On Fri, May 17, 2024 at 03:48:49PM +0800, YangYang wrote:
> On 2024/5/16 23:29, Benjamin Marzinski wrote:
> > On Thu, May 16, 2024 at 09:55:53AM +0800, YangYang wrote:
> > > On 2024/5/15 23:42, Benjamin Marzinski wrote:
> > > > On Tue, May 14, 2024 at 05:04:42PM +0800, Yang Yang wrote:
> > >
> > > If I understand correctly, you are suggesting to iterate through all the
> > > targets, handling those with sends_pass_around_flush set, and skipping
> > > those where sends_pass_around_flush is not set. I believe this approach
> > > may result in some CPU wastage.
> > >
> > > for i in {0..1023}; do
> > > echo $((8000*$i)) 8000 linear /dev/sda2 $((16384*$i))
> > > done | sudo dmsetup create example
> > >
> > > In this specific scenario, a single iteration of the loop is all that
> > > is needed.
> >
> > It's just one iteration of the loop either way. You either loop through
> > the targets or the devices. It's true that if you have lots of targets
> > all mapped to the same device, you would waste time looping through all
> > the targets instead of looping through the devices. But if you only had
> > one striped target mapped to lots of devices, you would waste time
> > looping through all of the devices instead of looping through the
> > targets.
>
> Yes, I get your point. This patchset may make things even worse for
> the striped target.
> I am just curious, in what scenario is the "dm-strip" target mapped to
> a large number of underlying devices from the same block device.
>
I don't think anyone in the real world does create dm-stripe devices with a
huge number of stripe table devices. My point was that it didn't seem
obvious me that looping through the targets was a significant problem
compared to looping through the devices.
At any rate, Mikulas's patch already does this optimally, even for
targets like dm-stripe, so it doesn't really matter now.
-Ben
