Re: [PATCH] Provide ways of crashing the kernel through debugfs

[ebiederm@xmission.com (Eric W. Biederman)]

Andrew Morton <akpm@linux-foundation.org> writes:

> On Fri, 29 Jan 2010 07:13:24 +0100 Simon Kagstrom <simon.kagstrom@netinsight.net> wrote:
>
>> On Thu, 28 Jan 2010 16:38:02 +0200
>> Artem Bityutskiy <dedekind1@gmail.com> wrote:
>> 
>> > On Wed, 2010-01-27 at 10:53 +0800, Am__rico Wang wrote:
>> > > > Well, it provides a few more ways of crashing the kernel. That's
>> > > > basically the only additional feature you'll get.
>> > > 
>> > > Yeah, I can see that, but why do I need to care how I crash the kernel
>> > > as long as I can crash it in a way.
>> > 
>> > But Simon did explain in his first e-mail why he cares. You or others
>> > might care for similar reasons.
>> 
>> Another argument for the patch is that it's simple and well-contained,
>> it doesn't touch any other code apart from the driver itself.
>> 
>> It is also easy to extend with other tests, e.g., provoking kernel
>> hangs to test watchdogs and so on.
>> 
>
> Yes, it's the sort of thing which lots of people have written
> throw-away ad-hoc versions of.  It probably makes sense to do it once,
> do it right to save people from having to rererereinvent that wheel.
>
> What do others think?

I think it makes sense, and in fact we have already merged one attempt
at doing this generically.  drivers/misc/lkdtm.c

I think Simon's patch adds some additional interesting failure modes.
write_after_free, corrupt_stack_write, unaligned_load_store.

Simon is there any chance you can change your patch to an enhancement of lkdtm?

lkdtm actually digs into the interesting failure points with a jprobe
to trigger the harder to reproduce scenarios.  Like stack overflow in
an interrupt handler.

Eric