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 a lot 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 #include <linux/types.h> 20 #include <linux/spinlock.h> 21 #include <linux/slab.h> 22 23 #include <linux/kernel.h> 24 #include <linux/module.h> 25 #include <linux/sched.h> 26 #include <linux/init.h> 27 #include <linux/interrupt.h> 28 #include <linux/smp.h> 29 #include <linux/kernel_stat.h> 30 #include <linux/mm.h> 31 #include <linux/err.h> 32 #include <linux/delay.h> 33 #include <linux/bitops.h> 34 35 #include <asm/system.h> 36 #include <asm/atomic.h> 37 #include <asm/current.h> 38 #include <asm/delay.h> 39 #include <asm/tlbflush.h> 40 41 #include <asm/io.h> 42 #include <asm/irq.h> /* for CPU_IRQ_REGION and friends */ 43 #include <asm/mmu_context.h> 44 #include <asm/page.h> 45 #include <asm/pgtable.h> 46 #include <asm/pgalloc.h> 47 #include <asm/processor.h> 48 #include <asm/ptrace.h> 49 #include <asm/unistd.h> 50 #include <asm/cacheflush.h> 51 52 #undef DEBUG_SMP 53 #ifdef DEBUG_SMP 54 static int smp_debug_lvl = 0; 55 #define smp_debug(lvl, printargs...) \ 56 if (lvl >= smp_debug_lvl) \ 57 printk(printargs); 58 #else 59 #define smp_debug(lvl, ...) 60 #endif /* DEBUG_SMP */ 61 62 DEFINE_SPINLOCK(smp_lock); 63 64 volatile struct task_struct *smp_init_current_idle_task; 65 66 static volatile int cpu_now_booting __read_mostly = 0; /* track which CPU is booting */ 67 68 static int parisc_max_cpus __read_mostly = 1; 69 70 /* online cpus are ones that we've managed to bring up completely 71 * possible cpus are all valid cpu 72 * present cpus are all detected cpu 73 * 74 * On startup we bring up the "possible" cpus. Since we discover 75 * CPUs later, we add them as hotplug, so the possible cpu mask is 76 * empty in the beginning. 77 */ 78 79 cpumask_t cpu_online_map __read_mostly = CPU_MASK_NONE; /* Bitmap of online CPUs */ 80 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; /* Bitmap of Present CPUs */ 81 82 EXPORT_SYMBOL(cpu_online_map); 83 EXPORT_SYMBOL(cpu_possible_map); 84 85 DEFINE_PER_CPU(spinlock_t, ipi_lock) = SPIN_LOCK_UNLOCKED; 86 87 struct smp_call_struct { 88 void (*func) (void *info); 89 void *info; 90 long wait; 91 atomic_t unstarted_count; 92 atomic_t unfinished_count; 93 }; 94 static volatile struct smp_call_struct *smp_call_function_data; 95 96 enum ipi_message_type { 97 IPI_NOP=0, 98 IPI_RESCHEDULE=1, 99 IPI_CALL_FUNC, 100 IPI_CPU_START, 101 IPI_CPU_STOP, 102 IPI_CPU_TEST 103 }; 104 105 106 /********** SMP inter processor interrupt and communication routines */ 107 108 #undef PER_CPU_IRQ_REGION 109 #ifdef PER_CPU_IRQ_REGION 110 /* XXX REVISIT Ignore for now. 111 ** *May* need this "hook" to register IPI handler 112 ** once we have perCPU ExtIntr switch tables. 113 */ 114 static void 115 ipi_init(int cpuid) 116 { 117 #error verify IRQ_OFFSET(IPI_IRQ) is ipi_interrupt() in new IRQ region 118 119 if(cpu_online(cpuid) ) 120 { 121 switch_to_idle_task(current); 122 } 123 124 return; 125 } 126 #endif 127 128 129 /* 130 ** Yoink this CPU from the runnable list... 131 ** 132 */ 133 static void 134 halt_processor(void) 135 { 136 /* REVISIT : redirect I/O Interrupts to another CPU? */ 137 /* REVISIT : does PM *know* this CPU isn't available? */ 138 cpu_clear(smp_processor_id(), cpu_online_map); 139 local_irq_disable(); 140 for (;;) 141 ; 142 } 143 144 145 irqreturn_t 146 ipi_interrupt(int irq, void *dev_id) 147 { 148 int this_cpu = smp_processor_id(); 149 struct cpuinfo_parisc *p = &cpu_data[this_cpu]; 150 unsigned long ops; 151 unsigned long flags; 152 153 /* Count this now; we may make a call that never returns. */ 154 p->ipi_count++; 155 156 mb(); /* Order interrupt and bit testing. */ 157 158 for (;;) { 159 spinlock_t *lock = &per_cpu(ipi_lock, this_cpu); 160 spin_lock_irqsave(lock, flags); 161 ops = p->pending_ipi; 162 p->pending_ipi = 0; 163 spin_unlock_irqrestore(lock, flags); 164 165 mb(); /* Order bit clearing and data access. */ 166 167 if (!ops) 168 break; 169 170 while (ops) { 171 unsigned long which = ffz(~ops); 172 173 ops &= ~(1 << which); 174 175 switch (which) { 176 case IPI_NOP: 177 smp_debug(100, KERN_DEBUG "CPU%d IPI_NOP\n", this_cpu); 178 break; 179 180 case IPI_RESCHEDULE: 181 smp_debug(100, KERN_DEBUG "CPU%d IPI_RESCHEDULE\n", this_cpu); 182 /* 183 * Reschedule callback. Everything to be 184 * done is done by the interrupt return path. 185 */ 186 break; 187 188 case IPI_CALL_FUNC: 189 smp_debug(100, KERN_DEBUG "CPU%d IPI_CALL_FUNC\n", this_cpu); 190 { 191 volatile struct smp_call_struct *data; 192 void (*func)(void *info); 193 void *info; 194 int wait; 195 196 data = smp_call_function_data; 197 func = data->func; 198 info = data->info; 199 wait = data->wait; 200 201 mb(); 202 atomic_dec ((atomic_t *)&data->unstarted_count); 203 204 /* At this point, *data can't 205 * be relied upon. 206 */ 207 208 (*func)(info); 209 210 /* Notify the sending CPU that the 211 * task is done. 212 */ 213 mb(); 214 if (wait) 215 atomic_dec ((atomic_t *)&data->unfinished_count); 216 } 217 break; 218 219 case IPI_CPU_START: 220 smp_debug(100, KERN_DEBUG "CPU%d IPI_CPU_START\n", this_cpu); 221 break; 222 223 case IPI_CPU_STOP: 224 smp_debug(100, KERN_DEBUG "CPU%d IPI_CPU_STOP\n", this_cpu); 225 halt_processor(); 226 break; 227 228 case IPI_CPU_TEST: 229 smp_debug(100, KERN_DEBUG "CPU%d is alive!\n", this_cpu); 230 break; 231 232 default: 233 printk(KERN_CRIT "Unknown IPI num on CPU%d: %lu\n", 234 this_cpu, which); 235 return IRQ_NONE; 236 } /* Switch */ 237 /* let in any pending interrupts */ 238 local_irq_enable(); 239 local_irq_disable(); 240 } /* while (ops) */ 241 } 242 return IRQ_HANDLED; 243 } 244 245 246 static inline void 247 ipi_send(int cpu, enum ipi_message_type op) 248 { 249 struct cpuinfo_parisc *p = &cpu_data[cpu]; 250 spinlock_t *lock = &per_cpu(ipi_lock, cpu); 251 unsigned long flags; 252 253 spin_lock_irqsave(lock, flags); 254 p->pending_ipi |= 1 << op; 255 gsc_writel(IPI_IRQ - CPU_IRQ_BASE, cpu_data[cpu].hpa); 256 spin_unlock_irqrestore(lock, flags); 257 } 258 259 260 static inline void 261 send_IPI_single(int dest_cpu, enum ipi_message_type op) 262 { 263 if (dest_cpu == NO_PROC_ID) { 264 BUG(); 265 return; 266 } 267 268 ipi_send(dest_cpu, op); 269 } 270 271 static inline void 272 send_IPI_allbutself(enum ipi_message_type op) 273 { 274 int i; 275 276 for_each_online_cpu(i) { 277 if (i != smp_processor_id()) 278 send_IPI_single(i, op); 279 } 280 } 281 282 283 inline void 284 smp_send_stop(void) { send_IPI_allbutself(IPI_CPU_STOP); } 285 286 static inline void 287 smp_send_start(void) { send_IPI_allbutself(IPI_CPU_START); } 288 289 void 290 smp_send_reschedule(int cpu) { send_IPI_single(cpu, IPI_RESCHEDULE); } 291 292 void 293 smp_send_all_nop(void) 294 { 295 send_IPI_allbutself(IPI_NOP); 296 } 297 298 299 /** 300 * Run a function on all other CPUs. 301 * <func> The function to run. This must be fast and non-blocking. 302 * <info> An arbitrary pointer to pass to the function. 303 * <retry> If true, keep retrying until ready. 304 * <wait> If true, wait until function has completed on other CPUs. 305 * [RETURNS] 0 on success, else a negative status code. 306 * 307 * Does not return until remote CPUs are nearly ready to execute <func> 308 * or have executed. 309 */ 310 311 int 312 smp_call_function (void (*func) (void *info), void *info, int retry, int wait) 313 { 314 struct smp_call_struct data; 315 unsigned long timeout; 316 static DEFINE_SPINLOCK(lock); 317 int retries = 0; 318 319 if (num_online_cpus() < 2) 320 return 0; 321 322 /* Can deadlock when called with interrupts disabled */ 323 WARN_ON(irqs_disabled()); 324 325 /* can also deadlock if IPIs are disabled */ 326 WARN_ON((get_eiem() & (1UL<<(CPU_IRQ_MAX - IPI_IRQ))) == 0); 327 328 329 data.func = func; 330 data.info = info; 331 data.wait = wait; 332 atomic_set(&data.unstarted_count, num_online_cpus() - 1); 333 atomic_set(&data.unfinished_count, num_online_cpus() - 1); 334 335 if (retry) { 336 spin_lock (&lock); 337 while (smp_call_function_data != 0) 338 barrier(); 339 } 340 else { 341 spin_lock (&lock); 342 if (smp_call_function_data) { 343 spin_unlock (&lock); 344 return -EBUSY; 345 } 346 } 347 348 smp_call_function_data = &data; 349 spin_unlock (&lock); 350 351 /* Send a message to all other CPUs and wait for them to respond */ 352 send_IPI_allbutself(IPI_CALL_FUNC); 353 354 retry: 355 /* Wait for response */ 356 timeout = jiffies + HZ; 357 while ( (atomic_read (&data.unstarted_count) > 0) && 358 time_before (jiffies, timeout) ) 359 barrier (); 360 361 if (atomic_read (&data.unstarted_count) > 0) { 362 printk(KERN_CRIT "SMP CALL FUNCTION TIMED OUT! (cpu=%d), try %d\n", 363 smp_processor_id(), ++retries); 364 goto retry; 365 } 366 /* We either got one or timed out. Release the lock */ 367 368 mb(); 369 smp_call_function_data = NULL; 370 371 while (wait && atomic_read (&data.unfinished_count) > 0) 372 barrier (); 373 374 return 0; 375 } 376 377 EXPORT_SYMBOL(smp_call_function); 378 379 /* 380 * Flush all other CPU's tlb and then mine. Do this with on_each_cpu() 381 * as we want to ensure all TLB's flushed before proceeding. 382 */ 383 384 void 385 smp_flush_tlb_all(void) 386 { 387 on_each_cpu(flush_tlb_all_local, NULL, 1, 1); 388 } 389 390 /* 391 * Called by secondaries to update state and initialize CPU registers. 392 */ 393 static void __init 394 smp_cpu_init(int cpunum) 395 { 396 extern int init_per_cpu(int); /* arch/parisc/kernel/processor.c */ 397 extern void init_IRQ(void); /* arch/parisc/kernel/irq.c */ 398 extern void start_cpu_itimer(void); /* arch/parisc/kernel/time.c */ 399 400 /* Set modes and Enable floating point coprocessor */ 401 (void) init_per_cpu(cpunum); 402 403 disable_sr_hashing(); 404 405 mb(); 406 407 /* Well, support 2.4 linux scheme as well. */ 408 if (cpu_test_and_set(cpunum, cpu_online_map)) 409 { 410 extern void machine_halt(void); /* arch/parisc.../process.c */ 411 412 printk(KERN_CRIT "CPU#%d already initialized!\n", cpunum); 413 machine_halt(); 414 } 415 416 /* Initialise the idle task for this CPU */ 417 atomic_inc(&init_mm.mm_count); 418 current->active_mm = &init_mm; 419 if(current->mm) 420 BUG(); 421 enter_lazy_tlb(&init_mm, current); 422 423 init_IRQ(); /* make sure no IRQs are enabled or pending */ 424 start_cpu_itimer(); 425 } 426 427 428 /* 429 * Slaves start using C here. Indirectly called from smp_slave_stext. 430 * Do what start_kernel() and main() do for boot strap processor (aka monarch) 431 */ 432 void __init smp_callin(void) 433 { 434 int slave_id = cpu_now_booting; 435 #if 0 436 void *istack; 437 #endif 438 439 smp_cpu_init(slave_id); 440 preempt_disable(); 441 442 #if 0 /* NOT WORKING YET - see entry.S */ 443 istack = (void *)__get_free_pages(GFP_KERNEL,ISTACK_ORDER); 444 if (istack == NULL) { 445 printk(KERN_CRIT "Failed to allocate interrupt stack for cpu %d\n",slave_id); 446 BUG(); 447 } 448 mtctl(istack,31); 449 #endif 450 451 flush_cache_all_local(); /* start with known state */ 452 flush_tlb_all_local(NULL); 453 454 local_irq_enable(); /* Interrupts have been off until now */ 455 456 cpu_idle(); /* Wait for timer to schedule some work */ 457 458 /* NOTREACHED */ 459 panic("smp_callin() AAAAaaaaahhhh....\n"); 460 } 461 462 /* 463 * Bring one cpu online. 464 */ 465 int __cpuinit smp_boot_one_cpu(int cpuid) 466 { 467 struct task_struct *idle; 468 long timeout; 469 470 /* 471 * Create an idle task for this CPU. Note the address wed* give 472 * to kernel_thread is irrelevant -- it's going to start 473 * where OS_BOOT_RENDEVZ vector in SAL says to start. But 474 * this gets all the other task-y sort of data structures set 475 * up like we wish. We need to pull the just created idle task 476 * off the run queue and stuff it into the init_tasks[] array. 477 * Sheesh . . . 478 */ 479 480 idle = fork_idle(cpuid); 481 if (IS_ERR(idle)) 482 panic("SMP: fork failed for CPU:%d", cpuid); 483 484 task_thread_info(idle)->cpu = cpuid; 485 486 /* Let _start know what logical CPU we're booting 487 ** (offset into init_tasks[],cpu_data[]) 488 */ 489 cpu_now_booting = cpuid; 490 491 /* 492 ** boot strap code needs to know the task address since 493 ** it also contains the process stack. 494 */ 495 smp_init_current_idle_task = idle ; 496 mb(); 497 498 printk("Releasing cpu %d now, hpa=%lx\n", cpuid, cpu_data[cpuid].hpa); 499 500 /* 501 ** This gets PDC to release the CPU from a very tight loop. 502 ** 503 ** From the PA-RISC 2.0 Firmware Architecture Reference Specification: 504 ** "The MEM_RENDEZ vector specifies the location of OS_RENDEZ which 505 ** is executed after receiving the rendezvous signal (an interrupt to 506 ** EIR{0}). MEM_RENDEZ is valid only when it is nonzero and the 507 ** contents of memory are valid." 508 */ 509 gsc_writel(TIMER_IRQ - CPU_IRQ_BASE, cpu_data[cpuid].hpa); 510 mb(); 511 512 /* 513 * OK, wait a bit for that CPU to finish staggering about. 514 * Slave will set a bit when it reaches smp_cpu_init(). 515 * Once the "monarch CPU" sees the bit change, it can move on. 516 */ 517 for (timeout = 0; timeout < 10000; timeout++) { 518 if(cpu_online(cpuid)) { 519 /* Which implies Slave has started up */ 520 cpu_now_booting = 0; 521 smp_init_current_idle_task = NULL; 522 goto alive ; 523 } 524 udelay(100); 525 barrier(); 526 } 527 528 put_task_struct(idle); 529 idle = NULL; 530 531 printk(KERN_CRIT "SMP: CPU:%d is stuck.\n", cpuid); 532 return -1; 533 534 alive: 535 /* Remember the Slave data */ 536 smp_debug(100, KERN_DEBUG "SMP: CPU:%d came alive after %ld _us\n", 537 cpuid, timeout * 100); 538 return 0; 539 } 540 541 void __devinit smp_prepare_boot_cpu(void) 542 { 543 int bootstrap_processor=cpu_data[0].cpuid; /* CPU ID of BSP */ 544 545 /* Setup BSP mappings */ 546 printk("SMP: bootstrap CPU ID is %d\n",bootstrap_processor); 547 548 cpu_set(bootstrap_processor, cpu_online_map); 549 cpu_set(bootstrap_processor, cpu_present_map); 550 } 551 552 553 554 /* 555 ** inventory.c:do_inventory() hasn't yet been run and thus we 556 ** don't 'discover' the additional CPUs until later. 557 */ 558 void __init smp_prepare_cpus(unsigned int max_cpus) 559 { 560 cpus_clear(cpu_present_map); 561 cpu_set(0, cpu_present_map); 562 563 parisc_max_cpus = max_cpus; 564 if (!max_cpus) 565 printk(KERN_INFO "SMP mode deactivated.\n"); 566 } 567 568 569 void smp_cpus_done(unsigned int cpu_max) 570 { 571 return; 572 } 573 574 575 int __cpuinit __cpu_up(unsigned int cpu) 576 { 577 if (cpu != 0 && cpu < parisc_max_cpus) 578 smp_boot_one_cpu(cpu); 579 580 return cpu_online(cpu) ? 0 : -ENOSYS; 581 } 582 583 #ifdef CONFIG_PROC_FS 584 int __init 585 setup_profiling_timer(unsigned int multiplier) 586 { 587 return -EINVAL; 588 } 589 #endif 590