xref: /openbmc/linux/arch/ia64/kernel/smpboot.c (revision 87c2ce3b)
1 /*
2  * SMP boot-related support
3  *
4  * Copyright (C) 1998-2003, 2005 Hewlett-Packard Co
5  *	David Mosberger-Tang <davidm@hpl.hp.com>
6  * Copyright (C) 2001, 2004-2005 Intel Corp
7  * 	Rohit Seth <rohit.seth@intel.com>
8  * 	Suresh Siddha <suresh.b.siddha@intel.com>
9  * 	Gordon Jin <gordon.jin@intel.com>
10  *	Ashok Raj  <ashok.raj@intel.com>
11  *
12  * 01/05/16 Rohit Seth <rohit.seth@intel.com>	Moved SMP booting functions from smp.c to here.
13  * 01/04/27 David Mosberger <davidm@hpl.hp.com>	Added ITC synching code.
14  * 02/07/31 David Mosberger <davidm@hpl.hp.com>	Switch over to hotplug-CPU boot-sequence.
15  *						smp_boot_cpus()/smp_commence() is replaced by
16  *						smp_prepare_cpus()/__cpu_up()/smp_cpus_done().
17  * 04/06/21 Ashok Raj		<ashok.raj@intel.com> Added CPU Hotplug Support
18  * 04/12/26 Jin Gordon <gordon.jin@intel.com>
19  * 04/12/26 Rohit Seth <rohit.seth@intel.com>
20  *						Add multi-threading and multi-core detection
21  * 05/01/30 Suresh Siddha <suresh.b.siddha@intel.com>
22  *						Setup cpu_sibling_map and cpu_core_map
23  */
24 #include <linux/config.h>
25 
26 #include <linux/module.h>
27 #include <linux/acpi.h>
28 #include <linux/bootmem.h>
29 #include <linux/cpu.h>
30 #include <linux/delay.h>
31 #include <linux/init.h>
32 #include <linux/interrupt.h>
33 #include <linux/irq.h>
34 #include <linux/kernel.h>
35 #include <linux/kernel_stat.h>
36 #include <linux/mm.h>
37 #include <linux/notifier.h>
38 #include <linux/smp.h>
39 #include <linux/smp_lock.h>
40 #include <linux/spinlock.h>
41 #include <linux/efi.h>
42 #include <linux/percpu.h>
43 #include <linux/bitops.h>
44 
45 #include <asm/atomic.h>
46 #include <asm/cache.h>
47 #include <asm/current.h>
48 #include <asm/delay.h>
49 #include <asm/ia32.h>
50 #include <asm/io.h>
51 #include <asm/irq.h>
52 #include <asm/machvec.h>
53 #include <asm/mca.h>
54 #include <asm/page.h>
55 #include <asm/pgalloc.h>
56 #include <asm/pgtable.h>
57 #include <asm/processor.h>
58 #include <asm/ptrace.h>
59 #include <asm/sal.h>
60 #include <asm/system.h>
61 #include <asm/tlbflush.h>
62 #include <asm/unistd.h>
63 
64 #define SMP_DEBUG 0
65 
66 #if SMP_DEBUG
67 #define Dprintk(x...)  printk(x)
68 #else
69 #define Dprintk(x...)
70 #endif
71 
72 #ifdef CONFIG_HOTPLUG_CPU
73 /*
74  * Store all idle threads, this can be reused instead of creating
75  * a new thread. Also avoids complicated thread destroy functionality
76  * for idle threads.
77  */
78 struct task_struct *idle_thread_array[NR_CPUS];
79 
80 /*
81  * Global array allocated for NR_CPUS at boot time
82  */
83 struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
84 
85 /*
86  * start_ap in head.S uses this to store current booting cpu
87  * info.
88  */
89 struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
90 
91 #define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
92 
93 #define get_idle_for_cpu(x)		(idle_thread_array[(x)])
94 #define set_idle_for_cpu(x,p)	(idle_thread_array[(x)] = (p))
95 
96 #else
97 
98 #define get_idle_for_cpu(x)		(NULL)
99 #define set_idle_for_cpu(x,p)
100 #define set_brendez_area(x)
101 #endif
102 
103 
104 /*
105  * ITC synchronization related stuff:
106  */
107 #define MASTER	0
108 #define SLAVE	(SMP_CACHE_BYTES/8)
109 
110 #define NUM_ROUNDS	64	/* magic value */
111 #define NUM_ITERS	5	/* likewise */
112 
113 static DEFINE_SPINLOCK(itc_sync_lock);
114 static volatile unsigned long go[SLAVE + 1];
115 
116 #define DEBUG_ITC_SYNC	0
117 
118 extern void __devinit calibrate_delay (void);
119 extern void start_ap (void);
120 extern unsigned long ia64_iobase;
121 
122 task_t *task_for_booting_cpu;
123 
124 /*
125  * State for each CPU
126  */
127 DEFINE_PER_CPU(int, cpu_state);
128 
129 /* Bitmasks of currently online, and possible CPUs */
130 cpumask_t cpu_online_map;
131 EXPORT_SYMBOL(cpu_online_map);
132 cpumask_t cpu_possible_map;
133 EXPORT_SYMBOL(cpu_possible_map);
134 
135 cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
136 cpumask_t cpu_sibling_map[NR_CPUS] __cacheline_aligned;
137 int smp_num_siblings = 1;
138 int smp_num_cpucores = 1;
139 
140 /* which logical CPU number maps to which CPU (physical APIC ID) */
141 volatile int ia64_cpu_to_sapicid[NR_CPUS];
142 EXPORT_SYMBOL(ia64_cpu_to_sapicid);
143 
144 static volatile cpumask_t cpu_callin_map;
145 
146 struct smp_boot_data smp_boot_data __initdata;
147 
148 unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
149 
150 char __initdata no_int_routing;
151 
152 unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
153 
154 static int __init
155 nointroute (char *str)
156 {
157 	no_int_routing = 1;
158 	printk ("no_int_routing on\n");
159 	return 1;
160 }
161 
162 __setup("nointroute", nointroute);
163 
164 void
165 sync_master (void *arg)
166 {
167 	unsigned long flags, i;
168 
169 	go[MASTER] = 0;
170 
171 	local_irq_save(flags);
172 	{
173 		for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
174 			while (!go[MASTER])
175 				cpu_relax();
176 			go[MASTER] = 0;
177 			go[SLAVE] = ia64_get_itc();
178 		}
179 	}
180 	local_irq_restore(flags);
181 }
182 
183 /*
184  * Return the number of cycles by which our itc differs from the itc on the master
185  * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
186  * negative that it is behind.
187  */
188 static inline long
189 get_delta (long *rt, long *master)
190 {
191 	unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
192 	unsigned long tcenter, t0, t1, tm;
193 	long i;
194 
195 	for (i = 0; i < NUM_ITERS; ++i) {
196 		t0 = ia64_get_itc();
197 		go[MASTER] = 1;
198 		while (!(tm = go[SLAVE]))
199 			cpu_relax();
200 		go[SLAVE] = 0;
201 		t1 = ia64_get_itc();
202 
203 		if (t1 - t0 < best_t1 - best_t0)
204 			best_t0 = t0, best_t1 = t1, best_tm = tm;
205 	}
206 
207 	*rt = best_t1 - best_t0;
208 	*master = best_tm - best_t0;
209 
210 	/* average best_t0 and best_t1 without overflow: */
211 	tcenter = (best_t0/2 + best_t1/2);
212 	if (best_t0 % 2 + best_t1 % 2 == 2)
213 		++tcenter;
214 	return tcenter - best_tm;
215 }
216 
217 /*
218  * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
219  * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
220  * unaccounted-for errors (such as getting a machine check in the middle of a calibration
221  * step).  The basic idea is for the slave to ask the master what itc value it has and to
222  * read its own itc before and after the master responds.  Each iteration gives us three
223  * timestamps:
224  *
225  *	slave		master
226  *
227  *	t0 ---\
228  *             ---\
229  *		   --->
230  *			tm
231  *		   /---
232  *	       /---
233  *	t1 <---
234  *
235  *
236  * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
237  * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
238  * between the slave and the master is symmetric.  Even if the interconnect were
239  * asymmetric, we would still know that the synchronization error is smaller than the
240  * roundtrip latency (t0 - t1).
241  *
242  * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
243  * within one or two cycles.  However, we can only *guarantee* that the synchronization is
244  * accurate to within a round-trip time, which is typically in the range of several
245  * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
246  * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
247  * than half a micro second or so.
248  */
249 void
250 ia64_sync_itc (unsigned int master)
251 {
252 	long i, delta, adj, adjust_latency = 0, done = 0;
253 	unsigned long flags, rt, master_time_stamp, bound;
254 #if DEBUG_ITC_SYNC
255 	struct {
256 		long rt;	/* roundtrip time */
257 		long master;	/* master's timestamp */
258 		long diff;	/* difference between midpoint and master's timestamp */
259 		long lat;	/* estimate of itc adjustment latency */
260 	} t[NUM_ROUNDS];
261 #endif
262 
263 	/*
264 	 * Make sure local timer ticks are disabled while we sync.  If
265 	 * they were enabled, we'd have to worry about nasty issues
266 	 * like setting the ITC ahead of (or a long time before) the
267 	 * next scheduled tick.
268 	 */
269 	BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
270 
271 	go[MASTER] = 1;
272 
273 	if (smp_call_function_single(master, sync_master, NULL, 1, 0) < 0) {
274 		printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
275 		return;
276 	}
277 
278 	while (go[MASTER])
279 		cpu_relax();	/* wait for master to be ready */
280 
281 	spin_lock_irqsave(&itc_sync_lock, flags);
282 	{
283 		for (i = 0; i < NUM_ROUNDS; ++i) {
284 			delta = get_delta(&rt, &master_time_stamp);
285 			if (delta == 0) {
286 				done = 1;	/* let's lock on to this... */
287 				bound = rt;
288 			}
289 
290 			if (!done) {
291 				if (i > 0) {
292 					adjust_latency += -delta;
293 					adj = -delta + adjust_latency/4;
294 				} else
295 					adj = -delta;
296 
297 				ia64_set_itc(ia64_get_itc() + adj);
298 			}
299 #if DEBUG_ITC_SYNC
300 			t[i].rt = rt;
301 			t[i].master = master_time_stamp;
302 			t[i].diff = delta;
303 			t[i].lat = adjust_latency/4;
304 #endif
305 		}
306 	}
307 	spin_unlock_irqrestore(&itc_sync_lock, flags);
308 
309 #if DEBUG_ITC_SYNC
310 	for (i = 0; i < NUM_ROUNDS; ++i)
311 		printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
312 		       t[i].rt, t[i].master, t[i].diff, t[i].lat);
313 #endif
314 
315 	printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
316 	       "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
317 }
318 
319 /*
320  * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
321  */
322 static inline void __devinit
323 smp_setup_percpu_timer (void)
324 {
325 }
326 
327 static void __devinit
328 smp_callin (void)
329 {
330 	int cpuid, phys_id;
331 	extern void ia64_init_itm(void);
332 
333 #ifdef CONFIG_PERFMON
334 	extern void pfm_init_percpu(void);
335 #endif
336 
337 	cpuid = smp_processor_id();
338 	phys_id = hard_smp_processor_id();
339 
340 	if (cpu_online(cpuid)) {
341 		printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
342 		       phys_id, cpuid);
343 		BUG();
344 	}
345 
346 	lock_ipi_calllock();
347 	cpu_set(cpuid, cpu_online_map);
348 	unlock_ipi_calllock();
349 	per_cpu(cpu_state, cpuid) = CPU_ONLINE;
350 
351 	smp_setup_percpu_timer();
352 
353 	ia64_mca_cmc_vector_setup();	/* Setup vector on AP */
354 
355 #ifdef CONFIG_PERFMON
356 	pfm_init_percpu();
357 #endif
358 
359 	local_irq_enable();
360 
361 	if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
362 		/*
363 		 * Synchronize the ITC with the BP.  Need to do this after irqs are
364 		 * enabled because ia64_sync_itc() calls smp_call_function_single(), which
365 		 * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
366 		 * local_bh_enable(), which bugs out if irqs are not enabled...
367 		 */
368 		Dprintk("Going to syncup ITC with BP.\n");
369 		ia64_sync_itc(0);
370 	}
371 
372 	/*
373 	 * Get our bogomips.
374 	 */
375 	ia64_init_itm();
376 	calibrate_delay();
377 	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
378 
379 #ifdef CONFIG_IA32_SUPPORT
380 	ia32_gdt_init();
381 #endif
382 
383 	/*
384 	 * Allow the master to continue.
385 	 */
386 	cpu_set(cpuid, cpu_callin_map);
387 	Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
388 }
389 
390 
391 /*
392  * Activate a secondary processor.  head.S calls this.
393  */
394 int __devinit
395 start_secondary (void *unused)
396 {
397 	/* Early console may use I/O ports */
398 	ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
399 	Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
400 	efi_map_pal_code();
401 	cpu_init();
402 	preempt_disable();
403 	smp_callin();
404 
405 	cpu_idle();
406 	return 0;
407 }
408 
409 struct pt_regs * __devinit idle_regs(struct pt_regs *regs)
410 {
411 	return NULL;
412 }
413 
414 struct create_idle {
415 	struct task_struct *idle;
416 	struct completion done;
417 	int cpu;
418 };
419 
420 void
421 do_fork_idle(void *_c_idle)
422 {
423 	struct create_idle *c_idle = _c_idle;
424 
425 	c_idle->idle = fork_idle(c_idle->cpu);
426 	complete(&c_idle->done);
427 }
428 
429 static int __devinit
430 do_boot_cpu (int sapicid, int cpu)
431 {
432 	int timeout;
433 	struct create_idle c_idle = {
434 		.cpu	= cpu,
435 		.done	= COMPLETION_INITIALIZER(c_idle.done),
436 	};
437 	DECLARE_WORK(work, do_fork_idle, &c_idle);
438 
439  	c_idle.idle = get_idle_for_cpu(cpu);
440  	if (c_idle.idle) {
441 		init_idle(c_idle.idle, cpu);
442  		goto do_rest;
443 	}
444 
445 	/*
446 	 * We can't use kernel_thread since we must avoid to reschedule the child.
447 	 */
448 	if (!keventd_up() || current_is_keventd())
449 		work.func(work.data);
450 	else {
451 		schedule_work(&work);
452 		wait_for_completion(&c_idle.done);
453 	}
454 
455 	if (IS_ERR(c_idle.idle))
456 		panic("failed fork for CPU %d", cpu);
457 
458 	set_idle_for_cpu(cpu, c_idle.idle);
459 
460 do_rest:
461 	task_for_booting_cpu = c_idle.idle;
462 
463 	Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
464 
465 	set_brendez_area(cpu);
466 	platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
467 
468 	/*
469 	 * Wait 10s total for the AP to start
470 	 */
471 	Dprintk("Waiting on callin_map ...");
472 	for (timeout = 0; timeout < 100000; timeout++) {
473 		if (cpu_isset(cpu, cpu_callin_map))
474 			break;  /* It has booted */
475 		udelay(100);
476 	}
477 	Dprintk("\n");
478 
479 	if (!cpu_isset(cpu, cpu_callin_map)) {
480 		printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
481 		ia64_cpu_to_sapicid[cpu] = -1;
482 		cpu_clear(cpu, cpu_online_map);  /* was set in smp_callin() */
483 		return -EINVAL;
484 	}
485 	return 0;
486 }
487 
488 static int __init
489 decay (char *str)
490 {
491 	int ticks;
492 	get_option (&str, &ticks);
493 	return 1;
494 }
495 
496 __setup("decay=", decay);
497 
498 /*
499  * Initialize the logical CPU number to SAPICID mapping
500  */
501 void __init
502 smp_build_cpu_map (void)
503 {
504 	int sapicid, cpu, i;
505 	int boot_cpu_id = hard_smp_processor_id();
506 
507 	for (cpu = 0; cpu < NR_CPUS; cpu++) {
508 		ia64_cpu_to_sapicid[cpu] = -1;
509 #ifdef CONFIG_HOTPLUG_CPU
510 		cpu_set(cpu, cpu_possible_map);
511 #endif
512 	}
513 
514 	ia64_cpu_to_sapicid[0] = boot_cpu_id;
515 	cpus_clear(cpu_present_map);
516 	cpu_set(0, cpu_present_map);
517 	cpu_set(0, cpu_possible_map);
518 	for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
519 		sapicid = smp_boot_data.cpu_phys_id[i];
520 		if (sapicid == boot_cpu_id)
521 			continue;
522 		cpu_set(cpu, cpu_present_map);
523 		cpu_set(cpu, cpu_possible_map);
524 		ia64_cpu_to_sapicid[cpu] = sapicid;
525 		cpu++;
526 	}
527 }
528 
529 /*
530  * Cycle through the APs sending Wakeup IPIs to boot each.
531  */
532 void __init
533 smp_prepare_cpus (unsigned int max_cpus)
534 {
535 	int boot_cpu_id = hard_smp_processor_id();
536 
537 	/*
538 	 * Initialize the per-CPU profiling counter/multiplier
539 	 */
540 
541 	smp_setup_percpu_timer();
542 
543 	/*
544 	 * We have the boot CPU online for sure.
545 	 */
546 	cpu_set(0, cpu_online_map);
547 	cpu_set(0, cpu_callin_map);
548 
549 	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
550 	ia64_cpu_to_sapicid[0] = boot_cpu_id;
551 
552 	printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
553 
554 	current_thread_info()->cpu = 0;
555 
556 	/*
557 	 * If SMP should be disabled, then really disable it!
558 	 */
559 	if (!max_cpus) {
560 		printk(KERN_INFO "SMP mode deactivated.\n");
561 		cpus_clear(cpu_online_map);
562 		cpus_clear(cpu_present_map);
563 		cpus_clear(cpu_possible_map);
564 		cpu_set(0, cpu_online_map);
565 		cpu_set(0, cpu_present_map);
566 		cpu_set(0, cpu_possible_map);
567 		return;
568 	}
569 }
570 
571 void __devinit smp_prepare_boot_cpu(void)
572 {
573 	cpu_set(smp_processor_id(), cpu_online_map);
574 	cpu_set(smp_processor_id(), cpu_callin_map);
575 	per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
576 }
577 
578 /*
579  * mt_info[] is a temporary store for all info returned by
580  * PAL_LOGICAL_TO_PHYSICAL, to be copied into cpuinfo_ia64 when the
581  * specific cpu comes.
582  */
583 static struct {
584 	__u32   socket_id;
585 	__u16   core_id;
586 	__u16   thread_id;
587 	__u16   proc_fixed_addr;
588 	__u8    valid;
589 } mt_info[NR_CPUS] __devinitdata;
590 
591 #ifdef CONFIG_HOTPLUG_CPU
592 static inline void
593 remove_from_mtinfo(int cpu)
594 {
595 	int i;
596 
597 	for_each_cpu(i)
598 		if (mt_info[i].valid &&  mt_info[i].socket_id ==
599 		    				cpu_data(cpu)->socket_id)
600 			mt_info[i].valid = 0;
601 }
602 
603 static inline void
604 clear_cpu_sibling_map(int cpu)
605 {
606 	int i;
607 
608 	for_each_cpu_mask(i, cpu_sibling_map[cpu])
609 		cpu_clear(cpu, cpu_sibling_map[i]);
610 	for_each_cpu_mask(i, cpu_core_map[cpu])
611 		cpu_clear(cpu, cpu_core_map[i]);
612 
613 	cpu_sibling_map[cpu] = cpu_core_map[cpu] = CPU_MASK_NONE;
614 }
615 
616 static void
617 remove_siblinginfo(int cpu)
618 {
619 	int last = 0;
620 
621 	if (cpu_data(cpu)->threads_per_core == 1 &&
622 	    cpu_data(cpu)->cores_per_socket == 1) {
623 		cpu_clear(cpu, cpu_core_map[cpu]);
624 		cpu_clear(cpu, cpu_sibling_map[cpu]);
625 		return;
626 	}
627 
628 	last = (cpus_weight(cpu_core_map[cpu]) == 1 ? 1 : 0);
629 
630 	/* remove it from all sibling map's */
631 	clear_cpu_sibling_map(cpu);
632 
633 	/* if this cpu is the last in the core group, remove all its info
634 	 * from mt_info structure
635 	 */
636 	if (last)
637 		remove_from_mtinfo(cpu);
638 }
639 
640 extern void fixup_irqs(void);
641 /* must be called with cpucontrol mutex held */
642 int __cpu_disable(void)
643 {
644 	int cpu = smp_processor_id();
645 
646 	/*
647 	 * dont permit boot processor for now
648 	 */
649 	if (cpu == 0)
650 		return -EBUSY;
651 
652 	remove_siblinginfo(cpu);
653 	cpu_clear(cpu, cpu_online_map);
654 	fixup_irqs();
655 	local_flush_tlb_all();
656 	cpu_clear(cpu, cpu_callin_map);
657 	return 0;
658 }
659 
660 void __cpu_die(unsigned int cpu)
661 {
662 	unsigned int i;
663 
664 	for (i = 0; i < 100; i++) {
665 		/* They ack this in play_dead by setting CPU_DEAD */
666 		if (per_cpu(cpu_state, cpu) == CPU_DEAD)
667 		{
668 			printk ("CPU %d is now offline\n", cpu);
669 			return;
670 		}
671 		msleep(100);
672 	}
673  	printk(KERN_ERR "CPU %u didn't die...\n", cpu);
674 }
675 #else /* !CONFIG_HOTPLUG_CPU */
676 int __cpu_disable(void)
677 {
678 	return -ENOSYS;
679 }
680 
681 void __cpu_die(unsigned int cpu)
682 {
683 	/* We said "no" in __cpu_disable */
684 	BUG();
685 }
686 #endif /* CONFIG_HOTPLUG_CPU */
687 
688 void
689 smp_cpus_done (unsigned int dummy)
690 {
691 	int cpu;
692 	unsigned long bogosum = 0;
693 
694 	/*
695 	 * Allow the user to impress friends.
696 	 */
697 
698 	for_each_online_cpu(cpu) {
699 		bogosum += cpu_data(cpu)->loops_per_jiffy;
700 	}
701 
702 	printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
703 	       (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
704 }
705 
706 static inline void __devinit
707 set_cpu_sibling_map(int cpu)
708 {
709 	int i;
710 
711 	for_each_online_cpu(i) {
712 		if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
713 			cpu_set(i, cpu_core_map[cpu]);
714 			cpu_set(cpu, cpu_core_map[i]);
715 			if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
716 				cpu_set(i, cpu_sibling_map[cpu]);
717 				cpu_set(cpu, cpu_sibling_map[i]);
718 			}
719 		}
720 	}
721 }
722 
723 int __devinit
724 __cpu_up (unsigned int cpu)
725 {
726 	int ret;
727 	int sapicid;
728 
729 	sapicid = ia64_cpu_to_sapicid[cpu];
730 	if (sapicid == -1)
731 		return -EINVAL;
732 
733 	/*
734 	 * Already booted cpu? not valid anymore since we dont
735 	 * do idle loop tightspin anymore.
736 	 */
737 	if (cpu_isset(cpu, cpu_callin_map))
738 		return -EINVAL;
739 
740 	per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
741 	/* Processor goes to start_secondary(), sets online flag */
742 	ret = do_boot_cpu(sapicid, cpu);
743 	if (ret < 0)
744 		return ret;
745 
746 	if (cpu_data(cpu)->threads_per_core == 1 &&
747 	    cpu_data(cpu)->cores_per_socket == 1) {
748 		cpu_set(cpu, cpu_sibling_map[cpu]);
749 		cpu_set(cpu, cpu_core_map[cpu]);
750 		return 0;
751 	}
752 
753 	set_cpu_sibling_map(cpu);
754 
755 	return 0;
756 }
757 
758 /*
759  * Assume that CPU's have been discovered by some platform-dependent interface.  For
760  * SoftSDV/Lion, that would be ACPI.
761  *
762  * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
763  */
764 void __init
765 init_smp_config(void)
766 {
767 	struct fptr {
768 		unsigned long fp;
769 		unsigned long gp;
770 	} *ap_startup;
771 	long sal_ret;
772 
773 	/* Tell SAL where to drop the AP's.  */
774 	ap_startup = (struct fptr *) start_ap;
775 	sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
776 				       ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
777 	if (sal_ret < 0)
778 		printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
779 		       ia64_sal_strerror(sal_ret));
780 }
781 
782 static inline int __devinit
783 check_for_mtinfo_index(void)
784 {
785 	int i;
786 
787 	for_each_cpu(i)
788 		if (!mt_info[i].valid)
789 			return i;
790 
791 	return -1;
792 }
793 
794 /*
795  * Search the mt_info to find out if this socket's cid/tid information is
796  * cached or not. If the socket exists, fill in the core_id and thread_id
797  * in cpuinfo
798  */
799 static int __devinit
800 check_for_new_socket(__u16 logical_address, struct cpuinfo_ia64 *c)
801 {
802 	int i;
803 	__u32 sid = c->socket_id;
804 
805 	for_each_cpu(i) {
806 		if (mt_info[i].valid && mt_info[i].proc_fixed_addr == logical_address
807 		    && mt_info[i].socket_id == sid) {
808 			c->core_id = mt_info[i].core_id;
809 			c->thread_id = mt_info[i].thread_id;
810 			return 1; /* not a new socket */
811 		}
812 	}
813 	return 0;
814 }
815 
816 /*
817  * identify_siblings(cpu) gets called from identify_cpu. This populates the
818  * information related to logical execution units in per_cpu_data structure.
819  */
820 void __devinit
821 identify_siblings(struct cpuinfo_ia64 *c)
822 {
823 	s64 status;
824 	u16 pltid;
825 	u64 proc_fixed_addr;
826 	int count, i;
827 	pal_logical_to_physical_t info;
828 
829 	if (smp_num_cpucores == 1 && smp_num_siblings == 1)
830 		return;
831 
832 	if ((status = ia64_pal_logical_to_phys(0, &info)) != PAL_STATUS_SUCCESS) {
833 		printk(KERN_ERR "ia64_pal_logical_to_phys failed with %ld\n",
834 		       status);
835 		return;
836 	}
837 	if ((status = ia64_sal_physical_id_info(&pltid)) != PAL_STATUS_SUCCESS) {
838 		printk(KERN_ERR "ia64_sal_pltid failed with %ld\n", status);
839 		return;
840 	}
841 	if ((status = ia64_pal_fixed_addr(&proc_fixed_addr)) != PAL_STATUS_SUCCESS) {
842 		printk(KERN_ERR "ia64_pal_fixed_addr failed with %ld\n", status);
843 		return;
844 	}
845 
846 	c->socket_id =  (pltid << 8) | info.overview_ppid;
847 	c->cores_per_socket = info.overview_cpp;
848 	c->threads_per_core = info.overview_tpc;
849 	count = c->num_log = info.overview_num_log;
850 
851 	/* If the thread and core id information is already cached, then
852 	 * we will simply update cpu_info and return. Otherwise, we will
853 	 * do the PAL calls and cache core and thread id's of all the siblings.
854 	 */
855 	if (check_for_new_socket(proc_fixed_addr, c))
856 		return;
857 
858 	for (i = 0; i < count; i++) {
859 		int index;
860 
861 		if (i && (status = ia64_pal_logical_to_phys(i, &info))
862 			  != PAL_STATUS_SUCCESS) {
863                 	printk(KERN_ERR "ia64_pal_logical_to_phys failed"
864 					" with %ld\n", status);
865                 	return;
866 		}
867 		if (info.log2_la == proc_fixed_addr) {
868 			c->core_id = info.log1_cid;
869 			c->thread_id = info.log1_tid;
870 		}
871 
872 		index = check_for_mtinfo_index();
873 		/* We will not do the mt_info caching optimization in this case.
874 		 */
875 		if (index < 0)
876 			continue;
877 
878 		mt_info[index].valid = 1;
879 		mt_info[index].socket_id = c->socket_id;
880 		mt_info[index].core_id = info.log1_cid;
881 		mt_info[index].thread_id = info.log1_tid;
882 		mt_info[index].proc_fixed_addr = info.log2_la;
883 	}
884 }
885