Re: [Xenomai-core] Buildbot cannot compile xenoma: rtcanconfig refences pthread_kill
Philippe Gerum wrote:
On Tue, 2006-08-15 at 10:15 +0200, Jan Kiszka wrote:
Niklaus Giger wrote:
Hi
All my PPC based compilation fail with something like
ccache gcc -rdynamic -o .libs/rtcanconfig
rtcanconfig.o
-L/mnt/data.ng/buildslave/buildbot/quick-ppc/build/ppc/src/skins/native
-L/mnt/data.ng/buildslave/buildbot/quick-ppc/build/ppc/src/skins/rtdm
/mnt/data.ng/buildslave/buildbot/quick-ppc/build/ppc/src/skins/rtdm/.libs/librtdm.so
-Wl,--rpath -Wl,/usr/xenomai/lib
/mnt/data.ng/buildslave/buildbot/quick-ppc/build/ppc/src/skins/rtdm/.libs/librtdm.so:
undefined reference to `pthread_kill'
http://ngiger.dyndns.org/buildbot/ppc/builds/229/step-mk_xeno/0
Is this only a PPC problem? It appears first with build 226, which referred
to
revisions 1433 and 1434. Details see
http://ngiger.dyndns.org/buildbot/ppc/builds/226
Jan, could you please have a look at the error, as it seems to me that you
reduced the lib dependencies a little bit too much?
Well, neither rtcanconfig nor librtdm have explicit dependencies on
libpthread. Unfortunately, the generic skin init code now drags in
pthread_kill when the linker fails to remove xeno_handle_mlock_alert for
librtdm. That doesn't happen with my gcc-4.1, likely older compilers are
less smart.
Anyone any ideas how to solve it WITHOUT adding -lpthread to librtdm?
Could you develop a bit more, i.e. what's the issue with adding this
dependency?
It's simply unneeded. This is only a single tiny tool, but we may see
more of them in the future, so I would like to clean this up now.
I
would say either move that functions to a separate module
The reason to have this code in nucleus/bind.h is that we don't have any
dependency on common libs aside of the ones providing the skin syscalls,
and we try to avoid code duplication by not implementing this stuff into
each and every init.c files.
or include it
via a separate header that all skins include except rtdm.
I'm still reluctant to make RTDM an exception to the common rule,
especially to fix a compiler issue.
Actually, the issue is slightly broader, though with varying
significance. The current approach drags xeno_handle_mlock_alert and
xeno_sigxcpu_no_mlock multiple times into the skin lib. Even with
gcc-4.1 that increases libnative by a few data bytes, under 3.3 by ~300
byte of dead code (overall increase of /lib with 3.3: 32 Kbyte).
The attached patch solves both the linking issue and avoids multiple
static includes of code and variables that are only required once. I
think it's cleaner this way, following the idea of asm/bits/ (BTW,
should it become a asm-generic/bits header?).
Jan
Index: include/nucleus/mlock_alert.h
===
--- include/nucleus/mlock_alert.h (revision 0)
+++ include/nucleus/mlock_alert.h (revision 0)
@@ -0,0 +1,37 @@
+#ifndef _XENO_NUCLEUS_MLOCK_ALERT_H
+#define _XENO_NUCLEUS_MLOCK_ALERT_H
+
+#ifndef __XENO_SIM__
+#include stdio.h
+#include stdlib.h
+#include signal.h
+#include pthread.h
+
+__attribute__ ((weak))
+int xeno_sigxcpu_no_mlock = 1;
+
+__attribute__ ((visibility (internal)))
+void xeno_handle_mlock_alert(int sig)
+{
+ struct sigaction sa;
+
+ if (xeno_sigxcpu_no_mlock) {
+ fprintf(stderr, Xenomai: process memory not locked
+ (missing mlockall?)\n);
+ fflush(stderr);
+ exit(4);
+ }
+
+ /* XNTRAPSW was set for the thread but no user-defined handler
+ has been set to override our internal handler, so let's
+ invoke the default signal action. */
+
+ sa.sa_handler = SIG_DFL;
+ sigemptyset(sa.sa_mask);
+ sa.sa_flags = 0;
+ sigaction(SIGXCPU, sa, NULL);
+ pthread_kill(pthread_self(), SIGXCPU);
+}
+#endif /* __XENO_SIM__ */
+
+#endif /* _XENO_NUCLEUS_MLOCK_ALERT_H */
Index: include/nucleus/bind.h
===
--- include/nucleus/bind.h (revision 1440)
+++ include/nucleus/bind.h (working copy)
@@ -6,33 +6,9 @@
#include string.h
#include errno.h
#include signal.h
-#include pthread.h
#include asm/xenomai/syscall.h
-__attribute__ ((weak))
-int xeno_sigxcpu_no_mlock = 1;
-
-static void xeno_handle_mlock_alert(int sig)
-{
- struct sigaction sa;
-
- if (xeno_sigxcpu_no_mlock) {
- fprintf(stderr,
- Xenomai: process memory not locked (missing
mlockall?)\n);
- fflush(stderr);
- exit(4);
- }
-
- /* XNTRAPSW was set for the thread but no user-defined handler
- has been set to override our internal handler, so let's
- invoke the default signal action. */
-
- sa.sa_handler = SIG_DFL;
- sigemptyset(sa.sa_mask);
- sa.sa_flags = 0;
- sigaction(SIGXCPU, sa, NULL);
- pthread_kill(pthread_self(), SIGXCPU);
-}
+void xeno_handle_mlock_alert(int sig);
Re: [Xenomai-core] Buildbot cannot compile xenoma: rtcanconfig refences pthread_kill
On Tue, 2006-08-15 at 14:20 +0200, Jan Kiszka wrote:
Philippe Gerum wrote:
On Tue, 2006-08-15 at 10:15 +0200, Jan Kiszka wrote:
Niklaus Giger wrote:
Hi
All my PPC based compilation fail with something like
ccache gcc -rdynamic -o .libs/rtcanconfig
rtcanconfig.o
-L/mnt/data.ng/buildslave/buildbot/quick-ppc/build/ppc/src/skins/native
-L/mnt/data.ng/buildslave/buildbot/quick-ppc/build/ppc/src/skins/rtdm
/mnt/data.ng/buildslave/buildbot/quick-ppc/build/ppc/src/skins/rtdm/.libs/librtdm.so
-Wl,--rpath -Wl,/usr/xenomai/lib
/mnt/data.ng/buildslave/buildbot/quick-ppc/build/ppc/src/skins/rtdm/.libs/librtdm.so:
undefined reference to `pthread_kill'
http://ngiger.dyndns.org/buildbot/ppc/builds/229/step-mk_xeno/0
Is this only a PPC problem? It appears first with build 226, which
referred to
revisions 1433 and 1434. Details see
http://ngiger.dyndns.org/buildbot/ppc/builds/226
Jan, could you please have a look at the error, as it seems to me that
you
reduced the lib dependencies a little bit too much?
Well, neither rtcanconfig nor librtdm have explicit dependencies on
libpthread. Unfortunately, the generic skin init code now drags in
pthread_kill when the linker fails to remove xeno_handle_mlock_alert for
librtdm. That doesn't happen with my gcc-4.1, likely older compilers are
less smart.
Anyone any ideas how to solve it WITHOUT adding -lpthread to librtdm?
Could you develop a bit more, i.e. what's the issue with adding this
dependency?
It's simply unneeded. This is only a single tiny tool, but we may see
more of them in the future, so I would like to clean this up now.
Sounds reasonable.
I
would say either move that functions to a separate module
The reason to have this code in nucleus/bind.h is that we don't have any
dependency on common libs aside of the ones providing the skin syscalls,
and we try to avoid code duplication by not implementing this stuff into
each and every init.c files.
or include it
via a separate header that all skins include except rtdm.
I'm still reluctant to make RTDM an exception to the common rule,
especially to fix a compiler issue.
Actually, the issue is slightly broader, though with varying
significance. The current approach drags xeno_handle_mlock_alert and
xeno_sigxcpu_no_mlock multiple times into the skin lib. Even with
gcc-4.1 that increases libnative by a few data bytes, under 3.3 by ~300
byte of dead code (overall increase of /lib with 3.3: 32 Kbyte).
The attached patch solves both the linking issue and avoids multiple
static includes of code and variables that are only required once. I
Ok, this patch looks ok and does not rely on funky file-scoped macros,
which is good (this is what I would not want to see in the tree anymore,
hence the bits/ scheme).
think it's cleaner this way, following the idea of asm/bits/ (BTW,
should it become a asm-generic/bits header?).
Probably, yes; and the same goes for nucleus/bind.h. In which case,
checking for __XENO_SIM__ would become useless. In any case, the various
init.c files are not parsed by the compiler when building the simulated
environment.
Jan
plain text document attachment (separate_mlock_alert.patch)
Index: include/nucleus/mlock_alert.h
===
--- include/nucleus/mlock_alert.h (revision 0)
+++ include/nucleus/mlock_alert.h (revision 0)
@@ -0,0 +1,37 @@
+#ifndef _XENO_NUCLEUS_MLOCK_ALERT_H
+#define _XENO_NUCLEUS_MLOCK_ALERT_H
+
+#ifndef __XENO_SIM__
+#include stdio.h
+#include stdlib.h
+#include signal.h
+#include pthread.h
+
+__attribute__ ((weak))
+int xeno_sigxcpu_no_mlock = 1;
+
+__attribute__ ((visibility (internal)))
+void xeno_handle_mlock_alert(int sig)
+{
+ struct sigaction sa;
+
+ if (xeno_sigxcpu_no_mlock) {
+ fprintf(stderr, Xenomai: process memory not locked
+ (missing mlockall?)\n);
+ fflush(stderr);
+ exit(4);
+ }
+
+ /* XNTRAPSW was set for the thread but no user-defined handler
+has been set to override our internal handler, so let's
+invoke the default signal action. */
+
+ sa.sa_handler = SIG_DFL;
+ sigemptyset(sa.sa_mask);
+ sa.sa_flags = 0;
+ sigaction(SIGXCPU, sa, NULL);
+ pthread_kill(pthread_self(), SIGXCPU);
+}
+#endif /* __XENO_SIM__ */
+
+#endif /* _XENO_NUCLEUS_MLOCK_ALERT_H */
Index: include/nucleus/bind.h
===
--- include/nucleus/bind.h(revision 1440)
+++ include/nucleus/bind.h(working copy)
@@ -6,33 +6,9 @@
#include string.h
#include errno.h
#include signal.h
-#include pthread.h
#include asm/xenomai/syscall.h
-__attribute__ ((weak))
-int xeno_sigxcpu_no_mlock = 1;
-
-static void xeno_handle_mlock_alert(int sig)
-{
- struct sigaction sa;
-
- if
Re: [Xenomai-core] Buildbot cannot compile xenoma: rtcanconfig refences pthread_kill
Niklaus Giger wrote: Hi All my PPC based compilation fail with something like ... Things start to work again, but other issues pop up now. o The installation step of xenomai on the tqm860 is broken. For that reason RTnet fails to configure. Please have a look, Niklaus. o The simulator fails in both the posix and vxworks part: http://ngiger.dyndns.org/buildbot/sim/builds/258/step-check_sim/0 Could anyone have a look? o The run test on the hcu_vx seems to cause a hang after the nucleus being loaded: http://ngiger.dyndns.org/buildbot/hcu_vx/builds/128/step-xeno-test/0 hcu3 is fine, a rebuild of ppc is currently pending. A few remarks and wishes: o Some list of the characteristics of those systems would be nice, just to gain overview what the differences are. o Maybe it's useful to strip the kernel configs of some systems (non-related drivers etc.) to reduce the turn-around time. o Some kind of allyesconfig for Xenomai would be nice. 2.4 doesn't seem to have this, but maybe one can derive that from a 2.6 .config, grab all XENO-related switches, and append them to the 2.4 .config. This way new stuff like the CAN stack now would get included faster. On final remark I can only repeat over and over again: Your buildbot is really great stuff! It significantly reduces the round-trip time for detecting and fixing (often trivial) build issues as long as there are not yet millions of Xenomai users updating their SVN checkout every day ;) Thanks, Jan signature.asc Description: OpenPGP digital signature ___ Xenomai-core mailing list Xenomai-core@gna.org https://mail.gna.org/listinfo/xenomai-core
Re: [Xenomai-core] Buildbot cannot compile xenoma: rtcanconfig refences pthread_kill
Am Dienstag, 15. August 2006 20:14 schrieb Jan Kiszka: Niklaus Giger wrote: Hi All my PPC based compilation fail with something like ... Things start to work again, but other issues pop up now. o The installation step of xenomai on the tqm860 is broken. For that reason RTnet fails to configure. Please have a look, Niklaus. I had noticed this just before the compile issues popped up. I will have a look, but the whole buildbot system is quite complex and confuses sometimes even its creator. o The simulator fails in both the posix and vxworks part: http://ngiger.dyndns.org/buildbot/sim/builds/258/step-check_sim/0 I reported the error vxworks-t010823-2:t010823-2.c:163: Expected sequence: SEQ(Test2,2009); got SEQ(Test2,2003) some time ago. Looks like a minor issue to me, as the simulator has its own internal timing. But I always prefer that Gilles fixes these kind of errors. Could anyone have a look? The other errors in the POSIX skin are new and need to be looked at. o The run test on the hcu_vx seems to cause a hang after the nucleus being loaded: http://ngiger.dyndns.org/buildbot/hcu_vx/builds/128/step-xeno-test/0 This is a long standing problem, where Philippe needs to get hand on a PPC40x system to find the error. I tried to give it a look hcu3 is fine, a rebuild of ppc is currently pending. hcu3 is the most important target (at least for me), as it is the only target which runs xeno-test. A few remarks and wishes: o Some list of the characteristics of those systems would be nice, just to gain overview what the differences are. I will try to fix my documentation. But from the beginning I wanted to introduce as much variance as possible in the different buildings, therefore I intentionally used different gcc versions, cross-compiling vs native, only xenomai vs xenomai+RTnet, POSIX vs VXWORKS, everything built-in vs everything as a module. Looking at all the different log files (I know that this is time consuming) usually gives me quite a few of the answers. o Maybe it's useful to strip the kernel configs of some systems (non-related drivers etc.) to reduce the turn-around time. I was thinking about it, but it is only the ppc build which takes a long time to build and I am running usually a specific build of this kernel to assure me that Xenomai is working all right in a real environment. Roundtrip time could also be improved if more buildbot slave were added. Though if anybody has a spare machine running all the time which could support the load of Xenomai builds, I would be glad to help integrate another slaves into my buildbot system. It would be also possible to add specific slave for anybody who wants to be assured that the Xenomai trunk continues to work for his specific board and build environment. o Some kind of allyesconfig for Xenomai would be nice. 2.4 doesn't seem to have this, but maybe one can derive that from a 2.6 .config, grab all XENO-related switches, and append them to the 2.4 .config. This way new stuff like the CAN stack now would get included faster. A allyesconfig is not possible as e.g. VXWORKS and the POSIX-skin have mutually exclusive requirements for the timing system (periodic vs aperiodic). If you want some specific CONFIG set please drop me a note. On final remark I can only repeat over and over again: Your buildbot is really great stuff! It significantly reduces the round-trip time for detecting and fixing (often trivial) build issues as long as there are not yet millions of Xenomai users updating their SVN checkout every day ;) Thanks for the compliment. As long as everybody on the mailing list responds so well to my complaints that something is broken, I will remain motivated to invest my time. Best regards -- Niklaus Giger ___ Xenomai-core mailing list Xenomai-core@gna.org https://mail.gna.org/listinfo/xenomai-core
Re: [Xenomai-core] Buildbot cannot compile xenoma: rtcanconfig refences pthread_kill
Niklaus Giger wrote: Am Dienstag, 15. August 2006 20:14 schrieb Jan Kiszka: Niklaus Giger wrote: Hi All my PPC based compilation fail with something like ... Things start to work again, but other issues pop up now. o The installation step of xenomai on the tqm860 is broken. For that reason RTnet fails to configure. Please have a look, Niklaus. I had noticed this just before the compile issues popped up. I will have a look, but the whole buildbot system is quite complex and confuses sometimes even its creator. :) o The simulator fails in both the posix and vxworks part: http://ngiger.dyndns.org/buildbot/sim/builds/258/step-check_sim/0 I reported the error vxworks-t010823-2:t010823-2.c:163: Expected sequence: SEQ(Test2,2009); got SEQ(Test2,2003) some time ago. Looks like a minor issue to me, as the simulator has its own internal timing. But I always prefer that Gilles fixes these kind of errors. Could anyone have a look? The other errors in the POSIX skin are new and need to be looked at. That looks to me (after finding out how to build that testsuite: make check!) like it is related to latest changes on pthread_cond_xxx. Gilles is on CC now. o The run test on the hcu_vx seems to cause a hang after the nucleus being loaded: http://ngiger.dyndns.org/buildbot/hcu_vx/builds/128/step-xeno-test/0 This is a long standing problem, where Philippe needs to get hand on a PPC40x system to find the error. I tried to give it a look If you happen to have time, I would suggest narrowing the cause down by switching synchronous printk on (ipipe_set_printk_sync(rthal_root); ipipe_set_printk_sync(rthal_root_domain);) and then spreading some printk to the init routines of the nucleus, trying the find the problematic line(s). This printk will even works under hard IRQ lock, so also endless loops in the timer handler can be detected that way. It's just a bit time-consuming for the operator... hcu3 is fine, a rebuild of ppc is currently pending. hcu3 is the most important target (at least for me), as it is the only target which runs xeno-test. A few remarks and wishes: o Some list of the characteristics of those systems would be nice, just to gain overview what the differences are. I will try to fix my documentation. But from the beginning I wanted to introduce as much variance as possible in the different buildings, therefore I intentionally used different gcc versions, cross-compiling vs native, only xenomai vs xenomai+RTnet, POSIX vs VXWORKS, everything built-in vs everything as a module. Looking at all the different log files (I know that this is time consuming) usually gives me quite a few of the answers. Is there already some doc somewhere? Maybe I was just blind. o Maybe it's useful to strip the kernel configs of some systems (non-related drivers etc.) to reduce the turn-around time. I was thinking about it, but it is only the ppc build which takes a long time to build and I am running usually a specific build of this kernel to assure me that Xenomai is working all right in a real environment. Roundtrip time could also be improved if more buildbot slave were added. Though if anybody has a spare machine running all the time which could support the load of Xenomai builds, I would be glad to help integrate another slaves into my buildbot system. It would be also possible to add specific slave for anybody who wants to be assured that the Xenomai trunk continues to work for his specific board and build environment. Well, I would offer to spend a few cycles of our development server, but our build environment and target boards are already the most common ones: latest compiler on x86 for x86. Boring, I know... o Some kind of allyesconfig for Xenomai would be nice. 2.4 doesn't seem to have this, but maybe one can derive that from a 2.6 .config, grab all XENO-related switches, and append them to the 2.4 .config. This way new stuff like the CAN stack now would get included faster. A allyesconfig is not possible as e.g. VXWORKS and the POSIX-skin have mutually exclusive requirements for the timing system (periodic vs aperiodic). If you want some specific CONFIG set please drop me a note. Then please check for the CAN drivers (same as with RTnet: some only build for PPC) and the latest switches of the POSIX skin. On final remark I can only repeat over and over again: Your buildbot is really great stuff! It significantly reduces the round-trip time for detecting and fixing (often trivial) build issues as long as there are not yet millions of Xenomai users updating their SVN checkout every day ;) Thanks for the compliment. As long as everybody on the mailing list responds so well to my complaints that something is broken, I will remain motivated to invest my time. Ok, and if we do not respond, just yell loader! ;) Jan ___ Xenomai-core mailing list
