Messages in this thread | | | From | Thomas Gleixner <> | Subject | Re: [PATCH 4/5] x86/mce: Fix all mce notifiers to update the mce->handled bitmask | Date | Fri, 14 Feb 2020 01:18:27 +0100 |
| |
Andy Lutomirski <luto@kernel.org> writes:
> On Thu, Feb 13, 2020 at 2:19 PM Luck, Tony <tony.luck@intel.com> wrote: >> >> On Thu, Feb 13, 2020 at 06:03:08PM +0100, Borislav Petkov wrote: >> > On Wed, Feb 12, 2020 at 12:46:51PM -0800, Tony Luck wrote: >> > > If the handler took any action to log or deal with the error, set >> > > a bit int mce->handled so that the default handler on the end of >> > > the machine check chain can see what has been done. >> > > >> > > [!!! What to do about NOTIFY_STOP ... any handler that returns this >> > > value short-circuits calling subsequent entries on the chain. In >> > > some cases this may be the right thing to do ... but it others we >> > > really want to keep calling other functions on the chain] >> > >> > Yes, we can kill that NOTIFY_STOP thing in the mce code since it is >> > nasty. >> >> Well, there are places where we want to keep NOTIFY_STOP. > > I very very strongly disagree.
Ack. The unholy mess of cpu hotplug notifiers and the at least 50 bugs which were unearthed by converting them to a comprehensible and symmetric state machine have documented the insanity of notifiers nicely.
>> 1) Default case for CEC. We want it to "hide" the corrected error. >> That was one of the main goals for CEC. We've discussed cases >> where CEC shouldn't hide (when internal threshold exceeded and >> it tries to take a page offline ... probably something related to >> CMCI storms ... though we didn't really come to any conclusion) > > Then put this logic in do_machine_check() or in some sensible place > that it calls via some ops structure or directly. Don't hide it in > some incomprehensible, possibly nondeterministic place in a notifier > chain. > >> 2) Errata. Perhaps a vendor/platform specific function at the head >> of the notify chain that weeds out errors that should never have >> been reported. > > No, do this before the notifier chain please.
Right. The amount of possible handlers is really not huge.
So having a well defined flow of explicit calls including the handling of magic workarounds in a central place makes tons of sense.
Thanks,
tglx
| |