xref: /openbmc/linux/arch/powerpc/kernel/smp.c (revision 11a163f2)
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 
38 #include <asm/ptrace.h>
39 #include <linux/atomic.h>
40 #include <asm/irq.h>
41 #include <asm/hw_irq.h>
42 #include <asm/kvm_ppc.h>
43 #include <asm/dbell.h>
44 #include <asm/page.h>
45 #include <asm/prom.h>
46 #include <asm/smp.h>
47 #include <asm/time.h>
48 #include <asm/machdep.h>
49 #include <asm/cputhreads.h>
50 #include <asm/cputable.h>
51 #include <asm/mpic.h>
52 #include <asm/vdso_datapage.h>
53 #ifdef CONFIG_PPC64
54 #include <asm/paca.h>
55 #endif
56 #include <asm/vdso.h>
57 #include <asm/debug.h>
58 #include <asm/kexec.h>
59 #include <asm/asm-prototypes.h>
60 #include <asm/cpu_has_feature.h>
61 #include <asm/ftrace.h>
62 #include <asm/kup.h>
63 
64 #ifdef DEBUG
65 #include <asm/udbg.h>
66 #define DBG(fmt...) udbg_printf(fmt)
67 #else
68 #define DBG(fmt...)
69 #endif
70 
71 #ifdef CONFIG_HOTPLUG_CPU
72 /* State of each CPU during hotplug phases */
73 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
74 #endif
75 
76 struct task_struct *secondary_current;
77 bool has_big_cores;
78 bool coregroup_enabled;
79 
80 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
81 DEFINE_PER_CPU(cpumask_var_t, cpu_smallcore_map);
82 DEFINE_PER_CPU(cpumask_var_t, cpu_l2_cache_map);
83 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
84 DEFINE_PER_CPU(cpumask_var_t, cpu_coregroup_map);
85 
86 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
87 EXPORT_PER_CPU_SYMBOL(cpu_l2_cache_map);
88 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
89 EXPORT_SYMBOL_GPL(has_big_cores);
90 
91 enum {
92 #ifdef CONFIG_SCHED_SMT
93 	smt_idx,
94 #endif
95 	cache_idx,
96 	mc_idx,
97 	die_idx,
98 };
99 
100 #define MAX_THREAD_LIST_SIZE	8
101 #define THREAD_GROUP_SHARE_L1   1
102 struct thread_groups {
103 	unsigned int property;
104 	unsigned int nr_groups;
105 	unsigned int threads_per_group;
106 	unsigned int thread_list[MAX_THREAD_LIST_SIZE];
107 };
108 
109 /*
110  * On big-cores system, cpu_l1_cache_map for each CPU corresponds to
111  * the set its siblings that share the L1-cache.
112  */
113 DEFINE_PER_CPU(cpumask_var_t, cpu_l1_cache_map);
114 
115 /* SMP operations for this machine */
116 struct smp_ops_t *smp_ops;
117 
118 /* Can't be static due to PowerMac hackery */
119 volatile unsigned int cpu_callin_map[NR_CPUS];
120 
121 int smt_enabled_at_boot = 1;
122 
123 /*
124  * Returns 1 if the specified cpu should be brought up during boot.
125  * Used to inhibit booting threads if they've been disabled or
126  * limited on the command line
127  */
128 int smp_generic_cpu_bootable(unsigned int nr)
129 {
130 	/* Special case - we inhibit secondary thread startup
131 	 * during boot if the user requests it.
132 	 */
133 	if (system_state < SYSTEM_RUNNING && cpu_has_feature(CPU_FTR_SMT)) {
134 		if (!smt_enabled_at_boot && cpu_thread_in_core(nr) != 0)
135 			return 0;
136 		if (smt_enabled_at_boot
137 		    && cpu_thread_in_core(nr) >= smt_enabled_at_boot)
138 			return 0;
139 	}
140 
141 	return 1;
142 }
143 
144 
145 #ifdef CONFIG_PPC64
146 int smp_generic_kick_cpu(int nr)
147 {
148 	if (nr < 0 || nr >= nr_cpu_ids)
149 		return -EINVAL;
150 
151 	/*
152 	 * The processor is currently spinning, waiting for the
153 	 * cpu_start field to become non-zero After we set cpu_start,
154 	 * the processor will continue on to secondary_start
155 	 */
156 	if (!paca_ptrs[nr]->cpu_start) {
157 		paca_ptrs[nr]->cpu_start = 1;
158 		smp_mb();
159 		return 0;
160 	}
161 
162 #ifdef CONFIG_HOTPLUG_CPU
163 	/*
164 	 * Ok it's not there, so it might be soft-unplugged, let's
165 	 * try to bring it back
166 	 */
167 	generic_set_cpu_up(nr);
168 	smp_wmb();
169 	smp_send_reschedule(nr);
170 #endif /* CONFIG_HOTPLUG_CPU */
171 
172 	return 0;
173 }
174 #endif /* CONFIG_PPC64 */
175 
176 static irqreturn_t call_function_action(int irq, void *data)
177 {
178 	generic_smp_call_function_interrupt();
179 	return IRQ_HANDLED;
180 }
181 
182 static irqreturn_t reschedule_action(int irq, void *data)
183 {
184 	scheduler_ipi();
185 	return IRQ_HANDLED;
186 }
187 
188 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
189 static irqreturn_t tick_broadcast_ipi_action(int irq, void *data)
190 {
191 	timer_broadcast_interrupt();
192 	return IRQ_HANDLED;
193 }
194 #endif
195 
196 #ifdef CONFIG_NMI_IPI
197 static irqreturn_t nmi_ipi_action(int irq, void *data)
198 {
199 	smp_handle_nmi_ipi(get_irq_regs());
200 	return IRQ_HANDLED;
201 }
202 #endif
203 
204 static irq_handler_t smp_ipi_action[] = {
205 	[PPC_MSG_CALL_FUNCTION] =  call_function_action,
206 	[PPC_MSG_RESCHEDULE] = reschedule_action,
207 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
208 	[PPC_MSG_TICK_BROADCAST] = tick_broadcast_ipi_action,
209 #endif
210 #ifdef CONFIG_NMI_IPI
211 	[PPC_MSG_NMI_IPI] = nmi_ipi_action,
212 #endif
213 };
214 
215 /*
216  * The NMI IPI is a fallback and not truly non-maskable. It is simpler
217  * than going through the call function infrastructure, and strongly
218  * serialized, so it is more appropriate for debugging.
219  */
220 const char *smp_ipi_name[] = {
221 	[PPC_MSG_CALL_FUNCTION] =  "ipi call function",
222 	[PPC_MSG_RESCHEDULE] = "ipi reschedule",
223 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
224 	[PPC_MSG_TICK_BROADCAST] = "ipi tick-broadcast",
225 #endif
226 #ifdef CONFIG_NMI_IPI
227 	[PPC_MSG_NMI_IPI] = "nmi ipi",
228 #endif
229 };
230 
231 /* optional function to request ipi, for controllers with >= 4 ipis */
232 int smp_request_message_ipi(int virq, int msg)
233 {
234 	int err;
235 
236 	if (msg < 0 || msg > PPC_MSG_NMI_IPI)
237 		return -EINVAL;
238 #ifndef CONFIG_NMI_IPI
239 	if (msg == PPC_MSG_NMI_IPI)
240 		return 1;
241 #endif
242 
243 	err = request_irq(virq, smp_ipi_action[msg],
244 			  IRQF_PERCPU | IRQF_NO_THREAD | IRQF_NO_SUSPEND,
245 			  smp_ipi_name[msg], NULL);
246 	WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
247 		virq, smp_ipi_name[msg], err);
248 
249 	return err;
250 }
251 
252 #ifdef CONFIG_PPC_SMP_MUXED_IPI
253 struct cpu_messages {
254 	long messages;			/* current messages */
255 };
256 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
257 
258 void smp_muxed_ipi_set_message(int cpu, int msg)
259 {
260 	struct cpu_messages *info = &per_cpu(ipi_message, cpu);
261 	char *message = (char *)&info->messages;
262 
263 	/*
264 	 * Order previous accesses before accesses in the IPI handler.
265 	 */
266 	smp_mb();
267 	message[msg] = 1;
268 }
269 
270 void smp_muxed_ipi_message_pass(int cpu, int msg)
271 {
272 	smp_muxed_ipi_set_message(cpu, msg);
273 
274 	/*
275 	 * cause_ipi functions are required to include a full barrier
276 	 * before doing whatever causes the IPI.
277 	 */
278 	smp_ops->cause_ipi(cpu);
279 }
280 
281 #ifdef __BIG_ENDIAN__
282 #define IPI_MESSAGE(A) (1uL << ((BITS_PER_LONG - 8) - 8 * (A)))
283 #else
284 #define IPI_MESSAGE(A) (1uL << (8 * (A)))
285 #endif
286 
287 irqreturn_t smp_ipi_demux(void)
288 {
289 	mb();	/* order any irq clear */
290 
291 	return smp_ipi_demux_relaxed();
292 }
293 
294 /* sync-free variant. Callers should ensure synchronization */
295 irqreturn_t smp_ipi_demux_relaxed(void)
296 {
297 	struct cpu_messages *info;
298 	unsigned long all;
299 
300 	info = this_cpu_ptr(&ipi_message);
301 	do {
302 		all = xchg(&info->messages, 0);
303 #if defined(CONFIG_KVM_XICS) && defined(CONFIG_KVM_BOOK3S_HV_POSSIBLE)
304 		/*
305 		 * Must check for PPC_MSG_RM_HOST_ACTION messages
306 		 * before PPC_MSG_CALL_FUNCTION messages because when
307 		 * a VM is destroyed, we call kick_all_cpus_sync()
308 		 * to ensure that any pending PPC_MSG_RM_HOST_ACTION
309 		 * messages have completed before we free any VCPUs.
310 		 */
311 		if (all & IPI_MESSAGE(PPC_MSG_RM_HOST_ACTION))
312 			kvmppc_xics_ipi_action();
313 #endif
314 		if (all & IPI_MESSAGE(PPC_MSG_CALL_FUNCTION))
315 			generic_smp_call_function_interrupt();
316 		if (all & IPI_MESSAGE(PPC_MSG_RESCHEDULE))
317 			scheduler_ipi();
318 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
319 		if (all & IPI_MESSAGE(PPC_MSG_TICK_BROADCAST))
320 			timer_broadcast_interrupt();
321 #endif
322 #ifdef CONFIG_NMI_IPI
323 		if (all & IPI_MESSAGE(PPC_MSG_NMI_IPI))
324 			nmi_ipi_action(0, NULL);
325 #endif
326 	} while (info->messages);
327 
328 	return IRQ_HANDLED;
329 }
330 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
331 
332 static inline void do_message_pass(int cpu, int msg)
333 {
334 	if (smp_ops->message_pass)
335 		smp_ops->message_pass(cpu, msg);
336 #ifdef CONFIG_PPC_SMP_MUXED_IPI
337 	else
338 		smp_muxed_ipi_message_pass(cpu, msg);
339 #endif
340 }
341 
342 void smp_send_reschedule(int cpu)
343 {
344 	if (likely(smp_ops))
345 		do_message_pass(cpu, PPC_MSG_RESCHEDULE);
346 }
347 EXPORT_SYMBOL_GPL(smp_send_reschedule);
348 
349 void arch_send_call_function_single_ipi(int cpu)
350 {
351 	do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
352 }
353 
354 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
355 {
356 	unsigned int cpu;
357 
358 	for_each_cpu(cpu, mask)
359 		do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
360 }
361 
362 #ifdef CONFIG_NMI_IPI
363 
364 /*
365  * "NMI IPI" system.
366  *
367  * NMI IPIs may not be recoverable, so should not be used as ongoing part of
368  * a running system. They can be used for crash, debug, halt/reboot, etc.
369  *
370  * The IPI call waits with interrupts disabled until all targets enter the
371  * NMI handler, then returns. Subsequent IPIs can be issued before targets
372  * have returned from their handlers, so there is no guarantee about
373  * concurrency or re-entrancy.
374  *
375  * A new NMI can be issued before all targets exit the handler.
376  *
377  * The IPI call may time out without all targets entering the NMI handler.
378  * In that case, there is some logic to recover (and ignore subsequent
379  * NMI interrupts that may eventually be raised), but the platform interrupt
380  * handler may not be able to distinguish this from other exception causes,
381  * which may cause a crash.
382  */
383 
384 static atomic_t __nmi_ipi_lock = ATOMIC_INIT(0);
385 static struct cpumask nmi_ipi_pending_mask;
386 static bool nmi_ipi_busy = false;
387 static void (*nmi_ipi_function)(struct pt_regs *) = NULL;
388 
389 static void nmi_ipi_lock_start(unsigned long *flags)
390 {
391 	raw_local_irq_save(*flags);
392 	hard_irq_disable();
393 	while (atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1) {
394 		raw_local_irq_restore(*flags);
395 		spin_until_cond(atomic_read(&__nmi_ipi_lock) == 0);
396 		raw_local_irq_save(*flags);
397 		hard_irq_disable();
398 	}
399 }
400 
401 static void nmi_ipi_lock(void)
402 {
403 	while (atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1)
404 		spin_until_cond(atomic_read(&__nmi_ipi_lock) == 0);
405 }
406 
407 static void nmi_ipi_unlock(void)
408 {
409 	smp_mb();
410 	WARN_ON(atomic_read(&__nmi_ipi_lock) != 1);
411 	atomic_set(&__nmi_ipi_lock, 0);
412 }
413 
414 static void nmi_ipi_unlock_end(unsigned long *flags)
415 {
416 	nmi_ipi_unlock();
417 	raw_local_irq_restore(*flags);
418 }
419 
420 /*
421  * Platform NMI handler calls this to ack
422  */
423 int smp_handle_nmi_ipi(struct pt_regs *regs)
424 {
425 	void (*fn)(struct pt_regs *) = NULL;
426 	unsigned long flags;
427 	int me = raw_smp_processor_id();
428 	int ret = 0;
429 
430 	/*
431 	 * Unexpected NMIs are possible here because the interrupt may not
432 	 * be able to distinguish NMI IPIs from other types of NMIs, or
433 	 * because the caller may have timed out.
434 	 */
435 	nmi_ipi_lock_start(&flags);
436 	if (cpumask_test_cpu(me, &nmi_ipi_pending_mask)) {
437 		cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
438 		fn = READ_ONCE(nmi_ipi_function);
439 		WARN_ON_ONCE(!fn);
440 		ret = 1;
441 	}
442 	nmi_ipi_unlock_end(&flags);
443 
444 	if (fn)
445 		fn(regs);
446 
447 	return ret;
448 }
449 
450 static void do_smp_send_nmi_ipi(int cpu, bool safe)
451 {
452 	if (!safe && smp_ops->cause_nmi_ipi && smp_ops->cause_nmi_ipi(cpu))
453 		return;
454 
455 	if (cpu >= 0) {
456 		do_message_pass(cpu, PPC_MSG_NMI_IPI);
457 	} else {
458 		int c;
459 
460 		for_each_online_cpu(c) {
461 			if (c == raw_smp_processor_id())
462 				continue;
463 			do_message_pass(c, PPC_MSG_NMI_IPI);
464 		}
465 	}
466 }
467 
468 /*
469  * - cpu is the target CPU (must not be this CPU), or NMI_IPI_ALL_OTHERS.
470  * - fn is the target callback function.
471  * - delay_us > 0 is the delay before giving up waiting for targets to
472  *   begin executing the handler, == 0 specifies indefinite delay.
473  */
474 static int __smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *),
475 				u64 delay_us, bool safe)
476 {
477 	unsigned long flags;
478 	int me = raw_smp_processor_id();
479 	int ret = 1;
480 
481 	BUG_ON(cpu == me);
482 	BUG_ON(cpu < 0 && cpu != NMI_IPI_ALL_OTHERS);
483 
484 	if (unlikely(!smp_ops))
485 		return 0;
486 
487 	nmi_ipi_lock_start(&flags);
488 	while (nmi_ipi_busy) {
489 		nmi_ipi_unlock_end(&flags);
490 		spin_until_cond(!nmi_ipi_busy);
491 		nmi_ipi_lock_start(&flags);
492 	}
493 	nmi_ipi_busy = true;
494 	nmi_ipi_function = fn;
495 
496 	WARN_ON_ONCE(!cpumask_empty(&nmi_ipi_pending_mask));
497 
498 	if (cpu < 0) {
499 		/* ALL_OTHERS */
500 		cpumask_copy(&nmi_ipi_pending_mask, cpu_online_mask);
501 		cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
502 	} else {
503 		cpumask_set_cpu(cpu, &nmi_ipi_pending_mask);
504 	}
505 
506 	nmi_ipi_unlock();
507 
508 	/* Interrupts remain hard disabled */
509 
510 	do_smp_send_nmi_ipi(cpu, safe);
511 
512 	nmi_ipi_lock();
513 	/* nmi_ipi_busy is set here, so unlock/lock is okay */
514 	while (!cpumask_empty(&nmi_ipi_pending_mask)) {
515 		nmi_ipi_unlock();
516 		udelay(1);
517 		nmi_ipi_lock();
518 		if (delay_us) {
519 			delay_us--;
520 			if (!delay_us)
521 				break;
522 		}
523 	}
524 
525 	if (!cpumask_empty(&nmi_ipi_pending_mask)) {
526 		/* Timeout waiting for CPUs to call smp_handle_nmi_ipi */
527 		ret = 0;
528 		cpumask_clear(&nmi_ipi_pending_mask);
529 	}
530 
531 	nmi_ipi_function = NULL;
532 	nmi_ipi_busy = false;
533 
534 	nmi_ipi_unlock_end(&flags);
535 
536 	return ret;
537 }
538 
539 int smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
540 {
541 	return __smp_send_nmi_ipi(cpu, fn, delay_us, false);
542 }
543 
544 int smp_send_safe_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
545 {
546 	return __smp_send_nmi_ipi(cpu, fn, delay_us, true);
547 }
548 #endif /* CONFIG_NMI_IPI */
549 
550 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
551 void tick_broadcast(const struct cpumask *mask)
552 {
553 	unsigned int cpu;
554 
555 	for_each_cpu(cpu, mask)
556 		do_message_pass(cpu, PPC_MSG_TICK_BROADCAST);
557 }
558 #endif
559 
560 #ifdef CONFIG_DEBUGGER
561 void debugger_ipi_callback(struct pt_regs *regs)
562 {
563 	debugger_ipi(regs);
564 }
565 
566 void smp_send_debugger_break(void)
567 {
568 	smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, debugger_ipi_callback, 1000000);
569 }
570 #endif
571 
572 #ifdef CONFIG_KEXEC_CORE
573 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
574 {
575 	int cpu;
576 
577 	smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, crash_ipi_callback, 1000000);
578 	if (kdump_in_progress() && crash_wake_offline) {
579 		for_each_present_cpu(cpu) {
580 			if (cpu_online(cpu))
581 				continue;
582 			/*
583 			 * crash_ipi_callback will wait for
584 			 * all cpus, including offline CPUs.
585 			 * We don't care about nmi_ipi_function.
586 			 * Offline cpus will jump straight into
587 			 * crash_ipi_callback, we can skip the
588 			 * entire NMI dance and waiting for
589 			 * cpus to clear pending mask, etc.
590 			 */
591 			do_smp_send_nmi_ipi(cpu, false);
592 		}
593 	}
594 }
595 #endif
596 
597 #ifdef CONFIG_NMI_IPI
598 static void nmi_stop_this_cpu(struct pt_regs *regs)
599 {
600 	/*
601 	 * IRQs are already hard disabled by the smp_handle_nmi_ipi.
602 	 */
603 	spin_begin();
604 	while (1)
605 		spin_cpu_relax();
606 }
607 
608 void smp_send_stop(void)
609 {
610 	smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, nmi_stop_this_cpu, 1000000);
611 }
612 
613 #else /* CONFIG_NMI_IPI */
614 
615 static void stop_this_cpu(void *dummy)
616 {
617 	hard_irq_disable();
618 	spin_begin();
619 	while (1)
620 		spin_cpu_relax();
621 }
622 
623 void smp_send_stop(void)
624 {
625 	static bool stopped = false;
626 
627 	/*
628 	 * Prevent waiting on csd lock from a previous smp_send_stop.
629 	 * This is racy, but in general callers try to do the right
630 	 * thing and only fire off one smp_send_stop (e.g., see
631 	 * kernel/panic.c)
632 	 */
633 	if (stopped)
634 		return;
635 
636 	stopped = true;
637 
638 	smp_call_function(stop_this_cpu, NULL, 0);
639 }
640 #endif /* CONFIG_NMI_IPI */
641 
642 struct task_struct *current_set[NR_CPUS];
643 
644 static void smp_store_cpu_info(int id)
645 {
646 	per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
647 #ifdef CONFIG_PPC_FSL_BOOK3E
648 	per_cpu(next_tlbcam_idx, id)
649 		= (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
650 #endif
651 }
652 
653 /*
654  * Relationships between CPUs are maintained in a set of per-cpu cpumasks so
655  * rather than just passing around the cpumask we pass around a function that
656  * returns the that cpumask for the given CPU.
657  */
658 static void set_cpus_related(int i, int j, struct cpumask *(*get_cpumask)(int))
659 {
660 	cpumask_set_cpu(i, get_cpumask(j));
661 	cpumask_set_cpu(j, get_cpumask(i));
662 }
663 
664 #ifdef CONFIG_HOTPLUG_CPU
665 static void set_cpus_unrelated(int i, int j,
666 		struct cpumask *(*get_cpumask)(int))
667 {
668 	cpumask_clear_cpu(i, get_cpumask(j));
669 	cpumask_clear_cpu(j, get_cpumask(i));
670 }
671 #endif
672 
673 /*
674  * Extends set_cpus_related. Instead of setting one CPU at a time in
675  * dstmask, set srcmask at oneshot. dstmask should be super set of srcmask.
676  */
677 static void or_cpumasks_related(int i, int j, struct cpumask *(*srcmask)(int),
678 				struct cpumask *(*dstmask)(int))
679 {
680 	struct cpumask *mask;
681 	int k;
682 
683 	mask = srcmask(j);
684 	for_each_cpu(k, srcmask(i))
685 		cpumask_or(dstmask(k), dstmask(k), mask);
686 
687 	if (i == j)
688 		return;
689 
690 	mask = srcmask(i);
691 	for_each_cpu(k, srcmask(j))
692 		cpumask_or(dstmask(k), dstmask(k), mask);
693 }
694 
695 /*
696  * parse_thread_groups: Parses the "ibm,thread-groups" device tree
697  *                      property for the CPU device node @dn and stores
698  *                      the parsed output in the thread_groups
699  *                      structure @tg if the ibm,thread-groups[0]
700  *                      matches @property.
701  *
702  * @dn: The device node of the CPU device.
703  * @tg: Pointer to a thread group structure into which the parsed
704  *      output of "ibm,thread-groups" is stored.
705  * @property: The property of the thread-group that the caller is
706  *            interested in.
707  *
708  * ibm,thread-groups[0..N-1] array defines which group of threads in
709  * the CPU-device node can be grouped together based on the property.
710  *
711  * ibm,thread-groups[0] tells us the property based on which the
712  * threads are being grouped together. If this value is 1, it implies
713  * that the threads in the same group share L1, translation cache.
714  *
715  * ibm,thread-groups[1] tells us how many such thread groups exist.
716  *
717  * ibm,thread-groups[2] tells us the number of threads in each such
718  * group.
719  *
720  * ibm,thread-groups[3..N-1] is the list of threads identified by
721  * "ibm,ppc-interrupt-server#s" arranged as per their membership in
722  * the grouping.
723  *
724  * Example: If ibm,thread-groups = [1,2,4,5,6,7,8,9,10,11,12] it
725  * implies that there are 2 groups of 4 threads each, where each group
726  * of threads share L1, translation cache.
727  *
728  * The "ibm,ppc-interrupt-server#s" of the first group is {5,6,7,8}
729  * and the "ibm,ppc-interrupt-server#s" of the second group is {9, 10,
730  * 11, 12} structure
731  *
732  * Returns 0 on success, -EINVAL if the property does not exist,
733  * -ENODATA if property does not have a value, and -EOVERFLOW if the
734  * property data isn't large enough.
735  */
736 static int parse_thread_groups(struct device_node *dn,
737 			       struct thread_groups *tg,
738 			       unsigned int property)
739 {
740 	int i;
741 	u32 thread_group_array[3 + MAX_THREAD_LIST_SIZE];
742 	u32 *thread_list;
743 	size_t total_threads;
744 	int ret;
745 
746 	ret = of_property_read_u32_array(dn, "ibm,thread-groups",
747 					 thread_group_array, 3);
748 	if (ret)
749 		return ret;
750 
751 	tg->property = thread_group_array[0];
752 	tg->nr_groups = thread_group_array[1];
753 	tg->threads_per_group = thread_group_array[2];
754 	if (tg->property != property ||
755 	    tg->nr_groups < 1 ||
756 	    tg->threads_per_group < 1)
757 		return -ENODATA;
758 
759 	total_threads = tg->nr_groups * tg->threads_per_group;
760 
761 	ret = of_property_read_u32_array(dn, "ibm,thread-groups",
762 					 thread_group_array,
763 					 3 + total_threads);
764 	if (ret)
765 		return ret;
766 
767 	thread_list = &thread_group_array[3];
768 
769 	for (i = 0 ; i < total_threads; i++)
770 		tg->thread_list[i] = thread_list[i];
771 
772 	return 0;
773 }
774 
775 /*
776  * get_cpu_thread_group_start : Searches the thread group in tg->thread_list
777  *                              that @cpu belongs to.
778  *
779  * @cpu : The logical CPU whose thread group is being searched.
780  * @tg : The thread-group structure of the CPU node which @cpu belongs
781  *       to.
782  *
783  * Returns the index to tg->thread_list that points to the the start
784  * of the thread_group that @cpu belongs to.
785  *
786  * Returns -1 if cpu doesn't belong to any of the groups pointed to by
787  * tg->thread_list.
788  */
789 static int get_cpu_thread_group_start(int cpu, struct thread_groups *tg)
790 {
791 	int hw_cpu_id = get_hard_smp_processor_id(cpu);
792 	int i, j;
793 
794 	for (i = 0; i < tg->nr_groups; i++) {
795 		int group_start = i * tg->threads_per_group;
796 
797 		for (j = 0; j < tg->threads_per_group; j++) {
798 			int idx = group_start + j;
799 
800 			if (tg->thread_list[idx] == hw_cpu_id)
801 				return group_start;
802 		}
803 	}
804 
805 	return -1;
806 }
807 
808 static int init_cpu_l1_cache_map(int cpu)
809 
810 {
811 	struct device_node *dn = of_get_cpu_node(cpu, NULL);
812 	struct thread_groups tg = {.property = 0,
813 				   .nr_groups = 0,
814 				   .threads_per_group = 0};
815 	int first_thread = cpu_first_thread_sibling(cpu);
816 	int i, cpu_group_start = -1, err = 0;
817 
818 	if (!dn)
819 		return -ENODATA;
820 
821 	err = parse_thread_groups(dn, &tg, THREAD_GROUP_SHARE_L1);
822 	if (err)
823 		goto out;
824 
825 	cpu_group_start = get_cpu_thread_group_start(cpu, &tg);
826 
827 	if (unlikely(cpu_group_start == -1)) {
828 		WARN_ON_ONCE(1);
829 		err = -ENODATA;
830 		goto out;
831 	}
832 
833 	zalloc_cpumask_var_node(&per_cpu(cpu_l1_cache_map, cpu),
834 				GFP_KERNEL, cpu_to_node(cpu));
835 
836 	for (i = first_thread; i < first_thread + threads_per_core; i++) {
837 		int i_group_start = get_cpu_thread_group_start(i, &tg);
838 
839 		if (unlikely(i_group_start == -1)) {
840 			WARN_ON_ONCE(1);
841 			err = -ENODATA;
842 			goto out;
843 		}
844 
845 		if (i_group_start == cpu_group_start)
846 			cpumask_set_cpu(i, per_cpu(cpu_l1_cache_map, cpu));
847 	}
848 
849 out:
850 	of_node_put(dn);
851 	return err;
852 }
853 
854 static bool shared_caches;
855 
856 #ifdef CONFIG_SCHED_SMT
857 /* cpumask of CPUs with asymmetric SMT dependency */
858 static int powerpc_smt_flags(void)
859 {
860 	int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
861 
862 	if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
863 		printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
864 		flags |= SD_ASYM_PACKING;
865 	}
866 	return flags;
867 }
868 #endif
869 
870 /*
871  * P9 has a slightly odd architecture where pairs of cores share an L2 cache.
872  * This topology makes it *much* cheaper to migrate tasks between adjacent cores
873  * since the migrated task remains cache hot. We want to take advantage of this
874  * at the scheduler level so an extra topology level is required.
875  */
876 static int powerpc_shared_cache_flags(void)
877 {
878 	return SD_SHARE_PKG_RESOURCES;
879 }
880 
881 /*
882  * We can't just pass cpu_l2_cache_mask() directly because
883  * returns a non-const pointer and the compiler barfs on that.
884  */
885 static const struct cpumask *shared_cache_mask(int cpu)
886 {
887 	return per_cpu(cpu_l2_cache_map, cpu);
888 }
889 
890 #ifdef CONFIG_SCHED_SMT
891 static const struct cpumask *smallcore_smt_mask(int cpu)
892 {
893 	return cpu_smallcore_mask(cpu);
894 }
895 #endif
896 
897 static struct cpumask *cpu_coregroup_mask(int cpu)
898 {
899 	return per_cpu(cpu_coregroup_map, cpu);
900 }
901 
902 static bool has_coregroup_support(void)
903 {
904 	return coregroup_enabled;
905 }
906 
907 static const struct cpumask *cpu_mc_mask(int cpu)
908 {
909 	return cpu_coregroup_mask(cpu);
910 }
911 
912 static struct sched_domain_topology_level powerpc_topology[] = {
913 #ifdef CONFIG_SCHED_SMT
914 	{ cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
915 #endif
916 	{ shared_cache_mask, powerpc_shared_cache_flags, SD_INIT_NAME(CACHE) },
917 	{ cpu_mc_mask, SD_INIT_NAME(MC) },
918 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
919 	{ NULL, },
920 };
921 
922 static int init_big_cores(void)
923 {
924 	int cpu;
925 
926 	for_each_possible_cpu(cpu) {
927 		int err = init_cpu_l1_cache_map(cpu);
928 
929 		if (err)
930 			return err;
931 
932 		zalloc_cpumask_var_node(&per_cpu(cpu_smallcore_map, cpu),
933 					GFP_KERNEL,
934 					cpu_to_node(cpu));
935 	}
936 
937 	has_big_cores = true;
938 	return 0;
939 }
940 
941 void __init smp_prepare_cpus(unsigned int max_cpus)
942 {
943 	unsigned int cpu;
944 
945 	DBG("smp_prepare_cpus\n");
946 
947 	/*
948 	 * setup_cpu may need to be called on the boot cpu. We havent
949 	 * spun any cpus up but lets be paranoid.
950 	 */
951 	BUG_ON(boot_cpuid != smp_processor_id());
952 
953 	/* Fixup boot cpu */
954 	smp_store_cpu_info(boot_cpuid);
955 	cpu_callin_map[boot_cpuid] = 1;
956 
957 	for_each_possible_cpu(cpu) {
958 		zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
959 					GFP_KERNEL, cpu_to_node(cpu));
960 		zalloc_cpumask_var_node(&per_cpu(cpu_l2_cache_map, cpu),
961 					GFP_KERNEL, cpu_to_node(cpu));
962 		zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
963 					GFP_KERNEL, cpu_to_node(cpu));
964 		if (has_coregroup_support())
965 			zalloc_cpumask_var_node(&per_cpu(cpu_coregroup_map, cpu),
966 						GFP_KERNEL, cpu_to_node(cpu));
967 
968 #ifdef CONFIG_NEED_MULTIPLE_NODES
969 		/*
970 		 * numa_node_id() works after this.
971 		 */
972 		if (cpu_present(cpu)) {
973 			set_cpu_numa_node(cpu, numa_cpu_lookup_table[cpu]);
974 			set_cpu_numa_mem(cpu,
975 				local_memory_node(numa_cpu_lookup_table[cpu]));
976 		}
977 #endif
978 		/*
979 		 * cpu_core_map is now more updated and exists only since
980 		 * its been exported for long. It only will have a snapshot
981 		 * of cpu_cpu_mask.
982 		 */
983 		cpumask_copy(per_cpu(cpu_core_map, cpu), cpu_cpu_mask(cpu));
984 	}
985 
986 	/* Init the cpumasks so the boot CPU is related to itself */
987 	cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
988 	cpumask_set_cpu(boot_cpuid, cpu_l2_cache_mask(boot_cpuid));
989 
990 	if (has_coregroup_support())
991 		cpumask_set_cpu(boot_cpuid, cpu_coregroup_mask(boot_cpuid));
992 
993 	init_big_cores();
994 	if (has_big_cores) {
995 		cpumask_set_cpu(boot_cpuid,
996 				cpu_smallcore_mask(boot_cpuid));
997 	}
998 
999 	if (smp_ops && smp_ops->probe)
1000 		smp_ops->probe();
1001 }
1002 
1003 void smp_prepare_boot_cpu(void)
1004 {
1005 	BUG_ON(smp_processor_id() != boot_cpuid);
1006 #ifdef CONFIG_PPC64
1007 	paca_ptrs[boot_cpuid]->__current = current;
1008 #endif
1009 	set_numa_node(numa_cpu_lookup_table[boot_cpuid]);
1010 	current_set[boot_cpuid] = current;
1011 }
1012 
1013 #ifdef CONFIG_HOTPLUG_CPU
1014 
1015 int generic_cpu_disable(void)
1016 {
1017 	unsigned int cpu = smp_processor_id();
1018 
1019 	if (cpu == boot_cpuid)
1020 		return -EBUSY;
1021 
1022 	set_cpu_online(cpu, false);
1023 #ifdef CONFIG_PPC64
1024 	vdso_data->processorCount--;
1025 #endif
1026 	/* Update affinity of all IRQs previously aimed at this CPU */
1027 	irq_migrate_all_off_this_cpu();
1028 
1029 	/*
1030 	 * Depending on the details of the interrupt controller, it's possible
1031 	 * that one of the interrupts we just migrated away from this CPU is
1032 	 * actually already pending on this CPU. If we leave it in that state
1033 	 * the interrupt will never be EOI'ed, and will never fire again. So
1034 	 * temporarily enable interrupts here, to allow any pending interrupt to
1035 	 * be received (and EOI'ed), before we take this CPU offline.
1036 	 */
1037 	local_irq_enable();
1038 	mdelay(1);
1039 	local_irq_disable();
1040 
1041 	return 0;
1042 }
1043 
1044 void generic_cpu_die(unsigned int cpu)
1045 {
1046 	int i;
1047 
1048 	for (i = 0; i < 100; i++) {
1049 		smp_rmb();
1050 		if (is_cpu_dead(cpu))
1051 			return;
1052 		msleep(100);
1053 	}
1054 	printk(KERN_ERR "CPU%d didn't die...\n", cpu);
1055 }
1056 
1057 void generic_set_cpu_dead(unsigned int cpu)
1058 {
1059 	per_cpu(cpu_state, cpu) = CPU_DEAD;
1060 }
1061 
1062 /*
1063  * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
1064  * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
1065  * which makes the delay in generic_cpu_die() not happen.
1066  */
1067 void generic_set_cpu_up(unsigned int cpu)
1068 {
1069 	per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1070 }
1071 
1072 int generic_check_cpu_restart(unsigned int cpu)
1073 {
1074 	return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
1075 }
1076 
1077 int is_cpu_dead(unsigned int cpu)
1078 {
1079 	return per_cpu(cpu_state, cpu) == CPU_DEAD;
1080 }
1081 
1082 static bool secondaries_inhibited(void)
1083 {
1084 	return kvm_hv_mode_active();
1085 }
1086 
1087 #else /* HOTPLUG_CPU */
1088 
1089 #define secondaries_inhibited()		0
1090 
1091 #endif
1092 
1093 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
1094 {
1095 #ifdef CONFIG_PPC64
1096 	paca_ptrs[cpu]->__current = idle;
1097 	paca_ptrs[cpu]->kstack = (unsigned long)task_stack_page(idle) +
1098 				 THREAD_SIZE - STACK_FRAME_OVERHEAD;
1099 #endif
1100 	idle->cpu = cpu;
1101 	secondary_current = current_set[cpu] = idle;
1102 }
1103 
1104 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
1105 {
1106 	int rc, c;
1107 
1108 	/*
1109 	 * Don't allow secondary threads to come online if inhibited
1110 	 */
1111 	if (threads_per_core > 1 && secondaries_inhibited() &&
1112 	    cpu_thread_in_subcore(cpu))
1113 		return -EBUSY;
1114 
1115 	if (smp_ops == NULL ||
1116 	    (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
1117 		return -EINVAL;
1118 
1119 	cpu_idle_thread_init(cpu, tidle);
1120 
1121 	/*
1122 	 * The platform might need to allocate resources prior to bringing
1123 	 * up the CPU
1124 	 */
1125 	if (smp_ops->prepare_cpu) {
1126 		rc = smp_ops->prepare_cpu(cpu);
1127 		if (rc)
1128 			return rc;
1129 	}
1130 
1131 	/* Make sure callin-map entry is 0 (can be leftover a CPU
1132 	 * hotplug
1133 	 */
1134 	cpu_callin_map[cpu] = 0;
1135 
1136 	/* The information for processor bringup must
1137 	 * be written out to main store before we release
1138 	 * the processor.
1139 	 */
1140 	smp_mb();
1141 
1142 	/* wake up cpus */
1143 	DBG("smp: kicking cpu %d\n", cpu);
1144 	rc = smp_ops->kick_cpu(cpu);
1145 	if (rc) {
1146 		pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
1147 		return rc;
1148 	}
1149 
1150 	/*
1151 	 * wait to see if the cpu made a callin (is actually up).
1152 	 * use this value that I found through experimentation.
1153 	 * -- Cort
1154 	 */
1155 	if (system_state < SYSTEM_RUNNING)
1156 		for (c = 50000; c && !cpu_callin_map[cpu]; c--)
1157 			udelay(100);
1158 #ifdef CONFIG_HOTPLUG_CPU
1159 	else
1160 		/*
1161 		 * CPUs can take much longer to come up in the
1162 		 * hotplug case.  Wait five seconds.
1163 		 */
1164 		for (c = 5000; c && !cpu_callin_map[cpu]; c--)
1165 			msleep(1);
1166 #endif
1167 
1168 	if (!cpu_callin_map[cpu]) {
1169 		printk(KERN_ERR "Processor %u is stuck.\n", cpu);
1170 		return -ENOENT;
1171 	}
1172 
1173 	DBG("Processor %u found.\n", cpu);
1174 
1175 	if (smp_ops->give_timebase)
1176 		smp_ops->give_timebase();
1177 
1178 	/* Wait until cpu puts itself in the online & active maps */
1179 	spin_until_cond(cpu_online(cpu));
1180 
1181 	return 0;
1182 }
1183 
1184 /* Return the value of the reg property corresponding to the given
1185  * logical cpu.
1186  */
1187 int cpu_to_core_id(int cpu)
1188 {
1189 	struct device_node *np;
1190 	const __be32 *reg;
1191 	int id = -1;
1192 
1193 	np = of_get_cpu_node(cpu, NULL);
1194 	if (!np)
1195 		goto out;
1196 
1197 	reg = of_get_property(np, "reg", NULL);
1198 	if (!reg)
1199 		goto out;
1200 
1201 	id = be32_to_cpup(reg);
1202 out:
1203 	of_node_put(np);
1204 	return id;
1205 }
1206 EXPORT_SYMBOL_GPL(cpu_to_core_id);
1207 
1208 /* Helper routines for cpu to core mapping */
1209 int cpu_core_index_of_thread(int cpu)
1210 {
1211 	return cpu >> threads_shift;
1212 }
1213 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
1214 
1215 int cpu_first_thread_of_core(int core)
1216 {
1217 	return core << threads_shift;
1218 }
1219 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
1220 
1221 /* Must be called when no change can occur to cpu_present_mask,
1222  * i.e. during cpu online or offline.
1223  */
1224 static struct device_node *cpu_to_l2cache(int cpu)
1225 {
1226 	struct device_node *np;
1227 	struct device_node *cache;
1228 
1229 	if (!cpu_present(cpu))
1230 		return NULL;
1231 
1232 	np = of_get_cpu_node(cpu, NULL);
1233 	if (np == NULL)
1234 		return NULL;
1235 
1236 	cache = of_find_next_cache_node(np);
1237 
1238 	of_node_put(np);
1239 
1240 	return cache;
1241 }
1242 
1243 static bool update_mask_by_l2(int cpu, cpumask_var_t *mask)
1244 {
1245 	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1246 	struct device_node *l2_cache, *np;
1247 	int i;
1248 
1249 	if (has_big_cores)
1250 		submask_fn = cpu_smallcore_mask;
1251 
1252 	l2_cache = cpu_to_l2cache(cpu);
1253 	if (!l2_cache || !*mask) {
1254 		/* Assume only core siblings share cache with this CPU */
1255 		for_each_cpu(i, submask_fn(cpu))
1256 			set_cpus_related(cpu, i, cpu_l2_cache_mask);
1257 
1258 		return false;
1259 	}
1260 
1261 	cpumask_and(*mask, cpu_online_mask, cpu_cpu_mask(cpu));
1262 
1263 	/* Update l2-cache mask with all the CPUs that are part of submask */
1264 	or_cpumasks_related(cpu, cpu, submask_fn, cpu_l2_cache_mask);
1265 
1266 	/* Skip all CPUs already part of current CPU l2-cache mask */
1267 	cpumask_andnot(*mask, *mask, cpu_l2_cache_mask(cpu));
1268 
1269 	for_each_cpu(i, *mask) {
1270 		/*
1271 		 * when updating the marks the current CPU has not been marked
1272 		 * online, but we need to update the cache masks
1273 		 */
1274 		np = cpu_to_l2cache(i);
1275 
1276 		/* Skip all CPUs already part of current CPU l2-cache */
1277 		if (np == l2_cache) {
1278 			or_cpumasks_related(cpu, i, submask_fn, cpu_l2_cache_mask);
1279 			cpumask_andnot(*mask, *mask, submask_fn(i));
1280 		} else {
1281 			cpumask_andnot(*mask, *mask, cpu_l2_cache_mask(i));
1282 		}
1283 
1284 		of_node_put(np);
1285 	}
1286 	of_node_put(l2_cache);
1287 
1288 	return true;
1289 }
1290 
1291 #ifdef CONFIG_HOTPLUG_CPU
1292 static void remove_cpu_from_masks(int cpu)
1293 {
1294 	struct cpumask *(*mask_fn)(int) = cpu_sibling_mask;
1295 	int i;
1296 
1297 	if (shared_caches)
1298 		mask_fn = cpu_l2_cache_mask;
1299 
1300 	for_each_cpu(i, mask_fn(cpu)) {
1301 		set_cpus_unrelated(cpu, i, cpu_l2_cache_mask);
1302 		set_cpus_unrelated(cpu, i, cpu_sibling_mask);
1303 		if (has_big_cores)
1304 			set_cpus_unrelated(cpu, i, cpu_smallcore_mask);
1305 	}
1306 
1307 	if (has_coregroup_support()) {
1308 		for_each_cpu(i, cpu_coregroup_mask(cpu))
1309 			set_cpus_unrelated(cpu, i, cpu_coregroup_mask);
1310 	}
1311 }
1312 #endif
1313 
1314 static inline void add_cpu_to_smallcore_masks(int cpu)
1315 {
1316 	int i;
1317 
1318 	if (!has_big_cores)
1319 		return;
1320 
1321 	cpumask_set_cpu(cpu, cpu_smallcore_mask(cpu));
1322 
1323 	for_each_cpu(i, per_cpu(cpu_l1_cache_map, cpu)) {
1324 		if (cpu_online(i))
1325 			set_cpus_related(i, cpu, cpu_smallcore_mask);
1326 	}
1327 }
1328 
1329 static void update_coregroup_mask(int cpu, cpumask_var_t *mask)
1330 {
1331 	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1332 	int coregroup_id = cpu_to_coregroup_id(cpu);
1333 	int i;
1334 
1335 	if (shared_caches)
1336 		submask_fn = cpu_l2_cache_mask;
1337 
1338 	if (!*mask) {
1339 		/* Assume only siblings are part of this CPU's coregroup */
1340 		for_each_cpu(i, submask_fn(cpu))
1341 			set_cpus_related(cpu, i, cpu_coregroup_mask);
1342 
1343 		return;
1344 	}
1345 
1346 	cpumask_and(*mask, cpu_online_mask, cpu_cpu_mask(cpu));
1347 
1348 	/* Update coregroup mask with all the CPUs that are part of submask */
1349 	or_cpumasks_related(cpu, cpu, submask_fn, cpu_coregroup_mask);
1350 
1351 	/* Skip all CPUs already part of coregroup mask */
1352 	cpumask_andnot(*mask, *mask, cpu_coregroup_mask(cpu));
1353 
1354 	for_each_cpu(i, *mask) {
1355 		/* Skip all CPUs not part of this coregroup */
1356 		if (coregroup_id == cpu_to_coregroup_id(i)) {
1357 			or_cpumasks_related(cpu, i, submask_fn, cpu_coregroup_mask);
1358 			cpumask_andnot(*mask, *mask, submask_fn(i));
1359 		} else {
1360 			cpumask_andnot(*mask, *mask, cpu_coregroup_mask(i));
1361 		}
1362 	}
1363 }
1364 
1365 static void add_cpu_to_masks(int cpu)
1366 {
1367 	int first_thread = cpu_first_thread_sibling(cpu);
1368 	cpumask_var_t mask;
1369 	int i;
1370 
1371 	/*
1372 	 * This CPU will not be in the online mask yet so we need to manually
1373 	 * add it to it's own thread sibling mask.
1374 	 */
1375 	cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
1376 
1377 	for (i = first_thread; i < first_thread + threads_per_core; i++)
1378 		if (cpu_online(i))
1379 			set_cpus_related(i, cpu, cpu_sibling_mask);
1380 
1381 	add_cpu_to_smallcore_masks(cpu);
1382 
1383 	/* In CPU-hotplug path, hence use GFP_ATOMIC */
1384 	alloc_cpumask_var_node(&mask, GFP_ATOMIC, cpu_to_node(cpu));
1385 	update_mask_by_l2(cpu, &mask);
1386 
1387 	if (has_coregroup_support())
1388 		update_coregroup_mask(cpu, &mask);
1389 
1390 	free_cpumask_var(mask);
1391 }
1392 
1393 /* Activate a secondary processor. */
1394 void start_secondary(void *unused)
1395 {
1396 	unsigned int cpu = smp_processor_id();
1397 
1398 	mmgrab(&init_mm);
1399 	current->active_mm = &init_mm;
1400 
1401 	smp_store_cpu_info(cpu);
1402 	set_dec(tb_ticks_per_jiffy);
1403 	preempt_disable();
1404 	cpu_callin_map[cpu] = 1;
1405 
1406 	if (smp_ops->setup_cpu)
1407 		smp_ops->setup_cpu(cpu);
1408 	if (smp_ops->take_timebase)
1409 		smp_ops->take_timebase();
1410 
1411 	secondary_cpu_time_init();
1412 
1413 #ifdef CONFIG_PPC64
1414 	if (system_state == SYSTEM_RUNNING)
1415 		vdso_data->processorCount++;
1416 
1417 	vdso_getcpu_init();
1418 #endif
1419 	/* Update topology CPU masks */
1420 	add_cpu_to_masks(cpu);
1421 
1422 	/*
1423 	 * Check for any shared caches. Note that this must be done on a
1424 	 * per-core basis because one core in the pair might be disabled.
1425 	 */
1426 	if (!shared_caches) {
1427 		struct cpumask *(*sibling_mask)(int) = cpu_sibling_mask;
1428 		struct cpumask *mask = cpu_l2_cache_mask(cpu);
1429 
1430 		if (has_big_cores)
1431 			sibling_mask = cpu_smallcore_mask;
1432 
1433 		if (cpumask_weight(mask) > cpumask_weight(sibling_mask(cpu)))
1434 			shared_caches = true;
1435 	}
1436 
1437 	set_numa_node(numa_cpu_lookup_table[cpu]);
1438 	set_numa_mem(local_memory_node(numa_cpu_lookup_table[cpu]));
1439 
1440 	smp_wmb();
1441 	notify_cpu_starting(cpu);
1442 	set_cpu_online(cpu, true);
1443 
1444 	boot_init_stack_canary();
1445 
1446 	local_irq_enable();
1447 
1448 	/* We can enable ftrace for secondary cpus now */
1449 	this_cpu_enable_ftrace();
1450 
1451 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
1452 
1453 	BUG();
1454 }
1455 
1456 int setup_profiling_timer(unsigned int multiplier)
1457 {
1458 	return 0;
1459 }
1460 
1461 static void fixup_topology(void)
1462 {
1463 	int i;
1464 
1465 #ifdef CONFIG_SCHED_SMT
1466 	if (has_big_cores) {
1467 		pr_info("Big cores detected but using small core scheduling\n");
1468 		powerpc_topology[smt_idx].mask = smallcore_smt_mask;
1469 	}
1470 #endif
1471 
1472 	if (!has_coregroup_support())
1473 		powerpc_topology[mc_idx].mask = powerpc_topology[cache_idx].mask;
1474 
1475 	/*
1476 	 * Try to consolidate topology levels here instead of
1477 	 * allowing scheduler to degenerate.
1478 	 * - Dont consolidate if masks are different.
1479 	 * - Dont consolidate if sd_flags exists and are different.
1480 	 */
1481 	for (i = 1; i <= die_idx; i++) {
1482 		if (powerpc_topology[i].mask != powerpc_topology[i - 1].mask)
1483 			continue;
1484 
1485 		if (powerpc_topology[i].sd_flags && powerpc_topology[i - 1].sd_flags &&
1486 				powerpc_topology[i].sd_flags != powerpc_topology[i - 1].sd_flags)
1487 			continue;
1488 
1489 		if (!powerpc_topology[i - 1].sd_flags)
1490 			powerpc_topology[i - 1].sd_flags = powerpc_topology[i].sd_flags;
1491 
1492 		powerpc_topology[i].mask = powerpc_topology[i + 1].mask;
1493 		powerpc_topology[i].sd_flags = powerpc_topology[i + 1].sd_flags;
1494 #ifdef CONFIG_SCHED_DEBUG
1495 		powerpc_topology[i].name = powerpc_topology[i + 1].name;
1496 #endif
1497 	}
1498 }
1499 
1500 void __init smp_cpus_done(unsigned int max_cpus)
1501 {
1502 	/*
1503 	 * We are running pinned to the boot CPU, see rest_init().
1504 	 */
1505 	if (smp_ops && smp_ops->setup_cpu)
1506 		smp_ops->setup_cpu(boot_cpuid);
1507 
1508 	if (smp_ops && smp_ops->bringup_done)
1509 		smp_ops->bringup_done();
1510 
1511 	dump_numa_cpu_topology();
1512 
1513 	fixup_topology();
1514 	set_sched_topology(powerpc_topology);
1515 }
1516 
1517 #ifdef CONFIG_HOTPLUG_CPU
1518 int __cpu_disable(void)
1519 {
1520 	int cpu = smp_processor_id();
1521 	int err;
1522 
1523 	if (!smp_ops->cpu_disable)
1524 		return -ENOSYS;
1525 
1526 	this_cpu_disable_ftrace();
1527 
1528 	err = smp_ops->cpu_disable();
1529 	if (err)
1530 		return err;
1531 
1532 	/* Update sibling maps */
1533 	remove_cpu_from_masks(cpu);
1534 
1535 	return 0;
1536 }
1537 
1538 void __cpu_die(unsigned int cpu)
1539 {
1540 	if (smp_ops->cpu_die)
1541 		smp_ops->cpu_die(cpu);
1542 }
1543 
1544 void arch_cpu_idle_dead(void)
1545 {
1546 	sched_preempt_enable_no_resched();
1547 
1548 	/*
1549 	 * Disable on the down path. This will be re-enabled by
1550 	 * start_secondary() via start_secondary_resume() below
1551 	 */
1552 	this_cpu_disable_ftrace();
1553 
1554 	if (smp_ops->cpu_offline_self)
1555 		smp_ops->cpu_offline_self();
1556 
1557 	/* If we return, we re-enter start_secondary */
1558 	start_secondary_resume();
1559 }
1560 
1561 #endif
1562