4 * Written by Cort Dougan (cort@cs.nmt.edu) borrowing a great
5 * deal of code from the sparc and intel versions.
7 * Copyright (C) 1999 Cort Dougan <cort@cs.nmt.edu>
9 * PowerPC-64 Support added by Dave Engebretsen, Peter Bergner, and
10 * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
12 * This program is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU General Public License
14 * as published by the Free Software Foundation; either version
15 * 2 of the License, or (at your option) any later version.
20 #include <linux/kernel.h>
21 #include <linux/export.h>
22 #include <linux/sched.h>
23 #include <linux/smp.h>
24 #include <linux/interrupt.h>
25 #include <linux/delay.h>
26 #include <linux/init.h>
27 #include <linux/spinlock.h>
28 #include <linux/cache.h>
29 #include <linux/err.h>
30 #include <linux/device.h>
31 #include <linux/cpu.h>
32 #include <linux/notifier.h>
33 #include <linux/topology.h>
35 #include <asm/ptrace.h>
36 #include <linux/atomic.h>
38 #include <asm/hw_irq.h>
39 #include <asm/kvm_ppc.h>
41 #include <asm/pgtable.h>
45 #include <asm/machdep.h>
46 #include <asm/cputhreads.h>
47 #include <asm/cputable.h>
49 #include <asm/vdso_datapage.h>
54 #include <asm/debug.h>
55 #include <asm/kexec.h>
59 #define DBG(fmt...) udbg_printf(fmt)
64 #ifdef CONFIG_HOTPLUG_CPU
65 /* State of each CPU during hotplug phases */
66 static DEFINE_PER_CPU(int, cpu_state
) = { 0 };
69 struct thread_info
*secondary_ti
;
71 DEFINE_PER_CPU(cpumask_var_t
, cpu_sibling_map
);
72 DEFINE_PER_CPU(cpumask_var_t
, cpu_core_map
);
74 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map
);
75 EXPORT_PER_CPU_SYMBOL(cpu_core_map
);
77 /* SMP operations for this machine */
78 struct smp_ops_t
*smp_ops
;
80 /* Can't be static due to PowerMac hackery */
81 volatile unsigned int cpu_callin_map
[NR_CPUS
];
83 int smt_enabled_at_boot
= 1;
85 static void (*crash_ipi_function_ptr
)(struct pt_regs
*) = NULL
;
88 * Returns 1 if the specified cpu should be brought up during boot.
89 * Used to inhibit booting threads if they've been disabled or
90 * limited on the command line
92 int smp_generic_cpu_bootable(unsigned int nr
)
94 /* Special case - we inhibit secondary thread startup
95 * during boot if the user requests it.
97 if (system_state
== SYSTEM_BOOTING
&& cpu_has_feature(CPU_FTR_SMT
)) {
98 if (!smt_enabled_at_boot
&& cpu_thread_in_core(nr
) != 0)
100 if (smt_enabled_at_boot
101 && cpu_thread_in_core(nr
) >= smt_enabled_at_boot
)
110 int smp_generic_kick_cpu(int nr
)
112 BUG_ON(nr
< 0 || nr
>= NR_CPUS
);
115 * The processor is currently spinning, waiting for the
116 * cpu_start field to become non-zero After we set cpu_start,
117 * the processor will continue on to secondary_start
119 if (!paca
[nr
].cpu_start
) {
120 paca
[nr
].cpu_start
= 1;
125 #ifdef CONFIG_HOTPLUG_CPU
127 * Ok it's not there, so it might be soft-unplugged, let's
128 * try to bring it back
130 generic_set_cpu_up(nr
);
132 smp_send_reschedule(nr
);
133 #endif /* CONFIG_HOTPLUG_CPU */
137 #endif /* CONFIG_PPC64 */
139 static irqreturn_t
call_function_action(int irq
, void *data
)
141 generic_smp_call_function_interrupt();
145 static irqreturn_t
reschedule_action(int irq
, void *data
)
151 static irqreturn_t
tick_broadcast_ipi_action(int irq
, void *data
)
153 tick_broadcast_ipi_handler();
157 static irqreturn_t
debug_ipi_action(int irq
, void *data
)
159 if (crash_ipi_function_ptr
) {
160 crash_ipi_function_ptr(get_irq_regs());
164 #ifdef CONFIG_DEBUGGER
165 debugger_ipi(get_irq_regs());
166 #endif /* CONFIG_DEBUGGER */
171 static irq_handler_t smp_ipi_action
[] = {
172 [PPC_MSG_CALL_FUNCTION
] = call_function_action
,
173 [PPC_MSG_RESCHEDULE
] = reschedule_action
,
174 [PPC_MSG_TICK_BROADCAST
] = tick_broadcast_ipi_action
,
175 [PPC_MSG_DEBUGGER_BREAK
] = debug_ipi_action
,
178 const char *smp_ipi_name
[] = {
179 [PPC_MSG_CALL_FUNCTION
] = "ipi call function",
180 [PPC_MSG_RESCHEDULE
] = "ipi reschedule",
181 [PPC_MSG_TICK_BROADCAST
] = "ipi tick-broadcast",
182 [PPC_MSG_DEBUGGER_BREAK
] = "ipi debugger",
185 /* optional function to request ipi, for controllers with >= 4 ipis */
186 int smp_request_message_ipi(int virq
, int msg
)
190 if (msg
< 0 || msg
> PPC_MSG_DEBUGGER_BREAK
) {
193 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
194 if (msg
== PPC_MSG_DEBUGGER_BREAK
) {
198 err
= request_irq(virq
, smp_ipi_action
[msg
],
199 IRQF_PERCPU
| IRQF_NO_THREAD
| IRQF_NO_SUSPEND
,
200 smp_ipi_name
[msg
], NULL
);
201 WARN(err
< 0, "unable to request_irq %d for %s (rc %d)\n",
202 virq
, smp_ipi_name
[msg
], err
);
207 #ifdef CONFIG_PPC_SMP_MUXED_IPI
208 struct cpu_messages
{
209 long messages
; /* current messages */
210 unsigned long data
; /* data for cause ipi */
212 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages
, ipi_message
);
214 void smp_muxed_ipi_set_data(int cpu
, unsigned long data
)
216 struct cpu_messages
*info
= &per_cpu(ipi_message
, cpu
);
221 void smp_muxed_ipi_set_message(int cpu
, int msg
)
223 struct cpu_messages
*info
= &per_cpu(ipi_message
, cpu
);
224 char *message
= (char *)&info
->messages
;
227 * Order previous accesses before accesses in the IPI handler.
233 void smp_muxed_ipi_message_pass(int cpu
, int msg
)
235 struct cpu_messages
*info
= &per_cpu(ipi_message
, cpu
);
237 smp_muxed_ipi_set_message(cpu
, msg
);
239 * cause_ipi functions are required to include a full barrier
240 * before doing whatever causes the IPI.
242 smp_ops
->cause_ipi(cpu
, info
->data
);
245 #ifdef __BIG_ENDIAN__
246 #define IPI_MESSAGE(A) (1uL << ((BITS_PER_LONG - 8) - 8 * (A)))
248 #define IPI_MESSAGE(A) (1uL << (8 * (A)))
251 irqreturn_t
smp_ipi_demux(void)
253 struct cpu_messages
*info
= this_cpu_ptr(&ipi_message
);
256 mb(); /* order any irq clear */
259 all
= xchg(&info
->messages
, 0);
260 #if defined(CONFIG_KVM_XICS) && defined(CONFIG_KVM_BOOK3S_HV_POSSIBLE)
262 * Must check for PPC_MSG_RM_HOST_ACTION messages
263 * before PPC_MSG_CALL_FUNCTION messages because when
264 * a VM is destroyed, we call kick_all_cpus_sync()
265 * to ensure that any pending PPC_MSG_RM_HOST_ACTION
266 * messages have completed before we free any VCPUs.
268 if (all
& IPI_MESSAGE(PPC_MSG_RM_HOST_ACTION
))
269 kvmppc_xics_ipi_action();
271 if (all
& IPI_MESSAGE(PPC_MSG_CALL_FUNCTION
))
272 generic_smp_call_function_interrupt();
273 if (all
& IPI_MESSAGE(PPC_MSG_RESCHEDULE
))
275 if (all
& IPI_MESSAGE(PPC_MSG_TICK_BROADCAST
))
276 tick_broadcast_ipi_handler();
277 if (all
& IPI_MESSAGE(PPC_MSG_DEBUGGER_BREAK
))
278 debug_ipi_action(0, NULL
);
279 } while (info
->messages
);
283 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
285 static inline void do_message_pass(int cpu
, int msg
)
287 if (smp_ops
->message_pass
)
288 smp_ops
->message_pass(cpu
, msg
);
289 #ifdef CONFIG_PPC_SMP_MUXED_IPI
291 smp_muxed_ipi_message_pass(cpu
, msg
);
295 void smp_send_reschedule(int cpu
)
298 do_message_pass(cpu
, PPC_MSG_RESCHEDULE
);
300 EXPORT_SYMBOL_GPL(smp_send_reschedule
);
302 void arch_send_call_function_single_ipi(int cpu
)
304 do_message_pass(cpu
, PPC_MSG_CALL_FUNCTION
);
307 void arch_send_call_function_ipi_mask(const struct cpumask
*mask
)
311 for_each_cpu(cpu
, mask
)
312 do_message_pass(cpu
, PPC_MSG_CALL_FUNCTION
);
315 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
316 void tick_broadcast(const struct cpumask
*mask
)
320 for_each_cpu(cpu
, mask
)
321 do_message_pass(cpu
, PPC_MSG_TICK_BROADCAST
);
325 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
326 void smp_send_debugger_break(void)
329 int me
= raw_smp_processor_id();
331 if (unlikely(!smp_ops
))
334 for_each_online_cpu(cpu
)
336 do_message_pass(cpu
, PPC_MSG_DEBUGGER_BREAK
);
341 void crash_send_ipi(void (*crash_ipi_callback
)(struct pt_regs
*))
343 crash_ipi_function_ptr
= crash_ipi_callback
;
344 if (crash_ipi_callback
) {
346 smp_send_debugger_break();
351 static void stop_this_cpu(void *dummy
)
353 /* Remove this CPU */
354 set_cpu_online(smp_processor_id(), false);
361 void smp_send_stop(void)
363 smp_call_function(stop_this_cpu
, NULL
, 0);
366 struct thread_info
*current_set
[NR_CPUS
];
368 static void smp_store_cpu_info(int id
)
370 per_cpu(cpu_pvr
, id
) = mfspr(SPRN_PVR
);
371 #ifdef CONFIG_PPC_FSL_BOOK3E
372 per_cpu(next_tlbcam_idx
, id
)
373 = (mfspr(SPRN_TLB1CFG
) & TLBnCFG_N_ENTRY
) - 1;
377 void __init
smp_prepare_cpus(unsigned int max_cpus
)
381 DBG("smp_prepare_cpus\n");
384 * setup_cpu may need to be called on the boot cpu. We havent
385 * spun any cpus up but lets be paranoid.
387 BUG_ON(boot_cpuid
!= smp_processor_id());
390 smp_store_cpu_info(boot_cpuid
);
391 cpu_callin_map
[boot_cpuid
] = 1;
393 for_each_possible_cpu(cpu
) {
394 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map
, cpu
),
395 GFP_KERNEL
, cpu_to_node(cpu
));
396 zalloc_cpumask_var_node(&per_cpu(cpu_core_map
, cpu
),
397 GFP_KERNEL
, cpu_to_node(cpu
));
399 * numa_node_id() works after this.
401 if (cpu_present(cpu
)) {
402 set_cpu_numa_node(cpu
, numa_cpu_lookup_table
[cpu
]);
403 set_cpu_numa_mem(cpu
,
404 local_memory_node(numa_cpu_lookup_table
[cpu
]));
408 cpumask_set_cpu(boot_cpuid
, cpu_sibling_mask(boot_cpuid
));
409 cpumask_set_cpu(boot_cpuid
, cpu_core_mask(boot_cpuid
));
411 if (smp_ops
&& smp_ops
->probe
)
415 void smp_prepare_boot_cpu(void)
417 BUG_ON(smp_processor_id() != boot_cpuid
);
419 paca
[boot_cpuid
].__current
= current
;
421 set_numa_node(numa_cpu_lookup_table
[boot_cpuid
]);
422 current_set
[boot_cpuid
] = task_thread_info(current
);
425 #ifdef CONFIG_HOTPLUG_CPU
427 int generic_cpu_disable(void)
429 unsigned int cpu
= smp_processor_id();
431 if (cpu
== boot_cpuid
)
434 set_cpu_online(cpu
, false);
436 vdso_data
->processorCount
--;
442 void generic_cpu_die(unsigned int cpu
)
446 for (i
= 0; i
< 100; i
++) {
448 if (is_cpu_dead(cpu
))
452 printk(KERN_ERR
"CPU%d didn't die...\n", cpu
);
455 void generic_set_cpu_dead(unsigned int cpu
)
457 per_cpu(cpu_state
, cpu
) = CPU_DEAD
;
461 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
462 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
463 * which makes the delay in generic_cpu_die() not happen.
465 void generic_set_cpu_up(unsigned int cpu
)
467 per_cpu(cpu_state
, cpu
) = CPU_UP_PREPARE
;
470 int generic_check_cpu_restart(unsigned int cpu
)
472 return per_cpu(cpu_state
, cpu
) == CPU_UP_PREPARE
;
475 int is_cpu_dead(unsigned int cpu
)
477 return per_cpu(cpu_state
, cpu
) == CPU_DEAD
;
480 static bool secondaries_inhibited(void)
482 return kvm_hv_mode_active();
485 #else /* HOTPLUG_CPU */
487 #define secondaries_inhibited() 0
491 static void cpu_idle_thread_init(unsigned int cpu
, struct task_struct
*idle
)
493 struct thread_info
*ti
= task_thread_info(idle
);
496 paca
[cpu
].__current
= idle
;
497 paca
[cpu
].kstack
= (unsigned long)ti
+ THREAD_SIZE
- STACK_FRAME_OVERHEAD
;
500 secondary_ti
= current_set
[cpu
] = ti
;
503 int __cpu_up(unsigned int cpu
, struct task_struct
*tidle
)
508 * Don't allow secondary threads to come online if inhibited
510 if (threads_per_core
> 1 && secondaries_inhibited() &&
511 cpu_thread_in_subcore(cpu
))
514 if (smp_ops
== NULL
||
515 (smp_ops
->cpu_bootable
&& !smp_ops
->cpu_bootable(cpu
)))
518 cpu_idle_thread_init(cpu
, tidle
);
520 /* Make sure callin-map entry is 0 (can be leftover a CPU
523 cpu_callin_map
[cpu
] = 0;
525 /* The information for processor bringup must
526 * be written out to main store before we release
532 DBG("smp: kicking cpu %d\n", cpu
);
533 rc
= smp_ops
->kick_cpu(cpu
);
535 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu
, rc
);
540 * wait to see if the cpu made a callin (is actually up).
541 * use this value that I found through experimentation.
544 if (system_state
< SYSTEM_RUNNING
)
545 for (c
= 50000; c
&& !cpu_callin_map
[cpu
]; c
--)
547 #ifdef CONFIG_HOTPLUG_CPU
550 * CPUs can take much longer to come up in the
551 * hotplug case. Wait five seconds.
553 for (c
= 5000; c
&& !cpu_callin_map
[cpu
]; c
--)
557 if (!cpu_callin_map
[cpu
]) {
558 printk(KERN_ERR
"Processor %u is stuck.\n", cpu
);
562 DBG("Processor %u found.\n", cpu
);
564 if (smp_ops
->give_timebase
)
565 smp_ops
->give_timebase();
567 /* Wait until cpu puts itself in the online & active maps */
568 while (!cpu_online(cpu
) || !cpu_active(cpu
))
574 /* Return the value of the reg property corresponding to the given
577 int cpu_to_core_id(int cpu
)
579 struct device_node
*np
;
583 np
= of_get_cpu_node(cpu
, NULL
);
587 reg
= of_get_property(np
, "reg", NULL
);
591 id
= be32_to_cpup(reg
);
597 /* Helper routines for cpu to core mapping */
598 int cpu_core_index_of_thread(int cpu
)
600 return cpu
>> threads_shift
;
602 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread
);
604 int cpu_first_thread_of_core(int core
)
606 return core
<< threads_shift
;
608 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core
);
610 static void traverse_siblings_chip_id(int cpu
, bool add
, int chipid
)
612 const struct cpumask
*mask
;
613 struct device_node
*np
;
617 mask
= add
? cpu_online_mask
: cpu_present_mask
;
618 for_each_cpu(i
, mask
) {
619 np
= of_get_cpu_node(i
, NULL
);
622 prop
= of_get_property(np
, "ibm,chip-id", &plen
);
623 if (prop
&& plen
== sizeof(int) &&
624 of_read_number(prop
, 1) == chipid
) {
626 cpumask_set_cpu(cpu
, cpu_core_mask(i
));
627 cpumask_set_cpu(i
, cpu_core_mask(cpu
));
629 cpumask_clear_cpu(cpu
, cpu_core_mask(i
));
630 cpumask_clear_cpu(i
, cpu_core_mask(cpu
));
637 /* Must be called when no change can occur to cpu_present_mask,
638 * i.e. during cpu online or offline.
640 static struct device_node
*cpu_to_l2cache(int cpu
)
642 struct device_node
*np
;
643 struct device_node
*cache
;
645 if (!cpu_present(cpu
))
648 np
= of_get_cpu_node(cpu
, NULL
);
652 cache
= of_find_next_cache_node(np
);
659 static void traverse_core_siblings(int cpu
, bool add
)
661 struct device_node
*l2_cache
, *np
;
662 const struct cpumask
*mask
;
666 /* First see if we have ibm,chip-id properties in cpu nodes */
667 np
= of_get_cpu_node(cpu
, NULL
);
670 prop
= of_get_property(np
, "ibm,chip-id", &plen
);
671 if (prop
&& plen
== sizeof(int))
672 chip
= of_read_number(prop
, 1);
675 traverse_siblings_chip_id(cpu
, add
, chip
);
680 l2_cache
= cpu_to_l2cache(cpu
);
681 mask
= add
? cpu_online_mask
: cpu_present_mask
;
682 for_each_cpu(i
, mask
) {
683 np
= cpu_to_l2cache(i
);
686 if (np
== l2_cache
) {
688 cpumask_set_cpu(cpu
, cpu_core_mask(i
));
689 cpumask_set_cpu(i
, cpu_core_mask(cpu
));
691 cpumask_clear_cpu(cpu
, cpu_core_mask(i
));
692 cpumask_clear_cpu(i
, cpu_core_mask(cpu
));
697 of_node_put(l2_cache
);
700 /* Activate a secondary processor. */
701 void start_secondary(void *unused
)
703 unsigned int cpu
= smp_processor_id();
706 atomic_inc(&init_mm
.mm_count
);
707 current
->active_mm
= &init_mm
;
709 smp_store_cpu_info(cpu
);
710 set_dec(tb_ticks_per_jiffy
);
712 cpu_callin_map
[cpu
] = 1;
714 if (smp_ops
->setup_cpu
)
715 smp_ops
->setup_cpu(cpu
);
716 if (smp_ops
->take_timebase
)
717 smp_ops
->take_timebase();
719 secondary_cpu_time_init();
722 if (system_state
== SYSTEM_RUNNING
)
723 vdso_data
->processorCount
++;
727 /* Update sibling maps */
728 base
= cpu_first_thread_sibling(cpu
);
729 for (i
= 0; i
< threads_per_core
; i
++) {
730 if (cpu_is_offline(base
+ i
) && (cpu
!= base
+ i
))
732 cpumask_set_cpu(cpu
, cpu_sibling_mask(base
+ i
));
733 cpumask_set_cpu(base
+ i
, cpu_sibling_mask(cpu
));
735 /* cpu_core_map should be a superset of
736 * cpu_sibling_map even if we don't have cache
737 * information, so update the former here, too.
739 cpumask_set_cpu(cpu
, cpu_core_mask(base
+ i
));
740 cpumask_set_cpu(base
+ i
, cpu_core_mask(cpu
));
742 traverse_core_siblings(cpu
, true);
744 set_numa_node(numa_cpu_lookup_table
[cpu
]);
745 set_numa_mem(local_memory_node(numa_cpu_lookup_table
[cpu
]));
748 notify_cpu_starting(cpu
);
749 set_cpu_online(cpu
, true);
753 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE
);
758 int setup_profiling_timer(unsigned int multiplier
)
763 #ifdef CONFIG_SCHED_SMT
764 /* cpumask of CPUs with asymetric SMT dependancy */
765 static int powerpc_smt_flags(void)
767 int flags
= SD_SHARE_CPUCAPACITY
| SD_SHARE_PKG_RESOURCES
;
769 if (cpu_has_feature(CPU_FTR_ASYM_SMT
)) {
770 printk_once(KERN_INFO
"Enabling Asymmetric SMT scheduling\n");
771 flags
|= SD_ASYM_PACKING
;
777 static struct sched_domain_topology_level powerpc_topology
[] = {
778 #ifdef CONFIG_SCHED_SMT
779 { cpu_smt_mask
, powerpc_smt_flags
, SD_INIT_NAME(SMT
) },
781 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
785 void __init
smp_cpus_done(unsigned int max_cpus
)
787 cpumask_var_t old_mask
;
789 /* We want the setup_cpu() here to be called from CPU 0, but our
790 * init thread may have been "borrowed" by another CPU in the meantime
791 * se we pin us down to CPU 0 for a short while
793 alloc_cpumask_var(&old_mask
, GFP_NOWAIT
);
794 cpumask_copy(old_mask
, tsk_cpus_allowed(current
));
795 set_cpus_allowed_ptr(current
, cpumask_of(boot_cpuid
));
797 if (smp_ops
&& smp_ops
->setup_cpu
)
798 smp_ops
->setup_cpu(boot_cpuid
);
800 set_cpus_allowed_ptr(current
, old_mask
);
802 free_cpumask_var(old_mask
);
804 if (smp_ops
&& smp_ops
->bringup_done
)
805 smp_ops
->bringup_done();
807 dump_numa_cpu_topology();
809 set_sched_topology(powerpc_topology
);
813 #ifdef CONFIG_HOTPLUG_CPU
814 int __cpu_disable(void)
816 int cpu
= smp_processor_id();
820 if (!smp_ops
->cpu_disable
)
823 err
= smp_ops
->cpu_disable();
827 /* Update sibling maps */
828 base
= cpu_first_thread_sibling(cpu
);
829 for (i
= 0; i
< threads_per_core
; i
++) {
830 cpumask_clear_cpu(cpu
, cpu_sibling_mask(base
+ i
));
831 cpumask_clear_cpu(base
+ i
, cpu_sibling_mask(cpu
));
832 cpumask_clear_cpu(cpu
, cpu_core_mask(base
+ i
));
833 cpumask_clear_cpu(base
+ i
, cpu_core_mask(cpu
));
835 traverse_core_siblings(cpu
, false);
840 void __cpu_die(unsigned int cpu
)
842 if (smp_ops
->cpu_die
)
843 smp_ops
->cpu_die(cpu
);
851 /* If we return, we re-enter start_secondary */
852 start_secondary_resume();