Skip to content
GitLab
Closed
Issue created by Mike Perry@mikeperryDeveloper24 of 24 checklist items completed24/24 checklist items

Implement Prop#324 Flow Control

There's enough pieces of Flow Control that need to be done to get it merge-ready that a checklist and ticket seem wise at this point.

Here's stuff I will do:

  • Improve monotime checks to export a global monotime status
  • Remove mobile xon/xoff limits
  • Advertise the average and max edge conn drain rates in XON (after first XOFF, or if above a "low watermark" queue length)
  • Low watermark params to send advisory XON before XOFF
  • Send periodic XONs if the drain rate changes significantly
  • Perform checks for semantically valid XON/XOFF and call circuit_read_valid_data()
  • Turn CELL_QUEUE_HIGHWATER_SIZE (and others?) into consensus params
  • Implement ways to restrict when advisory XONs can be sent, to reduce side channels from exits
  • Alter half-open edge connection checks to work with XON/XOFF wrt valid data
  • Fix edge case where XON/XOFF can arrive after stream close (depends on half-open fix)
  • Additional consensus parameter for CircEWMA's EWMA_TICK_LEN and edge ratelimit low/high change
  • Preliminary tuning of new consensus params over onion svcs
  • Misc XXX's
  • Update Prop#324 spec with above
  • Debug log removal and other log message cleanups/improvements
  • Squash branch for review

Here's stuff @dgoulet can do:

  • Preliminary code review
  • Rate limit packaging data on edge connection's circuits (or reading on edge source sockets?) to match the advertised rate using token buckets
  • Improve oomkiller/circuit closing wrt total edge connection outbuf lengths on circs
  • Determine if we can use any kernel info from KIST to improve buffer length calls (there's some XXX's to note where in the flow control code this may help)
  • Spotcheck existing oomkiller, KIST, and CircEWMA code to see if we should parameterize or tighten anything else for congestion control generally
  • Determine if we should parameterize any other buffer lengths inside channel handling, cell handling, and the circuitmux dragon
  • Help me figure out why flow_control_decide_xon() is almost always only called when the outbuf is 0, where the data is going to, and what we should do about it (KIST, ask the socket if it wouldblock?)
  • Test out the flow control branch on some rate limited onions for weirdness. Help decide some tuning parameter values.

Cc: @dgoulet

Edited by Mike Perry
To upload designs, you'll need to enable LFS and have an admin enable hashed storage. More information