On 4/12/25 5:56 AM, Benjamin Marzinski wrote:
> When using a kthread to delay the IOs, dm-delay would continuously loop,
> checking if IOs were ready to submit. It had a cond_resched() call in
> the loop, but might still loop hundreds of millions of times waiting for
> an IO that was scheduled to be submitted 10s of ms in the future. With
> the change to make dm-delay over zoned devices always use kthreads
> regardless of the length of the delay, this wasted work only gets worse.
>
> To solve this and still keep roughly the same precision for very short
> delays, dm-delay now calls fsleep() for 1/8th of the smallest non-zero
> delay it will place on IOs, or 1 ms, whichever is smaller. The reason
> that dm-delay doesn't just use the actual expiration time of the next
> delayed IO to calculated the sleep time is that delay_dtr() must wait
> for the kthread to finish before deleting the table. If a zoned device
> with a long delay queued an IO shortly before being suspended and
> removed, the IO would be flushed in delay_presuspend(), but the removing
> the device would still have to wait for the remainder of the long delay.
> This time is now capped at 1 ms.
>
> Fixes: 70bbeb29fab09 ("dm delay: for short delays, use kthread instead of
> timers and wq")
> Signed-off-by: Benjamin Marzinski <bmarz...@redhat.com>
Nice !
Reviewed-by: Damien Le Moal <dlem...@kernel.org>
--
Damien Le Moal
Western Digital Research
