remove BKL from ext2's readdir
PATCH 1/1 remove BKL from ext2's readdir
hi!
I took a look at readdir() in 2.5.64's ext2 and found it serialized by BKL. As far as I see in sources, there is no need in this BKL. So, here is small patch and some benchmarks. of course, the result was expected pretty much ;)
500 1000 readdir+BKL: 0m11.793s 0m23.403s readdir-BKL: 0m6.060s 0m12.113s
description: two processes read own dir populated by files (500 and 1000 files). this repeats for 100000 times. the iron is dual 1GHz P3.
fs/ext2/dir.c +0 -3
@@ -259,8 +259,6 @@int need_revalidate = (filp->f_version != inode->i_version);int ret =0 ; -lock_kernel (); -if (pos > inode->i_size -EXT2_DIR_REC_LEN (1 )) goto done;
@@ -313,7 +311,6 @@ filp->f_pos = (n <<PAGE_CACHE_SHIFT ) | offset; filp->f_version = inode->i_version;UPDATE_ATIME (inode); -unlock_kernel ();return 0 ; }
Quoted 5 lines · Alex Tomas
hi!
I took a look at
readdir()in 2.5.64's ext2 and found it serialized by BKL.
Yes, I had this in -mm for ages, seem to have lost it. Also removal of BKL from lseek().
The theory is that the lock is there to avoid f_pos races against lseek. We ended up deciding that the way to address this is to ensure that all readdir implementations do:
foo_readdir()
{
loff_t pos = file->f_pos;
....
<code which doesn't touch file->f_pos, but which modifies pos>
...
file->f_pos = pos;
}ext2 does this right and does not need the lock_kernel(). Once all filesystems have been audited (and, if necessary, fixed) we can remove the BKL from lseek also.
Quoted 6 lines · Andrew Morton
Quoted 4 lines · Alex Tomas
hi!
Quoted 9 lines
Quoted 1 line
I took a look at
readdir()in 2.5.64's ext2 and found it serialized by BKL.Yes, I had this in -mm for ages, seem to have lost it. Also removal of BKL from
lseek().The theory is that the lock is there to avoid
f_posraces against lseek. We ended up deciding that the way to address this is to ensure that all readdir implementations do:
hm, no. lseek is using the directory's i_sem now, and readdir runs under that too. So we should be able to remove lock_kernel() from the readdir implementation of all filesystems which are using generic_file_llseek().
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Andrew Morton (AM) writes:
AM> hm, no. lseek is using the directory's i_sem now, and readdir AM> runs under that too. So we should be able to remove AM> lock_kernel() from the readdir implementation of all filesystems AM> which are using generic_file_llseek().
looks like only coda and ntfs use generic_file_llseek(). other use default_llseek(). what's the reason do not use generic_file_llseek(). historical only? or not?
Quoted 11 lines · Alex Tomas
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Andrew Morton (AM) writes:
AM> hm, no. lseek is using the directory's
i_semnow, and readdir AM> runs under that too. So we should be able to remove AM>lock_kernel()from the readdir implementation of all filesystems AM> which are usinggeneric_file_llseek().looks like only coda and ntfs use
generic_file_llseek(). other usedefault_llseek(). what's the reason do not usegeneric_file_llseek(). historical only? or not?
grep again.
I've made the change to ext2 and ext3. Other filesystems will require some thought to verify that the lock_kernel()s are not protecting against some other random codepath.
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Andrew Morton (AM) writes:
AM> grep again.
AM> I've made the change to ext2 and ext3. Other filesystems will AM> require some thought to verify that the lock_kernel()s are not AM> protecting against some other random codepath.
hmm:
[root@proto edited]# head Makefile
VERSION = 2
PATCHLEVEL = 5
SUBLEVEL = 64
EXTRAVERSION =
fs/ext2/dir.c:
struct file_operations ext2_dir_operations = {
.read = generic_read_dir,
.readdir = ext2_readdir,
.ioctl = ext2_ioctl,
.fsync = ext2_sync_file,};
fs/read_write.c:
static inline loff_t llseek(struct file *file, loff_t offset, int origin)
{
loff_t (*fn)(struct file *, loff_t, int);
fn = default_llseek;
if (file->f_op && file->f_op->llseek)
fn = file->f_op->llseek;
return fn(file, offset, origin);
}Quoted 9 lines · Alex Tomas
fs/ext2/dir.c:struct
file_operationsext2_dir_operations= {.read = generic_read_dir, .readdir = ext2_readdir, .ioctl = ext2_ioctl, .fsync = ext2_sync_file,};
ah, OK. How about this?
fs/ext2/dir.c renamed from fs/ext2/dir.c~lseek-ext2_readdir +1 -3
@@ -259,8 +259,6 @@ ext2_readdir (struct file * filp, void *int need_revalidate = (filp->f_version != inode->i_version);int ret =0 ; -lock_kernel (); -if (pos > inode->i_size -EXT2_DIR_REC_LEN (1 )) goto done;
@@ -313,7 +311,6 @@ done: filp->f_pos = (n <<PAGE_CACHE_SHIFT ) | offset; filp->f_version = inode->i_version;UPDATE_ATIME (inode); -unlock_kernel ();return 0 ; }
@@ -660,6 +657,7 @@ not_empty: }struct file_operations ext2_dir_operations = { + .llseek = generic_file_llseek, .read = generic_read_dir, .readdir = ext2_readdir, .ioctl = ext2_ioctl,
fs/ext3/dir.c renamed from fs/ext3/dir.c~lseek-ext2_readdir +9 -9
@@ -37,10 +37,11 @@ static int ext3_release_dir (struct inod struct file * filp);struct file_operations ext3_dir_operations = { + .llseek = generic_file_llseek, .read = generic_read_dir, .readdir = ext3_readdir,/* we take BKL. needed?*/ .ioctl = ext3_ioctl,/* BKL held */ - .fsync = ext3_sync_file,/* BKL held */ + .fsync = ext3_sync_file,/* BKL held */ #ifdef CONFIG_EXT3_INDEX .release = ext3_release_dir, #endif
@@ -98,8 +99,7 @@ static int ext3_readdir(struct file * fistruct super_block * sb;int err;struct inode *inode = filp->f_dentry->d_inode; - -lock_kernel (); +int ret =0 ; sb = inode->i_sb;
@@ -110,8 +110,8 @@ static int ext3_readdir(struct file * fi ((inode->i_size >> sb->s_blocksize_bits) == 1))) { err =ext3_dx_readdir (filp, dirent, filldir);if (err !=ERR_BAD_DX_DIR ) { -unlock_kernel (); -return err; + ret = err; + goto out; }/* * We don't set the inode dirty flag since it's not
@@ -191,8 +191,8 @@ revalidate: filp->f_pos = (filp->f_pos | (sb->s_blocksize -1 )) +1 ;brelse (bh); -unlock_kernel (); -return stored; + ret = stored; + goto out; } offset +=le16_to_cpu (de->rec_len);if (le32_to_cpu (de->inode)) {
@@ -222,8 +222,8 @@ revalidate:brelse (bh); }UPDATE_ATIME (inode); -unlock_kernel (); -return 0 ; +out: +return ret; }#ifdef CONFIG_EXT3_INDEX
diff -puN fs/ext2/file.c~lseek-ext2_readdir fs/ext2/file.c
_
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Andrew Morton (AM) writes:
AM> ah, OK. How about this?
looks fine
Quoted 4 lines · Andrew Morton
Quoted 1 line
I took a look at
readdir()in 2.5.64's ext2 and found it serialized by BKL.Yes, I had this in -mm for ages, seem to have lost it. Also removal of BKL from
lseek().
I moved lseek() from the BKL to the i_sem in early 2.5.
Robert Love