mirror of
https://github.com/jmcorgan/fips.git
synced 2026-10-05 19:18:25 +00:00
Johnathan Corgan
32bf476660
A tree announce was counted as delivered once the transport accepted it, so a lost datagram left the peer on our old tree position until the periodic re-broadcast up to a minute later, and indefinitely on a node with a single peer, which has no periodic re-broadcast. Meanwhile the peer could leave destinations out of discovery or route toward them by stale coordinates. The tree state now keeps, per peer, the tree announce still awaiting confirmation from the link's receiver reports, using the same delivery check as filter announces. The declaration sequence is the lineage: a new sequence refills the resend budgets, while a resend or periodic re-broadcast of the same declaration spends from them. The tracking is dropped with the peer. The node resends on a reported loss, and once after 30 s when the reports cannot confirm the announce. Per declaration and session it resends at most once unchecked and three times on loss, and at most six times a minute per peer. The resend is an ordinary announce of the current declaration through the rate-limited send path, and the "Sent TreeAnnounce" trace line now carries the declaration sequence. The node tests run on a converged line with every role read from the converged tree, and share the loopback MMP, loss and rekey helpers of the filter announce tests, now generalised to node indices. The Ethernet mesh scenario's 60 s delivery window no longer sits at the ordinary recovery bound, since tree and filter announces lost to a link flap are now resent within seconds of the link returning, or after about 30 s when the reports cannot confirm them. A comment beside the assertion says so, so a red there is investigated rather than expected.
118 lines
3.9 KiB
YAML
118 lines
3.9 KiB
YAML
# Mixed transport mesh: 6 nodes, UDP + Ethernet edges
|
|
#
|
|
# Exercises both transports in a single mesh. UDP edges use static
|
|
# peer config; Ethernet edges use beacon discovery. Tests that the
|
|
# spanning tree converges across heterogeneous transports with netem
|
|
# and link flaps active, and that datagrams are delivered over the
|
|
# Ethernet links (the delivery assertion below).
|
|
#
|
|
# Topology:
|
|
#
|
|
# n01 ---udp--- n02 ---udp--- n03
|
|
# | | |
|
|
# eth eth udp
|
|
# | | |
|
|
# n04 ---eth--- n05 ---udp--- n06
|
|
#
|
|
# UDP edges: n01-n02, n02-n03, n03-n06, n05-n06
|
|
# Ethernet edges: n01-n04, n02-n05, n04-n05
|
|
|
|
scenario:
|
|
name: "ethernet-mesh"
|
|
seed: 42
|
|
duration_secs: 120
|
|
|
|
topology:
|
|
algorithm: explicit
|
|
num_nodes: 6
|
|
params:
|
|
adjacency:
|
|
- [n01, n02, udp]
|
|
- [n02, n03, udp]
|
|
- [n03, n06, udp]
|
|
- [n05, n06, udp]
|
|
- [n01, n04, ethernet]
|
|
- [n02, n05, ethernet]
|
|
- [n04, n05, ethernet]
|
|
|
|
netem:
|
|
enabled: true
|
|
default_policy:
|
|
delay_ms: [1, 10]
|
|
jitter_ms: [0, 2]
|
|
loss_pct: [0, 1]
|
|
mutation:
|
|
interval_secs: {min: 20, max: 40}
|
|
fraction: 0.3
|
|
policies:
|
|
normal:
|
|
delay_ms: [1, 10]
|
|
loss_pct: [0, 1]
|
|
degraded:
|
|
delay_ms: [30, 80]
|
|
jitter_ms: [5, 20]
|
|
loss_pct: [3, 8]
|
|
|
|
link_flaps:
|
|
enabled: true
|
|
interval_secs: {min: 20, max: 40}
|
|
max_down_links: 1
|
|
down_duration_secs: {min: 10, max: 20}
|
|
protect_connectivity: true
|
|
|
|
traffic:
|
|
enabled: false
|
|
|
|
# Baseline: the mesh came up, agreed on a root, and took parents. It
|
|
# says nothing about the data plane; it exists so that a run in which
|
|
# the mesh never formed cannot report success. Six nodes, one root, five
|
|
# parented in all six provably-completed archived runs.
|
|
#
|
|
# Delivery: the one assertion anywhere in CI that a datagram crossed an
|
|
# Ethernet link. n04's only edges are Ethernet (n01-n04, n04-n05), so a
|
|
# probe to or from n04 must cross one, provided n04 has no peer on
|
|
# another transport; its container also sits on the docker network, so
|
|
# that is checked, not assumed: n04's peers are read before and after
|
|
# the probe, and the assertion fails if it has none or any is not
|
|
# Ethernet. n04 -> n06 and n06 -> n04 cross Ethernet and then UDP;
|
|
# n04 -> n05 is a direct Ethernet hop.
|
|
#
|
|
# Load-robust by shape: each pair and size is pinged one packet at a
|
|
# time until 3 replies or 60 s, and the verdict is whether that
|
|
# happened, not a loss ratio, so a busy host slows it without failing
|
|
# it. It runs at teardown, after flapped links are restored.
|
|
#
|
|
# Recovery: a tree or bloom announce lost to a flap is resent from the
|
|
# link's receiver reports a few seconds after the link returns, or within
|
|
# about 30 s when those reports cannot confirm it; only a per-peer backoff
|
|
# built up by earlier resends on a lossy link can hold a resend longer, up
|
|
# to 60 s after the previous one. The 60 s window therefore no longer sits
|
|
# at the ordinary recovery bound, and a red here is investigated, not
|
|
# expected.
|
|
#
|
|
# Sizes: payload 0 is the smallest ICMPv6 echo and 1200 is near the
|
|
# 1280-byte TUN MTU. Neither can reach the receiver's padding trim on a
|
|
# veth link. Measured on a live run of this scenario (2026-09-19): an
|
|
# empty echo is a 226-byte Ethernet frame and a 1200-byte one 1342
|
|
# bytes, far above the 60-byte minimum; and the veth does not pad the
|
|
# frames that are shorter (54-byte control frames and 48-byte beacons
|
|
# arrived as sent; none of 250 received data frames carried bytes past
|
|
# its length field). That trim is covered instead by the unit tests on the receive loop's
|
|
# data-frame parse (data_payload in src/transport/ethernet/mod.rs).
|
|
assertions:
|
|
baseline:
|
|
min_nodes_reporting: 6
|
|
max_roots: 1
|
|
min_nodes_parented: 5
|
|
delivery:
|
|
pairs: [[n04, n06], [n06, n04], [n04, n05]]
|
|
payload_bytes: [0, 1200]
|
|
min_replies: 3
|
|
deadline_secs: 60
|
|
require_transport: ethernet
|
|
transport_node: n04
|
|
|
|
logging:
|
|
rust_log: "info"
|
|
output_dir: "./sim-results"
|