svn commit: r270444 - in head/sys: kern sys

[Konstantin Belousov]

On Tue, Aug 26, 2014 at 05:23:10PM -0400, John Baldwin wrote:
> On Tuesday, August 26, 2014 3:32:10 pm John-Mark Gurney wrote:
> > John Baldwin wrote this message on Tue, Aug 26, 2014 at 15:09 -0400:
> > > On Monday, August 25, 2014 6:30:34 pm John-Mark Gurney wrote:
> > > > John Baldwin wrote this message on Mon, Aug 25, 2014 at 13:35 -0400:
> > > > > On Monday, August 25, 2014 07:02:41 PM Mateusz Guzik wrote:
> > > > > > On Mon, Aug 25, 2014 at 10:23:19AM -0400, John Baldwin wrote:
> > > > > > > On Sunday, August 24, 2014 09:04:09 AM Mateusz Guzik wrote:
> > > > > > > > Author: mjg
> > > > > > > > Date: Sun Aug 24 09:04:09 2014
> > > > > > > > New Revision: 270444
> > > > > > > > URL: http://svnweb.freebsd.org/changeset/base/270444
> > > > > > > > 
> > > > > > > > Log:
> > > > > > > >   Fix getppid for traced processes.
> > > > > > > >   
> > > > > > > >   Traced processes always have the tracer set as the parent.
> > > > > > > >   Utilize proc_realparent to obtain the right process when needed.
> > > > > > > 
> > > > > > > Are you sure this won't break things?  I know of several applications that
> > > > > > > expect a debugger to be the parent when attached and change behavior as a
> > > > > > > result (e.g. inserting a breakpoint on an assertion failure rather than
> > > > > > > generating a core).
Shouldn't such applications use a breakpoint instruction like INT3
unconditionally then ? Detection of the attached debugger is inherently
racy, the debugger might have detached after the test. This, and the
fact that default action for the SIGTRAP is coredumping.

> > > > > > 
> > > > > > Well, this is what linux and solaris do.
> > > > > 
> > > > > Interesting.
> > > > > 
> > > > > > I don't feel strongly about this change. If you really want I'm happy to
> > > > > > revert.
> > > > > 
> > > > > In general I'd like to someday have the debugger-debuggee relationship not 
> > > > > override parent-child and this is a step in that direction.  However, this 
> > > > > will break existing applications, so this needs to be clearly documented in 
> > > > > the release notes.  In addition, we should probably advertise how a process 
> > > > > can correctly determine if it is being run under a debugger (right now you can 
> > > > > do 'getppid()' and use strcmp or strstr on the p_comm of that pid so you can 
> > > > > do different things for "gdb" vs "gcore", etc. so just checking P_TRACED from 
> > > > > kinfo_proc wouldn't be equivalent in functionality)
> > > > 
> > > > But what about when you attach gdb to a running process...  That
> > > > doesn't magicly make the now debugged process a child of gdb does it?
> > > 
> > > % cat hello.c
> > > #include <stdio.h>
> > > 
> > > int
> > > main()
> > > {
> > >         printf("hello world\n");
> > >         (void)getchar();
> > >         return (0);
> > > }
> > > % cc -g hello.c -o hello
> > > % ./hello 
> > > hello world
> > > load: 9.81  cmd: hello 42599 [ttyin] 1.67r 0.00u 0.00s 0% 1056k
> > > 
> > >  < different window >
> > > 
> > > % ps -O ppid -p `pgrep hello`
> > >   PID  PPID  TT  STAT      TIME COMMAND
> > > 42599  5340  16  I+     0:00.00 ./hello
> > > % gdb hello `pgrep hello`
> > > GNU gdb 6.1.1 [FreeBSD]
> > > ...
> > > (gdb) <hit Ctrl-Z>
> > > Suspended
> > > % ps -O ppid -p `pgrep hello`
> > >   PID  PPID  TT  STAT      TIME COMMAND
> > > 42599 45079  16  TX+    0:00.00 ./hello
> > 
> > Wow, learn something new every day...
> > 
> > But doesn't that break apps that use getppid to signal their parent
> > that forked them?
> 
> Until mjg@'s commit, yes.  It's been that way in FreeBSD at least for
> as long as I can remember.  Certainly back to 4.x.

The ps(1) trick continues to work after the commit, since kern_proc
sysctl directly accesses p_pptr to fill ki_ppid. I simply forgot about
it during the review.

Anyway, checking the parent pid is definitely not the right way to
see if the process is under ptrace debugging.  What if the parent
is the debugger ?  The p_flag AKA ki_flag P_TRACED bit seems to be
the correct indicator.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 819 bytes
Desc: not available
URL: <http://lists.freebsd.org/pipermail/svn-src-head/attachments/20140827/ce121a68/attachment-0001.sig>