Re: [patch] use cheaper elv_queue_empty when unplug a device

[Nick Piggin <nickpiggin@yahoo.com.au>]

Arjan van de Ven wrote:
> On Tue, 2005-03-29 at 19:19 +1000, Nick Piggin wrote:
> 
>>- removes the relock/retry merge mechanism in __make_request if we
>>   aren't able to get the GFP_ATOMIC allocation. Just fall through
>>   and assume the chances of getting a merge will be small (is this
>>   a valid assumption? Should measure it I guess).
> 
> 
> this may have a potential problem; if the vm gets in trouble, you
> suddenly start to generate worse IO patterns, which means IO performance
> goes down right when it's most needed.....
> 

Sorry my wording was incorrect. It currently *always* retries the
merge if it had at first failed, and after the patch, we never retry.
So it should not result in behavioural shifts when there is a VM load
is high.

It seems to be a clear source of problems for Kenneth though, because
his workload appears to have almost zero merges, so he'll always be
invoking the merge logic twice.

I agree there is potential for subtle interactions. But generally the
block layer is surprisingly well behaved in my experience.

As Jens said, the complete removal of the GFP_ATOMIC allocation probably
has the most potential for problems in this regard, although bios are not
using GFP_ATOMIC allocations, so I would be a little surprised if it made
a really noticable difference.