Re: [PATCH] Retry commands with UNIT_ATTENTION sense codes

[Christoph Hellwig <hch@infradead.org>]

On Wed, May 05, 2010 at 09:19:53AM -0400, James Bottomley wrote:
> Unfortunately, not.  A long time ago we used to use SPECIAL for prepared
> commands, but we stripped that path out.  If you look at the prep
> functions today, they only accept either REQ_TYPE_FS or
> REQ_TYPE_BLOCK_PC ... they kill everything else, so we'd have to do
> surgery on all the input paths to get this to work.
> 
> Looking at all of this, I'm really unhappy with the way barriers are
> working in SCSI ... using BLOCK_PC for them was a big mistake, since
> block can't handle the errors and BLOCK_PC by its very nature requests
> that all errors, however trivially fixable, be returned.
> 
> It looks like we have precisely the same problems with discard too,
> except there at least we'll ignore the error.
> 
> Someone needs to take a good look at fixing this ... or at least just
> making discard and barrier use the same submit paths, so we can fix it
> up in the lower layers.

Both barriers and dicards are FS requests as far as the block layer is
concerned.  We turn them into BLOCK_PC requests very late in the game,
either in the prepare_flush function for barriers, or directly inside
prep_fn for discard.  I tried to handle discard requests as FS all
the way through, but the various different ways in which we account for
a request length didn't play nicely with that.

Handling cache flushes as FS requests all the way through should be
easier than that.