Convert cpu_sibling_map to be a per cpu variable
[linux-3.10.git] / 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/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 #include <asm/sn/arch.h>
62
63 #define SMP_DEBUG 0
64
65 #if SMP_DEBUG
66 #define Dprintk(x...)  printk(x)
67 #else
68 #define Dprintk(x...)
69 #endif
70
71 #ifdef CONFIG_HOTPLUG_CPU
72 #ifdef CONFIG_PERMIT_BSP_REMOVE
73 #define bsp_remove_ok   1
74 #else
75 #define bsp_remove_ok   0
76 #endif
77
78 /*
79  * Store all idle threads, this can be reused instead of creating
80  * a new thread. Also avoids complicated thread destroy functionality
81  * for idle threads.
82  */
83 struct task_struct *idle_thread_array[NR_CPUS];
84
85 /*
86  * Global array allocated for NR_CPUS at boot time
87  */
88 struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
89
90 /*
91  * start_ap in head.S uses this to store current booting cpu
92  * info.
93  */
94 struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
95
96 #define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
97
98 #define get_idle_for_cpu(x)             (idle_thread_array[(x)])
99 #define set_idle_for_cpu(x,p)   (idle_thread_array[(x)] = (p))
100
101 #else
102
103 #define get_idle_for_cpu(x)             (NULL)
104 #define set_idle_for_cpu(x,p)
105 #define set_brendez_area(x)
106 #endif
107
108
109 /*
110  * ITC synchronization related stuff:
111  */
112 #define MASTER  (0)
113 #define SLAVE   (SMP_CACHE_BYTES/8)
114
115 #define NUM_ROUNDS      64      /* magic value */
116 #define NUM_ITERS       5       /* likewise */
117
118 static DEFINE_SPINLOCK(itc_sync_lock);
119 static volatile unsigned long go[SLAVE + 1];
120
121 #define DEBUG_ITC_SYNC  0
122
123 extern void __devinit calibrate_delay (void);
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 int smp_num_cpucores = 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, 1, 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         lock_ipi_calllock();
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         unlock_ipi_calllock();
406         per_cpu(cpu_state, cpuid) = CPU_ONLINE;
407         spin_unlock(&vector_lock);
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         Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
472         efi_map_pal_code();
473         cpu_init();
474         preempt_disable();
475         smp_callin();
476
477         cpu_idle();
478         return 0;
479 }
480
481 struct pt_regs * __devinit idle_regs(struct pt_regs *regs)
482 {
483         return NULL;
484 }
485
486 struct create_idle {
487         struct work_struct work;
488         struct task_struct *idle;
489         struct completion done;
490         int cpu;
491 };
492
493 void __cpuinit
494 do_fork_idle(struct work_struct *work)
495 {
496         struct create_idle *c_idle =
497                 container_of(work, struct create_idle, work);
498
499         c_idle->idle = fork_idle(c_idle->cpu);
500         complete(&c_idle->done);
501 }
502
503 static int __cpuinit
504 do_boot_cpu (int sapicid, int cpu)
505 {
506         int timeout;
507         struct create_idle c_idle = {
508                 .work = __WORK_INITIALIZER(c_idle.work, do_fork_idle),
509                 .cpu    = cpu,
510                 .done   = COMPLETION_INITIALIZER(c_idle.done),
511         };
512
513         c_idle.idle = get_idle_for_cpu(cpu);
514         if (c_idle.idle) {
515                 init_idle(c_idle.idle, cpu);
516                 goto do_rest;
517         }
518
519         /*
520          * We can't use kernel_thread since we must avoid to reschedule the child.
521          */
522         if (!keventd_up() || current_is_keventd())
523                 c_idle.work.func(&c_idle.work);
524         else {
525                 schedule_work(&c_idle.work);
526                 wait_for_completion(&c_idle.done);
527         }
528
529         if (IS_ERR(c_idle.idle))
530                 panic("failed fork for CPU %d", cpu);
531
532         set_idle_for_cpu(cpu, c_idle.idle);
533
534 do_rest:
535         task_for_booting_cpu = c_idle.idle;
536
537         Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
538
539         set_brendez_area(cpu);
540         platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
541
542         /*
543          * Wait 10s total for the AP to start
544          */
545         Dprintk("Waiting on callin_map ...");
546         for (timeout = 0; timeout < 100000; timeout++) {
547                 if (cpu_isset(cpu, cpu_callin_map))
548                         break;  /* It has booted */
549                 udelay(100);
550         }
551         Dprintk("\n");
552
553         if (!cpu_isset(cpu, cpu_callin_map)) {
554                 printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
555                 ia64_cpu_to_sapicid[cpu] = -1;
556                 cpu_clear(cpu, cpu_online_map);  /* was set in smp_callin() */
557                 return -EINVAL;
558         }
559         return 0;
560 }
561
562 static int __init
563 decay (char *str)
564 {
565         int ticks;
566         get_option (&str, &ticks);
567         return 1;
568 }
569
570 __setup("decay=", decay);
571
572 /*
573  * Initialize the logical CPU number to SAPICID mapping
574  */
575 void __init
576 smp_build_cpu_map (void)
577 {
578         int sapicid, cpu, i;
579         int boot_cpu_id = hard_smp_processor_id();
580
581         for (cpu = 0; cpu < NR_CPUS; cpu++) {
582                 ia64_cpu_to_sapicid[cpu] = -1;
583         }
584
585         ia64_cpu_to_sapicid[0] = boot_cpu_id;
586         cpus_clear(cpu_present_map);
587         cpu_set(0, cpu_present_map);
588         cpu_set(0, cpu_possible_map);
589         for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
590                 sapicid = smp_boot_data.cpu_phys_id[i];
591                 if (sapicid == boot_cpu_id)
592                         continue;
593                 cpu_set(cpu, cpu_present_map);
594                 cpu_set(cpu, cpu_possible_map);
595                 ia64_cpu_to_sapicid[cpu] = sapicid;
596                 cpu++;
597         }
598 }
599
600 /*
601  * Cycle through the APs sending Wakeup IPIs to boot each.
602  */
603 void __init
604 smp_prepare_cpus (unsigned int max_cpus)
605 {
606         int boot_cpu_id = hard_smp_processor_id();
607
608         /*
609          * Initialize the per-CPU profiling counter/multiplier
610          */
611
612         smp_setup_percpu_timer();
613
614         /*
615          * We have the boot CPU online for sure.
616          */
617         cpu_set(0, cpu_online_map);
618         cpu_set(0, cpu_callin_map);
619
620         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
621         ia64_cpu_to_sapicid[0] = boot_cpu_id;
622
623         printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
624
625         current_thread_info()->cpu = 0;
626
627         /*
628          * If SMP should be disabled, then really disable it!
629          */
630         if (!max_cpus) {
631                 printk(KERN_INFO "SMP mode deactivated.\n");
632                 cpus_clear(cpu_online_map);
633                 cpus_clear(cpu_present_map);
634                 cpus_clear(cpu_possible_map);
635                 cpu_set(0, cpu_online_map);
636                 cpu_set(0, cpu_present_map);
637                 cpu_set(0, cpu_possible_map);
638                 return;
639         }
640 }
641
642 void __devinit smp_prepare_boot_cpu(void)
643 {
644         cpu_set(smp_processor_id(), cpu_online_map);
645         cpu_set(smp_processor_id(), cpu_callin_map);
646         per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
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 #else /* !CONFIG_HOTPLUG_CPU */
772 int __cpu_disable(void)
773 {
774         return -ENOSYS;
775 }
776
777 void __cpu_die(unsigned int cpu)
778 {
779         /* We said "no" in __cpu_disable */
780         BUG();
781 }
782 #endif /* CONFIG_HOTPLUG_CPU */
783
784 void
785 smp_cpus_done (unsigned int dummy)
786 {
787         int cpu;
788         unsigned long bogosum = 0;
789
790         /*
791          * Allow the user to impress friends.
792          */
793
794         for_each_online_cpu(cpu) {
795                 bogosum += cpu_data(cpu)->loops_per_jiffy;
796         }
797
798         printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
799                (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
800 }
801
802 static inline void __devinit
803 set_cpu_sibling_map(int cpu)
804 {
805         int i;
806
807         for_each_online_cpu(i) {
808                 if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
809                         cpu_set(i, cpu_core_map[cpu]);
810                         cpu_set(cpu, cpu_core_map[i]);
811                         if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
812                                 cpu_set(i, per_cpu(cpu_sibling_map, cpu));
813                                 cpu_set(cpu, per_cpu(cpu_sibling_map, i));
814                         }
815                 }
816         }
817 }
818
819 int __cpuinit
820 __cpu_up (unsigned int cpu)
821 {
822         int ret;
823         int sapicid;
824
825         sapicid = ia64_cpu_to_sapicid[cpu];
826         if (sapicid == -1)
827                 return -EINVAL;
828
829         /*
830          * Already booted cpu? not valid anymore since we dont
831          * do idle loop tightspin anymore.
832          */
833         if (cpu_isset(cpu, cpu_callin_map))
834                 return -EINVAL;
835
836         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
837         /* Processor goes to start_secondary(), sets online flag */
838         ret = do_boot_cpu(sapicid, cpu);
839         if (ret < 0)
840                 return ret;
841
842         if (cpu_data(cpu)->threads_per_core == 1 &&
843             cpu_data(cpu)->cores_per_socket == 1) {
844                 cpu_set(cpu, per_cpu(cpu_sibling_map, cpu));
845                 cpu_set(cpu, cpu_core_map[cpu]);
846                 return 0;
847         }
848
849         set_cpu_sibling_map(cpu);
850
851         return 0;
852 }
853
854 /*
855  * Assume that CPUs have been discovered by some platform-dependent interface.  For
856  * SoftSDV/Lion, that would be ACPI.
857  *
858  * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
859  */
860 void __init
861 init_smp_config(void)
862 {
863         struct fptr {
864                 unsigned long fp;
865                 unsigned long gp;
866         } *ap_startup;
867         long sal_ret;
868
869         /* Tell SAL where to drop the APs.  */
870         ap_startup = (struct fptr *) start_ap;
871         sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
872                                        ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
873         if (sal_ret < 0)
874                 printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
875                        ia64_sal_strerror(sal_ret));
876 }
877
878 /*
879  * identify_siblings(cpu) gets called from identify_cpu. This populates the 
880  * information related to logical execution units in per_cpu_data structure.
881  */
882 void __devinit
883 identify_siblings(struct cpuinfo_ia64 *c)
884 {
885         s64 status;
886         u16 pltid;
887         pal_logical_to_physical_t info;
888
889         if (smp_num_cpucores == 1 && smp_num_siblings == 1)
890                 return;
891
892         if ((status = ia64_pal_logical_to_phys(-1, &info)) != PAL_STATUS_SUCCESS) {
893                 printk(KERN_ERR "ia64_pal_logical_to_phys failed with %ld\n",
894                        status);
895                 return;
896         }
897         if ((status = ia64_sal_physical_id_info(&pltid)) != PAL_STATUS_SUCCESS) {
898                 printk(KERN_ERR "ia64_sal_pltid failed with %ld\n", status);
899                 return;
900         }
901
902         c->socket_id =  (pltid << 8) | info.overview_ppid;
903         c->cores_per_socket = info.overview_cpp;
904         c->threads_per_core = info.overview_tpc;
905         c->num_log = info.overview_num_log;
906
907         c->core_id = info.log1_cid;
908         c->thread_id = info.log1_tid;
909 }
910
911 /*
912  * returns non zero, if multi-threading is enabled
913  * on at least one physical package. Due to hotplug cpu
914  * and (maxcpus=), all threads may not necessarily be enabled
915  * even though the processor supports multi-threading.
916  */
917 int is_multithreading_enabled(void)
918 {
919         int i, j;
920
921         for_each_present_cpu(i) {
922                 for_each_present_cpu(j) {
923                         if (j == i)
924                                 continue;
925                         if ((cpu_data(j)->socket_id == cpu_data(i)->socket_id)) {
926                                 if (cpu_data(j)->core_id == cpu_data(i)->core_id)
927                                         return 1;
928                         }
929                 }
930         }
931         return 0;
932 }
933 EXPORT_SYMBOL_GPL(is_multithreading_enabled);