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/virt.h>
  47 #include <asm/mach/arch.h>
  48
  49 /*
  50  * as from 2.5, kernels no longer have an init_tasks structure
  51  * so we need some other way of telling a new secondary core
  52  * where to place its SVC stack
  53  */
  54 struct secondary_data secondary_data;
  55
  56 /*
  57  * control for which core is the next to come out of the secondary
  58  * boot "holding pen"
  59  */
  60 volatile int __cpuinitdata pen_release = -1;
  61
  62 enum ipi_msg_type {
  63         IPI_WAKEUP,
  64         IPI_TIMER,
  65         IPI_RESCHEDULE,
  66         IPI_CALL_FUNC,
  67         IPI_CALL_FUNC_SINGLE,
  68         IPI_CPU_STOP,
  69 };
  70
  71 static DECLARE_COMPLETION(cpu_running);
  72
  73 static struct smp_operations smp_ops;
  74
  75 void __init smp_set_ops(struct smp_operations *ops)
  76 {
  77         if (ops)
  78                 smp_ops = *ops;
  79 };
  80
  81 int __cpuinit __cpu_up(unsigned int cpu, struct task_struct *idle)
  82 {
  83         int ret;
  84
  85         /*
  86          * We need to tell the secondary core where to find
  87          * its stack and the page tables.
  88          */
  89         secondary_data.stack = task_stack_page(idle) + THREAD_START_SP;
  90         secondary_data.pgdir = virt_to_phys(idmap_pgd);
  91         secondary_data.swapper_pg_dir = virt_to_phys(swapper_pg_dir);
  92         __cpuc_flush_dcache_area(&secondary_data, sizeof(secondary_data));
  93         outer_clean_range(__pa(&secondary_data), __pa(&secondary_data + 1));
  94
  95         /*
  96          * Now bring the CPU into our world.
  97          */
  98         ret = boot_secondary(cpu, idle);
  99         if (ret == 0) {
 100                 /*
 101                  * CPU was successfully started, wait for it
 102                  * to come online or time out.
 103                  */
 104                 wait_for_completion_timeout(&cpu_running,
 105                                                  msecs_to_jiffies(1000));
 106
 107                 if (!cpu_online(cpu)) {
 108                         pr_crit("CPU%u: failed to come online\n", cpu);
 109                         ret = -EIO;
 110                 }
 111         } else {
 112                 pr_err("CPU%u: failed to boot: %d\n", cpu, ret);
 113         }
 114
 115         secondary_data.stack = NULL;
 116         secondary_data.pgdir = 0;
 117
 118         return ret;
 119 }
 120
 121 /* platform specific SMP operations */
 122 void __init smp_init_cpus(void)
 123 {
 124         if (smp_ops.smp_init_cpus)
 125                 smp_ops.smp_init_cpus();
 126 }
 127
 128 int __cpuinit boot_secondary(unsigned int cpu, struct task_struct *idle)
 129 {
 130         if (smp_ops.smp_boot_secondary)
 131                 return smp_ops.smp_boot_secondary(cpu, idle);
 132         return -ENOSYS;
 133 }
 134
 135 #ifdef CONFIG_HOTPLUG_CPU
 136 static void percpu_timer_stop(void);
 137
 138 static int platform_cpu_kill(unsigned int cpu)
 139 {
 140         if (smp_ops.cpu_kill)
 141                 return smp_ops.cpu_kill(cpu);
 142         return 1;
 143 }
 144
 145 static int platform_cpu_disable(unsigned int cpu)
 146 {
 147         if (smp_ops.cpu_disable)
 148                 return smp_ops.cpu_disable(cpu);
 149
 150         /*
 151          * By default, allow disabling all CPUs except the first one,
 152          * since this is special on a lot of platforms, e.g. because
 153          * of clock tick interrupts.
 154          */
 155         return cpu == 0 ? -EPERM : 0;
 156 }
 157 /*
 158  * __cpu_disable runs on the processor to be shutdown.
 159  */
 160 int __cpuinit __cpu_disable(void)
 161 {
 162         unsigned int cpu = smp_processor_id();
 163         int ret;
 164
 165         ret = platform_cpu_disable(cpu);
 166         if (ret)
 167                 return ret;
 168
 169         /*
 170          * Take this CPU offline.  Once we clear this, we can't return,
 171          * and we must not schedule until we're ready to give up the cpu.
 172          */
 173         set_cpu_online(cpu, false);
 174
 175         /*
 176          * OK - migrate IRQs away from this CPU
 177          */
 178         migrate_irqs();
 179
 180         /*
 181          * Stop the local timer for this CPU.
 182          */
 183         percpu_timer_stop();
 184
 185         /*
 186          * Flush user cache and TLB mappings, and then remove this CPU
 187          * from the vm mask set of all processes.
 188          *
 189          * Caches are flushed to the Level of Unification Inner Shareable
 190          * to write-back dirty lines to unified caches shared by all CPUs.
 191          */
 192         flush_cache_louis();
 193         local_flush_tlb_all();
 194
 195         clear_tasks_mm_cpumask(cpu);
 196
 197         return 0;
 198 }
 199
 200 static DECLARE_COMPLETION(cpu_died);
 201
 202 /*
 203  * called on the thread which is asking for a CPU to be shutdown -
 204  * waits until shutdown has completed, or it is timed out.
 205  */
 206 void __cpuinit __cpu_die(unsigned int cpu)
 207 {
 208         if (!wait_for_completion_timeout(&cpu_died, msecs_to_jiffies(5000))) {
 209                 pr_err("CPU%u: cpu didn't die\n", cpu);
 210                 return;
 211         }
 212         printk(KERN_NOTICE "CPU%u: shutdown\n", cpu);
 213
 214         if (!platform_cpu_kill(cpu))
 215                 printk("CPU%u: unable to kill\n", cpu);
 216 }
 217
 218 /*
 219  * Called from the idle thread for the CPU which has been shutdown.
 220  *
 221  * Note that we disable IRQs here, but do not re-enable them
 222  * before returning to the caller. This is also the behaviour
 223  * of the other hotplug-cpu capable cores, so presumably coming
 224  * out of idle fixes this.
 225  */
 226 void __ref cpu_die(void)
 227 {
 228         unsigned int cpu = smp_processor_id();
 229
 230         idle_task_exit();
 231
 232         local_irq_disable();
 233         mb();
 234
 235         /* Tell __cpu_die() that this CPU is now safe to dispose of */
 236         RCU_NONIDLE(complete(&cpu_died));
 237
 238         /*
 239          * actual CPU shutdown procedure is at least platform (if not
 240          * CPU) specific.
 241          */
 242         if (smp_ops.cpu_die)
 243                 smp_ops.cpu_die(cpu);
 244
 245         /*
 246          * Do not return to the idle loop - jump back to the secondary
 247          * cpu initialisation.  There's some initialisation which needs
 248          * to be repeated to undo the effects of taking the CPU offline.
 249          */
 250         __asm__("mov    sp, %0\n"
 251         "       mov     fp, #0\n"
 252         "       b       secondary_start_kernel"
 253                 :
 254                 : "r" (task_stack_page(current) + THREAD_SIZE - 8));
 255 }
 256 #endif /* CONFIG_HOTPLUG_CPU */
 257
 258 /*
 259  * Called by both boot and secondaries to move global data into
 260  * per-processor storage.
 261  */
 262 static void __cpuinit smp_store_cpu_info(unsigned int cpuid)
 263 {
 264         struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid);
 265
 266         cpu_info->loops_per_jiffy = loops_per_jiffy;
 267         cpu_info->cpuid = read_cpuid_id();
 268
 269         store_cpu_topology(cpuid);
 270 }
 271
 272 static void percpu_timer_setup(void);
 273
 274 /*
 275  * This is the secondary CPU boot entry.  We're using this CPUs
 276  * idle thread stack, but a set of temporary page tables.
 277  */
 278 asmlinkage void __cpuinit secondary_start_kernel(void)
 279 {
 280         struct mm_struct *mm = &init_mm;
 281         unsigned int cpu;
 282
 283         /*
 284          * The identity mapping is uncached (strongly ordered), so
 285          * switch away from it before attempting any exclusive accesses.
 286          */
 287         cpu_switch_mm(mm->pgd, mm);
 288         enter_lazy_tlb(mm, current);
 289         local_flush_tlb_all();
 290
 291         /*
 292          * All kernel threads share the same mm context; grab a
 293          * reference and switch to it.
 294          */
 295         cpu = smp_processor_id();
 296         atomic_inc(&mm->mm_count);
 297         current->active_mm = mm;
 298         cpumask_set_cpu(cpu, mm_cpumask(mm));
 299
 300         cpu_init();
 301
 302         printk("CPU%u: Booted secondary processor\n", cpu);
 303
 304         preempt_disable();
 305         trace_hardirqs_off();
 306
 307         /*
 308          * Give the platform a chance to do its own initialisation.
 309          */
 310         if (smp_ops.smp_secondary_init)
 311                 smp_ops.smp_secondary_init(cpu);
 312
 313         notify_cpu_starting(cpu);
 314
 315         calibrate_delay();
 316
 317         smp_store_cpu_info(cpu);
 318
 319         /*
 320          * OK, now it's safe to let the boot CPU continue.  Wait for
 321          * the CPU migration code to notice that the CPU is online
 322          * before we continue - which happens after __cpu_up returns.
 323          */
 324         set_cpu_online(cpu, true);
 325         complete(&cpu_running);
 326
 327         /*
 328          * Setup the percpu timer for this CPU.
 329          */
 330         percpu_timer_setup();
 331
 332         local_irq_enable();
 333         local_fiq_enable();
 334
 335         /*
 336          * OK, it's off to the idle thread for us
 337          */
 338         cpu_idle();
 339 }
 340
 341 void __init smp_cpus_done(unsigned int max_cpus)
 342 {
 343         int cpu;
 344         unsigned long bogosum = 0;
 345
 346         for_each_online_cpu(cpu)
 347                 bogosum += per_cpu(cpu_data, cpu).loops_per_jiffy;
 348
 349         printk(KERN_INFO "SMP: Total of %d processors activated "
 350                "(%lu.%02lu BogoMIPS).\n",
 351                num_online_cpus(),
 352                bogosum / (500000/HZ),
 353                (bogosum / (5000/HZ)) % 100);
 354
 355         hyp_mode_check();
 356 }
 357
 358 void __init smp_prepare_boot_cpu(void)
 359 {
 360         set_my_cpu_offset(per_cpu_offset(smp_processor_id()));
 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 the platforms smp_prepare_cpus()
 387                  * if 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                 if (smp_ops.smp_prepare_cpus)
 396                         smp_ops.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         if (!smp_cross_call)
 405                 smp_cross_call = fn;
 406 }
 407
 408 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
 409 {
 410         smp_cross_call(mask, IPI_CALL_FUNC);
 411 }
 412
 413 void arch_send_wakeup_ipi_mask(const struct cpumask *mask)
 414 {
 415         smp_cross_call(mask, IPI_WAKEUP);
 416 }
 417
 418 void arch_send_call_function_single_ipi(int cpu)
 419 {
 420         smp_cross_call(cpumask_of(cpu), IPI_CALL_FUNC_SINGLE);
 421 }
 422
 423 static const char *ipi_types[NR_IPI] = {
 424 #define S(x,s)  [x] = s
 425         S(IPI_WAKEUP, "CPU wakeup interrupts"),
 426         S(IPI_TIMER, "Timer broadcast interrupts"),
 427         S(IPI_RESCHEDULE, "Rescheduling interrupts"),
 428         S(IPI_CALL_FUNC, "Function call interrupts"),
 429         S(IPI_CALL_FUNC_SINGLE, "Single function call interrupts"),
 430         S(IPI_CPU_STOP, "CPU stop interrupts"),
 431 };
 432
 433 void show_ipi_list(struct seq_file *p, int prec)
 434 {
 435         unsigned int cpu, i;
 436
 437         for (i = 0; i < NR_IPI; i++) {
 438                 seq_printf(p, "%*s%u: ", prec - 1, "IPI", i);
 439
 440                 for_each_online_cpu(cpu)
 441                         seq_printf(p, "%10u ",
 442                                    __get_irq_stat(cpu, ipi_irqs[i]));
 443
 444                 seq_printf(p, " %s\n", ipi_types[i]);
 445         }
 446 }
 447
 448 u64 smp_irq_stat_cpu(unsigned int cpu)
 449 {
 450         u64 sum = 0;
 451         int i;
 452
 453         for (i = 0; i < NR_IPI; i++)
 454                 sum += __get_irq_stat(cpu, ipi_irqs[i]);
 455
 456         return sum;
 457 }
 458
 459 /*
 460  * Timer (local or broadcast) support
 461  */
 462 static DEFINE_PER_CPU(struct clock_event_device, percpu_clockevent);
 463
 464 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 465 void tick_broadcast(const struct cpumask *mask)
 466 {
 467         smp_cross_call(mask, IPI_TIMER);
 468 }
 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
 512         if (!lt_ops || lt_ops->setup(evt))
 513                 broadcast_timer_setup(evt);
 514 }
 515
 516 #ifdef CONFIG_HOTPLUG_CPU
 517 /*
 518  * The generic clock events code purposely does not stop the local timer
 519  * on CPU_DEAD/CPU_DEAD_FROZEN hotplug events, so we have to do it
 520  * manually here.
 521  */
 522 static void percpu_timer_stop(void)
 523 {
 524         unsigned int cpu = smp_processor_id();
 525         struct clock_event_device *evt = &per_cpu(percpu_clockevent, cpu);
 526
 527         if (lt_ops)
 528                 lt_ops->stop(evt);
 529 }
 530 #endif
 531
 532 static DEFINE_RAW_SPINLOCK(stop_lock);
 533
 534 /*
 535  * ipi_cpu_stop - handle IPI from smp_send_stop()
 536  */
 537 static void ipi_cpu_stop(unsigned int cpu)
 538 {
 539         if (system_state == SYSTEM_BOOTING ||
 540             system_state == SYSTEM_RUNNING) {
 541                 raw_spin_lock(&stop_lock);
 542                 printk(KERN_CRIT "CPU%u: stopping\n", cpu);
 543                 dump_stack();
 544                 raw_spin_unlock(&stop_lock);
 545         }
 546
 547         set_cpu_online(cpu, false);
 548
 549         local_fiq_disable();
 550         local_irq_disable();
 551
 552         while (1)
 553                 cpu_relax();
 554 }
 555
 556 /*
 557  * Main handler for inter-processor interrupts
 558  */
 559 asmlinkage void __exception_irq_entry do_IPI(int ipinr, struct pt_regs *regs)
 560 {
 561         handle_IPI(ipinr, regs);
 562 }
 563
 564 void handle_IPI(int ipinr, struct pt_regs *regs)
 565 {
 566         unsigned int cpu = smp_processor_id();
 567         struct pt_regs *old_regs = set_irq_regs(regs);
 568
 569         if (ipinr < NR_IPI)
 570                 __inc_irq_stat(cpu, ipi_irqs[ipinr]);
 571
 572         switch (ipinr) {
 573         case IPI_WAKEUP:
 574                 break;
 575
 576 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 577         case IPI_TIMER:
 578                 irq_enter();
 579                 tick_receive_broadcast();
 580                 irq_exit();
 581                 break;
 582 #endif
 583
 584         case IPI_RESCHEDULE:
 585                 scheduler_ipi();
 586                 break;
 587
 588         case IPI_CALL_FUNC:
 589                 irq_enter();
 590                 generic_smp_call_function_interrupt();
 591                 irq_exit();
 592                 break;
 593
 594         case IPI_CALL_FUNC_SINGLE:
 595                 irq_enter();
 596                 generic_smp_call_function_single_interrupt();
 597                 irq_exit();
 598                 break;
 599
 600         case IPI_CPU_STOP:
 601                 irq_enter();
 602                 ipi_cpu_stop(cpu);
 603                 irq_exit();
 604                 break;
 605
 606         default:
 607                 printk(KERN_CRIT "CPU%u: Unknown IPI message 0x%x\n",
 608                        cpu, ipinr);
 609                 break;
 610         }
 611         set_irq_regs(old_regs);
 612 }
 613
 614 void smp_send_reschedule(int cpu)
 615 {
 616         smp_cross_call(cpumask_of(cpu), IPI_RESCHEDULE);
 617 }
 618
 619 #ifdef CONFIG_HOTPLUG_CPU
 620 static void smp_kill_cpus(cpumask_t *mask)
 621 {
 622         unsigned int cpu;
 623         for_each_cpu(cpu, mask)
 624                 platform_cpu_kill(cpu);
 625 }
 626 #else
 627 static void smp_kill_cpus(cpumask_t *mask) { }
 628 #endif
 629
 630 void smp_send_stop(void)
 631 {
 632         unsigned long timeout;
 633         struct cpumask mask;
 634
 635         cpumask_copy(&mask, cpu_online_mask);
 636         cpumask_clear_cpu(smp_processor_id(), &mask);
 637         if (!cpumask_empty(&mask))
 638                 smp_cross_call(&mask, IPI_CPU_STOP);
 639
 640         /* Wait up to one second for other CPUs to stop */
 641         timeout = USEC_PER_SEC;
 642         while (num_online_cpus() > 1 && timeout--)
 643                 udelay(1);
 644
 645         if (num_online_cpus() > 1)
 646                 pr_warning("SMP: failed to stop secondary CPUs\n");
 647
 648         smp_kill_cpus(&mask);
 649 }
 650
 651 /*
 652  * not supported here
 653  */
 654 int setup_profiling_timer(unsigned int multiplier)
 655 {
 656         return -EINVAL;
 657 }
 658
 659 #ifdef CONFIG_CPU_FREQ
 660
 661 static DEFINE_PER_CPU(unsigned long, l_p_j_ref);
 662 static DEFINE_PER_CPU(unsigned long, l_p_j_ref_freq);
 663 static unsigned long global_l_p_j_ref;
 664 static unsigned long global_l_p_j_ref_freq;
 665
 666 static int cpufreq_callback(struct notifier_block *nb,
 667                                         unsigned long val, void *data)
 668 {
 669         struct cpufreq_freqs *freq = data;
 670         int cpu = freq->cpu;
 671
 672         if (freq->flags & CPUFREQ_CONST_LOOPS)
 673                 return NOTIFY_OK;
 674
 675         if (arm_delay_ops.const_clock)
 676                 return NOTIFY_OK;
 677
 678         if (!per_cpu(l_p_j_ref, cpu)) {
 679                 per_cpu(l_p_j_ref, cpu) =
 680                         per_cpu(cpu_data, cpu).loops_per_jiffy;
 681                 per_cpu(l_p_j_ref_freq, cpu) = freq->old;
 682                 if (!global_l_p_j_ref) {
 683                         global_l_p_j_ref = loops_per_jiffy;
 684                         global_l_p_j_ref_freq = freq->old;
 685                 }
 686         }
 687
 688         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 689             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 690             (val == CPUFREQ_RESUMECHANGE || val == CPUFREQ_SUSPENDCHANGE)) {
 691                 loops_per_jiffy = cpufreq_scale(global_l_p_j_ref,
 692                                                 global_l_p_j_ref_freq,
 693                                                 freq->new);
 694                 per_cpu(cpu_data, cpu).loops_per_jiffy =
 695                         cpufreq_scale(per_cpu(l_p_j_ref, cpu),
 696                                         per_cpu(l_p_j_ref_freq, cpu),
 697                                         freq->new);
 698         }
 699         return NOTIFY_OK;
 700 }
 701
 702 static struct notifier_block cpufreq_notifier = {
 703         .notifier_call  = cpufreq_callback,
 704 };
 705
 706 static int __init register_cpufreq_notifier(void)
 707 {
 708         return cpufreq_register_notifier(&cpufreq_notifier,
 709                                                 CPUFREQ_TRANSITION_NOTIFIER);
 710 }
 711 core_initcall(register_cpufreq_notifier);
 712
 713 #endif