Merge http://ftp.arm.linux.org.uk/pub/linux/arm/kernel/git-cur/linux-2.6-arm into...
[linux-2.6] / arch / ia64 / kernel / smpboot.c
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/paravirt.h>
54 #include <asm/pgalloc.h>
55 #include <asm/pgtable.h>
56 #include <asm/processor.h>
57 #include <asm/ptrace.h>
58 #include <asm/sal.h>
59 #include <asm/system.h>
60 #include <asm/tlbflush.h>
61 #include <asm/unistd.h>
62 #include <asm/sn/arch.h>
63
64 #define SMP_DEBUG 0
65
66 #if SMP_DEBUG
67 #define Dprintk(x...)  printk(x)
68 #else
69 #define Dprintk(x...)
70 #endif
71
72 #ifdef CONFIG_HOTPLUG_CPU
73 #ifdef CONFIG_PERMIT_BSP_REMOVE
74 #define bsp_remove_ok   1
75 #else
76 #define bsp_remove_ok   0
77 #endif
78
79 /*
80  * Store all idle threads, this can be reused instead of creating
81  * a new thread. Also avoids complicated thread destroy functionality
82  * for idle threads.
83  */
84 struct task_struct *idle_thread_array[NR_CPUS];
85
86 /*
87  * Global array allocated for NR_CPUS at boot time
88  */
89 struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
90
91 /*
92  * start_ap in head.S uses this to store current booting cpu
93  * info.
94  */
95 struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
96
97 #define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
98
99 #define get_idle_for_cpu(x)             (idle_thread_array[(x)])
100 #define set_idle_for_cpu(x,p)   (idle_thread_array[(x)] = (p))
101
102 #else
103
104 #define get_idle_for_cpu(x)             (NULL)
105 #define set_idle_for_cpu(x,p)
106 #define set_brendez_area(x)
107 #endif
108
109
110 /*
111  * ITC synchronization related stuff:
112  */
113 #define MASTER  (0)
114 #define SLAVE   (SMP_CACHE_BYTES/8)
115
116 #define NUM_ROUNDS      64      /* magic value */
117 #define NUM_ITERS       5       /* likewise */
118
119 static DEFINE_SPINLOCK(itc_sync_lock);
120 static volatile unsigned long go[SLAVE + 1];
121
122 #define DEBUG_ITC_SYNC  0
123
124 extern void start_ap (void);
125 extern unsigned long ia64_iobase;
126
127 struct task_struct *task_for_booting_cpu;
128
129 /*
130  * State for each CPU
131  */
132 DEFINE_PER_CPU(int, cpu_state);
133
134 /* Bitmasks of currently online, and possible CPUs */
135 cpumask_t cpu_online_map;
136 EXPORT_SYMBOL(cpu_online_map);
137 cpumask_t cpu_possible_map = CPU_MASK_NONE;
138 EXPORT_SYMBOL(cpu_possible_map);
139
140 cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
141 DEFINE_PER_CPU_SHARED_ALIGNED(cpumask_t, cpu_sibling_map);
142 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
143
144 int smp_num_siblings = 1;
145
146 /* which logical CPU number maps to which CPU (physical APIC ID) */
147 volatile int ia64_cpu_to_sapicid[NR_CPUS];
148 EXPORT_SYMBOL(ia64_cpu_to_sapicid);
149
150 static volatile cpumask_t cpu_callin_map;
151
152 struct smp_boot_data smp_boot_data __initdata;
153
154 unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
155
156 char __initdata no_int_routing;
157
158 unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
159
160 #ifdef CONFIG_FORCE_CPEI_RETARGET
161 #define CPEI_OVERRIDE_DEFAULT   (1)
162 #else
163 #define CPEI_OVERRIDE_DEFAULT   (0)
164 #endif
165
166 unsigned int force_cpei_retarget = CPEI_OVERRIDE_DEFAULT;
167
168 static int __init
169 cmdl_force_cpei(char *str)
170 {
171         int value=0;
172
173         get_option (&str, &value);
174         force_cpei_retarget = value;
175
176         return 1;
177 }
178
179 __setup("force_cpei=", cmdl_force_cpei);
180
181 static int __init
182 nointroute (char *str)
183 {
184         no_int_routing = 1;
185         printk ("no_int_routing on\n");
186         return 1;
187 }
188
189 __setup("nointroute", nointroute);
190
191 static void fix_b0_for_bsp(void)
192 {
193 #ifdef CONFIG_HOTPLUG_CPU
194         int cpuid;
195         static int fix_bsp_b0 = 1;
196
197         cpuid = smp_processor_id();
198
199         /*
200          * Cache the b0 value on the first AP that comes up
201          */
202         if (!(fix_bsp_b0 && cpuid))
203                 return;
204
205         sal_boot_rendez_state[0].br[0] = sal_boot_rendez_state[cpuid].br[0];
206         printk ("Fixed BSP b0 value from CPU %d\n", cpuid);
207
208         fix_bsp_b0 = 0;
209 #endif
210 }
211
212 void
213 sync_master (void *arg)
214 {
215         unsigned long flags, i;
216
217         go[MASTER] = 0;
218
219         local_irq_save(flags);
220         {
221                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
222                         while (!go[MASTER])
223                                 cpu_relax();
224                         go[MASTER] = 0;
225                         go[SLAVE] = ia64_get_itc();
226                 }
227         }
228         local_irq_restore(flags);
229 }
230
231 /*
232  * Return the number of cycles by which our itc differs from the itc on the master
233  * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
234  * negative that it is behind.
235  */
236 static inline long
237 get_delta (long *rt, long *master)
238 {
239         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
240         unsigned long tcenter, t0, t1, tm;
241         long i;
242
243         for (i = 0; i < NUM_ITERS; ++i) {
244                 t0 = ia64_get_itc();
245                 go[MASTER] = 1;
246                 while (!(tm = go[SLAVE]))
247                         cpu_relax();
248                 go[SLAVE] = 0;
249                 t1 = ia64_get_itc();
250
251                 if (t1 - t0 < best_t1 - best_t0)
252                         best_t0 = t0, best_t1 = t1, best_tm = tm;
253         }
254
255         *rt = best_t1 - best_t0;
256         *master = best_tm - best_t0;
257
258         /* average best_t0 and best_t1 without overflow: */
259         tcenter = (best_t0/2 + best_t1/2);
260         if (best_t0 % 2 + best_t1 % 2 == 2)
261                 ++tcenter;
262         return tcenter - best_tm;
263 }
264
265 /*
266  * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
267  * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
268  * unaccounted-for errors (such as getting a machine check in the middle of a calibration
269  * step).  The basic idea is for the slave to ask the master what itc value it has and to
270  * read its own itc before and after the master responds.  Each iteration gives us three
271  * timestamps:
272  *
273  *      slave           master
274  *
275  *      t0 ---\
276  *             ---\
277  *                 --->
278  *                      tm
279  *                 /---
280  *             /---
281  *      t1 <---
282  *
283  *
284  * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
285  * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
286  * between the slave and the master is symmetric.  Even if the interconnect were
287  * asymmetric, we would still know that the synchronization error is smaller than the
288  * roundtrip latency (t0 - t1).
289  *
290  * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
291  * within one or two cycles.  However, we can only *guarantee* that the synchronization is
292  * accurate to within a round-trip time, which is typically in the range of several
293  * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
294  * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
295  * than half a micro second or so.
296  */
297 void
298 ia64_sync_itc (unsigned int master)
299 {
300         long i, delta, adj, adjust_latency = 0, done = 0;
301         unsigned long flags, rt, master_time_stamp, bound;
302 #if DEBUG_ITC_SYNC
303         struct {
304                 long rt;        /* roundtrip time */
305                 long master;    /* master's timestamp */
306                 long diff;      /* difference between midpoint and master's timestamp */
307                 long lat;       /* estimate of itc adjustment latency */
308         } t[NUM_ROUNDS];
309 #endif
310
311         /*
312          * Make sure local timer ticks are disabled while we sync.  If
313          * they were enabled, we'd have to worry about nasty issues
314          * like setting the ITC ahead of (or a long time before) the
315          * next scheduled tick.
316          */
317         BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
318
319         go[MASTER] = 1;
320
321         if (smp_call_function_single(master, sync_master, NULL, 0) < 0) {
322                 printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
323                 return;
324         }
325
326         while (go[MASTER])
327                 cpu_relax();    /* wait for master to be ready */
328
329         spin_lock_irqsave(&itc_sync_lock, flags);
330         {
331                 for (i = 0; i < NUM_ROUNDS; ++i) {
332                         delta = get_delta(&rt, &master_time_stamp);
333                         if (delta == 0) {
334                                 done = 1;       /* let's lock on to this... */
335                                 bound = rt;
336                         }
337
338                         if (!done) {
339                                 if (i > 0) {
340                                         adjust_latency += -delta;
341                                         adj = -delta + adjust_latency/4;
342                                 } else
343                                         adj = -delta;
344
345                                 ia64_set_itc(ia64_get_itc() + adj);
346                         }
347 #if DEBUG_ITC_SYNC
348                         t[i].rt = rt;
349                         t[i].master = master_time_stamp;
350                         t[i].diff = delta;
351                         t[i].lat = adjust_latency/4;
352 #endif
353                 }
354         }
355         spin_unlock_irqrestore(&itc_sync_lock, flags);
356
357 #if DEBUG_ITC_SYNC
358         for (i = 0; i < NUM_ROUNDS; ++i)
359                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
360                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
361 #endif
362
363         printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
364                "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
365 }
366
367 /*
368  * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
369  */
370 static inline void __devinit
371 smp_setup_percpu_timer (void)
372 {
373 }
374
375 static void __cpuinit
376 smp_callin (void)
377 {
378         int cpuid, phys_id, itc_master;
379         struct cpuinfo_ia64 *last_cpuinfo, *this_cpuinfo;
380         extern void ia64_init_itm(void);
381         extern volatile int time_keeper_id;
382
383 #ifdef CONFIG_PERFMON
384         extern void pfm_init_percpu(void);
385 #endif
386
387         cpuid = smp_processor_id();
388         phys_id = hard_smp_processor_id();
389         itc_master = time_keeper_id;
390
391         if (cpu_online(cpuid)) {
392                 printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
393                        phys_id, cpuid);
394                 BUG();
395         }
396
397         fix_b0_for_bsp();
398
399         ipi_call_lock_irq();
400         spin_lock(&vector_lock);
401         /* Setup the per cpu irq handling data structures */
402         __setup_vector_irq(cpuid);
403         cpu_set(cpuid, cpu_online_map);
404         per_cpu(cpu_state, cpuid) = CPU_ONLINE;
405         spin_unlock(&vector_lock);
406         ipi_call_unlock_irq();
407
408         smp_setup_percpu_timer();
409
410         ia64_mca_cmc_vector_setup();    /* Setup vector on AP */
411
412 #ifdef CONFIG_PERFMON
413         pfm_init_percpu();
414 #endif
415
416         local_irq_enable();
417
418         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
419                 /*
420                  * Synchronize the ITC with the BP.  Need to do this after irqs are
421                  * enabled because ia64_sync_itc() calls smp_call_function_single(), which
422                  * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
423                  * local_bh_enable(), which bugs out if irqs are not enabled...
424                  */
425                 Dprintk("Going to syncup ITC with ITC Master.\n");
426                 ia64_sync_itc(itc_master);
427         }
428
429         /*
430          * Get our bogomips.
431          */
432         ia64_init_itm();
433
434         /*
435          * Delay calibration can be skipped if new processor is identical to the
436          * previous processor.
437          */
438         last_cpuinfo = cpu_data(cpuid - 1);
439         this_cpuinfo = local_cpu_data;
440         if (last_cpuinfo->itc_freq != this_cpuinfo->itc_freq ||
441             last_cpuinfo->proc_freq != this_cpuinfo->proc_freq ||
442             last_cpuinfo->features != this_cpuinfo->features ||
443             last_cpuinfo->revision != this_cpuinfo->revision ||
444             last_cpuinfo->family != this_cpuinfo->family ||
445             last_cpuinfo->archrev != this_cpuinfo->archrev ||
446             last_cpuinfo->model != this_cpuinfo->model)
447                 calibrate_delay();
448         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
449
450 #ifdef CONFIG_IA32_SUPPORT
451         ia32_gdt_init();
452 #endif
453
454         /*
455          * Allow the master to continue.
456          */
457         cpu_set(cpuid, cpu_callin_map);
458         Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
459 }
460
461
462 /*
463  * Activate a secondary processor.  head.S calls this.
464  */
465 int __cpuinit
466 start_secondary (void *unused)
467 {
468         /* Early console may use I/O ports */
469         ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
470         Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
471         efi_map_pal_code();
472         cpu_init();
473         preempt_disable();
474         smp_callin();
475
476         cpu_idle();
477         return 0;
478 }
479
480 struct pt_regs * __cpuinit idle_regs(struct pt_regs *regs)
481 {
482         return NULL;
483 }
484
485 struct create_idle {
486         struct work_struct work;
487         struct task_struct *idle;
488         struct completion done;
489         int cpu;
490 };
491
492 void __cpuinit
493 do_fork_idle(struct work_struct *work)
494 {
495         struct create_idle *c_idle =
496                 container_of(work, struct create_idle, work);
497
498         c_idle->idle = fork_idle(c_idle->cpu);
499         complete(&c_idle->done);
500 }
501
502 static int __cpuinit
503 do_boot_cpu (int sapicid, int cpu)
504 {
505         int timeout;
506         struct create_idle c_idle = {
507                 .work = __WORK_INITIALIZER(c_idle.work, do_fork_idle),
508                 .cpu    = cpu,
509                 .done   = COMPLETION_INITIALIZER(c_idle.done),
510         };
511
512         c_idle.idle = get_idle_for_cpu(cpu);
513         if (c_idle.idle) {
514                 init_idle(c_idle.idle, cpu);
515                 goto do_rest;
516         }
517
518         /*
519          * We can't use kernel_thread since we must avoid to reschedule the child.
520          */
521         if (!keventd_up() || current_is_keventd())
522                 c_idle.work.func(&c_idle.work);
523         else {
524                 schedule_work(&c_idle.work);
525                 wait_for_completion(&c_idle.done);
526         }
527
528         if (IS_ERR(c_idle.idle))
529                 panic("failed fork for CPU %d", cpu);
530
531         set_idle_for_cpu(cpu, c_idle.idle);
532
533 do_rest:
534         task_for_booting_cpu = c_idle.idle;
535
536         Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
537
538         set_brendez_area(cpu);
539         platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
540
541         /*
542          * Wait 10s total for the AP to start
543          */
544         Dprintk("Waiting on callin_map ...");
545         for (timeout = 0; timeout < 100000; timeout++) {
546                 if (cpu_isset(cpu, cpu_callin_map))
547                         break;  /* It has booted */
548                 udelay(100);
549         }
550         Dprintk("\n");
551
552         if (!cpu_isset(cpu, cpu_callin_map)) {
553                 printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
554                 ia64_cpu_to_sapicid[cpu] = -1;
555                 cpu_clear(cpu, cpu_online_map);  /* was set in smp_callin() */
556                 return -EINVAL;
557         }
558         return 0;
559 }
560
561 static int __init
562 decay (char *str)
563 {
564         int ticks;
565         get_option (&str, &ticks);
566         return 1;
567 }
568
569 __setup("decay=", decay);
570
571 /*
572  * Initialize the logical CPU number to SAPICID mapping
573  */
574 void __init
575 smp_build_cpu_map (void)
576 {
577         int sapicid, cpu, i;
578         int boot_cpu_id = hard_smp_processor_id();
579
580         for (cpu = 0; cpu < NR_CPUS; cpu++) {
581                 ia64_cpu_to_sapicid[cpu] = -1;
582         }
583
584         ia64_cpu_to_sapicid[0] = boot_cpu_id;
585         cpus_clear(cpu_present_map);
586         cpu_set(0, cpu_present_map);
587         cpu_set(0, cpu_possible_map);
588         for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
589                 sapicid = smp_boot_data.cpu_phys_id[i];
590                 if (sapicid == boot_cpu_id)
591                         continue;
592                 cpu_set(cpu, cpu_present_map);
593                 cpu_set(cpu, cpu_possible_map);
594                 ia64_cpu_to_sapicid[cpu] = sapicid;
595                 cpu++;
596         }
597 }
598
599 /*
600  * Cycle through the APs sending Wakeup IPIs to boot each.
601  */
602 void __init
603 smp_prepare_cpus (unsigned int max_cpus)
604 {
605         int boot_cpu_id = hard_smp_processor_id();
606
607         /*
608          * Initialize the per-CPU profiling counter/multiplier
609          */
610
611         smp_setup_percpu_timer();
612
613         /*
614          * We have the boot CPU online for sure.
615          */
616         cpu_set(0, cpu_online_map);
617         cpu_set(0, cpu_callin_map);
618
619         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
620         ia64_cpu_to_sapicid[0] = boot_cpu_id;
621
622         printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
623
624         current_thread_info()->cpu = 0;
625
626         /*
627          * If SMP should be disabled, then really disable it!
628          */
629         if (!max_cpus) {
630                 printk(KERN_INFO "SMP mode deactivated.\n");
631                 cpus_clear(cpu_online_map);
632                 cpus_clear(cpu_present_map);
633                 cpus_clear(cpu_possible_map);
634                 cpu_set(0, cpu_online_map);
635                 cpu_set(0, cpu_present_map);
636                 cpu_set(0, cpu_possible_map);
637                 return;
638         }
639 }
640
641 void __devinit smp_prepare_boot_cpu(void)
642 {
643         cpu_set(smp_processor_id(), cpu_online_map);
644         cpu_set(smp_processor_id(), cpu_callin_map);
645         per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
646         paravirt_post_smp_prepare_boot_cpu();
647 }
648
649 #ifdef CONFIG_HOTPLUG_CPU
650 static inline void
651 clear_cpu_sibling_map(int cpu)
652 {
653         int i;
654
655         for_each_cpu_mask(i, per_cpu(cpu_sibling_map, cpu))
656                 cpu_clear(cpu, per_cpu(cpu_sibling_map, i));
657         for_each_cpu_mask(i, cpu_core_map[cpu])
658                 cpu_clear(cpu, cpu_core_map[i]);
659
660         per_cpu(cpu_sibling_map, cpu) = cpu_core_map[cpu] = CPU_MASK_NONE;
661 }
662
663 static void
664 remove_siblinginfo(int cpu)
665 {
666         int last = 0;
667
668         if (cpu_data(cpu)->threads_per_core == 1 &&
669             cpu_data(cpu)->cores_per_socket == 1) {
670                 cpu_clear(cpu, cpu_core_map[cpu]);
671                 cpu_clear(cpu, per_cpu(cpu_sibling_map, cpu));
672                 return;
673         }
674
675         last = (cpus_weight(cpu_core_map[cpu]) == 1 ? 1 : 0);
676
677         /* remove it from all sibling map's */
678         clear_cpu_sibling_map(cpu);
679 }
680
681 extern void fixup_irqs(void);
682
683 int migrate_platform_irqs(unsigned int cpu)
684 {
685         int new_cpei_cpu;
686         irq_desc_t *desc = NULL;
687         cpumask_t       mask;
688         int             retval = 0;
689
690         /*
691          * dont permit CPEI target to removed.
692          */
693         if (cpe_vector > 0 && is_cpu_cpei_target(cpu)) {
694                 printk ("CPU (%d) is CPEI Target\n", cpu);
695                 if (can_cpei_retarget()) {
696                         /*
697                          * Now re-target the CPEI to a different processor
698                          */
699                         new_cpei_cpu = any_online_cpu(cpu_online_map);
700                         mask = cpumask_of_cpu(new_cpei_cpu);
701                         set_cpei_target_cpu(new_cpei_cpu);
702                         desc = irq_desc + ia64_cpe_irq;
703                         /*
704                          * Switch for now, immediately, we need to do fake intr
705                          * as other interrupts, but need to study CPEI behaviour with
706                          * polling before making changes.
707                          */
708                         if (desc) {
709                                 desc->chip->disable(ia64_cpe_irq);
710                                 desc->chip->set_affinity(ia64_cpe_irq, mask);
711                                 desc->chip->enable(ia64_cpe_irq);
712                                 printk ("Re-targetting CPEI to cpu %d\n", new_cpei_cpu);
713                         }
714                 }
715                 if (!desc) {
716                         printk ("Unable to retarget CPEI, offline cpu [%d] failed\n", cpu);
717                         retval = -EBUSY;
718                 }
719         }
720         return retval;
721 }
722
723 /* must be called with cpucontrol mutex held */
724 int __cpu_disable(void)
725 {
726         int cpu = smp_processor_id();
727
728         /*
729          * dont permit boot processor for now
730          */
731         if (cpu == 0 && !bsp_remove_ok) {
732                 printk ("Your platform does not support removal of BSP\n");
733                 return (-EBUSY);
734         }
735
736         if (ia64_platform_is("sn2")) {
737                 if (!sn_cpu_disable_allowed(cpu))
738                         return -EBUSY;
739         }
740
741         cpu_clear(cpu, cpu_online_map);
742
743         if (migrate_platform_irqs(cpu)) {
744                 cpu_set(cpu, cpu_online_map);
745                 return (-EBUSY);
746         }
747
748         remove_siblinginfo(cpu);
749         cpu_clear(cpu, cpu_online_map);
750         fixup_irqs();
751         local_flush_tlb_all();
752         cpu_clear(cpu, cpu_callin_map);
753         return 0;
754 }
755
756 void __cpu_die(unsigned int cpu)
757 {
758         unsigned int i;
759
760         for (i = 0; i < 100; i++) {
761                 /* They ack this in play_dead by setting CPU_DEAD */
762                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
763                 {
764                         printk ("CPU %d is now offline\n", cpu);
765                         return;
766                 }
767                 msleep(100);
768         }
769         printk(KERN_ERR "CPU %u didn't die...\n", cpu);
770 }
771 #endif /* CONFIG_HOTPLUG_CPU */
772
773 void
774 smp_cpus_done (unsigned int dummy)
775 {
776         int cpu;
777         unsigned long bogosum = 0;
778
779         /*
780          * Allow the user to impress friends.
781          */
782
783         for_each_online_cpu(cpu) {
784                 bogosum += cpu_data(cpu)->loops_per_jiffy;
785         }
786
787         printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
788                (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
789 }
790
791 static inline void __devinit
792 set_cpu_sibling_map(int cpu)
793 {
794         int i;
795
796         for_each_online_cpu(i) {
797                 if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
798                         cpu_set(i, cpu_core_map[cpu]);
799                         cpu_set(cpu, cpu_core_map[i]);
800                         if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
801                                 cpu_set(i, per_cpu(cpu_sibling_map, cpu));
802                                 cpu_set(cpu, per_cpu(cpu_sibling_map, i));
803                         }
804                 }
805         }
806 }
807
808 int __cpuinit
809 __cpu_up (unsigned int cpu)
810 {
811         int ret;
812         int sapicid;
813
814         sapicid = ia64_cpu_to_sapicid[cpu];
815         if (sapicid == -1)
816                 return -EINVAL;
817
818         /*
819          * Already booted cpu? not valid anymore since we dont
820          * do idle loop tightspin anymore.
821          */
822         if (cpu_isset(cpu, cpu_callin_map))
823                 return -EINVAL;
824
825         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
826         /* Processor goes to start_secondary(), sets online flag */
827         ret = do_boot_cpu(sapicid, cpu);
828         if (ret < 0)
829                 return ret;
830
831         if (cpu_data(cpu)->threads_per_core == 1 &&
832             cpu_data(cpu)->cores_per_socket == 1) {
833                 cpu_set(cpu, per_cpu(cpu_sibling_map, cpu));
834                 cpu_set(cpu, cpu_core_map[cpu]);
835                 return 0;
836         }
837
838         set_cpu_sibling_map(cpu);
839
840         return 0;
841 }
842
843 /*
844  * Assume that CPUs have been discovered by some platform-dependent interface.  For
845  * SoftSDV/Lion, that would be ACPI.
846  *
847  * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
848  */
849 void __init
850 init_smp_config(void)
851 {
852         struct fptr {
853                 unsigned long fp;
854                 unsigned long gp;
855         } *ap_startup;
856         long sal_ret;
857
858         /* Tell SAL where to drop the APs.  */
859         ap_startup = (struct fptr *) start_ap;
860         sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
861                                        ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
862         if (sal_ret < 0)
863                 printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
864                        ia64_sal_strerror(sal_ret));
865 }
866
867 /*
868  * identify_siblings(cpu) gets called from identify_cpu. This populates the 
869  * information related to logical execution units in per_cpu_data structure.
870  */
871 void __devinit
872 identify_siblings(struct cpuinfo_ia64 *c)
873 {
874         s64 status;
875         u16 pltid;
876         pal_logical_to_physical_t info;
877
878         status = ia64_pal_logical_to_phys(-1, &info);
879         if (status != PAL_STATUS_SUCCESS) {
880                 if (status != PAL_STATUS_UNIMPLEMENTED) {
881                         printk(KERN_ERR
882                                 "ia64_pal_logical_to_phys failed with %ld\n",
883                                 status);
884                         return;
885                 }
886
887                 info.overview_ppid = 0;
888                 info.overview_cpp  = 1;
889                 info.overview_tpc  = 1;
890         }
891
892         status = ia64_sal_physical_id_info(&pltid);
893         if (status != PAL_STATUS_SUCCESS) {
894                 if (status != PAL_STATUS_UNIMPLEMENTED)
895                         printk(KERN_ERR
896                                 "ia64_sal_pltid failed with %ld\n",
897                                 status);
898                 return;
899         }
900
901         c->socket_id =  (pltid << 8) | info.overview_ppid;
902
903         if (info.overview_cpp == 1 && info.overview_tpc == 1)
904                 return;
905
906         c->cores_per_socket = info.overview_cpp;
907         c->threads_per_core = info.overview_tpc;
908         c->num_log = info.overview_num_log;
909
910         c->core_id = info.log1_cid;
911         c->thread_id = info.log1_tid;
912 }
913
914 /*
915  * returns non zero, if multi-threading is enabled
916  * on at least one physical package. Due to hotplug cpu
917  * and (maxcpus=), all threads may not necessarily be enabled
918  * even though the processor supports multi-threading.
919  */
920 int is_multithreading_enabled(void)
921 {
922         int i, j;
923
924         for_each_present_cpu(i) {
925                 for_each_present_cpu(j) {
926                         if (j == i)
927                                 continue;
928                         if ((cpu_data(j)->socket_id == cpu_data(i)->socket_id)) {
929                                 if (cpu_data(j)->core_id == cpu_data(i)->core_id)
930                                         return 1;
931                         }
932                 }
933         }
934         return 0;
935 }
936 EXPORT_SYMBOL_GPL(is_multithreading_enabled);