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)
45 #define LPSTATE_SHIFT 48
46 #define GPSTATE_SHIFT 56
47 #define GET_LPSTATE(x) (((x) >> LPSTATE_SHIFT) & 0xFF)
48 #define GET_GPSTATE(x) (((x) >> GPSTATE_SHIFT) & 0xFF)
50 #define MAX_RAMP_DOWN_TIME 5120
52 * On an idle system we want the global pstate to ramp-down from max value to
53 * min over a span of ~5 secs. Also we want it to initially ramp-down slowly and
54 * then ramp-down rapidly later on.
56 * This gives a percentage rampdown for time elapsed in milliseconds.
57 * ramp_down_percentage = ((ms * ms) >> 18)
58 * ~= 3.8 * (sec * sec)
60 * At 0 ms ramp_down_percent = 0
61 * At 5120 ms ramp_down_percent = 100
63 #define ramp_down_percent(time) ((time * time) >> 18)
65 /* Interval after which the timer is queued to bring down global pstate */
66 #define GPSTATE_TIMER_INTERVAL 2000
69 * struct global_pstate_info - Per policy data structure to maintain history of
71 * @highest_lpstate_idx: The local pstate index from which we are
73 * @elapsed_time: Time in ms spent in ramping down from
75 * @last_sampled_time: Time from boot in ms when global pstates were
77 * @last_lpstate_idx, Last set value of local pstate and global
78 * last_gpstate_idx pstate in terms of cpufreq table index
79 * @timer: Is used for ramping down if cpu goes idle for
80 * a long time with global pstate held high
81 * @gpstate_lock: A spinlock to maintain synchronization between
82 * routines called by the timer handler and
83 * governer's target_index calls
85 struct global_pstate_info {
86 int highest_lpstate_idx;
87 unsigned int elapsed_time;
88 unsigned int last_sampled_time;
91 spinlock_t gpstate_lock;
92 struct timer_list timer;
93 struct cpufreq_policy *policy;
96 static struct cpufreq_frequency_table powernv_freqs[POWERNV_MAX_PSTATES+1];
97 static bool rebooting, throttled, occ_reset;
99 static const char * const throttle_reason[] = {
102 "Processor Over Temperature",
103 "Power Supply Failure",
108 enum throttle_reason_type {
112 POWER_SUPPLY_FAILURE,
124 struct work_struct throttle;
126 int throttle_sub_turbo;
127 int reason[OCC_MAX_REASON];
131 static DEFINE_PER_CPU(struct chip *, chip_info);
135 * The set of pstates consists of contiguous integers.
136 * powernv_pstate_info stores the index of the frequency table for
137 * max, min and nominal frequencies. It also stores number of
138 * available frequencies.
140 * powernv_pstate_info.nominal indicates the index to the highest
141 * non-turbo frequency.
143 static struct powernv_pstate_info {
146 unsigned int nominal;
147 unsigned int nr_pstates;
149 } powernv_pstate_info;
151 /* Use following macros for conversions between pstate_id and index */
152 static inline int idx_to_pstate(unsigned int i)
154 if (unlikely(i >= powernv_pstate_info.nr_pstates)) {
155 pr_warn_once("index %u is out of bound\n", i);
156 return powernv_freqs[powernv_pstate_info.nominal].driver_data;
159 return powernv_freqs[i].driver_data;
162 static inline unsigned int pstate_to_idx(int pstate)
164 int min = powernv_freqs[powernv_pstate_info.min].driver_data;
165 int max = powernv_freqs[powernv_pstate_info.max].driver_data;
168 if (unlikely((pstate < max) || (pstate > min))) {
169 pr_warn_once("pstate %d is out of bound\n", pstate);
170 return powernv_pstate_info.nominal;
173 if (unlikely((pstate > max) || (pstate < min))) {
174 pr_warn_once("pstate %d is out of bound\n", pstate);
175 return powernv_pstate_info.nominal;
179 * abs() is deliberately used so that is works with
180 * both monotonically increasing and decreasing
183 return abs(pstate - idx_to_pstate(powernv_pstate_info.max));
186 static inline void reset_gpstates(struct cpufreq_policy *policy)
188 struct global_pstate_info *gpstates = policy->driver_data;
190 gpstates->highest_lpstate_idx = 0;
191 gpstates->elapsed_time = 0;
192 gpstates->last_sampled_time = 0;
193 gpstates->last_lpstate_idx = 0;
194 gpstates->last_gpstate_idx = 0;
198 * Initialize the freq table based on data obtained
199 * from the firmware passed via device-tree
201 static int init_powernv_pstates(void)
203 struct device_node *power_mgt;
204 int i, nr_pstates = 0;
205 const __be32 *pstate_ids, *pstate_freqs;
206 u32 len_ids, len_freqs;
207 u32 pstate_min, pstate_max, pstate_nominal;
208 u32 pstate_turbo, pstate_ultra_turbo;
210 power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
212 pr_warn("power-mgt node not found\n");
216 if (of_property_read_u32(power_mgt, "ibm,pstate-min", &pstate_min)) {
217 pr_warn("ibm,pstate-min node not found\n");
221 if (of_property_read_u32(power_mgt, "ibm,pstate-max", &pstate_max)) {
222 pr_warn("ibm,pstate-max node not found\n");
226 if (of_property_read_u32(power_mgt, "ibm,pstate-nominal",
228 pr_warn("ibm,pstate-nominal not found\n");
232 if (of_property_read_u32(power_mgt, "ibm,pstate-ultra-turbo",
233 &pstate_ultra_turbo)) {
234 powernv_pstate_info.wof_enabled = false;
238 if (of_property_read_u32(power_mgt, "ibm,pstate-turbo",
240 powernv_pstate_info.wof_enabled = false;
244 if (pstate_turbo == pstate_ultra_turbo)
245 powernv_pstate_info.wof_enabled = false;
247 powernv_pstate_info.wof_enabled = true;
250 pr_info("cpufreq pstate min %d nominal %d max %d\n", pstate_min,
251 pstate_nominal, pstate_max);
252 pr_info("Workload Optimized Frequency is %s in the platform\n",
253 (powernv_pstate_info.wof_enabled) ? "enabled" : "disabled");
255 pstate_ids = of_get_property(power_mgt, "ibm,pstate-ids", &len_ids);
257 pr_warn("ibm,pstate-ids not found\n");
261 pstate_freqs = of_get_property(power_mgt, "ibm,pstate-frequencies-mhz",
264 pr_warn("ibm,pstate-frequencies-mhz not found\n");
268 if (len_ids != len_freqs) {
269 pr_warn("Entries in ibm,pstate-ids and "
270 "ibm,pstate-frequencies-mhz does not match\n");
273 nr_pstates = min(len_ids, len_freqs) / sizeof(u32);
275 pr_warn("No PStates found\n");
279 powernv_pstate_info.nr_pstates = nr_pstates;
280 pr_debug("NR PStates %d\n", nr_pstates);
281 for (i = 0; i < nr_pstates; i++) {
282 u32 id = be32_to_cpu(pstate_ids[i]);
283 u32 freq = be32_to_cpu(pstate_freqs[i]);
285 pr_debug("PState id %d freq %d MHz\n", id, freq);
286 powernv_freqs[i].frequency = freq * 1000; /* kHz */
287 powernv_freqs[i].driver_data = id;
289 if (id == pstate_max)
290 powernv_pstate_info.max = i;
291 else if (id == pstate_nominal)
292 powernv_pstate_info.nominal = i;
293 else if (id == pstate_min)
294 powernv_pstate_info.min = i;
296 if (powernv_pstate_info.wof_enabled && id == pstate_turbo) {
299 for (j = i - 1; j >= (int)powernv_pstate_info.max; j--)
300 powernv_freqs[j].flags = CPUFREQ_BOOST_FREQ;
304 /* End of list marker entry */
305 powernv_freqs[i].frequency = CPUFREQ_TABLE_END;
309 /* Returns the CPU frequency corresponding to the pstate_id. */
310 static unsigned int pstate_id_to_freq(int pstate_id)
314 i = pstate_to_idx(pstate_id);
315 if (i >= powernv_pstate_info.nr_pstates || i < 0) {
316 pr_warn("PState id %d outside of PState table, "
317 "reporting nominal id %d instead\n",
318 pstate_id, idx_to_pstate(powernv_pstate_info.nominal));
319 i = powernv_pstate_info.nominal;
322 return powernv_freqs[i].frequency;
326 * cpuinfo_nominal_freq_show - Show the nominal CPU frequency as indicated by
329 static ssize_t cpuinfo_nominal_freq_show(struct cpufreq_policy *policy,
332 return sprintf(buf, "%u\n",
333 powernv_freqs[powernv_pstate_info.nominal].frequency);
336 struct freq_attr cpufreq_freq_attr_cpuinfo_nominal_freq =
337 __ATTR_RO(cpuinfo_nominal_freq);
339 #define SCALING_BOOST_FREQS_ATTR_INDEX 2
341 static struct freq_attr *powernv_cpu_freq_attr[] = {
342 &cpufreq_freq_attr_scaling_available_freqs,
343 &cpufreq_freq_attr_cpuinfo_nominal_freq,
344 &cpufreq_freq_attr_scaling_boost_freqs,
348 #define throttle_attr(name, member) \
349 static ssize_t name##_show(struct cpufreq_policy *policy, char *buf) \
351 struct chip *chip = per_cpu(chip_info, policy->cpu); \
353 return sprintf(buf, "%u\n", chip->member); \
356 static struct freq_attr throttle_attr_##name = __ATTR_RO(name) \
358 throttle_attr(unthrottle, reason[NO_THROTTLE]);
359 throttle_attr(powercap, reason[POWERCAP]);
360 throttle_attr(overtemp, reason[CPU_OVERTEMP]);
361 throttle_attr(supply_fault, reason[POWER_SUPPLY_FAILURE]);
362 throttle_attr(overcurrent, reason[OVERCURRENT]);
363 throttle_attr(occ_reset, reason[OCC_RESET_THROTTLE]);
364 throttle_attr(turbo_stat, throttle_turbo);
365 throttle_attr(sub_turbo_stat, throttle_sub_turbo);
367 static struct attribute *throttle_attrs[] = {
368 &throttle_attr_unthrottle.attr,
369 &throttle_attr_powercap.attr,
370 &throttle_attr_overtemp.attr,
371 &throttle_attr_supply_fault.attr,
372 &throttle_attr_overcurrent.attr,
373 &throttle_attr_occ_reset.attr,
374 &throttle_attr_turbo_stat.attr,
375 &throttle_attr_sub_turbo_stat.attr,
379 static const struct attribute_group throttle_attr_grp = {
380 .name = "throttle_stats",
381 .attrs = throttle_attrs,
384 /* Helper routines */
386 /* Access helpers to power mgt SPR */
388 static inline unsigned long get_pmspr(unsigned long sprn)
392 return mfspr(SPRN_PMCR);
395 return mfspr(SPRN_PMICR);
398 return mfspr(SPRN_PMSR);
403 static inline void set_pmspr(unsigned long sprn, unsigned long val)
407 mtspr(SPRN_PMCR, val);
411 mtspr(SPRN_PMICR, val);
418 * Use objects of this type to query/update
419 * pstates on a remote CPU via smp_call_function.
421 struct powernv_smp_call_data {
428 * powernv_read_cpu_freq: Reads the current frequency on this CPU.
430 * Called via smp_call_function.
432 * Note: The caller of the smp_call_function should pass an argument of
433 * the type 'struct powernv_smp_call_data *' along with this function.
435 * The current frequency on this CPU will be returned via
436 * ((struct powernv_smp_call_data *)arg)->freq;
438 static void powernv_read_cpu_freq(void *arg)
440 unsigned long pmspr_val;
442 struct powernv_smp_call_data *freq_data = arg;
444 pmspr_val = get_pmspr(SPRN_PMSR);
447 * The local pstate id corresponds bits 48..55 in the PMSR.
448 * Note: Watch out for the sign!
450 local_pstate_id = (pmspr_val >> 48) & 0xFF;
451 freq_data->pstate_id = local_pstate_id;
452 freq_data->freq = pstate_id_to_freq(freq_data->pstate_id);
454 pr_debug("cpu %d pmsr %016lX pstate_id %d frequency %d kHz\n",
455 raw_smp_processor_id(), pmspr_val, freq_data->pstate_id,
460 * powernv_cpufreq_get: Returns the CPU frequency as reported by the
461 * firmware for CPU 'cpu'. This value is reported through the sysfs
462 * file cpuinfo_cur_freq.
464 static unsigned int powernv_cpufreq_get(unsigned int cpu)
466 struct powernv_smp_call_data freq_data;
468 smp_call_function_any(cpu_sibling_mask(cpu), powernv_read_cpu_freq,
471 return freq_data.freq;
475 * set_pstate: Sets the pstate on this CPU.
477 * This is called via an smp_call_function.
479 * The caller must ensure that freq_data is of the type
480 * (struct powernv_smp_call_data *) and the pstate_id which needs to be set
481 * on this CPU should be present in freq_data->pstate_id.
483 static void set_pstate(void *data)
486 struct powernv_smp_call_data *freq_data = data;
487 unsigned long pstate_ul = freq_data->pstate_id;
488 unsigned long gpstate_ul = freq_data->gpstate_id;
490 val = get_pmspr(SPRN_PMCR);
491 val = val & 0x0000FFFFFFFFFFFFULL;
493 pstate_ul = pstate_ul & 0xFF;
494 gpstate_ul = gpstate_ul & 0xFF;
496 /* Set both global(bits 56..63) and local(bits 48..55) PStates */
497 val = val | (gpstate_ul << 56) | (pstate_ul << 48);
499 pr_debug("Setting cpu %d pmcr to %016lX\n",
500 raw_smp_processor_id(), val);
501 set_pmspr(SPRN_PMCR, val);
505 * get_nominal_index: Returns the index corresponding to the nominal
506 * pstate in the cpufreq table
508 static inline unsigned int get_nominal_index(void)
510 return powernv_pstate_info.nominal;
513 static void powernv_cpufreq_throttle_check(void *data)
516 unsigned int cpu = smp_processor_id();
519 unsigned int pmsr_pmax_idx;
521 pmsr = get_pmspr(SPRN_PMSR);
522 chip = this_cpu_read(chip_info);
524 /* Check for Pmax Capping */
525 pmsr_pmax = (s8)PMSR_MAX(pmsr);
526 pmsr_pmax_idx = pstate_to_idx(pmsr_pmax);
527 if (pmsr_pmax_idx != powernv_pstate_info.max) {
530 chip->throttled = true;
531 if (pmsr_pmax_idx > powernv_pstate_info.nominal) {
532 pr_warn_once("CPU %d on Chip %u has Pmax(%d) reduced below nominal frequency(%d)\n",
533 cpu, chip->id, pmsr_pmax,
534 idx_to_pstate(powernv_pstate_info.nominal));
535 chip->throttle_sub_turbo++;
537 chip->throttle_turbo++;
539 trace_powernv_throttle(chip->id,
540 throttle_reason[chip->throttle_reason],
542 } else if (chip->throttled) {
543 chip->throttled = false;
544 trace_powernv_throttle(chip->id,
545 throttle_reason[chip->throttle_reason],
549 /* Check if Psafe_mode_active is set in PMSR. */
551 if (pmsr & PMSR_PSAFE_ENABLE) {
553 pr_info("Pstate set to safe frequency\n");
556 /* Check if SPR_EM_DISABLE is set in PMSR */
557 if (pmsr & PMSR_SPR_EM_DISABLE) {
559 pr_info("Frequency Control disabled from OS\n");
563 pr_info("PMSR = %16lx\n", pmsr);
564 pr_warn("CPU Frequency could be throttled\n");
569 * calc_global_pstate - Calculate global pstate
570 * @elapsed_time: Elapsed time in milliseconds
571 * @local_pstate_idx: New local pstate
572 * @highest_lpstate_idx: pstate from which its ramping down
574 * Finds the appropriate global pstate based on the pstate from which its
575 * ramping down and the time elapsed in ramping down. It follows a quadratic
576 * equation which ensures that it reaches ramping down to pmin in 5sec.
578 static inline int calc_global_pstate(unsigned int elapsed_time,
579 int highest_lpstate_idx,
580 int local_pstate_idx)
585 * Using ramp_down_percent we get the percentage of rampdown
586 * that we are expecting to be dropping. Difference between
587 * highest_lpstate_idx and powernv_pstate_info.min will give a absolute
588 * number of how many pstates we will drop eventually by the end of
589 * 5 seconds, then just scale it get the number pstates to be dropped.
591 index_diff = ((int)ramp_down_percent(elapsed_time) *
592 (powernv_pstate_info.min - highest_lpstate_idx)) / 100;
594 /* Ensure that global pstate is >= to local pstate */
595 if (highest_lpstate_idx + index_diff >= local_pstate_idx)
596 return local_pstate_idx;
598 return highest_lpstate_idx + index_diff;
601 static inline void queue_gpstate_timer(struct global_pstate_info *gpstates)
603 unsigned int timer_interval;
606 * Setting up timer to fire after GPSTATE_TIMER_INTERVAL ms, But
607 * if it exceeds MAX_RAMP_DOWN_TIME ms for ramp down time.
608 * Set timer such that it fires exactly at MAX_RAMP_DOWN_TIME
609 * seconds of ramp down time.
611 if ((gpstates->elapsed_time + GPSTATE_TIMER_INTERVAL)
612 > MAX_RAMP_DOWN_TIME)
613 timer_interval = MAX_RAMP_DOWN_TIME - gpstates->elapsed_time;
615 timer_interval = GPSTATE_TIMER_INTERVAL;
617 mod_timer(&gpstates->timer, jiffies + msecs_to_jiffies(timer_interval));
621 * gpstate_timer_handler
623 * @data: pointer to cpufreq_policy on which timer was queued
625 * This handler brings down the global pstate closer to the local pstate
626 * according quadratic equation. Queues a new timer if it is still not equal
629 void gpstate_timer_handler(struct timer_list *t)
631 struct global_pstate_info *gpstates = from_timer(gpstates, t, timer);
632 struct cpufreq_policy *policy = gpstates->policy;
633 int gpstate_idx, lpstate_idx;
635 unsigned int time_diff = jiffies_to_msecs(jiffies)
636 - gpstates->last_sampled_time;
637 struct powernv_smp_call_data freq_data;
639 if (!spin_trylock(&gpstates->gpstate_lock))
643 * If PMCR was last updated was using fast_swtich then
644 * We may have wrong in gpstate->last_lpstate_idx
645 * value. Hence, read from PMCR to get correct data.
647 val = get_pmspr(SPRN_PMCR);
648 freq_data.gpstate_id = (s8)GET_GPSTATE(val);
649 freq_data.pstate_id = (s8)GET_LPSTATE(val);
650 if (freq_data.gpstate_id == freq_data.pstate_id) {
651 reset_gpstates(policy);
652 spin_unlock(&gpstates->gpstate_lock);
656 gpstates->last_sampled_time += time_diff;
657 gpstates->elapsed_time += time_diff;
659 if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
660 gpstate_idx = pstate_to_idx(freq_data.pstate_id);
661 lpstate_idx = gpstate_idx;
662 reset_gpstates(policy);
663 gpstates->highest_lpstate_idx = gpstate_idx;
665 lpstate_idx = pstate_to_idx(freq_data.pstate_id);
666 gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
667 gpstates->highest_lpstate_idx,
670 freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
671 gpstates->last_gpstate_idx = gpstate_idx;
672 gpstates->last_lpstate_idx = lpstate_idx;
674 * If local pstate is equal to global pstate, rampdown is over
675 * So timer is not required to be queued.
677 if (gpstate_idx != gpstates->last_lpstate_idx)
678 queue_gpstate_timer(gpstates);
680 spin_unlock(&gpstates->gpstate_lock);
682 /* Timer may get migrated to a different cpu on cpu hot unplug */
683 smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
687 * powernv_cpufreq_target_index: Sets the frequency corresponding to
688 * the cpufreq table entry indexed by new_index on the cpus in the
691 static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
692 unsigned int new_index)
694 struct powernv_smp_call_data freq_data;
695 unsigned int cur_msec, gpstate_idx;
696 struct global_pstate_info *gpstates = policy->driver_data;
698 if (unlikely(rebooting) && new_index != get_nominal_index())
702 /* we don't want to be preempted while
703 * checking if the CPU frequency has been throttled
706 powernv_cpufreq_throttle_check(NULL);
710 cur_msec = jiffies_to_msecs(get_jiffies_64());
712 spin_lock(&gpstates->gpstate_lock);
713 freq_data.pstate_id = idx_to_pstate(new_index);
715 if (!gpstates->last_sampled_time) {
716 gpstate_idx = new_index;
717 gpstates->highest_lpstate_idx = new_index;
721 if (gpstates->last_gpstate_idx < new_index) {
722 gpstates->elapsed_time += cur_msec -
723 gpstates->last_sampled_time;
726 * If its has been ramping down for more than MAX_RAMP_DOWN_TIME
727 * we should be resetting all global pstate related data. Set it
728 * equal to local pstate to start fresh.
730 if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
731 reset_gpstates(policy);
732 gpstates->highest_lpstate_idx = new_index;
733 gpstate_idx = new_index;
735 /* Elaspsed_time is less than 5 seconds, continue to rampdown */
736 gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
737 gpstates->highest_lpstate_idx,
741 reset_gpstates(policy);
742 gpstates->highest_lpstate_idx = new_index;
743 gpstate_idx = new_index;
747 * If local pstate is equal to global pstate, rampdown is over
748 * So timer is not required to be queued.
750 if (gpstate_idx != new_index)
751 queue_gpstate_timer(gpstates);
753 del_timer_sync(&gpstates->timer);
756 freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
757 gpstates->last_sampled_time = cur_msec;
758 gpstates->last_gpstate_idx = gpstate_idx;
759 gpstates->last_lpstate_idx = new_index;
761 spin_unlock(&gpstates->gpstate_lock);
764 * Use smp_call_function to send IPI and execute the
765 * mtspr on target CPU. We could do that without IPI
766 * if current CPU is within policy->cpus (core)
768 smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
772 static int powernv_cpufreq_cpu_init(struct cpufreq_policy *policy)
775 struct kernfs_node *kn;
776 struct global_pstate_info *gpstates;
778 base = cpu_first_thread_sibling(policy->cpu);
780 for (i = 0; i < threads_per_core; i++)
781 cpumask_set_cpu(base + i, policy->cpus);
783 kn = kernfs_find_and_get(policy->kobj.sd, throttle_attr_grp.name);
787 ret = sysfs_create_group(&policy->kobj, &throttle_attr_grp);
789 pr_info("Failed to create throttle stats directory for cpu %d\n",
797 gpstates = kzalloc(sizeof(*gpstates), GFP_KERNEL);
801 policy->driver_data = gpstates;
803 /* initialize timer */
804 gpstates->policy = policy;
805 timer_setup(&gpstates->timer, gpstate_timer_handler,
806 TIMER_PINNED | TIMER_DEFERRABLE);
807 gpstates->timer.expires = jiffies +
808 msecs_to_jiffies(GPSTATE_TIMER_INTERVAL);
809 spin_lock_init(&gpstates->gpstate_lock);
810 ret = cpufreq_table_validate_and_show(policy, powernv_freqs);
813 kfree(policy->driver_data);
817 policy->fast_switch_possible = true;
821 static int powernv_cpufreq_cpu_exit(struct cpufreq_policy *policy)
823 /* timer is deleted in cpufreq_cpu_stop() */
824 kfree(policy->driver_data);
829 static int powernv_cpufreq_reboot_notifier(struct notifier_block *nb,
830 unsigned long action, void *unused)
833 struct cpufreq_policy cpu_policy;
836 for_each_online_cpu(cpu) {
837 cpufreq_get_policy(&cpu_policy, cpu);
838 powernv_cpufreq_target_index(&cpu_policy, get_nominal_index());
844 static struct notifier_block powernv_cpufreq_reboot_nb = {
845 .notifier_call = powernv_cpufreq_reboot_notifier,
848 void powernv_cpufreq_work_fn(struct work_struct *work)
850 struct chip *chip = container_of(work, struct chip, throttle);
855 cpumask_and(&mask, &chip->mask, cpu_online_mask);
856 smp_call_function_any(&mask,
857 powernv_cpufreq_throttle_check, NULL, 0);
862 chip->restore = false;
863 for_each_cpu(cpu, &mask) {
865 struct cpufreq_policy policy;
867 cpufreq_get_policy(&policy, cpu);
868 index = cpufreq_table_find_index_c(&policy, policy.cur);
869 powernv_cpufreq_target_index(&policy, index);
870 cpumask_andnot(&mask, &mask, policy.cpus);
876 static int powernv_cpufreq_occ_msg(struct notifier_block *nb,
877 unsigned long msg_type, void *_msg)
879 struct opal_msg *msg = _msg;
880 struct opal_occ_msg omsg;
883 if (msg_type != OPAL_MSG_OCC)
886 omsg.type = be64_to_cpu(msg->params[0]);
891 pr_info("OCC (On Chip Controller - enforces hard thermal/power limits) Resetting\n");
893 * powernv_cpufreq_throttle_check() is called in
894 * target() callback which can detect the throttle state
895 * for governors like ondemand.
896 * But static governors will not call target() often thus
897 * report throttling here.
901 pr_warn("CPU frequency is throttled for duration\n");
906 pr_info("OCC Loading, CPU frequency is throttled until OCC is started\n");
909 omsg.chip = be64_to_cpu(msg->params[1]);
910 omsg.throttle_status = be64_to_cpu(msg->params[2]);
915 pr_info("OCC Active, CPU frequency is no longer throttled\n");
917 for (i = 0; i < nr_chips; i++) {
918 chips[i].restore = true;
919 schedule_work(&chips[i].throttle);
925 for (i = 0; i < nr_chips; i++)
926 if (chips[i].id == omsg.chip)
929 if (omsg.throttle_status >= 0 &&
930 omsg.throttle_status <= OCC_MAX_THROTTLE_STATUS) {
931 chips[i].throttle_reason = omsg.throttle_status;
932 chips[i].reason[omsg.throttle_status]++;
935 if (!omsg.throttle_status)
936 chips[i].restore = true;
938 schedule_work(&chips[i].throttle);
943 static struct notifier_block powernv_cpufreq_opal_nb = {
944 .notifier_call = powernv_cpufreq_occ_msg,
949 static void powernv_cpufreq_stop_cpu(struct cpufreq_policy *policy)
951 struct powernv_smp_call_data freq_data;
952 struct global_pstate_info *gpstates = policy->driver_data;
954 freq_data.pstate_id = idx_to_pstate(powernv_pstate_info.min);
955 freq_data.gpstate_id = idx_to_pstate(powernv_pstate_info.min);
956 smp_call_function_single(policy->cpu, set_pstate, &freq_data, 1);
957 del_timer_sync(&gpstates->timer);
960 static unsigned int powernv_fast_switch(struct cpufreq_policy *policy,
961 unsigned int target_freq)
964 struct powernv_smp_call_data freq_data;
966 index = cpufreq_table_find_index_dl(policy, target_freq);
967 freq_data.pstate_id = powernv_freqs[index].driver_data;
968 freq_data.gpstate_id = powernv_freqs[index].driver_data;
969 set_pstate(&freq_data);
971 return powernv_freqs[index].frequency;
974 static struct cpufreq_driver powernv_cpufreq_driver = {
975 .name = "powernv-cpufreq",
976 .flags = CPUFREQ_CONST_LOOPS,
977 .init = powernv_cpufreq_cpu_init,
978 .exit = powernv_cpufreq_cpu_exit,
979 .verify = cpufreq_generic_frequency_table_verify,
980 .target_index = powernv_cpufreq_target_index,
981 .fast_switch = powernv_fast_switch,
982 .get = powernv_cpufreq_get,
983 .stop_cpu = powernv_cpufreq_stop_cpu,
984 .attr = powernv_cpu_freq_attr,
987 static int init_chip_info(void)
989 unsigned int chip[256];
991 unsigned int prev_chip_id = UINT_MAX;
993 for_each_possible_cpu(cpu) {
994 unsigned int id = cpu_to_chip_id(cpu);
996 if (prev_chip_id != id) {
998 chip[nr_chips++] = id;
1002 chips = kcalloc(nr_chips, sizeof(struct chip), GFP_KERNEL);
1006 for (i = 0; i < nr_chips; i++) {
1007 chips[i].id = chip[i];
1008 cpumask_copy(&chips[i].mask, cpumask_of_node(chip[i]));
1009 INIT_WORK(&chips[i].throttle, powernv_cpufreq_work_fn);
1010 for_each_cpu(cpu, &chips[i].mask)
1011 per_cpu(chip_info, cpu) = &chips[i];
1017 static inline void clean_chip_info(void)
1022 static inline void unregister_all_notifiers(void)
1024 opal_message_notifier_unregister(OPAL_MSG_OCC,
1025 &powernv_cpufreq_opal_nb);
1026 unregister_reboot_notifier(&powernv_cpufreq_reboot_nb);
1029 static int __init powernv_cpufreq_init(void)
1033 /* Don't probe on pseries (guest) platforms */
1034 if (!firmware_has_feature(FW_FEATURE_OPAL))
1037 /* Discover pstates from device tree and init */
1038 rc = init_powernv_pstates();
1042 /* Populate chip info */
1043 rc = init_chip_info();
1047 register_reboot_notifier(&powernv_cpufreq_reboot_nb);
1048 opal_message_notifier_register(OPAL_MSG_OCC, &powernv_cpufreq_opal_nb);
1050 if (powernv_pstate_info.wof_enabled)
1051 powernv_cpufreq_driver.boost_enabled = true;
1053 powernv_cpu_freq_attr[SCALING_BOOST_FREQS_ATTR_INDEX] = NULL;
1055 rc = cpufreq_register_driver(&powernv_cpufreq_driver);
1057 pr_info("Failed to register the cpufreq driver (%d)\n", rc);
1058 goto cleanup_notifiers;
1061 if (powernv_pstate_info.wof_enabled)
1062 cpufreq_enable_boost_support();
1066 unregister_all_notifiers();
1069 pr_info("Platform driver disabled. System does not support PState control\n");
1072 module_init(powernv_cpufreq_init);
1074 static void __exit powernv_cpufreq_exit(void)
1076 cpufreq_unregister_driver(&powernv_cpufreq_driver);
1077 unregister_all_notifiers();
1080 module_exit(powernv_cpufreq_exit);
1082 MODULE_LICENSE("GPL");
1083 MODULE_AUTHOR("Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>");