tcp: add support for reassembling out-of-order segments

Currently out-of-order segments will result in cutting a stream into
two pieces while the out-of-order segment itself is ignored. For
example, a stream of segments "ABDCE" is interpreted as "AB", "DE" with
"C" ignored. This behavior breaks TLS decryption or prevent application
layer PDUs (such as HTTP requests/responses) from being reconstructed.
To fix this, buffer segments when a gap is detected.

The proposed approach extends the "multi-segment PDU" (MSP) mechanism
which is normally used for linking multiple, sequential TCP segments
into a single PDU. When a gap is detected between segments, it is
assumed that the segments within this gap are out-of-order and will be
received (or retransmitted) later.

The current implementation has a limitation though, if multiple gaps
exist, then the subdissector will only be called when all gaps are
filled (the subdissector will receive segments later than necessary).
For example with "ACEBD", "ABC" can already be processed after "B" is
received (with "E" still buffered), but due to how MSP are extended, it
must receive "D" too before it reassembles "ABCDE". In practice this
could mean that the request/response times between HTTP requests and
responses are slightly off, but at least the stream is correct now.
(These limitations are documented in the User's Guide.)

As the feature fails at least the 802.11 decryption test where packets
are missing (instead of OoO), hide this feature behind a preference.

Tested with captures containing out-of-order TCP segments from the
linked bug reports, comparing the effect of toggling the preference on
the summary output of tshark, the verbose output (-V) and the two-pass
output (-2 or -2V). Captures marked with "ok" just needed "simple"
out-of-order handling. Captures marked with "ok2" additionally required
the reassembly API change to set the correct reassembled length.

This change does "regress" on bug 10289 though when the preference is
enabled as retransmitted single-segment PDUs are now passed to
subdissectors. I added a TODO comment for this unrelated cosmetic issue.

Bug: 3389   # capture 2907 (HTTP) ok
Bug: 4727   # capture 4590 (HTTP) ok
Bug: 9461   # capture 12130 (TLS/HTTP/RPC-over-HTTP +key 12131) ok
Bug: 12006  # capture 14236 (HTTP) ok2; capture 15261 (HTTP) ok
Bug: 13517  # capture 15370 (HTTP) ok; capture 16059 (MQ) ok
Bug: 13754  # capture 15593 (MySQL) ok2
Bug: 14649  # capture 16305 (WebSocket) ok
Change-Id: If3938c5c1c96db8f7f50e39ea779f623ce657d56
Reviewed-on: https://code.wireshark.org/review/27943
Petri-Dish: Peter Wu <peter@lekensteyn.nl>
Tested-by: Petri Dish Buildbot
Reviewed-by: Anders Broman <a.broman58@gmail.com>

1 files changed, 12 insertions, 0 deletions

diff --git a/epan/reassemble.h b/epan/reassemble.h
index 449d336b19..3c234e8ab1 100644
--- a/epan/reassemble.h
+++ b/epan/reassemble.h
@@ -365,6 +365,18 @@ fragment_set_tot_len(reassembly_table *table, const packet_info *pinfo,
 		     const guint32 id, const void *data, const guint32 tot_len);
 
 /*
+ * Similar to fragment_set_tot_len, it sets the expected number of bytes (for
+ * fragment_add functions) for a previously started reassembly. If the specified
+ * length already matches the reassembled length, then nothing will be done.
+ *
+ * If the fragments were previously reassembled, then this state will be
+ * cleared, allowing new fragments to extend the reassembled result again.
+ */
+void
+fragment_reset_tot_len(reassembly_table *table, const packet_info *pinfo,
+		       const guint32 id, const void *data, const guint32 tot_len);
+
+/*
  * Return the expected index for the last block (for fragment_add_seq functions)
  * or the expected number of bytes (for fragment_add functions).
  */