Loading...
1/*
2 * Local APIC related interfaces to support IOAPIC, MSI, etc.
3 *
4 * Copyright (C) 1997, 1998, 1999, 2000, 2009 Ingo Molnar, Hajnalka Szabo
5 * Moved from arch/x86/kernel/apic/io_apic.c.
6 * Jiang Liu <jiang.liu@linux.intel.com>
7 * Enable support of hierarchical irqdomains
8 *
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License version 2 as
11 * published by the Free Software Foundation.
12 */
13#include <linux/interrupt.h>
14#include <linux/seq_file.h>
15#include <linux/init.h>
16#include <linux/compiler.h>
17#include <linux/slab.h>
18#include <asm/irqdomain.h>
19#include <asm/hw_irq.h>
20#include <asm/apic.h>
21#include <asm/i8259.h>
22#include <asm/desc.h>
23#include <asm/irq_remapping.h>
24
25#include <asm/trace/irq_vectors.h>
26
27struct apic_chip_data {
28 struct irq_cfg hw_irq_cfg;
29 unsigned int vector;
30 unsigned int prev_vector;
31 unsigned int cpu;
32 unsigned int prev_cpu;
33 unsigned int irq;
34 struct hlist_node clist;
35 unsigned int move_in_progress : 1,
36 is_managed : 1,
37 can_reserve : 1,
38 has_reserved : 1;
39};
40
41struct irq_domain *x86_vector_domain;
42EXPORT_SYMBOL_GPL(x86_vector_domain);
43static DEFINE_RAW_SPINLOCK(vector_lock);
44static cpumask_var_t vector_searchmask;
45static struct irq_chip lapic_controller;
46static struct irq_matrix *vector_matrix;
47#ifdef CONFIG_SMP
48static DEFINE_PER_CPU(struct hlist_head, cleanup_list);
49#endif
50
51void lock_vector_lock(void)
52{
53 /* Used to the online set of cpus does not change
54 * during assign_irq_vector.
55 */
56 raw_spin_lock(&vector_lock);
57}
58
59void unlock_vector_lock(void)
60{
61 raw_spin_unlock(&vector_lock);
62}
63
64void init_irq_alloc_info(struct irq_alloc_info *info,
65 const struct cpumask *mask)
66{
67 memset(info, 0, sizeof(*info));
68 info->mask = mask;
69}
70
71void copy_irq_alloc_info(struct irq_alloc_info *dst, struct irq_alloc_info *src)
72{
73 if (src)
74 *dst = *src;
75 else
76 memset(dst, 0, sizeof(*dst));
77}
78
79static struct apic_chip_data *apic_chip_data(struct irq_data *irqd)
80{
81 if (!irqd)
82 return NULL;
83
84 while (irqd->parent_data)
85 irqd = irqd->parent_data;
86
87 return irqd->chip_data;
88}
89
90struct irq_cfg *irqd_cfg(struct irq_data *irqd)
91{
92 struct apic_chip_data *apicd = apic_chip_data(irqd);
93
94 return apicd ? &apicd->hw_irq_cfg : NULL;
95}
96EXPORT_SYMBOL_GPL(irqd_cfg);
97
98struct irq_cfg *irq_cfg(unsigned int irq)
99{
100 return irqd_cfg(irq_get_irq_data(irq));
101}
102
103static struct apic_chip_data *alloc_apic_chip_data(int node)
104{
105 struct apic_chip_data *apicd;
106
107 apicd = kzalloc_node(sizeof(*apicd), GFP_KERNEL, node);
108 if (apicd)
109 INIT_HLIST_NODE(&apicd->clist);
110 return apicd;
111}
112
113static void free_apic_chip_data(struct apic_chip_data *apicd)
114{
115 kfree(apicd);
116}
117
118static void apic_update_irq_cfg(struct irq_data *irqd, unsigned int vector,
119 unsigned int cpu)
120{
121 struct apic_chip_data *apicd = apic_chip_data(irqd);
122
123 lockdep_assert_held(&vector_lock);
124
125 apicd->hw_irq_cfg.vector = vector;
126 apicd->hw_irq_cfg.dest_apicid = apic->calc_dest_apicid(cpu);
127 irq_data_update_effective_affinity(irqd, cpumask_of(cpu));
128 trace_vector_config(irqd->irq, vector, cpu,
129 apicd->hw_irq_cfg.dest_apicid);
130}
131
132static void apic_update_vector(struct irq_data *irqd, unsigned int newvec,
133 unsigned int newcpu)
134{
135 struct apic_chip_data *apicd = apic_chip_data(irqd);
136 struct irq_desc *desc = irq_data_to_desc(irqd);
137 bool managed = irqd_affinity_is_managed(irqd);
138
139 lockdep_assert_held(&vector_lock);
140
141 trace_vector_update(irqd->irq, newvec, newcpu, apicd->vector,
142 apicd->cpu);
143
144 /*
145 * If there is no vector associated or if the associated vector is
146 * the shutdown vector, which is associated to make PCI/MSI
147 * shutdown mode work, then there is nothing to release. Clear out
148 * prev_vector for this and the offlined target case.
149 */
150 apicd->prev_vector = 0;
151 if (!apicd->vector || apicd->vector == MANAGED_IRQ_SHUTDOWN_VECTOR)
152 goto setnew;
153 /*
154 * If the target CPU of the previous vector is online, then mark
155 * the vector as move in progress and store it for cleanup when the
156 * first interrupt on the new vector arrives. If the target CPU is
157 * offline then the regular release mechanism via the cleanup
158 * vector is not possible and the vector can be immediately freed
159 * in the underlying matrix allocator.
160 */
161 if (cpu_online(apicd->cpu)) {
162 apicd->move_in_progress = true;
163 apicd->prev_vector = apicd->vector;
164 apicd->prev_cpu = apicd->cpu;
165 } else {
166 irq_matrix_free(vector_matrix, apicd->cpu, apicd->vector,
167 managed);
168 }
169
170setnew:
171 apicd->vector = newvec;
172 apicd->cpu = newcpu;
173 BUG_ON(!IS_ERR_OR_NULL(per_cpu(vector_irq, newcpu)[newvec]));
174 per_cpu(vector_irq, newcpu)[newvec] = desc;
175}
176
177static void vector_assign_managed_shutdown(struct irq_data *irqd)
178{
179 unsigned int cpu = cpumask_first(cpu_online_mask);
180
181 apic_update_irq_cfg(irqd, MANAGED_IRQ_SHUTDOWN_VECTOR, cpu);
182}
183
184static int reserve_managed_vector(struct irq_data *irqd)
185{
186 const struct cpumask *affmsk = irq_data_get_affinity_mask(irqd);
187 struct apic_chip_data *apicd = apic_chip_data(irqd);
188 unsigned long flags;
189 int ret;
190
191 raw_spin_lock_irqsave(&vector_lock, flags);
192 apicd->is_managed = true;
193 ret = irq_matrix_reserve_managed(vector_matrix, affmsk);
194 raw_spin_unlock_irqrestore(&vector_lock, flags);
195 trace_vector_reserve_managed(irqd->irq, ret);
196 return ret;
197}
198
199static void reserve_irq_vector_locked(struct irq_data *irqd)
200{
201 struct apic_chip_data *apicd = apic_chip_data(irqd);
202
203 irq_matrix_reserve(vector_matrix);
204 apicd->can_reserve = true;
205 apicd->has_reserved = true;
206 irqd_set_can_reserve(irqd);
207 trace_vector_reserve(irqd->irq, 0);
208 vector_assign_managed_shutdown(irqd);
209}
210
211static int reserve_irq_vector(struct irq_data *irqd)
212{
213 unsigned long flags;
214
215 raw_spin_lock_irqsave(&vector_lock, flags);
216 reserve_irq_vector_locked(irqd);
217 raw_spin_unlock_irqrestore(&vector_lock, flags);
218 return 0;
219}
220
221static int allocate_vector(struct irq_data *irqd, const struct cpumask *dest)
222{
223 struct apic_chip_data *apicd = apic_chip_data(irqd);
224 bool resvd = apicd->has_reserved;
225 unsigned int cpu = apicd->cpu;
226 int vector = apicd->vector;
227
228 lockdep_assert_held(&vector_lock);
229
230 /*
231 * If the current target CPU is online and in the new requested
232 * affinity mask, there is no point in moving the interrupt from
233 * one CPU to another.
234 */
235 if (vector && cpu_online(cpu) && cpumask_test_cpu(cpu, dest))
236 return 0;
237
238 vector = irq_matrix_alloc(vector_matrix, dest, resvd, &cpu);
239 if (vector > 0)
240 apic_update_vector(irqd, vector, cpu);
241 trace_vector_alloc(irqd->irq, vector, resvd, vector);
242 return vector;
243}
244
245static int assign_vector_locked(struct irq_data *irqd,
246 const struct cpumask *dest)
247{
248 struct apic_chip_data *apicd = apic_chip_data(irqd);
249 int vector = allocate_vector(irqd, dest);
250
251 if (vector < 0)
252 return vector;
253
254 apic_update_irq_cfg(irqd, apicd->vector, apicd->cpu);
255 return 0;
256}
257
258static int assign_irq_vector(struct irq_data *irqd, const struct cpumask *dest)
259{
260 unsigned long flags;
261 int ret;
262
263 raw_spin_lock_irqsave(&vector_lock, flags);
264 cpumask_and(vector_searchmask, dest, cpu_online_mask);
265 ret = assign_vector_locked(irqd, vector_searchmask);
266 raw_spin_unlock_irqrestore(&vector_lock, flags);
267 return ret;
268}
269
270static int assign_irq_vector_any_locked(struct irq_data *irqd)
271{
272 /* Get the affinity mask - either irq_default_affinity or (user) set */
273 const struct cpumask *affmsk = irq_data_get_affinity_mask(irqd);
274 int node = irq_data_get_node(irqd);
275
276 if (node == NUMA_NO_NODE)
277 goto all;
278 /* Try the intersection of @affmsk and node mask */
279 cpumask_and(vector_searchmask, cpumask_of_node(node), affmsk);
280 if (!assign_vector_locked(irqd, vector_searchmask))
281 return 0;
282 /* Try the node mask */
283 if (!assign_vector_locked(irqd, cpumask_of_node(node)))
284 return 0;
285all:
286 /* Try the full affinity mask */
287 cpumask_and(vector_searchmask, affmsk, cpu_online_mask);
288 if (!assign_vector_locked(irqd, vector_searchmask))
289 return 0;
290 /* Try the full online mask */
291 return assign_vector_locked(irqd, cpu_online_mask);
292}
293
294static int
295assign_irq_vector_policy(struct irq_data *irqd, struct irq_alloc_info *info)
296{
297 if (irqd_affinity_is_managed(irqd))
298 return reserve_managed_vector(irqd);
299 if (info->mask)
300 return assign_irq_vector(irqd, info->mask);
301 /*
302 * Make only a global reservation with no guarantee. A real vector
303 * is associated at activation time.
304 */
305 return reserve_irq_vector(irqd);
306}
307
308static int
309assign_managed_vector(struct irq_data *irqd, const struct cpumask *dest)
310{
311 const struct cpumask *affmsk = irq_data_get_affinity_mask(irqd);
312 struct apic_chip_data *apicd = apic_chip_data(irqd);
313 int vector, cpu;
314
315 cpumask_and(vector_searchmask, vector_searchmask, affmsk);
316 cpu = cpumask_first(vector_searchmask);
317 if (cpu >= nr_cpu_ids)
318 return -EINVAL;
319 /* set_affinity might call here for nothing */
320 if (apicd->vector && cpumask_test_cpu(apicd->cpu, vector_searchmask))
321 return 0;
322 vector = irq_matrix_alloc_managed(vector_matrix, cpu);
323 trace_vector_alloc_managed(irqd->irq, vector, vector);
324 if (vector < 0)
325 return vector;
326 apic_update_vector(irqd, vector, cpu);
327 apic_update_irq_cfg(irqd, vector, cpu);
328 return 0;
329}
330
331static void clear_irq_vector(struct irq_data *irqd)
332{
333 struct apic_chip_data *apicd = apic_chip_data(irqd);
334 bool managed = irqd_affinity_is_managed(irqd);
335 unsigned int vector = apicd->vector;
336
337 lockdep_assert_held(&vector_lock);
338
339 if (!vector)
340 return;
341
342 trace_vector_clear(irqd->irq, vector, apicd->cpu, apicd->prev_vector,
343 apicd->prev_cpu);
344
345 per_cpu(vector_irq, apicd->cpu)[vector] = VECTOR_UNUSED;
346 irq_matrix_free(vector_matrix, apicd->cpu, vector, managed);
347 apicd->vector = 0;
348
349 /* Clean up move in progress */
350 vector = apicd->prev_vector;
351 if (!vector)
352 return;
353
354 per_cpu(vector_irq, apicd->prev_cpu)[vector] = VECTOR_UNUSED;
355 irq_matrix_free(vector_matrix, apicd->prev_cpu, vector, managed);
356 apicd->prev_vector = 0;
357 apicd->move_in_progress = 0;
358 hlist_del_init(&apicd->clist);
359}
360
361static void x86_vector_deactivate(struct irq_domain *dom, struct irq_data *irqd)
362{
363 struct apic_chip_data *apicd = apic_chip_data(irqd);
364 unsigned long flags;
365
366 trace_vector_deactivate(irqd->irq, apicd->is_managed,
367 apicd->can_reserve, false);
368
369 /* Regular fixed assigned interrupt */
370 if (!apicd->is_managed && !apicd->can_reserve)
371 return;
372 /* If the interrupt has a global reservation, nothing to do */
373 if (apicd->has_reserved)
374 return;
375
376 raw_spin_lock_irqsave(&vector_lock, flags);
377 clear_irq_vector(irqd);
378 if (apicd->can_reserve)
379 reserve_irq_vector_locked(irqd);
380 else
381 vector_assign_managed_shutdown(irqd);
382 raw_spin_unlock_irqrestore(&vector_lock, flags);
383}
384
385static int activate_reserved(struct irq_data *irqd)
386{
387 struct apic_chip_data *apicd = apic_chip_data(irqd);
388 int ret;
389
390 ret = assign_irq_vector_any_locked(irqd);
391 if (!ret) {
392 apicd->has_reserved = false;
393 /*
394 * Core might have disabled reservation mode after
395 * allocating the irq descriptor. Ideally this should
396 * happen before allocation time, but that would require
397 * completely convoluted ways of transporting that
398 * information.
399 */
400 if (!irqd_can_reserve(irqd))
401 apicd->can_reserve = false;
402 }
403 return ret;
404}
405
406static int activate_managed(struct irq_data *irqd)
407{
408 const struct cpumask *dest = irq_data_get_affinity_mask(irqd);
409 int ret;
410
411 cpumask_and(vector_searchmask, dest, cpu_online_mask);
412 if (WARN_ON_ONCE(cpumask_empty(vector_searchmask))) {
413 /* Something in the core code broke! Survive gracefully */
414 pr_err("Managed startup for irq %u, but no CPU\n", irqd->irq);
415 return EINVAL;
416 }
417
418 ret = assign_managed_vector(irqd, vector_searchmask);
419 /*
420 * This should not happen. The vector reservation got buggered. Handle
421 * it gracefully.
422 */
423 if (WARN_ON_ONCE(ret < 0)) {
424 pr_err("Managed startup irq %u, no vector available\n",
425 irqd->irq);
426 }
427 return ret;
428}
429
430static int x86_vector_activate(struct irq_domain *dom, struct irq_data *irqd,
431 bool reserve)
432{
433 struct apic_chip_data *apicd = apic_chip_data(irqd);
434 unsigned long flags;
435 int ret = 0;
436
437 trace_vector_activate(irqd->irq, apicd->is_managed,
438 apicd->can_reserve, reserve);
439
440 /* Nothing to do for fixed assigned vectors */
441 if (!apicd->can_reserve && !apicd->is_managed)
442 return 0;
443
444 raw_spin_lock_irqsave(&vector_lock, flags);
445 if (reserve || irqd_is_managed_and_shutdown(irqd))
446 vector_assign_managed_shutdown(irqd);
447 else if (apicd->is_managed)
448 ret = activate_managed(irqd);
449 else if (apicd->has_reserved)
450 ret = activate_reserved(irqd);
451 raw_spin_unlock_irqrestore(&vector_lock, flags);
452 return ret;
453}
454
455static void vector_free_reserved_and_managed(struct irq_data *irqd)
456{
457 const struct cpumask *dest = irq_data_get_affinity_mask(irqd);
458 struct apic_chip_data *apicd = apic_chip_data(irqd);
459
460 trace_vector_teardown(irqd->irq, apicd->is_managed,
461 apicd->has_reserved);
462
463 if (apicd->has_reserved)
464 irq_matrix_remove_reserved(vector_matrix);
465 if (apicd->is_managed)
466 irq_matrix_remove_managed(vector_matrix, dest);
467}
468
469static void x86_vector_free_irqs(struct irq_domain *domain,
470 unsigned int virq, unsigned int nr_irqs)
471{
472 struct apic_chip_data *apicd;
473 struct irq_data *irqd;
474 unsigned long flags;
475 int i;
476
477 for (i = 0; i < nr_irqs; i++) {
478 irqd = irq_domain_get_irq_data(x86_vector_domain, virq + i);
479 if (irqd && irqd->chip_data) {
480 raw_spin_lock_irqsave(&vector_lock, flags);
481 clear_irq_vector(irqd);
482 vector_free_reserved_and_managed(irqd);
483 apicd = irqd->chip_data;
484 irq_domain_reset_irq_data(irqd);
485 raw_spin_unlock_irqrestore(&vector_lock, flags);
486 free_apic_chip_data(apicd);
487 }
488 }
489}
490
491static bool vector_configure_legacy(unsigned int virq, struct irq_data *irqd,
492 struct apic_chip_data *apicd)
493{
494 unsigned long flags;
495 bool realloc = false;
496
497 apicd->vector = ISA_IRQ_VECTOR(virq);
498 apicd->cpu = 0;
499
500 raw_spin_lock_irqsave(&vector_lock, flags);
501 /*
502 * If the interrupt is activated, then it must stay at this vector
503 * position. That's usually the timer interrupt (0).
504 */
505 if (irqd_is_activated(irqd)) {
506 trace_vector_setup(virq, true, 0);
507 apic_update_irq_cfg(irqd, apicd->vector, apicd->cpu);
508 } else {
509 /* Release the vector */
510 apicd->can_reserve = true;
511 irqd_set_can_reserve(irqd);
512 clear_irq_vector(irqd);
513 realloc = true;
514 }
515 raw_spin_unlock_irqrestore(&vector_lock, flags);
516 return realloc;
517}
518
519static int x86_vector_alloc_irqs(struct irq_domain *domain, unsigned int virq,
520 unsigned int nr_irqs, void *arg)
521{
522 struct irq_alloc_info *info = arg;
523 struct apic_chip_data *apicd;
524 struct irq_data *irqd;
525 int i, err, node;
526
527 if (disable_apic)
528 return -ENXIO;
529
530 /* Currently vector allocator can't guarantee contiguous allocations */
531 if ((info->flags & X86_IRQ_ALLOC_CONTIGUOUS_VECTORS) && nr_irqs > 1)
532 return -ENOSYS;
533
534 for (i = 0; i < nr_irqs; i++) {
535 irqd = irq_domain_get_irq_data(domain, virq + i);
536 BUG_ON(!irqd);
537 node = irq_data_get_node(irqd);
538 WARN_ON_ONCE(irqd->chip_data);
539 apicd = alloc_apic_chip_data(node);
540 if (!apicd) {
541 err = -ENOMEM;
542 goto error;
543 }
544
545 apicd->irq = virq + i;
546 irqd->chip = &lapic_controller;
547 irqd->chip_data = apicd;
548 irqd->hwirq = virq + i;
549 irqd_set_single_target(irqd);
550 /*
551 * Legacy vectors are already assigned when the IOAPIC
552 * takes them over. They stay on the same vector. This is
553 * required for check_timer() to work correctly as it might
554 * switch back to legacy mode. Only update the hardware
555 * config.
556 */
557 if (info->flags & X86_IRQ_ALLOC_LEGACY) {
558 if (!vector_configure_legacy(virq + i, irqd, apicd))
559 continue;
560 }
561
562 err = assign_irq_vector_policy(irqd, info);
563 trace_vector_setup(virq + i, false, err);
564 if (err) {
565 irqd->chip_data = NULL;
566 free_apic_chip_data(apicd);
567 goto error;
568 }
569 }
570
571 return 0;
572
573error:
574 x86_vector_free_irqs(domain, virq, i);
575 return err;
576}
577
578#ifdef CONFIG_GENERIC_IRQ_DEBUGFS
579static void x86_vector_debug_show(struct seq_file *m, struct irq_domain *d,
580 struct irq_data *irqd, int ind)
581{
582 unsigned int cpu, vector, prev_cpu, prev_vector;
583 struct apic_chip_data *apicd;
584 unsigned long flags;
585 int irq;
586
587 if (!irqd) {
588 irq_matrix_debug_show(m, vector_matrix, ind);
589 return;
590 }
591
592 irq = irqd->irq;
593 if (irq < nr_legacy_irqs() && !test_bit(irq, &io_apic_irqs)) {
594 seq_printf(m, "%*sVector: %5d\n", ind, "", ISA_IRQ_VECTOR(irq));
595 seq_printf(m, "%*sTarget: Legacy PIC all CPUs\n", ind, "");
596 return;
597 }
598
599 apicd = irqd->chip_data;
600 if (!apicd) {
601 seq_printf(m, "%*sVector: Not assigned\n", ind, "");
602 return;
603 }
604
605 raw_spin_lock_irqsave(&vector_lock, flags);
606 cpu = apicd->cpu;
607 vector = apicd->vector;
608 prev_cpu = apicd->prev_cpu;
609 prev_vector = apicd->prev_vector;
610 raw_spin_unlock_irqrestore(&vector_lock, flags);
611 seq_printf(m, "%*sVector: %5u\n", ind, "", vector);
612 seq_printf(m, "%*sTarget: %5u\n", ind, "", cpu);
613 if (prev_vector) {
614 seq_printf(m, "%*sPrevious vector: %5u\n", ind, "", prev_vector);
615 seq_printf(m, "%*sPrevious target: %5u\n", ind, "", prev_cpu);
616 }
617}
618#endif
619
620static const struct irq_domain_ops x86_vector_domain_ops = {
621 .alloc = x86_vector_alloc_irqs,
622 .free = x86_vector_free_irqs,
623 .activate = x86_vector_activate,
624 .deactivate = x86_vector_deactivate,
625#ifdef CONFIG_GENERIC_IRQ_DEBUGFS
626 .debug_show = x86_vector_debug_show,
627#endif
628};
629
630int __init arch_probe_nr_irqs(void)
631{
632 int nr;
633
634 if (nr_irqs > (NR_VECTORS * nr_cpu_ids))
635 nr_irqs = NR_VECTORS * nr_cpu_ids;
636
637 nr = (gsi_top + nr_legacy_irqs()) + 8 * nr_cpu_ids;
638#if defined(CONFIG_PCI_MSI)
639 /*
640 * for MSI and HT dyn irq
641 */
642 if (gsi_top <= NR_IRQS_LEGACY)
643 nr += 8 * nr_cpu_ids;
644 else
645 nr += gsi_top * 16;
646#endif
647 if (nr < nr_irqs)
648 nr_irqs = nr;
649
650 /*
651 * We don't know if PIC is present at this point so we need to do
652 * probe() to get the right number of legacy IRQs.
653 */
654 return legacy_pic->probe();
655}
656
657void lapic_assign_legacy_vector(unsigned int irq, bool replace)
658{
659 /*
660 * Use assign system here so it wont get accounted as allocated
661 * and moveable in the cpu hotplug check and it prevents managed
662 * irq reservation from touching it.
663 */
664 irq_matrix_assign_system(vector_matrix, ISA_IRQ_VECTOR(irq), replace);
665}
666
667void __init lapic_assign_system_vectors(void)
668{
669 unsigned int i, vector = 0;
670
671 for_each_set_bit_from(vector, system_vectors, NR_VECTORS)
672 irq_matrix_assign_system(vector_matrix, vector, false);
673
674 if (nr_legacy_irqs() > 1)
675 lapic_assign_legacy_vector(PIC_CASCADE_IR, false);
676
677 /* System vectors are reserved, online it */
678 irq_matrix_online(vector_matrix);
679
680 /* Mark the preallocated legacy interrupts */
681 for (i = 0; i < nr_legacy_irqs(); i++) {
682 if (i != PIC_CASCADE_IR)
683 irq_matrix_assign(vector_matrix, ISA_IRQ_VECTOR(i));
684 }
685}
686
687int __init arch_early_irq_init(void)
688{
689 struct fwnode_handle *fn;
690
691 fn = irq_domain_alloc_named_fwnode("VECTOR");
692 BUG_ON(!fn);
693 x86_vector_domain = irq_domain_create_tree(fn, &x86_vector_domain_ops,
694 NULL);
695 BUG_ON(x86_vector_domain == NULL);
696 irq_domain_free_fwnode(fn);
697 irq_set_default_host(x86_vector_domain);
698
699 arch_init_msi_domain(x86_vector_domain);
700
701 BUG_ON(!alloc_cpumask_var(&vector_searchmask, GFP_KERNEL));
702
703 /*
704 * Allocate the vector matrix allocator data structure and limit the
705 * search area.
706 */
707 vector_matrix = irq_alloc_matrix(NR_VECTORS, FIRST_EXTERNAL_VECTOR,
708 FIRST_SYSTEM_VECTOR);
709 BUG_ON(!vector_matrix);
710
711 return arch_early_ioapic_init();
712}
713
714#ifdef CONFIG_SMP
715
716static struct irq_desc *__setup_vector_irq(int vector)
717{
718 int isairq = vector - ISA_IRQ_VECTOR(0);
719
720 /* Check whether the irq is in the legacy space */
721 if (isairq < 0 || isairq >= nr_legacy_irqs())
722 return VECTOR_UNUSED;
723 /* Check whether the irq is handled by the IOAPIC */
724 if (test_bit(isairq, &io_apic_irqs))
725 return VECTOR_UNUSED;
726 return irq_to_desc(isairq);
727}
728
729/* Online the local APIC infrastructure and initialize the vectors */
730void lapic_online(void)
731{
732 unsigned int vector;
733
734 lockdep_assert_held(&vector_lock);
735
736 /* Online the vector matrix array for this CPU */
737 irq_matrix_online(vector_matrix);
738
739 /*
740 * The interrupt affinity logic never targets interrupts to offline
741 * CPUs. The exception are the legacy PIC interrupts. In general
742 * they are only targeted to CPU0, but depending on the platform
743 * they can be distributed to any online CPU in hardware. The
744 * kernel has no influence on that. So all active legacy vectors
745 * must be installed on all CPUs. All non legacy interrupts can be
746 * cleared.
747 */
748 for (vector = 0; vector < NR_VECTORS; vector++)
749 this_cpu_write(vector_irq[vector], __setup_vector_irq(vector));
750}
751
752void lapic_offline(void)
753{
754 lock_vector_lock();
755 irq_matrix_offline(vector_matrix);
756 unlock_vector_lock();
757}
758
759static int apic_set_affinity(struct irq_data *irqd,
760 const struct cpumask *dest, bool force)
761{
762 struct apic_chip_data *apicd = apic_chip_data(irqd);
763 int err;
764
765 /*
766 * Core code can call here for inactive interrupts. For inactive
767 * interrupts which use managed or reservation mode there is no
768 * point in going through the vector assignment right now as the
769 * activation will assign a vector which fits the destination
770 * cpumask. Let the core code store the destination mask and be
771 * done with it.
772 */
773 if (!irqd_is_activated(irqd) &&
774 (apicd->is_managed || apicd->can_reserve))
775 return IRQ_SET_MASK_OK;
776
777 raw_spin_lock(&vector_lock);
778 cpumask_and(vector_searchmask, dest, cpu_online_mask);
779 if (irqd_affinity_is_managed(irqd))
780 err = assign_managed_vector(irqd, vector_searchmask);
781 else
782 err = assign_vector_locked(irqd, vector_searchmask);
783 raw_spin_unlock(&vector_lock);
784 return err ? err : IRQ_SET_MASK_OK;
785}
786
787#else
788# define apic_set_affinity NULL
789#endif
790
791static int apic_retrigger_irq(struct irq_data *irqd)
792{
793 struct apic_chip_data *apicd = apic_chip_data(irqd);
794 unsigned long flags;
795
796 raw_spin_lock_irqsave(&vector_lock, flags);
797 apic->send_IPI(apicd->cpu, apicd->vector);
798 raw_spin_unlock_irqrestore(&vector_lock, flags);
799
800 return 1;
801}
802
803void apic_ack_edge(struct irq_data *irqd)
804{
805 irq_complete_move(irqd_cfg(irqd));
806 irq_move_irq(irqd);
807 ack_APIC_irq();
808}
809
810static struct irq_chip lapic_controller = {
811 .name = "APIC",
812 .irq_ack = apic_ack_edge,
813 .irq_set_affinity = apic_set_affinity,
814 .irq_retrigger = apic_retrigger_irq,
815};
816
817#ifdef CONFIG_SMP
818
819static void free_moved_vector(struct apic_chip_data *apicd)
820{
821 unsigned int vector = apicd->prev_vector;
822 unsigned int cpu = apicd->prev_cpu;
823 bool managed = apicd->is_managed;
824
825 /*
826 * This should never happen. Managed interrupts are not
827 * migrated except on CPU down, which does not involve the
828 * cleanup vector. But try to keep the accounting correct
829 * nevertheless.
830 */
831 WARN_ON_ONCE(managed);
832
833 trace_vector_free_moved(apicd->irq, cpu, vector, managed);
834 irq_matrix_free(vector_matrix, cpu, vector, managed);
835 per_cpu(vector_irq, cpu)[vector] = VECTOR_UNUSED;
836 hlist_del_init(&apicd->clist);
837 apicd->prev_vector = 0;
838 apicd->move_in_progress = 0;
839}
840
841asmlinkage __visible void __irq_entry smp_irq_move_cleanup_interrupt(void)
842{
843 struct hlist_head *clhead = this_cpu_ptr(&cleanup_list);
844 struct apic_chip_data *apicd;
845 struct hlist_node *tmp;
846
847 entering_ack_irq();
848 /* Prevent vectors vanishing under us */
849 raw_spin_lock(&vector_lock);
850
851 hlist_for_each_entry_safe(apicd, tmp, clhead, clist) {
852 unsigned int irr, vector = apicd->prev_vector;
853
854 /*
855 * Paranoia: Check if the vector that needs to be cleaned
856 * up is registered at the APICs IRR. If so, then this is
857 * not the best time to clean it up. Clean it up in the
858 * next attempt by sending another IRQ_MOVE_CLEANUP_VECTOR
859 * to this CPU. IRQ_MOVE_CLEANUP_VECTOR is the lowest
860 * priority external vector, so on return from this
861 * interrupt the device interrupt will happen first.
862 */
863 irr = apic_read(APIC_IRR + (vector / 32 * 0x10));
864 if (irr & (1U << (vector % 32))) {
865 apic->send_IPI_self(IRQ_MOVE_CLEANUP_VECTOR);
866 continue;
867 }
868 free_moved_vector(apicd);
869 }
870
871 raw_spin_unlock(&vector_lock);
872 exiting_irq();
873}
874
875static void __send_cleanup_vector(struct apic_chip_data *apicd)
876{
877 unsigned int cpu;
878
879 raw_spin_lock(&vector_lock);
880 apicd->move_in_progress = 0;
881 cpu = apicd->prev_cpu;
882 if (cpu_online(cpu)) {
883 hlist_add_head(&apicd->clist, per_cpu_ptr(&cleanup_list, cpu));
884 apic->send_IPI(cpu, IRQ_MOVE_CLEANUP_VECTOR);
885 } else {
886 apicd->prev_vector = 0;
887 }
888 raw_spin_unlock(&vector_lock);
889}
890
891void send_cleanup_vector(struct irq_cfg *cfg)
892{
893 struct apic_chip_data *apicd;
894
895 apicd = container_of(cfg, struct apic_chip_data, hw_irq_cfg);
896 if (apicd->move_in_progress)
897 __send_cleanup_vector(apicd);
898}
899
900static void __irq_complete_move(struct irq_cfg *cfg, unsigned vector)
901{
902 struct apic_chip_data *apicd;
903
904 apicd = container_of(cfg, struct apic_chip_data, hw_irq_cfg);
905 if (likely(!apicd->move_in_progress))
906 return;
907
908 if (vector == apicd->vector && apicd->cpu == smp_processor_id())
909 __send_cleanup_vector(apicd);
910}
911
912void irq_complete_move(struct irq_cfg *cfg)
913{
914 __irq_complete_move(cfg, ~get_irq_regs()->orig_ax);
915}
916
917/*
918 * Called from fixup_irqs() with @desc->lock held and interrupts disabled.
919 */
920void irq_force_complete_move(struct irq_desc *desc)
921{
922 struct apic_chip_data *apicd;
923 struct irq_data *irqd;
924 unsigned int vector;
925
926 /*
927 * The function is called for all descriptors regardless of which
928 * irqdomain they belong to. For example if an IRQ is provided by
929 * an irq_chip as part of a GPIO driver, the chip data for that
930 * descriptor is specific to the irq_chip in question.
931 *
932 * Check first that the chip_data is what we expect
933 * (apic_chip_data) before touching it any further.
934 */
935 irqd = irq_domain_get_irq_data(x86_vector_domain,
936 irq_desc_get_irq(desc));
937 if (!irqd)
938 return;
939
940 raw_spin_lock(&vector_lock);
941 apicd = apic_chip_data(irqd);
942 if (!apicd)
943 goto unlock;
944
945 /*
946 * If prev_vector is empty, no action required.
947 */
948 vector = apicd->prev_vector;
949 if (!vector)
950 goto unlock;
951
952 /*
953 * This is tricky. If the cleanup of the old vector has not been
954 * done yet, then the following setaffinity call will fail with
955 * -EBUSY. This can leave the interrupt in a stale state.
956 *
957 * All CPUs are stuck in stop machine with interrupts disabled so
958 * calling __irq_complete_move() would be completely pointless.
959 *
960 * 1) The interrupt is in move_in_progress state. That means that we
961 * have not seen an interrupt since the io_apic was reprogrammed to
962 * the new vector.
963 *
964 * 2) The interrupt has fired on the new vector, but the cleanup IPIs
965 * have not been processed yet.
966 */
967 if (apicd->move_in_progress) {
968 /*
969 * In theory there is a race:
970 *
971 * set_ioapic(new_vector) <-- Interrupt is raised before update
972 * is effective, i.e. it's raised on
973 * the old vector.
974 *
975 * So if the target cpu cannot handle that interrupt before
976 * the old vector is cleaned up, we get a spurious interrupt
977 * and in the worst case the ioapic irq line becomes stale.
978 *
979 * But in case of cpu hotplug this should be a non issue
980 * because if the affinity update happens right before all
981 * cpus rendevouz in stop machine, there is no way that the
982 * interrupt can be blocked on the target cpu because all cpus
983 * loops first with interrupts enabled in stop machine, so the
984 * old vector is not yet cleaned up when the interrupt fires.
985 *
986 * So the only way to run into this issue is if the delivery
987 * of the interrupt on the apic/system bus would be delayed
988 * beyond the point where the target cpu disables interrupts
989 * in stop machine. I doubt that it can happen, but at least
990 * there is a theroretical chance. Virtualization might be
991 * able to expose this, but AFAICT the IOAPIC emulation is not
992 * as stupid as the real hardware.
993 *
994 * Anyway, there is nothing we can do about that at this point
995 * w/o refactoring the whole fixup_irq() business completely.
996 * We print at least the irq number and the old vector number,
997 * so we have the necessary information when a problem in that
998 * area arises.
999 */
1000 pr_warn("IRQ fixup: irq %d move in progress, old vector %d\n",
1001 irqd->irq, vector);
1002 }
1003 free_moved_vector(apicd);
1004unlock:
1005 raw_spin_unlock(&vector_lock);
1006}
1007
1008#ifdef CONFIG_HOTPLUG_CPU
1009/*
1010 * Note, this is not accurate accounting, but at least good enough to
1011 * prevent that the actual interrupt move will run out of vectors.
1012 */
1013int lapic_can_unplug_cpu(void)
1014{
1015 unsigned int rsvd, avl, tomove, cpu = smp_processor_id();
1016 int ret = 0;
1017
1018 raw_spin_lock(&vector_lock);
1019 tomove = irq_matrix_allocated(vector_matrix);
1020 avl = irq_matrix_available(vector_matrix, true);
1021 if (avl < tomove) {
1022 pr_warn("CPU %u has %u vectors, %u available. Cannot disable CPU\n",
1023 cpu, tomove, avl);
1024 ret = -ENOSPC;
1025 goto out;
1026 }
1027 rsvd = irq_matrix_reserved(vector_matrix);
1028 if (avl < rsvd) {
1029 pr_warn("Reserved vectors %u > available %u. IRQ request may fail\n",
1030 rsvd, avl);
1031 }
1032out:
1033 raw_spin_unlock(&vector_lock);
1034 return ret;
1035}
1036#endif /* HOTPLUG_CPU */
1037#endif /* SMP */
1038
1039static void __init print_APIC_field(int base)
1040{
1041 int i;
1042
1043 printk(KERN_DEBUG);
1044
1045 for (i = 0; i < 8; i++)
1046 pr_cont("%08x", apic_read(base + i*0x10));
1047
1048 pr_cont("\n");
1049}
1050
1051static void __init print_local_APIC(void *dummy)
1052{
1053 unsigned int i, v, ver, maxlvt;
1054 u64 icr;
1055
1056 pr_debug("printing local APIC contents on CPU#%d/%d:\n",
1057 smp_processor_id(), hard_smp_processor_id());
1058 v = apic_read(APIC_ID);
1059 pr_info("... APIC ID: %08x (%01x)\n", v, read_apic_id());
1060 v = apic_read(APIC_LVR);
1061 pr_info("... APIC VERSION: %08x\n", v);
1062 ver = GET_APIC_VERSION(v);
1063 maxlvt = lapic_get_maxlvt();
1064
1065 v = apic_read(APIC_TASKPRI);
1066 pr_debug("... APIC TASKPRI: %08x (%02x)\n", v, v & APIC_TPRI_MASK);
1067
1068 /* !82489DX */
1069 if (APIC_INTEGRATED(ver)) {
1070 if (!APIC_XAPIC(ver)) {
1071 v = apic_read(APIC_ARBPRI);
1072 pr_debug("... APIC ARBPRI: %08x (%02x)\n",
1073 v, v & APIC_ARBPRI_MASK);
1074 }
1075 v = apic_read(APIC_PROCPRI);
1076 pr_debug("... APIC PROCPRI: %08x\n", v);
1077 }
1078
1079 /*
1080 * Remote read supported only in the 82489DX and local APIC for
1081 * Pentium processors.
1082 */
1083 if (!APIC_INTEGRATED(ver) || maxlvt == 3) {
1084 v = apic_read(APIC_RRR);
1085 pr_debug("... APIC RRR: %08x\n", v);
1086 }
1087
1088 v = apic_read(APIC_LDR);
1089 pr_debug("... APIC LDR: %08x\n", v);
1090 if (!x2apic_enabled()) {
1091 v = apic_read(APIC_DFR);
1092 pr_debug("... APIC DFR: %08x\n", v);
1093 }
1094 v = apic_read(APIC_SPIV);
1095 pr_debug("... APIC SPIV: %08x\n", v);
1096
1097 pr_debug("... APIC ISR field:\n");
1098 print_APIC_field(APIC_ISR);
1099 pr_debug("... APIC TMR field:\n");
1100 print_APIC_field(APIC_TMR);
1101 pr_debug("... APIC IRR field:\n");
1102 print_APIC_field(APIC_IRR);
1103
1104 /* !82489DX */
1105 if (APIC_INTEGRATED(ver)) {
1106 /* Due to the Pentium erratum 3AP. */
1107 if (maxlvt > 3)
1108 apic_write(APIC_ESR, 0);
1109
1110 v = apic_read(APIC_ESR);
1111 pr_debug("... APIC ESR: %08x\n", v);
1112 }
1113
1114 icr = apic_icr_read();
1115 pr_debug("... APIC ICR: %08x\n", (u32)icr);
1116 pr_debug("... APIC ICR2: %08x\n", (u32)(icr >> 32));
1117
1118 v = apic_read(APIC_LVTT);
1119 pr_debug("... APIC LVTT: %08x\n", v);
1120
1121 if (maxlvt > 3) {
1122 /* PC is LVT#4. */
1123 v = apic_read(APIC_LVTPC);
1124 pr_debug("... APIC LVTPC: %08x\n", v);
1125 }
1126 v = apic_read(APIC_LVT0);
1127 pr_debug("... APIC LVT0: %08x\n", v);
1128 v = apic_read(APIC_LVT1);
1129 pr_debug("... APIC LVT1: %08x\n", v);
1130
1131 if (maxlvt > 2) {
1132 /* ERR is LVT#3. */
1133 v = apic_read(APIC_LVTERR);
1134 pr_debug("... APIC LVTERR: %08x\n", v);
1135 }
1136
1137 v = apic_read(APIC_TMICT);
1138 pr_debug("... APIC TMICT: %08x\n", v);
1139 v = apic_read(APIC_TMCCT);
1140 pr_debug("... APIC TMCCT: %08x\n", v);
1141 v = apic_read(APIC_TDCR);
1142 pr_debug("... APIC TDCR: %08x\n", v);
1143
1144 if (boot_cpu_has(X86_FEATURE_EXTAPIC)) {
1145 v = apic_read(APIC_EFEAT);
1146 maxlvt = (v >> 16) & 0xff;
1147 pr_debug("... APIC EFEAT: %08x\n", v);
1148 v = apic_read(APIC_ECTRL);
1149 pr_debug("... APIC ECTRL: %08x\n", v);
1150 for (i = 0; i < maxlvt; i++) {
1151 v = apic_read(APIC_EILVTn(i));
1152 pr_debug("... APIC EILVT%d: %08x\n", i, v);
1153 }
1154 }
1155 pr_cont("\n");
1156}
1157
1158static void __init print_local_APICs(int maxcpu)
1159{
1160 int cpu;
1161
1162 if (!maxcpu)
1163 return;
1164
1165 preempt_disable();
1166 for_each_online_cpu(cpu) {
1167 if (cpu >= maxcpu)
1168 break;
1169 smp_call_function_single(cpu, print_local_APIC, NULL, 1);
1170 }
1171 preempt_enable();
1172}
1173
1174static void __init print_PIC(void)
1175{
1176 unsigned int v;
1177 unsigned long flags;
1178
1179 if (!nr_legacy_irqs())
1180 return;
1181
1182 pr_debug("\nprinting PIC contents\n");
1183
1184 raw_spin_lock_irqsave(&i8259A_lock, flags);
1185
1186 v = inb(0xa1) << 8 | inb(0x21);
1187 pr_debug("... PIC IMR: %04x\n", v);
1188
1189 v = inb(0xa0) << 8 | inb(0x20);
1190 pr_debug("... PIC IRR: %04x\n", v);
1191
1192 outb(0x0b, 0xa0);
1193 outb(0x0b, 0x20);
1194 v = inb(0xa0) << 8 | inb(0x20);
1195 outb(0x0a, 0xa0);
1196 outb(0x0a, 0x20);
1197
1198 raw_spin_unlock_irqrestore(&i8259A_lock, flags);
1199
1200 pr_debug("... PIC ISR: %04x\n", v);
1201
1202 v = inb(0x4d1) << 8 | inb(0x4d0);
1203 pr_debug("... PIC ELCR: %04x\n", v);
1204}
1205
1206static int show_lapic __initdata = 1;
1207static __init int setup_show_lapic(char *arg)
1208{
1209 int num = -1;
1210
1211 if (strcmp(arg, "all") == 0) {
1212 show_lapic = CONFIG_NR_CPUS;
1213 } else {
1214 get_option(&arg, &num);
1215 if (num >= 0)
1216 show_lapic = num;
1217 }
1218
1219 return 1;
1220}
1221__setup("show_lapic=", setup_show_lapic);
1222
1223static int __init print_ICs(void)
1224{
1225 if (apic_verbosity == APIC_QUIET)
1226 return 0;
1227
1228 print_PIC();
1229
1230 /* don't print out if apic is not there */
1231 if (!boot_cpu_has(X86_FEATURE_APIC) && !apic_from_smp_config())
1232 return 0;
1233
1234 print_local_APICs(show_lapic);
1235 print_IO_APICs();
1236
1237 return 0;
1238}
1239
1240late_initcall(print_ICs);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Local APIC related interfaces to support IOAPIC, MSI, etc.
4 *
5 * Copyright (C) 1997, 1998, 1999, 2000, 2009 Ingo Molnar, Hajnalka Szabo
6 * Moved from arch/x86/kernel/apic/io_apic.c.
7 * Jiang Liu <jiang.liu@linux.intel.com>
8 * Enable support of hierarchical irqdomains
9 */
10#include <linux/interrupt.h>
11#include <linux/irq.h>
12#include <linux/seq_file.h>
13#include <linux/init.h>
14#include <linux/compiler.h>
15#include <linux/slab.h>
16#include <asm/irqdomain.h>
17#include <asm/hw_irq.h>
18#include <asm/traps.h>
19#include <asm/apic.h>
20#include <asm/i8259.h>
21#include <asm/desc.h>
22#include <asm/irq_remapping.h>
23
24#include <asm/trace/irq_vectors.h>
25
26struct apic_chip_data {
27 struct irq_cfg hw_irq_cfg;
28 unsigned int vector;
29 unsigned int prev_vector;
30 unsigned int cpu;
31 unsigned int prev_cpu;
32 unsigned int irq;
33 struct hlist_node clist;
34 unsigned int move_in_progress : 1,
35 is_managed : 1,
36 can_reserve : 1,
37 has_reserved : 1;
38};
39
40struct irq_domain *x86_vector_domain;
41EXPORT_SYMBOL_GPL(x86_vector_domain);
42static DEFINE_RAW_SPINLOCK(vector_lock);
43static cpumask_var_t vector_searchmask;
44static struct irq_chip lapic_controller;
45static struct irq_matrix *vector_matrix;
46#ifdef CONFIG_SMP
47static DEFINE_PER_CPU(struct hlist_head, cleanup_list);
48#endif
49
50void lock_vector_lock(void)
51{
52 /* Used to the online set of cpus does not change
53 * during assign_irq_vector.
54 */
55 raw_spin_lock(&vector_lock);
56}
57
58void unlock_vector_lock(void)
59{
60 raw_spin_unlock(&vector_lock);
61}
62
63void init_irq_alloc_info(struct irq_alloc_info *info,
64 const struct cpumask *mask)
65{
66 memset(info, 0, sizeof(*info));
67 info->mask = mask;
68}
69
70void copy_irq_alloc_info(struct irq_alloc_info *dst, struct irq_alloc_info *src)
71{
72 if (src)
73 *dst = *src;
74 else
75 memset(dst, 0, sizeof(*dst));
76}
77
78static struct apic_chip_data *apic_chip_data(struct irq_data *irqd)
79{
80 if (!irqd)
81 return NULL;
82
83 while (irqd->parent_data)
84 irqd = irqd->parent_data;
85
86 return irqd->chip_data;
87}
88
89struct irq_cfg *irqd_cfg(struct irq_data *irqd)
90{
91 struct apic_chip_data *apicd = apic_chip_data(irqd);
92
93 return apicd ? &apicd->hw_irq_cfg : NULL;
94}
95EXPORT_SYMBOL_GPL(irqd_cfg);
96
97struct irq_cfg *irq_cfg(unsigned int irq)
98{
99 return irqd_cfg(irq_get_irq_data(irq));
100}
101
102static struct apic_chip_data *alloc_apic_chip_data(int node)
103{
104 struct apic_chip_data *apicd;
105
106 apicd = kzalloc_node(sizeof(*apicd), GFP_KERNEL, node);
107 if (apicd)
108 INIT_HLIST_NODE(&apicd->clist);
109 return apicd;
110}
111
112static void free_apic_chip_data(struct apic_chip_data *apicd)
113{
114 kfree(apicd);
115}
116
117static void apic_update_irq_cfg(struct irq_data *irqd, unsigned int vector,
118 unsigned int cpu)
119{
120 struct apic_chip_data *apicd = apic_chip_data(irqd);
121
122 lockdep_assert_held(&vector_lock);
123
124 apicd->hw_irq_cfg.vector = vector;
125 apicd->hw_irq_cfg.dest_apicid = apic->calc_dest_apicid(cpu);
126 irq_data_update_effective_affinity(irqd, cpumask_of(cpu));
127 trace_vector_config(irqd->irq, vector, cpu,
128 apicd->hw_irq_cfg.dest_apicid);
129}
130
131static void apic_update_vector(struct irq_data *irqd, unsigned int newvec,
132 unsigned int newcpu)
133{
134 struct apic_chip_data *apicd = apic_chip_data(irqd);
135 struct irq_desc *desc = irq_data_to_desc(irqd);
136 bool managed = irqd_affinity_is_managed(irqd);
137
138 lockdep_assert_held(&vector_lock);
139
140 trace_vector_update(irqd->irq, newvec, newcpu, apicd->vector,
141 apicd->cpu);
142
143 /*
144 * If there is no vector associated or if the associated vector is
145 * the shutdown vector, which is associated to make PCI/MSI
146 * shutdown mode work, then there is nothing to release. Clear out
147 * prev_vector for this and the offlined target case.
148 */
149 apicd->prev_vector = 0;
150 if (!apicd->vector || apicd->vector == MANAGED_IRQ_SHUTDOWN_VECTOR)
151 goto setnew;
152 /*
153 * If the target CPU of the previous vector is online, then mark
154 * the vector as move in progress and store it for cleanup when the
155 * first interrupt on the new vector arrives. If the target CPU is
156 * offline then the regular release mechanism via the cleanup
157 * vector is not possible and the vector can be immediately freed
158 * in the underlying matrix allocator.
159 */
160 if (cpu_online(apicd->cpu)) {
161 apicd->move_in_progress = true;
162 apicd->prev_vector = apicd->vector;
163 apicd->prev_cpu = apicd->cpu;
164 WARN_ON_ONCE(apicd->cpu == newcpu);
165 } else {
166 irq_matrix_free(vector_matrix, apicd->cpu, apicd->vector,
167 managed);
168 }
169
170setnew:
171 apicd->vector = newvec;
172 apicd->cpu = newcpu;
173 BUG_ON(!IS_ERR_OR_NULL(per_cpu(vector_irq, newcpu)[newvec]));
174 per_cpu(vector_irq, newcpu)[newvec] = desc;
175}
176
177static void vector_assign_managed_shutdown(struct irq_data *irqd)
178{
179 unsigned int cpu = cpumask_first(cpu_online_mask);
180
181 apic_update_irq_cfg(irqd, MANAGED_IRQ_SHUTDOWN_VECTOR, cpu);
182}
183
184static int reserve_managed_vector(struct irq_data *irqd)
185{
186 const struct cpumask *affmsk = irq_data_get_affinity_mask(irqd);
187 struct apic_chip_data *apicd = apic_chip_data(irqd);
188 unsigned long flags;
189 int ret;
190
191 raw_spin_lock_irqsave(&vector_lock, flags);
192 apicd->is_managed = true;
193 ret = irq_matrix_reserve_managed(vector_matrix, affmsk);
194 raw_spin_unlock_irqrestore(&vector_lock, flags);
195 trace_vector_reserve_managed(irqd->irq, ret);
196 return ret;
197}
198
199static void reserve_irq_vector_locked(struct irq_data *irqd)
200{
201 struct apic_chip_data *apicd = apic_chip_data(irqd);
202
203 irq_matrix_reserve(vector_matrix);
204 apicd->can_reserve = true;
205 apicd->has_reserved = true;
206 irqd_set_can_reserve(irqd);
207 trace_vector_reserve(irqd->irq, 0);
208 vector_assign_managed_shutdown(irqd);
209}
210
211static int reserve_irq_vector(struct irq_data *irqd)
212{
213 unsigned long flags;
214
215 raw_spin_lock_irqsave(&vector_lock, flags);
216 reserve_irq_vector_locked(irqd);
217 raw_spin_unlock_irqrestore(&vector_lock, flags);
218 return 0;
219}
220
221static int
222assign_vector_locked(struct irq_data *irqd, const struct cpumask *dest)
223{
224 struct apic_chip_data *apicd = apic_chip_data(irqd);
225 bool resvd = apicd->has_reserved;
226 unsigned int cpu = apicd->cpu;
227 int vector = apicd->vector;
228
229 lockdep_assert_held(&vector_lock);
230
231 /*
232 * If the current target CPU is online and in the new requested
233 * affinity mask, there is no point in moving the interrupt from
234 * one CPU to another.
235 */
236 if (vector && cpu_online(cpu) && cpumask_test_cpu(cpu, dest))
237 return 0;
238
239 /*
240 * Careful here. @apicd might either have move_in_progress set or
241 * be enqueued for cleanup. Assigning a new vector would either
242 * leave a stale vector on some CPU around or in case of a pending
243 * cleanup corrupt the hlist.
244 */
245 if (apicd->move_in_progress || !hlist_unhashed(&apicd->clist))
246 return -EBUSY;
247
248 vector = irq_matrix_alloc(vector_matrix, dest, resvd, &cpu);
249 trace_vector_alloc(irqd->irq, vector, resvd, vector);
250 if (vector < 0)
251 return vector;
252 apic_update_vector(irqd, vector, cpu);
253 apic_update_irq_cfg(irqd, vector, cpu);
254
255 return 0;
256}
257
258static int assign_irq_vector(struct irq_data *irqd, const struct cpumask *dest)
259{
260 unsigned long flags;
261 int ret;
262
263 raw_spin_lock_irqsave(&vector_lock, flags);
264 cpumask_and(vector_searchmask, dest, cpu_online_mask);
265 ret = assign_vector_locked(irqd, vector_searchmask);
266 raw_spin_unlock_irqrestore(&vector_lock, flags);
267 return ret;
268}
269
270static int assign_irq_vector_any_locked(struct irq_data *irqd)
271{
272 /* Get the affinity mask - either irq_default_affinity or (user) set */
273 const struct cpumask *affmsk = irq_data_get_affinity_mask(irqd);
274 int node = irq_data_get_node(irqd);
275
276 if (node != NUMA_NO_NODE) {
277 /* Try the intersection of @affmsk and node mask */
278 cpumask_and(vector_searchmask, cpumask_of_node(node), affmsk);
279 if (!assign_vector_locked(irqd, vector_searchmask))
280 return 0;
281 }
282
283 /* Try the full affinity mask */
284 cpumask_and(vector_searchmask, affmsk, cpu_online_mask);
285 if (!assign_vector_locked(irqd, vector_searchmask))
286 return 0;
287
288 if (node != NUMA_NO_NODE) {
289 /* Try the node mask */
290 if (!assign_vector_locked(irqd, cpumask_of_node(node)))
291 return 0;
292 }
293
294 /* Try the full online mask */
295 return assign_vector_locked(irqd, cpu_online_mask);
296}
297
298static int
299assign_irq_vector_policy(struct irq_data *irqd, struct irq_alloc_info *info)
300{
301 if (irqd_affinity_is_managed(irqd))
302 return reserve_managed_vector(irqd);
303 if (info->mask)
304 return assign_irq_vector(irqd, info->mask);
305 /*
306 * Make only a global reservation with no guarantee. A real vector
307 * is associated at activation time.
308 */
309 return reserve_irq_vector(irqd);
310}
311
312static int
313assign_managed_vector(struct irq_data *irqd, const struct cpumask *dest)
314{
315 const struct cpumask *affmsk = irq_data_get_affinity_mask(irqd);
316 struct apic_chip_data *apicd = apic_chip_data(irqd);
317 int vector, cpu;
318
319 cpumask_and(vector_searchmask, dest, affmsk);
320
321 /* set_affinity might call here for nothing */
322 if (apicd->vector && cpumask_test_cpu(apicd->cpu, vector_searchmask))
323 return 0;
324 vector = irq_matrix_alloc_managed(vector_matrix, vector_searchmask,
325 &cpu);
326 trace_vector_alloc_managed(irqd->irq, vector, vector);
327 if (vector < 0)
328 return vector;
329 apic_update_vector(irqd, vector, cpu);
330 apic_update_irq_cfg(irqd, vector, cpu);
331 return 0;
332}
333
334static void clear_irq_vector(struct irq_data *irqd)
335{
336 struct apic_chip_data *apicd = apic_chip_data(irqd);
337 bool managed = irqd_affinity_is_managed(irqd);
338 unsigned int vector = apicd->vector;
339
340 lockdep_assert_held(&vector_lock);
341
342 if (!vector)
343 return;
344
345 trace_vector_clear(irqd->irq, vector, apicd->cpu, apicd->prev_vector,
346 apicd->prev_cpu);
347
348 per_cpu(vector_irq, apicd->cpu)[vector] = VECTOR_SHUTDOWN;
349 irq_matrix_free(vector_matrix, apicd->cpu, vector, managed);
350 apicd->vector = 0;
351
352 /* Clean up move in progress */
353 vector = apicd->prev_vector;
354 if (!vector)
355 return;
356
357 per_cpu(vector_irq, apicd->prev_cpu)[vector] = VECTOR_SHUTDOWN;
358 irq_matrix_free(vector_matrix, apicd->prev_cpu, vector, managed);
359 apicd->prev_vector = 0;
360 apicd->move_in_progress = 0;
361 hlist_del_init(&apicd->clist);
362}
363
364static void x86_vector_deactivate(struct irq_domain *dom, struct irq_data *irqd)
365{
366 struct apic_chip_data *apicd = apic_chip_data(irqd);
367 unsigned long flags;
368
369 trace_vector_deactivate(irqd->irq, apicd->is_managed,
370 apicd->can_reserve, false);
371
372 /* Regular fixed assigned interrupt */
373 if (!apicd->is_managed && !apicd->can_reserve)
374 return;
375 /* If the interrupt has a global reservation, nothing to do */
376 if (apicd->has_reserved)
377 return;
378
379 raw_spin_lock_irqsave(&vector_lock, flags);
380 clear_irq_vector(irqd);
381 if (apicd->can_reserve)
382 reserve_irq_vector_locked(irqd);
383 else
384 vector_assign_managed_shutdown(irqd);
385 raw_spin_unlock_irqrestore(&vector_lock, flags);
386}
387
388static int activate_reserved(struct irq_data *irqd)
389{
390 struct apic_chip_data *apicd = apic_chip_data(irqd);
391 int ret;
392
393 ret = assign_irq_vector_any_locked(irqd);
394 if (!ret) {
395 apicd->has_reserved = false;
396 /*
397 * Core might have disabled reservation mode after
398 * allocating the irq descriptor. Ideally this should
399 * happen before allocation time, but that would require
400 * completely convoluted ways of transporting that
401 * information.
402 */
403 if (!irqd_can_reserve(irqd))
404 apicd->can_reserve = false;
405 }
406
407 /*
408 * Check to ensure that the effective affinity mask is a subset
409 * the user supplied affinity mask, and warn the user if it is not
410 */
411 if (!cpumask_subset(irq_data_get_effective_affinity_mask(irqd),
412 irq_data_get_affinity_mask(irqd))) {
413 pr_warn("irq %u: Affinity broken due to vector space exhaustion.\n",
414 irqd->irq);
415 }
416
417 return ret;
418}
419
420static int activate_managed(struct irq_data *irqd)
421{
422 const struct cpumask *dest = irq_data_get_affinity_mask(irqd);
423 int ret;
424
425 cpumask_and(vector_searchmask, dest, cpu_online_mask);
426 if (WARN_ON_ONCE(cpumask_empty(vector_searchmask))) {
427 /* Something in the core code broke! Survive gracefully */
428 pr_err("Managed startup for irq %u, but no CPU\n", irqd->irq);
429 return -EINVAL;
430 }
431
432 ret = assign_managed_vector(irqd, vector_searchmask);
433 /*
434 * This should not happen. The vector reservation got buggered. Handle
435 * it gracefully.
436 */
437 if (WARN_ON_ONCE(ret < 0)) {
438 pr_err("Managed startup irq %u, no vector available\n",
439 irqd->irq);
440 }
441 return ret;
442}
443
444static int x86_vector_activate(struct irq_domain *dom, struct irq_data *irqd,
445 bool reserve)
446{
447 struct apic_chip_data *apicd = apic_chip_data(irqd);
448 unsigned long flags;
449 int ret = 0;
450
451 trace_vector_activate(irqd->irq, apicd->is_managed,
452 apicd->can_reserve, reserve);
453
454 raw_spin_lock_irqsave(&vector_lock, flags);
455 if (!apicd->can_reserve && !apicd->is_managed)
456 assign_irq_vector_any_locked(irqd);
457 else if (reserve || irqd_is_managed_and_shutdown(irqd))
458 vector_assign_managed_shutdown(irqd);
459 else if (apicd->is_managed)
460 ret = activate_managed(irqd);
461 else if (apicd->has_reserved)
462 ret = activate_reserved(irqd);
463 raw_spin_unlock_irqrestore(&vector_lock, flags);
464 return ret;
465}
466
467static void vector_free_reserved_and_managed(struct irq_data *irqd)
468{
469 const struct cpumask *dest = irq_data_get_affinity_mask(irqd);
470 struct apic_chip_data *apicd = apic_chip_data(irqd);
471
472 trace_vector_teardown(irqd->irq, apicd->is_managed,
473 apicd->has_reserved);
474
475 if (apicd->has_reserved)
476 irq_matrix_remove_reserved(vector_matrix);
477 if (apicd->is_managed)
478 irq_matrix_remove_managed(vector_matrix, dest);
479}
480
481static void x86_vector_free_irqs(struct irq_domain *domain,
482 unsigned int virq, unsigned int nr_irqs)
483{
484 struct apic_chip_data *apicd;
485 struct irq_data *irqd;
486 unsigned long flags;
487 int i;
488
489 for (i = 0; i < nr_irqs; i++) {
490 irqd = irq_domain_get_irq_data(x86_vector_domain, virq + i);
491 if (irqd && irqd->chip_data) {
492 raw_spin_lock_irqsave(&vector_lock, flags);
493 clear_irq_vector(irqd);
494 vector_free_reserved_and_managed(irqd);
495 apicd = irqd->chip_data;
496 irq_domain_reset_irq_data(irqd);
497 raw_spin_unlock_irqrestore(&vector_lock, flags);
498 free_apic_chip_data(apicd);
499 }
500 }
501}
502
503static bool vector_configure_legacy(unsigned int virq, struct irq_data *irqd,
504 struct apic_chip_data *apicd)
505{
506 unsigned long flags;
507 bool realloc = false;
508
509 apicd->vector = ISA_IRQ_VECTOR(virq);
510 apicd->cpu = 0;
511
512 raw_spin_lock_irqsave(&vector_lock, flags);
513 /*
514 * If the interrupt is activated, then it must stay at this vector
515 * position. That's usually the timer interrupt (0).
516 */
517 if (irqd_is_activated(irqd)) {
518 trace_vector_setup(virq, true, 0);
519 apic_update_irq_cfg(irqd, apicd->vector, apicd->cpu);
520 } else {
521 /* Release the vector */
522 apicd->can_reserve = true;
523 irqd_set_can_reserve(irqd);
524 clear_irq_vector(irqd);
525 realloc = true;
526 }
527 raw_spin_unlock_irqrestore(&vector_lock, flags);
528 return realloc;
529}
530
531static int x86_vector_alloc_irqs(struct irq_domain *domain, unsigned int virq,
532 unsigned int nr_irqs, void *arg)
533{
534 struct irq_alloc_info *info = arg;
535 struct apic_chip_data *apicd;
536 struct irq_data *irqd;
537 int i, err, node;
538
539 if (disable_apic)
540 return -ENXIO;
541
542 /* Currently vector allocator can't guarantee contiguous allocations */
543 if ((info->flags & X86_IRQ_ALLOC_CONTIGUOUS_VECTORS) && nr_irqs > 1)
544 return -ENOSYS;
545
546 /*
547 * Catch any attempt to touch the cascade interrupt on a PIC
548 * equipped system.
549 */
550 if (WARN_ON_ONCE(info->flags & X86_IRQ_ALLOC_LEGACY &&
551 virq == PIC_CASCADE_IR))
552 return -EINVAL;
553
554 for (i = 0; i < nr_irqs; i++) {
555 irqd = irq_domain_get_irq_data(domain, virq + i);
556 BUG_ON(!irqd);
557 node = irq_data_get_node(irqd);
558 WARN_ON_ONCE(irqd->chip_data);
559 apicd = alloc_apic_chip_data(node);
560 if (!apicd) {
561 err = -ENOMEM;
562 goto error;
563 }
564
565 apicd->irq = virq + i;
566 irqd->chip = &lapic_controller;
567 irqd->chip_data = apicd;
568 irqd->hwirq = virq + i;
569 irqd_set_single_target(irqd);
570 /*
571 * Prevent that any of these interrupts is invoked in
572 * non interrupt context via e.g. generic_handle_irq()
573 * as that can corrupt the affinity move state.
574 */
575 irqd_set_handle_enforce_irqctx(irqd);
576
577 /* Don't invoke affinity setter on deactivated interrupts */
578 irqd_set_affinity_on_activate(irqd);
579
580 /*
581 * Legacy vectors are already assigned when the IOAPIC
582 * takes them over. They stay on the same vector. This is
583 * required for check_timer() to work correctly as it might
584 * switch back to legacy mode. Only update the hardware
585 * config.
586 */
587 if (info->flags & X86_IRQ_ALLOC_LEGACY) {
588 if (!vector_configure_legacy(virq + i, irqd, apicd))
589 continue;
590 }
591
592 err = assign_irq_vector_policy(irqd, info);
593 trace_vector_setup(virq + i, false, err);
594 if (err) {
595 irqd->chip_data = NULL;
596 free_apic_chip_data(apicd);
597 goto error;
598 }
599 }
600
601 return 0;
602
603error:
604 x86_vector_free_irqs(domain, virq, i);
605 return err;
606}
607
608#ifdef CONFIG_GENERIC_IRQ_DEBUGFS
609static void x86_vector_debug_show(struct seq_file *m, struct irq_domain *d,
610 struct irq_data *irqd, int ind)
611{
612 struct apic_chip_data apicd;
613 unsigned long flags;
614 int irq;
615
616 if (!irqd) {
617 irq_matrix_debug_show(m, vector_matrix, ind);
618 return;
619 }
620
621 irq = irqd->irq;
622 if (irq < nr_legacy_irqs() && !test_bit(irq, &io_apic_irqs)) {
623 seq_printf(m, "%*sVector: %5d\n", ind, "", ISA_IRQ_VECTOR(irq));
624 seq_printf(m, "%*sTarget: Legacy PIC all CPUs\n", ind, "");
625 return;
626 }
627
628 if (!irqd->chip_data) {
629 seq_printf(m, "%*sVector: Not assigned\n", ind, "");
630 return;
631 }
632
633 raw_spin_lock_irqsave(&vector_lock, flags);
634 memcpy(&apicd, irqd->chip_data, sizeof(apicd));
635 raw_spin_unlock_irqrestore(&vector_lock, flags);
636
637 seq_printf(m, "%*sVector: %5u\n", ind, "", apicd.vector);
638 seq_printf(m, "%*sTarget: %5u\n", ind, "", apicd.cpu);
639 if (apicd.prev_vector) {
640 seq_printf(m, "%*sPrevious vector: %5u\n", ind, "", apicd.prev_vector);
641 seq_printf(m, "%*sPrevious target: %5u\n", ind, "", apicd.prev_cpu);
642 }
643 seq_printf(m, "%*smove_in_progress: %u\n", ind, "", apicd.move_in_progress ? 1 : 0);
644 seq_printf(m, "%*sis_managed: %u\n", ind, "", apicd.is_managed ? 1 : 0);
645 seq_printf(m, "%*scan_reserve: %u\n", ind, "", apicd.can_reserve ? 1 : 0);
646 seq_printf(m, "%*shas_reserved: %u\n", ind, "", apicd.has_reserved ? 1 : 0);
647 seq_printf(m, "%*scleanup_pending: %u\n", ind, "", !hlist_unhashed(&apicd.clist));
648}
649#endif
650
651int x86_fwspec_is_ioapic(struct irq_fwspec *fwspec)
652{
653 if (fwspec->param_count != 1)
654 return 0;
655
656 if (is_fwnode_irqchip(fwspec->fwnode)) {
657 const char *fwname = fwnode_get_name(fwspec->fwnode);
658 return fwname && !strncmp(fwname, "IO-APIC-", 8) &&
659 simple_strtol(fwname+8, NULL, 10) == fwspec->param[0];
660 }
661 return to_of_node(fwspec->fwnode) &&
662 of_device_is_compatible(to_of_node(fwspec->fwnode),
663 "intel,ce4100-ioapic");
664}
665
666int x86_fwspec_is_hpet(struct irq_fwspec *fwspec)
667{
668 if (fwspec->param_count != 1)
669 return 0;
670
671 if (is_fwnode_irqchip(fwspec->fwnode)) {
672 const char *fwname = fwnode_get_name(fwspec->fwnode);
673 return fwname && !strncmp(fwname, "HPET-MSI-", 9) &&
674 simple_strtol(fwname+9, NULL, 10) == fwspec->param[0];
675 }
676 return 0;
677}
678
679static int x86_vector_select(struct irq_domain *d, struct irq_fwspec *fwspec,
680 enum irq_domain_bus_token bus_token)
681{
682 /*
683 * HPET and I/OAPIC cannot be parented in the vector domain
684 * if IRQ remapping is enabled. APIC IDs above 15 bits are
685 * only permitted if IRQ remapping is enabled, so check that.
686 */
687 if (apic->apic_id_valid(32768))
688 return 0;
689
690 return x86_fwspec_is_ioapic(fwspec) || x86_fwspec_is_hpet(fwspec);
691}
692
693static const struct irq_domain_ops x86_vector_domain_ops = {
694 .select = x86_vector_select,
695 .alloc = x86_vector_alloc_irqs,
696 .free = x86_vector_free_irqs,
697 .activate = x86_vector_activate,
698 .deactivate = x86_vector_deactivate,
699#ifdef CONFIG_GENERIC_IRQ_DEBUGFS
700 .debug_show = x86_vector_debug_show,
701#endif
702};
703
704int __init arch_probe_nr_irqs(void)
705{
706 int nr;
707
708 if (nr_irqs > (NR_VECTORS * nr_cpu_ids))
709 nr_irqs = NR_VECTORS * nr_cpu_ids;
710
711 nr = (gsi_top + nr_legacy_irqs()) + 8 * nr_cpu_ids;
712#if defined(CONFIG_PCI_MSI)
713 /*
714 * for MSI and HT dyn irq
715 */
716 if (gsi_top <= NR_IRQS_LEGACY)
717 nr += 8 * nr_cpu_ids;
718 else
719 nr += gsi_top * 16;
720#endif
721 if (nr < nr_irqs)
722 nr_irqs = nr;
723
724 /*
725 * We don't know if PIC is present at this point so we need to do
726 * probe() to get the right number of legacy IRQs.
727 */
728 return legacy_pic->probe();
729}
730
731void lapic_assign_legacy_vector(unsigned int irq, bool replace)
732{
733 /*
734 * Use assign system here so it wont get accounted as allocated
735 * and moveable in the cpu hotplug check and it prevents managed
736 * irq reservation from touching it.
737 */
738 irq_matrix_assign_system(vector_matrix, ISA_IRQ_VECTOR(irq), replace);
739}
740
741void __init lapic_update_legacy_vectors(void)
742{
743 unsigned int i;
744
745 if (IS_ENABLED(CONFIG_X86_IO_APIC) && nr_ioapics > 0)
746 return;
747
748 /*
749 * If the IO/APIC is disabled via config, kernel command line or
750 * lack of enumeration then all legacy interrupts are routed
751 * through the PIC. Make sure that they are marked as legacy
752 * vectors. PIC_CASCADE_IRQ has already been marked in
753 * lapic_assign_system_vectors().
754 */
755 for (i = 0; i < nr_legacy_irqs(); i++) {
756 if (i != PIC_CASCADE_IR)
757 lapic_assign_legacy_vector(i, true);
758 }
759}
760
761void __init lapic_assign_system_vectors(void)
762{
763 unsigned int i, vector = 0;
764
765 for_each_set_bit_from(vector, system_vectors, NR_VECTORS)
766 irq_matrix_assign_system(vector_matrix, vector, false);
767
768 if (nr_legacy_irqs() > 1)
769 lapic_assign_legacy_vector(PIC_CASCADE_IR, false);
770
771 /* System vectors are reserved, online it */
772 irq_matrix_online(vector_matrix);
773
774 /* Mark the preallocated legacy interrupts */
775 for (i = 0; i < nr_legacy_irqs(); i++) {
776 /*
777 * Don't touch the cascade interrupt. It's unusable
778 * on PIC equipped machines. See the large comment
779 * in the IO/APIC code.
780 */
781 if (i != PIC_CASCADE_IR)
782 irq_matrix_assign(vector_matrix, ISA_IRQ_VECTOR(i));
783 }
784}
785
786int __init arch_early_irq_init(void)
787{
788 struct fwnode_handle *fn;
789
790 fn = irq_domain_alloc_named_fwnode("VECTOR");
791 BUG_ON(!fn);
792 x86_vector_domain = irq_domain_create_tree(fn, &x86_vector_domain_ops,
793 NULL);
794 BUG_ON(x86_vector_domain == NULL);
795 irq_set_default_host(x86_vector_domain);
796
797 BUG_ON(!alloc_cpumask_var(&vector_searchmask, GFP_KERNEL));
798
799 /*
800 * Allocate the vector matrix allocator data structure and limit the
801 * search area.
802 */
803 vector_matrix = irq_alloc_matrix(NR_VECTORS, FIRST_EXTERNAL_VECTOR,
804 FIRST_SYSTEM_VECTOR);
805 BUG_ON(!vector_matrix);
806
807 return arch_early_ioapic_init();
808}
809
810#ifdef CONFIG_SMP
811
812static struct irq_desc *__setup_vector_irq(int vector)
813{
814 int isairq = vector - ISA_IRQ_VECTOR(0);
815
816 /* Check whether the irq is in the legacy space */
817 if (isairq < 0 || isairq >= nr_legacy_irqs())
818 return VECTOR_UNUSED;
819 /* Check whether the irq is handled by the IOAPIC */
820 if (test_bit(isairq, &io_apic_irqs))
821 return VECTOR_UNUSED;
822 return irq_to_desc(isairq);
823}
824
825/* Online the local APIC infrastructure and initialize the vectors */
826void lapic_online(void)
827{
828 unsigned int vector;
829
830 lockdep_assert_held(&vector_lock);
831
832 /* Online the vector matrix array for this CPU */
833 irq_matrix_online(vector_matrix);
834
835 /*
836 * The interrupt affinity logic never targets interrupts to offline
837 * CPUs. The exception are the legacy PIC interrupts. In general
838 * they are only targeted to CPU0, but depending on the platform
839 * they can be distributed to any online CPU in hardware. The
840 * kernel has no influence on that. So all active legacy vectors
841 * must be installed on all CPUs. All non legacy interrupts can be
842 * cleared.
843 */
844 for (vector = 0; vector < NR_VECTORS; vector++)
845 this_cpu_write(vector_irq[vector], __setup_vector_irq(vector));
846}
847
848void lapic_offline(void)
849{
850 lock_vector_lock();
851 irq_matrix_offline(vector_matrix);
852 unlock_vector_lock();
853}
854
855static int apic_set_affinity(struct irq_data *irqd,
856 const struct cpumask *dest, bool force)
857{
858 int err;
859
860 if (WARN_ON_ONCE(!irqd_is_activated(irqd)))
861 return -EIO;
862
863 raw_spin_lock(&vector_lock);
864 cpumask_and(vector_searchmask, dest, cpu_online_mask);
865 if (irqd_affinity_is_managed(irqd))
866 err = assign_managed_vector(irqd, vector_searchmask);
867 else
868 err = assign_vector_locked(irqd, vector_searchmask);
869 raw_spin_unlock(&vector_lock);
870 return err ? err : IRQ_SET_MASK_OK;
871}
872
873#else
874# define apic_set_affinity NULL
875#endif
876
877static int apic_retrigger_irq(struct irq_data *irqd)
878{
879 struct apic_chip_data *apicd = apic_chip_data(irqd);
880 unsigned long flags;
881
882 raw_spin_lock_irqsave(&vector_lock, flags);
883 apic->send_IPI(apicd->cpu, apicd->vector);
884 raw_spin_unlock_irqrestore(&vector_lock, flags);
885
886 return 1;
887}
888
889void apic_ack_irq(struct irq_data *irqd)
890{
891 irq_move_irq(irqd);
892 ack_APIC_irq();
893}
894
895void apic_ack_edge(struct irq_data *irqd)
896{
897 irq_complete_move(irqd_cfg(irqd));
898 apic_ack_irq(irqd);
899}
900
901static void x86_vector_msi_compose_msg(struct irq_data *data,
902 struct msi_msg *msg)
903{
904 __irq_msi_compose_msg(irqd_cfg(data), msg, false);
905}
906
907static struct irq_chip lapic_controller = {
908 .name = "APIC",
909 .irq_ack = apic_ack_edge,
910 .irq_set_affinity = apic_set_affinity,
911 .irq_compose_msi_msg = x86_vector_msi_compose_msg,
912 .irq_retrigger = apic_retrigger_irq,
913};
914
915#ifdef CONFIG_SMP
916
917static void free_moved_vector(struct apic_chip_data *apicd)
918{
919 unsigned int vector = apicd->prev_vector;
920 unsigned int cpu = apicd->prev_cpu;
921 bool managed = apicd->is_managed;
922
923 /*
924 * Managed interrupts are usually not migrated away
925 * from an online CPU, but CPU isolation 'managed_irq'
926 * can make that happen.
927 * 1) Activation does not take the isolation into account
928 * to keep the code simple
929 * 2) Migration away from an isolated CPU can happen when
930 * a non-isolated CPU which is in the calculated
931 * affinity mask comes online.
932 */
933 trace_vector_free_moved(apicd->irq, cpu, vector, managed);
934 irq_matrix_free(vector_matrix, cpu, vector, managed);
935 per_cpu(vector_irq, cpu)[vector] = VECTOR_UNUSED;
936 hlist_del_init(&apicd->clist);
937 apicd->prev_vector = 0;
938 apicd->move_in_progress = 0;
939}
940
941DEFINE_IDTENTRY_SYSVEC(sysvec_irq_move_cleanup)
942{
943 struct hlist_head *clhead = this_cpu_ptr(&cleanup_list);
944 struct apic_chip_data *apicd;
945 struct hlist_node *tmp;
946
947 ack_APIC_irq();
948 /* Prevent vectors vanishing under us */
949 raw_spin_lock(&vector_lock);
950
951 hlist_for_each_entry_safe(apicd, tmp, clhead, clist) {
952 unsigned int irr, vector = apicd->prev_vector;
953
954 /*
955 * Paranoia: Check if the vector that needs to be cleaned
956 * up is registered at the APICs IRR. If so, then this is
957 * not the best time to clean it up. Clean it up in the
958 * next attempt by sending another IRQ_MOVE_CLEANUP_VECTOR
959 * to this CPU. IRQ_MOVE_CLEANUP_VECTOR is the lowest
960 * priority external vector, so on return from this
961 * interrupt the device interrupt will happen first.
962 */
963 irr = apic_read(APIC_IRR + (vector / 32 * 0x10));
964 if (irr & (1U << (vector % 32))) {
965 apic->send_IPI_self(IRQ_MOVE_CLEANUP_VECTOR);
966 continue;
967 }
968 free_moved_vector(apicd);
969 }
970
971 raw_spin_unlock(&vector_lock);
972}
973
974static void __send_cleanup_vector(struct apic_chip_data *apicd)
975{
976 unsigned int cpu;
977
978 raw_spin_lock(&vector_lock);
979 apicd->move_in_progress = 0;
980 cpu = apicd->prev_cpu;
981 if (cpu_online(cpu)) {
982 hlist_add_head(&apicd->clist, per_cpu_ptr(&cleanup_list, cpu));
983 apic->send_IPI(cpu, IRQ_MOVE_CLEANUP_VECTOR);
984 } else {
985 apicd->prev_vector = 0;
986 }
987 raw_spin_unlock(&vector_lock);
988}
989
990void send_cleanup_vector(struct irq_cfg *cfg)
991{
992 struct apic_chip_data *apicd;
993
994 apicd = container_of(cfg, struct apic_chip_data, hw_irq_cfg);
995 if (apicd->move_in_progress)
996 __send_cleanup_vector(apicd);
997}
998
999void irq_complete_move(struct irq_cfg *cfg)
1000{
1001 struct apic_chip_data *apicd;
1002
1003 apicd = container_of(cfg, struct apic_chip_data, hw_irq_cfg);
1004 if (likely(!apicd->move_in_progress))
1005 return;
1006
1007 /*
1008 * If the interrupt arrived on the new target CPU, cleanup the
1009 * vector on the old target CPU. A vector check is not required
1010 * because an interrupt can never move from one vector to another
1011 * on the same CPU.
1012 */
1013 if (apicd->cpu == smp_processor_id())
1014 __send_cleanup_vector(apicd);
1015}
1016
1017/*
1018 * Called from fixup_irqs() with @desc->lock held and interrupts disabled.
1019 */
1020void irq_force_complete_move(struct irq_desc *desc)
1021{
1022 struct apic_chip_data *apicd;
1023 struct irq_data *irqd;
1024 unsigned int vector;
1025
1026 /*
1027 * The function is called for all descriptors regardless of which
1028 * irqdomain they belong to. For example if an IRQ is provided by
1029 * an irq_chip as part of a GPIO driver, the chip data for that
1030 * descriptor is specific to the irq_chip in question.
1031 *
1032 * Check first that the chip_data is what we expect
1033 * (apic_chip_data) before touching it any further.
1034 */
1035 irqd = irq_domain_get_irq_data(x86_vector_domain,
1036 irq_desc_get_irq(desc));
1037 if (!irqd)
1038 return;
1039
1040 raw_spin_lock(&vector_lock);
1041 apicd = apic_chip_data(irqd);
1042 if (!apicd)
1043 goto unlock;
1044
1045 /*
1046 * If prev_vector is empty, no action required.
1047 */
1048 vector = apicd->prev_vector;
1049 if (!vector)
1050 goto unlock;
1051
1052 /*
1053 * This is tricky. If the cleanup of the old vector has not been
1054 * done yet, then the following setaffinity call will fail with
1055 * -EBUSY. This can leave the interrupt in a stale state.
1056 *
1057 * All CPUs are stuck in stop machine with interrupts disabled so
1058 * calling __irq_complete_move() would be completely pointless.
1059 *
1060 * 1) The interrupt is in move_in_progress state. That means that we
1061 * have not seen an interrupt since the io_apic was reprogrammed to
1062 * the new vector.
1063 *
1064 * 2) The interrupt has fired on the new vector, but the cleanup IPIs
1065 * have not been processed yet.
1066 */
1067 if (apicd->move_in_progress) {
1068 /*
1069 * In theory there is a race:
1070 *
1071 * set_ioapic(new_vector) <-- Interrupt is raised before update
1072 * is effective, i.e. it's raised on
1073 * the old vector.
1074 *
1075 * So if the target cpu cannot handle that interrupt before
1076 * the old vector is cleaned up, we get a spurious interrupt
1077 * and in the worst case the ioapic irq line becomes stale.
1078 *
1079 * But in case of cpu hotplug this should be a non issue
1080 * because if the affinity update happens right before all
1081 * cpus rendezvous in stop machine, there is no way that the
1082 * interrupt can be blocked on the target cpu because all cpus
1083 * loops first with interrupts enabled in stop machine, so the
1084 * old vector is not yet cleaned up when the interrupt fires.
1085 *
1086 * So the only way to run into this issue is if the delivery
1087 * of the interrupt on the apic/system bus would be delayed
1088 * beyond the point where the target cpu disables interrupts
1089 * in stop machine. I doubt that it can happen, but at least
1090 * there is a theoretical chance. Virtualization might be
1091 * able to expose this, but AFAICT the IOAPIC emulation is not
1092 * as stupid as the real hardware.
1093 *
1094 * Anyway, there is nothing we can do about that at this point
1095 * w/o refactoring the whole fixup_irq() business completely.
1096 * We print at least the irq number and the old vector number,
1097 * so we have the necessary information when a problem in that
1098 * area arises.
1099 */
1100 pr_warn("IRQ fixup: irq %d move in progress, old vector %d\n",
1101 irqd->irq, vector);
1102 }
1103 free_moved_vector(apicd);
1104unlock:
1105 raw_spin_unlock(&vector_lock);
1106}
1107
1108#ifdef CONFIG_HOTPLUG_CPU
1109/*
1110 * Note, this is not accurate accounting, but at least good enough to
1111 * prevent that the actual interrupt move will run out of vectors.
1112 */
1113int lapic_can_unplug_cpu(void)
1114{
1115 unsigned int rsvd, avl, tomove, cpu = smp_processor_id();
1116 int ret = 0;
1117
1118 raw_spin_lock(&vector_lock);
1119 tomove = irq_matrix_allocated(vector_matrix);
1120 avl = irq_matrix_available(vector_matrix, true);
1121 if (avl < tomove) {
1122 pr_warn("CPU %u has %u vectors, %u available. Cannot disable CPU\n",
1123 cpu, tomove, avl);
1124 ret = -ENOSPC;
1125 goto out;
1126 }
1127 rsvd = irq_matrix_reserved(vector_matrix);
1128 if (avl < rsvd) {
1129 pr_warn("Reserved vectors %u > available %u. IRQ request may fail\n",
1130 rsvd, avl);
1131 }
1132out:
1133 raw_spin_unlock(&vector_lock);
1134 return ret;
1135}
1136#endif /* HOTPLUG_CPU */
1137#endif /* SMP */
1138
1139static void __init print_APIC_field(int base)
1140{
1141 int i;
1142
1143 printk(KERN_DEBUG);
1144
1145 for (i = 0; i < 8; i++)
1146 pr_cont("%08x", apic_read(base + i*0x10));
1147
1148 pr_cont("\n");
1149}
1150
1151static void __init print_local_APIC(void *dummy)
1152{
1153 unsigned int i, v, ver, maxlvt;
1154 u64 icr;
1155
1156 pr_debug("printing local APIC contents on CPU#%d/%d:\n",
1157 smp_processor_id(), hard_smp_processor_id());
1158 v = apic_read(APIC_ID);
1159 pr_info("... APIC ID: %08x (%01x)\n", v, read_apic_id());
1160 v = apic_read(APIC_LVR);
1161 pr_info("... APIC VERSION: %08x\n", v);
1162 ver = GET_APIC_VERSION(v);
1163 maxlvt = lapic_get_maxlvt();
1164
1165 v = apic_read(APIC_TASKPRI);
1166 pr_debug("... APIC TASKPRI: %08x (%02x)\n", v, v & APIC_TPRI_MASK);
1167
1168 /* !82489DX */
1169 if (APIC_INTEGRATED(ver)) {
1170 if (!APIC_XAPIC(ver)) {
1171 v = apic_read(APIC_ARBPRI);
1172 pr_debug("... APIC ARBPRI: %08x (%02x)\n",
1173 v, v & APIC_ARBPRI_MASK);
1174 }
1175 v = apic_read(APIC_PROCPRI);
1176 pr_debug("... APIC PROCPRI: %08x\n", v);
1177 }
1178
1179 /*
1180 * Remote read supported only in the 82489DX and local APIC for
1181 * Pentium processors.
1182 */
1183 if (!APIC_INTEGRATED(ver) || maxlvt == 3) {
1184 v = apic_read(APIC_RRR);
1185 pr_debug("... APIC RRR: %08x\n", v);
1186 }
1187
1188 v = apic_read(APIC_LDR);
1189 pr_debug("... APIC LDR: %08x\n", v);
1190 if (!x2apic_enabled()) {
1191 v = apic_read(APIC_DFR);
1192 pr_debug("... APIC DFR: %08x\n", v);
1193 }
1194 v = apic_read(APIC_SPIV);
1195 pr_debug("... APIC SPIV: %08x\n", v);
1196
1197 pr_debug("... APIC ISR field:\n");
1198 print_APIC_field(APIC_ISR);
1199 pr_debug("... APIC TMR field:\n");
1200 print_APIC_field(APIC_TMR);
1201 pr_debug("... APIC IRR field:\n");
1202 print_APIC_field(APIC_IRR);
1203
1204 /* !82489DX */
1205 if (APIC_INTEGRATED(ver)) {
1206 /* Due to the Pentium erratum 3AP. */
1207 if (maxlvt > 3)
1208 apic_write(APIC_ESR, 0);
1209
1210 v = apic_read(APIC_ESR);
1211 pr_debug("... APIC ESR: %08x\n", v);
1212 }
1213
1214 icr = apic_icr_read();
1215 pr_debug("... APIC ICR: %08x\n", (u32)icr);
1216 pr_debug("... APIC ICR2: %08x\n", (u32)(icr >> 32));
1217
1218 v = apic_read(APIC_LVTT);
1219 pr_debug("... APIC LVTT: %08x\n", v);
1220
1221 if (maxlvt > 3) {
1222 /* PC is LVT#4. */
1223 v = apic_read(APIC_LVTPC);
1224 pr_debug("... APIC LVTPC: %08x\n", v);
1225 }
1226 v = apic_read(APIC_LVT0);
1227 pr_debug("... APIC LVT0: %08x\n", v);
1228 v = apic_read(APIC_LVT1);
1229 pr_debug("... APIC LVT1: %08x\n", v);
1230
1231 if (maxlvt > 2) {
1232 /* ERR is LVT#3. */
1233 v = apic_read(APIC_LVTERR);
1234 pr_debug("... APIC LVTERR: %08x\n", v);
1235 }
1236
1237 v = apic_read(APIC_TMICT);
1238 pr_debug("... APIC TMICT: %08x\n", v);
1239 v = apic_read(APIC_TMCCT);
1240 pr_debug("... APIC TMCCT: %08x\n", v);
1241 v = apic_read(APIC_TDCR);
1242 pr_debug("... APIC TDCR: %08x\n", v);
1243
1244 if (boot_cpu_has(X86_FEATURE_EXTAPIC)) {
1245 v = apic_read(APIC_EFEAT);
1246 maxlvt = (v >> 16) & 0xff;
1247 pr_debug("... APIC EFEAT: %08x\n", v);
1248 v = apic_read(APIC_ECTRL);
1249 pr_debug("... APIC ECTRL: %08x\n", v);
1250 for (i = 0; i < maxlvt; i++) {
1251 v = apic_read(APIC_EILVTn(i));
1252 pr_debug("... APIC EILVT%d: %08x\n", i, v);
1253 }
1254 }
1255 pr_cont("\n");
1256}
1257
1258static void __init print_local_APICs(int maxcpu)
1259{
1260 int cpu;
1261
1262 if (!maxcpu)
1263 return;
1264
1265 preempt_disable();
1266 for_each_online_cpu(cpu) {
1267 if (cpu >= maxcpu)
1268 break;
1269 smp_call_function_single(cpu, print_local_APIC, NULL, 1);
1270 }
1271 preempt_enable();
1272}
1273
1274static void __init print_PIC(void)
1275{
1276 unsigned int v;
1277 unsigned long flags;
1278
1279 if (!nr_legacy_irqs())
1280 return;
1281
1282 pr_debug("\nprinting PIC contents\n");
1283
1284 raw_spin_lock_irqsave(&i8259A_lock, flags);
1285
1286 v = inb(0xa1) << 8 | inb(0x21);
1287 pr_debug("... PIC IMR: %04x\n", v);
1288
1289 v = inb(0xa0) << 8 | inb(0x20);
1290 pr_debug("... PIC IRR: %04x\n", v);
1291
1292 outb(0x0b, 0xa0);
1293 outb(0x0b, 0x20);
1294 v = inb(0xa0) << 8 | inb(0x20);
1295 outb(0x0a, 0xa0);
1296 outb(0x0a, 0x20);
1297
1298 raw_spin_unlock_irqrestore(&i8259A_lock, flags);
1299
1300 pr_debug("... PIC ISR: %04x\n", v);
1301
1302 v = inb(0x4d1) << 8 | inb(0x4d0);
1303 pr_debug("... PIC ELCR: %04x\n", v);
1304}
1305
1306static int show_lapic __initdata = 1;
1307static __init int setup_show_lapic(char *arg)
1308{
1309 int num = -1;
1310
1311 if (strcmp(arg, "all") == 0) {
1312 show_lapic = CONFIG_NR_CPUS;
1313 } else {
1314 get_option(&arg, &num);
1315 if (num >= 0)
1316 show_lapic = num;
1317 }
1318
1319 return 1;
1320}
1321__setup("show_lapic=", setup_show_lapic);
1322
1323static int __init print_ICs(void)
1324{
1325 if (apic_verbosity == APIC_QUIET)
1326 return 0;
1327
1328 print_PIC();
1329
1330 /* don't print out if apic is not there */
1331 if (!boot_cpu_has(X86_FEATURE_APIC) && !apic_from_smp_config())
1332 return 0;
1333
1334 print_local_APICs(show_lapic);
1335 print_IO_APICs();
1336
1337 return 0;
1338}
1339
1340late_initcall(print_ICs);