Re: [RFC PATCH v2 15/21] loop: use aio to perform io on the underlying file

[Jeff Moyer <jmoyer@redhat.com>]

Zach Brown <zab@zabbo.net> writes:

> On 04/20/2012 11:20 AM, Jeff Moyer wrote:
>> Dave Kleikamp<dave.kleikamp@oracle.com>  writes:
>>
>>> On 04/20/2012 09:48 AM, Maxim V. Patlasov wrote:
>>>> On 03/30/2012 07:43 PM, Dave Kleikamp wrote:
>>>>> From: Zach Brown<zab@zabbo.net>
>>>>>
>>>>> This uses the new kernel aio interface to process loopback IO by
>>>>> submitting concurrent direct aio.  Previously loop's IO was serialized
>>>>> by synchronous processing in a thread.
>>>>>
>>>>
>>>> The patch ignores REQ_FLUSH bit of bi_rw. Is it simply overlook?
>>>
>>> Good question. Since the loop device is sending only direct IO requests,
>>> it shouldn't be necessary to explicitly flush page cache, but REQ_FLUSH
>>
>> REQ_FLUSH isn't about the page cache, it's about flushing the volatile
>> disk write cache.  You need to handle that.
>
> I guess O_DIRECT doesn't routinely issue flushes simply because it's too
> expensive?  

Bypassing the page cache is different from bypassing the underlying
device's cache.  O_DIRECT does not mean "straight to platter".

> Apps that care about consistent IO and O_DIRECT are expected to not
> have writeback caching enabled?  'cause there's no way they're issuing
> syncs themselves.

They most certainly should be!  The app should be written with the
assumption that there is a write-back cache on the storage.  Turning
those flushes into noops is an optimization the OS performs.  See this
lwn article: http://lwn.net/Articles/457667/.

Cheers,
Jeff