When Zeek flips roles of a HTTP connection subsequent to the HTTP analyzer
being attached, that analyzer would not update its own ContentLine analyzer
state, resulting in the wrong ContentLine analyzer being switched into
plain delivery mode.
In debug builds, this would result in assertion failures, in production
builds, the HTTP analyzer would receive HTTP bodies as individual header
lines, or conversely, individual header lines would be delivered as a
large chunk from the ContentLine analyzer.
PCAPs were generated locally using tcprewrite to select well-known-http ports
for both endpoints, then editcap to drop the first SYN packet.
Kudos to @JordanBarnartt for keeping at it.
Closes#3789
This reverts part of commit a0888b7e36 due
to inhibiting analyzer violations when parsing non SSH traffic when
the &restofdata path is entered.
@J-Gras reported the analyzer not being disabled when sending HTTP
traffic on port 22.
This adds the verbose analyzer.log baselines such that future improvements
of these scenarios become visible.
The Spicy analyzer is added as a child analyzer when enabled and the
WebSocket.cc logic dispatches between the BinPac and Spicy version.
It substantially slower when tested against a somewhat artificial
2.4GB PCAP. The first flamegraph indicates that the unmask() function
stands out with 35% of all samples, and above it shared_ptr samples.
A continuation frame has the same type as the first frame, but that
information wasn't used nor kept, resulting payload of continuation
frames not being forwarded. The pcap was created with a fake Python
server and a bit of message crafting.
* origin/topic/awelzel/3424-http-upgrade-websocket-v1:
websocket: Handle breaking from WebSocket::configure_analyzer()
websocket: Address review feedback for BinPac code
fuzzers: Add WebSocket fuzzer
websocket: Fix crash for fragmented messages
websocket: Verify Sec-WebSocket-Key/Accept headers and review feedback
btest/websocket: Test for coalesced reply-ping
HTTP/CONNECT: Also weird on extra data in reply
HTTP/Upgrade: Weird when more data is available
ContentLine: Add GetDeliverStreamRemainingLength() accessor
HTTP: Drain event queue after instantiating upgrade analyzer
btest/http: Explain switching-protocols test change as comment
WebSocket: Introduce new analyzer and log
HTTP: Add mechanism to instantiate Upgrade analyzer
The &transient attribute does not work well with $element as that won't
be available within &until anymore apparently.
Found after a few seconds building out the fuzzer.
Don't log them, they are random and arbitrary in the normal case. Users
can do the following to log them if wanted.
redef += WebSocket::Info$client_key += { &log };
redef += WebSocket::Info$server_accept += { &log };
Add a constructed PCAP where the HTTP/websocket server send a WebSocket
ping message directly with the packet of the HTTP reply. Ensure this is
interpreted the same as if the WebSocket message is in a separate packet
following the HTTP reply.
For the server side this should work, for the client side we'd need to
synchronize suspend parsing the client side as we currently cannot quite
know whether it's a pipelined HTTP request following, or upgraded protocol
data and we don't have "suspend parsing" functionality here.
DPD enables HTTP based on the content of the WebSocket frames. However,
it's not HTTP, the protocol is x-kaazing-handshake and the server sends
some form of status/acknowledge to the client first, so the HTTP and the
HTTP analyzer receives that as the first bytes of the response and
bails, oh well.
This adds a new WebSocket analyzer that is enabled with the HTTP upgrade
mechanism introduced previously. It is a first implementation in BinPac with
manual chunking of frame payload. Configuration of the analyzer is sketched
via the new websocket_handshake() event and a configuration BiF called
WebSocket::__configure_analyzer(). In short, script land collects WebSocket
related HTTP headers and can forward these to the analyzer to change its
parsing behavior at websocket_handshake() time. For now, however, there's
no actual logic that would change behavior based on agreed upon extensions
exchanged via HTTP headers (e.g. frame compression). WebSocket::Configure()
simply attaches a PIA_TCP analyzer to the WebSocket analyzer for dynamic
protocol detection (or a custom analyzer if set). The added pcaps show this
in action for tunneled ssh, http and https using wstunnel. One test pcap is
Broker's WebSocket traffic from our own test suite, the other is the
Jupyter websocket traffic from the ticket/discussion.
This commit further adds a basic websocket.log that aggregates the WebSocket
specific headers (Sec-WebSocket-*) headers into a single log.
Closes#3424
The BDAT analyzer should be supporting uint64_t sized chunks reasonably well,
but the ContentLine analyzer does not, And also, I totally got types for
RemainingChunkSize() and in DeliverStream() wrong, resulting in overflows
and segfaults when very large chunk sizes were used.
Tickled by OSS-Fuzz. Actually running the fuzzer locally only took a
few minutes to find the crash, too. Embarrassing.
OSS-Fuzz managed to produce a MIME multipart message construction with
thousands of nested entities (or that's what Zeek makes out of it anyhow).
Prevent such deep analysis by capping at a nesting depth of 100,
preventing unnecessary resource usage. A new weird named exceeded_mime_max_depth
is reported when this limit is reached.
This change reduces the runtime of the OSS-Fuzz reproducer from ~45 seconds
to ~2.5 seconds.
The test PCAP was produced from a Python script using the email package
and sending the rendered version via POST to a HTTP server.
Closes#208
OSS-Fuzz found that providing an invalid BDAT line would tickle an
assert in UpdateState(). The BDAT state was never initialized, but
within UpdateState() that was expected.
This also removes the AnalyzerViolation() call for bad BDAT commands
and instead raises a weird. The SMTP analyzer is very lax and not triggering
the violation allows to parse the server's response to such an invalid
command.
PCAP files produced by a custom Python SMTP client against Postfix.
The initial (prefix) and final (suffix) strings are specified individually
with a variable number of "any" matches that can occur between these.
The previous implementation assumed a single string and rendered it
as *<string>*.
Reported and PCAP provided by @martinvanhensbergen, thanks!
Closeszeek/spicy-ldap#27
Skimming through the RFC, the previous approach of having containers for most
fields seems unfounded for normal protocol operation. The new weirds could just
as well be considered protocol violations. Outside of duplicated or missed data
they just shouldn't happen for well-behaved client/server behavior.
Additionally, with non-conformant traffic it would be trivial to cause
unbounded state growth and immense log record sizes.
Unfortunately, things have become a bit clunky now.
Closes#3504
We already had these declared in dns/const.zeek, so extend the parser
as well to avoid raising weirds and add some test pcaps:
$ dig @8.8.8.8 DNSKEY ed448.no
$ dig @8.8.8.8 ed448.no +dnssec
And the same for the ed25519.no domain.
Closes#3453
This commit adds a multitude of new extension types that were added in
the last few years; it also adds grease values to extensions, curves,
and ciphersuites.
Furthermore, it adds a test that contains a encrypted-client-hello
key-exchange (which uses several extension types that we do not have in
our baseline so far).
The ssl_history field may grow unbounded (e.g., ssl_alert event). Prevent this
by capping using a configurable limit (default 100) and raise a weird once reached.
When http_reply events are received before http_request events, either
through faking traffic or possible re-ordering, it is possible to trigger
unbounded state growth due to later http_requests never being matched
again with responses.
Prevent this by synchronizing request/response counters when late
requests come in.
Also forcefully flush pending requests when http_replies are never
observed either due to the analyzer having been disabled or because
half-duplex traffic.
Fixes#1705
This works around the new semantics of is_orig=T for "connections"
from DHCP servers to broadcast addresses. IMO, having the server address
as originator in the conn.log is still more intuitive.
