Re: [PATCH v2] blk-throtl: optimize IOPS throttle for large IO scenarios

From: Tejun Heo
Date: Fri Jul 30 2021 - 12:07:11 EST


On Fri, Jul 30, 2021 at 10:09:34AM +0800, brookxu wrote:
> >> @@ -877,10 +900,19 @@ static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
> >> else
> >> tg->bytes_disp[rw] = 0;
> >>
> >> - if (tg->io_disp[rw] >= io_trim)
> >> + if (tg_io_disp(tg, rw) >= io_trim) {
> >
> > Instead of checking this in multiple places, would it be simpler to transfer
> > the atomic counters to the existing counters whenever we enter blk-throtl
> > and leave the rest of the code as-is?
>
> If we do this, we need to do similar processing on the bio submission path and the bio
> resubmission path in pending_timer. It seems that the code is more complicated?

Yeah, basically whenever we enter blk-throtl. Factored to a function,
calling it on entry should be fairly clean, right? I wonder whether it'd be
better to consolidate all atomic counter handling in a single location and
all it does is transferring whatever's accumulated to the usual counters.
Also, when you're reading & resetting the atomic counters, can you use a
pattern like the following?

main_counter += atomic_xchg(counter, 0);

Right now, there's a race window between reading and resetting.

Thanks.

--
tejun