Thanks for your comment.
> > On Dec 8, 2022, at 19:30, Hiroshi Shimamoto <h-shimamoto@xxxxxxx> wrote:
> >
> > From: minoura <minoura@xxxxxxxxxxxxx>
> >
> > Commit 9130b8dbc6ac ("SUNRPC: allow for upcalls for the same uid
> > but different gss service") introduced `auth` argument to
> > __gss_find_upcall(), but in gss_pipe_downcall() it was left as NULL
> > since it (and auth->service) was not (yet) determined.
> >
> > When multiple upcalls with the same uid and different service are
> > ongoing, it could happen that __gss_find_upcall(), which returns the
> > first match found in the pipe->in_downcall list, could not find the
> > correct gss_msg corresponding to the downcall we are looking for due
> > to two reasons:
> >
> > - the order of the msgs in pipe->in_downcall and those in pipe->pipe
> > (that is, the order of the upcalls sent to rpc.gssd) might be
> > different because pipe->lock is dropped between adding one to each
> > list.
> > - rpc.gssd uses threads to write responses, which means we cannot
> > guarantee the order of responses.
> >
> > We could see mount.nfs process hung in D state with multiple mount.nfs
> > are executed in parallel. The call trace below is of CentOS 7.9
> > kernel-3.10.0-1160.24.1.el7.x86_64 but we observed the same hang w/
> > elrepo kernel-ml-6.0.7-1.el7.
> >
> > PID: 71258 TASK: ffff91ebd4be0000 CPU: 36 COMMAND: "mount.nfs"
> > #0 [ffff9203ca3234f8] __schedule at ffffffffa3b8899f
> > #1 [ffff9203ca323580] schedule at ffffffffa3b88eb9
> > #2 [ffff9203ca323590] gss_cred_init at ffffffffc0355818 [auth_rpcgss]
> > #3 [ffff9203ca323658] rpcauth_lookup_credcache at ffffffffc0421ebc [sunrpc]
> > #4 [ffff9203ca3236d8] gss_lookup_cred at ffffffffc0353633 [auth_rpcgss]
> > #5 [ffff9203ca3236e8] rpcauth_lookupcred at ffffffffc0421581 [sunrpc]
> > #6 [ffff9203ca323740] rpcauth_refreshcred at ffffffffc04223d3 [sunrpc]
> > #7 [ffff9203ca3237a0] call_refresh at ffffffffc04103dc [sunrpc]
> > #8 [ffff9203ca3237b8] __rpc_execute at ffffffffc041e1c9 [sunrpc]
> > #9 [ffff9203ca323820] rpc_execute at ffffffffc0420a48 [sunrpc]
> >
> > The scenario is like this. Let's say there are two upcalls for
> > services A and B, A -> B in pipe->in_downcall, B -> A in pipe->pipe.
> >
> > When rpc.gssd reads pipe to get the upcall msg corresponding to
> > service B from pipe->pipe and then writes the response, in
> > gss_pipe_downcall the msg corresponding to service A will be picked
> > because only uid is used to find the msg and it is before the one for
> > B in pipe->in_downcall. And the process waiting for the msg
> > corresponding to service A will be woken up.
> Wait a minute… The ‘service’ here is one of krb5, krb5i, or
> krb5p. What is being pushed down from user space is a RPCSEC_GSS
> context that can be used for any one of those services. So the
> ordering of A and B is not supposed to matter. Any one of those
> requests can take the context and make use of it.
> However once the context has been used with one of the krb5, krb5i or
> krb5p services then it cannot be used with any of the others. This is
> why commit 9130b8dbc6ac that you referenced above separates the
> services in gss_add_msg().
Right. (BTW I am wondering why we do not have to compare
->service in gss_match(), during looking up the cred cache...)
> >
> > Actual scheduing of that process might be after rpc.gssd processes the
> > next msg. In rpc_pipe_generic_upcall it clears msg->errno (for A).
> > The process is scheduled to see gss_msg->ctx == NULL and
> > gss_msg->msg.errno == 0, therefore it cannot break the loop in
> > gss_create_upcall and is never woken up after that.
> >
> > This patch introduces wait and retry at gss_add_msg() to serialize
> > when requests with the same uid but different service comes.
> As long as rpc.gssd returns one context downcall for each upcall (or
> it breaks the connection in order to force a retransmission) then we
> shouldn’t have to serialise anything.
> However what we could do to fix the race you appear to be describing
> is to check if the upcall has completed yet before we accept the
> message as a candidate for the downcall. That could be just a simple
> check for (msg->copied != 0 && list_empty(&msg->list)). Maybe add a
> helper for that in include/linux/sunrpc/rpc_pipe_fs.h?
This makes perfect sense. The race occurs because downcall
(mistakingly) picks a msg which is not sent to gssd yet, and
it can be distinguished with that condition. Will revise
the patch. Thanks!
> _________________________________
> Trond Myklebust
> Linux NFS client maintainer, Hammerspace
> trond.myklebust@xxxxxxxxxxxxxxx
