Repository navigation
burrow v1.1.5
Fixed
-
The proxy now actually uses the TCP tuning. Every client connection, and every
connection to Telegram, ends inside the mtg container's own network namespace. That
namespace doesn't inherit most of the host's settings, so the relay was running on
kernel defaults: slow start after idle, no MTU probing, a 128-entry SYN queue and
two-hour keepalives. The container now sets these itself:- no slow start after idle, so media sent after a pause doesn't start slowly;
- MTU probing, which avoids stalls on mobile networks;
- 8,192-entry accept and SYN queues;
- faster cleanup of dead connections;
- a 128 KiB cap on unsent data per socket.
A new test keeps the container and host profiles in step.
-
fqnow takes effect at install time.net.core.default_qdisconly applies to
interfaces brought up after it's set, so the network card stayed onfq_codeluntil a
reboot. The installer now switches the uplink tofqimmediately: onefqfor a
single-queue card, or one per queue undermq. -
About 80 unit tests could not fail. burrow's own
failstatus helper shadowed the
test framework'sfail, so every assertion in tests that load burrow's libraries
passed even when it was wrong. The helper is renamed tocross. Turning the assertions
back on found one stale expectation and no product bugs.
Changed
- The host profile also caps unsent data per socket (
tcp_notsent_lowat, 128 KiB): less
memory per connection and less queueing delay, with no loss of throughput under BBR.