NFSv4 over RDMA

[Talpey, Thomas]

At 09:57 AM 9/25/2008, Lukas Hejtmanek wrote:
>well, it does happen only if I interrupt the mount command, not during regular
>traffic...

Correct - because the ^C causes the RPC layer to close, NULLing out the
upcall pointer. I believe it can also happen during an idle timeout, but I have
never seen it actually occur there.

>Using Mellanox ib_mthca driver.
>
>this is what I got:
>
>Aug 25 21:05:21 skirit59 kernel: [13170.041433]  [<ffffffffa023b05c>]
>? :xprtrdma:rpcrdma_conn_upcall+0x10c/0x290
>...
>Aug 25 21:05:21 skirit59 kernel: [13170.041433]  [<ffffffffa022a3c1>]
>? :rdma_cm:cma_comp_exch+0x51/0x70
>Aug 25 21:05:21 skirit59 kernel: [13170.041433]  [<ffffffffa022c08b>]
>? :rdma_cm:cma_ib_handler+0xab/0x210
>Aug 25 21:05:21 skirit59 kernel: [13170.041433]  [<ffffffffa0204558>]
>? :ib_cm:cm_process_work+0x18/0xd0

Yep, that's the one. Interesting that you see it on mthca, I have only
seen it on cxgb3. Perhaps your network has slightly different latencies.
In any case, the NULL test is an effective workaround.

Tom.