Re: [PATCH v2 3/5] of: unittest: hashed phandles unitest

[Rob Herring]

On Mon, May 16, 2016 at 11:52 AM, Pantelis Antoniou
<pantelis.antoniou@xxxxxxxxxxxx> wrote:
> Add a benchmarking hashed phandles unittest which report what kind
> of speed up we get switching to hashed phandle lookups.
>
>  ### dt-test ### the hash method is 8.2 times faster than the original
>
> On the beaglebone we perform about 1877 phandle lookups until that
> point in the unittest. Each non-hashed lookup takes about 23us when
> the cash is hot, while the hash lookup takes about 3us.
>
> For those 1877 lookup we get a speedup in the boot sequence of
> 1877 * (23 - 3) = 37.5ms, which is not spectacular but there's no
> point in wasting cycles and energy.
>
> Signed-off-by: Pantelis Antoniou <pantelis.antoniou@xxxxxxxxxxxx>
> ---
>  drivers/of/unittest.c | 68 +++++++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 68 insertions(+)
>
> diff --git a/drivers/of/unittest.c b/drivers/of/unittest.c
> index 7ea3689..59cad84 100644
> --- a/drivers/of/unittest.c
> +++ b/drivers/of/unittest.c
> @@ -25,6 +25,9 @@
>
>  #include <linux/bitops.h>
>
> +#include <linux/timekeeping.h>
> +#include <linux/random.h>
> +
>  #include "of_private.h"
>
>  static struct unittest_results {
> @@ -2266,6 +2269,70 @@ out:
>  static inline void __init of_unittest_overlay(void) { }
>  #endif
>
> +#define PHANDLE_LOOKUPS        1000
> +
> +static void __init of_unittest_phandle_hash(void)
> +{
> +       struct device_node *node;
> +       phandle max_phandle;
> +       u32 ph;
> +       unsigned long flags;
> +       int i, j, total;
> +       ktime_t start, end;
> +       s64 dur[2];
> +       int dec, frac;
> +
> +       /* test only available when hashing is available */
> +       if (!of_phandle_ht_available()) {
> +               pr_warn("phandle hash test requires hash to be initialized\n");
> +               return;

As the point of the unittest is to test the core DT code, this should
be a test fail.

> +       }
> +
> +       /* find the maximum phandle of the tree */
> +       raw_spin_lock_irqsave(&devtree_lock, flags);
> +       max_phandle = 0;
> +       total = 0;
> +       for_each_of_allnodes(node) {
> +               if (node->phandle != (phandle)-1U &&
> +                               node->phandle > max_phandle)
> +                       max_phandle = node->phandle;
> +               total++;
> +       }
> +       raw_spin_unlock_irqrestore(&devtree_lock, flags);
> +       max_phandle++;
> +
> +       pr_debug("phandle: max-phandle #%u, #%d total nodes\n",
> +                       (u32)max_phandle, total);
> +
> +       /* perform random lookups using the hash */
> +       for (j = 0; j < 2; j++) {
> +
> +               /* disabled for pass #0, enabled for pass #1 */
> +               of_phandle_ht_is_disabled = j == 0;

I'm not wild about having this variable leaked from the core code just
for the unit test. Yet another step away from the unittest being a
module.

I think you should just measure current performance and users can
revert the hash table if they want to measure the slow path.

Rob