-*** Version 3.0.17 Stable ***
-
-- Fixed last-chunk-detection for responses where the content was small
- enough to be read with the body, causing Privoxy to wait for the
- end of the content until the server closed the connection or the
- request timed out. Reported by "Karsten" in #3028326.
-- Responses with status code 204 weren't properly detected as body-less
- like RFC2616 mandates. Like the previous bug, this caused Privoxy to
- wait for the end of the content until the server closed the connection
- or the request timed out. Fixes #3022042 and #3025553, reported by a
- user with no visible name. Most likely also fixes a bunch of other
- AJAX-related problem reports that got closed in the past due to
- insufficient information and lack of feedback.
-- Fixed an ACL bug that made it impossible to build a blacklist.
- Usually the ACL directives are used in a whitelist, which worked
- as expected, but blacklisting is still useful for public proxies
- where one only needs to deny known abusers access.
-- Added LOG_LEVEL_RECEIVED to log the not-yet-parsed data read from the
- network. This should make debugging various parsing issues a lot easier.
-- The IPv6 code is enabled by default on Windows versions that support it.
- Patch submitted by oCameLo in #2942729.
-- In mingw32 versions, the user.filter file is reachable through the
- GUI, just like default.filter is. Feature request 3040263.
-- Added the configure option --enable-large-file-support to set a few
- defines that are required by platforms like GNU/Linux to support files
- larger then 2GB. Mainly interesting for users without proper logfile
- management.
-- Logging with "debug 16" no longer stops at the first nul byte which is
- pretty useless. Non-printable characters are replaced with their hex value
- so the result can't span multiple lines making parsing them harder then
- necessary.
-- Privoxy logs when reading an action, filter or trust file.
-- Fixed incorrect regression test markup which caused a test in
- 3.0.16 to fail while Privoxy itself was working correctly.
- While Privoxy accepts hide-referer, too, the action name is actually
- hide-referrer which is also the name used one the final results page,
- where the test expected the alias.
-
-- CGI interface improvements:
- - In finish_http_response(), continue to add the 'Connection: close'
- header if the client connection will not be kept alive.
- Anonymously pointed out in #2987454.
- - Apostrophes in block messages no longer cause parse errors
- when the blocked page is viewed with JavaScript enabled.
- Reported by dg1727 in #3062296.
- - Fix a bunch of anchors that used underscores instead of dashes.
- - Allow to keep the client connection alive after crunching the previous request.
- Already opened server connections can be kept alive, too.
- - In cgi_show_url_info(), don't forget to prefix URLs that only contain
- http:// or https:// in the path. Fixes #2975765 reported by Adam Piggott.
- - Show the 404 CGI page if cgi_send_user_manual() is called while
- local user manual delivery is disabled.
+- General improvements:
+ - Add a receive-buffer-size directive which can be used to
+ set the size of the previously statically allocated buffer
+ in handle_established_connection().
+ Increasing the buffer size increases Privoxy's memory usage but
+ can lower the number of context switches and thereby reduce the
+ CPU usage and potentially increase the throughput.
+ This is mostly relevant for fast network connections and
+ large downloads that don't require filtering.
+ Sponsored by: Robert Klemme
+ - Add a listen-backlog directive which specifies the backlog
+ value passed to listen().
+ Sponsored by: Robert Klemme
+ - Add an enable-accept-filter directive which allows to
+ toggle accept filter support at run time when compiled
+ with FEATURE_ACCEPT_FILTER support.
+ It makes testing more convenient and now that it's
+ optional we can emit an error message if enabling
+ the accept filter fails.
+ Sponsored by: Robert Klemme
+ - Add a delay-response{} action.
+ This is useful to tar pit JavaScript requests that
+ are endlessly retried in case of blocks. It can also
+ be used to simulate a slow Internet connection.
+ Sponsored by: Robert Klemme
+ - Add a 'trusted-cgi-referrer' directive.
+ It allows to configure another page or site that can be used
+ to reach sensitive CGI resources.
+ Sponsored by: Robert Klemme
+ - Add a --fuzz mode which exposes Privoxy internals to input
+ from files or stdout.
+ Mainly tested with American Fuzzy Lop. For details see:
+ https://www.fabiankeil.de/talks/fuzzing-on-freebsd/
+ This work was partially funded with donations and done
+ as part of the Privoxy month in 2015.
+ - Consistently use the U(ngreedy) flag in the 'img-reorder' filter.
+ - listen_loop(): Reuse a single thread attribute object
+ The object doesn't change and creating a new one for
+ every thread is a waste of (CPU) time.
+ Sponsored by: Robert Klemme
+ - Free csp resources in the thread that belongs to the csp instead
+ of the main thread which has enough on its plate already.
+ Sponsored by: Robert Klemme
+ - Improve 'socket timeout reached' message.
+ Log the timeout that was triggered and downgrade the
+ log level to LOG_LEVEL_CONNECT to reduce the log noise
+ with common debug settings.
+ The timeout isn't necessary the result of an error and
+ usually merely indicates that Privoxy's socket timeout
+ is lower than the relevant timeouts used by client and
+ server.
+ Sponsored by: Robert Klemme
+ - Explicitly taint the server socket in case of CONNECT requests.
+ This doesn't fix any known problems, but makes
+ some log messages less confusing.
+ - Let write_pid_file() terminate if the pid file can't be opened.
+ Logging the issue at info level is unlikely to help.
+ - log_error(): Reduce the mutex-protected area by not using a
+ heap-allocated buffer that is shared between all threads.
+ This increases performance and reduces the latency with
+ verbose debug settings and multiple concurrent connections.
+ Sponsored by: Robert Klemme
+ - Let zalloc() use calloc() if it's available.
+ In some situations using calloc() can be faster than
+ malloc() + memset() and it should never be slower.
+ In the real world the impact of this change is not
+ expected to be noticeable.
+ Sponsored by: Robert Klemme
+ - Never use select() when poll() is available.
+ On most platforms select() is limited by FD_SETSIZE while
+ poll() is not. This was a scaling issue for multi-user setups.
+ Using poll() has no downside other than the usual risk
+ that code modifications may introduce new bugs that have
+ yet to be found and fixed.
+ At least in theory this commit could also reduce the latency
+ when there are lots of connections and select() would use
+ "bit fields in arrays of integers" to store file descriptors.
+ Another side effect is that Privoxy no longer has to stop
+ monitoring the client sockets when pipelined requests are
+ waiting but can't be read yet.
+ This code keeps the select()-based code behind ifdefs for
+ now but hopefully it can be removed soonish to make the
+ code more readable.
+ Sponsored by: Robert Klemme
+ - Add a 'reproducible-tarball-dist' target.
+ It's currently separate from the "tarball-dist" target
+ because it requires a tar implementation with mtree spec
+ support.
+ It's far from being perfect and does not enforce a
+ reproducible mode, but it's better than nothing.
+ - Use arc4random() if it's available.
+ While Privoxy doesn't need high quality pseudo-random numbers
+ there's no reason not to use them when we can and this silences
+ a warning emitted by code checkers that can't tell whether or not
+ the quality matters.
+ - Show the FEATURE_EXTERNAL_FILTERS status on the status page.
+ Better late than never. Previously a couple of tests weren't
+ executed as Privoxy-Regression-Test couldn't detect that the
+ FEATURE_EXTERNAL_FILTERS dependency was satisfied.
+ - Ditch FEATURE_IMAGE_DETECT_MSIE.
+ It's an obsolete workaround we inherited from Junkbuster
+ and was already disabled by default.
+ Users that feel the urge to work around issues with
+ image requests coming from an Internet Explorer version
+ from more than 15 years ago can still do this using tags.
+ - Consistently use strdup_or_die() instead of strdup() in
+ cases where allocation failures aren't expected.
+ Using strdup_or_die() allows to remove a couple of explicit
+ error checks which slightly reduces the size of the binary.
+ - Insert a refresh tag into the /client-tags CGI page when
+ serving it while a client-specific tag is temporarily enabled.
+ This makes it less likely that the user ends up
+ looking at tag state that is out of date.
+ - Use absolute URLs in the client-tag forms.
+ It's more consistent with the rest of the CGI page
+ URLs and makes it more convenient to copy the forms
+ to external pages.
+ - cgi_error_disabled(): Use status code 403 and an appropriate response line
+ - Use a dedicated CGI handler to deal with tag-toggle requests
+ As a result the /client-tags page is now safe to reach without
+ trusted Referer header which makes bookmarking or linking to
+ it more convenient.
+ Finally, refreshing the /client-tags page to show the
+ current state can no longer unintentionally repeat the
+ previous toggle request.
+ - Don't add a "Connection" header for CONNECT requests.
+ Explicitly sending "Connection: close" is not necessary and
+ apparently it causes problems with some forwarding proxies
+ that will close the connection prematurely.
+ Reported by Marc Thomas.
+ - Fix compiler warnings.
+
+- Bug fixes:
+ - rfc2553_connect_to(): Properly detect and log when poll()
+ reached the time out. Previously this was logged as:
+ Could not connect to [...]: No error: 0.
+ which isn't very helpful.
+ Sponsored by: Robert Klemme
+ - add_tag_for_client(): Set time_to_live properly.
+ Previously the time_to_live was always set for the first tag.
+ Attempts to temporarily enable a tag would result in enabling
+ it permanently unless no tag was enabled already.
+ - Revert r1.165 which didn't perform as advertised.
+ While the idea was to use "https:// when creating links
+ for the user manual on the website", the actual effect
+ was to use "https://" when Privoxy was supposed to serve
+ the user manual itself.
+ Reported by Yossi Zahn on Privoxy-devel@.
+ - socks5_connect(): Fail in case of unsupported address types.
+ Previously they would not be detected right away and
+ Privoxy would fail later on with an error message that
+ didn't make it obvious that the problem was socks-related.
+ So far, no such problems have actually been reported.
+ - socks5_connect(): Properly deal with socks replies that
+ contain IPv6 addresses.
+ Previously parts of the reply were left unread and
+ later on treated as invalid HTTP response data.
+ Fixes #904 reported by Danny Goossen who also provided
+ the initial version of this patch.