Re: [PATCH] cfq-iosched: don't allow aliased requests to starve others

[Jens Axboe <axboe@kernel.dk>]

On 2010-07-26 15:17, Jeff Moyer wrote:
> Jens Axboe <axboe@kernel.dk> writes:
> 
>> On 07/24/2010 10:04 AM, Heinz Diehl wrote:
>>> On 14.07.2010, Jeff Moyer wrote: 
>>>
>>>> Comments, as always, are welcome.
>>>
>>> This patch, applied to 2.6.35-rc6, increases desktop interactivity
>>> _NOTICEABLY_ on my quadcore machine, and the machine stays rock-stable.
>>> I have now tested this patch with the latest 2.6.35-rc kernels over 
>>> 1 week.
>>>
>>> Unfortunately, I can't provide some testing results which makes this
>>> statement more objective, but I'll do some synthetic testing in the next
>>> days.
>>
>> It is extremely unlikely that this patch will have any impact on
>> "normal" workloads. To even hit a code path where it would make a
>> difference, you would need to use O_DIRECT IO, otherwise you cannot have
>> aliases in the IO scheduler.
> 
> I agree that it shouldn't help normal workloads at all.  I do think
> there is one other case where you can get aliases: doing I/O both
> through the file system and the underlying device.  However, that's
> obviously a bad idea (and maybe open_bdev_exclusive will keep that from
> happening?).

That's correct, you could construct such a test case since you would
get page cache synchronization from different mappings. But again,
not something that the casual user would run into :-)

Exclusive opens only guard against each other, not against "normal"
opens.

-- 
Jens Axboe