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