[PATCH] VFS unlink performance enhancement

Sang-yong Suh (sysuh@kigam.re.kr)
Thu, 12 Aug 1999 14:07:53 +0900 (GMT+0900)

Patch-Name: fast-unlink.patch
Kernel-Version: linux-2.2.10
Target-file: linux/fs/namei.c
Abstract: The do_unlink() function locks parent directory and calls
vfs_unlink() without checking the existence of the target.
This patch fixes the problem.

I was comparing the news server performances running kernels 2.0.36
and 2.2.10, before upgrading the main service machine to the newer kernel.
The following table shows a part of INND timer report of the servers.
The hardware specs are almost same except the number of news readers.
They serve approximately 20 and 500 readers per day, respectively.

host-1 host-2
------------- ------ ------
Linux kernel 2.2.10 2.0.36
avg.readers 20 500
------------- ------ ------

INND code avg. msec/call
------------- ------ ------
artwrite 14.956 53.826
open 12.779 40.805
write 1.975 4.505
close 0.038 8.305
artexpire 20.981 14.628

The above result shows that 2.2.10 is faster except artexpire.
Naturally, I had to study the reason why the unlink() performance
is so bad on 2.2.10.

Reading sys_unlink(), do_unlink(), and lookup_dentry(), I realized
that do_unlink() does something unnecessary processes. It looks up
the target but does not check the existence of the target which is
stored within the returned dentry. It simply locks the parent
directory and calls vfs_unlink().

There are many application programs which do not check the existence
of a file and relies on the kernel to delete. For example, during the
boot, my Slackware-4.0 Linux tries to delete following files without
checking the status:

/etc/initrunlvl
/var/log/initrunlvl
/dev/log
/etc/ld.so.cache~
/dev/log
/var/run/named.pid
/dev/printer
cron.update

And they are repeatededly requested. :-(

And, finally, here is the patch. It changes only one file,
linux/fs/namei.c. The diff is against kernel-2.2.10.

--
Sang-yong Suh


Patch-Name: fast-unlink.patch
Kernel-Version: linux-2.2.10
Target-files: linux/fs/namei.c
Abstract: The do_unlink() function locks parent directory and calls
	vfs_unlink() without checking the existence of the target.
	This patch fixes the problem.
Author: Sang-yong Suh <sysuh@kigam.re.kr>
Date: Aug. 12, 1998


*** namei.c.orig	Sun May  9 12:46:08 1999
--- namei.c	Thu Aug 12 12:13:59 1999
***************
*** 1040,1051 ****
  	if (IS_ERR(dentry))
  		goto exit;
  
- 	dir = lock_parent(dentry);
  	error = -ENOENT;
! 	if (check_parent(dir, dentry))
! 		error = vfs_unlink(dir->d_inode, dentry);
! 
          unlock_dir(dir);
  	dput(dentry);
  exit:
  	return error;
--- 1040,1052 ----
  	if (IS_ERR(dentry))
  		goto exit;
  
  	error = -ENOENT;
! 	if (dentry->d_inode) {
! 		dir = lock_parent(dentry);
! 		if (check_parent(dir, dentry))
! 			error = vfs_unlink(dir->d_inode, dentry);
          unlock_dir(dir);
+ 	}
  	dput(dentry);
  exit:
  	return error;


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/