Messages in this thread |  | | Date | Tue, 24 May 2011 10:14:34 +0200 | From | Borislav Petkov <> | Subject | Re: [RFC 0/9] mce recovery for Sandy Bridge server |
| |
On Tue, May 24, 2011 at 05:40:23AM +0200, Ingo Molnar wrote: > So we *really* want to promote this code to a higher level of abstraction. > Everyone would benefit from doing that: Intel hardware error handling features > would be enabled much more richly and i suspect they would also be *used* in a > much more meaningful way - driving the hw cycle as well.
Absolutely agreed. The RAS architecture should look like this, IMHO:
I. Event collection: #MC handler and pollers, no queueing or buffering crap.
II. Pluggable and extensible filters which are * per vendor * configurable from userspace * easily extensible * decide whether action should be taken in the kernel or error is non-critical and should go to RAS daemon
III. Error handling callback(s) * also extensible * also per vendor * also configurable from userspace
Advantages: * reuse perf code - no need for ad-hoc new buffers and lockless thingies when we have it all already
* easy code and even hw testing with perf inject or ras inject ** this gives us also the different injection methods per vendor in an unified way instead of interfaces in /sys or debugfs or mcelog or ...
* keep code design sane instead of letting it needlessly fiddle with other parts of the kernel
* ...
Now I should better go and put my patches where my mouth is :).
-- Regards/Gruss, Boris.
|  |