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>
37 #include <linux/clockchips.h>
38 #include <linux/kexec.h>
40 #include <asm/ptrace.h>
41 #include <linux/atomic.h>
43 #include <asm/hw_irq.h>
44 #include <asm/kvm_ppc.h>
45 #include <asm/dbell.h>
49 #include <asm/machdep.h>
50 #include <asm/mmu_context.h>
51 #include <asm/cputhreads.h>
52 #include <asm/cputable.h>
54 #include <asm/vdso_datapage.h>
59 #include <asm/debug.h>
60 #include <asm/cpu_has_feature.h>
61 #include <asm/ftrace.h>
63 #include <asm/fadump.h>
65 #include <trace/events/ipi.h>
69 #define DBG(fmt...) udbg_printf(fmt)
74 #ifdef CONFIG_HOTPLUG_CPU
75 /* State of each CPU during hotplug phases */
76 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
79 struct task_struct *secondary_current;
81 bool coregroup_enabled;
82 bool thread_group_shares_l2;
83 bool thread_group_shares_l3;
85 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
86 DEFINE_PER_CPU(cpumask_var_t, cpu_smallcore_map);
87 DEFINE_PER_CPU(cpumask_var_t, cpu_l2_cache_map);
88 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
89 static DEFINE_PER_CPU(cpumask_var_t, cpu_coregroup_map);
91 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
92 EXPORT_PER_CPU_SYMBOL(cpu_l2_cache_map);
93 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
94 EXPORT_SYMBOL_GPL(has_big_cores);
97 #ifdef CONFIG_SCHED_SMT
105 #define MAX_THREAD_LIST_SIZE 8
106 #define THREAD_GROUP_SHARE_L1 1
107 #define THREAD_GROUP_SHARE_L2_L3 2
108 struct thread_groups {
109 unsigned int property;
110 unsigned int nr_groups;
111 unsigned int threads_per_group;
112 unsigned int thread_list[MAX_THREAD_LIST_SIZE];
115 /* Maximum number of properties that groups of threads within a core can share */
116 #define MAX_THREAD_GROUP_PROPERTIES 2
118 struct thread_groups_list {
119 unsigned int nr_properties;
120 struct thread_groups property_tgs[MAX_THREAD_GROUP_PROPERTIES];
123 static struct thread_groups_list tgl[NR_CPUS] __initdata;
125 * On big-cores system, thread_group_l1_cache_map for each CPU corresponds to
126 * the set its siblings that share the L1-cache.
128 DEFINE_PER_CPU(cpumask_var_t, thread_group_l1_cache_map);
131 * On some big-cores system, thread_group_l2_cache_map for each CPU
132 * corresponds to the set its siblings within the core that share the
135 DEFINE_PER_CPU(cpumask_var_t, thread_group_l2_cache_map);
138 * On P10, thread_group_l3_cache_map for each CPU is equal to the
139 * thread_group_l2_cache_map
141 DEFINE_PER_CPU(cpumask_var_t, thread_group_l3_cache_map);
143 /* SMP operations for this machine */
144 struct smp_ops_t *smp_ops;
146 /* Can't be static due to PowerMac hackery */
147 volatile unsigned int cpu_callin_map[NR_CPUS];
149 int smt_enabled_at_boot = 1;
152 * Returns 1 if the specified cpu should be brought up during boot.
153 * Used to inhibit booting threads if they've been disabled or
154 * limited on the command line
156 int smp_generic_cpu_bootable(unsigned int nr)
158 /* Special case - we inhibit secondary thread startup
159 * during boot if the user requests it.
161 if (system_state < SYSTEM_RUNNING && cpu_has_feature(CPU_FTR_SMT)) {
162 if (!smt_enabled_at_boot && cpu_thread_in_core(nr) != 0)
164 if (smt_enabled_at_boot
165 && cpu_thread_in_core(nr) >= smt_enabled_at_boot)
174 int smp_generic_kick_cpu(int nr)
176 if (nr < 0 || nr >= nr_cpu_ids)
180 * The processor is currently spinning, waiting for the
181 * cpu_start field to become non-zero After we set cpu_start,
182 * the processor will continue on to secondary_start
184 if (!paca_ptrs[nr]->cpu_start) {
185 paca_ptrs[nr]->cpu_start = 1;
190 #ifdef CONFIG_HOTPLUG_CPU
192 * Ok it's not there, so it might be soft-unplugged, let's
193 * try to bring it back
195 generic_set_cpu_up(nr);
197 smp_send_reschedule(nr);
198 #endif /* CONFIG_HOTPLUG_CPU */
202 #endif /* CONFIG_PPC64 */
204 static irqreturn_t call_function_action(int irq, void *data)
206 generic_smp_call_function_interrupt();
210 static irqreturn_t reschedule_action(int irq, void *data)
216 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
217 static irqreturn_t tick_broadcast_ipi_action(int irq, void *data)
219 timer_broadcast_interrupt();
224 #ifdef CONFIG_NMI_IPI
225 static irqreturn_t nmi_ipi_action(int irq, void *data)
227 smp_handle_nmi_ipi(get_irq_regs());
232 static irq_handler_t smp_ipi_action[] = {
233 [PPC_MSG_CALL_FUNCTION] = call_function_action,
234 [PPC_MSG_RESCHEDULE] = reschedule_action,
235 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
236 [PPC_MSG_TICK_BROADCAST] = tick_broadcast_ipi_action,
238 #ifdef CONFIG_NMI_IPI
239 [PPC_MSG_NMI_IPI] = nmi_ipi_action,
244 * The NMI IPI is a fallback and not truly non-maskable. It is simpler
245 * than going through the call function infrastructure, and strongly
246 * serialized, so it is more appropriate for debugging.
248 const char *smp_ipi_name[] = {
249 [PPC_MSG_CALL_FUNCTION] = "ipi call function",
250 [PPC_MSG_RESCHEDULE] = "ipi reschedule",
251 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
252 [PPC_MSG_TICK_BROADCAST] = "ipi tick-broadcast",
254 #ifdef CONFIG_NMI_IPI
255 [PPC_MSG_NMI_IPI] = "nmi ipi",
259 /* optional function to request ipi, for controllers with >= 4 ipis */
260 int smp_request_message_ipi(int virq, int msg)
264 if (msg < 0 || msg > PPC_MSG_NMI_IPI)
266 #ifndef CONFIG_NMI_IPI
267 if (msg == PPC_MSG_NMI_IPI)
271 err = request_irq(virq, smp_ipi_action[msg],
272 IRQF_PERCPU | IRQF_NO_THREAD | IRQF_NO_SUSPEND,
273 smp_ipi_name[msg], NULL);
274 WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
275 virq, smp_ipi_name[msg], err);
280 #ifdef CONFIG_PPC_SMP_MUXED_IPI
281 struct cpu_messages {
282 long messages; /* current messages */
284 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
286 void smp_muxed_ipi_set_message(int cpu, int msg)
288 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
289 char *message = (char *)&info->messages;
292 * Order previous accesses before accesses in the IPI handler.
295 WRITE_ONCE(message[msg], 1);
298 void smp_muxed_ipi_message_pass(int cpu, int msg)
300 smp_muxed_ipi_set_message(cpu, msg);
303 * cause_ipi functions are required to include a full barrier
304 * before doing whatever causes the IPI.
306 smp_ops->cause_ipi(cpu);
309 #ifdef __BIG_ENDIAN__
310 #define IPI_MESSAGE(A) (1uL << ((BITS_PER_LONG - 8) - 8 * (A)))
312 #define IPI_MESSAGE(A) (1uL << (8 * (A)))
315 irqreturn_t smp_ipi_demux(void)
317 mb(); /* order any irq clear */
319 return smp_ipi_demux_relaxed();
322 /* sync-free variant. Callers should ensure synchronization */
323 irqreturn_t smp_ipi_demux_relaxed(void)
325 struct cpu_messages *info;
328 info = this_cpu_ptr(&ipi_message);
330 all = xchg(&info->messages, 0);
331 #if defined(CONFIG_KVM_XICS) && defined(CONFIG_KVM_BOOK3S_HV_POSSIBLE)
333 * Must check for PPC_MSG_RM_HOST_ACTION messages
334 * before PPC_MSG_CALL_FUNCTION messages because when
335 * a VM is destroyed, we call kick_all_cpus_sync()
336 * to ensure that any pending PPC_MSG_RM_HOST_ACTION
337 * messages have completed before we free any VCPUs.
339 if (all & IPI_MESSAGE(PPC_MSG_RM_HOST_ACTION))
340 kvmppc_xics_ipi_action();
342 if (all & IPI_MESSAGE(PPC_MSG_CALL_FUNCTION))
343 generic_smp_call_function_interrupt();
344 if (all & IPI_MESSAGE(PPC_MSG_RESCHEDULE))
346 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
347 if (all & IPI_MESSAGE(PPC_MSG_TICK_BROADCAST))
348 timer_broadcast_interrupt();
350 #ifdef CONFIG_NMI_IPI
351 if (all & IPI_MESSAGE(PPC_MSG_NMI_IPI))
352 nmi_ipi_action(0, NULL);
354 } while (READ_ONCE(info->messages));
358 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
360 static inline void do_message_pass(int cpu, int msg)
362 if (smp_ops->message_pass)
363 smp_ops->message_pass(cpu, msg);
364 #ifdef CONFIG_PPC_SMP_MUXED_IPI
366 smp_muxed_ipi_message_pass(cpu, msg);
370 void arch_smp_send_reschedule(int cpu)
373 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
375 EXPORT_SYMBOL_GPL(arch_smp_send_reschedule);
377 void arch_send_call_function_single_ipi(int cpu)
379 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
382 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
386 for_each_cpu(cpu, mask)
387 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
390 #ifdef CONFIG_NMI_IPI
395 * NMI IPIs may not be recoverable, so should not be used as ongoing part of
396 * a running system. They can be used for crash, debug, halt/reboot, etc.
398 * The IPI call waits with interrupts disabled until all targets enter the
399 * NMI handler, then returns. Subsequent IPIs can be issued before targets
400 * have returned from their handlers, so there is no guarantee about
401 * concurrency or re-entrancy.
403 * A new NMI can be issued before all targets exit the handler.
405 * The IPI call may time out without all targets entering the NMI handler.
406 * In that case, there is some logic to recover (and ignore subsequent
407 * NMI interrupts that may eventually be raised), but the platform interrupt
408 * handler may not be able to distinguish this from other exception causes,
409 * which may cause a crash.
412 static atomic_t __nmi_ipi_lock = ATOMIC_INIT(0);
413 static struct cpumask nmi_ipi_pending_mask;
414 static bool nmi_ipi_busy = false;
415 static void (*nmi_ipi_function)(struct pt_regs *) = NULL;
417 noinstr static void nmi_ipi_lock_start(unsigned long *flags)
419 raw_local_irq_save(*flags);
421 while (raw_atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1) {
422 raw_local_irq_restore(*flags);
423 spin_until_cond(raw_atomic_read(&__nmi_ipi_lock) == 0);
424 raw_local_irq_save(*flags);
429 noinstr static void nmi_ipi_lock(void)
431 while (raw_atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1)
432 spin_until_cond(raw_atomic_read(&__nmi_ipi_lock) == 0);
435 noinstr static void nmi_ipi_unlock(void)
438 WARN_ON(raw_atomic_read(&__nmi_ipi_lock) != 1);
439 raw_atomic_set(&__nmi_ipi_lock, 0);
442 noinstr static void nmi_ipi_unlock_end(unsigned long *flags)
445 raw_local_irq_restore(*flags);
449 * Platform NMI handler calls this to ack
451 noinstr int smp_handle_nmi_ipi(struct pt_regs *regs)
453 void (*fn)(struct pt_regs *) = NULL;
455 int me = raw_smp_processor_id();
459 * Unexpected NMIs are possible here because the interrupt may not
460 * be able to distinguish NMI IPIs from other types of NMIs, or
461 * because the caller may have timed out.
463 nmi_ipi_lock_start(&flags);
464 if (cpumask_test_cpu(me, &nmi_ipi_pending_mask)) {
465 cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
466 fn = READ_ONCE(nmi_ipi_function);
470 nmi_ipi_unlock_end(&flags);
478 static void do_smp_send_nmi_ipi(int cpu, bool safe)
480 if (!safe && smp_ops->cause_nmi_ipi && smp_ops->cause_nmi_ipi(cpu))
484 do_message_pass(cpu, PPC_MSG_NMI_IPI);
488 for_each_online_cpu(c) {
489 if (c == raw_smp_processor_id())
491 do_message_pass(c, PPC_MSG_NMI_IPI);
497 * - cpu is the target CPU (must not be this CPU), or NMI_IPI_ALL_OTHERS.
498 * - fn is the target callback function.
499 * - delay_us > 0 is the delay before giving up waiting for targets to
500 * begin executing the handler, == 0 specifies indefinite delay.
502 static int __smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *),
503 u64 delay_us, bool safe)
506 int me = raw_smp_processor_id();
510 BUG_ON(cpu < 0 && cpu != NMI_IPI_ALL_OTHERS);
512 if (unlikely(!smp_ops))
515 nmi_ipi_lock_start(&flags);
516 while (nmi_ipi_busy) {
517 nmi_ipi_unlock_end(&flags);
518 spin_until_cond(!nmi_ipi_busy);
519 nmi_ipi_lock_start(&flags);
522 nmi_ipi_function = fn;
524 WARN_ON_ONCE(!cpumask_empty(&nmi_ipi_pending_mask));
528 cpumask_copy(&nmi_ipi_pending_mask, cpu_online_mask);
529 cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
531 cpumask_set_cpu(cpu, &nmi_ipi_pending_mask);
536 /* Interrupts remain hard disabled */
538 do_smp_send_nmi_ipi(cpu, safe);
541 /* nmi_ipi_busy is set here, so unlock/lock is okay */
542 while (!cpumask_empty(&nmi_ipi_pending_mask)) {
553 if (!cpumask_empty(&nmi_ipi_pending_mask)) {
554 /* Timeout waiting for CPUs to call smp_handle_nmi_ipi */
556 cpumask_clear(&nmi_ipi_pending_mask);
559 nmi_ipi_function = NULL;
560 nmi_ipi_busy = false;
562 nmi_ipi_unlock_end(&flags);
567 int smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
569 return __smp_send_nmi_ipi(cpu, fn, delay_us, false);
572 int smp_send_safe_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
574 return __smp_send_nmi_ipi(cpu, fn, delay_us, true);
576 #endif /* CONFIG_NMI_IPI */
578 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
579 void tick_broadcast(const struct cpumask *mask)
583 for_each_cpu(cpu, mask)
584 do_message_pass(cpu, PPC_MSG_TICK_BROADCAST);
588 #ifdef CONFIG_DEBUGGER
589 static void debugger_ipi_callback(struct pt_regs *regs)
594 void smp_send_debugger_break(void)
596 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, debugger_ipi_callback, 1000000);
600 #ifdef CONFIG_KEXEC_CORE
601 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
605 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, crash_ipi_callback, 1000000);
606 if (kdump_in_progress() && crash_wake_offline) {
607 for_each_present_cpu(cpu) {
611 * crash_ipi_callback will wait for
612 * all cpus, including offline CPUs.
613 * We don't care about nmi_ipi_function.
614 * Offline cpus will jump straight into
615 * crash_ipi_callback, we can skip the
616 * entire NMI dance and waiting for
617 * cpus to clear pending mask, etc.
619 do_smp_send_nmi_ipi(cpu, false);
625 void crash_smp_send_stop(void)
627 static bool stopped = false;
630 * In case of fadump, register data for all CPUs is captured by f/w
631 * on ibm,os-term rtas call. Skip IPI callbacks to other CPUs before
632 * this rtas call to avoid tricky post processing of those CPUs'
635 if (should_fadump_crash())
643 #ifdef CONFIG_KEXEC_CORE
644 if (kexec_crash_image) {
645 crash_kexec_prepare();
653 #ifdef CONFIG_NMI_IPI
654 static void nmi_stop_this_cpu(struct pt_regs *regs)
657 * IRQs are already hard disabled by the smp_handle_nmi_ipi.
659 set_cpu_online(smp_processor_id(), false);
666 void smp_send_stop(void)
668 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, nmi_stop_this_cpu, 1000000);
671 #else /* CONFIG_NMI_IPI */
673 static void stop_this_cpu(void *dummy)
678 * Offlining CPUs in stop_this_cpu can result in scheduler warnings,
679 * (see commit de6e5d38417e), but printk_safe_flush_on_panic() wants
680 * to know other CPUs are offline before it breaks locks to flush
681 * printk buffers, in case we panic()ed while holding the lock.
683 set_cpu_online(smp_processor_id(), false);
690 void smp_send_stop(void)
692 static bool stopped = false;
695 * Prevent waiting on csd lock from a previous smp_send_stop.
696 * This is racy, but in general callers try to do the right
697 * thing and only fire off one smp_send_stop (e.g., see
705 smp_call_function(stop_this_cpu, NULL, 0);
707 #endif /* CONFIG_NMI_IPI */
709 static struct task_struct *current_set[NR_CPUS];
711 static void smp_store_cpu_info(int id)
713 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
714 #ifdef CONFIG_PPC_E500
715 per_cpu(next_tlbcam_idx, id)
716 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
721 * Relationships between CPUs are maintained in a set of per-cpu cpumasks so
722 * rather than just passing around the cpumask we pass around a function that
723 * returns the that cpumask for the given CPU.
725 static void set_cpus_related(int i, int j, struct cpumask *(*get_cpumask)(int))
727 cpumask_set_cpu(i, get_cpumask(j));
728 cpumask_set_cpu(j, get_cpumask(i));
731 #ifdef CONFIG_HOTPLUG_CPU
732 static void set_cpus_unrelated(int i, int j,
733 struct cpumask *(*get_cpumask)(int))
735 cpumask_clear_cpu(i, get_cpumask(j));
736 cpumask_clear_cpu(j, get_cpumask(i));
741 * Extends set_cpus_related. Instead of setting one CPU at a time in
742 * dstmask, set srcmask at oneshot. dstmask should be super set of srcmask.
744 static void or_cpumasks_related(int i, int j, struct cpumask *(*srcmask)(int),
745 struct cpumask *(*dstmask)(int))
747 struct cpumask *mask;
751 for_each_cpu(k, srcmask(i))
752 cpumask_or(dstmask(k), dstmask(k), mask);
758 for_each_cpu(k, srcmask(j))
759 cpumask_or(dstmask(k), dstmask(k), mask);
763 * parse_thread_groups: Parses the "ibm,thread-groups" device tree
764 * property for the CPU device node @dn and stores
765 * the parsed output in the thread_groups_list
768 * @dn: The device node of the CPU device.
769 * @tglp: Pointer to a thread group list structure into which the parsed
770 * output of "ibm,thread-groups" is stored.
772 * ibm,thread-groups[0..N-1] array defines which group of threads in
773 * the CPU-device node can be grouped together based on the property.
775 * This array can represent thread groupings for multiple properties.
777 * ibm,thread-groups[i + 0] tells us the property based on which the
778 * threads are being grouped together. If this value is 1, it implies
779 * that the threads in the same group share L1, translation cache. If
780 * the value is 2, it implies that the threads in the same group share
783 * ibm,thread-groups[i+1] tells us how many such thread groups exist for the
784 * property ibm,thread-groups[i]
786 * ibm,thread-groups[i+2] tells us the number of threads in each such
788 * Suppose k = (ibm,thread-groups[i+1] * ibm,thread-groups[i+2]), then,
790 * ibm,thread-groups[i+3..i+k+2] (is the list of threads identified by
791 * "ibm,ppc-interrupt-server#s" arranged as per their membership in
795 * If "ibm,thread-groups" = [1,2,4,8,10,12,14,9,11,13,15,2,2,4,8,10,12,14,9,11,13,15]
796 * This can be decomposed up into two consecutive arrays:
797 * a) [1,2,4,8,10,12,14,9,11,13,15]
798 * b) [2,2,4,8,10,12,14,9,11,13,15]
802 * a) provides information of Property "1" being shared by "2" groups,
803 * each with "4" threads each. The "ibm,ppc-interrupt-server#s" of
804 * the first group is {8,10,12,14} and the
805 * "ibm,ppc-interrupt-server#s" of the second group is
806 * {9,11,13,15}. Property "1" is indicative of the thread in the
807 * group sharing L1 cache, translation cache and Instruction Data
810 * b) provides information of Property "2" being shared by "2" groups,
811 * each group with "4" threads. The "ibm,ppc-interrupt-server#s" of
812 * the first group is {8,10,12,14} and the
813 * "ibm,ppc-interrupt-server#s" of the second group is
814 * {9,11,13,15}. Property "2" indicates that the threads in each
815 * group share the L2-cache.
817 * Returns 0 on success, -EINVAL if the property does not exist,
818 * -ENODATA if property does not have a value, and -EOVERFLOW if the
819 * property data isn't large enough.
821 static int parse_thread_groups(struct device_node *dn,
822 struct thread_groups_list *tglp)
824 unsigned int property_idx = 0;
825 u32 *thread_group_array;
826 size_t total_threads;
831 count = of_property_count_u32_elems(dn, "ibm,thread-groups");
832 thread_group_array = kcalloc(count, sizeof(u32), GFP_KERNEL);
833 ret = of_property_read_u32_array(dn, "ibm,thread-groups",
834 thread_group_array, count);
838 while (i < count && property_idx < MAX_THREAD_GROUP_PROPERTIES) {
840 struct thread_groups *tg = &tglp->property_tgs[property_idx++];
842 tg->property = thread_group_array[i];
843 tg->nr_groups = thread_group_array[i + 1];
844 tg->threads_per_group = thread_group_array[i + 2];
845 total_threads = tg->nr_groups * tg->threads_per_group;
847 thread_list = &thread_group_array[i + 3];
849 for (j = 0; j < total_threads; j++)
850 tg->thread_list[j] = thread_list[j];
851 i = i + 3 + total_threads;
854 tglp->nr_properties = property_idx;
857 kfree(thread_group_array);
862 * get_cpu_thread_group_start : Searches the thread group in tg->thread_list
863 * that @cpu belongs to.
865 * @cpu : The logical CPU whose thread group is being searched.
866 * @tg : The thread-group structure of the CPU node which @cpu belongs
869 * Returns the index to tg->thread_list that points to the start
870 * of the thread_group that @cpu belongs to.
872 * Returns -1 if cpu doesn't belong to any of the groups pointed to by
875 static int get_cpu_thread_group_start(int cpu, struct thread_groups *tg)
877 int hw_cpu_id = get_hard_smp_processor_id(cpu);
880 for (i = 0; i < tg->nr_groups; i++) {
881 int group_start = i * tg->threads_per_group;
883 for (j = 0; j < tg->threads_per_group; j++) {
884 int idx = group_start + j;
886 if (tg->thread_list[idx] == hw_cpu_id)
894 static struct thread_groups *__init get_thread_groups(int cpu,
898 struct device_node *dn = of_get_cpu_node(cpu, NULL);
899 struct thread_groups_list *cpu_tgl = &tgl[cpu];
900 struct thread_groups *tg = NULL;
909 if (!cpu_tgl->nr_properties) {
910 *err = parse_thread_groups(dn, cpu_tgl);
915 for (i = 0; i < cpu_tgl->nr_properties; i++) {
916 if (cpu_tgl->property_tgs[i].property == group_property) {
917 tg = &cpu_tgl->property_tgs[i];
929 static int __init update_mask_from_threadgroup(cpumask_var_t *mask, struct thread_groups *tg,
930 int cpu, int cpu_group_start)
932 int first_thread = cpu_first_thread_sibling(cpu);
935 zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cpu));
937 for (i = first_thread; i < first_thread + threads_per_core; i++) {
938 int i_group_start = get_cpu_thread_group_start(i, tg);
940 if (unlikely(i_group_start == -1)) {
945 if (i_group_start == cpu_group_start)
946 cpumask_set_cpu(i, *mask);
952 static int __init init_thread_group_cache_map(int cpu, int cache_property)
955 int cpu_group_start = -1, err = 0;
956 struct thread_groups *tg = NULL;
957 cpumask_var_t *mask = NULL;
959 if (cache_property != THREAD_GROUP_SHARE_L1 &&
960 cache_property != THREAD_GROUP_SHARE_L2_L3)
963 tg = get_thread_groups(cpu, cache_property, &err);
968 cpu_group_start = get_cpu_thread_group_start(cpu, tg);
970 if (unlikely(cpu_group_start == -1)) {
975 if (cache_property == THREAD_GROUP_SHARE_L1) {
976 mask = &per_cpu(thread_group_l1_cache_map, cpu);
977 update_mask_from_threadgroup(mask, tg, cpu, cpu_group_start);
979 else if (cache_property == THREAD_GROUP_SHARE_L2_L3) {
980 mask = &per_cpu(thread_group_l2_cache_map, cpu);
981 update_mask_from_threadgroup(mask, tg, cpu, cpu_group_start);
982 mask = &per_cpu(thread_group_l3_cache_map, cpu);
983 update_mask_from_threadgroup(mask, tg, cpu, cpu_group_start);
990 static bool shared_caches;
992 #ifdef CONFIG_SCHED_SMT
993 /* cpumask of CPUs with asymmetric SMT dependency */
994 static int powerpc_smt_flags(void)
996 int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
998 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
999 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
1000 flags |= SD_ASYM_PACKING;
1007 * P9 has a slightly odd architecture where pairs of cores share an L2 cache.
1008 * This topology makes it *much* cheaper to migrate tasks between adjacent cores
1009 * since the migrated task remains cache hot. We want to take advantage of this
1010 * at the scheduler level so an extra topology level is required.
1012 static int powerpc_shared_cache_flags(void)
1014 return SD_SHARE_PKG_RESOURCES;
1018 * We can't just pass cpu_l2_cache_mask() directly because
1019 * returns a non-const pointer and the compiler barfs on that.
1021 static const struct cpumask *shared_cache_mask(int cpu)
1023 return per_cpu(cpu_l2_cache_map, cpu);
1026 #ifdef CONFIG_SCHED_SMT
1027 static const struct cpumask *smallcore_smt_mask(int cpu)
1029 return cpu_smallcore_mask(cpu);
1033 static struct cpumask *cpu_coregroup_mask(int cpu)
1035 return per_cpu(cpu_coregroup_map, cpu);
1038 static bool has_coregroup_support(void)
1040 return coregroup_enabled;
1043 static const struct cpumask *cpu_mc_mask(int cpu)
1045 return cpu_coregroup_mask(cpu);
1048 static struct sched_domain_topology_level powerpc_topology[] = {
1049 #ifdef CONFIG_SCHED_SMT
1050 { cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
1052 { shared_cache_mask, powerpc_shared_cache_flags, SD_INIT_NAME(CACHE) },
1053 { cpu_mc_mask, SD_INIT_NAME(MC) },
1054 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1058 static int __init init_big_cores(void)
1062 for_each_possible_cpu(cpu) {
1063 int err = init_thread_group_cache_map(cpu, THREAD_GROUP_SHARE_L1);
1068 zalloc_cpumask_var_node(&per_cpu(cpu_smallcore_map, cpu),
1073 has_big_cores = true;
1075 for_each_possible_cpu(cpu) {
1076 int err = init_thread_group_cache_map(cpu, THREAD_GROUP_SHARE_L2_L3);
1082 thread_group_shares_l2 = true;
1083 thread_group_shares_l3 = true;
1084 pr_debug("L2/L3 cache only shared by the threads in the small core\n");
1089 void __init smp_prepare_cpus(unsigned int max_cpus)
1091 unsigned int cpu, num_threads;
1093 DBG("smp_prepare_cpus\n");
1096 * setup_cpu may need to be called on the boot cpu. We haven't
1097 * spun any cpus up but lets be paranoid.
1099 BUG_ON(boot_cpuid != smp_processor_id());
1101 /* Fixup boot cpu */
1102 smp_store_cpu_info(boot_cpuid);
1103 cpu_callin_map[boot_cpuid] = 1;
1105 for_each_possible_cpu(cpu) {
1106 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
1107 GFP_KERNEL, cpu_to_node(cpu));
1108 zalloc_cpumask_var_node(&per_cpu(cpu_l2_cache_map, cpu),
1109 GFP_KERNEL, cpu_to_node(cpu));
1110 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
1111 GFP_KERNEL, cpu_to_node(cpu));
1112 if (has_coregroup_support())
1113 zalloc_cpumask_var_node(&per_cpu(cpu_coregroup_map, cpu),
1114 GFP_KERNEL, cpu_to_node(cpu));
1118 * numa_node_id() works after this.
1120 if (cpu_present(cpu)) {
1121 set_cpu_numa_node(cpu, numa_cpu_lookup_table[cpu]);
1122 set_cpu_numa_mem(cpu,
1123 local_memory_node(numa_cpu_lookup_table[cpu]));
1128 /* Init the cpumasks so the boot CPU is related to itself */
1129 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
1130 cpumask_set_cpu(boot_cpuid, cpu_l2_cache_mask(boot_cpuid));
1131 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
1133 if (has_coregroup_support())
1134 cpumask_set_cpu(boot_cpuid, cpu_coregroup_mask(boot_cpuid));
1137 if (has_big_cores) {
1138 cpumask_set_cpu(boot_cpuid,
1139 cpu_smallcore_mask(boot_cpuid));
1142 if (cpu_to_chip_id(boot_cpuid) != -1) {
1143 int idx = DIV_ROUND_UP(num_possible_cpus(), threads_per_core);
1146 * All threads of a core will all belong to the same core,
1147 * chip_id_lookup_table will have one entry per core.
1148 * Assumption: if boot_cpuid doesn't have a chip-id, then no
1149 * other CPUs, will also not have chip-id.
1151 chip_id_lookup_table = kcalloc(idx, sizeof(int), GFP_KERNEL);
1152 if (chip_id_lookup_table)
1153 memset(chip_id_lookup_table, -1, sizeof(int) * idx);
1156 if (smp_ops && smp_ops->probe)
1159 // Initalise the generic SMT topology support
1161 if (smt_enabled_at_boot)
1162 num_threads = smt_enabled_at_boot;
1163 cpu_smt_set_num_threads(num_threads, threads_per_core);
1166 void smp_prepare_boot_cpu(void)
1168 BUG_ON(smp_processor_id() != boot_cpuid);
1170 paca_ptrs[boot_cpuid]->__current = current;
1172 set_numa_node(numa_cpu_lookup_table[boot_cpuid]);
1173 current_set[boot_cpuid] = current;
1176 #ifdef CONFIG_HOTPLUG_CPU
1178 int generic_cpu_disable(void)
1180 unsigned int cpu = smp_processor_id();
1182 if (cpu == boot_cpuid)
1185 set_cpu_online(cpu, false);
1187 vdso_data->processorCount--;
1189 /* Update affinity of all IRQs previously aimed at this CPU */
1190 irq_migrate_all_off_this_cpu();
1193 * Depending on the details of the interrupt controller, it's possible
1194 * that one of the interrupts we just migrated away from this CPU is
1195 * actually already pending on this CPU. If we leave it in that state
1196 * the interrupt will never be EOI'ed, and will never fire again. So
1197 * temporarily enable interrupts here, to allow any pending interrupt to
1198 * be received (and EOI'ed), before we take this CPU offline.
1202 local_irq_disable();
1207 void generic_cpu_die(unsigned int cpu)
1211 for (i = 0; i < 100; i++) {
1213 if (is_cpu_dead(cpu))
1217 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
1220 void generic_set_cpu_dead(unsigned int cpu)
1222 per_cpu(cpu_state, cpu) = CPU_DEAD;
1226 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
1227 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
1228 * which makes the delay in generic_cpu_die() not happen.
1230 void generic_set_cpu_up(unsigned int cpu)
1232 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1235 int generic_check_cpu_restart(unsigned int cpu)
1237 return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
1240 int is_cpu_dead(unsigned int cpu)
1242 return per_cpu(cpu_state, cpu) == CPU_DEAD;
1245 static bool secondaries_inhibited(void)
1247 return kvm_hv_mode_active();
1250 #else /* HOTPLUG_CPU */
1252 #define secondaries_inhibited() 0
1256 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
1259 paca_ptrs[cpu]->__current = idle;
1260 paca_ptrs[cpu]->kstack = (unsigned long)task_stack_page(idle) +
1261 THREAD_SIZE - STACK_FRAME_MIN_SIZE;
1263 task_thread_info(idle)->cpu = cpu;
1264 secondary_current = current_set[cpu] = idle;
1267 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
1269 const unsigned long boot_spin_ms = 5 * MSEC_PER_SEC;
1270 const bool booting = system_state < SYSTEM_RUNNING;
1271 const unsigned long hp_spin_ms = 1;
1272 unsigned long deadline;
1274 const unsigned long spin_wait_ms = booting ? boot_spin_ms : hp_spin_ms;
1277 * Don't allow secondary threads to come online if inhibited
1279 if (threads_per_core > 1 && secondaries_inhibited() &&
1280 cpu_thread_in_subcore(cpu))
1283 if (smp_ops == NULL ||
1284 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
1287 cpu_idle_thread_init(cpu, tidle);
1290 * The platform might need to allocate resources prior to bringing
1293 if (smp_ops->prepare_cpu) {
1294 rc = smp_ops->prepare_cpu(cpu);
1299 /* Make sure callin-map entry is 0 (can be leftover a CPU
1302 cpu_callin_map[cpu] = 0;
1304 /* The information for processor bringup must
1305 * be written out to main store before we release
1311 DBG("smp: kicking cpu %d\n", cpu);
1312 rc = smp_ops->kick_cpu(cpu);
1314 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
1319 * At boot time, simply spin on the callin word until the
1322 * At run time, spin for an optimistic amount of time to avoid
1323 * sleeping in the common case.
1325 deadline = jiffies + msecs_to_jiffies(spin_wait_ms);
1326 spin_until_cond(cpu_callin_map[cpu] || time_is_before_jiffies(deadline));
1328 if (!cpu_callin_map[cpu] && system_state >= SYSTEM_RUNNING) {
1329 const unsigned long sleep_interval_us = 10 * USEC_PER_MSEC;
1330 const unsigned long sleep_wait_ms = 100 * MSEC_PER_SEC;
1332 deadline = jiffies + msecs_to_jiffies(sleep_wait_ms);
1333 while (!cpu_callin_map[cpu] && time_is_after_jiffies(deadline))
1334 fsleep(sleep_interval_us);
1337 if (!cpu_callin_map[cpu]) {
1338 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
1342 DBG("Processor %u found.\n", cpu);
1344 if (smp_ops->give_timebase)
1345 smp_ops->give_timebase();
1347 /* Wait until cpu puts itself in the online & active maps */
1348 spin_until_cond(cpu_online(cpu));
1353 /* Return the value of the reg property corresponding to the given
1356 int cpu_to_core_id(int cpu)
1358 struct device_node *np;
1361 np = of_get_cpu_node(cpu, NULL);
1365 id = of_get_cpu_hwid(np, 0);
1370 EXPORT_SYMBOL_GPL(cpu_to_core_id);
1372 /* Helper routines for cpu to core mapping */
1373 int cpu_core_index_of_thread(int cpu)
1375 return cpu >> threads_shift;
1377 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
1379 int cpu_first_thread_of_core(int core)
1381 return core << threads_shift;
1383 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
1385 /* Must be called when no change can occur to cpu_present_mask,
1386 * i.e. during cpu online or offline.
1388 static struct device_node *cpu_to_l2cache(int cpu)
1390 struct device_node *np;
1391 struct device_node *cache;
1393 if (!cpu_present(cpu))
1396 np = of_get_cpu_node(cpu, NULL);
1400 cache = of_find_next_cache_node(np);
1407 static bool update_mask_by_l2(int cpu, cpumask_var_t *mask)
1409 struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1410 struct device_node *l2_cache, *np;
1414 submask_fn = cpu_smallcore_mask;
1417 * If the threads in a thread-group share L2 cache, then the
1418 * L2-mask can be obtained from thread_group_l2_cache_map.
1420 if (thread_group_shares_l2) {
1421 cpumask_set_cpu(cpu, cpu_l2_cache_mask(cpu));
1423 for_each_cpu(i, per_cpu(thread_group_l2_cache_map, cpu)) {
1425 set_cpus_related(i, cpu, cpu_l2_cache_mask);
1428 /* Verify that L1-cache siblings are a subset of L2 cache-siblings */
1429 if (!cpumask_equal(submask_fn(cpu), cpu_l2_cache_mask(cpu)) &&
1430 !cpumask_subset(submask_fn(cpu), cpu_l2_cache_mask(cpu))) {
1431 pr_warn_once("CPU %d : Inconsistent L1 and L2 cache siblings\n",
1438 l2_cache = cpu_to_l2cache(cpu);
1439 if (!l2_cache || !*mask) {
1440 /* Assume only core siblings share cache with this CPU */
1441 for_each_cpu(i, cpu_sibling_mask(cpu))
1442 set_cpus_related(cpu, i, cpu_l2_cache_mask);
1447 cpumask_and(*mask, cpu_online_mask, cpu_cpu_mask(cpu));
1449 /* Update l2-cache mask with all the CPUs that are part of submask */
1450 or_cpumasks_related(cpu, cpu, submask_fn, cpu_l2_cache_mask);
1452 /* Skip all CPUs already part of current CPU l2-cache mask */
1453 cpumask_andnot(*mask, *mask, cpu_l2_cache_mask(cpu));
1455 for_each_cpu(i, *mask) {
1457 * when updating the marks the current CPU has not been marked
1458 * online, but we need to update the cache masks
1460 np = cpu_to_l2cache(i);
1462 /* Skip all CPUs already part of current CPU l2-cache */
1463 if (np == l2_cache) {
1464 or_cpumasks_related(cpu, i, submask_fn, cpu_l2_cache_mask);
1465 cpumask_andnot(*mask, *mask, submask_fn(i));
1467 cpumask_andnot(*mask, *mask, cpu_l2_cache_mask(i));
1472 of_node_put(l2_cache);
1477 #ifdef CONFIG_HOTPLUG_CPU
1478 static void remove_cpu_from_masks(int cpu)
1480 struct cpumask *(*mask_fn)(int) = cpu_sibling_mask;
1483 unmap_cpu_from_node(cpu);
1486 mask_fn = cpu_l2_cache_mask;
1488 for_each_cpu(i, mask_fn(cpu)) {
1489 set_cpus_unrelated(cpu, i, cpu_l2_cache_mask);
1490 set_cpus_unrelated(cpu, i, cpu_sibling_mask);
1492 set_cpus_unrelated(cpu, i, cpu_smallcore_mask);
1495 for_each_cpu(i, cpu_core_mask(cpu))
1496 set_cpus_unrelated(cpu, i, cpu_core_mask);
1498 if (has_coregroup_support()) {
1499 for_each_cpu(i, cpu_coregroup_mask(cpu))
1500 set_cpus_unrelated(cpu, i, cpu_coregroup_mask);
1505 static inline void add_cpu_to_smallcore_masks(int cpu)
1512 cpumask_set_cpu(cpu, cpu_smallcore_mask(cpu));
1514 for_each_cpu(i, per_cpu(thread_group_l1_cache_map, cpu)) {
1516 set_cpus_related(i, cpu, cpu_smallcore_mask);
1520 static void update_coregroup_mask(int cpu, cpumask_var_t *mask)
1522 struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1523 int coregroup_id = cpu_to_coregroup_id(cpu);
1527 submask_fn = cpu_l2_cache_mask;
1530 /* Assume only siblings are part of this CPU's coregroup */
1531 for_each_cpu(i, submask_fn(cpu))
1532 set_cpus_related(cpu, i, cpu_coregroup_mask);
1537 cpumask_and(*mask, cpu_online_mask, cpu_cpu_mask(cpu));
1539 /* Update coregroup mask with all the CPUs that are part of submask */
1540 or_cpumasks_related(cpu, cpu, submask_fn, cpu_coregroup_mask);
1542 /* Skip all CPUs already part of coregroup mask */
1543 cpumask_andnot(*mask, *mask, cpu_coregroup_mask(cpu));
1545 for_each_cpu(i, *mask) {
1546 /* Skip all CPUs not part of this coregroup */
1547 if (coregroup_id == cpu_to_coregroup_id(i)) {
1548 or_cpumasks_related(cpu, i, submask_fn, cpu_coregroup_mask);
1549 cpumask_andnot(*mask, *mask, submask_fn(i));
1551 cpumask_andnot(*mask, *mask, cpu_coregroup_mask(i));
1556 static void add_cpu_to_masks(int cpu)
1558 struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1559 int first_thread = cpu_first_thread_sibling(cpu);
1566 * This CPU will not be in the online mask yet so we need to manually
1567 * add it to it's own thread sibling mask.
1569 map_cpu_to_node(cpu, cpu_to_node(cpu));
1570 cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
1571 cpumask_set_cpu(cpu, cpu_core_mask(cpu));
1573 for (i = first_thread; i < first_thread + threads_per_core; i++)
1575 set_cpus_related(i, cpu, cpu_sibling_mask);
1577 add_cpu_to_smallcore_masks(cpu);
1579 /* In CPU-hotplug path, hence use GFP_ATOMIC */
1580 ret = alloc_cpumask_var_node(&mask, GFP_ATOMIC, cpu_to_node(cpu));
1581 update_mask_by_l2(cpu, &mask);
1583 if (has_coregroup_support())
1584 update_coregroup_mask(cpu, &mask);
1586 if (chip_id_lookup_table && ret)
1587 chip_id = cpu_to_chip_id(cpu);
1590 submask_fn = cpu_l2_cache_mask;
1592 /* Update core_mask with all the CPUs that are part of submask */
1593 or_cpumasks_related(cpu, cpu, submask_fn, cpu_core_mask);
1595 /* Skip all CPUs already part of current CPU core mask */
1596 cpumask_andnot(mask, cpu_online_mask, cpu_core_mask(cpu));
1598 /* If chip_id is -1; limit the cpu_core_mask to within DIE*/
1600 cpumask_and(mask, mask, cpu_cpu_mask(cpu));
1602 for_each_cpu(i, mask) {
1603 if (chip_id == cpu_to_chip_id(i)) {
1604 or_cpumasks_related(cpu, i, submask_fn, cpu_core_mask);
1605 cpumask_andnot(mask, mask, submask_fn(i));
1607 cpumask_andnot(mask, mask, cpu_core_mask(i));
1611 free_cpumask_var(mask);
1614 /* Activate a secondary processor. */
1615 __no_stack_protector
1616 void start_secondary(void *unused)
1618 unsigned int cpu = raw_smp_processor_id();
1620 /* PPC64 calls setup_kup() in early_setup_secondary() */
1621 if (IS_ENABLED(CONFIG_PPC32))
1624 mmgrab_lazy_tlb(&init_mm);
1625 current->active_mm = &init_mm;
1626 VM_WARN_ON(cpumask_test_cpu(smp_processor_id(), mm_cpumask(&init_mm)));
1627 cpumask_set_cpu(cpu, mm_cpumask(&init_mm));
1628 inc_mm_active_cpus(&init_mm);
1630 smp_store_cpu_info(cpu);
1631 set_dec(tb_ticks_per_jiffy);
1632 rcu_cpu_starting(cpu);
1633 cpu_callin_map[cpu] = 1;
1635 if (smp_ops->setup_cpu)
1636 smp_ops->setup_cpu(cpu);
1637 if (smp_ops->take_timebase)
1638 smp_ops->take_timebase();
1640 secondary_cpu_time_init();
1643 if (system_state == SYSTEM_RUNNING)
1644 vdso_data->processorCount++;
1648 set_numa_node(numa_cpu_lookup_table[cpu]);
1649 set_numa_mem(local_memory_node(numa_cpu_lookup_table[cpu]));
1651 /* Update topology CPU masks */
1652 add_cpu_to_masks(cpu);
1655 * Check for any shared caches. Note that this must be done on a
1656 * per-core basis because one core in the pair might be disabled.
1658 if (!shared_caches) {
1659 struct cpumask *(*sibling_mask)(int) = cpu_sibling_mask;
1660 struct cpumask *mask = cpu_l2_cache_mask(cpu);
1663 sibling_mask = cpu_smallcore_mask;
1665 if (cpumask_weight(mask) > cpumask_weight(sibling_mask(cpu)))
1666 shared_caches = true;
1670 notify_cpu_starting(cpu);
1671 set_cpu_online(cpu, true);
1673 boot_init_stack_canary();
1677 /* We can enable ftrace for secondary cpus now */
1678 this_cpu_enable_ftrace();
1680 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
1685 static void __init fixup_topology(void)
1689 #ifdef CONFIG_SCHED_SMT
1690 if (has_big_cores) {
1691 pr_info("Big cores detected but using small core scheduling\n");
1692 powerpc_topology[smt_idx].mask = smallcore_smt_mask;
1696 if (!has_coregroup_support())
1697 powerpc_topology[mc_idx].mask = powerpc_topology[cache_idx].mask;
1700 * Try to consolidate topology levels here instead of
1701 * allowing scheduler to degenerate.
1702 * - Dont consolidate if masks are different.
1703 * - Dont consolidate if sd_flags exists and are different.
1705 for (i = 1; i <= die_idx; i++) {
1706 if (powerpc_topology[i].mask != powerpc_topology[i - 1].mask)
1709 if (powerpc_topology[i].sd_flags && powerpc_topology[i - 1].sd_flags &&
1710 powerpc_topology[i].sd_flags != powerpc_topology[i - 1].sd_flags)
1713 if (!powerpc_topology[i - 1].sd_flags)
1714 powerpc_topology[i - 1].sd_flags = powerpc_topology[i].sd_flags;
1716 powerpc_topology[i].mask = powerpc_topology[i + 1].mask;
1717 powerpc_topology[i].sd_flags = powerpc_topology[i + 1].sd_flags;
1718 #ifdef CONFIG_SCHED_DEBUG
1719 powerpc_topology[i].name = powerpc_topology[i + 1].name;
1724 void __init smp_cpus_done(unsigned int max_cpus)
1727 * We are running pinned to the boot CPU, see rest_init().
1729 if (smp_ops && smp_ops->setup_cpu)
1730 smp_ops->setup_cpu(boot_cpuid);
1732 if (smp_ops && smp_ops->bringup_done)
1733 smp_ops->bringup_done();
1735 dump_numa_cpu_topology();
1738 set_sched_topology(powerpc_topology);
1741 #ifdef CONFIG_HOTPLUG_CPU
1742 int __cpu_disable(void)
1744 int cpu = smp_processor_id();
1747 if (!smp_ops->cpu_disable)
1750 this_cpu_disable_ftrace();
1752 err = smp_ops->cpu_disable();
1756 /* Update sibling maps */
1757 remove_cpu_from_masks(cpu);
1762 void __cpu_die(unsigned int cpu)
1765 * This could perhaps be a generic call in idlea_task_dead(), but
1766 * that requires testing from all archs, so first put it here to
1768 VM_WARN_ON_ONCE(!cpumask_test_cpu(cpu, mm_cpumask(&init_mm)));
1769 dec_mm_active_cpus(&init_mm);
1770 cpumask_clear_cpu(cpu, mm_cpumask(&init_mm));
1772 if (smp_ops->cpu_die)
1773 smp_ops->cpu_die(cpu);
1776 void __noreturn arch_cpu_idle_dead(void)
1779 * Disable on the down path. This will be re-enabled by
1780 * start_secondary() via start_secondary_resume() below
1782 this_cpu_disable_ftrace();
1784 if (smp_ops->cpu_offline_self)
1785 smp_ops->cpu_offline_self();
1787 /* If we return, we re-enter start_secondary */
1788 start_secondary_resume();