[Top] [All Lists]

Re: [RFC, PATCH 0/12] xfs: compound buffers for directory blocks

To: Christoph Hellwig <hch@xxxxxxxxxxxxx>
Subject: Re: [RFC, PATCH 0/12] xfs: compound buffers for directory blocks
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Mon, 19 Dec 2011 10:01:03 +1100
Cc: xfs@xxxxxxxxxxx
In-reply-to: <20111214183314.GA18724@xxxxxxxxxxxxx>
References: <1323238703-13198-1-git-send-email-david@xxxxxxxxxxxxx> <20111214183314.GA18724@xxxxxxxxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Wed, Dec 14, 2011 at 01:33:14PM -0500, Christoph Hellwig wrote:
> I've been really busy the last days and only managed to look over this
> a bit while waiting for compiles and benchmarks.
> The only really major comment I have is that I start to really dislike
> the way struct xfs_bmbt_irec is used in the compound buffers.  The
> XFS_EXT_DADDR flag basically gives the data types in there a totally
> different meaning and is really confusing for anyone reading the code:
> either we're using a "normal" one passes in from the directory code,
> or an artificial one used internally, which uses different units.
> I suspect we'd really be much better off introducing a new, smaller type
> here, especially as it's only seen inside the buffer code:
> struct xfs_buf_map {
>       sector_t        bm_bno;
>       size_t          bm_len;
> };
> any use it consistently, that is also use it to replace
> b_file_offset/b_buffer_length and b_bn.    This will also make my number
> two suggestion a lot cleaner, that is make sure the new abstraction can
> be implemented for reading/writing log buffers that wrap around, and
> allow to kill off xfs_buf_associate_memory entirely.

Yeah, I was a bit unsure of the best way to do this. I started with
the bmbt_irec code because it already existed and was easy to pass
in from the directory code. It wasn't until I realised the sector
addressing issue later on that the XFS_EXT_DADDR came about. That
was a simple way of making it work, rather than rewriting the
patches again...

Ok, I'll go back and rewrite them to use a xfs_buf_map/vec and try
to address all your other points as well. Stay tuned.


Dave Chinner

<Prev in Thread] Current Thread [Next in Thread>