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/smp_lock.h>
  39 #include <linux/spinlock.h>
  40 #include <linux/efi.h>
  41 #include <linux/percpu.h>
  42 #include <linux/bitops.h>
  43
  44 #include <asm/atomic.h>
  45 #include <asm/cache.h>
  46 #include <asm/current.h>
  47 #include <asm/delay.h>
  48 #include <asm/ia32.h>
  49 #include <asm/io.h>
  50 #include <asm/irq.h>
  51 #include <asm/machvec.h>
  52 #include <asm/mca.h>
  53 #include <asm/page.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
  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 __devinit calibrate_delay (void);
 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 cpumask_t cpu_sibling_map[NR_CPUS] __cacheline_aligned;
 142 int smp_num_siblings = 1;
 143 int smp_num_cpucores = 1;
 144
 145 /* which logical CPU number maps to which CPU (physical APIC ID) */
 146 volatile int ia64_cpu_to_sapicid[NR_CPUS];
 147 EXPORT_SYMBOL(ia64_cpu_to_sapicid);
 148
 149 static volatile cpumask_t cpu_callin_map;
 150
 151 struct smp_boot_data smp_boot_data __initdata;
 152
 153 unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
 154
 155 char __initdata no_int_routing;
 156
 157 unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
 158
 159 #ifdef CONFIG_FORCE_CPEI_RETARGET
 160 #define CPEI_OVERRIDE_DEFAULT   (1)
 161 #else
 162 #define CPEI_OVERRIDE_DEFAULT   (0)
 163 #endif
 164
 165 unsigned int force_cpei_retarget = CPEI_OVERRIDE_DEFAULT;
 166
 167 static int __init
 168 cmdl_force_cpei(char *str)
 169 {
 170         int value=0;
 171
 172         get_option (&str, &value);
 173         force_cpei_retarget = value;
 174
 175         return 1;
 176 }
 177
 178 __setup("force_cpei=", cmdl_force_cpei);
 179
 180 static int __init
 181 nointroute (char *str)
 182 {
 183         no_int_routing = 1;
 184         printk ("no_int_routing on\n");
 185         return 1;
 186 }
 187
 188 __setup("nointroute", nointroute);
 189
 190 static void fix_b0_for_bsp(void)
 191 {
 192 #ifdef CONFIG_HOTPLUG_CPU
 193         int cpuid;
 194         static int fix_bsp_b0 = 1;
 195
 196         cpuid = smp_processor_id();
 197
 198         /*
 199          * Cache the b0 value on the first AP that comes up
 200          */
 201         if (!(fix_bsp_b0 && cpuid))
 202                 return;
 203
 204         sal_boot_rendez_state[0].br[0] = sal_boot_rendez_state[cpuid].br[0];
 205         printk ("Fixed BSP b0 value from CPU %d\n", cpuid);
 206
 207         fix_bsp_b0 = 0;
 208 #endif
 209 }
 210
 211 void
 212 sync_master (void *arg)
 213 {
 214         unsigned long flags, i;
 215
 216         go[MASTER] = 0;
 217
 218         local_irq_save(flags);
 219         {
 220                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
 221                         while (!go[MASTER])
 222                                 cpu_relax();
 223                         go[MASTER] = 0;
 224                         go[SLAVE] = ia64_get_itc();
 225                 }
 226         }
 227         local_irq_restore(flags);
 228 }
 229
 230 /*
 231  * Return the number of cycles by which our itc differs from the itc on the master
 232  * (time-keeper) CPU.  A positive number indicates our itc is ahead of the master,
 233  * negative that it is behind.
 234  */
 235 static inline long
 236 get_delta (long *rt, long *master)
 237 {
 238         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
 239         unsigned long tcenter, t0, t1, tm;
 240         long i;
 241
 242         for (i = 0; i < NUM_ITERS; ++i) {
 243                 t0 = ia64_get_itc();
 244                 go[MASTER] = 1;
 245                 while (!(tm = go[SLAVE]))
 246                         cpu_relax();
 247                 go[SLAVE] = 0;
 248                 t1 = ia64_get_itc();
 249
 250                 if (t1 - t0 < best_t1 - best_t0)
 251                         best_t0 = t0, best_t1 = t1, best_tm = tm;
 252         }
 253
 254         *rt = best_t1 - best_t0;
 255         *master = best_tm - best_t0;
 256
 257         /* average best_t0 and best_t1 without overflow: */
 258         tcenter = (best_t0/2 + best_t1/2);
 259         if (best_t0 % 2 + best_t1 % 2 == 2)
 260                 ++tcenter;
 261         return tcenter - best_tm;
 262 }
 263
 264 /*
 265  * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
 266  * (normally the time-keeper CPU).  We use a closed loop to eliminate the possibility of
 267  * unaccounted-for errors (such as getting a machine check in the middle of a calibration
 268  * step).  The basic idea is for the slave to ask the master what itc value it has and to
 269  * read its own itc before and after the master responds.  Each iteration gives us three
 270  * timestamps:
 271  *
 272  *      slave           master
 273  *
 274  *      t0 ---\
 275  *             ---\
 276  *                 --->
 277  *                      tm
 278  *                 /---
 279  *             /---
 280  *      t1 <---
 281  *
 282  *
 283  * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
 284  * and t1.  If we achieve this, the clocks are synchronized provided the interconnect
 285  * between the slave and the master is symmetric.  Even if the interconnect were
 286  * asymmetric, we would still know that the synchronization error is smaller than the
 287  * roundtrip latency (t0 - t1).
 288  *
 289  * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
 290  * within one or two cycles.  However, we can only *guarantee* that the synchronization is
 291  * accurate to within a round-trip time, which is typically in the range of several
 292  * hundred cycles (e.g., ~500 cycles).  In practice, this means that the itc's are usually
 293  * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
 294  * than half a micro second or so.
 295  */
 296 void
 297 ia64_sync_itc (unsigned int master)
 298 {
 299         long i, delta, adj, adjust_latency = 0, done = 0;
 300         unsigned long flags, rt, master_time_stamp, bound;
 301 #if DEBUG_ITC_SYNC
 302         struct {
 303                 long rt;        /* roundtrip time */
 304                 long master;    /* master's timestamp */
 305                 long diff;      /* difference between midpoint and master's timestamp */
 306                 long lat;       /* estimate of itc adjustment latency */
 307         } t[NUM_ROUNDS];
 308 #endif
 309
 310         /*
 311          * Make sure local timer ticks are disabled while we sync.  If
 312          * they were enabled, we'd have to worry about nasty issues
 313          * like setting the ITC ahead of (or a long time before) the
 314          * next scheduled tick.
 315          */
 316         BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
 317
 318         go[MASTER] = 1;
 319
 320         if (smp_call_function_single(master, sync_master, NULL, 1, 0) < 0) {
 321                 printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
 322                 return;
 323         }
 324
 325         while (go[MASTER])
 326                 cpu_relax();    /* wait for master to be ready */
 327
 328         spin_lock_irqsave(&itc_sync_lock, flags);
 329         {
 330                 for (i = 0; i < NUM_ROUNDS; ++i) {
 331                         delta = get_delta(&rt, &master_time_stamp);
 332                         if (delta == 0) {
 333                                 done = 1;       /* let's lock on to this... */
 334                                 bound = rt;
 335                         }
 336
 337                         if (!done) {
 338                                 if (i > 0) {
 339                                         adjust_latency += -delta;
 340                                         adj = -delta + adjust_latency/4;
 341                                 } else
 342                                         adj = -delta;
 343
 344                                 ia64_set_itc(ia64_get_itc() + adj);
 345                         }
 346 #if DEBUG_ITC_SYNC
 347                         t[i].rt = rt;
 348                         t[i].master = master_time_stamp;
 349                         t[i].diff = delta;
 350                         t[i].lat = adjust_latency/4;
 351 #endif
 352                 }
 353         }
 354         spin_unlock_irqrestore(&itc_sync_lock, flags);
 355
 356 #if DEBUG_ITC_SYNC
 357         for (i = 0; i < NUM_ROUNDS; ++i)
 358                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
 359                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
 360 #endif
 361
 362         printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
 363                "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
 364 }
 365
 366 /*
 367  * Ideally sets up per-cpu profiling hooks.  Doesn't do much now...
 368  */
 369 static inline void __devinit
 370 smp_setup_percpu_timer (void)
 371 {
 372 }
 373
 374 static void __devinit
 375 smp_callin (void)
 376 {
 377         int cpuid, phys_id, itc_master;
 378         extern void ia64_init_itm(void);
 379         extern volatile int time_keeper_id;
 380
 381 #ifdef CONFIG_PERFMON
 382         extern void pfm_init_percpu(void);
 383 #endif
 384
 385         cpuid = smp_processor_id();
 386         phys_id = hard_smp_processor_id();
 387         itc_master = time_keeper_id;
 388
 389         if (cpu_online(cpuid)) {
 390                 printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
 391                        phys_id, cpuid);
 392                 BUG();
 393         }
 394
 395         fix_b0_for_bsp();
 396
 397         lock_ipi_calllock();
 398         cpu_set(cpuid, cpu_online_map);
 399         unlock_ipi_calllock();
 400         per_cpu(cpu_state, cpuid) = CPU_ONLINE;
 401
 402         smp_setup_percpu_timer();
 403
 404         ia64_mca_cmc_vector_setup();    /* Setup vector on AP */
 405
 406 #ifdef CONFIG_PERFMON
 407         pfm_init_percpu();
 408 #endif
 409
 410         local_irq_enable();
 411
 412         if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
 413                 /*
 414                  * Synchronize the ITC with the BP.  Need to do this after irqs are
 415                  * enabled because ia64_sync_itc() calls smp_call_function_single(), which
 416                  * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
 417                  * local_bh_enable(), which bugs out if irqs are not enabled...
 418                  */
 419                 Dprintk("Going to syncup ITC with ITC Master.\n");
 420                 ia64_sync_itc(itc_master);
 421         }
 422
 423         /*
 424          * Get our bogomips.
 425          */
 426         ia64_init_itm();
 427         calibrate_delay();
 428         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
 429
 430 #ifdef CONFIG_IA32_SUPPORT
 431         ia32_gdt_init();
 432 #endif
 433
 434         /*
 435          * Allow the master to continue.
 436          */
 437         cpu_set(cpuid, cpu_callin_map);
 438         Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
 439 }
 440
 441
 442 /*
 443  * Activate a secondary processor.  head.S calls this.
 444  */
 445 int __devinit
 446 start_secondary (void *unused)
 447 {
 448         /* Early console may use I/O ports */
 449         ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
 450         Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
 451         efi_map_pal_code();
 452         cpu_init();
 453         preempt_disable();
 454         smp_callin();
 455
 456         cpu_idle();
 457         return 0;
 458 }
 459
 460 struct pt_regs * __devinit idle_regs(struct pt_regs *regs)
 461 {
 462         return NULL;
 463 }
 464
 465 struct create_idle {
 466         struct work_struct work;
 467         struct task_struct *idle;
 468         struct completion done;
 469         int cpu;
 470 };
 471
 472 void
 473 do_fork_idle(struct work_struct *work)
 474 {
 475         struct create_idle *c_idle =
 476                 container_of(work, struct create_idle, work);
 477
 478         c_idle->idle = fork_idle(c_idle->cpu);
 479         complete(&c_idle->done);
 480 }
 481
 482 static int __devinit
 483 do_boot_cpu (int sapicid, int cpu)
 484 {
 485         int timeout;
 486         struct create_idle c_idle = {
 487                 .work = __WORK_INITIALIZER(c_idle.work, do_fork_idle),
 488                 .cpu    = cpu,
 489                 .done   = COMPLETION_INITIALIZER(c_idle.done),
 490         };
 491
 492         c_idle.idle = get_idle_for_cpu(cpu);
 493         if (c_idle.idle) {
 494                 init_idle(c_idle.idle, cpu);
 495                 goto do_rest;
 496         }
 497
 498         /*
 499          * We can't use kernel_thread since we must avoid to reschedule the child.
 500          */
 501         if (!keventd_up() || current_is_keventd())
 502                 c_idle.work.func(&c_idle.work);
 503         else {
 504                 schedule_work(&c_idle.work);
 505                 wait_for_completion(&c_idle.done);
 506         }
 507
 508         if (IS_ERR(c_idle.idle))
 509                 panic("failed fork for CPU %d", cpu);
 510
 511         set_idle_for_cpu(cpu, c_idle.idle);
 512
 513 do_rest:
 514         task_for_booting_cpu = c_idle.idle;
 515
 516         Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
 517
 518         set_brendez_area(cpu);
 519         platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
 520
 521         /*
 522          * Wait 10s total for the AP to start
 523          */
 524         Dprintk("Waiting on callin_map ...");
 525         for (timeout = 0; timeout < 100000; timeout++) {
 526                 if (cpu_isset(cpu, cpu_callin_map))
 527                         break;  /* It has booted */
 528                 udelay(100);
 529         }
 530         Dprintk("\n");
 531
 532         if (!cpu_isset(cpu, cpu_callin_map)) {
 533                 printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
 534                 ia64_cpu_to_sapicid[cpu] = -1;
 535                 cpu_clear(cpu, cpu_online_map);  /* was set in smp_callin() */
 536                 return -EINVAL;
 537         }
 538         return 0;
 539 }
 540
 541 static int __init
 542 decay (char *str)
 543 {
 544         int ticks;
 545         get_option (&str, &ticks);
 546         return 1;
 547 }
 548
 549 __setup("decay=", decay);
 550
 551 /*
 552  * Initialize the logical CPU number to SAPICID mapping
 553  */
 554 void __init
 555 smp_build_cpu_map (void)
 556 {
 557         int sapicid, cpu, i;
 558         int boot_cpu_id = hard_smp_processor_id();
 559
 560         for (cpu = 0; cpu < NR_CPUS; cpu++) {
 561                 ia64_cpu_to_sapicid[cpu] = -1;
 562         }
 563
 564         ia64_cpu_to_sapicid[0] = boot_cpu_id;
 565         cpus_clear(cpu_present_map);
 566         cpu_set(0, cpu_present_map);
 567         cpu_set(0, cpu_possible_map);
 568         for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
 569                 sapicid = smp_boot_data.cpu_phys_id[i];
 570                 if (sapicid == boot_cpu_id)
 571                         continue;
 572                 cpu_set(cpu, cpu_present_map);
 573                 cpu_set(cpu, cpu_possible_map);
 574                 ia64_cpu_to_sapicid[cpu] = sapicid;
 575                 cpu++;
 576         }
 577 }
 578
 579 /*
 580  * Cycle through the APs sending Wakeup IPIs to boot each.
 581  */
 582 void __init
 583 smp_prepare_cpus (unsigned int max_cpus)
 584 {
 585         int boot_cpu_id = hard_smp_processor_id();
 586
 587         /*
 588          * Initialize the per-CPU profiling counter/multiplier
 589          */
 590
 591         smp_setup_percpu_timer();
 592
 593         /*
 594          * We have the boot CPU online for sure.
 595          */
 596         cpu_set(0, cpu_online_map);
 597         cpu_set(0, cpu_callin_map);
 598
 599         local_cpu_data->loops_per_jiffy = loops_per_jiffy;
 600         ia64_cpu_to_sapicid[0] = boot_cpu_id;
 601
 602         printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
 603
 604         current_thread_info()->cpu = 0;
 605
 606         /*
 607          * If SMP should be disabled, then really disable it!
 608          */
 609         if (!max_cpus) {
 610                 printk(KERN_INFO "SMP mode deactivated.\n");
 611                 cpus_clear(cpu_online_map);
 612                 cpus_clear(cpu_present_map);
 613                 cpus_clear(cpu_possible_map);
 614                 cpu_set(0, cpu_online_map);
 615                 cpu_set(0, cpu_present_map);
 616                 cpu_set(0, cpu_possible_map);
 617                 return;
 618         }
 619 }
 620
 621 void __devinit smp_prepare_boot_cpu(void)
 622 {
 623         cpu_set(smp_processor_id(), cpu_online_map);
 624         cpu_set(smp_processor_id(), cpu_callin_map);
 625         per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
 626 }
 627
 628 #ifdef CONFIG_HOTPLUG_CPU
 629 static inline void
 630 clear_cpu_sibling_map(int cpu)
 631 {
 632         int i;
 633
 634         for_each_cpu_mask(i, cpu_sibling_map[cpu])
 635                 cpu_clear(cpu, cpu_sibling_map[i]);
 636         for_each_cpu_mask(i, cpu_core_map[cpu])
 637                 cpu_clear(cpu, cpu_core_map[i]);
 638
 639         cpu_sibling_map[cpu] = cpu_core_map[cpu] = CPU_MASK_NONE;
 640 }
 641
 642 static void
 643 remove_siblinginfo(int cpu)
 644 {
 645         int last = 0;
 646
 647         if (cpu_data(cpu)->threads_per_core == 1 &&
 648             cpu_data(cpu)->cores_per_socket == 1) {
 649                 cpu_clear(cpu, cpu_core_map[cpu]);
 650                 cpu_clear(cpu, cpu_sibling_map[cpu]);
 651                 return;
 652         }
 653
 654         last = (cpus_weight(cpu_core_map[cpu]) == 1 ? 1 : 0);
 655
 656         /* remove it from all sibling map's */
 657         clear_cpu_sibling_map(cpu);
 658 }
 659
 660 extern void fixup_irqs(void);
 661
 662 int migrate_platform_irqs(unsigned int cpu)
 663 {
 664         int new_cpei_cpu;
 665         irq_desc_t *desc = NULL;
 666         cpumask_t       mask;
 667         int             retval = 0;
 668
 669         /*
 670          * dont permit CPEI target to removed.
 671          */
 672         if (cpe_vector > 0 && is_cpu_cpei_target(cpu)) {
 673                 printk ("CPU (%d) is CPEI Target\n", cpu);
 674                 if (can_cpei_retarget()) {
 675                         /*
 676                          * Now re-target the CPEI to a different processor
 677                          */
 678                         new_cpei_cpu = any_online_cpu(cpu_online_map);
 679                         mask = cpumask_of_cpu(new_cpei_cpu);
 680                         set_cpei_target_cpu(new_cpei_cpu);
 681                         desc = irq_desc + ia64_cpe_irq;
 682                         /*
 683                          * Switch for now, immediatly, we need to do fake intr
 684                          * as other interrupts, but need to study CPEI behaviour with
 685                          * polling before making changes.
 686                          */
 687                         if (desc) {
 688                                 desc->chip->disable(ia64_cpe_irq);
 689                                 desc->chip->set_affinity(ia64_cpe_irq, mask);
 690                                 desc->chip->enable(ia64_cpe_irq);
 691                                 printk ("Re-targetting CPEI to cpu %d\n", new_cpei_cpu);
 692                         }
 693                 }
 694                 if (!desc) {
 695                         printk ("Unable to retarget CPEI, offline cpu [%d] failed\n", cpu);
 696                         retval = -EBUSY;
 697                 }
 698         }
 699         return retval;
 700 }
 701
 702 /* must be called with cpucontrol mutex held */
 703 int __cpu_disable(void)
 704 {
 705         int cpu = smp_processor_id();
 706
 707         /*
 708          * dont permit boot processor for now
 709          */
 710         if (cpu == 0 && !bsp_remove_ok) {
 711                 printk ("Your platform does not support removal of BSP\n");
 712                 return (-EBUSY);
 713         }
 714
 715         cpu_clear(cpu, cpu_online_map);
 716
 717         if (migrate_platform_irqs(cpu)) {
 718                 cpu_set(cpu, cpu_online_map);
 719                 return (-EBUSY);
 720         }
 721
 722         remove_siblinginfo(cpu);
 723         cpu_clear(cpu, cpu_online_map);
 724         fixup_irqs();
 725         local_flush_tlb_all();
 726         cpu_clear(cpu, cpu_callin_map);
 727         return 0;
 728 }
 729
 730 void __cpu_die(unsigned int cpu)
 731 {
 732         unsigned int i;
 733
 734         for (i = 0; i < 100; i++) {
 735                 /* They ack this in play_dead by setting CPU_DEAD */
 736                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
 737                 {
 738                         printk ("CPU %d is now offline\n", cpu);
 739                         return;
 740                 }
 741                 msleep(100);
 742         }
 743         printk(KERN_ERR "CPU %u didn't die...\n", cpu);
 744 }
 745 #else /* !CONFIG_HOTPLUG_CPU */
 746 int __cpu_disable(void)
 747 {
 748         return -ENOSYS;
 749 }
 750
 751 void __cpu_die(unsigned int cpu)
 752 {
 753         /* We said "no" in __cpu_disable */
 754         BUG();
 755 }
 756 #endif /* CONFIG_HOTPLUG_CPU */
 757
 758 void
 759 smp_cpus_done (unsigned int dummy)
 760 {
 761         int cpu;
 762         unsigned long bogosum = 0;
 763
 764         /*
 765          * Allow the user to impress friends.
 766          */
 767
 768         for_each_online_cpu(cpu) {
 769                 bogosum += cpu_data(cpu)->loops_per_jiffy;
 770         }
 771
 772         printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
 773                (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
 774 }
 775
 776 static inline void __devinit
 777 set_cpu_sibling_map(int cpu)
 778 {
 779         int i;
 780
 781         for_each_online_cpu(i) {
 782                 if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
 783                         cpu_set(i, cpu_core_map[cpu]);
 784                         cpu_set(cpu, cpu_core_map[i]);
 785                         if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
 786                                 cpu_set(i, cpu_sibling_map[cpu]);
 787                                 cpu_set(cpu, cpu_sibling_map[i]);
 788                         }
 789                 }
 790         }
 791 }
 792
 793 int __devinit
 794 __cpu_up (unsigned int cpu)
 795 {
 796         int ret;
 797         int sapicid;
 798
 799         sapicid = ia64_cpu_to_sapicid[cpu];
 800         if (sapicid == -1)
 801                 return -EINVAL;
 802
 803         /*
 804          * Already booted cpu? not valid anymore since we dont
 805          * do idle loop tightspin anymore.
 806          */
 807         if (cpu_isset(cpu, cpu_callin_map))
 808                 return -EINVAL;
 809
 810         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 811         /* Processor goes to start_secondary(), sets online flag */
 812         ret = do_boot_cpu(sapicid, cpu);
 813         if (ret < 0)
 814                 return ret;
 815
 816         if (cpu_data(cpu)->threads_per_core == 1 &&
 817             cpu_data(cpu)->cores_per_socket == 1) {
 818                 cpu_set(cpu, cpu_sibling_map[cpu]);
 819                 cpu_set(cpu, cpu_core_map[cpu]);
 820                 return 0;
 821         }
 822
 823         set_cpu_sibling_map(cpu);
 824
 825         return 0;
 826 }
 827
 828 /*
 829  * Assume that CPU's have been discovered by some platform-dependent interface.  For
 830  * SoftSDV/Lion, that would be ACPI.
 831  *
 832  * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
 833  */
 834 void __init
 835 init_smp_config(void)
 836 {
 837         struct fptr {
 838                 unsigned long fp;
 839                 unsigned long gp;
 840         } *ap_startup;
 841         long sal_ret;
 842
 843         /* Tell SAL where to drop the AP's.  */
 844         ap_startup = (struct fptr *) start_ap;
 845         sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
 846                                        ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
 847         if (sal_ret < 0)
 848                 printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
 849                        ia64_sal_strerror(sal_ret));
 850 }
 851
 852 /*
 853  * identify_siblings(cpu) gets called from identify_cpu. This populates the
 854  * information related to logical execution units in per_cpu_data structure.
 855  */
 856 void __devinit
 857 identify_siblings(struct cpuinfo_ia64 *c)
 858 {
 859         s64 status;
 860         u16 pltid;
 861         pal_logical_to_physical_t info;
 862
 863         if (smp_num_cpucores == 1 && smp_num_siblings == 1)
 864                 return;
 865
 866         if ((status = ia64_pal_logical_to_phys(-1, &info)) != PAL_STATUS_SUCCESS) {
 867                 printk(KERN_ERR "ia64_pal_logical_to_phys failed with %ld\n",
 868                        status);
 869                 return;
 870         }
 871         if ((status = ia64_sal_physical_id_info(&pltid)) != PAL_STATUS_SUCCESS) {
 872                 printk(KERN_ERR "ia64_sal_pltid failed with %ld\n", status);
 873                 return;
 874         }
 875
 876         c->socket_id =  (pltid << 8) | info.overview_ppid;
 877         c->cores_per_socket = info.overview_cpp;
 878         c->threads_per_core = info.overview_tpc;
 879         c->num_log = info.overview_num_log;
 880
 881         c->core_id = info.log1_cid;
 882         c->thread_id = info.log1_tid;
 883 }
 884
 885 /*
 886  * returns non zero, if multi-threading is enabled
 887  * on at least one physical package. Due to hotplug cpu
 888  * and (maxcpus=), all threads may not necessarily be enabled
 889  * even though the processor supports multi-threading.
 890  */
 891 int is_multithreading_enabled(void)
 892 {
 893         int i, j;
 894
 895         for_each_present_cpu(i) {
 896                 for_each_present_cpu(j) {
 897                         if (j == i)
 898                                 continue;
 899                         if ((cpu_data(j)->socket_id == cpu_data(i)->socket_id)) {
 900                                 if (cpu_data(j)->core_id == cpu_data(i)->core_id)
 901                                         return 1;
 902                         }
 903                 }
 904         }
 905         return 0;
 906 }
 907 EXPORT_SYMBOL_GPL(is_multithreading_enabled);