1/*
2 * I/O SAPIC support.
3 *
4 * Copyright (C) 1999 Intel Corp.
5 * Copyright (C) 1999 Asit Mallick <asit.k.mallick@intel.com>
6 * Copyright (C) 2000-2002 J.I. Lee <jung-ik.lee@intel.com>
7 * Copyright (C) 1999-2000, 2002-2003 Hewlett-Packard Co.
8 *	David Mosberger-Tang <davidm@hpl.hp.com>
9 * Copyright (C) 1999 VA Linux Systems
10 * Copyright (C) 1999,2000 Walt Drummond <drummond@valinux.com>
11 *
12 * 00/04/19	D. Mosberger	Rewritten to mirror more closely the x86 I/O
13 *				APIC code.  In particular, we now have separate
14 *				handlers for edge and level triggered
15 *				interrupts.
16 * 00/10/27	Asit Mallick, Goutham Rao <goutham.rao@intel.com> IRQ vector
17 *				allocation PCI to vector mapping, shared PCI
18 *				interrupts.
19 * 00/10/27	D. Mosberger	Document things a bit more to make them more
20 *				understandable.  Clean up much of the old
21 *				IOSAPIC cruft.
22 * 01/07/27	J.I. Lee	PCI irq routing, Platform/Legacy interrupts
23 *				and fixes for ACPI S5(SoftOff) support.
24 * 02/01/23	J.I. Lee	iosapic pgm fixes for PCI irq routing from _PRT
25 * 02/01/07     E. Focht        <efocht@ess.nec.de> Redirectable interrupt
26 *				vectors in iosapic_set_affinity(),
27 *				initializations for /proc/irq/#/smp_affinity
28 * 02/04/02	P. Diefenbaugh	Cleaned up ACPI PCI IRQ routing.
29 * 02/04/18	J.I. Lee	bug fix in iosapic_init_pci_irq
30 * 02/04/30	J.I. Lee	bug fix in find_iosapic to fix ACPI PCI IRQ to
31 *				IOSAPIC mapping error
32 * 02/07/29	T. Kochi	Allocate interrupt vectors dynamically
33 * 02/08/04	T. Kochi	Cleaned up terminology (irq, global system
34 *				interrupt, vector, etc.)
35 * 02/09/20	D. Mosberger	Simplified by taking advantage of ACPI's
36 *				pci_irq code.
37 * 03/02/19	B. Helgaas	Make pcat_compat system-wide, not per-IOSAPIC.
38 *				Remove iosapic_address & gsi_base from
39 *				external interfaces.  Rationalize
40 *				__init/__devinit attributes.
41 * 04/12/04 Ashok Raj	<ashok.raj@intel.com> Intel Corporation 2004
42 *				Updated to work with irq migration necessary
43 *				for CPU Hotplug
44 */
45/*
46 * Here is what the interrupt logic between a PCI device and the kernel looks
47 * like:
48 *
49 * (1) A PCI device raises one of the four interrupt pins (INTA, INTB, INTC,
50 *     INTD).  The device is uniquely identified by its bus-, and slot-number
51 *     (the function number does not matter here because all functions share
52 *     the same interrupt lines).
53 *
54 * (2) The motherboard routes the interrupt line to a pin on a IOSAPIC
55 *     controller.  Multiple interrupt lines may have to share the same
56 *     IOSAPIC pin (if they're level triggered and use the same polarity).
57 *     Each interrupt line has a unique Global System Interrupt (GSI) number
58 *     which can be calculated as the sum of the controller's base GSI number
59 *     and the IOSAPIC pin number to which the line connects.
60 *
61 * (3) The IOSAPIC uses an internal routing table entries (RTEs) to map the
62 * IOSAPIC pin into the IA-64 interrupt vector.  This interrupt vector is then
63 * sent to the CPU.
64 *
65 * (4) The kernel recognizes an interrupt as an IRQ.  The IRQ interface is
66 *     used as architecture-independent interrupt handling mechanism in Linux.
67 *     As an IRQ is a number, we have to have
68 *     IA-64 interrupt vector number <-> IRQ number mapping.  On smaller
69 *     systems, we use one-to-one mapping between IA-64 vector and IRQ.  A
70 *     platform can implement platform_irq_to_vector(irq) and
71 *     platform_local_vector_to_irq(vector) APIs to differentiate the mapping.
72 *     Please see also arch/ia64/include/asm/hw_irq.h for those APIs.
73 *
74 * To sum up, there are three levels of mappings involved:
75 *
76 *	PCI pin -> global system interrupt (GSI) -> IA-64 vector <-> IRQ
77 *
78 * Note: The term "IRQ" is loosely used everywhere in Linux kernel to
79 * describe interrupts.  Now we use "IRQ" only for Linux IRQ's.  ISA IRQ
80 * (isa_irq) is the only exception in this source code.
81 */
82
83#include <linux/acpi.h>
84#include <linux/init.h>
85#include <linux/irq.h>
86#include <linux/kernel.h>
87#include <linux/list.h>
88#include <linux/pci.h>
89#include <linux/slab.h>
90#include <linux/smp.h>
91#include <linux/string.h>
92#include <linux/bootmem.h>
93
94#include <asm/delay.h>
95#include <asm/hw_irq.h>
96#include <asm/io.h>
97#include <asm/iosapic.h>
98#include <asm/machvec.h>
99#include <asm/processor.h>
100#include <asm/ptrace.h>
101
102#undef DEBUG_INTERRUPT_ROUTING
103
104#ifdef DEBUG_INTERRUPT_ROUTING
105#define DBG(fmt...)	printk(fmt)
106#else
107#define DBG(fmt...)
108#endif
109
110static DEFINE_SPINLOCK(iosapic_lock);
111
112/*
113 * These tables map IA-64 vectors to the IOSAPIC pin that generates this
114 * vector.
115 */
116
117#define NO_REF_RTE	0
118
119static struct iosapic {
120	char __iomem	*addr;		/* base address of IOSAPIC */
121	unsigned int	gsi_base;	/* GSI base */
122	unsigned short	num_rte;	/* # of RTEs on this IOSAPIC */
123	int		rtes_inuse;	/* # of RTEs in use on this IOSAPIC */
124#ifdef CONFIG_NUMA
125	unsigned short	node;		/* numa node association via pxm */
126#endif
127	spinlock_t	lock;		/* lock for indirect reg access */
128} iosapic_lists[NR_IOSAPICS];
129
130struct iosapic_rte_info {
131	struct list_head rte_list;	/* RTEs sharing the same vector */
132	char		rte_index;	/* IOSAPIC RTE index */
133	int		refcnt;		/* reference counter */
134	struct iosapic	*iosapic;
135} ____cacheline_aligned;
136
137static struct iosapic_intr_info {
138	struct list_head rtes;		/* RTEs using this vector (empty =>
139					 * not an IOSAPIC interrupt) */
140	int		count;		/* # of registered RTEs */
141	u32		low32;		/* current value of low word of
142					 * Redirection table entry */
143	unsigned int	dest;		/* destination CPU physical ID */
144	unsigned char	dmode	: 3;	/* delivery mode (see iosapic.h) */
145	unsigned char 	polarity: 1;	/* interrupt polarity
146					 * (see iosapic.h) */
147	unsigned char	trigger	: 1;	/* trigger mode (see iosapic.h) */
148} iosapic_intr_info[NR_IRQS];
149
150static unsigned char pcat_compat;	/* 8259 compatibility flag */
151
152static inline void
153iosapic_write(struct iosapic *iosapic, unsigned int reg, u32 val)
154{
155	unsigned long flags;
156
157	spin_lock_irqsave(&iosapic->lock, flags);
158	__iosapic_write(iosapic->addr, reg, val);
159	spin_unlock_irqrestore(&iosapic->lock, flags);
160}
161
162/*
163 * Find an IOSAPIC associated with a GSI
164 */
165static inline int
166find_iosapic (unsigned int gsi)
167{
168	int i;
169
170	for (i = 0; i < NR_IOSAPICS; i++) {
171		if ((unsigned) (gsi - iosapic_lists[i].gsi_base) <
172		    iosapic_lists[i].num_rte)
173			return i;
174	}
175
176	return -1;
177}
178
179static inline int __gsi_to_irq(unsigned int gsi)
180{
181	int irq;
182	struct iosapic_intr_info *info;
183	struct iosapic_rte_info *rte;
184
185	for (irq = 0; irq < NR_IRQS; irq++) {
186		info = &iosapic_intr_info[irq];
187		list_for_each_entry(rte, &info->rtes, rte_list)
188			if (rte->iosapic->gsi_base + rte->rte_index == gsi)
189				return irq;
190	}
191	return -1;
192}
193
194int
195gsi_to_irq (unsigned int gsi)
196{
197	unsigned long flags;
198	int irq;
199
200	spin_lock_irqsave(&iosapic_lock, flags);
201	irq = __gsi_to_irq(gsi);
202	spin_unlock_irqrestore(&iosapic_lock, flags);
203	return irq;
204}
205
206static struct iosapic_rte_info *find_rte(unsigned int irq, unsigned int gsi)
207{
208	struct iosapic_rte_info *rte;
209
210	list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
211		if (rte->iosapic->gsi_base + rte->rte_index == gsi)
212			return rte;
213	return NULL;
214}
215
216static void
217set_rte (unsigned int gsi, unsigned int irq, unsigned int dest, int mask)
218{
219	unsigned long pol, trigger, dmode;
220	u32 low32, high32;
221	int rte_index;
222	char redir;
223	struct iosapic_rte_info *rte;
224	ia64_vector vector = irq_to_vector(irq);
225
226	DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
227
228	rte = find_rte(irq, gsi);
229	if (!rte)
230		return;		/* not an IOSAPIC interrupt */
231
232	rte_index = rte->rte_index;
233	pol     = iosapic_intr_info[irq].polarity;
234	trigger = iosapic_intr_info[irq].trigger;
235	dmode   = iosapic_intr_info[irq].dmode;
236
237	redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
238
239#ifdef CONFIG_SMP
240	set_irq_affinity_info(irq, (int)(dest & 0xffff), redir);
241#endif
242
243	low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
244		 (trigger << IOSAPIC_TRIGGER_SHIFT) |
245		 (dmode << IOSAPIC_DELIVERY_SHIFT) |
246		 ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
247		 vector);
248
249	/* dest contains both id and eid */
250	high32 = (dest << IOSAPIC_DEST_SHIFT);
251
252	iosapic_write(rte->iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
253	iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
254	iosapic_intr_info[irq].low32 = low32;
255	iosapic_intr_info[irq].dest = dest;
256}
257
258static void
259nop (struct irq_data *data)
260{
261	/* do nothing... */
262}
263
264
265#ifdef CONFIG_KEXEC
266void
267kexec_disable_iosapic(void)
268{
269	struct iosapic_intr_info *info;
270	struct iosapic_rte_info *rte;
271	ia64_vector vec;
272	int irq;
273
274	for (irq = 0; irq < NR_IRQS; irq++) {
275		info = &iosapic_intr_info[irq];
276		vec = irq_to_vector(irq);
277		list_for_each_entry(rte, &info->rtes,
278				rte_list) {
279			iosapic_write(rte->iosapic,
280					IOSAPIC_RTE_LOW(rte->rte_index),
281					IOSAPIC_MASK|vec);
282			iosapic_eoi(rte->iosapic->addr, vec);
283		}
284	}
285}
286#endif
287
288static void
289mask_irq (struct irq_data *data)
290{
291	unsigned int irq = data->irq;
292	u32 low32;
293	int rte_index;
294	struct iosapic_rte_info *rte;
295
296	if (!iosapic_intr_info[irq].count)
297		return;			/* not an IOSAPIC interrupt! */
298
299	/* set only the mask bit */
300	low32 = iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
301	list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
302		rte_index = rte->rte_index;
303		iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
304	}
305}
306
307static void
308unmask_irq (struct irq_data *data)
309{
310	unsigned int irq = data->irq;
311	u32 low32;
312	int rte_index;
313	struct iosapic_rte_info *rte;
314
315	if (!iosapic_intr_info[irq].count)
316		return;			/* not an IOSAPIC interrupt! */
317
318	low32 = iosapic_intr_info[irq].low32 &= ~IOSAPIC_MASK;
319	list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
320		rte_index = rte->rte_index;
321		iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
322	}
323}
324
325
326static int
327iosapic_set_affinity(struct irq_data *data, const struct cpumask *mask,
328		     bool force)
329{
330#ifdef CONFIG_SMP
331	unsigned int irq = data->irq;
332	u32 high32, low32;
333	int cpu, dest, rte_index;
334	int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
335	struct iosapic_rte_info *rte;
336	struct iosapic *iosapic;
337
338	irq &= (~IA64_IRQ_REDIRECTED);
339
340	cpu = cpumask_first_and(cpu_online_mask, mask);
341	if (cpu >= nr_cpu_ids)
342		return -1;
343
344	if (irq_prepare_move(irq, cpu))
345		return -1;
346
347	dest = cpu_physical_id(cpu);
348
349	if (!iosapic_intr_info[irq].count)
350		return -1;			/* not an IOSAPIC interrupt */
351
352	set_irq_affinity_info(irq, dest, redir);
353
354	/* dest contains both id and eid */
355	high32 = dest << IOSAPIC_DEST_SHIFT;
356
357	low32 = iosapic_intr_info[irq].low32 & ~(7 << IOSAPIC_DELIVERY_SHIFT);
358	if (redir)
359		/* change delivery mode to lowest priority */
360		low32 |= (IOSAPIC_LOWEST_PRIORITY << IOSAPIC_DELIVERY_SHIFT);
361	else
362		/* change delivery mode to fixed */
363		low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
364	low32 &= IOSAPIC_VECTOR_MASK;
365	low32 |= irq_to_vector(irq);
366
367	iosapic_intr_info[irq].low32 = low32;
368	iosapic_intr_info[irq].dest = dest;
369	list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
370		iosapic = rte->iosapic;
371		rte_index = rte->rte_index;
372		iosapic_write(iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
373		iosapic_write(iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
374	}
375
376#endif
377	return 0;
378}
379
380/*
381 * Handlers for level-triggered interrupts.
382 */
383
384static unsigned int
385iosapic_startup_level_irq (struct irq_data *data)
386{
387	unmask_irq(data);
388	return 0;
389}
390
391static void
392iosapic_unmask_level_irq (struct irq_data *data)
393{
394	unsigned int irq = data->irq;
395	ia64_vector vec = irq_to_vector(irq);
396	struct iosapic_rte_info *rte;
397	int do_unmask_irq = 0;
398
399	irq_complete_move(irq);
400	if (unlikely(irqd_is_setaffinity_pending(data))) {
401		do_unmask_irq = 1;
402		mask_irq(data);
403	} else
404		unmask_irq(data);
405
406	list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
407		iosapic_eoi(rte->iosapic->addr, vec);
408
409	if (unlikely(do_unmask_irq)) {
410		irq_move_masked_irq(data);
411		unmask_irq(data);
412	}
413}
414
415#define iosapic_shutdown_level_irq	mask_irq
416#define iosapic_enable_level_irq	unmask_irq
417#define iosapic_disable_level_irq	mask_irq
418#define iosapic_ack_level_irq		nop
419
420static struct irq_chip irq_type_iosapic_level = {
421	.name =			"IO-SAPIC-level",
422	.irq_startup =		iosapic_startup_level_irq,
423	.irq_shutdown =		iosapic_shutdown_level_irq,
424	.irq_enable =		iosapic_enable_level_irq,
425	.irq_disable =		iosapic_disable_level_irq,
426	.irq_ack =		iosapic_ack_level_irq,
427	.irq_mask =		mask_irq,
428	.irq_unmask =		iosapic_unmask_level_irq,
429	.irq_set_affinity =	iosapic_set_affinity
430};
431
432/*
433 * Handlers for edge-triggered interrupts.
434 */
435
436static unsigned int
437iosapic_startup_edge_irq (struct irq_data *data)
438{
439	unmask_irq(data);
440	/*
441	 * IOSAPIC simply drops interrupts pended while the
442	 * corresponding pin was masked, so we can't know if an
443	 * interrupt is pending already.  Let's hope not...
444	 */
445	return 0;
446}
447
448static void
449iosapic_ack_edge_irq (struct irq_data *data)
450{
451	irq_complete_move(data->irq);
452	irq_move_irq(data);
453}
454
455#define iosapic_enable_edge_irq		unmask_irq
456#define iosapic_disable_edge_irq	nop
457
458static struct irq_chip irq_type_iosapic_edge = {
459	.name =			"IO-SAPIC-edge",
460	.irq_startup =		iosapic_startup_edge_irq,
461	.irq_shutdown =		iosapic_disable_edge_irq,
462	.irq_enable =		iosapic_enable_edge_irq,
463	.irq_disable =		iosapic_disable_edge_irq,
464	.irq_ack =		iosapic_ack_edge_irq,
465	.irq_mask =		mask_irq,
466	.irq_unmask =		unmask_irq,
467	.irq_set_affinity =	iosapic_set_affinity
468};
469
470static unsigned int
471iosapic_version (char __iomem *addr)
472{
473	/*
474	 * IOSAPIC Version Register return 32 bit structure like:
475	 * {
476	 *	unsigned int version   : 8;
477	 *	unsigned int reserved1 : 8;
478	 *	unsigned int max_redir : 8;
479	 *	unsigned int reserved2 : 8;
480	 * }
481	 */
482	return __iosapic_read(addr, IOSAPIC_VERSION);
483}
484
485static int iosapic_find_sharable_irq(unsigned long trigger, unsigned long pol)
486{
487	int i, irq = -ENOSPC, min_count = -1;
488	struct iosapic_intr_info *info;
489
490	/*
491	 * shared vectors for edge-triggered interrupts are not
492	 * supported yet
493	 */
494	if (trigger == IOSAPIC_EDGE)
495		return -EINVAL;
496
497	for (i = 0; i < NR_IRQS; i++) {
498		info = &iosapic_intr_info[i];
499		if (info->trigger == trigger && info->polarity == pol &&
500		    (info->dmode == IOSAPIC_FIXED ||
501		     info->dmode == IOSAPIC_LOWEST_PRIORITY) &&
502		    can_request_irq(i, IRQF_SHARED)) {
503			if (min_count == -1 || info->count < min_count) {
504				irq = i;
505				min_count = info->count;
506			}
507		}
508	}
509	return irq;
510}
511
512/*
513 * if the given vector is already owned by other,
514 *  assign a new vector for the other and make the vector available
515 */
516static void __init
517iosapic_reassign_vector (int irq)
518{
519	int new_irq;
520
521	if (iosapic_intr_info[irq].count) {
522		new_irq = create_irq();
523		if (new_irq < 0)
524			panic("%s: out of interrupt vectors!\n", __func__);
525		printk(KERN_INFO "Reassigning vector %d to %d\n",
526		       irq_to_vector(irq), irq_to_vector(new_irq));
527		memcpy(&iosapic_intr_info[new_irq], &iosapic_intr_info[irq],
528		       sizeof(struct iosapic_intr_info));
529		INIT_LIST_HEAD(&iosapic_intr_info[new_irq].rtes);
530		list_move(iosapic_intr_info[irq].rtes.next,
531			  &iosapic_intr_info[new_irq].rtes);
532		memset(&iosapic_intr_info[irq], 0,
533		       sizeof(struct iosapic_intr_info));
534		iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
535		INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
536	}
537}
538
539static inline int irq_is_shared (int irq)
540{
541	return (iosapic_intr_info[irq].count > 1);
542}
543
544struct irq_chip*
545ia64_native_iosapic_get_irq_chip(unsigned long trigger)
546{
547	if (trigger == IOSAPIC_EDGE)
548		return &irq_type_iosapic_edge;
549	else
550		return &irq_type_iosapic_level;
551}
552
553static int
554register_intr (unsigned int gsi, int irq, unsigned char delivery,
555	       unsigned long polarity, unsigned long trigger)
556{
557	struct irq_chip *chip, *irq_type;
558	int index;
559	struct iosapic_rte_info *rte;
560
561	index = find_iosapic(gsi);
562	if (index < 0) {
563		printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
564		       __func__, gsi);
565		return -ENODEV;
566	}
567
568	rte = find_rte(irq, gsi);
569	if (!rte) {
570		rte = kzalloc(sizeof (*rte), GFP_ATOMIC);
571		if (!rte) {
572			printk(KERN_WARNING "%s: cannot allocate memory\n",
573			       __func__);
574			return -ENOMEM;
575		}
576
577		rte->iosapic	= &iosapic_lists[index];
578		rte->rte_index	= gsi - rte->iosapic->gsi_base;
579		rte->refcnt++;
580		list_add_tail(&rte->rte_list, &iosapic_intr_info[irq].rtes);
581		iosapic_intr_info[irq].count++;
582		iosapic_lists[index].rtes_inuse++;
583	}
584	else if (rte->refcnt == NO_REF_RTE) {
585		struct iosapic_intr_info *info = &iosapic_intr_info[irq];
586		if (info->count > 0 &&
587		    (info->trigger != trigger || info->polarity != polarity)){
588			printk (KERN_WARNING
589				"%s: cannot override the interrupt\n",
590				__func__);
591			return -EINVAL;
592		}
593		rte->refcnt++;
594		iosapic_intr_info[irq].count++;
595		iosapic_lists[index].rtes_inuse++;
596	}
597
598	iosapic_intr_info[irq].polarity = polarity;
599	iosapic_intr_info[irq].dmode    = delivery;
600	iosapic_intr_info[irq].trigger  = trigger;
601
602	irq_type = iosapic_get_irq_chip(trigger);
603
604	chip = irq_get_chip(irq);
605	if (irq_type != NULL && chip != irq_type) {
606		if (chip != &no_irq_chip)
607			printk(KERN_WARNING
608			       "%s: changing vector %d from %s to %s\n",
609			       __func__, irq_to_vector(irq),
610			       chip->name, irq_type->name);
611		chip = irq_type;
612	}
613	__irq_set_chip_handler_name_locked(irq, chip, trigger == IOSAPIC_EDGE ?
614					   handle_edge_irq : handle_level_irq,
615					   NULL);
616	return 0;
617}
618
619static unsigned int
620get_target_cpu (unsigned int gsi, int irq)
621{
622#ifdef CONFIG_SMP
623	static int cpu = -1;
624	extern int cpe_vector;
625	cpumask_t domain = irq_to_domain(irq);
626
627	/*
628	 * In case of vector shared by multiple RTEs, all RTEs that
629	 * share the vector need to use the same destination CPU.
630	 */
631	if (iosapic_intr_info[irq].count)
632		return iosapic_intr_info[irq].dest;
633
634	/*
635	 * If the platform supports redirection via XTP, let it
636	 * distribute interrupts.
637	 */
638	if (smp_int_redirect & SMP_IRQ_REDIRECTION)
639		return cpu_physical_id(smp_processor_id());
640
641	/*
642	 * Some interrupts (ACPI SCI, for instance) are registered
643	 * before the BSP is marked as online.
644	 */
645	if (!cpu_online(smp_processor_id()))
646		return cpu_physical_id(smp_processor_id());
647
648#ifdef CONFIG_ACPI
649	if (cpe_vector > 0 && irq_to_vector(irq) == IA64_CPEP_VECTOR)
650		return get_cpei_target_cpu();
651#endif
652
653#ifdef CONFIG_NUMA
654	{
655		int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
656		const struct cpumask *cpu_mask;
657
658		iosapic_index = find_iosapic(gsi);
659		if (iosapic_index < 0 ||
660		    iosapic_lists[iosapic_index].node == MAX_NUMNODES)
661			goto skip_numa_setup;
662
663		cpu_mask = cpumask_of_node(iosapic_lists[iosapic_index].node);
664		num_cpus = 0;
665		for_each_cpu_and(numa_cpu, cpu_mask, &domain) {
666			if (cpu_online(numa_cpu))
667				num_cpus++;
668		}
669
670		if (!num_cpus)
671			goto skip_numa_setup;
672
673		/* Use irq assignment to distribute across cpus in node */
674		cpu_index = irq % num_cpus;
675
676		for_each_cpu_and(numa_cpu, cpu_mask, &domain)
677			if (cpu_online(numa_cpu) && i++ >= cpu_index)
678				break;
679
680		if (numa_cpu < nr_cpu_ids)
681			return cpu_physical_id(numa_cpu);
682	}
683skip_numa_setup:
684#endif
685	/*
686	 * Otherwise, round-robin interrupt vectors across all the
687	 * processors.  (It'd be nice if we could be smarter in the
688	 * case of NUMA.)
689	 */
690	do {
691		if (++cpu >= nr_cpu_ids)
692			cpu = 0;
693	} while (!cpu_online(cpu) || !cpumask_test_cpu(cpu, &domain));
694
695	return cpu_physical_id(cpu);
696#else  /* CONFIG_SMP */
697	return cpu_physical_id(smp_processor_id());
698#endif
699}
700
701static inline unsigned char choose_dmode(void)
702{
703#ifdef CONFIG_SMP
704	if (smp_int_redirect & SMP_IRQ_REDIRECTION)
705		return IOSAPIC_LOWEST_PRIORITY;
706#endif
707	return IOSAPIC_FIXED;
708}
709
710/*
711 * ACPI can describe IOSAPIC interrupts via static tables and namespace
712 * methods.  This provides an interface to register those interrupts and
713 * program the IOSAPIC RTE.
714 */
715int
716iosapic_register_intr (unsigned int gsi,
717		       unsigned long polarity, unsigned long trigger)
718{
719	int irq, mask = 1, err;
720	unsigned int dest;
721	unsigned long flags;
722	struct iosapic_rte_info *rte;
723	u32 low32;
724	unsigned char dmode;
725	struct irq_desc *desc;
726
727	/*
728	 * If this GSI has already been registered (i.e., it's a
729	 * shared interrupt, or we lost a race to register it),
730	 * don't touch the RTE.
731	 */
732	spin_lock_irqsave(&iosapic_lock, flags);
733	irq = __gsi_to_irq(gsi);
734	if (irq > 0) {
735		rte = find_rte(irq, gsi);
736		if(iosapic_intr_info[irq].count == 0) {
737			assign_irq_vector(irq);
738			irq_init_desc(irq);
739		} else if (rte->refcnt != NO_REF_RTE) {
740			rte->refcnt++;
741			goto unlock_iosapic_lock;
742		}
743	} else
744		irq = create_irq();
745
746	/* If vector is running out, we try to find a sharable vector */
747	if (irq < 0) {
748		irq = iosapic_find_sharable_irq(trigger, polarity);
749		if (irq < 0)
750			goto unlock_iosapic_lock;
751	}
752
753	desc = irq_to_desc(irq);
754	raw_spin_lock(&desc->lock);
755	dest = get_target_cpu(gsi, irq);
756	dmode = choose_dmode();
757	err = register_intr(gsi, irq, dmode, polarity, trigger);
758	if (err < 0) {
759		raw_spin_unlock(&desc->lock);
760		irq = err;
761		goto unlock_iosapic_lock;
762	}
763
764	/*
765	 * If the vector is shared and already unmasked for other
766	 * interrupt sources, don't mask it.
767	 */
768	low32 = iosapic_intr_info[irq].low32;
769	if (irq_is_shared(irq) && !(low32 & IOSAPIC_MASK))
770		mask = 0;
771	set_rte(gsi, irq, dest, mask);
772
773	printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
774	       gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
775	       (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
776	       cpu_logical_id(dest), dest, irq_to_vector(irq));
777
778	raw_spin_unlock(&desc->lock);
779 unlock_iosapic_lock:
780	spin_unlock_irqrestore(&iosapic_lock, flags);
781	return irq;
782}
783
784void
785iosapic_unregister_intr (unsigned int gsi)
786{
787	unsigned long flags;
788	int irq, index;
789	u32 low32;
790	unsigned long trigger, polarity;
791	unsigned int dest;
792	struct iosapic_rte_info *rte;
793
794	/*
795	 * If the irq associated with the gsi is not found,
796	 * iosapic_unregister_intr() is unbalanced. We need to check
797	 * this again after getting locks.
798	 */
799	irq = gsi_to_irq(gsi);
800	if (irq < 0) {
801		printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
802		       gsi);
803		WARN_ON(1);
804		return;
805	}
806
807	spin_lock_irqsave(&iosapic_lock, flags);
808	if ((rte = find_rte(irq, gsi)) == NULL) {
809		printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
810		       gsi);
811		WARN_ON(1);
812		goto out;
813	}
814
815	if (--rte->refcnt > 0)
816		goto out;
817
818	rte->refcnt = NO_REF_RTE;
819
820	/* Mask the interrupt */
821	low32 = iosapic_intr_info[irq].low32 | IOSAPIC_MASK;
822	iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte->rte_index), low32);
823
824	iosapic_intr_info[irq].count--;
825	index = find_iosapic(gsi);
826	iosapic_lists[index].rtes_inuse--;
827	WARN_ON(iosapic_lists[index].rtes_inuse < 0);
828
829	trigger  = iosapic_intr_info[irq].trigger;
830	polarity = iosapic_intr_info[irq].polarity;
831	dest     = iosapic_intr_info[irq].dest;
832	printk(KERN_INFO
833	       "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d unregistered\n",
834	       gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
835	       (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
836	       cpu_logical_id(dest), dest, irq_to_vector(irq));
837
838	if (iosapic_intr_info[irq].count == 0) {
839#ifdef CONFIG_SMP
840		/* Clear affinity */
841		cpumask_setall(irq_get_irq_data(irq)->affinity);
842#endif
843		/* Clear the interrupt information */
844		iosapic_intr_info[irq].dest = 0;
845		iosapic_intr_info[irq].dmode = 0;
846		iosapic_intr_info[irq].polarity = 0;
847		iosapic_intr_info[irq].trigger = 0;
848		iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
849
850		/* Destroy and reserve IRQ */
851		destroy_and_reserve_irq(irq);
852	}
853 out:
854	spin_unlock_irqrestore(&iosapic_lock, flags);
855}
856
857/*
858 * ACPI calls this when it finds an entry for a platform interrupt.
859 */
860int __init
861iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
862				int iosapic_vector, u16 eid, u16 id,
863				unsigned long polarity, unsigned long trigger)
864{
865	static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
866	unsigned char delivery;
867	int irq, vector, mask = 0;
868	unsigned int dest = ((id << 8) | eid) & 0xffff;
869
870	switch (int_type) {
871	      case ACPI_INTERRUPT_PMI:
872		irq = vector = iosapic_vector;
873		bind_irq_vector(irq, vector, CPU_MASK_ALL);
874		/*
875		 * since PMI vector is alloc'd by FW(ACPI) not by kernel,
876		 * we need to make sure the vector is available
877		 */
878		iosapic_reassign_vector(irq);
879		delivery = IOSAPIC_PMI;
880		break;
881	      case ACPI_INTERRUPT_INIT:
882		irq = create_irq();
883		if (irq < 0)
884			panic("%s: out of interrupt vectors!\n", __func__);
885		vector = irq_to_vector(irq);
886		delivery = IOSAPIC_INIT;
887		break;
888	      case ACPI_INTERRUPT_CPEI:
889		irq = vector = IA64_CPE_VECTOR;
890		BUG_ON(bind_irq_vector(irq, vector, CPU_MASK_ALL));
891		delivery = IOSAPIC_FIXED;
892		mask = 1;
893		break;
894	      default:
895		printk(KERN_ERR "%s: invalid int type 0x%x\n", __func__,
896		       int_type);
897		return -1;
898	}
899
900	register_intr(gsi, irq, delivery, polarity, trigger);
901
902	printk(KERN_INFO
903	       "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x)"
904	       " vector %d\n",
905	       int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
906	       int_type, gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
907	       (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
908	       cpu_logical_id(dest), dest, vector);
909
910	set_rte(gsi, irq, dest, mask);
911	return vector;
912}
913
914/*
915 * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
916 */
917void iosapic_override_isa_irq(unsigned int isa_irq, unsigned int gsi,
918			      unsigned long polarity, unsigned long trigger)
919{
920	int vector, irq;
921	unsigned int dest = cpu_physical_id(smp_processor_id());
922	unsigned char dmode;
923
924	irq = vector = isa_irq_to_vector(isa_irq);
925	BUG_ON(bind_irq_vector(irq, vector, CPU_MASK_ALL));
926	dmode = choose_dmode();
927	register_intr(gsi, irq, dmode, polarity, trigger);
928
929	DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
930	    isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
931	    polarity == IOSAPIC_POL_HIGH ? "high" : "low",
932	    cpu_logical_id(dest), dest, vector);
933
934	set_rte(gsi, irq, dest, 1);
935}
936
937void __init
938ia64_native_iosapic_pcat_compat_init(void)
939{
940	if (pcat_compat) {
941		/*
942		 * Disable the compatibility mode interrupts (8259 style),
943		 * needs IN/OUT support enabled.
944		 */
945		printk(KERN_INFO
946		       "%s: Disabling PC-AT compatible 8259 interrupts\n",
947		       __func__);
948		outb(0xff, 0xA1);
949		outb(0xff, 0x21);
950	}
951}
952
953void __init
954iosapic_system_init (int system_pcat_compat)
955{
956	int irq;
957
958	for (irq = 0; irq < NR_IRQS; ++irq) {
959		iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
960		/* mark as unused */
961		INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
962
963		iosapic_intr_info[irq].count = 0;
964	}
965
966	pcat_compat = system_pcat_compat;
967	if (pcat_compat)
968		iosapic_pcat_compat_init();
969}
970
971static inline int
972iosapic_alloc (void)
973{
974	int index;
975
976	for (index = 0; index < NR_IOSAPICS; index++)
977		if (!iosapic_lists[index].addr)
978			return index;
979
980	printk(KERN_WARNING "%s: failed to allocate iosapic\n", __func__);
981	return -1;
982}
983
984static inline void
985iosapic_free (int index)
986{
987	memset(&iosapic_lists[index], 0, sizeof(iosapic_lists[0]));
988}
989
990static inline int
991iosapic_check_gsi_range (unsigned int gsi_base, unsigned int ver)
992{
993	int index;
994	unsigned int gsi_end, base, end;
995
996	/* check gsi range */
997	gsi_end = gsi_base + ((ver >> 16) & 0xff);
998	for (index = 0; index < NR_IOSAPICS; index++) {
999		if (!iosapic_lists[index].addr)
1000			continue;
1001
1002		base = iosapic_lists[index].gsi_base;
1003		end  = base + iosapic_lists[index].num_rte - 1;
1004
1005		if (gsi_end < base || end < gsi_base)
1006			continue; /* OK */
1007
1008		return -EBUSY;
1009	}
1010	return 0;
1011}
1012
1013static int
1014iosapic_delete_rte(unsigned int irq, unsigned int gsi)
1015{
1016	struct iosapic_rte_info *rte, *temp;
1017
1018	list_for_each_entry_safe(rte, temp, &iosapic_intr_info[irq].rtes,
1019								rte_list) {
1020		if (rte->iosapic->gsi_base + rte->rte_index == gsi) {
1021			if (rte->refcnt)
1022				return -EBUSY;
1023
1024			list_del(&rte->rte_list);
1025			kfree(rte);
1026			return 0;
1027		}
1028	}
1029
1030	return -EINVAL;
1031}
1032
1033int iosapic_init(unsigned long phys_addr, unsigned int gsi_base)
1034{
1035	int num_rte, err, index;
1036	unsigned int isa_irq, ver;
1037	char __iomem *addr;
1038	unsigned long flags;
1039
1040	spin_lock_irqsave(&iosapic_lock, flags);
1041	index = find_iosapic(gsi_base);
1042	if (index >= 0) {
1043		spin_unlock_irqrestore(&iosapic_lock, flags);
1044		return -EBUSY;
1045	}
1046
1047	addr = ioremap(phys_addr, 0);
1048	if (addr == NULL) {
1049		spin_unlock_irqrestore(&iosapic_lock, flags);
1050		return -ENOMEM;
1051	}
1052	ver = iosapic_version(addr);
1053	if ((err = iosapic_check_gsi_range(gsi_base, ver))) {
1054		iounmap(addr);
1055		spin_unlock_irqrestore(&iosapic_lock, flags);
1056		return err;
1057	}
1058
1059	/*
1060	 * The MAX_REDIR register holds the highest input pin number
1061	 * (starting from 0).  We add 1 so that we can use it for
1062	 * number of pins (= RTEs)
1063	 */
1064	num_rte = ((ver >> 16) & 0xff) + 1;
1065
1066	index = iosapic_alloc();
1067	iosapic_lists[index].addr = addr;
1068	iosapic_lists[index].gsi_base = gsi_base;
1069	iosapic_lists[index].num_rte = num_rte;
1070#ifdef CONFIG_NUMA
1071	iosapic_lists[index].node = MAX_NUMNODES;
1072#endif
1073	spin_lock_init(&iosapic_lists[index].lock);
1074	spin_unlock_irqrestore(&iosapic_lock, flags);
1075
1076	if ((gsi_base == 0) && pcat_compat) {
1077		/*
1078		 * Map the legacy ISA devices into the IOSAPIC data.  Some of
1079		 * these may get reprogrammed later on with data from the ACPI
1080		 * Interrupt Source Override table.
1081		 */
1082		for (isa_irq = 0; isa_irq < 16; ++isa_irq)
1083			iosapic_override_isa_irq(isa_irq, isa_irq,
1084						 IOSAPIC_POL_HIGH,
1085						 IOSAPIC_EDGE);
1086	}
1087	return 0;
1088}
1089
1090int iosapic_remove(unsigned int gsi_base)
1091{
1092	int i, irq, index, err = 0;
1093	unsigned long flags;
1094
1095	spin_lock_irqsave(&iosapic_lock, flags);
1096	index = find_iosapic(gsi_base);
1097	if (index < 0) {
1098		printk(KERN_WARNING "%s: No IOSAPIC for GSI base %u\n",
1099		       __func__, gsi_base);
1100		goto out;
1101	}
1102
1103	if (iosapic_lists[index].rtes_inuse) {
1104		err = -EBUSY;
1105		printk(KERN_WARNING "%s: IOSAPIC for GSI base %u is busy\n",
1106		       __func__, gsi_base);
1107		goto out;
1108	}
1109
1110	for (i = gsi_base; i < gsi_base + iosapic_lists[index].num_rte; i++) {
1111		irq = __gsi_to_irq(i);
1112		if (irq < 0)
1113			continue;
1114
1115		err = iosapic_delete_rte(irq, i);
1116		if (err)
1117			goto out;
1118	}
1119
1120	iounmap(iosapic_lists[index].addr);
1121	iosapic_free(index);
1122 out:
1123	spin_unlock_irqrestore(&iosapic_lock, flags);
1124	return err;
1125}
1126
1127#ifdef CONFIG_NUMA
1128void map_iosapic_to_node(unsigned int gsi_base, int node)
1129{
1130	int index;
1131
1132	index = find_iosapic(gsi_base);
1133	if (index < 0) {
1134		printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1135		       __func__, gsi_base);
1136		return;
1137	}
1138	iosapic_lists[index].node = node;
1139	return;
1140}
1141#endif
1142