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/seq_file.h>
  23 #include <linux/irq.h>
  24 #include <linux/percpu.h>
  25 #include <linux/clockchips.h>
  26 #include <linux/completion.h>
  27 #include <linux/cpufreq.h>
  28
  29 #include <linux/atomic.h>
  30 #include <asm/smp.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 #include <asm/mach/arch.h>
  47
  48 /*
  49  * as from 2.5, kernels no longer have an init_tasks structure
  50  * so we need some other way of telling a new secondary core
  51  * where to place its SVC stack
  52  */
  53 struct secondary_data secondary_data;
  54
  55 /*
  56  * control for which core is the next to come out of the secondary
  57  * boot "holding pen"
  58  */
  59 volatile int __cpuinitdata pen_release = -1;
  60
  61 enum ipi_msg_type {
  62         IPI_WAKEUP,
  63         IPI_TIMER,
  64         IPI_RESCHEDULE,
  65         IPI_CALL_FUNC,
  66         IPI_CALL_FUNC_SINGLE,
  67         IPI_CPU_STOP,
  68 };
  69
  70 static DECLARE_COMPLETION(cpu_running);
  71
  72 static struct smp_operations smp_ops;
  73
  74 void __init smp_set_ops(struct smp_operations *ops)
  75 {
  76         if (ops)
  77                 smp_ops = *ops;
  78 };
  79
  80 int __cpuinit __cpu_up(unsigned int cpu, struct task_struct *idle)
  81 {
  82         int ret;
  83
  84         /*
  85          * We need to tell the secondary core where to find
  86          * its stack and the page tables.
  87          */
  88         secondary_data.stack = task_stack_page(idle) + THREAD_START_SP;
  89         secondary_data.pgdir = virt_to_phys(idmap_pgd);
  90         secondary_data.swapper_pg_dir = virt_to_phys(swapper_pg_dir);
  91         __cpuc_flush_dcache_area(&secondary_data, sizeof(secondary_data));
  92         outer_clean_range(__pa(&secondary_data), __pa(&secondary_data + 1));
  93
  94         /*
  95          * Now bring the CPU into our world.
  96          */
  97         ret = boot_secondary(cpu, idle);
  98         if (ret == 0) {
  99                 /*
 100                  * CPU was successfully started, wait for it
 101                  * to come online or time out.
 102                  */
 103                 wait_for_completion_timeout(&cpu_running,
 104                                                  msecs_to_jiffies(1000));
 105
 106                 if (!cpu_online(cpu)) {
 107                         pr_crit("CPU%u: failed to come online\n", cpu);
 108                         ret = -EIO;
 109                 }
 110         } else {
 111                 pr_err("CPU%u: failed to boot: %d\n", cpu, ret);
 112         }
 113
 114         secondary_data.stack = NULL;
 115         secondary_data.pgdir = 0;
 116
 117         return ret;
 118 }
 119
 120 /* platform specific SMP operations */
 121 void __init smp_init_cpus(void)
 122 {
 123         if (smp_ops.smp_init_cpus)
 124                 smp_ops.smp_init_cpus();
 125 }
 126
 127 static void __init platform_smp_prepare_cpus(unsigned int max_cpus)
 128 {
 129         if (smp_ops.smp_prepare_cpus)
 130                 smp_ops.smp_prepare_cpus(max_cpus);
 131 }
 132
 133 static void __cpuinit platform_secondary_init(unsigned int cpu)
 134 {
 135         if (smp_ops.smp_secondary_init)
 136                 smp_ops.smp_secondary_init(cpu);
 137 }
 138
 139 int __cpuinit boot_secondary(unsigned int cpu, struct task_struct *idle)
 140 {
 141         if (smp_ops.smp_boot_secondary)
 142                 return smp_ops.smp_boot_secondary(cpu, idle);
 143         return -ENOSYS;
 144 }
 145
 146 #ifdef CONFIG_HOTPLUG_CPU
 147 static void percpu_timer_stop(void);
 148
 149 static int platform_cpu_kill(unsigned int cpu)
 150 {
 151         if (smp_ops.cpu_kill)
 152                 return smp_ops.cpu_kill(cpu);
 153         return 1;
 154 }
 155
 156 static void platform_cpu_die(unsigned int cpu)
 157 {
 158         if (smp_ops.cpu_die)
 159                 smp_ops.cpu_die(cpu);
 160 }
 161
 162 static int platform_cpu_disable(unsigned int cpu)
 163 {
 164         if (smp_ops.cpu_disable)
 165                 return smp_ops.cpu_disable(cpu);
 166
 167         /*
 168          * By default, allow disabling all CPUs except the first one,
 169          * since this is special on a lot of platforms, e.g. because
 170          * of clock tick interrupts.
 171          */
 172         return cpu == 0 ? -EPERM : 0;
 173 }
 174 /*
 175  * __cpu_disable runs on the processor to be shutdown.
 176  */
 177 int __cpuinit __cpu_disable(void)
 178 {
 179         unsigned int cpu = smp_processor_id();
 180         int ret;
 181
 182         ret = platform_cpu_disable(cpu);
 183         if (ret)
 184                 return ret;
 185
 186         /*
 187          * Take this CPU offline.  Once we clear this, we can't return,
 188          * and we must not schedule until we're ready to give up the cpu.
 189          */
 190         set_cpu_online(cpu, false);
 191
 192         /*
 193          * OK - migrate IRQs away from this CPU
 194          */
 195         migrate_irqs();
 196
 197         /*
 198          * Stop the local timer for this CPU.
 199          */
 200         percpu_timer_stop();
 201
 202         /*
 203          * Flush user cache and TLB mappings, and then remove this CPU
 204          * from the vm mask set of all processes.
 205          */
 206         flush_cache_all();
 207         local_flush_tlb_all();
 208
 209         clear_tasks_mm_cpumask(cpu);
 210
 211         return 0;
 212 }
 213
 214 static DECLARE_COMPLETION(cpu_died);
 215
 216 /*
 217  * called on the thread which is asking for a CPU to be shutdown -
 218  * waits until shutdown has completed, or it is timed out.
 219  */
 220 void __cpuinit __cpu_die(unsigned int cpu)
 221 {
 222         if (!wait_for_completion_timeout(&cpu_died, msecs_to_jiffies(5000))) {
 223                 pr_err("CPU%u: cpu didn't die\n", cpu);
 224                 return;
 225         }
 226         printk(KERN_NOTICE "CPU%u: shutdown\n", cpu);
 227
 228         if (!platform_cpu_kill(cpu))
 229                 printk("CPU%u: unable to kill\n", cpu);
 230 }
 231
 232 /*
 233  * Called from the idle thread for the CPU which has been shutdown.
 234  *
 235  * Note that we disable IRQs here, but do not re-enable them
 236  * before returning to the caller. This is also the behaviour
 237  * of the other hotplug-cpu capable cores, so presumably coming
 238  * out of idle fixes this.
 239  */
 240 void __ref cpu_die(void)
 241 {
 242         unsigned int cpu = smp_processor_id();
 243
 244         idle_task_exit();
 245
 246         local_irq_disable();
 247         mb();
 248
 249         /* Tell __cpu_die() that this CPU is now safe to dispose of */
 250         RCU_NONIDLE(complete(&cpu_died));
 251
 252         /*
 253          * actual CPU shutdown procedure is at least platform (if not
 254          * CPU) specific.
 255          */
 256         platform_cpu_die(cpu);
 257
 258         /*
 259          * Do not return to the idle loop - jump back to the secondary
 260          * cpu initialisation.  There's some initialisation which needs
 261          * to be repeated to undo the effects of taking the CPU offline.
 262          */
 263         __asm__("mov    sp, %0\n"
 264         "       mov     fp, #0\n"
 265         "       b       secondary_start_kernel"
 266                 :
 267                 : "r" (task_stack_page(current) + THREAD_SIZE - 8));
 268 }
 269 #endif /* CONFIG_HOTPLUG_CPU */
 270
 271 /*
 272  * Called by both boot and secondaries to move global data into
 273  * per-processor storage.
 274  */
 275 static void __cpuinit smp_store_cpu_info(unsigned int cpuid)
 276 {
 277         struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid);
 278
 279         cpu_info->loops_per_jiffy = loops_per_jiffy;
 280
 281         store_cpu_topology(cpuid);
 282 }
 283
 284 static void percpu_timer_setup(void);
 285
 286 /*
 287  * This is the secondary CPU boot entry.  We're using this CPUs
 288  * idle thread stack, but a set of temporary page tables.
 289  */
 290 asmlinkage void __cpuinit secondary_start_kernel(void)
 291 {
 292         struct mm_struct *mm = &init_mm;
 293         unsigned int cpu = smp_processor_id();
 294
 295         /*
 296          * All kernel threads share the same mm context; grab a
 297          * reference and switch to it.
 298          */
 299         atomic_inc(&mm->mm_count);
 300         current->active_mm = mm;
 301         cpumask_set_cpu(cpu, mm_cpumask(mm));
 302         cpu_switch_mm(mm->pgd, mm);
 303         enter_lazy_tlb(mm, current);
 304         local_flush_tlb_all();
 305
 306         printk("CPU%u: Booted secondary processor\n", cpu);
 307
 308         cpu_init();
 309         preempt_disable();
 310         trace_hardirqs_off();
 311
 312         /*
 313          * Give the platform a chance to do its own initialisation.
 314          */
 315         platform_secondary_init(cpu);
 316
 317         notify_cpu_starting(cpu);
 318
 319         calibrate_delay();
 320
 321         smp_store_cpu_info(cpu);
 322
 323         /*
 324          * OK, now it's safe to let the boot CPU continue.  Wait for
 325          * the CPU migration code to notice that the CPU is online
 326          * before we continue - which happens after __cpu_up returns.
 327          */
 328         set_cpu_online(cpu, true);
 329         complete(&cpu_running);
 330
 331         /*
 332          * Setup the percpu timer for this CPU.
 333          */
 334         percpu_timer_setup();
 335
 336         local_irq_enable();
 337         local_fiq_enable();
 338
 339         /*
 340          * OK, it's off to the idle thread for us
 341          */
 342         cpu_idle();
 343 }
 344
 345 void __init smp_cpus_done(unsigned int max_cpus)
 346 {
 347         int cpu;
 348         unsigned long bogosum = 0;
 349
 350         for_each_online_cpu(cpu)
 351                 bogosum += per_cpu(cpu_data, cpu).loops_per_jiffy;
 352
 353         printk(KERN_INFO "SMP: Total of %d processors activated "
 354                "(%lu.%02lu BogoMIPS).\n",
 355                num_online_cpus(),
 356                bogosum / (500000/HZ),
 357                (bogosum / (5000/HZ)) % 100);
 358 }
 359
 360 void __init smp_prepare_boot_cpu(void)
 361 {
 362 }
 363
 364 void __init smp_prepare_cpus(unsigned int max_cpus)
 365 {
 366         unsigned int ncores = num_possible_cpus();
 367
 368         init_cpu_topology();
 369
 370         smp_store_cpu_info(smp_processor_id());
 371
 372         /*
 373          * are we trying to boot more cores than exist?
 374          */
 375         if (max_cpus > ncores)
 376                 max_cpus = ncores;
 377         if (ncores > 1 && max_cpus) {
 378                 /*
 379                  * Enable the local timer or broadcast device for the
 380                  * boot CPU, but only if we have more than one CPU.
 381                  */
 382                 percpu_timer_setup();
 383
 384                 /*
 385                  * Initialise the present map, which describes the set of CPUs
 386                  * actually populated at the present time. A platform should
 387                  * re-initialize the map in platform_smp_prepare_cpus() if
 388                  * present != possible (e.g. physical hotplug).
 389                  */
 390                 init_cpu_present(cpu_possible_mask);
 391
 392                 /*
 393                  * Initialise the SCU if there are more than one CPU
 394                  * and let them know where to start.
 395                  */
 396                 platform_smp_prepare_cpus(max_cpus);
 397         }
 398 }
 399
 400 static void (*smp_cross_call)(const struct cpumask *, unsigned int);
 401
 402 void __init set_smp_cross_call(void (*fn)(const struct cpumask *, unsigned int))
 403 {
 404         smp_cross_call = fn;
 405 }
 406
 407 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
 408 {
 409         smp_cross_call(mask, IPI_CALL_FUNC);
 410 }
 411
 412 void arch_send_call_function_single_ipi(int cpu)
 413 {
 414         smp_cross_call(cpumask_of(cpu), IPI_CALL_FUNC_SINGLE);
 415 }
 416
 417 static const char *ipi_types[NR_IPI] = {
 418 #define S(x,s)  [x] = s
 419         S(IPI_WAKEUP, "CPU wakeup interrupts"),
 420         S(IPI_TIMER, "Timer broadcast interrupts"),
 421         S(IPI_RESCHEDULE, "Rescheduling interrupts"),
 422         S(IPI_CALL_FUNC, "Function call interrupts"),
 423         S(IPI_CALL_FUNC_SINGLE, "Single function call interrupts"),
 424         S(IPI_CPU_STOP, "CPU stop interrupts"),
 425 };
 426
 427 void show_ipi_list(struct seq_file *p, int prec)
 428 {
 429         unsigned int cpu, i;
 430
 431         for (i = 0; i < NR_IPI; i++) {
 432                 seq_printf(p, "%*s%u: ", prec - 1, "IPI", i);
 433
 434                 for_each_present_cpu(cpu)
 435                         seq_printf(p, "%10u ",
 436                                    __get_irq_stat(cpu, ipi_irqs[i]));
 437
 438                 seq_printf(p, " %s\n", ipi_types[i]);
 439         }
 440 }
 441
 442 u64 smp_irq_stat_cpu(unsigned int cpu)
 443 {
 444         u64 sum = 0;
 445         int i;
 446
 447         for (i = 0; i < NR_IPI; i++)
 448                 sum += __get_irq_stat(cpu, ipi_irqs[i]);
 449
 450         return sum;
 451 }
 452
 453 /*
 454  * Timer (local or broadcast) support
 455  */
 456 static DEFINE_PER_CPU(struct clock_event_device, percpu_clockevent);
 457
 458 static void ipi_timer(void)
 459 {
 460         struct clock_event_device *evt = &__get_cpu_var(percpu_clockevent);
 461         evt->event_handler(evt);
 462 }
 463
 464 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 465 static void smp_timer_broadcast(const struct cpumask *mask)
 466 {
 467         smp_cross_call(mask, IPI_TIMER);
 468 }
 469 #else
 470 #define smp_timer_broadcast     NULL
 471 #endif
 472
 473 static void broadcast_timer_set_mode(enum clock_event_mode mode,
 474         struct clock_event_device *evt)
 475 {
 476 }
 477
 478 static void __cpuinit broadcast_timer_setup(struct clock_event_device *evt)
 479 {
 480         evt->name       = "dummy_timer";
 481         evt->features   = CLOCK_EVT_FEAT_ONESHOT |
 482                           CLOCK_EVT_FEAT_PERIODIC |
 483                           CLOCK_EVT_FEAT_DUMMY;
 484         evt->rating     = 400;
 485         evt->mult       = 1;
 486         evt->set_mode   = broadcast_timer_set_mode;
 487
 488         clockevents_register_device(evt);
 489 }
 490
 491 static struct local_timer_ops *lt_ops;
 492
 493 #ifdef CONFIG_LOCAL_TIMERS
 494 int local_timer_register(struct local_timer_ops *ops)
 495 {
 496         if (!is_smp() || !setup_max_cpus)
 497                 return -ENXIO;
 498
 499         if (lt_ops)
 500                 return -EBUSY;
 501
 502         lt_ops = ops;
 503         return 0;
 504 }
 505 #endif
 506
 507 static void __cpuinit percpu_timer_setup(void)
 508 {
 509         unsigned int cpu = smp_processor_id();
 510         struct clock_event_device *evt = &per_cpu(percpu_clockevent, cpu);
 511
 512         evt->cpumask = cpumask_of(cpu);
 513         evt->broadcast = smp_timer_broadcast;
 514
 515         if (!lt_ops || lt_ops->setup(evt))
 516                 broadcast_timer_setup(evt);
 517 }
 518
 519 #ifdef CONFIG_HOTPLUG_CPU
 520 /*
 521  * The generic clock events code purposely does not stop the local timer
 522  * on CPU_DEAD/CPU_DEAD_FROZEN hotplug events, so we have to do it
 523  * manually here.
 524  */
 525 static void percpu_timer_stop(void)
 526 {
 527         unsigned int cpu = smp_processor_id();
 528         struct clock_event_device *evt = &per_cpu(percpu_clockevent, cpu);
 529
 530         if (lt_ops)
 531                 lt_ops->stop(evt);
 532 }
 533 #endif
 534
 535 static DEFINE_RAW_SPINLOCK(stop_lock);
 536
 537 /*
 538  * ipi_cpu_stop - handle IPI from smp_send_stop()
 539  */
 540 static void ipi_cpu_stop(unsigned int cpu)
 541 {
 542         if (system_state == SYSTEM_BOOTING ||
 543             system_state == SYSTEM_RUNNING) {
 544                 raw_spin_lock(&stop_lock);
 545                 printk(KERN_CRIT "CPU%u: stopping\n", cpu);
 546                 dump_stack();
 547                 raw_spin_unlock(&stop_lock);
 548         }
 549
 550         set_cpu_online(cpu, false);
 551
 552         local_fiq_disable();
 553         local_irq_disable();
 554
 555         while (1)
 556                 cpu_relax();
 557 }
 558
 559 /*
 560  * Main handler for inter-processor interrupts
 561  */
 562 asmlinkage void __exception_irq_entry do_IPI(int ipinr, struct pt_regs *regs)
 563 {
 564         handle_IPI(ipinr, regs);
 565 }
 566
 567 void handle_IPI(int ipinr, struct pt_regs *regs)
 568 {
 569         unsigned int cpu = smp_processor_id();
 570         struct pt_regs *old_regs = set_irq_regs(regs);
 571
 572         if (ipinr < NR_IPI)
 573                 __inc_irq_stat(cpu, ipi_irqs[ipinr]);
 574
 575         switch (ipinr) {
 576         case IPI_WAKEUP:
 577                 break;
 578
 579         case IPI_TIMER:
 580                 irq_enter();
 581                 ipi_timer();
 582                 irq_exit();
 583                 break;
 584
 585         case IPI_RESCHEDULE:
 586                 scheduler_ipi();
 587                 break;
 588
 589         case IPI_CALL_FUNC:
 590                 irq_enter();
 591                 generic_smp_call_function_interrupt();
 592                 irq_exit();
 593                 break;
 594
 595         case IPI_CALL_FUNC_SINGLE:
 596                 irq_enter();
 597                 generic_smp_call_function_single_interrupt();
 598                 irq_exit();
 599                 break;
 600
 601         case IPI_CPU_STOP:
 602                 irq_enter();
 603                 ipi_cpu_stop(cpu);
 604                 irq_exit();
 605                 break;
 606
 607         default:
 608                 printk(KERN_CRIT "CPU%u: Unknown IPI message 0x%x\n",
 609                        cpu, ipinr);
 610                 break;
 611         }
 612         set_irq_regs(old_regs);
 613 }
 614
 615 void smp_send_reschedule(int cpu)
 616 {
 617         smp_cross_call(cpumask_of(cpu), IPI_RESCHEDULE);
 618 }
 619
 620 #ifdef CONFIG_HOTPLUG_CPU
 621 static void smp_kill_cpus(cpumask_t *mask)
 622 {
 623         unsigned int cpu;
 624         for_each_cpu(cpu, mask)
 625                 platform_cpu_kill(cpu);
 626 }
 627 #else
 628 static void smp_kill_cpus(cpumask_t *mask) { }
 629 #endif
 630
 631 void smp_send_stop(void)
 632 {
 633         unsigned long timeout;
 634         struct cpumask mask;
 635
 636         cpumask_copy(&mask, cpu_online_mask);
 637         cpumask_clear_cpu(smp_processor_id(), &mask);
 638         if (!cpumask_empty(&mask))
 639                 smp_cross_call(&mask, IPI_CPU_STOP);
 640
 641         /* Wait up to one second for other CPUs to stop */
 642         timeout = USEC_PER_SEC;
 643         while (num_online_cpus() > 1 && timeout--)
 644                 udelay(1);
 645
 646         if (num_online_cpus() > 1)
 647                 pr_warning("SMP: failed to stop secondary CPUs\n");
 648
 649         smp_kill_cpus(&mask);
 650 }
 651
 652 /*
 653  * not supported here
 654  */
 655 int setup_profiling_timer(unsigned int multiplier)
 656 {
 657         return -EINVAL;
 658 }
 659
 660 #ifdef CONFIG_CPU_FREQ
 661
 662 static DEFINE_PER_CPU(unsigned long, l_p_j_ref);
 663 static DEFINE_PER_CPU(unsigned long, l_p_j_ref_freq);
 664 static unsigned long global_l_p_j_ref;
 665 static unsigned long global_l_p_j_ref_freq;
 666
 667 static int cpufreq_callback(struct notifier_block *nb,
 668                                         unsigned long val, void *data)
 669 {
 670         struct cpufreq_freqs *freq = data;
 671         int cpu = freq->cpu;
 672
 673         if (freq->flags & CPUFREQ_CONST_LOOPS)
 674                 return NOTIFY_OK;
 675
 676         if (!per_cpu(l_p_j_ref, cpu)) {
 677                 per_cpu(l_p_j_ref, cpu) =
 678                         per_cpu(cpu_data, cpu).loops_per_jiffy;
 679                 per_cpu(l_p_j_ref_freq, cpu) = freq->old;
 680                 if (!global_l_p_j_ref) {
 681                         global_l_p_j_ref = loops_per_jiffy;
 682                         global_l_p_j_ref_freq = freq->old;
 683                 }
 684         }
 685
 686         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 687             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 688             (val == CPUFREQ_RESUMECHANGE || val == CPUFREQ_SUSPENDCHANGE)) {
 689                 loops_per_jiffy = cpufreq_scale(global_l_p_j_ref,
 690                                                 global_l_p_j_ref_freq,
 691                                                 freq->new);
 692                 per_cpu(cpu_data, cpu).loops_per_jiffy =
 693                         cpufreq_scale(per_cpu(l_p_j_ref, cpu),
 694                                         per_cpu(l_p_j_ref_freq, cpu),
 695                                         freq->new);
 696         }
 697         return NOTIFY_OK;
 698 }
 699
 700 static struct notifier_block cpufreq_notifier = {
 701         .notifier_call  = cpufreq_callback,
 702 };
 703
 704 static int __init register_cpufreq_notifier(void)
 705 {
 706         return cpufreq_register_notifier(&cpufreq_notifier,
 707                                                 CPUFREQ_TRANSITION_NOTIFIER);
 708 }
 709 core_initcall(register_cpufreq_notifier);
 710
 711 #endif