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>
40 #include <asm/pgtable.h>
44 #include <asm/machdep.h>
45 #include <asm/cputhreads.h>
46 #include <asm/cputable.h>
48 #include <asm/vdso_datapage.h>
53 #include <asm/debug.h>
57 #define DBG(fmt...) udbg_printf(fmt)
62 #ifdef CONFIG_HOTPLUG_CPU
63 /* State of each CPU during hotplug phases */
64 static DEFINE_PER_CPU(int, cpu_state
) = { 0 };
67 struct thread_info
*secondary_ti
;
69 DEFINE_PER_CPU(cpumask_var_t
, cpu_sibling_map
);
70 DEFINE_PER_CPU(cpumask_var_t
, cpu_core_map
);
72 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map
);
73 EXPORT_PER_CPU_SYMBOL(cpu_core_map
);
75 /* SMP operations for this machine */
76 struct smp_ops_t
*smp_ops
;
78 /* Can't be static due to PowerMac hackery */
79 volatile unsigned int cpu_callin_map
[NR_CPUS
];
81 int smt_enabled_at_boot
= 1;
83 static void (*crash_ipi_function_ptr
)(struct pt_regs
*) = NULL
;
86 * Returns 1 if the specified cpu should be brought up during boot.
87 * Used to inhibit booting threads if they've been disabled or
88 * limited on the command line
90 int smp_generic_cpu_bootable(unsigned int nr
)
92 /* Special case - we inhibit secondary thread startup
93 * during boot if the user requests it.
95 if (system_state
== SYSTEM_BOOTING
&& cpu_has_feature(CPU_FTR_SMT
)) {
96 if (!smt_enabled_at_boot
&& cpu_thread_in_core(nr
) != 0)
98 if (smt_enabled_at_boot
99 && cpu_thread_in_core(nr
) >= smt_enabled_at_boot
)
108 int smp_generic_kick_cpu(int nr
)
110 BUG_ON(nr
< 0 || nr
>= NR_CPUS
);
113 * The processor is currently spinning, waiting for the
114 * cpu_start field to become non-zero After we set cpu_start,
115 * the processor will continue on to secondary_start
117 if (!paca
[nr
].cpu_start
) {
118 paca
[nr
].cpu_start
= 1;
123 #ifdef CONFIG_HOTPLUG_CPU
125 * Ok it's not there, so it might be soft-unplugged, let's
126 * try to bring it back
128 generic_set_cpu_up(nr
);
130 smp_send_reschedule(nr
);
131 #endif /* CONFIG_HOTPLUG_CPU */
135 #endif /* CONFIG_PPC64 */
137 static irqreturn_t
call_function_action(int irq
, void *data
)
139 generic_smp_call_function_interrupt();
143 static irqreturn_t
reschedule_action(int irq
, void *data
)
149 static irqreturn_t
tick_broadcast_ipi_action(int irq
, void *data
)
151 tick_broadcast_ipi_handler();
155 static irqreturn_t
debug_ipi_action(int irq
, void *data
)
157 if (crash_ipi_function_ptr
) {
158 crash_ipi_function_ptr(get_irq_regs());
162 #ifdef CONFIG_DEBUGGER
163 debugger_ipi(get_irq_regs());
164 #endif /* CONFIG_DEBUGGER */
169 static irq_handler_t smp_ipi_action
[] = {
170 [PPC_MSG_CALL_FUNCTION
] = call_function_action
,
171 [PPC_MSG_RESCHEDULE
] = reschedule_action
,
172 [PPC_MSG_TICK_BROADCAST
] = tick_broadcast_ipi_action
,
173 [PPC_MSG_DEBUGGER_BREAK
] = debug_ipi_action
,
176 const char *smp_ipi_name
[] = {
177 [PPC_MSG_CALL_FUNCTION
] = "ipi call function",
178 [PPC_MSG_RESCHEDULE
] = "ipi reschedule",
179 [PPC_MSG_TICK_BROADCAST
] = "ipi tick-broadcast",
180 [PPC_MSG_DEBUGGER_BREAK
] = "ipi debugger",
183 /* optional function to request ipi, for controllers with >= 4 ipis */
184 int smp_request_message_ipi(int virq
, int msg
)
188 if (msg
< 0 || msg
> PPC_MSG_DEBUGGER_BREAK
) {
191 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
192 if (msg
== PPC_MSG_DEBUGGER_BREAK
) {
196 err
= request_irq(virq
, smp_ipi_action
[msg
],
197 IRQF_PERCPU
| IRQF_NO_THREAD
| IRQF_NO_SUSPEND
,
198 smp_ipi_name
[msg
], NULL
);
199 WARN(err
< 0, "unable to request_irq %d for %s (rc %d)\n",
200 virq
, smp_ipi_name
[msg
], err
);
205 #ifdef CONFIG_PPC_SMP_MUXED_IPI
206 struct cpu_messages
{
207 int messages
; /* current messages */
208 unsigned long data
; /* data for cause ipi */
210 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages
, ipi_message
);
212 void smp_muxed_ipi_set_data(int cpu
, unsigned long data
)
214 struct cpu_messages
*info
= &per_cpu(ipi_message
, cpu
);
219 void smp_muxed_ipi_message_pass(int cpu
, int msg
)
221 struct cpu_messages
*info
= &per_cpu(ipi_message
, cpu
);
222 char *message
= (char *)&info
->messages
;
225 * Order previous accesses before accesses in the IPI handler.
230 * cause_ipi functions are required to include a full barrier
231 * before doing whatever causes the IPI.
233 smp_ops
->cause_ipi(cpu
, info
->data
);
236 #ifdef __BIG_ENDIAN__
237 #define IPI_MESSAGE(A) (1 << (24 - 8 * (A)))
239 #define IPI_MESSAGE(A) (1 << (8 * (A)))
242 irqreturn_t
smp_ipi_demux(void)
244 struct cpu_messages
*info
= &__get_cpu_var(ipi_message
);
247 mb(); /* order any irq clear */
250 all
= xchg(&info
->messages
, 0);
251 if (all
& IPI_MESSAGE(PPC_MSG_CALL_FUNCTION
))
252 generic_smp_call_function_interrupt();
253 if (all
& IPI_MESSAGE(PPC_MSG_RESCHEDULE
))
255 if (all
& IPI_MESSAGE(PPC_MSG_TICK_BROADCAST
))
256 tick_broadcast_ipi_handler();
257 if (all
& IPI_MESSAGE(PPC_MSG_DEBUGGER_BREAK
))
258 debug_ipi_action(0, NULL
);
259 } while (info
->messages
);
263 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
265 static inline void do_message_pass(int cpu
, int msg
)
267 if (smp_ops
->message_pass
)
268 smp_ops
->message_pass(cpu
, msg
);
269 #ifdef CONFIG_PPC_SMP_MUXED_IPI
271 smp_muxed_ipi_message_pass(cpu
, msg
);
275 void smp_send_reschedule(int cpu
)
278 do_message_pass(cpu
, PPC_MSG_RESCHEDULE
);
280 EXPORT_SYMBOL_GPL(smp_send_reschedule
);
282 void arch_send_call_function_single_ipi(int cpu
)
284 do_message_pass(cpu
, PPC_MSG_CALL_FUNCTION
);
287 void arch_send_call_function_ipi_mask(const struct cpumask
*mask
)
291 for_each_cpu(cpu
, mask
)
292 do_message_pass(cpu
, PPC_MSG_CALL_FUNCTION
);
295 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
296 void tick_broadcast(const struct cpumask
*mask
)
300 for_each_cpu(cpu
, mask
)
301 do_message_pass(cpu
, PPC_MSG_TICK_BROADCAST
);
305 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
306 void smp_send_debugger_break(void)
309 int me
= raw_smp_processor_id();
311 if (unlikely(!smp_ops
))
314 for_each_online_cpu(cpu
)
316 do_message_pass(cpu
, PPC_MSG_DEBUGGER_BREAK
);
321 void crash_send_ipi(void (*crash_ipi_callback
)(struct pt_regs
*))
323 crash_ipi_function_ptr
= crash_ipi_callback
;
324 if (crash_ipi_callback
) {
326 smp_send_debugger_break();
331 static void stop_this_cpu(void *dummy
)
333 /* Remove this CPU */
334 set_cpu_online(smp_processor_id(), false);
341 void smp_send_stop(void)
343 smp_call_function(stop_this_cpu
, NULL
, 0);
346 struct thread_info
*current_set
[NR_CPUS
];
348 static void smp_store_cpu_info(int id
)
350 per_cpu(cpu_pvr
, id
) = mfspr(SPRN_PVR
);
351 #ifdef CONFIG_PPC_FSL_BOOK3E
352 per_cpu(next_tlbcam_idx
, id
)
353 = (mfspr(SPRN_TLB1CFG
) & TLBnCFG_N_ENTRY
) - 1;
357 void __init
smp_prepare_cpus(unsigned int max_cpus
)
361 DBG("smp_prepare_cpus\n");
364 * setup_cpu may need to be called on the boot cpu. We havent
365 * spun any cpus up but lets be paranoid.
367 BUG_ON(boot_cpuid
!= smp_processor_id());
370 smp_store_cpu_info(boot_cpuid
);
371 cpu_callin_map
[boot_cpuid
] = 1;
373 for_each_possible_cpu(cpu
) {
374 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map
, cpu
),
375 GFP_KERNEL
, cpu_to_node(cpu
));
376 zalloc_cpumask_var_node(&per_cpu(cpu_core_map
, cpu
),
377 GFP_KERNEL
, cpu_to_node(cpu
));
380 cpumask_set_cpu(boot_cpuid
, cpu_sibling_mask(boot_cpuid
));
381 cpumask_set_cpu(boot_cpuid
, cpu_core_mask(boot_cpuid
));
383 if (smp_ops
&& smp_ops
->probe
)
387 void smp_prepare_boot_cpu(void)
389 BUG_ON(smp_processor_id() != boot_cpuid
);
391 paca
[boot_cpuid
].__current
= current
;
393 current_set
[boot_cpuid
] = task_thread_info(current
);
396 #ifdef CONFIG_HOTPLUG_CPU
398 int generic_cpu_disable(void)
400 unsigned int cpu
= smp_processor_id();
402 if (cpu
== boot_cpuid
)
405 set_cpu_online(cpu
, false);
407 vdso_data
->processorCount
--;
413 void generic_cpu_die(unsigned int cpu
)
417 for (i
= 0; i
< 100; i
++) {
419 if (per_cpu(cpu_state
, cpu
) == CPU_DEAD
)
423 printk(KERN_ERR
"CPU%d didn't die...\n", cpu
);
426 void generic_mach_cpu_die(void)
432 cpu
= smp_processor_id();
433 printk(KERN_DEBUG
"CPU%d offline\n", cpu
);
434 __get_cpu_var(cpu_state
) = CPU_DEAD
;
436 while (__get_cpu_var(cpu_state
) != CPU_UP_PREPARE
)
440 void generic_set_cpu_dead(unsigned int cpu
)
442 per_cpu(cpu_state
, cpu
) = CPU_DEAD
;
446 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
447 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
448 * which makes the delay in generic_cpu_die() not happen.
450 void generic_set_cpu_up(unsigned int cpu
)
452 per_cpu(cpu_state
, cpu
) = CPU_UP_PREPARE
;
455 int generic_check_cpu_restart(unsigned int cpu
)
457 return per_cpu(cpu_state
, cpu
) == CPU_UP_PREPARE
;
460 static atomic_t secondary_inhibit_count
;
463 * Don't allow secondary CPU threads to come online
465 void inhibit_secondary_onlining(void)
468 * This makes secondary_inhibit_count stable during cpu
469 * online/offline operations.
473 atomic_inc(&secondary_inhibit_count
);
476 EXPORT_SYMBOL_GPL(inhibit_secondary_onlining
);
479 * Allow secondary CPU threads to come online again
481 void uninhibit_secondary_onlining(void)
484 atomic_dec(&secondary_inhibit_count
);
487 EXPORT_SYMBOL_GPL(uninhibit_secondary_onlining
);
489 static int secondaries_inhibited(void)
491 return atomic_read(&secondary_inhibit_count
);
494 #else /* HOTPLUG_CPU */
496 #define secondaries_inhibited() 0
500 static void cpu_idle_thread_init(unsigned int cpu
, struct task_struct
*idle
)
502 struct thread_info
*ti
= task_thread_info(idle
);
505 paca
[cpu
].__current
= idle
;
506 paca
[cpu
].kstack
= (unsigned long)ti
+ THREAD_SIZE
- STACK_FRAME_OVERHEAD
;
509 secondary_ti
= current_set
[cpu
] = ti
;
512 int __cpu_up(unsigned int cpu
, struct task_struct
*tidle
)
517 * Don't allow secondary threads to come online if inhibited
519 if (threads_per_core
> 1 && secondaries_inhibited() &&
520 cpu
% threads_per_core
!= 0)
523 if (smp_ops
== NULL
||
524 (smp_ops
->cpu_bootable
&& !smp_ops
->cpu_bootable(cpu
)))
527 cpu_idle_thread_init(cpu
, tidle
);
529 /* Make sure callin-map entry is 0 (can be leftover a CPU
532 cpu_callin_map
[cpu
] = 0;
534 /* The information for processor bringup must
535 * be written out to main store before we release
541 DBG("smp: kicking cpu %d\n", cpu
);
542 rc
= smp_ops
->kick_cpu(cpu
);
544 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu
, rc
);
549 * wait to see if the cpu made a callin (is actually up).
550 * use this value that I found through experimentation.
553 if (system_state
< SYSTEM_RUNNING
)
554 for (c
= 50000; c
&& !cpu_callin_map
[cpu
]; c
--)
556 #ifdef CONFIG_HOTPLUG_CPU
559 * CPUs can take much longer to come up in the
560 * hotplug case. Wait five seconds.
562 for (c
= 5000; c
&& !cpu_callin_map
[cpu
]; c
--)
566 if (!cpu_callin_map
[cpu
]) {
567 printk(KERN_ERR
"Processor %u is stuck.\n", cpu
);
571 DBG("Processor %u found.\n", cpu
);
573 if (smp_ops
->give_timebase
)
574 smp_ops
->give_timebase();
576 /* Wait until cpu puts itself in the online map */
577 while (!cpu_online(cpu
))
583 /* Return the value of the reg property corresponding to the given
586 int cpu_to_core_id(int cpu
)
588 struct device_node
*np
;
592 np
= of_get_cpu_node(cpu
, NULL
);
596 reg
= of_get_property(np
, "reg", NULL
);
600 id
= be32_to_cpup(reg
);
606 /* Helper routines for cpu to core mapping */
607 int cpu_core_index_of_thread(int cpu
)
609 return cpu
>> threads_shift
;
611 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread
);
613 int cpu_first_thread_of_core(int core
)
615 return core
<< threads_shift
;
617 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core
);
619 static void traverse_siblings_chip_id(int cpu
, bool add
, int chipid
)
621 const struct cpumask
*mask
;
622 struct device_node
*np
;
626 mask
= add
? cpu_online_mask
: cpu_present_mask
;
627 for_each_cpu(i
, mask
) {
628 np
= of_get_cpu_node(i
, NULL
);
631 prop
= of_get_property(np
, "ibm,chip-id", &plen
);
632 if (prop
&& plen
== sizeof(int) &&
633 of_read_number(prop
, 1) == chipid
) {
635 cpumask_set_cpu(cpu
, cpu_core_mask(i
));
636 cpumask_set_cpu(i
, cpu_core_mask(cpu
));
638 cpumask_clear_cpu(cpu
, cpu_core_mask(i
));
639 cpumask_clear_cpu(i
, cpu_core_mask(cpu
));
646 /* Must be called when no change can occur to cpu_present_mask,
647 * i.e. during cpu online or offline.
649 static struct device_node
*cpu_to_l2cache(int cpu
)
651 struct device_node
*np
;
652 struct device_node
*cache
;
654 if (!cpu_present(cpu
))
657 np
= of_get_cpu_node(cpu
, NULL
);
661 cache
= of_find_next_cache_node(np
);
668 static void traverse_core_siblings(int cpu
, bool add
)
670 struct device_node
*l2_cache
, *np
;
671 const struct cpumask
*mask
;
675 /* First see if we have ibm,chip-id properties in cpu nodes */
676 np
= of_get_cpu_node(cpu
, NULL
);
679 prop
= of_get_property(np
, "ibm,chip-id", &plen
);
680 if (prop
&& plen
== sizeof(int))
681 chip
= of_read_number(prop
, 1);
684 traverse_siblings_chip_id(cpu
, add
, chip
);
689 l2_cache
= cpu_to_l2cache(cpu
);
690 mask
= add
? cpu_online_mask
: cpu_present_mask
;
691 for_each_cpu(i
, mask
) {
692 np
= cpu_to_l2cache(i
);
695 if (np
== l2_cache
) {
697 cpumask_set_cpu(cpu
, cpu_core_mask(i
));
698 cpumask_set_cpu(i
, cpu_core_mask(cpu
));
700 cpumask_clear_cpu(cpu
, cpu_core_mask(i
));
701 cpumask_clear_cpu(i
, cpu_core_mask(cpu
));
706 of_node_put(l2_cache
);
709 /* Activate a secondary processor. */
710 void start_secondary(void *unused
)
712 unsigned int cpu
= smp_processor_id();
715 atomic_inc(&init_mm
.mm_count
);
716 current
->active_mm
= &init_mm
;
718 smp_store_cpu_info(cpu
);
719 set_dec(tb_ticks_per_jiffy
);
721 cpu_callin_map
[cpu
] = 1;
723 if (smp_ops
->setup_cpu
)
724 smp_ops
->setup_cpu(cpu
);
725 if (smp_ops
->take_timebase
)
726 smp_ops
->take_timebase();
728 secondary_cpu_time_init();
731 if (system_state
== SYSTEM_RUNNING
)
732 vdso_data
->processorCount
++;
736 /* Update sibling maps */
737 base
= cpu_first_thread_sibling(cpu
);
738 for (i
= 0; i
< threads_per_core
; i
++) {
739 if (cpu_is_offline(base
+ i
) && (cpu
!= base
+ i
))
741 cpumask_set_cpu(cpu
, cpu_sibling_mask(base
+ i
));
742 cpumask_set_cpu(base
+ i
, cpu_sibling_mask(cpu
));
744 /* cpu_core_map should be a superset of
745 * cpu_sibling_map even if we don't have cache
746 * information, so update the former here, too.
748 cpumask_set_cpu(cpu
, cpu_core_mask(base
+ i
));
749 cpumask_set_cpu(base
+ i
, cpu_core_mask(cpu
));
751 traverse_core_siblings(cpu
, true);
754 notify_cpu_starting(cpu
);
755 set_cpu_online(cpu
, true);
759 cpu_startup_entry(CPUHP_ONLINE
);
764 int setup_profiling_timer(unsigned int multiplier
)
769 void __init
smp_cpus_done(unsigned int max_cpus
)
771 cpumask_var_t old_mask
;
773 /* We want the setup_cpu() here to be called from CPU 0, but our
774 * init thread may have been "borrowed" by another CPU in the meantime
775 * se we pin us down to CPU 0 for a short while
777 alloc_cpumask_var(&old_mask
, GFP_NOWAIT
);
778 cpumask_copy(old_mask
, tsk_cpus_allowed(current
));
779 set_cpus_allowed_ptr(current
, cpumask_of(boot_cpuid
));
781 if (smp_ops
&& smp_ops
->setup_cpu
)
782 smp_ops
->setup_cpu(boot_cpuid
);
784 set_cpus_allowed_ptr(current
, old_mask
);
786 free_cpumask_var(old_mask
);
788 if (smp_ops
&& smp_ops
->bringup_done
)
789 smp_ops
->bringup_done();
791 dump_numa_cpu_topology();
795 int arch_sd_sibling_asym_packing(void)
797 if (cpu_has_feature(CPU_FTR_ASYM_SMT
)) {
798 printk_once(KERN_INFO
"Enabling Asymmetric SMT scheduling\n");
799 return SD_ASYM_PACKING
;
804 #ifdef CONFIG_HOTPLUG_CPU
805 int __cpu_disable(void)
807 int cpu
= smp_processor_id();
811 if (!smp_ops
->cpu_disable
)
814 err
= smp_ops
->cpu_disable();
818 /* Update sibling maps */
819 base
= cpu_first_thread_sibling(cpu
);
820 for (i
= 0; i
< threads_per_core
; i
++) {
821 cpumask_clear_cpu(cpu
, cpu_sibling_mask(base
+ i
));
822 cpumask_clear_cpu(base
+ i
, cpu_sibling_mask(cpu
));
823 cpumask_clear_cpu(cpu
, cpu_core_mask(base
+ i
));
824 cpumask_clear_cpu(base
+ i
, cpu_core_mask(cpu
));
826 traverse_core_siblings(cpu
, false);
831 void __cpu_die(unsigned int cpu
)
833 if (smp_ops
->cpu_die
)
834 smp_ops
->cpu_die(cpu
);
842 /* If we return, we re-enter start_secondary */
843 start_secondary_resume();