aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/cpufreq
diff options
context:
space:
mode:
authorTejun Heo <tj@kernel.org>2009-08-14 01:41:02 -0400
committerTejun Heo <tj@kernel.org>2009-08-14 01:45:31 -0400
commit384be2b18a5f9475eab9ca2bdfa95cc1a04ef59c (patch)
tree04c93f391a1b65c8bf8d7ba8643c07d26c26590a /drivers/cpufreq
parenta76761b621bcd8336065c4fe3a74f046858bc34c (diff)
parent142d44b0dd6741a64a7bdbe029110e7c1dcf1d23 (diff)
Merge branch 'percpu-for-linus' into percpu-for-next
Conflicts: arch/sparc/kernel/smp_64.c arch/x86/kernel/cpu/perf_counter.c arch/x86/kernel/setup_percpu.c drivers/cpufreq/cpufreq_ondemand.c mm/percpu.c Conflicts in core and arch percpu codes are mostly from commit ed78e1e078dd44249f88b1dd8c76dafb39567161 which substituted many num_possible_cpus() with nr_cpu_ids. As for-next branch has moved all the first chunk allocators into mm/percpu.c, the changes are moved from arch code to mm/percpu.c. Signed-off-by: Tejun Heo <tj@kernel.org>
Diffstat (limited to 'drivers/cpufreq')
-rw-r--r--drivers/cpufreq/cpufreq.c99
-rw-r--r--drivers/cpufreq/cpufreq_conservative.c43
-rw-r--r--drivers/cpufreq/cpufreq_ondemand.c77
3 files changed, 118 insertions, 101 deletions
diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
index 6e2ec0b18948..fd69086d08d5 100644
--- a/drivers/cpufreq/cpufreq.c
+++ b/drivers/cpufreq/cpufreq.c
@@ -761,6 +761,10 @@ static struct kobj_type ktype_cpufreq = {
761 * cpufreq_add_dev - add a CPU device 761 * cpufreq_add_dev - add a CPU device
762 * 762 *
763 * Adds the cpufreq interface for a CPU device. 763 * Adds the cpufreq interface for a CPU device.
764 *
765 * The Oracle says: try running cpufreq registration/unregistration concurrently
766 * with with cpu hotplugging and all hell will break loose. Tried to clean this
767 * mess up, but more thorough testing is needed. - Mathieu
764 */ 768 */
765static int cpufreq_add_dev(struct sys_device *sys_dev) 769static int cpufreq_add_dev(struct sys_device *sys_dev)
766{ 770{
@@ -772,9 +776,6 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
772 struct sys_device *cpu_sys_dev; 776 struct sys_device *cpu_sys_dev;
773 unsigned long flags; 777 unsigned long flags;
774 unsigned int j; 778 unsigned int j;
775#ifdef CONFIG_SMP
776 struct cpufreq_policy *managed_policy;
777#endif
778 779
779 if (cpu_is_offline(cpu)) 780 if (cpu_is_offline(cpu))
780 return 0; 781 return 0;
@@ -804,15 +805,12 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
804 goto nomem_out; 805 goto nomem_out;
805 } 806 }
806 if (!alloc_cpumask_var(&policy->cpus, GFP_KERNEL)) { 807 if (!alloc_cpumask_var(&policy->cpus, GFP_KERNEL)) {
807 kfree(policy);
808 ret = -ENOMEM; 808 ret = -ENOMEM;
809 goto nomem_out; 809 goto err_free_policy;
810 } 810 }
811 if (!zalloc_cpumask_var(&policy->related_cpus, GFP_KERNEL)) { 811 if (!zalloc_cpumask_var(&policy->related_cpus, GFP_KERNEL)) {
812 free_cpumask_var(policy->cpus);
813 kfree(policy);
814 ret = -ENOMEM; 812 ret = -ENOMEM;
815 goto nomem_out; 813 goto err_free_cpumask;
816 } 814 }
817 815
818 policy->cpu = cpu; 816 policy->cpu = cpu;
@@ -820,7 +818,8 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
820 818
821 /* Initially set CPU itself as the policy_cpu */ 819 /* Initially set CPU itself as the policy_cpu */
822 per_cpu(policy_cpu, cpu) = cpu; 820 per_cpu(policy_cpu, cpu) = cpu;
823 lock_policy_rwsem_write(cpu); 821 ret = (lock_policy_rwsem_write(cpu) < 0);
822 WARN_ON(ret);
824 823
825 init_completion(&policy->kobj_unregister); 824 init_completion(&policy->kobj_unregister);
826 INIT_WORK(&policy->update, handle_update); 825 INIT_WORK(&policy->update, handle_update);
@@ -833,7 +832,7 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
833 ret = cpufreq_driver->init(policy); 832 ret = cpufreq_driver->init(policy);
834 if (ret) { 833 if (ret) {
835 dprintk("initialization failed\n"); 834 dprintk("initialization failed\n");
836 goto err_out; 835 goto err_unlock_policy;
837 } 836 }
838 policy->user_policy.min = policy->min; 837 policy->user_policy.min = policy->min;
839 policy->user_policy.max = policy->max; 838 policy->user_policy.max = policy->max;
@@ -852,21 +851,31 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
852#endif 851#endif
853 852
854 for_each_cpu(j, policy->cpus) { 853 for_each_cpu(j, policy->cpus) {
854 struct cpufreq_policy *managed_policy;
855
855 if (cpu == j) 856 if (cpu == j)
856 continue; 857 continue;
857 858
858 /* Check for existing affected CPUs. 859 /* Check for existing affected CPUs.
859 * They may not be aware of it due to CPU Hotplug. 860 * They may not be aware of it due to CPU Hotplug.
861 * cpufreq_cpu_put is called when the device is removed
862 * in __cpufreq_remove_dev()
860 */ 863 */
861 managed_policy = cpufreq_cpu_get(j); /* FIXME: Where is this released? What about error paths? */ 864 managed_policy = cpufreq_cpu_get(j);
862 if (unlikely(managed_policy)) { 865 if (unlikely(managed_policy)) {
863 866
864 /* Set proper policy_cpu */ 867 /* Set proper policy_cpu */
865 unlock_policy_rwsem_write(cpu); 868 unlock_policy_rwsem_write(cpu);
866 per_cpu(policy_cpu, cpu) = managed_policy->cpu; 869 per_cpu(policy_cpu, cpu) = managed_policy->cpu;
867 870
868 if (lock_policy_rwsem_write(cpu) < 0) 871 if (lock_policy_rwsem_write(cpu) < 0) {
869 goto err_out_driver_exit; 872 /* Should not go through policy unlock path */
873 if (cpufreq_driver->exit)
874 cpufreq_driver->exit(policy);
875 ret = -EBUSY;
876 cpufreq_cpu_put(managed_policy);
877 goto err_free_cpumask;
878 }
870 879
871 spin_lock_irqsave(&cpufreq_driver_lock, flags); 880 spin_lock_irqsave(&cpufreq_driver_lock, flags);
872 cpumask_copy(managed_policy->cpus, policy->cpus); 881 cpumask_copy(managed_policy->cpus, policy->cpus);
@@ -878,11 +887,13 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
878 &managed_policy->kobj, 887 &managed_policy->kobj,
879 "cpufreq"); 888 "cpufreq");
880 if (ret) 889 if (ret)
881 goto err_out_driver_exit; 890 cpufreq_cpu_put(managed_policy);
882 891 /*
883 cpufreq_debug_enable_ratelimit(); 892 * Success. We only needed to be added to the mask.
884 ret = 0; 893 * Call driver->exit() because only the cpu parent of
885 goto err_out_driver_exit; /* call driver->exit() */ 894 * the kobj needed to call init().
895 */
896 goto out_driver_exit; /* call driver->exit() */
886 } 897 }
887 } 898 }
888#endif 899#endif
@@ -892,29 +903,31 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
892 ret = kobject_init_and_add(&policy->kobj, &ktype_cpufreq, &sys_dev->kobj, 903 ret = kobject_init_and_add(&policy->kobj, &ktype_cpufreq, &sys_dev->kobj,
893 "cpufreq"); 904 "cpufreq");
894 if (ret) 905 if (ret)
895 goto err_out_driver_exit; 906 goto out_driver_exit;
896 907
897 /* set up files for this cpu device */ 908 /* set up files for this cpu device */
898 drv_attr = cpufreq_driver->attr; 909 drv_attr = cpufreq_driver->attr;
899 while ((drv_attr) && (*drv_attr)) { 910 while ((drv_attr) && (*drv_attr)) {
900 ret = sysfs_create_file(&policy->kobj, &((*drv_attr)->attr)); 911 ret = sysfs_create_file(&policy->kobj, &((*drv_attr)->attr));
901 if (ret) 912 if (ret)
902 goto err_out_driver_exit; 913 goto err_out_kobj_put;
903 drv_attr++; 914 drv_attr++;
904 } 915 }
905 if (cpufreq_driver->get) { 916 if (cpufreq_driver->get) {
906 ret = sysfs_create_file(&policy->kobj, &cpuinfo_cur_freq.attr); 917 ret = sysfs_create_file(&policy->kobj, &cpuinfo_cur_freq.attr);
907 if (ret) 918 if (ret)
908 goto err_out_driver_exit; 919 goto err_out_kobj_put;
909 } 920 }
910 if (cpufreq_driver->target) { 921 if (cpufreq_driver->target) {
911 ret = sysfs_create_file(&policy->kobj, &scaling_cur_freq.attr); 922 ret = sysfs_create_file(&policy->kobj, &scaling_cur_freq.attr);
912 if (ret) 923 if (ret)
913 goto err_out_driver_exit; 924 goto err_out_kobj_put;
914 } 925 }
915 926
916 spin_lock_irqsave(&cpufreq_driver_lock, flags); 927 spin_lock_irqsave(&cpufreq_driver_lock, flags);
917 for_each_cpu(j, policy->cpus) { 928 for_each_cpu(j, policy->cpus) {
929 if (!cpu_online(j))
930 continue;
918 per_cpu(cpufreq_cpu_data, j) = policy; 931 per_cpu(cpufreq_cpu_data, j) = policy;
919 per_cpu(policy_cpu, j) = policy->cpu; 932 per_cpu(policy_cpu, j) = policy->cpu;
920 } 933 }
@@ -922,18 +935,22 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
922 935
923 /* symlink affected CPUs */ 936 /* symlink affected CPUs */
924 for_each_cpu(j, policy->cpus) { 937 for_each_cpu(j, policy->cpus) {
938 struct cpufreq_policy *managed_policy;
939
925 if (j == cpu) 940 if (j == cpu)
926 continue; 941 continue;
927 if (!cpu_online(j)) 942 if (!cpu_online(j))
928 continue; 943 continue;
929 944
930 dprintk("CPU %u already managed, adding link\n", j); 945 dprintk("CPU %u already managed, adding link\n", j);
931 cpufreq_cpu_get(cpu); 946 managed_policy = cpufreq_cpu_get(cpu);
932 cpu_sys_dev = get_cpu_sysdev(j); 947 cpu_sys_dev = get_cpu_sysdev(j);
933 ret = sysfs_create_link(&cpu_sys_dev->kobj, &policy->kobj, 948 ret = sysfs_create_link(&cpu_sys_dev->kobj, &policy->kobj,
934 "cpufreq"); 949 "cpufreq");
935 if (ret) 950 if (ret) {
951 cpufreq_cpu_put(managed_policy);
936 goto err_out_unregister; 952 goto err_out_unregister;
953 }
937 } 954 }
938 955
939 policy->governor = NULL; /* to assure that the starting sequence is 956 policy->governor = NULL; /* to assure that the starting sequence is
@@ -965,17 +982,20 @@ err_out_unregister:
965 per_cpu(cpufreq_cpu_data, j) = NULL; 982 per_cpu(cpufreq_cpu_data, j) = NULL;
966 spin_unlock_irqrestore(&cpufreq_driver_lock, flags); 983 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
967 984
985err_out_kobj_put:
968 kobject_put(&policy->kobj); 986 kobject_put(&policy->kobj);
969 wait_for_completion(&policy->kobj_unregister); 987 wait_for_completion(&policy->kobj_unregister);
970 988
971err_out_driver_exit: 989out_driver_exit:
972 if (cpufreq_driver->exit) 990 if (cpufreq_driver->exit)
973 cpufreq_driver->exit(policy); 991 cpufreq_driver->exit(policy);
974 992
975err_out: 993err_unlock_policy:
976 unlock_policy_rwsem_write(cpu); 994 unlock_policy_rwsem_write(cpu);
995err_free_cpumask:
996 free_cpumask_var(policy->cpus);
997err_free_policy:
977 kfree(policy); 998 kfree(policy);
978
979nomem_out: 999nomem_out:
980 module_put(cpufreq_driver->owner); 1000 module_put(cpufreq_driver->owner);
981module_out: 1001module_out:
@@ -1070,8 +1090,6 @@ static int __cpufreq_remove_dev(struct sys_device *sys_dev)
1070 spin_unlock_irqrestore(&cpufreq_driver_lock, flags); 1090 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1071#endif 1091#endif
1072 1092
1073 unlock_policy_rwsem_write(cpu);
1074
1075 if (cpufreq_driver->target) 1093 if (cpufreq_driver->target)
1076 __cpufreq_governor(data, CPUFREQ_GOV_STOP); 1094 __cpufreq_governor(data, CPUFREQ_GOV_STOP);
1077 1095
@@ -1088,6 +1106,8 @@ static int __cpufreq_remove_dev(struct sys_device *sys_dev)
1088 if (cpufreq_driver->exit) 1106 if (cpufreq_driver->exit)
1089 cpufreq_driver->exit(data); 1107 cpufreq_driver->exit(data);
1090 1108
1109 unlock_policy_rwsem_write(cpu);
1110
1091 free_cpumask_var(data->related_cpus); 1111 free_cpumask_var(data->related_cpus);
1092 free_cpumask_var(data->cpus); 1112 free_cpumask_var(data->cpus);
1093 kfree(data); 1113 kfree(data);
@@ -1228,13 +1248,22 @@ EXPORT_SYMBOL(cpufreq_get);
1228 1248
1229static int cpufreq_suspend(struct sys_device *sysdev, pm_message_t pmsg) 1249static int cpufreq_suspend(struct sys_device *sysdev, pm_message_t pmsg)
1230{ 1250{
1231 int cpu = sysdev->id;
1232 int ret = 0; 1251 int ret = 0;
1252
1253#ifdef __powerpc__
1254 int cpu = sysdev->id;
1233 unsigned int cur_freq = 0; 1255 unsigned int cur_freq = 0;
1234 struct cpufreq_policy *cpu_policy; 1256 struct cpufreq_policy *cpu_policy;
1235 1257
1236 dprintk("suspending cpu %u\n", cpu); 1258 dprintk("suspending cpu %u\n", cpu);
1237 1259
1260 /*
1261 * This whole bogosity is here because Powerbooks are made of fail.
1262 * No sane platform should need any of the code below to be run.
1263 * (it's entirely the wrong thing to do, as driver->get may
1264 * reenable interrupts on some architectures).
1265 */
1266
1238 if (!cpu_online(cpu)) 1267 if (!cpu_online(cpu))
1239 return 0; 1268 return 0;
1240 1269
@@ -1293,6 +1322,7 @@ static int cpufreq_suspend(struct sys_device *sysdev, pm_message_t pmsg)
1293 1322
1294out: 1323out:
1295 cpufreq_cpu_put(cpu_policy); 1324 cpufreq_cpu_put(cpu_policy);
1325#endif /* __powerpc__ */
1296 return ret; 1326 return ret;
1297} 1327}
1298 1328
@@ -1306,12 +1336,18 @@ out:
1306 */ 1336 */
1307static int cpufreq_resume(struct sys_device *sysdev) 1337static int cpufreq_resume(struct sys_device *sysdev)
1308{ 1338{
1309 int cpu = sysdev->id;
1310 int ret = 0; 1339 int ret = 0;
1340
1341#ifdef __powerpc__
1342 int cpu = sysdev->id;
1311 struct cpufreq_policy *cpu_policy; 1343 struct cpufreq_policy *cpu_policy;
1312 1344
1313 dprintk("resuming cpu %u\n", cpu); 1345 dprintk("resuming cpu %u\n", cpu);
1314 1346
1347 /* As with the ->suspend method, all the code below is
1348 * only necessary because Powerbooks suck.
1349 * See commit 42d4dc3f4e1e for jokes. */
1350
1315 if (!cpu_online(cpu)) 1351 if (!cpu_online(cpu))
1316 return 0; 1352 return 0;
1317 1353
@@ -1375,6 +1411,7 @@ out:
1375 schedule_work(&cpu_policy->update); 1411 schedule_work(&cpu_policy->update);
1376fail: 1412fail:
1377 cpufreq_cpu_put(cpu_policy); 1413 cpufreq_cpu_put(cpu_policy);
1414#endif /* __powerpc__ */
1378 return ret; 1415 return ret;
1379} 1416}
1380 1417
diff --git a/drivers/cpufreq/cpufreq_conservative.c b/drivers/cpufreq/cpufreq_conservative.c
index a7ef465c83b9..bc33ddc9c97c 100644
--- a/drivers/cpufreq/cpufreq_conservative.c
+++ b/drivers/cpufreq/cpufreq_conservative.c
@@ -64,21 +64,20 @@ struct cpu_dbs_info_s {
64 unsigned int requested_freq; 64 unsigned int requested_freq;
65 int cpu; 65 int cpu;
66 unsigned int enable:1; 66 unsigned int enable:1;
67 /*
68 * percpu mutex that serializes governor limit change with
69 * do_dbs_timer invocation. We do not want do_dbs_timer to run
70 * when user is changing the governor or limits.
71 */
72 struct mutex timer_mutex;
67}; 73};
68static DEFINE_PER_CPU(struct cpu_dbs_info_s, cs_cpu_dbs_info); 74static DEFINE_PER_CPU(struct cpu_dbs_info_s, cs_cpu_dbs_info);
69 75
70static unsigned int dbs_enable; /* number of CPUs using this policy */ 76static unsigned int dbs_enable; /* number of CPUs using this policy */
71 77
72/* 78/*
73 * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug 79 * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
74 * lock and dbs_mutex. cpu_hotplug lock should always be held before 80 * different CPUs. It protects dbs_enable in governor start/stop.
75 * dbs_mutex. If any function that can potentially take cpu_hotplug lock
76 * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
77 * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
78 * is recursive for the same process. -Venki
79 * DEADLOCK ALERT! (2) : do_dbs_timer() must not take the dbs_mutex, because it
80 * would deadlock with cancel_delayed_work_sync(), which is needed for proper
81 * raceless workqueue teardown.
82 */ 81 */
83static DEFINE_MUTEX(dbs_mutex); 82static DEFINE_MUTEX(dbs_mutex);
84 83
@@ -488,18 +487,12 @@ static void do_dbs_timer(struct work_struct *work)
488 487
489 delay -= jiffies % delay; 488 delay -= jiffies % delay;
490 489
491 if (lock_policy_rwsem_write(cpu) < 0) 490 mutex_lock(&dbs_info->timer_mutex);
492 return;
493
494 if (!dbs_info->enable) {
495 unlock_policy_rwsem_write(cpu);
496 return;
497 }
498 491
499 dbs_check_cpu(dbs_info); 492 dbs_check_cpu(dbs_info);
500 493
501 queue_delayed_work_on(cpu, kconservative_wq, &dbs_info->work, delay); 494 queue_delayed_work_on(cpu, kconservative_wq, &dbs_info->work, delay);
502 unlock_policy_rwsem_write(cpu); 495 mutex_unlock(&dbs_info->timer_mutex);
503} 496}
504 497
505static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info) 498static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
@@ -535,9 +528,6 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
535 if ((!cpu_online(cpu)) || (!policy->cur)) 528 if ((!cpu_online(cpu)) || (!policy->cur))
536 return -EINVAL; 529 return -EINVAL;
537 530
538 if (this_dbs_info->enable) /* Already enabled */
539 break;
540
541 mutex_lock(&dbs_mutex); 531 mutex_lock(&dbs_mutex);
542 532
543 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group); 533 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
@@ -561,6 +551,7 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
561 this_dbs_info->down_skip = 0; 551 this_dbs_info->down_skip = 0;
562 this_dbs_info->requested_freq = policy->cur; 552 this_dbs_info->requested_freq = policy->cur;
563 553
554 mutex_init(&this_dbs_info->timer_mutex);
564 dbs_enable++; 555 dbs_enable++;
565 /* 556 /*
566 * Start the timerschedule work, when this governor 557 * Start the timerschedule work, when this governor
@@ -590,17 +581,19 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
590 &dbs_cpufreq_notifier_block, 581 &dbs_cpufreq_notifier_block,
591 CPUFREQ_TRANSITION_NOTIFIER); 582 CPUFREQ_TRANSITION_NOTIFIER);
592 } 583 }
593 dbs_timer_init(this_dbs_info);
594
595 mutex_unlock(&dbs_mutex); 584 mutex_unlock(&dbs_mutex);
596 585
586 dbs_timer_init(this_dbs_info);
587
597 break; 588 break;
598 589
599 case CPUFREQ_GOV_STOP: 590 case CPUFREQ_GOV_STOP:
600 mutex_lock(&dbs_mutex);
601 dbs_timer_exit(this_dbs_info); 591 dbs_timer_exit(this_dbs_info);
592
593 mutex_lock(&dbs_mutex);
602 sysfs_remove_group(&policy->kobj, &dbs_attr_group); 594 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
603 dbs_enable--; 595 dbs_enable--;
596 mutex_destroy(&this_dbs_info->timer_mutex);
604 597
605 /* 598 /*
606 * Stop the timerschedule work, when this governor 599 * Stop the timerschedule work, when this governor
@@ -616,7 +609,7 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
616 break; 609 break;
617 610
618 case CPUFREQ_GOV_LIMITS: 611 case CPUFREQ_GOV_LIMITS:
619 mutex_lock(&dbs_mutex); 612 mutex_lock(&this_dbs_info->timer_mutex);
620 if (policy->max < this_dbs_info->cur_policy->cur) 613 if (policy->max < this_dbs_info->cur_policy->cur)
621 __cpufreq_driver_target( 614 __cpufreq_driver_target(
622 this_dbs_info->cur_policy, 615 this_dbs_info->cur_policy,
@@ -625,7 +618,7 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
625 __cpufreq_driver_target( 618 __cpufreq_driver_target(
626 this_dbs_info->cur_policy, 619 this_dbs_info->cur_policy,
627 policy->min, CPUFREQ_RELATION_L); 620 policy->min, CPUFREQ_RELATION_L);
628 mutex_unlock(&dbs_mutex); 621 mutex_unlock(&this_dbs_info->timer_mutex);
629 622
630 break; 623 break;
631 } 624 }
diff --git a/drivers/cpufreq/cpufreq_ondemand.c b/drivers/cpufreq/cpufreq_ondemand.c
index 36f292a7bd01..d7a528c80de8 100644
--- a/drivers/cpufreq/cpufreq_ondemand.c
+++ b/drivers/cpufreq/cpufreq_ondemand.c
@@ -70,23 +70,21 @@ struct cpu_dbs_info_s {
70 unsigned int freq_lo_jiffies; 70 unsigned int freq_lo_jiffies;
71 unsigned int freq_hi_jiffies; 71 unsigned int freq_hi_jiffies;
72 int cpu; 72 int cpu;
73 unsigned int enable:1, 73 unsigned int sample_type:1;
74 sample_type:1; 74 /*
75 * percpu mutex that serializes governor limit change with
76 * do_dbs_timer invocation. We do not want do_dbs_timer to run
77 * when user is changing the governor or limits.
78 */
79 struct mutex timer_mutex;
75}; 80};
76static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info); 81static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info);
77 82
78static unsigned int dbs_enable; /* number of CPUs using this policy */ 83static unsigned int dbs_enable; /* number of CPUs using this policy */
79 84
80/* 85/*
81 * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug 86 * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
82 * lock and dbs_mutex. cpu_hotplug lock should always be held before 87 * different CPUs. It protects dbs_enable in governor start/stop.
83 * dbs_mutex. If any function that can potentially take cpu_hotplug lock
84 * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
85 * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
86 * is recursive for the same process. -Venki
87 * DEADLOCK ALERT! (2) : do_dbs_timer() must not take the dbs_mutex, because it
88 * would deadlock with cancel_delayed_work_sync(), which is needed for proper
89 * raceless workqueue teardown.
90 */ 88 */
91static DEFINE_MUTEX(dbs_mutex); 89static DEFINE_MUTEX(dbs_mutex);
92 90
@@ -193,13 +191,18 @@ static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
193 return freq_hi; 191 return freq_hi;
194} 192}
195 193
194static void ondemand_powersave_bias_init_cpu(int cpu)
195{
196 struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
197 dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
198 dbs_info->freq_lo = 0;
199}
200
196static void ondemand_powersave_bias_init(void) 201static void ondemand_powersave_bias_init(void)
197{ 202{
198 int i; 203 int i;
199 for_each_online_cpu(i) { 204 for_each_online_cpu(i) {
200 struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, i); 205 ondemand_powersave_bias_init_cpu(i);
201 dbs_info->freq_table = cpufreq_frequency_get_table(i);
202 dbs_info->freq_lo = 0;
203 } 206 }
204} 207}
205 208
@@ -241,12 +244,10 @@ static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
241 unsigned int input; 244 unsigned int input;
242 int ret; 245 int ret;
243 ret = sscanf(buf, "%u", &input); 246 ret = sscanf(buf, "%u", &input);
247 if (ret != 1)
248 return -EINVAL;
244 249
245 mutex_lock(&dbs_mutex); 250 mutex_lock(&dbs_mutex);
246 if (ret != 1) {
247 mutex_unlock(&dbs_mutex);
248 return -EINVAL;
249 }
250 dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate); 251 dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
251 mutex_unlock(&dbs_mutex); 252 mutex_unlock(&dbs_mutex);
252 253
@@ -260,13 +261,12 @@ static ssize_t store_up_threshold(struct cpufreq_policy *unused,
260 int ret; 261 int ret;
261 ret = sscanf(buf, "%u", &input); 262 ret = sscanf(buf, "%u", &input);
262 263
263 mutex_lock(&dbs_mutex);
264 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || 264 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
265 input < MIN_FREQUENCY_UP_THRESHOLD) { 265 input < MIN_FREQUENCY_UP_THRESHOLD) {
266 mutex_unlock(&dbs_mutex);
267 return -EINVAL; 266 return -EINVAL;
268 } 267 }
269 268
269 mutex_lock(&dbs_mutex);
270 dbs_tuners_ins.up_threshold = input; 270 dbs_tuners_ins.up_threshold = input;
271 mutex_unlock(&dbs_mutex); 271 mutex_unlock(&dbs_mutex);
272 272
@@ -364,9 +364,6 @@ static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
364 struct cpufreq_policy *policy; 364 struct cpufreq_policy *policy;
365 unsigned int j; 365 unsigned int j;
366 366
367 if (!this_dbs_info->enable)
368 return;
369
370 this_dbs_info->freq_lo = 0; 367 this_dbs_info->freq_lo = 0;
371 policy = this_dbs_info->cur_policy; 368 policy = this_dbs_info->cur_policy;
372 369
@@ -494,14 +491,7 @@ static void do_dbs_timer(struct work_struct *work)
494 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); 491 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
495 492
496 delay -= jiffies % delay; 493 delay -= jiffies % delay;
497 494 mutex_lock(&dbs_info->timer_mutex);
498 if (lock_policy_rwsem_write(cpu) < 0)
499 return;
500
501 if (!dbs_info->enable) {
502 unlock_policy_rwsem_write(cpu);
503 return;
504 }
505 495
506 /* Common NORMAL_SAMPLE setup */ 496 /* Common NORMAL_SAMPLE setup */
507 dbs_info->sample_type = DBS_NORMAL_SAMPLE; 497 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
@@ -518,7 +508,7 @@ static void do_dbs_timer(struct work_struct *work)
518 dbs_info->freq_lo, CPUFREQ_RELATION_H); 508 dbs_info->freq_lo, CPUFREQ_RELATION_H);
519 } 509 }
520 queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay); 510 queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
521 unlock_policy_rwsem_write(cpu); 511 mutex_unlock(&dbs_info->timer_mutex);
522} 512}
523 513
524static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info) 514static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
@@ -527,8 +517,6 @@ static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
527 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); 517 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
528 delay -= jiffies % delay; 518 delay -= jiffies % delay;
529 519
530 dbs_info->enable = 1;
531 ondemand_powersave_bias_init();
532 dbs_info->sample_type = DBS_NORMAL_SAMPLE; 520 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
533 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer); 521 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
534 queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work, 522 queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
@@ -537,7 +525,6 @@ static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
537 525
538static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info) 526static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
539{ 527{
540 dbs_info->enable = 0;
541 cancel_delayed_work_sync(&dbs_info->work); 528 cancel_delayed_work_sync(&dbs_info->work);
542} 529}
543 530
@@ -556,19 +543,15 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
556 if ((!cpu_online(cpu)) || (!policy->cur)) 543 if ((!cpu_online(cpu)) || (!policy->cur))
557 return -EINVAL; 544 return -EINVAL;
558 545
559 if (this_dbs_info->enable) /* Already enabled */
560 break;
561
562 mutex_lock(&dbs_mutex); 546 mutex_lock(&dbs_mutex);
563 dbs_enable++;
564 547
565 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group); 548 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
566 if (rc) { 549 if (rc) {
567 dbs_enable--;
568 mutex_unlock(&dbs_mutex); 550 mutex_unlock(&dbs_mutex);
569 return rc; 551 return rc;
570 } 552 }
571 553
554 dbs_enable++;
572 for_each_cpu(j, policy->cpus) { 555 for_each_cpu(j, policy->cpus) {
573 struct cpu_dbs_info_s *j_dbs_info; 556 struct cpu_dbs_info_s *j_dbs_info;
574 j_dbs_info = &per_cpu(od_cpu_dbs_info, j); 557 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
@@ -582,6 +565,8 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
582 } 565 }
583 } 566 }
584 this_dbs_info->cpu = cpu; 567 this_dbs_info->cpu = cpu;
568 ondemand_powersave_bias_init_cpu(cpu);
569 mutex_init(&this_dbs_info->timer_mutex);
585 /* 570 /*
586 * Start the timerschedule work, when this governor 571 * Start the timerschedule work, when this governor
587 * is used for first time 572 * is used for first time
@@ -599,29 +584,31 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
599 max(min_sampling_rate, 584 max(min_sampling_rate,
600 latency * LATENCY_MULTIPLIER); 585 latency * LATENCY_MULTIPLIER);
601 } 586 }
602 dbs_timer_init(this_dbs_info);
603
604 mutex_unlock(&dbs_mutex); 587 mutex_unlock(&dbs_mutex);
588
589 dbs_timer_init(this_dbs_info);
605 break; 590 break;
606 591
607 case CPUFREQ_GOV_STOP: 592 case CPUFREQ_GOV_STOP:
608 mutex_lock(&dbs_mutex);
609 dbs_timer_exit(this_dbs_info); 593 dbs_timer_exit(this_dbs_info);
594
595 mutex_lock(&dbs_mutex);
610 sysfs_remove_group(&policy->kobj, &dbs_attr_group); 596 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
597 mutex_destroy(&this_dbs_info->timer_mutex);
611 dbs_enable--; 598 dbs_enable--;
612 mutex_unlock(&dbs_mutex); 599 mutex_unlock(&dbs_mutex);
613 600
614 break; 601 break;
615 602
616 case CPUFREQ_GOV_LIMITS: 603 case CPUFREQ_GOV_LIMITS:
617 mutex_lock(&dbs_mutex); 604 mutex_lock(&this_dbs_info->timer_mutex);
618 if (policy->max < this_dbs_info->cur_policy->cur) 605 if (policy->max < this_dbs_info->cur_policy->cur)
619 __cpufreq_driver_target(this_dbs_info->cur_policy, 606 __cpufreq_driver_target(this_dbs_info->cur_policy,
620 policy->max, CPUFREQ_RELATION_H); 607 policy->max, CPUFREQ_RELATION_H);
621 else if (policy->min > this_dbs_info->cur_policy->cur) 608 else if (policy->min > this_dbs_info->cur_policy->cur)
622 __cpufreq_driver_target(this_dbs_info->cur_policy, 609 __cpufreq_driver_target(this_dbs_info->cur_policy,
623 policy->min, CPUFREQ_RELATION_L); 610 policy->min, CPUFREQ_RELATION_L);
624 mutex_unlock(&dbs_mutex); 611 mutex_unlock(&this_dbs_info->timer_mutex);
625 break; 612 break;
626 } 613 }
627 return 0; 614 return 0;