On Mon, May 13, 2019 at 11:56:31AM +0100, Will Deacon wrote:
> Hi Sudeep,
>
> On Thu, May 09, 2019 at 03:13:50PM +0100, Sudeep Holla wrote:
> > On Thu, May 09, 2019 at 11:35:59AM +0100, Sudeep Holla wrote:
> > > On Thu, May 09, 2019 at 10:28:11AM +0100, Will Deacon wrote:
> >
> > [...]
> >
> > > >
> > > > Since SPE uses virtual addressing, we don't really care about the underlying
> > > > page layout so there's no need to use higher-order allocations. I suppose we
> > > > could theoretically map them at the pmd level in some cases, but ignoring
> > > > them should also be harmless and I suspect you can delete the check.
> > > >
> > >
> > > Yes, I did a quick look to see if we can do that, but couldn't find a clue.
> > > Not sure if that's any optimisation, we can use order from page_private
> > > and set the values accordingly ?
> > >
> > And I forgot to add the diff that I mentioned above, something like the
> > patch below.
> >
> > Regards,
> > Sudeep
> >
> > -->8
> >
> > diff --git i/drivers/perf/arm_spe_pmu.c w/drivers/perf/arm_spe_pmu.c
> > index 7cb766dafe85..45cd62517080 100644
> > --- i/drivers/perf/arm_spe_pmu.c
> > +++ w/drivers/perf/arm_spe_pmu.c
> > @@ -827,7 +827,7 @@ static void arm_spe_pmu_read(struct perf_event *event)
> > static void *arm_spe_pmu_setup_aux(struct perf_event *event, void **pages,
> > int nr_pages, bool snapshot)
> > {
> > - int i, cpu = event->cpu;
> > + int i, j, cpu = event->cpu;
> > struct page **pglist;
> > struct arm_spe_pmu_buf *buf;
> >
> > @@ -859,11 +859,12 @@ static void *arm_spe_pmu_setup_aux(struct perf_event *event, void **pages,
> > struct page *page = virt_to_page(pages[i]);
> >
> > if (PagePrivate(page)) {
> > - pr_warn("unexpected high-order page for auxbuf!");
> > - goto out_free_pglist;
> > + for (j = 0; j < 1 << page_private(page); j++)
> > + pglist[i + j] = page++;
> > + i += j - 1;
> > + } else {
> > + pglist[i] = page;
>
> Hmm. Given that vmap() doesn't do anything special for high-order pages
> and rb_alloc_aux()/rb_alloc_aux_page() already split the allocation up
> for the page array, what does your change accomplish on top of that?
>
Not much, instead of computing page ptr for each page using virt_to_page,
we jump pointers automatically for all the pages that are private.
page_private(page) holds the order. i.e. for 2MB high order allocation
we can skip calling virt_to_page for 511 pages that are contiguous.
--
Regards,
Sudeep
