1 // SPDX-License-Identifier: GPL-2.0-or-later
10 #define pr_fmt(fmt)	"powernv-cpufreq: " fmt
44  * On an idle system we want the global pstate to ramp-down from max value to
45  * min over a span of ~5 secs. Also we want it to initially ramp-down slowly and
46  * then ramp-down rapidly later on.
57 /* Interval after which the timer is queued to bring down global pstate */
61  * struct global_pstate_info -	Per policy data structure to maintain history of
64  *				ramping down
65  * @elapsed_time:		Time in ms spent in ramping down from
71  * @timer:			Is used for ramping down if cpu goes idle for
155  * non-turbo frequency.
209 		if (revmap_data->pstate_id == pstate)  in pstate_to_idx()
210 			return revmap_data->cpufreq_table_idx;  in pstate_to_idx()
219 	struct global_pstate_info *gpstates = policy->driver_data;  in reset_gpstates()
221 	gpstates->highest_lpstate_idx = 0;  in reset_gpstates()
222 	gpstates->elapsed_time = 0;  in reset_gpstates()
223 	gpstates->last_sampled_time = 0;  in reset_gpstates()
224 	gpstates->last_lpstate_idx = 0;  in reset_gpstates()
225 	gpstates->last_gpstate_idx = 0;  in reset_gpstates()
230  * from the firmware passed via device-tree
240 	int rc = -ENODEV;  in init_powernv_pstates()
242 	power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");  in init_powernv_pstates()
244 		pr_warn("power-mgt node not found\n");  in init_powernv_pstates()
245 		return -ENODEV;  in init_powernv_pstates()
248 	if (of_property_read_u32(power_mgt, "ibm,pstate-min", &pstate_min)) {  in init_powernv_pstates()
249 		pr_warn("ibm,pstate-min node not found\n");  in init_powernv_pstates()
253 	if (of_property_read_u32(power_mgt, "ibm,pstate-max", &pstate_max)) {  in init_powernv_pstates()
254 		pr_warn("ibm,pstate-max node not found\n");  in init_powernv_pstates()
258 	if (of_property_read_u32(power_mgt, "ibm,pstate-nominal",  in init_powernv_pstates()
260 		pr_warn("ibm,pstate-nominal not found\n");  in init_powernv_pstates()
264 	if (of_property_read_u32(power_mgt, "ibm,pstate-ultra-turbo",  in init_powernv_pstates()
270 	if (of_property_read_u32(power_mgt, "ibm,pstate-turbo",  in init_powernv_pstates()
287 	pstate_ids = of_get_property(power_mgt, "ibm,pstate-ids", &len_ids);  in init_powernv_pstates()
289 		pr_warn("ibm,pstate-ids not found\n");  in init_powernv_pstates()
293 	pstate_freqs = of_get_property(power_mgt, "ibm,pstate-frequencies-mhz",  in init_powernv_pstates()
296 		pr_warn("ibm,pstate-frequencies-mhz not found\n");  in init_powernv_pstates()
301 		pr_warn("Entries in ibm,pstate-ids and "  in init_powernv_pstates()
302 			"ibm,pstate-frequencies-mhz does not match\n");  in init_powernv_pstates()
326 			rc = -ENOMEM;  in init_powernv_pstates()
330 		revmap_data->pstate_id = id & 0xFF;  in init_powernv_pstates()
331 		revmap_data->cpufreq_table_idx = i;  in init_powernv_pstates()
332 		key = (revmap_data->pstate_id) % POWERNV_MAX_PSTATES;  in init_powernv_pstates()
333 		hash_add(pstate_revmap, &revmap_data->hentry, key);  in init_powernv_pstates()
345 			for (j = i - 1; j >= (int)powernv_pstate_info.max; j--)  in init_powernv_pstates()
376  * cpuinfo_nominal_freq_show - Show the nominal CPU frequency as indicated by
397 	struct chip *chip = per_cpu(chip_info, policy->cpu);		\
399 	return sprintf(buf, "%u\n", chip->member);			\
482  * ((struct powernv_smp_call_data *)arg)->freq;
490 	freq_data->pstate_id = extract_local_pstate(pmspr_val);  in powernv_read_cpu_freq()
491 	freq_data->freq = pstate_id_to_freq(freq_data->pstate_id);  in powernv_read_cpu_freq()
494 		 raw_smp_processor_id(), pmspr_val, freq_data->pstate_id,  in powernv_read_cpu_freq()
495 		 freq_data->freq);  in powernv_read_cpu_freq()
520  * on this CPU should be present in freq_data->pstate_id.
526 	unsigned long pstate_ul = freq_data->pstate_id;  in set_pstate()
527 	unsigned long gpstate_ul = freq_data->gpstate_id;  in set_pstate()
567 		if (chip->throttled)  in powernv_cpufreq_throttle_check()
569 		chip->throttled = true;  in powernv_cpufreq_throttle_check()
572 				     cpu, chip->id, pmsr_pmax,  in powernv_cpufreq_throttle_check()
574 			chip->throttle_sub_turbo++;  in powernv_cpufreq_throttle_check()
576 			chip->throttle_turbo++;  in powernv_cpufreq_throttle_check()
578 		trace_powernv_throttle(chip->id,  in powernv_cpufreq_throttle_check()
579 				      throttle_reason[chip->throttle_reason],  in powernv_cpufreq_throttle_check()
581 	} else if (chip->throttled) {  in powernv_cpufreq_throttle_check()
582 		chip->throttled = false;  in powernv_cpufreq_throttle_check()
583 		trace_powernv_throttle(chip->id,  in powernv_cpufreq_throttle_check()
584 				      throttle_reason[chip->throttle_reason],  in powernv_cpufreq_throttle_check()
608  * calc_global_pstate - Calculate global pstate
611  * @highest_lpstate_idx:	pstate from which its ramping down
614  * ramping down and the time elapsed in ramping down. It follows a quadratic
615  * equation which ensures that it reaches ramping down to pmin in 5sec.
631 			(powernv_pstate_info.min - highest_lpstate_idx)) / 100;  in calc_global_pstate()
646 	 * if it exceeds MAX_RAMP_DOWN_TIME ms for ramp down time.  in queue_gpstate_timer()
648 	 * seconds of ramp down time.  in queue_gpstate_timer()
650 	if ((gpstates->elapsed_time + GPSTATE_TIMER_INTERVAL)  in queue_gpstate_timer()
652 		timer_interval = MAX_RAMP_DOWN_TIME - gpstates->elapsed_time;  in queue_gpstate_timer()
656 	mod_timer(&gpstates->timer, jiffies + msecs_to_jiffies(timer_interval));  in queue_gpstate_timer()
664  * This handler brings down the global pstate closer to the local pstate
671 	struct cpufreq_policy *policy = gpstates->policy;  in gpstate_timer_handler()
675 					- gpstates->last_sampled_time;  in gpstate_timer_handler()
678 	if (!spin_trylock(&gpstates->gpstate_lock))  in gpstate_timer_handler()
682 	 * it back to one of the policy->cpus  in gpstate_timer_handler()
684 	if (!cpumask_test_cpu(raw_smp_processor_id(), policy->cpus)) {  in gpstate_timer_handler()
685 		gpstates->timer.expires = jiffies + msecs_to_jiffies(1);  in gpstate_timer_handler()
686 		add_timer_on(&gpstates->timer, cpumask_first(policy->cpus));  in gpstate_timer_handler()
687 		spin_unlock(&gpstates->gpstate_lock);  in gpstate_timer_handler()
693 	 * We may have wrong in gpstate->last_lpstate_idx  in gpstate_timer_handler()
701 		spin_unlock(&gpstates->gpstate_lock);  in gpstate_timer_handler()
705 	gpstates->last_sampled_time += time_diff;  in gpstate_timer_handler()
706 	gpstates->elapsed_time += time_diff;  in gpstate_timer_handler()
708 	if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {  in gpstate_timer_handler()
712 		gpstates->highest_lpstate_idx = gpstate_idx;  in gpstate_timer_handler()
715 		gpstate_idx = calc_global_pstate(gpstates->elapsed_time,  in gpstate_timer_handler()
716 						 gpstates->highest_lpstate_idx,  in gpstate_timer_handler()
720 	gpstates->last_gpstate_idx = gpstate_idx;  in gpstate_timer_handler()
721 	gpstates->last_lpstate_idx = lpstate_idx;  in gpstate_timer_handler()
726 	if (gpstate_idx != gpstates->last_lpstate_idx)  in gpstate_timer_handler()
730 	spin_unlock(&gpstates->gpstate_lock);  in gpstate_timer_handler()
736  * mask policy->cpus
743 	struct global_pstate_info *gpstates = policy->driver_data;  in powernv_cpufreq_target_index()
765 	spin_lock(&gpstates->gpstate_lock);  in powernv_cpufreq_target_index()
767 	if (!gpstates->last_sampled_time) {  in powernv_cpufreq_target_index()
769 		gpstates->highest_lpstate_idx = new_index;  in powernv_cpufreq_target_index()
773 	if (gpstates->last_gpstate_idx < new_index) {  in powernv_cpufreq_target_index()
774 		gpstates->elapsed_time += cur_msec -  in powernv_cpufreq_target_index()
775 						 gpstates->last_sampled_time;  in powernv_cpufreq_target_index()
778 		 * If its has been ramping down for more than MAX_RAMP_DOWN_TIME  in powernv_cpufreq_target_index()
782 		if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {  in powernv_cpufreq_target_index()
784 			gpstates->highest_lpstate_idx = new_index;  in powernv_cpufreq_target_index()
788 			gpstate_idx = calc_global_pstate(gpstates->elapsed_time,  in powernv_cpufreq_target_index()
789 							 gpstates->highest_lpstate_idx,  in powernv_cpufreq_target_index()
794 		gpstates->highest_lpstate_idx = new_index;  in powernv_cpufreq_target_index()
805 		timer_delete_sync(&gpstates->timer);  in powernv_cpufreq_target_index()
809 	gpstates->last_sampled_time = cur_msec;  in powernv_cpufreq_target_index()
810 	gpstates->last_gpstate_idx = gpstate_idx;  in powernv_cpufreq_target_index()
811 	gpstates->last_lpstate_idx = new_index;  in powernv_cpufreq_target_index()
813 	spin_unlock(&gpstates->gpstate_lock);  in powernv_cpufreq_target_index()
819 	 * if current CPU is within policy->cpus (core)  in powernv_cpufreq_target_index()
821 	smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);  in powernv_cpufreq_target_index()
831 	base = cpu_first_thread_sibling(policy->cpu);  in powernv_cpufreq_cpu_init()
834 		cpumask_set_cpu(base + i, policy->cpus);  in powernv_cpufreq_cpu_init()
836 	kn = kernfs_find_and_get(policy->kobj.sd, throttle_attr_grp.name);  in powernv_cpufreq_cpu_init()
840 		ret = sysfs_create_group(&policy->kobj, &throttle_attr_grp);  in powernv_cpufreq_cpu_init()
843 				policy->cpu);  in powernv_cpufreq_cpu_init()
850 	policy->freq_table = powernv_freqs;  in powernv_cpufreq_cpu_init()
851 	policy->fast_switch_possible = true;  in powernv_cpufreq_cpu_init()
856 	/* Initialise Gpstate ramp-down timer only on POWER8 */  in powernv_cpufreq_cpu_init()
859 		return -ENOMEM;  in powernv_cpufreq_cpu_init()
861 	policy->driver_data = gpstates;  in powernv_cpufreq_cpu_init()
864 	gpstates->policy = policy;  in powernv_cpufreq_cpu_init()
865 	timer_setup(&gpstates->timer, gpstate_timer_handler,  in powernv_cpufreq_cpu_init()
867 	gpstates->timer.expires = jiffies +  in powernv_cpufreq_cpu_init()
869 	spin_lock_init(&gpstates->gpstate_lock);  in powernv_cpufreq_cpu_init()
877 	struct global_pstate_info *gpstates = policy->driver_data;  in powernv_cpufreq_cpu_exit()
881 	smp_call_function_single(policy->cpu, set_pstate, &freq_data, 1);  in powernv_cpufreq_cpu_exit()
883 		timer_delete_sync(&gpstates->timer);  in powernv_cpufreq_cpu_exit()
885 	kfree(policy->driver_data);  in powernv_cpufreq_cpu_exit()
918 	cpumask_and(&mask, &chip->mask, cpu_online_mask);  in powernv_cpufreq_work_fn()
922 	if (!chip->restore)  in powernv_cpufreq_work_fn()
925 	chip->restore = false;  in powernv_cpufreq_work_fn()
932 		index = cpufreq_table_find_index_c(policy, policy->cur, false);  in powernv_cpufreq_work_fn()
934 		cpumask_andnot(&mask, &mask, policy->cpus);  in powernv_cpufreq_work_fn()
951 	omsg.type = be64_to_cpu(msg->params[0]);  in powernv_cpufreq_occ_msg()
956 		pr_info("OCC (On Chip Controller - enforces hard thermal/power limits) Resetting\n");  in powernv_cpufreq_occ_msg()
974 		omsg.chip = be64_to_cpu(msg->params[1]);  in powernv_cpufreq_occ_msg()
975 		omsg.throttle_status = be64_to_cpu(msg->params[2]);  in powernv_cpufreq_occ_msg()
1029 	.name		= "powernv-cpufreq",
1050 		return -ENOMEM;  in init_chip_info()
1055 		ret = -ENOMEM;  in init_chip_info()
1066 		cpumask_set_cpu(cpu, &chip_cpu_mask[nr_chips-1]);  in init_chip_info()
1071 		ret = -ENOMEM;  in init_chip_info()
1114 		return -ENODEV;  in powernv_cpufreq_init()