Re: [PATCH -tip] remove the BKL: Replace BKL in mount/umount syscalls with a mutex

[Ingo Molnar <mingo@elte.hu>]

* Al Viro <viro@ZenIV.linux.org.uk> wrote:

> On Fri, Apr 17, 2009 at 02:01:42AM +0200, Ingo Molnar wrote:
> > 
> > * Al Viro <viro@ZenIV.linux.org.uk> wrote:
> > 
> > > remount is potentially nastier, but then it *is* nasty.  Again, 
> > > it's only per-fs stuff, so the obvious first step is taking BKL 
> > > down into the instances. It doesn't protect anything in VFS; all 
> > > uses are fs internal, so that'll take review of individual 
> > > filesystems.
> > > 
> > > NOTE: do not assume that code in fs/foo/* is correct; "it doesn't 
> > > take BKL elsewhere" does _not_ mean that we don't have races.  
> > > IOW, the same review ought to look for such beasts and deal with 
> > > them.  Mere "oh, no BKL anywhere in that fs" is not enough to 
> > > discard the ->remount_fs() instance.
> > 
> > what kind of races do you mean? Timing sensitive ones that are there 
> > just are not easy to trigger with the BKL held?
> > 
> > Or actual locking interaction between that body of BKL code and 
> > all other BKL using code?
> 
> Old foo_read_super/foo_write_super/foo_put_super/foo_remount_fs 
> for the same foo.  IOW, per-driver (and not per-fs - that's taken 
> care of) data structures.  Arbitrary weird ones.

Could you give an example of such a weird interaction?

I fear, unless i'm misunderstanding your feedback, that you are 
setting the purist's irrealistically high burden to get rid of the 
BKL from the VFS here.

"Arbitrary weird ones" means all BKL using sites in the kernel - all 
~800 ones - up to 800x800 == close to a million interactions to 
check.

That's not manageable humanly, nor does it really matter: most BKL 
assumptions have bit-rotted anyway already (by the sheer entropy of 
kernel facilities growing new schedule() sites or moving them, over 
a period of 10 years or more) and many BKL using drivers are broken 
already or dont need it at all.

Thanks,

	Ingo