Re: [PATCH linux-next v2] SUNRPC: rpcrdma_register_default_external: Dynamically allocate ib_phys_buf

["J. Bruce Fields" <bfields@fieldses.org>]

On Mon, Mar 11, 2013 at 07:48:51PM +0000, Myklebust, Trond wrote:
> On Mon, 2013-03-11 at 15:15 -0400, J. Bruce Fields wrote:
> > On Mon, Mar 11, 2013 at 12:51:44PM -0600, Tim Gardner wrote:
> > > On 03/11/2013 12:14 PM, J. Bruce Fields wrote:
> > > <snip>
> > > >>
> > > >> v2 - Move the array of 'struct ib_phys_buf' objects into struct rpcrdma_req
> > > >> and pass this request down through rpcrdma_register_external() and
> > > >> rpcrdma_register_default_external(). This is less overhead then using
> > > >> kmalloc() and requires no extra error checking as the allocation burden is
> > > >> shifted to the transport client.
> > > > 
> > > > Oh good--so that works, and the req is the right place to put this?  How
> > > > are you testing this?
> > > > 
> > > > (Just want to make it clear: I'm *not* an expert on the rdma code, so my
> > > > suggestion to put this in the rpcrdma_req was a suggestion for something
> > > > to look into, not a claim that it's correct.)
> > > > 
> > > 
> > > Just compile tested so far. Incidentally, I've been through the call stack:
> > > 
> > > call_transmit
> > >  xprt_transmit
> > >   xprt->ops->send_request(task)
> > >    xprt_rdma_send_request
> > >     rpcrdma_marshal_req
> > >      rpcrdma_create_chunks
> > >       rpcrdma_register_external
> > >        rpcrdma_register_default_external
> > > 
> > > It appears that the context for kmalloc() should be fine unless there is
> > > a spinlock held around call_transmit() (which seems unlikely).
> > 
> > Right, though I think it shouldn't be GFP_KERNEL--looks like writes
> > could wait on it.
> 
> Nothing inside the RPC client should be using anything heavier than
> GFP_NOWAIT (unless done at setup).
> 
> > In any case, the embedding-in-rpcrdma_req solution does look cleaner if
> > that's correct (e.g. if we can be sure there won't be two simultaneous
> > users of that array).
> 
> Putting it in the rpcrdma_req means that you have one copy per transport
> slot. Why not rather put it in the rpcrdma_xprt?
> AFAICS you only need this array at transmit time for registering memory
> for RDMA, at which time the transport XPRT_LOCK guarantees that nobody
> else is competing for these resources.

Oh, good.  If that works, Steve might want to look back at how that
array size was chosen?  I seem to recall there being some compromise due
to this array being on the stack, and that there might have been some
performance advantage to increasing it further, but I can't find the bug
right now....  (And I might be misremembering.)

--b.