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 
248 	power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
249 	if (!power_mgt) {
250 		pr_warn("power-mgt node not found\n");
251 		return -ENODEV;
252 	}
253 
254 	if (of_property_read_u32(power_mgt, "ibm,pstate-min", &pstate_min)) {
255 		pr_warn("ibm,pstate-min node not found\n");
256 		return -ENODEV;
257 	}
258 
259 	if (of_property_read_u32(power_mgt, "ibm,pstate-max", &pstate_max)) {
260 		pr_warn("ibm,pstate-max node not found\n");
261 		return -ENODEV;
262 	}
263 
264 	if (of_property_read_u32(power_mgt, "ibm,pstate-nominal",
265 				 &pstate_nominal)) {
266 		pr_warn("ibm,pstate-nominal not found\n");
267 		return -ENODEV;
268 	}
269 
270 	if (of_property_read_u32(power_mgt, "ibm,pstate-ultra-turbo",
271 				 &pstate_ultra_turbo)) {
272 		powernv_pstate_info.wof_enabled = false;
273 		goto next;
274 	}
275 
276 	if (of_property_read_u32(power_mgt, "ibm,pstate-turbo",
277 				 &pstate_turbo)) {
278 		powernv_pstate_info.wof_enabled = false;
279 		goto next;
280 	}
281 
282 	if (pstate_turbo == pstate_ultra_turbo)
283 		powernv_pstate_info.wof_enabled = false;
284 	else
285 		powernv_pstate_info.wof_enabled = true;
286 
287 next:
288 	pr_info("cpufreq pstate min 0x%x nominal 0x%x max 0x%x\n", pstate_min,
289 		pstate_nominal, pstate_max);
290 	pr_info("Workload Optimized Frequency is %s in the platform\n",
291 		(powernv_pstate_info.wof_enabled) ? "enabled" : "disabled");
292 
293 	pstate_ids = of_get_property(power_mgt, "ibm,pstate-ids", &len_ids);
294 	if (!pstate_ids) {
295 		pr_warn("ibm,pstate-ids not found\n");
296 		return -ENODEV;
297 	}
298 
299 	pstate_freqs = of_get_property(power_mgt, "ibm,pstate-frequencies-mhz",
300 				      &len_freqs);
301 	if (!pstate_freqs) {
302 		pr_warn("ibm,pstate-frequencies-mhz not found\n");
303 		return -ENODEV;
304 	}
305 
306 	if (len_ids != len_freqs) {
307 		pr_warn("Entries in ibm,pstate-ids and "
308 			"ibm,pstate-frequencies-mhz does not match\n");
309 	}
310 
311 	nr_pstates = min(len_ids, len_freqs) / sizeof(u32);
312 	if (!nr_pstates) {
313 		pr_warn("No PStates found\n");
314 		return -ENODEV;
315 	}
316 
317 	powernv_pstate_info.nr_pstates = nr_pstates;
318 	pr_debug("NR PStates %d\n", nr_pstates);
319 
320 	for (i = 0; i < nr_pstates; i++) {
321 		u32 id = be32_to_cpu(pstate_ids[i]);
322 		u32 freq = be32_to_cpu(pstate_freqs[i]);
323 		struct pstate_idx_revmap_data *revmap_data;
324 		unsigned int key;
325 
326 		pr_debug("PState id %d freq %d MHz\n", id, freq);
327 		powernv_freqs[i].frequency = freq * 1000; /* kHz */
328 		powernv_freqs[i].driver_data = id & 0xFF;
329 
330 		revmap_data = (struct pstate_idx_revmap_data *)
331 			      kmalloc(sizeof(*revmap_data), GFP_KERNEL);
332 
333 		revmap_data->pstate_id = id & 0xFF;
334 		revmap_data->cpufreq_table_idx = i;
335 		key = (revmap_data->pstate_id) % POWERNV_MAX_PSTATES;
336 		hash_add(pstate_revmap, &revmap_data->hentry, key);
337 
338 		if (id == pstate_max)
339 			powernv_pstate_info.max = i;
340 		if (id == pstate_nominal)
341 			powernv_pstate_info.nominal = i;
342 		if (id == pstate_min)
343 			powernv_pstate_info.min = i;
344 
345 		if (powernv_pstate_info.wof_enabled && id == pstate_turbo) {
346 			int j;
347 
348 			for (j = i - 1; j >= (int)powernv_pstate_info.max; j--)
349 				powernv_freqs[j].flags = CPUFREQ_BOOST_FREQ;
350 		}
351 	}
352 
353 	/* End of list marker entry */
354 	powernv_freqs[i].frequency = CPUFREQ_TABLE_END;
355 	return 0;
356 }
357 
358 /* Returns the CPU frequency corresponding to the pstate_id. */
359 static unsigned int pstate_id_to_freq(u8 pstate_id)
360 {
361 	int i;
362 
363 	i = pstate_to_idx(pstate_id);
364 	if (i >= powernv_pstate_info.nr_pstates || i < 0) {
365 		pr_warn("PState id 0x%x outside of PState table, reporting nominal id 0x%x instead\n",
366 			pstate_id, idx_to_pstate(powernv_pstate_info.nominal));
367 		i = powernv_pstate_info.nominal;
368 	}
369 
370 	return powernv_freqs[i].frequency;
371 }
372 
373 /*
374  * cpuinfo_nominal_freq_show - Show the nominal CPU frequency as indicated by
375  * the firmware
376  */
377 static ssize_t cpuinfo_nominal_freq_show(struct cpufreq_policy *policy,
378 					char *buf)
379 {
380 	return sprintf(buf, "%u\n",
381 		powernv_freqs[powernv_pstate_info.nominal].frequency);
382 }
383 
384 struct freq_attr cpufreq_freq_attr_cpuinfo_nominal_freq =
385 	__ATTR_RO(cpuinfo_nominal_freq);
386 
387 #define SCALING_BOOST_FREQS_ATTR_INDEX		2
388 
389 static struct freq_attr *powernv_cpu_freq_attr[] = {
390 	&cpufreq_freq_attr_scaling_available_freqs,
391 	&cpufreq_freq_attr_cpuinfo_nominal_freq,
392 	&cpufreq_freq_attr_scaling_boost_freqs,
393 	NULL,
394 };
395 
396 #define throttle_attr(name, member)					\
397 static ssize_t name##_show(struct cpufreq_policy *policy, char *buf)	\
398 {									\
399 	struct chip *chip = per_cpu(chip_info, policy->cpu);		\
400 									\
401 	return sprintf(buf, "%u\n", chip->member);			\
402 }									\
403 									\
404 static struct freq_attr throttle_attr_##name = __ATTR_RO(name)		\
405 
406 throttle_attr(unthrottle, reason[NO_THROTTLE]);
407 throttle_attr(powercap, reason[POWERCAP]);
408 throttle_attr(overtemp, reason[CPU_OVERTEMP]);
409 throttle_attr(supply_fault, reason[POWER_SUPPLY_FAILURE]);
410 throttle_attr(overcurrent, reason[OVERCURRENT]);
411 throttle_attr(occ_reset, reason[OCC_RESET_THROTTLE]);
412 throttle_attr(turbo_stat, throttle_turbo);
413 throttle_attr(sub_turbo_stat, throttle_sub_turbo);
414 
415 static struct attribute *throttle_attrs[] = {
416 	&throttle_attr_unthrottle.attr,
417 	&throttle_attr_powercap.attr,
418 	&throttle_attr_overtemp.attr,
419 	&throttle_attr_supply_fault.attr,
420 	&throttle_attr_overcurrent.attr,
421 	&throttle_attr_occ_reset.attr,
422 	&throttle_attr_turbo_stat.attr,
423 	&throttle_attr_sub_turbo_stat.attr,
424 	NULL,
425 };
426 
427 static const struct attribute_group throttle_attr_grp = {
428 	.name	= "throttle_stats",
429 	.attrs	= throttle_attrs,
430 };
431 
432 /* Helper routines */
433 
434 /* Access helpers to power mgt SPR */
435 
436 static inline unsigned long get_pmspr(unsigned long sprn)
437 {
438 	switch (sprn) {
439 	case SPRN_PMCR:
440 		return mfspr(SPRN_PMCR);
441 
442 	case SPRN_PMICR:
443 		return mfspr(SPRN_PMICR);
444 
445 	case SPRN_PMSR:
446 		return mfspr(SPRN_PMSR);
447 	}
448 	BUG();
449 }
450 
451 static inline void set_pmspr(unsigned long sprn, unsigned long val)
452 {
453 	switch (sprn) {
454 	case SPRN_PMCR:
455 		mtspr(SPRN_PMCR, val);
456 		return;
457 
458 	case SPRN_PMICR:
459 		mtspr(SPRN_PMICR, val);
460 		return;
461 	}
462 	BUG();
463 }
464 
465 /*
466  * Use objects of this type to query/update
467  * pstates on a remote CPU via smp_call_function.
468  */
469 struct powernv_smp_call_data {
470 	unsigned int freq;
471 	u8 pstate_id;
472 	u8 gpstate_id;
473 };
474 
475 /*
476  * powernv_read_cpu_freq: Reads the current frequency on this CPU.
477  *
478  * Called via smp_call_function.
479  *
480  * Note: The caller of the smp_call_function should pass an argument of
481  * the type 'struct powernv_smp_call_data *' along with this function.
482  *
483  * The current frequency on this CPU will be returned via
484  * ((struct powernv_smp_call_data *)arg)->freq;
485  */
486 static void powernv_read_cpu_freq(void *arg)
487 {
488 	unsigned long pmspr_val;
489 	struct powernv_smp_call_data *freq_data = arg;
490 
491 	pmspr_val = get_pmspr(SPRN_PMSR);
492 	freq_data->pstate_id = extract_local_pstate(pmspr_val);
493 	freq_data->freq = pstate_id_to_freq(freq_data->pstate_id);
494 
495 	pr_debug("cpu %d pmsr %016lX pstate_id 0x%x frequency %d kHz\n",
496 		 raw_smp_processor_id(), pmspr_val, freq_data->pstate_id,
497 		 freq_data->freq);
498 }
499 
500 /*
501  * powernv_cpufreq_get: Returns the CPU frequency as reported by the
502  * firmware for CPU 'cpu'. This value is reported through the sysfs
503  * file cpuinfo_cur_freq.
504  */
505 static unsigned int powernv_cpufreq_get(unsigned int cpu)
506 {
507 	struct powernv_smp_call_data freq_data;
508 
509 	smp_call_function_any(cpu_sibling_mask(cpu), powernv_read_cpu_freq,
510 			&freq_data, 1);
511 
512 	return freq_data.freq;
513 }
514 
515 /*
516  * set_pstate: Sets the pstate on this CPU.
517  *
518  * This is called via an smp_call_function.
519  *
520  * The caller must ensure that freq_data is of the type
521  * (struct powernv_smp_call_data *) and the pstate_id which needs to be set
522  * on this CPU should be present in freq_data->pstate_id.
523  */
524 static void set_pstate(void *data)
525 {
526 	unsigned long val;
527 	struct powernv_smp_call_data *freq_data = data;
528 	unsigned long pstate_ul = freq_data->pstate_id;
529 	unsigned long gpstate_ul = freq_data->gpstate_id;
530 
531 	val = get_pmspr(SPRN_PMCR);
532 	val = val & 0x0000FFFFFFFFFFFFULL;
533 
534 	pstate_ul = pstate_ul & 0xFF;
535 	gpstate_ul = gpstate_ul & 0xFF;
536 
537 	/* Set both global(bits 56..63) and local(bits 48..55) PStates */
538 	val = val | (gpstate_ul << 56) | (pstate_ul << 48);
539 
540 	pr_debug("Setting cpu %d pmcr to %016lX\n",
541 			raw_smp_processor_id(), val);
542 	set_pmspr(SPRN_PMCR, val);
543 }
544 
545 /*
546  * get_nominal_index: Returns the index corresponding to the nominal
547  * pstate in the cpufreq table
548  */
549 static inline unsigned int get_nominal_index(void)
550 {
551 	return powernv_pstate_info.nominal;
552 }
553 
554 static void powernv_cpufreq_throttle_check(void *data)
555 {
556 	struct chip *chip;
557 	unsigned int cpu = smp_processor_id();
558 	unsigned long pmsr;
559 	u8 pmsr_pmax;
560 	unsigned int pmsr_pmax_idx;
561 
562 	pmsr = get_pmspr(SPRN_PMSR);
563 	chip = this_cpu_read(chip_info);
564 
565 	/* Check for Pmax Capping */
566 	pmsr_pmax = extract_max_pstate(pmsr);
567 	pmsr_pmax_idx = pstate_to_idx(pmsr_pmax);
568 	if (pmsr_pmax_idx != powernv_pstate_info.max) {
569 		if (chip->throttled)
570 			goto next;
571 		chip->throttled = true;
572 		if (pmsr_pmax_idx > powernv_pstate_info.nominal) {
573 			pr_warn_once("CPU %d on Chip %u has Pmax(0x%x) reduced below that of nominal frequency(0x%x)\n",
574 				     cpu, chip->id, pmsr_pmax,
575 				     idx_to_pstate(powernv_pstate_info.nominal));
576 			chip->throttle_sub_turbo++;
577 		} else {
578 			chip->throttle_turbo++;
579 		}
580 		trace_powernv_throttle(chip->id,
581 				      throttle_reason[chip->throttle_reason],
582 				      pmsr_pmax);
583 	} else if (chip->throttled) {
584 		chip->throttled = false;
585 		trace_powernv_throttle(chip->id,
586 				      throttle_reason[chip->throttle_reason],
587 				      pmsr_pmax);
588 	}
589 
590 	/* Check if Psafe_mode_active is set in PMSR. */
591 next:
592 	if (pmsr & PMSR_PSAFE_ENABLE) {
593 		throttled = true;
594 		pr_info("Pstate set to safe frequency\n");
595 	}
596 
597 	/* Check if SPR_EM_DISABLE is set in PMSR */
598 	if (pmsr & PMSR_SPR_EM_DISABLE) {
599 		throttled = true;
600 		pr_info("Frequency Control disabled from OS\n");
601 	}
602 
603 	if (throttled) {
604 		pr_info("PMSR = %16lx\n", pmsr);
605 		pr_warn("CPU Frequency could be throttled\n");
606 	}
607 }
608 
609 /**
610  * calc_global_pstate - Calculate global pstate
611  * @elapsed_time:		Elapsed time in milliseconds
612  * @local_pstate_idx:		New local pstate
613  * @highest_lpstate_idx:	pstate from which its ramping down
614  *
615  * Finds the appropriate global pstate based on the pstate from which its
616  * ramping down and the time elapsed in ramping down. It follows a quadratic
617  * equation which ensures that it reaches ramping down to pmin in 5sec.
618  */
619 static inline int calc_global_pstate(unsigned int elapsed_time,
620 				     int highest_lpstate_idx,
621 				     int local_pstate_idx)
622 {
623 	int index_diff;
624 
625 	/*
626 	 * Using ramp_down_percent we get the percentage of rampdown
627 	 * that we are expecting to be dropping. Difference between
628 	 * highest_lpstate_idx and powernv_pstate_info.min will give a absolute
629 	 * number of how many pstates we will drop eventually by the end of
630 	 * 5 seconds, then just scale it get the number pstates to be dropped.
631 	 */
632 	index_diff =  ((int)ramp_down_percent(elapsed_time) *
633 			(powernv_pstate_info.min - highest_lpstate_idx)) / 100;
634 
635 	/* Ensure that global pstate is >= to local pstate */
636 	if (highest_lpstate_idx + index_diff >= local_pstate_idx)
637 		return local_pstate_idx;
638 	else
639 		return highest_lpstate_idx + index_diff;
640 }
641 
642 static inline void  queue_gpstate_timer(struct global_pstate_info *gpstates)
643 {
644 	unsigned int timer_interval;
645 
646 	/*
647 	 * Setting up timer to fire after GPSTATE_TIMER_INTERVAL ms, But
648 	 * if it exceeds MAX_RAMP_DOWN_TIME ms for ramp down time.
649 	 * Set timer such that it fires exactly at MAX_RAMP_DOWN_TIME
650 	 * seconds of ramp down time.
651 	 */
652 	if ((gpstates->elapsed_time + GPSTATE_TIMER_INTERVAL)
653 	     > MAX_RAMP_DOWN_TIME)
654 		timer_interval = MAX_RAMP_DOWN_TIME - gpstates->elapsed_time;
655 	else
656 		timer_interval = GPSTATE_TIMER_INTERVAL;
657 
658 	mod_timer(&gpstates->timer, jiffies + msecs_to_jiffies(timer_interval));
659 }
660 
661 /**
662  * gpstate_timer_handler
663  *
664  * @data: pointer to cpufreq_policy on which timer was queued
665  *
666  * This handler brings down the global pstate closer to the local pstate
667  * according quadratic equation. Queues a new timer if it is still not equal
668  * to local pstate
669  */
670 void gpstate_timer_handler(struct timer_list *t)
671 {
672 	struct global_pstate_info *gpstates = from_timer(gpstates, t, timer);
673 	struct cpufreq_policy *policy = gpstates->policy;
674 	int gpstate_idx, lpstate_idx;
675 	unsigned long val;
676 	unsigned int time_diff = jiffies_to_msecs(jiffies)
677 					- gpstates->last_sampled_time;
678 	struct powernv_smp_call_data freq_data;
679 
680 	if (!spin_trylock(&gpstates->gpstate_lock))
681 		return;
682 
683 	/*
684 	 * If PMCR was last updated was using fast_swtich then
685 	 * We may have wrong in gpstate->last_lpstate_idx
686 	 * value. Hence, read from PMCR to get correct data.
687 	 */
688 	val = get_pmspr(SPRN_PMCR);
689 	freq_data.gpstate_id = extract_global_pstate(val);
690 	freq_data.pstate_id = extract_local_pstate(val);
691 	if (freq_data.gpstate_id  == freq_data.pstate_id) {
692 		reset_gpstates(policy);
693 		spin_unlock(&gpstates->gpstate_lock);
694 		return;
695 	}
696 
697 	gpstates->last_sampled_time += time_diff;
698 	gpstates->elapsed_time += time_diff;
699 
700 	if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
701 		gpstate_idx = pstate_to_idx(freq_data.pstate_id);
702 		lpstate_idx = gpstate_idx;
703 		reset_gpstates(policy);
704 		gpstates->highest_lpstate_idx = gpstate_idx;
705 	} else {
706 		lpstate_idx = pstate_to_idx(freq_data.pstate_id);
707 		gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
708 						 gpstates->highest_lpstate_idx,
709 						 lpstate_idx);
710 	}
711 	freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
712 	gpstates->last_gpstate_idx = gpstate_idx;
713 	gpstates->last_lpstate_idx = lpstate_idx;
714 	/*
715 	 * If local pstate is equal to global pstate, rampdown is over
716 	 * So timer is not required to be queued.
717 	 */
718 	if (gpstate_idx != gpstates->last_lpstate_idx)
719 		queue_gpstate_timer(gpstates);
720 
721 	spin_unlock(&gpstates->gpstate_lock);
722 
723 	/* Timer may get migrated to a different cpu on cpu hot unplug */
724 	smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
725 }
726 
727 /*
728  * powernv_cpufreq_target_index: Sets the frequency corresponding to
729  * the cpufreq table entry indexed by new_index on the cpus in the
730  * mask policy->cpus
731  */
732 static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
733 					unsigned int new_index)
734 {
735 	struct powernv_smp_call_data freq_data;
736 	unsigned int cur_msec, gpstate_idx;
737 	struct global_pstate_info *gpstates = policy->driver_data;
738 
739 	if (unlikely(rebooting) && new_index != get_nominal_index())
740 		return 0;
741 
742 	if (!throttled) {
743 		/* we don't want to be preempted while
744 		 * checking if the CPU frequency has been throttled
745 		 */
746 		preempt_disable();
747 		powernv_cpufreq_throttle_check(NULL);
748 		preempt_enable();
749 	}
750 
751 	cur_msec = jiffies_to_msecs(get_jiffies_64());
752 
753 	spin_lock(&gpstates->gpstate_lock);
754 	freq_data.pstate_id = idx_to_pstate(new_index);
755 
756 	if (!gpstates->last_sampled_time) {
757 		gpstate_idx = new_index;
758 		gpstates->highest_lpstate_idx = new_index;
759 		goto gpstates_done;
760 	}
761 
762 	if (gpstates->last_gpstate_idx < new_index) {
763 		gpstates->elapsed_time += cur_msec -
764 						 gpstates->last_sampled_time;
765 
766 		/*
767 		 * If its has been ramping down for more than MAX_RAMP_DOWN_TIME
768 		 * we should be resetting all global pstate related data. Set it
769 		 * equal to local pstate to start fresh.
770 		 */
771 		if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
772 			reset_gpstates(policy);
773 			gpstates->highest_lpstate_idx = new_index;
774 			gpstate_idx = new_index;
775 		} else {
776 		/* Elaspsed_time is less than 5 seconds, continue to rampdown */
777 			gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
778 							 gpstates->highest_lpstate_idx,
779 							 new_index);
780 		}
781 	} else {
782 		reset_gpstates(policy);
783 		gpstates->highest_lpstate_idx = new_index;
784 		gpstate_idx = new_index;
785 	}
786 
787 	/*
788 	 * If local pstate is equal to global pstate, rampdown is over
789 	 * So timer is not required to be queued.
790 	 */
791 	if (gpstate_idx != new_index)
792 		queue_gpstate_timer(gpstates);
793 	else
794 		del_timer_sync(&gpstates->timer);
795 
796 gpstates_done:
797 	freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
798 	gpstates->last_sampled_time = cur_msec;
799 	gpstates->last_gpstate_idx = gpstate_idx;
800 	gpstates->last_lpstate_idx = new_index;
801 
802 	spin_unlock(&gpstates->gpstate_lock);
803 
804 	/*
805 	 * Use smp_call_function to send IPI and execute the
806 	 * mtspr on target CPU.  We could do that without IPI
807 	 * if current CPU is within policy->cpus (core)
808 	 */
809 	smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
810 	return 0;
811 }
812 
813 static int powernv_cpufreq_cpu_init(struct cpufreq_policy *policy)
814 {
815 	int base, i, ret;
816 	struct kernfs_node *kn;
817 	struct global_pstate_info *gpstates;
818 
819 	base = cpu_first_thread_sibling(policy->cpu);
820 
821 	for (i = 0; i < threads_per_core; i++)
822 		cpumask_set_cpu(base + i, policy->cpus);
823 
824 	kn = kernfs_find_and_get(policy->kobj.sd, throttle_attr_grp.name);
825 	if (!kn) {
826 		int ret;
827 
828 		ret = sysfs_create_group(&policy->kobj, &throttle_attr_grp);
829 		if (ret) {
830 			pr_info("Failed to create throttle stats directory for cpu %d\n",
831 				policy->cpu);
832 			return ret;
833 		}
834 	} else {
835 		kernfs_put(kn);
836 	}
837 
838 	gpstates =  kzalloc(sizeof(*gpstates), GFP_KERNEL);
839 	if (!gpstates)
840 		return -ENOMEM;
841 
842 	policy->driver_data = gpstates;
843 
844 	/* initialize timer */
845 	gpstates->policy = policy;
846 	timer_setup(&gpstates->timer, gpstate_timer_handler,
847 		    TIMER_PINNED | TIMER_DEFERRABLE);
848 	gpstates->timer.expires = jiffies +
849 				msecs_to_jiffies(GPSTATE_TIMER_INTERVAL);
850 	spin_lock_init(&gpstates->gpstate_lock);
851 	ret = cpufreq_table_validate_and_show(policy, powernv_freqs);
852 
853 	if (ret < 0) {
854 		kfree(policy->driver_data);
855 		return ret;
856 	}
857 
858 	policy->fast_switch_possible = true;
859 	return ret;
860 }
861 
862 static int powernv_cpufreq_cpu_exit(struct cpufreq_policy *policy)
863 {
864 	/* timer is deleted in cpufreq_cpu_stop() */
865 	kfree(policy->driver_data);
866 
867 	return 0;
868 }
869 
870 static int powernv_cpufreq_reboot_notifier(struct notifier_block *nb,
871 				unsigned long action, void *unused)
872 {
873 	int cpu;
874 	struct cpufreq_policy cpu_policy;
875 
876 	rebooting = true;
877 	for_each_online_cpu(cpu) {
878 		cpufreq_get_policy(&cpu_policy, cpu);
879 		powernv_cpufreq_target_index(&cpu_policy, get_nominal_index());
880 	}
881 
882 	return NOTIFY_DONE;
883 }
884 
885 static struct notifier_block powernv_cpufreq_reboot_nb = {
886 	.notifier_call = powernv_cpufreq_reboot_notifier,
887 };
888 
889 void powernv_cpufreq_work_fn(struct work_struct *work)
890 {
891 	struct chip *chip = container_of(work, struct chip, throttle);
892 	unsigned int cpu;
893 	cpumask_t mask;
894 
895 	get_online_cpus();
896 	cpumask_and(&mask, &chip->mask, cpu_online_mask);
897 	smp_call_function_any(&mask,
898 			      powernv_cpufreq_throttle_check, NULL, 0);
899 
900 	if (!chip->restore)
901 		goto out;
902 
903 	chip->restore = false;
904 	for_each_cpu(cpu, &mask) {
905 		int index;
906 		struct cpufreq_policy policy;
907 
908 		cpufreq_get_policy(&policy, cpu);
909 		index = cpufreq_table_find_index_c(&policy, policy.cur);
910 		powernv_cpufreq_target_index(&policy, index);
911 		cpumask_andnot(&mask, &mask, policy.cpus);
912 	}
913 out:
914 	put_online_cpus();
915 }
916 
917 static int powernv_cpufreq_occ_msg(struct notifier_block *nb,
918 				   unsigned long msg_type, void *_msg)
919 {
920 	struct opal_msg *msg = _msg;
921 	struct opal_occ_msg omsg;
922 	int i;
923 
924 	if (msg_type != OPAL_MSG_OCC)
925 		return 0;
926 
927 	omsg.type = be64_to_cpu(msg->params[0]);
928 
929 	switch (omsg.type) {
930 	case OCC_RESET:
931 		occ_reset = true;
932 		pr_info("OCC (On Chip Controller - enforces hard thermal/power limits) Resetting\n");
933 		/*
934 		 * powernv_cpufreq_throttle_check() is called in
935 		 * target() callback which can detect the throttle state
936 		 * for governors like ondemand.
937 		 * But static governors will not call target() often thus
938 		 * report throttling here.
939 		 */
940 		if (!throttled) {
941 			throttled = true;
942 			pr_warn("CPU frequency is throttled for duration\n");
943 		}
944 
945 		break;
946 	case OCC_LOAD:
947 		pr_info("OCC Loading, CPU frequency is throttled until OCC is started\n");
948 		break;
949 	case OCC_THROTTLE:
950 		omsg.chip = be64_to_cpu(msg->params[1]);
951 		omsg.throttle_status = be64_to_cpu(msg->params[2]);
952 
953 		if (occ_reset) {
954 			occ_reset = false;
955 			throttled = false;
956 			pr_info("OCC Active, CPU frequency is no longer throttled\n");
957 
958 			for (i = 0; i < nr_chips; i++) {
959 				chips[i].restore = true;
960 				schedule_work(&chips[i].throttle);
961 			}
962 
963 			return 0;
964 		}
965 
966 		for (i = 0; i < nr_chips; i++)
967 			if (chips[i].id == omsg.chip)
968 				break;
969 
970 		if (omsg.throttle_status >= 0 &&
971 		    omsg.throttle_status <= OCC_MAX_THROTTLE_STATUS) {
972 			chips[i].throttle_reason = omsg.throttle_status;
973 			chips[i].reason[omsg.throttle_status]++;
974 		}
975 
976 		if (!omsg.throttle_status)
977 			chips[i].restore = true;
978 
979 		schedule_work(&chips[i].throttle);
980 	}
981 	return 0;
982 }
983 
984 static struct notifier_block powernv_cpufreq_opal_nb = {
985 	.notifier_call	= powernv_cpufreq_occ_msg,
986 	.next		= NULL,
987 	.priority	= 0,
988 };
989 
990 static void powernv_cpufreq_stop_cpu(struct cpufreq_policy *policy)
991 {
992 	struct powernv_smp_call_data freq_data;
993 	struct global_pstate_info *gpstates = policy->driver_data;
994 
995 	freq_data.pstate_id = idx_to_pstate(powernv_pstate_info.min);
996 	freq_data.gpstate_id = idx_to_pstate(powernv_pstate_info.min);
997 	smp_call_function_single(policy->cpu, set_pstate, &freq_data, 1);
998 	del_timer_sync(&gpstates->timer);
999 }
1000 
1001 static unsigned int powernv_fast_switch(struct cpufreq_policy *policy,
1002 					unsigned int target_freq)
1003 {
1004 	int index;
1005 	struct powernv_smp_call_data freq_data;
1006 
1007 	index = cpufreq_table_find_index_dl(policy, target_freq);
1008 	freq_data.pstate_id = powernv_freqs[index].driver_data;
1009 	freq_data.gpstate_id = powernv_freqs[index].driver_data;
1010 	set_pstate(&freq_data);
1011 
1012 	return powernv_freqs[index].frequency;
1013 }
1014 
1015 static struct cpufreq_driver powernv_cpufreq_driver = {
1016 	.name		= "powernv-cpufreq",
1017 	.flags		= CPUFREQ_CONST_LOOPS,
1018 	.init		= powernv_cpufreq_cpu_init,
1019 	.exit		= powernv_cpufreq_cpu_exit,
1020 	.verify		= cpufreq_generic_frequency_table_verify,
1021 	.target_index	= powernv_cpufreq_target_index,
1022 	.fast_switch	= powernv_fast_switch,
1023 	.get		= powernv_cpufreq_get,
1024 	.stop_cpu	= powernv_cpufreq_stop_cpu,
1025 	.attr		= powernv_cpu_freq_attr,
1026 };
1027 
1028 static int init_chip_info(void)
1029 {
1030 	unsigned int chip[256];
1031 	unsigned int cpu, i;
1032 	unsigned int prev_chip_id = UINT_MAX;
1033 
1034 	for_each_possible_cpu(cpu) {
1035 		unsigned int id = cpu_to_chip_id(cpu);
1036 
1037 		if (prev_chip_id != id) {
1038 			prev_chip_id = id;
1039 			chip[nr_chips++] = id;
1040 		}
1041 	}
1042 
1043 	chips = kcalloc(nr_chips, sizeof(struct chip), GFP_KERNEL);
1044 	if (!chips)
1045 		return -ENOMEM;
1046 
1047 	for (i = 0; i < nr_chips; i++) {
1048 		chips[i].id = chip[i];
1049 		cpumask_copy(&chips[i].mask, cpumask_of_node(chip[i]));
1050 		INIT_WORK(&chips[i].throttle, powernv_cpufreq_work_fn);
1051 		for_each_cpu(cpu, &chips[i].mask)
1052 			per_cpu(chip_info, cpu) =  &chips[i];
1053 	}
1054 
1055 	return 0;
1056 }
1057 
1058 static inline void clean_chip_info(void)
1059 {
1060 	kfree(chips);
1061 }
1062 
1063 static inline void unregister_all_notifiers(void)
1064 {
1065 	opal_message_notifier_unregister(OPAL_MSG_OCC,
1066 					 &powernv_cpufreq_opal_nb);
1067 	unregister_reboot_notifier(&powernv_cpufreq_reboot_nb);
1068 }
1069 
1070 static int __init powernv_cpufreq_init(void)
1071 {
1072 	int rc = 0;
1073 
1074 	/* Don't probe on pseries (guest) platforms */
1075 	if (!firmware_has_feature(FW_FEATURE_OPAL))
1076 		return -ENODEV;
1077 
1078 	/* Discover pstates from device tree and init */
1079 	rc = init_powernv_pstates();
1080 	if (rc)
1081 		goto out;
1082 
1083 	/* Populate chip info */
1084 	rc = init_chip_info();
1085 	if (rc)
1086 		goto out;
1087 
1088 	register_reboot_notifier(&powernv_cpufreq_reboot_nb);
1089 	opal_message_notifier_register(OPAL_MSG_OCC, &powernv_cpufreq_opal_nb);
1090 
1091 	if (powernv_pstate_info.wof_enabled)
1092 		powernv_cpufreq_driver.boost_enabled = true;
1093 	else
1094 		powernv_cpu_freq_attr[SCALING_BOOST_FREQS_ATTR_INDEX] = NULL;
1095 
1096 	rc = cpufreq_register_driver(&powernv_cpufreq_driver);
1097 	if (rc) {
1098 		pr_info("Failed to register the cpufreq driver (%d)\n", rc);
1099 		goto cleanup_notifiers;
1100 	}
1101 
1102 	if (powernv_pstate_info.wof_enabled)
1103 		cpufreq_enable_boost_support();
1104 
1105 	return 0;
1106 cleanup_notifiers:
1107 	unregister_all_notifiers();
1108 	clean_chip_info();
1109 out:
1110 	pr_info("Platform driver disabled. System does not support PState control\n");
1111 	return rc;
1112 }
1113 module_init(powernv_cpufreq_init);
1114 
1115 static void __exit powernv_cpufreq_exit(void)
1116 {
1117 	cpufreq_unregister_driver(&powernv_cpufreq_driver);
1118 	unregister_all_notifiers();
1119 	clean_chip_info();
1120 }
1121 module_exit(powernv_cpufreq_exit);
1122 
1123 MODULE_LICENSE("GPL");
1124 MODULE_AUTHOR("Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>");
1125