1 // SPDX-License-Identifier: GPL-2.0-or-later
4 * Copyright (C) 2000, 2001 Kanoj Sarcar
5 * Copyright (C) 2000, 2001 Ralf Baechle
6 * Copyright (C) 2000, 2001 Silicon Graphics, Inc.
7 * Copyright (C) 2000, 2001, 2003 Broadcom Corporation
9 #include <linux/cache.h>
10 #include <linux/delay.h>
11 #include <linux/init.h>
12 #include <linux/interrupt.h>
13 #include <linux/smp.h>
14 #include <linux/spinlock.h>
15 #include <linux/threads.h>
16 #include <linux/export.h>
17 #include <linux/time.h>
18 #include <linux/timex.h>
19 #include <linux/sched/mm.h>
20 #include <linux/cpumask.h>
21 #include <linux/cpu.h>
22 #include <linux/err.h>
23 #include <linux/ftrace.h>
24 #include <linux/irqdomain.h>
26 #include <linux/of_irq.h>
28 #include <linux/atomic.h>
30 #include <asm/ginvt.h>
31 #include <asm/processor.h>
33 #include <asm/r4k-timer.h>
34 #include <asm/mips-cps.h>
35 #include <asm/mmu_context.h>
37 #include <asm/setup.h>
40 int __cpu_number_map[CONFIG_MIPS_NR_CPU_NR_MAP]; /* Map physical to logical */
41 EXPORT_SYMBOL(__cpu_number_map);
43 int __cpu_logical_map[NR_CPUS]; /* Map logical to physical */
44 EXPORT_SYMBOL(__cpu_logical_map);
46 /* Number of TCs (or siblings in Intel speak) per CPU core */
47 int smp_num_siblings = 1;
48 EXPORT_SYMBOL(smp_num_siblings);
50 /* representing the TCs (or siblings in Intel speak) of each logical CPU */
51 cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly;
52 EXPORT_SYMBOL(cpu_sibling_map);
54 /* representing the core map of multi-core chips of each logical CPU */
55 cpumask_t cpu_core_map[NR_CPUS] __read_mostly;
56 EXPORT_SYMBOL(cpu_core_map);
58 static DECLARE_COMPLETION(cpu_starting);
59 static DECLARE_COMPLETION(cpu_running);
62 * A logical cpu mask containing only one VPE per core to
63 * reduce the number of IPIs on large MT systems.
65 cpumask_t cpu_foreign_map[NR_CPUS] __read_mostly;
66 EXPORT_SYMBOL(cpu_foreign_map);
68 /* representing cpus for which sibling maps can be computed */
69 static cpumask_t cpu_sibling_setup_map;
71 /* representing cpus for which core maps can be computed */
72 static cpumask_t cpu_core_setup_map;
74 cpumask_t cpu_coherent_mask;
76 #ifdef CONFIG_GENERIC_IRQ_IPI
77 static struct irq_desc *call_desc;
78 static struct irq_desc *sched_desc;
81 static inline void set_cpu_sibling_map(int cpu)
85 cpumask_set_cpu(cpu, &cpu_sibling_setup_map);
87 if (smp_num_siblings > 1) {
88 for_each_cpu(i, &cpu_sibling_setup_map) {
89 if (cpus_are_siblings(cpu, i)) {
90 cpumask_set_cpu(i, &cpu_sibling_map[cpu]);
91 cpumask_set_cpu(cpu, &cpu_sibling_map[i]);
95 cpumask_set_cpu(cpu, &cpu_sibling_map[cpu]);
98 static inline void set_cpu_core_map(int cpu)
102 cpumask_set_cpu(cpu, &cpu_core_setup_map);
104 for_each_cpu(i, &cpu_core_setup_map) {
105 if (cpu_data[cpu].package == cpu_data[i].package) {
106 cpumask_set_cpu(i, &cpu_core_map[cpu]);
107 cpumask_set_cpu(cpu, &cpu_core_map[i]);
113 * Calculate a new cpu_foreign_map mask whenever a
114 * new cpu appears or disappears.
116 void calculate_cpu_foreign_map(void)
118 int i, k, core_present;
119 cpumask_t temp_foreign_map;
121 /* Re-calculate the mask */
122 cpumask_clear(&temp_foreign_map);
123 for_each_online_cpu(i) {
125 for_each_cpu(k, &temp_foreign_map)
126 if (cpus_are_siblings(i, k))
129 cpumask_set_cpu(i, &temp_foreign_map);
132 for_each_online_cpu(i)
133 cpumask_andnot(&cpu_foreign_map[i],
134 &temp_foreign_map, &cpu_sibling_map[i]);
137 const struct plat_smp_ops *mp_ops;
138 EXPORT_SYMBOL(mp_ops);
140 void register_smp_ops(const struct plat_smp_ops *ops)
143 printk(KERN_WARNING "Overriding previously set SMP ops\n");
148 #ifdef CONFIG_GENERIC_IRQ_IPI
149 void mips_smp_send_ipi_single(int cpu, unsigned int action)
151 mips_smp_send_ipi_mask(cpumask_of(cpu), action);
154 void mips_smp_send_ipi_mask(const struct cpumask *mask, unsigned int action)
160 local_irq_save(flags);
163 case SMP_CALL_FUNCTION:
164 __ipi_send_mask(call_desc, mask);
167 case SMP_RESCHEDULE_YOURSELF:
168 __ipi_send_mask(sched_desc, mask);
175 if (mips_cpc_present()) {
176 for_each_cpu(cpu, mask) {
177 if (cpus_are_siblings(cpu, smp_processor_id()))
180 core = cpu_core(&cpu_data[cpu]);
182 while (!cpumask_test_cpu(cpu, &cpu_coherent_mask)) {
183 mips_cm_lock_other_cpu(cpu, CM_GCR_Cx_OTHER_BLOCK_LOCAL);
184 mips_cpc_lock_other(core);
185 write_cpc_co_cmd(CPC_Cx_CMD_PWRUP);
186 mips_cpc_unlock_other();
187 mips_cm_unlock_other();
192 local_irq_restore(flags);
196 static irqreturn_t ipi_resched_interrupt(int irq, void *dev_id)
203 static irqreturn_t ipi_call_interrupt(int irq, void *dev_id)
205 generic_smp_call_function_interrupt();
210 static void smp_ipi_init_one(unsigned int virq, const char *name,
211 irq_handler_t handler)
215 irq_set_handler(virq, handle_percpu_irq);
216 ret = request_irq(virq, handler, IRQF_PERCPU, name, NULL);
220 static unsigned int call_virq, sched_virq;
222 int mips_smp_ipi_allocate(const struct cpumask *mask)
225 struct irq_domain *ipidomain;
226 struct device_node *node;
228 node = of_irq_find_parent(of_root);
229 ipidomain = irq_find_matching_host(node, DOMAIN_BUS_IPI);
232 * Some platforms have half DT setup. So if we found irq node but
233 * didn't find an ipidomain, try to search for one that is not in the
236 if (node && !ipidomain)
237 ipidomain = irq_find_matching_host(NULL, DOMAIN_BUS_IPI);
240 * There are systems which use IPI IRQ domains, but only have one
241 * registered when some runtime condition is met. For example a Malta
242 * kernel may include support for GIC & CPU interrupt controller IPI
243 * IRQ domains, but if run on a system with no GIC & no MT ASE then
244 * neither will be supported or registered.
246 * We only have a problem if we're actually using multiple CPUs so fail
247 * loudly if that is the case. Otherwise simply return, skipping IPI
248 * setup, if we're running with only a single CPU.
251 BUG_ON(num_present_cpus() > 1);
255 virq = irq_reserve_ipi(ipidomain, mask);
260 virq = irq_reserve_ipi(ipidomain, mask);
265 if (irq_domain_is_ipi_per_cpu(ipidomain)) {
268 for_each_cpu(cpu, mask) {
269 smp_ipi_init_one(call_virq + cpu, "IPI call",
271 smp_ipi_init_one(sched_virq + cpu, "IPI resched",
272 ipi_resched_interrupt);
275 smp_ipi_init_one(call_virq, "IPI call", ipi_call_interrupt);
276 smp_ipi_init_one(sched_virq, "IPI resched",
277 ipi_resched_interrupt);
283 int mips_smp_ipi_free(const struct cpumask *mask)
285 struct irq_domain *ipidomain;
286 struct device_node *node;
288 node = of_irq_find_parent(of_root);
289 ipidomain = irq_find_matching_host(node, DOMAIN_BUS_IPI);
292 * Some platforms have half DT setup. So if we found irq node but
293 * didn't find an ipidomain, try to search for one that is not in the
296 if (node && !ipidomain)
297 ipidomain = irq_find_matching_host(NULL, DOMAIN_BUS_IPI);
301 if (irq_domain_is_ipi_per_cpu(ipidomain)) {
304 for_each_cpu(cpu, mask) {
305 free_irq(call_virq + cpu, NULL);
306 free_irq(sched_virq + cpu, NULL);
309 irq_destroy_ipi(call_virq, mask);
310 irq_destroy_ipi(sched_virq, mask);
315 static int __init mips_smp_ipi_init(void)
317 if (num_possible_cpus() == 1)
320 mips_smp_ipi_allocate(cpu_possible_mask);
322 call_desc = irq_to_desc(call_virq);
323 sched_desc = irq_to_desc(sched_virq);
327 early_initcall(mips_smp_ipi_init);
331 * First C code run on the secondary CPUs after being started up by
334 asmlinkage void start_secondary(void)
339 per_cpu_trap_init(false);
340 mips_clockevent_init();
341 mp_ops->init_secondary();
346 * XXX parity protection should be folded in here when it's converted
347 * to an option instead of something based on .cputype
351 cpu = smp_processor_id();
352 cpu_data[cpu].udelay_val = loops_per_jiffy;
354 set_cpu_sibling_map(cpu);
355 set_cpu_core_map(cpu);
357 cpumask_set_cpu(cpu, &cpu_coherent_mask);
358 notify_cpu_starting(cpu);
360 /* Notify boot CPU that we're starting & ready to sync counters */
361 complete(&cpu_starting);
363 synchronise_count_slave(cpu);
365 /* The CPU is running and counters synchronised, now mark it online */
366 set_cpu_online(cpu, true);
368 calculate_cpu_foreign_map();
371 * Notify boot CPU that we're up & online and it can safely return
374 complete(&cpu_running);
377 * irq will be enabled in ->smp_finish(), enabling it too early
380 WARN_ON_ONCE(!irqs_disabled());
381 mp_ops->smp_finish();
383 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
386 static void stop_this_cpu(void *dummy)
392 set_cpu_online(smp_processor_id(), false);
393 calculate_cpu_foreign_map();
398 void smp_send_stop(void)
400 smp_call_function(stop_this_cpu, NULL, 0);
403 void __init smp_cpus_done(unsigned int max_cpus)
407 /* called from main before smp_init() */
408 void __init smp_prepare_cpus(unsigned int max_cpus)
410 init_new_context(current, &init_mm);
411 current_thread_info()->cpu = 0;
412 mp_ops->prepare_cpus(max_cpus);
413 set_cpu_sibling_map(0);
415 calculate_cpu_foreign_map();
416 #ifndef CONFIG_HOTPLUG_CPU
417 init_cpu_present(cpu_possible_mask);
419 cpumask_copy(&cpu_coherent_mask, cpu_possible_mask);
422 /* preload SMP state for boot cpu */
423 void smp_prepare_boot_cpu(void)
425 if (mp_ops->prepare_boot_cpu)
426 mp_ops->prepare_boot_cpu();
427 set_cpu_possible(0, true);
428 set_cpu_online(0, true);
431 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
435 err = mp_ops->boot_secondary(cpu, tidle);
439 /* Wait for CPU to start and be ready to sync counters */
440 if (!wait_for_completion_timeout(&cpu_starting,
441 msecs_to_jiffies(1000))) {
442 pr_crit("CPU%u: failed to start\n", cpu);
446 synchronise_count_master(cpu);
448 /* Wait for CPU to finish startup & mark itself online before return */
449 wait_for_completion(&cpu_running);
453 /* Not really SMP stuff ... */
454 int setup_profiling_timer(unsigned int multiplier)
459 static void flush_tlb_all_ipi(void *info)
461 local_flush_tlb_all();
464 void flush_tlb_all(void)
470 instruction_hazard();
475 on_each_cpu(flush_tlb_all_ipi, NULL, 1);
478 static void flush_tlb_mm_ipi(void *mm)
480 drop_mmu_context((struct mm_struct *)mm);
484 * Special Variant of smp_call_function for use by TLB functions:
487 * o collapses to normal function call on UP kernels
488 * o collapses to normal function call on systems with a single shared
491 static inline void smp_on_other_tlbs(void (*func) (void *info), void *info)
493 smp_call_function(func, info, 1);
496 static inline void smp_on_each_tlb(void (*func) (void *info), void *info)
500 smp_on_other_tlbs(func, info);
507 * The following tlb flush calls are invoked when old translations are
508 * being torn down, or pte attributes are changing. For single threaded
509 * address spaces, a new context is obtained on the current cpu, and tlb
510 * context on other cpus are invalidated to force a new context allocation
511 * at switch_mm time, should the mm ever be used on other cpus. For
512 * multithreaded address spaces, inter-CPU interrupts have to be sent.
513 * Another case where inter-CPU interrupts are required is when the target
514 * mm might be active on another cpu (eg debuggers doing the flushes on
515 * behalf of debugees, kswapd stealing pages from another process etc).
519 void flush_tlb_mm(struct mm_struct *mm)
525 * No need to worry about other CPUs - the ginvt in
526 * drop_mmu_context() will be globalized.
528 } else if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
529 smp_on_other_tlbs(flush_tlb_mm_ipi, mm);
533 for_each_online_cpu(cpu) {
534 if (cpu != smp_processor_id() && cpu_context(cpu, mm))
535 set_cpu_context(cpu, mm, 0);
538 drop_mmu_context(mm);
543 struct flush_tlb_data {
544 struct vm_area_struct *vma;
549 static void flush_tlb_range_ipi(void *info)
551 struct flush_tlb_data *fd = info;
553 local_flush_tlb_range(fd->vma, fd->addr1, fd->addr2);
556 void flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
558 struct mm_struct *mm = vma->vm_mm;
565 old_mmid = read_c0_memorymapid();
566 write_c0_memorymapid(cpu_asid(0, mm));
568 addr = round_down(start, PAGE_SIZE * 2);
569 end = round_up(end, PAGE_SIZE * 2);
573 addr += PAGE_SIZE * 2;
574 } while (addr < end);
575 write_c0_memorymapid(old_mmid);
576 instruction_hazard();
578 } else if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) {
579 struct flush_tlb_data fd = {
585 smp_on_other_tlbs(flush_tlb_range_ipi, &fd);
586 local_flush_tlb_range(vma, start, end);
589 int exec = vma->vm_flags & VM_EXEC;
591 for_each_online_cpu(cpu) {
593 * flush_cache_range() will only fully flush icache if
594 * the VMA is executable, otherwise we must invalidate
595 * ASID without it appearing to has_valid_asid() as if
596 * mm has been completely unused by that CPU.
598 if (cpu != smp_processor_id() && cpu_context(cpu, mm))
599 set_cpu_context(cpu, mm, !exec);
601 local_flush_tlb_range(vma, start, end);
606 static void flush_tlb_kernel_range_ipi(void *info)
608 struct flush_tlb_data *fd = info;
610 local_flush_tlb_kernel_range(fd->addr1, fd->addr2);
613 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
615 struct flush_tlb_data fd = {
620 on_each_cpu(flush_tlb_kernel_range_ipi, &fd, 1);
623 static void flush_tlb_page_ipi(void *info)
625 struct flush_tlb_data *fd = info;
627 local_flush_tlb_page(fd->vma, fd->addr1);
630 void flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
637 old_mmid = read_c0_memorymapid();
638 write_c0_memorymapid(cpu_asid(0, vma->vm_mm));
642 write_c0_memorymapid(old_mmid);
643 instruction_hazard();
645 } else if ((atomic_read(&vma->vm_mm->mm_users) != 1) ||
646 (current->mm != vma->vm_mm)) {
647 struct flush_tlb_data fd = {
652 smp_on_other_tlbs(flush_tlb_page_ipi, &fd);
653 local_flush_tlb_page(vma, page);
657 for_each_online_cpu(cpu) {
659 * flush_cache_page() only does partial flushes, so
660 * invalidate ASID without it appearing to
661 * has_valid_asid() as if mm has been completely unused
664 if (cpu != smp_processor_id() && cpu_context(cpu, vma->vm_mm))
665 set_cpu_context(cpu, vma->vm_mm, 1);
667 local_flush_tlb_page(vma, page);
672 static void flush_tlb_one_ipi(void *info)
674 unsigned long vaddr = (unsigned long) info;
676 local_flush_tlb_one(vaddr);
679 void flush_tlb_one(unsigned long vaddr)
681 smp_on_each_tlb(flush_tlb_one_ipi, (void *) vaddr);
684 EXPORT_SYMBOL(flush_tlb_page);
685 EXPORT_SYMBOL(flush_tlb_one);
687 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
689 static void tick_broadcast_callee(void *info)
691 tick_receive_broadcast();
694 static DEFINE_PER_CPU(call_single_data_t, tick_broadcast_csd) =
695 CSD_INIT(tick_broadcast_callee, NULL);
697 void tick_broadcast(const struct cpumask *mask)
699 call_single_data_t *csd;
702 for_each_cpu(cpu, mask) {
703 csd = &per_cpu(tick_broadcast_csd, cpu);
704 smp_call_function_single_async(cpu, csd);
708 #endif /* CONFIG_GENERIC_CLOCKEVENTS_BROADCAST */