Linux Audio

Check our new training course

Loading...
v5.4
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * SMP boot-related support
  4 *
  5 * Copyright (C) 1998-2003, 2005 Hewlett-Packard Co
  6 *	David Mosberger-Tang <davidm@hpl.hp.com>
  7 * Copyright (C) 2001, 2004-2005 Intel Corp
  8 * 	Rohit Seth <rohit.seth@intel.com>
  9 * 	Suresh Siddha <suresh.b.siddha@intel.com>
 10 * 	Gordon Jin <gordon.jin@intel.com>
 11 *	Ashok Raj  <ashok.raj@intel.com>
 12 *
 13 * 01/05/16 Rohit Seth <rohit.seth@intel.com>	Moved SMP booting functions from smp.c to here.
 14 * 01/04/27 David Mosberger <davidm@hpl.hp.com>	Added ITC synching code.
 15 * 02/07/31 David Mosberger <davidm@hpl.hp.com>	Switch over to hotplug-CPU boot-sequence.
 16 *						smp_boot_cpus()/smp_commence() is replaced by
 17 *						smp_prepare_cpus()/__cpu_up()/smp_cpus_done().
 18 * 04/06/21 Ashok Raj		<ashok.raj@intel.com> Added CPU Hotplug Support
 19 * 04/12/26 Jin Gordon <gordon.jin@intel.com>
 20 * 04/12/26 Rohit Seth <rohit.seth@intel.com>
 21 *						Add multi-threading and multi-core detection
 22 * 05/01/30 Suresh Siddha <suresh.b.siddha@intel.com>
 23 *						Setup cpu_sibling_map and cpu_core_map
 24 */
 25
 26#include <linux/module.h>
 27#include <linux/acpi.h>
 28#include <linux/memblock.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/spinlock.h>
 40#include <linux/efi.h>
 41#include <linux/percpu.h>
 42#include <linux/bitops.h>
 43
 44#include <linux/atomic.h>
 45#include <asm/cache.h>
 46#include <asm/current.h>
 47#include <asm/delay.h>
 48#include <asm/io.h>
 49#include <asm/irq.h>
 50#include <asm/mca.h>
 51#include <asm/page.h>
 52#include <asm/pgalloc.h>
 53#include <asm/pgtable.h>
 54#include <asm/processor.h>
 55#include <asm/ptrace.h>
 56#include <asm/sal.h>
 57#include <asm/tlbflush.h>
 58#include <asm/unistd.h>
 59
 60#define SMP_DEBUG 0
 61
 62#if SMP_DEBUG
 63#define Dprintk(x...)  printk(x)
 64#else
 65#define Dprintk(x...)
 66#endif
 67
 68#ifdef CONFIG_HOTPLUG_CPU
 69#ifdef CONFIG_PERMIT_BSP_REMOVE
 70#define bsp_remove_ok	1
 71#else
 72#define bsp_remove_ok	0
 73#endif
 74
 75/*
 76 * Global array allocated for NR_CPUS at boot time
 77 */
 78struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
 79
 80/*
 81 * start_ap in head.S uses this to store current booting cpu
 82 * info.
 83 */
 84struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
 85
 86#define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
 87
 88#else
 89#define set_brendez_area(x)
 90#endif
 91
 92
 93/*
 94 * ITC synchronization related stuff:
 95 */
 96#define MASTER	(0)
 97#define SLAVE	(SMP_CACHE_BYTES/8)
 98
 99#define NUM_ROUNDS	64	/* magic value */
100#define NUM_ITERS	5	/* likewise */
101
102static DEFINE_SPINLOCK(itc_sync_lock);
103static volatile unsigned long go[SLAVE + 1];
104
105#define DEBUG_ITC_SYNC	0
106
107extern void start_ap (void);
108extern unsigned long ia64_iobase;
109
110struct task_struct *task_for_booting_cpu;
111
112/*
113 * State for each CPU
114 */
115DEFINE_PER_CPU(int, cpu_state);
116
117cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
118EXPORT_SYMBOL(cpu_core_map);
119DEFINE_PER_CPU_SHARED_ALIGNED(cpumask_t, cpu_sibling_map);
120EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
121
122int smp_num_siblings = 1;
123
124/* which logical CPU number maps to which CPU (physical APIC ID) */
125volatile int ia64_cpu_to_sapicid[NR_CPUS];
126EXPORT_SYMBOL(ia64_cpu_to_sapicid);
127
128static cpumask_t cpu_callin_map;
129
130struct smp_boot_data smp_boot_data __initdata;
131
132unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
133
134char __initdata no_int_routing;
135
136unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
137
138#ifdef CONFIG_FORCE_CPEI_RETARGET
139#define CPEI_OVERRIDE_DEFAULT	(1)
140#else
141#define CPEI_OVERRIDE_DEFAULT	(0)
142#endif
143
144unsigned int force_cpei_retarget = CPEI_OVERRIDE_DEFAULT;
145
146static int __init
147cmdl_force_cpei(char *str)
148{
149	int value=0;
150
151	get_option (&str, &value);
152	force_cpei_retarget = value;
153
154	return 1;
155}
156
157__setup("force_cpei=", cmdl_force_cpei);
158
159static int __init
160nointroute (char *str)
161{
162	no_int_routing = 1;
163	printk ("no_int_routing on\n");
164	return 1;
165}
166
167__setup("nointroute", nointroute);
168
169static void fix_b0_for_bsp(void)
170{
171#ifdef CONFIG_HOTPLUG_CPU
172	int cpuid;
173	static int fix_bsp_b0 = 1;
174
175	cpuid = smp_processor_id();
176
177	/*
178	 * Cache the b0 value on the first AP that comes up
179	 */
180	if (!(fix_bsp_b0 && cpuid))
181		return;
182
183	sal_boot_rendez_state[0].br[0] = sal_boot_rendez_state[cpuid].br[0];
184	printk ("Fixed BSP b0 value from CPU %d\n", cpuid);
185
186	fix_bsp_b0 = 0;
187#endif
188}
189
190void
191sync_master (void *arg)
192{
193	unsigned long flags, i;
194
195	go[MASTER] = 0;
196
197	local_irq_save(flags);
198	{
199		for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
200			while (!go[MASTER])
201				cpu_relax();
202			go[MASTER] = 0;
203			go[SLAVE] = ia64_get_itc();
204		}
205	}
206	local_irq_restore(flags);
207}
208
209/*
210 * Return the number of cycles by which our itc differs from the itc on the master
211 * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
212 * negative that it is behind.
213 */
214static inline long
215get_delta (long *rt, long *master)
216{
217	unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
218	unsigned long tcenter, t0, t1, tm;
219	long i;
220
221	for (i = 0; i < NUM_ITERS; ++i) {
222		t0 = ia64_get_itc();
223		go[MASTER] = 1;
224		while (!(tm = go[SLAVE]))
225			cpu_relax();
226		go[SLAVE] = 0;
227		t1 = ia64_get_itc();
228
229		if (t1 - t0 < best_t1 - best_t0)
230			best_t0 = t0, best_t1 = t1, best_tm = tm;
231	}
232
233	*rt = best_t1 - best_t0;
234	*master = best_tm - best_t0;
235
236	/* average best_t0 and best_t1 without overflow: */
237	tcenter = (best_t0/2 + best_t1/2);
238	if (best_t0 % 2 + best_t1 % 2 == 2)
239		++tcenter;
240	return tcenter - best_tm;
241}
242
243/*
244 * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
245 * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
246 * unaccounted-for errors (such as getting a machine check in the middle of a calibration
247 * step).  The basic idea is for the slave to ask the master what itc value it has and to
248 * read its own itc before and after the master responds.  Each iteration gives us three
249 * timestamps:
250 *
251 *	slave		master
252 *
253 *	t0 ---\
254 *             ---\
255 *		   --->
256 *			tm
257 *		   /---
258 *	       /---
259 *	t1 <---
260 *
261 *
262 * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
263 * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
264 * between the slave and the master is symmetric.  Even if the interconnect were
265 * asymmetric, we would still know that the synchronization error is smaller than the
266 * roundtrip latency (t0 - t1).
267 *
268 * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
269 * within one or two cycles.  However, we can only *guarantee* that the synchronization is
270 * accurate to within a round-trip time, which is typically in the range of several
271 * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
272 * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
273 * than half a micro second or so.
274 */
275void
276ia64_sync_itc (unsigned int master)
277{
278	long i, delta, adj, adjust_latency = 0, done = 0;
279	unsigned long flags, rt, master_time_stamp, bound;
280#if DEBUG_ITC_SYNC
281	struct {
282		long rt;	/* roundtrip time */
283		long master;	/* master's timestamp */
284		long diff;	/* difference between midpoint and master's timestamp */
285		long lat;	/* estimate of itc adjustment latency */
286	} t[NUM_ROUNDS];
287#endif
288
289	/*
290	 * Make sure local timer ticks are disabled while we sync.  If
291	 * they were enabled, we'd have to worry about nasty issues
292	 * like setting the ITC ahead of (or a long time before) the
293	 * next scheduled tick.
294	 */
295	BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
296
297	go[MASTER] = 1;
298
299	if (smp_call_function_single(master, sync_master, NULL, 0) < 0) {
300		printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
301		return;
302	}
303
304	while (go[MASTER])
305		cpu_relax();	/* wait for master to be ready */
306
307	spin_lock_irqsave(&itc_sync_lock, flags);
308	{
309		for (i = 0; i < NUM_ROUNDS; ++i) {
310			delta = get_delta(&rt, &master_time_stamp);
311			if (delta == 0) {
312				done = 1;	/* let's lock on to this... */
313				bound = rt;
314			}
315
316			if (!done) {
317				if (i > 0) {
318					adjust_latency += -delta;
319					adj = -delta + adjust_latency/4;
320				} else
321					adj = -delta;
322
323				ia64_set_itc(ia64_get_itc() + adj);
324			}
325#if DEBUG_ITC_SYNC
326			t[i].rt = rt;
327			t[i].master = master_time_stamp;
328			t[i].diff = delta;
329			t[i].lat = adjust_latency/4;
330#endif
331		}
332	}
333	spin_unlock_irqrestore(&itc_sync_lock, flags);
334
335#if DEBUG_ITC_SYNC
336	for (i = 0; i < NUM_ROUNDS; ++i)
337		printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
338		       t[i].rt, t[i].master, t[i].diff, t[i].lat);
339#endif
340
341	printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
342	       "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
343}
344
345/*
346 * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
347 */
348static inline void smp_setup_percpu_timer(void)
349{
350}
351
352static void
353smp_callin (void)
354{
355	int cpuid, phys_id, itc_master;
356	struct cpuinfo_ia64 *last_cpuinfo, *this_cpuinfo;
357	extern void ia64_init_itm(void);
358	extern volatile int time_keeper_id;
359
360#ifdef CONFIG_PERFMON
361	extern void pfm_init_percpu(void);
362#endif
363
364	cpuid = smp_processor_id();
365	phys_id = hard_smp_processor_id();
366	itc_master = time_keeper_id;
367
368	if (cpu_online(cpuid)) {
369		printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
370		       phys_id, cpuid);
371		BUG();
372	}
373
374	fix_b0_for_bsp();
375
376	/*
377	 * numa_node_id() works after this.
378	 */
379	set_numa_node(cpu_to_node_map[cpuid]);
380	set_numa_mem(local_memory_node(cpu_to_node_map[cpuid]));
381
382	spin_lock(&vector_lock);
383	/* Setup the per cpu irq handling data structures */
384	__setup_vector_irq(cpuid);
385	notify_cpu_starting(cpuid);
386	set_cpu_online(cpuid, true);
387	per_cpu(cpu_state, cpuid) = CPU_ONLINE;
388	spin_unlock(&vector_lock);
389
390	smp_setup_percpu_timer();
391
392	ia64_mca_cmc_vector_setup();	/* Setup vector on AP */
393
394#ifdef CONFIG_PERFMON
395	pfm_init_percpu();
396#endif
397
398	local_irq_enable();
399
400	if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
401		/*
402		 * Synchronize the ITC with the BP.  Need to do this after irqs are
403		 * enabled because ia64_sync_itc() calls smp_call_function_single(), which
404		 * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
405		 * local_bh_enable(), which bugs out if irqs are not enabled...
406		 */
407		Dprintk("Going to syncup ITC with ITC Master.\n");
408		ia64_sync_itc(itc_master);
409	}
410
411	/*
412	 * Get our bogomips.
413	 */
414	ia64_init_itm();
415
416	/*
417	 * Delay calibration can be skipped if new processor is identical to the
418	 * previous processor.
419	 */
420	last_cpuinfo = cpu_data(cpuid - 1);
421	this_cpuinfo = local_cpu_data;
422	if (last_cpuinfo->itc_freq != this_cpuinfo->itc_freq ||
423	    last_cpuinfo->proc_freq != this_cpuinfo->proc_freq ||
424	    last_cpuinfo->features != this_cpuinfo->features ||
425	    last_cpuinfo->revision != this_cpuinfo->revision ||
426	    last_cpuinfo->family != this_cpuinfo->family ||
427	    last_cpuinfo->archrev != this_cpuinfo->archrev ||
428	    last_cpuinfo->model != this_cpuinfo->model)
429		calibrate_delay();
430	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
431
432	/*
433	 * Allow the master to continue.
434	 */
435	cpumask_set_cpu(cpuid, &cpu_callin_map);
436	Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
437}
438
439
440/*
441 * Activate a secondary processor.  head.S calls this.
442 */
443int
444start_secondary (void *unused)
445{
446	/* Early console may use I/O ports */
447	ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
448#ifndef CONFIG_PRINTK_TIME
449	Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
450#endif
451	efi_map_pal_code();
452	cpu_init();
453	preempt_disable();
454	smp_callin();
455
456	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
457	return 0;
458}
459
460static int
461do_boot_cpu (int sapicid, int cpu, struct task_struct *idle)
462{
463	int timeout;
464
465	task_for_booting_cpu = idle;
466	Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
467
468	set_brendez_area(cpu);
469	ia64_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
470
471	/*
472	 * Wait 10s total for the AP to start
473	 */
474	Dprintk("Waiting on callin_map ...");
475	for (timeout = 0; timeout < 100000; timeout++) {
476		if (cpumask_test_cpu(cpu, &cpu_callin_map))
477			break;  /* It has booted */
478		barrier(); /* Make sure we re-read cpu_callin_map */
479		udelay(100);
480	}
481	Dprintk("\n");
482
483	if (!cpumask_test_cpu(cpu, &cpu_callin_map)) {
484		printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
485		ia64_cpu_to_sapicid[cpu] = -1;
486		set_cpu_online(cpu, false);  /* was set in smp_callin() */
487		return -EINVAL;
488	}
489	return 0;
490}
491
492static int __init
493decay (char *str)
494{
495	int ticks;
496	get_option (&str, &ticks);
497	return 1;
498}
499
500__setup("decay=", decay);
501
502/*
503 * Initialize the logical CPU number to SAPICID mapping
504 */
505void __init
506smp_build_cpu_map (void)
507{
508	int sapicid, cpu, i;
509	int boot_cpu_id = hard_smp_processor_id();
510
511	for (cpu = 0; cpu < NR_CPUS; cpu++) {
512		ia64_cpu_to_sapicid[cpu] = -1;
513	}
514
515	ia64_cpu_to_sapicid[0] = boot_cpu_id;
516	init_cpu_present(cpumask_of(0));
517	set_cpu_possible(0, true);
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		set_cpu_present(cpu, true);
523		set_cpu_possible(cpu, true);
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 */
532void __init
533smp_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	cpumask_set_cpu(0, &cpu_callin_map);
544
545	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
546	ia64_cpu_to_sapicid[0] = boot_cpu_id;
547
548	printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
549
550	current_thread_info()->cpu = 0;
551
552	/*
553	 * If SMP should be disabled, then really disable it!
554	 */
555	if (!max_cpus) {
556		printk(KERN_INFO "SMP mode deactivated.\n");
557		init_cpu_online(cpumask_of(0));
558		init_cpu_present(cpumask_of(0));
559		init_cpu_possible(cpumask_of(0));
560		return;
561	}
562}
563
564void smp_prepare_boot_cpu(void)
565{
566	set_cpu_online(smp_processor_id(), true);
567	cpumask_set_cpu(smp_processor_id(), &cpu_callin_map);
568	set_numa_node(cpu_to_node_map[smp_processor_id()]);
569	per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
570}
571
572#ifdef CONFIG_HOTPLUG_CPU
573static inline void
574clear_cpu_sibling_map(int cpu)
575{
576	int i;
577
578	for_each_cpu(i, &per_cpu(cpu_sibling_map, cpu))
579		cpumask_clear_cpu(cpu, &per_cpu(cpu_sibling_map, i));
580	for_each_cpu(i, &cpu_core_map[cpu])
581		cpumask_clear_cpu(cpu, &cpu_core_map[i]);
582
583	per_cpu(cpu_sibling_map, cpu) = cpu_core_map[cpu] = CPU_MASK_NONE;
584}
585
586static void
587remove_siblinginfo(int cpu)
588{
589	int last = 0;
590
591	if (cpu_data(cpu)->threads_per_core == 1 &&
592	    cpu_data(cpu)->cores_per_socket == 1) {
593		cpumask_clear_cpu(cpu, &cpu_core_map[cpu]);
594		cpumask_clear_cpu(cpu, &per_cpu(cpu_sibling_map, cpu));
595		return;
596	}
597
598	last = (cpumask_weight(&cpu_core_map[cpu]) == 1 ? 1 : 0);
599
600	/* remove it from all sibling map's */
601	clear_cpu_sibling_map(cpu);
602}
603
604extern void fixup_irqs(void);
605
606int migrate_platform_irqs(unsigned int cpu)
607{
608	int new_cpei_cpu;
609	struct irq_data *data = NULL;
610	const struct cpumask *mask;
611	int 		retval = 0;
612
613	/*
614	 * dont permit CPEI target to removed.
615	 */
616	if (cpe_vector > 0 && is_cpu_cpei_target(cpu)) {
617		printk ("CPU (%d) is CPEI Target\n", cpu);
618		if (can_cpei_retarget()) {
619			/*
620			 * Now re-target the CPEI to a different processor
621			 */
622			new_cpei_cpu = cpumask_any(cpu_online_mask);
623			mask = cpumask_of(new_cpei_cpu);
624			set_cpei_target_cpu(new_cpei_cpu);
625			data = irq_get_irq_data(ia64_cpe_irq);
626			/*
627			 * Switch for now, immediately, we need to do fake intr
628			 * as other interrupts, but need to study CPEI behaviour with
629			 * polling before making changes.
630			 */
631			if (data && data->chip) {
632				data->chip->irq_disable(data);
633				data->chip->irq_set_affinity(data, mask, false);
634				data->chip->irq_enable(data);
635				printk ("Re-targeting CPEI to cpu %d\n", new_cpei_cpu);
636			}
637		}
638		if (!data) {
639			printk ("Unable to retarget CPEI, offline cpu [%d] failed\n", cpu);
640			retval = -EBUSY;
641		}
642	}
643	return retval;
644}
645
646/* must be called with cpucontrol mutex held */
647int __cpu_disable(void)
648{
649	int cpu = smp_processor_id();
650
651	/*
652	 * dont permit boot processor for now
653	 */
654	if (cpu == 0 && !bsp_remove_ok) {
655		printk ("Your platform does not support removal of BSP\n");
656		return (-EBUSY);
657	}
658
659	set_cpu_online(cpu, false);
660
661	if (migrate_platform_irqs(cpu)) {
662		set_cpu_online(cpu, true);
663		return -EBUSY;
664	}
665
666	remove_siblinginfo(cpu);
667	fixup_irqs();
668	local_flush_tlb_all();
669	cpumask_clear_cpu(cpu, &cpu_callin_map);
670	return 0;
671}
672
673void __cpu_die(unsigned int cpu)
674{
675	unsigned int i;
676
677	for (i = 0; i < 100; i++) {
678		/* They ack this in play_dead by setting CPU_DEAD */
679		if (per_cpu(cpu_state, cpu) == CPU_DEAD)
680		{
681			printk ("CPU %d is now offline\n", cpu);
682			return;
683		}
684		msleep(100);
685	}
686 	printk(KERN_ERR "CPU %u didn't die...\n", cpu);
687}
688#endif /* CONFIG_HOTPLUG_CPU */
689
690void
691smp_cpus_done (unsigned int dummy)
692{
693	int cpu;
694	unsigned long bogosum = 0;
695
696	/*
697	 * Allow the user to impress friends.
698	 */
699
700	for_each_online_cpu(cpu) {
701		bogosum += cpu_data(cpu)->loops_per_jiffy;
702	}
703
704	printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
705	       (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
706}
707
708static inline void set_cpu_sibling_map(int cpu)
709{
710	int i;
711
712	for_each_online_cpu(i) {
713		if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
714			cpumask_set_cpu(i, &cpu_core_map[cpu]);
715			cpumask_set_cpu(cpu, &cpu_core_map[i]);
716			if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
717				cpumask_set_cpu(i,
718						&per_cpu(cpu_sibling_map, cpu));
719				cpumask_set_cpu(cpu,
720						&per_cpu(cpu_sibling_map, i));
721			}
722		}
723	}
724}
725
726int
727__cpu_up(unsigned int cpu, struct task_struct *tidle)
728{
729	int ret;
730	int sapicid;
731
732	sapicid = ia64_cpu_to_sapicid[cpu];
733	if (sapicid == -1)
734		return -EINVAL;
735
736	/*
737	 * Already booted cpu? not valid anymore since we dont
738	 * do idle loop tightspin anymore.
739	 */
740	if (cpumask_test_cpu(cpu, &cpu_callin_map))
741		return -EINVAL;
742
743	per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
744	/* Processor goes to start_secondary(), sets online flag */
745	ret = do_boot_cpu(sapicid, cpu, tidle);
746	if (ret < 0)
747		return ret;
748
749	if (cpu_data(cpu)->threads_per_core == 1 &&
750	    cpu_data(cpu)->cores_per_socket == 1) {
751		cpumask_set_cpu(cpu, &per_cpu(cpu_sibling_map, cpu));
752		cpumask_set_cpu(cpu, &cpu_core_map[cpu]);
753		return 0;
754	}
755
756	set_cpu_sibling_map(cpu);
757
758	return 0;
759}
760
761/*
762 * Assume that CPUs have been discovered by some platform-dependent interface.  For
763 * SoftSDV/Lion, that would be ACPI.
764 *
765 * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
766 */
767void __init
768init_smp_config(void)
769{
770	struct fptr {
771		unsigned long fp;
772		unsigned long gp;
773	} *ap_startup;
774	long sal_ret;
775
776	/* Tell SAL where to drop the APs.  */
777	ap_startup = (struct fptr *) start_ap;
778	sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
779				       ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
780	if (sal_ret < 0)
781		printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
782		       ia64_sal_strerror(sal_ret));
783}
784
785/*
786 * identify_siblings(cpu) gets called from identify_cpu. This populates the 
787 * information related to logical execution units in per_cpu_data structure.
788 */
789void identify_siblings(struct cpuinfo_ia64 *c)
790{
791	long status;
792	u16 pltid;
793	pal_logical_to_physical_t info;
794
795	status = ia64_pal_logical_to_phys(-1, &info);
796	if (status != PAL_STATUS_SUCCESS) {
797		if (status != PAL_STATUS_UNIMPLEMENTED) {
798			printk(KERN_ERR
799				"ia64_pal_logical_to_phys failed with %ld\n",
800				status);
801			return;
802		}
803
804		info.overview_ppid = 0;
805		info.overview_cpp  = 1;
806		info.overview_tpc  = 1;
807	}
808
809	status = ia64_sal_physical_id_info(&pltid);
810	if (status != PAL_STATUS_SUCCESS) {
811		if (status != PAL_STATUS_UNIMPLEMENTED)
812			printk(KERN_ERR
813				"ia64_sal_pltid failed with %ld\n",
814				status);
815		return;
816	}
817
818	c->socket_id =  (pltid << 8) | info.overview_ppid;
819
820	if (info.overview_cpp == 1 && info.overview_tpc == 1)
821		return;
822
823	c->cores_per_socket = info.overview_cpp;
824	c->threads_per_core = info.overview_tpc;
825	c->num_log = info.overview_num_log;
826
827	c->core_id = info.log1_cid;
828	c->thread_id = info.log1_tid;
829}
830
831/*
832 * returns non zero, if multi-threading is enabled
833 * on at least one physical package. Due to hotplug cpu
834 * and (maxcpus=), all threads may not necessarily be enabled
835 * even though the processor supports multi-threading.
836 */
837int is_multithreading_enabled(void)
838{
839	int i, j;
840
841	for_each_present_cpu(i) {
842		for_each_present_cpu(j) {
843			if (j == i)
844				continue;
845			if ((cpu_data(j)->socket_id == cpu_data(i)->socket_id)) {
846				if (cpu_data(j)->core_id == cpu_data(i)->core_id)
847					return 1;
848			}
849		}
850	}
851	return 0;
852}
853EXPORT_SYMBOL_GPL(is_multithreading_enabled);
v5.9
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * SMP boot-related support
  4 *
  5 * Copyright (C) 1998-2003, 2005 Hewlett-Packard Co
  6 *	David Mosberger-Tang <davidm@hpl.hp.com>
  7 * Copyright (C) 2001, 2004-2005 Intel Corp
  8 * 	Rohit Seth <rohit.seth@intel.com>
  9 * 	Suresh Siddha <suresh.b.siddha@intel.com>
 10 * 	Gordon Jin <gordon.jin@intel.com>
 11 *	Ashok Raj  <ashok.raj@intel.com>
 12 *
 13 * 01/05/16 Rohit Seth <rohit.seth@intel.com>	Moved SMP booting functions from smp.c to here.
 14 * 01/04/27 David Mosberger <davidm@hpl.hp.com>	Added ITC synching code.
 15 * 02/07/31 David Mosberger <davidm@hpl.hp.com>	Switch over to hotplug-CPU boot-sequence.
 16 *						smp_boot_cpus()/smp_commence() is replaced by
 17 *						smp_prepare_cpus()/__cpu_up()/smp_cpus_done().
 18 * 04/06/21 Ashok Raj		<ashok.raj@intel.com> Added CPU Hotplug Support
 19 * 04/12/26 Jin Gordon <gordon.jin@intel.com>
 20 * 04/12/26 Rohit Seth <rohit.seth@intel.com>
 21 *						Add multi-threading and multi-core detection
 22 * 05/01/30 Suresh Siddha <suresh.b.siddha@intel.com>
 23 *						Setup cpu_sibling_map and cpu_core_map
 24 */
 25
 26#include <linux/module.h>
 27#include <linux/acpi.h>
 28#include <linux/memblock.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/spinlock.h>
 40#include <linux/efi.h>
 41#include <linux/percpu.h>
 42#include <linux/bitops.h>
 43
 44#include <linux/atomic.h>
 45#include <asm/cache.h>
 46#include <asm/current.h>
 47#include <asm/delay.h>
 48#include <asm/io.h>
 49#include <asm/irq.h>
 50#include <asm/mca.h>
 51#include <asm/page.h>
 
 
 52#include <asm/processor.h>
 53#include <asm/ptrace.h>
 54#include <asm/sal.h>
 55#include <asm/tlbflush.h>
 56#include <asm/unistd.h>
 57
 58#define SMP_DEBUG 0
 59
 60#if SMP_DEBUG
 61#define Dprintk(x...)  printk(x)
 62#else
 63#define Dprintk(x...)
 64#endif
 65
 66#ifdef CONFIG_HOTPLUG_CPU
 67#ifdef CONFIG_PERMIT_BSP_REMOVE
 68#define bsp_remove_ok	1
 69#else
 70#define bsp_remove_ok	0
 71#endif
 72
 73/*
 74 * Global array allocated for NR_CPUS at boot time
 75 */
 76struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
 77
 78/*
 79 * start_ap in head.S uses this to store current booting cpu
 80 * info.
 81 */
 82struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
 83
 84#define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
 85
 86#else
 87#define set_brendez_area(x)
 88#endif
 89
 90
 91/*
 92 * ITC synchronization related stuff:
 93 */
 94#define MASTER	(0)
 95#define SLAVE	(SMP_CACHE_BYTES/8)
 96
 97#define NUM_ROUNDS	64	/* magic value */
 98#define NUM_ITERS	5	/* likewise */
 99
100static DEFINE_SPINLOCK(itc_sync_lock);
101static volatile unsigned long go[SLAVE + 1];
102
103#define DEBUG_ITC_SYNC	0
104
105extern void start_ap (void);
106extern unsigned long ia64_iobase;
107
108struct task_struct *task_for_booting_cpu;
109
110/*
111 * State for each CPU
112 */
113DEFINE_PER_CPU(int, cpu_state);
114
115cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
116EXPORT_SYMBOL(cpu_core_map);
117DEFINE_PER_CPU_SHARED_ALIGNED(cpumask_t, cpu_sibling_map);
118EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
119
120int smp_num_siblings = 1;
121
122/* which logical CPU number maps to which CPU (physical APIC ID) */
123volatile int ia64_cpu_to_sapicid[NR_CPUS];
124EXPORT_SYMBOL(ia64_cpu_to_sapicid);
125
126static cpumask_t cpu_callin_map;
127
128struct smp_boot_data smp_boot_data __initdata;
129
130unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
131
132char __initdata no_int_routing;
133
134unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
135
136#ifdef CONFIG_FORCE_CPEI_RETARGET
137#define CPEI_OVERRIDE_DEFAULT	(1)
138#else
139#define CPEI_OVERRIDE_DEFAULT	(0)
140#endif
141
142unsigned int force_cpei_retarget = CPEI_OVERRIDE_DEFAULT;
143
144static int __init
145cmdl_force_cpei(char *str)
146{
147	int value=0;
148
149	get_option (&str, &value);
150	force_cpei_retarget = value;
151
152	return 1;
153}
154
155__setup("force_cpei=", cmdl_force_cpei);
156
157static int __init
158nointroute (char *str)
159{
160	no_int_routing = 1;
161	printk ("no_int_routing on\n");
162	return 1;
163}
164
165__setup("nointroute", nointroute);
166
167static void fix_b0_for_bsp(void)
168{
169#ifdef CONFIG_HOTPLUG_CPU
170	int cpuid;
171	static int fix_bsp_b0 = 1;
172
173	cpuid = smp_processor_id();
174
175	/*
176	 * Cache the b0 value on the first AP that comes up
177	 */
178	if (!(fix_bsp_b0 && cpuid))
179		return;
180
181	sal_boot_rendez_state[0].br[0] = sal_boot_rendez_state[cpuid].br[0];
182	printk ("Fixed BSP b0 value from CPU %d\n", cpuid);
183
184	fix_bsp_b0 = 0;
185#endif
186}
187
188void
189sync_master (void *arg)
190{
191	unsigned long flags, i;
192
193	go[MASTER] = 0;
194
195	local_irq_save(flags);
196	{
197		for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
198			while (!go[MASTER])
199				cpu_relax();
200			go[MASTER] = 0;
201			go[SLAVE] = ia64_get_itc();
202		}
203	}
204	local_irq_restore(flags);
205}
206
207/*
208 * Return the number of cycles by which our itc differs from the itc on the master
209 * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
210 * negative that it is behind.
211 */
212static inline long
213get_delta (long *rt, long *master)
214{
215	unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
216	unsigned long tcenter, t0, t1, tm;
217	long i;
218
219	for (i = 0; i < NUM_ITERS; ++i) {
220		t0 = ia64_get_itc();
221		go[MASTER] = 1;
222		while (!(tm = go[SLAVE]))
223			cpu_relax();
224		go[SLAVE] = 0;
225		t1 = ia64_get_itc();
226
227		if (t1 - t0 < best_t1 - best_t0)
228			best_t0 = t0, best_t1 = t1, best_tm = tm;
229	}
230
231	*rt = best_t1 - best_t0;
232	*master = best_tm - best_t0;
233
234	/* average best_t0 and best_t1 without overflow: */
235	tcenter = (best_t0/2 + best_t1/2);
236	if (best_t0 % 2 + best_t1 % 2 == 2)
237		++tcenter;
238	return tcenter - best_tm;
239}
240
241/*
242 * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
243 * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
244 * unaccounted-for errors (such as getting a machine check in the middle of a calibration
245 * step).  The basic idea is for the slave to ask the master what itc value it has and to
246 * read its own itc before and after the master responds.  Each iteration gives us three
247 * timestamps:
248 *
249 *	slave		master
250 *
251 *	t0 ---\
252 *             ---\
253 *		   --->
254 *			tm
255 *		   /---
256 *	       /---
257 *	t1 <---
258 *
259 *
260 * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
261 * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
262 * between the slave and the master is symmetric.  Even if the interconnect were
263 * asymmetric, we would still know that the synchronization error is smaller than the
264 * roundtrip latency (t0 - t1).
265 *
266 * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
267 * within one or two cycles.  However, we can only *guarantee* that the synchronization is
268 * accurate to within a round-trip time, which is typically in the range of several
269 * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
270 * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
271 * than half a micro second or so.
272 */
273void
274ia64_sync_itc (unsigned int master)
275{
276	long i, delta, adj, adjust_latency = 0, done = 0;
277	unsigned long flags, rt, master_time_stamp, bound;
278#if DEBUG_ITC_SYNC
279	struct {
280		long rt;	/* roundtrip time */
281		long master;	/* master's timestamp */
282		long diff;	/* difference between midpoint and master's timestamp */
283		long lat;	/* estimate of itc adjustment latency */
284	} t[NUM_ROUNDS];
285#endif
286
287	/*
288	 * Make sure local timer ticks are disabled while we sync.  If
289	 * they were enabled, we'd have to worry about nasty issues
290	 * like setting the ITC ahead of (or a long time before) the
291	 * next scheduled tick.
292	 */
293	BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
294
295	go[MASTER] = 1;
296
297	if (smp_call_function_single(master, sync_master, NULL, 0) < 0) {
298		printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
299		return;
300	}
301
302	while (go[MASTER])
303		cpu_relax();	/* wait for master to be ready */
304
305	spin_lock_irqsave(&itc_sync_lock, flags);
306	{
307		for (i = 0; i < NUM_ROUNDS; ++i) {
308			delta = get_delta(&rt, &master_time_stamp);
309			if (delta == 0) {
310				done = 1;	/* let's lock on to this... */
311				bound = rt;
312			}
313
314			if (!done) {
315				if (i > 0) {
316					adjust_latency += -delta;
317					adj = -delta + adjust_latency/4;
318				} else
319					adj = -delta;
320
321				ia64_set_itc(ia64_get_itc() + adj);
322			}
323#if DEBUG_ITC_SYNC
324			t[i].rt = rt;
325			t[i].master = master_time_stamp;
326			t[i].diff = delta;
327			t[i].lat = adjust_latency/4;
328#endif
329		}
330	}
331	spin_unlock_irqrestore(&itc_sync_lock, flags);
332
333#if DEBUG_ITC_SYNC
334	for (i = 0; i < NUM_ROUNDS; ++i)
335		printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
336		       t[i].rt, t[i].master, t[i].diff, t[i].lat);
337#endif
338
339	printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
340	       "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
341}
342
343/*
344 * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
345 */
346static inline void smp_setup_percpu_timer(void)
347{
348}
349
350static void
351smp_callin (void)
352{
353	int cpuid, phys_id, itc_master;
354	struct cpuinfo_ia64 *last_cpuinfo, *this_cpuinfo;
355	extern void ia64_init_itm(void);
356	extern volatile int time_keeper_id;
357
358#ifdef CONFIG_PERFMON
359	extern void pfm_init_percpu(void);
360#endif
361
362	cpuid = smp_processor_id();
363	phys_id = hard_smp_processor_id();
364	itc_master = time_keeper_id;
365
366	if (cpu_online(cpuid)) {
367		printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
368		       phys_id, cpuid);
369		BUG();
370	}
371
372	fix_b0_for_bsp();
373
374	/*
375	 * numa_node_id() works after this.
376	 */
377	set_numa_node(cpu_to_node_map[cpuid]);
378	set_numa_mem(local_memory_node(cpu_to_node_map[cpuid]));
379
380	spin_lock(&vector_lock);
381	/* Setup the per cpu irq handling data structures */
382	__setup_vector_irq(cpuid);
383	notify_cpu_starting(cpuid);
384	set_cpu_online(cpuid, true);
385	per_cpu(cpu_state, cpuid) = CPU_ONLINE;
386	spin_unlock(&vector_lock);
387
388	smp_setup_percpu_timer();
389
390	ia64_mca_cmc_vector_setup();	/* Setup vector on AP */
391
392#ifdef CONFIG_PERFMON
393	pfm_init_percpu();
394#endif
395
396	local_irq_enable();
397
398	if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
399		/*
400		 * Synchronize the ITC with the BP.  Need to do this after irqs are
401		 * enabled because ia64_sync_itc() calls smp_call_function_single(), which
402		 * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
403		 * local_bh_enable(), which bugs out if irqs are not enabled...
404		 */
405		Dprintk("Going to syncup ITC with ITC Master.\n");
406		ia64_sync_itc(itc_master);
407	}
408
409	/*
410	 * Get our bogomips.
411	 */
412	ia64_init_itm();
413
414	/*
415	 * Delay calibration can be skipped if new processor is identical to the
416	 * previous processor.
417	 */
418	last_cpuinfo = cpu_data(cpuid - 1);
419	this_cpuinfo = local_cpu_data;
420	if (last_cpuinfo->itc_freq != this_cpuinfo->itc_freq ||
421	    last_cpuinfo->proc_freq != this_cpuinfo->proc_freq ||
422	    last_cpuinfo->features != this_cpuinfo->features ||
423	    last_cpuinfo->revision != this_cpuinfo->revision ||
424	    last_cpuinfo->family != this_cpuinfo->family ||
425	    last_cpuinfo->archrev != this_cpuinfo->archrev ||
426	    last_cpuinfo->model != this_cpuinfo->model)
427		calibrate_delay();
428	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
429
430	/*
431	 * Allow the master to continue.
432	 */
433	cpumask_set_cpu(cpuid, &cpu_callin_map);
434	Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
435}
436
437
438/*
439 * Activate a secondary processor.  head.S calls this.
440 */
441int
442start_secondary (void *unused)
443{
444	/* Early console may use I/O ports */
445	ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
446#ifndef CONFIG_PRINTK_TIME
447	Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
448#endif
449	efi_map_pal_code();
450	cpu_init();
451	preempt_disable();
452	smp_callin();
453
454	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
455	return 0;
456}
457
458static int
459do_boot_cpu (int sapicid, int cpu, struct task_struct *idle)
460{
461	int timeout;
462
463	task_for_booting_cpu = idle;
464	Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
465
466	set_brendez_area(cpu);
467	ia64_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
468
469	/*
470	 * Wait 10s total for the AP to start
471	 */
472	Dprintk("Waiting on callin_map ...");
473	for (timeout = 0; timeout < 100000; timeout++) {
474		if (cpumask_test_cpu(cpu, &cpu_callin_map))
475			break;  /* It has booted */
476		barrier(); /* Make sure we re-read cpu_callin_map */
477		udelay(100);
478	}
479	Dprintk("\n");
480
481	if (!cpumask_test_cpu(cpu, &cpu_callin_map)) {
482		printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
483		ia64_cpu_to_sapicid[cpu] = -1;
484		set_cpu_online(cpu, false);  /* was set in smp_callin() */
485		return -EINVAL;
486	}
487	return 0;
488}
489
490static int __init
491decay (char *str)
492{
493	int ticks;
494	get_option (&str, &ticks);
495	return 1;
496}
497
498__setup("decay=", decay);
499
500/*
501 * Initialize the logical CPU number to SAPICID mapping
502 */
503void __init
504smp_build_cpu_map (void)
505{
506	int sapicid, cpu, i;
507	int boot_cpu_id = hard_smp_processor_id();
508
509	for (cpu = 0; cpu < NR_CPUS; cpu++) {
510		ia64_cpu_to_sapicid[cpu] = -1;
511	}
512
513	ia64_cpu_to_sapicid[0] = boot_cpu_id;
514	init_cpu_present(cpumask_of(0));
515	set_cpu_possible(0, true);
516	for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
517		sapicid = smp_boot_data.cpu_phys_id[i];
518		if (sapicid == boot_cpu_id)
519			continue;
520		set_cpu_present(cpu, true);
521		set_cpu_possible(cpu, true);
522		ia64_cpu_to_sapicid[cpu] = sapicid;
523		cpu++;
524	}
525}
526
527/*
528 * Cycle through the APs sending Wakeup IPIs to boot each.
529 */
530void __init
531smp_prepare_cpus (unsigned int max_cpus)
532{
533	int boot_cpu_id = hard_smp_processor_id();
534
535	/*
536	 * Initialize the per-CPU profiling counter/multiplier
537	 */
538
539	smp_setup_percpu_timer();
540
541	cpumask_set_cpu(0, &cpu_callin_map);
542
543	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
544	ia64_cpu_to_sapicid[0] = boot_cpu_id;
545
546	printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
547
548	current_thread_info()->cpu = 0;
549
550	/*
551	 * If SMP should be disabled, then really disable it!
552	 */
553	if (!max_cpus) {
554		printk(KERN_INFO "SMP mode deactivated.\n");
555		init_cpu_online(cpumask_of(0));
556		init_cpu_present(cpumask_of(0));
557		init_cpu_possible(cpumask_of(0));
558		return;
559	}
560}
561
562void smp_prepare_boot_cpu(void)
563{
564	set_cpu_online(smp_processor_id(), true);
565	cpumask_set_cpu(smp_processor_id(), &cpu_callin_map);
566	set_numa_node(cpu_to_node_map[smp_processor_id()]);
567	per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
568}
569
570#ifdef CONFIG_HOTPLUG_CPU
571static inline void
572clear_cpu_sibling_map(int cpu)
573{
574	int i;
575
576	for_each_cpu(i, &per_cpu(cpu_sibling_map, cpu))
577		cpumask_clear_cpu(cpu, &per_cpu(cpu_sibling_map, i));
578	for_each_cpu(i, &cpu_core_map[cpu])
579		cpumask_clear_cpu(cpu, &cpu_core_map[i]);
580
581	per_cpu(cpu_sibling_map, cpu) = cpu_core_map[cpu] = CPU_MASK_NONE;
582}
583
584static void
585remove_siblinginfo(int cpu)
586{
587	int last = 0;
588
589	if (cpu_data(cpu)->threads_per_core == 1 &&
590	    cpu_data(cpu)->cores_per_socket == 1) {
591		cpumask_clear_cpu(cpu, &cpu_core_map[cpu]);
592		cpumask_clear_cpu(cpu, &per_cpu(cpu_sibling_map, cpu));
593		return;
594	}
595
596	last = (cpumask_weight(&cpu_core_map[cpu]) == 1 ? 1 : 0);
597
598	/* remove it from all sibling map's */
599	clear_cpu_sibling_map(cpu);
600}
601
602extern void fixup_irqs(void);
603
604int migrate_platform_irqs(unsigned int cpu)
605{
606	int new_cpei_cpu;
607	struct irq_data *data = NULL;
608	const struct cpumask *mask;
609	int 		retval = 0;
610
611	/*
612	 * dont permit CPEI target to removed.
613	 */
614	if (cpe_vector > 0 && is_cpu_cpei_target(cpu)) {
615		printk ("CPU (%d) is CPEI Target\n", cpu);
616		if (can_cpei_retarget()) {
617			/*
618			 * Now re-target the CPEI to a different processor
619			 */
620			new_cpei_cpu = cpumask_any(cpu_online_mask);
621			mask = cpumask_of(new_cpei_cpu);
622			set_cpei_target_cpu(new_cpei_cpu);
623			data = irq_get_irq_data(ia64_cpe_irq);
624			/*
625			 * Switch for now, immediately, we need to do fake intr
626			 * as other interrupts, but need to study CPEI behaviour with
627			 * polling before making changes.
628			 */
629			if (data && data->chip) {
630				data->chip->irq_disable(data);
631				data->chip->irq_set_affinity(data, mask, false);
632				data->chip->irq_enable(data);
633				printk ("Re-targeting CPEI to cpu %d\n", new_cpei_cpu);
634			}
635		}
636		if (!data) {
637			printk ("Unable to retarget CPEI, offline cpu [%d] failed\n", cpu);
638			retval = -EBUSY;
639		}
640	}
641	return retval;
642}
643
644/* must be called with cpucontrol mutex held */
645int __cpu_disable(void)
646{
647	int cpu = smp_processor_id();
648
649	/*
650	 * dont permit boot processor for now
651	 */
652	if (cpu == 0 && !bsp_remove_ok) {
653		printk ("Your platform does not support removal of BSP\n");
654		return (-EBUSY);
655	}
656
657	set_cpu_online(cpu, false);
658
659	if (migrate_platform_irqs(cpu)) {
660		set_cpu_online(cpu, true);
661		return -EBUSY;
662	}
663
664	remove_siblinginfo(cpu);
665	fixup_irqs();
666	local_flush_tlb_all();
667	cpumask_clear_cpu(cpu, &cpu_callin_map);
668	return 0;
669}
670
671void __cpu_die(unsigned int cpu)
672{
673	unsigned int i;
674
675	for (i = 0; i < 100; i++) {
676		/* They ack this in play_dead by setting CPU_DEAD */
677		if (per_cpu(cpu_state, cpu) == CPU_DEAD)
678		{
679			printk ("CPU %d is now offline\n", cpu);
680			return;
681		}
682		msleep(100);
683	}
684 	printk(KERN_ERR "CPU %u didn't die...\n", cpu);
685}
686#endif /* CONFIG_HOTPLUG_CPU */
687
688void
689smp_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
706static inline void set_cpu_sibling_map(int cpu)
707{
708	int i;
709
710	for_each_online_cpu(i) {
711		if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
712			cpumask_set_cpu(i, &cpu_core_map[cpu]);
713			cpumask_set_cpu(cpu, &cpu_core_map[i]);
714			if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
715				cpumask_set_cpu(i,
716						&per_cpu(cpu_sibling_map, cpu));
717				cpumask_set_cpu(cpu,
718						&per_cpu(cpu_sibling_map, i));
719			}
720		}
721	}
722}
723
724int
725__cpu_up(unsigned int cpu, struct task_struct *tidle)
726{
727	int ret;
728	int sapicid;
729
730	sapicid = ia64_cpu_to_sapicid[cpu];
731	if (sapicid == -1)
732		return -EINVAL;
733
734	/*
735	 * Already booted cpu? not valid anymore since we dont
736	 * do idle loop tightspin anymore.
737	 */
738	if (cpumask_test_cpu(cpu, &cpu_callin_map))
739		return -EINVAL;
740
741	per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
742	/* Processor goes to start_secondary(), sets online flag */
743	ret = do_boot_cpu(sapicid, cpu, tidle);
744	if (ret < 0)
745		return ret;
746
747	if (cpu_data(cpu)->threads_per_core == 1 &&
748	    cpu_data(cpu)->cores_per_socket == 1) {
749		cpumask_set_cpu(cpu, &per_cpu(cpu_sibling_map, cpu));
750		cpumask_set_cpu(cpu, &cpu_core_map[cpu]);
751		return 0;
752	}
753
754	set_cpu_sibling_map(cpu);
755
756	return 0;
757}
758
759/*
760 * Assume that CPUs have been discovered by some platform-dependent interface.  For
761 * SoftSDV/Lion, that would be ACPI.
762 *
763 * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
764 */
765void __init
766init_smp_config(void)
767{
768	struct fptr {
769		unsigned long fp;
770		unsigned long gp;
771	} *ap_startup;
772	long sal_ret;
773
774	/* Tell SAL where to drop the APs.  */
775	ap_startup = (struct fptr *) start_ap;
776	sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
777				       ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
778	if (sal_ret < 0)
779		printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
780		       ia64_sal_strerror(sal_ret));
781}
782
783/*
784 * identify_siblings(cpu) gets called from identify_cpu. This populates the 
785 * information related to logical execution units in per_cpu_data structure.
786 */
787void identify_siblings(struct cpuinfo_ia64 *c)
788{
789	long status;
790	u16 pltid;
791	pal_logical_to_physical_t info;
792
793	status = ia64_pal_logical_to_phys(-1, &info);
794	if (status != PAL_STATUS_SUCCESS) {
795		if (status != PAL_STATUS_UNIMPLEMENTED) {
796			printk(KERN_ERR
797				"ia64_pal_logical_to_phys failed with %ld\n",
798				status);
799			return;
800		}
801
802		info.overview_ppid = 0;
803		info.overview_cpp  = 1;
804		info.overview_tpc  = 1;
805	}
806
807	status = ia64_sal_physical_id_info(&pltid);
808	if (status != PAL_STATUS_SUCCESS) {
809		if (status != PAL_STATUS_UNIMPLEMENTED)
810			printk(KERN_ERR
811				"ia64_sal_pltid failed with %ld\n",
812				status);
813		return;
814	}
815
816	c->socket_id =  (pltid << 8) | info.overview_ppid;
817
818	if (info.overview_cpp == 1 && info.overview_tpc == 1)
819		return;
820
821	c->cores_per_socket = info.overview_cpp;
822	c->threads_per_core = info.overview_tpc;
823	c->num_log = info.overview_num_log;
824
825	c->core_id = info.log1_cid;
826	c->thread_id = info.log1_tid;
827}
828
829/*
830 * returns non zero, if multi-threading is enabled
831 * on at least one physical package. Due to hotplug cpu
832 * and (maxcpus=), all threads may not necessarily be enabled
833 * even though the processor supports multi-threading.
834 */
835int is_multithreading_enabled(void)
836{
837	int i, j;
838
839	for_each_present_cpu(i) {
840		for_each_present_cpu(j) {
841			if (j == i)
842				continue;
843			if ((cpu_data(j)->socket_id == cpu_data(i)->socket_id)) {
844				if (cpu_data(j)->core_id == cpu_data(i)->core_id)
845					return 1;
846			}
847		}
848	}
849	return 0;
850}
851EXPORT_SYMBOL_GPL(is_multithreading_enabled);