Re: linux-next regression: SNP Guest boot hangs with certain cpu/mem config combination

[Ard Biesheuvel]

On Fri, 28 Mar 2025 at 10:26, Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote:
>
> On Fri, Mar 28, 2025 at 10:17:16AM +0100, Ard Biesheuvel wrote:
> > On Fri, 28 Mar 2025 at 10:09, Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote:
> > >
> > > On Fri, Mar 28, 2025 at 10:28:19AM +0200, Kirill A. Shutemov wrote:
> > > > On Thu, Mar 27, 2025 at 07:39:22PM +0200, Kirill A. Shutemov wrote:
> > > > > On Thu, Mar 27, 2025 at 11:02:24AM -0400, Steven Rostedt wrote:
> > > > > > On Thu, 27 Mar 2025 16:43:43 +0200
> > > > > > "Kirill A. Shutemov" <kirill.shutemov@xxxxxxxxxxxxxxx> wrote:
> > > > > >
> > > > > > > > > The only option I see so far is to drop static branch from this path.
> > > > > > > > >
> > > > > > > > > But I am not sure if it the only case were we use static branch from CPU
> > > > > > > > > hotplug callbacks.
> > > > > > > > >
> > > > > > > > > Any other ideas?
> > > > > > > >
> > > > > > > >
> > > > > > > > Hmmm, didn't take too close a look here, but there is the
> > > > > > > > static_key_slow_dec_cpuslocked() variant, would that work here? Is the issue
> > > > > > > > the caller may or may not have the cpu_hotplug lock?
> > > > > > >
> > > > > > > Yes. This is generic page alloc path and it can be called with and without
> > > > > > > the lock.
> > > > > >
> > > > > > Note, it's not the static_branch that is an issue, it's enabling/disabling
> > > > > > the static branch that is. Changing a static branch takes a bit of work as
> > > > > > it does modify the kernel text.
> > > > > >
> > > > > > Is it possible to delay the update via a workqueue?
> > > > >
> > > > > Ah. Good point. Should work. I'll give it try.
> > > >
> > > > The patch below fixes problem for me.
> > >
> > > Ah. No, it won't work. We can get there before workqueues are initialized:
> > > mm_core_init() is called before workqueue_init_early().
> > >
> > > We cannot queue a work. :/
> > >
> > > Steven, any other ideas?
> > >
> >
> > Why is the use of a static key required here?
>
> It is not required, but nice to have. We don't want extra conditional in
> page alloc path.
>

This would be a __read_mostly global variable that rarely changes, and
therefore gets predicted correctly 99% of the time. So I wonder if the
static key is worth the hassle.

Can you quantify the speedup?