Date: Mon, 24 Nov 2008 19:57:47 +0100
From: Jens Axboe <jens.axboe@oracle.com>
To: David Woodhouse <dwmw2@infradead.org>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>,
       Matthew Wilcox <matthew@wil.cx>, Tejun Heo <teheo@suse.de>,
       Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
       Nick Piggin <npiggin@suse.de>,
       IDE/ATA development list <linux-ide@vger.kernel.org>,
       Jeff Garzik <jeff@garzik.org>, Dongjun Shin <djshin90@gmail.com>,
       chris.mason@oracle.com
Subject: Re: about TRIM/DISCARD support and barriers
Message-ID: <20081124185747.GN26308@kernel.dk>
References: <4928E010.4090801@kernel.org> <4929023C.2060302@suse.de> <20081123123514.GI5707@parisc-linux.org> <1227447584.4901.405.camel@macbook.infradead.org> <1227480776.25499.3.camel@localhost.localdomain> <1227517430.26957.20.camel@macbook.infradead.org> <1227552171.25499.30.camel@localhost.localdomain> <1227552765.25385.129.camel@macbook.infradead.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <1227552765.25385.129.camel@macbook.infradead.org>
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2687
Lines: 56

On Mon, Nov 24 2008, David Woodhouse wrote:
> On Mon, 2008-11-24 at 13:42 -0500, James Bottomley wrote:
> > On Mon, 2008-11-24 at 09:03 +0000, David Woodhouse wrote:
> > > On Mon, 2008-11-24 at 07:52 +0900, James Bottomley wrote:
> > > > On Sun, 2008-11-23 at 13:39 +0000, David Woodhouse wrote:
> > > > > > We don't attempt to put non-contiguous ranges into a single TRIM yet.
> > > > > 
> > > > > We don't even merge contiguous ranges -- I still need to fix the
> > > > > elevators to stop writes crossing writes,
> > > > 
> > > > I don't think we want to do that ... it's legal if the write isn't a
> > > > barrier and it will inhibit merging.  That may be just fine for a SSD,
> > > > but it's not for spinning media since they get better performance out of
> > > > merged writes.
> > > 
> > > No, I just mean writes _to the same sector_. At the moment, we happily
> > > let those cross each other in the queue.
>  ...
> > It's not a bug ... but changing it might be feasible ... as long as it
> > doesn't affect write performance too much (which I don't think it will),
> > since it is in the critical path.
> 
> We could argue about how much sense it makes to let two writes to the
> same sector actually happen in reverse order.
> 
> Especially given the fact that we actually _do_ preserve ordering in
> some cases; just not in others. (We preserve ordering only if the start
> and end of the duplicate writes are _precisely_ matching; if it's just
> overlapping (which may well happen in the presence of merges), then this
> check doesn't trigger.
> 
> But that's just semantics. Yes, changing it should be feasible. I talked
> to Jens about that at the kernel summit, and we agreed that it should
> probably be done.

The way this currently works is that we sort based on the first sector
in the request. So if you have have an overlap condition between rq1 and
rq2 and then a write gets merged into rq1, then you may have passing
writes. Linux has never guarenteed any write ordering for non-barriers,
so we've never attempted to handle it. Direct aliases (matching first
sectors) are handled as you mention, but that's more of an algorithmic
thing than by design.

My main worry is that this will add considerable overhead to request
sorting. For the rbtree based sorting, we'd have to do a rb_next/rb_prev
to look at adjacent requests. For CFQ it's even worse, since there's no
per-queue big rbtree for sorting.

-- 
Jens Axboe

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/