Add DEBUG_PROXY directive around all log_dbg_level_printf() and related lines
Log stats to syslog, similar to error logs, so that it is simpler to rotate and parse
-O w/o -g is failing bufferevent_socket_connect for parent dst, so either enable -O w/ -g, or disable -O w/o -g (-O2 is failing too)
Refactoring
Check if srv_dst bev is NULL or not: Since both eventcb and writecb for srv_dst are enabled, either eventcb or writecb may get a NULL srv_dst bev, causing a crash with signal 10.
Remove unnecessary evutil_closesocket() calls, they could close the fds twice, causing conn stalls
Fix child eof bug
Flickr keeps redirecting to https with 301 unless we remove the Via line of squid, so apparently flickr assumes the existence of Via header field or squid keyword a sign of plain http, even if we are using https
Also do not send the loopback address to the Internet
Other fixes
Got rid of now unnecessary pxy_conn_is_ready_to_free*() functions
Got rid of now unnecessary duplicate child and parent free functions: pxy_*_conn_free()
Fix error handling
Insert our header line right after the first header line in the packet in case the packet is fragmented, otherwise Squid is confused when it cannot find our header line in the first packet
Fix segmentation fault upon receiving BEV_EVENT_EOF on the inbound
bufferevent while the outbound bufferevent has not received
BEV_EVENT_CONNECTED yet.
Issue: #124
Patch by: Eun Soo Park
Enable full strength DHE and ECDHE by default in order to allow modern
browsers to connect without weak crypto warnings.
Issue: #119
Reported by: @curioustwo
When other->closed is set, the bufferevent write handler accesses
other->bev even though it is invalid. Fix this access, and as added
layer of defense against future bugs, set ->bev to NULL whenever
invalidating it, except where the connection is torn down completely.
Reported by: Eun Soo Park
Introduced in: 2bcfaf417d753f
Issue: #109
Only manipulate other->bev if it is not NULL to avoid a NULL pointer
dereference in the proxy bufferevent write handler when only one
direction is fully established, for example during connection shutdown.
Reported by: @david-holonet
Introduced in: 2bcfaf4
Issue: #109
The event buffer write handler failes to re-enable the corresponding
read event of the opposite connection if the buffer is not only down to
less than half the limit, but completely emptied. In that case, the
read event would never be re-enabled and the connection would stall and
time out.
Issue: #109
Patch by: Eun Soo Park
Introduce HAVE_SSLV2, HAVE_SSLV3, HAVE_TLSV10, HAVE_TLSV11 and
HAVE_TLSV12 to indicate that support for the respective protocol is
available in OpenSSL. This was necessary due to the increased
complexity of testing version support following the phasing out of SSLv2
and SSLv3 from OpenSSL implementations. This fixes the build with
OpenSSL versions which have SSLv3 support removed.
While here, de-duplicate code for setting SSL_CTX options and do not set
SSL_OP_NO_SESSION_RESUMPTION_ON_RENEGOTIATION anymore; it has no benefit
in the context of splitting SSL/TLS for analysis.
Reported by: Jérémie Courrèges-Anglas
Refactor and unify ssl_tls_clienthello_identify() and the earlier
ssl_tls_clienthello_parse_sni() into a single
ssl_tls_clienthello_parse() function that handles parsing ClientHello
messages for different purposes. As a result, rename the debug knob
DEBUG_SNI_PARSER into DEBUG_CLIENTHELLO_PARSER.
This code looks at the beginning of each read from the src for something
that looks like an ssl client hello message; if it finds one it tries to
upgrade the connection to proxied ssl. So it works only in the simple
case where the connection has no binary data before the upgrade attempt
(so there are no false positives), and where the client hello comes at
the beginning of a packet from the source.
Store host and port in separate strings internally and get rid of the
[host]:port representation where separate host and port would be
cleaner. This includes the following user-visible changes:
- Generated filenames that contain host and port, such as by -S and
-F %d and %s, now use a host,port format instead of [host]:port.
- Connect log now uses separate fields for host and port.
Issue: #69#74
Reported by: Adam Jacob Muller
- Update copyright to 2015
- Remove the non-standard "unmodified" from the 2-clause BSD license
- Remove scalable from the tagline to avoid misinterpretations
Make -w and -W work in conjunction with dropping privileges and
chrooting by moving the cert writer code to a separate logger thread and
using the privsep framework to open the files if they do not exist
already.
Issue: #70
Local process lookup is independent of the NAT engine used, it depends
only on the operating system's process enumeration API. Moving the code
out of NAT lookup also makes it work for static and SNI proxyspecs.
Add -r to force a specific SSL/TLS protocol version.
Add -R to disable one or several SSL/TLS protocol versions.
Replace WANT_SSLV2_CLIENT and WANT_SSLV2_SERVER to WITH_SSLV2.
Issue: #30
Reported by: @Apollo2342
Migrate knobs from DISABLE_SSLV2_CLIENT and DISABLE_SSLV2_SERVER to
WANT_SSLV2_CLIENT and WANT_SSLV2_SERVER and remove the
DISABLE_SSLV2_SESSION_CACHE knob entirely, automatically including the
respective code if SSLv2 support is compiled into SSLsplit.
Also remove HTTP Strict Transport Security (HSTS, RFC 6797) headers from
HTTP responses. With HSTS active, the user is not allowed to accept
untrusted certificates.
This exposes the pid lookup code as a standard attribute
of NAT lookup -- if a matching process cannot be found,
or if pid lookup isn't supported by the NAT backend,
a pid of -1 is returned.
This also adds the local_pid to the pxyconn context; this
will be used to populate log strings.
The key type checks which are used to optimize the loading of DH and
ECDH parameters should check the type of the supplied server key, not
the global options key.
The code in pxy_ossl_servername_cb() which generated the forged
certificates did not call SSL_CTX_free() on the newly allocated SSL_CTX
struct after associating it with the SSL struct, which increments the
reference count internally. Also add some comments explaining OpenSSL
reference counting behaviour to be more explicit on what happens to the
instances that OpenSSL keeps track of.
When using passthrough mode, if a connection to a server fails with an
SSL error, sslsplit falls back to plain TCP passthrough. When
reconnecting with plain TCP, the SSL context was freed, but the file
descriptor was never closed. The fix remedies that by calling the
proper cleanup function for the dst bev before reconnecting.
Reported by: Peter Haag
Remove a superfluous call to SSL_CTX_free() which caused SSL_CTX
structures to be free'd twice under some circumstances.
Issue: #16
Reported by: Amit Chowdhary
This removes the spurious "Unknown bufferevent 0x80" debug message but
does not have any change in functionality, since return would have been
called anyway after falling down the debug message.
While here, remove the useless "ignoring event" debug message unless
DEBUG_PROXY is defined, and also print the timeout flag in debug mode.
Always explicitly name which arguments are non-null, even if all
arguments are non-null. This is to avoid bugs where newly added
arguments are automatically non-null by accident, possibly leading to
optimisation errors.
This also fixes a few potential errors related to non-null arguments;
specifically it prevents the compiler optimising away a test for sni
being NULL in cachedsess_mkkey().
Issue: #14
Reported by: kythyria
Properly reset connection state when reconnecting the dst part of the
connection. This fixes the fallback to passthrough when no certficates
are present which can be used to split the SSL.
Issue: #9
Reported by: ceear
