cpufreq: powernv: Remove cpu_to_chip_id() from hot-path
authorShilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
Tue, 2 Feb 2016 19:41:39 +0000 (01:11 +0530)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Fri, 5 Feb 2016 01:38:02 +0000 (02:38 +0100)
cpu_to_chip_id() does a DT walk through to find out the chip id by
taking a contended device tree lock. This adds an unnecessary overhead
in a hot path. So instead of calling cpu_to_chip_id() everytime cache
the chip ids for all cores in the array 'core_to_chip_map' and use it
in the hotpath.

Reported-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
drivers/cpufreq/powernv-cpufreq.c

index a271b0f..c670314 100644 (file)
@@ -43,6 +43,7 @@
 
 static struct cpufreq_frequency_table powernv_freqs[POWERNV_MAX_PSTATES+1];
 static bool rebooting, throttled, occ_reset;
+static unsigned int *core_to_chip_map;
 
 static struct chip {
        unsigned int id;
@@ -313,13 +314,14 @@ static inline unsigned int get_nominal_index(void)
 static void powernv_cpufreq_throttle_check(void *data)
 {
        unsigned int cpu = smp_processor_id();
+       unsigned int chip_id = core_to_chip_map[cpu_core_index_of_thread(cpu)];
        unsigned long pmsr;
        int pmsr_pmax, i;
 
        pmsr = get_pmspr(SPRN_PMSR);
 
        for (i = 0; i < nr_chips; i++)
-               if (chips[i].id == cpu_to_chip_id(cpu))
+               if (chips[i].id == chip_id)
                        break;
 
        /* Check for Pmax Capping */
@@ -559,19 +561,29 @@ static int init_chip_info(void)
        unsigned int chip[256];
        unsigned int cpu, i;
        unsigned int prev_chip_id = UINT_MAX;
+       cpumask_t cpu_mask;
+       int ret = -ENOMEM;
 
-       for_each_possible_cpu(cpu) {
+       core_to_chip_map = kcalloc(cpu_nr_cores(), sizeof(unsigned int),
+                                  GFP_KERNEL);
+       if (!core_to_chip_map)
+               goto out;
+
+       cpumask_copy(&cpu_mask, cpu_possible_mask);
+       for_each_cpu(cpu, &cpu_mask) {
                unsigned int id = cpu_to_chip_id(cpu);
 
                if (prev_chip_id != id) {
                        prev_chip_id = id;
                        chip[nr_chips++] = id;
                }
+               core_to_chip_map[cpu_core_index_of_thread(cpu)] = id;
+               cpumask_andnot(&cpu_mask, &cpu_mask, cpu_sibling_mask(cpu));
        }
 
        chips = kmalloc_array(nr_chips, sizeof(struct chip), GFP_KERNEL);
        if (!chips)
-               return -ENOMEM;
+               goto free_chip_map;
 
        for (i = 0; i < nr_chips; i++) {
                chips[i].id = chip[i];
@@ -582,6 +594,10 @@ static int init_chip_info(void)
        }
 
        return 0;
+free_chip_map:
+       kfree(core_to_chip_map);
+out:
+       return ret;
 }
 
 static int __init powernv_cpufreq_init(void)
@@ -616,6 +632,7 @@ static void __exit powernv_cpufreq_exit(void)
        opal_message_notifier_unregister(OPAL_MSG_OCC,
                                         &powernv_cpufreq_opal_nb);
        kfree(chips);
+       kfree(core_to_chip_map);
        cpufreq_unregister_driver(&powernv_cpufreq_driver);
 }
 module_exit(powernv_cpufreq_exit);