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_mask(const struct cpumask *mask)
 193 {
 194         unsigned int cpu;
 195
 196         for_each_cpu(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                 if (smp_ops->probe)
 273                         max_cpus = smp_ops->probe();
 274                 else
 275                         max_cpus = NR_CPUS;
 276         else
 277                 max_cpus = 1;
 278
 279         smp_space_timers(max_cpus);
 280
 281         for_each_possible_cpu(cpu)
 282                 if (cpu != boot_cpuid)
 283                         smp_create_idle(cpu);
 284 }
 285
 286 void __devinit smp_prepare_boot_cpu(void)
 287 {
 288         BUG_ON(smp_processor_id() != boot_cpuid);
 289
 290         set_cpu_online(boot_cpuid, true);
 291         cpu_set(boot_cpuid, per_cpu(cpu_sibling_map, boot_cpuid));
 292         cpu_set(boot_cpuid, per_cpu(cpu_core_map, boot_cpuid));
 293 #ifdef CONFIG_PPC64
 294         paca[boot_cpuid].__current = current;
 295 #endif
 296         current_set[boot_cpuid] = task_thread_info(current);
 297 }
 298
 299 #ifdef CONFIG_HOTPLUG_CPU
 300 /* State of each CPU during hotplug phases */
 301 DEFINE_PER_CPU(int, cpu_state) = { 0 };
 302
 303 int generic_cpu_disable(void)
 304 {
 305         unsigned int cpu = smp_processor_id();
 306
 307         if (cpu == boot_cpuid)
 308                 return -EBUSY;
 309
 310         set_cpu_online(cpu, false);
 311 #ifdef CONFIG_PPC64
 312         vdso_data->processorCount--;
 313         fixup_irqs(cpu_online_map);
 314 #endif
 315         return 0;
 316 }
 317
 318 int generic_cpu_enable(unsigned int cpu)
 319 {
 320         /* Do the normal bootup if we haven't
 321          * already bootstrapped. */
 322         if (system_state != SYSTEM_RUNNING)
 323                 return -ENOSYS;
 324
 325         /* get the target out of it's holding state */
 326         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 327         smp_wmb();
 328
 329         while (!cpu_online(cpu))
 330                 cpu_relax();
 331
 332 #ifdef CONFIG_PPC64
 333         fixup_irqs(cpu_online_map);
 334         /* counter the irq disable in fixup_irqs */
 335         local_irq_enable();
 336 #endif
 337         return 0;
 338 }
 339
 340 void generic_cpu_die(unsigned int cpu)
 341 {
 342         int i;
 343
 344         for (i = 0; i < 100; i++) {
 345                 smp_rmb();
 346                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
 347                         return;
 348                 msleep(100);
 349         }
 350         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
 351 }
 352
 353 void generic_mach_cpu_die(void)
 354 {
 355         unsigned int cpu;
 356
 357         local_irq_disable();
 358         cpu = smp_processor_id();
 359         printk(KERN_DEBUG "CPU%d offline\n", cpu);
 360         __get_cpu_var(cpu_state) = CPU_DEAD;
 361         smp_wmb();
 362         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
 363                 cpu_relax();
 364         set_cpu_online(cpu, true);
 365         local_irq_enable();
 366 }
 367 #endif
 368
 369 static int __devinit cpu_enable(unsigned int cpu)
 370 {
 371         if (smp_ops && smp_ops->cpu_enable)
 372                 return smp_ops->cpu_enable(cpu);
 373
 374         return -ENOSYS;
 375 }
 376
 377 int __cpuinit __cpu_up(unsigned int cpu)
 378 {
 379         int c;
 380
 381         secondary_ti = current_set[cpu];
 382         if (!cpu_enable(cpu))
 383                 return 0;
 384
 385         if (smp_ops == NULL ||
 386             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
 387                 return -EINVAL;
 388
 389         /* Make sure callin-map entry is 0 (can be leftover a CPU
 390          * hotplug
 391          */
 392         cpu_callin_map[cpu] = 0;
 393
 394         /* The information for processor bringup must
 395          * be written out to main store before we release
 396          * the processor.
 397          */
 398         smp_mb();
 399
 400         /* wake up cpus */
 401         DBG("smp: kicking cpu %d\n", cpu);
 402         smp_ops->kick_cpu(cpu);
 403
 404         /*
 405          * wait to see if the cpu made a callin (is actually up).
 406          * use this value that I found through experimentation.
 407          * -- Cort
 408          */
 409         if (system_state < SYSTEM_RUNNING)
 410                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
 411                         udelay(100);
 412 #ifdef CONFIG_HOTPLUG_CPU
 413         else
 414                 /*
 415                  * CPUs can take much longer to come up in the
 416                  * hotplug case.  Wait five seconds.
 417                  */
 418                 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
 419                         msleep(1);
 420 #endif
 421
 422         if (!cpu_callin_map[cpu]) {
 423                 printk("Processor %u is stuck.\n", cpu);
 424                 return -ENOENT;
 425         }
 426
 427         printk("Processor %u found.\n", cpu);
 428
 429         if (smp_ops->give_timebase)
 430                 smp_ops->give_timebase();
 431
 432         /* Wait until cpu puts itself in the online map */
 433         while (!cpu_online(cpu))
 434                 cpu_relax();
 435
 436         return 0;
 437 }
 438
 439 /* Return the value of the reg property corresponding to the given
 440  * logical cpu.
 441  */
 442 int cpu_to_core_id(int cpu)
 443 {
 444         struct device_node *np;
 445         const int *reg;
 446         int id = -1;
 447
 448         np = of_get_cpu_node(cpu, NULL);
 449         if (!np)
 450                 goto out;
 451
 452         reg = of_get_property(np, "reg", NULL);
 453         if (!reg)
 454                 goto out;
 455
 456         id = *reg;
 457 out:
 458         of_node_put(np);
 459         return id;
 460 }
 461
 462 /* Must be called when no change can occur to cpu_present_map,
 463  * i.e. during cpu online or offline.
 464  */
 465 static struct device_node *cpu_to_l2cache(int cpu)
 466 {
 467         struct device_node *np;
 468         struct device_node *cache;
 469
 470         if (!cpu_present(cpu))
 471                 return NULL;
 472
 473         np = of_get_cpu_node(cpu, NULL);
 474         if (np == NULL)
 475                 return NULL;
 476
 477         cache = of_find_next_cache_node(np);
 478
 479         of_node_put(np);
 480
 481         return cache;
 482 }
 483
 484 /* Activate a secondary processor. */
 485 int __devinit start_secondary(void *unused)
 486 {
 487         unsigned int cpu = smp_processor_id();
 488         struct device_node *l2_cache;
 489         int i, base;
 490
 491         atomic_inc(&init_mm.mm_count);
 492         current->active_mm = &init_mm;
 493
 494         smp_store_cpu_info(cpu);
 495         set_dec(tb_ticks_per_jiffy);
 496         preempt_disable();
 497         cpu_callin_map[cpu] = 1;
 498
 499         if (smp_ops->setup_cpu)
 500                 smp_ops->setup_cpu(cpu);
 501         if (smp_ops->take_timebase)
 502                 smp_ops->take_timebase();
 503
 504         if (system_state > SYSTEM_BOOTING)
 505                 snapshot_timebase();
 506
 507         secondary_cpu_time_init();
 508
 509         ipi_call_lock();
 510         notify_cpu_starting(cpu);
 511         set_cpu_online(cpu, true);
 512         /* Update sibling maps */
 513         base = cpu_first_thread_in_core(cpu);
 514         for (i = 0; i < threads_per_core; i++) {
 515                 if (cpu_is_offline(base + i))
 516                         continue;
 517                 cpu_set(cpu, per_cpu(cpu_sibling_map, base + i));
 518                 cpu_set(base + i, per_cpu(cpu_sibling_map, cpu));
 519
 520                 /* cpu_core_map should be a superset of
 521                  * cpu_sibling_map even if we don't have cache
 522                  * information, so update the former here, too.
 523                  */
 524                 cpu_set(cpu, per_cpu(cpu_core_map, base +i));
 525                 cpu_set(base + i, per_cpu(cpu_core_map, cpu));
 526         }
 527         l2_cache = cpu_to_l2cache(cpu);
 528         for_each_online_cpu(i) {
 529                 struct device_node *np = cpu_to_l2cache(i);
 530                 if (!np)
 531                         continue;
 532                 if (np == l2_cache) {
 533                         cpu_set(cpu, per_cpu(cpu_core_map, i));
 534                         cpu_set(i, per_cpu(cpu_core_map, cpu));
 535                 }
 536                 of_node_put(np);
 537         }
 538         of_node_put(l2_cache);
 539         ipi_call_unlock();
 540
 541         local_irq_enable();
 542
 543         cpu_idle();
 544         return 0;
 545 }
 546
 547 int setup_profiling_timer(unsigned int multiplier)
 548 {
 549         return 0;
 550 }
 551
 552 void __init smp_cpus_done(unsigned int max_cpus)
 553 {
 554         cpumask_t old_mask;
 555
 556         /* We want the setup_cpu() here to be called from CPU 0, but our
 557          * init thread may have been "borrowed" by another CPU in the meantime
 558          * se we pin us down to CPU 0 for a short while
 559          */
 560         old_mask = current->cpus_allowed;
 561         set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid));
 562
 563         if (smp_ops && smp_ops->setup_cpu)
 564                 smp_ops->setup_cpu(boot_cpuid);
 565
 566         set_cpus_allowed(current, old_mask);
 567
 568         snapshot_timebases();
 569
 570         dump_numa_cpu_topology();
 571 }
 572
 573 #ifdef CONFIG_HOTPLUG_CPU
 574 int __cpu_disable(void)
 575 {
 576         struct device_node *l2_cache;
 577         int cpu = smp_processor_id();
 578         int base, i;
 579         int err;
 580
 581         if (!smp_ops->cpu_disable)
 582                 return -ENOSYS;
 583
 584         err = smp_ops->cpu_disable();
 585         if (err)
 586                 return err;
 587
 588         /* Update sibling maps */
 589         base = cpu_first_thread_in_core(cpu);
 590         for (i = 0; i < threads_per_core; i++) {
 591                 cpu_clear(cpu, per_cpu(cpu_sibling_map, base + i));
 592                 cpu_clear(base + i, per_cpu(cpu_sibling_map, cpu));
 593                 cpu_clear(cpu, per_cpu(cpu_core_map, base +i));
 594                 cpu_clear(base + i, per_cpu(cpu_core_map, cpu));
 595         }
 596
 597         l2_cache = cpu_to_l2cache(cpu);
 598         for_each_present_cpu(i) {
 599                 struct device_node *np = cpu_to_l2cache(i);
 600                 if (!np)
 601                         continue;
 602                 if (np == l2_cache) {
 603                         cpu_clear(cpu, per_cpu(cpu_core_map, i));
 604                         cpu_clear(i, per_cpu(cpu_core_map, cpu));
 605                 }
 606                 of_node_put(np);
 607         }
 608         of_node_put(l2_cache);
 609
 610
 611         return 0;
 612 }
 613
 614 void __cpu_die(unsigned int cpu)
 615 {
 616         if (smp_ops->cpu_die)
 617                 smp_ops->cpu_die(cpu);
 618 }
 619 #endif