kernel/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         cpumask_clear(&temp_foreign_map);
 125         for_each_online_cpu(i) {
 126                 core_present = 0;
 127                 for_each_cpu(k, &temp_foreign_map)
 128                         if (cpu_data[i].package == cpu_data[k].package &&
 129                             cpu_data[i].core == cpu_data[k].core)
 130                                 core_present = 1;
 131                 if (!core_present)
 132                         cpumask_set_cpu(i, &temp_foreign_map);
 133         }
 134
 135         cpumask_copy(&cpu_foreign_map, &temp_foreign_map);
 136 }
 137
 138 struct plat_smp_ops *mp_ops;
 139 EXPORT_SYMBOL(mp_ops);
 140
 141 void register_smp_ops(struct plat_smp_ops *ops)
 142 {
 143         if (mp_ops)
 144                 printk(KERN_WARNING "Overriding previously set SMP ops\n");
 145
 146         mp_ops = ops;
 147 }
 148
 149 /*
 150  * First C code run on the secondary CPUs after being started up by
 151  * the master.
 152  */
 153 asmlinkage void start_secondary(void)
 154 {
 155         unsigned int cpu;
 156
 157         cpu_probe();
 158         per_cpu_trap_init(false);
 159         mips_clockevent_init();
 160         mp_ops->init_secondary();
 161         cpu_report();
 162         maar_init();
 163
 164         /*
 165          * XXX parity protection should be folded in here when it's converted
 166          * to an option instead of something based on .cputype
 167          */
 168
 169         calibrate_delay();
 170         preempt_disable();
 171         cpu = smp_processor_id();
 172         cpu_data[cpu].udelay_val = loops_per_jiffy;
 173
 174         cpumask_set_cpu(cpu, &cpu_coherent_mask);
 175         notify_cpu_starting(cpu);
 176
 177         set_cpu_online(cpu, true);
 178
 179         set_cpu_sibling_map(cpu);
 180         set_cpu_core_map(cpu);
 181
 182         calculate_cpu_foreign_map();
 183
 184         cpumask_set_cpu(cpu, &cpu_callin_map);
 185
 186         synchronise_count_slave(cpu);
 187
 188         /*
 189          * irq will be enabled in ->smp_finish(), enabling it too early
 190          * is dangerous.
 191          */
 192         WARN_ON_ONCE(!irqs_disabled());
 193         mp_ops->smp_finish();
 194
 195         cpu_startup_entry(CPUHP_ONLINE);
 196 }
 197
 198 static void stop_this_cpu(void *dummy)
 199 {
 200         /*
 201          * Remove this CPU. Be a bit slow here and
 202          * set the bits for every online CPU so we don't miss
 203          * any IPI whilst taking this VPE down.
 204          */
 205
 206         cpumask_copy(&cpu_foreign_map, cpu_online_mask);
 207
 208         /* Make it visible to every other CPU */
 209         smp_mb();
 210
 211         set_cpu_online(smp_processor_id(), false);
 212         calculate_cpu_foreign_map();
 213         local_irq_disable();
 214         while (1);
 215 }
 216
 217 void smp_send_stop(void)
 218 {
 219         smp_call_function(stop_this_cpu, NULL, 0);
 220 }
 221
 222 void __init smp_cpus_done(unsigned int max_cpus)
 223 {
 224 }
 225
 226 /* called from main before smp_init() */
 227 void __init smp_prepare_cpus(unsigned int max_cpus)
 228 {
 229         init_new_context(current, &init_mm);
 230         current_thread_info()->cpu = 0;
 231         mp_ops->prepare_cpus(max_cpus);
 232         set_cpu_sibling_map(0);
 233         set_cpu_core_map(0);
 234         calculate_cpu_foreign_map();
 235 #ifndef CONFIG_HOTPLUG_CPU
 236         init_cpu_present(cpu_possible_mask);
 237 #endif
 238         cpumask_copy(&cpu_coherent_mask, cpu_possible_mask);
 239 }
 240
 241 /* preload SMP state for boot cpu */
 242 void smp_prepare_boot_cpu(void)
 243 {
 244         set_cpu_possible(0, true);
 245         set_cpu_online(0, true);
 246         cpumask_set_cpu(0, &cpu_callin_map);
 247 }
 248
 249 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
 250 {
 251         mp_ops->boot_secondary(cpu, tidle);
 252
 253         /*
 254          * Trust is futile.  We should really have timeouts ...
 255          */
 256         while (!cpumask_test_cpu(cpu, &cpu_callin_map)) {
 257                 udelay(100);
 258                 schedule();
 259         }
 260
 261         synchronise_count_master(cpu);
 262         return 0;
 263 }
 264
 265 /* Not really SMP stuff ... */
 266 int setup_profiling_timer(unsigned int multiplier)
 267 {
 268         return 0;
 269 }
 270
 271 static void flush_tlb_all_ipi(void *info)
 272 {
 273         local_flush_tlb_all();
 274 }
 275
 276 void flush_tlb_all(void)
 277 {
 278         on_each_cpu(flush_tlb_all_ipi, NULL, 1);
 279 }
 280
 281 static void flush_tlb_mm_ipi(void *mm)
 282 {
 283         local_flush_tlb_mm((struct mm_struct *)mm);
 284 }
 285
 286 /*
 287  * Special Variant of smp_call_function for use by TLB functions:
 288  *
 289  *  o No return value
 290  *  o collapses to normal function call on UP kernels
 291  *  o collapses to normal function call on systems with a single shared
 292  *    primary cache.
 293  */
 294 static inline void smp_on_other_tlbs(void (*func) (void *info), void *info)
 295 {
 296         smp_call_function(func, info, 1);
 297 }
 298
 299 static inline void smp_on_each_tlb(void (*func) (void *info), void *info)
 300 {
 301         preempt_disable();
 302
 303         smp_on_other_tlbs(func, info);
 304         func(info);
 305
 306         preempt_enable();
 307 }
 308
 309 /*
 310  * The following tlb flush calls are invoked when old translations are
 311  * being torn down, or pte attributes are changing. For single threaded
 312  * address spaces, a new context is obtained on the current cpu, and tlb
 313  * context on other cpus are invalidated to force a new context allocation
 314  * at switch_mm time, should the mm ever be used on other cpus. For
 315  * multithreaded address spaces, intercpu interrupts have to be sent.
 316  * Another case where intercpu interrupts are required is when the target
 317  * mm might be active on another cpu (eg debuggers doing the flushes on
 318  * behalf of debugees, kswapd stealing pages from another process etc).
 319  * Kanoj 07/00.
 320  */
 321
 322 void flush_tlb_mm(struct mm_struct *mm)
 323 {
 324         preempt_disable();
 325
 326         if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
 327                 smp_on_other_tlbs(flush_tlb_mm_ipi, mm);
 328         } else {
 329                 unsigned int cpu;
 330
 331                 for_each_online_cpu(cpu) {
 332                         if (cpu != smp_processor_id() && cpu_context(cpu, mm))
 333                                 cpu_context(cpu, mm) = 0;
 334                 }
 335         }
 336         local_flush_tlb_mm(mm);
 337
 338         preempt_enable();
 339 }
 340
 341 struct flush_tlb_data {
 342         struct vm_area_struct *vma;
 343         unsigned long addr1;
 344         unsigned long addr2;
 345 };
 346
 347 static void flush_tlb_range_ipi(void *info)
 348 {
 349         struct flush_tlb_data *fd = info;
 350
 351         local_flush_tlb_range(fd->vma, fd->addr1, fd->addr2);
 352 }
 353
 354 void flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
 355 {
 356         struct mm_struct *mm = vma->vm_mm;
 357
 358         preempt_disable();
 359         if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
 360                 struct flush_tlb_data fd = {
 361                         .vma = vma,
 362                         .addr1 = start,
 363                         .addr2 = end,
 364                 };
 365
 366                 smp_on_other_tlbs(flush_tlb_range_ipi, &fd);
 367         } else {
 368                 unsigned int cpu;
 369
 370                 for_each_online_cpu(cpu) {
 371                         if (cpu != smp_processor_id() && cpu_context(cpu, mm))
 372                                 cpu_context(cpu, mm) = 0;
 373                 }
 374         }
 375         local_flush_tlb_range(vma, start, end);
 376         preempt_enable();
 377 }
 378
 379 static void flush_tlb_kernel_range_ipi(void *info)
 380 {
 381         struct flush_tlb_data *fd = info;
 382
 383         local_flush_tlb_kernel_range(fd->addr1, fd->addr2);
 384 }
 385
 386 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 387 {
 388         struct flush_tlb_data fd = {
 389                 .addr1 = start,
 390                 .addr2 = end,
 391         };
 392
 393         on_each_cpu(flush_tlb_kernel_range_ipi, &fd, 1);
 394 }
 395
 396 static void flush_tlb_page_ipi(void *info)
 397 {
 398         struct flush_tlb_data *fd = info;
 399
 400         local_flush_tlb_page(fd->vma, fd->addr1);
 401 }
 402
 403 void flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 404 {
 405         preempt_disable();
 406         if ((atomic_read(&vma->vm_mm->mm_users) != 1) || (current->mm != vma->vm_mm)) {
 407                 struct flush_tlb_data fd = {
 408                         .vma = vma,
 409                         .addr1 = page,
 410                 };
 411
 412                 smp_on_other_tlbs(flush_tlb_page_ipi, &fd);
 413         } else {
 414                 unsigned int cpu;
 415
 416                 for_each_online_cpu(cpu) {
 417                         if (cpu != smp_processor_id() && cpu_context(cpu, vma->vm_mm))
 418                                 cpu_context(cpu, vma->vm_mm) = 0;
 419                 }
 420         }
 421         local_flush_tlb_page(vma, page);
 422         preempt_enable();
 423 }
 424
 425 static void flush_tlb_one_ipi(void *info)
 426 {
 427         unsigned long vaddr = (unsigned long) info;
 428
 429         local_flush_tlb_one(vaddr);
 430 }
 431
 432 void flush_tlb_one(unsigned long vaddr)
 433 {
 434         smp_on_each_tlb(flush_tlb_one_ipi, (void *) vaddr);
 435 }
 436
 437 EXPORT_SYMBOL(flush_tlb_page);
 438 EXPORT_SYMBOL(flush_tlb_one);
 439
 440 #if defined(CONFIG_KEXEC)
 441 void (*dump_ipi_function_ptr)(void *) = NULL;
 442 void dump_send_ipi(void (*dump_ipi_callback)(void *))
 443 {
 444         int i;
 445         int cpu = smp_processor_id();
 446
 447         dump_ipi_function_ptr = dump_ipi_callback;
 448         smp_mb();
 449         for_each_online_cpu(i)
 450                 if (i != cpu)
 451                         mp_ops->send_ipi_single(i, SMP_DUMP);
 452
 453 }
 454 EXPORT_SYMBOL(dump_send_ipi);
 455 #endif
 456
 457 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 458
 459 static DEFINE_PER_CPU(atomic_t, tick_broadcast_count);
 460 static DEFINE_PER_CPU(struct call_single_data, tick_broadcast_csd);
 461
 462 void tick_broadcast(const struct cpumask *mask)
 463 {
 464         atomic_t *count;
 465         struct call_single_data *csd;
 466         int cpu;
 467
 468         for_each_cpu(cpu, mask) {
 469                 count = &per_cpu(tick_broadcast_count, cpu);
 470                 csd = &per_cpu(tick_broadcast_csd, cpu);
 471
 472                 if (atomic_inc_return(count) == 1)
 473                         smp_call_function_single_async(cpu, csd);
 474         }
 475 }
 476
 477 static void tick_broadcast_callee(void *info)
 478 {
 479         int cpu = smp_processor_id();
 480         tick_receive_broadcast();
 481         atomic_set(&per_cpu(tick_broadcast_count, cpu), 0);
 482 }
 483
 484 static int __init tick_broadcast_init(void)
 485 {
 486         struct call_single_data *csd;
 487         int cpu;
 488
 489         for (cpu = 0; cpu < NR_CPUS; cpu++) {
 490                 csd = &per_cpu(tick_broadcast_csd, cpu);
 491                 csd->func = tick_broadcast_callee;
 492         }
 493
 494         return 0;
 495 }
 496 early_initcall(tick_broadcast_init);
 497
 498 #endif /* CONFIG_GENERIC_CLOCKEVENTS_BROADCAST */