On Mon, 2019-03-04 at 10:26 -0500, Chuck Lever wrote: > Hi Trond- > > > > On Mar 3, 2019, at 7:19 PM, Trond Myklebust <trondmy@xxxxxxxxx> > > wrote: > > > > Now that the reads happen in a process context rather than a > > softirq, > > it is safe to allocate back channel slots using a reclaiming > > allocation. > > Is this a required change for transports, or simply an optimization? > Wondering if a similar change is needed for RPC-over-RDMA. Now that we no longer need to be atomic for the socket receive code, I figure we can afford the call. The alternative is that we have to drop the connection, which is also very costly. Cheers Trond -- Trond Myklebust Linux NFS client maintainer, Hammerspace trond.myklebust@xxxxxxxxxxxxxxx