]> rtime.felk.cvut.cz Git - linux-imx.git/blob - arch/ia64/kernel/iosapic.c
[IA64] Use dynamic irq for iosapic interrupts
[linux-imx.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 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 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         unsigned int    flags;          /* flags */
136         struct iosapic  *iosapic;
137 } ____cacheline_aligned;
138
139 static struct iosapic_intr_info {
140         struct list_head rtes;          /* RTEs using this vector (empty =>
141                                          * not an IOSAPIC interrupt) */
142         int             count;          /* # of RTEs that shares this vector */
143         u32             low32;          /* current value of low word of
144                                          * Redirection table entry */
145         unsigned int    dest;           /* destination CPU physical ID */
146         unsigned char   dmode   : 3;    /* delivery mode (see iosapic.h) */
147         unsigned char   polarity: 1;    /* interrupt polarity
148                                          * (see iosapic.h) */
149         unsigned char   trigger : 1;    /* trigger mode (see iosapic.h) */
150 } iosapic_intr_info[IA64_NUM_VECTORS];
151
152 static unsigned char pcat_compat __devinitdata; /* 8259 compatibility flag */
153
154 static int iosapic_kmalloc_ok;
155 static LIST_HEAD(free_rte_list);
156
157 static inline void
158 iosapic_write(struct iosapic *iosapic, unsigned int reg, u32 val)
159 {
160         unsigned long flags;
161
162         spin_lock_irqsave(&iosapic->lock, flags);
163         __iosapic_write(iosapic->addr, reg, val);
164         spin_unlock_irqrestore(&iosapic->lock, flags);
165 }
166
167 /*
168  * Find an IOSAPIC associated with a GSI
169  */
170 static inline int
171 find_iosapic (unsigned int gsi)
172 {
173         int i;
174
175         for (i = 0; i < NR_IOSAPICS; i++) {
176                 if ((unsigned) (gsi - iosapic_lists[i].gsi_base) <
177                     iosapic_lists[i].num_rte)
178                         return i;
179         }
180
181         return -1;
182 }
183
184 static inline int
185 _gsi_to_vector (unsigned int gsi)
186 {
187         struct iosapic_intr_info *info;
188         struct iosapic_rte_info *rte;
189
190         for (info = iosapic_intr_info; info <
191                      iosapic_intr_info + IA64_NUM_VECTORS; ++info)
192                 list_for_each_entry(rte, &info->rtes, rte_list)
193                         if (rte->iosapic->gsi_base + rte->rte_index == gsi)
194                                 return info - iosapic_intr_info;
195         return -1;
196 }
197
198 /*
199  * Translate GSI number to the corresponding IA-64 interrupt vector.  If no
200  * entry exists, return -1.
201  */
202 inline int
203 gsi_to_vector (unsigned int gsi)
204 {
205         return _gsi_to_vector(gsi);
206 }
207
208 int
209 gsi_to_irq (unsigned int gsi)
210 {
211         unsigned long flags;
212         int irq;
213         /*
214          * XXX fix me: this assumes an identity mapping between IA-64 vector
215          * and Linux irq numbers...
216          */
217         spin_lock_irqsave(&iosapic_lock, flags);
218         irq = _gsi_to_vector(gsi);
219         spin_unlock_irqrestore(&iosapic_lock, flags);
220
221         return irq;
222 }
223
224 static struct iosapic_rte_info *gsi_vector_to_rte(unsigned int gsi,
225                                                   unsigned int vec)
226 {
227         struct iosapic_rte_info *rte;
228
229         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list)
230                 if (rte->iosapic->gsi_base + rte->rte_index == gsi)
231                         return rte;
232         return NULL;
233 }
234
235 static void
236 set_rte (unsigned int gsi, unsigned int vector, unsigned int dest, int mask)
237 {
238         unsigned long pol, trigger, dmode;
239         u32 low32, high32;
240         int rte_index;
241         char redir;
242         struct iosapic_rte_info *rte;
243
244         DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
245
246         rte = gsi_vector_to_rte(gsi, vector);
247         if (!rte)
248                 return;         /* not an IOSAPIC interrupt */
249
250         rte_index = rte->rte_index;
251         pol     = iosapic_intr_info[vector].polarity;
252         trigger = iosapic_intr_info[vector].trigger;
253         dmode   = iosapic_intr_info[vector].dmode;
254
255         redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
256
257 #ifdef CONFIG_SMP
258         {
259                 unsigned int irq;
260
261                 for (irq = 0; irq < NR_IRQS; ++irq)
262                         if (irq_to_vector(irq) == vector) {
263                                 set_irq_affinity_info(irq,
264                                                       (int)(dest & 0xffff),
265                                                       redir);
266                                 break;
267                         }
268         }
269 #endif
270
271         low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
272                  (trigger << IOSAPIC_TRIGGER_SHIFT) |
273                  (dmode << IOSAPIC_DELIVERY_SHIFT) |
274                  ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
275                  vector);
276
277         /* dest contains both id and eid */
278         high32 = (dest << IOSAPIC_DEST_SHIFT);
279
280         iosapic_write(rte->iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
281         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
282         iosapic_intr_info[vector].low32 = low32;
283         iosapic_intr_info[vector].dest = dest;
284 }
285
286 static void
287 nop (unsigned int irq)
288 {
289         /* do nothing... */
290 }
291
292
293 #ifdef CONFIG_KEXEC
294 void
295 kexec_disable_iosapic(void)
296 {
297         struct iosapic_intr_info *info;
298         struct iosapic_rte_info *rte;
299         u8 vec = 0;
300         for (info = iosapic_intr_info; info <
301                         iosapic_intr_info + IA64_NUM_VECTORS; ++info, ++vec) {
302                 list_for_each_entry(rte, &info->rtes,
303                                 rte_list) {
304                         iosapic_write(rte->iosapic,
305                                         IOSAPIC_RTE_LOW(rte->rte_index),
306                                         IOSAPIC_MASK|vec);
307                         iosapic_eoi(rte->iosapic->addr, vec);
308                 }
309         }
310 }
311 #endif
312
313 static void
314 mask_irq (unsigned int irq)
315 {
316         u32 low32;
317         int rte_index;
318         ia64_vector vec = irq_to_vector(irq);
319         struct iosapic_rte_info *rte;
320
321         if (list_empty(&iosapic_intr_info[vec].rtes))
322                 return;                 /* not an IOSAPIC interrupt! */
323
324         /* set only the mask bit */
325         low32 = iosapic_intr_info[vec].low32 |= IOSAPIC_MASK;
326         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list) {
327                 rte_index = rte->rte_index;
328                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
329         }
330 }
331
332 static void
333 unmask_irq (unsigned int irq)
334 {
335         u32 low32;
336         int rte_index;
337         ia64_vector vec = irq_to_vector(irq);
338         struct iosapic_rte_info *rte;
339
340         if (list_empty(&iosapic_intr_info[vec].rtes))
341                 return;                 /* not an IOSAPIC interrupt! */
342
343         low32 = iosapic_intr_info[vec].low32 &= ~IOSAPIC_MASK;
344         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list) {
345                 rte_index = rte->rte_index;
346                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
347         }
348 }
349
350
351 static void
352 iosapic_set_affinity (unsigned int irq, cpumask_t mask)
353 {
354 #ifdef CONFIG_SMP
355         u32 high32, low32;
356         int dest, rte_index;
357         int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
358         ia64_vector vec;
359         struct iosapic_rte_info *rte;
360         struct iosapic *iosapic;
361
362         irq &= (~IA64_IRQ_REDIRECTED);
363         vec = irq_to_vector(irq);
364
365         if (cpus_empty(mask))
366                 return;
367
368         dest = cpu_physical_id(first_cpu(mask));
369
370         if (list_empty(&iosapic_intr_info[vec].rtes))
371                 return;                 /* not an IOSAPIC interrupt */
372
373         set_irq_affinity_info(irq, dest, redir);
374
375         /* dest contains both id and eid */
376         high32 = dest << IOSAPIC_DEST_SHIFT;
377
378         low32 = iosapic_intr_info[vec].low32 & ~(7 << IOSAPIC_DELIVERY_SHIFT);
379         if (redir)
380                 /* change delivery mode to lowest priority */
381                 low32 |= (IOSAPIC_LOWEST_PRIORITY << IOSAPIC_DELIVERY_SHIFT);
382         else
383                 /* change delivery mode to fixed */
384                 low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
385
386         iosapic_intr_info[vec].low32 = low32;
387         iosapic_intr_info[vec].dest = dest;
388         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list) {
389                 iosapic = rte->iosapic;
390                 rte_index = rte->rte_index;
391                 iosapic_write(iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
392                 iosapic_write(iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
393         }
394 #endif
395 }
396
397 /*
398  * Handlers for level-triggered interrupts.
399  */
400
401 static unsigned int
402 iosapic_startup_level_irq (unsigned int irq)
403 {
404         unmask_irq(irq);
405         return 0;
406 }
407
408 static void
409 iosapic_end_level_irq (unsigned int irq)
410 {
411         ia64_vector vec = irq_to_vector(irq);
412         struct iosapic_rte_info *rte;
413
414         move_native_irq(irq);
415         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list)
416                 iosapic_eoi(rte->iosapic->addr, vec);
417 }
418
419 #define iosapic_shutdown_level_irq      mask_irq
420 #define iosapic_enable_level_irq        unmask_irq
421 #define iosapic_disable_level_irq       mask_irq
422 #define iosapic_ack_level_irq           nop
423
424 struct irq_chip irq_type_iosapic_level = {
425         .name =         "IO-SAPIC-level",
426         .startup =      iosapic_startup_level_irq,
427         .shutdown =     iosapic_shutdown_level_irq,
428         .enable =       iosapic_enable_level_irq,
429         .disable =      iosapic_disable_level_irq,
430         .ack =          iosapic_ack_level_irq,
431         .end =          iosapic_end_level_irq,
432         .mask =         mask_irq,
433         .unmask =       unmask_irq,
434         .set_affinity = iosapic_set_affinity
435 };
436
437 /*
438  * Handlers for edge-triggered interrupts.
439  */
440
441 static unsigned int
442 iosapic_startup_edge_irq (unsigned int irq)
443 {
444         unmask_irq(irq);
445         /*
446          * IOSAPIC simply drops interrupts pended while the
447          * corresponding pin was masked, so we can't know if an
448          * interrupt is pending already.  Let's hope not...
449          */
450         return 0;
451 }
452
453 static void
454 iosapic_ack_edge_irq (unsigned int irq)
455 {
456         irq_desc_t *idesc = irq_desc + irq;
457
458         move_native_irq(irq);
459         /*
460          * Once we have recorded IRQ_PENDING already, we can mask the
461          * interrupt for real. This prevents IRQ storms from unhandled
462          * devices.
463          */
464         if ((idesc->status & (IRQ_PENDING|IRQ_DISABLED)) ==
465             (IRQ_PENDING|IRQ_DISABLED))
466                 mask_irq(irq);
467 }
468
469 #define iosapic_enable_edge_irq         unmask_irq
470 #define iosapic_disable_edge_irq        nop
471 #define iosapic_end_edge_irq            nop
472
473 struct irq_chip irq_type_iosapic_edge = {
474         .name =         "IO-SAPIC-edge",
475         .startup =      iosapic_startup_edge_irq,
476         .shutdown =     iosapic_disable_edge_irq,
477         .enable =       iosapic_enable_edge_irq,
478         .disable =      iosapic_disable_edge_irq,
479         .ack =          iosapic_ack_edge_irq,
480         .end =          iosapic_end_edge_irq,
481         .mask =         mask_irq,
482         .unmask =       unmask_irq,
483         .set_affinity = iosapic_set_affinity
484 };
485
486 unsigned int
487 iosapic_version (char __iomem *addr)
488 {
489         /*
490          * IOSAPIC Version Register return 32 bit structure like:
491          * {
492          *      unsigned int version   : 8;
493          *      unsigned int reserved1 : 8;
494          *      unsigned int max_redir : 8;
495          *      unsigned int reserved2 : 8;
496          * }
497          */
498         return __iosapic_read(addr, IOSAPIC_VERSION);
499 }
500
501 static int iosapic_find_sharable_vector (unsigned long trigger,
502                                          unsigned long pol)
503 {
504         int i, vector = -ENOSPC, min_count = -1;
505         struct iosapic_intr_info *info;
506
507         /*
508          * shared vectors for edge-triggered interrupts are not
509          * supported yet
510          */
511         if (trigger == IOSAPIC_EDGE)
512                 return -EINVAL;
513
514         for (i = IA64_FIRST_DEVICE_VECTOR; i <= IA64_LAST_DEVICE_VECTOR; i++) {
515                 info = &iosapic_intr_info[i];
516                 if (info->trigger == trigger && info->polarity == pol &&
517                     (info->dmode == IOSAPIC_FIXED || info->dmode ==
518                      IOSAPIC_LOWEST_PRIORITY)) {
519                         if (min_count == -1 || info->count < min_count) {
520                                 vector = i;
521                                 min_count = info->count;
522                         }
523                 }
524         }
525
526         return vector;
527 }
528
529 /*
530  * if the given vector is already owned by other,
531  *  assign a new vector for the other and make the vector available
532  */
533 static void __init
534 iosapic_reassign_vector (int vector)
535 {
536         int irq, new_vector;
537
538         if (!list_empty(&iosapic_intr_info[vector].rtes)) {
539                 irq = create_irq();
540                 if (irq < 0)
541                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
542                 new_vector = irq_to_vector(irq);
543                 printk(KERN_INFO "Reassigning vector %d to %d\n",
544                        vector, new_vector);
545                 memcpy(&iosapic_intr_info[new_vector], &iosapic_intr_info[vector],
546                        sizeof(struct iosapic_intr_info));
547                 INIT_LIST_HEAD(&iosapic_intr_info[new_vector].rtes);
548                 list_move(iosapic_intr_info[vector].rtes.next,
549                           &iosapic_intr_info[new_vector].rtes);
550                 memset(&iosapic_intr_info[vector], 0,
551                        sizeof(struct iosapic_intr_info));
552                 iosapic_intr_info[vector].low32 = IOSAPIC_MASK;
553                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
554         }
555 }
556
557 static struct iosapic_rte_info *iosapic_alloc_rte (void)
558 {
559         int i;
560         struct iosapic_rte_info *rte;
561         int preallocated = 0;
562
563         if (!iosapic_kmalloc_ok && list_empty(&free_rte_list)) {
564                 rte = alloc_bootmem(sizeof(struct iosapic_rte_info) *
565                                     NR_PREALLOCATE_RTE_ENTRIES);
566                 if (!rte)
567                         return NULL;
568                 for (i = 0; i < NR_PREALLOCATE_RTE_ENTRIES; i++, rte++)
569                         list_add(&rte->rte_list, &free_rte_list);
570         }
571
572         if (!list_empty(&free_rte_list)) {
573                 rte = list_entry(free_rte_list.next, struct iosapic_rte_info,
574                                  rte_list);
575                 list_del(&rte->rte_list);
576                 preallocated++;
577         } else {
578                 rte = kmalloc(sizeof(struct iosapic_rte_info), GFP_ATOMIC);
579                 if (!rte)
580                         return NULL;
581         }
582
583         memset(rte, 0, sizeof(struct iosapic_rte_info));
584         if (preallocated)
585                 rte->flags |= RTE_PREALLOCATED;
586
587         return rte;
588 }
589
590 static void iosapic_free_rte (struct iosapic_rte_info *rte)
591 {
592         if (rte->flags & RTE_PREALLOCATED)
593                 list_add_tail(&rte->rte_list, &free_rte_list);
594         else
595                 kfree(rte);
596 }
597
598 static inline int vector_is_shared (int vector)
599 {
600         return (iosapic_intr_info[vector].count > 1);
601 }
602
603 static int
604 register_intr (unsigned int gsi, int vector, unsigned char delivery,
605                unsigned long polarity, unsigned long trigger)
606 {
607         irq_desc_t *idesc;
608         struct hw_interrupt_type *irq_type;
609         int index;
610         struct iosapic_rte_info *rte;
611
612         index = find_iosapic(gsi);
613         if (index < 0) {
614                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
615                        __FUNCTION__, gsi);
616                 return -ENODEV;
617         }
618
619         rte = gsi_vector_to_rte(gsi, vector);
620         if (!rte) {
621                 rte = iosapic_alloc_rte();
622                 if (!rte) {
623                         printk(KERN_WARNING "%s: cannot allocate memory\n",
624                                __FUNCTION__);
625                         return -ENOMEM;
626                 }
627
628                 rte->iosapic    = &iosapic_lists[index];
629                 rte->rte_index  = gsi - rte->iosapic->gsi_base;
630                 rte->refcnt++;
631                 list_add_tail(&rte->rte_list, &iosapic_intr_info[vector].rtes);
632                 iosapic_intr_info[vector].count++;
633                 iosapic_lists[index].rtes_inuse++;
634         }
635         else if (vector_is_shared(vector)) {
636                 struct iosapic_intr_info *info = &iosapic_intr_info[vector];
637                 if (info->trigger != trigger || info->polarity != polarity) {
638                         printk (KERN_WARNING
639                                 "%s: cannot override the interrupt\n",
640                                 __FUNCTION__);
641                         return -EINVAL;
642                 }
643         }
644
645         iosapic_intr_info[vector].polarity = polarity;
646         iosapic_intr_info[vector].dmode    = delivery;
647         iosapic_intr_info[vector].trigger  = trigger;
648
649         if (trigger == IOSAPIC_EDGE)
650                 irq_type = &irq_type_iosapic_edge;
651         else
652                 irq_type = &irq_type_iosapic_level;
653
654         idesc = irq_desc + vector;
655         if (idesc->chip != irq_type) {
656                 if (idesc->chip != &no_irq_type)
657                         printk(KERN_WARNING
658                                "%s: changing vector %d from %s to %s\n",
659                                __FUNCTION__, vector,
660                                idesc->chip->name, irq_type->name);
661                 idesc->chip = irq_type;
662         }
663         return 0;
664 }
665
666 static unsigned int
667 get_target_cpu (unsigned int gsi, int vector)
668 {
669 #ifdef CONFIG_SMP
670         static int cpu = -1;
671         extern int cpe_vector;
672
673         /*
674          * In case of vector shared by multiple RTEs, all RTEs that
675          * share the vector need to use the same destination CPU.
676          */
677         if (!list_empty(&iosapic_intr_info[vector].rtes))
678                 return iosapic_intr_info[vector].dest;
679
680         /*
681          * If the platform supports redirection via XTP, let it
682          * distribute interrupts.
683          */
684         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
685                 return cpu_physical_id(smp_processor_id());
686
687         /*
688          * Some interrupts (ACPI SCI, for instance) are registered
689          * before the BSP is marked as online.
690          */
691         if (!cpu_online(smp_processor_id()))
692                 return cpu_physical_id(smp_processor_id());
693
694 #ifdef CONFIG_ACPI
695         if (cpe_vector > 0 && vector == IA64_CPEP_VECTOR)
696                 return get_cpei_target_cpu();
697 #endif
698
699 #ifdef CONFIG_NUMA
700         {
701                 int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
702                 cpumask_t cpu_mask;
703
704                 iosapic_index = find_iosapic(gsi);
705                 if (iosapic_index < 0 ||
706                     iosapic_lists[iosapic_index].node == MAX_NUMNODES)
707                         goto skip_numa_setup;
708
709                 cpu_mask = node_to_cpumask(iosapic_lists[iosapic_index].node);
710
711                 for_each_cpu_mask(numa_cpu, cpu_mask) {
712                         if (!cpu_online(numa_cpu))
713                                 cpu_clear(numa_cpu, cpu_mask);
714                 }
715
716                 num_cpus = cpus_weight(cpu_mask);
717
718                 if (!num_cpus)
719                         goto skip_numa_setup;
720
721                 /* Use vector assignment to distribute across cpus in node */
722                 cpu_index = vector % num_cpus;
723
724                 for (numa_cpu = first_cpu(cpu_mask) ; i < cpu_index ; i++)
725                         numa_cpu = next_cpu(numa_cpu, cpu_mask);
726
727                 if (numa_cpu != NR_CPUS)
728                         return cpu_physical_id(numa_cpu);
729         }
730 skip_numa_setup:
731 #endif
732         /*
733          * Otherwise, round-robin interrupt vectors across all the
734          * processors.  (It'd be nice if we could be smarter in the
735          * case of NUMA.)
736          */
737         do {
738                 if (++cpu >= NR_CPUS)
739                         cpu = 0;
740         } while (!cpu_online(cpu));
741
742         return cpu_physical_id(cpu);
743 #else  /* CONFIG_SMP */
744         return cpu_physical_id(smp_processor_id());
745 #endif
746 }
747
748 /*
749  * ACPI can describe IOSAPIC interrupts via static tables and namespace
750  * methods.  This provides an interface to register those interrupts and
751  * program the IOSAPIC RTE.
752  */
753 int
754 iosapic_register_intr (unsigned int gsi,
755                        unsigned long polarity, unsigned long trigger)
756 {
757         int irq, vector, mask = 1, err;
758         unsigned int dest;
759         unsigned long flags;
760         struct iosapic_rte_info *rte;
761         u32 low32;
762
763         /*
764          * If this GSI has already been registered (i.e., it's a
765          * shared interrupt, or we lost a race to register it),
766          * don't touch the RTE.
767          */
768         spin_lock_irqsave(&iosapic_lock, flags);
769         vector = gsi_to_vector(gsi);
770         if (vector > 0) {
771                 rte = gsi_vector_to_rte(gsi, vector);
772                 rte->refcnt++;
773                 goto unlock_iosapic_lock;
774         }
775
776         /* If vector is running out, we try to find a sharable vector */
777         irq = create_irq();
778         if (irq < 0) {
779                 vector = iosapic_find_sharable_vector(trigger, polarity);
780                 if (vector < 0)
781                         goto unlock_iosapic_lock;
782         } else
783                 vector = irq_to_vector(irq);
784
785         spin_lock(&irq_desc[vector].lock);
786         dest = get_target_cpu(gsi, vector);
787         err = register_intr(gsi, vector, IOSAPIC_LOWEST_PRIORITY,
788                             polarity, trigger);
789         if (err < 0) {
790                 vector = err;
791                 goto unlock_all;
792         }
793
794         /*
795          * If the vector is shared and already unmasked for other
796          * interrupt sources, don't mask it.
797          */
798         low32 = iosapic_intr_info[vector].low32;
799         if (vector_is_shared(vector) && !(low32 & IOSAPIC_MASK))
800                 mask = 0;
801         set_rte(gsi, vector, dest, mask);
802
803         printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
804                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
805                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
806                cpu_logical_id(dest), dest, vector);
807  unlock_all:
808         spin_unlock(&irq_desc[vector].lock);
809  unlock_iosapic_lock:
810         spin_unlock_irqrestore(&iosapic_lock, flags);
811         return vector;
812 }
813
814 void
815 iosapic_unregister_intr (unsigned int gsi)
816 {
817         unsigned long flags;
818         int irq, vector, index;
819         irq_desc_t *idesc;
820         u32 low32;
821         unsigned long trigger, polarity;
822         unsigned int dest;
823         struct iosapic_rte_info *rte;
824
825         /*
826          * If the irq associated with the gsi is not found,
827          * iosapic_unregister_intr() is unbalanced. We need to check
828          * this again after getting locks.
829          */
830         irq = gsi_to_irq(gsi);
831         if (irq < 0) {
832                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
833                        gsi);
834                 WARN_ON(1);
835                 return;
836         }
837         vector = irq_to_vector(irq);
838
839         spin_lock_irqsave(&iosapic_lock, flags);
840         if ((rte = gsi_vector_to_rte(gsi, vector)) == NULL) {
841                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
842                        gsi);
843                 WARN_ON(1);
844                 goto out;
845         }
846
847         if (--rte->refcnt > 0)
848                 goto out;
849
850         /* Remove the rte entry from the list */
851         idesc = irq_desc + irq;
852         spin_lock(&idesc->lock);
853         list_del(&rte->rte_list);
854         spin_unlock(&idesc->lock);
855
856         /* Mask the interrupt */
857         low32 = iosapic_intr_info[vector].low32 | IOSAPIC_MASK;
858         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte->rte_index), low32);
859
860         iosapic_intr_info[vector].count--;
861         iosapic_free_rte(rte);
862         index = find_iosapic(gsi);
863         iosapic_lists[index].rtes_inuse--;
864         WARN_ON(iosapic_lists[index].rtes_inuse < 0);
865
866         trigger  = iosapic_intr_info[vector].trigger;
867         polarity = iosapic_intr_info[vector].polarity;
868         dest     = iosapic_intr_info[vector].dest;
869         printk(KERN_INFO
870                "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d unregistered\n",
871                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
872                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
873                cpu_logical_id(dest), dest, vector);
874
875         if (list_empty(&iosapic_intr_info[vector].rtes)) {
876                 /* Sanity check */
877                 BUG_ON(iosapic_intr_info[vector].count);
878 #ifdef CONFIG_SMP
879                 /* Clear affinity */
880                 cpus_setall(idesc->affinity);
881 #endif
882                 /* Clear the interrupt information */
883                 memset(&iosapic_intr_info[vector], 0,
884                        sizeof(struct iosapic_intr_info));
885                 iosapic_intr_info[vector].low32 |= IOSAPIC_MASK;
886                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
887
888                 /* Destroy IRQ */
889                 destroy_irq(irq);
890         }
891  out:
892         spin_unlock_irqrestore(&iosapic_lock, flags);
893 }
894
895 /*
896  * ACPI calls this when it finds an entry for a platform interrupt.
897  */
898 int __init
899 iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
900                                 int iosapic_vector, u16 eid, u16 id,
901                                 unsigned long polarity, unsigned long trigger)
902 {
903         static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
904         unsigned char delivery;
905         int irq, vector, mask = 0;
906         unsigned int dest = ((id << 8) | eid) & 0xffff;
907
908         switch (int_type) {
909               case ACPI_INTERRUPT_PMI:
910                 vector = iosapic_vector;
911                 /*
912                  * since PMI vector is alloc'd by FW(ACPI) not by kernel,
913                  * we need to make sure the vector is available
914                  */
915                 iosapic_reassign_vector(vector);
916                 delivery = IOSAPIC_PMI;
917                 break;
918               case ACPI_INTERRUPT_INIT:
919                 irq = create_irq();
920                 if (irq < 0)
921                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
922                 vector = irq_to_vector(irq);
923                 delivery = IOSAPIC_INIT;
924                 break;
925               case ACPI_INTERRUPT_CPEI:
926                 vector = IA64_CPE_VECTOR;
927                 delivery = IOSAPIC_LOWEST_PRIORITY;
928                 mask = 1;
929                 break;
930               default:
931                 printk(KERN_ERR "%s: invalid int type 0x%x\n", __FUNCTION__,
932                        int_type);
933                 return -1;
934         }
935
936         register_intr(gsi, vector, delivery, polarity, trigger);
937
938         printk(KERN_INFO
939                "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x)"
940                " vector %d\n",
941                int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
942                int_type, gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
943                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
944                cpu_logical_id(dest), dest, vector);
945
946         set_rte(gsi, vector, dest, mask);
947         return vector;
948 }
949
950 /*
951  * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
952  */
953 void __devinit
954 iosapic_override_isa_irq (unsigned int isa_irq, unsigned int gsi,
955                           unsigned long polarity,
956                           unsigned long trigger)
957 {
958         int vector;
959         unsigned int dest = cpu_physical_id(smp_processor_id());
960
961         vector = isa_irq_to_vector(isa_irq);
962
963         register_intr(gsi, vector, IOSAPIC_LOWEST_PRIORITY, polarity, trigger);
964
965         DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
966             isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
967             polarity == IOSAPIC_POL_HIGH ? "high" : "low",
968             cpu_logical_id(dest), dest, vector);
969
970         set_rte(gsi, vector, dest, 1);
971 }
972
973 void __init
974 iosapic_system_init (int system_pcat_compat)
975 {
976         int vector;
977
978         for (vector = 0; vector < IA64_NUM_VECTORS; ++vector) {
979                 iosapic_intr_info[vector].low32 = IOSAPIC_MASK;
980                 /* mark as unused */
981                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
982         }
983
984         pcat_compat = system_pcat_compat;
985         if (pcat_compat) {
986                 /*
987                  * Disable the compatibility mode interrupts (8259 style),
988                  * needs IN/OUT support enabled.
989                  */
990                 printk(KERN_INFO
991                        "%s: Disabling PC-AT compatible 8259 interrupts\n",
992                        __FUNCTION__);
993                 outb(0xff, 0xA1);
994                 outb(0xff, 0x21);
995         }
996 }
997
998 static inline int
999 iosapic_alloc (void)
1000 {
1001         int index;
1002
1003         for (index = 0; index < NR_IOSAPICS; index++)
1004                 if (!iosapic_lists[index].addr)
1005                         return index;
1006
1007         printk(KERN_WARNING "%s: failed to allocate iosapic\n", __FUNCTION__);
1008         return -1;
1009 }
1010
1011 static inline void
1012 iosapic_free (int index)
1013 {
1014         memset(&iosapic_lists[index], 0, sizeof(iosapic_lists[0]));
1015 }
1016
1017 static inline int
1018 iosapic_check_gsi_range (unsigned int gsi_base, unsigned int ver)
1019 {
1020         int index;
1021         unsigned int gsi_end, base, end;
1022
1023         /* check gsi range */
1024         gsi_end = gsi_base + ((ver >> 16) & 0xff);
1025         for (index = 0; index < NR_IOSAPICS; index++) {
1026                 if (!iosapic_lists[index].addr)
1027                         continue;
1028
1029                 base = iosapic_lists[index].gsi_base;
1030                 end  = base + iosapic_lists[index].num_rte - 1;
1031
1032                 if (gsi_end < base || end < gsi_base)
1033                         continue; /* OK */
1034
1035                 return -EBUSY;
1036         }
1037         return 0;
1038 }
1039
1040 int __devinit
1041 iosapic_init (unsigned long phys_addr, unsigned int gsi_base)
1042 {
1043         int num_rte, err, index;
1044         unsigned int isa_irq, ver;
1045         char __iomem *addr;
1046         unsigned long flags;
1047
1048         spin_lock_irqsave(&iosapic_lock, flags);
1049         index = find_iosapic(gsi_base);
1050         if (index >= 0) {
1051                 spin_unlock_irqrestore(&iosapic_lock, flags);
1052                 return -EBUSY;
1053         }
1054
1055         addr = ioremap(phys_addr, 0);
1056         ver = iosapic_version(addr);
1057         if ((err = iosapic_check_gsi_range(gsi_base, ver))) {
1058                 iounmap(addr);
1059                 spin_unlock_irqrestore(&iosapic_lock, flags);
1060                 return err;
1061         }
1062
1063         /*
1064          * The MAX_REDIR register holds the highest input pin number
1065          * (starting from 0).  We add 1 so that we can use it for
1066          * number of pins (= RTEs)
1067          */
1068         num_rte = ((ver >> 16) & 0xff) + 1;
1069
1070         index = iosapic_alloc();
1071         iosapic_lists[index].addr = addr;
1072         iosapic_lists[index].gsi_base = gsi_base;
1073         iosapic_lists[index].num_rte = num_rte;
1074 #ifdef CONFIG_NUMA
1075         iosapic_lists[index].node = MAX_NUMNODES;
1076 #endif
1077         spin_lock_init(&iosapic_lists[index].lock);
1078         spin_unlock_irqrestore(&iosapic_lock, flags);
1079
1080         if ((gsi_base == 0) && pcat_compat) {
1081                 /*
1082                  * Map the legacy ISA devices into the IOSAPIC data.  Some of
1083                  * these may get reprogrammed later on with data from the ACPI
1084                  * Interrupt Source Override table.
1085                  */
1086                 for (isa_irq = 0; isa_irq < 16; ++isa_irq)
1087                         iosapic_override_isa_irq(isa_irq, isa_irq,
1088                                                  IOSAPIC_POL_HIGH,
1089                                                  IOSAPIC_EDGE);
1090         }
1091         return 0;
1092 }
1093
1094 #ifdef CONFIG_HOTPLUG
1095 int
1096 iosapic_remove (unsigned int gsi_base)
1097 {
1098         int index, err = 0;
1099         unsigned long flags;
1100
1101         spin_lock_irqsave(&iosapic_lock, flags);
1102         index = find_iosapic(gsi_base);
1103         if (index < 0) {
1104                 printk(KERN_WARNING "%s: No IOSAPIC for GSI base %u\n",
1105                        __FUNCTION__, gsi_base);
1106                 goto out;
1107         }
1108
1109         if (iosapic_lists[index].rtes_inuse) {
1110                 err = -EBUSY;
1111                 printk(KERN_WARNING "%s: IOSAPIC for GSI base %u is busy\n",
1112                        __FUNCTION__, gsi_base);
1113                 goto out;
1114         }
1115
1116         iounmap(iosapic_lists[index].addr);
1117         iosapic_free(index);
1118  out:
1119         spin_unlock_irqrestore(&iosapic_lock, flags);
1120         return err;
1121 }
1122 #endif /* CONFIG_HOTPLUG */
1123
1124 #ifdef CONFIG_NUMA
1125 void __devinit
1126 map_iosapic_to_node(unsigned int gsi_base, int node)
1127 {
1128         int index;
1129
1130         index = find_iosapic(gsi_base);
1131         if (index < 0) {
1132                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1133                        __FUNCTION__, gsi_base);
1134                 return;
1135         }
1136         iosapic_lists[index].node = node;
1137         return;
1138 }
1139 #endif
1140
1141 static int __init iosapic_enable_kmalloc (void)
1142 {
1143         iosapic_kmalloc_ok = 1;
1144         return 0;
1145 }
1146 core_initcall (iosapic_enable_kmalloc);