Re: Re: de-BKLing blkfront

[Jeremy Fitzhardinge <jeremy@goop.org>]

On 07/21/2010 02:12 PM, Daniel Stodden wrote:
> On Wed, 2010-07-21 at 15:49 -0400, Jeremy Fitzhardinge wrote:
>   
>> When I was preparing the latest set of blkfront patches to send upstream
>> to Jens Axboe, he pointed out there were conflicts with what he
>> currently has queued.
>>
>> It turns out the conflict was from pushing the BKL (lock/unlock_kernel)
>> into the open and release functions.  I did the merge keeping them
>> around all the new stuff you added to those functions, but I wonder if
>> its actually necessary.  Do we rely on open/release being globally
>> serialized in there?
>>
>> I've pushed what I have into the upstream/blkfront branch in xen.git.
>>     
> Whatever it was, a BLK presumably fixed it.
>   

There's an ongoing project to remove the BKL; part of that is to remove
implicit use of the BKL from the core kernel and push uses down to
drivers which need it.  That basically means mechanically adding
lock_kernel/unlock_kernel pairs to driver functions as they're removed
from the core kernel.  blkfront got hit with that at some point (haven't
identified precisely where), so we have the option to remove those
lock/unlocks if they're not necessary.

> Anyway, it should not be necessary any more.
>
> Next I made the common mistake of looking into my code again, and
> immediately found I would send an unlucky opener transiently spinning in
> blkdev_get. Sometimes I wonder what I'm thinking.
>   

What's the effect of that?

> Hold on for a patch to both.
>   

Thanks.  BTW, did you get a change to look into those barrier comments?

    J