]> rtime.felk.cvut.cz Git - linux-imx.git/blob - arch/ia64/kernel/iosapic.c
29ec86c241941da05fcb497a3ac79afd0fb385b5
[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 } iosapic_lists[NR_IOSAPICS];
129
130 struct iosapic_rte_info {
131         struct list_head rte_list;      /* RTEs sharing the same vector */
132         char            rte_index;      /* IOSAPIC RTE index */
133         int             refcnt;         /* reference counter */
134         unsigned int    flags;          /* flags */
135         struct iosapic  *iosapic;
136 } ____cacheline_aligned;
137
138 static struct iosapic_intr_info {
139         struct list_head rtes;          /* RTEs using this vector (empty =>
140                                          * not an IOSAPIC interrupt) */
141         int             count;          /* # of RTEs that shares this vector */
142         u32             low32;          /* current value of low word of
143                                          * Redirection table entry */
144         unsigned int    dest;           /* destination CPU physical ID */
145         unsigned char   dmode   : 3;    /* delivery mode (see iosapic.h) */
146         unsigned char   polarity: 1;    /* interrupt polarity
147                                          * (see iosapic.h) */
148         unsigned char   trigger : 1;    /* trigger mode (see iosapic.h) */
149 } iosapic_intr_info[IA64_NUM_VECTORS];
150
151 static unsigned char pcat_compat __devinitdata; /* 8259 compatibility flag */
152
153 static int iosapic_kmalloc_ok;
154 static LIST_HEAD(free_rte_list);
155
156 /*
157  * Find an IOSAPIC associated with a GSI
158  */
159 static inline int
160 find_iosapic (unsigned int gsi)
161 {
162         int i;
163
164         for (i = 0; i < NR_IOSAPICS; i++) {
165                 if ((unsigned) (gsi - iosapic_lists[i].gsi_base) <
166                     iosapic_lists[i].num_rte)
167                         return i;
168         }
169
170         return -1;
171 }
172
173 static inline int
174 _gsi_to_vector (unsigned int gsi)
175 {
176         struct iosapic_intr_info *info;
177         struct iosapic_rte_info *rte;
178
179         for (info = iosapic_intr_info; info <
180                      iosapic_intr_info + IA64_NUM_VECTORS; ++info)
181                 list_for_each_entry(rte, &info->rtes, rte_list)
182                         if (rte->iosapic->gsi_base + rte->rte_index == gsi)
183                                 return info - iosapic_intr_info;
184         return -1;
185 }
186
187 /*
188  * Translate GSI number to the corresponding IA-64 interrupt vector.  If no
189  * entry exists, return -1.
190  */
191 inline int
192 gsi_to_vector (unsigned int gsi)
193 {
194         return _gsi_to_vector(gsi);
195 }
196
197 int
198 gsi_to_irq (unsigned int gsi)
199 {
200         unsigned long flags;
201         int irq;
202         /*
203          * XXX fix me: this assumes an identity mapping between IA-64 vector
204          * and Linux irq numbers...
205          */
206         spin_lock_irqsave(&iosapic_lock, flags);
207         irq = _gsi_to_vector(gsi);
208         spin_unlock_irqrestore(&iosapic_lock, flags);
209
210         return irq;
211 }
212
213 static struct iosapic_rte_info *gsi_vector_to_rte(unsigned int gsi,
214                                                   unsigned int vec)
215 {
216         struct iosapic_rte_info *rte;
217
218         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list)
219                 if (rte->iosapic->gsi_base + rte->rte_index == gsi)
220                         return rte;
221         return NULL;
222 }
223
224 static void
225 set_rte (unsigned int gsi, unsigned int vector, unsigned int dest, int mask)
226 {
227         unsigned long pol, trigger, dmode;
228         u32 low32, high32;
229         char __iomem *addr;
230         int rte_index;
231         char redir;
232         struct iosapic_rte_info *rte;
233
234         DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
235
236         rte = gsi_vector_to_rte(gsi, vector);
237         if (!rte)
238                 return;         /* not an IOSAPIC interrupt */
239
240         rte_index = rte->rte_index;
241         addr    = rte->iosapic->addr;
242         pol     = iosapic_intr_info[vector].polarity;
243         trigger = iosapic_intr_info[vector].trigger;
244         dmode   = iosapic_intr_info[vector].dmode;
245
246         redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
247
248 #ifdef CONFIG_SMP
249         {
250                 unsigned int irq;
251
252                 for (irq = 0; irq < NR_IRQS; ++irq)
253                         if (irq_to_vector(irq) == vector) {
254                                 set_irq_affinity_info(irq,
255                                                       (int)(dest & 0xffff),
256                                                       redir);
257                                 break;
258                         }
259         }
260 #endif
261
262         low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
263                  (trigger << IOSAPIC_TRIGGER_SHIFT) |
264                  (dmode << IOSAPIC_DELIVERY_SHIFT) |
265                  ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
266                  vector);
267
268         /* dest contains both id and eid */
269         high32 = (dest << IOSAPIC_DEST_SHIFT);
270
271         iosapic_write(addr, IOSAPIC_RTE_HIGH(rte_index), high32);
272         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
273         iosapic_intr_info[vector].low32 = low32;
274         iosapic_intr_info[vector].dest = dest;
275 }
276
277 static void
278 nop (unsigned int irq)
279 {
280         /* do nothing... */
281 }
282
283
284 #ifdef CONFIG_KEXEC
285 void
286 kexec_disable_iosapic(void)
287 {
288         struct iosapic_intr_info *info;
289         struct iosapic_rte_info *rte;
290         u8 vec = 0;
291         for (info = iosapic_intr_info; info <
292                         iosapic_intr_info + IA64_NUM_VECTORS; ++info, ++vec) {
293                 list_for_each_entry(rte, &info->rtes,
294                                 rte_list) {
295                         iosapic_write(rte->iosapic->addr,
296                                         IOSAPIC_RTE_LOW(rte->rte_index),
297                                         IOSAPIC_MASK|vec);
298                         iosapic_eoi(rte->iosapic->addr, vec);
299                 }
300         }
301 }
302 #endif
303
304 static void
305 mask_irq (unsigned int irq)
306 {
307         unsigned long flags;
308         char __iomem *addr;
309         u32 low32;
310         int rte_index;
311         ia64_vector vec = irq_to_vector(irq);
312         struct iosapic_rte_info *rte;
313
314         if (list_empty(&iosapic_intr_info[vec].rtes))
315                 return;                 /* not an IOSAPIC interrupt! */
316
317         spin_lock_irqsave(&iosapic_lock, flags);
318         /* set only the mask bit */
319         low32 = iosapic_intr_info[vec].low32 |= IOSAPIC_MASK;
320         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list) {
321                 addr = rte->iosapic->addr;
322                 rte_index = rte->rte_index;
323                 iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
324         }
325         spin_unlock_irqrestore(&iosapic_lock, flags);
326 }
327
328 static void
329 unmask_irq (unsigned int irq)
330 {
331         unsigned long flags;
332         char __iomem *addr;
333         u32 low32;
334         int rte_index;
335         ia64_vector vec = irq_to_vector(irq);
336         struct iosapic_rte_info *rte;
337
338         if (list_empty(&iosapic_intr_info[vec].rtes))
339                 return;                 /* not an IOSAPIC interrupt! */
340
341         spin_lock_irqsave(&iosapic_lock, flags);
342         low32 = iosapic_intr_info[vec].low32 &= ~IOSAPIC_MASK;
343         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list) {
344                 addr = rte->iosapic->addr;
345                 rte_index = rte->rte_index;
346                 iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
347         }
348         spin_unlock_irqrestore(&iosapic_lock, flags);
349 }
350
351
352 static void
353 iosapic_set_affinity (unsigned int irq, cpumask_t mask)
354 {
355 #ifdef CONFIG_SMP
356         unsigned long flags;
357         u32 high32, low32;
358         int dest, rte_index;
359         char __iomem *addr;
360         int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
361         ia64_vector vec;
362         struct iosapic_rte_info *rte;
363
364         irq &= (~IA64_IRQ_REDIRECTED);
365         vec = irq_to_vector(irq);
366
367         if (cpus_empty(mask))
368                 return;
369
370         dest = cpu_physical_id(first_cpu(mask));
371
372         if (list_empty(&iosapic_intr_info[vec].rtes))
373                 return;                 /* not an IOSAPIC interrupt */
374
375         set_irq_affinity_info(irq, dest, redir);
376
377         /* dest contains both id and eid */
378         high32 = dest << IOSAPIC_DEST_SHIFT;
379
380         spin_lock_irqsave(&iosapic_lock, flags);
381         low32 = iosapic_intr_info[vec].low32 & ~(7 << IOSAPIC_DELIVERY_SHIFT);
382         if (redir)
383                 /* change delivery mode to lowest priority */
384                 low32 |= (IOSAPIC_LOWEST_PRIORITY << IOSAPIC_DELIVERY_SHIFT);
385         else
386                 /* change delivery mode to fixed */
387                 low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
388
389         iosapic_intr_info[vec].low32 = low32;
390         iosapic_intr_info[vec].dest = dest;
391         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list) {
392                 addr = rte->iosapic->addr;
393                 rte_index = rte->rte_index;
394                 iosapic_write(addr, IOSAPIC_RTE_HIGH(rte_index), high32);
395                 iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
396         }
397         spin_unlock_irqrestore(&iosapic_lock, flags);
398 #endif
399 }
400
401 /*
402  * Handlers for level-triggered interrupts.
403  */
404
405 static unsigned int
406 iosapic_startup_level_irq (unsigned int irq)
407 {
408         unmask_irq(irq);
409         return 0;
410 }
411
412 static void
413 iosapic_end_level_irq (unsigned int irq)
414 {
415         ia64_vector vec = irq_to_vector(irq);
416         struct iosapic_rte_info *rte;
417
418         move_native_irq(irq);
419         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list)
420                 iosapic_eoi(rte->iosapic->addr, vec);
421 }
422
423 #define iosapic_shutdown_level_irq      mask_irq
424 #define iosapic_enable_level_irq        unmask_irq
425 #define iosapic_disable_level_irq       mask_irq
426 #define iosapic_ack_level_irq           nop
427
428 struct irq_chip irq_type_iosapic_level = {
429         .name =         "IO-SAPIC-level",
430         .startup =      iosapic_startup_level_irq,
431         .shutdown =     iosapic_shutdown_level_irq,
432         .enable =       iosapic_enable_level_irq,
433         .disable =      iosapic_disable_level_irq,
434         .ack =          iosapic_ack_level_irq,
435         .end =          iosapic_end_level_irq,
436         .mask =         mask_irq,
437         .unmask =       unmask_irq,
438         .set_affinity = iosapic_set_affinity
439 };
440
441 /*
442  * Handlers for edge-triggered interrupts.
443  */
444
445 static unsigned int
446 iosapic_startup_edge_irq (unsigned int irq)
447 {
448         unmask_irq(irq);
449         /*
450          * IOSAPIC simply drops interrupts pended while the
451          * corresponding pin was masked, so we can't know if an
452          * interrupt is pending already.  Let's hope not...
453          */
454         return 0;
455 }
456
457 static void
458 iosapic_ack_edge_irq (unsigned int irq)
459 {
460         irq_desc_t *idesc = irq_desc + irq;
461
462         move_native_irq(irq);
463         /*
464          * Once we have recorded IRQ_PENDING already, we can mask the
465          * interrupt for real. This prevents IRQ storms from unhandled
466          * devices.
467          */
468         if ((idesc->status & (IRQ_PENDING|IRQ_DISABLED)) ==
469             (IRQ_PENDING|IRQ_DISABLED))
470                 mask_irq(irq);
471 }
472
473 #define iosapic_enable_edge_irq         unmask_irq
474 #define iosapic_disable_edge_irq        nop
475 #define iosapic_end_edge_irq            nop
476
477 struct irq_chip irq_type_iosapic_edge = {
478         .name =         "IO-SAPIC-edge",
479         .startup =      iosapic_startup_edge_irq,
480         .shutdown =     iosapic_disable_edge_irq,
481         .enable =       iosapic_enable_edge_irq,
482         .disable =      iosapic_disable_edge_irq,
483         .ack =          iosapic_ack_edge_irq,
484         .end =          iosapic_end_edge_irq,
485         .mask =         mask_irq,
486         .unmask =       unmask_irq,
487         .set_affinity = iosapic_set_affinity
488 };
489
490 unsigned int
491 iosapic_version (char __iomem *addr)
492 {
493         /*
494          * IOSAPIC Version Register return 32 bit structure like:
495          * {
496          *      unsigned int version   : 8;
497          *      unsigned int reserved1 : 8;
498          *      unsigned int max_redir : 8;
499          *      unsigned int reserved2 : 8;
500          * }
501          */
502         return iosapic_read(addr, IOSAPIC_VERSION);
503 }
504
505 static int iosapic_find_sharable_vector (unsigned long trigger,
506                                          unsigned long pol)
507 {
508         int i, vector = -1, min_count = -1;
509         struct iosapic_intr_info *info;
510
511         /*
512          * shared vectors for edge-triggered interrupts are not
513          * supported yet
514          */
515         if (trigger == IOSAPIC_EDGE)
516                 return -1;
517
518         for (i = IA64_FIRST_DEVICE_VECTOR; i <= IA64_LAST_DEVICE_VECTOR; i++) {
519                 info = &iosapic_intr_info[i];
520                 if (info->trigger == trigger && info->polarity == pol &&
521                     (info->dmode == IOSAPIC_FIXED || info->dmode ==
522                      IOSAPIC_LOWEST_PRIORITY)) {
523                         if (min_count == -1 || info->count < min_count) {
524                                 vector = i;
525                                 min_count = info->count;
526                         }
527                 }
528         }
529
530         return vector;
531 }
532
533 /*
534  * if the given vector is already owned by other,
535  *  assign a new vector for the other and make the vector available
536  */
537 static void __init
538 iosapic_reassign_vector (int vector)
539 {
540         int new_vector;
541
542         if (!list_empty(&iosapic_intr_info[vector].rtes)) {
543                 new_vector = assign_irq_vector(AUTO_ASSIGN);
544                 if (new_vector < 0)
545                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
546                 printk(KERN_INFO "Reassigning vector %d to %d\n",
547                        vector, new_vector);
548                 memcpy(&iosapic_intr_info[new_vector], &iosapic_intr_info[vector],
549                        sizeof(struct iosapic_intr_info));
550                 INIT_LIST_HEAD(&iosapic_intr_info[new_vector].rtes);
551                 list_move(iosapic_intr_info[vector].rtes.next,
552                           &iosapic_intr_info[new_vector].rtes);
553                 memset(&iosapic_intr_info[vector], 0,
554                        sizeof(struct iosapic_intr_info));
555                 iosapic_intr_info[vector].low32 = IOSAPIC_MASK;
556                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
557         }
558 }
559
560 static struct iosapic_rte_info *iosapic_alloc_rte (void)
561 {
562         int i;
563         struct iosapic_rte_info *rte;
564         int preallocated = 0;
565
566         if (!iosapic_kmalloc_ok && list_empty(&free_rte_list)) {
567                 rte = alloc_bootmem(sizeof(struct iosapic_rte_info) *
568                                     NR_PREALLOCATE_RTE_ENTRIES);
569                 if (!rte)
570                         return NULL;
571                 for (i = 0; i < NR_PREALLOCATE_RTE_ENTRIES; i++, rte++)
572                         list_add(&rte->rte_list, &free_rte_list);
573         }
574
575         if (!list_empty(&free_rte_list)) {
576                 rte = list_entry(free_rte_list.next, struct iosapic_rte_info,
577                                  rte_list);
578                 list_del(&rte->rte_list);
579                 preallocated++;
580         } else {
581                 rte = kmalloc(sizeof(struct iosapic_rte_info), GFP_ATOMIC);
582                 if (!rte)
583                         return NULL;
584         }
585
586         memset(rte, 0, sizeof(struct iosapic_rte_info));
587         if (preallocated)
588                 rte->flags |= RTE_PREALLOCATED;
589
590         return rte;
591 }
592
593 static void iosapic_free_rte (struct iosapic_rte_info *rte)
594 {
595         if (rte->flags & RTE_PREALLOCATED)
596                 list_add_tail(&rte->rte_list, &free_rte_list);
597         else
598                 kfree(rte);
599 }
600
601 static inline int vector_is_shared (int vector)
602 {
603         return (iosapic_intr_info[vector].count > 1);
604 }
605
606 static int
607 register_intr (unsigned int gsi, int vector, unsigned char delivery,
608                unsigned long polarity, unsigned long trigger)
609 {
610         irq_desc_t *idesc;
611         struct hw_interrupt_type *irq_type;
612         int index;
613         struct iosapic_rte_info *rte;
614
615         index = find_iosapic(gsi);
616         if (index < 0) {
617                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
618                        __FUNCTION__, gsi);
619                 return -ENODEV;
620         }
621
622         rte = gsi_vector_to_rte(gsi, vector);
623         if (!rte) {
624                 rte = iosapic_alloc_rte();
625                 if (!rte) {
626                         printk(KERN_WARNING "%s: cannot allocate memory\n",
627                                __FUNCTION__);
628                         return -ENOMEM;
629                 }
630
631                 rte->iosapic    = &iosapic_lists[index];
632                 rte->rte_index  = gsi - rte->iosapic->gsi_base;
633                 rte->refcnt++;
634                 list_add_tail(&rte->rte_list, &iosapic_intr_info[vector].rtes);
635                 iosapic_intr_info[vector].count++;
636                 iosapic_lists[index].rtes_inuse++;
637         }
638         else if (vector_is_shared(vector)) {
639                 struct iosapic_intr_info *info = &iosapic_intr_info[vector];
640                 if (info->trigger != trigger || info->polarity != polarity) {
641                         printk (KERN_WARNING
642                                 "%s: cannot override the interrupt\n",
643                                 __FUNCTION__);
644                         return -EINVAL;
645                 }
646         }
647
648         iosapic_intr_info[vector].polarity = polarity;
649         iosapic_intr_info[vector].dmode    = delivery;
650         iosapic_intr_info[vector].trigger  = trigger;
651
652         if (trigger == IOSAPIC_EDGE)
653                 irq_type = &irq_type_iosapic_edge;
654         else
655                 irq_type = &irq_type_iosapic_level;
656
657         idesc = irq_desc + vector;
658         if (idesc->chip != irq_type) {
659                 if (idesc->chip != &no_irq_type)
660                         printk(KERN_WARNING
661                                "%s: changing vector %d from %s to %s\n",
662                                __FUNCTION__, vector,
663                                idesc->chip->name, irq_type->name);
664                 idesc->chip = irq_type;
665         }
666         return 0;
667 }
668
669 static unsigned int
670 get_target_cpu (unsigned int gsi, int vector)
671 {
672 #ifdef CONFIG_SMP
673         static int cpu = -1;
674         extern int cpe_vector;
675
676         /*
677          * In case of vector shared by multiple RTEs, all RTEs that
678          * share the vector need to use the same destination CPU.
679          */
680         if (!list_empty(&iosapic_intr_info[vector].rtes))
681                 return iosapic_intr_info[vector].dest;
682
683         /*
684          * If the platform supports redirection via XTP, let it
685          * distribute interrupts.
686          */
687         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
688                 return cpu_physical_id(smp_processor_id());
689
690         /*
691          * Some interrupts (ACPI SCI, for instance) are registered
692          * before the BSP is marked as online.
693          */
694         if (!cpu_online(smp_processor_id()))
695                 return cpu_physical_id(smp_processor_id());
696
697 #ifdef CONFIG_ACPI
698         if (cpe_vector > 0 && vector == IA64_CPEP_VECTOR)
699                 return get_cpei_target_cpu();
700 #endif
701
702 #ifdef CONFIG_NUMA
703         {
704                 int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
705                 cpumask_t cpu_mask;
706
707                 iosapic_index = find_iosapic(gsi);
708                 if (iosapic_index < 0 ||
709                     iosapic_lists[iosapic_index].node == MAX_NUMNODES)
710                         goto skip_numa_setup;
711
712                 cpu_mask = node_to_cpumask(iosapic_lists[iosapic_index].node);
713
714                 for_each_cpu_mask(numa_cpu, cpu_mask) {
715                         if (!cpu_online(numa_cpu))
716                                 cpu_clear(numa_cpu, cpu_mask);
717                 }
718
719                 num_cpus = cpus_weight(cpu_mask);
720
721                 if (!num_cpus)
722                         goto skip_numa_setup;
723
724                 /* Use vector assignment to distribute across cpus in node */
725                 cpu_index = vector % num_cpus;
726
727                 for (numa_cpu = first_cpu(cpu_mask) ; i < cpu_index ; i++)
728                         numa_cpu = next_cpu(numa_cpu, cpu_mask);
729
730                 if (numa_cpu != NR_CPUS)
731                         return cpu_physical_id(numa_cpu);
732         }
733 skip_numa_setup:
734 #endif
735         /*
736          * Otherwise, round-robin interrupt vectors across all the
737          * processors.  (It'd be nice if we could be smarter in the
738          * case of NUMA.)
739          */
740         do {
741                 if (++cpu >= NR_CPUS)
742                         cpu = 0;
743         } while (!cpu_online(cpu));
744
745         return cpu_physical_id(cpu);
746 #else  /* CONFIG_SMP */
747         return cpu_physical_id(smp_processor_id());
748 #endif
749 }
750
751 /*
752  * ACPI can describe IOSAPIC interrupts via static tables and namespace
753  * methods.  This provides an interface to register those interrupts and
754  * program the IOSAPIC RTE.
755  */
756 int
757 iosapic_register_intr (unsigned int gsi,
758                        unsigned long polarity, unsigned long trigger)
759 {
760         int vector, mask = 1, err;
761         unsigned int dest;
762         unsigned long flags;
763         struct iosapic_rte_info *rte;
764         u32 low32;
765 again:
766         /*
767          * If this GSI has already been registered (i.e., it's a
768          * shared interrupt, or we lost a race to register it),
769          * don't touch the RTE.
770          */
771         spin_lock_irqsave(&iosapic_lock, flags);
772         vector = gsi_to_vector(gsi);
773         if (vector > 0) {
774                 rte = gsi_vector_to_rte(gsi, vector);
775                 rte->refcnt++;
776                 spin_unlock_irqrestore(&iosapic_lock, flags);
777                 return vector;
778         }
779         spin_unlock_irqrestore(&iosapic_lock, flags);
780
781         /* If vector is running out, we try to find a sharable vector */
782         vector = assign_irq_vector(AUTO_ASSIGN);
783         if (vector < 0) {
784                 vector = iosapic_find_sharable_vector(trigger, polarity);
785                 if (vector < 0)
786                         return -ENOSPC;
787         }
788
789         spin_lock_irqsave(&irq_desc[vector].lock, flags);
790         spin_lock(&iosapic_lock);
791         if (gsi_to_vector(gsi) > 0) {
792                 if (list_empty(&iosapic_intr_info[vector].rtes))
793                         free_irq_vector(vector);
794                 spin_unlock(&iosapic_lock);
795                 spin_unlock_irqrestore(&irq_desc[vector].lock, flags);
796                 goto again;
797         }
798
799         dest = get_target_cpu(gsi, vector);
800         err = register_intr(gsi, vector, IOSAPIC_LOWEST_PRIORITY,
801                             polarity, trigger);
802         if (err < 0) {
803                 spin_unlock(&iosapic_lock);
804                 spin_unlock_irqrestore(&irq_desc[vector].lock, flags);
805                 return err;
806         }
807
808         /*
809          * If the vector is shared and already unmasked for other
810          * interrupt sources, don't mask it.
811          */
812         low32 = iosapic_intr_info[vector].low32;
813         if (vector_is_shared(vector) && !(low32 & IOSAPIC_MASK))
814                 mask = 0;
815         set_rte(gsi, vector, dest, mask);
816         spin_unlock(&iosapic_lock);
817         spin_unlock_irqrestore(&irq_desc[vector].lock, flags);
818
819         printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
820                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
821                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
822                cpu_logical_id(dest), dest, vector);
823
824         return vector;
825 }
826
827 void
828 iosapic_unregister_intr (unsigned int gsi)
829 {
830         unsigned long flags;
831         int irq, vector, index;
832         irq_desc_t *idesc;
833         u32 low32;
834         unsigned long trigger, polarity;
835         unsigned int dest;
836         struct iosapic_rte_info *rte;
837
838         /*
839          * If the irq associated with the gsi is not found,
840          * iosapic_unregister_intr() is unbalanced. We need to check
841          * this again after getting locks.
842          */
843         irq = gsi_to_irq(gsi);
844         if (irq < 0) {
845                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
846                        gsi);
847                 WARN_ON(1);
848                 return;
849         }
850         vector = irq_to_vector(irq);
851
852         idesc = irq_desc + irq;
853         spin_lock_irqsave(&idesc->lock, flags);
854         spin_lock(&iosapic_lock);
855         if ((rte = gsi_vector_to_rte(gsi, vector)) == NULL) {
856                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
857                        gsi);
858                 WARN_ON(1);
859                 goto out;
860         }
861
862         if (--rte->refcnt > 0)
863                 goto out;
864
865         /* Mask the interrupt */
866         low32 = iosapic_intr_info[vector].low32 | IOSAPIC_MASK;
867         iosapic_write(rte->iosapic->addr,
868                       IOSAPIC_RTE_LOW(rte->rte_index), low32);
869
870         /* Remove the rte entry from the list */
871         list_del(&rte->rte_list);
872         iosapic_intr_info[vector].count--;
873         iosapic_free_rte(rte);
874         index = find_iosapic(gsi);
875         iosapic_lists[index].rtes_inuse--;
876         WARN_ON(iosapic_lists[index].rtes_inuse < 0);
877
878         trigger  = iosapic_intr_info[vector].trigger;
879         polarity = iosapic_intr_info[vector].polarity;
880         dest     = iosapic_intr_info[vector].dest;
881         printk(KERN_INFO
882                "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d unregistered\n",
883                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
884                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
885                cpu_logical_id(dest), dest, vector);
886
887         if (list_empty(&iosapic_intr_info[vector].rtes)) {
888                 /* Sanity check */
889                 BUG_ON(iosapic_intr_info[vector].count);
890
891                 /* Clear the interrupt controller descriptor */
892                 idesc->chip = &no_irq_type;
893
894 #ifdef CONFIG_SMP
895                 /* Clear affinity */
896                 cpus_setall(idesc->affinity);
897 #endif
898
899                 /* Clear the interrupt information */
900                 memset(&iosapic_intr_info[vector], 0,
901                        sizeof(struct iosapic_intr_info));
902                 iosapic_intr_info[vector].low32 |= IOSAPIC_MASK;
903                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
904
905                 if (idesc->action) {
906                         printk(KERN_ERR
907                                "interrupt handlers still exist on IRQ %u\n",
908                                irq);
909                         WARN_ON(1);
910                 }
911
912                 /* Free the interrupt vector */
913                 free_irq_vector(vector);
914         }
915  out:
916         spin_unlock(&iosapic_lock);
917         spin_unlock_irqrestore(&idesc->lock, flags);
918 }
919
920 /*
921  * ACPI calls this when it finds an entry for a platform interrupt.
922  */
923 int __init
924 iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
925                                 int iosapic_vector, u16 eid, u16 id,
926                                 unsigned long polarity, unsigned long trigger)
927 {
928         static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
929         unsigned char delivery;
930         int vector, mask = 0;
931         unsigned int dest = ((id << 8) | eid) & 0xffff;
932
933         switch (int_type) {
934               case ACPI_INTERRUPT_PMI:
935                 vector = iosapic_vector;
936                 /*
937                  * since PMI vector is alloc'd by FW(ACPI) not by kernel,
938                  * we need to make sure the vector is available
939                  */
940                 iosapic_reassign_vector(vector);
941                 delivery = IOSAPIC_PMI;
942                 break;
943               case ACPI_INTERRUPT_INIT:
944                 vector = assign_irq_vector(AUTO_ASSIGN);
945                 if (vector < 0)
946                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
947                 delivery = IOSAPIC_INIT;
948                 break;
949               case ACPI_INTERRUPT_CPEI:
950                 vector = IA64_CPE_VECTOR;
951                 delivery = IOSAPIC_LOWEST_PRIORITY;
952                 mask = 1;
953                 break;
954               default:
955                 printk(KERN_ERR "%s: invalid int type 0x%x\n", __FUNCTION__,
956                        int_type);
957                 return -1;
958         }
959
960         register_intr(gsi, vector, delivery, polarity, trigger);
961
962         printk(KERN_INFO
963                "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x)"
964                " vector %d\n",
965                int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
966                int_type, 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, mask);
971         return vector;
972 }
973
974 /*
975  * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
976  */
977 void __devinit
978 iosapic_override_isa_irq (unsigned int isa_irq, unsigned int gsi,
979                           unsigned long polarity,
980                           unsigned long trigger)
981 {
982         int vector;
983         unsigned int dest = cpu_physical_id(smp_processor_id());
984
985         vector = isa_irq_to_vector(isa_irq);
986
987         register_intr(gsi, vector, IOSAPIC_LOWEST_PRIORITY, polarity, trigger);
988
989         DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
990             isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
991             polarity == IOSAPIC_POL_HIGH ? "high" : "low",
992             cpu_logical_id(dest), dest, vector);
993
994         set_rte(gsi, vector, dest, 1);
995 }
996
997 void __init
998 iosapic_system_init (int system_pcat_compat)
999 {
1000         int vector;
1001
1002         for (vector = 0; vector < IA64_NUM_VECTORS; ++vector) {
1003                 iosapic_intr_info[vector].low32 = IOSAPIC_MASK;
1004                 /* mark as unused */
1005                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
1006         }
1007
1008         pcat_compat = system_pcat_compat;
1009         if (pcat_compat) {
1010                 /*
1011                  * Disable the compatibility mode interrupts (8259 style),
1012                  * needs IN/OUT support enabled.
1013                  */
1014                 printk(KERN_INFO
1015                        "%s: Disabling PC-AT compatible 8259 interrupts\n",
1016                        __FUNCTION__);
1017                 outb(0xff, 0xA1);
1018                 outb(0xff, 0x21);
1019         }
1020 }
1021
1022 static inline int
1023 iosapic_alloc (void)
1024 {
1025         int index;
1026
1027         for (index = 0; index < NR_IOSAPICS; index++)
1028                 if (!iosapic_lists[index].addr)
1029                         return index;
1030
1031         printk(KERN_WARNING "%s: failed to allocate iosapic\n", __FUNCTION__);
1032         return -1;
1033 }
1034
1035 static inline void
1036 iosapic_free (int index)
1037 {
1038         memset(&iosapic_lists[index], 0, sizeof(iosapic_lists[0]));
1039 }
1040
1041 static inline int
1042 iosapic_check_gsi_range (unsigned int gsi_base, unsigned int ver)
1043 {
1044         int index;
1045         unsigned int gsi_end, base, end;
1046
1047         /* check gsi range */
1048         gsi_end = gsi_base + ((ver >> 16) & 0xff);
1049         for (index = 0; index < NR_IOSAPICS; index++) {
1050                 if (!iosapic_lists[index].addr)
1051                         continue;
1052
1053                 base = iosapic_lists[index].gsi_base;
1054                 end  = base + iosapic_lists[index].num_rte - 1;
1055
1056                 if (gsi_end < base || end < gsi_base)
1057                         continue; /* OK */
1058
1059                 return -EBUSY;
1060         }
1061         return 0;
1062 }
1063
1064 int __devinit
1065 iosapic_init (unsigned long phys_addr, unsigned int gsi_base)
1066 {
1067         int num_rte, err, index;
1068         unsigned int isa_irq, ver;
1069         char __iomem *addr;
1070         unsigned long flags;
1071
1072         spin_lock_irqsave(&iosapic_lock, flags);
1073         addr = ioremap(phys_addr, 0);
1074         ver = iosapic_version(addr);
1075
1076         if ((err = iosapic_check_gsi_range(gsi_base, ver))) {
1077                 iounmap(addr);
1078                 spin_unlock_irqrestore(&iosapic_lock, flags);
1079                 return err;
1080         }
1081
1082         /*
1083          * The MAX_REDIR register holds the highest input pin number
1084          * (starting from 0).  We add 1 so that we can use it for
1085          * number of pins (= RTEs)
1086          */
1087         num_rte = ((ver >> 16) & 0xff) + 1;
1088
1089         index = iosapic_alloc();
1090         iosapic_lists[index].addr = addr;
1091         iosapic_lists[index].gsi_base = gsi_base;
1092         iosapic_lists[index].num_rte = num_rte;
1093 #ifdef CONFIG_NUMA
1094         iosapic_lists[index].node = MAX_NUMNODES;
1095 #endif
1096         spin_unlock_irqrestore(&iosapic_lock, flags);
1097
1098         if ((gsi_base == 0) && pcat_compat) {
1099                 /*
1100                  * Map the legacy ISA devices into the IOSAPIC data.  Some of
1101                  * these may get reprogrammed later on with data from the ACPI
1102                  * Interrupt Source Override table.
1103                  */
1104                 for (isa_irq = 0; isa_irq < 16; ++isa_irq)
1105                         iosapic_override_isa_irq(isa_irq, isa_irq,
1106                                                  IOSAPIC_POL_HIGH,
1107                                                  IOSAPIC_EDGE);
1108         }
1109         return 0;
1110 }
1111
1112 #ifdef CONFIG_HOTPLUG
1113 int
1114 iosapic_remove (unsigned int gsi_base)
1115 {
1116         int index, err = 0;
1117         unsigned long flags;
1118
1119         spin_lock_irqsave(&iosapic_lock, flags);
1120         index = find_iosapic(gsi_base);
1121         if (index < 0) {
1122                 printk(KERN_WARNING "%s: No IOSAPIC for GSI base %u\n",
1123                        __FUNCTION__, gsi_base);
1124                 goto out;
1125         }
1126
1127         if (iosapic_lists[index].rtes_inuse) {
1128                 err = -EBUSY;
1129                 printk(KERN_WARNING "%s: IOSAPIC for GSI base %u is busy\n",
1130                        __FUNCTION__, gsi_base);
1131                 goto out;
1132         }
1133
1134         iounmap(iosapic_lists[index].addr);
1135         iosapic_free(index);
1136  out:
1137         spin_unlock_irqrestore(&iosapic_lock, flags);
1138         return err;
1139 }
1140 #endif /* CONFIG_HOTPLUG */
1141
1142 #ifdef CONFIG_NUMA
1143 void __devinit
1144 map_iosapic_to_node(unsigned int gsi_base, int node)
1145 {
1146         int index;
1147
1148         index = find_iosapic(gsi_base);
1149         if (index < 0) {
1150                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1151                        __FUNCTION__, gsi_base);
1152                 return;
1153         }
1154         iosapic_lists[index].node = node;
1155         return;
1156 }
1157 #endif
1158
1159 static int __init iosapic_enable_kmalloc (void)
1160 {
1161         iosapic_kmalloc_ok = 1;
1162         return 0;
1163 }
1164 core_initcall (iosapic_enable_kmalloc);