[IA64] Allow smp_call_function_single() to current cpu
[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/pgalloc.h>
54 #include <asm/pgtable.h>
55 #include <asm/processor.h>
56 #include <asm/ptrace.h>
57 #include <asm/sal.h>
58 #include <asm/system.h>
59 #include <asm/tlbflush.h>
60 #include <asm/unistd.h>
61
62 #define SMP_DEBUG 0
63
64 #if SMP_DEBUG
65 #define Dprintk(x...)  printk(x)
66 #else
67 #define Dprintk(x...)
68 #endif
69
70 #ifdef CONFIG_HOTPLUG_CPU
71 #ifdef CONFIG_PERMIT_BSP_REMOVE
72 #define bsp_remove_ok   1
73 #else
74 #define bsp_remove_ok   0
75 #endif
76
77 /*
78  * Store all idle threads, this can be reused instead of creating
79  * a new thread. Also avoids complicated thread destroy functionality
80  * for idle threads.
81  */
82 struct task_struct *idle_thread_array[NR_CPUS];
83
84 /*
85  * Global array allocated for NR_CPUS at boot time
86  */
87 struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
88
89 /*
90  * start_ap in head.S uses this to store current booting cpu
91  * info.
92  */
93 struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
94
95 #define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
96
97 #define get_idle_for_cpu(x)             (idle_thread_array[(x)])
98 #define set_idle_for_cpu(x,p)   (idle_thread_array[(x)] = (p))
99
100 #else
101
102 #define get_idle_for_cpu(x)             (NULL)
103 #define set_idle_for_cpu(x,p)
104 #define set_brendez_area(x)
105 #endif
106
107
108 /*
109  * ITC synchronization related stuff:
110  */
111 #define MASTER  (0)
112 #define SLAVE   (SMP_CACHE_BYTES/8)
113
114 #define NUM_ROUNDS      64      /* magic value */
115 #define NUM_ITERS       5       /* likewise */
116
117 static DEFINE_SPINLOCK(itc_sync_lock);
118 static volatile unsigned long go[SLAVE + 1];
119
120 #define DEBUG_ITC_SYNC  0
121
122 extern void __devinit calibrate_delay (void);
123 extern void start_ap (void);
124 extern unsigned long ia64_iobase;
125
126 struct task_struct *task_for_booting_cpu;
127
128 /*
129  * State for each CPU
130  */
131 DEFINE_PER_CPU(int, cpu_state);
132
133 /* Bitmasks of currently online, and possible CPUs */
134 cpumask_t cpu_online_map;
135 EXPORT_SYMBOL(cpu_online_map);
136 cpumask_t cpu_possible_map = CPU_MASK_NONE;
137 EXPORT_SYMBOL(cpu_possible_map);
138
139 cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
140 cpumask_t cpu_sibling_map[NR_CPUS] __cacheline_aligned;
141 int smp_num_siblings = 1;
142 int smp_num_cpucores = 1;
143
144 /* which logical CPU number maps to which CPU (physical APIC ID) */
145 volatile int ia64_cpu_to_sapicid[NR_CPUS];
146 EXPORT_SYMBOL(ia64_cpu_to_sapicid);
147
148 static volatile cpumask_t cpu_callin_map;
149
150 struct smp_boot_data smp_boot_data __initdata;
151
152 unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
153
154 char __initdata no_int_routing;
155
156 unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
157
158 #ifdef CONFIG_FORCE_CPEI_RETARGET
159 #define CPEI_OVERRIDE_DEFAULT   (1)
160 #else
161 #define CPEI_OVERRIDE_DEFAULT   (0)
162 #endif
163
164 unsigned int force_cpei_retarget = CPEI_OVERRIDE_DEFAULT;
165
166 static int __init
167 cmdl_force_cpei(char *str)
168 {
169         int value=0;
170
171         get_option (&str, &value);
172         force_cpei_retarget = value;
173
174         return 1;
175 }
176
177 __setup("force_cpei=", cmdl_force_cpei);
178
179 static int __init
180 nointroute (char *str)
181 {
182         no_int_routing = 1;
183         printk ("no_int_routing on\n");
184         return 1;
185 }
186
187 __setup("nointroute", nointroute);
188
189 static void fix_b0_for_bsp(void)
190 {
191 #ifdef CONFIG_HOTPLUG_CPU
192         int cpuid;
193         static int fix_bsp_b0 = 1;
194
195         cpuid = smp_processor_id();
196
197         /*
198          * Cache the b0 value on the first AP that comes up
199          */
200         if (!(fix_bsp_b0 && cpuid))
201                 return;
202
203         sal_boot_rendez_state[0].br[0] = sal_boot_rendez_state[cpuid].br[0];
204         printk ("Fixed BSP b0 value from CPU %d\n", cpuid);
205
206         fix_bsp_b0 = 0;
207 #endif
208 }
209
210 void
211 sync_master (void *arg)
212 {
213         unsigned long flags, i;
214
215         go[MASTER] = 0;
216
217         local_irq_save(flags);
218         {
219                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
220                         while (!go[MASTER])
221                                 cpu_relax();
222                         go[MASTER] = 0;
223                         go[SLAVE] = ia64_get_itc();
224                 }
225         }
226         local_irq_restore(flags);
227 }
228
229 /*
230  * Return the number of cycles by which our itc differs from the itc on the master
231  * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
232  * negative that it is behind.
233  */
234 static inline long
235 get_delta (long *rt, long *master)
236 {
237         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
238         unsigned long tcenter, t0, t1, tm;
239         long i;
240
241         for (i = 0; i < NUM_ITERS; ++i) {
242                 t0 = ia64_get_itc();
243                 go[MASTER] = 1;
244                 while (!(tm = go[SLAVE]))
245                         cpu_relax();
246                 go[SLAVE] = 0;
247                 t1 = ia64_get_itc();
248
249                 if (t1 - t0 < best_t1 - best_t0)
250                         best_t0 = t0, best_t1 = t1, best_tm = tm;
251         }
252
253         *rt = best_t1 - best_t0;
254         *master = best_tm - best_t0;
255
256         /* average best_t0 and best_t1 without overflow: */
257         tcenter = (best_t0/2 + best_t1/2);
258         if (best_t0 % 2 + best_t1 % 2 == 2)
259                 ++tcenter;
260         return tcenter - best_tm;
261 }
262
263 /*
264  * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
265  * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
266  * unaccounted-for errors (such as getting a machine check in the middle of a calibration
267  * step).  The basic idea is for the slave to ask the master what itc value it has and to
268  * read its own itc before and after the master responds.  Each iteration gives us three
269  * timestamps:
270  *
271  *      slave           master
272  *
273  *      t0 ---\
274  *             ---\
275  *                 --->
276  *                      tm
277  *                 /---
278  *             /---
279  *      t1 <---
280  *
281  *
282  * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
283  * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
284  * between the slave and the master is symmetric.  Even if the interconnect were
285  * asymmetric, we would still know that the synchronization error is smaller than the
286  * roundtrip latency (t0 - t1).
287  *
288  * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
289  * within one or two cycles.  However, we can only *guarantee* that the synchronization is
290  * accurate to within a round-trip time, which is typically in the range of several
291  * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
292  * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
293  * than half a micro second or so.
294  */
295 void
296 ia64_sync_itc (unsigned int master)
297 {
298         long i, delta, adj, adjust_latency = 0, done = 0;
299         unsigned long flags, rt, master_time_stamp, bound;
300 #if DEBUG_ITC_SYNC
301         struct {
302                 long rt;        /* roundtrip time */
303                 long master;    /* master's timestamp */
304                 long diff;      /* difference between midpoint and master's timestamp */
305                 long lat;       /* estimate of itc adjustment latency */
306         } t[NUM_ROUNDS];
307 #endif
308
309         /*
310          * Make sure local timer ticks are disabled while we sync.  If
311          * they were enabled, we'd have to worry about nasty issues
312          * like setting the ITC ahead of (or a long time before) the
313          * next scheduled tick.
314          */
315         BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
316
317         go[MASTER] = 1;
318
319         if (smp_call_function_single(master, sync_master, NULL, 1, 0) < 0) {
320                 printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
321                 return;
322         }
323
324         while (go[MASTER])
325                 cpu_relax();    /* wait for master to be ready */
326
327         spin_lock_irqsave(&itc_sync_lock, flags);
328         {
329                 for (i = 0; i < NUM_ROUNDS; ++i) {
330                         delta = get_delta(&rt, &master_time_stamp);
331                         if (delta == 0) {
332                                 done = 1;       /* let's lock on to this... */
333                                 bound = rt;
334                         }
335
336                         if (!done) {
337                                 if (i > 0) {
338                                         adjust_latency += -delta;
339                                         adj = -delta + adjust_latency/4;
340                                 } else
341                                         adj = -delta;
342
343                                 ia64_set_itc(ia64_get_itc() + adj);
344                         }
345 #if DEBUG_ITC_SYNC
346                         t[i].rt = rt;
347                         t[i].master = master_time_stamp;
348                         t[i].diff = delta;
349                         t[i].lat = adjust_latency/4;
350 #endif
351                 }
352         }
353         spin_unlock_irqrestore(&itc_sync_lock, flags);
354
355 #if DEBUG_ITC_SYNC
356         for (i = 0; i < NUM_ROUNDS; ++i)
357                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
358                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
359 #endif
360
361         printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
362                "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
363 }
364
365 /*
366  * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
367  */
368 static inline void __devinit
369 smp_setup_percpu_timer (void)
370 {
371 }
372
373 static void __cpuinit
374 smp_callin (void)
375 {
376         int cpuid, phys_id, itc_master;
377         struct cpuinfo_ia64 *last_cpuinfo, *this_cpuinfo;
378         extern void ia64_init_itm(void);
379         extern volatile int time_keeper_id;
380
381 #ifdef CONFIG_PERFMON
382         extern void pfm_init_percpu(void);
383 #endif
384
385         cpuid = smp_processor_id();
386         phys_id = hard_smp_processor_id();
387         itc_master = time_keeper_id;
388
389         if (cpu_online(cpuid)) {
390                 printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
391                        phys_id, cpuid);
392                 BUG();
393         }
394
395         fix_b0_for_bsp();
396
397         lock_ipi_calllock();
398         spin_lock(&vector_lock);
399         /* Setup the per cpu irq handling data structures */
400         __setup_vector_irq(cpuid);
401         cpu_set(cpuid, cpu_online_map);
402         unlock_ipi_calllock();
403         per_cpu(cpu_state, cpuid) = CPU_ONLINE;
404         spin_unlock(&vector_lock);
405
406         smp_setup_percpu_timer();
407
408         ia64_mca_cmc_vector_setup();    /* Setup vector on AP */
409
410 #ifdef CONFIG_PERFMON
411         pfm_init_percpu();
412 #endif
413
414         local_irq_enable();
415
416         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
417                 /*
418                  * Synchronize the ITC with the BP.  Need to do this after irqs are
419                  * enabled because ia64_sync_itc() calls smp_call_function_single(), which
420                  * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
421                  * local_bh_enable(), which bugs out if irqs are not enabled...
422                  */
423                 Dprintk("Going to syncup ITC with ITC Master.\n");
424                 ia64_sync_itc(itc_master);
425         }
426
427         /*
428          * Get our bogomips.
429          */
430         ia64_init_itm();
431
432         /*
433          * Delay calibration can be skipped if new processor is identical to the
434          * previous processor.
435          */
436         last_cpuinfo = cpu_data(cpuid - 1);
437         this_cpuinfo = local_cpu_data;
438         if (last_cpuinfo->itc_freq != this_cpuinfo->itc_freq ||
439             last_cpuinfo->proc_freq != this_cpuinfo->proc_freq ||
440             last_cpuinfo->features != this_cpuinfo->features ||
441             last_cpuinfo->revision != this_cpuinfo->revision ||
442             last_cpuinfo->family != this_cpuinfo->family ||
443             last_cpuinfo->archrev != this_cpuinfo->archrev ||
444             last_cpuinfo->model != this_cpuinfo->model)
445                 calibrate_delay();
446         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
447
448 #ifdef CONFIG_IA32_SUPPORT
449         ia32_gdt_init();
450 #endif
451
452         /*
453          * Allow the master to continue.
454          */
455         cpu_set(cpuid, cpu_callin_map);
456         Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
457 }
458
459
460 /*
461  * Activate a secondary processor.  head.S calls this.
462  */
463 int __cpuinit
464 start_secondary (void *unused)
465 {
466         /* Early console may use I/O ports */
467         ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
468         Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
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 * __devinit 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 }
645
646 #ifdef CONFIG_HOTPLUG_CPU
647 static inline void
648 clear_cpu_sibling_map(int cpu)
649 {
650         int i;
651
652         for_each_cpu_mask(i, cpu_sibling_map[cpu])
653                 cpu_clear(cpu, cpu_sibling_map[i]);
654         for_each_cpu_mask(i, cpu_core_map[cpu])
655                 cpu_clear(cpu, cpu_core_map[i]);
656
657         cpu_sibling_map[cpu] = cpu_core_map[cpu] = CPU_MASK_NONE;
658 }
659
660 static void
661 remove_siblinginfo(int cpu)
662 {
663         int last = 0;
664
665         if (cpu_data(cpu)->threads_per_core == 1 &&
666             cpu_data(cpu)->cores_per_socket == 1) {
667                 cpu_clear(cpu, cpu_core_map[cpu]);
668                 cpu_clear(cpu, cpu_sibling_map[cpu]);
669                 return;
670         }
671
672         last = (cpus_weight(cpu_core_map[cpu]) == 1 ? 1 : 0);
673
674         /* remove it from all sibling map's */
675         clear_cpu_sibling_map(cpu);
676 }
677
678 extern void fixup_irqs(void);
679
680 int migrate_platform_irqs(unsigned int cpu)
681 {
682         int new_cpei_cpu;
683         irq_desc_t *desc = NULL;
684         cpumask_t       mask;
685         int             retval = 0;
686
687         /*
688          * dont permit CPEI target to removed.
689          */
690         if (cpe_vector > 0 && is_cpu_cpei_target(cpu)) {
691                 printk ("CPU (%d) is CPEI Target\n", cpu);
692                 if (can_cpei_retarget()) {
693                         /*
694                          * Now re-target the CPEI to a different processor
695                          */
696                         new_cpei_cpu = any_online_cpu(cpu_online_map);
697                         mask = cpumask_of_cpu(new_cpei_cpu);
698                         set_cpei_target_cpu(new_cpei_cpu);
699                         desc = irq_desc + ia64_cpe_irq;
700                         /*
701                          * Switch for now, immediately, we need to do fake intr
702                          * as other interrupts, but need to study CPEI behaviour with
703                          * polling before making changes.
704                          */
705                         if (desc) {
706                                 desc->chip->disable(ia64_cpe_irq);
707                                 desc->chip->set_affinity(ia64_cpe_irq, mask);
708                                 desc->chip->enable(ia64_cpe_irq);
709                                 printk ("Re-targetting CPEI to cpu %d\n", new_cpei_cpu);
710                         }
711                 }
712                 if (!desc) {
713                         printk ("Unable to retarget CPEI, offline cpu [%d] failed\n", cpu);
714                         retval = -EBUSY;
715                 }
716         }
717         return retval;
718 }
719
720 /* must be called with cpucontrol mutex held */
721 int __cpu_disable(void)
722 {
723         int cpu = smp_processor_id();
724
725         /*
726          * dont permit boot processor for now
727          */
728         if (cpu == 0 && !bsp_remove_ok) {
729                 printk ("Your platform does not support removal of BSP\n");
730                 return (-EBUSY);
731         }
732
733         cpu_clear(cpu, cpu_online_map);
734
735         if (migrate_platform_irqs(cpu)) {
736                 cpu_set(cpu, cpu_online_map);
737                 return (-EBUSY);
738         }
739
740         remove_siblinginfo(cpu);
741         cpu_clear(cpu, cpu_online_map);
742         fixup_irqs();
743         local_flush_tlb_all();
744         cpu_clear(cpu, cpu_callin_map);
745         return 0;
746 }
747
748 void __cpu_die(unsigned int cpu)
749 {
750         unsigned int i;
751
752         for (i = 0; i < 100; i++) {
753                 /* They ack this in play_dead by setting CPU_DEAD */
754                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
755                 {
756                         printk ("CPU %d is now offline\n", cpu);
757                         return;
758                 }
759                 msleep(100);
760         }
761         printk(KERN_ERR "CPU %u didn't die...\n", cpu);
762 }
763 #else /* !CONFIG_HOTPLUG_CPU */
764 int __cpu_disable(void)
765 {
766         return -ENOSYS;
767 }
768
769 void __cpu_die(unsigned int cpu)
770 {
771         /* We said "no" in __cpu_disable */
772         BUG();
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, cpu_sibling_map[cpu]);
805                                 cpu_set(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, 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         if (smp_num_cpucores == 1 && smp_num_siblings == 1)
882                 return;
883
884         if ((status = ia64_pal_logical_to_phys(-1, &info)) != PAL_STATUS_SUCCESS) {
885                 printk(KERN_ERR "ia64_pal_logical_to_phys failed with %ld\n",
886                        status);
887                 return;
888         }
889         if ((status = ia64_sal_physical_id_info(&pltid)) != PAL_STATUS_SUCCESS) {
890                 printk(KERN_ERR "ia64_sal_pltid failed with %ld\n", status);
891                 return;
892         }
893
894         c->socket_id =  (pltid << 8) | info.overview_ppid;
895         c->cores_per_socket = info.overview_cpp;
896         c->threads_per_core = info.overview_tpc;
897         c->num_log = info.overview_num_log;
898
899         c->core_id = info.log1_cid;
900         c->thread_id = info.log1_tid;
901 }
902
903 /*
904  * returns non zero, if multi-threading is enabled
905  * on at least one physical package. Due to hotplug cpu
906  * and (maxcpus=), all threads may not necessarily be enabled
907  * even though the processor supports multi-threading.
908  */
909 int is_multithreading_enabled(void)
910 {
911         int i, j;
912
913         for_each_present_cpu(i) {
914                 for_each_present_cpu(j) {
915                         if (j == i)
916                                 continue;
917                         if ((cpu_data(j)->socket_id == cpu_data(i)->socket_id)) {
918                                 if (cpu_data(j)->core_id == cpu_data(i)->core_id)
919                                         return 1;
920                         }
921                 }
922         }
923         return 0;
924 }
925 EXPORT_SYMBOL_GPL(is_multithreading_enabled);