So you disallow timeouts shorter than a second ?

I felt that timeouts less than 1sec or of a resolution greater than the number of
seconds would not make much sense, since I thought there would be delays on
the order of 1sec during normal operation, when a server is slow.

The use I make of it is setting it to 15-30sec.

I also used a rather arbitrary 600sec as a sanity upper bound in the related
change in clnt_vc.c.

If anyone thinks this should be changed (either including the tv_usec field in the
timeout calculation or allowing a fraction of 1sec timeout) I can easily change it?

No, I think this is fine, just add an explanation in a comment, please.

Why not use 1 sec there for timeout ? You make the sleep timeout proportional to the socket timeout, but I do not see why it would be useful. We want to get out of loop early, your method causes pause up to 2* timeout.

Well, the function called to do an RPC is clnt_reconnect_call(). In it there is a loop:
do while

• if not connected, call clnt_reconnect_call(), which ends up here.
  • if this times out, count it as a retry (variable is "tries" and compared with a limit of rc_retries. If at rc_retries, exit loop.
• calls CLNT_MBUF_CALL(), which does the sosend()
• if this returns an error due to timeout, count it as a retry, same as above.

As such, a timeout for either case counts as a retry, so it seems to me that they should
be the same amount of time. If this one is reduced to 1sec, then the rc_retries needs
to be increased (it is the "retrans=N" mount option) to result in the same amount of
time before the function gives up and fails the RPC, but then a timeout at sosend()
would only count as one of these retries.

Well, ideally you would check the real time spent in the wait. There could be spurious wakeups, which would trigger the undeserved error. Also, if timeout has non-zero usec part, you are ignoring it.

IMO doing the sleeps in 1 sec (hz) steps would eliminate all that problems, of course EWOULDBLOCK needs to be ignored and real time check used as the timeout indication, instead.

I can easily add code to include the tv_usec component (actually I have already done that
in the code I have by using tvtohz() which is what the code uses in other places).

Do you really think spurious wakup()s are a problem. I grep'd for wakeup()s on so_timeo
and found 4 of them in uipc_socket.c. Three of them definitely clear SS_ISCONNECTING
before the wakeup(), so those cases will exit the loop. The only case that might not is
in soshutdown(), which I don't think would be used and I can't see why you'd want to
keep looping in that case.