Hi Aaron,

Thanks for the detailed review. Responses inline:

Robot/infra issue: Understood, no action needed on my end -- thanks for
the context.

USDT probes: OVS_USDT_PROBE() fits live tracing well, but gives no
durable history -- nothing's recorded unless a consumer is attached --
so I'd keep the vlog line, optionally adding a probe at purge. A
creation probe isn't needed for lifecycle info, since conn->created is
stored and available at purge. And using probes as the persistent
record would mean a permanently attached tracer, so a uprobe trap on
every new connection, on the PMD, under ct_lock, versus a single field
write today.

ct_log_take_budget() concurrency bug: Confirmed -- separate relaxed
atomics let ct_clean/main-thread callers mis-admit lines. Honestly,
overlooked this in the original design. Fixed with a mutex-protected
token_bucket (same primitive vlog_should_drop()/old ct_print_conn_info()
used), still configurable via rate=N.

That said -- the rate limit is a leftover from an earlier version that
logged at creation time, on the datapath. Now that logging only happens
at purge, on ct_clean or the main thread, that reason is gone. Is it
still worth keeping just to cap log volume, or should I drop it (and
the rate= option)?

Unexercised test: Not intentional, and a fair catch. I've wired
benchmark-new-conns into tests/library.at, and added an ofproto-dpif.at
test that flushes more connections than the rate allows and checks
conntrack_log_ratelimited. It immediately caught a burst-size bug in
the rewritten limiter, now fixed. Let me know if you meant something
else.

ct_log_parse_zones() accepting zones=5,all,15123-31234: Agreed -- made
"all" and an explicit list mutually exclusive, rejected outright
instead of silently preferring one.

Upper bound on rate=: Yes, needed -- in this version, (int) rate turns
any rate above INT_MAX into a negative budget, so nothing gets logged
at all. The rewritten limiter takes rate= as-is and saturates the burst
size via OVS_SAT_MUL (same helper VLOG_RATE_LIMIT_INIT uses), so an
oversized value clamps instead of wrapping.

CT_LOG_R_EXPIRED's "replaced by new connection" comment: Confirmed
inaccurate -- conn_clean() has only three kinds of callers (destroy,
sweep, flush), no "replaced" path exists. Fixed to just
/* Timed out. */

ct_print_conn_info(): Looked at its last version before removal (commit
99ae06105) -- structurally close to what I want, just missing the
creation timestamp. I'll look at basing v5's logging on that precedent
rather than the from-scratch version.

I'll send v5 once we've settled the rate-limit question.

Thanks
Thilak Raj S
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to