On Wed, Jul 29, 2026 at 02:37:37PM -0700, Doug Anderson wrote: > Hi, > > On Tue, Jul 28, 2026 at 5:18 AM Kiryl Shutsemau <[email protected]> wrote: > > > > The piece that isn't boot-time yet is the hardlockup detector that drives > > all this. perf and buddy are mutually exclusive at compile time. We want > > to use perf watchdog for FEAT_NMI while SDEI needs buddy. > > FWIW: can I ask why? You're worried about the "all CPUs lock up at the > same time" problem? I am aware of some work towards solving that (even > with buddy) by letting the hardware watchdog driver's "bark" interrupt > run at NMI-level (or pseudo-NMI-level). With that change, I'm not > aware of any benefits of the "perf" detector over the buddy one.
Honestly, no strong reason. I flagged perf mostly as "the detector you'd reach for once NMIs are back", not from a concrete need. The all-CPUs-wedge-at-once case is the only gap I was vaguely worried about, since there's no live buddy left to notice it. If the NMI-level bark work you mention closes that for buddy, then I don't see a reason to want perf at all -- and as you say, buddy is cheaper. Buddy everywhere would be strictly better for us. -- Kiryl Shutsemau / Kirill A. Shutemov
