Thilak Raj Surendra Babu via dev <[email protected]> writes:

> Add optional logging of userspace connection tracking lifecycles.
> Each connection logs exactly one line, written when it is purged,
> with its tuple, when it was created, when and why it was purged,
> and how long it lived.
>
> A PMD thread must never block, format a string, or write a log
> line, so no part of this runs on one.  The thread that already
> removes the connection -- the 'ct_clean' sweep thread for a normal
> expiry, or whichever thread handles an explicit flush -- does the
> formatting and the vlog write itself, inline, right there.  No ring
> buffer, no dedicated consumer thread, and no new memory: the only
> addition to 'struct conn' is the 'created' timestamp, set once at
> creation and never touched again.  A connection's removal from the
> table is unconditional and is never delayed by logging; a line can
> only ever be skipped (and counted by the
> 'conntrack_log_ratelimited' coverage counter), never the removal
> itself.
>
> Logging is disabled by default.  When disabled, the cost on the
> datapath is a single atomic load.  When enabled, output can be
> restricted to a list of conntrack zones and is capped at a
> configurable number of lines per second (default 1000, 0 for
> unlimited) so that a connection storm cannot flood the log;
> reclaiming connections continues at full speed regardless of the
> cap.  Configure with 'ovs-appctl conntrack/log-set [enable|disable]
> [zones=LIST] [rate=N]' and inspect the current configuration with
> 'ovs-appctl conntrack/log-show'.
>
> Earlier revisions of this change used a lock-free ring buffer and a
> separate consumer thread to move logging work off any thread that
> could be a PMD.  That is unnecessary here: the sweep and flush
> threads already touch every connection they purge, are never PMDs,
> and run far below packet rate, so moving the work there directly
> needs no new thread, no ring, and no additional memory.
>
> Tested with 'make check', including a new ofproto-dpif.at test case
> covering enabling, disabling, rate-limiting, and zone-filtering,
> and with the new 'ovstest test-conntrack benchmark-new-conns'
> micro-benchmark on real 25G hardware: logging disabled, or capped
> at the default rate, is indistinguishable from logging off for
> both connection creation and flush; only an explicit, non-default
> 'rate=0' (unlimited) measurably slows a bulk flush, trading flush
> time for a guarantee that no line is dropped.
>
> Signed-off-by: Thilak Raj Surendra Babu <[email protected]>
> Assisted-by: Claude Sonnet 5, Claude Code
> ---
> v2 -> v3: No code changes. Rebased onto current master: two unrelated
> NEWS entries landed upstream at the same insertion point since v2 and
> made git-am's context match fail (0-day robot: "Failed to merge in
> the changes"). Resolved the NEWS conflict only; everything else is
> unchanged from v2.
>
> v1 -> v2: No code changes. Resending because the 0-day robot reported
> a git-am "Patch format detection failed" error on v1; this did not
> reproduce locally (git am applies cleanly at the current master
> merge-base, checkpatch is clean), and a transient 502 from
> patchwork's own mbox endpoint around the same time suggests an
> infra-side hiccup fetching the patch rather than an issue with it.

I sent feedback on v2.  As you note, the code is the same.
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to