gazagnaire.org
User-mode network gateway (SLIRP / vpnkit) in pure OCaml
slirp: correct the guest TSQ default to the measured 4 MiB
The first live /stats read-back showed the pinned 6.18.39 guest defaults
tcp_limit_output_bytes to 4 MiB (eight BIG TCP frames), not the 1 MiB the
E26 analysis assumed. State the measured value where the knob is
documented.
author
Thomas Gazagnaire
date
(Jul 22, 2026, 8:05 PM UTC)
commit
88c5129a
88c5129a4bceae39c102179ef44ec5d8b0296016
parent
3f46cce6
3f46cce62326487eb255223c4c9c1dc30e40d9ce
+5
-4
+2
-1
bench/vz_bench.sh
+2
-1
bench/vz_bench.sh
···
51
51
# TCP_WRITE_BATCH_BYTES host-socket logical batch target (default 1048576)
52
52
# TCP_LIMIT_OUTPUT_BYTES guest tcp_limit_output_bytes (TSQ allowance),
53
53
# applied through the guest agent once it is up; unset keeps the
54
-
# kernel default (~1 MiB = two BIG TCP frames, PROFILING.md E26)
54
+
# kernel default (4 MiB = eight BIG TCP frames on the pinned
55
+
# 6.18.39 guest, read back in /stats; PROFILING.md E26)
55
56
# GC_SPACE_OVERHEAD / VIRTIO_SHARD_INGEST / VM_CPUS / VM_MEMORY_MB /
56
57
# VIRTIO_MERGEABLE
57
58
# forwarded to the matching vz flags; OCAMLRUNPARAM (e.g. s=4M)
+3
-3
guest/vz_net_agent.ml
+3
-3
guest/vz_net_agent.ml
···
27
27
| () -> Ok ()
28
28
| exception Sys_error message -> Error message
29
29
30
-
(* The kernel's per-flow TSQ allowance. Its 1 MiB default is two 524280-byte
31
-
BIG TCP frames in flight, so a host that holds transmit completions across
32
-
one write cycle stalls the sender (PROFILING.md E26). *)
30
+
(* The kernel's per-flow TSQ allowance: the in-flight budget that prices
31
+
PROFILING.md E26's release cliff. The pinned 6.18.39 guest defaults to
32
+
4 MiB, eight 524280-byte BIG TCP frames; /stats reports the live value. *)
33
33
let tcp_limit_output_path = "/proc/sys/net/ipv4/tcp_limit_output_bytes"
34
34
35
35
let set_tcp_limit_output bytes =