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/export.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/device.h> 31 #include <linux/cpu.h> 32 #include <linux/notifier.h> 33 #include <linux/topology.h> 34 35 #include <asm/ptrace.h> 36 #include <linux/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/mpic.h> 47 #include <asm/vdso_datapage.h> 48 #ifdef CONFIG_PPC64 49 #include <asm/paca.h> 50 #endif 51 #include <asm/vdso.h> 52 #include <asm/debug.h> 53 54 #ifdef DEBUG 55 #include <asm/udbg.h> 56 #define DBG(fmt...) udbg_printf(fmt) 57 #else 58 #define DBG(fmt...) 59 #endif 60 61 #ifdef CONFIG_HOTPLUG_CPU 62 /* State of each CPU during hotplug phases */ 63 static DEFINE_PER_CPU(int, cpu_state) = { 0 }; 64 #endif 65 66 struct thread_info *secondary_ti; 67 68 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map); 69 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map); 70 71 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map); 72 EXPORT_PER_CPU_SYMBOL(cpu_core_map); 73 74 /* SMP operations for this machine */ 75 struct smp_ops_t *smp_ops; 76 77 /* Can't be static due to PowerMac hackery */ 78 volatile unsigned int cpu_callin_map[NR_CPUS]; 79 80 int smt_enabled_at_boot = 1; 81 82 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL; 83 84 #ifdef CONFIG_PPC64 85 int __devinit smp_generic_kick_cpu(int nr) 86 { 87 BUG_ON(nr < 0 || nr >= NR_CPUS); 88 89 /* 90 * The processor is currently spinning, waiting for the 91 * cpu_start field to become non-zero After we set cpu_start, 92 * the processor will continue on to secondary_start 93 */ 94 if (!paca[nr].cpu_start) { 95 paca[nr].cpu_start = 1; 96 smp_mb(); 97 return 0; 98 } 99 100 #ifdef CONFIG_HOTPLUG_CPU 101 /* 102 * Ok it's not there, so it might be soft-unplugged, let's 103 * try to bring it back 104 */ 105 per_cpu(cpu_state, nr) = CPU_UP_PREPARE; 106 smp_wmb(); 107 smp_send_reschedule(nr); 108 #endif /* CONFIG_HOTPLUG_CPU */ 109 110 return 0; 111 } 112 #endif /* CONFIG_PPC64 */ 113 114 static irqreturn_t call_function_action(int irq, void *data) 115 { 116 generic_smp_call_function_interrupt(); 117 return IRQ_HANDLED; 118 } 119 120 static irqreturn_t reschedule_action(int irq, void *data) 121 { 122 scheduler_ipi(); 123 return IRQ_HANDLED; 124 } 125 126 static irqreturn_t call_function_single_action(int irq, void *data) 127 { 128 generic_smp_call_function_single_interrupt(); 129 return IRQ_HANDLED; 130 } 131 132 static irqreturn_t debug_ipi_action(int irq, void *data) 133 { 134 if (crash_ipi_function_ptr) { 135 crash_ipi_function_ptr(get_irq_regs()); 136 return IRQ_HANDLED; 137 } 138 139 #ifdef CONFIG_DEBUGGER 140 debugger_ipi(get_irq_regs()); 141 #endif /* CONFIG_DEBUGGER */ 142 143 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], 174 IRQF_PERCPU | IRQF_NO_THREAD, 175 smp_ipi_name[msg], 0); 176 WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n", 177 virq, smp_ipi_name[msg], err); 178 179 return err; 180 } 181 182 #ifdef CONFIG_PPC_SMP_MUXED_IPI 183 struct cpu_messages { 184 int messages; /* current messages */ 185 unsigned long data; /* data for cause ipi */ 186 }; 187 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message); 188 189 void smp_muxed_ipi_set_data(int cpu, unsigned long data) 190 { 191 struct cpu_messages *info = &per_cpu(ipi_message, cpu); 192 193 info->data = data; 194 } 195 196 void smp_muxed_ipi_message_pass(int cpu, int msg) 197 { 198 struct cpu_messages *info = &per_cpu(ipi_message, cpu); 199 char *message = (char *)&info->messages; 200 201 message[msg] = 1; 202 mb(); 203 smp_ops->cause_ipi(cpu, info->data); 204 } 205 206 irqreturn_t smp_ipi_demux(void) 207 { 208 struct cpu_messages *info = &__get_cpu_var(ipi_message); 209 unsigned int all; 210 211 mb(); /* order any irq clear */ 212 213 do { 214 all = xchg_local(&info->messages, 0); 215 216 #ifdef __BIG_ENDIAN 217 if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNCTION))) 218 generic_smp_call_function_interrupt(); 219 if (all & (1 << (24 - 8 * PPC_MSG_RESCHEDULE))) 220 scheduler_ipi(); 221 if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNC_SINGLE))) 222 generic_smp_call_function_single_interrupt(); 223 if (all & (1 << (24 - 8 * PPC_MSG_DEBUGGER_BREAK))) 224 debug_ipi_action(0, NULL); 225 #else 226 #error Unsupported ENDIAN 227 #endif 228 } while (info->messages); 229 230 return IRQ_HANDLED; 231 } 232 #endif /* CONFIG_PPC_SMP_MUXED_IPI */ 233 234 static inline void do_message_pass(int cpu, int msg) 235 { 236 if (smp_ops->message_pass) 237 smp_ops->message_pass(cpu, msg); 238 #ifdef CONFIG_PPC_SMP_MUXED_IPI 239 else 240 smp_muxed_ipi_message_pass(cpu, msg); 241 #endif 242 } 243 244 void smp_send_reschedule(int cpu) 245 { 246 if (likely(smp_ops)) 247 do_message_pass(cpu, PPC_MSG_RESCHEDULE); 248 } 249 EXPORT_SYMBOL_GPL(smp_send_reschedule); 250 251 void arch_send_call_function_single_ipi(int cpu) 252 { 253 do_message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE); 254 } 255 256 void arch_send_call_function_ipi_mask(const struct cpumask *mask) 257 { 258 unsigned int cpu; 259 260 for_each_cpu(cpu, mask) 261 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION); 262 } 263 264 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC) 265 void smp_send_debugger_break(void) 266 { 267 int cpu; 268 int me = raw_smp_processor_id(); 269 270 if (unlikely(!smp_ops)) 271 return; 272 273 for_each_online_cpu(cpu) 274 if (cpu != me) 275 do_message_pass(cpu, PPC_MSG_DEBUGGER_BREAK); 276 } 277 #endif 278 279 #ifdef CONFIG_KEXEC 280 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *)) 281 { 282 crash_ipi_function_ptr = crash_ipi_callback; 283 if (crash_ipi_callback) { 284 mb(); 285 smp_send_debugger_break(); 286 } 287 } 288 #endif 289 290 static void stop_this_cpu(void *dummy) 291 { 292 /* Remove this CPU */ 293 set_cpu_online(smp_processor_id(), false); 294 295 local_irq_disable(); 296 while (1) 297 ; 298 } 299 300 void smp_send_stop(void) 301 { 302 smp_call_function(stop_this_cpu, NULL, 0); 303 } 304 305 struct thread_info *current_set[NR_CPUS]; 306 307 static void __devinit smp_store_cpu_info(int id) 308 { 309 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR); 310 #ifdef CONFIG_PPC_FSL_BOOK3E 311 per_cpu(next_tlbcam_idx, id) 312 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1; 313 #endif 314 } 315 316 void __init smp_prepare_cpus(unsigned int max_cpus) 317 { 318 unsigned int cpu; 319 320 DBG("smp_prepare_cpus\n"); 321 322 /* 323 * setup_cpu may need to be called on the boot cpu. We havent 324 * spun any cpus up but lets be paranoid. 325 */ 326 BUG_ON(boot_cpuid != smp_processor_id()); 327 328 /* Fixup boot cpu */ 329 smp_store_cpu_info(boot_cpuid); 330 cpu_callin_map[boot_cpuid] = 1; 331 332 for_each_possible_cpu(cpu) { 333 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu), 334 GFP_KERNEL, cpu_to_node(cpu)); 335 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu), 336 GFP_KERNEL, cpu_to_node(cpu)); 337 } 338 339 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid)); 340 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid)); 341 342 if (smp_ops) 343 if (smp_ops->probe) 344 max_cpus = smp_ops->probe(); 345 else 346 max_cpus = NR_CPUS; 347 else 348 max_cpus = 1; 349 } 350 351 void __devinit smp_prepare_boot_cpu(void) 352 { 353 BUG_ON(smp_processor_id() != boot_cpuid); 354 #ifdef CONFIG_PPC64 355 paca[boot_cpuid].__current = current; 356 #endif 357 current_set[boot_cpuid] = task_thread_info(current); 358 } 359 360 #ifdef CONFIG_HOTPLUG_CPU 361 362 int generic_cpu_disable(void) 363 { 364 unsigned int cpu = smp_processor_id(); 365 366 if (cpu == boot_cpuid) 367 return -EBUSY; 368 369 set_cpu_online(cpu, false); 370 #ifdef CONFIG_PPC64 371 vdso_data->processorCount--; 372 #endif 373 migrate_irqs(); 374 return 0; 375 } 376 377 void generic_cpu_die(unsigned int cpu) 378 { 379 int i; 380 381 for (i = 0; i < 100; i++) { 382 smp_rmb(); 383 if (per_cpu(cpu_state, cpu) == CPU_DEAD) 384 return; 385 msleep(100); 386 } 387 printk(KERN_ERR "CPU%d didn't die...\n", cpu); 388 } 389 390 void generic_mach_cpu_die(void) 391 { 392 unsigned int cpu; 393 394 local_irq_disable(); 395 idle_task_exit(); 396 cpu = smp_processor_id(); 397 printk(KERN_DEBUG "CPU%d offline\n", cpu); 398 __get_cpu_var(cpu_state) = CPU_DEAD; 399 smp_wmb(); 400 while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE) 401 cpu_relax(); 402 } 403 404 void generic_set_cpu_dead(unsigned int cpu) 405 { 406 per_cpu(cpu_state, cpu) = CPU_DEAD; 407 } 408 409 int generic_check_cpu_restart(unsigned int cpu) 410 { 411 return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE; 412 } 413 #endif 414 415 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle) 416 { 417 struct thread_info *ti = task_thread_info(idle); 418 419 #ifdef CONFIG_PPC64 420 paca[cpu].__current = idle; 421 paca[cpu].kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD; 422 #endif 423 ti->cpu = cpu; 424 secondary_ti = current_set[cpu] = ti; 425 } 426 427 int __cpuinit __cpu_up(unsigned int cpu, struct task_struct *tidle) 428 { 429 int rc, c; 430 431 if (smp_ops == NULL || 432 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu))) 433 return -EINVAL; 434 435 cpu_idle_thread_init(cpu, tidle); 436 437 /* Make sure callin-map entry is 0 (can be leftover a CPU 438 * hotplug 439 */ 440 cpu_callin_map[cpu] = 0; 441 442 /* The information for processor bringup must 443 * be written out to main store before we release 444 * the processor. 445 */ 446 smp_mb(); 447 448 /* wake up cpus */ 449 DBG("smp: kicking cpu %d\n", cpu); 450 rc = smp_ops->kick_cpu(cpu); 451 if (rc) { 452 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc); 453 return rc; 454 } 455 456 /* 457 * wait to see if the cpu made a callin (is actually up). 458 * use this value that I found through experimentation. 459 * -- Cort 460 */ 461 if (system_state < SYSTEM_RUNNING) 462 for (c = 50000; c && !cpu_callin_map[cpu]; c--) 463 udelay(100); 464 #ifdef CONFIG_HOTPLUG_CPU 465 else 466 /* 467 * CPUs can take much longer to come up in the 468 * hotplug case. Wait five seconds. 469 */ 470 for (c = 5000; c && !cpu_callin_map[cpu]; c--) 471 msleep(1); 472 #endif 473 474 if (!cpu_callin_map[cpu]) { 475 printk(KERN_ERR "Processor %u is stuck.\n", cpu); 476 return -ENOENT; 477 } 478 479 DBG("Processor %u found.\n", cpu); 480 481 if (smp_ops->give_timebase) 482 smp_ops->give_timebase(); 483 484 /* Wait until cpu puts itself in the online map */ 485 while (!cpu_online(cpu)) 486 cpu_relax(); 487 488 return 0; 489 } 490 491 /* Return the value of the reg property corresponding to the given 492 * logical cpu. 493 */ 494 int cpu_to_core_id(int cpu) 495 { 496 struct device_node *np; 497 const int *reg; 498 int id = -1; 499 500 np = of_get_cpu_node(cpu, NULL); 501 if (!np) 502 goto out; 503 504 reg = of_get_property(np, "reg", NULL); 505 if (!reg) 506 goto out; 507 508 id = *reg; 509 out: 510 of_node_put(np); 511 return id; 512 } 513 514 /* Helper routines for cpu to core mapping */ 515 int cpu_core_index_of_thread(int cpu) 516 { 517 return cpu >> threads_shift; 518 } 519 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread); 520 521 int cpu_first_thread_of_core(int core) 522 { 523 return core << threads_shift; 524 } 525 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core); 526 527 /* Must be called when no change can occur to cpu_present_mask, 528 * i.e. during cpu online or offline. 529 */ 530 static struct device_node *cpu_to_l2cache(int cpu) 531 { 532 struct device_node *np; 533 struct device_node *cache; 534 535 if (!cpu_present(cpu)) 536 return NULL; 537 538 np = of_get_cpu_node(cpu, NULL); 539 if (np == NULL) 540 return NULL; 541 542 cache = of_find_next_cache_node(np); 543 544 of_node_put(np); 545 546 return cache; 547 } 548 549 /* Activate a secondary processor. */ 550 void __devinit start_secondary(void *unused) 551 { 552 unsigned int cpu = smp_processor_id(); 553 struct device_node *l2_cache; 554 int i, base; 555 556 atomic_inc(&init_mm.mm_count); 557 current->active_mm = &init_mm; 558 559 smp_store_cpu_info(cpu); 560 set_dec(tb_ticks_per_jiffy); 561 preempt_disable(); 562 cpu_callin_map[cpu] = 1; 563 564 if (smp_ops->setup_cpu) 565 smp_ops->setup_cpu(cpu); 566 if (smp_ops->take_timebase) 567 smp_ops->take_timebase(); 568 569 secondary_cpu_time_init(); 570 571 #ifdef CONFIG_PPC64 572 if (system_state == SYSTEM_RUNNING) 573 vdso_data->processorCount++; 574 575 vdso_getcpu_init(); 576 #endif 577 notify_cpu_starting(cpu); 578 set_cpu_online(cpu, true); 579 /* Update sibling maps */ 580 base = cpu_first_thread_sibling(cpu); 581 for (i = 0; i < threads_per_core; i++) { 582 if (cpu_is_offline(base + i)) 583 continue; 584 cpumask_set_cpu(cpu, cpu_sibling_mask(base + i)); 585 cpumask_set_cpu(base + i, cpu_sibling_mask(cpu)); 586 587 /* cpu_core_map should be a superset of 588 * cpu_sibling_map even if we don't have cache 589 * information, so update the former here, too. 590 */ 591 cpumask_set_cpu(cpu, cpu_core_mask(base + i)); 592 cpumask_set_cpu(base + i, cpu_core_mask(cpu)); 593 } 594 l2_cache = cpu_to_l2cache(cpu); 595 for_each_online_cpu(i) { 596 struct device_node *np = cpu_to_l2cache(i); 597 if (!np) 598 continue; 599 if (np == l2_cache) { 600 cpumask_set_cpu(cpu, cpu_core_mask(i)); 601 cpumask_set_cpu(i, cpu_core_mask(cpu)); 602 } 603 of_node_put(np); 604 } 605 of_node_put(l2_cache); 606 607 local_irq_enable(); 608 609 cpu_idle(); 610 611 BUG(); 612 } 613 614 int setup_profiling_timer(unsigned int multiplier) 615 { 616 return 0; 617 } 618 619 void __init smp_cpus_done(unsigned int max_cpus) 620 { 621 cpumask_var_t old_mask; 622 623 /* We want the setup_cpu() here to be called from CPU 0, but our 624 * init thread may have been "borrowed" by another CPU in the meantime 625 * se we pin us down to CPU 0 for a short while 626 */ 627 alloc_cpumask_var(&old_mask, GFP_NOWAIT); 628 cpumask_copy(old_mask, tsk_cpus_allowed(current)); 629 set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid)); 630 631 if (smp_ops && smp_ops->setup_cpu) 632 smp_ops->setup_cpu(boot_cpuid); 633 634 set_cpus_allowed_ptr(current, old_mask); 635 636 free_cpumask_var(old_mask); 637 638 if (smp_ops && smp_ops->bringup_done) 639 smp_ops->bringup_done(); 640 641 dump_numa_cpu_topology(); 642 643 } 644 645 int arch_sd_sibling_asym_packing(void) 646 { 647 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) { 648 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n"); 649 return SD_ASYM_PACKING; 650 } 651 return 0; 652 } 653 654 #ifdef CONFIG_HOTPLUG_CPU 655 int __cpu_disable(void) 656 { 657 struct device_node *l2_cache; 658 int cpu = smp_processor_id(); 659 int base, i; 660 int err; 661 662 if (!smp_ops->cpu_disable) 663 return -ENOSYS; 664 665 err = smp_ops->cpu_disable(); 666 if (err) 667 return err; 668 669 /* Update sibling maps */ 670 base = cpu_first_thread_sibling(cpu); 671 for (i = 0; i < threads_per_core; i++) { 672 cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i)); 673 cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu)); 674 cpumask_clear_cpu(cpu, cpu_core_mask(base + i)); 675 cpumask_clear_cpu(base + i, cpu_core_mask(cpu)); 676 } 677 678 l2_cache = cpu_to_l2cache(cpu); 679 for_each_present_cpu(i) { 680 struct device_node *np = cpu_to_l2cache(i); 681 if (!np) 682 continue; 683 if (np == l2_cache) { 684 cpumask_clear_cpu(cpu, cpu_core_mask(i)); 685 cpumask_clear_cpu(i, cpu_core_mask(cpu)); 686 } 687 of_node_put(np); 688 } 689 of_node_put(l2_cache); 690 691 692 return 0; 693 } 694 695 void __cpu_die(unsigned int cpu) 696 { 697 if (smp_ops->cpu_die) 698 smp_ops->cpu_die(cpu); 699 } 700 701 static DEFINE_MUTEX(powerpc_cpu_hotplug_driver_mutex); 702 703 void cpu_hotplug_driver_lock() 704 { 705 mutex_lock(&powerpc_cpu_hotplug_driver_mutex); 706 } 707 708 void cpu_hotplug_driver_unlock() 709 { 710 mutex_unlock(&powerpc_cpu_hotplug_driver_mutex); 711 } 712 713 void cpu_die(void) 714 { 715 if (ppc_md.cpu_die) 716 ppc_md.cpu_die(); 717 718 /* If we return, we re-enter start_secondary */ 719 start_secondary_resume(); 720 } 721 722 #endif 723