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