feat(docker): expose the eBPF time-window filter on kernel links

uBridge's window mode gained the semantics the B.2 spec asked for —
`tc window_drop <if> <start> <outage> <pct> [<period> [<jitter>]]`: a
single outage that traffic passes before *and* after, recurring flaps
with period >= outage, and a per-cycle jitter — so the kernel-only
`window_drop` filter type is now exposed:

- validation mirrors the tc grammar (3-5 params, ranges, period >=
  outage, jitter cap) and adds window_drop to KERNEL_ONLY_FILTERS;
  filter_inactive_filters keeps start=0 — unlike every other filter a
  zero first parameter means "the outage starts now", not "disabled".
- `_ubridge_apply_ebpf_drops` sends `tc window_drop` (positional
  pass-through) and turns the mode off explicitly when absent, under the
  same ebpf=1 gate.
- FILTERS entry (Start/Outage/Chance/Period/Jitter) + available_filters.

The schedule is relative to the apply: reconcile re-sets the classifier
on every apply, so any filter change, node restart or link reset restarts
the window.

Verified on a two-container kernel link, 15/15: a single outage
[1000, 1500, 100] arrives per-seq as [0..4, 13..19] (passing before the
window, dropped inside, passing after); recurring 1000/2000 ms = 52 %
round trip (duty cycle); period == outage = permanent outage; period <
outage = 409 before uBridge; jittered flap keeps the ~50 % duty cycle;
30 % chance on a permanent window = 53 % round trip against the
per-direction theory 1-0.7^2 = 51 %; coexist with netem; node restart
restores; available_filters kernel vs relay (409 + hidden); teardown
clean. P6b re-run as a regression on the same uBridge build: 13/13.

Worth recording for the next rebuild: the first delivered window program
loaded as root but reported ebpf=0 through gns3-server — the verifier's
non-root path loses the loop counter's bound, treats the catch-up loop as
unbounded and rejects the *whole* program (E2BIG, 8193 pending states),
which also took frequency_drop and quota down, with ebpf=0 as the only
symptom. Smaller trip caps did not help; the loop had to go (uBridge
572ca0c), on top of a Makefile header-dependency fix (53e34bb) that had
been shipping a stale program object.
This commit is contained in:
YueGuobin committed 2026-10-06 00:21:08 +08:00
1 parent 0c4ec8c38c
commit 316e7002c1
8 files changed
+300 -48

No files matched your search

+28 -8
View File
@@ -1950,7 +1950,8 @@ class DockerVM(BaseNode):
# Impairment filters become one tc netem qdisc on the veth host
# end (restored here on node restart, like the capture above);
# bpf expressions become cls_bpf match-drop classifiers and
# frequency_drop/quota become the eBPF stateful classifier.
# frequency_drop/quota/window_drop become the eBPF stateful
# classifier.
await self._ubridge_apply_netem(host_ifc, nio.filters)
await self._ubridge_apply_bpf_drops(host_ifc, nio.filters)
await self._ubridge_apply_ebpf_drops(host_ifc, nio.filters)
@@ -2379,20 +2380,28 @@ class DockerVM(BaseNode):
Translate the NIO's stateful filters into uBridge's eBPF impairment
classifier on the veth host end's clsact egress (prio 1, below
bpf_drop's 10-99): ``frequency_drop`` becomes the exact every-Nth
mode (``tc nth_drop``; -1 = drop everything maps to every 1st), and
the kernel-only ``quota`` type the byte-cap mode (``tc quota_drop``).
mode (``tc nth_drop``; -1 = drop everything maps to every 1st), the
kernel-only ``quota`` type the byte-cap mode (``tc quota_drop``),
and the kernel-only ``window_drop`` type the time-window mode
(``tc window_drop``: single outage, or recurring flaps with an
optional per-cycle jitter). The program evaluates the modes in the
fixed order nth → quota → window.
Reconcile is a plain re-set on every apply: the preceding netem
apply resets the interface (which detaches the program and drops
uBridge's per-interface registry entry), so a mode set re-loads the
program from scratch, and a mode left absent is turned off
explicitly (idempotent) so a removed filter stops dropping.
explicitly (idempotent) so a removed filter stops dropping. For
window_drop this also means the schedule restarts on every apply:
the first window opens Start-ms from the moment of this call.
Semantic difference vs the relay, documented server-side: the relay
counts packets of BOTH directions through its single filtered
bridge, while each veth end's classifier counts only the traffic
entering that container — with both endpoints applying theirs, each
direction drops every Nth independently.
direction drops every Nth independently (their window schedules
start ~simultaneously, so a round trip survives only when both
directions are outside their windows).
:param host_ifc: veth host-end interface name
:param filters: NIO filters dictionary
@@ -2406,24 +2415,28 @@ class DockerVM(BaseNode):
frequency = values_of("frequency_drop")
quota = values_of("quota")
if not frequency and not quota and self._ubridge_tc_caps is None:
window = values_of("window_drop")
if not frequency and not quota and not window and self._ubridge_tc_caps is None:
# Nothing to set and nothing was ever installed (no probe yet) —
# skip touching an old uBridge entirely.
return
caps = await self._ubridge_tc_capabilities()
if not frequency and not quota:
if not frequency and not quota and not window:
if caps.get("ebpf") == "1":
# Only when the eBPF module answers can anything be
# installed; "off" is idempotent in uBridge.
await self._ubridge_send(f'tc nth_drop "{host_ifc}" off')
await self._ubridge_send(f'tc quota_drop "{host_ifc}" off')
await self._ubridge_send(f'tc window_drop "{host_ifc}" off')
return
if caps.get("ebpf") != "1":
raise DockerError(
"Packet filter(s) {} on a kernel-datapath link need a uBridge with eBPF support "
"(tc capabilities reports no ebpf; setcap cap_bpf,cap_net_admin,cap_net_raw=ep on "
"the uBridge binary); upgrade uBridge on this compute or keep the link on the "
"relay datapath".format(", ".join(sorted(f for f in ("frequency_drop", "quota") if f in filters)))
"relay datapath".format(
", ".join(sorted(f for f in ("frequency_drop", "quota", "window_drop") if f in filters))
)
)
if frequency:
# relay semantics: -1 = drop everything, N = every Nth packet
@@ -2437,6 +2450,13 @@ class DockerVM(BaseNode):
)
else:
await self._ubridge_send(f'tc quota_drop "{host_ifc}" off')
if window:
# [start, outage, chance, period?, jitter?] — start is relative
# to this apply (the schedule restarts on every reconcile).
params = " ".join(str(int(v)) for v in window)
await self._ubridge_send(f'tc window_drop "{host_ifc}" {params}')
else:
await self._ubridge_send(f'tc window_drop "{host_ifc}" off')
async def _remove_kernel_markers(self, host_ifc):
"""
+18 -2
View File
@@ -147,6 +147,20 @@ FILTERS = [
{"name": "Chance", "minimum": 0, "maximum": 100, "type": "int", "unit": "%"},
],
},
{
"type": "window_drop",
"name": "Time window drop",
"description": "Drop packets with the given chance inside a time window starting Start ms from the "
"moment the filter is applied — a single outage (traffic passes before and after), or with a "
"period recurring flaps whose per-cycle timing a jitter randomizes (kernel-datapath links only)",
"parameters": [
{"name": "Start", "minimum": 0, "maximum": 1000000000000, "type": "int", "unit": "ms"},
{"name": "Outage", "minimum": 1, "maximum": 1000000000000, "type": "int", "unit": "ms"},
{"name": "Chance", "minimum": 0, "maximum": 100, "type": "int", "unit": "%"},
{"name": "Period", "minimum": 1, "maximum": 1000000000000, "type": "int", "unit": "ms"},
{"name": "Jitter", "minimum": 0, "maximum": 1000000000, "type": "int", "unit": "ms"},
],
},
]
@@ -308,7 +322,8 @@ class Link:
raise ControllerError(f"Invalid packet filter parameters: {e!s}")
if self._created and not self.kernel_datapath:
# Kernel-only filters (the netem extensions, the eBPF quota mode)
# Kernel-only filters (the netem extensions, the eBPF quota and
# window modes)
# have no relay equivalent. Only enforced on created links — while
# loading a project the datapath is not decided yet (a link that
# will be wired on the kernel datapath must accept them), and the
@@ -723,7 +738,8 @@ class Link:
if self.kernel_datapath:
# Kernel-datapath links serve every filter type: netem for
# delay/loss/corrupt and the netem extensions, cls_bpf for
# bpf, the eBPF classifier for frequency_drop and quota.
# bpf, the eBPF classifier for frequency_drop, quota and
# window_drop.
return FILTERS
# Relay links: hide the kernel-only types the relay cannot run
# (they would be rejected with 409 on update).
+27 -5
View File
@@ -24,10 +24,10 @@ class FilterValidationError(Exception):
# kernel/relay choice is purely topological (same compute, docker/docker).
#
# Packet filters with no uBridge *relay* equivalent: the netem extensions
# plus the eBPF quota mode run on the kernel datapath only. The relay's
# packet-filter registry only knows frequency_drop / packet_loss / delay /
# corrupt / bpf / mark.
KERNEL_ONLY_FILTERS = frozenset({"rate", "reorder", "gemodel", "duplicate", "seed", "limit", "quota"})
# plus the eBPF quota/window modes run on the kernel datapath only. The
# relay's packet-filter registry only knows frequency_drop / packet_loss /
# delay / corrupt / bpf / mark.
KERNEL_ONLY_FILTERS = frozenset({"rate", "reorder", "gemodel", "duplicate", "seed", "limit", "quota", "window_drop"})
# Jitter distributions embedded in the netem-extension uBridge (tc_netem_dist).
NETEM_DISTRIBUTIONS = ("uniform", "normal", "pareto", "paretonormal")
@@ -229,6 +229,19 @@ def validate_filter_parameters(filter_type: str, values: List[Any]) -> None:
"names": ["Quota", "Chance"],
"units": ["bytes", "%"],
},
"window_drop": {
# eBPF stateful classifier (uBridge tc window_drop): packets drop
# with the given chance inside [start, start+outage) measured from
# the moment the filter is applied — a single outage, traffic
# passes before AND after. A period makes the outage recur every
# cycle (period >= outage); a jitter re-draws each cycle's outage
# and period uniformly in nominal ± jitter (0 = the fixed
# schedule). Kernel-datapath only.
"params_count": (3, 5),
"ranges": [(0, 10**12), (1, 10**12), (0, 100), (1, 10**12), (0, 10**9)],
"names": ["Start", "Outage", "Chance", "Period", "Jitter"],
"units": ["ms", "ms", "%", "ms", "ms"],
},
"bpf": {"params_count": (1, 1), "is_text": True, "names": ["Filters"]},
}
@@ -335,6 +348,12 @@ def filter_inactive_filters(filters: Dict[str, List[Any]]) -> Dict[str, List[Any
else:
# latency>0, normal configuration
active_filters[filter_type] = values
# window_drop's first parameter is a start offset: 0 means "the
# outage starts now", not "disabled" — the filter is active whenever
# present (an outage of 0 is rejected by validation, so there is no
# natural zero-value encoding for "off").
elif filter_type == "window_drop":
active_filters[filter_type] = values
# For other filters, skip if first value is 0 or empty string (means "disabled")
elif values[0] != 0 and values[0] != "":
active_filters[filter_type] = values
@@ -398,7 +417,7 @@ def validate_all_filters(filters: Dict[str, List[Any]]) -> None:
Validate all packet filters, including cross-filter dependencies that
mirror the tc netem grammar (reorder requires delay; gemodel and
packet_loss both translate to the netem loss keyword and are mutually
exclusive).
exclusive) and the eBPF window grammar (period >= outage).
Args:
filters: Dictionary mapping filter types to their values
@@ -420,3 +439,6 @@ def validate_all_filters(filters: Dict[str, List[Any]]) -> None:
raise FilterValidationError("reorder requires delay")
if "gemodel" in filters and "packet_loss" in filters:
raise FilterValidationError("gemodel and packet_loss are mutually exclusive (both map to the netem loss keyword)")
window = filters.get("window_drop")
if window and len(window) >= 4 and int(window[3]) < int(window[1]):
raise FilterValidationError("window_drop period must be greater than or equal to the outage length")