xen: implement xen_hvm_register_pirq
[pandora-kernel.git] / drivers / xen / events.c
1 /*
2  * Xen event channels
3  *
4  * Xen models interrupts with abstract event channels.  Because each
5  * domain gets 1024 event channels, but NR_IRQ is not that large, we
6  * must dynamically map irqs<->event channels.  The event channels
7  * interface with the rest of the kernel by defining a xen interrupt
8  * chip.  When an event is recieved, it is mapped to an irq and sent
9  * through the normal interrupt processing path.
10  *
11  * There are four kinds of events which can be mapped to an event
12  * channel:
13  *
14  * 1. Inter-domain notifications.  This includes all the virtual
15  *    device events, since they're driven by front-ends in another domain
16  *    (typically dom0).
17  * 2. VIRQs, typically used for timers.  These are per-cpu events.
18  * 3. IPIs.
19  * 4. PIRQs - Hardware interrupts.
20  *
21  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
22  */
23
24 #include <linux/linkage.h>
25 #include <linux/interrupt.h>
26 #include <linux/irq.h>
27 #include <linux/module.h>
28 #include <linux/string.h>
29 #include <linux/bootmem.h>
30 #include <linux/slab.h>
31 #include <linux/irqnr.h>
32
33 #include <asm/desc.h>
34 #include <asm/ptrace.h>
35 #include <asm/irq.h>
36 #include <asm/idle.h>
37 #include <asm/io_apic.h>
38 #include <asm/sync_bitops.h>
39 #include <asm/xen/pci.h>
40 #include <asm/xen/hypercall.h>
41 #include <asm/xen/hypervisor.h>
42
43 #include <xen/xen.h>
44 #include <xen/hvm.h>
45 #include <xen/xen-ops.h>
46 #include <xen/events.h>
47 #include <xen/interface/xen.h>
48 #include <xen/interface/event_channel.h>
49 #include <xen/interface/hvm/hvm_op.h>
50 #include <xen/interface/hvm/params.h>
51
52 /*
53  * This lock protects updates to the following mapping and reference-count
54  * arrays. The lock does not need to be acquired to read the mapping tables.
55  */
56 static DEFINE_SPINLOCK(irq_mapping_update_lock);
57
58 /* IRQ <-> VIRQ mapping. */
59 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
60
61 /* IRQ <-> IPI mapping */
62 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
63
64 /* Interrupt types. */
65 enum xen_irq_type {
66         IRQT_UNBOUND = 0,
67         IRQT_PIRQ,
68         IRQT_VIRQ,
69         IRQT_IPI,
70         IRQT_EVTCHN
71 };
72
73 /*
74  * Packed IRQ information:
75  * type - enum xen_irq_type
76  * event channel - irq->event channel mapping
77  * cpu - cpu this event channel is bound to
78  * index - type-specific information:
79  *    PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
80  *           guest, or GSI (real passthrough IRQ) of the device.
81  *    VIRQ - virq number
82  *    IPI - IPI vector
83  *    EVTCHN -
84  */
85 struct irq_info
86 {
87         enum xen_irq_type type; /* type */
88         unsigned short evtchn;  /* event channel */
89         unsigned short cpu;     /* cpu bound */
90
91         union {
92                 unsigned short virq;
93                 enum ipi_vector ipi;
94                 struct {
95                         unsigned short pirq;
96                         unsigned short gsi;
97                         unsigned char vector;
98                         unsigned char flags;
99                 } pirq;
100         } u;
101 };
102 #define PIRQ_NEEDS_EOI  (1 << 0)
103 #define PIRQ_SHAREABLE  (1 << 1)
104
105 static struct irq_info *irq_info;
106 static int *pirq_to_irq;
107 static int nr_pirqs;
108
109 static int *evtchn_to_irq;
110 struct cpu_evtchn_s {
111         unsigned long bits[NR_EVENT_CHANNELS/BITS_PER_LONG];
112 };
113
114 static __initdata struct cpu_evtchn_s init_evtchn_mask = {
115         .bits[0 ... (NR_EVENT_CHANNELS/BITS_PER_LONG)-1] = ~0ul,
116 };
117 static struct cpu_evtchn_s *cpu_evtchn_mask_p = &init_evtchn_mask;
118
119 static inline unsigned long *cpu_evtchn_mask(int cpu)
120 {
121         return cpu_evtchn_mask_p[cpu].bits;
122 }
123
124 /* Xen will never allocate port zero for any purpose. */
125 #define VALID_EVTCHN(chn)       ((chn) != 0)
126
127 static struct irq_chip xen_dynamic_chip;
128 static struct irq_chip xen_percpu_chip;
129 static struct irq_chip xen_pirq_chip;
130
131 /* Constructor for packed IRQ information. */
132 static struct irq_info mk_unbound_info(void)
133 {
134         return (struct irq_info) { .type = IRQT_UNBOUND };
135 }
136
137 static struct irq_info mk_evtchn_info(unsigned short evtchn)
138 {
139         return (struct irq_info) { .type = IRQT_EVTCHN, .evtchn = evtchn,
140                         .cpu = 0 };
141 }
142
143 static struct irq_info mk_ipi_info(unsigned short evtchn, enum ipi_vector ipi)
144 {
145         return (struct irq_info) { .type = IRQT_IPI, .evtchn = evtchn,
146                         .cpu = 0, .u.ipi = ipi };
147 }
148
149 static struct irq_info mk_virq_info(unsigned short evtchn, unsigned short virq)
150 {
151         return (struct irq_info) { .type = IRQT_VIRQ, .evtchn = evtchn,
152                         .cpu = 0, .u.virq = virq };
153 }
154
155 static struct irq_info mk_pirq_info(unsigned short evtchn, unsigned short pirq,
156                                     unsigned short gsi, unsigned short vector)
157 {
158         return (struct irq_info) { .type = IRQT_PIRQ, .evtchn = evtchn,
159                         .cpu = 0,
160                         .u.pirq = { .pirq = pirq, .gsi = gsi, .vector = vector } };
161 }
162
163 /*
164  * Accessors for packed IRQ information.
165  */
166 static struct irq_info *info_for_irq(unsigned irq)
167 {
168         return &irq_info[irq];
169 }
170
171 static unsigned int evtchn_from_irq(unsigned irq)
172 {
173         return info_for_irq(irq)->evtchn;
174 }
175
176 unsigned irq_from_evtchn(unsigned int evtchn)
177 {
178         return evtchn_to_irq[evtchn];
179 }
180 EXPORT_SYMBOL_GPL(irq_from_evtchn);
181
182 static enum ipi_vector ipi_from_irq(unsigned irq)
183 {
184         struct irq_info *info = info_for_irq(irq);
185
186         BUG_ON(info == NULL);
187         BUG_ON(info->type != IRQT_IPI);
188
189         return info->u.ipi;
190 }
191
192 static unsigned virq_from_irq(unsigned irq)
193 {
194         struct irq_info *info = info_for_irq(irq);
195
196         BUG_ON(info == NULL);
197         BUG_ON(info->type != IRQT_VIRQ);
198
199         return info->u.virq;
200 }
201
202 static unsigned pirq_from_irq(unsigned irq)
203 {
204         struct irq_info *info = info_for_irq(irq);
205
206         BUG_ON(info == NULL);
207         BUG_ON(info->type != IRQT_PIRQ);
208
209         return info->u.pirq.pirq;
210 }
211
212 static unsigned gsi_from_irq(unsigned irq)
213 {
214         struct irq_info *info = info_for_irq(irq);
215
216         BUG_ON(info == NULL);
217         BUG_ON(info->type != IRQT_PIRQ);
218
219         return info->u.pirq.gsi;
220 }
221
222 static unsigned vector_from_irq(unsigned irq)
223 {
224         struct irq_info *info = info_for_irq(irq);
225
226         BUG_ON(info == NULL);
227         BUG_ON(info->type != IRQT_PIRQ);
228
229         return info->u.pirq.vector;
230 }
231
232 static enum xen_irq_type type_from_irq(unsigned irq)
233 {
234         return info_for_irq(irq)->type;
235 }
236
237 static unsigned cpu_from_irq(unsigned irq)
238 {
239         return info_for_irq(irq)->cpu;
240 }
241
242 static unsigned int cpu_from_evtchn(unsigned int evtchn)
243 {
244         int irq = evtchn_to_irq[evtchn];
245         unsigned ret = 0;
246
247         if (irq != -1)
248                 ret = cpu_from_irq(irq);
249
250         return ret;
251 }
252
253 static bool pirq_needs_eoi(unsigned irq)
254 {
255         struct irq_info *info = info_for_irq(irq);
256
257         BUG_ON(info->type != IRQT_PIRQ);
258
259         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
260 }
261
262 static inline unsigned long active_evtchns(unsigned int cpu,
263                                            struct shared_info *sh,
264                                            unsigned int idx)
265 {
266         return (sh->evtchn_pending[idx] &
267                 cpu_evtchn_mask(cpu)[idx] &
268                 ~sh->evtchn_mask[idx]);
269 }
270
271 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
272 {
273         int irq = evtchn_to_irq[chn];
274
275         BUG_ON(irq == -1);
276 #ifdef CONFIG_SMP
277         cpumask_copy(irq_to_desc(irq)->affinity, cpumask_of(cpu));
278 #endif
279
280         __clear_bit(chn, cpu_evtchn_mask(cpu_from_irq(irq)));
281         __set_bit(chn, cpu_evtchn_mask(cpu));
282
283         irq_info[irq].cpu = cpu;
284 }
285
286 static void init_evtchn_cpu_bindings(void)
287 {
288 #ifdef CONFIG_SMP
289         struct irq_desc *desc;
290         int i;
291
292         /* By default all event channels notify CPU#0. */
293         for_each_irq_desc(i, desc) {
294                 cpumask_copy(desc->affinity, cpumask_of(0));
295         }
296 #endif
297
298         memset(cpu_evtchn_mask(0), ~0, sizeof(cpu_evtchn_mask(0)));
299 }
300
301 static inline void clear_evtchn(int port)
302 {
303         struct shared_info *s = HYPERVISOR_shared_info;
304         sync_clear_bit(port, &s->evtchn_pending[0]);
305 }
306
307 static inline void set_evtchn(int port)
308 {
309         struct shared_info *s = HYPERVISOR_shared_info;
310         sync_set_bit(port, &s->evtchn_pending[0]);
311 }
312
313 static inline int test_evtchn(int port)
314 {
315         struct shared_info *s = HYPERVISOR_shared_info;
316         return sync_test_bit(port, &s->evtchn_pending[0]);
317 }
318
319
320 /**
321  * notify_remote_via_irq - send event to remote end of event channel via irq
322  * @irq: irq of event channel to send event to
323  *
324  * Unlike notify_remote_via_evtchn(), this is safe to use across
325  * save/restore. Notifications on a broken connection are silently
326  * dropped.
327  */
328 void notify_remote_via_irq(int irq)
329 {
330         int evtchn = evtchn_from_irq(irq);
331
332         if (VALID_EVTCHN(evtchn))
333                 notify_remote_via_evtchn(evtchn);
334 }
335 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
336
337 static void mask_evtchn(int port)
338 {
339         struct shared_info *s = HYPERVISOR_shared_info;
340         sync_set_bit(port, &s->evtchn_mask[0]);
341 }
342
343 static void unmask_evtchn(int port)
344 {
345         struct shared_info *s = HYPERVISOR_shared_info;
346         unsigned int cpu = get_cpu();
347
348         BUG_ON(!irqs_disabled());
349
350         /* Slow path (hypercall) if this is a non-local port. */
351         if (unlikely(cpu != cpu_from_evtchn(port))) {
352                 struct evtchn_unmask unmask = { .port = port };
353                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
354         } else {
355                 struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
356
357                 sync_clear_bit(port, &s->evtchn_mask[0]);
358
359                 /*
360                  * The following is basically the equivalent of
361                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
362                  * the interrupt edge' if the channel is masked.
363                  */
364                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
365                     !sync_test_and_set_bit(port / BITS_PER_LONG,
366                                            &vcpu_info->evtchn_pending_sel))
367                         vcpu_info->evtchn_upcall_pending = 1;
368         }
369
370         put_cpu();
371 }
372
373 static int get_nr_hw_irqs(void)
374 {
375         int ret = 1;
376
377 #ifdef CONFIG_X86_IO_APIC
378         ret = get_nr_irqs_gsi();
379 #endif
380
381         return ret;
382 }
383
384 /* callers of this function should make sure that PHYSDEVOP_get_nr_pirqs
385  * succeeded otherwise nr_pirqs won't hold the right value */
386 static int find_unbound_pirq(void)
387 {
388         int i;
389         for (i = nr_pirqs-1; i >= 0; i--) {
390                 if (pirq_to_irq[i] < 0)
391                         return i;
392         }
393         return -1;
394 }
395
396 static int find_unbound_irq(void)
397 {
398         struct irq_data *data;
399         int irq, res;
400         int start = get_nr_hw_irqs();
401
402         if (start == nr_irqs)
403                 goto no_irqs;
404
405         /* nr_irqs is a magic value. Must not use it.*/
406         for (irq = nr_irqs-1; irq > start; irq--) {
407                 data = irq_get_irq_data(irq);
408                 /* only 0->15 have init'd desc; handle irq > 16 */
409                 if (!data)
410                         break;
411                 if (data->chip == &no_irq_chip)
412                         break;
413                 if (data->chip != &xen_dynamic_chip)
414                         continue;
415                 if (irq_info[irq].type == IRQT_UNBOUND)
416                         return irq;
417         }
418
419         if (irq == start)
420                 goto no_irqs;
421
422         res = irq_alloc_desc_at(irq, 0);
423
424         if (WARN_ON(res != irq))
425                 return -1;
426
427         return irq;
428
429 no_irqs:
430         panic("No available IRQ to bind to: increase nr_irqs!\n");
431 }
432
433 static bool identity_mapped_irq(unsigned irq)
434 {
435         /* identity map all the hardware irqs */
436         return irq < get_nr_hw_irqs();
437 }
438
439 static void pirq_unmask_notify(int irq)
440 {
441         struct physdev_eoi eoi = { .irq = pirq_from_irq(irq) };
442
443         if (unlikely(pirq_needs_eoi(irq))) {
444                 int rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
445                 WARN_ON(rc);
446         }
447 }
448
449 static void pirq_query_unmask(int irq)
450 {
451         struct physdev_irq_status_query irq_status;
452         struct irq_info *info = info_for_irq(irq);
453
454         BUG_ON(info->type != IRQT_PIRQ);
455
456         irq_status.irq = pirq_from_irq(irq);
457         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
458                 irq_status.flags = 0;
459
460         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
461         if (irq_status.flags & XENIRQSTAT_needs_eoi)
462                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
463 }
464
465 static bool probing_irq(int irq)
466 {
467         struct irq_desc *desc = irq_to_desc(irq);
468
469         return desc && desc->action == NULL;
470 }
471
472 static unsigned int startup_pirq(unsigned int irq)
473 {
474         struct evtchn_bind_pirq bind_pirq;
475         struct irq_info *info = info_for_irq(irq);
476         int evtchn = evtchn_from_irq(irq);
477         int rc;
478
479         BUG_ON(info->type != IRQT_PIRQ);
480
481         if (VALID_EVTCHN(evtchn))
482                 goto out;
483
484         bind_pirq.pirq = pirq_from_irq(irq);
485         /* NB. We are happy to share unless we are probing. */
486         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
487                                         BIND_PIRQ__WILL_SHARE : 0;
488         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
489         if (rc != 0) {
490                 if (!probing_irq(irq))
491                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
492                                irq);
493                 return 0;
494         }
495         evtchn = bind_pirq.port;
496
497         pirq_query_unmask(irq);
498
499         evtchn_to_irq[evtchn] = irq;
500         bind_evtchn_to_cpu(evtchn, 0);
501         info->evtchn = evtchn;
502
503 out:
504         unmask_evtchn(evtchn);
505         pirq_unmask_notify(irq);
506
507         return 0;
508 }
509
510 static void shutdown_pirq(unsigned int irq)
511 {
512         struct evtchn_close close;
513         struct irq_info *info = info_for_irq(irq);
514         int evtchn = evtchn_from_irq(irq);
515
516         BUG_ON(info->type != IRQT_PIRQ);
517
518         if (!VALID_EVTCHN(evtchn))
519                 return;
520
521         mask_evtchn(evtchn);
522
523         close.port = evtchn;
524         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
525                 BUG();
526
527         bind_evtchn_to_cpu(evtchn, 0);
528         evtchn_to_irq[evtchn] = -1;
529         info->evtchn = 0;
530 }
531
532 static void enable_pirq(unsigned int irq)
533 {
534         startup_pirq(irq);
535 }
536
537 static void disable_pirq(unsigned int irq)
538 {
539 }
540
541 static void ack_pirq(unsigned int irq)
542 {
543         int evtchn = evtchn_from_irq(irq);
544
545         move_native_irq(irq);
546
547         if (VALID_EVTCHN(evtchn)) {
548                 mask_evtchn(evtchn);
549                 clear_evtchn(evtchn);
550         }
551 }
552
553 static void end_pirq(unsigned int irq)
554 {
555         int evtchn = evtchn_from_irq(irq);
556         struct irq_desc *desc = irq_to_desc(irq);
557
558         if (WARN_ON(!desc))
559                 return;
560
561         if ((desc->status & (IRQ_DISABLED|IRQ_PENDING)) ==
562             (IRQ_DISABLED|IRQ_PENDING)) {
563                 shutdown_pirq(irq);
564         } else if (VALID_EVTCHN(evtchn)) {
565                 unmask_evtchn(evtchn);
566                 pirq_unmask_notify(irq);
567         }
568 }
569
570 static int find_irq_by_gsi(unsigned gsi)
571 {
572         int irq;
573
574         for (irq = 0; irq < nr_irqs; irq++) {
575                 struct irq_info *info = info_for_irq(irq);
576
577                 if (info == NULL || info->type != IRQT_PIRQ)
578                         continue;
579
580                 if (gsi_from_irq(irq) == gsi)
581                         return irq;
582         }
583
584         return -1;
585 }
586
587 int xen_allocate_pirq(unsigned gsi, int shareable, char *name)
588 {
589         return xen_map_pirq_gsi(gsi, gsi, shareable, name);
590 }
591
592 /* xen_map_pirq_gsi might allocate irqs from the top down, as a
593  * consequence don't assume that the irq number returned has a low value
594  * or can be used as a pirq number unless you know otherwise.
595  *
596  * One notable exception is when xen_map_pirq_gsi is called passing an
597  * hardware gsi as argument, in that case the irq number returned
598  * matches the gsi number passed as second argument.
599  *
600  * Note: We don't assign an event channel until the irq actually started
601  * up.  Return an existing irq if we've already got one for the gsi.
602  */
603 int xen_map_pirq_gsi(unsigned pirq, unsigned gsi, int shareable, char *name)
604 {
605         int irq = 0;
606         struct physdev_irq irq_op;
607
608         spin_lock(&irq_mapping_update_lock);
609
610         if ((pirq > nr_pirqs) || (gsi > nr_irqs)) {
611                 printk(KERN_WARNING "xen_map_pirq_gsi: %s %s is incorrect!\n",
612                         pirq > nr_pirqs ? "nr_pirqs" :"",
613                         gsi > nr_irqs ? "nr_irqs" : "");
614                 goto out;
615         }
616
617         irq = find_irq_by_gsi(gsi);
618         if (irq != -1) {
619                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
620                        irq, gsi);
621                 goto out;       /* XXX need refcount? */
622         }
623
624         /* If we are a PV guest, we don't have GSIs (no ACPI passed). Therefore
625          * we are using the !xen_initial_domain() to drop in the function.*/
626         if (identity_mapped_irq(gsi) || !xen_initial_domain()) {
627                 irq = gsi;
628                 irq_alloc_desc_at(irq, 0);
629         } else
630                 irq = find_unbound_irq();
631
632         set_irq_chip_and_handler_name(irq, &xen_pirq_chip,
633                                       handle_level_irq, name);
634
635         irq_op.irq = irq;
636         irq_op.vector = 0;
637
638         /* Only the privileged domain can do this. For non-priv, the pcifront
639          * driver provides a PCI bus that does the call to do exactly
640          * this in the priv domain. */
641         if (xen_initial_domain() &&
642             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
643                 irq_free_desc(irq);
644                 irq = -ENOSPC;
645                 goto out;
646         }
647
648         irq_info[irq] = mk_pirq_info(0, pirq, gsi, irq_op.vector);
649         irq_info[irq].u.pirq.flags |= shareable ? PIRQ_SHAREABLE : 0;
650         pirq_to_irq[pirq] = irq;
651
652 out:
653         spin_unlock(&irq_mapping_update_lock);
654
655         return irq;
656 }
657
658 int xen_destroy_irq(int irq)
659 {
660         struct irq_desc *desc;
661         int rc = -ENOENT;
662
663         spin_lock(&irq_mapping_update_lock);
664
665         desc = irq_to_desc(irq);
666         if (!desc)
667                 goto out;
668
669         irq_info[irq] = mk_unbound_info();
670
671         irq_free_desc(irq);
672
673 out:
674         spin_unlock(&irq_mapping_update_lock);
675         return rc;
676 }
677
678 int xen_vector_from_irq(unsigned irq)
679 {
680         return vector_from_irq(irq);
681 }
682
683 int xen_gsi_from_irq(unsigned irq)
684 {
685         return gsi_from_irq(irq);
686 }
687
688 int bind_evtchn_to_irq(unsigned int evtchn)
689 {
690         int irq;
691
692         spin_lock(&irq_mapping_update_lock);
693
694         irq = evtchn_to_irq[evtchn];
695
696         if (irq == -1) {
697                 irq = find_unbound_irq();
698
699                 set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
700                                               handle_edge_irq, "event");
701
702                 evtchn_to_irq[evtchn] = irq;
703                 irq_info[irq] = mk_evtchn_info(evtchn);
704         }
705
706         spin_unlock(&irq_mapping_update_lock);
707
708         return irq;
709 }
710 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
711
712 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
713 {
714         struct evtchn_bind_ipi bind_ipi;
715         int evtchn, irq;
716
717         spin_lock(&irq_mapping_update_lock);
718
719         irq = per_cpu(ipi_to_irq, cpu)[ipi];
720
721         if (irq == -1) {
722                 irq = find_unbound_irq();
723                 if (irq < 0)
724                         goto out;
725
726                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
727                                               handle_percpu_irq, "ipi");
728
729                 bind_ipi.vcpu = cpu;
730                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
731                                                 &bind_ipi) != 0)
732                         BUG();
733                 evtchn = bind_ipi.port;
734
735                 evtchn_to_irq[evtchn] = irq;
736                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
737                 per_cpu(ipi_to_irq, cpu)[ipi] = irq;
738
739                 bind_evtchn_to_cpu(evtchn, cpu);
740         }
741
742  out:
743         spin_unlock(&irq_mapping_update_lock);
744         return irq;
745 }
746
747
748 static int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
749 {
750         struct evtchn_bind_virq bind_virq;
751         int evtchn, irq;
752
753         spin_lock(&irq_mapping_update_lock);
754
755         irq = per_cpu(virq_to_irq, cpu)[virq];
756
757         if (irq == -1) {
758                 bind_virq.virq = virq;
759                 bind_virq.vcpu = cpu;
760                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
761                                                 &bind_virq) != 0)
762                         BUG();
763                 evtchn = bind_virq.port;
764
765                 irq = find_unbound_irq();
766
767                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
768                                               handle_percpu_irq, "virq");
769
770                 evtchn_to_irq[evtchn] = irq;
771                 irq_info[irq] = mk_virq_info(evtchn, virq);
772
773                 per_cpu(virq_to_irq, cpu)[virq] = irq;
774
775                 bind_evtchn_to_cpu(evtchn, cpu);
776         }
777
778         spin_unlock(&irq_mapping_update_lock);
779
780         return irq;
781 }
782
783 static void unbind_from_irq(unsigned int irq)
784 {
785         struct evtchn_close close;
786         int evtchn = evtchn_from_irq(irq);
787
788         spin_lock(&irq_mapping_update_lock);
789
790         if (VALID_EVTCHN(evtchn)) {
791                 close.port = evtchn;
792                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
793                         BUG();
794
795                 switch (type_from_irq(irq)) {
796                 case IRQT_VIRQ:
797                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
798                                 [virq_from_irq(irq)] = -1;
799                         break;
800                 case IRQT_IPI:
801                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
802                                 [ipi_from_irq(irq)] = -1;
803                         break;
804                 default:
805                         break;
806                 }
807
808                 /* Closed ports are implicitly re-bound to VCPU0. */
809                 bind_evtchn_to_cpu(evtchn, 0);
810
811                 evtchn_to_irq[evtchn] = -1;
812         }
813
814         if (irq_info[irq].type != IRQT_UNBOUND) {
815                 irq_info[irq] = mk_unbound_info();
816
817                 irq_free_desc(irq);
818         }
819
820         spin_unlock(&irq_mapping_update_lock);
821 }
822
823 int bind_evtchn_to_irqhandler(unsigned int evtchn,
824                               irq_handler_t handler,
825                               unsigned long irqflags,
826                               const char *devname, void *dev_id)
827 {
828         unsigned int irq;
829         int retval;
830
831         irq = bind_evtchn_to_irq(evtchn);
832         retval = request_irq(irq, handler, irqflags, devname, dev_id);
833         if (retval != 0) {
834                 unbind_from_irq(irq);
835                 return retval;
836         }
837
838         return irq;
839 }
840 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
841
842 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
843                             irq_handler_t handler,
844                             unsigned long irqflags, const char *devname, void *dev_id)
845 {
846         unsigned int irq;
847         int retval;
848
849         irq = bind_virq_to_irq(virq, cpu);
850         retval = request_irq(irq, handler, irqflags, devname, dev_id);
851         if (retval != 0) {
852                 unbind_from_irq(irq);
853                 return retval;
854         }
855
856         return irq;
857 }
858 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
859
860 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
861                            unsigned int cpu,
862                            irq_handler_t handler,
863                            unsigned long irqflags,
864                            const char *devname,
865                            void *dev_id)
866 {
867         int irq, retval;
868
869         irq = bind_ipi_to_irq(ipi, cpu);
870         if (irq < 0)
871                 return irq;
872
873         irqflags |= IRQF_NO_SUSPEND;
874         retval = request_irq(irq, handler, irqflags, devname, dev_id);
875         if (retval != 0) {
876                 unbind_from_irq(irq);
877                 return retval;
878         }
879
880         return irq;
881 }
882
883 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
884 {
885         free_irq(irq, dev_id);
886         unbind_from_irq(irq);
887 }
888 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
889
890 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
891 {
892         int irq = per_cpu(ipi_to_irq, cpu)[vector];
893         BUG_ON(irq < 0);
894         notify_remote_via_irq(irq);
895 }
896
897 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
898 {
899         struct shared_info *sh = HYPERVISOR_shared_info;
900         int cpu = smp_processor_id();
901         int i;
902         unsigned long flags;
903         static DEFINE_SPINLOCK(debug_lock);
904
905         spin_lock_irqsave(&debug_lock, flags);
906
907         printk("vcpu %d\n  ", cpu);
908
909         for_each_online_cpu(i) {
910                 struct vcpu_info *v = per_cpu(xen_vcpu, i);
911                 printk("%d: masked=%d pending=%d event_sel %08lx\n  ", i,
912                         (get_irq_regs() && i == cpu) ? xen_irqs_disabled(get_irq_regs()) : v->evtchn_upcall_mask,
913                         v->evtchn_upcall_pending,
914                         v->evtchn_pending_sel);
915         }
916         printk("pending:\n   ");
917         for(i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
918                 printk("%08lx%s", sh->evtchn_pending[i],
919                         i % 8 == 0 ? "\n   " : " ");
920         printk("\nmasks:\n   ");
921         for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
922                 printk("%08lx%s", sh->evtchn_mask[i],
923                         i % 8 == 0 ? "\n   " : " ");
924
925         printk("\nunmasked:\n   ");
926         for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
927                 printk("%08lx%s", sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
928                         i % 8 == 0 ? "\n   " : " ");
929
930         printk("\npending list:\n");
931         for(i = 0; i < NR_EVENT_CHANNELS; i++) {
932                 if (sync_test_bit(i, sh->evtchn_pending)) {
933                         printk("  %d: event %d -> irq %d\n",
934                                cpu_from_evtchn(i), i,
935                                evtchn_to_irq[i]);
936                 }
937         }
938
939         spin_unlock_irqrestore(&debug_lock, flags);
940
941         return IRQ_HANDLED;
942 }
943
944 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
945
946 /*
947  * Search the CPUs pending events bitmasks.  For each one found, map
948  * the event number to an irq, and feed it into do_IRQ() for
949  * handling.
950  *
951  * Xen uses a two-level bitmap to speed searching.  The first level is
952  * a bitset of words which contain pending event bits.  The second
953  * level is a bitset of pending events themselves.
954  */
955 static void __xen_evtchn_do_upcall(void)
956 {
957         int cpu = get_cpu();
958         struct shared_info *s = HYPERVISOR_shared_info;
959         struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
960         unsigned count;
961
962         do {
963                 unsigned long pending_words;
964
965                 vcpu_info->evtchn_upcall_pending = 0;
966
967                 if (__get_cpu_var(xed_nesting_count)++)
968                         goto out;
969
970 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
971                 /* Clear master flag /before/ clearing selector flag. */
972                 wmb();
973 #endif
974                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
975                 while (pending_words != 0) {
976                         unsigned long pending_bits;
977                         int word_idx = __ffs(pending_words);
978                         pending_words &= ~(1UL << word_idx);
979
980                         while ((pending_bits = active_evtchns(cpu, s, word_idx)) != 0) {
981                                 int bit_idx = __ffs(pending_bits);
982                                 int port = (word_idx * BITS_PER_LONG) + bit_idx;
983                                 int irq = evtchn_to_irq[port];
984                                 struct irq_desc *desc;
985
986                                 if (irq != -1) {
987                                         desc = irq_to_desc(irq);
988                                         if (desc)
989                                                 generic_handle_irq_desc(irq, desc);
990                                 }
991                         }
992                 }
993
994                 BUG_ON(!irqs_disabled());
995
996                 count = __get_cpu_var(xed_nesting_count);
997                 __get_cpu_var(xed_nesting_count) = 0;
998         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
999
1000 out:
1001
1002         put_cpu();
1003 }
1004
1005 void xen_evtchn_do_upcall(struct pt_regs *regs)
1006 {
1007         struct pt_regs *old_regs = set_irq_regs(regs);
1008
1009         exit_idle();
1010         irq_enter();
1011
1012         __xen_evtchn_do_upcall();
1013
1014         irq_exit();
1015         set_irq_regs(old_regs);
1016 }
1017
1018 void xen_hvm_evtchn_do_upcall(void)
1019 {
1020         __xen_evtchn_do_upcall();
1021 }
1022 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1023
1024 /* Rebind a new event channel to an existing irq. */
1025 void rebind_evtchn_irq(int evtchn, int irq)
1026 {
1027         struct irq_info *info = info_for_irq(irq);
1028
1029         /* Make sure the irq is masked, since the new event channel
1030            will also be masked. */
1031         disable_irq(irq);
1032
1033         spin_lock(&irq_mapping_update_lock);
1034
1035         /* After resume the irq<->evtchn mappings are all cleared out */
1036         BUG_ON(evtchn_to_irq[evtchn] != -1);
1037         /* Expect irq to have been bound before,
1038            so there should be a proper type */
1039         BUG_ON(info->type == IRQT_UNBOUND);
1040
1041         evtchn_to_irq[evtchn] = irq;
1042         irq_info[irq] = mk_evtchn_info(evtchn);
1043
1044         spin_unlock(&irq_mapping_update_lock);
1045
1046         /* new event channels are always bound to cpu 0 */
1047         irq_set_affinity(irq, cpumask_of(0));
1048
1049         /* Unmask the event channel. */
1050         enable_irq(irq);
1051 }
1052
1053 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1054 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1055 {
1056         struct evtchn_bind_vcpu bind_vcpu;
1057         int evtchn = evtchn_from_irq(irq);
1058
1059         /* events delivered via platform PCI interrupts are always
1060          * routed to vcpu 0 */
1061         if (!VALID_EVTCHN(evtchn) ||
1062                 (xen_hvm_domain() && !xen_have_vector_callback))
1063                 return -1;
1064
1065         /* Send future instances of this interrupt to other vcpu. */
1066         bind_vcpu.port = evtchn;
1067         bind_vcpu.vcpu = tcpu;
1068
1069         /*
1070          * If this fails, it usually just indicates that we're dealing with a
1071          * virq or IPI channel, which don't actually need to be rebound. Ignore
1072          * it, but don't do the xenlinux-level rebind in that case.
1073          */
1074         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1075                 bind_evtchn_to_cpu(evtchn, tcpu);
1076
1077         return 0;
1078 }
1079
1080 static int set_affinity_irq(unsigned irq, const struct cpumask *dest)
1081 {
1082         unsigned tcpu = cpumask_first(dest);
1083
1084         return rebind_irq_to_cpu(irq, tcpu);
1085 }
1086
1087 int resend_irq_on_evtchn(unsigned int irq)
1088 {
1089         int masked, evtchn = evtchn_from_irq(irq);
1090         struct shared_info *s = HYPERVISOR_shared_info;
1091
1092         if (!VALID_EVTCHN(evtchn))
1093                 return 1;
1094
1095         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1096         sync_set_bit(evtchn, s->evtchn_pending);
1097         if (!masked)
1098                 unmask_evtchn(evtchn);
1099
1100         return 1;
1101 }
1102
1103 static void enable_dynirq(unsigned int irq)
1104 {
1105         int evtchn = evtchn_from_irq(irq);
1106
1107         if (VALID_EVTCHN(evtchn))
1108                 unmask_evtchn(evtchn);
1109 }
1110
1111 static void disable_dynirq(unsigned int irq)
1112 {
1113         int evtchn = evtchn_from_irq(irq);
1114
1115         if (VALID_EVTCHN(evtchn))
1116                 mask_evtchn(evtchn);
1117 }
1118
1119 static void ack_dynirq(unsigned int irq)
1120 {
1121         int evtchn = evtchn_from_irq(irq);
1122
1123         move_native_irq(irq);
1124
1125         if (VALID_EVTCHN(evtchn))
1126                 clear_evtchn(evtchn);
1127 }
1128
1129 static int retrigger_dynirq(unsigned int irq)
1130 {
1131         int evtchn = evtchn_from_irq(irq);
1132         struct shared_info *sh = HYPERVISOR_shared_info;
1133         int ret = 0;
1134
1135         if (VALID_EVTCHN(evtchn)) {
1136                 int masked;
1137
1138                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1139                 sync_set_bit(evtchn, sh->evtchn_pending);
1140                 if (!masked)
1141                         unmask_evtchn(evtchn);
1142                 ret = 1;
1143         }
1144
1145         return ret;
1146 }
1147
1148 static void restore_cpu_virqs(unsigned int cpu)
1149 {
1150         struct evtchn_bind_virq bind_virq;
1151         int virq, irq, evtchn;
1152
1153         for (virq = 0; virq < NR_VIRQS; virq++) {
1154                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1155                         continue;
1156
1157                 BUG_ON(virq_from_irq(irq) != virq);
1158
1159                 /* Get a new binding from Xen. */
1160                 bind_virq.virq = virq;
1161                 bind_virq.vcpu = cpu;
1162                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1163                                                 &bind_virq) != 0)
1164                         BUG();
1165                 evtchn = bind_virq.port;
1166
1167                 /* Record the new mapping. */
1168                 evtchn_to_irq[evtchn] = irq;
1169                 irq_info[irq] = mk_virq_info(evtchn, virq);
1170                 bind_evtchn_to_cpu(evtchn, cpu);
1171
1172                 /* Ready for use. */
1173                 unmask_evtchn(evtchn);
1174         }
1175 }
1176
1177 static void restore_cpu_ipis(unsigned int cpu)
1178 {
1179         struct evtchn_bind_ipi bind_ipi;
1180         int ipi, irq, evtchn;
1181
1182         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1183                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1184                         continue;
1185
1186                 BUG_ON(ipi_from_irq(irq) != ipi);
1187
1188                 /* Get a new binding from Xen. */
1189                 bind_ipi.vcpu = cpu;
1190                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1191                                                 &bind_ipi) != 0)
1192                         BUG();
1193                 evtchn = bind_ipi.port;
1194
1195                 /* Record the new mapping. */
1196                 evtchn_to_irq[evtchn] = irq;
1197                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
1198                 bind_evtchn_to_cpu(evtchn, cpu);
1199
1200                 /* Ready for use. */
1201                 unmask_evtchn(evtchn);
1202
1203         }
1204 }
1205
1206 /* Clear an irq's pending state, in preparation for polling on it */
1207 void xen_clear_irq_pending(int irq)
1208 {
1209         int evtchn = evtchn_from_irq(irq);
1210
1211         if (VALID_EVTCHN(evtchn))
1212                 clear_evtchn(evtchn);
1213 }
1214 EXPORT_SYMBOL(xen_clear_irq_pending);
1215 void xen_set_irq_pending(int irq)
1216 {
1217         int evtchn = evtchn_from_irq(irq);
1218
1219         if (VALID_EVTCHN(evtchn))
1220                 set_evtchn(evtchn);
1221 }
1222
1223 bool xen_test_irq_pending(int irq)
1224 {
1225         int evtchn = evtchn_from_irq(irq);
1226         bool ret = false;
1227
1228         if (VALID_EVTCHN(evtchn))
1229                 ret = test_evtchn(evtchn);
1230
1231         return ret;
1232 }
1233
1234 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1235  * the irq will be disabled so it won't deliver an interrupt. */
1236 void xen_poll_irq_timeout(int irq, u64 timeout)
1237 {
1238         evtchn_port_t evtchn = evtchn_from_irq(irq);
1239
1240         if (VALID_EVTCHN(evtchn)) {
1241                 struct sched_poll poll;
1242
1243                 poll.nr_ports = 1;
1244                 poll.timeout = timeout;
1245                 set_xen_guest_handle(poll.ports, &evtchn);
1246
1247                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1248                         BUG();
1249         }
1250 }
1251 EXPORT_SYMBOL(xen_poll_irq_timeout);
1252 /* Poll waiting for an irq to become pending.  In the usual case, the
1253  * irq will be disabled so it won't deliver an interrupt. */
1254 void xen_poll_irq(int irq)
1255 {
1256         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1257 }
1258
1259 void xen_irq_resume(void)
1260 {
1261         unsigned int cpu, irq, evtchn;
1262
1263         init_evtchn_cpu_bindings();
1264
1265         /* New event-channel space is not 'live' yet. */
1266         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1267                 mask_evtchn(evtchn);
1268
1269         /* No IRQ <-> event-channel mappings. */
1270         for (irq = 0; irq < nr_irqs; irq++)
1271                 irq_info[irq].evtchn = 0; /* zap event-channel binding */
1272
1273         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1274                 evtchn_to_irq[evtchn] = -1;
1275
1276         for_each_possible_cpu(cpu) {
1277                 restore_cpu_virqs(cpu);
1278                 restore_cpu_ipis(cpu);
1279         }
1280 }
1281
1282 static struct irq_chip xen_dynamic_chip __read_mostly = {
1283         .name           = "xen-dyn",
1284
1285         .disable        = disable_dynirq,
1286         .mask           = disable_dynirq,
1287         .unmask         = enable_dynirq,
1288
1289         .ack            = ack_dynirq,
1290         .set_affinity   = set_affinity_irq,
1291         .retrigger      = retrigger_dynirq,
1292 };
1293
1294 static struct irq_chip xen_pirq_chip __read_mostly = {
1295         .name           = "xen-pirq",
1296
1297         .startup        = startup_pirq,
1298         .shutdown       = shutdown_pirq,
1299
1300         .enable         = enable_pirq,
1301         .unmask         = enable_pirq,
1302
1303         .disable        = disable_pirq,
1304         .mask           = disable_pirq,
1305
1306         .ack            = ack_pirq,
1307         .end            = end_pirq,
1308
1309         .set_affinity   = set_affinity_irq,
1310
1311         .retrigger      = retrigger_dynirq,
1312 };
1313
1314 static struct irq_chip xen_percpu_chip __read_mostly = {
1315         .name           = "xen-percpu",
1316
1317         .disable        = disable_dynirq,
1318         .mask           = disable_dynirq,
1319         .unmask         = enable_dynirq,
1320
1321         .ack            = ack_dynirq,
1322 };
1323
1324 int xen_set_callback_via(uint64_t via)
1325 {
1326         struct xen_hvm_param a;
1327         a.domid = DOMID_SELF;
1328         a.index = HVM_PARAM_CALLBACK_IRQ;
1329         a.value = via;
1330         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1331 }
1332 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1333
1334 #ifdef CONFIG_XEN_PVHVM
1335 /* Vector callbacks are better than PCI interrupts to receive event
1336  * channel notifications because we can receive vector callbacks on any
1337  * vcpu and we don't need PCI support or APIC interactions. */
1338 void xen_callback_vector(void)
1339 {
1340         int rc;
1341         uint64_t callback_via;
1342         if (xen_have_vector_callback) {
1343                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1344                 rc = xen_set_callback_via(callback_via);
1345                 if (rc) {
1346                         printk(KERN_ERR "Request for Xen HVM callback vector"
1347                                         " failed.\n");
1348                         xen_have_vector_callback = 0;
1349                         return;
1350                 }
1351                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1352                                 "enabled\n");
1353                 /* in the restore case the vector has already been allocated */
1354                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1355                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1356         }
1357 }
1358 #else
1359 void xen_callback_vector(void) {}
1360 #endif
1361
1362 void __init xen_init_IRQ(void)
1363 {
1364         int i, rc;
1365         struct physdev_nr_pirqs op_nr_pirqs;
1366
1367         cpu_evtchn_mask_p = kcalloc(nr_cpu_ids, sizeof(struct cpu_evtchn_s),
1368                                     GFP_KERNEL);
1369         irq_info = kcalloc(nr_irqs, sizeof(*irq_info), GFP_KERNEL);
1370
1371         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_nr_pirqs, &op_nr_pirqs);
1372         if (rc < 0) {
1373                 nr_pirqs = nr_irqs;
1374                 if (rc != -ENOSYS)
1375                         printk(KERN_WARNING "PHYSDEVOP_get_nr_pirqs returned rc=%d\n", rc);
1376         } else {
1377                 if (xen_pv_domain() && !xen_initial_domain())
1378                         nr_pirqs = max((int)op_nr_pirqs.nr_pirqs, nr_irqs);
1379                 else
1380                         nr_pirqs = op_nr_pirqs.nr_pirqs;
1381         }
1382         pirq_to_irq = kcalloc(nr_pirqs, sizeof(*pirq_to_irq), GFP_KERNEL);
1383         for (i = 0; i < nr_pirqs; i++)
1384                 pirq_to_irq[i] = -1;
1385
1386         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1387                                     GFP_KERNEL);
1388         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1389                 evtchn_to_irq[i] = -1;
1390
1391         init_evtchn_cpu_bindings();
1392
1393         /* No event channels are 'live' right now. */
1394         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1395                 mask_evtchn(i);
1396
1397         if (xen_hvm_domain()) {
1398                 xen_callback_vector();
1399                 native_init_IRQ();
1400         } else {
1401                 irq_ctx_init(smp_processor_id());
1402         }
1403 }