2021-09-29 20:10:37

by Kees Cook

[permalink] [raw]
Subject: Re: [PATCH] Introduce the pkill_on_warn boot parameter

On Wed, Sep 29, 2021 at 12:03:36PM -0700, Dave Hansen wrote:
> On 9/29/21 11:58 AM, Alexander Popov wrote:
> > --- a/kernel/panic.c
> > +++ b/kernel/panic.c
> > @@ -53,6 +53,7 @@ static int pause_on_oops_flag;
> > static DEFINE_SPINLOCK(pause_on_oops_lock);
> > bool crash_kexec_post_notifiers;
> > int panic_on_warn __read_mostly;
> > +int pkill_on_warn __read_mostly;

I like this idea. I can't tell if Linus would tolerate it, though. But I
really have wanted a middle ground like BUG(). Having only WARN() and
panic() is not very friendly. :(

> > unsigned long panic_on_taint;
> > bool panic_on_taint_nousertaint = false;
> >
> > @@ -610,6 +611,9 @@ void __warn(const char *file, int line, void *caller, unsigned taint,
> >
> > print_oops_end_marker();
> >
> > + if (pkill_on_warn && system_state >= SYSTEM_RUNNING)
> > + do_group_exit(SIGKILL);
> > +
> > /* Just a warning, don't kill lockdep. */
> > add_taint(taint, LOCKDEP_STILL_OK);
> > }
>
> Doesn't this tie into the warning *printing* code? That's better than
> nothing, for sure. But, if we're doing this for hardening, I think we
> would want to kill anyone provoking a warning, not just the first one
> that triggered *printing* the warning.

Right, this needs to be moved into the callers of __warn() (i.e.
report_bug(), and warn_slowpath_fmt()), likely with some small
refactoring in report_bug().

--
Kees Cook


2021-09-30 18:35:14

by Alexander Popov

[permalink] [raw]
Subject: Re: [PATCH] Introduce the pkill_on_warn boot parameter

On 29.09.2021 23:06, Kees Cook wrote:
> On Wed, Sep 29, 2021 at 12:03:36PM -0700, Dave Hansen wrote:
>> On 9/29/21 11:58 AM, Alexander Popov wrote:
>>> --- a/kernel/panic.c
>>> +++ b/kernel/panic.c
>>> @@ -53,6 +53,7 @@ static int pause_on_oops_flag;
>>> static DEFINE_SPINLOCK(pause_on_oops_lock);
>>> bool crash_kexec_post_notifiers;
>>> int panic_on_warn __read_mostly;
>>> +int pkill_on_warn __read_mostly;
>
> I like this idea. I can't tell if Linus would tolerate it, though. But I
> really have wanted a middle ground like BUG(). Having only WARN() and
> panic() is not very friendly. :(

Ok, let's see.

Kees, could you also share your thoughts on the good questions by Petr Mladek in
this thread?

>>> unsigned long panic_on_taint;
>>> bool panic_on_taint_nousertaint = false;
>>>
>>> @@ -610,6 +611,9 @@ void __warn(const char *file, int line, void *caller, unsigned taint,
>>>
>>> print_oops_end_marker();
>>>
>>> + if (pkill_on_warn && system_state >= SYSTEM_RUNNING)
>>> + do_group_exit(SIGKILL);
>>> +
>>> /* Just a warning, don't kill lockdep. */
>>> add_taint(taint, LOCKDEP_STILL_OK);
>>> }
>>
>> Doesn't this tie into the warning *printing* code? That's better than
>> nothing, for sure. But, if we're doing this for hardening, I think we
>> would want to kill anyone provoking a warning, not just the first one
>> that triggered *printing* the warning.
>
> Right, this needs to be moved into the callers of __warn() (i.e.
> report_bug(), and warn_slowpath_fmt()), likely with some small
> refactoring in report_bug().

Yes, I see now. Thanks, Dave, Peter and Kees.
The kernel can hit warning and omit calling __warn() that prints the message.
But pkill_on_warn action should be taken each time.

As I can understand now, include/asm-generic/bug.h defines three warning
implementations:
1. CONFIG_BUG=y and the arch provides __WARN_FLAGS. In that case pkill_on_warn
should be checked in report_bug() that you mention.
2. CONFIG_BUG=y and the arch doesn't have __WARN_FLAGS. In that case
pkill_on_warn should be checked in warn_slowpath_fmt().
3. CONFIG_BUG is not set. In that case pkill_on_warn should not be considered.

Please, correct me if needed.

Best regards,
Alexander

2021-09-30 21:54:47

by Kees Cook

[permalink] [raw]
Subject: Re: [PATCH] Introduce the pkill_on_warn boot parameter

On Thu, Sep 30, 2021 at 04:55:37PM +0300, Alexander Popov wrote:
> The kernel can hit warning and omit calling __warn() that prints the message.
> But pkill_on_warn action should be taken each time.
>
> As I can understand now, include/asm-generic/bug.h defines three warning
> implementations:
> 1. CONFIG_BUG=y and the arch provides __WARN_FLAGS. In that case pkill_on_warn
> should be checked in report_bug() that you mention.
> 2. CONFIG_BUG=y and the arch doesn't have __WARN_FLAGS. In that case
> pkill_on_warn should be checked in warn_slowpath_fmt().
> 3. CONFIG_BUG is not set. In that case pkill_on_warn should not be considered.
>
> Please, correct me if needed.

That looks correct to me, yes.

--
Kees Cook