arch/loongarch/kernel/smp.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Copyright (C) 2020-2022 Loongson Technology Corporation Limited
   4  *
   5  * Derived from MIPS:
   6  * Copyright (C) 2000, 2001 Kanoj Sarcar
   7  * Copyright (C) 2000, 2001 Ralf Baechle
   8  * Copyright (C) 2000, 2001 Silicon Graphics, Inc.
   9  * Copyright (C) 2000, 2001, 2003 Broadcom Corporation
  10  */
  11 #include <linux/cpu.h>
  12 #include <linux/cpumask.h>
  13 #include <linux/init.h>
  14 #include <linux/interrupt.h>
  15 #include <linux/seq_file.h>
  16 #include <linux/smp.h>
  17 #include <linux/threads.h>
  18 #include <linux/export.h>
  19 #include <linux/time.h>
  20 #include <linux/tracepoint.h>
  21 #include <linux/sched/hotplug.h>
  22 #include <linux/sched/task_stack.h>
  23
  24 #include <asm/cpu.h>
  25 #include <asm/idle.h>
  26 #include <asm/loongson.h>
  27 #include <asm/mmu_context.h>
  28 #include <asm/numa.h>
  29 #include <asm/processor.h>
  30 #include <asm/setup.h>
  31 #include <asm/time.h>
  32
  33 int __cpu_number_map[NR_CPUS];   /* Map physical to logical */
  34 EXPORT_SYMBOL(__cpu_number_map);
  35
  36 int __cpu_logical_map[NR_CPUS];         /* Map logical to physical */
  37 EXPORT_SYMBOL(__cpu_logical_map);
  38
  39 /* Number of threads (siblings) per CPU core */
  40 int smp_num_siblings = 1;
  41 EXPORT_SYMBOL(smp_num_siblings);
  42
  43 /* Representing the threads (siblings) of each logical CPU */
  44 cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly;
  45 EXPORT_SYMBOL(cpu_sibling_map);
  46
  47 /* Representing the core map of multi-core chips of each logical CPU */
  48 cpumask_t cpu_core_map[NR_CPUS] __read_mostly;
  49 EXPORT_SYMBOL(cpu_core_map);
  50
  51 static DECLARE_COMPLETION(cpu_starting);
  52 static DECLARE_COMPLETION(cpu_running);
  53
  54 /*
  55  * A logcal cpu mask containing only one VPE per core to
  56  * reduce the number of IPIs on large MT systems.
  57  */
  58 cpumask_t cpu_foreign_map[NR_CPUS] __read_mostly;
  59 EXPORT_SYMBOL(cpu_foreign_map);
  60
  61 /* representing cpus for which sibling maps can be computed */
  62 static cpumask_t cpu_sibling_setup_map;
  63
  64 /* representing cpus for which core maps can be computed */
  65 static cpumask_t cpu_core_setup_map;
  66
  67 struct secondary_data cpuboot_data;
  68 static DEFINE_PER_CPU(int, cpu_state);
  69
  70 enum ipi_msg_type {
  71         IPI_RESCHEDULE,
  72         IPI_CALL_FUNCTION,
  73 };
  74
  75 static const char *ipi_types[NR_IPI] __tracepoint_string = {
  76         [IPI_RESCHEDULE] = "Rescheduling interrupts",
  77         [IPI_CALL_FUNCTION] = "Function call interrupts",
  78 };
  79
  80 void show_ipi_list(struct seq_file *p, int prec)
  81 {
  82         unsigned int cpu, i;
  83
  84         for (i = 0; i < NR_IPI; i++) {
  85                 seq_printf(p, "%*s%u:%s", prec - 1, "IPI", i, prec >= 4 ? " " : "");
  86                 for_each_online_cpu(cpu)
  87                         seq_printf(p, "%10u ", per_cpu(irq_stat, cpu).ipi_irqs[i]);
  88                 seq_printf(p, " LoongArch  %d  %s\n", i + 1, ipi_types[i]);
  89         }
  90 }
  91
  92 /* Send mailbox buffer via Mail_Send */
  93 static void csr_mail_send(uint64_t data, int cpu, int mailbox)
  94 {
  95         uint64_t val;
  96
  97         /* Send high 32 bits */
  98         val = IOCSR_MBUF_SEND_BLOCKING;
  99         val |= (IOCSR_MBUF_SEND_BOX_HI(mailbox) << IOCSR_MBUF_SEND_BOX_SHIFT);
 100         val |= (cpu << IOCSR_MBUF_SEND_CPU_SHIFT);
 101         val |= (data & IOCSR_MBUF_SEND_H32_MASK);
 102         iocsr_write64(val, LOONGARCH_IOCSR_MBUF_SEND);
 103
 104         /* Send low 32 bits */
 105         val = IOCSR_MBUF_SEND_BLOCKING;
 106         val |= (IOCSR_MBUF_SEND_BOX_LO(mailbox) << IOCSR_MBUF_SEND_BOX_SHIFT);
 107         val |= (cpu << IOCSR_MBUF_SEND_CPU_SHIFT);
 108         val |= (data << IOCSR_MBUF_SEND_BUF_SHIFT);
 109         iocsr_write64(val, LOONGARCH_IOCSR_MBUF_SEND);
 110 };
 111
 112 static u32 ipi_read_clear(int cpu)
 113 {
 114         u32 action;
 115
 116         /* Load the ipi register to figure out what we're supposed to do */
 117         action = iocsr_read32(LOONGARCH_IOCSR_IPI_STATUS);
 118         /* Clear the ipi register to clear the interrupt */
 119         iocsr_write32(action, LOONGARCH_IOCSR_IPI_CLEAR);
 120         smp_mb();
 121
 122         return action;
 123 }
 124
 125 static void ipi_write_action(int cpu, u32 action)
 126 {
 127         unsigned int irq = 0;
 128
 129         while ((irq = ffs(action))) {
 130                 uint32_t val = IOCSR_IPI_SEND_BLOCKING;
 131
 132                 val |= (irq - 1);
 133                 val |= (cpu << IOCSR_IPI_SEND_CPU_SHIFT);
 134                 iocsr_write32(val, LOONGARCH_IOCSR_IPI_SEND);
 135                 action &= ~BIT(irq - 1);
 136         }
 137 }
 138
 139 void loongson_send_ipi_single(int cpu, unsigned int action)
 140 {
 141         ipi_write_action(cpu_logical_map(cpu), (u32)action);
 142 }
 143
 144 void loongson_send_ipi_mask(const struct cpumask *mask, unsigned int action)
 145 {
 146         unsigned int i;
 147
 148         for_each_cpu(i, mask)
 149                 ipi_write_action(cpu_logical_map(i), (u32)action);
 150 }
 151
 152 /*
 153  * This function sends a 'reschedule' IPI to another CPU.
 154  * it goes straight through and wastes no time serializing
 155  * anything. Worst case is that we lose a reschedule ...
 156  */
 157 void smp_send_reschedule(int cpu)
 158 {
 159         loongson_send_ipi_single(cpu, SMP_RESCHEDULE);
 160 }
 161 EXPORT_SYMBOL_GPL(smp_send_reschedule);
 162
 163 irqreturn_t loongson_ipi_interrupt(int irq, void *dev)
 164 {
 165         unsigned int action;
 166         unsigned int cpu = smp_processor_id();
 167
 168         action = ipi_read_clear(cpu_logical_map(cpu));
 169
 170         if (action & SMP_RESCHEDULE) {
 171                 scheduler_ipi();
 172                 per_cpu(irq_stat, cpu).ipi_irqs[IPI_RESCHEDULE]++;
 173         }
 174
 175         if (action & SMP_CALL_FUNCTION) {
 176                 generic_smp_call_function_interrupt();
 177                 per_cpu(irq_stat, cpu).ipi_irqs[IPI_CALL_FUNCTION]++;
 178         }
 179
 180         return IRQ_HANDLED;
 181 }
 182
 183 void __init loongson_smp_setup(void)
 184 {
 185         cpu_data[0].core = cpu_logical_map(0) % loongson_sysconf.cores_per_package;
 186         cpu_data[0].package = cpu_logical_map(0) / loongson_sysconf.cores_per_package;
 187
 188         iocsr_write32(0xffffffff, LOONGARCH_IOCSR_IPI_EN);
 189         pr_info("Detected %i available CPU(s)\n", loongson_sysconf.nr_cpus);
 190 }
 191
 192 void __init loongson_prepare_cpus(unsigned int max_cpus)
 193 {
 194         int i = 0;
 195
 196         for (i = 0; i < loongson_sysconf.nr_cpus; i++) {
 197                 set_cpu_present(i, true);
 198                 csr_mail_send(0, __cpu_logical_map[i], 0);
 199         }
 200
 201         per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
 202 }
 203
 204 /*
 205  * Setup the PC, SP, and TP of a secondary processor and start it running!
 206  */
 207 void loongson_boot_secondary(int cpu, struct task_struct *idle)
 208 {
 209         unsigned long entry;
 210
 211         pr_info("Booting CPU#%d...\n", cpu);
 212
 213         entry = __pa_symbol((unsigned long)&smpboot_entry);
 214         cpuboot_data.stack = (unsigned long)__KSTK_TOS(idle);
 215         cpuboot_data.thread_info = (unsigned long)task_thread_info(idle);
 216
 217         csr_mail_send(entry, cpu_logical_map(cpu), 0);
 218
 219         loongson_send_ipi_single(cpu, SMP_BOOT_CPU);
 220 }
 221
 222 /*
 223  * SMP init and finish on secondary CPUs
 224  */
 225 void loongson_init_secondary(void)
 226 {
 227         unsigned int cpu = smp_processor_id();
 228         unsigned int imask = ECFGF_IP0 | ECFGF_IP1 | ECFGF_IP2 |
 229                              ECFGF_IPI | ECFGF_PMC | ECFGF_TIMER;
 230
 231         change_csr_ecfg(ECFG0_IM, imask);
 232
 233         iocsr_write32(0xffffffff, LOONGARCH_IOCSR_IPI_EN);
 234
 235 #ifdef CONFIG_NUMA
 236         numa_add_cpu(cpu);
 237 #endif
 238         per_cpu(cpu_state, cpu) = CPU_ONLINE;
 239         cpu_data[cpu].core =
 240                      cpu_logical_map(cpu) % loongson_sysconf.cores_per_package;
 241         cpu_data[cpu].package =
 242                      cpu_logical_map(cpu) / loongson_sysconf.cores_per_package;
 243 }
 244
 245 void loongson_smp_finish(void)
 246 {
 247         local_irq_enable();
 248         iocsr_write64(0, LOONGARCH_IOCSR_MBUF0);
 249         pr_info("CPU#%d finished\n", smp_processor_id());
 250 }
 251
 252 #ifdef CONFIG_HOTPLUG_CPU
 253
 254 int loongson_cpu_disable(void)
 255 {
 256         unsigned long flags;
 257         unsigned int cpu = smp_processor_id();
 258
 259         if (io_master(cpu))
 260                 return -EBUSY;
 261
 262 #ifdef CONFIG_NUMA
 263         numa_remove_cpu(cpu);
 264 #endif
 265         set_cpu_online(cpu, false);
 266         calculate_cpu_foreign_map();
 267         local_irq_save(flags);
 268         irq_migrate_all_off_this_cpu();
 269         clear_csr_ecfg(ECFG0_IM);
 270         local_irq_restore(flags);
 271         local_flush_tlb_all();
 272
 273         return 0;
 274 }
 275
 276 void loongson_cpu_die(unsigned int cpu)
 277 {
 278         while (per_cpu(cpu_state, cpu) != CPU_DEAD)
 279                 cpu_relax();
 280
 281         mb();
 282 }
 283
 284 void play_dead(void)
 285 {
 286         register uint64_t addr;
 287         register void (*init_fn)(void);
 288
 289         idle_task_exit();
 290         local_irq_enable();
 291         set_csr_ecfg(ECFGF_IPI);
 292         __this_cpu_write(cpu_state, CPU_DEAD);
 293
 294         __smp_mb();
 295         do {
 296                 __asm__ __volatile__("idle 0\n\t");
 297                 addr = iocsr_read64(LOONGARCH_IOCSR_MBUF0);
 298         } while (addr == 0);
 299
 300         init_fn = (void *)TO_CACHE(addr);
 301         iocsr_write32(0xffffffff, LOONGARCH_IOCSR_IPI_CLEAR);
 302
 303         init_fn();
 304         unreachable();
 305 }
 306
 307 #endif
 308
 309 /*
 310  * Power management
 311  */
 312 #ifdef CONFIG_PM
 313
 314 static int loongson_ipi_suspend(void)
 315 {
 316         return 0;
 317 }
 318
 319 static void loongson_ipi_resume(void)
 320 {
 321         iocsr_write32(0xffffffff, LOONGARCH_IOCSR_IPI_EN);
 322 }
 323
 324 static struct syscore_ops loongson_ipi_syscore_ops = {
 325         .resume         = loongson_ipi_resume,
 326         .suspend        = loongson_ipi_suspend,
 327 };
 328
 329 /*
 330  * Enable boot cpu ipi before enabling nonboot cpus
 331  * during syscore_resume.
 332  */
 333 static int __init ipi_pm_init(void)
 334 {
 335         register_syscore_ops(&loongson_ipi_syscore_ops);
 336         return 0;
 337 }
 338
 339 core_initcall(ipi_pm_init);
 340 #endif
 341
 342 static inline void set_cpu_sibling_map(int cpu)
 343 {
 344         int i;
 345
 346         cpumask_set_cpu(cpu, &cpu_sibling_setup_map);
 347
 348         if (smp_num_siblings <= 1)
 349                 cpumask_set_cpu(cpu, &cpu_sibling_map[cpu]);
 350         else {
 351                 for_each_cpu(i, &cpu_sibling_setup_map) {
 352                         if (cpus_are_siblings(cpu, i)) {
 353                                 cpumask_set_cpu(i, &cpu_sibling_map[cpu]);
 354                                 cpumask_set_cpu(cpu, &cpu_sibling_map[i]);
 355                         }
 356                 }
 357         }
 358 }
 359
 360 static inline void set_cpu_core_map(int cpu)
 361 {
 362         int i;
 363
 364         cpumask_set_cpu(cpu, &cpu_core_setup_map);
 365
 366         for_each_cpu(i, &cpu_core_setup_map) {
 367                 if (cpu_data[cpu].package == cpu_data[i].package) {
 368                         cpumask_set_cpu(i, &cpu_core_map[cpu]);
 369                         cpumask_set_cpu(cpu, &cpu_core_map[i]);
 370                 }
 371         }
 372 }
 373
 374 /*
 375  * Calculate a new cpu_foreign_map mask whenever a
 376  * new cpu appears or disappears.
 377  */
 378 void calculate_cpu_foreign_map(void)
 379 {
 380         int i, k, core_present;
 381         cpumask_t temp_foreign_map;
 382
 383         /* Re-calculate the mask */
 384         cpumask_clear(&temp_foreign_map);
 385         for_each_online_cpu(i) {
 386                 core_present = 0;
 387                 for_each_cpu(k, &temp_foreign_map)
 388                         if (cpus_are_siblings(i, k))
 389                                 core_present = 1;
 390                 if (!core_present)
 391                         cpumask_set_cpu(i, &temp_foreign_map);
 392         }
 393
 394         for_each_online_cpu(i)
 395                 cpumask_andnot(&cpu_foreign_map[i],
 396                                &temp_foreign_map, &cpu_sibling_map[i]);
 397 }
 398
 399 /* Preload SMP state for boot cpu */
 400 void smp_prepare_boot_cpu(void)
 401 {
 402         unsigned int cpu, node, rr_node;
 403
 404         set_cpu_possible(0, true);
 405         set_cpu_online(0, true);
 406         set_my_cpu_offset(per_cpu_offset(0));
 407
 408         rr_node = first_node(node_online_map);
 409         for_each_possible_cpu(cpu) {
 410                 node = early_cpu_to_node(cpu);
 411
 412                 /*
 413                  * The mapping between present cpus and nodes has been
 414                  * built during MADT and SRAT parsing.
 415                  *
 416                  * If possible cpus = present cpus here, early_cpu_to_node
 417                  * will return valid node.
 418                  *
 419                  * If possible cpus > present cpus here (e.g. some possible
 420                  * cpus will be added by cpu-hotplug later), for possible but
 421                  * not present cpus, early_cpu_to_node will return NUMA_NO_NODE,
 422                  * and we just map them to online nodes in round-robin way.
 423                  * Once hotplugged, new correct mapping will be built for them.
 424                  */
 425                 if (node != NUMA_NO_NODE)
 426                         set_cpu_numa_node(cpu, node);
 427                 else {
 428                         set_cpu_numa_node(cpu, rr_node);
 429                         rr_node = next_node_in(rr_node, node_online_map);
 430                 }
 431         }
 432 }
 433
 434 /* called from main before smp_init() */
 435 void __init smp_prepare_cpus(unsigned int max_cpus)
 436 {
 437         init_new_context(current, &init_mm);
 438         current_thread_info()->cpu = 0;
 439         loongson_prepare_cpus(max_cpus);
 440         set_cpu_sibling_map(0);
 441         set_cpu_core_map(0);
 442         calculate_cpu_foreign_map();
 443 #ifndef CONFIG_HOTPLUG_CPU
 444         init_cpu_present(cpu_possible_mask);
 445 #endif
 446 }
 447
 448 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
 449 {
 450         loongson_boot_secondary(cpu, tidle);
 451
 452         /* Wait for CPU to start and be ready to sync counters */
 453         if (!wait_for_completion_timeout(&cpu_starting,
 454                                          msecs_to_jiffies(5000))) {
 455                 pr_crit("CPU%u: failed to start\n", cpu);
 456                 return -EIO;
 457         }
 458
 459         /* Wait for CPU to finish startup & mark itself online before return */
 460         wait_for_completion(&cpu_running);
 461
 462         return 0;
 463 }
 464
 465 /*
 466  * First C code run on the secondary CPUs after being started up by
 467  * the master.
 468  */
 469 asmlinkage void start_secondary(void)
 470 {
 471         unsigned int cpu;
 472
 473         sync_counter();
 474         cpu = smp_processor_id();
 475         set_my_cpu_offset(per_cpu_offset(cpu));
 476
 477         cpu_probe();
 478         constant_clockevent_init();
 479         loongson_init_secondary();
 480
 481         set_cpu_sibling_map(cpu);
 482         set_cpu_core_map(cpu);
 483
 484         notify_cpu_starting(cpu);
 485
 486         /* Notify boot CPU that we're starting */
 487         complete(&cpu_starting);
 488
 489         /* The CPU is running, now mark it online */
 490         set_cpu_online(cpu, true);
 491
 492         calculate_cpu_foreign_map();
 493
 494         /*
 495          * Notify boot CPU that we're up & online and it can safely return
 496          * from __cpu_up()
 497          */
 498         complete(&cpu_running);
 499
 500         /*
 501          * irq will be enabled in loongson_smp_finish(), enabling it too
 502          * early is dangerous.
 503          */
 504         WARN_ON_ONCE(!irqs_disabled());
 505         loongson_smp_finish();
 506
 507         cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
 508 }
 509
 510 void __init smp_cpus_done(unsigned int max_cpus)
 511 {
 512 }
 513
 514 static void stop_this_cpu(void *dummy)
 515 {
 516         set_cpu_online(smp_processor_id(), false);
 517         calculate_cpu_foreign_map();
 518         local_irq_disable();
 519         while (true);
 520 }
 521
 522 void smp_send_stop(void)
 523 {
 524         smp_call_function(stop_this_cpu, NULL, 0);
 525 }
 526
 527 int setup_profiling_timer(unsigned int multiplier)
 528 {
 529         return 0;
 530 }
 531
 532 static void flush_tlb_all_ipi(void *info)
 533 {
 534         local_flush_tlb_all();
 535 }
 536
 537 void flush_tlb_all(void)
 538 {
 539         on_each_cpu(flush_tlb_all_ipi, NULL, 1);
 540 }
 541
 542 static void flush_tlb_mm_ipi(void *mm)
 543 {
 544         local_flush_tlb_mm((struct mm_struct *)mm);
 545 }
 546
 547 void flush_tlb_mm(struct mm_struct *mm)
 548 {
 549         if (atomic_read(&mm->mm_users) == 0)
 550                 return;         /* happens as a result of exit_mmap() */
 551
 552         preempt_disable();
 553
 554         if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
 555                 on_each_cpu_mask(mm_cpumask(mm), flush_tlb_mm_ipi, mm, 1);
 556         } else {
 557                 unsigned int cpu;
 558
 559                 for_each_online_cpu(cpu) {
 560                         if (cpu != smp_processor_id() && cpu_context(cpu, mm))
 561                                 cpu_context(cpu, mm) = 0;
 562                 }
 563                 local_flush_tlb_mm(mm);
 564         }
 565
 566         preempt_enable();
 567 }
 568
 569 struct flush_tlb_data {
 570         struct vm_area_struct *vma;
 571         unsigned long addr1;
 572         unsigned long addr2;
 573 };
 574
 575 static void flush_tlb_range_ipi(void *info)
 576 {
 577         struct flush_tlb_data *fd = info;
 578
 579         local_flush_tlb_range(fd->vma, fd->addr1, fd->addr2);
 580 }
 581
 582 void flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
 583 {
 584         struct mm_struct *mm = vma->vm_mm;
 585
 586         preempt_disable();
 587         if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
 588                 struct flush_tlb_data fd = {
 589                         .vma = vma,
 590                         .addr1 = start,
 591                         .addr2 = end,
 592                 };
 593
 594                 on_each_cpu_mask(mm_cpumask(mm), flush_tlb_range_ipi, &fd, 1);
 595         } else {
 596                 unsigned int cpu;
 597
 598                 for_each_online_cpu(cpu) {
 599                         if (cpu != smp_processor_id() && cpu_context(cpu, mm))
 600                                 cpu_context(cpu, mm) = 0;
 601                 }
 602                 local_flush_tlb_range(vma, start, end);
 603         }
 604         preempt_enable();
 605 }
 606
 607 static void flush_tlb_kernel_range_ipi(void *info)
 608 {
 609         struct flush_tlb_data *fd = info;
 610
 611         local_flush_tlb_kernel_range(fd->addr1, fd->addr2);
 612 }
 613
 614 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 615 {
 616         struct flush_tlb_data fd = {
 617                 .addr1 = start,
 618                 .addr2 = end,
 619         };
 620
 621         on_each_cpu(flush_tlb_kernel_range_ipi, &fd, 1);
 622 }
 623
 624 static void flush_tlb_page_ipi(void *info)
 625 {
 626         struct flush_tlb_data *fd = info;
 627
 628         local_flush_tlb_page(fd->vma, fd->addr1);
 629 }
 630
 631 void flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 632 {
 633         preempt_disable();
 634         if ((atomic_read(&vma->vm_mm->mm_users) != 1) || (current->mm != vma->vm_mm)) {
 635                 struct flush_tlb_data fd = {
 636                         .vma = vma,
 637                         .addr1 = page,
 638                 };
 639
 640                 on_each_cpu_mask(mm_cpumask(vma->vm_mm), flush_tlb_page_ipi, &fd, 1);
 641         } else {
 642                 unsigned int cpu;
 643
 644                 for_each_online_cpu(cpu) {
 645                         if (cpu != smp_processor_id() && cpu_context(cpu, vma->vm_mm))
 646                                 cpu_context(cpu, vma->vm_mm) = 0;
 647                 }
 648                 local_flush_tlb_page(vma, page);
 649         }
 650         preempt_enable();
 651 }
 652 EXPORT_SYMBOL(flush_tlb_page);
 653
 654 static void flush_tlb_one_ipi(void *info)
 655 {
 656         unsigned long vaddr = (unsigned long) info;
 657
 658         local_flush_tlb_one(vaddr);
 659 }
 660
 661 void flush_tlb_one(unsigned long vaddr)
 662 {
 663         on_each_cpu(flush_tlb_one_ipi, (void *)vaddr, 1);
 664 }
 665 EXPORT_SYMBOL(flush_tlb_one);