Messages in this thread | | | From | Naoya Horiguchi <> | Subject | Re: [PATCH v8] x86: mce: kexec: switch MCE handler for kexec/kdump | Date | Fri, 10 Apr 2015 00:49:33 +0000 |
| |
On Thu, Apr 09, 2015 at 09:05:51PM +0200, Borislav Petkov wrote: > On Thu, Apr 09, 2015 at 06:22:02PM +0000, Luck, Tony wrote: > > > Why? Those CPUs are offlined and num_online_cpus() in mce_start() should > > > account for that, no? > > > > > > And if those are offlined, they're very very unlikely to trigger an MCE > > > as they're idle and not executing code. > > > > Let's step back a few feet and look at the big picture. There are three main classes of machine check > > that we might see while trying to run kdump - an remember that all machine checks are currently > > broadcast, so all cpus whether online or offline will see them > > > > 1) Fatal > > We have to crash - lose the dump. Having a new machine check handler will make things a bit easier > > to see what happened because we won't have any synchronization failed messages from the offline > > cpus. > > But this should not be a problem if kdump path keeps cpu_online_mask > uptodate. I'm looking at kdump_nmi_callback() or crash_nmi_callback() or > so. Those should clear cpu_online_mask and then mce_start() will work > fine on the crashing CPU. > > IMHO, of course.
Sorry, I misread you. With clearing cpu_online_mask in shootdown (not done yet,) raising tolerance should work without timeout message. So I think you are right.
> > 2) Execution path recoverable (SRAR in SDM parlance). > > Also going to be fatal (kdump is all running in ring0, and we can't recover from errors in ring 0). Cleaner > > messages as above. Potentially in the future we might be able to make the kdump machine check handler > > actually recover by just skipping a page - if the location of the error was in the old kernel image. > > > > 3) Non-execution path recoverable (SRAO in SDM) > > We ought to be able to keep kdump running if this happens - the "AO" stands for "action optional", > > so we are going to choose to not take an action. Wherever the error was, it won't affect correctness > > of execution of the current context. > > Those could be simply made to go to dmesg during kdump, i.e. decouple > any MCE consumers. And we do that now anyway, i.e. box without mcelog or > some other ras daemon running. > > So we could reuse the normal handler - we just need to do some tweaking > first... AFAICT, of course. I believe in that endeavor, the devil will > be in the detail.
OK, I'll try this approach with updating cpu_online_mask.
Thanks, Naoya Horiguchi
| |