Too many security check reads to getxattr?

Too many security check reads to getxattr?

[Charles Manning]

I'm investigating a performance degradation in yaffs when xattribs are 
enabled. Yes, I know yaffs is not part of mainline Linux yet, but this is 
really a vfs question.

What I notice is that on every write the "security.capability" xattribute gets 
checked. Surely once the vfs has tried reading this once it should not try 
again on every write.

Herewith a call stack

[ 2317.288740]  [<f85b3b17>] yaffs_getxattr+0x27/0xa0 [yaffs2]
[ 2317.288746]  [<f85b3af0>] ? yaffs_getxattr+0x0/0xa0 [yaffs2]
[ 2317.288750]  [<c03091c7>] cap_inode_need_killpriv+0x37/0x50
[ 2317.288753]  [<c030a434>] security_inode_need_killpriv+0x14/0x20
[ 2317.288757]  [<c01dda2d>] file_remove_suid+0x1d/0x60
[ 2317.288761]  [<c01dfd8a>] __generic_file_aio_write+0x25a/0x550
[ 2317.288764]  [<c0229eef>] ? link_path_walk+0x42f/0x9b0
[ 2317.288768]  [<c023075f>] ? dput+0x6f/0xd0
[ 2317.288772]  [<c01e00d8>] generic_file_aio_write+0x58/0xc0
[ 2317.288775]  [<c021e819>] do_sync_write+0xb9/0xf0
[ 2317.288779]  [<c0222954>] ? cp_new_stat64+0xe4/0x100
[ 2317.288783]  [<c0338346>] ? apparmor_file_permission+0x16/0x20
[ 2317.288786]  [<c030ba4d>] ? security_file_permission+0x1d/0x80
[ 2317.288790]  [<c021e9b1>] ? rw_verify_area+0x71/0xe0
[ 2317.288793]  [<c021eabb>] vfs_write+0x9b/0x160
[ 2317.288796]  [<c021e760>] ? do_sync_write+0x0/0xf0
[ 2317.288800]  [<c021f4a2>] sys_write+0x42/0x70
[ 2317.288803]  [<c0102cdf>] sysenter_do_call+0x12/0x28

[ 2317.288806] yaffs_getxattr "security.capability" from object 257

Any hints as to why this is happening and how to stop this from happening?

Thanks

-- Charles

Re: Too many security check reads to getxattr?

[Andreas Dilger]

On 2010-08-30, at 19:24, Charles Manning wrote:
> I'm investigating a performance degradation in yaffs when xattribs are 
> enabled. Yes, I know yaffs is not part of mainline Linux yet, but this is 
> really a vfs question.
> 
> What I notice is that on every write the "security.capability" xattribute gets 
> checked. Surely once the vfs has tried reading this once it should not try 
> again on every write.

This was also discussed in the thread "remove_suid bangs on xattrs" a couple of weeks ago.  We've noticed this on btrfs and Lustre also, and it really hurts performance over the network.

Cheers, Andreas