Re: [PATCH 1/2] tracing: syscall_*regfunc() can race with copy_process()

[Oleg Nesterov <oleg@redhat.com>]

On 03/17, Steven Rostedt wrote:
>
> On Sun, 2013-03-17 at 20:00 +0100, Oleg Nesterov wrote:
> > On 03/17, Steven Rostedt wrote:
> > >
> > > > Change copy_process() to update the child's TIF_SYSCALL_TRACEPOINT
> > > > under tasklist.
> > >
> > > Is this because "p = dup_task_struct(current);" is outside the lock?
> > > Probably should state this in the change log.
> >
> > Not only, syscall_regfunc/syscall_unregfunc can miss the new child.
> >
> > Just suppose that syscall_regfunc() takes tasklist right before the
> > forking task tries to take it for writing and and the child to the
> > list.
>
> I'm a bit confused by the above. Maybe it's the typo with the "and and"
> that's confusing me.

Yes, "and and" was supposed to be "and add".

But probably I misunderstood you before... Well yes, this is because
"p = dup_task_struct(current)" copies TIF_SYSCALL_TRACEPOINT outside
of the tasklist-protected section which also makes the new task visible
for do_each_thread().

IOW, the state of TIF_SYSCALL_TRACEPOINT bit can be correct after
dup_task_struct(), but it can't be updated until copy_process() add
the child to the list.

> > OK, thanks, will do. But perhaps tracepoint_fork() would be better?
>
> tracepoint_fork() is similar to being called trace_fork() which would be
> considered a tracepoint. Seeing tracepoint_fork() would make me think it
> has something to do with the fork tracepoint.
>
> Do we plan on doing anything other than updating the syscall tracepoint
> flag here? I find the "syscall_tracepoint_update()" very descriptive to
> what is actually happening. While reading the fork code, seeing
> 'syscall_tracepoint_update()' would tell me that this has something to
> do with syscall tracepoints, which it does. But tracepoint_fork() would
> have me think something completely different.

OK, thanks, I am sending v2 in reply to v1.

Oleg.