blk-throttle: fix lower bps rate by throtl_trim_slice()

The bio submission time may be a few jiffies more than the expected
waiting time, due to 'extra_bytes' can't be divided in
tg_within_bps_limit(), and also due to timer wakeup delay.
In this case, adjust slice_start to jiffies will discard the extra wait time,
causing lower rate than expected.

Current in-tree code already covers deviation by rounddown(), but turns
out it is not enough, because jiffies - slice_start can be a multiple of
throtl_slice.

For example, assume bps_limit is 1000bytes, 1 jiffes is 10ms, and
slice is 20ms(2 jiffies), expected rate is 1000 / 1000 * 20 = 20 bytes
per slice.

If user issues two 21 bytes IO, then wait time will be 30ms for the
first IO:

bytes_allowed = 20, extra_bytes = 1;
jiffy_wait = 1 + 2 = 3 jiffies

and consider
extra 1 jiffies by timer, throtl_trim_slice() will be called at:

jiffies = 40ms
slice_start = 0ms, slice_end= 40ms
bytes_disp = 21

In this case, before the patch, real rate in the first two slices is
10.5 bytes per slice, and slice will be updated to:

jiffies = 40ms
slice_start = 40ms, slice_end = 60ms,
bytes_disp = 0;

Hence the second IO will have to wait another 30ms;

With the patch, the real rate in the first slice is 20 bytes per slice,
which is the same as expected, and slice will be updated:

jiffies=40ms,
slice_start = 20ms, slice_end = 60ms,
bytes_disp = 1;

And now, there is still 19 bytes allowed in the second slice, and the
second IO will only have to wait 10ms;

This problem will cause blktests throtl/001 failure in case of
CONFIG_HZ_100=y, fix it by preserving one extra finished slice in
throtl_trim_slice().

Fixes: e43473b7f223 ("blkio: Core implementation of throttle policy")
Reported-by: Ming Lei <ming.lei@redhat.com>
Closes: https://lore.kernel.org/linux-block/20250222092823.210318-3-yukuai1@huaweicloud.com/
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Link: https://lore.kernel.org/r/20250227120645.812815-1-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

diff --git a/block/blk-throttle.c b/block/blk-throttle.c
index 8d149aff9fd0b7fe23d51116b2353b6cf41801e1..a52f0d6b40ad4ee4aeff589d538e424999ee05c1 100644 (file)
--- a/block/blk-throttle.c
+++ b/block/blk-throttle.c
@@ -599,14 +599,23 @@ static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
         * sooner, then we need to reduce slice_end. A high bogus slice_end
         * is bad because it does not allow new slice to start.
         */
-
        throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
 
        time_elapsed = rounddown(jiffies - tg->slice_start[rw],
                                 tg->td->throtl_slice);
-       if (!time_elapsed)
+       /* Don't trim slice until at least 2 slices are used */
+       if (time_elapsed < tg->td->throtl_slice * 2)
                return;
 
+       /*
+        * The bio submission time may be a few jiffies more than the expected
+        * waiting time, due to 'extra_bytes' can't be divided in
+        * tg_within_bps_limit(), and also due to timer wakeup delay. In this
+        * case, adjust slice_start will discard the extra wait time, causing
+        * lower rate than expected. Therefore, other than the above rounddown,
+        * one extra slice is preserved for deviation.
+        */
+       time_elapsed -= tg->td->throtl_slice;
        bytes_trim = calculate_bytes_allowed(tg_bps_limit(tg, rw),
                                             time_elapsed) +
                     tg->carryover_bytes[rw];