I'm having some trouble getting Unbound 1.4.2 (and every other
version) to run with libevent with multiple threads on Solaris 10
10/09 x86. I am using the Sun Studio 12 compiler to building from
source (I have tried with many different versions of gcc but the
compile always errors out).
I have tried multiple scenarios:
1. Compiling with the options --without-pthreads
--without-solaris-threads works fine. The server starts and works as
expected.
2. Compiling with no options works fine. I can run with 1 thread or
many and the server works as expected.
3. Compiling with --with-libevent and using libevent-1.4.13. This
works fine when the configuration is set to start only 1 thread. If
more than 1 thread is started, then the logs fill up with messages
"unbound: [ID 993594 daemon.error] [10258:1] error: could not
event_del on close". When I stop the test, a few seconds later when
the threads stop working, unbound crashes. I could only find one
previous thread related to this while using select in Solaris, but
this does not appear to be the same. libevent should be using
/dev/poll and I have kept the file descriptors low and still get the
messages and crash: "[1268198366] unbound[10262:0] warning: increased
limit(open files) from 256 to 734".
I've seen better performance from 2.0.4-alpha, but it also crashes the
server and it does not pass the regression tests after compiling. I
plan to post on the libevent mailing list, but looks like there is a
routing loop preventing me from accessing it right now.
4. Compiling with --with-libevent and using libev-3.9. The server
starts fine and responds to queries. If I run performance tests where
I hit with multiple queries (slow or fast) each thread will bring a
CPU to 100% and stay that way until the thread is done working. When
I stop the test, a few seconds later when the threads stop working,
unbound crashes. No errors messages are every produced.
Does anyone have any ideas on how I could get this working? I would
prefer to have a multithreaded Unbound over the multi-process Unbound
(I am able to get this working).
Couple of data points:
* compile with gcc on solaris works fine for me. What is the gcc
compile error?
* /dev/poll support has been flakey in libevent and libev in the past.
There is also evports on solaris, can you try to use that?
If you compile with solaris-threads, but not with libevent (or get
libevent to use select() instead) does that work? I.e. is there some
interference between unbounds use of kernel-mutexes and dev/poll ?
* The warning: increased limit(open files): is harmless here, it signals
that the default ulimit for you is 256 open files and obviously with
several threads it needs more open (network-)sockets. It notifies you
that it is overriding the ulimit.
Here's the error I get (with every version of gcc I've tried). I
previously compiled openssl0.9.8m into /opt/unbound (thus the include
flags).
./libtool --quiet --tag=CC --mode=compile /usr/sfw/bin/gcc -I.
-Ildns-src/include -I/opt/unbound/include -I/opt/install/lib
event-1.4.13-stable -I/opt/install/libevent-1.4.13-stable/include
-lnsl -lsocket -I/opt/unbound/include -g -O2 -D_REENTRANT -pthreads -o
build/util/rtt.lo -c util/rtt.c
util/netevent.c: In function `comm_point_send_udp_msg_if':
util/netevent.c:373: error: structure has no member named `msg_control'
util/netevent.c:375: error: structure has no member named `msg_controllen'
util/netevent.c:377: error: structure has no member named `msg_flags'
util/netevent.c:380: warning: assignment makes pointer from integer
without a cast
util/netevent.c:386: warning: passing arg 1 of `memmove' makes pointer
from integer without a cast
util/netevent.c:399: warning: passing arg 1 of `memmove' makes pointer
from integer without a cast
util/netevent.c:405: warning: passing arg 1 of `memset' makes pointer
from integer without a cast
util/netevent.c:408: error: structure has no member named `msg_controllen'
util/netevent.c: In function `comm_point_udp_ancil_callback':
util/netevent.c:468: error: structure has no member named `msg_control'
util/netevent.c:470: error: structure has no member named `msg_controllen'
util/netevent.c:472: error: structure has no member named `msg_flags'
util/netevent.c:485: warning: assignment makes pointer from integer
without a cast
util/netevent.c:486: warning: assignment makes pointer from integer
without a cast
util/netevent.c:491: warning: passing arg 2 of `memmove' makes pointer
from integer without a cast
util/netevent.c:498: warning: passing arg 2 of `memmove' makes pointer
from integer without a cast
gmake: *** [build/util/netevent.lo] Error 1
I was mistaken--the crashes occur when using event ports. Libevent
seems to work fine with /dev/poll (even with lots of threads), but I
would still like it to work with evports /dev/poll works when
using --without-pthreads --with-solaris-threads as well as with
--with-pthreads. Event ports crashes on both.
Gosh, this commandline is missing a lot of options that would have been
set by ./configure. -std=c99, -D__EXTENSIONS__ and a lot more. This
causes solaris to not export certain API, and those errors appear below.
What was your configure output? (also config.log can be interesting)
You did ./configure CC=/usr/sfw/bin/gcc right? Or it can find gcc
itself if /usr/sfw/bin is in $PATH.
/dev/poll works when