From mboxrd@z Thu Jan  1 00:00:00 1970
From: willy@linux.intel.com (Matthew Wilcox)
Date: Mon, 24 Jun 2013 23:01:40 -0400
Subject: RFC: Allow block drivers to poll for I/O instead of sleeping
In-Reply-To: <20130624071544.GR9422@kernel.dk>
References: <20130620201713.GV8211@linux.intel.com>
 <20130623100920.GA19021@gmail.com>
 <20130624071544.GR9422@kernel.dk>
Message-ID: <20130625030140.GZ8211@linux.intel.com>

On Mon, Jun 24, 2013@09:15:45AM +0200, Jens Axboe wrote:
> Willy, I think the general design is fine, hooking in via the bdi is the
> only way to get back to the right place from where you need to sleep.
> Some thoughts:
> 
> - This should be hooked in via blk-iopoll, both of them should call into
>   the same driver hook for polling completions.

I actually started working on this, then I realised that it's actually
a bad idea.  blk-iopoll's poll function is to poll the single I/O queue
closest to this CPU.  The iowait poll function is to poll all queues
that the I/O for this address_space might complete on.

I'm reluctant to ask drivers to define two poll functions, but I'm even
more reluctant to ask them to define one function with two purposes.

> - It needs to be more intelligent in when you want to poll and when you
>   want regular irq driven IO.

Oh yeah, absolutely.  While the example patch didn't show it, I wouldn't
enable it for all NVMe devices; only ones with sufficiently low latency.
There's also the ability for the driver to look at the number of
outstanding I/Os and return an error (eg -EBUSY) to stop spinning.

> - With the former note, the app either needs to opt in (and hence
>   willingly sacrifice CPU cycles of its scheduling slice) or it needs to
>   be nicer in when it gives up and goes back to irq driven IO.

Yup.  I like the way you framed it.  If the task *wants* to spend its
CPU cycles on polling for I/O instead of giving up the remainder of its
time slice, then it should be able to do that.  After all, it already can;
it can submit an I/O request via AIO, and then call io_getevents in a
tight loop.

So maybe the right way to do this is with a task flag?  If we go that
route, I'd like to further develop this option to allow I/Os to be
designated as "low latency" vs "normal".  Taking a page fault would be
"low latency" for all tasks, not just ones that choose to spin for I/O.