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