regmap: mmio: convert some error returns to BUG()
[linux-2.6.git] / mm / hugetlb.c
index 4c97c17..b8ce6f4 100644 (file)
@@ -2,11 +2,11 @@
  * 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/mm.h>
+#include <linux/seq_file.h>
 #include <linux/sysctl.h>
 #include <linux/highmem.h>
 #include <linux/mmu_notifier.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;
@@ -48,6 +53,84 @@ static unsigned long __initdata default_hstate_size;
  */
 static DEFINE_SPINLOCK(hugetlb_lock);
 
+static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
+{
+       bool free = (spool->count == 0) && (spool->used_hpages == 0);
+
+       spin_unlock(&spool->lock);
+
+       /* If no pages are used, and no other handles to the subpool
+        * remain, free the subpool the subpool remain */
+       if (free)
+               kfree(spool);
+}
+
+struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
+{
+       struct hugepage_subpool *spool;
+
+       spool = kmalloc(sizeof(*spool), GFP_KERNEL);
+       if (!spool)
+               return NULL;
+
+       spin_lock_init(&spool->lock);
+       spool->count = 1;
+       spool->max_hpages = nr_blocks;
+       spool->used_hpages = 0;
+
+       return spool;
+}
+
+void hugepage_put_subpool(struct hugepage_subpool *spool)
+{
+       spin_lock(&spool->lock);
+       BUG_ON(!spool->count);
+       spool->count--;
+       unlock_or_release_subpool(spool);
+}
+
+static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
+                                     long delta)
+{
+       int ret = 0;
+
+       if (!spool)
+               return 0;
+
+       spin_lock(&spool->lock);
+       if ((spool->used_hpages + delta) <= spool->max_hpages) {
+               spool->used_hpages += delta;
+       } else {
+               ret = -ENOMEM;
+       }
+       spin_unlock(&spool->lock);
+
+       return ret;
+}
+
+static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
+                                      long delta)
+{
+       if (!spool)
+               return;
+
+       spin_lock(&spool->lock);
+       spool->used_hpages -= delta;
+       /* If hugetlbfs_put_super couldn't free spool due to
+       * an outstanding quota reference, free it now. */
+       unlock_or_release_subpool(spool);
+}
+
+static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
+{
+       return HUGETLBFS_SB(inode->i_sb)->spool;
+}
+
+static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
+{
+       return subpool_inode(vma->vm_file->f_dentry->d_inode);
+}
+
 /*
  * Region tracking -- allows tracking of reservations and instantiated pages
  *                    across the pages in a mapping.
@@ -57,10 +140,10 @@ 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;
@@ -141,7 +224,7 @@ 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) {
@@ -218,6 +301,42 @@ 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.
+ */
+unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
+{
+       struct hstate *hstate;
+
+       if (!is_vm_hugetlb_page(vma))
+               return PAGE_SIZE;
+
+       hstate = hstate_vma(vma);
+
+       return 1UL << (hstate->order + PAGE_SHIFT);
+}
+EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
+
+/*
+ * Return the page size being used by the MMU to back a VMA. In the majority
+ * of cases, the page size used by the kernel matches the MMU size. On
+ * architectures where it differs, an architecture-specific version of this
+ * function is required.
+ */
+#ifndef vma_mmu_pagesize
+unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
+{
+       return vma_kernel_pagesize(vma);
+}
+#endif
+
 /*
  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
  * bits of the reservation map pointer, which are always clear due to
@@ -262,7 +381,7 @@ struct resv_map {
        struct list_head regions;
 };
 
-struct resv_map *resv_map_alloc(void)
+static struct resv_map *resv_map_alloc(void)
 {
        struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
        if (!resv_map)
@@ -274,7 +393,7 @@ struct resv_map *resv_map_alloc(void)
        return resv_map;
 }
 
-void resv_map_release(struct kref *ref)
+static void resv_map_release(struct kref *ref)
 {
        struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
 
@@ -286,16 +405,16 @@ void resv_map_release(struct kref *ref)
 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
 {
        VM_BUG_ON(!is_vm_hugetlb_page(vma));
-       if (!(vma->vm_flags & VM_SHARED))
+       if (!(vma->vm_flags & VM_MAYSHARE))
                return (struct resv_map *)(get_vma_private_data(vma) &
                                                        ~HPAGE_RESV_MASK);
-       return 0;
+       return NULL;
 }
 
 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
 {
        VM_BUG_ON(!is_vm_hugetlb_page(vma));
-       VM_BUG_ON(vma->vm_flags & VM_SHARED);
+       VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 
        set_vma_private_data(vma, (get_vma_private_data(vma) &
                                HPAGE_RESV_MASK) | (unsigned long)map);
@@ -304,7 +423,7 @@ static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
 {
        VM_BUG_ON(!is_vm_hugetlb_page(vma));
-       VM_BUG_ON(vma->vm_flags & VM_SHARED);
+       VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 
        set_vma_private_data(vma, get_vma_private_data(vma) | flags);
 }
@@ -323,7 +442,7 @@ static void decrement_hugepage_resv_vma(struct hstate *h,
        if (vma->vm_flags & VM_NORESERVE)
                return;
 
-       if (vma->vm_flags & VM_SHARED) {
+       if (vma->vm_flags & VM_MAYSHARE) {
                /* Shared mappings always use reserves */
                h->resv_huge_pages--;
        } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
@@ -339,42 +458,51 @@ static void decrement_hugepage_resv_vma(struct hstate *h,
 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
 {
        VM_BUG_ON(!is_vm_hugetlb_page(vma));
-       if (!(vma->vm_flags & VM_SHARED))
+       if (!(vma->vm_flags & VM_MAYSHARE))
                vma->vm_private_data = (void *)0;
 }
 
 /* Returns true if the VMA has associated reserve pages */
 static int vma_has_reserves(struct vm_area_struct *vma)
 {
-       if (vma->vm_flags & VM_SHARED)
+       if (vma->vm_flags & VM_MAYSHARE)
                return 1;
        if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
                return 1;
        return 0;
 }
 
-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;
+       struct hstate *h = page_hstate(src);
+       struct page *dst_base = dst;
+       struct page *src_base = src;
 
-       might_sleep();
-       for (i = 0; i < sz/PAGE_SIZE; i++) {
+       for (i = 0; i < pages_per_huge_page(h); ) {
                cond_resched();
-               clear_user_highpage(page + i, addr + i * PAGE_SIZE);
+               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);
+               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);
        }
 }
 
@@ -386,21 +514,17 @@ static void enqueue_huge_page(struct hstate *h, struct page *page)
        h->free_huge_pages_node[nid]++;
 }
 
-static struct page *dequeue_huge_page(struct hstate *h)
+static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
 {
-       int nid;
-       struct page *page = NULL;
+       struct page *page;
 
-       for (nid = 0; nid < MAX_NUMNODES; ++nid) {
-               if (!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]--;
-                       break;
-               }
-       }
+       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;
 }
 
@@ -408,15 +532,18 @@ 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 page *page;
        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;
+       unsigned int cpuset_mems_cookie;
 
+retry_cpuset:
+       cpuset_mems_cookie = 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
@@ -424,43 +551,47 @@ 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;
+                       }
                }
        }
+
        mpol_cond_put(mpol);
+       if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
+               goto retry_cpuset;
        return page;
+
+err:
+       mpol_cond_put(mpol);
+       return NULL;
 }
 
 static void update_and_free_page(struct hstate *h, struct page *page)
 {
        int i;
 
+       VM_BUG_ON(h->order >= MAX_ORDER);
+
        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);
@@ -487,11 +618,13 @@ static void free_huge_page(struct page *page)
         */
        struct hstate *h = page_hstate(page);
        int nid = page_to_nid(page);
-       struct address_space *mapping;
+       struct hugepage_subpool *spool =
+               (struct hugepage_subpool *)page_private(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);
@@ -503,43 +636,7 @@ static void free_huge_page(struct page *page)
                enqueue_huge_page(h, page);
        }
        spin_unlock(&hugetlb_lock);
-       if (mapping)
-               hugetlb_put_quota(mapping, 1);
-}
-
-/*
- * Increment or decrement surplus_huge_pages.  Keep node-specific counters
- * balanced by operating on them in a round-robin fashion.
- * Returns 1 if an adjustment was made.
- */
-static int adjust_pool_surplus(struct hstate *h, int delta)
-{
-       static int prev_nid;
-       int nid = prev_nid;
-       int ret = 0;
-
-       VM_BUG_ON(delta != -1 && delta != 1);
-       do {
-               nid = next_node(nid, node_online_map);
-               if (nid == MAX_NUMNODES)
-                       nid = first_node(node_online_map);
-
-               /* To shrink on this node, there must be a surplus page */
-               if (delta < 0 && !h->surplus_huge_pages_node[nid])
-                       continue;
-               /* Surplus cannot exceed the total number of pages */
-               if (delta > 0 && h->surplus_huge_pages_node[nid] >=
-                                               h->nr_huge_pages_node[nid])
-                       continue;
-
-               h->surplus_huge_pages += delta;
-               h->surplus_huge_pages_node[nid] += delta;
-               ret = 1;
-               break;
-       } while (nid != prev_nid);
-
-       prev_nid = nid;
-       return ret;
+       hugepage_subpool_put_pages(spool, 1);
 }
 
 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
@@ -552,6 +649,36 @@ static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
        put_page(page); /* free it into the hugepage allocator */
 }
 
+static void prep_compound_gigantic_page(struct page *page, unsigned long order)
+{
+       int i;
+       int nr_pages = 1 << order;
+       struct page *p = page + 1;
+
+       /* we rely on prep_new_huge_page to set the destructor */
+       set_compound_order(page, 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;
+       }
+}
+
+int PageHuge(struct page *page)
+{
+       compound_page_dtor *dtor;
+
+       if (!PageCompound(page))
+               return 0;
+
+       page = compound_head(page);
+       dtor = get_compound_page_dtor(page);
+
+       return dtor == free_huge_page;
+}
+EXPORT_SYMBOL_GPL(PageHuge);
+
 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
 {
        struct page *page;
@@ -559,7 +686,7 @@ static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
        if (h->order >= MAX_ORDER)
                return NULL;
 
-       page = alloc_pages_node(nid,
+       page = alloc_pages_exact_node(nid,
                htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
                                                __GFP_REPEAT|__GFP_NOWARN,
                huge_page_order(h));
@@ -575,41 +702,66 @@ static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
 }
 
 /*
- * Use a helper variable to find the next node and then
- * copy it back to hugetlb_next_nid afterwards:
- * otherwise there's a window in which a racer might
- * pass invalid nid MAX_NUMNODES to alloc_pages_node.
- * But we don't need to use a spin_lock here: it really
- * doesn't matter if occasionally a racer chooses the
- * same nid as we do.  Move nid forward in the mask even
- * if we just successfully allocated a hugepage so that
- * the next caller gets hugepages on the next node.
+ * common helper functions for hstate_next_node_to_{alloc|free}.
+ * We may have allocated or freed a huge page based on a different
+ * nodes_allowed previously, so h->next_node_to_{alloc|free} might
+ * be outside of *nodes_allowed.  Ensure that we use an allowed
+ * node for alloc or free.
  */
-static int hstate_next_node(struct hstate *h)
+static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
 {
-       int next_nid;
-       next_nid = next_node(h->hugetlb_next_nid, node_online_map);
-       if (next_nid == MAX_NUMNODES)
-               next_nid = first_node(node_online_map);
-       h->hugetlb_next_nid = next_nid;
-       return next_nid;
+       nid = next_node(nid, *nodes_allowed);
+       if (nid == MAX_NUMNODES)
+               nid = first_node(*nodes_allowed);
+       VM_BUG_ON(nid >= MAX_NUMNODES);
+
+       return nid;
+}
+
+static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
+{
+       if (!node_isset(nid, *nodes_allowed))
+               nid = next_node_allowed(nid, nodes_allowed);
+       return nid;
+}
+
+/*
+ * returns the previously saved node ["this node"] from which to
+ * allocate a persistent huge page for the pool and advance the
+ * next node from which to allocate, handling wrap at end of node
+ * mask.
+ */
+static int hstate_next_node_to_alloc(struct hstate *h,
+                                       nodemask_t *nodes_allowed)
+{
+       int nid;
+
+       VM_BUG_ON(!nodes_allowed);
+
+       nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
+       h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
+
+       return nid;
 }
 
-static int alloc_fresh_huge_page(struct hstate *h)
+static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
 {
        struct page *page;
        int start_nid;
        int next_nid;
        int ret = 0;
 
-       start_nid = h->hugetlb_next_nid;
+       start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
+       next_nid = start_nid;
 
        do {
-               page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
-               if (page)
+               page = alloc_fresh_huge_page_node(h, next_nid);
+               if (page) {
                        ret = 1;
-               next_nid = hstate_next_node(h);
-       } while (!page && h->hugetlb_next_nid != start_nid);
+                       break;
+               }
+               next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
+       } while (next_nid != start_nid);
 
        if (ret)
                count_vm_event(HTLB_BUDDY_PGALLOC);
@@ -619,11 +771,71 @@ static int alloc_fresh_huge_page(struct hstate *h)
        return ret;
 }
 
-static struct page *alloc_buddy_huge_page(struct hstate *h,
-                       struct vm_area_struct *vma, unsigned long address)
+/*
+ * helper for free_pool_huge_page() - return the previously saved
+ * node ["this node"] from which to free a huge page.  Advance the
+ * next node id whether or not we find a free huge page to free so
+ * that the next attempt to free addresses the next node.
+ */
+static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
+{
+       int nid;
+
+       VM_BUG_ON(!nodes_allowed);
+
+       nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
+       h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
+
+       return nid;
+}
+
+/*
+ * Free huge page from pool from next node to free.
+ * Attempt to keep persistent huge pages more or less
+ * balanced over allowed nodes.
+ * Called with hugetlb_lock locked.
+ */
+static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
+                                                        bool acct_surplus)
+{
+       int start_nid;
+       int next_nid;
+       int ret = 0;
+
+       start_nid = hstate_next_node_to_free(h, nodes_allowed);
+       next_nid = start_nid;
+
+       do {
+               /*
+                * If we're returning unused surplus pages, only examine
+                * nodes with surplus pages.
+                */
+               if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
+                   !list_empty(&h->hugepage_freelists[next_nid])) {
+                       struct page *page =
+                               list_entry(h->hugepage_freelists[next_nid].next,
+                                         struct page, lru);
+                       list_del(&page->lru);
+                       h->free_huge_pages--;
+                       h->free_huge_pages_node[next_nid]--;
+                       if (acct_surplus) {
+                               h->surplus_huge_pages--;
+                               h->surplus_huge_pages_node[next_nid]--;
+                       }
+                       update_and_free_page(h, page);
+                       ret = 1;
+                       break;
+               }
+               next_nid = hstate_next_node_to_free(h, nodes_allowed);
+       } while (next_nid != start_nid);
+
+       return ret;
+}
+
+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;
@@ -661,30 +873,29 @@ 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));
-               return NULL;
+               page = NULL;
        }
 
        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--;
@@ -697,7 +908,26 @@ 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)
@@ -706,6 +936,7 @@ static int gather_surplus_pages(struct hstate *h, int delta)
        struct page *page, *tmp;
        int ret, i;
        int needed, allocated;
+       bool alloc_ok = true;
 
        needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
        if (needed <= 0) {
@@ -720,21 +951,14 @@ 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);
+               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;
+                       alloc_ok = false;
+                       break;
                }
-
                list_add(&page->lru, &surplus_list);
        }
-       allocated += needed;
+       allocated += i;
 
        /*
         * After retaking hugetlb_lock, we need to recalculate 'needed'
@@ -743,12 +967,19 @@ retry:
        spin_lock(&hugetlb_lock);
        needed = (h->resv_huge_pages + delta) -
                        (h->free_huge_pages + allocated);
-       if (needed > 0)
-               goto retry;
-
+       if (needed > 0) {
+               if (alloc_ok)
+                       goto retry;
+               /*
+                * We were not able to allocate enough pages to
+                * satisfy the entire reservation so we free what
+                * we've allocated so far.
+                */
+               goto free;
+       }
        /*
         * 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
@@ -757,31 +988,31 @@ retry:
        needed += allocated;
        h->resv_huge_pages += delta;
        ret = 0;
-free:
+
        /* 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:
+       spin_unlock(&hugetlb_lock);
 
        /* Free unnecessary surplus pages to the buddy allocator */
        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;
 }
@@ -790,22 +1021,13 @@ free:
  * When releasing a hugetlb pool reservation, any surplus pages that were
  * allocated to satisfy the reservation must be explicitly freed if they were
  * never used.
+ * Called with hugetlb_lock held.
  */
 static void return_unused_surplus_pages(struct hstate *h,
                                        unsigned long unused_resv_pages)
 {
-       static int nid = -1;
-       struct page *page;
        unsigned long nr_pages;
 
-       /*
-        * 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.
-        */
-       unsigned long remaining_iterations = num_online_nodes();
-
        /* Uncommit the reservation */
        h->resv_huge_pages -= unused_resv_pages;
 
@@ -815,45 +1037,37 @@ static void return_unused_surplus_pages(struct hstate *h,
 
        nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
 
-       while (remaining_iterations-- && nr_pages) {
-               nid = next_node(nid, node_online_map);
-               if (nid == MAX_NUMNODES)
-                       nid = first_node(node_online_map);
-
-               if (!h->surplus_huge_pages_node[nid])
-                       continue;
-
-               if (!list_empty(&h->hugepage_freelists[nid])) {
-                       page = list_entry(h->hugepage_freelists[nid].next,
-                                         struct page, lru);
-                       list_del(&page->lru);
-                       update_and_free_page(h, page);
-                       h->free_huge_pages--;
-                       h->free_huge_pages_node[nid]--;
-                       h->surplus_huge_pages--;
-                       h->surplus_huge_pages_node[nid]--;
-                       nr_pages--;
-                       remaining_iterations = num_online_nodes();
-               }
+       /*
+        * We want to release as many surplus pages as possible, spread
+        * 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_states[N_HIGH_MEMORY], 1))
+                       break;
        }
 }
 
 /*
  * Determine if the huge page at addr within the vma has an associated
  * reservation.  Where it does not we will need to logically increase
- * reservation and actually increase quota before an allocation can occur.
- * Where any new reservation would be required the reservation change is
- * prepared, but not committed.  Once the page has been quota'd allocated
- * an instantiated the change should be committed via vma_commit_reservation.
- * No action is required on failure.
+ * reservation and actually increase subpool usage before an allocation
+ * can occur.  Where any new reservation would be required the
+ * reservation change is prepared, but not committed.  Once the page
+ * has been allocated from the subpool and instantiated the change should
+ * be committed via vma_commit_reservation.  No action is required on
+ * failure.
  */
-static int vma_needs_reservation(struct hstate *h,
+static long vma_needs_reservation(struct hstate *h,
                        struct vm_area_struct *vma, unsigned long addr)
 {
        struct address_space *mapping = vma->vm_file->f_mapping;
        struct inode *inode = mapping->host;
 
-       if (vma->vm_flags & VM_SHARED) {
+       if (vma->vm_flags & VM_MAYSHARE) {
                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
                return region_chg(&inode->i_mapping->private_list,
                                                        idx, idx + 1);
@@ -862,7 +1076,7 @@ static int vma_needs_reservation(struct hstate *h,
                return 1;
 
        } else  {
-               int err;
+               long err;
                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
                struct resv_map *reservations = vma_resv_map(vma);
 
@@ -878,7 +1092,7 @@ static void vma_commit_reservation(struct hstate *h,
        struct address_space *mapping = vma->vm_file->f_mapping;
        struct inode *inode = mapping->host;
 
-       if (vma->vm_flags & VM_SHARED) {
+       if (vma->vm_flags & VM_MAYSHARE) {
                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
                region_add(&inode->i_mapping->private_list, idx, idx + 1);
 
@@ -894,56 +1108,56 @@ static void vma_commit_reservation(struct hstate *h,
 static struct page *alloc_huge_page(struct vm_area_struct *vma,
                                    unsigned long addr, int avoid_reserve)
 {
+       struct hugepage_subpool *spool = subpool_vma(vma);
        struct hstate *h = hstate_vma(vma);
        struct page *page;
-       struct address_space *mapping = vma->vm_file->f_mapping;
-       struct inode *inode = mapping->host;
-       unsigned int chg;
+       long chg;
 
        /*
-        * Processes that did not create the mapping will have no reserves and
-        * will not have accounted against quota. Check that the quota can be
-        * made before satisfying the allocation
-        * MAP_NORESERVE mappings may also need pages and quota allocated
-        * if no reserve mapping overlaps.
+        * Processes that did not create the mapping will have no
+        * reserves and will not have accounted against subpool
+        * limit. Check that the subpool limit can be made before
+        * satisfying the allocation MAP_NORESERVE mappings may also
+        * need pages and subpool limit allocated allocated if no reserve
+        * mapping overlaps.
         */
        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);
+               if (hugepage_subpool_get_pages(spool, chg))
+                       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);
+                       hugepage_subpool_put_pages(spool, chg);
+                       return ERR_PTR(-VM_FAULT_SIGBUS);
                }
        }
 
-       set_page_refcounted(page);
-       set_page_private(page, (unsigned long) mapping);
+       set_page_private(page, (unsigned long)spool);
 
        vma_commit_reservation(h, vma, addr);
 
        return page;
 }
 
-__attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h)
+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(h->hugetlb_next_nid),
+                               NODE_DATA(hstate_next_node_to_alloc(h,
+                                               &node_states[N_HIGH_MEMORY])),
                                huge_page_size(h), huge_page_size(h), 0);
 
                if (addr) {
@@ -953,10 +1167,8 @@ __attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h)
                         * puts them into the mem_map).
                         */
                        m = addr;
-                       if (m)
-                               goto found;
+                       goto found;
                }
-               hstate_next_node(h);
                nr_nodes--;
        }
        return 0;
@@ -969,18 +1181,42 @@ found:
        return 1;
 }
 
+static void prep_compound_huge_page(struct page *page, int order)
+{
+       if (unlikely(order > (MAX_ORDER - 1)))
+               prep_compound_gigantic_page(page, order);
+       else
+               prep_compound_page(page, order);
+}
+
 /* Put bootmem huge pages into the standard lists after mem_map is up */
 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_page(page, h->order);
+               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;
        }
 }
 
@@ -992,7 +1228,8 @@ 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))
+               } else if (!alloc_fresh_huge_page(h,
+                                        &node_states[N_HIGH_MEMORY]))
                        break;
        }
        h->max_huge_pages = i;
@@ -1034,14 +1271,15 @@ static void __init report_hugepages(void)
 }
 
 #ifdef CONFIG_HIGHMEM
-static void try_to_free_low(struct hstate *h, unsigned long count)
+static void try_to_free_low(struct hstate *h, unsigned long count,
+                                               nodemask_t *nodes_allowed)
 {
        int i;
 
        if (h->order >= MAX_ORDER)
                return;
 
-       for (i = 0; i < MAX_NUMNODES; ++i) {
+       for_each_node_mask(i, *nodes_allowed) {
                struct page *page, *next;
                struct list_head *freel = &h->hugepage_freelists[i];
                list_for_each_entry_safe(page, next, freel, lru) {
@@ -1057,13 +1295,67 @@ static void try_to_free_low(struct hstate *h, unsigned long count)
        }
 }
 #else
-static inline void try_to_free_low(struct hstate *h, unsigned long count)
+static inline void try_to_free_low(struct hstate *h, unsigned long count,
+                                               nodemask_t *nodes_allowed)
 {
 }
 #endif
 
+/*
+ * Increment or decrement surplus_huge_pages.  Keep node-specific counters
+ * balanced by operating on them in a round-robin fashion.
+ * Returns 1 if an adjustment was made.
+ */
+static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
+                               int delta)
+{
+       int start_nid, next_nid;
+       int ret = 0;
+
+       VM_BUG_ON(delta != -1 && delta != 1);
+
+       if (delta < 0)
+               start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
+       else
+               start_nid = hstate_next_node_to_free(h, nodes_allowed);
+       next_nid = start_nid;
+
+       do {
+               int nid = next_nid;
+               if (delta < 0)  {
+                       /*
+                        * To shrink on this node, there must be a surplus page
+                        */
+                       if (!h->surplus_huge_pages_node[nid]) {
+                               next_nid = hstate_next_node_to_alloc(h,
+                                                               nodes_allowed);
+                               continue;
+                       }
+               }
+               if (delta > 0) {
+                       /*
+                        * Surplus cannot exceed the total number of pages
+                        */
+                       if (h->surplus_huge_pages_node[nid] >=
+                                               h->nr_huge_pages_node[nid]) {
+                               next_nid = hstate_next_node_to_free(h,
+                                                               nodes_allowed);
+                               continue;
+                       }
+               }
+
+               h->surplus_huge_pages += delta;
+               h->surplus_huge_pages_node[nid] += delta;
+               ret = 1;
+               break;
+       } while (next_nid != start_nid);
+
+       return ret;
+}
+
 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
-static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
+static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
+                                               nodemask_t *nodes_allowed)
 {
        unsigned long min_count, ret;
 
@@ -1083,7 +1375,7 @@ static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
         */
        spin_lock(&hugetlb_lock);
        while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
-               if (!adjust_pool_surplus(h, -1))
+               if (!adjust_pool_surplus(h, nodes_allowed, -1))
                        break;
        }
 
@@ -1094,11 +1386,14 @@ static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
                 * and reducing the surplus.
                 */
                spin_unlock(&hugetlb_lock);
-               ret = alloc_fresh_huge_page(h);
+               ret = alloc_fresh_huge_page(h, nodes_allowed);
                spin_lock(&hugetlb_lock);
                if (!ret)
                        goto out;
 
+               /* Bail for signals. Probably ctrl-c from user */
+               if (signal_pending(current))
+                       goto out;
        }
 
        /*
@@ -1118,15 +1413,13 @@ static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
         */
        min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
        min_count = max(count, min_count);
-       try_to_free_low(h, min_count);
+       try_to_free_low(h, min_count, nodes_allowed);
        while (min_count < persistent_huge_pages(h)) {
-               struct page *page = dequeue_huge_page(h);
-               if (!page)
+               if (!free_pool_huge_page(h, nodes_allowed, 0))
                        break;
-               update_and_free_page(h, page);
        }
        while (count < persistent_huge_pages(h)) {
-               if (!adjust_pool_surplus(h, 1))
+               if (!adjust_pool_surplus(h, nodes_allowed, 1))
                        break;
        }
 out:
@@ -1145,55 +1438,142 @@ 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(struct kobject *kobj,
+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(struct kobject *kobj,
-               struct kobj_attribute *attr, const char *buf, size_t count)
+
+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;
-       unsigned long input;
-       struct hstate *h = kobj_to_hstate(kobj);
+       int nid;
+       unsigned long count;
+       struct hstate *h;
+       NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
 
-       err = strict_strtoul(buf, 10, &input);
+       err = strict_strtoul(buf, 10, &count);
        if (err)
-               return 0;
+               goto out;
 
-       h->max_huge_pages = set_max_huge_pages(h, input);
+       h = kobj_to_hstate(kobj, &nid);
+       if (h->order >= MAX_ORDER) {
+               err = -EINVAL;
+               goto out;
+       }
 
-       return count;
+       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_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,
+                                      struct kobj_attribute *attr, char *buf)
+{
+       return nr_hugepages_show_common(kobj, attr, buf);
+}
+
+static ssize_t nr_hugepages_store(struct kobject *kobj,
+              struct kobj_attribute *attr, const char *buf, size_t len)
+{
+       return nr_hugepages_store_common(false, kobj, attr, buf, len);
 }
 HSTATE_ATTR(nr_hugepages);
 
+#ifdef CONFIG_NUMA
+
+/*
+ * hstate attribute for optionally mempolicy-based constraint on persistent
+ * huge page alloc/free.
+ */
+static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
+                                      struct kobj_attribute *attr, char *buf)
+{
+       return nr_hugepages_show_common(kobj, attr, buf);
+}
+
+static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
+              struct kobj_attribute *attr, const char *buf, size_t len)
+{
+       return nr_hugepages_store_common(true, kobj, attr, buf, len);
+}
+HSTATE_ATTR(nr_hugepages_mempolicy);
+#endif
+
+
 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;
@@ -1206,15 +1586,24 @@ 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);
@@ -1222,8 +1611,17 @@ 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);
 
@@ -1233,6 +1631,9 @@ static struct attribute *hstate_attrs[] = {
        &free_hugepages_attr.attr,
        &resv_hugepages_attr.attr,
        &surplus_hugepages_attr.attr,
+#ifdef CONFIG_NUMA
+       &nr_hugepages_mempolicy_attr.attr,
+#endif
        NULL,
 };
 
@@ -1240,19 +1641,20 @@ 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;
 }
@@ -1267,17 +1669,184 @@ 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 devices in node_devices[] using a parallel array.  The array
+ * index of a node device or _hstate == node id.
+ * This is here to avoid any static dependency of the node device 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 devices
+ */
+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 device 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 device.
+ * No-op if no hstate attributes attached.
+ */
+void hugetlb_unregister_node(struct node *node)
+{
+       struct hstate *h;
+       struct node_hstate *nhs = &node_hstates[node->dev.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 devices
+ * that have them.
+ */
+static void hugetlb_unregister_all_nodes(void)
+{
+       int nid;
+
+       /*
+        * disable node device 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 device.
+ * No-op if attributes already registered.
+ */
+void hugetlb_register_node(struct node *node)
+{
+       struct hstate *h;
+       struct node_hstate *nhs = &node_hstates[node->dev.id];
+       int err;
+
+       if (nhs->hugepages_kobj)
+               return;         /* already allocated */
+
+       nhs->hugepages_kobj = kobject_create_and_add("hugepages",
+                                                       &node->dev.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->dev.id);
+                       hugetlb_unregister_node(node);
+                       break;
+               }
+       }
+}
+
+/*
+ * hugetlb init time:  register hstate attributes for all registered node
+ * devices of nodes that have memory.  All on-line nodes should have
+ * registered their associated device 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->dev.id == nid)
+                       hugetlb_register_node(node);
+       }
+
+       /*
+        * Let the node device 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]);
        }
@@ -1312,6 +1881,8 @@ static int __init hugetlb_init(void)
 
        hugetlb_sysfs_init();
 
+       hugetlb_register_all_nodes();
+
        return 0;
 }
 module_init(hugetlb_init);
@@ -1335,7 +1906,8 @@ 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->hugetlb_next_nid = 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);
 
@@ -1398,31 +1970,64 @@ static unsigned int cpuset_mems_nr(unsigned int *array)
 }
 
 #ifdef CONFIG_SYSCTL
-int hugetlb_sysctl_handler(struct ctl_table *table, int write,
-                          struct file *file, void __user *buffer,
-                          size_t *length, loff_t *ppos)
+static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
+                        struct ctl_table *table, int write,
+                        void __user *buffer, size_t *length, loff_t *ppos)
 {
        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, file, buffer, length, ppos);
+       ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
+       if (ret)
+               goto out;
 
-       if (write)
-               h->max_huge_pages = set_max_huge_pages(h, tmp);
+       if (write) {
+               NODEMASK_ALLOC(nodemask_t, nodes_allowed,
+                                               GFP_KERNEL | __GFP_NORETRY);
+               if (!(obey_mempolicy &&
+                              init_nodemask_of_mempolicy(nodes_allowed))) {
+                       NODEMASK_FREE(nodes_allowed);
+                       nodes_allowed = &node_states[N_HIGH_MEMORY];
+               }
+               h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
 
-       return 0;
+               if (nodes_allowed != &node_states[N_HIGH_MEMORY])
+                       NODEMASK_FREE(nodes_allowed);
+       }
+out:
+       return ret;
 }
 
+int hugetlb_sysctl_handler(struct ctl_table *table, int write,
+                         void __user *buffer, size_t *length, loff_t *ppos)
+{
+
+       return hugetlb_sysctl_handler_common(false, table, write,
+                                                       buffer, length, ppos);
+}
+
+#ifdef CONFIG_NUMA
+int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
+                         void __user *buffer, size_t *length, loff_t *ppos)
+{
+       return hugetlb_sysctl_handler_common(true, table, write,
+                                                       buffer, length, ppos);
+}
+#endif /* CONFIG_NUMA */
+
 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
-                       struct file *file, void __user *buffer,
+                       void __user *buffer,
                        size_t *length, loff_t *ppos)
 {
-       proc_dointvec(table, write, file, buffer, length, ppos);
+       proc_dointvec(table, write, buffer, length, ppos);
        if (hugepages_treat_as_movable)
                htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
        else
@@ -1431,39 +2036,44 @@ int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
 }
 
 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
-                       struct file *file, void __user *buffer,
+                       void __user *buffer,
                        size_t *length, loff_t *ppos)
 {
        struct hstate *h = &default_hstate;
        unsigned long tmp;
+       int ret;
+
+       tmp = h->nr_overcommit_huge_pages;
 
-       if (!write)
-               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, file, 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 */
 
-int hugetlb_report_meminfo(char *buf)
+void hugetlb_report_meminfo(struct seq_file *m)
 {
        struct hstate *h = &default_hstate;
-       return sprintf(buf,
-                       "HugePages_Total: %5lu\n"
-                       "HugePages_Free:  %5lu\n"
-                       "HugePages_Rsvd:  %5lu\n"
-                       "HugePages_Surp:  %5lu\n"
-                       "Hugepagesize:    %5lu kB\n",
+       seq_printf(m,
+                       "HugePages_Total:   %5lu\n"
+                       "HugePages_Free:    %5lu\n"
+                       "HugePages_Rsvd:    %5lu\n"
+                       "HugePages_Surp:    %5lu\n"
+                       "Hugepagesize:   %8lu kB\n",
                        h->nr_huge_pages,
                        h->free_huge_pages,
                        h->resv_huge_pages,
@@ -1539,7 +2149,7 @@ 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.
         */
@@ -1551,6 +2161,7 @@ static void hugetlb_vm_op_close(struct vm_area_struct *vma)
 {
        struct hstate *h = hstate_vma(vma);
        struct resv_map *reservations = vma_resv_map(vma);
+       struct hugepage_subpool *spool = subpool_vma(vma);
        unsigned long reserve;
        unsigned long start;
        unsigned long end;
@@ -1566,7 +2177,7 @@ static void hugetlb_vm_op_close(struct vm_area_struct *vma)
 
                if (reserve) {
                        hugetlb_acct_memory(h, -reserve);
-                       hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
+                       hugepage_subpool_put_pages(spool, reserve);
                }
        }
 }
@@ -1583,7 +2194,7 @@ static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
        return 0;
 }
 
-struct vm_operations_struct hugetlb_vm_ops = {
+const struct vm_operations_struct hugetlb_vm_ops = {
        .fault = hugetlb_vm_op_fault,
        .open = hugetlb_vm_op_open,
        .close = hugetlb_vm_op_close,
@@ -1612,9 +2223,8 @@ 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);
 }
 
 
@@ -1650,6 +2260,7 @@ 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);
@@ -1661,6 +2272,32 @@ 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)
 {
@@ -1674,7 +2311,7 @@ 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.
         */
@@ -1694,16 +2331,23 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
                if (huge_pmd_unshare(mm, &address, ptep))
                        continue;
 
+               pte = huge_ptep_get(ptep);
+               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 a reference page is supplied, it is because a specific
                 * page is being unmapped, not a range. Ensure the page we
                 * are about to unmap is the actual page of interest.
                 */
                if (ref_page) {
-                       pte = huge_ptep_get(ptep);
-                       if (huge_pte_none(pte))
-                               continue;
-                       page = pte_page(pte);
                        if (page != ref_page)
                                continue;
 
@@ -1716,18 +2360,19 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
                }
 
                pte = huge_ptep_get_and_clear(mm, address, ptep);
-               if (huge_pte_none(pte))
-                       continue;
-
-               page = pte_page(pte);
                if (pte_dirty(pte))
                        set_page_dirty(page);
                list_add(&page->lru, &page_list);
+
+               /* Bail out after unmapping reference page if supplied */
+               if (ref_page)
+                       break;
        }
-       spin_unlock(&mm->page_table_lock);
        flush_tlb_range(vma, start, end);
+       spin_unlock(&mm->page_table_lock);
        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);
        }
@@ -1736,9 +2381,9 @@ 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);
 }
 
 /*
@@ -1747,11 +2392,10 @@ void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
  * from other VMAs and let the children be SIGKILLed if they are faulting the
  * same region.
  */
-int unmap_ref_private(struct mm_struct *mm,
-                                       struct vm_area_struct *vma,
-                                       struct page *page,
-                                       unsigned long address)
+static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
+                               struct page *page, unsigned long address)
 {
+       struct hstate *h = hstate_vma(vma);
        struct vm_area_struct *iter_vma;
        struct address_space *mapping;
        struct prio_tree_iter iter;
@@ -1761,11 +2405,16 @@ int unmap_ref_private(struct mm_struct *mm,
         * vm_pgoff is in PAGE_SIZE units, hence the different calculation
         * from page cache lookup which is in HPAGE_SIZE units.
         */
-       address = address & huge_page_mask(hstate_vma(vma));
-       pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
-               + (vma->vm_pgoff >> PAGE_SHIFT);
-       mapping = (struct address_space *)page_private(page);
+       address = address & huge_page_mask(h);
+       pgoff = vma_hugecache_offset(h, vma, address);
+       mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
 
+       /*
+        * 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)
@@ -1779,14 +2428,21 @@ int unmap_ref_private(struct mm_struct *mm,
                 * 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,
-                               address, address + HPAGE_SIZE,
+                       __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.
+ * Called with hugetlb_instantiation_mutex held and pte_page locked so we
+ * cannot race with other handlers or page migration.
+ * Keep the pte_same checks anyway to make transition from the mutex easier.
+ */
 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)
@@ -1801,8 +2457,10 @@ 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;
        }
@@ -1816,12 +2474,15 @@ retry_avoidcopy:
         * at the time of fork() could consume its reserves on COW instead
         * of the full address range.
         */
-       if (!(vma->vm_flags & VM_SHARED) &&
+       if (!(vma->vm_flags & VM_MAYSHARE) &&
                        is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
                        old_page != pagecache_page)
                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)) {
@@ -1839,27 +2500,61 @@ retry_avoidcopy:
                        if (unmap_ref_private(mm, vma, old_page, address)) {
                                BUG_ON(page_count(old_page) != 1);
                                BUG_ON(huge_pte_none(pte));
-                               goto retry_avoidcopy;
+                               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)))
+                                       goto retry_avoidcopy;
+                               /*
+                                * race occurs while re-acquiring page_table_lock, and
+                                * our job is done.
+                                */
+                               return 0;
                        }
                        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);
@@ -1879,11 +2574,32 @@ static struct page *hugetlbfs_pagecache_page(struct hstate *h,
        return find_lock_page(mapping, idx);
 }
 
+/*
+ * Return whether there is a pagecache page to back given address within VMA.
+ * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
+ */
+static bool hugetlbfs_pagecache_present(struct hstate *h,
+                       struct vm_area_struct *vma, unsigned long address)
+{
+       struct address_space *mapping;
+       pgoff_t idx;
+       struct page *page;
+
+       mapping = vma->vm_file->f_mapping;
+       idx = vma_hugecache_offset(h, vma, address);
+
+       page = find_get_page(mapping, idx);
+       if (page)
+               put_page(page);
+       return page != NULL;
+}
+
 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
-                       unsigned long address, pte_t *ptep, int write_access)
+                       unsigned long address, pte_t *ptep, unsigned int flags)
 {
        struct hstate *h = hstate_vma(vma);
        int ret = VM_FAULT_SIGBUS;
+       int anon_rmap = 0;
        pgoff_t idx;
        unsigned long size;
        struct page *page;
@@ -1893,7 +2609,7 @@ 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
@@ -1920,10 +2636,10 @@ 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_SHARED) {
+               if (vma->vm_flags & VM_MAYSHARE) {
                        int err;
                        struct inode *inode = mapping->host;
 
@@ -1938,8 +2654,25 @@ retry:
                        spin_lock(&inode->i_lock);
                        inode->i_blocks += blocks_per_huge_page(h);
                        spin_unlock(&inode->i_lock);
-               } else
+               } else {
                        lock_page(page);
+                       if (unlikely(anon_vma_prepare(vma))) {
+                               ret = VM_FAULT_OOM;
+                               goto backout_unlocked;
+                       }
+                       anon_rmap = 1;
+               }
+       } 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;
+               }
        }
 
        /*
@@ -1948,8 +2681,11 @@ retry:
         * any allocations necessary to record that reservation occur outside
         * the spinlock.
         */
-       if (write_access && !(vma->vm_flags & VM_SHARED))
-               vma_needs_reservation(h, vma, address);
+       if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
+               if (vma_needs_reservation(h, vma, address) < 0) {
+                       ret = VM_FAULT_OOM;
+                       goto backout_unlocked;
+               }
 
        spin_lock(&mm->page_table_lock);
        size = i_size_read(mapping->host) >> huge_page_shift(h);
@@ -1960,11 +2696,15 @@ retry:
        if (!huge_pte_none(huge_ptep_get(ptep)))
                goto backout;
 
+       if (anon_rmap)
+               hugepage_add_new_anon_rmap(page, vma, address);
+       else
+               page_dup_rmap(page);
        new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
                                && (vma->vm_flags & VM_SHARED)));
        set_huge_pte_at(mm, address, ptep, new_pte);
 
-       if (write_access && !(vma->vm_flags & VM_SHARED)) {
+       if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
                /* Optimization, do the COW without a second fault */
                ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
        }
@@ -1976,21 +2716,36 @@ out:
 
 backout:
        spin_unlock(&mm->page_table_lock);
+backout_unlocked:
        unlock_page(page);
        put_page(page);
        goto out;
 }
 
 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
-                       unsigned long address, int write_access)
+                       unsigned long address, unsigned int flags)
 {
        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);
 
+       address &= huge_page_mask(h);
+
+       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;
@@ -2003,9 +2758,8 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        mutex_lock(&hugetlb_instantiation_mutex);
        entry = huge_ptep_get(ptep);
        if (huge_pte_none(entry)) {
-               ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
-               mutex_unlock(&hugetlb_instantiation_mutex);
-               return ret;
+               ret = hugetlb_no_page(mm, vma, address, ptep, flags);
+               goto out_mutex;
        }
 
        ret = 0;
@@ -2018,27 +2772,58 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
         * page now as it is used to determine if a reservation has been
         * consumed.
         */
-       if (write_access && !pte_write(entry)) {
-               vma_needs_reservation(h, vma, address);
+       if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
+               if (vma_needs_reservation(h, vma, address) < 0) {
+                       ret = VM_FAULT_OOM;
+                       goto out_mutex;
+               }
 
-               if (!(vma->vm_flags & VM_SHARED))
+               if (!(vma->vm_flags & VM_MAYSHARE))
                        pagecache_page = hugetlbfs_pagecache_page(h,
                                                                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 (likely(pte_same(entry, huge_ptep_get(ptep))))
-               if (write_access && !pte_write(entry))
+       if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
+               goto out_page_table_lock;
+
+
+       if (flags & FAULT_FLAG_WRITE) {
+               if (!pte_write(entry)) {
                        ret = hugetlb_cow(mm, vma, address, ptep, entry,
                                                        pagecache_page);
+                       goto out_page_table_lock;
+               }
+               entry = pte_mkdirty(entry);
+       }
+       entry = pte_mkyoung(entry);
+       if (huge_ptep_set_access_flags(vma, address, ptep, entry,
+                                               flags & FAULT_FLAG_WRITE))
+               update_mmu_cache(vma, address, ptep);
+
+out_page_table_lock:
        spin_unlock(&mm->page_table_lock);
 
        if (pagecache_page) {
                unlock_page(pagecache_page);
                put_page(pagecache_page);
        }
+       if (page != pagecache_page)
+               unlock_page(page);
 
+out_mutex:
        mutex_unlock(&hugetlb_instantiation_mutex);
 
        return ret;
@@ -2056,7 +2841,7 @@ follow_huge_pud(struct mm_struct *mm, unsigned long address,
 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
                        struct page **pages, struct vm_area_struct **vmas,
                        unsigned long *position, int *length, int i,
-                       int write)
+                       unsigned int flags)
 {
        unsigned long pfn_offset;
        unsigned long vaddr = *position;
@@ -2066,28 +2851,42 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
        spin_lock(&mm->page_table_lock);
        while (vaddr < vma->vm_end && remainder) {
                pte_t *pte;
+               int absent;
                struct page *page;
 
                /*
                 * Some archs (sparc64, sh*) have multiple pte_ts to
-                * each hugepage.  We have to make * sure we get the
+                * each hugepage.  We have to make sure we get the
                 * first, for the page indexing below to work.
                 */
                pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
+               absent = !pte || huge_pte_none(huge_ptep_get(pte));
 
-               if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
-                   (write && !pte_write(huge_ptep_get(pte)))) {
+               /*
+                * When coredumping, it suits get_dump_page if we just return
+                * an error where there's an empty slot with no huge pagecache
+                * to back it.  This way, we avoid allocating a hugepage, and
+                * the sparse dumpfile avoids allocating disk blocks, but its
+                * huge holes still show up with zeroes where they need to be.
+                */
+               if (absent && (flags & FOLL_DUMP) &&
+                   !hugetlbfs_pagecache_present(h, vma, vaddr)) {
+                       remainder = 0;
+                       break;
+               }
+
+               if (absent ||
+                   ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
                        int ret;
 
                        spin_unlock(&mm->page_table_lock);
-                       ret = hugetlb_fault(mm, vma, vaddr, write);
+                       ret = hugetlb_fault(mm, vma, vaddr,
+                               (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
                        spin_lock(&mm->page_table_lock);
                        if (!(ret & VM_FAULT_ERROR))
                                continue;
 
                        remainder = 0;
-                       if (!i)
-                               i = -EFAULT;
                        break;
                }
 
@@ -2095,8 +2894,8 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
                page = pte_page(huge_ptep_get(pte));
 same_page:
                if (pages) {
-                       get_page(page);
-                       pages[i] = page + pfn_offset;
+                       pages[i] = mem_map_offset(page, pfn_offset);
+                       get_page(pages[i]);
                }
 
                if (vmas)
@@ -2119,7 +2918,7 @@ same_page:
        *length = remainder;
        *position = vaddr;
 
-       return i;
+       return i ? i : -EFAULT;
 }
 
 void hugetlb_change_protection(struct vm_area_struct *vma,
@@ -2134,7 +2933,7 @@ 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);
@@ -2149,19 +2948,26 @@ 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);
 }
 
 int hugetlb_reserve_pages(struct inode *inode,
                                        long from, long to,
-                                       struct vm_area_struct *vma)
+                                       struct vm_area_struct *vma,
+                                       vm_flags_t vm_flags)
 {
        long ret, chg;
        struct hstate *h = hstate_inode(inode);
+       struct hugepage_subpool *spool = subpool_inode(inode);
 
-       if (vma && vma->vm_flags & VM_NORESERVE)
+       /*
+        * Only apply hugepage reservation if asked. At fault time, an
+        * attempt will be made for VM_NORESERVE to allocate a page
+        * without using reserves
+        */
+       if (vm_flags & VM_NORESERVE)
                return 0;
 
        /*
@@ -2170,7 +2976,7 @@ int hugetlb_reserve_pages(struct inode *inode,
         * to reserve the full area even if read-only as mprotect() may be
         * called to make the mapping read-write. Assume !vma is a shm mapping
         */
-       if (!vma || vma->vm_flags & VM_SHARED)
+       if (!vma || vma->vm_flags & VM_MAYSHARE)
                chg = region_chg(&inode->i_mapping->private_list, from, to);
        else {
                struct resv_map *resv_map = resv_map_alloc();
@@ -2186,14 +2992,32 @@ int hugetlb_reserve_pages(struct inode *inode,
        if (chg < 0)
                return chg;
 
-       if (hugetlb_get_quota(inode->i_mapping, chg))
+       /* There must be enough pages in the subpool for the mapping */
+       if (hugepage_subpool_get_pages(spool, chg))
                return -ENOSPC;
+
+       /*
+        * Check enough hugepages are available for the reservation.
+        * Hand the pages back to the subpool if there are not
+        */
        ret = hugetlb_acct_memory(h, chg);
        if (ret < 0) {
-               hugetlb_put_quota(inode->i_mapping, chg);
+               hugepage_subpool_put_pages(spool, chg);
                return ret;
        }
-       if (!vma || vma->vm_flags & VM_SHARED)
+
+       /*
+        * Account for the reservations made. Shared mappings record regions
+        * that have reservations as they are shared by multiple VMAs.
+        * When the last VMA disappears, the region map says how much
+        * the reservation was and the page cache tells how much of
+        * the reservation was consumed. Private mappings are per-VMA and
+        * only the consumed reservations are tracked. When the VMA
+        * disappears, the original reservation is the VMA size and the
+        * consumed reservations are stored in the map. Hence, nothing
+        * else has to be done for private mappings here
+        */
+       if (!vma || vma->vm_flags & VM_MAYSHARE)
                region_add(&inode->i_mapping->private_list, from, to);
        return 0;
 }
@@ -2202,11 +3026,51 @@ void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
 {
        struct hstate *h = hstate_inode(inode);
        long chg = region_truncate(&inode->i_mapping->private_list, offset);
+       struct hugepage_subpool *spool = subpool_inode(inode);
 
        spin_lock(&inode->i_lock);
-       inode->i_blocks -= blocks_per_huge_page(h);
+       inode->i_blocks -= (blocks_per_huge_page(h) * freed);
        spin_unlock(&inode->i_lock);
 
-       hugetlb_put_quota(inode->i_mapping, (chg - freed));
+       hugepage_subpool_put_pages(spool, (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