1 /*
2  * Local APIC related interfaces to support IOAPIC, MSI, HT_IRQ 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/init.h>
15 #include <linux/compiler.h>
16 #include <linux/slab.h>
17 #include <asm/irqdomain.h>
18 #include <asm/hw_irq.h>
19 #include <asm/apic.h>
20 #include <asm/i8259.h>
21 #include <asm/desc.h>
22 #include <asm/irq_remapping.h>
23 
24 struct apic_chip_data {
25 	struct irq_cfg		cfg;
26 	cpumask_var_t		domain;
27 	cpumask_var_t		old_domain;
28 	u8			move_in_progress : 1;
29 };
30 
31 struct irq_domain *x86_vector_domain;
32 static DEFINE_RAW_SPINLOCK(vector_lock);
33 static cpumask_var_t vector_cpumask, vector_searchmask, searched_cpumask;
34 static struct irq_chip lapic_controller;
35 #ifdef	CONFIG_X86_IO_APIC
36 static struct apic_chip_data *legacy_irq_data[NR_IRQS_LEGACY];
37 #endif
38 
lock_vector_lock(void)39 void lock_vector_lock(void)
40 {
41 	/* Used to the online set of cpus does not change
42 	 * during assign_irq_vector.
43 	 */
44 	raw_spin_lock(&vector_lock);
45 }
46 
unlock_vector_lock(void)47 void unlock_vector_lock(void)
48 {
49 	raw_spin_unlock(&vector_lock);
50 }
51 
apic_chip_data(struct irq_data * irq_data)52 static struct apic_chip_data *apic_chip_data(struct irq_data *irq_data)
53 {
54 	if (!irq_data)
55 		return NULL;
56 
57 	while (irq_data->parent_data)
58 		irq_data = irq_data->parent_data;
59 
60 	return irq_data->chip_data;
61 }
62 
irqd_cfg(struct irq_data * irq_data)63 struct irq_cfg *irqd_cfg(struct irq_data *irq_data)
64 {
65 	struct apic_chip_data *data = apic_chip_data(irq_data);
66 
67 	return data ? &data->cfg : NULL;
68 }
69 
irq_cfg(unsigned int irq)70 struct irq_cfg *irq_cfg(unsigned int irq)
71 {
72 	return irqd_cfg(irq_get_irq_data(irq));
73 }
74 
alloc_apic_chip_data(int node)75 static struct apic_chip_data *alloc_apic_chip_data(int node)
76 {
77 	struct apic_chip_data *data;
78 
79 	data = kzalloc_node(sizeof(*data), GFP_KERNEL, node);
80 	if (!data)
81 		return NULL;
82 	if (!zalloc_cpumask_var_node(&data->domain, GFP_KERNEL, node))
83 		goto out_data;
84 	if (!zalloc_cpumask_var_node(&data->old_domain, GFP_KERNEL, node))
85 		goto out_domain;
86 	return data;
87 out_domain:
88 	free_cpumask_var(data->domain);
89 out_data:
90 	kfree(data);
91 	return NULL;
92 }
93 
free_apic_chip_data(struct apic_chip_data * data)94 static void free_apic_chip_data(struct apic_chip_data *data)
95 {
96 	if (data) {
97 		free_cpumask_var(data->domain);
98 		free_cpumask_var(data->old_domain);
99 		kfree(data);
100 	}
101 }
102 
__assign_irq_vector(int irq,struct apic_chip_data * d,const struct cpumask * mask)103 static int __assign_irq_vector(int irq, struct apic_chip_data *d,
104 			       const struct cpumask *mask)
105 {
106 	/*
107 	 * NOTE! The local APIC isn't very good at handling
108 	 * multiple interrupts at the same interrupt level.
109 	 * As the interrupt level is determined by taking the
110 	 * vector number and shifting that right by 4, we
111 	 * want to spread these out a bit so that they don't
112 	 * all fall in the same interrupt level.
113 	 *
114 	 * Also, we've got to be careful not to trash gate
115 	 * 0x80, because int 0x80 is hm, kind of importantish. ;)
116 	 */
117 	static int current_vector = FIRST_EXTERNAL_VECTOR + VECTOR_OFFSET_START;
118 	static int current_offset = VECTOR_OFFSET_START % 16;
119 	int cpu, vector;
120 
121 	/*
122 	 * If there is still a move in progress or the previous move has not
123 	 * been cleaned up completely, tell the caller to come back later.
124 	 */
125 	if (d->move_in_progress ||
126 	    cpumask_intersects(d->old_domain, cpu_online_mask))
127 		return -EBUSY;
128 
129 	/* Only try and allocate irqs on cpus that are present */
130 	cpumask_clear(d->old_domain);
131 	cpumask_clear(searched_cpumask);
132 	cpu = cpumask_first_and(mask, cpu_online_mask);
133 	while (cpu < nr_cpu_ids) {
134 		int new_cpu, offset;
135 
136 		/* Get the possible target cpus for @mask/@cpu from the apic */
137 		apic->vector_allocation_domain(cpu, vector_cpumask, mask);
138 
139 		/*
140 		 * Clear the offline cpus from @vector_cpumask for searching
141 		 * and verify whether the result overlaps with @mask. If true,
142 		 * then the call to apic->cpu_mask_to_apicid_and() will
143 		 * succeed as well. If not, no point in trying to find a
144 		 * vector in this mask.
145 		 */
146 		cpumask_and(vector_searchmask, vector_cpumask, cpu_online_mask);
147 		if (!cpumask_intersects(vector_searchmask, mask))
148 			goto next_cpu;
149 
150 		if (cpumask_subset(vector_cpumask, d->domain)) {
151 			if (cpumask_equal(vector_cpumask, d->domain))
152 				goto success;
153 			/*
154 			 * Mark the cpus which are not longer in the mask for
155 			 * cleanup.
156 			 */
157 			cpumask_andnot(d->old_domain, d->domain, vector_cpumask);
158 			vector = d->cfg.vector;
159 			goto update;
160 		}
161 
162 		vector = current_vector;
163 		offset = current_offset;
164 next:
165 		vector += 16;
166 		if (vector >= first_system_vector) {
167 			offset = (offset + 1) % 16;
168 			vector = FIRST_EXTERNAL_VECTOR + offset;
169 		}
170 
171 		/* If the search wrapped around, try the next cpu */
172 		if (unlikely(current_vector == vector))
173 			goto next_cpu;
174 
175 		if (test_bit(vector, used_vectors))
176 			goto next;
177 
178 		for_each_cpu(new_cpu, vector_searchmask) {
179 			if (!IS_ERR_OR_NULL(per_cpu(vector_irq, new_cpu)[vector]))
180 				goto next;
181 		}
182 		/* Found one! */
183 		current_vector = vector;
184 		current_offset = offset;
185 		/* Schedule the old vector for cleanup on all cpus */
186 		if (d->cfg.vector)
187 			cpumask_copy(d->old_domain, d->domain);
188 		for_each_cpu(new_cpu, vector_searchmask)
189 			per_cpu(vector_irq, new_cpu)[vector] = irq_to_desc(irq);
190 		goto update;
191 
192 next_cpu:
193 		/*
194 		 * We exclude the current @vector_cpumask from the requested
195 		 * @mask and try again with the next online cpu in the
196 		 * result. We cannot modify @mask, so we use @vector_cpumask
197 		 * as a temporary buffer here as it will be reassigned when
198 		 * calling apic->vector_allocation_domain() above.
199 		 */
200 		cpumask_or(searched_cpumask, searched_cpumask, vector_cpumask);
201 		cpumask_andnot(vector_cpumask, mask, searched_cpumask);
202 		cpu = cpumask_first_and(vector_cpumask, cpu_online_mask);
203 		continue;
204 	}
205 	return -ENOSPC;
206 
207 update:
208 	/*
209 	 * Exclude offline cpus from the cleanup mask and set the
210 	 * move_in_progress flag when the result is not empty.
211 	 */
212 	cpumask_and(d->old_domain, d->old_domain, cpu_online_mask);
213 	d->move_in_progress = !cpumask_empty(d->old_domain);
214 	d->cfg.old_vector = d->move_in_progress ? d->cfg.vector : 0;
215 	d->cfg.vector = vector;
216 	cpumask_copy(d->domain, vector_cpumask);
217 success:
218 	/*
219 	 * Cache destination APIC IDs into cfg->dest_apicid. This cannot fail
220 	 * as we already established, that mask & d->domain & cpu_online_mask
221 	 * is not empty.
222 	 */
223 	BUG_ON(apic->cpu_mask_to_apicid_and(mask, d->domain,
224 					    &d->cfg.dest_apicid));
225 	return 0;
226 }
227 
assign_irq_vector(int irq,struct apic_chip_data * data,const struct cpumask * mask)228 static int assign_irq_vector(int irq, struct apic_chip_data *data,
229 			     const struct cpumask *mask)
230 {
231 	int err;
232 	unsigned long flags;
233 
234 	raw_spin_lock_irqsave(&vector_lock, flags);
235 	err = __assign_irq_vector(irq, data, mask);
236 	raw_spin_unlock_irqrestore(&vector_lock, flags);
237 	return err;
238 }
239 
assign_irq_vector_policy(int irq,int node,struct apic_chip_data * data,struct irq_alloc_info * info)240 static int assign_irq_vector_policy(int irq, int node,
241 				    struct apic_chip_data *data,
242 				    struct irq_alloc_info *info)
243 {
244 	if (info && info->mask)
245 		return assign_irq_vector(irq, data, info->mask);
246 	if (node != NUMA_NO_NODE &&
247 	    assign_irq_vector(irq, data, cpumask_of_node(node)) == 0)
248 		return 0;
249 	return assign_irq_vector(irq, data, apic->target_cpus());
250 }
251 
clear_irq_vector(int irq,struct apic_chip_data * data)252 static void clear_irq_vector(int irq, struct apic_chip_data *data)
253 {
254 	struct irq_desc *desc;
255 	int cpu, vector;
256 
257 	if (!data->cfg.vector)
258 		return;
259 
260 	vector = data->cfg.vector;
261 	for_each_cpu_and(cpu, data->domain, cpu_online_mask)
262 		per_cpu(vector_irq, cpu)[vector] = VECTOR_UNUSED;
263 
264 	data->cfg.vector = 0;
265 	cpumask_clear(data->domain);
266 
267 	/*
268 	 * If move is in progress or the old_domain mask is not empty,
269 	 * i.e. the cleanup IPI has not been processed yet, we need to remove
270 	 * the old references to desc from all cpus vector tables.
271 	 */
272 	if (!data->move_in_progress && cpumask_empty(data->old_domain))
273 		return;
274 
275 	desc = irq_to_desc(irq);
276 	for_each_cpu_and(cpu, data->old_domain, cpu_online_mask) {
277 		for (vector = FIRST_EXTERNAL_VECTOR; vector < NR_VECTORS;
278 		     vector++) {
279 			if (per_cpu(vector_irq, cpu)[vector] != desc)
280 				continue;
281 			per_cpu(vector_irq, cpu)[vector] = VECTOR_UNUSED;
282 			break;
283 		}
284 	}
285 	data->move_in_progress = 0;
286 }
287 
init_irq_alloc_info(struct irq_alloc_info * info,const struct cpumask * mask)288 void init_irq_alloc_info(struct irq_alloc_info *info,
289 			 const struct cpumask *mask)
290 {
291 	memset(info, 0, sizeof(*info));
292 	info->mask = mask;
293 }
294 
copy_irq_alloc_info(struct irq_alloc_info * dst,struct irq_alloc_info * src)295 void copy_irq_alloc_info(struct irq_alloc_info *dst, struct irq_alloc_info *src)
296 {
297 	if (src)
298 		*dst = *src;
299 	else
300 		memset(dst, 0, sizeof(*dst));
301 }
302 
x86_vector_free_irqs(struct irq_domain * domain,unsigned int virq,unsigned int nr_irqs)303 static void x86_vector_free_irqs(struct irq_domain *domain,
304 				 unsigned int virq, unsigned int nr_irqs)
305 {
306 	struct apic_chip_data *apic_data;
307 	struct irq_data *irq_data;
308 	unsigned long flags;
309 	int i;
310 
311 	for (i = 0; i < nr_irqs; i++) {
312 		irq_data = irq_domain_get_irq_data(x86_vector_domain, virq + i);
313 		if (irq_data && irq_data->chip_data) {
314 			raw_spin_lock_irqsave(&vector_lock, flags);
315 			clear_irq_vector(virq + i, irq_data->chip_data);
316 			apic_data = irq_data->chip_data;
317 			irq_domain_reset_irq_data(irq_data);
318 			raw_spin_unlock_irqrestore(&vector_lock, flags);
319 			free_apic_chip_data(apic_data);
320 #ifdef	CONFIG_X86_IO_APIC
321 			if (virq + i < nr_legacy_irqs())
322 				legacy_irq_data[virq + i] = NULL;
323 #endif
324 		}
325 	}
326 }
327 
x86_vector_alloc_irqs(struct irq_domain * domain,unsigned int virq,unsigned int nr_irqs,void * arg)328 static int x86_vector_alloc_irqs(struct irq_domain *domain, unsigned int virq,
329 				 unsigned int nr_irqs, void *arg)
330 {
331 	struct irq_alloc_info *info = arg;
332 	struct apic_chip_data *data;
333 	struct irq_data *irq_data;
334 	int i, err, node;
335 
336 	if (disable_apic)
337 		return -ENXIO;
338 
339 	/* Currently vector allocator can't guarantee contiguous allocations */
340 	if ((info->flags & X86_IRQ_ALLOC_CONTIGUOUS_VECTORS) && nr_irqs > 1)
341 		return -ENOSYS;
342 
343 	for (i = 0; i < nr_irqs; i++) {
344 		irq_data = irq_domain_get_irq_data(domain, virq + i);
345 		BUG_ON(!irq_data);
346 		node = irq_data_get_node(irq_data);
347 #ifdef	CONFIG_X86_IO_APIC
348 		if (virq + i < nr_legacy_irqs() && legacy_irq_data[virq + i])
349 			data = legacy_irq_data[virq + i];
350 		else
351 #endif
352 			data = alloc_apic_chip_data(node);
353 		if (!data) {
354 			err = -ENOMEM;
355 			goto error;
356 		}
357 
358 		irq_data->chip = &lapic_controller;
359 		irq_data->chip_data = data;
360 		irq_data->hwirq = virq + i;
361 		err = assign_irq_vector_policy(virq + i, node, data, info);
362 		if (err)
363 			goto error;
364 	}
365 
366 	return 0;
367 
368 error:
369 	x86_vector_free_irqs(domain, virq, i + 1);
370 	return err;
371 }
372 
373 static const struct irq_domain_ops x86_vector_domain_ops = {
374 	.alloc	= x86_vector_alloc_irqs,
375 	.free	= x86_vector_free_irqs,
376 };
377 
arch_probe_nr_irqs(void)378 int __init arch_probe_nr_irqs(void)
379 {
380 	int nr;
381 
382 	if (nr_irqs > (NR_VECTORS * nr_cpu_ids))
383 		nr_irqs = NR_VECTORS * nr_cpu_ids;
384 
385 	nr = (gsi_top + nr_legacy_irqs()) + 8 * nr_cpu_ids;
386 #if defined(CONFIG_PCI_MSI) || defined(CONFIG_HT_IRQ)
387 	/*
388 	 * for MSI and HT dyn irq
389 	 */
390 	if (gsi_top <= NR_IRQS_LEGACY)
391 		nr +=  8 * nr_cpu_ids;
392 	else
393 		nr += gsi_top * 16;
394 #endif
395 	if (nr < nr_irqs)
396 		nr_irqs = nr;
397 
398 	/*
399 	 * We don't know if PIC is present at this point so we need to do
400 	 * probe() to get the right number of legacy IRQs.
401 	 */
402 	return legacy_pic->probe();
403 }
404 
405 #ifdef	CONFIG_X86_IO_APIC
init_legacy_irqs(void)406 static void init_legacy_irqs(void)
407 {
408 	int i, node = cpu_to_node(0);
409 	struct apic_chip_data *data;
410 
411 	/*
412 	 * For legacy IRQ's, start with assigning irq0 to irq15 to
413 	 * ISA_IRQ_VECTOR(i) for all cpu's.
414 	 */
415 	for (i = 0; i < nr_legacy_irqs(); i++) {
416 		data = legacy_irq_data[i] = alloc_apic_chip_data(node);
417 		BUG_ON(!data);
418 
419 		data->cfg.vector = ISA_IRQ_VECTOR(i);
420 		cpumask_setall(data->domain);
421 		irq_set_chip_data(i, data);
422 	}
423 }
424 #else
init_legacy_irqs(void)425 static void init_legacy_irqs(void) { }
426 #endif
427 
arch_early_irq_init(void)428 int __init arch_early_irq_init(void)
429 {
430 	init_legacy_irqs();
431 
432 	x86_vector_domain = irq_domain_add_tree(NULL, &x86_vector_domain_ops,
433 						NULL);
434 	BUG_ON(x86_vector_domain == NULL);
435 	irq_set_default_host(x86_vector_domain);
436 
437 	arch_init_msi_domain(x86_vector_domain);
438 	arch_init_htirq_domain(x86_vector_domain);
439 
440 	BUG_ON(!alloc_cpumask_var(&vector_cpumask, GFP_KERNEL));
441 	BUG_ON(!alloc_cpumask_var(&vector_searchmask, GFP_KERNEL));
442 	BUG_ON(!alloc_cpumask_var(&searched_cpumask, GFP_KERNEL));
443 
444 	return arch_early_ioapic_init();
445 }
446 
447 /* Initialize vector_irq on a new cpu */
__setup_vector_irq(int cpu)448 static void __setup_vector_irq(int cpu)
449 {
450 	struct apic_chip_data *data;
451 	struct irq_desc *desc;
452 	int irq, vector;
453 
454 	/* Mark the inuse vectors */
455 	for_each_irq_desc(irq, desc) {
456 		struct irq_data *idata = irq_desc_get_irq_data(desc);
457 
458 		data = apic_chip_data(idata);
459 		if (!data || !cpumask_test_cpu(cpu, data->domain))
460 			continue;
461 		vector = data->cfg.vector;
462 		per_cpu(vector_irq, cpu)[vector] = desc;
463 	}
464 	/* Mark the free vectors */
465 	for (vector = 0; vector < NR_VECTORS; ++vector) {
466 		desc = per_cpu(vector_irq, cpu)[vector];
467 		if (IS_ERR_OR_NULL(desc))
468 			continue;
469 
470 		data = apic_chip_data(irq_desc_get_irq_data(desc));
471 		if (!cpumask_test_cpu(cpu, data->domain))
472 			per_cpu(vector_irq, cpu)[vector] = VECTOR_UNUSED;
473 	}
474 }
475 
476 /*
477  * Setup the vector to irq mappings. Must be called with vector_lock held.
478  */
setup_vector_irq(int cpu)479 void setup_vector_irq(int cpu)
480 {
481 	int irq;
482 
483 	lockdep_assert_held(&vector_lock);
484 	/*
485 	 * On most of the platforms, legacy PIC delivers the interrupts on the
486 	 * boot cpu. But there are certain platforms where PIC interrupts are
487 	 * delivered to multiple cpu's. If the legacy IRQ is handled by the
488 	 * legacy PIC, for the new cpu that is coming online, setup the static
489 	 * legacy vector to irq mapping:
490 	 */
491 	for (irq = 0; irq < nr_legacy_irqs(); irq++)
492 		per_cpu(vector_irq, cpu)[ISA_IRQ_VECTOR(irq)] = irq_to_desc(irq);
493 
494 	__setup_vector_irq(cpu);
495 }
496 
apic_retrigger_irq(struct irq_data * irq_data)497 static int apic_retrigger_irq(struct irq_data *irq_data)
498 {
499 	struct apic_chip_data *data = apic_chip_data(irq_data);
500 	unsigned long flags;
501 	int cpu;
502 
503 	raw_spin_lock_irqsave(&vector_lock, flags);
504 	cpu = cpumask_first_and(data->domain, cpu_online_mask);
505 	apic->send_IPI_mask(cpumask_of(cpu), data->cfg.vector);
506 	raw_spin_unlock_irqrestore(&vector_lock, flags);
507 
508 	return 1;
509 }
510 
apic_ack_edge(struct irq_data * data)511 void apic_ack_edge(struct irq_data *data)
512 {
513 	irq_complete_move(irqd_cfg(data));
514 	irq_move_irq(data);
515 	ack_APIC_irq();
516 }
517 
apic_set_affinity(struct irq_data * irq_data,const struct cpumask * dest,bool force)518 static int apic_set_affinity(struct irq_data *irq_data,
519 			     const struct cpumask *dest, bool force)
520 {
521 	struct apic_chip_data *data = irq_data->chip_data;
522 	int err, irq = irq_data->irq;
523 
524 	if (!config_enabled(CONFIG_SMP))
525 		return -EPERM;
526 
527 	if (!cpumask_intersects(dest, cpu_online_mask))
528 		return -EINVAL;
529 
530 	err = assign_irq_vector(irq, data, dest);
531 	return err ? err : IRQ_SET_MASK_OK;
532 }
533 
534 static struct irq_chip lapic_controller = {
535 	.irq_ack		= apic_ack_edge,
536 	.irq_set_affinity	= apic_set_affinity,
537 	.irq_retrigger		= apic_retrigger_irq,
538 };
539 
540 #ifdef CONFIG_SMP
__send_cleanup_vector(struct apic_chip_data * data)541 static void __send_cleanup_vector(struct apic_chip_data *data)
542 {
543 	raw_spin_lock(&vector_lock);
544 	cpumask_and(data->old_domain, data->old_domain, cpu_online_mask);
545 	data->move_in_progress = 0;
546 	if (!cpumask_empty(data->old_domain))
547 		apic->send_IPI_mask(data->old_domain, IRQ_MOVE_CLEANUP_VECTOR);
548 	raw_spin_unlock(&vector_lock);
549 }
550 
send_cleanup_vector(struct irq_cfg * cfg)551 void send_cleanup_vector(struct irq_cfg *cfg)
552 {
553 	struct apic_chip_data *data;
554 
555 	data = container_of(cfg, struct apic_chip_data, cfg);
556 	if (data->move_in_progress)
557 		__send_cleanup_vector(data);
558 }
559 
smp_irq_move_cleanup_interrupt(void)560 asmlinkage __visible void smp_irq_move_cleanup_interrupt(void)
561 {
562 	unsigned vector, me;
563 
564 	entering_ack_irq();
565 
566 	/* Prevent vectors vanishing under us */
567 	raw_spin_lock(&vector_lock);
568 
569 	me = smp_processor_id();
570 	for (vector = FIRST_EXTERNAL_VECTOR; vector < NR_VECTORS; vector++) {
571 		struct apic_chip_data *data;
572 		struct irq_desc *desc;
573 		unsigned int irr;
574 
575 	retry:
576 		desc = __this_cpu_read(vector_irq[vector]);
577 		if (IS_ERR_OR_NULL(desc))
578 			continue;
579 
580 		if (!raw_spin_trylock(&desc->lock)) {
581 			raw_spin_unlock(&vector_lock);
582 			cpu_relax();
583 			raw_spin_lock(&vector_lock);
584 			goto retry;
585 		}
586 
587 		data = apic_chip_data(irq_desc_get_irq_data(desc));
588 		if (!data)
589 			goto unlock;
590 
591 		/*
592 		 * Nothing to cleanup if irq migration is in progress
593 		 * or this cpu is not set in the cleanup mask.
594 		 */
595 		if (data->move_in_progress ||
596 		    !cpumask_test_cpu(me, data->old_domain))
597 			goto unlock;
598 
599 		/*
600 		 * We have two cases to handle here:
601 		 * 1) vector is unchanged but the target mask got reduced
602 		 * 2) vector and the target mask has changed
603 		 *
604 		 * #1 is obvious, but in #2 we have two vectors with the same
605 		 * irq descriptor: the old and the new vector. So we need to
606 		 * make sure that we only cleanup the old vector. The new
607 		 * vector has the current @vector number in the config and
608 		 * this cpu is part of the target mask. We better leave that
609 		 * one alone.
610 		 */
611 		if (vector == data->cfg.vector &&
612 		    cpumask_test_cpu(me, data->domain))
613 			goto unlock;
614 
615 		irr = apic_read(APIC_IRR + (vector / 32 * 0x10));
616 		/*
617 		 * Check if the vector that needs to be cleanedup is
618 		 * registered at the cpu's IRR. If so, then this is not
619 		 * the best time to clean it up. Lets clean it up in the
620 		 * next attempt by sending another IRQ_MOVE_CLEANUP_VECTOR
621 		 * to myself.
622 		 */
623 		if (irr  & (1 << (vector % 32))) {
624 			apic->send_IPI_self(IRQ_MOVE_CLEANUP_VECTOR);
625 			goto unlock;
626 		}
627 		__this_cpu_write(vector_irq[vector], VECTOR_UNUSED);
628 		cpumask_clear_cpu(me, data->old_domain);
629 unlock:
630 		raw_spin_unlock(&desc->lock);
631 	}
632 
633 	raw_spin_unlock(&vector_lock);
634 
635 	exiting_irq();
636 }
637 
__irq_complete_move(struct irq_cfg * cfg,unsigned vector)638 static void __irq_complete_move(struct irq_cfg *cfg, unsigned vector)
639 {
640 	unsigned me;
641 	struct apic_chip_data *data;
642 
643 	data = container_of(cfg, struct apic_chip_data, cfg);
644 	if (likely(!data->move_in_progress))
645 		return;
646 
647 	me = smp_processor_id();
648 	if (vector == data->cfg.vector && cpumask_test_cpu(me, data->domain))
649 		__send_cleanup_vector(data);
650 }
651 
irq_complete_move(struct irq_cfg * cfg)652 void irq_complete_move(struct irq_cfg *cfg)
653 {
654 	__irq_complete_move(cfg, ~get_irq_regs()->orig_ax);
655 }
656 
657 /*
658  * Called from fixup_irqs() with @desc->lock held and interrupts disabled.
659  */
irq_force_complete_move(struct irq_desc * desc)660 void irq_force_complete_move(struct irq_desc *desc)
661 {
662 	struct irq_data *irqdata = irq_desc_get_irq_data(desc);
663 	struct apic_chip_data *data = apic_chip_data(irqdata);
664 	struct irq_cfg *cfg = data ? &data->cfg : NULL;
665 	unsigned int cpu;
666 
667 	if (!cfg)
668 		return;
669 
670 	/*
671 	 * This is tricky. If the cleanup of @data->old_domain has not been
672 	 * done yet, then the following setaffinity call will fail with
673 	 * -EBUSY. This can leave the interrupt in a stale state.
674 	 *
675 	 * All CPUs are stuck in stop machine with interrupts disabled so
676 	 * calling __irq_complete_move() would be completely pointless.
677 	 */
678 	raw_spin_lock(&vector_lock);
679 	/*
680 	 * Clean out all offline cpus (including the outgoing one) from the
681 	 * old_domain mask.
682 	 */
683 	cpumask_and(data->old_domain, data->old_domain, cpu_online_mask);
684 
685 	/*
686 	 * If move_in_progress is cleared and the old_domain mask is empty,
687 	 * then there is nothing to cleanup. fixup_irqs() will take care of
688 	 * the stale vectors on the outgoing cpu.
689 	 */
690 	if (!data->move_in_progress && cpumask_empty(data->old_domain)) {
691 		raw_spin_unlock(&vector_lock);
692 		return;
693 	}
694 
695 	/*
696 	 * 1) The interrupt is in move_in_progress state. That means that we
697 	 *    have not seen an interrupt since the io_apic was reprogrammed to
698 	 *    the new vector.
699 	 *
700 	 * 2) The interrupt has fired on the new vector, but the cleanup IPIs
701 	 *    have not been processed yet.
702 	 */
703 	if (data->move_in_progress) {
704 		/*
705 		 * In theory there is a race:
706 		 *
707 		 * set_ioapic(new_vector) <-- Interrupt is raised before update
708 		 *			      is effective, i.e. it's raised on
709 		 *			      the old vector.
710 		 *
711 		 * So if the target cpu cannot handle that interrupt before
712 		 * the old vector is cleaned up, we get a spurious interrupt
713 		 * and in the worst case the ioapic irq line becomes stale.
714 		 *
715 		 * But in case of cpu hotplug this should be a non issue
716 		 * because if the affinity update happens right before all
717 		 * cpus rendevouz in stop machine, there is no way that the
718 		 * interrupt can be blocked on the target cpu because all cpus
719 		 * loops first with interrupts enabled in stop machine, so the
720 		 * old vector is not yet cleaned up when the interrupt fires.
721 		 *
722 		 * So the only way to run into this issue is if the delivery
723 		 * of the interrupt on the apic/system bus would be delayed
724 		 * beyond the point where the target cpu disables interrupts
725 		 * in stop machine. I doubt that it can happen, but at least
726 		 * there is a theroretical chance. Virtualization might be
727 		 * able to expose this, but AFAICT the IOAPIC emulation is not
728 		 * as stupid as the real hardware.
729 		 *
730 		 * Anyway, there is nothing we can do about that at this point
731 		 * w/o refactoring the whole fixup_irq() business completely.
732 		 * We print at least the irq number and the old vector number,
733 		 * so we have the necessary information when a problem in that
734 		 * area arises.
735 		 */
736 		pr_warn("IRQ fixup: irq %d move in progress, old vector %d\n",
737 			irqdata->irq, cfg->old_vector);
738 	}
739 	/*
740 	 * If old_domain is not empty, then other cpus still have the irq
741 	 * descriptor set in their vector array. Clean it up.
742 	 */
743 	for_each_cpu(cpu, data->old_domain)
744 		per_cpu(vector_irq, cpu)[cfg->old_vector] = VECTOR_UNUSED;
745 
746 	/* Cleanup the left overs of the (half finished) move */
747 	cpumask_clear(data->old_domain);
748 	data->move_in_progress = 0;
749 	raw_spin_unlock(&vector_lock);
750 }
751 #endif
752 
print_APIC_field(int base)753 static void __init print_APIC_field(int base)
754 {
755 	int i;
756 
757 	printk(KERN_DEBUG);
758 
759 	for (i = 0; i < 8; i++)
760 		pr_cont("%08x", apic_read(base + i*0x10));
761 
762 	pr_cont("\n");
763 }
764 
print_local_APIC(void * dummy)765 static void __init print_local_APIC(void *dummy)
766 {
767 	unsigned int i, v, ver, maxlvt;
768 	u64 icr;
769 
770 	pr_debug("printing local APIC contents on CPU#%d/%d:\n",
771 		 smp_processor_id(), hard_smp_processor_id());
772 	v = apic_read(APIC_ID);
773 	pr_info("... APIC ID:      %08x (%01x)\n", v, read_apic_id());
774 	v = apic_read(APIC_LVR);
775 	pr_info("... APIC VERSION: %08x\n", v);
776 	ver = GET_APIC_VERSION(v);
777 	maxlvt = lapic_get_maxlvt();
778 
779 	v = apic_read(APIC_TASKPRI);
780 	pr_debug("... APIC TASKPRI: %08x (%02x)\n", v, v & APIC_TPRI_MASK);
781 
782 	/* !82489DX */
783 	if (APIC_INTEGRATED(ver)) {
784 		if (!APIC_XAPIC(ver)) {
785 			v = apic_read(APIC_ARBPRI);
786 			pr_debug("... APIC ARBPRI: %08x (%02x)\n",
787 				 v, v & APIC_ARBPRI_MASK);
788 		}
789 		v = apic_read(APIC_PROCPRI);
790 		pr_debug("... APIC PROCPRI: %08x\n", v);
791 	}
792 
793 	/*
794 	 * Remote read supported only in the 82489DX and local APIC for
795 	 * Pentium processors.
796 	 */
797 	if (!APIC_INTEGRATED(ver) || maxlvt == 3) {
798 		v = apic_read(APIC_RRR);
799 		pr_debug("... APIC RRR: %08x\n", v);
800 	}
801 
802 	v = apic_read(APIC_LDR);
803 	pr_debug("... APIC LDR: %08x\n", v);
804 	if (!x2apic_enabled()) {
805 		v = apic_read(APIC_DFR);
806 		pr_debug("... APIC DFR: %08x\n", v);
807 	}
808 	v = apic_read(APIC_SPIV);
809 	pr_debug("... APIC SPIV: %08x\n", v);
810 
811 	pr_debug("... APIC ISR field:\n");
812 	print_APIC_field(APIC_ISR);
813 	pr_debug("... APIC TMR field:\n");
814 	print_APIC_field(APIC_TMR);
815 	pr_debug("... APIC IRR field:\n");
816 	print_APIC_field(APIC_IRR);
817 
818 	/* !82489DX */
819 	if (APIC_INTEGRATED(ver)) {
820 		/* Due to the Pentium erratum 3AP. */
821 		if (maxlvt > 3)
822 			apic_write(APIC_ESR, 0);
823 
824 		v = apic_read(APIC_ESR);
825 		pr_debug("... APIC ESR: %08x\n", v);
826 	}
827 
828 	icr = apic_icr_read();
829 	pr_debug("... APIC ICR: %08x\n", (u32)icr);
830 	pr_debug("... APIC ICR2: %08x\n", (u32)(icr >> 32));
831 
832 	v = apic_read(APIC_LVTT);
833 	pr_debug("... APIC LVTT: %08x\n", v);
834 
835 	if (maxlvt > 3) {
836 		/* PC is LVT#4. */
837 		v = apic_read(APIC_LVTPC);
838 		pr_debug("... APIC LVTPC: %08x\n", v);
839 	}
840 	v = apic_read(APIC_LVT0);
841 	pr_debug("... APIC LVT0: %08x\n", v);
842 	v = apic_read(APIC_LVT1);
843 	pr_debug("... APIC LVT1: %08x\n", v);
844 
845 	if (maxlvt > 2) {
846 		/* ERR is LVT#3. */
847 		v = apic_read(APIC_LVTERR);
848 		pr_debug("... APIC LVTERR: %08x\n", v);
849 	}
850 
851 	v = apic_read(APIC_TMICT);
852 	pr_debug("... APIC TMICT: %08x\n", v);
853 	v = apic_read(APIC_TMCCT);
854 	pr_debug("... APIC TMCCT: %08x\n", v);
855 	v = apic_read(APIC_TDCR);
856 	pr_debug("... APIC TDCR: %08x\n", v);
857 
858 	if (boot_cpu_has(X86_FEATURE_EXTAPIC)) {
859 		v = apic_read(APIC_EFEAT);
860 		maxlvt = (v >> 16) & 0xff;
861 		pr_debug("... APIC EFEAT: %08x\n", v);
862 		v = apic_read(APIC_ECTRL);
863 		pr_debug("... APIC ECTRL: %08x\n", v);
864 		for (i = 0; i < maxlvt; i++) {
865 			v = apic_read(APIC_EILVTn(i));
866 			pr_debug("... APIC EILVT%d: %08x\n", i, v);
867 		}
868 	}
869 	pr_cont("\n");
870 }
871 
print_local_APICs(int maxcpu)872 static void __init print_local_APICs(int maxcpu)
873 {
874 	int cpu;
875 
876 	if (!maxcpu)
877 		return;
878 
879 	preempt_disable();
880 	for_each_online_cpu(cpu) {
881 		if (cpu >= maxcpu)
882 			break;
883 		smp_call_function_single(cpu, print_local_APIC, NULL, 1);
884 	}
885 	preempt_enable();
886 }
887 
print_PIC(void)888 static void __init print_PIC(void)
889 {
890 	unsigned int v;
891 	unsigned long flags;
892 
893 	if (!nr_legacy_irqs())
894 		return;
895 
896 	pr_debug("\nprinting PIC contents\n");
897 
898 	raw_spin_lock_irqsave(&i8259A_lock, flags);
899 
900 	v = inb(0xa1) << 8 | inb(0x21);
901 	pr_debug("... PIC  IMR: %04x\n", v);
902 
903 	v = inb(0xa0) << 8 | inb(0x20);
904 	pr_debug("... PIC  IRR: %04x\n", v);
905 
906 	outb(0x0b, 0xa0);
907 	outb(0x0b, 0x20);
908 	v = inb(0xa0) << 8 | inb(0x20);
909 	outb(0x0a, 0xa0);
910 	outb(0x0a, 0x20);
911 
912 	raw_spin_unlock_irqrestore(&i8259A_lock, flags);
913 
914 	pr_debug("... PIC  ISR: %04x\n", v);
915 
916 	v = inb(0x4d1) << 8 | inb(0x4d0);
917 	pr_debug("... PIC ELCR: %04x\n", v);
918 }
919 
920 static int show_lapic __initdata = 1;
setup_show_lapic(char * arg)921 static __init int setup_show_lapic(char *arg)
922 {
923 	int num = -1;
924 
925 	if (strcmp(arg, "all") == 0) {
926 		show_lapic = CONFIG_NR_CPUS;
927 	} else {
928 		get_option(&arg, &num);
929 		if (num >= 0)
930 			show_lapic = num;
931 	}
932 
933 	return 1;
934 }
935 __setup("show_lapic=", setup_show_lapic);
936 
print_ICs(void)937 static int __init print_ICs(void)
938 {
939 	if (apic_verbosity == APIC_QUIET)
940 		return 0;
941 
942 	print_PIC();
943 
944 	/* don't print out if apic is not there */
945 	if (!cpu_has_apic && !apic_from_smp_config())
946 		return 0;
947 
948 	print_local_APICs(show_lapic);
949 	print_IO_APICs();
950 
951 	return 0;
952 }
953 
954 late_initcall(print_ICs);
955