Re: [PATCH 18/30] x86, kaiser: map virtually-addressed performance monitoring buffers

[Andy Lutomirski]

On Tue, Nov 14, 2017 at 11:10 AM, Hugh Dickins <hughd@xxxxxxxxxx> wrote:
> On Tue, 14 Nov 2017, Dave Hansen wrote:
>> On 11/14/2017 10:20 AM, Peter Zijlstra wrote:
>> > On Fri, Nov 10, 2017 at 11:31:39AM -0800, Dave Hansen wrote:
>> >>  static int alloc_ds_buffer(int cpu)
>> >>  {
>> >> +  struct debug_store *ds = per_cpu_ptr(&cpu_debug_store, cpu);
>> >>
>> >> +  memset(ds, 0, sizeof(*ds));
>> > Still wondering about that memset...
>
> Sorry, my attention is far away at the moment.
>
>>
>> My guess is that it was done to mirror the zeroing done by the original
>> kzalloc().
>
> You guess right.
>
>> But, I think you're right that it's zero'd already by virtue
>> of being static:
>>
>> static
>> DEFINE_PER_CPU_SHARED_ALIGNED_USER_MAPPED(struct debug_store,
>> cpu_debug_store);
>>
>> I'll queue a cleanup, or update it if I re-post the set.
>
> I was about to agree, but now I'm not so sure.  I don't know much
> about these PMC things, but at a glance it looks like what is reserved
> by x86_reserve_hardware() may later be released by x86_release_hardware(),
> and then later reserved again by x86_reserve_hardware().  And although
> the static per-cpu area would be zeroed the first time, the second time
> it will contain data left over from before, so really needs the memset?
>

For an upstream solution, I would really really like to see
DEFINE_PER_CPU_SHARED_ALIGNED_USER_MAPPED and friends completely gone
and to use cpu_entry_area instead.  I don't know whether this has any
material impact on this particular discussion, though.

--Andy

> Hugh