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         node = of_irq_find_parent(of_root);
 247         ipidomain = irq_find_matching_host(node, DOMAIN_BUS_IPI);
 248
 249         /*
 250          * Some platforms have half DT setup. So if we found irq node but
 251          * didn't find an ipidomain, try to search for one that is not in the
 252          * DT.
 253          */
 254         if (node && !ipidomain)
 255                 ipidomain = irq_find_matching_host(NULL, DOMAIN_BUS_IPI);
 256
 257         /*
 258          * There are systems which only use IPI domains some of the time,
 259          * depending upon configuration we don't know until runtime. An
 260          * example is Malta where we may compile in support for GIC & the
 261          * MT ASE, but run on a system which has multiple VPEs in a single
 262          * core and doesn't include a GIC. Until all IPI implementations
 263          * have been converted to use IPI domains the best we can do here
 264          * is to return & hope some other code sets up the IPIs.
 265          */
 266         if (!ipidomain)
 267                 return 0;
 268
 269         call_virq = irq_reserve_ipi(ipidomain, cpu_possible_mask);
 270         BUG_ON(!call_virq);
 271
 272         sched_virq = irq_reserve_ipi(ipidomain, cpu_possible_mask);
 273         BUG_ON(!sched_virq);
 274
 275         if (irq_domain_is_ipi_per_cpu(ipidomain)) {
 276                 int cpu;
 277
 278                 for_each_cpu(cpu, cpu_possible_mask) {
 279                         smp_ipi_init_one(call_virq + cpu, &irq_call);
 280                         smp_ipi_init_one(sched_virq + cpu, &irq_resched);
 281                 }
 282         } else {
 283                 smp_ipi_init_one(call_virq, &irq_call);
 284                 smp_ipi_init_one(sched_virq, &irq_resched);
 285         }
 286
 287         call_desc = irq_to_desc(call_virq);
 288         sched_desc = irq_to_desc(sched_virq);
 289
 290         return 0;
 291 }
 292 early_initcall(mips_smp_ipi_init);
 293 #endif
 294
 295 /*
 296  * First C code run on the secondary CPUs after being started up by
 297  * the master.
 298  */
 299 asmlinkage void start_secondary(void)
 300 {
 301         unsigned int cpu;
 302
 303         cpu_probe();
 304         per_cpu_trap_init(false);
 305         mips_clockevent_init();
 306         mp_ops->init_secondary();
 307         cpu_report();
 308         maar_init();
 309
 310         /*
 311          * XXX parity protection should be folded in here when it's converted
 312          * to an option instead of something based on .cputype
 313          */
 314
 315         calibrate_delay();
 316         preempt_disable();
 317         cpu = smp_processor_id();
 318         cpu_data[cpu].udelay_val = loops_per_jiffy;
 319
 320         cpumask_set_cpu(cpu, &cpu_coherent_mask);
 321         notify_cpu_starting(cpu);
 322
 323         set_cpu_online(cpu, true);
 324
 325         set_cpu_sibling_map(cpu);
 326         set_cpu_core_map(cpu);
 327
 328         calculate_cpu_foreign_map();
 329
 330         cpumask_set_cpu(cpu, &cpu_callin_map);
 331
 332         synchronise_count_slave(cpu);
 333
 334         /*
 335          * irq will be enabled in ->smp_finish(), enabling it too early
 336          * is dangerous.
 337          */
 338         WARN_ON_ONCE(!irqs_disabled());
 339         mp_ops->smp_finish();
 340
 341         cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
 342 }
 343
 344 static void stop_this_cpu(void *dummy)
 345 {
 346         /*
 347          * Remove this CPU. Be a bit slow here and
 348          * set the bits for every online CPU so we don't miss
 349          * any IPI whilst taking this VPE down.
 350          */
 351
 352         cpumask_copy(&cpu_foreign_map, cpu_online_mask);
 353
 354         /* Make it visible to every other CPU */
 355         smp_mb();
 356
 357         set_cpu_online(smp_processor_id(), false);
 358         calculate_cpu_foreign_map();
 359         local_irq_disable();
 360         while (1);
 361 }
 362
 363 void smp_send_stop(void)
 364 {
 365         smp_call_function(stop_this_cpu, NULL, 0);
 366 }
 367
 368 void __init smp_cpus_done(unsigned int max_cpus)
 369 {
 370 }
 371
 372 /* called from main before smp_init() */
 373 void __init smp_prepare_cpus(unsigned int max_cpus)
 374 {
 375         init_new_context(current, &init_mm);
 376         current_thread_info()->cpu = 0;
 377         mp_ops->prepare_cpus(max_cpus);
 378         set_cpu_sibling_map(0);
 379         set_cpu_core_map(0);
 380         calculate_cpu_foreign_map();
 381 #ifndef CONFIG_HOTPLUG_CPU
 382         init_cpu_present(cpu_possible_mask);
 383 #endif
 384         cpumask_copy(&cpu_coherent_mask, cpu_possible_mask);
 385 }
 386
 387 /* preload SMP state for boot cpu */
 388 void smp_prepare_boot_cpu(void)
 389 {
 390         set_cpu_possible(0, true);
 391         set_cpu_online(0, true);
 392         cpumask_set_cpu(0, &cpu_callin_map);
 393 }
 394
 395 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
 396 {
 397         mp_ops->boot_secondary(cpu, tidle);
 398
 399         /*
 400          * Trust is futile.  We should really have timeouts ...
 401          */
 402         while (!cpumask_test_cpu(cpu, &cpu_callin_map)) {
 403                 udelay(100);
 404                 schedule();
 405         }
 406
 407         synchronise_count_master(cpu);
 408         return 0;
 409 }
 410
 411 /* Not really SMP stuff ... */
 412 int setup_profiling_timer(unsigned int multiplier)
 413 {
 414         return 0;
 415 }
 416
 417 static void flush_tlb_all_ipi(void *info)
 418 {
 419         local_flush_tlb_all();
 420 }
 421
 422 void flush_tlb_all(void)
 423 {
 424         on_each_cpu(flush_tlb_all_ipi, NULL, 1);
 425 }
 426
 427 static void flush_tlb_mm_ipi(void *mm)
 428 {
 429         local_flush_tlb_mm((struct mm_struct *)mm);
 430 }
 431
 432 /*
 433  * Special Variant of smp_call_function for use by TLB functions:
 434  *
 435  *  o No return value
 436  *  o collapses to normal function call on UP kernels
 437  *  o collapses to normal function call on systems with a single shared
 438  *    primary cache.
 439  */
 440 static inline void smp_on_other_tlbs(void (*func) (void *info), void *info)
 441 {
 442         smp_call_function(func, info, 1);
 443 }
 444
 445 static inline void smp_on_each_tlb(void (*func) (void *info), void *info)
 446 {
 447         preempt_disable();
 448
 449         smp_on_other_tlbs(func, info);
 450         func(info);
 451
 452         preempt_enable();
 453 }
 454
 455 /*
 456  * The following tlb flush calls are invoked when old translations are
 457  * being torn down, or pte attributes are changing. For single threaded
 458  * address spaces, a new context is obtained on the current cpu, and tlb
 459  * context on other cpus are invalidated to force a new context allocation
 460  * at switch_mm time, should the mm ever be used on other cpus. For
 461  * multithreaded address spaces, intercpu interrupts have to be sent.
 462  * Another case where intercpu interrupts are required is when the target
 463  * mm might be active on another cpu (eg debuggers doing the flushes on
 464  * behalf of debugees, kswapd stealing pages from another process etc).
 465  * Kanoj 07/00.
 466  */
 467
 468 void flush_tlb_mm(struct mm_struct *mm)
 469 {
 470         preempt_disable();
 471
 472         if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
 473                 smp_on_other_tlbs(flush_tlb_mm_ipi, mm);
 474         } else {
 475                 unsigned int cpu;
 476
 477                 for_each_online_cpu(cpu) {
 478                         if (cpu != smp_processor_id() && cpu_context(cpu, mm))
 479                                 cpu_context(cpu, mm) = 0;
 480                 }
 481         }
 482         local_flush_tlb_mm(mm);
 483
 484         preempt_enable();
 485 }
 486
 487 struct flush_tlb_data {
 488         struct vm_area_struct *vma;
 489         unsigned long addr1;
 490         unsigned long addr2;
 491 };
 492
 493 static void flush_tlb_range_ipi(void *info)
 494 {
 495         struct flush_tlb_data *fd = info;
 496
 497         local_flush_tlb_range(fd->vma, fd->addr1, fd->addr2);
 498 }
 499
 500 void flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
 501 {
 502         struct mm_struct *mm = vma->vm_mm;
 503
 504         preempt_disable();
 505         if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
 506                 struct flush_tlb_data fd = {
 507                         .vma = vma,
 508                         .addr1 = start,
 509                         .addr2 = end,
 510                 };
 511
 512                 smp_on_other_tlbs(flush_tlb_range_ipi, &fd);
 513         } else {
 514                 unsigned int cpu;
 515
 516                 for_each_online_cpu(cpu) {
 517                         if (cpu != smp_processor_id() && cpu_context(cpu, mm))
 518                                 cpu_context(cpu, mm) = 0;
 519                 }
 520         }
 521         local_flush_tlb_range(vma, start, end);
 522         preempt_enable();
 523 }
 524
 525 static void flush_tlb_kernel_range_ipi(void *info)
 526 {
 527         struct flush_tlb_data *fd = info;
 528
 529         local_flush_tlb_kernel_range(fd->addr1, fd->addr2);
 530 }
 531
 532 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 533 {
 534         struct flush_tlb_data fd = {
 535                 .addr1 = start,
 536                 .addr2 = end,
 537         };
 538
 539         on_each_cpu(flush_tlb_kernel_range_ipi, &fd, 1);
 540 }
 541
 542 static void flush_tlb_page_ipi(void *info)
 543 {
 544         struct flush_tlb_data *fd = info;
 545
 546         local_flush_tlb_page(fd->vma, fd->addr1);
 547 }
 548
 549 void flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 550 {
 551         preempt_disable();
 552         if ((atomic_read(&vma->vm_mm->mm_users) != 1) || (current->mm != vma->vm_mm)) {
 553                 struct flush_tlb_data fd = {
 554                         .vma = vma,
 555                         .addr1 = page,
 556                 };
 557
 558                 smp_on_other_tlbs(flush_tlb_page_ipi, &fd);
 559         } else {
 560                 unsigned int cpu;
 561
 562                 for_each_online_cpu(cpu) {
 563                         if (cpu != smp_processor_id() && cpu_context(cpu, vma->vm_mm))
 564                                 cpu_context(cpu, vma->vm_mm) = 0;
 565                 }
 566         }
 567         local_flush_tlb_page(vma, page);
 568         preempt_enable();
 569 }
 570
 571 static void flush_tlb_one_ipi(void *info)
 572 {
 573         unsigned long vaddr = (unsigned long) info;
 574
 575         local_flush_tlb_one(vaddr);
 576 }
 577
 578 void flush_tlb_one(unsigned long vaddr)
 579 {
 580         smp_on_each_tlb(flush_tlb_one_ipi, (void *) vaddr);
 581 }
 582
 583 EXPORT_SYMBOL(flush_tlb_page);
 584 EXPORT_SYMBOL(flush_tlb_one);
 585
 586 #if defined(CONFIG_KEXEC)
 587 void (*dump_ipi_function_ptr)(void *) = NULL;
 588 void dump_send_ipi(void (*dump_ipi_callback)(void *))
 589 {
 590         int i;
 591         int cpu = smp_processor_id();
 592
 593         dump_ipi_function_ptr = dump_ipi_callback;
 594         smp_mb();
 595         for_each_online_cpu(i)
 596                 if (i != cpu)
 597                         mp_ops->send_ipi_single(i, SMP_DUMP);
 598
 599 }
 600 EXPORT_SYMBOL(dump_send_ipi);
 601 #endif
 602
 603 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 604
 605 static DEFINE_PER_CPU(atomic_t, tick_broadcast_count);
 606 static DEFINE_PER_CPU(struct call_single_data, tick_broadcast_csd);
 607
 608 void tick_broadcast(const struct cpumask *mask)
 609 {
 610         atomic_t *count;
 611         struct call_single_data *csd;
 612         int cpu;
 613
 614         for_each_cpu(cpu, mask) {
 615                 count = &per_cpu(tick_broadcast_count, cpu);
 616                 csd = &per_cpu(tick_broadcast_csd, cpu);
 617
 618                 if (atomic_inc_return(count) == 1)
 619                         smp_call_function_single_async(cpu, csd);
 620         }
 621 }
 622
 623 static void tick_broadcast_callee(void *info)
 624 {
 625         int cpu = smp_processor_id();
 626         tick_receive_broadcast();
 627         atomic_set(&per_cpu(tick_broadcast_count, cpu), 0);
 628 }
 629
 630 static int __init tick_broadcast_init(void)
 631 {
 632         struct call_single_data *csd;
 633         int cpu;
 634
 635         for (cpu = 0; cpu < NR_CPUS; cpu++) {
 636                 csd = &per_cpu(tick_broadcast_csd, cpu);
 637                 csd->func = tick_broadcast_callee;
 638         }
 639
 640         return 0;
 641 }
 642 early_initcall(tick_broadcast_init);
 643
 644 #endif /* CONFIG_GENERIC_CLOCKEVENTS_BROADCAST */