CPU hotplug issue in dm-crypt driver

CPU hotplug issue in dm-crypt driver

[mittals]

Hi,

We are using dm-crypt driver to encrypt Android Phone on a 2 core target.
We saw random failures (kernel panics) in scatterwalk code when we were
trying to bring secondary cpu online and offline randomly.

I looked into the dm-crypt driver(drivers/md/dm-crypt.c) and put a hack in
the kcryptd_queue_crypt function to acquire hotplug lock (using
get_online_cpus()), just before adding new work in work queue and release
this lock (using put_online_cpus) after completing work in kcryptd_crypt
function.
With this i didn't see any issue during encryption.

So my questions are -

1) Is it a know problem that there is some issue with dm-crypt driver with
cpu hotplug?

2) If yes, then can someone suggest me a cleaner way to resolve this issue?

Thanks in advance!

--Shashank

Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.

Re: CPU hotplug issue in dm-crypt driver

[Milan Broz]

On 03/26/2012 09:28 PM, mittals@codeaurora.org wrote:
> 1) Is it a know problem that there is some issue with dm-crypt driver with
> cpu hotplug?

It is quite possible that there is a bug on CPU hotplugging.

Can you please sent your patch/workaround and how you can reproduce it?

(The code in dmcrypt will change in future. Anyway Mikulas already
had some complaints here regarding CPU hotplug, adding him to cc.)

Milan

Re: CPU hotplug issue in dm-crypt driver

[Mikulas Patocka]

Hi

On Mon, 26 Mar 2012, mittals@codeaurora.org wrote:

> Hi,
> 
> We are using dm-crypt driver to encrypt Android Phone on a 2 core target.
> We saw random failures (kernel panics) in scatterwalk code when we were
> trying to bring secondary cpu online and offline randomly.
> 
> I looked into the dm-crypt driver(drivers/md/dm-crypt.c) and put a hack in
> the kcryptd_queue_crypt function to acquire hotplug lock (using
> get_online_cpus()), just before adding new work in work queue and release
> this lock (using put_online_cpus) after completing work in kcryptd_crypt
> function.
> With this i didn't see any issue during encryption.
> 
> So my questions are -
> 
> 1) Is it a know problem that there is some issue with dm-crypt driver with
> cpu hotplug?

Yes it is a known problem. Not only dm-crypt is affected, there are other 
kernel parts that can crash too, see: 
http://www.redhat.com/archives/dm-devel/2012-March/msg00035.html

> 2) If yes, then can someone suggest me a cleaner way to resolve this issue?

Another possible solution is to flush dm-crypt workqueues in an unplug 
hook. Another possible solution is to use these patches: 
http://people.redhat.com/mpatocka/patches/kernel/dm-crypt-paralelizace/current/ 
--- they add paralelization to dm-crypt and as a side effect, they remove 
the dm-crypt bug. You can actually use only the first five patches: 
dm-crypt-fix-mempool-deadlock.patch, dm-crypt-test-error.patch, 
dm-crypt-rename-pending.patch, dm-crypt-global-static-cipher-data.patch, 
dm-crypt-remove-percpu.patch.

But, as a general solution, I'd recommend to not use CPU unplug at all. 
There are other kernel parts that have the same bug as dm-crypt and it is 
not trivial to find them.

Mikulas

> Thanks in advance!
> 
> --Shashank
> 
> Sent by an employee of the Qualcomm Innovation Center, Inc.
> The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.
> 
> --
> dm-devel mailing list
> dm-devel@redhat.com
> https://www.redhat.com/mailman/listinfo/dm-devel
>