Messages in this thread | | | Date | Mon, 21 Mar 2011 12:34:13 -0500 | From | Jack Steiner <> | Subject | Re: [PATCH] x86, UV: Fix NMI handler for UV platforms |
| |
On Mon, Mar 21, 2011 at 08:19:09PM +0300, Cyrill Gorcunov wrote: > On 03/21/2011 08:08 PM, Jack Steiner wrote: > > On Mon, Mar 21, 2011 at 08:00:53PM +0300, Cyrill Gorcunov wrote: > >> On 03/21/2011 07:43 PM, Cyrill Gorcunov wrote: > >> ... > >>> > >>> I think Jack might need to setup priority for his notifier, like > >>> > >>> static struct notifier_block uv_dump_stack_nmi_nb = { > >>> .notifier_call = uv_handle_nmi, > >>> .priority = NMI_LOCAL_HIGH_PRIOR+1, > >>> }; > >>> > >>> so it would be called before perf nmi. Don, am I right? > >>> > >>> Since for perf nmis we do have > >>> > >>> static __read_mostly struct notifier_block perf_event_nmi_notifier = { > >>> .notifier_call = perf_event_nmi_handler, > >>> .next = NULL, > >>> .priority = NMI_LOCAL_LOW_PRIOR, > >>> }; > >>> > >> > >> I must admit I've missed the fact that Jack has tried NMIs priorities, right? > >> x86_platform_ops seems to be a cleaner indeed (btw I think p4 pmu kgdb issue > >> is exactly the same problem) but same time this might end up in over-swelled > >> ideas behind this small code snippet. Dunno. Probably we need some per-cpu > >> system status for nmi reasons other than unknown nmis... > > > > We use KDB internally, and yes, it has the same issue. The version of the > > patch that uses KDB OR's the "handled" status for both KDB & the UV NMI handler. > > If either KDB or the UV NMI handler returns "handled", the code in traps.c exits > > after the call to the first die notifier. > > > > Not particularily pretty but I could not find a better way to do it. > > > > --- jack > > Another option might be to add pre-nmi notifier chain, which of course > not much differ from platform ops but I guess platform ops stands mostly > for one-shot events while chain might be more flexible. Ie I mean something > like > > if (notify_pre_die(DIE_NMI, "nmi", regs, 0, 2, SIGINT) == NOTIFY_STOP) > return;
You still need to process both chains in order to handle the case where both hw_perf & the SGI BMC raise NMIs at about the same time.
--- jack
| |