Re: [PATCH] video4linux: Push down the BKL

[Arjan van de Ven <arjan@infradead.org>]

On Tue, 27 May 2008 15:59:42 -0300
Mauro Carvalho Chehab <mchehab@infradead.org> wrote:

> On Tue, 27 May 2008 10:37:55 -0600
> Jonathan Corbet <corbet@lwn.net> wrote:
> 
> > On Tue, 27 May 2008 13:31:00 -0300
> > Mauro Carvalho Chehab <mchehab@infradead.org> wrote:
> > 
> > > Since the other methods don't explicitly call BKL (and, AFAIK,
> > > kernel open handler don't call it neither), if a program 1 is
> > > opening a device and initializing some data, and a program 2
> > > starts doing ioctl, interrupting program 1 execution in the
> > > middle of a data initialization procedure, you may have a race
> > > condition, since some devices initialize some device global data
> > > during open [1].
> > 
> > In fact, 2.6.26 and prior kernels *do* acquire the BKL on open (for
> > char devices) - that's the behavior that the bkl-removal tree is
> > there to do away with.  So, for example, I've pushed that
> > acquisition down into video_open() instead. 
> > 
> > So, for now, open() is serialized against ioctl() in video
> > drivers.  As soon as you take the BKL out of ioctl(), though, that
> > won't happen, unless the mutex you use is also acquired in the open
> > path.
> 
> Ok.
> 
> A few drivers seem to be almost read to work without BKL. 
> 
> For example, em28xx has already a lock at the operations that change
> values at "dev" struct, including open() method. However, since the
> lock is not called at get operations, it needs to be fixed. I would
> also change it from mutex to a read/write semaphore, since two (or
> more) get operations can safely happen in parallel.
> 
\

please don't use rw/sems just because there MIGHT be parallel.
THey're more expensive than mutexes by quite a bit and you get a lot
less checking from lockdep. They make sense for very specific, very
read biased, contended cases. But please don't use them "just
because"...