Messages in this thread | | | Date | Sun, 12 Feb 2012 15:21:42 -0200 | From | Mauro Carvalho Chehab <> | Subject | Re: [PATCH v3 01/31] events/hw_event: Create a Hardware Events Report Mecanism (HERM) |
| |
Em 12-02-2012 10:48, Borislav Petkov escreveu: > On Fri, Feb 10, 2012 at 12:17:51PM -0200, Mauro Carvalho Chehab wrote: >> Em 10-02-2012 11:41, Borislav Petkov escreveu: >>> On Thu, Feb 09, 2012 at 10:01:00PM -0200, Mauro Carvalho Chehab wrote: >>>> In order to provide a proper hardware event subsystem, let's >>>> encapsulate hardware events into a common trace facility, and >>>> make both edac and mce drivers to use it. After that, common >>>> facilities can be moved into a new core for hardware events >>>> reporting subsystem. This patch is the first of a series, and just >>>> touches at mce. >>> >>> I think it would work too if you had only one event: >>> >>> * trace_hw_error(...) >>> >>> which would have as an argument a string describing it, like >>> "Uncorrected Memory Read Error", "Memory Read Error (out of range)" "TLB >>> Multimatch Error" etc., followed by the rest of the error info. >>> >>> Currently, you're introducing at least 5 trace_* calls _only_ for memory >>> errors. What about the remaining couples of tens of errors which haven't >>> been addressed yet? >> >> Good point. >> >> The way I see it is that: >> >> - a non-memory related, non-parsed MCE event would generate a "mce_record" trace >> (we need an additional patch to disable it when the error is parsed. >> I'll address it after finishing the tests with a few other platforms); >> >> As more MCE parsers are added at the core, the situations where such event will >> be generated will reduce, and will eventually disappear in long term. >> >> - a non-x86 event (or a x86 event for a memory controller that is not addressed >> by MCE events) will use a "mc_error"; >> >> - a x86 event generated via MCE will use a "mc_error_mce". >> >> There are two special events defined when there's a memory error _and_ a driver >> bug: >> >> "mc_out_of_range_mce" and "mc_out_of_range". >> >> While the name of them and one of the parameters are memory-controller specific, >> it should be easy to make it generic enough to be used by other types of errors. >> >> The previous EDAC logic were to generate an out of range printk and return. With >> the changes I made, it is possible to let the EDAC to provide the information >> parsed, just discarding the bad parsed value. That's the approach I took, as the >> other information there may be useful. By taking such approach, the MCE information >> will be shown by the "mc_error_mce" trace. So, we can remove the "mc_out_of_range_mce" >> without loosing any information. >> >> In any case, we can't merge the *_mce with the non-mce variant, as the mce.h header >> is arch specific and doesn't exist on PPC and tilera architectures. >> >> So, the only event that we can actually remove is "mc_out_of_range_mce", if we let >> the core generate two events for badly parsed error events. What do you think? > > As I said already, error messages from the drivers should be something > very seldom so they don't need a special trace event. > > But most importantly, _ALL_ hw errors could use a single > trace_hw_error() macro which has a single string argument containing all > the required error info as a string since the error format is different > based on the error type. In any case, memory errors are not special! As > I said also before, we cannot have a trace-call for every error type > which adds additional information or which might generate an error while > producing that error info.
All trace events could be resumed into a single string. That's what the TK_printk macro does.
As I said before, there's just one trace call for memory error events (hw_event:mc_error) on my second RFC.
I've added the mce-variant (mc_error_mce) for the version 3 because of what I understood from the feedback you've provided me in priv, that the mce_record event should be merged with it. I'm more than happy to remove it, if I miss-understood you.
Regards, Mauro.
>
| |