cvs commit: src/sys/kern kern_sig.c

[John Baldwin]

On Wednesday 20 June 2007 03:22:39 am David Xu wrote:
> John Baldwin wrote:
> > On Monday 06 June 2005 01:13:11 am David Xu wrote:
> > 
> >>davidxu     2005-06-06 05:13:11 UTC
> >>
> >>  FreeBSD src repository
> >>
> >>  Modified files:
> >>    sys/kern             kern_sig.c 
> >>  Log:
> >>  Fix a bug relavant to debugging, a masked signal unexpectedly interrupts
> >>  a sleeping thread when process is being debugged.
> >>  
> >>  PR: GNU/77818
> >>  Tested by: Sean C. Farley <sean-freebsd at farley org>
> > 
> > 
> > This actually breaks other debugging as now debuggers or other processes 
using 
> > procfs/ptrace to catch signals can no longer see ignored signals or 
> > SIGSTOP/SIGCONT.  The latter breaks strace when execing a new child 
process 
> > as it opens a race where the child process hangs because the parent 
doesn't 
> > ever see that the child process has stopped itself with SIGSTOP (the 
parent 
> > resumes it with SIGCONT when it sees that).  The signal shouldn't make it 
to 
> > the target thread if it is ignored, but the process should be stopped and 
the 
> > debugger notified of all signals.
> > 
> 
> Where can you find the place SIGSTOP can be masked or ignored ?
> Though SIGCONT can be ignored but it still can resume a suspended
> process. if ignored signals can be seen by debugger, then a sleep(10)
> will be interrupted by ignored signals when the process is being
> debugged, but will work correctly if it is not being debugged, this
> becauses issignal() must be called by debugged thread to report any
> signals to debugger, the thread should be woken up.

I added a printf to stopevent() for S_SIG and it wasn't invoked for SIGSTOP, 
so PIOCSTATUS ioctl via procfs never returns a status saying the process is 
stopped on SIGSTOP.   Rather than try to untangle the mess that is the signal 
code I just patched strace to use the same algo truss does for exec'ing a new 
child process.  4.x worked fine though both for gdb and the existing strace 
algo, so 6.x as it currently stands is a regression.

-- 
John Baldwin