Re: Kernel broken on processors without performance counters

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Jul 23, 2015 at 10:19:52AM -0400, Jason Baron wrote:
> > And I think it'll all work. Hmm?
> 
> Cool. This also gives an extra degree of freedom in that it allows keys to
> be arbitrarily mixed with the likely/unlikely branch types. I'm not sure that's
> come up as a use-case, but seems like it would be good. It also implies
> that the LABEL_TYPE_{TRUE,FALSE}, is no longer associated with the key
> b/c a key could be used in both and unlikely() or likely() branch. So that
> would eventually go away, and the 'struct static_key()', I guess could point
> to its relevant entries in both tables. Although, that means an extra
> pointer in the 'struct static_key'. It may be simpler to simply add another
> field to the jump table that specifies if the branch is likely/unlikely,
> and then we are back to one table? IE  arch_static_branch() could add
> that field to the jump table.

Way ahead of you, while implementing the dual section I ran into trouble
and found that it would be far easier to indeed stick it in the
jump_entry.

However, instead of growing the thing, I've used the LSB of the key
field, that's a pointer so it has at least two bits free anyhow.

I've also implemented it for all archs (+- compile failures, I've not
gotten that far).

Lemme finish this and I'll post it.