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 int smp_hw_index[NR_CPUS];
  61 struct thread_info *secondary_ti;
  62
  63 cpumask_t cpu_possible_map = CPU_MASK_NONE;
  64 cpumask_t cpu_online_map = CPU_MASK_NONE;
  65 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
  66 DEFINE_PER_CPU(cpumask_t, cpu_core_map) = CPU_MASK_NONE;
  67
  68 EXPORT_SYMBOL(cpu_online_map);
  69 EXPORT_SYMBOL(cpu_possible_map);
  70 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
  71 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
  72
  73 /* SMP operations for this machine */
  74 struct smp_ops_t *smp_ops;
  75
  76 static volatile unsigned int cpu_callin_map[NR_CPUS];
  77
  78 int smt_enabled_at_boot = 1;
  79
  80 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
  81
  82 #ifdef CONFIG_PPC64
  83 void __devinit smp_generic_kick_cpu(int nr)
  84 {
  85         BUG_ON(nr < 0 || nr >= NR_CPUS);
  86
  87         /*
  88          * The processor is currently spinning, waiting for the
  89          * cpu_start field to become non-zero After we set cpu_start,
  90          * the processor will continue on to secondary_start
  91          */
  92         paca[nr].cpu_start = 1;
  93         smp_mb();
  94 }
  95 #endif
  96
  97 void smp_message_recv(int msg)
  98 {
  99         switch(msg) {
 100         case PPC_MSG_CALL_FUNCTION:
 101                 generic_smp_call_function_interrupt();
 102                 break;
 103         case PPC_MSG_RESCHEDULE:
 104                 /* we notice need_resched on exit */
 105                 break;
 106         case PPC_MSG_CALL_FUNC_SINGLE:
 107                 generic_smp_call_function_single_interrupt();
 108                 break;
 109         case PPC_MSG_DEBUGGER_BREAK:
 110                 if (crash_ipi_function_ptr) {
 111                         crash_ipi_function_ptr(get_irq_regs());
 112                         break;
 113                 }
 114 #ifdef CONFIG_DEBUGGER
 115                 debugger_ipi(get_irq_regs());
 116                 break;
 117 #endif /* CONFIG_DEBUGGER */
 118                 /* FALLTHROUGH */
 119         default:
 120                 printk("SMP %d: smp_message_recv(): unknown msg %d\n",
 121                        smp_processor_id(), msg);
 122                 break;
 123         }
 124 }
 125
 126 static irqreturn_t call_function_action(int irq, void *data)
 127 {
 128         generic_smp_call_function_interrupt();
 129         return IRQ_HANDLED;
 130 }
 131
 132 static irqreturn_t reschedule_action(int irq, void *data)
 133 {
 134         /* we just need the return path side effect of checking need_resched */
 135         return IRQ_HANDLED;
 136 }
 137
 138 static irqreturn_t call_function_single_action(int irq, void *data)
 139 {
 140         generic_smp_call_function_single_interrupt();
 141         return IRQ_HANDLED;
 142 }
 143
 144 static irqreturn_t debug_ipi_action(int irq, void *data)
 145 {
 146         smp_message_recv(PPC_MSG_DEBUGGER_BREAK);
 147         return IRQ_HANDLED;
 148 }
 149
 150 static irq_handler_t smp_ipi_action[] = {
 151         [PPC_MSG_CALL_FUNCTION] =  call_function_action,
 152         [PPC_MSG_RESCHEDULE] = reschedule_action,
 153         [PPC_MSG_CALL_FUNC_SINGLE] = call_function_single_action,
 154         [PPC_MSG_DEBUGGER_BREAK] = debug_ipi_action,
 155 };
 156
 157 const char *smp_ipi_name[] = {
 158         [PPC_MSG_CALL_FUNCTION] =  "ipi call function",
 159         [PPC_MSG_RESCHEDULE] = "ipi reschedule",
 160         [PPC_MSG_CALL_FUNC_SINGLE] = "ipi call function single",
 161         [PPC_MSG_DEBUGGER_BREAK] = "ipi debugger",
 162 };
 163
 164 /* optional function to request ipi, for controllers with >= 4 ipis */
 165 int smp_request_message_ipi(int virq, int msg)
 166 {
 167         int err;
 168
 169         if (msg < 0 || msg > PPC_MSG_DEBUGGER_BREAK) {
 170                 return -EINVAL;
 171         }
 172 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
 173         if (msg == PPC_MSG_DEBUGGER_BREAK) {
 174                 return 1;
 175         }
 176 #endif
 177         err = request_irq(virq, smp_ipi_action[msg], IRQF_DISABLED|IRQF_PERCPU,
 178                           smp_ipi_name[msg], 0);
 179         WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
 180                 virq, smp_ipi_name[msg], err);
 181
 182         return err;
 183 }
 184
 185 void smp_send_reschedule(int cpu)
 186 {
 187         if (likely(smp_ops))
 188                 smp_ops->message_pass(cpu, PPC_MSG_RESCHEDULE);
 189 }
 190
 191 void arch_send_call_function_single_ipi(int cpu)
 192 {
 193         smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
 194 }
 195
 196 void arch_send_call_function_ipi(cpumask_t mask)
 197 {
 198         unsigned int cpu;
 199
 200         for_each_cpu_mask(cpu, mask)
 201                 smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNCTION);
 202 }
 203
 204 #ifdef CONFIG_DEBUGGER
 205 void smp_send_debugger_break(int cpu)
 206 {
 207         if (likely(smp_ops))
 208                 smp_ops->message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
 209 }
 210 #endif
 211
 212 #ifdef CONFIG_KEXEC
 213 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
 214 {
 215         crash_ipi_function_ptr = crash_ipi_callback;
 216         if (crash_ipi_callback && smp_ops) {
 217                 mb();
 218                 smp_ops->message_pass(MSG_ALL_BUT_SELF, PPC_MSG_DEBUGGER_BREAK);
 219         }
 220 }
 221 #endif
 222
 223 static void stop_this_cpu(void *dummy)
 224 {
 225         local_irq_disable();
 226         while (1)
 227                 ;
 228 }
 229
 230 void smp_send_stop(void)
 231 {
 232         smp_call_function(stop_this_cpu, NULL, 0);
 233 }
 234
 235 struct thread_info *current_set[NR_CPUS];
 236
 237 static void __devinit smp_store_cpu_info(int id)
 238 {
 239         per_cpu(pvr, id) = mfspr(SPRN_PVR);
 240 }
 241
 242 static void __init smp_create_idle(unsigned int cpu)
 243 {
 244         struct task_struct *p;
 245
 246         /* create a process for the processor */
 247         p = fork_idle(cpu);
 248         if (IS_ERR(p))
 249                 panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p));
 250 #ifdef CONFIG_PPC64
 251         paca[cpu].__current = p;
 252         paca[cpu].kstack = (unsigned long) task_thread_info(p)
 253                 + THREAD_SIZE - STACK_FRAME_OVERHEAD;
 254 #endif
 255         current_set[cpu] = task_thread_info(p);
 256         task_thread_info(p)->cpu = cpu;
 257 }
 258
 259 void __init smp_prepare_cpus(unsigned int max_cpus)
 260 {
 261         unsigned int cpu;
 262
 263         DBG("smp_prepare_cpus\n");
 264
 265         /*
 266          * setup_cpu may need to be called on the boot cpu. We havent
 267          * spun any cpus up but lets be paranoid.
 268          */
 269         BUG_ON(boot_cpuid != smp_processor_id());
 270
 271         /* Fixup boot cpu */
 272         smp_store_cpu_info(boot_cpuid);
 273         cpu_callin_map[boot_cpuid] = 1;
 274
 275         if (smp_ops)
 276                 max_cpus = smp_ops->probe();
 277         else
 278                 max_cpus = 1;
 279
 280         smp_space_timers(max_cpus);
 281
 282         for_each_possible_cpu(cpu)
 283                 if (cpu != boot_cpuid)
 284                         smp_create_idle(cpu);
 285 }
 286
 287 void __devinit smp_prepare_boot_cpu(void)
 288 {
 289         BUG_ON(smp_processor_id() != boot_cpuid);
 290
 291         cpu_set(boot_cpuid, cpu_online_map);
 292         cpu_set(boot_cpuid, per_cpu(cpu_sibling_map, boot_cpuid));
 293         cpu_set(boot_cpuid, per_cpu(cpu_core_map, boot_cpuid));
 294 #ifdef CONFIG_PPC64
 295         paca[boot_cpuid].__current = current;
 296 #endif
 297         current_set[boot_cpuid] = task_thread_info(current);
 298 }
 299
 300 #ifdef CONFIG_HOTPLUG_CPU
 301 /* State of each CPU during hotplug phases */
 302 DEFINE_PER_CPU(int, cpu_state) = { 0 };
 303
 304 int generic_cpu_disable(void)
 305 {
 306         unsigned int cpu = smp_processor_id();
 307
 308         if (cpu == boot_cpuid)
 309                 return -EBUSY;
 310
 311         cpu_clear(cpu, cpu_online_map);
 312 #ifdef CONFIG_PPC64
 313         vdso_data->processorCount--;
 314         fixup_irqs(cpu_online_map);
 315 #endif
 316         return 0;
 317 }
 318
 319 int generic_cpu_enable(unsigned int cpu)
 320 {
 321         /* Do the normal bootup if we haven't
 322          * already bootstrapped. */
 323         if (system_state != SYSTEM_RUNNING)
 324                 return -ENOSYS;
 325
 326         /* get the target out of it's holding state */
 327         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 328         smp_wmb();
 329
 330         while (!cpu_online(cpu))
 331                 cpu_relax();
 332
 333 #ifdef CONFIG_PPC64
 334         fixup_irqs(cpu_online_map);
 335         /* counter the irq disable in fixup_irqs */
 336         local_irq_enable();
 337 #endif
 338         return 0;
 339 }
 340
 341 void generic_cpu_die(unsigned int cpu)
 342 {
 343         int i;
 344
 345         for (i = 0; i < 100; i++) {
 346                 smp_rmb();
 347                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
 348                         return;
 349                 msleep(100);
 350         }
 351         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
 352 }
 353
 354 void generic_mach_cpu_die(void)
 355 {
 356         unsigned int cpu;
 357
 358         local_irq_disable();
 359         cpu = smp_processor_id();
 360         printk(KERN_DEBUG "CPU%d offline\n", cpu);
 361         __get_cpu_var(cpu_state) = CPU_DEAD;
 362         smp_wmb();
 363         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
 364                 cpu_relax();
 365         cpu_set(cpu, cpu_online_map);
 366         local_irq_enable();
 367 }
 368 #endif
 369
 370 static int __devinit cpu_enable(unsigned int cpu)
 371 {
 372         if (smp_ops && smp_ops->cpu_enable)
 373                 return smp_ops->cpu_enable(cpu);
 374
 375         return -ENOSYS;
 376 }
 377
 378 int __cpuinit __cpu_up(unsigned int cpu)
 379 {
 380         int c;
 381
 382         secondary_ti = current_set[cpu];
 383         if (!cpu_enable(cpu))
 384                 return 0;
 385
 386         if (smp_ops == NULL ||
 387             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
 388                 return -EINVAL;
 389
 390         /* Make sure callin-map entry is 0 (can be leftover a CPU
 391          * hotplug
 392          */
 393         cpu_callin_map[cpu] = 0;
 394
 395         /* The information for processor bringup must
 396          * be written out to main store before we release
 397          * the processor.
 398          */
 399         smp_mb();
 400
 401         /* wake up cpus */
 402         DBG("smp: kicking cpu %d\n", cpu);
 403         smp_ops->kick_cpu(cpu);
 404
 405         /*
 406          * wait to see if the cpu made a callin (is actually up).
 407          * use this value that I found through experimentation.
 408          * -- Cort
 409          */
 410         if (system_state < SYSTEM_RUNNING)
 411                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
 412                         udelay(100);
 413 #ifdef CONFIG_HOTPLUG_CPU
 414         else
 415                 /*
 416                  * CPUs can take much longer to come up in the
 417                  * hotplug case.  Wait five seconds.
 418                  */
 419                 for (c = 25; c && !cpu_callin_map[cpu]; c--) {
 420                         msleep(200);
 421                 }
 422 #endif
 423
 424         if (!cpu_callin_map[cpu]) {
 425                 printk("Processor %u is stuck.\n", cpu);
 426                 return -ENOENT;
 427         }
 428
 429         printk("Processor %u found.\n", cpu);
 430
 431         if (smp_ops->give_timebase)
 432                 smp_ops->give_timebase();
 433
 434         /* Wait until cpu puts itself in the online map */
 435         while (!cpu_online(cpu))
 436                 cpu_relax();
 437
 438         return 0;
 439 }
 440
 441 /* Return the value of the reg property corresponding to the given
 442  * logical cpu.
 443  */
 444 int cpu_to_core_id(int cpu)
 445 {
 446         struct device_node *np;
 447         const int *reg;
 448         int id = -1;
 449
 450         np = of_get_cpu_node(cpu, NULL);
 451         if (!np)
 452                 goto out;
 453
 454         reg = of_get_property(np, "reg", NULL);
 455         if (!reg)
 456                 goto out;
 457
 458         id = *reg;
 459 out:
 460         of_node_put(np);
 461         return id;
 462 }
 463
 464 /* Must be called when no change can occur to cpu_present_map,
 465  * i.e. during cpu online or offline.
 466  */
 467 static struct device_node *cpu_to_l2cache(int cpu)
 468 {
 469         struct device_node *np;
 470         const phandle *php;
 471         phandle ph;
 472
 473         if (!cpu_present(cpu))
 474                 return NULL;
 475
 476         np = of_get_cpu_node(cpu, NULL);
 477         if (np == NULL)
 478                 return NULL;
 479
 480         php = of_get_property(np, "l2-cache", NULL);
 481         if (php == NULL)
 482                 return NULL;
 483         ph = *php;
 484         of_node_put(np);
 485
 486         return of_find_node_by_phandle(ph);
 487 }
 488
 489 /* Activate a secondary processor. */
 490 int __devinit start_secondary(void *unused)
 491 {
 492         unsigned int cpu = smp_processor_id();
 493         struct device_node *l2_cache;
 494         int i, base;
 495
 496         atomic_inc(&init_mm.mm_count);
 497         current->active_mm = &init_mm;
 498
 499         smp_store_cpu_info(cpu);
 500         set_dec(tb_ticks_per_jiffy);
 501         preempt_disable();
 502         cpu_callin_map[cpu] = 1;
 503
 504         smp_ops->setup_cpu(cpu);
 505         if (smp_ops->take_timebase)
 506                 smp_ops->take_timebase();
 507
 508         if (system_state > SYSTEM_BOOTING)
 509                 snapshot_timebase();
 510
 511         secondary_cpu_time_init();
 512
 513         ipi_call_lock();
 514         notify_cpu_starting(cpu);
 515         cpu_set(cpu, cpu_online_map);
 516         /* Update sibling maps */
 517         base = cpu_first_thread_in_core(cpu);
 518         for (i = 0; i < threads_per_core; i++) {
 519                 if (cpu_is_offline(base + i))
 520                         continue;
 521                 cpu_set(cpu, per_cpu(cpu_sibling_map, base + i));
 522                 cpu_set(base + i, per_cpu(cpu_sibling_map, cpu));
 523
 524                 /* cpu_core_map should be a superset of
 525                  * cpu_sibling_map even if we don't have cache
 526                  * information, so update the former here, too.
 527                  */
 528                 cpu_set(cpu, per_cpu(cpu_core_map, base +i));
 529                 cpu_set(base + i, per_cpu(cpu_core_map, cpu));
 530         }
 531         l2_cache = cpu_to_l2cache(cpu);
 532         for_each_online_cpu(i) {
 533                 struct device_node *np = cpu_to_l2cache(i);
 534                 if (!np)
 535                         continue;
 536                 if (np == l2_cache) {
 537                         cpu_set(cpu, per_cpu(cpu_core_map, i));
 538                         cpu_set(i, per_cpu(cpu_core_map, cpu));
 539                 }
 540                 of_node_put(np);
 541         }
 542         of_node_put(l2_cache);
 543         ipi_call_unlock();
 544
 545         local_irq_enable();
 546
 547         cpu_idle();
 548         return 0;
 549 }
 550
 551 int setup_profiling_timer(unsigned int multiplier)
 552 {
 553         return 0;
 554 }
 555
 556 void __init smp_cpus_done(unsigned int max_cpus)
 557 {
 558         cpumask_t old_mask;
 559
 560         /* We want the setup_cpu() here to be called from CPU 0, but our
 561          * init thread may have been "borrowed" by another CPU in the meantime
 562          * se we pin us down to CPU 0 for a short while
 563          */
 564         old_mask = current->cpus_allowed;
 565         set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid));
 566
 567         if (smp_ops)
 568                 smp_ops->setup_cpu(boot_cpuid);
 569
 570         set_cpus_allowed(current, old_mask);
 571
 572         snapshot_timebases();
 573
 574         dump_numa_cpu_topology();
 575 }
 576
 577 #ifdef CONFIG_HOTPLUG_CPU
 578 int __cpu_disable(void)
 579 {
 580         struct device_node *l2_cache;
 581         int cpu = smp_processor_id();
 582         int base, i;
 583         int err;
 584
 585         if (!smp_ops->cpu_disable)
 586                 return -ENOSYS;
 587
 588         err = smp_ops->cpu_disable();
 589         if (err)
 590                 return err;
 591
 592         /* Update sibling maps */
 593         base = cpu_first_thread_in_core(cpu);
 594         for (i = 0; i < threads_per_core; i++) {
 595                 cpu_clear(cpu, per_cpu(cpu_sibling_map, base + i));
 596                 cpu_clear(base + i, per_cpu(cpu_sibling_map, cpu));
 597                 cpu_clear(cpu, per_cpu(cpu_core_map, base +i));
 598                 cpu_clear(base + i, per_cpu(cpu_core_map, cpu));
 599         }
 600
 601         l2_cache = cpu_to_l2cache(cpu);
 602         for_each_present_cpu(i) {
 603                 struct device_node *np = cpu_to_l2cache(i);
 604                 if (!np)
 605                         continue;
 606                 if (np == l2_cache) {
 607                         cpu_clear(cpu, per_cpu(cpu_core_map, i));
 608                         cpu_clear(i, per_cpu(cpu_core_map, cpu));
 609                 }
 610                 of_node_put(np);
 611         }
 612         of_node_put(l2_cache);
 613
 614
 615         return 0;
 616 }
 617
 618 void __cpu_die(unsigned int cpu)
 619 {
 620         if (smp_ops->cpu_die)
 621                 smp_ops->cpu_die(cpu);
 622 }
 623 #endif