]> granicus.if.org Git - libevent/log
libevent
5 years agohttp: cleanup of the request-line parsing
Azat Khuzhin [Mon, 22 Oct 2018 20:25:01 +0000 (23:25 +0300)]
http: cleanup of the request-line parsing

6 years agoDisable parallel jobs for the osx (due to CPU time deficit) in travis-ci
Azat Khuzhin [Sun, 21 Oct 2018 22:06:48 +0000 (01:06 +0300)]
Disable parallel jobs for the osx (due to CPU time deficit) in travis-ci

As you can see right now linux workers has zero failed tests, while osx
workers has 18 failed tests:
  [bufferevent_connect_hostname_emfile FAILED]
  [bufferevent_pair_release_lock FAILED]
  [bufferevent_timeout FAILED]
  [bufferevent_timeout_filter FAILED]
  [bufferevent_timeout_pair FAILED]
  [common_timeout FAILED]
  [del_wait FAILED]
  [immediatesignal FAILED]
  [loopexit FAILED]
  [loopexit_multiple FAILED]
  [monotonic_res FAILED]
  [no_events FAILED]
  [persistent_active_timeout FAILED]
  [persistent_timeout_jump FAILED]
  [signal_switchbase FAILED]
  [signal_while_processing FAILED]
  [simpletimeout FAILED]
  [usleep FAILED]

And this patch should remove from this list time related failures
(though maybe not all of them).

6 years agoUse matrix.fast_finish for the travis-ci to mark the build ASAP
Azat Khuzhin [Sun, 21 Oct 2018 22:00:54 +0000 (01:00 +0300)]
Use matrix.fast_finish for the travis-ci to mark the build ASAP

AFAIR there is shortage of osx workers on travis-ci, IOW builds that
requires them can wait fair amount of time in the queue by just waiting.

Plus linux workers AFAICS can run multiple jobs in parallel (4-5), while
osx does not.

Hence if we do allow failures for osx (and right now they have a lot of
failed tests) let's mark build result based on tests under linux only.

So in a nut shell this will reduce build time from 5 hours to 20-30
minutes.

6 years agoSimplify bufferevent timeout tests to reduce CPU usage in between start/compare
Azat Khuzhin [Sun, 21 Oct 2018 15:31:01 +0000 (18:31 +0300)]
Simplify bufferevent timeout tests to reduce CPU usage in between start/compare

Between start (setting "started_at") and comparing the time when
timeouts triggered with the start (test_timeval_diff_eq), there is too
much various things that can introduce extra delays and eventually could
fail the test on machine with shortage of CPU.

And this is exactly what happend on:
- travis-ci
- #262

Here is a simple reproducer that I came up with for this issue:
  docker run --cpus=0.01 -e LD_LIBRARY_PATH=$PWD/lib -e PATH=/usr/bin:/bin:$PWD/bin -v $PWD:$PWD --rm -it debian:testing regress --no-fork --verbose bufferevent/bufferevent_timeout

Under limited CPU (see reproducer) the test almost always has problems
with that "write_timeout_at" exceed default timeval diff tolerance
(test_timeval_diff_eq() has 50 tolerance), i.e.:
  FAIL ../test/regress_bufferevent.c:1040: assert(labs(timeval_msec_diff(((&started_at)), ((&res1.write_timeout_at))) - (100)) <= 50): 101 vs 50

But under some setup write timeout can even not triggered, and the
reason for this is that we write to the bufferevent 1024*1024 bytes, and
hence if evbuffer_write_iovec() will has some delay after writev() and
not send more then one vector at a time [1], it is pretty simple to
trigger, i.e.:
  FAIL ../test/regress_bufferevent.c:1040: assert(labs(timeval_msec_diff(((&started_at)), ((&res1.write_timeout_at))) - (100)) <= 50): 1540155888478 vs 50

  [1]: https://gist.github.com/azat/b72773dfe7549fed865d439e03de05c1

So this patch just send static small payload for all cases (plus a few
more asserts added).

The outcome of this patch is that all regression tests passed on
travis-ci for linux box [2]. While before it fails almost always [3].
Also reproducer with CPU limiting via docker also survive some
iterations (and strictly speaking it should has less CPU then travis-ci
workers I guess).

  [2]: https://travis-ci.org/azat/libevent/builds/444391481
  [3]: https://travis-ci.org/libevent/libevent/builds/444336505

6 years agoevdns: handle NULL filename explicitly
Bogdan Harjoc [Thu, 9 Aug 2018 10:04:52 +0000 (13:04 +0300)]
evdns: handle NULL filename explicitly

Otherwise we will try to open NULL filename and got EFAULT and anyway
will got the same return code from evdns_base_resolv_conf_parse_impl()

Closes: #680 (cherry-picked)
6 years agoRevert "travis-ci: use travis_retry to reduce number of problems with flacky tests"
Azat Khuzhin [Sun, 21 Oct 2018 00:18:48 +0000 (03:18 +0300)]
Revert "travis-ci: use travis_retry to reduce number of problems with flacky tests"

That was a bad idea, because all this patch did is just uses extra CPU
time on travis-ci workers and eventually fail the build anyway.

It also increases time of executing one matrix entry (x3).

This reverts commit 7004ee8e869faabe3a284fb3d7eb859fb50ef90c.

6 years agoMerge branch 'regress-dns-fixes'
Azat Khuzhin [Sun, 21 Oct 2018 00:15:34 +0000 (03:15 +0300)]
Merge branch 'regress-dns-fixes'

* regress-dns-fixes:
  Do not rely on getservbyname() for most of the dns regression tests
  Turn off dns/getaddrinfo_race_gotresolve by default
  Fix an error for debug locking in dns/getaddrinfo_race_gotresolve

6 years agoDo not rely on getservbyname() for most of the dns regression tests
Azat Khuzhin [Sun, 21 Oct 2018 00:03:25 +0000 (03:03 +0300)]
Do not rely on getservbyname() for most of the dns regression tests

There is only one test that uses service name getaddrinfo_async, which
manually check whether it works or not, other should not assume that it
is available and works.

There was already an attempt to overcome some possible limitations, like
lack of "http" in /etc/services in
d6bafbbeb27ff3943d6f3b6783bcded76384c31e ("test/dns: replace servname
since solaris does not have "http"")

6 years agoTurn off dns/getaddrinfo_race_gotresolve by default
Azat Khuzhin [Sun, 21 Oct 2018 00:03:57 +0000 (03:03 +0300)]
Turn off dns/getaddrinfo_race_gotresolve by default

It is:
- pretty internal regression
- CPU bound
- right now failed on travis-ci machines

6 years agoFix an error for debug locking in dns/getaddrinfo_race_gotresolve
Azat Khuzhin [Sat, 20 Oct 2018 23:50:04 +0000 (02:50 +0300)]
Fix an error for debug locking in dns/getaddrinfo_race_gotresolve

When there is no /etc/services file evdns_getaddrinfo() will fail (with
service="ssh") and hence it will go to then "end" label with locked
rp.lock which in case of debug locking checks will bail with:
  [err] ../evthread.c:220: Assertion lock->count == 0 failed in debug_lock_free

So add rp.locked flag, and unlock the lock before freeing it if it is in
locked state.

And here is how you can reproduce the issue:
  $ docker run -e LD_LIBRARY_PATH=$PWD/lib -e PATH=/usr/bin:/bin:$PWD/bin -v $PWD:$PWD --rm -it debian:testing regress dns/getaddrinfo_race_gotresolve
(since debian:testing does not have /etc/services)

6 years agoFix missing LIST_HEAD
Jiri Luznicky [Wed, 23 May 2018 13:39:13 +0000 (15:39 +0200)]
Fix missing LIST_HEAD

Despite the presence of 'sys/queue.h' in some stdlib implementations
(i.e. uclibc) 'LIST_HEAD' macro can be missing.  This fix defines this
macro in the same manner as was done previously for 'TAILQ_'.

Fixes: #539
Closes: #639 (cherry-picked)
Backport: 2.1.9

6 years agoMake cmake's check_const_exists() rely on cache
Azat Khuzhin [Thu, 18 Oct 2018 20:16:05 +0000 (23:16 +0300)]
Make cmake's check_const_exists() rely on cache

6 years agoMerge branch 'be-wm-overrun-v2'
Azat Khuzhin [Wed, 17 Oct 2018 20:21:32 +0000 (23:21 +0300)]
Merge branch 'be-wm-overrun-v2'

* be-wm-overrun-v2:
  Fix hangs due to watermarks overruns in bufferevents implementations
  test: cover watermarks (with some corner cases) in ssl bufferevent

Fixes: #690
6 years agoFix hangs due to watermarks overruns in bufferevents implementations
Azat Khuzhin [Wed, 17 Oct 2018 20:21:17 +0000 (23:21 +0300)]
Fix hangs due to watermarks overruns in bufferevents implementations

Some implementations of bufferevents (for example openssl) can overrun
read high watermark.
And after this if user callback will not drain enough data it will be
suspended (i.e. it will not be runned again anymore).
This is not the expecting behaviour as one may guess, since in this case
the data will never be read. Hence once we detected that the watermark
exceeded (even after calling user callback) we will schedule the
callback again.

This also can be fixed in bufferevent openssl implementation (by
strictly limiting how much data is added to the read buffer according to
read high watermark), but since this data is already available (and in
memory) there is no point in doing so.

6 years agotest: cover watermarks (with some corner cases) in ssl bufferevent
Azat Khuzhin [Wed, 3 Oct 2018 22:03:10 +0000 (01:03 +0300)]
test: cover watermarks (with some corner cases) in ssl bufferevent

6 years agoAdd -fdiagnostics-color=always for Ninja
Azat Khuzhin [Wed, 17 Oct 2018 08:14:48 +0000 (11:14 +0300)]
Add -fdiagnostics-color=always for Ninja

6 years agoevrpc: avoid NULL dereference on request is not EVHTTP_REQ_POST
Azat Khuzhin [Thu, 13 Sep 2018 23:24:44 +0000 (02:24 +0300)]
evrpc: avoid NULL dereference on request is not EVHTTP_REQ_POST

Fixes: #660
6 years agotest: make sure pthread is defined
Carlo Marcelo Arenas Belón [Wed, 15 Aug 2018 20:57:01 +0000 (13:57 -0700)]
test: make sure pthread is defined

avoid warnings with any modern C99 compiler due to implicit function
declaration for pthread_create, as shown by the following :

test/regress_dns.c:2226:2: warning: implicit declaration of function
      'pthread_create' is invalid in C99 [-Wimplicit-function-declaration]
        THREAD_START(thread[0], race_base_run, &rp);
        ^
test/regress_thread.h:35:2: note: expanded from macro 'THREAD_START'
        pthread_create(&(threadvar), NULL, fn, arg)
        ^
test/regress_dns.c:2226:2: warning: this function declaration is not a prototype
      [-Wstrict-prototypes]
test/regress_thread.h:35:2: note: expanded from macro 'THREAD_START'
        pthread_create(&(threadvar), NULL, fn, arg)
        ^

$ clang --version
Apple LLVM version 9.1.0 (clang-902.0.39.2)
Target: x86_64-apple-darwin17.7.0
Thread model: posix
InstalledDir: /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin

Closes: #686 (cherry-picked)
6 years agoappveyor: add all build tree into artifacts in case of failure
Azat Khuzhin [Sat, 18 Aug 2018 20:50:10 +0000 (23:50 +0300)]
appveyor: add all build tree into artifacts in case of failure

This will help in investigations of win32 failures...

P.S. looks like there is no better way, since
"artifacts.*.publish_on_failure" hadn't been implemented yet (while was
requiested in 2015).

6 years agodns-example: free result in getaddrinfo callback
Bogdan Harjoc [Thu, 9 Aug 2018 11:47:17 +0000 (14:47 +0300)]
dns-example: free result in getaddrinfo callback

According to evdns.c, the result not freed by libevent after
the callback runs:

evdns_getaddrinfo_gotresolve()
{
    ...
    data->user_cb(0, data->pending_result, data->user_data);
    data->pending_result = NULL;
    ...
}

To reproduce, build with -fsanitize=address, add -g to the getopt
list in dns-example.c like in the current commit and run

  dns-example -g google.com

Closes: #681 # cherry-picked
6 years agoEliminate compiler warnings (at least for gcc/linux)
Philip Prindeville [Sat, 16 Jun 2018 23:16:10 +0000 (17:16 -0600)]
Eliminate compiler warnings (at least for gcc/linux)

Signed-off-by: Philip Prindeville <philipp@redfish-solutions.com>
Closes: #646 # cherry-picked from the PR
6 years agoMerge branch 'evdns_getaddrinfo-race-fix'
Azat Khuzhin [Thu, 2 Aug 2018 06:29:29 +0000 (09:29 +0300)]
Merge branch 'evdns_getaddrinfo-race-fix'

* evdns_getaddrinfo-race-fix:
  evdns: add regress test for getaddrinfo race
  evdns: fix race condition in evdns_getaddrinfo()

Merges: #673

6 years agoevdns: add regress test for getaddrinfo race
Sergey Fionov [Thu, 2 Aug 2018 06:13:33 +0000 (09:13 +0300)]
evdns: add regress test for getaddrinfo race

6 years agoevdns: fix race condition in evdns_getaddrinfo()
Sergey Fionov [Wed, 1 Aug 2018 21:35:28 +0000 (00:35 +0300)]
evdns: fix race condition in evdns_getaddrinfo()

evdns_getaddrinfo() starts two parallel requests for A and AAAA record.
But if request is created from thread different from dns_base's, request of A record is
started immediately and may result in calling free_getaddrinfo_request() from
evdns_getaddrinfo_gotresolve() because `other_req' doesn't exist yet.

After that, request of AAAA record starts and finishes, and evdns_getaddrinfo_gotresolve()
is called again for structure that is already freed.

This commits adds locking into evdns_getaddrinfo() function.

6 years agoCleanup __func__ detection
Azat Khuzhin [Tue, 19 Jun 2018 07:15:08 +0000 (10:15 +0300)]
Cleanup __func__ detection

First of all __func__ is not a macro, it is char[] array, so the code
that we had before in cmake, was incorrect, i.e.:
  #if defined (__func__)
  #define EVENT____func__ __func__
  #elif defined(__FUNCTION__)
  #define EVENT____func__  __FUNCTION__
  #else
  #define EVENT____func__ __FILE__
  #endif

So just detect do we have __func__/__FUNCTION__ in configure/cmake
before build and define EVENT__HAVE___func__/EVENT__HAVE___FUNCTION__
to use the later to choose which should be used as a __func__ (if it is
not presented).

Closes: #644
6 years agoMerge branch 'official/pr/671' -- README cleanup
Azat Khuzhin [Wed, 1 Aug 2018 06:48:42 +0000 (09:48 +0300)]
Merge branch 'official/pr/671' -- README cleanup

* official/pr/671:
  Capitalise project names consistently in README.md
  Indent configure flag section to make markdown format them as code
  Use https for resources that support it
  Rewords awkward sentences in README.md
  Fix typos in README.md

6 years agoautotools: include win32 specific headers for socklen_t detection on win32/mingw
Azat Khuzhin [Tue, 31 Jul 2018 21:58:02 +0000 (00:58 +0300)]
autotools: include win32 specific headers for socklen_t detection on win32/mingw

The [1] removes EVENT__ prefix, and now if we will incorrectly detect
that "foobar" (or socklen_t in our case) type is not available, but
somewhere later it will be available then we will get next error [2]:
  error: two or more data types in declaration specifiers

According to [3]:
- Compile something in Cygwin and you are compiling it for Cygwin.
- Compile something in MinGW and you are compiling it for Windows.
  And I can confirm this, since there is _WIN32 defined (according to [4])

And since according to [5] our image in appveyour (Visual Studion 2015)
has mingw (and we use it, not cygwin) we need ws2tcpip.h (over
sys/socket.h -- which does not exist in win32) header to detect
socklen_t existence.

[1]: 587e9f5828c4eb3c6e79e9eb29f174e4ae0c05bd ("config.h can't be prefixed unconditionally")
[2]: https://ci.appveyor.com/project/nmathewson/libevent/build/job/yonukoc5q3tr3e5e#L372
[3]: https://stackoverflow.com/questions/771756/what-is-the-difference-between-cygwin-and-mingw
[4]: https://www.appveyor.com/docs/build-environment/#mingw-msys-cygwin
[5]: https://ci.appveyor.com/project/nmathewson/libevent/build/job/yonukoc5q3tr3e5e#L164

Fixes: #649
6 years agoCapitalise project names consistently in README.md
Aapeli [Wed, 1 Aug 2018 00:29:43 +0000 (10:29 +1000)]
Capitalise project names consistently in README.md

6 years agoIndent configure flag section to make markdown format them as code
Aapeli [Wed, 1 Aug 2018 00:24:50 +0000 (10:24 +1000)]
Indent configure flag section to make markdown format them as code

6 years agoUse https for resources that support it
Aapeli [Wed, 1 Aug 2018 00:21:57 +0000 (10:21 +1000)]
Use https for resources that support it

6 years agoRewords awkward sentences in README.md
Aapeli [Wed, 1 Aug 2018 00:20:28 +0000 (10:20 +1000)]
Rewords awkward sentences in README.md

6 years agoFix typos in README.md
Aapeli [Wed, 1 Aug 2018 00:17:42 +0000 (10:17 +1000)]
Fix typos in README.md

6 years agoappveyor: disable almost all configurations (except autotools/cmake basic)
Azat Khuzhin [Wed, 1 Aug 2018 00:01:43 +0000 (03:01 +0300)]
appveyor: disable almost all configurations (except autotools/cmake basic)

Since we have FREE plan, we have a lot of limits on our builds, so let's
reduce their amount to make them run/start faster (I hope).

6 years agovagrant/win32: reduce amount of threads
Azat Khuzhin [Tue, 31 Jul 2018 21:44:58 +0000 (00:44 +0300)]
vagrant/win32: reduce amount of threads

Otherwise I got:
    win:     855 [main] sh 840 fork: child -1 - CreateProcessW failed for 'C:\tools\cygwin\bin\sh.exe', errno 5
    win: /bin/sh: /usr/bin/sed: Input/output error
    win: /bin/sh: line 1: /bin/sh: Input/output error
    win:   CC       bufferevent_pair.lo
    win: /bin/sh: fork: Input/output error
    win: make[1]: *** [Makefile:1722: buffer.lo] Error 126
    win: make[1]: *** Waiting for unfinished jobs....
    win:   CC       bufferevent_ratelim.lo
    win:     176 [main] sh 2440 fork: child -1 - CreateProcessW failed for 'C:\tools\cygwin\bin\sh.exe', errno 5
    win: /bin/sh: fork: Input/output error
    win:       0 [main] sh 3684 fork: child -1 - CreateProcessW failed for 'C:\tools\cygwin\bin\sh.exe', errno 5
    win: ./libtool: fork: Input/output error
    win: make[1]: *** [Makefile:1722: bufferevent_filter.lo] Error 254
    win:       0 [main] sh 1356 fork: child -1 - CreateProcessW failed for 'C:\tools\cygwin\bin\sh.exe', errno 5
    win: ./libtool: fork: Input/output error
    win:       0 [main] sh 3640 fork: child -1 - CreateProcessW failed for 'C:\tools\cygwin\bin\sh.exe', errno 5
    win: ./libtool: line 922: /usr/bin/sed: Input/output error
    win: ./libtool: fork: Input/output error
    win:       0 [main] sh 3444 fork: child -1 - CreateProcessW failed for 'C:\tools\cygwin\bin\sh.exe', errno 5
    win:       0 [main] sh 3796 fork: child -1 - CreateProcessW failed for 'C:\tools\cygwin\bin\sh.exe', errno 5
    win: ./libtool: fork: Input/output error
    win:     693 [main] sh 304 fork: child -1 - CreateProcessW failed for 'C:\tools\cygwin\bin\sh.exe', errno 5
    win: ./libtool: fork: Input/output error
    win: ./libtool: fork: Input/output error
    win: ./libtool: line 1377: /usr/bin/sed: Input/output error
    win: :   error: Failed to create '.libs'

6 years agovagrant/win32: add missing aclocal
Azat Khuzhin [Tue, 31 Jul 2018 21:28:48 +0000 (00:28 +0300)]
vagrant/win32: add missing aclocal

6 years agoFix typo
Xiaozhou Liu [Thu, 5 Jul 2018 03:25:24 +0000 (11:25 +0800)]
Fix typo

Closes: #658
6 years ago[http] fix C90 warnings
Nathan French [Thu, 5 Jul 2018 16:46:51 +0000 (12:46 -0400)]
[http] fix C90 warnings

6 years agoTell Travis to use quiet git clone
Philip Prindeville [Thu, 21 Jun 2018 05:47:51 +0000 (23:47 -0600)]
Tell Travis to use quiet git clone

Otherwise the logs show a bunch of pointless progress, which is
really only useful for interactive sessions.

Signed-off-by: Philip Prindeville <philipp@redfish-solutions.com>
Closes: #651 (cherry-picked)
6 years agotest: make regress_dns C89 compatible
Leo Zhang [Tue, 15 May 2018 04:14:05 +0000 (12:14 +0800)]
test: make regress_dns C89 compatible

Closes: #635 (cherry-picked)
6 years agoIgnore evconfig-private.h for autotools
Azat Khuzhin [Wed, 20 Jun 2018 23:37:24 +0000 (02:37 +0300)]
Ignore evconfig-private.h for autotools

Fixes: e2874d95af584b3657eb5dc1d2e5404811dce117 ("Adopt ignore rules for
cmake + ninja")

6 years agoReplace all EVENT_HAVE with EVENT__HAVE (EVENT__HAVE_USLEEP)
Azat Khuzhin [Tue, 19 Jun 2018 07:18:46 +0000 (10:18 +0300)]
Replace all EVENT_HAVE with EVENT__HAVE (EVENT__HAVE_USLEEP)

6 years agoAdd convenience macros for user-triggered events
Philip Prindeville [Sat, 16 Jun 2018 23:39:34 +0000 (17:39 -0600)]
Add convenience macros for user-triggered events

Signed-off-by: Philip Prindeville <philipp@redfish-solutions.com>
Closes: #647 (picked)
6 years agoGet rid of macros which are never used
Philip Prindeville [Sat, 16 Jun 2018 22:44:06 +0000 (16:44 -0600)]
Get rid of macros which are never used

There are no expansions of these macros or tests for their existence.

Signed-off-by: Philip Prindeville <philipp@redfish-solutions.com>
6 years agofix spelling mistakes
an-tao [Fri, 25 May 2018 10:07:28 +0000 (18:07 +0800)]
fix spelling mistakes

Closes: nmathewson/Libevent#162
6 years agoconfig.h can't be prefixed unconditionally
Philip Prindeville [Wed, 18 Apr 2018 21:09:35 +0000 (15:09 -0600)]
config.h can't be prefixed unconditionally

The script make-event-config.sed was mangling all the symbols by
prefixing them with "EVENT__".  The problem here is that some
symbols aren't for local consumption within libevent, but rather
influence other system header files (ex: __USE_FILE_OFFSET64 is
used by dozens of header files including <sys/sendfile.h>).

As a workaround, all symbols starting with a capital letter only
(with the exception of STDC_HEADERS which must also be left
untouched) will be mangled.

Future contributors will need to be aware of this distinction.

6 years agoMake rpc headers self-compilable
Azat Khuzhin [Sun, 6 May 2018 23:39:44 +0000 (02:39 +0300)]
Make rpc headers self-compilable

Fixes: #633
6 years agoMerge pull request #632 from NathanFrench/realign_struct_event
Nathan French [Wed, 2 May 2018 23:34:04 +0000 (19:34 -0400)]
Merge pull request #632 from NathanFrench/realign_struct_event

Reorder fields in struct event for memory efficiency

6 years ago[core] re-order fields in struct event for memory efficiency
Nathan French [Mon, 30 Apr 2018 22:13:45 +0000 (18:13 -0400)]
[core] re-order fields in struct event for memory efficiency

The sizeof `struct event` can reduced on both 32 bit and 64 bit systems
by moving the 4 bytes that make up `ev_events` and `ev_res` below `ev_fd`,
before `struct event_base * ev_base;` since our compiler wouldn't dare do
such a thing (it instead will pad twice, whereas it only needs to be padded
once)

```C
struct event {
                                                         /*  OFFS |  SZ   Bytes  | Total Bytes | START - END  */
    struct event_callback ev_evcallback;                 /*   0x0 |  40          | 40          | 0x0   - 0x28 */

    union {                                              /*  0x28 |  ----------- | ----------- | ------------ */
        TAILQ_ENTRY(event) ev_next_with_common_timeout;  /*       |       ((16)) |             |              */
        int min_heap_idx;                                /*       |       ((04)) |             |              */
    }   ev_timeout_pos;                                  /*       |  16          | 56          | 0x28  - 0x38 */
    int ev_fd;                                           /*  0x38 |  04          | 60          | 0x38  - 0x3c */
```

Since the next field is 8 bytes in length, and we are up to 60 bytes, `ev_fd` ends up being padded (4 more bytes on 64b).

```C
                              /* --- 1 byte gap HERE --->            1           | <61>                      */
                              /* --- 1 byte gap HERE --->            1           | <62>                      */
                              /* --- 1 byte gap HERE --->            1           | <63>                      */
                              /* --- 1 byte gap HERE --->            1           | <64>                      */
    struct event_base * ev_base;                         /*  0x3c |  8           | 68         | 0x3c - 0x40  */
    union {                                              /*  0x40 | ------------ | ---------- | ------------ */
        struct {                                         /*       | ------------ |            |              */
            LIST_ENTRY (event) ev_io_next;               /*       |       ((16+  |            |              */
            struct timeval ev_timeout;                   /*       |         16)) |            |              */
        } ev_io;                                         /*       | ((32))       |            |              */
        struct {                                         /*       | ------------ |            |              */
            LIST_ENTRY (event) ev_signal_next;           /*       |      ((16+   |            |              */
            short   ev_ncalls;                           /*       |        02+   |            |              */
            short * ev_pncalls;                          /*       |        08))  |            |              */
        } ev_signal;                                     /*       | ((26))       |            |              */
    } ev_;                                               /*  0x60 | 32           | 100        | 0x40 - 0x60  */

    short ev_events;                                     /*  0x60 | 2            | 102        | 0x60 - 0x62  */
    short ev_res;                                        /*  0x62 | 2            | 104        | 0x62 - 0x64  */
```

We now hit another line, `struct timeval` is 16 bytes on 64b arch, so we have 4 more bytes
of padding on `ev_res`.

```C
    /* --- 1 byte gap HERE --- */
    /* --- 1 byte gap HERE --- */
    /* --- 1 byte gap HERE --- */
    /* --- 1 byte gap HERE --- */
    struct timeval ev_timeout;                           /*  0x64 | 16           | 120        | 0x64 - 0x74  */
};
```

After moving `ev_events` and `ev_res` below `ev_fd` we have something
a bit more optimal:

```C
struct event2 {
                                                         /*  OFFS |  SZ / Bytes  | RSUM Bytes | START - END  */
    struct event_callback ev_evcallback;                 /*   0x0 | 40           | 40         | 0x0   - 0x28 */
    union {                                              /*  0x28 | ------------ | ---------- | ------------ */
        TAILQ_ENTRY(event) ev_next_with_common_timeout;  /*       |       ((16)) |            |              */
        int min_heap_idx;                                /*       |       ((04)) |            |              */
    } ev_timeout_pos;                                    /*       | 16           | 56         | 0x28  - 0x38 */
    int ev_fd;                                           /*  0x38 | 4            | 60         | 0x38  - 0x3c */
    short ev_events;                                     /*  0x3c | 2            | 62         | 0x3c  - 0x3e */
    short ev_res;                                        /*  0x3e | 2            | 64         | 0x3e  - 0x40 */

    struct event_base * ev_base;                         /*  0x40 | 8            | 74         | 0x40  - 0x48 */
    union {                                              /*  0x48 | ------------ | ---------- | ------------ */
        struct {                                         /*       | ------------ |            |              */
            LIST_ENTRY (event) ev_io_next;               /*       |        ((16+ |            |              */
            struct timeval ev_timeout;                   /*       |         16)) |            |              */
        } ev_io;                                         /*       | ((32))       |            |              */
        struct {                                         /*       | ------------ |            |              */
            LIST_ENTRY (event) ev_signal_next;           /*       |       ((16+  |            |              */
            short   ev_ncalls;                           /*       |         02+  |            |              */
            short * ev_pncalls;                          /*       |         08)) |            |              */
        } ev_signal;                                     /*       | ((26))       |            |              */
    } ev_;                                               /*       | 32           | 106        | 0x48 - 0x68  */
    struct timeval ev_timeout;                           /* 0x68  | 16           | 120        | 0x68 - 0x78  */
};
```

We still have a gap here, but the first was removed.

Again, we can save 8 bytes on both 32 and 64 word sizes (32/64 byte cacheline).

Below are the results for testing v2.1.6 -> master -> master + this patch (Release/-O3)

Code:

```C
  #include <event2/event.h>

  int
  main(int argc, char ** argv) {
      printf("%zu\n", event_get_struct_event_size());
      return 0;
  }
```

Branch: `master` (2.2.x)

```
  $ gcc -O3 -Wall -Wl,-R/usr/local/lib bleh.c -L/usr/local/lib -o bleh  -levent
  $ ldd bleh
      linux-vdso.so.1 =>  (0x00007ffc3df50000)
      libevent.so.2.2.0 => /usr/local/lib/libevent.so.2.2.0 (0x00007f91fd781000)
      libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f91fd3a1000)
      libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f91fd182000)
      /lib64/ld-linux-x86-64.so.2 (0x00007f91fdbcc000)
  $ ./bleh
  128
```

Release: `2.1.6`

```
  $ gcc -O3 bleh.c -o bleh  -levent
  $ ldd bleh
      linux-vdso.so.1 =>  (0x00007ffd43773000)
      libevent-2.1.so.6 => /usr/lib/x86_64-linux-gnu/libevent-2.1.so.6 (0x00007feb3add6000)
      libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007feb3a9f6000)
      libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007feb3a7d7000)
      /lib64/ld-linux-x86-64.so.2 (0x00007feb3b22a000)
  $ ./bleh
  128
```

Branch: `this one`

```
  $ gcc -O3 -Wl,-R./lib bleh.c -o bleh -L./lib -levent
  $ ldd bleh
      linux-vdso.so.1 =>  (0x00007ffff55f7000)
      libevent.so.2.2.0 => ./lib/libevent.so.2.2.0 (0x00007ff8e5c82000)
      libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007ff8e58a2000)
      libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007ff8e5683000)
      /lib64/ld-linux-x86-64.so.2 (0x00007ff8e60cd000)
  $ ./bleh
  120
```

6 years agoNotify event base if there are no more events, so it can exit without delay
Azat Khuzhin [Mon, 23 Apr 2018 21:59:11 +0000 (00:59 +0300)]
Notify event base if there are no more events, so it can exit without delay

Fixes: #623
6 years agoFix CheckFunctionExistsEx() cmake macro on win32
Azat Khuzhin [Sun, 22 Apr 2018 22:50:55 +0000 (01:50 +0300)]
Fix CheckFunctionExistsEx() cmake macro on win32

For example under mingw64 it could not detect that strtok_r() exists,
because it checks with:
  void *p = func_name;

And for this you need the function to be defined, so just sync our
CheckFunctionExistsEx.c with CheckFunctionExists.c from cmake (and later
we should drop them out) since it does correct things to detech
functions existence.

Also for WIN32 there is -FIwinsock2.h -FIws2tcpip.h, and I guess that is
not works for mingw gcc (since -F in gcc is framework, and in windows
-FI is like -include in gcc). But looks like we do not need them
already (due to fixed CheckFunctionExistsEx()).

Refs: #605

6 years agotest/dns: install correct RLIMIT_NOFILE in bufferevent_connect_hostname_emfile
Azat Khuzhin [Tue, 24 Apr 2018 11:46:06 +0000 (14:46 +0300)]
test/dns: install correct RLIMIT_NOFILE in bufferevent_connect_hostname_emfile

Otherwise poll() will fail with EINVAL:
       EINVAL The nfds value exceeds the RLIMIT_NOFILE value.

P.S. and cleanup this test a little, with early-return.

CI: https://travis-ci.org/libevent/libevent/jobs/370350426

6 years agotravis-ci: allow tests under osx to fail (temporary, until we will fix them)
Azat Khuzhin [Mon, 23 Apr 2018 23:56:08 +0000 (02:56 +0300)]
travis-ci: allow tests under osx to fail (temporary, until we will fix them)

6 years agotest/http: cover evhttp_connection_get_addr() for incomming connections
Azat Khuzhin [Mon, 23 Apr 2018 23:22:58 +0000 (02:22 +0300)]
test/http: cover evhttp_connection_get_addr() for incomming connections

Refs: #510

6 years agoFix evhttp_connection_get_addr() fox incomming http connections
Greg Hazel [Mon, 12 Feb 2018 00:28:58 +0000 (16:28 -0800)]
Fix evhttp_connection_get_addr() fox incomming http connections

Install conn_address of the bufferevent on incomping http connections
(even though this is kind of subsytem violation, so let's fix it in a
simplest way and thinkg about long-term solution).

Fixes: #510
Closes: #595 (pick)
6 years agotest/dns: verify bufferevent_socket_connect() errorcb invoking if socket() fails
Azat Khuzhin [Mon, 23 Apr 2018 22:45:00 +0000 (01:45 +0300)]
test/dns: verify bufferevent_socket_connect() errorcb invoking if socket() fails

Refs: #600

6 years agobufferevent_socket_connect{,_hostname}() missing event callback and use ret code
Jesse Fang [Fri, 23 Feb 2018 11:15:12 +0000 (19:15 +0800)]
bufferevent_socket_connect{,_hostname}() missing event callback and use ret code

- When socket() failed in bufferevent_socket_connect() , the event
  callback should be called also in
  bufferevent_socket_connect_hostname().  eg. when use
  bufferevent_socket_connect_hostname() to resolve and connect an IP
  address but process have a smaller ulimit open files, socket() fails
  always but caller is not notified.

- Make bufferevent_socket_connect()'s behavior more consistent: function
  return error then no callback, function return ok then error passed by
  event callback.

Fixes: #597
Closes: #599
Closes: #600
6 years agotest/dns: cleanup test_bufferevent_connect_hostname()
Azat Khuzhin [Mon, 23 Apr 2018 22:34:06 +0000 (01:34 +0300)]
test/dns: cleanup test_bufferevent_connect_hostname()

6 years agoMerge branch 'pull-628'
Azat Khuzhin [Sun, 22 Apr 2018 21:52:57 +0000 (00:52 +0300)]
Merge branch 'pull-628'

* pull-628:
  Give priority to the build directory headers
  Do not ship evconfig-private.h in dist archive

Closes: #628
6 years agoGive priority to the build directory headers
Cristian Morales Vega [Thu, 19 Apr 2018 08:39:39 +0000 (09:39 +0100)]
Give priority to the build directory headers

Not changing anything right now AFAIK. But if for any reason in the
future we end up with two headers with the same name in the source and
build directories, chances are we want to use the one in the build
directory.

6 years agoDo not ship evconfig-private.h in dist archive
Cristian Morales Vega [Thu, 19 Apr 2018 08:28:36 +0000 (09:28 +0100)]
Do not ship evconfig-private.h in dist archive

It will be generated by autotools, so there is not reason to include it.
And infact this breaks compilation with out-of-tree builds (VPATH),
since, for the quote form of the include directive, headers in the
directory of the file with the #include line have priority over those
named in -I options, the copy of evconfig-private.h from the source
directory had priority over the one in the build directory.

6 years agoAdopt ignore rules for cmake + ninja
Azat Khuzhin [Sun, 22 Apr 2018 21:26:08 +0000 (00:26 +0300)]
Adopt ignore rules for cmake + ninja

In case we have build directory differs from source directory there will be
bunch of files we should ignore, so just remove leading "/" for some or rules.
And fix others.

6 years agoFix assert() condition in evbuffer_drain() for IOCP
SuckShit [Fri, 20 Apr 2018 07:48:32 +0000 (15:48 +0800)]
Fix assert() condition in evbuffer_drain() for IOCP

In the case of iocp, in the for loop above, there is a situation where:
  remaining == chain->off == 0

And this happens due to CHAIN_PINNED_R() case (that is used only in
buffer_iocp.c)

Closes: #630 (picked)
6 years agoAdd Gitter badge
The Gitter Badger [Sun, 15 Apr 2018 23:02:05 +0000 (23:02 +0000)]
Add Gitter badge

6 years agocmake: ensure windows dll's are installed as well as lib files
Philip Herron [Thu, 5 Apr 2018 14:17:06 +0000 (15:17 +0100)]
cmake: ensure windows dll's are installed as well as lib files

Closes: #621
6 years agoGenerating evdns_base_config_windows_nameservers docs on all platforms
dpayne [Tue, 3 Apr 2018 22:43:22 +0000 (15:43 -0700)]
Generating evdns_base_config_windows_nameservers docs on all platforms

6 years agoFixing doxygen docs for evdns_base_search_clear when generated on non-windows machines
dpayne [Tue, 3 Apr 2018 22:17:51 +0000 (15:17 -0700)]
Fixing doxygen docs for evdns_base_search_clear when generated on non-windows machines

6 years agoFix build with LibreSSL 2.7
Bernard Spil [Mon, 2 Apr 2018 11:18:27 +0000 (13:18 +0200)]
Fix build with LibreSSL 2.7

LibreSSL 2.7 implements OpenSSL 1.1 API except for BIO_get_init()

See also: https://bugs.freebsd.org/226900
Signed-off-by: Bernard Spil <brnrd@FreeBSD.org>
Closes: #617 (cherry-pick)
6 years agoAvoid possible SEGVs in select() (in unit tests)
Philip Prindeville [Wed, 28 Mar 2018 14:25:28 +0000 (08:25 -0600)]
Avoid possible SEGVs in select() (in unit tests)

Per the POSIX definition of select():

http://pubs.opengroup.org/onlinepubs/009696699/functions/pselect.html

"Upon successful completion, the select() function may modify the object
pointed to by the timout argument."

If "struct timeval" pointer is a "static const", it could potentially
be allocated in a RO text segment.  The kernel would then try to copy
back the modified value (with the time remaining) into a read-only
address and SEGV.

Signed-off-by: Philip Prindeville <philipp@redfish-solutions.com>
Closes: #614
6 years agodon't fail be_null_filter if bytes are copied
Greg Hazel [Wed, 21 Mar 2018 01:04:53 +0000 (18:04 -0700)]
don't fail be_null_filter if bytes are copied

Otherwise it will not reset timeouts for the IO, but other things should
still works correctly.

Also evbuffer_remove_buffer() could return 0, due to empty buffers, for
example during flushing bufferevent, so let's count this is BEV_OK too.

6 years agoCall underlying bev ctrl GET_FD on filtered bufferevents
Greg Hazel [Tue, 20 Mar 2018 22:12:07 +0000 (15:12 -0700)]
Call underlying bev ctrl GET_FD on filtered bufferevents

Fixes: #611
Fixes: #610
6 years agoFix base unlocking in event_del() if event_base_set() runned in another thread
Azat Khuzhin [Tue, 27 Feb 2018 18:12:14 +0000 (21:12 +0300)]
Fix base unlocking in event_del() if event_base_set() runned in another thread

Image next situation:
  T1:                                        T2:
   event_del_()
     lock the event.ev_base.th_base_lock
     event_del_nolock_()                     event_set_base()
     unlock the event.ev_base.th_base_lock

In this case we will unlock the wrong base after event_del_nolock_()
returns, and deadlock is likely to happens, since event_base_set() do
not check any mutexes (due to it is possible to do this only if event is
not inserted anywhere).

So event_del_() has to cache the base before removing the event, and
cached base.th_base_lock after.

6 years agoConvert event_debug_*() helpers from macros to static functions
Azat Khuzhin [Tue, 27 Feb 2018 18:08:35 +0000 (21:08 +0300)]
Convert event_debug_*() helpers from macros to static functions

6 years agoProvide Makefile variables LIBEVENT_{CFLAGS,CPPFLAGS,LDFLAGS}
stenn [Sun, 14 May 2017 10:22:11 +0000 (06:22 -0400)]
Provide Makefile variables LIBEVENT_{CFLAGS,CPPFLAGS,LDFLAGS}

This will allow a parent package to specify compiler, CPP, and linker
flags to a libevent built as a sub-package. Document this in
whatsnew-2.2.txt.

Picked-from: #506
Closes: #506
6 years agohttp: remove message in case !Content-Length and Connection!=close
Azat Khuzhin [Tue, 13 Feb 2018 14:43:02 +0000 (17:43 +0300)]
http: remove message in case !Content-Length and Connection!=close

Since [1] GET can have body, and hence for every incomming connection it
will print this error.

[1] db483e3b002b33890fc88cadd77f6fd1fccad2d2 ("Allow bodies for
GET/DELETE/OPTIONS/CONNECT")

Noticed-by: BotoX (irc)
Refs: #408

6 years agobuffer: fix incorrect unlock of the buffer mutex (for deferred callbacks)
Azat Khuzhin [Mon, 12 Feb 2018 16:43:13 +0000 (19:43 +0300)]
buffer: fix incorrect unlock of the buffer mutex (for deferred callbacks)

TSAN reports:
  WARNING: ThreadSanitizer: unlock of an unlocked mutex (or by a wrong thread) (pid=17111)
      #0 pthread_mutex_unlock /build/gcc/src/gcc/libsanitizer/sanitizer_common/sanitizer_common_interceptors.inc:3621 (libtsan.so.0+0x00000003b71c)
      #1 evbuffer_add <null> (libevent_core-2.2.so.1+0x00000000ddb6)
      ...

    Mutex M392 (0x7b0c00000f00) created at:
      #0 pthread_mutex_init /build/gcc/src/gcc/libsanitizer/tsan/tsan_interceptors.cc:1117 (libtsan.so.0+0x0000000291af)
      #1 <null> <null> (libevent_pthreads-2.2.so.1+0x000000000d46)
      ...

$ addr2line -e /lib/libevent_core-2.2.so.1 0x00000000ddb6
/src/libevent/buffer.c:1815 (discriminator 1)

Introduced-in: ae2b84b2575be93d0aebba5c0b78453836f89f3c ("Replace
deferred_cbs with event_callback-based implementation.")

6 years agoFix typos in comments
Dmitry Alimov [Mon, 15 Jan 2018 14:30:08 +0000 (17:30 +0300)]
Fix typos in comments

6 years agohttp: fix leaks in evhttp_uriencode()
Azat Khuzhin [Tue, 9 Jan 2018 18:44:57 +0000 (21:44 +0300)]
http: fix leaks in evhttp_uriencode()

Fixes: #584
6 years agoMerge branch 'listener-immediate-close'
Azat Khuzhin [Thu, 4 Jan 2018 16:28:59 +0000 (19:28 +0300)]
Merge branch 'listener-immediate-close'

* listener-immediate-close:
  test/listener: cover immediate-close logic
  Immediately stop trying to accept more connections if listener disabled

6 years agotest/listener: cover immediate-close logic
Azat Khuzhin [Thu, 4 Jan 2018 16:26:50 +0000 (19:26 +0300)]
test/listener: cover immediate-close logic

6 years agoImmediately stop trying to accept more connections if listener disabled
John Fremlin [Mon, 18 Dec 2017 03:43:00 +0000 (22:43 -0500)]
Immediately stop trying to accept more connections if listener disabled

This is a refined version of the logic previously in #578

The rationale is that the consumer of sockets may wish to temporarily
delay accepting for some reason (e.g. being out of file-descriptors).
The kernel will then queue them up. The kernel queue is bounded and
programs like NodeJS will actually try to quickly accept and then close
(as the current behaviour before this PR).

However, it seems that libevent should allow the user to choose whether
to accept and respond correctly if the listener is disabled.

6 years agoFixed last_with_datap description
Marcin Szewczyk [Mon, 18 Dec 2017 13:01:15 +0000 (14:01 +0100)]
Fixed last_with_datap description

Grammar and typo fix

6 years agohttp: add callback to allow server to decline (and thereby close) incoming connections.
John Fremlin [Fri, 1 Dec 2017 01:29:32 +0000 (01:29 +0000)]
http: add callback to allow server to decline (and thereby close) incoming connections.

This is important, as otherwise clients can easily exhaust the file
descriptors available on a libevent HTTP server, which can cause
problems in other code which does not handle EMFILE well: for example,
see https://github.com/bitcoin/bitcoin/issues/11368

Closes: #578 (patch cherry picked)
6 years agoFix typo in cmake because of which EVENT__SIZEOF_SIZE_T was wrong
Azat Khuzhin [Sat, 16 Dec 2017 20:28:41 +0000 (23:28 +0300)]
Fix typo in cmake because of which EVENT__SIZEOF_SIZE_T was wrong

Fixes: #580
6 years agoCONNECT method only takes an authority
Greg Hazel [Sat, 2 Dec 2017 20:53:57 +0000 (12:53 -0800)]
CONNECT method only takes an authority

6 years agoIf precise_time is false, we should not set EVENT_BASE_FLAG_PRECISE_TIMER
yongqing.jiao [Sat, 9 Dec 2017 06:39:07 +0000 (14:39 +0800)]
If precise_time is false, we should not set EVENT_BASE_FLAG_PRECISE_TIMER

Fixes: 630f077c296de61c7b99ed83bf30de11e75e2740 ("Simple unit tests for
monotonic timers")

6 years agoMerge branch 'evconnlistener-do-not-close-client-fd'
Azat Khuzhin [Sun, 10 Dec 2017 20:57:19 +0000 (23:57 +0300)]
Merge branch 'evconnlistener-do-not-close-client-fd'

Fixes: #577
* evconnlistener-do-not-close-client-fd:
  listener: cover closing of fd in case evconnlistener_free() called from acceptcb
  Revert "Fix potential fd leak in listener_read_cb()"

6 years agolistener: cover closing of fd in case evconnlistener_free() called from acceptcb
Azat Khuzhin [Tue, 5 Dec 2017 20:26:28 +0000 (23:26 +0300)]
listener: cover closing of fd in case evconnlistener_free() called from acceptcb

6 years agoRevert "Fix potential fd leak in listener_read_cb()"
Azat Khuzhin [Tue, 5 Dec 2017 19:59:20 +0000 (22:59 +0300)]
Revert "Fix potential fd leak in listener_read_cb()"

This reverts commit a695a720cda892c270736d127333d73553842094.

@kgraefe:
  "I believe that this commit is just wrong: if lev->cnt is not 1 after
  the callback, new_fd will still never be closed in listener_read_cb().
  So in that case it is the responsibility of the user's code to close
  the file descriptor (which is fine). But why shouldn't it be in the
  other case? And how does the user's code know?"

And I agree

6 years agoFix generation of LibeventConfig.cmake for the installation tree
Andrey Okoshkin [Wed, 29 Nov 2017 08:13:51 +0000 (11:13 +0300)]
Fix generation of LibeventConfig.cmake for the installation tree

'LIBEVENT_INCLUDE_DIRS' is properly initialized in 'LibeventConfig.cmake' as
'LibeventConfig.cmake.in' contains usage of 'LIBEVENT_CMAKE_DIR' and
'EVENT_INSTALL_INCLUDE_DIR' variables but not 'EVENT_CMAKE_DIR' and
'EVENT__INCLUDE_DIRS'.
Related typos are fixed.

6 years agoAdd missing includes into openssl-compat.h
Azat Khuzhin [Wed, 22 Nov 2017 07:33:15 +0000 (10:33 +0300)]
Add missing includes into openssl-compat.h

Before it depends from the caller #include appropriate headers (at least
for OPENSSL_VERSION_NUMBER), but let's make it independent.

Fixes: #574
6 years agoFree dns/event bases in dns-example to avoid leaks
Azat Khuzhin [Sun, 19 Nov 2017 23:07:54 +0000 (02:07 +0300)]
Free dns/event bases in dns-example to avoid leaks

6 years agoFix incorrect ref to evhttp_get_decoded_uri in http.h
ejurgensen [Sun, 5 Nov 2017 11:18:49 +0000 (12:18 +0100)]
Fix incorrect ref to evhttp_get_decoded_uri in http.h

Replaces reference in the http.h include header file to evhttp_get_decoded_uri
with evhttp_uridecode. There is no function called evhttp_get_decoded_uri.

6 years agoFix crashing http server when callback do not reply in place from *gencb*
Azat Khuzhin [Sat, 4 Nov 2017 16:13:28 +0000 (19:13 +0300)]
Fix crashing http server when callback do not reply in place from *gencb*

This is the second hunk of the first patch
5ff8eb26371c4dc56f384b2de35bea2d87814779 ("Fix crashing http server when
callback do not reply in place")

Fixes: #567
6 years agocmake doesn't has POLICY CMP0054 in low version
lightningkay [Thu, 2 Nov 2017 06:58:17 +0000 (14:58 +0800)]
cmake doesn't has POLICY CMP0054 in low version

6 years agoAllow bodies for GET/DELETE/OPTIONS/CONNECT
Azat Khuzhin [Sun, 29 Oct 2017 19:53:41 +0000 (22:53 +0300)]
Allow bodies for GET/DELETE/OPTIONS/CONNECT

I checked with nginx, and via it's lua bindings it allows body for all
this methods. Also everybody knows that some of web-servers allows body
for GET even though this is not RFC conformant.

Refs: #408

6 years agoDo not crash when evhttp_send_reply_start() is called after a timeout.
Andreas Gustafsson [Tue, 26 Sep 2017 16:23:29 +0000 (19:23 +0300)]
Do not crash when evhttp_send_reply_start() is called after a timeout.

This fixes the crash reported in issue #509.  The "would be nice"
items discussed in #509 can be addressed separately.

6 years agoFix crashing http server when callback do not reply in place
Azat Khuzhin [Sun, 22 Oct 2017 21:13:37 +0000 (00:13 +0300)]
Fix crashing http server when callback do not reply in place

General http callback looks like:
  static void http_cb(struct evhttp_request *req, void *arg)
  {
    evhttp_send_reply(req, HTTP_OK, "Everything is fine", NULL);
  }

And they will work fine becuase in this case http will write request
first, and during write preparation it will disable *read callback* (in
evhttp_write_buffer()), but if we don't reply immediately, for example:
  static void http_cb(struct evhttp_request *req, void *arg)
  {
    return;
  }

This will leave connection in incorrect state, and if another request
will be written to the same connection libevent will abort with:
  [err] ../http.c: illegal connection state 7

Because it thinks that read for now is not possible, since there were no
write.

Fix this by disabling EV_READ entirely. We couldn't just reset callbacks
because this will leave EOF detection, which we don't need, since user
hasn't replied to callback yet.

Reported-by: Cory Fields <cory@coryfields.com>
7 years agofix handling of close_notify (ssl) in http with openssl bufferevents
Azat Khuzhin [Wed, 13 Sep 2017 18:39:32 +0000 (21:39 +0300)]
fix handling of close_notify (ssl) in http with openssl bufferevents

Since it can arrive after we disabled events in that bufferevent and
reseted fd, hence evhttp_error_cb() could be called after
SSL_RECEIVED_SHUTDOWN.

Closes: #557
7 years agoPort `event_rpcgen.py` and `test/check-dumpevents.py` to Python 3.
Kiyoshi Aman [Thu, 14 Sep 2017 22:39:20 +0000 (17:39 -0500)]
Port `event_rpcgen.py` and `test/check-dumpevents.py` to Python 3.

These scripts remain compatible with Python 2.

7 years agoRemove OpenSSL paragram from README
Azat Khuzhin [Sun, 24 Sep 2017 12:12:13 +0000 (15:12 +0300)]
Remove OpenSSL paragram from README

Because it is mauvais ton to use binaries instead of normal packages
(like apt-get in debian, pacman in arch, and others).

Plus that link was borken and according to [1] OpenSSL do not ship
binaries officially.

And personally I don't think that this is not obvious that you need
openssl libraries to build libevent with it's support, and BTW you need
headers too (of course).

[1]: https://www.openssl.org/community/binaries.html
Fixes: #562