From: Jonathan Matthew <jonathan@d14n.org>
Subject: Re: nvme: use I/O submission queue entry size reported by controller
To: tech@openbsd.org
Cc: jcs@jcs.org
Date: Tue, 19 May 2026 15:14:12 +1000

On Mon, May 18, 2026 at 08:29:47PM -0500, joshua stein wrote:
> On Tue, 19 May 2026 at 10:35:03 +1000, Jonathan Matthew wrote:
> > On Sun, May 17, 2026 at 12:25:57PM -0500, joshua stein wrote:
> > > On at least the Apple T2 NVMe, 128-byte submission queue entries on
> > > I/O queues are required instead of the standard 64 bytes.
> > > 
> > > This gets NVMe working on the 2018 Mac Mini.  Also tested on a 
> > > non-Apple NVMe but more tests would be helpful.
> > 
> > I've tested this on a few different nvme devices (real and virtual) and
> > haven't seen any problems.
> > 
> > I'm curious as to why you're using the maximum supported sqe size rather
> > than the minimum.  Does the apple T2 nvme not report 128 as the minimum
> > size?
> 
> No, it reports 64.
>

How unfortunate.  Even the fanciest nvme devices I can check at the
moment only support 64 byte sqes, so I don't think this will cause any
harm.  At worst, with a device supporting 32kB sqes, we'd allocate 4MB
of memory where we only need 2kB.  Maybe it'd be worth ensuring
identify->sqes >> 4 is >= 6 just to be safe?

ok jmatthew@