]> rtime.felk.cvut.cz Git - linux-imx.git/blob - arch/ia64/kernel/iosapic.c
bcf91dceaf5d2d29485cced1dc5e4fbc2f1313b1
[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[NR_IRQS];
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 __gsi_to_irq(unsigned int gsi)
185 {
186         int irq;
187         struct iosapic_intr_info *info;
188         struct iosapic_rte_info *rte;
189
190         for (irq = 0; irq < NR_IRQS; irq++) {
191                 info = &iosapic_intr_info[irq];
192                 list_for_each_entry(rte, &info->rtes, rte_list)
193                         if (rte->iosapic->gsi_base + rte->rte_index == gsi)
194                                 return irq;
195         }
196         return -1;
197 }
198
199 /*
200  * Translate GSI number to the corresponding IA-64 interrupt vector.  If no
201  * entry exists, return -1.
202  */
203 inline int
204 gsi_to_vector (unsigned int gsi)
205 {
206         int irq = __gsi_to_irq(gsi);
207         if (irq < 0)
208                 return -1;
209         return irq_to_vector(irq);
210 }
211
212 int
213 gsi_to_irq (unsigned int gsi)
214 {
215         unsigned long flags;
216         int irq;
217
218         spin_lock_irqsave(&iosapic_lock, flags);
219         irq = __gsi_to_irq(gsi);
220         spin_unlock_irqrestore(&iosapic_lock, flags);
221         return irq;
222 }
223
224 static struct iosapic_rte_info *find_rte(unsigned int irq, unsigned int gsi)
225 {
226         struct iosapic_rte_info *rte;
227
228         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
229                 if (rte->iosapic->gsi_base + rte->rte_index == gsi)
230                         return rte;
231         return NULL;
232 }
233
234 static void
235 set_rte (unsigned int gsi, unsigned int irq, unsigned int dest, int mask)
236 {
237         unsigned long pol, trigger, dmode;
238         u32 low32, high32;
239         int rte_index;
240         char redir;
241         struct iosapic_rte_info *rte;
242         ia64_vector vector = irq_to_vector(irq);
243
244         DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
245
246         rte = find_rte(irq, gsi);
247         if (!rte)
248                 return;         /* not an IOSAPIC interrupt */
249
250         rte_index = rte->rte_index;
251         pol     = iosapic_intr_info[irq].polarity;
252         trigger = iosapic_intr_info[irq].trigger;
253         dmode   = iosapic_intr_info[irq].dmode;
254
255         redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
256
257 #ifdef CONFIG_SMP
258         set_irq_affinity_info(irq, (int)(dest & 0xffff), redir);
259 #endif
260
261         low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
262                  (trigger << IOSAPIC_TRIGGER_SHIFT) |
263                  (dmode << IOSAPIC_DELIVERY_SHIFT) |
264                  ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
265                  vector);
266
267         /* dest contains both id and eid */
268         high32 = (dest << IOSAPIC_DEST_SHIFT);
269
270         iosapic_write(rte->iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
271         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
272         iosapic_intr_info[irq].low32 = low32;
273         iosapic_intr_info[irq].dest = dest;
274 }
275
276 static void
277 nop (unsigned int irq)
278 {
279         /* do nothing... */
280 }
281
282
283 #ifdef CONFIG_KEXEC
284 void
285 kexec_disable_iosapic(void)
286 {
287         struct iosapic_intr_info *info;
288         struct iosapic_rte_info *rte;
289         ia64_vector vec;
290         int irq;
291
292         for (irq = 0; irq < NR_IRQS; irq++) {
293                 info = &iosapic_intr_info[irq];
294                 vec = irq_to_vector(irq);
295                 list_for_each_entry(rte, &info->rtes,
296                                 rte_list) {
297                         iosapic_write(rte->iosapic,
298                                         IOSAPIC_RTE_LOW(rte->rte_index),
299                                         IOSAPIC_MASK|vec);
300                         iosapic_eoi(rte->iosapic->addr, vec);
301                 }
302         }
303 }
304 #endif
305
306 static void
307 mask_irq (unsigned int irq)
308 {
309         u32 low32;
310         int rte_index;
311         struct iosapic_rte_info *rte;
312
313         if (list_empty(&iosapic_intr_info[irq].rtes))
314                 return;                 /* not an IOSAPIC interrupt! */
315
316         /* set only the mask bit */
317         low32 = iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
318         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
319                 rte_index = rte->rte_index;
320                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
321         }
322 }
323
324 static void
325 unmask_irq (unsigned int irq)
326 {
327         u32 low32;
328         int rte_index;
329         struct iosapic_rte_info *rte;
330
331         if (list_empty(&iosapic_intr_info[irq].rtes))
332                 return;                 /* not an IOSAPIC interrupt! */
333
334         low32 = iosapic_intr_info[irq].low32 &= ~IOSAPIC_MASK;
335         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
336                 rte_index = rte->rte_index;
337                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
338         }
339 }
340
341
342 static void
343 iosapic_set_affinity (unsigned int irq, cpumask_t mask)
344 {
345 #ifdef CONFIG_SMP
346         u32 high32, low32;
347         int dest, rte_index;
348         int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
349         struct iosapic_rte_info *rte;
350         struct iosapic *iosapic;
351
352         irq &= (~IA64_IRQ_REDIRECTED);
353
354         if (cpus_empty(mask))
355                 return;
356
357         dest = cpu_physical_id(first_cpu(mask));
358
359         if (list_empty(&iosapic_intr_info[irq].rtes))
360                 return;                 /* not an IOSAPIC interrupt */
361
362         set_irq_affinity_info(irq, dest, redir);
363
364         /* dest contains both id and eid */
365         high32 = dest << IOSAPIC_DEST_SHIFT;
366
367         low32 = iosapic_intr_info[irq].low32 & ~(7 << IOSAPIC_DELIVERY_SHIFT);
368         if (redir)
369                 /* change delivery mode to lowest priority */
370                 low32 |= (IOSAPIC_LOWEST_PRIORITY << IOSAPIC_DELIVERY_SHIFT);
371         else
372                 /* change delivery mode to fixed */
373                 low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
374
375         iosapic_intr_info[irq].low32 = low32;
376         iosapic_intr_info[irq].dest = dest;
377         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
378                 iosapic = rte->iosapic;
379                 rte_index = rte->rte_index;
380                 iosapic_write(iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
381                 iosapic_write(iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
382         }
383 #endif
384 }
385
386 /*
387  * Handlers for level-triggered interrupts.
388  */
389
390 static unsigned int
391 iosapic_startup_level_irq (unsigned int irq)
392 {
393         unmask_irq(irq);
394         return 0;
395 }
396
397 static void
398 iosapic_end_level_irq (unsigned int irq)
399 {
400         ia64_vector vec = irq_to_vector(irq);
401         struct iosapic_rte_info *rte;
402
403         move_native_irq(irq);
404         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
405                 iosapic_eoi(rte->iosapic->addr, vec);
406 }
407
408 #define iosapic_shutdown_level_irq      mask_irq
409 #define iosapic_enable_level_irq        unmask_irq
410 #define iosapic_disable_level_irq       mask_irq
411 #define iosapic_ack_level_irq           nop
412
413 struct irq_chip irq_type_iosapic_level = {
414         .name =         "IO-SAPIC-level",
415         .startup =      iosapic_startup_level_irq,
416         .shutdown =     iosapic_shutdown_level_irq,
417         .enable =       iosapic_enable_level_irq,
418         .disable =      iosapic_disable_level_irq,
419         .ack =          iosapic_ack_level_irq,
420         .end =          iosapic_end_level_irq,
421         .mask =         mask_irq,
422         .unmask =       unmask_irq,
423         .set_affinity = iosapic_set_affinity
424 };
425
426 /*
427  * Handlers for edge-triggered interrupts.
428  */
429
430 static unsigned int
431 iosapic_startup_edge_irq (unsigned int irq)
432 {
433         unmask_irq(irq);
434         /*
435          * IOSAPIC simply drops interrupts pended while the
436          * corresponding pin was masked, so we can't know if an
437          * interrupt is pending already.  Let's hope not...
438          */
439         return 0;
440 }
441
442 static void
443 iosapic_ack_edge_irq (unsigned int irq)
444 {
445         irq_desc_t *idesc = irq_desc + irq;
446
447         move_native_irq(irq);
448         /*
449          * Once we have recorded IRQ_PENDING already, we can mask the
450          * interrupt for real. This prevents IRQ storms from unhandled
451          * devices.
452          */
453         if ((idesc->status & (IRQ_PENDING|IRQ_DISABLED)) ==
454             (IRQ_PENDING|IRQ_DISABLED))
455                 mask_irq(irq);
456 }
457
458 #define iosapic_enable_edge_irq         unmask_irq
459 #define iosapic_disable_edge_irq        nop
460 #define iosapic_end_edge_irq            nop
461
462 struct irq_chip irq_type_iosapic_edge = {
463         .name =         "IO-SAPIC-edge",
464         .startup =      iosapic_startup_edge_irq,
465         .shutdown =     iosapic_disable_edge_irq,
466         .enable =       iosapic_enable_edge_irq,
467         .disable =      iosapic_disable_edge_irq,
468         .ack =          iosapic_ack_edge_irq,
469         .end =          iosapic_end_edge_irq,
470         .mask =         mask_irq,
471         .unmask =       unmask_irq,
472         .set_affinity = iosapic_set_affinity
473 };
474
475 unsigned int
476 iosapic_version (char __iomem *addr)
477 {
478         /*
479          * IOSAPIC Version Register return 32 bit structure like:
480          * {
481          *      unsigned int version   : 8;
482          *      unsigned int reserved1 : 8;
483          *      unsigned int max_redir : 8;
484          *      unsigned int reserved2 : 8;
485          * }
486          */
487         return __iosapic_read(addr, IOSAPIC_VERSION);
488 }
489
490 static int iosapic_find_sharable_irq(unsigned long trigger, unsigned long pol)
491 {
492         int i, irq = -ENOSPC, min_count = -1;
493         struct iosapic_intr_info *info;
494
495         /*
496          * shared vectors for edge-triggered interrupts are not
497          * supported yet
498          */
499         if (trigger == IOSAPIC_EDGE)
500                 return -EINVAL;
501
502         for (i = 0; i <= NR_IRQS; i++) {
503                 info = &iosapic_intr_info[i];
504                 if (info->trigger == trigger && info->polarity == pol &&
505                     (info->dmode == IOSAPIC_FIXED || info->dmode ==
506                      IOSAPIC_LOWEST_PRIORITY)) {
507                         if (min_count == -1 || info->count < min_count) {
508                                 irq = i;
509                                 min_count = info->count;
510                         }
511                 }
512         }
513         return irq;
514 }
515
516 /*
517  * if the given vector is already owned by other,
518  *  assign a new vector for the other and make the vector available
519  */
520 static void __init
521 iosapic_reassign_vector (int irq)
522 {
523         int new_irq;
524
525         if (!list_empty(&iosapic_intr_info[irq].rtes)) {
526                 new_irq = create_irq();
527                 if (new_irq < 0)
528                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
529                 printk(KERN_INFO "Reassigning vector %d to %d\n",
530                        irq_to_vector(irq), irq_to_vector(new_irq));
531                 memcpy(&iosapic_intr_info[new_irq], &iosapic_intr_info[irq],
532                        sizeof(struct iosapic_intr_info));
533                 INIT_LIST_HEAD(&iosapic_intr_info[new_irq].rtes);
534                 list_move(iosapic_intr_info[irq].rtes.next,
535                           &iosapic_intr_info[new_irq].rtes);
536                 memset(&iosapic_intr_info[irq], 0,
537                        sizeof(struct iosapic_intr_info));
538                 iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
539                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
540         }
541 }
542
543 static struct iosapic_rte_info *iosapic_alloc_rte (void)
544 {
545         int i;
546         struct iosapic_rte_info *rte;
547         int preallocated = 0;
548
549         if (!iosapic_kmalloc_ok && list_empty(&free_rte_list)) {
550                 rte = alloc_bootmem(sizeof(struct iosapic_rte_info) *
551                                     NR_PREALLOCATE_RTE_ENTRIES);
552                 if (!rte)
553                         return NULL;
554                 for (i = 0; i < NR_PREALLOCATE_RTE_ENTRIES; i++, rte++)
555                         list_add(&rte->rte_list, &free_rte_list);
556         }
557
558         if (!list_empty(&free_rte_list)) {
559                 rte = list_entry(free_rte_list.next, struct iosapic_rte_info,
560                                  rte_list);
561                 list_del(&rte->rte_list);
562                 preallocated++;
563         } else {
564                 rte = kmalloc(sizeof(struct iosapic_rte_info), GFP_ATOMIC);
565                 if (!rte)
566                         return NULL;
567         }
568
569         memset(rte, 0, sizeof(struct iosapic_rte_info));
570         if (preallocated)
571                 rte->flags |= RTE_PREALLOCATED;
572
573         return rte;
574 }
575
576 static void iosapic_free_rte (struct iosapic_rte_info *rte)
577 {
578         if (rte->flags & RTE_PREALLOCATED)
579                 list_add_tail(&rte->rte_list, &free_rte_list);
580         else
581                 kfree(rte);
582 }
583
584 static inline int irq_is_shared (int irq)
585 {
586         return (iosapic_intr_info[irq].count > 1);
587 }
588
589 static int
590 register_intr (unsigned int gsi, int irq, unsigned char delivery,
591                unsigned long polarity, unsigned long trigger)
592 {
593         irq_desc_t *idesc;
594         struct hw_interrupt_type *irq_type;
595         int index;
596         struct iosapic_rte_info *rte;
597
598         index = find_iosapic(gsi);
599         if (index < 0) {
600                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
601                        __FUNCTION__, gsi);
602                 return -ENODEV;
603         }
604
605         rte = find_rte(irq, gsi);
606         if (!rte) {
607                 rte = iosapic_alloc_rte();
608                 if (!rte) {
609                         printk(KERN_WARNING "%s: cannot allocate memory\n",
610                                __FUNCTION__);
611                         return -ENOMEM;
612                 }
613
614                 rte->iosapic    = &iosapic_lists[index];
615                 rte->rte_index  = gsi - rte->iosapic->gsi_base;
616                 rte->refcnt++;
617                 list_add_tail(&rte->rte_list, &iosapic_intr_info[irq].rtes);
618                 iosapic_intr_info[irq].count++;
619                 iosapic_lists[index].rtes_inuse++;
620         }
621         else if (irq_is_shared(irq)) {
622                 struct iosapic_intr_info *info = &iosapic_intr_info[irq];
623                 if (info->trigger != trigger || info->polarity != polarity) {
624                         printk (KERN_WARNING
625                                 "%s: cannot override the interrupt\n",
626                                 __FUNCTION__);
627                         return -EINVAL;
628                 }
629         }
630
631         iosapic_intr_info[irq].polarity = polarity;
632         iosapic_intr_info[irq].dmode    = delivery;
633         iosapic_intr_info[irq].trigger  = trigger;
634
635         if (trigger == IOSAPIC_EDGE)
636                 irq_type = &irq_type_iosapic_edge;
637         else
638                 irq_type = &irq_type_iosapic_level;
639
640         idesc = irq_desc + irq;
641         if (idesc->chip != irq_type) {
642                 if (idesc->chip != &no_irq_type)
643                         printk(KERN_WARNING
644                                "%s: changing vector %d from %s to %s\n",
645                                __FUNCTION__, irq_to_vector(irq),
646                                idesc->chip->name, irq_type->name);
647                 idesc->chip = irq_type;
648         }
649         return 0;
650 }
651
652 static unsigned int
653 get_target_cpu (unsigned int gsi, int irq)
654 {
655 #ifdef CONFIG_SMP
656         static int cpu = -1;
657         extern int cpe_vector;
658
659         /*
660          * In case of vector shared by multiple RTEs, all RTEs that
661          * share the vector need to use the same destination CPU.
662          */
663         if (!list_empty(&iosapic_intr_info[irq].rtes))
664                 return iosapic_intr_info[irq].dest;
665
666         /*
667          * If the platform supports redirection via XTP, let it
668          * distribute interrupts.
669          */
670         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
671                 return cpu_physical_id(smp_processor_id());
672
673         /*
674          * Some interrupts (ACPI SCI, for instance) are registered
675          * before the BSP is marked as online.
676          */
677         if (!cpu_online(smp_processor_id()))
678                 return cpu_physical_id(smp_processor_id());
679
680 #ifdef CONFIG_ACPI
681         if (cpe_vector > 0 && irq_to_vector(irq) == IA64_CPEP_VECTOR)
682                 return get_cpei_target_cpu();
683 #endif
684
685 #ifdef CONFIG_NUMA
686         {
687                 int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
688                 cpumask_t cpu_mask;
689
690                 iosapic_index = find_iosapic(gsi);
691                 if (iosapic_index < 0 ||
692                     iosapic_lists[iosapic_index].node == MAX_NUMNODES)
693                         goto skip_numa_setup;
694
695                 cpu_mask = node_to_cpumask(iosapic_lists[iosapic_index].node);
696
697                 for_each_cpu_mask(numa_cpu, cpu_mask) {
698                         if (!cpu_online(numa_cpu))
699                                 cpu_clear(numa_cpu, cpu_mask);
700                 }
701
702                 num_cpus = cpus_weight(cpu_mask);
703
704                 if (!num_cpus)
705                         goto skip_numa_setup;
706
707                 /* Use irq assignment to distribute across cpus in node */
708                 cpu_index = irq % num_cpus;
709
710                 for (numa_cpu = first_cpu(cpu_mask) ; i < cpu_index ; i++)
711                         numa_cpu = next_cpu(numa_cpu, cpu_mask);
712
713                 if (numa_cpu != NR_CPUS)
714                         return cpu_physical_id(numa_cpu);
715         }
716 skip_numa_setup:
717 #endif
718         /*
719          * Otherwise, round-robin interrupt vectors across all the
720          * processors.  (It'd be nice if we could be smarter in the
721          * case of NUMA.)
722          */
723         do {
724                 if (++cpu >= NR_CPUS)
725                         cpu = 0;
726         } while (!cpu_online(cpu));
727
728         return cpu_physical_id(cpu);
729 #else  /* CONFIG_SMP */
730         return cpu_physical_id(smp_processor_id());
731 #endif
732 }
733
734 /*
735  * ACPI can describe IOSAPIC interrupts via static tables and namespace
736  * methods.  This provides an interface to register those interrupts and
737  * program the IOSAPIC RTE.
738  */
739 int
740 iosapic_register_intr (unsigned int gsi,
741                        unsigned long polarity, unsigned long trigger)
742 {
743         int irq, mask = 1, err;
744         unsigned int dest;
745         unsigned long flags;
746         struct iosapic_rte_info *rte;
747         u32 low32;
748
749         /*
750          * If this GSI has already been registered (i.e., it's a
751          * shared interrupt, or we lost a race to register it),
752          * don't touch the RTE.
753          */
754         spin_lock_irqsave(&iosapic_lock, flags);
755         irq = __gsi_to_irq(gsi);
756         if (irq > 0) {
757                 rte = find_rte(irq, gsi);
758                 rte->refcnt++;
759                 goto unlock_iosapic_lock;
760         }
761
762         /* If vector is running out, we try to find a sharable vector */
763         irq = create_irq();
764         if (irq < 0) {
765                 irq = iosapic_find_sharable_irq(trigger, polarity);
766                 if (irq < 0)
767                         goto unlock_iosapic_lock;
768         }
769
770         spin_lock(&irq_desc[irq].lock);
771         dest = get_target_cpu(gsi, irq);
772         err = register_intr(gsi, irq, IOSAPIC_LOWEST_PRIORITY,
773                             polarity, trigger);
774         if (err < 0) {
775                 irq = err;
776                 goto unlock_all;
777         }
778
779         /*
780          * If the vector is shared and already unmasked for other
781          * interrupt sources, don't mask it.
782          */
783         low32 = iosapic_intr_info[irq].low32;
784         if (irq_is_shared(irq) && !(low32 & IOSAPIC_MASK))
785                 mask = 0;
786         set_rte(gsi, irq, dest, mask);
787
788         printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
789                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
790                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
791                cpu_logical_id(dest), dest, irq_to_vector(irq));
792  unlock_all:
793         spin_unlock(&irq_desc[irq].lock);
794  unlock_iosapic_lock:
795         spin_unlock_irqrestore(&iosapic_lock, flags);
796         return irq;
797 }
798
799 void
800 iosapic_unregister_intr (unsigned int gsi)
801 {
802         unsigned long flags;
803         int irq, index;
804         irq_desc_t *idesc;
805         u32 low32;
806         unsigned long trigger, polarity;
807         unsigned int dest;
808         struct iosapic_rte_info *rte;
809
810         /*
811          * If the irq associated with the gsi is not found,
812          * iosapic_unregister_intr() is unbalanced. We need to check
813          * this again after getting locks.
814          */
815         irq = gsi_to_irq(gsi);
816         if (irq < 0) {
817                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
818                        gsi);
819                 WARN_ON(1);
820                 return;
821         }
822
823         spin_lock_irqsave(&iosapic_lock, flags);
824         if ((rte = find_rte(irq, gsi)) == NULL) {
825                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
826                        gsi);
827                 WARN_ON(1);
828                 goto out;
829         }
830
831         if (--rte->refcnt > 0)
832                 goto out;
833
834         /* Remove the rte entry from the list */
835         idesc = irq_desc + irq;
836         spin_lock(&idesc->lock);
837         list_del(&rte->rte_list);
838         spin_unlock(&idesc->lock);
839
840         /* Mask the interrupt */
841         low32 = iosapic_intr_info[irq].low32 | IOSAPIC_MASK;
842         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte->rte_index), low32);
843
844         iosapic_intr_info[irq].count--;
845         iosapic_free_rte(rte);
846         index = find_iosapic(gsi);
847         iosapic_lists[index].rtes_inuse--;
848         WARN_ON(iosapic_lists[index].rtes_inuse < 0);
849
850         trigger  = iosapic_intr_info[irq].trigger;
851         polarity = iosapic_intr_info[irq].polarity;
852         dest     = iosapic_intr_info[irq].dest;
853         printk(KERN_INFO
854                "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d unregistered\n",
855                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
856                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
857                cpu_logical_id(dest), dest, irq_to_vector(irq));
858
859         if (list_empty(&iosapic_intr_info[irq].rtes)) {
860                 /* Sanity check */
861                 BUG_ON(iosapic_intr_info[irq].count);
862 #ifdef CONFIG_SMP
863                 /* Clear affinity */
864                 cpus_setall(idesc->affinity);
865 #endif
866                 /* Clear the interrupt information */
867                 memset(&iosapic_intr_info[irq], 0,
868                        sizeof(struct iosapic_intr_info));
869                 iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
870                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
871
872                 /* Destroy IRQ */
873                 destroy_irq(irq);
874         }
875  out:
876         spin_unlock_irqrestore(&iosapic_lock, flags);
877 }
878
879 /*
880  * ACPI calls this when it finds an entry for a platform interrupt.
881  */
882 int __init
883 iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
884                                 int iosapic_vector, u16 eid, u16 id,
885                                 unsigned long polarity, unsigned long trigger)
886 {
887         static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
888         unsigned char delivery;
889         int irq, vector, mask = 0;
890         unsigned int dest = ((id << 8) | eid) & 0xffff;
891
892         switch (int_type) {
893               case ACPI_INTERRUPT_PMI:
894                 vector = iosapic_vector;
895                 irq = vector;           /* FIXME */
896                 /*
897                  * since PMI vector is alloc'd by FW(ACPI) not by kernel,
898                  * we need to make sure the vector is available
899                  */
900                 iosapic_reassign_vector(irq);
901                 delivery = IOSAPIC_PMI;
902                 break;
903               case ACPI_INTERRUPT_INIT:
904                 irq = create_irq();
905                 if (irq < 0)
906                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
907                 vector = irq_to_vector(irq);
908                 delivery = IOSAPIC_INIT;
909                 break;
910               case ACPI_INTERRUPT_CPEI:
911                 vector = IA64_CPE_VECTOR;
912                 irq = vector;           /* FIXME */
913                 delivery = IOSAPIC_LOWEST_PRIORITY;
914                 mask = 1;
915                 break;
916               default:
917                 printk(KERN_ERR "%s: invalid int type 0x%x\n", __FUNCTION__,
918                        int_type);
919                 return -1;
920         }
921
922         register_intr(gsi, irq, delivery, polarity, trigger);
923
924         printk(KERN_INFO
925                "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x)"
926                " vector %d\n",
927                int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
928                int_type, gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
929                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
930                cpu_logical_id(dest), dest, vector);
931
932         set_rte(gsi, irq, dest, mask);
933         return vector;
934 }
935
936 /*
937  * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
938  */
939 void __devinit
940 iosapic_override_isa_irq (unsigned int isa_irq, unsigned int gsi,
941                           unsigned long polarity,
942                           unsigned long trigger)
943 {
944         int vector, irq;
945         unsigned int dest = cpu_physical_id(smp_processor_id());
946
947         vector = isa_irq_to_vector(isa_irq);
948         irq = vector;   /* FIXME */
949         register_intr(gsi, irq, IOSAPIC_LOWEST_PRIORITY, polarity, trigger);
950
951         DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
952             isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
953             polarity == IOSAPIC_POL_HIGH ? "high" : "low",
954             cpu_logical_id(dest), dest, vector);
955
956         set_rte(gsi, irq, dest, 1);
957 }
958
959 void __init
960 iosapic_system_init (int system_pcat_compat)
961 {
962         int irq;
963
964         for (irq = 0; irq < NR_IRQS; ++irq) {
965                 iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
966                 /* mark as unused */
967                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
968         }
969
970         pcat_compat = system_pcat_compat;
971         if (pcat_compat) {
972                 /*
973                  * Disable the compatibility mode interrupts (8259 style),
974                  * needs IN/OUT support enabled.
975                  */
976                 printk(KERN_INFO
977                        "%s: Disabling PC-AT compatible 8259 interrupts\n",
978                        __FUNCTION__);
979                 outb(0xff, 0xA1);
980                 outb(0xff, 0x21);
981         }
982 }
983
984 static inline int
985 iosapic_alloc (void)
986 {
987         int index;
988
989         for (index = 0; index < NR_IOSAPICS; index++)
990                 if (!iosapic_lists[index].addr)
991                         return index;
992
993         printk(KERN_WARNING "%s: failed to allocate iosapic\n", __FUNCTION__);
994         return -1;
995 }
996
997 static inline void
998 iosapic_free (int index)
999 {
1000         memset(&iosapic_lists[index], 0, sizeof(iosapic_lists[0]));
1001 }
1002
1003 static inline int
1004 iosapic_check_gsi_range (unsigned int gsi_base, unsigned int ver)
1005 {
1006         int index;
1007         unsigned int gsi_end, base, end;
1008
1009         /* check gsi range */
1010         gsi_end = gsi_base + ((ver >> 16) & 0xff);
1011         for (index = 0; index < NR_IOSAPICS; index++) {
1012                 if (!iosapic_lists[index].addr)
1013                         continue;
1014
1015                 base = iosapic_lists[index].gsi_base;
1016                 end  = base + iosapic_lists[index].num_rte - 1;
1017
1018                 if (gsi_end < base || end < gsi_base)
1019                         continue; /* OK */
1020
1021                 return -EBUSY;
1022         }
1023         return 0;
1024 }
1025
1026 int __devinit
1027 iosapic_init (unsigned long phys_addr, unsigned int gsi_base)
1028 {
1029         int num_rte, err, index;
1030         unsigned int isa_irq, ver;
1031         char __iomem *addr;
1032         unsigned long flags;
1033
1034         spin_lock_irqsave(&iosapic_lock, flags);
1035         index = find_iosapic(gsi_base);
1036         if (index >= 0) {
1037                 spin_unlock_irqrestore(&iosapic_lock, flags);
1038                 return -EBUSY;
1039         }
1040
1041         addr = ioremap(phys_addr, 0);
1042         ver = iosapic_version(addr);
1043         if ((err = iosapic_check_gsi_range(gsi_base, ver))) {
1044                 iounmap(addr);
1045                 spin_unlock_irqrestore(&iosapic_lock, flags);
1046                 return err;
1047         }
1048
1049         /*
1050          * The MAX_REDIR register holds the highest input pin number
1051          * (starting from 0).  We add 1 so that we can use it for
1052          * number of pins (= RTEs)
1053          */
1054         num_rte = ((ver >> 16) & 0xff) + 1;
1055
1056         index = iosapic_alloc();
1057         iosapic_lists[index].addr = addr;
1058         iosapic_lists[index].gsi_base = gsi_base;
1059         iosapic_lists[index].num_rte = num_rte;
1060 #ifdef CONFIG_NUMA
1061         iosapic_lists[index].node = MAX_NUMNODES;
1062 #endif
1063         spin_lock_init(&iosapic_lists[index].lock);
1064         spin_unlock_irqrestore(&iosapic_lock, flags);
1065
1066         if ((gsi_base == 0) && pcat_compat) {
1067                 /*
1068                  * Map the legacy ISA devices into the IOSAPIC data.  Some of
1069                  * these may get reprogrammed later on with data from the ACPI
1070                  * Interrupt Source Override table.
1071                  */
1072                 for (isa_irq = 0; isa_irq < 16; ++isa_irq)
1073                         iosapic_override_isa_irq(isa_irq, isa_irq,
1074                                                  IOSAPIC_POL_HIGH,
1075                                                  IOSAPIC_EDGE);
1076         }
1077         return 0;
1078 }
1079
1080 #ifdef CONFIG_HOTPLUG
1081 int
1082 iosapic_remove (unsigned int gsi_base)
1083 {
1084         int index, err = 0;
1085         unsigned long flags;
1086
1087         spin_lock_irqsave(&iosapic_lock, flags);
1088         index = find_iosapic(gsi_base);
1089         if (index < 0) {
1090                 printk(KERN_WARNING "%s: No IOSAPIC for GSI base %u\n",
1091                        __FUNCTION__, gsi_base);
1092                 goto out;
1093         }
1094
1095         if (iosapic_lists[index].rtes_inuse) {
1096                 err = -EBUSY;
1097                 printk(KERN_WARNING "%s: IOSAPIC for GSI base %u is busy\n",
1098                        __FUNCTION__, gsi_base);
1099                 goto out;
1100         }
1101
1102         iounmap(iosapic_lists[index].addr);
1103         iosapic_free(index);
1104  out:
1105         spin_unlock_irqrestore(&iosapic_lock, flags);
1106         return err;
1107 }
1108 #endif /* CONFIG_HOTPLUG */
1109
1110 #ifdef CONFIG_NUMA
1111 void __devinit
1112 map_iosapic_to_node(unsigned int gsi_base, int node)
1113 {
1114         int index;
1115
1116         index = find_iosapic(gsi_base);
1117         if (index < 0) {
1118                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1119                        __FUNCTION__, gsi_base);
1120                 return;
1121         }
1122         iosapic_lists[index].node = node;
1123         return;
1124 }
1125 #endif
1126
1127 static int __init iosapic_enable_kmalloc (void)
1128 {
1129         iosapic_kmalloc_ok = 1;
1130         return 0;
1131 }
1132 core_initcall (iosapic_enable_kmalloc);