[PATCH] Update cfq io scheduler to time sliced design
[linux-2.6] / arch / ia64 / kernel / iosapic.c
1 /*
2  * I/O SAPIC support.
3  *
4  * Copyright (C) 1999 Intel Corp.
5  * Copyright (C) 1999 Asit Mallick <asit.k.mallick@intel.com>
6  * Copyright (C) 2000-2002 J.I. Lee <jung-ik.lee@intel.com>
7  * Copyright (C) 1999-2000, 2002-2003 Hewlett-Packard Co.
8  *      David Mosberger-Tang <davidm@hpl.hp.com>
9  * Copyright (C) 1999 VA Linux Systems
10  * Copyright (C) 1999,2000 Walt Drummond <drummond@valinux.com>
11  *
12  * 00/04/19     D. Mosberger    Rewritten to mirror more closely the x86 I/O 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 #ifdef CONFIG_NUMA
133         unsigned short  node;           /* numa node association via pxm */
134 #endif
135 } iosapic_lists[NR_IOSAPICS];
136
137 static int num_iosapic;
138
139 static unsigned char pcat_compat __initdata;    /* 8259 compatibility flag */
140
141 static int iosapic_kmalloc_ok;
142 static LIST_HEAD(free_rte_list);
143
144 /*
145  * Find an IOSAPIC associated with a GSI
146  */
147 static inline int
148 find_iosapic (unsigned int gsi)
149 {
150         int i;
151
152         for (i = 0; i < num_iosapic; i++) {
153                 if ((unsigned) (gsi - iosapic_lists[i].gsi_base) < iosapic_lists[i].num_rte)
154                         return i;
155         }
156
157         return -1;
158 }
159
160 static inline int
161 _gsi_to_vector (unsigned int gsi)
162 {
163         struct iosapic_intr_info *info;
164         struct iosapic_rte_info *rte;
165
166         for (info = iosapic_intr_info; info < iosapic_intr_info + IA64_NUM_VECTORS; ++info)
167                 list_for_each_entry(rte, &info->rtes, rte_list)
168                         if (rte->gsi_base + rte->rte_index == gsi)
169                                 return info - iosapic_intr_info;
170         return -1;
171 }
172
173 /*
174  * Translate GSI number to the corresponding IA-64 interrupt vector.  If no
175  * entry exists, return -1.
176  */
177 inline int
178 gsi_to_vector (unsigned int gsi)
179 {
180         return _gsi_to_vector(gsi);
181 }
182
183 int
184 gsi_to_irq (unsigned int gsi)
185 {
186         unsigned long flags;
187         int irq;
188         /*
189          * XXX fix me: this assumes an identity mapping vetween IA-64 vector and Linux irq
190          * numbers...
191          */
192         spin_lock_irqsave(&iosapic_lock, flags);
193         {
194                 irq = _gsi_to_vector(gsi);
195         }
196         spin_unlock_irqrestore(&iosapic_lock, flags);
197
198         return irq;
199 }
200
201 static struct iosapic_rte_info *gsi_vector_to_rte(unsigned int gsi, unsigned int vec)
202 {
203         struct iosapic_rte_info *rte;
204
205         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list)
206                 if (rte->gsi_base + rte->rte_index == gsi)
207                         return rte;
208         return NULL;
209 }
210
211 static void
212 set_rte (unsigned int gsi, unsigned int vector, unsigned int dest, int mask)
213 {
214         unsigned long pol, trigger, dmode;
215         u32 low32, high32;
216         char __iomem *addr;
217         int rte_index;
218         char redir;
219         struct iosapic_rte_info *rte;
220
221         DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
222
223         rte = gsi_vector_to_rte(gsi, vector);
224         if (!rte)
225                 return;         /* not an IOSAPIC interrupt */
226
227         rte_index = rte->rte_index;
228         addr    = rte->addr;
229         pol     = iosapic_intr_info[vector].polarity;
230         trigger = iosapic_intr_info[vector].trigger;
231         dmode   = iosapic_intr_info[vector].dmode;
232
233         redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
234
235 #ifdef CONFIG_SMP
236         {
237                 unsigned int irq;
238
239                 for (irq = 0; irq < NR_IRQS; ++irq)
240                         if (irq_to_vector(irq) == vector) {
241                                 set_irq_affinity_info(irq, (int)(dest & 0xffff), redir);
242                                 break;
243                         }
244         }
245 #endif
246
247         low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
248                  (trigger << IOSAPIC_TRIGGER_SHIFT) |
249                  (dmode << IOSAPIC_DELIVERY_SHIFT) |
250                  ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
251                  vector);
252
253         /* dest contains both id and eid */
254         high32 = (dest << IOSAPIC_DEST_SHIFT);
255
256         iosapic_write(addr, IOSAPIC_RTE_HIGH(rte_index), high32);
257         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
258         iosapic_intr_info[vector].low32 = low32;
259         iosapic_intr_info[vector].dest = dest;
260 }
261
262 static void
263 nop (unsigned int vector)
264 {
265         /* do nothing... */
266 }
267
268 static void
269 mask_irq (unsigned int irq)
270 {
271         unsigned long flags;
272         char __iomem *addr;
273         u32 low32;
274         int rte_index;
275         ia64_vector vec = irq_to_vector(irq);
276         struct iosapic_rte_info *rte;
277
278         if (list_empty(&iosapic_intr_info[vec].rtes))
279                 return;                 /* not an IOSAPIC interrupt! */
280
281         spin_lock_irqsave(&iosapic_lock, flags);
282         {
283                 /* set only the mask bit */
284                 low32 = iosapic_intr_info[vec].low32 |= IOSAPIC_MASK;
285                 list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list) {
286                         addr = rte->addr;
287                         rte_index = rte->rte_index;
288                         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
289                 }
290         }
291         spin_unlock_irqrestore(&iosapic_lock, flags);
292 }
293
294 static void
295 unmask_irq (unsigned int irq)
296 {
297         unsigned long flags;
298         char __iomem *addr;
299         u32 low32;
300         int rte_index;
301         ia64_vector vec = irq_to_vector(irq);
302         struct iosapic_rte_info *rte;
303
304         if (list_empty(&iosapic_intr_info[vec].rtes))
305                 return;                 /* not an IOSAPIC interrupt! */
306
307         spin_lock_irqsave(&iosapic_lock, flags);
308         {
309                 low32 = iosapic_intr_info[vec].low32 &= ~IOSAPIC_MASK;
310                 list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list) {
311                         addr = rte->addr;
312                         rte_index = rte->rte_index;
313                         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
314                 }
315         }
316         spin_unlock_irqrestore(&iosapic_lock, flags);
317 }
318
319
320 static void
321 iosapic_set_affinity (unsigned int irq, cpumask_t mask)
322 {
323 #ifdef CONFIG_SMP
324         unsigned long flags;
325         u32 high32, low32;
326         int dest, rte_index;
327         char __iomem *addr;
328         int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
329         ia64_vector vec;
330         struct iosapic_rte_info *rte;
331
332         irq &= (~IA64_IRQ_REDIRECTED);
333         vec = irq_to_vector(irq);
334
335         if (cpus_empty(mask))
336                 return;
337
338         dest = cpu_physical_id(first_cpu(mask));
339
340         if (list_empty(&iosapic_intr_info[vec].rtes))
341                 return;                 /* not an IOSAPIC interrupt */
342
343         set_irq_affinity_info(irq, dest, redir);
344
345         /* dest contains both id and eid */
346         high32 = dest << IOSAPIC_DEST_SHIFT;
347
348         spin_lock_irqsave(&iosapic_lock, flags);
349         {
350                 low32 = iosapic_intr_info[vec].low32 & ~(7 << IOSAPIC_DELIVERY_SHIFT);
351
352                 if (redir)
353                         /* change delivery mode to lowest priority */
354                         low32 |= (IOSAPIC_LOWEST_PRIORITY << IOSAPIC_DELIVERY_SHIFT);
355                 else
356                         /* change delivery mode to fixed */
357                         low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
358
359                 iosapic_intr_info[vec].low32 = low32;
360                 iosapic_intr_info[vec].dest = dest;
361                 list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list) {
362                         addr = rte->addr;
363                         rte_index = rte->rte_index;
364                         iosapic_write(addr, IOSAPIC_RTE_HIGH(rte_index), high32);
365                         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
366                 }
367         }
368         spin_unlock_irqrestore(&iosapic_lock, flags);
369 #endif
370 }
371
372 /*
373  * Handlers for level-triggered interrupts.
374  */
375
376 static unsigned int
377 iosapic_startup_level_irq (unsigned int irq)
378 {
379         unmask_irq(irq);
380         return 0;
381 }
382
383 static void
384 iosapic_end_level_irq (unsigned int irq)
385 {
386         ia64_vector vec = irq_to_vector(irq);
387         struct iosapic_rte_info *rte;
388
389         move_irq(irq);
390         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list)
391                 iosapic_eoi(rte->addr, vec);
392 }
393
394 #define iosapic_shutdown_level_irq      mask_irq
395 #define iosapic_enable_level_irq        unmask_irq
396 #define iosapic_disable_level_irq       mask_irq
397 #define iosapic_ack_level_irq           nop
398
399 struct hw_interrupt_type irq_type_iosapic_level = {
400         .typename =     "IO-SAPIC-level",
401         .startup =      iosapic_startup_level_irq,
402         .shutdown =     iosapic_shutdown_level_irq,
403         .enable =       iosapic_enable_level_irq,
404         .disable =      iosapic_disable_level_irq,
405         .ack =          iosapic_ack_level_irq,
406         .end =          iosapic_end_level_irq,
407         .set_affinity = iosapic_set_affinity
408 };
409
410 /*
411  * Handlers for edge-triggered interrupts.
412  */
413
414 static unsigned int
415 iosapic_startup_edge_irq (unsigned int irq)
416 {
417         unmask_irq(irq);
418         /*
419          * IOSAPIC simply drops interrupts pended while the
420          * corresponding pin was masked, so we can't know if an
421          * interrupt is pending already.  Let's hope not...
422          */
423         return 0;
424 }
425
426 static void
427 iosapic_ack_edge_irq (unsigned int irq)
428 {
429         irq_desc_t *idesc = irq_descp(irq);
430
431         move_irq(irq);
432         /*
433          * Once we have recorded IRQ_PENDING already, we can mask the
434          * interrupt for real. This prevents IRQ storms from unhandled
435          * devices.
436          */
437         if ((idesc->status & (IRQ_PENDING|IRQ_DISABLED)) == (IRQ_PENDING|IRQ_DISABLED))
438                 mask_irq(irq);
439 }
440
441 #define iosapic_enable_edge_irq         unmask_irq
442 #define iosapic_disable_edge_irq        nop
443 #define iosapic_end_edge_irq            nop
444
445 struct hw_interrupt_type irq_type_iosapic_edge = {
446         .typename =     "IO-SAPIC-edge",
447         .startup =      iosapic_startup_edge_irq,
448         .shutdown =     iosapic_disable_edge_irq,
449         .enable =       iosapic_enable_edge_irq,
450         .disable =      iosapic_disable_edge_irq,
451         .ack =          iosapic_ack_edge_irq,
452         .end =          iosapic_end_edge_irq,
453         .set_affinity = iosapic_set_affinity
454 };
455
456 unsigned int
457 iosapic_version (char __iomem *addr)
458 {
459         /*
460          * IOSAPIC Version Register return 32 bit structure like:
461          * {
462          *      unsigned int version   : 8;
463          *      unsigned int reserved1 : 8;
464          *      unsigned int max_redir : 8;
465          *      unsigned int reserved2 : 8;
466          * }
467          */
468         return iosapic_read(addr, IOSAPIC_VERSION);
469 }
470
471 static int iosapic_find_sharable_vector (unsigned long trigger, unsigned long pol)
472 {
473         int i, vector = -1, min_count = -1;
474         struct iosapic_intr_info *info;
475
476         /*
477          * shared vectors for edge-triggered interrupts are not
478          * supported yet
479          */
480         if (trigger == IOSAPIC_EDGE)
481                 return -1;
482
483         for (i = IA64_FIRST_DEVICE_VECTOR; i <= IA64_LAST_DEVICE_VECTOR; i++) {
484                 info = &iosapic_intr_info[i];
485                 if (info->trigger == trigger && info->polarity == pol &&
486                     (info->dmode == IOSAPIC_FIXED || info->dmode == IOSAPIC_LOWEST_PRIORITY)) {
487                         if (min_count == -1 || info->count < min_count) {
488                                 vector = i;
489                                 min_count = info->count;
490                         }
491                 }
492         }
493         if (vector < 0)
494                 panic("%s: out of interrupt vectors!\n", __FUNCTION__);
495
496         return vector;
497 }
498
499 /*
500  * if the given vector is already owned by other,
501  *  assign a new vector for the other and make the vector available
502  */
503 static void __init
504 iosapic_reassign_vector (int vector)
505 {
506         int new_vector;
507
508         if (!list_empty(&iosapic_intr_info[vector].rtes)) {
509                 new_vector = assign_irq_vector(AUTO_ASSIGN);
510                 printk(KERN_INFO "Reassigning vector %d to %d\n", vector, new_vector);
511                 memcpy(&iosapic_intr_info[new_vector], &iosapic_intr_info[vector],
512                        sizeof(struct iosapic_intr_info));
513                 INIT_LIST_HEAD(&iosapic_intr_info[new_vector].rtes);
514                 list_move(iosapic_intr_info[vector].rtes.next, &iosapic_intr_info[new_vector].rtes);
515                 memset(&iosapic_intr_info[vector], 0, sizeof(struct iosapic_intr_info));
516                 iosapic_intr_info[vector].low32 = IOSAPIC_MASK;
517                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
518         }
519 }
520
521 static struct iosapic_rte_info *iosapic_alloc_rte (void)
522 {
523         int i;
524         struct iosapic_rte_info *rte;
525         int preallocated = 0;
526
527         if (!iosapic_kmalloc_ok && list_empty(&free_rte_list)) {
528                 rte = alloc_bootmem(sizeof(struct iosapic_rte_info) * NR_PREALLOCATE_RTE_ENTRIES);
529                 if (!rte)
530                         return NULL;
531                 for (i = 0; i < NR_PREALLOCATE_RTE_ENTRIES; i++, rte++)
532                         list_add(&rte->rte_list, &free_rte_list);
533         }
534
535         if (!list_empty(&free_rte_list)) {
536                 rte = list_entry(free_rte_list.next, struct iosapic_rte_info, rte_list);
537                 list_del(&rte->rte_list);
538                 preallocated++;
539         } else {
540                 rte = kmalloc(sizeof(struct iosapic_rte_info), GFP_ATOMIC);
541                 if (!rte)
542                         return NULL;
543         }
544
545         memset(rte, 0, sizeof(struct iosapic_rte_info));
546         if (preallocated)
547                 rte->flags |= RTE_PREALLOCATED;
548
549         return rte;
550 }
551
552 static void iosapic_free_rte (struct iosapic_rte_info *rte)
553 {
554         if (rte->flags & RTE_PREALLOCATED)
555                 list_add_tail(&rte->rte_list, &free_rte_list);
556         else
557                 kfree(rte);
558 }
559
560 static inline int vector_is_shared (int vector)
561 {
562         return (iosapic_intr_info[vector].count > 1);
563 }
564
565 static void
566 register_intr (unsigned int gsi, int vector, unsigned char delivery,
567                unsigned long polarity, unsigned long trigger)
568 {
569         irq_desc_t *idesc;
570         struct hw_interrupt_type *irq_type;
571         int rte_index;
572         int index;
573         unsigned long gsi_base;
574         void __iomem *iosapic_address;
575         struct iosapic_rte_info *rte;
576
577         index = find_iosapic(gsi);
578         if (index < 0) {
579                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n", __FUNCTION__, gsi);
580                 return;
581         }
582
583         iosapic_address = iosapic_lists[index].addr;
584         gsi_base = iosapic_lists[index].gsi_base;
585
586         rte = gsi_vector_to_rte(gsi, vector);
587         if (!rte) {
588                 rte = iosapic_alloc_rte();
589                 if (!rte) {
590                         printk(KERN_WARNING "%s: cannot allocate memory\n", __FUNCTION__);
591                         return;
592                 }
593
594                 rte_index = gsi - gsi_base;
595                 rte->rte_index  = rte_index;
596                 rte->addr       = iosapic_address;
597                 rte->gsi_base   = gsi_base;
598                 rte->refcnt++;
599                 list_add_tail(&rte->rte_list, &iosapic_intr_info[vector].rtes);
600                 iosapic_intr_info[vector].count++;
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;
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 }
627
628 static unsigned int
629 get_target_cpu (unsigned int gsi, int vector)
630 {
631 #ifdef CONFIG_SMP
632         static int cpu = -1;
633
634         /*
635          * In case of vector shared by multiple RTEs, all RTEs that
636          * share the vector need to use the same destination CPU.
637          */
638         if (!list_empty(&iosapic_intr_info[vector].rtes))
639                 return iosapic_intr_info[vector].dest;
640
641         /*
642          * If the platform supports redirection via XTP, let it
643          * distribute interrupts.
644          */
645         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
646                 return cpu_physical_id(smp_processor_id());
647
648         /*
649          * Some interrupts (ACPI SCI, for instance) are registered
650          * before the BSP is marked as online.
651          */
652         if (!cpu_online(smp_processor_id()))
653                 return cpu_physical_id(smp_processor_id());
654
655 #ifdef CONFIG_NUMA
656         {
657                 int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
658                 cpumask_t cpu_mask;
659
660                 iosapic_index = find_iosapic(gsi);
661                 if (iosapic_index < 0 ||
662                     iosapic_lists[iosapic_index].node == MAX_NUMNODES)
663                         goto skip_numa_setup;
664
665                 cpu_mask = node_to_cpumask(iosapic_lists[iosapic_index].node);
666
667                 for_each_cpu_mask(numa_cpu, cpu_mask) {
668                         if (!cpu_online(numa_cpu))
669                                 cpu_clear(numa_cpu, cpu_mask);
670                 }
671
672                 num_cpus = cpus_weight(cpu_mask);
673
674                 if (!num_cpus)
675                         goto skip_numa_setup;
676
677                 /* Use vector assigment to distribute across cpus in node */
678                 cpu_index = vector % num_cpus;
679
680                 for (numa_cpu = first_cpu(cpu_mask) ; i < cpu_index ; i++)
681                         numa_cpu = next_cpu(numa_cpu, cpu_mask);
682
683                 if (numa_cpu != NR_CPUS)
684                         return cpu_physical_id(numa_cpu);
685         }
686 skip_numa_setup:
687 #endif
688         /*
689          * Otherwise, round-robin interrupt vectors across all the
690          * processors.  (It'd be nice if we could be smarter in the
691          * case of NUMA.)
692          */
693         do {
694                 if (++cpu >= NR_CPUS)
695                         cpu = 0;
696         } while (!cpu_online(cpu));
697
698         return cpu_physical_id(cpu);
699 #else
700         return cpu_physical_id(smp_processor_id());
701 #endif
702 }
703
704 /*
705  * ACPI can describe IOSAPIC interrupts via static tables and namespace
706  * methods.  This provides an interface to register those interrupts and
707  * program the IOSAPIC RTE.
708  */
709 int
710 iosapic_register_intr (unsigned int gsi,
711                        unsigned long polarity, unsigned long trigger)
712 {
713         int vector, mask = 1;
714         unsigned int dest;
715         unsigned long flags;
716         struct iosapic_rte_info *rte;
717         u32 low32;
718 again:
719         /*
720          * If this GSI has already been registered (i.e., it's a
721          * shared interrupt, or we lost a race to register it),
722          * don't touch the RTE.
723          */
724         spin_lock_irqsave(&iosapic_lock, flags);
725         {
726                 vector = gsi_to_vector(gsi);
727                 if (vector > 0) {
728                         rte = gsi_vector_to_rte(gsi, vector);
729                         rte->refcnt++;
730                         spin_unlock_irqrestore(&iosapic_lock, flags);
731                         return vector;
732                 }
733         }
734         spin_unlock_irqrestore(&iosapic_lock, flags);
735
736         /* If vector is running out, we try to find a sharable vector */
737         vector = assign_irq_vector_nopanic(AUTO_ASSIGN);
738         if (vector < 0)
739                 vector = iosapic_find_sharable_vector(trigger, polarity);
740
741         spin_lock_irqsave(&irq_descp(vector)->lock, flags);
742         spin_lock(&iosapic_lock);
743         {
744                 if (gsi_to_vector(gsi) > 0) {
745                         if (list_empty(&iosapic_intr_info[vector].rtes))
746                                 free_irq_vector(vector);
747                         spin_unlock(&iosapic_lock);
748                         spin_unlock_irqrestore(&irq_descp(vector)->lock, flags);
749                         goto again;
750                 }
751
752                 dest = get_target_cpu(gsi, vector);
753                 register_intr(gsi, vector, IOSAPIC_LOWEST_PRIORITY,
754                               polarity, trigger);
755
756                 /*
757                  * If the vector is shared and already unmasked for
758                  * other interrupt sources, don't mask it.
759                  */
760                 low32 = iosapic_intr_info[vector].low32;
761                 if (vector_is_shared(vector) && !(low32 & IOSAPIC_MASK))
762                         mask = 0;
763                 set_rte(gsi, vector, dest, mask);
764         }
765         spin_unlock(&iosapic_lock);
766         spin_unlock_irqrestore(&irq_descp(vector)->lock, flags);
767
768         printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
769                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
770                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
771                cpu_logical_id(dest), dest, vector);
772
773         return vector;
774 }
775
776 #ifdef CONFIG_ACPI_DEALLOCATE_IRQ
777 void
778 iosapic_unregister_intr (unsigned int gsi)
779 {
780         unsigned long flags;
781         int irq, vector;
782         irq_desc_t *idesc;
783         u32 low32;
784         unsigned long trigger, polarity;
785         unsigned int dest;
786         struct iosapic_rte_info *rte;
787
788         /*
789          * If the irq associated with the gsi is not found,
790          * iosapic_unregister_intr() is unbalanced. We need to check
791          * this again after getting locks.
792          */
793         irq = gsi_to_irq(gsi);
794         if (irq < 0) {
795                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n", gsi);
796                 WARN_ON(1);
797                 return;
798         }
799         vector = irq_to_vector(irq);
800
801         idesc = irq_descp(irq);
802         spin_lock_irqsave(&idesc->lock, flags);
803         spin_lock(&iosapic_lock);
804         {
805                 if ((rte = gsi_vector_to_rte(gsi, vector)) == NULL) {
806                         printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n", gsi);
807                         WARN_ON(1);
808                         goto out;
809                 }
810
811                 if (--rte->refcnt > 0)
812                         goto out;
813
814                 /* Mask the interrupt */
815                 low32 = iosapic_intr_info[vector].low32 | IOSAPIC_MASK;
816                 iosapic_write(rte->addr, IOSAPIC_RTE_LOW(rte->rte_index), low32);
817
818                 /* Remove the rte entry from the list */
819                 list_del(&rte->rte_list);
820                 iosapic_intr_info[vector].count--;
821                 iosapic_free_rte(rte);
822
823                 trigger  = iosapic_intr_info[vector].trigger;
824                 polarity = iosapic_intr_info[vector].polarity;
825                 dest     = iosapic_intr_info[vector].dest;
826                 printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d unregistered\n",
827                        gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
828                        (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
829                        cpu_logical_id(dest), dest, vector);
830
831                 if (list_empty(&iosapic_intr_info[vector].rtes)) {
832                         /* Sanity check */
833                         BUG_ON(iosapic_intr_info[vector].count);
834
835                         /* Clear the interrupt controller descriptor */
836                         idesc->handler = &no_irq_type;
837
838                         /* Clear the interrupt information */
839                         memset(&iosapic_intr_info[vector], 0, sizeof(struct iosapic_intr_info));
840                         iosapic_intr_info[vector].low32 |= IOSAPIC_MASK;
841                         INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
842
843                         if (idesc->action) {
844                                 printk(KERN_ERR "interrupt handlers still exist on IRQ %u\n", irq);
845                                 WARN_ON(1);
846                         }
847
848                         /* Free the interrupt vector */
849                         free_irq_vector(vector);
850                 }
851         }
852  out:
853         spin_unlock(&iosapic_lock);
854         spin_unlock_irqrestore(&idesc->lock, flags);
855 }
856 #endif /* CONFIG_ACPI_DEALLOCATE_IRQ */
857
858 /*
859  * ACPI calls this when it finds an entry for a platform interrupt.
860  * Note that the irq_base and IOSAPIC address must be set in iosapic_init().
861  */
862 int __init
863 iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
864                                 int iosapic_vector, u16 eid, u16 id,
865                                 unsigned long polarity, unsigned long trigger)
866 {
867         static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
868         unsigned char delivery;
869         int vector, mask = 0;
870         unsigned int dest = ((id << 8) | eid) & 0xffff;
871
872         switch (int_type) {
873               case ACPI_INTERRUPT_PMI:
874                 vector = iosapic_vector;
875                 /*
876                  * since PMI vector is alloc'd by FW(ACPI) not by kernel,
877                  * we need to make sure the vector is available
878                  */
879                 iosapic_reassign_vector(vector);
880                 delivery = IOSAPIC_PMI;
881                 break;
882               case ACPI_INTERRUPT_INIT:
883                 vector = assign_irq_vector(AUTO_ASSIGN);
884                 delivery = IOSAPIC_INIT;
885                 break;
886               case ACPI_INTERRUPT_CPEI:
887                 vector = IA64_CPE_VECTOR;
888                 delivery = IOSAPIC_LOWEST_PRIORITY;
889                 mask = 1;
890                 break;
891               default:
892                 printk(KERN_ERR "iosapic_register_platform_irq(): invalid int type 0x%x\n", int_type);
893                 return -1;
894         }
895
896         register_intr(gsi, vector, delivery, polarity, trigger);
897
898         printk(KERN_INFO "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
899                int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
900                int_type, gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
901                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
902                cpu_logical_id(dest), dest, vector);
903
904         set_rte(gsi, vector, dest, mask);
905         return vector;
906 }
907
908
909 /*
910  * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
911  * Note that the gsi_base and IOSAPIC address must be set in iosapic_init().
912  */
913 void __init
914 iosapic_override_isa_irq (unsigned int isa_irq, unsigned int gsi,
915                           unsigned long polarity,
916                           unsigned long trigger)
917 {
918         int vector;
919         unsigned int dest = cpu_physical_id(smp_processor_id());
920
921         vector = isa_irq_to_vector(isa_irq);
922
923         register_intr(gsi, vector, IOSAPIC_LOWEST_PRIORITY, polarity, trigger);
924
925         DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
926             isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
927             polarity == IOSAPIC_POL_HIGH ? "high" : "low",
928             cpu_logical_id(dest), dest, vector);
929
930         set_rte(gsi, vector, dest, 1);
931 }
932
933 void __init
934 iosapic_system_init (int system_pcat_compat)
935 {
936         int vector;
937
938         for (vector = 0; vector < IA64_NUM_VECTORS; ++vector) {
939                 iosapic_intr_info[vector].low32 = IOSAPIC_MASK;
940                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);        /* mark as unused */
941         }
942
943         pcat_compat = system_pcat_compat;
944         if (pcat_compat) {
945                 /*
946                  * Disable the compatibility mode interrupts (8259 style), needs IN/OUT support
947                  * enabled.
948                  */
949                 printk(KERN_INFO "%s: Disabling PC-AT compatible 8259 interrupts\n", __FUNCTION__);
950                 outb(0xff, 0xA1);
951                 outb(0xff, 0x21);
952         }
953 }
954
955 void __init
956 iosapic_init (unsigned long phys_addr, unsigned int gsi_base)
957 {
958         int num_rte;
959         unsigned int isa_irq, ver;
960         char __iomem *addr;
961
962         addr = ioremap(phys_addr, 0);
963         ver = iosapic_version(addr);
964
965         /*
966          * The MAX_REDIR register holds the highest input pin
967          * number (starting from 0).
968          * We add 1 so that we can use it for number of pins (= RTEs)
969          */
970         num_rte = ((ver >> 16) & 0xff) + 1;
971
972         iosapic_lists[num_iosapic].addr = addr;
973         iosapic_lists[num_iosapic].gsi_base = gsi_base;
974         iosapic_lists[num_iosapic].num_rte = num_rte;
975 #ifdef CONFIG_NUMA
976         iosapic_lists[num_iosapic].node = MAX_NUMNODES;
977 #endif
978         num_iosapic++;
979
980         if ((gsi_base == 0) && pcat_compat) {
981                 /*
982                  * Map the legacy ISA devices into the IOSAPIC data.  Some of these may
983                  * get reprogrammed later on with data from the ACPI Interrupt Source
984                  * Override table.
985                  */
986                 for (isa_irq = 0; isa_irq < 16; ++isa_irq)
987                         iosapic_override_isa_irq(isa_irq, isa_irq, IOSAPIC_POL_HIGH, IOSAPIC_EDGE);
988         }
989 }
990
991 #ifdef CONFIG_NUMA
992 void __init
993 map_iosapic_to_node(unsigned int gsi_base, int node)
994 {
995         int index;
996
997         index = find_iosapic(gsi_base);
998         if (index < 0) {
999                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1000                        __FUNCTION__, gsi_base);
1001                 return;
1002         }
1003         iosapic_lists[index].node = node;
1004         return;
1005 }
1006 #endif
1007
1008 static int __init iosapic_enable_kmalloc (void)
1009 {
1010         iosapic_kmalloc_ok = 1;
1011         return 0;
1012 }
1013 core_initcall (iosapic_enable_kmalloc);