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/cputhreads.h>
  45 #include <asm/cputable.h>
  46 #include <asm/system.h>
  47 #include <asm/mpic.h>
  48 #include <asm/vdso_datapage.h>
  49 #ifdef CONFIG_PPC64
  50 #include <asm/paca.h>
  51 #endif
  52
  53 #ifdef DEBUG
  54 #include <asm/udbg.h>
  55 #define DBG(fmt...) udbg_printf(fmt)
  56 #else
  57 #define DBG(fmt...)
  58 #endif
  59
  60 struct thread_info *secondary_ti;
  61
  62 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
  63 DEFINE_PER_CPU(cpumask_t, cpu_core_map) = CPU_MASK_NONE;
  64
  65 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
  66 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
  67
  68 /* SMP operations for this machine */
  69 struct smp_ops_t *smp_ops;
  70
  71 /* Can't be static due to PowerMac hackery */
  72 volatile unsigned int cpu_callin_map[NR_CPUS];
  73
  74 int smt_enabled_at_boot = 1;
  75
  76 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
  77
  78 #ifdef CONFIG_PPC64
  79 void __devinit smp_generic_kick_cpu(int nr)
  80 {
  81         BUG_ON(nr < 0 || nr >= NR_CPUS);
  82
  83         /*
  84          * The processor is currently spinning, waiting for the
  85          * cpu_start field to become non-zero After we set cpu_start,
  86          * the processor will continue on to secondary_start
  87          */
  88         paca[nr].cpu_start = 1;
  89         smp_mb();
  90 }
  91 #endif
  92
  93 void smp_message_recv(int msg)
  94 {
  95         switch(msg) {
  96         case PPC_MSG_CALL_FUNCTION:
  97                 generic_smp_call_function_interrupt();
  98                 break;
  99         case PPC_MSG_RESCHEDULE:
 100                 /* we notice need_resched on exit */
 101                 break;
 102         case PPC_MSG_CALL_FUNC_SINGLE:
 103                 generic_smp_call_function_single_interrupt();
 104                 break;
 105         case PPC_MSG_DEBUGGER_BREAK:
 106                 if (crash_ipi_function_ptr) {
 107                         crash_ipi_function_ptr(get_irq_regs());
 108                         break;
 109                 }
 110 #ifdef CONFIG_DEBUGGER
 111                 debugger_ipi(get_irq_regs());
 112                 break;
 113 #endif /* CONFIG_DEBUGGER */
 114                 /* FALLTHROUGH */
 115         default:
 116                 printk("SMP %d: smp_message_recv(): unknown msg %d\n",
 117                        smp_processor_id(), msg);
 118                 break;
 119         }
 120 }
 121
 122 static irqreturn_t call_function_action(int irq, void *data)
 123 {
 124         generic_smp_call_function_interrupt();
 125         return IRQ_HANDLED;
 126 }
 127
 128 static irqreturn_t reschedule_action(int irq, void *data)
 129 {
 130         /* we just need the return path side effect of checking need_resched */
 131         return IRQ_HANDLED;
 132 }
 133
 134 static irqreturn_t call_function_single_action(int irq, void *data)
 135 {
 136         generic_smp_call_function_single_interrupt();
 137         return IRQ_HANDLED;
 138 }
 139
 140 static irqreturn_t debug_ipi_action(int irq, void *data)
 141 {
 142         smp_message_recv(PPC_MSG_DEBUGGER_BREAK);
 143         return IRQ_HANDLED;
 144 }
 145
 146 static irq_handler_t smp_ipi_action[] = {
 147         [PPC_MSG_CALL_FUNCTION] =  call_function_action,
 148         [PPC_MSG_RESCHEDULE] = reschedule_action,
 149         [PPC_MSG_CALL_FUNC_SINGLE] = call_function_single_action,
 150         [PPC_MSG_DEBUGGER_BREAK] = debug_ipi_action,
 151 };
 152
 153 const char *smp_ipi_name[] = {
 154         [PPC_MSG_CALL_FUNCTION] =  "ipi call function",
 155         [PPC_MSG_RESCHEDULE] = "ipi reschedule",
 156         [PPC_MSG_CALL_FUNC_SINGLE] = "ipi call function single",
 157         [PPC_MSG_DEBUGGER_BREAK] = "ipi debugger",
 158 };
 159
 160 /* optional function to request ipi, for controllers with >= 4 ipis */
 161 int smp_request_message_ipi(int virq, int msg)
 162 {
 163         int err;
 164
 165         if (msg < 0 || msg > PPC_MSG_DEBUGGER_BREAK) {
 166                 return -EINVAL;
 167         }
 168 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
 169         if (msg == PPC_MSG_DEBUGGER_BREAK) {
 170                 return 1;
 171         }
 172 #endif
 173         err = request_irq(virq, smp_ipi_action[msg], IRQF_DISABLED|IRQF_PERCPU,
 174                           smp_ipi_name[msg], 0);
 175         WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
 176                 virq, smp_ipi_name[msg], err);
 177
 178         return err;
 179 }
 180
 181 void smp_send_reschedule(int cpu)
 182 {
 183         if (likely(smp_ops))
 184                 smp_ops->message_pass(cpu, PPC_MSG_RESCHEDULE);
 185 }
 186
 187 void arch_send_call_function_single_ipi(int cpu)
 188 {
 189         smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
 190 }
 191
 192 void arch_send_call_function_ipi(cpumask_t mask)
 193 {
 194         unsigned int cpu;
 195
 196         for_each_cpu_mask(cpu, mask)
 197                 smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNCTION);
 198 }
 199
 200 #ifdef CONFIG_DEBUGGER
 201 void smp_send_debugger_break(int cpu)
 202 {
 203         if (likely(smp_ops))
 204                 smp_ops->message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
 205 }
 206 #endif
 207
 208 #ifdef CONFIG_KEXEC
 209 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
 210 {
 211         crash_ipi_function_ptr = crash_ipi_callback;
 212         if (crash_ipi_callback && smp_ops) {
 213                 mb();
 214                 smp_ops->message_pass(MSG_ALL_BUT_SELF, PPC_MSG_DEBUGGER_BREAK);
 215         }
 216 }
 217 #endif
 218
 219 static void stop_this_cpu(void *dummy)
 220 {
 221         local_irq_disable();
 222         while (1)
 223                 ;
 224 }
 225
 226 void smp_send_stop(void)
 227 {
 228         smp_call_function(stop_this_cpu, NULL, 0);
 229 }
 230
 231 struct thread_info *current_set[NR_CPUS];
 232
 233 static void __devinit smp_store_cpu_info(int id)
 234 {
 235         per_cpu(pvr, id) = mfspr(SPRN_PVR);
 236 }
 237
 238 static void __init smp_create_idle(unsigned int cpu)
 239 {
 240         struct task_struct *p;
 241
 242         /* create a process for the processor */
 243         p = fork_idle(cpu);
 244         if (IS_ERR(p))
 245                 panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p));
 246 #ifdef CONFIG_PPC64
 247         paca[cpu].__current = p;
 248         paca[cpu].kstack = (unsigned long) task_thread_info(p)
 249                 + THREAD_SIZE - STACK_FRAME_OVERHEAD;
 250 #endif
 251         current_set[cpu] = task_thread_info(p);
 252         task_thread_info(p)->cpu = cpu;
 253 }
 254
 255 void __init smp_prepare_cpus(unsigned int max_cpus)
 256 {
 257         unsigned int cpu;
 258
 259         DBG("smp_prepare_cpus\n");
 260
 261         /*
 262          * setup_cpu may need to be called on the boot cpu. We havent
 263          * spun any cpus up but lets be paranoid.
 264          */
 265         BUG_ON(boot_cpuid != smp_processor_id());
 266
 267         /* Fixup boot cpu */
 268         smp_store_cpu_info(boot_cpuid);
 269         cpu_callin_map[boot_cpuid] = 1;
 270
 271         if (smp_ops)
 272                 max_cpus = smp_ops->probe();
 273         else
 274                 max_cpus = 1;
 275
 276         smp_space_timers(max_cpus);
 277
 278         for_each_possible_cpu(cpu)
 279                 if (cpu != boot_cpuid)
 280                         smp_create_idle(cpu);
 281 }
 282
 283 void __devinit smp_prepare_boot_cpu(void)
 284 {
 285         BUG_ON(smp_processor_id() != boot_cpuid);
 286
 287         cpu_set(boot_cpuid, cpu_online_map);
 288         cpu_set(boot_cpuid, per_cpu(cpu_sibling_map, boot_cpuid));
 289         cpu_set(boot_cpuid, per_cpu(cpu_core_map, boot_cpuid));
 290 #ifdef CONFIG_PPC64
 291         paca[boot_cpuid].__current = current;
 292 #endif
 293         current_set[boot_cpuid] = task_thread_info(current);
 294 }
 295
 296 #ifdef CONFIG_HOTPLUG_CPU
 297 /* State of each CPU during hotplug phases */
 298 DEFINE_PER_CPU(int, cpu_state) = { 0 };
 299
 300 int generic_cpu_disable(void)
 301 {
 302         unsigned int cpu = smp_processor_id();
 303
 304         if (cpu == boot_cpuid)
 305                 return -EBUSY;
 306
 307         cpu_clear(cpu, cpu_online_map);
 308 #ifdef CONFIG_PPC64
 309         vdso_data->processorCount--;
 310         fixup_irqs(cpu_online_map);
 311 #endif
 312         return 0;
 313 }
 314
 315 int generic_cpu_enable(unsigned int cpu)
 316 {
 317         /* Do the normal bootup if we haven't
 318          * already bootstrapped. */
 319         if (system_state != SYSTEM_RUNNING)
 320                 return -ENOSYS;
 321
 322         /* get the target out of it's holding state */
 323         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 324         smp_wmb();
 325
 326         while (!cpu_online(cpu))
 327                 cpu_relax();
 328
 329 #ifdef CONFIG_PPC64
 330         fixup_irqs(cpu_online_map);
 331         /* counter the irq disable in fixup_irqs */
 332         local_irq_enable();
 333 #endif
 334         return 0;
 335 }
 336
 337 void generic_cpu_die(unsigned int cpu)
 338 {
 339         int i;
 340
 341         for (i = 0; i < 100; i++) {
 342                 smp_rmb();
 343                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
 344                         return;
 345                 msleep(100);
 346         }
 347         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
 348 }
 349
 350 void generic_mach_cpu_die(void)
 351 {
 352         unsigned int cpu;
 353
 354         local_irq_disable();
 355         cpu = smp_processor_id();
 356         printk(KERN_DEBUG "CPU%d offline\n", cpu);
 357         __get_cpu_var(cpu_state) = CPU_DEAD;
 358         smp_wmb();
 359         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
 360                 cpu_relax();
 361         cpu_set(cpu, cpu_online_map);
 362         local_irq_enable();
 363 }
 364 #endif
 365
 366 static int __devinit cpu_enable(unsigned int cpu)
 367 {
 368         if (smp_ops && smp_ops->cpu_enable)
 369                 return smp_ops->cpu_enable(cpu);
 370
 371         return -ENOSYS;
 372 }
 373
 374 int __cpuinit __cpu_up(unsigned int cpu)
 375 {
 376         int c;
 377
 378         secondary_ti = current_set[cpu];
 379         if (!cpu_enable(cpu))
 380                 return 0;
 381
 382         if (smp_ops == NULL ||
 383             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
 384                 return -EINVAL;
 385
 386         /* Make sure callin-map entry is 0 (can be leftover a CPU
 387          * hotplug
 388          */
 389         cpu_callin_map[cpu] = 0;
 390
 391         /* The information for processor bringup must
 392          * be written out to main store before we release
 393          * the processor.
 394          */
 395         smp_mb();
 396
 397         /* wake up cpus */
 398         DBG("smp: kicking cpu %d\n", cpu);
 399         smp_ops->kick_cpu(cpu);
 400
 401         /*
 402          * wait to see if the cpu made a callin (is actually up).
 403          * use this value that I found through experimentation.
 404          * -- Cort
 405          */
 406         if (system_state < SYSTEM_RUNNING)
 407                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
 408                         udelay(100);
 409 #ifdef CONFIG_HOTPLUG_CPU
 410         else
 411                 /*
 412                  * CPUs can take much longer to come up in the
 413                  * hotplug case.  Wait five seconds.
 414                  */
 415                 for (c = 25; c && !cpu_callin_map[cpu]; c--) {
 416                         msleep(200);
 417                 }
 418 #endif
 419
 420         if (!cpu_callin_map[cpu]) {
 421                 printk("Processor %u is stuck.\n", cpu);
 422                 return -ENOENT;
 423         }
 424
 425         printk("Processor %u found.\n", cpu);
 426
 427         if (smp_ops->give_timebase)
 428                 smp_ops->give_timebase();
 429
 430         /* Wait until cpu puts itself in the online map */
 431         while (!cpu_online(cpu))
 432                 cpu_relax();
 433
 434         return 0;
 435 }
 436
 437 /* Return the value of the reg property corresponding to the given
 438  * logical cpu.
 439  */
 440 int cpu_to_core_id(int cpu)
 441 {
 442         struct device_node *np;
 443         const int *reg;
 444         int id = -1;
 445
 446         np = of_get_cpu_node(cpu, NULL);
 447         if (!np)
 448                 goto out;
 449
 450         reg = of_get_property(np, "reg", NULL);
 451         if (!reg)
 452                 goto out;
 453
 454         id = *reg;
 455 out:
 456         of_node_put(np);
 457         return id;
 458 }
 459
 460 /* Must be called when no change can occur to cpu_present_map,
 461  * i.e. during cpu online or offline.
 462  */
 463 static struct device_node *cpu_to_l2cache(int cpu)
 464 {
 465         struct device_node *np;
 466         struct device_node *cache;
 467
 468         if (!cpu_present(cpu))
 469                 return NULL;
 470
 471         np = of_get_cpu_node(cpu, NULL);
 472         if (np == NULL)
 473                 return NULL;
 474
 475         cache = of_find_next_cache_node(np);
 476
 477         of_node_put(np);
 478
 479         return cache;
 480 }
 481
 482 /* Activate a secondary processor. */
 483 int __devinit start_secondary(void *unused)
 484 {
 485         unsigned int cpu = smp_processor_id();
 486         struct device_node *l2_cache;
 487         int i, base;
 488
 489         atomic_inc(&init_mm.mm_count);
 490         current->active_mm = &init_mm;
 491
 492         smp_store_cpu_info(cpu);
 493         set_dec(tb_ticks_per_jiffy);
 494         preempt_disable();
 495         cpu_callin_map[cpu] = 1;
 496
 497         smp_ops->setup_cpu(cpu);
 498         if (smp_ops->take_timebase)
 499                 smp_ops->take_timebase();
 500
 501         if (system_state > SYSTEM_BOOTING)
 502                 snapshot_timebase();
 503
 504         secondary_cpu_time_init();
 505
 506         ipi_call_lock();
 507         notify_cpu_starting(cpu);
 508         cpu_set(cpu, cpu_online_map);
 509         /* Update sibling maps */
 510         base = cpu_first_thread_in_core(cpu);
 511         for (i = 0; i < threads_per_core; i++) {
 512                 if (cpu_is_offline(base + i))
 513                         continue;
 514                 cpu_set(cpu, per_cpu(cpu_sibling_map, base + i));
 515                 cpu_set(base + i, per_cpu(cpu_sibling_map, cpu));
 516
 517                 /* cpu_core_map should be a superset of
 518                  * cpu_sibling_map even if we don't have cache
 519                  * information, so update the former here, too.
 520                  */
 521                 cpu_set(cpu, per_cpu(cpu_core_map, base +i));
 522                 cpu_set(base + i, per_cpu(cpu_core_map, cpu));
 523         }
 524         l2_cache = cpu_to_l2cache(cpu);
 525         for_each_online_cpu(i) {
 526                 struct device_node *np = cpu_to_l2cache(i);
 527                 if (!np)
 528                         continue;
 529                 if (np == l2_cache) {
 530                         cpu_set(cpu, per_cpu(cpu_core_map, i));
 531                         cpu_set(i, per_cpu(cpu_core_map, cpu));
 532                 }
 533                 of_node_put(np);
 534         }
 535         of_node_put(l2_cache);
 536         ipi_call_unlock();
 537
 538         local_irq_enable();
 539
 540         cpu_idle();
 541         return 0;
 542 }
 543
 544 int setup_profiling_timer(unsigned int multiplier)
 545 {
 546         return 0;
 547 }
 548
 549 void __init smp_cpus_done(unsigned int max_cpus)
 550 {
 551         cpumask_t old_mask;
 552
 553         /* We want the setup_cpu() here to be called from CPU 0, but our
 554          * init thread may have been "borrowed" by another CPU in the meantime
 555          * se we pin us down to CPU 0 for a short while
 556          */
 557         old_mask = current->cpus_allowed;
 558         set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid));
 559
 560         if (smp_ops)
 561                 smp_ops->setup_cpu(boot_cpuid);
 562
 563         set_cpus_allowed(current, old_mask);
 564
 565         snapshot_timebases();
 566
 567         dump_numa_cpu_topology();
 568 }
 569
 570 #ifdef CONFIG_HOTPLUG_CPU
 571 int __cpu_disable(void)
 572 {
 573         struct device_node *l2_cache;
 574         int cpu = smp_processor_id();
 575         int base, i;
 576         int err;
 577
 578         if (!smp_ops->cpu_disable)
 579                 return -ENOSYS;
 580
 581         err = smp_ops->cpu_disable();
 582         if (err)
 583                 return err;
 584
 585         /* Update sibling maps */
 586         base = cpu_first_thread_in_core(cpu);
 587         for (i = 0; i < threads_per_core; i++) {
 588                 cpu_clear(cpu, per_cpu(cpu_sibling_map, base + i));
 589                 cpu_clear(base + i, per_cpu(cpu_sibling_map, cpu));
 590                 cpu_clear(cpu, per_cpu(cpu_core_map, base +i));
 591                 cpu_clear(base + i, per_cpu(cpu_core_map, cpu));
 592         }
 593
 594         l2_cache = cpu_to_l2cache(cpu);
 595         for_each_present_cpu(i) {
 596                 struct device_node *np = cpu_to_l2cache(i);
 597                 if (!np)
 598                         continue;
 599                 if (np == l2_cache) {
 600                         cpu_clear(cpu, per_cpu(cpu_core_map, i));
 601                         cpu_clear(i, per_cpu(cpu_core_map, cpu));
 602                 }
 603                 of_node_put(np);
 604         }
 605         of_node_put(l2_cache);
 606
 607
 608         return 0;
 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