHWPoison(hpage))) {
1095                     num_poisoned_pages_sub(nr_pages);
1096                     unlock_page(hpage);
1097                     return 0;
1098                 }
1099             }
1100             set_page_hwpoison_huge_page(hpage);
1101             res = dequeue_hwpoisoned_huge_page(hpage);
1102             action_result(pfn, MF_MSG_FREE_HUGE,
1103                       res ? MF_IGNORED : MF_DELAYED);
1104             unlock_page(hpage);
1105             return res;
1106         } else {
1107             action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
1108             return -EBUSY;
1109         }
1110     }
1111 
1112     if (!PageHuge(p) && PageTransHuge(hpage)) {
1113         lock_page(p);
1114         if (!PageAnon(p) || unlikely(split_huge_page(p))) {
1115             unlock_page(p);
1116             if (!PageAnon(p))
1117                 pr_err("Memory failure: %#lx: non anonymous thp\n",
1118                     pfn);
1119             else
1120                 pr_err("Memory failure: %#lx: thp split failed\n",
1121                     pfn);
1122             if (TestClearPageHWPoison(p))
1123                 num_poisoned_pages_sub(nr_pages);
1124             put_hwpoison_page(p);
1125             return -EBUSY;
1126         }
1127         unlock_page(p);
1128         VM_BUG_ON_PAGE(!page_count(p), p);
1129         hpage = compound_head(p);
1130     }
1131 
1132     /*
1133      * We ignore non-LRU pages for good reasons.
1134      * - PG_locked is only well defined for LRU pages and a few others
1135      * - to avoid races with __SetPageLocked()
1136      * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1137      * The check (unnecessarily) ignores LRU pages being isolated and
1138      * walked by the page reclaim code, however that's not a big loss.
1139      */
1140     if (!PageHuge(p)) {
1141         if (!PageLRU(p))
1142             shake_page(p, 0);
1143         if (!PageLRU(p)) {
1144             /*
1145              * shake_page could have turned it free.
1146              */
1147             if (is_free_buddy_page(p)) {
1148                 if (flags & MF_COUNT_INCREASED)
1149                     action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1150                 else
1151                     action_result(pfn, MF_MSG_BUDDY_2ND,
1152                               MF_DELAYED);
1153                 return 0;
1154             }
1155         }
1156     }
1157 
1158     lock_page(hpage);
1159 
1160     /*
1161      * The page could have changed compound pages during the locking.
1162      * If this happens just bail out.
1163      */
1164     if (PageCompound(p) && compound_head(p) != orig_head) {
1165         action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
1166         res = -EBUSY;
1167         goto out;
1168     }
1169 
1170     /*
1171      * We use page flags to determine what action should be taken, but
1172      * the flags can be modified by the error containment action.  One
1173      * example is an mlocked page, where PG_mlocked is cleared by
1174      * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1175      * correctly, we save a copy of the page flags at this time.
1176      */
1177     page_flags = p->flags;
1178 
1179     /*
1180      * unpoison always clear PG_hwpoison inside page lock
1181      */
1182     if (!PageHWPoison(p)) {
1183         pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1184         num_poisoned_pages_sub(nr_pages);
1185         unlock_page(hpage);
1186         put_hwpoison_page(hpage);
1187         return 0;
1188     }
1189     if (hwpoison_filter(p)) {
1190         if (TestClearPageHWPoison(p))
1191             num_poisoned_pages_sub(nr_pages);
1192         unlock_page(hpage);
1193         put_hwpoison_page(hpage);
1194         return 0;
1195     }
1196 
1197     if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p))
1198         goto identify_page_state;
1199 
1200     /*
1201      * For error on the tail page, we should set PG_hwpoison
1202      * on the head page to show that the hugepage is hwpoisoned
1203      */
1204     if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1205         action_result(pfn, MF_MSG_POISONED_HUGE, MF_IGNORED);
1206         unlock_page(hpage);
1207         put_hwpoison_page(hpage);
1208         return 0;
1209     }
1210     /*
1211      * Set PG_hwpoison on all pages in an error hugepage,
1212      * because containment is done in hugepage unit for now.
1213      * Since we have done TestSetPageHWPoison() for the head page with
1214      * page lock held, we can safely set PG_hwpoison bits on tail pages.
1215      */
1216     if (PageHuge(p))
1217         set_page_hwpoison_huge_page(hpage);
1218 
1219     /*
1220      * It's very difficult to mess with pages currently under IO
1221      * and in many cases impossible, so we just avoid it here.
1222      */
1223     wait_on_page_writeback(p);
1224 
1225     /*
1226      * Now take care of user space mappings.
1227      * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1228      *
1229      * When the raw error page is thp tail page, hpage points to the raw
1230      * page after thp split.
1231      */
1232     if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)
1233         != SWAP_SUCCESS) {
1234         action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1235         res = -EBUSY;
1236         goto out;
1237     }
1238 
1239     /*
1240      * Torn down by someone else?
1241      */
1242     if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1243         action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
1244         res = -EBUSY;
1245         goto out;
1246     }
1247 
1248 identify_page_state:
1249     res = -EBUSY;
1250     /*
1251      * The first check uses the current page flags which may not have any
1252      * relevant information. The second check with the saved page flagss is
1253      * carried out only if the first check can't determine the page status.
1254      */
1255     for (ps = error_states;; ps++)
1256         if ((p->flags & ps->mask) == ps->res)
1257             break;
1258 
1259     page_flags |= (p->flags & (1UL << PG_dirty));
1260 
1261     if (!ps->mask)
1262         for (ps = error_states;; ps++)
1263             if ((page_flags & ps->mask) == ps->res)
1264                 break;
1265     res = page_action(ps, p, pfn);
1266 out:
1267     unlock_page(hpage);
1268     return res;
1269 }
1270 EXPORT_SYMBOL_GPL(memory_failure);
1271 
1272 #define MEMORY_FAILURE_FIFO_ORDER   4
1273 #define MEMORY_FAILURE_FIFO_SIZE    (1 << MEMORY_FAILURE_FIFO_ORDER)
1274 
1275 struct memory_failure_entry {
1276     unsigned long pfn;
1277     int trapno;
1278     int flags;
1279 };
1280 
1281 struct memory_failure_cpu {
1282     DECLARE_KFIFO(fifo, struct memory_failure_entry,
1283               MEMORY_FAILURE_FIFO_SIZE);
1284     spinlock_t lock;
1285     struct work_struct work;
1286 };
1287 
1288 static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1289 
1290 /**
1291  * memory_failure_queue - Schedule handling memory failure of a page.
1292  * @pfn: Page Number of the corrupted page
1293  * @trapno: Trap number reported in the signal to user space.
1294  * @flags: Flags for memory failure handling
1295  *
1296  * This function is called by the low level hardware error handler
1297  * when it detects hardware memory corruption of a page. It schedules
1298  * the recovering of error page, including dropping pages, killing
1299  * processes etc.
1300  *
1301  * The function is primarily of use for corruptions that
1302  * happen outside the current execution context (e.g. when
1303  * detected by a background scrubber)
1304  *
1305  * Can run in IRQ context.
1306  */
1307 void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1308 {
1309     struct memory_failure_cpu *mf_cpu;
1310     unsigned long proc_flags;
1311     struct memory_failure_entry entry = {
1312         .pfn =      pfn,
1313         .trapno =   trapno,
1314         .flags =    flags,
1315     };
1316 
1317     mf_cpu = &get_cpu_var(memory_failure_cpu);
1318     spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1319     if (kfifo_put(&mf_cpu->fifo, entry))
1320         schedule_work_on(smp_processor_id(), &mf_cpu->work);
1321     else
1322         pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1323                pfn);
1324     spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1325     put_cpu_var(memory_failure_cpu);
1326 }
1327 EXPORT_SYMBOL_GPL(memory_failure_queue);
1328 
1329 static void memory_failure_work_func(struct work_struct *work)
1330 {
1331     struct memory_failure_cpu *mf_cpu;
1332     struct memory_failure_entry entry = { 0, };
1333     unsigned long proc_flags;
1334     int gotten;
1335 
1336     mf_cpu = this_cpu_ptr(&memory_failure_cpu);
1337     for (;;) {
1338         spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1339         gotten = kfifo_get(&mf_cpu->fifo, &entry);
1340         spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1341         if (!gotten)
1342             break;
1343         if (entry.flags & MF_SOFT_OFFLINE)
1344             soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1345         else
1346             memory_failure(entry.pfn, entry.trapno, entry.flags);
1347     }
1348 }
1349 
1350 static int __init memory_failure_init(void)
1351 {
1352     struct memory_failure_cpu *mf_cpu;
1353     int cpu;
1354 
1355     for_each_possible_cpu(cpu) {
1356         mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1357         spin_lock_init(&mf_cpu->lock);
1358         INIT_KFIFO(mf_cpu->fifo);
1359         INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1360     }
1361 
1362     return 0;
1363 }
1364 core_initcall(memory_failure_init);
1365 
1366 #define unpoison_pr_info(fmt, pfn, rs)          \
1367 ({                          \
1368     if (__ratelimit(rs))                \
1369         pr_info(fmt, pfn);          \
1370 })
1371 
1372 /**
1373  * unpoison_memory - Unpoison a previously poisoned page
1374  * @pfn: Page number of the to be unpoisoned page
1375  *
1376  * Software-unpoison a page that has been poisoned by
1377  * memory_failure() earlier.
1378  *
1379  * This is only done on the software-level, so it only works
1380  * for linux injected failures, not real hardware failures
1381  *
1382  * Returns 0 for success, otherwise -errno.
1383  */
1384 int unpoison_memory(unsigned long pfn)
1385 {
1386     struct page *page;
1387     struct page *p;
1388     int freeit = 0;
1389     unsigned int nr_pages;
1390     static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
1391                     DEFAULT_RATELIMIT_BURST);
1392 
1393     if (!pfn_valid(pfn))
1394         return -ENXIO;
1395 
1396     p = pfn_to_page(pfn);
1397     page = compound_head(p);
1398 
1399     if (!PageHWPoison(p)) {
1400         unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
1401                  pfn, &unpoison_rs);
1402         return 0;
1403     }
1404 
1405     if (page_count(page) > 1) {
1406         unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
1407                  pfn, &unpoison_rs);
1408         return 0;
1409     }
1410 
1411     if (page_mapped(page)) {
1412         unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
1413                  pfn, &unpoison_rs);
1414         return 0;
1415     }
1416 
1417     if (page_mapping(page)) {
1418         unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
1419                  pfn, &unpoison_rs);
1420         return 0;
1421     }
1422 
1423     /*
1424      * unpoison_memory() can encounter thp only when the thp is being
1425      * worked by memory_failure() and the page lock is not held yet.
1426      * In such case, we yield to memory_failure() and make unpoison fail.
1427      */
1428     if (!PageHuge(page) && PageTransHuge(page)) {
1429         unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n",
1430                  pfn, &unpoison_rs);
1431         return 0;
1432     }
1433 
1434     nr_pages = 1 << compound_order(page);
1435 
1436     if (!get_hwpoison_page(p)) {
1437         /*
1438          * Since HWPoisoned hugepage should have non-zero refcount,
1439          * race between memory failure and unpoison seems to happen.
1440          * In such case unpoison fails and memory failure runs
1441          * to the end.
1442          */
1443         if (PageHuge(page)) {
1444             unpoison_pr_info("Unpoison: Memory failure is now running on free hugepage %#lx\n",
1445                      pfn, &unpoison_rs);
1446             return 0;
1447         }
1448         if (TestClearPageHWPoison(p))
1449             num_poisoned_pages_dec();
1450         unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n",
1451                  pfn, &unpoison_rs);
1452         return 0;
1453     }
1454 
1455     lock_page(page);
1456     /*
1457      * This test is racy because PG_hwpoison is set outside of page lock.
1458      * That's acceptable because that won't trigger kernel panic. Instead,
1459      * the PG_hwpoison page will be caught and isolated on the entrance to
1460      * the free buddy page pool.
1461      */
1462     if (TestClearPageHWPoison(page)) {
1463         unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
1464                  pfn, &unpoison_rs);
1465         num_poisoned_pages_sub(nr_pages);
1466         freeit = 1;
1467         if (PageHuge(page))
1468             clear_page_hwpoison_huge_page(page);
1469     }
1470     unlock_page(page);
1471 
1472     put_hwpoison_page(page);
1473     if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
1474         put_hwpoison_page(page);
1475 
1476     return 0;
1477 }
1478 EXPORT_SYMBOL(unpoison_memory);
1479 
1480 static struct page *new_page(struct page *p, unsigned long private, int **x)
1481 {
1482     int nid = page_to_nid(p);
1483     if (PageHuge(p))
1484         return alloc_huge_page_node(page_hstate(compound_head(p)),
1485                            nid);
1486     else
1487         return __alloc_pages_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1488 }
1489 
1490 /*
1491  * Safely get reference count of an arbitrary page.
1492  * Returns 0 for a free page, -EIO for a zero refcount page
1493  * that is not free, and 1 for any other page type.
1494  * For 1 the page is returned with increased page count, otherwise not.
1495  */
1496 static int __get_any_page(struct page *p, unsigned long pfn, int flags)
1497 {
1498     int ret;
1499 
1500     if (flags & MF_COUNT_INCREASED)
1501         return 1;
1502 
1503     /*
1504      * When the target page is a free hugepage, just remove it
1505      * from free hugepage list.
1506      */
1507     if (!get_hwpoison_page(p)) {
1508         if (PageHuge(p)) {
1509             pr_info("%s: %#lx free huge page\n", __func__, pfn);
1510             ret = 0;
1511         } else if (is_free_buddy_page(p)) {
1512             pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1513             ret = 0;
1514         } else {
1515             pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1516                 __func__, pfn, p->flags);
1517             ret = -EIO;
1518         }
1519     } else {
1520         /* Not a free page */
1521         ret = 1;
1522     }
1523     return ret;
1524 }
1525 
1526 static int get_any_page(struct page *page, unsigned long pfn, int flags)
1527 {
1528     int ret = __get_any_page(page, pfn, flags);
1529 
1530     if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1531         /*
1532          * Try to free it.
1533          */
1534         put_hwpoison_page(page);
1535         shake_page(page, 1);
1536 
1537         /*
1538          * Did it turn free?
1539          */
1540         ret = __get_any_page(page, pfn, 0);
1541         if (ret == 1 && !PageLRU(page)) {
1542             /* Drop page reference which is from __get_any_page() */
1543             put_hwpoison_page(page);
1544             pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1545                 pfn, page->flags);
1546             return -EIO;
1547         }
1548     }
1549     return ret;
1550 }
1551 
1552 static int soft_offline_huge_page(struct page *page, int flags)
1553 {
1554     int ret;
1555     unsigned long pfn = page_to_pfn(page);
1556     struct page *hpage = compound_head(page);
1557     LIST_HEAD(pagelist);
1558 
1559     /*
1560      * This double-check of PageHWPoison is to avoid the race with
1561      * memory_failure(). See also comment in __soft_offline_page().
1562      */
1563     lock_page(hpage);
1564     if (PageHWPoison(hpage)) {
1565         unlock_page(hpage);
1566         put_hwpoison_page(hpage);
1567         pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1568         return -EBUSY;
1569     }
1570     unlock_page(hpage);
1571 
1572     ret = isolate_huge_page(hpage, &pagelist);
1573     /*
1574      * get_any_page() and isolate_huge_page() takes a refcount each,
1575      * so need to drop one here.
1576      */
1577     put_hwpoison_page(hpage);
1578     if (!ret) {
1579         pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn);
1580         return -EBUSY;
1581     }
1582 
1583     ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1584                 MIGRATE_SYNC, MR_MEMORY_FAILURE);
1585     if (ret) {
1586         pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1587             pfn, ret, page->flags);
1588         /*
1589          * We know that soft_offline_huge_page() tries to migrate
1590          * only one hugepage pointed to by hpage, so we need not
1591          * run through the pagelist here.
1592          */
1593         putback_active_hugepage(hpage);
1594         if (ret > 0)
1595             ret = -EIO;
1596     } else {
1597         /* overcommit hugetlb page will be freed to buddy */
1598         if (PageHuge(page)) {
1599             set_page_hwpoison_huge_page(hpage);
1600             dequeue_hwpoisoned_huge_page(hpage);
1601             num_poisoned_pages_add(1 << compound_order(hpage));
1602         } else {
1603             SetPageHWPoison(page);
1604             num_poisoned_pages_inc();
1605         }
1606     }
1607     return ret;
1608 }
1609 
1610 static int __soft_offline_page(struct page *page, int flags)
1611 {
1612     int ret;
1613     unsigned long pfn = page_to_pfn(page);
1614 
1615     /*
1616      * Check PageHWPoison again inside page lock because PageHWPoison
1617      * is set by memory_failure() outside page lock. Note that
1618      * memory_failure() also double-checks PageHWPoison inside page lock,
1619      * so there's no race between soft_offline_page() and memory_failure().
1620      */
1621     lock_page(page);
1622     wait_on_page_writeback(page);
1623     if (PageHWPoison(page)) {
1624         unlock_page(page);
1625         put_hwpoison_page(page);
1626         pr_info("soft offline: %#lx page already poisoned\n", pfn);
1627         return -EBUSY;
1628     }
1629     /*
1630      * Try to invalidate first. This should work for
1631      * non dirty unmapped page cache pages.
1632      */
1633     ret = invalidate_inode_page(page);
1634     unlock_page(page);
1635     /*
1636      * RED-PEN would be better to keep it isolated here, but we
1637      * would need to fix isolation locking first.
1638      */
1639     if (ret == 1) {
1640         put_hwpoison_page(page);
1641         pr_info("soft_offline: %#lx: invalidated\n", pfn);
1642         SetPageHWPoison(page);
1643         num_poisoned_pages_inc();
1644         return 0;
1645     }
1646 
1647     /*
1648      * Simple invalidation didn't work.
1649      * Try to migrate to a new page instead. migrate.c
1650      * handles a large number of cases for us.
1651      */
1652     ret = isolate_lru_page(page);
1653     /*
1654      * Drop page reference which is came from get_any_page()
1655      * successful isolate_lru_page() already took another one.
1656      */
1657     put_hwpoison_page(page);
1658     if (!ret) {
1659         LIST_HEAD(pagelist);
1660         inc_node_page_state(page, NR_ISOLATED_ANON +
1661                     page_is_file_cache(page));
1662         list_add(&page->lru, &pagelist);
1663         ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1664                     MIGRATE_SYNC, MR_MEMORY_FAILURE);
1665         if (ret) {
1666             if (!list_empty(&pagelist)) {
1667                 list_del(&page->lru);
1668                 dec_node_page_state(page, NR_ISOLATED_ANON +
1669                         page_is_file_cache(page));
1670                 putback_lru_page(page);
1671             }
1672 
1673             pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1674                 pfn, ret, page->flags);
1675             if (ret > 0)
1676                 ret = -EIO;
1677         }
1678     } else {
1679         pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1680             pfn, ret, page_count(page), page->flags);
1681     }
1682     return ret;
1683 }
1684 
1685 static int soft_offline_in_use_page(struct page *page, int flags)
1686 {
1687     int ret;
1688     struct page *hpage = compound_head(page);
1689 
1690     if (!PageHuge(page) && PageTransHuge(hpage)) {
1691         lock_page(hpage);
1692         if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) {
1693             unlock_page(hpage);
1694             if (!PageAnon(hpage))
1695                 pr_info("soft offline: %#lx: non anonymous thp\n", page_to_pfn(page));
1696             else
1697                 pr_info("soft offline: %#lx: thp split failed\n", page_to_pfn(page));
1698             put_hwpoison_page(hpage);
1699             return -EBUSY;
1700         }
1701         unlock_page(hpage);
1702         get_hwpoison_page(page);
1703         put_hwpoison_page(hpage);
1704     }
1705 
1706     if (PageHuge(page))
1707         ret = soft_offline_huge_page(page, flags);
1708     else
1709         ret = __soft_offline_page(page, flags);
1710 
1711     return ret;
1712 }
1713 
1714 static void soft_offline_free_page(struct page *page)
1715 {
1716     if (PageHuge(page)) {
1717         struct page *hpage = compound_head(page);
1718 
1719         set_page_hwpoison_huge_page(hpage);
1720         if (!dequeue_hwpoisoned_huge_page(hpage))
1721             num_poisoned_pages_add(1 << compound_order(hpage));
1722     } else {
1723         if (!TestSetPageHWPoison(page))
1724             num_poisoned_pages_inc();
1725     }
1726 }
1727 
1728 /**
1729  * soft_offline_page - Soft offline a page.
1730  * @page: page to offline
1731  * @flags: flags. Same as memory_failure().
1732  *
1733  * Returns 0 on success, otherwise negated errno.
1734  *
1735  * Soft offline a page, by migration or invalidation,
1736  * without killing anything. This is for the case when
1737  * a page is not corrupted yet (so it's still valid to access),
1738  * but has had a number of corrected errors and is better taken
1739  * out.
1740  *
1741  * The actual policy on when to do that is maintained by
1742  * user space.
1743  *
1744  * This should never impact any application or cause data loss,
1745  * however it might take some time.
1746  *
1747  * This is not a 100% solution for all memory, but tries to be
1748  * ``good enough'' for the majority of memory.
1749  */
1750 int soft_offline_page(struct page *page, int flags)
1751 {
1752     int ret;
1753     unsigned long pfn = page_to_pfn(page);
1754 
1755     if (PageHWPoison(page)) {
1756         pr_info("soft offline: %#lx page already poisoned\n", pfn);
1757         if (flags & MF_COUNT_INCREASED)
1758             put_hwpoison_page(page);
1759         return -EBUSY;
1760     }
1761 
1762     get_online_mems();
1763     ret = get_any_page(page, pfn, flags);
1764     put_online_mems();
1765 
1766     if (ret > 0)
1767         ret = soft_offline_in_use_page(page, flags);
1768     else if (ret == 0)
1769         soft_offline_free_page(page);
1770 
1771     return ret;
1772 }