arch/ia64/kernel/smpboot.c

   1 /*
   2  * SMP boot-related support
   3  *
   4  * Copyright (C) 1998-2003, 2005 Hewlett-Packard Co
   5  *      David Mosberger-Tang <davidm@hpl.hp.com>
   6  * Copyright (C) 2001, 2004-2005 Intel Corp
   7  *      Rohit Seth <rohit.seth@intel.com>
   8  *      Suresh Siddha <suresh.b.siddha@intel.com>
   9  *      Gordon Jin <gordon.jin@intel.com>
  10  *      Ashok Raj  <ashok.raj@intel.com>
  11  *
  12  * 01/05/16 Rohit Seth <rohit.seth@intel.com>   Moved SMP booting functions from smp.c to here.
  13  * 01/04/27 David Mosberger <davidm@hpl.hp.com> Added ITC synching code.
  14  * 02/07/31 David Mosberger <davidm@hpl.hp.com> Switch over to hotplug-CPU boot-sequence.
  15  *                                              smp_boot_cpus()/smp_commence() is replaced by
  16  *                                              smp_prepare_cpus()/__cpu_up()/smp_cpus_done().
  17  * 04/06/21 Ashok Raj           <ashok.raj@intel.com> Added CPU Hotplug Support
  18  * 04/12/26 Jin Gordon <gordon.jin@intel.com>
  19  * 04/12/26 Rohit Seth <rohit.seth@intel.com>
  20  *                                              Add multi-threading and multi-core detection
  21  * 05/01/30 Suresh Siddha <suresh.b.siddha@intel.com>
  22  *                                              Setup cpu_sibling_map and cpu_core_map
  23  */
  24
  25 #include <linux/module.h>
  26 #include <linux/acpi.h>
  27 #include <linux/bootmem.h>
  28 #include <linux/cpu.h>
  29 #include <linux/delay.h>
  30 #include <linux/init.h>
  31 #include <linux/interrupt.h>
  32 #include <linux/irq.h>
  33 #include <linux/kernel.h>
  34 #include <linux/kernel_stat.h>
  35 #include <linux/mm.h>
  36 #include <linux/notifier.h>
  37 #include <linux/smp.h>
  38 #include <linux/spinlock.h>
  39 #include <linux/efi.h>
  40 #include <linux/percpu.h>
  41 #include <linux/bitops.h>
  42
  43 #include <asm/atomic.h>
  44 #include <asm/cache.h>
  45 #include <asm/current.h>
  46 #include <asm/delay.h>
  47 #include <asm/ia32.h>
  48 #include <asm/io.h>
  49 #include <asm/irq.h>
  50 #include <asm/machvec.h>
  51 #include <asm/mca.h>
  52 #include <asm/page.h>
  53 #include <asm/paravirt.h>
  54 #include <asm/pgalloc.h>
  55 #include <asm/pgtable.h>
  56 #include <asm/processor.h>
  57 #include <asm/ptrace.h>
  58 #include <asm/sal.h>
  59 #include <asm/system.h>
  60 #include <asm/tlbflush.h>
  61 #include <asm/unistd.h>
  62 #include <asm/sn/arch.h>
  63
  64 #define SMP_DEBUG 0
  65
  66 #if SMP_DEBUG
  67 #define Dprintk(x...)  printk(x)
  68 #else
  69 #define Dprintk(x...)
  70 #endif
  71
  72 #ifdef CONFIG_HOTPLUG_CPU
  73 #ifdef CONFIG_PERMIT_BSP_REMOVE
  74 #define bsp_remove_ok   1
  75 #else
  76 #define bsp_remove_ok   0
  77 #endif
  78
  79 /*
  80  * Store all idle threads, this can be reused instead of creating
  81  * a new thread. Also avoids complicated thread destroy functionality
  82  * for idle threads.
  83  */
  84 struct task_struct *idle_thread_array[NR_CPUS];
  85
  86 /*
  87  * Global array allocated for NR_CPUS at boot time
  88  */
  89 struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
  90
  91 /*
  92  * start_ap in head.S uses this to store current booting cpu
  93  * info.
  94  */
  95 struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
  96
  97 #define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
  98
  99 #define get_idle_for_cpu(x)             (idle_thread_array[(x)])
 100 #define set_idle_for_cpu(x,p)   (idle_thread_array[(x)] = (p))
 101
 102 #else
 103
 104 #define get_idle_for_cpu(x)             (NULL)
 105 #define set_idle_for_cpu(x,p)
 106 #define set_brendez_area(x)
 107 #endif
 108
 109
 110 /*
 111  * ITC synchronization related stuff:
 112  */
 113 #define MASTER  (0)
 114 #define SLAVE   (SMP_CACHE_BYTES/8)
 115
 116 #define NUM_ROUNDS      64      /* magic value */
 117 #define NUM_ITERS       5       /* likewise */
 118
 119 static DEFINE_SPINLOCK(itc_sync_lock);
 120 static volatile unsigned long go[SLAVE + 1];
 121
 122 #define DEBUG_ITC_SYNC  0
 123
 124 extern void start_ap (void);
 125 extern unsigned long ia64_iobase;
 126
 127 struct task_struct *task_for_booting_cpu;
 128
 129 /*
 130  * State for each CPU
 131  */
 132 DEFINE_PER_CPU(int, cpu_state);
 133
 134 /* Bitmasks of currently online, and possible CPUs */
 135 cpumask_t cpu_online_map;
 136 EXPORT_SYMBOL(cpu_online_map);
 137 cpumask_t cpu_possible_map = CPU_MASK_NONE;
 138 EXPORT_SYMBOL(cpu_possible_map);
 139
 140 cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
 141 DEFINE_PER_CPU_SHARED_ALIGNED(cpumask_t, cpu_sibling_map);
 142 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
 143
 144 int smp_num_siblings = 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, 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 __cpuinit
 376 smp_callin (void)
 377 {
 378         int cpuid, phys_id, itc_master;
 379         struct cpuinfo_ia64 *last_cpuinfo, *this_cpuinfo;
 380         extern void ia64_init_itm(void);
 381         extern volatile int time_keeper_id;
 382
 383 #ifdef CONFIG_PERFMON
 384         extern void pfm_init_percpu(void);
 385 #endif
 386
 387         cpuid = smp_processor_id();
 388         phys_id = hard_smp_processor_id();
 389         itc_master = time_keeper_id;
 390
 391         if (cpu_online(cpuid)) {
 392                 printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
 393                        phys_id, cpuid);
 394                 BUG();
 395         }
 396
 397         fix_b0_for_bsp();
 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         cpu_set(cpuid, cpu_online_map);
 404         per_cpu(cpu_state, cpuid) = CPU_ONLINE;
 405         spin_unlock(&vector_lock);
 406         ipi_call_unlock_irq();
 407
 408         smp_setup_percpu_timer();
 409
 410         ia64_mca_cmc_vector_setup();    /* Setup vector on AP */
 411
 412 #ifdef CONFIG_PERFMON
 413         pfm_init_percpu();
 414 #endif
 415
 416         local_irq_enable();
 417
 418         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
 419                 /*
 420                  * Synchronize the ITC with the BP.  Need to do this after irqs are
 421                  * enabled because ia64_sync_itc() calls smp_call_function_single(), which
 422                  * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
 423                  * local_bh_enable(), which bugs out if irqs are not enabled...
 424                  */
 425                 Dprintk("Going to syncup ITC with ITC Master.\n");
 426                 ia64_sync_itc(itc_master);
 427         }
 428
 429         /*
 430          * Get our bogomips.
 431          */
 432         ia64_init_itm();
 433
 434         /*
 435          * Delay calibration can be skipped if new processor is identical to the
 436          * previous processor.
 437          */
 438         last_cpuinfo = cpu_data(cpuid - 1);
 439         this_cpuinfo = local_cpu_data;
 440         if (last_cpuinfo->itc_freq != this_cpuinfo->itc_freq ||
 441             last_cpuinfo->proc_freq != this_cpuinfo->proc_freq ||
 442             last_cpuinfo->features != this_cpuinfo->features ||
 443             last_cpuinfo->revision != this_cpuinfo->revision ||
 444             last_cpuinfo->family != this_cpuinfo->family ||
 445             last_cpuinfo->archrev != this_cpuinfo->archrev ||
 446             last_cpuinfo->model != this_cpuinfo->model)
 447                 calibrate_delay();
 448         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
 449
 450 #ifdef CONFIG_IA32_SUPPORT
 451         ia32_gdt_init();
 452 #endif
 453
 454         /*
 455          * Allow the master to continue.
 456          */
 457         cpu_set(cpuid, cpu_callin_map);
 458         Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
 459 }
 460
 461
 462 /*
 463  * Activate a secondary processor.  head.S calls this.
 464  */
 465 int __cpuinit
 466 start_secondary (void *unused)
 467 {
 468         /* Early console may use I/O ports */
 469         ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
 470 #ifndef CONFIG_PRINTK_TIME
 471         Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
 472 #endif
 473         efi_map_pal_code();
 474         cpu_init();
 475         preempt_disable();
 476         smp_callin();
 477
 478         cpu_idle();
 479         return 0;
 480 }
 481
 482 struct pt_regs * __cpuinit idle_regs(struct pt_regs *regs)
 483 {
 484         return NULL;
 485 }
 486
 487 struct create_idle {
 488         struct work_struct work;
 489         struct task_struct *idle;
 490         struct completion done;
 491         int cpu;
 492 };
 493
 494 void __cpuinit
 495 do_fork_idle(struct work_struct *work)
 496 {
 497         struct create_idle *c_idle =
 498                 container_of(work, struct create_idle, work);
 499
 500         c_idle->idle = fork_idle(c_idle->cpu);
 501         complete(&c_idle->done);
 502 }
 503
 504 static int __cpuinit
 505 do_boot_cpu (int sapicid, int cpu)
 506 {
 507         int timeout;
 508         struct create_idle c_idle = {
 509                 .work = __WORK_INITIALIZER(c_idle.work, do_fork_idle),
 510                 .cpu    = cpu,
 511                 .done   = COMPLETION_INITIALIZER(c_idle.done),
 512         };
 513
 514         c_idle.idle = get_idle_for_cpu(cpu);
 515         if (c_idle.idle) {
 516                 init_idle(c_idle.idle, cpu);
 517                 goto do_rest;
 518         }
 519
 520         /*
 521          * We can't use kernel_thread since we must avoid to reschedule the child.
 522          */
 523         if (!keventd_up() || current_is_keventd())
 524                 c_idle.work.func(&c_idle.work);
 525         else {
 526                 schedule_work(&c_idle.work);
 527                 wait_for_completion(&c_idle.done);
 528         }
 529
 530         if (IS_ERR(c_idle.idle))
 531                 panic("failed fork for CPU %d", cpu);
 532
 533         set_idle_for_cpu(cpu, c_idle.idle);
 534
 535 do_rest:
 536         task_for_booting_cpu = c_idle.idle;
 537
 538         Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
 539
 540         set_brendez_area(cpu);
 541         platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
 542
 543         /*
 544          * Wait 10s total for the AP to start
 545          */
 546         Dprintk("Waiting on callin_map ...");
 547         for (timeout = 0; timeout < 100000; timeout++) {
 548                 if (cpu_isset(cpu, cpu_callin_map))
 549                         break;  /* It has booted */
 550                 udelay(100);
 551         }
 552         Dprintk("\n");
 553
 554         if (!cpu_isset(cpu, cpu_callin_map)) {
 555                 printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
 556                 ia64_cpu_to_sapicid[cpu] = -1;
 557                 cpu_clear(cpu, cpu_online_map);  /* was set in smp_callin() */
 558                 return -EINVAL;
 559         }
 560         return 0;
 561 }
 562
 563 static int __init
 564 decay (char *str)
 565 {
 566         int ticks;
 567         get_option (&str, &ticks);
 568         return 1;
 569 }
 570
 571 __setup("decay=", decay);
 572
 573 /*
 574  * Initialize the logical CPU number to SAPICID mapping
 575  */
 576 void __init
 577 smp_build_cpu_map (void)
 578 {
 579         int sapicid, cpu, i;
 580         int boot_cpu_id = hard_smp_processor_id();
 581
 582         for (cpu = 0; cpu < NR_CPUS; cpu++) {
 583                 ia64_cpu_to_sapicid[cpu] = -1;
 584         }
 585
 586         ia64_cpu_to_sapicid[0] = boot_cpu_id;
 587         cpus_clear(cpu_present_map);
 588         cpu_set(0, cpu_present_map);
 589         cpu_set(0, cpu_possible_map);
 590         for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
 591                 sapicid = smp_boot_data.cpu_phys_id[i];
 592                 if (sapicid == boot_cpu_id)
 593                         continue;
 594                 cpu_set(cpu, cpu_present_map);
 595                 cpu_set(cpu, cpu_possible_map);
 596                 ia64_cpu_to_sapicid[cpu] = sapicid;
 597                 cpu++;
 598         }
 599 }
 600
 601 /*
 602  * Cycle through the APs sending Wakeup IPIs to boot each.
 603  */
 604 void __init
 605 smp_prepare_cpus (unsigned int max_cpus)
 606 {
 607         int boot_cpu_id = hard_smp_processor_id();
 608
 609         /*
 610          * Initialize the per-CPU profiling counter/multiplier
 611          */
 612
 613         smp_setup_percpu_timer();
 614
 615         /*
 616          * We have the boot CPU online for sure.
 617          */
 618         cpu_set(0, cpu_online_map);
 619         cpu_set(0, cpu_callin_map);
 620
 621         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
 622         ia64_cpu_to_sapicid[0] = boot_cpu_id;
 623
 624         printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
 625
 626         current_thread_info()->cpu = 0;
 627
 628         /*
 629          * If SMP should be disabled, then really disable it!
 630          */
 631         if (!max_cpus) {
 632                 printk(KERN_INFO "SMP mode deactivated.\n");
 633                 cpus_clear(cpu_online_map);
 634                 cpus_clear(cpu_present_map);
 635                 cpus_clear(cpu_possible_map);
 636                 cpu_set(0, cpu_online_map);
 637                 cpu_set(0, cpu_present_map);
 638                 cpu_set(0, cpu_possible_map);
 639                 return;
 640         }
 641 }
 642
 643 void __devinit smp_prepare_boot_cpu(void)
 644 {
 645         cpu_set(smp_processor_id(), cpu_online_map);
 646         cpu_set(smp_processor_id(), cpu_callin_map);
 647         per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
 648         paravirt_post_smp_prepare_boot_cpu();
 649 }
 650
 651 #ifdef CONFIG_HOTPLUG_CPU
 652 static inline void
 653 clear_cpu_sibling_map(int cpu)
 654 {
 655         int i;
 656
 657         for_each_cpu_mask(i, per_cpu(cpu_sibling_map, cpu))
 658                 cpu_clear(cpu, per_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         per_cpu(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, per_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
 683 extern void fixup_irqs(void);
 684
 685 int migrate_platform_irqs(unsigned int cpu)
 686 {
 687         int new_cpei_cpu;
 688         irq_desc_t *desc = NULL;
 689         cpumask_t       mask;
 690         int             retval = 0;
 691
 692         /*
 693          * dont permit CPEI target to removed.
 694          */
 695         if (cpe_vector > 0 && is_cpu_cpei_target(cpu)) {
 696                 printk ("CPU (%d) is CPEI Target\n", cpu);
 697                 if (can_cpei_retarget()) {
 698                         /*
 699                          * Now re-target the CPEI to a different processor
 700                          */
 701                         new_cpei_cpu = any_online_cpu(cpu_online_map);
 702                         mask = cpumask_of_cpu(new_cpei_cpu);
 703                         set_cpei_target_cpu(new_cpei_cpu);
 704                         desc = irq_desc + ia64_cpe_irq;
 705                         /*
 706                          * Switch for now, immediately, we need to do fake intr
 707                          * as other interrupts, but need to study CPEI behaviour with
 708                          * polling before making changes.
 709                          */
 710                         if (desc) {
 711                                 desc->chip->disable(ia64_cpe_irq);
 712                                 desc->chip->set_affinity(ia64_cpe_irq, mask);
 713                                 desc->chip->enable(ia64_cpe_irq);
 714                                 printk ("Re-targetting CPEI to cpu %d\n", new_cpei_cpu);
 715                         }
 716                 }
 717                 if (!desc) {
 718                         printk ("Unable to retarget CPEI, offline cpu [%d] failed\n", cpu);
 719                         retval = -EBUSY;
 720                 }
 721         }
 722         return retval;
 723 }
 724
 725 /* must be called with cpucontrol mutex held */
 726 int __cpu_disable(void)
 727 {
 728         int cpu = smp_processor_id();
 729
 730         /*
 731          * dont permit boot processor for now
 732          */
 733         if (cpu == 0 && !bsp_remove_ok) {
 734                 printk ("Your platform does not support removal of BSP\n");
 735                 return (-EBUSY);
 736         }
 737
 738         if (ia64_platform_is("sn2")) {
 739                 if (!sn_cpu_disable_allowed(cpu))
 740                         return -EBUSY;
 741         }
 742
 743         cpu_clear(cpu, cpu_online_map);
 744
 745         if (migrate_platform_irqs(cpu)) {
 746                 cpu_set(cpu, cpu_online_map);
 747                 return (-EBUSY);
 748         }
 749
 750         remove_siblinginfo(cpu);
 751         cpu_clear(cpu, cpu_online_map);
 752         fixup_irqs();
 753         local_flush_tlb_all();
 754         cpu_clear(cpu, cpu_callin_map);
 755         return 0;
 756 }
 757
 758 void __cpu_die(unsigned int cpu)
 759 {
 760         unsigned int i;
 761
 762         for (i = 0; i < 100; i++) {
 763                 /* They ack this in play_dead by setting CPU_DEAD */
 764                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
 765                 {
 766                         printk ("CPU %d is now offline\n", cpu);
 767                         return;
 768                 }
 769                 msleep(100);
 770         }
 771         printk(KERN_ERR "CPU %u didn't die...\n", cpu);
 772 }
 773 #endif /* CONFIG_HOTPLUG_CPU */
 774
 775 void
 776 smp_cpus_done (unsigned int dummy)
 777 {
 778         int cpu;
 779         unsigned long bogosum = 0;
 780
 781         /*
 782          * Allow the user to impress friends.
 783          */
 784
 785         for_each_online_cpu(cpu) {
 786                 bogosum += cpu_data(cpu)->loops_per_jiffy;
 787         }
 788
 789         printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
 790                (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
 791 }
 792
 793 static inline void __devinit
 794 set_cpu_sibling_map(int cpu)
 795 {
 796         int i;
 797
 798         for_each_online_cpu(i) {
 799                 if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
 800                         cpu_set(i, cpu_core_map[cpu]);
 801                         cpu_set(cpu, cpu_core_map[i]);
 802                         if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
 803                                 cpu_set(i, per_cpu(cpu_sibling_map, cpu));
 804                                 cpu_set(cpu, per_cpu(cpu_sibling_map, i));
 805                         }
 806                 }
 807         }
 808 }
 809
 810 int __cpuinit
 811 __cpu_up (unsigned int cpu)
 812 {
 813         int ret;
 814         int sapicid;
 815
 816         sapicid = ia64_cpu_to_sapicid[cpu];
 817         if (sapicid == -1)
 818                 return -EINVAL;
 819
 820         /*
 821          * Already booted cpu? not valid anymore since we dont
 822          * do idle loop tightspin anymore.
 823          */
 824         if (cpu_isset(cpu, cpu_callin_map))
 825                 return -EINVAL;
 826
 827         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 828         /* Processor goes to start_secondary(), sets online flag */
 829         ret = do_boot_cpu(sapicid, cpu);
 830         if (ret < 0)
 831                 return ret;
 832
 833         if (cpu_data(cpu)->threads_per_core == 1 &&
 834             cpu_data(cpu)->cores_per_socket == 1) {
 835                 cpu_set(cpu, per_cpu(cpu_sibling_map, cpu));
 836                 cpu_set(cpu, cpu_core_map[cpu]);
 837                 return 0;
 838         }
 839
 840         set_cpu_sibling_map(cpu);
 841
 842         return 0;
 843 }
 844
 845 /*
 846  * Assume that CPUs have been discovered by some platform-dependent interface.  For
 847  * SoftSDV/Lion, that would be ACPI.
 848  *
 849  * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
 850  */
 851 void __init
 852 init_smp_config(void)
 853 {
 854         struct fptr {
 855                 unsigned long fp;
 856                 unsigned long gp;
 857         } *ap_startup;
 858         long sal_ret;
 859
 860         /* Tell SAL where to drop the APs.  */
 861         ap_startup = (struct fptr *) start_ap;
 862         sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
 863                                        ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
 864         if (sal_ret < 0)
 865                 printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
 866                        ia64_sal_strerror(sal_ret));
 867 }
 868
 869 /*
 870  * identify_siblings(cpu) gets called from identify_cpu. This populates the
 871  * information related to logical execution units in per_cpu_data structure.
 872  */
 873 void __devinit
 874 identify_siblings(struct cpuinfo_ia64 *c)
 875 {
 876         s64 status;
 877         u16 pltid;
 878         pal_logical_to_physical_t info;
 879
 880         status = ia64_pal_logical_to_phys(-1, &info);
 881         if (status != PAL_STATUS_SUCCESS) {
 882                 if (status != PAL_STATUS_UNIMPLEMENTED) {
 883                         printk(KERN_ERR
 884                                 "ia64_pal_logical_to_phys failed with %ld\n",
 885                                 status);
 886                         return;
 887                 }
 888
 889                 info.overview_ppid = 0;
 890                 info.overview_cpp  = 1;
 891                 info.overview_tpc  = 1;
 892         }
 893
 894         status = ia64_sal_physical_id_info(&pltid);
 895         if (status != PAL_STATUS_SUCCESS) {
 896                 if (status != PAL_STATUS_UNIMPLEMENTED)
 897                         printk(KERN_ERR
 898                                 "ia64_sal_pltid failed with %ld\n",
 899                                 status);
 900                 return;
 901         }
 902
 903         c->socket_id =  (pltid << 8) | info.overview_ppid;
 904
 905         if (info.overview_cpp == 1 && info.overview_tpc == 1)
 906                 return;
 907
 908         c->cores_per_socket = info.overview_cpp;
 909         c->threads_per_core = info.overview_tpc;
 910         c->num_log = info.overview_num_log;
 911
 912         c->core_id = info.log1_cid;
 913         c->thread_id = info.log1_tid;
 914 }
 915
 916 /*
 917  * returns non zero, if multi-threading is enabled
 918  * on at least one physical package. Due to hotplug cpu
 919  * and (maxcpus=), all threads may not necessarily be enabled
 920  * even though the processor supports multi-threading.
 921  */
 922 int is_multithreading_enabled(void)
 923 {
 924         int i, j;
 925
 926         for_each_present_cpu(i) {
 927                 for_each_present_cpu(j) {
 928                         if (j == i)
 929                                 continue;
 930                         if ((cpu_data(j)->socket_id == cpu_data(i)->socket_id)) {
 931                                 if (cpu_data(j)->core_id == cpu_data(i)->core_id)
 932                                         return 1;
 933                         }
 934                 }
 935         }
 936         return 0;
 937 }
 938 EXPORT_SYMBOL_GPL(is_multithreading_enabled);