Merge branch 'for_rmk' of git://git.mnementh.co.uk/linux-2.6-im
[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 EXPORT_SYMBOL(cpu_core_map);
142 DEFINE_PER_CPU_SHARED_ALIGNED(cpumask_t, cpu_sibling_map);
143 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
144
145 int smp_num_siblings = 1;
146
147 /* which logical CPU number maps to which CPU (physical APIC ID) */
148 volatile int ia64_cpu_to_sapicid[NR_CPUS];
149 EXPORT_SYMBOL(ia64_cpu_to_sapicid);
150
151 static volatile cpumask_t cpu_callin_map;
152
153 struct smp_boot_data smp_boot_data __initdata;
154
155 unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
156
157 char __initdata no_int_routing;
158
159 unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
160
161 #ifdef CONFIG_FORCE_CPEI_RETARGET
162 #define CPEI_OVERRIDE_DEFAULT   (1)
163 #else
164 #define CPEI_OVERRIDE_DEFAULT   (0)
165 #endif
166
167 unsigned int force_cpei_retarget = CPEI_OVERRIDE_DEFAULT;
168
169 static int __init
170 cmdl_force_cpei(char *str)
171 {
172         int value=0;
173
174         get_option (&str, &value);
175         force_cpei_retarget = value;
176
177         return 1;
178 }
179
180 __setup("force_cpei=", cmdl_force_cpei);
181
182 static int __init
183 nointroute (char *str)
184 {
185         no_int_routing = 1;
186         printk ("no_int_routing on\n");
187         return 1;
188 }
189
190 __setup("nointroute", nointroute);
191
192 static void fix_b0_for_bsp(void)
193 {
194 #ifdef CONFIG_HOTPLUG_CPU
195         int cpuid;
196         static int fix_bsp_b0 = 1;
197
198         cpuid = smp_processor_id();
199
200         /*
201          * Cache the b0 value on the first AP that comes up
202          */
203         if (!(fix_bsp_b0 && cpuid))
204                 return;
205
206         sal_boot_rendez_state[0].br[0] = sal_boot_rendez_state[cpuid].br[0];
207         printk ("Fixed BSP b0 value from CPU %d\n", cpuid);
208
209         fix_bsp_b0 = 0;
210 #endif
211 }
212
213 void
214 sync_master (void *arg)
215 {
216         unsigned long flags, i;
217
218         go[MASTER] = 0;
219
220         local_irq_save(flags);
221         {
222                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
223                         while (!go[MASTER])
224                                 cpu_relax();
225                         go[MASTER] = 0;
226                         go[SLAVE] = ia64_get_itc();
227                 }
228         }
229         local_irq_restore(flags);
230 }
231
232 /*
233  * Return the number of cycles by which our itc differs from the itc on the master
234  * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
235  * negative that it is behind.
236  */
237 static inline long
238 get_delta (long *rt, long *master)
239 {
240         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
241         unsigned long tcenter, t0, t1, tm;
242         long i;
243
244         for (i = 0; i < NUM_ITERS; ++i) {
245                 t0 = ia64_get_itc();
246                 go[MASTER] = 1;
247                 while (!(tm = go[SLAVE]))
248                         cpu_relax();
249                 go[SLAVE] = 0;
250                 t1 = ia64_get_itc();
251
252                 if (t1 - t0 < best_t1 - best_t0)
253                         best_t0 = t0, best_t1 = t1, best_tm = tm;
254         }
255
256         *rt = best_t1 - best_t0;
257         *master = best_tm - best_t0;
258
259         /* average best_t0 and best_t1 without overflow: */
260         tcenter = (best_t0/2 + best_t1/2);
261         if (best_t0 % 2 + best_t1 % 2 == 2)
262                 ++tcenter;
263         return tcenter - best_tm;
264 }
265
266 /*
267  * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
268  * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
269  * unaccounted-for errors (such as getting a machine check in the middle of a calibration
270  * step).  The basic idea is for the slave to ask the master what itc value it has and to
271  * read its own itc before and after the master responds.  Each iteration gives us three
272  * timestamps:
273  *
274  *      slave           master
275  *
276  *      t0 ---\
277  *             ---\
278  *                 --->
279  *                      tm
280  *                 /---
281  *             /---
282  *      t1 <---
283  *
284  *
285  * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
286  * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
287  * between the slave and the master is symmetric.  Even if the interconnect were
288  * asymmetric, we would still know that the synchronization error is smaller than the
289  * roundtrip latency (t0 - t1).
290  *
291  * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
292  * within one or two cycles.  However, we can only *guarantee* that the synchronization is
293  * accurate to within a round-trip time, which is typically in the range of several
294  * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
295  * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
296  * than half a micro second or so.
297  */
298 void
299 ia64_sync_itc (unsigned int master)
300 {
301         long i, delta, adj, adjust_latency = 0, done = 0;
302         unsigned long flags, rt, master_time_stamp, bound;
303 #if DEBUG_ITC_SYNC
304         struct {
305                 long rt;        /* roundtrip time */
306                 long master;    /* master's timestamp */
307                 long diff;      /* difference between midpoint and master's timestamp */
308                 long lat;       /* estimate of itc adjustment latency */
309         } t[NUM_ROUNDS];
310 #endif
311
312         /*
313          * Make sure local timer ticks are disabled while we sync.  If
314          * they were enabled, we'd have to worry about nasty issues
315          * like setting the ITC ahead of (or a long time before) the
316          * next scheduled tick.
317          */
318         BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
319
320         go[MASTER] = 1;
321
322         if (smp_call_function_single(master, sync_master, NULL, 0) < 0) {
323                 printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
324                 return;
325         }
326
327         while (go[MASTER])
328                 cpu_relax();    /* wait for master to be ready */
329
330         spin_lock_irqsave(&itc_sync_lock, flags);
331         {
332                 for (i = 0; i < NUM_ROUNDS; ++i) {
333                         delta = get_delta(&rt, &master_time_stamp);
334                         if (delta == 0) {
335                                 done = 1;       /* let's lock on to this... */
336                                 bound = rt;
337                         }
338
339                         if (!done) {
340                                 if (i > 0) {
341                                         adjust_latency += -delta;
342                                         adj = -delta + adjust_latency/4;
343                                 } else
344                                         adj = -delta;
345
346                                 ia64_set_itc(ia64_get_itc() + adj);
347                         }
348 #if DEBUG_ITC_SYNC
349                         t[i].rt = rt;
350                         t[i].master = master_time_stamp;
351                         t[i].diff = delta;
352                         t[i].lat = adjust_latency/4;
353 #endif
354                 }
355         }
356         spin_unlock_irqrestore(&itc_sync_lock, flags);
357
358 #if DEBUG_ITC_SYNC
359         for (i = 0; i < NUM_ROUNDS; ++i)
360                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
361                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
362 #endif
363
364         printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
365                "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
366 }
367
368 /*
369  * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
370  */
371 static inline void __devinit
372 smp_setup_percpu_timer (void)
373 {
374 }
375
376 static void __cpuinit
377 smp_callin (void)
378 {
379         int cpuid, phys_id, itc_master;
380         struct cpuinfo_ia64 *last_cpuinfo, *this_cpuinfo;
381         extern void ia64_init_itm(void);
382         extern volatile int time_keeper_id;
383
384 #ifdef CONFIG_PERFMON
385         extern void pfm_init_percpu(void);
386 #endif
387
388         cpuid = smp_processor_id();
389         phys_id = hard_smp_processor_id();
390         itc_master = time_keeper_id;
391
392         if (cpu_online(cpuid)) {
393                 printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
394                        phys_id, cpuid);
395                 BUG();
396         }
397
398         fix_b0_for_bsp();
399
400         ipi_call_lock_irq();
401         spin_lock(&vector_lock);
402         /* Setup the per cpu irq handling data structures */
403         __setup_vector_irq(cpuid);
404         cpu_set(cpuid, cpu_online_map);
405         per_cpu(cpu_state, cpuid) = CPU_ONLINE;
406         spin_unlock(&vector_lock);
407         ipi_call_unlock_irq();
408
409         smp_setup_percpu_timer();
410
411         ia64_mca_cmc_vector_setup();    /* Setup vector on AP */
412
413 #ifdef CONFIG_PERFMON
414         pfm_init_percpu();
415 #endif
416
417         local_irq_enable();
418
419         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
420                 /*
421                  * Synchronize the ITC with the BP.  Need to do this after irqs are
422                  * enabled because ia64_sync_itc() calls smp_call_function_single(), which
423                  * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
424                  * local_bh_enable(), which bugs out if irqs are not enabled...
425                  */
426                 Dprintk("Going to syncup ITC with ITC Master.\n");
427                 ia64_sync_itc(itc_master);
428         }
429
430         /*
431          * Get our bogomips.
432          */
433         ia64_init_itm();
434
435         /*
436          * Delay calibration can be skipped if new processor is identical to the
437          * previous processor.
438          */
439         last_cpuinfo = cpu_data(cpuid - 1);
440         this_cpuinfo = local_cpu_data;
441         if (last_cpuinfo->itc_freq != this_cpuinfo->itc_freq ||
442             last_cpuinfo->proc_freq != this_cpuinfo->proc_freq ||
443             last_cpuinfo->features != this_cpuinfo->features ||
444             last_cpuinfo->revision != this_cpuinfo->revision ||
445             last_cpuinfo->family != this_cpuinfo->family ||
446             last_cpuinfo->archrev != this_cpuinfo->archrev ||
447             last_cpuinfo->model != this_cpuinfo->model)
448                 calibrate_delay();
449         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
450
451 #ifdef CONFIG_IA32_SUPPORT
452         ia32_gdt_init();
453 #endif
454
455         /*
456          * Allow the master to continue.
457          */
458         cpu_set(cpuid, cpu_callin_map);
459         Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
460 }
461
462
463 /*
464  * Activate a secondary processor.  head.S calls this.
465  */
466 int __cpuinit
467 start_secondary (void *unused)
468 {
469         /* Early console may use I/O ports */
470         ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
471 #ifndef CONFIG_PRINTK_TIME
472         Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
473 #endif
474         efi_map_pal_code();
475         cpu_init();
476         preempt_disable();
477         smp_callin();
478
479         cpu_idle();
480         return 0;
481 }
482
483 struct pt_regs * __cpuinit idle_regs(struct pt_regs *regs)
484 {
485         return NULL;
486 }
487
488 struct create_idle {
489         struct work_struct work;
490         struct task_struct *idle;
491         struct completion done;
492         int cpu;
493 };
494
495 void __cpuinit
496 do_fork_idle(struct work_struct *work)
497 {
498         struct create_idle *c_idle =
499                 container_of(work, struct create_idle, work);
500
501         c_idle->idle = fork_idle(c_idle->cpu);
502         complete(&c_idle->done);
503 }
504
505 static int __cpuinit
506 do_boot_cpu (int sapicid, int cpu)
507 {
508         int timeout;
509         struct create_idle c_idle = {
510                 .work = __WORK_INITIALIZER(c_idle.work, do_fork_idle),
511                 .cpu    = cpu,
512                 .done   = COMPLETION_INITIALIZER(c_idle.done),
513         };
514
515         c_idle.idle = get_idle_for_cpu(cpu);
516         if (c_idle.idle) {
517                 init_idle(c_idle.idle, cpu);
518                 goto do_rest;
519         }
520
521         /*
522          * We can't use kernel_thread since we must avoid to reschedule the child.
523          */
524         if (!keventd_up() || current_is_keventd())
525                 c_idle.work.func(&c_idle.work);
526         else {
527                 schedule_work(&c_idle.work);
528                 wait_for_completion(&c_idle.done);
529         }
530
531         if (IS_ERR(c_idle.idle))
532                 panic("failed fork for CPU %d", cpu);
533
534         set_idle_for_cpu(cpu, c_idle.idle);
535
536 do_rest:
537         task_for_booting_cpu = c_idle.idle;
538
539         Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
540
541         set_brendez_area(cpu);
542         platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
543
544         /*
545          * Wait 10s total for the AP to start
546          */
547         Dprintk("Waiting on callin_map ...");
548         for (timeout = 0; timeout < 100000; timeout++) {
549                 if (cpu_isset(cpu, cpu_callin_map))
550                         break;  /* It has booted */
551                 udelay(100);
552         }
553         Dprintk("\n");
554
555         if (!cpu_isset(cpu, cpu_callin_map)) {
556                 printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
557                 ia64_cpu_to_sapicid[cpu] = -1;
558                 cpu_clear(cpu, cpu_online_map);  /* was set in smp_callin() */
559                 return -EINVAL;
560         }
561         return 0;
562 }
563
564 static int __init
565 decay (char *str)
566 {
567         int ticks;
568         get_option (&str, &ticks);
569         return 1;
570 }
571
572 __setup("decay=", decay);
573
574 /*
575  * Initialize the logical CPU number to SAPICID mapping
576  */
577 void __init
578 smp_build_cpu_map (void)
579 {
580         int sapicid, cpu, i;
581         int boot_cpu_id = hard_smp_processor_id();
582
583         for (cpu = 0; cpu < NR_CPUS; cpu++) {
584                 ia64_cpu_to_sapicid[cpu] = -1;
585         }
586
587         ia64_cpu_to_sapicid[0] = boot_cpu_id;
588         cpus_clear(cpu_present_map);
589         cpu_set(0, cpu_present_map);
590         cpu_set(0, cpu_possible_map);
591         for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
592                 sapicid = smp_boot_data.cpu_phys_id[i];
593                 if (sapicid == boot_cpu_id)
594                         continue;
595                 cpu_set(cpu, cpu_present_map);
596                 cpu_set(cpu, cpu_possible_map);
597                 ia64_cpu_to_sapicid[cpu] = sapicid;
598                 cpu++;
599         }
600 }
601
602 /*
603  * Cycle through the APs sending Wakeup IPIs to boot each.
604  */
605 void __init
606 smp_prepare_cpus (unsigned int max_cpus)
607 {
608         int boot_cpu_id = hard_smp_processor_id();
609
610         /*
611          * Initialize the per-CPU profiling counter/multiplier
612          */
613
614         smp_setup_percpu_timer();
615
616         /*
617          * We have the boot CPU online for sure.
618          */
619         cpu_set(0, cpu_online_map);
620         cpu_set(0, cpu_callin_map);
621
622         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
623         ia64_cpu_to_sapicid[0] = boot_cpu_id;
624
625         printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
626
627         current_thread_info()->cpu = 0;
628
629         /*
630          * If SMP should be disabled, then really disable it!
631          */
632         if (!max_cpus) {
633                 printk(KERN_INFO "SMP mode deactivated.\n");
634                 cpus_clear(cpu_online_map);
635                 cpus_clear(cpu_present_map);
636                 cpus_clear(cpu_possible_map);
637                 cpu_set(0, cpu_online_map);
638                 cpu_set(0, cpu_present_map);
639                 cpu_set(0, cpu_possible_map);
640                 return;
641         }
642 }
643
644 void __devinit smp_prepare_boot_cpu(void)
645 {
646         cpu_set(smp_processor_id(), cpu_online_map);
647         cpu_set(smp_processor_id(), cpu_callin_map);
648         per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
649         paravirt_post_smp_prepare_boot_cpu();
650 }
651
652 #ifdef CONFIG_HOTPLUG_CPU
653 static inline void
654 clear_cpu_sibling_map(int cpu)
655 {
656         int i;
657
658         for_each_cpu_mask(i, per_cpu(cpu_sibling_map, cpu))
659                 cpu_clear(cpu, per_cpu(cpu_sibling_map, i));
660         for_each_cpu_mask(i, cpu_core_map[cpu])
661                 cpu_clear(cpu, cpu_core_map[i]);
662
663         per_cpu(cpu_sibling_map, cpu) = cpu_core_map[cpu] = CPU_MASK_NONE;
664 }
665
666 static void
667 remove_siblinginfo(int cpu)
668 {
669         int last = 0;
670
671         if (cpu_data(cpu)->threads_per_core == 1 &&
672             cpu_data(cpu)->cores_per_socket == 1) {
673                 cpu_clear(cpu, cpu_core_map[cpu]);
674                 cpu_clear(cpu, per_cpu(cpu_sibling_map, cpu));
675                 return;
676         }
677
678         last = (cpus_weight(cpu_core_map[cpu]) == 1 ? 1 : 0);
679
680         /* remove it from all sibling map's */
681         clear_cpu_sibling_map(cpu);
682 }
683
684 extern void fixup_irqs(void);
685
686 int migrate_platform_irqs(unsigned int cpu)
687 {
688         int new_cpei_cpu;
689         irq_desc_t *desc = NULL;
690         cpumask_t       mask;
691         int             retval = 0;
692
693         /*
694          * dont permit CPEI target to removed.
695          */
696         if (cpe_vector > 0 && is_cpu_cpei_target(cpu)) {
697                 printk ("CPU (%d) is CPEI Target\n", cpu);
698                 if (can_cpei_retarget()) {
699                         /*
700                          * Now re-target the CPEI to a different processor
701                          */
702                         new_cpei_cpu = any_online_cpu(cpu_online_map);
703                         mask = cpumask_of_cpu(new_cpei_cpu);
704                         set_cpei_target_cpu(new_cpei_cpu);
705                         desc = irq_desc + ia64_cpe_irq;
706                         /*
707                          * Switch for now, immediately, we need to do fake intr
708                          * as other interrupts, but need to study CPEI behaviour with
709                          * polling before making changes.
710                          */
711                         if (desc) {
712                                 desc->chip->disable(ia64_cpe_irq);
713                                 desc->chip->set_affinity(ia64_cpe_irq, mask);
714                                 desc->chip->enable(ia64_cpe_irq);
715                                 printk ("Re-targetting CPEI to cpu %d\n", new_cpei_cpu);
716                         }
717                 }
718                 if (!desc) {
719                         printk ("Unable to retarget CPEI, offline cpu [%d] failed\n", cpu);
720                         retval = -EBUSY;
721                 }
722         }
723         return retval;
724 }
725
726 /* must be called with cpucontrol mutex held */
727 int __cpu_disable(void)
728 {
729         int cpu = smp_processor_id();
730
731         /*
732          * dont permit boot processor for now
733          */
734         if (cpu == 0 && !bsp_remove_ok) {
735                 printk ("Your platform does not support removal of BSP\n");
736                 return (-EBUSY);
737         }
738
739         if (ia64_platform_is("sn2")) {
740                 if (!sn_cpu_disable_allowed(cpu))
741                         return -EBUSY;
742         }
743
744         cpu_clear(cpu, cpu_online_map);
745
746         if (migrate_platform_irqs(cpu)) {
747                 cpu_set(cpu, cpu_online_map);
748                 return (-EBUSY);
749         }
750
751         remove_siblinginfo(cpu);
752         cpu_clear(cpu, cpu_online_map);
753         fixup_irqs();
754         local_flush_tlb_all();
755         cpu_clear(cpu, cpu_callin_map);
756         return 0;
757 }
758
759 void __cpu_die(unsigned int cpu)
760 {
761         unsigned int i;
762
763         for (i = 0; i < 100; i++) {
764                 /* They ack this in play_dead by setting CPU_DEAD */
765                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
766                 {
767                         printk ("CPU %d is now offline\n", cpu);
768                         return;
769                 }
770                 msleep(100);
771         }
772         printk(KERN_ERR "CPU %u didn't die...\n", cpu);
773 }
774 #endif /* CONFIG_HOTPLUG_CPU */
775
776 void
777 smp_cpus_done (unsigned int dummy)
778 {
779         int cpu;
780         unsigned long bogosum = 0;
781
782         /*
783          * Allow the user to impress friends.
784          */
785
786         for_each_online_cpu(cpu) {
787                 bogosum += cpu_data(cpu)->loops_per_jiffy;
788         }
789
790         printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
791                (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
792 }
793
794 static inline void __devinit
795 set_cpu_sibling_map(int cpu)
796 {
797         int i;
798
799         for_each_online_cpu(i) {
800                 if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
801                         cpu_set(i, cpu_core_map[cpu]);
802                         cpu_set(cpu, cpu_core_map[i]);
803                         if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
804                                 cpu_set(i, per_cpu(cpu_sibling_map, cpu));
805                                 cpu_set(cpu, per_cpu(cpu_sibling_map, i));
806                         }
807                 }
808         }
809 }
810
811 int __cpuinit
812 __cpu_up (unsigned int cpu)
813 {
814         int ret;
815         int sapicid;
816
817         sapicid = ia64_cpu_to_sapicid[cpu];
818         if (sapicid == -1)
819                 return -EINVAL;
820
821         /*
822          * Already booted cpu? not valid anymore since we dont
823          * do idle loop tightspin anymore.
824          */
825         if (cpu_isset(cpu, cpu_callin_map))
826                 return -EINVAL;
827
828         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
829         /* Processor goes to start_secondary(), sets online flag */
830         ret = do_boot_cpu(sapicid, cpu);
831         if (ret < 0)
832                 return ret;
833
834         if (cpu_data(cpu)->threads_per_core == 1 &&
835             cpu_data(cpu)->cores_per_socket == 1) {
836                 cpu_set(cpu, per_cpu(cpu_sibling_map, cpu));
837                 cpu_set(cpu, cpu_core_map[cpu]);
838                 return 0;
839         }
840
841         set_cpu_sibling_map(cpu);
842
843         return 0;
844 }
845
846 /*
847  * Assume that CPUs have been discovered by some platform-dependent interface.  For
848  * SoftSDV/Lion, that would be ACPI.
849  *
850  * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
851  */
852 void __init
853 init_smp_config(void)
854 {
855         struct fptr {
856                 unsigned long fp;
857                 unsigned long gp;
858         } *ap_startup;
859         long sal_ret;
860
861         /* Tell SAL where to drop the APs.  */
862         ap_startup = (struct fptr *) start_ap;
863         sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
864                                        ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
865         if (sal_ret < 0)
866                 printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
867                        ia64_sal_strerror(sal_ret));
868 }
869
870 /*
871  * identify_siblings(cpu) gets called from identify_cpu. This populates the 
872  * information related to logical execution units in per_cpu_data structure.
873  */
874 void __devinit
875 identify_siblings(struct cpuinfo_ia64 *c)
876 {
877         s64 status;
878         u16 pltid;
879         pal_logical_to_physical_t info;
880
881         status = ia64_pal_logical_to_phys(-1, &info);
882         if (status != PAL_STATUS_SUCCESS) {
883                 if (status != PAL_STATUS_UNIMPLEMENTED) {
884                         printk(KERN_ERR
885                                 "ia64_pal_logical_to_phys failed with %ld\n",
886                                 status);
887                         return;
888                 }
889
890                 info.overview_ppid = 0;
891                 info.overview_cpp  = 1;
892                 info.overview_tpc  = 1;
893         }
894
895         status = ia64_sal_physical_id_info(&pltid);
896         if (status != PAL_STATUS_SUCCESS) {
897                 if (status != PAL_STATUS_UNIMPLEMENTED)
898                         printk(KERN_ERR
899                                 "ia64_sal_pltid failed with %ld\n",
900                                 status);
901                 return;
902         }
903
904         c->socket_id =  (pltid << 8) | info.overview_ppid;
905
906         if (info.overview_cpp == 1 && info.overview_tpc == 1)
907                 return;
908
909         c->cores_per_socket = info.overview_cpp;
910         c->threads_per_core = info.overview_tpc;
911         c->num_log = info.overview_num_log;
912
913         c->core_id = info.log1_cid;
914         c->thread_id = info.log1_tid;
915 }
916
917 /*
918  * returns non zero, if multi-threading is enabled
919  * on at least one physical package. Due to hotplug cpu
920  * and (maxcpus=), all threads may not necessarily be enabled
921  * even though the processor supports multi-threading.
922  */
923 int is_multithreading_enabled(void)
924 {
925         int i, j;
926
927         for_each_present_cpu(i) {
928                 for_each_present_cpu(j) {
929                         if (j == i)
930                                 continue;
931                         if ((cpu_data(j)->socket_id == cpu_data(i)->socket_id)) {
932                                 if (cpu_data(j)->core_id == cpu_data(i)->core_id)
933                                         return 1;
934                         }
935                 }
936         }
937         return 0;
938 }
939 EXPORT_SYMBOL_GPL(is_multithreading_enabled);