Re: [PATCH] blk-throttle: support io merge over iops_limit

[Ming Lei]

On Mon, Mar 10, 2025 at 09:58:57AM +0800, Yu Kuai wrote:
> Hi,
> 
> 在 2025/03/08 0:07, Tejun Heo 写道:
> > Hello,
> > 
> > On Fri, Mar 07, 2025 at 05:01:52PM +0800, Yu Kuai wrote:
> > > From: Yu Kuai <yukuai3@xxxxxxxxxx>
> > > 
> > > Commit 9f5ede3c01f9 ("block: throttle split bio in case of iops limit")
> > > support to account split IO for iops limit, because block layer provides
> > > io accounting against split bio.
> > > 
> > > However, io merge is still not handled, while block layer doesn't
> > > account merged io for iops. Fix this problem by decreasing io_disp
> > > if bio is merged, and following IO can use the extra budget. If io merge
> > > concurrent with iops throttling, it's not handled if one more or one
> > > less bio is dispatched, this is fine because as long as new slice is not
> > > started, blk-throttle already preserve one extra slice for deviation,
> > > and it's not worth it to handle the case that iops_limit rate is less than
> > > one per slice.
> > > 
> > > A regression test will be added for this case [1], before this patch,
> > > the test will fail:
> > > 
> > > +++ /root/blktests-mainline/results/nodev/throtl/007.out.bad
> > > @@ -1,4 +1,4 @@
> > >   Running throtl/007
> > >   1
> > > -1
> > > +11
> > >   Test complete
> > > 
> > > [1] https://lore.kernel.org/all/20250307080318.3860858-2-yukuai1@xxxxxxxxxxxxxxx/
> > 
> > For blk-throtl, iops limit has meant the number of bios issued. I'm not
> 
> Yes, but it's a litter hard to explain to users the differece between IO
> split and IO merge, they just think IO split is the numer of IOs issued
> to disk, and IO merge is the number of IOs issued from user.

Here it is really one trouble.

a) Sometimes people think IOs wrt. IOPS means that the read/write IO
submitted from application, one typical example is `fio`.

b) Sometimes people think it is the data observed from `iostat`.

In a), io merge/split isn't taken into account, but b) does cover io
merge and split.

So question is that what is the correct way for user to observe IOPS
wrt. iops throttle?


Thanks,
Ming