mm: fix wrong vmap address calculations with odd NR_CPUS values
[linux-2.6.git] / mm / memory-failure.c
index 745f610..2b43ba0 100644 (file)
@@ -7,21 +7,26 @@
  * Free Software Foundation.
  *
  * High level machine check handler. Handles pages reported by the
- * hardware as being corrupted usually due to a 2bit ECC memory or cache
+ * hardware as being corrupted usually due to a multi-bit ECC memory or cache
  * failure.
+ * 
+ * In addition there is a "soft offline" entry point that allows stop using
+ * not-yet-corrupted-by-suspicious pages without killing anything.
  *
  * Handles page cache pages in various states. The tricky part
- * here is that we can access any page asynchronous to other VM
- * users, because memory failures could happen anytime and anywhere,
- * possibly violating some of their assumptions. This is why this code
- * has to be extremely careful. Generally it tries to use normal locking
- * rules, as in get the standard locks, even if that means the
- * error handling takes potentially a long time.
- *
- * The operation to map back from RMAP chains to processes has to walk
- * the complete process list and has non linear complexity with the number
- * mappings. In short it can be quite slow. But since memory corruptions
- * are rare we hope to get away with this.
+ * here is that we can access any page asynchronously in respect to 
+ * other VM users, because memory failures could happen anytime and 
+ * anywhere. This could violate some of their assumptions. This is why 
+ * this code has to be extremely careful. Generally it tries to use 
+ * normal locking rules, as in get the standard locks, even if that means 
+ * the error handling takes potentially a long time.
+ * 
+ * There are several operations here with exponential complexity because
+ * of unsuitable VM data structures. For example the operation to map back 
+ * from RMAP chains to processes has to walk the complete process list and 
+ * has non linear complexity with the number. But since memory corruptions
+ * are rare we hope to get away with this. This avoids impacting the core 
+ * VM.
  */
 
 /*
  * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
  * - pass bad pages to kdump next kernel
  */
-#define DEBUG 1                /* remove me in 2.6.34 */
 #include <linux/kernel.h>
 #include <linux/mm.h>
 #include <linux/page-flags.h>
+#include <linux/kernel-page-flags.h>
 #include <linux/sched.h>
 #include <linux/ksm.h>
 #include <linux/rmap.h>
 #include <linux/pagemap.h>
 #include <linux/swap.h>
 #include <linux/backing-dev.h>
+#include <linux/migrate.h>
+#include <linux/page-isolation.h>
+#include <linux/suspend.h>
+#include <linux/slab.h>
+#include <linux/swapops.h>
+#include <linux/hugetlb.h>
+#include <linux/memory_hotplug.h>
+#include <linux/mm_inline.h>
+#include <linux/kfifo.h>
 #include "internal.h"
 
 int sysctl_memory_failure_early_kill __read_mostly = 0;
@@ -48,12 +62,135 @@ int sysctl_memory_failure_recovery __read_mostly = 1;
 
 atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
 
+#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
+
+u32 hwpoison_filter_enable = 0;
+u32 hwpoison_filter_dev_major = ~0U;
+u32 hwpoison_filter_dev_minor = ~0U;
+u64 hwpoison_filter_flags_mask;
+u64 hwpoison_filter_flags_value;
+EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
+EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
+EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
+EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
+EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
+
+static int hwpoison_filter_dev(struct page *p)
+{
+       struct address_space *mapping;
+       dev_t dev;
+
+       if (hwpoison_filter_dev_major == ~0U &&
+           hwpoison_filter_dev_minor == ~0U)
+               return 0;
+
+       /*
+        * page_mapping() does not accept slab pages.
+        */
+       if (PageSlab(p))
+               return -EINVAL;
+
+       mapping = page_mapping(p);
+       if (mapping == NULL || mapping->host == NULL)
+               return -EINVAL;
+
+       dev = mapping->host->i_sb->s_dev;
+       if (hwpoison_filter_dev_major != ~0U &&
+           hwpoison_filter_dev_major != MAJOR(dev))
+               return -EINVAL;
+       if (hwpoison_filter_dev_minor != ~0U &&
+           hwpoison_filter_dev_minor != MINOR(dev))
+               return -EINVAL;
+
+       return 0;
+}
+
+static int hwpoison_filter_flags(struct page *p)
+{
+       if (!hwpoison_filter_flags_mask)
+               return 0;
+
+       if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
+                                   hwpoison_filter_flags_value)
+               return 0;
+       else
+               return -EINVAL;
+}
+
+/*
+ * This allows stress tests to limit test scope to a collection of tasks
+ * by putting them under some memcg. This prevents killing unrelated/important
+ * processes such as /sbin/init. Note that the target task may share clean
+ * pages with init (eg. libc text), which is harmless. If the target task
+ * share _dirty_ pages with another task B, the test scheme must make sure B
+ * is also included in the memcg. At last, due to race conditions this filter
+ * can only guarantee that the page either belongs to the memcg tasks, or is
+ * a freed page.
+ */
+#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
+u64 hwpoison_filter_memcg;
+EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
+static int hwpoison_filter_task(struct page *p)
+{
+       struct mem_cgroup *mem;
+       struct cgroup_subsys_state *css;
+       unsigned long ino;
+
+       if (!hwpoison_filter_memcg)
+               return 0;
+
+       mem = try_get_mem_cgroup_from_page(p);
+       if (!mem)
+               return -EINVAL;
+
+       css = mem_cgroup_css(mem);
+       /* root_mem_cgroup has NULL dentries */
+       if (!css->cgroup->dentry)
+               return -EINVAL;
+
+       ino = css->cgroup->dentry->d_inode->i_ino;
+       css_put(css);
+
+       if (ino != hwpoison_filter_memcg)
+               return -EINVAL;
+
+       return 0;
+}
+#else
+static int hwpoison_filter_task(struct page *p) { return 0; }
+#endif
+
+int hwpoison_filter(struct page *p)
+{
+       if (!hwpoison_filter_enable)
+               return 0;
+
+       if (hwpoison_filter_dev(p))
+               return -EINVAL;
+
+       if (hwpoison_filter_flags(p))
+               return -EINVAL;
+
+       if (hwpoison_filter_task(p))
+               return -EINVAL;
+
+       return 0;
+}
+#else
+int hwpoison_filter(struct page *p)
+{
+       return 0;
+}
+#endif
+
+EXPORT_SYMBOL_GPL(hwpoison_filter);
+
 /*
  * Send all the processes who have the page mapped an ``action optional''
  * signal.
  */
 static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
-                       unsigned long pfn)
+                       unsigned long pfn, struct page *page)
 {
        struct siginfo si;
        int ret;
@@ -68,12 +205,12 @@ static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
 #ifdef __ARCH_SI_TRAPNO
        si.si_trapno = trapno;
 #endif
-       si.si_addr_lsb = PAGE_SHIFT;
+       si.si_addr_lsb = compound_trans_order(compound_head(page)) + PAGE_SHIFT;
        /*
         * Don't use force here, it's convenient if the signal
         * can be temporarily blocked.
         * This could cause a loop when the user sets SIGBUS
-        * to SIG_IGN, but hopefully noone will do that?
+        * to SIG_IGN, but hopefully no one will do that?
         */
        ret = send_sig_info(SIGBUS, &si, t);  /* synchronous? */
        if (ret < 0)
@@ -86,7 +223,7 @@ static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
  * When a unknown page type is encountered drain as many buffers as possible
  * in the hope to turn the page into a LRU or free page, which we can handle.
  */
-void shake_page(struct page *p)
+void shake_page(struct page *p, int access)
 {
        if (!PageSlab(p)) {
                lru_add_drain_all();
@@ -96,11 +233,23 @@ void shake_page(struct page *p)
                if (PageLRU(p) || is_free_buddy_page(p))
                        return;
        }
+
        /*
-        * Could call shrink_slab here (which would also
-        * shrink other caches). Unfortunately that might
-        * also access the corrupted page, which could be fatal.
+        * Only call shrink_slab here (which would also shrink other caches) if
+        * access is not potentially fatal.
         */
+       if (access) {
+               int nr;
+               do {
+                       struct shrink_control shrink = {
+                               .gfp_mask = GFP_KERNEL,
+                       };
+
+                       nr = shrink_slab(&shrink, 1000, 1000);
+                       if (page_count(p) == 1)
+                               break;
+               } while (nr > 10);
+       }
 }
 EXPORT_SYMBOL_GPL(shake_page);
 
@@ -130,7 +279,7 @@ struct to_kill {
        struct list_head nd;
        struct task_struct *tsk;
        unsigned long addr;
-       unsigned addr_valid:1;
+       char addr_valid;
 };
 
 /*
@@ -171,7 +320,7 @@ static void add_to_kill(struct task_struct *tsk, struct page *p,
         * a SIGKILL because the error is not contained anymore.
         */
        if (tk->addr == -EFAULT) {
-               pr_debug("MCE: Unable to find user space address %lx in %s\n",
+               pr_info("MCE: Unable to find user space address %lx in %s\n",
                        page_to_pfn(p), tsk->comm);
                tk->addr_valid = 0;
        }
@@ -189,7 +338,7 @@ static void add_to_kill(struct task_struct *tsk, struct page *p,
  * wrong earlier.
  */
 static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
-                         int fail, unsigned long pfn)
+                         int fail, struct page *page, unsigned long pfn)
 {
        struct to_kill *tk, *next;
 
@@ -199,7 +348,6 @@ static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
                         * In case something went wrong with munmapping
                         * make sure the process doesn't catch the
                         * signal and then access the memory. Just kill it.
-                        * the signal handlers
                         */
                        if (fail || tk->addr_valid == 0) {
                                printk(KERN_ERR
@@ -215,7 +363,7 @@ static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
                         * process anyways.
                         */
                        else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
-                                             pfn) < 0)
+                                             pfn, page) < 0)
                                printk(KERN_ERR
                "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
                                        pfn, tk->tsk->comm, tk->tsk->pid);
@@ -244,23 +392,26 @@ static void collect_procs_anon(struct page *page, struct list_head *to_kill,
        struct task_struct *tsk;
        struct anon_vma *av;
 
-       read_lock(&tasklist_lock);
        av = page_lock_anon_vma(page);
        if (av == NULL) /* Not actually mapped anymore */
-               goto out;
+               return;
+
+       read_lock(&tasklist_lock);
        for_each_process (tsk) {
+               struct anon_vma_chain *vmac;
+
                if (!task_early_kill(tsk))
                        continue;
-               list_for_each_entry (vma, &av->head, anon_vma_node) {
+               list_for_each_entry(vmac, &av->head, same_anon_vma) {
+                       vma = vmac->vma;
                        if (!page_mapped_in_vma(page, vma))
                                continue;
                        if (vma->vm_mm == tsk->mm)
                                add_to_kill(tsk, page, vma, to_kill, tkc);
                }
        }
-       page_unlock_anon_vma(av);
-out:
        read_unlock(&tasklist_lock);
+       page_unlock_anon_vma(av);
 }
 
 /*
@@ -274,17 +425,8 @@ static void collect_procs_file(struct page *page, struct list_head *to_kill,
        struct prio_tree_iter iter;
        struct address_space *mapping = page->mapping;
 
-       /*
-        * A note on the locking order between the two locks.
-        * We don't rely on this particular order.
-        * If you have some other code that needs a different order
-        * feel free to switch them around. Or add a reverse link
-        * from mm_struct to task_struct, then this could be all
-        * done without taking tasklist_lock and looping over all tasks.
-        */
-
+       mutex_lock(&mapping->i_mmap_mutex);
        read_lock(&tasklist_lock);
-       spin_lock(&mapping->i_mmap_lock);
        for_each_process(tsk) {
                pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 
@@ -304,8 +446,8 @@ static void collect_procs_file(struct page *page, struct list_head *to_kill,
                                add_to_kill(tsk, page, vma, to_kill, tkc);
                }
        }
-       spin_unlock(&mapping->i_mmap_lock);
        read_unlock(&tasklist_lock);
+       mutex_unlock(&mapping->i_mmap_mutex);
 }
 
 /*
@@ -336,33 +478,49 @@ static void collect_procs(struct page *page, struct list_head *tokill)
  */
 
 enum outcome {
-       FAILED,         /* Error handling failed */
+       IGNORED,        /* Error: cannot be handled */
+       FAILED,         /* Error: handling failed */
        DELAYED,        /* Will be handled later */
-       IGNORED,        /* Error safely ignored */
        RECOVERED,      /* Successfully recovered */
 };
 
 static const char *action_name[] = {
+       [IGNORED] = "Ignored",
        [FAILED] = "Failed",
        [DELAYED] = "Delayed",
-       [IGNORED] = "Ignored",
        [RECOVERED] = "Recovered",
 };
 
 /*
- * Error hit kernel page.
- * Do nothing, try to be lucky and not touch this instead. For a few cases we
- * could be more sophisticated.
+ * XXX: It is possible that a page is isolated from LRU cache,
+ * and then kept in swap cache or failed to remove from page cache.
+ * The page count will stop it from being freed by unpoison.
+ * Stress tests should be aware of this memory leak problem.
  */
-static int me_kernel(struct page *p, unsigned long pfn)
+static int delete_from_lru_cache(struct page *p)
 {
-       return DELAYED;
+       if (!isolate_lru_page(p)) {
+               /*
+                * Clear sensible page flags, so that the buddy system won't
+                * complain when the page is unpoison-and-freed.
+                */
+               ClearPageActive(p);
+               ClearPageUnevictable(p);
+               /*
+                * drop the page count elevated by isolate_lru_page()
+                */
+               page_cache_release(p);
+               return 0;
+       }
+       return -EIO;
 }
 
 /*
- * Already poisoned page.
+ * Error hit kernel page.
+ * Do nothing, try to be lucky and not touch this instead. For a few cases we
+ * could be more sophisticated.
  */
-static int me_ignore(struct page *p, unsigned long pfn)
+static int me_kernel(struct page *p, unsigned long pfn)
 {
        return IGNORED;
 }
@@ -377,14 +535,6 @@ static int me_unknown(struct page *p, unsigned long pfn)
 }
 
 /*
- * Free memory
- */
-static int me_free(struct page *p, unsigned long pfn)
-{
-       return DELAYED;
-}
-
-/*
  * Clean (or cleaned) page cache page.
  */
 static int me_pagecache_clean(struct page *p, unsigned long pfn)
@@ -393,6 +543,8 @@ static int me_pagecache_clean(struct page *p, unsigned long pfn)
        int ret = FAILED;
        struct address_space *mapping;
 
+       delete_from_lru_cache(p);
+
        /*
         * For anonymous pages we're done the only reference left
         * should be the one m_f() holds.
@@ -427,7 +579,7 @@ static int me_pagecache_clean(struct page *p, unsigned long pfn)
                                        pfn, err);
                } else if (page_has_private(p) &&
                                !try_to_release_page(p, GFP_NOIO)) {
-                       pr_debug("MCE %#lx: failed to release buffers\n", pfn);
+                       pr_info("MCE %#lx: failed to release buffers\n", pfn);
                } else {
                        ret = RECOVERED;
                }
@@ -479,7 +631,7 @@ static int me_pagecache_dirty(struct page *p, unsigned long pfn)
                 * when the page is reread or dropped.  If an
                 * application assumes it will always get error on
                 * fsync, but does other operations on the fd before
-                * and the page is dropped inbetween then the error
+                * and the page is dropped between then the error
                 * will not be properly reported.
                 *
                 * This can already happen even without hwpoisoned
@@ -522,30 +674,48 @@ static int me_swapcache_dirty(struct page *p, unsigned long pfn)
        /* Trigger EIO in shmem: */
        ClearPageUptodate(p);
 
-       return DELAYED;
+       if (!delete_from_lru_cache(p))
+               return DELAYED;
+       else
+               return FAILED;
 }
 
 static int me_swapcache_clean(struct page *p, unsigned long pfn)
 {
        delete_from_swap_cache(p);
 
-       return RECOVERED;
+       if (!delete_from_lru_cache(p))
+               return RECOVERED;
+       else
+               return FAILED;
 }
 
 /*
  * Huge pages. Needs work.
  * Issues:
- * No rmap support so we cannot find the original mapper. In theory could walk
- * all MMs and look for the mappings, but that would be non atomic and racy.
- * Need rmap for hugepages for this. Alternatively we could employ a heuristic,
- * like just walking the current process and hoping it has it mapped (that
- * should be usually true for the common "shared database cache" case)
- * Should handle free huge pages and dequeue them too, but this needs to
- * handle huge page accounting correctly.
+ * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
+ *   To narrow down kill region to one page, we need to break up pmd.
  */
 static int me_huge_page(struct page *p, unsigned long pfn)
 {
-       return FAILED;
+       int res = 0;
+       struct page *hpage = compound_head(p);
+       /*
+        * We can safely recover from error on free or reserved (i.e.
+        * not in-use) hugepage by dequeuing it from freelist.
+        * To check whether a hugepage is in-use or not, we can't use
+        * page->lru because it can be used in other hugepage operations,
+        * such as __unmap_hugepage_range() and gather_surplus_pages().
+        * So instead we use page_mapping() and PageAnon().
+        * We assume that this function is called with page lock held,
+        * so there is no race between isolation and mapping/unmapping.
+        */
+       if (!(page_mapping(hpage) || PageAnon(hpage))) {
+               res = dequeue_hwpoisoned_huge_page(hpage);
+               if (!res)
+                       return RECOVERED;
+       }
+       return DELAYED;
 }
 
 /*
@@ -555,7 +725,7 @@ static int me_huge_page(struct page *p, unsigned long pfn)
  * The table matches them in order and calls the right handler.
  *
  * This is quite tricky because we can access page at any time
- * in its live cycle, so all accesses have to be extremly careful.
+ * in its live cycle, so all accesses have to be extremely careful.
  *
  * This is not complete. More states could be added.
  * For any missing state don't attempt recovery.
@@ -572,7 +742,6 @@ static int me_huge_page(struct page *p, unsigned long pfn)
 #define tail           (1UL << PG_tail)
 #define compound       (1UL << PG_compound)
 #define slab           (1UL << PG_slab)
-#define buddy          (1UL << PG_buddy)
 #define reserved       (1UL << PG_reserved)
 
 static struct page_state {
@@ -581,8 +750,11 @@ static struct page_state {
        char *msg;
        int (*action)(struct page *p, unsigned long pfn);
 } error_states[] = {
-       { reserved,     reserved,       "reserved kernel",      me_ignore },
-       { buddy,        buddy,          "free kernel",  me_free },
+       { reserved,     reserved,       "reserved kernel",      me_kernel },
+       /*
+        * free pages are specially detected outside this table:
+        * PG_buddy pages only make a small fraction of all free pages.
+        */
 
        /*
         * Could in theory check if slab page is free or if we can drop
@@ -616,20 +788,31 @@ static struct page_state {
        { 0,            0,              "unknown page state",   me_unknown },
 };
 
+#undef dirty
+#undef sc
+#undef unevict
+#undef mlock
+#undef writeback
+#undef lru
+#undef swapbacked
+#undef head
+#undef tail
+#undef compound
+#undef slab
+#undef reserved
+
 static void action_result(unsigned long pfn, char *msg, int result)
 {
-       struct page *page = NULL;
-       if (pfn_valid(pfn))
-               page = pfn_to_page(pfn);
+       struct page *page = pfn_to_page(pfn);
 
        printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
                pfn,
-               page && PageDirty(page) ? "dirty " : "",
+               PageDirty(page) ? "dirty " : "",
                msg, action_name[result]);
 }
 
 static int page_action(struct page_state *ps, struct page *p,
-                       unsigned long pfn, int ref)
+                       unsigned long pfn)
 {
        int result;
        int count;
@@ -637,45 +820,51 @@ static int page_action(struct page_state *ps, struct page *p,
        result = ps->action(p, pfn);
        action_result(pfn, ps->msg, result);
 
-       count = page_count(p) - 1 - ref;
-       if (count != 0)
+       count = page_count(p) - 1;
+       if (ps->action == me_swapcache_dirty && result == DELAYED)
+               count--;
+       if (count != 0) {
                printk(KERN_ERR
                       "MCE %#lx: %s page still referenced by %d users\n",
                       pfn, ps->msg, count);
+               result = FAILED;
+       }
 
        /* Could do more checks here if page looks ok */
        /*
         * Could adjust zone counters here to correct for the missing page.
         */
 
-       return result == RECOVERED ? 0 : -EBUSY;
+       return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
 }
 
-#define N_UNMAP_TRIES 5
-
 /*
  * Do all that is necessary to remove user space mappings. Unmap
  * the pages and send SIGBUS to the processes if the data was dirty.
  */
-static void hwpoison_user_mappings(struct page *p, unsigned long pfn,
+static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
                                  int trapno)
 {
        enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
        struct address_space *mapping;
        LIST_HEAD(tokill);
        int ret;
-       int i;
        int kill = 1;
+       struct page *hpage = compound_head(p);
+       struct page *ppage;
 
-       if (PageReserved(p) || PageCompound(p) || PageSlab(p) || PageKsm(p))
-               return;
+       if (PageReserved(p) || PageSlab(p))
+               return SWAP_SUCCESS;
 
        /*
         * This check implies we don't kill processes if their pages
         * are in the swap cache early. Those are always late kills.
         */
-       if (!page_mapped(p))
-               return;
+       if (!page_mapped(hpage))
+               return SWAP_SUCCESS;
+
+       if (PageKsm(p))
+               return SWAP_FAIL;
 
        if (PageSwapCache(p)) {
                printk(KERN_ERR
@@ -686,11 +875,14 @@ static void hwpoison_user_mappings(struct page *p, unsigned long pfn,
        /*
         * Propagate the dirty bit from PTEs to struct page first, because we
         * need this to decide if we should kill or just drop the page.
+        * XXX: the dirty test could be racy: set_page_dirty() may not always
+        * be called inside page lock (it's recommended but not enforced).
         */
-       mapping = page_mapping(p);
-       if (!PageDirty(p) && mapping && mapping_cap_writeback_dirty(mapping)) {
-               if (page_mkclean(p)) {
-                       SetPageDirty(p);
+       mapping = page_mapping(hpage);
+       if (!PageDirty(hpage) && mapping &&
+           mapping_cap_writeback_dirty(mapping)) {
+               if (page_mkclean(hpage)) {
+                       SetPageDirty(hpage);
                } else {
                        kill = 0;
                        ttu |= TTU_IGNORE_HWPOISON;
@@ -701,6 +893,44 @@ static void hwpoison_user_mappings(struct page *p, unsigned long pfn,
        }
 
        /*
+        * ppage: poisoned page
+        *   if p is regular page(4k page)
+        *        ppage == real poisoned page;
+        *   else p is hugetlb or THP, ppage == head page.
+        */
+       ppage = hpage;
+
+       if (PageTransHuge(hpage)) {
+               /*
+                * Verify that this isn't a hugetlbfs head page, the check for
+                * PageAnon is just for avoid tripping a split_huge_page
+                * internal debug check, as split_huge_page refuses to deal with
+                * anything that isn't an anon page. PageAnon can't go away fro
+                * under us because we hold a refcount on the hpage, without a
+                * refcount on the hpage. split_huge_page can't be safely called
+                * in the first place, having a refcount on the tail isn't
+                * enough * to be safe.
+                */
+               if (!PageHuge(hpage) && PageAnon(hpage)) {
+                       if (unlikely(split_huge_page(hpage))) {
+                               /*
+                                * FIXME: if splitting THP is failed, it is
+                                * better to stop the following operation rather
+                                * than causing panic by unmapping. System might
+                                * survive if the page is freed later.
+                                */
+                               printk(KERN_INFO
+                                       "MCE %#lx: failed to split THP\n", pfn);
+
+                               BUG_ON(!PageHWPoison(p));
+                               return SWAP_FAIL;
+                       }
+                       /* THP is split, so ppage should be the real poisoned page. */
+                       ppage = p;
+               }
+       }
+
+       /*
         * First collect all the processes that have the page
         * mapped in dirty form.  This has to be done before try_to_unmap,
         * because ttu takes the rmap data structures down.
@@ -709,22 +939,18 @@ static void hwpoison_user_mappings(struct page *p, unsigned long pfn,
         * there's nothing that can be done.
         */
        if (kill)
-               collect_procs(p, &tokill);
+               collect_procs(ppage, &tokill);
 
-       /*
-        * try_to_unmap can fail temporarily due to races.
-        * Try a few times (RED-PEN better strategy?)
-        */
-       for (i = 0; i < N_UNMAP_TRIES; i++) {
-               ret = try_to_unmap(p, ttu);
-               if (ret == SWAP_SUCCESS)
-                       break;
-               pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn,  ret);
-       }
+       if (hpage != ppage)
+               lock_page(ppage);
 
+       ret = try_to_unmap(ppage, ttu);
        if (ret != SWAP_SUCCESS)
                printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
-                               pfn, page_mapcount(p));
+                               pfn, page_mapcount(ppage));
+
+       if (hpage != ppage)
+               unlock_page(ppage);
 
        /*
         * Now that the dirty bit has been propagated to the
@@ -735,47 +961,97 @@ static void hwpoison_user_mappings(struct page *p, unsigned long pfn,
         * use a more force-full uncatchable kill to prevent
         * any accesses to the poisoned memory.
         */
-       kill_procs_ao(&tokill, !!PageDirty(p), trapno,
-                     ret != SWAP_SUCCESS, pfn);
+       kill_procs_ao(&tokill, !!PageDirty(ppage), trapno,
+                     ret != SWAP_SUCCESS, p, pfn);
+
+       return ret;
 }
 
-int __memory_failure(unsigned long pfn, int trapno, int ref)
+static void set_page_hwpoison_huge_page(struct page *hpage)
+{
+       int i;
+       int nr_pages = 1 << compound_trans_order(hpage);
+       for (i = 0; i < nr_pages; i++)
+               SetPageHWPoison(hpage + i);
+}
+
+static void clear_page_hwpoison_huge_page(struct page *hpage)
+{
+       int i;
+       int nr_pages = 1 << compound_trans_order(hpage);
+       for (i = 0; i < nr_pages; i++)
+               ClearPageHWPoison(hpage + i);
+}
+
+int __memory_failure(unsigned long pfn, int trapno, int flags)
 {
-       unsigned long lru_flag;
        struct page_state *ps;
        struct page *p;
+       struct page *hpage;
        int res;
+       unsigned int nr_pages;
 
        if (!sysctl_memory_failure_recovery)
                panic("Memory failure from trap %d on page %lx", trapno, pfn);
 
        if (!pfn_valid(pfn)) {
-               action_result(pfn, "memory outside kernel control", IGNORED);
-               return -EIO;
+               printk(KERN_ERR
+                      "MCE %#lx: memory outside kernel control\n",
+                      pfn);
+               return -ENXIO;
        }
 
        p = pfn_to_page(pfn);
+       hpage = compound_head(p);
        if (TestSetPageHWPoison(p)) {
-               action_result(pfn, "already hardware poisoned", IGNORED);
+               printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
                return 0;
        }
 
-       atomic_long_add(1, &mce_bad_pages);
+       nr_pages = 1 << compound_trans_order(hpage);
+       atomic_long_add(nr_pages, &mce_bad_pages);
 
        /*
         * We need/can do nothing about count=0 pages.
         * 1) it's a free page, and therefore in safe hand:
         *    prep_new_page() will be the gate keeper.
-        * 2) it's part of a non-compound high order page.
+        * 2) it's a free hugepage, which is also safe:
+        *    an affected hugepage will be dequeued from hugepage freelist,
+        *    so there's no concern about reusing it ever after.
+        * 3) it's part of a non-compound high order page.
         *    Implies some kernel user: cannot stop them from
         *    R/W the page; let's pray that the page has been
         *    used and will be freed some time later.
         * In fact it's dangerous to directly bump up page count from 0,
         * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
         */
-       if (!get_page_unless_zero(compound_head(p))) {
-               action_result(pfn, "free or high order kernel", IGNORED);
-               return PageBuddy(compound_head(p)) ? 0 : -EBUSY;
+       if (!(flags & MF_COUNT_INCREASED) &&
+               !get_page_unless_zero(hpage)) {
+               if (is_free_buddy_page(p)) {
+                       action_result(pfn, "free buddy", DELAYED);
+                       return 0;
+               } else if (PageHuge(hpage)) {
+                       /*
+                        * Check "just unpoisoned", "filter hit", and
+                        * "race with other subpage."
+                        */
+                       lock_page(hpage);
+                       if (!PageHWPoison(hpage)
+                           || (hwpoison_filter(p) && TestClearPageHWPoison(p))
+                           || (p != hpage && TestSetPageHWPoison(hpage))) {
+                               atomic_long_sub(nr_pages, &mce_bad_pages);
+                               return 0;
+                       }
+                       set_page_hwpoison_huge_page(hpage);
+                       res = dequeue_hwpoisoned_huge_page(hpage);
+                       action_result(pfn, "free huge",
+                                     res ? IGNORED : DELAYED);
+                       unlock_page(hpage);
+                       return res;
+               } else {
+                       action_result(pfn, "high order kernel", IGNORED);
+                       return -EBUSY;
+               }
        }
 
        /*
@@ -786,47 +1062,97 @@ int __memory_failure(unsigned long pfn, int trapno, int ref)
         * The check (unnecessarily) ignores LRU pages being isolated and
         * walked by the page reclaim code, however that's not a big loss.
         */
-       if (!PageLRU(p))
-               lru_add_drain_all();
-       lru_flag = p->flags & lru;
-       if (isolate_lru_page(p)) {
-               action_result(pfn, "non LRU", IGNORED);
-               put_page(p);
-               return -EBUSY;
+       if (!PageHuge(p) && !PageTransCompound(p)) {
+               if (!PageLRU(p))
+                       shake_page(p, 0);
+               if (!PageLRU(p)) {
+                       /*
+                        * shake_page could have turned it free.
+                        */
+                       if (is_free_buddy_page(p)) {
+                               action_result(pfn, "free buddy, 2nd try",
+                                               DELAYED);
+                               return 0;
+                       }
+                       action_result(pfn, "non LRU", IGNORED);
+                       put_page(p);
+                       return -EBUSY;
+               }
        }
-       page_cache_release(p);
 
        /*
         * Lock the page and wait for writeback to finish.
         * It's very difficult to mess with pages currently under IO
         * and in many cases impossible, so we just avoid it here.
         */
-       lock_page_nosync(p);
+       lock_page(hpage);
+
+       /*
+        * unpoison always clear PG_hwpoison inside page lock
+        */
+       if (!PageHWPoison(p)) {
+               printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
+               res = 0;
+               goto out;
+       }
+       if (hwpoison_filter(p)) {
+               if (TestClearPageHWPoison(p))
+                       atomic_long_sub(nr_pages, &mce_bad_pages);
+               unlock_page(hpage);
+               put_page(hpage);
+               return 0;
+       }
+
+       /*
+        * For error on the tail page, we should set PG_hwpoison
+        * on the head page to show that the hugepage is hwpoisoned
+        */
+       if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
+               action_result(pfn, "hugepage already hardware poisoned",
+                               IGNORED);
+               unlock_page(hpage);
+               put_page(hpage);
+               return 0;
+       }
+       /*
+        * Set PG_hwpoison on all pages in an error hugepage,
+        * because containment is done in hugepage unit for now.
+        * Since we have done TestSetPageHWPoison() for the head page with
+        * page lock held, we can safely set PG_hwpoison bits on tail pages.
+        */
+       if (PageHuge(p))
+               set_page_hwpoison_huge_page(hpage);
+
        wait_on_page_writeback(p);
 
        /*
         * Now take care of user space mappings.
+        * Abort on fail: __delete_from_page_cache() assumes unmapped page.
         */
-       hwpoison_user_mappings(p, pfn, trapno);
+       if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
+               printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
+               res = -EBUSY;
+               goto out;
+       }
 
        /*
         * Torn down by someone else?
         */
-       if ((lru_flag & lru) && !PageSwapCache(p) && p->mapping == NULL) {
+       if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
                action_result(pfn, "already truncated LRU", IGNORED);
-               res = 0;
+               res = -EBUSY;
                goto out;
        }
 
        res = -EBUSY;
        for (ps = error_states;; ps++) {
-               if (((p->flags | lru_flag)& ps->mask) == ps->res) {
-                       res = page_action(ps, p, pfn, ref);
+               if ((p->flags & ps->mask) == ps->res) {
+                       res = page_action(ps, p, pfn);
                        break;
                }
        }
 out:
-       unlock_page(p);
+       unlock_page(hpage);
        return res;
 }
 EXPORT_SYMBOL_GPL(__memory_failure);
@@ -852,3 +1178,402 @@ void memory_failure(unsigned long pfn, int trapno)
 {
        __memory_failure(pfn, trapno, 0);
 }
+
+#define MEMORY_FAILURE_FIFO_ORDER      4
+#define MEMORY_FAILURE_FIFO_SIZE       (1 << MEMORY_FAILURE_FIFO_ORDER)
+
+struct memory_failure_entry {
+       unsigned long pfn;
+       int trapno;
+       int flags;
+};
+
+struct memory_failure_cpu {
+       DECLARE_KFIFO(fifo, struct memory_failure_entry,
+                     MEMORY_FAILURE_FIFO_SIZE);
+       spinlock_t lock;
+       struct work_struct work;
+};
+
+static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
+
+/**
+ * memory_failure_queue - Schedule handling memory failure of a page.
+ * @pfn: Page Number of the corrupted page
+ * @trapno: Trap number reported in the signal to user space.
+ * @flags: Flags for memory failure handling
+ *
+ * This function is called by the low level hardware error handler
+ * when it detects hardware memory corruption of a page. It schedules
+ * the recovering of error page, including dropping pages, killing
+ * processes etc.
+ *
+ * The function is primarily of use for corruptions that
+ * happen outside the current execution context (e.g. when
+ * detected by a background scrubber)
+ *
+ * Can run in IRQ context.
+ */
+void memory_failure_queue(unsigned long pfn, int trapno, int flags)
+{
+       struct memory_failure_cpu *mf_cpu;
+       unsigned long proc_flags;
+       struct memory_failure_entry entry = {
+               .pfn =          pfn,
+               .trapno =       trapno,
+               .flags =        flags,
+       };
+
+       mf_cpu = &get_cpu_var(memory_failure_cpu);
+       spin_lock_irqsave(&mf_cpu->lock, proc_flags);
+       if (kfifo_put(&mf_cpu->fifo, &entry))
+               schedule_work_on(smp_processor_id(), &mf_cpu->work);
+       else
+               pr_err("Memory failure: buffer overflow when queuing memory failure at 0x%#lx\n",
+                      pfn);
+       spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
+       put_cpu_var(memory_failure_cpu);
+}
+EXPORT_SYMBOL_GPL(memory_failure_queue);
+
+static void memory_failure_work_func(struct work_struct *work)
+{
+       struct memory_failure_cpu *mf_cpu;
+       struct memory_failure_entry entry = { 0, };
+       unsigned long proc_flags;
+       int gotten;
+
+       mf_cpu = &__get_cpu_var(memory_failure_cpu);
+       for (;;) {
+               spin_lock_irqsave(&mf_cpu->lock, proc_flags);
+               gotten = kfifo_get(&mf_cpu->fifo, &entry);
+               spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
+               if (!gotten)
+                       break;
+               __memory_failure(entry.pfn, entry.trapno, entry.flags);
+       }
+}
+
+static int __init memory_failure_init(void)
+{
+       struct memory_failure_cpu *mf_cpu;
+       int cpu;
+
+       for_each_possible_cpu(cpu) {
+               mf_cpu = &per_cpu(memory_failure_cpu, cpu);
+               spin_lock_init(&mf_cpu->lock);
+               INIT_KFIFO(mf_cpu->fifo);
+               INIT_WORK(&mf_cpu->work, memory_failure_work_func);
+       }
+
+       return 0;
+}
+core_initcall(memory_failure_init);
+
+/**
+ * unpoison_memory - Unpoison a previously poisoned page
+ * @pfn: Page number of the to be unpoisoned page
+ *
+ * Software-unpoison a page that has been poisoned by
+ * memory_failure() earlier.
+ *
+ * This is only done on the software-level, so it only works
+ * for linux injected failures, not real hardware failures
+ *
+ * Returns 0 for success, otherwise -errno.
+ */
+int unpoison_memory(unsigned long pfn)
+{
+       struct page *page;
+       struct page *p;
+       int freeit = 0;
+       unsigned int nr_pages;
+
+       if (!pfn_valid(pfn))
+               return -ENXIO;
+
+       p = pfn_to_page(pfn);
+       page = compound_head(p);
+
+       if (!PageHWPoison(p)) {
+               pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
+               return 0;
+       }
+
+       nr_pages = 1 << compound_trans_order(page);
+
+       if (!get_page_unless_zero(page)) {
+               /*
+                * Since HWPoisoned hugepage should have non-zero refcount,
+                * race between memory failure and unpoison seems to happen.
+                * In such case unpoison fails and memory failure runs
+                * to the end.
+                */
+               if (PageHuge(page)) {
+                       pr_debug("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
+                       return 0;
+               }
+               if (TestClearPageHWPoison(p))
+                       atomic_long_sub(nr_pages, &mce_bad_pages);
+               pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
+               return 0;
+       }
+
+       lock_page(page);
+       /*
+        * This test is racy because PG_hwpoison is set outside of page lock.
+        * That's acceptable because that won't trigger kernel panic. Instead,
+        * the PG_hwpoison page will be caught and isolated on the entrance to
+        * the free buddy page pool.
+        */
+       if (TestClearPageHWPoison(page)) {
+               pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
+               atomic_long_sub(nr_pages, &mce_bad_pages);
+               freeit = 1;
+               if (PageHuge(page))
+                       clear_page_hwpoison_huge_page(page);
+       }
+       unlock_page(page);
+
+       put_page(page);
+       if (freeit)
+               put_page(page);
+
+       return 0;
+}
+EXPORT_SYMBOL(unpoison_memory);
+
+static struct page *new_page(struct page *p, unsigned long private, int **x)
+{
+       int nid = page_to_nid(p);
+       if (PageHuge(p))
+               return alloc_huge_page_node(page_hstate(compound_head(p)),
+                                                  nid);
+       else
+               return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
+}
+
+/*
+ * Safely get reference count of an arbitrary page.
+ * Returns 0 for a free page, -EIO for a zero refcount page
+ * that is not free, and 1 for any other page type.
+ * For 1 the page is returned with increased page count, otherwise not.
+ */
+static int get_any_page(struct page *p, unsigned long pfn, int flags)
+{
+       int ret;
+
+       if (flags & MF_COUNT_INCREASED)
+               return 1;
+
+       /*
+        * The lock_memory_hotplug prevents a race with memory hotplug.
+        * This is a big hammer, a better would be nicer.
+        */
+       lock_memory_hotplug();
+
+       /*
+        * Isolate the page, so that it doesn't get reallocated if it
+        * was free.
+        */
+       set_migratetype_isolate(p);
+       /*
+        * When the target page is a free hugepage, just remove it
+        * from free hugepage list.
+        */
+       if (!get_page_unless_zero(compound_head(p))) {
+               if (PageHuge(p)) {
+                       pr_info("get_any_page: %#lx free huge page\n", pfn);
+                       ret = dequeue_hwpoisoned_huge_page(compound_head(p));
+               } else if (is_free_buddy_page(p)) {
+                       pr_info("get_any_page: %#lx free buddy page\n", pfn);
+                       /* Set hwpoison bit while page is still isolated */
+                       SetPageHWPoison(p);
+                       ret = 0;
+               } else {
+                       pr_info("get_any_page: %#lx: unknown zero refcount page type %lx\n",
+                               pfn, p->flags);
+                       ret = -EIO;
+               }
+       } else {
+               /* Not a free page */
+               ret = 1;
+       }
+       unset_migratetype_isolate(p);
+       unlock_memory_hotplug();
+       return ret;
+}
+
+static int soft_offline_huge_page(struct page *page, int flags)
+{
+       int ret;
+       unsigned long pfn = page_to_pfn(page);
+       struct page *hpage = compound_head(page);
+       LIST_HEAD(pagelist);
+
+       ret = get_any_page(page, pfn, flags);
+       if (ret < 0)
+               return ret;
+       if (ret == 0)
+               goto done;
+
+       if (PageHWPoison(hpage)) {
+               put_page(hpage);
+               pr_debug("soft offline: %#lx hugepage already poisoned\n", pfn);
+               return -EBUSY;
+       }
+
+       /* Keep page count to indicate a given hugepage is isolated. */
+
+       list_add(&hpage->lru, &pagelist);
+       ret = migrate_huge_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0,
+                               true);
+       if (ret) {
+               struct page *page1, *page2;
+               list_for_each_entry_safe(page1, page2, &pagelist, lru)
+                       put_page(page1);
+
+               pr_debug("soft offline: %#lx: migration failed %d, type %lx\n",
+                        pfn, ret, page->flags);
+               if (ret > 0)
+                       ret = -EIO;
+               return ret;
+       }
+done:
+       if (!PageHWPoison(hpage))
+               atomic_long_add(1 << compound_trans_order(hpage), &mce_bad_pages);
+       set_page_hwpoison_huge_page(hpage);
+       dequeue_hwpoisoned_huge_page(hpage);
+       /* keep elevated page count for bad page */
+       return ret;
+}
+
+/**
+ * soft_offline_page - Soft offline a page.
+ * @page: page to offline
+ * @flags: flags. Same as memory_failure().
+ *
+ * Returns 0 on success, otherwise negated errno.
+ *
+ * Soft offline a page, by migration or invalidation,
+ * without killing anything. This is for the case when
+ * a page is not corrupted yet (so it's still valid to access),
+ * but has had a number of corrected errors and is better taken
+ * out.
+ *
+ * The actual policy on when to do that is maintained by
+ * user space.
+ *
+ * This should never impact any application or cause data loss,
+ * however it might take some time.
+ *
+ * This is not a 100% solution for all memory, but tries to be
+ * ``good enough'' for the majority of memory.
+ */
+int soft_offline_page(struct page *page, int flags)
+{
+       int ret;
+       unsigned long pfn = page_to_pfn(page);
+
+       if (PageHuge(page))
+               return soft_offline_huge_page(page, flags);
+
+       ret = get_any_page(page, pfn, flags);
+       if (ret < 0)
+               return ret;
+       if (ret == 0)
+               goto done;
+
+       /*
+        * Page cache page we can handle?
+        */
+       if (!PageLRU(page)) {
+               /*
+                * Try to free it.
+                */
+               put_page(page);
+               shake_page(page, 1);
+
+               /*
+                * Did it turn free?
+                */
+               ret = get_any_page(page, pfn, 0);
+               if (ret < 0)
+                       return ret;
+               if (ret == 0)
+                       goto done;
+       }
+       if (!PageLRU(page)) {
+               pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
+                               pfn, page->flags);
+               return -EIO;
+       }
+
+       lock_page(page);
+       wait_on_page_writeback(page);
+
+       /*
+        * Synchronized using the page lock with memory_failure()
+        */
+       if (PageHWPoison(page)) {
+               unlock_page(page);
+               put_page(page);
+               pr_info("soft offline: %#lx page already poisoned\n", pfn);
+               return -EBUSY;
+       }
+
+       /*
+        * Try to invalidate first. This should work for
+        * non dirty unmapped page cache pages.
+        */
+       ret = invalidate_inode_page(page);
+       unlock_page(page);
+       /*
+        * RED-PEN would be better to keep it isolated here, but we
+        * would need to fix isolation locking first.
+        */
+       if (ret == 1) {
+               put_page(page);
+               ret = 0;
+               pr_info("soft_offline: %#lx: invalidated\n", pfn);
+               goto done;
+       }
+
+       /*
+        * Simple invalidation didn't work.
+        * Try to migrate to a new page instead. migrate.c
+        * handles a large number of cases for us.
+        */
+       ret = isolate_lru_page(page);
+       /*
+        * Drop page reference which is came from get_any_page()
+        * successful isolate_lru_page() already took another one.
+        */
+       put_page(page);
+       if (!ret) {
+               LIST_HEAD(pagelist);
+               inc_zone_page_state(page, NR_ISOLATED_ANON +
+                                           page_is_file_cache(page));
+               list_add(&page->lru, &pagelist);
+               ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
+                                                               0, true);
+               if (ret) {
+                       putback_lru_pages(&pagelist);
+                       pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
+                               pfn, ret, page->flags);
+                       if (ret > 0)
+                               ret = -EIO;
+               }
+       } else {
+               pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
+                               pfn, ret, page_count(page), page->flags);
+       }
+       if (ret)
+               return ret;
+
+done:
+       atomic_long_add(1, &mce_bad_pages);
+       SetPageHWPoison(page);
+       /* keep elevated page count for bad page */
+       return ret;
+}