Automatic merge of /spare/repo/netdev-2.6 branch remove-drivers
[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 #include <linux/config.h>
25
26 #include <linux/module.h>
27 #include <linux/acpi.h>
28 #include <linux/bootmem.h>
29 #include <linux/cpu.h>
30 #include <linux/delay.h>
31 #include <linux/init.h>
32 #include <linux/interrupt.h>
33 #include <linux/irq.h>
34 #include <linux/kernel.h>
35 #include <linux/kernel_stat.h>
36 #include <linux/mm.h>
37 #include <linux/notifier.h>
38 #include <linux/smp.h>
39 #include <linux/smp_lock.h>
40 #include <linux/spinlock.h>
41 #include <linux/efi.h>
42 #include <linux/percpu.h>
43 #include <linux/bitops.h>
44
45 #include <asm/atomic.h>
46 #include <asm/cache.h>
47 #include <asm/current.h>
48 #include <asm/delay.h>
49 #include <asm/ia32.h>
50 #include <asm/io.h>
51 #include <asm/irq.h>
52 #include <asm/machvec.h>
53 #include <asm/mca.h>
54 #include <asm/page.h>
55 #include <asm/pgalloc.h>
56 #include <asm/pgtable.h>
57 #include <asm/processor.h>
58 #include <asm/ptrace.h>
59 #include <asm/sal.h>
60 #include <asm/system.h>
61 #include <asm/tlbflush.h>
62 #include <asm/unistd.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 /*
74  * Store all idle threads, this can be reused instead of creating
75  * a new thread. Also avoids complicated thread destroy functionality
76  * for idle threads.
77  */
78 struct task_struct *idle_thread_array[NR_CPUS];
79
80 /*
81  * Global array allocated for NR_CPUS at boot time
82  */
83 struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
84
85 /*
86  * start_ap in head.S uses this to store current booting cpu
87  * info.
88  */
89 struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
90
91 #define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
92
93 #define get_idle_for_cpu(x)             (idle_thread_array[(x)])
94 #define set_idle_for_cpu(x,p)   (idle_thread_array[(x)] = (p))
95
96 #else
97
98 #define get_idle_for_cpu(x)             (NULL)
99 #define set_idle_for_cpu(x,p)
100 #define set_brendez_area(x)
101 #endif
102
103
104 /*
105  * ITC synchronization related stuff:
106  */
107 #define MASTER  0
108 #define SLAVE   (SMP_CACHE_BYTES/8)
109
110 #define NUM_ROUNDS      64      /* magic value */
111 #define NUM_ITERS       5       /* likewise */
112
113 static DEFINE_SPINLOCK(itc_sync_lock);
114 static volatile unsigned long go[SLAVE + 1];
115
116 #define DEBUG_ITC_SYNC  0
117
118 extern void __devinit calibrate_delay (void);
119 extern void start_ap (void);
120 extern unsigned long ia64_iobase;
121
122 task_t *task_for_booting_cpu;
123
124 /*
125  * State for each CPU
126  */
127 DEFINE_PER_CPU(int, cpu_state);
128
129 /* Bitmasks of currently online, and possible CPUs */
130 cpumask_t cpu_online_map;
131 EXPORT_SYMBOL(cpu_online_map);
132 cpumask_t cpu_possible_map;
133 EXPORT_SYMBOL(cpu_possible_map);
134
135 cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
136 cpumask_t cpu_sibling_map[NR_CPUS] __cacheline_aligned;
137 int smp_num_siblings = 1;
138 int smp_num_cpucores = 1;
139
140 /* which logical CPU number maps to which CPU (physical APIC ID) */
141 volatile int ia64_cpu_to_sapicid[NR_CPUS];
142 EXPORT_SYMBOL(ia64_cpu_to_sapicid);
143
144 static volatile cpumask_t cpu_callin_map;
145
146 struct smp_boot_data smp_boot_data __initdata;
147
148 unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
149
150 char __initdata no_int_routing;
151
152 unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
153
154 static int __init
155 nointroute (char *str)
156 {
157         no_int_routing = 1;
158         printk ("no_int_routing on\n");
159         return 1;
160 }
161
162 __setup("nointroute", nointroute);
163
164 void
165 sync_master (void *arg)
166 {
167         unsigned long flags, i;
168
169         go[MASTER] = 0;
170
171         local_irq_save(flags);
172         {
173                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
174                         while (!go[MASTER])
175                                 cpu_relax();
176                         go[MASTER] = 0;
177                         go[SLAVE] = ia64_get_itc();
178                 }
179         }
180         local_irq_restore(flags);
181 }
182
183 /*
184  * Return the number of cycles by which our itc differs from the itc on the master
185  * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
186  * negative that it is behind.
187  */
188 static inline long
189 get_delta (long *rt, long *master)
190 {
191         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
192         unsigned long tcenter, t0, t1, tm;
193         long i;
194
195         for (i = 0; i < NUM_ITERS; ++i) {
196                 t0 = ia64_get_itc();
197                 go[MASTER] = 1;
198                 while (!(tm = go[SLAVE]))
199                         cpu_relax();
200                 go[SLAVE] = 0;
201                 t1 = ia64_get_itc();
202
203                 if (t1 - t0 < best_t1 - best_t0)
204                         best_t0 = t0, best_t1 = t1, best_tm = tm;
205         }
206
207         *rt = best_t1 - best_t0;
208         *master = best_tm - best_t0;
209
210         /* average best_t0 and best_t1 without overflow: */
211         tcenter = (best_t0/2 + best_t1/2);
212         if (best_t0 % 2 + best_t1 % 2 == 2)
213                 ++tcenter;
214         return tcenter - best_tm;
215 }
216
217 /*
218  * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
219  * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
220  * unaccounted-for errors (such as getting a machine check in the middle of a calibration
221  * step).  The basic idea is for the slave to ask the master what itc value it has and to
222  * read its own itc before and after the master responds.  Each iteration gives us three
223  * timestamps:
224  *
225  *      slave           master
226  *
227  *      t0 ---\
228  *             ---\
229  *                 --->
230  *                      tm
231  *                 /---
232  *             /---
233  *      t1 <---
234  *
235  *
236  * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
237  * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
238  * between the slave and the master is symmetric.  Even if the interconnect were
239  * asymmetric, we would still know that the synchronization error is smaller than the
240  * roundtrip latency (t0 - t1).
241  *
242  * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
243  * within one or two cycles.  However, we can only *guarantee* that the synchronization is
244  * accurate to within a round-trip time, which is typically in the range of several
245  * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
246  * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
247  * than half a micro second or so.
248  */
249 void
250 ia64_sync_itc (unsigned int master)
251 {
252         long i, delta, adj, adjust_latency = 0, done = 0;
253         unsigned long flags, rt, master_time_stamp, bound;
254 #if DEBUG_ITC_SYNC
255         struct {
256                 long rt;        /* roundtrip time */
257                 long master;    /* master's timestamp */
258                 long diff;      /* difference between midpoint and master's timestamp */
259                 long lat;       /* estimate of itc adjustment latency */
260         } t[NUM_ROUNDS];
261 #endif
262
263         /*
264          * Make sure local timer ticks are disabled while we sync.  If
265          * they were enabled, we'd have to worry about nasty issues
266          * like setting the ITC ahead of (or a long time before) the
267          * next scheduled tick.
268          */
269         BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
270
271         go[MASTER] = 1;
272
273         if (smp_call_function_single(master, sync_master, NULL, 1, 0) < 0) {
274                 printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
275                 return;
276         }
277
278         while (go[MASTER])
279                 cpu_relax();    /* wait for master to be ready */
280
281         spin_lock_irqsave(&itc_sync_lock, flags);
282         {
283                 for (i = 0; i < NUM_ROUNDS; ++i) {
284                         delta = get_delta(&rt, &master_time_stamp);
285                         if (delta == 0) {
286                                 done = 1;       /* let's lock on to this... */
287                                 bound = rt;
288                         }
289
290                         if (!done) {
291                                 if (i > 0) {
292                                         adjust_latency += -delta;
293                                         adj = -delta + adjust_latency/4;
294                                 } else
295                                         adj = -delta;
296
297                                 ia64_set_itc(ia64_get_itc() + adj);
298                         }
299 #if DEBUG_ITC_SYNC
300                         t[i].rt = rt;
301                         t[i].master = master_time_stamp;
302                         t[i].diff = delta;
303                         t[i].lat = adjust_latency/4;
304 #endif
305                 }
306         }
307         spin_unlock_irqrestore(&itc_sync_lock, flags);
308
309 #if DEBUG_ITC_SYNC
310         for (i = 0; i < NUM_ROUNDS; ++i)
311                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
312                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
313 #endif
314
315         printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
316                "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
317 }
318
319 /*
320  * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
321  */
322 static inline void __devinit
323 smp_setup_percpu_timer (void)
324 {
325 }
326
327 static void __devinit
328 smp_callin (void)
329 {
330         int cpuid, phys_id;
331         extern void ia64_init_itm(void);
332
333 #ifdef CONFIG_PERFMON
334         extern void pfm_init_percpu(void);
335 #endif
336
337         cpuid = smp_processor_id();
338         phys_id = hard_smp_processor_id();
339
340         if (cpu_online(cpuid)) {
341                 printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
342                        phys_id, cpuid);
343                 BUG();
344         }
345
346         lock_ipi_calllock();
347         cpu_set(cpuid, cpu_online_map);
348         unlock_ipi_calllock();
349
350         smp_setup_percpu_timer();
351
352         ia64_mca_cmc_vector_setup();    /* Setup vector on AP */
353
354 #ifdef CONFIG_PERFMON
355         pfm_init_percpu();
356 #endif
357
358         local_irq_enable();
359
360         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
361                 /*
362                  * Synchronize the ITC with the BP.  Need to do this after irqs are
363                  * enabled because ia64_sync_itc() calls smp_call_function_single(), which
364                  * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
365                  * local_bh_enable(), which bugs out if irqs are not enabled...
366                  */
367                 Dprintk("Going to syncup ITC with BP.\n");
368                 ia64_sync_itc(0);
369         }
370
371         /*
372          * Get our bogomips.
373          */
374         ia64_init_itm();
375         calibrate_delay();
376         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
377
378 #ifdef CONFIG_IA32_SUPPORT
379         ia32_gdt_init();
380 #endif
381
382         /*
383          * Allow the master to continue.
384          */
385         cpu_set(cpuid, cpu_callin_map);
386         Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
387 }
388
389
390 /*
391  * Activate a secondary processor.  head.S calls this.
392  */
393 int __devinit
394 start_secondary (void *unused)
395 {
396         /* Early console may use I/O ports */
397         ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
398         Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
399         efi_map_pal_code();
400         cpu_init();
401         smp_callin();
402
403         cpu_idle();
404         return 0;
405 }
406
407 struct pt_regs * __devinit idle_regs(struct pt_regs *regs)
408 {
409         return NULL;
410 }
411
412 struct create_idle {
413         struct task_struct *idle;
414         struct completion done;
415         int cpu;
416 };
417
418 void
419 do_fork_idle(void *_c_idle)
420 {
421         struct create_idle *c_idle = _c_idle;
422
423         c_idle->idle = fork_idle(c_idle->cpu);
424         complete(&c_idle->done);
425 }
426
427 static int __devinit
428 do_boot_cpu (int sapicid, int cpu)
429 {
430         int timeout;
431         struct create_idle c_idle = {
432                 .cpu    = cpu,
433                 .done   = COMPLETION_INITIALIZER(c_idle.done),
434         };
435         DECLARE_WORK(work, do_fork_idle, &c_idle);
436
437         c_idle.idle = get_idle_for_cpu(cpu);
438         if (c_idle.idle) {
439                 init_idle(c_idle.idle, cpu);
440                 goto do_rest;
441         }
442
443         /*
444          * We can't use kernel_thread since we must avoid to reschedule the child.
445          */
446         if (!keventd_up() || current_is_keventd())
447                 work.func(work.data);
448         else {
449                 schedule_work(&work);
450                 wait_for_completion(&c_idle.done);
451         }
452
453         if (IS_ERR(c_idle.idle))
454                 panic("failed fork for CPU %d", cpu);
455
456         set_idle_for_cpu(cpu, c_idle.idle);
457
458 do_rest:
459         task_for_booting_cpu = c_idle.idle;
460
461         Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
462
463         set_brendez_area(cpu);
464         platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
465
466         /*
467          * Wait 10s total for the AP to start
468          */
469         Dprintk("Waiting on callin_map ...");
470         for (timeout = 0; timeout < 100000; timeout++) {
471                 if (cpu_isset(cpu, cpu_callin_map))
472                         break;  /* It has booted */
473                 udelay(100);
474         }
475         Dprintk("\n");
476
477         if (!cpu_isset(cpu, cpu_callin_map)) {
478                 printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
479                 ia64_cpu_to_sapicid[cpu] = -1;
480                 cpu_clear(cpu, cpu_online_map);  /* was set in smp_callin() */
481                 return -EINVAL;
482         }
483         return 0;
484 }
485
486 static int __init
487 decay (char *str)
488 {
489         int ticks;
490         get_option (&str, &ticks);
491         return 1;
492 }
493
494 __setup("decay=", decay);
495
496 /*
497  * Initialize the logical CPU number to SAPICID mapping
498  */
499 void __init
500 smp_build_cpu_map (void)
501 {
502         int sapicid, cpu, i;
503         int boot_cpu_id = hard_smp_processor_id();
504
505         for (cpu = 0; cpu < NR_CPUS; cpu++) {
506                 ia64_cpu_to_sapicid[cpu] = -1;
507 #ifdef CONFIG_HOTPLUG_CPU
508                 cpu_set(cpu, cpu_possible_map);
509 #endif
510         }
511
512         ia64_cpu_to_sapicid[0] = boot_cpu_id;
513         cpus_clear(cpu_present_map);
514         cpu_set(0, cpu_present_map);
515         cpu_set(0, cpu_possible_map);
516         for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
517                 sapicid = smp_boot_data.cpu_phys_id[i];
518                 if (sapicid == boot_cpu_id)
519                         continue;
520                 cpu_set(cpu, cpu_present_map);
521                 cpu_set(cpu, cpu_possible_map);
522                 ia64_cpu_to_sapicid[cpu] = sapicid;
523                 cpu++;
524         }
525 }
526
527 #ifdef CONFIG_NUMA
528
529 /* on which node is each logical CPU (one cacheline even for 64 CPUs) */
530 u8 cpu_to_node_map[NR_CPUS] __cacheline_aligned;
531 EXPORT_SYMBOL(cpu_to_node_map);
532 /* which logical CPUs are on which nodes */
533 cpumask_t node_to_cpu_mask[MAX_NUMNODES] __cacheline_aligned;
534
535 /*
536  * Build cpu to node mapping and initialize the per node cpu masks.
537  */
538 void __init
539 build_cpu_to_node_map (void)
540 {
541         int cpu, i, node;
542
543         for(node=0; node<MAX_NUMNODES; node++)
544                 cpus_clear(node_to_cpu_mask[node]);
545         for(cpu = 0; cpu < NR_CPUS; ++cpu) {
546                 /*
547                  * All Itanium NUMA platforms I know use ACPI, so maybe we
548                  * can drop this ifdef completely.                    [EF]
549                  */
550 #ifdef CONFIG_ACPI_NUMA
551                 node = -1;
552                 for (i = 0; i < NR_CPUS; ++i)
553                         if (cpu_physical_id(cpu) == node_cpuid[i].phys_id) {
554                                 node = node_cpuid[i].nid;
555                                 break;
556                         }
557 #else
558 #               error Fixme: Dunno how to build CPU-to-node map.
559 #endif
560                 cpu_to_node_map[cpu] = (node >= 0) ? node : 0;
561                 if (node >= 0)
562                         cpu_set(cpu, node_to_cpu_mask[node]);
563         }
564 }
565
566 #endif /* CONFIG_NUMA */
567
568 /*
569  * Cycle through the APs sending Wakeup IPIs to boot each.
570  */
571 void __init
572 smp_prepare_cpus (unsigned int max_cpus)
573 {
574         int boot_cpu_id = hard_smp_processor_id();
575
576         /*
577          * Initialize the per-CPU profiling counter/multiplier
578          */
579
580         smp_setup_percpu_timer();
581
582         /*
583          * We have the boot CPU online for sure.
584          */
585         cpu_set(0, cpu_online_map);
586         cpu_set(0, cpu_callin_map);
587
588         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
589         ia64_cpu_to_sapicid[0] = boot_cpu_id;
590
591         printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
592
593         current_thread_info()->cpu = 0;
594
595         /*
596          * If SMP should be disabled, then really disable it!
597          */
598         if (!max_cpus) {
599                 printk(KERN_INFO "SMP mode deactivated.\n");
600                 cpus_clear(cpu_online_map);
601                 cpus_clear(cpu_present_map);
602                 cpus_clear(cpu_possible_map);
603                 cpu_set(0, cpu_online_map);
604                 cpu_set(0, cpu_present_map);
605                 cpu_set(0, cpu_possible_map);
606                 return;
607         }
608 }
609
610 void __devinit smp_prepare_boot_cpu(void)
611 {
612         cpu_set(smp_processor_id(), cpu_online_map);
613         cpu_set(smp_processor_id(), cpu_callin_map);
614 }
615
616 /*
617  * mt_info[] is a temporary store for all info returned by
618  * PAL_LOGICAL_TO_PHYSICAL, to be copied into cpuinfo_ia64 when the
619  * specific cpu comes.
620  */
621 static struct {
622         __u32   socket_id;
623         __u16   core_id;
624         __u16   thread_id;
625         __u16   proc_fixed_addr;
626         __u8    valid;
627 } mt_info[NR_CPUS] __devinitdata;
628
629 #ifdef CONFIG_HOTPLUG_CPU
630 static inline void
631 remove_from_mtinfo(int cpu)
632 {
633         int i;
634
635         for_each_cpu(i)
636                 if (mt_info[i].valid &&  mt_info[i].socket_id ==
637                                                 cpu_data(cpu)->socket_id)
638                         mt_info[i].valid = 0;
639 }
640
641 static inline void
642 clear_cpu_sibling_map(int cpu)
643 {
644         int i;
645
646         for_each_cpu_mask(i, cpu_sibling_map[cpu])
647                 cpu_clear(cpu, cpu_sibling_map[i]);
648         for_each_cpu_mask(i, cpu_core_map[cpu])
649                 cpu_clear(cpu, cpu_core_map[i]);
650
651         cpu_sibling_map[cpu] = cpu_core_map[cpu] = CPU_MASK_NONE;
652 }
653
654 static void
655 remove_siblinginfo(int cpu)
656 {
657         int last = 0;
658
659         if (cpu_data(cpu)->threads_per_core == 1 &&
660             cpu_data(cpu)->cores_per_socket == 1) {
661                 cpu_clear(cpu, cpu_core_map[cpu]);
662                 cpu_clear(cpu, cpu_sibling_map[cpu]);
663                 return;
664         }
665
666         last = (cpus_weight(cpu_core_map[cpu]) == 1 ? 1 : 0);
667
668         /* remove it from all sibling map's */
669         clear_cpu_sibling_map(cpu);
670
671         /* if this cpu is the last in the core group, remove all its info 
672          * from mt_info structure
673          */
674         if (last)
675                 remove_from_mtinfo(cpu);
676 }
677
678 extern void fixup_irqs(void);
679 /* must be called with cpucontrol mutex held */
680 int __cpu_disable(void)
681 {
682         int cpu = smp_processor_id();
683
684         /*
685          * dont permit boot processor for now
686          */
687         if (cpu == 0)
688                 return -EBUSY;
689
690         remove_siblinginfo(cpu);
691         fixup_irqs();
692         local_flush_tlb_all();
693         cpu_clear(cpu, cpu_callin_map);
694         return 0;
695 }
696
697 void __cpu_die(unsigned int cpu)
698 {
699         unsigned int i;
700
701         for (i = 0; i < 100; i++) {
702                 /* They ack this in play_dead by setting CPU_DEAD */
703                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
704                 {
705                         printk ("CPU %d is now offline\n", cpu);
706                         return;
707                 }
708                 msleep(100);
709         }
710         printk(KERN_ERR "CPU %u didn't die...\n", cpu);
711 }
712 #else /* !CONFIG_HOTPLUG_CPU */
713 int __cpu_disable(void)
714 {
715         return -ENOSYS;
716 }
717
718 void __cpu_die(unsigned int cpu)
719 {
720         /* We said "no" in __cpu_disable */
721         BUG();
722 }
723 #endif /* CONFIG_HOTPLUG_CPU */
724
725 void
726 smp_cpus_done (unsigned int dummy)
727 {
728         int cpu;
729         unsigned long bogosum = 0;
730
731         /*
732          * Allow the user to impress friends.
733          */
734
735         for (cpu = 0; cpu < NR_CPUS; cpu++)
736                 if (cpu_online(cpu))
737                         bogosum += cpu_data(cpu)->loops_per_jiffy;
738
739         printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
740                (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
741 }
742
743 static inline void __devinit
744 set_cpu_sibling_map(int cpu)
745 {
746         int i;
747
748         for_each_online_cpu(i) {
749                 if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
750                         cpu_set(i, cpu_core_map[cpu]);
751                         cpu_set(cpu, cpu_core_map[i]);
752                         if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
753                                 cpu_set(i, cpu_sibling_map[cpu]);
754                                 cpu_set(cpu, cpu_sibling_map[i]);
755                         }
756                 }
757         }
758 }
759
760 int __devinit
761 __cpu_up (unsigned int cpu)
762 {
763         int ret;
764         int sapicid;
765
766         sapicid = ia64_cpu_to_sapicid[cpu];
767         if (sapicid == -1)
768                 return -EINVAL;
769
770         /*
771          * Already booted cpu? not valid anymore since we dont
772          * do idle loop tightspin anymore.
773          */
774         if (cpu_isset(cpu, cpu_callin_map))
775                 return -EINVAL;
776
777         /* Processor goes to start_secondary(), sets online flag */
778         ret = do_boot_cpu(sapicid, cpu);
779         if (ret < 0)
780                 return ret;
781
782         if (cpu_data(cpu)->threads_per_core == 1 &&
783             cpu_data(cpu)->cores_per_socket == 1) {
784                 cpu_set(cpu, cpu_sibling_map[cpu]);
785                 cpu_set(cpu, cpu_core_map[cpu]);
786                 return 0;
787         }
788
789         set_cpu_sibling_map(cpu);
790
791         return 0;
792 }
793
794 /*
795  * Assume that CPU's have been discovered by some platform-dependent interface.  For
796  * SoftSDV/Lion, that would be ACPI.
797  *
798  * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
799  */
800 void __init
801 init_smp_config(void)
802 {
803         struct fptr {
804                 unsigned long fp;
805                 unsigned long gp;
806         } *ap_startup;
807         long sal_ret;
808
809         /* Tell SAL where to drop the AP's.  */
810         ap_startup = (struct fptr *) start_ap;
811         sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
812                                        ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
813         if (sal_ret < 0)
814                 printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
815                        ia64_sal_strerror(sal_ret));
816 }
817
818 static inline int __devinit
819 check_for_mtinfo_index(void)
820 {
821         int i;
822         
823         for_each_cpu(i)
824                 if (!mt_info[i].valid)
825                         return i;
826
827         return -1;
828 }
829
830 /*
831  * Search the mt_info to find out if this socket's cid/tid information is
832  * cached or not. If the socket exists, fill in the core_id and thread_id 
833  * in cpuinfo
834  */
835 static int __devinit
836 check_for_new_socket(__u16 logical_address, struct cpuinfo_ia64 *c)
837 {
838         int i;
839         __u32 sid = c->socket_id;
840
841         for_each_cpu(i) {
842                 if (mt_info[i].valid && mt_info[i].proc_fixed_addr == logical_address
843                     && mt_info[i].socket_id == sid) {
844                         c->core_id = mt_info[i].core_id;
845                         c->thread_id = mt_info[i].thread_id;
846                         return 1; /* not a new socket */
847                 }
848         }
849         return 0;
850 }
851
852 /*
853  * identify_siblings(cpu) gets called from identify_cpu. This populates the 
854  * information related to logical execution units in per_cpu_data structure.
855  */
856 void __devinit
857 identify_siblings(struct cpuinfo_ia64 *c)
858 {
859         s64 status;
860         u16 pltid;
861         u64 proc_fixed_addr;
862         int count, i;
863         pal_logical_to_physical_t info;
864
865         if (smp_num_cpucores == 1 && smp_num_siblings == 1)
866                 return;
867
868         if ((status = ia64_pal_logical_to_phys(0, &info)) != PAL_STATUS_SUCCESS) {
869                 printk(KERN_ERR "ia64_pal_logical_to_phys failed with %ld\n",
870                        status);
871                 return;
872         }
873         if ((status = ia64_sal_physical_id_info(&pltid)) != PAL_STATUS_SUCCESS) {
874                 printk(KERN_ERR "ia64_sal_pltid failed with %ld\n", status);
875                 return;
876         }
877         if ((status = ia64_pal_fixed_addr(&proc_fixed_addr)) != PAL_STATUS_SUCCESS) {
878                 printk(KERN_ERR "ia64_pal_fixed_addr failed with %ld\n", status);
879                 return;
880         }
881
882         c->socket_id =  (pltid << 8) | info.overview_ppid;
883         c->cores_per_socket = info.overview_cpp;
884         c->threads_per_core = info.overview_tpc;
885         count = c->num_log = info.overview_num_log;
886
887         /* If the thread and core id information is already cached, then
888          * we will simply update cpu_info and return. Otherwise, we will
889          * do the PAL calls and cache core and thread id's of all the siblings.
890          */
891         if (check_for_new_socket(proc_fixed_addr, c))
892                 return;
893
894         for (i = 0; i < count; i++) {
895                 int index;
896
897                 if (i && (status = ia64_pal_logical_to_phys(i, &info))
898                           != PAL_STATUS_SUCCESS) {
899                         printk(KERN_ERR "ia64_pal_logical_to_phys failed"
900                                         " with %ld\n", status);
901                         return;
902                 }
903                 if (info.log2_la == proc_fixed_addr) {
904                         c->core_id = info.log1_cid;
905                         c->thread_id = info.log1_tid;
906                 }
907
908                 index = check_for_mtinfo_index();
909                 /* We will not do the mt_info caching optimization in this case.
910                  */
911                 if (index < 0)
912                         continue;
913
914                 mt_info[index].valid = 1;
915                 mt_info[index].socket_id = c->socket_id;
916                 mt_info[index].core_id = info.log1_cid;
917                 mt_info[index].thread_id = info.log1_tid;
918                 mt_info[index].proc_fixed_addr = info.log2_la;
919         }
920 }