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 #include <linux/irqdomain.h>
  37 #include <linux/of.h>
  38 #include <linux/of_irq.h>
  39
  40 #include <linux/atomic.h>
  41 #include <asm/cpu.h>
  42 #include <asm/processor.h>
  43 #include <asm/idle.h>
  44 #include <asm/r4k-timer.h>
  45 #include <asm/mips-cpc.h>
  46 #include <asm/mmu_context.h>
  47 #include <asm/time.h>
  48 #include <asm/setup.h>
  49 #include <asm/maar.h>
  50
  51 cpumask_t cpu_callin_map;               /* Bitmask of started secondaries */
  52
  53 int __cpu_number_map[NR_CPUS];          /* Map physical to logical */
  54 EXPORT_SYMBOL(__cpu_number_map);
  55
  56 int __cpu_logical_map[NR_CPUS];         /* Map logical to physical */
  57 EXPORT_SYMBOL(__cpu_logical_map);
  58
  59 /* Number of TCs (or siblings in Intel speak) per CPU core */
  60 int smp_num_siblings = 1;
  61 EXPORT_SYMBOL(smp_num_siblings);
  62
  63 /* representing the TCs (or siblings in Intel speak) of each logical CPU */
  64 cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly;
  65 EXPORT_SYMBOL(cpu_sibling_map);
  66
  67 /* representing the core map of multi-core chips of each logical CPU */
  68 cpumask_t cpu_core_map[NR_CPUS] __read_mostly;
  69 EXPORT_SYMBOL(cpu_core_map);
  70
  71 /*
  72  * A logcal cpu mask containing only one VPE per core to
  73  * reduce the number of IPIs on large MT systems.
  74  */
  75 cpumask_t cpu_foreign_map __read_mostly;
  76 EXPORT_SYMBOL(cpu_foreign_map);
  77
  78 /* representing cpus for which sibling maps can be computed */
  79 static cpumask_t cpu_sibling_setup_map;
  80
  81 /* representing cpus for which core maps can be computed */
  82 static cpumask_t cpu_core_setup_map;
  83
  84 cpumask_t cpu_coherent_mask;
  85
  86 #ifdef CONFIG_GENERIC_IRQ_IPI
  87 static struct irq_desc *call_desc;
  88 static struct irq_desc *sched_desc;
  89 #endif
  90
  91 static inline void set_cpu_sibling_map(int cpu)
  92 {
  93         int i;
  94
  95         cpumask_set_cpu(cpu, &cpu_sibling_setup_map);
  96
  97         if (smp_num_siblings > 1) {
  98                 for_each_cpu(i, &cpu_sibling_setup_map) {
  99                         if (cpu_data[cpu].package == cpu_data[i].package &&
 100                                     cpu_data[cpu].core == cpu_data[i].core) {
 101                                 cpumask_set_cpu(i, &cpu_sibling_map[cpu]);
 102                                 cpumask_set_cpu(cpu, &cpu_sibling_map[i]);
 103                         }
 104                 }
 105         } else
 106                 cpumask_set_cpu(cpu, &cpu_sibling_map[cpu]);
 107 }
 108
 109 static inline void set_cpu_core_map(int cpu)
 110 {
 111         int i;
 112
 113         cpumask_set_cpu(cpu, &cpu_core_setup_map);
 114
 115         for_each_cpu(i, &cpu_core_setup_map) {
 116                 if (cpu_data[cpu].package == cpu_data[i].package) {
 117                         cpumask_set_cpu(i, &cpu_core_map[cpu]);
 118                         cpumask_set_cpu(cpu, &cpu_core_map[i]);
 119                 }
 120         }
 121 }
 122
 123 /*
 124  * Calculate a new cpu_foreign_map mask whenever a
 125  * new cpu appears or disappears.
 126  */
 127 static inline void calculate_cpu_foreign_map(void)
 128 {
 129         int i, k, core_present;
 130         cpumask_t temp_foreign_map;
 131
 132         /* Re-calculate the mask */
 133         cpumask_clear(&temp_foreign_map);
 134         for_each_online_cpu(i) {
 135                 core_present = 0;
 136                 for_each_cpu(k, &temp_foreign_map)
 137                         if (cpu_data[i].package == cpu_data[k].package &&
 138                             cpu_data[i].core == cpu_data[k].core)
 139                                 core_present = 1;
 140                 if (!core_present)
 141                         cpumask_set_cpu(i, &temp_foreign_map);
 142         }
 143
 144         cpumask_copy(&cpu_foreign_map, &temp_foreign_map);
 145 }
 146
 147 struct plat_smp_ops *mp_ops;
 148 EXPORT_SYMBOL(mp_ops);
 149
 150 void register_smp_ops(struct plat_smp_ops *ops)
 151 {
 152         if (mp_ops)
 153                 printk(KERN_WARNING "Overriding previously set SMP ops\n");
 154
 155         mp_ops = ops;
 156 }
 157
 158 #ifdef CONFIG_GENERIC_IRQ_IPI
 159 void mips_smp_send_ipi_single(int cpu, unsigned int action)
 160 {
 161         mips_smp_send_ipi_mask(cpumask_of(cpu), action);
 162 }
 163
 164 void mips_smp_send_ipi_mask(const struct cpumask *mask, unsigned int action)
 165 {
 166         unsigned long flags;
 167         unsigned int core;
 168         int cpu;
 169
 170         local_irq_save(flags);
 171
 172         switch (action) {
 173         case SMP_CALL_FUNCTION:
 174                 __ipi_send_mask(call_desc, mask);
 175                 break;
 176
 177         case SMP_RESCHEDULE_YOURSELF:
 178                 __ipi_send_mask(sched_desc, mask);
 179                 break;
 180
 181         default:
 182                 BUG();
 183         }
 184
 185         if (mips_cpc_present()) {
 186                 for_each_cpu(cpu, mask) {
 187                         core = cpu_data[cpu].core;
 188
 189                         if (core == current_cpu_data.core)
 190                                 continue;
 191
 192                         while (!cpumask_test_cpu(cpu, &cpu_coherent_mask)) {
 193                                 mips_cpc_lock_other(core);
 194                                 write_cpc_co_cmd(CPC_Cx_CMD_PWRUP);
 195                                 mips_cpc_unlock_other();
 196                         }
 197                 }
 198         }
 199
 200         local_irq_restore(flags);
 201 }
 202
 203
 204 static irqreturn_t ipi_resched_interrupt(int irq, void *dev_id)
 205 {
 206         scheduler_ipi();
 207
 208         return IRQ_HANDLED;
 209 }
 210
 211 static irqreturn_t ipi_call_interrupt(int irq, void *dev_id)
 212 {
 213         generic_smp_call_function_interrupt();
 214
 215         return IRQ_HANDLED;
 216 }
 217
 218 static struct irqaction irq_resched = {
 219         .handler        = ipi_resched_interrupt,
 220         .flags          = IRQF_PERCPU,
 221         .name           = "IPI resched"
 222 };
 223
 224 static struct irqaction irq_call = {
 225         .handler        = ipi_call_interrupt,
 226         .flags          = IRQF_PERCPU,
 227         .name           = "IPI call"
 228 };
 229
 230 static __init void smp_ipi_init_one(unsigned int virq,
 231                                     struct irqaction *action)
 232 {
 233         int ret;
 234
 235         irq_set_handler(virq, handle_percpu_irq);
 236         ret = setup_irq(virq, action);
 237         BUG_ON(ret);
 238 }
 239
 240 static int __init mips_smp_ipi_init(void)
 241 {
 242         unsigned int call_virq, sched_virq;
 243         struct irq_domain *ipidomain;
 244         struct device_node *node;
 245
 246         /*
 247          * In some cases like qemu-malta, it is desired to try SMP with
 248          * a single core. Qemu-malta has no GIC, so an attempt to set any IPIs
 249          * would cause a BUG_ON() to be triggered since there's no ipidomain.
 250          *
 251          * Since for a single core system IPIs aren't required really, skip the
 252          * initialisation which should generally keep any such configurations
 253          * happy and only fail hard when trying to truely run SMP.
 254          */
 255         if (cpumask_weight(cpu_possible_mask) == 1)
 256                 return 0;
 257
 258         node = of_irq_find_parent(of_root);
 259         ipidomain = irq_find_matching_host(node, DOMAIN_BUS_IPI);
 260
 261         /*
 262          * Some platforms have half DT setup. So if we found irq node but
 263          * didn't find an ipidomain, try to search for one that is not in the
 264          * DT.
 265          */
 266         if (node && !ipidomain)
 267                 ipidomain = irq_find_matching_host(NULL, DOMAIN_BUS_IPI);
 268
 269         BUG_ON(!ipidomain);
 270
 271         call_virq = irq_reserve_ipi(ipidomain, cpu_possible_mask);
 272         BUG_ON(!call_virq);
 273
 274         sched_virq = irq_reserve_ipi(ipidomain, cpu_possible_mask);
 275         BUG_ON(!sched_virq);
 276
 277         if (irq_domain_is_ipi_per_cpu(ipidomain)) {
 278                 int cpu;
 279
 280                 for_each_cpu(cpu, cpu_possible_mask) {
 281                         smp_ipi_init_one(call_virq + cpu, &irq_call);
 282                         smp_ipi_init_one(sched_virq + cpu, &irq_resched);
 283                 }
 284         } else {
 285                 smp_ipi_init_one(call_virq, &irq_call);
 286                 smp_ipi_init_one(sched_virq, &irq_resched);
 287         }
 288
 289         call_desc = irq_to_desc(call_virq);
 290         sched_desc = irq_to_desc(sched_virq);
 291
 292         return 0;
 293 }
 294 early_initcall(mips_smp_ipi_init);
 295 #endif
 296
 297 /*
 298  * First C code run on the secondary CPUs after being started up by
 299  * the master.
 300  */
 301 asmlinkage void start_secondary(void)
 302 {
 303         unsigned int cpu;
 304
 305         cpu_probe();
 306         per_cpu_trap_init(false);
 307         mips_clockevent_init();
 308         mp_ops->init_secondary();
 309         cpu_report();
 310         maar_init();
 311
 312         /*
 313          * XXX parity protection should be folded in here when it's converted
 314          * to an option instead of something based on .cputype
 315          */
 316
 317         calibrate_delay();
 318         preempt_disable();
 319         cpu = smp_processor_id();
 320         cpu_data[cpu].udelay_val = loops_per_jiffy;
 321
 322         cpumask_set_cpu(cpu, &cpu_coherent_mask);
 323         notify_cpu_starting(cpu);
 324
 325         set_cpu_online(cpu, true);
 326
 327         set_cpu_sibling_map(cpu);
 328         set_cpu_core_map(cpu);
 329
 330         calculate_cpu_foreign_map();
 331
 332         cpumask_set_cpu(cpu, &cpu_callin_map);
 333
 334         synchronise_count_slave(cpu);
 335
 336         /*
 337          * irq will be enabled in ->smp_finish(), enabling it too early
 338          * is dangerous.
 339          */
 340         WARN_ON_ONCE(!irqs_disabled());
 341         mp_ops->smp_finish();
 342
 343         cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
 344 }
 345
 346 static void stop_this_cpu(void *dummy)
 347 {
 348         /*
 349          * Remove this CPU. Be a bit slow here and
 350          * set the bits for every online CPU so we don't miss
 351          * any IPI whilst taking this VPE down.
 352          */
 353
 354         cpumask_copy(&cpu_foreign_map, cpu_online_mask);
 355
 356         /* Make it visible to every other CPU */
 357         smp_mb();
 358
 359         set_cpu_online(smp_processor_id(), false);
 360         calculate_cpu_foreign_map();
 361         local_irq_disable();
 362         while (1);
 363 }
 364
 365 void smp_send_stop(void)
 366 {
 367         smp_call_function(stop_this_cpu, NULL, 0);
 368 }
 369
 370 void __init smp_cpus_done(unsigned int max_cpus)
 371 {
 372 }
 373
 374 /* called from main before smp_init() */
 375 void __init smp_prepare_cpus(unsigned int max_cpus)
 376 {
 377         init_new_context(current, &init_mm);
 378         current_thread_info()->cpu = 0;
 379         mp_ops->prepare_cpus(max_cpus);
 380         set_cpu_sibling_map(0);
 381         set_cpu_core_map(0);
 382         calculate_cpu_foreign_map();
 383 #ifndef CONFIG_HOTPLUG_CPU
 384         init_cpu_present(cpu_possible_mask);
 385 #endif
 386         cpumask_copy(&cpu_coherent_mask, cpu_possible_mask);
 387 }
 388
 389 /* preload SMP state for boot cpu */
 390 void smp_prepare_boot_cpu(void)
 391 {
 392         set_cpu_possible(0, true);
 393         set_cpu_online(0, true);
 394         cpumask_set_cpu(0, &cpu_callin_map);
 395 }
 396
 397 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
 398 {
 399         mp_ops->boot_secondary(cpu, tidle);
 400
 401         /*
 402          * Trust is futile.  We should really have timeouts ...
 403          */
 404         while (!cpumask_test_cpu(cpu, &cpu_callin_map)) {
 405                 udelay(100);
 406                 schedule();
 407         }
 408
 409         synchronise_count_master(cpu);
 410         return 0;
 411 }
 412
 413 /* Not really SMP stuff ... */
 414 int setup_profiling_timer(unsigned int multiplier)
 415 {
 416         return 0;
 417 }
 418
 419 static void flush_tlb_all_ipi(void *info)
 420 {
 421         local_flush_tlb_all();
 422 }
 423
 424 void flush_tlb_all(void)
 425 {
 426         on_each_cpu(flush_tlb_all_ipi, NULL, 1);
 427 }
 428
 429 static void flush_tlb_mm_ipi(void *mm)
 430 {
 431         local_flush_tlb_mm((struct mm_struct *)mm);
 432 }
 433
 434 /*
 435  * Special Variant of smp_call_function for use by TLB functions:
 436  *
 437  *  o No return value
 438  *  o collapses to normal function call on UP kernels
 439  *  o collapses to normal function call on systems with a single shared
 440  *    primary cache.
 441  */
 442 static inline void smp_on_other_tlbs(void (*func) (void *info), void *info)
 443 {
 444         smp_call_function(func, info, 1);
 445 }
 446
 447 static inline void smp_on_each_tlb(void (*func) (void *info), void *info)
 448 {
 449         preempt_disable();
 450
 451         smp_on_other_tlbs(func, info);
 452         func(info);
 453
 454         preempt_enable();
 455 }
 456
 457 /*
 458  * The following tlb flush calls are invoked when old translations are
 459  * being torn down, or pte attributes are changing. For single threaded
 460  * address spaces, a new context is obtained on the current cpu, and tlb
 461  * context on other cpus are invalidated to force a new context allocation
 462  * at switch_mm time, should the mm ever be used on other cpus. For
 463  * multithreaded address spaces, intercpu interrupts have to be sent.
 464  * Another case where intercpu interrupts are required is when the target
 465  * mm might be active on another cpu (eg debuggers doing the flushes on
 466  * behalf of debugees, kswapd stealing pages from another process etc).
 467  * Kanoj 07/00.
 468  */
 469
 470 void flush_tlb_mm(struct mm_struct *mm)
 471 {
 472         preempt_disable();
 473
 474         if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
 475                 smp_on_other_tlbs(flush_tlb_mm_ipi, mm);
 476         } else {
 477                 unsigned int cpu;
 478
 479                 for_each_online_cpu(cpu) {
 480                         if (cpu != smp_processor_id() && cpu_context(cpu, mm))
 481                                 cpu_context(cpu, mm) = 0;
 482                 }
 483         }
 484         local_flush_tlb_mm(mm);
 485
 486         preempt_enable();
 487 }
 488
 489 struct flush_tlb_data {
 490         struct vm_area_struct *vma;
 491         unsigned long addr1;
 492         unsigned long addr2;
 493 };
 494
 495 static void flush_tlb_range_ipi(void *info)
 496 {
 497         struct flush_tlb_data *fd = info;
 498
 499         local_flush_tlb_range(fd->vma, fd->addr1, fd->addr2);
 500 }
 501
 502 void flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
 503 {
 504         struct mm_struct *mm = vma->vm_mm;
 505
 506         preempt_disable();
 507         if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
 508                 struct flush_tlb_data fd = {
 509                         .vma = vma,
 510                         .addr1 = start,
 511                         .addr2 = end,
 512                 };
 513
 514                 smp_on_other_tlbs(flush_tlb_range_ipi, &fd);
 515         } else {
 516                 unsigned int cpu;
 517
 518                 for_each_online_cpu(cpu) {
 519                         if (cpu != smp_processor_id() && cpu_context(cpu, mm))
 520                                 cpu_context(cpu, mm) = 0;
 521                 }
 522         }
 523         local_flush_tlb_range(vma, start, end);
 524         preempt_enable();
 525 }
 526
 527 static void flush_tlb_kernel_range_ipi(void *info)
 528 {
 529         struct flush_tlb_data *fd = info;
 530
 531         local_flush_tlb_kernel_range(fd->addr1, fd->addr2);
 532 }
 533
 534 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 535 {
 536         struct flush_tlb_data fd = {
 537                 .addr1 = start,
 538                 .addr2 = end,
 539         };
 540
 541         on_each_cpu(flush_tlb_kernel_range_ipi, &fd, 1);
 542 }
 543
 544 static void flush_tlb_page_ipi(void *info)
 545 {
 546         struct flush_tlb_data *fd = info;
 547
 548         local_flush_tlb_page(fd->vma, fd->addr1);
 549 }
 550
 551 void flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 552 {
 553         preempt_disable();
 554         if ((atomic_read(&vma->vm_mm->mm_users) != 1) || (current->mm != vma->vm_mm)) {
 555                 struct flush_tlb_data fd = {
 556                         .vma = vma,
 557                         .addr1 = page,
 558                 };
 559
 560                 smp_on_other_tlbs(flush_tlb_page_ipi, &fd);
 561         } else {
 562                 unsigned int cpu;
 563
 564                 for_each_online_cpu(cpu) {
 565                         if (cpu != smp_processor_id() && cpu_context(cpu, vma->vm_mm))
 566                                 cpu_context(cpu, vma->vm_mm) = 0;
 567                 }
 568         }
 569         local_flush_tlb_page(vma, page);
 570         preempt_enable();
 571 }
 572
 573 static void flush_tlb_one_ipi(void *info)
 574 {
 575         unsigned long vaddr = (unsigned long) info;
 576
 577         local_flush_tlb_one(vaddr);
 578 }
 579
 580 void flush_tlb_one(unsigned long vaddr)
 581 {
 582         smp_on_each_tlb(flush_tlb_one_ipi, (void *) vaddr);
 583 }
 584
 585 EXPORT_SYMBOL(flush_tlb_page);
 586 EXPORT_SYMBOL(flush_tlb_one);
 587
 588 #if defined(CONFIG_KEXEC)
 589 void (*dump_ipi_function_ptr)(void *) = NULL;
 590 void dump_send_ipi(void (*dump_ipi_callback)(void *))
 591 {
 592         int i;
 593         int cpu = smp_processor_id();
 594
 595         dump_ipi_function_ptr = dump_ipi_callback;
 596         smp_mb();
 597         for_each_online_cpu(i)
 598                 if (i != cpu)
 599                         mp_ops->send_ipi_single(i, SMP_DUMP);
 600
 601 }
 602 EXPORT_SYMBOL(dump_send_ipi);
 603 #endif
 604
 605 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 606
 607 static DEFINE_PER_CPU(atomic_t, tick_broadcast_count);
 608 static DEFINE_PER_CPU(struct call_single_data, tick_broadcast_csd);
 609
 610 void tick_broadcast(const struct cpumask *mask)
 611 {
 612         atomic_t *count;
 613         struct call_single_data *csd;
 614         int cpu;
 615
 616         for_each_cpu(cpu, mask) {
 617                 count = &per_cpu(tick_broadcast_count, cpu);
 618                 csd = &per_cpu(tick_broadcast_csd, cpu);
 619
 620                 if (atomic_inc_return(count) == 1)
 621                         smp_call_function_single_async(cpu, csd);
 622         }
 623 }
 624
 625 static void tick_broadcast_callee(void *info)
 626 {
 627         int cpu = smp_processor_id();
 628         tick_receive_broadcast();
 629         atomic_set(&per_cpu(tick_broadcast_count, cpu), 0);
 630 }
 631
 632 static int __init tick_broadcast_init(void)
 633 {
 634         struct call_single_data *csd;
 635         int cpu;
 636
 637         for (cpu = 0; cpu < NR_CPUS; cpu++) {
 638                 csd = &per_cpu(tick_broadcast_csd, cpu);
 639                 csd->func = tick_broadcast_callee;
 640         }
 641
 642         return 0;
 643 }
 644 early_initcall(tick_broadcast_init);
 645
 646 #endif /* CONFIG_GENERIC_CLOCKEVENTS_BROADCAST */