Re: [RFC 00/15] x86_64: Optimize percpu accesses

[Jeremy Fitzhardinge]

Eric W. Biederman wrote:
> Jeremy Fitzhardinge <jeremy@xxxxxxxx> writes:
>
>   
> > No, that sounds like a bad idea.  For one, how would you enforce it?  How would
> > you check for it?  It's one of those things that would mostly work and then fail
> > very rarely.
> >     
>
> Well the easiest way would be to avoid the letting people take the address of
> per cpu memory, and just provide macros to read/write it.  We are 90% of the
> way there already so it isn't a big jump.
>   

Well, the x86_X_percpu api is there.  But per_cpu() and get_cpu_var() 
both explicitly return lvalues which can have their addresses taken.

> > Secondly, I depend on it.  I register a percpu structure with Xen to share
> > per-vcpu specific information (interrupt mask, time info, runstate stats, etc).
> >     
>
> Well even virtual allocation is likely to break the Xen sharing case as you
> would at least need to compute the physical address and pass it to Xen.
>   

Right.  At the moment it assumes that the percpu variable is in the 
linear mapping, but it could easily do a pagetable walk if necessary.

   J
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/