Re: vfat BKL/lock_super regression in v2.6.26-rc3-g8f59342

[Bart Trojanowski <bart@jukie.net>]

* Linus Torvalds <torvalds@linux-foundation.org> [080819 18:19]:
> And I think you hit this issue because you probably mounted the USB stick 
> as a "sync" (or dirsync) mount - which is what some distros do by default, 
> even if it is known to cause problems for some flash cards that don't do a 
> good job at wear levelling.

You're right.  I turned on sync a while back so I could just pull the
stick out w/o the need to do a manual sync... probably not a good idea
on my part with respect to the wear of the media.

> But it's good that you did that, because all _my_ testing (which was 
> admittedly very deficient) had been done with a default mount without that 
> thing.

My ignorance paid off!

> Btw, quite often, the right solution may be to remove one of the locks 
> entirely. FAT should actually have been largely BKL free, and my 
> conversion of BKL to super-lock was "overly eager" exactly because it's 
> easier to find deadlocks (and debug things carefully and handle them as 
> they pop up) than it is to find races (which are almost impossible to 
> debug and pinpoint).

I totally agree.  After spending a couple of hours (while doing other
things) bisecting the tree, I found the commit that introduced the
regression.  In contrast, I was able to find the double-lock in 5
minutes.

> In particular, I think fat_write_inode() really is safe.

I wasn't sure there.

> So I'm pretty sure the right fix is to just remove [un]lock_super() 
> entirely from fat_write_inode(). 

Ok, I will test your patch and let you know in a few minutes.

-Bart

-- 
				WebSig: http://www.jukie.net/~bart/sig/