Messages in this thread | | | Date | Tue, 23 Dec 2014 00:19:29 +0100 | From | Borislav Petkov <> | Subject | Re: [PATCH 0/3] Fix MCE handling for AMD multi-node processors |
| |
On Mon, Dec 22, 2014 at 02:56:47PM -0600, Aravind Gopalakrishnan wrote: > On 12/22/2014 2:15 PM, Borislav Petkov wrote: > >On Mon, Dec 22, 2014 at 02:10:09PM -0600, Aravind Gopalakrishnan wrote: > >>When a MCE happens that is to be logged onto bank 4 of AMD multi-node > >>processors, they are reported only to corresponding node base core of > >>the cpu on which the error occurred. > >> > >>Refer D18F3x44[NbMcaToMstCpuEn] on BKDGs of Fam10h and later for > >Let me try to understand this correctly: > > > >Does that mean that we could fix this by simply doing: > > > >D18F3x44[NbMcaToMstCpuEn]=0b > > > >on each NB? > > > > Not quite.. > When this field is 0, BKDG says the error may be reported to the core that > originated the request *if applicable and known* > Looking at the error signatures table for MC4 (Part 2), > we can see only some errors have 'ErrCoreId' column as valid > > Besides, if IO originated the request, then it is reported only to NBC. > > So, to take care of all these cases, I am just following one approach here: > and that is to look at NBC MSRs for any bank 4 errors. > (It seems to be what the BKDG recommends anyway as BIOS by default should > set D18F3x44[NbMcaToMstCpuEn])
Then in that case you have to check the case where D18F3x44[NbMcaToMstCpuEn] is 0 for whatever reason (some BIOS forgot to set it or whatever) and to set it again.
Then, upon a quick scan, your patches are adding a lot of vendor-specific stuff which doesn't belong in the #MC handler, should probably be wrapped or so, no good idea right now.
Then, you're using rd/wrmsr_on_cpu which does smp_call_function_single() which can deadlock in atomic context and #MC is one.
Also, the math in amd_get_nbc_for_node() is too fragile and will break the moment some BIOS renumbers cores to accomodate some other OS.
In any case, I won't be able to take a detailed look soon with the holidays coming up.
Also, I'm wondering if this can't be solved much more elegantly by detecting that condition (bank == 4) in the #MC handler and issuing an IPI before exiting it using irq_work which will schedule do_machine_check on the NBC. And that should be even easier to do since we're moving the #MC handler out of the IST and to the normal kernel stack for 3.20, which would make this endeavor pretty cheap.
Anyway, just a couple of thoughts...
-- Regards/Gruss, Boris.
Sent from a fat crate under my desk. Formatting is fine. --
| |