arch/mips/kernel/smp.c

   1 /*
   2  * This program is free software; you can redistribute it and/or
   3  * modify it under the terms of the GNU General Public License
   4  * as published by the Free Software Foundation; either version 2
   5  * of the License, or (at your option) any later version.
   6  *
   7  * This program is distributed in the hope that it will be useful,
   8  * but WITHOUT ANY WARRANTY; without even the implied warranty of
   9  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  10  * GNU General Public License for more details.
  11  *
  12  * You should have received a copy of the GNU General Public License
  13  * along with this program; if not, write to the Free Software
  14  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
  15  *
  16  * Copyright (C) 2000, 2001 Kanoj Sarcar
  17  * Copyright (C) 2000, 2001 Ralf Baechle
  18  * Copyright (C) 2000, 2001 Silicon Graphics, Inc.
  19  * Copyright (C) 2000, 2001, 2003 Broadcom Corporation
  20  */
  21 #include <linux/cache.h>
  22 #include <linux/delay.h>
  23 #include <linux/init.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/smp.h>
  26 #include <linux/spinlock.h>
  27 #include <linux/threads.h>
  28 #include <linux/module.h>
  29 #include <linux/time.h>
  30 #include <linux/timex.h>
  31 #include <linux/sched.h>
  32 #include <linux/cpumask.h>
  33 #include <linux/cpu.h>
  34 #include <linux/err.h>
  35 #include <linux/ftrace.h>
  36
  37 #include <linux/atomic.h>
  38 #include <asm/cpu.h>
  39 #include <asm/processor.h>
  40 #include <asm/idle.h>
  41 #include <asm/r4k-timer.h>
  42 #include <asm/mmu_context.h>
  43 #include <asm/time.h>
  44 #include <asm/setup.h>
  45 #include <asm/maar.h>
  46
  47 cpumask_t cpu_callin_map;               /* Bitmask of started secondaries */
  48
  49 int __cpu_number_map[NR_CPUS];          /* Map physical to logical */
  50 EXPORT_SYMBOL(__cpu_number_map);
  51
  52 int __cpu_logical_map[NR_CPUS];         /* Map logical to physical */
  53 EXPORT_SYMBOL(__cpu_logical_map);
  54
  55 /* Number of TCs (or siblings in Intel speak) per CPU core */
  56 int smp_num_siblings = 1;
  57 EXPORT_SYMBOL(smp_num_siblings);
  58
  59 /* representing the TCs (or siblings in Intel speak) of each logical CPU */
  60 cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly;
  61 EXPORT_SYMBOL(cpu_sibling_map);
  62
  63 /* representing the core map of multi-core chips of each logical CPU */
  64 cpumask_t cpu_core_map[NR_CPUS] __read_mostly;
  65 EXPORT_SYMBOL(cpu_core_map);
  66
  67 /*
  68  * A logcal cpu mask containing only one VPE per core to
  69  * reduce the number of IPIs on large MT systems.
  70  */
  71 cpumask_t cpu_foreign_map __read_mostly;
  72 EXPORT_SYMBOL(cpu_foreign_map);
  73
  74 /* representing cpus for which sibling maps can be computed */
  75 static cpumask_t cpu_sibling_setup_map;
  76
  77 /* representing cpus for which core maps can be computed */
  78 static cpumask_t cpu_core_setup_map;
  79
  80 cpumask_t cpu_coherent_mask;
  81
  82 static inline void set_cpu_sibling_map(int cpu)
  83 {
  84         int i;
  85
  86         cpumask_set_cpu(cpu, &cpu_sibling_setup_map);
  87
  88         if (smp_num_siblings > 1) {
  89                 for_each_cpu(i, &cpu_sibling_setup_map) {
  90                         if (cpu_data[cpu].package == cpu_data[i].package &&
  91                                     cpu_data[cpu].core == cpu_data[i].core) {
  92                                 cpumask_set_cpu(i, &cpu_sibling_map[cpu]);
  93                                 cpumask_set_cpu(cpu, &cpu_sibling_map[i]);
  94                         }
  95                 }
  96         } else
  97                 cpumask_set_cpu(cpu, &cpu_sibling_map[cpu]);
  98 }
  99
 100 static inline void set_cpu_core_map(int cpu)
 101 {
 102         int i;
 103
 104         cpumask_set_cpu(cpu, &cpu_core_setup_map);
 105
 106         for_each_cpu(i, &cpu_core_setup_map) {
 107                 if (cpu_data[cpu].package == cpu_data[i].package) {
 108                         cpumask_set_cpu(i, &cpu_core_map[cpu]);
 109                         cpumask_set_cpu(cpu, &cpu_core_map[i]);
 110                 }
 111         }
 112 }
 113
 114 /*
 115  * Calculate a new cpu_foreign_map mask whenever a
 116  * new cpu appears or disappears.
 117  */
 118 static inline void calculate_cpu_foreign_map(void)
 119 {
 120         int i, k, core_present;
 121         cpumask_t temp_foreign_map;
 122
 123         /* Re-calculate the mask */
 124         for_each_online_cpu(i) {
 125                 core_present = 0;
 126                 for_each_cpu(k, &temp_foreign_map)
 127                         if (cpu_data[i].package == cpu_data[k].package &&
 128                             cpu_data[i].core == cpu_data[k].core)
 129                                 core_present = 1;
 130                 if (!core_present)
 131                         cpumask_set_cpu(i, &temp_foreign_map);
 132         }
 133
 134         cpumask_copy(&cpu_foreign_map, &temp_foreign_map);
 135 }
 136
 137 struct plat_smp_ops *mp_ops;
 138 EXPORT_SYMBOL(mp_ops);
 139
 140 void register_smp_ops(struct plat_smp_ops *ops)
 141 {
 142         if (mp_ops)
 143                 printk(KERN_WARNING "Overriding previously set SMP ops\n");
 144
 145         mp_ops = ops;
 146 }
 147
 148 /*
 149  * First C code run on the secondary CPUs after being started up by
 150  * the master.
 151  */
 152 asmlinkage void start_secondary(void)
 153 {
 154         unsigned int cpu;
 155
 156         cpu_probe();
 157         per_cpu_trap_init(false);
 158         mips_clockevent_init();
 159         mp_ops->init_secondary();
 160         cpu_report();
 161         maar_init();
 162
 163         /*
 164          * XXX parity protection should be folded in here when it's converted
 165          * to an option instead of something based on .cputype
 166          */
 167
 168         calibrate_delay();
 169         preempt_disable();
 170         cpu = smp_processor_id();
 171         cpu_data[cpu].udelay_val = loops_per_jiffy;
 172
 173         cpumask_set_cpu(cpu, &cpu_coherent_mask);
 174         notify_cpu_starting(cpu);
 175
 176         set_cpu_online(cpu, true);
 177
 178         set_cpu_sibling_map(cpu);
 179         set_cpu_core_map(cpu);
 180
 181         calculate_cpu_foreign_map();
 182
 183         cpumask_set_cpu(cpu, &cpu_callin_map);
 184
 185         synchronise_count_slave(cpu);
 186
 187         /*
 188          * irq will be enabled in ->smp_finish(), enabling it too early
 189          * is dangerous.
 190          */
 191         WARN_ON_ONCE(!irqs_disabled());
 192         mp_ops->smp_finish();
 193
 194         cpu_startup_entry(CPUHP_ONLINE);
 195 }
 196
 197 static void stop_this_cpu(void *dummy)
 198 {
 199         /*
 200          * Remove this CPU. Be a bit slow here and
 201          * set the bits for every online CPU so we don't miss
 202          * any IPI whilst taking this VPE down.
 203          */
 204
 205         cpumask_copy(&cpu_foreign_map, cpu_online_mask);
 206
 207         /* Make it visible to every other CPU */
 208         smp_mb();
 209
 210         set_cpu_online(smp_processor_id(), false);
 211         calculate_cpu_foreign_map();
 212         local_irq_disable();
 213         while (1);
 214 }
 215
 216 void smp_send_stop(void)
 217 {
 218         smp_call_function(stop_this_cpu, NULL, 0);
 219 }
 220
 221 void __init smp_cpus_done(unsigned int max_cpus)
 222 {
 223 }
 224
 225 /* called from main before smp_init() */
 226 void __init smp_prepare_cpus(unsigned int max_cpus)
 227 {
 228         init_new_context(current, &init_mm);
 229         current_thread_info()->cpu = 0;
 230         mp_ops->prepare_cpus(max_cpus);
 231         set_cpu_sibling_map(0);
 232         set_cpu_core_map(0);
 233         calculate_cpu_foreign_map();
 234 #ifndef CONFIG_HOTPLUG_CPU
 235         init_cpu_present(cpu_possible_mask);
 236 #endif
 237         cpumask_copy(&cpu_coherent_mask, cpu_possible_mask);
 238 }
 239
 240 /* preload SMP state for boot cpu */
 241 void smp_prepare_boot_cpu(void)
 242 {
 243         set_cpu_possible(0, true);
 244         set_cpu_online(0, true);
 245         cpumask_set_cpu(0, &cpu_callin_map);
 246 }
 247
 248 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
 249 {
 250         mp_ops->boot_secondary(cpu, tidle);
 251
 252         /*
 253          * Trust is futile.  We should really have timeouts ...
 254          */
 255         while (!cpumask_test_cpu(cpu, &cpu_callin_map)) {
 256                 udelay(100);
 257                 schedule();
 258         }
 259
 260         synchronise_count_master(cpu);
 261         return 0;
 262 }
 263
 264 /* Not really SMP stuff ... */
 265 int setup_profiling_timer(unsigned int multiplier)
 266 {
 267         return 0;
 268 }
 269
 270 static void flush_tlb_all_ipi(void *info)
 271 {
 272         local_flush_tlb_all();
 273 }
 274
 275 void flush_tlb_all(void)
 276 {
 277         on_each_cpu(flush_tlb_all_ipi, NULL, 1);
 278 }
 279
 280 static void flush_tlb_mm_ipi(void *mm)
 281 {
 282         local_flush_tlb_mm((struct mm_struct *)mm);
 283 }
 284
 285 /*
 286  * Special Variant of smp_call_function for use by TLB functions:
 287  *
 288  *  o No return value
 289  *  o collapses to normal function call on UP kernels
 290  *  o collapses to normal function call on systems with a single shared
 291  *    primary cache.
 292  */
 293 static inline void smp_on_other_tlbs(void (*func) (void *info), void *info)
 294 {
 295         smp_call_function(func, info, 1);
 296 }
 297
 298 static inline void smp_on_each_tlb(void (*func) (void *info), void *info)
 299 {
 300         preempt_disable();
 301
 302         smp_on_other_tlbs(func, info);
 303         func(info);
 304
 305         preempt_enable();
 306 }
 307
 308 /*
 309  * The following tlb flush calls are invoked when old translations are
 310  * being torn down, or pte attributes are changing. For single threaded
 311  * address spaces, a new context is obtained on the current cpu, and tlb
 312  * context on other cpus are invalidated to force a new context allocation
 313  * at switch_mm time, should the mm ever be used on other cpus. For
 314  * multithreaded address spaces, intercpu interrupts have to be sent.
 315  * Another case where intercpu interrupts are required is when the target
 316  * mm might be active on another cpu (eg debuggers doing the flushes on
 317  * behalf of debugees, kswapd stealing pages from another process etc).
 318  * Kanoj 07/00.
 319  */
 320
 321 void flush_tlb_mm(struct mm_struct *mm)
 322 {
 323         preempt_disable();
 324
 325         if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
 326                 smp_on_other_tlbs(flush_tlb_mm_ipi, mm);
 327         } else {
 328                 unsigned int cpu;
 329
 330                 for_each_online_cpu(cpu) {
 331                         if (cpu != smp_processor_id() && cpu_context(cpu, mm))
 332                                 cpu_context(cpu, mm) = 0;
 333                 }
 334         }
 335         local_flush_tlb_mm(mm);
 336
 337         preempt_enable();
 338 }
 339
 340 struct flush_tlb_data {
 341         struct vm_area_struct *vma;
 342         unsigned long addr1;
 343         unsigned long addr2;
 344 };
 345
 346 static void flush_tlb_range_ipi(void *info)
 347 {
 348         struct flush_tlb_data *fd = info;
 349
 350         local_flush_tlb_range(fd->vma, fd->addr1, fd->addr2);
 351 }
 352
 353 void flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
 354 {
 355         struct mm_struct *mm = vma->vm_mm;
 356
 357         preempt_disable();
 358         if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
 359                 struct flush_tlb_data fd = {
 360                         .vma = vma,
 361                         .addr1 = start,
 362                         .addr2 = end,
 363                 };
 364
 365                 smp_on_other_tlbs(flush_tlb_range_ipi, &fd);
 366         } else {
 367                 unsigned int cpu;
 368
 369                 for_each_online_cpu(cpu) {
 370                         if (cpu != smp_processor_id() && cpu_context(cpu, mm))
 371                                 cpu_context(cpu, mm) = 0;
 372                 }
 373         }
 374         local_flush_tlb_range(vma, start, end);
 375         preempt_enable();
 376 }
 377
 378 static void flush_tlb_kernel_range_ipi(void *info)
 379 {
 380         struct flush_tlb_data *fd = info;
 381
 382         local_flush_tlb_kernel_range(fd->addr1, fd->addr2);
 383 }
 384
 385 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 386 {
 387         struct flush_tlb_data fd = {
 388                 .addr1 = start,
 389                 .addr2 = end,
 390         };
 391
 392         on_each_cpu(flush_tlb_kernel_range_ipi, &fd, 1);
 393 }
 394
 395 static void flush_tlb_page_ipi(void *info)
 396 {
 397         struct flush_tlb_data *fd = info;
 398
 399         local_flush_tlb_page(fd->vma, fd->addr1);
 400 }
 401
 402 void flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 403 {
 404         preempt_disable();
 405         if ((atomic_read(&vma->vm_mm->mm_users) != 1) || (current->mm != vma->vm_mm)) {
 406                 struct flush_tlb_data fd = {
 407                         .vma = vma,
 408                         .addr1 = page,
 409                 };
 410
 411                 smp_on_other_tlbs(flush_tlb_page_ipi, &fd);
 412         } else {
 413                 unsigned int cpu;
 414
 415                 for_each_online_cpu(cpu) {
 416                         if (cpu != smp_processor_id() && cpu_context(cpu, vma->vm_mm))
 417                                 cpu_context(cpu, vma->vm_mm) = 0;
 418                 }
 419         }
 420         local_flush_tlb_page(vma, page);
 421         preempt_enable();
 422 }
 423
 424 static void flush_tlb_one_ipi(void *info)
 425 {
 426         unsigned long vaddr = (unsigned long) info;
 427
 428         local_flush_tlb_one(vaddr);
 429 }
 430
 431 void flush_tlb_one(unsigned long vaddr)
 432 {
 433         smp_on_each_tlb(flush_tlb_one_ipi, (void *) vaddr);
 434 }
 435
 436 EXPORT_SYMBOL(flush_tlb_page);
 437 EXPORT_SYMBOL(flush_tlb_one);
 438
 439 #if defined(CONFIG_KEXEC)
 440 void (*dump_ipi_function_ptr)(void *) = NULL;
 441 void dump_send_ipi(void (*dump_ipi_callback)(void *))
 442 {
 443         int i;
 444         int cpu = smp_processor_id();
 445
 446         dump_ipi_function_ptr = dump_ipi_callback;
 447         smp_mb();
 448         for_each_online_cpu(i)
 449                 if (i != cpu)
 450                         mp_ops->send_ipi_single(i, SMP_DUMP);
 451
 452 }
 453 EXPORT_SYMBOL(dump_send_ipi);
 454 #endif
 455
 456 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 457
 458 static DEFINE_PER_CPU(atomic_t, tick_broadcast_count);
 459 static DEFINE_PER_CPU(struct call_single_data, tick_broadcast_csd);
 460
 461 void tick_broadcast(const struct cpumask *mask)
 462 {
 463         atomic_t *count;
 464         struct call_single_data *csd;
 465         int cpu;
 466
 467         for_each_cpu(cpu, mask) {
 468                 count = &per_cpu(tick_broadcast_count, cpu);
 469                 csd = &per_cpu(tick_broadcast_csd, cpu);
 470
 471                 if (atomic_inc_return(count) == 1)
 472                         smp_call_function_single_async(cpu, csd);
 473         }
 474 }
 475
 476 static void tick_broadcast_callee(void *info)
 477 {
 478         int cpu = smp_processor_id();
 479         tick_receive_broadcast();
 480         atomic_set(&per_cpu(tick_broadcast_count, cpu), 0);
 481 }
 482
 483 static int __init tick_broadcast_init(void)
 484 {
 485         struct call_single_data *csd;
 486         int cpu;
 487
 488         for (cpu = 0; cpu < NR_CPUS; cpu++) {
 489                 csd = &per_cpu(tick_broadcast_csd, cpu);
 490                 csd->func = tick_broadcast_callee;
 491         }
 492
 493         return 0;
 494 }
 495 early_initcall(tick_broadcast_init);
 496
 497 #endif /* CONFIG_GENERIC_CLOCKEVENTS_BROADCAST */