3. What's New in this Release

Privoxy 3.0.19 is a stable release. The changes since 3.0.18 stable are:

• Bug fixes:

  • Prevent a segmentation fault when de-chunking buffered content. It could be triggered by malicious web servers if Privoxy was configured to filter the content and running on a platform where SIZE_T_MAX isn't larger than UINT_MAX, which probably includes most 32-bit systems. On those platforms, all Privoxy versions before 3.0.19 appear to be affected. To be on the safe side, this bug should be presumed to allow code execution as proving that it doesn't seems unrealistic.

  • Do not expect a response from the SOCKS4/4A server until it got something to respond to. This regression was introduced in 3.0.18 and prevented the SOCKS4/4A negotiation from working. Reported by qqqqqw in #3459781.

• General improvements:

  • Fix an off-by-one in an error message about connect failures.

  • Use a GNUMakefile variable for the webserver root directory and update the path. Sourceforge changed it which broke various web-related targets.

  • Update the CODE_STATUS description.

The following changes were made between 3.0.17 and 3.0.18:

• Bug fixes:

  • If a generated redirect URL contains characters RFC 3986 doesn't permit, they are (re)encoded. Not doing this makes Privoxy versions from 3.0.5 to 3.0.17 susceptible to HTTP response splitting (CWE-113) attacks if the +fast-redirects{check-decoded-url} action is used.

  • Fix a logic bug that could cause Privoxy to reuse a server socket after it got tainted by a server-header-tagger-induced block that was triggered before the whole server response had been read. If keep-alive was enabled and the request following the blocked one was to the same host and using the same forwarding settings, Privoxy would send it on the tainted server socket. While the server would simply treat it as a pipelined request, Privoxy would later on fail to properly parse the server's response as it would try to parse the unread data from the first response as server headers for the second one. Regression introduced in 3.0.17.

  • When implying keep-alive in client_connection(), remember that the client didn't. Fixes a regression introduced in 3.0.13 that would cause Privoxy to wait for additional client requests after receiving a HTTP/1.1 request with "Connection: close" set and connection sharing enabled. With clients which terminates the client connection after detecting that the whole body has been received it doesn't really matter, but with clients that don't the connection would be kept open until it timed out.

  • Fix a subtle race condition between prepare_csp_for_next_request() and sweep(). A thread preparing itself for the next client request could briefly appear to be inactive. If all other threads were already using more recent files, the thread could get its files swept away under its feet. So far this has only been reproduced while stress testing in valgrind while touching action files in a loop. It's unlikely to have caused any actual problems in the real world.

  • Disable filters if SDCH compression is used unless filtering is forced. If SDCH was combined with a supported compression algorithm, Privoxy previously could try to decompress it and ditch the Content-Encoding header even though the SDCH compression wasn't dealt with. Reported by zebul666 in #3225863.

  • Make a copy of the --user value and only mess with that when splitting user and group. On some operating systems modifying the value directly is reflected in the output of ps and friends and can be misleading. Reported by zepard in #3292710.

  • If forwarded-connect-retries is set, only retry if Privoxy is actually forwarding the request. Previously direct connections would be retried as well.

  • Fixed a small memory leak when retrying connections with IPv6 support enabled.

  • Remove an incorrect assertion in compile_dynamic_pcrs_job_list() It could be triggered by a pcrs job with an invalid pcre pattern (for example one that contains a lone quantifier).

  • If the --user argument user[.group] contains a dot, always bail out if no group has been specified. Previously the intended, but undocumented (and apparently untested), behaviour was to try interpreting the whole argument as user name, but the detection was flawed and checked for '0' instead of '\0', thus merely preventing group names beginning with a zero.

  • In html_code_map[], use a numeric character reference instead of ' which wasn't standardized before XHTML 1.0.

  • Fix an invalid free when compiled with FEATURE_GRACEFUL_TERMINATION and shut down through http://config.privoxy.org/die

  • In get_actions(), fix the "temporary" backwards compatibility hack to accept block actions without reason. It also covered other actions that should be rejected as invalid. Reported by Billy Crook.

• General improvements:

  • Privoxy can (re)compress buffered content before delivering it to the client. Disabled by default as most users wouldn't benefit from it.

  • The +fast-redirects{check-decoded-url} action checks URL segments separately. If there are other parameters behind the redirect URL, this makes it unnecessary to cut them off by additionally using a +redirect{} pcrs command. Initial patch submitted by Jamie Zawinski in #3429848.

  • When loading action sections, verify that the referenced filters exist. Currently missing filters only result in an error message, but eventually the severity will be upgraded to fatal.

  • Allow to bind to multiple separate addresses. Patch set submitted by Petr Pisar in #3354485.

  • Set socket_error to errno if connecting fails in rfc2553_connect_to(). Previously rejected direct connections could be incorrectly reported as DNS issues if Privoxy was compiled with IPv6 support.

  • Adjust url_code_map[] so spaces are replaced with %20 instead of '+' While '+' can be used by client's submitting form data, this is not actually what Privoxy is using the lookups for. This is more of a cosmetic issue and doesn't fix any known problems.

  • When compiled without FEATURE_FAST_REDIRECTS, do not silently ignore +fast-redirect{} directives

  • Added a workaround for GNU libc's strptime() reporting negative year values when the parsed year is only specified with two digits. On affected systems cookies with such a date would not be turned into session cookies by the +session-cookies-only action. Reported by Vaeinoe in #3403560

  • Fixed bind failures with certain GNU libc versions if no non-loopback IP address has been configured on the system. This is mainly an issue if the system is using DHCP and Privoxy is started before the network is completely configured. Reported by Raphael Marichez in #3349356. Additional insight from Petr Pisar.

  • Privoxy log messages now use the ISO 8601 date format %Y-%m-%d. It's only slightly longer than the old format, but contains the full date including the year and allows sorting by date (when grepping in multiple log files) without hassle.

  • In get_last_url(), do not bother trying to decode URLs that do not contain at least one '%' sign. It reduces the log noise and a number of unnecessary memory allocations.

  • In case of SOCKS5 failures, dump the socks response in the log message.

  • Simplify the signal setup in main().

  • Streamline socks5_connect() slightly.

  • In socks5_connect(), require a complete socks response from the server. Previously Privoxy didn't care how much data the server response contained as long as the first two bytes contained the expected values. While at it, shrink the buffer size so Privoxy can't read more than a whole socks response.

  • In chat(), do not bother to generate a client request in case of direct CONNECT requests. It will not be used anyway.

  • Reduce server_last_modified()'s stack size.

  • Shorten get_http_time() by using strftime().

  • Constify the known_http_methods pointers in unknown_method().

  • Constify the time_formats pointers in parse_header_time().

  • Constify the formerly_valid_actions pointers in action_used_to_be_valid().

  • Introduce a GNUMakefile MAN_PAGE variable that defaults to privoxy.1. The Debian package uses section 8 for the man page and this should simplify the patch.

  • Deduplicate the INADDR_NONE definition for Solaris by moving it to jbsockets.h

  • In block_url(), ditch the obsolete workaround for ancient Netscape versions that supposedly couldn't properly deal with status code 403.

  • Remove a useless NULL pointer check in load_trustfile().

  • Remove two useless NULL pointer checks in load_one_re_filterfile().

  • Change url_code_map[] from an array of pointers to an array of arrays It removes an unnecessary layer of indirection and on 64bit system reduces the size of the binary a bit.

  • Fix various typos. Fixes taken from Debian's 29_typos.dpatch by Roland Rosenfeld.

  • Add a dok-tidy GNUMakefile target to clean up the messy HTML generated by the other dok targets.

  • GNUisms in the GNUMakefile have been removed.

  • Change the HTTP version in static responses to 1.1

  • Synced config.sub and config.guess with upstream 2011-11-11/386c7218162c145f5f9e1ff7f558a3fbb66c37c5.

  • Add a dedicated function to parse the values of toggles. Reduces duplicated code in load_config() and provides better error handling. Invalid or missing toggle values are now a fatal error instead of being silently ignored.

  • Terminate HTML lines in static error messages with \n instead of \r\n.

  • Simplify cgi_error_unknown() a bit.

  • In LogPutString(), don't bother looking at pszText when not actually logging anything.

  • Change ssplit()'s fourth parameter from int to size_t. Fixes a clang complaint.

  • Add a warning that the statistics currently can't be trusted. Mention Privoxy-Log-Parser's --statistics option as an alternative for the time being.

  • In rfc2553_connect_to(), start setting cgi->error_message on error.

  • Change the expected status code returned for http://p.p/die depending on whether or not FEATURE_GRACEFUL_TERMINATION is available.

  • In cgi_die(), mark the client connection for closing. If the client will fetch the style sheet through another connection it gets the main thread out of the accept() state and should thus trigger the actual shutdown.

  • Add a proper CGI message for cgi_die().

  • Don't enforce a logical line length limit in read_config_line().

  • Slightly refactor server_last_modified() to remove useless gmtime*() calls.

  • In get_content_type(), also recognize '.jpeg' as JPEG extension.

  • Add '.png' to the list of recognized file extensions in get_content_type().

  • In block_url(), consistently use the block reason "Request blocked by Privoxy" In two places the reason was "Request for blocked URL" which hides the fact that the request got blocked by Privoxy and isn't necessarily correct as the block may be due to tags.

  • In listen_loop(), reload the configuration files after accepting a new connection instead of before. Previously the first connection that arrived after a configuration change would still be handled with the old configuration.

  • In chat()'s receive-data loop, skip a client socket check if the socket will be written to right away anyway. This can increase the transfer speed for unfiltered content on fast network connections.

  • The socket timeout is used for SOCKS negotiations as well which previously couldn't timeout.

  • Don't keep the client connection alive if any configuration file changed since the time the connection came in. This is closer to Privoxy's behaviour before keep-alive support for client connection has been added and also less confusing in general.

  • Treat all Content-Type header values containing the pattern 'script' as a sign of text. Reported by pribog in #3134970.

• Action file improvements:

  • Moved the site-specific block pattern section below the one for the generic patterns so for requests that are matched in both, the block reason for the domain is shown which is usually more useful than showing the one for the generic pattern.

  • Remove -prevent-compression from the fragile alias. It's no longer used anywhere by default and isn't known to break stuff anyway.

  • Add a (disabled) section to block various Facebook tracking URLs. Reported by Dan Stahlke in #3421764.

  • Add a (disabled) section to rewrite and redirect click-tracking URLs used on news.google.com. Reported by Dan Stahlke in #3421755.

  • Unblock linuxcounter.net/. Reported by Dan Stahlke in #3422612.

  • Block 'www91.intel.com/' which is used by Omniture. Reported by Adam Piggott in #3167370.

  • Disable the handle-as-empty-doc-returns-ok option and mark it as deprecated. Reminded by tceverling in #2790091.

  • Add ".ivwbox.de/" to the "Cross-site user tracking" section. Reported by Nettozahler in #3172525.

  • Unblock and fast-redirect ".awin1.com/.*=http://". Reported by Adam Piggott in #3170921.

  • Block "b.collective-media.net/".

  • Widen the Debian popcon exception to "qa.debian.org/popcon". Seen in Debian's 05_default_action.dpatch by Roland Rosenfeld.

  • Block ".gemius.pl/" which only seems to be used for user tracking. Reported by johnd16 in #3002731. Additional input from Lee and movax.

  • Disable banners-by-size filters for '.thinkgeek.com/'. The filter only seems to catch pictures of the inventory.

  • Block requests for 'go.idmnet.bbelements.com/please/showit/'. Reported by kacperdominik in #3372959.

  • Unblock adainitiative.org/.

  • Add a fast-redirects exception for '.googleusercontent.com/.*=cache'.

  • Add a fast-redirects exception for webcache.googleusercontent.com/.

  • Unblock http://adassier.wordpress.com/ and http://adassier.files.wordpress.com/.

• Filter file improvements:

  • Let the yahoo filter hide '.ads'.

  • Let the msn filter hide overlay ads for Facebook 'likes' in search results and elements with the id 's_notf_div'. They only seem to be used to advertise site 'enhancements'.

  • Let the js-events filter additionally disarm setInterval(). Suggested by dg1727 in #3423775.

• Documentation improvements:

  • Clarify the effect of compiling Privoxy with zlib support. Suggested by dg1727 in #3423782.

  • Point out that the SourceForge messaging system works like a black hole and should thus not be used to contact individual developers.

  • Mention some of the problems one can experience when not explicitly configuring an IP addresses as listen address.

  • Explicitly mention that hostnames can be used instead of IP addresses for the listen-address, that only the first address returned will be used and what happens if the address is invalid. Requested by Calestyo in #3302213.

• Log message improvements:

  • If only the server connection is kept alive, do not pretend to wait for a new client request.

  • Remove a superfluous log message in forget_connection().

  • In chat(), properly report missing server responses as such instead of calling them empty.

  • In forwarded_connect(), fix a log message nobody should ever see.

  • Fix a log message in socks5_connect(), a failed write operation was logged as failed read operation.

  • Let load_one_actions_file() properly complain about a missing '{' at the beginning of the file. Simply stating that a line is invalid isn't particularly helpful.

  • Do not claim to listen on a socket until Privoxy actually does. Patch submitted by Petr Pisar #3354485

  • Prevent a duplicated LOG_LEVEL_CLF message when sending out the "no-server-data" response.

  • Also log the client socket when dropping a connection.

  • Include the destination host in the 'Request ... marked for blocking. limit-connect{...} doesn't allow CONNECT ...' message Patch submitted by Saperski in #3296250.

  • Prevent a duplicated log message if none of the resolved IP addresses were reachable.

  • In connect_to(), do not pretend to retry if forwarded-connect-retries is zero or unset.

  • When a specified user or group can't be found, put the name in single-quotes when logging it.

  • In rfc2553_connect_to(), explain getnameinfo() errors better.

  • Remove a useless log message in chat().

  • When retrying to connect, also log the maximum number of connection attempts.

  • Rephrase a log message in compile_dynamic_pcrs_job_list(). Divide the error code and its meaning with a colon. Call the pcrs job dynamic and not the filter. Filters may contain dynamic and non-dynamic pcrs jobs at the same time. Only mention the name of the filter or tagger, but don't claim it's a filter when it could be a tagger.

  • In a fatal error message in load_one_actions_file(), cover both URL and TAG patterns.

  • In pcrs_strerror(), properly report unknown positive error code values as such. Previously they were handled like 0 (no error).

  • In compile_dynamic_pcrs_job_list(), also log the actual error code as pcrs_strerror() doesn't handle all errors reported by pcre.

  • Don't bother trying to continue chatting if the client didn't ask for it. Reduces log noise a bit.

  • Make two fatal error message in load_one_actions_file() more descriptive.

  • In cgi_send_user_manual(), log when rejecting a file name due to '/' or '..'.

  • In load_file(), log a message if opening a file failed. The CGI error message alone isn't too helpful.

  • In connection_destination_matches(), improve two log messages to help understand why the destinations don't match.

  • Rephrase a log message in serve(). Client request arrival should be differentiated from closed client connections now.

  • In serve(), log if a client connection isn't reused due to a configuration file change.

  • Let mark_server_socket_tainted() always mark the server socket tainted, just don't talk about it in cases where it has no effect. It doesn't change Privoxy's behaviour, but makes understanding the log file easier.

• configure:

  • Added a --disable-ipv6-support switch for platforms where support is detected but doesn't actually work.

  • Do not check for the existence of strerror() and memmove() twice

  • Remove a useless test for setpgrp(2). Privoxy doesn't need it and it can cause problems when cross-compiling.

  • Rename the --disable-acl-files switch to --disable-acl-support. Since about 2001, ACL directives are specified in the standard config file.

  • Update the URL of the 'Removing outdated PCRE version after the next stable release' posting. The old URL stopped working after one of SF's recent site "optimizations". Reported by Han Liu.

• Privoxy-Regression-Test:

  • Added --shuffle-tests option to increase the chances of detection race conditions.

  • Added a --local-test-file option that allows to use Privoxy-Regression-Test without Privoxy.

  • Added tests for missing socks4 and socks4a forwarders.

  • The --privoxy-address option now works with IPv6 addresses containing brackets, too.

  • Perform limited sanity checks for parameters that are supposed to have numerical values.

  • Added a --sleep-time option to specify a number of seconds to sleep between tests, defaults to 0.

  • Disable the range-requests tagger for tests that break if it's enabled.

  • Log messages use the ISO 8601 date format %Y-%m-%d.

  • Fix spelling in two error messages.

  • In the --help output, include a list of supported tests and their default levels.

  • Adjust the tests to properly deal with FEATURE_TOGGLE being disabled.

• Privoxy-Log-Parser:

  • Perform limited sanity checks for command line parameters that are supposed to have numerical values.

  • Implement a --unbreak-lines-only option to try to revert MUA breakage.

  • Accept and highlight: Added header: Content-Encoding: deflate

  • Accept and highlight: Compressed content from 29258 to 8630 bytes.

  • Accept and highlight: Client request arrived in time on socket 21.

  • Highlight: Didn't receive data in time: a.fsdn.com:443

  • Accept log messages with ISO 8601 time stamps, too.

• uagen:

  • Bump generated Firefox version to 8.0.

  • Only randomize the release date if the new --randomize-release-date option is enabled. Firefox versions after 4 use a fixed date string without meaning.