Run queue corruption issue

Run queue corruption issue

[Jerrin Shaji George]

Hi All,

I wanted help with a piece of code that I have been working on.

Please see -

https://gist.github.com/jerrinsg/333e584d1f65dc95b9f13b61dcebdaa7

I have written two function, migrate_to and migrate_back. migrate_to is used
to remove a process from the run queue, and migrate_back is used to insert this
process back into the run queue.

The gist is from a taken from a larger project, where we are working on building
a mechanism to support thread migration across heterogeneous processors.
migrate_to_call() will be called by a thread which wants to remove itself from
the run queue (hence, it will pass the current task struct as the migration
argument). Once the other processor completes execution of the assigned task, it
will interrupt the main processor, which runs an interrupt handler, which in
turn calls the migrate_back_call() function. It passes the task struct of the
process that was removed from the run queue earlier to this function.

This mechanism works fine the first few times, but when this process is repeated
many times in a loop, I am seeing a run queue corruption:
https://gist.github.com/jerrinsg/0ab09cd435d8d2cb6ae692c7e6f4f26b

Is there anything wrong in the process dequeue or enqueue function that I have
written? Please help!

Kernel used: Linux 3.13

Thanks,
Jerrin

Run queue corruption issue

[Greg KH]

On Tue, May 17, 2016 at 06:55:07PM -0400, Jerrin Shaji George wrote:
> Hi All,
> 
> I wanted help with a piece of code that I have been working on.
> 
> Please see -
> 
> https://gist.github.com/jerrinsg/333e584d1f65dc95b9f13b61dcebdaa7
> 
> I have written two function, migrate_to and migrate_back. migrate_to is used
> to remove a process from the run queue, and migrate_back is used to insert this
> process back into the run queue.
> 
> The gist is from a taken from a larger project, where we are working on building
> a mechanism to support thread migration across heterogeneous processors.
> migrate_to_call() will be called by a thread which wants to remove itself from
> the run queue (hence, it will pass the current task struct as the migration
> argument). Once the other processor completes execution of the assigned task, it
> will interrupt the main processor, which runs an interrupt handler, which in
> turn calls the migrate_back_call() function. It passes the task struct of the
> process that was removed from the run queue earlier to this function.
> 
> This mechanism works fine the first few times, but when this process is repeated
> many times in a loop, I am seeing a run queue corruption:
> https://gist.github.com/jerrinsg/0ab09cd435d8d2cb6ae692c7e6f4f26b
> 
> Is there anything wrong in the process dequeue or enqueue function that I have
> written? Please help!

volatile doesn't mean what you think it does, please don't use it in the
kernel.

And why are you using "raw_spin_lock()"?

> Kernel used: Linux 3.13

Wow that's obsolete and buggy, why use such an old thing?

greg k-h

Run queue corruption issue

[Jerrin Shaji George]

Hi Greg,

Thanks for your response.

On Tue, May 17, 2016 at 7:20 PM, Greg KH <greg@kroah.com> wrote:
> On Tue, May 17, 2016 at 06:55:07PM -0400, Jerrin Shaji George wrote:
>> Hi All,
>>
>> I wanted help with a piece of code that I have been working on.
>>
>> Please see -
>>
>> https://gist.github.com/jerrinsg/333e584d1f65dc95b9f13b61dcebdaa7
>>
>> I have written two function, migrate_to and migrate_back. migrate_to is used
>> to remove a process from the run queue, and migrate_back is used to insert this
>> process back into the run queue.
>>
>> The gist is from a taken from a larger project, where we are working on building
>> a mechanism to support thread migration across heterogeneous processors.
>> migrate_to_call() will be called by a thread which wants to remove itself from
>> the run queue (hence, it will pass the current task struct as the migration
>> argument). Once the other processor completes execution of the assigned task, it
>> will interrupt the main processor, which runs an interrupt handler, which in
>> turn calls the migrate_back_call() function. It passes the task struct of the
>> process that was removed from the run queue earlier to this function.
>>
>> This mechanism works fine the first few times, but when this process is repeated
>> many times in a loop, I am seeing a run queue corruption:
>> https://gist.github.com/jerrinsg/0ab09cd435d8d2cb6ae692c7e6f4f26b
>>
>> Is there anything wrong in the process dequeue or enqueue function that I have
>> written? Please help!
>
> volatile doesn't mean what you think it does, please don't use it in the
> kernel.
>

This flag was to be used for synchronization. I will change this.

> And why are you using "raw_spin_lock()"?

I used this seeing other usage in sched/core.c. Can please you let me know if I
should instead use a different function to lock the run queue?

>> Kernel used: Linux 3.13
>
> Wow that's obsolete and buggy, why use such an old thing?

This is the codebase that I inherited. Once I get the basic prototype working, I
will be working to port it to a newer version of the kernel.

>
> greg k-h

Run queue corruption issue

[Greg KH]

On Wed, May 18, 2016 at 01:29:55AM -0400, Jerrin Shaji George wrote:
> Hi Greg,
> 
> Thanks for your response.
> 
> On Tue, May 17, 2016 at 7:20 PM, Greg KH <greg@kroah.com> wrote:
> > On Tue, May 17, 2016 at 06:55:07PM -0400, Jerrin Shaji George wrote:
> >> Hi All,
> >>
> >> I wanted help with a piece of code that I have been working on.
> >>
> >> Please see -
> >>
> >> https://gist.github.com/jerrinsg/333e584d1f65dc95b9f13b61dcebdaa7
> >>
> >> I have written two function, migrate_to and migrate_back. migrate_to is used
> >> to remove a process from the run queue, and migrate_back is used to insert this
> >> process back into the run queue.
> >>
> >> The gist is from a taken from a larger project, where we are working on building
> >> a mechanism to support thread migration across heterogeneous processors.
> >> migrate_to_call() will be called by a thread which wants to remove itself from
> >> the run queue (hence, it will pass the current task struct as the migration
> >> argument). Once the other processor completes execution of the assigned task, it
> >> will interrupt the main processor, which runs an interrupt handler, which in
> >> turn calls the migrate_back_call() function. It passes the task struct of the
> >> process that was removed from the run queue earlier to this function.
> >>
> >> This mechanism works fine the first few times, but when this process is repeated
> >> many times in a loop, I am seeing a run queue corruption:
> >> https://gist.github.com/jerrinsg/0ab09cd435d8d2cb6ae692c7e6f4f26b
> >>
> >> Is there anything wrong in the process dequeue or enqueue function that I have
> >> written? Please help!
> >
> > volatile doesn't mean what you think it does, please don't use it in the
> > kernel.
> >
> 
> This flag was to be used for synchronization. I will change this.
> 
> > And why are you using "raw_spin_lock()"?
> 
> I used this seeing other usage in sched/core.c. Can please you let me know if I
> should instead use a different function to lock the run queue?

Ah, don't know, don't mess with the scheduler, thankfully :)

> >> Kernel used: Linux 3.13
> >
> > Wow that's obsolete and buggy, why use such an old thing?
> 
> This is the codebase that I inherited. Once I get the basic prototype working, I
> will be working to port it to a newer version of the kernel.

Try porting it to a modern kernel and then posting your real patch for
review, that would make things a bit more obvious and probably show your
bug better.

good luck,

greg k-h