1 /* 2 * linux/arch/arm/kernel/smp.c 3 * 4 * Copyright (C) 2002 ARM Limited, All Rights Reserved. 5 * 6 * This program is free software; you can redistribute it and/or modify 7 * it under the terms of the GNU General Public License version 2 as 8 * published by the Free Software Foundation. 9 */ 10 #include <linux/config.h> 11 #include <linux/delay.h> 12 #include <linux/init.h> 13 #include <linux/spinlock.h> 14 #include <linux/sched.h> 15 #include <linux/interrupt.h> 16 #include <linux/cache.h> 17 #include <linux/profile.h> 18 #include <linux/errno.h> 19 #include <linux/mm.h> 20 #include <linux/cpu.h> 21 #include <linux/smp.h> 22 #include <linux/seq_file.h> 23 24 #include <asm/atomic.h> 25 #include <asm/cacheflush.h> 26 #include <asm/cpu.h> 27 #include <asm/mmu_context.h> 28 #include <asm/pgtable.h> 29 #include <asm/pgalloc.h> 30 #include <asm/processor.h> 31 #include <asm/tlbflush.h> 32 #include <asm/ptrace.h> 33 34 /* 35 * bitmask of present and online CPUs. 36 * The present bitmask indicates that the CPU is physically present. 37 * The online bitmask indicates that the CPU is up and running. 38 */ 39 cpumask_t cpu_possible_map; 40 cpumask_t cpu_online_map; 41 42 /* 43 * as from 2.5, kernels no longer have an init_tasks structure 44 * so we need some other way of telling a new secondary core 45 * where to place its SVC stack 46 */ 47 struct secondary_data secondary_data; 48 49 /* 50 * structures for inter-processor calls 51 * - A collection of single bit ipi messages. 52 */ 53 struct ipi_data { 54 spinlock_t lock; 55 unsigned long ipi_count; 56 unsigned long bits; 57 }; 58 59 static DEFINE_PER_CPU(struct ipi_data, ipi_data) = { 60 .lock = SPIN_LOCK_UNLOCKED, 61 }; 62 63 enum ipi_msg_type { 64 IPI_TIMER, 65 IPI_RESCHEDULE, 66 IPI_CALL_FUNC, 67 IPI_CPU_STOP, 68 }; 69 70 struct smp_call_struct { 71 void (*func)(void *info); 72 void *info; 73 int wait; 74 cpumask_t pending; 75 cpumask_t unfinished; 76 }; 77 78 static struct smp_call_struct * volatile smp_call_function_data; 79 static DEFINE_SPINLOCK(smp_call_function_lock); 80 81 int __cpuinit __cpu_up(unsigned int cpu) 82 { 83 struct cpuinfo_arm *ci = &per_cpu(cpu_data, cpu); 84 struct task_struct *idle = ci->idle; 85 pgd_t *pgd; 86 pmd_t *pmd; 87 int ret; 88 89 /* 90 * Spawn a new process manually, if not already done. 91 * Grab a pointer to its task struct so we can mess with it 92 */ 93 if (!idle) { 94 idle = fork_idle(cpu); 95 if (IS_ERR(idle)) { 96 printk(KERN_ERR "CPU%u: fork() failed\n", cpu); 97 return PTR_ERR(idle); 98 } 99 ci->idle = idle; 100 } 101 102 /* 103 * Allocate initial page tables to allow the new CPU to 104 * enable the MMU safely. This essentially means a set 105 * of our "standard" page tables, with the addition of 106 * a 1:1 mapping for the physical address of the kernel. 107 */ 108 pgd = pgd_alloc(&init_mm); 109 pmd = pmd_offset(pgd, PHYS_OFFSET); 110 *pmd = __pmd((PHYS_OFFSET & PGDIR_MASK) | 111 PMD_TYPE_SECT | PMD_SECT_AP_WRITE); 112 113 /* 114 * We need to tell the secondary core where to find 115 * its stack and the page tables. 116 */ 117 secondary_data.stack = (void *)idle->thread_info + THREAD_START_SP; 118 secondary_data.pgdir = virt_to_phys(pgd); 119 wmb(); 120 121 /* 122 * Now bring the CPU into our world. 123 */ 124 ret = boot_secondary(cpu, idle); 125 if (ret == 0) { 126 unsigned long timeout; 127 128 /* 129 * CPU was successfully started, wait for it 130 * to come online or time out. 131 */ 132 timeout = jiffies + HZ; 133 while (time_before(jiffies, timeout)) { 134 if (cpu_online(cpu)) 135 break; 136 137 udelay(10); 138 barrier(); 139 } 140 141 if (!cpu_online(cpu)) 142 ret = -EIO; 143 } 144 145 secondary_data.stack = 0; 146 secondary_data.pgdir = 0; 147 148 *pmd_offset(pgd, PHYS_OFFSET) = __pmd(0); 149 pgd_free(pgd); 150 151 if (ret) { 152 printk(KERN_CRIT "CPU%u: processor failed to boot\n", cpu); 153 154 /* 155 * FIXME: We need to clean up the new idle thread. --rmk 156 */ 157 } 158 159 return ret; 160 } 161 162 #ifdef CONFIG_HOTPLUG_CPU 163 /* 164 * __cpu_disable runs on the processor to be shutdown. 165 */ 166 int __cpuexit __cpu_disable(void) 167 { 168 unsigned int cpu = smp_processor_id(); 169 struct task_struct *p; 170 int ret; 171 172 ret = mach_cpu_disable(cpu); 173 if (ret) 174 return ret; 175 176 /* 177 * Take this CPU offline. Once we clear this, we can't return, 178 * and we must not schedule until we're ready to give up the cpu. 179 */ 180 cpu_clear(cpu, cpu_online_map); 181 182 /* 183 * OK - migrate IRQs away from this CPU 184 */ 185 migrate_irqs(); 186 187 /* 188 * Flush user cache and TLB mappings, and then remove this CPU 189 * from the vm mask set of all processes. 190 */ 191 flush_cache_all(); 192 local_flush_tlb_all(); 193 194 read_lock(&tasklist_lock); 195 for_each_process(p) { 196 if (p->mm) 197 cpu_clear(cpu, p->mm->cpu_vm_mask); 198 } 199 read_unlock(&tasklist_lock); 200 201 return 0; 202 } 203 204 /* 205 * called on the thread which is asking for a CPU to be shutdown - 206 * waits until shutdown has completed, or it is timed out. 207 */ 208 void __cpuexit __cpu_die(unsigned int cpu) 209 { 210 if (!platform_cpu_kill(cpu)) 211 printk("CPU%u: unable to kill\n", cpu); 212 } 213 214 /* 215 * Called from the idle thread for the CPU which has been shutdown. 216 * 217 * Note that we disable IRQs here, but do not re-enable them 218 * before returning to the caller. This is also the behaviour 219 * of the other hotplug-cpu capable cores, so presumably coming 220 * out of idle fixes this. 221 */ 222 void __cpuexit cpu_die(void) 223 { 224 unsigned int cpu = smp_processor_id(); 225 226 local_irq_disable(); 227 idle_task_exit(); 228 229 /* 230 * actual CPU shutdown procedure is at least platform (if not 231 * CPU) specific 232 */ 233 platform_cpu_die(cpu); 234 235 /* 236 * Do not return to the idle loop - jump back to the secondary 237 * cpu initialisation. There's some initialisation which needs 238 * to be repeated to undo the effects of taking the CPU offline. 239 */ 240 __asm__("mov sp, %0\n" 241 " b secondary_start_kernel" 242 : 243 : "r" ((void *)current->thread_info + THREAD_SIZE - 8)); 244 } 245 #endif /* CONFIG_HOTPLUG_CPU */ 246 247 /* 248 * This is the secondary CPU boot entry. We're using this CPUs 249 * idle thread stack, but a set of temporary page tables. 250 */ 251 asmlinkage void __cpuinit secondary_start_kernel(void) 252 { 253 struct mm_struct *mm = &init_mm; 254 unsigned int cpu = smp_processor_id(); 255 256 printk("CPU%u: Booted secondary processor\n", cpu); 257 258 /* 259 * All kernel threads share the same mm context; grab a 260 * reference and switch to it. 261 */ 262 atomic_inc(&mm->mm_users); 263 atomic_inc(&mm->mm_count); 264 current->active_mm = mm; 265 cpu_set(cpu, mm->cpu_vm_mask); 266 cpu_switch_mm(mm->pgd, mm); 267 enter_lazy_tlb(mm, current); 268 local_flush_tlb_all(); 269 270 cpu_init(); 271 272 /* 273 * Give the platform a chance to do its own initialisation. 274 */ 275 platform_secondary_init(cpu); 276 277 /* 278 * Enable local interrupts. 279 */ 280 local_irq_enable(); 281 local_fiq_enable(); 282 283 calibrate_delay(); 284 285 smp_store_cpu_info(cpu); 286 287 /* 288 * OK, now it's safe to let the boot CPU continue 289 */ 290 cpu_set(cpu, cpu_online_map); 291 292 /* 293 * OK, it's off to the idle thread for us 294 */ 295 cpu_idle(); 296 } 297 298 /* 299 * Called by both boot and secondaries to move global data into 300 * per-processor storage. 301 */ 302 void __cpuinit smp_store_cpu_info(unsigned int cpuid) 303 { 304 struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid); 305 306 cpu_info->loops_per_jiffy = loops_per_jiffy; 307 } 308 309 void __init smp_cpus_done(unsigned int max_cpus) 310 { 311 int cpu; 312 unsigned long bogosum = 0; 313 314 for_each_online_cpu(cpu) 315 bogosum += per_cpu(cpu_data, cpu).loops_per_jiffy; 316 317 printk(KERN_INFO "SMP: Total of %d processors activated " 318 "(%lu.%02lu BogoMIPS).\n", 319 num_online_cpus(), 320 bogosum / (500000/HZ), 321 (bogosum / (5000/HZ)) % 100); 322 } 323 324 void __init smp_prepare_boot_cpu(void) 325 { 326 unsigned int cpu = smp_processor_id(); 327 328 per_cpu(cpu_data, cpu).idle = current; 329 330 cpu_set(cpu, cpu_possible_map); 331 cpu_set(cpu, cpu_present_map); 332 cpu_set(cpu, cpu_online_map); 333 } 334 335 static void send_ipi_message(cpumask_t callmap, enum ipi_msg_type msg) 336 { 337 unsigned long flags; 338 unsigned int cpu; 339 340 local_irq_save(flags); 341 342 for_each_cpu_mask(cpu, callmap) { 343 struct ipi_data *ipi = &per_cpu(ipi_data, cpu); 344 345 spin_lock(&ipi->lock); 346 ipi->bits |= 1 << msg; 347 spin_unlock(&ipi->lock); 348 } 349 350 /* 351 * Call the platform specific cross-CPU call function. 352 */ 353 smp_cross_call(callmap); 354 355 local_irq_restore(flags); 356 } 357 358 /* 359 * You must not call this function with disabled interrupts, from a 360 * hardware interrupt handler, nor from a bottom half handler. 361 */ 362 int smp_call_function_on_cpu(void (*func)(void *info), void *info, int retry, 363 int wait, cpumask_t callmap) 364 { 365 struct smp_call_struct data; 366 unsigned long timeout; 367 int ret = 0; 368 369 data.func = func; 370 data.info = info; 371 data.wait = wait; 372 373 cpu_clear(smp_processor_id(), callmap); 374 if (cpus_empty(callmap)) 375 goto out; 376 377 data.pending = callmap; 378 if (wait) 379 data.unfinished = callmap; 380 381 /* 382 * try to get the mutex on smp_call_function_data 383 */ 384 spin_lock(&smp_call_function_lock); 385 smp_call_function_data = &data; 386 387 send_ipi_message(callmap, IPI_CALL_FUNC); 388 389 timeout = jiffies + HZ; 390 while (!cpus_empty(data.pending) && time_before(jiffies, timeout)) 391 barrier(); 392 393 /* 394 * did we time out? 395 */ 396 if (!cpus_empty(data.pending)) { 397 /* 398 * this may be causing our panic - report it 399 */ 400 printk(KERN_CRIT 401 "CPU%u: smp_call_function timeout for %p(%p)\n" 402 " callmap %lx pending %lx, %swait\n", 403 smp_processor_id(), func, info, *cpus_addr(callmap), 404 *cpus_addr(data.pending), wait ? "" : "no "); 405 406 /* 407 * TRACE 408 */ 409 timeout = jiffies + (5 * HZ); 410 while (!cpus_empty(data.pending) && time_before(jiffies, timeout)) 411 barrier(); 412 413 if (cpus_empty(data.pending)) 414 printk(KERN_CRIT " RESOLVED\n"); 415 else 416 printk(KERN_CRIT " STILL STUCK\n"); 417 } 418 419 /* 420 * whatever happened, we're done with the data, so release it 421 */ 422 smp_call_function_data = NULL; 423 spin_unlock(&smp_call_function_lock); 424 425 if (!cpus_empty(data.pending)) { 426 ret = -ETIMEDOUT; 427 goto out; 428 } 429 430 if (wait) 431 while (!cpus_empty(data.unfinished)) 432 barrier(); 433 out: 434 435 return 0; 436 } 437 438 int smp_call_function(void (*func)(void *info), void *info, int retry, 439 int wait) 440 { 441 return smp_call_function_on_cpu(func, info, retry, wait, 442 cpu_online_map); 443 } 444 445 void show_ipi_list(struct seq_file *p) 446 { 447 unsigned int cpu; 448 449 seq_puts(p, "IPI:"); 450 451 for_each_present_cpu(cpu) 452 seq_printf(p, " %10lu", per_cpu(ipi_data, cpu).ipi_count); 453 454 seq_putc(p, '\n'); 455 } 456 457 static void ipi_timer(struct pt_regs *regs) 458 { 459 int user = user_mode(regs); 460 461 irq_enter(); 462 profile_tick(CPU_PROFILING, regs); 463 update_process_times(user); 464 irq_exit(); 465 } 466 467 /* 468 * ipi_call_function - handle IPI from smp_call_function() 469 * 470 * Note that we copy data out of the cross-call structure and then 471 * let the caller know that we're here and have done with their data 472 */ 473 static void ipi_call_function(unsigned int cpu) 474 { 475 struct smp_call_struct *data = smp_call_function_data; 476 void (*func)(void *info) = data->func; 477 void *info = data->info; 478 int wait = data->wait; 479 480 cpu_clear(cpu, data->pending); 481 482 func(info); 483 484 if (wait) 485 cpu_clear(cpu, data->unfinished); 486 } 487 488 static DEFINE_SPINLOCK(stop_lock); 489 490 /* 491 * ipi_cpu_stop - handle IPI from smp_send_stop() 492 */ 493 static void ipi_cpu_stop(unsigned int cpu) 494 { 495 spin_lock(&stop_lock); 496 printk(KERN_CRIT "CPU%u: stopping\n", cpu); 497 dump_stack(); 498 spin_unlock(&stop_lock); 499 500 cpu_clear(cpu, cpu_online_map); 501 502 local_fiq_disable(); 503 local_irq_disable(); 504 505 while (1) 506 cpu_relax(); 507 } 508 509 /* 510 * Main handler for inter-processor interrupts 511 * 512 * For ARM, the ipimask now only identifies a single 513 * category of IPI (Bit 1 IPIs have been replaced by a 514 * different mechanism): 515 * 516 * Bit 0 - Inter-processor function call 517 */ 518 void do_IPI(struct pt_regs *regs) 519 { 520 unsigned int cpu = smp_processor_id(); 521 struct ipi_data *ipi = &per_cpu(ipi_data, cpu); 522 523 ipi->ipi_count++; 524 525 for (;;) { 526 unsigned long msgs; 527 528 spin_lock(&ipi->lock); 529 msgs = ipi->bits; 530 ipi->bits = 0; 531 spin_unlock(&ipi->lock); 532 533 if (!msgs) 534 break; 535 536 do { 537 unsigned nextmsg; 538 539 nextmsg = msgs & -msgs; 540 msgs &= ~nextmsg; 541 nextmsg = ffz(~nextmsg); 542 543 switch (nextmsg) { 544 case IPI_TIMER: 545 ipi_timer(regs); 546 break; 547 548 case IPI_RESCHEDULE: 549 /* 550 * nothing more to do - eveything is 551 * done on the interrupt return path 552 */ 553 break; 554 555 case IPI_CALL_FUNC: 556 ipi_call_function(cpu); 557 break; 558 559 case IPI_CPU_STOP: 560 ipi_cpu_stop(cpu); 561 break; 562 563 default: 564 printk(KERN_CRIT "CPU%u: Unknown IPI message 0x%x\n", 565 cpu, nextmsg); 566 break; 567 } 568 } while (msgs); 569 } 570 } 571 572 void smp_send_reschedule(int cpu) 573 { 574 send_ipi_message(cpumask_of_cpu(cpu), IPI_RESCHEDULE); 575 } 576 577 void smp_send_timer(void) 578 { 579 cpumask_t mask = cpu_online_map; 580 cpu_clear(smp_processor_id(), mask); 581 send_ipi_message(mask, IPI_TIMER); 582 } 583 584 void smp_send_stop(void) 585 { 586 cpumask_t mask = cpu_online_map; 587 cpu_clear(smp_processor_id(), mask); 588 send_ipi_message(mask, IPI_CPU_STOP); 589 } 590 591 /* 592 * not supported here 593 */ 594 int __init setup_profiling_timer(unsigned int multiplier) 595 { 596 return -EINVAL; 597 } 598 599 static int 600 on_each_cpu_mask(void (*func)(void *), void *info, int retry, int wait, 601 cpumask_t mask) 602 { 603 int ret = 0; 604 605 preempt_disable(); 606 607 ret = smp_call_function_on_cpu(func, info, retry, wait, mask); 608 if (cpu_isset(smp_processor_id(), mask)) 609 func(info); 610 611 preempt_enable(); 612 613 return ret; 614 } 615 616 /**********************************************************************/ 617 618 /* 619 * TLB operations 620 */ 621 struct tlb_args { 622 struct vm_area_struct *ta_vma; 623 unsigned long ta_start; 624 unsigned long ta_end; 625 }; 626 627 static inline void ipi_flush_tlb_all(void *ignored) 628 { 629 local_flush_tlb_all(); 630 } 631 632 static inline void ipi_flush_tlb_mm(void *arg) 633 { 634 struct mm_struct *mm = (struct mm_struct *)arg; 635 636 local_flush_tlb_mm(mm); 637 } 638 639 static inline void ipi_flush_tlb_page(void *arg) 640 { 641 struct tlb_args *ta = (struct tlb_args *)arg; 642 643 local_flush_tlb_page(ta->ta_vma, ta->ta_start); 644 } 645 646 static inline void ipi_flush_tlb_kernel_page(void *arg) 647 { 648 struct tlb_args *ta = (struct tlb_args *)arg; 649 650 local_flush_tlb_kernel_page(ta->ta_start); 651 } 652 653 static inline void ipi_flush_tlb_range(void *arg) 654 { 655 struct tlb_args *ta = (struct tlb_args *)arg; 656 657 local_flush_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end); 658 } 659 660 static inline void ipi_flush_tlb_kernel_range(void *arg) 661 { 662 struct tlb_args *ta = (struct tlb_args *)arg; 663 664 local_flush_tlb_kernel_range(ta->ta_start, ta->ta_end); 665 } 666 667 void flush_tlb_all(void) 668 { 669 on_each_cpu(ipi_flush_tlb_all, NULL, 1, 1); 670 } 671 672 void flush_tlb_mm(struct mm_struct *mm) 673 { 674 cpumask_t mask = mm->cpu_vm_mask; 675 676 on_each_cpu_mask(ipi_flush_tlb_mm, mm, 1, 1, mask); 677 } 678 679 void flush_tlb_page(struct vm_area_struct *vma, unsigned long uaddr) 680 { 681 cpumask_t mask = vma->vm_mm->cpu_vm_mask; 682 struct tlb_args ta; 683 684 ta.ta_vma = vma; 685 ta.ta_start = uaddr; 686 687 on_each_cpu_mask(ipi_flush_tlb_page, &ta, 1, 1, mask); 688 } 689 690 void flush_tlb_kernel_page(unsigned long kaddr) 691 { 692 struct tlb_args ta; 693 694 ta.ta_start = kaddr; 695 696 on_each_cpu(ipi_flush_tlb_kernel_page, &ta, 1, 1); 697 } 698 699 void flush_tlb_range(struct vm_area_struct *vma, 700 unsigned long start, unsigned long end) 701 { 702 cpumask_t mask = vma->vm_mm->cpu_vm_mask; 703 struct tlb_args ta; 704 705 ta.ta_vma = vma; 706 ta.ta_start = start; 707 ta.ta_end = end; 708 709 on_each_cpu_mask(ipi_flush_tlb_range, &ta, 1, 1, mask); 710 } 711 712 void flush_tlb_kernel_range(unsigned long start, unsigned long end) 713 { 714 struct tlb_args ta; 715 716 ta.ta_start = start; 717 ta.ta_end = end; 718 719 on_each_cpu(ipi_flush_tlb_kernel_range, &ta, 1, 1); 720 } 721