Skip to content
Snippets Groups Projects
  1. Sep 22, 2017
  2. Sep 21, 2017
  3. Sep 19, 2017
  4. Sep 17, 2017
  5. Sep 16, 2017
  6. Sep 15, 2017
  7. Sep 14, 2017
  8. Sep 13, 2017
    • Tatsuhiro Tsujikawa's avatar
      ngtcp2_rtb: O(1) removal · 8ad4e356
      Tatsuhiro Tsujikawa authored
      8ad4e356
    • Tatsuhiro Tsujikawa's avatar
      Run missing test · b4dae38e
      Tatsuhiro Tsujikawa authored
      b4dae38e
    • Tatsuhiro Tsujikawa's avatar
      4715049c
    • Tatsuhiro Tsujikawa's avatar
      Don't take down server with fatal socket error · b3f7955f
      Tatsuhiro Tsujikawa authored
      Instead just delete the corresponding handler.
      b3f7955f
    • Tatsuhiro Tsujikawa's avatar
      Fix compiler warning · c1a0bf2b
      Tatsuhiro Tsujikawa authored
      c1a0bf2b
    • Tatsuhiro Tsujikawa's avatar
      Merge pull request #42 from nickfajones/send_retry · f1163ddd
      Tatsuhiro Tsujikawa authored
      Add packet send error checking and retry
      f1163ddd
    • nickfajones's avatar
      Add packet send error checking and retry · 3200cdc5
      nickfajones authored
      This patch represents phase 2 of the "Close all
      connections on exit" feature, as well as including
      a number of other benefits for both client and server:
      
      - Packet writing is now moved to a function: send_packet
        this function will sense a number of different errors
        on the sendto system call, and categorise these as
        either fatal or non fatal (able to be retried).
        Non fatal errors will most likely be caused by the
        kernel tx queue being full, which means the packet
        should be writable at a later time.
        This code is loosely based on similar code in the
        Berkeley Internet Name Daemon (BIND), which I
        assume would know how to send UDP packets.
      - Functions that write packets will observe the
        send_packet return value and if the network error is
        non fatal, the packet that was to be sent will be
        retained and writability on the socket file
        descriptor will be polled through the event reactor.
      - When writes are once again possible on the socket,
        the stored packets will be sent first, to preserve
        ordering at the send side, and only once this
        packet is sent will the core library be queried
        for more new transport level and application level
        packets, which themselves may need to be queued.
      3200cdc5
  9. Sep 11, 2017