| 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_MPIC |
| 82 | int __init smp_mpic_probe(void) |
| 83 | { |
| 84 | int nr_cpus; |
| 85 | |
| 86 | DBG("smp_mpic_probe()...\n"); |
| 87 | |
| 88 | nr_cpus = cpus_weight(cpu_possible_map); |
| 89 | |
| 90 | DBG("nr_cpus: %d\n", nr_cpus); |
| 91 | |
| 92 | if (nr_cpus > 1) |
| 93 | mpic_request_ipis(); |
| 94 | |
| 95 | return nr_cpus; |
| 96 | } |
| 97 | |
| 98 | void __devinit smp_mpic_setup_cpu(int cpu) |
| 99 | { |
| 100 | mpic_setup_this_cpu(); |
| 101 | } |
| 102 | #endif /* CONFIG_MPIC */ |
| 103 | |
| 104 | #ifdef CONFIG_PPC64 |
| 105 | void __devinit smp_generic_kick_cpu(int nr) |
| 106 | { |
| 107 | BUG_ON(nr < 0 || nr >= NR_CPUS); |
| 108 | |
| 109 | /* |
| 110 | * The processor is currently spinning, waiting for the |
| 111 | * cpu_start field to become non-zero After we set cpu_start, |
| 112 | * the processor will continue on to secondary_start |
| 113 | */ |
| 114 | paca[nr].cpu_start = 1; |
| 115 | smp_mb(); |
| 116 | } |
| 117 | #endif |
| 118 | |
| 119 | void smp_message_recv(int msg) |
| 120 | { |
| 121 | switch(msg) { |
| 122 | case PPC_MSG_CALL_FUNCTION: |
| 123 | smp_call_function_interrupt(); |
| 124 | break; |
| 125 | case PPC_MSG_RESCHEDULE: |
| 126 | /* XXX Do we have to do this? */ |
| 127 | set_need_resched(); |
| 128 | break; |
| 129 | case PPC_MSG_DEBUGGER_BREAK: |
| 130 | if (crash_ipi_function_ptr) { |
| 131 | crash_ipi_function_ptr(get_irq_regs()); |
| 132 | break; |
| 133 | } |
| 134 | #ifdef CONFIG_DEBUGGER |
| 135 | debugger_ipi(get_irq_regs()); |
| 136 | break; |
| 137 | #endif /* CONFIG_DEBUGGER */ |
| 138 | /* FALLTHROUGH */ |
| 139 | default: |
| 140 | printk("SMP %d: smp_message_recv(): unknown msg %d\n", |
| 141 | smp_processor_id(), msg); |
| 142 | break; |
| 143 | } |
| 144 | } |
| 145 | |
| 146 | void smp_send_reschedule(int cpu) |
| 147 | { |
| 148 | if (likely(smp_ops)) |
| 149 | smp_ops->message_pass(cpu, PPC_MSG_RESCHEDULE); |
| 150 | } |
| 151 | |
| 152 | #ifdef CONFIG_DEBUGGER |
| 153 | void smp_send_debugger_break(int cpu) |
| 154 | { |
| 155 | if (likely(smp_ops)) |
| 156 | smp_ops->message_pass(cpu, PPC_MSG_DEBUGGER_BREAK); |
| 157 | } |
| 158 | #endif |
| 159 | |
| 160 | #ifdef CONFIG_KEXEC |
| 161 | void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *)) |
| 162 | { |
| 163 | crash_ipi_function_ptr = crash_ipi_callback; |
| 164 | if (crash_ipi_callback && smp_ops) { |
| 165 | mb(); |
| 166 | smp_ops->message_pass(MSG_ALL_BUT_SELF, PPC_MSG_DEBUGGER_BREAK); |
| 167 | } |
| 168 | } |
| 169 | #endif |
| 170 | |
| 171 | static void stop_this_cpu(void *dummy) |
| 172 | { |
| 173 | local_irq_disable(); |
| 174 | while (1) |
| 175 | ; |
| 176 | } |
| 177 | |
| 178 | void smp_send_stop(void) |
| 179 | { |
| 180 | smp_call_function(stop_this_cpu, NULL, 1, 0); |
| 181 | } |
| 182 | |
| 183 | /* |
| 184 | * Structure and data for smp_call_function(). This is designed to minimise |
| 185 | * static memory requirements. It also looks cleaner. |
| 186 | * Stolen from the i386 version. |
| 187 | */ |
| 188 | static __cacheline_aligned_in_smp DEFINE_SPINLOCK(call_lock); |
| 189 | |
| 190 | static struct call_data_struct { |
| 191 | void (*func) (void *info); |
| 192 | void *info; |
| 193 | atomic_t started; |
| 194 | atomic_t finished; |
| 195 | int wait; |
| 196 | } *call_data; |
| 197 | |
| 198 | /* delay of at least 8 seconds */ |
| 199 | #define SMP_CALL_TIMEOUT 8 |
| 200 | |
| 201 | /* |
| 202 | * This function sends a 'generic call function' IPI to all other CPUs |
| 203 | * in the system. |
| 204 | * |
| 205 | * [SUMMARY] Run a function on all other CPUs. |
| 206 | * <func> The function to run. This must be fast and non-blocking. |
| 207 | * <info> An arbitrary pointer to pass to the function. |
| 208 | * <nonatomic> currently unused. |
| 209 | * <wait> If true, wait (atomically) until function has completed on other CPUs. |
| 210 | * [RETURNS] 0 on success, else a negative status code. Does not return until |
| 211 | * remote CPUs are nearly ready to execute <<func>> or are or have executed. |
| 212 | * |
| 213 | * You must not call this function with disabled interrupts or from a |
| 214 | * hardware interrupt handler or from a bottom half handler. |
| 215 | */ |
| 216 | int smp_call_function (void (*func) (void *info), void *info, int nonatomic, |
| 217 | int wait) |
| 218 | { |
| 219 | struct call_data_struct data; |
| 220 | int ret = -1, cpus; |
| 221 | u64 timeout; |
| 222 | |
| 223 | /* Can deadlock when called with interrupts disabled */ |
| 224 | WARN_ON(irqs_disabled()); |
| 225 | |
| 226 | if (unlikely(smp_ops == NULL)) |
| 227 | return -1; |
| 228 | |
| 229 | data.func = func; |
| 230 | data.info = info; |
| 231 | atomic_set(&data.started, 0); |
| 232 | data.wait = wait; |
| 233 | if (wait) |
| 234 | atomic_set(&data.finished, 0); |
| 235 | |
| 236 | spin_lock(&call_lock); |
| 237 | /* Must grab online cpu count with preempt disabled, otherwise |
| 238 | * it can change. */ |
| 239 | cpus = num_online_cpus() - 1; |
| 240 | if (!cpus) { |
| 241 | ret = 0; |
| 242 | goto out; |
| 243 | } |
| 244 | |
| 245 | call_data = &data; |
| 246 | smp_wmb(); |
| 247 | /* Send a message to all other CPUs and wait for them to respond */ |
| 248 | smp_ops->message_pass(MSG_ALL_BUT_SELF, PPC_MSG_CALL_FUNCTION); |
| 249 | |
| 250 | timeout = get_tb() + (u64) SMP_CALL_TIMEOUT * tb_ticks_per_sec; |
| 251 | |
| 252 | /* Wait for response */ |
| 253 | while (atomic_read(&data.started) != cpus) { |
| 254 | HMT_low(); |
| 255 | if (get_tb() >= timeout) { |
| 256 | printk("smp_call_function on cpu %d: other cpus not " |
| 257 | "responding (%d)\n", smp_processor_id(), |
| 258 | atomic_read(&data.started)); |
| 259 | debugger(NULL); |
| 260 | goto out; |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | if (wait) { |
| 265 | while (atomic_read(&data.finished) != cpus) { |
| 266 | HMT_low(); |
| 267 | if (get_tb() >= timeout) { |
| 268 | printk("smp_call_function on cpu %d: other " |
| 269 | "cpus not finishing (%d/%d)\n", |
| 270 | smp_processor_id(), |
| 271 | atomic_read(&data.finished), |
| 272 | atomic_read(&data.started)); |
| 273 | debugger(NULL); |
| 274 | goto out; |
| 275 | } |
| 276 | } |
| 277 | } |
| 278 | |
| 279 | ret = 0; |
| 280 | |
| 281 | out: |
| 282 | call_data = NULL; |
| 283 | HMT_medium(); |
| 284 | spin_unlock(&call_lock); |
| 285 | return ret; |
| 286 | } |
| 287 | |
| 288 | EXPORT_SYMBOL(smp_call_function); |
| 289 | |
| 290 | void smp_call_function_interrupt(void) |
| 291 | { |
| 292 | void (*func) (void *info); |
| 293 | void *info; |
| 294 | int wait; |
| 295 | |
| 296 | /* call_data will be NULL if the sender timed out while |
| 297 | * waiting on us to receive the call. |
| 298 | */ |
| 299 | if (!call_data) |
| 300 | return; |
| 301 | |
| 302 | func = call_data->func; |
| 303 | info = call_data->info; |
| 304 | wait = call_data->wait; |
| 305 | |
| 306 | if (!wait) |
| 307 | smp_mb__before_atomic_inc(); |
| 308 | |
| 309 | /* |
| 310 | * Notify initiating CPU that I've grabbed the data and am |
| 311 | * about to execute the function |
| 312 | */ |
| 313 | atomic_inc(&call_data->started); |
| 314 | /* |
| 315 | * At this point the info structure may be out of scope unless wait==1 |
| 316 | */ |
| 317 | (*func)(info); |
| 318 | if (wait) { |
| 319 | smp_mb__before_atomic_inc(); |
| 320 | atomic_inc(&call_data->finished); |
| 321 | } |
| 322 | } |
| 323 | |
| 324 | extern struct gettimeofday_struct do_gtod; |
| 325 | |
| 326 | struct thread_info *current_set[NR_CPUS]; |
| 327 | |
| 328 | DECLARE_PER_CPU(unsigned int, pvr); |
| 329 | |
| 330 | static void __devinit smp_store_cpu_info(int id) |
| 331 | { |
| 332 | per_cpu(pvr, id) = mfspr(SPRN_PVR); |
| 333 | } |
| 334 | |
| 335 | static void __init smp_create_idle(unsigned int cpu) |
| 336 | { |
| 337 | struct task_struct *p; |
| 338 | |
| 339 | /* create a process for the processor */ |
| 340 | p = fork_idle(cpu); |
| 341 | if (IS_ERR(p)) |
| 342 | panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p)); |
| 343 | #ifdef CONFIG_PPC64 |
| 344 | paca[cpu].__current = p; |
| 345 | #endif |
| 346 | current_set[cpu] = task_thread_info(p); |
| 347 | task_thread_info(p)->cpu = cpu; |
| 348 | } |
| 349 | |
| 350 | void __init smp_prepare_cpus(unsigned int max_cpus) |
| 351 | { |
| 352 | unsigned int cpu; |
| 353 | |
| 354 | DBG("smp_prepare_cpus\n"); |
| 355 | |
| 356 | /* |
| 357 | * setup_cpu may need to be called on the boot cpu. We havent |
| 358 | * spun any cpus up but lets be paranoid. |
| 359 | */ |
| 360 | BUG_ON(boot_cpuid != smp_processor_id()); |
| 361 | |
| 362 | /* Fixup boot cpu */ |
| 363 | smp_store_cpu_info(boot_cpuid); |
| 364 | cpu_callin_map[boot_cpuid] = 1; |
| 365 | |
| 366 | if (smp_ops) |
| 367 | max_cpus = smp_ops->probe(); |
| 368 | else |
| 369 | max_cpus = 1; |
| 370 | |
| 371 | smp_space_timers(max_cpus); |
| 372 | |
| 373 | for_each_possible_cpu(cpu) |
| 374 | if (cpu != boot_cpuid) |
| 375 | smp_create_idle(cpu); |
| 376 | } |
| 377 | |
| 378 | void __devinit smp_prepare_boot_cpu(void) |
| 379 | { |
| 380 | BUG_ON(smp_processor_id() != boot_cpuid); |
| 381 | |
| 382 | cpu_set(boot_cpuid, cpu_online_map); |
| 383 | #ifdef CONFIG_PPC64 |
| 384 | paca[boot_cpuid].__current = current; |
| 385 | #endif |
| 386 | current_set[boot_cpuid] = task_thread_info(current); |
| 387 | } |
| 388 | |
| 389 | #ifdef CONFIG_HOTPLUG_CPU |
| 390 | /* State of each CPU during hotplug phases */ |
| 391 | DEFINE_PER_CPU(int, cpu_state) = { 0 }; |
| 392 | |
| 393 | int generic_cpu_disable(void) |
| 394 | { |
| 395 | unsigned int cpu = smp_processor_id(); |
| 396 | |
| 397 | if (cpu == boot_cpuid) |
| 398 | return -EBUSY; |
| 399 | |
| 400 | cpu_clear(cpu, cpu_online_map); |
| 401 | #ifdef CONFIG_PPC64 |
| 402 | vdso_data->processorCount--; |
| 403 | fixup_irqs(cpu_online_map); |
| 404 | #endif |
| 405 | return 0; |
| 406 | } |
| 407 | |
| 408 | int generic_cpu_enable(unsigned int cpu) |
| 409 | { |
| 410 | /* Do the normal bootup if we haven't |
| 411 | * already bootstrapped. */ |
| 412 | if (system_state != SYSTEM_RUNNING) |
| 413 | return -ENOSYS; |
| 414 | |
| 415 | /* get the target out of it's holding state */ |
| 416 | per_cpu(cpu_state, cpu) = CPU_UP_PREPARE; |
| 417 | smp_wmb(); |
| 418 | |
| 419 | while (!cpu_online(cpu)) |
| 420 | cpu_relax(); |
| 421 | |
| 422 | #ifdef CONFIG_PPC64 |
| 423 | fixup_irqs(cpu_online_map); |
| 424 | /* counter the irq disable in fixup_irqs */ |
| 425 | local_irq_enable(); |
| 426 | #endif |
| 427 | return 0; |
| 428 | } |
| 429 | |
| 430 | void generic_cpu_die(unsigned int cpu) |
| 431 | { |
| 432 | int i; |
| 433 | |
| 434 | for (i = 0; i < 100; i++) { |
| 435 | smp_rmb(); |
| 436 | if (per_cpu(cpu_state, cpu) == CPU_DEAD) |
| 437 | return; |
| 438 | msleep(100); |
| 439 | } |
| 440 | printk(KERN_ERR "CPU%d didn't die...\n", cpu); |
| 441 | } |
| 442 | |
| 443 | void generic_mach_cpu_die(void) |
| 444 | { |
| 445 | unsigned int cpu; |
| 446 | |
| 447 | local_irq_disable(); |
| 448 | cpu = smp_processor_id(); |
| 449 | printk(KERN_DEBUG "CPU%d offline\n", cpu); |
| 450 | __get_cpu_var(cpu_state) = CPU_DEAD; |
| 451 | smp_wmb(); |
| 452 | while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE) |
| 453 | cpu_relax(); |
| 454 | |
| 455 | #ifdef CONFIG_PPC64 |
| 456 | flush_tlb_pending(); |
| 457 | #endif |
| 458 | cpu_set(cpu, cpu_online_map); |
| 459 | local_irq_enable(); |
| 460 | } |
| 461 | #endif |
| 462 | |
| 463 | static int __devinit cpu_enable(unsigned int cpu) |
| 464 | { |
| 465 | if (smp_ops && smp_ops->cpu_enable) |
| 466 | return smp_ops->cpu_enable(cpu); |
| 467 | |
| 468 | return -ENOSYS; |
| 469 | } |
| 470 | |
| 471 | int __devinit __cpu_up(unsigned int cpu) |
| 472 | { |
| 473 | int c; |
| 474 | |
| 475 | secondary_ti = current_set[cpu]; |
| 476 | if (!cpu_enable(cpu)) |
| 477 | return 0; |
| 478 | |
| 479 | if (smp_ops == NULL || |
| 480 | (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu))) |
| 481 | return -EINVAL; |
| 482 | |
| 483 | /* Make sure callin-map entry is 0 (can be leftover a CPU |
| 484 | * hotplug |
| 485 | */ |
| 486 | cpu_callin_map[cpu] = 0; |
| 487 | |
| 488 | /* The information for processor bringup must |
| 489 | * be written out to main store before we release |
| 490 | * the processor. |
| 491 | */ |
| 492 | smp_mb(); |
| 493 | |
| 494 | /* wake up cpus */ |
| 495 | DBG("smp: kicking cpu %d\n", cpu); |
| 496 | smp_ops->kick_cpu(cpu); |
| 497 | |
| 498 | /* |
| 499 | * wait to see if the cpu made a callin (is actually up). |
| 500 | * use this value that I found through experimentation. |
| 501 | * -- Cort |
| 502 | */ |
| 503 | if (system_state < SYSTEM_RUNNING) |
| 504 | for (c = 50000; c && !cpu_callin_map[cpu]; c--) |
| 505 | udelay(100); |
| 506 | #ifdef CONFIG_HOTPLUG_CPU |
| 507 | else |
| 508 | /* |
| 509 | * CPUs can take much longer to come up in the |
| 510 | * hotplug case. Wait five seconds. |
| 511 | */ |
| 512 | for (c = 25; c && !cpu_callin_map[cpu]; c--) { |
| 513 | msleep(200); |
| 514 | } |
| 515 | #endif |
| 516 | |
| 517 | if (!cpu_callin_map[cpu]) { |
| 518 | printk("Processor %u is stuck.\n", cpu); |
| 519 | return -ENOENT; |
| 520 | } |
| 521 | |
| 522 | printk("Processor %u found.\n", cpu); |
| 523 | |
| 524 | if (smp_ops->give_timebase) |
| 525 | smp_ops->give_timebase(); |
| 526 | |
| 527 | /* Wait until cpu puts itself in the online map */ |
| 528 | while (!cpu_online(cpu)) |
| 529 | cpu_relax(); |
| 530 | |
| 531 | return 0; |
| 532 | } |
| 533 | |
| 534 | |
| 535 | /* Activate a secondary processor. */ |
| 536 | int __devinit start_secondary(void *unused) |
| 537 | { |
| 538 | unsigned int cpu = smp_processor_id(); |
| 539 | |
| 540 | atomic_inc(&init_mm.mm_count); |
| 541 | current->active_mm = &init_mm; |
| 542 | |
| 543 | smp_store_cpu_info(cpu); |
| 544 | set_dec(tb_ticks_per_jiffy); |
| 545 | preempt_disable(); |
| 546 | cpu_callin_map[cpu] = 1; |
| 547 | |
| 548 | smp_ops->setup_cpu(cpu); |
| 549 | if (smp_ops->take_timebase) |
| 550 | smp_ops->take_timebase(); |
| 551 | |
| 552 | if (system_state > SYSTEM_BOOTING) |
| 553 | snapshot_timebase(); |
| 554 | |
| 555 | spin_lock(&call_lock); |
| 556 | cpu_set(cpu, cpu_online_map); |
| 557 | spin_unlock(&call_lock); |
| 558 | |
| 559 | local_irq_enable(); |
| 560 | |
| 561 | cpu_idle(); |
| 562 | return 0; |
| 563 | } |
| 564 | |
| 565 | int setup_profiling_timer(unsigned int multiplier) |
| 566 | { |
| 567 | return 0; |
| 568 | } |
| 569 | |
| 570 | void __init smp_cpus_done(unsigned int max_cpus) |
| 571 | { |
| 572 | cpumask_t old_mask; |
| 573 | |
| 574 | /* We want the setup_cpu() here to be called from CPU 0, but our |
| 575 | * init thread may have been "borrowed" by another CPU in the meantime |
| 576 | * se we pin us down to CPU 0 for a short while |
| 577 | */ |
| 578 | old_mask = current->cpus_allowed; |
| 579 | set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid)); |
| 580 | |
| 581 | if (smp_ops) |
| 582 | smp_ops->setup_cpu(boot_cpuid); |
| 583 | |
| 584 | set_cpus_allowed(current, old_mask); |
| 585 | |
| 586 | snapshot_timebases(); |
| 587 | |
| 588 | dump_numa_cpu_topology(); |
| 589 | } |
| 590 | |
| 591 | #ifdef CONFIG_HOTPLUG_CPU |
| 592 | int __cpu_disable(void) |
| 593 | { |
| 594 | if (smp_ops->cpu_disable) |
| 595 | return smp_ops->cpu_disable(); |
| 596 | |
| 597 | return -ENOSYS; |
| 598 | } |
| 599 | |
| 600 | void __cpu_die(unsigned int cpu) |
| 601 | { |
| 602 | if (smp_ops->cpu_die) |
| 603 | smp_ops->cpu_die(cpu); |
| 604 | } |
| 605 | #endif |