xref: /openbmc/linux/arch/ia64/kernel/smpboot.c (revision 64c70b1c)
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 <asm/atomic.h>
44 #include <asm/cache.h>
45 #include <asm/current.h>
46 #include <asm/delay.h>
47 #include <asm/ia32.h>
48 #include <asm/io.h>
49 #include <asm/irq.h>
50 #include <asm/machvec.h>
51 #include <asm/mca.h>
52 #include <asm/page.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/system.h>
59 #include <asm/tlbflush.h>
60 #include <asm/unistd.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  * Store all idle threads, this can be reused instead of creating
79  * a new thread. Also avoids complicated thread destroy functionality
80  * for idle threads.
81  */
82 struct task_struct *idle_thread_array[NR_CPUS];
83 
84 /*
85  * Global array allocated for NR_CPUS at boot time
86  */
87 struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
88 
89 /*
90  * start_ap in head.S uses this to store current booting cpu
91  * info.
92  */
93 struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
94 
95 #define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
96 
97 #define get_idle_for_cpu(x)		(idle_thread_array[(x)])
98 #define set_idle_for_cpu(x,p)	(idle_thread_array[(x)] = (p))
99 
100 #else
101 
102 #define get_idle_for_cpu(x)		(NULL)
103 #define set_idle_for_cpu(x,p)
104 #define set_brendez_area(x)
105 #endif
106 
107 
108 /*
109  * ITC synchronization related stuff:
110  */
111 #define MASTER	(0)
112 #define SLAVE	(SMP_CACHE_BYTES/8)
113 
114 #define NUM_ROUNDS	64	/* magic value */
115 #define NUM_ITERS	5	/* likewise */
116 
117 static DEFINE_SPINLOCK(itc_sync_lock);
118 static volatile unsigned long go[SLAVE + 1];
119 
120 #define DEBUG_ITC_SYNC	0
121 
122 extern void __devinit calibrate_delay (void);
123 extern void start_ap (void);
124 extern unsigned long ia64_iobase;
125 
126 struct task_struct *task_for_booting_cpu;
127 
128 /*
129  * State for each CPU
130  */
131 DEFINE_PER_CPU(int, cpu_state);
132 
133 /* Bitmasks of currently online, and possible CPUs */
134 cpumask_t cpu_online_map;
135 EXPORT_SYMBOL(cpu_online_map);
136 cpumask_t cpu_possible_map = CPU_MASK_NONE;
137 EXPORT_SYMBOL(cpu_possible_map);
138 
139 cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
140 cpumask_t cpu_sibling_map[NR_CPUS] __cacheline_aligned;
141 int smp_num_siblings = 1;
142 int smp_num_cpucores = 1;
143 
144 /* which logical CPU number maps to which CPU (physical APIC ID) */
145 volatile int ia64_cpu_to_sapicid[NR_CPUS];
146 EXPORT_SYMBOL(ia64_cpu_to_sapicid);
147 
148 static volatile cpumask_t cpu_callin_map;
149 
150 struct smp_boot_data smp_boot_data __initdata;
151 
152 unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
153 
154 char __initdata no_int_routing;
155 
156 unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
157 
158 #ifdef CONFIG_FORCE_CPEI_RETARGET
159 #define CPEI_OVERRIDE_DEFAULT	(1)
160 #else
161 #define CPEI_OVERRIDE_DEFAULT	(0)
162 #endif
163 
164 unsigned int force_cpei_retarget = CPEI_OVERRIDE_DEFAULT;
165 
166 static int __init
167 cmdl_force_cpei(char *str)
168 {
169 	int value=0;
170 
171 	get_option (&str, &value);
172 	force_cpei_retarget = value;
173 
174 	return 1;
175 }
176 
177 __setup("force_cpei=", cmdl_force_cpei);
178 
179 static int __init
180 nointroute (char *str)
181 {
182 	no_int_routing = 1;
183 	printk ("no_int_routing on\n");
184 	return 1;
185 }
186 
187 __setup("nointroute", nointroute);
188 
189 static void fix_b0_for_bsp(void)
190 {
191 #ifdef CONFIG_HOTPLUG_CPU
192 	int cpuid;
193 	static int fix_bsp_b0 = 1;
194 
195 	cpuid = smp_processor_id();
196 
197 	/*
198 	 * Cache the b0 value on the first AP that comes up
199 	 */
200 	if (!(fix_bsp_b0 && cpuid))
201 		return;
202 
203 	sal_boot_rendez_state[0].br[0] = sal_boot_rendez_state[cpuid].br[0];
204 	printk ("Fixed BSP b0 value from CPU %d\n", cpuid);
205 
206 	fix_bsp_b0 = 0;
207 #endif
208 }
209 
210 void
211 sync_master (void *arg)
212 {
213 	unsigned long flags, i;
214 
215 	go[MASTER] = 0;
216 
217 	local_irq_save(flags);
218 	{
219 		for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
220 			while (!go[MASTER])
221 				cpu_relax();
222 			go[MASTER] = 0;
223 			go[SLAVE] = ia64_get_itc();
224 		}
225 	}
226 	local_irq_restore(flags);
227 }
228 
229 /*
230  * Return the number of cycles by which our itc differs from the itc on the master
231  * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
232  * negative that it is behind.
233  */
234 static inline long
235 get_delta (long *rt, long *master)
236 {
237 	unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
238 	unsigned long tcenter, t0, t1, tm;
239 	long i;
240 
241 	for (i = 0; i < NUM_ITERS; ++i) {
242 		t0 = ia64_get_itc();
243 		go[MASTER] = 1;
244 		while (!(tm = go[SLAVE]))
245 			cpu_relax();
246 		go[SLAVE] = 0;
247 		t1 = ia64_get_itc();
248 
249 		if (t1 - t0 < best_t1 - best_t0)
250 			best_t0 = t0, best_t1 = t1, best_tm = tm;
251 	}
252 
253 	*rt = best_t1 - best_t0;
254 	*master = best_tm - best_t0;
255 
256 	/* average best_t0 and best_t1 without overflow: */
257 	tcenter = (best_t0/2 + best_t1/2);
258 	if (best_t0 % 2 + best_t1 % 2 == 2)
259 		++tcenter;
260 	return tcenter - best_tm;
261 }
262 
263 /*
264  * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
265  * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
266  * unaccounted-for errors (such as getting a machine check in the middle of a calibration
267  * step).  The basic idea is for the slave to ask the master what itc value it has and to
268  * read its own itc before and after the master responds.  Each iteration gives us three
269  * timestamps:
270  *
271  *	slave		master
272  *
273  *	t0 ---\
274  *             ---\
275  *		   --->
276  *			tm
277  *		   /---
278  *	       /---
279  *	t1 <---
280  *
281  *
282  * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
283  * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
284  * between the slave and the master is symmetric.  Even if the interconnect were
285  * asymmetric, we would still know that the synchronization error is smaller than the
286  * roundtrip latency (t0 - t1).
287  *
288  * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
289  * within one or two cycles.  However, we can only *guarantee* that the synchronization is
290  * accurate to within a round-trip time, which is typically in the range of several
291  * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
292  * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
293  * than half a micro second or so.
294  */
295 void
296 ia64_sync_itc (unsigned int master)
297 {
298 	long i, delta, adj, adjust_latency = 0, done = 0;
299 	unsigned long flags, rt, master_time_stamp, bound;
300 #if DEBUG_ITC_SYNC
301 	struct {
302 		long rt;	/* roundtrip time */
303 		long master;	/* master's timestamp */
304 		long diff;	/* difference between midpoint and master's timestamp */
305 		long lat;	/* estimate of itc adjustment latency */
306 	} t[NUM_ROUNDS];
307 #endif
308 
309 	/*
310 	 * Make sure local timer ticks are disabled while we sync.  If
311 	 * they were enabled, we'd have to worry about nasty issues
312 	 * like setting the ITC ahead of (or a long time before) the
313 	 * next scheduled tick.
314 	 */
315 	BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
316 
317 	go[MASTER] = 1;
318 
319 	if (smp_call_function_single(master, sync_master, NULL, 1, 0) < 0) {
320 		printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
321 		return;
322 	}
323 
324 	while (go[MASTER])
325 		cpu_relax();	/* wait for master to be ready */
326 
327 	spin_lock_irqsave(&itc_sync_lock, flags);
328 	{
329 		for (i = 0; i < NUM_ROUNDS; ++i) {
330 			delta = get_delta(&rt, &master_time_stamp);
331 			if (delta == 0) {
332 				done = 1;	/* let's lock on to this... */
333 				bound = rt;
334 			}
335 
336 			if (!done) {
337 				if (i > 0) {
338 					adjust_latency += -delta;
339 					adj = -delta + adjust_latency/4;
340 				} else
341 					adj = -delta;
342 
343 				ia64_set_itc(ia64_get_itc() + adj);
344 			}
345 #if DEBUG_ITC_SYNC
346 			t[i].rt = rt;
347 			t[i].master = master_time_stamp;
348 			t[i].diff = delta;
349 			t[i].lat = adjust_latency/4;
350 #endif
351 		}
352 	}
353 	spin_unlock_irqrestore(&itc_sync_lock, flags);
354 
355 #if DEBUG_ITC_SYNC
356 	for (i = 0; i < NUM_ROUNDS; ++i)
357 		printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
358 		       t[i].rt, t[i].master, t[i].diff, t[i].lat);
359 #endif
360 
361 	printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
362 	       "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
363 }
364 
365 /*
366  * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
367  */
368 static inline void __devinit
369 smp_setup_percpu_timer (void)
370 {
371 }
372 
373 static void __cpuinit
374 smp_callin (void)
375 {
376 	int cpuid, phys_id, itc_master;
377 	struct cpuinfo_ia64 *last_cpuinfo, *this_cpuinfo;
378 	extern void ia64_init_itm(void);
379 	extern volatile int time_keeper_id;
380 
381 #ifdef CONFIG_PERFMON
382 	extern void pfm_init_percpu(void);
383 #endif
384 
385 	cpuid = smp_processor_id();
386 	phys_id = hard_smp_processor_id();
387 	itc_master = time_keeper_id;
388 
389 	if (cpu_online(cpuid)) {
390 		printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
391 		       phys_id, cpuid);
392 		BUG();
393 	}
394 
395 	fix_b0_for_bsp();
396 
397 	lock_ipi_calllock();
398 	cpu_set(cpuid, cpu_online_map);
399 	unlock_ipi_calllock();
400 	per_cpu(cpu_state, cpuid) = CPU_ONLINE;
401 
402 	smp_setup_percpu_timer();
403 
404 	ia64_mca_cmc_vector_setup();	/* Setup vector on AP */
405 
406 #ifdef CONFIG_PERFMON
407 	pfm_init_percpu();
408 #endif
409 
410 	local_irq_enable();
411 
412 	if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
413 		/*
414 		 * Synchronize the ITC with the BP.  Need to do this after irqs are
415 		 * enabled because ia64_sync_itc() calls smp_call_function_single(), which
416 		 * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
417 		 * local_bh_enable(), which bugs out if irqs are not enabled...
418 		 */
419 		Dprintk("Going to syncup ITC with ITC Master.\n");
420 		ia64_sync_itc(itc_master);
421 	}
422 
423 	/*
424 	 * Get our bogomips.
425 	 */
426 	ia64_init_itm();
427 
428 	/*
429 	 * Delay calibration can be skipped if new processor is identical to the
430 	 * previous processor.
431 	 */
432 	last_cpuinfo = cpu_data(cpuid - 1);
433 	this_cpuinfo = local_cpu_data;
434 	if (last_cpuinfo->itc_freq != this_cpuinfo->itc_freq ||
435 	    last_cpuinfo->proc_freq != this_cpuinfo->proc_freq ||
436 	    last_cpuinfo->features != this_cpuinfo->features ||
437 	    last_cpuinfo->revision != this_cpuinfo->revision ||
438 	    last_cpuinfo->family != this_cpuinfo->family ||
439 	    last_cpuinfo->archrev != this_cpuinfo->archrev ||
440 	    last_cpuinfo->model != this_cpuinfo->model)
441 		calibrate_delay();
442 	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
443 
444 #ifdef CONFIG_IA32_SUPPORT
445 	ia32_gdt_init();
446 #endif
447 
448 	/*
449 	 * Allow the master to continue.
450 	 */
451 	cpu_set(cpuid, cpu_callin_map);
452 	Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
453 }
454 
455 
456 /*
457  * Activate a secondary processor.  head.S calls this.
458  */
459 int __cpuinit
460 start_secondary (void *unused)
461 {
462 	/* Early console may use I/O ports */
463 	ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
464 	Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
465 	efi_map_pal_code();
466 	cpu_init();
467 	preempt_disable();
468 	smp_callin();
469 
470 	cpu_idle();
471 	return 0;
472 }
473 
474 struct pt_regs * __devinit idle_regs(struct pt_regs *regs)
475 {
476 	return NULL;
477 }
478 
479 struct create_idle {
480 	struct work_struct work;
481 	struct task_struct *idle;
482 	struct completion done;
483 	int cpu;
484 };
485 
486 void
487 do_fork_idle(struct work_struct *work)
488 {
489 	struct create_idle *c_idle =
490 		container_of(work, struct create_idle, work);
491 
492 	c_idle->idle = fork_idle(c_idle->cpu);
493 	complete(&c_idle->done);
494 }
495 
496 static int __devinit
497 do_boot_cpu (int sapicid, int cpu)
498 {
499 	int timeout;
500 	struct create_idle c_idle = {
501 		.work = __WORK_INITIALIZER(c_idle.work, do_fork_idle),
502 		.cpu	= cpu,
503 		.done	= COMPLETION_INITIALIZER(c_idle.done),
504 	};
505 
506  	c_idle.idle = get_idle_for_cpu(cpu);
507  	if (c_idle.idle) {
508 		init_idle(c_idle.idle, cpu);
509  		goto do_rest;
510 	}
511 
512 	/*
513 	 * We can't use kernel_thread since we must avoid to reschedule the child.
514 	 */
515 	if (!keventd_up() || current_is_keventd())
516 		c_idle.work.func(&c_idle.work);
517 	else {
518 		schedule_work(&c_idle.work);
519 		wait_for_completion(&c_idle.done);
520 	}
521 
522 	if (IS_ERR(c_idle.idle))
523 		panic("failed fork for CPU %d", cpu);
524 
525 	set_idle_for_cpu(cpu, c_idle.idle);
526 
527 do_rest:
528 	task_for_booting_cpu = c_idle.idle;
529 
530 	Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
531 
532 	set_brendez_area(cpu);
533 	platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
534 
535 	/*
536 	 * Wait 10s total for the AP to start
537 	 */
538 	Dprintk("Waiting on callin_map ...");
539 	for (timeout = 0; timeout < 100000; timeout++) {
540 		if (cpu_isset(cpu, cpu_callin_map))
541 			break;  /* It has booted */
542 		udelay(100);
543 	}
544 	Dprintk("\n");
545 
546 	if (!cpu_isset(cpu, cpu_callin_map)) {
547 		printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
548 		ia64_cpu_to_sapicid[cpu] = -1;
549 		cpu_clear(cpu, cpu_online_map);  /* was set in smp_callin() */
550 		return -EINVAL;
551 	}
552 	return 0;
553 }
554 
555 static int __init
556 decay (char *str)
557 {
558 	int ticks;
559 	get_option (&str, &ticks);
560 	return 1;
561 }
562 
563 __setup("decay=", decay);
564 
565 /*
566  * Initialize the logical CPU number to SAPICID mapping
567  */
568 void __init
569 smp_build_cpu_map (void)
570 {
571 	int sapicid, cpu, i;
572 	int boot_cpu_id = hard_smp_processor_id();
573 
574 	for (cpu = 0; cpu < NR_CPUS; cpu++) {
575 		ia64_cpu_to_sapicid[cpu] = -1;
576 	}
577 
578 	ia64_cpu_to_sapicid[0] = boot_cpu_id;
579 	cpus_clear(cpu_present_map);
580 	cpu_set(0, cpu_present_map);
581 	cpu_set(0, cpu_possible_map);
582 	for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
583 		sapicid = smp_boot_data.cpu_phys_id[i];
584 		if (sapicid == boot_cpu_id)
585 			continue;
586 		cpu_set(cpu, cpu_present_map);
587 		cpu_set(cpu, cpu_possible_map);
588 		ia64_cpu_to_sapicid[cpu] = sapicid;
589 		cpu++;
590 	}
591 }
592 
593 /*
594  * Cycle through the APs sending Wakeup IPIs to boot each.
595  */
596 void __init
597 smp_prepare_cpus (unsigned int max_cpus)
598 {
599 	int boot_cpu_id = hard_smp_processor_id();
600 
601 	/*
602 	 * Initialize the per-CPU profiling counter/multiplier
603 	 */
604 
605 	smp_setup_percpu_timer();
606 
607 	/*
608 	 * We have the boot CPU online for sure.
609 	 */
610 	cpu_set(0, cpu_online_map);
611 	cpu_set(0, cpu_callin_map);
612 
613 	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
614 	ia64_cpu_to_sapicid[0] = boot_cpu_id;
615 
616 	printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
617 
618 	current_thread_info()->cpu = 0;
619 
620 	/*
621 	 * If SMP should be disabled, then really disable it!
622 	 */
623 	if (!max_cpus) {
624 		printk(KERN_INFO "SMP mode deactivated.\n");
625 		cpus_clear(cpu_online_map);
626 		cpus_clear(cpu_present_map);
627 		cpus_clear(cpu_possible_map);
628 		cpu_set(0, cpu_online_map);
629 		cpu_set(0, cpu_present_map);
630 		cpu_set(0, cpu_possible_map);
631 		return;
632 	}
633 }
634 
635 void __devinit smp_prepare_boot_cpu(void)
636 {
637 	cpu_set(smp_processor_id(), cpu_online_map);
638 	cpu_set(smp_processor_id(), cpu_callin_map);
639 	per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
640 }
641 
642 #ifdef CONFIG_HOTPLUG_CPU
643 static inline void
644 clear_cpu_sibling_map(int cpu)
645 {
646 	int i;
647 
648 	for_each_cpu_mask(i, cpu_sibling_map[cpu])
649 		cpu_clear(cpu, cpu_sibling_map[i]);
650 	for_each_cpu_mask(i, cpu_core_map[cpu])
651 		cpu_clear(cpu, cpu_core_map[i]);
652 
653 	cpu_sibling_map[cpu] = cpu_core_map[cpu] = CPU_MASK_NONE;
654 }
655 
656 static void
657 remove_siblinginfo(int cpu)
658 {
659 	int last = 0;
660 
661 	if (cpu_data(cpu)->threads_per_core == 1 &&
662 	    cpu_data(cpu)->cores_per_socket == 1) {
663 		cpu_clear(cpu, cpu_core_map[cpu]);
664 		cpu_clear(cpu, cpu_sibling_map[cpu]);
665 		return;
666 	}
667 
668 	last = (cpus_weight(cpu_core_map[cpu]) == 1 ? 1 : 0);
669 
670 	/* remove it from all sibling map's */
671 	clear_cpu_sibling_map(cpu);
672 }
673 
674 extern void fixup_irqs(void);
675 
676 int migrate_platform_irqs(unsigned int cpu)
677 {
678 	int new_cpei_cpu;
679 	irq_desc_t *desc = NULL;
680 	cpumask_t 	mask;
681 	int 		retval = 0;
682 
683 	/*
684 	 * dont permit CPEI target to removed.
685 	 */
686 	if (cpe_vector > 0 && is_cpu_cpei_target(cpu)) {
687 		printk ("CPU (%d) is CPEI Target\n", cpu);
688 		if (can_cpei_retarget()) {
689 			/*
690 			 * Now re-target the CPEI to a different processor
691 			 */
692 			new_cpei_cpu = any_online_cpu(cpu_online_map);
693 			mask = cpumask_of_cpu(new_cpei_cpu);
694 			set_cpei_target_cpu(new_cpei_cpu);
695 			desc = irq_desc + ia64_cpe_irq;
696 			/*
697 			 * Switch for now, immediately, we need to do fake intr
698 			 * as other interrupts, but need to study CPEI behaviour with
699 			 * polling before making changes.
700 			 */
701 			if (desc) {
702 				desc->chip->disable(ia64_cpe_irq);
703 				desc->chip->set_affinity(ia64_cpe_irq, mask);
704 				desc->chip->enable(ia64_cpe_irq);
705 				printk ("Re-targetting CPEI to cpu %d\n", new_cpei_cpu);
706 			}
707 		}
708 		if (!desc) {
709 			printk ("Unable to retarget CPEI, offline cpu [%d] failed\n", cpu);
710 			retval = -EBUSY;
711 		}
712 	}
713 	return retval;
714 }
715 
716 /* must be called with cpucontrol mutex held */
717 int __cpu_disable(void)
718 {
719 	int cpu = smp_processor_id();
720 
721 	/*
722 	 * dont permit boot processor for now
723 	 */
724 	if (cpu == 0 && !bsp_remove_ok) {
725 		printk ("Your platform does not support removal of BSP\n");
726 		return (-EBUSY);
727 	}
728 
729 	cpu_clear(cpu, cpu_online_map);
730 
731 	if (migrate_platform_irqs(cpu)) {
732 		cpu_set(cpu, cpu_online_map);
733 		return (-EBUSY);
734 	}
735 
736 	remove_siblinginfo(cpu);
737 	cpu_clear(cpu, cpu_online_map);
738 	fixup_irqs();
739 	local_flush_tlb_all();
740 	cpu_clear(cpu, cpu_callin_map);
741 	return 0;
742 }
743 
744 void __cpu_die(unsigned int cpu)
745 {
746 	unsigned int i;
747 
748 	for (i = 0; i < 100; i++) {
749 		/* They ack this in play_dead by setting CPU_DEAD */
750 		if (per_cpu(cpu_state, cpu) == CPU_DEAD)
751 		{
752 			printk ("CPU %d is now offline\n", cpu);
753 			return;
754 		}
755 		msleep(100);
756 	}
757  	printk(KERN_ERR "CPU %u didn't die...\n", cpu);
758 }
759 #else /* !CONFIG_HOTPLUG_CPU */
760 int __cpu_disable(void)
761 {
762 	return -ENOSYS;
763 }
764 
765 void __cpu_die(unsigned int cpu)
766 {
767 	/* We said "no" in __cpu_disable */
768 	BUG();
769 }
770 #endif /* CONFIG_HOTPLUG_CPU */
771 
772 void
773 smp_cpus_done (unsigned int dummy)
774 {
775 	int cpu;
776 	unsigned long bogosum = 0;
777 
778 	/*
779 	 * Allow the user to impress friends.
780 	 */
781 
782 	for_each_online_cpu(cpu) {
783 		bogosum += cpu_data(cpu)->loops_per_jiffy;
784 	}
785 
786 	printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
787 	       (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
788 }
789 
790 static inline void __devinit
791 set_cpu_sibling_map(int cpu)
792 {
793 	int i;
794 
795 	for_each_online_cpu(i) {
796 		if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
797 			cpu_set(i, cpu_core_map[cpu]);
798 			cpu_set(cpu, cpu_core_map[i]);
799 			if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
800 				cpu_set(i, cpu_sibling_map[cpu]);
801 				cpu_set(cpu, cpu_sibling_map[i]);
802 			}
803 		}
804 	}
805 }
806 
807 int __devinit
808 __cpu_up (unsigned int cpu)
809 {
810 	int ret;
811 	int sapicid;
812 
813 	sapicid = ia64_cpu_to_sapicid[cpu];
814 	if (sapicid == -1)
815 		return -EINVAL;
816 
817 	/*
818 	 * Already booted cpu? not valid anymore since we dont
819 	 * do idle loop tightspin anymore.
820 	 */
821 	if (cpu_isset(cpu, cpu_callin_map))
822 		return -EINVAL;
823 
824 	per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
825 	/* Processor goes to start_secondary(), sets online flag */
826 	ret = do_boot_cpu(sapicid, cpu);
827 	if (ret < 0)
828 		return ret;
829 
830 	if (cpu_data(cpu)->threads_per_core == 1 &&
831 	    cpu_data(cpu)->cores_per_socket == 1) {
832 		cpu_set(cpu, cpu_sibling_map[cpu]);
833 		cpu_set(cpu, cpu_core_map[cpu]);
834 		return 0;
835 	}
836 
837 	set_cpu_sibling_map(cpu);
838 
839 	return 0;
840 }
841 
842 /*
843  * Assume that CPUs have been discovered by some platform-dependent interface.  For
844  * SoftSDV/Lion, that would be ACPI.
845  *
846  * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
847  */
848 void __init
849 init_smp_config(void)
850 {
851 	struct fptr {
852 		unsigned long fp;
853 		unsigned long gp;
854 	} *ap_startup;
855 	long sal_ret;
856 
857 	/* Tell SAL where to drop the APs.  */
858 	ap_startup = (struct fptr *) start_ap;
859 	sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
860 				       ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
861 	if (sal_ret < 0)
862 		printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
863 		       ia64_sal_strerror(sal_ret));
864 }
865 
866 /*
867  * identify_siblings(cpu) gets called from identify_cpu. This populates the
868  * information related to logical execution units in per_cpu_data structure.
869  */
870 void __devinit
871 identify_siblings(struct cpuinfo_ia64 *c)
872 {
873 	s64 status;
874 	u16 pltid;
875 	pal_logical_to_physical_t info;
876 
877 	if (smp_num_cpucores == 1 && smp_num_siblings == 1)
878 		return;
879 
880 	if ((status = ia64_pal_logical_to_phys(-1, &info)) != PAL_STATUS_SUCCESS) {
881 		printk(KERN_ERR "ia64_pal_logical_to_phys failed with %ld\n",
882 		       status);
883 		return;
884 	}
885 	if ((status = ia64_sal_physical_id_info(&pltid)) != PAL_STATUS_SUCCESS) {
886 		printk(KERN_ERR "ia64_sal_pltid failed with %ld\n", status);
887 		return;
888 	}
889 
890 	c->socket_id =  (pltid << 8) | info.overview_ppid;
891 	c->cores_per_socket = info.overview_cpp;
892 	c->threads_per_core = info.overview_tpc;
893 	c->num_log = info.overview_num_log;
894 
895 	c->core_id = info.log1_cid;
896 	c->thread_id = info.log1_tid;
897 }
898 
899 /*
900  * returns non zero, if multi-threading is enabled
901  * on at least one physical package. Due to hotplug cpu
902  * and (maxcpus=), all threads may not necessarily be enabled
903  * even though the processor supports multi-threading.
904  */
905 int is_multithreading_enabled(void)
906 {
907 	int i, j;
908 
909 	for_each_present_cpu(i) {
910 		for_each_present_cpu(j) {
911 			if (j == i)
912 				continue;
913 			if ((cpu_data(j)->socket_id == cpu_data(i)->socket_id)) {
914 				if (cpu_data(j)->core_id == cpu_data(i)->core_id)
915 					return 1;
916 			}
917 		}
918 	}
919 	return 0;
920 }
921 EXPORT_SYMBOL_GPL(is_multithreading_enabled);
922