Re: [PATCH] print_worker_info: Handle pointer with more care

Re: [PATCH] print_worker_info: Handle pointer with more care

[Tejun Heo]

On Fri, Aug 16, 2013 at 06:15:07PM +0200, Richard Weinberger wrote:
> On UML I hit the case that pwq is NULL.
> Then we oops at &pwq->wq...

Hmmm?  I'm confused.  &pwq->wq is pwq's pointer + wq's offset in pwq.
It doesn't involve dereferencing pwq->wq.  Maybe uml isn't
implementing probe_kernel_thread()?  Now that I think about it, I'm
not sure how it could.

cc'ing uml people.  Hey, guys, workqueue uses proble_kernel_read() to
print out workqueue related information during oops because those
events are completely asynchronous and workqueue states may not be
consistently accessible.  It seems like uml doesn't implement
probe_kernel_read() and tries direct derference of incorrect pointers
leading to its own oops.  Maybe uml should check whether the memory is
mapped from probe_kernel_read()?

Thanks.

-- 
tejun

Re: [PATCH] print_worker_info: Handle pointer with more care

[Richard Weinberger]

Am 16.08.2013 18:28, schrieb Tejun Heo:
> On Fri, Aug 16, 2013 at 06:15:07PM +0200, Richard Weinberger wrote:
>> On UML I hit the case that pwq is NULL.
>> Then we oops at &pwq->wq...
> 
> Hmmm?  I'm confused.  &pwq->wq is pwq's pointer + wq's offset in pwq.
> It doesn't involve dereferencing pwq->wq.  Maybe uml isn't
> implementing probe_kernel_thread()?  Now that I think about it, I'm
> not sure how it could.
> 
> cc'ing uml people.  Hey, guys, workqueue uses proble_kernel_read() to
> print out workqueue related information during oops because those
> events are completely asynchronous and workqueue states may not be
> consistently accessible.  It seems like uml doesn't implement
> probe_kernel_read() and tries direct derference of incorrect pointers
> leading to its own oops.  Maybe uml should check whether the memory is
> mapped from probe_kernel_read()?

You are already talking to UML people. ;)
Anyway, I'll investigate into that.
What I see so far is that pwq is NULL after probe_kernel_read().

Thanks,
//richard

Re: [PATCH] print_worker_info: Handle pointer with more care

[Tejun Heo]

Hello,

On Fri, Aug 16, 2013 at 06:38:58PM +0200, Richard Weinberger wrote:
> > cc'ing uml people.  Hey, guys, workqueue uses proble_kernel_read() to
> > print out workqueue related information during oops because those
> > events are completely asynchronous and workqueue states may not be
> > consistently accessible.  It seems like uml doesn't implement
> > probe_kernel_read() and tries direct derference of incorrect pointers
> > leading to its own oops.  Maybe uml should check whether the memory is
> > mapped from probe_kernel_read()?
> 
> You are already talking to UML people. ;)

Ooh... :)

> Anyway, I'll investigate into that.
> What I see so far is that pwq is NULL after probe_kernel_read().

Yeah, and that should be fine.  &pwq->wq would be just an offset of wq
from NULL which is an invalid pointer but probe_kernel_read() should
be able to handle that and probably just return 0 or -1 (all bits
set).  I *think* what's necessary is making probe_kernel_read() use
mincore() to fine out whether the requested address is mapped (it
should return -EFAULT if not) and try to dereference the address iff
it's mapped.

Thanks.

-- 
tejun

Re: [PATCH] print_worker_info: Handle pointer with more care

[Richard Weinberger]

Am 16.08.2013 18:45, schrieb Tejun Heo:
> Hello,
> 
> On Fri, Aug 16, 2013 at 06:38:58PM +0200, Richard Weinberger wrote:
>>> cc'ing uml people.  Hey, guys, workqueue uses proble_kernel_read() to
>>> print out workqueue related information during oops because those
>>> events are completely asynchronous and workqueue states may not be
>>> consistently accessible.  It seems like uml doesn't implement
>>> probe_kernel_read() and tries direct derference of incorrect pointers
>>> leading to its own oops.  Maybe uml should check whether the memory is
>>> mapped from probe_kernel_read()?
>>
>> You are already talking to UML people. ;)
> 
> Ooh... :)
> 
>> Anyway, I'll investigate into that.
>> What I see so far is that pwq is NULL after probe_kernel_read().
> 
> Yeah, and that should be fine.  &pwq->wq would be just an offset of wq

Yep. Now my brain also parsed the C notation correctly.
Sorry for the completely wrong patch.

> from NULL which is an invalid pointer but probe_kernel_read() should
> be able to handle that and probably just return 0 or -1 (all bits
> set).  I *think* what's necessary is making probe_kernel_read() use
> mincore() to fine out whether the requested address is mapped (it
> should return -EFAULT if not) and try to dereference the address iff
> it's mapped.

UML needs a custom probe_kernel_read()? Fine. :)

Thanks,
//richard

Re: [PATCH] print_worker_info: Handle pointer with more care

[Richard Weinberger]

Am 16.08.2013 18:53, schrieb Richard Weinberger:
>> Yeah, and that should be fine.  &pwq->wq would be just an offset of wq
> 
> Yep. Now my brain also parsed the C notation correctly.
> Sorry for the completely wrong patch.
> 
>> from NULL which is an invalid pointer but probe_kernel_read() should
>> be able to handle that and probably just return 0 or -1 (all bits
>> set).  I *think* what's necessary is making probe_kernel_read() use
>> mincore() to fine out whether the requested address is mapped (it
>> should return -EFAULT if not) and try to dereference the address iff
>> it's mapped.
> 
> UML needs a custom probe_kernel_read()? Fine. :)

With my own probe_kernel_read() that uses mincore()
it works fine so far. :-)

Thanks,
//richard