Re: [PATCH v2 0/5] Implement referring call lists for CB_OFFLOAD

[Chuck Lever <chuck.lever@oracle.com>]

On 3/7/25 10:00 AM, Jeff Layton wrote:
> On Sat, 2025-03-01 at 13:31 -0500, cel@kernel.org wrote:
>> From: Chuck Lever <chuck.lever@oracle.com>
>>
>> I've built a naive proof-of-concept of the csa_referring_call_list
>> argument of the CB_SEQUENCE operation, and hooked it up for the
>> CB_OFFLOAD callback operation.
>>
>> This has been pushed to my kernel.org "fix-async-copy" branch for
>> folks to play around with.
>>
>> I've done some basic testing with a server that ensures the
>> CB_OFFLOAD callback is sent before the COPY reply, while running a
>> network capture. Operation appears correct, Wireshark is happy
>> with the construction of the XDR, and the CB_SEQUENCE arguments
>> match the SEQUENCE operation in the COPY COMPOUND.
>>
>> I'd like to include this series in nfsd-testing.
>>
>> Changes since RFC:
>> - Add a field to struct nfsd4_slot that records its table index
>> - Include a few additional COPY-related fixes
>> - Some operational testing has been done
>>
>> Chuck Lever (5):
>>   NFSD: OFFLOAD_CANCEL should mark an async COPY as completed
>>   NFSD: Shorten CB_OFFLOAD response to NFS4ERR_DELAY
>>   NFSD: Implement CB_SEQUENCE referring call lists
>>   NFSD: Record each NFSv4 call's session slot index
>>   NFSD: Use a referring call list for CB_OFFLOAD
>>
>>  fs/nfsd/nfs4callback.c | 132 +++++++++++++++++++++++++++++++++++++++--
>>  fs/nfsd/nfs4proc.c     |  16 ++++-
>>  fs/nfsd/nfs4state.c    |  38 ++++++------
>>  fs/nfsd/state.h        |  23 +++++++
>>  fs/nfsd/xdr4.h         |   4 ++
>>  fs/nfsd/xdr4cb.h       |   5 +-
>>  6 files changed, 193 insertions(+), 25 deletions(-)
>>
> 
> I think this all looks good for a first pass, and should be OK for
> COPY. You can add:
> 
> 
>     Reviewed-by: Jeff Layton <jlayton@kernel.org>

Thanks!


> I think we'll eventually want this for longer-lived stateids too.
> Specifically:
> 
>     CB_RECALL
>     CB_LAYOUTRECALL
>     CB_NOTIFY_LOCK
> 
> The main thing missing for that is the ability to free referring call
> records once we ensure that the client has seen the reply. For
> instance, if nfsd records a referring call on slot:seq 1:2, then once
> it sees a SEQUENCE for 1:3, then it doesn't need to keep around the
> referring call for 1:2. The server knows that call is no longer in
> flight so it's no longer needed.
> 
> If we don't do that, then we could end up with rather long referring
> call lists, with a bunch of long-completed calls.

Agreed that RCLs will have uses outside of COPY. I don't have a good
understanding of the use cases you mention above, so it would delay
RCL support in CB_OFFLOAD quite a bit if we were to wait for the
implementation of the other use cases.


-- 
Chuck Lever