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
  69 /* SMP operations for this machine */
  70 struct smp_ops_t *smp_ops;
  71
  72 static volatile unsigned int cpu_callin_map[NR_CPUS];
  73
  74 void smp_call_function_interrupt(void);
  75
  76 int smt_enabled_at_boot = 1;
  77
  78 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
  79
  80 #ifdef CONFIG_MPIC
  81 int __init smp_mpic_probe(void)
  82 {
  83         int nr_cpus;
  84
  85         DBG("smp_mpic_probe()...\n");
  86
  87         nr_cpus = cpus_weight(cpu_possible_map);
  88
  89         DBG("nr_cpus: %d\n", nr_cpus);
  90
  91         if (nr_cpus > 1)
  92                 mpic_request_ipis();
  93
  94         return nr_cpus;
  95 }
  96
  97 void __devinit smp_mpic_setup_cpu(int cpu)
  98 {
  99         mpic_setup_this_cpu();
 100 }
 101 #endif /* CONFIG_MPIC */
 102
 103 #ifdef CONFIG_PPC64
 104 void __devinit smp_generic_kick_cpu(int nr)
 105 {
 106         BUG_ON(nr < 0 || nr >= NR_CPUS);
 107
 108         /*
 109          * The processor is currently spinning, waiting for the
 110          * cpu_start field to become non-zero After we set cpu_start,
 111          * the processor will continue on to secondary_start
 112          */
 113         paca[nr].cpu_start = 1;
 114         smp_mb();
 115 }
 116 #endif
 117
 118 void smp_message_recv(int msg)
 119 {
 120         switch(msg) {
 121         case PPC_MSG_CALL_FUNCTION:
 122                 smp_call_function_interrupt();
 123                 break;
 124         case PPC_MSG_RESCHEDULE:
 125                 /* XXX Do we have to do this? */
 126                 set_need_resched();
 127                 break;
 128         case PPC_MSG_DEBUGGER_BREAK:
 129                 if (crash_ipi_function_ptr) {
 130                         crash_ipi_function_ptr(get_irq_regs());
 131                         break;
 132                 }
 133 #ifdef CONFIG_DEBUGGER
 134                 debugger_ipi(get_irq_regs());
 135                 break;
 136 #endif /* CONFIG_DEBUGGER */
 137                 /* FALLTHROUGH */
 138         default:
 139                 printk("SMP %d: smp_message_recv(): unknown msg %d\n",
 140                        smp_processor_id(), msg);
 141                 break;
 142         }
 143 }
 144
 145 void smp_send_reschedule(int cpu)
 146 {
 147         if (likely(smp_ops))
 148                 smp_ops->message_pass(cpu, PPC_MSG_RESCHEDULE);
 149 }
 150
 151 #ifdef CONFIG_DEBUGGER
 152 void smp_send_debugger_break(int cpu)
 153 {
 154         if (likely(smp_ops))
 155                 smp_ops->message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
 156 }
 157 #endif
 158
 159 #ifdef CONFIG_KEXEC
 160 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
 161 {
 162         crash_ipi_function_ptr = crash_ipi_callback;
 163         if (crash_ipi_callback && smp_ops) {
 164                 mb();
 165                 smp_ops->message_pass(MSG_ALL_BUT_SELF, PPC_MSG_DEBUGGER_BREAK);
 166         }
 167 }
 168 #endif
 169
 170 static void stop_this_cpu(void *dummy)
 171 {
 172         local_irq_disable();
 173         while (1)
 174                 ;
 175 }
 176
 177 void smp_send_stop(void)
 178 {
 179         smp_call_function(stop_this_cpu, NULL, 1, 0);
 180 }
 181
 182 /*
 183  * Structure and data for smp_call_function(). This is designed to minimise
 184  * static memory requirements. It also looks cleaner.
 185  * Stolen from the i386 version.
 186  */
 187 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(call_lock);
 188
 189 static struct call_data_struct {
 190         void (*func) (void *info);
 191         void *info;
 192         atomic_t started;
 193         atomic_t finished;
 194         int wait;
 195 } *call_data;
 196
 197 /* delay of at least 8 seconds */
 198 #define SMP_CALL_TIMEOUT        8
 199
 200 /*
 201  * This function sends a 'generic call function' IPI to all other CPUs
 202  * in the system.
 203  *
 204  * [SUMMARY] Run a function on all other CPUs.
 205  * <func> The function to run. This must be fast and non-blocking.
 206  * <info> An arbitrary pointer to pass to the function.
 207  * <nonatomic> currently unused.
 208  * <wait> If true, wait (atomically) until function has completed on other CPUs.
 209  * [RETURNS] 0 on success, else a negative status code. Does not return until
 210  * remote CPUs are nearly ready to execute <<func>> or are or have executed.
 211  *
 212  * You must not call this function with disabled interrupts or from a
 213  * hardware interrupt handler or from a bottom half handler.
 214  */
 215 int smp_call_function (void (*func) (void *info), void *info, int nonatomic,
 216                        int wait)
 217 {
 218         struct call_data_struct data;
 219         int ret = -1, cpus;
 220         u64 timeout;
 221
 222         /* Can deadlock when called with interrupts disabled */
 223         WARN_ON(irqs_disabled());
 224
 225         if (unlikely(smp_ops == NULL))
 226                 return -1;
 227
 228         data.func = func;
 229         data.info = info;
 230         atomic_set(&data.started, 0);
 231         data.wait = wait;
 232         if (wait)
 233                 atomic_set(&data.finished, 0);
 234
 235         spin_lock(&call_lock);
 236         /* Must grab online cpu count with preempt disabled, otherwise
 237          * it can change. */
 238         cpus = num_online_cpus() - 1;
 239         if (!cpus) {
 240                 ret = 0;
 241                 goto out;
 242         }
 243
 244         call_data = &data;
 245         smp_wmb();
 246         /* Send a message to all other CPUs and wait for them to respond */
 247         smp_ops->message_pass(MSG_ALL_BUT_SELF, PPC_MSG_CALL_FUNCTION);
 248
 249         timeout = get_tb() + (u64) SMP_CALL_TIMEOUT * tb_ticks_per_sec;
 250
 251         /* Wait for response */
 252         while (atomic_read(&data.started) != cpus) {
 253                 HMT_low();
 254                 if (get_tb() >= timeout) {
 255                         printk("smp_call_function on cpu %d: other cpus not "
 256                                "responding (%d)\n", smp_processor_id(),
 257                                atomic_read(&data.started));
 258                         debugger(NULL);
 259                         goto out;
 260                 }
 261         }
 262
 263         if (wait) {
 264                 while (atomic_read(&data.finished) != cpus) {
 265                         HMT_low();
 266                         if (get_tb() >= timeout) {
 267                                 printk("smp_call_function on cpu %d: other "
 268                                        "cpus not finishing (%d/%d)\n",
 269                                        smp_processor_id(),
 270                                        atomic_read(&data.finished),
 271                                        atomic_read(&data.started));
 272                                 debugger(NULL);
 273                                 goto out;
 274                         }
 275                 }
 276         }
 277
 278         ret = 0;
 279
 280  out:
 281         call_data = NULL;
 282         HMT_medium();
 283         spin_unlock(&call_lock);
 284         return ret;
 285 }
 286
 287 EXPORT_SYMBOL(smp_call_function);
 288
 289 void smp_call_function_interrupt(void)
 290 {
 291         void (*func) (void *info);
 292         void *info;
 293         int wait;
 294
 295         /* call_data will be NULL if the sender timed out while
 296          * waiting on us to receive the call.
 297          */
 298         if (!call_data)
 299                 return;
 300
 301         func = call_data->func;
 302         info = call_data->info;
 303         wait = call_data->wait;
 304
 305         if (!wait)
 306                 smp_mb__before_atomic_inc();
 307
 308         /*
 309          * Notify initiating CPU that I've grabbed the data and am
 310          * about to execute the function
 311          */
 312         atomic_inc(&call_data->started);
 313         /*
 314          * At this point the info structure may be out of scope unless wait==1
 315          */
 316         (*func)(info);
 317         if (wait) {
 318                 smp_mb__before_atomic_inc();
 319                 atomic_inc(&call_data->finished);
 320         }
 321 }
 322
 323 extern struct gettimeofday_struct do_gtod;
 324
 325 struct thread_info *current_set[NR_CPUS];
 326
 327 DECLARE_PER_CPU(unsigned int, pvr);
 328
 329 static void __devinit smp_store_cpu_info(int id)
 330 {
 331         per_cpu(pvr, id) = mfspr(SPRN_PVR);
 332 }
 333
 334 static void __init smp_create_idle(unsigned int cpu)
 335 {
 336         struct task_struct *p;
 337
 338         /* create a process for the processor */
 339         p = fork_idle(cpu);
 340         if (IS_ERR(p))
 341                 panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p));
 342 #ifdef CONFIG_PPC64
 343         paca[cpu].__current = p;
 344 #endif
 345         current_set[cpu] = task_thread_info(p);
 346         task_thread_info(p)->cpu = cpu;
 347 }
 348
 349 void __init smp_prepare_cpus(unsigned int max_cpus)
 350 {
 351         unsigned int cpu;
 352
 353         DBG("smp_prepare_cpus\n");
 354
 355         /*
 356          * setup_cpu may need to be called on the boot cpu. We havent
 357          * spun any cpus up but lets be paranoid.
 358          */
 359         BUG_ON(boot_cpuid != smp_processor_id());
 360
 361         /* Fixup boot cpu */
 362         smp_store_cpu_info(boot_cpuid);
 363         cpu_callin_map[boot_cpuid] = 1;
 364
 365         if (smp_ops)
 366                 max_cpus = smp_ops->probe();
 367         else
 368                 max_cpus = 1;
 369
 370         smp_space_timers(max_cpus);
 371
 372         for_each_possible_cpu(cpu)
 373                 if (cpu != boot_cpuid)
 374                         smp_create_idle(cpu);
 375 }
 376
 377 void __devinit smp_prepare_boot_cpu(void)
 378 {
 379         BUG_ON(smp_processor_id() != boot_cpuid);
 380
 381         cpu_set(boot_cpuid, cpu_online_map);
 382 #ifdef CONFIG_PPC64
 383         paca[boot_cpuid].__current = current;
 384 #endif
 385         current_set[boot_cpuid] = task_thread_info(current);
 386 }
 387
 388 #ifdef CONFIG_HOTPLUG_CPU
 389 /* State of each CPU during hotplug phases */
 390 DEFINE_PER_CPU(int, cpu_state) = { 0 };
 391
 392 int generic_cpu_disable(void)
 393 {
 394         unsigned int cpu = smp_processor_id();
 395
 396         if (cpu == boot_cpuid)
 397                 return -EBUSY;
 398
 399         cpu_clear(cpu, cpu_online_map);
 400 #ifdef CONFIG_PPC64
 401         vdso_data->processorCount--;
 402         fixup_irqs(cpu_online_map);
 403 #endif
 404         return 0;
 405 }
 406
 407 int generic_cpu_enable(unsigned int cpu)
 408 {
 409         /* Do the normal bootup if we haven't
 410          * already bootstrapped. */
 411         if (system_state != SYSTEM_RUNNING)
 412                 return -ENOSYS;
 413
 414         /* get the target out of it's holding state */
 415         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 416         smp_wmb();
 417
 418         while (!cpu_online(cpu))
 419                 cpu_relax();
 420
 421 #ifdef CONFIG_PPC64
 422         fixup_irqs(cpu_online_map);
 423         /* counter the irq disable in fixup_irqs */
 424         local_irq_enable();
 425 #endif
 426         return 0;
 427 }
 428
 429 void generic_cpu_die(unsigned int cpu)
 430 {
 431         int i;
 432
 433         for (i = 0; i < 100; i++) {
 434                 smp_rmb();
 435                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
 436                         return;
 437                 msleep(100);
 438         }
 439         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
 440 }
 441
 442 void generic_mach_cpu_die(void)
 443 {
 444         unsigned int cpu;
 445
 446         local_irq_disable();
 447         cpu = smp_processor_id();
 448         printk(KERN_DEBUG "CPU%d offline\n", cpu);
 449         __get_cpu_var(cpu_state) = CPU_DEAD;
 450         smp_wmb();
 451         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
 452                 cpu_relax();
 453
 454 #ifdef CONFIG_PPC64
 455         flush_tlb_pending();
 456 #endif
 457         cpu_set(cpu, cpu_online_map);
 458         local_irq_enable();
 459 }
 460 #endif
 461
 462 static int __devinit cpu_enable(unsigned int cpu)
 463 {
 464         if (smp_ops && smp_ops->cpu_enable)
 465                 return smp_ops->cpu_enable(cpu);
 466
 467         return -ENOSYS;
 468 }
 469
 470 int __devinit __cpu_up(unsigned int cpu)
 471 {
 472         int c;
 473
 474         secondary_ti = current_set[cpu];
 475         if (!cpu_enable(cpu))
 476                 return 0;
 477
 478         if (smp_ops == NULL ||
 479             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
 480                 return -EINVAL;
 481
 482         /* Make sure callin-map entry is 0 (can be leftover a CPU
 483          * hotplug
 484          */
 485         cpu_callin_map[cpu] = 0;
 486
 487         /* The information for processor bringup must
 488          * be written out to main store before we release
 489          * the processor.
 490          */
 491         smp_mb();
 492
 493         /* wake up cpus */
 494         DBG("smp: kicking cpu %d\n", cpu);
 495         smp_ops->kick_cpu(cpu);
 496
 497         /*
 498          * wait to see if the cpu made a callin (is actually up).
 499          * use this value that I found through experimentation.
 500          * -- Cort
 501          */
 502         if (system_state < SYSTEM_RUNNING)
 503                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
 504                         udelay(100);
 505 #ifdef CONFIG_HOTPLUG_CPU
 506         else
 507                 /*
 508                  * CPUs can take much longer to come up in the
 509                  * hotplug case.  Wait five seconds.
 510                  */
 511                 for (c = 25; c && !cpu_callin_map[cpu]; c--) {
 512                         msleep(200);
 513                 }
 514 #endif
 515
 516         if (!cpu_callin_map[cpu]) {
 517                 printk("Processor %u is stuck.\n", cpu);
 518                 return -ENOENT;
 519         }
 520
 521         printk("Processor %u found.\n", cpu);
 522
 523         if (smp_ops->give_timebase)
 524                 smp_ops->give_timebase();
 525
 526         /* Wait until cpu puts itself in the online map */
 527         while (!cpu_online(cpu))
 528                 cpu_relax();
 529
 530         return 0;
 531 }
 532
 533
 534 /* Activate a secondary processor. */
 535 int __devinit start_secondary(void *unused)
 536 {
 537         unsigned int cpu = smp_processor_id();
 538
 539         atomic_inc(&init_mm.mm_count);
 540         current->active_mm = &init_mm;
 541
 542         smp_store_cpu_info(cpu);
 543         set_dec(tb_ticks_per_jiffy);
 544         preempt_disable();
 545         cpu_callin_map[cpu] = 1;
 546
 547         smp_ops->setup_cpu(cpu);
 548         if (smp_ops->take_timebase)
 549                 smp_ops->take_timebase();
 550
 551         if (system_state > SYSTEM_BOOTING)
 552                 snapshot_timebase();
 553
 554         spin_lock(&call_lock);
 555         cpu_set(cpu, cpu_online_map);
 556         spin_unlock(&call_lock);
 557
 558         local_irq_enable();
 559
 560         cpu_idle();
 561         return 0;
 562 }
 563
 564 int setup_profiling_timer(unsigned int multiplier)
 565 {
 566         return 0;
 567 }
 568
 569 void __init smp_cpus_done(unsigned int max_cpus)
 570 {
 571         cpumask_t old_mask;
 572
 573         /* We want the setup_cpu() here to be called from CPU 0, but our
 574          * init thread may have been "borrowed" by another CPU in the meantime
 575          * se we pin us down to CPU 0 for a short while
 576          */
 577         old_mask = current->cpus_allowed;
 578         set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid));
 579
 580         if (smp_ops)
 581                 smp_ops->setup_cpu(boot_cpuid);
 582
 583         set_cpus_allowed(current, old_mask);
 584
 585         snapshot_timebases();
 586
 587         dump_numa_cpu_topology();
 588 }
 589
 590 #ifdef CONFIG_HOTPLUG_CPU
 591 int __cpu_disable(void)
 592 {
 593         if (smp_ops->cpu_disable)
 594                 return smp_ops->cpu_disable();
 595
 596         return -ENOSYS;
 597 }
 598
 599 void __cpu_die(unsigned int cpu)
 600 {
 601         if (smp_ops->cpu_die)
 602                 smp_ops->cpu_die(cpu);
 603 }
 604 #endif