Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Thu, 18 Jul 2019 17:36:55 -0700
From:   Bjorn Andersson <bjorn.andersson@linaro.org>
To:     Benjamin LaHaise <ben@communityfibre.ca>
Cc:     Al Viro <viro@zeniv.linux.org.uk>, linux-fsdevel@vger.kernel.org,
        linux-aio@kvack.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] aio: Support read/write with non-iter file-ops
Message-ID: <20190719003655.GO30636@minitux>
References: <20190718231054.8175-1-bjorn.andersson@linaro.org>
 <20190718231751.GV17978@ZenIV.linux.org.uk>
 <20190718234352.GN30636@minitux>
 <20190718235616.GM29731@kvack.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20190718235616.GM29731@kvack.org>
User-Agent: Mutt/1.12.0 (2019-05-25)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu 18 Jul 16:56 PDT 2019, Benjamin LaHaise wrote:

> On Thu, Jul 18, 2019 at 04:43:52PM -0700, Bjorn Andersson wrote:
> > On Thu 18 Jul 16:17 PDT 2019, Al Viro wrote:
> > 
> > > On Thu, Jul 18, 2019 at 04:10:54PM -0700, Bjorn Andersson wrote:
> > > > Implement a wrapper for aio_read()/write() to allow async IO on files
> > > > not implementing the iter version of read/write, such as sysfs. This
> > > > mimics how readv/writev uses non-iter ops in do_loop_readv_writev().
> > > 
> > > IDGI.  How would that IO manage to be async?  And what's the point
> > > using aio in such situations in the first place?
> > 
> > The point is that an application using aio to submit io operations on a
> > set of files, can use the same mechanism to read/write files that
> > happens to be implemented by driver only implementing read/write (not
> > read_iter/write_iter) in the registered file_operations struct, such as
> > kernfs.
> > 
> > In this particular case I have a sysfs file that is accessing hardware
> > and hence will block for a while and using this patch I can io_submit()
> > a write and handle the completion of this in my normal event loop.
> > 
> > 
> > Each individual io operation will be just as synchronous as the current
> > iter-based mechanism - for the drivers that implement that.
> 
> Just adding the fops is not enough.  I have patches floating around at
> Solace that add thread based fallbacks for files that don't have an aio
> read / write implementation, but I'm not working on that code any more.

My bad. Took another look and now I see the bigger picture of how this
is currently implemented and why just adding the fops would defeat the
purpose of the api.

Sorry for the noise.

> The thread based methods were quite useful in applications that had a need
> for using other kernel infrastructure in their main event loops.
> 

Yes indeed.

Regards,
Bjorn