ext4: dynamically allocate the jbd2_inode in ext4_inode_info as necessary
Theodore Ts'o [Mon, 10 Jan 2011 17:29:43 +0000 (12:29 -0500)]
Replace the jbd2_inode structure (which is 48 bytes) with a pointer
and only allocate the jbd2_inode when it is needed --- that is, when
the file system has a journal present and the inode has been opened
for writing.  This allows us to further slim down the ext4_inode_info
structure.

Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>

fs/ext4/ext4.h
fs/ext4/ext4_jbd2.h
fs/ext4/file.c
fs/ext4/inode.c
fs/ext4/super.c
fs/jbd2/journal.c
include/linux/jbd2.h

index 2fb531c..32b7daa 100644 (file)
@@ -811,7 +811,7 @@ struct ext4_inode_info {
         */
        struct rw_semaphore i_data_sem;
        struct inode vfs_inode;
-       struct jbd2_inode jinode;
+       struct jbd2_inode *jinode;
 
        struct ext4_ext_cache i_cached_extent;
        /*
index b0bd792..d8b992e 100644 (file)
@@ -253,7 +253,7 @@ static inline int ext4_journal_force_commit(journal_t *journal)
 static inline int ext4_jbd2_file_inode(handle_t *handle, struct inode *inode)
 {
        if (ext4_handle_valid(handle))
-               return jbd2_journal_file_inode(handle, &EXT4_I(inode)->jinode);
+               return jbd2_journal_file_inode(handle, EXT4_I(inode)->jinode);
        return 0;
 }
 
index 5a5c55d..bb003dc 100644 (file)
@@ -104,6 +104,7 @@ static int ext4_file_open(struct inode * inode, struct file * filp)
 {
        struct super_block *sb = inode->i_sb;
        struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
+       struct ext4_inode_info *ei = EXT4_I(inode);
        struct vfsmount *mnt = filp->f_path.mnt;
        struct path path;
        char buf[64], *cp;
@@ -127,6 +128,27 @@ static int ext4_file_open(struct inode * inode, struct file * filp)
                        ext4_mark_super_dirty(sb);
                }
        }
+       /*
+        * Set up the jbd2_inode if we are opening the inode for
+        * writing and the journal is present
+        */
+       if (sbi->s_journal && !ei->jinode && (filp->f_mode & FMODE_WRITE)) {
+               struct jbd2_inode *jinode = jbd2_alloc_inode(GFP_KERNEL);
+
+               spin_lock(&inode->i_lock);
+               if (!ei->jinode) {
+                       if (!jinode) {
+                               spin_unlock(&inode->i_lock);
+                               return -ENOMEM;
+                       }
+                       ei->jinode = jinode;
+                       jbd2_journal_init_jbd_inode(ei->jinode, inode);
+                       jinode = NULL;
+               }
+               spin_unlock(&inode->i_lock);
+               if (unlikely(jinode != NULL))
+                       jbd2_free_inode(jinode);
+       }
        return dquot_file_open(inode, filp);
 }
 
index 0801ee6..2693fcd 100644 (file)
@@ -55,10 +55,17 @@ static inline int ext4_begin_ordered_truncate(struct inode *inode,
                                              loff_t new_size)
 {
        trace_ext4_begin_ordered_truncate(inode, new_size);
-       return jbd2_journal_begin_ordered_truncate(
-                                       EXT4_SB(inode->i_sb)->s_journal,
-                                       &EXT4_I(inode)->jinode,
-                                       new_size);
+       /*
+        * If jinode is zero, then we never opened the file for
+        * writing, so there's no need to call
+        * jbd2_journal_begin_ordered_truncate() since there's no
+        * outstanding writes we need to flush.
+        */
+       if (!EXT4_I(inode)->jinode)
+               return 0;
+       return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
+                                                  EXT4_I(inode)->jinode,
+                                                  new_size);
 }
 
 static void ext4_invalidatepage(struct page *page, unsigned long offset);
@@ -4054,7 +4061,7 @@ int ext4_block_truncate_page(handle_t *handle,
        if (ext4_should_journal_data(inode)) {
                err = ext4_handle_dirty_metadata(handle, inode, bh);
        } else {
-               if (ext4_should_order_data(inode))
+               if (ext4_should_order_data(inode) && EXT4_I(inode)->jinode)
                        err = ext4_jbd2_file_inode(handle, inode);
                mark_buffer_dirty(bh);
        }
index f5960d6..1cd4326 100644 (file)
@@ -818,12 +818,6 @@ static struct inode *ext4_alloc_inode(struct super_block *sb)
        memset(&ei->i_cached_extent, 0, sizeof(struct ext4_ext_cache));
        INIT_LIST_HEAD(&ei->i_prealloc_list);
        spin_lock_init(&ei->i_prealloc_lock);
-       /*
-        * Note:  We can be called before EXT4_SB(sb)->s_journal is set,
-        * therefore it can be null here.  Don't check it, just initialize
-        * jinode.
-        */
-       jbd2_journal_init_jbd_inode(&ei->jinode, &ei->vfs_inode);
        ei->i_reserved_data_blocks = 0;
        ei->i_reserved_meta_blocks = 0;
        ei->i_allocated_meta_blocks = 0;
@@ -832,6 +826,7 @@ static struct inode *ext4_alloc_inode(struct super_block *sb)
 #ifdef CONFIG_QUOTA
        ei->i_reserved_quota = 0;
 #endif
+       ei->jinode = NULL;
        INIT_LIST_HEAD(&ei->i_completed_io_list);
        spin_lock_init(&ei->i_completed_io_lock);
        ei->cur_aio_dio = NULL;
@@ -900,9 +895,12 @@ void ext4_clear_inode(struct inode *inode)
        end_writeback(inode);
        dquot_drop(inode);
        ext4_discard_preallocations(inode);
-       if (EXT4_JOURNAL(inode))
-               jbd2_journal_release_jbd_inode(EXT4_SB(inode->i_sb)->s_journal,
-                                      &EXT4_I(inode)->jinode);
+       if (EXT4_I(inode)->jinode) {
+               jbd2_journal_release_jbd_inode(EXT4_JOURNAL(inode),
+                                              EXT4_I(inode)->jinode);
+               jbd2_free_inode(EXT4_I(inode)->jinode);
+               EXT4_I(inode)->jinode = NULL;
+       }
 }
 
 static inline void ext4_show_quota_options(struct seq_file *seq,
index 2447bd8..9e46869 100644 (file)
@@ -94,6 +94,7 @@ EXPORT_SYMBOL(jbd2_journal_file_inode);
 EXPORT_SYMBOL(jbd2_journal_init_jbd_inode);
 EXPORT_SYMBOL(jbd2_journal_release_jbd_inode);
 EXPORT_SYMBOL(jbd2_journal_begin_ordered_truncate);
+EXPORT_SYMBOL(jbd2_inode_cache);
 
 static int journal_convert_superblock_v1(journal_t *, journal_superblock_t *);
 static void __journal_abort_soft (journal_t *journal, int errno);
@@ -2286,17 +2287,19 @@ static void __exit jbd2_remove_jbd_stats_proc_entry(void)
 
 #endif
 
-struct kmem_cache *jbd2_handle_cache;
+struct kmem_cache *jbd2_handle_cache, *jbd2_inode_cache;
 
 static int __init journal_init_handle_cache(void)
 {
-       jbd2_handle_cache = kmem_cache_create("jbd2_journal_handle",
-                               sizeof(handle_t),
-                               0,              /* offset */
-                               SLAB_TEMPORARY, /* flags */
-                               NULL);          /* ctor */
+       jbd2_handle_cache = KMEM_CACHE(jbd2_journal_handle, SLAB_TEMPORARY);
        if (jbd2_handle_cache == NULL) {
-               printk(KERN_EMERG "JBD: failed to create handle cache\n");
+               printk(KERN_EMERG "JBD2: failed to create handle cache\n");
+               return -ENOMEM;
+       }
+       jbd2_inode_cache = KMEM_CACHE(jbd2_inode, 0);
+       if (jbd2_inode_cache == NULL) {
+               printk(KERN_EMERG "JBD2: failed to create inode cache\n");
+               kmem_cache_destroy(jbd2_handle_cache);
                return -ENOMEM;
        }
        return 0;
@@ -2306,6 +2309,9 @@ static void jbd2_journal_destroy_handle_cache(void)
 {
        if (jbd2_handle_cache)
                kmem_cache_destroy(jbd2_handle_cache);
+       if (jbd2_inode_cache)
+               kmem_cache_destroy(jbd2_inode_cache);
+
 }
 
 /*
index 2ae86aa..27e79c2 100644 (file)
@@ -94,7 +94,7 @@ extern void jbd2_free(void *ptr, size_t size);
  *
  * This is an opaque datatype.
  **/
-typedef struct handle_s                handle_t;       /* Atomic operation type */
+typedef struct jbd2_journal_handle handle_t;   /* Atomic operation type */
 
 
 /**
@@ -416,7 +416,7 @@ struct jbd2_revoke_table_s;
  * in so it can be fixed later.
  */
 
-struct handle_s
+struct jbd2_journal_handle
 {
        /* Which compound transaction is this update a part of? */
        transaction_t           *h_transaction;
@@ -1158,6 +1158,22 @@ static inline void jbd2_free_handle(handle_t *handle)
        kmem_cache_free(jbd2_handle_cache, handle);
 }
 
+/*
+ * jbd2_inode management (optional, for those file systems that want to use
+ * dynamically allocated jbd2_inode structures)
+ */
+extern struct kmem_cache *jbd2_inode_cache;
+
+static inline struct jbd2_inode *jbd2_alloc_inode(gfp_t gfp_flags)
+{
+       return kmem_cache_alloc(jbd2_inode_cache, gfp_flags);
+}
+
+static inline void jbd2_free_inode(struct jbd2_inode *jinode)
+{
+       kmem_cache_free(jbd2_inode_cache, jinode);
+}
+
 /* Primary revoke support */
 #define JOURNAL_REVOKE_DEFAULT_HASH 256
 extern int        jbd2_journal_init_revoke(journal_t *, int);