Re: Call for testers with NVMe

Hello,

Abdelkader Boudih <freebsd@seuros.com> escreveu (quarta, 2/09/2026 à(s)
14:46):

>
> So it a stack
>
> https://reviews.freebsd.org/D59151 and https://reviews.freebsd.org/D59152
> are bugs that need to be backported
>
> They cause a wedge/stall under certain conditions
>
> The next 2 53/54 are for all NVMe's, 54 trigger in controllers that
> advertise a non zero MAXCMD.
>
> Then 55 this is the one that touches every brand. It programs the
> Arbitration feature from Identify RAB on all
> non-quirked controllers. Both Apple controllers have QUIRK_IGNORE_RAB
> because they wedged under load.
>
> 56 are just Apple dequirkage for T1 drives...  When i fixed the Apple T2
> Nvme in D57087, it broke T1 (that used to work), so i had to patch it and
> depatch it.
>
> So if you want benchmark, you need to test between 15.1 and 16-Current
> with these patches.
>
> I observed the improvement across multiple setup, most consumer drive were
> ok.
>
> run 'sudo nvmecontrol identify nvme0'
>
> You will see something like 'Recommended Arb Burst:       2'  <-- here 2
>
> From 0-7 (except 4 (i dont have 4)) i tested all.  6-7  are the top
> performers with the new code. Anything over it is not tested but it already
> clamped. (Also my tests were just copying a lot of files across the drive)
>
> So in short this stack is 2 bug fixes, 2 conformance commits and 2 fixes
> for apple nvme that don't respect the spec.
>
> Best.
> Abdelkader
>
>
Do you have any sugestion about what tool to test before and after nvme
patch so a comparison could be done?

dd, fio, bonnie, etc, It would be nice that people share his experience
using same tools.

Thanks for your work,

-- 
Nuno Teixeira
FreeBSD UNIX:  <eduardo@FreeBSD.org>   Web:  https://FreeBSD.org