Pull bsp-removal into release branch
[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 #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 #ifdef CONFIG_PERMIT_BSP_REMOVE
74 #define bsp_remove_ok   1
75 #else
76 #define bsp_remove_ok   0
77 #endif
78
79 /*
80  * Store all idle threads, this can be reused instead of creating
81  * a new thread. Also avoids complicated thread destroy functionality
82  * for idle threads.
83  */
84 struct task_struct *idle_thread_array[NR_CPUS];
85
86 /*
87  * Global array allocated for NR_CPUS at boot time
88  */
89 struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
90
91 /*
92  * start_ap in head.S uses this to store current booting cpu
93  * info.
94  */
95 struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
96
97 #define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
98
99 #define get_idle_for_cpu(x)             (idle_thread_array[(x)])
100 #define set_idle_for_cpu(x,p)   (idle_thread_array[(x)] = (p))
101
102 #else
103
104 #define get_idle_for_cpu(x)             (NULL)
105 #define set_idle_for_cpu(x,p)
106 #define set_brendez_area(x)
107 #endif
108
109
110 /*
111  * ITC synchronization related stuff:
112  */
113 #define MASTER  (0)
114 #define SLAVE   (SMP_CACHE_BYTES/8)
115
116 #define NUM_ROUNDS      64      /* magic value */
117 #define NUM_ITERS       5       /* likewise */
118
119 static DEFINE_SPINLOCK(itc_sync_lock);
120 static volatile unsigned long go[SLAVE + 1];
121
122 #define DEBUG_ITC_SYNC  0
123
124 extern void __devinit calibrate_delay (void);
125 extern void start_ap (void);
126 extern unsigned long ia64_iobase;
127
128 task_t *task_for_booting_cpu;
129
130 /*
131  * State for each CPU
132  */
133 DEFINE_PER_CPU(int, cpu_state);
134
135 /* Bitmasks of currently online, and possible CPUs */
136 cpumask_t cpu_online_map;
137 EXPORT_SYMBOL(cpu_online_map);
138 cpumask_t cpu_possible_map = CPU_MASK_NONE;
139 EXPORT_SYMBOL(cpu_possible_map);
140
141 cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
142 cpumask_t cpu_sibling_map[NR_CPUS] __cacheline_aligned;
143 int smp_num_siblings = 1;
144 int smp_num_cpucores = 1;
145
146 /* which logical CPU number maps to which CPU (physical APIC ID) */
147 volatile int ia64_cpu_to_sapicid[NR_CPUS];
148 EXPORT_SYMBOL(ia64_cpu_to_sapicid);
149
150 static volatile cpumask_t cpu_callin_map;
151
152 struct smp_boot_data smp_boot_data __initdata;
153
154 unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
155
156 char __initdata no_int_routing;
157
158 unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
159
160 #ifdef CONFIG_FORCE_CPEI_RETARGET
161 #define CPEI_OVERRIDE_DEFAULT   (1)
162 #else
163 #define CPEI_OVERRIDE_DEFAULT   (0)
164 #endif
165
166 unsigned int force_cpei_retarget = CPEI_OVERRIDE_DEFAULT;
167
168 static int __init
169 cmdl_force_cpei(char *str)
170 {
171         int value=0;
172
173         get_option (&str, &value);
174         force_cpei_retarget = value;
175
176         return 1;
177 }
178
179 __setup("force_cpei=", cmdl_force_cpei);
180
181 static int __init
182 nointroute (char *str)
183 {
184         no_int_routing = 1;
185         printk ("no_int_routing on\n");
186         return 1;
187 }
188
189 __setup("nointroute", nointroute);
190
191 static void fix_b0_for_bsp(void)
192 {
193 #ifdef CONFIG_HOTPLUG_CPU
194         int cpuid;
195         static int fix_bsp_b0 = 1;
196
197         cpuid = smp_processor_id();
198
199         /*
200          * Cache the b0 value on the first AP that comes up
201          */
202         if (!(fix_bsp_b0 && cpuid))
203                 return;
204
205         sal_boot_rendez_state[0].br[0] = sal_boot_rendez_state[cpuid].br[0];
206         printk ("Fixed BSP b0 value from CPU %d\n", cpuid);
207
208         fix_bsp_b0 = 0;
209 #endif
210 }
211
212 void
213 sync_master (void *arg)
214 {
215         unsigned long flags, i;
216
217         go[MASTER] = 0;
218
219         local_irq_save(flags);
220         {
221                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
222                         while (!go[MASTER])
223                                 cpu_relax();
224                         go[MASTER] = 0;
225                         go[SLAVE] = ia64_get_itc();
226                 }
227         }
228         local_irq_restore(flags);
229 }
230
231 /*
232  * Return the number of cycles by which our itc differs from the itc on the master
233  * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
234  * negative that it is behind.
235  */
236 static inline long
237 get_delta (long *rt, long *master)
238 {
239         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
240         unsigned long tcenter, t0, t1, tm;
241         long i;
242
243         for (i = 0; i < NUM_ITERS; ++i) {
244                 t0 = ia64_get_itc();
245                 go[MASTER] = 1;
246                 while (!(tm = go[SLAVE]))
247                         cpu_relax();
248                 go[SLAVE] = 0;
249                 t1 = ia64_get_itc();
250
251                 if (t1 - t0 < best_t1 - best_t0)
252                         best_t0 = t0, best_t1 = t1, best_tm = tm;
253         }
254
255         *rt = best_t1 - best_t0;
256         *master = best_tm - best_t0;
257
258         /* average best_t0 and best_t1 without overflow: */
259         tcenter = (best_t0/2 + best_t1/2);
260         if (best_t0 % 2 + best_t1 % 2 == 2)
261                 ++tcenter;
262         return tcenter - best_tm;
263 }
264
265 /*
266  * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
267  * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
268  * unaccounted-for errors (such as getting a machine check in the middle of a calibration
269  * step).  The basic idea is for the slave to ask the master what itc value it has and to
270  * read its own itc before and after the master responds.  Each iteration gives us three
271  * timestamps:
272  *
273  *      slave           master
274  *
275  *      t0 ---\
276  *             ---\
277  *                 --->
278  *                      tm
279  *                 /---
280  *             /---
281  *      t1 <---
282  *
283  *
284  * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
285  * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
286  * between the slave and the master is symmetric.  Even if the interconnect were
287  * asymmetric, we would still know that the synchronization error is smaller than the
288  * roundtrip latency (t0 - t1).
289  *
290  * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
291  * within one or two cycles.  However, we can only *guarantee* that the synchronization is
292  * accurate to within a round-trip time, which is typically in the range of several
293  * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
294  * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
295  * than half a micro second or so.
296  */
297 void
298 ia64_sync_itc (unsigned int master)
299 {
300         long i, delta, adj, adjust_latency = 0, done = 0;
301         unsigned long flags, rt, master_time_stamp, bound;
302 #if DEBUG_ITC_SYNC
303         struct {
304                 long rt;        /* roundtrip time */
305                 long master;    /* master's timestamp */
306                 long diff;      /* difference between midpoint and master's timestamp */
307                 long lat;       /* estimate of itc adjustment latency */
308         } t[NUM_ROUNDS];
309 #endif
310
311         /*
312          * Make sure local timer ticks are disabled while we sync.  If
313          * they were enabled, we'd have to worry about nasty issues
314          * like setting the ITC ahead of (or a long time before) the
315          * next scheduled tick.
316          */
317         BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
318
319         go[MASTER] = 1;
320
321         if (smp_call_function_single(master, sync_master, NULL, 1, 0) < 0) {
322                 printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
323                 return;
324         }
325
326         while (go[MASTER])
327                 cpu_relax();    /* wait for master to be ready */
328
329         spin_lock_irqsave(&itc_sync_lock, flags);
330         {
331                 for (i = 0; i < NUM_ROUNDS; ++i) {
332                         delta = get_delta(&rt, &master_time_stamp);
333                         if (delta == 0) {
334                                 done = 1;       /* let's lock on to this... */
335                                 bound = rt;
336                         }
337
338                         if (!done) {
339                                 if (i > 0) {
340                                         adjust_latency += -delta;
341                                         adj = -delta + adjust_latency/4;
342                                 } else
343                                         adj = -delta;
344
345                                 ia64_set_itc(ia64_get_itc() + adj);
346                         }
347 #if DEBUG_ITC_SYNC
348                         t[i].rt = rt;
349                         t[i].master = master_time_stamp;
350                         t[i].diff = delta;
351                         t[i].lat = adjust_latency/4;
352 #endif
353                 }
354         }
355         spin_unlock_irqrestore(&itc_sync_lock, flags);
356
357 #if DEBUG_ITC_SYNC
358         for (i = 0; i < NUM_ROUNDS; ++i)
359                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
360                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
361 #endif
362
363         printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
364                "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
365 }
366
367 /*
368  * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
369  */
370 static inline void __devinit
371 smp_setup_percpu_timer (void)
372 {
373 }
374
375 static void __devinit
376 smp_callin (void)
377 {
378         int cpuid, phys_id, itc_master;
379         extern void ia64_init_itm(void);
380         extern volatile int time_keeper_id;
381
382 #ifdef CONFIG_PERFMON
383         extern void pfm_init_percpu(void);
384 #endif
385
386         cpuid = smp_processor_id();
387         phys_id = hard_smp_processor_id();
388         itc_master = time_keeper_id;
389
390         if (cpu_online(cpuid)) {
391                 printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
392                        phys_id, cpuid);
393                 BUG();
394         }
395
396         fix_b0_for_bsp();
397
398         lock_ipi_calllock();
399         cpu_set(cpuid, cpu_online_map);
400         unlock_ipi_calllock();
401         per_cpu(cpu_state, cpuid) = CPU_ONLINE;
402
403         smp_setup_percpu_timer();
404
405         ia64_mca_cmc_vector_setup();    /* Setup vector on AP */
406
407 #ifdef CONFIG_PERFMON
408         pfm_init_percpu();
409 #endif
410
411         local_irq_enable();
412
413         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
414                 /*
415                  * Synchronize the ITC with the BP.  Need to do this after irqs are
416                  * enabled because ia64_sync_itc() calls smp_call_function_single(), which
417                  * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
418                  * local_bh_enable(), which bugs out if irqs are not enabled...
419                  */
420                 Dprintk("Going to syncup ITC with ITC Master.\n");
421                 ia64_sync_itc(itc_master);
422         }
423
424         /*
425          * Get our bogomips.
426          */
427         ia64_init_itm();
428         calibrate_delay();
429         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
430
431 #ifdef CONFIG_IA32_SUPPORT
432         ia32_gdt_init();
433 #endif
434
435         /*
436          * Allow the master to continue.
437          */
438         cpu_set(cpuid, cpu_callin_map);
439         Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
440 }
441
442
443 /*
444  * Activate a secondary processor.  head.S calls this.
445  */
446 int __devinit
447 start_secondary (void *unused)
448 {
449         /* Early console may use I/O ports */
450         ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
451         Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
452         efi_map_pal_code();
453         cpu_init();
454         preempt_disable();
455         smp_callin();
456
457         cpu_idle();
458         return 0;
459 }
460
461 struct pt_regs * __devinit idle_regs(struct pt_regs *regs)
462 {
463         return NULL;
464 }
465
466 struct create_idle {
467         struct task_struct *idle;
468         struct completion done;
469         int cpu;
470 };
471
472 void
473 do_fork_idle(void *_c_idle)
474 {
475         struct create_idle *c_idle = _c_idle;
476
477         c_idle->idle = fork_idle(c_idle->cpu);
478         complete(&c_idle->done);
479 }
480
481 static int __devinit
482 do_boot_cpu (int sapicid, int cpu)
483 {
484         int timeout;
485         struct create_idle c_idle = {
486                 .cpu    = cpu,
487                 .done   = COMPLETION_INITIALIZER(c_idle.done),
488         };
489         DECLARE_WORK(work, do_fork_idle, &c_idle);
490
491         c_idle.idle = get_idle_for_cpu(cpu);
492         if (c_idle.idle) {
493                 init_idle(c_idle.idle, cpu);
494                 goto do_rest;
495         }
496
497         /*
498          * We can't use kernel_thread since we must avoid to reschedule the child.
499          */
500         if (!keventd_up() || current_is_keventd())
501                 work.func(work.data);
502         else {
503                 schedule_work(&work);
504                 wait_for_completion(&c_idle.done);
505         }
506
507         if (IS_ERR(c_idle.idle))
508                 panic("failed fork for CPU %d", cpu);
509
510         set_idle_for_cpu(cpu, c_idle.idle);
511
512 do_rest:
513         task_for_booting_cpu = c_idle.idle;
514
515         Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
516
517         set_brendez_area(cpu);
518         platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
519
520         /*
521          * Wait 10s total for the AP to start
522          */
523         Dprintk("Waiting on callin_map ...");
524         for (timeout = 0; timeout < 100000; timeout++) {
525                 if (cpu_isset(cpu, cpu_callin_map))
526                         break;  /* It has booted */
527                 udelay(100);
528         }
529         Dprintk("\n");
530
531         if (!cpu_isset(cpu, cpu_callin_map)) {
532                 printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
533                 ia64_cpu_to_sapicid[cpu] = -1;
534                 cpu_clear(cpu, cpu_online_map);  /* was set in smp_callin() */
535                 return -EINVAL;
536         }
537         return 0;
538 }
539
540 static int __init
541 decay (char *str)
542 {
543         int ticks;
544         get_option (&str, &ticks);
545         return 1;
546 }
547
548 __setup("decay=", decay);
549
550 /*
551  * Initialize the logical CPU number to SAPICID mapping
552  */
553 void __init
554 smp_build_cpu_map (void)
555 {
556         int sapicid, cpu, i;
557         int boot_cpu_id = hard_smp_processor_id();
558
559         for (cpu = 0; cpu < NR_CPUS; cpu++) {
560                 ia64_cpu_to_sapicid[cpu] = -1;
561         }
562
563         ia64_cpu_to_sapicid[0] = boot_cpu_id;
564         cpus_clear(cpu_present_map);
565         cpu_set(0, cpu_present_map);
566         cpu_set(0, cpu_possible_map);
567         for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
568                 sapicid = smp_boot_data.cpu_phys_id[i];
569                 if (sapicid == boot_cpu_id)
570                         continue;
571                 cpu_set(cpu, cpu_present_map);
572                 cpu_set(cpu, cpu_possible_map);
573                 ia64_cpu_to_sapicid[cpu] = sapicid;
574                 cpu++;
575         }
576 }
577
578 /*
579  * Cycle through the APs sending Wakeup IPIs to boot each.
580  */
581 void __init
582 smp_prepare_cpus (unsigned int max_cpus)
583 {
584         int boot_cpu_id = hard_smp_processor_id();
585
586         /*
587          * Initialize the per-CPU profiling counter/multiplier
588          */
589
590         smp_setup_percpu_timer();
591
592         /*
593          * We have the boot CPU online for sure.
594          */
595         cpu_set(0, cpu_online_map);
596         cpu_set(0, cpu_callin_map);
597
598         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
599         ia64_cpu_to_sapicid[0] = boot_cpu_id;
600
601         printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
602
603         current_thread_info()->cpu = 0;
604
605         /*
606          * If SMP should be disabled, then really disable it!
607          */
608         if (!max_cpus) {
609                 printk(KERN_INFO "SMP mode deactivated.\n");
610                 cpus_clear(cpu_online_map);
611                 cpus_clear(cpu_present_map);
612                 cpus_clear(cpu_possible_map);
613                 cpu_set(0, cpu_online_map);
614                 cpu_set(0, cpu_present_map);
615                 cpu_set(0, cpu_possible_map);
616                 return;
617         }
618 }
619
620 void __devinit smp_prepare_boot_cpu(void)
621 {
622         cpu_set(smp_processor_id(), cpu_online_map);
623         cpu_set(smp_processor_id(), cpu_callin_map);
624         per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
625 }
626
627 /*
628  * mt_info[] is a temporary store for all info returned by
629  * PAL_LOGICAL_TO_PHYSICAL, to be copied into cpuinfo_ia64 when the
630  * specific cpu comes.
631  */
632 static struct {
633         __u32   socket_id;
634         __u16   core_id;
635         __u16   thread_id;
636         __u16   proc_fixed_addr;
637         __u8    valid;
638 } mt_info[NR_CPUS] __devinitdata;
639
640 #ifdef CONFIG_HOTPLUG_CPU
641 static inline void
642 remove_from_mtinfo(int cpu)
643 {
644         int i;
645
646         for_each_cpu(i)
647                 if (mt_info[i].valid &&  mt_info[i].socket_id ==
648                                                 cpu_data(cpu)->socket_id)
649                         mt_info[i].valid = 0;
650 }
651
652 static inline void
653 clear_cpu_sibling_map(int cpu)
654 {
655         int i;
656
657         for_each_cpu_mask(i, cpu_sibling_map[cpu])
658                 cpu_clear(cpu, cpu_sibling_map[i]);
659         for_each_cpu_mask(i, cpu_core_map[cpu])
660                 cpu_clear(cpu, cpu_core_map[i]);
661
662         cpu_sibling_map[cpu] = cpu_core_map[cpu] = CPU_MASK_NONE;
663 }
664
665 static void
666 remove_siblinginfo(int cpu)
667 {
668         int last = 0;
669
670         if (cpu_data(cpu)->threads_per_core == 1 &&
671             cpu_data(cpu)->cores_per_socket == 1) {
672                 cpu_clear(cpu, cpu_core_map[cpu]);
673                 cpu_clear(cpu, cpu_sibling_map[cpu]);
674                 return;
675         }
676
677         last = (cpus_weight(cpu_core_map[cpu]) == 1 ? 1 : 0);
678
679         /* remove it from all sibling map's */
680         clear_cpu_sibling_map(cpu);
681
682         /* if this cpu is the last in the core group, remove all its info 
683          * from mt_info structure
684          */
685         if (last)
686                 remove_from_mtinfo(cpu);
687 }
688
689 extern void fixup_irqs(void);
690
691 int migrate_platform_irqs(unsigned int cpu)
692 {
693         int new_cpei_cpu;
694         irq_desc_t *desc = NULL;
695         cpumask_t       mask;
696         int             retval = 0;
697
698         /*
699          * dont permit CPEI target to removed.
700          */
701         if (cpe_vector > 0 && is_cpu_cpei_target(cpu)) {
702                 printk ("CPU (%d) is CPEI Target\n", cpu);
703                 if (can_cpei_retarget()) {
704                         /*
705                          * Now re-target the CPEI to a different processor
706                          */
707                         new_cpei_cpu = any_online_cpu(cpu_online_map);
708                         mask = cpumask_of_cpu(new_cpei_cpu);
709                         set_cpei_target_cpu(new_cpei_cpu);
710                         desc = irq_descp(ia64_cpe_irq);
711                         /*
712                          * Switch for now, immediatly, we need to do fake intr
713                          * as other interrupts, but need to study CPEI behaviour with
714                          * polling before making changes.
715                          */
716                         if (desc) {
717                                 desc->handler->disable(ia64_cpe_irq);
718                                 desc->handler->set_affinity(ia64_cpe_irq, mask);
719                                 desc->handler->enable(ia64_cpe_irq);
720                                 printk ("Re-targetting CPEI to cpu %d\n", new_cpei_cpu);
721                         }
722                 }
723                 if (!desc) {
724                         printk ("Unable to retarget CPEI, offline cpu [%d] failed\n", cpu);
725                         retval = -EBUSY;
726                 }
727         }
728         return retval;
729 }
730
731 /* must be called with cpucontrol mutex held */
732 int __cpu_disable(void)
733 {
734         int cpu = smp_processor_id();
735
736         /*
737          * dont permit boot processor for now
738          */
739         if (cpu == 0 && !bsp_remove_ok) {
740                 printk ("Your platform does not support removal of BSP\n");
741                 return (-EBUSY);
742         }
743
744         cpu_clear(cpu, cpu_online_map);
745
746         if (migrate_platform_irqs(cpu)) {
747                 cpu_set(cpu, cpu_online_map);
748                 return (-EBUSY);
749         }
750
751         remove_siblinginfo(cpu);
752         cpu_clear(cpu, cpu_online_map);
753         fixup_irqs();
754         local_flush_tlb_all();
755         cpu_clear(cpu, cpu_callin_map);
756         return 0;
757 }
758
759 void __cpu_die(unsigned int cpu)
760 {
761         unsigned int i;
762
763         for (i = 0; i < 100; i++) {
764                 /* They ack this in play_dead by setting CPU_DEAD */
765                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
766                 {
767                         printk ("CPU %d is now offline\n", cpu);
768                         return;
769                 }
770                 msleep(100);
771         }
772         printk(KERN_ERR "CPU %u didn't die...\n", cpu);
773 }
774 #else /* !CONFIG_HOTPLUG_CPU */
775 int __cpu_disable(void)
776 {
777         return -ENOSYS;
778 }
779
780 void __cpu_die(unsigned int cpu)
781 {
782         /* We said "no" in __cpu_disable */
783         BUG();
784 }
785 #endif /* CONFIG_HOTPLUG_CPU */
786
787 void
788 smp_cpus_done (unsigned int dummy)
789 {
790         int cpu;
791         unsigned long bogosum = 0;
792
793         /*
794          * Allow the user to impress friends.
795          */
796
797         for_each_online_cpu(cpu) {
798                 bogosum += cpu_data(cpu)->loops_per_jiffy;
799         }
800
801         printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
802                (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
803 }
804
805 static inline void __devinit
806 set_cpu_sibling_map(int cpu)
807 {
808         int i;
809
810         for_each_online_cpu(i) {
811                 if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
812                         cpu_set(i, cpu_core_map[cpu]);
813                         cpu_set(cpu, cpu_core_map[i]);
814                         if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
815                                 cpu_set(i, cpu_sibling_map[cpu]);
816                                 cpu_set(cpu, cpu_sibling_map[i]);
817                         }
818                 }
819         }
820 }
821
822 int __devinit
823 __cpu_up (unsigned int cpu)
824 {
825         int ret;
826         int sapicid;
827
828         sapicid = ia64_cpu_to_sapicid[cpu];
829         if (sapicid == -1)
830                 return -EINVAL;
831
832         /*
833          * Already booted cpu? not valid anymore since we dont
834          * do idle loop tightspin anymore.
835          */
836         if (cpu_isset(cpu, cpu_callin_map))
837                 return -EINVAL;
838
839         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
840         /* Processor goes to start_secondary(), sets online flag */
841         ret = do_boot_cpu(sapicid, cpu);
842         if (ret < 0)
843                 return ret;
844
845         if (cpu_data(cpu)->threads_per_core == 1 &&
846             cpu_data(cpu)->cores_per_socket == 1) {
847                 cpu_set(cpu, cpu_sibling_map[cpu]);
848                 cpu_set(cpu, cpu_core_map[cpu]);
849                 return 0;
850         }
851
852         set_cpu_sibling_map(cpu);
853
854         return 0;
855 }
856
857 /*
858  * Assume that CPU's have been discovered by some platform-dependent interface.  For
859  * SoftSDV/Lion, that would be ACPI.
860  *
861  * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
862  */
863 void __init
864 init_smp_config(void)
865 {
866         struct fptr {
867                 unsigned long fp;
868                 unsigned long gp;
869         } *ap_startup;
870         long sal_ret;
871
872         /* Tell SAL where to drop the AP's.  */
873         ap_startup = (struct fptr *) start_ap;
874         sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
875                                        ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
876         if (sal_ret < 0)
877                 printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
878                        ia64_sal_strerror(sal_ret));
879 }
880
881 static inline int __devinit
882 check_for_mtinfo_index(void)
883 {
884         int i;
885         
886         for_each_cpu(i)
887                 if (!mt_info[i].valid)
888                         return i;
889
890         return -1;
891 }
892
893 /*
894  * Search the mt_info to find out if this socket's cid/tid information is
895  * cached or not. If the socket exists, fill in the core_id and thread_id 
896  * in cpuinfo
897  */
898 static int __devinit
899 check_for_new_socket(__u16 logical_address, struct cpuinfo_ia64 *c)
900 {
901         int i;
902         __u32 sid = c->socket_id;
903
904         for_each_cpu(i) {
905                 if (mt_info[i].valid && mt_info[i].proc_fixed_addr == logical_address
906                     && mt_info[i].socket_id == sid) {
907                         c->core_id = mt_info[i].core_id;
908                         c->thread_id = mt_info[i].thread_id;
909                         return 1; /* not a new socket */
910                 }
911         }
912         return 0;
913 }
914
915 /*
916  * identify_siblings(cpu) gets called from identify_cpu. This populates the 
917  * information related to logical execution units in per_cpu_data structure.
918  */
919 void __devinit
920 identify_siblings(struct cpuinfo_ia64 *c)
921 {
922         s64 status;
923         u16 pltid;
924         u64 proc_fixed_addr;
925         int count, i;
926         pal_logical_to_physical_t info;
927
928         if (smp_num_cpucores == 1 && smp_num_siblings == 1)
929                 return;
930
931         if ((status = ia64_pal_logical_to_phys(0, &info)) != PAL_STATUS_SUCCESS) {
932                 printk(KERN_ERR "ia64_pal_logical_to_phys failed with %ld\n",
933                        status);
934                 return;
935         }
936         if ((status = ia64_sal_physical_id_info(&pltid)) != PAL_STATUS_SUCCESS) {
937                 printk(KERN_ERR "ia64_sal_pltid failed with %ld\n", status);
938                 return;
939         }
940         if ((status = ia64_pal_fixed_addr(&proc_fixed_addr)) != PAL_STATUS_SUCCESS) {
941                 printk(KERN_ERR "ia64_pal_fixed_addr failed with %ld\n", status);
942                 return;
943         }
944
945         c->socket_id =  (pltid << 8) | info.overview_ppid;
946         c->cores_per_socket = info.overview_cpp;
947         c->threads_per_core = info.overview_tpc;
948         count = c->num_log = info.overview_num_log;
949
950         /* If the thread and core id information is already cached, then
951          * we will simply update cpu_info and return. Otherwise, we will
952          * do the PAL calls and cache core and thread id's of all the siblings.
953          */
954         if (check_for_new_socket(proc_fixed_addr, c))
955                 return;
956
957         for (i = 0; i < count; i++) {
958                 int index;
959
960                 if (i && (status = ia64_pal_logical_to_phys(i, &info))
961                           != PAL_STATUS_SUCCESS) {
962                         printk(KERN_ERR "ia64_pal_logical_to_phys failed"
963                                         " with %ld\n", status);
964                         return;
965                 }
966                 if (info.log2_la == proc_fixed_addr) {
967                         c->core_id = info.log1_cid;
968                         c->thread_id = info.log1_tid;
969                 }
970
971                 index = check_for_mtinfo_index();
972                 /* We will not do the mt_info caching optimization in this case.
973                  */
974                 if (index < 0)
975                         continue;
976
977                 mt_info[index].valid = 1;
978                 mt_info[index].socket_id = c->socket_id;
979                 mt_info[index].core_id = info.log1_cid;
980                 mt_info[index].thread_id = info.log1_tid;
981                 mt_info[index].proc_fixed_addr = info.log2_la;
982         }
983 }