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>
58 #define DBG(fmt...) udbg_printf(fmt)
63 #ifdef CONFIG_HOTPLUG_CPU
64 /* State of each CPU during hotplug phases */
65 static DEFINE_PER_CPU(int, cpu_state
) = { 0 };
68 struct thread_info
*secondary_ti
;
70 DEFINE_PER_CPU(cpumask_var_t
, cpu_sibling_map
);
71 DEFINE_PER_CPU(cpumask_var_t
, cpu_core_map
);
73 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map
);
74 EXPORT_PER_CPU_SYMBOL(cpu_core_map
);
76 /* SMP operations for this machine */
77 struct smp_ops_t
*smp_ops
;
79 /* Can't be static due to PowerMac hackery */
80 volatile unsigned int cpu_callin_map
[NR_CPUS
];
82 int smt_enabled_at_boot
= 1;
84 static void (*crash_ipi_function_ptr
)(struct pt_regs
*) = NULL
;
87 * Returns 1 if the specified cpu should be brought up during boot.
88 * Used to inhibit booting threads if they've been disabled or
89 * limited on the command line
91 int smp_generic_cpu_bootable(unsigned int nr
)
93 /* Special case - we inhibit secondary thread startup
94 * during boot if the user requests it.
96 if (system_state
== SYSTEM_BOOTING
&& cpu_has_feature(CPU_FTR_SMT
)) {
97 if (!smt_enabled_at_boot
&& cpu_thread_in_core(nr
) != 0)
99 if (smt_enabled_at_boot
100 && cpu_thread_in_core(nr
) >= smt_enabled_at_boot
)
109 int smp_generic_kick_cpu(int nr
)
111 BUG_ON(nr
< 0 || nr
>= NR_CPUS
);
114 * The processor is currently spinning, waiting for the
115 * cpu_start field to become non-zero After we set cpu_start,
116 * the processor will continue on to secondary_start
118 if (!paca
[nr
].cpu_start
) {
119 paca
[nr
].cpu_start
= 1;
124 #ifdef CONFIG_HOTPLUG_CPU
126 * Ok it's not there, so it might be soft-unplugged, let's
127 * try to bring it back
129 generic_set_cpu_up(nr
);
131 smp_send_reschedule(nr
);
132 #endif /* CONFIG_HOTPLUG_CPU */
136 #endif /* CONFIG_PPC64 */
138 static irqreturn_t
call_function_action(int irq
, void *data
)
140 generic_smp_call_function_interrupt();
144 static irqreturn_t
reschedule_action(int irq
, void *data
)
150 static irqreturn_t
tick_broadcast_ipi_action(int irq
, void *data
)
152 tick_broadcast_ipi_handler();
156 static irqreturn_t
debug_ipi_action(int irq
, void *data
)
158 if (crash_ipi_function_ptr
) {
159 crash_ipi_function_ptr(get_irq_regs());
163 #ifdef CONFIG_DEBUGGER
164 debugger_ipi(get_irq_regs());
165 #endif /* CONFIG_DEBUGGER */
170 static irq_handler_t smp_ipi_action
[] = {
171 [PPC_MSG_CALL_FUNCTION
] = call_function_action
,
172 [PPC_MSG_RESCHEDULE
] = reschedule_action
,
173 [PPC_MSG_TICK_BROADCAST
] = tick_broadcast_ipi_action
,
174 [PPC_MSG_DEBUGGER_BREAK
] = debug_ipi_action
,
177 const char *smp_ipi_name
[] = {
178 [PPC_MSG_CALL_FUNCTION
] = "ipi call function",
179 [PPC_MSG_RESCHEDULE
] = "ipi reschedule",
180 [PPC_MSG_TICK_BROADCAST
] = "ipi tick-broadcast",
181 [PPC_MSG_DEBUGGER_BREAK
] = "ipi debugger",
184 /* optional function to request ipi, for controllers with >= 4 ipis */
185 int smp_request_message_ipi(int virq
, int msg
)
189 if (msg
< 0 || msg
> PPC_MSG_DEBUGGER_BREAK
) {
192 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
193 if (msg
== PPC_MSG_DEBUGGER_BREAK
) {
197 err
= request_irq(virq
, smp_ipi_action
[msg
],
198 IRQF_PERCPU
| IRQF_NO_THREAD
| IRQF_NO_SUSPEND
,
199 smp_ipi_name
[msg
], NULL
);
200 WARN(err
< 0, "unable to request_irq %d for %s (rc %d)\n",
201 virq
, smp_ipi_name
[msg
], err
);
206 #ifdef CONFIG_PPC_SMP_MUXED_IPI
207 struct cpu_messages
{
208 int messages
; /* current messages */
209 unsigned long data
; /* data for cause ipi */
211 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages
, ipi_message
);
213 void smp_muxed_ipi_set_data(int cpu
, unsigned long data
)
215 struct cpu_messages
*info
= &per_cpu(ipi_message
, cpu
);
220 void smp_muxed_ipi_message_pass(int cpu
, int msg
)
222 struct cpu_messages
*info
= &per_cpu(ipi_message
, cpu
);
223 char *message
= (char *)&info
->messages
;
226 * Order previous accesses before accesses in the IPI handler.
231 * cause_ipi functions are required to include a full barrier
232 * before doing whatever causes the IPI.
234 smp_ops
->cause_ipi(cpu
, info
->data
);
237 #ifdef __BIG_ENDIAN__
238 #define IPI_MESSAGE(A) (1 << (24 - 8 * (A)))
240 #define IPI_MESSAGE(A) (1 << (8 * (A)))
243 irqreturn_t
smp_ipi_demux(void)
245 struct cpu_messages
*info
= &__get_cpu_var(ipi_message
);
248 mb(); /* order any irq clear */
251 all
= xchg(&info
->messages
, 0);
252 if (all
& IPI_MESSAGE(PPC_MSG_CALL_FUNCTION
))
253 generic_smp_call_function_interrupt();
254 if (all
& IPI_MESSAGE(PPC_MSG_RESCHEDULE
))
256 if (all
& IPI_MESSAGE(PPC_MSG_TICK_BROADCAST
))
257 tick_broadcast_ipi_handler();
258 if (all
& IPI_MESSAGE(PPC_MSG_DEBUGGER_BREAK
))
259 debug_ipi_action(0, NULL
);
260 } while (info
->messages
);
264 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
266 static inline void do_message_pass(int cpu
, int msg
)
268 if (smp_ops
->message_pass
)
269 smp_ops
->message_pass(cpu
, msg
);
270 #ifdef CONFIG_PPC_SMP_MUXED_IPI
272 smp_muxed_ipi_message_pass(cpu
, msg
);
276 void smp_send_reschedule(int cpu
)
279 do_message_pass(cpu
, PPC_MSG_RESCHEDULE
);
281 EXPORT_SYMBOL_GPL(smp_send_reschedule
);
283 void arch_send_call_function_single_ipi(int cpu
)
285 do_message_pass(cpu
, PPC_MSG_CALL_FUNCTION
);
288 void arch_send_call_function_ipi_mask(const struct cpumask
*mask
)
292 for_each_cpu(cpu
, mask
)
293 do_message_pass(cpu
, PPC_MSG_CALL_FUNCTION
);
296 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
297 void tick_broadcast(const struct cpumask
*mask
)
301 for_each_cpu(cpu
, mask
)
302 do_message_pass(cpu
, PPC_MSG_TICK_BROADCAST
);
306 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
307 void smp_send_debugger_break(void)
310 int me
= raw_smp_processor_id();
312 if (unlikely(!smp_ops
))
315 for_each_online_cpu(cpu
)
317 do_message_pass(cpu
, PPC_MSG_DEBUGGER_BREAK
);
322 void crash_send_ipi(void (*crash_ipi_callback
)(struct pt_regs
*))
324 crash_ipi_function_ptr
= crash_ipi_callback
;
325 if (crash_ipi_callback
) {
327 smp_send_debugger_break();
332 static void stop_this_cpu(void *dummy
)
334 /* Remove this CPU */
335 set_cpu_online(smp_processor_id(), false);
342 void smp_send_stop(void)
344 smp_call_function(stop_this_cpu
, NULL
, 0);
347 struct thread_info
*current_set
[NR_CPUS
];
349 static void smp_store_cpu_info(int id
)
351 per_cpu(cpu_pvr
, id
) = mfspr(SPRN_PVR
);
352 #ifdef CONFIG_PPC_FSL_BOOK3E
353 per_cpu(next_tlbcam_idx
, id
)
354 = (mfspr(SPRN_TLB1CFG
) & TLBnCFG_N_ENTRY
) - 1;
358 void __init
smp_prepare_cpus(unsigned int max_cpus
)
362 DBG("smp_prepare_cpus\n");
365 * setup_cpu may need to be called on the boot cpu. We havent
366 * spun any cpus up but lets be paranoid.
368 BUG_ON(boot_cpuid
!= smp_processor_id());
371 smp_store_cpu_info(boot_cpuid
);
372 cpu_callin_map
[boot_cpuid
] = 1;
374 for_each_possible_cpu(cpu
) {
375 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map
, cpu
),
376 GFP_KERNEL
, cpu_to_node(cpu
));
377 zalloc_cpumask_var_node(&per_cpu(cpu_core_map
, cpu
),
378 GFP_KERNEL
, cpu_to_node(cpu
));
380 * numa_node_id() works after this.
382 set_cpu_numa_node(cpu
, numa_cpu_lookup_table
[cpu
]);
383 set_cpu_numa_mem(cpu
, local_memory_node(numa_cpu_lookup_table
[cpu
]));
386 cpumask_set_cpu(boot_cpuid
, cpu_sibling_mask(boot_cpuid
));
387 cpumask_set_cpu(boot_cpuid
, cpu_core_mask(boot_cpuid
));
389 if (smp_ops
&& smp_ops
->probe
)
393 void smp_prepare_boot_cpu(void)
395 BUG_ON(smp_processor_id() != boot_cpuid
);
397 paca
[boot_cpuid
].__current
= current
;
399 set_numa_node(numa_cpu_lookup_table
[boot_cpuid
]);
400 current_set
[boot_cpuid
] = task_thread_info(current
);
403 #ifdef CONFIG_HOTPLUG_CPU
405 int generic_cpu_disable(void)
407 unsigned int cpu
= smp_processor_id();
409 if (cpu
== boot_cpuid
)
412 set_cpu_online(cpu
, false);
414 vdso_data
->processorCount
--;
420 void generic_cpu_die(unsigned int cpu
)
424 for (i
= 0; i
< 100; i
++) {
426 if (per_cpu(cpu_state
, cpu
) == CPU_DEAD
)
430 printk(KERN_ERR
"CPU%d didn't die...\n", cpu
);
433 void generic_mach_cpu_die(void)
439 cpu
= smp_processor_id();
440 printk(KERN_DEBUG
"CPU%d offline\n", cpu
);
441 __get_cpu_var(cpu_state
) = CPU_DEAD
;
443 while (__get_cpu_var(cpu_state
) != CPU_UP_PREPARE
)
447 void generic_set_cpu_dead(unsigned int cpu
)
449 per_cpu(cpu_state
, cpu
) = CPU_DEAD
;
453 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
454 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
455 * which makes the delay in generic_cpu_die() not happen.
457 void generic_set_cpu_up(unsigned int cpu
)
459 per_cpu(cpu_state
, cpu
) = CPU_UP_PREPARE
;
462 int generic_check_cpu_restart(unsigned int cpu
)
464 return per_cpu(cpu_state
, cpu
) == CPU_UP_PREPARE
;
467 static bool secondaries_inhibited(void)
469 return kvm_hv_mode_active();
472 #else /* HOTPLUG_CPU */
474 #define secondaries_inhibited() 0
478 static void cpu_idle_thread_init(unsigned int cpu
, struct task_struct
*idle
)
480 struct thread_info
*ti
= task_thread_info(idle
);
483 paca
[cpu
].__current
= idle
;
484 paca
[cpu
].kstack
= (unsigned long)ti
+ THREAD_SIZE
- STACK_FRAME_OVERHEAD
;
487 secondary_ti
= current_set
[cpu
] = ti
;
490 int __cpu_up(unsigned int cpu
, struct task_struct
*tidle
)
495 * Don't allow secondary threads to come online if inhibited
497 if (threads_per_core
> 1 && secondaries_inhibited() &&
498 cpu_thread_in_subcore(cpu
))
501 if (smp_ops
== NULL
||
502 (smp_ops
->cpu_bootable
&& !smp_ops
->cpu_bootable(cpu
)))
505 cpu_idle_thread_init(cpu
, tidle
);
507 /* Make sure callin-map entry is 0 (can be leftover a CPU
510 cpu_callin_map
[cpu
] = 0;
512 /* The information for processor bringup must
513 * be written out to main store before we release
519 DBG("smp: kicking cpu %d\n", cpu
);
520 rc
= smp_ops
->kick_cpu(cpu
);
522 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu
, rc
);
527 * wait to see if the cpu made a callin (is actually up).
528 * use this value that I found through experimentation.
531 if (system_state
< SYSTEM_RUNNING
)
532 for (c
= 50000; c
&& !cpu_callin_map
[cpu
]; c
--)
534 #ifdef CONFIG_HOTPLUG_CPU
537 * CPUs can take much longer to come up in the
538 * hotplug case. Wait five seconds.
540 for (c
= 5000; c
&& !cpu_callin_map
[cpu
]; c
--)
544 if (!cpu_callin_map
[cpu
]) {
545 printk(KERN_ERR
"Processor %u is stuck.\n", cpu
);
549 DBG("Processor %u found.\n", cpu
);
551 if (smp_ops
->give_timebase
)
552 smp_ops
->give_timebase();
554 /* Wait until cpu puts itself in the online map */
555 while (!cpu_online(cpu
))
561 /* Return the value of the reg property corresponding to the given
564 int cpu_to_core_id(int cpu
)
566 struct device_node
*np
;
570 np
= of_get_cpu_node(cpu
, NULL
);
574 reg
= of_get_property(np
, "reg", NULL
);
578 id
= be32_to_cpup(reg
);
584 /* Helper routines for cpu to core mapping */
585 int cpu_core_index_of_thread(int cpu
)
587 return cpu
>> threads_shift
;
589 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread
);
591 int cpu_first_thread_of_core(int core
)
593 return core
<< threads_shift
;
595 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core
);
597 static void traverse_siblings_chip_id(int cpu
, bool add
, int chipid
)
599 const struct cpumask
*mask
;
600 struct device_node
*np
;
604 mask
= add
? cpu_online_mask
: cpu_present_mask
;
605 for_each_cpu(i
, mask
) {
606 np
= of_get_cpu_node(i
, NULL
);
609 prop
= of_get_property(np
, "ibm,chip-id", &plen
);
610 if (prop
&& plen
== sizeof(int) &&
611 of_read_number(prop
, 1) == chipid
) {
613 cpumask_set_cpu(cpu
, cpu_core_mask(i
));
614 cpumask_set_cpu(i
, cpu_core_mask(cpu
));
616 cpumask_clear_cpu(cpu
, cpu_core_mask(i
));
617 cpumask_clear_cpu(i
, cpu_core_mask(cpu
));
624 /* Must be called when no change can occur to cpu_present_mask,
625 * i.e. during cpu online or offline.
627 static struct device_node
*cpu_to_l2cache(int cpu
)
629 struct device_node
*np
;
630 struct device_node
*cache
;
632 if (!cpu_present(cpu
))
635 np
= of_get_cpu_node(cpu
, NULL
);
639 cache
= of_find_next_cache_node(np
);
646 static void traverse_core_siblings(int cpu
, bool add
)
648 struct device_node
*l2_cache
, *np
;
649 const struct cpumask
*mask
;
653 /* First see if we have ibm,chip-id properties in cpu nodes */
654 np
= of_get_cpu_node(cpu
, NULL
);
657 prop
= of_get_property(np
, "ibm,chip-id", &plen
);
658 if (prop
&& plen
== sizeof(int))
659 chip
= of_read_number(prop
, 1);
662 traverse_siblings_chip_id(cpu
, add
, chip
);
667 l2_cache
= cpu_to_l2cache(cpu
);
668 mask
= add
? cpu_online_mask
: cpu_present_mask
;
669 for_each_cpu(i
, mask
) {
670 np
= cpu_to_l2cache(i
);
673 if (np
== l2_cache
) {
675 cpumask_set_cpu(cpu
, cpu_core_mask(i
));
676 cpumask_set_cpu(i
, cpu_core_mask(cpu
));
678 cpumask_clear_cpu(cpu
, cpu_core_mask(i
));
679 cpumask_clear_cpu(i
, cpu_core_mask(cpu
));
684 of_node_put(l2_cache
);
687 /* Activate a secondary processor. */
688 void start_secondary(void *unused
)
690 unsigned int cpu
= smp_processor_id();
693 atomic_inc(&init_mm
.mm_count
);
694 current
->active_mm
= &init_mm
;
696 smp_store_cpu_info(cpu
);
697 set_dec(tb_ticks_per_jiffy
);
699 cpu_callin_map
[cpu
] = 1;
701 if (smp_ops
->setup_cpu
)
702 smp_ops
->setup_cpu(cpu
);
703 if (smp_ops
->take_timebase
)
704 smp_ops
->take_timebase();
706 secondary_cpu_time_init();
709 if (system_state
== SYSTEM_RUNNING
)
710 vdso_data
->processorCount
++;
714 /* Update sibling maps */
715 base
= cpu_first_thread_sibling(cpu
);
716 for (i
= 0; i
< threads_per_core
; i
++) {
717 if (cpu_is_offline(base
+ i
) && (cpu
!= base
+ i
))
719 cpumask_set_cpu(cpu
, cpu_sibling_mask(base
+ i
));
720 cpumask_set_cpu(base
+ i
, cpu_sibling_mask(cpu
));
722 /* cpu_core_map should be a superset of
723 * cpu_sibling_map even if we don't have cache
724 * information, so update the former here, too.
726 cpumask_set_cpu(cpu
, cpu_core_mask(base
+ i
));
727 cpumask_set_cpu(base
+ i
, cpu_core_mask(cpu
));
729 traverse_core_siblings(cpu
, true);
732 notify_cpu_starting(cpu
);
733 set_cpu_online(cpu
, true);
737 cpu_startup_entry(CPUHP_ONLINE
);
742 int setup_profiling_timer(unsigned int multiplier
)
747 #ifdef CONFIG_SCHED_SMT
748 /* cpumask of CPUs with asymetric SMT dependancy */
749 static int powerpc_smt_flags(void)
751 int flags
= SD_SHARE_CPUCAPACITY
| SD_SHARE_PKG_RESOURCES
;
753 if (cpu_has_feature(CPU_FTR_ASYM_SMT
)) {
754 printk_once(KERN_INFO
"Enabling Asymmetric SMT scheduling\n");
755 flags
|= SD_ASYM_PACKING
;
761 static struct sched_domain_topology_level powerpc_topology
[] = {
762 #ifdef CONFIG_SCHED_SMT
763 { cpu_smt_mask
, powerpc_smt_flags
, SD_INIT_NAME(SMT
) },
765 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
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();
793 set_sched_topology(powerpc_topology
);
797 #ifdef CONFIG_HOTPLUG_CPU
798 int __cpu_disable(void)
800 int cpu
= smp_processor_id();
804 if (!smp_ops
->cpu_disable
)
807 err
= smp_ops
->cpu_disable();
811 /* Update sibling maps */
812 base
= cpu_first_thread_sibling(cpu
);
813 for (i
= 0; i
< threads_per_core
; i
++) {
814 cpumask_clear_cpu(cpu
, cpu_sibling_mask(base
+ i
));
815 cpumask_clear_cpu(base
+ i
, cpu_sibling_mask(cpu
));
816 cpumask_clear_cpu(cpu
, cpu_core_mask(base
+ i
));
817 cpumask_clear_cpu(base
+ i
, cpu_core_mask(cpu
));
819 traverse_core_siblings(cpu
, false);
824 void __cpu_die(unsigned int cpu
)
826 if (smp_ops
->cpu_die
)
827 smp_ops
->cpu_die(cpu
);
835 /* If we return, we re-enter start_secondary */
836 start_secondary_resume();