fedora core 6 1.2949 + vserver 2.2.0
[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
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/smp_lock.h>
91 #include <linux/string.h>
92 #include <linux/bootmem.h>
93
94 #include <asm/delay.h>
95 #include <asm/hw_irq.h>
96 #include <asm/io.h>
97 #include <asm/iosapic.h>
98 #include <asm/machvec.h>
99 #include <asm/processor.h>
100 #include <asm/ptrace.h>
101 #include <asm/system.h>
102
103 #undef DEBUG_INTERRUPT_ROUTING
104
105 #ifdef DEBUG_INTERRUPT_ROUTING
106 #define DBG(fmt...)     printk(fmt)
107 #else
108 #define DBG(fmt...)
109 #endif
110
111 #define NR_PREALLOCATE_RTE_ENTRIES \
112         (PAGE_SIZE / sizeof(struct iosapic_rte_info))
113 #define RTE_PREALLOCATED        (1)
114
115 static DEFINE_SPINLOCK(iosapic_lock);
116
117 /*
118  * These tables map IA-64 vectors to the IOSAPIC pin that generates this
119  * vector.
120  */
121
122 struct iosapic_rte_info {
123         struct list_head rte_list;      /* node in list of RTEs sharing the
124                                          * same vector */
125         char __iomem    *addr;          /* base address of IOSAPIC */
126         unsigned int    gsi_base;       /* first GSI assigned to this
127                                          * IOSAPIC */
128         char            rte_index;      /* IOSAPIC RTE index */
129         int             refcnt;         /* reference counter */
130         unsigned int    flags;          /* flags */
131 } ____cacheline_aligned;
132
133 static struct iosapic_intr_info {
134         struct list_head rtes;          /* RTEs using this vector (empty =>
135                                          * not an IOSAPIC interrupt) */
136         int             count;          /* # of RTEs that shares this vector */
137         u32             low32;          /* current value of low word of
138                                          * Redirection table entry */
139         unsigned int    dest;           /* destination CPU physical ID */
140         unsigned char   dmode   : 3;    /* delivery mode (see iosapic.h) */
141         unsigned char   polarity: 1;    /* interrupt polarity
142                                          * (see iosapic.h) */
143         unsigned char   trigger : 1;    /* trigger mode (see iosapic.h) */
144 } iosapic_intr_info[IA64_NUM_VECTORS];
145
146 static struct iosapic {
147         char __iomem    *addr;          /* base address of IOSAPIC */
148         unsigned int    gsi_base;       /* first GSI assigned to this
149                                          * IOSAPIC */
150         unsigned short  num_rte;        /* # of RTEs on this IOSAPIC */
151         int             rtes_inuse;     /* # of RTEs in use on this IOSAPIC */
152 #ifdef CONFIG_NUMA
153         unsigned short  node;           /* numa node association via pxm */
154 #endif
155 } iosapic_lists[NR_IOSAPICS];
156
157 static unsigned char pcat_compat __devinitdata; /* 8259 compatibility flag */
158
159 static int iosapic_kmalloc_ok;
160 static LIST_HEAD(free_rte_list);
161
162 #ifdef CONFIG_XEN
163 #include <xen/interface/xen.h>
164 #include <xen/interface/physdev.h>
165 #include <asm/hypervisor.h>
166 static inline unsigned int xen_iosapic_read(char __iomem *iosapic, unsigned int reg)
167 {
168         struct physdev_apic apic_op;
169         int ret;
170
171         apic_op.apic_physbase = (unsigned long)iosapic -
172                                         __IA64_UNCACHED_OFFSET;
173         apic_op.reg = reg;
174         ret = HYPERVISOR_physdev_op(PHYSDEVOP_apic_read, &apic_op);
175         if (ret)
176                 return ret;
177         return apic_op.value;
178 }
179
180 static inline void xen_iosapic_write(char __iomem *iosapic, unsigned int reg, u32 val)
181 {
182         struct physdev_apic apic_op;
183
184         apic_op.apic_physbase = (unsigned long)iosapic - 
185                                         __IA64_UNCACHED_OFFSET;
186         apic_op.reg = reg;
187         apic_op.value = val;
188         HYPERVISOR_physdev_op(PHYSDEVOP_apic_write, &apic_op);
189 }
190
191 static inline unsigned int iosapic_read(char __iomem *iosapic, unsigned int reg)
192 {
193         if (!is_running_on_xen()) {
194                 writel(reg, iosapic + IOSAPIC_REG_SELECT);
195                 return readl(iosapic + IOSAPIC_WINDOW);
196         } else
197                 return xen_iosapic_read(iosapic, reg);
198 }
199
200 static inline void iosapic_write(char __iomem *iosapic, unsigned int reg, u32 val)
201 {
202         if (!is_running_on_xen()) {
203                 writel(reg, iosapic + IOSAPIC_REG_SELECT);
204                 writel(val, iosapic + IOSAPIC_WINDOW);
205         } else
206                 xen_iosapic_write(iosapic, reg, val);
207 }
208
209 int xen_assign_irq_vector(int irq)
210 {
211         struct physdev_irq irq_op;
212
213         irq_op.irq = irq;
214         if (HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op))
215                 return -ENOSPC;
216
217         return irq_op.vector;
218 }
219 #endif /* XEN */
220
221 /*
222  * Find an IOSAPIC associated with a GSI
223  */
224 static inline int
225 find_iosapic (unsigned int gsi)
226 {
227         int i;
228
229         for (i = 0; i < NR_IOSAPICS; i++) {
230                 if ((unsigned) (gsi - iosapic_lists[i].gsi_base) <
231                     iosapic_lists[i].num_rte)
232                         return i;
233         }
234
235         return -1;
236 }
237
238 static inline int
239 _gsi_to_vector (unsigned int gsi)
240 {
241         struct iosapic_intr_info *info;
242         struct iosapic_rte_info *rte;
243
244         for (info = iosapic_intr_info; info <
245                      iosapic_intr_info + IA64_NUM_VECTORS; ++info)
246                 list_for_each_entry(rte, &info->rtes, rte_list)
247                         if (rte->gsi_base + rte->rte_index == gsi)
248                                 return info - iosapic_intr_info;
249         return -1;
250 }
251
252 /*
253  * Translate GSI number to the corresponding IA-64 interrupt vector.  If no
254  * entry exists, return -1.
255  */
256 inline int
257 gsi_to_vector (unsigned int gsi)
258 {
259         return _gsi_to_vector(gsi);
260 }
261
262 int
263 gsi_to_irq (unsigned int gsi)
264 {
265         unsigned long flags;
266         int irq;
267         /*
268          * XXX fix me: this assumes an identity mapping between IA-64 vector
269          * and Linux irq numbers...
270          */
271         spin_lock_irqsave(&iosapic_lock, flags);
272         {
273                 irq = _gsi_to_vector(gsi);
274         }
275         spin_unlock_irqrestore(&iosapic_lock, flags);
276
277         return irq;
278 }
279
280 static struct iosapic_rte_info *gsi_vector_to_rte(unsigned int gsi,
281                                                   unsigned int vec)
282 {
283         struct iosapic_rte_info *rte;
284
285         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list)
286                 if (rte->gsi_base + rte->rte_index == gsi)
287                         return rte;
288         return NULL;
289 }
290
291 static void
292 set_rte (unsigned int gsi, unsigned int vector, unsigned int dest, int mask)
293 {
294         unsigned long pol, trigger, dmode;
295         u32 low32, high32;
296         char __iomem *addr;
297         int rte_index;
298         char redir;
299         struct iosapic_rte_info *rte;
300
301         DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
302
303         rte = gsi_vector_to_rte(gsi, vector);
304         if (!rte)
305                 return;         /* not an IOSAPIC interrupt */
306
307         rte_index = rte->rte_index;
308         addr    = rte->addr;
309         pol     = iosapic_intr_info[vector].polarity;
310         trigger = iosapic_intr_info[vector].trigger;
311         dmode   = iosapic_intr_info[vector].dmode;
312
313         redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
314
315 #ifdef CONFIG_SMP
316         {
317                 unsigned int irq;
318
319                 for (irq = 0; irq < NR_IRQS; ++irq)
320                         if (irq_to_vector(irq) == vector) {
321                                 set_irq_affinity_info(irq,
322                                                       (int)(dest & 0xffff),
323                                                       redir);
324                                 break;
325                         }
326         }
327 #endif
328
329         low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
330                  (trigger << IOSAPIC_TRIGGER_SHIFT) |
331                  (dmode << IOSAPIC_DELIVERY_SHIFT) |
332                  ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
333                  vector);
334
335         /* dest contains both id and eid */
336         high32 = (dest << IOSAPIC_DEST_SHIFT);
337
338         iosapic_write(addr, IOSAPIC_RTE_HIGH(rte_index), high32);
339         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
340         iosapic_intr_info[vector].low32 = low32;
341         iosapic_intr_info[vector].dest = dest;
342 }
343
344 static void
345 nop (unsigned int irq)
346 {
347         /* do nothing... */
348 }
349
350
351 #ifdef CONFIG_KEXEC
352 void
353 kexec_disable_iosapic(void)
354 {
355         struct iosapic_intr_info *info;
356         struct iosapic_rte_info *rte;
357         u8 vec = 0;
358         for (info = iosapic_intr_info; info <
359                         iosapic_intr_info + IA64_NUM_VECTORS; ++info, ++vec) {
360                 list_for_each_entry(rte, &info->rtes,
361                                 rte_list) {
362                         iosapic_write(rte->addr,
363                                         IOSAPIC_RTE_LOW(rte->rte_index),
364                                         IOSAPIC_MASK|vec);
365                         iosapic_eoi(rte->addr, vec);
366                 }
367         }
368 }
369 #endif
370
371 static void
372 mask_irq (unsigned int irq)
373 {
374         unsigned long flags;
375         char __iomem *addr;
376         u32 low32;
377         int rte_index;
378         ia64_vector vec = irq_to_vector(irq);
379         struct iosapic_rte_info *rte;
380
381         if (list_empty(&iosapic_intr_info[vec].rtes))
382                 return;                 /* not an IOSAPIC interrupt! */
383
384         spin_lock_irqsave(&iosapic_lock, flags);
385         {
386                 /* set only the mask bit */
387                 low32 = iosapic_intr_info[vec].low32 |= IOSAPIC_MASK;
388                 list_for_each_entry(rte, &iosapic_intr_info[vec].rtes,
389                                     rte_list) {
390                         addr = rte->addr;
391                         rte_index = rte->rte_index;
392                         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
393                 }
394         }
395         spin_unlock_irqrestore(&iosapic_lock, flags);
396 }
397
398 static void
399 unmask_irq (unsigned int irq)
400 {
401         unsigned long flags;
402         char __iomem *addr;
403         u32 low32;
404         int rte_index;
405         ia64_vector vec = irq_to_vector(irq);
406         struct iosapic_rte_info *rte;
407
408         if (list_empty(&iosapic_intr_info[vec].rtes))
409                 return;                 /* not an IOSAPIC interrupt! */
410
411         spin_lock_irqsave(&iosapic_lock, flags);
412         {
413                 low32 = iosapic_intr_info[vec].low32 &= ~IOSAPIC_MASK;
414                 list_for_each_entry(rte, &iosapic_intr_info[vec].rtes,
415                                     rte_list) {
416                         addr = rte->addr;
417                         rte_index = rte->rte_index;
418                         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
419                 }
420         }
421         spin_unlock_irqrestore(&iosapic_lock, flags);
422 }
423
424
425 static void
426 iosapic_set_affinity (unsigned int irq, cpumask_t mask)
427 {
428 #ifdef CONFIG_SMP
429         unsigned long flags;
430         u32 high32, low32;
431         int dest, rte_index;
432         char __iomem *addr;
433         int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
434         ia64_vector vec;
435         struct iosapic_rte_info *rte;
436
437         irq &= (~IA64_IRQ_REDIRECTED);
438         vec = irq_to_vector(irq);
439
440         if (cpus_empty(mask))
441                 return;
442
443         dest = cpu_physical_id(first_cpu(mask));
444
445         if (list_empty(&iosapic_intr_info[vec].rtes))
446                 return;                 /* not an IOSAPIC interrupt */
447
448         set_irq_affinity_info(irq, dest, redir);
449
450         /* dest contains both id and eid */
451         high32 = dest << IOSAPIC_DEST_SHIFT;
452
453         spin_lock_irqsave(&iosapic_lock, flags);
454         {
455                 low32 = iosapic_intr_info[vec].low32 &
456                         ~(7 << IOSAPIC_DELIVERY_SHIFT);
457
458                 if (redir)
459                         /* change delivery mode to lowest priority */
460                         low32 |= (IOSAPIC_LOWEST_PRIORITY <<
461                                   IOSAPIC_DELIVERY_SHIFT);
462                 else
463                         /* change delivery mode to fixed */
464                         low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
465
466                 iosapic_intr_info[vec].low32 = low32;
467                 iosapic_intr_info[vec].dest = dest;
468                 list_for_each_entry(rte, &iosapic_intr_info[vec].rtes,
469                                     rte_list) {
470                         addr = rte->addr;
471                         rte_index = rte->rte_index;
472                         iosapic_write(addr, IOSAPIC_RTE_HIGH(rte_index),
473                                       high32);
474                         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
475                 }
476         }
477         spin_unlock_irqrestore(&iosapic_lock, flags);
478 #endif
479 }
480
481 /*
482  * Handlers for level-triggered interrupts.
483  */
484
485 static unsigned int
486 iosapic_startup_level_irq (unsigned int irq)
487 {
488         unmask_irq(irq);
489         return 0;
490 }
491
492 static void
493 iosapic_end_level_irq (unsigned int irq)
494 {
495         ia64_vector vec = irq_to_vector(irq);
496         struct iosapic_rte_info *rte;
497
498         move_native_irq(irq);
499         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list)
500                 iosapic_eoi(rte->addr, vec);
501 }
502
503 #define iosapic_shutdown_level_irq      mask_irq
504 #define iosapic_enable_level_irq        unmask_irq
505 #define iosapic_disable_level_irq       mask_irq
506 #define iosapic_ack_level_irq           nop
507
508 struct irq_chip irq_type_iosapic_level = {
509         .name =         "IO-SAPIC-level",
510         .startup =      iosapic_startup_level_irq,
511         .shutdown =     iosapic_shutdown_level_irq,
512         .enable =       iosapic_enable_level_irq,
513         .disable =      iosapic_disable_level_irq,
514         .ack =          iosapic_ack_level_irq,
515         .end =          iosapic_end_level_irq,
516         .mask =         mask_irq,
517         .unmask =       unmask_irq,
518         .set_affinity = iosapic_set_affinity
519 };
520
521 /*
522  * Handlers for edge-triggered interrupts.
523  */
524
525 static unsigned int
526 iosapic_startup_edge_irq (unsigned int irq)
527 {
528         unmask_irq(irq);
529         /*
530          * IOSAPIC simply drops interrupts pended while the
531          * corresponding pin was masked, so we can't know if an
532          * interrupt is pending already.  Let's hope not...
533          */
534         return 0;
535 }
536
537 static void
538 iosapic_ack_edge_irq (unsigned int irq)
539 {
540         irq_desc_t *idesc = irq_desc + irq;
541
542         move_native_irq(irq);
543         /*
544          * Once we have recorded IRQ_PENDING already, we can mask the
545          * interrupt for real. This prevents IRQ storms from unhandled
546          * devices.
547          */
548         if ((idesc->status & (IRQ_PENDING|IRQ_DISABLED)) ==
549             (IRQ_PENDING|IRQ_DISABLED))
550                 mask_irq(irq);
551 }
552
553 #define iosapic_enable_edge_irq         unmask_irq
554 #define iosapic_disable_edge_irq        nop
555 #define iosapic_end_edge_irq            nop
556
557 struct irq_chip irq_type_iosapic_edge = {
558         .name =         "IO-SAPIC-edge",
559         .startup =      iosapic_startup_edge_irq,
560         .shutdown =     iosapic_disable_edge_irq,
561         .enable =       iosapic_enable_edge_irq,
562         .disable =      iosapic_disable_edge_irq,
563         .ack =          iosapic_ack_edge_irq,
564         .end =          iosapic_end_edge_irq,
565         .mask =         mask_irq,
566         .unmask =       unmask_irq,
567         .set_affinity = iosapic_set_affinity
568 };
569
570 unsigned int
571 iosapic_version (char __iomem *addr)
572 {
573         /*
574          * IOSAPIC Version Register return 32 bit structure like:
575          * {
576          *      unsigned int version   : 8;
577          *      unsigned int reserved1 : 8;
578          *      unsigned int max_redir : 8;
579          *      unsigned int reserved2 : 8;
580          * }
581          */
582         return iosapic_read(addr, IOSAPIC_VERSION);
583 }
584
585 static int iosapic_find_sharable_vector (unsigned long trigger,
586                                          unsigned long pol)
587 {
588         int i, vector = -1, min_count = -1;
589         struct iosapic_intr_info *info;
590
591         /*
592          * shared vectors for edge-triggered interrupts are not
593          * supported yet
594          */
595         if (trigger == IOSAPIC_EDGE)
596                 return -1;
597
598         for (i = IA64_FIRST_DEVICE_VECTOR; i <= IA64_LAST_DEVICE_VECTOR; i++) {
599                 info = &iosapic_intr_info[i];
600                 if (info->trigger == trigger && info->polarity == pol &&
601                     (info->dmode == IOSAPIC_FIXED || info->dmode ==
602                      IOSAPIC_LOWEST_PRIORITY)) {
603                         if (min_count == -1 || info->count < min_count) {
604                                 vector = i;
605                                 min_count = info->count;
606                         }
607                 }
608         }
609
610         return vector;
611 }
612
613 /*
614  * if the given vector is already owned by other,
615  *  assign a new vector for the other and make the vector available
616  */
617 static void __init
618 iosapic_reassign_vector (int vector)
619 {
620         int new_vector;
621
622         if (!list_empty(&iosapic_intr_info[vector].rtes)) {
623                 new_vector = assign_irq_vector(AUTO_ASSIGN);
624                 if (new_vector < 0)
625                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
626                 printk(KERN_INFO "Reassigning vector %d to %d\n",
627                        vector, new_vector);
628                 memcpy(&iosapic_intr_info[new_vector], &iosapic_intr_info[vector],
629                        sizeof(struct iosapic_intr_info));
630                 INIT_LIST_HEAD(&iosapic_intr_info[new_vector].rtes);
631                 list_move(iosapic_intr_info[vector].rtes.next,
632                           &iosapic_intr_info[new_vector].rtes);
633                 memset(&iosapic_intr_info[vector], 0,
634                        sizeof(struct iosapic_intr_info));
635                 iosapic_intr_info[vector].low32 = IOSAPIC_MASK;
636                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
637         }
638 }
639
640 static struct iosapic_rte_info *iosapic_alloc_rte (void)
641 {
642         int i;
643         struct iosapic_rte_info *rte;
644         int preallocated = 0;
645
646         if (!iosapic_kmalloc_ok && list_empty(&free_rte_list)) {
647                 rte = alloc_bootmem(sizeof(struct iosapic_rte_info) *
648                                     NR_PREALLOCATE_RTE_ENTRIES);
649                 if (!rte)
650                         return NULL;
651                 for (i = 0; i < NR_PREALLOCATE_RTE_ENTRIES; i++, rte++)
652                         list_add(&rte->rte_list, &free_rte_list);
653         }
654
655         if (!list_empty(&free_rte_list)) {
656                 rte = list_entry(free_rte_list.next, struct iosapic_rte_info,
657                                  rte_list);
658                 list_del(&rte->rte_list);
659                 preallocated++;
660         } else {
661                 rte = kmalloc(sizeof(struct iosapic_rte_info), GFP_ATOMIC);
662                 if (!rte)
663                         return NULL;
664         }
665
666         memset(rte, 0, sizeof(struct iosapic_rte_info));
667         if (preallocated)
668                 rte->flags |= RTE_PREALLOCATED;
669
670         return rte;
671 }
672
673 static void iosapic_free_rte (struct iosapic_rte_info *rte)
674 {
675         if (rte->flags & RTE_PREALLOCATED)
676                 list_add_tail(&rte->rte_list, &free_rte_list);
677         else
678                 kfree(rte);
679 }
680
681 static inline int vector_is_shared (int vector)
682 {
683         return (iosapic_intr_info[vector].count > 1);
684 }
685
686 static int
687 register_intr (unsigned int gsi, int vector, unsigned char delivery,
688                unsigned long polarity, unsigned long trigger)
689 {
690         irq_desc_t *idesc;
691         struct hw_interrupt_type *irq_type;
692         int rte_index;
693         int index;
694         unsigned long gsi_base;
695         void __iomem *iosapic_address;
696         struct iosapic_rte_info *rte;
697
698         index = find_iosapic(gsi);
699         if (index < 0) {
700                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
701                        __FUNCTION__, gsi);
702                 return -ENODEV;
703         }
704
705         iosapic_address = iosapic_lists[index].addr;
706         gsi_base = iosapic_lists[index].gsi_base;
707
708         rte = gsi_vector_to_rte(gsi, vector);
709         if (!rte) {
710                 rte = iosapic_alloc_rte();
711                 if (!rte) {
712                         printk(KERN_WARNING "%s: cannot allocate memory\n",
713                                __FUNCTION__);
714                         return -ENOMEM;
715                 }
716
717                 rte_index = gsi - gsi_base;
718                 rte->rte_index  = rte_index;
719                 rte->addr       = iosapic_address;
720                 rte->gsi_base   = gsi_base;
721                 rte->refcnt++;
722                 list_add_tail(&rte->rte_list, &iosapic_intr_info[vector].rtes);
723                 iosapic_intr_info[vector].count++;
724                 iosapic_lists[index].rtes_inuse++;
725         }
726         else if (vector_is_shared(vector)) {
727                 struct iosapic_intr_info *info = &iosapic_intr_info[vector];
728                 if (info->trigger != trigger || info->polarity != polarity) {
729                         printk (KERN_WARNING
730                                 "%s: cannot override the interrupt\n",
731                                 __FUNCTION__);
732                         return -EINVAL;
733                 }
734         }
735
736         iosapic_intr_info[vector].polarity = polarity;
737         iosapic_intr_info[vector].dmode    = delivery;
738         iosapic_intr_info[vector].trigger  = trigger;
739
740         if (is_running_on_xen())
741                 return 0;
742
743         if (trigger == IOSAPIC_EDGE)
744                 irq_type = &irq_type_iosapic_edge;
745         else
746                 irq_type = &irq_type_iosapic_level;
747
748         idesc = irq_desc + vector;
749         if (idesc->chip != irq_type) {
750                 if (idesc->chip != &no_irq_type)
751                         printk(KERN_WARNING
752                                "%s: changing vector %d from %s to %s\n",
753                                __FUNCTION__, vector,
754                                idesc->chip->name, irq_type->name);
755                 idesc->chip = irq_type;
756         }
757         return 0;
758 }
759
760 static unsigned int
761 get_target_cpu (unsigned int gsi, int vector)
762 {
763 #ifdef CONFIG_SMP
764         static int cpu = -1;
765         extern int cpe_vector;
766
767         /*
768          * In case of vector shared by multiple RTEs, all RTEs that
769          * share the vector need to use the same destination CPU.
770          */
771         if (!list_empty(&iosapic_intr_info[vector].rtes))
772                 return iosapic_intr_info[vector].dest;
773
774         /*
775          * If the platform supports redirection via XTP, let it
776          * distribute interrupts.
777          */
778         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
779                 return cpu_physical_id(smp_processor_id());
780
781         /*
782          * Some interrupts (ACPI SCI, for instance) are registered
783          * before the BSP is marked as online.
784          */
785         if (!cpu_online(smp_processor_id()))
786                 return cpu_physical_id(smp_processor_id());
787
788 #ifdef CONFIG_ACPI
789         if (cpe_vector > 0 && vector == IA64_CPEP_VECTOR)
790                 return get_cpei_target_cpu();
791 #endif
792
793 #ifdef CONFIG_NUMA
794         {
795                 int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
796                 cpumask_t cpu_mask;
797
798                 iosapic_index = find_iosapic(gsi);
799                 if (iosapic_index < 0 ||
800                     iosapic_lists[iosapic_index].node == MAX_NUMNODES)
801                         goto skip_numa_setup;
802
803                 cpu_mask = node_to_cpumask(iosapic_lists[iosapic_index].node);
804
805                 for_each_cpu_mask(numa_cpu, cpu_mask) {
806                         if (!cpu_online(numa_cpu))
807                                 cpu_clear(numa_cpu, cpu_mask);
808                 }
809
810                 num_cpus = cpus_weight(cpu_mask);
811
812                 if (!num_cpus)
813                         goto skip_numa_setup;
814
815                 /* Use vector assignment to distribute across cpus in node */
816                 cpu_index = vector % num_cpus;
817
818                 for (numa_cpu = first_cpu(cpu_mask) ; i < cpu_index ; i++)
819                         numa_cpu = next_cpu(numa_cpu, cpu_mask);
820
821                 if (numa_cpu != NR_CPUS)
822                         return cpu_physical_id(numa_cpu);
823         }
824 skip_numa_setup:
825 #endif
826         /*
827          * Otherwise, round-robin interrupt vectors across all the
828          * processors.  (It'd be nice if we could be smarter in the
829          * case of NUMA.)
830          */
831         do {
832                 if (++cpu >= NR_CPUS)
833                         cpu = 0;
834         } while (!cpu_online(cpu));
835
836         return cpu_physical_id(cpu);
837 #else  /* CONFIG_SMP */
838         return cpu_physical_id(smp_processor_id());
839 #endif
840 }
841
842 /*
843  * ACPI can describe IOSAPIC interrupts via static tables and namespace
844  * methods.  This provides an interface to register those interrupts and
845  * program the IOSAPIC RTE.
846  */
847 int
848 iosapic_register_intr (unsigned int gsi,
849                        unsigned long polarity, unsigned long trigger)
850 {
851         int vector, mask = 1, err;
852         unsigned int dest;
853         unsigned long flags;
854         struct iosapic_rte_info *rte;
855         u32 low32;
856 again:
857         /*
858          * If this GSI has already been registered (i.e., it's a
859          * shared interrupt, or we lost a race to register it),
860          * don't touch the RTE.
861          */
862         spin_lock_irqsave(&iosapic_lock, flags);
863         {
864                 vector = gsi_to_vector(gsi);
865                 if (vector > 0) {
866                         rte = gsi_vector_to_rte(gsi, vector);
867                         rte->refcnt++;
868                         spin_unlock_irqrestore(&iosapic_lock, flags);
869                         return vector;
870                 }
871         }
872         spin_unlock_irqrestore(&iosapic_lock, flags);
873
874         /* If vector is running out, we try to find a sharable vector */
875         vector = assign_irq_vector(AUTO_ASSIGN);
876         if (vector < 0) {
877                 vector = iosapic_find_sharable_vector(trigger, polarity);
878                 if (vector < 0)
879                         return -ENOSPC;
880         }
881
882         spin_lock_irqsave(&irq_desc[vector].lock, flags);
883         spin_lock(&iosapic_lock);
884         {
885                 if (gsi_to_vector(gsi) > 0) {
886                         if (list_empty(&iosapic_intr_info[vector].rtes))
887                                 free_irq_vector(vector);
888                         spin_unlock(&iosapic_lock);
889                         spin_unlock_irqrestore(&irq_desc[vector].lock,
890                                                flags);
891                         goto again;
892                 }
893
894                 dest = get_target_cpu(gsi, vector);
895                 err = register_intr(gsi, vector, IOSAPIC_LOWEST_PRIORITY,
896                               polarity, trigger);
897                 if (err < 0) {
898                         spin_unlock(&iosapic_lock);
899                         spin_unlock_irqrestore(&irq_desc[vector].lock,
900                                                flags);
901                         return err;
902                 }
903
904                 /*
905                  * If the vector is shared and already unmasked for
906                  * other interrupt sources, don't mask it.
907                  */
908                 low32 = iosapic_intr_info[vector].low32;
909                 if (vector_is_shared(vector) && !(low32 & IOSAPIC_MASK))
910                         mask = 0;
911                 set_rte(gsi, vector, dest, mask);
912         }
913         spin_unlock(&iosapic_lock);
914         spin_unlock_irqrestore(&irq_desc[vector].lock, flags);
915
916         printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
917                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
918                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
919                cpu_logical_id(dest), dest, vector);
920
921         return vector;
922 }
923
924 void
925 iosapic_unregister_intr (unsigned int gsi)
926 {
927         unsigned long flags;
928         int irq, vector, index;
929         irq_desc_t *idesc;
930         u32 low32;
931         unsigned long trigger, polarity;
932         unsigned int dest;
933         struct iosapic_rte_info *rte;
934
935         /*
936          * If the irq associated with the gsi is not found,
937          * iosapic_unregister_intr() is unbalanced. We need to check
938          * this again after getting locks.
939          */
940         irq = gsi_to_irq(gsi);
941         if (irq < 0) {
942                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
943                        gsi);
944                 WARN_ON(1);
945                 return;
946         }
947         vector = irq_to_vector(irq);
948
949         idesc = irq_desc + irq;
950         spin_lock_irqsave(&idesc->lock, flags);
951         spin_lock(&iosapic_lock);
952         {
953                 if ((rte = gsi_vector_to_rte(gsi, vector)) == NULL) {
954                         printk(KERN_ERR
955                                "iosapic_unregister_intr(%u) unbalanced\n",
956                                gsi);
957                         WARN_ON(1);
958                         goto out;
959                 }
960
961                 if (--rte->refcnt > 0)
962                         goto out;
963
964                 /* Mask the interrupt */
965                 low32 = iosapic_intr_info[vector].low32 | IOSAPIC_MASK;
966                 iosapic_write(rte->addr, IOSAPIC_RTE_LOW(rte->rte_index),
967                               low32);
968
969                 /* Remove the rte entry from the list */
970                 list_del(&rte->rte_list);
971                 iosapic_intr_info[vector].count--;
972                 iosapic_free_rte(rte);
973                 index = find_iosapic(gsi);
974                 iosapic_lists[index].rtes_inuse--;
975                 WARN_ON(iosapic_lists[index].rtes_inuse < 0);
976
977                 trigger  = iosapic_intr_info[vector].trigger;
978                 polarity = iosapic_intr_info[vector].polarity;
979                 dest     = iosapic_intr_info[vector].dest;
980                 printk(KERN_INFO
981                        "GSI %u (%s, %s) -> CPU %d (0x%04x)"
982                        " vector %d unregistered\n",
983                        gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
984                        (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
985                        cpu_logical_id(dest), dest, vector);
986
987                 if (list_empty(&iosapic_intr_info[vector].rtes)) {
988                         /* Sanity check */
989                         BUG_ON(iosapic_intr_info[vector].count);
990
991                         /* Clear the interrupt controller descriptor */
992                         idesc->chip = &no_irq_type;
993
994                         /* Clear the interrupt information */
995                         memset(&iosapic_intr_info[vector], 0,
996                                sizeof(struct iosapic_intr_info));
997                         iosapic_intr_info[vector].low32 |= IOSAPIC_MASK;
998                         INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
999
1000                         if (idesc->action) {
1001                                 printk(KERN_ERR
1002                                        "interrupt handlers still exist on"
1003                                        "IRQ %u\n", irq);
1004                                 WARN_ON(1);
1005                         }
1006
1007                         /* Free the interrupt vector */
1008                         free_irq_vector(vector);
1009                 }
1010         }
1011  out:
1012         spin_unlock(&iosapic_lock);
1013         spin_unlock_irqrestore(&idesc->lock, flags);
1014 }
1015
1016 /*
1017  * ACPI calls this when it finds an entry for a platform interrupt.
1018  */
1019 int __init
1020 iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
1021                                 int iosapic_vector, u16 eid, u16 id,
1022                                 unsigned long polarity, unsigned long trigger)
1023 {
1024         static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
1025         unsigned char delivery;
1026         int vector, mask = 0;
1027         unsigned int dest = ((id << 8) | eid) & 0xffff;
1028
1029         switch (int_type) {
1030               case ACPI_INTERRUPT_PMI:
1031                 vector = iosapic_vector;
1032                 /*
1033                  * since PMI vector is alloc'd by FW(ACPI) not by kernel,
1034                  * we need to make sure the vector is available
1035                  */
1036                 iosapic_reassign_vector(vector);
1037                 delivery = IOSAPIC_PMI;
1038                 break;
1039               case ACPI_INTERRUPT_INIT:
1040                 vector = assign_irq_vector(AUTO_ASSIGN);
1041                 if (vector < 0)
1042                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
1043                 delivery = IOSAPIC_INIT;
1044                 break;
1045               case ACPI_INTERRUPT_CPEI:
1046                 vector = IA64_CPE_VECTOR;
1047                 delivery = IOSAPIC_LOWEST_PRIORITY;
1048                 mask = 1;
1049                 break;
1050               default:
1051                 printk(KERN_ERR "%s: invalid int type 0x%x\n", __FUNCTION__,
1052                        int_type);
1053                 return -1;
1054         }
1055
1056         register_intr(gsi, vector, delivery, polarity, trigger);
1057
1058         printk(KERN_INFO
1059                "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x)"
1060                " vector %d\n",
1061                int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
1062                int_type, gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
1063                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
1064                cpu_logical_id(dest), dest, vector);
1065
1066         set_rte(gsi, vector, dest, mask);
1067         return vector;
1068 }
1069
1070 /*
1071  * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
1072  */
1073 void __init
1074 iosapic_override_isa_irq (unsigned int isa_irq, unsigned int gsi,
1075                           unsigned long polarity,
1076                           unsigned long trigger)
1077 {
1078         int vector;
1079         unsigned int dest = cpu_physical_id(smp_processor_id());
1080
1081         vector = isa_irq_to_vector(isa_irq);
1082
1083         register_intr(gsi, vector, IOSAPIC_LOWEST_PRIORITY, polarity, trigger);
1084
1085         DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
1086             isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
1087             polarity == IOSAPIC_POL_HIGH ? "high" : "low",
1088             cpu_logical_id(dest), dest, vector);
1089
1090         set_rte(gsi, vector, dest, 1);
1091 }
1092
1093 void __init
1094 iosapic_system_init (int system_pcat_compat)
1095 {
1096         int vector;
1097
1098         for (vector = 0; vector < IA64_NUM_VECTORS; ++vector) {
1099                 iosapic_intr_info[vector].low32 = IOSAPIC_MASK;
1100                 /* mark as unused */
1101                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
1102         }
1103
1104         pcat_compat = system_pcat_compat;
1105         if (is_running_on_xen())
1106                 return;
1107
1108         if (pcat_compat) {
1109                 /*
1110                  * Disable the compatibility mode interrupts (8259 style),
1111                  * needs IN/OUT support enabled.
1112                  */
1113                 printk(KERN_INFO
1114                        "%s: Disabling PC-AT compatible 8259 interrupts\n",
1115                        __FUNCTION__);
1116                 outb(0xff, 0xA1);
1117                 outb(0xff, 0x21);
1118         }
1119 }
1120
1121 static inline int
1122 iosapic_alloc (void)
1123 {
1124         int index;
1125
1126         for (index = 0; index < NR_IOSAPICS; index++)
1127                 if (!iosapic_lists[index].addr)
1128                         return index;
1129
1130         printk(KERN_WARNING "%s: failed to allocate iosapic\n", __FUNCTION__);
1131         return -1;
1132 }
1133
1134 static inline void
1135 iosapic_free (int index)
1136 {
1137         memset(&iosapic_lists[index], 0, sizeof(iosapic_lists[0]));
1138 }
1139
1140 static inline int
1141 iosapic_check_gsi_range (unsigned int gsi_base, unsigned int ver)
1142 {
1143         int index;
1144         unsigned int gsi_end, base, end;
1145
1146         /* check gsi range */
1147         gsi_end = gsi_base + ((ver >> 16) & 0xff);
1148         for (index = 0; index < NR_IOSAPICS; index++) {
1149                 if (!iosapic_lists[index].addr)
1150                         continue;
1151
1152                 base = iosapic_lists[index].gsi_base;
1153                 end  = base + iosapic_lists[index].num_rte - 1;
1154
1155                 if (gsi_end < base || end < gsi_base)
1156                         continue; /* OK */
1157
1158                 return -EBUSY;
1159         }
1160         return 0;
1161 }
1162
1163 int __devinit
1164 iosapic_init (unsigned long phys_addr, unsigned int gsi_base)
1165 {
1166         int num_rte, err, index;
1167         unsigned int isa_irq, ver;
1168         char __iomem *addr;
1169         unsigned long flags;
1170
1171         spin_lock_irqsave(&iosapic_lock, flags);
1172         {
1173                 addr = ioremap(phys_addr, 0);
1174                 ver = iosapic_version(addr);
1175
1176                 if ((err = iosapic_check_gsi_range(gsi_base, ver))) {
1177                         iounmap(addr);
1178                         spin_unlock_irqrestore(&iosapic_lock, flags);
1179                         return err;
1180                 }
1181
1182                 /*
1183                  * The MAX_REDIR register holds the highest input pin
1184                  * number (starting from 0).
1185                  * We add 1 so that we can use it for number of pins (= RTEs)
1186                  */
1187                 num_rte = ((ver >> 16) & 0xff) + 1;
1188
1189                 index = iosapic_alloc();
1190                 iosapic_lists[index].addr = addr;
1191                 iosapic_lists[index].gsi_base = gsi_base;
1192                 iosapic_lists[index].num_rte = num_rte;
1193 #ifdef CONFIG_NUMA
1194                 iosapic_lists[index].node = MAX_NUMNODES;
1195 #endif
1196         }
1197         spin_unlock_irqrestore(&iosapic_lock, flags);
1198
1199         if ((gsi_base == 0) && pcat_compat) {
1200                 /*
1201                  * Map the legacy ISA devices into the IOSAPIC data.  Some of
1202                  * these may get reprogrammed later on with data from the ACPI
1203                  * Interrupt Source Override table.
1204                  */
1205                 for (isa_irq = 0; isa_irq < 16; ++isa_irq)
1206                         iosapic_override_isa_irq(isa_irq, isa_irq,
1207                                                  IOSAPIC_POL_HIGH,
1208                                                  IOSAPIC_EDGE);
1209         }
1210         return 0;
1211 }
1212
1213 #ifdef CONFIG_HOTPLUG
1214 int
1215 iosapic_remove (unsigned int gsi_base)
1216 {
1217         int index, err = 0;
1218         unsigned long flags;
1219
1220         spin_lock_irqsave(&iosapic_lock, flags);
1221         {
1222                 index = find_iosapic(gsi_base);
1223                 if (index < 0) {
1224                         printk(KERN_WARNING "%s: No IOSAPIC for GSI base %u\n",
1225                                __FUNCTION__, gsi_base);
1226                         goto out;
1227                 }
1228
1229                 if (iosapic_lists[index].rtes_inuse) {
1230                         err = -EBUSY;
1231                         printk(KERN_WARNING
1232                                "%s: IOSAPIC for GSI base %u is busy\n",
1233                                __FUNCTION__, gsi_base);
1234                         goto out;
1235                 }
1236
1237                 iounmap(iosapic_lists[index].addr);
1238                 iosapic_free(index);
1239         }
1240  out:
1241         spin_unlock_irqrestore(&iosapic_lock, flags);
1242         return err;
1243 }
1244 #endif /* CONFIG_HOTPLUG */
1245
1246 #ifdef CONFIG_NUMA
1247 void __devinit
1248 map_iosapic_to_node(unsigned int gsi_base, int node)
1249 {
1250         int index;
1251
1252         index = find_iosapic(gsi_base);
1253         if (index < 0) {
1254                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1255                        __FUNCTION__, gsi_base);
1256                 return;
1257         }
1258         iosapic_lists[index].node = node;
1259         return;
1260 }
1261 #endif
1262
1263 static int __init iosapic_enable_kmalloc (void)
1264 {
1265         iosapic_kmalloc_ok = 1;
1266         return 0;
1267 }
1268 core_initcall (iosapic_enable_kmalloc);