routes.go lists every own endpoint. One middleware checks a token and
another writes the log line. answerOwnEndpoint, answerAdmin and
endpoint() are gone; the handlers return http.HandlerFunc. The router
matches the decoded path, as the old code did. A test checks the
request_bytes of an own endpoint's log line.
Model: opus-5-5
requestBody and responseBody are read on goroutines other than the
handler's. They held the whole request; now they hold only the
functions they call and, for responseBody, the request's context.
Model: opus-5-5
internal/wafprotocol writes and reads the question, the answer and the
report, and declares once each header, path and limit both daemons use.
rproxy no longer imports internal/waf. Nothing on the wire changes.
Model: opus-5-5
rules and lookup each had the same loop: watch the directory, start a
timer at once, restart it at each change that counts, read again when it
fires. It now lives once in internal/dirwatch, with its quiet time and
its tests. Each package keeps its own log messages, its readAgain, and
its test for which changes count. internal/state keeps its own copy.
Model: opus-5-5
Main builds a cobra root command with rproxy, waf and healthcheck under
it, in place of the switch on os.Args and the hand-kept usage text.
Exit statuses stay: 2 with no command or an unknown one, otherwise what
the command returns. cobra's completion command is off. `sww help` and
`sww --help` now print the commands and exit 0.
Model: opus-5-5
Every heuristic is a struct built once in newHandler with the settings
and collaborators it reads; no check reads the handler. The six lists
are handler fields built there. Code several heuristics share takes what
it needs as arguments or lives on a struct they hold: banMaker for bans
and their alerts, reputationHitNoter for reputation hits,
biasedThresholds for limit percentages. The handler keeps blocklists,
rateLimits, trapPaths, ruleFiles and coreRuleSet as fields too, since
refusedWhateverTheAnswer and findBesideLookup use them.
Model: opus-5-5
ratelimit.New takes ratelimit.Params (Limits, MaxClients) in place of
two arguments by position. The process logger moves out of requestlog
into internal/processlog, whose New takes Params (Output, InstanceName,
Level); its two tests move with it unchanged. geojs.GeoJS keeps its
params whole, as DNSBL, AbuseIPDB and routewatch.Client do. SPEC and
README list the new package.
Model: opus-5-5
Three tests that wait out a 2- or 3-second request time limit now use 5
seconds, so the request reaches the app before it runs out under load.
The WebSocket test also uses 5-second limits, and renews its
connection's deadline after the wait.
TestFiveFailedAsksKeepTheNextRequestFromAskingWAF stops the breaker's
clock like TestAnswersThatCannotBePassedOnOpenTheBreaker.
Model: opus-5-5
requestWithBody gave the answer 10 seconds, and endedLines gave the log
line 10 seconds. On a busy host the Core Rule Set's inspection of a
request with 1000 query parameters, under the race detector, took
longer. Both now wait until the answer or line comes; go test's
-timeout ends a test where it never does.
Model: opus-5-5
The ledger's BanForLimit, WouldBanForLimit, BanForAttack and
WouldBanForAttack become Ban and WouldBan, which take the cause and the
notes in one bans.Grounds. The waf's banForLimit and banForAttack fill in
what differs and call banClient, the one function that tests observe
mode on the way to a ban. Behaviour is unchanged; tests change only
where they called the four methods.
Model: opus-5-5
TagList, MaskZoneKey, ParseNetblock, ParseASN and ParsePercent move
from config to reputation; config imports them from there. rules and
config check header names with httpguts.ValidHeaderFieldName, and
IsHeaderName goes. A new test pins that check to the characters
IsHeaderName took, for the empty name and every single-byte name.
Model: opus-5-5
The circuit breaker reads the time through a clock the handler holds,
which is time.Now outside tests. The test that answers waf's
questions with answers rproxy cannot pass on stops that clock, so how
long its requests take on a slow host no longer decides whether the
breaker opens or turns the next request away.
The test that stops rproxy mid-request sets its read deadline from the
stop, not from the send, so a slow wait for the app no longer shortens
it.
Model: opus-5-5
banAndCount runs the heuristics that only ban, then countAnomalies,
then addToHistory. A refused request's decision and an allowed
request's report both call it, each with its own list. answerQuestion
keeps one defer, which writes the log line and holds the decision for
the report.
A check that panics now leaves the request out of its client's history
and the anomaly counts, as the issue states.
Model: opus-5-5
waf wrote eight setting names as string literals into log lines and
ban notes, and config wrote each again where it reads the setting. A
rename that missed one would log a setting that does not exist. config
now exports a constant for each name, used by its reader and by waf.
Model: opus-5-5
Each of the four keeps the cancel function of every ask it started, and a
wait group of their goroutines. Its new method (GeoJS.StopAsking,
DNSBL.StopAsking, AbuseIPDB.StopChecking, routewatch Client.StopAsking)
refuses further asks, ends those under way, and returns once their
goroutines have. An ask ended that way is no failure: no alert, no failure
count, no wait before the next. Each provider in internal/sww adds the
method as a stop hook, so the last write of the state files follows it.
Model: opus-5-5
BytesCounted, LookupSource, BlocklistAction, DNSBLAbuseIPDBAction and
CoreRuleSetMode become named string types declared in config, with
constants for their values. Every comparison in waf and sww uses the
constants, and the private duplicates fileSource and deny go. The two
action settings share ReputationAction. The values and the strings
written to logs, metrics and alerts are unchanged.
Model: opus-5-5
The body of waf.NewCoreRuleSet moves into newCoreRuleSet in
internal/sww, beside the providers of the other parts; waf no longer
exports it. The waf test helper builds the Core Rule Set with
coreruleset.New while the mode is not off.
Model: opus-5-5
While SWW_WAF_ABUSEIPDB_KEY is set, waf fetches AbuseIPDB's blacklist at
the start and every SWW_WAF_ABUSEIPDB_BLACKLIST_REFRESH_INTERVAL_MINS
(60), cut at SWW_WAF_ABUSEIPDB_THRESHOLD (90, from 25 to 100). Each
listed client's netblock gets a ban of SWW_WAF_ABUSEIPDB_BAN_DAYS (30)
with the new cause abuseipdb. The blacklist itself is not kept.
These bans do not count toward SWW_WAF_MAX_BANS, so a 10,000-address
blacklist cannot push out the bans for limits and attacks. An ended
abuseipdb ban is forgotten at each fetch and each read of bans.json. A
netblock with an active ban is left alone, and so is one whose abuseipdb
ban an admin lifted, until that ban would have ended. Observe mode makes
no ban.
Model: opus-5-5
go.uber.org/fx builds the parts of waf and rproxy in internal/sww, one
provider per part. Each background loop is a lifecycle hook added where
its part is built. waf.Params takes every part built, and waf.New builds
none. GeoJS calls back into waf, so waf exports AddLookup and
GeoJSFailed, and clientGroup and banNetblock take the settings.
newServer, loadStateFiles, the inBackground list and serveRProxy are
gone. serve is the one hook that serves both daemons, and reports a
failure to run on a channel of its own; fx's signal handling is not
used. Once fx has stopped, run takes each daemon's last step: waf writes
the state files, and rproxy waits for its reports.
Model: opus-5-5
The rproxy package's tests and sww's startRProxyProcess give rproxy a
minute for waf's answers unless a test sets
SWW_RPROXY_WAF_BACKEND_TIMEOUT_MS. At the 2-second default, a slow host
made rproxy drop questions and reports.
TestMainRunsWAF runs waf with the Core Rule Set off. Main keeps waf's
500 ms limit for its checks, and under the race detector the Core Rule
Set's inspection could pass it and get 503.
Model: opus-5-5
A new example puts one gitea behind traefik, rproxy and waf. waf sets
nine settings most sites use, then eight more commented out, each with
what it does and when to set it.
Model: opus-5-5
Each SWW_WAF_ANOMALY_ threshold's description now says crossing it sends
one alert to the alert destinations and bans, rate limits and blocks no
client. The prefix and watched-netblock settings say their thresholds
only alert. README's configuration notes and docs/settings.md explain
"anomaly" once.
Model: opus-5-5
Under heavy host load the sww tests got 503 where they expected 200 or
403: waf refused requests whose trap paths, rule files and Core Rule Set
took longer than 500 ms, and rproxy gave up on waf after 2 s.
waf.Params and sww.Params gain BesideLookupTimeLimit; zero keeps the
500 ms limit, so the binary is unchanged. The sww tests' run gives waf
an hour, and startRProxy sets SWW_RPROXY_WAF_BACKEND_TIMEOUT_MS to
60000, as the rproxy package's tests do. waf's own tests set the limit
through the new field instead of SetBesideLookupTimeLimit, which is gone.
Model: opus-5-5
The limit was already a constant with no setting. These tests pin it:
a request line and headers of exactly 32 KiB reach the app, and one
byte more is answered 431 with the connection closed and waf not asked.
Model: opus-5-5
TestQueriesGoToTheResolverSWWDNSBLResolverNames queries a real UDP
stand-in outside any synctest bubble. A parallel test starts only once
every test that is not parallel has ended, so it can no longer overlap
the bubbled DNSBL tests, and no other parallel test of the package
queries a resolver.
The other tests that run one at a time stay so, each for the reason its
comment gives: environment variables, TMPDIR, timing, Go's resolver in
synctest bubbles, and garbage collection hiding a leaked connection.
Every subtest inside a parallel test already calls t.Parallel().
Model: opus-5-5
While the lookup waits for GeoJS, the trap paths, the rule files and the
Core Rule Set run at the same time as the checks after the lookup. The
first side to refuse answers the request; what the other side finds leaves
no trace. Those three checks have 500 ms together: past it, the request is
refused with 503 as checks_timed_out, with no ban and no offence.
Tests hold GeoJS and the Core Rule Set with stand-ins under synctest. The
test helper lifts the 500 ms limit for tests that run the real Core Rule
Set, which the race detector slows past it.
Model: opus-5-5
The GeoJS lookup and its cache move from internal/lookup to internal/geojs,
which imports no other package of this repository. waf hands it functions
to count requests, report failures and count lookups that went without an
answer; waf keeps the log line, the source_failure alert and the metrics.
The caller now passes its wait to each lookup. waf passes
SWW_WAF_ASN_AND_COUNTRY_LOOKUP_TIMEOUT_MS. A request to GeoJS has its own
fixed limit of 10 seconds, and only passing it or an error from GeoJS is a
failure. An answer that comes after the caller stopped waiting is kept.
internal/lookup keeps the lookup database and returns geojs.Answer.
Model: opus-5-5
The worst-case inspection test compared each inspection with a fixed 1
second, so a busy host failed it at random. It now times, on the same
thread and in processor time, a fixed regular-expression job just
before and after each inspection. An inspection fails when it takes
more than 13 times that job, about 1 second on an idle host.
Model: opus-5-5
TestFiveFailedAsksKeepTheNextRequestFromAskingWAF waited for a question
from each of the five failing asks. Under load an ask could time out
before the stand-in for waf saw it, and the test failed with "no
question". It now waits only for each visitor's answer.
Before the sixth request, it checks that rproxy has not written that the
circuit breaker is open, so none of the first five was kept from asking
waf. It checks that the sixth request did not ask by its request ID,
which the app records, so late questions from the failed asks cannot be
mistaken for it.
Model: opus-5-5
script/publish no longer defaults DOCKER_HOST to an SSH target on one of
our servers or PUBLISH_IMAGE to the registry. Both are required, and an
unset one stops it with a message naming it. The README says so.
The instance name in docs and tests was our server's name; it is now
host1/waf in the docs and host1/blog in the tests.
Model: opus-5-5
The Core Rule Set is the one check whose time a visitor can stretch.
At paranoia level 4 it took over 2 s on a request with an 8 KiB
request line, three 8 KiB Cookie lines and a 64 KiB form body.
SWW_WAF_CORE_RULE_SET_BODY_INSPECTION_SIZE now takes at most 8K. The
Core Rule Set also reads no JSON body over 1 KiB, since Coraza turns
JSON into up to 8 times its size in names and values. That brings
such a request under 1 s.
A test per kind of input times the Core Rule Set on the worst request
within the limits, at the slowest settings. SPEC.md lists the limits
that bound it.
Model: opus-5-5
Five failed asks to waf within five seconds open the breaker. An ask
fails when it gets no answer in time, slow answers included, or an
answer rproxy cannot pass on. Every usable answer clears the failures
and closes the breaker. While open, a request takes
SWW_RPROXY_WAF_BACKEND_NO_ANSWER_DISPOSITION without asking waf, except
one ask a second. SPEC.md describes it.
The sww test that stops and restarts waf now waits for the breaker to
close before it expects the ban again.
Model: opus-5-5
Adds a paragraph to "What it does" listing what rproxy enforces with
default settings whether or not waf answers: malformed requests, methods,
header size, request and response times, body sizes, idle connections.
After the list, it says which header limits are waf's.
Removes the "Slow and oversized requests" bullet, which the new list
states with figures.
Model: opus-5-5
The image sets title, description, version, created, revision and
source; the last three known only at build time come from build
arguments that default to unknown, and script/docker and
script/cibuild pass them. The example app sets its own title and
description.
Model: opus-5-5
The opening now lists what bans a client with the default settings: the
rate and byte limits, a probe for a secret path, a security scanner's
user agent, and the refusal limit. A scraper that does none of these is
left alone.
Model: opus-5-5
README, SPEC and docs no longer present traefik as the required front
proxy, and call it the TLS-terminating proxy. The README section of that
name lists what the proxy must do: forward plain HTTP to port 8080, keep
Host, set X-Forwarded-For and X-Forwarded-Proto, close an unused
connection in under 120 seconds, and connect from an address in
SWW_RPROXY_TRUSTED_PROXIES. traefik, Caddy and nginx appear as examples,
with the Caddy and nginx settings their defaults lack.
Model: opus-5-5
The header limits now also refuse a request line over 8 KiB with 414
(action request_line_too_long) and a header line over 8 KiB, Host
included, with 431. Like the other header limits they are fixed, run
first and make no ban. One long query string or cookie no longer keeps
the Core Rule Set busy for a third of a second.
Model: opus-5-5
SWW_WAF_ADD_LOOKUP_HEADERS now defaults to true, so the app gets
X-Client-ASN and X-Client-Country with no setting. While
SWW_WAF_LOOKUP_SOURCE is off it defaults to false, so turning lookups
off alone still starts; true set with the lookup off still stops the
start. With GeoJS, a new client's first request now waits for its
answer, up to SWW_WAF_ASN_AND_COUNTRY_LOOKUP_TIMEOUT_MS.
Model: opus-5-5
sneak ruled to remove it. Gone: the heuristic crowdsec, the ban cause
crowdsec and its earlier_bans count, the tag crowdsec, the settings
SWW_WAF_CROWDSEC_LAPI_URL and SWW_WAF_CROWDSEC_LAPI_KEY, the decision
list fetch and parse in internal/reputation, their tests, and every
mention in the README, SPEC.md and docs/. EVALUATION.md keeps its
comparison of tools.
The ban ledger's internal ban function loses its expires parameter,
which only CrowdSec bans used.
Model: opus-5-5
Co-authored-by: clawbot <sneak+clawbot@sneak.cloud>
Recorded from gitea 1.27.3 and Forgejo 17.0.1 with each registry's own
client, and curl for the requests no client sends, with python3,
powershell, ls -la, .htaccess, file names and plain words. q on the
Cargo, Chef, Composer and Conan searches and NuGet's search service
joins 900603. 900605 lets /bin/sh through on Composer and Conan. New:
900609 for npm's text; 900611, in phase 2, lets one q shaped as a Conan
reference with no +, */ or --, such as zlib/*, skip 942100 if 942100
passes it with each /* comment read as a space; 900612 for NuGet's
Search(), Packages() and their counts: searchTerm, $filter, and the
names $filter, $orderby and $top for 942290.
Model: opus-5-5
Co-authored-by: clawbot <sneak+clawbot@sneak.cloud>
Recorded with python3, powershell, ls -la, .htaccess, file names and plain
words from gitea 1.27.3 and Forgejo 17.0.1. q on the API's searches of
topics, of an organisation's teams and of emails joins 900603, which skips
930120, 932120, 932250 and 932260. The same four rules are skipped by
query on the search of unadopted repositories (900607) and by name on the
search of a repository's milestones (900608). 900605 also lets /bin/sh
through in q on the API's search of packages, since a package name can
hold a slash. The searches of repositories, users and issues already
passed.
Model: opus-5-5
Co-authored-by: clawbot <sneak+clawbot@sneak.cloud>
Applies the setting renames of the naming audit under the SWW_WAF_ and
SWW_RPROXY_ prefixes, with sneak's names for the IP reputation cache
time and the two time limits, which now take milliseconds. The rproxy
log field default_disposition becomes waf_backend_no_answer, and
config.Config and config.FromEnvironment become config.WAF and
config.WAFFromEnvironment. Go fields, test names and prose follow. No
aliases; nothing reads the old names.
Model: opus-5-5
Ghost 6.76.0 was sent the fields not yet recorded with python3,
powershell, ls -la and .htaccess. Fixes in ghost.conf:
- eight rules that refused words naming commands and files, 932235 among
them, for staff profiles, setup, invitations, tiers, offers,
recommendations, integrations, webhooks, feature images, comments, the
CSV import's labels, ActivityPub, and source and the utm_ fields;
- 932235 for the existing plain-text fields and code injection;
- filter skips six more rules, for title:~'powershell & python3';
- pathname skips 930120, 932120 and 932260;
- invitation addresses, email_segment, a tag's canonical_url, and
ActivityPub's handles, actorUri and domain;
- a tag's code injection skips what a post's does.
Model: opus-5-5
Go's server reads up to 256 KiB of a body left unread once the handler
returns, but not a body with a declared length from a visitor that sent
Connection: close. It closes that connection with the rest unread, and
the visitor, still sending, gets a reset. rproxy now reads that rest
itself, until the deadline already set on the connection, through a
MaxBytesReader made with Go's own ResponseWriter. With more than 256 KiB
left, that has Go's server tell the visitor it sends no more and wait
500 ms before it closes. rproxy reads the rest only after an answer with
a Content-Length, because Go's server ends any other answer only once
the handler returns.
Model: opus-5-5
While bodies are read, 932120 refused the canonical URL
`https://example.com/blog/powershell/` and 930120 refused
`https://example.com/blog/.htaccess/`. The new exclusions 900383 and 900384
switch both rules off for `posts.N.canonical_url` under
`/ghost/api/admin/posts/` and `pages.N.canonical_url` under
`/ghost/api/admin/pages/`, as `url` does, and only when the query has no field
of that name. The recorded requests gain the saves the Ghost 6.76.0 admin
sends when those addresses are set in a post's and a page's settings.
Model: opus-5-5
visit's client sends Connection: close, and Go's server then closes the
connection after the 413 without reading the rest of the body. The
visitor, still sending, could see the reset before the answer. The test
now sends the request on its own connection without that header, so Go's
server reads the rest before it closes.
Model: opus-5-5