arch/arm/kernel/smp.c

   1 /*
   2  *  linux/arch/arm/kernel/smp.c
   3  *
   4  *  Copyright (C) 2002 ARM Limited, All Rights Reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 as
   8  * published by the Free Software Foundation.
   9  */
  10 #include <linux/module.h>
  11 #include <linux/delay.h>
  12 #include <linux/init.h>
  13 #include <linux/spinlock.h>
  14 #include <linux/sched.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/cache.h>
  17 #include <linux/profile.h>
  18 #include <linux/errno.h>
  19 #include <linux/mm.h>
  20 #include <linux/err.h>
  21 #include <linux/cpu.h>
  22 #include <linux/smp.h>
  23 #include <linux/seq_file.h>
  24 #include <linux/irq.h>
  25 #include <linux/percpu.h>
  26 #include <linux/clockchips.h>
  27 #include <linux/completion.h>
  28 #include <linux/cpufreq.h>
  29
  30 #include <linux/atomic.h>
  31 #include <asm/cacheflush.h>
  32 #include <asm/cpu.h>
  33 #include <asm/cputype.h>
  34 #include <asm/exception.h>
  35 #include <asm/idmap.h>
  36 #include <asm/topology.h>
  37 #include <asm/mmu_context.h>
  38 #include <asm/pgtable.h>
  39 #include <asm/pgalloc.h>
  40 #include <asm/processor.h>
  41 #include <asm/sections.h>
  42 #include <asm/tlbflush.h>
  43 #include <asm/ptrace.h>
  44 #include <asm/localtimer.h>
  45 #include <asm/smp_plat.h>
  46
  47 /*
  48  * as from 2.5, kernels no longer have an init_tasks structure
  49  * so we need some other way of telling a new secondary core
  50  * where to place its SVC stack
  51  */
  52 struct secondary_data secondary_data;
  53
  54 enum ipi_msg_type {
  55         IPI_TIMER = 2,
  56         IPI_RESCHEDULE,
  57         IPI_CALL_FUNC,
  58         IPI_CALL_FUNC_SINGLE,
  59         IPI_CPU_STOP,
  60 };
  61
  62 static DECLARE_COMPLETION(cpu_running);
  63
  64 int __cpuinit __cpu_up(unsigned int cpu, struct task_struct *idle)
  65 {
  66         int ret;
  67
  68         /*
  69          * We need to tell the secondary core where to find
  70          * its stack and the page tables.
  71          */
  72         secondary_data.stack = task_stack_page(idle) + THREAD_START_SP;
  73         secondary_data.pgdir = virt_to_phys(idmap_pgd);
  74         secondary_data.swapper_pg_dir = virt_to_phys(swapper_pg_dir);
  75         __cpuc_flush_dcache_area(&secondary_data, sizeof(secondary_data));
  76         outer_clean_range(__pa(&secondary_data), __pa(&secondary_data + 1));
  77
  78         /*
  79          * Now bring the CPU into our world.
  80          */
  81         ret = boot_secondary(cpu, idle);
  82         if (ret == 0) {
  83                 /*
  84                  * CPU was successfully started, wait for it
  85                  * to come online or time out.
  86                  */
  87                 wait_for_completion_timeout(&cpu_running,
  88                                                  msecs_to_jiffies(1000));
  89
  90                 if (!cpu_online(cpu)) {
  91                         pr_crit("CPU%u: failed to come online\n", cpu);
  92                         ret = -EIO;
  93                 }
  94         } else {
  95                 pr_err("CPU%u: failed to boot: %d\n", cpu, ret);
  96         }
  97
  98         secondary_data.stack = NULL;
  99         secondary_data.pgdir = 0;
 100
 101         return ret;
 102 }
 103
 104 #ifdef CONFIG_HOTPLUG_CPU
 105 static void percpu_timer_stop(void);
 106
 107 /*
 108  * __cpu_disable runs on the processor to be shutdown.
 109  */
 110 int __cpu_disable(void)
 111 {
 112         unsigned int cpu = smp_processor_id();
 113         int ret;
 114
 115         ret = platform_cpu_disable(cpu);
 116         if (ret)
 117                 return ret;
 118
 119         /*
 120          * Take this CPU offline.  Once we clear this, we can't return,
 121          * and we must not schedule until we're ready to give up the cpu.
 122          */
 123         set_cpu_online(cpu, false);
 124
 125         /*
 126          * OK - migrate IRQs away from this CPU
 127          */
 128         migrate_irqs();
 129
 130         /*
 131          * Stop the local timer for this CPU.
 132          */
 133         percpu_timer_stop();
 134
 135         /*
 136          * Flush user cache and TLB mappings, and then remove this CPU
 137          * from the vm mask set of all processes.
 138          */
 139         flush_cache_all();
 140         local_flush_tlb_all();
 141
 142         clear_tasks_mm_cpumask(cpu);
 143
 144         return 0;
 145 }
 146
 147 static DECLARE_COMPLETION(cpu_died);
 148
 149 /*
 150  * called on the thread which is asking for a CPU to be shutdown -
 151  * waits until shutdown has completed, or it is timed out.
 152  */
 153 void __cpu_die(unsigned int cpu)
 154 {
 155         if (!wait_for_completion_timeout(&cpu_died, msecs_to_jiffies(5000))) {
 156                 pr_err("CPU%u: cpu didn't die\n", cpu);
 157                 return;
 158         }
 159         printk(KERN_NOTICE "CPU%u: shutdown\n", cpu);
 160
 161         if (!platform_cpu_kill(cpu))
 162                 printk("CPU%u: unable to kill\n", cpu);
 163 }
 164
 165 /*
 166  * Called from the idle thread for the CPU which has been shutdown.
 167  *
 168  * Note that we disable IRQs here, but do not re-enable them
 169  * before returning to the caller. This is also the behaviour
 170  * of the other hotplug-cpu capable cores, so presumably coming
 171  * out of idle fixes this.
 172  */
 173 void __ref cpu_die(void)
 174 {
 175         unsigned int cpu = smp_processor_id();
 176
 177         idle_task_exit();
 178
 179         local_irq_disable();
 180         mb();
 181
 182         /* Tell __cpu_die() that this CPU is now safe to dispose of */
 183         RCU_NONIDLE(complete(&cpu_died));
 184
 185         /*
 186          * actual CPU shutdown procedure is at least platform (if not
 187          * CPU) specific.
 188          */
 189         platform_cpu_die(cpu);
 190
 191         /*
 192          * Do not return to the idle loop - jump back to the secondary
 193          * cpu initialisation.  There's some initialisation which needs
 194          * to be repeated to undo the effects of taking the CPU offline.
 195          */
 196         __asm__("mov    sp, %0\n"
 197         "       mov     fp, #0\n"
 198         "       b       secondary_start_kernel"
 199                 :
 200                 : "r" (task_stack_page(current) + THREAD_SIZE - 8));
 201 }
 202 #endif /* CONFIG_HOTPLUG_CPU */
 203
 204 /*
 205  * Called by both boot and secondaries to move global data into
 206  * per-processor storage.
 207  */
 208 static void __cpuinit smp_store_cpu_info(unsigned int cpuid)
 209 {
 210         struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid);
 211
 212         cpu_info->loops_per_jiffy = loops_per_jiffy;
 213
 214         store_cpu_topology(cpuid);
 215 }
 216
 217 static void percpu_timer_setup(void);
 218
 219 /*
 220  * This is the secondary CPU boot entry.  We're using this CPUs
 221  * idle thread stack, but a set of temporary page tables.
 222  */
 223 asmlinkage void __cpuinit secondary_start_kernel(void)
 224 {
 225         struct mm_struct *mm = &init_mm;
 226         unsigned int cpu = smp_processor_id();
 227
 228         /*
 229          * All kernel threads share the same mm context; grab a
 230          * reference and switch to it.
 231          */
 232         atomic_inc(&mm->mm_count);
 233         current->active_mm = mm;
 234         cpumask_set_cpu(cpu, mm_cpumask(mm));
 235         cpu_switch_mm(mm->pgd, mm);
 236         enter_lazy_tlb(mm, current);
 237         local_flush_tlb_all();
 238
 239         printk("CPU%u: Booted secondary processor\n", cpu);
 240
 241         cpu_init();
 242         preempt_disable();
 243         trace_hardirqs_off();
 244
 245         /*
 246          * Give the platform a chance to do its own initialisation.
 247          */
 248         platform_secondary_init(cpu);
 249
 250         notify_cpu_starting(cpu);
 251
 252         calibrate_delay();
 253
 254         smp_store_cpu_info(cpu);
 255
 256         /*
 257          * OK, now it's safe to let the boot CPU continue.  Wait for
 258          * the CPU migration code to notice that the CPU is online
 259          * before we continue - which happens after __cpu_up returns.
 260          */
 261         set_cpu_online(cpu, true);
 262         complete(&cpu_running);
 263
 264         /*
 265          * Setup the percpu timer for this CPU.
 266          */
 267         percpu_timer_setup();
 268
 269         local_irq_enable();
 270         local_fiq_enable();
 271
 272         /*
 273          * OK, it's off to the idle thread for us
 274          */
 275         cpu_idle();
 276 }
 277
 278 void __init smp_cpus_done(unsigned int max_cpus)
 279 {
 280         int cpu;
 281         unsigned long bogosum = 0;
 282
 283         for_each_online_cpu(cpu)
 284                 bogosum += per_cpu(cpu_data, cpu).loops_per_jiffy;
 285
 286         printk(KERN_INFO "SMP: Total of %d processors activated "
 287                "(%lu.%02lu BogoMIPS).\n",
 288                num_online_cpus(),
 289                bogosum / (500000/HZ),
 290                (bogosum / (5000/HZ)) % 100);
 291 }
 292
 293 void __init smp_prepare_boot_cpu(void)
 294 {
 295 }
 296
 297 void __init smp_prepare_cpus(unsigned int max_cpus)
 298 {
 299         unsigned int ncores = num_possible_cpus();
 300
 301         init_cpu_topology();
 302
 303         smp_store_cpu_info(smp_processor_id());
 304
 305         /*
 306          * are we trying to boot more cores than exist?
 307          */
 308         if (max_cpus > ncores)
 309                 max_cpus = ncores;
 310         if (ncores > 1 && max_cpus) {
 311                 /*
 312                  * Enable the local timer or broadcast device for the
 313                  * boot CPU, but only if we have more than one CPU.
 314                  */
 315                 percpu_timer_setup();
 316
 317                 /*
 318                  * Initialise the present map, which describes the set of CPUs
 319                  * actually populated at the present time. A platform should
 320                  * re-initialize the map in platform_smp_prepare_cpus() if
 321                  * present != possible (e.g. physical hotplug).
 322                  */
 323                 init_cpu_present(cpu_possible_mask);
 324
 325                 /*
 326                  * Initialise the SCU if there are more than one CPU
 327                  * and let them know where to start.
 328                  */
 329                 platform_smp_prepare_cpus(max_cpus);
 330         }
 331 }
 332
 333 static void (*smp_cross_call)(const struct cpumask *, unsigned int);
 334
 335 void __init set_smp_cross_call(void (*fn)(const struct cpumask *, unsigned int))
 336 {
 337         smp_cross_call = fn;
 338 }
 339
 340 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
 341 {
 342         smp_cross_call(mask, IPI_CALL_FUNC);
 343 }
 344
 345 void arch_send_call_function_single_ipi(int cpu)
 346 {
 347         smp_cross_call(cpumask_of(cpu), IPI_CALL_FUNC_SINGLE);
 348 }
 349
 350 static const char *ipi_types[NR_IPI] = {
 351 #define S(x,s)  [x - IPI_TIMER] = s
 352         S(IPI_TIMER, "Timer broadcast interrupts"),
 353         S(IPI_RESCHEDULE, "Rescheduling interrupts"),
 354         S(IPI_CALL_FUNC, "Function call interrupts"),
 355         S(IPI_CALL_FUNC_SINGLE, "Single function call interrupts"),
 356         S(IPI_CPU_STOP, "CPU stop interrupts"),
 357 };
 358
 359 void show_ipi_list(struct seq_file *p, int prec)
 360 {
 361         unsigned int cpu, i;
 362
 363         for (i = 0; i < NR_IPI; i++) {
 364                 seq_printf(p, "%*s%u: ", prec - 1, "IPI", i);
 365
 366                 for_each_present_cpu(cpu)
 367                         seq_printf(p, "%10u ",
 368                                    __get_irq_stat(cpu, ipi_irqs[i]));
 369
 370                 seq_printf(p, " %s\n", ipi_types[i]);
 371         }
 372 }
 373
 374 u64 smp_irq_stat_cpu(unsigned int cpu)
 375 {
 376         u64 sum = 0;
 377         int i;
 378
 379         for (i = 0; i < NR_IPI; i++)
 380                 sum += __get_irq_stat(cpu, ipi_irqs[i]);
 381
 382         return sum;
 383 }
 384
 385 /*
 386  * Timer (local or broadcast) support
 387  */
 388 static DEFINE_PER_CPU(struct clock_event_device, percpu_clockevent);
 389
 390 static void ipi_timer(void)
 391 {
 392         struct clock_event_device *evt = &__get_cpu_var(percpu_clockevent);
 393         evt->event_handler(evt);
 394 }
 395
 396 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 397 static void smp_timer_broadcast(const struct cpumask *mask)
 398 {
 399         smp_cross_call(mask, IPI_TIMER);
 400 }
 401 #else
 402 #define smp_timer_broadcast     NULL
 403 #endif
 404
 405 static void broadcast_timer_set_mode(enum clock_event_mode mode,
 406         struct clock_event_device *evt)
 407 {
 408 }
 409
 410 static void __cpuinit broadcast_timer_setup(struct clock_event_device *evt)
 411 {
 412         evt->name       = "dummy_timer";
 413         evt->features   = CLOCK_EVT_FEAT_ONESHOT |
 414                           CLOCK_EVT_FEAT_PERIODIC |
 415                           CLOCK_EVT_FEAT_DUMMY;
 416         evt->rating     = 400;
 417         evt->mult       = 1;
 418         evt->set_mode   = broadcast_timer_set_mode;
 419
 420         clockevents_register_device(evt);
 421 }
 422
 423 static struct local_timer_ops *lt_ops;
 424
 425 #ifdef CONFIG_LOCAL_TIMERS
 426 int local_timer_register(struct local_timer_ops *ops)
 427 {
 428         if (!is_smp() || !setup_max_cpus)
 429                 return -ENXIO;
 430
 431         if (lt_ops)
 432                 return -EBUSY;
 433
 434         lt_ops = ops;
 435         return 0;
 436 }
 437 #endif
 438
 439 static void __cpuinit percpu_timer_setup(void)
 440 {
 441         unsigned int cpu = smp_processor_id();
 442         struct clock_event_device *evt = &per_cpu(percpu_clockevent, cpu);
 443
 444         evt->cpumask = cpumask_of(cpu);
 445         evt->broadcast = smp_timer_broadcast;
 446
 447         if (!lt_ops || lt_ops->setup(evt))
 448                 broadcast_timer_setup(evt);
 449 }
 450
 451 #ifdef CONFIG_HOTPLUG_CPU
 452 /*
 453  * The generic clock events code purposely does not stop the local timer
 454  * on CPU_DEAD/CPU_DEAD_FROZEN hotplug events, so we have to do it
 455  * manually here.
 456  */
 457 static void percpu_timer_stop(void)
 458 {
 459         unsigned int cpu = smp_processor_id();
 460         struct clock_event_device *evt = &per_cpu(percpu_clockevent, cpu);
 461
 462         if (lt_ops)
 463                 lt_ops->stop(evt);
 464 }
 465 #endif
 466
 467 static DEFINE_RAW_SPINLOCK(stop_lock);
 468
 469 /*
 470  * ipi_cpu_stop - handle IPI from smp_send_stop()
 471  */
 472 static void ipi_cpu_stop(unsigned int cpu)
 473 {
 474         if (system_state == SYSTEM_BOOTING ||
 475             system_state == SYSTEM_RUNNING) {
 476                 raw_spin_lock(&stop_lock);
 477                 printk(KERN_CRIT "CPU%u: stopping\n", cpu);
 478                 dump_stack();
 479                 raw_spin_unlock(&stop_lock);
 480         }
 481
 482         set_cpu_online(cpu, false);
 483
 484         local_fiq_disable();
 485         local_irq_disable();
 486
 487         while (1)
 488                 cpu_relax();
 489 }
 490
 491 /*
 492  * Main handler for inter-processor interrupts
 493  */
 494 asmlinkage void __exception_irq_entry do_IPI(int ipinr, struct pt_regs *regs)
 495 {
 496         handle_IPI(ipinr, regs);
 497 }
 498
 499 void handle_IPI(int ipinr, struct pt_regs *regs)
 500 {
 501         unsigned int cpu = smp_processor_id();
 502         struct pt_regs *old_regs = set_irq_regs(regs);
 503
 504         if (ipinr >= IPI_TIMER && ipinr < IPI_TIMER + NR_IPI)
 505                 __inc_irq_stat(cpu, ipi_irqs[ipinr - IPI_TIMER]);
 506
 507         switch (ipinr) {
 508         case IPI_TIMER:
 509                 irq_enter();
 510                 ipi_timer();
 511                 irq_exit();
 512                 break;
 513
 514         case IPI_RESCHEDULE:
 515                 scheduler_ipi();
 516                 break;
 517
 518         case IPI_CALL_FUNC:
 519                 irq_enter();
 520                 generic_smp_call_function_interrupt();
 521                 irq_exit();
 522                 break;
 523
 524         case IPI_CALL_FUNC_SINGLE:
 525                 irq_enter();
 526                 generic_smp_call_function_single_interrupt();
 527                 irq_exit();
 528                 break;
 529
 530         case IPI_CPU_STOP:
 531                 irq_enter();
 532                 ipi_cpu_stop(cpu);
 533                 irq_exit();
 534                 break;
 535
 536         default:
 537                 printk(KERN_CRIT "CPU%u: Unknown IPI message 0x%x\n",
 538                        cpu, ipinr);
 539                 break;
 540         }
 541         set_irq_regs(old_regs);
 542 }
 543
 544 void smp_send_reschedule(int cpu)
 545 {
 546         smp_cross_call(cpumask_of(cpu), IPI_RESCHEDULE);
 547 }
 548
 549 #ifdef CONFIG_HOTPLUG_CPU
 550 static void smp_kill_cpus(cpumask_t *mask)
 551 {
 552         unsigned int cpu;
 553         for_each_cpu(cpu, mask)
 554                 platform_cpu_kill(cpu);
 555 }
 556 #else
 557 static void smp_kill_cpus(cpumask_t *mask) { }
 558 #endif
 559
 560 void smp_send_stop(void)
 561 {
 562         unsigned long timeout;
 563         struct cpumask mask;
 564
 565         cpumask_copy(&mask, cpu_online_mask);
 566         cpumask_clear_cpu(smp_processor_id(), &mask);
 567         if (!cpumask_empty(&mask))
 568                 smp_cross_call(&mask, IPI_CPU_STOP);
 569
 570         /* Wait up to one second for other CPUs to stop */
 571         timeout = USEC_PER_SEC;
 572         while (num_online_cpus() > 1 && timeout--)
 573                 udelay(1);
 574
 575         if (num_online_cpus() > 1)
 576                 pr_warning("SMP: failed to stop secondary CPUs\n");
 577
 578         smp_kill_cpus(&mask);
 579 }
 580
 581 /*
 582  * not supported here
 583  */
 584 int setup_profiling_timer(unsigned int multiplier)
 585 {
 586         return -EINVAL;
 587 }
 588
 589 #ifdef CONFIG_CPU_FREQ
 590
 591 static DEFINE_PER_CPU(unsigned long, l_p_j_ref);
 592 static DEFINE_PER_CPU(unsigned long, l_p_j_ref_freq);
 593 static unsigned long global_l_p_j_ref;
 594 static unsigned long global_l_p_j_ref_freq;
 595
 596 static int cpufreq_callback(struct notifier_block *nb,
 597                                         unsigned long val, void *data)
 598 {
 599         struct cpufreq_freqs *freq = data;
 600         int cpu = freq->cpu;
 601
 602         if (freq->flags & CPUFREQ_CONST_LOOPS)
 603                 return NOTIFY_OK;
 604
 605         if (!per_cpu(l_p_j_ref, cpu)) {
 606                 per_cpu(l_p_j_ref, cpu) =
 607                         per_cpu(cpu_data, cpu).loops_per_jiffy;
 608                 per_cpu(l_p_j_ref_freq, cpu) = freq->old;
 609                 if (!global_l_p_j_ref) {
 610                         global_l_p_j_ref = loops_per_jiffy;
 611                         global_l_p_j_ref_freq = freq->old;
 612                 }
 613         }
 614
 615         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 616             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 617             (val == CPUFREQ_RESUMECHANGE || val == CPUFREQ_SUSPENDCHANGE)) {
 618                 loops_per_jiffy = cpufreq_scale(global_l_p_j_ref,
 619                                                 global_l_p_j_ref_freq,
 620                                                 freq->new);
 621                 per_cpu(cpu_data, cpu).loops_per_jiffy =
 622                         cpufreq_scale(per_cpu(l_p_j_ref, cpu),
 623                                         per_cpu(l_p_j_ref_freq, cpu),
 624                                         freq->new);
 625         }
 626         return NOTIFY_OK;
 627 }
 628
 629 static struct notifier_block cpufreq_notifier = {
 630         .notifier_call  = cpufreq_callback,
 631 };
 632
 633 static int __init register_cpufreq_notifier(void)
 634 {
 635         return cpufreq_register_notifier(&cpufreq_notifier,
 636                                                 CPUFREQ_TRANSITION_NOTIFIER);
 637 }
 638 core_initcall(register_cpufreq_notifier);
 639
 640 #endif