do_syslog_sudo: clamp the split length to avoid an infinite loop

[iefa-m]

do_syslog_sudo() splits a long log line into several syslog(3) calls, computing the per-line budget as maxlen = syslog_maxlen - (strlen(fmt) - 5 + strlen(submituser)). On sudo_logsrvd the submit user comes straight from the client's AcceptMessage (evlog_new() fills evlog->submituser from the submituser key), so its length is attacker controlled. When the escaped submit user is exactly syslog_maxlen minus the continuation overhead (23 bytes, so 937 at the default syslog_maxlen of 960) the continuation maxlen works out to 0. memrchr(p, ' ', 0) returns NULL, tmp is set to p, the line is truncated to empty and p never advances, so the for loop spins forever emitting empty "(command continued)" records. It is reachable on the default logsrvd config (syslog, sudo format) with no TLS client certificate, and hangs the single event-loop daemon for every client.

Compute the overhead separately and clamp maxlen to a minimum of 1 at both the first-line and continuation sites, so the loop always advances. The same guard also avoids the size_t wrap when syslog_maxlen is smaller than the overhead. I confirmed the hang by replaying the split loop with the 937-byte submit user: unpacked it never advances p, and with the clamp it terminates. make check passes.

[millert]

This is mitigated by 5e3b8b0, but it still seems like a good idea to guarantee the loop can make progress in all cases.