xref: /linux/drivers/cpufreq/powernv-cpufreq.c (revision 611ac726f9ebbb12f2113e5345ef109660954eeb)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * POWERNV cpufreq driver for the IBM POWER processors
4  *
5  * (C) Copyright IBM 2014
6  *
7  * Author: Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>
8  */
9 
10 #define pr_fmt(fmt)	"powernv-cpufreq: " fmt
11 
12 #include <linux/kernel.h>
13 #include <linux/sysfs.h>
14 #include <linux/cpumask.h>
15 #include <linux/module.h>
16 #include <linux/cpufreq.h>
17 #include <linux/smp.h>
18 #include <linux/of.h>
19 #include <linux/reboot.h>
20 #include <linux/slab.h>
21 #include <linux/cpu.h>
22 #include <linux/hashtable.h>
23 #include <trace/events/power.h>
24 
25 #include <asm/cputhreads.h>
26 #include <asm/firmware.h>
27 #include <asm/reg.h>
28 #include <asm/smp.h> /* Required for cpu_sibling_mask() in UP configs */
29 #include <asm/opal.h>
30 #include <linux/timer.h>
31 
32 #define POWERNV_MAX_PSTATES_ORDER  8
33 #define POWERNV_MAX_PSTATES	(1UL << (POWERNV_MAX_PSTATES_ORDER))
34 #define PMSR_PSAFE_ENABLE	(1UL << 30)
35 #define PMSR_SPR_EM_DISABLE	(1UL << 31)
36 #define MAX_PSTATE_SHIFT	32
37 #define LPSTATE_SHIFT		48
38 #define GPSTATE_SHIFT		56
39 
40 #define MAX_RAMP_DOWN_TIME				5120
41 /*
42  * On an idle system we want the global pstate to ramp-down from max value to
43  * min over a span of ~5 secs. Also we want it to initially ramp-down slowly and
44  * then ramp-down rapidly later on.
45  *
46  * This gives a percentage rampdown for time elapsed in milliseconds.
47  * ramp_down_percentage = ((ms * ms) >> 18)
48  *			~= 3.8 * (sec * sec)
49  *
50  * At 0 ms	ramp_down_percent = 0
51  * At 5120 ms	ramp_down_percent = 100
52  */
53 #define ramp_down_percent(time)		((time * time) >> 18)
54 
55 /* Interval after which the timer is queued to bring down global pstate */
56 #define GPSTATE_TIMER_INTERVAL				2000
57 
58 /**
59  * struct global_pstate_info -	Per policy data structure to maintain history of
60  *				global pstates
61  * @highest_lpstate_idx:	The local pstate index from which we are
62  *				ramping down
63  * @elapsed_time:		Time in ms spent in ramping down from
64  *				highest_lpstate_idx
65  * @last_sampled_time:		Time from boot in ms when global pstates were
66  *				last set
67  * @last_lpstate_idx:		Last set value of local pstate and global
68  * @last_gpstate_idx:		pstate in terms of cpufreq table index
69  * @timer:			Is used for ramping down if cpu goes idle for
70  *				a long time with global pstate held high
71  * @gpstate_lock:		A spinlock to maintain synchronization between
72  *				routines called by the timer handler and
73  *				governer's target_index calls
74  * @policy:			Associated CPUFreq policy
75  */
76 struct global_pstate_info {
77 	int highest_lpstate_idx;
78 	unsigned int elapsed_time;
79 	unsigned int last_sampled_time;
80 	int last_lpstate_idx;
81 	int last_gpstate_idx;
82 	spinlock_t gpstate_lock;
83 	struct timer_list timer;
84 	struct cpufreq_policy *policy;
85 };
86 
87 static struct cpufreq_frequency_table powernv_freqs[POWERNV_MAX_PSTATES+1];
88 
89 static DEFINE_HASHTABLE(pstate_revmap, POWERNV_MAX_PSTATES_ORDER);
90 /**
91  * struct pstate_idx_revmap_data: Entry in the hashmap pstate_revmap
92  *				  indexed by a function of pstate id.
93  *
94  * @pstate_id: pstate id for this entry.
95  *
96  * @cpufreq_table_idx: Index into the powernv_freqs
97  *		       cpufreq_frequency_table for frequency
98  *		       corresponding to pstate_id.
99  *
100  * @hentry: hlist_node that hooks this entry into the pstate_revmap
101  *	    hashtable
102  */
103 struct pstate_idx_revmap_data {
104 	u8 pstate_id;
105 	unsigned int cpufreq_table_idx;
106 	struct hlist_node hentry;
107 };
108 
109 static bool rebooting, throttled, occ_reset;
110 
111 static const char * const throttle_reason[] = {
112 	"No throttling",
113 	"Power Cap",
114 	"Processor Over Temperature",
115 	"Power Supply Failure",
116 	"Over Current",
117 	"OCC Reset"
118 };
119 
120 enum throttle_reason_type {
121 	NO_THROTTLE = 0,
122 	POWERCAP,
123 	CPU_OVERTEMP,
124 	POWER_SUPPLY_FAILURE,
125 	OVERCURRENT,
126 	OCC_RESET_THROTTLE,
127 	OCC_MAX_REASON
128 };
129 
130 static struct chip {
131 	unsigned int id;
132 	bool throttled;
133 	bool restore;
134 	u8 throttle_reason;
135 	cpumask_t mask;
136 	struct work_struct throttle;
137 	int throttle_turbo;
138 	int throttle_sub_turbo;
139 	int reason[OCC_MAX_REASON];
140 } *chips;
141 
142 static int nr_chips;
143 static DEFINE_PER_CPU(struct chip *, chip_info);
144 
145 /*
146  * Note:
147  * The set of pstates consists of contiguous integers.
148  * powernv_pstate_info stores the index of the frequency table for
149  * max, min and nominal frequencies. It also stores number of
150  * available frequencies.
151  *
152  * powernv_pstate_info.nominal indicates the index to the highest
153  * non-turbo frequency.
154  */
155 static struct powernv_pstate_info {
156 	unsigned int min;
157 	unsigned int max;
158 	unsigned int nominal;
159 	unsigned int nr_pstates;
160 	bool wof_enabled;
161 } powernv_pstate_info;
162 
163 static inline u8 extract_pstate(u64 pmsr_val, unsigned int shift)
164 {
165 	return ((pmsr_val >> shift) & 0xFF);
166 }
167 
168 #define extract_local_pstate(x) extract_pstate(x, LPSTATE_SHIFT)
169 #define extract_global_pstate(x) extract_pstate(x, GPSTATE_SHIFT)
170 #define extract_max_pstate(x)  extract_pstate(x, MAX_PSTATE_SHIFT)
171 
172 /* Use following functions for conversions between pstate_id and index */
173 
174 /*
175  * idx_to_pstate : Returns the pstate id corresponding to the
176  *		   frequency in the cpufreq frequency table
177  *		   powernv_freqs indexed by @i.
178  *
179  *		   If @i is out of bound, this will return the pstate
180  *		   corresponding to the nominal frequency.
181  */
182 static inline u8 idx_to_pstate(unsigned int i)
183 {
184 	if (unlikely(i >= powernv_pstate_info.nr_pstates)) {
185 		pr_warn_once("idx_to_pstate: index %u is out of bound\n", i);
186 		return powernv_freqs[powernv_pstate_info.nominal].driver_data;
187 	}
188 
189 	return powernv_freqs[i].driver_data;
190 }
191 
192 /*
193  * pstate_to_idx : Returns the index in the cpufreq frequencytable
194  *		   powernv_freqs for the frequency whose corresponding
195  *		   pstate id is @pstate.
196  *
197  *		   If no frequency corresponding to @pstate is found,
198  *		   this will return the index of the nominal
199  *		   frequency.
200  */
201 static unsigned int pstate_to_idx(u8 pstate)
202 {
203 	unsigned int key = pstate % POWERNV_MAX_PSTATES;
204 	struct pstate_idx_revmap_data *revmap_data;
205 
206 	hash_for_each_possible(pstate_revmap, revmap_data, hentry, key) {
207 		if (revmap_data->pstate_id == pstate)
208 			return revmap_data->cpufreq_table_idx;
209 	}
210 
211 	pr_warn_once("pstate_to_idx: pstate 0x%x not found\n", pstate);
212 	return powernv_pstate_info.nominal;
213 }
214 
215 static inline void reset_gpstates(struct cpufreq_policy *policy)
216 {
217 	struct global_pstate_info *gpstates = policy->driver_data;
218 
219 	gpstates->highest_lpstate_idx = 0;
220 	gpstates->elapsed_time = 0;
221 	gpstates->last_sampled_time = 0;
222 	gpstates->last_lpstate_idx = 0;
223 	gpstates->last_gpstate_idx = 0;
224 }
225 
226 /*
227  * Initialize the freq table based on data obtained
228  * from the firmware passed via device-tree
229  */
230 static int init_powernv_pstates(void)
231 {
232 	struct device_node *power_mgt;
233 	int i, nr_pstates = 0;
234 	const __be32 *pstate_ids, *pstate_freqs;
235 	u32 len_ids, len_freqs;
236 	u32 pstate_min, pstate_max, pstate_nominal;
237 	u32 pstate_turbo, pstate_ultra_turbo;
238 	int rc = -ENODEV;
239 
240 	power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
241 	if (!power_mgt) {
242 		pr_warn("power-mgt node not found\n");
243 		return -ENODEV;
244 	}
245 
246 	if (of_property_read_u32(power_mgt, "ibm,pstate-min", &pstate_min)) {
247 		pr_warn("ibm,pstate-min node not found\n");
248 		goto out;
249 	}
250 
251 	if (of_property_read_u32(power_mgt, "ibm,pstate-max", &pstate_max)) {
252 		pr_warn("ibm,pstate-max node not found\n");
253 		goto out;
254 	}
255 
256 	if (of_property_read_u32(power_mgt, "ibm,pstate-nominal",
257 				 &pstate_nominal)) {
258 		pr_warn("ibm,pstate-nominal not found\n");
259 		goto out;
260 	}
261 
262 	if (of_property_read_u32(power_mgt, "ibm,pstate-ultra-turbo",
263 				 &pstate_ultra_turbo)) {
264 		powernv_pstate_info.wof_enabled = false;
265 		goto next;
266 	}
267 
268 	if (of_property_read_u32(power_mgt, "ibm,pstate-turbo",
269 				 &pstate_turbo)) {
270 		powernv_pstate_info.wof_enabled = false;
271 		goto next;
272 	}
273 
274 	if (pstate_turbo == pstate_ultra_turbo)
275 		powernv_pstate_info.wof_enabled = false;
276 	else
277 		powernv_pstate_info.wof_enabled = true;
278 
279 next:
280 	pr_info("cpufreq pstate min 0x%x nominal 0x%x max 0x%x\n", pstate_min,
281 		pstate_nominal, pstate_max);
282 	pr_info("Workload Optimized Frequency is %s in the platform\n",
283 		(powernv_pstate_info.wof_enabled) ? "enabled" : "disabled");
284 
285 	pstate_ids = of_get_property(power_mgt, "ibm,pstate-ids", &len_ids);
286 	if (!pstate_ids) {
287 		pr_warn("ibm,pstate-ids not found\n");
288 		goto out;
289 	}
290 
291 	pstate_freqs = of_get_property(power_mgt, "ibm,pstate-frequencies-mhz",
292 				      &len_freqs);
293 	if (!pstate_freqs) {
294 		pr_warn("ibm,pstate-frequencies-mhz not found\n");
295 		goto out;
296 	}
297 
298 	if (len_ids != len_freqs) {
299 		pr_warn("Entries in ibm,pstate-ids and "
300 			"ibm,pstate-frequencies-mhz does not match\n");
301 	}
302 
303 	nr_pstates = min(len_ids, len_freqs) / sizeof(u32);
304 	if (!nr_pstates) {
305 		pr_warn("No PStates found\n");
306 		goto out;
307 	}
308 
309 	powernv_pstate_info.nr_pstates = nr_pstates;
310 	pr_debug("NR PStates %d\n", nr_pstates);
311 
312 	for (i = 0; i < nr_pstates; i++) {
313 		u32 id = be32_to_cpu(pstate_ids[i]);
314 		u32 freq = be32_to_cpu(pstate_freqs[i]);
315 		struct pstate_idx_revmap_data *revmap_data;
316 		unsigned int key;
317 
318 		pr_debug("PState id %d freq %d MHz\n", id, freq);
319 		powernv_freqs[i].frequency = freq * 1000; /* kHz */
320 		powernv_freqs[i].driver_data = id & 0xFF;
321 
322 		revmap_data = kmalloc(sizeof(*revmap_data), GFP_KERNEL);
323 		if (!revmap_data) {
324 			rc = -ENOMEM;
325 			goto out;
326 		}
327 
328 		revmap_data->pstate_id = id & 0xFF;
329 		revmap_data->cpufreq_table_idx = i;
330 		key = (revmap_data->pstate_id) % POWERNV_MAX_PSTATES;
331 		hash_add(pstate_revmap, &revmap_data->hentry, key);
332 
333 		if (id == pstate_max)
334 			powernv_pstate_info.max = i;
335 		if (id == pstate_nominal)
336 			powernv_pstate_info.nominal = i;
337 		if (id == pstate_min)
338 			powernv_pstate_info.min = i;
339 
340 		if (powernv_pstate_info.wof_enabled && id == pstate_turbo) {
341 			int j;
342 
343 			for (j = i - 1; j >= (int)powernv_pstate_info.max; j--)
344 				powernv_freqs[j].flags = CPUFREQ_BOOST_FREQ;
345 		}
346 	}
347 
348 	/* End of list marker entry */
349 	powernv_freqs[i].frequency = CPUFREQ_TABLE_END;
350 
351 	of_node_put(power_mgt);
352 	return 0;
353 out:
354 	of_node_put(power_mgt);
355 	return rc;
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 static 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  * @t: Timer context used to fetch global pstate info struct
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 static 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 	 * If the timer has migrated to the different cpu then bring
684 	 * it back to one of the policy->cpus
685 	 */
686 	if (!cpumask_test_cpu(raw_smp_processor_id(), policy->cpus)) {
687 		gpstates->timer.expires = jiffies + msecs_to_jiffies(1);
688 		add_timer_on(&gpstates->timer, cpumask_first(policy->cpus));
689 		spin_unlock(&gpstates->gpstate_lock);
690 		return;
691 	}
692 
693 	/*
694 	 * If PMCR was last updated was using fast_swtich then
695 	 * We may have wrong in gpstate->last_lpstate_idx
696 	 * value. Hence, read from PMCR to get correct data.
697 	 */
698 	val = get_pmspr(SPRN_PMCR);
699 	freq_data.gpstate_id = extract_global_pstate(val);
700 	freq_data.pstate_id = extract_local_pstate(val);
701 	if (freq_data.gpstate_id  == freq_data.pstate_id) {
702 		reset_gpstates(policy);
703 		spin_unlock(&gpstates->gpstate_lock);
704 		return;
705 	}
706 
707 	gpstates->last_sampled_time += time_diff;
708 	gpstates->elapsed_time += time_diff;
709 
710 	if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
711 		gpstate_idx = pstate_to_idx(freq_data.pstate_id);
712 		lpstate_idx = gpstate_idx;
713 		reset_gpstates(policy);
714 		gpstates->highest_lpstate_idx = gpstate_idx;
715 	} else {
716 		lpstate_idx = pstate_to_idx(freq_data.pstate_id);
717 		gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
718 						 gpstates->highest_lpstate_idx,
719 						 lpstate_idx);
720 	}
721 	freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
722 	gpstates->last_gpstate_idx = gpstate_idx;
723 	gpstates->last_lpstate_idx = lpstate_idx;
724 	/*
725 	 * If local pstate is equal to global pstate, rampdown is over
726 	 * So timer is not required to be queued.
727 	 */
728 	if (gpstate_idx != gpstates->last_lpstate_idx)
729 		queue_gpstate_timer(gpstates);
730 
731 	set_pstate(&freq_data);
732 	spin_unlock(&gpstates->gpstate_lock);
733 }
734 
735 /*
736  * powernv_cpufreq_target_index: Sets the frequency corresponding to
737  * the cpufreq table entry indexed by new_index on the cpus in the
738  * mask policy->cpus
739  */
740 static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
741 					unsigned int new_index)
742 {
743 	struct powernv_smp_call_data freq_data;
744 	unsigned int cur_msec, gpstate_idx;
745 	struct global_pstate_info *gpstates = policy->driver_data;
746 
747 	if (unlikely(rebooting) && new_index != get_nominal_index())
748 		return 0;
749 
750 	if (!throttled) {
751 		/* we don't want to be preempted while
752 		 * checking if the CPU frequency has been throttled
753 		 */
754 		preempt_disable();
755 		powernv_cpufreq_throttle_check(NULL);
756 		preempt_enable();
757 	}
758 
759 	cur_msec = jiffies_to_msecs(get_jiffies_64());
760 
761 	freq_data.pstate_id = idx_to_pstate(new_index);
762 	if (!gpstates) {
763 		freq_data.gpstate_id = freq_data.pstate_id;
764 		goto no_gpstate;
765 	}
766 
767 	spin_lock(&gpstates->gpstate_lock);
768 
769 	if (!gpstates->last_sampled_time) {
770 		gpstate_idx = new_index;
771 		gpstates->highest_lpstate_idx = new_index;
772 		goto gpstates_done;
773 	}
774 
775 	if (gpstates->last_gpstate_idx < new_index) {
776 		gpstates->elapsed_time += cur_msec -
777 						 gpstates->last_sampled_time;
778 
779 		/*
780 		 * If its has been ramping down for more than MAX_RAMP_DOWN_TIME
781 		 * we should be resetting all global pstate related data. Set it
782 		 * equal to local pstate to start fresh.
783 		 */
784 		if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
785 			reset_gpstates(policy);
786 			gpstates->highest_lpstate_idx = new_index;
787 			gpstate_idx = new_index;
788 		} else {
789 		/* Elaspsed_time is less than 5 seconds, continue to rampdown */
790 			gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
791 							 gpstates->highest_lpstate_idx,
792 							 new_index);
793 		}
794 	} else {
795 		reset_gpstates(policy);
796 		gpstates->highest_lpstate_idx = new_index;
797 		gpstate_idx = new_index;
798 	}
799 
800 	/*
801 	 * If local pstate is equal to global pstate, rampdown is over
802 	 * So timer is not required to be queued.
803 	 */
804 	if (gpstate_idx != new_index)
805 		queue_gpstate_timer(gpstates);
806 	else
807 		del_timer_sync(&gpstates->timer);
808 
809 gpstates_done:
810 	freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
811 	gpstates->last_sampled_time = cur_msec;
812 	gpstates->last_gpstate_idx = gpstate_idx;
813 	gpstates->last_lpstate_idx = new_index;
814 
815 	spin_unlock(&gpstates->gpstate_lock);
816 
817 no_gpstate:
818 	/*
819 	 * Use smp_call_function to send IPI and execute the
820 	 * mtspr on target CPU.  We could do that without IPI
821 	 * if current CPU is within policy->cpus (core)
822 	 */
823 	smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
824 	return 0;
825 }
826 
827 static int powernv_cpufreq_cpu_init(struct cpufreq_policy *policy)
828 {
829 	int base, i;
830 	struct kernfs_node *kn;
831 	struct global_pstate_info *gpstates;
832 
833 	base = cpu_first_thread_sibling(policy->cpu);
834 
835 	for (i = 0; i < threads_per_core; i++)
836 		cpumask_set_cpu(base + i, policy->cpus);
837 
838 	kn = kernfs_find_and_get(policy->kobj.sd, throttle_attr_grp.name);
839 	if (!kn) {
840 		int ret;
841 
842 		ret = sysfs_create_group(&policy->kobj, &throttle_attr_grp);
843 		if (ret) {
844 			pr_info("Failed to create throttle stats directory for cpu %d\n",
845 				policy->cpu);
846 			return ret;
847 		}
848 	} else {
849 		kernfs_put(kn);
850 	}
851 
852 	policy->freq_table = powernv_freqs;
853 	policy->fast_switch_possible = true;
854 
855 	if (pvr_version_is(PVR_POWER9))
856 		return 0;
857 
858 	/* Initialise Gpstate ramp-down timer only on POWER8 */
859 	gpstates =  kzalloc(sizeof(*gpstates), GFP_KERNEL);
860 	if (!gpstates)
861 		return -ENOMEM;
862 
863 	policy->driver_data = gpstates;
864 
865 	/* initialize timer */
866 	gpstates->policy = policy;
867 	timer_setup(&gpstates->timer, gpstate_timer_handler,
868 		    TIMER_PINNED | TIMER_DEFERRABLE);
869 	gpstates->timer.expires = jiffies +
870 				msecs_to_jiffies(GPSTATE_TIMER_INTERVAL);
871 	spin_lock_init(&gpstates->gpstate_lock);
872 
873 	return 0;
874 }
875 
876 static int powernv_cpufreq_cpu_exit(struct cpufreq_policy *policy)
877 {
878 	struct powernv_smp_call_data freq_data;
879 	struct global_pstate_info *gpstates = policy->driver_data;
880 
881 	freq_data.pstate_id = idx_to_pstate(powernv_pstate_info.min);
882 	freq_data.gpstate_id = idx_to_pstate(powernv_pstate_info.min);
883 	smp_call_function_single(policy->cpu, set_pstate, &freq_data, 1);
884 	if (gpstates)
885 		del_timer_sync(&gpstates->timer);
886 
887 	kfree(policy->driver_data);
888 
889 	return 0;
890 }
891 
892 static int powernv_cpufreq_reboot_notifier(struct notifier_block *nb,
893 				unsigned long action, void *unused)
894 {
895 	int cpu;
896 	struct cpufreq_policy *cpu_policy;
897 
898 	rebooting = true;
899 	for_each_online_cpu(cpu) {
900 		cpu_policy = cpufreq_cpu_get(cpu);
901 		if (!cpu_policy)
902 			continue;
903 		powernv_cpufreq_target_index(cpu_policy, get_nominal_index());
904 		cpufreq_cpu_put(cpu_policy);
905 	}
906 
907 	return NOTIFY_DONE;
908 }
909 
910 static struct notifier_block powernv_cpufreq_reboot_nb = {
911 	.notifier_call = powernv_cpufreq_reboot_notifier,
912 };
913 
914 static void powernv_cpufreq_work_fn(struct work_struct *work)
915 {
916 	struct chip *chip = container_of(work, struct chip, throttle);
917 	struct cpufreq_policy *policy;
918 	unsigned int cpu;
919 	cpumask_t mask;
920 
921 	get_online_cpus();
922 	cpumask_and(&mask, &chip->mask, cpu_online_mask);
923 	smp_call_function_any(&mask,
924 			      powernv_cpufreq_throttle_check, NULL, 0);
925 
926 	if (!chip->restore)
927 		goto out;
928 
929 	chip->restore = false;
930 	for_each_cpu(cpu, &mask) {
931 		int index;
932 
933 		policy = cpufreq_cpu_get(cpu);
934 		if (!policy)
935 			continue;
936 		index = cpufreq_table_find_index_c(policy, policy->cur);
937 		powernv_cpufreq_target_index(policy, index);
938 		cpumask_andnot(&mask, &mask, policy->cpus);
939 		cpufreq_cpu_put(policy);
940 	}
941 out:
942 	put_online_cpus();
943 }
944 
945 static int powernv_cpufreq_occ_msg(struct notifier_block *nb,
946 				   unsigned long msg_type, void *_msg)
947 {
948 	struct opal_msg *msg = _msg;
949 	struct opal_occ_msg omsg;
950 	int i;
951 
952 	if (msg_type != OPAL_MSG_OCC)
953 		return 0;
954 
955 	omsg.type = be64_to_cpu(msg->params[0]);
956 
957 	switch (omsg.type) {
958 	case OCC_RESET:
959 		occ_reset = true;
960 		pr_info("OCC (On Chip Controller - enforces hard thermal/power limits) Resetting\n");
961 		/*
962 		 * powernv_cpufreq_throttle_check() is called in
963 		 * target() callback which can detect the throttle state
964 		 * for governors like ondemand.
965 		 * But static governors will not call target() often thus
966 		 * report throttling here.
967 		 */
968 		if (!throttled) {
969 			throttled = true;
970 			pr_warn("CPU frequency is throttled for duration\n");
971 		}
972 
973 		break;
974 	case OCC_LOAD:
975 		pr_info("OCC Loading, CPU frequency is throttled until OCC is started\n");
976 		break;
977 	case OCC_THROTTLE:
978 		omsg.chip = be64_to_cpu(msg->params[1]);
979 		omsg.throttle_status = be64_to_cpu(msg->params[2]);
980 
981 		if (occ_reset) {
982 			occ_reset = false;
983 			throttled = false;
984 			pr_info("OCC Active, CPU frequency is no longer throttled\n");
985 
986 			for (i = 0; i < nr_chips; i++) {
987 				chips[i].restore = true;
988 				schedule_work(&chips[i].throttle);
989 			}
990 
991 			return 0;
992 		}
993 
994 		for (i = 0; i < nr_chips; i++)
995 			if (chips[i].id == omsg.chip)
996 				break;
997 
998 		if (omsg.throttle_status >= 0 &&
999 		    omsg.throttle_status <= OCC_MAX_THROTTLE_STATUS) {
1000 			chips[i].throttle_reason = omsg.throttle_status;
1001 			chips[i].reason[omsg.throttle_status]++;
1002 		}
1003 
1004 		if (!omsg.throttle_status)
1005 			chips[i].restore = true;
1006 
1007 		schedule_work(&chips[i].throttle);
1008 	}
1009 	return 0;
1010 }
1011 
1012 static struct notifier_block powernv_cpufreq_opal_nb = {
1013 	.notifier_call	= powernv_cpufreq_occ_msg,
1014 	.next		= NULL,
1015 	.priority	= 0,
1016 };
1017 
1018 static unsigned int powernv_fast_switch(struct cpufreq_policy *policy,
1019 					unsigned int target_freq)
1020 {
1021 	int index;
1022 	struct powernv_smp_call_data freq_data;
1023 
1024 	index = cpufreq_table_find_index_dl(policy, target_freq);
1025 	freq_data.pstate_id = powernv_freqs[index].driver_data;
1026 	freq_data.gpstate_id = powernv_freqs[index].driver_data;
1027 	set_pstate(&freq_data);
1028 
1029 	return powernv_freqs[index].frequency;
1030 }
1031 
1032 static struct cpufreq_driver powernv_cpufreq_driver = {
1033 	.name		= "powernv-cpufreq",
1034 	.flags		= CPUFREQ_CONST_LOOPS,
1035 	.init		= powernv_cpufreq_cpu_init,
1036 	.exit		= powernv_cpufreq_cpu_exit,
1037 	.verify		= cpufreq_generic_frequency_table_verify,
1038 	.target_index	= powernv_cpufreq_target_index,
1039 	.fast_switch	= powernv_fast_switch,
1040 	.get		= powernv_cpufreq_get,
1041 	.attr		= powernv_cpu_freq_attr,
1042 };
1043 
1044 static int init_chip_info(void)
1045 {
1046 	unsigned int *chip;
1047 	unsigned int cpu, i;
1048 	unsigned int prev_chip_id = UINT_MAX;
1049 	int ret = 0;
1050 
1051 	chip = kcalloc(num_possible_cpus(), sizeof(*chip), GFP_KERNEL);
1052 	if (!chip)
1053 		return -ENOMEM;
1054 
1055 	for_each_possible_cpu(cpu) {
1056 		unsigned int id = cpu_to_chip_id(cpu);
1057 
1058 		if (prev_chip_id != id) {
1059 			prev_chip_id = id;
1060 			chip[nr_chips++] = id;
1061 		}
1062 	}
1063 
1064 	chips = kcalloc(nr_chips, sizeof(struct chip), GFP_KERNEL);
1065 	if (!chips) {
1066 		ret = -ENOMEM;
1067 		goto free_and_return;
1068 	}
1069 
1070 	for (i = 0; i < nr_chips; i++) {
1071 		chips[i].id = chip[i];
1072 		cpumask_copy(&chips[i].mask, cpumask_of_node(chip[i]));
1073 		INIT_WORK(&chips[i].throttle, powernv_cpufreq_work_fn);
1074 		for_each_cpu(cpu, &chips[i].mask)
1075 			per_cpu(chip_info, cpu) =  &chips[i];
1076 	}
1077 
1078 free_and_return:
1079 	kfree(chip);
1080 	return ret;
1081 }
1082 
1083 static inline void clean_chip_info(void)
1084 {
1085 	int i;
1086 
1087 	/* flush any pending work items */
1088 	if (chips)
1089 		for (i = 0; i < nr_chips; i++)
1090 			cancel_work_sync(&chips[i].throttle);
1091 	kfree(chips);
1092 }
1093 
1094 static inline void unregister_all_notifiers(void)
1095 {
1096 	opal_message_notifier_unregister(OPAL_MSG_OCC,
1097 					 &powernv_cpufreq_opal_nb);
1098 	unregister_reboot_notifier(&powernv_cpufreq_reboot_nb);
1099 }
1100 
1101 static int __init powernv_cpufreq_init(void)
1102 {
1103 	int rc = 0;
1104 
1105 	/* Don't probe on pseries (guest) platforms */
1106 	if (!firmware_has_feature(FW_FEATURE_OPAL))
1107 		return -ENODEV;
1108 
1109 	/* Discover pstates from device tree and init */
1110 	rc = init_powernv_pstates();
1111 	if (rc)
1112 		goto out;
1113 
1114 	/* Populate chip info */
1115 	rc = init_chip_info();
1116 	if (rc)
1117 		goto out;
1118 
1119 	if (powernv_pstate_info.wof_enabled)
1120 		powernv_cpufreq_driver.boost_enabled = true;
1121 	else
1122 		powernv_cpu_freq_attr[SCALING_BOOST_FREQS_ATTR_INDEX] = NULL;
1123 
1124 	rc = cpufreq_register_driver(&powernv_cpufreq_driver);
1125 	if (rc) {
1126 		pr_info("Failed to register the cpufreq driver (%d)\n", rc);
1127 		goto cleanup;
1128 	}
1129 
1130 	if (powernv_pstate_info.wof_enabled)
1131 		cpufreq_enable_boost_support();
1132 
1133 	register_reboot_notifier(&powernv_cpufreq_reboot_nb);
1134 	opal_message_notifier_register(OPAL_MSG_OCC, &powernv_cpufreq_opal_nb);
1135 
1136 	return 0;
1137 cleanup:
1138 	clean_chip_info();
1139 out:
1140 	pr_info("Platform driver disabled. System does not support PState control\n");
1141 	return rc;
1142 }
1143 module_init(powernv_cpufreq_init);
1144 
1145 static void __exit powernv_cpufreq_exit(void)
1146 {
1147 	cpufreq_unregister_driver(&powernv_cpufreq_driver);
1148 	unregister_all_notifiers();
1149 	clean_chip_info();
1150 }
1151 module_exit(powernv_cpufreq_exit);
1152 
1153 MODULE_LICENSE("GPL");
1154 MODULE_AUTHOR("Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>");
1155