1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * PowerNV cpuidle code
4 *
5 * Copyright 2015 IBM Corp.
6 */
7
8 #include <linux/types.h>
9 #include <linux/mm.h>
10 #include <linux/slab.h>
11 #include <linux/of.h>
12 #include <linux/device.h>
13 #include <linux/cpu.h>
14
15 #include <asm/asm-prototypes.h>
16 #include <asm/firmware.h>
17 #include <asm/interrupt.h>
18 #include <asm/machdep.h>
19 #include <asm/opal.h>
20 #include <asm/cputhreads.h>
21 #include <asm/cpuidle.h>
22 #include <asm/code-patching.h>
23 #include <asm/smp.h>
24 #include <asm/runlatch.h>
25 #include <asm/dbell.h>
26
27 #include "powernv.h"
28 #include "subcore.h"
29
30 /* Power ISA 3.0 allows for stop states 0x0 - 0xF */
31 #define MAX_STOP_STATE 0xF
32
33 #define P9_STOP_SPR_MSR 2000
34 #define P9_STOP_SPR_PSSCR 855
35
36 static u32 supported_cpuidle_states;
37 struct pnv_idle_states_t *pnv_idle_states;
38 int nr_pnv_idle_states;
39
40 /*
41 * The default stop state that will be used by ppc_md.power_save
42 * function on platforms that support stop instruction.
43 */
44 static u64 pnv_default_stop_val;
45 static u64 pnv_default_stop_mask;
46 static bool default_stop_found;
47
48 /*
49 * First stop state levels when SPR and TB loss can occur.
50 */
51 static u64 pnv_first_tb_loss_level = MAX_STOP_STATE + 1;
52 static u64 deep_spr_loss_state = MAX_STOP_STATE + 1;
53
54 /*
55 * psscr value and mask of the deepest stop idle state.
56 * Used when a cpu is offlined.
57 */
58 static u64 pnv_deepest_stop_psscr_val;
59 static u64 pnv_deepest_stop_psscr_mask;
60 static u64 pnv_deepest_stop_flag;
61 static bool deepest_stop_found;
62
63 static unsigned long power7_offline_type;
64
pnv_save_sprs_for_deep_states(void)65 static int pnv_save_sprs_for_deep_states(void)
66 {
67 int cpu;
68 int rc;
69
70 /*
71 * hid0, hid1, hid4, hid5, hmeer and lpcr values are symmetric across
72 * all cpus at boot. Get these reg values of current cpu and use the
73 * same across all cpus.
74 */
75 uint64_t lpcr_val = mfspr(SPRN_LPCR);
76 uint64_t hid0_val = mfspr(SPRN_HID0);
77 uint64_t hmeer_val = mfspr(SPRN_HMEER);
78 uint64_t msr_val = MSR_IDLE;
79 uint64_t psscr_val = pnv_deepest_stop_psscr_val;
80
81 for_each_present_cpu(cpu) {
82 uint64_t pir = get_hard_smp_processor_id(cpu);
83 uint64_t hsprg0_val = (uint64_t)paca_ptrs[cpu];
84
85 rc = opal_slw_set_reg(pir, SPRN_HSPRG0, hsprg0_val);
86 if (rc != 0)
87 return rc;
88
89 rc = opal_slw_set_reg(pir, SPRN_LPCR, lpcr_val);
90 if (rc != 0)
91 return rc;
92
93 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
94 rc = opal_slw_set_reg(pir, P9_STOP_SPR_MSR, msr_val);
95 if (rc)
96 return rc;
97
98 rc = opal_slw_set_reg(pir,
99 P9_STOP_SPR_PSSCR, psscr_val);
100
101 if (rc)
102 return rc;
103 }
104
105 /* HIDs are per core registers */
106 if (cpu_thread_in_core(cpu) == 0) {
107
108 rc = opal_slw_set_reg(pir, SPRN_HMEER, hmeer_val);
109 if (rc != 0)
110 return rc;
111
112 rc = opal_slw_set_reg(pir, SPRN_HID0, hid0_val);
113 if (rc != 0)
114 return rc;
115
116 /* Only p8 needs to set extra HID regiters */
117 if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
118 uint64_t hid1_val = mfspr(SPRN_HID1);
119 uint64_t hid4_val = mfspr(SPRN_HID4);
120 uint64_t hid5_val = mfspr(SPRN_HID5);
121
122 rc = opal_slw_set_reg(pir, SPRN_HID1, hid1_val);
123 if (rc != 0)
124 return rc;
125
126 rc = opal_slw_set_reg(pir, SPRN_HID4, hid4_val);
127 if (rc != 0)
128 return rc;
129
130 rc = opal_slw_set_reg(pir, SPRN_HID5, hid5_val);
131 if (rc != 0)
132 return rc;
133 }
134 }
135 }
136
137 return 0;
138 }
139
pnv_get_supported_cpuidle_states(void)140 u32 pnv_get_supported_cpuidle_states(void)
141 {
142 return supported_cpuidle_states;
143 }
144 EXPORT_SYMBOL_GPL(pnv_get_supported_cpuidle_states);
145
pnv_fastsleep_workaround_apply(void * info)146 static void pnv_fastsleep_workaround_apply(void *info)
147
148 {
149 int rc;
150 int *err = info;
151
152 rc = opal_config_cpu_idle_state(OPAL_CONFIG_IDLE_FASTSLEEP,
153 OPAL_CONFIG_IDLE_APPLY);
154 if (rc)
155 *err = 1;
156 }
157
158 static bool power7_fastsleep_workaround_entry = true;
159 static bool power7_fastsleep_workaround_exit = true;
160
161 /*
162 * Used to store fastsleep workaround state
163 * 0 - Workaround applied/undone at fastsleep entry/exit path (Default)
164 * 1 - Workaround applied once, never undone.
165 */
166 static u8 fastsleep_workaround_applyonce;
167
show_fastsleep_workaround_applyonce(struct device * dev,struct device_attribute * attr,char * buf)168 static ssize_t show_fastsleep_workaround_applyonce(struct device *dev,
169 struct device_attribute *attr, char *buf)
170 {
171 return sprintf(buf, "%u\n", fastsleep_workaround_applyonce);
172 }
173
store_fastsleep_workaround_applyonce(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)174 static ssize_t store_fastsleep_workaround_applyonce(struct device *dev,
175 struct device_attribute *attr, const char *buf,
176 size_t count)
177 {
178 cpumask_t primary_thread_mask;
179 int err;
180 u8 val;
181
182 if (kstrtou8(buf, 0, &val) || val != 1)
183 return -EINVAL;
184
185 if (fastsleep_workaround_applyonce == 1)
186 return count;
187
188 /*
189 * fastsleep_workaround_applyonce = 1 implies
190 * fastsleep workaround needs to be left in 'applied' state on all
191 * the cores. Do this by-
192 * 1. Disable the 'undo' workaround in fastsleep exit path
193 * 2. Sendi IPIs to all the cores which have at least one online thread
194 * 3. Disable the 'apply' workaround in fastsleep entry path
195 *
196 * There is no need to send ipi to cores which have all threads
197 * offlined, as last thread of the core entering fastsleep or deeper
198 * state would have applied workaround.
199 */
200 power7_fastsleep_workaround_exit = false;
201
202 cpus_read_lock();
203 primary_thread_mask = cpu_online_cores_map();
204 on_each_cpu_mask(&primary_thread_mask,
205 pnv_fastsleep_workaround_apply,
206 &err, 1);
207 cpus_read_unlock();
208 if (err) {
209 pr_err("fastsleep_workaround_applyonce change failed while running pnv_fastsleep_workaround_apply");
210 goto fail;
211 }
212
213 power7_fastsleep_workaround_entry = false;
214
215 fastsleep_workaround_applyonce = 1;
216
217 return count;
218 fail:
219 return -EIO;
220 }
221
222 static DEVICE_ATTR(fastsleep_workaround_applyonce, 0600,
223 show_fastsleep_workaround_applyonce,
224 store_fastsleep_workaround_applyonce);
225
atomic_start_thread_idle(void)226 static inline void atomic_start_thread_idle(void)
227 {
228 int cpu = raw_smp_processor_id();
229 int first = cpu_first_thread_sibling(cpu);
230 int thread_nr = cpu_thread_in_core(cpu);
231 unsigned long *state = &paca_ptrs[first]->idle_state;
232
233 clear_bit(thread_nr, state);
234 }
235
atomic_stop_thread_idle(void)236 static inline void atomic_stop_thread_idle(void)
237 {
238 int cpu = raw_smp_processor_id();
239 int first = cpu_first_thread_sibling(cpu);
240 int thread_nr = cpu_thread_in_core(cpu);
241 unsigned long *state = &paca_ptrs[first]->idle_state;
242
243 set_bit(thread_nr, state);
244 }
245
atomic_lock_thread_idle(void)246 static inline void atomic_lock_thread_idle(void)
247 {
248 int cpu = raw_smp_processor_id();
249 int first = cpu_first_thread_sibling(cpu);
250 unsigned long *state = &paca_ptrs[first]->idle_state;
251
252 while (unlikely(test_and_set_bit_lock(NR_PNV_CORE_IDLE_LOCK_BIT, state)))
253 barrier();
254 }
255
atomic_unlock_and_stop_thread_idle(void)256 static inline void atomic_unlock_and_stop_thread_idle(void)
257 {
258 int cpu = raw_smp_processor_id();
259 int first = cpu_first_thread_sibling(cpu);
260 unsigned long thread = 1UL << cpu_thread_in_core(cpu);
261 unsigned long *state = &paca_ptrs[first]->idle_state;
262 u64 s = READ_ONCE(*state);
263 u64 new, tmp;
264
265 BUG_ON(!(s & PNV_CORE_IDLE_LOCK_BIT));
266 BUG_ON(s & thread);
267
268 again:
269 new = (s | thread) & ~PNV_CORE_IDLE_LOCK_BIT;
270 tmp = cmpxchg(state, s, new);
271 if (unlikely(tmp != s)) {
272 s = tmp;
273 goto again;
274 }
275 }
276
atomic_unlock_thread_idle(void)277 static inline void atomic_unlock_thread_idle(void)
278 {
279 int cpu = raw_smp_processor_id();
280 int first = cpu_first_thread_sibling(cpu);
281 unsigned long *state = &paca_ptrs[first]->idle_state;
282
283 BUG_ON(!test_bit(NR_PNV_CORE_IDLE_LOCK_BIT, state));
284 clear_bit_unlock(NR_PNV_CORE_IDLE_LOCK_BIT, state);
285 }
286
287 /* P7 and P8 */
288 struct p7_sprs {
289 /* per core */
290 u64 tscr;
291 u64 worc;
292
293 /* per subcore */
294 u64 sdr1;
295 u64 rpr;
296
297 /* per thread */
298 u64 lpcr;
299 u64 hfscr;
300 u64 fscr;
301 u64 purr;
302 u64 spurr;
303 u64 dscr;
304 u64 wort;
305
306 /* per thread SPRs that get lost in shallow states */
307 u64 amr;
308 u64 iamr;
309 u64 amor;
310 u64 uamor;
311 };
312
power7_idle_insn(unsigned long type)313 static unsigned long power7_idle_insn(unsigned long type)
314 {
315 int cpu = raw_smp_processor_id();
316 int first = cpu_first_thread_sibling(cpu);
317 unsigned long *state = &paca_ptrs[first]->idle_state;
318 unsigned long thread = 1UL << cpu_thread_in_core(cpu);
319 unsigned long core_thread_mask = (1UL << threads_per_core) - 1;
320 unsigned long srr1;
321 bool full_winkle;
322 struct p7_sprs sprs = {}; /* avoid false use-uninitialised */
323 bool sprs_saved = false;
324 int rc;
325
326 if (unlikely(type != PNV_THREAD_NAP)) {
327 atomic_lock_thread_idle();
328
329 BUG_ON(!(*state & thread));
330 *state &= ~thread;
331
332 if (power7_fastsleep_workaround_entry) {
333 if ((*state & core_thread_mask) == 0) {
334 rc = opal_config_cpu_idle_state(
335 OPAL_CONFIG_IDLE_FASTSLEEP,
336 OPAL_CONFIG_IDLE_APPLY);
337 BUG_ON(rc);
338 }
339 }
340
341 if (type == PNV_THREAD_WINKLE) {
342 sprs.tscr = mfspr(SPRN_TSCR);
343 sprs.worc = mfspr(SPRN_WORC);
344
345 sprs.sdr1 = mfspr(SPRN_SDR1);
346 sprs.rpr = mfspr(SPRN_RPR);
347
348 sprs.lpcr = mfspr(SPRN_LPCR);
349 if (cpu_has_feature(CPU_FTR_ARCH_207S)) {
350 sprs.hfscr = mfspr(SPRN_HFSCR);
351 sprs.fscr = mfspr(SPRN_FSCR);
352 }
353 sprs.purr = mfspr(SPRN_PURR);
354 sprs.spurr = mfspr(SPRN_SPURR);
355 sprs.dscr = mfspr(SPRN_DSCR);
356 sprs.wort = mfspr(SPRN_WORT);
357
358 sprs_saved = true;
359
360 /*
361 * Increment winkle counter and set all winkle bits if
362 * all threads are winkling. This allows wakeup side to
363 * distinguish between fast sleep and winkle state
364 * loss. Fast sleep still has to resync the timebase so
365 * this may not be a really big win.
366 */
367 *state += 1 << PNV_CORE_IDLE_WINKLE_COUNT_SHIFT;
368 if ((*state & PNV_CORE_IDLE_WINKLE_COUNT_BITS)
369 >> PNV_CORE_IDLE_WINKLE_COUNT_SHIFT
370 == threads_per_core)
371 *state |= PNV_CORE_IDLE_THREAD_WINKLE_BITS;
372 WARN_ON((*state & PNV_CORE_IDLE_WINKLE_COUNT_BITS) == 0);
373 }
374
375 atomic_unlock_thread_idle();
376 }
377
378 if (cpu_has_feature(CPU_FTR_ARCH_207S)) {
379 sprs.amr = mfspr(SPRN_AMR);
380 sprs.iamr = mfspr(SPRN_IAMR);
381 sprs.amor = mfspr(SPRN_AMOR);
382 sprs.uamor = mfspr(SPRN_UAMOR);
383 }
384
385 local_paca->thread_idle_state = type;
386 srr1 = isa206_idle_insn_mayloss(type); /* go idle */
387 local_paca->thread_idle_state = PNV_THREAD_RUNNING;
388
389 WARN_ON_ONCE(!srr1);
390 WARN_ON_ONCE(mfmsr() & (MSR_IR|MSR_DR));
391
392 if (cpu_has_feature(CPU_FTR_ARCH_207S)) {
393 if ((srr1 & SRR1_WAKESTATE) != SRR1_WS_NOLOSS) {
394 /*
395 * We don't need an isync after the mtsprs here because
396 * the upcoming mtmsrd is execution synchronizing.
397 */
398 mtspr(SPRN_AMR, sprs.amr);
399 mtspr(SPRN_IAMR, sprs.iamr);
400 mtspr(SPRN_AMOR, sprs.amor);
401 mtspr(SPRN_UAMOR, sprs.uamor);
402 }
403 }
404
405 if (unlikely((srr1 & SRR1_WAKEMASK_P8) == SRR1_WAKEHMI))
406 hmi_exception_realmode(NULL);
407
408 if (likely((srr1 & SRR1_WAKESTATE) != SRR1_WS_HVLOSS)) {
409 if (unlikely(type != PNV_THREAD_NAP)) {
410 atomic_lock_thread_idle();
411 if (type == PNV_THREAD_WINKLE) {
412 WARN_ON((*state & PNV_CORE_IDLE_WINKLE_COUNT_BITS) == 0);
413 *state -= 1 << PNV_CORE_IDLE_WINKLE_COUNT_SHIFT;
414 *state &= ~(thread << PNV_CORE_IDLE_THREAD_WINKLE_BITS_SHIFT);
415 }
416 atomic_unlock_and_stop_thread_idle();
417 }
418 return srr1;
419 }
420
421 /* HV state loss */
422 BUG_ON(type == PNV_THREAD_NAP);
423
424 atomic_lock_thread_idle();
425
426 full_winkle = false;
427 if (type == PNV_THREAD_WINKLE) {
428 WARN_ON((*state & PNV_CORE_IDLE_WINKLE_COUNT_BITS) == 0);
429 *state -= 1 << PNV_CORE_IDLE_WINKLE_COUNT_SHIFT;
430 if (*state & (thread << PNV_CORE_IDLE_THREAD_WINKLE_BITS_SHIFT)) {
431 *state &= ~(thread << PNV_CORE_IDLE_THREAD_WINKLE_BITS_SHIFT);
432 full_winkle = true;
433 BUG_ON(!sprs_saved);
434 }
435 }
436
437 WARN_ON(*state & thread);
438
439 if ((*state & core_thread_mask) != 0)
440 goto core_woken;
441
442 /* Per-core SPRs */
443 if (full_winkle) {
444 mtspr(SPRN_TSCR, sprs.tscr);
445 mtspr(SPRN_WORC, sprs.worc);
446 }
447
448 if (power7_fastsleep_workaround_exit) {
449 rc = opal_config_cpu_idle_state(OPAL_CONFIG_IDLE_FASTSLEEP,
450 OPAL_CONFIG_IDLE_UNDO);
451 BUG_ON(rc);
452 }
453
454 /* TB */
455 if (opal_resync_timebase() != OPAL_SUCCESS)
456 BUG();
457
458 core_woken:
459 if (!full_winkle)
460 goto subcore_woken;
461
462 if ((*state & local_paca->subcore_sibling_mask) != 0)
463 goto subcore_woken;
464
465 /* Per-subcore SPRs */
466 mtspr(SPRN_SDR1, sprs.sdr1);
467 mtspr(SPRN_RPR, sprs.rpr);
468
469 subcore_woken:
470 /*
471 * isync after restoring shared SPRs and before unlocking. Unlock
472 * only contains hwsync which does not necessarily do the right
473 * thing for SPRs.
474 */
475 isync();
476 atomic_unlock_and_stop_thread_idle();
477
478 /* Fast sleep does not lose SPRs */
479 if (!full_winkle)
480 return srr1;
481
482 /* Per-thread SPRs */
483 mtspr(SPRN_LPCR, sprs.lpcr);
484 if (cpu_has_feature(CPU_FTR_ARCH_207S)) {
485 mtspr(SPRN_HFSCR, sprs.hfscr);
486 mtspr(SPRN_FSCR, sprs.fscr);
487 }
488 mtspr(SPRN_PURR, sprs.purr);
489 mtspr(SPRN_SPURR, sprs.spurr);
490 mtspr(SPRN_DSCR, sprs.dscr);
491 mtspr(SPRN_WORT, sprs.wort);
492
493 mtspr(SPRN_SPRG3, local_paca->sprg_vdso);
494
495 /*
496 * The SLB has to be restored here, but it sometimes still
497 * contains entries, so the __ variant must be used to prevent
498 * multi hits.
499 */
500 __slb_restore_bolted_realmode();
501
502 return srr1;
503 }
504
505 extern unsigned long idle_kvm_start_guest(unsigned long srr1);
506
507 #ifdef CONFIG_HOTPLUG_CPU
power7_offline(void)508 static unsigned long power7_offline(void)
509 {
510 unsigned long srr1;
511
512 mtmsr(MSR_IDLE);
513
514 #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
515 /* Tell KVM we're entering idle. */
516 /******************************************************/
517 /* N O T E W E L L ! ! ! N O T E W E L L */
518 /* The following store to HSTATE_HWTHREAD_STATE(r13) */
519 /* MUST occur in real mode, i.e. with the MMU off, */
520 /* and the MMU must stay off until we clear this flag */
521 /* and test HSTATE_HWTHREAD_REQ(r13) in */
522 /* pnv_powersave_wakeup in this file. */
523 /* The reason is that another thread can switch the */
524 /* MMU to a guest context whenever this flag is set */
525 /* to KVM_HWTHREAD_IN_IDLE, and if the MMU was on, */
526 /* that would potentially cause this thread to start */
527 /* executing instructions from guest memory in */
528 /* hypervisor mode, leading to a host crash or data */
529 /* corruption, or worse. */
530 /******************************************************/
531 local_paca->kvm_hstate.hwthread_state = KVM_HWTHREAD_IN_IDLE;
532 #endif
533
534 __ppc64_runlatch_off();
535 srr1 = power7_idle_insn(power7_offline_type);
536 __ppc64_runlatch_on();
537
538 #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
539 local_paca->kvm_hstate.hwthread_state = KVM_HWTHREAD_IN_KERNEL;
540 /* Order setting hwthread_state vs. testing hwthread_req */
541 smp_mb();
542 if (local_paca->kvm_hstate.hwthread_req)
543 srr1 = idle_kvm_start_guest(srr1);
544 #endif
545
546 mtmsr(MSR_KERNEL);
547
548 return srr1;
549 }
550 #endif
551
power7_idle_type(unsigned long type)552 void power7_idle_type(unsigned long type)
553 {
554 unsigned long srr1;
555
556 if (!prep_irq_for_idle_irqsoff())
557 return;
558
559 mtmsr(MSR_IDLE);
560 __ppc64_runlatch_off();
561 srr1 = power7_idle_insn(type);
562 __ppc64_runlatch_on();
563 mtmsr(MSR_KERNEL);
564
565 fini_irq_for_idle_irqsoff();
566 irq_set_pending_from_srr1(srr1);
567 }
568
power7_idle(void)569 static void power7_idle(void)
570 {
571 if (!powersave_nap)
572 return;
573
574 power7_idle_type(PNV_THREAD_NAP);
575 }
576
577 struct p9_sprs {
578 /* per core */
579 u64 ptcr;
580 u64 rpr;
581 u64 tscr;
582 u64 ldbar;
583
584 /* per thread */
585 u64 lpcr;
586 u64 hfscr;
587 u64 fscr;
588 u64 pid;
589 u64 purr;
590 u64 spurr;
591 u64 dscr;
592 u64 wort;
593 u64 ciabr;
594
595 u64 mmcra;
596 u32 mmcr0;
597 u32 mmcr1;
598 u64 mmcr2;
599
600 /* per thread SPRs that get lost in shallow states */
601 u64 amr;
602 u64 iamr;
603 u64 amor;
604 u64 uamor;
605 };
606
power9_idle_stop(unsigned long psscr)607 static unsigned long power9_idle_stop(unsigned long psscr)
608 {
609 int cpu = raw_smp_processor_id();
610 int first = cpu_first_thread_sibling(cpu);
611 unsigned long *state = &paca_ptrs[first]->idle_state;
612 unsigned long core_thread_mask = (1UL << threads_per_core) - 1;
613 unsigned long srr1;
614 unsigned long pls;
615 unsigned long mmcr0 = 0;
616 unsigned long mmcra = 0;
617 struct p9_sprs sprs = {}; /* avoid false used-uninitialised */
618 bool sprs_saved = false;
619
620 if (!(psscr & (PSSCR_EC|PSSCR_ESL))) {
621 /* EC=ESL=0 case */
622
623 /*
624 * Wake synchronously. SRESET via xscom may still cause
625 * a 0x100 powersave wakeup with SRR1 reason!
626 */
627 srr1 = isa300_idle_stop_noloss(psscr); /* go idle */
628 if (likely(!srr1))
629 return 0;
630
631 /*
632 * Registers not saved, can't recover!
633 * This would be a hardware bug
634 */
635 BUG_ON((srr1 & SRR1_WAKESTATE) != SRR1_WS_NOLOSS);
636
637 goto out;
638 }
639
640 /* EC=ESL=1 case */
641 #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
642 if (cpu_has_feature(CPU_FTR_P9_TM_XER_SO_BUG)) {
643 local_paca->requested_psscr = psscr;
644 /* order setting requested_psscr vs testing dont_stop */
645 smp_mb();
646 if (atomic_read(&local_paca->dont_stop)) {
647 local_paca->requested_psscr = 0;
648 return 0;
649 }
650 }
651 #endif
652
653 if (!cpu_has_feature(CPU_FTR_POWER9_DD2_1)) {
654 /*
655 * POWER9 DD2 can incorrectly set PMAO when waking up
656 * after a state-loss idle. Saving and restoring MMCR0
657 * over idle is a workaround.
658 */
659 mmcr0 = mfspr(SPRN_MMCR0);
660 }
661
662 if ((psscr & PSSCR_RL_MASK) >= deep_spr_loss_state) {
663 sprs.lpcr = mfspr(SPRN_LPCR);
664 sprs.hfscr = mfspr(SPRN_HFSCR);
665 sprs.fscr = mfspr(SPRN_FSCR);
666 sprs.pid = mfspr(SPRN_PID);
667 sprs.purr = mfspr(SPRN_PURR);
668 sprs.spurr = mfspr(SPRN_SPURR);
669 sprs.dscr = mfspr(SPRN_DSCR);
670 sprs.ciabr = mfspr(SPRN_CIABR);
671
672 sprs.mmcra = mfspr(SPRN_MMCRA);
673 sprs.mmcr0 = mfspr(SPRN_MMCR0);
674 sprs.mmcr1 = mfspr(SPRN_MMCR1);
675 sprs.mmcr2 = mfspr(SPRN_MMCR2);
676
677 sprs.ptcr = mfspr(SPRN_PTCR);
678 sprs.rpr = mfspr(SPRN_RPR);
679 sprs.tscr = mfspr(SPRN_TSCR);
680 if (!firmware_has_feature(FW_FEATURE_ULTRAVISOR))
681 sprs.ldbar = mfspr(SPRN_LDBAR);
682
683 sprs_saved = true;
684
685 atomic_start_thread_idle();
686 }
687
688 sprs.amr = mfspr(SPRN_AMR);
689 sprs.iamr = mfspr(SPRN_IAMR);
690 sprs.amor = mfspr(SPRN_AMOR);
691 sprs.uamor = mfspr(SPRN_UAMOR);
692
693 srr1 = isa300_idle_stop_mayloss(psscr); /* go idle */
694
695 #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
696 local_paca->requested_psscr = 0;
697 #endif
698
699 psscr = mfspr(SPRN_PSSCR);
700
701 WARN_ON_ONCE(!srr1);
702 WARN_ON_ONCE(mfmsr() & (MSR_IR|MSR_DR));
703
704 if ((srr1 & SRR1_WAKESTATE) != SRR1_WS_NOLOSS) {
705 /*
706 * We don't need an isync after the mtsprs here because the
707 * upcoming mtmsrd is execution synchronizing.
708 */
709 mtspr(SPRN_AMR, sprs.amr);
710 mtspr(SPRN_IAMR, sprs.iamr);
711 mtspr(SPRN_AMOR, sprs.amor);
712 mtspr(SPRN_UAMOR, sprs.uamor);
713
714 /*
715 * Workaround for POWER9 DD2.0, if we lost resources, the ERAT
716 * might have been corrupted and needs flushing. We also need
717 * to reload MMCR0 (see mmcr0 comment above).
718 */
719 if (!cpu_has_feature(CPU_FTR_POWER9_DD2_1)) {
720 asm volatile(PPC_ISA_3_0_INVALIDATE_ERAT);
721 mtspr(SPRN_MMCR0, mmcr0);
722 }
723
724 /*
725 * DD2.2 and earlier need to set then clear bit 60 in MMCRA
726 * to ensure the PMU starts running.
727 */
728 mmcra = mfspr(SPRN_MMCRA);
729 mmcra |= PPC_BIT(60);
730 mtspr(SPRN_MMCRA, mmcra);
731 mmcra &= ~PPC_BIT(60);
732 mtspr(SPRN_MMCRA, mmcra);
733 }
734
735 if (unlikely((srr1 & SRR1_WAKEMASK_P8) == SRR1_WAKEHMI))
736 hmi_exception_realmode(NULL);
737
738 /*
739 * On POWER9, SRR1 bits do not match exactly as expected.
740 * SRR1_WS_GPRLOSS (10b) can also result in SPR loss, so
741 * just always test PSSCR for SPR/TB state loss.
742 */
743 pls = (psscr & PSSCR_PLS) >> PSSCR_PLS_SHIFT;
744 if (likely(pls < deep_spr_loss_state)) {
745 if (sprs_saved)
746 atomic_stop_thread_idle();
747 goto out;
748 }
749
750 /* HV state loss */
751 BUG_ON(!sprs_saved);
752
753 atomic_lock_thread_idle();
754
755 if ((*state & core_thread_mask) != 0)
756 goto core_woken;
757
758 /* Per-core SPRs */
759 mtspr(SPRN_PTCR, sprs.ptcr);
760 mtspr(SPRN_RPR, sprs.rpr);
761 mtspr(SPRN_TSCR, sprs.tscr);
762
763 if (pls >= pnv_first_tb_loss_level) {
764 /* TB loss */
765 if (opal_resync_timebase() != OPAL_SUCCESS)
766 BUG();
767 }
768
769 /*
770 * isync after restoring shared SPRs and before unlocking. Unlock
771 * only contains hwsync which does not necessarily do the right
772 * thing for SPRs.
773 */
774 isync();
775
776 core_woken:
777 atomic_unlock_and_stop_thread_idle();
778
779 /* Per-thread SPRs */
780 mtspr(SPRN_LPCR, sprs.lpcr);
781 mtspr(SPRN_HFSCR, sprs.hfscr);
782 mtspr(SPRN_FSCR, sprs.fscr);
783 mtspr(SPRN_PID, sprs.pid);
784 mtspr(SPRN_PURR, sprs.purr);
785 mtspr(SPRN_SPURR, sprs.spurr);
786 mtspr(SPRN_DSCR, sprs.dscr);
787 mtspr(SPRN_CIABR, sprs.ciabr);
788
789 mtspr(SPRN_MMCRA, sprs.mmcra);
790 mtspr(SPRN_MMCR0, sprs.mmcr0);
791 mtspr(SPRN_MMCR1, sprs.mmcr1);
792 mtspr(SPRN_MMCR2, sprs.mmcr2);
793 if (!firmware_has_feature(FW_FEATURE_ULTRAVISOR))
794 mtspr(SPRN_LDBAR, sprs.ldbar);
795
796 mtspr(SPRN_SPRG3, local_paca->sprg_vdso);
797
798 if (!radix_enabled())
799 __slb_restore_bolted_realmode();
800
801 out:
802 mtmsr(MSR_KERNEL);
803
804 return srr1;
805 }
806
807 #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
808 /*
809 * This is used in working around bugs in thread reconfiguration
810 * on POWER9 (at least up to Nimbus DD2.2) relating to transactional
811 * memory and the way that XER[SO] is checkpointed.
812 * This function forces the core into SMT4 in order by asking
813 * all other threads not to stop, and sending a message to any
814 * that are in a stop state.
815 * Must be called with preemption disabled.
816 */
pnv_power9_force_smt4_catch(void)817 void pnv_power9_force_smt4_catch(void)
818 {
819 int cpu, cpu0, thr;
820 int awake_threads = 1; /* this thread is awake */
821 int poke_threads = 0;
822 int need_awake = threads_per_core;
823
824 cpu = smp_processor_id();
825 cpu0 = cpu & ~(threads_per_core - 1);
826 for (thr = 0; thr < threads_per_core; ++thr) {
827 if (cpu != cpu0 + thr)
828 atomic_inc(&paca_ptrs[cpu0+thr]->dont_stop);
829 }
830 /* order setting dont_stop vs testing requested_psscr */
831 smp_mb();
832 for (thr = 0; thr < threads_per_core; ++thr) {
833 if (!paca_ptrs[cpu0+thr]->requested_psscr)
834 ++awake_threads;
835 else
836 poke_threads |= (1 << thr);
837 }
838
839 /* If at least 3 threads are awake, the core is in SMT4 already */
840 if (awake_threads < need_awake) {
841 /* We have to wake some threads; we'll use msgsnd */
842 for (thr = 0; thr < threads_per_core; ++thr) {
843 if (poke_threads & (1 << thr)) {
844 ppc_msgsnd_sync();
845 ppc_msgsnd(PPC_DBELL_MSGTYPE, 0,
846 paca_ptrs[cpu0+thr]->hw_cpu_id);
847 }
848 }
849 /* now spin until at least 3 threads are awake */
850 do {
851 for (thr = 0; thr < threads_per_core; ++thr) {
852 if ((poke_threads & (1 << thr)) &&
853 !paca_ptrs[cpu0+thr]->requested_psscr) {
854 ++awake_threads;
855 poke_threads &= ~(1 << thr);
856 }
857 }
858 } while (awake_threads < need_awake);
859 }
860 }
861 EXPORT_SYMBOL_GPL(pnv_power9_force_smt4_catch);
862
pnv_power9_force_smt4_release(void)863 void pnv_power9_force_smt4_release(void)
864 {
865 int cpu, cpu0, thr;
866
867 cpu = smp_processor_id();
868 cpu0 = cpu & ~(threads_per_core - 1);
869
870 /* clear all the dont_stop flags */
871 for (thr = 0; thr < threads_per_core; ++thr) {
872 if (cpu != cpu0 + thr)
873 atomic_dec(&paca_ptrs[cpu0+thr]->dont_stop);
874 }
875 }
876 EXPORT_SYMBOL_GPL(pnv_power9_force_smt4_release);
877 #endif /* CONFIG_KVM_BOOK3S_HV_POSSIBLE */
878
879 struct p10_sprs {
880 /*
881 * SPRs that get lost in shallow states:
882 *
883 * P10 loses CR, LR, CTR, FPSCR, VSCR, XER, TAR, SPRG2, and HSPRG1
884 * isa300 idle routines restore CR, LR.
885 * CTR is volatile
886 * idle thread doesn't use FP or VEC
887 * kernel doesn't use TAR
888 * HSPRG1 is only live in HV interrupt entry
889 * SPRG2 is only live in KVM guests, KVM handles it.
890 */
891 };
892
power10_idle_stop(unsigned long psscr)893 static unsigned long power10_idle_stop(unsigned long psscr)
894 {
895 int cpu = raw_smp_processor_id();
896 int first = cpu_first_thread_sibling(cpu);
897 unsigned long *state = &paca_ptrs[first]->idle_state;
898 unsigned long core_thread_mask = (1UL << threads_per_core) - 1;
899 unsigned long srr1;
900 unsigned long pls;
901 // struct p10_sprs sprs = {}; /* avoid false used-uninitialised */
902 bool sprs_saved = false;
903
904 if (!(psscr & (PSSCR_EC|PSSCR_ESL))) {
905 /* EC=ESL=0 case */
906
907 /*
908 * Wake synchronously. SRESET via xscom may still cause
909 * a 0x100 powersave wakeup with SRR1 reason!
910 */
911 srr1 = isa300_idle_stop_noloss(psscr); /* go idle */
912 if (likely(!srr1))
913 return 0;
914
915 /*
916 * Registers not saved, can't recover!
917 * This would be a hardware bug
918 */
919 BUG_ON((srr1 & SRR1_WAKESTATE) != SRR1_WS_NOLOSS);
920
921 goto out;
922 }
923
924 /* EC=ESL=1 case */
925 if ((psscr & PSSCR_RL_MASK) >= deep_spr_loss_state) {
926 /* XXX: save SPRs for deep state loss here. */
927
928 sprs_saved = true;
929
930 atomic_start_thread_idle();
931 }
932
933 srr1 = isa300_idle_stop_mayloss(psscr); /* go idle */
934
935 psscr = mfspr(SPRN_PSSCR);
936
937 WARN_ON_ONCE(!srr1);
938 WARN_ON_ONCE(mfmsr() & (MSR_IR|MSR_DR));
939
940 if (unlikely((srr1 & SRR1_WAKEMASK_P8) == SRR1_WAKEHMI))
941 hmi_exception_realmode(NULL);
942
943 /*
944 * On POWER10, SRR1 bits do not match exactly as expected.
945 * SRR1_WS_GPRLOSS (10b) can also result in SPR loss, so
946 * just always test PSSCR for SPR/TB state loss.
947 */
948 pls = (psscr & PSSCR_PLS) >> PSSCR_PLS_SHIFT;
949 if (likely(pls < deep_spr_loss_state)) {
950 if (sprs_saved)
951 atomic_stop_thread_idle();
952 goto out;
953 }
954
955 /* HV state loss */
956 BUG_ON(!sprs_saved);
957
958 atomic_lock_thread_idle();
959
960 if ((*state & core_thread_mask) != 0)
961 goto core_woken;
962
963 /* XXX: restore per-core SPRs here */
964
965 if (pls >= pnv_first_tb_loss_level) {
966 /* TB loss */
967 if (opal_resync_timebase() != OPAL_SUCCESS)
968 BUG();
969 }
970
971 /*
972 * isync after restoring shared SPRs and before unlocking. Unlock
973 * only contains hwsync which does not necessarily do the right
974 * thing for SPRs.
975 */
976 isync();
977
978 core_woken:
979 atomic_unlock_and_stop_thread_idle();
980
981 /* XXX: restore per-thread SPRs here */
982
983 if (!radix_enabled())
984 __slb_restore_bolted_realmode();
985
986 out:
987 mtmsr(MSR_KERNEL);
988
989 return srr1;
990 }
991
992 #ifdef CONFIG_HOTPLUG_CPU
arch300_offline_stop(unsigned long psscr)993 static unsigned long arch300_offline_stop(unsigned long psscr)
994 {
995 unsigned long srr1;
996
997 if (cpu_has_feature(CPU_FTR_ARCH_31))
998 srr1 = power10_idle_stop(psscr);
999 else
1000 srr1 = power9_idle_stop(psscr);
1001
1002 return srr1;
1003 }
1004 #endif
1005
arch300_idle_type(unsigned long stop_psscr_val,unsigned long stop_psscr_mask)1006 void arch300_idle_type(unsigned long stop_psscr_val,
1007 unsigned long stop_psscr_mask)
1008 {
1009 unsigned long psscr;
1010 unsigned long srr1;
1011
1012 if (!prep_irq_for_idle_irqsoff())
1013 return;
1014
1015 psscr = mfspr(SPRN_PSSCR);
1016 psscr = (psscr & ~stop_psscr_mask) | stop_psscr_val;
1017
1018 __ppc64_runlatch_off();
1019 if (cpu_has_feature(CPU_FTR_ARCH_31))
1020 srr1 = power10_idle_stop(psscr);
1021 else
1022 srr1 = power9_idle_stop(psscr);
1023 __ppc64_runlatch_on();
1024
1025 fini_irq_for_idle_irqsoff();
1026
1027 irq_set_pending_from_srr1(srr1);
1028 }
1029
1030 /*
1031 * Used for ppc_md.power_save which needs a function with no parameters
1032 */
arch300_idle(void)1033 static void arch300_idle(void)
1034 {
1035 arch300_idle_type(pnv_default_stop_val, pnv_default_stop_mask);
1036 }
1037
1038 #ifdef CONFIG_HOTPLUG_CPU
1039
pnv_program_cpu_hotplug_lpcr(unsigned int cpu,u64 lpcr_val)1040 void pnv_program_cpu_hotplug_lpcr(unsigned int cpu, u64 lpcr_val)
1041 {
1042 u64 pir = get_hard_smp_processor_id(cpu);
1043
1044 mtspr(SPRN_LPCR, lpcr_val);
1045
1046 /*
1047 * Program the LPCR via stop-api only if the deepest stop state
1048 * can lose hypervisor context.
1049 */
1050 if (supported_cpuidle_states & OPAL_PM_LOSE_FULL_CONTEXT)
1051 opal_slw_set_reg(pir, SPRN_LPCR, lpcr_val);
1052 }
1053
1054 /*
1055 * pnv_cpu_offline: A function that puts the CPU into the deepest
1056 * available platform idle state on a CPU-Offline.
1057 * interrupts hard disabled and no lazy irq pending.
1058 */
pnv_cpu_offline(unsigned int cpu)1059 unsigned long pnv_cpu_offline(unsigned int cpu)
1060 {
1061 unsigned long srr1;
1062
1063 __ppc64_runlatch_off();
1064
1065 if (cpu_has_feature(CPU_FTR_ARCH_300) && deepest_stop_found) {
1066 unsigned long psscr;
1067
1068 psscr = mfspr(SPRN_PSSCR);
1069 psscr = (psscr & ~pnv_deepest_stop_psscr_mask) |
1070 pnv_deepest_stop_psscr_val;
1071 srr1 = arch300_offline_stop(psscr);
1072 } else if (cpu_has_feature(CPU_FTR_ARCH_206) && power7_offline_type) {
1073 srr1 = power7_offline();
1074 } else {
1075 /* This is the fallback method. We emulate snooze */
1076 while (!generic_check_cpu_restart(cpu)) {
1077 HMT_low();
1078 HMT_very_low();
1079 }
1080 srr1 = 0;
1081 HMT_medium();
1082 }
1083
1084 __ppc64_runlatch_on();
1085
1086 return srr1;
1087 }
1088 #endif
1089
1090 /*
1091 * Power ISA 3.0 idle initialization.
1092 *
1093 * POWER ISA 3.0 defines a new SPR Processor stop Status and Control
1094 * Register (PSSCR) to control idle behavior.
1095 *
1096 * PSSCR layout:
1097 * ----------------------------------------------------------
1098 * | PLS | /// | SD | ESL | EC | PSLL | /// | TR | MTL | RL |
1099 * ----------------------------------------------------------
1100 * 0 4 41 42 43 44 48 54 56 60
1101 *
1102 * PSSCR key fields:
1103 * Bits 0:3 - Power-Saving Level Status (PLS). This field indicates the
1104 * lowest power-saving state the thread entered since stop instruction was
1105 * last executed.
1106 *
1107 * Bit 41 - Status Disable(SD)
1108 * 0 - Shows PLS entries
1109 * 1 - PLS entries are all 0
1110 *
1111 * Bit 42 - Enable State Loss
1112 * 0 - No state is lost irrespective of other fields
1113 * 1 - Allows state loss
1114 *
1115 * Bit 43 - Exit Criterion
1116 * 0 - Exit from power-save mode on any interrupt
1117 * 1 - Exit from power-save mode controlled by LPCR's PECE bits
1118 *
1119 * Bits 44:47 - Power-Saving Level Limit
1120 * This limits the power-saving level that can be entered into.
1121 *
1122 * Bits 60:63 - Requested Level
1123 * Used to specify which power-saving level must be entered on executing
1124 * stop instruction
1125 */
1126
validate_psscr_val_mask(u64 * psscr_val,u64 * psscr_mask,u32 flags)1127 int validate_psscr_val_mask(u64 *psscr_val, u64 *psscr_mask, u32 flags)
1128 {
1129 int err = 0;
1130
1131 /*
1132 * psscr_mask == 0xf indicates an older firmware.
1133 * Set remaining fields of psscr to the default values.
1134 * See NOTE above definition of PSSCR_HV_DEFAULT_VAL
1135 */
1136 if (*psscr_mask == 0xf) {
1137 *psscr_val = *psscr_val | PSSCR_HV_DEFAULT_VAL;
1138 *psscr_mask = PSSCR_HV_DEFAULT_MASK;
1139 return err;
1140 }
1141
1142 /*
1143 * New firmware is expected to set the psscr_val bits correctly.
1144 * Validate that the following invariants are correctly maintained by
1145 * the new firmware.
1146 * - ESL bit value matches the EC bit value.
1147 * - ESL bit is set for all the deep stop states.
1148 */
1149 if (GET_PSSCR_ESL(*psscr_val) != GET_PSSCR_EC(*psscr_val)) {
1150 err = ERR_EC_ESL_MISMATCH;
1151 } else if ((flags & OPAL_PM_LOSE_FULL_CONTEXT) &&
1152 GET_PSSCR_ESL(*psscr_val) == 0) {
1153 err = ERR_DEEP_STATE_ESL_MISMATCH;
1154 }
1155
1156 return err;
1157 }
1158
1159 /*
1160 * pnv_arch300_idle_init: Initializes the default idle state, first
1161 * deep idle state and deepest idle state on
1162 * ISA 3.0 CPUs.
1163 *
1164 * @np: /ibm,opal/power-mgt device node
1165 * @flags: cpu-idle-state-flags array
1166 * @dt_idle_states: Number of idle state entries
1167 * Returns 0 on success
1168 */
pnv_arch300_idle_init(void)1169 static void __init pnv_arch300_idle_init(void)
1170 {
1171 u64 max_residency_ns = 0;
1172 int i;
1173
1174 /* stop is not really architected, we only have p9,p10 drivers */
1175 if (!pvr_version_is(PVR_POWER10) && !pvr_version_is(PVR_POWER9))
1176 return;
1177
1178 /*
1179 * pnv_deepest_stop_{val,mask} should be set to values corresponding to
1180 * the deepest stop state.
1181 *
1182 * pnv_default_stop_{val,mask} should be set to values corresponding to
1183 * the deepest loss-less (OPAL_PM_STOP_INST_FAST) stop state.
1184 */
1185 pnv_first_tb_loss_level = MAX_STOP_STATE + 1;
1186 deep_spr_loss_state = MAX_STOP_STATE + 1;
1187 for (i = 0; i < nr_pnv_idle_states; i++) {
1188 int err;
1189 struct pnv_idle_states_t *state = &pnv_idle_states[i];
1190 u64 psscr_rl = state->psscr_val & PSSCR_RL_MASK;
1191
1192 /* No deep loss driver implemented for POWER10 yet */
1193 if (pvr_version_is(PVR_POWER10) &&
1194 state->flags & (OPAL_PM_TIMEBASE_STOP|OPAL_PM_LOSE_FULL_CONTEXT))
1195 continue;
1196
1197 if ((state->flags & OPAL_PM_TIMEBASE_STOP) &&
1198 (pnv_first_tb_loss_level > psscr_rl))
1199 pnv_first_tb_loss_level = psscr_rl;
1200
1201 if ((state->flags & OPAL_PM_LOSE_FULL_CONTEXT) &&
1202 (deep_spr_loss_state > psscr_rl))
1203 deep_spr_loss_state = psscr_rl;
1204
1205 /*
1206 * The idle code does not deal with TB loss occurring
1207 * in a shallower state than SPR loss, so force it to
1208 * behave like SPRs are lost if TB is lost. POWER9 would
1209 * never encouter this, but a POWER8 core would if it
1210 * implemented the stop instruction. So this is for forward
1211 * compatibility.
1212 */
1213 if ((state->flags & OPAL_PM_TIMEBASE_STOP) &&
1214 (deep_spr_loss_state > psscr_rl))
1215 deep_spr_loss_state = psscr_rl;
1216
1217 err = validate_psscr_val_mask(&state->psscr_val,
1218 &state->psscr_mask,
1219 state->flags);
1220 if (err) {
1221 report_invalid_psscr_val(state->psscr_val, err);
1222 continue;
1223 }
1224
1225 state->valid = true;
1226
1227 if (max_residency_ns < state->residency_ns) {
1228 max_residency_ns = state->residency_ns;
1229 pnv_deepest_stop_psscr_val = state->psscr_val;
1230 pnv_deepest_stop_psscr_mask = state->psscr_mask;
1231 pnv_deepest_stop_flag = state->flags;
1232 deepest_stop_found = true;
1233 }
1234
1235 if (!default_stop_found &&
1236 (state->flags & OPAL_PM_STOP_INST_FAST)) {
1237 pnv_default_stop_val = state->psscr_val;
1238 pnv_default_stop_mask = state->psscr_mask;
1239 default_stop_found = true;
1240 WARN_ON(state->flags & OPAL_PM_LOSE_FULL_CONTEXT);
1241 }
1242 }
1243
1244 if (unlikely(!default_stop_found)) {
1245 pr_warn("cpuidle-powernv: No suitable default stop state found. Disabling platform idle.\n");
1246 } else {
1247 ppc_md.power_save = arch300_idle;
1248 pr_info("cpuidle-powernv: Default stop: psscr = 0x%016llx,mask=0x%016llx\n",
1249 pnv_default_stop_val, pnv_default_stop_mask);
1250 }
1251
1252 if (unlikely(!deepest_stop_found)) {
1253 pr_warn("cpuidle-powernv: No suitable stop state for CPU-Hotplug. Offlined CPUs will busy wait");
1254 } else {
1255 pr_info("cpuidle-powernv: Deepest stop: psscr = 0x%016llx,mask=0x%016llx\n",
1256 pnv_deepest_stop_psscr_val,
1257 pnv_deepest_stop_psscr_mask);
1258 }
1259
1260 pr_info("cpuidle-powernv: First stop level that may lose SPRs = 0x%llx\n",
1261 deep_spr_loss_state);
1262
1263 pr_info("cpuidle-powernv: First stop level that may lose timebase = 0x%llx\n",
1264 pnv_first_tb_loss_level);
1265 }
1266
pnv_disable_deep_states(void)1267 static void __init pnv_disable_deep_states(void)
1268 {
1269 /*
1270 * The stop-api is unable to restore hypervisor
1271 * resources on wakeup from platform idle states which
1272 * lose full context. So disable such states.
1273 */
1274 supported_cpuidle_states &= ~OPAL_PM_LOSE_FULL_CONTEXT;
1275 pr_warn("cpuidle-powernv: Disabling idle states that lose full context\n");
1276 pr_warn("cpuidle-powernv: Idle power-savings, CPU-Hotplug affected\n");
1277
1278 if (cpu_has_feature(CPU_FTR_ARCH_300) &&
1279 (pnv_deepest_stop_flag & OPAL_PM_LOSE_FULL_CONTEXT)) {
1280 /*
1281 * Use the default stop state for CPU-Hotplug
1282 * if available.
1283 */
1284 if (default_stop_found) {
1285 pnv_deepest_stop_psscr_val = pnv_default_stop_val;
1286 pnv_deepest_stop_psscr_mask = pnv_default_stop_mask;
1287 pr_warn("cpuidle-powernv: Offlined CPUs will stop with psscr = 0x%016llx\n",
1288 pnv_deepest_stop_psscr_val);
1289 } else { /* Fallback to snooze loop for CPU-Hotplug */
1290 deepest_stop_found = false;
1291 pr_warn("cpuidle-powernv: Offlined CPUs will busy wait\n");
1292 }
1293 }
1294 }
1295
1296 /*
1297 * Probe device tree for supported idle states
1298 */
pnv_probe_idle_states(void)1299 static void __init pnv_probe_idle_states(void)
1300 {
1301 int i;
1302
1303 if (nr_pnv_idle_states < 0) {
1304 pr_warn("cpuidle-powernv: no idle states found in the DT\n");
1305 return;
1306 }
1307
1308 if (cpu_has_feature(CPU_FTR_ARCH_300))
1309 pnv_arch300_idle_init();
1310
1311 for (i = 0; i < nr_pnv_idle_states; i++)
1312 supported_cpuidle_states |= pnv_idle_states[i].flags;
1313 }
1314
1315 /*
1316 * This function parses device-tree and populates all the information
1317 * into pnv_idle_states structure. It also sets up nr_pnv_idle_states
1318 * which is the number of cpuidle states discovered through device-tree.
1319 */
1320
pnv_parse_cpuidle_dt(void)1321 static int pnv_parse_cpuidle_dt(void)
1322 {
1323 struct device_node *np;
1324 int nr_idle_states, i;
1325 int rc = 0;
1326 u32 *temp_u32;
1327 u64 *temp_u64;
1328 const char **temp_string;
1329
1330 np = of_find_node_by_path("/ibm,opal/power-mgt");
1331 if (!np) {
1332 pr_warn("opal: PowerMgmt Node not found\n");
1333 return -ENODEV;
1334 }
1335 nr_idle_states = of_property_count_u32_elems(np,
1336 "ibm,cpu-idle-state-flags");
1337
1338 pnv_idle_states = kcalloc(nr_idle_states, sizeof(*pnv_idle_states),
1339 GFP_KERNEL);
1340 temp_u32 = kcalloc(nr_idle_states, sizeof(u32), GFP_KERNEL);
1341 temp_u64 = kcalloc(nr_idle_states, sizeof(u64), GFP_KERNEL);
1342 temp_string = kcalloc(nr_idle_states, sizeof(char *), GFP_KERNEL);
1343
1344 if (!(pnv_idle_states && temp_u32 && temp_u64 && temp_string)) {
1345 pr_err("Could not allocate memory for dt parsing\n");
1346 rc = -ENOMEM;
1347 goto out;
1348 }
1349
1350 /* Read flags */
1351 if (of_property_read_u32_array(np, "ibm,cpu-idle-state-flags",
1352 temp_u32, nr_idle_states)) {
1353 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-flags in DT\n");
1354 rc = -EINVAL;
1355 goto out;
1356 }
1357 for (i = 0; i < nr_idle_states; i++)
1358 pnv_idle_states[i].flags = temp_u32[i];
1359
1360 /* Read latencies */
1361 if (of_property_read_u32_array(np, "ibm,cpu-idle-state-latencies-ns",
1362 temp_u32, nr_idle_states)) {
1363 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-latencies-ns in DT\n");
1364 rc = -EINVAL;
1365 goto out;
1366 }
1367 for (i = 0; i < nr_idle_states; i++)
1368 pnv_idle_states[i].latency_ns = temp_u32[i];
1369
1370 /* Read residencies */
1371 if (of_property_read_u32_array(np, "ibm,cpu-idle-state-residency-ns",
1372 temp_u32, nr_idle_states)) {
1373 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-residency-ns in DT\n");
1374 rc = -EINVAL;
1375 goto out;
1376 }
1377 for (i = 0; i < nr_idle_states; i++)
1378 pnv_idle_states[i].residency_ns = temp_u32[i];
1379
1380 /* For power9 and later */
1381 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1382 /* Read pm_crtl_val */
1383 if (of_property_read_u64_array(np, "ibm,cpu-idle-state-psscr",
1384 temp_u64, nr_idle_states)) {
1385 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-psscr in DT\n");
1386 rc = -EINVAL;
1387 goto out;
1388 }
1389 for (i = 0; i < nr_idle_states; i++)
1390 pnv_idle_states[i].psscr_val = temp_u64[i];
1391
1392 /* Read pm_crtl_mask */
1393 if (of_property_read_u64_array(np, "ibm,cpu-idle-state-psscr-mask",
1394 temp_u64, nr_idle_states)) {
1395 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-psscr-mask in DT\n");
1396 rc = -EINVAL;
1397 goto out;
1398 }
1399 for (i = 0; i < nr_idle_states; i++)
1400 pnv_idle_states[i].psscr_mask = temp_u64[i];
1401 }
1402
1403 /*
1404 * power8 specific properties ibm,cpu-idle-state-pmicr-mask and
1405 * ibm,cpu-idle-state-pmicr-val were never used and there is no
1406 * plan to use it in near future. Hence, not parsing these properties
1407 */
1408
1409 if (of_property_read_string_array(np, "ibm,cpu-idle-state-names",
1410 temp_string, nr_idle_states) < 0) {
1411 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-names in DT\n");
1412 rc = -EINVAL;
1413 goto out;
1414 }
1415 for (i = 0; i < nr_idle_states; i++)
1416 strlcpy(pnv_idle_states[i].name, temp_string[i],
1417 PNV_IDLE_NAME_LEN);
1418 nr_pnv_idle_states = nr_idle_states;
1419 rc = 0;
1420 out:
1421 kfree(temp_u32);
1422 kfree(temp_u64);
1423 kfree(temp_string);
1424 return rc;
1425 }
1426
pnv_init_idle_states(void)1427 static int __init pnv_init_idle_states(void)
1428 {
1429 int cpu;
1430 int rc = 0;
1431
1432 /* Set up PACA fields */
1433 for_each_present_cpu(cpu) {
1434 struct paca_struct *p = paca_ptrs[cpu];
1435
1436 p->idle_state = 0;
1437 if (cpu == cpu_first_thread_sibling(cpu))
1438 p->idle_state = (1 << threads_per_core) - 1;
1439
1440 if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
1441 /* P7/P8 nap */
1442 p->thread_idle_state = PNV_THREAD_RUNNING;
1443 } else if (pvr_version_is(PVR_POWER9)) {
1444 /* P9 stop workarounds */
1445 #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
1446 p->requested_psscr = 0;
1447 atomic_set(&p->dont_stop, 0);
1448 #endif
1449 }
1450 }
1451
1452 /* In case we error out nr_pnv_idle_states will be zero */
1453 nr_pnv_idle_states = 0;
1454 supported_cpuidle_states = 0;
1455
1456 if (cpuidle_disable != IDLE_NO_OVERRIDE)
1457 goto out;
1458 rc = pnv_parse_cpuidle_dt();
1459 if (rc)
1460 return rc;
1461 pnv_probe_idle_states();
1462
1463 if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
1464 if (!(supported_cpuidle_states & OPAL_PM_SLEEP_ENABLED_ER1)) {
1465 power7_fastsleep_workaround_entry = false;
1466 power7_fastsleep_workaround_exit = false;
1467 } else {
1468 /*
1469 * OPAL_PM_SLEEP_ENABLED_ER1 is set. It indicates that
1470 * workaround is needed to use fastsleep. Provide sysfs
1471 * control to choose how this workaround has to be
1472 * applied.
1473 */
1474 device_create_file(cpu_subsys.dev_root,
1475 &dev_attr_fastsleep_workaround_applyonce);
1476 }
1477
1478 update_subcore_sibling_mask();
1479
1480 if (supported_cpuidle_states & OPAL_PM_NAP_ENABLED) {
1481 ppc_md.power_save = power7_idle;
1482 power7_offline_type = PNV_THREAD_NAP;
1483 }
1484
1485 if ((supported_cpuidle_states & OPAL_PM_WINKLE_ENABLED) &&
1486 (supported_cpuidle_states & OPAL_PM_LOSE_FULL_CONTEXT))
1487 power7_offline_type = PNV_THREAD_WINKLE;
1488 else if ((supported_cpuidle_states & OPAL_PM_SLEEP_ENABLED) ||
1489 (supported_cpuidle_states & OPAL_PM_SLEEP_ENABLED_ER1))
1490 power7_offline_type = PNV_THREAD_SLEEP;
1491 }
1492
1493 if (supported_cpuidle_states & OPAL_PM_LOSE_FULL_CONTEXT) {
1494 if (pnv_save_sprs_for_deep_states())
1495 pnv_disable_deep_states();
1496 }
1497
1498 out:
1499 return 0;
1500 }
1501 machine_subsys_initcall(powernv, pnv_init_idle_states);
1502