1 /* 2 ** SMP Support 3 ** 4 ** Copyright (C) 1999 Walt Drummond <drummond@valinux.com> 5 ** Copyright (C) 1999 David Mosberger-Tang <davidm@hpl.hp.com> 6 ** Copyright (C) 2001,2004 Grant Grundler <grundler@parisc-linux.org> 7 ** 8 ** Lots of stuff stolen from arch/alpha/kernel/smp.c 9 ** ...and then parisc stole from arch/ia64/kernel/smp.c. Thanks David! :^) 10 ** 11 ** Thanks to John Curry and Ullas Ponnadi. I learned alot from their work. 12 ** -grant (1/12/2001) 13 ** 14 ** This program is free software; you can redistribute it and/or modify 15 ** it under the terms of the GNU General Public License as published by 16 ** the Free Software Foundation; either version 2 of the License, or 17 ** (at your option) any later version. 18 */ 19 #undef ENTRY_SYS_CPUS /* syscall support for iCOD-like functionality */ 20 21 #include <linux/config.h> 22 23 #include <linux/types.h> 24 #include <linux/spinlock.h> 25 #include <linux/slab.h> 26 27 #include <linux/kernel.h> 28 #include <linux/module.h> 29 #include <linux/sched.h> 30 #include <linux/init.h> 31 #include <linux/interrupt.h> 32 #include <linux/smp.h> 33 #include <linux/kernel_stat.h> 34 #include <linux/mm.h> 35 #include <linux/delay.h> 36 #include <linux/bitops.h> 37 38 #include <asm/system.h> 39 #include <asm/atomic.h> 40 #include <asm/current.h> 41 #include <asm/delay.h> 42 #include <asm/pgalloc.h> /* for flush_tlb_all() proto/macro */ 43 44 #include <asm/io.h> 45 #include <asm/irq.h> /* for CPU_IRQ_REGION and friends */ 46 #include <asm/mmu_context.h> 47 #include <asm/page.h> 48 #include <asm/pgtable.h> 49 #include <asm/pgalloc.h> 50 #include <asm/processor.h> 51 #include <asm/ptrace.h> 52 #include <asm/unistd.h> 53 #include <asm/cacheflush.h> 54 55 #define kDEBUG 0 56 57 DEFINE_SPINLOCK(smp_lock); 58 59 volatile struct task_struct *smp_init_current_idle_task; 60 61 static volatile int cpu_now_booting = 0; /* track which CPU is booting */ 62 63 static int parisc_max_cpus = 1; 64 65 /* online cpus are ones that we've managed to bring up completely 66 * possible cpus are all valid cpu 67 * present cpus are all detected cpu 68 * 69 * On startup we bring up the "possible" cpus. Since we discover 70 * CPUs later, we add them as hotplug, so the possible cpu mask is 71 * empty in the beginning. 72 */ 73 74 cpumask_t cpu_online_map = CPU_MASK_NONE; /* Bitmap of online CPUs */ 75 cpumask_t cpu_possible_map = CPU_MASK_ALL; /* Bitmap of Present CPUs */ 76 77 EXPORT_SYMBOL(cpu_online_map); 78 EXPORT_SYMBOL(cpu_possible_map); 79 80 81 struct smp_call_struct { 82 void (*func) (void *info); 83 void *info; 84 long wait; 85 atomic_t unstarted_count; 86 atomic_t unfinished_count; 87 }; 88 static volatile struct smp_call_struct *smp_call_function_data; 89 90 enum ipi_message_type { 91 IPI_NOP=0, 92 IPI_RESCHEDULE=1, 93 IPI_CALL_FUNC, 94 IPI_CPU_START, 95 IPI_CPU_STOP, 96 IPI_CPU_TEST 97 }; 98 99 100 /********** SMP inter processor interrupt and communication routines */ 101 102 #undef PER_CPU_IRQ_REGION 103 #ifdef PER_CPU_IRQ_REGION 104 /* XXX REVISIT Ignore for now. 105 ** *May* need this "hook" to register IPI handler 106 ** once we have perCPU ExtIntr switch tables. 107 */ 108 static void 109 ipi_init(int cpuid) 110 { 111 112 /* If CPU is present ... */ 113 #ifdef ENTRY_SYS_CPUS 114 /* *and* running (not stopped) ... */ 115 #error iCOD support wants state checked here. 116 #endif 117 118 #error verify IRQ_OFFSET(IPI_IRQ) is ipi_interrupt() in new IRQ region 119 120 if(cpu_online(cpuid) ) 121 { 122 switch_to_idle_task(current); 123 } 124 125 return; 126 } 127 #endif 128 129 130 /* 131 ** Yoink this CPU from the runnable list... 132 ** 133 */ 134 static void 135 halt_processor(void) 136 { 137 #ifdef ENTRY_SYS_CPUS 138 #error halt_processor() needs rework 139 /* 140 ** o migrate I/O interrupts off this CPU. 141 ** o leave IPI enabled - __cli() will disable IPI. 142 ** o leave CPU in online map - just change the state 143 */ 144 cpu_data[this_cpu].state = STATE_STOPPED; 145 mark_bh(IPI_BH); 146 #else 147 /* REVISIT : redirect I/O Interrupts to another CPU? */ 148 /* REVISIT : does PM *know* this CPU isn't available? */ 149 cpu_clear(smp_processor_id(), cpu_online_map); 150 local_irq_disable(); 151 for (;;) 152 ; 153 #endif 154 } 155 156 157 irqreturn_t 158 ipi_interrupt(int irq, void *dev_id, struct pt_regs *regs) 159 { 160 int this_cpu = smp_processor_id(); 161 struct cpuinfo_parisc *p = &cpu_data[this_cpu]; 162 unsigned long ops; 163 unsigned long flags; 164 165 /* Count this now; we may make a call that never returns. */ 166 p->ipi_count++; 167 168 mb(); /* Order interrupt and bit testing. */ 169 170 for (;;) { 171 spin_lock_irqsave(&(p->lock),flags); 172 ops = p->pending_ipi; 173 p->pending_ipi = 0; 174 spin_unlock_irqrestore(&(p->lock),flags); 175 176 mb(); /* Order bit clearing and data access. */ 177 178 if (!ops) 179 break; 180 181 while (ops) { 182 unsigned long which = ffz(~ops); 183 184 ops &= ~(1 << which); 185 186 switch (which) { 187 case IPI_NOP: 188 #if (kDEBUG>=100) 189 printk(KERN_DEBUG "CPU%d IPI_NOP\n",this_cpu); 190 #endif /* kDEBUG */ 191 break; 192 193 case IPI_RESCHEDULE: 194 #if (kDEBUG>=100) 195 printk(KERN_DEBUG "CPU%d IPI_RESCHEDULE\n",this_cpu); 196 #endif /* kDEBUG */ 197 /* 198 * Reschedule callback. Everything to be 199 * done is done by the interrupt return path. 200 */ 201 break; 202 203 case IPI_CALL_FUNC: 204 #if (kDEBUG>=100) 205 printk(KERN_DEBUG "CPU%d IPI_CALL_FUNC\n",this_cpu); 206 #endif /* kDEBUG */ 207 { 208 volatile struct smp_call_struct *data; 209 void (*func)(void *info); 210 void *info; 211 int wait; 212 213 data = smp_call_function_data; 214 func = data->func; 215 info = data->info; 216 wait = data->wait; 217 218 mb(); 219 atomic_dec ((atomic_t *)&data->unstarted_count); 220 221 /* At this point, *data can't 222 * be relied upon. 223 */ 224 225 (*func)(info); 226 227 /* Notify the sending CPU that the 228 * task is done. 229 */ 230 mb(); 231 if (wait) 232 atomic_dec ((atomic_t *)&data->unfinished_count); 233 } 234 break; 235 236 case IPI_CPU_START: 237 #if (kDEBUG>=100) 238 printk(KERN_DEBUG "CPU%d IPI_CPU_START\n",this_cpu); 239 #endif /* kDEBUG */ 240 #ifdef ENTRY_SYS_CPUS 241 p->state = STATE_RUNNING; 242 #endif 243 break; 244 245 case IPI_CPU_STOP: 246 #if (kDEBUG>=100) 247 printk(KERN_DEBUG "CPU%d IPI_CPU_STOP\n",this_cpu); 248 #endif /* kDEBUG */ 249 #ifdef ENTRY_SYS_CPUS 250 #else 251 halt_processor(); 252 #endif 253 break; 254 255 case IPI_CPU_TEST: 256 #if (kDEBUG>=100) 257 printk(KERN_DEBUG "CPU%d is alive!\n",this_cpu); 258 #endif /* kDEBUG */ 259 break; 260 261 default: 262 printk(KERN_CRIT "Unknown IPI num on CPU%d: %lu\n", 263 this_cpu, which); 264 return IRQ_NONE; 265 } /* Switch */ 266 } /* while (ops) */ 267 } 268 return IRQ_HANDLED; 269 } 270 271 272 static inline void 273 ipi_send(int cpu, enum ipi_message_type op) 274 { 275 struct cpuinfo_parisc *p = &cpu_data[cpu]; 276 unsigned long flags; 277 278 spin_lock_irqsave(&(p->lock),flags); 279 p->pending_ipi |= 1 << op; 280 gsc_writel(IPI_IRQ - CPU_IRQ_BASE, cpu_data[cpu].hpa); 281 spin_unlock_irqrestore(&(p->lock),flags); 282 } 283 284 285 static inline void 286 send_IPI_single(int dest_cpu, enum ipi_message_type op) 287 { 288 if (dest_cpu == NO_PROC_ID) { 289 BUG(); 290 return; 291 } 292 293 ipi_send(dest_cpu, op); 294 } 295 296 static inline void 297 send_IPI_allbutself(enum ipi_message_type op) 298 { 299 int i; 300 301 for (i = 0; i < NR_CPUS; i++) { 302 if (cpu_online(i) && i != smp_processor_id()) 303 send_IPI_single(i, op); 304 } 305 } 306 307 308 inline void 309 smp_send_stop(void) { send_IPI_allbutself(IPI_CPU_STOP); } 310 311 static inline void 312 smp_send_start(void) { send_IPI_allbutself(IPI_CPU_START); } 313 314 void 315 smp_send_reschedule(int cpu) { send_IPI_single(cpu, IPI_RESCHEDULE); } 316 317 void 318 smp_send_all_nop(void) 319 { 320 send_IPI_allbutself(IPI_NOP); 321 } 322 323 324 /** 325 * Run a function on all other CPUs. 326 * <func> The function to run. This must be fast and non-blocking. 327 * <info> An arbitrary pointer to pass to the function. 328 * <retry> If true, keep retrying until ready. 329 * <wait> If true, wait until function has completed on other CPUs. 330 * [RETURNS] 0 on success, else a negative status code. 331 * 332 * Does not return until remote CPUs are nearly ready to execute <func> 333 * or have executed. 334 */ 335 336 int 337 smp_call_function (void (*func) (void *info), void *info, int retry, int wait) 338 { 339 struct smp_call_struct data; 340 unsigned long timeout; 341 static DEFINE_SPINLOCK(lock); 342 int retries = 0; 343 344 if (num_online_cpus() < 2) 345 return 0; 346 347 /* Can deadlock when called with interrupts disabled */ 348 WARN_ON(irqs_disabled()); 349 350 /* can also deadlock if IPIs are disabled */ 351 WARN_ON((get_eiem() & (1UL<<(CPU_IRQ_MAX - IPI_IRQ))) == 0); 352 353 354 data.func = func; 355 data.info = info; 356 data.wait = wait; 357 atomic_set(&data.unstarted_count, num_online_cpus() - 1); 358 atomic_set(&data.unfinished_count, num_online_cpus() - 1); 359 360 if (retry) { 361 spin_lock (&lock); 362 while (smp_call_function_data != 0) 363 barrier(); 364 } 365 else { 366 spin_lock (&lock); 367 if (smp_call_function_data) { 368 spin_unlock (&lock); 369 return -EBUSY; 370 } 371 } 372 373 smp_call_function_data = &data; 374 spin_unlock (&lock); 375 376 /* Send a message to all other CPUs and wait for them to respond */ 377 send_IPI_allbutself(IPI_CALL_FUNC); 378 379 retry: 380 /* Wait for response */ 381 timeout = jiffies + HZ; 382 while ( (atomic_read (&data.unstarted_count) > 0) && 383 time_before (jiffies, timeout) ) 384 barrier (); 385 386 if (atomic_read (&data.unstarted_count) > 0) { 387 printk(KERN_CRIT "SMP CALL FUNCTION TIMED OUT! (cpu=%d), try %d\n", 388 smp_processor_id(), ++retries); 389 goto retry; 390 } 391 /* We either got one or timed out. Release the lock */ 392 393 mb(); 394 smp_call_function_data = NULL; 395 396 while (wait && atomic_read (&data.unfinished_count) > 0) 397 barrier (); 398 399 return 0; 400 } 401 402 EXPORT_SYMBOL(smp_call_function); 403 404 /* 405 * Flush all other CPU's tlb and then mine. Do this with on_each_cpu() 406 * as we want to ensure all TLB's flushed before proceeding. 407 */ 408 409 extern void flush_tlb_all_local(void); 410 411 void 412 smp_flush_tlb_all(void) 413 { 414 on_each_cpu((void (*)(void *))flush_tlb_all_local, NULL, 1, 1); 415 } 416 417 418 void 419 smp_do_timer(struct pt_regs *regs) 420 { 421 int cpu = smp_processor_id(); 422 struct cpuinfo_parisc *data = &cpu_data[cpu]; 423 424 if (!--data->prof_counter) { 425 data->prof_counter = data->prof_multiplier; 426 update_process_times(user_mode(regs)); 427 } 428 } 429 430 /* 431 * Called by secondaries to update state and initialize CPU registers. 432 */ 433 static void __init 434 smp_cpu_init(int cpunum) 435 { 436 extern int init_per_cpu(int); /* arch/parisc/kernel/setup.c */ 437 extern void init_IRQ(void); /* arch/parisc/kernel/irq.c */ 438 439 /* Set modes and Enable floating point coprocessor */ 440 (void) init_per_cpu(cpunum); 441 442 disable_sr_hashing(); 443 444 mb(); 445 446 /* Well, support 2.4 linux scheme as well. */ 447 if (cpu_test_and_set(cpunum, cpu_online_map)) 448 { 449 extern void machine_halt(void); /* arch/parisc.../process.c */ 450 451 printk(KERN_CRIT "CPU#%d already initialized!\n", cpunum); 452 machine_halt(); 453 } 454 455 /* Initialise the idle task for this CPU */ 456 atomic_inc(&init_mm.mm_count); 457 current->active_mm = &init_mm; 458 if(current->mm) 459 BUG(); 460 enter_lazy_tlb(&init_mm, current); 461 462 init_IRQ(); /* make sure no IRQ's are enabled or pending */ 463 } 464 465 466 /* 467 * Slaves start using C here. Indirectly called from smp_slave_stext. 468 * Do what start_kernel() and main() do for boot strap processor (aka monarch) 469 */ 470 void __init smp_callin(void) 471 { 472 int slave_id = cpu_now_booting; 473 #if 0 474 void *istack; 475 #endif 476 477 smp_cpu_init(slave_id); 478 preempt_disable(); 479 480 #if 0 /* NOT WORKING YET - see entry.S */ 481 istack = (void *)__get_free_pages(GFP_KERNEL,ISTACK_ORDER); 482 if (istack == NULL) { 483 printk(KERN_CRIT "Failed to allocate interrupt stack for cpu %d\n",slave_id); 484 BUG(); 485 } 486 mtctl(istack,31); 487 #endif 488 489 flush_cache_all_local(); /* start with known state */ 490 flush_tlb_all_local(); 491 492 local_irq_enable(); /* Interrupts have been off until now */ 493 494 cpu_idle(); /* Wait for timer to schedule some work */ 495 496 /* NOTREACHED */ 497 panic("smp_callin() AAAAaaaaahhhh....\n"); 498 } 499 500 /* 501 * Bring one cpu online. 502 */ 503 int __init smp_boot_one_cpu(int cpuid) 504 { 505 struct task_struct *idle; 506 long timeout; 507 508 /* 509 * Create an idle task for this CPU. Note the address wed* give 510 * to kernel_thread is irrelevant -- it's going to start 511 * where OS_BOOT_RENDEVZ vector in SAL says to start. But 512 * this gets all the other task-y sort of data structures set 513 * up like we wish. We need to pull the just created idle task 514 * off the run queue and stuff it into the init_tasks[] array. 515 * Sheesh . . . 516 */ 517 518 idle = fork_idle(cpuid); 519 if (IS_ERR(idle)) 520 panic("SMP: fork failed for CPU:%d", cpuid); 521 522 idle->thread_info->cpu = cpuid; 523 524 /* Let _start know what logical CPU we're booting 525 ** (offset into init_tasks[],cpu_data[]) 526 */ 527 cpu_now_booting = cpuid; 528 529 /* 530 ** boot strap code needs to know the task address since 531 ** it also contains the process stack. 532 */ 533 smp_init_current_idle_task = idle ; 534 mb(); 535 536 printk("Releasing cpu %d now, hpa=%lx\n", cpuid, cpu_data[cpuid].hpa); 537 538 /* 539 ** This gets PDC to release the CPU from a very tight loop. 540 ** 541 ** From the PA-RISC 2.0 Firmware Architecture Reference Specification: 542 ** "The MEM_RENDEZ vector specifies the location of OS_RENDEZ which 543 ** is executed after receiving the rendezvous signal (an interrupt to 544 ** EIR{0}). MEM_RENDEZ is valid only when it is nonzero and the 545 ** contents of memory are valid." 546 */ 547 gsc_writel(TIMER_IRQ - CPU_IRQ_BASE, cpu_data[cpuid].hpa); 548 mb(); 549 550 /* 551 * OK, wait a bit for that CPU to finish staggering about. 552 * Slave will set a bit when it reaches smp_cpu_init(). 553 * Once the "monarch CPU" sees the bit change, it can move on. 554 */ 555 for (timeout = 0; timeout < 10000; timeout++) { 556 if(cpu_online(cpuid)) { 557 /* Which implies Slave has started up */ 558 cpu_now_booting = 0; 559 smp_init_current_idle_task = NULL; 560 goto alive ; 561 } 562 udelay(100); 563 barrier(); 564 } 565 566 put_task_struct(idle); 567 idle = NULL; 568 569 printk(KERN_CRIT "SMP: CPU:%d is stuck.\n", cpuid); 570 return -1; 571 572 alive: 573 /* Remember the Slave data */ 574 #if (kDEBUG>=100) 575 printk(KERN_DEBUG "SMP: CPU:%d came alive after %ld _us\n", 576 cpuid, timeout * 100); 577 #endif /* kDEBUG */ 578 #ifdef ENTRY_SYS_CPUS 579 cpu_data[cpuid].state = STATE_RUNNING; 580 #endif 581 return 0; 582 } 583 584 void __devinit smp_prepare_boot_cpu(void) 585 { 586 int bootstrap_processor=cpu_data[0].cpuid; /* CPU ID of BSP */ 587 588 #ifdef ENTRY_SYS_CPUS 589 cpu_data[0].state = STATE_RUNNING; 590 #endif 591 592 /* Setup BSP mappings */ 593 printk("SMP: bootstrap CPU ID is %d\n",bootstrap_processor); 594 595 cpu_set(bootstrap_processor, cpu_online_map); 596 cpu_set(bootstrap_processor, cpu_present_map); 597 } 598 599 600 601 /* 602 ** inventory.c:do_inventory() hasn't yet been run and thus we 603 ** don't 'discover' the additional CPU's until later. 604 */ 605 void __init smp_prepare_cpus(unsigned int max_cpus) 606 { 607 cpus_clear(cpu_present_map); 608 cpu_set(0, cpu_present_map); 609 610 parisc_max_cpus = max_cpus; 611 if (!max_cpus) 612 printk(KERN_INFO "SMP mode deactivated.\n"); 613 } 614 615 616 void smp_cpus_done(unsigned int cpu_max) 617 { 618 return; 619 } 620 621 622 int __devinit __cpu_up(unsigned int cpu) 623 { 624 if (cpu != 0 && cpu < parisc_max_cpus) 625 smp_boot_one_cpu(cpu); 626 627 return cpu_online(cpu) ? 0 : -ENOSYS; 628 } 629 630 631 632 #ifdef ENTRY_SYS_CPUS 633 /* Code goes along with: 634 ** entry.s: ENTRY_NAME(sys_cpus) / * 215, for cpu stat * / 635 */ 636 int sys_cpus(int argc, char **argv) 637 { 638 int i,j=0; 639 extern int current_pid(int cpu); 640 641 if( argc > 2 ) { 642 printk("sys_cpus:Only one argument supported\n"); 643 return (-1); 644 } 645 if ( argc == 1 ){ 646 647 #ifdef DUMP_MORE_STATE 648 for(i=0; i<NR_CPUS; i++) { 649 int cpus_per_line = 4; 650 if(cpu_online(i)) { 651 if (j++ % cpus_per_line) 652 printk(" %3d",i); 653 else 654 printk("\n %3d",i); 655 } 656 } 657 printk("\n"); 658 #else 659 printk("\n 0\n"); 660 #endif 661 } else if((argc==2) && !(strcmp(argv[1],"-l"))) { 662 printk("\nCPUSTATE TASK CPUNUM CPUID HARDCPU(HPA)\n"); 663 #ifdef DUMP_MORE_STATE 664 for(i=0;i<NR_CPUS;i++) { 665 if (!cpu_online(i)) 666 continue; 667 if (cpu_data[i].cpuid != NO_PROC_ID) { 668 switch(cpu_data[i].state) { 669 case STATE_RENDEZVOUS: 670 printk("RENDEZVS "); 671 break; 672 case STATE_RUNNING: 673 printk((current_pid(i)!=0) ? "RUNNING " : "IDLING "); 674 break; 675 case STATE_STOPPED: 676 printk("STOPPED "); 677 break; 678 case STATE_HALTED: 679 printk("HALTED "); 680 break; 681 default: 682 printk("%08x?", cpu_data[i].state); 683 break; 684 } 685 if(cpu_online(i)) { 686 printk(" %4d",current_pid(i)); 687 } 688 printk(" %6d",cpu_number_map(i)); 689 printk(" %5d",i); 690 printk(" 0x%lx\n",cpu_data[i].hpa); 691 } 692 } 693 #else 694 printk("\n%s %4d 0 0 --------", 695 (current->pid)?"RUNNING ": "IDLING ",current->pid); 696 #endif 697 } else if ((argc==2) && !(strcmp(argv[1],"-s"))) { 698 #ifdef DUMP_MORE_STATE 699 printk("\nCPUSTATE CPUID\n"); 700 for (i=0;i<NR_CPUS;i++) { 701 if (!cpu_online(i)) 702 continue; 703 if (cpu_data[i].cpuid != NO_PROC_ID) { 704 switch(cpu_data[i].state) { 705 case STATE_RENDEZVOUS: 706 printk("RENDEZVS");break; 707 case STATE_RUNNING: 708 printk((current_pid(i)!=0) ? "RUNNING " : "IDLING"); 709 break; 710 case STATE_STOPPED: 711 printk("STOPPED ");break; 712 case STATE_HALTED: 713 printk("HALTED ");break; 714 default: 715 } 716 printk(" %5d\n",i); 717 } 718 } 719 #else 720 printk("\n%s CPU0",(current->pid==0)?"RUNNING ":"IDLING "); 721 #endif 722 } else { 723 printk("sys_cpus:Unknown request\n"); 724 return (-1); 725 } 726 return 0; 727 } 728 #endif /* ENTRY_SYS_CPUS */ 729 730 #ifdef CONFIG_PROC_FS 731 int __init 732 setup_profiling_timer(unsigned int multiplier) 733 { 734 return -EINVAL; 735 } 736 #endif 737