1 // SPDX-License-Identifier: GPL-2.0-or-later
6 * deal of code from the sparc and intel versions.
10 * PowerPC-64 Support added by Dave Engebretsen, Peter Bergner, and
11 * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
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>
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>
48 #include <asm/machdep.h>
49 #include <asm/cputhreads.h>
50 #include <asm/cputable.h>
52 #include <asm/vdso_datapage.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>
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 task_struct *secondary_current;
79 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
80 DEFINE_PER_CPU(cpumask_var_t, cpu_smallcore_map);
81 DEFINE_PER_CPU(cpumask_var_t, cpu_l2_cache_map);
82 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
84 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
85 EXPORT_PER_CPU_SYMBOL(cpu_l2_cache_map);
86 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
87 EXPORT_SYMBOL_GPL(has_big_cores);
89 #define MAX_THREAD_LIST_SIZE 8
90 #define THREAD_GROUP_SHARE_L1 1
91 struct thread_groups {
92 unsigned int property;
93 unsigned int nr_groups;
94 unsigned int threads_per_group;
95 unsigned int thread_list[MAX_THREAD_LIST_SIZE];
99 * On big-cores system, cpu_l1_cache_map for each CPU corresponds to
100 * the set its siblings that share the L1-cache.
102 DEFINE_PER_CPU(cpumask_var_t, cpu_l1_cache_map);
104 /* SMP operations for this machine */
105 struct smp_ops_t *smp_ops;
107 /* Can't be static due to PowerMac hackery */
108 volatile unsigned int cpu_callin_map[NR_CPUS];
110 int smt_enabled_at_boot = 1;
113 * Returns 1 if the specified cpu should be brought up during boot.
114 * Used to inhibit booting threads if they've been disabled or
115 * limited on the command line
117 int smp_generic_cpu_bootable(unsigned int nr)
119 /* Special case - we inhibit secondary thread startup
120 * during boot if the user requests it.
122 if (system_state < SYSTEM_RUNNING && cpu_has_feature(CPU_FTR_SMT)) {
123 if (!smt_enabled_at_boot && cpu_thread_in_core(nr) != 0)
125 if (smt_enabled_at_boot
126 && cpu_thread_in_core(nr) >= smt_enabled_at_boot)
135 int smp_generic_kick_cpu(int nr)
137 if (nr < 0 || nr >= nr_cpu_ids)
141 * The processor is currently spinning, waiting for the
142 * cpu_start field to become non-zero After we set cpu_start,
143 * the processor will continue on to secondary_start
145 if (!paca_ptrs[nr]->cpu_start) {
146 paca_ptrs[nr]->cpu_start = 1;
151 #ifdef CONFIG_HOTPLUG_CPU
153 * Ok it's not there, so it might be soft-unplugged, let's
154 * try to bring it back
156 generic_set_cpu_up(nr);
158 smp_send_reschedule(nr);
159 #endif /* CONFIG_HOTPLUG_CPU */
163 #endif /* CONFIG_PPC64 */
165 static irqreturn_t call_function_action(int irq, void *data)
167 generic_smp_call_function_interrupt();
171 static irqreturn_t reschedule_action(int irq, void *data)
177 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
178 static irqreturn_t tick_broadcast_ipi_action(int irq, void *data)
180 timer_broadcast_interrupt();
185 #ifdef CONFIG_NMI_IPI
186 static irqreturn_t nmi_ipi_action(int irq, void *data)
188 smp_handle_nmi_ipi(get_irq_regs());
193 static irq_handler_t smp_ipi_action[] = {
194 [PPC_MSG_CALL_FUNCTION] = call_function_action,
195 [PPC_MSG_RESCHEDULE] = reschedule_action,
196 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
197 [PPC_MSG_TICK_BROADCAST] = tick_broadcast_ipi_action,
199 #ifdef CONFIG_NMI_IPI
200 [PPC_MSG_NMI_IPI] = nmi_ipi_action,
205 * The NMI IPI is a fallback and not truly non-maskable. It is simpler
206 * than going through the call function infrastructure, and strongly
207 * serialized, so it is more appropriate for debugging.
209 const char *smp_ipi_name[] = {
210 [PPC_MSG_CALL_FUNCTION] = "ipi call function",
211 [PPC_MSG_RESCHEDULE] = "ipi reschedule",
212 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
213 [PPC_MSG_TICK_BROADCAST] = "ipi tick-broadcast",
215 #ifdef CONFIG_NMI_IPI
216 [PPC_MSG_NMI_IPI] = "nmi ipi",
220 /* optional function to request ipi, for controllers with >= 4 ipis */
221 int smp_request_message_ipi(int virq, int msg)
225 if (msg < 0 || msg > PPC_MSG_NMI_IPI)
227 #ifndef CONFIG_NMI_IPI
228 if (msg == PPC_MSG_NMI_IPI)
232 err = request_irq(virq, smp_ipi_action[msg],
233 IRQF_PERCPU | IRQF_NO_THREAD | IRQF_NO_SUSPEND,
234 smp_ipi_name[msg], NULL);
235 WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
236 virq, smp_ipi_name[msg], err);
241 #ifdef CONFIG_PPC_SMP_MUXED_IPI
242 struct cpu_messages {
243 long messages; /* current messages */
245 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
247 void smp_muxed_ipi_set_message(int cpu, int msg)
249 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
250 char *message = (char *)&info->messages;
253 * Order previous accesses before accesses in the IPI handler.
259 void smp_muxed_ipi_message_pass(int cpu, int msg)
261 smp_muxed_ipi_set_message(cpu, msg);
264 * cause_ipi functions are required to include a full barrier
265 * before doing whatever causes the IPI.
267 smp_ops->cause_ipi(cpu);
270 #ifdef __BIG_ENDIAN__
271 #define IPI_MESSAGE(A) (1uL << ((BITS_PER_LONG - 8) - 8 * (A)))
273 #define IPI_MESSAGE(A) (1uL << (8 * (A)))
276 irqreturn_t smp_ipi_demux(void)
278 mb(); /* order any irq clear */
280 return smp_ipi_demux_relaxed();
283 /* sync-free variant. Callers should ensure synchronization */
284 irqreturn_t smp_ipi_demux_relaxed(void)
286 struct cpu_messages *info;
289 info = this_cpu_ptr(&ipi_message);
291 all = xchg(&info->messages, 0);
292 #if defined(CONFIG_KVM_XICS) && defined(CONFIG_KVM_BOOK3S_HV_POSSIBLE)
294 * Must check for PPC_MSG_RM_HOST_ACTION messages
295 * before PPC_MSG_CALL_FUNCTION messages because when
296 * a VM is destroyed, we call kick_all_cpus_sync()
297 * to ensure that any pending PPC_MSG_RM_HOST_ACTION
298 * messages have completed before we free any VCPUs.
300 if (all & IPI_MESSAGE(PPC_MSG_RM_HOST_ACTION))
301 kvmppc_xics_ipi_action();
303 if (all & IPI_MESSAGE(PPC_MSG_CALL_FUNCTION))
304 generic_smp_call_function_interrupt();
305 if (all & IPI_MESSAGE(PPC_MSG_RESCHEDULE))
307 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
308 if (all & IPI_MESSAGE(PPC_MSG_TICK_BROADCAST))
309 timer_broadcast_interrupt();
311 #ifdef CONFIG_NMI_IPI
312 if (all & IPI_MESSAGE(PPC_MSG_NMI_IPI))
313 nmi_ipi_action(0, NULL);
315 } while (info->messages);
319 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
321 static inline void do_message_pass(int cpu, int msg)
323 if (smp_ops->message_pass)
324 smp_ops->message_pass(cpu, msg);
325 #ifdef CONFIG_PPC_SMP_MUXED_IPI
327 smp_muxed_ipi_message_pass(cpu, msg);
331 void smp_send_reschedule(int cpu)
334 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
336 EXPORT_SYMBOL_GPL(smp_send_reschedule);
338 void arch_send_call_function_single_ipi(int cpu)
340 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
343 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
347 for_each_cpu(cpu, mask)
348 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
351 #ifdef CONFIG_NMI_IPI
356 * NMI IPIs may not be recoverable, so should not be used as ongoing part of
357 * a running system. They can be used for crash, debug, halt/reboot, etc.
359 * The IPI call waits with interrupts disabled until all targets enter the
360 * NMI handler, then returns. Subsequent IPIs can be issued before targets
361 * have returned from their handlers, so there is no guarantee about
362 * concurrency or re-entrancy.
364 * A new NMI can be issued before all targets exit the handler.
366 * The IPI call may time out without all targets entering the NMI handler.
367 * In that case, there is some logic to recover (and ignore subsequent
368 * NMI interrupts that may eventually be raised), but the platform interrupt
369 * handler may not be able to distinguish this from other exception causes,
370 * which may cause a crash.
373 static atomic_t __nmi_ipi_lock = ATOMIC_INIT(0);
374 static struct cpumask nmi_ipi_pending_mask;
375 static bool nmi_ipi_busy = false;
376 static void (*nmi_ipi_function)(struct pt_regs *) = NULL;
378 static void nmi_ipi_lock_start(unsigned long *flags)
380 raw_local_irq_save(*flags);
382 while (atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1) {
383 raw_local_irq_restore(*flags);
384 spin_until_cond(atomic_read(&__nmi_ipi_lock) == 0);
385 raw_local_irq_save(*flags);
390 static void nmi_ipi_lock(void)
392 while (atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1)
393 spin_until_cond(atomic_read(&__nmi_ipi_lock) == 0);
396 static void nmi_ipi_unlock(void)
399 WARN_ON(atomic_read(&__nmi_ipi_lock) != 1);
400 atomic_set(&__nmi_ipi_lock, 0);
403 static void nmi_ipi_unlock_end(unsigned long *flags)
406 raw_local_irq_restore(*flags);
410 * Platform NMI handler calls this to ack
412 int smp_handle_nmi_ipi(struct pt_regs *regs)
414 void (*fn)(struct pt_regs *) = NULL;
416 int me = raw_smp_processor_id();
420 * Unexpected NMIs are possible here because the interrupt may not
421 * be able to distinguish NMI IPIs from other types of NMIs, or
422 * because the caller may have timed out.
424 nmi_ipi_lock_start(&flags);
425 if (cpumask_test_cpu(me, &nmi_ipi_pending_mask)) {
426 cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
427 fn = READ_ONCE(nmi_ipi_function);
431 nmi_ipi_unlock_end(&flags);
439 static void do_smp_send_nmi_ipi(int cpu, bool safe)
441 if (!safe && smp_ops->cause_nmi_ipi && smp_ops->cause_nmi_ipi(cpu))
445 do_message_pass(cpu, PPC_MSG_NMI_IPI);
449 for_each_online_cpu(c) {
450 if (c == raw_smp_processor_id())
452 do_message_pass(c, PPC_MSG_NMI_IPI);
458 * - cpu is the target CPU (must not be this CPU), or NMI_IPI_ALL_OTHERS.
459 * - fn is the target callback function.
460 * - delay_us > 0 is the delay before giving up waiting for targets to
461 * begin executing the handler, == 0 specifies indefinite delay.
463 static int __smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *),
464 u64 delay_us, bool safe)
467 int me = raw_smp_processor_id();
471 BUG_ON(cpu < 0 && cpu != NMI_IPI_ALL_OTHERS);
473 if (unlikely(!smp_ops))
476 nmi_ipi_lock_start(&flags);
477 while (nmi_ipi_busy) {
478 nmi_ipi_unlock_end(&flags);
479 spin_until_cond(!nmi_ipi_busy);
480 nmi_ipi_lock_start(&flags);
483 nmi_ipi_function = fn;
485 WARN_ON_ONCE(!cpumask_empty(&nmi_ipi_pending_mask));
489 cpumask_copy(&nmi_ipi_pending_mask, cpu_online_mask);
490 cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
492 cpumask_set_cpu(cpu, &nmi_ipi_pending_mask);
497 /* Interrupts remain hard disabled */
499 do_smp_send_nmi_ipi(cpu, safe);
502 /* nmi_ipi_busy is set here, so unlock/lock is okay */
503 while (!cpumask_empty(&nmi_ipi_pending_mask)) {
514 if (!cpumask_empty(&nmi_ipi_pending_mask)) {
515 /* Timeout waiting for CPUs to call smp_handle_nmi_ipi */
517 cpumask_clear(&nmi_ipi_pending_mask);
520 nmi_ipi_function = NULL;
521 nmi_ipi_busy = false;
523 nmi_ipi_unlock_end(&flags);
528 int smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
530 return __smp_send_nmi_ipi(cpu, fn, delay_us, false);
533 int smp_send_safe_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
535 return __smp_send_nmi_ipi(cpu, fn, delay_us, true);
537 #endif /* CONFIG_NMI_IPI */
539 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
540 void tick_broadcast(const struct cpumask *mask)
544 for_each_cpu(cpu, mask)
545 do_message_pass(cpu, PPC_MSG_TICK_BROADCAST);
549 #ifdef CONFIG_DEBUGGER
550 void debugger_ipi_callback(struct pt_regs *regs)
555 void smp_send_debugger_break(void)
557 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, debugger_ipi_callback, 1000000);
561 #ifdef CONFIG_KEXEC_CORE
562 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
566 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, crash_ipi_callback, 1000000);
567 if (kdump_in_progress() && crash_wake_offline) {
568 for_each_present_cpu(cpu) {
572 * crash_ipi_callback will wait for
573 * all cpus, including offline CPUs.
574 * We don't care about nmi_ipi_function.
575 * Offline cpus will jump straight into
576 * crash_ipi_callback, we can skip the
577 * entire NMI dance and waiting for
578 * cpus to clear pending mask, etc.
580 do_smp_send_nmi_ipi(cpu, false);
586 #ifdef CONFIG_NMI_IPI
587 static void nmi_stop_this_cpu(struct pt_regs *regs)
590 * IRQs are already hard disabled by the smp_handle_nmi_ipi.
597 void smp_send_stop(void)
599 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, nmi_stop_this_cpu, 1000000);
602 #else /* CONFIG_NMI_IPI */
604 static void stop_this_cpu(void *dummy)
612 void smp_send_stop(void)
614 static bool stopped = false;
617 * Prevent waiting on csd lock from a previous smp_send_stop.
618 * This is racy, but in general callers try to do the right
619 * thing and only fire off one smp_send_stop (e.g., see
627 smp_call_function(stop_this_cpu, NULL, 0);
629 #endif /* CONFIG_NMI_IPI */
631 struct task_struct *current_set[NR_CPUS];
633 static void smp_store_cpu_info(int id)
635 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
636 #ifdef CONFIG_PPC_FSL_BOOK3E
637 per_cpu(next_tlbcam_idx, id)
638 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
643 * Relationships between CPUs are maintained in a set of per-cpu cpumasks so
644 * rather than just passing around the cpumask we pass around a function that
645 * returns the that cpumask for the given CPU.
647 static void set_cpus_related(int i, int j, struct cpumask *(*get_cpumask)(int))
649 cpumask_set_cpu(i, get_cpumask(j));
650 cpumask_set_cpu(j, get_cpumask(i));
653 #ifdef CONFIG_HOTPLUG_CPU
654 static void set_cpus_unrelated(int i, int j,
655 struct cpumask *(*get_cpumask)(int))
657 cpumask_clear_cpu(i, get_cpumask(j));
658 cpumask_clear_cpu(j, get_cpumask(i));
663 * parse_thread_groups: Parses the "ibm,thread-groups" device tree
664 * property for the CPU device node @dn and stores
665 * the parsed output in the thread_groups
666 * structure @tg if the ibm,thread-groups[0]
669 * @dn: The device node of the CPU device.
670 * @tg: Pointer to a thread group structure into which the parsed
671 * output of "ibm,thread-groups" is stored.
672 * @property: The property of the thread-group that the caller is
675 * ibm,thread-groups[0..N-1] array defines which group of threads in
676 * the CPU-device node can be grouped together based on the property.
678 * ibm,thread-groups[0] tells us the property based on which the
679 * threads are being grouped together. If this value is 1, it implies
680 * that the threads in the same group share L1, translation cache.
682 * ibm,thread-groups[1] tells us how many such thread groups exist.
684 * ibm,thread-groups[2] tells us the number of threads in each such
687 * ibm,thread-groups[3..N-1] is the list of threads identified by
688 * "ibm,ppc-interrupt-server#s" arranged as per their membership in
691 * Example: If ibm,thread-groups = [1,2,4,5,6,7,8,9,10,11,12] it
692 * implies that there are 2 groups of 4 threads each, where each group
693 * of threads share L1, translation cache.
695 * The "ibm,ppc-interrupt-server#s" of the first group is {5,6,7,8}
696 * and the "ibm,ppc-interrupt-server#s" of the second group is {9, 10,
699 * Returns 0 on success, -EINVAL if the property does not exist,
700 * -ENODATA if property does not have a value, and -EOVERFLOW if the
701 * property data isn't large enough.
703 static int parse_thread_groups(struct device_node *dn,
704 struct thread_groups *tg,
705 unsigned int property)
708 u32 thread_group_array[3 + MAX_THREAD_LIST_SIZE];
710 size_t total_threads;
713 ret = of_property_read_u32_array(dn, "ibm,thread-groups",
714 thread_group_array, 3);
718 tg->property = thread_group_array[0];
719 tg->nr_groups = thread_group_array[1];
720 tg->threads_per_group = thread_group_array[2];
721 if (tg->property != property ||
723 tg->threads_per_group < 1)
726 total_threads = tg->nr_groups * tg->threads_per_group;
728 ret = of_property_read_u32_array(dn, "ibm,thread-groups",
734 thread_list = &thread_group_array[3];
736 for (i = 0 ; i < total_threads; i++)
737 tg->thread_list[i] = thread_list[i];
743 * get_cpu_thread_group_start : Searches the thread group in tg->thread_list
744 * that @cpu belongs to.
746 * @cpu : The logical CPU whose thread group is being searched.
747 * @tg : The thread-group structure of the CPU node which @cpu belongs
750 * Returns the index to tg->thread_list that points to the the start
751 * of the thread_group that @cpu belongs to.
753 * Returns -1 if cpu doesn't belong to any of the groups pointed to by
756 static int get_cpu_thread_group_start(int cpu, struct thread_groups *tg)
758 int hw_cpu_id = get_hard_smp_processor_id(cpu);
761 for (i = 0; i < tg->nr_groups; i++) {
762 int group_start = i * tg->threads_per_group;
764 for (j = 0; j < tg->threads_per_group; j++) {
765 int idx = group_start + j;
767 if (tg->thread_list[idx] == hw_cpu_id)
775 static int init_cpu_l1_cache_map(int cpu)
778 struct device_node *dn = of_get_cpu_node(cpu, NULL);
779 struct thread_groups tg = {.property = 0,
781 .threads_per_group = 0};
782 int first_thread = cpu_first_thread_sibling(cpu);
783 int i, cpu_group_start = -1, err = 0;
788 err = parse_thread_groups(dn, &tg, THREAD_GROUP_SHARE_L1);
792 zalloc_cpumask_var_node(&per_cpu(cpu_l1_cache_map, cpu),
796 cpu_group_start = get_cpu_thread_group_start(cpu, &tg);
798 if (unlikely(cpu_group_start == -1)) {
804 for (i = first_thread; i < first_thread + threads_per_core; i++) {
805 int i_group_start = get_cpu_thread_group_start(i, &tg);
807 if (unlikely(i_group_start == -1)) {
813 if (i_group_start == cpu_group_start)
814 cpumask_set_cpu(i, per_cpu(cpu_l1_cache_map, cpu));
822 static int init_big_cores(void)
826 for_each_possible_cpu(cpu) {
827 int err = init_cpu_l1_cache_map(cpu);
832 zalloc_cpumask_var_node(&per_cpu(cpu_smallcore_map, cpu),
837 has_big_cores = true;
841 void __init smp_prepare_cpus(unsigned int max_cpus)
845 DBG("smp_prepare_cpus\n");
848 * setup_cpu may need to be called on the boot cpu. We havent
849 * spun any cpus up but lets be paranoid.
851 BUG_ON(boot_cpuid != smp_processor_id());
854 smp_store_cpu_info(boot_cpuid);
855 cpu_callin_map[boot_cpuid] = 1;
857 for_each_possible_cpu(cpu) {
858 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
859 GFP_KERNEL, cpu_to_node(cpu));
860 zalloc_cpumask_var_node(&per_cpu(cpu_l2_cache_map, cpu),
861 GFP_KERNEL, cpu_to_node(cpu));
862 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
863 GFP_KERNEL, cpu_to_node(cpu));
865 * numa_node_id() works after this.
867 if (cpu_present(cpu)) {
868 set_cpu_numa_node(cpu, numa_cpu_lookup_table[cpu]);
869 set_cpu_numa_mem(cpu,
870 local_memory_node(numa_cpu_lookup_table[cpu]));
874 /* Init the cpumasks so the boot CPU is related to itself */
875 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
876 cpumask_set_cpu(boot_cpuid, cpu_l2_cache_mask(boot_cpuid));
877 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
881 cpumask_set_cpu(boot_cpuid,
882 cpu_smallcore_mask(boot_cpuid));
885 if (smp_ops && smp_ops->probe)
889 void smp_prepare_boot_cpu(void)
891 BUG_ON(smp_processor_id() != boot_cpuid);
893 paca_ptrs[boot_cpuid]->__current = current;
895 set_numa_node(numa_cpu_lookup_table[boot_cpuid]);
896 current_set[boot_cpuid] = current;
899 #ifdef CONFIG_HOTPLUG_CPU
901 int generic_cpu_disable(void)
903 unsigned int cpu = smp_processor_id();
905 if (cpu == boot_cpuid)
908 set_cpu_online(cpu, false);
910 vdso_data->processorCount--;
912 /* Update affinity of all IRQs previously aimed at this CPU */
913 irq_migrate_all_off_this_cpu();
916 * Depending on the details of the interrupt controller, it's possible
917 * that one of the interrupts we just migrated away from this CPU is
918 * actually already pending on this CPU. If we leave it in that state
919 * the interrupt will never be EOI'ed, and will never fire again. So
920 * temporarily enable interrupts here, to allow any pending interrupt to
921 * be received (and EOI'ed), before we take this CPU offline.
930 void generic_cpu_die(unsigned int cpu)
934 for (i = 0; i < 100; i++) {
936 if (is_cpu_dead(cpu))
940 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
943 void generic_set_cpu_dead(unsigned int cpu)
945 per_cpu(cpu_state, cpu) = CPU_DEAD;
949 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
950 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
951 * which makes the delay in generic_cpu_die() not happen.
953 void generic_set_cpu_up(unsigned int cpu)
955 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
958 int generic_check_cpu_restart(unsigned int cpu)
960 return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
963 int is_cpu_dead(unsigned int cpu)
965 return per_cpu(cpu_state, cpu) == CPU_DEAD;
968 static bool secondaries_inhibited(void)
970 return kvm_hv_mode_active();
973 #else /* HOTPLUG_CPU */
975 #define secondaries_inhibited() 0
979 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
982 paca_ptrs[cpu]->__current = idle;
983 paca_ptrs[cpu]->kstack = (unsigned long)task_stack_page(idle) +
984 THREAD_SIZE - STACK_FRAME_OVERHEAD;
987 secondary_current = current_set[cpu] = idle;
990 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
995 * Don't allow secondary threads to come online if inhibited
997 if (threads_per_core > 1 && secondaries_inhibited() &&
998 cpu_thread_in_subcore(cpu))
1001 if (smp_ops == NULL ||
1002 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
1005 cpu_idle_thread_init(cpu, tidle);
1008 * The platform might need to allocate resources prior to bringing
1011 if (smp_ops->prepare_cpu) {
1012 rc = smp_ops->prepare_cpu(cpu);
1017 /* Make sure callin-map entry is 0 (can be leftover a CPU
1020 cpu_callin_map[cpu] = 0;
1022 /* The information for processor bringup must
1023 * be written out to main store before we release
1029 DBG("smp: kicking cpu %d\n", cpu);
1030 rc = smp_ops->kick_cpu(cpu);
1032 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
1037 * wait to see if the cpu made a callin (is actually up).
1038 * use this value that I found through experimentation.
1041 if (system_state < SYSTEM_RUNNING)
1042 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
1044 #ifdef CONFIG_HOTPLUG_CPU
1047 * CPUs can take much longer to come up in the
1048 * hotplug case. Wait five seconds.
1050 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
1054 if (!cpu_callin_map[cpu]) {
1055 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
1059 DBG("Processor %u found.\n", cpu);
1061 if (smp_ops->give_timebase)
1062 smp_ops->give_timebase();
1064 /* Wait until cpu puts itself in the online & active maps */
1065 spin_until_cond(cpu_online(cpu));
1070 /* Return the value of the reg property corresponding to the given
1073 int cpu_to_core_id(int cpu)
1075 struct device_node *np;
1079 np = of_get_cpu_node(cpu, NULL);
1083 reg = of_get_property(np, "reg", NULL);
1087 id = be32_to_cpup(reg);
1092 EXPORT_SYMBOL_GPL(cpu_to_core_id);
1094 /* Helper routines for cpu to core mapping */
1095 int cpu_core_index_of_thread(int cpu)
1097 return cpu >> threads_shift;
1099 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
1101 int cpu_first_thread_of_core(int core)
1103 return core << threads_shift;
1105 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
1107 /* Must be called when no change can occur to cpu_present_mask,
1108 * i.e. during cpu online or offline.
1110 static struct device_node *cpu_to_l2cache(int cpu)
1112 struct device_node *np;
1113 struct device_node *cache;
1115 if (!cpu_present(cpu))
1118 np = of_get_cpu_node(cpu, NULL);
1122 cache = of_find_next_cache_node(np);
1129 static bool update_mask_by_l2(int cpu, struct cpumask *(*mask_fn)(int))
1131 struct device_node *l2_cache, *np;
1134 l2_cache = cpu_to_l2cache(cpu);
1138 for_each_cpu(i, cpu_online_mask) {
1140 * when updating the marks the current CPU has not been marked
1141 * online, but we need to update the cache masks
1143 np = cpu_to_l2cache(i);
1148 set_cpus_related(cpu, i, mask_fn);
1152 of_node_put(l2_cache);
1157 #ifdef CONFIG_HOTPLUG_CPU
1158 static void remove_cpu_from_masks(int cpu)
1162 /* NB: cpu_core_mask is a superset of the others */
1163 for_each_cpu(i, cpu_core_mask(cpu)) {
1164 set_cpus_unrelated(cpu, i, cpu_core_mask);
1165 set_cpus_unrelated(cpu, i, cpu_l2_cache_mask);
1166 set_cpus_unrelated(cpu, i, cpu_sibling_mask);
1168 set_cpus_unrelated(cpu, i, cpu_smallcore_mask);
1173 static inline void add_cpu_to_smallcore_masks(int cpu)
1175 struct cpumask *this_l1_cache_map = per_cpu(cpu_l1_cache_map, cpu);
1176 int i, first_thread = cpu_first_thread_sibling(cpu);
1181 cpumask_set_cpu(cpu, cpu_smallcore_mask(cpu));
1183 for (i = first_thread; i < first_thread + threads_per_core; i++) {
1184 if (cpu_online(i) && cpumask_test_cpu(i, this_l1_cache_map))
1185 set_cpus_related(i, cpu, cpu_smallcore_mask);
1189 int get_physical_package_id(int cpu)
1191 int pkg_id = cpu_to_chip_id(cpu);
1194 * If the platform is PowerNV or Guest on KVM, ibm,chip-id is
1195 * defined. Hence we would return the chip-id as the result of
1196 * get_physical_package_id.
1198 if (pkg_id == -1 && firmware_has_feature(FW_FEATURE_LPAR) &&
1199 IS_ENABLED(CONFIG_PPC_SPLPAR)) {
1200 struct device_node *np = of_get_cpu_node(cpu, NULL);
1201 pkg_id = of_node_to_nid(np);
1207 EXPORT_SYMBOL_GPL(get_physical_package_id);
1209 static void add_cpu_to_masks(int cpu)
1211 int first_thread = cpu_first_thread_sibling(cpu);
1212 int pkg_id = get_physical_package_id(cpu);
1216 * This CPU will not be in the online mask yet so we need to manually
1217 * add it to it's own thread sibling mask.
1219 cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
1221 for (i = first_thread; i < first_thread + threads_per_core; i++)
1223 set_cpus_related(i, cpu, cpu_sibling_mask);
1225 add_cpu_to_smallcore_masks(cpu);
1227 * Copy the thread sibling mask into the cache sibling mask
1228 * and mark any CPUs that share an L2 with this CPU.
1230 for_each_cpu(i, cpu_sibling_mask(cpu))
1231 set_cpus_related(cpu, i, cpu_l2_cache_mask);
1232 update_mask_by_l2(cpu, cpu_l2_cache_mask);
1235 * Copy the cache sibling mask into core sibling mask and mark
1236 * any CPUs on the same chip as this CPU.
1238 for_each_cpu(i, cpu_l2_cache_mask(cpu))
1239 set_cpus_related(cpu, i, cpu_core_mask);
1244 for_each_cpu(i, cpu_online_mask)
1245 if (get_physical_package_id(i) == pkg_id)
1246 set_cpus_related(cpu, i, cpu_core_mask);
1249 static bool shared_caches;
1251 /* Activate a secondary processor. */
1252 void start_secondary(void *unused)
1254 unsigned int cpu = smp_processor_id();
1255 struct cpumask *(*sibling_mask)(int) = cpu_sibling_mask;
1258 current->active_mm = &init_mm;
1260 smp_store_cpu_info(cpu);
1261 set_dec(tb_ticks_per_jiffy);
1263 cpu_callin_map[cpu] = 1;
1265 if (smp_ops->setup_cpu)
1266 smp_ops->setup_cpu(cpu);
1267 if (smp_ops->take_timebase)
1268 smp_ops->take_timebase();
1270 secondary_cpu_time_init();
1273 if (system_state == SYSTEM_RUNNING)
1274 vdso_data->processorCount++;
1278 /* Update topology CPU masks */
1279 add_cpu_to_masks(cpu);
1282 sibling_mask = cpu_smallcore_mask;
1284 * Check for any shared caches. Note that this must be done on a
1285 * per-core basis because one core in the pair might be disabled.
1287 if (!cpumask_equal(cpu_l2_cache_mask(cpu), sibling_mask(cpu)))
1288 shared_caches = true;
1290 set_numa_node(numa_cpu_lookup_table[cpu]);
1291 set_numa_mem(local_memory_node(numa_cpu_lookup_table[cpu]));
1294 notify_cpu_starting(cpu);
1295 set_cpu_online(cpu, true);
1297 boot_init_stack_canary();
1301 /* We can enable ftrace for secondary cpus now */
1302 this_cpu_enable_ftrace();
1304 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
1309 int setup_profiling_timer(unsigned int multiplier)
1314 #ifdef CONFIG_SCHED_SMT
1315 /* cpumask of CPUs with asymetric SMT dependancy */
1316 static int powerpc_smt_flags(void)
1318 int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
1320 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
1321 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
1322 flags |= SD_ASYM_PACKING;
1328 static struct sched_domain_topology_level powerpc_topology[] = {
1329 #ifdef CONFIG_SCHED_SMT
1330 { cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
1332 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1337 * P9 has a slightly odd architecture where pairs of cores share an L2 cache.
1338 * This topology makes it *much* cheaper to migrate tasks between adjacent cores
1339 * since the migrated task remains cache hot. We want to take advantage of this
1340 * at the scheduler level so an extra topology level is required.
1342 static int powerpc_shared_cache_flags(void)
1344 return SD_SHARE_PKG_RESOURCES;
1348 * We can't just pass cpu_l2_cache_mask() directly because
1349 * returns a non-const pointer and the compiler barfs on that.
1351 static const struct cpumask *shared_cache_mask(int cpu)
1353 return cpu_l2_cache_mask(cpu);
1356 #ifdef CONFIG_SCHED_SMT
1357 static const struct cpumask *smallcore_smt_mask(int cpu)
1359 return cpu_smallcore_mask(cpu);
1363 static struct sched_domain_topology_level power9_topology[] = {
1364 #ifdef CONFIG_SCHED_SMT
1365 { cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
1367 { shared_cache_mask, powerpc_shared_cache_flags, SD_INIT_NAME(CACHE) },
1368 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1372 void __init smp_cpus_done(unsigned int max_cpus)
1375 * We are running pinned to the boot CPU, see rest_init().
1377 if (smp_ops && smp_ops->setup_cpu)
1378 smp_ops->setup_cpu(boot_cpuid);
1380 if (smp_ops && smp_ops->bringup_done)
1381 smp_ops->bringup_done();
1383 dump_numa_cpu_topology();
1385 #ifdef CONFIG_SCHED_SMT
1386 if (has_big_cores) {
1387 pr_info("Big cores detected but using small core scheduling\n");
1388 power9_topology[0].mask = smallcore_smt_mask;
1389 powerpc_topology[0].mask = smallcore_smt_mask;
1393 * If any CPU detects that it's sharing a cache with another CPU then
1394 * use the deeper topology that is aware of this sharing.
1396 if (shared_caches) {
1397 pr_info("Using shared cache scheduler topology\n");
1398 set_sched_topology(power9_topology);
1400 pr_info("Using standard scheduler topology\n");
1401 set_sched_topology(powerpc_topology);
1405 #ifdef CONFIG_HOTPLUG_CPU
1406 int __cpu_disable(void)
1408 int cpu = smp_processor_id();
1411 if (!smp_ops->cpu_disable)
1414 this_cpu_disable_ftrace();
1416 err = smp_ops->cpu_disable();
1420 /* Update sibling maps */
1421 remove_cpu_from_masks(cpu);
1426 void __cpu_die(unsigned int cpu)
1428 if (smp_ops->cpu_die)
1429 smp_ops->cpu_die(cpu);
1435 * Disable on the down path. This will be re-enabled by
1436 * start_secondary() via start_secondary_resume() below
1438 this_cpu_disable_ftrace();
1443 /* If we return, we re-enter start_secondary */
1444 start_secondary_resume();