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