mm: account for MAP_SHARED mappings using VM_MAYSHARE and not VM_SHARED in hugetlbfs
[linux-2.6.git] / mm / hugetlb.c
index 034617f..e83ad2c 100644 (file)
 #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/nodemask.h>
 #include <linux/pagemap.h>
 #include <linux/mempolicy.h>
 #include <linux/cpuset.h>
 #include <linux/mutex.h>
+#include <linux/bootmem.h>
+#include <linux/sysfs.h>
 
 #include <asm/page.h>
 #include <asm/pgtable.h>
+#include <asm/io.h>
 
 #include <linux/hugetlb.h>
 #include "internal.h"
 
 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
-static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
-static unsigned long surplus_huge_pages;
-unsigned long max_huge_pages;
-static struct list_head hugepage_freelists[MAX_NUMNODES];
-static unsigned int nr_huge_pages_node[MAX_NUMNODES];
-static unsigned int free_huge_pages_node[MAX_NUMNODES];
-static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
 unsigned long hugepages_treat_as_movable;
-int hugetlb_dynamic_pool;
-static int hugetlb_next_nid;
+
+static int max_hstate;
+unsigned int default_hstate_idx;
+struct hstate hstates[HUGE_MAX_HSTATE];
+
+__initdata LIST_HEAD(huge_boot_pages);
+
+/* for command line parsing */
+static struct hstate * __initdata parsed_hstate;
+static unsigned long __initdata default_hstate_max_huge_pages;
+static unsigned long __initdata default_hstate_size;
+
+#define for_each_hstate(h) \
+       for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
 
 /*
  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  */
 static DEFINE_SPINLOCK(hugetlb_lock);
 
-static void clear_huge_page(struct page *page, unsigned long addr)
+/*
+ * Region tracking -- allows tracking of reservations and instantiated pages
+ *                    across the pages in a mapping.
+ *
+ * The region data structures are protected by a combination of the mmap_sem
+ * and the hugetlb_instantion_mutex.  To access or modify a region the caller
+ * must either hold the mmap_sem for write, or the mmap_sem for read and
+ * the hugetlb_instantiation mutex:
+ *
+ *     down_write(&mm->mmap_sem);
+ * or
+ *     down_read(&mm->mmap_sem);
+ *     mutex_lock(&hugetlb_instantiation_mutex);
+ */
+struct file_region {
+       struct list_head link;
+       long from;
+       long to;
+};
+
+static long region_add(struct list_head *head, long f, long t)
+{
+       struct file_region *rg, *nrg, *trg;
+
+       /* Locate the region we are either in or before. */
+       list_for_each_entry(rg, head, link)
+               if (f <= rg->to)
+                       break;
+
+       /* Round our left edge to the current segment if it encloses us. */
+       if (f > rg->from)
+               f = rg->from;
+
+       /* Check for and consume any regions we now overlap with. */
+       nrg = rg;
+       list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
+               if (&rg->link == head)
+                       break;
+               if (rg->from > t)
+                       break;
+
+               /* If this area reaches higher then extend our area to
+                * include it completely.  If this is not the first area
+                * which we intend to reuse, free it. */
+               if (rg->to > t)
+                       t = rg->to;
+               if (rg != nrg) {
+                       list_del(&rg->link);
+                       kfree(rg);
+               }
+       }
+       nrg->from = f;
+       nrg->to = t;
+       return 0;
+}
+
+static long region_chg(struct list_head *head, long f, long t)
+{
+       struct file_region *rg, *nrg;
+       long chg = 0;
+
+       /* Locate the region we are before or in. */
+       list_for_each_entry(rg, head, link)
+               if (f <= rg->to)
+                       break;
+
+       /* If we are below the current region then a new region is required.
+        * Subtle, allocate a new region at the position but make it zero
+        * size such that we can guarantee to record the reservation. */
+       if (&rg->link == head || t < rg->from) {
+               nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
+               if (!nrg)
+                       return -ENOMEM;
+               nrg->from = f;
+               nrg->to   = f;
+               INIT_LIST_HEAD(&nrg->link);
+               list_add(&nrg->link, rg->link.prev);
+
+               return t - f;
+       }
+
+       /* Round our left edge to the current segment if it encloses us. */
+       if (f > rg->from)
+               f = rg->from;
+       chg = t - f;
+
+       /* Check for and consume any regions we now overlap with. */
+       list_for_each_entry(rg, rg->link.prev, link) {
+               if (&rg->link == head)
+                       break;
+               if (rg->from > t)
+                       return chg;
+
+               /* We overlap with this area, if it extends futher than
+                * us then we must extend ourselves.  Account for its
+                * existing reservation. */
+               if (rg->to > t) {
+                       chg += rg->to - t;
+                       t = rg->to;
+               }
+               chg -= rg->to - rg->from;
+       }
+       return chg;
+}
+
+static long region_truncate(struct list_head *head, long end)
+{
+       struct file_region *rg, *trg;
+       long chg = 0;
+
+       /* Locate the region we are either in or before. */
+       list_for_each_entry(rg, head, link)
+               if (end <= rg->to)
+                       break;
+       if (&rg->link == head)
+               return 0;
+
+       /* If we are in the middle of a region then adjust it. */
+       if (end > rg->from) {
+               chg = rg->to - end;
+               rg->to = end;
+               rg = list_entry(rg->link.next, typeof(*rg), link);
+       }
+
+       /* Drop any remaining regions. */
+       list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
+               if (&rg->link == head)
+                       break;
+               chg += rg->to - rg->from;
+               list_del(&rg->link);
+               kfree(rg);
+       }
+       return chg;
+}
+
+static long region_count(struct list_head *head, long f, long t)
+{
+       struct file_region *rg;
+       long chg = 0;
+
+       /* Locate each segment we overlap with, and count that overlap. */
+       list_for_each_entry(rg, head, link) {
+               int seg_from;
+               int seg_to;
+
+               if (rg->to <= f)
+                       continue;
+               if (rg->from >= t)
+                       break;
+
+               seg_from = max(rg->from, f);
+               seg_to = min(rg->to, t);
+
+               chg += seg_to - seg_from;
+       }
+
+       return chg;
+}
+
+/*
+ * Convert the address within this vma to the page offset within
+ * the mapping, in pagecache page units; huge pages here.
+ */
+static pgoff_t vma_hugecache_offset(struct hstate *h,
+                       struct vm_area_struct *vma, unsigned long address)
+{
+       return ((address - vma->vm_start) >> huge_page_shift(h)) +
+                       (vma->vm_pgoff >> huge_page_order(h));
+}
+
+/*
+ * 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);
+}
+
+/*
+ * 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
+ * alignment.
+ */
+#define HPAGE_RESV_OWNER    (1UL << 0)
+#define HPAGE_RESV_UNMAPPED (1UL << 1)
+#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
+
+/*
+ * These helpers are used to track how many pages are reserved for
+ * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
+ * is guaranteed to have their future faults succeed.
+ *
+ * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
+ * the reserve counters are updated with the hugetlb_lock held. It is safe
+ * to reset the VMA at fork() time as it is not in use yet and there is no
+ * chance of the global counters getting corrupted as a result of the values.
+ *
+ * The private mapping reservation is represented in a subtly different
+ * manner to a shared mapping.  A shared mapping has a region map associated
+ * with the underlying file, this region map represents the backing file
+ * pages which have ever had a reservation assigned which this persists even
+ * after the page is instantiated.  A private mapping has a region map
+ * associated with the original mmap which is attached to all VMAs which
+ * reference it, this region map represents those offsets which have consumed
+ * reservation ie. where pages have been instantiated.
+ */
+static unsigned long get_vma_private_data(struct vm_area_struct *vma)
+{
+       return (unsigned long)vma->vm_private_data;
+}
+
+static void set_vma_private_data(struct vm_area_struct *vma,
+                                                       unsigned long value)
+{
+       vma->vm_private_data = (void *)value;
+}
+
+struct resv_map {
+       struct kref refs;
+       struct list_head regions;
+};
+
+static struct resv_map *resv_map_alloc(void)
+{
+       struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
+       if (!resv_map)
+               return NULL;
+
+       kref_init(&resv_map->refs);
+       INIT_LIST_HEAD(&resv_map->regions);
+
+       return resv_map;
+}
+
+static void resv_map_release(struct kref *ref)
+{
+       struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
+
+       /* Clear out any active regions before we release the map. */
+       region_truncate(&resv_map->regions, 0);
+       kfree(resv_map);
+}
+
+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_MAYSHARE))
+               return (struct resv_map *)(get_vma_private_data(vma) &
+                                                       ~HPAGE_RESV_MASK);
+       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_MAYSHARE);
+
+       set_vma_private_data(vma, (get_vma_private_data(vma) &
+                               HPAGE_RESV_MASK) | (unsigned long)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_MAYSHARE);
+
+       set_vma_private_data(vma, get_vma_private_data(vma) | flags);
+}
+
+static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
+{
+       VM_BUG_ON(!is_vm_hugetlb_page(vma));
+
+       return (get_vma_private_data(vma) & flag) != 0;
+}
+
+/* Decrement the reserved pages in the hugepage pool by one */
+static void decrement_hugepage_resv_vma(struct hstate *h,
+                       struct vm_area_struct *vma)
+{
+       if (vma->vm_flags & VM_NORESERVE)
+               return;
+
+       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)) {
+               /*
+                * Only the process that called mmap() has reserves for
+                * private mappings.
+                */
+               h->resv_huge_pages--;
+       }
+}
+
+/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
+void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
+{
+       VM_BUG_ON(!is_vm_hugetlb_page(vma));
+       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_MAYSHARE)
+               return 1;
+       if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
+               return 1;
+       return 0;
+}
+
+static void clear_gigantic_page(struct page *page,
+                       unsigned long addr, unsigned long sz)
 {
        int i;
+       struct page *p = page;
 
        might_sleep();
-       for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
+       for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
+               cond_resched();
+               clear_user_highpage(p, addr + i * PAGE_SIZE);
+       }
+}
+static void clear_huge_page(struct page *page,
+                       unsigned long addr, unsigned long sz)
+{
+       int i;
+
+       if (unlikely(sz > MAX_ORDER_NR_PAGES)) {
+               clear_gigantic_page(page, addr, sz);
+               return;
+       }
+
+       might_sleep();
+       for (i = 0; i < sz/PAGE_SIZE; i++) {
                cond_resched();
                clear_user_highpage(page + i, addr + i * PAGE_SIZE);
        }
 }
 
+static void copy_gigantic_page(struct page *dst, struct page *src,
+                          unsigned long addr, struct vm_area_struct *vma)
+{
+       int i;
+       struct hstate *h = hstate_vma(vma);
+       struct page *dst_base = dst;
+       struct page *src_base = src;
+       might_sleep();
+       for (i = 0; i < pages_per_huge_page(h); ) {
+               cond_resched();
+               copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
+
+               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)
 {
        int i;
+       struct hstate *h = hstate_vma(vma);
+
+       if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
+               copy_gigantic_page(dst, src, addr, vma);
+               return;
+       }
 
        might_sleep();
-       for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
+       for (i = 0; i < pages_per_huge_page(h); i++) {
                cond_resched();
                copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
        }
 }
 
-static void enqueue_huge_page(struct page *page)
+static void enqueue_huge_page(struct hstate *h, struct page *page)
 {
        int nid = page_to_nid(page);
-       list_add(&page->lru, &hugepage_freelists[nid]);
-       free_huge_pages++;
-       free_huge_pages_node[nid]++;
+       list_add(&page->lru, &h->hugepage_freelists[nid]);
+       h->free_huge_pages++;
+       h->free_huge_pages_node[nid]++;
+}
+
+static struct page *dequeue_huge_page(struct hstate *h)
+{
+       int nid;
+       struct page *page = NULL;
+
+       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;
+               }
+       }
+       return page;
 }
 
-static struct page *dequeue_huge_page(struct vm_area_struct *vma,
-                               unsigned long address)
+static struct page *dequeue_huge_page_vma(struct hstate *h,
+                               struct vm_area_struct *vma,
+                               unsigned long address, int avoid_reserve)
 {
        int nid;
        struct page *page = NULL;
        struct mempolicy *mpol;
+       nodemask_t *nodemask;
        struct zonelist *zonelist = huge_zonelist(vma, address,
-                                       htlb_alloc_mask, &mpol);
-       struct zone **z;
-
-       for (z = zonelist->zones; *z; z++) {
-               nid = zone_to_nid(*z);
-               if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
-                   !list_empty(&hugepage_freelists[nid])) {
-                       page = list_entry(hugepage_freelists[nid].next,
+                                       htlb_alloc_mask, &mpol, &nodemask);
+       struct zone *zone;
+       struct zoneref *z;
+
+       /*
+        * A child process with MAP_PRIVATE mappings created by their parent
+        * have no page reserves. This check ensures that reservations are
+        * not "stolen". The child may still get SIGKILLed
+        */
+       if (!vma_has_reserves(vma) &&
+                       h->free_huge_pages - h->resv_huge_pages == 0)
+               return NULL;
+
+       /* 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;
+
+       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);
-                       free_huge_pages--;
-                       free_huge_pages_node[nid]--;
-                       if (vma && vma->vm_flags & VM_MAYSHARE)
-                               resv_huge_pages--;
+                       h->free_huge_pages--;
+                       h->free_huge_pages_node[nid]--;
+
+                       if (!avoid_reserve)
+                               decrement_hugepage_resv_vma(h, vma);
+
                        break;
                }
        }
-       mpol_free(mpol);        /* unref if mpol !NULL */
+       mpol_cond_put(mpol);
        return page;
 }
 
-static void update_and_free_page(struct page *page)
+static void update_and_free_page(struct hstate *h, struct page *page)
 {
        int i;
-       nr_huge_pages--;
-       nr_huge_pages_node[page_to_nid(page)]--;
-       for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); 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);
        }
        set_compound_page_dtor(page, NULL);
        set_page_refcounted(page);
-       __free_pages(page, HUGETLB_PAGE_ORDER);
+       arch_release_hugepage(page);
+       __free_pages(page, huge_page_order(h));
+}
+
+struct hstate *size_to_hstate(unsigned long size)
+{
+       struct hstate *h;
+
+       for_each_hstate(h) {
+               if (huge_page_size(h) == size)
+                       return h;
+       }
+       return NULL;
 }
 
 static void free_huge_page(struct page *page)
 {
+       /*
+        * Can't pass hstate in here because it is called from the
+        * compound page destructor.
+        */
+       struct hstate *h = page_hstate(page);
        int nid = page_to_nid(page);
+       struct address_space *mapping;
 
+       mapping = (struct address_space *) page_private(page);
+       set_page_private(page, 0);
        BUG_ON(page_count(page));
        INIT_LIST_HEAD(&page->lru);
 
        spin_lock(&hugetlb_lock);
-       if (surplus_huge_pages_node[nid]) {
-               update_and_free_page(page);
-               surplus_huge_pages--;
-               surplus_huge_pages_node[nid]--;
+       if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
+               update_and_free_page(h, page);
+               h->surplus_huge_pages--;
+               h->surplus_huge_pages_node[nid]--;
        } else {
-               enqueue_huge_page(page);
+               enqueue_huge_page(h, page);
        }
        spin_unlock(&hugetlb_lock);
+       if (mapping)
+               hugetlb_put_quota(mapping, 1);
 }
 
 /*
@@ -136,7 +583,7 @@ static void free_huge_page(struct page *page)
  * balanced by operating on them in a round-robin fashion.
  * Returns 1 if an adjustment was made.
  */
-static int adjust_pool_surplus(int delta)
+static int adjust_pool_surplus(struct hstate *h, int delta)
 {
        static int prev_nid;
        int nid = prev_nid;
@@ -149,15 +596,15 @@ static int adjust_pool_surplus(int delta)
                        nid = first_node(node_online_map);
 
                /* To shrink on this node, there must be a surplus page */
-               if (delta < 0 && !surplus_huge_pages_node[nid])
+               if (delta < 0 && !h->surplus_huge_pages_node[nid])
                        continue;
                /* Surplus cannot exceed the total number of pages */
-               if (delta > 0 && surplus_huge_pages_node[nid] >=
-                                               nr_huge_pages_node[nid])
+               if (delta > 0 && h->surplus_huge_pages_node[nid] >=
+                                               h->nr_huge_pages_node[nid])
                        continue;
 
-               surplus_huge_pages += delta;
-               surplus_huge_pages_node[nid] += delta;
+               h->surplus_huge_pages += delta;
+               h->surplus_huge_pages_node[nid] += delta;
                ret = 1;
                break;
        } while (nid != prev_nid);
@@ -166,78 +613,156 @@ static int adjust_pool_surplus(int delta)
        return ret;
 }
 
-static struct page *alloc_fresh_huge_page_node(int nid)
+static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
+{
+       set_compound_page_dtor(page, free_huge_page);
+       spin_lock(&hugetlb_lock);
+       h->nr_huge_pages++;
+       h->nr_huge_pages_node[nid]++;
+       spin_unlock(&hugetlb_lock);
+       put_page(page); /* free it into the hugepage allocator */
+}
+
+static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
 {
        struct page *page;
 
+       if (h->order >= MAX_ORDER)
+               return NULL;
+
        page = alloc_pages_node(nid,
-               htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
-               HUGETLB_PAGE_ORDER);
+               htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
+                                               __GFP_REPEAT|__GFP_NOWARN,
+               huge_page_order(h));
        if (page) {
-               set_compound_page_dtor(page, free_huge_page);
-               spin_lock(&hugetlb_lock);
-               nr_huge_pages++;
-               nr_huge_pages_node[nid]++;
-               spin_unlock(&hugetlb_lock);
-               put_page(page); /* free it into the hugepage allocator */
+               if (arch_prepare_hugepage(page)) {
+                       __free_pages(page, huge_page_order(h));
+                       return NULL;
+               }
+               prep_new_huge_page(h, page, nid);
        }
 
        return page;
 }
 
-static int alloc_fresh_huge_page(void)
+/*
+ * 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.
+ */
+static int hstate_next_node(struct hstate *h)
+{
+       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;
+}
+
+static int alloc_fresh_huge_page(struct hstate *h)
 {
        struct page *page;
        int start_nid;
        int next_nid;
        int ret = 0;
 
-       start_nid = hugetlb_next_nid;
+       start_nid = h->hugetlb_next_nid;
 
        do {
-               page = alloc_fresh_huge_page_node(hugetlb_next_nid);
+               page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
                if (page)
                        ret = 1;
-               /*
-                * 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.
-                */
-               next_nid = next_node(hugetlb_next_nid, node_online_map);
-               if (next_nid == MAX_NUMNODES)
-                       next_nid = first_node(node_online_map);
-               hugetlb_next_nid = next_nid;
-       } while (!page && hugetlb_next_nid != start_nid);
+               next_nid = hstate_next_node(h);
+       } while (!page && h->hugetlb_next_nid != start_nid);
+
+       if (ret)
+               count_vm_event(HTLB_BUDDY_PGALLOC);
+       else
+               count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 
        return ret;
 }
 
-static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
-                                               unsigned long address)
+static struct page *alloc_buddy_huge_page(struct hstate *h,
+                       struct vm_area_struct *vma, unsigned long address)
 {
        struct page *page;
+       unsigned int nid;
 
-       /* Check if the dynamic pool is enabled */
-       if (!hugetlb_dynamic_pool)
+       if (h->order >= MAX_ORDER)
                return NULL;
 
-       page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
-                                       HUGETLB_PAGE_ORDER);
+       /*
+        * Assume we will successfully allocate the surplus page to
+        * prevent racing processes from causing the surplus to exceed
+        * overcommit
+        *
+        * This however introduces a different race, where a process B
+        * tries to grow the static hugepage pool while alloc_pages() is
+        * called by process A. B will only examine the per-node
+        * counters in determining if surplus huge pages can be
+        * converted to normal huge pages in adjust_pool_surplus(). A
+        * won't be able to increment the per-node counter, until the
+        * lock is dropped by B, but B doesn't drop hugetlb_lock until
+        * no more huge pages can be converted from surplus to normal
+        * state (and doesn't try to convert again). Thus, we have a
+        * case where a surplus huge page exists, the pool is grown, and
+        * the surplus huge page still exists after, even though it
+        * should just have been converted to a normal huge page. This
+        * does not leak memory, though, as the hugepage will be freed
+        * once it is out of use. It also does not allow the counters to
+        * go out of whack in adjust_pool_surplus() as we don't modify
+        * the node values until we've gotten the hugepage and only the
+        * per-node value is checked there.
+        */
+       spin_lock(&hugetlb_lock);
+       if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
+               spin_unlock(&hugetlb_lock);
+               return NULL;
+       } else {
+               h->nr_huge_pages++;
+               h->surplus_huge_pages++;
+       }
+       spin_unlock(&hugetlb_lock);
+
+       page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
+                                       __GFP_REPEAT|__GFP_NOWARN,
+                                       huge_page_order(h));
+
+       if (page && arch_prepare_hugepage(page)) {
+               __free_pages(page, huge_page_order(h));
+               return 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);
                set_compound_page_dtor(page, free_huge_page);
-               spin_lock(&hugetlb_lock);
-               nr_huge_pages++;
-               nr_huge_pages_node[page_to_nid(page)]++;
-               surplus_huge_pages++;
-               surplus_huge_pages_node[page_to_nid(page)]++;
-               spin_unlock(&hugetlb_lock);
+               /*
+                * We incremented the global counters already
+                */
+               h->nr_huge_pages_node[nid]++;
+               h->surplus_huge_pages_node[nid]++;
+               __count_vm_event(HTLB_BUDDY_PGALLOC);
+       } else {
+               h->nr_huge_pages--;
+               h->surplus_huge_pages--;
+               __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
        }
+       spin_unlock(&hugetlb_lock);
 
        return page;
 }
@@ -246,16 +771,18 @@ static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
  * Increase the hugetlb pool such that it can accomodate a reservation
  * of size 'delta'.
  */
-static int gather_surplus_pages(int delta)
+static int gather_surplus_pages(struct hstate *h, int delta)
 {
        struct list_head surplus_list;
        struct page *page, *tmp;
        int ret, i;
        int needed, allocated;
 
-       needed = (resv_huge_pages + delta) - free_huge_pages;
-       if (needed <= 0)
+       needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
+       if (needed <= 0) {
+               h->resv_huge_pages += delta;
                return 0;
+       }
 
        allocated = 0;
        INIT_LIST_HEAD(&surplus_list);
@@ -264,7 +791,7 @@ static int gather_surplus_pages(int delta)
 retry:
        spin_unlock(&hugetlb_lock);
        for (i = 0; i < needed; i++) {
-               page = alloc_buddy_huge_page(NULL, 0);
+               page = alloc_buddy_huge_page(h, NULL, 0);
                if (!page) {
                        /*
                         * We were not able to allocate enough pages to
@@ -285,7 +812,8 @@ retry:
         * because either resv_huge_pages or free_huge_pages may have changed.
         */
        spin_lock(&hugetlb_lock);
-       needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
+       needed = (h->resv_huge_pages + delta) -
+                       (h->free_huge_pages + allocated);
        if (needed > 0)
                goto retry;
 
@@ -293,26 +821,37 @@ retry:
         * The surplus_list now contains _at_least_ the number of extra pages
         * needed to accomodate the reservation.  Add the appropriate number
         * of pages to the hugetlb pool and free the extras back to the buddy
-        * allocator.
+        * allocator.  Commit the entire reservation here to prevent another
+        * process from stealing the pages as they are added to the pool but
+        * before they are reserved.
         */
        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);
-               if ((--needed) >= 0)
-                       enqueue_huge_page(page);
-               else {
+               enqueue_huge_page(h, page);
+       }
+
+       /* 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);
                        /*
-                        * Decrement the refcount and free the page using its
-                        * destructor.  This must be done with hugetlb_lock
+                        * 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.
                         */
-                       spin_unlock(&hugetlb_lock);
-                       put_page(page);
-                       spin_lock(&hugetlb_lock);
+                       free_huge_page(page);
                }
+               spin_lock(&hugetlb_lock);
        }
 
        return ret;
@@ -323,160 +862,317 @@ free:
  * allocated to satisfy the reservation must be explicitly freed if they were
  * never used.
  */
-void return_unused_surplus_pages(unsigned long unused_resv_pages)
+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;
 
-       nr_pages = min(unused_resv_pages, surplus_huge_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;
+
+       /* Cannot return gigantic pages currently */
+       if (h->order >= MAX_ORDER)
+               return;
+
+       nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
 
-       while (nr_pages) {
+       while (remaining_iterations-- && nr_pages) {
                nid = next_node(nid, node_online_map);
                if (nid == MAX_NUMNODES)
                        nid = first_node(node_online_map);
 
-               if (!surplus_huge_pages_node[nid])
+               if (!h->surplus_huge_pages_node[nid])
                        continue;
 
-               if (!list_empty(&hugepage_freelists[nid])) {
-                       page = list_entry(hugepage_freelists[nid].next,
+               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(page);
-                       free_huge_pages--;
-                       free_huge_pages_node[nid]--;
-                       surplus_huge_pages--;
-                       surplus_huge_pages_node[nid]--;
+                       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();
                }
        }
 }
 
-static struct page *alloc_huge_page(struct vm_area_struct *vma,
-                                   unsigned long addr)
+/*
+ * 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.
+ */
+static long vma_needs_reservation(struct hstate *h,
+                       struct vm_area_struct *vma, unsigned long addr)
 {
-       struct page *page = NULL;
-       int use_reserved_page = vma->vm_flags & VM_MAYSHARE;
+       struct address_space *mapping = vma->vm_file->f_mapping;
+       struct inode *inode = mapping->host;
 
-       spin_lock(&hugetlb_lock);
-       if (!use_reserved_page && (free_huge_pages <= resv_huge_pages))
-               goto fail;
+       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);
 
-       page = dequeue_huge_page(vma, addr);
-       if (!page)
-               goto fail;
+       } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
+               return 1;
 
-       spin_unlock(&hugetlb_lock);
-       set_page_refcounted(page);
-       return page;
+       } else  {
+               long err;
+               pgoff_t idx = vma_hugecache_offset(h, vma, addr);
+               struct resv_map *reservations = vma_resv_map(vma);
 
-fail:
-       spin_unlock(&hugetlb_lock);
+               err = region_chg(&reservations->regions, idx, idx + 1);
+               if (err < 0)
+                       return err;
+               return 0;
+       }
+}
+static void vma_commit_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_MAYSHARE) {
+               pgoff_t idx = vma_hugecache_offset(h, vma, addr);
+               region_add(&inode->i_mapping->private_list, idx, idx + 1);
+
+       } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
+               pgoff_t idx = vma_hugecache_offset(h, vma, addr);
+               struct resv_map *reservations = vma_resv_map(vma);
+
+               /* Mark this page used in the map. */
+               region_add(&reservations->regions, idx, idx + 1);
+       }
+}
+
+static struct page *alloc_huge_page(struct vm_area_struct *vma,
+                                   unsigned long addr, int avoid_reserve)
+{
+       struct hstate *h = hstate_vma(vma);
+       struct page *page;
+       struct address_space *mapping = vma->vm_file->f_mapping;
+       struct inode *inode = mapping->host;
+       long chg;
 
        /*
-        * Private mappings do not use reserved huge pages so the allocation
-        * may have failed due to an undersized hugetlb pool.  Try to grab a
-        * surplus huge page from the buddy allocator.
+        * 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.
         */
-       if (!use_reserved_page)
-               page = alloc_buddy_huge_page(vma, addr);
+       chg = vma_needs_reservation(h, vma, addr);
+       if (chg < 0)
+               return ERR_PTR(chg);
+       if (chg)
+               if (hugetlb_get_quota(inode->i_mapping, chg))
+                       return ERR_PTR(-ENOSPC);
+
+       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);
+               if (!page) {
+                       hugetlb_put_quota(inode->i_mapping, chg);
+                       return ERR_PTR(-VM_FAULT_OOM);
+               }
+       }
+
+       set_page_refcounted(page);
+       set_page_private(page, (unsigned long) mapping);
+
+       vma_commit_reservation(h, vma, addr);
 
        return page;
 }
 
-static int __init hugetlb_init(void)
+int __weak alloc_bootmem_huge_page(struct hstate *h)
 {
-       unsigned long i;
+       struct huge_bootmem_page *m;
+       int nr_nodes = nodes_weight(node_online_map);
 
-       if (HPAGE_SHIFT == 0)
-               return 0;
+       while (nr_nodes) {
+               void *addr;
 
-       for (i = 0; i < MAX_NUMNODES; ++i)
-               INIT_LIST_HEAD(&hugepage_freelists[i]);
+               addr = __alloc_bootmem_node_nopanic(
+                               NODE_DATA(h->hugetlb_next_nid),
+                               huge_page_size(h), huge_page_size(h), 0);
 
-       hugetlb_next_nid = first_node(node_online_map);
-
-       for (i = 0; i < max_huge_pages; ++i) {
-               if (!alloc_fresh_huge_page())
-                       break;
+               if (addr) {
+                       /*
+                        * Use the beginning of the huge page to store the
+                        * huge_bootmem_page struct (until gather_bootmem
+                        * puts them into the mem_map).
+                        */
+                       m = addr;
+                       goto found;
+               }
+               hstate_next_node(h);
+               nr_nodes--;
        }
-       max_huge_pages = free_huge_pages = nr_huge_pages = i;
-       printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
        return 0;
+
+found:
+       BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
+       /* Put them into a private list first because mem_map is not up yet */
+       list_add(&m->list, &huge_boot_pages);
+       m->hstate = h;
+       return 1;
 }
-module_init(hugetlb_init);
 
-static int __init hugetlb_setup(char *s)
+static void prep_compound_huge_page(struct page *page, int order)
 {
-       if (sscanf(s, "%lu", &max_huge_pages) <= 0)
-               max_huge_pages = 0;
-       return 1;
+       if (unlikely(order > (MAX_ORDER - 1)))
+               prep_compound_gigantic_page(page, order);
+       else
+               prep_compound_page(page, order);
 }
-__setup("hugepages=", hugetlb_setup);
 
-static unsigned int cpuset_mems_nr(unsigned int *array)
+/* Put bootmem huge pages into the standard lists after mem_map is up */
+static void __init gather_bootmem_prealloc(void)
 {
-       int node;
-       unsigned int nr = 0;
+       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;
+               __ClearPageReserved(page);
+               WARN_ON(page_count(page) != 1);
+               prep_compound_huge_page(page, h->order);
+               prep_new_huge_page(h, page, page_to_nid(page));
+       }
+}
 
-       for_each_node_mask(node, cpuset_current_mems_allowed)
-               nr += array[node];
+static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
+{
+       unsigned long i;
 
-       return nr;
+       for (i = 0; i < h->max_huge_pages; ++i) {
+               if (h->order >= MAX_ORDER) {
+                       if (!alloc_bootmem_huge_page(h))
+                               break;
+               } else if (!alloc_fresh_huge_page(h))
+                       break;
+       }
+       h->max_huge_pages = i;
+}
+
+static void __init hugetlb_init_hstates(void)
+{
+       struct hstate *h;
+
+       for_each_hstate(h) {
+               /* oversize hugepages were init'ed in early boot */
+               if (h->order < MAX_ORDER)
+                       hugetlb_hstate_alloc_pages(h);
+       }
+}
+
+static char * __init memfmt(char *buf, unsigned long n)
+{
+       if (n >= (1UL << 30))
+               sprintf(buf, "%lu GB", n >> 30);
+       else if (n >= (1UL << 20))
+               sprintf(buf, "%lu MB", n >> 20);
+       else
+               sprintf(buf, "%lu KB", n >> 10);
+       return buf;
+}
+
+static void __init report_hugepages(void)
+{
+       struct hstate *h;
+
+       for_each_hstate(h) {
+               char buf[32];
+               printk(KERN_INFO "HugeTLB registered %s page size, "
+                                "pre-allocated %ld pages\n",
+                       memfmt(buf, huge_page_size(h)),
+                       h->free_huge_pages);
+       }
 }
 
-#ifdef CONFIG_SYSCTL
 #ifdef CONFIG_HIGHMEM
-static void try_to_free_low(unsigned long count)
+static void try_to_free_low(struct hstate *h, unsigned long count)
 {
        int i;
 
+       if (h->order >= MAX_ORDER)
+               return;
+
        for (i = 0; i < MAX_NUMNODES; ++i) {
                struct page *page, *next;
-               list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
-                       if (count >= nr_huge_pages)
+               struct list_head *freel = &h->hugepage_freelists[i];
+               list_for_each_entry_safe(page, next, freel, lru) {
+                       if (count >= h->nr_huge_pages)
                                return;
                        if (PageHighMem(page))
                                continue;
                        list_del(&page->lru);
-                       update_and_free_page(page);
-                       free_huge_pages--;
-                       free_huge_pages_node[page_to_nid(page)]--;
+                       update_and_free_page(h, page);
+                       h->free_huge_pages--;
+                       h->free_huge_pages_node[page_to_nid(page)]--;
                }
        }
 }
 #else
-static inline void try_to_free_low(unsigned long count)
+static inline void try_to_free_low(struct hstate *h, unsigned long count)
 {
 }
 #endif
 
-#define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
-static unsigned long set_max_huge_pages(unsigned long count)
+#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)
 {
        unsigned long min_count, ret;
 
+       if (h->order >= MAX_ORDER)
+               return h->max_huge_pages;
+
        /*
         * Increase the pool size
         * First take pages out of surplus state.  Then make up the
         * remaining difference by allocating fresh huge pages.
+        *
+        * We might race with alloc_buddy_huge_page() here and be unable
+        * to convert a surplus huge page to a normal huge page. That is
+        * not critical, though, it just means the overall size of the
+        * pool might be one hugepage larger than it needs to be, but
+        * within all the constraints specified by the sysctls.
         */
        spin_lock(&hugetlb_lock);
-       while (surplus_huge_pages && count > persistent_huge_pages) {
-               if (!adjust_pool_surplus(-1))
+       while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
+               if (!adjust_pool_surplus(h, -1))
                        break;
        }
 
-       while (count > persistent_huge_pages) {
-               int ret;
+       while (count > persistent_huge_pages(h)) {
                /*
                 * If this allocation races such that we no longer need the
                 * page, free_huge_page will handle it by freeing the page
                 * and reducing the surplus.
                 */
                spin_unlock(&hugetlb_lock);
-               ret = alloc_fresh_huge_page();
+               ret = alloc_fresh_huge_page(h);
                spin_lock(&hugetlb_lock);
                if (!ret)
                        goto out;
@@ -489,32 +1185,314 @@ static unsigned long set_max_huge_pages(unsigned long count)
         * to keep enough around to satisfy reservations).  Then place
         * pages into surplus state as needed so the pool will shrink
         * to the desired size as pages become free.
+        *
+        * By placing pages into the surplus state independent of the
+        * overcommit value, we are allowing the surplus pool size to
+        * exceed overcommit. There are few sane options here. Since
+        * alloc_buddy_huge_page() is checking the global counter,
+        * though, we'll note that we're not allowed to exceed surplus
+        * and won't grow the pool anywhere else. Not until one of the
+        * sysctls are changed, or the surplus pages go out of use.
         */
-       min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
+       min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
        min_count = max(count, min_count);
-       try_to_free_low(min_count);
-       while (min_count < persistent_huge_pages) {
-               struct page *page = dequeue_huge_page(NULL, 0);
+       try_to_free_low(h, min_count);
+       while (min_count < persistent_huge_pages(h)) {
+               struct page *page = dequeue_huge_page(h);
                if (!page)
                        break;
-               update_and_free_page(page);
+               update_and_free_page(h, page);
        }
-       while (count < persistent_huge_pages) {
-               if (!adjust_pool_surplus(1))
+       while (count < persistent_huge_pages(h)) {
+               if (!adjust_pool_surplus(h, 1))
                        break;
        }
 out:
-       ret = persistent_huge_pages;
+       ret = persistent_huge_pages(h);
        spin_unlock(&hugetlb_lock);
        return ret;
 }
 
+#define HSTATE_ATTR_RO(_name) \
+       static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
+
+#define HSTATE_ATTR(_name) \
+       static struct kobj_attribute _name##_attr = \
+               __ATTR(_name, 0644, _name##_show, _name##_store)
+
+static struct kobject *hugepages_kobj;
+static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
+
+static struct hstate *kobj_to_hstate(struct kobject *kobj)
+{
+       int i;
+       for (i = 0; i < HUGE_MAX_HSTATE; i++)
+               if (hstate_kobjs[i] == kobj)
+                       return &hstates[i];
+       BUG();
+       return NULL;
+}
+
+static ssize_t nr_hugepages_show(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);
+}
+static ssize_t nr_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);
+
+       err = strict_strtoul(buf, 10, &input);
+       if (err)
+               return 0;
+
+       h->max_huge_pages = set_max_huge_pages(h, input);
+
+       return count;
+}
+HSTATE_ATTR(nr_hugepages);
+
+static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
+                                       struct kobj_attribute *attr, char *buf)
+{
+       struct hstate *h = kobj_to_hstate(kobj);
+       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);
+
+       err = strict_strtoul(buf, 10, &input);
+       if (err)
+               return 0;
+
+       spin_lock(&hugetlb_lock);
+       h->nr_overcommit_huge_pages = input;
+       spin_unlock(&hugetlb_lock);
+
+       return count;
+}
+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);
+}
+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);
+       return sprintf(buf, "%lu\n", h->resv_huge_pages);
+}
+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);
+}
+HSTATE_ATTR_RO(surplus_hugepages);
+
+static struct attribute *hstate_attrs[] = {
+       &nr_hugepages_attr.attr,
+       &nr_overcommit_hugepages_attr.attr,
+       &free_hugepages_attr.attr,
+       &resv_hugepages_attr.attr,
+       &surplus_hugepages_attr.attr,
+       NULL,
+};
+
+static struct attribute_group hstate_attr_group = {
+       .attrs = hstate_attrs,
+};
+
+static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
+{
+       int retval;
+
+       hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
+                                                       hugepages_kobj);
+       if (!hstate_kobjs[h - hstates])
+               return -ENOMEM;
+
+       retval = sysfs_create_group(hstate_kobjs[h - hstates],
+                                                       &hstate_attr_group);
+       if (retval)
+               kobject_put(hstate_kobjs[h - hstates]);
+
+       return retval;
+}
+
+static void __init hugetlb_sysfs_init(void)
+{
+       struct hstate *h;
+       int err;
+
+       hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
+       if (!hugepages_kobj)
+               return;
+
+       for_each_hstate(h) {
+               err = hugetlb_sysfs_add_hstate(h);
+               if (err)
+                       printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
+                                                               h->name);
+       }
+}
+
+static void __exit hugetlb_exit(void)
+{
+       struct hstate *h;
+
+       for_each_hstate(h) {
+               kobject_put(hstate_kobjs[h - hstates]);
+       }
+
+       kobject_put(hugepages_kobj);
+}
+module_exit(hugetlb_exit);
+
+static int __init hugetlb_init(void)
+{
+       /* Some platform decide whether they support huge pages at boot
+        * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
+        * there is no such support
+        */
+       if (HPAGE_SHIFT == 0)
+               return 0;
+
+       if (!size_to_hstate(default_hstate_size)) {
+               default_hstate_size = HPAGE_SIZE;
+               if (!size_to_hstate(default_hstate_size))
+                       hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
+       }
+       default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
+       if (default_hstate_max_huge_pages)
+               default_hstate.max_huge_pages = default_hstate_max_huge_pages;
+
+       hugetlb_init_hstates();
+
+       gather_bootmem_prealloc();
+
+       report_hugepages();
+
+       hugetlb_sysfs_init();
+
+       return 0;
+}
+module_init(hugetlb_init);
+
+/* Should be called on processing a hugepagesz=... option */
+void __init hugetlb_add_hstate(unsigned order)
+{
+       struct hstate *h;
+       unsigned long i;
+
+       if (size_to_hstate(PAGE_SIZE << order)) {
+               printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
+               return;
+       }
+       BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
+       BUG_ON(order == 0);
+       h = &hstates[max_hstate++];
+       h->order = order;
+       h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
+       h->nr_huge_pages = 0;
+       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);
+       snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
+                                       huge_page_size(h)/1024);
+
+       parsed_hstate = h;
+}
+
+static int __init hugetlb_nrpages_setup(char *s)
+{
+       unsigned long *mhp;
+       static unsigned long *last_mhp;
+
+       /*
+        * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
+        * so this hugepages= parameter goes to the "default hstate".
+        */
+       if (!max_hstate)
+               mhp = &default_hstate_max_huge_pages;
+       else
+               mhp = &parsed_hstate->max_huge_pages;
+
+       if (mhp == last_mhp) {
+               printk(KERN_WARNING "hugepages= specified twice without "
+                       "interleaving hugepagesz=, ignoring\n");
+               return 1;
+       }
+
+       if (sscanf(s, "%lu", mhp) <= 0)
+               *mhp = 0;
+
+       /*
+        * Global state is always initialized later in hugetlb_init.
+        * But we need to allocate >= MAX_ORDER hstates here early to still
+        * use the bootmem allocator.
+        */
+       if (max_hstate && parsed_hstate->order >= MAX_ORDER)
+               hugetlb_hstate_alloc_pages(parsed_hstate);
+
+       last_mhp = mhp;
+
+       return 1;
+}
+__setup("hugepages=", hugetlb_nrpages_setup);
+
+static int __init hugetlb_default_setup(char *s)
+{
+       default_hstate_size = memparse(s, &s);
+       return 1;
+}
+__setup("default_hugepagesz=", hugetlb_default_setup);
+
+static unsigned int cpuset_mems_nr(unsigned int *array)
+{
+       int node;
+       unsigned int nr = 0;
+
+       for_each_node_mask(node, cpuset_current_mems_allowed)
+               nr += array[node];
+
+       return nr;
+}
+
+#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)
 {
+       struct hstate *h = &default_hstate;
+       unsigned long tmp;
+
+       if (!write)
+               tmp = h->max_huge_pages;
+
+       table->data = &tmp;
+       table->maxlen = sizeof(unsigned long);
        proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
-       max_huge_pages = set_max_huge_pages(max_huge_pages);
+
+       if (write)
+               h->max_huge_pages = set_max_huge_pages(h, tmp);
+
        return 0;
 }
 
@@ -530,36 +1508,145 @@ int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
        return 0;
 }
 
+int hugetlb_overcommit_handler(struct ctl_table *table, int write,
+                       struct file *file, void __user *buffer,
+                       size_t *length, loff_t *ppos)
+{
+       struct hstate *h = &default_hstate;
+       unsigned long tmp;
+
+       if (!write)
+               tmp = h->nr_overcommit_huge_pages;
+
+       table->data = &tmp;
+       table->maxlen = sizeof(unsigned long);
+       proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
+
+       if (write) {
+               spin_lock(&hugetlb_lock);
+               h->nr_overcommit_huge_pages = tmp;
+               spin_unlock(&hugetlb_lock);
+       }
+
+       return 0;
+}
+
 #endif /* CONFIG_SYSCTL */
 
-int hugetlb_report_meminfo(char *buf)
+void hugetlb_report_meminfo(struct seq_file *m)
 {
-       return sprintf(buf,
-                       "HugePages_Total: %5lu\n"
-                       "HugePages_Free:  %5lu\n"
-                       "HugePages_Rsvd:  %5lu\n"
-                       "HugePages_Surp:  %5lu\n"
-                       "Hugepagesize:    %5lu kB\n",
-                       nr_huge_pages,
-                       free_huge_pages,
-                       resv_huge_pages,
-                       surplus_huge_pages,
-                       HPAGE_SIZE/1024);
+       struct hstate *h = &default_hstate;
+       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,
+                       h->surplus_huge_pages,
+                       1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
 }
 
 int hugetlb_report_node_meminfo(int nid, char *buf)
 {
+       struct hstate *h = &default_hstate;
        return sprintf(buf,
                "Node %d HugePages_Total: %5u\n"
-               "Node %d HugePages_Free:  %5u\n",
-               nid, nr_huge_pages_node[nid],
-               nid, free_huge_pages_node[nid]);
+               "Node %d HugePages_Free:  %5u\n"
+               "Node %d HugePages_Surp:  %5u\n",
+               nid, h->nr_huge_pages_node[nid],
+               nid, h->free_huge_pages_node[nid],
+               nid, h->surplus_huge_pages_node[nid]);
 }
 
 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 unsigned long hugetlb_total_pages(void)
 {
-       return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
+       struct hstate *h = &default_hstate;
+       return h->nr_huge_pages * pages_per_huge_page(h);
+}
+
+static int hugetlb_acct_memory(struct hstate *h, long delta)
+{
+       int ret = -ENOMEM;
+
+       spin_lock(&hugetlb_lock);
+       /*
+        * When cpuset is configured, it breaks the strict hugetlb page
+        * reservation as the accounting is done on a global variable. Such
+        * reservation is completely rubbish in the presence of cpuset because
+        * the reservation is not checked against page availability for the
+        * current cpuset. Application can still potentially OOM'ed by kernel
+        * with lack of free htlb page in cpuset that the task is in.
+        * Attempt to enforce strict accounting with cpuset is almost
+        * impossible (or too ugly) because cpuset is too fluid that
+        * task or memory node can be dynamically moved between cpusets.
+        *
+        * The change of semantics for shared hugetlb mapping with cpuset is
+        * undesirable. However, in order to preserve some of the semantics,
+        * we fall back to check against current free page availability as
+        * a best attempt and hopefully to minimize the impact of changing
+        * semantics that cpuset has.
+        */
+       if (delta > 0) {
+               if (gather_surplus_pages(h, delta) < 0)
+                       goto out;
+
+               if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
+                       return_unused_surplus_pages(h, delta);
+                       goto out;
+               }
+       }
+
+       ret = 0;
+       if (delta < 0)
+               return_unused_surplus_pages(h, (unsigned long) -delta);
+
+out:
+       spin_unlock(&hugetlb_lock);
+       return ret;
+}
+
+static void hugetlb_vm_op_open(struct vm_area_struct *vma)
+{
+       struct resv_map *reservations = vma_resv_map(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
+        * after this open call completes.  It is therefore safe to take a
+        * new reference here without additional locking.
+        */
+       if (reservations)
+               kref_get(&reservations->refs);
+}
+
+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);
+       unsigned long reserve;
+       unsigned long start;
+       unsigned long end;
+
+       if (reservations) {
+               start = vma_hugecache_offset(h, vma, vma->vm_start);
+               end = vma_hugecache_offset(h, vma, vma->vm_end);
+
+               reserve = (end - start) -
+                       region_count(&reservations->regions, start, end);
+
+               kref_put(&reservations->refs, resv_map_release);
+
+               if (reserve) {
+                       hugetlb_acct_memory(h, -reserve);
+                       hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
+               }
+       }
 }
 
 /*
@@ -576,6 +1663,8 @@ static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 
 struct vm_operations_struct hugetlb_vm_ops = {
        .fault = hugetlb_vm_op_fault,
+       .open = hugetlb_vm_op_open,
+       .close = hugetlb_vm_op_close,
 };
 
 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
@@ -587,7 +1676,7 @@ static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
                entry =
                    pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
        } else {
-               entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
+               entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
        }
        entry = pte_mkyoung(entry);
        entry = pte_mkhuge(entry);
@@ -600,8 +1689,8 @@ static void set_huge_ptep_writable(struct vm_area_struct *vma,
 {
        pte_t entry;
 
-       entry = pte_mkwrite(pte_mkdirty(*ptep));
-       if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
+       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);
        }
 }
@@ -614,22 +1703,29 @@ int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
        struct page *ptepage;
        unsigned long addr;
        int cow;
+       struct hstate *h = hstate_vma(vma);
+       unsigned long sz = huge_page_size(h);
 
        cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 
-       for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
+       for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
                src_pte = huge_pte_offset(src, addr);
                if (!src_pte)
                        continue;
-               dst_pte = huge_pte_alloc(dst, addr);
+               dst_pte = huge_pte_alloc(dst, addr, sz);
                if (!dst_pte)
                        goto nomem;
+
+               /* If the pagetables are shared don't copy or take references */
+               if (dst_pte == src_pte)
+                       continue;
+
                spin_lock(&dst->page_table_lock);
-               spin_lock(&src->page_table_lock);
-               if (!pte_none(*src_pte)) {
+               spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
+               if (!huge_pte_none(huge_ptep_get(src_pte))) {
                        if (cow)
-                               ptep_set_wrprotect(src, addr, src_pte);
-                       entry = *src_pte;
+                               huge_ptep_set_wrprotect(src, addr, src_pte);
+                       entry = huge_ptep_get(src_pte);
                        ptepage = pte_page(entry);
                        get_page(ptepage);
                        set_huge_pte_at(dst, addr, dst_pte, entry);
@@ -644,7 +1740,7 @@ nomem:
 }
 
 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
-                           unsigned long end)
+                           unsigned long end, struct page *ref_page)
 {
        struct mm_struct *mm = vma->vm_mm;
        unsigned long address;
@@ -652,6 +1748,9 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
        pte_t pte;
        struct page *page;
        struct page *tmp;
+       struct hstate *h = hstate_vma(vma);
+       unsigned long sz = huge_page_size(h);
+
        /*
         * A page gathering list, protected by per file i_mmap_lock. The
         * lock is used to avoid list corruption from multiple unmapping
@@ -660,11 +1759,12 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
        LIST_HEAD(page_list);
 
        WARN_ON(!is_vm_hugetlb_page(vma));
-       BUG_ON(start & ~HPAGE_MASK);
-       BUG_ON(end & ~HPAGE_MASK);
+       BUG_ON(start & ~huge_page_mask(h));
+       BUG_ON(end & ~huge_page_mask(h));
 
+       mmu_notifier_invalidate_range_start(mm, start, end);
        spin_lock(&mm->page_table_lock);
-       for (address = start; address < end; address += HPAGE_SIZE) {
+       for (address = start; address < end; address += sz) {
                ptep = huge_pte_offset(mm, address);
                if (!ptep)
                        continue;
@@ -672,8 +1772,29 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
                if (huge_pmd_unshare(mm, &address, ptep))
                        continue;
 
+               /*
+                * 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;
+
+                       /*
+                        * Mark the VMA as having unmapped its page so that
+                        * future faults in this VMA will fail rather than
+                        * looking like data was lost
+                        */
+                       set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
+               }
+
                pte = huge_ptep_get_and_clear(mm, address, ptep);
-               if (pte_none(pte))
+               if (huge_pte_none(pte))
                        continue;
 
                page = pte_page(pte);
@@ -683,6 +1804,7 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
        }
        spin_unlock(&mm->page_table_lock);
        flush_tlb_range(vma, start, end);
+       mmu_notifier_invalidate_range_end(mm, start, end);
        list_for_each_entry_safe(page, tmp, &page_list, lru) {
                list_del(&page->lru);
                put_page(page);
@@ -690,31 +1812,70 @@ 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)
+                         unsigned long end, struct page *ref_page)
 {
+       spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
+       __unmap_hugepage_range(vma, start, end, ref_page);
+       spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
+}
+
+/*
+ * This is called when the original mapper is failing to COW a MAP_PRIVATE
+ * mappping it owns the reserve page for. The intention is to unmap the page
+ * from other VMAs and let the children be SIGKILLed if they are faulting the
+ * same region.
+ */
+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;
+       pgoff_t pgoff;
+
        /*
-        * It is undesirable to test vma->vm_file as it should be non-null
-        * for valid hugetlb area. However, vm_file will be NULL in the error
-        * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
-        * do_mmap_pgoff() nullifies vma->vm_file before calling this function
-        * to clean up. Since no pte has actually been setup, it is safe to
-        * do nothing in this case.
+        * vm_pgoff is in PAGE_SIZE units, hence the different calculation
+        * from page cache lookup which is in HPAGE_SIZE units.
         */
-       if (vma->vm_file) {
-               spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
-               __unmap_hugepage_range(vma, start, end);
-               spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
+       address = address & huge_page_mask(h);
+       pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
+               + (vma->vm_pgoff >> PAGE_SHIFT);
+       mapping = (struct address_space *)page_private(page);
+
+       vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
+               /* Do not unmap the current VMA */
+               if (iter_vma == vma)
+                       continue;
+
+               /*
+                * Unmap the page from other VMAs without their own reserves.
+                * They get marked to be SIGKILLed if they fault in these
+                * areas. This is because a future no-page fault on this VMA
+                * could insert a zeroed page instead of the data existing
+                * 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 + huge_page_size(h),
+                               page);
        }
+
+       return 1;
 }
 
 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
-                       unsigned long address, pte_t *ptep, pte_t pte)
+                       unsigned long address, pte_t *ptep, pte_t pte,
+                       struct page *pagecache_page)
 {
+       struct hstate *h = hstate_vma(vma);
        struct page *old_page, *new_page;
        int avoidcopy;
+       int outside_reserve = 0;
 
        old_page = pte_page(pte);
 
+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);
@@ -723,21 +1884,55 @@ static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
                return 0;
        }
 
+       /*
+        * If the process that created a MAP_PRIVATE mapping is about to
+        * perform a COW due to a shared page count, attempt to satisfy
+        * the allocation without using the existing reserves. The pagecache
+        * page is used to determine if the reserve at this address was
+        * consumed or not. If reserves were used, a partial faulted mapping
+        * at the time of fork() could consume its reserves on COW instead
+        * of the full address range.
+        */
+       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);
-       new_page = alloc_huge_page(vma, address);
+       new_page = alloc_huge_page(vma, address, outside_reserve);
 
-       if (!new_page) {
+       if (IS_ERR(new_page)) {
                page_cache_release(old_page);
-               return VM_FAULT_OOM;
+
+               /*
+                * If a process owning a MAP_PRIVATE mapping fails to COW,
+                * it is due to references held by a child and an insufficient
+                * huge page pool. To guarantee the original mappers
+                * reliability, unmap the page from child processes. The child
+                * may get SIGKILLed if it later faults.
+                */
+               if (outside_reserve) {
+                       BUG_ON(huge_pte_none(pte));
+                       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;
+                       }
+                       WARN_ON_ONCE(1);
+               }
+
+               return -PTR_ERR(new_page);
        }
 
        spin_unlock(&mm->page_table_lock);
        copy_huge_page(new_page, old_page, address, vma);
+       __SetPageUptodate(new_page);
        spin_lock(&mm->page_table_lock);
 
-       ptep = huge_pte_offset(mm, address & HPAGE_MASK);
-       if (likely(pte_same(*ptep, pte))) {
+       ptep = huge_pte_offset(mm, address & huge_page_mask(h));
+       if (likely(pte_same(huge_ptep_get(ptep), pte))) {
                /* Break COW */
+               huge_ptep_clear_flush(vma, address, ptep);
                set_huge_pte_at(mm, address, ptep,
                                make_huge_pte(vma, new_page, 1));
                /* Make the old page be freed below */
@@ -748,19 +1943,44 @@ static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
        return 0;
 }
 
+/* Return the pagecache page at a given address within a VMA */
+static struct page *hugetlbfs_pagecache_page(struct hstate *h,
+                       struct vm_area_struct *vma, unsigned long address)
+{
+       struct address_space *mapping;
+       pgoff_t idx;
+
+       mapping = vma->vm_file->f_mapping;
+       idx = vma_hugecache_offset(h, vma, address);
+
+       return find_lock_page(mapping, idx);
+}
+
 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
                        unsigned long address, pte_t *ptep, int write_access)
 {
+       struct hstate *h = hstate_vma(vma);
        int ret = VM_FAULT_SIGBUS;
-       unsigned long idx;
+       pgoff_t idx;
        unsigned long size;
        struct page *page;
        struct address_space *mapping;
        pte_t new_pte;
 
+       /*
+        * 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
+        */
+       if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
+               printk(KERN_WARNING
+                       "PID %d killed due to inadequate hugepage pool\n",
+                       current->pid);
+               return ret;
+       }
+
        mapping = vma->vm_file->f_mapping;
-       idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
-               + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
+       idx = vma_hugecache_offset(h, vma, address);
 
        /*
         * Use page lock to guard against racing truncation
@@ -769,41 +1989,55 @@ static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 retry:
        page = find_lock_page(mapping, idx);
        if (!page) {
-               size = i_size_read(mapping->host) >> HPAGE_SHIFT;
+               size = i_size_read(mapping->host) >> huge_page_shift(h);
                if (idx >= size)
                        goto out;
-               if (hugetlb_get_quota(mapping))
-                       goto out;
-               page = alloc_huge_page(vma, address);
-               if (!page) {
-                       hugetlb_put_quota(mapping);
-                       ret = VM_FAULT_OOM;
+               page = alloc_huge_page(vma, address, 0);
+               if (IS_ERR(page)) {
+                       ret = -PTR_ERR(page);
                        goto out;
                }
-               clear_huge_page(page, address);
+               clear_huge_page(page, address, huge_page_size(h));
+               __SetPageUptodate(page);
 
-               if (vma->vm_flags & VM_SHARED) {
+               if (vma->vm_flags & VM_MAYSHARE) {
                        int err;
+                       struct inode *inode = mapping->host;
 
                        err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
                        if (err) {
                                put_page(page);
-                               hugetlb_put_quota(mapping);
                                if (err == -EEXIST)
                                        goto retry;
                                goto out;
                        }
+
+                       spin_lock(&inode->i_lock);
+                       inode->i_blocks += blocks_per_huge_page(h);
+                       spin_unlock(&inode->i_lock);
                } else
                        lock_page(page);
        }
 
+       /*
+        * If we are going to COW a private mapping later, we examine the
+        * pending reservations for this page now. This will ensure that
+        * any allocations necessary to record that reservation occur outside
+        * the spinlock.
+        */
+       if (write_access && !(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) >> HPAGE_SHIFT;
+       size = i_size_read(mapping->host) >> huge_page_shift(h);
        if (idx >= size)
                goto backout;
 
        ret = 0;
-       if (!pte_none(*ptep))
+       if (!huge_pte_none(huge_ptep_get(ptep)))
                goto backout;
 
        new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
@@ -812,7 +2046,7 @@ retry:
 
        if (write_access && !(vma->vm_flags & VM_SHARED)) {
                /* Optimization, do the COW without a second fault */
-               ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
+               ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
        }
 
        spin_unlock(&mm->page_table_lock);
@@ -822,7 +2056,7 @@ out:
 
 backout:
        spin_unlock(&mm->page_table_lock);
-       hugetlb_put_quota(mapping);
+backout_unlocked:
        unlock_page(page);
        put_page(page);
        goto out;
@@ -834,9 +2068,11 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        pte_t *ptep;
        pte_t entry;
        int ret;
+       struct page *pagecache_page = NULL;
        static DEFINE_MUTEX(hugetlb_instantiation_mutex);
+       struct hstate *h = hstate_vma(vma);
 
-       ptep = huge_pte_alloc(mm, address);
+       ptep = huge_pte_alloc(mm, address, huge_page_size(h));
        if (!ptep)
                return VM_FAULT_OOM;
 
@@ -846,33 +2082,93 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
         * the same page in the page cache.
         */
        mutex_lock(&hugetlb_instantiation_mutex);
-       entry = *ptep;
-       if (pte_none(entry)) {
+       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;
+               goto out_mutex;
        }
 
        ret = 0;
 
+       /*
+        * If we are going to COW the mapping later, we examine the pending
+        * reservations for this page now. This will ensure that any
+        * allocations necessary to record that reservation occur outside the
+        * spinlock. For private mappings, we also lookup the pagecache
+        * page now as it is used to determine if a reservation has been
+        * consumed.
+        */
+       if (write_access && !pte_write(entry)) {
+               if (vma_needs_reservation(h, vma, address) < 0) {
+                       ret = VM_FAULT_OOM;
+                       goto out_mutex;
+               }
+
+               if (!(vma->vm_flags & VM_MAYSHARE))
+                       pagecache_page = hugetlbfs_pagecache_page(h,
+                                                               vma, address);
+       }
+
        spin_lock(&mm->page_table_lock);
        /* Check for a racing update before calling hugetlb_cow */
-       if (likely(pte_same(entry, *ptep)))
-               if (write_access && !pte_write(entry))
-                       ret = hugetlb_cow(mm, vma, address, ptep, entry);
+       if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
+               goto out_page_table_lock;
+
+
+       if (write_access) {
+               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, write_access))
+               update_mmu_cache(vma, address, entry);
+
+out_page_table_lock:
        spin_unlock(&mm->page_table_lock);
+
+       if (pagecache_page) {
+               unlock_page(pagecache_page);
+               put_page(pagecache_page);
+       }
+
+out_mutex:
        mutex_unlock(&hugetlb_instantiation_mutex);
 
        return ret;
 }
 
+/* Can be overriden by architectures */
+__attribute__((weak)) struct page *
+follow_huge_pud(struct mm_struct *mm, unsigned long address,
+              pud_t *pud, int write)
+{
+       BUG();
+       return NULL;
+}
+
+static int huge_zeropage_ok(pte_t *ptep, int write, int shared)
+{
+       if (!ptep || write || shared)
+               return 0;
+       else
+               return huge_pte_none(huge_ptep_get(ptep));
+}
+
 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)
+                       unsigned long *position, int *length, int i,
+                       int write)
 {
        unsigned long pfn_offset;
        unsigned long vaddr = *position;
        int remainder = *length;
+       struct hstate *h = hstate_vma(vma);
+       int zeropage_ok = 0;
+       int shared = vma->vm_flags & VM_SHARED;
 
        spin_lock(&mm->page_table_lock);
        while (vaddr < vma->vm_end && remainder) {
@@ -884,13 +2180,17 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
                 * each hugepage.  We have to make * sure we get the
                 * first, for the page indexing below to work.
                 */
-               pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
+               pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
+               if (huge_zeropage_ok(pte, write, shared))
+                       zeropage_ok = 1;
 
-               if (!pte || pte_none(*pte)) {
+               if (!pte ||
+                   (huge_pte_none(huge_ptep_get(pte)) && !zeropage_ok) ||
+                   (write && !pte_write(huge_ptep_get(pte)))) {
                        int ret;
 
                        spin_unlock(&mm->page_table_lock);
-                       ret = hugetlb_fault(mm, vma, vaddr, 0);
+                       ret = hugetlb_fault(mm, vma, vaddr, write);
                        spin_lock(&mm->page_table_lock);
                        if (!(ret & VM_FAULT_ERROR))
                                continue;
@@ -901,12 +2201,15 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
                        break;
                }
 
-               pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
-               page = pte_page(*pte);
+               pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
+               page = pte_page(huge_ptep_get(pte));
 same_page:
                if (pages) {
-                       get_page(page);
-                       pages[i] = page + pfn_offset;
+                       if (zeropage_ok)
+                               pages[i] = ZERO_PAGE(0);
+                       else
+                               pages[i] = mem_map_offset(page, pfn_offset);
+                       get_page(pages[i]);
                }
 
                if (vmas)
@@ -917,7 +2220,7 @@ same_page:
                --remainder;
                ++i;
                if (vaddr < vma->vm_end && remainder &&
-                               pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
+                               pfn_offset < pages_per_huge_page(h)) {
                        /*
                         * We use pfn_offset to avoid touching the pageframes
                         * of this compound page.
@@ -939,19 +2242,20 @@ void hugetlb_change_protection(struct vm_area_struct *vma,
        unsigned long start = address;
        pte_t *ptep;
        pte_t pte;
+       struct hstate *h = hstate_vma(vma);
 
        BUG_ON(address >= end);
        flush_cache_range(vma, address, end);
 
        spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
        spin_lock(&mm->page_table_lock);
-       for (; address < end; address += HPAGE_SIZE) {
+       for (; address < end; address += huge_page_size(h)) {
                ptep = huge_pte_offset(mm, address);
                if (!ptep)
                        continue;
                if (huge_pmd_unshare(mm, &address, ptep))
                        continue;
-               if (!pte_none(*ptep)) {
+               if (!huge_pte_none(huge_ptep_get(ptep))) {
                        pte = huge_ptep_get_and_clear(mm, address, ptep);
                        pte = pte_mkhuge(pte_modify(pte, newprot));
                        set_huge_pte_at(mm, address, ptep, pte);
@@ -963,184 +2267,83 @@ void hugetlb_change_protection(struct vm_area_struct *vma,
        flush_tlb_range(vma, start, end);
 }
 
-struct file_region {
-       struct list_head link;
-       long from;
-       long to;
-};
-
-static long region_add(struct list_head *head, long f, long t)
+int hugetlb_reserve_pages(struct inode *inode,
+                                       long from, long to,
+                                       struct vm_area_struct *vma,
+                                       int acctflag)
 {
-       struct file_region *rg, *nrg, *trg;
-
-       /* Locate the region we are either in or before. */
-       list_for_each_entry(rg, head, link)
-               if (f <= rg->to)
-                       break;
-
-       /* Round our left edge to the current segment if it encloses us. */
-       if (f > rg->from)
-               f = rg->from;
-
-       /* Check for and consume any regions we now overlap with. */
-       nrg = rg;
-       list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
-               if (&rg->link == head)
-                       break;
-               if (rg->from > t)
-                       break;
-
-               /* If this area reaches higher then extend our area to
-                * include it completely.  If this is not the first area
-                * which we intend to reuse, free it. */
-               if (rg->to > t)
-                       t = rg->to;
-               if (rg != nrg) {
-                       list_del(&rg->link);
-                       kfree(rg);
-               }
-       }
-       nrg->from = f;
-       nrg->to = t;
-       return 0;
-}
-
-static long region_chg(struct list_head *head, long f, long t)
-{
-       struct file_region *rg, *nrg;
-       long chg = 0;
+       long ret, chg;
+       struct hstate *h = hstate_inode(inode);
 
-       /* Locate the region we are before or in. */
-       list_for_each_entry(rg, head, link)
-               if (f <= rg->to)
-                       break;
+       /*
+        * Only apply hugepage reservation if asked. At fault time, an
+        * attempt will be made for VM_NORESERVE to allocate a page
+        * and filesystem quota without using reserves
+        */
+       if (acctflag & VM_NORESERVE)
+               return 0;
 
-       /* If we are below the current region then a new region is required.
-        * Subtle, allocate a new region at the position but make it zero
-        * size such that we can guarentee to record the reservation. */
-       if (&rg->link == head || t < rg->from) {
-               nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
-               if (!nrg)
+       /*
+        * Shared mappings base their reservation on the number of pages that
+        * are already allocated on behalf of the file. Private mappings need
+        * 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_MAYSHARE)
+               chg = region_chg(&inode->i_mapping->private_list, from, to);
+       else {
+               struct resv_map *resv_map = resv_map_alloc();
+               if (!resv_map)
                        return -ENOMEM;
-               nrg->from = f;
-               nrg->to   = f;
-               INIT_LIST_HEAD(&nrg->link);
-               list_add(&nrg->link, rg->link.prev);
-
-               return t - f;
-       }
-
-       /* Round our left edge to the current segment if it encloses us. */
-       if (f > rg->from)
-               f = rg->from;
-       chg = t - f;
-
-       /* Check for and consume any regions we now overlap with. */
-       list_for_each_entry(rg, rg->link.prev, link) {
-               if (&rg->link == head)
-                       break;
-               if (rg->from > t)
-                       return chg;
-
-               /* We overlap with this area, if it extends futher than
-                * us then we must extend ourselves.  Account for its
-                * existing reservation. */
-               if (rg->to > t) {
-                       chg += rg->to - t;
-                       t = rg->to;
-               }
-               chg -= rg->to - rg->from;
-       }
-       return chg;
-}
 
-static long region_truncate(struct list_head *head, long end)
-{
-       struct file_region *rg, *trg;
-       long chg = 0;
-
-       /* Locate the region we are either in or before. */
-       list_for_each_entry(rg, head, link)
-               if (end <= rg->to)
-                       break;
-       if (&rg->link == head)
-               return 0;
+               chg = to - from;
 
-       /* If we are in the middle of a region then adjust it. */
-       if (end > rg->from) {
-               chg = rg->to - end;
-               rg->to = end;
-               rg = list_entry(rg->link.next, typeof(*rg), link);
+               set_vma_resv_map(vma, resv_map);
+               set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
        }
 
-       /* Drop any remaining regions. */
-       list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
-               if (&rg->link == head)
-                       break;
-               chg += rg->to - rg->from;
-               list_del(&rg->link);
-               kfree(rg);
-       }
-       return chg;
-}
+       if (chg < 0)
+               return chg;
 
-static int hugetlb_acct_memory(long delta)
-{
-       int ret = -ENOMEM;
+       /* There must be enough filesystem quota for the mapping */
+       if (hugetlb_get_quota(inode->i_mapping, chg))
+               return -ENOSPC;
 
-       spin_lock(&hugetlb_lock);
        /*
-        * When cpuset is configured, it breaks the strict hugetlb page
-        * reservation as the accounting is done on a global variable. Such
-        * reservation is completely rubbish in the presence of cpuset because
-        * the reservation is not checked against page availability for the
-        * current cpuset. Application can still potentially OOM'ed by kernel
-        * with lack of free htlb page in cpuset that the task is in.
-        * Attempt to enforce strict accounting with cpuset is almost
-        * impossible (or too ugly) because cpuset is too fluid that
-        * task or memory node can be dynamically moved between cpusets.
-        *
-        * The change of semantics for shared hugetlb mapping with cpuset is
-        * undesirable. However, in order to preserve some of the semantics,
-        * we fall back to check against current free page availability as
-        * a best attempt and hopefully to minimize the impact of changing
-        * semantics that cpuset has.
+        * Check enough hugepages are available for the reservation.
+        * Hand back the quota if there are not
         */
-       if (delta > 0) {
-               if (gather_surplus_pages(delta) < 0)
-                       goto out;
-
-               if (delta > cpuset_mems_nr(free_huge_pages_node))
-                       goto out;
+       ret = hugetlb_acct_memory(h, chg);
+       if (ret < 0) {
+               hugetlb_put_quota(inode->i_mapping, chg);
+               return ret;
        }
 
-       ret = 0;
-       resv_huge_pages += delta;
-       if (delta < 0)
-               return_unused_surplus_pages((unsigned long) -delta);
-
-out:
-       spin_unlock(&hugetlb_lock);
-       return ret;
-}
-
-int hugetlb_reserve_pages(struct inode *inode, long from, long to)
-{
-       long ret, chg;
-
-       chg = region_chg(&inode->i_mapping->private_list, from, to);
-       if (chg < 0)
-               return chg;
-
-       ret = hugetlb_acct_memory(chg);
-       if (ret < 0)
-               return ret;
-       region_add(&inode->i_mapping->private_list, from, to);
+       /*
+        * 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;
 }
 
 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);
-       hugetlb_acct_memory(freed - chg);
+
+       spin_lock(&inode->i_lock);
+       inode->i_blocks -= blocks_per_huge_page(h);
+       spin_unlock(&inode->i_lock);
+
+       hugetlb_put_quota(inode->i_mapping, (chg - freed));
+       hugetlb_acct_memory(h, -(chg - freed));
 }