1 /*
2  * POWERNV cpufreq driver for the IBM POWER processors
3  *
4  * (C) Copyright IBM 2014
5  *
6  * Author: Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>
7  *
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)
11  * any later version.
12  *
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.
17  *
18  */
19 
20 #define pr_fmt(fmt)	"powernv-cpufreq: " fmt
21 
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>
28 #include <linux/of.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>
34 
35 #include <asm/cputhreads.h>
36 #include <asm/firmware.h>
37 #include <asm/reg.h>
38 #include <asm/smp.h> /* Required for cpu_sibling_mask() in UP configs */
39 #include <asm/opal.h>
40 #include <linux/timer.h>
41 
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
49 
50 #define MAX_RAMP_DOWN_TIME				5120
51 /*
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.
55  *
56  * This gives a percentage rampdown for time elapsed in milliseconds.
57  * ramp_down_percentage = ((ms * ms) >> 18)
58  *			~= 3.8 * (sec * sec)
59  *
60  * At 0 ms	ramp_down_percent = 0
61  * At 5120 ms	ramp_down_percent = 100
62  */
63 #define ramp_down_percent(time)		((time * time) >> 18)
64 
65 /* Interval after which the timer is queued to bring down global pstate */
66 #define GPSTATE_TIMER_INTERVAL				2000
67 
68 /**
69  * struct global_pstate_info -	Per policy data structure to maintain history of
70  *				global pstates
71  * @highest_lpstate_idx:	The local pstate index from which we are
72  *				ramping down
73  * @elapsed_time:		Time in ms spent in ramping down from
74  *				highest_lpstate_idx
75  * @last_sampled_time:		Time from boot in ms when global pstates were
76  *				last set
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
84  */
85 struct global_pstate_info {
86 	int highest_lpstate_idx;
87 	unsigned int elapsed_time;
88 	unsigned int last_sampled_time;
89 	int last_lpstate_idx;
90 	int last_gpstate_idx;
91 	spinlock_t gpstate_lock;
92 	struct timer_list timer;
93 	struct cpufreq_policy *policy;
94 };
95 
96 static struct cpufreq_frequency_table powernv_freqs[POWERNV_MAX_PSTATES+1];
97 
98 DEFINE_HASHTABLE(pstate_revmap, POWERNV_MAX_PSTATES_ORDER);
99 /**
100  * struct pstate_idx_revmap_data: Entry in the hashmap pstate_revmap
101  *				  indexed by a function of pstate id.
102  *
103  * @pstate_id: pstate id for this entry.
104  *
105  * @cpufreq_table_idx: Index into the powernv_freqs
106  *		       cpufreq_frequency_table for frequency
107  *		       corresponding to pstate_id.
108  *
109  * @hentry: hlist_node that hooks this entry into the pstate_revmap
110  *	    hashtable
111  */
112 struct pstate_idx_revmap_data {
113 	u8 pstate_id;
114 	unsigned int cpufreq_table_idx;
115 	struct hlist_node hentry;
116 };
117 
118 static bool rebooting, throttled, occ_reset;
119 
120 static const char * const throttle_reason[] = {
121 	"No throttling",
122 	"Power Cap",
123 	"Processor Over Temperature",
124 	"Power Supply Failure",
125 	"Over Current",
126 	"OCC Reset"
127 };
128 
129 enum throttle_reason_type {
130 	NO_THROTTLE = 0,
131 	POWERCAP,
132 	CPU_OVERTEMP,
133 	POWER_SUPPLY_FAILURE,
134 	OVERCURRENT,
135 	OCC_RESET_THROTTLE,
136 	OCC_MAX_REASON
137 };
138 
139 static struct chip {
140 	unsigned int id;
141 	bool throttled;
142 	bool restore;
143 	u8 throttle_reason;
144 	cpumask_t mask;
145 	struct work_struct throttle;
146 	int throttle_turbo;
147 	int throttle_sub_turbo;
148 	int reason[OCC_MAX_REASON];
149 } *chips;
150 
151 static int nr_chips;
152 static DEFINE_PER_CPU(struct chip *, chip_info);
153 
154 /*
155  * Note:
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.
160  *
161  * powernv_pstate_info.nominal indicates the index to the highest
162  * non-turbo frequency.
163  */
164 static struct powernv_pstate_info {
165 	unsigned int min;
166 	unsigned int max;
167 	unsigned int nominal;
168 	unsigned int nr_pstates;
169 	bool wof_enabled;
170 } powernv_pstate_info;
171 
172 static inline u8 extract_pstate(u64 pmsr_val, unsigned int shift)
173 {
174 	return ((pmsr_val >> shift) & 0xFF);
175 }
176 
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)
180 
181 /* Use following functions for conversions between pstate_id and index */
182 
183 /**
184  * idx_to_pstate : Returns the pstate id corresponding to the
185  *		   frequency in the cpufreq frequency table
186  *		   powernv_freqs indexed by @i.
187  *
188  *		   If @i is out of bound, this will return the pstate
189  *		   corresponding to the nominal frequency.
190  */
191 static inline u8 idx_to_pstate(unsigned int i)
192 {
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;
196 	}
197 
198 	return powernv_freqs[i].driver_data;
199 }
200 
201 /**
202  * pstate_to_idx : Returns the index in the cpufreq frequencytable
203  *		   powernv_freqs for the frequency whose corresponding
204  *		   pstate id is @pstate.
205  *
206  *		   If no frequency corresponding to @pstate is found,
207  *		   this will return the index of the nominal
208  *		   frequency.
209  */
210 static unsigned int pstate_to_idx(u8 pstate)
211 {
212 	unsigned int key = pstate % POWERNV_MAX_PSTATES;
213 	struct pstate_idx_revmap_data *revmap_data;
214 
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;
218 	}
219 
220 	pr_warn_once("pstate_to_idx: pstate 0x%x not found\n", pstate);
221 	return powernv_pstate_info.nominal;
222 }
223 
224 static inline void reset_gpstates(struct cpufreq_policy *policy)
225 {
226 	struct global_pstate_info *gpstates = policy->driver_data;
227 
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;
233 }
234 
235 /*
236  * Initialize the freq table based on data obtained
237  * from the firmware passed via device-tree
238  */
239 static int init_powernv_pstates(void)
240 {
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;
247 	int rc = -ENODEV;
248 
249 	power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
250 	if (!power_mgt) {
251 		pr_warn("power-mgt node not found\n");
252 		return -ENODEV;
253 	}
254 
255 	if (of_property_read_u32(power_mgt, "ibm,pstate-min", &pstate_min)) {
256 		pr_warn("ibm,pstate-min node not found\n");
257 		goto out;
258 	}
259 
260 	if (of_property_read_u32(power_mgt, "ibm,pstate-max", &pstate_max)) {
261 		pr_warn("ibm,pstate-max node not found\n");
262 		goto out;
263 	}
264 
265 	if (of_property_read_u32(power_mgt, "ibm,pstate-nominal",
266 				 &pstate_nominal)) {
267 		pr_warn("ibm,pstate-nominal not found\n");
268 		goto out;
269 	}
270 
271 	if (of_property_read_u32(power_mgt, "ibm,pstate-ultra-turbo",
272 				 &pstate_ultra_turbo)) {
273 		powernv_pstate_info.wof_enabled = false;
274 		goto next;
275 	}
276 
277 	if (of_property_read_u32(power_mgt, "ibm,pstate-turbo",
278 				 &pstate_turbo)) {
279 		powernv_pstate_info.wof_enabled = false;
280 		goto next;
281 	}
282 
283 	if (pstate_turbo == pstate_ultra_turbo)
284 		powernv_pstate_info.wof_enabled = false;
285 	else
286 		powernv_pstate_info.wof_enabled = true;
287 
288 next:
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");
293 
294 	pstate_ids = of_get_property(power_mgt, "ibm,pstate-ids", &len_ids);
295 	if (!pstate_ids) {
296 		pr_warn("ibm,pstate-ids not found\n");
297 		goto out;
298 	}
299 
300 	pstate_freqs = of_get_property(power_mgt, "ibm,pstate-frequencies-mhz",
301 				      &len_freqs);
302 	if (!pstate_freqs) {
303 		pr_warn("ibm,pstate-frequencies-mhz not found\n");
304 		goto out;
305 	}
306 
307 	if (len_ids != len_freqs) {
308 		pr_warn("Entries in ibm,pstate-ids and "
309 			"ibm,pstate-frequencies-mhz does not match\n");
310 	}
311 
312 	nr_pstates = min(len_ids, len_freqs) / sizeof(u32);
313 	if (!nr_pstates) {
314 		pr_warn("No PStates found\n");
315 		goto out;
316 	}
317 
318 	powernv_pstate_info.nr_pstates = nr_pstates;
319 	pr_debug("NR PStates %d\n", nr_pstates);
320 
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;
325 		unsigned int key;
326 
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;
330 
331 		revmap_data = kmalloc(sizeof(*revmap_data), GFP_KERNEL);
332 		if (!revmap_data) {
333 			rc = -ENOMEM;
334 			goto out;
335 		}
336 
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);
341 
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;
348 
349 		if (powernv_pstate_info.wof_enabled && id == pstate_turbo) {
350 			int j;
351 
352 			for (j = i - 1; j >= (int)powernv_pstate_info.max; j--)
353 				powernv_freqs[j].flags = CPUFREQ_BOOST_FREQ;
354 		}
355 	}
356 
357 	/* End of list marker entry */
358 	powernv_freqs[i].frequency = CPUFREQ_TABLE_END;
359 
360 	of_node_put(power_mgt);
361 	return 0;
362 out:
363 	of_node_put(power_mgt);
364 	return rc;
365 }
366 
367 /* Returns the CPU frequency corresponding to the pstate_id. */
368 static unsigned int pstate_id_to_freq(u8 pstate_id)
369 {
370 	int i;
371 
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;
377 	}
378 
379 	return powernv_freqs[i].frequency;
380 }
381 
382 /*
383  * cpuinfo_nominal_freq_show - Show the nominal CPU frequency as indicated by
384  * the firmware
385  */
386 static ssize_t cpuinfo_nominal_freq_show(struct cpufreq_policy *policy,
387 					char *buf)
388 {
389 	return sprintf(buf, "%u\n",
390 		powernv_freqs[powernv_pstate_info.nominal].frequency);
391 }
392 
393 struct freq_attr cpufreq_freq_attr_cpuinfo_nominal_freq =
394 	__ATTR_RO(cpuinfo_nominal_freq);
395 
396 #define SCALING_BOOST_FREQS_ATTR_INDEX		2
397 
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,
402 	NULL,
403 };
404 
405 #define throttle_attr(name, member)					\
406 static ssize_t name##_show(struct cpufreq_policy *policy, char *buf)	\
407 {									\
408 	struct chip *chip = per_cpu(chip_info, policy->cpu);		\
409 									\
410 	return sprintf(buf, "%u\n", chip->member);			\
411 }									\
412 									\
413 static struct freq_attr throttle_attr_##name = __ATTR_RO(name)		\
414 
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);
423 
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,
433 	NULL,
434 };
435 
436 static const struct attribute_group throttle_attr_grp = {
437 	.name	= "throttle_stats",
438 	.attrs	= throttle_attrs,
439 };
440 
441 /* Helper routines */
442 
443 /* Access helpers to power mgt SPR */
444 
445 static inline unsigned long get_pmspr(unsigned long sprn)
446 {
447 	switch (sprn) {
448 	case SPRN_PMCR:
449 		return mfspr(SPRN_PMCR);
450 
451 	case SPRN_PMICR:
452 		return mfspr(SPRN_PMICR);
453 
454 	case SPRN_PMSR:
455 		return mfspr(SPRN_PMSR);
456 	}
457 	BUG();
458 }
459 
460 static inline void set_pmspr(unsigned long sprn, unsigned long val)
461 {
462 	switch (sprn) {
463 	case SPRN_PMCR:
464 		mtspr(SPRN_PMCR, val);
465 		return;
466 
467 	case SPRN_PMICR:
468 		mtspr(SPRN_PMICR, val);
469 		return;
470 	}
471 	BUG();
472 }
473 
474 /*
475  * Use objects of this type to query/update
476  * pstates on a remote CPU via smp_call_function.
477  */
478 struct powernv_smp_call_data {
479 	unsigned int freq;
480 	u8 pstate_id;
481 	u8 gpstate_id;
482 };
483 
484 /*
485  * powernv_read_cpu_freq: Reads the current frequency on this CPU.
486  *
487  * Called via smp_call_function.
488  *
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.
491  *
492  * The current frequency on this CPU will be returned via
493  * ((struct powernv_smp_call_data *)arg)->freq;
494  */
495 static void powernv_read_cpu_freq(void *arg)
496 {
497 	unsigned long pmspr_val;
498 	struct powernv_smp_call_data *freq_data = arg;
499 
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);
503 
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,
506 		 freq_data->freq);
507 }
508 
509 /*
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.
513  */
514 static unsigned int powernv_cpufreq_get(unsigned int cpu)
515 {
516 	struct powernv_smp_call_data freq_data;
517 
518 	smp_call_function_any(cpu_sibling_mask(cpu), powernv_read_cpu_freq,
519 			&freq_data, 1);
520 
521 	return freq_data.freq;
522 }
523 
524 /*
525  * set_pstate: Sets the pstate on this CPU.
526  *
527  * This is called via an smp_call_function.
528  *
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.
532  */
533 static void set_pstate(void *data)
534 {
535 	unsigned long val;
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;
539 
540 	val = get_pmspr(SPRN_PMCR);
541 	val = val & 0x0000FFFFFFFFFFFFULL;
542 
543 	pstate_ul = pstate_ul & 0xFF;
544 	gpstate_ul = gpstate_ul & 0xFF;
545 
546 	/* Set both global(bits 56..63) and local(bits 48..55) PStates */
547 	val = val | (gpstate_ul << 56) | (pstate_ul << 48);
548 
549 	pr_debug("Setting cpu %d pmcr to %016lX\n",
550 			raw_smp_processor_id(), val);
551 	set_pmspr(SPRN_PMCR, val);
552 }
553 
554 /*
555  * get_nominal_index: Returns the index corresponding to the nominal
556  * pstate in the cpufreq table
557  */
558 static inline unsigned int get_nominal_index(void)
559 {
560 	return powernv_pstate_info.nominal;
561 }
562 
563 static void powernv_cpufreq_throttle_check(void *data)
564 {
565 	struct chip *chip;
566 	unsigned int cpu = smp_processor_id();
567 	unsigned long pmsr;
568 	u8 pmsr_pmax;
569 	unsigned int pmsr_pmax_idx;
570 
571 	pmsr = get_pmspr(SPRN_PMSR);
572 	chip = this_cpu_read(chip_info);
573 
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) {
578 		if (chip->throttled)
579 			goto next;
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++;
586 		} else {
587 			chip->throttle_turbo++;
588 		}
589 		trace_powernv_throttle(chip->id,
590 				      throttle_reason[chip->throttle_reason],
591 				      pmsr_pmax);
592 	} else if (chip->throttled) {
593 		chip->throttled = false;
594 		trace_powernv_throttle(chip->id,
595 				      throttle_reason[chip->throttle_reason],
596 				      pmsr_pmax);
597 	}
598 
599 	/* Check if Psafe_mode_active is set in PMSR. */
600 next:
601 	if (pmsr & PMSR_PSAFE_ENABLE) {
602 		throttled = true;
603 		pr_info("Pstate set to safe frequency\n");
604 	}
605 
606 	/* Check if SPR_EM_DISABLE is set in PMSR */
607 	if (pmsr & PMSR_SPR_EM_DISABLE) {
608 		throttled = true;
609 		pr_info("Frequency Control disabled from OS\n");
610 	}
611 
612 	if (throttled) {
613 		pr_info("PMSR = %16lx\n", pmsr);
614 		pr_warn("CPU Frequency could be throttled\n");
615 	}
616 }
617 
618 /**
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
623  *
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.
627  */
628 static inline int calc_global_pstate(unsigned int elapsed_time,
629 				     int highest_lpstate_idx,
630 				     int local_pstate_idx)
631 {
632 	int index_diff;
633 
634 	/*
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.
640 	 */
641 	index_diff =  ((int)ramp_down_percent(elapsed_time) *
642 			(powernv_pstate_info.min - highest_lpstate_idx)) / 100;
643 
644 	/* Ensure that global pstate is >= to local pstate */
645 	if (highest_lpstate_idx + index_diff >= local_pstate_idx)
646 		return local_pstate_idx;
647 	else
648 		return highest_lpstate_idx + index_diff;
649 }
650 
651 static inline void  queue_gpstate_timer(struct global_pstate_info *gpstates)
652 {
653 	unsigned int timer_interval;
654 
655 	/*
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.
660 	 */
661 	if ((gpstates->elapsed_time + GPSTATE_TIMER_INTERVAL)
662 	     > MAX_RAMP_DOWN_TIME)
663 		timer_interval = MAX_RAMP_DOWN_TIME - gpstates->elapsed_time;
664 	else
665 		timer_interval = GPSTATE_TIMER_INTERVAL;
666 
667 	mod_timer(&gpstates->timer, jiffies + msecs_to_jiffies(timer_interval));
668 }
669 
670 /**
671  * gpstate_timer_handler
672  *
673  * @data: pointer to cpufreq_policy on which timer was queued
674  *
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
677  * to local pstate
678  */
679 void gpstate_timer_handler(struct timer_list *t)
680 {
681 	struct global_pstate_info *gpstates = from_timer(gpstates, t, timer);
682 	struct cpufreq_policy *policy = gpstates->policy;
683 	int gpstate_idx, lpstate_idx;
684 	unsigned long val;
685 	unsigned int time_diff = jiffies_to_msecs(jiffies)
686 					- gpstates->last_sampled_time;
687 	struct powernv_smp_call_data freq_data;
688 
689 	if (!spin_trylock(&gpstates->gpstate_lock))
690 		return;
691 	/*
692 	 * If the timer has migrated to the different cpu then bring
693 	 * it back to one of the policy->cpus
694 	 */
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);
699 		return;
700 	}
701 
702 	/*
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.
706 	 */
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);
713 		return;
714 	}
715 
716 	gpstates->last_sampled_time += time_diff;
717 	gpstates->elapsed_time += time_diff;
718 
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;
724 	} else {
725 		lpstate_idx = pstate_to_idx(freq_data.pstate_id);
726 		gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
727 						 gpstates->highest_lpstate_idx,
728 						 lpstate_idx);
729 	}
730 	freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
731 	gpstates->last_gpstate_idx = gpstate_idx;
732 	gpstates->last_lpstate_idx = lpstate_idx;
733 	/*
734 	 * If local pstate is equal to global pstate, rampdown is over
735 	 * So timer is not required to be queued.
736 	 */
737 	if (gpstate_idx != gpstates->last_lpstate_idx)
738 		queue_gpstate_timer(gpstates);
739 
740 	set_pstate(&freq_data);
741 	spin_unlock(&gpstates->gpstate_lock);
742 }
743 
744 /*
745  * powernv_cpufreq_target_index: Sets the frequency corresponding to
746  * the cpufreq table entry indexed by new_index on the cpus in the
747  * mask policy->cpus
748  */
749 static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
750 					unsigned int new_index)
751 {
752 	struct powernv_smp_call_data freq_data;
753 	unsigned int cur_msec, gpstate_idx;
754 	struct global_pstate_info *gpstates = policy->driver_data;
755 
756 	if (unlikely(rebooting) && new_index != get_nominal_index())
757 		return 0;
758 
759 	if (!throttled) {
760 		/* we don't want to be preempted while
761 		 * checking if the CPU frequency has been throttled
762 		 */
763 		preempt_disable();
764 		powernv_cpufreq_throttle_check(NULL);
765 		preempt_enable();
766 	}
767 
768 	cur_msec = jiffies_to_msecs(get_jiffies_64());
769 
770 	freq_data.pstate_id = idx_to_pstate(new_index);
771 	if (!gpstates) {
772 		freq_data.gpstate_id = freq_data.pstate_id;
773 		goto no_gpstate;
774 	}
775 
776 	spin_lock(&gpstates->gpstate_lock);
777 
778 	if (!gpstates->last_sampled_time) {
779 		gpstate_idx = new_index;
780 		gpstates->highest_lpstate_idx = new_index;
781 		goto gpstates_done;
782 	}
783 
784 	if (gpstates->last_gpstate_idx < new_index) {
785 		gpstates->elapsed_time += cur_msec -
786 						 gpstates->last_sampled_time;
787 
788 		/*
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.
792 		 */
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;
797 		} else {
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,
801 							 new_index);
802 		}
803 	} else {
804 		reset_gpstates(policy);
805 		gpstates->highest_lpstate_idx = new_index;
806 		gpstate_idx = new_index;
807 	}
808 
809 	/*
810 	 * If local pstate is equal to global pstate, rampdown is over
811 	 * So timer is not required to be queued.
812 	 */
813 	if (gpstate_idx != new_index)
814 		queue_gpstate_timer(gpstates);
815 	else
816 		del_timer_sync(&gpstates->timer);
817 
818 gpstates_done:
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;
823 
824 	spin_unlock(&gpstates->gpstate_lock);
825 
826 no_gpstate:
827 	/*
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)
831 	 */
832 	smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
833 	return 0;
834 }
835 
836 static int powernv_cpufreq_cpu_init(struct cpufreq_policy *policy)
837 {
838 	int base, i;
839 	struct kernfs_node *kn;
840 	struct global_pstate_info *gpstates;
841 
842 	base = cpu_first_thread_sibling(policy->cpu);
843 
844 	for (i = 0; i < threads_per_core; i++)
845 		cpumask_set_cpu(base + i, policy->cpus);
846 
847 	kn = kernfs_find_and_get(policy->kobj.sd, throttle_attr_grp.name);
848 	if (!kn) {
849 		int ret;
850 
851 		ret = sysfs_create_group(&policy->kobj, &throttle_attr_grp);
852 		if (ret) {
853 			pr_info("Failed to create throttle stats directory for cpu %d\n",
854 				policy->cpu);
855 			return ret;
856 		}
857 	} else {
858 		kernfs_put(kn);
859 	}
860 
861 	policy->freq_table = powernv_freqs;
862 	policy->fast_switch_possible = true;
863 
864 	if (pvr_version_is(PVR_POWER9))
865 		return 0;
866 
867 	/* Initialise Gpstate ramp-down timer only on POWER8 */
868 	gpstates =  kzalloc(sizeof(*gpstates), GFP_KERNEL);
869 	if (!gpstates)
870 		return -ENOMEM;
871 
872 	policy->driver_data = gpstates;
873 
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);
881 
882 	return 0;
883 }
884 
885 static int powernv_cpufreq_cpu_exit(struct cpufreq_policy *policy)
886 {
887 	/* timer is deleted in cpufreq_cpu_stop() */
888 	kfree(policy->driver_data);
889 
890 	return 0;
891 }
892 
893 static int powernv_cpufreq_reboot_notifier(struct notifier_block *nb,
894 				unsigned long action, void *unused)
895 {
896 	int cpu;
897 	struct cpufreq_policy cpu_policy;
898 
899 	rebooting = true;
900 	for_each_online_cpu(cpu) {
901 		cpufreq_get_policy(&cpu_policy, cpu);
902 		powernv_cpufreq_target_index(&cpu_policy, get_nominal_index());
903 	}
904 
905 	return NOTIFY_DONE;
906 }
907 
908 static struct notifier_block powernv_cpufreq_reboot_nb = {
909 	.notifier_call = powernv_cpufreq_reboot_notifier,
910 };
911 
912 void powernv_cpufreq_work_fn(struct work_struct *work)
913 {
914 	struct chip *chip = container_of(work, struct chip, throttle);
915 	unsigned int cpu;
916 	cpumask_t mask;
917 
918 	get_online_cpus();
919 	cpumask_and(&mask, &chip->mask, cpu_online_mask);
920 	smp_call_function_any(&mask,
921 			      powernv_cpufreq_throttle_check, NULL, 0);
922 
923 	if (!chip->restore)
924 		goto out;
925 
926 	chip->restore = false;
927 	for_each_cpu(cpu, &mask) {
928 		int index;
929 		struct cpufreq_policy policy;
930 
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);
935 	}
936 out:
937 	put_online_cpus();
938 }
939 
940 static int powernv_cpufreq_occ_msg(struct notifier_block *nb,
941 				   unsigned long msg_type, void *_msg)
942 {
943 	struct opal_msg *msg = _msg;
944 	struct opal_occ_msg omsg;
945 	int i;
946 
947 	if (msg_type != OPAL_MSG_OCC)
948 		return 0;
949 
950 	omsg.type = be64_to_cpu(msg->params[0]);
951 
952 	switch (omsg.type) {
953 	case OCC_RESET:
954 		occ_reset = true;
955 		pr_info("OCC (On Chip Controller - enforces hard thermal/power limits) Resetting\n");
956 		/*
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.
962 		 */
963 		if (!throttled) {
964 			throttled = true;
965 			pr_warn("CPU frequency is throttled for duration\n");
966 		}
967 
968 		break;
969 	case OCC_LOAD:
970 		pr_info("OCC Loading, CPU frequency is throttled until OCC is started\n");
971 		break;
972 	case OCC_THROTTLE:
973 		omsg.chip = be64_to_cpu(msg->params[1]);
974 		omsg.throttle_status = be64_to_cpu(msg->params[2]);
975 
976 		if (occ_reset) {
977 			occ_reset = false;
978 			throttled = false;
979 			pr_info("OCC Active, CPU frequency is no longer throttled\n");
980 
981 			for (i = 0; i < nr_chips; i++) {
982 				chips[i].restore = true;
983 				schedule_work(&chips[i].throttle);
984 			}
985 
986 			return 0;
987 		}
988 
989 		for (i = 0; i < nr_chips; i++)
990 			if (chips[i].id == omsg.chip)
991 				break;
992 
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]++;
997 		}
998 
999 		if (!omsg.throttle_status)
1000 			chips[i].restore = true;
1001 
1002 		schedule_work(&chips[i].throttle);
1003 	}
1004 	return 0;
1005 }
1006 
1007 static struct notifier_block powernv_cpufreq_opal_nb = {
1008 	.notifier_call	= powernv_cpufreq_occ_msg,
1009 	.next		= NULL,
1010 	.priority	= 0,
1011 };
1012 
1013 static void powernv_cpufreq_stop_cpu(struct cpufreq_policy *policy)
1014 {
1015 	struct powernv_smp_call_data freq_data;
1016 	struct global_pstate_info *gpstates = policy->driver_data;
1017 
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);
1021 	if (gpstates)
1022 		del_timer_sync(&gpstates->timer);
1023 }
1024 
1025 static unsigned int powernv_fast_switch(struct cpufreq_policy *policy,
1026 					unsigned int target_freq)
1027 {
1028 	int index;
1029 	struct powernv_smp_call_data freq_data;
1030 
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);
1035 
1036 	return powernv_freqs[index].frequency;
1037 }
1038 
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,
1050 };
1051 
1052 static int init_chip_info(void)
1053 {
1054 	unsigned int chip[256];
1055 	unsigned int cpu, i;
1056 	unsigned int prev_chip_id = UINT_MAX;
1057 
1058 	for_each_possible_cpu(cpu) {
1059 		unsigned int id = cpu_to_chip_id(cpu);
1060 
1061 		if (prev_chip_id != id) {
1062 			prev_chip_id = id;
1063 			chip[nr_chips++] = id;
1064 		}
1065 	}
1066 
1067 	chips = kcalloc(nr_chips, sizeof(struct chip), GFP_KERNEL);
1068 	if (!chips)
1069 		return -ENOMEM;
1070 
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];
1077 	}
1078 
1079 	return 0;
1080 }
1081 
1082 static inline void clean_chip_info(void)
1083 {
1084 	kfree(chips);
1085 }
1086 
1087 static inline void unregister_all_notifiers(void)
1088 {
1089 	opal_message_notifier_unregister(OPAL_MSG_OCC,
1090 					 &powernv_cpufreq_opal_nb);
1091 	unregister_reboot_notifier(&powernv_cpufreq_reboot_nb);
1092 }
1093 
1094 static int __init powernv_cpufreq_init(void)
1095 {
1096 	int rc = 0;
1097 
1098 	/* Don't probe on pseries (guest) platforms */
1099 	if (!firmware_has_feature(FW_FEATURE_OPAL))
1100 		return -ENODEV;
1101 
1102 	/* Discover pstates from device tree and init */
1103 	rc = init_powernv_pstates();
1104 	if (rc)
1105 		goto out;
1106 
1107 	/* Populate chip info */
1108 	rc = init_chip_info();
1109 	if (rc)
1110 		goto out;
1111 
1112 	register_reboot_notifier(&powernv_cpufreq_reboot_nb);
1113 	opal_message_notifier_register(OPAL_MSG_OCC, &powernv_cpufreq_opal_nb);
1114 
1115 	if (powernv_pstate_info.wof_enabled)
1116 		powernv_cpufreq_driver.boost_enabled = true;
1117 	else
1118 		powernv_cpu_freq_attr[SCALING_BOOST_FREQS_ATTR_INDEX] = NULL;
1119 
1120 	rc = cpufreq_register_driver(&powernv_cpufreq_driver);
1121 	if (rc) {
1122 		pr_info("Failed to register the cpufreq driver (%d)\n", rc);
1123 		goto cleanup_notifiers;
1124 	}
1125 
1126 	if (powernv_pstate_info.wof_enabled)
1127 		cpufreq_enable_boost_support();
1128 
1129 	return 0;
1130 cleanup_notifiers:
1131 	unregister_all_notifiers();
1132 	clean_chip_info();
1133 out:
1134 	pr_info("Platform driver disabled. System does not support PState control\n");
1135 	return rc;
1136 }
1137 module_init(powernv_cpufreq_init);
1138 
1139 static void __exit powernv_cpufreq_exit(void)
1140 {
1141 	cpufreq_unregister_driver(&powernv_cpufreq_driver);
1142 	unregister_all_notifiers();
1143 	clean_chip_info();
1144 }
1145 module_exit(powernv_cpufreq_exit);
1146 
1147 MODULE_LICENSE("GPL");
1148 MODULE_AUTHOR("Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>");
1149