在 2022/05/20 0:10, Michal Koutný 写道:

On Thu, May 19, 2022 at 08:14:28PM +0800, "yukuai (C)" [off-list ref] wrote:

quoted

tg_with_in_bps_limit:
  jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
  tmp = bps_limit * jiffy_elapsed_rnd;
  do_div(tmp, HZ);
  bytes_allowed = tmp; -> how many bytes are allowed in this slice,
		         incluing dispatched.
  if (tg->bytes_disp[rw] + bio_size <= bytes_allowed)
   *wait = 0 -> no need to wait if this bio is within limit

  extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
  -> extra_bytes is based on 'bytes_disp'

For example:

1) bps_limit is 2k, we issue two io, (1k and 9k)
2) the first io(1k) will be dispatched, bytes_disp = 1k, slice_start = 0
    the second io(9k) is waiting for (9 - (2 - 1)) / 2 = 4 s

The 2nd io arrived at 1s, the wait time is 4s, i.e. it can be dispatched
at 5s (i.e. 10k/*2kB/s = 5s).

No, the example is that the second io arrived together with first io.

quoted

3) after 3 s, we update bps_limit to 1k, then new waiting is caculated:

without this patch:  bytes_disp = 0, slict_start =3:
bytes_allowed = 1k	                            <--- why 1k and not 0?

Because slice_start == jiffies, bytes_allowed is equal to bps_limit

quoted

extra_bytes = 9k - 1k = 8k
wait = 8s

This looks like it was calculated at time 4s (1s after new config was
set).

No... it was caculated at time 3s:

jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);

jiffies should be greater than 3s here, thus jiffy_elapsed_rnd is
3s + throtl_slice (I'm using throtl_slice = 1s here, it should not
affect result)

quoted

whth this patch: bytes_disp = 0.5k, slice_start =  0,
bytes_allowed = 1k * 3 + 1k = 4k
extra_bytes =  0.5k + 9k - 4k = 5.5k
wait = 5.5s

This looks like calculated at 4s, so the IO would be waiting till
4s+5.5s = 9.5s.

wait time is based on extra_bytes, this is really 5.5s, add 4s is
wrong here.

bytes_allowed = ((jiffies - slice_start) / Hz + 1) * bps_limit
extra_bytes = bio_size + bytes_disp - bytes_allowed
wait = extra_bytes / bps_limit

As I don't know why using time 4s, I'll shift this calculation to the
time 3s (when the config changes):

bytes_disp = 0.5k, slice_start =  0,
bytes_allowed = 1k * 3  = 3k
extra_bytes =  0.5k + 9k - 3k = 7.5k

6.5k

wait = 7.5s

In absolute time, the IO would wait till 3s+7.5s = 10.5s

Like I said above, wait time should not add (jiffies - slice_start)

OK, either your 9.5s or my 10.5s looks weird (although earlier than
original 4s+8s=12s).
However, the IO should ideally only wait till

     3s + (9k -   (6k    -    1k)     ) / 1k/s =
          bio - (allowed - dispatched)  / new_limit

    =3s + 4k / 1k/s = 7s

    ('allowed' is based on old limit)

Or in another example, what if you change the config from 2k/s to ∞k/s
(unlimited, let's neglect the arithmetic overflow that you handle
explicitly, imagine a big number but not so big to be greater than
division result).

In such a case, the wait time should be zero, i.e. IO should be
dispatched right at the time of config change.

I thought about it, however, IMO, this is not a good idea. If user
updated config quite frequently, io throttle will be invalid.

Thanks,
Kuai

(With your patch that still calculates >0 wait time (and the original
behavior gives >0 wait too.)

quoted

I hope I can expliain it clearly...

Yes, thanks for pointing me to relevant parts.
I hope I grasped them correctly.

IOW, your patch and formula make the wait time shorter but still IO can
be delayed indefinitely if you pass a sequence of new configs. (AFAIU)

Regards,
Michal
.