arch/powerpc/kernel/smp.c

   1 /*
   2  * SMP support for ppc.
   3  *
   4  * Written by Cort Dougan (cort@cs.nmt.edu) borrowing a great
   5  * deal of code from the sparc and intel versions.
   6  *
   7  * Copyright (C) 1999 Cort Dougan <cort@cs.nmt.edu>
   8  *
   9  * PowerPC-64 Support added by Dave Engebretsen, Peter Bergner, and
  10  * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
  11  *
  12  *      This program is free software; you can redistribute it and/or
  13  *      modify it under the terms of the GNU General Public License
  14  *      as published by the Free Software Foundation; either version
  15  *      2 of the License, or (at your option) any later version.
  16  */
  17
  18 #undef DEBUG
  19
  20 #include <linux/kernel.h>
  21 #include <linux/module.h>
  22 #include <linux/sched.h>
  23 #include <linux/smp.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/delay.h>
  26 #include <linux/init.h>
  27 #include <linux/spinlock.h>
  28 #include <linux/cache.h>
  29 #include <linux/err.h>
  30 #include <linux/sysdev.h>
  31 #include <linux/cpu.h>
  32 #include <linux/notifier.h>
  33 #include <linux/topology.h>
  34
  35 #include <asm/ptrace.h>
  36 #include <asm/atomic.h>
  37 #include <asm/irq.h>
  38 #include <asm/page.h>
  39 #include <asm/pgtable.h>
  40 #include <asm/prom.h>
  41 #include <asm/smp.h>
  42 #include <asm/time.h>
  43 #include <asm/machdep.h>
  44 #include <asm/cputable.h>
  45 #include <asm/system.h>
  46 #include <asm/mpic.h>
  47 #include <asm/vdso_datapage.h>
  48 #ifdef CONFIG_PPC64
  49 #include <asm/paca.h>
  50 #endif
  51
  52 #ifdef DEBUG
  53 #include <asm/udbg.h>
  54 #define DBG(fmt...) udbg_printf(fmt)
  55 #else
  56 #define DBG(fmt...)
  57 #endif
  58
  59 int smp_hw_index[NR_CPUS];
  60 struct thread_info *secondary_ti;
  61
  62 cpumask_t cpu_possible_map = CPU_MASK_NONE;
  63 cpumask_t cpu_online_map = CPU_MASK_NONE;
  64 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
  65
  66 EXPORT_SYMBOL(cpu_online_map);
  67 EXPORT_SYMBOL(cpu_possible_map);
  68 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
  69
  70 /* SMP operations for this machine */
  71 struct smp_ops_t *smp_ops;
  72
  73 static volatile unsigned int cpu_callin_map[NR_CPUS];
  74
  75 void smp_call_function_interrupt(void);
  76
  77 int smt_enabled_at_boot = 1;
  78
  79 static int ipi_fail_ok;
  80
  81 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
  82
  83 #ifdef CONFIG_PPC64
  84 void __devinit smp_generic_kick_cpu(int nr)
  85 {
  86         BUG_ON(nr < 0 || nr >= NR_CPUS);
  87
  88         /*
  89          * The processor is currently spinning, waiting for the
  90          * cpu_start field to become non-zero After we set cpu_start,
  91          * the processor will continue on to secondary_start
  92          */
  93         paca[nr].cpu_start = 1;
  94         smp_mb();
  95 }
  96 #endif
  97
  98 void smp_message_recv(int msg)
  99 {
 100         switch(msg) {
 101         case PPC_MSG_CALL_FUNCTION:
 102                 smp_call_function_interrupt();
 103                 break;
 104         case PPC_MSG_RESCHEDULE:
 105                 /* XXX Do we have to do this? */
 106                 set_need_resched();
 107                 break;
 108         case PPC_MSG_DEBUGGER_BREAK:
 109                 if (crash_ipi_function_ptr) {
 110                         crash_ipi_function_ptr(get_irq_regs());
 111                         break;
 112                 }
 113 #ifdef CONFIG_DEBUGGER
 114                 debugger_ipi(get_irq_regs());
 115                 break;
 116 #endif /* CONFIG_DEBUGGER */
 117                 /* FALLTHROUGH */
 118         default:
 119                 printk("SMP %d: smp_message_recv(): unknown msg %d\n",
 120                        smp_processor_id(), msg);
 121                 break;
 122         }
 123 }
 124
 125 void smp_send_reschedule(int cpu)
 126 {
 127         if (likely(smp_ops))
 128                 smp_ops->message_pass(cpu, PPC_MSG_RESCHEDULE);
 129 }
 130
 131 #ifdef CONFIG_DEBUGGER
 132 void smp_send_debugger_break(int cpu)
 133 {
 134         if (likely(smp_ops))
 135                 smp_ops->message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
 136 }
 137 #endif
 138
 139 #ifdef CONFIG_KEXEC
 140 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
 141 {
 142         crash_ipi_function_ptr = crash_ipi_callback;
 143         if (crash_ipi_callback && smp_ops) {
 144                 mb();
 145                 smp_ops->message_pass(MSG_ALL_BUT_SELF, PPC_MSG_DEBUGGER_BREAK);
 146         }
 147 }
 148 #endif
 149
 150 static void stop_this_cpu(void *dummy)
 151 {
 152         local_irq_disable();
 153         while (1)
 154                 ;
 155 }
 156
 157 /*
 158  * Structure and data for smp_call_function(). This is designed to minimise
 159  * static memory requirements. It also looks cleaner.
 160  * Stolen from the i386 version.
 161  */
 162 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(call_lock);
 163
 164 static struct call_data_struct {
 165         void (*func) (void *info);
 166         void *info;
 167         atomic_t started;
 168         atomic_t finished;
 169         int wait;
 170 } *call_data;
 171
 172 /* delay of at least 8 seconds */
 173 #define SMP_CALL_TIMEOUT        8
 174
 175 /*
 176  * These functions send a 'generic call function' IPI to other online
 177  * CPUS in the system.
 178  *
 179  * [SUMMARY] Run a function on other CPUs.
 180  * <func> The function to run. This must be fast and non-blocking.
 181  * <info> An arbitrary pointer to pass to the function.
 182  * <nonatomic> currently unused.
 183  * <wait> If true, wait (atomically) until function has completed on other CPUs.
 184  * [RETURNS] 0 on success, else a negative status code. Does not return until
 185  * remote CPUs are nearly ready to execute <<func>> or are or have executed.
 186  * <map> is a cpu map of the cpus to send IPI to.
 187  *
 188  * You must not call this function with disabled interrupts or from a
 189  * hardware interrupt handler or from a bottom half handler.
 190  */
 191 static int __smp_call_function_map(void (*func) (void *info), void *info,
 192                                    int nonatomic, int wait, cpumask_t map)
 193 {
 194         struct call_data_struct data;
 195         int ret = -1, num_cpus;
 196         int cpu;
 197         u64 timeout;
 198
 199         if (unlikely(smp_ops == NULL))
 200                 return ret;
 201
 202         data.func = func;
 203         data.info = info;
 204         atomic_set(&data.started, 0);
 205         data.wait = wait;
 206         if (wait)
 207                 atomic_set(&data.finished, 0);
 208
 209         /* remove 'self' from the map */
 210         if (cpu_isset(smp_processor_id(), map))
 211                 cpu_clear(smp_processor_id(), map);
 212
 213         /* sanity check the map, remove any non-online processors. */
 214         cpus_and(map, map, cpu_online_map);
 215
 216         num_cpus = cpus_weight(map);
 217         if (!num_cpus)
 218                 goto done;
 219
 220         call_data = &data;
 221         smp_wmb();
 222         /* Send a message to all CPUs in the map */
 223         for_each_cpu_mask(cpu, map)
 224                 smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNCTION);
 225
 226         timeout = get_tb() + (u64) SMP_CALL_TIMEOUT * tb_ticks_per_sec;
 227
 228         /* Wait for indication that they have received the message */
 229         while (atomic_read(&data.started) != num_cpus) {
 230                 HMT_low();
 231                 if (get_tb() >= timeout) {
 232                         printk("smp_call_function on cpu %d: other cpus not "
 233                                 "responding (%d)\n", smp_processor_id(),
 234                                 atomic_read(&data.started));
 235                         if (!ipi_fail_ok)
 236                                 debugger(NULL);
 237                         goto out;
 238                 }
 239         }
 240
 241         /* optionally wait for the CPUs to complete */
 242         if (wait) {
 243                 while (atomic_read(&data.finished) != num_cpus) {
 244                         HMT_low();
 245                         if (get_tb() >= timeout) {
 246                                 printk("smp_call_function on cpu %d: other "
 247                                         "cpus not finishing (%d/%d)\n",
 248                                         smp_processor_id(),
 249                                         atomic_read(&data.finished),
 250                                         atomic_read(&data.started));
 251                                 debugger(NULL);
 252                                 goto out;
 253                         }
 254                 }
 255         }
 256
 257  done:
 258         ret = 0;
 259
 260  out:
 261         call_data = NULL;
 262         HMT_medium();
 263         return ret;
 264 }
 265
 266 static int __smp_call_function(void (*func)(void *info), void *info,
 267                                int nonatomic, int wait)
 268 {
 269         int ret;
 270         spin_lock(&call_lock);
 271         ret =__smp_call_function_map(func, info, nonatomic, wait,
 272                                        cpu_online_map);
 273         spin_unlock(&call_lock);
 274         return ret;
 275 }
 276
 277 int smp_call_function(void (*func) (void *info), void *info, int nonatomic,
 278                         int wait)
 279 {
 280         /* Can deadlock when called with interrupts disabled */
 281         WARN_ON(irqs_disabled());
 282
 283         return __smp_call_function(func, info, nonatomic, wait);
 284 }
 285 EXPORT_SYMBOL(smp_call_function);
 286
 287 int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
 288                              int nonatomic, int wait)
 289 {
 290         cpumask_t map = CPU_MASK_NONE;
 291         int ret = 0;
 292
 293         /* Can deadlock when called with interrupts disabled */
 294         WARN_ON(irqs_disabled());
 295
 296         if (!cpu_online(cpu))
 297                 return -EINVAL;
 298
 299         cpu_set(cpu, map);
 300         if (cpu != get_cpu()) {
 301                 spin_lock(&call_lock);
 302                 ret = __smp_call_function_map(func, info, nonatomic, wait, map);
 303                 spin_unlock(&call_lock);
 304         } else {
 305                 local_irq_disable();
 306                 func(info);
 307                 local_irq_enable();
 308         }
 309         put_cpu();
 310         return ret;
 311 }
 312 EXPORT_SYMBOL(smp_call_function_single);
 313
 314 void smp_send_stop(void)
 315 {
 316         int nolock;
 317
 318         /* It's OK to fail sending the IPI, since the alternative is to
 319          * be stuck forever waiting on the other CPU to take the interrupt.
 320          *
 321          * It's better to at least continue and go through reboot, since this
 322          * function is usually called at panic or reboot time in the first
 323          * place.
 324          */
 325         ipi_fail_ok = 1;
 326
 327         /* Don't deadlock in case we got called through panic */
 328         nolock = !spin_trylock(&call_lock);
 329         __smp_call_function_map(stop_this_cpu, NULL, 1, 0, cpu_online_map);
 330         if (!nolock)
 331                 spin_unlock(&call_lock);
 332 }
 333
 334 void smp_call_function_interrupt(void)
 335 {
 336         void (*func) (void *info);
 337         void *info;
 338         int wait;
 339
 340         /* call_data will be NULL if the sender timed out while
 341          * waiting on us to receive the call.
 342          */
 343         if (!call_data)
 344                 return;
 345
 346         func = call_data->func;
 347         info = call_data->info;
 348         wait = call_data->wait;
 349
 350         if (!wait)
 351                 smp_mb__before_atomic_inc();
 352
 353         /*
 354          * Notify initiating CPU that I've grabbed the data and am
 355          * about to execute the function
 356          */
 357         atomic_inc(&call_data->started);
 358         /*
 359          * At this point the info structure may be out of scope unless wait==1
 360          */
 361         (*func)(info);
 362         if (wait) {
 363                 smp_mb__before_atomic_inc();
 364                 atomic_inc(&call_data->finished);
 365         }
 366 }
 367
 368 extern struct gettimeofday_struct do_gtod;
 369
 370 struct thread_info *current_set[NR_CPUS];
 371
 372 DECLARE_PER_CPU(unsigned int, pvr);
 373
 374 static void __devinit smp_store_cpu_info(int id)
 375 {
 376         per_cpu(pvr, id) = mfspr(SPRN_PVR);
 377 }
 378
 379 static void __init smp_create_idle(unsigned int cpu)
 380 {
 381         struct task_struct *p;
 382
 383         /* create a process for the processor */
 384         p = fork_idle(cpu);
 385         if (IS_ERR(p))
 386                 panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p));
 387 #ifdef CONFIG_PPC64
 388         paca[cpu].__current = p;
 389 #endif
 390         current_set[cpu] = task_thread_info(p);
 391         task_thread_info(p)->cpu = cpu;
 392 }
 393
 394 void __init smp_prepare_cpus(unsigned int max_cpus)
 395 {
 396         unsigned int cpu;
 397
 398         DBG("smp_prepare_cpus\n");
 399
 400         /*
 401          * setup_cpu may need to be called on the boot cpu. We havent
 402          * spun any cpus up but lets be paranoid.
 403          */
 404         BUG_ON(boot_cpuid != smp_processor_id());
 405
 406         /* Fixup boot cpu */
 407         smp_store_cpu_info(boot_cpuid);
 408         cpu_callin_map[boot_cpuid] = 1;
 409
 410         if (smp_ops)
 411                 max_cpus = smp_ops->probe();
 412         else
 413                 max_cpus = 1;
 414
 415         smp_space_timers(max_cpus);
 416
 417         for_each_possible_cpu(cpu)
 418                 if (cpu != boot_cpuid)
 419                         smp_create_idle(cpu);
 420 }
 421
 422 void __devinit smp_prepare_boot_cpu(void)
 423 {
 424         BUG_ON(smp_processor_id() != boot_cpuid);
 425
 426         cpu_set(boot_cpuid, cpu_online_map);
 427 #ifdef CONFIG_PPC64
 428         paca[boot_cpuid].__current = current;
 429 #endif
 430         current_set[boot_cpuid] = task_thread_info(current);
 431 }
 432
 433 #ifdef CONFIG_HOTPLUG_CPU
 434 /* State of each CPU during hotplug phases */
 435 DEFINE_PER_CPU(int, cpu_state) = { 0 };
 436
 437 int generic_cpu_disable(void)
 438 {
 439         unsigned int cpu = smp_processor_id();
 440
 441         if (cpu == boot_cpuid)
 442                 return -EBUSY;
 443
 444         cpu_clear(cpu, cpu_online_map);
 445 #ifdef CONFIG_PPC64
 446         vdso_data->processorCount--;
 447         fixup_irqs(cpu_online_map);
 448 #endif
 449         return 0;
 450 }
 451
 452 int generic_cpu_enable(unsigned int cpu)
 453 {
 454         /* Do the normal bootup if we haven't
 455          * already bootstrapped. */
 456         if (system_state != SYSTEM_RUNNING)
 457                 return -ENOSYS;
 458
 459         /* get the target out of it's holding state */
 460         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 461         smp_wmb();
 462
 463         while (!cpu_online(cpu))
 464                 cpu_relax();
 465
 466 #ifdef CONFIG_PPC64
 467         fixup_irqs(cpu_online_map);
 468         /* counter the irq disable in fixup_irqs */
 469         local_irq_enable();
 470 #endif
 471         return 0;
 472 }
 473
 474 void generic_cpu_die(unsigned int cpu)
 475 {
 476         int i;
 477
 478         for (i = 0; i < 100; i++) {
 479                 smp_rmb();
 480                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
 481                         return;
 482                 msleep(100);
 483         }
 484         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
 485 }
 486
 487 void generic_mach_cpu_die(void)
 488 {
 489         unsigned int cpu;
 490
 491         local_irq_disable();
 492         cpu = smp_processor_id();
 493         printk(KERN_DEBUG "CPU%d offline\n", cpu);
 494         __get_cpu_var(cpu_state) = CPU_DEAD;
 495         smp_wmb();
 496         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
 497                 cpu_relax();
 498         cpu_set(cpu, cpu_online_map);
 499         local_irq_enable();
 500 }
 501 #endif
 502
 503 static int __devinit cpu_enable(unsigned int cpu)
 504 {
 505         if (smp_ops && smp_ops->cpu_enable)
 506                 return smp_ops->cpu_enable(cpu);
 507
 508         return -ENOSYS;
 509 }
 510
 511 int __cpuinit __cpu_up(unsigned int cpu)
 512 {
 513         int c;
 514
 515         secondary_ti = current_set[cpu];
 516         if (!cpu_enable(cpu))
 517                 return 0;
 518
 519         if (smp_ops == NULL ||
 520             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
 521                 return -EINVAL;
 522
 523         /* Make sure callin-map entry is 0 (can be leftover a CPU
 524          * hotplug
 525          */
 526         cpu_callin_map[cpu] = 0;
 527
 528         /* The information for processor bringup must
 529          * be written out to main store before we release
 530          * the processor.
 531          */
 532         smp_mb();
 533
 534         /* wake up cpus */
 535         DBG("smp: kicking cpu %d\n", cpu);
 536         smp_ops->kick_cpu(cpu);
 537
 538         /*
 539          * wait to see if the cpu made a callin (is actually up).
 540          * use this value that I found through experimentation.
 541          * -- Cort
 542          */
 543         if (system_state < SYSTEM_RUNNING)
 544                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
 545                         udelay(100);
 546 #ifdef CONFIG_HOTPLUG_CPU
 547         else
 548                 /*
 549                  * CPUs can take much longer to come up in the
 550                  * hotplug case.  Wait five seconds.
 551                  */
 552                 for (c = 25; c && !cpu_callin_map[cpu]; c--) {
 553                         msleep(200);
 554                 }
 555 #endif
 556
 557         if (!cpu_callin_map[cpu]) {
 558                 printk("Processor %u is stuck.\n", cpu);
 559                 return -ENOENT;
 560         }
 561
 562         printk("Processor %u found.\n", cpu);
 563
 564         if (smp_ops->give_timebase)
 565                 smp_ops->give_timebase();
 566
 567         /* Wait until cpu puts itself in the online map */
 568         while (!cpu_online(cpu))
 569                 cpu_relax();
 570
 571         return 0;
 572 }
 573
 574
 575 /* Activate a secondary processor. */
 576 int __devinit start_secondary(void *unused)
 577 {
 578         unsigned int cpu = smp_processor_id();
 579
 580         atomic_inc(&init_mm.mm_count);
 581         current->active_mm = &init_mm;
 582
 583         smp_store_cpu_info(cpu);
 584         set_dec(tb_ticks_per_jiffy);
 585         preempt_disable();
 586         cpu_callin_map[cpu] = 1;
 587
 588         smp_ops->setup_cpu(cpu);
 589         if (smp_ops->take_timebase)
 590                 smp_ops->take_timebase();
 591
 592         if (system_state > SYSTEM_BOOTING)
 593                 snapshot_timebase();
 594
 595         secondary_cpu_time_init();
 596
 597         spin_lock(&call_lock);
 598         cpu_set(cpu, cpu_online_map);
 599         spin_unlock(&call_lock);
 600
 601         local_irq_enable();
 602
 603         cpu_idle();
 604         return 0;
 605 }
 606
 607 int setup_profiling_timer(unsigned int multiplier)
 608 {
 609         return 0;
 610 }
 611
 612 void __init smp_cpus_done(unsigned int max_cpus)
 613 {
 614         cpumask_t old_mask;
 615
 616         /* We want the setup_cpu() here to be called from CPU 0, but our
 617          * init thread may have been "borrowed" by another CPU in the meantime
 618          * se we pin us down to CPU 0 for a short while
 619          */
 620         old_mask = current->cpus_allowed;
 621         set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid));
 622
 623         if (smp_ops)
 624                 smp_ops->setup_cpu(boot_cpuid);
 625
 626         set_cpus_allowed(current, old_mask);
 627
 628         snapshot_timebases();
 629
 630         dump_numa_cpu_topology();
 631 }
 632
 633 #ifdef CONFIG_HOTPLUG_CPU
 634 int __cpu_disable(void)
 635 {
 636         if (smp_ops->cpu_disable)
 637                 return smp_ops->cpu_disable();
 638
 639         return -ENOSYS;
 640 }
 641
 642 void __cpu_die(unsigned int cpu)
 643 {
 644         if (smp_ops->cpu_die)
 645                 smp_ops->cpu_die(cpu);
 646 }
 647 #endif