xref: /linux/arch/powerpc/kernel/smp.c (revision 05dc8c02bf40090e9ed23932b1980ead48eb8870)
1 /*
2  * SMP support for ppc.
3  *
4  * Written by Cort Dougan (cort@cs.nmt.edu) borrowing a great
5  * deal of code from the sparc and intel versions.
6  *
7  * Copyright (C) 1999 Cort Dougan <cort@cs.nmt.edu>
8  *
9  * PowerPC-64 Support added by Dave Engebretsen, Peter Bergner, and
10  * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
11  *
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.
16  */
17 
18 #undef DEBUG
19 
20 #include <linux/kernel.h>
21 #include <linux/module.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/sysdev.h>
31 #include <linux/cpu.h>
32 #include <linux/notifier.h>
33 #include <linux/topology.h>
34 
35 #include <asm/ptrace.h>
36 #include <asm/atomic.h>
37 #include <asm/irq.h>
38 #include <asm/page.h>
39 #include <asm/pgtable.h>
40 #include <asm/prom.h>
41 #include <asm/smp.h>
42 #include <asm/time.h>
43 #include <asm/machdep.h>
44 #include <asm/cputable.h>
45 #include <asm/system.h>
46 #include <asm/mpic.h>
47 #include <asm/vdso_datapage.h>
48 #ifdef CONFIG_PPC64
49 #include <asm/paca.h>
50 #endif
51 
52 #ifdef DEBUG
53 #include <asm/udbg.h>
54 #define DBG(fmt...) udbg_printf(fmt)
55 #else
56 #define DBG(fmt...)
57 #endif
58 
59 int smp_hw_index[NR_CPUS];
60 struct thread_info *secondary_ti;
61 
62 cpumask_t cpu_possible_map = CPU_MASK_NONE;
63 cpumask_t cpu_online_map = CPU_MASK_NONE;
64 cpumask_t cpu_sibling_map[NR_CPUS] = { [0 ... NR_CPUS-1] = CPU_MASK_NONE };
65 
66 EXPORT_SYMBOL(cpu_online_map);
67 EXPORT_SYMBOL(cpu_possible_map);
68 EXPORT_SYMBOL(cpu_sibling_map);
69 
70 /* SMP operations for this machine */
71 struct smp_ops_t *smp_ops;
72 
73 static volatile unsigned int cpu_callin_map[NR_CPUS];
74 
75 void smp_call_function_interrupt(void);
76 
77 int smt_enabled_at_boot = 1;
78 
79 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
80 
81 #ifdef CONFIG_PPC64
82 void __devinit smp_generic_kick_cpu(int nr)
83 {
84 	BUG_ON(nr < 0 || nr >= NR_CPUS);
85 
86 	/*
87 	 * The processor is currently spinning, waiting for the
88 	 * cpu_start field to become non-zero After we set cpu_start,
89 	 * the processor will continue on to secondary_start
90 	 */
91 	paca[nr].cpu_start = 1;
92 	smp_mb();
93 }
94 #endif
95 
96 void smp_message_recv(int msg)
97 {
98 	switch(msg) {
99 	case PPC_MSG_CALL_FUNCTION:
100 		smp_call_function_interrupt();
101 		break;
102 	case PPC_MSG_RESCHEDULE:
103 		/* XXX Do we have to do this? */
104 		set_need_resched();
105 		break;
106 	case PPC_MSG_DEBUGGER_BREAK:
107 		if (crash_ipi_function_ptr) {
108 			crash_ipi_function_ptr(get_irq_regs());
109 			break;
110 		}
111 #ifdef CONFIG_DEBUGGER
112 		debugger_ipi(get_irq_regs());
113 		break;
114 #endif /* CONFIG_DEBUGGER */
115 		/* FALLTHROUGH */
116 	default:
117 		printk("SMP %d: smp_message_recv(): unknown msg %d\n",
118 		       smp_processor_id(), msg);
119 		break;
120 	}
121 }
122 
123 void smp_send_reschedule(int cpu)
124 {
125 	if (likely(smp_ops))
126 		smp_ops->message_pass(cpu, PPC_MSG_RESCHEDULE);
127 }
128 
129 #ifdef CONFIG_DEBUGGER
130 void smp_send_debugger_break(int cpu)
131 {
132 	if (likely(smp_ops))
133 		smp_ops->message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
134 }
135 #endif
136 
137 #ifdef CONFIG_KEXEC
138 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
139 {
140 	crash_ipi_function_ptr = crash_ipi_callback;
141 	if (crash_ipi_callback && smp_ops) {
142 		mb();
143 		smp_ops->message_pass(MSG_ALL_BUT_SELF, PPC_MSG_DEBUGGER_BREAK);
144 	}
145 }
146 #endif
147 
148 static void stop_this_cpu(void *dummy)
149 {
150 	local_irq_disable();
151 	while (1)
152 		;
153 }
154 
155 void smp_send_stop(void)
156 {
157 	smp_call_function(stop_this_cpu, NULL, 1, 0);
158 }
159 
160 /*
161  * Structure and data for smp_call_function(). This is designed to minimise
162  * static memory requirements. It also looks cleaner.
163  * Stolen from the i386 version.
164  */
165 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(call_lock);
166 
167 static struct call_data_struct {
168 	void (*func) (void *info);
169 	void *info;
170 	atomic_t started;
171 	atomic_t finished;
172 	int wait;
173 } *call_data;
174 
175 /* delay of at least 8 seconds */
176 #define SMP_CALL_TIMEOUT	8
177 
178 /*
179  * These functions send a 'generic call function' IPI to other online
180  * CPUS in the system.
181  *
182  * [SUMMARY] Run a function on other CPUs.
183  * <func> The function to run. This must be fast and non-blocking.
184  * <info> An arbitrary pointer to pass to the function.
185  * <nonatomic> currently unused.
186  * <wait> If true, wait (atomically) until function has completed on other CPUs.
187  * [RETURNS] 0 on success, else a negative status code. Does not return until
188  * remote CPUs are nearly ready to execute <<func>> or are or have executed.
189  *
190  * You must not call this function with disabled interrupts or from a
191  * hardware interrupt handler or from a bottom half handler.
192  */
193 int smp_call_function_map(void (*func) (void *info), void *info, int nonatomic,
194 			int wait, cpumask_t map)
195 {
196 	struct call_data_struct data;
197 	int ret = -1, num_cpus;
198 	int cpu;
199 	u64 timeout;
200 
201 	/* Can deadlock when called with interrupts disabled */
202 	WARN_ON(irqs_disabled());
203 
204 	if (unlikely(smp_ops == NULL))
205 		return ret;
206 
207 	data.func = func;
208 	data.info = info;
209 	atomic_set(&data.started, 0);
210 	data.wait = wait;
211 	if (wait)
212 		atomic_set(&data.finished, 0);
213 
214 	spin_lock(&call_lock);
215 
216 	/* remove 'self' from the map */
217 	if (cpu_isset(smp_processor_id(), map))
218 		cpu_clear(smp_processor_id(), map);
219 
220 	/* sanity check the map, remove any non-online processors. */
221 	cpus_and(map, map, cpu_online_map);
222 
223 	num_cpus = cpus_weight(map);
224 	if (!num_cpus)
225 		goto done;
226 
227 	call_data = &data;
228 	smp_wmb();
229 	/* Send a message to all CPUs in the map */
230 	for_each_cpu_mask(cpu, map)
231 		smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNCTION);
232 
233 	timeout = get_tb() + (u64) SMP_CALL_TIMEOUT * tb_ticks_per_sec;
234 
235 	/* Wait for indication that they have received the message */
236 	while (atomic_read(&data.started) != num_cpus) {
237 		HMT_low();
238 		if (get_tb() >= timeout) {
239 			printk("smp_call_function on cpu %d: other cpus not "
240 				"responding (%d)\n", smp_processor_id(),
241 				atomic_read(&data.started));
242 			debugger(NULL);
243 			goto out;
244 		}
245 	}
246 
247 	/* optionally wait for the CPUs to complete */
248 	if (wait) {
249 		while (atomic_read(&data.finished) != num_cpus) {
250 			HMT_low();
251 			if (get_tb() >= timeout) {
252 				printk("smp_call_function on cpu %d: other "
253 					"cpus not finishing (%d/%d)\n",
254 					smp_processor_id(),
255 					atomic_read(&data.finished),
256 					atomic_read(&data.started));
257 				debugger(NULL);
258 				goto out;
259 			}
260 		}
261 	}
262 
263  done:
264 	ret = 0;
265 
266  out:
267 	call_data = NULL;
268 	HMT_medium();
269 	spin_unlock(&call_lock);
270 	return ret;
271 }
272 
273 int smp_call_function(void (*func) (void *info), void *info, int nonatomic,
274 			int wait)
275 {
276 	return smp_call_function_map(func,info,nonatomic,wait,cpu_online_map);
277 }
278 EXPORT_SYMBOL(smp_call_function);
279 
280 int smp_call_function_single(int cpu, void (*func) (void *info), void *info, int nonatomic,
281 			int wait)
282 {
283 	cpumask_t map = CPU_MASK_NONE;
284 	int ret = 0;
285 
286 	if (!cpu_online(cpu))
287 		return -EINVAL;
288 
289 	cpu_set(cpu, map);
290 	if (cpu != get_cpu())
291 		ret = smp_call_function_map(func,info,nonatomic,wait,map);
292 	else {
293 		local_irq_disable();
294 		func(info);
295 		local_irq_enable();
296 	}
297 	put_cpu();
298 	return ret;
299 }
300 EXPORT_SYMBOL(smp_call_function_single);
301 
302 void smp_call_function_interrupt(void)
303 {
304 	void (*func) (void *info);
305 	void *info;
306 	int wait;
307 
308 	/* call_data will be NULL if the sender timed out while
309 	 * waiting on us to receive the call.
310 	 */
311 	if (!call_data)
312 		return;
313 
314 	func = call_data->func;
315 	info = call_data->info;
316 	wait = call_data->wait;
317 
318 	if (!wait)
319 		smp_mb__before_atomic_inc();
320 
321 	/*
322 	 * Notify initiating CPU that I've grabbed the data and am
323 	 * about to execute the function
324 	 */
325 	atomic_inc(&call_data->started);
326 	/*
327 	 * At this point the info structure may be out of scope unless wait==1
328 	 */
329 	(*func)(info);
330 	if (wait) {
331 		smp_mb__before_atomic_inc();
332 		atomic_inc(&call_data->finished);
333 	}
334 }
335 
336 extern struct gettimeofday_struct do_gtod;
337 
338 struct thread_info *current_set[NR_CPUS];
339 
340 DECLARE_PER_CPU(unsigned int, pvr);
341 
342 static void __devinit smp_store_cpu_info(int id)
343 {
344 	per_cpu(pvr, id) = mfspr(SPRN_PVR);
345 }
346 
347 static void __init smp_create_idle(unsigned int cpu)
348 {
349 	struct task_struct *p;
350 
351 	/* create a process for the processor */
352 	p = fork_idle(cpu);
353 	if (IS_ERR(p))
354 		panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p));
355 #ifdef CONFIG_PPC64
356 	paca[cpu].__current = p;
357 #endif
358 	current_set[cpu] = task_thread_info(p);
359 	task_thread_info(p)->cpu = cpu;
360 }
361 
362 void __init smp_prepare_cpus(unsigned int max_cpus)
363 {
364 	unsigned int cpu;
365 
366 	DBG("smp_prepare_cpus\n");
367 
368 	/*
369 	 * setup_cpu may need to be called on the boot cpu. We havent
370 	 * spun any cpus up but lets be paranoid.
371 	 */
372 	BUG_ON(boot_cpuid != smp_processor_id());
373 
374 	/* Fixup boot cpu */
375 	smp_store_cpu_info(boot_cpuid);
376 	cpu_callin_map[boot_cpuid] = 1;
377 
378 	if (smp_ops)
379 		max_cpus = smp_ops->probe();
380 	else
381 		max_cpus = 1;
382 
383 	smp_space_timers(max_cpus);
384 
385 	for_each_possible_cpu(cpu)
386 		if (cpu != boot_cpuid)
387 			smp_create_idle(cpu);
388 }
389 
390 void __devinit smp_prepare_boot_cpu(void)
391 {
392 	BUG_ON(smp_processor_id() != boot_cpuid);
393 
394 	cpu_set(boot_cpuid, cpu_online_map);
395 #ifdef CONFIG_PPC64
396 	paca[boot_cpuid].__current = current;
397 #endif
398 	current_set[boot_cpuid] = task_thread_info(current);
399 }
400 
401 #ifdef CONFIG_HOTPLUG_CPU
402 /* State of each CPU during hotplug phases */
403 DEFINE_PER_CPU(int, cpu_state) = { 0 };
404 
405 int generic_cpu_disable(void)
406 {
407 	unsigned int cpu = smp_processor_id();
408 
409 	if (cpu == boot_cpuid)
410 		return -EBUSY;
411 
412 	cpu_clear(cpu, cpu_online_map);
413 #ifdef CONFIG_PPC64
414 	vdso_data->processorCount--;
415 	fixup_irqs(cpu_online_map);
416 #endif
417 	return 0;
418 }
419 
420 int generic_cpu_enable(unsigned int cpu)
421 {
422 	/* Do the normal bootup if we haven't
423 	 * already bootstrapped. */
424 	if (system_state != SYSTEM_RUNNING)
425 		return -ENOSYS;
426 
427 	/* get the target out of it's holding state */
428 	per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
429 	smp_wmb();
430 
431 	while (!cpu_online(cpu))
432 		cpu_relax();
433 
434 #ifdef CONFIG_PPC64
435 	fixup_irqs(cpu_online_map);
436 	/* counter the irq disable in fixup_irqs */
437 	local_irq_enable();
438 #endif
439 	return 0;
440 }
441 
442 void generic_cpu_die(unsigned int cpu)
443 {
444 	int i;
445 
446 	for (i = 0; i < 100; i++) {
447 		smp_rmb();
448 		if (per_cpu(cpu_state, cpu) == CPU_DEAD)
449 			return;
450 		msleep(100);
451 	}
452 	printk(KERN_ERR "CPU%d didn't die...\n", cpu);
453 }
454 
455 void generic_mach_cpu_die(void)
456 {
457 	unsigned int cpu;
458 
459 	local_irq_disable();
460 	cpu = smp_processor_id();
461 	printk(KERN_DEBUG "CPU%d offline\n", cpu);
462 	__get_cpu_var(cpu_state) = CPU_DEAD;
463 	smp_wmb();
464 	while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
465 		cpu_relax();
466 	cpu_set(cpu, cpu_online_map);
467 	local_irq_enable();
468 }
469 #endif
470 
471 static int __devinit cpu_enable(unsigned int cpu)
472 {
473 	if (smp_ops && smp_ops->cpu_enable)
474 		return smp_ops->cpu_enable(cpu);
475 
476 	return -ENOSYS;
477 }
478 
479 int __cpuinit __cpu_up(unsigned int cpu)
480 {
481 	int c;
482 
483 	secondary_ti = current_set[cpu];
484 	if (!cpu_enable(cpu))
485 		return 0;
486 
487 	if (smp_ops == NULL ||
488 	    (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
489 		return -EINVAL;
490 
491 	/* Make sure callin-map entry is 0 (can be leftover a CPU
492 	 * hotplug
493 	 */
494 	cpu_callin_map[cpu] = 0;
495 
496 	/* The information for processor bringup must
497 	 * be written out to main store before we release
498 	 * the processor.
499 	 */
500 	smp_mb();
501 
502 	/* wake up cpus */
503 	DBG("smp: kicking cpu %d\n", cpu);
504 	smp_ops->kick_cpu(cpu);
505 
506 	/*
507 	 * wait to see if the cpu made a callin (is actually up).
508 	 * use this value that I found through experimentation.
509 	 * -- Cort
510 	 */
511 	if (system_state < SYSTEM_RUNNING)
512 		for (c = 50000; c && !cpu_callin_map[cpu]; c--)
513 			udelay(100);
514 #ifdef CONFIG_HOTPLUG_CPU
515 	else
516 		/*
517 		 * CPUs can take much longer to come up in the
518 		 * hotplug case.  Wait five seconds.
519 		 */
520 		for (c = 25; c && !cpu_callin_map[cpu]; c--) {
521 			msleep(200);
522 		}
523 #endif
524 
525 	if (!cpu_callin_map[cpu]) {
526 		printk("Processor %u is stuck.\n", cpu);
527 		return -ENOENT;
528 	}
529 
530 	printk("Processor %u found.\n", cpu);
531 
532 	if (smp_ops->give_timebase)
533 		smp_ops->give_timebase();
534 
535 	/* Wait until cpu puts itself in the online map */
536 	while (!cpu_online(cpu))
537 		cpu_relax();
538 
539 	return 0;
540 }
541 
542 
543 /* Activate a secondary processor. */
544 int __devinit start_secondary(void *unused)
545 {
546 	unsigned int cpu = smp_processor_id();
547 
548 	atomic_inc(&init_mm.mm_count);
549 	current->active_mm = &init_mm;
550 
551 	smp_store_cpu_info(cpu);
552 	set_dec(tb_ticks_per_jiffy);
553 	preempt_disable();
554 	cpu_callin_map[cpu] = 1;
555 
556 	smp_ops->setup_cpu(cpu);
557 	if (smp_ops->take_timebase)
558 		smp_ops->take_timebase();
559 
560 	if (system_state > SYSTEM_BOOTING)
561 		snapshot_timebase();
562 
563 	spin_lock(&call_lock);
564 	cpu_set(cpu, cpu_online_map);
565 	spin_unlock(&call_lock);
566 
567 	local_irq_enable();
568 
569 	cpu_idle();
570 	return 0;
571 }
572 
573 int setup_profiling_timer(unsigned int multiplier)
574 {
575 	return 0;
576 }
577 
578 void __init smp_cpus_done(unsigned int max_cpus)
579 {
580 	cpumask_t old_mask;
581 
582 	/* We want the setup_cpu() here to be called from CPU 0, but our
583 	 * init thread may have been "borrowed" by another CPU in the meantime
584 	 * se we pin us down to CPU 0 for a short while
585 	 */
586 	old_mask = current->cpus_allowed;
587 	set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid));
588 
589 	if (smp_ops)
590 		smp_ops->setup_cpu(boot_cpuid);
591 
592 	set_cpus_allowed(current, old_mask);
593 
594 	snapshot_timebases();
595 
596 	dump_numa_cpu_topology();
597 }
598 
599 #ifdef CONFIG_HOTPLUG_CPU
600 int __cpu_disable(void)
601 {
602 	if (smp_ops->cpu_disable)
603 		return smp_ops->cpu_disable();
604 
605 	return -ENOSYS;
606 }
607 
608 void __cpu_die(unsigned int cpu)
609 {
610 	if (smp_ops->cpu_die)
611 		smp_ops->cpu_die(cpu);
612 }
613 #endif
614