[dm-devel] block: revert to using min_not_zero() when stacking chunk_sectors

[Mike Snitzer]

On Mon, Nov 30 2020 at  7:21pm -0500,
John Dorminy <jdorminy at redhat.com> wrote:

> > If you're going to cherry pick a portion of a commit header please
> > reference the commit id and use quotes or indentation to make it clear
> > what is being referenced, etc.
> Apologies.
> 
> > Quite the tangent just to setup an a toy example of say: thinp with 256K
> > blocksize/chunk_sectors ontop of a RAID6 with a chunk_sectors of 128K
> > and stripesize of 1280K.
> 
> I screwed up my math ... many apologies :/
> 
> Consider a thinp of chunk_sectors 512K atop a RAID6 with chunk_sectors 1280K.
> (Previously, this RAID6 would be disallowed because chunk_sectors
> could only be a power of 2, but 07d098e6bba removed this constraint.)

Think you have your example messed up still.  RAID 10+2 with 128K
chunk_sectors, 1280K full stripe (io_opt). Then thinp stacked ontop of
it with chunk_sectors of 1280K was usecase that wasn't supported before.

So stacked chunk_sectors = min_not_zero(128K, 1280K) = 128K

> -With lcm_not_zero(), a full-device IO would be split into 2560K IOs,
> which obviously spans both 512K and 1280K chunk boundaries.

Sure, think we both agree lcm_not_zero() shouldn't be used.

> -With min_not_zero(), a full-device IO would be split into 512K IOs,
> some of which would span 1280k chunk boundaries. For instance, one IO
> would span from offset 1024K to 1536K.

RAID6 with chunk_sectors of 1280K is pretty insane...
And yet you're saying full device IO is 1280K...
So something still isn't adding up.

Anyway, if we run with your example of chunk_sectors (512K, 1280K), yes
there is serious potential for IO to span the RAID6 layer's chunk_sector
boundary.

> -With the hypothetical gcd_not_zero(), a full-device IO would be split
> into 256K IOs, which span neither 512K nor 1280K chunk boundaries.

Yeap, I see.

> > To be clear, you are _not_ saying using lcm_not_zero() is correct.
> > You're saying that simply reverting block core back to using
> > min_not_zero() may not be as good as using gcd().
> 
> Assuming my understanding of chunk_sectors is correct -- which as per
> blk-settings.c seems to be "a driver will not receive a bio that spans
> a chunk_sector boundary, except in single-page cases" -- I believe
> using lcm_not_zero() and min_not_zero() can both violate this
> requirement. The current lcm_not_zero() is not correct, but also
> reverting block core back to using min_not_zero() leaves edge cases as
> above.

But your chunk_sectors (512K, 1280K) example is a misconfigured IO
stack.  Really not sure it worth being concerned about it.

> I believe gcd provides the requirement, but min_not_zero() +
> disallowing non-power-of-2 chunk_sectors also provides the
> requirement.

Kind of on the fence on this... think I'd like to get Martin's take.

Using gcd() instead of min_not_zero() to stack chunk_sectors isn't a big
deal; given the nature of chunk_sectors coupled with it being able to be
a non-power-of-2 _does_ add a new wrinkle.

So you had a valid point all along, just that you made me work pretty
hard to understand you.

> > > But it's possible I'm misunderstanding the purpose of chunk_sectors,
> > > or there should be a check that the one of the two devices' chunk
> > > sizes divides the other.
> >
> > Seriously not amused by your response, I now have to do damage control
> > because you have a concern that you really weren't able to communicate
> > very effectively.
> 
> Apologies.

Eh, I need to build my pain threshold back up.. been away from it all
for more than a week.. ;)

Mike