Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * SMP support for ppc.
4 *
5 * Written by Cort Dougan (cort@cs.nmt.edu) borrowing a great
6 * deal of code from the sparc and intel versions.
7 *
8 * Copyright (C) 1999 Cort Dougan <cort@cs.nmt.edu>
9 *
10 * PowerPC-64 Support added by Dave Engebretsen, Peter Bergner, and
11 * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
12 */
13
14#undef DEBUG
15
16#include <linux/kernel.h>
17#include <linux/export.h>
18#include <linux/sched/mm.h>
19#include <linux/sched/task_stack.h>
20#include <linux/sched/topology.h>
21#include <linux/smp.h>
22#include <linux/interrupt.h>
23#include <linux/delay.h>
24#include <linux/init.h>
25#include <linux/spinlock.h>
26#include <linux/cache.h>
27#include <linux/err.h>
28#include <linux/device.h>
29#include <linux/cpu.h>
30#include <linux/notifier.h>
31#include <linux/topology.h>
32#include <linux/profile.h>
33#include <linux/processor.h>
34#include <linux/random.h>
35#include <linux/stackprotector.h>
36#include <linux/pgtable.h>
37
38#include <asm/ptrace.h>
39#include <linux/atomic.h>
40#include <asm/irq.h>
41#include <asm/hw_irq.h>
42#include <asm/kvm_ppc.h>
43#include <asm/dbell.h>
44#include <asm/page.h>
45#include <asm/prom.h>
46#include <asm/smp.h>
47#include <asm/time.h>
48#include <asm/machdep.h>
49#include <asm/cputhreads.h>
50#include <asm/cputable.h>
51#include <asm/mpic.h>
52#include <asm/vdso_datapage.h>
53#ifdef CONFIG_PPC64
54#include <asm/paca.h>
55#endif
56#include <asm/vdso.h>
57#include <asm/debug.h>
58#include <asm/kexec.h>
59#include <asm/asm-prototypes.h>
60#include <asm/cpu_has_feature.h>
61#include <asm/ftrace.h>
62#include <asm/kup.h>
63
64#ifdef DEBUG
65#include <asm/udbg.h>
66#define DBG(fmt...) udbg_printf(fmt)
67#else
68#define DBG(fmt...)
69#endif
70
71#ifdef CONFIG_HOTPLUG_CPU
72/* State of each CPU during hotplug phases */
73static DEFINE_PER_CPU(int, cpu_state) = { 0 };
74#endif
75
76struct task_struct *secondary_current;
77bool has_big_cores;
78
79DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
80DEFINE_PER_CPU(cpumask_var_t, cpu_smallcore_map);
81DEFINE_PER_CPU(cpumask_var_t, cpu_l2_cache_map);
82DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
83
84EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
85EXPORT_PER_CPU_SYMBOL(cpu_l2_cache_map);
86EXPORT_PER_CPU_SYMBOL(cpu_core_map);
87EXPORT_SYMBOL_GPL(has_big_cores);
88
89#define MAX_THREAD_LIST_SIZE 8
90#define THREAD_GROUP_SHARE_L1 1
91struct 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];
96};
97
98/*
99 * On big-cores system, cpu_l1_cache_map for each CPU corresponds to
100 * the set its siblings that share the L1-cache.
101 */
102DEFINE_PER_CPU(cpumask_var_t, cpu_l1_cache_map);
103
104/* SMP operations for this machine */
105struct smp_ops_t *smp_ops;
106
107/* Can't be static due to PowerMac hackery */
108volatile unsigned int cpu_callin_map[NR_CPUS];
109
110int smt_enabled_at_boot = 1;
111
112/*
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
116 */
117int smp_generic_cpu_bootable(unsigned int nr)
118{
119 /* Special case - we inhibit secondary thread startup
120 * during boot if the user requests it.
121 */
122 if (system_state < SYSTEM_RUNNING && cpu_has_feature(CPU_FTR_SMT)) {
123 if (!smt_enabled_at_boot && cpu_thread_in_core(nr) != 0)
124 return 0;
125 if (smt_enabled_at_boot
126 && cpu_thread_in_core(nr) >= smt_enabled_at_boot)
127 return 0;
128 }
129
130 return 1;
131}
132
133
134#ifdef CONFIG_PPC64
135int smp_generic_kick_cpu(int nr)
136{
137 if (nr < 0 || nr >= nr_cpu_ids)
138 return -EINVAL;
139
140 /*
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
144 */
145 if (!paca_ptrs[nr]->cpu_start) {
146 paca_ptrs[nr]->cpu_start = 1;
147 smp_mb();
148 return 0;
149 }
150
151#ifdef CONFIG_HOTPLUG_CPU
152 /*
153 * Ok it's not there, so it might be soft-unplugged, let's
154 * try to bring it back
155 */
156 generic_set_cpu_up(nr);
157 smp_wmb();
158 smp_send_reschedule(nr);
159#endif /* CONFIG_HOTPLUG_CPU */
160
161 return 0;
162}
163#endif /* CONFIG_PPC64 */
164
165static irqreturn_t call_function_action(int irq, void *data)
166{
167 generic_smp_call_function_interrupt();
168 return IRQ_HANDLED;
169}
170
171static irqreturn_t reschedule_action(int irq, void *data)
172{
173 scheduler_ipi();
174 return IRQ_HANDLED;
175}
176
177#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
178static irqreturn_t tick_broadcast_ipi_action(int irq, void *data)
179{
180 timer_broadcast_interrupt();
181 return IRQ_HANDLED;
182}
183#endif
184
185#ifdef CONFIG_NMI_IPI
186static irqreturn_t nmi_ipi_action(int irq, void *data)
187{
188 smp_handle_nmi_ipi(get_irq_regs());
189 return IRQ_HANDLED;
190}
191#endif
192
193static 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,
198#endif
199#ifdef CONFIG_NMI_IPI
200 [PPC_MSG_NMI_IPI] = nmi_ipi_action,
201#endif
202};
203
204/*
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.
208 */
209const 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",
214#endif
215#ifdef CONFIG_NMI_IPI
216 [PPC_MSG_NMI_IPI] = "nmi ipi",
217#endif
218};
219
220/* optional function to request ipi, for controllers with >= 4 ipis */
221int smp_request_message_ipi(int virq, int msg)
222{
223 int err;
224
225 if (msg < 0 || msg > PPC_MSG_NMI_IPI)
226 return -EINVAL;
227#ifndef CONFIG_NMI_IPI
228 if (msg == PPC_MSG_NMI_IPI)
229 return 1;
230#endif
231
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);
237
238 return err;
239}
240
241#ifdef CONFIG_PPC_SMP_MUXED_IPI
242struct cpu_messages {
243 long messages; /* current messages */
244};
245static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
246
247void smp_muxed_ipi_set_message(int cpu, int msg)
248{
249 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
250 char *message = (char *)&info->messages;
251
252 /*
253 * Order previous accesses before accesses in the IPI handler.
254 */
255 smp_mb();
256 message[msg] = 1;
257}
258
259void smp_muxed_ipi_message_pass(int cpu, int msg)
260{
261 smp_muxed_ipi_set_message(cpu, msg);
262
263 /*
264 * cause_ipi functions are required to include a full barrier
265 * before doing whatever causes the IPI.
266 */
267 smp_ops->cause_ipi(cpu);
268}
269
270#ifdef __BIG_ENDIAN__
271#define IPI_MESSAGE(A) (1uL << ((BITS_PER_LONG - 8) - 8 * (A)))
272#else
273#define IPI_MESSAGE(A) (1uL << (8 * (A)))
274#endif
275
276irqreturn_t smp_ipi_demux(void)
277{
278 mb(); /* order any irq clear */
279
280 return smp_ipi_demux_relaxed();
281}
282
283/* sync-free variant. Callers should ensure synchronization */
284irqreturn_t smp_ipi_demux_relaxed(void)
285{
286 struct cpu_messages *info;
287 unsigned long all;
288
289 info = this_cpu_ptr(&ipi_message);
290 do {
291 all = xchg(&info->messages, 0);
292#if defined(CONFIG_KVM_XICS) && defined(CONFIG_KVM_BOOK3S_HV_POSSIBLE)
293 /*
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.
299 */
300 if (all & IPI_MESSAGE(PPC_MSG_RM_HOST_ACTION))
301 kvmppc_xics_ipi_action();
302#endif
303 if (all & IPI_MESSAGE(PPC_MSG_CALL_FUNCTION))
304 generic_smp_call_function_interrupt();
305 if (all & IPI_MESSAGE(PPC_MSG_RESCHEDULE))
306 scheduler_ipi();
307#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
308 if (all & IPI_MESSAGE(PPC_MSG_TICK_BROADCAST))
309 timer_broadcast_interrupt();
310#endif
311#ifdef CONFIG_NMI_IPI
312 if (all & IPI_MESSAGE(PPC_MSG_NMI_IPI))
313 nmi_ipi_action(0, NULL);
314#endif
315 } while (info->messages);
316
317 return IRQ_HANDLED;
318}
319#endif /* CONFIG_PPC_SMP_MUXED_IPI */
320
321static inline void do_message_pass(int cpu, int msg)
322{
323 if (smp_ops->message_pass)
324 smp_ops->message_pass(cpu, msg);
325#ifdef CONFIG_PPC_SMP_MUXED_IPI
326 else
327 smp_muxed_ipi_message_pass(cpu, msg);
328#endif
329}
330
331void smp_send_reschedule(int cpu)
332{
333 if (likely(smp_ops))
334 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
335}
336EXPORT_SYMBOL_GPL(smp_send_reschedule);
337
338void arch_send_call_function_single_ipi(int cpu)
339{
340 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
341}
342
343void arch_send_call_function_ipi_mask(const struct cpumask *mask)
344{
345 unsigned int cpu;
346
347 for_each_cpu(cpu, mask)
348 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
349}
350
351#ifdef CONFIG_NMI_IPI
352
353/*
354 * "NMI IPI" system.
355 *
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.
358 *
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.
363 *
364 * A new NMI can be issued before all targets exit the handler.
365 *
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.
371 */
372
373static atomic_t __nmi_ipi_lock = ATOMIC_INIT(0);
374static struct cpumask nmi_ipi_pending_mask;
375static bool nmi_ipi_busy = false;
376static void (*nmi_ipi_function)(struct pt_regs *) = NULL;
377
378static void nmi_ipi_lock_start(unsigned long *flags)
379{
380 raw_local_irq_save(*flags);
381 hard_irq_disable();
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);
386 hard_irq_disable();
387 }
388}
389
390static void nmi_ipi_lock(void)
391{
392 while (atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1)
393 spin_until_cond(atomic_read(&__nmi_ipi_lock) == 0);
394}
395
396static void nmi_ipi_unlock(void)
397{
398 smp_mb();
399 WARN_ON(atomic_read(&__nmi_ipi_lock) != 1);
400 atomic_set(&__nmi_ipi_lock, 0);
401}
402
403static void nmi_ipi_unlock_end(unsigned long *flags)
404{
405 nmi_ipi_unlock();
406 raw_local_irq_restore(*flags);
407}
408
409/*
410 * Platform NMI handler calls this to ack
411 */
412int smp_handle_nmi_ipi(struct pt_regs *regs)
413{
414 void (*fn)(struct pt_regs *) = NULL;
415 unsigned long flags;
416 int me = raw_smp_processor_id();
417 int ret = 0;
418
419 /*
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.
423 */
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);
428 WARN_ON_ONCE(!fn);
429 ret = 1;
430 }
431 nmi_ipi_unlock_end(&flags);
432
433 if (fn)
434 fn(regs);
435
436 return ret;
437}
438
439static void do_smp_send_nmi_ipi(int cpu, bool safe)
440{
441 if (!safe && smp_ops->cause_nmi_ipi && smp_ops->cause_nmi_ipi(cpu))
442 return;
443
444 if (cpu >= 0) {
445 do_message_pass(cpu, PPC_MSG_NMI_IPI);
446 } else {
447 int c;
448
449 for_each_online_cpu(c) {
450 if (c == raw_smp_processor_id())
451 continue;
452 do_message_pass(c, PPC_MSG_NMI_IPI);
453 }
454 }
455}
456
457/*
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.
462 */
463static int __smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *),
464 u64 delay_us, bool safe)
465{
466 unsigned long flags;
467 int me = raw_smp_processor_id();
468 int ret = 1;
469
470 BUG_ON(cpu == me);
471 BUG_ON(cpu < 0 && cpu != NMI_IPI_ALL_OTHERS);
472
473 if (unlikely(!smp_ops))
474 return 0;
475
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);
481 }
482 nmi_ipi_busy = true;
483 nmi_ipi_function = fn;
484
485 WARN_ON_ONCE(!cpumask_empty(&nmi_ipi_pending_mask));
486
487 if (cpu < 0) {
488 /* ALL_OTHERS */
489 cpumask_copy(&nmi_ipi_pending_mask, cpu_online_mask);
490 cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
491 } else {
492 cpumask_set_cpu(cpu, &nmi_ipi_pending_mask);
493 }
494
495 nmi_ipi_unlock();
496
497 /* Interrupts remain hard disabled */
498
499 do_smp_send_nmi_ipi(cpu, safe);
500
501 nmi_ipi_lock();
502 /* nmi_ipi_busy is set here, so unlock/lock is okay */
503 while (!cpumask_empty(&nmi_ipi_pending_mask)) {
504 nmi_ipi_unlock();
505 udelay(1);
506 nmi_ipi_lock();
507 if (delay_us) {
508 delay_us--;
509 if (!delay_us)
510 break;
511 }
512 }
513
514 if (!cpumask_empty(&nmi_ipi_pending_mask)) {
515 /* Timeout waiting for CPUs to call smp_handle_nmi_ipi */
516 ret = 0;
517 cpumask_clear(&nmi_ipi_pending_mask);
518 }
519
520 nmi_ipi_function = NULL;
521 nmi_ipi_busy = false;
522
523 nmi_ipi_unlock_end(&flags);
524
525 return ret;
526}
527
528int smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
529{
530 return __smp_send_nmi_ipi(cpu, fn, delay_us, false);
531}
532
533int smp_send_safe_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
534{
535 return __smp_send_nmi_ipi(cpu, fn, delay_us, true);
536}
537#endif /* CONFIG_NMI_IPI */
538
539#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
540void tick_broadcast(const struct cpumask *mask)
541{
542 unsigned int cpu;
543
544 for_each_cpu(cpu, mask)
545 do_message_pass(cpu, PPC_MSG_TICK_BROADCAST);
546}
547#endif
548
549#ifdef CONFIG_DEBUGGER
550void debugger_ipi_callback(struct pt_regs *regs)
551{
552 debugger_ipi(regs);
553}
554
555void smp_send_debugger_break(void)
556{
557 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, debugger_ipi_callback, 1000000);
558}
559#endif
560
561#ifdef CONFIG_KEXEC_CORE
562void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
563{
564 int cpu;
565
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) {
569 if (cpu_online(cpu))
570 continue;
571 /*
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.
579 */
580 do_smp_send_nmi_ipi(cpu, false);
581 }
582 }
583}
584#endif
585
586#ifdef CONFIG_NMI_IPI
587static void nmi_stop_this_cpu(struct pt_regs *regs)
588{
589 /*
590 * IRQs are already hard disabled by the smp_handle_nmi_ipi.
591 */
592 spin_begin();
593 while (1)
594 spin_cpu_relax();
595}
596
597void smp_send_stop(void)
598{
599 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, nmi_stop_this_cpu, 1000000);
600}
601
602#else /* CONFIG_NMI_IPI */
603
604static void stop_this_cpu(void *dummy)
605{
606 hard_irq_disable();
607 spin_begin();
608 while (1)
609 spin_cpu_relax();
610}
611
612void smp_send_stop(void)
613{
614 static bool stopped = false;
615
616 /*
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
620 * kernel/panic.c)
621 */
622 if (stopped)
623 return;
624
625 stopped = true;
626
627 smp_call_function(stop_this_cpu, NULL, 0);
628}
629#endif /* CONFIG_NMI_IPI */
630
631struct task_struct *current_set[NR_CPUS];
632
633static void smp_store_cpu_info(int id)
634{
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;
639#endif
640}
641
642/*
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.
646 */
647static void set_cpus_related(int i, int j, struct cpumask *(*get_cpumask)(int))
648{
649 cpumask_set_cpu(i, get_cpumask(j));
650 cpumask_set_cpu(j, get_cpumask(i));
651}
652
653#ifdef CONFIG_HOTPLUG_CPU
654static void set_cpus_unrelated(int i, int j,
655 struct cpumask *(*get_cpumask)(int))
656{
657 cpumask_clear_cpu(i, get_cpumask(j));
658 cpumask_clear_cpu(j, get_cpumask(i));
659}
660#endif
661
662/*
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]
667 * matches @property.
668 *
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
673 * interested in.
674 *
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.
677 *
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.
681 *
682 * ibm,thread-groups[1] tells us how many such thread groups exist.
683 *
684 * ibm,thread-groups[2] tells us the number of threads in each such
685 * group.
686 *
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
689 * the grouping.
690 *
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.
694 *
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,
697 * 11, 12} structure
698 *
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.
702 */
703static int parse_thread_groups(struct device_node *dn,
704 struct thread_groups *tg,
705 unsigned int property)
706{
707 int i;
708 u32 thread_group_array[3 + MAX_THREAD_LIST_SIZE];
709 u32 *thread_list;
710 size_t total_threads;
711 int ret;
712
713 ret = of_property_read_u32_array(dn, "ibm,thread-groups",
714 thread_group_array, 3);
715 if (ret)
716 return ret;
717
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 ||
722 tg->nr_groups < 1 ||
723 tg->threads_per_group < 1)
724 return -ENODATA;
725
726 total_threads = tg->nr_groups * tg->threads_per_group;
727
728 ret = of_property_read_u32_array(dn, "ibm,thread-groups",
729 thread_group_array,
730 3 + total_threads);
731 if (ret)
732 return ret;
733
734 thread_list = &thread_group_array[3];
735
736 for (i = 0 ; i < total_threads; i++)
737 tg->thread_list[i] = thread_list[i];
738
739 return 0;
740}
741
742/*
743 * get_cpu_thread_group_start : Searches the thread group in tg->thread_list
744 * that @cpu belongs to.
745 *
746 * @cpu : The logical CPU whose thread group is being searched.
747 * @tg : The thread-group structure of the CPU node which @cpu belongs
748 * to.
749 *
750 * Returns the index to tg->thread_list that points to the the start
751 * of the thread_group that @cpu belongs to.
752 *
753 * Returns -1 if cpu doesn't belong to any of the groups pointed to by
754 * tg->thread_list.
755 */
756static int get_cpu_thread_group_start(int cpu, struct thread_groups *tg)
757{
758 int hw_cpu_id = get_hard_smp_processor_id(cpu);
759 int i, j;
760
761 for (i = 0; i < tg->nr_groups; i++) {
762 int group_start = i * tg->threads_per_group;
763
764 for (j = 0; j < tg->threads_per_group; j++) {
765 int idx = group_start + j;
766
767 if (tg->thread_list[idx] == hw_cpu_id)
768 return group_start;
769 }
770 }
771
772 return -1;
773}
774
775static int init_cpu_l1_cache_map(int cpu)
776
777{
778 struct device_node *dn = of_get_cpu_node(cpu, NULL);
779 struct thread_groups tg = {.property = 0,
780 .nr_groups = 0,
781 .threads_per_group = 0};
782 int first_thread = cpu_first_thread_sibling(cpu);
783 int i, cpu_group_start = -1, err = 0;
784
785 if (!dn)
786 return -ENODATA;
787
788 err = parse_thread_groups(dn, &tg, THREAD_GROUP_SHARE_L1);
789 if (err)
790 goto out;
791
792 zalloc_cpumask_var_node(&per_cpu(cpu_l1_cache_map, cpu),
793 GFP_KERNEL,
794 cpu_to_node(cpu));
795
796 cpu_group_start = get_cpu_thread_group_start(cpu, &tg);
797
798 if (unlikely(cpu_group_start == -1)) {
799 WARN_ON_ONCE(1);
800 err = -ENODATA;
801 goto out;
802 }
803
804 for (i = first_thread; i < first_thread + threads_per_core; i++) {
805 int i_group_start = get_cpu_thread_group_start(i, &tg);
806
807 if (unlikely(i_group_start == -1)) {
808 WARN_ON_ONCE(1);
809 err = -ENODATA;
810 goto out;
811 }
812
813 if (i_group_start == cpu_group_start)
814 cpumask_set_cpu(i, per_cpu(cpu_l1_cache_map, cpu));
815 }
816
817out:
818 of_node_put(dn);
819 return err;
820}
821
822static int init_big_cores(void)
823{
824 int cpu;
825
826 for_each_possible_cpu(cpu) {
827 int err = init_cpu_l1_cache_map(cpu);
828
829 if (err)
830 return err;
831
832 zalloc_cpumask_var_node(&per_cpu(cpu_smallcore_map, cpu),
833 GFP_KERNEL,
834 cpu_to_node(cpu));
835 }
836
837 has_big_cores = true;
838 return 0;
839}
840
841void __init smp_prepare_cpus(unsigned int max_cpus)
842{
843 unsigned int cpu;
844
845 DBG("smp_prepare_cpus\n");
846
847 /*
848 * setup_cpu may need to be called on the boot cpu. We havent
849 * spun any cpus up but lets be paranoid.
850 */
851 BUG_ON(boot_cpuid != smp_processor_id());
852
853 /* Fixup boot cpu */
854 smp_store_cpu_info(boot_cpuid);
855 cpu_callin_map[boot_cpuid] = 1;
856
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));
864 /*
865 * numa_node_id() works after this.
866 */
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]));
871 }
872 }
873
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));
878
879 init_big_cores();
880 if (has_big_cores) {
881 cpumask_set_cpu(boot_cpuid,
882 cpu_smallcore_mask(boot_cpuid));
883 }
884
885 if (smp_ops && smp_ops->probe)
886 smp_ops->probe();
887}
888
889void smp_prepare_boot_cpu(void)
890{
891 BUG_ON(smp_processor_id() != boot_cpuid);
892#ifdef CONFIG_PPC64
893 paca_ptrs[boot_cpuid]->__current = current;
894#endif
895 set_numa_node(numa_cpu_lookup_table[boot_cpuid]);
896 current_set[boot_cpuid] = current;
897}
898
899#ifdef CONFIG_HOTPLUG_CPU
900
901int generic_cpu_disable(void)
902{
903 unsigned int cpu = smp_processor_id();
904
905 if (cpu == boot_cpuid)
906 return -EBUSY;
907
908 set_cpu_online(cpu, false);
909#ifdef CONFIG_PPC64
910 vdso_data->processorCount--;
911#endif
912 /* Update affinity of all IRQs previously aimed at this CPU */
913 irq_migrate_all_off_this_cpu();
914
915 /*
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.
922 */
923 local_irq_enable();
924 mdelay(1);
925 local_irq_disable();
926
927 return 0;
928}
929
930void generic_cpu_die(unsigned int cpu)
931{
932 int i;
933
934 for (i = 0; i < 100; i++) {
935 smp_rmb();
936 if (is_cpu_dead(cpu))
937 return;
938 msleep(100);
939 }
940 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
941}
942
943void generic_set_cpu_dead(unsigned int cpu)
944{
945 per_cpu(cpu_state, cpu) = CPU_DEAD;
946}
947
948/*
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.
952 */
953void generic_set_cpu_up(unsigned int cpu)
954{
955 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
956}
957
958int generic_check_cpu_restart(unsigned int cpu)
959{
960 return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
961}
962
963int is_cpu_dead(unsigned int cpu)
964{
965 return per_cpu(cpu_state, cpu) == CPU_DEAD;
966}
967
968static bool secondaries_inhibited(void)
969{
970 return kvm_hv_mode_active();
971}
972
973#else /* HOTPLUG_CPU */
974
975#define secondaries_inhibited() 0
976
977#endif
978
979static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
980{
981#ifdef CONFIG_PPC64
982 paca_ptrs[cpu]->__current = idle;
983 paca_ptrs[cpu]->kstack = (unsigned long)task_stack_page(idle) +
984 THREAD_SIZE - STACK_FRAME_OVERHEAD;
985#endif
986 idle->cpu = cpu;
987 secondary_current = current_set[cpu] = idle;
988}
989
990int __cpu_up(unsigned int cpu, struct task_struct *tidle)
991{
992 int rc, c;
993
994 /*
995 * Don't allow secondary threads to come online if inhibited
996 */
997 if (threads_per_core > 1 && secondaries_inhibited() &&
998 cpu_thread_in_subcore(cpu))
999 return -EBUSY;
1000
1001 if (smp_ops == NULL ||
1002 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
1003 return -EINVAL;
1004
1005 cpu_idle_thread_init(cpu, tidle);
1006
1007 /*
1008 * The platform might need to allocate resources prior to bringing
1009 * up the CPU
1010 */
1011 if (smp_ops->prepare_cpu) {
1012 rc = smp_ops->prepare_cpu(cpu);
1013 if (rc)
1014 return rc;
1015 }
1016
1017 /* Make sure callin-map entry is 0 (can be leftover a CPU
1018 * hotplug
1019 */
1020 cpu_callin_map[cpu] = 0;
1021
1022 /* The information for processor bringup must
1023 * be written out to main store before we release
1024 * the processor.
1025 */
1026 smp_mb();
1027
1028 /* wake up cpus */
1029 DBG("smp: kicking cpu %d\n", cpu);
1030 rc = smp_ops->kick_cpu(cpu);
1031 if (rc) {
1032 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
1033 return rc;
1034 }
1035
1036 /*
1037 * wait to see if the cpu made a callin (is actually up).
1038 * use this value that I found through experimentation.
1039 * -- Cort
1040 */
1041 if (system_state < SYSTEM_RUNNING)
1042 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
1043 udelay(100);
1044#ifdef CONFIG_HOTPLUG_CPU
1045 else
1046 /*
1047 * CPUs can take much longer to come up in the
1048 * hotplug case. Wait five seconds.
1049 */
1050 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
1051 msleep(1);
1052#endif
1053
1054 if (!cpu_callin_map[cpu]) {
1055 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
1056 return -ENOENT;
1057 }
1058
1059 DBG("Processor %u found.\n", cpu);
1060
1061 if (smp_ops->give_timebase)
1062 smp_ops->give_timebase();
1063
1064 /* Wait until cpu puts itself in the online & active maps */
1065 spin_until_cond(cpu_online(cpu));
1066
1067 return 0;
1068}
1069
1070/* Return the value of the reg property corresponding to the given
1071 * logical cpu.
1072 */
1073int cpu_to_core_id(int cpu)
1074{
1075 struct device_node *np;
1076 const __be32 *reg;
1077 int id = -1;
1078
1079 np = of_get_cpu_node(cpu, NULL);
1080 if (!np)
1081 goto out;
1082
1083 reg = of_get_property(np, "reg", NULL);
1084 if (!reg)
1085 goto out;
1086
1087 id = be32_to_cpup(reg);
1088out:
1089 of_node_put(np);
1090 return id;
1091}
1092EXPORT_SYMBOL_GPL(cpu_to_core_id);
1093
1094/* Helper routines for cpu to core mapping */
1095int cpu_core_index_of_thread(int cpu)
1096{
1097 return cpu >> threads_shift;
1098}
1099EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
1100
1101int cpu_first_thread_of_core(int core)
1102{
1103 return core << threads_shift;
1104}
1105EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
1106
1107/* Must be called when no change can occur to cpu_present_mask,
1108 * i.e. during cpu online or offline.
1109 */
1110static struct device_node *cpu_to_l2cache(int cpu)
1111{
1112 struct device_node *np;
1113 struct device_node *cache;
1114
1115 if (!cpu_present(cpu))
1116 return NULL;
1117
1118 np = of_get_cpu_node(cpu, NULL);
1119 if (np == NULL)
1120 return NULL;
1121
1122 cache = of_find_next_cache_node(np);
1123
1124 of_node_put(np);
1125
1126 return cache;
1127}
1128
1129static bool update_mask_by_l2(int cpu, struct cpumask *(*mask_fn)(int))
1130{
1131 struct device_node *l2_cache, *np;
1132 int i;
1133
1134 l2_cache = cpu_to_l2cache(cpu);
1135 if (!l2_cache)
1136 return false;
1137
1138 for_each_cpu(i, cpu_online_mask) {
1139 /*
1140 * when updating the marks the current CPU has not been marked
1141 * online, but we need to update the cache masks
1142 */
1143 np = cpu_to_l2cache(i);
1144 if (!np)
1145 continue;
1146
1147 if (np == l2_cache)
1148 set_cpus_related(cpu, i, mask_fn);
1149
1150 of_node_put(np);
1151 }
1152 of_node_put(l2_cache);
1153
1154 return true;
1155}
1156
1157#ifdef CONFIG_HOTPLUG_CPU
1158static void remove_cpu_from_masks(int cpu)
1159{
1160 int i;
1161
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);
1167 if (has_big_cores)
1168 set_cpus_unrelated(cpu, i, cpu_smallcore_mask);
1169 }
1170}
1171#endif
1172
1173static inline void add_cpu_to_smallcore_masks(int cpu)
1174{
1175 struct cpumask *this_l1_cache_map = per_cpu(cpu_l1_cache_map, cpu);
1176 int i, first_thread = cpu_first_thread_sibling(cpu);
1177
1178 if (!has_big_cores)
1179 return;
1180
1181 cpumask_set_cpu(cpu, cpu_smallcore_mask(cpu));
1182
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);
1186 }
1187}
1188
1189int get_physical_package_id(int cpu)
1190{
1191 int pkg_id = cpu_to_chip_id(cpu);
1192
1193 /*
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.
1197 */
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);
1202 of_node_put(np);
1203 }
1204
1205 return pkg_id;
1206}
1207EXPORT_SYMBOL_GPL(get_physical_package_id);
1208
1209static void add_cpu_to_masks(int cpu)
1210{
1211 int first_thread = cpu_first_thread_sibling(cpu);
1212 int pkg_id = get_physical_package_id(cpu);
1213 int i;
1214
1215 /*
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.
1218 */
1219 cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
1220
1221 for (i = first_thread; i < first_thread + threads_per_core; i++)
1222 if (cpu_online(i))
1223 set_cpus_related(i, cpu, cpu_sibling_mask);
1224
1225 add_cpu_to_smallcore_masks(cpu);
1226 /*
1227 * Copy the thread sibling mask into the cache sibling mask
1228 * and mark any CPUs that share an L2 with this CPU.
1229 */
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);
1233
1234 /*
1235 * Copy the cache sibling mask into core sibling mask and mark
1236 * any CPUs on the same chip as this CPU.
1237 */
1238 for_each_cpu(i, cpu_l2_cache_mask(cpu))
1239 set_cpus_related(cpu, i, cpu_core_mask);
1240
1241 if (pkg_id == -1)
1242 return;
1243
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);
1247}
1248
1249static bool shared_caches;
1250
1251/* Activate a secondary processor. */
1252void start_secondary(void *unused)
1253{
1254 unsigned int cpu = smp_processor_id();
1255 struct cpumask *(*sibling_mask)(int) = cpu_sibling_mask;
1256
1257 mmgrab(&init_mm);
1258 current->active_mm = &init_mm;
1259
1260 smp_store_cpu_info(cpu);
1261 set_dec(tb_ticks_per_jiffy);
1262 preempt_disable();
1263 cpu_callin_map[cpu] = 1;
1264
1265 if (smp_ops->setup_cpu)
1266 smp_ops->setup_cpu(cpu);
1267 if (smp_ops->take_timebase)
1268 smp_ops->take_timebase();
1269
1270 secondary_cpu_time_init();
1271
1272#ifdef CONFIG_PPC64
1273 if (system_state == SYSTEM_RUNNING)
1274 vdso_data->processorCount++;
1275
1276 vdso_getcpu_init();
1277#endif
1278 /* Update topology CPU masks */
1279 add_cpu_to_masks(cpu);
1280
1281 if (has_big_cores)
1282 sibling_mask = cpu_smallcore_mask;
1283 /*
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.
1286 */
1287 if (!cpumask_equal(cpu_l2_cache_mask(cpu), sibling_mask(cpu)))
1288 shared_caches = true;
1289
1290 set_numa_node(numa_cpu_lookup_table[cpu]);
1291 set_numa_mem(local_memory_node(numa_cpu_lookup_table[cpu]));
1292
1293 smp_wmb();
1294 notify_cpu_starting(cpu);
1295 set_cpu_online(cpu, true);
1296
1297 boot_init_stack_canary();
1298
1299 local_irq_enable();
1300
1301 /* We can enable ftrace for secondary cpus now */
1302 this_cpu_enable_ftrace();
1303
1304 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
1305
1306 BUG();
1307}
1308
1309int setup_profiling_timer(unsigned int multiplier)
1310{
1311 return 0;
1312}
1313
1314#ifdef CONFIG_SCHED_SMT
1315/* cpumask of CPUs with asymetric SMT dependancy */
1316static int powerpc_smt_flags(void)
1317{
1318 int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
1319
1320 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
1321 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
1322 flags |= SD_ASYM_PACKING;
1323 }
1324 return flags;
1325}
1326#endif
1327
1328static struct sched_domain_topology_level powerpc_topology[] = {
1329#ifdef CONFIG_SCHED_SMT
1330 { cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
1331#endif
1332 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1333 { NULL, },
1334};
1335
1336/*
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.
1341 */
1342static int powerpc_shared_cache_flags(void)
1343{
1344 return SD_SHARE_PKG_RESOURCES;
1345}
1346
1347/*
1348 * We can't just pass cpu_l2_cache_mask() directly because
1349 * returns a non-const pointer and the compiler barfs on that.
1350 */
1351static const struct cpumask *shared_cache_mask(int cpu)
1352{
1353 return cpu_l2_cache_mask(cpu);
1354}
1355
1356#ifdef CONFIG_SCHED_SMT
1357static const struct cpumask *smallcore_smt_mask(int cpu)
1358{
1359 return cpu_smallcore_mask(cpu);
1360}
1361#endif
1362
1363static struct sched_domain_topology_level power9_topology[] = {
1364#ifdef CONFIG_SCHED_SMT
1365 { cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
1366#endif
1367 { shared_cache_mask, powerpc_shared_cache_flags, SD_INIT_NAME(CACHE) },
1368 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1369 { NULL, },
1370};
1371
1372void __init smp_cpus_done(unsigned int max_cpus)
1373{
1374 /*
1375 * We are running pinned to the boot CPU, see rest_init().
1376 */
1377 if (smp_ops && smp_ops->setup_cpu)
1378 smp_ops->setup_cpu(boot_cpuid);
1379
1380 if (smp_ops && smp_ops->bringup_done)
1381 smp_ops->bringup_done();
1382
1383 dump_numa_cpu_topology();
1384
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;
1390 }
1391#endif
1392 /*
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.
1395 */
1396 if (shared_caches) {
1397 pr_info("Using shared cache scheduler topology\n");
1398 set_sched_topology(power9_topology);
1399 } else {
1400 pr_info("Using standard scheduler topology\n");
1401 set_sched_topology(powerpc_topology);
1402 }
1403}
1404
1405#ifdef CONFIG_HOTPLUG_CPU
1406int __cpu_disable(void)
1407{
1408 int cpu = smp_processor_id();
1409 int err;
1410
1411 if (!smp_ops->cpu_disable)
1412 return -ENOSYS;
1413
1414 this_cpu_disable_ftrace();
1415
1416 err = smp_ops->cpu_disable();
1417 if (err)
1418 return err;
1419
1420 /* Update sibling maps */
1421 remove_cpu_from_masks(cpu);
1422
1423 return 0;
1424}
1425
1426void __cpu_die(unsigned int cpu)
1427{
1428 if (smp_ops->cpu_die)
1429 smp_ops->cpu_die(cpu);
1430}
1431
1432void cpu_die(void)
1433{
1434 /*
1435 * Disable on the down path. This will be re-enabled by
1436 * start_secondary() via start_secondary_resume() below
1437 */
1438 this_cpu_disable_ftrace();
1439
1440 if (ppc_md.cpu_die)
1441 ppc_md.cpu_die();
1442
1443 /* If we return, we re-enter start_secondary */
1444 start_secondary_resume();
1445}
1446
1447#endif
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * SMP support for ppc.
4 *
5 * Written by Cort Dougan (cort@cs.nmt.edu) borrowing a great
6 * deal of code from the sparc and intel versions.
7 *
8 * Copyright (C) 1999 Cort Dougan <cort@cs.nmt.edu>
9 *
10 * PowerPC-64 Support added by Dave Engebretsen, Peter Bergner, and
11 * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
12 */
13
14#undef DEBUG
15
16#include <linux/kernel.h>
17#include <linux/export.h>
18#include <linux/sched/mm.h>
19#include <linux/sched/task_stack.h>
20#include <linux/sched/topology.h>
21#include <linux/smp.h>
22#include <linux/interrupt.h>
23#include <linux/delay.h>
24#include <linux/init.h>
25#include <linux/spinlock.h>
26#include <linux/cache.h>
27#include <linux/err.h>
28#include <linux/device.h>
29#include <linux/cpu.h>
30#include <linux/notifier.h>
31#include <linux/topology.h>
32#include <linux/profile.h>
33#include <linux/processor.h>
34#include <linux/random.h>
35#include <linux/stackprotector.h>
36#include <linux/pgtable.h>
37#include <linux/clockchips.h>
38#include <linux/kexec.h>
39
40#include <asm/ptrace.h>
41#include <linux/atomic.h>
42#include <asm/irq.h>
43#include <asm/hw_irq.h>
44#include <asm/kvm_ppc.h>
45#include <asm/dbell.h>
46#include <asm/page.h>
47#include <asm/smp.h>
48#include <asm/time.h>
49#include <asm/machdep.h>
50#include <asm/cputhreads.h>
51#include <asm/cputable.h>
52#include <asm/mpic.h>
53#include <asm/vdso_datapage.h>
54#ifdef CONFIG_PPC64
55#include <asm/paca.h>
56#endif
57#include <asm/vdso.h>
58#include <asm/debug.h>
59#include <asm/cpu_has_feature.h>
60#include <asm/ftrace.h>
61#include <asm/kup.h>
62#include <asm/fadump.h>
63
64#ifdef DEBUG
65#include <asm/udbg.h>
66#define DBG(fmt...) udbg_printf(fmt)
67#else
68#define DBG(fmt...)
69#endif
70
71#ifdef CONFIG_HOTPLUG_CPU
72/* State of each CPU during hotplug phases */
73static DEFINE_PER_CPU(int, cpu_state) = { 0 };
74#endif
75
76struct task_struct *secondary_current;
77bool has_big_cores;
78bool coregroup_enabled;
79bool thread_group_shares_l2;
80bool thread_group_shares_l3;
81
82DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
83DEFINE_PER_CPU(cpumask_var_t, cpu_smallcore_map);
84DEFINE_PER_CPU(cpumask_var_t, cpu_l2_cache_map);
85DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
86static DEFINE_PER_CPU(cpumask_var_t, cpu_coregroup_map);
87
88EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
89EXPORT_PER_CPU_SYMBOL(cpu_l2_cache_map);
90EXPORT_PER_CPU_SYMBOL(cpu_core_map);
91EXPORT_SYMBOL_GPL(has_big_cores);
92
93enum {
94#ifdef CONFIG_SCHED_SMT
95 smt_idx,
96#endif
97 cache_idx,
98 mc_idx,
99 die_idx,
100};
101
102#define MAX_THREAD_LIST_SIZE 8
103#define THREAD_GROUP_SHARE_L1 1
104#define THREAD_GROUP_SHARE_L2_L3 2
105struct thread_groups {
106 unsigned int property;
107 unsigned int nr_groups;
108 unsigned int threads_per_group;
109 unsigned int thread_list[MAX_THREAD_LIST_SIZE];
110};
111
112/* Maximum number of properties that groups of threads within a core can share */
113#define MAX_THREAD_GROUP_PROPERTIES 2
114
115struct thread_groups_list {
116 unsigned int nr_properties;
117 struct thread_groups property_tgs[MAX_THREAD_GROUP_PROPERTIES];
118};
119
120static struct thread_groups_list tgl[NR_CPUS] __initdata;
121/*
122 * On big-cores system, thread_group_l1_cache_map for each CPU corresponds to
123 * the set its siblings that share the L1-cache.
124 */
125DEFINE_PER_CPU(cpumask_var_t, thread_group_l1_cache_map);
126
127/*
128 * On some big-cores system, thread_group_l2_cache_map for each CPU
129 * corresponds to the set its siblings within the core that share the
130 * L2-cache.
131 */
132DEFINE_PER_CPU(cpumask_var_t, thread_group_l2_cache_map);
133
134/*
135 * On P10, thread_group_l3_cache_map for each CPU is equal to the
136 * thread_group_l2_cache_map
137 */
138DEFINE_PER_CPU(cpumask_var_t, thread_group_l3_cache_map);
139
140/* SMP operations for this machine */
141struct smp_ops_t *smp_ops;
142
143/* Can't be static due to PowerMac hackery */
144volatile unsigned int cpu_callin_map[NR_CPUS];
145
146int smt_enabled_at_boot = 1;
147
148/*
149 * Returns 1 if the specified cpu should be brought up during boot.
150 * Used to inhibit booting threads if they've been disabled or
151 * limited on the command line
152 */
153int smp_generic_cpu_bootable(unsigned int nr)
154{
155 /* Special case - we inhibit secondary thread startup
156 * during boot if the user requests it.
157 */
158 if (system_state < SYSTEM_RUNNING && cpu_has_feature(CPU_FTR_SMT)) {
159 if (!smt_enabled_at_boot && cpu_thread_in_core(nr) != 0)
160 return 0;
161 if (smt_enabled_at_boot
162 && cpu_thread_in_core(nr) >= smt_enabled_at_boot)
163 return 0;
164 }
165
166 return 1;
167}
168
169
170#ifdef CONFIG_PPC64
171int smp_generic_kick_cpu(int nr)
172{
173 if (nr < 0 || nr >= nr_cpu_ids)
174 return -EINVAL;
175
176 /*
177 * The processor is currently spinning, waiting for the
178 * cpu_start field to become non-zero After we set cpu_start,
179 * the processor will continue on to secondary_start
180 */
181 if (!paca_ptrs[nr]->cpu_start) {
182 paca_ptrs[nr]->cpu_start = 1;
183 smp_mb();
184 return 0;
185 }
186
187#ifdef CONFIG_HOTPLUG_CPU
188 /*
189 * Ok it's not there, so it might be soft-unplugged, let's
190 * try to bring it back
191 */
192 generic_set_cpu_up(nr);
193 smp_wmb();
194 smp_send_reschedule(nr);
195#endif /* CONFIG_HOTPLUG_CPU */
196
197 return 0;
198}
199#endif /* CONFIG_PPC64 */
200
201static irqreturn_t call_function_action(int irq, void *data)
202{
203 generic_smp_call_function_interrupt();
204 return IRQ_HANDLED;
205}
206
207static irqreturn_t reschedule_action(int irq, void *data)
208{
209 scheduler_ipi();
210 return IRQ_HANDLED;
211}
212
213#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
214static irqreturn_t tick_broadcast_ipi_action(int irq, void *data)
215{
216 timer_broadcast_interrupt();
217 return IRQ_HANDLED;
218}
219#endif
220
221#ifdef CONFIG_NMI_IPI
222static irqreturn_t nmi_ipi_action(int irq, void *data)
223{
224 smp_handle_nmi_ipi(get_irq_regs());
225 return IRQ_HANDLED;
226}
227#endif
228
229static irq_handler_t smp_ipi_action[] = {
230 [PPC_MSG_CALL_FUNCTION] = call_function_action,
231 [PPC_MSG_RESCHEDULE] = reschedule_action,
232#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
233 [PPC_MSG_TICK_BROADCAST] = tick_broadcast_ipi_action,
234#endif
235#ifdef CONFIG_NMI_IPI
236 [PPC_MSG_NMI_IPI] = nmi_ipi_action,
237#endif
238};
239
240/*
241 * The NMI IPI is a fallback and not truly non-maskable. It is simpler
242 * than going through the call function infrastructure, and strongly
243 * serialized, so it is more appropriate for debugging.
244 */
245const char *smp_ipi_name[] = {
246 [PPC_MSG_CALL_FUNCTION] = "ipi call function",
247 [PPC_MSG_RESCHEDULE] = "ipi reschedule",
248#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
249 [PPC_MSG_TICK_BROADCAST] = "ipi tick-broadcast",
250#endif
251#ifdef CONFIG_NMI_IPI
252 [PPC_MSG_NMI_IPI] = "nmi ipi",
253#endif
254};
255
256/* optional function to request ipi, for controllers with >= 4 ipis */
257int smp_request_message_ipi(int virq, int msg)
258{
259 int err;
260
261 if (msg < 0 || msg > PPC_MSG_NMI_IPI)
262 return -EINVAL;
263#ifndef CONFIG_NMI_IPI
264 if (msg == PPC_MSG_NMI_IPI)
265 return 1;
266#endif
267
268 err = request_irq(virq, smp_ipi_action[msg],
269 IRQF_PERCPU | IRQF_NO_THREAD | IRQF_NO_SUSPEND,
270 smp_ipi_name[msg], NULL);
271 WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
272 virq, smp_ipi_name[msg], err);
273
274 return err;
275}
276
277#ifdef CONFIG_PPC_SMP_MUXED_IPI
278struct cpu_messages {
279 long messages; /* current messages */
280};
281static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
282
283void smp_muxed_ipi_set_message(int cpu, int msg)
284{
285 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
286 char *message = (char *)&info->messages;
287
288 /*
289 * Order previous accesses before accesses in the IPI handler.
290 */
291 smp_mb();
292 message[msg] = 1;
293}
294
295void smp_muxed_ipi_message_pass(int cpu, int msg)
296{
297 smp_muxed_ipi_set_message(cpu, msg);
298
299 /*
300 * cause_ipi functions are required to include a full barrier
301 * before doing whatever causes the IPI.
302 */
303 smp_ops->cause_ipi(cpu);
304}
305
306#ifdef __BIG_ENDIAN__
307#define IPI_MESSAGE(A) (1uL << ((BITS_PER_LONG - 8) - 8 * (A)))
308#else
309#define IPI_MESSAGE(A) (1uL << (8 * (A)))
310#endif
311
312irqreturn_t smp_ipi_demux(void)
313{
314 mb(); /* order any irq clear */
315
316 return smp_ipi_demux_relaxed();
317}
318
319/* sync-free variant. Callers should ensure synchronization */
320irqreturn_t smp_ipi_demux_relaxed(void)
321{
322 struct cpu_messages *info;
323 unsigned long all;
324
325 info = this_cpu_ptr(&ipi_message);
326 do {
327 all = xchg(&info->messages, 0);
328#if defined(CONFIG_KVM_XICS) && defined(CONFIG_KVM_BOOK3S_HV_POSSIBLE)
329 /*
330 * Must check for PPC_MSG_RM_HOST_ACTION messages
331 * before PPC_MSG_CALL_FUNCTION messages because when
332 * a VM is destroyed, we call kick_all_cpus_sync()
333 * to ensure that any pending PPC_MSG_RM_HOST_ACTION
334 * messages have completed before we free any VCPUs.
335 */
336 if (all & IPI_MESSAGE(PPC_MSG_RM_HOST_ACTION))
337 kvmppc_xics_ipi_action();
338#endif
339 if (all & IPI_MESSAGE(PPC_MSG_CALL_FUNCTION))
340 generic_smp_call_function_interrupt();
341 if (all & IPI_MESSAGE(PPC_MSG_RESCHEDULE))
342 scheduler_ipi();
343#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
344 if (all & IPI_MESSAGE(PPC_MSG_TICK_BROADCAST))
345 timer_broadcast_interrupt();
346#endif
347#ifdef CONFIG_NMI_IPI
348 if (all & IPI_MESSAGE(PPC_MSG_NMI_IPI))
349 nmi_ipi_action(0, NULL);
350#endif
351 } while (info->messages);
352
353 return IRQ_HANDLED;
354}
355#endif /* CONFIG_PPC_SMP_MUXED_IPI */
356
357static inline void do_message_pass(int cpu, int msg)
358{
359 if (smp_ops->message_pass)
360 smp_ops->message_pass(cpu, msg);
361#ifdef CONFIG_PPC_SMP_MUXED_IPI
362 else
363 smp_muxed_ipi_message_pass(cpu, msg);
364#endif
365}
366
367void smp_send_reschedule(int cpu)
368{
369 if (likely(smp_ops))
370 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
371}
372EXPORT_SYMBOL_GPL(smp_send_reschedule);
373
374void arch_send_call_function_single_ipi(int cpu)
375{
376 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
377}
378
379void arch_send_call_function_ipi_mask(const struct cpumask *mask)
380{
381 unsigned int cpu;
382
383 for_each_cpu(cpu, mask)
384 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
385}
386
387#ifdef CONFIG_NMI_IPI
388
389/*
390 * "NMI IPI" system.
391 *
392 * NMI IPIs may not be recoverable, so should not be used as ongoing part of
393 * a running system. They can be used for crash, debug, halt/reboot, etc.
394 *
395 * The IPI call waits with interrupts disabled until all targets enter the
396 * NMI handler, then returns. Subsequent IPIs can be issued before targets
397 * have returned from their handlers, so there is no guarantee about
398 * concurrency or re-entrancy.
399 *
400 * A new NMI can be issued before all targets exit the handler.
401 *
402 * The IPI call may time out without all targets entering the NMI handler.
403 * In that case, there is some logic to recover (and ignore subsequent
404 * NMI interrupts that may eventually be raised), but the platform interrupt
405 * handler may not be able to distinguish this from other exception causes,
406 * which may cause a crash.
407 */
408
409static atomic_t __nmi_ipi_lock = ATOMIC_INIT(0);
410static struct cpumask nmi_ipi_pending_mask;
411static bool nmi_ipi_busy = false;
412static void (*nmi_ipi_function)(struct pt_regs *) = NULL;
413
414noinstr static void nmi_ipi_lock_start(unsigned long *flags)
415{
416 raw_local_irq_save(*flags);
417 hard_irq_disable();
418 while (arch_atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1) {
419 raw_local_irq_restore(*flags);
420 spin_until_cond(arch_atomic_read(&__nmi_ipi_lock) == 0);
421 raw_local_irq_save(*flags);
422 hard_irq_disable();
423 }
424}
425
426noinstr static void nmi_ipi_lock(void)
427{
428 while (arch_atomic_cmpxchg(&__nmi_ipi_lock, 0, 1) == 1)
429 spin_until_cond(arch_atomic_read(&__nmi_ipi_lock) == 0);
430}
431
432noinstr static void nmi_ipi_unlock(void)
433{
434 smp_mb();
435 WARN_ON(arch_atomic_read(&__nmi_ipi_lock) != 1);
436 arch_atomic_set(&__nmi_ipi_lock, 0);
437}
438
439noinstr static void nmi_ipi_unlock_end(unsigned long *flags)
440{
441 nmi_ipi_unlock();
442 raw_local_irq_restore(*flags);
443}
444
445/*
446 * Platform NMI handler calls this to ack
447 */
448noinstr int smp_handle_nmi_ipi(struct pt_regs *regs)
449{
450 void (*fn)(struct pt_regs *) = NULL;
451 unsigned long flags;
452 int me = raw_smp_processor_id();
453 int ret = 0;
454
455 /*
456 * Unexpected NMIs are possible here because the interrupt may not
457 * be able to distinguish NMI IPIs from other types of NMIs, or
458 * because the caller may have timed out.
459 */
460 nmi_ipi_lock_start(&flags);
461 if (cpumask_test_cpu(me, &nmi_ipi_pending_mask)) {
462 cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
463 fn = READ_ONCE(nmi_ipi_function);
464 WARN_ON_ONCE(!fn);
465 ret = 1;
466 }
467 nmi_ipi_unlock_end(&flags);
468
469 if (fn)
470 fn(regs);
471
472 return ret;
473}
474
475static void do_smp_send_nmi_ipi(int cpu, bool safe)
476{
477 if (!safe && smp_ops->cause_nmi_ipi && smp_ops->cause_nmi_ipi(cpu))
478 return;
479
480 if (cpu >= 0) {
481 do_message_pass(cpu, PPC_MSG_NMI_IPI);
482 } else {
483 int c;
484
485 for_each_online_cpu(c) {
486 if (c == raw_smp_processor_id())
487 continue;
488 do_message_pass(c, PPC_MSG_NMI_IPI);
489 }
490 }
491}
492
493/*
494 * - cpu is the target CPU (must not be this CPU), or NMI_IPI_ALL_OTHERS.
495 * - fn is the target callback function.
496 * - delay_us > 0 is the delay before giving up waiting for targets to
497 * begin executing the handler, == 0 specifies indefinite delay.
498 */
499static int __smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *),
500 u64 delay_us, bool safe)
501{
502 unsigned long flags;
503 int me = raw_smp_processor_id();
504 int ret = 1;
505
506 BUG_ON(cpu == me);
507 BUG_ON(cpu < 0 && cpu != NMI_IPI_ALL_OTHERS);
508
509 if (unlikely(!smp_ops))
510 return 0;
511
512 nmi_ipi_lock_start(&flags);
513 while (nmi_ipi_busy) {
514 nmi_ipi_unlock_end(&flags);
515 spin_until_cond(!nmi_ipi_busy);
516 nmi_ipi_lock_start(&flags);
517 }
518 nmi_ipi_busy = true;
519 nmi_ipi_function = fn;
520
521 WARN_ON_ONCE(!cpumask_empty(&nmi_ipi_pending_mask));
522
523 if (cpu < 0) {
524 /* ALL_OTHERS */
525 cpumask_copy(&nmi_ipi_pending_mask, cpu_online_mask);
526 cpumask_clear_cpu(me, &nmi_ipi_pending_mask);
527 } else {
528 cpumask_set_cpu(cpu, &nmi_ipi_pending_mask);
529 }
530
531 nmi_ipi_unlock();
532
533 /* Interrupts remain hard disabled */
534
535 do_smp_send_nmi_ipi(cpu, safe);
536
537 nmi_ipi_lock();
538 /* nmi_ipi_busy is set here, so unlock/lock is okay */
539 while (!cpumask_empty(&nmi_ipi_pending_mask)) {
540 nmi_ipi_unlock();
541 udelay(1);
542 nmi_ipi_lock();
543 if (delay_us) {
544 delay_us--;
545 if (!delay_us)
546 break;
547 }
548 }
549
550 if (!cpumask_empty(&nmi_ipi_pending_mask)) {
551 /* Timeout waiting for CPUs to call smp_handle_nmi_ipi */
552 ret = 0;
553 cpumask_clear(&nmi_ipi_pending_mask);
554 }
555
556 nmi_ipi_function = NULL;
557 nmi_ipi_busy = false;
558
559 nmi_ipi_unlock_end(&flags);
560
561 return ret;
562}
563
564int smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
565{
566 return __smp_send_nmi_ipi(cpu, fn, delay_us, false);
567}
568
569int smp_send_safe_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us)
570{
571 return __smp_send_nmi_ipi(cpu, fn, delay_us, true);
572}
573#endif /* CONFIG_NMI_IPI */
574
575#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
576void tick_broadcast(const struct cpumask *mask)
577{
578 unsigned int cpu;
579
580 for_each_cpu(cpu, mask)
581 do_message_pass(cpu, PPC_MSG_TICK_BROADCAST);
582}
583#endif
584
585#ifdef CONFIG_DEBUGGER
586static void debugger_ipi_callback(struct pt_regs *regs)
587{
588 debugger_ipi(regs);
589}
590
591void smp_send_debugger_break(void)
592{
593 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, debugger_ipi_callback, 1000000);
594}
595#endif
596
597#ifdef CONFIG_KEXEC_CORE
598void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
599{
600 int cpu;
601
602 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, crash_ipi_callback, 1000000);
603 if (kdump_in_progress() && crash_wake_offline) {
604 for_each_present_cpu(cpu) {
605 if (cpu_online(cpu))
606 continue;
607 /*
608 * crash_ipi_callback will wait for
609 * all cpus, including offline CPUs.
610 * We don't care about nmi_ipi_function.
611 * Offline cpus will jump straight into
612 * crash_ipi_callback, we can skip the
613 * entire NMI dance and waiting for
614 * cpus to clear pending mask, etc.
615 */
616 do_smp_send_nmi_ipi(cpu, false);
617 }
618 }
619}
620#endif
621
622void crash_smp_send_stop(void)
623{
624 static bool stopped = false;
625
626 /*
627 * In case of fadump, register data for all CPUs is captured by f/w
628 * on ibm,os-term rtas call. Skip IPI callbacks to other CPUs before
629 * this rtas call to avoid tricky post processing of those CPUs'
630 * backtraces.
631 */
632 if (should_fadump_crash())
633 return;
634
635 if (stopped)
636 return;
637
638 stopped = true;
639
640#ifdef CONFIG_KEXEC_CORE
641 if (kexec_crash_image) {
642 crash_kexec_prepare();
643 return;
644 }
645#endif
646
647 smp_send_stop();
648}
649
650#ifdef CONFIG_NMI_IPI
651static void nmi_stop_this_cpu(struct pt_regs *regs)
652{
653 /*
654 * IRQs are already hard disabled by the smp_handle_nmi_ipi.
655 */
656 set_cpu_online(smp_processor_id(), false);
657
658 spin_begin();
659 while (1)
660 spin_cpu_relax();
661}
662
663void smp_send_stop(void)
664{
665 smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, nmi_stop_this_cpu, 1000000);
666}
667
668#else /* CONFIG_NMI_IPI */
669
670static void stop_this_cpu(void *dummy)
671{
672 hard_irq_disable();
673
674 /*
675 * Offlining CPUs in stop_this_cpu can result in scheduler warnings,
676 * (see commit de6e5d38417e), but printk_safe_flush_on_panic() wants
677 * to know other CPUs are offline before it breaks locks to flush
678 * printk buffers, in case we panic()ed while holding the lock.
679 */
680 set_cpu_online(smp_processor_id(), false);
681
682 spin_begin();
683 while (1)
684 spin_cpu_relax();
685}
686
687void smp_send_stop(void)
688{
689 static bool stopped = false;
690
691 /*
692 * Prevent waiting on csd lock from a previous smp_send_stop.
693 * This is racy, but in general callers try to do the right
694 * thing and only fire off one smp_send_stop (e.g., see
695 * kernel/panic.c)
696 */
697 if (stopped)
698 return;
699
700 stopped = true;
701
702 smp_call_function(stop_this_cpu, NULL, 0);
703}
704#endif /* CONFIG_NMI_IPI */
705
706static struct task_struct *current_set[NR_CPUS];
707
708static void smp_store_cpu_info(int id)
709{
710 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
711#ifdef CONFIG_PPC_E500
712 per_cpu(next_tlbcam_idx, id)
713 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
714#endif
715}
716
717/*
718 * Relationships between CPUs are maintained in a set of per-cpu cpumasks so
719 * rather than just passing around the cpumask we pass around a function that
720 * returns the that cpumask for the given CPU.
721 */
722static void set_cpus_related(int i, int j, struct cpumask *(*get_cpumask)(int))
723{
724 cpumask_set_cpu(i, get_cpumask(j));
725 cpumask_set_cpu(j, get_cpumask(i));
726}
727
728#ifdef CONFIG_HOTPLUG_CPU
729static void set_cpus_unrelated(int i, int j,
730 struct cpumask *(*get_cpumask)(int))
731{
732 cpumask_clear_cpu(i, get_cpumask(j));
733 cpumask_clear_cpu(j, get_cpumask(i));
734}
735#endif
736
737/*
738 * Extends set_cpus_related. Instead of setting one CPU at a time in
739 * dstmask, set srcmask at oneshot. dstmask should be super set of srcmask.
740 */
741static void or_cpumasks_related(int i, int j, struct cpumask *(*srcmask)(int),
742 struct cpumask *(*dstmask)(int))
743{
744 struct cpumask *mask;
745 int k;
746
747 mask = srcmask(j);
748 for_each_cpu(k, srcmask(i))
749 cpumask_or(dstmask(k), dstmask(k), mask);
750
751 if (i == j)
752 return;
753
754 mask = srcmask(i);
755 for_each_cpu(k, srcmask(j))
756 cpumask_or(dstmask(k), dstmask(k), mask);
757}
758
759/*
760 * parse_thread_groups: Parses the "ibm,thread-groups" device tree
761 * property for the CPU device node @dn and stores
762 * the parsed output in the thread_groups_list
763 * structure @tglp.
764 *
765 * @dn: The device node of the CPU device.
766 * @tglp: Pointer to a thread group list structure into which the parsed
767 * output of "ibm,thread-groups" is stored.
768 *
769 * ibm,thread-groups[0..N-1] array defines which group of threads in
770 * the CPU-device node can be grouped together based on the property.
771 *
772 * This array can represent thread groupings for multiple properties.
773 *
774 * ibm,thread-groups[i + 0] tells us the property based on which the
775 * threads are being grouped together. If this value is 1, it implies
776 * that the threads in the same group share L1, translation cache. If
777 * the value is 2, it implies that the threads in the same group share
778 * the same L2 cache.
779 *
780 * ibm,thread-groups[i+1] tells us how many such thread groups exist for the
781 * property ibm,thread-groups[i]
782 *
783 * ibm,thread-groups[i+2] tells us the number of threads in each such
784 * group.
785 * Suppose k = (ibm,thread-groups[i+1] * ibm,thread-groups[i+2]), then,
786 *
787 * ibm,thread-groups[i+3..i+k+2] (is the list of threads identified by
788 * "ibm,ppc-interrupt-server#s" arranged as per their membership in
789 * the grouping.
790 *
791 * Example:
792 * 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]
793 * This can be decomposed up into two consecutive arrays:
794 * a) [1,2,4,8,10,12,14,9,11,13,15]
795 * b) [2,2,4,8,10,12,14,9,11,13,15]
796 *
797 * where in,
798 *
799 * a) provides information of Property "1" being shared by "2" groups,
800 * each with "4" threads each. The "ibm,ppc-interrupt-server#s" of
801 * the first group is {8,10,12,14} and the
802 * "ibm,ppc-interrupt-server#s" of the second group is
803 * {9,11,13,15}. Property "1" is indicative of the thread in the
804 * group sharing L1 cache, translation cache and Instruction Data
805 * flow.
806 *
807 * b) provides information of Property "2" being shared by "2" groups,
808 * each group with "4" threads. The "ibm,ppc-interrupt-server#s" of
809 * the first group is {8,10,12,14} and the
810 * "ibm,ppc-interrupt-server#s" of the second group is
811 * {9,11,13,15}. Property "2" indicates that the threads in each
812 * group share the L2-cache.
813 *
814 * Returns 0 on success, -EINVAL if the property does not exist,
815 * -ENODATA if property does not have a value, and -EOVERFLOW if the
816 * property data isn't large enough.
817 */
818static int parse_thread_groups(struct device_node *dn,
819 struct thread_groups_list *tglp)
820{
821 unsigned int property_idx = 0;
822 u32 *thread_group_array;
823 size_t total_threads;
824 int ret = 0, count;
825 u32 *thread_list;
826 int i = 0;
827
828 count = of_property_count_u32_elems(dn, "ibm,thread-groups");
829 thread_group_array = kcalloc(count, sizeof(u32), GFP_KERNEL);
830 ret = of_property_read_u32_array(dn, "ibm,thread-groups",
831 thread_group_array, count);
832 if (ret)
833 goto out_free;
834
835 while (i < count && property_idx < MAX_THREAD_GROUP_PROPERTIES) {
836 int j;
837 struct thread_groups *tg = &tglp->property_tgs[property_idx++];
838
839 tg->property = thread_group_array[i];
840 tg->nr_groups = thread_group_array[i + 1];
841 tg->threads_per_group = thread_group_array[i + 2];
842 total_threads = tg->nr_groups * tg->threads_per_group;
843
844 thread_list = &thread_group_array[i + 3];
845
846 for (j = 0; j < total_threads; j++)
847 tg->thread_list[j] = thread_list[j];
848 i = i + 3 + total_threads;
849 }
850
851 tglp->nr_properties = property_idx;
852
853out_free:
854 kfree(thread_group_array);
855 return ret;
856}
857
858/*
859 * get_cpu_thread_group_start : Searches the thread group in tg->thread_list
860 * that @cpu belongs to.
861 *
862 * @cpu : The logical CPU whose thread group is being searched.
863 * @tg : The thread-group structure of the CPU node which @cpu belongs
864 * to.
865 *
866 * Returns the index to tg->thread_list that points to the start
867 * of the thread_group that @cpu belongs to.
868 *
869 * Returns -1 if cpu doesn't belong to any of the groups pointed to by
870 * tg->thread_list.
871 */
872static int get_cpu_thread_group_start(int cpu, struct thread_groups *tg)
873{
874 int hw_cpu_id = get_hard_smp_processor_id(cpu);
875 int i, j;
876
877 for (i = 0; i < tg->nr_groups; i++) {
878 int group_start = i * tg->threads_per_group;
879
880 for (j = 0; j < tg->threads_per_group; j++) {
881 int idx = group_start + j;
882
883 if (tg->thread_list[idx] == hw_cpu_id)
884 return group_start;
885 }
886 }
887
888 return -1;
889}
890
891static struct thread_groups *__init get_thread_groups(int cpu,
892 int group_property,
893 int *err)
894{
895 struct device_node *dn = of_get_cpu_node(cpu, NULL);
896 struct thread_groups_list *cpu_tgl = &tgl[cpu];
897 struct thread_groups *tg = NULL;
898 int i;
899 *err = 0;
900
901 if (!dn) {
902 *err = -ENODATA;
903 return NULL;
904 }
905
906 if (!cpu_tgl->nr_properties) {
907 *err = parse_thread_groups(dn, cpu_tgl);
908 if (*err)
909 goto out;
910 }
911
912 for (i = 0; i < cpu_tgl->nr_properties; i++) {
913 if (cpu_tgl->property_tgs[i].property == group_property) {
914 tg = &cpu_tgl->property_tgs[i];
915 break;
916 }
917 }
918
919 if (!tg)
920 *err = -EINVAL;
921out:
922 of_node_put(dn);
923 return tg;
924}
925
926static int __init update_mask_from_threadgroup(cpumask_var_t *mask, struct thread_groups *tg,
927 int cpu, int cpu_group_start)
928{
929 int first_thread = cpu_first_thread_sibling(cpu);
930 int i;
931
932 zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cpu));
933
934 for (i = first_thread; i < first_thread + threads_per_core; i++) {
935 int i_group_start = get_cpu_thread_group_start(i, tg);
936
937 if (unlikely(i_group_start == -1)) {
938 WARN_ON_ONCE(1);
939 return -ENODATA;
940 }
941
942 if (i_group_start == cpu_group_start)
943 cpumask_set_cpu(i, *mask);
944 }
945
946 return 0;
947}
948
949static int __init init_thread_group_cache_map(int cpu, int cache_property)
950
951{
952 int cpu_group_start = -1, err = 0;
953 struct thread_groups *tg = NULL;
954 cpumask_var_t *mask = NULL;
955
956 if (cache_property != THREAD_GROUP_SHARE_L1 &&
957 cache_property != THREAD_GROUP_SHARE_L2_L3)
958 return -EINVAL;
959
960 tg = get_thread_groups(cpu, cache_property, &err);
961
962 if (!tg)
963 return err;
964
965 cpu_group_start = get_cpu_thread_group_start(cpu, tg);
966
967 if (unlikely(cpu_group_start == -1)) {
968 WARN_ON_ONCE(1);
969 return -ENODATA;
970 }
971
972 if (cache_property == THREAD_GROUP_SHARE_L1) {
973 mask = &per_cpu(thread_group_l1_cache_map, cpu);
974 update_mask_from_threadgroup(mask, tg, cpu, cpu_group_start);
975 }
976 else if (cache_property == THREAD_GROUP_SHARE_L2_L3) {
977 mask = &per_cpu(thread_group_l2_cache_map, cpu);
978 update_mask_from_threadgroup(mask, tg, cpu, cpu_group_start);
979 mask = &per_cpu(thread_group_l3_cache_map, cpu);
980 update_mask_from_threadgroup(mask, tg, cpu, cpu_group_start);
981 }
982
983
984 return 0;
985}
986
987static bool shared_caches;
988
989#ifdef CONFIG_SCHED_SMT
990/* cpumask of CPUs with asymmetric SMT dependency */
991static int powerpc_smt_flags(void)
992{
993 int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
994
995 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
996 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
997 flags |= SD_ASYM_PACKING;
998 }
999 return flags;
1000}
1001#endif
1002
1003/*
1004 * P9 has a slightly odd architecture where pairs of cores share an L2 cache.
1005 * This topology makes it *much* cheaper to migrate tasks between adjacent cores
1006 * since the migrated task remains cache hot. We want to take advantage of this
1007 * at the scheduler level so an extra topology level is required.
1008 */
1009static int powerpc_shared_cache_flags(void)
1010{
1011 return SD_SHARE_PKG_RESOURCES;
1012}
1013
1014/*
1015 * We can't just pass cpu_l2_cache_mask() directly because
1016 * returns a non-const pointer and the compiler barfs on that.
1017 */
1018static const struct cpumask *shared_cache_mask(int cpu)
1019{
1020 return per_cpu(cpu_l2_cache_map, cpu);
1021}
1022
1023#ifdef CONFIG_SCHED_SMT
1024static const struct cpumask *smallcore_smt_mask(int cpu)
1025{
1026 return cpu_smallcore_mask(cpu);
1027}
1028#endif
1029
1030static struct cpumask *cpu_coregroup_mask(int cpu)
1031{
1032 return per_cpu(cpu_coregroup_map, cpu);
1033}
1034
1035static bool has_coregroup_support(void)
1036{
1037 return coregroup_enabled;
1038}
1039
1040static const struct cpumask *cpu_mc_mask(int cpu)
1041{
1042 return cpu_coregroup_mask(cpu);
1043}
1044
1045static struct sched_domain_topology_level powerpc_topology[] = {
1046#ifdef CONFIG_SCHED_SMT
1047 { cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
1048#endif
1049 { shared_cache_mask, powerpc_shared_cache_flags, SD_INIT_NAME(CACHE) },
1050 { cpu_mc_mask, SD_INIT_NAME(MC) },
1051 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1052 { NULL, },
1053};
1054
1055static int __init init_big_cores(void)
1056{
1057 int cpu;
1058
1059 for_each_possible_cpu(cpu) {
1060 int err = init_thread_group_cache_map(cpu, THREAD_GROUP_SHARE_L1);
1061
1062 if (err)
1063 return err;
1064
1065 zalloc_cpumask_var_node(&per_cpu(cpu_smallcore_map, cpu),
1066 GFP_KERNEL,
1067 cpu_to_node(cpu));
1068 }
1069
1070 has_big_cores = true;
1071
1072 for_each_possible_cpu(cpu) {
1073 int err = init_thread_group_cache_map(cpu, THREAD_GROUP_SHARE_L2_L3);
1074
1075 if (err)
1076 return err;
1077 }
1078
1079 thread_group_shares_l2 = true;
1080 thread_group_shares_l3 = true;
1081 pr_debug("L2/L3 cache only shared by the threads in the small core\n");
1082
1083 return 0;
1084}
1085
1086void __init smp_prepare_cpus(unsigned int max_cpus)
1087{
1088 unsigned int cpu;
1089
1090 DBG("smp_prepare_cpus\n");
1091
1092 /*
1093 * setup_cpu may need to be called on the boot cpu. We haven't
1094 * spun any cpus up but lets be paranoid.
1095 */
1096 BUG_ON(boot_cpuid != smp_processor_id());
1097
1098 /* Fixup boot cpu */
1099 smp_store_cpu_info(boot_cpuid);
1100 cpu_callin_map[boot_cpuid] = 1;
1101
1102 for_each_possible_cpu(cpu) {
1103 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
1104 GFP_KERNEL, cpu_to_node(cpu));
1105 zalloc_cpumask_var_node(&per_cpu(cpu_l2_cache_map, cpu),
1106 GFP_KERNEL, cpu_to_node(cpu));
1107 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
1108 GFP_KERNEL, cpu_to_node(cpu));
1109 if (has_coregroup_support())
1110 zalloc_cpumask_var_node(&per_cpu(cpu_coregroup_map, cpu),
1111 GFP_KERNEL, cpu_to_node(cpu));
1112
1113#ifdef CONFIG_NUMA
1114 /*
1115 * numa_node_id() works after this.
1116 */
1117 if (cpu_present(cpu)) {
1118 set_cpu_numa_node(cpu, numa_cpu_lookup_table[cpu]);
1119 set_cpu_numa_mem(cpu,
1120 local_memory_node(numa_cpu_lookup_table[cpu]));
1121 }
1122#endif
1123 }
1124
1125 /* Init the cpumasks so the boot CPU is related to itself */
1126 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
1127 cpumask_set_cpu(boot_cpuid, cpu_l2_cache_mask(boot_cpuid));
1128 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
1129
1130 if (has_coregroup_support())
1131 cpumask_set_cpu(boot_cpuid, cpu_coregroup_mask(boot_cpuid));
1132
1133 init_big_cores();
1134 if (has_big_cores) {
1135 cpumask_set_cpu(boot_cpuid,
1136 cpu_smallcore_mask(boot_cpuid));
1137 }
1138
1139 if (cpu_to_chip_id(boot_cpuid) != -1) {
1140 int idx = DIV_ROUND_UP(num_possible_cpus(), threads_per_core);
1141
1142 /*
1143 * All threads of a core will all belong to the same core,
1144 * chip_id_lookup_table will have one entry per core.
1145 * Assumption: if boot_cpuid doesn't have a chip-id, then no
1146 * other CPUs, will also not have chip-id.
1147 */
1148 chip_id_lookup_table = kcalloc(idx, sizeof(int), GFP_KERNEL);
1149 if (chip_id_lookup_table)
1150 memset(chip_id_lookup_table, -1, sizeof(int) * idx);
1151 }
1152
1153 if (smp_ops && smp_ops->probe)
1154 smp_ops->probe();
1155}
1156
1157void smp_prepare_boot_cpu(void)
1158{
1159 BUG_ON(smp_processor_id() != boot_cpuid);
1160#ifdef CONFIG_PPC64
1161 paca_ptrs[boot_cpuid]->__current = current;
1162#endif
1163 set_numa_node(numa_cpu_lookup_table[boot_cpuid]);
1164 current_set[boot_cpuid] = current;
1165}
1166
1167#ifdef CONFIG_HOTPLUG_CPU
1168
1169int generic_cpu_disable(void)
1170{
1171 unsigned int cpu = smp_processor_id();
1172
1173 if (cpu == boot_cpuid)
1174 return -EBUSY;
1175
1176 set_cpu_online(cpu, false);
1177#ifdef CONFIG_PPC64
1178 vdso_data->processorCount--;
1179#endif
1180 /* Update affinity of all IRQs previously aimed at this CPU */
1181 irq_migrate_all_off_this_cpu();
1182
1183 /*
1184 * Depending on the details of the interrupt controller, it's possible
1185 * that one of the interrupts we just migrated away from this CPU is
1186 * actually already pending on this CPU. If we leave it in that state
1187 * the interrupt will never be EOI'ed, and will never fire again. So
1188 * temporarily enable interrupts here, to allow any pending interrupt to
1189 * be received (and EOI'ed), before we take this CPU offline.
1190 */
1191 local_irq_enable();
1192 mdelay(1);
1193 local_irq_disable();
1194
1195 return 0;
1196}
1197
1198void generic_cpu_die(unsigned int cpu)
1199{
1200 int i;
1201
1202 for (i = 0; i < 100; i++) {
1203 smp_rmb();
1204 if (is_cpu_dead(cpu))
1205 return;
1206 msleep(100);
1207 }
1208 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
1209}
1210
1211void generic_set_cpu_dead(unsigned int cpu)
1212{
1213 per_cpu(cpu_state, cpu) = CPU_DEAD;
1214}
1215
1216/*
1217 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
1218 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
1219 * which makes the delay in generic_cpu_die() not happen.
1220 */
1221void generic_set_cpu_up(unsigned int cpu)
1222{
1223 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1224}
1225
1226int generic_check_cpu_restart(unsigned int cpu)
1227{
1228 return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
1229}
1230
1231int is_cpu_dead(unsigned int cpu)
1232{
1233 return per_cpu(cpu_state, cpu) == CPU_DEAD;
1234}
1235
1236static bool secondaries_inhibited(void)
1237{
1238 return kvm_hv_mode_active();
1239}
1240
1241#else /* HOTPLUG_CPU */
1242
1243#define secondaries_inhibited() 0
1244
1245#endif
1246
1247static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
1248{
1249#ifdef CONFIG_PPC64
1250 paca_ptrs[cpu]->__current = idle;
1251 paca_ptrs[cpu]->kstack = (unsigned long)task_stack_page(idle) +
1252 THREAD_SIZE - STACK_FRAME_MIN_SIZE;
1253#endif
1254 task_thread_info(idle)->cpu = cpu;
1255 secondary_current = current_set[cpu] = idle;
1256}
1257
1258int __cpu_up(unsigned int cpu, struct task_struct *tidle)
1259{
1260 const unsigned long boot_spin_ms = 5 * MSEC_PER_SEC;
1261 const bool booting = system_state < SYSTEM_RUNNING;
1262 const unsigned long hp_spin_ms = 1;
1263 unsigned long deadline;
1264 int rc;
1265 const unsigned long spin_wait_ms = booting ? boot_spin_ms : hp_spin_ms;
1266
1267 /*
1268 * Don't allow secondary threads to come online if inhibited
1269 */
1270 if (threads_per_core > 1 && secondaries_inhibited() &&
1271 cpu_thread_in_subcore(cpu))
1272 return -EBUSY;
1273
1274 if (smp_ops == NULL ||
1275 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
1276 return -EINVAL;
1277
1278 cpu_idle_thread_init(cpu, tidle);
1279
1280 /*
1281 * The platform might need to allocate resources prior to bringing
1282 * up the CPU
1283 */
1284 if (smp_ops->prepare_cpu) {
1285 rc = smp_ops->prepare_cpu(cpu);
1286 if (rc)
1287 return rc;
1288 }
1289
1290 /* Make sure callin-map entry is 0 (can be leftover a CPU
1291 * hotplug
1292 */
1293 cpu_callin_map[cpu] = 0;
1294
1295 /* The information for processor bringup must
1296 * be written out to main store before we release
1297 * the processor.
1298 */
1299 smp_mb();
1300
1301 /* wake up cpus */
1302 DBG("smp: kicking cpu %d\n", cpu);
1303 rc = smp_ops->kick_cpu(cpu);
1304 if (rc) {
1305 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
1306 return rc;
1307 }
1308
1309 /*
1310 * At boot time, simply spin on the callin word until the
1311 * deadline passes.
1312 *
1313 * At run time, spin for an optimistic amount of time to avoid
1314 * sleeping in the common case.
1315 */
1316 deadline = jiffies + msecs_to_jiffies(spin_wait_ms);
1317 spin_until_cond(cpu_callin_map[cpu] || time_is_before_jiffies(deadline));
1318
1319 if (!cpu_callin_map[cpu] && system_state >= SYSTEM_RUNNING) {
1320 const unsigned long sleep_interval_us = 10 * USEC_PER_MSEC;
1321 const unsigned long sleep_wait_ms = 100 * MSEC_PER_SEC;
1322
1323 deadline = jiffies + msecs_to_jiffies(sleep_wait_ms);
1324 while (!cpu_callin_map[cpu] && time_is_after_jiffies(deadline))
1325 fsleep(sleep_interval_us);
1326 }
1327
1328 if (!cpu_callin_map[cpu]) {
1329 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
1330 return -ENOENT;
1331 }
1332
1333 DBG("Processor %u found.\n", cpu);
1334
1335 if (smp_ops->give_timebase)
1336 smp_ops->give_timebase();
1337
1338 /* Wait until cpu puts itself in the online & active maps */
1339 spin_until_cond(cpu_online(cpu));
1340
1341 return 0;
1342}
1343
1344/* Return the value of the reg property corresponding to the given
1345 * logical cpu.
1346 */
1347int cpu_to_core_id(int cpu)
1348{
1349 struct device_node *np;
1350 int id = -1;
1351
1352 np = of_get_cpu_node(cpu, NULL);
1353 if (!np)
1354 goto out;
1355
1356 id = of_get_cpu_hwid(np, 0);
1357out:
1358 of_node_put(np);
1359 return id;
1360}
1361EXPORT_SYMBOL_GPL(cpu_to_core_id);
1362
1363/* Helper routines for cpu to core mapping */
1364int cpu_core_index_of_thread(int cpu)
1365{
1366 return cpu >> threads_shift;
1367}
1368EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
1369
1370int cpu_first_thread_of_core(int core)
1371{
1372 return core << threads_shift;
1373}
1374EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
1375
1376/* Must be called when no change can occur to cpu_present_mask,
1377 * i.e. during cpu online or offline.
1378 */
1379static struct device_node *cpu_to_l2cache(int cpu)
1380{
1381 struct device_node *np;
1382 struct device_node *cache;
1383
1384 if (!cpu_present(cpu))
1385 return NULL;
1386
1387 np = of_get_cpu_node(cpu, NULL);
1388 if (np == NULL)
1389 return NULL;
1390
1391 cache = of_find_next_cache_node(np);
1392
1393 of_node_put(np);
1394
1395 return cache;
1396}
1397
1398static bool update_mask_by_l2(int cpu, cpumask_var_t *mask)
1399{
1400 struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1401 struct device_node *l2_cache, *np;
1402 int i;
1403
1404 if (has_big_cores)
1405 submask_fn = cpu_smallcore_mask;
1406
1407 /*
1408 * If the threads in a thread-group share L2 cache, then the
1409 * L2-mask can be obtained from thread_group_l2_cache_map.
1410 */
1411 if (thread_group_shares_l2) {
1412 cpumask_set_cpu(cpu, cpu_l2_cache_mask(cpu));
1413
1414 for_each_cpu(i, per_cpu(thread_group_l2_cache_map, cpu)) {
1415 if (cpu_online(i))
1416 set_cpus_related(i, cpu, cpu_l2_cache_mask);
1417 }
1418
1419 /* Verify that L1-cache siblings are a subset of L2 cache-siblings */
1420 if (!cpumask_equal(submask_fn(cpu), cpu_l2_cache_mask(cpu)) &&
1421 !cpumask_subset(submask_fn(cpu), cpu_l2_cache_mask(cpu))) {
1422 pr_warn_once("CPU %d : Inconsistent L1 and L2 cache siblings\n",
1423 cpu);
1424 }
1425
1426 return true;
1427 }
1428
1429 l2_cache = cpu_to_l2cache(cpu);
1430 if (!l2_cache || !*mask) {
1431 /* Assume only core siblings share cache with this CPU */
1432 for_each_cpu(i, cpu_sibling_mask(cpu))
1433 set_cpus_related(cpu, i, cpu_l2_cache_mask);
1434
1435 return false;
1436 }
1437
1438 cpumask_and(*mask, cpu_online_mask, cpu_cpu_mask(cpu));
1439
1440 /* Update l2-cache mask with all the CPUs that are part of submask */
1441 or_cpumasks_related(cpu, cpu, submask_fn, cpu_l2_cache_mask);
1442
1443 /* Skip all CPUs already part of current CPU l2-cache mask */
1444 cpumask_andnot(*mask, *mask, cpu_l2_cache_mask(cpu));
1445
1446 for_each_cpu(i, *mask) {
1447 /*
1448 * when updating the marks the current CPU has not been marked
1449 * online, but we need to update the cache masks
1450 */
1451 np = cpu_to_l2cache(i);
1452
1453 /* Skip all CPUs already part of current CPU l2-cache */
1454 if (np == l2_cache) {
1455 or_cpumasks_related(cpu, i, submask_fn, cpu_l2_cache_mask);
1456 cpumask_andnot(*mask, *mask, submask_fn(i));
1457 } else {
1458 cpumask_andnot(*mask, *mask, cpu_l2_cache_mask(i));
1459 }
1460
1461 of_node_put(np);
1462 }
1463 of_node_put(l2_cache);
1464
1465 return true;
1466}
1467
1468#ifdef CONFIG_HOTPLUG_CPU
1469static void remove_cpu_from_masks(int cpu)
1470{
1471 struct cpumask *(*mask_fn)(int) = cpu_sibling_mask;
1472 int i;
1473
1474 unmap_cpu_from_node(cpu);
1475
1476 if (shared_caches)
1477 mask_fn = cpu_l2_cache_mask;
1478
1479 for_each_cpu(i, mask_fn(cpu)) {
1480 set_cpus_unrelated(cpu, i, cpu_l2_cache_mask);
1481 set_cpus_unrelated(cpu, i, cpu_sibling_mask);
1482 if (has_big_cores)
1483 set_cpus_unrelated(cpu, i, cpu_smallcore_mask);
1484 }
1485
1486 for_each_cpu(i, cpu_core_mask(cpu))
1487 set_cpus_unrelated(cpu, i, cpu_core_mask);
1488
1489 if (has_coregroup_support()) {
1490 for_each_cpu(i, cpu_coregroup_mask(cpu))
1491 set_cpus_unrelated(cpu, i, cpu_coregroup_mask);
1492 }
1493}
1494#endif
1495
1496static inline void add_cpu_to_smallcore_masks(int cpu)
1497{
1498 int i;
1499
1500 if (!has_big_cores)
1501 return;
1502
1503 cpumask_set_cpu(cpu, cpu_smallcore_mask(cpu));
1504
1505 for_each_cpu(i, per_cpu(thread_group_l1_cache_map, cpu)) {
1506 if (cpu_online(i))
1507 set_cpus_related(i, cpu, cpu_smallcore_mask);
1508 }
1509}
1510
1511static void update_coregroup_mask(int cpu, cpumask_var_t *mask)
1512{
1513 struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1514 int coregroup_id = cpu_to_coregroup_id(cpu);
1515 int i;
1516
1517 if (shared_caches)
1518 submask_fn = cpu_l2_cache_mask;
1519
1520 if (!*mask) {
1521 /* Assume only siblings are part of this CPU's coregroup */
1522 for_each_cpu(i, submask_fn(cpu))
1523 set_cpus_related(cpu, i, cpu_coregroup_mask);
1524
1525 return;
1526 }
1527
1528 cpumask_and(*mask, cpu_online_mask, cpu_cpu_mask(cpu));
1529
1530 /* Update coregroup mask with all the CPUs that are part of submask */
1531 or_cpumasks_related(cpu, cpu, submask_fn, cpu_coregroup_mask);
1532
1533 /* Skip all CPUs already part of coregroup mask */
1534 cpumask_andnot(*mask, *mask, cpu_coregroup_mask(cpu));
1535
1536 for_each_cpu(i, *mask) {
1537 /* Skip all CPUs not part of this coregroup */
1538 if (coregroup_id == cpu_to_coregroup_id(i)) {
1539 or_cpumasks_related(cpu, i, submask_fn, cpu_coregroup_mask);
1540 cpumask_andnot(*mask, *mask, submask_fn(i));
1541 } else {
1542 cpumask_andnot(*mask, *mask, cpu_coregroup_mask(i));
1543 }
1544 }
1545}
1546
1547static void add_cpu_to_masks(int cpu)
1548{
1549 struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1550 int first_thread = cpu_first_thread_sibling(cpu);
1551 cpumask_var_t mask;
1552 int chip_id = -1;
1553 bool ret;
1554 int i;
1555
1556 /*
1557 * This CPU will not be in the online mask yet so we need to manually
1558 * add it to it's own thread sibling mask.
1559 */
1560 map_cpu_to_node(cpu, cpu_to_node(cpu));
1561 cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
1562 cpumask_set_cpu(cpu, cpu_core_mask(cpu));
1563
1564 for (i = first_thread; i < first_thread + threads_per_core; i++)
1565 if (cpu_online(i))
1566 set_cpus_related(i, cpu, cpu_sibling_mask);
1567
1568 add_cpu_to_smallcore_masks(cpu);
1569
1570 /* In CPU-hotplug path, hence use GFP_ATOMIC */
1571 ret = alloc_cpumask_var_node(&mask, GFP_ATOMIC, cpu_to_node(cpu));
1572 update_mask_by_l2(cpu, &mask);
1573
1574 if (has_coregroup_support())
1575 update_coregroup_mask(cpu, &mask);
1576
1577 if (chip_id_lookup_table && ret)
1578 chip_id = cpu_to_chip_id(cpu);
1579
1580 if (shared_caches)
1581 submask_fn = cpu_l2_cache_mask;
1582
1583 /* Update core_mask with all the CPUs that are part of submask */
1584 or_cpumasks_related(cpu, cpu, submask_fn, cpu_core_mask);
1585
1586 /* Skip all CPUs already part of current CPU core mask */
1587 cpumask_andnot(mask, cpu_online_mask, cpu_core_mask(cpu));
1588
1589 /* If chip_id is -1; limit the cpu_core_mask to within DIE*/
1590 if (chip_id == -1)
1591 cpumask_and(mask, mask, cpu_cpu_mask(cpu));
1592
1593 for_each_cpu(i, mask) {
1594 if (chip_id == cpu_to_chip_id(i)) {
1595 or_cpumasks_related(cpu, i, submask_fn, cpu_core_mask);
1596 cpumask_andnot(mask, mask, submask_fn(i));
1597 } else {
1598 cpumask_andnot(mask, mask, cpu_core_mask(i));
1599 }
1600 }
1601
1602 free_cpumask_var(mask);
1603}
1604
1605/* Activate a secondary processor. */
1606void start_secondary(void *unused)
1607{
1608 unsigned int cpu = raw_smp_processor_id();
1609
1610 /* PPC64 calls setup_kup() in early_setup_secondary() */
1611 if (IS_ENABLED(CONFIG_PPC32))
1612 setup_kup();
1613
1614 mmgrab(&init_mm);
1615 current->active_mm = &init_mm;
1616
1617 smp_store_cpu_info(cpu);
1618 set_dec(tb_ticks_per_jiffy);
1619 rcu_cpu_starting(cpu);
1620 cpu_callin_map[cpu] = 1;
1621
1622 if (smp_ops->setup_cpu)
1623 smp_ops->setup_cpu(cpu);
1624 if (smp_ops->take_timebase)
1625 smp_ops->take_timebase();
1626
1627 secondary_cpu_time_init();
1628
1629#ifdef CONFIG_PPC64
1630 if (system_state == SYSTEM_RUNNING)
1631 vdso_data->processorCount++;
1632
1633 vdso_getcpu_init();
1634#endif
1635 set_numa_node(numa_cpu_lookup_table[cpu]);
1636 set_numa_mem(local_memory_node(numa_cpu_lookup_table[cpu]));
1637
1638 /* Update topology CPU masks */
1639 add_cpu_to_masks(cpu);
1640
1641 /*
1642 * Check for any shared caches. Note that this must be done on a
1643 * per-core basis because one core in the pair might be disabled.
1644 */
1645 if (!shared_caches) {
1646 struct cpumask *(*sibling_mask)(int) = cpu_sibling_mask;
1647 struct cpumask *mask = cpu_l2_cache_mask(cpu);
1648
1649 if (has_big_cores)
1650 sibling_mask = cpu_smallcore_mask;
1651
1652 if (cpumask_weight(mask) > cpumask_weight(sibling_mask(cpu)))
1653 shared_caches = true;
1654 }
1655
1656 smp_wmb();
1657 notify_cpu_starting(cpu);
1658 set_cpu_online(cpu, true);
1659
1660 boot_init_stack_canary();
1661
1662 local_irq_enable();
1663
1664 /* We can enable ftrace for secondary cpus now */
1665 this_cpu_enable_ftrace();
1666
1667 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
1668
1669 BUG();
1670}
1671
1672static void __init fixup_topology(void)
1673{
1674 int i;
1675
1676#ifdef CONFIG_SCHED_SMT
1677 if (has_big_cores) {
1678 pr_info("Big cores detected but using small core scheduling\n");
1679 powerpc_topology[smt_idx].mask = smallcore_smt_mask;
1680 }
1681#endif
1682
1683 if (!has_coregroup_support())
1684 powerpc_topology[mc_idx].mask = powerpc_topology[cache_idx].mask;
1685
1686 /*
1687 * Try to consolidate topology levels here instead of
1688 * allowing scheduler to degenerate.
1689 * - Dont consolidate if masks are different.
1690 * - Dont consolidate if sd_flags exists and are different.
1691 */
1692 for (i = 1; i <= die_idx; i++) {
1693 if (powerpc_topology[i].mask != powerpc_topology[i - 1].mask)
1694 continue;
1695
1696 if (powerpc_topology[i].sd_flags && powerpc_topology[i - 1].sd_flags &&
1697 powerpc_topology[i].sd_flags != powerpc_topology[i - 1].sd_flags)
1698 continue;
1699
1700 if (!powerpc_topology[i - 1].sd_flags)
1701 powerpc_topology[i - 1].sd_flags = powerpc_topology[i].sd_flags;
1702
1703 powerpc_topology[i].mask = powerpc_topology[i + 1].mask;
1704 powerpc_topology[i].sd_flags = powerpc_topology[i + 1].sd_flags;
1705#ifdef CONFIG_SCHED_DEBUG
1706 powerpc_topology[i].name = powerpc_topology[i + 1].name;
1707#endif
1708 }
1709}
1710
1711void __init smp_cpus_done(unsigned int max_cpus)
1712{
1713 /*
1714 * We are running pinned to the boot CPU, see rest_init().
1715 */
1716 if (smp_ops && smp_ops->setup_cpu)
1717 smp_ops->setup_cpu(boot_cpuid);
1718
1719 if (smp_ops && smp_ops->bringup_done)
1720 smp_ops->bringup_done();
1721
1722 dump_numa_cpu_topology();
1723
1724 fixup_topology();
1725 set_sched_topology(powerpc_topology);
1726}
1727
1728#ifdef CONFIG_HOTPLUG_CPU
1729int __cpu_disable(void)
1730{
1731 int cpu = smp_processor_id();
1732 int err;
1733
1734 if (!smp_ops->cpu_disable)
1735 return -ENOSYS;
1736
1737 this_cpu_disable_ftrace();
1738
1739 err = smp_ops->cpu_disable();
1740 if (err)
1741 return err;
1742
1743 /* Update sibling maps */
1744 remove_cpu_from_masks(cpu);
1745
1746 return 0;
1747}
1748
1749void __cpu_die(unsigned int cpu)
1750{
1751 if (smp_ops->cpu_die)
1752 smp_ops->cpu_die(cpu);
1753}
1754
1755void arch_cpu_idle_dead(void)
1756{
1757 /*
1758 * Disable on the down path. This will be re-enabled by
1759 * start_secondary() via start_secondary_resume() below
1760 */
1761 this_cpu_disable_ftrace();
1762
1763 if (smp_ops->cpu_offline_self)
1764 smp_ops->cpu_offline_self();
1765
1766 /* If we return, we re-enter start_secondary */
1767 start_secondary_resume();
1768}
1769
1770#endif