Re: [PATCHSET x86/core/percpu] improve the first percpu chunk allocation

[Tejun Heo <tj@kernel.org>]

Hello, Ingo.

Patches posted to fix the build failure and warning.  Patches are also
available in the usual git tree.

  git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git tj-percpu

Ingo Molnar wrote:
> and one tip:master merge impact due to API change:
> 
>  kernel/trace/trace_functions_graph.c: In function ‘graph_trace_close’:
>  kernel/trace/trace_functions_graph.c:836: error: implicit declaration of function ‘percpu_free’
> 
> that's free_percpu() now, right?
> 
> Btw., why was this rename done? We generally standardize on 
> hierarchical names, going from the more general to the more 
> specific names, left to right. So we have 
> <subsystem>_<functionality>_<subtype> sort of names generally.

It's a strange story and not really a rename.

We had __percpu_alloc_mask(), percpu_alloc(), __alloc_percpu() and
alloc_percpu() and of course matching frees.  The percpu_*() stuff was
introduced so that allocations can take online cpus into
consideration.  So, percpu_alloc() uses cpu_online_map as the default
allocation mask while alloc_percpu() uses cpu_possible_map.  The
allocation mask thing never really took off and there virtually was no
user and got killed.

If the only merge impact was percpu_free() and it's not missing
percpu_alloc(), it could be that allocation path used alloc_percpu()
but free path used percpu_free().

Thanks.

-- 
tejun