mm: account for MAP_SHARED mappings using VM_MAYSHARE and not VM_SHARED in hugetlbfs
[linux-2.6.git] / mm / hugetlb.c
index 243a868..e83ad2c 100644 (file)
@@ -7,8 +7,10 @@
 #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>
@@ -19,6 +21,7 @@
 
 #include <asm/page.h>
 #include <asm/pgtable.h>
+#include <asm/io.h>
 
 #include <linux/hugetlb.h>
 #include "internal.h"
@@ -31,9 +34,12 @@ 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)++)
@@ -214,6 +220,35 @@ static pgoff_t vma_hugecache_offset(struct hstate *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.
@@ -257,7 +292,7 @@ struct resv_map {
        struct list_head regions;
 };
 
-struct resv_map *resv_map_alloc(void)
+static struct resv_map *resv_map_alloc(void)
 {
        struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
        if (!resv_map)
@@ -269,7 +304,7 @@ struct resv_map *resv_map_alloc(void)
        return resv_map;
 }
 
-void resv_map_release(struct kref *ref)
+static void resv_map_release(struct kref *ref)
 {
        struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
 
@@ -281,16 +316,16 @@ void resv_map_release(struct kref *ref)
 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
 {
        VM_BUG_ON(!is_vm_hugetlb_page(vma));
-       if (!(vma->vm_flags & VM_SHARED))
+       if (!(vma->vm_flags & VM_MAYSHARE))
                return (struct resv_map *)(get_vma_private_data(vma) &
                                                        ~HPAGE_RESV_MASK);
-       return 0;
+       return NULL;
 }
 
 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
 {
        VM_BUG_ON(!is_vm_hugetlb_page(vma));
-       VM_BUG_ON(vma->vm_flags & VM_SHARED);
+       VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 
        set_vma_private_data(vma, (get_vma_private_data(vma) &
                                HPAGE_RESV_MASK) | (unsigned long)map);
@@ -299,7 +334,7 @@ static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
 {
        VM_BUG_ON(!is_vm_hugetlb_page(vma));
-       VM_BUG_ON(vma->vm_flags & VM_SHARED);
+       VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 
        set_vma_private_data(vma, get_vma_private_data(vma) | flags);
 }
@@ -318,7 +353,7 @@ static void decrement_hugepage_resv_vma(struct hstate *h,
        if (vma->vm_flags & VM_NORESERVE)
                return;
 
-       if (vma->vm_flags & VM_SHARED) {
+       if (vma->vm_flags & VM_MAYSHARE) {
                /* Shared mappings always use reserves */
                h->resv_huge_pages--;
        } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
@@ -334,25 +369,42 @@ static void decrement_hugepage_resv_vma(struct hstate *h,
 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
 {
        VM_BUG_ON(!is_vm_hugetlb_page(vma));
-       if (!(vma->vm_flags & VM_SHARED))
+       if (!(vma->vm_flags & VM_MAYSHARE))
                vma->vm_private_data = (void *)0;
 }
 
 /* Returns true if the VMA has associated reserve pages */
-static int vma_has_private_reserves(struct vm_area_struct *vma)
+static int vma_has_reserves(struct vm_area_struct *vma)
 {
-       if (vma->vm_flags & VM_SHARED)
-               return 0;
-       if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER))
-               return 0;
-       return 1;
+       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 < 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();
@@ -360,12 +412,34 @@ static void clear_huge_page(struct page *page,
        }
 }
 
+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 < pages_per_huge_page(h); i++) {
                cond_resched();
@@ -417,7 +491,7 @@ static struct page *dequeue_huge_page_vma(struct hstate *h,
         * have no page reserves. This check ensures that reservations are
         * not "stolen". The child may still get SIGKILLed
         */
-       if (!vma_has_private_reserves(vma) &&
+       if (!vma_has_reserves(vma) &&
                        h->free_huge_pages - h->resv_huge_pages == 0)
                return NULL;
 
@@ -450,6 +524,8 @@ static void update_and_free_page(struct hstate *h, struct page *page)
 {
        int i;
 
+       VM_BUG_ON(h->order >= MAX_ORDER);
+
        h->nr_huge_pages--;
        h->nr_huge_pages_node[page_to_nid(page)]--;
        for (i = 0; i < pages_per_huge_page(h); i++) {
@@ -560,7 +636,7 @@ static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
                huge_page_order(h));
        if (page) {
                if (arch_prepare_hugepage(page)) {
-                       __free_pages(page, HUGETLB_PAGE_ORDER);
+                       __free_pages(page, huge_page_order(h));
                        return NULL;
                }
                prep_new_huge_page(h, page, nid);
@@ -660,6 +736,11 @@ static struct page *alloc_buddy_huge_page(struct hstate *h,
                                        __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) {
                /*
@@ -837,13 +918,13 @@ static void return_unused_surplus_pages(struct hstate *h,
  * an instantiated the change should be committed via vma_commit_reservation.
  * No action is required on failure.
  */
-static int vma_needs_reservation(struct hstate *h,
+static long vma_needs_reservation(struct hstate *h,
                        struct vm_area_struct *vma, unsigned long addr)
 {
        struct address_space *mapping = vma->vm_file->f_mapping;
        struct inode *inode = mapping->host;
 
-       if (vma->vm_flags & VM_SHARED) {
+       if (vma->vm_flags & VM_MAYSHARE) {
                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
                return region_chg(&inode->i_mapping->private_list,
                                                        idx, idx + 1);
@@ -852,7 +933,7 @@ static int vma_needs_reservation(struct hstate *h,
                return 1;
 
        } else  {
-               int err;
+               long err;
                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
                struct resv_map *reservations = vma_resv_map(vma);
 
@@ -868,7 +949,7 @@ static void vma_commit_reservation(struct hstate *h,
        struct address_space *mapping = vma->vm_file->f_mapping;
        struct inode *inode = mapping->host;
 
-       if (vma->vm_flags & VM_SHARED) {
+       if (vma->vm_flags & VM_MAYSHARE) {
                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
                region_add(&inode->i_mapping->private_list, idx, idx + 1);
 
@@ -888,7 +969,7 @@ static struct page *alloc_huge_page(struct vm_area_struct *vma,
        struct page *page;
        struct address_space *mapping = vma->vm_file->f_mapping;
        struct inode *inode = mapping->host;
-       unsigned int chg;
+       long chg;
 
        /*
         * Processes that did not create the mapping will have no reserves and
@@ -924,14 +1005,7 @@ static struct page *alloc_huge_page(struct vm_area_struct *vma,
        return page;
 }
 
-static __initdata LIST_HEAD(huge_boot_pages);
-
-struct huge_bootmem_page {
-       struct list_head list;
-       struct hstate *hstate;
-};
-
-static int __init alloc_bootmem_huge_page(struct hstate *h)
+int __weak alloc_bootmem_huge_page(struct hstate *h)
 {
        struct huge_bootmem_page *m;
        int nr_nodes = nodes_weight(node_online_map);
@@ -950,8 +1024,7 @@ static int __init alloc_bootmem_huge_page(struct hstate *h)
                         * puts them into the mem_map).
                         */
                        m = addr;
-                       if (m)
-                               goto found;
+                       goto found;
                }
                hstate_next_node(h);
                nr_nodes--;
@@ -966,6 +1039,14 @@ found:
        return 1;
 }
 
+static void prep_compound_huge_page(struct page *page, int order)
+{
+       if (unlikely(order > (MAX_ORDER - 1)))
+               prep_compound_gigantic_page(page, order);
+       else
+               prep_compound_page(page, order);
+}
+
 /* Put bootmem huge pages into the standard lists after mem_map is up */
 static void __init gather_bootmem_prealloc(void)
 {
@@ -976,7 +1057,7 @@ static void __init gather_bootmem_prealloc(void)
                struct hstate *h = m->hstate;
                __ClearPageReserved(page);
                WARN_ON(page_count(page) != 1);
-               prep_compound_page(page, h->order);
+               prep_compound_huge_page(page, h->order);
                prep_new_huge_page(h, page, page_to_nid(page));
        }
 }
@@ -1006,19 +1087,30 @@ static void __init hugetlb_init_hstates(void)
        }
 }
 
+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) {
-               printk(KERN_INFO "Total HugeTLB memory allocated, "
-                               "%ld %dMB pages\n",
-                               h->free_huge_pages,
-                               1 << (h->order + PAGE_SHIFT - 20));
+               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(struct hstate *h, unsigned long count)
 {
@@ -1274,13 +1366,21 @@ module_exit(hugetlb_exit);
 
 static int __init hugetlb_init(void)
 {
-       BUILD_BUG_ON(HPAGE_SHIFT == 0);
+       /* 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(HPAGE_SIZE)) {
-               hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
-               parsed_hstate->max_huge_pages = default_hstate_max_huge_pages;
+       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(HPAGE_SIZE) - hstates;
+       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();
 
@@ -1320,7 +1420,7 @@ void __init hugetlb_add_hstate(unsigned order)
        parsed_hstate = h;
 }
 
-static int __init hugetlb_setup(char *s)
+static int __init hugetlb_nrpages_setup(char *s)
 {
        unsigned long *mhp;
        static unsigned long *last_mhp;
@@ -1355,7 +1455,14 @@ static int __init hugetlb_setup(char *s)
 
        return 1;
 }
-__setup("hugepages=", hugetlb_setup);
+__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)
 {
@@ -1368,6 +1475,7 @@ static unsigned int cpuset_mems_nr(unsigned int *array)
        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)
@@ -1425,15 +1533,15 @@ int hugetlb_overcommit_handler(struct ctl_table *table, int write,
 
 #endif /* CONFIG_SYSCTL */
 
-int hugetlb_report_meminfo(char *buf)
+void hugetlb_report_meminfo(struct seq_file *m)
 {
        struct hstate *h = &default_hstate;
-       return sprintf(buf,
-                       "HugePages_Total: %5lu\n"
-                       "HugePages_Free:  %5lu\n"
-                       "HugePages_Rsvd:  %5lu\n"
-                       "HugePages_Surp:  %5lu\n"
-                       "Hugepagesize:    %5lu kB\n",
+       seq_printf(m,
+                       "HugePages_Total:   %5lu\n"
+                       "HugePages_Free:    %5lu\n"
+                       "HugePages_Rsvd:    %5lu\n"
+                       "HugePages_Surp:    %5lu\n"
+                       "Hugepagesize:   %8lu kB\n",
                        h->nr_huge_pages,
                        h->free_huge_pages,
                        h->resv_huge_pages,
@@ -1534,8 +1642,10 @@ static void hugetlb_vm_op_close(struct vm_area_struct *vma)
 
                kref_put(&reservations->refs, resv_map_release);
 
-               if (reserve)
+               if (reserve) {
                        hugetlb_acct_memory(h, -reserve);
+                       hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
+               }
        }
 }
 
@@ -1652,6 +1762,7 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
        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 += sz) {
                ptep = huge_pte_offset(mm, address);
@@ -1693,6 +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);
@@ -1713,11 +1825,10 @@ void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
  * from other VMAs and let the children be SIGKILLed if they are faulting the
  * same region.
  */
-int unmap_ref_private(struct mm_struct *mm,
-                                       struct vm_area_struct *vma,
-                                       struct page *page,
-                                       unsigned long address)
+static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
+                               struct page *page, unsigned long address)
 {
+       struct hstate *h = hstate_vma(vma);
        struct vm_area_struct *iter_vma;
        struct address_space *mapping;
        struct prio_tree_iter iter;
@@ -1727,7 +1838,7 @@ int unmap_ref_private(struct mm_struct *mm,
         * vm_pgoff is in PAGE_SIZE units, hence the different calculation
         * from page cache lookup which is in HPAGE_SIZE units.
         */
-       address = address & huge_page_mask(hstate_vma(vma));
+       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);
@@ -1746,7 +1857,7 @@ int unmap_ref_private(struct mm_struct *mm,
                 */
                if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
                        unmap_hugepage_range(iter_vma,
-                               address, address + HPAGE_SIZE,
+                               address, address + huge_page_size(h),
                                page);
        }
 
@@ -1782,7 +1893,7 @@ retry_avoidcopy:
         * at the time of fork() could consume its reserves on COW instead
         * of the full address range.
         */
-       if (!(vma->vm_flags & VM_SHARED) &&
+       if (!(vma->vm_flags & VM_MAYSHARE) &&
                        is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
                        old_page != pagecache_page)
                outside_reserve = 1;
@@ -1889,7 +2000,7 @@ retry:
                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;
 
@@ -1908,6 +2019,18 @@ retry:
                        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) >> huge_page_shift(h);
        if (idx >= size)
@@ -1933,6 +2056,7 @@ out:
 
 backout:
        spin_unlock(&mm->page_table_lock);
+backout_unlocked:
        unlock_page(page);
        put_page(page);
        goto out;
@@ -1944,6 +2068,7 @@ 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);
 
@@ -1960,30 +2085,79 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        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, huge_ptep_get(ptep))))
-               if (write_access && !pte_write(entry)) {
-                       struct page *page;
-                       page = hugetlbfs_pagecache_page(h, vma, address);
-                       ret = hugetlb_cow(mm, vma, address, ptep, entry, page);
-                       if (page) {
-                               unlock_page(page);
-                               put_page(page);
-                       }
+       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,
@@ -1993,6 +2167,8 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
        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) {
@@ -2005,8 +2181,11 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
                 * first, for the page indexing below to work.
                 */
                pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
+               if (huge_zeropage_ok(pte, write, shared))
+                       zeropage_ok = 1;
 
-               if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
+               if (!pte ||
+                   (huge_pte_none(huge_ptep_get(pte)) && !zeropage_ok) ||
                    (write && !pte_write(huge_ptep_get(pte)))) {
                        int ret;
 
@@ -2026,8 +2205,11 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
                page = pte_page(huge_ptep_get(pte));
 same_page:
                if (pages) {
-                       get_page(page);
-                       pages[i] = page + pfn_offset;
+                       if (zeropage_ok)
+                               pages[i] = ZERO_PAGE(0);
+                       else
+                               pages[i] = mem_map_offset(page, pfn_offset);
+                       get_page(pages[i]);
                }
 
                if (vmas)
@@ -2087,12 +2269,18 @@ void hugetlb_change_protection(struct vm_area_struct *vma,
 
 int hugetlb_reserve_pages(struct inode *inode,
                                        long from, long to,
-                                       struct vm_area_struct *vma)
+                                       struct vm_area_struct *vma,
+                                       int acctflag)
 {
        long ret, chg;
        struct hstate *h = hstate_inode(inode);
 
-       if (vma && vma->vm_flags & VM_NORESERVE)
+       /*
+        * Only apply hugepage reservation if asked. At fault time, an
+        * attempt will be made for VM_NORESERVE to allocate a page
+        * and filesystem quota without using reserves
+        */
+       if (acctflag & VM_NORESERVE)
                return 0;
 
        /*
@@ -2101,7 +2289,7 @@ int hugetlb_reserve_pages(struct inode *inode,
         * to reserve the full area even if read-only as mprotect() may be
         * called to make the mapping read-write. Assume !vma is a shm mapping
         */
-       if (!vma || vma->vm_flags & VM_SHARED)
+       if (!vma || vma->vm_flags & VM_MAYSHARE)
                chg = region_chg(&inode->i_mapping->private_list, from, to);
        else {
                struct resv_map *resv_map = resv_map_alloc();
@@ -2117,14 +2305,32 @@ int hugetlb_reserve_pages(struct inode *inode,
        if (chg < 0)
                return chg;
 
+       /* There must be enough filesystem quota for the mapping */
        if (hugetlb_get_quota(inode->i_mapping, chg))
                return -ENOSPC;
+
+       /*
+        * Check enough hugepages are available for the reservation.
+        * Hand back the quota if there are not
+        */
        ret = hugetlb_acct_memory(h, chg);
        if (ret < 0) {
                hugetlb_put_quota(inode->i_mapping, chg);
                return ret;
        }
-       if (!vma || vma->vm_flags & VM_SHARED)
+
+       /*
+        * Account for the reservations made. Shared mappings record regions
+        * that have reservations as they are shared by multiple VMAs.
+        * When the last VMA disappears, the region map says how much
+        * the reservation was and the page cache tells how much of
+        * the reservation was consumed. Private mappings are per-VMA and
+        * only the consumed reservations are tracked. When the VMA
+        * disappears, the original reservation is the VMA size and the
+        * consumed reservations are stored in the map. Hence, nothing
+        * else has to be done for private mappings here
+        */
+       if (!vma || vma->vm_flags & VM_MAYSHARE)
                region_add(&inode->i_mapping->private_list, from, to);
        return 0;
 }