Re: NFS over RDMA issues on Linux 5.4

[Leon Romanovsky <leon@kernel.org>]

On Tue, Aug 04, 2020 at 11:34:05AM -0400, Chuck Lever wrote:
>
>
> > On Aug 4, 2020, at 9:53 AM, Chuck Lever <chuck.lever@oracle.com> wrote:
> >
> >
> >
> >> On Aug 4, 2020, at 9:46 AM, Leon Romanovsky <leon@kernel.org> wrote:
> >>
> >> On Tue, Aug 04, 2020 at 09:12:55AM -0400, Chuck Lever wrote:
> >>>
> >>>
> >>>> On Aug 4, 2020, at 9:08 AM, Timo Rothenpieler <timo@rothenpieler.org> wrote:
> >>>>
> >>>> On 04.08.2020 14:49, Chuck Lever wrote:
> >>>>> Timo, I tend to think this is not a configuration issue.
> >>>>> Do you know of a known working kernel?
> >>>>
> >>>> This is a brand new system, it's never been running with any kernel older than 5.4, and downgrading it to 4.19 or something else while in operation is unfortunately not easily possible. For a client it would definitely not be out of the question, but the main nfs server I cannot easily downgrade.
> >>>>
> >>>> Also keep in mind that the dmesg spam happens on both server and client simultaneously.
> >>>
> >>> Let's start with the client only, since restarting it seems to clear the problem.
> >>
> >> It is client because according to the server CQE errors, it is Remote_Invalid_Request_Error
> >> with "9.7.5.2.2 NAK CODES" from IBTA.
> >
> > Thanks! OK, then let's use ftrace.
> >
> > Timo, can you install trace-cmd on your client? Then:
> >
> > 1. # trace-cmd record -e rpcrdma -e sunrpc
> >
> > 2. Trigger the problem
> >
> > 3. Control-C the trace-cmd, and copy the trace.dat file to another system
> >
> > 4. reboot your client
> >
> > Then send me your trace.dat. You don't have to cc the mailing lists.
>
> I see a LOC_LEN_ERR on a Receive. Leon, doesn't that mean the server's
> Send was too large?

1.
We have local_length_error counter, it can help to run it on server and clients.
[leonro@vm ~]$ cat /sys/class/infiniband/ibp0s9/ports/1/hw_counters/resp_local_length_error
0

resp_local_length_error - "Number of times responder detected local length errors."
2.
LOC_LEN_ERR supports that is written in CQE error on the client.
This is what is written in our HW document:
 IB compliant completion with error syndrome
	0x1: Local_Length_Error
3.
From IBTA, 11.6.2 COMPLETION RETURN STATUS
Local Length Error - Generated for a Work Request posted to the local
Send Queue when the sum of the Data Segment lengths exceeds the message
length for the channel adapter port. Generated for a Work Request posted
to the local Receive Queue when the sum of the Data Segment lengths is
too small to receive a valid incoming message or the length of the incoming
message is greater than the maximum message size supported by the HCA port
that received the message.


So if "1" works :), we will be able to distinguish if client sends too
large WR or recieves too large.

Thanks

>
> Timo, what filesystem are you sharing on your NFS server? The thing that
> comes to mind is https://bugzilla.kernel.org/show_bug.cgi?id=198053
>
>
> --
> Chuck Lever
>
>
>