Linux Audio

Check our new training course

Open-source upstreaming

Need help get the support for your hardware in upstream Linux?
Loading...
v4.17
  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
 25#include <linux/module.h>
 26#include <linux/acpi.h>
 27#include <linux/bootmem.h>
 28#include <linux/cpu.h>
 29#include <linux/delay.h>
 30#include <linux/init.h>
 31#include <linux/interrupt.h>
 32#include <linux/irq.h>
 33#include <linux/kernel.h>
 34#include <linux/kernel_stat.h>
 35#include <linux/mm.h>
 36#include <linux/notifier.h>
 37#include <linux/smp.h>
 38#include <linux/spinlock.h>
 39#include <linux/efi.h>
 40#include <linux/percpu.h>
 41#include <linux/bitops.h>
 42
 43#include <linux/atomic.h>
 44#include <asm/cache.h>
 45#include <asm/current.h>
 46#include <asm/delay.h>
 47#include <asm/io.h>
 48#include <asm/irq.h>
 49#include <asm/machvec.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#include <asm/sn/arch.h>
 60
 61#define SMP_DEBUG 0
 62
 63#if SMP_DEBUG
 64#define Dprintk(x...)  printk(x)
 65#else
 66#define Dprintk(x...)
 67#endif
 68
 69#ifdef CONFIG_HOTPLUG_CPU
 70#ifdef CONFIG_PERMIT_BSP_REMOVE
 71#define bsp_remove_ok	1
 72#else
 73#define bsp_remove_ok	0
 74#endif
 75
 76/*
 77 * Global array allocated for NR_CPUS at boot time
 78 */
 79struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
 80
 81/*
 82 * start_ap in head.S uses this to store current booting cpu
 83 * info.
 84 */
 85struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
 86
 87#define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
 88
 89#else
 90#define set_brendez_area(x)
 91#endif
 92
 93
 94/*
 95 * ITC synchronization related stuff:
 96 */
 97#define MASTER	(0)
 98#define SLAVE	(SMP_CACHE_BYTES/8)
 99
100#define NUM_ROUNDS	64	/* magic value */
101#define NUM_ITERS	5	/* likewise */
102
103static DEFINE_SPINLOCK(itc_sync_lock);
104static volatile unsigned long go[SLAVE + 1];
105
106#define DEBUG_ITC_SYNC	0
107
108extern void start_ap (void);
109extern unsigned long ia64_iobase;
110
111struct task_struct *task_for_booting_cpu;
112
113/*
114 * State for each CPU
115 */
116DEFINE_PER_CPU(int, cpu_state);
117
118cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
119EXPORT_SYMBOL(cpu_core_map);
120DEFINE_PER_CPU_SHARED_ALIGNED(cpumask_t, cpu_sibling_map);
121EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
122
123int smp_num_siblings = 1;
124
125/* which logical CPU number maps to which CPU (physical APIC ID) */
126volatile int ia64_cpu_to_sapicid[NR_CPUS];
127EXPORT_SYMBOL(ia64_cpu_to_sapicid);
128
129static cpumask_t cpu_callin_map;
130
131struct smp_boot_data smp_boot_data __initdata;
132
133unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
134
135char __initdata no_int_routing;
136
137unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
138
139#ifdef CONFIG_FORCE_CPEI_RETARGET
140#define CPEI_OVERRIDE_DEFAULT	(1)
141#else
142#define CPEI_OVERRIDE_DEFAULT	(0)
143#endif
144
145unsigned int force_cpei_retarget = CPEI_OVERRIDE_DEFAULT;
146
147static int __init
148cmdl_force_cpei(char *str)
149{
150	int value=0;
151
152	get_option (&str, &value);
153	force_cpei_retarget = value;
154
155	return 1;
156}
157
158__setup("force_cpei=", cmdl_force_cpei);
159
160static int __init
161nointroute (char *str)
162{
163	no_int_routing = 1;
164	printk ("no_int_routing on\n");
165	return 1;
166}
167
168__setup("nointroute", nointroute);
169
170static void fix_b0_for_bsp(void)
171{
172#ifdef CONFIG_HOTPLUG_CPU
173	int cpuid;
174	static int fix_bsp_b0 = 1;
175
176	cpuid = smp_processor_id();
177
178	/*
179	 * Cache the b0 value on the first AP that comes up
180	 */
181	if (!(fix_bsp_b0 && cpuid))
182		return;
183
184	sal_boot_rendez_state[0].br[0] = sal_boot_rendez_state[cpuid].br[0];
185	printk ("Fixed BSP b0 value from CPU %d\n", cpuid);
186
187	fix_bsp_b0 = 0;
188#endif
189}
190
191void
192sync_master (void *arg)
193{
194	unsigned long flags, i;
195
196	go[MASTER] = 0;
197
198	local_irq_save(flags);
199	{
200		for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
201			while (!go[MASTER])
202				cpu_relax();
203			go[MASTER] = 0;
204			go[SLAVE] = ia64_get_itc();
205		}
206	}
207	local_irq_restore(flags);
208}
209
210/*
211 * Return the number of cycles by which our itc differs from the itc on the master
212 * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
213 * negative that it is behind.
214 */
215static inline long
216get_delta (long *rt, long *master)
217{
218	unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
219	unsigned long tcenter, t0, t1, tm;
220	long i;
221
222	for (i = 0; i < NUM_ITERS; ++i) {
223		t0 = ia64_get_itc();
224		go[MASTER] = 1;
225		while (!(tm = go[SLAVE]))
226			cpu_relax();
227		go[SLAVE] = 0;
228		t1 = ia64_get_itc();
229
230		if (t1 - t0 < best_t1 - best_t0)
231			best_t0 = t0, best_t1 = t1, best_tm = tm;
232	}
233
234	*rt = best_t1 - best_t0;
235	*master = best_tm - best_t0;
236
237	/* average best_t0 and best_t1 without overflow: */
238	tcenter = (best_t0/2 + best_t1/2);
239	if (best_t0 % 2 + best_t1 % 2 == 2)
240		++tcenter;
241	return tcenter - best_tm;
242}
243
244/*
245 * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
246 * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
247 * unaccounted-for errors (such as getting a machine check in the middle of a calibration
248 * step).  The basic idea is for the slave to ask the master what itc value it has and to
249 * read its own itc before and after the master responds.  Each iteration gives us three
250 * timestamps:
251 *
252 *	slave		master
253 *
254 *	t0 ---\
255 *             ---\
256 *		   --->
257 *			tm
258 *		   /---
259 *	       /---
260 *	t1 <---
261 *
262 *
263 * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
264 * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
265 * between the slave and the master is symmetric.  Even if the interconnect were
266 * asymmetric, we would still know that the synchronization error is smaller than the
267 * roundtrip latency (t0 - t1).
268 *
269 * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
270 * within one or two cycles.  However, we can only *guarantee* that the synchronization is
271 * accurate to within a round-trip time, which is typically in the range of several
272 * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
273 * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
274 * than half a micro second or so.
275 */
276void
277ia64_sync_itc (unsigned int master)
278{
279	long i, delta, adj, adjust_latency = 0, done = 0;
280	unsigned long flags, rt, master_time_stamp, bound;
281#if DEBUG_ITC_SYNC
282	struct {
283		long rt;	/* roundtrip time */
284		long master;	/* master's timestamp */
285		long diff;	/* difference between midpoint and master's timestamp */
286		long lat;	/* estimate of itc adjustment latency */
287	} t[NUM_ROUNDS];
288#endif
289
290	/*
291	 * Make sure local timer ticks are disabled while we sync.  If
292	 * they were enabled, we'd have to worry about nasty issues
293	 * like setting the ITC ahead of (or a long time before) the
294	 * next scheduled tick.
295	 */
296	BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
297
298	go[MASTER] = 1;
299
300	if (smp_call_function_single(master, sync_master, NULL, 0) < 0) {
301		printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
302		return;
303	}
304
305	while (go[MASTER])
306		cpu_relax();	/* wait for master to be ready */
307
308	spin_lock_irqsave(&itc_sync_lock, flags);
309	{
310		for (i = 0; i < NUM_ROUNDS; ++i) {
311			delta = get_delta(&rt, &master_time_stamp);
312			if (delta == 0) {
313				done = 1;	/* let's lock on to this... */
314				bound = rt;
315			}
316
317			if (!done) {
318				if (i > 0) {
319					adjust_latency += -delta;
320					adj = -delta + adjust_latency/4;
321				} else
322					adj = -delta;
323
324				ia64_set_itc(ia64_get_itc() + adj);
325			}
326#if DEBUG_ITC_SYNC
327			t[i].rt = rt;
328			t[i].master = master_time_stamp;
329			t[i].diff = delta;
330			t[i].lat = adjust_latency/4;
331#endif
332		}
333	}
334	spin_unlock_irqrestore(&itc_sync_lock, flags);
335
336#if DEBUG_ITC_SYNC
337	for (i = 0; i < NUM_ROUNDS; ++i)
338		printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
339		       t[i].rt, t[i].master, t[i].diff, t[i].lat);
340#endif
341
342	printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
343	       "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
344}
345
346/*
347 * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
348 */
349static inline void smp_setup_percpu_timer(void)
 
350{
351}
352
353static void
354smp_callin (void)
355{
356	int cpuid, phys_id, itc_master;
357	struct cpuinfo_ia64 *last_cpuinfo, *this_cpuinfo;
358	extern void ia64_init_itm(void);
359	extern volatile int time_keeper_id;
360
361#ifdef CONFIG_PERFMON
362	extern void pfm_init_percpu(void);
363#endif
364
365	cpuid = smp_processor_id();
366	phys_id = hard_smp_processor_id();
367	itc_master = time_keeper_id;
368
369	if (cpu_online(cpuid)) {
370		printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
371		       phys_id, cpuid);
372		BUG();
373	}
374
375	fix_b0_for_bsp();
376
377	/*
378	 * numa_node_id() works after this.
379	 */
380	set_numa_node(cpu_to_node_map[cpuid]);
381	set_numa_mem(local_memory_node(cpu_to_node_map[cpuid]));
382
 
383	spin_lock(&vector_lock);
384	/* Setup the per cpu irq handling data structures */
385	__setup_vector_irq(cpuid);
386	notify_cpu_starting(cpuid);
387	set_cpu_online(cpuid, true);
388	per_cpu(cpu_state, cpuid) = CPU_ONLINE;
389	spin_unlock(&vector_lock);
 
390
391	smp_setup_percpu_timer();
392
393	ia64_mca_cmc_vector_setup();	/* Setup vector on AP */
394
395#ifdef CONFIG_PERFMON
396	pfm_init_percpu();
397#endif
398
399	local_irq_enable();
400
401	if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
402		/*
403		 * Synchronize the ITC with the BP.  Need to do this after irqs are
404		 * enabled because ia64_sync_itc() calls smp_call_function_single(), which
405		 * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
406		 * local_bh_enable(), which bugs out if irqs are not enabled...
407		 */
408		Dprintk("Going to syncup ITC with ITC Master.\n");
409		ia64_sync_itc(itc_master);
410	}
411
412	/*
413	 * Get our bogomips.
414	 */
415	ia64_init_itm();
416
417	/*
418	 * Delay calibration can be skipped if new processor is identical to the
419	 * previous processor.
420	 */
421	last_cpuinfo = cpu_data(cpuid - 1);
422	this_cpuinfo = local_cpu_data;
423	if (last_cpuinfo->itc_freq != this_cpuinfo->itc_freq ||
424	    last_cpuinfo->proc_freq != this_cpuinfo->proc_freq ||
425	    last_cpuinfo->features != this_cpuinfo->features ||
426	    last_cpuinfo->revision != this_cpuinfo->revision ||
427	    last_cpuinfo->family != this_cpuinfo->family ||
428	    last_cpuinfo->archrev != this_cpuinfo->archrev ||
429	    last_cpuinfo->model != this_cpuinfo->model)
430		calibrate_delay();
431	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
432
433	/*
434	 * Allow the master to continue.
435	 */
436	cpumask_set_cpu(cpuid, &cpu_callin_map);
437	Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
438}
439
440
441/*
442 * Activate a secondary processor.  head.S calls this.
443 */
444int
445start_secondary (void *unused)
446{
447	/* Early console may use I/O ports */
448	ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
449#ifndef CONFIG_PRINTK_TIME
450	Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
451#endif
452	efi_map_pal_code();
453	cpu_init();
454	preempt_disable();
455	smp_callin();
456
457	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
458	return 0;
459}
460
461static int
 
 
 
 
 
462do_boot_cpu (int sapicid, int cpu, struct task_struct *idle)
463{
464	int timeout;
465
466	task_for_booting_cpu = idle;
467	Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
468
469	set_brendez_area(cpu);
470	platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
471
472	/*
473	 * Wait 10s total for the AP to start
474	 */
475	Dprintk("Waiting on callin_map ...");
476	for (timeout = 0; timeout < 100000; timeout++) {
477		if (cpumask_test_cpu(cpu, &cpu_callin_map))
478			break;  /* It has booted */
479		barrier(); /* Make sure we re-read cpu_callin_map */
480		udelay(100);
481	}
482	Dprintk("\n");
483
484	if (!cpumask_test_cpu(cpu, &cpu_callin_map)) {
485		printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
486		ia64_cpu_to_sapicid[cpu] = -1;
487		set_cpu_online(cpu, false);  /* was set in smp_callin() */
488		return -EINVAL;
489	}
490	return 0;
491}
492
493static int __init
494decay (char *str)
495{
496	int ticks;
497	get_option (&str, &ticks);
498	return 1;
499}
500
501__setup("decay=", decay);
502
503/*
504 * Initialize the logical CPU number to SAPICID mapping
505 */
506void __init
507smp_build_cpu_map (void)
508{
509	int sapicid, cpu, i;
510	int boot_cpu_id = hard_smp_processor_id();
511
512	for (cpu = 0; cpu < NR_CPUS; cpu++) {
513		ia64_cpu_to_sapicid[cpu] = -1;
514	}
515
516	ia64_cpu_to_sapicid[0] = boot_cpu_id;
517	init_cpu_present(cpumask_of(0));
518	set_cpu_possible(0, true);
519	for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
520		sapicid = smp_boot_data.cpu_phys_id[i];
521		if (sapicid == boot_cpu_id)
522			continue;
523		set_cpu_present(cpu, true);
524		set_cpu_possible(cpu, true);
525		ia64_cpu_to_sapicid[cpu] = sapicid;
526		cpu++;
527	}
528}
529
530/*
531 * Cycle through the APs sending Wakeup IPIs to boot each.
532 */
533void __init
534smp_prepare_cpus (unsigned int max_cpus)
535{
536	int boot_cpu_id = hard_smp_processor_id();
537
538	/*
539	 * Initialize the per-CPU profiling counter/multiplier
540	 */
541
542	smp_setup_percpu_timer();
543
544	cpumask_set_cpu(0, &cpu_callin_map);
545
546	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
547	ia64_cpu_to_sapicid[0] = boot_cpu_id;
548
549	printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
550
551	current_thread_info()->cpu = 0;
552
553	/*
554	 * If SMP should be disabled, then really disable it!
555	 */
556	if (!max_cpus) {
557		printk(KERN_INFO "SMP mode deactivated.\n");
558		init_cpu_online(cpumask_of(0));
559		init_cpu_present(cpumask_of(0));
560		init_cpu_possible(cpumask_of(0));
561		return;
562	}
563}
564
565void smp_prepare_boot_cpu(void)
566{
567	set_cpu_online(smp_processor_id(), true);
568	cpumask_set_cpu(smp_processor_id(), &cpu_callin_map);
569	set_numa_node(cpu_to_node_map[smp_processor_id()]);
570	per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
 
571}
572
573#ifdef CONFIG_HOTPLUG_CPU
574static inline void
575clear_cpu_sibling_map(int cpu)
576{
577	int i;
578
579	for_each_cpu(i, &per_cpu(cpu_sibling_map, cpu))
580		cpumask_clear_cpu(cpu, &per_cpu(cpu_sibling_map, i));
581	for_each_cpu(i, &cpu_core_map[cpu])
582		cpumask_clear_cpu(cpu, &cpu_core_map[i]);
583
584	per_cpu(cpu_sibling_map, cpu) = cpu_core_map[cpu] = CPU_MASK_NONE;
585}
586
587static void
588remove_siblinginfo(int cpu)
589{
590	int last = 0;
591
592	if (cpu_data(cpu)->threads_per_core == 1 &&
593	    cpu_data(cpu)->cores_per_socket == 1) {
594		cpumask_clear_cpu(cpu, &cpu_core_map[cpu]);
595		cpumask_clear_cpu(cpu, &per_cpu(cpu_sibling_map, cpu));
596		return;
597	}
598
599	last = (cpumask_weight(&cpu_core_map[cpu]) == 1 ? 1 : 0);
600
601	/* remove it from all sibling map's */
602	clear_cpu_sibling_map(cpu);
603}
604
605extern void fixup_irqs(void);
606
607int migrate_platform_irqs(unsigned int cpu)
608{
609	int new_cpei_cpu;
610	struct irq_data *data = NULL;
611	const struct cpumask *mask;
612	int 		retval = 0;
613
614	/*
615	 * dont permit CPEI target to removed.
616	 */
617	if (cpe_vector > 0 && is_cpu_cpei_target(cpu)) {
618		printk ("CPU (%d) is CPEI Target\n", cpu);
619		if (can_cpei_retarget()) {
620			/*
621			 * Now re-target the CPEI to a different processor
622			 */
623			new_cpei_cpu = cpumask_any(cpu_online_mask);
624			mask = cpumask_of(new_cpei_cpu);
625			set_cpei_target_cpu(new_cpei_cpu);
626			data = irq_get_irq_data(ia64_cpe_irq);
627			/*
628			 * Switch for now, immediately, we need to do fake intr
629			 * as other interrupts, but need to study CPEI behaviour with
630			 * polling before making changes.
631			 */
632			if (data && data->chip) {
633				data->chip->irq_disable(data);
634				data->chip->irq_set_affinity(data, mask, false);
635				data->chip->irq_enable(data);
636				printk ("Re-targeting CPEI to cpu %d\n", new_cpei_cpu);
637			}
638		}
639		if (!data) {
640			printk ("Unable to retarget CPEI, offline cpu [%d] failed\n", cpu);
641			retval = -EBUSY;
642		}
643	}
644	return retval;
645}
646
647/* must be called with cpucontrol mutex held */
648int __cpu_disable(void)
649{
650	int cpu = smp_processor_id();
651
652	/*
653	 * dont permit boot processor for now
654	 */
655	if (cpu == 0 && !bsp_remove_ok) {
656		printk ("Your platform does not support removal of BSP\n");
657		return (-EBUSY);
658	}
659
660	if (ia64_platform_is("sn2")) {
661		if (!sn_cpu_disable_allowed(cpu))
662			return -EBUSY;
663	}
664
665	set_cpu_online(cpu, false);
666
667	if (migrate_platform_irqs(cpu)) {
668		set_cpu_online(cpu, true);
669		return -EBUSY;
670	}
671
672	remove_siblinginfo(cpu);
673	fixup_irqs();
674	local_flush_tlb_all();
675	cpumask_clear_cpu(cpu, &cpu_callin_map);
676	return 0;
677}
678
679void __cpu_die(unsigned int cpu)
680{
681	unsigned int i;
682
683	for (i = 0; i < 100; i++) {
684		/* They ack this in play_dead by setting CPU_DEAD */
685		if (per_cpu(cpu_state, cpu) == CPU_DEAD)
686		{
687			printk ("CPU %d is now offline\n", cpu);
688			return;
689		}
690		msleep(100);
691	}
692 	printk(KERN_ERR "CPU %u didn't die...\n", cpu);
693}
694#endif /* CONFIG_HOTPLUG_CPU */
695
696void
697smp_cpus_done (unsigned int dummy)
698{
699	int cpu;
700	unsigned long bogosum = 0;
701
702	/*
703	 * Allow the user to impress friends.
704	 */
705
706	for_each_online_cpu(cpu) {
707		bogosum += cpu_data(cpu)->loops_per_jiffy;
708	}
709
710	printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
711	       (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
712}
713
714static inline void set_cpu_sibling_map(int cpu)
 
715{
716	int i;
717
718	for_each_online_cpu(i) {
719		if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
720			cpumask_set_cpu(i, &cpu_core_map[cpu]);
721			cpumask_set_cpu(cpu, &cpu_core_map[i]);
722			if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
723				cpumask_set_cpu(i,
724						&per_cpu(cpu_sibling_map, cpu));
725				cpumask_set_cpu(cpu,
726						&per_cpu(cpu_sibling_map, i));
727			}
728		}
729	}
730}
731
732int
733__cpu_up(unsigned int cpu, struct task_struct *tidle)
734{
735	int ret;
736	int sapicid;
737
738	sapicid = ia64_cpu_to_sapicid[cpu];
739	if (sapicid == -1)
740		return -EINVAL;
741
742	/*
743	 * Already booted cpu? not valid anymore since we dont
744	 * do idle loop tightspin anymore.
745	 */
746	if (cpumask_test_cpu(cpu, &cpu_callin_map))
747		return -EINVAL;
748
749	per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
750	/* Processor goes to start_secondary(), sets online flag */
751	ret = do_boot_cpu(sapicid, cpu, tidle);
752	if (ret < 0)
753		return ret;
754
755	if (cpu_data(cpu)->threads_per_core == 1 &&
756	    cpu_data(cpu)->cores_per_socket == 1) {
757		cpumask_set_cpu(cpu, &per_cpu(cpu_sibling_map, cpu));
758		cpumask_set_cpu(cpu, &cpu_core_map[cpu]);
759		return 0;
760	}
761
762	set_cpu_sibling_map(cpu);
763
764	return 0;
765}
766
767/*
768 * Assume that CPUs have been discovered by some platform-dependent interface.  For
769 * SoftSDV/Lion, that would be ACPI.
770 *
771 * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
772 */
773void __init
774init_smp_config(void)
775{
776	struct fptr {
777		unsigned long fp;
778		unsigned long gp;
779	} *ap_startup;
780	long sal_ret;
781
782	/* Tell SAL where to drop the APs.  */
783	ap_startup = (struct fptr *) start_ap;
784	sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
785				       ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
786	if (sal_ret < 0)
787		printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
788		       ia64_sal_strerror(sal_ret));
789}
790
791/*
792 * identify_siblings(cpu) gets called from identify_cpu. This populates the 
793 * information related to logical execution units in per_cpu_data structure.
794 */
795void identify_siblings(struct cpuinfo_ia64 *c)
 
796{
797	long status;
798	u16 pltid;
799	pal_logical_to_physical_t info;
800
801	status = ia64_pal_logical_to_phys(-1, &info);
802	if (status != PAL_STATUS_SUCCESS) {
803		if (status != PAL_STATUS_UNIMPLEMENTED) {
804			printk(KERN_ERR
805				"ia64_pal_logical_to_phys failed with %ld\n",
806				status);
807			return;
808		}
809
810		info.overview_ppid = 0;
811		info.overview_cpp  = 1;
812		info.overview_tpc  = 1;
813	}
814
815	status = ia64_sal_physical_id_info(&pltid);
816	if (status != PAL_STATUS_SUCCESS) {
817		if (status != PAL_STATUS_UNIMPLEMENTED)
818			printk(KERN_ERR
819				"ia64_sal_pltid failed with %ld\n",
820				status);
821		return;
822	}
823
824	c->socket_id =  (pltid << 8) | info.overview_ppid;
825
826	if (info.overview_cpp == 1 && info.overview_tpc == 1)
827		return;
828
829	c->cores_per_socket = info.overview_cpp;
830	c->threads_per_core = info.overview_tpc;
831	c->num_log = info.overview_num_log;
832
833	c->core_id = info.log1_cid;
834	c->thread_id = info.log1_tid;
835}
836
837/*
838 * returns non zero, if multi-threading is enabled
839 * on at least one physical package. Due to hotplug cpu
840 * and (maxcpus=), all threads may not necessarily be enabled
841 * even though the processor supports multi-threading.
842 */
843int is_multithreading_enabled(void)
844{
845	int i, j;
846
847	for_each_present_cpu(i) {
848		for_each_present_cpu(j) {
849			if (j == i)
850				continue;
851			if ((cpu_data(j)->socket_id == cpu_data(i)->socket_id)) {
852				if (cpu_data(j)->core_id == cpu_data(i)->core_id)
853					return 1;
854			}
855		}
856	}
857	return 0;
858}
859EXPORT_SYMBOL_GPL(is_multithreading_enabled);
v3.5.6
  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
 25#include <linux/module.h>
 26#include <linux/acpi.h>
 27#include <linux/bootmem.h>
 28#include <linux/cpu.h>
 29#include <linux/delay.h>
 30#include <linux/init.h>
 31#include <linux/interrupt.h>
 32#include <linux/irq.h>
 33#include <linux/kernel.h>
 34#include <linux/kernel_stat.h>
 35#include <linux/mm.h>
 36#include <linux/notifier.h>
 37#include <linux/smp.h>
 38#include <linux/spinlock.h>
 39#include <linux/efi.h>
 40#include <linux/percpu.h>
 41#include <linux/bitops.h>
 42
 43#include <linux/atomic.h>
 44#include <asm/cache.h>
 45#include <asm/current.h>
 46#include <asm/delay.h>
 47#include <asm/io.h>
 48#include <asm/irq.h>
 49#include <asm/machvec.h>
 50#include <asm/mca.h>
 51#include <asm/page.h>
 52#include <asm/paravirt.h>
 53#include <asm/pgalloc.h>
 54#include <asm/pgtable.h>
 55#include <asm/processor.h>
 56#include <asm/ptrace.h>
 57#include <asm/sal.h>
 58#include <asm/tlbflush.h>
 59#include <asm/unistd.h>
 60#include <asm/sn/arch.h>
 61
 62#define SMP_DEBUG 0
 63
 64#if SMP_DEBUG
 65#define Dprintk(x...)  printk(x)
 66#else
 67#define Dprintk(x...)
 68#endif
 69
 70#ifdef CONFIG_HOTPLUG_CPU
 71#ifdef CONFIG_PERMIT_BSP_REMOVE
 72#define bsp_remove_ok	1
 73#else
 74#define bsp_remove_ok	0
 75#endif
 76
 77/*
 78 * Global array allocated for NR_CPUS at boot time
 79 */
 80struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
 81
 82/*
 83 * start_ap in head.S uses this to store current booting cpu
 84 * info.
 85 */
 86struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
 87
 88#define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
 89
 90#else
 91#define set_brendez_area(x)
 92#endif
 93
 94
 95/*
 96 * ITC synchronization related stuff:
 97 */
 98#define MASTER	(0)
 99#define SLAVE	(SMP_CACHE_BYTES/8)
100
101#define NUM_ROUNDS	64	/* magic value */
102#define NUM_ITERS	5	/* likewise */
103
104static DEFINE_SPINLOCK(itc_sync_lock);
105static volatile unsigned long go[SLAVE + 1];
106
107#define DEBUG_ITC_SYNC	0
108
109extern void start_ap (void);
110extern unsigned long ia64_iobase;
111
112struct task_struct *task_for_booting_cpu;
113
114/*
115 * State for each CPU
116 */
117DEFINE_PER_CPU(int, cpu_state);
118
119cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
120EXPORT_SYMBOL(cpu_core_map);
121DEFINE_PER_CPU_SHARED_ALIGNED(cpumask_t, cpu_sibling_map);
122EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
123
124int smp_num_siblings = 1;
125
126/* which logical CPU number maps to which CPU (physical APIC ID) */
127volatile int ia64_cpu_to_sapicid[NR_CPUS];
128EXPORT_SYMBOL(ia64_cpu_to_sapicid);
129
130static volatile cpumask_t cpu_callin_map;
131
132struct smp_boot_data smp_boot_data __initdata;
133
134unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
135
136char __initdata no_int_routing;
137
138unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
139
140#ifdef CONFIG_FORCE_CPEI_RETARGET
141#define CPEI_OVERRIDE_DEFAULT	(1)
142#else
143#define CPEI_OVERRIDE_DEFAULT	(0)
144#endif
145
146unsigned int force_cpei_retarget = CPEI_OVERRIDE_DEFAULT;
147
148static int __init
149cmdl_force_cpei(char *str)
150{
151	int value=0;
152
153	get_option (&str, &value);
154	force_cpei_retarget = value;
155
156	return 1;
157}
158
159__setup("force_cpei=", cmdl_force_cpei);
160
161static int __init
162nointroute (char *str)
163{
164	no_int_routing = 1;
165	printk ("no_int_routing on\n");
166	return 1;
167}
168
169__setup("nointroute", nointroute);
170
171static void fix_b0_for_bsp(void)
172{
173#ifdef CONFIG_HOTPLUG_CPU
174	int cpuid;
175	static int fix_bsp_b0 = 1;
176
177	cpuid = smp_processor_id();
178
179	/*
180	 * Cache the b0 value on the first AP that comes up
181	 */
182	if (!(fix_bsp_b0 && cpuid))
183		return;
184
185	sal_boot_rendez_state[0].br[0] = sal_boot_rendez_state[cpuid].br[0];
186	printk ("Fixed BSP b0 value from CPU %d\n", cpuid);
187
188	fix_bsp_b0 = 0;
189#endif
190}
191
192void
193sync_master (void *arg)
194{
195	unsigned long flags, i;
196
197	go[MASTER] = 0;
198
199	local_irq_save(flags);
200	{
201		for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
202			while (!go[MASTER])
203				cpu_relax();
204			go[MASTER] = 0;
205			go[SLAVE] = ia64_get_itc();
206		}
207	}
208	local_irq_restore(flags);
209}
210
211/*
212 * Return the number of cycles by which our itc differs from the itc on the master
213 * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
214 * negative that it is behind.
215 */
216static inline long
217get_delta (long *rt, long *master)
218{
219	unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
220	unsigned long tcenter, t0, t1, tm;
221	long i;
222
223	for (i = 0; i < NUM_ITERS; ++i) {
224		t0 = ia64_get_itc();
225		go[MASTER] = 1;
226		while (!(tm = go[SLAVE]))
227			cpu_relax();
228		go[SLAVE] = 0;
229		t1 = ia64_get_itc();
230
231		if (t1 - t0 < best_t1 - best_t0)
232			best_t0 = t0, best_t1 = t1, best_tm = tm;
233	}
234
235	*rt = best_t1 - best_t0;
236	*master = best_tm - best_t0;
237
238	/* average best_t0 and best_t1 without overflow: */
239	tcenter = (best_t0/2 + best_t1/2);
240	if (best_t0 % 2 + best_t1 % 2 == 2)
241		++tcenter;
242	return tcenter - best_tm;
243}
244
245/*
246 * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
247 * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
248 * unaccounted-for errors (such as getting a machine check in the middle of a calibration
249 * step).  The basic idea is for the slave to ask the master what itc value it has and to
250 * read its own itc before and after the master responds.  Each iteration gives us three
251 * timestamps:
252 *
253 *	slave		master
254 *
255 *	t0 ---\
256 *             ---\
257 *		   --->
258 *			tm
259 *		   /---
260 *	       /---
261 *	t1 <---
262 *
263 *
264 * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
265 * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
266 * between the slave and the master is symmetric.  Even if the interconnect were
267 * asymmetric, we would still know that the synchronization error is smaller than the
268 * roundtrip latency (t0 - t1).
269 *
270 * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
271 * within one or two cycles.  However, we can only *guarantee* that the synchronization is
272 * accurate to within a round-trip time, which is typically in the range of several
273 * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
274 * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
275 * than half a micro second or so.
276 */
277void
278ia64_sync_itc (unsigned int master)
279{
280	long i, delta, adj, adjust_latency = 0, done = 0;
281	unsigned long flags, rt, master_time_stamp, bound;
282#if DEBUG_ITC_SYNC
283	struct {
284		long rt;	/* roundtrip time */
285		long master;	/* master's timestamp */
286		long diff;	/* difference between midpoint and master's timestamp */
287		long lat;	/* estimate of itc adjustment latency */
288	} t[NUM_ROUNDS];
289#endif
290
291	/*
292	 * Make sure local timer ticks are disabled while we sync.  If
293	 * they were enabled, we'd have to worry about nasty issues
294	 * like setting the ITC ahead of (or a long time before) the
295	 * next scheduled tick.
296	 */
297	BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
298
299	go[MASTER] = 1;
300
301	if (smp_call_function_single(master, sync_master, NULL, 0) < 0) {
302		printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
303		return;
304	}
305
306	while (go[MASTER])
307		cpu_relax();	/* wait for master to be ready */
308
309	spin_lock_irqsave(&itc_sync_lock, flags);
310	{
311		for (i = 0; i < NUM_ROUNDS; ++i) {
312			delta = get_delta(&rt, &master_time_stamp);
313			if (delta == 0) {
314				done = 1;	/* let's lock on to this... */
315				bound = rt;
316			}
317
318			if (!done) {
319				if (i > 0) {
320					adjust_latency += -delta;
321					adj = -delta + adjust_latency/4;
322				} else
323					adj = -delta;
324
325				ia64_set_itc(ia64_get_itc() + adj);
326			}
327#if DEBUG_ITC_SYNC
328			t[i].rt = rt;
329			t[i].master = master_time_stamp;
330			t[i].diff = delta;
331			t[i].lat = adjust_latency/4;
332#endif
333		}
334	}
335	spin_unlock_irqrestore(&itc_sync_lock, flags);
336
337#if DEBUG_ITC_SYNC
338	for (i = 0; i < NUM_ROUNDS; ++i)
339		printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
340		       t[i].rt, t[i].master, t[i].diff, t[i].lat);
341#endif
342
343	printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
344	       "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
345}
346
347/*
348 * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
349 */
350static inline void __devinit
351smp_setup_percpu_timer (void)
352{
353}
354
355static void __cpuinit
356smp_callin (void)
357{
358	int cpuid, phys_id, itc_master;
359	struct cpuinfo_ia64 *last_cpuinfo, *this_cpuinfo;
360	extern void ia64_init_itm(void);
361	extern volatile int time_keeper_id;
362
363#ifdef CONFIG_PERFMON
364	extern void pfm_init_percpu(void);
365#endif
366
367	cpuid = smp_processor_id();
368	phys_id = hard_smp_processor_id();
369	itc_master = time_keeper_id;
370
371	if (cpu_online(cpuid)) {
372		printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
373		       phys_id, cpuid);
374		BUG();
375	}
376
377	fix_b0_for_bsp();
378
379	/*
380	 * numa_node_id() works after this.
381	 */
382	set_numa_node(cpu_to_node_map[cpuid]);
383	set_numa_mem(local_memory_node(cpu_to_node_map[cpuid]));
384
385	ipi_call_lock_irq();
386	spin_lock(&vector_lock);
387	/* Setup the per cpu irq handling data structures */
388	__setup_vector_irq(cpuid);
389	notify_cpu_starting(cpuid);
390	set_cpu_online(cpuid, true);
391	per_cpu(cpu_state, cpuid) = CPU_ONLINE;
392	spin_unlock(&vector_lock);
393	ipi_call_unlock_irq();
394
395	smp_setup_percpu_timer();
396
397	ia64_mca_cmc_vector_setup();	/* Setup vector on AP */
398
399#ifdef CONFIG_PERFMON
400	pfm_init_percpu();
401#endif
402
403	local_irq_enable();
404
405	if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
406		/*
407		 * Synchronize the ITC with the BP.  Need to do this after irqs are
408		 * enabled because ia64_sync_itc() calls smp_call_function_single(), which
409		 * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
410		 * local_bh_enable(), which bugs out if irqs are not enabled...
411		 */
412		Dprintk("Going to syncup ITC with ITC Master.\n");
413		ia64_sync_itc(itc_master);
414	}
415
416	/*
417	 * Get our bogomips.
418	 */
419	ia64_init_itm();
420
421	/*
422	 * Delay calibration can be skipped if new processor is identical to the
423	 * previous processor.
424	 */
425	last_cpuinfo = cpu_data(cpuid - 1);
426	this_cpuinfo = local_cpu_data;
427	if (last_cpuinfo->itc_freq != this_cpuinfo->itc_freq ||
428	    last_cpuinfo->proc_freq != this_cpuinfo->proc_freq ||
429	    last_cpuinfo->features != this_cpuinfo->features ||
430	    last_cpuinfo->revision != this_cpuinfo->revision ||
431	    last_cpuinfo->family != this_cpuinfo->family ||
432	    last_cpuinfo->archrev != this_cpuinfo->archrev ||
433	    last_cpuinfo->model != this_cpuinfo->model)
434		calibrate_delay();
435	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
436
437	/*
438	 * Allow the master to continue.
439	 */
440	cpu_set(cpuid, cpu_callin_map);
441	Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
442}
443
444
445/*
446 * Activate a secondary processor.  head.S calls this.
447 */
448int __cpuinit
449start_secondary (void *unused)
450{
451	/* Early console may use I/O ports */
452	ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
453#ifndef CONFIG_PRINTK_TIME
454	Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
455#endif
456	efi_map_pal_code();
457	cpu_init();
458	preempt_disable();
459	smp_callin();
460
461	cpu_idle();
462	return 0;
463}
464
465struct pt_regs * __cpuinit idle_regs(struct pt_regs *regs)
466{
467	return NULL;
468}
469
470static int __cpuinit
471do_boot_cpu (int sapicid, int cpu, struct task_struct *idle)
472{
473	int timeout;
474
475	task_for_booting_cpu = idle;
476	Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
477
478	set_brendez_area(cpu);
479	platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
480
481	/*
482	 * Wait 10s total for the AP to start
483	 */
484	Dprintk("Waiting on callin_map ...");
485	for (timeout = 0; timeout < 100000; timeout++) {
486		if (cpu_isset(cpu, cpu_callin_map))
487			break;  /* It has booted */
 
488		udelay(100);
489	}
490	Dprintk("\n");
491
492	if (!cpu_isset(cpu, cpu_callin_map)) {
493		printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
494		ia64_cpu_to_sapicid[cpu] = -1;
495		set_cpu_online(cpu, false);  /* was set in smp_callin() */
496		return -EINVAL;
497	}
498	return 0;
499}
500
501static int __init
502decay (char *str)
503{
504	int ticks;
505	get_option (&str, &ticks);
506	return 1;
507}
508
509__setup("decay=", decay);
510
511/*
512 * Initialize the logical CPU number to SAPICID mapping
513 */
514void __init
515smp_build_cpu_map (void)
516{
517	int sapicid, cpu, i;
518	int boot_cpu_id = hard_smp_processor_id();
519
520	for (cpu = 0; cpu < NR_CPUS; cpu++) {
521		ia64_cpu_to_sapicid[cpu] = -1;
522	}
523
524	ia64_cpu_to_sapicid[0] = boot_cpu_id;
525	init_cpu_present(cpumask_of(0));
526	set_cpu_possible(0, true);
527	for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
528		sapicid = smp_boot_data.cpu_phys_id[i];
529		if (sapicid == boot_cpu_id)
530			continue;
531		set_cpu_present(cpu, true);
532		set_cpu_possible(cpu, true);
533		ia64_cpu_to_sapicid[cpu] = sapicid;
534		cpu++;
535	}
536}
537
538/*
539 * Cycle through the APs sending Wakeup IPIs to boot each.
540 */
541void __init
542smp_prepare_cpus (unsigned int max_cpus)
543{
544	int boot_cpu_id = hard_smp_processor_id();
545
546	/*
547	 * Initialize the per-CPU profiling counter/multiplier
548	 */
549
550	smp_setup_percpu_timer();
551
552	cpu_set(0, cpu_callin_map);
553
554	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
555	ia64_cpu_to_sapicid[0] = boot_cpu_id;
556
557	printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
558
559	current_thread_info()->cpu = 0;
560
561	/*
562	 * If SMP should be disabled, then really disable it!
563	 */
564	if (!max_cpus) {
565		printk(KERN_INFO "SMP mode deactivated.\n");
566		init_cpu_online(cpumask_of(0));
567		init_cpu_present(cpumask_of(0));
568		init_cpu_possible(cpumask_of(0));
569		return;
570	}
571}
572
573void __devinit smp_prepare_boot_cpu(void)
574{
575	set_cpu_online(smp_processor_id(), true);
576	cpu_set(smp_processor_id(), cpu_callin_map);
577	set_numa_node(cpu_to_node_map[smp_processor_id()]);
578	per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
579	paravirt_post_smp_prepare_boot_cpu();
580}
581
582#ifdef CONFIG_HOTPLUG_CPU
583static inline void
584clear_cpu_sibling_map(int cpu)
585{
586	int i;
587
588	for_each_cpu_mask(i, per_cpu(cpu_sibling_map, cpu))
589		cpu_clear(cpu, per_cpu(cpu_sibling_map, i));
590	for_each_cpu_mask(i, cpu_core_map[cpu])
591		cpu_clear(cpu, cpu_core_map[i]);
592
593	per_cpu(cpu_sibling_map, cpu) = cpu_core_map[cpu] = CPU_MASK_NONE;
594}
595
596static void
597remove_siblinginfo(int cpu)
598{
599	int last = 0;
600
601	if (cpu_data(cpu)->threads_per_core == 1 &&
602	    cpu_data(cpu)->cores_per_socket == 1) {
603		cpu_clear(cpu, cpu_core_map[cpu]);
604		cpu_clear(cpu, per_cpu(cpu_sibling_map, cpu));
605		return;
606	}
607
608	last = (cpus_weight(cpu_core_map[cpu]) == 1 ? 1 : 0);
609
610	/* remove it from all sibling map's */
611	clear_cpu_sibling_map(cpu);
612}
613
614extern void fixup_irqs(void);
615
616int migrate_platform_irqs(unsigned int cpu)
617{
618	int new_cpei_cpu;
619	struct irq_data *data = NULL;
620	const struct cpumask *mask;
621	int 		retval = 0;
622
623	/*
624	 * dont permit CPEI target to removed.
625	 */
626	if (cpe_vector > 0 && is_cpu_cpei_target(cpu)) {
627		printk ("CPU (%d) is CPEI Target\n", cpu);
628		if (can_cpei_retarget()) {
629			/*
630			 * Now re-target the CPEI to a different processor
631			 */
632			new_cpei_cpu = cpumask_any(cpu_online_mask);
633			mask = cpumask_of(new_cpei_cpu);
634			set_cpei_target_cpu(new_cpei_cpu);
635			data = irq_get_irq_data(ia64_cpe_irq);
636			/*
637			 * Switch for now, immediately, we need to do fake intr
638			 * as other interrupts, but need to study CPEI behaviour with
639			 * polling before making changes.
640			 */
641			if (data && data->chip) {
642				data->chip->irq_disable(data);
643				data->chip->irq_set_affinity(data, mask, false);
644				data->chip->irq_enable(data);
645				printk ("Re-targeting CPEI to cpu %d\n", new_cpei_cpu);
646			}
647		}
648		if (!data) {
649			printk ("Unable to retarget CPEI, offline cpu [%d] failed\n", cpu);
650			retval = -EBUSY;
651		}
652	}
653	return retval;
654}
655
656/* must be called with cpucontrol mutex held */
657int __cpu_disable(void)
658{
659	int cpu = smp_processor_id();
660
661	/*
662	 * dont permit boot processor for now
663	 */
664	if (cpu == 0 && !bsp_remove_ok) {
665		printk ("Your platform does not support removal of BSP\n");
666		return (-EBUSY);
667	}
668
669	if (ia64_platform_is("sn2")) {
670		if (!sn_cpu_disable_allowed(cpu))
671			return -EBUSY;
672	}
673
674	set_cpu_online(cpu, false);
675
676	if (migrate_platform_irqs(cpu)) {
677		set_cpu_online(cpu, true);
678		return -EBUSY;
679	}
680
681	remove_siblinginfo(cpu);
682	fixup_irqs();
683	local_flush_tlb_all();
684	cpu_clear(cpu, cpu_callin_map);
685	return 0;
686}
687
688void __cpu_die(unsigned int cpu)
689{
690	unsigned int i;
691
692	for (i = 0; i < 100; i++) {
693		/* They ack this in play_dead by setting CPU_DEAD */
694		if (per_cpu(cpu_state, cpu) == CPU_DEAD)
695		{
696			printk ("CPU %d is now offline\n", cpu);
697			return;
698		}
699		msleep(100);
700	}
701 	printk(KERN_ERR "CPU %u didn't die...\n", cpu);
702}
703#endif /* CONFIG_HOTPLUG_CPU */
704
705void
706smp_cpus_done (unsigned int dummy)
707{
708	int cpu;
709	unsigned long bogosum = 0;
710
711	/*
712	 * Allow the user to impress friends.
713	 */
714
715	for_each_online_cpu(cpu) {
716		bogosum += cpu_data(cpu)->loops_per_jiffy;
717	}
718
719	printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
720	       (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
721}
722
723static inline void __devinit
724set_cpu_sibling_map(int cpu)
725{
726	int i;
727
728	for_each_online_cpu(i) {
729		if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
730			cpu_set(i, cpu_core_map[cpu]);
731			cpu_set(cpu, cpu_core_map[i]);
732			if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
733				cpu_set(i, per_cpu(cpu_sibling_map, cpu));
734				cpu_set(cpu, per_cpu(cpu_sibling_map, i));
 
 
735			}
736		}
737	}
738}
739
740int __cpuinit
741__cpu_up(unsigned int cpu, struct task_struct *tidle)
742{
743	int ret;
744	int sapicid;
745
746	sapicid = ia64_cpu_to_sapicid[cpu];
747	if (sapicid == -1)
748		return -EINVAL;
749
750	/*
751	 * Already booted cpu? not valid anymore since we dont
752	 * do idle loop tightspin anymore.
753	 */
754	if (cpu_isset(cpu, cpu_callin_map))
755		return -EINVAL;
756
757	per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
758	/* Processor goes to start_secondary(), sets online flag */
759	ret = do_boot_cpu(sapicid, cpu, tidle);
760	if (ret < 0)
761		return ret;
762
763	if (cpu_data(cpu)->threads_per_core == 1 &&
764	    cpu_data(cpu)->cores_per_socket == 1) {
765		cpu_set(cpu, per_cpu(cpu_sibling_map, cpu));
766		cpu_set(cpu, cpu_core_map[cpu]);
767		return 0;
768	}
769
770	set_cpu_sibling_map(cpu);
771
772	return 0;
773}
774
775/*
776 * Assume that CPUs have been discovered by some platform-dependent interface.  For
777 * SoftSDV/Lion, that would be ACPI.
778 *
779 * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
780 */
781void __init
782init_smp_config(void)
783{
784	struct fptr {
785		unsigned long fp;
786		unsigned long gp;
787	} *ap_startup;
788	long sal_ret;
789
790	/* Tell SAL where to drop the APs.  */
791	ap_startup = (struct fptr *) start_ap;
792	sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
793				       ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
794	if (sal_ret < 0)
795		printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
796		       ia64_sal_strerror(sal_ret));
797}
798
799/*
800 * identify_siblings(cpu) gets called from identify_cpu. This populates the 
801 * information related to logical execution units in per_cpu_data structure.
802 */
803void __devinit
804identify_siblings(struct cpuinfo_ia64 *c)
805{
806	long status;
807	u16 pltid;
808	pal_logical_to_physical_t info;
809
810	status = ia64_pal_logical_to_phys(-1, &info);
811	if (status != PAL_STATUS_SUCCESS) {
812		if (status != PAL_STATUS_UNIMPLEMENTED) {
813			printk(KERN_ERR
814				"ia64_pal_logical_to_phys failed with %ld\n",
815				status);
816			return;
817		}
818
819		info.overview_ppid = 0;
820		info.overview_cpp  = 1;
821		info.overview_tpc  = 1;
822	}
823
824	status = ia64_sal_physical_id_info(&pltid);
825	if (status != PAL_STATUS_SUCCESS) {
826		if (status != PAL_STATUS_UNIMPLEMENTED)
827			printk(KERN_ERR
828				"ia64_sal_pltid failed with %ld\n",
829				status);
830		return;
831	}
832
833	c->socket_id =  (pltid << 8) | info.overview_ppid;
834
835	if (info.overview_cpp == 1 && info.overview_tpc == 1)
836		return;
837
838	c->cores_per_socket = info.overview_cpp;
839	c->threads_per_core = info.overview_tpc;
840	c->num_log = info.overview_num_log;
841
842	c->core_id = info.log1_cid;
843	c->thread_id = info.log1_tid;
844}
845
846/*
847 * returns non zero, if multi-threading is enabled
848 * on at least one physical package. Due to hotplug cpu
849 * and (maxcpus=), all threads may not necessarily be enabled
850 * even though the processor supports multi-threading.
851 */
852int is_multithreading_enabled(void)
853{
854	int i, j;
855
856	for_each_present_cpu(i) {
857		for_each_present_cpu(j) {
858			if (j == i)
859				continue;
860			if ((cpu_data(j)->socket_id == cpu_data(i)->socket_id)) {
861				if (cpu_data(j)->core_id == cpu_data(i)->core_id)
862					return 1;
863			}
864		}
865	}
866	return 0;
867}
868EXPORT_SYMBOL_GPL(is_multithreading_enabled);