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
  46 cpumask_t cpu_callin_map;               /* Bitmask of started secondaries */
  47
  48 int __cpu_number_map[NR_CPUS];          /* Map physical to logical */
  49 EXPORT_SYMBOL(__cpu_number_map);
  50
  51 int __cpu_logical_map[NR_CPUS];         /* Map logical to physical */
  52 EXPORT_SYMBOL(__cpu_logical_map);
  53
  54 /* Number of TCs (or siblings in Intel speak) per CPU core */
  55 int smp_num_siblings = 1;
  56 EXPORT_SYMBOL(smp_num_siblings);
  57
  58 /* representing the TCs (or siblings in Intel speak) of each logical CPU */
  59 cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly;
  60 EXPORT_SYMBOL(cpu_sibling_map);
  61
  62 /* representing the core map of multi-core chips of each logical CPU */
  63 cpumask_t cpu_core_map[NR_CPUS] __read_mostly;
  64 EXPORT_SYMBOL(cpu_core_map);
  65
  66 /*
  67  * A logcal cpu mask containing only one VPE per core to
  68  * reduce the number of IPIs on large MT systems.
  69  */
  70 cpumask_t cpu_foreign_map __read_mostly;
  71 EXPORT_SYMBOL(cpu_foreign_map);
  72
  73 /* representing cpus for which sibling maps can be computed */
  74 static cpumask_t cpu_sibling_setup_map;
  75
  76 /* representing cpus for which core maps can be computed */
  77 static cpumask_t cpu_core_setup_map;
  78
  79 cpumask_t cpu_coherent_mask;
  80
  81 static inline void set_cpu_sibling_map(int cpu)
  82 {
  83         int i;
  84
  85         cpumask_set_cpu(cpu, &cpu_sibling_setup_map);
  86
  87         if (smp_num_siblings > 1) {
  88                 for_each_cpu(i, &cpu_sibling_setup_map) {
  89                         if (cpu_data[cpu].package == cpu_data[i].package &&
  90                                     cpu_data[cpu].core == cpu_data[i].core) {
  91                                 cpumask_set_cpu(i, &cpu_sibling_map[cpu]);
  92                                 cpumask_set_cpu(cpu, &cpu_sibling_map[i]);
  93                         }
  94                 }
  95         } else
  96                 cpumask_set_cpu(cpu, &cpu_sibling_map[cpu]);
  97 }
  98
  99 static inline void set_cpu_core_map(int cpu)
 100 {
 101         int i;
 102
 103         cpumask_set_cpu(cpu, &cpu_core_setup_map);
 104
 105         for_each_cpu(i, &cpu_core_setup_map) {
 106                 if (cpu_data[cpu].package == cpu_data[i].package) {
 107                         cpumask_set_cpu(i, &cpu_core_map[cpu]);
 108                         cpumask_set_cpu(cpu, &cpu_core_map[i]);
 109                 }
 110         }
 111 }
 112
 113 /*
 114  * Calculate a new cpu_foreign_map mask whenever a
 115  * new cpu appears or disappears.
 116  */
 117 static inline void calculate_cpu_foreign_map(void)
 118 {
 119         int i, k, core_present;
 120         cpumask_t temp_foreign_map;
 121
 122         /* Re-calculate the mask */
 123         for_each_online_cpu(i) {
 124                 core_present = 0;
 125                 for_each_cpu(k, &temp_foreign_map)
 126                         if (cpu_data[i].package == cpu_data[k].package &&
 127                             cpu_data[i].core == cpu_data[k].core)
 128                                 core_present = 1;
 129                 if (!core_present)
 130                         cpumask_set_cpu(i, &temp_foreign_map);
 131         }
 132
 133         cpumask_copy(&cpu_foreign_map, &temp_foreign_map);
 134 }
 135
 136 struct plat_smp_ops *mp_ops;
 137 EXPORT_SYMBOL(mp_ops);
 138
 139 void register_smp_ops(struct plat_smp_ops *ops)
 140 {
 141         if (mp_ops)
 142                 printk(KERN_WARNING "Overriding previously set SMP ops\n");
 143
 144         mp_ops = ops;
 145 }
 146
 147 /*
 148  * First C code run on the secondary CPUs after being started up by
 149  * the master.
 150  */
 151 asmlinkage void start_secondary(void)
 152 {
 153         unsigned int cpu;
 154
 155         cpu_probe();
 156         per_cpu_trap_init(false);
 157         mips_clockevent_init();
 158         mp_ops->init_secondary();
 159         cpu_report();
 160
 161         /*
 162          * XXX parity protection should be folded in here when it's converted
 163          * to an option instead of something based on .cputype
 164          */
 165
 166         calibrate_delay();
 167         preempt_disable();
 168         cpu = smp_processor_id();
 169         cpu_data[cpu].udelay_val = loops_per_jiffy;
 170
 171         cpumask_set_cpu(cpu, &cpu_coherent_mask);
 172         notify_cpu_starting(cpu);
 173
 174         set_cpu_online(cpu, true);
 175
 176         set_cpu_sibling_map(cpu);
 177         set_cpu_core_map(cpu);
 178
 179         calculate_cpu_foreign_map();
 180
 181         cpumask_set_cpu(cpu, &cpu_callin_map);
 182
 183         synchronise_count_slave(cpu);
 184
 185         /*
 186          * irq will be enabled in ->smp_finish(), enabling it too early
 187          * is dangerous.
 188          */
 189         WARN_ON_ONCE(!irqs_disabled());
 190         mp_ops->smp_finish();
 191
 192         cpu_startup_entry(CPUHP_ONLINE);
 193 }
 194
 195 /*
 196  * Call into both interrupt handlers, as we share the IPI for them
 197  */
 198 void __irq_entry smp_call_function_interrupt(void)
 199 {
 200         irq_enter();
 201         generic_smp_call_function_interrupt();
 202         irq_exit();
 203 }
 204
 205 static void stop_this_cpu(void *dummy)
 206 {
 207         /*
 208          * Remove this CPU. Be a bit slow here and
 209          * set the bits for every online CPU so we don't miss
 210          * any IPI whilst taking this VPE down.
 211          */
 212
 213         cpumask_copy(&cpu_foreign_map, cpu_online_mask);
 214
 215         /* Make it visible to every other CPU */
 216         smp_mb();
 217
 218         set_cpu_online(smp_processor_id(), false);
 219         calculate_cpu_foreign_map();
 220         local_irq_disable();
 221         while (1);
 222 }
 223
 224 void smp_send_stop(void)
 225 {
 226         smp_call_function(stop_this_cpu, NULL, 0);
 227 }
 228
 229 void __init smp_cpus_done(unsigned int max_cpus)
 230 {
 231 }
 232
 233 /* called from main before smp_init() */
 234 void __init smp_prepare_cpus(unsigned int max_cpus)
 235 {
 236         init_new_context(current, &init_mm);
 237         current_thread_info()->cpu = 0;
 238         mp_ops->prepare_cpus(max_cpus);
 239         set_cpu_sibling_map(0);
 240         set_cpu_core_map(0);
 241         calculate_cpu_foreign_map();
 242 #ifndef CONFIG_HOTPLUG_CPU
 243         init_cpu_present(cpu_possible_mask);
 244 #endif
 245         cpumask_copy(&cpu_coherent_mask, cpu_possible_mask);
 246 }
 247
 248 /* preload SMP state for boot cpu */
 249 void smp_prepare_boot_cpu(void)
 250 {
 251         set_cpu_possible(0, true);
 252         set_cpu_online(0, true);
 253         cpumask_set_cpu(0, &cpu_callin_map);
 254 }
 255
 256 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
 257 {
 258         mp_ops->boot_secondary(cpu, tidle);
 259
 260         /*
 261          * Trust is futile.  We should really have timeouts ...
 262          */
 263         while (!cpumask_test_cpu(cpu, &cpu_callin_map)) {
 264                 udelay(100);
 265                 schedule();
 266         }
 267
 268         synchronise_count_master(cpu);
 269         return 0;
 270 }
 271
 272 /* Not really SMP stuff ... */
 273 int setup_profiling_timer(unsigned int multiplier)
 274 {
 275         return 0;
 276 }
 277
 278 static void flush_tlb_all_ipi(void *info)
 279 {
 280         local_flush_tlb_all();
 281 }
 282
 283 void flush_tlb_all(void)
 284 {
 285         on_each_cpu(flush_tlb_all_ipi, NULL, 1);
 286 }
 287
 288 static void flush_tlb_mm_ipi(void *mm)
 289 {
 290         local_flush_tlb_mm((struct mm_struct *)mm);
 291 }
 292
 293 /*
 294  * Special Variant of smp_call_function for use by TLB functions:
 295  *
 296  *  o No return value
 297  *  o collapses to normal function call on UP kernels
 298  *  o collapses to normal function call on systems with a single shared
 299  *    primary cache.
 300  */
 301 static inline void smp_on_other_tlbs(void (*func) (void *info), void *info)
 302 {
 303         smp_call_function(func, info, 1);
 304 }
 305
 306 static inline void smp_on_each_tlb(void (*func) (void *info), void *info)
 307 {
 308         preempt_disable();
 309
 310         smp_on_other_tlbs(func, info);
 311         func(info);
 312
 313         preempt_enable();
 314 }
 315
 316 /*
 317  * The following tlb flush calls are invoked when old translations are
 318  * being torn down, or pte attributes are changing. For single threaded
 319  * address spaces, a new context is obtained on the current cpu, and tlb
 320  * context on other cpus are invalidated to force a new context allocation
 321  * at switch_mm time, should the mm ever be used on other cpus. For
 322  * multithreaded address spaces, intercpu interrupts have to be sent.
 323  * Another case where intercpu interrupts are required is when the target
 324  * mm might be active on another cpu (eg debuggers doing the flushes on
 325  * behalf of debugees, kswapd stealing pages from another process etc).
 326  * Kanoj 07/00.
 327  */
 328
 329 void flush_tlb_mm(struct mm_struct *mm)
 330 {
 331         preempt_disable();
 332
 333         if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
 334                 smp_on_other_tlbs(flush_tlb_mm_ipi, mm);
 335         } else {
 336                 unsigned int cpu;
 337
 338                 for_each_online_cpu(cpu) {
 339                         if (cpu != smp_processor_id() && cpu_context(cpu, mm))
 340                                 cpu_context(cpu, mm) = 0;
 341                 }
 342         }
 343         local_flush_tlb_mm(mm);
 344
 345         preempt_enable();
 346 }
 347
 348 struct flush_tlb_data {
 349         struct vm_area_struct *vma;
 350         unsigned long addr1;
 351         unsigned long addr2;
 352 };
 353
 354 static void flush_tlb_range_ipi(void *info)
 355 {
 356         struct flush_tlb_data *fd = info;
 357
 358         local_flush_tlb_range(fd->vma, fd->addr1, fd->addr2);
 359 }
 360
 361 void flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
 362 {
 363         struct mm_struct *mm = vma->vm_mm;
 364
 365         preempt_disable();
 366         if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
 367                 struct flush_tlb_data fd = {
 368                         .vma = vma,
 369                         .addr1 = start,
 370                         .addr2 = end,
 371                 };
 372
 373                 smp_on_other_tlbs(flush_tlb_range_ipi, &fd);
 374         } else {
 375                 unsigned int cpu;
 376
 377                 for_each_online_cpu(cpu) {
 378                         if (cpu != smp_processor_id() && cpu_context(cpu, mm))
 379                                 cpu_context(cpu, mm) = 0;
 380                 }
 381         }
 382         local_flush_tlb_range(vma, start, end);
 383         preempt_enable();
 384 }
 385
 386 static void flush_tlb_kernel_range_ipi(void *info)
 387 {
 388         struct flush_tlb_data *fd = info;
 389
 390         local_flush_tlb_kernel_range(fd->addr1, fd->addr2);
 391 }
 392
 393 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 394 {
 395         struct flush_tlb_data fd = {
 396                 .addr1 = start,
 397                 .addr2 = end,
 398         };
 399
 400         on_each_cpu(flush_tlb_kernel_range_ipi, &fd, 1);
 401 }
 402
 403 static void flush_tlb_page_ipi(void *info)
 404 {
 405         struct flush_tlb_data *fd = info;
 406
 407         local_flush_tlb_page(fd->vma, fd->addr1);
 408 }
 409
 410 void flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 411 {
 412         preempt_disable();
 413         if ((atomic_read(&vma->vm_mm->mm_users) != 1) || (current->mm != vma->vm_mm)) {
 414                 struct flush_tlb_data fd = {
 415                         .vma = vma,
 416                         .addr1 = page,
 417                 };
 418
 419                 smp_on_other_tlbs(flush_tlb_page_ipi, &fd);
 420         } else {
 421                 unsigned int cpu;
 422
 423                 for_each_online_cpu(cpu) {
 424                         if (cpu != smp_processor_id() && cpu_context(cpu, vma->vm_mm))
 425                                 cpu_context(cpu, vma->vm_mm) = 0;
 426                 }
 427         }
 428         local_flush_tlb_page(vma, page);
 429         preempt_enable();
 430 }
 431
 432 static void flush_tlb_one_ipi(void *info)
 433 {
 434         unsigned long vaddr = (unsigned long) info;
 435
 436         local_flush_tlb_one(vaddr);
 437 }
 438
 439 void flush_tlb_one(unsigned long vaddr)
 440 {
 441         smp_on_each_tlb(flush_tlb_one_ipi, (void *) vaddr);
 442 }
 443
 444 EXPORT_SYMBOL(flush_tlb_page);
 445 EXPORT_SYMBOL(flush_tlb_one);
 446
 447 #if defined(CONFIG_KEXEC)
 448 void (*dump_ipi_function_ptr)(void *) = NULL;
 449 void dump_send_ipi(void (*dump_ipi_callback)(void *))
 450 {
 451         int i;
 452         int cpu = smp_processor_id();
 453
 454         dump_ipi_function_ptr = dump_ipi_callback;
 455         smp_mb();
 456         for_each_online_cpu(i)
 457                 if (i != cpu)
 458                         mp_ops->send_ipi_single(i, SMP_DUMP);
 459
 460 }
 461 EXPORT_SYMBOL(dump_send_ipi);
 462 #endif
 463
 464 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 465
 466 static DEFINE_PER_CPU(atomic_t, tick_broadcast_count);
 467 static DEFINE_PER_CPU(struct call_single_data, tick_broadcast_csd);
 468
 469 void tick_broadcast(const struct cpumask *mask)
 470 {
 471         atomic_t *count;
 472         struct call_single_data *csd;
 473         int cpu;
 474
 475         for_each_cpu(cpu, mask) {
 476                 count = &per_cpu(tick_broadcast_count, cpu);
 477                 csd = &per_cpu(tick_broadcast_csd, cpu);
 478
 479                 if (atomic_inc_return(count) == 1)
 480                         smp_call_function_single_async(cpu, csd);
 481         }
 482 }
 483
 484 static void tick_broadcast_callee(void *info)
 485 {
 486         int cpu = smp_processor_id();
 487         tick_receive_broadcast();
 488         atomic_set(&per_cpu(tick_broadcast_count, cpu), 0);
 489 }
 490
 491 static int __init tick_broadcast_init(void)
 492 {
 493         struct call_single_data *csd;
 494         int cpu;
 495
 496         for (cpu = 0; cpu < NR_CPUS; cpu++) {
 497                 csd = &per_cpu(tick_broadcast_csd, cpu);
 498                 csd->func = tick_broadcast_callee;
 499         }
 500
 501         return 0;
 502 }
 503 early_initcall(tick_broadcast_init);
 504
 505 #endif /* CONFIG_GENERIC_CLOCKEVENTS_BROADCAST */