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.
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).
Vincent JARDIN [Mon, 11 Sep 2017 19:56:30 +0000 (21:56 +0200)]
test: fix warning
In function ‘send_a_byte_cb’:
test/regress.c:1853:2: warning: ignoring return value of ‘write’, declared with
attribute warn_unused_result [-Wunused-result]
(void) write(*sockp, "A", 1);
by default, the max buffer size is 16K and histeresis is at 50%, so
a bigger read is needed to unlock writes than you would expect from
other BSD (512 bytes)
this doesn't introduce any regression on FreeBSD 11.1, OpenBSD 6.1, NetBSD 7.1,
macOS 10.12.6 and of course DragonFlyBSD 4.8.1, and most of them show
a max pipe size of 64K, so the read call should drain them all regardless
of how conservative they are on the free pipe space they will require
(usually 512 bytes) before kevent reports the fd as ready for write.
I couldn't find a reference to which bug this code was trying to look for
and it seems to be there from the beginning of git history so it might be
no longer relevant.
James Synge [Mon, 7 Aug 2017 15:06:28 +0000 (11:06 -0400)]
Fix race in access to ev_res from event loop with event_active()
Detected using ThreadSanitizer, resolved by capturing the value
of ev_res in a local variable while the event is locked, then
passing that captured variable to the callback.
Azat Khuzhin [Mon, 14 Aug 2017 20:11:53 +0000 (23:11 +0300)]
Remove check against passed bufferevent in bufferevent_socket_connect()
It is a bit confusing to check against the result of EVUTIL_UPCAST(),
because it can return not NULL for NULL pointers, but for even though
with bufferevent we are fine (because bufferevent is the first field in
bufferevent_private), there are no checks for "bufev" in bufferevent's
API, so just remove it to make it generic.
Fixes: #542 Signed-off-by: Ivan Maidanski <i.maidanski@samsung.com> Signed-off-by: Azat Khuzhin <a3at.mail@gmail.com>
(cherry picked from commit a10a6f4ed918ea1432820d99e9373f37f906d6f0)
Azat Khuzhin [Mon, 29 May 2017 15:48:48 +0000 (18:48 +0300)]
cmake: set CMP0054 to NEW to avoid variables over expansion (since cmake 3.8)
==> win: CMake Warning (dev) at CMakeLists.txt:782 (elseif):
==> win: Policy CMP0054 is not set: Only interpret if() arguments as variables or
==> win: keywords when unquoted. Run "cmake --help-policy CMP0054" for policy
==> win: details. Use the cmake_policy command to set the policy and suppress this
==> win: warning.
==> win:
==> win: Quoted variables like "MSVC" will no longer be dereferenced when the policy
==> win: is set to NEW. Since the policy is not set the OLD behavior will be used.
==> win: This warning is for project developers. Use -Wno-dev to suppress it.
José Luis Millán [Mon, 13 Apr 2015 06:54:52 +0000 (08:54 +0200)]
Return from event_del() after the last event callback termination
Delete the event from the queue before blocking for the current
event callback termination.
Ensures that no callback is being executed when event_del() returns,
hence making this function a secure mechanism to access data which is
handled in the event callack.
David Benjamin [Thu, 20 Apr 2017 22:03:50 +0000 (18:03 -0400)]
Explicitly call SSL_clear when reseting the fd.
If reconnecting the via BEV_CTRL_SET_FD, bufferevent_openssl.c expects
OpenSSL to reuse the configuration state in the SSL object but retain
connection state. This corresponds to the SSL_clear API.
The code currently only calls SSL_set_connect_state or
SSL_set_accept_state. Due to a quirk in OpenSSL, doing this causes the
handshake to implicitly SSL_clear the next time it is entered. However,
this, in the intervening time, leaves the SSL object in an odd state as
the connection state has not been dropped yet. This behavior also does
not appear to be documented by OpenSSL.
Azat Khuzhin [Tue, 14 Mar 2017 10:33:31 +0000 (13:33 +0300)]
cmake: use APPEND during exporting targets (for old cmake)
On centos with cmake 2.8.12.2:
CMake Error at cmake/AddEventLibrary.cmake:92 (export):
export called with target "event_extra_shared" which requires target
"event_core_shared" that is not in the export list.
If the required target is not easy to reference in this call, consider
using the APPEND option with multiple separate calls.
Azat Khuzhin [Tue, 14 Mar 2017 10:31:02 +0000 (13:31 +0300)]
Merge branch 'win32-fixes'
This patchset fixes win32 builds after some previous patches (referenced
in particular commits), and also removes some quirks for win32.
* win32-fixes:
test: do not return void
log/win32: fix exporting extern variable
log-internal: missing extern "C"
log: remove USE_GLOBAL_FOR_DEBUG_LOGGING
cmake: Export missing symbols for win32
cmake: eliminate EVENT_BUILDING_REGRESS_TEST, since we link with shared libs
test: windows doesn't have WNOWAIT
cmake: clean not used #defines from event-config.h
cmake: add <pthread.h> only for non-win32
Azat Khuzhin [Mon, 13 Mar 2017 18:05:32 +0000 (21:05 +0300)]
Merge branch 'cmake-missing-bits'
This patchset adds next missing things (in compare to autotools):
- pkgconfig
- event_pthreads/event_openssl
- compile shared/static libraries always
And some fixes, because it will not build after fixing other things:
- export missing symbols for cmake (-fvisibility=hidden)
* cmake-missing-bits:
cmake: support visibility for AppleClang too
cmake: fix export absolute path and relative path and cleanup a bit
cmake: generate and install pkgconfig files
cmake: build SHARED and STATIC libraries (like autoconf does)
cmake: add missing event_openssl/event_pthreads libraries
Export symbols for -fvisibility=hidden (under cmake)
Azat Khuzhin [Sun, 12 Mar 2017 10:12:53 +0000 (13:12 +0300)]
travis-ci: do not run coverage on osx
First of all we don't really need two of them, and also with apple-clang
it will fail (because of our checks in cmake), so just exclude it from
the build matrix.
Azat Khuzhin [Wed, 8 Mar 2017 20:15:33 +0000 (23:15 +0300)]
Do not add epoll_sub (syscall wrappers) for epoll in cmake
Nowadays mostly all supported OS'es has this wrappers, and some of them (like
SmartOS) has wrappers but doesn't have __NR_epoll* defines for syscall numbers,
so just drop them (instead of adding yet another check int cmake like autotools
has, since this will break building in cross-compile environment).
Also one minor note, configure doesn't added epoll_sub.c either, since it check
epoll_create() in runtime.
And I tested it in SmartOS and it even works (`make verify`).
Azat Khuzhin [Wed, 8 Mar 2017 14:58:26 +0000 (17:58 +0300)]
Merge branch 'cmake-configure-fixes-v2'
Fixes in cmake, to make it more like configure and support some
cross-compiling.
* cmake-configure-fixes-v2:
cmake: fix extracting of the version from git (check for number of matches)
Detect arch4random_addrandom() existence
Use off_t instead of ev_off_t for sendfile() (fixes android build)
cmake: detect _GNU_SOURCE not by __GNU_LIBRARY__ only (fallback to _GNU_SOURCE)
Check for WNOWAIT in waitpid() in runtime (not in cmake/configure)
cmake: add <pthread.h> into CMAKE_REQUIRED_INCLUDES for sizeof(pthread_t)
cmake: fix values for #cmakedefine
cmake: drop duplicates from event-config template
cmake: add value for the #cmakedefine macros (like autoconf)
cmake: Fix checking of enum values from sysctl.h
Azat Khuzhin [Sun, 5 Mar 2017 23:46:23 +0000 (02:46 +0300)]
Merge branch 'fix-openssl-linking'
* fix-openssl-linking:
sample/https-client: use ERR_remove_*state() when we have them
Do not check for ERR_remove_thread_state() (do not link ssl into every library)
Trond Norbye [Tue, 21 Feb 2017 11:54:02 +0000 (12:54 +0100)]
Fix RPATH for APPLE
By setting the CMake minimum version to 3.1 CMake automatically
adds the correct magic to make the library relocatable on
the filesystem (instead of burning the location of the library
at link time into the binary).
ex:
otool -L bin/http-connect
bin/http-connect:
@rpath/libevent_extra.2.2.0.dylib (compatibility version 2.2.0, current version 0.0.0)
/usr/lib/libz.1.dylib (compatibility version 1.0.0, current version 1.2.8)
/usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1238.0.0)
David Disseldorp [Mon, 20 Feb 2017 13:47:59 +0000 (14:47 +0100)]
sample/https-client: use host SSL certificate store by default
Currently a static (Debian) certificate path is used by default, which
can be overridden using the -crt parameter. This commit changes the
default behaviour such that the openssl default certificate store is
used, unless overridden by -crt.
Azat Khuzhin [Sun, 29 Jan 2017 20:07:40 +0000 (23:07 +0300)]
Merge branch 'fix-struct-linger'
@jbech
"Accidentally disabled by 0dda56a due to confusion between struct
linger vs. SO_LINGER and #define vs. AC_DEFINE. Try adding synthetic
#error test to confirm."
* fix-struct-linger:
cmake: check for 'struct linger' existence
test/bench*: prefix event-config.h macros after 0dda56a48e94
test/bench_httpclient: restore SO_LINGER usage after 0dda56a48e94
Azat Khuzhin [Wed, 25 Jan 2017 23:37:15 +0000 (02:37 +0300)]
Merge branch 'release-2.1.8-stable-pull'
* release-2.1.8-stable-pull:
Bump version to 2.1.8-stable everywhere
VERSION_INFO should me equal to library version
Bump whatsnew-2.1 document
Update ChangeLog for 2.1.8-stable
README: update AUTHORS
Azat Khuzhin [Sun, 22 Jan 2017 00:59:02 +0000 (03:59 +0300)]
Merge branch 'automake-tests-parallel-v4' (*includes ci bits also*)
This patch set runs tests in parallel (on travis/appveyor/vagrant), it
includes cmake/autotools(automake).
It should significantly decrease time that tests tooks on travis-ci (3-4
times lower, right now it is about 14-17 hours - too long!), but not
without downsides, now because travis-ci workers has limited resources
we will have more timing-related failures, but this is another storry,
anyway ~16 hours is not acceptable. Anyway if machine has enough
resources it is great to have ability to run tests in parallel (which
automake couldn't do before).
Changes for common test env:
- autotools: before 17m / after 3m
- cmake: before 15m / after 87 sec
* automake-tests-parallel-v4:
Run tests in parallel (they are lightweight), on CI and vagrant boxes
travis-ci: install cmake from xenial (CTEST_PARALLEL_LEVEL support)
test: register different tests in automake
test: run different tests under different options (in a wrapper)
automake: do not use serial-tests if parallel-test-harness available
Azat Khuzhin [Fri, 20 Jan 2017 13:31:17 +0000 (16:31 +0300)]
Run tests in parallel (they are lightweight), on CI and vagrant boxes
By default 20 parllel jobs, but one caveat for travis-ci, osx boxes
there slower and have less resources then linux (discovered during
testing), so limit number of parallel jobs to 4 there, and also install
travis_wait, so that travis-ci will not fail the build when there is no
output for 10 minutes, since for osx boxes it is very likely.
Azat Khuzhin [Fri, 20 Jan 2017 13:50:04 +0000 (16:50 +0300)]
travis-ci: install cmake from xenial (CTEST_PARALLEL_LEVEL support)
cmake supports it since 3.0 I guess.
And it must be trusty (not precise), otherwise it fails:
dpkg: error processing
/var/cache/apt/archives/ifupdown_0.8.10ubuntu1_amd64.deb (--unpack):
trying to overwrite '/etc/init.d/networking', which is also in package netbase 4.47ubuntu1
Azat Khuzhin [Fri, 20 Jan 2017 13:29:19 +0000 (16:29 +0300)]
test: register different tests in automake
Before this patch we have one test.sh (well test-script.sh), and tooks
very long to run it sequentially, but they are pretty lightweight, so we
should run then in parallel.
Azat Khuzhin [Thu, 19 Jan 2017 17:53:18 +0000 (20:53 +0300)]
Merge branch 'openssl-filter-fixes-v4'
* openssl-filter-fixes-v4:
be: fix with filtered bufferevents and connect() without EAGAIN
test/ssl: fix bufferevent_getfd() for bufferevent_openssl_filter_new()
be_openssl: Fix writing into filted openssl bufferevent after connected
test/https: separate cases for https client with filtered openssl bufferevent
test/ssl: cover case when we writing to be_openssl after connecting
Azat Khuzhin [Sun, 15 Jan 2017 23:31:54 +0000 (02:31 +0300)]
be: fix with filtered bufferevents and connect() without EAGAIN
With filtered bufferevents (i.e. not real one, that have socket), we can
trigger incorrect callback in this case. Let's look at example with http
and bufferevent_openssl_filter_new():
- bev = bufferevent_openssl_filter_new()
- http layer trying to connect() to localhost with bev
# at this time, bev have writecb/readcb NULL but ev_write/ev_read has
# timeout with 45 secs, default HTTP connect timeout
- and when connect() retruns without EAGAIN (BSD'ism) we called
event_active() before (with EV_WRITE), and this will call ev_write
timeout only, while it is more correct to act on bufferevent instead
of plain event, so let's trigger EV_WRITE for bufferevent which will
do the job (and let's do this deferred).
Azat Khuzhin [Tue, 6 Dec 2016 23:53:07 +0000 (02:53 +0300)]
be_openssl: Fix writing into filted openssl bufferevent after connected
The main problems was due to when bufferevent_openssl has underlying (i.e.
created with bufferevent_openssl_filter_new()) some events was
disabled/suspended, while with openssl, READ can require WRITE and vice-versa
hence this issues.
The BEV_CTRL_GET_FD hunk to fix http subsystem, since it depends from what
bufferevent_getfd() returns.
Azat Khuzhin [Fri, 2 Dec 2016 15:32:03 +0000 (18:32 +0300)]
test/ssl: cover case when we writing to be_openssl after connecting
Right now it fails because of regression for filtered openssl
bufferevent, and by it I mean ssl/bufferevent_filter_write_after_connect
test, and by fails - hang.
Azat Khuzhin [Thu, 19 Jan 2017 17:34:41 +0000 (20:34 +0300)]
test/dns: run async resolving after sync one (to avoid timeouts)
If system resolver (sync one) will respond too slow, then we can fail async
request and evdns will retransmit tham again, but evdns server will accept that
failed requets, so we will have not 2 requests but 4.
Azat Khuzhin [Mon, 19 Dec 2016 07:22:51 +0000 (10:22 +0300)]
Fix UB in evutil_date_rfc1123()
As pointed in https://github.com/libevent/libevent/pull/417#issuecomment-267860738
"code is unsafe because in evutil_date_rfc1123() the pointer to the
automatic variable struct tm cur is used outside the scope it defined."
Checked with `clang -fsanitize=address -fsanitize-address-use-after-scope`
and test that call evutil_date_rfc1123() with tm==NULL
Azat Khuzhin [Sun, 18 Dec 2016 22:26:43 +0000 (01:26 +0300)]
Merge branch 'date-rfc1123'
Follow RFC2616 (Date header should be in RFC1123 format).
But to reproduce this bug, program that uses libevent http-server should
call `setlocale(LC_ALL, "")` to properly initliaze locale, so that
strftime() retruns locale-specific dates.
* date-rfc1123:
add tests for evutil_date_rfc1123().
http: do not use local settings for Date header
Azat Khuzhin [Tue, 6 Dec 2016 22:43:35 +0000 (01:43 +0300)]
Merge branch 'be-openssl-fd-reset-fix-v2'
Two issues:
- dirty_shutdown for openssl 1.1
- BEV_CTRL_SET_FD for bufferevent_openssl didn't reset state
* be-openssl-fd-reset-fix-v2:
Fix dirty_shutdown for openssl 1.1
Fix reusing bufferevent_openssl after fd was reseted (i.e. on new connection)
test/https: fix ssl dirty bypass for https_simple
test/https: cover multiple request over the same connection
test/http: sanity check for http_request_empty_done()
Azat Khuzhin [Sat, 19 Nov 2016 14:08:53 +0000 (17:08 +0300)]
Fix reusing bufferevent_openssl after fd was reseted (i.e. on new connection)
For example if you trying to issue multiple requests over the same
evhttp_conneciton, and if connection already closed (IOW it should be
re-connected), than you will get into trouble since it will got wrong
openssl state. This patch addresses this issue by restoring state to
initial if SETFD called with -1 fd.
Azat Khuzhin [Tue, 6 Dec 2016 22:14:24 +0000 (01:14 +0300)]
Merge branch 'openssl-1.1-init-fixes-v2'
* openssl-1.1-init-fixes-v2:
sample/le-proxy: use TLS_method instead of SSLv23_method (latest deprecated)
test: fix building under openssl 1.1 (init functions has been deprecated)
le-proxy: fix building under openssl 1.1 (init functions has been deprecated)
projects / libevent / log