Re: [PATCH V8 00/33] loop: Issue O_DIRECT aio using bio_vec

[Dave Kleikamp <dave.kleikamp@oracle.com>]

On 08/21/2013 11:39 AM, Benjamin LaHaise wrote:
> On Wed, Aug 21, 2013 at 11:30:22AM -0500, Dave Kleikamp wrote:
>> Ben,
>> First, let me apologize for neglecting to copy you and linux-aio on the
>> applicable patches. I've been carrying along this patchset, assuming I
>> had gotten the proper cc's correct a while back, but I somehow missed
>> the aio pieces.
> 
> Thanks.  Let's figure out how to tackle this best.
> 
>> On 08/21/2013 08:02 AM, Benjamin LaHaise wrote:
> ...
>>> First off, have you tested that this series actually works when merged with 
>>> the pending AIO changes from Kent?  There a git tree with those pending 
>>> changes at git://git.kvack.org/~bcrl/aio-next.git , and they're in 
>>> linux-next.
>>
>> I've lightly tested the patchset against the linux-next tree, running a
>> fio job on loop-mounted filesystems of different fs types.
> 
> Good to hear.
>>
>>> One of the major problems your changeset continues to carry is that your 
>>> new read_iter/write_iter operations permit blocking (implicitely), which 
>>> really isn't what we want for aio.  If you're going to introduce a new api, 
>>> it should be made non-blocking, and enforce that non-blocking requirement 
>>> (ie warn when read_iter/write_iter methods perform blockin operations, 
>>> similar to the warnings when scheduling in atomic mode).  This means more 
>>> changes for some filesystem code involved, something that people have been 
>>> avoiding for years, but which really needs to be done.
>>
>> I'm not really sure how the read_iter and write_iter operations are more
>> likely to block than the current aio_read and aio_write operations. Am I
>> missing something?
> 
> What you say is true, however, my point is more that it will be far easier 
> to fix this issue by making it a hard constraint of a new API than it is 
> to do a system-wide retrofit.  You're converting code over to use the new 
> API one by one, so adding a little bit more work to try and finally sort 
> out this issue while making those conversions would be vrey helpful.
> 
> I'm not saying that you should be required to write the code to cope with 
> this additional requirement (I'm perfectly happy to help with that, and 
> can probably get some time for that at $work), but more that if we're 
> going to be changing all of the filesystems, we might as well try to get 
> things right.

I don't really intend to make the patchset any more complicated than it
already is. The read/write_iter operations are intended to be as near a
replacement as possible to aio_read/write with the added ability to deal
with both kernel and user pages. A completely non-blocking interface
would be great, but that's a bit of work I'd rather not have to wait
for. Maybe that requirement can be added later.