Re: 2.6.4-mm2

[markw@osdl.org]

On 18 Mar, Jens Axboe wrote:
> On Thu, Mar 18 2004, Andrew Morton wrote:
>> > Comparing one pair of readprofile results, I find it curious that
>> > dm_table_unplug_all and dm_table_any_congested show up near the top of a
>> > 2.6.4-mm2 profile when they haven't shown up before in 2.6.3.
>> 
>> 14015190 poll_idle                                241641.2069
>> 175162 generic_unplug_device                    1317.0075
>> 165480 __copy_from_user_ll                      1272.9231
>> 161151 __copy_to_user_ll                        1342.9250
>> 152106 schedule                                  85.0705
>> 142395 DAC960_LP_InterruptHandler               761.4706
>> 113677 dm_table_unplug_all                      1386.3049
>>  65420 __make_request                            45.5571
>>  64832 dm_table_any_congested                   697.1183
>>  37913 try_to_wake_up                            32.2939
>> 
>> That's broken.  How many disks are involve in the DM stack?
>> 
>> The relevant code was reworked subsequent to 2.6.4-mm2.  Maybe we fixed
>> this, but I cannot immediately explain what you're seeing here.
> 
> Ugh that looks really bad, I wonder how it could possibly ever be this
> bad. Mark, please do do a run with 2.6.5-rc1-mm2, I'd very much like to
> see the profile there. If things get this bad, I need to think some more
> about how to best handle the 'when to invoke request_fn on unplug calls'
> logic again.
> 
> Actually, please also do a run with 2.6.5-rc1-mm2 + inlined patch. For
> non-stacked dm on dm it should work and could make a lot of difference
> for you.
> 
> --- drivers/block/ll_rw_blk.c~	2004-03-18 20:26:17.088531084 +0100
> +++ drivers/block/ll_rw_blk.c	2004-03-18 20:26:44.773554953 +0100
> @@ -1134,11 +1134,8 @@
>  	if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
>  		return;
>  
> -	/*
> -	 * always call down, since we can race now with setting the plugged
> -	 * bit outside of the queue lock
> -	 */
> -	blk_remove_plug(q);
> +	if (!blk_remove_plug(q))
> +		return;
>  
>  	/*
>  	 * was plugged, fire request_fn if queue has stuff to do
> 

http://developer.osdl.org/markw/dbt2-pgsql/409/

There are the results with 2.6.5-rc1-mm2 with your patch with a 512kb
stripe width.  The throughput didn't improve but it did cut the ticks
for dm_table_unplug_all in half.  Should I continue with smaller stripe
widths?