Обсуждение: BUG #3728: pthread autoconf hangs
The following bug has been logged online:
Bug reference: 3728
Logged by: heasley
Email address: heas@shrubbery.net
PostgreSQL version: 8.1.10
Operating system: sparc-sun-solaris2.9 gcc-4.1.1
Description: pthread autoconf hangs
Details:
If the autoconf script ends-up with a CFLAGS that includes pthreads knobs,
for example -DREENTRANT, it will try to find pthread_join() without linking
in libpthread (ie: without gcc -pthread). The result is a conftest that
hangs because it has some half-baked stub for (among others)
pthread_create() that does nothing and thus it spins waiting for the
thread{1,2}_done variables to be set by a threads that do not exist.
squashing the result of the first check in config/acx_pthread.m4 allows it
find the proper results.
Sorry, I do not know the proper fix given all the platforms that need to
coexist.
heasley napsal(a):
> The following bug has been logged online:
>
> Bug reference: 3728
> Logged by: heasley
> Email address: heas@shrubbery.net
> PostgreSQL version: 8.1.10
> Operating system: sparc-sun-solaris2.9 gcc-4.1.1
> Description: pthread autoconf hangs
> Details:
>
> If the autoconf script ends-up with a CFLAGS that includes pthreads knobs,
> for example -DREENTRANT, it will try to find pthread_join() without linking
> in libpthread (ie: without gcc -pthread). The result is a conftest that
> hangs because it has some half-baked stub for (among others)
> pthread_create() that does nothing and thus it spins waiting for the
> thread{1,2}_done variables to be set by a threads that do not exist.
>
> squashing the result of the first check in config/acx_pthread.m4 allows it
> find the proper results.
>
> Sorry, I do not know the proper fix given all the platforms that need to
> coexist.
Please, can you provide ./configure cmd line and CFLAGS and other related
setting? Also some error output could help. Another question is gcc 4.1 is not
in Solaris 9. How did you install this version?
Zdenek
Thu, Nov 08, 2007 at 11:04:01AM +0100, Zdenek Kotala:
> heasley napsal(a):
> >The following bug has been logged online:
> >
> >Bug reference: 3728
> >Logged by: heasley
> >Email address: heas@shrubbery.net
> >PostgreSQL version: 8.1.10
> >Operating system: sparc-sun-solaris2.9 gcc-4.1.1
> >Description: pthread autoconf hangs
> >Details:
> >
> >If the autoconf script ends-up with a CFLAGS that includes pthreads knobs,
> >for example -DREENTRANT, it will try to find pthread_join() without linking
> >in libpthread (ie: without gcc -pthread). The result is a conftest that
> >hangs because it has some half-baked stub for (among others)
> >pthread_create() that does nothing and thus it spins waiting for the
> >thread{1,2}_done variables to be set by a threads that do not exist.
> >
> >squashing the result of the first check in config/acx_pthread.m4 allows it
> >find the proper results.
> >
> >Sorry, I do not know the proper fix given all the platforms that need to
> >coexist.
>
> Please, can you provide ./configure cmd line and CFLAGS and other related
> setting? Also some error output could help. Another question is gcc 4.1 is
> not in Solaris 9. How did you install this version?
The configure is via NetBSD's pkgsrc system.
./configure --sysconfdir=/usr/pkg/etc/postgresql --datadir=/usr/pkg/share/po
stgresql --with-docdir=/usr/pkg/share/doc/postgresql --with-template=solaris --w
ithout-readline --without-zlib --enable-nls --without-java --without-perl --with
out-python --without-tcl --with-openssl --with-readline --with-zlib --enable-thr
ead-safety --with-libiconv-prefix=/usr/pkg --with-libintl-prefix=/usr/pkg --pref
ix=/usr/pkg --host=sparc-sun-solaris2 --mandir=/usr/pkg/man
from config.log:
CFLAGS=-g -static-libgcc -static-libgcc -D_LARGEFILE64_SOURCE -mcpu=v9 -mtune=ultrasparc -m64 -D__sparc_v9__ -pipe
-I/usr/pkg/include-I/usr/include -Wall -Wmissing-prototypes -Wpointer-arith -Winline -Wdeclaration-after-statement
-Wendif-labels-fno-strict-aliasing
www.shrubbery.net/~heas/pg_config.log.txt
www.shrubbery.net/~heas/conftest.c.txt
I built gcc 4.1, after having some difficulty with other versions. It is a
fairly straight forward build, expect that it has a hack to avoid a libiconv
bug and is within a wrapper script that forces 64-bit options. I did not have
this problem with pre-8.1.5 postgres as I recall; 8.1.4 built fine and I do
not believe there have been any pkgsrc changes that would affect this.
I did try adding an if () around the pthread_create()s, but it would seem that
they return 0. And, indeed conftest did not have libpthread linked.
yew [22] ldd conftest
libintl.so.3 => /usr/pkg/lib/libintl.so.3
libssl.so.0.9.7 => /usr/pkg/lib/libssl.so.0.9.7
libcrypto.so.0.9.7 => /usr/pkg/lib/libcrypto.so.0.9.7
libz.so.1 => /usr/pkg/lib/libz.so.1
libreadline.so.5 => /usr/pkg/lib/libreadline.so.5
librt.so.1 => /usr/lib/64/librt.so.1
libresolv.so.2 => /usr/lib/64/libresolv.so.2
libgen.so.1 => /usr/lib/64/libgen.so.1
libsocket.so.1 => /usr/lib/64/libsocket.so.1
libnsl.so.1 => /usr/lib/64/libnsl.so.1
libdl.so.1 => /usr/lib/64/libdl.so.1
libm.so.1 => /usr/lib/64/libm.so.1
libc.so.1 => /usr/lib/64/libc.so.1
libiconv.so.2 => /usr/pkg/lib/libiconv.so.2
libcurses.so.1 => /usr/lib/64/libcurses.so.1
libaio.so.1 => /usr/lib/64/libaio.so.1
libmd5.so.1 => /usr/lib/64/libmd5.so.1
libmp.so.2 => /usr/lib/64/libmp.so.2
/usr/platform/SUNW,Sun-Fire-280R/lib/sparcv9/libc_psr.so.1
/usr/platform/SUNW,Sun-Fire-280R/lib/sparcv9/libmd5_psr.so.1
indeed there are weak symbols in libc:
[2776] | 676344| 36|FUNC |WEAK |0 |9 |_pthread_create
[4109] | 676580| 40|FUNC |WEAK |0 |9 |_pthread_join
[3999] | 674772| 36|FUNC |WEAK |0 |9 |_pthread_mutex_lock
heasley wrote: > Thu, Nov 08, 2007 at 11:04:01AM +0100, Zdenek Kotala: >> heasley napsal(a): > > The configure is via NetBSD's pkgsrc system. > > ./configure --sysconfdir=/usr/pkg/etc/postgresql --datadir=/usr/pkg/share/po > stgresql --with-docdir=/usr/pkg/share/doc/postgresql --with-template=solaris --w > ithout-readline --without-zlib --enable-nls --without-java --without-perl --with > out-python --without-tcl --with-openssl --with-readline --with-zlib --enable-thr > ead-safety --with-libiconv-prefix=/usr/pkg --with-libintl-prefix=/usr/pkg --pref > ix=/usr/pkg --host=sparc-sun-solaris2 --mandir=/usr/pkg/man It is really strange configure. See --with-readline/--without-readline. Unfrotunately I currently does not have system with S9 similar to yours configuration :(. I tested Sun Studio compiler with following configure switch ./configure --without-readline --enable-thread-safety and it works fine. Can you try Sun studio? http://developers.sun.com/sunstudio/downloads/thankyou.jsp?submit=%A0FREE+Download%A0%BB%A0 > from config.log: > CFLAGS=-g -static-libgcc -static-libgcc -D_LARGEFILE64_SOURCE -mcpu=v9 -mtune=ultrasparc -m64 -D__sparc_v9__ -pipe -I/usr/pkg/include-I/usr/include -Wall -Wmissing-prototypes -Wpointer-arith -Winline -Wdeclaration-after-statement -Wendif-labels-fno-strict-aliasing Can you try build it as 32bit application? If there is not problem with 64bit libraries. > www.shrubbery.net/~heas/pg_config.log.txt > www.shrubbery.net/~heas/conftest.c.txt > > I built gcc 4.1, after having some difficulty with other versions. It is a > fairly straight forward build, expect that it has a hack to avoid a libiconv > bug and is within a wrapper script that forces 64-bit options. I did not have > this problem with pre-8.1.5 postgres as I recall; 8.1.4 built fine and I do > not believe there have been any pkgsrc changes that would affect this. I don't see any difference in ./configure related to pthread. Do you use same process for building? Do you have same version of all libraries, GCC or did you update some version? Can you compile 8.1.4 with same configuration? It seems to me that it could be something wrong with your GCC compilation. Zdenek
Fri, Nov 09, 2007 at 12:17:42PM +0100, Zdenek Kotala: > heasley wrote: > >Thu, Nov 08, 2007 at 11:04:01AM +0100, Zdenek Kotala: > >>heasley napsal(a): > > > > >The configure is via NetBSD's pkgsrc system. > > > >./configure --sysconfdir=/usr/pkg/etc/postgresql > >--datadir=/usr/pkg/share/po > >stgresql --with-docdir=/usr/pkg/share/doc/postgresql > >--with-template=solaris --w > >ithout-readline --without-zlib --enable-nls --without-java --without-perl > >--with > >out-python --without-tcl --with-openssl --with-readline --with-zlib > >--enable-thr > >ead-safety --with-libiconv-prefix=/usr/pkg --with-libintl-prefix=/usr/pkg > >--pref > >ix=/usr/pkg --host=sparc-sun-solaris2 --mandir=/usr/pkg/man > > It is really strange configure. See --with-readline/--without-readline. > Unfrotunately I currently does not have system with S9 similar to yours > configuration :(. I tested Sun Studio compiler with following configure > switch ./configure --without-readline --enable-thread-safety and it > works fine. Can you try Sun studio? > > http://developers.sun.com/sunstudio/downloads/thankyou.jsp?submit=%A0FREE+Download%A0%BB%A0 > > >from config.log: > >CFLAGS=-g -static-libgcc -static-libgcc -D_LARGEFILE64_SOURCE -mcpu=v9 > >-mtune=ultrasparc -m64 -D__sparc_v9__ -pipe -I/usr/pkg/include > >-I/usr/include -Wall -Wmissing-prototypes -Wpointer-arith -Winline > >-Wdeclaration-after-statement -Wendif-labels -fno-strict-aliasing > > Can you try build it as 32bit application? If there is not problem with > 64bit libraries. sort of; I do not have 32bit readline, openssl, etc. and I do not have sunstudio ATM [sorry, I have very little free time]. I can try sunCC later. just using gcc, outside of the pkgsrc system, it will still be 64-bit (since I have a wrapper around gcc that forces the 64bit options), it builds. But, this success is an accident. It claims to be "checking without flags", but in fact it has -pthread [from where?]. configure:16405: checking whether pthreads work without any flags configure:16498: gcc -o conftest -g -static-libgcc -Wall -Wmissing-prototypes -W pointer-arith -Winline -Wdeclaration-after-statement -Wendif-labels -fno-strict- aliasing -Wall -Wmissing-prototypes -Wpointer-arith -Winline -Wdeclaration-after -statement -Wendif-labels -fno-strict-aliasing -pthread -pthreads conftes t.c -lpthread -lz -lrt -lresolv -lgen -lsocket -lnsl -ldl -lm >&5 configure:16504: $? = 0 configure:16508: test -z || test ! -s conftest.err configure:16511: $? = 0 configure:16514: test -s conftest configure:16517: $? = 0 configure:16555: result: yes www.shrubbery.net/~heas/pg_config.log2.txt > >www.shrubbery.net/~heas/pg_config.log.txt > >www.shrubbery.net/~heas/conftest.c.txt > > > >I built gcc 4.1, after having some difficulty with other versions. It is a > >fairly straight forward build, expect that it has a hack to avoid a > >libiconv > >bug and is within a wrapper script that forces 64-bit options. I did not > >have > >this problem with pre-8.1.5 postgres as I recall; 8.1.4 built fine and I do > >not believe there have been any pkgsrc changes that would affect this. > > I don't see any difference in ./configure related to pthread. Do you use > same process for building? Do you have same version of all libraries, > GCC or did you update some version? Can you compile 8.1.4 with same > configuration? yes, it does build, but for the reason as below. > It seems to me that it could be something wrong with your GCC compilation. I wouldn't put anything past gcc, but I do not think it is at fault here. My understanding, at least for NetBSD, FreeBSD and Solaris, is that they all require -pthread; which essentially adds -lpthread. So, unless you have some dependency that picks-up libpthread, you get these bogus (I agree) stubs from libc which I'd suspect are the cause of the spinning.