Merge branch 'master' of ssh://master.kernel.org/pub/scm/linux/kernel/git/kaber/nf-2.6
[pandora-kernel.git] / arch / ia64 / kernel / iosapic.c
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  * describeinterrupts.  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 #include <asm/system.h>
102
103 #undef DEBUG_INTERRUPT_ROUTING
104
105 #ifdef DEBUG_INTERRUPT_ROUTING
106 #define DBG(fmt...)     printk(fmt)
107 #else
108 #define DBG(fmt...)
109 #endif
110
111 static DEFINE_SPINLOCK(iosapic_lock);
112
113 /*
114  * These tables map IA-64 vectors to the IOSAPIC pin that generates this
115  * vector.
116  */
117
118 #define NO_REF_RTE      0
119
120 static struct iosapic {
121         char __iomem    *addr;          /* base address of IOSAPIC */
122         unsigned int    gsi_base;       /* GSI base */
123         unsigned short  num_rte;        /* # of RTEs on this IOSAPIC */
124         int             rtes_inuse;     /* # of RTEs in use on this IOSAPIC */
125 #ifdef CONFIG_NUMA
126         unsigned short  node;           /* numa node association via pxm */
127 #endif
128         spinlock_t      lock;           /* lock for indirect reg access */
129 } iosapic_lists[NR_IOSAPICS];
130
131 struct iosapic_rte_info {
132         struct list_head rte_list;      /* RTEs sharing the same vector */
133         char            rte_index;      /* IOSAPIC RTE index */
134         int             refcnt;         /* reference counter */
135         struct iosapic  *iosapic;
136 } ____cacheline_aligned;
137
138 static struct iosapic_intr_info {
139         struct list_head rtes;          /* RTEs using this vector (empty =>
140                                          * not an IOSAPIC interrupt) */
141         int             count;          /* # of registered RTEs */
142         u32             low32;          /* current value of low word of
143                                          * Redirection table entry */
144         unsigned int    dest;           /* destination CPU physical ID */
145         unsigned char   dmode   : 3;    /* delivery mode (see iosapic.h) */
146         unsigned char   polarity: 1;    /* interrupt polarity
147                                          * (see iosapic.h) */
148         unsigned char   trigger : 1;    /* trigger mode (see iosapic.h) */
149 } iosapic_intr_info[NR_IRQS];
150
151 static unsigned char pcat_compat __devinitdata; /* 8259 compatibility flag */
152
153 static inline void
154 iosapic_write(struct iosapic *iosapic, unsigned int reg, u32 val)
155 {
156         unsigned long flags;
157
158         spin_lock_irqsave(&iosapic->lock, flags);
159         __iosapic_write(iosapic->addr, reg, val);
160         spin_unlock_irqrestore(&iosapic->lock, flags);
161 }
162
163 /*
164  * Find an IOSAPIC associated with a GSI
165  */
166 static inline int
167 find_iosapic (unsigned int gsi)
168 {
169         int i;
170
171         for (i = 0; i < NR_IOSAPICS; i++) {
172                 if ((unsigned) (gsi - iosapic_lists[i].gsi_base) <
173                     iosapic_lists[i].num_rte)
174                         return i;
175         }
176
177         return -1;
178 }
179
180 static inline int __gsi_to_irq(unsigned int gsi)
181 {
182         int irq;
183         struct iosapic_intr_info *info;
184         struct iosapic_rte_info *rte;
185
186         for (irq = 0; irq < NR_IRQS; irq++) {
187                 info = &iosapic_intr_info[irq];
188                 list_for_each_entry(rte, &info->rtes, rte_list)
189                         if (rte->iosapic->gsi_base + rte->rte_index == gsi)
190                                 return irq;
191         }
192         return -1;
193 }
194
195 int
196 gsi_to_irq (unsigned int gsi)
197 {
198         unsigned long flags;
199         int irq;
200
201         spin_lock_irqsave(&iosapic_lock, flags);
202         irq = __gsi_to_irq(gsi);
203         spin_unlock_irqrestore(&iosapic_lock, flags);
204         return irq;
205 }
206
207 static struct iosapic_rte_info *find_rte(unsigned int irq, unsigned int gsi)
208 {
209         struct iosapic_rte_info *rte;
210
211         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
212                 if (rte->iosapic->gsi_base + rte->rte_index == gsi)
213                         return rte;
214         return NULL;
215 }
216
217 static void
218 set_rte (unsigned int gsi, unsigned int irq, unsigned int dest, int mask)
219 {
220         unsigned long pol, trigger, dmode;
221         u32 low32, high32;
222         int rte_index;
223         char redir;
224         struct iosapic_rte_info *rte;
225         ia64_vector vector = irq_to_vector(irq);
226
227         DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
228
229         rte = find_rte(irq, gsi);
230         if (!rte)
231                 return;         /* not an IOSAPIC interrupt */
232
233         rte_index = rte->rte_index;
234         pol     = iosapic_intr_info[irq].polarity;
235         trigger = iosapic_intr_info[irq].trigger;
236         dmode   = iosapic_intr_info[irq].dmode;
237
238         redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
239
240 #ifdef CONFIG_SMP
241         set_irq_affinity_info(irq, (int)(dest & 0xffff), redir);
242 #endif
243
244         low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
245                  (trigger << IOSAPIC_TRIGGER_SHIFT) |
246                  (dmode << IOSAPIC_DELIVERY_SHIFT) |
247                  ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
248                  vector);
249
250         /* dest contains both id and eid */
251         high32 = (dest << IOSAPIC_DEST_SHIFT);
252
253         iosapic_write(rte->iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
254         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
255         iosapic_intr_info[irq].low32 = low32;
256         iosapic_intr_info[irq].dest = dest;
257 }
258
259 static void
260 nop (struct irq_data *data)
261 {
262         /* do nothing... */
263 }
264
265
266 #ifdef CONFIG_KEXEC
267 void
268 kexec_disable_iosapic(void)
269 {
270         struct iosapic_intr_info *info;
271         struct iosapic_rte_info *rte;
272         ia64_vector vec;
273         int irq;
274
275         for (irq = 0; irq < NR_IRQS; irq++) {
276                 info = &iosapic_intr_info[irq];
277                 vec = irq_to_vector(irq);
278                 list_for_each_entry(rte, &info->rtes,
279                                 rte_list) {
280                         iosapic_write(rte->iosapic,
281                                         IOSAPIC_RTE_LOW(rte->rte_index),
282                                         IOSAPIC_MASK|vec);
283                         iosapic_eoi(rte->iosapic->addr, vec);
284                 }
285         }
286 }
287 #endif
288
289 static void
290 mask_irq (struct irq_data *data)
291 {
292         unsigned int irq = data->irq;
293         u32 low32;
294         int rte_index;
295         struct iosapic_rte_info *rte;
296
297         if (!iosapic_intr_info[irq].count)
298                 return;                 /* not an IOSAPIC interrupt! */
299
300         /* set only the mask bit */
301         low32 = iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
302         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
303                 rte_index = rte->rte_index;
304                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
305         }
306 }
307
308 static void
309 unmask_irq (struct irq_data *data)
310 {
311         unsigned int irq = data->irq;
312         u32 low32;
313         int rte_index;
314         struct iosapic_rte_info *rte;
315
316         if (!iosapic_intr_info[irq].count)
317                 return;                 /* not an IOSAPIC interrupt! */
318
319         low32 = iosapic_intr_info[irq].low32 &= ~IOSAPIC_MASK;
320         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
321                 rte_index = rte->rte_index;
322                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
323         }
324 }
325
326
327 static int
328 iosapic_set_affinity(struct irq_data *data, const struct cpumask *mask,
329                      bool force)
330 {
331 #ifdef CONFIG_SMP
332         unsigned int irq = data->irq;
333         u32 high32, low32;
334         int cpu, dest, rte_index;
335         int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
336         struct iosapic_rte_info *rte;
337         struct iosapic *iosapic;
338
339         irq &= (~IA64_IRQ_REDIRECTED);
340
341         cpu = cpumask_first_and(cpu_online_mask, mask);
342         if (cpu >= nr_cpu_ids)
343                 return -1;
344
345         if (irq_prepare_move(irq, cpu))
346                 return -1;
347
348         dest = cpu_physical_id(cpu);
349
350         if (!iosapic_intr_info[irq].count)
351                 return -1;                      /* not an IOSAPIC interrupt */
352
353         set_irq_affinity_info(irq, dest, redir);
354
355         /* dest contains both id and eid */
356         high32 = dest << IOSAPIC_DEST_SHIFT;
357
358         low32 = iosapic_intr_info[irq].low32 & ~(7 << IOSAPIC_DELIVERY_SHIFT);
359         if (redir)
360                 /* change delivery mode to lowest priority */
361                 low32 |= (IOSAPIC_LOWEST_PRIORITY << IOSAPIC_DELIVERY_SHIFT);
362         else
363                 /* change delivery mode to fixed */
364                 low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
365         low32 &= IOSAPIC_VECTOR_MASK;
366         low32 |= irq_to_vector(irq);
367
368         iosapic_intr_info[irq].low32 = low32;
369         iosapic_intr_info[irq].dest = dest;
370         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
371                 iosapic = rte->iosapic;
372                 rte_index = rte->rte_index;
373                 iosapic_write(iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
374                 iosapic_write(iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
375         }
376
377 #endif
378         return 0;
379 }
380
381 /*
382  * Handlers for level-triggered interrupts.
383  */
384
385 static unsigned int
386 iosapic_startup_level_irq (struct irq_data *data)
387 {
388         unmask_irq(data);
389         return 0;
390 }
391
392 static void
393 iosapic_unmask_level_irq (struct irq_data *data)
394 {
395         unsigned int irq = data->irq;
396         ia64_vector vec = irq_to_vector(irq);
397         struct iosapic_rte_info *rte;
398         int do_unmask_irq = 0;
399
400         irq_complete_move(irq);
401         if (unlikely(irqd_is_setaffinity_pending(data))) {
402                 do_unmask_irq = 1;
403                 mask_irq(data);
404         } else
405                 unmask_irq(data);
406
407         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
408                 iosapic_eoi(rte->iosapic->addr, vec);
409
410         if (unlikely(do_unmask_irq)) {
411                 irq_move_masked_irq(data);
412                 unmask_irq(data);
413         }
414 }
415
416 #define iosapic_shutdown_level_irq      mask_irq
417 #define iosapic_enable_level_irq        unmask_irq
418 #define iosapic_disable_level_irq       mask_irq
419 #define iosapic_ack_level_irq           nop
420
421 static struct irq_chip irq_type_iosapic_level = {
422         .name =                 "IO-SAPIC-level",
423         .irq_startup =          iosapic_startup_level_irq,
424         .irq_shutdown =         iosapic_shutdown_level_irq,
425         .irq_enable =           iosapic_enable_level_irq,
426         .irq_disable =          iosapic_disable_level_irq,
427         .irq_ack =              iosapic_ack_level_irq,
428         .irq_mask =             mask_irq,
429         .irq_unmask =           iosapic_unmask_level_irq,
430         .irq_set_affinity =     iosapic_set_affinity
431 };
432
433 /*
434  * Handlers for edge-triggered interrupts.
435  */
436
437 static unsigned int
438 iosapic_startup_edge_irq (struct irq_data *data)
439 {
440         unmask_irq(data);
441         /*
442          * IOSAPIC simply drops interrupts pended while the
443          * corresponding pin was masked, so we can't know if an
444          * interrupt is pending already.  Let's hope not...
445          */
446         return 0;
447 }
448
449 static void
450 iosapic_ack_edge_irq (struct irq_data *data)
451 {
452         irq_complete_move(data->irq);
453         irq_move_irq(data);
454 }
455
456 #define iosapic_enable_edge_irq         unmask_irq
457 #define iosapic_disable_edge_irq        nop
458
459 static struct irq_chip irq_type_iosapic_edge = {
460         .name =                 "IO-SAPIC-edge",
461         .irq_startup =          iosapic_startup_edge_irq,
462         .irq_shutdown =         iosapic_disable_edge_irq,
463         .irq_enable =           iosapic_enable_edge_irq,
464         .irq_disable =          iosapic_disable_edge_irq,
465         .irq_ack =              iosapic_ack_edge_irq,
466         .irq_mask =             mask_irq,
467         .irq_unmask =           unmask_irq,
468         .irq_set_affinity =     iosapic_set_affinity
469 };
470
471 static unsigned int
472 iosapic_version (char __iomem *addr)
473 {
474         /*
475          * IOSAPIC Version Register return 32 bit structure like:
476          * {
477          *      unsigned int version   : 8;
478          *      unsigned int reserved1 : 8;
479          *      unsigned int max_redir : 8;
480          *      unsigned int reserved2 : 8;
481          * }
482          */
483         return __iosapic_read(addr, IOSAPIC_VERSION);
484 }
485
486 static int iosapic_find_sharable_irq(unsigned long trigger, unsigned long pol)
487 {
488         int i, irq = -ENOSPC, min_count = -1;
489         struct iosapic_intr_info *info;
490
491         /*
492          * shared vectors for edge-triggered interrupts are not
493          * supported yet
494          */
495         if (trigger == IOSAPIC_EDGE)
496                 return -EINVAL;
497
498         for (i = 0; i < NR_IRQS; i++) {
499                 info = &iosapic_intr_info[i];
500                 if (info->trigger == trigger && info->polarity == pol &&
501                     (info->dmode == IOSAPIC_FIXED ||
502                      info->dmode == IOSAPIC_LOWEST_PRIORITY) &&
503                     can_request_irq(i, IRQF_SHARED)) {
504                         if (min_count == -1 || info->count < min_count) {
505                                 irq = i;
506                                 min_count = info->count;
507                         }
508                 }
509         }
510         return irq;
511 }
512
513 /*
514  * if the given vector is already owned by other,
515  *  assign a new vector for the other and make the vector available
516  */
517 static void __init
518 iosapic_reassign_vector (int irq)
519 {
520         int new_irq;
521
522         if (iosapic_intr_info[irq].count) {
523                 new_irq = create_irq();
524                 if (new_irq < 0)
525                         panic("%s: out of interrupt vectors!\n", __func__);
526                 printk(KERN_INFO "Reassigning vector %d to %d\n",
527                        irq_to_vector(irq), irq_to_vector(new_irq));
528                 memcpy(&iosapic_intr_info[new_irq], &iosapic_intr_info[irq],
529                        sizeof(struct iosapic_intr_info));
530                 INIT_LIST_HEAD(&iosapic_intr_info[new_irq].rtes);
531                 list_move(iosapic_intr_info[irq].rtes.next,
532                           &iosapic_intr_info[new_irq].rtes);
533                 memset(&iosapic_intr_info[irq], 0,
534                        sizeof(struct iosapic_intr_info));
535                 iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
536                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
537         }
538 }
539
540 static inline int irq_is_shared (int irq)
541 {
542         return (iosapic_intr_info[irq].count > 1);
543 }
544
545 struct irq_chip*
546 ia64_native_iosapic_get_irq_chip(unsigned long trigger)
547 {
548         if (trigger == IOSAPIC_EDGE)
549                 return &irq_type_iosapic_edge;
550         else
551                 return &irq_type_iosapic_level;
552 }
553
554 static int
555 register_intr (unsigned int gsi, int irq, unsigned char delivery,
556                unsigned long polarity, unsigned long trigger)
557 {
558         struct irq_chip *chip, *irq_type;
559         int index;
560         struct iosapic_rte_info *rte;
561
562         index = find_iosapic(gsi);
563         if (index < 0) {
564                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
565                        __func__, gsi);
566                 return -ENODEV;
567         }
568
569         rte = find_rte(irq, gsi);
570         if (!rte) {
571                 rte = kzalloc(sizeof (*rte), GFP_ATOMIC);
572                 if (!rte) {
573                         printk(KERN_WARNING "%s: cannot allocate memory\n",
574                                __func__);
575                         return -ENOMEM;
576                 }
577
578                 rte->iosapic    = &iosapic_lists[index];
579                 rte->rte_index  = gsi - rte->iosapic->gsi_base;
580                 rte->refcnt++;
581                 list_add_tail(&rte->rte_list, &iosapic_intr_info[irq].rtes);
582                 iosapic_intr_info[irq].count++;
583                 iosapic_lists[index].rtes_inuse++;
584         }
585         else if (rte->refcnt == NO_REF_RTE) {
586                 struct iosapic_intr_info *info = &iosapic_intr_info[irq];
587                 if (info->count > 0 &&
588                     (info->trigger != trigger || info->polarity != polarity)){
589                         printk (KERN_WARNING
590                                 "%s: cannot override the interrupt\n",
591                                 __func__);
592                         return -EINVAL;
593                 }
594                 rte->refcnt++;
595                 iosapic_intr_info[irq].count++;
596                 iosapic_lists[index].rtes_inuse++;
597         }
598
599         iosapic_intr_info[irq].polarity = polarity;
600         iosapic_intr_info[irq].dmode    = delivery;
601         iosapic_intr_info[irq].trigger  = trigger;
602
603         irq_type = iosapic_get_irq_chip(trigger);
604
605         chip = irq_get_chip(irq);
606         if (irq_type != NULL && chip != irq_type) {
607                 if (chip != &no_irq_chip)
608                         printk(KERN_WARNING
609                                "%s: changing vector %d from %s to %s\n",
610                                __func__, irq_to_vector(irq),
611                                chip->name, irq_type->name);
612                 chip = irq_type;
613         }
614         __irq_set_chip_handler_name_locked(irq, chip, trigger == IOSAPIC_EDGE ?
615                                            handle_edge_irq : handle_level_irq,
616                                            NULL);
617         return 0;
618 }
619
620 static unsigned int
621 get_target_cpu (unsigned int gsi, int irq)
622 {
623 #ifdef CONFIG_SMP
624         static int cpu = -1;
625         extern int cpe_vector;
626         cpumask_t domain = irq_to_domain(irq);
627
628         /*
629          * In case of vector shared by multiple RTEs, all RTEs that
630          * share the vector need to use the same destination CPU.
631          */
632         if (iosapic_intr_info[irq].count)
633                 return iosapic_intr_info[irq].dest;
634
635         /*
636          * If the platform supports redirection via XTP, let it
637          * distribute interrupts.
638          */
639         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
640                 return cpu_physical_id(smp_processor_id());
641
642         /*
643          * Some interrupts (ACPI SCI, for instance) are registered
644          * before the BSP is marked as online.
645          */
646         if (!cpu_online(smp_processor_id()))
647                 return cpu_physical_id(smp_processor_id());
648
649 #ifdef CONFIG_ACPI
650         if (cpe_vector > 0 && irq_to_vector(irq) == IA64_CPEP_VECTOR)
651                 return get_cpei_target_cpu();
652 #endif
653
654 #ifdef CONFIG_NUMA
655         {
656                 int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
657                 const struct cpumask *cpu_mask;
658
659                 iosapic_index = find_iosapic(gsi);
660                 if (iosapic_index < 0 ||
661                     iosapic_lists[iosapic_index].node == MAX_NUMNODES)
662                         goto skip_numa_setup;
663
664                 cpu_mask = cpumask_of_node(iosapic_lists[iosapic_index].node);
665                 num_cpus = 0;
666                 for_each_cpu_and(numa_cpu, cpu_mask, &domain) {
667                         if (cpu_online(numa_cpu))
668                                 num_cpus++;
669                 }
670
671                 if (!num_cpus)
672                         goto skip_numa_setup;
673
674                 /* Use irq assignment to distribute across cpus in node */
675                 cpu_index = irq % num_cpus;
676
677                 for_each_cpu_and(numa_cpu, cpu_mask, &domain)
678                         if (cpu_online(numa_cpu) && i++ >= cpu_index)
679                                 break;
680
681                 if (numa_cpu < nr_cpu_ids)
682                         return cpu_physical_id(numa_cpu);
683         }
684 skip_numa_setup:
685 #endif
686         /*
687          * Otherwise, round-robin interrupt vectors across all the
688          * processors.  (It'd be nice if we could be smarter in the
689          * case of NUMA.)
690          */
691         do {
692                 if (++cpu >= nr_cpu_ids)
693                         cpu = 0;
694         } while (!cpu_online(cpu) || !cpu_isset(cpu, domain));
695
696         return cpu_physical_id(cpu);
697 #else  /* CONFIG_SMP */
698         return cpu_physical_id(smp_processor_id());
699 #endif
700 }
701
702 static inline unsigned char choose_dmode(void)
703 {
704 #ifdef CONFIG_SMP
705         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
706                 return IOSAPIC_LOWEST_PRIORITY;
707 #endif
708         return IOSAPIC_FIXED;
709 }
710
711 /*
712  * ACPI can describe IOSAPIC interrupts via static tables and namespace
713  * methods.  This provides an interface to register those interrupts and
714  * program the IOSAPIC RTE.
715  */
716 int
717 iosapic_register_intr (unsigned int gsi,
718                        unsigned long polarity, unsigned long trigger)
719 {
720         int irq, mask = 1, err;
721         unsigned int dest;
722         unsigned long flags;
723         struct iosapic_rte_info *rte;
724         u32 low32;
725         unsigned char dmode;
726         struct irq_desc *desc;
727
728         /*
729          * If this GSI has already been registered (i.e., it's a
730          * shared interrupt, or we lost a race to register it),
731          * don't touch the RTE.
732          */
733         spin_lock_irqsave(&iosapic_lock, flags);
734         irq = __gsi_to_irq(gsi);
735         if (irq > 0) {
736                 rte = find_rte(irq, gsi);
737                 if(iosapic_intr_info[irq].count == 0) {
738                         assign_irq_vector(irq);
739                         dynamic_irq_init(irq);
740                 } else if (rte->refcnt != NO_REF_RTE) {
741                         rte->refcnt++;
742                         goto unlock_iosapic_lock;
743                 }
744         } else
745                 irq = create_irq();
746
747         /* If vector is running out, we try to find a sharable vector */
748         if (irq < 0) {
749                 irq = iosapic_find_sharable_irq(trigger, polarity);
750                 if (irq < 0)
751                         goto unlock_iosapic_lock;
752         }
753
754         desc = irq_to_desc(irq);
755         raw_spin_lock(&desc->lock);
756         dest = get_target_cpu(gsi, irq);
757         dmode = choose_dmode();
758         err = register_intr(gsi, irq, dmode, polarity, trigger);
759         if (err < 0) {
760                 raw_spin_unlock(&desc->lock);
761                 irq = err;
762                 goto unlock_iosapic_lock;
763         }
764
765         /*
766          * If the vector is shared and already unmasked for other
767          * interrupt sources, don't mask it.
768          */
769         low32 = iosapic_intr_info[irq].low32;
770         if (irq_is_shared(irq) && !(low32 & IOSAPIC_MASK))
771                 mask = 0;
772         set_rte(gsi, irq, dest, mask);
773
774         printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
775                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
776                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
777                cpu_logical_id(dest), dest, irq_to_vector(irq));
778
779         raw_spin_unlock(&desc->lock);
780  unlock_iosapic_lock:
781         spin_unlock_irqrestore(&iosapic_lock, flags);
782         return irq;
783 }
784
785 void
786 iosapic_unregister_intr (unsigned int gsi)
787 {
788         unsigned long flags;
789         int irq, index;
790         u32 low32;
791         unsigned long trigger, polarity;
792         unsigned int dest;
793         struct iosapic_rte_info *rte;
794
795         /*
796          * If the irq associated with the gsi is not found,
797          * iosapic_unregister_intr() is unbalanced. We need to check
798          * this again after getting locks.
799          */
800         irq = gsi_to_irq(gsi);
801         if (irq < 0) {
802                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
803                        gsi);
804                 WARN_ON(1);
805                 return;
806         }
807
808         spin_lock_irqsave(&iosapic_lock, flags);
809         if ((rte = find_rte(irq, gsi)) == NULL) {
810                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
811                        gsi);
812                 WARN_ON(1);
813                 goto out;
814         }
815
816         if (--rte->refcnt > 0)
817                 goto out;
818
819         rte->refcnt = NO_REF_RTE;
820
821         /* Mask the interrupt */
822         low32 = iosapic_intr_info[irq].low32 | IOSAPIC_MASK;
823         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte->rte_index), low32);
824
825         iosapic_intr_info[irq].count--;
826         index = find_iosapic(gsi);
827         iosapic_lists[index].rtes_inuse--;
828         WARN_ON(iosapic_lists[index].rtes_inuse < 0);
829
830         trigger  = iosapic_intr_info[irq].trigger;
831         polarity = iosapic_intr_info[irq].polarity;
832         dest     = iosapic_intr_info[irq].dest;
833         printk(KERN_INFO
834                "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d unregistered\n",
835                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
836                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
837                cpu_logical_id(dest), dest, irq_to_vector(irq));
838
839         if (iosapic_intr_info[irq].count == 0) {
840 #ifdef CONFIG_SMP
841                 /* Clear affinity */
842                 cpumask_setall(irq_get_irq_data(irq)->affinity);
843 #endif
844                 /* Clear the interrupt information */
845                 iosapic_intr_info[irq].dest = 0;
846                 iosapic_intr_info[irq].dmode = 0;
847                 iosapic_intr_info[irq].polarity = 0;
848                 iosapic_intr_info[irq].trigger = 0;
849                 iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
850
851                 /* Destroy and reserve IRQ */
852                 destroy_and_reserve_irq(irq);
853         }
854  out:
855         spin_unlock_irqrestore(&iosapic_lock, flags);
856 }
857
858 /*
859  * ACPI calls this when it finds an entry for a platform interrupt.
860  */
861 int __init
862 iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
863                                 int iosapic_vector, u16 eid, u16 id,
864                                 unsigned long polarity, unsigned long trigger)
865 {
866         static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
867         unsigned char delivery;
868         int irq, vector, mask = 0;
869         unsigned int dest = ((id << 8) | eid) & 0xffff;
870
871         switch (int_type) {
872               case ACPI_INTERRUPT_PMI:
873                 irq = vector = iosapic_vector;
874                 bind_irq_vector(irq, vector, CPU_MASK_ALL);
875                 /*
876                  * since PMI vector is alloc'd by FW(ACPI) not by kernel,
877                  * we need to make sure the vector is available
878                  */
879                 iosapic_reassign_vector(irq);
880                 delivery = IOSAPIC_PMI;
881                 break;
882               case ACPI_INTERRUPT_INIT:
883                 irq = create_irq();
884                 if (irq < 0)
885                         panic("%s: out of interrupt vectors!\n", __func__);
886                 vector = irq_to_vector(irq);
887                 delivery = IOSAPIC_INIT;
888                 break;
889               case ACPI_INTERRUPT_CPEI:
890                 irq = vector = IA64_CPE_VECTOR;
891                 BUG_ON(bind_irq_vector(irq, vector, CPU_MASK_ALL));
892                 delivery = IOSAPIC_FIXED;
893                 mask = 1;
894                 break;
895               default:
896                 printk(KERN_ERR "%s: invalid int type 0x%x\n", __func__,
897                        int_type);
898                 return -1;
899         }
900
901         register_intr(gsi, irq, delivery, polarity, trigger);
902
903         printk(KERN_INFO
904                "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x)"
905                " vector %d\n",
906                int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
907                int_type, gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
908                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
909                cpu_logical_id(dest), dest, vector);
910
911         set_rte(gsi, irq, dest, mask);
912         return vector;
913 }
914
915 /*
916  * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
917  */
918 void __devinit
919 iosapic_override_isa_irq (unsigned int isa_irq, unsigned int gsi,
920                           unsigned long polarity,
921                           unsigned long trigger)
922 {
923         int vector, irq;
924         unsigned int dest = cpu_physical_id(smp_processor_id());
925         unsigned char dmode;
926
927         irq = vector = isa_irq_to_vector(isa_irq);
928         BUG_ON(bind_irq_vector(irq, vector, CPU_MASK_ALL));
929         dmode = choose_dmode();
930         register_intr(gsi, irq, dmode, polarity, trigger);
931
932         DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
933             isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
934             polarity == IOSAPIC_POL_HIGH ? "high" : "low",
935             cpu_logical_id(dest), dest, vector);
936
937         set_rte(gsi, irq, dest, 1);
938 }
939
940 void __init
941 ia64_native_iosapic_pcat_compat_init(void)
942 {
943         if (pcat_compat) {
944                 /*
945                  * Disable the compatibility mode interrupts (8259 style),
946                  * needs IN/OUT support enabled.
947                  */
948                 printk(KERN_INFO
949                        "%s: Disabling PC-AT compatible 8259 interrupts\n",
950                        __func__);
951                 outb(0xff, 0xA1);
952                 outb(0xff, 0x21);
953         }
954 }
955
956 void __init
957 iosapic_system_init (int system_pcat_compat)
958 {
959         int irq;
960
961         for (irq = 0; irq < NR_IRQS; ++irq) {
962                 iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
963                 /* mark as unused */
964                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
965
966                 iosapic_intr_info[irq].count = 0;
967         }
968
969         pcat_compat = system_pcat_compat;
970         if (pcat_compat)
971                 iosapic_pcat_compat_init();
972 }
973
974 static inline int
975 iosapic_alloc (void)
976 {
977         int index;
978
979         for (index = 0; index < NR_IOSAPICS; index++)
980                 if (!iosapic_lists[index].addr)
981                         return index;
982
983         printk(KERN_WARNING "%s: failed to allocate iosapic\n", __func__);
984         return -1;
985 }
986
987 static inline void
988 iosapic_free (int index)
989 {
990         memset(&iosapic_lists[index], 0, sizeof(iosapic_lists[0]));
991 }
992
993 static inline int
994 iosapic_check_gsi_range (unsigned int gsi_base, unsigned int ver)
995 {
996         int index;
997         unsigned int gsi_end, base, end;
998
999         /* check gsi range */
1000         gsi_end = gsi_base + ((ver >> 16) & 0xff);
1001         for (index = 0; index < NR_IOSAPICS; index++) {
1002                 if (!iosapic_lists[index].addr)
1003                         continue;
1004
1005                 base = iosapic_lists[index].gsi_base;
1006                 end  = base + iosapic_lists[index].num_rte - 1;
1007
1008                 if (gsi_end < base || end < gsi_base)
1009                         continue; /* OK */
1010
1011                 return -EBUSY;
1012         }
1013         return 0;
1014 }
1015
1016 int __devinit
1017 iosapic_init (unsigned long phys_addr, unsigned int gsi_base)
1018 {
1019         int num_rte, err, index;
1020         unsigned int isa_irq, ver;
1021         char __iomem *addr;
1022         unsigned long flags;
1023
1024         spin_lock_irqsave(&iosapic_lock, flags);
1025         index = find_iosapic(gsi_base);
1026         if (index >= 0) {
1027                 spin_unlock_irqrestore(&iosapic_lock, flags);
1028                 return -EBUSY;
1029         }
1030
1031         addr = ioremap(phys_addr, 0);
1032         if (addr == NULL) {
1033                 spin_unlock_irqrestore(&iosapic_lock, flags);
1034                 return -ENOMEM;
1035         }
1036         ver = iosapic_version(addr);
1037         if ((err = iosapic_check_gsi_range(gsi_base, ver))) {
1038                 iounmap(addr);
1039                 spin_unlock_irqrestore(&iosapic_lock, flags);
1040                 return err;
1041         }
1042
1043         /*
1044          * The MAX_REDIR register holds the highest input pin number
1045          * (starting from 0).  We add 1 so that we can use it for
1046          * number of pins (= RTEs)
1047          */
1048         num_rte = ((ver >> 16) & 0xff) + 1;
1049
1050         index = iosapic_alloc();
1051         iosapic_lists[index].addr = addr;
1052         iosapic_lists[index].gsi_base = gsi_base;
1053         iosapic_lists[index].num_rte = num_rte;
1054 #ifdef CONFIG_NUMA
1055         iosapic_lists[index].node = MAX_NUMNODES;
1056 #endif
1057         spin_lock_init(&iosapic_lists[index].lock);
1058         spin_unlock_irqrestore(&iosapic_lock, flags);
1059
1060         if ((gsi_base == 0) && pcat_compat) {
1061                 /*
1062                  * Map the legacy ISA devices into the IOSAPIC data.  Some of
1063                  * these may get reprogrammed later on with data from the ACPI
1064                  * Interrupt Source Override table.
1065                  */
1066                 for (isa_irq = 0; isa_irq < 16; ++isa_irq)
1067                         iosapic_override_isa_irq(isa_irq, isa_irq,
1068                                                  IOSAPIC_POL_HIGH,
1069                                                  IOSAPIC_EDGE);
1070         }
1071         return 0;
1072 }
1073
1074 #ifdef CONFIG_HOTPLUG
1075 int
1076 iosapic_remove (unsigned int gsi_base)
1077 {
1078         int index, err = 0;
1079         unsigned long flags;
1080
1081         spin_lock_irqsave(&iosapic_lock, flags);
1082         index = find_iosapic(gsi_base);
1083         if (index < 0) {
1084                 printk(KERN_WARNING "%s: No IOSAPIC for GSI base %u\n",
1085                        __func__, gsi_base);
1086                 goto out;
1087         }
1088
1089         if (iosapic_lists[index].rtes_inuse) {
1090                 err = -EBUSY;
1091                 printk(KERN_WARNING "%s: IOSAPIC for GSI base %u is busy\n",
1092                        __func__, gsi_base);
1093                 goto out;
1094         }
1095
1096         iounmap(iosapic_lists[index].addr);
1097         iosapic_free(index);
1098  out:
1099         spin_unlock_irqrestore(&iosapic_lock, flags);
1100         return err;
1101 }
1102 #endif /* CONFIG_HOTPLUG */
1103
1104 #ifdef CONFIG_NUMA
1105 void __devinit
1106 map_iosapic_to_node(unsigned int gsi_base, int node)
1107 {
1108         int index;
1109
1110         index = find_iosapic(gsi_base);
1111         if (index < 0) {
1112                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1113                        __func__, gsi_base);
1114                 return;
1115         }
1116         iosapic_lists[index].node = node;
1117         return;
1118 }
1119 #endif