1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 *
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
8 */
9#include <linux/cache.h>
10#include <linux/delay.h>
11#include <linux/init.h>
12#include <linux/interrupt.h>
13#include <linux/profile.h>
14#include <linux/smp.h>
15#include <linux/spinlock.h>
16#include <linux/threads.h>
17#include <linux/export.h>
18#include <linux/time.h>
19#include <linux/timex.h>
20#include <linux/sched/mm.h>
21#include <linux/cpumask.h>
22#include <linux/cpu.h>
23#include <linux/err.h>
24#include <linux/ftrace.h>
25#include <linux/irqdomain.h>
26#include <linux/of.h>
27#include <linux/of_irq.h>
28
29#include <linux/atomic.h>
30#include <asm/cpu.h>
31#include <asm/ginvt.h>
32#include <asm/processor.h>
33#include <asm/idle.h>
34#include <asm/r4k-timer.h>
35#include <asm/mips-cps.h>
36#include <asm/mmu_context.h>
37#include <asm/time.h>
38#include <asm/setup.h>
39#include <asm/maar.h>
40
41int __cpu_number_map[CONFIG_MIPS_NR_CPU_NR_MAP]; /* Map physical to logical */
42EXPORT_SYMBOL(__cpu_number_map);
43
44int __cpu_logical_map[NR_CPUS]; /* Map logical to physical */
45EXPORT_SYMBOL(__cpu_logical_map);
46
47/* Number of TCs (or siblings in Intel speak) per CPU core */
48int smp_num_siblings = 1;
49EXPORT_SYMBOL(smp_num_siblings);
50
51/* representing the TCs (or siblings in Intel speak) of each logical CPU */
52cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly;
53EXPORT_SYMBOL(cpu_sibling_map);
54
55/* representing the core map of multi-core chips of each logical CPU */
56cpumask_t cpu_core_map[NR_CPUS] __read_mostly;
57EXPORT_SYMBOL(cpu_core_map);
58
59static DECLARE_COMPLETION(cpu_starting);
60static DECLARE_COMPLETION(cpu_running);
61
62/*
63 * A logical cpu mask containing only one VPE per core to
64 * reduce the number of IPIs on large MT systems.
65 */
66cpumask_t cpu_foreign_map[NR_CPUS] __read_mostly;
67EXPORT_SYMBOL(cpu_foreign_map);
68
69/* representing cpus for which sibling maps can be computed */
70static cpumask_t cpu_sibling_setup_map;
71
72/* representing cpus for which core maps can be computed */
73static cpumask_t cpu_core_setup_map;
74
75cpumask_t cpu_coherent_mask;
76
77unsigned int smp_max_threads __initdata = UINT_MAX;
78
79static int __init early_nosmt(char *s)
80{
81 smp_max_threads = 1;
82 return 0;
83}
84early_param("nosmt", early_nosmt);
85
86static int __init early_smt(char *s)
87{
88 get_option(str: &s, pint: &smp_max_threads);
89 /* Ensure at least one thread is available */
90 smp_max_threads = clamp_val(smp_max_threads, 1U, UINT_MAX);
91 return 0;
92}
93early_param("smt", early_smt);
94
95#ifdef CONFIG_GENERIC_IRQ_IPI
96static struct irq_desc *call_desc;
97static struct irq_desc *sched_desc;
98#endif
99
100static inline void set_cpu_sibling_map(int cpu)
101{
102 int i;
103
104 cpumask_set_cpu(cpu, dstp: &cpu_sibling_setup_map);
105
106 if (smp_num_siblings > 1) {
107 for_each_cpu(i, &cpu_sibling_setup_map) {
108 if (cpus_are_siblings(cpu, i)) {
109 cpumask_set_cpu(cpu: i, dstp: &cpu_sibling_map[cpu]);
110 cpumask_set_cpu(cpu, dstp: &cpu_sibling_map[i]);
111 }
112 }
113 } else
114 cpumask_set_cpu(cpu, dstp: &cpu_sibling_map[cpu]);
115}
116
117static inline void set_cpu_core_map(int cpu)
118{
119 int i;
120
121 cpumask_set_cpu(cpu, dstp: &cpu_core_setup_map);
122
123 for_each_cpu(i, &cpu_core_setup_map) {
124 if (cpu_data[cpu].package == cpu_data[i].package) {
125 cpumask_set_cpu(cpu: i, dstp: &cpu_core_map[cpu]);
126 cpumask_set_cpu(cpu, dstp: &cpu_core_map[i]);
127 }
128 }
129}
130
131/*
132 * Calculate a new cpu_foreign_map mask whenever a
133 * new cpu appears or disappears.
134 */
135void calculate_cpu_foreign_map(void)
136{
137 int i, k, core_present;
138 cpumask_t temp_foreign_map;
139
140 /* Re-calculate the mask */
141 cpumask_clear(dstp: &temp_foreign_map);
142 for_each_online_cpu(i) {
143 core_present = 0;
144 for_each_cpu(k, &temp_foreign_map)
145 if (cpus_are_siblings(i, k))
146 core_present = 1;
147 if (!core_present)
148 cpumask_set_cpu(cpu: i, dstp: &temp_foreign_map);
149 }
150
151 for_each_online_cpu(i)
152 cpumask_andnot(dstp: &cpu_foreign_map[i],
153 src1p: &temp_foreign_map, src2p: &cpu_sibling_map[i]);
154}
155
156const struct plat_smp_ops *mp_ops;
157EXPORT_SYMBOL(mp_ops);
158
159void register_smp_ops(const struct plat_smp_ops *ops)
160{
161 if (mp_ops)
162 printk(KERN_WARNING "Overriding previously set SMP ops\n");
163
164 mp_ops = ops;
165}
166
167#ifdef CONFIG_GENERIC_IRQ_IPI
168void mips_smp_send_ipi_single(int cpu, unsigned int action)
169{
170 mips_smp_send_ipi_mask(cpumask_of(cpu), action);
171}
172
173void mips_smp_send_ipi_mask(const struct cpumask *mask, unsigned int action)
174{
175 unsigned long flags;
176 unsigned int core;
177 int cpu;
178
179 local_irq_save(flags);
180
181 switch (action) {
182 case SMP_CALL_FUNCTION:
183 __ipi_send_mask(call_desc, mask);
184 break;
185
186 case SMP_RESCHEDULE_YOURSELF:
187 __ipi_send_mask(sched_desc, mask);
188 break;
189
190 default:
191 BUG();
192 }
193
194 if (mips_cpc_present()) {
195 for_each_cpu(cpu, mask) {
196 if (cpus_are_siblings(cpu, smp_processor_id()))
197 continue;
198
199 core = cpu_core(&cpu_data[cpu]);
200
201 while (!cpumask_test_cpu(cpu, &cpu_coherent_mask)) {
202 mips_cm_lock_other_cpu(cpu, CM_GCR_Cx_OTHER_BLOCK_LOCAL);
203 mips_cpc_lock_other(core);
204 write_cpc_co_cmd(CPC_Cx_CMD_PWRUP);
205 mips_cpc_unlock_other();
206 mips_cm_unlock_other();
207 }
208 }
209 }
210
211 local_irq_restore(flags);
212}
213
214
215static irqreturn_t ipi_resched_interrupt(int irq, void *dev_id)
216{
217 scheduler_ipi();
218
219 return IRQ_HANDLED;
220}
221
222static irqreturn_t ipi_call_interrupt(int irq, void *dev_id)
223{
224 generic_smp_call_function_interrupt();
225
226 return IRQ_HANDLED;
227}
228
229static void smp_ipi_init_one(unsigned int virq, const char *name,
230 irq_handler_t handler)
231{
232 int ret;
233
234 irq_set_handler(virq, handle_percpu_irq);
235 ret = request_irq(virq, handler, IRQF_PERCPU, name, NULL);
236 BUG_ON(ret);
237}
238
239static unsigned int call_virq, sched_virq;
240
241int mips_smp_ipi_allocate(const struct cpumask *mask)
242{
243 int virq;
244 struct irq_domain *ipidomain;
245 struct device_node *node;
246
247 node = of_irq_find_parent(of_root);
248 ipidomain = irq_find_matching_host(node, DOMAIN_BUS_IPI);
249
250 /*
251 * Some platforms have half DT setup. So if we found irq node but
252 * didn't find an ipidomain, try to search for one that is not in the
253 * DT.
254 */
255 if (node && !ipidomain)
256 ipidomain = irq_find_matching_host(NULL, DOMAIN_BUS_IPI);
257
258 /*
259 * There are systems which use IPI IRQ domains, but only have one
260 * registered when some runtime condition is met. For example a Malta
261 * kernel may include support for GIC & CPU interrupt controller IPI
262 * IRQ domains, but if run on a system with no GIC & no MT ASE then
263 * neither will be supported or registered.
264 *
265 * We only have a problem if we're actually using multiple CPUs so fail
266 * loudly if that is the case. Otherwise simply return, skipping IPI
267 * setup, if we're running with only a single CPU.
268 */
269 if (!ipidomain) {
270 BUG_ON(num_present_cpus() > 1);
271 return 0;
272 }
273
274 virq = irq_reserve_ipi(ipidomain, mask);
275 BUG_ON(!virq);
276 if (!call_virq)
277 call_virq = virq;
278
279 virq = irq_reserve_ipi(ipidomain, mask);
280 BUG_ON(!virq);
281 if (!sched_virq)
282 sched_virq = virq;
283
284 if (irq_domain_is_ipi_per_cpu(ipidomain)) {
285 int cpu;
286
287 for_each_cpu(cpu, mask) {
288 smp_ipi_init_one(call_virq + cpu, "IPI call",
289 ipi_call_interrupt);
290 smp_ipi_init_one(sched_virq + cpu, "IPI resched",
291 ipi_resched_interrupt);
292 }
293 } else {
294 smp_ipi_init_one(call_virq, "IPI call", ipi_call_interrupt);
295 smp_ipi_init_one(sched_virq, "IPI resched",
296 ipi_resched_interrupt);
297 }
298
299 return 0;
300}
301
302int mips_smp_ipi_free(const struct cpumask *mask)
303{
304 struct irq_domain *ipidomain;
305 struct device_node *node;
306
307 node = of_irq_find_parent(of_root);
308 ipidomain = irq_find_matching_host(node, DOMAIN_BUS_IPI);
309
310 /*
311 * Some platforms have half DT setup. So if we found irq node but
312 * didn't find an ipidomain, try to search for one that is not in the
313 * DT.
314 */
315 if (node && !ipidomain)
316 ipidomain = irq_find_matching_host(NULL, DOMAIN_BUS_IPI);
317
318 BUG_ON(!ipidomain);
319
320 if (irq_domain_is_ipi_per_cpu(ipidomain)) {
321 int cpu;
322
323 for_each_cpu(cpu, mask) {
324 free_irq(call_virq + cpu, NULL);
325 free_irq(sched_virq + cpu, NULL);
326 }
327 }
328 irq_destroy_ipi(call_virq, mask);
329 irq_destroy_ipi(sched_virq, mask);
330 return 0;
331}
332
333
334static int __init mips_smp_ipi_init(void)
335{
336 if (num_possible_cpus() == 1)
337 return 0;
338
339 mips_smp_ipi_allocate(cpu_possible_mask);
340
341 call_desc = irq_to_desc(call_virq);
342 sched_desc = irq_to_desc(sched_virq);
343
344 return 0;
345}
346early_initcall(mips_smp_ipi_init);
347#endif
348
349/*
350 * First C code run on the secondary CPUs after being started up by
351 * the master.
352 */
353asmlinkage void start_secondary(void)
354{
355 unsigned int cpu = raw_smp_processor_id();
356
357 cpu_probe();
358 per_cpu_trap_init(false);
359 rcutree_report_cpu_starting(cpu);
360 mips_clockevent_init();
361 mp_ops->init_secondary();
362 cpu_report();
363 maar_init();
364
365 /*
366 * XXX parity protection should be folded in here when it's converted
367 * to an option instead of something based on .cputype
368 */
369
370 calibrate_delay();
371 cpu_data[cpu].udelay_val = loops_per_jiffy;
372
373 set_cpu_sibling_map(cpu);
374 set_cpu_core_map(cpu);
375
376 cpumask_set_cpu(cpu, dstp: &cpu_coherent_mask);
377 notify_cpu_starting(cpu);
378
379 /* Notify boot CPU that we're starting & ready to sync counters */
380 complete(&cpu_starting);
381
382 synchronise_count_slave(cpu);
383
384 /* The CPU is running and counters synchronised, now mark it online */
385 set_cpu_online(cpu, online: true);
386
387 calculate_cpu_foreign_map();
388
389 /*
390 * Notify boot CPU that we're up & online and it can safely return
391 * from __cpu_up
392 */
393 complete(&cpu_running);
394
395 /*
396 * irq will be enabled in ->smp_finish(), enabling it too early
397 * is dangerous.
398 */
399 WARN_ON_ONCE(!irqs_disabled());
400 mp_ops->smp_finish();
401
402 cpu_startup_entry(state: CPUHP_AP_ONLINE_IDLE);
403}
404
405static void stop_this_cpu(void *dummy)
406{
407 /*
408 * Remove this CPU:
409 */
410
411 set_cpu_online(smp_processor_id(), online: false);
412 calculate_cpu_foreign_map();
413 local_irq_disable();
414 while (1);
415}
416
417void smp_send_stop(void)
418{
419 smp_call_function(func: stop_this_cpu, NULL, wait: 0);
420}
421
422void __init smp_cpus_done(unsigned int max_cpus)
423{
424}
425
426/* called from main before smp_init() */
427void __init smp_prepare_cpus(unsigned int max_cpus)
428{
429 init_new_context(current, mm: &init_mm);
430 current_thread_info()->cpu = 0;
431 mp_ops->prepare_cpus(max_cpus);
432 set_cpu_sibling_map(0);
433 set_cpu_core_map(0);
434 calculate_cpu_foreign_map();
435#ifndef CONFIG_HOTPLUG_CPU
436 init_cpu_present(cpu_possible_mask);
437#endif
438 cpumask_copy(dstp: &cpu_coherent_mask, cpu_possible_mask);
439}
440
441/* preload SMP state for boot cpu */
442void smp_prepare_boot_cpu(void)
443{
444 if (mp_ops->prepare_boot_cpu)
445 mp_ops->prepare_boot_cpu();
446 set_cpu_possible(cpu: 0, possible: true);
447 set_cpu_online(cpu: 0, online: true);
448}
449
450int __cpu_up(unsigned int cpu, struct task_struct *tidle)
451{
452 int err;
453
454 err = mp_ops->boot_secondary(cpu, tidle);
455 if (err)
456 return err;
457
458 /* Wait for CPU to start and be ready to sync counters */
459 if (!wait_for_completion_timeout(x: &cpu_starting,
460 timeout: msecs_to_jiffies(m: 1000))) {
461 pr_crit("CPU%u: failed to start\n", cpu);
462 return -EIO;
463 }
464
465 synchronise_count_master(cpu);
466
467 /* Wait for CPU to finish startup & mark itself online before return */
468 wait_for_completion(&cpu_running);
469 return 0;
470}
471
472#ifdef CONFIG_PROFILING
473/* Not really SMP stuff ... */
474int setup_profiling_timer(unsigned int multiplier)
475{
476 return 0;
477}
478#endif
479
480static void flush_tlb_all_ipi(void *info)
481{
482 local_flush_tlb_all();
483}
484
485void flush_tlb_all(void)
486{
487 if (cpu_has_mmid) {
488 htw_stop();
489 ginvt_full();
490 sync_ginv();
491 instruction_hazard();
492 htw_start();
493 return;
494 }
495
496 on_each_cpu(func: flush_tlb_all_ipi, NULL, wait: 1);
497}
498
499static void flush_tlb_mm_ipi(void *mm)
500{
501 drop_mmu_context((struct mm_struct *)mm);
502}
503
504/*
505 * Special Variant of smp_call_function for use by TLB functions:
506 *
507 * o No return value
508 * o collapses to normal function call on UP kernels
509 * o collapses to normal function call on systems with a single shared
510 * primary cache.
511 */
512static inline void smp_on_other_tlbs(void (*func) (void *info), void *info)
513{
514 smp_call_function(func, info, wait: 1);
515}
516
517static inline void smp_on_each_tlb(void (*func) (void *info), void *info)
518{
519 preempt_disable();
520
521 smp_on_other_tlbs(func, info);
522 func(info);
523
524 preempt_enable();
525}
526
527/*
528 * The following tlb flush calls are invoked when old translations are
529 * being torn down, or pte attributes are changing. For single threaded
530 * address spaces, a new context is obtained on the current cpu, and tlb
531 * context on other cpus are invalidated to force a new context allocation
532 * at switch_mm time, should the mm ever be used on other cpus. For
533 * multithreaded address spaces, inter-CPU interrupts have to be sent.
534 * Another case where inter-CPU interrupts are required is when the target
535 * mm might be active on another cpu (eg debuggers doing the flushes on
536 * behalf of debugees, kswapd stealing pages from another process etc).
537 * Kanoj 07/00.
538 */
539
540void flush_tlb_mm(struct mm_struct *mm)
541{
542 if (!mm)
543 return;
544
545 if (atomic_read(v: &mm->mm_users) == 0)
546 return; /* happens as a result of exit_mmap() */
547
548 preempt_disable();
549
550 if (cpu_has_mmid) {
551 /*
552 * No need to worry about other CPUs - the ginvt in
553 * drop_mmu_context() will be globalized.
554 */
555 } else if ((atomic_read(v: &mm->mm_users) != 1) || (current->mm != mm)) {
556 smp_on_other_tlbs(func: flush_tlb_mm_ipi, info: mm);
557 } else {
558 unsigned int cpu;
559
560 for_each_online_cpu(cpu) {
561 if (cpu != smp_processor_id() && cpu_context(cpu, mm))
562 set_cpu_context(cpu, mm, 0);
563 }
564 }
565 drop_mmu_context(mm);
566
567 preempt_enable();
568}
569
570struct flush_tlb_data {
571 struct vm_area_struct *vma;
572 unsigned long addr1;
573 unsigned long addr2;
574};
575
576static void flush_tlb_range_ipi(void *info)
577{
578 struct flush_tlb_data *fd = info;
579
580 local_flush_tlb_range(fd->vma, fd->addr1, fd->addr2);
581}
582
583void flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
584{
585 struct mm_struct *mm = vma->vm_mm;
586 unsigned long addr;
587 u32 old_mmid;
588
589 preempt_disable();
590 if (cpu_has_mmid) {
591 htw_stop();
592 old_mmid = read_c0_memorymapid();
593 write_c0_memorymapid(cpu_asid(0, mm));
594 mtc0_tlbw_hazard();
595 addr = round_down(start, PAGE_SIZE * 2);
596 end = round_up(end, PAGE_SIZE * 2);
597 do {
598 ginvt_va_mmid(addr);
599 sync_ginv();
600 addr += PAGE_SIZE * 2;
601 } while (addr < end);
602 write_c0_memorymapid(old_mmid);
603 instruction_hazard();
604 htw_start();
605 } else if ((atomic_read(v: &mm->mm_users) != 1) || (current->mm != mm)) {
606 struct flush_tlb_data fd = {
607 .vma = vma,
608 .addr1 = start,
609 .addr2 = end,
610 };
611
612 smp_on_other_tlbs(func: flush_tlb_range_ipi, info: &fd);
613 local_flush_tlb_range(vma, start, end);
614 } else {
615 unsigned int cpu;
616 int exec = vma->vm_flags & VM_EXEC;
617
618 for_each_online_cpu(cpu) {
619 /*
620 * flush_cache_range() will only fully flush icache if
621 * the VMA is executable, otherwise we must invalidate
622 * ASID without it appearing to has_valid_asid() as if
623 * mm has been completely unused by that CPU.
624 */
625 if (cpu != smp_processor_id() && cpu_context(cpu, mm))
626 set_cpu_context(cpu, mm, !exec);
627 }
628 local_flush_tlb_range(vma, start, end);
629 }
630 preempt_enable();
631}
632
633static void flush_tlb_kernel_range_ipi(void *info)
634{
635 struct flush_tlb_data *fd = info;
636
637 local_flush_tlb_kernel_range(fd->addr1, fd->addr2);
638}
639
640void flush_tlb_kernel_range(unsigned long start, unsigned long end)
641{
642 struct flush_tlb_data fd = {
643 .addr1 = start,
644 .addr2 = end,
645 };
646
647 on_each_cpu(func: flush_tlb_kernel_range_ipi, info: &fd, wait: 1);
648}
649
650static void flush_tlb_page_ipi(void *info)
651{
652 struct flush_tlb_data *fd = info;
653
654 local_flush_tlb_page(fd->vma, fd->addr1);
655}
656
657void flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
658{
659 u32 old_mmid;
660
661 preempt_disable();
662 if (cpu_has_mmid) {
663 htw_stop();
664 old_mmid = read_c0_memorymapid();
665 write_c0_memorymapid(cpu_asid(0, vma->vm_mm));
666 mtc0_tlbw_hazard();
667 ginvt_va_mmid(page);
668 sync_ginv();
669 write_c0_memorymapid(old_mmid);
670 instruction_hazard();
671 htw_start();
672 } else if ((atomic_read(v: &vma->vm_mm->mm_users) != 1) ||
673 (current->mm != vma->vm_mm)) {
674 struct flush_tlb_data fd = {
675 .vma = vma,
676 .addr1 = page,
677 };
678
679 smp_on_other_tlbs(func: flush_tlb_page_ipi, info: &fd);
680 local_flush_tlb_page(vma, page);
681 } else {
682 unsigned int cpu;
683
684 for_each_online_cpu(cpu) {
685 /*
686 * flush_cache_page() only does partial flushes, so
687 * invalidate ASID without it appearing to
688 * has_valid_asid() as if mm has been completely unused
689 * by that CPU.
690 */
691 if (cpu != smp_processor_id() && cpu_context(cpu, vma->vm_mm))
692 set_cpu_context(cpu, vma->vm_mm, 1);
693 }
694 local_flush_tlb_page(vma, page);
695 }
696 preempt_enable();
697}
698
699static void flush_tlb_one_ipi(void *info)
700{
701 unsigned long vaddr = (unsigned long) info;
702
703 local_flush_tlb_one(vaddr);
704}
705
706void flush_tlb_one(unsigned long vaddr)
707{
708 smp_on_each_tlb(func: flush_tlb_one_ipi, info: (void *) vaddr);
709}
710
711EXPORT_SYMBOL(flush_tlb_page);
712EXPORT_SYMBOL(flush_tlb_one);
713
714#ifdef CONFIG_HOTPLUG_CORE_SYNC_DEAD
715void arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)
716{
717 if (mp_ops->cleanup_dead_cpu)
718 mp_ops->cleanup_dead_cpu(cpu);
719}
720#endif
721
722#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
723
724static void tick_broadcast_callee(void *info)
725{
726 tick_receive_broadcast();
727}
728
729static DEFINE_PER_CPU(call_single_data_t, tick_broadcast_csd) =
730 CSD_INIT(tick_broadcast_callee, NULL);
731
732void tick_broadcast(const struct cpumask *mask)
733{
734 call_single_data_t *csd;
735 int cpu;
736
737 for_each_cpu(cpu, mask) {
738 csd = &per_cpu(tick_broadcast_csd, cpu);
739 smp_call_function_single_async(cpu, csd);
740 }
741}
742
743#endif /* CONFIG_GENERIC_CLOCKEVENTS_BROADCAST */
744

source code of linux/arch/mips/kernel/smp.c