Re: flush cache range proposal (was Re: ide errors in 7-rc1-mm1 and later)

[Jens Axboe <axboe@suse.de>]

On Fri, Jun 11 2004, Eric D. Mudama wrote:
> On Fri, Jun 11 at 12:31, Jeff Garzik wrote:
> >If queued-FUA is out of the question, this seems quite reasonable.  It 
> >appears to achieve the commit-block semantics described for barrier 
> >operation, AFAICS.
> 
> Queued FUA shouldn't be out of the question.
> 
> However, Queued FUA requires waiting for the queue to drain before
> sending more commands, since a pair of queued FUA commands doesn't
> guarantee the ordering of those two commands, which may or may not be
> acceptable semantics.

You can continue building and reordering requests behind the QUEUED_FUA
write(s).

> The barrier operation is basically a queueing-friendly flush+FUA,
> which may be better...  it lets the driver keep the queue in the drive

That's exactly correct.

> full, and also allows writes other than the commit block to not be
> done as FUA operations, which is potentially faster.  THe bigger the
> ratio of data to commit block, the better the performance would be
> with a barrier operation vs a purely queued FUA workload.

Just looking at how pre/write/post flush performs and I don't think it
will be that bad (it's already quite good). Depends on how sync
intensive the workload is of course.

But as long as it's the fastest possible implementation (and I think it
is), then arguing about performance is futile imo. Correctness comes
first.

-- 
Jens Axboe