Re: [patch] x86, ptrace: fix double-free on race

[Oleg Nesterov <oleg@redhat.com>]

On 02/12, Metzger, Markus T wrote:
>
> >-----Original Message-----
> >From: Oleg Nesterov [mailto:oleg@redhat.com]
> >
> >	static inline void arch_bts_init(struct task_struct *p)
> >	{
> >	#ifdef CONFIG_X86_PTRACE_BTS
> >		if (unlikely(p->bts)) {
> >			p->bts = NULL;
> >			p->bts_buffer = NULL;
> >			p->bts_size = 0;
> >			p->thread.bts_ovfl_signal = 0;
> >		}
> >	#endif /* CONFIG_X86_PTRACE_BTS */
> >	}
>
> It looked cleaner and more consistent to me, that way.
>
> Ptrace passes things down to arch in other cases already,
> e.g. ptrace_detach()->ptrace_disable(), ptrace()->arch_ptrace(),
> ptrace_attach()->arch_ptrace_attach().
>
> I think the function should be in arch/x86/ptrace.c - if we call it from
> ptrace or from arch_dup_task_struct() doesn't really matter.
>
> Do you want me to move the call to arch_dup_task_struct()?

Please do what you like more. We can call this from dup_task_struct()
or even from copy_process(). Actually, the initialization above is not
arch dependent, it only depends on CONFIG_BTS.

> >Btw, just curious, bts_ovfl_signal is not used, except the user can
> >set/get it via PTRACE_BTS_CONFIG/PTRACE_BTS_STATUS ?
>
> It's not really necessary for debugging, but we wanted to get the interface
> general enough to support other use cases.
>
> There's already a PMI interrupt handler in perfmon which handles the PEBS
> buffer overflow. Stephane was planning to extract it so it can be used by
> ds.c. Once that is done, we were planning to move the buffer overflow
> handling for BTS and PEBS into ds.c and enable the feature in ptrace.

OK, thanks. (of course I don't understand this magic, but hopefully
can understand your answer ;)

Oleg.