ext4: move quota initialization out of inode allocation transaction
authorJan Kara <jack@suse.cz>
Fri, 19 Apr 2013 17:38:14 +0000 (13:38 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Fri, 19 Apr 2013 17:38:14 +0000 (13:38 -0400)
Inode allocation transaction is pretty heavy (246 credits with quotas
and extents before previous patch, still around 200 after it).  This is
mostly due to credits required for allocation of quota structures
(credits there are heavily overestimated but it's difficult to make
better estimates if we don't want to wire non-trivial assumptions about
quota format into filesystem).

So move quota initialization out of allocation transaction. That way
transaction for quota structure allocation will be started only if we
need to look up quota structure on disk (rare) and furthermore it will
be started for each quota type separately, not for all of them at once.
This reduces maximum transaction size to 34 is most cases and to 73 in
the worst case.

[ Modified by tytso to clean up the cleanup paths for error handling.
  Also use a separate call to ext4_std_error() for each failure so it
  is easier for someone who is debugging a problem in this function to
  determine which function call failed. ]

Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
fs/ext4/ialloc.c
fs/ext4/namei.c

index 4c358f711cef25c584a90b6c34a10d7616cb1fde..18d36d85f5c9e1bd22286726ee6628e742c56612 100644 (file)
@@ -666,6 +666,23 @@ struct inode *__ext4_new_inode(handle_t *handle, struct inode *dir,
        ei = EXT4_I(inode);
        sbi = EXT4_SB(sb);
 
+       /*
+        * Initalize owners and quota early so that we don't have to account
+        * for quota initialization worst case in standard inode creating
+        * transaction
+        */
+       if (owner) {
+               inode->i_mode = mode;
+               i_uid_write(inode, owner[0]);
+               i_gid_write(inode, owner[1]);
+       } else if (test_opt(sb, GRPID)) {
+               inode->i_mode = mode;
+               inode->i_uid = current_fsuid();
+               inode->i_gid = dir->i_gid;
+       } else
+               inode_init_owner(inode, dir, mode);
+       dquot_initialize(inode);
+
        if (!goal)
                goal = sbi->s_inode_goal;
 
@@ -697,7 +714,7 @@ got_group:
 
                gdp = ext4_get_group_desc(sb, group, &group_desc_bh);
                if (!gdp)
-                       goto fail;
+                       goto out;
 
                /*
                 * Check free inodes count before loading bitmap.
@@ -711,7 +728,7 @@ got_group:
                brelse(inode_bitmap_bh);
                inode_bitmap_bh = ext4_read_inode_bitmap(sb, group);
                if (!inode_bitmap_bh)
-                       goto fail;
+                       goto out;
 
 repeat_in_this_group:
                ino = ext4_find_next_zero_bit((unsigned long *)
@@ -733,13 +750,16 @@ repeat_in_this_group:
                                                         handle_type, nblocks);
                        if (IS_ERR(handle)) {
                                err = PTR_ERR(handle);
-                               goto fail;
+                               ext4_std_error(sb, err);
+                               goto out;
                        }
                }
                BUFFER_TRACE(inode_bitmap_bh, "get_write_access");
                err = ext4_journal_get_write_access(handle, inode_bitmap_bh);
-               if (err)
-                       goto fail;
+               if (err) {
+                       ext4_std_error(sb, err);
+                       goto out;
+               }
                ext4_lock_group(sb, group);
                ret2 = ext4_test_and_set_bit(ino, inode_bitmap_bh->b_data);
                ext4_unlock_group(sb, group);
@@ -755,8 +775,10 @@ repeat_in_this_group:
 got:
        BUFFER_TRACE(inode_bitmap_bh, "call ext4_handle_dirty_metadata");
        err = ext4_handle_dirty_metadata(handle, NULL, inode_bitmap_bh);
-       if (err)
-               goto fail;
+       if (err) {
+               ext4_std_error(sb, err);
+               goto out;
+       }
 
        /* We may have to initialize the block bitmap if it isn't already */
        if (ext4_has_group_desc_csum(sb) &&
@@ -768,7 +790,8 @@ got:
                err = ext4_journal_get_write_access(handle, block_bitmap_bh);
                if (err) {
                        brelse(block_bitmap_bh);
-                       goto fail;
+                       ext4_std_error(sb, err);
+                       goto out;
                }
 
                BUFFER_TRACE(block_bitmap_bh, "dirty block bitmap");
@@ -787,14 +810,18 @@ got:
                ext4_unlock_group(sb, group);
                brelse(block_bitmap_bh);
 
-               if (err)
-                       goto fail;
+               if (err) {
+                       ext4_std_error(sb, err);
+                       goto out;
+               }
        }
 
        BUFFER_TRACE(group_desc_bh, "get_write_access");
        err = ext4_journal_get_write_access(handle, group_desc_bh);
-       if (err)
-               goto fail;
+       if (err) {
+               ext4_std_error(sb, err);
+               goto out;
+       }
 
        /* Update the relevant bg descriptor fields */
        if (ext4_has_group_desc_csum(sb)) {
@@ -840,8 +867,10 @@ got:
 
        BUFFER_TRACE(group_desc_bh, "call ext4_handle_dirty_metadata");
        err = ext4_handle_dirty_metadata(handle, NULL, group_desc_bh);
-       if (err)
-               goto fail;
+       if (err) {
+               ext4_std_error(sb, err);
+               goto out;
+       }
 
        percpu_counter_dec(&sbi->s_freeinodes_counter);
        if (S_ISDIR(mode))
@@ -851,16 +880,6 @@ got:
                flex_group = ext4_flex_group(sbi, group);
                atomic_dec(&sbi->s_flex_groups[flex_group].free_inodes);
        }
-       if (owner) {
-               inode->i_mode = mode;
-               i_uid_write(inode, owner[0]);
-               i_gid_write(inode, owner[1]);
-       } else if (test_opt(sb, GRPID)) {
-               inode->i_mode = mode;
-               inode->i_uid = current_fsuid();
-               inode->i_gid = dir->i_gid;
-       } else
-               inode_init_owner(inode, dir, mode);
 
        inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb);
        /* This is the optimal IO size (for stat), not the fs block size */
@@ -889,7 +908,9 @@ got:
                 * twice.
                 */
                err = -EIO;
-               goto fail;
+               ext4_error(sb, "failed to insert inode %lu: doubly allocated?",
+                          inode->i_ino);
+               goto out;
        }
        spin_lock(&sbi->s_next_gen_lock);
        inode->i_generation = sbi->s_next_generation++;
@@ -917,7 +938,6 @@ got:
                ext4_set_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
 
        ret = inode;
-       dquot_initialize(inode);
        err = dquot_alloc_inode(inode);
        if (err)
                goto fail_drop;
@@ -951,24 +971,17 @@ got:
 
        ext4_debug("allocating inode %lu\n", inode->i_ino);
        trace_ext4_allocate_inode(inode, dir, mode);
-       goto really_out;
-fail:
-       ext4_std_error(sb, err);
-out:
-       iput(inode);
-       ret = ERR_PTR(err);
-really_out:
        brelse(inode_bitmap_bh);
        return ret;
 
 fail_free_drop:
        dquot_free_inode(inode);
-
 fail_drop:
-       dquot_drop(inode);
-       inode->i_flags |= S_NOQUOTA;
        clear_nlink(inode);
        unlock_new_inode(inode);
+out:
+       dquot_drop(inode);
+       inode->i_flags |= S_NOQUOTA;
        iput(inode);
        brelse(inode_bitmap_bh);
        return ERR_PTR(err);
index 45a5ca89797fcdd663d831e7c3b8ce088778a9d9..955c907fc9800abd2960b44dc3a7d6b2d9ec239e 100644 (file)
@@ -2252,8 +2252,7 @@ static int ext4_create(struct inode *dir, struct dentry *dentry, umode_t mode,
        dquot_initialize(dir);
 
        credits = (EXT4_DATA_TRANS_BLOCKS(dir->i_sb) +
-                  EXT4_INDEX_EXTRA_TRANS_BLOCKS + 3 +
-                  EXT4_MAXQUOTAS_INIT_BLOCKS(dir->i_sb));
+                  EXT4_INDEX_EXTRA_TRANS_BLOCKS + 3);
 retry:
        inode = ext4_new_inode_start_handle(dir, mode, &dentry->d_name, 0,
                                            NULL, EXT4_HT_DIR, credits);
@@ -2287,8 +2286,7 @@ static int ext4_mknod(struct inode *dir, struct dentry *dentry,
        dquot_initialize(dir);
 
        credits = (EXT4_DATA_TRANS_BLOCKS(dir->i_sb) +
-                  EXT4_INDEX_EXTRA_TRANS_BLOCKS + 3 +
-                  EXT4_MAXQUOTAS_INIT_BLOCKS(dir->i_sb));
+                  EXT4_INDEX_EXTRA_TRANS_BLOCKS + 3);
 retry:
        inode = ext4_new_inode_start_handle(dir, mode, &dentry->d_name, 0,
                                            NULL, EXT4_HT_DIR, credits);
@@ -2397,8 +2395,7 @@ static int ext4_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
        dquot_initialize(dir);
 
        credits = (EXT4_DATA_TRANS_BLOCKS(dir->i_sb) +
-                  EXT4_INDEX_EXTRA_TRANS_BLOCKS + 3 +
-                  EXT4_MAXQUOTAS_INIT_BLOCKS(dir->i_sb));
+                  EXT4_INDEX_EXTRA_TRANS_BLOCKS + 3);
 retry:
        inode = ext4_new_inode_start_handle(dir, S_IFDIR | mode,
                                            &dentry->d_name,
@@ -2827,8 +2824,7 @@ static int ext4_symlink(struct inode *dir,
                 * quota blocks, sb is already counted in previous macros).
                 */
                credits = EXT4_DATA_TRANS_BLOCKS(dir->i_sb) +
-                         EXT4_INDEX_EXTRA_TRANS_BLOCKS + 3 +
-                         EXT4_MAXQUOTAS_INIT_BLOCKS(dir->i_sb);
+                         EXT4_INDEX_EXTRA_TRANS_BLOCKS + 3;
        }
 retry:
        inode = ext4_new_inode_start_handle(dir, S_IFLNK|S_IRWXUGO,