release/src-rt-6.x/linux/linux-2.6/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 = -EBUSY;
 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         put_cpu();
 296         return ret;
 297 }
 298 EXPORT_SYMBOL(smp_call_function_single);
 299
 300 void smp_call_function_interrupt(void)
 301 {
 302         void (*func) (void *info);
 303         void *info;
 304         int wait;
 305
 306         /* call_data will be NULL if the sender timed out while
 307          * waiting on us to receive the call.
 308          */
 309         if (!call_data)
 310                 return;
 311
 312         func = call_data->func;
 313         info = call_data->info;
 314         wait = call_data->wait;
 315
 316         if (!wait)
 317                 smp_mb__before_atomic_inc();
 318
 319         /*
 320          * Notify initiating CPU that I've grabbed the data and am
 321          * about to execute the function
 322          */
 323         atomic_inc(&call_data->started);
 324         /*
 325          * At this point the info structure may be out of scope unless wait==1
 326          */
 327         (*func)(info);
 328         if (wait) {
 329                 smp_mb__before_atomic_inc();
 330                 atomic_inc(&call_data->finished);
 331         }
 332 }
 333
 334 extern struct gettimeofday_struct do_gtod;
 335
 336 struct thread_info *current_set[NR_CPUS];
 337
 338 DECLARE_PER_CPU(unsigned int, pvr);
 339
 340 static void __devinit smp_store_cpu_info(int id)
 341 {
 342         per_cpu(pvr, id) = mfspr(SPRN_PVR);
 343 }
 344
 345 static void __init smp_create_idle(unsigned int cpu)
 346 {
 347         struct task_struct *p;
 348
 349         /* create a process for the processor */
 350         p = fork_idle(cpu);
 351         if (IS_ERR(p))
 352                 panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p));
 353 #ifdef CONFIG_PPC64
 354         paca[cpu].__current = p;
 355 #endif
 356         current_set[cpu] = task_thread_info(p);
 357         task_thread_info(p)->cpu = cpu;
 358 }
 359
 360 void __init smp_prepare_cpus(unsigned int max_cpus)
 361 {
 362         unsigned int cpu;
 363
 364         DBG("smp_prepare_cpus\n");
 365
 366         /*
 367          * setup_cpu may need to be called on the boot cpu. We havent
 368          * spun any cpus up but lets be paranoid.
 369          */
 370         BUG_ON(boot_cpuid != smp_processor_id());
 371
 372         /* Fixup boot cpu */
 373         smp_store_cpu_info(boot_cpuid);
 374         cpu_callin_map[boot_cpuid] = 1;
 375
 376         if (smp_ops)
 377                 max_cpus = smp_ops->probe();
 378         else
 379                 max_cpus = 1;
 380
 381         smp_space_timers(max_cpus);
 382
 383         for_each_possible_cpu(cpu)
 384                 if (cpu != boot_cpuid)
 385                         smp_create_idle(cpu);
 386 }
 387
 388 void __devinit smp_prepare_boot_cpu(void)
 389 {
 390         BUG_ON(smp_processor_id() != boot_cpuid);
 391
 392         cpu_set(boot_cpuid, cpu_online_map);
 393 #ifdef CONFIG_PPC64
 394         paca[boot_cpuid].__current = current;
 395 #endif
 396         current_set[boot_cpuid] = task_thread_info(current);
 397 }
 398
 399 #ifdef CONFIG_HOTPLUG_CPU
 400 /* State of each CPU during hotplug phases */
 401 DEFINE_PER_CPU(int, cpu_state) = { 0 };
 402
 403 int generic_cpu_disable(void)
 404 {
 405         unsigned int cpu = smp_processor_id();
 406
 407         if (cpu == boot_cpuid)
 408                 return -EBUSY;
 409
 410         cpu_clear(cpu, cpu_online_map);
 411 #ifdef CONFIG_PPC64
 412         vdso_data->processorCount--;
 413         fixup_irqs(cpu_online_map);
 414 #endif
 415         return 0;
 416 }
 417
 418 int generic_cpu_enable(unsigned int cpu)
 419 {
 420         /* Do the normal bootup if we haven't
 421          * already bootstrapped. */
 422         if (system_state != SYSTEM_RUNNING)
 423                 return -ENOSYS;
 424
 425         /* get the target out of it's holding state */
 426         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 427         smp_wmb();
 428
 429         while (!cpu_online(cpu))
 430                 cpu_relax();
 431
 432 #ifdef CONFIG_PPC64
 433         fixup_irqs(cpu_online_map);
 434         /* counter the irq disable in fixup_irqs */
 435         local_irq_enable();
 436 #endif
 437         return 0;
 438 }
 439
 440 void generic_cpu_die(unsigned int cpu)
 441 {
 442         int i;
 443
 444         for (i = 0; i < 100; i++) {
 445                 smp_rmb();
 446                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
 447                         return;
 448                 msleep(100);
 449         }
 450         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
 451 }
 452
 453 void generic_mach_cpu_die(void)
 454 {
 455         unsigned int cpu;
 456
 457         local_irq_disable();
 458         cpu = smp_processor_id();
 459         printk(KERN_DEBUG "CPU%d offline\n", cpu);
 460         __get_cpu_var(cpu_state) = CPU_DEAD;
 461         smp_wmb();
 462         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
 463                 cpu_relax();
 464         cpu_set(cpu, cpu_online_map);
 465         local_irq_enable();
 466 }
 467 #endif
 468
 469 static int __devinit cpu_enable(unsigned int cpu)
 470 {
 471         if (smp_ops && smp_ops->cpu_enable)
 472                 return smp_ops->cpu_enable(cpu);
 473
 474         return -ENOSYS;
 475 }
 476
 477 int __cpuinit __cpu_up(unsigned int cpu)
 478 {
 479         int c;
 480
 481         secondary_ti = current_set[cpu];
 482         if (!cpu_enable(cpu))
 483                 return 0;
 484
 485         if (smp_ops == NULL ||
 486             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
 487                 return -EINVAL;
 488
 489         /* Make sure callin-map entry is 0 (can be leftover a CPU
 490          * hotplug
 491          */
 492         cpu_callin_map[cpu] = 0;
 493
 494         /* The information for processor bringup must
 495          * be written out to main store before we release
 496          * the processor.
 497          */
 498         smp_mb();
 499
 500         /* wake up cpus */
 501         DBG("smp: kicking cpu %d\n", cpu);
 502         smp_ops->kick_cpu(cpu);
 503
 504         /*
 505          * wait to see if the cpu made a callin (is actually up).
 506          * use this value that I found through experimentation.
 507          * -- Cort
 508          */
 509         if (system_state < SYSTEM_RUNNING)
 510                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
 511                         udelay(100);
 512 #ifdef CONFIG_HOTPLUG_CPU
 513         else
 514                 /*
 515                  * CPUs can take much longer to come up in the
 516                  * hotplug case.  Wait five seconds.
 517                  */
 518                 for (c = 25; c && !cpu_callin_map[cpu]; c--) {
 519                         msleep(200);
 520                 }
 521 #endif
 522
 523         if (!cpu_callin_map[cpu]) {
 524                 printk("Processor %u is stuck.\n", cpu);
 525                 return -ENOENT;
 526         }
 527
 528         printk("Processor %u found.\n", cpu);
 529
 530         if (smp_ops->give_timebase)
 531                 smp_ops->give_timebase();
 532
 533         /* Wait until cpu puts itself in the online map */
 534         while (!cpu_online(cpu))
 535                 cpu_relax();
 536
 537         return 0;
 538 }
 539
 540
 541 /* Activate a secondary processor. */
 542 int __devinit start_secondary(void *unused)
 543 {
 544         unsigned int cpu = smp_processor_id();
 545
 546         atomic_inc(&init_mm.mm_count);
 547         current->active_mm = &init_mm;
 548
 549         smp_store_cpu_info(cpu);
 550         set_dec(tb_ticks_per_jiffy);
 551         preempt_disable();
 552         cpu_callin_map[cpu] = 1;
 553
 554         smp_ops->setup_cpu(cpu);
 555         if (smp_ops->take_timebase)
 556                 smp_ops->take_timebase();
 557
 558         if (system_state > SYSTEM_BOOTING)
 559                 snapshot_timebase();
 560
 561         spin_lock(&call_lock);
 562         cpu_set(cpu, cpu_online_map);
 563         spin_unlock(&call_lock);
 564
 565         local_irq_enable();
 566
 567         cpu_idle();
 568         return 0;
 569 }
 570
 571 int setup_profiling_timer(unsigned int multiplier)
 572 {
 573         return 0;
 574 }
 575
 576 void __init smp_cpus_done(unsigned int max_cpus)
 577 {
 578         cpumask_t old_mask;
 579
 580         /* We want the setup_cpu() here to be called from CPU 0, but our
 581          * init thread may have been "borrowed" by another CPU in the meantime
 582          * se we pin us down to CPU 0 for a short while
 583          */
 584         old_mask = current->cpus_allowed;
 585         set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid));
 586
 587         if (smp_ops)
 588                 smp_ops->setup_cpu(boot_cpuid);
 589
 590         set_cpus_allowed(current, old_mask);
 591
 592         snapshot_timebases();
 593
 594         dump_numa_cpu_topology();
 595 }
 596
 597 #ifdef CONFIG_HOTPLUG_CPU
 598 int __cpu_disable(void)
 599 {
 600         if (smp_ops->cpu_disable)
 601                 return smp_ops->cpu_disable();
 602
 603         return -ENOSYS;
 604 }
 605
 606 void __cpu_die(unsigned int cpu)
 607 {
 608         if (smp_ops->cpu_die)
 609                 smp_ops->cpu_die(cpu);
 610 }
 611 #endif