Re: [LTP/VFS] fcntl SETLEASE fails on ramfs/tmpfs

["J. Bruce Fields" <bfields@fieldses.org>]

On Thu, May 01, 2008 at 07:24:32AM +0100, Al Viro wrote:
> On Tue, Apr 29, 2008 at 05:42:31PM -0400, J. Bruce Fields wrote:
> > The most likely consequences are that a local reader gets out-of-date
> > data for a file that a Samba client has modified.
> > 
> > I suppose that re-checking the d_count and i_count after step 3 might
> > close the race.
>  
> The hell it might.

Yeah, looking back at the code, I suppose by the time we've added the
new lease to the inode's lock list, we've already broken conflicting
leases.  OK.

> Leases are broken, plain and simple.  Not to mention
> anything else, a couple of threads with shared descriptor table will
> bypass these checks happily.

I lost you there.

> 
> FWIW, that's far from the worst problem in fs/locks.c, and not even the
> worst one with leases.
> 
> That, BTW, is a fine demonstration of the reasons why application-specific
> kernel warts(tm) are bad.  Lease support is samba-only turd; so's dnotify,
> with its lovely problems.  And interfaces like that *suck*; they are
> developed with one application in mind and that leads to "we know how it
> will be used" mentality.  With obvious implications for quality of review
> they get from their developers...

I honestly don't understand how exactly Samba uses leases; it'd be
extremely useful to have a concise list of requirements from them.  I
know that they don't really meet the nfsv4 server's requirements (any
bugs aside).

> Al, currently crawling through struct file_lock review and extremely annoyed
> by the amount of turds being found...

Thanks, it's been long in need of more attention--details welcomed.

--b.