37ac742da8ede4de877b89018e92f144e284d2af
[linux-2.6.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 APIC code.
13  *                              In particular, we now have separate handlers for edge
14  *                              and level triggered interrupts.
15  * 00/10/27     Asit Mallick, Goutham Rao <goutham.rao@intel.com> IRQ vector allocation
16  *                              PCI to vector mapping, shared PCI interrupts.
17  * 00/10/27     D. Mosberger    Document things a bit more to make them more understandable.
18  *                              Clean up much of the old IOSAPIC cruft.
19  * 01/07/27     J.I. Lee        PCI irq routing, Platform/Legacy interrupts and fixes for
20  *                              ACPI S5(SoftOff) support.
21  * 02/01/23     J.I. Lee        iosapic pgm fixes for PCI irq routing from _PRT
22  * 02/01/07     E. Focht        <efocht@ess.nec.de> Redirectable interrupt vectors in
23  *                              iosapic_set_affinity(), initializations for
24  *                              /proc/irq/#/smp_affinity
25  * 02/04/02     P. Diefenbaugh  Cleaned up ACPI PCI IRQ routing.
26  * 02/04/18     J.I. Lee        bug fix in iosapic_init_pci_irq
27  * 02/04/30     J.I. Lee        bug fix in find_iosapic to fix ACPI PCI IRQ to IOSAPIC mapping
28  *                              error
29  * 02/07/29     T. Kochi        Allocate interrupt vectors dynamically
30  * 02/08/04     T. Kochi        Cleaned up terminology (irq, global system interrupt, vector, etc.)
31  * 02/09/20     D. Mosberger    Simplified by taking advantage of ACPI's pci_irq code.
32  * 03/02/19     B. Helgaas      Make pcat_compat system-wide, not per-IOSAPIC.
33  *                              Remove iosapic_address & gsi_base from external interfaces.
34  *                              Rationalize __init/__devinit attributes.
35  * 04/12/04 Ashok Raj   <ashok.raj@intel.com> Intel Corporation 2004
36  *                              Updated to work with irq migration necessary for CPU Hotplug
37  */
38 /*
39  * Here is what the interrupt logic between a PCI device and the kernel looks like:
40  *
41  * (1) A PCI device raises one of the four interrupt pins (INTA, INTB, INTC, INTD).  The
42  *     device is uniquely identified by its bus--, and slot-number (the function
43  *     number does not matter here because all functions share the same interrupt
44  *     lines).
45  *
46  * (2) The motherboard routes the interrupt line to a pin on a IOSAPIC controller.
47  *     Multiple interrupt lines may have to share the same IOSAPIC pin (if they're level
48  *     triggered and use the same polarity).  Each interrupt line has a unique Global
49  *     System Interrupt (GSI) number which can be calculated as the sum of the controller's
50  *     base GSI number and the IOSAPIC pin number to which the line connects.
51  *
52  * (3) The IOSAPIC uses an internal routing table entries (RTEs) to map the IOSAPIC pin
53  *     into the IA-64 interrupt vector.  This interrupt vector is then sent to the CPU.
54  *
55  * (4) The kernel recognizes an interrupt as an IRQ.  The IRQ interface is used as
56  *     architecture-independent interrupt handling mechanism in Linux.  As an
57  *     IRQ is a number, we have to have IA-64 interrupt vector number <-> IRQ number
58  *     mapping.  On smaller systems, we use one-to-one mapping between IA-64 vector and
59  *     IRQ.  A platform can implement platform_irq_to_vector(irq) and
60  *     platform_local_vector_to_irq(vector) APIs to differentiate the mapping.
61  *     Please see also include/asm-ia64/hw_irq.h for those APIs.
62  *
63  * To sum up, there are three levels of mappings involved:
64  *
65  *      PCI pin -> global system interrupt (GSI) -> IA-64 vector <-> IRQ
66  *
67  * Note: The term "IRQ" is loosely used everywhere in Linux kernel to describe interrupts.
68  * Now we use "IRQ" only for Linux IRQ's.  ISA IRQ (isa_irq) is the only exception in this
69  * source code.
70  */
71 #include <linux/config.h>
72
73 #include <linux/acpi.h>
74 #include <linux/init.h>
75 #include <linux/irq.h>
76 #include <linux/kernel.h>
77 #include <linux/list.h>
78 #include <linux/pci.h>
79 #include <linux/smp.h>
80 #include <linux/smp_lock.h>
81 #include <linux/string.h>
82 #include <linux/bootmem.h>
83
84 #include <asm/delay.h>
85 #include <asm/hw_irq.h>
86 #include <asm/io.h>
87 #include <asm/iosapic.h>
88 #include <asm/machvec.h>
89 #include <asm/processor.h>
90 #include <asm/ptrace.h>
91 #include <asm/system.h>
92
93
94 #undef DEBUG_INTERRUPT_ROUTING
95
96 #ifdef DEBUG_INTERRUPT_ROUTING
97 #define DBG(fmt...)     printk(fmt)
98 #else
99 #define DBG(fmt...)
100 #endif
101
102 #define NR_PREALLOCATE_RTE_ENTRIES      (PAGE_SIZE / sizeof(struct iosapic_rte_info))
103 #define RTE_PREALLOCATED        (1)
104
105 static DEFINE_SPINLOCK(iosapic_lock);
106
107 /* These tables map IA-64 vectors to the IOSAPIC pin that generates this vector. */
108
109 struct iosapic_rte_info {
110         struct list_head rte_list;      /* node in list of RTEs sharing the same vector */
111         char __iomem    *addr;          /* base address of IOSAPIC */
112         unsigned int    gsi_base;       /* first GSI assigned to this IOSAPIC */
113         char            rte_index;      /* IOSAPIC RTE index */
114         int             refcnt;         /* reference counter */
115         unsigned int    flags;          /* flags */
116 } ____cacheline_aligned;
117
118 static struct iosapic_intr_info {
119         struct list_head rtes;          /* RTEs using this vector (empty => not an IOSAPIC interrupt) */
120         int             count;          /* # of RTEs that shares this vector */
121         u32             low32;          /* current value of low word of Redirection table entry */
122         unsigned int    dest;           /* destination CPU physical ID */
123         unsigned char   dmode   : 3;    /* delivery mode (see iosapic.h) */
124         unsigned char   polarity: 1;    /* interrupt polarity (see iosapic.h) */
125         unsigned char   trigger : 1;    /* trigger mode (see iosapic.h) */
126 } iosapic_intr_info[IA64_NUM_VECTORS];
127
128 static struct iosapic {
129         char __iomem    *addr;          /* base address of IOSAPIC */
130         unsigned int    gsi_base;       /* first GSI assigned to this IOSAPIC */
131         unsigned short  num_rte;        /* number of RTE in this IOSAPIC */
132         int             rtes_inuse;     /* # of RTEs in use on this IOSAPIC */
133 #ifdef CONFIG_NUMA
134         unsigned short  node;           /* numa node association via pxm */
135 #endif
136 } iosapic_lists[NR_IOSAPICS];
137
138 static unsigned char pcat_compat __devinitdata; /* 8259 compatibility flag */
139
140 static int iosapic_kmalloc_ok;
141 static LIST_HEAD(free_rte_list);
142
143 /*
144  * Find an IOSAPIC associated with a GSI
145  */
146 static inline int
147 find_iosapic (unsigned int gsi)
148 {
149         int i;
150
151         for (i = 0; i < NR_IOSAPICS; i++) {
152                 if ((unsigned) (gsi - iosapic_lists[i].gsi_base) < iosapic_lists[i].num_rte)
153                         return i;
154         }
155
156         return -1;
157 }
158
159 static inline int
160 _gsi_to_vector (unsigned int gsi)
161 {
162         struct iosapic_intr_info *info;
163         struct iosapic_rte_info *rte;
164
165         for (info = iosapic_intr_info; info < iosapic_intr_info + IA64_NUM_VECTORS; ++info)
166                 list_for_each_entry(rte, &info->rtes, rte_list)
167                         if (rte->gsi_base + rte->rte_index == gsi)
168                                 return info - iosapic_intr_info;
169         return -1;
170 }
171
172 /*
173  * Translate GSI number to the corresponding IA-64 interrupt vector.  If no
174  * entry exists, return -1.
175  */
176 inline int
177 gsi_to_vector (unsigned int gsi)
178 {
179         return _gsi_to_vector(gsi);
180 }
181
182 int
183 gsi_to_irq (unsigned int gsi)
184 {
185         unsigned long flags;
186         int irq;
187         /*
188          * XXX fix me: this assumes an identity mapping vetween IA-64 vector and Linux irq
189          * numbers...
190          */
191         spin_lock_irqsave(&iosapic_lock, flags);
192         {
193                 irq = _gsi_to_vector(gsi);
194         }
195         spin_unlock_irqrestore(&iosapic_lock, flags);
196
197         return irq;
198 }
199
200 static struct iosapic_rte_info *gsi_vector_to_rte(unsigned int gsi, unsigned int vec)
201 {
202         struct iosapic_rte_info *rte;
203
204         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list)
205                 if (rte->gsi_base + rte->rte_index == gsi)
206                         return rte;
207         return NULL;
208 }
209
210 static void
211 set_rte (unsigned int gsi, unsigned int vector, unsigned int dest, int mask)
212 {
213         unsigned long pol, trigger, dmode;
214         u32 low32, high32;
215         char __iomem *addr;
216         int rte_index;
217         char redir;
218         struct iosapic_rte_info *rte;
219
220         DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
221
222         rte = gsi_vector_to_rte(gsi, vector);
223         if (!rte)
224                 return;         /* not an IOSAPIC interrupt */
225
226         rte_index = rte->rte_index;
227         addr    = rte->addr;
228         pol     = iosapic_intr_info[vector].polarity;
229         trigger = iosapic_intr_info[vector].trigger;
230         dmode   = iosapic_intr_info[vector].dmode;
231
232         redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
233
234 #ifdef CONFIG_SMP
235         {
236                 unsigned int irq;
237
238                 for (irq = 0; irq < NR_IRQS; ++irq)
239                         if (irq_to_vector(irq) == vector) {
240                                 set_irq_affinity_info(irq, (int)(dest & 0xffff), redir);
241                                 break;
242                         }
243         }
244 #endif
245
246         low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
247                  (trigger << IOSAPIC_TRIGGER_SHIFT) |
248                  (dmode << IOSAPIC_DELIVERY_SHIFT) |
249                  ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
250                  vector);
251
252         /* dest contains both id and eid */
253         high32 = (dest << IOSAPIC_DEST_SHIFT);
254
255         iosapic_write(addr, IOSAPIC_RTE_HIGH(rte_index), high32);
256         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
257         iosapic_intr_info[vector].low32 = low32;
258         iosapic_intr_info[vector].dest = dest;
259 }
260
261 static void
262 nop (unsigned int vector)
263 {
264         /* do nothing... */
265 }
266
267 static void
268 mask_irq (unsigned int irq)
269 {
270         unsigned long flags;
271         char __iomem *addr;
272         u32 low32;
273         int rte_index;
274         ia64_vector vec = irq_to_vector(irq);
275         struct iosapic_rte_info *rte;
276
277         if (list_empty(&iosapic_intr_info[vec].rtes))
278                 return;                 /* not an IOSAPIC interrupt! */
279
280         spin_lock_irqsave(&iosapic_lock, flags);
281         {
282                 /* set only the mask bit */
283                 low32 = iosapic_intr_info[vec].low32 |= IOSAPIC_MASK;
284                 list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list) {
285                         addr = rte->addr;
286                         rte_index = rte->rte_index;
287                         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
288                 }
289         }
290         spin_unlock_irqrestore(&iosapic_lock, flags);
291 }
292
293 static void
294 unmask_irq (unsigned int irq)
295 {
296         unsigned long flags;
297         char __iomem *addr;
298         u32 low32;
299         int rte_index;
300         ia64_vector vec = irq_to_vector(irq);
301         struct iosapic_rte_info *rte;
302
303         if (list_empty(&iosapic_intr_info[vec].rtes))
304                 return;                 /* not an IOSAPIC interrupt! */
305
306         spin_lock_irqsave(&iosapic_lock, flags);
307         {
308                 low32 = iosapic_intr_info[vec].low32 &= ~IOSAPIC_MASK;
309                 list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list) {
310                         addr = rte->addr;
311                         rte_index = rte->rte_index;
312                         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
313                 }
314         }
315         spin_unlock_irqrestore(&iosapic_lock, flags);
316 }
317
318
319 static void
320 iosapic_set_affinity (unsigned int irq, cpumask_t mask)
321 {
322 #ifdef CONFIG_SMP
323         unsigned long flags;
324         u32 high32, low32;
325         int dest, rte_index;
326         char __iomem *addr;
327         int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
328         ia64_vector vec;
329         struct iosapic_rte_info *rte;
330
331         irq &= (~IA64_IRQ_REDIRECTED);
332         vec = irq_to_vector(irq);
333
334         if (cpus_empty(mask))
335                 return;
336
337         dest = cpu_physical_id(first_cpu(mask));
338
339         if (list_empty(&iosapic_intr_info[vec].rtes))
340                 return;                 /* not an IOSAPIC interrupt */
341
342         set_irq_affinity_info(irq, dest, redir);
343
344         /* dest contains both id and eid */
345         high32 = dest << IOSAPIC_DEST_SHIFT;
346
347         spin_lock_irqsave(&iosapic_lock, flags);
348         {
349                 low32 = iosapic_intr_info[vec].low32 & ~(7 << IOSAPIC_DELIVERY_SHIFT);
350
351                 if (redir)
352                         /* change delivery mode to lowest priority */
353                         low32 |= (IOSAPIC_LOWEST_PRIORITY << IOSAPIC_DELIVERY_SHIFT);
354                 else
355                         /* change delivery mode to fixed */
356                         low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
357
358                 iosapic_intr_info[vec].low32 = low32;
359                 iosapic_intr_info[vec].dest = dest;
360                 list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list) {
361                         addr = rte->addr;
362                         rte_index = rte->rte_index;
363                         iosapic_write(addr, IOSAPIC_RTE_HIGH(rte_index), high32);
364                         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
365                 }
366         }
367         spin_unlock_irqrestore(&iosapic_lock, flags);
368 #endif
369 }
370
371 /*
372  * Handlers for level-triggered interrupts.
373  */
374
375 static unsigned int
376 iosapic_startup_level_irq (unsigned int irq)
377 {
378         unmask_irq(irq);
379         return 0;
380 }
381
382 static void
383 iosapic_end_level_irq (unsigned int irq)
384 {
385         ia64_vector vec = irq_to_vector(irq);
386         struct iosapic_rte_info *rte;
387
388         move_irq(irq);
389         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list)
390                 iosapic_eoi(rte->addr, vec);
391 }
392
393 #define iosapic_shutdown_level_irq      mask_irq
394 #define iosapic_enable_level_irq        unmask_irq
395 #define iosapic_disable_level_irq       mask_irq
396 #define iosapic_ack_level_irq           nop
397
398 struct hw_interrupt_type irq_type_iosapic_level = {
399         .typename =     "IO-SAPIC-level",
400         .startup =      iosapic_startup_level_irq,
401         .shutdown =     iosapic_shutdown_level_irq,
402         .enable =       iosapic_enable_level_irq,
403         .disable =      iosapic_disable_level_irq,
404         .ack =          iosapic_ack_level_irq,
405         .end =          iosapic_end_level_irq,
406         .set_affinity = iosapic_set_affinity
407 };
408
409 /*
410  * Handlers for edge-triggered interrupts.
411  */
412
413 static unsigned int
414 iosapic_startup_edge_irq (unsigned int irq)
415 {
416         unmask_irq(irq);
417         /*
418          * IOSAPIC simply drops interrupts pended while the
419          * corresponding pin was masked, so we can't know if an
420          * interrupt is pending already.  Let's hope not...
421          */
422         return 0;
423 }
424
425 static void
426 iosapic_ack_edge_irq (unsigned int irq)
427 {
428         irq_desc_t *idesc = irq_descp(irq);
429
430         move_irq(irq);
431         /*
432          * Once we have recorded IRQ_PENDING already, we can mask the
433          * interrupt for real. This prevents IRQ storms from unhandled
434          * devices.
435          */
436         if ((idesc->status & (IRQ_PENDING|IRQ_DISABLED)) == (IRQ_PENDING|IRQ_DISABLED))
437                 mask_irq(irq);
438 }
439
440 #define iosapic_enable_edge_irq         unmask_irq
441 #define iosapic_disable_edge_irq        nop
442 #define iosapic_end_edge_irq            nop
443
444 struct hw_interrupt_type irq_type_iosapic_edge = {
445         .typename =     "IO-SAPIC-edge",
446         .startup =      iosapic_startup_edge_irq,
447         .shutdown =     iosapic_disable_edge_irq,
448         .enable =       iosapic_enable_edge_irq,
449         .disable =      iosapic_disable_edge_irq,
450         .ack =          iosapic_ack_edge_irq,
451         .end =          iosapic_end_edge_irq,
452         .set_affinity = iosapic_set_affinity
453 };
454
455 unsigned int
456 iosapic_version (char __iomem *addr)
457 {
458         /*
459          * IOSAPIC Version Register return 32 bit structure like:
460          * {
461          *      unsigned int version   : 8;
462          *      unsigned int reserved1 : 8;
463          *      unsigned int max_redir : 8;
464          *      unsigned int reserved2 : 8;
465          * }
466          */
467         return iosapic_read(addr, IOSAPIC_VERSION);
468 }
469
470 static int iosapic_find_sharable_vector (unsigned long trigger, unsigned long pol)
471 {
472         int i, vector = -1, min_count = -1;
473         struct iosapic_intr_info *info;
474
475         /*
476          * shared vectors for edge-triggered interrupts are not
477          * supported yet
478          */
479         if (trigger == IOSAPIC_EDGE)
480                 return -1;
481
482         for (i = IA64_FIRST_DEVICE_VECTOR; i <= IA64_LAST_DEVICE_VECTOR; i++) {
483                 info = &iosapic_intr_info[i];
484                 if (info->trigger == trigger && info->polarity == pol &&
485                     (info->dmode == IOSAPIC_FIXED || info->dmode == IOSAPIC_LOWEST_PRIORITY)) {
486                         if (min_count == -1 || info->count < min_count) {
487                                 vector = i;
488                                 min_count = info->count;
489                         }
490                 }
491         }
492
493         return vector;
494 }
495
496 /*
497  * if the given vector is already owned by other,
498  *  assign a new vector for the other and make the vector available
499  */
500 static void __init
501 iosapic_reassign_vector (int vector)
502 {
503         int new_vector;
504
505         if (!list_empty(&iosapic_intr_info[vector].rtes)) {
506                 new_vector = assign_irq_vector(AUTO_ASSIGN);
507                 if (new_vector < 0)
508                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
509                 printk(KERN_INFO "Reassigning vector %d to %d\n", vector, new_vector);
510                 memcpy(&iosapic_intr_info[new_vector], &iosapic_intr_info[vector],
511                        sizeof(struct iosapic_intr_info));
512                 INIT_LIST_HEAD(&iosapic_intr_info[new_vector].rtes);
513                 list_move(iosapic_intr_info[vector].rtes.next, &iosapic_intr_info[new_vector].rtes);
514                 memset(&iosapic_intr_info[vector], 0, sizeof(struct iosapic_intr_info));
515                 iosapic_intr_info[vector].low32 = IOSAPIC_MASK;
516                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
517         }
518 }
519
520 static struct iosapic_rte_info *iosapic_alloc_rte (void)
521 {
522         int i;
523         struct iosapic_rte_info *rte;
524         int preallocated = 0;
525
526         if (!iosapic_kmalloc_ok && list_empty(&free_rte_list)) {
527                 rte = alloc_bootmem(sizeof(struct iosapic_rte_info) * NR_PREALLOCATE_RTE_ENTRIES);
528                 if (!rte)
529                         return NULL;
530                 for (i = 0; i < NR_PREALLOCATE_RTE_ENTRIES; i++, rte++)
531                         list_add(&rte->rte_list, &free_rte_list);
532         }
533
534         if (!list_empty(&free_rte_list)) {
535                 rte = list_entry(free_rte_list.next, struct iosapic_rte_info, rte_list);
536                 list_del(&rte->rte_list);
537                 preallocated++;
538         } else {
539                 rte = kmalloc(sizeof(struct iosapic_rte_info), GFP_ATOMIC);
540                 if (!rte)
541                         return NULL;
542         }
543
544         memset(rte, 0, sizeof(struct iosapic_rte_info));
545         if (preallocated)
546                 rte->flags |= RTE_PREALLOCATED;
547
548         return rte;
549 }
550
551 static void iosapic_free_rte (struct iosapic_rte_info *rte)
552 {
553         if (rte->flags & RTE_PREALLOCATED)
554                 list_add_tail(&rte->rte_list, &free_rte_list);
555         else
556                 kfree(rte);
557 }
558
559 static inline int vector_is_shared (int vector)
560 {
561         return (iosapic_intr_info[vector].count > 1);
562 }
563
564 static int
565 register_intr (unsigned int gsi, int vector, unsigned char delivery,
566                unsigned long polarity, unsigned long trigger)
567 {
568         irq_desc_t *idesc;
569         struct hw_interrupt_type *irq_type;
570         int rte_index;
571         int index;
572         unsigned long gsi_base;
573         void __iomem *iosapic_address;
574         struct iosapic_rte_info *rte;
575
576         index = find_iosapic(gsi);
577         if (index < 0) {
578                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n", __FUNCTION__, gsi);
579                 return -ENODEV;
580         }
581
582         iosapic_address = iosapic_lists[index].addr;
583         gsi_base = iosapic_lists[index].gsi_base;
584
585         rte = gsi_vector_to_rte(gsi, vector);
586         if (!rte) {
587                 rte = iosapic_alloc_rte();
588                 if (!rte) {
589                         printk(KERN_WARNING "%s: cannot allocate memory\n", __FUNCTION__);
590                         return -ENOMEM;
591                 }
592
593                 rte_index = gsi - gsi_base;
594                 rte->rte_index  = rte_index;
595                 rte->addr       = iosapic_address;
596                 rte->gsi_base   = gsi_base;
597                 rte->refcnt++;
598                 list_add_tail(&rte->rte_list, &iosapic_intr_info[vector].rtes);
599                 iosapic_intr_info[vector].count++;
600                 iosapic_lists[index].rtes_inuse++;
601         }
602         else if (vector_is_shared(vector)) {
603                 struct iosapic_intr_info *info = &iosapic_intr_info[vector];
604                 if (info->trigger != trigger || info->polarity != polarity) {
605                         printk (KERN_WARNING "%s: cannot override the interrupt\n", __FUNCTION__);
606                         return -EINVAL;
607                 }
608         }
609
610         iosapic_intr_info[vector].polarity = polarity;
611         iosapic_intr_info[vector].dmode    = delivery;
612         iosapic_intr_info[vector].trigger  = trigger;
613
614         if (trigger == IOSAPIC_EDGE)
615                 irq_type = &irq_type_iosapic_edge;
616         else
617                 irq_type = &irq_type_iosapic_level;
618
619         idesc = irq_descp(vector);
620         if (idesc->handler != irq_type) {
621                 if (idesc->handler != &no_irq_type)
622                         printk(KERN_WARNING "%s: changing vector %d from %s to %s\n",
623                                __FUNCTION__, vector, idesc->handler->typename, irq_type->typename);
624                 idesc->handler = irq_type;
625         }
626         return 0;
627 }
628
629 static unsigned int
630 get_target_cpu (unsigned int gsi, int vector)
631 {
632 #ifdef CONFIG_SMP
633         static int cpu = -1;
634         extern int cpe_vector;
635
636         /*
637          * In case of vector shared by multiple RTEs, all RTEs that
638          * share the vector need to use the same destination CPU.
639          */
640         if (!list_empty(&iosapic_intr_info[vector].rtes))
641                 return iosapic_intr_info[vector].dest;
642
643         /*
644          * If the platform supports redirection via XTP, let it
645          * distribute interrupts.
646          */
647         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
648                 return cpu_physical_id(smp_processor_id());
649
650         /*
651          * Some interrupts (ACPI SCI, for instance) are registered
652          * before the BSP is marked as online.
653          */
654         if (!cpu_online(smp_processor_id()))
655                 return cpu_physical_id(smp_processor_id());
656
657 #ifdef CONFIG_ACPI
658                 if (cpe_vector > 0 && vector == IA64_CPEP_VECTOR)
659                         return get_cpei_target_cpu();
660 #endif
661
662 #ifdef CONFIG_NUMA
663         {
664                 int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
665                 cpumask_t cpu_mask;
666
667                 iosapic_index = find_iosapic(gsi);
668                 if (iosapic_index < 0 ||
669                     iosapic_lists[iosapic_index].node == MAX_NUMNODES)
670                         goto skip_numa_setup;
671
672                 cpu_mask = node_to_cpumask(iosapic_lists[iosapic_index].node);
673
674                 for_each_cpu_mask(numa_cpu, cpu_mask) {
675                         if (!cpu_online(numa_cpu))
676                                 cpu_clear(numa_cpu, cpu_mask);
677                 }
678
679                 num_cpus = cpus_weight(cpu_mask);
680
681                 if (!num_cpus)
682                         goto skip_numa_setup;
683
684                 /* Use vector assigment to distribute across cpus in node */
685                 cpu_index = vector % num_cpus;
686
687                 for (numa_cpu = first_cpu(cpu_mask) ; i < cpu_index ; i++)
688                         numa_cpu = next_cpu(numa_cpu, cpu_mask);
689
690                 if (numa_cpu != NR_CPUS)
691                         return cpu_physical_id(numa_cpu);
692         }
693 skip_numa_setup:
694 #endif
695         /*
696          * Otherwise, round-robin interrupt vectors across all the
697          * processors.  (It'd be nice if we could be smarter in the
698          * case of NUMA.)
699          */
700         do {
701                 if (++cpu >= NR_CPUS)
702                         cpu = 0;
703         } while (!cpu_online(cpu));
704
705         return cpu_physical_id(cpu);
706 #else
707         return cpu_physical_id(smp_processor_id());
708 #endif
709 }
710
711 /*
712  * ACPI can describe IOSAPIC interrupts via static tables and namespace
713  * methods.  This provides an interface to register those interrupts and
714  * program the IOSAPIC RTE.
715  */
716 int
717 iosapic_register_intr (unsigned int gsi,
718                        unsigned long polarity, unsigned long trigger)
719 {
720         int vector, mask = 1, err;
721         unsigned int dest;
722         unsigned long flags;
723         struct iosapic_rte_info *rte;
724         u32 low32;
725 again:
726         /*
727          * If this GSI has already been registered (i.e., it's a
728          * shared interrupt, or we lost a race to register it),
729          * don't touch the RTE.
730          */
731         spin_lock_irqsave(&iosapic_lock, flags);
732         {
733                 vector = gsi_to_vector(gsi);
734                 if (vector > 0) {
735                         rte = gsi_vector_to_rte(gsi, vector);
736                         rte->refcnt++;
737                         spin_unlock_irqrestore(&iosapic_lock, flags);
738                         return vector;
739                 }
740         }
741         spin_unlock_irqrestore(&iosapic_lock, flags);
742
743         /* If vector is running out, we try to find a sharable vector */
744         vector = assign_irq_vector(AUTO_ASSIGN);
745         if (vector < 0) {
746                 vector = iosapic_find_sharable_vector(trigger, polarity);
747                 if (vector < 0)
748                         return -ENOSPC;
749         }
750
751         spin_lock_irqsave(&irq_descp(vector)->lock, flags);
752         spin_lock(&iosapic_lock);
753         {
754                 if (gsi_to_vector(gsi) > 0) {
755                         if (list_empty(&iosapic_intr_info[vector].rtes))
756                                 free_irq_vector(vector);
757                         spin_unlock(&iosapic_lock);
758                         spin_unlock_irqrestore(&irq_descp(vector)->lock, flags);
759                         goto again;
760                 }
761
762                 dest = get_target_cpu(gsi, vector);
763                 err = register_intr(gsi, vector, IOSAPIC_LOWEST_PRIORITY,
764                               polarity, trigger);
765                 if (err < 0) {
766                         spin_unlock(&iosapic_lock);
767                         spin_unlock_irqrestore(&irq_descp(vector)->lock, flags);
768                         return err;
769                 }
770
771                 /*
772                  * If the vector is shared and already unmasked for
773                  * other interrupt sources, don't mask it.
774                  */
775                 low32 = iosapic_intr_info[vector].low32;
776                 if (vector_is_shared(vector) && !(low32 & IOSAPIC_MASK))
777                         mask = 0;
778                 set_rte(gsi, vector, dest, mask);
779         }
780         spin_unlock(&iosapic_lock);
781         spin_unlock_irqrestore(&irq_descp(vector)->lock, flags);
782
783         printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
784                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
785                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
786                cpu_logical_id(dest), dest, vector);
787
788         return vector;
789 }
790
791 void
792 iosapic_unregister_intr (unsigned int gsi)
793 {
794         unsigned long flags;
795         int irq, vector, index;
796         irq_desc_t *idesc;
797         u32 low32;
798         unsigned long trigger, polarity;
799         unsigned int dest;
800         struct iosapic_rte_info *rte;
801
802         /*
803          * If the irq associated with the gsi is not found,
804          * iosapic_unregister_intr() is unbalanced. We need to check
805          * this again after getting locks.
806          */
807         irq = gsi_to_irq(gsi);
808         if (irq < 0) {
809                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n", gsi);
810                 WARN_ON(1);
811                 return;
812         }
813         vector = irq_to_vector(irq);
814
815         idesc = irq_descp(irq);
816         spin_lock_irqsave(&idesc->lock, flags);
817         spin_lock(&iosapic_lock);
818         {
819                 if ((rte = gsi_vector_to_rte(gsi, vector)) == NULL) {
820                         printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n", gsi);
821                         WARN_ON(1);
822                         goto out;
823                 }
824
825                 if (--rte->refcnt > 0)
826                         goto out;
827
828                 /* Mask the interrupt */
829                 low32 = iosapic_intr_info[vector].low32 | IOSAPIC_MASK;
830                 iosapic_write(rte->addr, IOSAPIC_RTE_LOW(rte->rte_index), low32);
831
832                 /* Remove the rte entry from the list */
833                 list_del(&rte->rte_list);
834                 iosapic_intr_info[vector].count--;
835                 iosapic_free_rte(rte);
836                 index = find_iosapic(gsi);
837                 iosapic_lists[index].rtes_inuse--;
838                 WARN_ON(iosapic_lists[index].rtes_inuse < 0);
839
840                 trigger  = iosapic_intr_info[vector].trigger;
841                 polarity = iosapic_intr_info[vector].polarity;
842                 dest     = iosapic_intr_info[vector].dest;
843                 printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d unregistered\n",
844                        gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
845                        (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
846                        cpu_logical_id(dest), dest, vector);
847
848                 if (list_empty(&iosapic_intr_info[vector].rtes)) {
849                         /* Sanity check */
850                         BUG_ON(iosapic_intr_info[vector].count);
851
852                         /* Clear the interrupt controller descriptor */
853                         idesc->handler = &no_irq_type;
854
855                         /* Clear the interrupt information */
856                         memset(&iosapic_intr_info[vector], 0, sizeof(struct iosapic_intr_info));
857                         iosapic_intr_info[vector].low32 |= IOSAPIC_MASK;
858                         INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
859
860                         if (idesc->action) {
861                                 printk(KERN_ERR "interrupt handlers still exist on IRQ %u\n", irq);
862                                 WARN_ON(1);
863                         }
864
865                         /* Free the interrupt vector */
866                         free_irq_vector(vector);
867                 }
868         }
869  out:
870         spin_unlock(&iosapic_lock);
871         spin_unlock_irqrestore(&idesc->lock, flags);
872 }
873
874 /*
875  * ACPI calls this when it finds an entry for a platform interrupt.
876  * Note that the irq_base and IOSAPIC address must be set in iosapic_init().
877  */
878 int __init
879 iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
880                                 int iosapic_vector, u16 eid, u16 id,
881                                 unsigned long polarity, unsigned long trigger)
882 {
883         static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
884         unsigned char delivery;
885         int vector, mask = 0;
886         unsigned int dest = ((id << 8) | eid) & 0xffff;
887
888         switch (int_type) {
889               case ACPI_INTERRUPT_PMI:
890                 vector = iosapic_vector;
891                 /*
892                  * since PMI vector is alloc'd by FW(ACPI) not by kernel,
893                  * we need to make sure the vector is available
894                  */
895                 iosapic_reassign_vector(vector);
896                 delivery = IOSAPIC_PMI;
897                 break;
898               case ACPI_INTERRUPT_INIT:
899                 vector = assign_irq_vector(AUTO_ASSIGN);
900                 if (vector < 0)
901                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
902                 delivery = IOSAPIC_INIT;
903                 break;
904               case ACPI_INTERRUPT_CPEI:
905                 vector = IA64_CPE_VECTOR;
906                 delivery = IOSAPIC_LOWEST_PRIORITY;
907                 mask = 1;
908                 break;
909               default:
910                 printk(KERN_ERR "iosapic_register_platform_irq(): invalid int type 0x%x\n", int_type);
911                 return -1;
912         }
913
914         register_intr(gsi, vector, delivery, polarity, trigger);
915
916         printk(KERN_INFO "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
917                int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
918                int_type, gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
919                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
920                cpu_logical_id(dest), dest, vector);
921
922         set_rte(gsi, vector, dest, mask);
923         return vector;
924 }
925
926
927 /*
928  * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
929  * Note that the gsi_base and IOSAPIC address must be set in iosapic_init().
930  */
931 void __init
932 iosapic_override_isa_irq (unsigned int isa_irq, unsigned int gsi,
933                           unsigned long polarity,
934                           unsigned long trigger)
935 {
936         int vector;
937         unsigned int dest = cpu_physical_id(smp_processor_id());
938
939         vector = isa_irq_to_vector(isa_irq);
940
941         register_intr(gsi, vector, IOSAPIC_LOWEST_PRIORITY, polarity, trigger);
942
943         DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
944             isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
945             polarity == IOSAPIC_POL_HIGH ? "high" : "low",
946             cpu_logical_id(dest), dest, vector);
947
948         set_rte(gsi, vector, dest, 1);
949 }
950
951 void __init
952 iosapic_system_init (int system_pcat_compat)
953 {
954         int vector;
955
956         for (vector = 0; vector < IA64_NUM_VECTORS; ++vector) {
957                 iosapic_intr_info[vector].low32 = IOSAPIC_MASK;
958                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);        /* mark as unused */
959         }
960
961         pcat_compat = system_pcat_compat;
962         if (pcat_compat) {
963                 /*
964                  * Disable the compatibility mode interrupts (8259 style), needs IN/OUT support
965                  * enabled.
966                  */
967                 printk(KERN_INFO "%s: Disabling PC-AT compatible 8259 interrupts\n", __FUNCTION__);
968                 outb(0xff, 0xA1);
969                 outb(0xff, 0x21);
970         }
971 }
972
973 static inline int
974 iosapic_alloc (void)
975 {
976         int index;
977
978         for (index = 0; index < NR_IOSAPICS; index++)
979                 if (!iosapic_lists[index].addr)
980                         return index;
981
982         printk(KERN_WARNING "%s: failed to allocate iosapic\n", __FUNCTION__);
983         return -1;
984 }
985
986 static inline void
987 iosapic_free (int index)
988 {
989         memset(&iosapic_lists[index], 0, sizeof(iosapic_lists[0]));
990 }
991
992 static inline int
993 iosapic_check_gsi_range (unsigned int gsi_base, unsigned int ver)
994 {
995         int index;
996         unsigned int gsi_end, base, end;
997
998         /* check gsi range */
999         gsi_end = gsi_base + ((ver >> 16) & 0xff);
1000         for (index = 0; index < NR_IOSAPICS; index++) {
1001                 if (!iosapic_lists[index].addr)
1002                         continue;
1003
1004                 base = iosapic_lists[index].gsi_base;
1005                 end  = base + iosapic_lists[index].num_rte - 1;
1006
1007                 if (gsi_base < base && gsi_end < base)
1008                         continue;/* OK */
1009
1010                 if (gsi_base > end && gsi_end > end)
1011                         continue; /* OK */
1012
1013                 return -EBUSY;
1014         }
1015         return 0;
1016 }
1017
1018 int __devinit
1019 iosapic_init (unsigned long phys_addr, unsigned int gsi_base)
1020 {
1021         int num_rte, err, index;
1022         unsigned int isa_irq, ver;
1023         char __iomem *addr;
1024         unsigned long flags;
1025
1026         spin_lock_irqsave(&iosapic_lock, flags);
1027         {
1028                 addr = ioremap(phys_addr, 0);
1029                 ver = iosapic_version(addr);
1030
1031                 if ((err = iosapic_check_gsi_range(gsi_base, ver))) {
1032                         iounmap(addr);
1033                         spin_unlock_irqrestore(&iosapic_lock, flags);
1034                         return err;
1035                 }
1036
1037                 /*
1038                  * The MAX_REDIR register holds the highest input pin
1039                  * number (starting from 0).
1040                  * We add 1 so that we can use it for number of pins (= RTEs)
1041                  */
1042                 num_rte = ((ver >> 16) & 0xff) + 1;
1043
1044                 index = iosapic_alloc();
1045                 iosapic_lists[index].addr = addr;
1046                 iosapic_lists[index].gsi_base = gsi_base;
1047                 iosapic_lists[index].num_rte = num_rte;
1048 #ifdef CONFIG_NUMA
1049                 iosapic_lists[index].node = MAX_NUMNODES;
1050 #endif
1051         }
1052         spin_unlock_irqrestore(&iosapic_lock, flags);
1053
1054         if ((gsi_base == 0) && pcat_compat) {
1055                 /*
1056                  * Map the legacy ISA devices into the IOSAPIC data.  Some of these may
1057                  * get reprogrammed later on with data from the ACPI Interrupt Source
1058                  * Override table.
1059                  */
1060                 for (isa_irq = 0; isa_irq < 16; ++isa_irq)
1061                         iosapic_override_isa_irq(isa_irq, isa_irq, IOSAPIC_POL_HIGH, IOSAPIC_EDGE);
1062         }
1063         return 0;
1064 }
1065
1066 #ifdef CONFIG_HOTPLUG
1067 int
1068 iosapic_remove (unsigned int gsi_base)
1069 {
1070         int index, err = 0;
1071         unsigned long flags;
1072
1073         spin_lock_irqsave(&iosapic_lock, flags);
1074         {
1075                 index = find_iosapic(gsi_base);
1076                 if (index < 0) {
1077                         printk(KERN_WARNING "%s: No IOSAPIC for GSI base %u\n",
1078                                __FUNCTION__, gsi_base);
1079                         goto out;
1080                 }
1081
1082                 if (iosapic_lists[index].rtes_inuse) {
1083                         err = -EBUSY;
1084                         printk(KERN_WARNING "%s: IOSAPIC for GSI base %u is busy\n",
1085                                __FUNCTION__, gsi_base);
1086                         goto out;
1087                 }
1088
1089                 iounmap(iosapic_lists[index].addr);
1090                 iosapic_free(index);
1091         }
1092  out:
1093         spin_unlock_irqrestore(&iosapic_lock, flags);
1094         return err;
1095 }
1096 #endif /* CONFIG_HOTPLUG */
1097
1098 #ifdef CONFIG_NUMA
1099 void __devinit
1100 map_iosapic_to_node(unsigned int gsi_base, int node)
1101 {
1102         int index;
1103
1104         index = find_iosapic(gsi_base);
1105         if (index < 0) {
1106                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1107                        __FUNCTION__, gsi_base);
1108                 return;
1109         }
1110         iosapic_lists[index].node = node;
1111         return;
1112 }
1113 #endif
1114
1115 static int __init iosapic_enable_kmalloc (void)
1116 {
1117         iosapic_kmalloc_ok = 1;
1118         return 0;
1119 }
1120 core_initcall (iosapic_enable_kmalloc);