4 * Written by Cort Dougan (cort@cs.nmt.edu) borrowing a great
5 * deal of code from the sparc and intel versions.
7 * Copyright (C) 1999 Cort Dougan <cort@cs.nmt.edu>
9 * PowerPC-64 Support added by Dave Engebretsen, Peter Bergner, and
10 * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
12 * This program is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU General Public License
14 * as published by the Free Software Foundation; either version
15 * 2 of the License, or (at your option) any later version.
20 #include <linux/kernel.h>
21 #include <linux/export.h>
22 #include <linux/sched/mm.h>
23 #include <linux/sched/topology.h>
24 #include <linux/smp.h>
25 #include <linux/interrupt.h>
26 #include <linux/delay.h>
27 #include <linux/init.h>
28 #include <linux/spinlock.h>
29 #include <linux/cache.h>
30 #include <linux/err.h>
31 #include <linux/device.h>
32 #include <linux/cpu.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/profile.h>
36 #include <linux/processor.h>
38 #include <asm/ptrace.h>
39 #include <linux/atomic.h>
41 #include <asm/hw_irq.h>
42 #include <asm/kvm_ppc.h>
43 #include <asm/dbell.h>
45 #include <asm/pgtable.h>
49 #include <asm/machdep.h>
50 #include <asm/cputhreads.h>
51 #include <asm/cputable.h>
53 #include <asm/vdso_datapage.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>
66 #define DBG(fmt...) udbg_printf(fmt)
71 #ifdef CONFIG_HOTPLUG_CPU
72 /* State of each CPU during hotplug phases */
73 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
76 struct thread_info *secondary_ti;
78 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
79 DEFINE_PER_CPU(cpumask_var_t, cpu_l2_cache_map);
80 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
82 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
83 EXPORT_PER_CPU_SYMBOL(cpu_l2_cache_map);
84 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
86 /* SMP operations for this machine */
87 struct smp_ops_t *smp_ops;
89 /* Can't be static due to PowerMac hackery */
90 volatile unsigned int cpu_callin_map[NR_CPUS];
92 int smt_enabled_at_boot = 1;
95 * Returns 1 if the specified cpu should be brought up during boot.
96 * Used to inhibit booting threads if they've been disabled or
97 * limited on the command line
99 int smp_generic_cpu_bootable(unsigned int nr)
101 /* Special case - we inhibit secondary thread startup
102 * during boot if the user requests it.
104 if (system_state < SYSTEM_RUNNING && cpu_has_feature(CPU_FTR_SMT)) {
105 if (!smt_enabled_at_boot && cpu_thread_in_core(nr) != 0)
107 if (smt_enabled_at_boot
108 && cpu_thread_in_core(nr) >= smt_enabled_at_boot)
117 int smp_generic_kick_cpu(int nr)
119 if (nr < 0 || nr >= nr_cpu_ids)
123 * The processor is currently spinning, waiting for the
124 * cpu_start field to become non-zero After we set cpu_start,
125 * the processor will continue on to secondary_start
127 if (!paca_ptrs[nr]->cpu_start) {
128 paca_ptrs[nr]->cpu_start = 1;
133 #ifdef CONFIG_HOTPLUG_CPU
135 * Ok it's not there, so it might be soft-unplugged, let's
136 * try to bring it back
138 generic_set_cpu_up(nr);
140 smp_send_reschedule(nr);
141 #endif /* CONFIG_HOTPLUG_CPU */
145 #endif /* CONFIG_PPC64 */
147 static irqreturn_t call_function_action(int irq, void *data)
149 generic_smp_call_function_interrupt();
153 static irqreturn_t reschedule_action(int irq, void *data)
159 static irqreturn_t tick_broadcast_ipi_action(int irq, void *data)
161 timer_broadcast_interrupt();
165 #ifdef CONFIG_NMI_IPI
166 static irqreturn_t nmi_ipi_action(int irq, void *data)
168 smp_handle_nmi_ipi(get_irq_regs());
173 static irq_handler_t smp_ipi_action[] = {
174 [PPC_MSG_CALL_FUNCTION] = call_function_action,
175 [PPC_MSG_RESCHEDULE] = reschedule_action,
176 [PPC_MSG_TICK_BROADCAST] = tick_broadcast_ipi_action,
177 #ifdef CONFIG_NMI_IPI
178 [PPC_MSG_NMI_IPI] = nmi_ipi_action,
183 * The NMI IPI is a fallback and not truly non-maskable. It is simpler
184 * than going through the call function infrastructure, and strongly
185 * serialized, so it is more appropriate for debugging.
187 const char *smp_ipi_name[] = {
188 [PPC_MSG_CALL_FUNCTION] = "ipi call function",
189 [PPC_MSG_RESCHEDULE] = "ipi reschedule",
190 [PPC_MSG_TICK_BROADCAST] = "ipi tick-broadcast",
191 [PPC_MSG_NMI_IPI] = "nmi ipi",
194 /* optional function to request ipi, for controllers with >= 4 ipis */
195 int smp_request_message_ipi(int virq, int msg)
199 if (msg < 0 || msg > PPC_MSG_NMI_IPI)
201 #ifndef CONFIG_NMI_IPI
202 if (msg == PPC_MSG_NMI_IPI)
206 err = request_irq(virq, smp_ipi_action[msg],
207 IRQF_PERCPU | IRQF_NO_THREAD | IRQF_NO_SUSPEND,
208 smp_ipi_name[msg], NULL);
209 WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
210 virq, smp_ipi_name[msg], err);
215 #ifdef CONFIG_PPC_SMP_MUXED_IPI
216 struct cpu_messages {
217 long messages; /* current messages */
219 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
221 void smp_muxed_ipi_set_message(int cpu, int msg)
223 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
224 char *message = (char *)&info->messages;
227 * Order previous accesses before accesses in the IPI handler.
233 void smp_muxed_ipi_message_pass(int cpu, int msg)
235 smp_muxed_ipi_set_message(cpu, msg);
238 * cause_ipi functions are required to include a full barrier
239 * before doing whatever causes the IPI.
241 smp_ops->cause_ipi(cpu);
244 #ifdef __BIG_ENDIAN__
245 #define IPI_MESSAGE(A) (1uL << ((BITS_PER_LONG - 8) - 8 * (A)))
247 #define IPI_MESSAGE(A) (1uL << (8 * (A)))
250 irqreturn_t smp_ipi_demux(void)
252 mb(); /* order any irq clear */
254 return smp_ipi_demux_relaxed();
257 /* sync-free variant. Callers should ensure synchronization */
258 irqreturn_t smp_ipi_demux_relaxed(void)
260 struct cpu_messages *info;
263 info = this_cpu_ptr(&ipi_message);
265 all = xchg(&info->messages, 0);
266 #if defined(CONFIG_KVM_XICS) && defined(CONFIG_KVM_BOOK3S_HV_POSSIBLE)
268 * Must check for PPC_MSG_RM_HOST_ACTION messages
269 * before PPC_MSG_CALL_FUNCTION messages because when
270 * a VM is destroyed, we call kick_all_cpus_sync()
271 * to ensure that any pending PPC_MSG_RM_HOST_ACTION
272 * messages have completed before we free any VCPUs.
274 if (all & IPI_MESSAGE(PPC_MSG_RM_HOST_ACTION))
275 kvmppc_xics_ipi_action();
277 if (all & IPI_MESSAGE(PPC_MSG_CALL_FUNCTION))
278 generic_smp_call_function_interrupt();
279 if (all & IPI_MESSAGE(PPC_MSG_RESCHEDULE))
281 if (all & IPI_MESSAGE(PPC_MSG_TICK_BROADCAST))
282 timer_broadcast_interrupt();
283 #ifdef CONFIG_NMI_IPI
284 if (all & IPI_MESSAGE(PPC_MSG_NMI_IPI))
285 nmi_ipi_action(0, NULL);
287 } while (info->messages);
291 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
293 static inline void do_message_pass(int cpu, int msg)
295 if (smp_ops->message_pass)
296 smp_ops->message_pass(cpu, msg);
297 #ifdef CONFIG_PPC_SMP_MUXED_IPI
299 smp_muxed_ipi_message_pass(cpu, msg);
303 void smp_send_reschedule(int cpu)
306 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
308 EXPORT_SYMBOL_GPL(smp_send_reschedule);
310 void arch_send_call_function_single_ipi(int cpu)
312 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
315 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
319 for_each_cpu(cpu, mask)
320 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
323 #ifdef CONFIG_NMI_IPI
328 * NMI IPIs may not be recoverable, so should not be used as ongoing part of
329 * a running system. They can be used for crash, debug, halt/reboot, etc.
331 * NMI IPIs are globally single threaded. No more than one in progress at
334 * The IPI call waits with interrupts disabled until all targets enter the
335 * NMI handler, then the call returns.
337 * No new NMI can be initiated until targets exit the handler.
339 * The IPI call may time out without all targets entering the NMI handler.
340 * In that case, there is some logic to recover (and ignore subsequent
341 * NMI interrupts that may eventually be raised), but the platform interrupt
342 * handler may not be able to distinguish this from other exception causes,
343 * which may cause a crash.
346 static atomic_t __nmi_ipi_lock = ATOMIC_INIT(0);
347 static struct cpumask nmi_ipi_pending_mask;
348 static int nmi_ipi_busy_count = 0;
349 static void (*nmi_ipi_function)(struct pt_regs *) = NULL;
351 static void nmi_ipi_lock_start(unsigned long *flags)
353 raw_local_irq_save(*flags);
355 while (atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1) {
356 raw_local_irq_restore(*flags);
357 spin_until_cond(atomic_read(&__nmi_ipi_lock) == 0);
358 raw_local_irq_save(*flags);
363 static void nmi_ipi_lock(void)
365 while (atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1)
366 spin_until_cond(atomic_read(&__nmi_ipi_lock) == 0);
369 static void nmi_ipi_unlock(void)
372 WARN_ON(atomic_read(&__nmi_ipi_lock) != 1);
373 atomic_set(&__nmi_ipi_lock, 0);
376 static void nmi_ipi_unlock_end(unsigned long *flags)
379 raw_local_irq_restore(*flags);
383 * Platform NMI handler calls this to ack
385 int smp_handle_nmi_ipi(struct pt_regs *regs)
387 void (*fn)(struct pt_regs *);
389 int me = raw_smp_processor_id();
393 * Unexpected NMIs are possible here because the interrupt may not
394 * be able to distinguish NMI IPIs from other types of NMIs, or
395 * because the caller may have timed out.
397 nmi_ipi_lock_start(&flags);
398 if (!nmi_ipi_busy_count)
400 if (!cpumask_test_cpu(me, &nmi_ipi_pending_mask))
403 fn = nmi_ipi_function;
407 cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
408 nmi_ipi_busy_count++;
416 nmi_ipi_busy_count--;
418 nmi_ipi_unlock_end(&flags);
423 static void do_smp_send_nmi_ipi(int cpu)
425 if (smp_ops->cause_nmi_ipi && smp_ops->cause_nmi_ipi(cpu))
429 do_message_pass(cpu, PPC_MSG_NMI_IPI);
433 for_each_online_cpu(c) {
434 if (c == raw_smp_processor_id())
436 do_message_pass(c, PPC_MSG_NMI_IPI);
441 void smp_flush_nmi_ipi(u64 delay_us)
445 nmi_ipi_lock_start(&flags);
446 while (nmi_ipi_busy_count) {
447 nmi_ipi_unlock_end(&flags);
454 nmi_ipi_lock_start(&flags);
456 nmi_ipi_unlock_end(&flags);
460 * - cpu is the target CPU (must not be this CPU), or NMI_IPI_ALL_OTHERS.
461 * - fn is the target callback function.
462 * - delay_us > 0 is the delay before giving up waiting for targets to
463 * enter the handler, == 0 specifies indefinite delay.
465 int smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
468 int me = raw_smp_processor_id();
472 BUG_ON(cpu < 0 && cpu != NMI_IPI_ALL_OTHERS);
474 if (unlikely(!smp_ops))
477 /* Take the nmi_ipi_busy count/lock with interrupts hard disabled */
478 nmi_ipi_lock_start(&flags);
479 while (nmi_ipi_busy_count) {
480 nmi_ipi_unlock_end(&flags);
481 spin_until_cond(nmi_ipi_busy_count == 0);
482 nmi_ipi_lock_start(&flags);
485 nmi_ipi_function = fn;
489 cpumask_copy(&nmi_ipi_pending_mask, cpu_online_mask);
490 cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
492 /* cpumask starts clear */
493 cpumask_set_cpu(cpu, &nmi_ipi_pending_mask);
495 nmi_ipi_busy_count++;
498 do_smp_send_nmi_ipi(cpu);
500 while (!cpumask_empty(&nmi_ipi_pending_mask)) {
510 if (!cpumask_empty(&nmi_ipi_pending_mask)) {
511 /* Could not gather all CPUs */
513 cpumask_clear(&nmi_ipi_pending_mask);
515 nmi_ipi_busy_count--;
516 nmi_ipi_unlock_end(&flags);
520 #endif /* CONFIG_NMI_IPI */
522 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
523 void tick_broadcast(const struct cpumask *mask)
527 for_each_cpu(cpu, mask)
528 do_message_pass(cpu, PPC_MSG_TICK_BROADCAST);
532 #ifdef CONFIG_DEBUGGER
533 void debugger_ipi_callback(struct pt_regs *regs)
538 void smp_send_debugger_break(void)
540 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, debugger_ipi_callback, 1000000);
544 #ifdef CONFIG_KEXEC_CORE
545 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
549 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, crash_ipi_callback, 1000000);
550 if (kdump_in_progress() && crash_wake_offline) {
551 for_each_present_cpu(cpu) {
555 * crash_ipi_callback will wait for
556 * all cpus, including offline CPUs.
557 * We don't care about nmi_ipi_function.
558 * Offline cpus will jump straight into
559 * crash_ipi_callback, we can skip the
560 * entire NMI dance and waiting for
561 * cpus to clear pending mask, etc.
563 do_smp_send_nmi_ipi(cpu);
569 #ifdef CONFIG_NMI_IPI
570 static void nmi_stop_this_cpu(struct pt_regs *regs)
573 * This is a special case because it never returns, so the NMI IPI
574 * handling would never mark it as done, which makes any later
575 * smp_send_nmi_ipi() call spin forever. Mark it done now.
577 * IRQs are already hard disabled by the smp_handle_nmi_ipi.
580 nmi_ipi_busy_count--;
583 /* Remove this CPU */
584 set_cpu_online(smp_processor_id(), false);
591 void smp_send_stop(void)
593 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, nmi_stop_this_cpu, 1000000);
596 #else /* CONFIG_NMI_IPI */
598 static void stop_this_cpu(void *dummy)
600 /* Remove this CPU */
601 set_cpu_online(smp_processor_id(), false);
609 void smp_send_stop(void)
611 static bool stopped = false;
614 * Prevent waiting on csd lock from a previous smp_send_stop.
615 * This is racy, but in general callers try to do the right
616 * thing and only fire off one smp_send_stop (e.g., see
624 smp_call_function(stop_this_cpu, NULL, 0);
626 #endif /* CONFIG_NMI_IPI */
628 struct thread_info *current_set[NR_CPUS];
630 static void smp_store_cpu_info(int id)
632 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
633 #ifdef CONFIG_PPC_FSL_BOOK3E
634 per_cpu(next_tlbcam_idx, id)
635 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
640 * Relationships between CPUs are maintained in a set of per-cpu cpumasks so
641 * rather than just passing around the cpumask we pass around a function that
642 * returns the that cpumask for the given CPU.
644 static void set_cpus_related(int i, int j, struct cpumask *(*get_cpumask)(int))
646 cpumask_set_cpu(i, get_cpumask(j));
647 cpumask_set_cpu(j, get_cpumask(i));
650 #ifdef CONFIG_HOTPLUG_CPU
651 static void set_cpus_unrelated(int i, int j,
652 struct cpumask *(*get_cpumask)(int))
654 cpumask_clear_cpu(i, get_cpumask(j));
655 cpumask_clear_cpu(j, get_cpumask(i));
659 void __init smp_prepare_cpus(unsigned int max_cpus)
663 DBG("smp_prepare_cpus\n");
666 * setup_cpu may need to be called on the boot cpu. We havent
667 * spun any cpus up but lets be paranoid.
669 BUG_ON(boot_cpuid != smp_processor_id());
672 smp_store_cpu_info(boot_cpuid);
673 cpu_callin_map[boot_cpuid] = 1;
675 for_each_possible_cpu(cpu) {
676 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
677 GFP_KERNEL, cpu_to_node(cpu));
678 zalloc_cpumask_var_node(&per_cpu(cpu_l2_cache_map, cpu),
679 GFP_KERNEL, cpu_to_node(cpu));
680 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
681 GFP_KERNEL, cpu_to_node(cpu));
683 * numa_node_id() works after this.
685 if (cpu_present(cpu)) {
686 set_cpu_numa_node(cpu, numa_cpu_lookup_table[cpu]);
687 set_cpu_numa_mem(cpu,
688 local_memory_node(numa_cpu_lookup_table[cpu]));
692 /* Init the cpumasks so the boot CPU is related to itself */
693 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
694 cpumask_set_cpu(boot_cpuid, cpu_l2_cache_mask(boot_cpuid));
695 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
697 if (smp_ops && smp_ops->probe)
701 void smp_prepare_boot_cpu(void)
703 BUG_ON(smp_processor_id() != boot_cpuid);
705 paca_ptrs[boot_cpuid]->__current = current;
707 set_numa_node(numa_cpu_lookup_table[boot_cpuid]);
708 current_set[boot_cpuid] = task_thread_info(current);
711 #ifdef CONFIG_HOTPLUG_CPU
713 int generic_cpu_disable(void)
715 unsigned int cpu = smp_processor_id();
717 if (cpu == boot_cpuid)
720 set_cpu_online(cpu, false);
722 vdso_data->processorCount--;
724 /* Update affinity of all IRQs previously aimed at this CPU */
725 irq_migrate_all_off_this_cpu();
728 * Depending on the details of the interrupt controller, it's possible
729 * that one of the interrupts we just migrated away from this CPU is
730 * actually already pending on this CPU. If we leave it in that state
731 * the interrupt will never be EOI'ed, and will never fire again. So
732 * temporarily enable interrupts here, to allow any pending interrupt to
733 * be received (and EOI'ed), before we take this CPU offline.
742 void generic_cpu_die(unsigned int cpu)
746 for (i = 0; i < 100; i++) {
748 if (is_cpu_dead(cpu))
752 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
755 void generic_set_cpu_dead(unsigned int cpu)
757 per_cpu(cpu_state, cpu) = CPU_DEAD;
761 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
762 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
763 * which makes the delay in generic_cpu_die() not happen.
765 void generic_set_cpu_up(unsigned int cpu)
767 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
770 int generic_check_cpu_restart(unsigned int cpu)
772 return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
775 int is_cpu_dead(unsigned int cpu)
777 return per_cpu(cpu_state, cpu) == CPU_DEAD;
780 static bool secondaries_inhibited(void)
782 return kvm_hv_mode_active();
785 #else /* HOTPLUG_CPU */
787 #define secondaries_inhibited() 0
791 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
793 struct thread_info *ti = task_thread_info(idle);
796 paca_ptrs[cpu]->__current = idle;
797 paca_ptrs[cpu]->kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
800 secondary_ti = current_set[cpu] = ti;
803 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
808 * Don't allow secondary threads to come online if inhibited
810 if (threads_per_core > 1 && secondaries_inhibited() &&
811 cpu_thread_in_subcore(cpu))
814 if (smp_ops == NULL ||
815 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
818 cpu_idle_thread_init(cpu, tidle);
821 * The platform might need to allocate resources prior to bringing
824 if (smp_ops->prepare_cpu) {
825 rc = smp_ops->prepare_cpu(cpu);
830 /* Make sure callin-map entry is 0 (can be leftover a CPU
833 cpu_callin_map[cpu] = 0;
835 /* The information for processor bringup must
836 * be written out to main store before we release
842 DBG("smp: kicking cpu %d\n", cpu);
843 rc = smp_ops->kick_cpu(cpu);
845 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
850 * wait to see if the cpu made a callin (is actually up).
851 * use this value that I found through experimentation.
854 if (system_state < SYSTEM_RUNNING)
855 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
857 #ifdef CONFIG_HOTPLUG_CPU
860 * CPUs can take much longer to come up in the
861 * hotplug case. Wait five seconds.
863 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
867 if (!cpu_callin_map[cpu]) {
868 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
872 DBG("Processor %u found.\n", cpu);
874 if (smp_ops->give_timebase)
875 smp_ops->give_timebase();
877 /* Wait until cpu puts itself in the online & active maps */
878 spin_until_cond(cpu_online(cpu));
883 /* Return the value of the reg property corresponding to the given
886 int cpu_to_core_id(int cpu)
888 struct device_node *np;
892 np = of_get_cpu_node(cpu, NULL);
896 reg = of_get_property(np, "reg", NULL);
900 id = be32_to_cpup(reg);
905 EXPORT_SYMBOL_GPL(cpu_to_core_id);
907 /* Helper routines for cpu to core mapping */
908 int cpu_core_index_of_thread(int cpu)
910 return cpu >> threads_shift;
912 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
914 int cpu_first_thread_of_core(int core)
916 return core << threads_shift;
918 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
920 /* Must be called when no change can occur to cpu_present_mask,
921 * i.e. during cpu online or offline.
923 static struct device_node *cpu_to_l2cache(int cpu)
925 struct device_node *np;
926 struct device_node *cache;
928 if (!cpu_present(cpu))
931 np = of_get_cpu_node(cpu, NULL);
935 cache = of_find_next_cache_node(np);
942 static bool update_mask_by_l2(int cpu, struct cpumask *(*mask_fn)(int))
944 struct device_node *l2_cache, *np;
947 l2_cache = cpu_to_l2cache(cpu);
951 for_each_cpu(i, cpu_online_mask) {
953 * when updating the marks the current CPU has not been marked
954 * online, but we need to update the cache masks
956 np = cpu_to_l2cache(i);
961 set_cpus_related(cpu, i, mask_fn);
965 of_node_put(l2_cache);
970 #ifdef CONFIG_HOTPLUG_CPU
971 static void remove_cpu_from_masks(int cpu)
975 /* NB: cpu_core_mask is a superset of the others */
976 for_each_cpu(i, cpu_core_mask(cpu)) {
977 set_cpus_unrelated(cpu, i, cpu_core_mask);
978 set_cpus_unrelated(cpu, i, cpu_l2_cache_mask);
979 set_cpus_unrelated(cpu, i, cpu_sibling_mask);
984 static void add_cpu_to_masks(int cpu)
986 int first_thread = cpu_first_thread_sibling(cpu);
987 int chipid = cpu_to_chip_id(cpu);
991 * This CPU will not be in the online mask yet so we need to manually
992 * add it to it's own thread sibling mask.
994 cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
996 for (i = first_thread; i < first_thread + threads_per_core; i++)
998 set_cpus_related(i, cpu, cpu_sibling_mask);
1001 * Copy the thread sibling mask into the cache sibling mask
1002 * and mark any CPUs that share an L2 with this CPU.
1004 for_each_cpu(i, cpu_sibling_mask(cpu))
1005 set_cpus_related(cpu, i, cpu_l2_cache_mask);
1006 update_mask_by_l2(cpu, cpu_l2_cache_mask);
1009 * Copy the cache sibling mask into core sibling mask and mark
1010 * any CPUs on the same chip as this CPU.
1012 for_each_cpu(i, cpu_l2_cache_mask(cpu))
1013 set_cpus_related(cpu, i, cpu_core_mask);
1018 for_each_cpu(i, cpu_online_mask)
1019 if (cpu_to_chip_id(i) == chipid)
1020 set_cpus_related(cpu, i, cpu_core_mask);
1023 static bool shared_caches;
1025 /* Activate a secondary processor. */
1026 void start_secondary(void *unused)
1028 unsigned int cpu = smp_processor_id();
1031 current->active_mm = &init_mm;
1033 smp_store_cpu_info(cpu);
1034 set_dec(tb_ticks_per_jiffy);
1036 cpu_callin_map[cpu] = 1;
1038 if (smp_ops->setup_cpu)
1039 smp_ops->setup_cpu(cpu);
1040 if (smp_ops->take_timebase)
1041 smp_ops->take_timebase();
1043 secondary_cpu_time_init();
1046 if (system_state == SYSTEM_RUNNING)
1047 vdso_data->processorCount++;
1051 /* Update topology CPU masks */
1052 add_cpu_to_masks(cpu);
1055 * Check for any shared caches. Note that this must be done on a
1056 * per-core basis because one core in the pair might be disabled.
1058 if (!cpumask_equal(cpu_l2_cache_mask(cpu), cpu_sibling_mask(cpu)))
1059 shared_caches = true;
1061 set_numa_node(numa_cpu_lookup_table[cpu]);
1062 set_numa_mem(local_memory_node(numa_cpu_lookup_table[cpu]));
1065 notify_cpu_starting(cpu);
1066 set_cpu_online(cpu, true);
1070 /* We can enable ftrace for secondary cpus now */
1071 this_cpu_enable_ftrace();
1073 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
1078 int setup_profiling_timer(unsigned int multiplier)
1083 #ifdef CONFIG_SCHED_SMT
1084 /* cpumask of CPUs with asymetric SMT dependancy */
1085 static int powerpc_smt_flags(void)
1087 int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
1089 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
1090 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
1091 flags |= SD_ASYM_PACKING;
1097 static struct sched_domain_topology_level powerpc_topology[] = {
1098 #ifdef CONFIG_SCHED_SMT
1099 { cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
1101 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1106 * P9 has a slightly odd architecture where pairs of cores share an L2 cache.
1107 * This topology makes it *much* cheaper to migrate tasks between adjacent cores
1108 * since the migrated task remains cache hot. We want to take advantage of this
1109 * at the scheduler level so an extra topology level is required.
1111 static int powerpc_shared_cache_flags(void)
1113 return SD_SHARE_PKG_RESOURCES;
1117 * We can't just pass cpu_l2_cache_mask() directly because
1118 * returns a non-const pointer and the compiler barfs on that.
1120 static const struct cpumask *shared_cache_mask(int cpu)
1122 return cpu_l2_cache_mask(cpu);
1125 static struct sched_domain_topology_level power9_topology[] = {
1126 #ifdef CONFIG_SCHED_SMT
1127 { cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
1129 { shared_cache_mask, powerpc_shared_cache_flags, SD_INIT_NAME(CACHE) },
1130 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1134 void __init smp_cpus_done(unsigned int max_cpus)
1137 * We are running pinned to the boot CPU, see rest_init().
1139 if (smp_ops && smp_ops->setup_cpu)
1140 smp_ops->setup_cpu(boot_cpuid);
1142 if (smp_ops && smp_ops->bringup_done)
1143 smp_ops->bringup_done();
1145 dump_numa_cpu_topology();
1148 * If any CPU detects that it's sharing a cache with another CPU then
1149 * use the deeper topology that is aware of this sharing.
1151 if (shared_caches) {
1152 pr_info("Using shared cache scheduler topology\n");
1153 set_sched_topology(power9_topology);
1155 pr_info("Using standard scheduler topology\n");
1156 set_sched_topology(powerpc_topology);
1160 #ifdef CONFIG_HOTPLUG_CPU
1161 int __cpu_disable(void)
1163 int cpu = smp_processor_id();
1166 if (!smp_ops->cpu_disable)
1169 this_cpu_disable_ftrace();
1171 err = smp_ops->cpu_disable();
1175 /* Update sibling maps */
1176 remove_cpu_from_masks(cpu);
1181 void __cpu_die(unsigned int cpu)
1183 if (smp_ops->cpu_die)
1184 smp_ops->cpu_die(cpu);
1190 * Disable on the down path. This will be re-enabled by
1191 * start_secondary() via start_secondary_resume() below
1193 this_cpu_disable_ftrace();
1198 /* If we return, we re-enter start_secondary */
1199 start_secondary_resume();