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/mm.h>
23 #include <linux/sched/topology.h>
24 #include <linux/smp.h>
25 #include <linux/interrupt.h>
26 #include <linux/delay.h>
27 #include <linux/init.h>
28 #include <linux/spinlock.h>
29 #include <linux/cache.h>
30 #include <linux/err.h>
31 #include <linux/device.h>
32 #include <linux/cpu.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/profile.h>
37 #include <asm/ptrace.h>
38 #include <linux/atomic.h>
40 #include <asm/hw_irq.h>
41 #include <asm/kvm_ppc.h>
42 #include <asm/dbell.h>
44 #include <asm/pgtable.h>
48 #include <asm/machdep.h>
49 #include <asm/cputhreads.h>
50 #include <asm/cputable.h>
52 #include <asm/vdso_datapage.h>
57 #include <asm/debug.h>
58 #include <asm/kexec.h>
59 #include <asm/asm-prototypes.h>
60 #include <asm/cpu_has_feature.h>
64 #define DBG(fmt...) udbg_printf(fmt)
69 #ifdef CONFIG_HOTPLUG_CPU
70 /* State of each CPU during hotplug phases */
71 static DEFINE_PER_CPU(int, cpu_state
) = { 0 };
74 struct thread_info
*secondary_ti
;
76 DEFINE_PER_CPU(cpumask_var_t
, cpu_sibling_map
);
77 DEFINE_PER_CPU(cpumask_var_t
, cpu_core_map
);
79 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map
);
80 EXPORT_PER_CPU_SYMBOL(cpu_core_map
);
82 /* SMP operations for this machine */
83 struct smp_ops_t
*smp_ops
;
85 /* Can't be static due to PowerMac hackery */
86 volatile unsigned int cpu_callin_map
[NR_CPUS
];
88 int smt_enabled_at_boot
= 1;
90 static void (*crash_ipi_function_ptr
)(struct pt_regs
*) = NULL
;
93 * Returns 1 if the specified cpu should be brought up during boot.
94 * Used to inhibit booting threads if they've been disabled or
95 * limited on the command line
97 int smp_generic_cpu_bootable(unsigned int nr
)
99 /* Special case - we inhibit secondary thread startup
100 * during boot if the user requests it.
102 if (system_state
== SYSTEM_BOOTING
&& cpu_has_feature(CPU_FTR_SMT
)) {
103 if (!smt_enabled_at_boot
&& cpu_thread_in_core(nr
) != 0)
105 if (smt_enabled_at_boot
106 && cpu_thread_in_core(nr
) >= smt_enabled_at_boot
)
115 int smp_generic_kick_cpu(int nr
)
117 BUG_ON(nr
< 0 || nr
>= NR_CPUS
);
120 * The processor is currently spinning, waiting for the
121 * cpu_start field to become non-zero After we set cpu_start,
122 * the processor will continue on to secondary_start
124 if (!paca
[nr
].cpu_start
) {
125 paca
[nr
].cpu_start
= 1;
130 #ifdef CONFIG_HOTPLUG_CPU
132 * Ok it's not there, so it might be soft-unplugged, let's
133 * try to bring it back
135 generic_set_cpu_up(nr
);
137 smp_send_reschedule(nr
);
138 #endif /* CONFIG_HOTPLUG_CPU */
142 #endif /* CONFIG_PPC64 */
144 static irqreturn_t
call_function_action(int irq
, void *data
)
146 generic_smp_call_function_interrupt();
150 static irqreturn_t
reschedule_action(int irq
, void *data
)
156 static irqreturn_t
tick_broadcast_ipi_action(int irq
, void *data
)
158 tick_broadcast_ipi_handler();
162 static irqreturn_t
debug_ipi_action(int irq
, void *data
)
164 if (crash_ipi_function_ptr
) {
165 crash_ipi_function_ptr(get_irq_regs());
169 #ifdef CONFIG_DEBUGGER
170 debugger_ipi(get_irq_regs());
171 #endif /* CONFIG_DEBUGGER */
176 static irq_handler_t smp_ipi_action
[] = {
177 [PPC_MSG_CALL_FUNCTION
] = call_function_action
,
178 [PPC_MSG_RESCHEDULE
] = reschedule_action
,
179 [PPC_MSG_TICK_BROADCAST
] = tick_broadcast_ipi_action
,
180 [PPC_MSG_DEBUGGER_BREAK
] = debug_ipi_action
,
183 const char *smp_ipi_name
[] = {
184 [PPC_MSG_CALL_FUNCTION
] = "ipi call function",
185 [PPC_MSG_RESCHEDULE
] = "ipi reschedule",
186 [PPC_MSG_TICK_BROADCAST
] = "ipi tick-broadcast",
187 [PPC_MSG_DEBUGGER_BREAK
] = "ipi debugger",
190 /* optional function to request ipi, for controllers with >= 4 ipis */
191 int smp_request_message_ipi(int virq
, int msg
)
195 if (msg
< 0 || msg
> PPC_MSG_DEBUGGER_BREAK
) {
198 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC_CORE)
199 if (msg
== PPC_MSG_DEBUGGER_BREAK
) {
203 err
= request_irq(virq
, smp_ipi_action
[msg
],
204 IRQF_PERCPU
| IRQF_NO_THREAD
| IRQF_NO_SUSPEND
,
205 smp_ipi_name
[msg
], NULL
);
206 WARN(err
< 0, "unable to request_irq %d for %s (rc %d)\n",
207 virq
, smp_ipi_name
[msg
], err
);
212 #ifdef CONFIG_PPC_SMP_MUXED_IPI
213 struct cpu_messages
{
214 long messages
; /* current messages */
216 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages
, ipi_message
);
218 void smp_muxed_ipi_set_message(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.
230 void smp_muxed_ipi_message_pass(int cpu
, int msg
)
232 smp_muxed_ipi_set_message(cpu
, msg
);
235 * cause_ipi functions are required to include a full barrier
236 * before doing whatever causes the IPI.
238 smp_ops
->cause_ipi(cpu
);
241 #ifdef __BIG_ENDIAN__
242 #define IPI_MESSAGE(A) (1uL << ((BITS_PER_LONG - 8) - 8 * (A)))
244 #define IPI_MESSAGE(A) (1uL << (8 * (A)))
247 irqreturn_t
smp_ipi_demux(void)
249 struct cpu_messages
*info
;
252 mb(); /* order any irq clear */
254 info
= this_cpu_ptr(&ipi_message
);
256 all
= xchg(&info
->messages
, 0);
257 #if defined(CONFIG_KVM_XICS) && defined(CONFIG_KVM_BOOK3S_HV_POSSIBLE)
259 * Must check for PPC_MSG_RM_HOST_ACTION messages
260 * before PPC_MSG_CALL_FUNCTION messages because when
261 * a VM is destroyed, we call kick_all_cpus_sync()
262 * to ensure that any pending PPC_MSG_RM_HOST_ACTION
263 * messages have completed before we free any VCPUs.
265 if (all
& IPI_MESSAGE(PPC_MSG_RM_HOST_ACTION
))
266 kvmppc_xics_ipi_action();
268 if (all
& IPI_MESSAGE(PPC_MSG_CALL_FUNCTION
))
269 generic_smp_call_function_interrupt();
270 if (all
& IPI_MESSAGE(PPC_MSG_RESCHEDULE
))
272 if (all
& IPI_MESSAGE(PPC_MSG_TICK_BROADCAST
))
273 tick_broadcast_ipi_handler();
274 if (all
& IPI_MESSAGE(PPC_MSG_DEBUGGER_BREAK
))
275 debug_ipi_action(0, NULL
);
276 } while (info
->messages
);
280 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
282 static inline void do_message_pass(int cpu
, int msg
)
284 if (smp_ops
->message_pass
)
285 smp_ops
->message_pass(cpu
, msg
);
286 #ifdef CONFIG_PPC_SMP_MUXED_IPI
288 smp_muxed_ipi_message_pass(cpu
, msg
);
292 void smp_send_reschedule(int cpu
)
295 do_message_pass(cpu
, PPC_MSG_RESCHEDULE
);
297 EXPORT_SYMBOL_GPL(smp_send_reschedule
);
299 void arch_send_call_function_single_ipi(int cpu
)
301 do_message_pass(cpu
, PPC_MSG_CALL_FUNCTION
);
304 void arch_send_call_function_ipi_mask(const struct cpumask
*mask
)
308 for_each_cpu(cpu
, mask
)
309 do_message_pass(cpu
, PPC_MSG_CALL_FUNCTION
);
312 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
313 void tick_broadcast(const struct cpumask
*mask
)
317 for_each_cpu(cpu
, mask
)
318 do_message_pass(cpu
, PPC_MSG_TICK_BROADCAST
);
322 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC_CORE)
323 void smp_send_debugger_break(void)
326 int me
= raw_smp_processor_id();
328 if (unlikely(!smp_ops
))
331 for_each_online_cpu(cpu
)
333 do_message_pass(cpu
, PPC_MSG_DEBUGGER_BREAK
);
337 #ifdef CONFIG_KEXEC_CORE
338 void crash_send_ipi(void (*crash_ipi_callback
)(struct pt_regs
*))
340 crash_ipi_function_ptr
= crash_ipi_callback
;
341 if (crash_ipi_callback
) {
343 smp_send_debugger_break();
348 static void stop_this_cpu(void *dummy
)
350 /* Remove this CPU */
351 set_cpu_online(smp_processor_id(), false);
358 void smp_send_stop(void)
360 smp_call_function(stop_this_cpu
, NULL
, 0);
363 struct thread_info
*current_set
[NR_CPUS
];
365 static void smp_store_cpu_info(int id
)
367 per_cpu(cpu_pvr
, id
) = mfspr(SPRN_PVR
);
368 #ifdef CONFIG_PPC_FSL_BOOK3E
369 per_cpu(next_tlbcam_idx
, id
)
370 = (mfspr(SPRN_TLB1CFG
) & TLBnCFG_N_ENTRY
) - 1;
374 void __init
smp_prepare_cpus(unsigned int max_cpus
)
378 DBG("smp_prepare_cpus\n");
381 * setup_cpu may need to be called on the boot cpu. We havent
382 * spun any cpus up but lets be paranoid.
384 BUG_ON(boot_cpuid
!= smp_processor_id());
387 smp_store_cpu_info(boot_cpuid
);
388 cpu_callin_map
[boot_cpuid
] = 1;
390 for_each_possible_cpu(cpu
) {
391 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map
, cpu
),
392 GFP_KERNEL
, cpu_to_node(cpu
));
393 zalloc_cpumask_var_node(&per_cpu(cpu_core_map
, cpu
),
394 GFP_KERNEL
, cpu_to_node(cpu
));
396 * numa_node_id() works after this.
398 if (cpu_present(cpu
)) {
399 set_cpu_numa_node(cpu
, numa_cpu_lookup_table
[cpu
]);
400 set_cpu_numa_mem(cpu
,
401 local_memory_node(numa_cpu_lookup_table
[cpu
]));
405 cpumask_set_cpu(boot_cpuid
, cpu_sibling_mask(boot_cpuid
));
406 cpumask_set_cpu(boot_cpuid
, cpu_core_mask(boot_cpuid
));
408 if (smp_ops
&& smp_ops
->probe
)
412 void smp_prepare_boot_cpu(void)
414 BUG_ON(smp_processor_id() != boot_cpuid
);
416 paca
[boot_cpuid
].__current
= current
;
418 set_numa_node(numa_cpu_lookup_table
[boot_cpuid
]);
419 current_set
[boot_cpuid
] = task_thread_info(current
);
422 #ifdef CONFIG_HOTPLUG_CPU
424 int generic_cpu_disable(void)
426 unsigned int cpu
= smp_processor_id();
428 if (cpu
== boot_cpuid
)
431 set_cpu_online(cpu
, false);
433 vdso_data
->processorCount
--;
435 /* Update affinity of all IRQs previously aimed at this CPU */
436 irq_migrate_all_off_this_cpu();
438 /* Give the CPU time to drain in-flight ones */
446 void generic_cpu_die(unsigned int cpu
)
450 for (i
= 0; i
< 100; i
++) {
452 if (is_cpu_dead(cpu
))
456 printk(KERN_ERR
"CPU%d didn't die...\n", cpu
);
459 void generic_set_cpu_dead(unsigned int cpu
)
461 per_cpu(cpu_state
, cpu
) = CPU_DEAD
;
465 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
466 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
467 * which makes the delay in generic_cpu_die() not happen.
469 void generic_set_cpu_up(unsigned int cpu
)
471 per_cpu(cpu_state
, cpu
) = CPU_UP_PREPARE
;
474 int generic_check_cpu_restart(unsigned int cpu
)
476 return per_cpu(cpu_state
, cpu
) == CPU_UP_PREPARE
;
479 int is_cpu_dead(unsigned int cpu
)
481 return per_cpu(cpu_state
, cpu
) == CPU_DEAD
;
484 static bool secondaries_inhibited(void)
486 return kvm_hv_mode_active();
489 #else /* HOTPLUG_CPU */
491 #define secondaries_inhibited() 0
495 static void cpu_idle_thread_init(unsigned int cpu
, struct task_struct
*idle
)
497 struct thread_info
*ti
= task_thread_info(idle
);
500 paca
[cpu
].__current
= idle
;
501 paca
[cpu
].kstack
= (unsigned long)ti
+ THREAD_SIZE
- STACK_FRAME_OVERHEAD
;
504 secondary_ti
= current_set
[cpu
] = ti
;
507 int __cpu_up(unsigned int cpu
, struct task_struct
*tidle
)
512 * Don't allow secondary threads to come online if inhibited
514 if (threads_per_core
> 1 && secondaries_inhibited() &&
515 cpu_thread_in_subcore(cpu
))
518 if (smp_ops
== NULL
||
519 (smp_ops
->cpu_bootable
&& !smp_ops
->cpu_bootable(cpu
)))
522 cpu_idle_thread_init(cpu
, tidle
);
525 * The platform might need to allocate resources prior to bringing
528 if (smp_ops
->prepare_cpu
) {
529 rc
= smp_ops
->prepare_cpu(cpu
);
534 /* Make sure callin-map entry is 0 (can be leftover a CPU
537 cpu_callin_map
[cpu
] = 0;
539 /* The information for processor bringup must
540 * be written out to main store before we release
546 DBG("smp: kicking cpu %d\n", cpu
);
547 rc
= smp_ops
->kick_cpu(cpu
);
549 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu
, rc
);
554 * wait to see if the cpu made a callin (is actually up).
555 * use this value that I found through experimentation.
558 if (system_state
< SYSTEM_RUNNING
)
559 for (c
= 50000; c
&& !cpu_callin_map
[cpu
]; c
--)
561 #ifdef CONFIG_HOTPLUG_CPU
564 * CPUs can take much longer to come up in the
565 * hotplug case. Wait five seconds.
567 for (c
= 5000; c
&& !cpu_callin_map
[cpu
]; c
--)
571 if (!cpu_callin_map
[cpu
]) {
572 printk(KERN_ERR
"Processor %u is stuck.\n", cpu
);
576 DBG("Processor %u found.\n", cpu
);
578 if (smp_ops
->give_timebase
)
579 smp_ops
->give_timebase();
581 /* Wait until cpu puts itself in the online & active maps */
582 while (!cpu_online(cpu
))
588 /* Return the value of the reg property corresponding to the given
591 int cpu_to_core_id(int cpu
)
593 struct device_node
*np
;
597 np
= of_get_cpu_node(cpu
, NULL
);
601 reg
= of_get_property(np
, "reg", NULL
);
605 id
= be32_to_cpup(reg
);
610 EXPORT_SYMBOL_GPL(cpu_to_core_id
);
612 /* Helper routines for cpu to core mapping */
613 int cpu_core_index_of_thread(int cpu
)
615 return cpu
>> threads_shift
;
617 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread
);
619 int cpu_first_thread_of_core(int core
)
621 return core
<< threads_shift
;
623 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core
);
625 static void traverse_siblings_chip_id(int cpu
, bool add
, int chipid
)
627 const struct cpumask
*mask
;
628 struct device_node
*np
;
632 mask
= add
? cpu_online_mask
: cpu_present_mask
;
633 for_each_cpu(i
, mask
) {
634 np
= of_get_cpu_node(i
, NULL
);
637 prop
= of_get_property(np
, "ibm,chip-id", &plen
);
638 if (prop
&& plen
== sizeof(int) &&
639 of_read_number(prop
, 1) == chipid
) {
641 cpumask_set_cpu(cpu
, cpu_core_mask(i
));
642 cpumask_set_cpu(i
, cpu_core_mask(cpu
));
644 cpumask_clear_cpu(cpu
, cpu_core_mask(i
));
645 cpumask_clear_cpu(i
, cpu_core_mask(cpu
));
652 /* Must be called when no change can occur to cpu_present_mask,
653 * i.e. during cpu online or offline.
655 static struct device_node
*cpu_to_l2cache(int cpu
)
657 struct device_node
*np
;
658 struct device_node
*cache
;
660 if (!cpu_present(cpu
))
663 np
= of_get_cpu_node(cpu
, NULL
);
667 cache
= of_find_next_cache_node(np
);
674 static void traverse_core_siblings(int cpu
, bool add
)
676 struct device_node
*l2_cache
, *np
;
677 const struct cpumask
*mask
;
681 /* First see if we have ibm,chip-id properties in cpu nodes */
682 np
= of_get_cpu_node(cpu
, NULL
);
685 prop
= of_get_property(np
, "ibm,chip-id", &plen
);
686 if (prop
&& plen
== sizeof(int))
687 chip
= of_read_number(prop
, 1);
690 traverse_siblings_chip_id(cpu
, add
, chip
);
695 l2_cache
= cpu_to_l2cache(cpu
);
696 mask
= add
? cpu_online_mask
: cpu_present_mask
;
697 for_each_cpu(i
, mask
) {
698 np
= cpu_to_l2cache(i
);
701 if (np
== l2_cache
) {
703 cpumask_set_cpu(cpu
, cpu_core_mask(i
));
704 cpumask_set_cpu(i
, cpu_core_mask(cpu
));
706 cpumask_clear_cpu(cpu
, cpu_core_mask(i
));
707 cpumask_clear_cpu(i
, cpu_core_mask(cpu
));
712 of_node_put(l2_cache
);
715 /* Activate a secondary processor. */
716 void start_secondary(void *unused
)
718 unsigned int cpu
= smp_processor_id();
722 current
->active_mm
= &init_mm
;
724 smp_store_cpu_info(cpu
);
725 set_dec(tb_ticks_per_jiffy
);
727 cpu_callin_map
[cpu
] = 1;
729 if (smp_ops
->setup_cpu
)
730 smp_ops
->setup_cpu(cpu
);
731 if (smp_ops
->take_timebase
)
732 smp_ops
->take_timebase();
734 secondary_cpu_time_init();
737 if (system_state
== SYSTEM_RUNNING
)
738 vdso_data
->processorCount
++;
742 /* Update sibling maps */
743 base
= cpu_first_thread_sibling(cpu
);
744 for (i
= 0; i
< threads_per_core
; i
++) {
745 if (cpu_is_offline(base
+ i
) && (cpu
!= base
+ i
))
747 cpumask_set_cpu(cpu
, cpu_sibling_mask(base
+ i
));
748 cpumask_set_cpu(base
+ i
, cpu_sibling_mask(cpu
));
750 /* cpu_core_map should be a superset of
751 * cpu_sibling_map even if we don't have cache
752 * information, so update the former here, too.
754 cpumask_set_cpu(cpu
, cpu_core_mask(base
+ i
));
755 cpumask_set_cpu(base
+ i
, cpu_core_mask(cpu
));
757 traverse_core_siblings(cpu
, true);
759 set_numa_node(numa_cpu_lookup_table
[cpu
]);
760 set_numa_mem(local_memory_node(numa_cpu_lookup_table
[cpu
]));
763 notify_cpu_starting(cpu
);
764 set_cpu_online(cpu
, true);
768 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE
);
773 int setup_profiling_timer(unsigned int multiplier
)
778 #ifdef CONFIG_SCHED_SMT
779 /* cpumask of CPUs with asymetric SMT dependancy */
780 static int powerpc_smt_flags(void)
782 int flags
= SD_SHARE_CPUCAPACITY
| SD_SHARE_PKG_RESOURCES
;
784 if (cpu_has_feature(CPU_FTR_ASYM_SMT
)) {
785 printk_once(KERN_INFO
"Enabling Asymmetric SMT scheduling\n");
786 flags
|= SD_ASYM_PACKING
;
792 static struct sched_domain_topology_level powerpc_topology
[] = {
793 #ifdef CONFIG_SCHED_SMT
794 { cpu_smt_mask
, powerpc_smt_flags
, SD_INIT_NAME(SMT
) },
796 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
800 void __init
smp_cpus_done(unsigned int max_cpus
)
802 cpumask_var_t old_mask
;
804 /* We want the setup_cpu() here to be called from CPU 0, but our
805 * init thread may have been "borrowed" by another CPU in the meantime
806 * se we pin us down to CPU 0 for a short while
808 alloc_cpumask_var(&old_mask
, GFP_NOWAIT
);
809 cpumask_copy(old_mask
, ¤t
->cpus_allowed
);
810 set_cpus_allowed_ptr(current
, cpumask_of(boot_cpuid
));
812 if (smp_ops
&& smp_ops
->setup_cpu
)
813 smp_ops
->setup_cpu(boot_cpuid
);
815 set_cpus_allowed_ptr(current
, old_mask
);
817 free_cpumask_var(old_mask
);
819 if (smp_ops
&& smp_ops
->bringup_done
)
820 smp_ops
->bringup_done();
822 dump_numa_cpu_topology();
824 set_sched_topology(powerpc_topology
);
828 #ifdef CONFIG_HOTPLUG_CPU
829 int __cpu_disable(void)
831 int cpu
= smp_processor_id();
835 if (!smp_ops
->cpu_disable
)
838 err
= smp_ops
->cpu_disable();
842 /* Update sibling maps */
843 base
= cpu_first_thread_sibling(cpu
);
844 for (i
= 0; i
< threads_per_core
&& base
+ i
< nr_cpu_ids
; i
++) {
845 cpumask_clear_cpu(cpu
, cpu_sibling_mask(base
+ i
));
846 cpumask_clear_cpu(base
+ i
, cpu_sibling_mask(cpu
));
847 cpumask_clear_cpu(cpu
, cpu_core_mask(base
+ i
));
848 cpumask_clear_cpu(base
+ i
, cpu_core_mask(cpu
));
850 traverse_core_siblings(cpu
, false);
855 void __cpu_die(unsigned int cpu
)
857 if (smp_ops
->cpu_die
)
858 smp_ops
->cpu_die(cpu
);
866 /* If we return, we re-enter start_secondary */
867 start_secondary_resume();