by Jon Medhurst (Tixy)

[permalink] [raw]

Subject: Re: Re: [PATCH v14 7/7] ARM: kprobes: enable OPTPROBES for ARM 32

On Tue, 2014-12-09 at 19:14 +0900, Masami Hiramatsu wrote:
> (2014/12/08 20:50), Jon Medhurst (Tixy) wrote:> arch_optimize_kprobes is calling __arch_optimize_kprobes, which is
> > iterating over a list of probes and removing each one in turn, if this
> > is happening on multiple cpu's simultaneously, it's not clear to me that
> > such an operation is safe. list_del_init calls __list_del which does
> >
> > next->prev = prev;
> > prev->next = next;
> >
> > so what happens if another cpu is at the same time updating any of those
> > list entries? Without even fully analysing the code I can see that with
> > the fact that the list handling helpers have no memory barriers, that
> > the above two lines could be seen to execute in the reverse order, e.g.
> >
> > prev->next = next;
> > next->prev = prev;
> >
> > so another CPU could find and delete next before this one has finished
> > doing so. Would the list end up in a consistent state where no loops
> > develop and no probes are missed? I don't know the answer and a full
> > analysis would be complicated, but my gut feeling is that if a cpu can
> > observe the links in the list in an inconsistent state then only bad
> > things can result.
>
> Just a comment, arch_optimize_kprobes() are only called under
> kprobe_mutex held. No concurrent update happens :)

Except in the case of the code I was commenting on which was using
stop_machine to make all cpu's simultaneously do the work of
arch_optimize_kprobes :-)

--
Tixy

2014-12-09 15:13:27

by 平松雅巳 / HIRAMATU，MASAMI

[permalink] [raw]

Subject: Re: [PATCH v14 7/7] ARM: kprobes: enable OPTPROBES for ARM 32

(2014/12/09 19:30), Jon Medhurst (Tixy) wrote:
> On Tue, 2014-12-09 at 19:14 +0900, Masami Hiramatsu wrote:
>> (2014/12/08 20:50), Jon Medhurst (Tixy) wrote:> arch_optimize_kprobes is calling __arch_optimize_kprobes, which is
>>> iterating over a list of probes and removing each one in turn, if this
>>> is happening on multiple cpu's simultaneously, it's not clear to me that
>>> such an operation is safe. list_del_init calls __list_del which does
>>>
>>> next->prev = prev;
>>> prev->next = next;
>>>
>>> so what happens if another cpu is at the same time updating any of those
>>> list entries? Without even fully analysing the code I can see that with
>>> the fact that the list handling helpers have no memory barriers, that
>>> the above two lines could be seen to execute in the reverse order, e.g.
>>>
>>> prev->next = next;
>>> next->prev = prev;
>>>
>>> so another CPU could find and delete next before this one has finished
>>> doing so. Would the list end up in a consistent state where no loops
>>> develop and no probes are missed? I don't know the answer and a full
>>> analysis would be complicated, but my gut feeling is that if a cpu can
>>> observe the links in the list in an inconsistent state then only bad
>>> things can result.
>>
>> Just a comment, arch_optimize_kprobes() are only called under
>> kprobe_mutex held. No concurrent update happens :)
>
> Except in the case of the code I was commenting on which was using
> stop_machine to make all cpu's simultaneously do the work of
> arch_optimize_kprobes :-)

Ah, right! stop_machine with cpu_online_mask cause that problem.

Thanks,

--
Masami HIRAMATSU
Software Platform Research Dept. Linux Technology Research Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: [email protected]