Re: [PATCH 7/9] aio: add aio_kernel_() interface

[Ming Lei <ming.lei@canonical.com>]

On Thu, Jul 24, 2014 at 11:16 PM, Dave Kleikamp
<dave.kleikamp@oracle.com> wrote:
> On 07/23/2014 08:57 PM, Ming Lei wrote:
>> On Thu, Jul 24, 2014 at 7:16 AM, Zach Brown <zab@zabbo.net> wrote:
>>> On Thu, Jul 24, 2014 at 06:55:28AM +0800, Ming Lei wrote:
>>>> From: Dave Kleikamp <dave.kleikamp@oracle.com>
>>>>
>>>> This adds an interface that lets kernel callers submit aio iocbs without
>>>> going through the user space syscalls.  This lets kernel callers avoid
>>>> the management limits and overhead of the context.  It will also let us
>>>> integrate aio operations with other kernel apis that the user space
>>>> interface doesn't have access to.
>>>>
>>>> This patch is based on Dave's posts in below links:
>>>>
>>>>       https://lkml.org/lkml/2013/10/16/365
>>>>       https://groups.google.com/forum/#!topic/linux.kernel/l7mogGJZoKQ
>>>
>>> This was originally written a billion years ago when dinosaurs roamed
>>> the earth.  Also, notably, before Kent and Ben reworked a bunch of the
>>
>> Not so far away, this patch is based on Dave's last version of V9, which
>> was posted in Oct, 2013, :-)
>
> Which was based on a much earlier patch from Zach. I regret that I left
> aio_kernel_submit entangled with aio_run_iocb when I reworked his patches.
>
>>> aio core.  I'd want them to take a look at this patch to make sure that
>>> it doesn't rely on any assumptions that have changed.
>>
>> Looks I missed to Cc Ken, :-(
>>
>>>
>>>> +/* opcode values not exposed to user space */
>>>> +enum {
>>>> +     IOCB_CMD_READ_ITER = 0x10000,
>>>> +     IOCB_CMD_WRITE_ITER = 0x10001,
>>>> +};
>>>
>>> And I think the consensus was that this isn't good enough.  Find a way
>>> to encode the kernel caller ops without polluting the uiocb cmd name
>>> space.
>>
>> That is easy, since the two cmd names are only for kernel AIO, whatever
>> should be OK, but looks I didn't see such comment.
>
> Agreed. These were added because the flags had been interpreted by
> aio_run_iocb(). I'm happy that is no longer the case.

We can remove the two cmd names completely, and just use one
read/write flag, will do it in V1.

>
>>>
>>> (I've now come to think that this entire approach of having loop use aio
>>> is misguided and that the way forward is to have dio consume what loop
>>> naturally produces -- bios, blk-mq rqs, whatever -- but I'm on to other
>>
>> Yes, that is what these patches are doing, and actually AIO's
>> model is a good match to driver's interface. Lots of drivers
>> use the asynchronous model(submit, complete, ...).
>>
>>> things these days.)
>>
>> At least, loop can improve its throughput much by kernel AIO
>> without big changes to fs/direct-io(attribute much to ITER_BVEC),
>> and vhost-scsi should benefit from it too.
>>
>> Thanks,
>>