Date: Wed, 17 Oct 2007 18:19:07 -0700 (PDT)
Message-Id: <20071017.181907.63126798.davem@davemloft.net>
To: torvalds@linux-foundation.org
Cc: fujita.tomonori@lab.ntt.co.jp, jens.axboe@oracle.com, mingo@elte.hu,
       linux-kernel@vger.kernel.org, jgarzik@pobox.com,
       alan@lxorguk.ukuu.org.uk, tomof@acm.org
Subject: Re: [bug] ata subsystem related crash with latest -git
From: David Miller <davem@davemloft.net>
In-Reply-To: <alpine.LFD.0.999.0710171801000.26902@woody.linux-foundation.org>
References: <alpine.LFD.0.999.0710171342550.26902@woody.linux-foundation.org>
	<20071018080048O.fujita.tomonori@lab.ntt.co.jp>
	<alpine.LFD.0.999.0710171801000.26902@woody.linux-foundation.org>
Mime-Version: 1.0
Content-Type: Text/Plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1892
Lines: 53

From: Linus Torvalds <torvalds@linux-foundation.org>
Date: Wed, 17 Oct 2007 18:07:19 -0700 (PDT)

> sg_next() - as it stands now - never actually looks at the SG that its 
> argument points to: it explicitly *only* looks at the next one.
> 
> That's the bug. If sg_next() looked at the actual *current* sg entry, we 
> wouldn't have any issues to begin with, and that's what I'm arguing we 
> should do in the longer run (where "longer run" is defined as "when Jens 
> does it asap").

What the thing really wants is some kind of indication of state,
without having to bloat up the scatterlist structure.

I believe that we have enough of a limited set of accessors to
sg->page that we can more aggressively encode things in the lower
bits.

I'm thinking of encoding the low two bits of sg->page as
follows:

1) bits == 0

   then the SG list is linear and sg_next() is sg++

2) bits == 1

   the nest SG is an indirect chunk, sg_next() is
   therefore something like:

	next = *((struct scatterlist **)(sg + 1));

3) bits == 2

   this is the last entry in the scatterlist, sg_next() is NULL

So for the cases of ARCH_HAS_SG_CHAIN not being set (ie. back
compatible), we can do no bit encoding in page->flags and just do
sg_next() == sg++, as is done now.

When doing SG chaining, in each non-linear chunk we have to allocate
one more pointer past the end of the scatterlist array (instead of a
full extra scatterlist entry for the indirect pointer encode).  Next,
all sg->page accesses have to be guarded to clear the state bits
out first.

I don't know, maybe it would work, and would make the loop termination
issues easier to handle properly.
-
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/