[lkml]   [2011]   [Feb]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH 1/1] ptrace: make sure do_wait() won't hang after PTRACE_ATTACH
On Wednesday 09 February 2011 22:25, Oleg Nesterov wrote:
> > Note that { the task is put into TASK_TRACED state and group stop
> > resume by SIGCONT is ignored. | the task is put into TASK_STOPPED
> > state and the following PTRACE request will transition it into
> > TASK_TRACED. If SIGCONT is received before transition to
> > TASK_TRACED is made, the task will resume execution. If PTRACE
> > request faces with SIGCONT, PTRACE request may fail. }
> To me, the first variant looks better. But, only because it is closer
> to the current behaviour. I mean, it is better to change the things
> incrementally.
> But in the longer term - I do not know. Personally, I like the
> TASK_STOPPED variant. To the point, I was thinking that (perhaps)
> we can change ptrace_stop() so that it simply calls do_signal_stop()
> if it notices ->group_stop_count != 0.
> > The ptracer may resume execution of the task using PTRACE_CONT
> > without affecting other tasks in the group.
> And this is what I do not like. I just can't accept the fact there
> is a running thread in the SIGNAL_STOP_STOPPED group.
> But yes: this is what the current code does, I am not sure we can
> change this, and both PTRACE_CONT-doesnt-resume-until-SIGCONT and
> PTRACE_CONT-acts-as-SIGCONT are not "perfect" too.

Can you enumerate reasons why each of them are not perfect?
I want to understand your thinking better here.

> > There exists a
> > fundamental race condition between SIGCONT and the next PTRACE call
> Yes, and this race is already here, ptracer should take care.

From the API POV, there is no race, if we assume Oleg's interpretation
that "stopped/not-stopped" and "traced/not-traced" states are
completely orthogonal:

As long as task is in "traced" state and it is in ptrace-stop, SIGCONT
delivered to it does not make it run. Only next PTRACE_CONT (or SYSCALL)
will. Neither will SIGCONT delivered to any other thread group member:
even though this will terminate group-stop state and all untraced
tasks will start running, all tasks which are in ptrace-stop will not:
they will wait for the next PTRACE_CONT (or SYSCALL).

I realize that currently it doesn't work like this, because
group-stop and ptrace-stop are intermingled concepts right now.
My point is, it can be made to work that way, and become free
of this particular race.

> > In either case, the fundamentals of ptrace operation don't really
> > change. All ptrace operations are still per-task and ptracer almost
> > always has control over execution of the tracee. Sure, it allows
> > ptraced task to escape group stop but it seems defined clear enough
> > and IMHO actually is a helpful debugging feature.
> Heh, I think we found the place where we can't convince each other.
> What if we toss a coin?

I'm with Oleg on this. If debugger wants to terminate group-stop,
it should just send SIGCONT, not depend on the obscure feature (it is not
documented, right?) that PTRACE_CONT somehow affects group-stop state.

> > > > What do you do about PTRACE requests while a task is group stopped?
> > > > Reject them? Block them?
> > >
> > > Yes, another known oddity. Of course we shouldn't reject or block.

Why they need to be rejected or blocked? Think again about
"strace sleep" interrupted by SIGSTOP (or SIGTSTP):

* sleep runs in nanosleep
* SIGSTOP arrives, strace sees it
* strace logs it and allows it via ptrace(PTRACE_SYSCALL, ..., SIGSTOP)
* sleep process enters group-stop
* nothing happens until some other signal arrives
* say, SIGCONT arrives
* strace logs it and allows it via ptrace(PTRACE_SYSCALL, ..., SIGCONT)

I believe your question is "what if tracer wants to do a ptrace op
on tracee while it is in group-stop" (step 4 above)?

The answer is simple:
the same as if tracer wants to do a ptrace op on tracee while it is running,
that is - ptrace() should return error. For the tracer (in my example,
strace) there is no difference in state after ptrace(PTRACE_SYSCALL, ..., SIGSTOP)
and ptrace(PTRACE_SYSCALL, ..., <0 or SIGWINCH or any other sig>):
in both cases tracer must wait for tracee to enter ptrace-stop before
any ptrace op is allowed.

Jan, from gdb developer's POV, do you have a problem with this?

> > Heh, I'm not asking for proof that it is more useful. :-) But I'm still
> > curious why you think it's important because the benefits aren't
> > apparent to me. Roland and you seem to share this opinion without
> > much dicussion so maybe I'm missing something?
> I can't!
> I hate this from the time when I noticed that the application doesn't
> respond to ^Z under strace. And I used strace exactly because I wanted
> do debug some (I can't recall exactly) problems with jctl. That is all.

Recently I had exactly this experience too. It's frustrating.

> > To me, it
> > isn't too objectionable to allow debuggers to diddle with tracees
> > behind the real parent's back. In fact, it would be quite useful when
> > debugging job control related behaviors. I wouldn't have much problem
> > accepting the other way around - ie. strict job control even while
> > being debugged, but given that it is already allowed and visible, I
> > fail to see why we should change the behavior. It doesn't seem to
> > have enough benefits to warrant such visible change.
> All I can say is: sure, I see your point, and perhaps you are right
> and I am wrong.
> I'd really like to force CC list to participate ;)

You just succeeded :)


 \ /
  Last update: 2011-02-14 00:05    [W:0.152 / U:18.072 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site