S3 regression on AMD in 4.20 (was: Re: [PATCH] ci: add yet another HW runner)

["Marek Marczykowski-Górecki" <marmarek@invisiblethingslab.com>]

[-- Attachment #1: Type: text/plain, Size: 2413 bytes --]

On Fri, Mar 14, 2025 at 02:19:19PM -0700, Stefano Stabellini wrote:
> On Fri, 14 Mar 2025, Marek Marczykowski-Górecki wrote:
> > This is AMD Zen2 (Ryzen 5 4500U specifically), in a HP Probook 445 G7.
> > 
> > This one has working S3, so add a test for it here.
> > 
> > Signed-off-by: Marek Marczykowski-Górecki <marmarek@invisiblethingslab.com>
> > ---
> > Cc: Jan Beulich <jbeulich@suse.com>
> > Cc: Andrew Cooper <andrew.cooper3@citrix.com>
> > 
> > The suspend test added here currently fails on staging[1], but passes on
> > staging-4.19[2]. So the regression wants fixing before committing this
> > patch.
> >
> > [1] https://gitlab.com/xen-project/people/marmarek/xen/-/jobs/9408437140
> > [2] https://gitlab.com/xen-project/people/marmarek/xen/-/jobs/9408943441
> 
> We could commit the patch now without the s3 test.
> 
> I don't know what the x86 maintainers think about fixing the suspend
> bug, but one idea would be to run a bisection between 4.20 and 4.19.

I'm on it already, but it's annoying. Lets convert this thread to
discussion about the issue:

So, I bisected it between staging-4.19 and master. The breakage is
somewhere between (inclusive):
eb21ce14d709 x86/boot: Rewrite EFI/MBI2 code partly in C
and
47990ecef286 x86/boot: Improve MBI2 structure check

But, the first one breaks booting on this system and it remains broken
until the second commit (or its parent) - at which point S3 is already
broken. So, there is a range of 71 commits that may be responsible...

But then, based on a matrix chat and Jan's observation I've tried
reverting f75780d26b2f "xen: move per-cpu area management into common
code" just on top of 47990ecef286, and that fixed suspend.
Applying "xen/percpu: don't initialize percpu on resume" on top of
47990ecef286 fixes suspend too.
But applying it on top of master
(91772f8420dfa2fcfe4db68480c216db5b79c512 specifically) does not fix it,
but the failure mode is different than without the patch - system resets
on S3 resume, with no crash message on the serial console (even with
sync_console), instead of hanging.
And one more data point: reverting f75780d26b2f on top of master is the
same as applying "xen/percpu: don't initialize percpu on resume" on
master - system reset on S3 resume.
So, it looks like there are more issues...

-- 
Best Regards,
Marek Marczykowski-Górecki
Invisible Things Lab

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]