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: Tue, 23 Jun 2015 22:56:41 +0100	[thread overview]
Message-ID: <20150623215641.GR17109@ZenIV.linux.org.uk> (raw)
In-Reply-To: <20150623164608.GR2427@quack.suse.cz>

On Tue, Jun 23, 2015 at 06:46:08PM +0200, Jan Kara wrote:

> Looks good to me. BTW also ext4 (with BIGALLOC feature) and OCFS2 can have
> block allocation unit (called cluster) larger than page size. However the
> block size of both filesystems is still <= page size. So at least ext4
> handles fun with partially initialized clusters by just marking parts
> of the cluster as uninitialized in the extent tree. But the code is still
> pretty messy to be honest.

Well, with UFS there's no place on disk to store such "this block is
uninitialized" marks - it uses a bog-standard Unix inode structure.

There are two units - fragments and blocks.  Block is an aligned group
of adjacent fragments; normal ratio is 8:1.  Block is at least 4Kb
(and always a power of two), fragment is at least a one sector and
block:fragment ratio is at most 8:1.  Inode structure is normal for a Unix
filesystem (12 direct + indirect + double indirect + triple indirect).
Each reference covers a block worth of file offsets and almost all
of them point to full blocks.  Indirects are full blocks as well.
Reference to a block is represented as the number of the first fragment
in it (i.e. with normal parameters bits 0..2 are clear).  Block bitmap is
actually a fragment bitmap (i.e. bit per fragment).  The only situation
when a reference is *not* to a full block is the last reference in a
file shorter than 12*block size (i.e. not requiring indirects at all).
In that case the last direct reference points to less than a full block
(unless the size in fragments is a multiple of block:fragment ratio,
that is).  One unusual thing is that holes can't extend to EOF - the last
byte *must* be allocated.  (BTW, the only difference between UFS2 and
UFS1 in that area is that fragment numbers are 64bit now.  There had been
talk about turning block:fragment ratio into a per-inode value, but so far
nobody has implemented that - ->di_blksize is there, but it's never used).