Merge branch 'davem-fixes' of master.kernel.org:/pub/scm/linux/kernel/git/jgarzik...
[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 int
203 gsi_to_irq (unsigned int gsi)
204 {
205         unsigned long flags;
206         int irq;
207
208         spin_lock_irqsave(&iosapic_lock, flags);
209         irq = __gsi_to_irq(gsi);
210         spin_unlock_irqrestore(&iosapic_lock, flags);
211         return irq;
212 }
213
214 static struct iosapic_rte_info *find_rte(unsigned int irq, unsigned int gsi)
215 {
216         struct iosapic_rte_info *rte;
217
218         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
219                 if (rte->iosapic->gsi_base + rte->rte_index == gsi)
220                         return rte;
221         return NULL;
222 }
223
224 static void
225 set_rte (unsigned int gsi, unsigned int irq, unsigned int dest, int mask)
226 {
227         unsigned long pol, trigger, dmode;
228         u32 low32, high32;
229         int rte_index;
230         char redir;
231         struct iosapic_rte_info *rte;
232         ia64_vector vector = irq_to_vector(irq);
233
234         DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
235
236         rte = find_rte(irq, gsi);
237         if (!rte)
238                 return;         /* not an IOSAPIC interrupt */
239
240         rte_index = rte->rte_index;
241         pol     = iosapic_intr_info[irq].polarity;
242         trigger = iosapic_intr_info[irq].trigger;
243         dmode   = iosapic_intr_info[irq].dmode;
244
245         redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
246
247 #ifdef CONFIG_SMP
248         set_irq_affinity_info(irq, (int)(dest & 0xffff), redir);
249 #endif
250
251         low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
252                  (trigger << IOSAPIC_TRIGGER_SHIFT) |
253                  (dmode << IOSAPIC_DELIVERY_SHIFT) |
254                  ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
255                  vector);
256
257         /* dest contains both id and eid */
258         high32 = (dest << IOSAPIC_DEST_SHIFT);
259
260         iosapic_write(rte->iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
261         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
262         iosapic_intr_info[irq].low32 = low32;
263         iosapic_intr_info[irq].dest = dest;
264 }
265
266 static void
267 nop (unsigned int irq)
268 {
269         /* do nothing... */
270 }
271
272
273 #ifdef CONFIG_KEXEC
274 void
275 kexec_disable_iosapic(void)
276 {
277         struct iosapic_intr_info *info;
278         struct iosapic_rte_info *rte;
279         ia64_vector vec;
280         int irq;
281
282         for (irq = 0; irq < NR_IRQS; irq++) {
283                 info = &iosapic_intr_info[irq];
284                 vec = irq_to_vector(irq);
285                 list_for_each_entry(rte, &info->rtes,
286                                 rte_list) {
287                         iosapic_write(rte->iosapic,
288                                         IOSAPIC_RTE_LOW(rte->rte_index),
289                                         IOSAPIC_MASK|vec);
290                         iosapic_eoi(rte->iosapic->addr, vec);
291                 }
292         }
293 }
294 #endif
295
296 static void
297 mask_irq (unsigned int irq)
298 {
299         u32 low32;
300         int rte_index;
301         struct iosapic_rte_info *rte;
302
303         if (!iosapic_intr_info[irq].count)
304                 return;                 /* not an IOSAPIC interrupt! */
305
306         /* set only the mask bit */
307         low32 = iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
308         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
309                 rte_index = rte->rte_index;
310                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
311         }
312 }
313
314 static void
315 unmask_irq (unsigned int irq)
316 {
317         u32 low32;
318         int rte_index;
319         struct iosapic_rte_info *rte;
320
321         if (!iosapic_intr_info[irq].count)
322                 return;                 /* not an IOSAPIC interrupt! */
323
324         low32 = iosapic_intr_info[irq].low32 &= ~IOSAPIC_MASK;
325         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
326                 rte_index = rte->rte_index;
327                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
328         }
329 }
330
331
332 static void
333 iosapic_set_affinity (unsigned int irq, cpumask_t mask)
334 {
335 #ifdef CONFIG_SMP
336         u32 high32, low32;
337         int dest, rte_index;
338         int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
339         struct iosapic_rte_info *rte;
340         struct iosapic *iosapic;
341
342         irq &= (~IA64_IRQ_REDIRECTED);
343
344         cpus_and(mask, mask, cpu_online_map);
345         if (cpus_empty(mask))
346                 return;
347
348         if (irq_prepare_move(irq, first_cpu(mask)))
349                 return;
350
351         dest = cpu_physical_id(first_cpu(mask));
352
353         if (!iosapic_intr_info[irq].count)
354                 return;                 /* not an IOSAPIC interrupt */
355
356         set_irq_affinity_info(irq, dest, redir);
357
358         /* dest contains both id and eid */
359         high32 = dest << IOSAPIC_DEST_SHIFT;
360
361         low32 = iosapic_intr_info[irq].low32 & ~(7 << IOSAPIC_DELIVERY_SHIFT);
362         if (redir)
363                 /* change delivery mode to lowest priority */
364                 low32 |= (IOSAPIC_LOWEST_PRIORITY << IOSAPIC_DELIVERY_SHIFT);
365         else
366                 /* change delivery mode to fixed */
367                 low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
368         low32 &= IOSAPIC_VECTOR_MASK;
369         low32 |= irq_to_vector(irq);
370
371         iosapic_intr_info[irq].low32 = low32;
372         iosapic_intr_info[irq].dest = dest;
373         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
374                 iosapic = rte->iosapic;
375                 rte_index = rte->rte_index;
376                 iosapic_write(iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
377                 iosapic_write(iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
378         }
379 #endif
380 }
381
382 /*
383  * Handlers for level-triggered interrupts.
384  */
385
386 static unsigned int
387 iosapic_startup_level_irq (unsigned int irq)
388 {
389         unmask_irq(irq);
390         return 0;
391 }
392
393 static void
394 iosapic_end_level_irq (unsigned int irq)
395 {
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(irq_desc[irq].status & IRQ_MOVE_PENDING)) {
402                 do_unmask_irq = 1;
403                 mask_irq(irq);
404         }
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                 move_masked_irq(irq);
411                 unmask_irq(irq);
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
420 static struct irq_chip irq_type_iosapic_level = {
421         .name =         "IO-SAPIC-level",
422         .startup =      iosapic_startup_level_irq,
423         .shutdown =     iosapic_shutdown_level_irq,
424         .enable =       iosapic_enable_level_irq,
425         .disable =      iosapic_disable_level_irq,
426         .ack =          iosapic_ack_level_irq,
427         .end =          iosapic_end_level_irq,
428         .mask =         mask_irq,
429         .unmask =       unmask_irq,
430         .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 (unsigned int irq)
439 {
440         unmask_irq(irq);
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 (unsigned int irq)
451 {
452         irq_desc_t *idesc = irq_desc + irq;
453
454         irq_complete_move(irq);
455         move_native_irq(irq);
456         /*
457          * Once we have recorded IRQ_PENDING already, we can mask the
458          * interrupt for real. This prevents IRQ storms from unhandled
459          * devices.
460          */
461         if ((idesc->status & (IRQ_PENDING|IRQ_DISABLED)) ==
462             (IRQ_PENDING|IRQ_DISABLED))
463                 mask_irq(irq);
464 }
465
466 #define iosapic_enable_edge_irq         unmask_irq
467 #define iosapic_disable_edge_irq        nop
468 #define iosapic_end_edge_irq            nop
469
470 static struct irq_chip irq_type_iosapic_edge = {
471         .name =         "IO-SAPIC-edge",
472         .startup =      iosapic_startup_edge_irq,
473         .shutdown =     iosapic_disable_edge_irq,
474         .enable =       iosapic_enable_edge_irq,
475         .disable =      iosapic_disable_edge_irq,
476         .ack =          iosapic_ack_edge_irq,
477         .end =          iosapic_end_edge_irq,
478         .mask =         mask_irq,
479         .unmask =       unmask_irq,
480         .set_affinity = iosapic_set_affinity
481 };
482
483 static unsigned int
484 iosapic_version (char __iomem *addr)
485 {
486         /*
487          * IOSAPIC Version Register return 32 bit structure like:
488          * {
489          *      unsigned int version   : 8;
490          *      unsigned int reserved1 : 8;
491          *      unsigned int max_redir : 8;
492          *      unsigned int reserved2 : 8;
493          * }
494          */
495         return __iosapic_read(addr, IOSAPIC_VERSION);
496 }
497
498 static int iosapic_find_sharable_irq(unsigned long trigger, unsigned long pol)
499 {
500         int i, irq = -ENOSPC, min_count = -1;
501         struct iosapic_intr_info *info;
502
503         /*
504          * shared vectors for edge-triggered interrupts are not
505          * supported yet
506          */
507         if (trigger == IOSAPIC_EDGE)
508                 return -EINVAL;
509
510         for (i = 0; i <= NR_IRQS; i++) {
511                 info = &iosapic_intr_info[i];
512                 if (info->trigger == trigger && info->polarity == pol &&
513                     (info->dmode == IOSAPIC_FIXED ||
514                      info->dmode == IOSAPIC_LOWEST_PRIORITY) &&
515                     can_request_irq(i, IRQF_SHARED)) {
516                         if (min_count == -1 || info->count < min_count) {
517                                 irq = i;
518                                 min_count = info->count;
519                         }
520                 }
521         }
522         return irq;
523 }
524
525 /*
526  * if the given vector is already owned by other,
527  *  assign a new vector for the other and make the vector available
528  */
529 static void __init
530 iosapic_reassign_vector (int irq)
531 {
532         int new_irq;
533
534         if (iosapic_intr_info[irq].count) {
535                 new_irq = create_irq();
536                 if (new_irq < 0)
537                         panic("%s: out of interrupt vectors!\n", __func__);
538                 printk(KERN_INFO "Reassigning vector %d to %d\n",
539                        irq_to_vector(irq), irq_to_vector(new_irq));
540                 memcpy(&iosapic_intr_info[new_irq], &iosapic_intr_info[irq],
541                        sizeof(struct iosapic_intr_info));
542                 INIT_LIST_HEAD(&iosapic_intr_info[new_irq].rtes);
543                 list_move(iosapic_intr_info[irq].rtes.next,
544                           &iosapic_intr_info[new_irq].rtes);
545                 memset(&iosapic_intr_info[irq], 0,
546                        sizeof(struct iosapic_intr_info));
547                 iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
548                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
549         }
550 }
551
552 static struct iosapic_rte_info * __init_refok iosapic_alloc_rte (void)
553 {
554         int i;
555         struct iosapic_rte_info *rte;
556         int preallocated = 0;
557
558         if (!iosapic_kmalloc_ok && list_empty(&free_rte_list)) {
559                 rte = alloc_bootmem(sizeof(struct iosapic_rte_info) *
560                                     NR_PREALLOCATE_RTE_ENTRIES);
561                 if (!rte)
562                         return NULL;
563                 for (i = 0; i < NR_PREALLOCATE_RTE_ENTRIES; i++, rte++)
564                         list_add(&rte->rte_list, &free_rte_list);
565         }
566
567         if (!list_empty(&free_rte_list)) {
568                 rte = list_entry(free_rte_list.next, struct iosapic_rte_info,
569                                  rte_list);
570                 list_del(&rte->rte_list);
571                 preallocated++;
572         } else {
573                 rte = kmalloc(sizeof(struct iosapic_rte_info), GFP_ATOMIC);
574                 if (!rte)
575                         return NULL;
576         }
577
578         memset(rte, 0, sizeof(struct iosapic_rte_info));
579         if (preallocated)
580                 rte->flags |= RTE_PREALLOCATED;
581
582         return rte;
583 }
584
585 static inline int irq_is_shared (int irq)
586 {
587         return (iosapic_intr_info[irq].count > 1);
588 }
589
590 static int
591 register_intr (unsigned int gsi, int irq, unsigned char delivery,
592                unsigned long polarity, unsigned long trigger)
593 {
594         irq_desc_t *idesc;
595         struct hw_interrupt_type *irq_type;
596         int index;
597         struct iosapic_rte_info *rte;
598
599         index = find_iosapic(gsi);
600         if (index < 0) {
601                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
602                        __func__, gsi);
603                 return -ENODEV;
604         }
605
606         rte = find_rte(irq, gsi);
607         if (!rte) {
608                 rte = iosapic_alloc_rte();
609                 if (!rte) {
610                         printk(KERN_WARNING "%s: cannot allocate memory\n",
611                                __func__);
612                         return -ENOMEM;
613                 }
614
615                 rte->iosapic    = &iosapic_lists[index];
616                 rte->rte_index  = gsi - rte->iosapic->gsi_base;
617                 rte->refcnt++;
618                 list_add_tail(&rte->rte_list, &iosapic_intr_info[irq].rtes);
619                 iosapic_intr_info[irq].count++;
620                 iosapic_lists[index].rtes_inuse++;
621         }
622         else if (rte->refcnt == NO_REF_RTE) {
623                 struct iosapic_intr_info *info = &iosapic_intr_info[irq];
624                 if (info->count > 0 &&
625                     (info->trigger != trigger || info->polarity != polarity)){
626                         printk (KERN_WARNING
627                                 "%s: cannot override the interrupt\n",
628                                 __func__);
629                         return -EINVAL;
630                 }
631                 rte->refcnt++;
632                 iosapic_intr_info[irq].count++;
633                 iosapic_lists[index].rtes_inuse++;
634         }
635
636         iosapic_intr_info[irq].polarity = polarity;
637         iosapic_intr_info[irq].dmode    = delivery;
638         iosapic_intr_info[irq].trigger  = trigger;
639
640         if (trigger == IOSAPIC_EDGE)
641                 irq_type = &irq_type_iosapic_edge;
642         else
643                 irq_type = &irq_type_iosapic_level;
644
645         idesc = irq_desc + irq;
646         if (idesc->chip != irq_type) {
647                 if (idesc->chip != &no_irq_type)
648                         printk(KERN_WARNING
649                                "%s: changing vector %d from %s to %s\n",
650                                __func__, irq_to_vector(irq),
651                                idesc->chip->name, irq_type->name);
652                 idesc->chip = irq_type;
653         }
654         return 0;
655 }
656
657 static unsigned int
658 get_target_cpu (unsigned int gsi, int irq)
659 {
660 #ifdef CONFIG_SMP
661         static int cpu = -1;
662         extern int cpe_vector;
663         cpumask_t domain = irq_to_domain(irq);
664
665         /*
666          * In case of vector shared by multiple RTEs, all RTEs that
667          * share the vector need to use the same destination CPU.
668          */
669         if (iosapic_intr_info[irq].count)
670                 return iosapic_intr_info[irq].dest;
671
672         /*
673          * If the platform supports redirection via XTP, let it
674          * distribute interrupts.
675          */
676         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
677                 return cpu_physical_id(smp_processor_id());
678
679         /*
680          * Some interrupts (ACPI SCI, for instance) are registered
681          * before the BSP is marked as online.
682          */
683         if (!cpu_online(smp_processor_id()))
684                 return cpu_physical_id(smp_processor_id());
685
686 #ifdef CONFIG_ACPI
687         if (cpe_vector > 0 && irq_to_vector(irq) == IA64_CPEP_VECTOR)
688                 return get_cpei_target_cpu();
689 #endif
690
691 #ifdef CONFIG_NUMA
692         {
693                 int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
694                 cpumask_t cpu_mask;
695
696                 iosapic_index = find_iosapic(gsi);
697                 if (iosapic_index < 0 ||
698                     iosapic_lists[iosapic_index].node == MAX_NUMNODES)
699                         goto skip_numa_setup;
700
701                 cpu_mask = node_to_cpumask(iosapic_lists[iosapic_index].node);
702                 cpus_and(cpu_mask, cpu_mask, domain);
703                 for_each_cpu_mask(numa_cpu, cpu_mask) {
704                         if (!cpu_online(numa_cpu))
705                                 cpu_clear(numa_cpu, cpu_mask);
706                 }
707
708                 num_cpus = cpus_weight(cpu_mask);
709
710                 if (!num_cpus)
711                         goto skip_numa_setup;
712
713                 /* Use irq assignment to distribute across cpus in node */
714                 cpu_index = irq % num_cpus;
715
716                 for (numa_cpu = first_cpu(cpu_mask) ; i < cpu_index ; i++)
717                         numa_cpu = next_cpu(numa_cpu, cpu_mask);
718
719                 if (numa_cpu != NR_CPUS)
720                         return cpu_physical_id(numa_cpu);
721         }
722 skip_numa_setup:
723 #endif
724         /*
725          * Otherwise, round-robin interrupt vectors across all the
726          * processors.  (It'd be nice if we could be smarter in the
727          * case of NUMA.)
728          */
729         do {
730                 if (++cpu >= NR_CPUS)
731                         cpu = 0;
732         } while (!cpu_online(cpu) || !cpu_isset(cpu, domain));
733
734         return cpu_physical_id(cpu);
735 #else  /* CONFIG_SMP */
736         return cpu_physical_id(smp_processor_id());
737 #endif
738 }
739
740 static inline unsigned char choose_dmode(void)
741 {
742 #ifdef CONFIG_SMP
743         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
744                 return IOSAPIC_LOWEST_PRIORITY;
745 #endif
746         return IOSAPIC_FIXED;
747 }
748
749 /*
750  * ACPI can describe IOSAPIC interrupts via static tables and namespace
751  * methods.  This provides an interface to register those interrupts and
752  * program the IOSAPIC RTE.
753  */
754 int
755 iosapic_register_intr (unsigned int gsi,
756                        unsigned long polarity, unsigned long trigger)
757 {
758         int irq, mask = 1, err;
759         unsigned int dest;
760         unsigned long flags;
761         struct iosapic_rte_info *rte;
762         u32 low32;
763         unsigned char dmode;
764
765         /*
766          * If this GSI has already been registered (i.e., it's a
767          * shared interrupt, or we lost a race to register it),
768          * don't touch the RTE.
769          */
770         spin_lock_irqsave(&iosapic_lock, flags);
771         irq = __gsi_to_irq(gsi);
772         if (irq > 0) {
773                 rte = find_rte(irq, gsi);
774                 if(iosapic_intr_info[irq].count == 0) {
775                         assign_irq_vector(irq);
776                         dynamic_irq_init(irq);
777                 } else if (rte->refcnt != NO_REF_RTE) {
778                         rte->refcnt++;
779                         goto unlock_iosapic_lock;
780                 }
781         } else
782                 irq = create_irq();
783
784         /* If vector is running out, we try to find a sharable vector */
785         if (irq < 0) {
786                 irq = iosapic_find_sharable_irq(trigger, polarity);
787                 if (irq < 0)
788                         goto unlock_iosapic_lock;
789         }
790
791         spin_lock(&irq_desc[irq].lock);
792         dest = get_target_cpu(gsi, irq);
793         dmode = choose_dmode();
794         err = register_intr(gsi, irq, dmode, polarity, trigger);
795         if (err < 0) {
796                 spin_unlock(&irq_desc[irq].lock);
797                 irq = err;
798                 goto unlock_iosapic_lock;
799         }
800
801         /*
802          * If the vector is shared and already unmasked for other
803          * interrupt sources, don't mask it.
804          */
805         low32 = iosapic_intr_info[irq].low32;
806         if (irq_is_shared(irq) && !(low32 & IOSAPIC_MASK))
807                 mask = 0;
808         set_rte(gsi, irq, dest, mask);
809
810         printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
811                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
812                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
813                cpu_logical_id(dest), dest, irq_to_vector(irq));
814
815         spin_unlock(&irq_desc[irq].lock);
816  unlock_iosapic_lock:
817         spin_unlock_irqrestore(&iosapic_lock, flags);
818         return irq;
819 }
820
821 void
822 iosapic_unregister_intr (unsigned int gsi)
823 {
824         unsigned long flags;
825         int irq, index;
826         irq_desc_t *idesc;
827         u32 low32;
828         unsigned long trigger, polarity;
829         unsigned int dest;
830         struct iosapic_rte_info *rte;
831
832         /*
833          * If the irq associated with the gsi is not found,
834          * iosapic_unregister_intr() is unbalanced. We need to check
835          * this again after getting locks.
836          */
837         irq = gsi_to_irq(gsi);
838         if (irq < 0) {
839                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
840                        gsi);
841                 WARN_ON(1);
842                 return;
843         }
844
845         spin_lock_irqsave(&iosapic_lock, flags);
846         if ((rte = find_rte(irq, gsi)) == NULL) {
847                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
848                        gsi);
849                 WARN_ON(1);
850                 goto out;
851         }
852
853         if (--rte->refcnt > 0)
854                 goto out;
855
856         idesc = irq_desc + irq;
857         rte->refcnt = NO_REF_RTE;
858
859         /* Mask the interrupt */
860         low32 = iosapic_intr_info[irq].low32 | IOSAPIC_MASK;
861         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte->rte_index), low32);
862
863         iosapic_intr_info[irq].count--;
864         index = find_iosapic(gsi);
865         iosapic_lists[index].rtes_inuse--;
866         WARN_ON(iosapic_lists[index].rtes_inuse < 0);
867
868         trigger  = iosapic_intr_info[irq].trigger;
869         polarity = iosapic_intr_info[irq].polarity;
870         dest     = iosapic_intr_info[irq].dest;
871         printk(KERN_INFO
872                "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d unregistered\n",
873                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
874                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
875                cpu_logical_id(dest), dest, irq_to_vector(irq));
876
877         if (iosapic_intr_info[irq].count == 0) {
878 #ifdef CONFIG_SMP
879                 /* Clear affinity */
880                 cpus_setall(idesc->affinity);
881 #endif
882                 /* Clear the interrupt information */
883                 iosapic_intr_info[irq].dest = 0;
884                 iosapic_intr_info[irq].dmode = 0;
885                 iosapic_intr_info[irq].polarity = 0;
886                 iosapic_intr_info[irq].trigger = 0;
887                 iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
888
889                 /* Destroy and reserve IRQ */
890                 destroy_and_reserve_irq(irq);
891         }
892  out:
893         spin_unlock_irqrestore(&iosapic_lock, flags);
894 }
895
896 /*
897  * ACPI calls this when it finds an entry for a platform interrupt.
898  */
899 int __init
900 iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
901                                 int iosapic_vector, u16 eid, u16 id,
902                                 unsigned long polarity, unsigned long trigger)
903 {
904         static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
905         unsigned char delivery;
906         int irq, vector, mask = 0;
907         unsigned int dest = ((id << 8) | eid) & 0xffff;
908
909         switch (int_type) {
910               case ACPI_INTERRUPT_PMI:
911                 irq = vector = iosapic_vector;
912                 bind_irq_vector(irq, vector, CPU_MASK_ALL);
913                 /*
914                  * since PMI vector is alloc'd by FW(ACPI) not by kernel,
915                  * we need to make sure the vector is available
916                  */
917                 iosapic_reassign_vector(irq);
918                 delivery = IOSAPIC_PMI;
919                 break;
920               case ACPI_INTERRUPT_INIT:
921                 irq = create_irq();
922                 if (irq < 0)
923                         panic("%s: out of interrupt vectors!\n", __func__);
924                 vector = irq_to_vector(irq);
925                 delivery = IOSAPIC_INIT;
926                 break;
927               case ACPI_INTERRUPT_CPEI:
928                 irq = vector = IA64_CPE_VECTOR;
929                 BUG_ON(bind_irq_vector(irq, vector, CPU_MASK_ALL));
930                 delivery = IOSAPIC_FIXED;
931                 mask = 1;
932                 break;
933               default:
934                 printk(KERN_ERR "%s: invalid int type 0x%x\n", __func__,
935                        int_type);
936                 return -1;
937         }
938
939         register_intr(gsi, irq, delivery, polarity, trigger);
940
941         printk(KERN_INFO
942                "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x)"
943                " vector %d\n",
944                int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
945                int_type, gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
946                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
947                cpu_logical_id(dest), dest, vector);
948
949         set_rte(gsi, irq, dest, mask);
950         return vector;
951 }
952
953 /*
954  * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
955  */
956 void __devinit
957 iosapic_override_isa_irq (unsigned int isa_irq, unsigned int gsi,
958                           unsigned long polarity,
959                           unsigned long trigger)
960 {
961         int vector, irq;
962         unsigned int dest = cpu_physical_id(smp_processor_id());
963         unsigned char dmode;
964
965         irq = vector = isa_irq_to_vector(isa_irq);
966         BUG_ON(bind_irq_vector(irq, vector, CPU_MASK_ALL));
967         dmode = choose_dmode();
968         register_intr(gsi, irq, dmode, polarity, trigger);
969
970         DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
971             isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
972             polarity == IOSAPIC_POL_HIGH ? "high" : "low",
973             cpu_logical_id(dest), dest, vector);
974
975         set_rte(gsi, irq, dest, 1);
976 }
977
978 void __init
979 iosapic_system_init (int system_pcat_compat)
980 {
981         int irq;
982
983         for (irq = 0; irq < NR_IRQS; ++irq) {
984                 iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
985                 /* mark as unused */
986                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
987
988                 iosapic_intr_info[irq].count = 0;
989         }
990
991         pcat_compat = system_pcat_compat;
992         if (pcat_compat) {
993                 /*
994                  * Disable the compatibility mode interrupts (8259 style),
995                  * needs IN/OUT support enabled.
996                  */
997                 printk(KERN_INFO
998                        "%s: Disabling PC-AT compatible 8259 interrupts\n",
999                        __func__);
1000                 outb(0xff, 0xA1);
1001                 outb(0xff, 0x21);
1002         }
1003 }
1004
1005 static inline int
1006 iosapic_alloc (void)
1007 {
1008         int index;
1009
1010         for (index = 0; index < NR_IOSAPICS; index++)
1011                 if (!iosapic_lists[index].addr)
1012                         return index;
1013
1014         printk(KERN_WARNING "%s: failed to allocate iosapic\n", __func__);
1015         return -1;
1016 }
1017
1018 static inline void
1019 iosapic_free (int index)
1020 {
1021         memset(&iosapic_lists[index], 0, sizeof(iosapic_lists[0]));
1022 }
1023
1024 static inline int
1025 iosapic_check_gsi_range (unsigned int gsi_base, unsigned int ver)
1026 {
1027         int index;
1028         unsigned int gsi_end, base, end;
1029
1030         /* check gsi range */
1031         gsi_end = gsi_base + ((ver >> 16) & 0xff);
1032         for (index = 0; index < NR_IOSAPICS; index++) {
1033                 if (!iosapic_lists[index].addr)
1034                         continue;
1035
1036                 base = iosapic_lists[index].gsi_base;
1037                 end  = base + iosapic_lists[index].num_rte - 1;
1038
1039                 if (gsi_end < base || end < gsi_base)
1040                         continue; /* OK */
1041
1042                 return -EBUSY;
1043         }
1044         return 0;
1045 }
1046
1047 int __devinit
1048 iosapic_init (unsigned long phys_addr, unsigned int gsi_base)
1049 {
1050         int num_rte, err, index;
1051         unsigned int isa_irq, ver;
1052         char __iomem *addr;
1053         unsigned long flags;
1054
1055         spin_lock_irqsave(&iosapic_lock, flags);
1056         index = find_iosapic(gsi_base);
1057         if (index >= 0) {
1058                 spin_unlock_irqrestore(&iosapic_lock, flags);
1059                 return -EBUSY;
1060         }
1061
1062         addr = ioremap(phys_addr, 0);
1063         ver = iosapic_version(addr);
1064         if ((err = iosapic_check_gsi_range(gsi_base, ver))) {
1065                 iounmap(addr);
1066                 spin_unlock_irqrestore(&iosapic_lock, flags);
1067                 return err;
1068         }
1069
1070         /*
1071          * The MAX_REDIR register holds the highest input pin number
1072          * (starting from 0).  We add 1 so that we can use it for
1073          * number of pins (= RTEs)
1074          */
1075         num_rte = ((ver >> 16) & 0xff) + 1;
1076
1077         index = iosapic_alloc();
1078         iosapic_lists[index].addr = addr;
1079         iosapic_lists[index].gsi_base = gsi_base;
1080         iosapic_lists[index].num_rte = num_rte;
1081 #ifdef CONFIG_NUMA
1082         iosapic_lists[index].node = MAX_NUMNODES;
1083 #endif
1084         spin_lock_init(&iosapic_lists[index].lock);
1085         spin_unlock_irqrestore(&iosapic_lock, flags);
1086
1087         if ((gsi_base == 0) && pcat_compat) {
1088                 /*
1089                  * Map the legacy ISA devices into the IOSAPIC data.  Some of
1090                  * these may get reprogrammed later on with data from the ACPI
1091                  * Interrupt Source Override table.
1092                  */
1093                 for (isa_irq = 0; isa_irq < 16; ++isa_irq)
1094                         iosapic_override_isa_irq(isa_irq, isa_irq,
1095                                                  IOSAPIC_POL_HIGH,
1096                                                  IOSAPIC_EDGE);
1097         }
1098         return 0;
1099 }
1100
1101 #ifdef CONFIG_HOTPLUG
1102 int
1103 iosapic_remove (unsigned int gsi_base)
1104 {
1105         int index, err = 0;
1106         unsigned long flags;
1107
1108         spin_lock_irqsave(&iosapic_lock, flags);
1109         index = find_iosapic(gsi_base);
1110         if (index < 0) {
1111                 printk(KERN_WARNING "%s: No IOSAPIC for GSI base %u\n",
1112                        __func__, gsi_base);
1113                 goto out;
1114         }
1115
1116         if (iosapic_lists[index].rtes_inuse) {
1117                 err = -EBUSY;
1118                 printk(KERN_WARNING "%s: IOSAPIC for GSI base %u is busy\n",
1119                        __func__, gsi_base);
1120                 goto out;
1121         }
1122
1123         iounmap(iosapic_lists[index].addr);
1124         iosapic_free(index);
1125  out:
1126         spin_unlock_irqrestore(&iosapic_lock, flags);
1127         return err;
1128 }
1129 #endif /* CONFIG_HOTPLUG */
1130
1131 #ifdef CONFIG_NUMA
1132 void __devinit
1133 map_iosapic_to_node(unsigned int gsi_base, int node)
1134 {
1135         int index;
1136
1137         index = find_iosapic(gsi_base);
1138         if (index < 0) {
1139                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1140                        __func__, gsi_base);
1141                 return;
1142         }
1143         iosapic_lists[index].node = node;
1144         return;
1145 }
1146 #endif
1147
1148 static int __init iosapic_enable_kmalloc (void)
1149 {
1150         iosapic_kmalloc_ok = 1;
1151         return 0;
1152 }
1153 core_initcall (iosapic_enable_kmalloc);