2 * POWERNV cpufreq driver for the IBM POWER processors
4 * (C) Copyright IBM 2014
6 * Author: Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2, or (at your option)
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
20 #define pr_fmt(fmt) "powernv-cpufreq: " fmt
22 #include <linux/kernel.h>
23 #include <linux/sysfs.h>
24 #include <linux/cpumask.h>
25 #include <linux/module.h>
26 #include <linux/cpufreq.h>
27 #include <linux/smp.h>
29 #include <linux/reboot.h>
30 #include <linux/slab.h>
31 #include <linux/cpu.h>
32 #include <trace/events/power.h>
34 #include <asm/cputhreads.h>
35 #include <asm/firmware.h>
37 #include <asm/smp.h> /* Required for cpu_sibling_mask() in UP configs */
39 #include <linux/timer.h>
41 #define POWERNV_MAX_PSTATES 256
42 #define PMSR_PSAFE_ENABLE (1UL << 30)
43 #define PMSR_SPR_EM_DISABLE (1UL << 31)
44 #define PMSR_MAX(x) ((x >> 32) & 0xFF)
46 #define MAX_RAMP_DOWN_TIME 5120
48 * On an idle system we want the global pstate to ramp-down from max value to
49 * min over a span of ~5 secs. Also we want it to initially ramp-down slowly and
50 * then ramp-down rapidly later on.
52 * This gives a percentage rampdown for time elapsed in milliseconds.
53 * ramp_down_percentage = ((ms * ms) >> 18)
54 * ~= 3.8 * (sec * sec)
56 * At 0 ms ramp_down_percent = 0
57 * At 5120 ms ramp_down_percent = 100
59 #define ramp_down_percent(time) ((time * time) >> 18)
61 /* Interval after which the timer is queued to bring down global pstate */
62 #define GPSTATE_TIMER_INTERVAL 2000
65 * struct global_pstate_info - Per policy data structure to maintain history of
67 * @highest_lpstate_idx: The local pstate index from which we are
69 * @elapsed_time: Time in ms spent in ramping down from
71 * @last_sampled_time: Time from boot in ms when global pstates were
73 * @last_lpstate_idx, Last set value of local pstate and global
74 * last_gpstate_idx pstate in terms of cpufreq table index
75 * @timer: Is used for ramping down if cpu goes idle for
76 * a long time with global pstate held high
77 * @gpstate_lock: A spinlock to maintain synchronization between
78 * routines called by the timer handler and
79 * governer's target_index calls
81 struct global_pstate_info {
82 int highest_lpstate_idx;
83 unsigned int elapsed_time;
84 unsigned int last_sampled_time;
87 spinlock_t gpstate_lock;
88 struct timer_list timer;
91 static struct cpufreq_frequency_table powernv_freqs[POWERNV_MAX_PSTATES+1];
92 static bool rebooting, throttled, occ_reset;
94 static const char * const throttle_reason[] = {
97 "Processor Over Temperature",
98 "Power Supply Failure",
103 enum throttle_reason_type {
107 POWER_SUPPLY_FAILURE,
119 struct work_struct throttle;
121 int throttle_sub_turbo;
122 int reason[OCC_MAX_REASON];
126 static DEFINE_PER_CPU(struct chip *, chip_info);
130 * The set of pstates consists of contiguous integers.
131 * powernv_pstate_info stores the index of the frequency table for
132 * max, min and nominal frequencies. It also stores number of
133 * available frequencies.
135 * powernv_pstate_info.nominal indicates the index to the highest
136 * non-turbo frequency.
138 static struct powernv_pstate_info {
141 unsigned int nominal;
142 unsigned int nr_pstates;
143 } powernv_pstate_info;
145 /* Use following macros for conversions between pstate_id and index */
146 static inline int idx_to_pstate(unsigned int i)
148 if (unlikely(i >= powernv_pstate_info.nr_pstates)) {
149 pr_warn_once("index %u is out of bound\n", i);
150 return powernv_freqs[powernv_pstate_info.nominal].driver_data;
153 return powernv_freqs[i].driver_data;
156 static inline unsigned int pstate_to_idx(int pstate)
158 int min = powernv_freqs[powernv_pstate_info.min].driver_data;
159 int max = powernv_freqs[powernv_pstate_info.max].driver_data;
162 if (unlikely((pstate < max) || (pstate > min))) {
163 pr_warn_once("pstate %d is out of bound\n", pstate);
164 return powernv_pstate_info.nominal;
167 if (unlikely((pstate > max) || (pstate < min))) {
168 pr_warn_once("pstate %d is out of bound\n", pstate);
169 return powernv_pstate_info.nominal;
173 * abs() is deliberately used so that is works with
174 * both monotonically increasing and decreasing
177 return abs(pstate - idx_to_pstate(powernv_pstate_info.max));
180 static inline void reset_gpstates(struct cpufreq_policy *policy)
182 struct global_pstate_info *gpstates = policy->driver_data;
184 gpstates->highest_lpstate_idx = 0;
185 gpstates->elapsed_time = 0;
186 gpstates->last_sampled_time = 0;
187 gpstates->last_lpstate_idx = 0;
188 gpstates->last_gpstate_idx = 0;
192 * Initialize the freq table based on data obtained
193 * from the firmware passed via device-tree
195 static int init_powernv_pstates(void)
197 struct device_node *power_mgt;
198 int i, nr_pstates = 0;
199 const __be32 *pstate_ids, *pstate_freqs;
200 u32 len_ids, len_freqs;
201 u32 pstate_min, pstate_max, pstate_nominal;
203 power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
205 pr_warn("power-mgt node not found\n");
209 if (of_property_read_u32(power_mgt, "ibm,pstate-min", &pstate_min)) {
210 pr_warn("ibm,pstate-min node not found\n");
214 if (of_property_read_u32(power_mgt, "ibm,pstate-max", &pstate_max)) {
215 pr_warn("ibm,pstate-max node not found\n");
219 if (of_property_read_u32(power_mgt, "ibm,pstate-nominal",
221 pr_warn("ibm,pstate-nominal not found\n");
224 pr_info("cpufreq pstate min %d nominal %d max %d\n", pstate_min,
225 pstate_nominal, pstate_max);
227 pstate_ids = of_get_property(power_mgt, "ibm,pstate-ids", &len_ids);
229 pr_warn("ibm,pstate-ids not found\n");
233 pstate_freqs = of_get_property(power_mgt, "ibm,pstate-frequencies-mhz",
236 pr_warn("ibm,pstate-frequencies-mhz not found\n");
240 if (len_ids != len_freqs) {
241 pr_warn("Entries in ibm,pstate-ids and "
242 "ibm,pstate-frequencies-mhz does not match\n");
245 nr_pstates = min(len_ids, len_freqs) / sizeof(u32);
247 pr_warn("No PStates found\n");
251 powernv_pstate_info.nr_pstates = nr_pstates;
252 pr_debug("NR PStates %d\n", nr_pstates);
253 for (i = 0; i < nr_pstates; i++) {
254 u32 id = be32_to_cpu(pstate_ids[i]);
255 u32 freq = be32_to_cpu(pstate_freqs[i]);
257 pr_debug("PState id %d freq %d MHz\n", id, freq);
258 powernv_freqs[i].frequency = freq * 1000; /* kHz */
259 powernv_freqs[i].driver_data = id;
261 if (id == pstate_max)
262 powernv_pstate_info.max = i;
263 else if (id == pstate_nominal)
264 powernv_pstate_info.nominal = i;
265 else if (id == pstate_min)
266 powernv_pstate_info.min = i;
269 /* End of list marker entry */
270 powernv_freqs[i].frequency = CPUFREQ_TABLE_END;
274 /* Returns the CPU frequency corresponding to the pstate_id. */
275 static unsigned int pstate_id_to_freq(int pstate_id)
279 i = pstate_to_idx(pstate_id);
280 if (i >= powernv_pstate_info.nr_pstates || i < 0) {
281 pr_warn("PState id %d outside of PState table, "
282 "reporting nominal id %d instead\n",
283 pstate_id, idx_to_pstate(powernv_pstate_info.nominal));
284 i = powernv_pstate_info.nominal;
287 return powernv_freqs[i].frequency;
291 * cpuinfo_nominal_freq_show - Show the nominal CPU frequency as indicated by
294 static ssize_t cpuinfo_nominal_freq_show(struct cpufreq_policy *policy,
297 return sprintf(buf, "%u\n",
298 powernv_freqs[powernv_pstate_info.nominal].frequency);
301 struct freq_attr cpufreq_freq_attr_cpuinfo_nominal_freq =
302 __ATTR_RO(cpuinfo_nominal_freq);
304 static struct freq_attr *powernv_cpu_freq_attr[] = {
305 &cpufreq_freq_attr_scaling_available_freqs,
306 &cpufreq_freq_attr_cpuinfo_nominal_freq,
310 #define throttle_attr(name, member) \
311 static ssize_t name##_show(struct cpufreq_policy *policy, char *buf) \
313 struct chip *chip = per_cpu(chip_info, policy->cpu); \
315 return sprintf(buf, "%u\n", chip->member); \
318 static struct freq_attr throttle_attr_##name = __ATTR_RO(name) \
320 throttle_attr(unthrottle, reason[NO_THROTTLE]);
321 throttle_attr(powercap, reason[POWERCAP]);
322 throttle_attr(overtemp, reason[CPU_OVERTEMP]);
323 throttle_attr(supply_fault, reason[POWER_SUPPLY_FAILURE]);
324 throttle_attr(overcurrent, reason[OVERCURRENT]);
325 throttle_attr(occ_reset, reason[OCC_RESET_THROTTLE]);
326 throttle_attr(turbo_stat, throttle_turbo);
327 throttle_attr(sub_turbo_stat, throttle_sub_turbo);
329 static struct attribute *throttle_attrs[] = {
330 &throttle_attr_unthrottle.attr,
331 &throttle_attr_powercap.attr,
332 &throttle_attr_overtemp.attr,
333 &throttle_attr_supply_fault.attr,
334 &throttle_attr_overcurrent.attr,
335 &throttle_attr_occ_reset.attr,
336 &throttle_attr_turbo_stat.attr,
337 &throttle_attr_sub_turbo_stat.attr,
341 static const struct attribute_group throttle_attr_grp = {
342 .name = "throttle_stats",
343 .attrs = throttle_attrs,
346 /* Helper routines */
348 /* Access helpers to power mgt SPR */
350 static inline unsigned long get_pmspr(unsigned long sprn)
354 return mfspr(SPRN_PMCR);
357 return mfspr(SPRN_PMICR);
360 return mfspr(SPRN_PMSR);
365 static inline void set_pmspr(unsigned long sprn, unsigned long val)
369 mtspr(SPRN_PMCR, val);
373 mtspr(SPRN_PMICR, val);
380 * Use objects of this type to query/update
381 * pstates on a remote CPU via smp_call_function.
383 struct powernv_smp_call_data {
390 * powernv_read_cpu_freq: Reads the current frequency on this CPU.
392 * Called via smp_call_function.
394 * Note: The caller of the smp_call_function should pass an argument of
395 * the type 'struct powernv_smp_call_data *' along with this function.
397 * The current frequency on this CPU will be returned via
398 * ((struct powernv_smp_call_data *)arg)->freq;
400 static void powernv_read_cpu_freq(void *arg)
402 unsigned long pmspr_val;
404 struct powernv_smp_call_data *freq_data = arg;
406 pmspr_val = get_pmspr(SPRN_PMSR);
409 * The local pstate id corresponds bits 48..55 in the PMSR.
410 * Note: Watch out for the sign!
412 local_pstate_id = (pmspr_val >> 48) & 0xFF;
413 freq_data->pstate_id = local_pstate_id;
414 freq_data->freq = pstate_id_to_freq(freq_data->pstate_id);
416 pr_debug("cpu %d pmsr %016lX pstate_id %d frequency %d kHz\n",
417 raw_smp_processor_id(), pmspr_val, freq_data->pstate_id,
422 * powernv_cpufreq_get: Returns the CPU frequency as reported by the
423 * firmware for CPU 'cpu'. This value is reported through the sysfs
424 * file cpuinfo_cur_freq.
426 static unsigned int powernv_cpufreq_get(unsigned int cpu)
428 struct powernv_smp_call_data freq_data;
430 smp_call_function_any(cpu_sibling_mask(cpu), powernv_read_cpu_freq,
433 return freq_data.freq;
437 * set_pstate: Sets the pstate on this CPU.
439 * This is called via an smp_call_function.
441 * The caller must ensure that freq_data is of the type
442 * (struct powernv_smp_call_data *) and the pstate_id which needs to be set
443 * on this CPU should be present in freq_data->pstate_id.
445 static void set_pstate(void *data)
448 struct powernv_smp_call_data *freq_data = data;
449 unsigned long pstate_ul = freq_data->pstate_id;
450 unsigned long gpstate_ul = freq_data->gpstate_id;
452 val = get_pmspr(SPRN_PMCR);
453 val = val & 0x0000FFFFFFFFFFFFULL;
455 pstate_ul = pstate_ul & 0xFF;
456 gpstate_ul = gpstate_ul & 0xFF;
458 /* Set both global(bits 56..63) and local(bits 48..55) PStates */
459 val = val | (gpstate_ul << 56) | (pstate_ul << 48);
461 pr_debug("Setting cpu %d pmcr to %016lX\n",
462 raw_smp_processor_id(), val);
463 set_pmspr(SPRN_PMCR, val);
467 * get_nominal_index: Returns the index corresponding to the nominal
468 * pstate in the cpufreq table
470 static inline unsigned int get_nominal_index(void)
472 return powernv_pstate_info.nominal;
475 static void powernv_cpufreq_throttle_check(void *data)
478 unsigned int cpu = smp_processor_id();
481 unsigned int pmsr_pmax_idx;
483 pmsr = get_pmspr(SPRN_PMSR);
484 chip = this_cpu_read(chip_info);
486 /* Check for Pmax Capping */
487 pmsr_pmax = (s8)PMSR_MAX(pmsr);
488 pmsr_pmax_idx = pstate_to_idx(pmsr_pmax);
489 if (pmsr_pmax_idx != powernv_pstate_info.max) {
492 chip->throttled = true;
493 if (pmsr_pmax_idx > powernv_pstate_info.nominal) {
494 pr_warn_once("CPU %d on Chip %u has Pmax(%d) reduced below nominal frequency(%d)\n",
495 cpu, chip->id, pmsr_pmax,
496 idx_to_pstate(powernv_pstate_info.nominal));
497 chip->throttle_sub_turbo++;
499 chip->throttle_turbo++;
501 trace_powernv_throttle(chip->id,
502 throttle_reason[chip->throttle_reason],
504 } else if (chip->throttled) {
505 chip->throttled = false;
506 trace_powernv_throttle(chip->id,
507 throttle_reason[chip->throttle_reason],
511 /* Check if Psafe_mode_active is set in PMSR. */
513 if (pmsr & PMSR_PSAFE_ENABLE) {
515 pr_info("Pstate set to safe frequency\n");
518 /* Check if SPR_EM_DISABLE is set in PMSR */
519 if (pmsr & PMSR_SPR_EM_DISABLE) {
521 pr_info("Frequency Control disabled from OS\n");
525 pr_info("PMSR = %16lx\n", pmsr);
526 pr_warn("CPU Frequency could be throttled\n");
531 * calc_global_pstate - Calculate global pstate
532 * @elapsed_time: Elapsed time in milliseconds
533 * @local_pstate_idx: New local pstate
534 * @highest_lpstate_idx: pstate from which its ramping down
536 * Finds the appropriate global pstate based on the pstate from which its
537 * ramping down and the time elapsed in ramping down. It follows a quadratic
538 * equation which ensures that it reaches ramping down to pmin in 5sec.
540 static inline int calc_global_pstate(unsigned int elapsed_time,
541 int highest_lpstate_idx,
542 int local_pstate_idx)
547 * Using ramp_down_percent we get the percentage of rampdown
548 * that we are expecting to be dropping. Difference between
549 * highest_lpstate_idx and powernv_pstate_info.min will give a absolute
550 * number of how many pstates we will drop eventually by the end of
551 * 5 seconds, then just scale it get the number pstates to be dropped.
553 index_diff = ((int)ramp_down_percent(elapsed_time) *
554 (powernv_pstate_info.min - highest_lpstate_idx)) / 100;
556 /* Ensure that global pstate is >= to local pstate */
557 if (highest_lpstate_idx + index_diff >= local_pstate_idx)
558 return local_pstate_idx;
560 return highest_lpstate_idx + index_diff;
563 static inline void queue_gpstate_timer(struct global_pstate_info *gpstates)
565 unsigned int timer_interval;
568 * Setting up timer to fire after GPSTATE_TIMER_INTERVAL ms, But
569 * if it exceeds MAX_RAMP_DOWN_TIME ms for ramp down time.
570 * Set timer such that it fires exactly at MAX_RAMP_DOWN_TIME
571 * seconds of ramp down time.
573 if ((gpstates->elapsed_time + GPSTATE_TIMER_INTERVAL)
574 > MAX_RAMP_DOWN_TIME)
575 timer_interval = MAX_RAMP_DOWN_TIME - gpstates->elapsed_time;
577 timer_interval = GPSTATE_TIMER_INTERVAL;
579 mod_timer_pinned(&gpstates->timer, jiffies +
580 msecs_to_jiffies(timer_interval));
584 * gpstate_timer_handler
586 * @data: pointer to cpufreq_policy on which timer was queued
588 * This handler brings down the global pstate closer to the local pstate
589 * according quadratic equation. Queues a new timer if it is still not equal
592 void gpstate_timer_handler(unsigned long data)
594 struct cpufreq_policy *policy = (struct cpufreq_policy *)data;
595 struct global_pstate_info *gpstates = policy->driver_data;
597 unsigned int time_diff = jiffies_to_msecs(jiffies)
598 - gpstates->last_sampled_time;
599 struct powernv_smp_call_data freq_data;
601 if (!spin_trylock(&gpstates->gpstate_lock))
604 gpstates->last_sampled_time += time_diff;
605 gpstates->elapsed_time += time_diff;
606 freq_data.pstate_id = idx_to_pstate(gpstates->last_lpstate_idx);
608 if ((gpstates->last_gpstate_idx == gpstates->last_lpstate_idx) ||
609 (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME)) {
610 gpstate_idx = pstate_to_idx(freq_data.pstate_id);
611 reset_gpstates(policy);
612 gpstates->highest_lpstate_idx = gpstate_idx;
614 gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
615 gpstates->highest_lpstate_idx,
616 gpstates->last_lpstate_idx);
620 * If local pstate is equal to global pstate, rampdown is over
621 * So timer is not required to be queued.
623 if (gpstate_idx != gpstates->last_lpstate_idx)
624 queue_gpstate_timer(gpstates);
626 freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
627 gpstates->last_gpstate_idx = pstate_to_idx(freq_data.gpstate_id);
628 gpstates->last_lpstate_idx = pstate_to_idx(freq_data.pstate_id);
630 spin_unlock(&gpstates->gpstate_lock);
632 /* Timer may get migrated to a different cpu on cpu hot unplug */
633 smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
637 * powernv_cpufreq_target_index: Sets the frequency corresponding to
638 * the cpufreq table entry indexed by new_index on the cpus in the
641 static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
642 unsigned int new_index)
644 struct powernv_smp_call_data freq_data;
645 unsigned int cur_msec, gpstate_idx;
646 struct global_pstate_info *gpstates = policy->driver_data;
648 if (unlikely(rebooting) && new_index != get_nominal_index())
652 powernv_cpufreq_throttle_check(NULL);
654 cur_msec = jiffies_to_msecs(get_jiffies_64());
656 spin_lock(&gpstates->gpstate_lock);
657 freq_data.pstate_id = idx_to_pstate(new_index);
659 if (!gpstates->last_sampled_time) {
660 gpstate_idx = new_index;
661 gpstates->highest_lpstate_idx = new_index;
665 if (gpstates->last_gpstate_idx < new_index) {
666 gpstates->elapsed_time += cur_msec -
667 gpstates->last_sampled_time;
670 * If its has been ramping down for more than MAX_RAMP_DOWN_TIME
671 * we should be resetting all global pstate related data. Set it
672 * equal to local pstate to start fresh.
674 if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
675 reset_gpstates(policy);
676 gpstates->highest_lpstate_idx = new_index;
677 gpstate_idx = new_index;
679 /* Elaspsed_time is less than 5 seconds, continue to rampdown */
680 gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
681 gpstates->highest_lpstate_idx,
685 reset_gpstates(policy);
686 gpstates->highest_lpstate_idx = new_index;
687 gpstate_idx = new_index;
691 * If local pstate is equal to global pstate, rampdown is over
692 * So timer is not required to be queued.
694 if (gpstate_idx != new_index)
695 queue_gpstate_timer(gpstates);
697 del_timer_sync(&gpstates->timer);
700 freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
701 gpstates->last_sampled_time = cur_msec;
702 gpstates->last_gpstate_idx = gpstate_idx;
703 gpstates->last_lpstate_idx = new_index;
705 spin_unlock(&gpstates->gpstate_lock);
708 * Use smp_call_function to send IPI and execute the
709 * mtspr on target CPU. We could do that without IPI
710 * if current CPU is within policy->cpus (core)
712 smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
716 static int powernv_cpufreq_cpu_init(struct cpufreq_policy *policy)
719 struct kernfs_node *kn;
720 struct global_pstate_info *gpstates;
722 base = cpu_first_thread_sibling(policy->cpu);
724 for (i = 0; i < threads_per_core; i++)
725 cpumask_set_cpu(base + i, policy->cpus);
727 kn = kernfs_find_and_get(policy->kobj.sd, throttle_attr_grp.name);
731 ret = sysfs_create_group(&policy->kobj, &throttle_attr_grp);
733 pr_info("Failed to create throttle stats directory for cpu %d\n",
741 gpstates = kzalloc(sizeof(*gpstates), GFP_KERNEL);
745 policy->driver_data = gpstates;
747 /* initialize timer */
748 init_timer_deferrable(&gpstates->timer);
749 gpstates->timer.data = (unsigned long)policy;
750 gpstates->timer.function = gpstate_timer_handler;
751 gpstates->timer.expires = jiffies +
752 msecs_to_jiffies(GPSTATE_TIMER_INTERVAL);
753 spin_lock_init(&gpstates->gpstate_lock);
754 ret = cpufreq_table_validate_and_show(policy, powernv_freqs);
757 kfree(policy->driver_data);
762 static int powernv_cpufreq_cpu_exit(struct cpufreq_policy *policy)
764 /* timer is deleted in cpufreq_cpu_stop() */
765 kfree(policy->driver_data);
770 static int powernv_cpufreq_reboot_notifier(struct notifier_block *nb,
771 unsigned long action, void *unused)
774 struct cpufreq_policy cpu_policy;
777 for_each_online_cpu(cpu) {
778 cpufreq_get_policy(&cpu_policy, cpu);
779 powernv_cpufreq_target_index(&cpu_policy, get_nominal_index());
785 static struct notifier_block powernv_cpufreq_reboot_nb = {
786 .notifier_call = powernv_cpufreq_reboot_notifier,
789 void powernv_cpufreq_work_fn(struct work_struct *work)
791 struct chip *chip = container_of(work, struct chip, throttle);
796 cpumask_and(&mask, &chip->mask, cpu_online_mask);
797 smp_call_function_any(&mask,
798 powernv_cpufreq_throttle_check, NULL, 0);
803 chip->restore = false;
804 for_each_cpu(cpu, &mask) {
806 struct cpufreq_policy policy;
808 cpufreq_get_policy(&policy, cpu);
809 index = cpufreq_table_find_index_c(&policy, policy.cur);
810 powernv_cpufreq_target_index(&policy, index);
811 cpumask_andnot(&mask, &mask, policy.cpus);
817 static int powernv_cpufreq_occ_msg(struct notifier_block *nb,
818 unsigned long msg_type, void *_msg)
820 struct opal_msg *msg = _msg;
821 struct opal_occ_msg omsg;
824 if (msg_type != OPAL_MSG_OCC)
827 omsg.type = be64_to_cpu(msg->params[0]);
832 pr_info("OCC (On Chip Controller - enforces hard thermal/power limits) Resetting\n");
834 * powernv_cpufreq_throttle_check() is called in
835 * target() callback which can detect the throttle state
836 * for governors like ondemand.
837 * But static governors will not call target() often thus
838 * report throttling here.
842 pr_warn("CPU frequency is throttled for duration\n");
847 pr_info("OCC Loading, CPU frequency is throttled until OCC is started\n");
850 omsg.chip = be64_to_cpu(msg->params[1]);
851 omsg.throttle_status = be64_to_cpu(msg->params[2]);
856 pr_info("OCC Active, CPU frequency is no longer throttled\n");
858 for (i = 0; i < nr_chips; i++) {
859 chips[i].restore = true;
860 schedule_work(&chips[i].throttle);
866 for (i = 0; i < nr_chips; i++)
867 if (chips[i].id == omsg.chip)
870 if (omsg.throttle_status >= 0 &&
871 omsg.throttle_status <= OCC_MAX_THROTTLE_STATUS) {
872 chips[i].throttle_reason = omsg.throttle_status;
873 chips[i].reason[omsg.throttle_status]++;
876 if (!omsg.throttle_status)
877 chips[i].restore = true;
879 schedule_work(&chips[i].throttle);
884 static struct notifier_block powernv_cpufreq_opal_nb = {
885 .notifier_call = powernv_cpufreq_occ_msg,
890 static void powernv_cpufreq_stop_cpu(struct cpufreq_policy *policy)
892 struct powernv_smp_call_data freq_data;
893 struct global_pstate_info *gpstates = policy->driver_data;
895 freq_data.pstate_id = idx_to_pstate(powernv_pstate_info.min);
896 freq_data.gpstate_id = idx_to_pstate(powernv_pstate_info.min);
897 smp_call_function_single(policy->cpu, set_pstate, &freq_data, 1);
898 del_timer_sync(&gpstates->timer);
901 static struct cpufreq_driver powernv_cpufreq_driver = {
902 .name = "powernv-cpufreq",
903 .flags = CPUFREQ_CONST_LOOPS,
904 .init = powernv_cpufreq_cpu_init,
905 .exit = powernv_cpufreq_cpu_exit,
906 .verify = cpufreq_generic_frequency_table_verify,
907 .target_index = powernv_cpufreq_target_index,
908 .get = powernv_cpufreq_get,
909 .stop_cpu = powernv_cpufreq_stop_cpu,
910 .attr = powernv_cpu_freq_attr,
913 static int init_chip_info(void)
915 unsigned int chip[256];
917 unsigned int prev_chip_id = UINT_MAX;
919 for_each_possible_cpu(cpu) {
920 unsigned int id = cpu_to_chip_id(cpu);
922 if (prev_chip_id != id) {
924 chip[nr_chips++] = id;
928 chips = kcalloc(nr_chips, sizeof(struct chip), GFP_KERNEL);
932 for (i = 0; i < nr_chips; i++) {
933 chips[i].id = chip[i];
934 cpumask_copy(&chips[i].mask, cpumask_of_node(chip[i]));
935 INIT_WORK(&chips[i].throttle, powernv_cpufreq_work_fn);
936 for_each_cpu(cpu, &chips[i].mask)
937 per_cpu(chip_info, cpu) = &chips[i];
943 static inline void clean_chip_info(void)
948 static inline void unregister_all_notifiers(void)
950 opal_message_notifier_unregister(OPAL_MSG_OCC,
951 &powernv_cpufreq_opal_nb);
952 unregister_reboot_notifier(&powernv_cpufreq_reboot_nb);
955 static int __init powernv_cpufreq_init(void)
959 /* Don't probe on pseries (guest) platforms */
960 if (!firmware_has_feature(FW_FEATURE_OPAL))
963 /* Discover pstates from device tree and init */
964 rc = init_powernv_pstates();
968 /* Populate chip info */
969 rc = init_chip_info();
973 register_reboot_notifier(&powernv_cpufreq_reboot_nb);
974 opal_message_notifier_register(OPAL_MSG_OCC, &powernv_cpufreq_opal_nb);
976 rc = cpufreq_register_driver(&powernv_cpufreq_driver);
980 pr_info("Failed to register the cpufreq driver (%d)\n", rc);
981 unregister_all_notifiers();
984 pr_info("Platform driver disabled. System does not support PState control\n");
987 module_init(powernv_cpufreq_init);
989 static void __exit powernv_cpufreq_exit(void)
991 cpufreq_unregister_driver(&powernv_cpufreq_driver);
992 unregister_all_notifiers();
995 module_exit(powernv_cpufreq_exit);
997 MODULE_LICENSE("GPL");
998 MODULE_AUTHOR("Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>");