xref: /linux/arch/powerpc/kernel/smp.c (revision f4cdf7ca9a1fdcca413157df19753f388a5a224e)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * SMP support for ppc.
4  *
5  * Written by Cort Dougan (cort@cs.nmt.edu) borrowing a great
6  * deal of code from the sparc and intel versions.
7  *
8  * Copyright (C) 1999 Cort Dougan <cort@cs.nmt.edu>
9  *
10  * PowerPC-64 Support added by Dave Engebretsen, Peter Bergner, and
11  * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
12  */
13 
14 #undef DEBUG
15 
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/sched/mm.h>
19 #include <linux/sched/task_stack.h>
20 #include <linux/sched/topology.h>
21 #include <linux/smp.h>
22 #include <linux/interrupt.h>
23 #include <linux/delay.h>
24 #include <linux/init.h>
25 #include <linux/spinlock.h>
26 #include <linux/cache.h>
27 #include <linux/err.h>
28 #include <linux/device.h>
29 #include <linux/cpu.h>
30 #include <linux/notifier.h>
31 #include <linux/topology.h>
32 #include <linux/profile.h>
33 #include <linux/processor.h>
34 #include <linux/random.h>
35 #include <linux/stackprotector.h>
36 #include <linux/pgtable.h>
37 #include <linux/clockchips.h>
38 #include <linux/kexec.h>
39 
40 #include <asm/ptrace.h>
41 #include <linux/atomic.h>
42 #include <asm/irq.h>
43 #include <asm/hw_irq.h>
44 #include <asm/kvm_ppc.h>
45 #include <asm/dbell.h>
46 #include <asm/page.h>
47 #include <asm/smp.h>
48 #include <asm/time.h>
49 #include <asm/machdep.h>
50 #include <asm/mmu_context.h>
51 #include <asm/cputhreads.h>
52 #include <asm/cputable.h>
53 #include <asm/mpic.h>
54 #include <asm/vdso_datapage.h>
55 #ifdef CONFIG_PPC64
56 #include <asm/paca.h>
57 #endif
58 #include <asm/vdso.h>
59 #include <asm/debug.h>
60 #include <asm/cpu_has_feature.h>
61 #include <asm/ftrace.h>
62 #include <asm/kup.h>
63 #include <asm/fadump.h>
64 #include <asm/systemcfg.h>
65 
66 #include <trace/events/ipi.h>
67 
68 #ifdef DEBUG
69 #include <asm/udbg.h>
70 #define DBG(fmt...) udbg_printf(fmt)
71 #else
72 #define DBG(fmt...)
73 #endif
74 
75 #ifdef CONFIG_HOTPLUG_CPU
76 /* State of each CPU during hotplug phases */
77 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
78 #endif
79 
80 struct task_struct *secondary_current;
81 bool has_big_cores __ro_after_init;
82 bool coregroup_enabled __ro_after_init;
83 bool thread_group_shares_l2 __ro_after_init;
84 bool thread_group_shares_l3 __ro_after_init;
85 
86 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
87 DEFINE_PER_CPU(cpumask_var_t, cpu_smallcore_map);
88 DEFINE_PER_CPU(cpumask_var_t, cpu_l2_cache_map);
89 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
90 static DEFINE_PER_CPU(cpumask_var_t, cpu_coregroup_map);
91 
92 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
93 EXPORT_PER_CPU_SYMBOL(cpu_l2_cache_map);
94 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
95 EXPORT_SYMBOL_GPL(has_big_cores);
96 
97 #define MAX_THREAD_LIST_SIZE	8
98 #define THREAD_GROUP_SHARE_L1   1
99 #define THREAD_GROUP_SHARE_L2_L3 2
100 struct thread_groups {
101 	unsigned int property;
102 	unsigned int nr_groups;
103 	unsigned int threads_per_group;
104 	unsigned int thread_list[MAX_THREAD_LIST_SIZE];
105 };
106 
107 /* Maximum number of properties that groups of threads within a core can share */
108 #define MAX_THREAD_GROUP_PROPERTIES 2
109 
110 struct thread_groups_list {
111 	unsigned int nr_properties;
112 	struct thread_groups property_tgs[MAX_THREAD_GROUP_PROPERTIES];
113 };
114 
115 static struct thread_groups_list tgl[NR_CPUS] __initdata;
116 /*
117  * On big-cores system, thread_group_l1_cache_map for each CPU corresponds to
118  * the set its siblings that share the L1-cache.
119  */
120 DEFINE_PER_CPU(cpumask_var_t, thread_group_l1_cache_map);
121 
122 /*
123  * On some big-cores system, thread_group_l2_cache_map for each CPU
124  * corresponds to the set its siblings within the core that share the
125  * L2-cache.
126  */
127 DEFINE_PER_CPU(cpumask_var_t, thread_group_l2_cache_map);
128 
129 /*
130  * On P10, thread_group_l3_cache_map for each CPU is equal to the
131  * thread_group_l2_cache_map
132  */
133 DEFINE_PER_CPU(cpumask_var_t, thread_group_l3_cache_map);
134 
135 /* SMP operations for this machine */
136 struct smp_ops_t *smp_ops;
137 
138 /* Can't be static due to PowerMac hackery */
139 volatile unsigned int cpu_callin_map[NR_CPUS];
140 
141 int smt_enabled_at_boot = 1;
142 
143 /*
144  * Returns 1 if the specified cpu should be brought up during boot.
145  * Used to inhibit booting threads if they've been disabled or
146  * limited on the command line
147  */
148 int smp_generic_cpu_bootable(unsigned int nr)
149 {
150 	/* Special case - we inhibit secondary thread startup
151 	 * during boot if the user requests it.
152 	 */
153 	if (system_state < SYSTEM_RUNNING && cpu_has_feature(CPU_FTR_SMT)) {
154 		if (!smt_enabled_at_boot && cpu_thread_in_core(nr) != 0)
155 			return 0;
156 		if (smt_enabled_at_boot
157 		    && cpu_thread_in_core(nr) >= smt_enabled_at_boot)
158 			return 0;
159 	}
160 
161 	return 1;
162 }
163 
164 
165 #ifdef CONFIG_PPC64
166 int smp_generic_kick_cpu(int nr)
167 {
168 	if (nr < 0 || nr >= nr_cpu_ids)
169 		return -EINVAL;
170 
171 	/*
172 	 * The processor is currently spinning, waiting for the
173 	 * cpu_start field to become non-zero After we set cpu_start,
174 	 * the processor will continue on to secondary_start
175 	 */
176 	if (!paca_ptrs[nr]->cpu_start) {
177 		paca_ptrs[nr]->cpu_start = 1;
178 		smp_mb();
179 		return 0;
180 	}
181 
182 #ifdef CONFIG_HOTPLUG_CPU
183 	/*
184 	 * Ok it's not there, so it might be soft-unplugged, let's
185 	 * try to bring it back
186 	 */
187 	generic_set_cpu_up(nr);
188 	smp_wmb();
189 	smp_send_reschedule(nr);
190 #endif /* CONFIG_HOTPLUG_CPU */
191 
192 	return 0;
193 }
194 #endif /* CONFIG_PPC64 */
195 
196 static irqreturn_t call_function_action(int irq, void *data)
197 {
198 	generic_smp_call_function_interrupt();
199 	return IRQ_HANDLED;
200 }
201 
202 static irqreturn_t reschedule_action(int irq, void *data)
203 {
204 	scheduler_ipi();
205 	return IRQ_HANDLED;
206 }
207 
208 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
209 static irqreturn_t tick_broadcast_ipi_action(int irq, void *data)
210 {
211 	timer_broadcast_interrupt();
212 	return IRQ_HANDLED;
213 }
214 #endif
215 
216 #ifdef CONFIG_NMI_IPI
217 static irqreturn_t nmi_ipi_action(int irq, void *data)
218 {
219 	smp_handle_nmi_ipi(get_irq_regs());
220 	return IRQ_HANDLED;
221 }
222 #endif
223 
224 static irq_handler_t smp_ipi_action[] = {
225 	[PPC_MSG_CALL_FUNCTION] =  call_function_action,
226 	[PPC_MSG_RESCHEDULE] = reschedule_action,
227 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
228 	[PPC_MSG_TICK_BROADCAST] = tick_broadcast_ipi_action,
229 #endif
230 #ifdef CONFIG_NMI_IPI
231 	[PPC_MSG_NMI_IPI] = nmi_ipi_action,
232 #endif
233 };
234 
235 /*
236  * The NMI IPI is a fallback and not truly non-maskable. It is simpler
237  * than going through the call function infrastructure, and strongly
238  * serialized, so it is more appropriate for debugging.
239  */
240 const char *smp_ipi_name[] = {
241 	[PPC_MSG_CALL_FUNCTION] =  "ipi call function",
242 	[PPC_MSG_RESCHEDULE] = "ipi reschedule",
243 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
244 	[PPC_MSG_TICK_BROADCAST] = "ipi tick-broadcast",
245 #endif
246 #ifdef CONFIG_NMI_IPI
247 	[PPC_MSG_NMI_IPI] = "nmi ipi",
248 #endif
249 };
250 
251 /* optional function to request ipi, for controllers with >= 4 ipis */
252 int smp_request_message_ipi(int virq, int msg)
253 {
254 	int err;
255 
256 	if (msg < 0 || msg > PPC_MSG_NMI_IPI)
257 		return -EINVAL;
258 #ifndef CONFIG_NMI_IPI
259 	if (msg == PPC_MSG_NMI_IPI)
260 		return 1;
261 #endif
262 
263 	err = request_irq(virq, smp_ipi_action[msg],
264 			  IRQF_PERCPU | IRQF_NO_THREAD | IRQF_NO_SUSPEND,
265 			  smp_ipi_name[msg], NULL);
266 	WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
267 		virq, smp_ipi_name[msg], err);
268 
269 	return err;
270 }
271 
272 #ifdef CONFIG_PPC_SMP_MUXED_IPI
273 struct cpu_messages {
274 	long messages;			/* current messages */
275 };
276 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
277 
278 void smp_muxed_ipi_set_message(int cpu, int msg)
279 {
280 	struct cpu_messages *info = &per_cpu(ipi_message, cpu);
281 	char *message = (char *)&info->messages;
282 
283 	/*
284 	 * Order previous accesses before accesses in the IPI handler.
285 	 */
286 	smp_mb();
287 	WRITE_ONCE(message[msg], 1);
288 }
289 
290 void smp_muxed_ipi_message_pass(int cpu, int msg)
291 {
292 	if (!smp_ops->cause_ipi)
293 		return;
294 
295 	smp_muxed_ipi_set_message(cpu, msg);
296 
297 	/*
298 	 * cause_ipi functions are required to include a full barrier
299 	 * before doing whatever causes the IPI.
300 	 */
301 	smp_ops->cause_ipi(cpu);
302 }
303 
304 #ifdef __BIG_ENDIAN__
305 #define IPI_MESSAGE(A) (1uL << ((BITS_PER_LONG - 8) - 8 * (A)))
306 #else
307 #define IPI_MESSAGE(A) (1uL << (8 * (A)))
308 #endif
309 
310 irqreturn_t smp_ipi_demux(void)
311 {
312 	mb();	/* order any irq clear */
313 
314 	return smp_ipi_demux_relaxed();
315 }
316 
317 /* sync-free variant. Callers should ensure synchronization */
318 irqreturn_t smp_ipi_demux_relaxed(void)
319 {
320 	struct cpu_messages *info;
321 	unsigned long all;
322 
323 	info = this_cpu_ptr(&ipi_message);
324 	do {
325 		all = xchg(&info->messages, 0);
326 #if defined(CONFIG_KVM_XICS) && defined(CONFIG_KVM_BOOK3S_HV_POSSIBLE)
327 		/*
328 		 * Must check for PPC_MSG_RM_HOST_ACTION messages
329 		 * before PPC_MSG_CALL_FUNCTION messages because when
330 		 * a VM is destroyed, we call kick_all_cpus_sync()
331 		 * to ensure that any pending PPC_MSG_RM_HOST_ACTION
332 		 * messages have completed before we free any VCPUs.
333 		 */
334 		if (all & IPI_MESSAGE(PPC_MSG_RM_HOST_ACTION))
335 			kvmppc_xics_ipi_action();
336 #endif
337 		if (all & IPI_MESSAGE(PPC_MSG_CALL_FUNCTION))
338 			generic_smp_call_function_interrupt();
339 		if (all & IPI_MESSAGE(PPC_MSG_RESCHEDULE))
340 			scheduler_ipi();
341 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
342 		if (all & IPI_MESSAGE(PPC_MSG_TICK_BROADCAST))
343 			timer_broadcast_interrupt();
344 #endif
345 #ifdef CONFIG_NMI_IPI
346 		if (all & IPI_MESSAGE(PPC_MSG_NMI_IPI))
347 			nmi_ipi_action(0, NULL);
348 #endif
349 	} while (READ_ONCE(info->messages));
350 
351 	return IRQ_HANDLED;
352 }
353 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
354 
355 static inline void do_message_pass(int cpu, int msg)
356 {
357 	if (smp_ops->message_pass)
358 		smp_ops->message_pass(cpu, msg);
359 #ifdef CONFIG_PPC_SMP_MUXED_IPI
360 	else
361 		smp_muxed_ipi_message_pass(cpu, msg);
362 #endif
363 }
364 
365 void arch_smp_send_reschedule(int cpu)
366 {
367 	if (likely(smp_ops))
368 		do_message_pass(cpu, PPC_MSG_RESCHEDULE);
369 }
370 EXPORT_SYMBOL_GPL(arch_smp_send_reschedule);
371 
372 void arch_send_call_function_single_ipi(int cpu)
373 {
374 	do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
375 }
376 
377 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
378 {
379 	unsigned int cpu;
380 
381 	for_each_cpu(cpu, mask)
382 		do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
383 }
384 
385 #ifdef CONFIG_NMI_IPI
386 
387 /*
388  * "NMI IPI" system.
389  *
390  * NMI IPIs may not be recoverable, so should not be used as ongoing part of
391  * a running system. They can be used for crash, debug, halt/reboot, etc.
392  *
393  * The IPI call waits with interrupts disabled until all targets enter the
394  * NMI handler, then returns. Subsequent IPIs can be issued before targets
395  * have returned from their handlers, so there is no guarantee about
396  * concurrency or re-entrancy.
397  *
398  * A new NMI can be issued before all targets exit the handler.
399  *
400  * The IPI call may time out without all targets entering the NMI handler.
401  * In that case, there is some logic to recover (and ignore subsequent
402  * NMI interrupts that may eventually be raised), but the platform interrupt
403  * handler may not be able to distinguish this from other exception causes,
404  * which may cause a crash.
405  */
406 
407 static atomic_t __nmi_ipi_lock = ATOMIC_INIT(0);
408 static struct cpumask nmi_ipi_pending_mask;
409 static bool nmi_ipi_busy = false;
410 static void (*nmi_ipi_function)(struct pt_regs *) = NULL;
411 
412 noinstr static void nmi_ipi_lock_start(unsigned long *flags)
413 {
414 	raw_local_irq_save(*flags);
415 	hard_irq_disable();
416 	while (raw_atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1) {
417 		raw_local_irq_restore(*flags);
418 		spin_until_cond(raw_atomic_read(&__nmi_ipi_lock) == 0);
419 		raw_local_irq_save(*flags);
420 		hard_irq_disable();
421 	}
422 }
423 
424 noinstr static void nmi_ipi_lock(void)
425 {
426 	while (raw_atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1)
427 		spin_until_cond(raw_atomic_read(&__nmi_ipi_lock) == 0);
428 }
429 
430 noinstr static void nmi_ipi_unlock(void)
431 {
432 	smp_mb();
433 	WARN_ON(raw_atomic_read(&__nmi_ipi_lock) != 1);
434 	raw_atomic_set(&__nmi_ipi_lock, 0);
435 }
436 
437 noinstr static void nmi_ipi_unlock_end(unsigned long *flags)
438 {
439 	nmi_ipi_unlock();
440 	raw_local_irq_restore(*flags);
441 }
442 
443 /*
444  * Platform NMI handler calls this to ack
445  */
446 noinstr int smp_handle_nmi_ipi(struct pt_regs *regs)
447 {
448 	void (*fn)(struct pt_regs *) = NULL;
449 	unsigned long flags;
450 	int me = raw_smp_processor_id();
451 	int ret = 0;
452 
453 	/*
454 	 * Unexpected NMIs are possible here because the interrupt may not
455 	 * be able to distinguish NMI IPIs from other types of NMIs, or
456 	 * because the caller may have timed out.
457 	 */
458 	nmi_ipi_lock_start(&flags);
459 	if (cpumask_test_cpu(me, &nmi_ipi_pending_mask)) {
460 		cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
461 		fn = READ_ONCE(nmi_ipi_function);
462 		WARN_ON_ONCE(!fn);
463 		ret = 1;
464 	}
465 	nmi_ipi_unlock_end(&flags);
466 
467 	if (fn)
468 		fn(regs);
469 
470 	return ret;
471 }
472 
473 static void do_smp_send_nmi_ipi(int cpu, bool safe)
474 {
475 	if (!safe && smp_ops->cause_nmi_ipi && smp_ops->cause_nmi_ipi(cpu))
476 		return;
477 
478 	if (cpu >= 0) {
479 		do_message_pass(cpu, PPC_MSG_NMI_IPI);
480 	} else {
481 		int c;
482 
483 		for_each_online_cpu(c) {
484 			if (c == raw_smp_processor_id())
485 				continue;
486 			do_message_pass(c, PPC_MSG_NMI_IPI);
487 		}
488 	}
489 }
490 
491 /*
492  * - cpu is the target CPU (must not be this CPU), or NMI_IPI_ALL_OTHERS.
493  * - fn is the target callback function.
494  * - delay_us > 0 is the delay before giving up waiting for targets to
495  *   begin executing the handler, == 0 specifies indefinite delay.
496  */
497 static int __smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *),
498 				u64 delay_us, bool safe)
499 {
500 	unsigned long flags;
501 	int me = raw_smp_processor_id();
502 	int ret = 1;
503 
504 	BUG_ON(cpu == me);
505 	BUG_ON(cpu < 0 && cpu != NMI_IPI_ALL_OTHERS);
506 
507 	if (unlikely(!smp_ops))
508 		return 0;
509 
510 	nmi_ipi_lock_start(&flags);
511 	while (nmi_ipi_busy) {
512 		nmi_ipi_unlock_end(&flags);
513 		spin_until_cond(!nmi_ipi_busy);
514 		nmi_ipi_lock_start(&flags);
515 	}
516 	nmi_ipi_busy = true;
517 	nmi_ipi_function = fn;
518 
519 	WARN_ON_ONCE(!cpumask_empty(&nmi_ipi_pending_mask));
520 
521 	if (cpu < 0) {
522 		/* ALL_OTHERS */
523 		cpumask_copy(&nmi_ipi_pending_mask, cpu_online_mask);
524 		cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
525 	} else {
526 		cpumask_set_cpu(cpu, &nmi_ipi_pending_mask);
527 	}
528 
529 	nmi_ipi_unlock();
530 
531 	/* Interrupts remain hard disabled */
532 
533 	do_smp_send_nmi_ipi(cpu, safe);
534 
535 	nmi_ipi_lock();
536 	/* nmi_ipi_busy is set here, so unlock/lock is okay */
537 	while (!cpumask_empty(&nmi_ipi_pending_mask)) {
538 		nmi_ipi_unlock();
539 		udelay(1);
540 		nmi_ipi_lock();
541 		if (delay_us) {
542 			delay_us--;
543 			if (!delay_us)
544 				break;
545 		}
546 	}
547 
548 	if (!cpumask_empty(&nmi_ipi_pending_mask)) {
549 		/* Timeout waiting for CPUs to call smp_handle_nmi_ipi */
550 		ret = 0;
551 		cpumask_clear(&nmi_ipi_pending_mask);
552 	}
553 
554 	nmi_ipi_function = NULL;
555 	nmi_ipi_busy = false;
556 
557 	nmi_ipi_unlock_end(&flags);
558 
559 	return ret;
560 }
561 
562 int smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
563 {
564 	return __smp_send_nmi_ipi(cpu, fn, delay_us, false);
565 }
566 
567 int smp_send_safe_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
568 {
569 	return __smp_send_nmi_ipi(cpu, fn, delay_us, true);
570 }
571 #endif /* CONFIG_NMI_IPI */
572 
573 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
574 void tick_broadcast(const struct cpumask *mask)
575 {
576 	unsigned int cpu;
577 
578 	for_each_cpu(cpu, mask)
579 		do_message_pass(cpu, PPC_MSG_TICK_BROADCAST);
580 }
581 #endif
582 
583 #ifdef CONFIG_DEBUGGER
584 static void debugger_ipi_callback(struct pt_regs *regs)
585 {
586 	debugger_ipi(regs);
587 }
588 
589 void smp_send_debugger_break(void)
590 {
591 	smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, debugger_ipi_callback, 1000000);
592 }
593 #endif
594 
595 #ifdef CONFIG_CRASH_DUMP
596 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
597 {
598 	int cpu;
599 
600 	smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, crash_ipi_callback, 1000000);
601 	if (kdump_in_progress() && crash_wake_offline) {
602 		for_each_present_cpu(cpu) {
603 			if (cpu_online(cpu))
604 				continue;
605 			/*
606 			 * crash_ipi_callback will wait for
607 			 * all cpus, including offline CPUs.
608 			 * We don't care about nmi_ipi_function.
609 			 * Offline cpus will jump straight into
610 			 * crash_ipi_callback, we can skip the
611 			 * entire NMI dance and waiting for
612 			 * cpus to clear pending mask, etc.
613 			 */
614 			do_smp_send_nmi_ipi(cpu, false);
615 		}
616 	}
617 }
618 #endif
619 
620 void crash_smp_send_stop(void)
621 {
622 	static bool stopped = false;
623 
624 	/*
625 	 * In case of fadump, register data for all CPUs is captured by f/w
626 	 * on ibm,os-term rtas call. Skip IPI callbacks to other CPUs before
627 	 * this rtas call to avoid tricky post processing of those CPUs'
628 	 * backtraces.
629 	 */
630 	if (should_fadump_crash())
631 		return;
632 
633 	if (stopped)
634 		return;
635 
636 	stopped = true;
637 
638 #ifdef CONFIG_CRASH_DUMP
639 	if (kexec_crash_image) {
640 		crash_kexec_prepare();
641 		return;
642 	}
643 #endif
644 
645 	smp_send_stop();
646 }
647 
648 #ifdef CONFIG_NMI_IPI
649 static void nmi_stop_this_cpu(struct pt_regs *regs)
650 {
651 	/*
652 	 * IRQs are already hard disabled by the smp_handle_nmi_ipi.
653 	 */
654 	set_cpu_online(smp_processor_id(), false);
655 
656 	spin_begin();
657 	while (1)
658 		spin_cpu_relax();
659 }
660 
661 void smp_send_stop(void)
662 {
663 	smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, nmi_stop_this_cpu, 1000000);
664 }
665 
666 #else /* CONFIG_NMI_IPI */
667 
668 static void stop_this_cpu(void *dummy)
669 {
670 	hard_irq_disable();
671 
672 	/*
673 	 * Offlining CPUs in stop_this_cpu can result in scheduler warnings,
674 	 * (see commit de6e5d38417e), but printk_safe_flush_on_panic() wants
675 	 * to know other CPUs are offline before it breaks locks to flush
676 	 * printk buffers, in case we panic()ed while holding the lock.
677 	 */
678 	set_cpu_online(smp_processor_id(), false);
679 
680 	spin_begin();
681 	while (1)
682 		spin_cpu_relax();
683 }
684 
685 void smp_send_stop(void)
686 {
687 	static bool stopped = false;
688 
689 	/*
690 	 * Prevent waiting on csd lock from a previous smp_send_stop.
691 	 * This is racy, but in general callers try to do the right
692 	 * thing and only fire off one smp_send_stop (e.g., see
693 	 * kernel/panic.c)
694 	 */
695 	if (stopped)
696 		return;
697 
698 	stopped = true;
699 
700 	smp_call_function(stop_this_cpu, NULL, 0);
701 }
702 #endif /* CONFIG_NMI_IPI */
703 
704 static struct task_struct *current_set[NR_CPUS];
705 
706 static void smp_store_cpu_info(int id)
707 {
708 	per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
709 #ifdef CONFIG_PPC_E500
710 	per_cpu(next_tlbcam_idx, id)
711 		= (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
712 #endif
713 }
714 
715 /*
716  * Relationships between CPUs are maintained in a set of per-cpu cpumasks so
717  * rather than just passing around the cpumask we pass around a function that
718  * returns the that cpumask for the given CPU.
719  */
720 static void set_cpus_related(int i, int j, struct cpumask *(*get_cpumask)(int))
721 {
722 	cpumask_set_cpu(i, get_cpumask(j));
723 	cpumask_set_cpu(j, get_cpumask(i));
724 }
725 
726 #ifdef CONFIG_HOTPLUG_CPU
727 static void set_cpus_unrelated(int i, int j,
728 		struct cpumask *(*get_cpumask)(int))
729 {
730 	cpumask_clear_cpu(i, get_cpumask(j));
731 	cpumask_clear_cpu(j, get_cpumask(i));
732 }
733 #endif
734 
735 /*
736  * Extends set_cpus_related. Instead of setting one CPU at a time in
737  * dstmask, set srcmask at oneshot. dstmask should be super set of srcmask.
738  */
739 static void or_cpumasks_related(int i, int j, struct cpumask *(*srcmask)(int),
740 				struct cpumask *(*dstmask)(int))
741 {
742 	struct cpumask *mask;
743 	int k;
744 
745 	mask = srcmask(j);
746 	for_each_cpu(k, srcmask(i))
747 		cpumask_or(dstmask(k), dstmask(k), mask);
748 
749 	if (i == j)
750 		return;
751 
752 	mask = srcmask(i);
753 	for_each_cpu(k, srcmask(j))
754 		cpumask_or(dstmask(k), dstmask(k), mask);
755 }
756 
757 /*
758  * parse_thread_groups: Parses the "ibm,thread-groups" device tree
759  *                      property for the CPU device node @dn and stores
760  *                      the parsed output in the thread_groups_list
761  *                      structure @tglp.
762  *
763  * @dn: The device node of the CPU device.
764  * @tglp: Pointer to a thread group list structure into which the parsed
765  *      output of "ibm,thread-groups" is stored.
766  *
767  * ibm,thread-groups[0..N-1] array defines which group of threads in
768  * the CPU-device node can be grouped together based on the property.
769  *
770  * This array can represent thread groupings for multiple properties.
771  *
772  * ibm,thread-groups[i + 0] tells us the property based on which the
773  * threads are being grouped together. If this value is 1, it implies
774  * that the threads in the same group share L1, translation cache. If
775  * the value is 2, it implies that the threads in the same group share
776  * the same L2 cache.
777  *
778  * ibm,thread-groups[i+1] tells us how many such thread groups exist for the
779  * property ibm,thread-groups[i]
780  *
781  * ibm,thread-groups[i+2] tells us the number of threads in each such
782  * group.
783  * Suppose k = (ibm,thread-groups[i+1] * ibm,thread-groups[i+2]), then,
784  *
785  * ibm,thread-groups[i+3..i+k+2] (is the list of threads identified by
786  * "ibm,ppc-interrupt-server#s" arranged as per their membership in
787  * the grouping.
788  *
789  * Example:
790  * If "ibm,thread-groups" = [1,2,4,8,10,12,14,9,11,13,15,2,2,4,8,10,12,14,9,11,13,15]
791  * This can be decomposed up into two consecutive arrays:
792  * a) [1,2,4,8,10,12,14,9,11,13,15]
793  * b) [2,2,4,8,10,12,14,9,11,13,15]
794  *
795  * where in,
796  *
797  * a) provides information of Property "1" being shared by "2" groups,
798  *  each with "4" threads each. The "ibm,ppc-interrupt-server#s" of
799  *  the first group is {8,10,12,14} and the
800  *  "ibm,ppc-interrupt-server#s" of the second group is
801  *  {9,11,13,15}. Property "1" is indicative of the thread in the
802  *  group sharing L1 cache, translation cache and Instruction Data
803  *  flow.
804  *
805  * b) provides information of Property "2" being shared by "2" groups,
806  *  each group with "4" threads. The "ibm,ppc-interrupt-server#s" of
807  *  the first group is {8,10,12,14} and the
808  *  "ibm,ppc-interrupt-server#s" of the second group is
809  *  {9,11,13,15}. Property "2" indicates that the threads in each
810  *  group share the L2-cache.
811  *
812  * Returns 0 on success, -EINVAL if the property does not exist,
813  * -ENODATA if property does not have a value, and -EOVERFLOW if the
814  * property data isn't large enough.
815  */
816 static int parse_thread_groups(struct device_node *dn,
817 			       struct thread_groups_list *tglp)
818 {
819 	unsigned int property_idx = 0;
820 	u32 *thread_group_array;
821 	size_t total_threads;
822 	int ret = 0, count;
823 	u32 *thread_list;
824 	int i = 0;
825 
826 	count = of_property_count_u32_elems(dn, "ibm,thread-groups");
827 	thread_group_array = kcalloc(count, sizeof(u32), GFP_KERNEL);
828 	if (!thread_group_array)
829 		return -ENOMEM;
830 	ret = of_property_read_u32_array(dn, "ibm,thread-groups",
831 					 thread_group_array, count);
832 	if (ret)
833 		goto out_free;
834 
835 	while (i < count && property_idx < MAX_THREAD_GROUP_PROPERTIES) {
836 		int j;
837 		struct thread_groups *tg = &tglp->property_tgs[property_idx++];
838 
839 		tg->property = thread_group_array[i];
840 		tg->nr_groups = thread_group_array[i + 1];
841 		tg->threads_per_group = thread_group_array[i + 2];
842 		total_threads = tg->nr_groups * tg->threads_per_group;
843 
844 		thread_list = &thread_group_array[i + 3];
845 
846 		for (j = 0; j < total_threads; j++)
847 			tg->thread_list[j] = thread_list[j];
848 		i = i + 3 + total_threads;
849 	}
850 
851 	tglp->nr_properties = property_idx;
852 
853 out_free:
854 	kfree(thread_group_array);
855 	return ret;
856 }
857 
858 /*
859  * get_cpu_thread_group_start : Searches the thread group in tg->thread_list
860  *                              that @cpu belongs to.
861  *
862  * @cpu : The logical CPU whose thread group is being searched.
863  * @tg : The thread-group structure of the CPU node which @cpu belongs
864  *       to.
865  *
866  * Returns the index to tg->thread_list that points to the start
867  * of the thread_group that @cpu belongs to.
868  *
869  * Returns -1 if cpu doesn't belong to any of the groups pointed to by
870  * tg->thread_list.
871  */
872 static int get_cpu_thread_group_start(int cpu, struct thread_groups *tg)
873 {
874 	int hw_cpu_id = get_hard_smp_processor_id(cpu);
875 	int i, j;
876 
877 	for (i = 0; i < tg->nr_groups; i++) {
878 		int group_start = i * tg->threads_per_group;
879 
880 		for (j = 0; j < tg->threads_per_group; j++) {
881 			int idx = group_start + j;
882 
883 			if (tg->thread_list[idx] == hw_cpu_id)
884 				return group_start;
885 		}
886 	}
887 
888 	return -1;
889 }
890 
891 static struct thread_groups *__init get_thread_groups(int cpu,
892 						      int group_property,
893 						      int *err)
894 {
895 	struct device_node *dn = of_get_cpu_node(cpu, NULL);
896 	struct thread_groups_list *cpu_tgl = &tgl[cpu];
897 	struct thread_groups *tg = NULL;
898 	int i;
899 	*err = 0;
900 
901 	if (!dn) {
902 		*err = -ENODATA;
903 		return NULL;
904 	}
905 
906 	if (!cpu_tgl->nr_properties) {
907 		*err = parse_thread_groups(dn, cpu_tgl);
908 		if (*err)
909 			goto out;
910 	}
911 
912 	for (i = 0; i < cpu_tgl->nr_properties; i++) {
913 		if (cpu_tgl->property_tgs[i].property == group_property) {
914 			tg = &cpu_tgl->property_tgs[i];
915 			break;
916 		}
917 	}
918 
919 	if (!tg)
920 		*err = -EINVAL;
921 out:
922 	of_node_put(dn);
923 	return tg;
924 }
925 
926 static int __init update_mask_from_threadgroup(cpumask_var_t *mask, struct thread_groups *tg,
927 					       int cpu, int cpu_group_start)
928 {
929 	int first_thread = cpu_first_thread_sibling(cpu);
930 	int i;
931 
932 	zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cpu));
933 
934 	for (i = first_thread; i < first_thread + threads_per_core; i++) {
935 		int i_group_start = get_cpu_thread_group_start(i, tg);
936 
937 		if (unlikely(i_group_start == -1)) {
938 			WARN_ON_ONCE(1);
939 			return -ENODATA;
940 		}
941 
942 		if (i_group_start == cpu_group_start)
943 			cpumask_set_cpu(i, *mask);
944 	}
945 
946 	return 0;
947 }
948 
949 static int __init init_thread_group_cache_map(int cpu, int cache_property)
950 
951 {
952 	int cpu_group_start = -1, err = 0;
953 	struct thread_groups *tg = NULL;
954 	cpumask_var_t *mask = NULL;
955 
956 	if (cache_property != THREAD_GROUP_SHARE_L1 &&
957 	    cache_property != THREAD_GROUP_SHARE_L2_L3)
958 		return -EINVAL;
959 
960 	tg = get_thread_groups(cpu, cache_property, &err);
961 
962 	if (!tg)
963 		return err;
964 
965 	cpu_group_start = get_cpu_thread_group_start(cpu, tg);
966 
967 	if (unlikely(cpu_group_start == -1)) {
968 		WARN_ON_ONCE(1);
969 		return -ENODATA;
970 	}
971 
972 	if (cache_property == THREAD_GROUP_SHARE_L1) {
973 		mask = &per_cpu(thread_group_l1_cache_map, cpu);
974 		update_mask_from_threadgroup(mask, tg, cpu, cpu_group_start);
975 	}
976 	else if (cache_property == THREAD_GROUP_SHARE_L2_L3) {
977 		mask = &per_cpu(thread_group_l2_cache_map, cpu);
978 		update_mask_from_threadgroup(mask, tg, cpu, cpu_group_start);
979 		mask = &per_cpu(thread_group_l3_cache_map, cpu);
980 		update_mask_from_threadgroup(mask, tg, cpu, cpu_group_start);
981 	}
982 
983 
984 	return 0;
985 }
986 
987 static bool shared_caches __ro_after_init;
988 
989 #ifdef CONFIG_SCHED_SMT
990 /* cpumask of CPUs with asymmetric SMT dependency */
991 static int powerpc_smt_flags(void)
992 {
993 	int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_LLC;
994 
995 	if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
996 		printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
997 		flags |= SD_ASYM_PACKING;
998 	}
999 	return flags;
1000 }
1001 #endif
1002 
1003 /*
1004  * On shared processor LPARs scheduled on a big core (which has two or more
1005  * independent thread groups per core), prefer lower numbered CPUs, so
1006  * that workload consolidates to lesser number of cores.
1007  */
1008 static __ro_after_init DEFINE_STATIC_KEY_FALSE(splpar_asym_pack);
1009 
1010 /*
1011  * P9 has a slightly odd architecture where pairs of cores share an L2 cache.
1012  * This topology makes it *much* cheaper to migrate tasks between adjacent cores
1013  * since the migrated task remains cache hot. We want to take advantage of this
1014  * at the scheduler level so an extra topology level is required.
1015  */
1016 static int powerpc_shared_cache_flags(void)
1017 {
1018 	if (static_branch_unlikely(&splpar_asym_pack))
1019 		return SD_SHARE_LLC | SD_ASYM_PACKING;
1020 
1021 	return SD_SHARE_LLC;
1022 }
1023 
1024 static int powerpc_shared_proc_flags(void)
1025 {
1026 	if (static_branch_unlikely(&splpar_asym_pack))
1027 		return SD_ASYM_PACKING;
1028 
1029 	return 0;
1030 }
1031 
1032 /*
1033  * We can't just pass cpu_l2_cache_mask() directly because
1034  * returns a non-const pointer and the compiler barfs on that.
1035  */
1036 static const struct cpumask *tl_cache_mask(struct sched_domain_topology_level *tl, int cpu)
1037 {
1038 	return per_cpu(cpu_l2_cache_map, cpu);
1039 }
1040 
1041 #ifdef CONFIG_SCHED_SMT
1042 static const struct cpumask *tl_smallcore_smt_mask(struct sched_domain_topology_level *tl, int cpu)
1043 {
1044 	return cpu_smallcore_mask(cpu);
1045 }
1046 #endif
1047 
1048 struct cpumask *cpu_coregroup_mask(int cpu)
1049 {
1050 	return per_cpu(cpu_coregroup_map, cpu);
1051 }
1052 
1053 static bool has_coregroup_support(void)
1054 {
1055 	/* Coregroup identification not available on shared systems */
1056 	if (is_shared_processor())
1057 		return 0;
1058 
1059 	return coregroup_enabled;
1060 }
1061 
1062 static int __init init_big_cores(void)
1063 {
1064 	int cpu;
1065 
1066 	for_each_possible_cpu(cpu) {
1067 		int err = init_thread_group_cache_map(cpu, THREAD_GROUP_SHARE_L1);
1068 
1069 		if (err)
1070 			return err;
1071 
1072 		zalloc_cpumask_var_node(&per_cpu(cpu_smallcore_map, cpu),
1073 					GFP_KERNEL,
1074 					cpu_to_node(cpu));
1075 	}
1076 
1077 	has_big_cores = true;
1078 
1079 	for_each_possible_cpu(cpu) {
1080 		int err = init_thread_group_cache_map(cpu, THREAD_GROUP_SHARE_L2_L3);
1081 
1082 		if (err)
1083 			return err;
1084 	}
1085 
1086 	thread_group_shares_l2 = true;
1087 	thread_group_shares_l3 = true;
1088 	pr_debug("L2/L3 cache only shared by the threads in the small core\n");
1089 
1090 	return 0;
1091 }
1092 
1093 /*
1094  * die_mask and die_id are only available on systems which support
1095  * multiple coregroups within a same package. On all other systems, die_mask
1096  * would be same as package mask and die_id would be set to -1.
1097  */
1098 const struct cpumask *cpu_die_mask(int cpu)
1099 {
1100 	if (has_coregroup_support())
1101 		return per_cpu(cpu_coregroup_map, cpu);
1102 	else
1103 		return cpu_node_mask(cpu);
1104 }
1105 EXPORT_SYMBOL_GPL(cpu_die_mask);
1106 
1107 int cpu_die_id(int cpu)
1108 {
1109 	if (has_coregroup_support())
1110 		return cpu_to_coregroup_id(cpu);
1111 	else
1112 		return -1;
1113 }
1114 EXPORT_SYMBOL_GPL(cpu_die_id);
1115 
1116 void __init smp_prepare_cpus(unsigned int max_cpus)
1117 {
1118 	unsigned int cpu, num_threads;
1119 
1120 	DBG("smp_prepare_cpus\n");
1121 
1122 	/*
1123 	 * setup_cpu may need to be called on the boot cpu. We haven't
1124 	 * spun any cpus up but lets be paranoid.
1125 	 */
1126 	BUG_ON(boot_cpuid != smp_processor_id());
1127 
1128 	/* Fixup boot cpu */
1129 	smp_store_cpu_info(boot_cpuid);
1130 	cpu_callin_map[boot_cpuid] = 1;
1131 
1132 	for_each_possible_cpu(cpu) {
1133 		zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
1134 					GFP_KERNEL, cpu_to_node(cpu));
1135 		zalloc_cpumask_var_node(&per_cpu(cpu_l2_cache_map, cpu),
1136 					GFP_KERNEL, cpu_to_node(cpu));
1137 		zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
1138 					GFP_KERNEL, cpu_to_node(cpu));
1139 		if (has_coregroup_support())
1140 			zalloc_cpumask_var_node(&per_cpu(cpu_coregroup_map, cpu),
1141 						GFP_KERNEL, cpu_to_node(cpu));
1142 
1143 #ifdef CONFIG_NUMA
1144 		/*
1145 		 * numa_node_id() works after this.
1146 		 */
1147 		if (cpu_present(cpu)) {
1148 			set_cpu_numa_node(cpu, numa_cpu_lookup_table[cpu]);
1149 			set_cpu_numa_mem(cpu,
1150 				local_memory_node(numa_cpu_lookup_table[cpu]));
1151 		}
1152 #endif
1153 	}
1154 
1155 	/* Init the cpumasks so the boot CPU is related to itself */
1156 	cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
1157 	cpumask_set_cpu(boot_cpuid, cpu_l2_cache_mask(boot_cpuid));
1158 	cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
1159 
1160 	if (has_coregroup_support())
1161 		cpumask_set_cpu(boot_cpuid, cpu_coregroup_mask(boot_cpuid));
1162 
1163 	init_big_cores();
1164 	if (has_big_cores) {
1165 		cpumask_set_cpu(boot_cpuid,
1166 				cpu_smallcore_mask(boot_cpuid));
1167 	}
1168 
1169 	if (cpu_to_chip_id(boot_cpuid) != -1) {
1170 		int idx = DIV_ROUND_UP(num_possible_cpus(), threads_per_core);
1171 
1172 		/*
1173 		 * All threads of a core will all belong to the same core,
1174 		 * chip_id_lookup_table will have one entry per core.
1175 		 * Assumption: if boot_cpuid doesn't have a chip-id, then no
1176 		 * other CPUs, will also not have chip-id.
1177 		 */
1178 		chip_id_lookup_table = kzalloc_objs(int, idx);
1179 		if (chip_id_lookup_table)
1180 			memset(chip_id_lookup_table, -1, sizeof(int) * idx);
1181 	}
1182 
1183 	if (smp_ops && smp_ops->probe)
1184 		smp_ops->probe();
1185 
1186 	// Initalise the generic SMT topology support
1187 	num_threads = 1;
1188 	if (smt_enabled_at_boot)
1189 		num_threads = smt_enabled_at_boot;
1190 	cpu_smt_set_num_threads(num_threads, threads_per_core);
1191 }
1192 
1193 void __init smp_prepare_boot_cpu(void)
1194 {
1195 	BUG_ON(smp_processor_id() != boot_cpuid);
1196 #ifdef CONFIG_PPC64
1197 	paca_ptrs[boot_cpuid]->__current = current;
1198 #endif
1199 	set_numa_node(numa_cpu_lookup_table[boot_cpuid]);
1200 	current_set[boot_cpuid] = current;
1201 }
1202 
1203 #ifdef CONFIG_HOTPLUG_CPU
1204 
1205 int generic_cpu_disable(void)
1206 {
1207 	unsigned int cpu = smp_processor_id();
1208 
1209 	if (cpu == boot_cpuid)
1210 		return -EBUSY;
1211 
1212 	set_cpu_online(cpu, false);
1213 #ifdef CONFIG_PPC64_PROC_SYSTEMCFG
1214 	systemcfg->processorCount--;
1215 #endif
1216 	/* Update affinity of all IRQs previously aimed at this CPU */
1217 	irq_migrate_all_off_this_cpu();
1218 
1219 	/*
1220 	 * Depending on the details of the interrupt controller, it's possible
1221 	 * that one of the interrupts we just migrated away from this CPU is
1222 	 * actually already pending on this CPU. If we leave it in that state
1223 	 * the interrupt will never be EOI'ed, and will never fire again. So
1224 	 * temporarily enable interrupts here, to allow any pending interrupt to
1225 	 * be received (and EOI'ed), before we take this CPU offline.
1226 	 */
1227 	local_irq_enable();
1228 	mdelay(1);
1229 	local_irq_disable();
1230 
1231 	return 0;
1232 }
1233 
1234 void generic_cpu_die(unsigned int cpu)
1235 {
1236 	int i;
1237 
1238 	for (i = 0; i < 100; i++) {
1239 		smp_rmb();
1240 		if (is_cpu_dead(cpu))
1241 			return;
1242 		msleep(100);
1243 	}
1244 	printk(KERN_ERR "CPU%d didn't die...\n", cpu);
1245 }
1246 
1247 void generic_set_cpu_dead(unsigned int cpu)
1248 {
1249 	per_cpu(cpu_state, cpu) = CPU_DEAD;
1250 }
1251 
1252 /*
1253  * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
1254  * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
1255  * which makes the delay in generic_cpu_die() not happen.
1256  */
1257 void generic_set_cpu_up(unsigned int cpu)
1258 {
1259 	per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1260 }
1261 
1262 int generic_check_cpu_restart(unsigned int cpu)
1263 {
1264 	return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
1265 }
1266 
1267 int is_cpu_dead(unsigned int cpu)
1268 {
1269 	return per_cpu(cpu_state, cpu) == CPU_DEAD;
1270 }
1271 
1272 static bool secondaries_inhibited(void)
1273 {
1274 	return kvm_hv_mode_active();
1275 }
1276 
1277 #else /* HOTPLUG_CPU */
1278 
1279 #define secondaries_inhibited()		0
1280 
1281 #endif
1282 
1283 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
1284 {
1285 #ifdef CONFIG_PPC64
1286 	paca_ptrs[cpu]->__current = idle;
1287 	paca_ptrs[cpu]->kstack = (unsigned long)task_stack_page(idle) +
1288 				 THREAD_SIZE - STACK_FRAME_MIN_SIZE;
1289 #endif
1290 	task_thread_info(idle)->cpu = cpu;
1291 	secondary_current = current_set[cpu] = idle;
1292 }
1293 
1294 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
1295 {
1296 	const unsigned long boot_spin_ms = 5 * MSEC_PER_SEC;
1297 	const bool booting = system_state < SYSTEM_RUNNING;
1298 	const unsigned long hp_spin_ms = 1;
1299 	unsigned long deadline;
1300 	int rc;
1301 	const unsigned long spin_wait_ms = booting ? boot_spin_ms : hp_spin_ms;
1302 
1303 	/*
1304 	 * Don't allow secondary threads to come online if inhibited
1305 	 */
1306 	if (threads_per_core > 1 && secondaries_inhibited() &&
1307 	    cpu_thread_in_subcore(cpu))
1308 		return -EBUSY;
1309 
1310 	if (smp_ops == NULL ||
1311 	    (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
1312 		return -EINVAL;
1313 
1314 	cpu_idle_thread_init(cpu, tidle);
1315 
1316 	/*
1317 	 * The platform might need to allocate resources prior to bringing
1318 	 * up the CPU
1319 	 */
1320 	if (smp_ops->prepare_cpu) {
1321 		rc = smp_ops->prepare_cpu(cpu);
1322 		if (rc)
1323 			return rc;
1324 	}
1325 
1326 	/* Make sure callin-map entry is 0 (can be leftover a CPU
1327 	 * hotplug
1328 	 */
1329 	cpu_callin_map[cpu] = 0;
1330 
1331 	/* The information for processor bringup must
1332 	 * be written out to main store before we release
1333 	 * the processor.
1334 	 */
1335 	smp_mb();
1336 
1337 	/* wake up cpus */
1338 	DBG("smp: kicking cpu %d\n", cpu);
1339 	rc = smp_ops->kick_cpu(cpu);
1340 	if (rc) {
1341 		pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
1342 		return rc;
1343 	}
1344 
1345 	/*
1346 	 * At boot time, simply spin on the callin word until the
1347 	 * deadline passes.
1348 	 *
1349 	 * At run time, spin for an optimistic amount of time to avoid
1350 	 * sleeping in the common case.
1351 	 */
1352 	deadline = jiffies + msecs_to_jiffies(spin_wait_ms);
1353 	spin_until_cond(cpu_callin_map[cpu] || time_is_before_jiffies(deadline));
1354 
1355 	if (!cpu_callin_map[cpu] && system_state >= SYSTEM_RUNNING) {
1356 		const unsigned long sleep_interval_us = 10 * USEC_PER_MSEC;
1357 		const unsigned long sleep_wait_ms = 100 * MSEC_PER_SEC;
1358 
1359 		deadline = jiffies + msecs_to_jiffies(sleep_wait_ms);
1360 		while (!cpu_callin_map[cpu] && time_is_after_jiffies(deadline))
1361 			fsleep(sleep_interval_us);
1362 	}
1363 
1364 	if (!cpu_callin_map[cpu]) {
1365 		printk(KERN_ERR "Processor %u is stuck.\n", cpu);
1366 		return -ENOENT;
1367 	}
1368 
1369 	DBG("Processor %u found.\n", cpu);
1370 
1371 	if (smp_ops->give_timebase)
1372 		smp_ops->give_timebase();
1373 
1374 	/* Wait until cpu puts itself in the online & active maps */
1375 	spin_until_cond(cpu_online(cpu));
1376 
1377 	return 0;
1378 }
1379 
1380 /* Return the value of the reg property corresponding to the given
1381  * logical cpu.
1382  */
1383 int cpu_to_core_id(int cpu)
1384 {
1385 	struct device_node *np;
1386 	int id = -1;
1387 
1388 	np = of_get_cpu_node(cpu, NULL);
1389 	if (!np)
1390 		goto out;
1391 
1392 	id = of_get_cpu_hwid(np, 0);
1393 out:
1394 	of_node_put(np);
1395 	return id;
1396 }
1397 EXPORT_SYMBOL_GPL(cpu_to_core_id);
1398 
1399 /* Helper routines for cpu to core mapping */
1400 int cpu_core_index_of_thread(int cpu)
1401 {
1402 	return cpu >> threads_shift;
1403 }
1404 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
1405 
1406 int cpu_first_thread_of_core(int core)
1407 {
1408 	return core << threads_shift;
1409 }
1410 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
1411 
1412 /* Must be called when no change can occur to cpu_present_mask,
1413  * i.e. during cpu online or offline.
1414  */
1415 static struct device_node *cpu_to_l2cache(int cpu)
1416 {
1417 	struct device_node *np;
1418 	struct device_node *cache;
1419 
1420 	if (!cpu_present(cpu))
1421 		return NULL;
1422 
1423 	np = of_get_cpu_node(cpu, NULL);
1424 	if (np == NULL)
1425 		return NULL;
1426 
1427 	cache = of_find_next_cache_node(np);
1428 
1429 	of_node_put(np);
1430 
1431 	return cache;
1432 }
1433 
1434 static bool update_mask_by_l2(int cpu, cpumask_var_t *mask)
1435 {
1436 	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1437 	struct device_node *l2_cache, *np;
1438 	int i;
1439 
1440 	if (has_big_cores)
1441 		submask_fn = cpu_smallcore_mask;
1442 
1443 	/*
1444 	 * If the threads in a thread-group share L2 cache, then the
1445 	 * L2-mask can be obtained from thread_group_l2_cache_map.
1446 	 */
1447 	if (thread_group_shares_l2) {
1448 		cpumask_set_cpu(cpu, cpu_l2_cache_mask(cpu));
1449 
1450 		for_each_cpu(i, per_cpu(thread_group_l2_cache_map, cpu)) {
1451 			if (cpu_online(i))
1452 				set_cpus_related(i, cpu, cpu_l2_cache_mask);
1453 		}
1454 
1455 		/* Verify that L1-cache siblings are a subset of L2 cache-siblings */
1456 		if (!cpumask_equal(submask_fn(cpu), cpu_l2_cache_mask(cpu)) &&
1457 		    !cpumask_subset(submask_fn(cpu), cpu_l2_cache_mask(cpu))) {
1458 			pr_warn_once("CPU %d : Inconsistent L1 and L2 cache siblings\n",
1459 				     cpu);
1460 		}
1461 
1462 		return true;
1463 	}
1464 
1465 	l2_cache = cpu_to_l2cache(cpu);
1466 	if (!l2_cache || !*mask) {
1467 		/* Assume only core siblings share cache with this CPU */
1468 		for_each_cpu(i, cpu_sibling_mask(cpu))
1469 			set_cpus_related(cpu, i, cpu_l2_cache_mask);
1470 
1471 		return false;
1472 	}
1473 
1474 	cpumask_and(*mask, cpu_online_mask, cpu_node_mask(cpu));
1475 
1476 	/* Update l2-cache mask with all the CPUs that are part of submask */
1477 	or_cpumasks_related(cpu, cpu, submask_fn, cpu_l2_cache_mask);
1478 
1479 	/* Skip all CPUs already part of current CPU l2-cache mask */
1480 	cpumask_andnot(*mask, *mask, cpu_l2_cache_mask(cpu));
1481 
1482 	for_each_cpu(i, *mask) {
1483 		/*
1484 		 * when updating the marks the current CPU has not been marked
1485 		 * online, but we need to update the cache masks
1486 		 */
1487 		np = cpu_to_l2cache(i);
1488 
1489 		/* Skip all CPUs already part of current CPU l2-cache */
1490 		if (np == l2_cache) {
1491 			or_cpumasks_related(cpu, i, submask_fn, cpu_l2_cache_mask);
1492 			cpumask_andnot(*mask, *mask, submask_fn(i));
1493 		} else {
1494 			cpumask_andnot(*mask, *mask, cpu_l2_cache_mask(i));
1495 		}
1496 
1497 		of_node_put(np);
1498 	}
1499 	of_node_put(l2_cache);
1500 
1501 	return true;
1502 }
1503 
1504 #ifdef CONFIG_HOTPLUG_CPU
1505 static void remove_cpu_from_masks(int cpu)
1506 {
1507 	struct cpumask *(*mask_fn)(int) = cpu_sibling_mask;
1508 	int i;
1509 
1510 	unmap_cpu_from_node(cpu);
1511 
1512 	if (shared_caches)
1513 		mask_fn = cpu_l2_cache_mask;
1514 
1515 	for_each_cpu(i, mask_fn(cpu)) {
1516 		set_cpus_unrelated(cpu, i, cpu_l2_cache_mask);
1517 		set_cpus_unrelated(cpu, i, cpu_sibling_mask);
1518 		if (has_big_cores)
1519 			set_cpus_unrelated(cpu, i, cpu_smallcore_mask);
1520 	}
1521 
1522 	for_each_cpu(i, cpu_core_mask(cpu))
1523 		set_cpus_unrelated(cpu, i, cpu_core_mask);
1524 
1525 	if (has_coregroup_support()) {
1526 		for_each_cpu(i, cpu_coregroup_mask(cpu))
1527 			set_cpus_unrelated(cpu, i, cpu_coregroup_mask);
1528 	}
1529 }
1530 #endif
1531 
1532 static inline void add_cpu_to_smallcore_masks(int cpu)
1533 {
1534 	int i;
1535 
1536 	if (!has_big_cores)
1537 		return;
1538 
1539 	cpumask_set_cpu(cpu, cpu_smallcore_mask(cpu));
1540 
1541 	for_each_cpu(i, per_cpu(thread_group_l1_cache_map, cpu)) {
1542 		if (cpu_online(i))
1543 			set_cpus_related(i, cpu, cpu_smallcore_mask);
1544 	}
1545 }
1546 
1547 static void update_coregroup_mask(int cpu, cpumask_var_t *mask)
1548 {
1549 	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1550 	int coregroup_id = cpu_to_coregroup_id(cpu);
1551 	int i;
1552 
1553 	if (shared_caches)
1554 		submask_fn = cpu_l2_cache_mask;
1555 
1556 	if (!*mask) {
1557 		/* Assume only siblings are part of this CPU's coregroup */
1558 		for_each_cpu(i, submask_fn(cpu))
1559 			set_cpus_related(cpu, i, cpu_coregroup_mask);
1560 
1561 		return;
1562 	}
1563 
1564 	cpumask_and(*mask, cpu_online_mask, cpu_node_mask(cpu));
1565 
1566 	/* Update coregroup mask with all the CPUs that are part of submask */
1567 	or_cpumasks_related(cpu, cpu, submask_fn, cpu_coregroup_mask);
1568 
1569 	/* Skip all CPUs already part of coregroup mask */
1570 	cpumask_andnot(*mask, *mask, cpu_coregroup_mask(cpu));
1571 
1572 	for_each_cpu(i, *mask) {
1573 		/* Skip all CPUs not part of this coregroup */
1574 		if (coregroup_id == cpu_to_coregroup_id(i)) {
1575 			or_cpumasks_related(cpu, i, submask_fn, cpu_coregroup_mask);
1576 			cpumask_andnot(*mask, *mask, submask_fn(i));
1577 		} else {
1578 			cpumask_andnot(*mask, *mask, cpu_coregroup_mask(i));
1579 		}
1580 	}
1581 }
1582 
1583 static void add_cpu_to_masks(int cpu)
1584 {
1585 	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1586 	int first_thread = cpu_first_thread_sibling(cpu);
1587 	cpumask_var_t mask;
1588 	int chip_id = -1;
1589 	bool ret;
1590 	int i;
1591 
1592 	/*
1593 	 * This CPU will not be in the online mask yet so we need to manually
1594 	 * add it to its own thread sibling mask.
1595 	 */
1596 	map_cpu_to_node(cpu, cpu_to_node(cpu));
1597 	cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
1598 	cpumask_set_cpu(cpu, cpu_core_mask(cpu));
1599 
1600 	for (i = first_thread; i < first_thread + threads_per_core; i++)
1601 		if (cpu_online(i))
1602 			set_cpus_related(i, cpu, cpu_sibling_mask);
1603 
1604 	add_cpu_to_smallcore_masks(cpu);
1605 
1606 	/* In CPU-hotplug path, hence use GFP_ATOMIC */
1607 	ret = alloc_cpumask_var_node(&mask, GFP_ATOMIC, cpu_to_node(cpu));
1608 	update_mask_by_l2(cpu, &mask);
1609 
1610 	if (has_coregroup_support())
1611 		update_coregroup_mask(cpu, &mask);
1612 
1613 	if (chip_id_lookup_table && ret)
1614 		chip_id = cpu_to_chip_id(cpu);
1615 
1616 	if (shared_caches)
1617 		submask_fn = cpu_l2_cache_mask;
1618 
1619 	/* Update core_mask with all the CPUs that are part of submask */
1620 	or_cpumasks_related(cpu, cpu, submask_fn, cpu_core_mask);
1621 
1622 	/* Skip all CPUs already part of current CPU core mask */
1623 	cpumask_andnot(mask, cpu_online_mask, cpu_core_mask(cpu));
1624 
1625 	/* If chip_id is -1; limit the cpu_core_mask to within PKG */
1626 	if (chip_id == -1)
1627 		cpumask_and(mask, mask, cpu_node_mask(cpu));
1628 
1629 	for_each_cpu(i, mask) {
1630 		if (chip_id == cpu_to_chip_id(i)) {
1631 			or_cpumasks_related(cpu, i, submask_fn, cpu_core_mask);
1632 			cpumask_andnot(mask, mask, submask_fn(i));
1633 		} else {
1634 			cpumask_andnot(mask, mask, cpu_core_mask(i));
1635 		}
1636 	}
1637 
1638 	free_cpumask_var(mask);
1639 }
1640 
1641 /* Activate a secondary processor. */
1642 __no_stack_protector
1643 void start_secondary(void *unused)
1644 {
1645 	unsigned int cpu = raw_smp_processor_id();
1646 
1647 	/* PPC64 calls setup_kup() in early_setup_secondary() */
1648 	if (IS_ENABLED(CONFIG_PPC32))
1649 		setup_kup();
1650 
1651 	mmgrab_lazy_tlb(&init_mm);
1652 	current->active_mm = &init_mm;
1653 	VM_WARN_ON(cpumask_test_cpu(smp_processor_id(), mm_cpumask(&init_mm)));
1654 	cpumask_set_cpu(cpu, mm_cpumask(&init_mm));
1655 	inc_mm_active_cpus(&init_mm);
1656 
1657 	smp_store_cpu_info(cpu);
1658 	set_dec(tb_ticks_per_jiffy);
1659 	rcutree_report_cpu_starting(cpu);
1660 	cpu_callin_map[cpu] = 1;
1661 
1662 	if (smp_ops->setup_cpu)
1663 		smp_ops->setup_cpu(cpu);
1664 	if (smp_ops->take_timebase)
1665 		smp_ops->take_timebase();
1666 
1667 	secondary_cpu_time_init();
1668 
1669 #ifdef CONFIG_PPC64_PROC_SYSTEMCFG
1670 	if (system_state == SYSTEM_RUNNING)
1671 		systemcfg->processorCount++;
1672 #endif
1673 
1674 #ifdef CONFIG_PPC64
1675 	vdso_getcpu_init();
1676 #endif
1677 	set_numa_node(numa_cpu_lookup_table[cpu]);
1678 	set_numa_mem(local_memory_node(numa_cpu_lookup_table[cpu]));
1679 
1680 	/* Update topology CPU masks */
1681 	add_cpu_to_masks(cpu);
1682 
1683 	/*
1684 	 * Check for any shared caches. Note that this must be done on a
1685 	 * per-core basis because one core in the pair might be disabled.
1686 	 */
1687 	if (!shared_caches) {
1688 		struct cpumask *(*sibling_mask)(int) = cpu_sibling_mask;
1689 		struct cpumask *mask = cpu_l2_cache_mask(cpu);
1690 
1691 		if (has_big_cores)
1692 			sibling_mask = cpu_smallcore_mask;
1693 
1694 		if (cpumask_weight(mask) > cpumask_weight(sibling_mask(cpu)))
1695 			shared_caches = true;
1696 	}
1697 
1698 	smp_wmb();
1699 	notify_cpu_starting(cpu);
1700 	set_cpu_online(cpu, true);
1701 
1702 	boot_init_stack_canary();
1703 
1704 	local_irq_enable();
1705 
1706 	/* We can enable ftrace for secondary cpus now */
1707 	this_cpu_enable_ftrace();
1708 
1709 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
1710 
1711 	BUG();
1712 }
1713 
1714 static struct sched_domain_topology_level powerpc_topology[6];
1715 
1716 static void __init build_sched_topology(void)
1717 {
1718 	int i = 0;
1719 
1720 	if (is_shared_processor() && has_big_cores)
1721 		static_branch_enable(&splpar_asym_pack);
1722 
1723 #ifdef CONFIG_SCHED_SMT
1724 	if (has_big_cores) {
1725 		pr_info("Big cores detected but using small core scheduling\n");
1726 		powerpc_topology[i++] =
1727 			SDTL_INIT(tl_smallcore_smt_mask, powerpc_smt_flags, SMT);
1728 	} else {
1729 		powerpc_topology[i++] = SDTL_INIT(tl_smt_mask, powerpc_smt_flags, SMT);
1730 	}
1731 #endif
1732 	if (shared_caches) {
1733 		powerpc_topology[i++] =
1734 			SDTL_INIT(tl_cache_mask, powerpc_shared_cache_flags, CACHE);
1735 	}
1736 
1737 	if (has_coregroup_support()) {
1738 		powerpc_topology[i++] =
1739 			SDTL_INIT(tl_mc_mask, powerpc_shared_proc_flags, MC);
1740 	}
1741 
1742 	powerpc_topology[i++] = SDTL_INIT(tl_pkg_mask, powerpc_shared_proc_flags, PKG);
1743 
1744 	/* There must be one trailing NULL entry left.  */
1745 	BUG_ON(i >= ARRAY_SIZE(powerpc_topology) - 1);
1746 
1747 	set_sched_topology(powerpc_topology);
1748 }
1749 
1750 void __init smp_cpus_done(unsigned int max_cpus)
1751 {
1752 	/*
1753 	 * We are running pinned to the boot CPU, see rest_init().
1754 	 */
1755 	if (smp_ops && smp_ops->setup_cpu)
1756 		smp_ops->setup_cpu(boot_cpuid);
1757 
1758 	if (smp_ops && smp_ops->bringup_done)
1759 		smp_ops->bringup_done();
1760 
1761 	dump_numa_cpu_topology();
1762 	build_sched_topology();
1763 }
1764 
1765 /*
1766  * For asym packing, by default lower numbered CPU has higher priority.
1767  * On shared processors, pack to lower numbered core. However avoid moving
1768  * between thread_groups within the same core.
1769  */
1770 int arch_asym_cpu_priority(int cpu)
1771 {
1772 	if (static_branch_unlikely(&splpar_asym_pack))
1773 		return -cpu / threads_per_core;
1774 
1775 	return -cpu;
1776 }
1777 
1778 #ifdef CONFIG_HOTPLUG_CPU
1779 int __cpu_disable(void)
1780 {
1781 	int cpu = smp_processor_id();
1782 	int err;
1783 
1784 	if (!smp_ops->cpu_disable)
1785 		return -ENOSYS;
1786 
1787 	this_cpu_disable_ftrace();
1788 
1789 	err = smp_ops->cpu_disable();
1790 	if (err)
1791 		return err;
1792 
1793 	/* Update sibling maps */
1794 	remove_cpu_from_masks(cpu);
1795 
1796 	return 0;
1797 }
1798 
1799 void __cpu_die(unsigned int cpu)
1800 {
1801 	/*
1802 	 * This could perhaps be a generic call in idlea_task_dead(), but
1803 	 * that requires testing from all archs, so first put it here to
1804 	 */
1805 	VM_WARN_ON_ONCE(!cpumask_test_cpu(cpu, mm_cpumask(&init_mm)));
1806 	dec_mm_active_cpus(&init_mm);
1807 	cpumask_clear_cpu(cpu, mm_cpumask(&init_mm));
1808 
1809 	if (smp_ops->cpu_die)
1810 		smp_ops->cpu_die(cpu);
1811 }
1812 
1813 void __noreturn arch_cpu_idle_dead(void)
1814 {
1815 	/*
1816 	 * Disable on the down path. This will be re-enabled by
1817 	 * start_secondary() via start_secondary_resume() below
1818 	 */
1819 	this_cpu_disable_ftrace();
1820 
1821 	if (smp_ops->cpu_offline_self)
1822 		smp_ops->cpu_offline_self();
1823 
1824 	/* If we return, we re-enter start_secondary */
1825 	start_secondary_resume();
1826 }
1827 
1828 #endif
1829