mirror of
https://github.com/soypat/lneto.git
synced 2026-09-09 16:19:41 +00:00
Pat Whittingslow
80ea9256e7
* fix(xnet): make blocking waits deadline-driven instead of iteration-capped
All six StackBlocking wait loops run 'for range maxIter' (1000). With a
non-sleeping backoff like BackoffFlagGosched, the sensible choice on
GOMAXPROCS=1 targets where sleeping starves the NIC poll, those 1000
iterations complete in ~20ms of wall time and silently replace the
caller's timeout: a dial with the default 2s timeout fails after ~21ms
with a spurious deadline error against any peer slower than that.
Observed dialing github.com from a single-core bare-metal (tamago)
target through go-net.
The deadline check inside every loop is the real guard, so it becomes the
loop condition itself and keeps the bounded lifetime visible on the for
statement:
for ok := true; ok; ok = s.checkDeadline(deadline) == nil {
DoDHCPv4 loses its separate deadline branch as a result. It used to check
only on iterations without a state change, which without the iteration
cap would let a peer that keeps feeding state transitions hold the loop
past any deadline; from the header the check runs every iteration.
The regression test drives a dial through 4000 no-progress wait
iterations on simulated time (<5% of the deadline elapsed) and requires
establishment once the handshake is finally serviced; on current main it
fails at exactly iteration 1000.
* go fix
---------
Co-authored-by: Derek den Haas <d.haas@directcode.com>