unix domain sockets on nullfs(5)
trociny at freebsd.org
Sat Feb 18 21:22:12 UTC 2012
On Mon, 23 Jan 2012 05:12:38 +0200 Kostik Belousov wrote:
KB> On Sun, Jan 22, 2012 at 08:33:45PM +0200, Mikolaj Golub wrote:
>> There was a bug in my patch: for vop_unpdetach it wanted the vnode to be
>> exclusively locked, while it was called from the context (uipc_detach) where
>> the vnode is not locked.
>> It looks it is OK that the vnode is not locked here: the operation is to null
>> vp->v_socket, and currently the only place where it is concurently accessed is
>> in unp_connect(), and it is protected by the linkage lock.
>> So I updated my patch to have "= = =" for unpdetach vp.
>> On Thu, 12 Jan 2012 23:51:06 +0200 Kostik Belousov wrote:
>> KB> On Thu, Jan 12, 2012 at 09:39:53PM +0000, Robert N. M. Watson wrote:
>> >> I still find myself worried by the fact that unp->unp_vnode points at the
>> >> nullfs vnode rather than the underlying vnode, but haven't yet managed to
>> >> identify any actual bugs that would result. I'll continue pondering it
>> >> over the weekend :-).
>> KB> I think I know what could go wrong there, but due to other bug, this
>> KB> wrongness cannot be realized now.
>> KB> Issue is that for the forced unmount, the unp_vnode is reclaimed, so that
>> KB> the unix domain sockets code references freed memory after reclaim.
>> Just to have this clear, as I understand this problem with reclaim is
>> orthogonal to the initial issue and would also exist without my patch too?
>> Could you please tell what is the other bug? I see that after force unmount,
>> in vflush() we call vgonel() for every vnode, and vgonel() does VOP_CLOSE(),
>> VOP_INACTIVE(), VOP_RECLAIM(), sets v_type = VBAD, but vnode's usecount is not
>> decreased so if a node was active it may not be freed when vdropl() is called.
KB> The lack of the cleanup in the reclamation code.
>> Was the usecount supposed to be dropped somewere in this path (when
>> VOP_CLOSE() is called?) and this is the bug you mean or it is something else?
KB> No, usecount must not be dropped. The hold count counts the owners of
KB> the pointer to the vnode, preventing the freeing of the struct vnode
KB> itself. Usecount is to avoid non-forced unmounts from reclaiming the
>> Currently the usecount (for both VREG and VSOCK) is deacreased when the
>> process finaly closes the discriptor.
>> KB> Probably, some helper should provided by uipc_usrreq, called from VOP_RECLAIM()
>> KB> implementations for VSOCK types of vnodes.
>> I would not be very happy with adding the helper to every fs's VOP_RECLAIM()
>> implementation :-). Couldn't it be somevere in the common code, e.g. in
>> vflush()? Here is the patch I tried:
KB> Not in the vflush(). I think vgonel() would be better place.
After collecting all suggestions and additional testing I have got this patch
It has survived some bind/connect/force umount stress testing revealing only
some issues that are also observed without patching.
More information about the freebsd-arch