arch/ppc64/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/config.h>
  21 #include <linux/kernel.h>
  22 #include <linux/module.h>
  23 #include <linux/sched.h>
  24 #include <linux/smp.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/delay.h>
  27 #include <linux/init.h>
  28 #include <linux/spinlock.h>
  29 #include <linux/cache.h>
  30 #include <linux/err.h>
  31 #include <linux/sysdev.h>
  32 #include <linux/cpu.h>
  33 #include <linux/notifier.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/paca.h>
  43 #include <asm/time.h>
  44 #include <asm/machdep.h>
  45 #include <asm/cputable.h>
  46 #include <asm/system.h>
  47 #include <asm/abs_addr.h>
  48
  49 #include "mpic.h"
  50
  51 #ifdef DEBUG
  52 #define DBG(fmt...) udbg_printf(fmt)
  53 #else
  54 #define DBG(fmt...)
  55 #endif
  56
  57 cpumask_t cpu_possible_map = CPU_MASK_NONE;
  58 cpumask_t cpu_online_map = CPU_MASK_NONE;
  59 cpumask_t cpu_sibling_map[NR_CPUS] = { [0 ... NR_CPUS-1] = CPU_MASK_NONE };
  60
  61 EXPORT_SYMBOL(cpu_online_map);
  62 EXPORT_SYMBOL(cpu_possible_map);
  63
  64 struct smp_ops_t *smp_ops;
  65
  66 static volatile unsigned int cpu_callin_map[NR_CPUS];
  67
  68 extern unsigned char stab_array[];
  69
  70 void smp_call_function_interrupt(void);
  71
  72 int smt_enabled_at_boot = 1;
  73
  74 #ifdef CONFIG_PPC_MULTIPLATFORM
  75 void smp_mpic_message_pass(int target, int msg)
  76 {
  77         /* make sure we're sending something that translates to an IPI */
  78         if ( msg > 0x3 ){
  79                 printk("SMP %d: smp_message_pass: unknown msg %d\n",
  80                        smp_processor_id(), msg);
  81                 return;
  82         }
  83         switch ( target )
  84         {
  85         case MSG_ALL:
  86                 mpic_send_ipi(msg, 0xffffffff);
  87                 break;
  88         case MSG_ALL_BUT_SELF:
  89                 mpic_send_ipi(msg, 0xffffffff & ~(1 << smp_processor_id()));
  90                 break;
  91         default:
  92                 mpic_send_ipi(msg, 1 << target);
  93                 break;
  94         }
  95 }
  96
  97 int __init smp_mpic_probe(void)
  98 {
  99         int nr_cpus;
 100
 101         DBG("smp_mpic_probe()...\n");
 102
 103         nr_cpus = cpus_weight(cpu_possible_map);
 104
 105         DBG("nr_cpus: %d\n", nr_cpus);
 106
 107         if (nr_cpus > 1)
 108                 mpic_request_ipis();
 109
 110         return nr_cpus;
 111 }
 112
 113 void __devinit smp_mpic_setup_cpu(int cpu)
 114 {
 115         mpic_setup_this_cpu();
 116 }
 117
 118 void __devinit smp_generic_kick_cpu(int nr)
 119 {
 120         BUG_ON(nr < 0 || nr >= NR_CPUS);
 121
 122         /*
 123          * The processor is currently spinning, waiting for the
 124          * cpu_start field to become non-zero After we set cpu_start,
 125          * the processor will continue on to secondary_start
 126          */
 127         paca[nr].cpu_start = 1;
 128         mb();
 129 }
 130
 131 #endif /* CONFIG_PPC_MULTIPLATFORM */
 132
 133 static void __init smp_space_timers(unsigned int max_cpus)
 134 {
 135         int i;
 136         unsigned long offset = tb_ticks_per_jiffy / max_cpus;
 137         unsigned long previous_tb = paca[boot_cpuid].next_jiffy_update_tb;
 138
 139         for_each_cpu(i) {
 140                 if (i != boot_cpuid) {
 141                         paca[i].next_jiffy_update_tb =
 142                                 previous_tb + offset;
 143                         previous_tb = paca[i].next_jiffy_update_tb;
 144                 }
 145         }
 146 }
 147
 148 void smp_message_recv(int msg, struct pt_regs *regs)
 149 {
 150         switch(msg) {
 151         case PPC_MSG_CALL_FUNCTION:
 152                 smp_call_function_interrupt();
 153                 break;
 154         case PPC_MSG_RESCHEDULE:
 155                 /* XXX Do we have to do this? */
 156                 set_need_resched();
 157                 break;
 158 #if 0
 159         case PPC_MSG_MIGRATE_TASK:
 160                 /* spare */
 161                 break;
 162 #endif
 163 #ifdef CONFIG_DEBUGGER
 164         case PPC_MSG_DEBUGGER_BREAK:
 165                 debugger_ipi(regs);
 166                 break;
 167 #endif
 168         default:
 169                 printk("SMP %d: smp_message_recv(): unknown msg %d\n",
 170                        smp_processor_id(), msg);
 171                 break;
 172         }
 173 }
 174
 175 void smp_send_reschedule(int cpu)
 176 {
 177         smp_ops->message_pass(cpu, PPC_MSG_RESCHEDULE);
 178 }
 179
 180 #ifdef CONFIG_DEBUGGER
 181 void smp_send_debugger_break(int cpu)
 182 {
 183         smp_ops->message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
 184 }
 185 #endif
 186
 187 static void stop_this_cpu(void *dummy)
 188 {
 189         local_irq_disable();
 190         while (1)
 191                 ;
 192 }
 193
 194 void smp_send_stop(void)
 195 {
 196         smp_call_function(stop_this_cpu, NULL, 1, 0);
 197 }
 198
 199 /*
 200  * Structure and data for smp_call_function(). This is designed to minimise
 201  * static memory requirements. It also looks cleaner.
 202  * Stolen from the i386 version.
 203  */
 204 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(call_lock);
 205
 206 static struct call_data_struct {
 207         void (*func) (void *info);
 208         void *info;
 209         atomic_t started;
 210         atomic_t finished;
 211         int wait;
 212 } *call_data;
 213
 214 /* delay of at least 8 seconds on 1GHz cpu */
 215 #define SMP_CALL_TIMEOUT (1UL << (30 + 3))
 216
 217 /*
 218  * This function sends a 'generic call function' IPI to all other CPUs
 219  * in the system.
 220  *
 221  * [SUMMARY] Run a function on all other CPUs.
 222  * <func> The function to run. This must be fast and non-blocking.
 223  * <info> An arbitrary pointer to pass to the function.
 224  * <nonatomic> currently unused.
 225  * <wait> If true, wait (atomically) until function has completed on other CPUs.
 226  * [RETURNS] 0 on success, else a negative status code. Does not return until
 227  * remote CPUs are nearly ready to execute <<func>> or are or have executed.
 228  *
 229  * You must not call this function with disabled interrupts or from a
 230  * hardware interrupt handler or from a bottom half handler.
 231  */
 232 int smp_call_function (void (*func) (void *info), void *info, int nonatomic,
 233                        int wait)
 234 {
 235         struct call_data_struct data;
 236         int ret = -1, cpus;
 237         unsigned long timeout;
 238
 239         /* Can deadlock when called with interrupts disabled */
 240         WARN_ON(irqs_disabled());
 241
 242         data.func = func;
 243         data.info = info;
 244         atomic_set(&data.started, 0);
 245         data.wait = wait;
 246         if (wait)
 247                 atomic_set(&data.finished, 0);
 248
 249         spin_lock(&call_lock);
 250         /* Must grab online cpu count with preempt disabled, otherwise
 251          * it can change. */
 252         cpus = num_online_cpus() - 1;
 253         if (!cpus) {
 254                 ret = 0;
 255                 goto out;
 256         }
 257
 258         call_data = &data;
 259         wmb();
 260         /* Send a message to all other CPUs and wait for them to respond */
 261         smp_ops->message_pass(MSG_ALL_BUT_SELF, PPC_MSG_CALL_FUNCTION);
 262
 263         /* Wait for response */
 264         timeout = SMP_CALL_TIMEOUT;
 265         while (atomic_read(&data.started) != cpus) {
 266                 HMT_low();
 267                 if (--timeout == 0) {
 268                         printk("smp_call_function on cpu %d: other cpus not "
 269                                "responding (%d)\n", smp_processor_id(),
 270                                atomic_read(&data.started));
 271                         debugger(NULL);
 272                         goto out;
 273                 }
 274         }
 275
 276         if (wait) {
 277                 timeout = SMP_CALL_TIMEOUT;
 278                 while (atomic_read(&data.finished) != cpus) {
 279                         HMT_low();
 280                         if (--timeout == 0) {
 281                                 printk("smp_call_function on cpu %d: other "
 282                                        "cpus not finishing (%d/%d)\n",
 283                                        smp_processor_id(),
 284                                        atomic_read(&data.finished),
 285                                        atomic_read(&data.started));
 286                                 debugger(NULL);
 287                                 goto out;
 288                         }
 289                 }
 290         }
 291
 292         ret = 0;
 293
 294 out:
 295         call_data = NULL;
 296         HMT_medium();
 297         spin_unlock(&call_lock);
 298         return ret;
 299 }
 300
 301 EXPORT_SYMBOL(smp_call_function);
 302
 303 void smp_call_function_interrupt(void)
 304 {
 305         void (*func) (void *info);
 306         void *info;
 307         int wait;
 308
 309         /* call_data will be NULL if the sender timed out while
 310          * waiting on us to receive the call.
 311          */
 312         if (!call_data)
 313                 return;
 314
 315         func = call_data->func;
 316         info = call_data->info;
 317         wait = call_data->wait;
 318
 319         if (!wait)
 320                 smp_mb__before_atomic_inc();
 321
 322         /*
 323          * Notify initiating CPU that I've grabbed the data and am
 324          * about to execute the function
 325          */
 326         atomic_inc(&call_data->started);
 327         /*
 328          * At this point the info structure may be out of scope unless wait==1
 329          */
 330         (*func)(info);
 331         if (wait) {
 332                 smp_mb__before_atomic_inc();
 333                 atomic_inc(&call_data->finished);
 334         }
 335 }
 336
 337 extern unsigned long decr_overclock;
 338 extern struct gettimeofday_struct do_gtod;
 339
 340 struct thread_info *current_set[NR_CPUS];
 341
 342 DECLARE_PER_CPU(unsigned int, pvr);
 343
 344 static void __devinit smp_store_cpu_info(int id)
 345 {
 346         per_cpu(pvr, id) = mfspr(SPRN_PVR);
 347 }
 348
 349 static void __init smp_create_idle(unsigned int cpu)
 350 {
 351         struct task_struct *p;
 352
 353         /* create a process for the processor */
 354         p = fork_idle(cpu);
 355         if (IS_ERR(p))
 356                 panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p));
 357         paca[cpu].__current = p;
 358         current_set[cpu] = p->thread_info;
 359 }
 360
 361 void __init smp_prepare_cpus(unsigned int max_cpus)
 362 {
 363         unsigned int cpu;
 364
 365         DBG("smp_prepare_cpus\n");
 366
 367         /*
 368          * setup_cpu may need to be called on the boot cpu. We havent
 369          * spun any cpus up but lets be paranoid.
 370          */
 371         BUG_ON(boot_cpuid != smp_processor_id());
 372
 373         /* Fixup boot cpu */
 374         smp_store_cpu_info(boot_cpuid);
 375         cpu_callin_map[boot_cpuid] = 1;
 376
 377 #ifndef CONFIG_PPC_ISERIES
 378         paca[boot_cpuid].next_jiffy_update_tb = tb_last_stamp = get_tb();
 379
 380         /*
 381          * Should update do_gtod.stamp_xsec.
 382          * For now we leave it which means the time can be some
 383          * number of msecs off until someone does a settimeofday()
 384          */
 385         do_gtod.varp->tb_orig_stamp = tb_last_stamp;
 386         systemcfg->tb_orig_stamp = tb_last_stamp;
 387 #endif
 388
 389         max_cpus = smp_ops->probe();
 390
 391         smp_space_timers(max_cpus);
 392
 393         for_each_cpu(cpu)
 394                 if (cpu != boot_cpuid)
 395                         smp_create_idle(cpu);
 396 }
 397
 398 void __devinit smp_prepare_boot_cpu(void)
 399 {
 400         BUG_ON(smp_processor_id() != boot_cpuid);
 401
 402         cpu_set(boot_cpuid, cpu_online_map);
 403
 404         paca[boot_cpuid].__current = current;
 405         current_set[boot_cpuid] = current->thread_info;
 406 }
 407
 408 #ifdef CONFIG_HOTPLUG_CPU
 409 /* State of each CPU during hotplug phases */
 410 DEFINE_PER_CPU(int, cpu_state) = { 0 };
 411
 412 int generic_cpu_disable(void)
 413 {
 414         unsigned int cpu = smp_processor_id();
 415
 416         if (cpu == boot_cpuid)
 417                 return -EBUSY;
 418
 419         systemcfg->processorCount--;
 420         cpu_clear(cpu, cpu_online_map);
 421         fixup_irqs(cpu_online_map);
 422         return 0;
 423 }
 424
 425 int generic_cpu_enable(unsigned int cpu)
 426 {
 427         /* Do the normal bootup if we haven't
 428          * already bootstrapped. */
 429         if (system_state != SYSTEM_RUNNING)
 430                 return -ENOSYS;
 431
 432         /* get the target out of it's holding state */
 433         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 434         wmb();
 435
 436         while (!cpu_online(cpu))
 437                 cpu_relax();
 438
 439         fixup_irqs(cpu_online_map);
 440         /* counter the irq disable in fixup_irqs */
 441         local_irq_enable();
 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                 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         wmb();
 467         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
 468                 cpu_relax();
 469
 470         flush_tlb_pending();
 471         cpu_set(cpu, cpu_online_map);
 472         local_irq_enable();
 473 }
 474 #endif
 475
 476 static int __devinit cpu_enable(unsigned int cpu)
 477 {
 478         if (smp_ops->cpu_enable)
 479                 return smp_ops->cpu_enable(cpu);
 480
 481         return -ENOSYS;
 482 }
 483
 484 int __devinit __cpu_up(unsigned int cpu)
 485 {
 486         int c;
 487
 488         if (!cpu_enable(cpu))
 489                 return 0;
 490
 491         if (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu))
 492                 return -EINVAL;
 493
 494         paca[cpu].default_decr = tb_ticks_per_jiffy / decr_overclock;
 495
 496         if (!cpu_has_feature(CPU_FTR_SLB)) {
 497                 void *tmp;
 498
 499                 /* maximum of 48 CPUs on machines with a segment table */
 500                 if (cpu >= 48)
 501                         BUG();
 502
 503                 tmp = &stab_array[PAGE_SIZE * cpu];
 504                 memset(tmp, 0, PAGE_SIZE);
 505                 paca[cpu].stab_addr = (unsigned long)tmp;
 506                 paca[cpu].stab_real = virt_to_abs(tmp);
 507         }
 508
 509         /* Make sure callin-map entry is 0 (can be leftover a CPU
 510          * hotplug
 511          */
 512         cpu_callin_map[cpu] = 0;
 513
 514         /* The information for processor bringup must
 515          * be written out to main store before we release
 516          * the processor.
 517          */
 518         mb();
 519
 520         /* wake up cpus */
 521         DBG("smp: kicking cpu %d\n", cpu);
 522         smp_ops->kick_cpu(cpu);
 523
 524         /*
 525          * wait to see if the cpu made a callin (is actually up).
 526          * use this value that I found through experimentation.
 527          * -- Cort
 528          */
 529         if (system_state < SYSTEM_RUNNING)
 530                 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
 531                         udelay(100);
 532 #ifdef CONFIG_HOTPLUG_CPU
 533         else
 534                 /*
 535                  * CPUs can take much longer to come up in the
 536                  * hotplug case.  Wait five seconds.
 537                  */
 538                 for (c = 25; c && !cpu_callin_map[cpu]; c--) {
 539                         msleep(200);
 540                 }
 541 #endif
 542
 543         if (!cpu_callin_map[cpu]) {
 544                 printk("Processor %u is stuck.\n", cpu);
 545                 return -ENOENT;
 546         }
 547
 548         printk("Processor %u found.\n", cpu);
 549
 550         if (smp_ops->give_timebase)
 551                 smp_ops->give_timebase();
 552
 553         /* Wait until cpu puts itself in the online map */
 554         while (!cpu_online(cpu))
 555                 cpu_relax();
 556
 557         return 0;
 558 }
 559
 560
 561 /* Activate a secondary processor. */
 562 int __devinit start_secondary(void *unused)
 563 {
 564         unsigned int cpu = smp_processor_id();
 565
 566         atomic_inc(&init_mm.mm_count);
 567         current->active_mm = &init_mm;
 568
 569         smp_store_cpu_info(cpu);
 570         set_dec(paca[cpu].default_decr);
 571         cpu_callin_map[cpu] = 1;
 572
 573         smp_ops->setup_cpu(cpu);
 574         if (smp_ops->take_timebase)
 575                 smp_ops->take_timebase();
 576
 577         spin_lock(&call_lock);
 578         cpu_set(cpu, cpu_online_map);
 579         spin_unlock(&call_lock);
 580
 581         local_irq_enable();
 582
 583         cpu_idle();
 584         return 0;
 585 }
 586
 587 int setup_profiling_timer(unsigned int multiplier)
 588 {
 589         return 0;
 590 }
 591
 592 void __init smp_cpus_done(unsigned int max_cpus)
 593 {
 594         cpumask_t old_mask;
 595
 596         /* We want the setup_cpu() here to be called from CPU 0, but our
 597          * init thread may have been "borrowed" by another CPU in the meantime
 598          * se we pin us down to CPU 0 for a short while
 599          */
 600         old_mask = current->cpus_allowed;
 601         set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid));
 602
 603         smp_ops->setup_cpu(boot_cpuid);
 604
 605         set_cpus_allowed(current, old_mask);
 606 }
 607
 608 #ifdef CONFIG_HOTPLUG_CPU
 609 int __cpu_disable(void)
 610 {
 611         if (smp_ops->cpu_disable)
 612                 return smp_ops->cpu_disable();
 613
 614         return -ENOSYS;
 615 }
 616
 617 void __cpu_die(unsigned int cpu)
 618 {
 619         if (smp_ops->cpu_die)
 620                 smp_ops->cpu_die(cpu);
 621 }
 622 #endif