Re: [PATCH 1/2] BKL: Remove BKL from default_llseek()

[Arnd Bergmann <arnd@arndb.de>]

On Wednesday 18 November 2009, Jamie Lokier wrote:
> Alan Cox wrote:
> > > Using the BKL in llseek() does not protect the inode's i_size from
> > > modification since the i_size is protected by a seqlock nowadays. Since
> > > default_llseek() is already using the i_size_read() wrapper it is not the
> > > BKL which is serializing the access here.
> > > The access to file->f_pos is not protected by the BKL either since its
> > > access in vfs_write()/vfs_read() is not protected by any lock. If the BKL
> > > is not protecting anything here it can clearly get removed.
> > 
> > No. Your logic is flawed
> > 
> > The BKL is protected something here - it protects the change of offset
> > with respect to other BKL users within drivers. The question is what if
> > anything in any other driver code depends upon the BKL and uses it to
> > protect f_pos. Probably very little if anything but a grep for f_pos
> > through the drivers might not be a bad idea before assuming this. Very
> > few touch f_pos except in their own llseek method.
> 
> Of course, drivers shouldn't be using f_pos outside their llseek
> method, as they should all behave the same with pread/pwrite as with
> llseek+read/write.
> 
> Is that mistaken?

There are drivers touching f_pos in ioctl() methods, which is vaguely
reasonable. There are also driver touching it in their read()/write()
methods, which has no effect whatsoever.

I started grepping through the kernel trying to find any instances
of the first case that uses the BKL, but I only found three instances
of the second case and got heavily demotivated by that.

	Arnd <><