[PATCH] hugepage: Strict page reservation for hugepage inodes
authorDavid Gibson <david@gibson.dropbear.id.au>
Wed, 22 Mar 2006 08:08:55 +0000 (00:08 -0800)
committerLinus Torvalds <torvalds@g5.osdl.org>
Wed, 22 Mar 2006 15:54:03 +0000 (07:54 -0800)
These days, hugepages are demand-allocated at first fault time.  There's a
somewhat dubious (and racy) heuristic when making a new mmap() to check if
there are enough available hugepages to fully satisfy that mapping.

A particularly obvious case where the heuristic breaks down is where a
process maps its hugepages not as a single chunk, but as a bunch of
individually mmap()ed (or shmat()ed) blocks without touching and
instantiating the pages in between allocations.  In this case the size of
each block is compared against the total number of available hugepages.
It's thus easy for the process to become overcommitted, because each block
mapping will succeed, although the total number of hugepages required by
all blocks exceeds the number available.  In particular, this defeats such
a program which will detect a mapping failure and adjust its hugepage usage
downward accordingly.

The patch below addresses this problem, by strictly reserving a number of
physical hugepages for hugepage inodes which have been mapped, but not
instatiated.  MAP_SHARED mappings are thus "safe" - they will fail on
mmap(), not later with an OOM SIGKILL.  MAP_PRIVATE mappings can still
trigger an OOM.  (Actually SHARED mappings can technically still OOM, but
only if the sysadmin explicitly reduces the hugepage pool between mapping
and instantiation)

This patch appears to address the problem at hand - it allows DB2 to start
correctly, for instance, which previously suffered the failure described
above.

This patch causes no regressions on the libhugetblfs testsuite, and makes a
test (designed to catch this problem) pass which previously failed (ppc64,
POWER5).

Signed-off-by: David Gibson <dwg@au1.ibm.com>
Cc: William Lee Irwin III <wli@holomorphy.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
fs/hugetlbfs/inode.c
include/linux/hugetlb.h
mm/hugetlb.c

index b351952899453be64165ba7315e5f59bd0c851ac..1a1c2fcb782329bbce007e8dc8f8af592e5176be 100644 (file)
@@ -56,48 +56,10 @@ static void huge_pagevec_release(struct pagevec *pvec)
        pagevec_reinit(pvec);
 }
 
-/*
- * huge_pages_needed tries to determine the number of new huge pages that
- * will be required to fully populate this VMA.  This will be equal to
- * the size of the VMA in huge pages minus the number of huge pages
- * (covered by this VMA) that are found in the page cache.
- *
- * Result is in bytes to be compatible with is_hugepage_mem_enough()
- */
-static unsigned long
-huge_pages_needed(struct address_space *mapping, struct vm_area_struct *vma)
-{
-       int i;
-       struct pagevec pvec;
-       unsigned long start = vma->vm_start;
-       unsigned long end = vma->vm_end;
-       unsigned long hugepages = (end - start) >> HPAGE_SHIFT;
-       pgoff_t next = vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT);
-       pgoff_t endpg = next + hugepages;
-
-       pagevec_init(&pvec, 0);
-       while (next < endpg) {
-               if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE))
-                       break;
-               for (i = 0; i < pagevec_count(&pvec); i++) {
-                       struct page *page = pvec.pages[i];
-                       if (page->index > next)
-                               next = page->index;
-                       if (page->index >= endpg)
-                               break;
-                       next++;
-                       hugepages--;
-               }
-               huge_pagevec_release(&pvec);
-       }
-       return hugepages << HPAGE_SHIFT;
-}
-
 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 {
        struct inode *inode = file->f_dentry->d_inode;
-       struct address_space *mapping = inode->i_mapping;
-       unsigned long bytes;
+       struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
        loff_t len, vma_len;
        int ret;
 
@@ -113,10 +75,6 @@ static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
        if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
                return -EINVAL;
 
-       bytes = huge_pages_needed(mapping, vma);
-       if (!is_hugepage_mem_enough(bytes))
-               return -ENOMEM;
-
        vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 
        mutex_lock(&inode->i_mutex);
@@ -129,6 +87,10 @@ static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
        if (!(vma->vm_flags & VM_WRITE) && len > inode->i_size)
                goto out;
 
+       if (vma->vm_flags & VM_MAYSHARE)
+               if (hugetlb_extend_reservation(info, len >> HPAGE_SHIFT) != 0)
+                       goto out;
+
        ret = 0;
        hugetlb_prefault_arch_hook(vma->vm_mm);
        if (inode->i_size < len)
@@ -227,13 +189,18 @@ static void truncate_huge_page(struct page *page)
        put_page(page);
 }
 
-static void truncate_hugepages(struct address_space *mapping, loff_t lstart)
+static void truncate_hugepages(struct inode *inode, loff_t lstart)
 {
+       struct address_space *mapping = &inode->i_data;
        const pgoff_t start = lstart >> HPAGE_SHIFT;
        struct pagevec pvec;
        pgoff_t next;
        int i;
 
+       hugetlb_truncate_reservation(HUGETLBFS_I(inode),
+                                    lstart >> HPAGE_SHIFT);
+       if (!mapping->nrpages)
+               return;
        pagevec_init(&pvec, 0);
        next = start;
        while (1) {
@@ -262,8 +229,7 @@ static void truncate_hugepages(struct address_space *mapping, loff_t lstart)
 
 static void hugetlbfs_delete_inode(struct inode *inode)
 {
-       if (inode->i_data.nrpages)
-               truncate_hugepages(&inode->i_data, 0);
+       truncate_hugepages(inode, 0);
        clear_inode(inode);
 }
 
@@ -296,8 +262,7 @@ static void hugetlbfs_forget_inode(struct inode *inode)
        inode->i_state |= I_FREEING;
        inodes_stat.nr_inodes--;
        spin_unlock(&inode_lock);
-       if (inode->i_data.nrpages)
-               truncate_hugepages(&inode->i_data, 0);
+       truncate_hugepages(inode, 0);
        clear_inode(inode);
        destroy_inode(inode);
 }
@@ -356,7 +321,7 @@ static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
        if (!prio_tree_empty(&mapping->i_mmap))
                hugetlb_vmtruncate_list(&mapping->i_mmap, pgoff);
        spin_unlock(&mapping->i_mmap_lock);
-       truncate_hugepages(mapping, offset);
+       truncate_hugepages(inode, offset);
        return 0;
 }
 
@@ -573,6 +538,7 @@ static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
                hugetlbfs_inc_free_inodes(sbinfo);
                return NULL;
        }
+       p->prereserved_hpages = 0;
        return &p->vfs_inode;
 }
 
@@ -805,9 +771,6 @@ struct file *hugetlb_zero_setup(size_t size)
        if (!can_do_hugetlb_shm())
                return ERR_PTR(-EPERM);
 
-       if (!is_hugepage_mem_enough(size))
-               return ERR_PTR(-ENOMEM);
-
        if (!user_shm_lock(size, current->user))
                return ERR_PTR(-ENOMEM);
 
@@ -831,6 +794,11 @@ struct file *hugetlb_zero_setup(size_t size)
        if (!inode)
                goto out_file;
 
+       error = -ENOMEM;
+       if (hugetlb_extend_reservation(HUGETLBFS_I(inode),
+                                      size >> HPAGE_SHIFT) != 0)
+               goto out_inode;
+
        d_instantiate(dentry, inode);
        inode->i_size = size;
        inode->i_nlink = 0;
@@ -841,6 +809,8 @@ struct file *hugetlb_zero_setup(size_t size)
        file->f_mode = FMODE_WRITE | FMODE_READ;
        return file;
 
+out_inode:
+       iput(inode);
 out_file:
        put_filp(file);
 out_dentry:
index fa83836b63d2098e0404a69c74d30d15f8402e6c..cafe73eecb056e64bdf21f589e7f4b7e4ceca9ed 100644 (file)
@@ -20,7 +20,6 @@ void unmap_hugepage_range(struct vm_area_struct *, unsigned long, unsigned long)
 int hugetlb_prefault(struct address_space *, struct vm_area_struct *);
 int hugetlb_report_meminfo(char *);
 int hugetlb_report_node_meminfo(int, char *);
-int is_hugepage_mem_enough(size_t);
 unsigned long hugetlb_total_pages(void);
 struct page *alloc_huge_page(struct vm_area_struct *, unsigned long);
 void free_huge_page(struct page *);
@@ -89,7 +88,6 @@ static inline unsigned long hugetlb_total_pages(void)
 #define copy_hugetlb_page_range(src, dst, vma) ({ BUG(); 0; })
 #define hugetlb_prefault(mapping, vma)         ({ BUG(); 0; })
 #define unmap_hugepage_range(vma, start, end)  BUG()
-#define is_hugepage_mem_enough(size)           0
 #define hugetlb_report_meminfo(buf)            0
 #define hugetlb_report_node_meminfo(n, buf)    0
 #define follow_huge_pmd(mm, addr, pmd, write)  NULL
@@ -132,6 +130,8 @@ struct hugetlbfs_sb_info {
 
 struct hugetlbfs_inode_info {
        struct shared_policy policy;
+       /* Protected by the (global) hugetlb_lock */
+       unsigned long prereserved_hpages;
        struct inode vfs_inode;
 };
 
@@ -148,6 +148,10 @@ static inline struct hugetlbfs_sb_info *HUGETLBFS_SB(struct super_block *sb)
 extern struct file_operations hugetlbfs_file_operations;
 extern struct vm_operations_struct hugetlb_vm_ops;
 struct file *hugetlb_zero_setup(size_t);
+int hugetlb_extend_reservation(struct hugetlbfs_inode_info *info,
+                              unsigned long atleast_hpages);
+void hugetlb_truncate_reservation(struct hugetlbfs_inode_info *info,
+                                 unsigned long atmost_hpages);
 int hugetlb_get_quota(struct address_space *mapping);
 void hugetlb_put_quota(struct address_space *mapping);
 
index d5987a87bbe536306a076796cc486a9078df3fe6..27fad5d9bcf648b27564465b35c3cd72969257fa 100644 (file)
@@ -22,7 +22,7 @@
 #include "internal.h"
 
 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
-static unsigned long nr_huge_pages, free_huge_pages;
+static unsigned long nr_huge_pages, free_huge_pages, reserved_huge_pages;
 unsigned long max_huge_pages;
 static struct list_head hugepage_freelists[MAX_NUMNODES];
 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
@@ -120,17 +120,136 @@ void free_huge_page(struct page *page)
 
 struct page *alloc_huge_page(struct vm_area_struct *vma, unsigned long addr)
 {
+       struct inode *inode = vma->vm_file->f_dentry->d_inode;
        struct page *page;
+       int use_reserve = 0;
+       unsigned long idx;
 
        spin_lock(&hugetlb_lock);
-       page = dequeue_huge_page(vma, addr);
-       if (!page) {
-               spin_unlock(&hugetlb_lock);
-               return NULL;
+
+       if (vma->vm_flags & VM_MAYSHARE) {
+
+               /* idx = radix tree index, i.e. offset into file in
+                * HPAGE_SIZE units */
+               idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
+                       + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
+
+               /* The hugetlbfs specific inode info stores the number
+                * of "guaranteed available" (huge) pages.  That is,
+                * the first 'prereserved_hpages' pages of the inode
+                * are either already instantiated, or have been
+                * pre-reserved (by hugetlb_reserve_for_inode()). Here
+                * we're in the process of instantiating the page, so
+                * we use this to determine whether to draw from the
+                * pre-reserved pool or the truly free pool. */
+               if (idx < HUGETLBFS_I(inode)->prereserved_hpages)
+                       use_reserve = 1;
+       }
+
+       if (!use_reserve) {
+               if (free_huge_pages <= reserved_huge_pages)
+                       goto fail;
+       } else {
+               BUG_ON(reserved_huge_pages == 0);
+               reserved_huge_pages--;
        }
+
+       page = dequeue_huge_page(vma, addr);
+       if (!page)
+               goto fail;
+
        spin_unlock(&hugetlb_lock);
        set_page_refcounted(page);
        return page;
+
+ fail:
+       WARN_ON(use_reserve); /* reserved allocations shouldn't fail */
+       spin_unlock(&hugetlb_lock);
+       return NULL;
+}
+
+/* hugetlb_extend_reservation()
+ *
+ * Ensure that at least 'atleast' hugepages are, and will remain,
+ * available to instantiate the first 'atleast' pages of the given
+ * inode.  If the inode doesn't already have this many pages reserved
+ * or instantiated, set aside some hugepages in the reserved pool to
+ * satisfy later faults (or fail now if there aren't enough, rather
+ * than getting the SIGBUS later).
+ */
+int hugetlb_extend_reservation(struct hugetlbfs_inode_info *info,
+                              unsigned long atleast)
+{
+       struct inode *inode = &info->vfs_inode;
+       unsigned long change_in_reserve = 0;
+       int ret = 0;
+
+       spin_lock(&hugetlb_lock);
+       read_lock_irq(&inode->i_mapping->tree_lock);
+
+       if (info->prereserved_hpages >= atleast)
+               goto out;
+
+       /* Because we always call this on shared mappings, none of the
+        * pages beyond info->prereserved_hpages can have been
+        * instantiated, so we need to reserve all of them now. */
+       change_in_reserve = atleast - info->prereserved_hpages;
+
+       if ((reserved_huge_pages + change_in_reserve) > free_huge_pages) {
+               ret = -ENOMEM;
+               goto out;
+       }
+
+       reserved_huge_pages += change_in_reserve;
+       info->prereserved_hpages = atleast;
+
+ out:
+       read_unlock_irq(&inode->i_mapping->tree_lock);
+       spin_unlock(&hugetlb_lock);
+
+       return ret;
+}
+
+/* hugetlb_truncate_reservation()
+ *
+ * This returns pages reserved for the given inode to the general free
+ * hugepage pool.  If the inode has any pages prereserved, but not
+ * instantiated, beyond offset (atmost << HPAGE_SIZE), then release
+ * them.
+ */
+void hugetlb_truncate_reservation(struct hugetlbfs_inode_info *info,
+                                 unsigned long atmost)
+{
+       struct inode *inode = &info->vfs_inode;
+       struct address_space *mapping = inode->i_mapping;
+       unsigned long idx;
+       unsigned long change_in_reserve = 0;
+       struct page *page;
+
+       spin_lock(&hugetlb_lock);
+       read_lock_irq(&inode->i_mapping->tree_lock);
+
+       if (info->prereserved_hpages <= atmost)
+               goto out;
+
+       /* Count pages which were reserved, but not instantiated, and
+        * which we can now release. */
+       for (idx = atmost; idx < info->prereserved_hpages; idx++) {
+               page = radix_tree_lookup(&mapping->page_tree, idx);
+               if (!page)
+                       /* Pages which are already instantiated can't
+                        * be unreserved (and in fact have already
+                        * been removed from the reserved pool) */
+                       change_in_reserve++;
+       }
+
+       BUG_ON(reserved_huge_pages < change_in_reserve);
+       reserved_huge_pages -= change_in_reserve;
+       info->prereserved_hpages = atmost;
+
+ out:
+       read_unlock_irq(&inode->i_mapping->tree_lock);
+       spin_unlock(&hugetlb_lock);
 }
 
 static int __init hugetlb_init(void)
@@ -238,9 +357,11 @@ int hugetlb_report_meminfo(char *buf)
        return sprintf(buf,
                        "HugePages_Total: %5lu\n"
                        "HugePages_Free:  %5lu\n"
+                       "HugePages_Rsvd:  %5lu\n"
                        "Hugepagesize:    %5lu kB\n",
                        nr_huge_pages,
                        free_huge_pages,
+                       reserved_huge_pages,
                        HPAGE_SIZE/1024);
 }
 
@@ -253,11 +374,6 @@ int hugetlb_report_node_meminfo(int nid, char *buf)
                nid, free_huge_pages_node[nid]);
 }
 
-int is_hugepage_mem_enough(size_t size)
-{
-       return (size + ~HPAGE_MASK)/HPAGE_SIZE <= free_huge_pages;
-}
-
 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 unsigned long hugetlb_total_pages(void)
 {