Pull thinkpad-2.6.24 into release branch
[linux-2.6] / 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 include/asm-ia64/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/smp.h>
90 #include <linux/string.h>
91 #include <linux/bootmem.h>
92
93 #include <asm/delay.h>
94 #include <asm/hw_irq.h>
95 #include <asm/io.h>
96 #include <asm/iosapic.h>
97 #include <asm/machvec.h>
98 #include <asm/processor.h>
99 #include <asm/ptrace.h>
100 #include <asm/system.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
110 #define NR_PREALLOCATE_RTE_ENTRIES \
111         (PAGE_SIZE / sizeof(struct iosapic_rte_info))
112 #define RTE_PREALLOCATED        (1)
113
114 static DEFINE_SPINLOCK(iosapic_lock);
115
116 /*
117  * These tables map IA-64 vectors to the IOSAPIC pin that generates this
118  * vector.
119  */
120
121 #define NO_REF_RTE      0
122
123 static struct iosapic {
124         char __iomem    *addr;          /* base address of IOSAPIC */
125         unsigned int    gsi_base;       /* GSI base */
126         unsigned short  num_rte;        /* # of RTEs on this IOSAPIC */
127         int             rtes_inuse;     /* # of RTEs in use on this IOSAPIC */
128 #ifdef CONFIG_NUMA
129         unsigned short  node;           /* numa node association via pxm */
130 #endif
131         spinlock_t      lock;           /* lock for indirect reg access */
132 } iosapic_lists[NR_IOSAPICS];
133
134 struct iosapic_rte_info {
135         struct list_head rte_list;      /* RTEs sharing the same vector */
136         char            rte_index;      /* IOSAPIC RTE index */
137         int             refcnt;         /* reference counter */
138         unsigned int    flags;          /* flags */
139         struct iosapic  *iosapic;
140 } ____cacheline_aligned;
141
142 static struct iosapic_intr_info {
143         struct list_head rtes;          /* RTEs using this vector (empty =>
144                                          * not an IOSAPIC interrupt) */
145         int             count;          /* # of registered RTEs */
146         u32             low32;          /* current value of low word of
147                                          * Redirection table entry */
148         unsigned int    dest;           /* destination CPU physical ID */
149         unsigned char   dmode   : 3;    /* delivery mode (see iosapic.h) */
150         unsigned char   polarity: 1;    /* interrupt polarity
151                                          * (see iosapic.h) */
152         unsigned char   trigger : 1;    /* trigger mode (see iosapic.h) */
153 } iosapic_intr_info[NR_IRQS];
154
155 static unsigned char pcat_compat __devinitdata; /* 8259 compatibility flag */
156
157 static int iosapic_kmalloc_ok;
158 static LIST_HEAD(free_rte_list);
159
160 static inline void
161 iosapic_write(struct iosapic *iosapic, unsigned int reg, u32 val)
162 {
163         unsigned long flags;
164
165         spin_lock_irqsave(&iosapic->lock, flags);
166         __iosapic_write(iosapic->addr, reg, val);
167         spin_unlock_irqrestore(&iosapic->lock, flags);
168 }
169
170 /*
171  * Find an IOSAPIC associated with a GSI
172  */
173 static inline int
174 find_iosapic (unsigned int gsi)
175 {
176         int i;
177
178         for (i = 0; i < NR_IOSAPICS; i++) {
179                 if ((unsigned) (gsi - iosapic_lists[i].gsi_base) <
180                     iosapic_lists[i].num_rte)
181                         return i;
182         }
183
184         return -1;
185 }
186
187 static inline int __gsi_to_irq(unsigned int gsi)
188 {
189         int irq;
190         struct iosapic_intr_info *info;
191         struct iosapic_rte_info *rte;
192
193         for (irq = 0; irq < NR_IRQS; irq++) {
194                 info = &iosapic_intr_info[irq];
195                 list_for_each_entry(rte, &info->rtes, rte_list)
196                         if (rte->iosapic->gsi_base + rte->rte_index == gsi)
197                                 return irq;
198         }
199         return -1;
200 }
201
202 /*
203  * Translate GSI number to the corresponding IA-64 interrupt vector.  If no
204  * entry exists, return -1.
205  */
206 inline int
207 gsi_to_vector (unsigned int gsi)
208 {
209         int irq = __gsi_to_irq(gsi);
210         if (check_irq_used(irq) < 0)
211                 return -1;
212         return irq_to_vector(irq);
213 }
214
215 int
216 gsi_to_irq (unsigned int gsi)
217 {
218         unsigned long flags;
219         int irq;
220
221         spin_lock_irqsave(&iosapic_lock, flags);
222         irq = __gsi_to_irq(gsi);
223         spin_unlock_irqrestore(&iosapic_lock, flags);
224         return irq;
225 }
226
227 static struct iosapic_rte_info *find_rte(unsigned int irq, unsigned int gsi)
228 {
229         struct iosapic_rte_info *rte;
230
231         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
232                 if (rte->iosapic->gsi_base + rte->rte_index == gsi)
233                         return rte;
234         return NULL;
235 }
236
237 static void
238 set_rte (unsigned int gsi, unsigned int irq, unsigned int dest, int mask)
239 {
240         unsigned long pol, trigger, dmode;
241         u32 low32, high32;
242         int rte_index;
243         char redir;
244         struct iosapic_rte_info *rte;
245         ia64_vector vector = irq_to_vector(irq);
246
247         DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
248
249         rte = find_rte(irq, gsi);
250         if (!rte)
251                 return;         /* not an IOSAPIC interrupt */
252
253         rte_index = rte->rte_index;
254         pol     = iosapic_intr_info[irq].polarity;
255         trigger = iosapic_intr_info[irq].trigger;
256         dmode   = iosapic_intr_info[irq].dmode;
257
258         redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
259
260 #ifdef CONFIG_SMP
261         set_irq_affinity_info(irq, (int)(dest & 0xffff), redir);
262 #endif
263
264         low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
265                  (trigger << IOSAPIC_TRIGGER_SHIFT) |
266                  (dmode << IOSAPIC_DELIVERY_SHIFT) |
267                  ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
268                  vector);
269
270         /* dest contains both id and eid */
271         high32 = (dest << IOSAPIC_DEST_SHIFT);
272
273         iosapic_write(rte->iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
274         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
275         iosapic_intr_info[irq].low32 = low32;
276         iosapic_intr_info[irq].dest = dest;
277 }
278
279 static void
280 nop (unsigned int irq)
281 {
282         /* do nothing... */
283 }
284
285
286 #ifdef CONFIG_KEXEC
287 void
288 kexec_disable_iosapic(void)
289 {
290         struct iosapic_intr_info *info;
291         struct iosapic_rte_info *rte;
292         ia64_vector vec;
293         int irq;
294
295         for (irq = 0; irq < NR_IRQS; irq++) {
296                 info = &iosapic_intr_info[irq];
297                 vec = irq_to_vector(irq);
298                 list_for_each_entry(rte, &info->rtes,
299                                 rte_list) {
300                         iosapic_write(rte->iosapic,
301                                         IOSAPIC_RTE_LOW(rte->rte_index),
302                                         IOSAPIC_MASK|vec);
303                         iosapic_eoi(rte->iosapic->addr, vec);
304                 }
305         }
306 }
307 #endif
308
309 static void
310 mask_irq (unsigned int irq)
311 {
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         /* set only the mask bit */
320         low32 = iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
321         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
322                 rte_index = rte->rte_index;
323                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
324         }
325 }
326
327 static void
328 unmask_irq (unsigned int irq)
329 {
330         u32 low32;
331         int rte_index;
332         struct iosapic_rte_info *rte;
333
334         if (!iosapic_intr_info[irq].count)
335                 return;                 /* not an IOSAPIC interrupt! */
336
337         low32 = iosapic_intr_info[irq].low32 &= ~IOSAPIC_MASK;
338         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
339                 rte_index = rte->rte_index;
340                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
341         }
342 }
343
344
345 static void
346 iosapic_set_affinity (unsigned int irq, cpumask_t mask)
347 {
348 #ifdef CONFIG_SMP
349         u32 high32, low32;
350         int dest, rte_index;
351         int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
352         struct iosapic_rte_info *rte;
353         struct iosapic *iosapic;
354
355         irq &= (~IA64_IRQ_REDIRECTED);
356
357         cpus_and(mask, mask, cpu_online_map);
358         if (cpus_empty(mask))
359                 return;
360
361         if (reassign_irq_vector(irq, first_cpu(mask)))
362                 return;
363
364         dest = cpu_physical_id(first_cpu(mask));
365
366         if (!iosapic_intr_info[irq].count)
367                 return;                 /* not an IOSAPIC interrupt */
368
369         set_irq_affinity_info(irq, dest, redir);
370
371         /* dest contains both id and eid */
372         high32 = dest << IOSAPIC_DEST_SHIFT;
373
374         low32 = iosapic_intr_info[irq].low32 & ~(7 << IOSAPIC_DELIVERY_SHIFT);
375         if (redir)
376                 /* change delivery mode to lowest priority */
377                 low32 |= (IOSAPIC_LOWEST_PRIORITY << IOSAPIC_DELIVERY_SHIFT);
378         else
379                 /* change delivery mode to fixed */
380                 low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
381         low32 &= IOSAPIC_VECTOR_MASK;
382         low32 |= irq_to_vector(irq);
383
384         iosapic_intr_info[irq].low32 = low32;
385         iosapic_intr_info[irq].dest = dest;
386         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
387                 iosapic = rte->iosapic;
388                 rte_index = rte->rte_index;
389                 iosapic_write(iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
390                 iosapic_write(iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
391         }
392 #endif
393 }
394
395 /*
396  * Handlers for level-triggered interrupts.
397  */
398
399 static unsigned int
400 iosapic_startup_level_irq (unsigned int irq)
401 {
402         unmask_irq(irq);
403         return 0;
404 }
405
406 static void
407 iosapic_end_level_irq (unsigned int irq)
408 {
409         ia64_vector vec = irq_to_vector(irq);
410         struct iosapic_rte_info *rte;
411         int do_unmask_irq = 0;
412
413         if (unlikely(irq_desc[irq].status & IRQ_MOVE_PENDING)) {
414                 do_unmask_irq = 1;
415                 mask_irq(irq);
416         }
417
418         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
419                 iosapic_eoi(rte->iosapic->addr, vec);
420
421         if (unlikely(do_unmask_irq)) {
422                 move_masked_irq(irq);
423                 unmask_irq(irq);
424         }
425 }
426
427 #define iosapic_shutdown_level_irq      mask_irq
428 #define iosapic_enable_level_irq        unmask_irq
429 #define iosapic_disable_level_irq       mask_irq
430 #define iosapic_ack_level_irq           nop
431
432 struct irq_chip irq_type_iosapic_level = {
433         .name =         "IO-SAPIC-level",
434         .startup =      iosapic_startup_level_irq,
435         .shutdown =     iosapic_shutdown_level_irq,
436         .enable =       iosapic_enable_level_irq,
437         .disable =      iosapic_disable_level_irq,
438         .ack =          iosapic_ack_level_irq,
439         .end =          iosapic_end_level_irq,
440         .mask =         mask_irq,
441         .unmask =       unmask_irq,
442         .set_affinity = iosapic_set_affinity
443 };
444
445 /*
446  * Handlers for edge-triggered interrupts.
447  */
448
449 static unsigned int
450 iosapic_startup_edge_irq (unsigned int irq)
451 {
452         unmask_irq(irq);
453         /*
454          * IOSAPIC simply drops interrupts pended while the
455          * corresponding pin was masked, so we can't know if an
456          * interrupt is pending already.  Let's hope not...
457          */
458         return 0;
459 }
460
461 static void
462 iosapic_ack_edge_irq (unsigned int irq)
463 {
464         irq_desc_t *idesc = irq_desc + irq;
465
466         move_native_irq(irq);
467         /*
468          * Once we have recorded IRQ_PENDING already, we can mask the
469          * interrupt for real. This prevents IRQ storms from unhandled
470          * devices.
471          */
472         if ((idesc->status & (IRQ_PENDING|IRQ_DISABLED)) ==
473             (IRQ_PENDING|IRQ_DISABLED))
474                 mask_irq(irq);
475 }
476
477 #define iosapic_enable_edge_irq         unmask_irq
478 #define iosapic_disable_edge_irq        nop
479 #define iosapic_end_edge_irq            nop
480
481 struct irq_chip irq_type_iosapic_edge = {
482         .name =         "IO-SAPIC-edge",
483         .startup =      iosapic_startup_edge_irq,
484         .shutdown =     iosapic_disable_edge_irq,
485         .enable =       iosapic_enable_edge_irq,
486         .disable =      iosapic_disable_edge_irq,
487         .ack =          iosapic_ack_edge_irq,
488         .end =          iosapic_end_edge_irq,
489         .mask =         mask_irq,
490         .unmask =       unmask_irq,
491         .set_affinity = iosapic_set_affinity
492 };
493
494 unsigned int
495 iosapic_version (char __iomem *addr)
496 {
497         /*
498          * IOSAPIC Version Register return 32 bit structure like:
499          * {
500          *      unsigned int version   : 8;
501          *      unsigned int reserved1 : 8;
502          *      unsigned int max_redir : 8;
503          *      unsigned int reserved2 : 8;
504          * }
505          */
506         return __iosapic_read(addr, IOSAPIC_VERSION);
507 }
508
509 static int iosapic_find_sharable_irq(unsigned long trigger, unsigned long pol)
510 {
511         int i, irq = -ENOSPC, min_count = -1;
512         struct iosapic_intr_info *info;
513
514         /*
515          * shared vectors for edge-triggered interrupts are not
516          * supported yet
517          */
518         if (trigger == IOSAPIC_EDGE)
519                 return -EINVAL;
520
521         for (i = 0; i <= NR_IRQS; i++) {
522                 info = &iosapic_intr_info[i];
523                 if (info->trigger == trigger && info->polarity == pol &&
524                     (info->dmode == IOSAPIC_FIXED ||
525                      info->dmode == IOSAPIC_LOWEST_PRIORITY) &&
526                     can_request_irq(i, IRQF_SHARED)) {
527                         if (min_count == -1 || info->count < min_count) {
528                                 irq = i;
529                                 min_count = info->count;
530                         }
531                 }
532         }
533         return irq;
534 }
535
536 /*
537  * if the given vector is already owned by other,
538  *  assign a new vector for the other and make the vector available
539  */
540 static void __init
541 iosapic_reassign_vector (int irq)
542 {
543         int new_irq;
544
545         if (iosapic_intr_info[irq].count) {
546                 new_irq = create_irq();
547                 if (new_irq < 0)
548                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
549                 printk(KERN_INFO "Reassigning vector %d to %d\n",
550                        irq_to_vector(irq), irq_to_vector(new_irq));
551                 memcpy(&iosapic_intr_info[new_irq], &iosapic_intr_info[irq],
552                        sizeof(struct iosapic_intr_info));
553                 INIT_LIST_HEAD(&iosapic_intr_info[new_irq].rtes);
554                 list_move(iosapic_intr_info[irq].rtes.next,
555                           &iosapic_intr_info[new_irq].rtes);
556                 memset(&iosapic_intr_info[irq], 0,
557                        sizeof(struct iosapic_intr_info));
558                 iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
559                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
560         }
561 }
562
563 static struct iosapic_rte_info * __init_refok iosapic_alloc_rte (void)
564 {
565         int i;
566         struct iosapic_rte_info *rte;
567         int preallocated = 0;
568
569         if (!iosapic_kmalloc_ok && list_empty(&free_rte_list)) {
570                 rte = alloc_bootmem(sizeof(struct iosapic_rte_info) *
571                                     NR_PREALLOCATE_RTE_ENTRIES);
572                 if (!rte)
573                         return NULL;
574                 for (i = 0; i < NR_PREALLOCATE_RTE_ENTRIES; i++, rte++)
575                         list_add(&rte->rte_list, &free_rte_list);
576         }
577
578         if (!list_empty(&free_rte_list)) {
579                 rte = list_entry(free_rte_list.next, struct iosapic_rte_info,
580                                  rte_list);
581                 list_del(&rte->rte_list);
582                 preallocated++;
583         } else {
584                 rte = kmalloc(sizeof(struct iosapic_rte_info), GFP_ATOMIC);
585                 if (!rte)
586                         return NULL;
587         }
588
589         memset(rte, 0, sizeof(struct iosapic_rte_info));
590         if (preallocated)
591                 rte->flags |= RTE_PREALLOCATED;
592
593         return rte;
594 }
595
596 static inline int irq_is_shared (int irq)
597 {
598         return (iosapic_intr_info[irq].count > 1);
599 }
600
601 static int
602 register_intr (unsigned int gsi, int irq, unsigned char delivery,
603                unsigned long polarity, unsigned long trigger)
604 {
605         irq_desc_t *idesc;
606         struct hw_interrupt_type *irq_type;
607         int index;
608         struct iosapic_rte_info *rte;
609
610         index = find_iosapic(gsi);
611         if (index < 0) {
612                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
613                        __FUNCTION__, gsi);
614                 return -ENODEV;
615         }
616
617         rte = find_rte(irq, gsi);
618         if (!rte) {
619                 rte = iosapic_alloc_rte();
620                 if (!rte) {
621                         printk(KERN_WARNING "%s: cannot allocate memory\n",
622                                __FUNCTION__);
623                         return -ENOMEM;
624                 }
625
626                 rte->iosapic    = &iosapic_lists[index];
627                 rte->rte_index  = gsi - rte->iosapic->gsi_base;
628                 rte->refcnt++;
629                 list_add_tail(&rte->rte_list, &iosapic_intr_info[irq].rtes);
630                 iosapic_intr_info[irq].count++;
631                 iosapic_lists[index].rtes_inuse++;
632         }
633         else if (rte->refcnt == NO_REF_RTE) {
634                 struct iosapic_intr_info *info = &iosapic_intr_info[irq];
635                 if (info->count > 0 &&
636                     (info->trigger != trigger || info->polarity != polarity)){
637                         printk (KERN_WARNING
638                                 "%s: cannot override the interrupt\n",
639                                 __FUNCTION__);
640                         return -EINVAL;
641                 }
642                 rte->refcnt++;
643                 iosapic_intr_info[irq].count++;
644                 iosapic_lists[index].rtes_inuse++;
645         }
646
647         iosapic_intr_info[irq].polarity = polarity;
648         iosapic_intr_info[irq].dmode    = delivery;
649         iosapic_intr_info[irq].trigger  = trigger;
650
651         if (trigger == IOSAPIC_EDGE)
652                 irq_type = &irq_type_iosapic_edge;
653         else
654                 irq_type = &irq_type_iosapic_level;
655
656         idesc = irq_desc + irq;
657         if (idesc->chip != irq_type) {
658                 if (idesc->chip != &no_irq_type)
659                         printk(KERN_WARNING
660                                "%s: changing vector %d from %s to %s\n",
661                                __FUNCTION__, irq_to_vector(irq),
662                                idesc->chip->name, irq_type->name);
663                 idesc->chip = irq_type;
664         }
665         return 0;
666 }
667
668 static unsigned int
669 get_target_cpu (unsigned int gsi, int irq)
670 {
671 #ifdef CONFIG_SMP
672         static int cpu = -1;
673         extern int cpe_vector;
674         cpumask_t domain = irq_to_domain(irq);
675
676         /*
677          * In case of vector shared by multiple RTEs, all RTEs that
678          * share the vector need to use the same destination CPU.
679          */
680         if (iosapic_intr_info[irq].count)
681                 return iosapic_intr_info[irq].dest;
682
683         /*
684          * If the platform supports redirection via XTP, let it
685          * distribute interrupts.
686          */
687         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
688                 return cpu_physical_id(smp_processor_id());
689
690         /*
691          * Some interrupts (ACPI SCI, for instance) are registered
692          * before the BSP is marked as online.
693          */
694         if (!cpu_online(smp_processor_id()))
695                 return cpu_physical_id(smp_processor_id());
696
697 #ifdef CONFIG_ACPI
698         if (cpe_vector > 0 && irq_to_vector(irq) == IA64_CPEP_VECTOR)
699                 return get_cpei_target_cpu();
700 #endif
701
702 #ifdef CONFIG_NUMA
703         {
704                 int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
705                 cpumask_t cpu_mask;
706
707                 iosapic_index = find_iosapic(gsi);
708                 if (iosapic_index < 0 ||
709                     iosapic_lists[iosapic_index].node == MAX_NUMNODES)
710                         goto skip_numa_setup;
711
712                 cpu_mask = node_to_cpumask(iosapic_lists[iosapic_index].node);
713                 cpus_and(cpu_mask, cpu_mask, domain);
714                 for_each_cpu_mask(numa_cpu, cpu_mask) {
715                         if (!cpu_online(numa_cpu))
716                                 cpu_clear(numa_cpu, cpu_mask);
717                 }
718
719                 num_cpus = cpus_weight(cpu_mask);
720
721                 if (!num_cpus)
722                         goto skip_numa_setup;
723
724                 /* Use irq assignment to distribute across cpus in node */
725                 cpu_index = irq % num_cpus;
726
727                 for (numa_cpu = first_cpu(cpu_mask) ; i < cpu_index ; i++)
728                         numa_cpu = next_cpu(numa_cpu, cpu_mask);
729
730                 if (numa_cpu != NR_CPUS)
731                         return cpu_physical_id(numa_cpu);
732         }
733 skip_numa_setup:
734 #endif
735         /*
736          * Otherwise, round-robin interrupt vectors across all the
737          * processors.  (It'd be nice if we could be smarter in the
738          * case of NUMA.)
739          */
740         do {
741                 if (++cpu >= NR_CPUS)
742                         cpu = 0;
743         } while (!cpu_online(cpu) || !cpu_isset(cpu, domain));
744
745         return cpu_physical_id(cpu);
746 #else  /* CONFIG_SMP */
747         return cpu_physical_id(smp_processor_id());
748 #endif
749 }
750
751 static inline unsigned char choose_dmode(void)
752 {
753 #ifdef CONFIG_SMP
754         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
755                 return IOSAPIC_LOWEST_PRIORITY;
756 #endif
757         return IOSAPIC_FIXED;
758 }
759
760 /*
761  * ACPI can describe IOSAPIC interrupts via static tables and namespace
762  * methods.  This provides an interface to register those interrupts and
763  * program the IOSAPIC RTE.
764  */
765 int
766 iosapic_register_intr (unsigned int gsi,
767                        unsigned long polarity, unsigned long trigger)
768 {
769         int irq, mask = 1, err;
770         unsigned int dest;
771         unsigned long flags;
772         struct iosapic_rte_info *rte;
773         u32 low32;
774         unsigned char dmode;
775
776         /*
777          * If this GSI has already been registered (i.e., it's a
778          * shared interrupt, or we lost a race to register it),
779          * don't touch the RTE.
780          */
781         spin_lock_irqsave(&iosapic_lock, flags);
782         irq = __gsi_to_irq(gsi);
783         if (irq > 0) {
784                 rte = find_rte(irq, gsi);
785                 if(iosapic_intr_info[irq].count == 0) {
786                         assign_irq_vector(irq);
787                         dynamic_irq_init(irq);
788                 } else if (rte->refcnt != NO_REF_RTE) {
789                         rte->refcnt++;
790                         goto unlock_iosapic_lock;
791                 }
792         } else
793                 irq = create_irq();
794
795         /* If vector is running out, we try to find a sharable vector */
796         if (irq < 0) {
797                 irq = iosapic_find_sharable_irq(trigger, polarity);
798                 if (irq < 0)
799                         goto unlock_iosapic_lock;
800         }
801
802         spin_lock(&irq_desc[irq].lock);
803         dest = get_target_cpu(gsi, irq);
804         dmode = choose_dmode();
805         err = register_intr(gsi, irq, dmode, polarity, trigger);
806         if (err < 0) {
807                 spin_unlock(&irq_desc[irq].lock);
808                 irq = err;
809                 goto unlock_iosapic_lock;
810         }
811
812         /*
813          * If the vector is shared and already unmasked for other
814          * interrupt sources, don't mask it.
815          */
816         low32 = iosapic_intr_info[irq].low32;
817         if (irq_is_shared(irq) && !(low32 & IOSAPIC_MASK))
818                 mask = 0;
819         set_rte(gsi, irq, dest, mask);
820
821         printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
822                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
823                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
824                cpu_logical_id(dest), dest, irq_to_vector(irq));
825
826         spin_unlock(&irq_desc[irq].lock);
827  unlock_iosapic_lock:
828         spin_unlock_irqrestore(&iosapic_lock, flags);
829         return irq;
830 }
831
832 void
833 iosapic_unregister_intr (unsigned int gsi)
834 {
835         unsigned long flags;
836         int irq, index;
837         irq_desc_t *idesc;
838         u32 low32;
839         unsigned long trigger, polarity;
840         unsigned int dest;
841         struct iosapic_rte_info *rte;
842
843         /*
844          * If the irq associated with the gsi is not found,
845          * iosapic_unregister_intr() is unbalanced. We need to check
846          * this again after getting locks.
847          */
848         irq = gsi_to_irq(gsi);
849         if (irq < 0) {
850                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
851                        gsi);
852                 WARN_ON(1);
853                 return;
854         }
855
856         spin_lock_irqsave(&iosapic_lock, flags);
857         if ((rte = find_rte(irq, gsi)) == NULL) {
858                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
859                        gsi);
860                 WARN_ON(1);
861                 goto out;
862         }
863
864         if (--rte->refcnt > 0)
865                 goto out;
866
867         idesc = irq_desc + irq;
868         rte->refcnt = NO_REF_RTE;
869
870         /* Mask the interrupt */
871         low32 = iosapic_intr_info[irq].low32 | IOSAPIC_MASK;
872         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte->rte_index), low32);
873
874         iosapic_intr_info[irq].count--;
875         index = find_iosapic(gsi);
876         iosapic_lists[index].rtes_inuse--;
877         WARN_ON(iosapic_lists[index].rtes_inuse < 0);
878
879         trigger  = iosapic_intr_info[irq].trigger;
880         polarity = iosapic_intr_info[irq].polarity;
881         dest     = iosapic_intr_info[irq].dest;
882         printk(KERN_INFO
883                "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d unregistered\n",
884                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
885                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
886                cpu_logical_id(dest), dest, irq_to_vector(irq));
887
888         if (iosapic_intr_info[irq].count == 0) {
889 #ifdef CONFIG_SMP
890                 /* Clear affinity */
891                 cpus_setall(idesc->affinity);
892 #endif
893                 /* Clear the interrupt information */
894                 iosapic_intr_info[irq].dest = 0;
895                 iosapic_intr_info[irq].dmode = 0;
896                 iosapic_intr_info[irq].polarity = 0;
897                 iosapic_intr_info[irq].trigger = 0;
898                 iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
899
900                 /* Destroy and reserve IRQ */
901                 destroy_and_reserve_irq(irq);
902         }
903  out:
904         spin_unlock_irqrestore(&iosapic_lock, flags);
905 }
906
907 /*
908  * ACPI calls this when it finds an entry for a platform interrupt.
909  */
910 int __init
911 iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
912                                 int iosapic_vector, u16 eid, u16 id,
913                                 unsigned long polarity, unsigned long trigger)
914 {
915         static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
916         unsigned char delivery;
917         int irq, vector, mask = 0;
918         unsigned int dest = ((id << 8) | eid) & 0xffff;
919
920         switch (int_type) {
921               case ACPI_INTERRUPT_PMI:
922                 irq = vector = iosapic_vector;
923                 bind_irq_vector(irq, vector, CPU_MASK_ALL);
924                 /*
925                  * since PMI vector is alloc'd by FW(ACPI) not by kernel,
926                  * we need to make sure the vector is available
927                  */
928                 iosapic_reassign_vector(irq);
929                 delivery = IOSAPIC_PMI;
930                 break;
931               case ACPI_INTERRUPT_INIT:
932                 irq = create_irq();
933                 if (irq < 0)
934                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
935                 vector = irq_to_vector(irq);
936                 delivery = IOSAPIC_INIT;
937                 break;
938               case ACPI_INTERRUPT_CPEI:
939                 irq = vector = IA64_CPE_VECTOR;
940                 BUG_ON(bind_irq_vector(irq, vector, CPU_MASK_ALL));
941                 delivery = IOSAPIC_LOWEST_PRIORITY;
942                 mask = 1;
943                 break;
944               default:
945                 printk(KERN_ERR "%s: invalid int type 0x%x\n", __FUNCTION__,
946                        int_type);
947                 return -1;
948         }
949
950         register_intr(gsi, irq, delivery, polarity, trigger);
951
952         printk(KERN_INFO
953                "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x)"
954                " vector %d\n",
955                int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
956                int_type, gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
957                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
958                cpu_logical_id(dest), dest, vector);
959
960         set_rte(gsi, irq, dest, mask);
961         return vector;
962 }
963
964 /*
965  * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
966  */
967 void __devinit
968 iosapic_override_isa_irq (unsigned int isa_irq, unsigned int gsi,
969                           unsigned long polarity,
970                           unsigned long trigger)
971 {
972         int vector, irq;
973         unsigned int dest = cpu_physical_id(smp_processor_id());
974         unsigned char dmode;
975
976         irq = vector = isa_irq_to_vector(isa_irq);
977         BUG_ON(bind_irq_vector(irq, vector, CPU_MASK_ALL));
978         dmode = choose_dmode();
979         register_intr(gsi, irq, dmode, polarity, trigger);
980
981         DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
982             isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
983             polarity == IOSAPIC_POL_HIGH ? "high" : "low",
984             cpu_logical_id(dest), dest, vector);
985
986         set_rte(gsi, irq, dest, 1);
987 }
988
989 void __init
990 iosapic_system_init (int system_pcat_compat)
991 {
992         int irq;
993
994         for (irq = 0; irq < NR_IRQS; ++irq) {
995                 iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
996                 /* mark as unused */
997                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
998
999                 iosapic_intr_info[irq].count = 0;
1000         }
1001
1002         pcat_compat = system_pcat_compat;
1003         if (pcat_compat) {
1004                 /*
1005                  * Disable the compatibility mode interrupts (8259 style),
1006                  * needs IN/OUT support enabled.
1007                  */
1008                 printk(KERN_INFO
1009                        "%s: Disabling PC-AT compatible 8259 interrupts\n",
1010                        __FUNCTION__);
1011                 outb(0xff, 0xA1);
1012                 outb(0xff, 0x21);
1013         }
1014 }
1015
1016 static inline int
1017 iosapic_alloc (void)
1018 {
1019         int index;
1020
1021         for (index = 0; index < NR_IOSAPICS; index++)
1022                 if (!iosapic_lists[index].addr)
1023                         return index;
1024
1025         printk(KERN_WARNING "%s: failed to allocate iosapic\n", __FUNCTION__);
1026         return -1;
1027 }
1028
1029 static inline void
1030 iosapic_free (int index)
1031 {
1032         memset(&iosapic_lists[index], 0, sizeof(iosapic_lists[0]));
1033 }
1034
1035 static inline int
1036 iosapic_check_gsi_range (unsigned int gsi_base, unsigned int ver)
1037 {
1038         int index;
1039         unsigned int gsi_end, base, end;
1040
1041         /* check gsi range */
1042         gsi_end = gsi_base + ((ver >> 16) & 0xff);
1043         for (index = 0; index < NR_IOSAPICS; index++) {
1044                 if (!iosapic_lists[index].addr)
1045                         continue;
1046
1047                 base = iosapic_lists[index].gsi_base;
1048                 end  = base + iosapic_lists[index].num_rte - 1;
1049
1050                 if (gsi_end < base || end < gsi_base)
1051                         continue; /* OK */
1052
1053                 return -EBUSY;
1054         }
1055         return 0;
1056 }
1057
1058 int __devinit
1059 iosapic_init (unsigned long phys_addr, unsigned int gsi_base)
1060 {
1061         int num_rte, err, index;
1062         unsigned int isa_irq, ver;
1063         char __iomem *addr;
1064         unsigned long flags;
1065
1066         spin_lock_irqsave(&iosapic_lock, flags);
1067         index = find_iosapic(gsi_base);
1068         if (index >= 0) {
1069                 spin_unlock_irqrestore(&iosapic_lock, flags);
1070                 return -EBUSY;
1071         }
1072
1073         addr = ioremap(phys_addr, 0);
1074         ver = iosapic_version(addr);
1075         if ((err = iosapic_check_gsi_range(gsi_base, ver))) {
1076                 iounmap(addr);
1077                 spin_unlock_irqrestore(&iosapic_lock, flags);
1078                 return err;
1079         }
1080
1081         /*
1082          * The MAX_REDIR register holds the highest input pin number
1083          * (starting from 0).  We add 1 so that we can use it for
1084          * number of pins (= RTEs)
1085          */
1086         num_rte = ((ver >> 16) & 0xff) + 1;
1087
1088         index = iosapic_alloc();
1089         iosapic_lists[index].addr = addr;
1090         iosapic_lists[index].gsi_base = gsi_base;
1091         iosapic_lists[index].num_rte = num_rte;
1092 #ifdef CONFIG_NUMA
1093         iosapic_lists[index].node = MAX_NUMNODES;
1094 #endif
1095         spin_lock_init(&iosapic_lists[index].lock);
1096         spin_unlock_irqrestore(&iosapic_lock, flags);
1097
1098         if ((gsi_base == 0) && pcat_compat) {
1099                 /*
1100                  * Map the legacy ISA devices into the IOSAPIC data.  Some of
1101                  * these may get reprogrammed later on with data from the ACPI
1102                  * Interrupt Source Override table.
1103                  */
1104                 for (isa_irq = 0; isa_irq < 16; ++isa_irq)
1105                         iosapic_override_isa_irq(isa_irq, isa_irq,
1106                                                  IOSAPIC_POL_HIGH,
1107                                                  IOSAPIC_EDGE);
1108         }
1109         return 0;
1110 }
1111
1112 #ifdef CONFIG_HOTPLUG
1113 int
1114 iosapic_remove (unsigned int gsi_base)
1115 {
1116         int index, err = 0;
1117         unsigned long flags;
1118
1119         spin_lock_irqsave(&iosapic_lock, flags);
1120         index = find_iosapic(gsi_base);
1121         if (index < 0) {
1122                 printk(KERN_WARNING "%s: No IOSAPIC for GSI base %u\n",
1123                        __FUNCTION__, gsi_base);
1124                 goto out;
1125         }
1126
1127         if (iosapic_lists[index].rtes_inuse) {
1128                 err = -EBUSY;
1129                 printk(KERN_WARNING "%s: IOSAPIC for GSI base %u is busy\n",
1130                        __FUNCTION__, gsi_base);
1131                 goto out;
1132         }
1133
1134         iounmap(iosapic_lists[index].addr);
1135         iosapic_free(index);
1136  out:
1137         spin_unlock_irqrestore(&iosapic_lock, flags);
1138         return err;
1139 }
1140 #endif /* CONFIG_HOTPLUG */
1141
1142 #ifdef CONFIG_NUMA
1143 void __devinit
1144 map_iosapic_to_node(unsigned int gsi_base, int node)
1145 {
1146         int index;
1147
1148         index = find_iosapic(gsi_base);
1149         if (index < 0) {
1150                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1151                        __FUNCTION__, gsi_base);
1152                 return;
1153         }
1154         iosapic_lists[index].node = node;
1155         return;
1156 }
1157 #endif
1158
1159 static int __init iosapic_enable_kmalloc (void)
1160 {
1161         iosapic_kmalloc_ok = 1;
1162         return 0;
1163 }
1164 core_initcall (iosapic_enable_kmalloc);