On Mon, Apr 14, 2025 at 12:53:15PM -0400, Benjamin Marzinski wrote:
> On Mon, Apr 14, 2025 at 03:52:18PM +0200, Mikulas Patocka wrote:
> >
> >
> > On Fri, 11 Apr 2025, 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>
> > > ---
> > > This patch is meant to apply on top of Damien Le Moal's "dm-delay:
> > > Prevent zoned write reordering on suspend" patch. If people think it's
> > > important to avoid either this much smaller amount of looping or the
> > > possible 1 ms delay on deleting a table, I can send a patch that uses
> > > usleep_range_state() and msleep_interruptible() to do an interruptible
> > > sleep with a duration based on the expiration time of the next delayed
> > > IO.
> >
> > Hi
> >
> > worker_sleep_ns should be worker_sleep_us - as the value is in
> > microseconds.
>
> Oops.
>
> > fsleep in flush_worker_fn should be called unconditionally, to not consume
> > 100% CPU when suspending.
>
> This is fine, but flush_worker_fn() won't busy-wait while suspending.
> Since it is calling flush_delayed_bios() with flush_all equal to true,
> it will handle all the bios on the list. As long as bios keep getting
> added to the list while it's flushing the last batch, it will keep
> looping to flush them. But it will be doing necessary work on each loop,
> and flush_delayed_bios() has a cond_resched(), so even if there are a
> flood of bios, it will still take breaks. As soon as bios stop
> continuously arriving, flush_worker_fn() will see the empty list and the
> kthread will stop.
>
> Unconditionally sleeping here makes it more likely that dm-delay will
> end up sleeping unnecessarily while a there are just a few remaining
> bios on the list. Like I said in my commit message, this sleep will be
> capped at 1 ms, so it's not that big of a deal. But it's a trade-off.
> I'm o.k. with your version. I just not sure that it is the better
> trade-off. Perhaps I'm overlooking something.
And the thing that I overlooked was you NAK'ing Damien's patch. Oops.
Please ignore this.
-Ben
> > cond_resched() shouldn't be removed because fsleep may fall back to
> > udelay.
>
> Again, your version is fine, but I'm not sure that cond_resched() was
> ever necessary, since there already is one in flush_delayed_bios().
> Also, at least the way it's currently coded, fsleep() will only resort
> to busy-waiting when the delay is 10 us or less, and the shortest it can
> be with this code is 62 us, so I don't think this cond_resched() will
> ever do anything.
>
> > The patch should increase target version.
> >
> > I fixed the patch so that it applies on the top Linus' tree and applied
> > it to the linux-dm tree.
> >
> > BTW. do we need to backport this to the stable kernels? I think not, but
> > if you have some reason why should we backport it, explain it.
>
> dm-delay is basically a testing target, so I agree that it seems
> unnecessary to backport this.
>
> -Ben
>
> >
> > Mikulas
>
