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>
39 #include <asm/pgtable.h>
43 #include <asm/machdep.h>
44 #include <asm/cputhreads.h>
45 #include <asm/cputable.h>
47 #include <asm/vdso_datapage.h>
52 #include <asm/debug.h>
56 #define DBG(fmt...) udbg_printf(fmt)
61 #ifdef CONFIG_HOTPLUG_CPU
62 /* State of each CPU during hotplug phases */
63 static DEFINE_PER_CPU(int, cpu_state
) = { 0 };
66 struct thread_info
*secondary_ti
;
68 DEFINE_PER_CPU(cpumask_var_t
, cpu_sibling_map
);
69 DEFINE_PER_CPU(cpumask_var_t
, cpu_core_map
);
71 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map
);
72 EXPORT_PER_CPU_SYMBOL(cpu_core_map
);
74 /* SMP operations for this machine */
75 struct smp_ops_t
*smp_ops
;
77 /* Can't be static due to PowerMac hackery */
78 volatile unsigned int cpu_callin_map
[NR_CPUS
];
80 int smt_enabled_at_boot
= 1;
82 static void (*crash_ipi_function_ptr
)(struct pt_regs
*) = NULL
;
85 * Returns 1 if the specified cpu should be brought up during boot.
86 * Used to inhibit booting threads if they've been disabled or
87 * limited on the command line
89 int smp_generic_cpu_bootable(unsigned int nr
)
91 /* Special case - we inhibit secondary thread startup
92 * during boot if the user requests it.
94 if (system_state
== SYSTEM_BOOTING
&& cpu_has_feature(CPU_FTR_SMT
)) {
95 if (!smt_enabled_at_boot
&& cpu_thread_in_core(nr
) != 0)
97 if (smt_enabled_at_boot
98 && cpu_thread_in_core(nr
) >= smt_enabled_at_boot
)
107 int smp_generic_kick_cpu(int nr
)
109 BUG_ON(nr
< 0 || nr
>= NR_CPUS
);
112 * The processor is currently spinning, waiting for the
113 * cpu_start field to become non-zero After we set cpu_start,
114 * the processor will continue on to secondary_start
116 if (!paca
[nr
].cpu_start
) {
117 paca
[nr
].cpu_start
= 1;
122 #ifdef CONFIG_HOTPLUG_CPU
124 * Ok it's not there, so it might be soft-unplugged, let's
125 * try to bring it back
127 generic_set_cpu_up(nr
);
129 smp_send_reschedule(nr
);
130 #endif /* CONFIG_HOTPLUG_CPU */
134 #endif /* CONFIG_PPC64 */
136 static irqreturn_t
call_function_action(int irq
, void *data
)
138 generic_smp_call_function_interrupt();
142 static irqreturn_t
reschedule_action(int irq
, void *data
)
148 static irqreturn_t
call_function_single_action(int irq
, void *data
)
150 generic_smp_call_function_single_interrupt();
154 static irqreturn_t
debug_ipi_action(int irq
, void *data
)
156 if (crash_ipi_function_ptr
) {
157 crash_ipi_function_ptr(get_irq_regs());
161 #ifdef CONFIG_DEBUGGER
162 debugger_ipi(get_irq_regs());
163 #endif /* CONFIG_DEBUGGER */
168 static irq_handler_t smp_ipi_action
[] = {
169 [PPC_MSG_CALL_FUNCTION
] = call_function_action
,
170 [PPC_MSG_RESCHEDULE
] = reschedule_action
,
171 [PPC_MSG_CALL_FUNC_SINGLE
] = call_function_single_action
,
172 [PPC_MSG_DEBUGGER_BREAK
] = debug_ipi_action
,
175 const char *smp_ipi_name
[] = {
176 [PPC_MSG_CALL_FUNCTION
] = "ipi call function",
177 [PPC_MSG_RESCHEDULE
] = "ipi reschedule",
178 [PPC_MSG_CALL_FUNC_SINGLE
] = "ipi call function single",
179 [PPC_MSG_DEBUGGER_BREAK
] = "ipi debugger",
182 /* optional function to request ipi, for controllers with >= 4 ipis */
183 int smp_request_message_ipi(int virq
, int msg
)
187 if (msg
< 0 || msg
> PPC_MSG_DEBUGGER_BREAK
) {
190 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
191 if (msg
== PPC_MSG_DEBUGGER_BREAK
) {
195 err
= request_irq(virq
, smp_ipi_action
[msg
],
196 IRQF_PERCPU
| IRQF_NO_THREAD
| IRQF_NO_SUSPEND
,
197 smp_ipi_name
[msg
], NULL
);
198 WARN(err
< 0, "unable to request_irq %d for %s (rc %d)\n",
199 virq
, smp_ipi_name
[msg
], err
);
204 #ifdef CONFIG_PPC_SMP_MUXED_IPI
205 struct cpu_messages
{
206 int messages
; /* current messages */
207 unsigned long data
; /* data for cause ipi */
209 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages
, ipi_message
);
211 void smp_muxed_ipi_set_data(int cpu
, unsigned long data
)
213 struct cpu_messages
*info
= &per_cpu(ipi_message
, cpu
);
218 void smp_muxed_ipi_message_pass(int cpu
, int msg
)
220 struct cpu_messages
*info
= &per_cpu(ipi_message
, cpu
);
221 char *message
= (char *)&info
->messages
;
224 * Order previous accesses before accesses in the IPI handler.
229 * cause_ipi functions are required to include a full barrier
230 * before doing whatever causes the IPI.
232 smp_ops
->cause_ipi(cpu
, info
->data
);
235 #ifdef __BIG_ENDIAN__
236 #define IPI_MESSAGE(A) (1 << (24 - 8 * (A)))
238 #define IPI_MESSAGE(A) (1 << (8 * (A)))
241 irqreturn_t
smp_ipi_demux(void)
243 struct cpu_messages
*info
= &__get_cpu_var(ipi_message
);
246 mb(); /* order any irq clear */
249 all
= xchg(&info
->messages
, 0);
250 if (all
& IPI_MESSAGE(PPC_MSG_CALL_FUNCTION
))
251 generic_smp_call_function_interrupt();
252 if (all
& IPI_MESSAGE(PPC_MSG_RESCHEDULE
))
254 if (all
& IPI_MESSAGE(PPC_MSG_CALL_FUNC_SINGLE
))
255 generic_smp_call_function_single_interrupt();
256 if (all
& IPI_MESSAGE(PPC_MSG_DEBUGGER_BREAK
))
257 debug_ipi_action(0, NULL
);
258 } while (info
->messages
);
262 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
264 static inline void do_message_pass(int cpu
, int msg
)
266 if (smp_ops
->message_pass
)
267 smp_ops
->message_pass(cpu
, msg
);
268 #ifdef CONFIG_PPC_SMP_MUXED_IPI
270 smp_muxed_ipi_message_pass(cpu
, msg
);
274 void smp_send_reschedule(int cpu
)
277 do_message_pass(cpu
, PPC_MSG_RESCHEDULE
);
279 EXPORT_SYMBOL_GPL(smp_send_reschedule
);
281 void arch_send_call_function_single_ipi(int cpu
)
283 do_message_pass(cpu
, PPC_MSG_CALL_FUNC_SINGLE
);
286 void arch_send_call_function_ipi_mask(const struct cpumask
*mask
)
290 for_each_cpu(cpu
, mask
)
291 do_message_pass(cpu
, PPC_MSG_CALL_FUNCTION
);
294 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
295 void smp_send_debugger_break(void)
298 int me
= raw_smp_processor_id();
300 if (unlikely(!smp_ops
))
303 for_each_online_cpu(cpu
)
305 do_message_pass(cpu
, PPC_MSG_DEBUGGER_BREAK
);
310 void crash_send_ipi(void (*crash_ipi_callback
)(struct pt_regs
*))
312 crash_ipi_function_ptr
= crash_ipi_callback
;
313 if (crash_ipi_callback
) {
315 smp_send_debugger_break();
320 static void stop_this_cpu(void *dummy
)
322 /* Remove this CPU */
323 set_cpu_online(smp_processor_id(), false);
330 void smp_send_stop(void)
332 smp_call_function(stop_this_cpu
, NULL
, 0);
335 struct thread_info
*current_set
[NR_CPUS
];
337 static void smp_store_cpu_info(int id
)
339 per_cpu(cpu_pvr
, id
) = mfspr(SPRN_PVR
);
340 #ifdef CONFIG_PPC_FSL_BOOK3E
341 per_cpu(next_tlbcam_idx
, id
)
342 = (mfspr(SPRN_TLB1CFG
) & TLBnCFG_N_ENTRY
) - 1;
346 void __init
smp_prepare_cpus(unsigned int max_cpus
)
350 DBG("smp_prepare_cpus\n");
353 * setup_cpu may need to be called on the boot cpu. We havent
354 * spun any cpus up but lets be paranoid.
356 BUG_ON(boot_cpuid
!= smp_processor_id());
359 smp_store_cpu_info(boot_cpuid
);
360 cpu_callin_map
[boot_cpuid
] = 1;
362 for_each_possible_cpu(cpu
) {
363 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map
, cpu
),
364 GFP_KERNEL
, cpu_to_node(cpu
));
365 zalloc_cpumask_var_node(&per_cpu(cpu_core_map
, cpu
),
366 GFP_KERNEL
, cpu_to_node(cpu
));
369 cpumask_set_cpu(boot_cpuid
, cpu_sibling_mask(boot_cpuid
));
370 cpumask_set_cpu(boot_cpuid
, cpu_core_mask(boot_cpuid
));
374 max_cpus
= smp_ops
->probe();
381 void smp_prepare_boot_cpu(void)
383 BUG_ON(smp_processor_id() != boot_cpuid
);
385 paca
[boot_cpuid
].__current
= current
;
387 current_set
[boot_cpuid
] = task_thread_info(current
);
390 #ifdef CONFIG_HOTPLUG_CPU
392 int generic_cpu_disable(void)
394 unsigned int cpu
= smp_processor_id();
396 if (cpu
== boot_cpuid
)
399 set_cpu_online(cpu
, false);
401 vdso_data
->processorCount
--;
407 void generic_cpu_die(unsigned int cpu
)
411 for (i
= 0; i
< 100; i
++) {
413 if (per_cpu(cpu_state
, cpu
) == CPU_DEAD
)
417 printk(KERN_ERR
"CPU%d didn't die...\n", cpu
);
420 void generic_mach_cpu_die(void)
426 cpu
= smp_processor_id();
427 printk(KERN_DEBUG
"CPU%d offline\n", cpu
);
428 __get_cpu_var(cpu_state
) = CPU_DEAD
;
430 while (__get_cpu_var(cpu_state
) != CPU_UP_PREPARE
)
434 void generic_set_cpu_dead(unsigned int cpu
)
436 per_cpu(cpu_state
, cpu
) = CPU_DEAD
;
440 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
441 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
442 * which makes the delay in generic_cpu_die() not happen.
444 void generic_set_cpu_up(unsigned int cpu
)
446 per_cpu(cpu_state
, cpu
) = CPU_UP_PREPARE
;
449 int generic_check_cpu_restart(unsigned int cpu
)
451 return per_cpu(cpu_state
, cpu
) == CPU_UP_PREPARE
;
454 static atomic_t secondary_inhibit_count
;
457 * Don't allow secondary CPU threads to come online
459 void inhibit_secondary_onlining(void)
462 * This makes secondary_inhibit_count stable during cpu
463 * online/offline operations.
467 atomic_inc(&secondary_inhibit_count
);
470 EXPORT_SYMBOL_GPL(inhibit_secondary_onlining
);
473 * Allow secondary CPU threads to come online again
475 void uninhibit_secondary_onlining(void)
478 atomic_dec(&secondary_inhibit_count
);
481 EXPORT_SYMBOL_GPL(uninhibit_secondary_onlining
);
483 static int secondaries_inhibited(void)
485 return atomic_read(&secondary_inhibit_count
);
488 #else /* HOTPLUG_CPU */
490 #define secondaries_inhibited() 0
494 static void cpu_idle_thread_init(unsigned int cpu
, struct task_struct
*idle
)
496 struct thread_info
*ti
= task_thread_info(idle
);
499 paca
[cpu
].__current
= idle
;
500 paca
[cpu
].kstack
= (unsigned long)ti
+ THREAD_SIZE
- STACK_FRAME_OVERHEAD
;
503 secondary_ti
= current_set
[cpu
] = ti
;
506 int __cpu_up(unsigned int cpu
, struct task_struct
*tidle
)
511 * Don't allow secondary threads to come online if inhibited
513 if (threads_per_core
> 1 && secondaries_inhibited() &&
514 cpu
% threads_per_core
!= 0)
517 if (smp_ops
== NULL
||
518 (smp_ops
->cpu_bootable
&& !smp_ops
->cpu_bootable(cpu
)))
521 cpu_idle_thread_init(cpu
, tidle
);
523 /* Make sure callin-map entry is 0 (can be leftover a CPU
526 cpu_callin_map
[cpu
] = 0;
528 /* The information for processor bringup must
529 * be written out to main store before we release
535 DBG("smp: kicking cpu %d\n", cpu
);
536 rc
= smp_ops
->kick_cpu(cpu
);
538 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu
, rc
);
543 * wait to see if the cpu made a callin (is actually up).
544 * use this value that I found through experimentation.
547 if (system_state
< SYSTEM_RUNNING
)
548 for (c
= 50000; c
&& !cpu_callin_map
[cpu
]; c
--)
550 #ifdef CONFIG_HOTPLUG_CPU
553 * CPUs can take much longer to come up in the
554 * hotplug case. Wait five seconds.
556 for (c
= 5000; c
&& !cpu_callin_map
[cpu
]; c
--)
560 if (!cpu_callin_map
[cpu
]) {
561 printk(KERN_ERR
"Processor %u is stuck.\n", cpu
);
565 DBG("Processor %u found.\n", cpu
);
567 if (smp_ops
->give_timebase
)
568 smp_ops
->give_timebase();
570 /* Wait until cpu puts itself in the online map */
571 while (!cpu_online(cpu
))
577 /* Return the value of the reg property corresponding to the given
580 int cpu_to_core_id(int cpu
)
582 struct device_node
*np
;
586 np
= of_get_cpu_node(cpu
, NULL
);
590 reg
= of_get_property(np
, "reg", NULL
);
600 /* Return the value of the chip-id property corresponding
601 * to the given logical cpu.
603 int cpu_to_chip_id(int cpu
)
605 struct device_node
*np
;
607 np
= of_get_cpu_node(cpu
, NULL
);
612 return of_get_ibm_chip_id(np
);
614 EXPORT_SYMBOL(cpu_to_chip_id
);
616 /* Helper routines for cpu to core mapping */
617 int cpu_core_index_of_thread(int cpu
)
619 return cpu
>> threads_shift
;
621 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread
);
623 int cpu_first_thread_of_core(int core
)
625 return core
<< threads_shift
;
627 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core
);
629 static void traverse_siblings_chip_id(int cpu
, bool add
, int chipid
)
631 const struct cpumask
*mask
;
632 struct device_node
*np
;
636 mask
= add
? cpu_online_mask
: cpu_present_mask
;
637 for_each_cpu(i
, mask
) {
638 np
= of_get_cpu_node(i
, NULL
);
641 prop
= of_get_property(np
, "ibm,chip-id", &plen
);
642 if (prop
&& plen
== sizeof(int) &&
643 of_read_number(prop
, 1) == chipid
) {
645 cpumask_set_cpu(cpu
, cpu_core_mask(i
));
646 cpumask_set_cpu(i
, cpu_core_mask(cpu
));
648 cpumask_clear_cpu(cpu
, cpu_core_mask(i
));
649 cpumask_clear_cpu(i
, cpu_core_mask(cpu
));
656 /* Must be called when no change can occur to cpu_present_mask,
657 * i.e. during cpu online or offline.
659 static struct device_node
*cpu_to_l2cache(int cpu
)
661 struct device_node
*np
;
662 struct device_node
*cache
;
664 if (!cpu_present(cpu
))
667 np
= of_get_cpu_node(cpu
, NULL
);
671 cache
= of_find_next_cache_node(np
);
678 static void traverse_core_siblings(int cpu
, bool add
)
680 struct device_node
*l2_cache
, *np
;
681 const struct cpumask
*mask
;
685 /* First see if we have ibm,chip-id properties in cpu nodes */
686 np
= of_get_cpu_node(cpu
, NULL
);
689 prop
= of_get_property(np
, "ibm,chip-id", &plen
);
690 if (prop
&& plen
== sizeof(int))
691 chip
= of_read_number(prop
, 1);
694 traverse_siblings_chip_id(cpu
, add
, chip
);
699 l2_cache
= cpu_to_l2cache(cpu
);
700 mask
= add
? cpu_online_mask
: cpu_present_mask
;
701 for_each_cpu(i
, mask
) {
702 np
= cpu_to_l2cache(i
);
705 if (np
== l2_cache
) {
707 cpumask_set_cpu(cpu
, cpu_core_mask(i
));
708 cpumask_set_cpu(i
, cpu_core_mask(cpu
));
710 cpumask_clear_cpu(cpu
, cpu_core_mask(i
));
711 cpumask_clear_cpu(i
, cpu_core_mask(cpu
));
716 of_node_put(l2_cache
);
719 /* Activate a secondary processor. */
720 void start_secondary(void *unused
)
722 unsigned int cpu
= smp_processor_id();
725 atomic_inc(&init_mm
.mm_count
);
726 current
->active_mm
= &init_mm
;
728 smp_store_cpu_info(cpu
);
729 set_dec(tb_ticks_per_jiffy
);
731 cpu_callin_map
[cpu
] = 1;
733 if (smp_ops
->setup_cpu
)
734 smp_ops
->setup_cpu(cpu
);
735 if (smp_ops
->take_timebase
)
736 smp_ops
->take_timebase();
738 secondary_cpu_time_init();
741 if (system_state
== SYSTEM_RUNNING
)
742 vdso_data
->processorCount
++;
746 /* Update sibling maps */
747 base
= cpu_first_thread_sibling(cpu
);
748 for (i
= 0; i
< threads_per_core
; i
++) {
749 if (cpu_is_offline(base
+ i
) && (cpu
!= base
+ i
))
751 cpumask_set_cpu(cpu
, cpu_sibling_mask(base
+ i
));
752 cpumask_set_cpu(base
+ i
, cpu_sibling_mask(cpu
));
754 /* cpu_core_map should be a superset of
755 * cpu_sibling_map even if we don't have cache
756 * information, so update the former here, too.
758 cpumask_set_cpu(cpu
, cpu_core_mask(base
+ i
));
759 cpumask_set_cpu(base
+ i
, cpu_core_mask(cpu
));
761 traverse_core_siblings(cpu
, true);
764 notify_cpu_starting(cpu
);
765 set_cpu_online(cpu
, true);
769 cpu_startup_entry(CPUHP_ONLINE
);
774 int setup_profiling_timer(unsigned int multiplier
)
779 void __init
smp_cpus_done(unsigned int max_cpus
)
781 cpumask_var_t old_mask
;
783 /* We want the setup_cpu() here to be called from CPU 0, but our
784 * init thread may have been "borrowed" by another CPU in the meantime
785 * se we pin us down to CPU 0 for a short while
787 alloc_cpumask_var(&old_mask
, GFP_NOWAIT
);
788 cpumask_copy(old_mask
, tsk_cpus_allowed(current
));
789 set_cpus_allowed_ptr(current
, cpumask_of(boot_cpuid
));
791 if (smp_ops
&& smp_ops
->setup_cpu
)
792 smp_ops
->setup_cpu(boot_cpuid
);
794 set_cpus_allowed_ptr(current
, old_mask
);
796 free_cpumask_var(old_mask
);
798 if (smp_ops
&& smp_ops
->bringup_done
)
799 smp_ops
->bringup_done();
801 dump_numa_cpu_topology();
805 int arch_sd_sibling_asym_packing(void)
807 if (cpu_has_feature(CPU_FTR_ASYM_SMT
)) {
808 printk_once(KERN_INFO
"Enabling Asymmetric SMT scheduling\n");
809 return SD_ASYM_PACKING
;
814 #ifdef CONFIG_HOTPLUG_CPU
815 int __cpu_disable(void)
817 int cpu
= smp_processor_id();
821 if (!smp_ops
->cpu_disable
)
824 err
= smp_ops
->cpu_disable();
828 /* Update sibling maps */
829 base
= cpu_first_thread_sibling(cpu
);
830 for (i
= 0; i
< threads_per_core
; i
++) {
831 cpumask_clear_cpu(cpu
, cpu_sibling_mask(base
+ i
));
832 cpumask_clear_cpu(base
+ i
, cpu_sibling_mask(cpu
));
833 cpumask_clear_cpu(cpu
, cpu_core_mask(base
+ i
));
834 cpumask_clear_cpu(base
+ i
, cpu_core_mask(cpu
));
836 traverse_core_siblings(cpu
, false);
841 void __cpu_die(unsigned int cpu
)
843 if (smp_ops
->cpu_die
)
844 smp_ops
->cpu_die(cpu
);
847 static DEFINE_MUTEX(powerpc_cpu_hotplug_driver_mutex
);
849 void cpu_hotplug_driver_lock()
851 mutex_lock(&powerpc_cpu_hotplug_driver_mutex
);
854 void cpu_hotplug_driver_unlock()
856 mutex_unlock(&powerpc_cpu_hotplug_driver_mutex
);
864 /* If we return, we re-enter start_secondary */
865 start_secondary_resume();