On Wed, 2025-01-22 at 11:06 -0500, Chuck Lever wrote: > On 1/22/25 10:44 AM, Jeff Layton wrote: > > On Wed, 2025-01-22 at 10:20 -0500, Chuck Lever wrote: > > > On 1/22/25 10:10 AM, Jeff Layton wrote: > > > > The v4.0 client always restarts the callback when the connection is shut > > > > down (which is indicated by RPC_SIGNALLED()). The RPC is then requeued > > > > and the result eventually should complete (or be aborted). > > > > > > > > The v4.1 code instead processes the result and only later decides to > > > > restart the call. Even more problematic is the fact that it releases the > > > > slot beforehand. The restarted call may get a new slot, which would > > > > could break DRC handling. > > > > > > "break DRC handling" -- I'd like to understand this. > > > > > > NFSD always sets cachethis to false in CB_SEQUENCE, so there is no DRC > > > for these operations. The only thing the client saves is the slot > > > sequence number IIUC. > > > > > > Is retrying an uncached operation via a different slot a problem? > > > > > > > Ahh, I missed that we always set cachethis to false. So, there is > > probably now a problem with the DRC after all. Still, I don't see a > > good argument for processing the CB_SEQUENCE result, when we intend to > > retransmit the call anyway. > > I expect that the rationale is that the slot sequence number needs to be > advanced appropriately before the slot can be used again. > > Once RPC_SIGNALLED returns true, the callback code can either trust the result of the rpc_task or not. If it's going to trust that result, then there is no need to restart the call. If it's not going to trust it, then the RPC call might as well have not happened, and there is no need to increment the slot sequence number or do anything else. Is my understanding wrong here? -- Jeff Layton <jlayton@xxxxxxxxxx>
