Re: [PATCH 1/2 linux-next] Revert "ufs: fix deadlocks introduced by sb mutex merge"

From: Al Viro <viro@ZenIV.linux.org.uk>
To: Jan Kara <jack@suse.cz>
Cc: Fabian Frederick <fabf@skynet.be>,
	Andrew Morton <akpm@linux-foundation.org>,
	Alexey Khoroshilov <khoroshilov@ispras.ru>,
	Ian Campbell <ian.campbell@citrix.com>,
	Roger Pau Monne <roger.pau@citrix.com>,
	Ian Jackson <Ian.Jackson@eu.citrix.com>,
	xen-devel <xen-devel@lists.xen.org>,
	Evgeniy Dushistov <dushistov@mail.ru>,
	linux-fsdevel@vger.kernel.org
Subject: Re: [PATCH 1/2 linux-next] Revert "ufs: fix deadlocks introduced by sb mutex merge"
Date: Wed, 17 Jun 2015 21:31:16 +0100	[thread overview]
Message-ID: <20150617203116.GG17109@ZenIV.linux.org.uk> (raw)
In-Reply-To: <20150617085715.GC1614@quack.suse.cz>

On Wed, Jun 17, 2015 at 10:57:15AM +0200, Jan Kara wrote:
> > Joy...  Folks, is anybody actively maintaining fs/ufs these days?
> 
> Looking into the changelog there wasn't anyone seriously looking into UFS
> for at least 5-6 years... Fabian did some cleanups recently but they were
> mostly cosmetic. So I don't think it's really maintained. OTOH clearly
> there are some users since occasionally someone comes with a bug report.
> Welcome to the world where we have ~50 on disk / network filesystems :-|

FWIW, I've a partial rewrite of that area (completely untested, etc.)
in vfs.git#ufs; it's still in progress, but hopefully it'll be done
by tomorrow.  Basically, it switches the sucker to locking scheme
similar to ext2 (modulo seqlock instead of rwlock for protection of
block pointers, having to be more careful due to 64bit assignments not
being atomic and being unable to fall back to ->truncate_mutex in
case of race on the read side) and starts cleaning the hell out of
the truncate (and eventually create side of get_block as well).

As far as I can see, the root of the problems with that sucker is the
misplaced handling of tail-packing logics.  That had ballooned into
a lot of convoluted codepaths ;-/  I'm not blaming Evgeniy - it *is*
painful to massage into saner shape and the damn thing kept missing
cleanups that were done to similar filesystems due to that...

One thing I'm somewhat unhappy about in what Evgeniy had done is dealing with
UFS vs. UFS2 differences at such a low level.  Sure, the way we did endianness
handling in there provoked exactly that, but it might have been better to
go the other way round; see e.g. fs/minix/itree*.c for opposite approach.

All this stuff is currently completely untested; I'll be using generic
parts of xfstests for beating it up, but any assistance would be welcome.
Note: all testing I'll be realistically able to do will be for FreeBSD
UFS variants - I don't have Solaris left on any boxen, to start with...