Re: [PATCH 01/22] x86: add methods for writing of an MSR on several CPUs

["H. Peter Anvin" <hpa@zytor.com>]

Borislav Petkov wrote:
>>>
>> smp_call_function_many() does allow concurrent execution.
> 
> Well, IMHO, not an absolutely concurrent execution since you can't
> control at what moment in time the IPIs will be executed on each core,
> IOW, the respective call function IPIs will generally not coincide on
> each core at a given moment in time. This is especially true if you're
> sending IPIs to cores across nodes.
> 
> And there's hypothetically a small window where you might get an
> inconsistent MSR content across cores. It's a whole another question
> whether this is going to be relevant.
> 

It *allows* concurrent execution.  It doesn't *guarantee* concurrent
execution.  Looping with smp_call_function_one() means we wait for
synchronization after ever one.

>> Looping over a list with smp_call_function_one() -- which you
>> currently have -- is serializing, at which point we might just push
>> the loop into the caller rather than worrying about a new interface.
> 
> So, the actual question is do we need that done in msr.c? If there are
> no potential users, I could easily do the whole thing in the driver and
> do not touch x86 code.

Only if we care about the additional performance of having it be
nonserializing.

	-hpa
-- 
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.