GET /_smallwebwaf/metrics answers in the Prometheus text format for a
request carrying SWWAF_METRICS_TOKEN, 401 without it and 404 while it is
unset. Every request under /_smallwebwaf/ but the health check now goes
through the checks and is answered where it would be forwarded, 404 for
any path but the metrics, so none reaches the app. SWWAF_METRICS_TOP_N
bounds the series by country, the rest counted as other.
Judgement call: a request answered at smallwebwaf's own endpoints is
neither forwarded nor refused in the client's history.
Deviation: go.mod and go.sum written by hand from the module proxy and
sum.golang.org, as go runs only through make.
Deviation: no metrics yet for state files read again after an edit or
edits set aside; that work is not merged.
Model: opus-5-5
The tests that have GeoJS asked now run in a synctest bubble, so a wait
lasts exactly as long as it should however slowly the test process runs:
a new client's wait is checked to be exactly one second, and its next
request exactly no wait. A request waiting on the network would stop the
bubble's clock, so the stand-in for GeoJS now answers in place of the
network, through a transport that a test-only file lets the tests set.
The test of the failure log uses an abandoned request instead of a
closed port.
Model: opus-5-5
smallwebwaf now copies its state to bans.json, clients.json and
lookups.json in SWWAF_STATE_DIR, as "Persistent state" in SPEC.md
describes, and reads them back at start, so a restart lifts no ban and
gives no client a fresh allowance. Each client gains a history, and a
ban's notes count the netblock's requests. bans.json is written
SWWAF_STATE_WRITE_DELAY after a ban, and every file every
SWWAF_STATE_COUNTER_INTERVAL and at the stop. A ban read back is masked
to its netblock and refuses every client in it. A file that does not
parse, an unknown version, an entry without a field it needs, or an
unwritable directory stops the start.
Deviation: no AS number or name, and no ban cause, reason or lifting yet.
Model: opus-5-5
A request over a rate limit is refused with SWWAF_BAN_RESPONSE and bans
the client's netblock: an hour at first, three times the last ban when
broken again within a day of its end, permanent past seven days. The
ban ledger in internal/bans is checked after the static lists and
before the lookup, and the requests it refuses are not counted. A ban
resets the client's counters and carries notes holding the request
that broke the limit, as SPEC.md now says. At most SWWAF_MAX_BANS are
held. SWWAF_BAN_RESPONSE also answers SWWAF_DENY_NETS and the country
lists.
Judgement call: the six ban settings cannot be off.
Judgement call: a permanent ban's ban_expires is "permanent".
Model: opus-5-5
SWWAF_CLIENT_REQUEST_HEADER_MAX_BYTES (default 32K) and
SWWAF_CLIENT_IDLE_TIMEOUT (default 120s) replace the two values the
proxy fixed. The idle time is read like the other durations, and can
be off.
Go's server reads 4K past the header limit it is given before it
refuses, so it is still given the setting less 4K. The header size
must be more than 4K and cannot be off; any other value stops the
start with a message that does not offer off.
SPEC.md and README.md say so. README.md lists both settings, no longer
calls them fixed, and names them as built.
Model: opus-5-5
The vendored files are fetched from sneak/prompts commit dd4027b. This
repository's own entries come after the canonical content, at the end of
each file: /bin in .dockerignore, the Go lines of .gitignore and [*.go]
in .editorconfig; the test-support deny list has no entries of its own.
The lint phase moves to golangci-lint v2.14.0. The build stage now takes
the version from git describe on the .git the build context carries,
unless VERSION is passed, and fails when .git is present but no version
comes out. The test phase drops -count=1, which the policy says it does
not need, and keeps its tmpfs build cache. One test calls Header.Get
with X-Real-IP, as canonicalheader asks.
Model: opus-5-5
Adds SWWAF_ALLOW_NETS, SWWAF_RATE_LIMIT_EXEMPT_NETS and SWWAF_DENY_NETS,
read like SWWAF_TRUSTED_PROXIES and empty by default, and checked against
the client's own address before its country is looked up. A client in
SWWAF_ALLOW_NETS skips the country lists and the rate limits and is not
looked up. One in SWWAF_DENY_NETS is refused with 403, logged as denied
and not counted. One in SWWAF_RATE_LIMIT_EXEMPT_NETS is neither counted
nor refused by the rate limits. SWWAF_EXCLUSIVELY_ALLOWED_COUNTRIES now
refuses a private, loopback or link-local client unless SWWAF_ALLOW_NETS
lists it.
Judgement call: an address in both SWWAF_ALLOW_NETS and SWWAF_DENY_NETS is let through.
Judgement call: the size and time limits still apply to SWWAF_ALLOW_NETS.
Model: opus-5-5
The test phase spends most of its time compiling with the race
detector from an empty build cache; then come writing the test image
and the internal/proxy tests.
- Go's build cache is on a tmpfs in the test phase, so its 137 MB are
no longer written into the test image.
- TestUpgradedConnectionOutlastsTheTimeouts waits until just past
shortTimeout after the answer to the upgrade was read, by when every
timeout has started, rather than 7.5 s, so it ends with the other
timing tests.
- shortTimeout is written as waitLimit / 2, as its comment says it is.
Model: opus-5-5
The Dockerfile's last stage is now the image of "Deployment" in SPEC.md:
Ubuntu 26.04 with ca-certificates, nix-bin and runit from a dated
snapshot whose InRelease files are checked by hash, nixpkgs from its
release file checked by SHA-256, runsvinit built at a fixed commit, and
smallwebwaf as a runit service. smallwebwaf answers
/_smallwebwaf/healthz, and `smallwebwaf healthcheck`, which takes no
further argument, is the image's HEALTHCHECK. script/example-app builds
an app on the image and checks it end to end.
The Nix profile comes last on the PATH: first, busybox from nixpkgs
replaced runit's own runsvdir and sv. SPEC.md is corrected to match
what was built.
Model: opus-5-5
SWWAF_DENIED_COUNTRIES and SWWAF_EXCLUSIVELY_ALLOWED_COUNTRIES refuse a
request with 403 before its body is read or rate-limited, logged as
country_denied. internal/lookup asks GeoJS only while a list is set, 200
clients per request, one at a time, keeping answers 7 days. Failures, a
redirect or an answer leaving an address out included, are logged without
addresses; GeoJS is then left alone a second, doubling to five minutes.
Private, loopback and link-local clients are never sent.
Deviation, per the issue: no SWWAF_LOOKUP_SOURCE or SWWAF_LOOKUP_TIMEOUT; 403, not SWWAF_BAN_RESPONSE.
Deviation: GeoJS's country endpoint, not geo.json.
Judgement call: an IPv6 /64 is asked about by its first address; at most 10,000 clients wait.
Judgement call: config.go lists the ISO 3166-1 codes; no widely used library holds them.
Model: opus-5-5
smallwebwaf starts each timeout as the request arrives, before the step
a test needs first: the upgrade answered, the app's buffers full, the
first part of an answer passed on. A hold-up of the test process longer
than the 300 ms timeout ran it out before that step. No test can make
that step come first, and in the "waiting on the app" cases it cannot
see which side smallwebwaf was waiting on, so the timeout is now 5 s,
the hold-up wantTimedOut already allows. The timeout tests set when it
must not run out goes from 10 s to 1 m to stay clear of it. The upgrade
test waits 7.5 s past the upgrade.
Judgement call: one shared value; the proxy tests take about 8 s, not 2 s.
Model: opus-5-5
TestAnswers502WhenTheAppCannotBeReached closed a listener and pointed
smallwebwaf at its port, which another test running in parallel could
open in between, so the test sometimes got that server's answer instead
of a 502. It now uses 127.0.0.1:1: nothing listens there, and a test
listening on port 0 is always given a port from 32768 up, so no test
can take it. No other test reuses a closed listener's port.
Model: opus-5-5
In the cases where the app reads and the client stops sending halfway,
smallwebwaf waits on the client only once it has passed the first bytes
of the body to the app. A test process held up for the whole 300 ms
timeout before then rightly gets 504, as SPEC.md asks, so the test was
wrong to expect 408 every time.
The app in those cases now records whether it received any of the body.
Once the app has finished with the request, the case expects 408 and its
log line if it did, and 504 and its log line if not.
Model: opus-5-5
Each client, one IPv4 address or one IPv6 /64, is counted in two buckets
per window, the earlier weighted by how much of it the window covers; at
most 20,000 clients are kept, least recently seen dropped first. A
request over SWWAF_RATE_LIMIT_PER_MINUTE, _HOUR or _DAY (1000, 10000,
50000, or off) gets 429 before reaching the app. Refused requests count,
413s included. A clock set back over a second behind a bucket's start
restarts that window. The log line gains limit_hit and the action
rate_limited.
Deviation from SPEC.md, per the issue: the 20,000 bound and /64 are fixed.
Judgement call: golang-lru/v2 holds the table; httprate does not count refused requests.
Deviation: go.mod and go.sum hand-written; no make target tidies them.
Model: opus-5-5
Milestone 1, the repo's first code. smallwebwaf passes each request to the app and the answer back unchanged, streaming bodies and WebSocket upgrades, within four timeouts (client and app, request and response) and two size limits, and writes one JSON line per request to stdout. Every setting has an SWWAF_ name and a default, and an invalid value stops the start. The repo gets the standard layout: script/ entrypoints, make targets that call them, a Dockerfile that runs the checks, and the Gitea workflow.
Disclosure: SPEC.md changed. Go's server reads the request line and headers before smallwebwaf sees the request, so slow headers are closed without an answer, and neither slow nor oversized headers get a log line.
Disclosure: standard library only.
Model: opus-5-5
clawbot