oom: fix integer overflow of points in oom_badness
[linux-2.6.git] / mm / hugetlb.c
index 1125d81..73f17c0 100644 (file)
@@ -2,7 +2,6 @@
  * Generic hugetlb support.
  * (C) William Irwin, April 2004
  */
-#include <linux/gfp.h>
 #include <linux/list.h>
 #include <linux/init.h>
 #include <linux/module.h>
 #include <linux/mutex.h>
 #include <linux/bootmem.h>
 #include <linux/sysfs.h>
+#include <linux/slab.h>
+#include <linux/rmap.h>
+#include <linux/swap.h>
+#include <linux/swapops.h>
 
 #include <asm/page.h>
 #include <asm/pgtable.h>
-#include <asm/io.h>
+#include <linux/io.h>
 
 #include <linux/hugetlb.h>
+#include <linux/node.h>
 #include "internal.h"
 
 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
@@ -58,10 +62,10 @@ static DEFINE_SPINLOCK(hugetlb_lock);
  * must either hold the mmap_sem for write, or the mmap_sem for read and
  * the hugetlb_instantiation mutex:
  *
- *     down_write(&mm->mmap_sem);
+ *     down_write(&mm->mmap_sem);
  * or
- *     down_read(&mm->mmap_sem);
- *     mutex_lock(&hugetlb_instantiation_mutex);
+ *     down_read(&mm->mmap_sem);
+ *     mutex_lock(&hugetlb_instantiation_mutex);
  */
 struct file_region {
        struct list_head link;
@@ -142,7 +146,7 @@ static long region_chg(struct list_head *head, long f, long t)
                if (rg->from > t)
                        return chg;
 
-               /* We overlap with this area, if it extends futher than
+               /* We overlap with this area, if it extends further than
                 * us then we must extend ourselves.  Account for its
                 * existing reservation. */
                if (rg->to > t) {
@@ -219,6 +223,12 @@ static pgoff_t vma_hugecache_offset(struct hstate *h,
                        (vma->vm_pgoff >> huge_page_order(h));
 }
 
+pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
+                                    unsigned long address)
+{
+       return vma_hugecache_offset(hstate_vma(vma), vma, address);
+}
+
 /*
  * Return the size of the pages allocated when backing a VMA. In the majority
  * cases this will be same size as used by the page table entries.
@@ -384,67 +394,37 @@ static int vma_has_reserves(struct vm_area_struct *vma)
        return 0;
 }
 
-static void clear_gigantic_page(struct page *page,
-                       unsigned long addr, unsigned long sz)
-{
-       int i;
-       struct page *p = page;
-
-       might_sleep();
-       for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
-               cond_resched();
-               clear_user_highpage(p, addr + i * PAGE_SIZE);
-       }
-}
-static void clear_huge_page(struct page *page,
-                       unsigned long addr, unsigned long sz)
+static void copy_gigantic_page(struct page *dst, struct page *src)
 {
        int i;
-
-       if (unlikely(sz > MAX_ORDER_NR_PAGES)) {
-               clear_gigantic_page(page, addr, sz);
-               return;
-       }
-
-       might_sleep();
-       for (i = 0; i < sz/PAGE_SIZE; i++) {
-               cond_resched();
-               clear_user_highpage(page + i, addr + i * PAGE_SIZE);
-       }
-}
-
-static void copy_gigantic_page(struct page *dst, struct page *src,
-                          unsigned long addr, struct vm_area_struct *vma)
-{
-       int i;
-       struct hstate *h = hstate_vma(vma);
+       struct hstate *h = page_hstate(src);
        struct page *dst_base = dst;
        struct page *src_base = src;
-       might_sleep();
+
        for (i = 0; i < pages_per_huge_page(h); ) {
                cond_resched();
-               copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
+               copy_highpage(dst, src);
 
                i++;
                dst = mem_map_next(dst, dst_base, i);
                src = mem_map_next(src, src_base, i);
        }
 }
-static void copy_huge_page(struct page *dst, struct page *src,
-                          unsigned long addr, struct vm_area_struct *vma)
+
+void copy_huge_page(struct page *dst, struct page *src)
 {
        int i;
-       struct hstate *h = hstate_vma(vma);
+       struct hstate *h = page_hstate(src);
 
        if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
-               copy_gigantic_page(dst, src, addr, vma);
+               copy_gigantic_page(dst, src);
                return;
        }
 
        might_sleep();
        for (i = 0; i < pages_per_huge_page(h); i++) {
                cond_resched();
-               copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
+               copy_highpage(dst + i, src + i);
        }
 }
 
@@ -456,19 +436,34 @@ static void enqueue_huge_page(struct hstate *h, struct page *page)
        h->free_huge_pages_node[nid]++;
 }
 
+static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
+{
+       struct page *page;
+
+       if (list_empty(&h->hugepage_freelists[nid]))
+               return NULL;
+       page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
+       list_del(&page->lru);
+       set_page_refcounted(page);
+       h->free_huge_pages--;
+       h->free_huge_pages_node[nid]--;
+       return page;
+}
+
 static struct page *dequeue_huge_page_vma(struct hstate *h,
                                struct vm_area_struct *vma,
                                unsigned long address, int avoid_reserve)
 {
-       int nid;
        struct page *page = NULL;
        struct mempolicy *mpol;
        nodemask_t *nodemask;
-       struct zonelist *zonelist = huge_zonelist(vma, address,
-                                       htlb_alloc_mask, &mpol, &nodemask);
+       struct zonelist *zonelist;
        struct zone *zone;
        struct zoneref *z;
 
+       get_mems_allowed();
+       zonelist = huge_zonelist(vma, address,
+                                       htlb_alloc_mask, &mpol, &nodemask);
        /*
         * A child process with MAP_PRIVATE mappings created by their parent
         * have no page reserves. This check ensures that reservations are
@@ -476,30 +471,26 @@ static struct page *dequeue_huge_page_vma(struct hstate *h,
         */
        if (!vma_has_reserves(vma) &&
                        h->free_huge_pages - h->resv_huge_pages == 0)
-               return NULL;
+               goto err;
 
        /* If reserves cannot be used, ensure enough pages are in the pool */
        if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
-               return NULL;
+               goto err;
 
        for_each_zone_zonelist_nodemask(zone, z, zonelist,
                                                MAX_NR_ZONES - 1, nodemask) {
-               nid = zone_to_nid(zone);
-               if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
-                   !list_empty(&h->hugepage_freelists[nid])) {
-                       page = list_entry(h->hugepage_freelists[nid].next,
-                                         struct page, lru);
-                       list_del(&page->lru);
-                       h->free_huge_pages--;
-                       h->free_huge_pages_node[nid]--;
-
-                       if (!avoid_reserve)
-                               decrement_hugepage_resv_vma(h, vma);
-
-                       break;
+               if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
+                       page = dequeue_huge_page_node(h, zone_to_nid(zone));
+                       if (page) {
+                               if (!avoid_reserve)
+                                       decrement_hugepage_resv_vma(h, vma);
+                               break;
+                       }
                }
        }
+err:
        mpol_cond_put(mpol);
+       put_mems_allowed();
        return page;
 }
 
@@ -512,9 +503,10 @@ static void update_and_free_page(struct hstate *h, struct page *page)
        h->nr_huge_pages--;
        h->nr_huge_pages_node[page_to_nid(page)]--;
        for (i = 0; i < pages_per_huge_page(h); i++) {
-               page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
-                               1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
-                               1 << PG_private | 1<< PG_writeback);
+               page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
+                               1 << PG_referenced | 1 << PG_dirty |
+                               1 << PG_active | 1 << PG_reserved |
+                               1 << PG_private | 1 << PG_writeback);
        }
        set_compound_page_dtor(page, NULL);
        set_page_refcounted(page);
@@ -545,7 +537,9 @@ static void free_huge_page(struct page *page)
 
        mapping = (struct address_space *) page_private(page);
        set_page_private(page, 0);
+       page->mapping = NULL;
        BUG_ON(page_count(page));
+       BUG_ON(page_mapcount(page));
        INIT_LIST_HEAD(&page->lru);
 
        spin_lock(&hugetlb_lock);
@@ -582,6 +576,7 @@ static void prep_compound_gigantic_page(struct page *page, unsigned long order)
        __SetPageHead(page);
        for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
                __SetPageTail(p);
+               set_page_count(p, 0);
                p->first_page = page;
        }
 }
@@ -598,6 +593,7 @@ int PageHuge(struct page *page)
 
        return dtor == free_huge_page;
 }
+EXPORT_SYMBOL_GPL(PageHuge);
 
 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
 {
@@ -752,11 +748,10 @@ static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
        return ret;
 }
 
-static struct page *alloc_buddy_huge_page(struct hstate *h,
-                       struct vm_area_struct *vma, unsigned long address)
+static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
 {
        struct page *page;
-       unsigned int nid;
+       unsigned int r_nid;
 
        if (h->order >= MAX_ORDER)
                return NULL;
@@ -794,9 +789,14 @@ static struct page *alloc_buddy_huge_page(struct hstate *h,
        }
        spin_unlock(&hugetlb_lock);
 
-       page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
-                                       __GFP_REPEAT|__GFP_NOWARN,
-                                       huge_page_order(h));
+       if (nid == NUMA_NO_NODE)
+               page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
+                                  __GFP_REPEAT|__GFP_NOWARN,
+                                  huge_page_order(h));
+       else
+               page = alloc_pages_exact_node(nid,
+                       htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
+                       __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
 
        if (page && arch_prepare_hugepage(page)) {
                __free_pages(page, huge_page_order(h));
@@ -805,19 +805,13 @@ static struct page *alloc_buddy_huge_page(struct hstate *h,
 
        spin_lock(&hugetlb_lock);
        if (page) {
-               /*
-                * This page is now managed by the hugetlb allocator and has
-                * no users -- drop the buddy allocator's reference.
-                */
-               put_page_testzero(page);
-               VM_BUG_ON(page_count(page));
-               nid = page_to_nid(page);
+               r_nid = page_to_nid(page);
                set_compound_page_dtor(page, free_huge_page);
                /*
                 * We incremented the global counters already
                 */
-               h->nr_huge_pages_node[nid]++;
-               h->surplus_huge_pages_node[nid]++;
+               h->nr_huge_pages_node[r_nid]++;
+               h->surplus_huge_pages_node[r_nid]++;
                __count_vm_event(HTLB_BUDDY_PGALLOC);
        } else {
                h->nr_huge_pages--;
@@ -830,7 +824,26 @@ static struct page *alloc_buddy_huge_page(struct hstate *h,
 }
 
 /*
- * Increase the hugetlb pool such that it can accomodate a reservation
+ * This allocation function is useful in the context where vma is irrelevant.
+ * E.g. soft-offlining uses this function because it only cares physical
+ * address of error page.
+ */
+struct page *alloc_huge_page_node(struct hstate *h, int nid)
+{
+       struct page *page;
+
+       spin_lock(&hugetlb_lock);
+       page = dequeue_huge_page_node(h, nid);
+       spin_unlock(&hugetlb_lock);
+
+       if (!page)
+               page = alloc_buddy_huge_page(h, nid);
+
+       return page;
+}
+
+/*
+ * Increase the hugetlb pool such that it can accommodate a reservation
  * of size 'delta'.
  */
 static int gather_surplus_pages(struct hstate *h, int delta)
@@ -853,17 +866,14 @@ static int gather_surplus_pages(struct hstate *h, int delta)
 retry:
        spin_unlock(&hugetlb_lock);
        for (i = 0; i < needed; i++) {
-               page = alloc_buddy_huge_page(h, NULL, 0);
-               if (!page) {
+               page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
+               if (!page)
                        /*
                         * We were not able to allocate enough pages to
                         * satisfy the entire reservation so we free what
                         * we've allocated so far.
                         */
-                       spin_lock(&hugetlb_lock);
-                       needed = 0;
                        goto free;
-               }
 
                list_add(&page->lru, &surplus_list);
        }
@@ -881,7 +891,7 @@ retry:
 
        /*
         * The surplus_list now contains _at_least_ the number of extra pages
-        * needed to accomodate the reservation.  Add the appropriate number
+        * needed to accommodate the reservation.  Add the appropriate number
         * of pages to the hugetlb pool and free the extras back to the buddy
         * allocator.  Commit the entire reservation here to prevent another
         * process from stealing the pages as they are added to the pool but
@@ -890,31 +900,31 @@ retry:
        needed += allocated;
        h->resv_huge_pages += delta;
        ret = 0;
-free:
+
+       spin_unlock(&hugetlb_lock);
        /* Free the needed pages to the hugetlb pool */
        list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
                if ((--needed) < 0)
                        break;
                list_del(&page->lru);
+               /*
+                * This page is now managed by the hugetlb allocator and has
+                * no users -- drop the buddy allocator's reference.
+                */
+               put_page_testzero(page);
+               VM_BUG_ON(page_count(page));
                enqueue_huge_page(h, page);
        }
 
        /* Free unnecessary surplus pages to the buddy allocator */
+free:
        if (!list_empty(&surplus_list)) {
-               spin_unlock(&hugetlb_lock);
                list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
                        list_del(&page->lru);
-                       /*
-                        * The page has a reference count of zero already, so
-                        * call free_huge_page directly instead of using
-                        * put_page.  This must be done with hugetlb_lock
-                        * unlocked which is safe because free_huge_page takes
-                        * hugetlb_lock before deciding how to free the page.
-                        */
-                       free_huge_page(page);
+                       put_page(page);
                }
-               spin_lock(&hugetlb_lock);
        }
+       spin_lock(&hugetlb_lock);
 
        return ret;
 }
@@ -941,14 +951,14 @@ static void return_unused_surplus_pages(struct hstate *h,
 
        /*
         * We want to release as many surplus pages as possible, spread
-        * evenly across all nodes. Iterate across all nodes until we
-        * can no longer free unreserved surplus pages. This occurs when
-        * the nodes with surplus pages have no free pages.
-        * free_pool_huge_page() will balance the the frees across the
-        * on-line nodes for us and will handle the hstate accounting.
+        * evenly across all nodes with memory. Iterate across these nodes
+        * until we can no longer free unreserved surplus pages. This occurs
+        * when the nodes with surplus pages have no free pages.
+        * free_pool_huge_page() will balance the the freed pages across the
+        * on-line nodes with memory and will handle the hstate accounting.
         */
        while (nr_pages--) {
-               if (!free_pool_huge_page(h, &node_online_map, 1))
+               if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
                        break;
        }
 }
@@ -1024,24 +1034,23 @@ static struct page *alloc_huge_page(struct vm_area_struct *vma,
         */
        chg = vma_needs_reservation(h, vma, addr);
        if (chg < 0)
-               return ERR_PTR(chg);
+               return ERR_PTR(-VM_FAULT_OOM);
        if (chg)
                if (hugetlb_get_quota(inode->i_mapping, chg))
-                       return ERR_PTR(-ENOSPC);
+                       return ERR_PTR(-VM_FAULT_SIGBUS);
 
        spin_lock(&hugetlb_lock);
        page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
        spin_unlock(&hugetlb_lock);
 
        if (!page) {
-               page = alloc_buddy_huge_page(h, vma, addr);
+               page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
                if (!page) {
                        hugetlb_put_quota(inode->i_mapping, chg);
-                       return ERR_PTR(-VM_FAULT_OOM);
+                       return ERR_PTR(-VM_FAULT_SIGBUS);
                }
        }
 
-       set_page_refcounted(page);
        set_page_private(page, (unsigned long) mapping);
 
        vma_commit_reservation(h, vma, addr);
@@ -1052,14 +1061,14 @@ static struct page *alloc_huge_page(struct vm_area_struct *vma,
 int __weak alloc_bootmem_huge_page(struct hstate *h)
 {
        struct huge_bootmem_page *m;
-       int nr_nodes = nodes_weight(node_online_map);
+       int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
 
        while (nr_nodes) {
                void *addr;
 
                addr = __alloc_bootmem_node_nopanic(
                                NODE_DATA(hstate_next_node_to_alloc(h,
-                                                       &node_online_map)),
+                                               &node_states[N_HIGH_MEMORY])),
                                huge_page_size(h), huge_page_size(h), 0);
 
                if (addr) {
@@ -1097,12 +1106,28 @@ static void __init gather_bootmem_prealloc(void)
        struct huge_bootmem_page *m;
 
        list_for_each_entry(m, &huge_boot_pages, list) {
-               struct page *page = virt_to_page(m);
                struct hstate *h = m->hstate;
+               struct page *page;
+
+#ifdef CONFIG_HIGHMEM
+               page = pfn_to_page(m->phys >> PAGE_SHIFT);
+               free_bootmem_late((unsigned long)m,
+                                 sizeof(struct huge_bootmem_page));
+#else
+               page = virt_to_page(m);
+#endif
                __ClearPageReserved(page);
                WARN_ON(page_count(page) != 1);
                prep_compound_huge_page(page, h->order);
                prep_new_huge_page(h, page, page_to_nid(page));
+               /*
+                * If we had gigantic hugepages allocated at boot time, we need
+                * to restore the 'stolen' pages to totalram_pages in order to
+                * fix confusing memory reports from free(1) and another
+                * side-effects, like CommitLimit going negative.
+                */
+               if (h->order > (MAX_ORDER - 1))
+                       totalram_pages += 1 << h->order;
        }
 }
 
@@ -1114,7 +1139,8 @@ static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
                if (h->order >= MAX_ORDER) {
                        if (!alloc_bootmem_huge_page(h))
                                break;
-               } else if (!alloc_fresh_huge_page(h, &node_online_map))
+               } else if (!alloc_fresh_huge_page(h,
+                                        &node_states[N_HIGH_MEMORY]))
                        break;
        }
        h->max_huge_pages = i;
@@ -1276,6 +1302,9 @@ static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
                if (!ret)
                        goto out;
 
+               /* Bail for signals. Probably ctrl-c from user */
+               if (signal_pending(current))
+                       goto out;
        }
 
        /*
@@ -1320,45 +1349,86 @@ out:
 static struct kobject *hugepages_kobj;
 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
 
-static struct hstate *kobj_to_hstate(struct kobject *kobj)
+static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
+
+static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
 {
        int i;
+
        for (i = 0; i < HUGE_MAX_HSTATE; i++)
-               if (hstate_kobjs[i] == kobj)
+               if (hstate_kobjs[i] == kobj) {
+                       if (nidp)
+                               *nidp = NUMA_NO_NODE;
                        return &hstates[i];
-       BUG();
-       return NULL;
+               }
+
+       return kobj_to_node_hstate(kobj, nidp);
 }
 
 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
                                        struct kobj_attribute *attr, char *buf)
 {
-       struct hstate *h = kobj_to_hstate(kobj);
-       return sprintf(buf, "%lu\n", h->nr_huge_pages);
+       struct hstate *h;
+       unsigned long nr_huge_pages;
+       int nid;
+
+       h = kobj_to_hstate(kobj, &nid);
+       if (nid == NUMA_NO_NODE)
+               nr_huge_pages = h->nr_huge_pages;
+       else
+               nr_huge_pages = h->nr_huge_pages_node[nid];
+
+       return sprintf(buf, "%lu\n", nr_huge_pages);
 }
+
 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
                        struct kobject *kobj, struct kobj_attribute *attr,
                        const char *buf, size_t len)
 {
        int err;
+       int nid;
        unsigned long count;
-       struct hstate *h = kobj_to_hstate(kobj);
-       NODEMASK_ALLOC(nodemask_t, nodes_allowed);
+       struct hstate *h;
+       NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
 
        err = strict_strtoul(buf, 10, &count);
        if (err)
-               return 0;
+               goto out;
 
-       if (!(obey_mempolicy && init_nodemask_of_mempolicy(nodes_allowed))) {
-               NODEMASK_FREE(nodes_allowed);
-               nodes_allowed = &node_online_map;
+       h = kobj_to_hstate(kobj, &nid);
+       if (h->order >= MAX_ORDER) {
+               err = -EINVAL;
+               goto out;
        }
+
+       if (nid == NUMA_NO_NODE) {
+               /*
+                * global hstate attribute
+                */
+               if (!(obey_mempolicy &&
+                               init_nodemask_of_mempolicy(nodes_allowed))) {
+                       NODEMASK_FREE(nodes_allowed);
+                       nodes_allowed = &node_states[N_HIGH_MEMORY];
+               }
+       } else if (nodes_allowed) {
+               /*
+                * per node hstate attribute: adjust count to global,
+                * but restrict alloc/free to the specified node.
+                */
+               count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
+               init_nodemask_of_node(nodes_allowed, nid);
+       } else
+               nodes_allowed = &node_states[N_HIGH_MEMORY];
+
        h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
 
-       if (nodes_allowed != &node_online_map)
+       if (nodes_allowed != &node_states[N_HIGH_MEMORY])
                NODEMASK_FREE(nodes_allowed);
 
        return len;
+out:
+       NODEMASK_FREE(nodes_allowed);
+       return err;
 }
 
 static ssize_t nr_hugepages_show(struct kobject *kobj,
@@ -1398,19 +1468,23 @@ HSTATE_ATTR(nr_hugepages_mempolicy);
 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
                                        struct kobj_attribute *attr, char *buf)
 {
-       struct hstate *h = kobj_to_hstate(kobj);
+       struct hstate *h = kobj_to_hstate(kobj, NULL);
        return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
 }
+
 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
                struct kobj_attribute *attr, const char *buf, size_t count)
 {
        int err;
        unsigned long input;
-       struct hstate *h = kobj_to_hstate(kobj);
+       struct hstate *h = kobj_to_hstate(kobj, NULL);
+
+       if (h->order >= MAX_ORDER)
+               return -EINVAL;
 
        err = strict_strtoul(buf, 10, &input);
        if (err)
-               return 0;
+               return err;
 
        spin_lock(&hugetlb_lock);
        h->nr_overcommit_huge_pages = input;
@@ -1423,15 +1497,24 @@ HSTATE_ATTR(nr_overcommit_hugepages);
 static ssize_t free_hugepages_show(struct kobject *kobj,
                                        struct kobj_attribute *attr, char *buf)
 {
-       struct hstate *h = kobj_to_hstate(kobj);
-       return sprintf(buf, "%lu\n", h->free_huge_pages);
+       struct hstate *h;
+       unsigned long free_huge_pages;
+       int nid;
+
+       h = kobj_to_hstate(kobj, &nid);
+       if (nid == NUMA_NO_NODE)
+               free_huge_pages = h->free_huge_pages;
+       else
+               free_huge_pages = h->free_huge_pages_node[nid];
+
+       return sprintf(buf, "%lu\n", free_huge_pages);
 }
 HSTATE_ATTR_RO(free_hugepages);
 
 static ssize_t resv_hugepages_show(struct kobject *kobj,
                                        struct kobj_attribute *attr, char *buf)
 {
-       struct hstate *h = kobj_to_hstate(kobj);
+       struct hstate *h = kobj_to_hstate(kobj, NULL);
        return sprintf(buf, "%lu\n", h->resv_huge_pages);
 }
 HSTATE_ATTR_RO(resv_hugepages);
@@ -1439,8 +1522,17 @@ HSTATE_ATTR_RO(resv_hugepages);
 static ssize_t surplus_hugepages_show(struct kobject *kobj,
                                        struct kobj_attribute *attr, char *buf)
 {
-       struct hstate *h = kobj_to_hstate(kobj);
-       return sprintf(buf, "%lu\n", h->surplus_huge_pages);
+       struct hstate *h;
+       unsigned long surplus_huge_pages;
+       int nid;
+
+       h = kobj_to_hstate(kobj, &nid);
+       if (nid == NUMA_NO_NODE)
+               surplus_huge_pages = h->surplus_huge_pages;
+       else
+               surplus_huge_pages = h->surplus_huge_pages_node[nid];
+
+       return sprintf(buf, "%lu\n", surplus_huge_pages);
 }
 HSTATE_ATTR_RO(surplus_hugepages);
 
@@ -1460,19 +1552,20 @@ static struct attribute_group hstate_attr_group = {
        .attrs = hstate_attrs,
 };
 
-static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
+static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
+                                   struct kobject **hstate_kobjs,
+                                   struct attribute_group *hstate_attr_group)
 {
        int retval;
+       int hi = h - hstates;
 
-       hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
-                                                       hugepages_kobj);
-       if (!hstate_kobjs[h - hstates])
+       hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
+       if (!hstate_kobjs[hi])
                return -ENOMEM;
 
-       retval = sysfs_create_group(hstate_kobjs[h - hstates],
-                                                       &hstate_attr_group);
+       retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
        if (retval)
-               kobject_put(hstate_kobjs[h - hstates]);
+               kobject_put(hstate_kobjs[hi]);
 
        return retval;
 }
@@ -1487,17 +1580,184 @@ static void __init hugetlb_sysfs_init(void)
                return;
 
        for_each_hstate(h) {
-               err = hugetlb_sysfs_add_hstate(h);
+               err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
+                                        hstate_kobjs, &hstate_attr_group);
                if (err)
                        printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
                                                                h->name);
        }
 }
 
+#ifdef CONFIG_NUMA
+
+/*
+ * node_hstate/s - associate per node hstate attributes, via their kobjects,
+ * with node sysdevs in node_devices[] using a parallel array.  The array
+ * index of a node sysdev or _hstate == node id.
+ * This is here to avoid any static dependency of the node sysdev driver, in
+ * the base kernel, on the hugetlb module.
+ */
+struct node_hstate {
+       struct kobject          *hugepages_kobj;
+       struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
+};
+struct node_hstate node_hstates[MAX_NUMNODES];
+
+/*
+ * A subset of global hstate attributes for node sysdevs
+ */
+static struct attribute *per_node_hstate_attrs[] = {
+       &nr_hugepages_attr.attr,
+       &free_hugepages_attr.attr,
+       &surplus_hugepages_attr.attr,
+       NULL,
+};
+
+static struct attribute_group per_node_hstate_attr_group = {
+       .attrs = per_node_hstate_attrs,
+};
+
+/*
+ * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
+ * Returns node id via non-NULL nidp.
+ */
+static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
+{
+       int nid;
+
+       for (nid = 0; nid < nr_node_ids; nid++) {
+               struct node_hstate *nhs = &node_hstates[nid];
+               int i;
+               for (i = 0; i < HUGE_MAX_HSTATE; i++)
+                       if (nhs->hstate_kobjs[i] == kobj) {
+                               if (nidp)
+                                       *nidp = nid;
+                               return &hstates[i];
+                       }
+       }
+
+       BUG();
+       return NULL;
+}
+
+/*
+ * Unregister hstate attributes from a single node sysdev.
+ * No-op if no hstate attributes attached.
+ */
+void hugetlb_unregister_node(struct node *node)
+{
+       struct hstate *h;
+       struct node_hstate *nhs = &node_hstates[node->sysdev.id];
+
+       if (!nhs->hugepages_kobj)
+               return;         /* no hstate attributes */
+
+       for_each_hstate(h)
+               if (nhs->hstate_kobjs[h - hstates]) {
+                       kobject_put(nhs->hstate_kobjs[h - hstates]);
+                       nhs->hstate_kobjs[h - hstates] = NULL;
+               }
+
+       kobject_put(nhs->hugepages_kobj);
+       nhs->hugepages_kobj = NULL;
+}
+
+/*
+ * hugetlb module exit:  unregister hstate attributes from node sysdevs
+ * that have them.
+ */
+static void hugetlb_unregister_all_nodes(void)
+{
+       int nid;
+
+       /*
+        * disable node sysdev registrations.
+        */
+       register_hugetlbfs_with_node(NULL, NULL);
+
+       /*
+        * remove hstate attributes from any nodes that have them.
+        */
+       for (nid = 0; nid < nr_node_ids; nid++)
+               hugetlb_unregister_node(&node_devices[nid]);
+}
+
+/*
+ * Register hstate attributes for a single node sysdev.
+ * No-op if attributes already registered.
+ */
+void hugetlb_register_node(struct node *node)
+{
+       struct hstate *h;
+       struct node_hstate *nhs = &node_hstates[node->sysdev.id];
+       int err;
+
+       if (nhs->hugepages_kobj)
+               return;         /* already allocated */
+
+       nhs->hugepages_kobj = kobject_create_and_add("hugepages",
+                                                       &node->sysdev.kobj);
+       if (!nhs->hugepages_kobj)
+               return;
+
+       for_each_hstate(h) {
+               err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
+                                               nhs->hstate_kobjs,
+                                               &per_node_hstate_attr_group);
+               if (err) {
+                       printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
+                                       " for node %d\n",
+                                               h->name, node->sysdev.id);
+                       hugetlb_unregister_node(node);
+                       break;
+               }
+       }
+}
+
+/*
+ * hugetlb init time:  register hstate attributes for all registered node
+ * sysdevs of nodes that have memory.  All on-line nodes should have
+ * registered their associated sysdev by this time.
+ */
+static void hugetlb_register_all_nodes(void)
+{
+       int nid;
+
+       for_each_node_state(nid, N_HIGH_MEMORY) {
+               struct node *node = &node_devices[nid];
+               if (node->sysdev.id == nid)
+                       hugetlb_register_node(node);
+       }
+
+       /*
+        * Let the node sysdev driver know we're here so it can
+        * [un]register hstate attributes on node hotplug.
+        */
+       register_hugetlbfs_with_node(hugetlb_register_node,
+                                    hugetlb_unregister_node);
+}
+#else  /* !CONFIG_NUMA */
+
+static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
+{
+       BUG();
+       if (nidp)
+               *nidp = -1;
+       return NULL;
+}
+
+static void hugetlb_unregister_all_nodes(void) { }
+
+static void hugetlb_register_all_nodes(void) { }
+
+#endif
+
 static void __exit hugetlb_exit(void)
 {
        struct hstate *h;
 
+       hugetlb_unregister_all_nodes();
+
        for_each_hstate(h) {
                kobject_put(hstate_kobjs[h - hstates]);
        }
@@ -1532,6 +1792,8 @@ static int __init hugetlb_init(void)
 
        hugetlb_sysfs_init();
 
+       hugetlb_register_all_nodes();
+
        return 0;
 }
 module_init(hugetlb_init);
@@ -1555,8 +1817,8 @@ void __init hugetlb_add_hstate(unsigned order)
        h->free_huge_pages = 0;
        for (i = 0; i < MAX_NUMNODES; ++i)
                INIT_LIST_HEAD(&h->hugepage_freelists[i]);
-       h->next_nid_to_alloc = first_node(node_online_map);
-       h->next_nid_to_free = first_node(node_online_map);
+       h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
+       h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
        snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
                                        huge_page_size(h)/1024);
 
@@ -1625,16 +1887,22 @@ static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
 {
        struct hstate *h = &default_hstate;
        unsigned long tmp;
+       int ret;
+
+       tmp = h->max_huge_pages;
 
-       if (!write)
-               tmp = h->max_huge_pages;
+       if (write && h->order >= MAX_ORDER)
+               return -EINVAL;
 
        table->data = &tmp;
        table->maxlen = sizeof(unsigned long);
-       proc_doulongvec_minmax(table, write, buffer, length, ppos);
+       ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
+       if (ret)
+               goto out;
 
        if (write) {
-               NODEMASK_ALLOC(nodemask_t, nodes_allowed);
+               NODEMASK_ALLOC(nodemask_t, nodes_allowed,
+                                               GFP_KERNEL | __GFP_NORETRY);
                if (!(obey_mempolicy &&
                               init_nodemask_of_mempolicy(nodes_allowed))) {
                        NODEMASK_FREE(nodes_allowed);
@@ -1645,8 +1913,8 @@ static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
                if (nodes_allowed != &node_states[N_HIGH_MEMORY])
                        NODEMASK_FREE(nodes_allowed);
        }
-
-       return 0;
+out:
+       return ret;
 }
 
 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
@@ -1684,21 +1952,26 @@ int hugetlb_overcommit_handler(struct ctl_table *table, int write,
 {
        struct hstate *h = &default_hstate;
        unsigned long tmp;
+       int ret;
 
-       if (!write)
-               tmp = h->nr_overcommit_huge_pages;
+       tmp = h->nr_overcommit_huge_pages;
+
+       if (write && h->order >= MAX_ORDER)
+               return -EINVAL;
 
        table->data = &tmp;
        table->maxlen = sizeof(unsigned long);
-       proc_doulongvec_minmax(table, write, buffer, length, ppos);
+       ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
+       if (ret)
+               goto out;
 
        if (write) {
                spin_lock(&hugetlb_lock);
                h->nr_overcommit_huge_pages = tmp;
                spin_unlock(&hugetlb_lock);
        }
-
-       return 0;
+out:
+       return ret;
 }
 
 #endif /* CONFIG_SYSCTL */
@@ -1787,7 +2060,7 @@ static void hugetlb_vm_op_open(struct vm_area_struct *vma)
         * This new VMA should share its siblings reservation map if present.
         * The VMA will only ever have a valid reservation map pointer where
         * it is being copied for another still existing VMA.  As that VMA
-        * has a reference to the reservation map it cannot dissappear until
+        * has a reference to the reservation map it cannot disappear until
         * after this open call completes.  It is therefore safe to take a
         * new reference here without additional locking.
         */
@@ -1860,9 +2133,8 @@ static void set_huge_ptep_writable(struct vm_area_struct *vma,
        pte_t entry;
 
        entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
-       if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
-               update_mmu_cache(vma, address, entry);
-       }
+       if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
+               update_mmu_cache(vma, address, ptep);
 }
 
 
@@ -1898,6 +2170,7 @@ int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
                        entry = huge_ptep_get(src_pte);
                        ptepage = pte_page(entry);
                        get_page(ptepage);
+                       page_dup_rmap(ptepage);
                        set_huge_pte_at(dst, addr, dst_pte, entry);
                }
                spin_unlock(&src->page_table_lock);
@@ -1909,6 +2182,32 @@ nomem:
        return -ENOMEM;
 }
 
+static int is_hugetlb_entry_migration(pte_t pte)
+{
+       swp_entry_t swp;
+
+       if (huge_pte_none(pte) || pte_present(pte))
+               return 0;
+       swp = pte_to_swp_entry(pte);
+       if (non_swap_entry(swp) && is_migration_entry(swp))
+               return 1;
+       else
+               return 0;
+}
+
+static int is_hugetlb_entry_hwpoisoned(pte_t pte)
+{
+       swp_entry_t swp;
+
+       if (huge_pte_none(pte) || pte_present(pte))
+               return 0;
+       swp = pte_to_swp_entry(pte);
+       if (non_swap_entry(swp) && is_hwpoison_entry(swp))
+               return 1;
+       else
+               return 0;
+}
+
 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
                            unsigned long end, struct page *ref_page)
 {
@@ -1922,7 +2221,7 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
        unsigned long sz = huge_page_size(h);
 
        /*
-        * A page gathering list, protected by per file i_mmap_lock. The
+        * A page gathering list, protected by per file i_mmap_mutex. The
         * lock is used to avoid list corruption from multiple unmapping
         * of the same page since we are using page->lru.
         */
@@ -1967,6 +2266,12 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
                if (huge_pte_none(pte))
                        continue;
 
+               /*
+                * HWPoisoned hugepage is already unmapped and dropped reference
+                */
+               if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
+                       continue;
+
                page = pte_page(pte);
                if (pte_dirty(pte))
                        set_page_dirty(page);
@@ -1976,6 +2281,7 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
        flush_tlb_range(vma, start, end);
        mmu_notifier_invalidate_range_end(mm, start, end);
        list_for_each_entry_safe(page, tmp, &page_list, lru) {
+               page_remove_rmap(page);
                list_del(&page->lru);
                put_page(page);
        }
@@ -1984,9 +2290,9 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
                          unsigned long end, struct page *ref_page)
 {
-       spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
+       mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
        __unmap_hugepage_range(vma, start, end, ref_page);
-       spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
+       mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
 }
 
 /*
@@ -2013,6 +2319,12 @@ static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
                + (vma->vm_pgoff >> PAGE_SHIFT);
        mapping = (struct address_space *)page_private(page);
 
+       /*
+        * Take the mapping lock for the duration of the table walk. As
+        * this mapping should be shared between all the VMAs,
+        * __unmap_hugepage_range() is called as the lock is already held
+        */
+       mutex_lock(&mapping->i_mmap_mutex);
        vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
                /* Do not unmap the current VMA */
                if (iter_vma == vma)
@@ -2026,14 +2338,18 @@ static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
                 * from the time of fork. This would look like data corruption
                 */
                if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
-                       unmap_hugepage_range(iter_vma,
+                       __unmap_hugepage_range(iter_vma,
                                address, address + huge_page_size(h),
                                page);
        }
+       mutex_unlock(&mapping->i_mmap_mutex);
 
        return 1;
 }
 
+/*
+ * Hugetlb_cow() should be called with page lock of the original hugepage held.
+ */
 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
                        unsigned long address, pte_t *ptep, pte_t pte,
                        struct page *pagecache_page)
@@ -2048,8 +2364,10 @@ static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 retry_avoidcopy:
        /* If no-one else is actually using this page, avoid the copy
         * and just make the page writable */
-       avoidcopy = (page_count(old_page) == 1);
+       avoidcopy = (page_mapcount(old_page) == 1);
        if (avoidcopy) {
+               if (PageAnon(old_page))
+                       page_move_anon_rmap(old_page, vma, address);
                set_huge_ptep_writable(vma, address, ptep);
                return 0;
        }
@@ -2069,6 +2387,9 @@ retry_avoidcopy:
                outside_reserve = 1;
 
        page_cache_get(old_page);
+
+       /* Drop page_table_lock as buddy allocator may be called */
+       spin_unlock(&mm->page_table_lock);
        new_page = alloc_huge_page(vma, address, outside_reserve);
 
        if (IS_ERR(new_page)) {
@@ -2086,27 +2407,54 @@ retry_avoidcopy:
                        if (unmap_ref_private(mm, vma, old_page, address)) {
                                BUG_ON(page_count(old_page) != 1);
                                BUG_ON(huge_pte_none(pte));
+                               spin_lock(&mm->page_table_lock);
                                goto retry_avoidcopy;
                        }
                        WARN_ON_ONCE(1);
                }
 
+               /* Caller expects lock to be held */
+               spin_lock(&mm->page_table_lock);
                return -PTR_ERR(new_page);
        }
 
-       spin_unlock(&mm->page_table_lock);
-       copy_huge_page(new_page, old_page, address, vma);
+       /*
+        * When the original hugepage is shared one, it does not have
+        * anon_vma prepared.
+        */
+       if (unlikely(anon_vma_prepare(vma))) {
+               page_cache_release(new_page);
+               page_cache_release(old_page);
+               /* Caller expects lock to be held */
+               spin_lock(&mm->page_table_lock);
+               return VM_FAULT_OOM;
+       }
+
+       copy_user_huge_page(new_page, old_page, address, vma,
+                           pages_per_huge_page(h));
        __SetPageUptodate(new_page);
-       spin_lock(&mm->page_table_lock);
 
+       /*
+        * Retake the page_table_lock to check for racing updates
+        * before the page tables are altered
+        */
+       spin_lock(&mm->page_table_lock);
        ptep = huge_pte_offset(mm, address & huge_page_mask(h));
        if (likely(pte_same(huge_ptep_get(ptep), pte))) {
                /* Break COW */
+               mmu_notifier_invalidate_range_start(mm,
+                       address & huge_page_mask(h),
+                       (address & huge_page_mask(h)) + huge_page_size(h));
                huge_ptep_clear_flush(vma, address, ptep);
                set_huge_pte_at(mm, address, ptep,
                                make_huge_pte(vma, new_page, 1));
+               page_remove_rmap(old_page);
+               hugepage_add_new_anon_rmap(new_page, vma, address);
                /* Make the old page be freed below */
                new_page = old_page;
+               mmu_notifier_invalidate_range_end(mm,
+                       address & huge_page_mask(h),
+                       (address & huge_page_mask(h)) + huge_page_size(h));
        }
        page_cache_release(new_page);
        page_cache_release(old_page);
@@ -2160,7 +2508,7 @@ static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
        /*
         * Currently, we are forced to kill the process in the event the
         * original mapper has unmapped pages from the child due to a failed
-        * COW. Warn that such a situation has occured as it may not be obvious
+        * COW. Warn that such a situation has occurred as it may not be obvious
         */
        if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
                printk(KERN_WARNING
@@ -2187,7 +2535,7 @@ retry:
                        ret = -PTR_ERR(page);
                        goto out;
                }
-               clear_huge_page(page, address, huge_page_size(h));
+               clear_huge_page(page, address, pages_per_huge_page(h));
                __SetPageUptodate(page);
 
                if (vma->vm_flags & VM_MAYSHARE) {
@@ -2205,8 +2553,27 @@ retry:
                        spin_lock(&inode->i_lock);
                        inode->i_blocks += blocks_per_huge_page(h);
                        spin_unlock(&inode->i_lock);
-               } else
+                       page_dup_rmap(page);
+               } else {
                        lock_page(page);
+                       if (unlikely(anon_vma_prepare(vma))) {
+                               ret = VM_FAULT_OOM;
+                               goto backout_unlocked;
+                       }
+                       hugepage_add_new_anon_rmap(page, vma, address);
+               }
+       } else {
+               /*
+                * If memory error occurs between mmap() and fault, some process
+                * don't have hwpoisoned swap entry for errored virtual address.
+                * So we need to block hugepage fault by PG_hwpoison bit check.
+                */
+               if (unlikely(PageHWPoison(page))) {
+                       ret = VM_FAULT_HWPOISON |
+                             VM_FAULT_SET_HINDEX(h - hstates);
+                       goto backout_unlocked;
+               }
+               page_dup_rmap(page);
        }
 
        /*
@@ -2258,10 +2625,22 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        pte_t *ptep;
        pte_t entry;
        int ret;
+       struct page *page = NULL;
        struct page *pagecache_page = NULL;
        static DEFINE_MUTEX(hugetlb_instantiation_mutex);
        struct hstate *h = hstate_vma(vma);
 
+       ptep = huge_pte_offset(mm, address);
+       if (ptep) {
+               entry = huge_ptep_get(ptep);
+               if (unlikely(is_hugetlb_entry_migration(entry))) {
+                       migration_entry_wait(mm, (pmd_t *)ptep, address);
+                       return 0;
+               } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
+                       return VM_FAULT_HWPOISON_LARGE |
+                              VM_FAULT_SET_HINDEX(h - hstates);
+       }
+
        ptep = huge_pte_alloc(mm, address, huge_page_size(h));
        if (!ptep)
                return VM_FAULT_OOM;
@@ -2299,6 +2678,17 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
                                                                vma, address);
        }
 
+       /*
+        * hugetlb_cow() requires page locks of pte_page(entry) and
+        * pagecache_page, so here we need take the former one
+        * when page != pagecache_page or !pagecache_page.
+        * Note that locking order is always pagecache_page -> page,
+        * so no worry about deadlock.
+        */
+       page = pte_page(entry);
+       if (page != pagecache_page)
+               lock_page(page);
+
        spin_lock(&mm->page_table_lock);
        /* Check for a racing update before calling hugetlb_cow */
        if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
@@ -2316,7 +2706,7 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        entry = pte_mkyoung(entry);
        if (huge_ptep_set_access_flags(vma, address, ptep, entry,
                                                flags & FAULT_FLAG_WRITE))
-               update_mmu_cache(vma, address, entry);
+               update_mmu_cache(vma, address, ptep);
 
 out_page_table_lock:
        spin_unlock(&mm->page_table_lock);
@@ -2325,6 +2715,8 @@ out_page_table_lock:
                unlock_page(pagecache_page);
                put_page(pagecache_page);
        }
+       if (page != pagecache_page)
+               unlock_page(page);
 
 out_mutex:
        mutex_unlock(&hugetlb_instantiation_mutex);
@@ -2436,7 +2828,7 @@ void hugetlb_change_protection(struct vm_area_struct *vma,
        BUG_ON(address >= end);
        flush_cache_range(vma, address, end);
 
-       spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
+       mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
        spin_lock(&mm->page_table_lock);
        for (; address < end; address += huge_page_size(h)) {
                ptep = huge_pte_offset(mm, address);
@@ -2451,7 +2843,7 @@ void hugetlb_change_protection(struct vm_area_struct *vma,
                }
        }
        spin_unlock(&mm->page_table_lock);
-       spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
+       mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
 
        flush_tlb_range(vma, start, end);
 }
@@ -2459,7 +2851,7 @@ void hugetlb_change_protection(struct vm_area_struct *vma,
 int hugetlb_reserve_pages(struct inode *inode,
                                        long from, long to,
                                        struct vm_area_struct *vma,
-                                       int acctflag)
+                                       vm_flags_t vm_flags)
 {
        long ret, chg;
        struct hstate *h = hstate_inode(inode);
@@ -2469,7 +2861,7 @@ int hugetlb_reserve_pages(struct inode *inode,
         * attempt will be made for VM_NORESERVE to allocate a page
         * and filesystem quota without using reserves
         */
-       if (acctflag & VM_NORESERVE)
+       if (vm_flags & VM_NORESERVE)
                return 0;
 
        /*
@@ -2536,3 +2928,42 @@ void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
        hugetlb_put_quota(inode->i_mapping, (chg - freed));
        hugetlb_acct_memory(h, -(chg - freed));
 }
+
+#ifdef CONFIG_MEMORY_FAILURE
+
+/* Should be called in hugetlb_lock */
+static int is_hugepage_on_freelist(struct page *hpage)
+{
+       struct page *page;
+       struct page *tmp;
+       struct hstate *h = page_hstate(hpage);
+       int nid = page_to_nid(hpage);
+
+       list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
+               if (page == hpage)
+                       return 1;
+       return 0;
+}
+
+/*
+ * This function is called from memory failure code.
+ * Assume the caller holds page lock of the head page.
+ */
+int dequeue_hwpoisoned_huge_page(struct page *hpage)
+{
+       struct hstate *h = page_hstate(hpage);
+       int nid = page_to_nid(hpage);
+       int ret = -EBUSY;
+
+       spin_lock(&hugetlb_lock);
+       if (is_hugepage_on_freelist(hpage)) {
+               list_del(&hpage->lru);
+               set_page_refcounted(hpage);
+               h->free_huge_pages--;
+               h->free_huge_pages_node[nid]--;
+               ret = 0;
+       }
+       spin_unlock(&hugetlb_lock);
+       return ret;
+}
+#endif