xref: /linux/arch/powerpc/kernel/smp.c (revision 9052e9c95d908d6c3d7570aadc8898e1d871c8bb)
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 
39 #include <asm/ptrace.h>
40 #include <linux/atomic.h>
41 #include <asm/irq.h>
42 #include <asm/hw_irq.h>
43 #include <asm/kvm_ppc.h>
44 #include <asm/dbell.h>
45 #include <asm/page.h>
46 #include <asm/prom.h>
47 #include <asm/smp.h>
48 #include <asm/time.h>
49 #include <asm/machdep.h>
50 #include <asm/cputhreads.h>
51 #include <asm/cputable.h>
52 #include <asm/mpic.h>
53 #include <asm/vdso_datapage.h>
54 #ifdef CONFIG_PPC64
55 #include <asm/paca.h>
56 #endif
57 #include <asm/vdso.h>
58 #include <asm/debug.h>
59 #include <asm/kexec.h>
60 #include <asm/asm-prototypes.h>
61 #include <asm/cpu_has_feature.h>
62 #include <asm/ftrace.h>
63 #include <asm/kup.h>
64 
65 #ifdef DEBUG
66 #include <asm/udbg.h>
67 #define DBG(fmt...) udbg_printf(fmt)
68 #else
69 #define DBG(fmt...)
70 #endif
71 
72 #ifdef CONFIG_HOTPLUG_CPU
73 /* State of each CPU during hotplug phases */
74 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
75 #endif
76 
77 struct task_struct *secondary_current;
78 bool has_big_cores;
79 bool coregroup_enabled;
80 bool thread_group_shares_l2;
81 bool thread_group_shares_l3;
82 
83 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
84 DEFINE_PER_CPU(cpumask_var_t, cpu_smallcore_map);
85 DEFINE_PER_CPU(cpumask_var_t, cpu_l2_cache_map);
86 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
87 static DEFINE_PER_CPU(cpumask_var_t, cpu_coregroup_map);
88 
89 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
90 EXPORT_PER_CPU_SYMBOL(cpu_l2_cache_map);
91 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
92 EXPORT_SYMBOL_GPL(has_big_cores);
93 
94 enum {
95 #ifdef CONFIG_SCHED_SMT
96 	smt_idx,
97 #endif
98 	cache_idx,
99 	mc_idx,
100 	die_idx,
101 };
102 
103 #define MAX_THREAD_LIST_SIZE	8
104 #define THREAD_GROUP_SHARE_L1   1
105 #define THREAD_GROUP_SHARE_L2_L3 2
106 struct thread_groups {
107 	unsigned int property;
108 	unsigned int nr_groups;
109 	unsigned int threads_per_group;
110 	unsigned int thread_list[MAX_THREAD_LIST_SIZE];
111 };
112 
113 /* Maximum number of properties that groups of threads within a core can share */
114 #define MAX_THREAD_GROUP_PROPERTIES 2
115 
116 struct thread_groups_list {
117 	unsigned int nr_properties;
118 	struct thread_groups property_tgs[MAX_THREAD_GROUP_PROPERTIES];
119 };
120 
121 static struct thread_groups_list tgl[NR_CPUS] __initdata;
122 /*
123  * On big-cores system, thread_group_l1_cache_map for each CPU corresponds to
124  * the set its siblings that share the L1-cache.
125  */
126 DEFINE_PER_CPU(cpumask_var_t, thread_group_l1_cache_map);
127 
128 /*
129  * On some big-cores system, thread_group_l2_cache_map for each CPU
130  * corresponds to the set its siblings within the core that share the
131  * L2-cache.
132  */
133 DEFINE_PER_CPU(cpumask_var_t, thread_group_l2_cache_map);
134 
135 /*
136  * On P10, thread_group_l3_cache_map for each CPU is equal to the
137  * thread_group_l2_cache_map
138  */
139 DEFINE_PER_CPU(cpumask_var_t, thread_group_l3_cache_map);
140 
141 /* SMP operations for this machine */
142 struct smp_ops_t *smp_ops;
143 
144 /* Can't be static due to PowerMac hackery */
145 volatile unsigned int cpu_callin_map[NR_CPUS];
146 
147 int smt_enabled_at_boot = 1;
148 
149 /*
150  * Returns 1 if the specified cpu should be brought up during boot.
151  * Used to inhibit booting threads if they've been disabled or
152  * limited on the command line
153  */
154 int smp_generic_cpu_bootable(unsigned int nr)
155 {
156 	/* Special case - we inhibit secondary thread startup
157 	 * during boot if the user requests it.
158 	 */
159 	if (system_state < SYSTEM_RUNNING && cpu_has_feature(CPU_FTR_SMT)) {
160 		if (!smt_enabled_at_boot && cpu_thread_in_core(nr) != 0)
161 			return 0;
162 		if (smt_enabled_at_boot
163 		    && cpu_thread_in_core(nr) >= smt_enabled_at_boot)
164 			return 0;
165 	}
166 
167 	return 1;
168 }
169 
170 
171 #ifdef CONFIG_PPC64
172 int smp_generic_kick_cpu(int nr)
173 {
174 	if (nr < 0 || nr >= nr_cpu_ids)
175 		return -EINVAL;
176 
177 	/*
178 	 * The processor is currently spinning, waiting for the
179 	 * cpu_start field to become non-zero After we set cpu_start,
180 	 * the processor will continue on to secondary_start
181 	 */
182 	if (!paca_ptrs[nr]->cpu_start) {
183 		paca_ptrs[nr]->cpu_start = 1;
184 		smp_mb();
185 		return 0;
186 	}
187 
188 #ifdef CONFIG_HOTPLUG_CPU
189 	/*
190 	 * Ok it's not there, so it might be soft-unplugged, let's
191 	 * try to bring it back
192 	 */
193 	generic_set_cpu_up(nr);
194 	smp_wmb();
195 	smp_send_reschedule(nr);
196 #endif /* CONFIG_HOTPLUG_CPU */
197 
198 	return 0;
199 }
200 #endif /* CONFIG_PPC64 */
201 
202 static irqreturn_t call_function_action(int irq, void *data)
203 {
204 	generic_smp_call_function_interrupt();
205 	return IRQ_HANDLED;
206 }
207 
208 static irqreturn_t reschedule_action(int irq, void *data)
209 {
210 	scheduler_ipi();
211 	return IRQ_HANDLED;
212 }
213 
214 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
215 static irqreturn_t tick_broadcast_ipi_action(int irq, void *data)
216 {
217 	timer_broadcast_interrupt();
218 	return IRQ_HANDLED;
219 }
220 #endif
221 
222 #ifdef CONFIG_NMI_IPI
223 static irqreturn_t nmi_ipi_action(int irq, void *data)
224 {
225 	smp_handle_nmi_ipi(get_irq_regs());
226 	return IRQ_HANDLED;
227 }
228 #endif
229 
230 static irq_handler_t smp_ipi_action[] = {
231 	[PPC_MSG_CALL_FUNCTION] =  call_function_action,
232 	[PPC_MSG_RESCHEDULE] = reschedule_action,
233 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
234 	[PPC_MSG_TICK_BROADCAST] = tick_broadcast_ipi_action,
235 #endif
236 #ifdef CONFIG_NMI_IPI
237 	[PPC_MSG_NMI_IPI] = nmi_ipi_action,
238 #endif
239 };
240 
241 /*
242  * The NMI IPI is a fallback and not truly non-maskable. It is simpler
243  * than going through the call function infrastructure, and strongly
244  * serialized, so it is more appropriate for debugging.
245  */
246 const char *smp_ipi_name[] = {
247 	[PPC_MSG_CALL_FUNCTION] =  "ipi call function",
248 	[PPC_MSG_RESCHEDULE] = "ipi reschedule",
249 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
250 	[PPC_MSG_TICK_BROADCAST] = "ipi tick-broadcast",
251 #endif
252 #ifdef CONFIG_NMI_IPI
253 	[PPC_MSG_NMI_IPI] = "nmi ipi",
254 #endif
255 };
256 
257 /* optional function to request ipi, for controllers with >= 4 ipis */
258 int smp_request_message_ipi(int virq, int msg)
259 {
260 	int err;
261 
262 	if (msg < 0 || msg > PPC_MSG_NMI_IPI)
263 		return -EINVAL;
264 #ifndef CONFIG_NMI_IPI
265 	if (msg == PPC_MSG_NMI_IPI)
266 		return 1;
267 #endif
268 
269 	err = request_irq(virq, smp_ipi_action[msg],
270 			  IRQF_PERCPU | IRQF_NO_THREAD | IRQF_NO_SUSPEND,
271 			  smp_ipi_name[msg], NULL);
272 	WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
273 		virq, smp_ipi_name[msg], err);
274 
275 	return err;
276 }
277 
278 #ifdef CONFIG_PPC_SMP_MUXED_IPI
279 struct cpu_messages {
280 	long messages;			/* current messages */
281 };
282 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
283 
284 void smp_muxed_ipi_set_message(int cpu, int msg)
285 {
286 	struct cpu_messages *info = &per_cpu(ipi_message, cpu);
287 	char *message = (char *)&info->messages;
288 
289 	/*
290 	 * Order previous accesses before accesses in the IPI handler.
291 	 */
292 	smp_mb();
293 	message[msg] = 1;
294 }
295 
296 void smp_muxed_ipi_message_pass(int cpu, int msg)
297 {
298 	smp_muxed_ipi_set_message(cpu, msg);
299 
300 	/*
301 	 * cause_ipi functions are required to include a full barrier
302 	 * before doing whatever causes the IPI.
303 	 */
304 	smp_ops->cause_ipi(cpu);
305 }
306 
307 #ifdef __BIG_ENDIAN__
308 #define IPI_MESSAGE(A) (1uL << ((BITS_PER_LONG - 8) - 8 * (A)))
309 #else
310 #define IPI_MESSAGE(A) (1uL << (8 * (A)))
311 #endif
312 
313 irqreturn_t smp_ipi_demux(void)
314 {
315 	mb();	/* order any irq clear */
316 
317 	return smp_ipi_demux_relaxed();
318 }
319 
320 /* sync-free variant. Callers should ensure synchronization */
321 irqreturn_t smp_ipi_demux_relaxed(void)
322 {
323 	struct cpu_messages *info;
324 	unsigned long all;
325 
326 	info = this_cpu_ptr(&ipi_message);
327 	do {
328 		all = xchg(&info->messages, 0);
329 #if defined(CONFIG_KVM_XICS) && defined(CONFIG_KVM_BOOK3S_HV_POSSIBLE)
330 		/*
331 		 * Must check for PPC_MSG_RM_HOST_ACTION messages
332 		 * before PPC_MSG_CALL_FUNCTION messages because when
333 		 * a VM is destroyed, we call kick_all_cpus_sync()
334 		 * to ensure that any pending PPC_MSG_RM_HOST_ACTION
335 		 * messages have completed before we free any VCPUs.
336 		 */
337 		if (all & IPI_MESSAGE(PPC_MSG_RM_HOST_ACTION))
338 			kvmppc_xics_ipi_action();
339 #endif
340 		if (all & IPI_MESSAGE(PPC_MSG_CALL_FUNCTION))
341 			generic_smp_call_function_interrupt();
342 		if (all & IPI_MESSAGE(PPC_MSG_RESCHEDULE))
343 			scheduler_ipi();
344 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
345 		if (all & IPI_MESSAGE(PPC_MSG_TICK_BROADCAST))
346 			timer_broadcast_interrupt();
347 #endif
348 #ifdef CONFIG_NMI_IPI
349 		if (all & IPI_MESSAGE(PPC_MSG_NMI_IPI))
350 			nmi_ipi_action(0, NULL);
351 #endif
352 	} while (info->messages);
353 
354 	return IRQ_HANDLED;
355 }
356 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
357 
358 static inline void do_message_pass(int cpu, int msg)
359 {
360 	if (smp_ops->message_pass)
361 		smp_ops->message_pass(cpu, msg);
362 #ifdef CONFIG_PPC_SMP_MUXED_IPI
363 	else
364 		smp_muxed_ipi_message_pass(cpu, msg);
365 #endif
366 }
367 
368 void smp_send_reschedule(int cpu)
369 {
370 	if (likely(smp_ops))
371 		do_message_pass(cpu, PPC_MSG_RESCHEDULE);
372 }
373 EXPORT_SYMBOL_GPL(smp_send_reschedule);
374 
375 void arch_send_call_function_single_ipi(int cpu)
376 {
377 	do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
378 }
379 
380 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
381 {
382 	unsigned int cpu;
383 
384 	for_each_cpu(cpu, mask)
385 		do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
386 }
387 
388 #ifdef CONFIG_NMI_IPI
389 
390 /*
391  * "NMI IPI" system.
392  *
393  * NMI IPIs may not be recoverable, so should not be used as ongoing part of
394  * a running system. They can be used for crash, debug, halt/reboot, etc.
395  *
396  * The IPI call waits with interrupts disabled until all targets enter the
397  * NMI handler, then returns. Subsequent IPIs can be issued before targets
398  * have returned from their handlers, so there is no guarantee about
399  * concurrency or re-entrancy.
400  *
401  * A new NMI can be issued before all targets exit the handler.
402  *
403  * The IPI call may time out without all targets entering the NMI handler.
404  * In that case, there is some logic to recover (and ignore subsequent
405  * NMI interrupts that may eventually be raised), but the platform interrupt
406  * handler may not be able to distinguish this from other exception causes,
407  * which may cause a crash.
408  */
409 
410 static atomic_t __nmi_ipi_lock = ATOMIC_INIT(0);
411 static struct cpumask nmi_ipi_pending_mask;
412 static bool nmi_ipi_busy = false;
413 static void (*nmi_ipi_function)(struct pt_regs *) = NULL;
414 
415 static void nmi_ipi_lock_start(unsigned long *flags)
416 {
417 	raw_local_irq_save(*flags);
418 	hard_irq_disable();
419 	while (atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1) {
420 		raw_local_irq_restore(*flags);
421 		spin_until_cond(atomic_read(&__nmi_ipi_lock) == 0);
422 		raw_local_irq_save(*flags);
423 		hard_irq_disable();
424 	}
425 }
426 
427 static void nmi_ipi_lock(void)
428 {
429 	while (atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1)
430 		spin_until_cond(atomic_read(&__nmi_ipi_lock) == 0);
431 }
432 
433 static void nmi_ipi_unlock(void)
434 {
435 	smp_mb();
436 	WARN_ON(atomic_read(&__nmi_ipi_lock) != 1);
437 	atomic_set(&__nmi_ipi_lock, 0);
438 }
439 
440 static void nmi_ipi_unlock_end(unsigned long *flags)
441 {
442 	nmi_ipi_unlock();
443 	raw_local_irq_restore(*flags);
444 }
445 
446 /*
447  * Platform NMI handler calls this to ack
448  */
449 int smp_handle_nmi_ipi(struct pt_regs *regs)
450 {
451 	void (*fn)(struct pt_regs *) = NULL;
452 	unsigned long flags;
453 	int me = raw_smp_processor_id();
454 	int ret = 0;
455 
456 	/*
457 	 * Unexpected NMIs are possible here because the interrupt may not
458 	 * be able to distinguish NMI IPIs from other types of NMIs, or
459 	 * because the caller may have timed out.
460 	 */
461 	nmi_ipi_lock_start(&flags);
462 	if (cpumask_test_cpu(me, &nmi_ipi_pending_mask)) {
463 		cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
464 		fn = READ_ONCE(nmi_ipi_function);
465 		WARN_ON_ONCE(!fn);
466 		ret = 1;
467 	}
468 	nmi_ipi_unlock_end(&flags);
469 
470 	if (fn)
471 		fn(regs);
472 
473 	return ret;
474 }
475 
476 static void do_smp_send_nmi_ipi(int cpu, bool safe)
477 {
478 	if (!safe && smp_ops->cause_nmi_ipi && smp_ops->cause_nmi_ipi(cpu))
479 		return;
480 
481 	if (cpu >= 0) {
482 		do_message_pass(cpu, PPC_MSG_NMI_IPI);
483 	} else {
484 		int c;
485 
486 		for_each_online_cpu(c) {
487 			if (c == raw_smp_processor_id())
488 				continue;
489 			do_message_pass(c, PPC_MSG_NMI_IPI);
490 		}
491 	}
492 }
493 
494 /*
495  * - cpu is the target CPU (must not be this CPU), or NMI_IPI_ALL_OTHERS.
496  * - fn is the target callback function.
497  * - delay_us > 0 is the delay before giving up waiting for targets to
498  *   begin executing the handler, == 0 specifies indefinite delay.
499  */
500 static int __smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *),
501 				u64 delay_us, bool safe)
502 {
503 	unsigned long flags;
504 	int me = raw_smp_processor_id();
505 	int ret = 1;
506 
507 	BUG_ON(cpu == me);
508 	BUG_ON(cpu < 0 && cpu != NMI_IPI_ALL_OTHERS);
509 
510 	if (unlikely(!smp_ops))
511 		return 0;
512 
513 	nmi_ipi_lock_start(&flags);
514 	while (nmi_ipi_busy) {
515 		nmi_ipi_unlock_end(&flags);
516 		spin_until_cond(!nmi_ipi_busy);
517 		nmi_ipi_lock_start(&flags);
518 	}
519 	nmi_ipi_busy = true;
520 	nmi_ipi_function = fn;
521 
522 	WARN_ON_ONCE(!cpumask_empty(&nmi_ipi_pending_mask));
523 
524 	if (cpu < 0) {
525 		/* ALL_OTHERS */
526 		cpumask_copy(&nmi_ipi_pending_mask, cpu_online_mask);
527 		cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
528 	} else {
529 		cpumask_set_cpu(cpu, &nmi_ipi_pending_mask);
530 	}
531 
532 	nmi_ipi_unlock();
533 
534 	/* Interrupts remain hard disabled */
535 
536 	do_smp_send_nmi_ipi(cpu, safe);
537 
538 	nmi_ipi_lock();
539 	/* nmi_ipi_busy is set here, so unlock/lock is okay */
540 	while (!cpumask_empty(&nmi_ipi_pending_mask)) {
541 		nmi_ipi_unlock();
542 		udelay(1);
543 		nmi_ipi_lock();
544 		if (delay_us) {
545 			delay_us--;
546 			if (!delay_us)
547 				break;
548 		}
549 	}
550 
551 	if (!cpumask_empty(&nmi_ipi_pending_mask)) {
552 		/* Timeout waiting for CPUs to call smp_handle_nmi_ipi */
553 		ret = 0;
554 		cpumask_clear(&nmi_ipi_pending_mask);
555 	}
556 
557 	nmi_ipi_function = NULL;
558 	nmi_ipi_busy = false;
559 
560 	nmi_ipi_unlock_end(&flags);
561 
562 	return ret;
563 }
564 
565 int smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
566 {
567 	return __smp_send_nmi_ipi(cpu, fn, delay_us, false);
568 }
569 
570 int smp_send_safe_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
571 {
572 	return __smp_send_nmi_ipi(cpu, fn, delay_us, true);
573 }
574 #endif /* CONFIG_NMI_IPI */
575 
576 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
577 void tick_broadcast(const struct cpumask *mask)
578 {
579 	unsigned int cpu;
580 
581 	for_each_cpu(cpu, mask)
582 		do_message_pass(cpu, PPC_MSG_TICK_BROADCAST);
583 }
584 #endif
585 
586 #ifdef CONFIG_DEBUGGER
587 static void debugger_ipi_callback(struct pt_regs *regs)
588 {
589 	debugger_ipi(regs);
590 }
591 
592 void smp_send_debugger_break(void)
593 {
594 	smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, debugger_ipi_callback, 1000000);
595 }
596 #endif
597 
598 #ifdef CONFIG_KEXEC_CORE
599 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
600 {
601 	int cpu;
602 
603 	smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, crash_ipi_callback, 1000000);
604 	if (kdump_in_progress() && crash_wake_offline) {
605 		for_each_present_cpu(cpu) {
606 			if (cpu_online(cpu))
607 				continue;
608 			/*
609 			 * crash_ipi_callback will wait for
610 			 * all cpus, including offline CPUs.
611 			 * We don't care about nmi_ipi_function.
612 			 * Offline cpus will jump straight into
613 			 * crash_ipi_callback, we can skip the
614 			 * entire NMI dance and waiting for
615 			 * cpus to clear pending mask, etc.
616 			 */
617 			do_smp_send_nmi_ipi(cpu, false);
618 		}
619 	}
620 }
621 #endif
622 
623 #ifdef CONFIG_NMI_IPI
624 static void nmi_stop_this_cpu(struct pt_regs *regs)
625 {
626 	/*
627 	 * IRQs are already hard disabled by the smp_handle_nmi_ipi.
628 	 */
629 	set_cpu_online(smp_processor_id(), false);
630 
631 	spin_begin();
632 	while (1)
633 		spin_cpu_relax();
634 }
635 
636 void smp_send_stop(void)
637 {
638 	smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, nmi_stop_this_cpu, 1000000);
639 }
640 
641 #else /* CONFIG_NMI_IPI */
642 
643 static void stop_this_cpu(void *dummy)
644 {
645 	hard_irq_disable();
646 
647 	/*
648 	 * Offlining CPUs in stop_this_cpu can result in scheduler warnings,
649 	 * (see commit de6e5d38417e), but printk_safe_flush_on_panic() wants
650 	 * to know other CPUs are offline before it breaks locks to flush
651 	 * printk buffers, in case we panic()ed while holding the lock.
652 	 */
653 	set_cpu_online(smp_processor_id(), false);
654 
655 	spin_begin();
656 	while (1)
657 		spin_cpu_relax();
658 }
659 
660 void smp_send_stop(void)
661 {
662 	static bool stopped = false;
663 
664 	/*
665 	 * Prevent waiting on csd lock from a previous smp_send_stop.
666 	 * This is racy, but in general callers try to do the right
667 	 * thing and only fire off one smp_send_stop (e.g., see
668 	 * kernel/panic.c)
669 	 */
670 	if (stopped)
671 		return;
672 
673 	stopped = true;
674 
675 	smp_call_function(stop_this_cpu, NULL, 0);
676 }
677 #endif /* CONFIG_NMI_IPI */
678 
679 struct task_struct *current_set[NR_CPUS];
680 
681 static void smp_store_cpu_info(int id)
682 {
683 	per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
684 #ifdef CONFIG_PPC_FSL_BOOK3E
685 	per_cpu(next_tlbcam_idx, id)
686 		= (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
687 #endif
688 }
689 
690 /*
691  * Relationships between CPUs are maintained in a set of per-cpu cpumasks so
692  * rather than just passing around the cpumask we pass around a function that
693  * returns the that cpumask for the given CPU.
694  */
695 static void set_cpus_related(int i, int j, struct cpumask *(*get_cpumask)(int))
696 {
697 	cpumask_set_cpu(i, get_cpumask(j));
698 	cpumask_set_cpu(j, get_cpumask(i));
699 }
700 
701 #ifdef CONFIG_HOTPLUG_CPU
702 static void set_cpus_unrelated(int i, int j,
703 		struct cpumask *(*get_cpumask)(int))
704 {
705 	cpumask_clear_cpu(i, get_cpumask(j));
706 	cpumask_clear_cpu(j, get_cpumask(i));
707 }
708 #endif
709 
710 /*
711  * Extends set_cpus_related. Instead of setting one CPU at a time in
712  * dstmask, set srcmask at oneshot. dstmask should be super set of srcmask.
713  */
714 static void or_cpumasks_related(int i, int j, struct cpumask *(*srcmask)(int),
715 				struct cpumask *(*dstmask)(int))
716 {
717 	struct cpumask *mask;
718 	int k;
719 
720 	mask = srcmask(j);
721 	for_each_cpu(k, srcmask(i))
722 		cpumask_or(dstmask(k), dstmask(k), mask);
723 
724 	if (i == j)
725 		return;
726 
727 	mask = srcmask(i);
728 	for_each_cpu(k, srcmask(j))
729 		cpumask_or(dstmask(k), dstmask(k), mask);
730 }
731 
732 /*
733  * parse_thread_groups: Parses the "ibm,thread-groups" device tree
734  *                      property for the CPU device node @dn and stores
735  *                      the parsed output in the thread_groups_list
736  *                      structure @tglp.
737  *
738  * @dn: The device node of the CPU device.
739  * @tglp: Pointer to a thread group list structure into which the parsed
740  *      output of "ibm,thread-groups" is stored.
741  *
742  * ibm,thread-groups[0..N-1] array defines which group of threads in
743  * the CPU-device node can be grouped together based on the property.
744  *
745  * This array can represent thread groupings for multiple properties.
746  *
747  * ibm,thread-groups[i + 0] tells us the property based on which the
748  * threads are being grouped together. If this value is 1, it implies
749  * that the threads in the same group share L1, translation cache. If
750  * the value is 2, it implies that the threads in the same group share
751  * the same L2 cache.
752  *
753  * ibm,thread-groups[i+1] tells us how many such thread groups exist for the
754  * property ibm,thread-groups[i]
755  *
756  * ibm,thread-groups[i+2] tells us the number of threads in each such
757  * group.
758  * Suppose k = (ibm,thread-groups[i+1] * ibm,thread-groups[i+2]), then,
759  *
760  * ibm,thread-groups[i+3..i+k+2] (is the list of threads identified by
761  * "ibm,ppc-interrupt-server#s" arranged as per their membership in
762  * the grouping.
763  *
764  * Example:
765  * 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]
766  * This can be decomposed up into two consecutive arrays:
767  * a) [1,2,4,8,10,12,14,9,11,13,15]
768  * b) [2,2,4,8,10,12,14,9,11,13,15]
769  *
770  * where in,
771  *
772  * a) provides information of Property "1" being shared by "2" groups,
773  *  each with "4" threads each. The "ibm,ppc-interrupt-server#s" of
774  *  the first group is {8,10,12,14} and the
775  *  "ibm,ppc-interrupt-server#s" of the second group is
776  *  {9,11,13,15}. Property "1" is indicative of the thread in the
777  *  group sharing L1 cache, translation cache and Instruction Data
778  *  flow.
779  *
780  * b) provides information of Property "2" being shared by "2" groups,
781  *  each group with "4" threads. The "ibm,ppc-interrupt-server#s" of
782  *  the first group is {8,10,12,14} and the
783  *  "ibm,ppc-interrupt-server#s" of the second group is
784  *  {9,11,13,15}. Property "2" indicates that the threads in each
785  *  group share the L2-cache.
786  *
787  * Returns 0 on success, -EINVAL if the property does not exist,
788  * -ENODATA if property does not have a value, and -EOVERFLOW if the
789  * property data isn't large enough.
790  */
791 static int parse_thread_groups(struct device_node *dn,
792 			       struct thread_groups_list *tglp)
793 {
794 	unsigned int property_idx = 0;
795 	u32 *thread_group_array;
796 	size_t total_threads;
797 	int ret = 0, count;
798 	u32 *thread_list;
799 	int i = 0;
800 
801 	count = of_property_count_u32_elems(dn, "ibm,thread-groups");
802 	thread_group_array = kcalloc(count, sizeof(u32), GFP_KERNEL);
803 	ret = of_property_read_u32_array(dn, "ibm,thread-groups",
804 					 thread_group_array, count);
805 	if (ret)
806 		goto out_free;
807 
808 	while (i < count && property_idx < MAX_THREAD_GROUP_PROPERTIES) {
809 		int j;
810 		struct thread_groups *tg = &tglp->property_tgs[property_idx++];
811 
812 		tg->property = thread_group_array[i];
813 		tg->nr_groups = thread_group_array[i + 1];
814 		tg->threads_per_group = thread_group_array[i + 2];
815 		total_threads = tg->nr_groups * tg->threads_per_group;
816 
817 		thread_list = &thread_group_array[i + 3];
818 
819 		for (j = 0; j < total_threads; j++)
820 			tg->thread_list[j] = thread_list[j];
821 		i = i + 3 + total_threads;
822 	}
823 
824 	tglp->nr_properties = property_idx;
825 
826 out_free:
827 	kfree(thread_group_array);
828 	return ret;
829 }
830 
831 /*
832  * get_cpu_thread_group_start : Searches the thread group in tg->thread_list
833  *                              that @cpu belongs to.
834  *
835  * @cpu : The logical CPU whose thread group is being searched.
836  * @tg : The thread-group structure of the CPU node which @cpu belongs
837  *       to.
838  *
839  * Returns the index to tg->thread_list that points to the the start
840  * of the thread_group that @cpu belongs to.
841  *
842  * Returns -1 if cpu doesn't belong to any of the groups pointed to by
843  * tg->thread_list.
844  */
845 static int get_cpu_thread_group_start(int cpu, struct thread_groups *tg)
846 {
847 	int hw_cpu_id = get_hard_smp_processor_id(cpu);
848 	int i, j;
849 
850 	for (i = 0; i < tg->nr_groups; i++) {
851 		int group_start = i * tg->threads_per_group;
852 
853 		for (j = 0; j < tg->threads_per_group; j++) {
854 			int idx = group_start + j;
855 
856 			if (tg->thread_list[idx] == hw_cpu_id)
857 				return group_start;
858 		}
859 	}
860 
861 	return -1;
862 }
863 
864 static struct thread_groups *__init get_thread_groups(int cpu,
865 						      int group_property,
866 						      int *err)
867 {
868 	struct device_node *dn = of_get_cpu_node(cpu, NULL);
869 	struct thread_groups_list *cpu_tgl = &tgl[cpu];
870 	struct thread_groups *tg = NULL;
871 	int i;
872 	*err = 0;
873 
874 	if (!dn) {
875 		*err = -ENODATA;
876 		return NULL;
877 	}
878 
879 	if (!cpu_tgl->nr_properties) {
880 		*err = parse_thread_groups(dn, cpu_tgl);
881 		if (*err)
882 			goto out;
883 	}
884 
885 	for (i = 0; i < cpu_tgl->nr_properties; i++) {
886 		if (cpu_tgl->property_tgs[i].property == group_property) {
887 			tg = &cpu_tgl->property_tgs[i];
888 			break;
889 		}
890 	}
891 
892 	if (!tg)
893 		*err = -EINVAL;
894 out:
895 	of_node_put(dn);
896 	return tg;
897 }
898 
899 static int update_mask_from_threadgroup(cpumask_var_t *mask, struct thread_groups *tg, int cpu, int cpu_group_start)
900 {
901 	int first_thread = cpu_first_thread_sibling(cpu);
902 	int i;
903 
904 	zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cpu));
905 
906 	for (i = first_thread; i < first_thread + threads_per_core; i++) {
907 		int i_group_start = get_cpu_thread_group_start(i, tg);
908 
909 		if (unlikely(i_group_start == -1)) {
910 			WARN_ON_ONCE(1);
911 			return -ENODATA;
912 		}
913 
914 		if (i_group_start == cpu_group_start)
915 			cpumask_set_cpu(i, *mask);
916 	}
917 
918 	return 0;
919 }
920 
921 static int __init init_thread_group_cache_map(int cpu, int cache_property)
922 
923 {
924 	int cpu_group_start = -1, err = 0;
925 	struct thread_groups *tg = NULL;
926 	cpumask_var_t *mask = NULL;
927 
928 	if (cache_property != THREAD_GROUP_SHARE_L1 &&
929 	    cache_property != THREAD_GROUP_SHARE_L2_L3)
930 		return -EINVAL;
931 
932 	tg = get_thread_groups(cpu, cache_property, &err);
933 
934 	if (!tg)
935 		return err;
936 
937 	cpu_group_start = get_cpu_thread_group_start(cpu, tg);
938 
939 	if (unlikely(cpu_group_start == -1)) {
940 		WARN_ON_ONCE(1);
941 		return -ENODATA;
942 	}
943 
944 	if (cache_property == THREAD_GROUP_SHARE_L1) {
945 		mask = &per_cpu(thread_group_l1_cache_map, cpu);
946 		update_mask_from_threadgroup(mask, tg, cpu, cpu_group_start);
947 	}
948 	else if (cache_property == THREAD_GROUP_SHARE_L2_L3) {
949 		mask = &per_cpu(thread_group_l2_cache_map, cpu);
950 		update_mask_from_threadgroup(mask, tg, cpu, cpu_group_start);
951 		mask = &per_cpu(thread_group_l3_cache_map, cpu);
952 		update_mask_from_threadgroup(mask, tg, cpu, cpu_group_start);
953 	}
954 
955 
956 	return 0;
957 }
958 
959 static bool shared_caches;
960 
961 #ifdef CONFIG_SCHED_SMT
962 /* cpumask of CPUs with asymmetric SMT dependency */
963 static int powerpc_smt_flags(void)
964 {
965 	int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
966 
967 	if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
968 		printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
969 		flags |= SD_ASYM_PACKING;
970 	}
971 	return flags;
972 }
973 #endif
974 
975 /*
976  * P9 has a slightly odd architecture where pairs of cores share an L2 cache.
977  * This topology makes it *much* cheaper to migrate tasks between adjacent cores
978  * since the migrated task remains cache hot. We want to take advantage of this
979  * at the scheduler level so an extra topology level is required.
980  */
981 static int powerpc_shared_cache_flags(void)
982 {
983 	return SD_SHARE_PKG_RESOURCES;
984 }
985 
986 /*
987  * We can't just pass cpu_l2_cache_mask() directly because
988  * returns a non-const pointer and the compiler barfs on that.
989  */
990 static const struct cpumask *shared_cache_mask(int cpu)
991 {
992 	return per_cpu(cpu_l2_cache_map, cpu);
993 }
994 
995 #ifdef CONFIG_SCHED_SMT
996 static const struct cpumask *smallcore_smt_mask(int cpu)
997 {
998 	return cpu_smallcore_mask(cpu);
999 }
1000 #endif
1001 
1002 static struct cpumask *cpu_coregroup_mask(int cpu)
1003 {
1004 	return per_cpu(cpu_coregroup_map, cpu);
1005 }
1006 
1007 static bool has_coregroup_support(void)
1008 {
1009 	return coregroup_enabled;
1010 }
1011 
1012 static const struct cpumask *cpu_mc_mask(int cpu)
1013 {
1014 	return cpu_coregroup_mask(cpu);
1015 }
1016 
1017 static struct sched_domain_topology_level powerpc_topology[] = {
1018 #ifdef CONFIG_SCHED_SMT
1019 	{ cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
1020 #endif
1021 	{ shared_cache_mask, powerpc_shared_cache_flags, SD_INIT_NAME(CACHE) },
1022 	{ cpu_mc_mask, SD_INIT_NAME(MC) },
1023 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
1024 	{ NULL, },
1025 };
1026 
1027 static int __init init_big_cores(void)
1028 {
1029 	int cpu;
1030 
1031 	for_each_possible_cpu(cpu) {
1032 		int err = init_thread_group_cache_map(cpu, THREAD_GROUP_SHARE_L1);
1033 
1034 		if (err)
1035 			return err;
1036 
1037 		zalloc_cpumask_var_node(&per_cpu(cpu_smallcore_map, cpu),
1038 					GFP_KERNEL,
1039 					cpu_to_node(cpu));
1040 	}
1041 
1042 	has_big_cores = true;
1043 
1044 	for_each_possible_cpu(cpu) {
1045 		int err = init_thread_group_cache_map(cpu, THREAD_GROUP_SHARE_L2_L3);
1046 
1047 		if (err)
1048 			return err;
1049 	}
1050 
1051 	thread_group_shares_l2 = true;
1052 	thread_group_shares_l3 = true;
1053 	pr_debug("L2/L3 cache only shared by the threads in the small core\n");
1054 
1055 	return 0;
1056 }
1057 
1058 void __init smp_prepare_cpus(unsigned int max_cpus)
1059 {
1060 	unsigned int cpu;
1061 
1062 	DBG("smp_prepare_cpus\n");
1063 
1064 	/*
1065 	 * setup_cpu may need to be called on the boot cpu. We havent
1066 	 * spun any cpus up but lets be paranoid.
1067 	 */
1068 	BUG_ON(boot_cpuid != smp_processor_id());
1069 
1070 	/* Fixup boot cpu */
1071 	smp_store_cpu_info(boot_cpuid);
1072 	cpu_callin_map[boot_cpuid] = 1;
1073 
1074 	for_each_possible_cpu(cpu) {
1075 		zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
1076 					GFP_KERNEL, cpu_to_node(cpu));
1077 		zalloc_cpumask_var_node(&per_cpu(cpu_l2_cache_map, cpu),
1078 					GFP_KERNEL, cpu_to_node(cpu));
1079 		zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
1080 					GFP_KERNEL, cpu_to_node(cpu));
1081 		if (has_coregroup_support())
1082 			zalloc_cpumask_var_node(&per_cpu(cpu_coregroup_map, cpu),
1083 						GFP_KERNEL, cpu_to_node(cpu));
1084 
1085 #ifdef CONFIG_NUMA
1086 		/*
1087 		 * numa_node_id() works after this.
1088 		 */
1089 		if (cpu_present(cpu)) {
1090 			set_cpu_numa_node(cpu, numa_cpu_lookup_table[cpu]);
1091 			set_cpu_numa_mem(cpu,
1092 				local_memory_node(numa_cpu_lookup_table[cpu]));
1093 		}
1094 #endif
1095 	}
1096 
1097 	/* Init the cpumasks so the boot CPU is related to itself */
1098 	cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
1099 	cpumask_set_cpu(boot_cpuid, cpu_l2_cache_mask(boot_cpuid));
1100 	cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
1101 
1102 	if (has_coregroup_support())
1103 		cpumask_set_cpu(boot_cpuid, cpu_coregroup_mask(boot_cpuid));
1104 
1105 	init_big_cores();
1106 	if (has_big_cores) {
1107 		cpumask_set_cpu(boot_cpuid,
1108 				cpu_smallcore_mask(boot_cpuid));
1109 	}
1110 
1111 	if (cpu_to_chip_id(boot_cpuid) != -1) {
1112 		int idx = DIV_ROUND_UP(num_possible_cpus(), threads_per_core);
1113 
1114 		/*
1115 		 * All threads of a core will all belong to the same core,
1116 		 * chip_id_lookup_table will have one entry per core.
1117 		 * Assumption: if boot_cpuid doesn't have a chip-id, then no
1118 		 * other CPUs, will also not have chip-id.
1119 		 */
1120 		chip_id_lookup_table = kcalloc(idx, sizeof(int), GFP_KERNEL);
1121 		if (chip_id_lookup_table)
1122 			memset(chip_id_lookup_table, -1, sizeof(int) * idx);
1123 	}
1124 
1125 	if (smp_ops && smp_ops->probe)
1126 		smp_ops->probe();
1127 }
1128 
1129 void smp_prepare_boot_cpu(void)
1130 {
1131 	BUG_ON(smp_processor_id() != boot_cpuid);
1132 #ifdef CONFIG_PPC64
1133 	paca_ptrs[boot_cpuid]->__current = current;
1134 #endif
1135 	set_numa_node(numa_cpu_lookup_table[boot_cpuid]);
1136 	current_set[boot_cpuid] = current;
1137 }
1138 
1139 #ifdef CONFIG_HOTPLUG_CPU
1140 
1141 int generic_cpu_disable(void)
1142 {
1143 	unsigned int cpu = smp_processor_id();
1144 
1145 	if (cpu == boot_cpuid)
1146 		return -EBUSY;
1147 
1148 	set_cpu_online(cpu, false);
1149 #ifdef CONFIG_PPC64
1150 	vdso_data->processorCount--;
1151 #endif
1152 	/* Update affinity of all IRQs previously aimed at this CPU */
1153 	irq_migrate_all_off_this_cpu();
1154 
1155 	/*
1156 	 * Depending on the details of the interrupt controller, it's possible
1157 	 * that one of the interrupts we just migrated away from this CPU is
1158 	 * actually already pending on this CPU. If we leave it in that state
1159 	 * the interrupt will never be EOI'ed, and will never fire again. So
1160 	 * temporarily enable interrupts here, to allow any pending interrupt to
1161 	 * be received (and EOI'ed), before we take this CPU offline.
1162 	 */
1163 	local_irq_enable();
1164 	mdelay(1);
1165 	local_irq_disable();
1166 
1167 	return 0;
1168 }
1169 
1170 void generic_cpu_die(unsigned int cpu)
1171 {
1172 	int i;
1173 
1174 	for (i = 0; i < 100; i++) {
1175 		smp_rmb();
1176 		if (is_cpu_dead(cpu))
1177 			return;
1178 		msleep(100);
1179 	}
1180 	printk(KERN_ERR "CPU%d didn't die...\n", cpu);
1181 }
1182 
1183 void generic_set_cpu_dead(unsigned int cpu)
1184 {
1185 	per_cpu(cpu_state, cpu) = CPU_DEAD;
1186 }
1187 
1188 /*
1189  * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
1190  * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
1191  * which makes the delay in generic_cpu_die() not happen.
1192  */
1193 void generic_set_cpu_up(unsigned int cpu)
1194 {
1195 	per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1196 }
1197 
1198 int generic_check_cpu_restart(unsigned int cpu)
1199 {
1200 	return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
1201 }
1202 
1203 int is_cpu_dead(unsigned int cpu)
1204 {
1205 	return per_cpu(cpu_state, cpu) == CPU_DEAD;
1206 }
1207 
1208 static bool secondaries_inhibited(void)
1209 {
1210 	return kvm_hv_mode_active();
1211 }
1212 
1213 #else /* HOTPLUG_CPU */
1214 
1215 #define secondaries_inhibited()		0
1216 
1217 #endif
1218 
1219 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
1220 {
1221 #ifdef CONFIG_PPC64
1222 	paca_ptrs[cpu]->__current = idle;
1223 	paca_ptrs[cpu]->kstack = (unsigned long)task_stack_page(idle) +
1224 				 THREAD_SIZE - STACK_FRAME_OVERHEAD;
1225 #endif
1226 	idle->cpu = cpu;
1227 	secondary_current = current_set[cpu] = idle;
1228 }
1229 
1230 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
1231 {
1232 	int rc, c;
1233 
1234 	/*
1235 	 * Don't allow secondary threads to come online if inhibited
1236 	 */
1237 	if (threads_per_core > 1 && secondaries_inhibited() &&
1238 	    cpu_thread_in_subcore(cpu))
1239 		return -EBUSY;
1240 
1241 	if (smp_ops == NULL ||
1242 	    (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
1243 		return -EINVAL;
1244 
1245 	cpu_idle_thread_init(cpu, tidle);
1246 
1247 	/*
1248 	 * The platform might need to allocate resources prior to bringing
1249 	 * up the CPU
1250 	 */
1251 	if (smp_ops->prepare_cpu) {
1252 		rc = smp_ops->prepare_cpu(cpu);
1253 		if (rc)
1254 			return rc;
1255 	}
1256 
1257 	/* Make sure callin-map entry is 0 (can be leftover a CPU
1258 	 * hotplug
1259 	 */
1260 	cpu_callin_map[cpu] = 0;
1261 
1262 	/* The information for processor bringup must
1263 	 * be written out to main store before we release
1264 	 * the processor.
1265 	 */
1266 	smp_mb();
1267 
1268 	/* wake up cpus */
1269 	DBG("smp: kicking cpu %d\n", cpu);
1270 	rc = smp_ops->kick_cpu(cpu);
1271 	if (rc) {
1272 		pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
1273 		return rc;
1274 	}
1275 
1276 	/*
1277 	 * wait to see if the cpu made a callin (is actually up).
1278 	 * use this value that I found through experimentation.
1279 	 * -- Cort
1280 	 */
1281 	if (system_state < SYSTEM_RUNNING)
1282 		for (c = 50000; c && !cpu_callin_map[cpu]; c--)
1283 			udelay(100);
1284 #ifdef CONFIG_HOTPLUG_CPU
1285 	else
1286 		/*
1287 		 * CPUs can take much longer to come up in the
1288 		 * hotplug case.  Wait five seconds.
1289 		 */
1290 		for (c = 5000; c && !cpu_callin_map[cpu]; c--)
1291 			msleep(1);
1292 #endif
1293 
1294 	if (!cpu_callin_map[cpu]) {
1295 		printk(KERN_ERR "Processor %u is stuck.\n", cpu);
1296 		return -ENOENT;
1297 	}
1298 
1299 	DBG("Processor %u found.\n", cpu);
1300 
1301 	if (smp_ops->give_timebase)
1302 		smp_ops->give_timebase();
1303 
1304 	/* Wait until cpu puts itself in the online & active maps */
1305 	spin_until_cond(cpu_online(cpu));
1306 
1307 	return 0;
1308 }
1309 
1310 /* Return the value of the reg property corresponding to the given
1311  * logical cpu.
1312  */
1313 int cpu_to_core_id(int cpu)
1314 {
1315 	struct device_node *np;
1316 	const __be32 *reg;
1317 	int id = -1;
1318 
1319 	np = of_get_cpu_node(cpu, NULL);
1320 	if (!np)
1321 		goto out;
1322 
1323 	reg = of_get_property(np, "reg", NULL);
1324 	if (!reg)
1325 		goto out;
1326 
1327 	id = be32_to_cpup(reg);
1328 out:
1329 	of_node_put(np);
1330 	return id;
1331 }
1332 EXPORT_SYMBOL_GPL(cpu_to_core_id);
1333 
1334 /* Helper routines for cpu to core mapping */
1335 int cpu_core_index_of_thread(int cpu)
1336 {
1337 	return cpu >> threads_shift;
1338 }
1339 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
1340 
1341 int cpu_first_thread_of_core(int core)
1342 {
1343 	return core << threads_shift;
1344 }
1345 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
1346 
1347 /* Must be called when no change can occur to cpu_present_mask,
1348  * i.e. during cpu online or offline.
1349  */
1350 static struct device_node *cpu_to_l2cache(int cpu)
1351 {
1352 	struct device_node *np;
1353 	struct device_node *cache;
1354 
1355 	if (!cpu_present(cpu))
1356 		return NULL;
1357 
1358 	np = of_get_cpu_node(cpu, NULL);
1359 	if (np == NULL)
1360 		return NULL;
1361 
1362 	cache = of_find_next_cache_node(np);
1363 
1364 	of_node_put(np);
1365 
1366 	return cache;
1367 }
1368 
1369 static bool update_mask_by_l2(int cpu, cpumask_var_t *mask)
1370 {
1371 	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1372 	struct device_node *l2_cache, *np;
1373 	int i;
1374 
1375 	if (has_big_cores)
1376 		submask_fn = cpu_smallcore_mask;
1377 
1378 	/*
1379 	 * If the threads in a thread-group share L2 cache, then the
1380 	 * L2-mask can be obtained from thread_group_l2_cache_map.
1381 	 */
1382 	if (thread_group_shares_l2) {
1383 		cpumask_set_cpu(cpu, cpu_l2_cache_mask(cpu));
1384 
1385 		for_each_cpu(i, per_cpu(thread_group_l2_cache_map, cpu)) {
1386 			if (cpu_online(i))
1387 				set_cpus_related(i, cpu, cpu_l2_cache_mask);
1388 		}
1389 
1390 		/* Verify that L1-cache siblings are a subset of L2 cache-siblings */
1391 		if (!cpumask_equal(submask_fn(cpu), cpu_l2_cache_mask(cpu)) &&
1392 		    !cpumask_subset(submask_fn(cpu), cpu_l2_cache_mask(cpu))) {
1393 			pr_warn_once("CPU %d : Inconsistent L1 and L2 cache siblings\n",
1394 				     cpu);
1395 		}
1396 
1397 		return true;
1398 	}
1399 
1400 	l2_cache = cpu_to_l2cache(cpu);
1401 	if (!l2_cache || !*mask) {
1402 		/* Assume only core siblings share cache with this CPU */
1403 		for_each_cpu(i, cpu_sibling_mask(cpu))
1404 			set_cpus_related(cpu, i, cpu_l2_cache_mask);
1405 
1406 		return false;
1407 	}
1408 
1409 	cpumask_and(*mask, cpu_online_mask, cpu_cpu_mask(cpu));
1410 
1411 	/* Update l2-cache mask with all the CPUs that are part of submask */
1412 	or_cpumasks_related(cpu, cpu, submask_fn, cpu_l2_cache_mask);
1413 
1414 	/* Skip all CPUs already part of current CPU l2-cache mask */
1415 	cpumask_andnot(*mask, *mask, cpu_l2_cache_mask(cpu));
1416 
1417 	for_each_cpu(i, *mask) {
1418 		/*
1419 		 * when updating the marks the current CPU has not been marked
1420 		 * online, but we need to update the cache masks
1421 		 */
1422 		np = cpu_to_l2cache(i);
1423 
1424 		/* Skip all CPUs already part of current CPU l2-cache */
1425 		if (np == l2_cache) {
1426 			or_cpumasks_related(cpu, i, submask_fn, cpu_l2_cache_mask);
1427 			cpumask_andnot(*mask, *mask, submask_fn(i));
1428 		} else {
1429 			cpumask_andnot(*mask, *mask, cpu_l2_cache_mask(i));
1430 		}
1431 
1432 		of_node_put(np);
1433 	}
1434 	of_node_put(l2_cache);
1435 
1436 	return true;
1437 }
1438 
1439 #ifdef CONFIG_HOTPLUG_CPU
1440 static void remove_cpu_from_masks(int cpu)
1441 {
1442 	struct cpumask *(*mask_fn)(int) = cpu_sibling_mask;
1443 	int i;
1444 
1445 	unmap_cpu_from_node(cpu);
1446 
1447 	if (shared_caches)
1448 		mask_fn = cpu_l2_cache_mask;
1449 
1450 	for_each_cpu(i, mask_fn(cpu)) {
1451 		set_cpus_unrelated(cpu, i, cpu_l2_cache_mask);
1452 		set_cpus_unrelated(cpu, i, cpu_sibling_mask);
1453 		if (has_big_cores)
1454 			set_cpus_unrelated(cpu, i, cpu_smallcore_mask);
1455 	}
1456 
1457 	for_each_cpu(i, cpu_core_mask(cpu))
1458 		set_cpus_unrelated(cpu, i, cpu_core_mask);
1459 
1460 	if (has_coregroup_support()) {
1461 		for_each_cpu(i, cpu_coregroup_mask(cpu))
1462 			set_cpus_unrelated(cpu, i, cpu_coregroup_mask);
1463 	}
1464 }
1465 #endif
1466 
1467 static inline void add_cpu_to_smallcore_masks(int cpu)
1468 {
1469 	int i;
1470 
1471 	if (!has_big_cores)
1472 		return;
1473 
1474 	cpumask_set_cpu(cpu, cpu_smallcore_mask(cpu));
1475 
1476 	for_each_cpu(i, per_cpu(thread_group_l1_cache_map, cpu)) {
1477 		if (cpu_online(i))
1478 			set_cpus_related(i, cpu, cpu_smallcore_mask);
1479 	}
1480 }
1481 
1482 static void update_coregroup_mask(int cpu, cpumask_var_t *mask)
1483 {
1484 	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1485 	int coregroup_id = cpu_to_coregroup_id(cpu);
1486 	int i;
1487 
1488 	if (shared_caches)
1489 		submask_fn = cpu_l2_cache_mask;
1490 
1491 	if (!*mask) {
1492 		/* Assume only siblings are part of this CPU's coregroup */
1493 		for_each_cpu(i, submask_fn(cpu))
1494 			set_cpus_related(cpu, i, cpu_coregroup_mask);
1495 
1496 		return;
1497 	}
1498 
1499 	cpumask_and(*mask, cpu_online_mask, cpu_cpu_mask(cpu));
1500 
1501 	/* Update coregroup mask with all the CPUs that are part of submask */
1502 	or_cpumasks_related(cpu, cpu, submask_fn, cpu_coregroup_mask);
1503 
1504 	/* Skip all CPUs already part of coregroup mask */
1505 	cpumask_andnot(*mask, *mask, cpu_coregroup_mask(cpu));
1506 
1507 	for_each_cpu(i, *mask) {
1508 		/* Skip all CPUs not part of this coregroup */
1509 		if (coregroup_id == cpu_to_coregroup_id(i)) {
1510 			or_cpumasks_related(cpu, i, submask_fn, cpu_coregroup_mask);
1511 			cpumask_andnot(*mask, *mask, submask_fn(i));
1512 		} else {
1513 			cpumask_andnot(*mask, *mask, cpu_coregroup_mask(i));
1514 		}
1515 	}
1516 }
1517 
1518 static void add_cpu_to_masks(int cpu)
1519 {
1520 	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1521 	int first_thread = cpu_first_thread_sibling(cpu);
1522 	cpumask_var_t mask;
1523 	int chip_id = -1;
1524 	bool ret;
1525 	int i;
1526 
1527 	/*
1528 	 * This CPU will not be in the online mask yet so we need to manually
1529 	 * add it to it's own thread sibling mask.
1530 	 */
1531 	map_cpu_to_node(cpu, cpu_to_node(cpu));
1532 	cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
1533 	cpumask_set_cpu(cpu, cpu_core_mask(cpu));
1534 
1535 	for (i = first_thread; i < first_thread + threads_per_core; i++)
1536 		if (cpu_online(i))
1537 			set_cpus_related(i, cpu, cpu_sibling_mask);
1538 
1539 	add_cpu_to_smallcore_masks(cpu);
1540 
1541 	/* In CPU-hotplug path, hence use GFP_ATOMIC */
1542 	ret = alloc_cpumask_var_node(&mask, GFP_ATOMIC, cpu_to_node(cpu));
1543 	update_mask_by_l2(cpu, &mask);
1544 
1545 	if (has_coregroup_support())
1546 		update_coregroup_mask(cpu, &mask);
1547 
1548 	if (chip_id_lookup_table && ret)
1549 		chip_id = cpu_to_chip_id(cpu);
1550 
1551 	if (shared_caches)
1552 		submask_fn = cpu_l2_cache_mask;
1553 
1554 	/* Update core_mask with all the CPUs that are part of submask */
1555 	or_cpumasks_related(cpu, cpu, submask_fn, cpu_core_mask);
1556 
1557 	/* Skip all CPUs already part of current CPU core mask */
1558 	cpumask_andnot(mask, cpu_online_mask, cpu_core_mask(cpu));
1559 
1560 	/* If chip_id is -1; limit the cpu_core_mask to within DIE*/
1561 	if (chip_id == -1)
1562 		cpumask_and(mask, mask, cpu_cpu_mask(cpu));
1563 
1564 	for_each_cpu(i, mask) {
1565 		if (chip_id == cpu_to_chip_id(i)) {
1566 			or_cpumasks_related(cpu, i, submask_fn, cpu_core_mask);
1567 			cpumask_andnot(mask, mask, submask_fn(i));
1568 		} else {
1569 			cpumask_andnot(mask, mask, cpu_core_mask(i));
1570 		}
1571 	}
1572 
1573 	free_cpumask_var(mask);
1574 }
1575 
1576 /* Activate a secondary processor. */
1577 void start_secondary(void *unused)
1578 {
1579 	unsigned int cpu = raw_smp_processor_id();
1580 
1581 	/* PPC64 calls setup_kup() in early_setup_secondary() */
1582 	if (IS_ENABLED(CONFIG_PPC32))
1583 		setup_kup();
1584 
1585 	mmgrab(&init_mm);
1586 	current->active_mm = &init_mm;
1587 
1588 	smp_store_cpu_info(cpu);
1589 	set_dec(tb_ticks_per_jiffy);
1590 	rcu_cpu_starting(cpu);
1591 	cpu_callin_map[cpu] = 1;
1592 
1593 	if (smp_ops->setup_cpu)
1594 		smp_ops->setup_cpu(cpu);
1595 	if (smp_ops->take_timebase)
1596 		smp_ops->take_timebase();
1597 
1598 	secondary_cpu_time_init();
1599 
1600 #ifdef CONFIG_PPC64
1601 	if (system_state == SYSTEM_RUNNING)
1602 		vdso_data->processorCount++;
1603 
1604 	vdso_getcpu_init();
1605 #endif
1606 	set_numa_node(numa_cpu_lookup_table[cpu]);
1607 	set_numa_mem(local_memory_node(numa_cpu_lookup_table[cpu]));
1608 
1609 	/* Update topology CPU masks */
1610 	add_cpu_to_masks(cpu);
1611 
1612 	/*
1613 	 * Check for any shared caches. Note that this must be done on a
1614 	 * per-core basis because one core in the pair might be disabled.
1615 	 */
1616 	if (!shared_caches) {
1617 		struct cpumask *(*sibling_mask)(int) = cpu_sibling_mask;
1618 		struct cpumask *mask = cpu_l2_cache_mask(cpu);
1619 
1620 		if (has_big_cores)
1621 			sibling_mask = cpu_smallcore_mask;
1622 
1623 		if (cpumask_weight(mask) > cpumask_weight(sibling_mask(cpu)))
1624 			shared_caches = true;
1625 	}
1626 
1627 	smp_wmb();
1628 	notify_cpu_starting(cpu);
1629 	set_cpu_online(cpu, true);
1630 
1631 	boot_init_stack_canary();
1632 
1633 	local_irq_enable();
1634 
1635 	/* We can enable ftrace for secondary cpus now */
1636 	this_cpu_enable_ftrace();
1637 
1638 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
1639 
1640 	BUG();
1641 }
1642 
1643 int setup_profiling_timer(unsigned int multiplier)
1644 {
1645 	return 0;
1646 }
1647 
1648 static void fixup_topology(void)
1649 {
1650 	int i;
1651 
1652 #ifdef CONFIG_SCHED_SMT
1653 	if (has_big_cores) {
1654 		pr_info("Big cores detected but using small core scheduling\n");
1655 		powerpc_topology[smt_idx].mask = smallcore_smt_mask;
1656 	}
1657 #endif
1658 
1659 	if (!has_coregroup_support())
1660 		powerpc_topology[mc_idx].mask = powerpc_topology[cache_idx].mask;
1661 
1662 	/*
1663 	 * Try to consolidate topology levels here instead of
1664 	 * allowing scheduler to degenerate.
1665 	 * - Dont consolidate if masks are different.
1666 	 * - Dont consolidate if sd_flags exists and are different.
1667 	 */
1668 	for (i = 1; i <= die_idx; i++) {
1669 		if (powerpc_topology[i].mask != powerpc_topology[i - 1].mask)
1670 			continue;
1671 
1672 		if (powerpc_topology[i].sd_flags && powerpc_topology[i - 1].sd_flags &&
1673 				powerpc_topology[i].sd_flags != powerpc_topology[i - 1].sd_flags)
1674 			continue;
1675 
1676 		if (!powerpc_topology[i - 1].sd_flags)
1677 			powerpc_topology[i - 1].sd_flags = powerpc_topology[i].sd_flags;
1678 
1679 		powerpc_topology[i].mask = powerpc_topology[i + 1].mask;
1680 		powerpc_topology[i].sd_flags = powerpc_topology[i + 1].sd_flags;
1681 #ifdef CONFIG_SCHED_DEBUG
1682 		powerpc_topology[i].name = powerpc_topology[i + 1].name;
1683 #endif
1684 	}
1685 }
1686 
1687 void __init smp_cpus_done(unsigned int max_cpus)
1688 {
1689 	/*
1690 	 * We are running pinned to the boot CPU, see rest_init().
1691 	 */
1692 	if (smp_ops && smp_ops->setup_cpu)
1693 		smp_ops->setup_cpu(boot_cpuid);
1694 
1695 	if (smp_ops && smp_ops->bringup_done)
1696 		smp_ops->bringup_done();
1697 
1698 	dump_numa_cpu_topology();
1699 
1700 	fixup_topology();
1701 	set_sched_topology(powerpc_topology);
1702 }
1703 
1704 #ifdef CONFIG_HOTPLUG_CPU
1705 int __cpu_disable(void)
1706 {
1707 	int cpu = smp_processor_id();
1708 	int err;
1709 
1710 	if (!smp_ops->cpu_disable)
1711 		return -ENOSYS;
1712 
1713 	this_cpu_disable_ftrace();
1714 
1715 	err = smp_ops->cpu_disable();
1716 	if (err)
1717 		return err;
1718 
1719 	/* Update sibling maps */
1720 	remove_cpu_from_masks(cpu);
1721 
1722 	return 0;
1723 }
1724 
1725 void __cpu_die(unsigned int cpu)
1726 {
1727 	if (smp_ops->cpu_die)
1728 		smp_ops->cpu_die(cpu);
1729 }
1730 
1731 void arch_cpu_idle_dead(void)
1732 {
1733 	sched_preempt_enable_no_resched();
1734 
1735 	/*
1736 	 * Disable on the down path. This will be re-enabled by
1737 	 * start_secondary() via start_secondary_resume() below
1738 	 */
1739 	this_cpu_disable_ftrace();
1740 
1741 	if (smp_ops->cpu_offline_self)
1742 		smp_ops->cpu_offline_self();
1743 
1744 	/* If we return, we re-enter start_secondary */
1745 	start_secondary_resume();
1746 }
1747 
1748 #endif
1749