]> www.pilppa.org Git - linux-2.6-omap-h63xx.git/commitdiff
sched: Define structure to store the sched_domain statistics for fbg()
authorGautham R Shenoy <ego@in.ibm.com>
Wed, 25 Mar 2009 09:13:56 +0000 (14:43 +0530)
committerIngo Molnar <mingo@elte.hu>
Wed, 25 Mar 2009 09:30:46 +0000 (10:30 +0100)
Impact: cleanup

Currently we use a lot of local variables in find_busiest_group()
to capture the various statistics related to the sched_domain.
Group them together into a single data structure.

This will help us to offload the job of updating the sched_domain
statistics to a helper function.

Credit: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>
Signed-off-by: Gautham R Shenoy <ego@in.ibm.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: "Balbir Singh" <balbir@in.ibm.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: "Dhaval Giani" <dhaval@linux.vnet.ibm.com>
Cc: Bharata B Rao <bharata@linux.vnet.ibm.com>
LKML-Reference: <20090325091356.13992.25970.stgit@sofia.in.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/sched.c

index 1893d5562f5fc3eb041116030874851aafe4e4c8..8198dbe8e4aa14c512d0a7cc37fb3b2b156ca37f 100644 (file)
@@ -3190,6 +3190,37 @@ static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
        return 0;
 }
 /********** Helpers for find_busiest_group ************************/
+/**
+ * sd_lb_stats - Structure to store the statistics of a sched_domain
+ *             during load balancing.
+ */
+struct sd_lb_stats {
+       struct sched_group *busiest; /* Busiest group in this sd */
+       struct sched_group *this;  /* Local group in this sd */
+       unsigned long total_load;  /* Total load of all groups in sd */
+       unsigned long total_pwr;   /*   Total power of all groups in sd */
+       unsigned long avg_load;    /* Average load across all groups in sd */
+
+       /** Statistics of this group */
+       unsigned long this_load;
+       unsigned long this_load_per_task;
+       unsigned long this_nr_running;
+
+       /* Statistics of the busiest group */
+       unsigned long max_load;
+       unsigned long busiest_load_per_task;
+       unsigned long busiest_nr_running;
+
+       int group_imb; /* Is there imbalance in this sd */
+#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
+       int power_savings_balance; /* Is powersave balance needed for this sd */
+       struct sched_group *group_min; /* Least loaded group in sd */
+       struct sched_group *group_leader; /* Group which relieves group_min */
+       unsigned long min_load_per_task; /* load_per_task in group_min */
+       unsigned long leader_nr_running; /* Nr running of group_leader */
+       unsigned long min_nr_running; /* Nr running of group_min */
+#endif
+};
 
 /**
  * sg_lb_stats - stats of a sched_group required for load_balancing
@@ -3346,23 +3377,16 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
                   unsigned long *imbalance, enum cpu_idle_type idle,
                   int *sd_idle, const struct cpumask *cpus, int *balance)
 {
-       struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
-       unsigned long max_load, avg_load, total_load, this_load, total_pwr;
+       struct sd_lb_stats sds;
+       struct sched_group *group = sd->groups;
        unsigned long max_pull;
-       unsigned long busiest_load_per_task, busiest_nr_running;
-       unsigned long this_load_per_task, this_nr_running;
-       int load_idx, group_imb = 0;
+       int load_idx;
+
+       memset(&sds, 0, sizeof(sds));
 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
-       int power_savings_balance = 1;
-       unsigned long leader_nr_running = 0, min_load_per_task = 0;
-       unsigned long min_nr_running = ULONG_MAX;
-       struct sched_group *group_min = NULL, *group_leader = NULL;
+       sds.power_savings_balance = 1;
+       sds.min_nr_running = ULONG_MAX;
 #endif
-
-       max_load = this_load = total_load = total_pwr = 0;
-       busiest_load_per_task = busiest_nr_running = 0;
-       this_load_per_task = this_nr_running = 0;
-
        load_idx = get_sd_load_idx(sd, idle);
 
        do {
@@ -3378,22 +3402,22 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
                if (balance && !(*balance))
                        goto ret;
 
-               total_load += sgs.group_load;
-               total_pwr += group->__cpu_power;
+               sds.total_load += sgs.group_load;
+               sds.total_pwr += group->__cpu_power;
 
                if (local_group) {
-                       this_load = sgs.avg_load;
-                       this = group;
-                       this_nr_running = sgs.sum_nr_running;
-                       this_load_per_task = sgs.sum_weighted_load;
-               } else if (sgs.avg_load > max_load &&
+                       sds.this_load = sgs.avg_load;
+                       sds.this = group;
+                       sds.this_nr_running = sgs.sum_nr_running;
+                       sds.this_load_per_task = sgs.sum_weighted_load;
+               } else if (sgs.avg_load > sds.max_load &&
                           (sgs.sum_nr_running > sgs.group_capacity ||
                                sgs.group_imb)) {
-                       max_load = sgs.avg_load;
-                       busiest = group;
-                       busiest_nr_running = sgs.sum_nr_running;
-                       busiest_load_per_task = sgs.sum_weighted_load;
-                       group_imb = sgs.group_imb;
+                       sds.max_load = sgs.avg_load;
+                       sds.busiest = group;
+                       sds.busiest_nr_running = sgs.sum_nr_running;
+                       sds.busiest_load_per_task = sgs.sum_weighted_load;
+                       sds.group_imb = sgs.group_imb;
                }
 
 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
@@ -3409,15 +3433,16 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
                 * If the local group is idle or completely loaded
                 * no need to do power savings balance at this domain
                 */
-               if (local_group && (this_nr_running >= sgs.group_capacity ||
-                                   !this_nr_running))
-                       power_savings_balance = 0;
+               if (local_group &&
+                       (sds.this_nr_running >= sgs.group_capacity ||
+                       !sds.this_nr_running))
+                       sds.power_savings_balance = 0;
 
                /*
                 * If a group is already running at full capacity or idle,
                 * don't include that group in power savings calculations
                 */
-               if (!power_savings_balance ||
+               if (!sds.power_savings_balance ||
                        sgs.sum_nr_running >= sgs.group_capacity ||
                        !sgs.sum_nr_running)
                        goto group_next;
@@ -3427,12 +3452,13 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
                 * This is the group from where we need to pick up the load
                 * for saving power
                 */
-               if ((sgs.sum_nr_running < min_nr_running) ||
-                   (sgs.sum_nr_running == min_nr_running &&
-                    group_first_cpu(group) > group_first_cpu(group_min))) {
-                       group_min = group;
-                       min_nr_running = sgs.sum_nr_running;
-                       min_load_per_task = sgs.sum_weighted_load /
+               if ((sgs.sum_nr_running < sds.min_nr_running) ||
+                   (sgs.sum_nr_running == sds.min_nr_running &&
+                    group_first_cpu(group) >
+                       group_first_cpu(sds.group_min))) {
+                       sds.group_min = group;
+                       sds.min_nr_running = sgs.sum_nr_running;
+                       sds.min_load_per_task = sgs.sum_weighted_load /
                                                sgs.sum_nr_running;
                }
 
@@ -3444,29 +3470,32 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
                if (sgs.sum_nr_running > sgs.group_capacity - 1)
                        goto group_next;
 
-               if (sgs.sum_nr_running > leader_nr_running ||
-                   (sgs.sum_nr_running == leader_nr_running &&
-                    group_first_cpu(group) < group_first_cpu(group_leader))) {
-                       group_leader = group;
-                       leader_nr_running = sgs.sum_nr_running;
+               if (sgs.sum_nr_running > sds.leader_nr_running ||
+                   (sgs.sum_nr_running == sds.leader_nr_running &&
+                    group_first_cpu(group) <
+                       group_first_cpu(sds.group_leader))) {
+                       sds.group_leader = group;
+                       sds.leader_nr_running = sgs.sum_nr_running;
                }
 group_next:
 #endif
                group = group->next;
        } while (group != sd->groups);
 
-       if (!busiest || this_load >= max_load || busiest_nr_running == 0)
+       if (!sds.busiest || sds.this_load >= sds.max_load
+               || sds.busiest_nr_running == 0)
                goto out_balanced;
 
-       avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
+       sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
 
-       if (this_load >= avg_load ||
-                       100*max_load <= sd->imbalance_pct*this_load)
+       if (sds.this_load >= sds.avg_load ||
+                       100*sds.max_load <= sd->imbalance_pct * sds.this_load)
                goto out_balanced;
 
-       busiest_load_per_task /= busiest_nr_running;
-       if (group_imb)
-               busiest_load_per_task = min(busiest_load_per_task, avg_load);
+       sds.busiest_load_per_task /= sds.busiest_nr_running;
+       if (sds.group_imb)
+               sds.busiest_load_per_task =
+                       min(sds.busiest_load_per_task, sds.avg_load);
 
        /*
         * We're trying to get all the cpus to the average_load, so we don't
@@ -3479,7 +3508,7 @@ group_next:
         * by pulling tasks to us. Be careful of negative numbers as they'll
         * appear as very large values with unsigned longs.
         */
-       if (max_load <= busiest_load_per_task)
+       if (sds.max_load <= sds.busiest_load_per_task)
                goto out_balanced;
 
        /*
@@ -3487,17 +3516,18 @@ group_next:
         * max load less than avg load(as we skip the groups at or below
         * its cpu_power, while calculating max_load..)
         */
-       if (max_load < avg_load) {
+       if (sds.max_load < sds.avg_load) {
                *imbalance = 0;
                goto small_imbalance;
        }
 
        /* Don't want to pull so many tasks that a group would go idle */
-       max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
+       max_pull = min(sds.max_load - sds.avg_load,
+                       sds.max_load - sds.busiest_load_per_task);
 
        /* How much load to actually move to equalise the imbalance */
-       *imbalance = min(max_pull * busiest->__cpu_power,
-                               (avg_load - this_load) * this->__cpu_power)
+       *imbalance = min(max_pull * sds.busiest->__cpu_power,
+                       (sds.avg_load - sds.this_load) * sds.this->__cpu_power)
                        / SCHED_LOAD_SCALE;
 
        /*
@@ -3506,24 +3536,27 @@ group_next:
         * a think about bumping its value to force at least one task to be
         * moved
         */
-       if (*imbalance < busiest_load_per_task) {
+       if (*imbalance < sds.busiest_load_per_task) {
                unsigned long tmp, pwr_now, pwr_move;
                unsigned int imbn;
 
 small_imbalance:
                pwr_move = pwr_now = 0;
                imbn = 2;
-               if (this_nr_running) {
-                       this_load_per_task /= this_nr_running;
-                       if (busiest_load_per_task > this_load_per_task)
+               if (sds.this_nr_running) {
+                       sds.this_load_per_task /= sds.this_nr_running;
+                       if (sds.busiest_load_per_task >
+                                       sds.this_load_per_task)
                                imbn = 1;
                } else
-                       this_load_per_task = cpu_avg_load_per_task(this_cpu);
-
-               if (max_load - this_load + busiest_load_per_task >=
-                                       busiest_load_per_task * imbn) {
-                       *imbalance = busiest_load_per_task;
-                       return busiest;
+                       sds.this_load_per_task =
+                               cpu_avg_load_per_task(this_cpu);
+
+               if (sds.max_load - sds.this_load +
+                       sds.busiest_load_per_task >=
+                               sds.busiest_load_per_task * imbn) {
+                       *imbalance = sds.busiest_load_per_task;
+                       return sds.busiest;
                }
 
                /*
@@ -3532,52 +3565,54 @@ small_imbalance:
                 * moving them.
                 */
 
-               pwr_now += busiest->__cpu_power *
-                               min(busiest_load_per_task, max_load);
-               pwr_now += this->__cpu_power *
-                               min(this_load_per_task, this_load);
+               pwr_now += sds.busiest->__cpu_power *
+                               min(sds.busiest_load_per_task, sds.max_load);
+               pwr_now += sds.this->__cpu_power *
+                               min(sds.this_load_per_task, sds.this_load);
                pwr_now /= SCHED_LOAD_SCALE;
 
                /* Amount of load we'd subtract */
-               tmp = sg_div_cpu_power(busiest,
-                               busiest_load_per_task * SCHED_LOAD_SCALE);
-               if (max_load > tmp)
-                       pwr_move += busiest->__cpu_power *
-                               min(busiest_load_per_task, max_load - tmp);
+               tmp = sg_div_cpu_power(sds.busiest,
+                               sds.busiest_load_per_task * SCHED_LOAD_SCALE);
+               if (sds.max_load > tmp)
+                       pwr_move += sds.busiest->__cpu_power *
+                               min(sds.busiest_load_per_task,
+                                               sds.max_load - tmp);
 
                /* Amount of load we'd add */
-               if (max_load * busiest->__cpu_power <
-                               busiest_load_per_task * SCHED_LOAD_SCALE)
-                       tmp = sg_div_cpu_power(this,
-                                       max_load * busiest->__cpu_power);
+               if (sds.max_load * sds.busiest->__cpu_power <
+                               sds.busiest_load_per_task * SCHED_LOAD_SCALE)
+                       tmp = sg_div_cpu_power(sds.this,
+                               sds.max_load * sds.busiest->__cpu_power);
                else
-                       tmp = sg_div_cpu_power(this,
-                               busiest_load_per_task * SCHED_LOAD_SCALE);
-               pwr_move += this->__cpu_power *
-                               min(this_load_per_task, this_load + tmp);
+                       tmp = sg_div_cpu_power(sds.this,
+                               sds.busiest_load_per_task * SCHED_LOAD_SCALE);
+               pwr_move += sds.this->__cpu_power *
+                               min(sds.this_load_per_task,
+                                       sds.this_load + tmp);
                pwr_move /= SCHED_LOAD_SCALE;
 
                /* Move if we gain throughput */
                if (pwr_move > pwr_now)
-                       *imbalance = busiest_load_per_task;
+                       *imbalance = sds.busiest_load_per_task;
        }
 
-       return busiest;
+       return sds.busiest;
 
 out_balanced:
 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
        if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
                goto ret;
 
-       if (this != group_leader || group_leader == group_min)
+       if (sds.this != sds.group_leader || sds.group_leader == sds.group_min)
                goto ret;
 
-       *imbalance = min_load_per_task;
+       *imbalance = sds.min_load_per_task;
        if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
                cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
-                       group_first_cpu(group_leader);
+                       group_first_cpu(sds.group_leader);
        }
-       return group_min;
+       return sds.group_min;
 
 #endif
 ret: