Il 28/06/21 16:33, Mauro S. via Xenomai ha scritto:
Hi all,
I'm using Xenomai3 (master branch, commit
d98c99798836dd3571144203b51ff787d2108f1b) on a X86_64bit Intel Atom
x5-E8000 with 2GB RAM, using kernel from Xenomai repos, in Cobalt
coniguration.
Just for completeness, I'm using a Yocto built filesystem (but I think
it doesn't matter to the issue that I'm going to describe).
Consider the simple test program attached: it creates 120 tasks that
does nothing than sleep. This test is compiled with the Yocto toolchain
using the attached makefile. Toolchain have Xenomai installed in.
(By the way, I needed to raise to 2048 the registry slots number in
kernel configuration: without raising this, I cannot create more than
110 tasks.)
If I kill this program with TERM signal, sighandler does its work and
the application exits cleanly.
If I kill this application with SEGV signal (no custom sighandler
installed), system freezes.
If I build this application with 50 tasks, then SEGV does not cause
system to freeze.
Bisecting the number of tasks, SIGSEGV freezes starts with a number of
tasks greater than 85.
I enabled watchdog (and almost all of the debug options) in Xenomai
section of the kernel configuration, without success.
Are there some system limits? I checked under /proc/xenomai but all
seems to be OK.
Thanks in advance, regards
Hi all,
update: I tested also latest commit of 5.4-x86 kernel + latest commit of
Xenomai master branch, same behavior.
In one test, with 86 tasks active (just one more than the maximum number
of tasks that does not freeze system), system didn't freeeze
immediately, so I had the time to launch a `cat
/proc/xenomai/sched/stat`. I saw that all the process task were in
000480c0 status, consuming 0.0% CPU, except the last one (the 86th one)
that were in 00048040 status, consuming 100% CPU (system is quad core).
In previous mail I also forgot to mention that:
- this CPU is actually used with success with Xenomai 2.6 (32bit mode)
- all dangerous settings (e.g. c-states) are off in BIOS (in fact
Xenomai 2.6 works well)
- xeno-test is executed without errors
- latency measurements are good
- SMI workaround is applied
Thanks in advance, regards
--
Mauro S.