mm: account for MAP_SHARED mappings using VM_MAYSHARE and not VM_SHARED in hugetlbfs
[linux-2.6.git] / mm / hugetlb.c
index 4cf7a90..e83ad2c 100644 (file)
@@ -7,16 +7,21 @@
 #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"
@@ -29,9 +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)++)
@@ -212,6 +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.
@@ -255,7 +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)
@@ -267,7 +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);
 
@@ -279,16 +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);
@@ -297,7 +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);
 }
@@ -316,7 +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)) {
@@ -332,25 +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();
@@ -358,12 +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();
@@ -415,7 +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;
 
@@ -448,6 +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++) {
@@ -488,7 +566,7 @@ static void free_huge_page(struct page *page)
        INIT_LIST_HEAD(&page->lru);
 
        spin_lock(&hugetlb_lock);
-       if (h->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]--;
@@ -549,13 +627,16 @@ 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_REPEAT|__GFP_NOWARN,
                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);
@@ -564,6 +645,27 @@ static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
        return page;
 }
 
+/*
+ * 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;
@@ -577,21 +679,7 @@ static int alloc_fresh_huge_page(struct hstate *h)
                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(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;
+               next_nid = hstate_next_node(h);
        } while (!page && h->hugetlb_next_nid != start_nid);
 
        if (ret)
@@ -608,6 +696,9 @@ static struct page *alloc_buddy_huge_page(struct hstate *h,
        struct page *page;
        unsigned int nid;
 
+       if (h->order >= MAX_ORDER)
+               return NULL;
+
        /*
         * Assume we will successfully allocate the surplus page to
         * prevent racing processes from causing the surplus to exceed
@@ -645,6 +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) {
                /*
@@ -784,6 +880,10 @@ static void return_unused_surplus_pages(struct hstate *h,
        /* 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 (remaining_iterations-- && nr_pages) {
@@ -818,13 +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);
@@ -833,7 +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);
 
@@ -849,7 +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);
 
@@ -869,7 +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
@@ -905,115 +1005,120 @@ static struct page *alloc_huge_page(struct vm_area_struct *vma,
        return page;
 }
 
-static void __init hugetlb_init_one_hstate(struct hstate *h)
+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);
 
-       for (i = 0; i < MAX_NUMNODES; ++i)
-               INIT_LIST_HEAD(&h->hugepage_freelists[i]);
+       while (nr_nodes) {
+               void *addr;
 
-       h->hugetlb_next_nid = first_node(node_online_map);
+               addr = __alloc_bootmem_node_nopanic(
+                               NODE_DATA(h->hugetlb_next_nid),
+                               huge_page_size(h), huge_page_size(h), 0);
 
-       for (i = 0; i < h->max_huge_pages; ++i) {
-               if (!alloc_fresh_huge_page(h))
-                       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--;
        }
-       h->max_huge_pages = h->free_huge_pages = h->nr_huge_pages = i;
+       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;
 }
 
-static void __init hugetlb_init_hstates(void)
+static void prep_compound_huge_page(struct page *page, int order)
 {
-       struct hstate *h;
-
-       for_each_hstate(h) {
-               hugetlb_init_one_hstate(h);
-       }
+       if (unlikely(order > (MAX_ORDER - 1)))
+               prep_compound_gigantic_page(page, order);
+       else
+               prep_compound_page(page, order);
 }
 
-static void __init report_hugepages(void)
+/* Put bootmem huge pages into the standard lists after mem_map is up */
+static void __init gather_bootmem_prealloc(void)
 {
-       struct hstate *h;
+       struct huge_bootmem_page *m;
 
-       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));
+       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));
        }
 }
 
-static int __init hugetlb_init(void)
+static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
 {
-       BUILD_BUG_ON(HPAGE_SHIFT == 0);
+       unsigned long i;
 
-       if (!size_to_hstate(HPAGE_SIZE)) {
-               hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
-               parsed_hstate->max_huge_pages = default_hstate_max_huge_pages;
+       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;
        }
-       default_hstate_idx = size_to_hstate(HPAGE_SIZE) - hstates;
-
-       hugetlb_init_hstates();
-
-       report_hugepages();
-
-       return 0;
+       h->max_huge_pages = i;
 }
-module_init(hugetlb_init);
 
-/* Should be called on processing a hugepagesz=... option */
-void __init hugetlb_add_hstate(unsigned order)
+static void __init hugetlb_init_hstates(void)
 {
        struct hstate *h;
-       if (size_to_hstate(PAGE_SIZE << order)) {
-               printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
-               return;
+
+       for_each_hstate(h) {
+               /* oversize hugepages were init'ed in early boot */
+               if (h->order < MAX_ORDER)
+                       hugetlb_hstate_alloc_pages(h);
        }
-       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);
-       hugetlb_init_one_hstate(h);
-       parsed_hstate = h;
 }
 
-static int __init hugetlb_setup(char *s)
+static char * __init memfmt(char *buf, unsigned long n)
 {
-       unsigned long *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;
+       if (n >= (1UL << 30))
+               sprintf(buf, "%lu GB", n >> 30);
+       else if (n >= (1UL << 20))
+               sprintf(buf, "%lu MB", n >> 20);
        else
-               mhp = &parsed_hstate->max_huge_pages;
-
-       if (sscanf(s, "%lu", mhp) <= 0)
-               *mhp = 0;
-
-       return 1;
+               sprintf(buf, "%lu KB", n >> 10);
+       return buf;
 }
-__setup("hugepages=", hugetlb_setup);
 
-static unsigned int cpuset_mems_nr(unsigned int *array)
+static void __init report_hugepages(void)
 {
-       int node;
-       unsigned int nr = 0;
-
-       for_each_node_mask(node, cpuset_current_mems_allowed)
-               nr += array[node];
+       struct hstate *h;
 
-       return nr;
+       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(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;
                struct list_head *freel = &h->hugepage_freelists[i];
@@ -1040,6 +1145,9 @@ 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
@@ -1105,6 +1213,269 @@ out:
        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)
@@ -1162,15 +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,
@@ -1271,8 +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);
+               }
        }
 }
 
@@ -1389,6 +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);
@@ -1430,6 +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);
@@ -1450,11 +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;
@@ -1464,7 +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);
@@ -1483,7 +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);
        }
 
@@ -1519,7 +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;
@@ -1626,7 +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;
 
@@ -1645,6 +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)
@@ -1670,6 +2056,7 @@ out:
 
 backout:
        spin_unlock(&mm->page_table_lock);
+backout_unlocked:
        unlock_page(page);
        put_page(page);
        goto out;
@@ -1681,6 +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);
 
@@ -1697,30 +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,
@@ -1730,6 +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) {
@@ -1742,8 +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;
 
@@ -1763,8 +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)
@@ -1824,12 +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;
 
        /*
@@ -1838,7 +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();
@@ -1854,14 +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;
 }