vfs: move inode stat information closer together
authorLinus Torvalds <torvalds@linux-foundation.org>
Sun, 3 Jun 2012 21:50:19 +0000 (14:50 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sun, 3 Jun 2012 21:50:19 +0000 (14:50 -0700)
The comment above it says "Stat data, not accessed from path walking",
but in fact some of inode fields we use for the common stat data was way
down at the end of the inode, causing unnecessary cache misses for the
common stat operations.

The inode structure is pretty big, and this can change padding depending
on field width, but at least on the common 64-bit configurations this
doesn't change the size.  Some of our inode layout has historically been
to tro to avoid unnecessary padding fields, but cache locality is at
least as important for layout, if not more.

Noticed by looking at kernel profiles, and noticing that the "i_blkbits"
access stood out like a sore thumb.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/fs.h

index 51978ed43e973ccf70996c19c33820fd39edeabe..17fd887c798f3fff64aaf27bf82a81a2d237ee31 100644 (file)
@@ -802,13 +802,14 @@ struct inode {
                unsigned int __i_nlink;
        };
        dev_t                   i_rdev;
+       loff_t                  i_size;
        struct timespec         i_atime;
        struct timespec         i_mtime;
        struct timespec         i_ctime;
        spinlock_t              i_lock; /* i_blocks, i_bytes, maybe i_size */
        unsigned short          i_bytes;
+       unsigned int            i_blkbits;
        blkcnt_t                i_blocks;
-       loff_t                  i_size;
 
 #ifdef __NEED_I_SIZE_ORDERED
        seqcount_t              i_size_seqcount;
@@ -828,9 +829,8 @@ struct inode {
                struct list_head        i_dentry;
                struct rcu_head         i_rcu;
        };
-       atomic_t                i_count;
-       unsigned int            i_blkbits;
        u64                     i_version;
+       atomic_t                i_count;
        atomic_t                i_dio_count;
        atomic_t                i_writecount;
        const struct file_operations    *i_fop; /* former ->i_op->default_file_ops */