tmpfs: pass gfp to shmem_getpage_gfp
authorHugh Dickins <hughd@google.com>
Tue, 26 Jul 2011 00:12:34 +0000 (17:12 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 26 Jul 2011 03:57:11 +0000 (20:57 -0700)
Make shmem_getpage() a wrapper, passing mapping_gfp_mask() down to
shmem_getpage_gfp(), which in turn passes gfp down to shmem_swp_alloc().

Change shmem_read_mapping_page_gfp() to use shmem_getpage_gfp() in the
CONFIG_SHMEM case; but leave tiny !SHMEM using read_cache_page_gfp().

Add a BUG_ON() in case anyone happens to call this on a non-shmem mapping;
though we might later want to let that case route to read_cache_page_gfp().

It annoys me to have these two almost-redundant args, gfp and fault_type:
I can't find a better way; but initialize fault_type only in shmem_fault().

Note that before, read_cache_page_gfp() was allocating i915_gem's pages
with __GFP_NORETRY as intended; but the corresponding swap vector pages
got allocated without it, leaving a small possibility of OOM.

Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/shmem.c

index f96614526d1c6d367d2883d48d4015ea13645f25..f6c94ba87808794eef9bae656659c66ab542c066 100644 (file)
@@ -127,8 +127,15 @@ static unsigned long shmem_default_max_inodes(void)
 }
 #endif
 
-static int shmem_getpage(struct inode *inode, unsigned long idx,
-                        struct page **pagep, enum sgp_type sgp, int *type);
+static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
+       struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
+
+static inline int shmem_getpage(struct inode *inode, pgoff_t index,
+       struct page **pagep, enum sgp_type sgp, int *fault_type)
+{
+       return shmem_getpage_gfp(inode, index, pagep, sgp,
+                       mapping_gfp_mask(inode->i_mapping), fault_type);
+}
 
 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
 {
@@ -404,10 +411,12 @@ static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, uns
  * @info:      info structure for the inode
  * @index:     index of the page to find
  * @sgp:       check and recheck i_size? skip allocation?
+ * @gfp:       gfp mask to use for any page allocation
  *
  * If the entry does not exist, allocate it.
  */
-static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
+static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info,
+                       unsigned long index, enum sgp_type sgp, gfp_t gfp)
 {
        struct inode *inode = &info->vfs_inode;
        struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
@@ -435,7 +444,7 @@ static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long
                }
 
                spin_unlock(&info->lock);
-               page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
+               page = shmem_dir_alloc(gfp);
                spin_lock(&info->lock);
 
                if (!page) {
@@ -1225,14 +1234,14 @@ static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 #endif
 
 /*
- * shmem_getpage - either get the page from swap or allocate a new one
+ * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
  *
  * If we allocate a new one we do not mark it dirty. That's up to the
  * vm. If we swap it in we mark it dirty since we also free the swap
  * entry since a page cannot live in both the swap and page cache
  */
-static int shmem_getpage(struct inode *inode, unsigned long idx,
-                       struct page **pagep, enum sgp_type sgp, int *type)
+static int shmem_getpage_gfp(struct inode *inode, pgoff_t idx,
+       struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
 {
        struct address_space *mapping = inode->i_mapping;
        struct shmem_inode_info *info = SHMEM_I(inode);
@@ -1242,15 +1251,11 @@ static int shmem_getpage(struct inode *inode, unsigned long idx,
        struct page *prealloc_page = NULL;
        swp_entry_t *entry;
        swp_entry_t swap;
-       gfp_t gfp;
        int error;
 
        if (idx >= SHMEM_MAX_INDEX)
                return -EFBIG;
 
-       if (type)
-               *type = 0;
-
        /*
         * Normally, filepage is NULL on entry, and either found
         * uptodate immediately, or allocated and zeroed, or read
@@ -1264,13 +1269,12 @@ repeat:
                filepage = find_lock_page(mapping, idx);
        if (filepage && PageUptodate(filepage))
                goto done;
-       gfp = mapping_gfp_mask(mapping);
        if (!filepage) {
                /*
                 * Try to preload while we can wait, to not make a habit of
                 * draining atomic reserves; but don't latch on to this cpu.
                 */
-               error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
+               error = radix_tree_preload(gfp & GFP_RECLAIM_MASK);
                if (error)
                        goto failed;
                radix_tree_preload_end();
@@ -1290,7 +1294,7 @@ repeat:
 
        spin_lock(&info->lock);
        shmem_recalc_inode(inode);
-       entry = shmem_swp_alloc(info, idx, sgp);
+       entry = shmem_swp_alloc(info, idx, sgp, gfp);
        if (IS_ERR(entry)) {
                spin_unlock(&info->lock);
                error = PTR_ERR(entry);
@@ -1305,12 +1309,12 @@ repeat:
                        shmem_swp_unmap(entry);
                        spin_unlock(&info->lock);
                        /* here we actually do the io */
-                       if (type)
-                               *type |= VM_FAULT_MAJOR;
+                       if (fault_type)
+                               *fault_type |= VM_FAULT_MAJOR;
                        swappage = shmem_swapin(swap, gfp, info, idx);
                        if (!swappage) {
                                spin_lock(&info->lock);
-                               entry = shmem_swp_alloc(info, idx, sgp);
+                               entry = shmem_swp_alloc(info, idx, sgp, gfp);
                                if (IS_ERR(entry))
                                        error = PTR_ERR(entry);
                                else {
@@ -1461,7 +1465,7 @@ repeat:
                                SetPageSwapBacked(filepage);
                        }
 
-                       entry = shmem_swp_alloc(info, idx, sgp);
+                       entry = shmem_swp_alloc(info, idx, sgp, gfp);
                        if (IS_ERR(entry))
                                error = PTR_ERR(entry);
                        else {
@@ -1539,7 +1543,7 @@ static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 {
        struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
        int error;
-       int ret;
+       int ret = VM_FAULT_LOCKED;
 
        if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
                return VM_FAULT_SIGBUS;
@@ -1547,11 +1551,12 @@ static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
        error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
        if (error)
                return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
+
        if (ret & VM_FAULT_MAJOR) {
                count_vm_event(PGMAJFAULT);
                mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
        }
-       return ret | VM_FAULT_LOCKED;
+       return ret;
 }
 
 #ifdef CONFIG_NUMA
@@ -3162,13 +3167,29 @@ int shmem_zero_setup(struct vm_area_struct *vma)
  * suit tmpfs, since it may have pages in swapcache, and needs to find those
  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
  *
- * Provide a stub for those callers to start using now, then later
- * flesh it out to call shmem_getpage() with additional gfp mask, when
- * shmem_file_splice_read() is added and shmem_readpage() is removed.
+ * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
+ * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
  */
 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
                                         pgoff_t index, gfp_t gfp)
 {
+#ifdef CONFIG_SHMEM
+       struct inode *inode = mapping->host;
+       struct page *page = NULL;
+       int error;
+
+       BUG_ON(mapping->a_ops != &shmem_aops);
+       error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
+       if (error)
+               page = ERR_PTR(error);
+       else
+               unlock_page(page);
+       return page;
+#else
+       /*
+        * The tiny !SHMEM case uses ramfs without swap
+        */
        return read_cache_page_gfp(mapping, index, gfp);
+#endif
 }
 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);