atomic: use <linux/atomic.h>
[linux-2.6.git] / arch / ia64 / kernel / smpboot.c
1 /*
2  * SMP boot-related support
3  *
4  * Copyright (C) 1998-2003, 2005 Hewlett-Packard Co
5  *      David Mosberger-Tang <davidm@hpl.hp.com>
6  * Copyright (C) 2001, 2004-2005 Intel Corp
7  *      Rohit Seth <rohit.seth@intel.com>
8  *      Suresh Siddha <suresh.b.siddha@intel.com>
9  *      Gordon Jin <gordon.jin@intel.com>
10  *      Ashok Raj  <ashok.raj@intel.com>
11  *
12  * 01/05/16 Rohit Seth <rohit.seth@intel.com>   Moved SMP booting functions from smp.c to here.
13  * 01/04/27 David Mosberger <davidm@hpl.hp.com> Added ITC synching code.
14  * 02/07/31 David Mosberger <davidm@hpl.hp.com> Switch over to hotplug-CPU boot-sequence.
15  *                                              smp_boot_cpus()/smp_commence() is replaced by
16  *                                              smp_prepare_cpus()/__cpu_up()/smp_cpus_done().
17  * 04/06/21 Ashok Raj           <ashok.raj@intel.com> Added CPU Hotplug Support
18  * 04/12/26 Jin Gordon <gordon.jin@intel.com>
19  * 04/12/26 Rohit Seth <rohit.seth@intel.com>
20  *                                              Add multi-threading and multi-core detection
21  * 05/01/30 Suresh Siddha <suresh.b.siddha@intel.com>
22  *                                              Setup cpu_sibling_map and cpu_core_map
23  */
24
25 #include <linux/module.h>
26 #include <linux/acpi.h>
27 #include <linux/bootmem.h>
28 #include <linux/cpu.h>
29 #include <linux/delay.h>
30 #include <linux/init.h>
31 #include <linux/interrupt.h>
32 #include <linux/irq.h>
33 #include <linux/kernel.h>
34 #include <linux/kernel_stat.h>
35 #include <linux/mm.h>
36 #include <linux/notifier.h>
37 #include <linux/smp.h>
38 #include <linux/spinlock.h>
39 #include <linux/efi.h>
40 #include <linux/percpu.h>
41 #include <linux/bitops.h>
42
43 #include <linux/atomic.h>
44 #include <asm/cache.h>
45 #include <asm/current.h>
46 #include <asm/delay.h>
47 #include <asm/io.h>
48 #include <asm/irq.h>
49 #include <asm/machvec.h>
50 #include <asm/mca.h>
51 #include <asm/page.h>
52 #include <asm/paravirt.h>
53 #include <asm/pgalloc.h>
54 #include <asm/pgtable.h>
55 #include <asm/processor.h>
56 #include <asm/ptrace.h>
57 #include <asm/sal.h>
58 #include <asm/system.h>
59 #include <asm/tlbflush.h>
60 #include <asm/unistd.h>
61 #include <asm/sn/arch.h>
62
63 #define SMP_DEBUG 0
64
65 #if SMP_DEBUG
66 #define Dprintk(x...)  printk(x)
67 #else
68 #define Dprintk(x...)
69 #endif
70
71 #ifdef CONFIG_HOTPLUG_CPU
72 #ifdef CONFIG_PERMIT_BSP_REMOVE
73 #define bsp_remove_ok   1
74 #else
75 #define bsp_remove_ok   0
76 #endif
77
78 /*
79  * Store all idle threads, this can be reused instead of creating
80  * a new thread. Also avoids complicated thread destroy functionality
81  * for idle threads.
82  */
83 struct task_struct *idle_thread_array[NR_CPUS];
84
85 /*
86  * Global array allocated for NR_CPUS at boot time
87  */
88 struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
89
90 /*
91  * start_ap in head.S uses this to store current booting cpu
92  * info.
93  */
94 struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
95
96 #define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
97
98 #define get_idle_for_cpu(x)             (idle_thread_array[(x)])
99 #define set_idle_for_cpu(x,p)   (idle_thread_array[(x)] = (p))
100
101 #else
102
103 #define get_idle_for_cpu(x)             (NULL)
104 #define set_idle_for_cpu(x,p)
105 #define set_brendez_area(x)
106 #endif
107
108
109 /*
110  * ITC synchronization related stuff:
111  */
112 #define MASTER  (0)
113 #define SLAVE   (SMP_CACHE_BYTES/8)
114
115 #define NUM_ROUNDS      64      /* magic value */
116 #define NUM_ITERS       5       /* likewise */
117
118 static DEFINE_SPINLOCK(itc_sync_lock);
119 static volatile unsigned long go[SLAVE + 1];
120
121 #define DEBUG_ITC_SYNC  0
122
123 extern void 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 cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
134 EXPORT_SYMBOL(cpu_core_map);
135 DEFINE_PER_CPU_SHARED_ALIGNED(cpumask_t, cpu_sibling_map);
136 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
137
138 int smp_num_siblings = 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 #ifdef CONFIG_FORCE_CPEI_RETARGET
155 #define CPEI_OVERRIDE_DEFAULT   (1)
156 #else
157 #define CPEI_OVERRIDE_DEFAULT   (0)
158 #endif
159
160 unsigned int force_cpei_retarget = CPEI_OVERRIDE_DEFAULT;
161
162 static int __init
163 cmdl_force_cpei(char *str)
164 {
165         int value=0;
166
167         get_option (&str, &value);
168         force_cpei_retarget = value;
169
170         return 1;
171 }
172
173 __setup("force_cpei=", cmdl_force_cpei);
174
175 static int __init
176 nointroute (char *str)
177 {
178         no_int_routing = 1;
179         printk ("no_int_routing on\n");
180         return 1;
181 }
182
183 __setup("nointroute", nointroute);
184
185 static void fix_b0_for_bsp(void)
186 {
187 #ifdef CONFIG_HOTPLUG_CPU
188         int cpuid;
189         static int fix_bsp_b0 = 1;
190
191         cpuid = smp_processor_id();
192
193         /*
194          * Cache the b0 value on the first AP that comes up
195          */
196         if (!(fix_bsp_b0 && cpuid))
197                 return;
198
199         sal_boot_rendez_state[0].br[0] = sal_boot_rendez_state[cpuid].br[0];
200         printk ("Fixed BSP b0 value from CPU %d\n", cpuid);
201
202         fix_bsp_b0 = 0;
203 #endif
204 }
205
206 void
207 sync_master (void *arg)
208 {
209         unsigned long flags, i;
210
211         go[MASTER] = 0;
212
213         local_irq_save(flags);
214         {
215                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
216                         while (!go[MASTER])
217                                 cpu_relax();
218                         go[MASTER] = 0;
219                         go[SLAVE] = ia64_get_itc();
220                 }
221         }
222         local_irq_restore(flags);
223 }
224
225 /*
226  * Return the number of cycles by which our itc differs from the itc on the master
227  * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
228  * negative that it is behind.
229  */
230 static inline long
231 get_delta (long *rt, long *master)
232 {
233         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
234         unsigned long tcenter, t0, t1, tm;
235         long i;
236
237         for (i = 0; i < NUM_ITERS; ++i) {
238                 t0 = ia64_get_itc();
239                 go[MASTER] = 1;
240                 while (!(tm = go[SLAVE]))
241                         cpu_relax();
242                 go[SLAVE] = 0;
243                 t1 = ia64_get_itc();
244
245                 if (t1 - t0 < best_t1 - best_t0)
246                         best_t0 = t0, best_t1 = t1, best_tm = tm;
247         }
248
249         *rt = best_t1 - best_t0;
250         *master = best_tm - best_t0;
251
252         /* average best_t0 and best_t1 without overflow: */
253         tcenter = (best_t0/2 + best_t1/2);
254         if (best_t0 % 2 + best_t1 % 2 == 2)
255                 ++tcenter;
256         return tcenter - best_tm;
257 }
258
259 /*
260  * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
261  * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
262  * unaccounted-for errors (such as getting a machine check in the middle of a calibration
263  * step).  The basic idea is for the slave to ask the master what itc value it has and to
264  * read its own itc before and after the master responds.  Each iteration gives us three
265  * timestamps:
266  *
267  *      slave           master
268  *
269  *      t0 ---\
270  *             ---\
271  *                 --->
272  *                      tm
273  *                 /---
274  *             /---
275  *      t1 <---
276  *
277  *
278  * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
279  * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
280  * between the slave and the master is symmetric.  Even if the interconnect were
281  * asymmetric, we would still know that the synchronization error is smaller than the
282  * roundtrip latency (t0 - t1).
283  *
284  * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
285  * within one or two cycles.  However, we can only *guarantee* that the synchronization is
286  * accurate to within a round-trip time, which is typically in the range of several
287  * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
288  * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
289  * than half a micro second or so.
290  */
291 void
292 ia64_sync_itc (unsigned int master)
293 {
294         long i, delta, adj, adjust_latency = 0, done = 0;
295         unsigned long flags, rt, master_time_stamp, bound;
296 #if DEBUG_ITC_SYNC
297         struct {
298                 long rt;        /* roundtrip time */
299                 long master;    /* master's timestamp */
300                 long diff;      /* difference between midpoint and master's timestamp */
301                 long lat;       /* estimate of itc adjustment latency */
302         } t[NUM_ROUNDS];
303 #endif
304
305         /*
306          * Make sure local timer ticks are disabled while we sync.  If
307          * they were enabled, we'd have to worry about nasty issues
308          * like setting the ITC ahead of (or a long time before) the
309          * next scheduled tick.
310          */
311         BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
312
313         go[MASTER] = 1;
314
315         if (smp_call_function_single(master, sync_master, NULL, 0) < 0) {
316                 printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
317                 return;
318         }
319
320         while (go[MASTER])
321                 cpu_relax();    /* wait for master to be ready */
322
323         spin_lock_irqsave(&itc_sync_lock, flags);
324         {
325                 for (i = 0; i < NUM_ROUNDS; ++i) {
326                         delta = get_delta(&rt, &master_time_stamp);
327                         if (delta == 0) {
328                                 done = 1;       /* let's lock on to this... */
329                                 bound = rt;
330                         }
331
332                         if (!done) {
333                                 if (i > 0) {
334                                         adjust_latency += -delta;
335                                         adj = -delta + adjust_latency/4;
336                                 } else
337                                         adj = -delta;
338
339                                 ia64_set_itc(ia64_get_itc() + adj);
340                         }
341 #if DEBUG_ITC_SYNC
342                         t[i].rt = rt;
343                         t[i].master = master_time_stamp;
344                         t[i].diff = delta;
345                         t[i].lat = adjust_latency/4;
346 #endif
347                 }
348         }
349         spin_unlock_irqrestore(&itc_sync_lock, flags);
350
351 #if DEBUG_ITC_SYNC
352         for (i = 0; i < NUM_ROUNDS; ++i)
353                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
354                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
355 #endif
356
357         printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
358                "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
359 }
360
361 /*
362  * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
363  */
364 static inline void __devinit
365 smp_setup_percpu_timer (void)
366 {
367 }
368
369 static void __cpuinit
370 smp_callin (void)
371 {
372         int cpuid, phys_id, itc_master;
373         struct cpuinfo_ia64 *last_cpuinfo, *this_cpuinfo;
374         extern void ia64_init_itm(void);
375         extern volatile int time_keeper_id;
376
377 #ifdef CONFIG_PERFMON
378         extern void pfm_init_percpu(void);
379 #endif
380
381         cpuid = smp_processor_id();
382         phys_id = hard_smp_processor_id();
383         itc_master = time_keeper_id;
384
385         if (cpu_online(cpuid)) {
386                 printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
387                        phys_id, cpuid);
388                 BUG();
389         }
390
391         fix_b0_for_bsp();
392
393         /*
394          * numa_node_id() works after this.
395          */
396         set_numa_node(cpu_to_node_map[cpuid]);
397         set_numa_mem(local_memory_node(cpu_to_node_map[cpuid]));
398
399         ipi_call_lock_irq();
400         spin_lock(&vector_lock);
401         /* Setup the per cpu irq handling data structures */
402         __setup_vector_irq(cpuid);
403         notify_cpu_starting(cpuid);
404         cpu_set(cpuid, cpu_online_map);
405         per_cpu(cpu_state, cpuid) = CPU_ONLINE;
406         spin_unlock(&vector_lock);
407         ipi_call_unlock_irq();
408
409         smp_setup_percpu_timer();
410
411         ia64_mca_cmc_vector_setup();    /* Setup vector on AP */
412
413 #ifdef CONFIG_PERFMON
414         pfm_init_percpu();
415 #endif
416
417         local_irq_enable();
418
419         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
420                 /*
421                  * Synchronize the ITC with the BP.  Need to do this after irqs are
422                  * enabled because ia64_sync_itc() calls smp_call_function_single(), which
423                  * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
424                  * local_bh_enable(), which bugs out if irqs are not enabled...
425                  */
426                 Dprintk("Going to syncup ITC with ITC Master.\n");
427                 ia64_sync_itc(itc_master);
428         }
429
430         /*
431          * Get our bogomips.
432          */
433         ia64_init_itm();
434
435         /*
436          * Delay calibration can be skipped if new processor is identical to the
437          * previous processor.
438          */
439         last_cpuinfo = cpu_data(cpuid - 1);
440         this_cpuinfo = local_cpu_data;
441         if (last_cpuinfo->itc_freq != this_cpuinfo->itc_freq ||
442             last_cpuinfo->proc_freq != this_cpuinfo->proc_freq ||
443             last_cpuinfo->features != this_cpuinfo->features ||
444             last_cpuinfo->revision != this_cpuinfo->revision ||
445             last_cpuinfo->family != this_cpuinfo->family ||
446             last_cpuinfo->archrev != this_cpuinfo->archrev ||
447             last_cpuinfo->model != this_cpuinfo->model)
448                 calibrate_delay();
449         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
450
451         /*
452          * Allow the master to continue.
453          */
454         cpu_set(cpuid, cpu_callin_map);
455         Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
456 }
457
458
459 /*
460  * Activate a secondary processor.  head.S calls this.
461  */
462 int __cpuinit
463 start_secondary (void *unused)
464 {
465         /* Early console may use I/O ports */
466         ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
467 #ifndef CONFIG_PRINTK_TIME
468         Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
469 #endif
470         efi_map_pal_code();
471         cpu_init();
472         preempt_disable();
473         smp_callin();
474
475         cpu_idle();
476         return 0;
477 }
478
479 struct pt_regs * __cpuinit idle_regs(struct pt_regs *regs)
480 {
481         return NULL;
482 }
483
484 struct create_idle {
485         struct work_struct work;
486         struct task_struct *idle;
487         struct completion done;
488         int cpu;
489 };
490
491 void __cpuinit
492 do_fork_idle(struct work_struct *work)
493 {
494         struct create_idle *c_idle =
495                 container_of(work, struct create_idle, work);
496
497         c_idle->idle = fork_idle(c_idle->cpu);
498         complete(&c_idle->done);
499 }
500
501 static int __cpuinit
502 do_boot_cpu (int sapicid, int cpu)
503 {
504         int timeout;
505         struct create_idle c_idle = {
506                 .work = __WORK_INITIALIZER(c_idle.work, do_fork_idle),
507                 .cpu    = cpu,
508                 .done   = COMPLETION_INITIALIZER(c_idle.done),
509         };
510
511         /*
512          * We can't use kernel_thread since we must avoid to
513          * reschedule the child.
514          */
515         c_idle.idle = get_idle_for_cpu(cpu);
516         if (c_idle.idle) {
517                 init_idle(c_idle.idle, cpu);
518                 goto do_rest;
519         }
520
521         schedule_work(&c_idle.work);
522         wait_for_completion(&c_idle.done);
523
524         if (IS_ERR(c_idle.idle))
525                 panic("failed fork for CPU %d", cpu);
526
527         set_idle_for_cpu(cpu, c_idle.idle);
528
529 do_rest:
530         task_for_booting_cpu = c_idle.idle;
531
532         Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
533
534         set_brendez_area(cpu);
535         platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
536
537         /*
538          * Wait 10s total for the AP to start
539          */
540         Dprintk("Waiting on callin_map ...");
541         for (timeout = 0; timeout < 100000; timeout++) {
542                 if (cpu_isset(cpu, cpu_callin_map))
543                         break;  /* It has booted */
544                 udelay(100);
545         }
546         Dprintk("\n");
547
548         if (!cpu_isset(cpu, cpu_callin_map)) {
549                 printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
550                 ia64_cpu_to_sapicid[cpu] = -1;
551                 cpu_clear(cpu, cpu_online_map);  /* was set in smp_callin() */
552                 return -EINVAL;
553         }
554         return 0;
555 }
556
557 static int __init
558 decay (char *str)
559 {
560         int ticks;
561         get_option (&str, &ticks);
562         return 1;
563 }
564
565 __setup("decay=", decay);
566
567 /*
568  * Initialize the logical CPU number to SAPICID mapping
569  */
570 void __init
571 smp_build_cpu_map (void)
572 {
573         int sapicid, cpu, i;
574         int boot_cpu_id = hard_smp_processor_id();
575
576         for (cpu = 0; cpu < NR_CPUS; cpu++) {
577                 ia64_cpu_to_sapicid[cpu] = -1;
578         }
579
580         ia64_cpu_to_sapicid[0] = boot_cpu_id;
581         cpus_clear(cpu_present_map);
582         set_cpu_present(0, true);
583         set_cpu_possible(0, true);
584         for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
585                 sapicid = smp_boot_data.cpu_phys_id[i];
586                 if (sapicid == boot_cpu_id)
587                         continue;
588                 set_cpu_present(cpu, true);
589                 set_cpu_possible(cpu, true);
590                 ia64_cpu_to_sapicid[cpu] = sapicid;
591                 cpu++;
592         }
593 }
594
595 /*
596  * Cycle through the APs sending Wakeup IPIs to boot each.
597  */
598 void __init
599 smp_prepare_cpus (unsigned int max_cpus)
600 {
601         int boot_cpu_id = hard_smp_processor_id();
602
603         /*
604          * Initialize the per-CPU profiling counter/multiplier
605          */
606
607         smp_setup_percpu_timer();
608
609         /*
610          * We have the boot CPU online for sure.
611          */
612         cpu_set(0, cpu_online_map);
613         cpu_set(0, cpu_callin_map);
614
615         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
616         ia64_cpu_to_sapicid[0] = boot_cpu_id;
617
618         printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
619
620         current_thread_info()->cpu = 0;
621
622         /*
623          * If SMP should be disabled, then really disable it!
624          */
625         if (!max_cpus) {
626                 printk(KERN_INFO "SMP mode deactivated.\n");
627                 init_cpu_online(cpumask_of(0));
628                 init_cpu_present(cpumask_of(0));
629                 init_cpu_possible(cpumask_of(0));
630                 return;
631         }
632 }
633
634 void __devinit smp_prepare_boot_cpu(void)
635 {
636         cpu_set(smp_processor_id(), cpu_online_map);
637         cpu_set(smp_processor_id(), cpu_callin_map);
638         set_numa_node(cpu_to_node_map[smp_processor_id()]);
639         per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
640         paravirt_post_smp_prepare_boot_cpu();
641 }
642
643 #ifdef CONFIG_HOTPLUG_CPU
644 static inline void
645 clear_cpu_sibling_map(int cpu)
646 {
647         int i;
648
649         for_each_cpu_mask(i, per_cpu(cpu_sibling_map, cpu))
650                 cpu_clear(cpu, per_cpu(cpu_sibling_map, i));
651         for_each_cpu_mask(i, cpu_core_map[cpu])
652                 cpu_clear(cpu, cpu_core_map[i]);
653
654         per_cpu(cpu_sibling_map, cpu) = cpu_core_map[cpu] = CPU_MASK_NONE;
655 }
656
657 static void
658 remove_siblinginfo(int cpu)
659 {
660         int last = 0;
661
662         if (cpu_data(cpu)->threads_per_core == 1 &&
663             cpu_data(cpu)->cores_per_socket == 1) {
664                 cpu_clear(cpu, cpu_core_map[cpu]);
665                 cpu_clear(cpu, per_cpu(cpu_sibling_map, cpu));
666                 return;
667         }
668
669         last = (cpus_weight(cpu_core_map[cpu]) == 1 ? 1 : 0);
670
671         /* remove it from all sibling map's */
672         clear_cpu_sibling_map(cpu);
673 }
674
675 extern void fixup_irqs(void);
676
677 int migrate_platform_irqs(unsigned int cpu)
678 {
679         int new_cpei_cpu;
680         struct irq_data *data = NULL;
681         const struct cpumask *mask;
682         int             retval = 0;
683
684         /*
685          * dont permit CPEI target to removed.
686          */
687         if (cpe_vector > 0 && is_cpu_cpei_target(cpu)) {
688                 printk ("CPU (%d) is CPEI Target\n", cpu);
689                 if (can_cpei_retarget()) {
690                         /*
691                          * Now re-target the CPEI to a different processor
692                          */
693                         new_cpei_cpu = any_online_cpu(cpu_online_map);
694                         mask = cpumask_of(new_cpei_cpu);
695                         set_cpei_target_cpu(new_cpei_cpu);
696                         data = irq_get_irq_data(ia64_cpe_irq);
697                         /*
698                          * Switch for now, immediately, we need to do fake intr
699                          * as other interrupts, but need to study CPEI behaviour with
700                          * polling before making changes.
701                          */
702                         if (data && data->chip) {
703                                 data->chip->irq_disable(data);
704                                 data->chip->irq_set_affinity(data, mask, false);
705                                 data->chip->irq_enable(data);
706                                 printk ("Re-targeting CPEI to cpu %d\n", new_cpei_cpu);
707                         }
708                 }
709                 if (!data) {
710                         printk ("Unable to retarget CPEI, offline cpu [%d] failed\n", cpu);
711                         retval = -EBUSY;
712                 }
713         }
714         return retval;
715 }
716
717 /* must be called with cpucontrol mutex held */
718 int __cpu_disable(void)
719 {
720         int cpu = smp_processor_id();
721
722         /*
723          * dont permit boot processor for now
724          */
725         if (cpu == 0 && !bsp_remove_ok) {
726                 printk ("Your platform does not support removal of BSP\n");
727                 return (-EBUSY);
728         }
729
730         if (ia64_platform_is("sn2")) {
731                 if (!sn_cpu_disable_allowed(cpu))
732                         return -EBUSY;
733         }
734
735         cpu_clear(cpu, cpu_online_map);
736
737         if (migrate_platform_irqs(cpu)) {
738                 cpu_set(cpu, cpu_online_map);
739                 return -EBUSY;
740         }
741
742         remove_siblinginfo(cpu);
743         fixup_irqs();
744         local_flush_tlb_all();
745         cpu_clear(cpu, cpu_callin_map);
746         return 0;
747 }
748
749 void __cpu_die(unsigned int cpu)
750 {
751         unsigned int i;
752
753         for (i = 0; i < 100; i++) {
754                 /* They ack this in play_dead by setting CPU_DEAD */
755                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
756                 {
757                         printk ("CPU %d is now offline\n", cpu);
758                         return;
759                 }
760                 msleep(100);
761         }
762         printk(KERN_ERR "CPU %u didn't die...\n", cpu);
763 }
764 #endif /* CONFIG_HOTPLUG_CPU */
765
766 void
767 smp_cpus_done (unsigned int dummy)
768 {
769         int cpu;
770         unsigned long bogosum = 0;
771
772         /*
773          * Allow the user to impress friends.
774          */
775
776         for_each_online_cpu(cpu) {
777                 bogosum += cpu_data(cpu)->loops_per_jiffy;
778         }
779
780         printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
781                (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
782 }
783
784 static inline void __devinit
785 set_cpu_sibling_map(int cpu)
786 {
787         int i;
788
789         for_each_online_cpu(i) {
790                 if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
791                         cpu_set(i, cpu_core_map[cpu]);
792                         cpu_set(cpu, cpu_core_map[i]);
793                         if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
794                                 cpu_set(i, per_cpu(cpu_sibling_map, cpu));
795                                 cpu_set(cpu, per_cpu(cpu_sibling_map, i));
796                         }
797                 }
798         }
799 }
800
801 int __cpuinit
802 __cpu_up (unsigned int cpu)
803 {
804         int ret;
805         int sapicid;
806
807         sapicid = ia64_cpu_to_sapicid[cpu];
808         if (sapicid == -1)
809                 return -EINVAL;
810
811         /*
812          * Already booted cpu? not valid anymore since we dont
813          * do idle loop tightspin anymore.
814          */
815         if (cpu_isset(cpu, cpu_callin_map))
816                 return -EINVAL;
817
818         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
819         /* Processor goes to start_secondary(), sets online flag */
820         ret = do_boot_cpu(sapicid, cpu);
821         if (ret < 0)
822                 return ret;
823
824         if (cpu_data(cpu)->threads_per_core == 1 &&
825             cpu_data(cpu)->cores_per_socket == 1) {
826                 cpu_set(cpu, per_cpu(cpu_sibling_map, cpu));
827                 cpu_set(cpu, cpu_core_map[cpu]);
828                 return 0;
829         }
830
831         set_cpu_sibling_map(cpu);
832
833         return 0;
834 }
835
836 /*
837  * Assume that CPUs have been discovered by some platform-dependent interface.  For
838  * SoftSDV/Lion, that would be ACPI.
839  *
840  * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
841  */
842 void __init
843 init_smp_config(void)
844 {
845         struct fptr {
846                 unsigned long fp;
847                 unsigned long gp;
848         } *ap_startup;
849         long sal_ret;
850
851         /* Tell SAL where to drop the APs.  */
852         ap_startup = (struct fptr *) start_ap;
853         sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
854                                        ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
855         if (sal_ret < 0)
856                 printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
857                        ia64_sal_strerror(sal_ret));
858 }
859
860 /*
861  * identify_siblings(cpu) gets called from identify_cpu. This populates the 
862  * information related to logical execution units in per_cpu_data structure.
863  */
864 void __devinit
865 identify_siblings(struct cpuinfo_ia64 *c)
866 {
867         long status;
868         u16 pltid;
869         pal_logical_to_physical_t info;
870
871         status = ia64_pal_logical_to_phys(-1, &info);
872         if (status != PAL_STATUS_SUCCESS) {
873                 if (status != PAL_STATUS_UNIMPLEMENTED) {
874                         printk(KERN_ERR
875                                 "ia64_pal_logical_to_phys failed with %ld\n",
876                                 status);
877                         return;
878                 }
879
880                 info.overview_ppid = 0;
881                 info.overview_cpp  = 1;
882                 info.overview_tpc  = 1;
883         }
884
885         status = ia64_sal_physical_id_info(&pltid);
886         if (status != PAL_STATUS_SUCCESS) {
887                 if (status != PAL_STATUS_UNIMPLEMENTED)
888                         printk(KERN_ERR
889                                 "ia64_sal_pltid failed with %ld\n",
890                                 status);
891                 return;
892         }
893
894         c->socket_id =  (pltid << 8) | info.overview_ppid;
895
896         if (info.overview_cpp == 1 && info.overview_tpc == 1)
897                 return;
898
899         c->cores_per_socket = info.overview_cpp;
900         c->threads_per_core = info.overview_tpc;
901         c->num_log = info.overview_num_log;
902
903         c->core_id = info.log1_cid;
904         c->thread_id = info.log1_tid;
905 }
906
907 /*
908  * returns non zero, if multi-threading is enabled
909  * on at least one physical package. Due to hotplug cpu
910  * and (maxcpus=), all threads may not necessarily be enabled
911  * even though the processor supports multi-threading.
912  */
913 int is_multithreading_enabled(void)
914 {
915         int i, j;
916
917         for_each_present_cpu(i) {
918                 for_each_present_cpu(j) {
919                         if (j == i)
920                                 continue;
921                         if ((cpu_data(j)->socket_id == cpu_data(i)->socket_id)) {
922                                 if (cpu_data(j)->core_id == cpu_data(i)->core_id)
923                                         return 1;
924                         }
925                 }
926         }
927         return 0;
928 }
929 EXPORT_SYMBOL_GPL(is_multithreading_enabled);