Re: [PATCH] Optimize struct task_delay_info

[Balbir Singh <balbir@linux.vnet.ibm.com>]

Zhang, Yanmin wrote:
> struct task_delay_info is used by per process block I/O delay statistics
> feature which is useful in kernel. This struct is not optimized.
> 
> My patch against kernel 2.6.22 shrinks it a half.
> 
> 1) Delete blkio_start and blkio_end. As the collection happens in
> io_schedule and io_schedule_timeout, we use local variables to
> replace them;

I am not sure if it's a good idea to push items on the stack.
Remember we are moving to 4K stacks.

> 2) Delete lock. The change to the protected data has no nested cases.
> In addition, the result is for performance data collection, so it’s
> unnecessary to add such lock. 

This is a cause of concern, we cannot afford to have incorrect data
collected. Incorrect/unreliable data which is worthless.

> 3) Delete flags. It just has one value. Use the most significant bit of
> blkio_delay (64 bits) to mark it..
> 

Yes, thats true right now, but I am not sure if we should go optimize
that so early. We could end up adding other accounting/extending the
framework, we'll need to add the flags back then.



> -static inline void delayacct_clear_flag(int flag)
> +static inline void delayacct_clear_swapin(void)
>  {
>  	if (current->delays)
> -		current->delays->flags &= ~flag;
> +		current->delays->blkio_delay |= DELAYACCT_PF_SWAPIN;

BTW, you should be clearing the flag here.


Overall, the lock removal is not acceptable. I don't like the bit
hacking for flags and moving counters to the stack either.

-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL