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
  28 #include <asm/atomic.h>
  29 #include <asm/cacheflush.h>
  30 #include <asm/cpu.h>
  31 #include <asm/cputype.h>
  32 #include <asm/mmu_context.h>
  33 #include <asm/pgtable.h>
  34 #include <asm/pgalloc.h>
  35 #include <asm/processor.h>
  36 #include <asm/tlbflush.h>
  37 #include <asm/ptrace.h>
  38 #include <asm/localtimer.h>
  39 #include <asm/smp_plat.h>
  40
  41 /*
  42  * as from 2.5, kernels no longer have an init_tasks structure
  43  * so we need some other way of telling a new secondary core
  44  * where to place its SVC stack
  45  */
  46 struct secondary_data secondary_data;
  47
  48 /*
  49  * structures for inter-processor calls
  50  * - A collection of single bit ipi messages.
  51  */
  52 struct ipi_data {
  53         spinlock_t lock;
  54         unsigned long ipi_count;
  55         unsigned long bits;
  56 };
  57
  58 static DEFINE_PER_CPU(struct ipi_data, ipi_data) = {
  59         .lock   = SPIN_LOCK_UNLOCKED,
  60 };
  61
  62 enum ipi_msg_type {
  63         IPI_TIMER,
  64         IPI_RESCHEDULE,
  65         IPI_CALL_FUNC,
  66         IPI_CALL_FUNC_SINGLE,
  67         IPI_CPU_STOP,
  68 };
  69
  70 int __cpuinit __cpu_up(unsigned int cpu)
  71 {
  72         struct cpuinfo_arm *ci = &per_cpu(cpu_data, cpu);
  73         struct task_struct *idle = ci->idle;
  74         pgd_t *pgd;
  75         pmd_t *pmd;
  76         int ret;
  77
  78         /*
  79          * Spawn a new process manually, if not already done.
  80          * Grab a pointer to its task struct so we can mess with it
  81          */
  82         if (!idle) {
  83                 idle = fork_idle(cpu);
  84                 if (IS_ERR(idle)) {
  85                         printk(KERN_ERR "CPU%u: fork() failed\n", cpu);
  86                         return PTR_ERR(idle);
  87                 }
  88                 ci->idle = idle;
  89         } else {
  90                 /*
  91                  * Since this idle thread is being re-used, call
  92                  * init_idle() to reinitialize the thread structure.
  93                  */
  94                 init_idle(idle, cpu);
  95         }
  96
  97         /*
  98          * Allocate initial page tables to allow the new CPU to
  99          * enable the MMU safely.  This essentially means a set
 100          * of our "standard" page tables, with the addition of
 101          * a 1:1 mapping for the physical address of the kernel.
 102          */
 103         pgd = pgd_alloc(&init_mm);
 104         pmd = pmd_offset(pgd + pgd_index(PHYS_OFFSET), PHYS_OFFSET);
 105         *pmd = __pmd((PHYS_OFFSET & PGDIR_MASK) |
 106                      PMD_TYPE_SECT | PMD_SECT_AP_WRITE);
 107         flush_pmd_entry(pmd);
 108         outer_clean_range(__pa(pmd), __pa(pmd + 1));
 109
 110         /*
 111          * We need to tell the secondary core where to find
 112          * its stack and the page tables.
 113          */
 114         secondary_data.stack = task_stack_page(idle) + THREAD_START_SP;
 115         secondary_data.pgdir = virt_to_phys(pgd);
 116         __cpuc_flush_dcache_area(&secondary_data, sizeof(secondary_data));
 117         outer_clean_range(__pa(&secondary_data), __pa(&secondary_data + 1));
 118
 119         /*
 120          * Now bring the CPU into our world.
 121          */
 122         ret = boot_secondary(cpu, idle);
 123         if (ret == 0) {
 124                 unsigned long timeout;
 125
 126                 /*
 127                  * CPU was successfully started, wait for it
 128                  * to come online or time out.
 129                  */
 130                 timeout = jiffies + HZ;
 131                 while (time_before(jiffies, timeout)) {
 132                         if (cpu_online(cpu))
 133                                 break;
 134
 135                         udelay(10);
 136                         barrier();
 137                 }
 138
 139                 if (!cpu_online(cpu))
 140                         ret = -EIO;
 141         }
 142
 143         secondary_data.stack = NULL;
 144         secondary_data.pgdir = 0;
 145
 146         *pmd = __pmd(0);
 147         clean_pmd_entry(pmd);
 148         pgd_free(&init_mm, pgd);
 149
 150         if (ret) {
 151                 printk(KERN_CRIT "CPU%u: processor failed to boot\n", cpu);
 152
 153                 /*
 154                  * FIXME: We need to clean up the new idle thread. --rmk
 155                  */
 156         }
 157
 158         return ret;
 159 }
 160
 161 #ifdef CONFIG_HOTPLUG_CPU
 162 /*
 163  * __cpu_disable runs on the processor to be shutdown.
 164  */
 165 int __cpu_disable(void)
 166 {
 167         unsigned int cpu = smp_processor_id();
 168         struct task_struct *p;
 169         int ret;
 170
 171         ret = platform_cpu_disable(cpu);
 172         if (ret)
 173                 return ret;
 174
 175         /*
 176          * Take this CPU offline.  Once we clear this, we can't return,
 177          * and we must not schedule until we're ready to give up the cpu.
 178          */
 179         set_cpu_online(cpu, false);
 180
 181         /*
 182          * OK - migrate IRQs away from this CPU
 183          */
 184         migrate_irqs();
 185
 186         /*
 187          * Stop the local timer for this CPU.
 188          */
 189         local_timer_stop();
 190
 191         /*
 192          * Flush user cache and TLB mappings, and then remove this CPU
 193          * from the vm mask set of all processes.
 194          */
 195         flush_cache_all();
 196         local_flush_tlb_all();
 197
 198         read_lock(&tasklist_lock);
 199         for_each_process(p) {
 200                 if (p->mm)
 201                         cpumask_clear_cpu(cpu, mm_cpumask(p->mm));
 202         }
 203         read_unlock(&tasklist_lock);
 204
 205         return 0;
 206 }
 207
 208 /*
 209  * called on the thread which is asking for a CPU to be shutdown -
 210  * waits until shutdown has completed, or it is timed out.
 211  */
 212 void __cpu_die(unsigned int 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         local_irq_disable();
 231         idle_task_exit();
 232
 233         /*
 234          * actual CPU shutdown procedure is at least platform (if not
 235          * CPU) specific
 236          */
 237         platform_cpu_die(cpu);
 238
 239         /*
 240          * Do not return to the idle loop - jump back to the secondary
 241          * cpu initialisation.  There's some initialisation which needs
 242          * to be repeated to undo the effects of taking the CPU offline.
 243          */
 244         __asm__("mov    sp, %0\n"
 245         "       b       secondary_start_kernel"
 246                 :
 247                 : "r" (task_stack_page(current) + THREAD_SIZE - 8));
 248 }
 249 #endif /* CONFIG_HOTPLUG_CPU */
 250
 251 /*
 252  * This is the secondary CPU boot entry.  We're using this CPUs
 253  * idle thread stack, but a set of temporary page tables.
 254  */
 255 asmlinkage void __cpuinit secondary_start_kernel(void)
 256 {
 257         struct mm_struct *mm = &init_mm;
 258         unsigned int cpu = smp_processor_id();
 259
 260         printk("CPU%u: Booted secondary processor\n", cpu);
 261
 262         /*
 263          * All kernel threads share the same mm context; grab a
 264          * reference and switch to it.
 265          */
 266         atomic_inc(&mm->mm_users);
 267         atomic_inc(&mm->mm_count);
 268         current->active_mm = mm;
 269         cpumask_set_cpu(cpu, mm_cpumask(mm));
 270         cpu_switch_mm(mm->pgd, mm);
 271         enter_lazy_tlb(mm, current);
 272         local_flush_tlb_all();
 273
 274         cpu_init();
 275         preempt_disable();
 276
 277         /*
 278          * Give the platform a chance to do its own initialisation.
 279          */
 280         platform_secondary_init(cpu);
 281
 282         /*
 283          * Enable local interrupts.
 284          */
 285         notify_cpu_starting(cpu);
 286         local_irq_enable();
 287         local_fiq_enable();
 288
 289         /*
 290          * Setup the percpu timer for this CPU.
 291          */
 292         percpu_timer_setup();
 293
 294         calibrate_delay();
 295
 296         smp_store_cpu_info(cpu);
 297
 298         /*
 299          * OK, now it's safe to let the boot CPU continue
 300          */
 301         set_cpu_online(cpu, true);
 302
 303         /*
 304          * OK, it's off to the idle thread for us
 305          */
 306         cpu_idle();
 307 }
 308
 309 /*
 310  * Called by both boot and secondaries to move global data into
 311  * per-processor storage.
 312  */
 313 void __cpuinit smp_store_cpu_info(unsigned int cpuid)
 314 {
 315         struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid);
 316
 317         cpu_info->loops_per_jiffy = loops_per_jiffy;
 318 }
 319
 320 void __init smp_cpus_done(unsigned int max_cpus)
 321 {
 322         int cpu;
 323         unsigned long bogosum = 0;
 324
 325         for_each_online_cpu(cpu)
 326                 bogosum += per_cpu(cpu_data, cpu).loops_per_jiffy;
 327
 328         printk(KERN_INFO "SMP: Total of %d processors activated "
 329                "(%lu.%02lu BogoMIPS).\n",
 330                num_online_cpus(),
 331                bogosum / (500000/HZ),
 332                (bogosum / (5000/HZ)) % 100);
 333 }
 334
 335 void __init smp_prepare_boot_cpu(void)
 336 {
 337         unsigned int cpu = smp_processor_id();
 338
 339         per_cpu(cpu_data, cpu).idle = current;
 340 }
 341
 342 static void send_ipi_message(const struct cpumask *mask, enum ipi_msg_type msg)
 343 {
 344         unsigned long flags;
 345         unsigned int cpu;
 346
 347         local_irq_save(flags);
 348
 349         for_each_cpu(cpu, mask) {
 350                 struct ipi_data *ipi = &per_cpu(ipi_data, cpu);
 351
 352                 spin_lock(&ipi->lock);
 353                 ipi->bits |= 1 << msg;
 354                 spin_unlock(&ipi->lock);
 355         }
 356
 357         /*
 358          * Call the platform specific cross-CPU call function.
 359          */
 360         smp_cross_call(mask);
 361
 362         local_irq_restore(flags);
 363 }
 364
 365 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
 366 {
 367         send_ipi_message(mask, IPI_CALL_FUNC);
 368 }
 369
 370 void arch_send_call_function_single_ipi(int cpu)
 371 {
 372         send_ipi_message(cpumask_of(cpu), IPI_CALL_FUNC_SINGLE);
 373 }
 374
 375 void show_ipi_list(struct seq_file *p)
 376 {
 377         unsigned int cpu;
 378
 379         seq_puts(p, "IPI:");
 380
 381         for_each_present_cpu(cpu)
 382                 seq_printf(p, " %10lu", per_cpu(ipi_data, cpu).ipi_count);
 383
 384         seq_putc(p, '\n');
 385 }
 386
 387 void show_local_irqs(struct seq_file *p)
 388 {
 389         unsigned int cpu;
 390
 391         seq_printf(p, "LOC: ");
 392
 393         for_each_present_cpu(cpu)
 394                 seq_printf(p, "%10u ", irq_stat[cpu].local_timer_irqs);
 395
 396         seq_putc(p, '\n');
 397 }
 398
 399 /*
 400  * Timer (local or broadcast) support
 401  */
 402 static DEFINE_PER_CPU(struct clock_event_device, percpu_clockevent);
 403
 404 static void ipi_timer(void)
 405 {
 406         struct clock_event_device *evt = &__get_cpu_var(percpu_clockevent);
 407         irq_enter();
 408         evt->event_handler(evt);
 409         irq_exit();
 410 }
 411
 412 #ifdef CONFIG_LOCAL_TIMERS
 413 asmlinkage void __exception do_local_timer(struct pt_regs *regs)
 414 {
 415         struct pt_regs *old_regs = set_irq_regs(regs);
 416         int cpu = smp_processor_id();
 417
 418         if (local_timer_ack()) {
 419                 irq_stat[cpu].local_timer_irqs++;
 420                 ipi_timer();
 421         }
 422
 423         set_irq_regs(old_regs);
 424 }
 425 #endif
 426
 427 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 428 static void smp_timer_broadcast(const struct cpumask *mask)
 429 {
 430         send_ipi_message(mask, IPI_TIMER);
 431 }
 432
 433 static void broadcast_timer_set_mode(enum clock_event_mode mode,
 434         struct clock_event_device *evt)
 435 {
 436 }
 437
 438 static void local_timer_setup(struct clock_event_device *evt)
 439 {
 440         evt->name       = "dummy_timer";
 441         evt->features   = CLOCK_EVT_FEAT_ONESHOT |
 442                           CLOCK_EVT_FEAT_PERIODIC |
 443                           CLOCK_EVT_FEAT_DUMMY;
 444         evt->rating     = 400;
 445         evt->mult       = 1;
 446         evt->set_mode   = broadcast_timer_set_mode;
 447         evt->broadcast  = smp_timer_broadcast;
 448
 449         clockevents_register_device(evt);
 450 }
 451 #endif
 452
 453 void __cpuinit percpu_timer_setup(void)
 454 {
 455         unsigned int cpu = smp_processor_id();
 456         struct clock_event_device *evt = &per_cpu(percpu_clockevent, cpu);
 457
 458         evt->cpumask = cpumask_of(cpu);
 459
 460         local_timer_setup(evt);
 461 }
 462
 463 static DEFINE_SPINLOCK(stop_lock);
 464
 465 /*
 466  * ipi_cpu_stop - handle IPI from smp_send_stop()
 467  */
 468 static void ipi_cpu_stop(unsigned int cpu)
 469 {
 470         spin_lock(&stop_lock);
 471         printk(KERN_CRIT "CPU%u: stopping\n", cpu);
 472         dump_stack();
 473         spin_unlock(&stop_lock);
 474
 475         set_cpu_online(cpu, false);
 476
 477         local_fiq_disable();
 478         local_irq_disable();
 479
 480         while (1)
 481                 cpu_relax();
 482 }
 483
 484 /*
 485  * Main handler for inter-processor interrupts
 486  *
 487  * For ARM, the ipimask now only identifies a single
 488  * category of IPI (Bit 1 IPIs have been replaced by a
 489  * different mechanism):
 490  *
 491  *  Bit 0 - Inter-processor function call
 492  */
 493 asmlinkage void __exception do_IPI(struct pt_regs *regs)
 494 {
 495         unsigned int cpu = smp_processor_id();
 496         struct ipi_data *ipi = &per_cpu(ipi_data, cpu);
 497         struct pt_regs *old_regs = set_irq_regs(regs);
 498
 499         ipi->ipi_count++;
 500
 501         for (;;) {
 502                 unsigned long msgs;
 503
 504                 spin_lock(&ipi->lock);
 505                 msgs = ipi->bits;
 506                 ipi->bits = 0;
 507                 spin_unlock(&ipi->lock);
 508
 509                 if (!msgs)
 510                         break;
 511
 512                 do {
 513                         unsigned nextmsg;
 514
 515                         nextmsg = msgs & -msgs;
 516                         msgs &= ~nextmsg;
 517                         nextmsg = ffz(~nextmsg);
 518
 519                         switch (nextmsg) {
 520                         case IPI_TIMER:
 521                                 ipi_timer();
 522                                 break;
 523
 524                         case IPI_RESCHEDULE:
 525                                 /*
 526                                  * nothing more to do - eveything is
 527                                  * done on the interrupt return path
 528                                  */
 529                                 break;
 530
 531                         case IPI_CALL_FUNC:
 532                                 generic_smp_call_function_interrupt();
 533                                 break;
 534
 535                         case IPI_CALL_FUNC_SINGLE:
 536                                 generic_smp_call_function_single_interrupt();
 537                                 break;
 538
 539                         case IPI_CPU_STOP:
 540                                 ipi_cpu_stop(cpu);
 541                                 break;
 542
 543                         default:
 544                                 printk(KERN_CRIT "CPU%u: Unknown IPI message 0x%x\n",
 545                                        cpu, nextmsg);
 546                                 break;
 547                         }
 548                 } while (msgs);
 549         }
 550
 551         set_irq_regs(old_regs);
 552 }
 553
 554 void smp_send_reschedule(int cpu)
 555 {
 556         send_ipi_message(cpumask_of(cpu), IPI_RESCHEDULE);
 557 }
 558
 559 void smp_send_stop(void)
 560 {
 561         cpumask_t mask = cpu_online_map;
 562         cpu_clear(smp_processor_id(), mask);
 563         send_ipi_message(&mask, IPI_CPU_STOP);
 564 }
 565
 566 /*
 567  * not supported here
 568  */
 569 int setup_profiling_timer(unsigned int multiplier)
 570 {
 571         return -EINVAL;
 572 }
 573
 574 static void
 575 on_each_cpu_mask(void (*func)(void *), void *info, int wait,
 576                 const struct cpumask *mask)
 577 {
 578         preempt_disable();
 579
 580         smp_call_function_many(mask, func, info, wait);
 581         if (cpumask_test_cpu(smp_processor_id(), mask))
 582                 func(info);
 583
 584         preempt_enable();
 585 }
 586
 587 /**********************************************************************/
 588
 589 /*
 590  * TLB operations
 591  */
 592 struct tlb_args {
 593         struct vm_area_struct *ta_vma;
 594         unsigned long ta_start;
 595         unsigned long ta_end;
 596 };
 597
 598 static inline void ipi_flush_tlb_all(void *ignored)
 599 {
 600         local_flush_tlb_all();
 601 }
 602
 603 static inline void ipi_flush_tlb_mm(void *arg)
 604 {
 605         struct mm_struct *mm = (struct mm_struct *)arg;
 606
 607         local_flush_tlb_mm(mm);
 608 }
 609
 610 static inline void ipi_flush_tlb_page(void *arg)
 611 {
 612         struct tlb_args *ta = (struct tlb_args *)arg;
 613
 614         local_flush_tlb_page(ta->ta_vma, ta->ta_start);
 615 }
 616
 617 static inline void ipi_flush_tlb_kernel_page(void *arg)
 618 {
 619         struct tlb_args *ta = (struct tlb_args *)arg;
 620
 621         local_flush_tlb_kernel_page(ta->ta_start);
 622 }
 623
 624 static inline void ipi_flush_tlb_range(void *arg)
 625 {
 626         struct tlb_args *ta = (struct tlb_args *)arg;
 627
 628         local_flush_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
 629 }
 630
 631 static inline void ipi_flush_tlb_kernel_range(void *arg)
 632 {
 633         struct tlb_args *ta = (struct tlb_args *)arg;
 634
 635         local_flush_tlb_kernel_range(ta->ta_start, ta->ta_end);
 636 }
 637
 638 void flush_tlb_all(void)
 639 {
 640         if (tlb_ops_need_broadcast())
 641                 on_each_cpu(ipi_flush_tlb_all, NULL, 1);
 642         else
 643                 local_flush_tlb_all();
 644 }
 645
 646 void flush_tlb_mm(struct mm_struct *mm)
 647 {
 648         if (tlb_ops_need_broadcast())
 649                 on_each_cpu_mask(ipi_flush_tlb_mm, mm, 1, mm_cpumask(mm));
 650         else
 651                 local_flush_tlb_mm(mm);
 652 }
 653
 654 void flush_tlb_page(struct vm_area_struct *vma, unsigned long uaddr)
 655 {
 656         if (tlb_ops_need_broadcast()) {
 657                 struct tlb_args ta;
 658                 ta.ta_vma = vma;
 659                 ta.ta_start = uaddr;
 660                 on_each_cpu_mask(ipi_flush_tlb_page, &ta, 1, mm_cpumask(vma->vm_mm));
 661         } else
 662                 local_flush_tlb_page(vma, uaddr);
 663 }
 664
 665 void flush_tlb_kernel_page(unsigned long kaddr)
 666 {
 667         if (tlb_ops_need_broadcast()) {
 668                 struct tlb_args ta;
 669                 ta.ta_start = kaddr;
 670                 on_each_cpu(ipi_flush_tlb_kernel_page, &ta, 1);
 671         } else
 672                 local_flush_tlb_kernel_page(kaddr);
 673 }
 674
 675 void flush_tlb_range(struct vm_area_struct *vma,
 676                      unsigned long start, unsigned long end)
 677 {
 678         if (tlb_ops_need_broadcast()) {
 679                 struct tlb_args ta;
 680                 ta.ta_vma = vma;
 681                 ta.ta_start = start;
 682                 ta.ta_end = end;
 683                 on_each_cpu_mask(ipi_flush_tlb_range, &ta, 1, mm_cpumask(vma->vm_mm));
 684         } else
 685                 local_flush_tlb_range(vma, start, end);
 686 }
 687
 688 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 689 {
 690         if (tlb_ops_need_broadcast()) {
 691                 struct tlb_args ta;
 692                 ta.ta_start = start;
 693                 ta.ta_end = end;
 694                 on_each_cpu(ipi_flush_tlb_kernel_range, &ta, 1);
 695         } else
 696                 local_flush_tlb_kernel_range(start, end);
 697 }