xfs
[Top] [All Lists]

Re: [PATCH 1/2] direct-io: move aio_complete into ->end_io

To: Jan Kara <jack@xxxxxxx>
Subject: Re: [PATCH 1/2] direct-io: move aio_complete into ->end_io
From: Christoph Hellwig <hch@xxxxxxxxxxxxx>
Date: Fri, 25 Jun 2010 02:36:10 -0400
Cc: Christoph Hellwig <hch@xxxxxxxxxxxxx>, linux-fsdevel@xxxxxxxxxxxxxxx, xfs@xxxxxxxxxxx, linux-ext4@xxxxxxxxxxxxxxx
In-reply-to: <20100624215922.GE3345@xxxxxxxxxxxxx>
References: <20100622122144.302857146@xxxxxxxxxxxxxxxxxxxxxx> <20100622123113.011371666@xxxxxxxxxxxxxxxxxxxxxx> <20100624215922.GE3345@xxxxxxxxxxxxx>
User-agent: Mutt/1.5.20 (2009-08-17)
On Thu, Jun 24, 2010 at 11:59:22PM +0200, Jan Kara wrote:
>   Umm, I don't get this. Looking at the ->end_io callback it has been
> always called with i_alloc_sem held. It's only aio_complete() which will
> be called with i_alloc_sem held after your changes. Or am I missing
> something?

No, that part of the commit message is flat out wrong.  Not sure what
I was thinking when I wrote it.

>   Moreover the async testing you do does not seem to be completely right.
> dio->is_async is a flag that controls whether dio code waits for IO to be
> completed or not. In particular it is not set for AIO that spans beyond
> current i_size so it does not seem to be exactly what you need (at least
> for ext4 it isn't). I think that is_sync_kiocb() is a test that should be
> used to recognize AIO - and that has an advantage that you don't have to
> pass the is_async flag around.

No.  is_sync_kiocb() means the ioctb was not intended as sync I/O from
the start.  But we can only call aio_complete when we returned
-EIOCBQUEUED from ->aio_read/write.  Take a look at the comment near the
end of direct_io_worker().

AIO beyond i_size is not supported using blockdev_direct_IO yet.  I
think I can add it fairly easily for XFS, but that will require
passing a new DIO_* flag to __blockdev_direct_IO which will make
is_async true for writes beyond i_size.

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