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_TIMER = 2,
  63         IPI_RESCHEDULE,
  64         IPI_CALL_FUNC,
  65         IPI_CALL_FUNC_SINGLE,
  66         IPI_CPU_STOP,
  67 };
  68
  69 static DECLARE_COMPLETION(cpu_running);
  70
  71 static struct smp_operations smp_ops;
  72
  73 void __init smp_set_ops(struct smp_operations *ops)
  74 {
  75         if (ops)
  76                 smp_ops = *ops;
  77 };
  78
  79 int __cpuinit __cpu_up(unsigned int cpu, struct task_struct *idle)
  80 {
  81         int ret;
  82
  83         /*
  84          * We need to tell the secondary core where to find
  85          * its stack and the page tables.
  86          */
  87         secondary_data.stack = task_stack_page(idle) + THREAD_START_SP;
  88         secondary_data.pgdir = virt_to_phys(idmap_pgd);
  89         secondary_data.swapper_pg_dir = virt_to_phys(swapper_pg_dir);
  90         __cpuc_flush_dcache_area(&secondary_data, sizeof(secondary_data));
  91         outer_clean_range(__pa(&secondary_data), __pa(&secondary_data + 1));
  92
  93         /*
  94          * Now bring the CPU into our world.
  95          */
  96         ret = boot_secondary(cpu, idle);
  97         if (ret == 0) {
  98                 /*
  99                  * CPU was successfully started, wait for it
 100                  * to come online or time out.
 101                  */
 102                 wait_for_completion_timeout(&cpu_running,
 103                                                  msecs_to_jiffies(1000));
 104
 105                 if (!cpu_online(cpu)) {
 106                         pr_crit("CPU%u: failed to come online\n", cpu);
 107                         ret = -EIO;
 108                 }
 109         } else {
 110                 pr_err("CPU%u: failed to boot: %d\n", cpu, ret);
 111         }
 112
 113         secondary_data.stack = NULL;
 114         secondary_data.pgdir = 0;
 115
 116         return ret;
 117 }
 118
 119 /* platform specific SMP operations */
 120 void __init smp_init_cpus(void)
 121 {
 122         if (smp_ops.smp_init_cpus)
 123                 smp_ops.smp_init_cpus();
 124 }
 125
 126 static void __init platform_smp_prepare_cpus(unsigned int max_cpus)
 127 {
 128         if (smp_ops.smp_prepare_cpus)
 129                 smp_ops.smp_prepare_cpus(max_cpus);
 130 }
 131
 132 static void __cpuinit platform_secondary_init(unsigned int cpu)
 133 {
 134         if (smp_ops.smp_secondary_init)
 135                 smp_ops.smp_secondary_init(cpu);
 136 }
 137
 138 int __cpuinit boot_secondary(unsigned int cpu, struct task_struct *idle)
 139 {
 140         if (smp_ops.smp_boot_secondary)
 141                 return smp_ops.smp_boot_secondary(cpu, idle);
 142         return -ENOSYS;
 143 }
 144
 145 #ifdef CONFIG_HOTPLUG_CPU
 146 static void percpu_timer_stop(void);
 147
 148 static int platform_cpu_kill(unsigned int cpu)
 149 {
 150         if (smp_ops.cpu_kill)
 151                 return smp_ops.cpu_kill(cpu);
 152         return 1;
 153 }
 154
 155 static void platform_cpu_die(unsigned int cpu)
 156 {
 157         if (smp_ops.cpu_die)
 158                 smp_ops.cpu_die(cpu);
 159 }
 160
 161 static int platform_cpu_disable(unsigned int cpu)
 162 {
 163         if (smp_ops.cpu_disable)
 164                 return smp_ops.cpu_disable(cpu);
 165
 166         /*
 167          * By default, allow disabling all CPUs except the first one,
 168          * since this is special on a lot of platforms, e.g. because
 169          * of clock tick interrupts.
 170          */
 171         return cpu == 0 ? -EPERM : 0;
 172 }
 173 /*
 174  * __cpu_disable runs on the processor to be shutdown.
 175  */
 176 int __cpuinit __cpu_disable(void)
 177 {
 178         unsigned int cpu = smp_processor_id();
 179         int ret;
 180
 181         ret = platform_cpu_disable(cpu);
 182         if (ret)
 183                 return ret;
 184
 185         /*
 186          * Take this CPU offline.  Once we clear this, we can't return,
 187          * and we must not schedule until we're ready to give up the cpu.
 188          */
 189         set_cpu_online(cpu, false);
 190
 191         /*
 192          * OK - migrate IRQs away from this CPU
 193          */
 194         migrate_irqs();
 195
 196         /*
 197          * Stop the local timer for this CPU.
 198          */
 199         percpu_timer_stop();
 200
 201         /*
 202          * Flush user cache and TLB mappings, and then remove this CPU
 203          * from the vm mask set of all processes.
 204          */
 205         flush_cache_all();
 206         local_flush_tlb_all();
 207
 208         clear_tasks_mm_cpumask(cpu);
 209
 210         return 0;
 211 }
 212
 213 static DECLARE_COMPLETION(cpu_died);
 214
 215 /*
 216  * called on the thread which is asking for a CPU to be shutdown -
 217  * waits until shutdown has completed, or it is timed out.
 218  */
 219 void __cpuinit __cpu_die(unsigned int cpu)
 220 {
 221         if (!wait_for_completion_timeout(&cpu_died, msecs_to_jiffies(5000))) {
 222                 pr_err("CPU%u: cpu didn't die\n", cpu);
 223                 return;
 224         }
 225         printk(KERN_NOTICE "CPU%u: shutdown\n", cpu);
 226
 227         if (!platform_cpu_kill(cpu))
 228                 printk("CPU%u: unable to kill\n", cpu);
 229 }
 230
 231 /*
 232  * Called from the idle thread for the CPU which has been shutdown.
 233  *
 234  * Note that we disable IRQs here, but do not re-enable them
 235  * before returning to the caller. This is also the behaviour
 236  * of the other hotplug-cpu capable cores, so presumably coming
 237  * out of idle fixes this.
 238  */
 239 void __ref cpu_die(void)
 240 {
 241         unsigned int cpu = smp_processor_id();
 242
 243         idle_task_exit();
 244
 245         local_irq_disable();
 246         mb();
 247
 248         /* Tell __cpu_die() that this CPU is now safe to dispose of */
 249         RCU_NONIDLE(complete(&cpu_died));
 250
 251         /*
 252          * actual CPU shutdown procedure is at least platform (if not
 253          * CPU) specific.
 254          */
 255         platform_cpu_die(cpu);
 256
 257         /*
 258          * Do not return to the idle loop - jump back to the secondary
 259          * cpu initialisation.  There's some initialisation which needs
 260          * to be repeated to undo the effects of taking the CPU offline.
 261          */
 262         __asm__("mov    sp, %0\n"
 263         "       mov     fp, #0\n"
 264         "       b       secondary_start_kernel"
 265                 :
 266                 : "r" (task_stack_page(current) + THREAD_SIZE - 8));
 267 }
 268 #endif /* CONFIG_HOTPLUG_CPU */
 269
 270 /*
 271  * Called by both boot and secondaries to move global data into
 272  * per-processor storage.
 273  */
 274 static void __cpuinit smp_store_cpu_info(unsigned int cpuid)
 275 {
 276         struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid);
 277
 278         cpu_info->loops_per_jiffy = loops_per_jiffy;
 279
 280         store_cpu_topology(cpuid);
 281 }
 282
 283 static void percpu_timer_setup(void);
 284
 285 /*
 286  * This is the secondary CPU boot entry.  We're using this CPUs
 287  * idle thread stack, but a set of temporary page tables.
 288  */
 289 asmlinkage void __cpuinit secondary_start_kernel(void)
 290 {
 291         struct mm_struct *mm = &init_mm;
 292         unsigned int cpu = smp_processor_id();
 293
 294         /*
 295          * All kernel threads share the same mm context; grab a
 296          * reference and switch to it.
 297          */
 298         atomic_inc(&mm->mm_count);
 299         current->active_mm = mm;
 300         cpumask_set_cpu(cpu, mm_cpumask(mm));
 301         cpu_switch_mm(mm->pgd, mm);
 302         enter_lazy_tlb(mm, current);
 303         local_flush_tlb_all();
 304
 305         printk("CPU%u: Booted secondary processor\n", cpu);
 306
 307         cpu_init();
 308         preempt_disable();
 309         trace_hardirqs_off();
 310
 311         /*
 312          * Give the platform a chance to do its own initialisation.
 313          */
 314         platform_secondary_init(cpu);
 315
 316         notify_cpu_starting(cpu);
 317
 318         calibrate_delay();
 319
 320         smp_store_cpu_info(cpu);
 321
 322         /*
 323          * OK, now it's safe to let the boot CPU continue.  Wait for
 324          * the CPU migration code to notice that the CPU is online
 325          * before we continue - which happens after __cpu_up returns.
 326          */
 327         set_cpu_online(cpu, true);
 328         complete(&cpu_running);
 329
 330         /*
 331          * Setup the percpu timer for this CPU.
 332          */
 333         percpu_timer_setup();
 334
 335         local_irq_enable();
 336         local_fiq_enable();
 337
 338         /*
 339          * OK, it's off to the idle thread for us
 340          */
 341         cpu_idle();
 342 }
 343
 344 void __init smp_cpus_done(unsigned int max_cpus)
 345 {
 346         int cpu;
 347         unsigned long bogosum = 0;
 348
 349         for_each_online_cpu(cpu)
 350                 bogosum += per_cpu(cpu_data, cpu).loops_per_jiffy;
 351
 352         printk(KERN_INFO "SMP: Total of %d processors activated "
 353                "(%lu.%02lu BogoMIPS).\n",
 354                num_online_cpus(),
 355                bogosum / (500000/HZ),
 356                (bogosum / (5000/HZ)) % 100);
 357 }
 358
 359 void __init smp_prepare_boot_cpu(void)
 360 {
 361 }
 362
 363 void __init smp_prepare_cpus(unsigned int max_cpus)
 364 {
 365         unsigned int ncores = num_possible_cpus();
 366
 367         init_cpu_topology();
 368
 369         smp_store_cpu_info(smp_processor_id());
 370
 371         /*
 372          * are we trying to boot more cores than exist?
 373          */
 374         if (max_cpus > ncores)
 375                 max_cpus = ncores;
 376         if (ncores > 1 && max_cpus) {
 377                 /*
 378                  * Enable the local timer or broadcast device for the
 379                  * boot CPU, but only if we have more than one CPU.
 380                  */
 381                 percpu_timer_setup();
 382
 383                 /*
 384                  * Initialise the present map, which describes the set of CPUs
 385                  * actually populated at the present time. A platform should
 386                  * re-initialize the map in platform_smp_prepare_cpus() if
 387                  * present != possible (e.g. physical hotplug).
 388                  */
 389                 init_cpu_present(cpu_possible_mask);
 390
 391                 /*
 392                  * Initialise the SCU if there are more than one CPU
 393                  * and let them know where to start.
 394                  */
 395                 platform_smp_prepare_cpus(max_cpus);
 396         }
 397 }
 398
 399 static void (*smp_cross_call)(const struct cpumask *, unsigned int);
 400
 401 void __init set_smp_cross_call(void (*fn)(const struct cpumask *, unsigned int))
 402 {
 403         smp_cross_call = fn;
 404 }
 405
 406 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
 407 {
 408         smp_cross_call(mask, IPI_CALL_FUNC);
 409 }
 410
 411 void arch_send_call_function_single_ipi(int cpu)
 412 {
 413         smp_cross_call(cpumask_of(cpu), IPI_CALL_FUNC_SINGLE);
 414 }
 415
 416 static const char *ipi_types[NR_IPI] = {
 417 #define S(x,s)  [x - IPI_TIMER] = s
 418         S(IPI_TIMER, "Timer broadcast interrupts"),
 419         S(IPI_RESCHEDULE, "Rescheduling interrupts"),
 420         S(IPI_CALL_FUNC, "Function call interrupts"),
 421         S(IPI_CALL_FUNC_SINGLE, "Single function call interrupts"),
 422         S(IPI_CPU_STOP, "CPU stop interrupts"),
 423 };
 424
 425 void show_ipi_list(struct seq_file *p, int prec)
 426 {
 427         unsigned int cpu, i;
 428
 429         for (i = 0; i < NR_IPI; i++) {
 430                 seq_printf(p, "%*s%u: ", prec - 1, "IPI", i);
 431
 432                 for_each_present_cpu(cpu)
 433                         seq_printf(p, "%10u ",
 434                                    __get_irq_stat(cpu, ipi_irqs[i]));
 435
 436                 seq_printf(p, " %s\n", ipi_types[i]);
 437         }
 438 }
 439
 440 u64 smp_irq_stat_cpu(unsigned int cpu)
 441 {
 442         u64 sum = 0;
 443         int i;
 444
 445         for (i = 0; i < NR_IPI; i++)
 446                 sum += __get_irq_stat(cpu, ipi_irqs[i]);
 447
 448         return sum;
 449 }
 450
 451 /*
 452  * Timer (local or broadcast) support
 453  */
 454 static DEFINE_PER_CPU(struct clock_event_device, percpu_clockevent);
 455
 456 static void ipi_timer(void)
 457 {
 458         struct clock_event_device *evt = &__get_cpu_var(percpu_clockevent);
 459         evt->event_handler(evt);
 460 }
 461
 462 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 463 static void smp_timer_broadcast(const struct cpumask *mask)
 464 {
 465         smp_cross_call(mask, IPI_TIMER);
 466 }
 467 #else
 468 #define smp_timer_broadcast     NULL
 469 #endif
 470
 471 static void broadcast_timer_set_mode(enum clock_event_mode mode,
 472         struct clock_event_device *evt)
 473 {
 474 }
 475
 476 static void __cpuinit broadcast_timer_setup(struct clock_event_device *evt)
 477 {
 478         evt->name       = "dummy_timer";
 479         evt->features   = CLOCK_EVT_FEAT_ONESHOT |
 480                           CLOCK_EVT_FEAT_PERIODIC |
 481                           CLOCK_EVT_FEAT_DUMMY;
 482         evt->rating     = 400;
 483         evt->mult       = 1;
 484         evt->set_mode   = broadcast_timer_set_mode;
 485
 486         clockevents_register_device(evt);
 487 }
 488
 489 static struct local_timer_ops *lt_ops;
 490
 491 #ifdef CONFIG_LOCAL_TIMERS
 492 int local_timer_register(struct local_timer_ops *ops)
 493 {
 494         if (!is_smp() || !setup_max_cpus)
 495                 return -ENXIO;
 496
 497         if (lt_ops)
 498                 return -EBUSY;
 499
 500         lt_ops = ops;
 501         return 0;
 502 }
 503 #endif
 504
 505 static void __cpuinit percpu_timer_setup(void)
 506 {
 507         unsigned int cpu = smp_processor_id();
 508         struct clock_event_device *evt = &per_cpu(percpu_clockevent, cpu);
 509
 510         evt->cpumask = cpumask_of(cpu);
 511         evt->broadcast = smp_timer_broadcast;
 512
 513         if (!lt_ops || lt_ops->setup(evt))
 514                 broadcast_timer_setup(evt);
 515 }
 516
 517 #ifdef CONFIG_HOTPLUG_CPU
 518 /*
 519  * The generic clock events code purposely does not stop the local timer
 520  * on CPU_DEAD/CPU_DEAD_FROZEN hotplug events, so we have to do it
 521  * manually here.
 522  */
 523 static void percpu_timer_stop(void)
 524 {
 525         unsigned int cpu = smp_processor_id();
 526         struct clock_event_device *evt = &per_cpu(percpu_clockevent, cpu);
 527
 528         if (lt_ops)
 529                 lt_ops->stop(evt);
 530 }
 531 #endif
 532
 533 static DEFINE_RAW_SPINLOCK(stop_lock);
 534
 535 /*
 536  * ipi_cpu_stop - handle IPI from smp_send_stop()
 537  */
 538 static void ipi_cpu_stop(unsigned int cpu)
 539 {
 540         if (system_state == SYSTEM_BOOTING ||
 541             system_state == SYSTEM_RUNNING) {
 542                 raw_spin_lock(&stop_lock);
 543                 printk(KERN_CRIT "CPU%u: stopping\n", cpu);
 544                 dump_stack();
 545                 raw_spin_unlock(&stop_lock);
 546         }
 547
 548         set_cpu_online(cpu, false);
 549
 550         local_fiq_disable();
 551         local_irq_disable();
 552
 553         while (1)
 554                 cpu_relax();
 555 }
 556
 557 /*
 558  * Main handler for inter-processor interrupts
 559  */
 560 asmlinkage void __exception_irq_entry do_IPI(int ipinr, struct pt_regs *regs)
 561 {
 562         handle_IPI(ipinr, regs);
 563 }
 564
 565 void handle_IPI(int ipinr, struct pt_regs *regs)
 566 {
 567         unsigned int cpu = smp_processor_id();
 568         struct pt_regs *old_regs = set_irq_regs(regs);
 569
 570         if (ipinr >= IPI_TIMER && ipinr < IPI_TIMER + NR_IPI)
 571                 __inc_irq_stat(cpu, ipi_irqs[ipinr - IPI_TIMER]);
 572
 573         switch (ipinr) {
 574         case IPI_TIMER:
 575                 irq_enter();
 576                 ipi_timer();
 577                 irq_exit();
 578                 break;
 579
 580         case IPI_RESCHEDULE:
 581                 scheduler_ipi();
 582                 break;
 583
 584         case IPI_CALL_FUNC:
 585                 irq_enter();
 586                 generic_smp_call_function_interrupt();
 587                 irq_exit();
 588                 break;
 589
 590         case IPI_CALL_FUNC_SINGLE:
 591                 irq_enter();
 592                 generic_smp_call_function_single_interrupt();
 593                 irq_exit();
 594                 break;
 595
 596         case IPI_CPU_STOP:
 597                 irq_enter();
 598                 ipi_cpu_stop(cpu);
 599                 irq_exit();
 600                 break;
 601
 602         default:
 603                 printk(KERN_CRIT "CPU%u: Unknown IPI message 0x%x\n",
 604                        cpu, ipinr);
 605                 break;
 606         }
 607         set_irq_regs(old_regs);
 608 }
 609
 610 void smp_send_reschedule(int cpu)
 611 {
 612         smp_cross_call(cpumask_of(cpu), IPI_RESCHEDULE);
 613 }
 614
 615 #ifdef CONFIG_HOTPLUG_CPU
 616 static void smp_kill_cpus(cpumask_t *mask)
 617 {
 618         unsigned int cpu;
 619         for_each_cpu(cpu, mask)
 620                 platform_cpu_kill(cpu);
 621 }
 622 #else
 623 static void smp_kill_cpus(cpumask_t *mask) { }
 624 #endif
 625
 626 void smp_send_stop(void)
 627 {
 628         unsigned long timeout;
 629         struct cpumask mask;
 630
 631         cpumask_copy(&mask, cpu_online_mask);
 632         cpumask_clear_cpu(smp_processor_id(), &mask);
 633         if (!cpumask_empty(&mask))
 634                 smp_cross_call(&mask, IPI_CPU_STOP);
 635
 636         /* Wait up to one second for other CPUs to stop */
 637         timeout = USEC_PER_SEC;
 638         while (num_online_cpus() > 1 && timeout--)
 639                 udelay(1);
 640
 641         if (num_online_cpus() > 1)
 642                 pr_warning("SMP: failed to stop secondary CPUs\n");
 643
 644         smp_kill_cpus(&mask);
 645 }
 646
 647 /*
 648  * not supported here
 649  */
 650 int setup_profiling_timer(unsigned int multiplier)
 651 {
 652         return -EINVAL;
 653 }
 654
 655 #ifdef CONFIG_CPU_FREQ
 656
 657 static DEFINE_PER_CPU(unsigned long, l_p_j_ref);
 658 static DEFINE_PER_CPU(unsigned long, l_p_j_ref_freq);
 659 static unsigned long global_l_p_j_ref;
 660 static unsigned long global_l_p_j_ref_freq;
 661
 662 static int cpufreq_callback(struct notifier_block *nb,
 663                                         unsigned long val, void *data)
 664 {
 665         struct cpufreq_freqs *freq = data;
 666         int cpu = freq->cpu;
 667
 668         if (freq->flags & CPUFREQ_CONST_LOOPS)
 669                 return NOTIFY_OK;
 670
 671         if (!per_cpu(l_p_j_ref, cpu)) {
 672                 per_cpu(l_p_j_ref, cpu) =
 673                         per_cpu(cpu_data, cpu).loops_per_jiffy;
 674                 per_cpu(l_p_j_ref_freq, cpu) = freq->old;
 675                 if (!global_l_p_j_ref) {
 676                         global_l_p_j_ref = loops_per_jiffy;
 677                         global_l_p_j_ref_freq = freq->old;
 678                 }
 679         }
 680
 681         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 682             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 683             (val == CPUFREQ_RESUMECHANGE || val == CPUFREQ_SUSPENDCHANGE)) {
 684                 loops_per_jiffy = cpufreq_scale(global_l_p_j_ref,
 685                                                 global_l_p_j_ref_freq,
 686                                                 freq->new);
 687                 per_cpu(cpu_data, cpu).loops_per_jiffy =
 688                         cpufreq_scale(per_cpu(l_p_j_ref, cpu),
 689                                         per_cpu(l_p_j_ref_freq, cpu),
 690                                         freq->new);
 691         }
 692         return NOTIFY_OK;
 693 }
 694
 695 static struct notifier_block cpufreq_notifier = {
 696         .notifier_call  = cpufreq_callback,
 697 };
 698
 699 static int __init register_cpufreq_notifier(void)
 700 {
 701         return cpufreq_register_notifier(&cpufreq_notifier,
 702                                                 CPUFREQ_TRANSITION_NOTIFIER);
 703 }
 704 core_initcall(register_cpufreq_notifier);
 705
 706 #endif