Signed-off-by: Rusty Russell <[email protected]>

Our CORK logic was wrong, and it's better to use Nagle anyway:
we disable Nagle just before sending timing-critical messages.

Time for 100 (failed) payments:

Before:
	148.8573575

After:
	10.7356977

Signed-off-by: Rusty Russell <[email protected]>
Changelog-Fixed: Protocol: Removed 200ms latency from sending commit/revoke messages.

To run, use:
	VALGRIND=0 TEST_BENCH=1 eatmydata pytest tests/test_connection.py::test_bench

Example of running on my laptop (without --enable-debugbuild, over of 10 runs):

	FAILED tests/test_connection.py::test_bench - assert 47.403406-49.191765(48.343+/-0.48) == 0

With eatmydata:

	FAILED tests/test_connection.py::test_bench - assert 25.867860-27.549878(26.5762+/-0.5) == 0

You can also run perf on l1 once it's running:

	perf record --call-graph dwarf -q -p $(cat /tmp/ltests-*/test_bench_1/lightning-1/lightningd-regtest.pid)

Then ^C after 10 seconds and run "perf report".

Things which stood out:

1. Tracing in db_exec_prepared_v2 (fixed in another PR)

   31.12%     0.04%  lightningd  lightningd            [.] db_exec_prepared_v2
   - 31.08% db_exec_prepared_v2
      + 22.96% db_sqlite3_exec
      + 4.46% trace_span_end
      + 1.77% trace_span_start
      + 1.11% trace_span_tag
      + 0.72% tal_free

2. Logging:

   - 16.03% logv
      - 8.15% maybe_print
         - log_to_files
            + 4.51% __GI__IO_fflush (inlined)
            + 1.97% tal_fmt_
            + 0.51% __GI___strftime_l (inlined)

3. Notification (when nothing is listening) in notify_log:

      - 6.84% maybe_notify_log
         - notify_log
            + 3.37% notify_send
            + 1.75% notify_start
            + 1.71% log_notification_serialize
        0.56% new_log_entry

Signed-off-by: Rusty Russell <[email protected]>

…hecking.

Before:
92.26076173782349
91.9576666355133
90.92732524871826

After:
90.5989830493927
88.10309219360352
90.07689118385315

Signed-off-by: Rusty Russell <[email protected]>

If nobody is subscribed, have notify_start return NULL and the caller
can skip serialization.  This is particularly useful for the "log"
notification which can get called a lot.

Signed-off-by: Rusty Russell <[email protected]>

This means we don't have to iterate through all plugins, making
our "do we even have to construct this notification" optimization
much more efficient.

Signed-off-by: Rusty Russell <[email protected]>

This speeds logging slightly, but most of the time is actually
spent in fflush() (which we need).

Result (10 runs, eatmydata):
	FAILED tests/test_connection.py::test_bench - assert 24.086638-25.347475(24.6901+/-0.4) == 0

This is an 8% performance improvement (over the entire series), which is not bad.

Signed-off-by: Rusty Russell <[email protected]>

We don't care about data persistence, and this suppresses all sync calls (speeding
sqlite3 specifically).

Signed-off-by: Rusty Russell <[email protected]>