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         if (unlikely(smp_ops == NULL))
 205                 return ret;
 206
 207         data.func = func;
 208         data.info = info;
 209         atomic_set(&data.started, 0);
 210         data.wait = wait;
 211         if (wait)
 212                 atomic_set(&data.finished, 0);
 213
 214         spin_lock(&call_lock);
 215         /* Must grab online cpu count with preempt disabled, otherwise
 216          * it can change. */
 217         num_cpus = num_online_cpus() - 1;
 218         if (!num_cpus)
 219                 goto done;
 220
 221         /* remove 'self' from the map */
 222         if (cpu_isset(smp_processor_id(), map))
 223                 cpu_clear(smp_processor_id(), map);
 224
 225         /* sanity check the map, remove any non-online processors. */
 226         cpus_and(map, map, cpu_online_map);
 227         if (cpus_empty(map))
 228                 goto done;
 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  done:
 267         ret = 0;
 268
 269  out:
 270         call_data = NULL;
 271         HMT_medium();
 272         spin_unlock(&call_lock);
 273         return ret;
 274 }
 275
 276 int smp_call_function(void (*func) (void *info), void *info, int nonatomic,
 277                         int wait)
 278 {
 279         return smp_call_function_map(func,info,nonatomic,wait,cpu_online_map);
 280 }
 281 EXPORT_SYMBOL(smp_call_function);
 282
 283 int smp_call_function_single(int cpu, void (*func) (void *info), void *info, int nonatomic,
 284                         int wait)
 285 {
 286         cpumask_t map = CPU_MASK_NONE;
 287         int ret = 0;
 288
 289         if (!cpu_online(cpu))
 290                 return -EINVAL;
 291
 292         cpu_set(cpu, map);
 293         if (cpu != get_cpu())
 294                 ret = smp_call_function_map(func,info,nonatomic,wait,map);
 295         else {
 296                 local_irq_disable();
 297                 func(info);
 298                 local_irq_enable();
 299         }
 300         put_cpu();
 301         return ret;
 302 }
 303 EXPORT_SYMBOL(smp_call_function_single);
 304
 305 void smp_call_function_interrupt(void)
 306 {
 307         void (*func) (void *info);
 308         void *info;
 309         int wait;
 310
 311         /* call_data will be NULL if the sender timed out while
 312          * waiting on us to receive the call.
 313          */
 314         if (!call_data)
 315                 return;
 316
 317         func = call_data->func;
 318         info = call_data->info;
 319         wait = call_data->wait;
 320
 321         if (!wait)
 322                 smp_mb__before_atomic_inc();
 323
 324         /*
 325          * Notify initiating CPU that I've grabbed the data and am
 326          * about to execute the function
 327          */
 328         atomic_inc(&call_data->started);
 329         /*
 330          * At this point the info structure may be out of scope unless wait==1
 331          */
 332         (*func)(info);
 333         if (wait) {
 334                 smp_mb__before_atomic_inc();
 335                 atomic_inc(&call_data->finished);
 336         }
 337 }
 338
 339 extern struct gettimeofday_struct do_gtod;
 340
 341 struct thread_info *current_set[NR_CPUS];
 342
 343 DECLARE_PER_CPU(unsigned int, pvr);
 344
 345 static void __devinit smp_store_cpu_info(int id)
 346 {
 347         per_cpu(pvr, id) = mfspr(SPRN_PVR);
 348 }
 349
 350 static void __init smp_create_idle(unsigned int cpu)
 351 {
 352         struct task_struct *p;
 353
 354         /* create a process for the processor */
 355         p = fork_idle(cpu);
 356         if (IS_ERR(p))
 357                 panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p));
 358 #ifdef CONFIG_PPC64
 359         paca[cpu].__current = p;
 360 #endif
 361         current_set[cpu] = task_thread_info(p);
 362         task_thread_info(p)->cpu = cpu;
 363 }
 364
 365 void __init smp_prepare_cpus(unsigned int max_cpus)
 366 {
 367         unsigned int cpu;
 368
 369         DBG("smp_prepare_cpus\n");
 370
 371         /*
 372          * setup_cpu may need to be called on the boot cpu. We havent
 373          * spun any cpus up but lets be paranoid.
 374          */
 375         BUG_ON(boot_cpuid != smp_processor_id());
 376
 377         /* Fixup boot cpu */
 378         smp_store_cpu_info(boot_cpuid);
 379         cpu_callin_map[boot_cpuid] = 1;
 380
 381         if (smp_ops)
 382                 max_cpus = smp_ops->probe();
 383         else
 384                 max_cpus = 1;
 385
 386         smp_space_timers(max_cpus);
 387
 388         for_each_possible_cpu(cpu)
 389                 if (cpu != boot_cpuid)
 390                         smp_create_idle(cpu);
 391 }
 392
 393 void __devinit smp_prepare_boot_cpu(void)
 394 {
 395         BUG_ON(smp_processor_id() != boot_cpuid);
 396
 397         cpu_set(boot_cpuid, cpu_online_map);
 398 #ifdef CONFIG_PPC64
 399         paca[boot_cpuid].__current = current;
 400 #endif
 401         current_set[boot_cpuid] = task_thread_info(current);
 402 }
 403
 404 #ifdef CONFIG_HOTPLUG_CPU
 405 /* State of each CPU during hotplug phases */
 406 DEFINE_PER_CPU(int, cpu_state) = { 0 };
 407
 408 int generic_cpu_disable(void)
 409 {
 410         unsigned int cpu = smp_processor_id();
 411
 412         if (cpu == boot_cpuid)
 413                 return -EBUSY;
 414
 415         cpu_clear(cpu, cpu_online_map);
 416 #ifdef CONFIG_PPC64
 417         vdso_data->processorCount--;
 418         fixup_irqs(cpu_online_map);
 419 #endif
 420         return 0;
 421 }
 422
 423 int generic_cpu_enable(unsigned int cpu)
 424 {
 425         /* Do the normal bootup if we haven't
 426          * already bootstrapped. */
 427         if (system_state != SYSTEM_RUNNING)
 428                 return -ENOSYS;
 429
 430         /* get the target out of it's holding state */
 431         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 432         smp_wmb();
 433
 434         while (!cpu_online(cpu))
 435                 cpu_relax();
 436
 437 #ifdef CONFIG_PPC64
 438         fixup_irqs(cpu_online_map);
 439         /* counter the irq disable in fixup_irqs */
 440         local_irq_enable();
 441 #endif
 442         return 0;
 443 }
 444
 445 void generic_cpu_die(unsigned int cpu)
 446 {
 447         int i;
 448
 449         for (i = 0; i < 100; i++) {
 450                 smp_rmb();
 451                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
 452                         return;
 453                 msleep(100);
 454         }
 455         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
 456 }
 457
 458 void generic_mach_cpu_die(void)
 459 {
 460         unsigned int cpu;
 461
 462         local_irq_disable();
 463         cpu = smp_processor_id();
 464         printk(KERN_DEBUG "CPU%d offline\n", cpu);
 465         __get_cpu_var(cpu_state) = CPU_DEAD;
 466         smp_wmb();
 467         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
 468                 cpu_relax();
 469         cpu_set(cpu, cpu_online_map);
 470         local_irq_enable();
 471 }
 472 #endif
 473
 474 static int __devinit cpu_enable(unsigned int cpu)
 475 {
 476         if (smp_ops && smp_ops->cpu_enable)
 477                 return smp_ops->cpu_enable(cpu);
 478
 479         return -ENOSYS;
 480 }
 481
 482 int __cpuinit __cpu_up(unsigned int cpu)
 483 {
 484         int c;
 485
 486         secondary_ti = current_set[cpu];
 487         if (!cpu_enable(cpu))
 488                 return 0;
 489
 490         if (smp_ops == NULL ||
 491             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
 492                 return -EINVAL;
 493
 494         /* Make sure callin-map entry is 0 (can be leftover a CPU
 495          * hotplug
 496          */
 497         cpu_callin_map[cpu] = 0;
 498
 499         /* The information for processor bringup must
 500          * be written out to main store before we release
 501          * the processor.
 502          */
 503         smp_mb();
 504
 505         /* wake up cpus */
 506         DBG("smp: kicking cpu %d\n", cpu);
 507         smp_ops->kick_cpu(cpu);
 508
 509         /*
 510          * wait to see if the cpu made a callin (is actually up).
 511          * use this value that I found through experimentation.
 512          * -- Cort
 513          */
 514         if (system_state < SYSTEM_RUNNING)
 515                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
 516                         udelay(100);
 517 #ifdef CONFIG_HOTPLUG_CPU
 518         else
 519                 /*
 520                  * CPUs can take much longer to come up in the
 521                  * hotplug case.  Wait five seconds.
 522                  */
 523                 for (c = 25; c && !cpu_callin_map[cpu]; c--) {
 524                         msleep(200);
 525                 }
 526 #endif
 527
 528         if (!cpu_callin_map[cpu]) {
 529                 printk("Processor %u is stuck.\n", cpu);
 530                 return -ENOENT;
 531         }
 532
 533         printk("Processor %u found.\n", cpu);
 534
 535         if (smp_ops->give_timebase)
 536                 smp_ops->give_timebase();
 537
 538         /* Wait until cpu puts itself in the online map */
 539         while (!cpu_online(cpu))
 540                 cpu_relax();
 541
 542         return 0;
 543 }
 544
 545
 546 /* Activate a secondary processor. */
 547 int __devinit start_secondary(void *unused)
 548 {
 549         unsigned int cpu = smp_processor_id();
 550
 551         atomic_inc(&init_mm.mm_count);
 552         current->active_mm = &init_mm;
 553
 554         smp_store_cpu_info(cpu);
 555         set_dec(tb_ticks_per_jiffy);
 556         preempt_disable();
 557         cpu_callin_map[cpu] = 1;
 558
 559         smp_ops->setup_cpu(cpu);
 560         if (smp_ops->take_timebase)
 561                 smp_ops->take_timebase();
 562
 563         if (system_state > SYSTEM_BOOTING)
 564                 snapshot_timebase();
 565
 566         spin_lock(&call_lock);
 567         cpu_set(cpu, cpu_online_map);
 568         spin_unlock(&call_lock);
 569
 570         local_irq_enable();
 571
 572         cpu_idle();
 573         return 0;
 574 }
 575
 576 int setup_profiling_timer(unsigned int multiplier)
 577 {
 578         return 0;
 579 }
 580
 581 void __init smp_cpus_done(unsigned int max_cpus)
 582 {
 583         cpumask_t old_mask;
 584
 585         /* We want the setup_cpu() here to be called from CPU 0, but our
 586          * init thread may have been "borrowed" by another CPU in the meantime
 587          * se we pin us down to CPU 0 for a short while
 588          */
 589         old_mask = current->cpus_allowed;
 590         set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid));
 591
 592         if (smp_ops)
 593                 smp_ops->setup_cpu(boot_cpuid);
 594
 595         set_cpus_allowed(current, old_mask);
 596
 597         snapshot_timebases();
 598
 599         dump_numa_cpu_topology();
 600 }
 601
 602 #ifdef CONFIG_HOTPLUG_CPU
 603 int __cpu_disable(void)
 604 {
 605         if (smp_ops->cpu_disable)
 606                 return smp_ops->cpu_disable();
 607
 608         return -ENOSYS;
 609 }
 610
 611 void __cpu_die(unsigned int cpu)
 612 {
 613         if (smp_ops->cpu_die)
 614                 smp_ops->cpu_die(cpu);
 615 }
 616 #endif