On Mon, 2003-03-31 at 12:21, HABBINGA,ERIK (HP-Loveland,ex1) wrote:
> Steve,
>
> The kernel I'm running does have a patch to drop the BKL around calls to
> xfs_permission, xfs_lookup, and xfs_readdir, based on the philosophy of a
> post of yours from Feb 2002:
>
> http://lists.insecure.org/linux-kernel/2002/Feb/4308.html
>
> permission, lookup, and readdir were the biggest hitters of BKL usage when
> profiling the SPEC SFS NFS test last year. I don't know if not having the
> BKL in the lookup code is the problem here, as we crash out of the lookup
> code in fh_compose. I don't drop the BKL before entering
> linvfs_dentry_to_fh. The stack trace from the oops isn't quite correct, the
> call tree doesn't go through lookup_hash before getting to
> linvfs_dentry_to_fh, it should go:
>
> nfsd_lookup -> fh_compose -> linvfs_dentry_to_fh
>
> I'll instrument linvfs_fh_to_dentry and xfs_inactive like you suggest and
> let you know the results. Let me know if you think dropping the BKL around
> permission, lookup, and readdir might be causing this problem.
It has been pointed out to me that there are cases where nfs has been
known to do things without the i_sem being held when it should be.
This in combination with dropping the BKL might the cause here.
I don't know what speedup you are seeing from BKL removals, but
I really think you may be introducing bugs by doing this.
Steve
--
Steve Lord voice: +1-651-683-3511
Principal Engineer, Filesystem Software email: lord@xxxxxxx
|