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 <linux/hashtable.h>
33 #include <trace/events/power.h>
35 #include <asm/cputhreads.h>
36 #include <asm/firmware.h>
38 #include <asm/smp.h> /* Required for cpu_sibling_mask() in UP configs */
40 #include <linux/timer.h>
42 #define POWERNV_MAX_PSTATES_ORDER 8
43 #define POWERNV_MAX_PSTATES (1UL << (POWERNV_MAX_PSTATES_ORDER))
44 #define PMSR_PSAFE_ENABLE (1UL << 30)
45 #define PMSR_SPR_EM_DISABLE (1UL << 31)
46 #define MAX_PSTATE_SHIFT 32
47 #define LPSTATE_SHIFT 48
48 #define GPSTATE_SHIFT 56
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];
98 DEFINE_HASHTABLE(pstate_revmap, POWERNV_MAX_PSTATES_ORDER);
100 * struct pstate_idx_revmap_data: Entry in the hashmap pstate_revmap
101 * indexed by a function of pstate id.
103 * @pstate_id: pstate id for this entry.
105 * @cpufreq_table_idx: Index into the powernv_freqs
106 * cpufreq_frequency_table for frequency
107 * corresponding to pstate_id.
109 * @hentry: hlist_node that hooks this entry into the pstate_revmap
112 struct pstate_idx_revmap_data {
114 unsigned int cpufreq_table_idx;
115 struct hlist_node hentry;
118 static bool rebooting, throttled, occ_reset;
120 static const char * const throttle_reason[] = {
123 "Processor Over Temperature",
124 "Power Supply Failure",
129 enum throttle_reason_type {
133 POWER_SUPPLY_FAILURE,
145 struct work_struct throttle;
147 int throttle_sub_turbo;
148 int reason[OCC_MAX_REASON];
152 static DEFINE_PER_CPU(struct chip *, chip_info);
156 * The set of pstates consists of contiguous integers.
157 * powernv_pstate_info stores the index of the frequency table for
158 * max, min and nominal frequencies. It also stores number of
159 * available frequencies.
161 * powernv_pstate_info.nominal indicates the index to the highest
162 * non-turbo frequency.
164 static struct powernv_pstate_info {
167 unsigned int nominal;
168 unsigned int nr_pstates;
170 } powernv_pstate_info;
172 static inline u8 extract_pstate(u64 pmsr_val, unsigned int shift)
174 return ((pmsr_val >> shift) & 0xFF);
177 #define extract_local_pstate(x) extract_pstate(x, LPSTATE_SHIFT)
178 #define extract_global_pstate(x) extract_pstate(x, GPSTATE_SHIFT)
179 #define extract_max_pstate(x) extract_pstate(x, MAX_PSTATE_SHIFT)
181 /* Use following functions for conversions between pstate_id and index */
184 * idx_to_pstate : Returns the pstate id corresponding to the
185 * frequency in the cpufreq frequency table
186 * powernv_freqs indexed by @i.
188 * If @i is out of bound, this will return the pstate
189 * corresponding to the nominal frequency.
191 static inline u8 idx_to_pstate(unsigned int i)
193 if (unlikely(i >= powernv_pstate_info.nr_pstates)) {
194 pr_warn_once("idx_to_pstate: index %u is out of bound\n", i);
195 return powernv_freqs[powernv_pstate_info.nominal].driver_data;
198 return powernv_freqs[i].driver_data;
202 * pstate_to_idx : Returns the index in the cpufreq frequencytable
203 * powernv_freqs for the frequency whose corresponding
204 * pstate id is @pstate.
206 * If no frequency corresponding to @pstate is found,
207 * this will return the index of the nominal
210 static unsigned int pstate_to_idx(u8 pstate)
212 unsigned int key = pstate % POWERNV_MAX_PSTATES;
213 struct pstate_idx_revmap_data *revmap_data;
215 hash_for_each_possible(pstate_revmap, revmap_data, hentry, key) {
216 if (revmap_data->pstate_id == pstate)
217 return revmap_data->cpufreq_table_idx;
220 pr_warn_once("pstate_to_idx: pstate 0x%x not found\n", pstate);
221 return powernv_pstate_info.nominal;
224 static inline void reset_gpstates(struct cpufreq_policy *policy)
226 struct global_pstate_info *gpstates = policy->driver_data;
228 gpstates->highest_lpstate_idx = 0;
229 gpstates->elapsed_time = 0;
230 gpstates->last_sampled_time = 0;
231 gpstates->last_lpstate_idx = 0;
232 gpstates->last_gpstate_idx = 0;
236 * Initialize the freq table based on data obtained
237 * from the firmware passed via device-tree
239 static int init_powernv_pstates(void)
241 struct device_node *power_mgt;
242 int i, nr_pstates = 0;
243 const __be32 *pstate_ids, *pstate_freqs;
244 u32 len_ids, len_freqs;
245 u32 pstate_min, pstate_max, pstate_nominal;
246 u32 pstate_turbo, pstate_ultra_turbo;
249 power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
251 pr_warn("power-mgt node not found\n");
255 if (of_property_read_u32(power_mgt, "ibm,pstate-min", &pstate_min)) {
256 pr_warn("ibm,pstate-min node not found\n");
260 if (of_property_read_u32(power_mgt, "ibm,pstate-max", &pstate_max)) {
261 pr_warn("ibm,pstate-max node not found\n");
265 if (of_property_read_u32(power_mgt, "ibm,pstate-nominal",
267 pr_warn("ibm,pstate-nominal not found\n");
271 if (of_property_read_u32(power_mgt, "ibm,pstate-ultra-turbo",
272 &pstate_ultra_turbo)) {
273 powernv_pstate_info.wof_enabled = false;
277 if (of_property_read_u32(power_mgt, "ibm,pstate-turbo",
279 powernv_pstate_info.wof_enabled = false;
283 if (pstate_turbo == pstate_ultra_turbo)
284 powernv_pstate_info.wof_enabled = false;
286 powernv_pstate_info.wof_enabled = true;
289 pr_info("cpufreq pstate min 0x%x nominal 0x%x max 0x%x\n", pstate_min,
290 pstate_nominal, pstate_max);
291 pr_info("Workload Optimized Frequency is %s in the platform\n",
292 (powernv_pstate_info.wof_enabled) ? "enabled" : "disabled");
294 pstate_ids = of_get_property(power_mgt, "ibm,pstate-ids", &len_ids);
296 pr_warn("ibm,pstate-ids not found\n");
300 pstate_freqs = of_get_property(power_mgt, "ibm,pstate-frequencies-mhz",
303 pr_warn("ibm,pstate-frequencies-mhz not found\n");
307 if (len_ids != len_freqs) {
308 pr_warn("Entries in ibm,pstate-ids and "
309 "ibm,pstate-frequencies-mhz does not match\n");
312 nr_pstates = min(len_ids, len_freqs) / sizeof(u32);
314 pr_warn("No PStates found\n");
318 powernv_pstate_info.nr_pstates = nr_pstates;
319 pr_debug("NR PStates %d\n", nr_pstates);
321 for (i = 0; i < nr_pstates; i++) {
322 u32 id = be32_to_cpu(pstate_ids[i]);
323 u32 freq = be32_to_cpu(pstate_freqs[i]);
324 struct pstate_idx_revmap_data *revmap_data;
327 pr_debug("PState id %d freq %d MHz\n", id, freq);
328 powernv_freqs[i].frequency = freq * 1000; /* kHz */
329 powernv_freqs[i].driver_data = id & 0xFF;
331 revmap_data = kmalloc(sizeof(*revmap_data), GFP_KERNEL);
337 revmap_data->pstate_id = id & 0xFF;
338 revmap_data->cpufreq_table_idx = i;
339 key = (revmap_data->pstate_id) % POWERNV_MAX_PSTATES;
340 hash_add(pstate_revmap, &revmap_data->hentry, key);
342 if (id == pstate_max)
343 powernv_pstate_info.max = i;
344 if (id == pstate_nominal)
345 powernv_pstate_info.nominal = i;
346 if (id == pstate_min)
347 powernv_pstate_info.min = i;
349 if (powernv_pstate_info.wof_enabled && id == pstate_turbo) {
352 for (j = i - 1; j >= (int)powernv_pstate_info.max; j--)
353 powernv_freqs[j].flags = CPUFREQ_BOOST_FREQ;
357 /* End of list marker entry */
358 powernv_freqs[i].frequency = CPUFREQ_TABLE_END;
360 of_node_put(power_mgt);
363 of_node_put(power_mgt);
367 /* Returns the CPU frequency corresponding to the pstate_id. */
368 static unsigned int pstate_id_to_freq(u8 pstate_id)
372 i = pstate_to_idx(pstate_id);
373 if (i >= powernv_pstate_info.nr_pstates || i < 0) {
374 pr_warn("PState id 0x%x outside of PState table, reporting nominal id 0x%x instead\n",
375 pstate_id, idx_to_pstate(powernv_pstate_info.nominal));
376 i = powernv_pstate_info.nominal;
379 return powernv_freqs[i].frequency;
383 * cpuinfo_nominal_freq_show - Show the nominal CPU frequency as indicated by
386 static ssize_t cpuinfo_nominal_freq_show(struct cpufreq_policy *policy,
389 return sprintf(buf, "%u\n",
390 powernv_freqs[powernv_pstate_info.nominal].frequency);
393 struct freq_attr cpufreq_freq_attr_cpuinfo_nominal_freq =
394 __ATTR_RO(cpuinfo_nominal_freq);
396 #define SCALING_BOOST_FREQS_ATTR_INDEX 2
398 static struct freq_attr *powernv_cpu_freq_attr[] = {
399 &cpufreq_freq_attr_scaling_available_freqs,
400 &cpufreq_freq_attr_cpuinfo_nominal_freq,
401 &cpufreq_freq_attr_scaling_boost_freqs,
405 #define throttle_attr(name, member) \
406 static ssize_t name##_show(struct cpufreq_policy *policy, char *buf) \
408 struct chip *chip = per_cpu(chip_info, policy->cpu); \
410 return sprintf(buf, "%u\n", chip->member); \
413 static struct freq_attr throttle_attr_##name = __ATTR_RO(name) \
415 throttle_attr(unthrottle, reason[NO_THROTTLE]);
416 throttle_attr(powercap, reason[POWERCAP]);
417 throttle_attr(overtemp, reason[CPU_OVERTEMP]);
418 throttle_attr(supply_fault, reason[POWER_SUPPLY_FAILURE]);
419 throttle_attr(overcurrent, reason[OVERCURRENT]);
420 throttle_attr(occ_reset, reason[OCC_RESET_THROTTLE]);
421 throttle_attr(turbo_stat, throttle_turbo);
422 throttle_attr(sub_turbo_stat, throttle_sub_turbo);
424 static struct attribute *throttle_attrs[] = {
425 &throttle_attr_unthrottle.attr,
426 &throttle_attr_powercap.attr,
427 &throttle_attr_overtemp.attr,
428 &throttle_attr_supply_fault.attr,
429 &throttle_attr_overcurrent.attr,
430 &throttle_attr_occ_reset.attr,
431 &throttle_attr_turbo_stat.attr,
432 &throttle_attr_sub_turbo_stat.attr,
436 static const struct attribute_group throttle_attr_grp = {
437 .name = "throttle_stats",
438 .attrs = throttle_attrs,
441 /* Helper routines */
443 /* Access helpers to power mgt SPR */
445 static inline unsigned long get_pmspr(unsigned long sprn)
449 return mfspr(SPRN_PMCR);
452 return mfspr(SPRN_PMICR);
455 return mfspr(SPRN_PMSR);
460 static inline void set_pmspr(unsigned long sprn, unsigned long val)
464 mtspr(SPRN_PMCR, val);
468 mtspr(SPRN_PMICR, val);
475 * Use objects of this type to query/update
476 * pstates on a remote CPU via smp_call_function.
478 struct powernv_smp_call_data {
485 * powernv_read_cpu_freq: Reads the current frequency on this CPU.
487 * Called via smp_call_function.
489 * Note: The caller of the smp_call_function should pass an argument of
490 * the type 'struct powernv_smp_call_data *' along with this function.
492 * The current frequency on this CPU will be returned via
493 * ((struct powernv_smp_call_data *)arg)->freq;
495 static void powernv_read_cpu_freq(void *arg)
497 unsigned long pmspr_val;
498 struct powernv_smp_call_data *freq_data = arg;
500 pmspr_val = get_pmspr(SPRN_PMSR);
501 freq_data->pstate_id = extract_local_pstate(pmspr_val);
502 freq_data->freq = pstate_id_to_freq(freq_data->pstate_id);
504 pr_debug("cpu %d pmsr %016lX pstate_id 0x%x frequency %d kHz\n",
505 raw_smp_processor_id(), pmspr_val, freq_data->pstate_id,
510 * powernv_cpufreq_get: Returns the CPU frequency as reported by the
511 * firmware for CPU 'cpu'. This value is reported through the sysfs
512 * file cpuinfo_cur_freq.
514 static unsigned int powernv_cpufreq_get(unsigned int cpu)
516 struct powernv_smp_call_data freq_data;
518 smp_call_function_any(cpu_sibling_mask(cpu), powernv_read_cpu_freq,
521 return freq_data.freq;
525 * set_pstate: Sets the pstate on this CPU.
527 * This is called via an smp_call_function.
529 * The caller must ensure that freq_data is of the type
530 * (struct powernv_smp_call_data *) and the pstate_id which needs to be set
531 * on this CPU should be present in freq_data->pstate_id.
533 static void set_pstate(void *data)
536 struct powernv_smp_call_data *freq_data = data;
537 unsigned long pstate_ul = freq_data->pstate_id;
538 unsigned long gpstate_ul = freq_data->gpstate_id;
540 val = get_pmspr(SPRN_PMCR);
541 val = val & 0x0000FFFFFFFFFFFFULL;
543 pstate_ul = pstate_ul & 0xFF;
544 gpstate_ul = gpstate_ul & 0xFF;
546 /* Set both global(bits 56..63) and local(bits 48..55) PStates */
547 val = val | (gpstate_ul << 56) | (pstate_ul << 48);
549 pr_debug("Setting cpu %d pmcr to %016lX\n",
550 raw_smp_processor_id(), val);
551 set_pmspr(SPRN_PMCR, val);
555 * get_nominal_index: Returns the index corresponding to the nominal
556 * pstate in the cpufreq table
558 static inline unsigned int get_nominal_index(void)
560 return powernv_pstate_info.nominal;
563 static void powernv_cpufreq_throttle_check(void *data)
566 unsigned int cpu = smp_processor_id();
569 unsigned int pmsr_pmax_idx;
571 pmsr = get_pmspr(SPRN_PMSR);
572 chip = this_cpu_read(chip_info);
574 /* Check for Pmax Capping */
575 pmsr_pmax = extract_max_pstate(pmsr);
576 pmsr_pmax_idx = pstate_to_idx(pmsr_pmax);
577 if (pmsr_pmax_idx != powernv_pstate_info.max) {
580 chip->throttled = true;
581 if (pmsr_pmax_idx > powernv_pstate_info.nominal) {
582 pr_warn_once("CPU %d on Chip %u has Pmax(0x%x) reduced below that of nominal frequency(0x%x)\n",
583 cpu, chip->id, pmsr_pmax,
584 idx_to_pstate(powernv_pstate_info.nominal));
585 chip->throttle_sub_turbo++;
587 chip->throttle_turbo++;
589 trace_powernv_throttle(chip->id,
590 throttle_reason[chip->throttle_reason],
592 } else if (chip->throttled) {
593 chip->throttled = false;
594 trace_powernv_throttle(chip->id,
595 throttle_reason[chip->throttle_reason],
599 /* Check if Psafe_mode_active is set in PMSR. */
601 if (pmsr & PMSR_PSAFE_ENABLE) {
603 pr_info("Pstate set to safe frequency\n");
606 /* Check if SPR_EM_DISABLE is set in PMSR */
607 if (pmsr & PMSR_SPR_EM_DISABLE) {
609 pr_info("Frequency Control disabled from OS\n");
613 pr_info("PMSR = %16lx\n", pmsr);
614 pr_warn("CPU Frequency could be throttled\n");
619 * calc_global_pstate - Calculate global pstate
620 * @elapsed_time: Elapsed time in milliseconds
621 * @local_pstate_idx: New local pstate
622 * @highest_lpstate_idx: pstate from which its ramping down
624 * Finds the appropriate global pstate based on the pstate from which its
625 * ramping down and the time elapsed in ramping down. It follows a quadratic
626 * equation which ensures that it reaches ramping down to pmin in 5sec.
628 static inline int calc_global_pstate(unsigned int elapsed_time,
629 int highest_lpstate_idx,
630 int local_pstate_idx)
635 * Using ramp_down_percent we get the percentage of rampdown
636 * that we are expecting to be dropping. Difference between
637 * highest_lpstate_idx and powernv_pstate_info.min will give a absolute
638 * number of how many pstates we will drop eventually by the end of
639 * 5 seconds, then just scale it get the number pstates to be dropped.
641 index_diff = ((int)ramp_down_percent(elapsed_time) *
642 (powernv_pstate_info.min - highest_lpstate_idx)) / 100;
644 /* Ensure that global pstate is >= to local pstate */
645 if (highest_lpstate_idx + index_diff >= local_pstate_idx)
646 return local_pstate_idx;
648 return highest_lpstate_idx + index_diff;
651 static inline void queue_gpstate_timer(struct global_pstate_info *gpstates)
653 unsigned int timer_interval;
656 * Setting up timer to fire after GPSTATE_TIMER_INTERVAL ms, But
657 * if it exceeds MAX_RAMP_DOWN_TIME ms for ramp down time.
658 * Set timer such that it fires exactly at MAX_RAMP_DOWN_TIME
659 * seconds of ramp down time.
661 if ((gpstates->elapsed_time + GPSTATE_TIMER_INTERVAL)
662 > MAX_RAMP_DOWN_TIME)
663 timer_interval = MAX_RAMP_DOWN_TIME - gpstates->elapsed_time;
665 timer_interval = GPSTATE_TIMER_INTERVAL;
667 mod_timer(&gpstates->timer, jiffies + msecs_to_jiffies(timer_interval));
671 * gpstate_timer_handler
673 * @data: pointer to cpufreq_policy on which timer was queued
675 * This handler brings down the global pstate closer to the local pstate
676 * according quadratic equation. Queues a new timer if it is still not equal
679 void gpstate_timer_handler(struct timer_list *t)
681 struct global_pstate_info *gpstates = from_timer(gpstates, t, timer);
682 struct cpufreq_policy *policy = gpstates->policy;
683 int gpstate_idx, lpstate_idx;
685 unsigned int time_diff = jiffies_to_msecs(jiffies)
686 - gpstates->last_sampled_time;
687 struct powernv_smp_call_data freq_data;
689 if (!spin_trylock(&gpstates->gpstate_lock))
692 * If the timer has migrated to the different cpu then bring
693 * it back to one of the policy->cpus
695 if (!cpumask_test_cpu(raw_smp_processor_id(), policy->cpus)) {
696 gpstates->timer.expires = jiffies + msecs_to_jiffies(1);
697 add_timer_on(&gpstates->timer, cpumask_first(policy->cpus));
698 spin_unlock(&gpstates->gpstate_lock);
703 * If PMCR was last updated was using fast_swtich then
704 * We may have wrong in gpstate->last_lpstate_idx
705 * value. Hence, read from PMCR to get correct data.
707 val = get_pmspr(SPRN_PMCR);
708 freq_data.gpstate_id = extract_global_pstate(val);
709 freq_data.pstate_id = extract_local_pstate(val);
710 if (freq_data.gpstate_id == freq_data.pstate_id) {
711 reset_gpstates(policy);
712 spin_unlock(&gpstates->gpstate_lock);
716 gpstates->last_sampled_time += time_diff;
717 gpstates->elapsed_time += time_diff;
719 if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
720 gpstate_idx = pstate_to_idx(freq_data.pstate_id);
721 lpstate_idx = gpstate_idx;
722 reset_gpstates(policy);
723 gpstates->highest_lpstate_idx = gpstate_idx;
725 lpstate_idx = pstate_to_idx(freq_data.pstate_id);
726 gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
727 gpstates->highest_lpstate_idx,
730 freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
731 gpstates->last_gpstate_idx = gpstate_idx;
732 gpstates->last_lpstate_idx = lpstate_idx;
734 * If local pstate is equal to global pstate, rampdown is over
735 * So timer is not required to be queued.
737 if (gpstate_idx != gpstates->last_lpstate_idx)
738 queue_gpstate_timer(gpstates);
740 set_pstate(&freq_data);
741 spin_unlock(&gpstates->gpstate_lock);
745 * powernv_cpufreq_target_index: Sets the frequency corresponding to
746 * the cpufreq table entry indexed by new_index on the cpus in the
749 static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
750 unsigned int new_index)
752 struct powernv_smp_call_data freq_data;
753 unsigned int cur_msec, gpstate_idx;
754 struct global_pstate_info *gpstates = policy->driver_data;
756 if (unlikely(rebooting) && new_index != get_nominal_index())
760 /* we don't want to be preempted while
761 * checking if the CPU frequency has been throttled
764 powernv_cpufreq_throttle_check(NULL);
768 cur_msec = jiffies_to_msecs(get_jiffies_64());
770 freq_data.pstate_id = idx_to_pstate(new_index);
772 freq_data.gpstate_id = freq_data.pstate_id;
776 spin_lock(&gpstates->gpstate_lock);
778 if (!gpstates->last_sampled_time) {
779 gpstate_idx = new_index;
780 gpstates->highest_lpstate_idx = new_index;
784 if (gpstates->last_gpstate_idx < new_index) {
785 gpstates->elapsed_time += cur_msec -
786 gpstates->last_sampled_time;
789 * If its has been ramping down for more than MAX_RAMP_DOWN_TIME
790 * we should be resetting all global pstate related data. Set it
791 * equal to local pstate to start fresh.
793 if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
794 reset_gpstates(policy);
795 gpstates->highest_lpstate_idx = new_index;
796 gpstate_idx = new_index;
798 /* Elaspsed_time is less than 5 seconds, continue to rampdown */
799 gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
800 gpstates->highest_lpstate_idx,
804 reset_gpstates(policy);
805 gpstates->highest_lpstate_idx = new_index;
806 gpstate_idx = new_index;
810 * If local pstate is equal to global pstate, rampdown is over
811 * So timer is not required to be queued.
813 if (gpstate_idx != new_index)
814 queue_gpstate_timer(gpstates);
816 del_timer_sync(&gpstates->timer);
819 freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
820 gpstates->last_sampled_time = cur_msec;
821 gpstates->last_gpstate_idx = gpstate_idx;
822 gpstates->last_lpstate_idx = new_index;
824 spin_unlock(&gpstates->gpstate_lock);
828 * Use smp_call_function to send IPI and execute the
829 * mtspr on target CPU. We could do that without IPI
830 * if current CPU is within policy->cpus (core)
832 smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
836 static int powernv_cpufreq_cpu_init(struct cpufreq_policy *policy)
839 struct kernfs_node *kn;
840 struct global_pstate_info *gpstates;
842 base = cpu_first_thread_sibling(policy->cpu);
844 for (i = 0; i < threads_per_core; i++)
845 cpumask_set_cpu(base + i, policy->cpus);
847 kn = kernfs_find_and_get(policy->kobj.sd, throttle_attr_grp.name);
851 ret = sysfs_create_group(&policy->kobj, &throttle_attr_grp);
853 pr_info("Failed to create throttle stats directory for cpu %d\n",
861 policy->freq_table = powernv_freqs;
862 policy->fast_switch_possible = true;
864 if (pvr_version_is(PVR_POWER9))
867 /* Initialise Gpstate ramp-down timer only on POWER8 */
868 gpstates = kzalloc(sizeof(*gpstates), GFP_KERNEL);
872 policy->driver_data = gpstates;
874 /* initialize timer */
875 gpstates->policy = policy;
876 timer_setup(&gpstates->timer, gpstate_timer_handler,
877 TIMER_PINNED | TIMER_DEFERRABLE);
878 gpstates->timer.expires = jiffies +
879 msecs_to_jiffies(GPSTATE_TIMER_INTERVAL);
880 spin_lock_init(&gpstates->gpstate_lock);
885 static int powernv_cpufreq_cpu_exit(struct cpufreq_policy *policy)
887 /* timer is deleted in cpufreq_cpu_stop() */
888 kfree(policy->driver_data);
893 static int powernv_cpufreq_reboot_notifier(struct notifier_block *nb,
894 unsigned long action, void *unused)
897 struct cpufreq_policy cpu_policy;
900 for_each_online_cpu(cpu) {
901 cpufreq_get_policy(&cpu_policy, cpu);
902 powernv_cpufreq_target_index(&cpu_policy, get_nominal_index());
908 static struct notifier_block powernv_cpufreq_reboot_nb = {
909 .notifier_call = powernv_cpufreq_reboot_notifier,
912 void powernv_cpufreq_work_fn(struct work_struct *work)
914 struct chip *chip = container_of(work, struct chip, throttle);
919 cpumask_and(&mask, &chip->mask, cpu_online_mask);
920 smp_call_function_any(&mask,
921 powernv_cpufreq_throttle_check, NULL, 0);
926 chip->restore = false;
927 for_each_cpu(cpu, &mask) {
929 struct cpufreq_policy policy;
931 cpufreq_get_policy(&policy, cpu);
932 index = cpufreq_table_find_index_c(&policy, policy.cur);
933 powernv_cpufreq_target_index(&policy, index);
934 cpumask_andnot(&mask, &mask, policy.cpus);
940 static int powernv_cpufreq_occ_msg(struct notifier_block *nb,
941 unsigned long msg_type, void *_msg)
943 struct opal_msg *msg = _msg;
944 struct opal_occ_msg omsg;
947 if (msg_type != OPAL_MSG_OCC)
950 omsg.type = be64_to_cpu(msg->params[0]);
955 pr_info("OCC (On Chip Controller - enforces hard thermal/power limits) Resetting\n");
957 * powernv_cpufreq_throttle_check() is called in
958 * target() callback which can detect the throttle state
959 * for governors like ondemand.
960 * But static governors will not call target() often thus
961 * report throttling here.
965 pr_warn("CPU frequency is throttled for duration\n");
970 pr_info("OCC Loading, CPU frequency is throttled until OCC is started\n");
973 omsg.chip = be64_to_cpu(msg->params[1]);
974 omsg.throttle_status = be64_to_cpu(msg->params[2]);
979 pr_info("OCC Active, CPU frequency is no longer throttled\n");
981 for (i = 0; i < nr_chips; i++) {
982 chips[i].restore = true;
983 schedule_work(&chips[i].throttle);
989 for (i = 0; i < nr_chips; i++)
990 if (chips[i].id == omsg.chip)
993 if (omsg.throttle_status >= 0 &&
994 omsg.throttle_status <= OCC_MAX_THROTTLE_STATUS) {
995 chips[i].throttle_reason = omsg.throttle_status;
996 chips[i].reason[omsg.throttle_status]++;
999 if (!omsg.throttle_status)
1000 chips[i].restore = true;
1002 schedule_work(&chips[i].throttle);
1007 static struct notifier_block powernv_cpufreq_opal_nb = {
1008 .notifier_call = powernv_cpufreq_occ_msg,
1013 static void powernv_cpufreq_stop_cpu(struct cpufreq_policy *policy)
1015 struct powernv_smp_call_data freq_data;
1016 struct global_pstate_info *gpstates = policy->driver_data;
1018 freq_data.pstate_id = idx_to_pstate(powernv_pstate_info.min);
1019 freq_data.gpstate_id = idx_to_pstate(powernv_pstate_info.min);
1020 smp_call_function_single(policy->cpu, set_pstate, &freq_data, 1);
1022 del_timer_sync(&gpstates->timer);
1025 static unsigned int powernv_fast_switch(struct cpufreq_policy *policy,
1026 unsigned int target_freq)
1029 struct powernv_smp_call_data freq_data;
1031 index = cpufreq_table_find_index_dl(policy, target_freq);
1032 freq_data.pstate_id = powernv_freqs[index].driver_data;
1033 freq_data.gpstate_id = powernv_freqs[index].driver_data;
1034 set_pstate(&freq_data);
1036 return powernv_freqs[index].frequency;
1039 static struct cpufreq_driver powernv_cpufreq_driver = {
1040 .name = "powernv-cpufreq",
1041 .flags = CPUFREQ_CONST_LOOPS,
1042 .init = powernv_cpufreq_cpu_init,
1043 .exit = powernv_cpufreq_cpu_exit,
1044 .verify = cpufreq_generic_frequency_table_verify,
1045 .target_index = powernv_cpufreq_target_index,
1046 .fast_switch = powernv_fast_switch,
1047 .get = powernv_cpufreq_get,
1048 .stop_cpu = powernv_cpufreq_stop_cpu,
1049 .attr = powernv_cpu_freq_attr,
1052 static int init_chip_info(void)
1054 unsigned int chip[256];
1055 unsigned int cpu, i;
1056 unsigned int prev_chip_id = UINT_MAX;
1058 for_each_possible_cpu(cpu) {
1059 unsigned int id = cpu_to_chip_id(cpu);
1061 if (prev_chip_id != id) {
1063 chip[nr_chips++] = id;
1067 chips = kcalloc(nr_chips, sizeof(struct chip), GFP_KERNEL);
1071 for (i = 0; i < nr_chips; i++) {
1072 chips[i].id = chip[i];
1073 cpumask_copy(&chips[i].mask, cpumask_of_node(chip[i]));
1074 INIT_WORK(&chips[i].throttle, powernv_cpufreq_work_fn);
1075 for_each_cpu(cpu, &chips[i].mask)
1076 per_cpu(chip_info, cpu) = &chips[i];
1082 static inline void clean_chip_info(void)
1087 static inline void unregister_all_notifiers(void)
1089 opal_message_notifier_unregister(OPAL_MSG_OCC,
1090 &powernv_cpufreq_opal_nb);
1091 unregister_reboot_notifier(&powernv_cpufreq_reboot_nb);
1094 static int __init powernv_cpufreq_init(void)
1098 /* Don't probe on pseries (guest) platforms */
1099 if (!firmware_has_feature(FW_FEATURE_OPAL))
1102 /* Discover pstates from device tree and init */
1103 rc = init_powernv_pstates();
1107 /* Populate chip info */
1108 rc = init_chip_info();
1112 register_reboot_notifier(&powernv_cpufreq_reboot_nb);
1113 opal_message_notifier_register(OPAL_MSG_OCC, &powernv_cpufreq_opal_nb);
1115 if (powernv_pstate_info.wof_enabled)
1116 powernv_cpufreq_driver.boost_enabled = true;
1118 powernv_cpu_freq_attr[SCALING_BOOST_FREQS_ATTR_INDEX] = NULL;
1120 rc = cpufreq_register_driver(&powernv_cpufreq_driver);
1122 pr_info("Failed to register the cpufreq driver (%d)\n", rc);
1123 goto cleanup_notifiers;
1126 if (powernv_pstate_info.wof_enabled)
1127 cpufreq_enable_boost_support();
1131 unregister_all_notifiers();
1134 pr_info("Platform driver disabled. System does not support PState control\n");
1137 module_init(powernv_cpufreq_init);
1139 static void __exit powernv_cpufreq_exit(void)
1141 cpufreq_unregister_driver(&powernv_cpufreq_driver);
1142 unregister_all_notifiers();
1145 module_exit(powernv_cpufreq_exit);
1147 MODULE_LICENSE("GPL");
1148 MODULE_AUTHOR("Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>");