On Mon, Apr 11, 2005 at 20:56:48 -0700, Jeff Rizzo wrote:

> I suspect (with virtually nothing to back this up but a hunch) that you
> got the lockups more quickly because of the code changes, but that you
> would have gotten them in any case.

These are the changes in sys/nfs:

nfs_subs.c:
----------------------------
revision 1.132.2.3.2.3
date: 2005/03/16 11:54:53;  author: tron;  state: Exp;  lines: +3 -3
Pull up revision 1.146 (requested by yamt in ticket #1148):
nfs_clearcommit: don't attempt to clear commit info (n_pushlo, etc)
unless the vnode is of VREG.  union members used to keep commit info
are used for other purposes in the case of !VREG.
----------------------------

nfs_vnops.c:
----------------------------
revision 1.186.2.6.2.4
date: 2005/03/16 12:50:34;  author: tron;  state: Exp;  lines: +22 -4
Pull up revision 1.219 (requested by yamt in ticket #1208):
nfs_readdirrpc, nfs_readdirplusrpc:
avoid infinite loops when getting readdir response without
any entries or eof.  PR/28971.
----------------------------
revision 1.186.2.6.2.3
date: 2005/03/16 12:09:46;  author: tron;  state: Exp;  lines: +3 -3
Pull up revision 1.206 (requested by yamt in ticket #1134):
nfs_commit: use NAC_NOTRUNC when loading an attribute
as we're called holding pages locked.
----------------------------

> I haven't looked into the problems lately - they were bad enough that I
> had to rearrange things to be able to avoid NFS in that situation.  I
> haven't had time to look into them, especially given my lack of
> experience in the NFS code.

I'll try to do some tcpdumping next Sunday.

	rvdp