xenbus: Xen paravirtualised PCI hotplug support.
[linux-2.6.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/hypercall.h>
40 #include <asm/xen/hypervisor.h>
41
42 #include <xen/xen.h>
43 #include <xen/hvm.h>
44 #include <xen/xen-ops.h>
45 #include <xen/events.h>
46 #include <xen/interface/xen.h>
47 #include <xen/interface/event_channel.h>
48 #include <xen/interface/hvm/hvm_op.h>
49 #include <xen/interface/hvm/params.h>
50
51 /*
52  * This lock protects updates to the following mapping and reference-count
53  * arrays. The lock does not need to be acquired to read the mapping tables.
54  */
55 static DEFINE_SPINLOCK(irq_mapping_update_lock);
56
57 /* IRQ <-> VIRQ mapping. */
58 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
59
60 /* IRQ <-> IPI mapping */
61 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
62
63 /* Interrupt types. */
64 enum xen_irq_type {
65         IRQT_UNBOUND = 0,
66         IRQT_PIRQ,
67         IRQT_VIRQ,
68         IRQT_IPI,
69         IRQT_EVTCHN
70 };
71
72 /*
73  * Packed IRQ information:
74  * type - enum xen_irq_type
75  * event channel - irq->event channel mapping
76  * cpu - cpu this event channel is bound to
77  * index - type-specific information:
78  *    PIRQ - vector, with MSB being "needs EIO"
79  *    VIRQ - virq number
80  *    IPI - IPI vector
81  *    EVTCHN -
82  */
83 struct irq_info
84 {
85         enum xen_irq_type type; /* type */
86         unsigned short evtchn;  /* event channel */
87         unsigned short cpu;     /* cpu bound */
88
89         union {
90                 unsigned short virq;
91                 enum ipi_vector ipi;
92                 struct {
93                         unsigned short gsi;
94                         unsigned char vector;
95                         unsigned char flags;
96                 } pirq;
97         } u;
98 };
99 #define PIRQ_NEEDS_EOI  (1 << 0)
100 #define PIRQ_SHAREABLE  (1 << 1)
101
102 static struct irq_info *irq_info;
103
104 static int *evtchn_to_irq;
105 struct cpu_evtchn_s {
106         unsigned long bits[NR_EVENT_CHANNELS/BITS_PER_LONG];
107 };
108
109 static __initdata struct cpu_evtchn_s init_evtchn_mask = {
110         .bits[0 ... (NR_EVENT_CHANNELS/BITS_PER_LONG)-1] = ~0ul,
111 };
112 static struct cpu_evtchn_s *cpu_evtchn_mask_p = &init_evtchn_mask;
113
114 static inline unsigned long *cpu_evtchn_mask(int cpu)
115 {
116         return cpu_evtchn_mask_p[cpu].bits;
117 }
118
119 /* Xen will never allocate port zero for any purpose. */
120 #define VALID_EVTCHN(chn)       ((chn) != 0)
121
122 static struct irq_chip xen_dynamic_chip;
123 static struct irq_chip xen_percpu_chip;
124 static struct irq_chip xen_pirq_chip;
125
126 /* Constructor for packed IRQ information. */
127 static struct irq_info mk_unbound_info(void)
128 {
129         return (struct irq_info) { .type = IRQT_UNBOUND };
130 }
131
132 static struct irq_info mk_evtchn_info(unsigned short evtchn)
133 {
134         return (struct irq_info) { .type = IRQT_EVTCHN, .evtchn = evtchn,
135                         .cpu = 0 };
136 }
137
138 static struct irq_info mk_ipi_info(unsigned short evtchn, enum ipi_vector ipi)
139 {
140         return (struct irq_info) { .type = IRQT_IPI, .evtchn = evtchn,
141                         .cpu = 0, .u.ipi = ipi };
142 }
143
144 static struct irq_info mk_virq_info(unsigned short evtchn, unsigned short virq)
145 {
146         return (struct irq_info) { .type = IRQT_VIRQ, .evtchn = evtchn,
147                         .cpu = 0, .u.virq = virq };
148 }
149
150 static struct irq_info mk_pirq_info(unsigned short evtchn,
151                                     unsigned short gsi, unsigned short vector)
152 {
153         return (struct irq_info) { .type = IRQT_PIRQ, .evtchn = evtchn,
154                         .cpu = 0, .u.pirq = { .gsi = gsi, .vector = vector } };
155 }
156
157 /*
158  * Accessors for packed IRQ information.
159  */
160 static struct irq_info *info_for_irq(unsigned irq)
161 {
162         return &irq_info[irq];
163 }
164
165 static unsigned int evtchn_from_irq(unsigned irq)
166 {
167         return info_for_irq(irq)->evtchn;
168 }
169
170 unsigned irq_from_evtchn(unsigned int evtchn)
171 {
172         return evtchn_to_irq[evtchn];
173 }
174 EXPORT_SYMBOL_GPL(irq_from_evtchn);
175
176 static enum ipi_vector ipi_from_irq(unsigned irq)
177 {
178         struct irq_info *info = info_for_irq(irq);
179
180         BUG_ON(info == NULL);
181         BUG_ON(info->type != IRQT_IPI);
182
183         return info->u.ipi;
184 }
185
186 static unsigned virq_from_irq(unsigned irq)
187 {
188         struct irq_info *info = info_for_irq(irq);
189
190         BUG_ON(info == NULL);
191         BUG_ON(info->type != IRQT_VIRQ);
192
193         return info->u.virq;
194 }
195
196 static unsigned gsi_from_irq(unsigned irq)
197 {
198         struct irq_info *info = info_for_irq(irq);
199
200         BUG_ON(info == NULL);
201         BUG_ON(info->type != IRQT_PIRQ);
202
203         return info->u.pirq.gsi;
204 }
205
206 static unsigned vector_from_irq(unsigned irq)
207 {
208         struct irq_info *info = info_for_irq(irq);
209
210         BUG_ON(info == NULL);
211         BUG_ON(info->type != IRQT_PIRQ);
212
213         return info->u.pirq.vector;
214 }
215
216 static enum xen_irq_type type_from_irq(unsigned irq)
217 {
218         return info_for_irq(irq)->type;
219 }
220
221 static unsigned cpu_from_irq(unsigned irq)
222 {
223         return info_for_irq(irq)->cpu;
224 }
225
226 static unsigned int cpu_from_evtchn(unsigned int evtchn)
227 {
228         int irq = evtchn_to_irq[evtchn];
229         unsigned ret = 0;
230
231         if (irq != -1)
232                 ret = cpu_from_irq(irq);
233
234         return ret;
235 }
236
237 static bool pirq_needs_eoi(unsigned irq)
238 {
239         struct irq_info *info = info_for_irq(irq);
240
241         BUG_ON(info->type != IRQT_PIRQ);
242
243         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
244 }
245
246 static inline unsigned long active_evtchns(unsigned int cpu,
247                                            struct shared_info *sh,
248                                            unsigned int idx)
249 {
250         return (sh->evtchn_pending[idx] &
251                 cpu_evtchn_mask(cpu)[idx] &
252                 ~sh->evtchn_mask[idx]);
253 }
254
255 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
256 {
257         int irq = evtchn_to_irq[chn];
258
259         BUG_ON(irq == -1);
260 #ifdef CONFIG_SMP
261         cpumask_copy(irq_to_desc(irq)->affinity, cpumask_of(cpu));
262 #endif
263
264         __clear_bit(chn, cpu_evtchn_mask(cpu_from_irq(irq)));
265         __set_bit(chn, cpu_evtchn_mask(cpu));
266
267         irq_info[irq].cpu = cpu;
268 }
269
270 static void init_evtchn_cpu_bindings(void)
271 {
272 #ifdef CONFIG_SMP
273         struct irq_desc *desc;
274         int i;
275
276         /* By default all event channels notify CPU#0. */
277         for_each_irq_desc(i, desc) {
278                 cpumask_copy(desc->affinity, cpumask_of(0));
279         }
280 #endif
281
282         memset(cpu_evtchn_mask(0), ~0, sizeof(cpu_evtchn_mask(0)));
283 }
284
285 static inline void clear_evtchn(int port)
286 {
287         struct shared_info *s = HYPERVISOR_shared_info;
288         sync_clear_bit(port, &s->evtchn_pending[0]);
289 }
290
291 static inline void set_evtchn(int port)
292 {
293         struct shared_info *s = HYPERVISOR_shared_info;
294         sync_set_bit(port, &s->evtchn_pending[0]);
295 }
296
297 static inline int test_evtchn(int port)
298 {
299         struct shared_info *s = HYPERVISOR_shared_info;
300         return sync_test_bit(port, &s->evtchn_pending[0]);
301 }
302
303
304 /**
305  * notify_remote_via_irq - send event to remote end of event channel via irq
306  * @irq: irq of event channel to send event to
307  *
308  * Unlike notify_remote_via_evtchn(), this is safe to use across
309  * save/restore. Notifications on a broken connection are silently
310  * dropped.
311  */
312 void notify_remote_via_irq(int irq)
313 {
314         int evtchn = evtchn_from_irq(irq);
315
316         if (VALID_EVTCHN(evtchn))
317                 notify_remote_via_evtchn(evtchn);
318 }
319 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
320
321 static void mask_evtchn(int port)
322 {
323         struct shared_info *s = HYPERVISOR_shared_info;
324         sync_set_bit(port, &s->evtchn_mask[0]);
325 }
326
327 static void unmask_evtchn(int port)
328 {
329         struct shared_info *s = HYPERVISOR_shared_info;
330         unsigned int cpu = get_cpu();
331
332         BUG_ON(!irqs_disabled());
333
334         /* Slow path (hypercall) if this is a non-local port. */
335         if (unlikely(cpu != cpu_from_evtchn(port))) {
336                 struct evtchn_unmask unmask = { .port = port };
337                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
338         } else {
339                 struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
340
341                 sync_clear_bit(port, &s->evtchn_mask[0]);
342
343                 /*
344                  * The following is basically the equivalent of
345                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
346                  * the interrupt edge' if the channel is masked.
347                  */
348                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
349                     !sync_test_and_set_bit(port / BITS_PER_LONG,
350                                            &vcpu_info->evtchn_pending_sel))
351                         vcpu_info->evtchn_upcall_pending = 1;
352         }
353
354         put_cpu();
355 }
356
357 static int get_nr_hw_irqs(void)
358 {
359         int ret = 1;
360
361 #ifdef CONFIG_X86_IO_APIC
362         ret = get_nr_irqs_gsi();
363 #endif
364
365         return ret;
366 }
367
368 static int find_unbound_irq(void)
369 {
370         struct irq_data *data;
371         int irq, res;
372         int start = get_nr_hw_irqs();
373
374         if (start == nr_irqs)
375                 goto no_irqs;
376
377         /* nr_irqs is a magic value. Must not use it.*/
378         for (irq = nr_irqs-1; irq > start; irq--) {
379                 data = irq_get_irq_data(irq);
380                 /* only 0->15 have init'd desc; handle irq > 16 */
381                 if (!data)
382                         break;
383                 if (data->chip == &no_irq_chip)
384                         break;
385                 if (data->chip != &xen_dynamic_chip)
386                         continue;
387                 if (irq_info[irq].type == IRQT_UNBOUND)
388                         return irq;
389         }
390
391         if (irq == start)
392                 goto no_irqs;
393
394         res = irq_alloc_desc_at(irq, 0);
395
396         if (WARN_ON(res != irq))
397                 return -1;
398
399         return irq;
400
401 no_irqs:
402         panic("No available IRQ to bind to: increase nr_irqs!\n");
403 }
404
405 static bool identity_mapped_irq(unsigned irq)
406 {
407         /* identity map all the hardware irqs */
408         return irq < get_nr_hw_irqs();
409 }
410
411 static void pirq_unmask_notify(int irq)
412 {
413         struct physdev_eoi eoi = { .irq = irq };
414
415         if (unlikely(pirq_needs_eoi(irq))) {
416                 int rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
417                 WARN_ON(rc);
418         }
419 }
420
421 static void pirq_query_unmask(int irq)
422 {
423         struct physdev_irq_status_query irq_status;
424         struct irq_info *info = info_for_irq(irq);
425
426         BUG_ON(info->type != IRQT_PIRQ);
427
428         irq_status.irq = irq;
429         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
430                 irq_status.flags = 0;
431
432         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
433         if (irq_status.flags & XENIRQSTAT_needs_eoi)
434                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
435 }
436
437 static bool probing_irq(int irq)
438 {
439         struct irq_desc *desc = irq_to_desc(irq);
440
441         return desc && desc->action == NULL;
442 }
443
444 static unsigned int startup_pirq(unsigned int irq)
445 {
446         struct evtchn_bind_pirq bind_pirq;
447         struct irq_info *info = info_for_irq(irq);
448         int evtchn = evtchn_from_irq(irq);
449         int rc;
450
451         BUG_ON(info->type != IRQT_PIRQ);
452
453         if (VALID_EVTCHN(evtchn))
454                 goto out;
455
456         bind_pirq.pirq = irq;
457         /* NB. We are happy to share unless we are probing. */
458         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
459                                         BIND_PIRQ__WILL_SHARE : 0;
460         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
461         if (rc != 0) {
462                 if (!probing_irq(irq))
463                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
464                                irq);
465                 return 0;
466         }
467         evtchn = bind_pirq.port;
468
469         pirq_query_unmask(irq);
470
471         evtchn_to_irq[evtchn] = irq;
472         bind_evtchn_to_cpu(evtchn, 0);
473         info->evtchn = evtchn;
474
475 out:
476         unmask_evtchn(evtchn);
477         pirq_unmask_notify(irq);
478
479         return 0;
480 }
481
482 static void shutdown_pirq(unsigned int irq)
483 {
484         struct evtchn_close close;
485         struct irq_info *info = info_for_irq(irq);
486         int evtchn = evtchn_from_irq(irq);
487
488         BUG_ON(info->type != IRQT_PIRQ);
489
490         if (!VALID_EVTCHN(evtchn))
491                 return;
492
493         mask_evtchn(evtchn);
494
495         close.port = evtchn;
496         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
497                 BUG();
498
499         bind_evtchn_to_cpu(evtchn, 0);
500         evtchn_to_irq[evtchn] = -1;
501         info->evtchn = 0;
502 }
503
504 static void enable_pirq(unsigned int irq)
505 {
506         startup_pirq(irq);
507 }
508
509 static void disable_pirq(unsigned int irq)
510 {
511 }
512
513 static void ack_pirq(unsigned int irq)
514 {
515         int evtchn = evtchn_from_irq(irq);
516
517         move_native_irq(irq);
518
519         if (VALID_EVTCHN(evtchn)) {
520                 mask_evtchn(evtchn);
521                 clear_evtchn(evtchn);
522         }
523 }
524
525 static void end_pirq(unsigned int irq)
526 {
527         int evtchn = evtchn_from_irq(irq);
528         struct irq_desc *desc = irq_to_desc(irq);
529
530         if (WARN_ON(!desc))
531                 return;
532
533         if ((desc->status & (IRQ_DISABLED|IRQ_PENDING)) ==
534             (IRQ_DISABLED|IRQ_PENDING)) {
535                 shutdown_pirq(irq);
536         } else if (VALID_EVTCHN(evtchn)) {
537                 unmask_evtchn(evtchn);
538                 pirq_unmask_notify(irq);
539         }
540 }
541
542 static int find_irq_by_gsi(unsigned gsi)
543 {
544         int irq;
545
546         for (irq = 0; irq < nr_irqs; irq++) {
547                 struct irq_info *info = info_for_irq(irq);
548
549                 if (info == NULL || info->type != IRQT_PIRQ)
550                         continue;
551
552                 if (gsi_from_irq(irq) == gsi)
553                         return irq;
554         }
555
556         return -1;
557 }
558
559 /* xen_allocate_irq might allocate irqs from the top down, as a
560  * consequence don't assume that the irq number returned has a low value
561  * or can be used as a pirq number unless you know otherwise.
562  *
563  * One notable exception is when xen_allocate_irq is called passing an
564  * hardware gsi as argument, in that case the irq number returned
565  * matches the gsi number passed as first argument.
566
567  * Note: We don't assign an
568  * event channel until the irq actually started up.  Return an
569  * existing irq if we've already got one for the gsi.
570  */
571 int xen_allocate_pirq(unsigned gsi, int shareable, char *name)
572 {
573         int irq;
574         struct physdev_irq irq_op;
575
576         spin_lock(&irq_mapping_update_lock);
577
578         irq = find_irq_by_gsi(gsi);
579         if (irq != -1) {
580                 printk(KERN_INFO "xen_allocate_pirq: returning irq %d for gsi %u\n",
581                        irq, gsi);
582                 goto out;       /* XXX need refcount? */
583         }
584
585         /* If we are a PV guest, we don't have GSIs (no ACPI passed). Therefore
586          * we are using the !xen_initial_domain() to drop in the function.*/
587         if (identity_mapped_irq(gsi) || !xen_initial_domain()) {
588                 irq = gsi;
589                 irq_to_desc_alloc_node(irq, 0);
590                 dynamic_irq_init(irq);
591         } else
592                 irq = find_unbound_irq();
593
594         set_irq_chip_and_handler_name(irq, &xen_pirq_chip,
595                                       handle_level_irq, name);
596
597         irq_op.irq = irq;
598         irq_op.vector = 0;
599
600         /* Only the privileged domain can do this. For non-priv, the pcifront
601          * driver provides a PCI bus that does the call to do exactly
602          * this in the priv domain. */
603         if (xen_initial_domain() &&
604             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
605                 dynamic_irq_cleanup(irq);
606                 irq = -ENOSPC;
607                 goto out;
608         }
609
610         irq_info[irq] = mk_pirq_info(0, gsi, irq_op.vector);
611         irq_info[irq].u.pirq.flags |= shareable ? PIRQ_SHAREABLE : 0;
612
613 out:
614         spin_unlock(&irq_mapping_update_lock);
615
616         return irq;
617 }
618
619 int xen_destroy_irq(int irq)
620 {
621         struct irq_desc *desc;
622         int rc = -ENOENT;
623
624         spin_lock(&irq_mapping_update_lock);
625
626         desc = irq_to_desc(irq);
627         if (!desc)
628                 goto out;
629
630         irq_info[irq] = mk_unbound_info();
631
632         dynamic_irq_cleanup(irq);
633
634 out:
635         spin_unlock(&irq_mapping_update_lock);
636         return rc;
637 }
638
639 int xen_vector_from_irq(unsigned irq)
640 {
641         return vector_from_irq(irq);
642 }
643
644 int xen_gsi_from_irq(unsigned irq)
645 {
646         return gsi_from_irq(irq);
647 }
648
649 int bind_evtchn_to_irq(unsigned int evtchn)
650 {
651         int irq;
652
653         spin_lock(&irq_mapping_update_lock);
654
655         irq = evtchn_to_irq[evtchn];
656
657         if (irq == -1) {
658                 irq = find_unbound_irq();
659
660                 set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
661                                               handle_edge_irq, "event");
662
663                 evtchn_to_irq[evtchn] = irq;
664                 irq_info[irq] = mk_evtchn_info(evtchn);
665         }
666
667         spin_unlock(&irq_mapping_update_lock);
668
669         return irq;
670 }
671 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
672
673 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
674 {
675         struct evtchn_bind_ipi bind_ipi;
676         int evtchn, irq;
677
678         spin_lock(&irq_mapping_update_lock);
679
680         irq = per_cpu(ipi_to_irq, cpu)[ipi];
681
682         if (irq == -1) {
683                 irq = find_unbound_irq();
684                 if (irq < 0)
685                         goto out;
686
687                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
688                                               handle_percpu_irq, "ipi");
689
690                 bind_ipi.vcpu = cpu;
691                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
692                                                 &bind_ipi) != 0)
693                         BUG();
694                 evtchn = bind_ipi.port;
695
696                 evtchn_to_irq[evtchn] = irq;
697                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
698                 per_cpu(ipi_to_irq, cpu)[ipi] = irq;
699
700                 bind_evtchn_to_cpu(evtchn, cpu);
701         }
702
703  out:
704         spin_unlock(&irq_mapping_update_lock);
705         return irq;
706 }
707
708
709 static int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
710 {
711         struct evtchn_bind_virq bind_virq;
712         int evtchn, irq;
713
714         spin_lock(&irq_mapping_update_lock);
715
716         irq = per_cpu(virq_to_irq, cpu)[virq];
717
718         if (irq == -1) {
719                 bind_virq.virq = virq;
720                 bind_virq.vcpu = cpu;
721                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
722                                                 &bind_virq) != 0)
723                         BUG();
724                 evtchn = bind_virq.port;
725
726                 irq = find_unbound_irq();
727
728                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
729                                               handle_percpu_irq, "virq");
730
731                 evtchn_to_irq[evtchn] = irq;
732                 irq_info[irq] = mk_virq_info(evtchn, virq);
733
734                 per_cpu(virq_to_irq, cpu)[virq] = irq;
735
736                 bind_evtchn_to_cpu(evtchn, cpu);
737         }
738
739         spin_unlock(&irq_mapping_update_lock);
740
741         return irq;
742 }
743
744 static void unbind_from_irq(unsigned int irq)
745 {
746         struct evtchn_close close;
747         int evtchn = evtchn_from_irq(irq);
748
749         spin_lock(&irq_mapping_update_lock);
750
751         if (VALID_EVTCHN(evtchn)) {
752                 close.port = evtchn;
753                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
754                         BUG();
755
756                 switch (type_from_irq(irq)) {
757                 case IRQT_VIRQ:
758                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
759                                 [virq_from_irq(irq)] = -1;
760                         break;
761                 case IRQT_IPI:
762                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
763                                 [ipi_from_irq(irq)] = -1;
764                         break;
765                 default:
766                         break;
767                 }
768
769                 /* Closed ports are implicitly re-bound to VCPU0. */
770                 bind_evtchn_to_cpu(evtchn, 0);
771
772                 evtchn_to_irq[evtchn] = -1;
773         }
774
775         if (irq_info[irq].type != IRQT_UNBOUND) {
776                 irq_info[irq] = mk_unbound_info();
777
778                 irq_free_desc(irq);
779         }
780
781         spin_unlock(&irq_mapping_update_lock);
782 }
783
784 int bind_evtchn_to_irqhandler(unsigned int evtchn,
785                               irq_handler_t handler,
786                               unsigned long irqflags,
787                               const char *devname, void *dev_id)
788 {
789         unsigned int irq;
790         int retval;
791
792         irq = bind_evtchn_to_irq(evtchn);
793         retval = request_irq(irq, handler, irqflags, devname, dev_id);
794         if (retval != 0) {
795                 unbind_from_irq(irq);
796                 return retval;
797         }
798
799         return irq;
800 }
801 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
802
803 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
804                             irq_handler_t handler,
805                             unsigned long irqflags, const char *devname, void *dev_id)
806 {
807         unsigned int irq;
808         int retval;
809
810         irq = bind_virq_to_irq(virq, cpu);
811         retval = request_irq(irq, handler, irqflags, devname, dev_id);
812         if (retval != 0) {
813                 unbind_from_irq(irq);
814                 return retval;
815         }
816
817         return irq;
818 }
819 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
820
821 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
822                            unsigned int cpu,
823                            irq_handler_t handler,
824                            unsigned long irqflags,
825                            const char *devname,
826                            void *dev_id)
827 {
828         int irq, retval;
829
830         irq = bind_ipi_to_irq(ipi, cpu);
831         if (irq < 0)
832                 return irq;
833
834         irqflags |= IRQF_NO_SUSPEND;
835         retval = request_irq(irq, handler, irqflags, devname, dev_id);
836         if (retval != 0) {
837                 unbind_from_irq(irq);
838                 return retval;
839         }
840
841         return irq;
842 }
843
844 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
845 {
846         free_irq(irq, dev_id);
847         unbind_from_irq(irq);
848 }
849 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
850
851 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
852 {
853         int irq = per_cpu(ipi_to_irq, cpu)[vector];
854         BUG_ON(irq < 0);
855         notify_remote_via_irq(irq);
856 }
857
858 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
859 {
860         struct shared_info *sh = HYPERVISOR_shared_info;
861         int cpu = smp_processor_id();
862         int i;
863         unsigned long flags;
864         static DEFINE_SPINLOCK(debug_lock);
865
866         spin_lock_irqsave(&debug_lock, flags);
867
868         printk("vcpu %d\n  ", cpu);
869
870         for_each_online_cpu(i) {
871                 struct vcpu_info *v = per_cpu(xen_vcpu, i);
872                 printk("%d: masked=%d pending=%d event_sel %08lx\n  ", i,
873                         (get_irq_regs() && i == cpu) ? xen_irqs_disabled(get_irq_regs()) : v->evtchn_upcall_mask,
874                         v->evtchn_upcall_pending,
875                         v->evtchn_pending_sel);
876         }
877         printk("pending:\n   ");
878         for(i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
879                 printk("%08lx%s", sh->evtchn_pending[i],
880                         i % 8 == 0 ? "\n   " : " ");
881         printk("\nmasks:\n   ");
882         for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
883                 printk("%08lx%s", sh->evtchn_mask[i],
884                         i % 8 == 0 ? "\n   " : " ");
885
886         printk("\nunmasked:\n   ");
887         for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
888                 printk("%08lx%s", sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
889                         i % 8 == 0 ? "\n   " : " ");
890
891         printk("\npending list:\n");
892         for(i = 0; i < NR_EVENT_CHANNELS; i++) {
893                 if (sync_test_bit(i, sh->evtchn_pending)) {
894                         printk("  %d: event %d -> irq %d\n",
895                                cpu_from_evtchn(i), i,
896                                evtchn_to_irq[i]);
897                 }
898         }
899
900         spin_unlock_irqrestore(&debug_lock, flags);
901
902         return IRQ_HANDLED;
903 }
904
905 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
906
907 /*
908  * Search the CPUs pending events bitmasks.  For each one found, map
909  * the event number to an irq, and feed it into do_IRQ() for
910  * handling.
911  *
912  * Xen uses a two-level bitmap to speed searching.  The first level is
913  * a bitset of words which contain pending event bits.  The second
914  * level is a bitset of pending events themselves.
915  */
916 static void __xen_evtchn_do_upcall(void)
917 {
918         int cpu = get_cpu();
919         struct shared_info *s = HYPERVISOR_shared_info;
920         struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
921         unsigned count;
922
923         do {
924                 unsigned long pending_words;
925
926                 vcpu_info->evtchn_upcall_pending = 0;
927
928                 if (__get_cpu_var(xed_nesting_count)++)
929                         goto out;
930
931 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
932                 /* Clear master flag /before/ clearing selector flag. */
933                 wmb();
934 #endif
935                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
936                 while (pending_words != 0) {
937                         unsigned long pending_bits;
938                         int word_idx = __ffs(pending_words);
939                         pending_words &= ~(1UL << word_idx);
940
941                         while ((pending_bits = active_evtchns(cpu, s, word_idx)) != 0) {
942                                 int bit_idx = __ffs(pending_bits);
943                                 int port = (word_idx * BITS_PER_LONG) + bit_idx;
944                                 int irq = evtchn_to_irq[port];
945                                 struct irq_desc *desc;
946
947                                 if (irq != -1) {
948                                         desc = irq_to_desc(irq);
949                                         if (desc)
950                                                 generic_handle_irq_desc(irq, desc);
951                                 }
952                         }
953                 }
954
955                 BUG_ON(!irqs_disabled());
956
957                 count = __get_cpu_var(xed_nesting_count);
958                 __get_cpu_var(xed_nesting_count) = 0;
959         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
960
961 out:
962
963         put_cpu();
964 }
965
966 void xen_evtchn_do_upcall(struct pt_regs *regs)
967 {
968         struct pt_regs *old_regs = set_irq_regs(regs);
969
970         exit_idle();
971         irq_enter();
972
973         __xen_evtchn_do_upcall();
974
975         irq_exit();
976         set_irq_regs(old_regs);
977 }
978
979 void xen_hvm_evtchn_do_upcall(void)
980 {
981         __xen_evtchn_do_upcall();
982 }
983 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
984
985 /* Rebind a new event channel to an existing irq. */
986 void rebind_evtchn_irq(int evtchn, int irq)
987 {
988         struct irq_info *info = info_for_irq(irq);
989
990         /* Make sure the irq is masked, since the new event channel
991            will also be masked. */
992         disable_irq(irq);
993
994         spin_lock(&irq_mapping_update_lock);
995
996         /* After resume the irq<->evtchn mappings are all cleared out */
997         BUG_ON(evtchn_to_irq[evtchn] != -1);
998         /* Expect irq to have been bound before,
999            so there should be a proper type */
1000         BUG_ON(info->type == IRQT_UNBOUND);
1001
1002         evtchn_to_irq[evtchn] = irq;
1003         irq_info[irq] = mk_evtchn_info(evtchn);
1004
1005         spin_unlock(&irq_mapping_update_lock);
1006
1007         /* new event channels are always bound to cpu 0 */
1008         irq_set_affinity(irq, cpumask_of(0));
1009
1010         /* Unmask the event channel. */
1011         enable_irq(irq);
1012 }
1013
1014 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1015 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1016 {
1017         struct evtchn_bind_vcpu bind_vcpu;
1018         int evtchn = evtchn_from_irq(irq);
1019
1020         /* events delivered via platform PCI interrupts are always
1021          * routed to vcpu 0 */
1022         if (!VALID_EVTCHN(evtchn) ||
1023                 (xen_hvm_domain() && !xen_have_vector_callback))
1024                 return -1;
1025
1026         /* Send future instances of this interrupt to other vcpu. */
1027         bind_vcpu.port = evtchn;
1028         bind_vcpu.vcpu = tcpu;
1029
1030         /*
1031          * If this fails, it usually just indicates that we're dealing with a
1032          * virq or IPI channel, which don't actually need to be rebound. Ignore
1033          * it, but don't do the xenlinux-level rebind in that case.
1034          */
1035         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1036                 bind_evtchn_to_cpu(evtchn, tcpu);
1037
1038         return 0;
1039 }
1040
1041 static int set_affinity_irq(unsigned irq, const struct cpumask *dest)
1042 {
1043         unsigned tcpu = cpumask_first(dest);
1044
1045         return rebind_irq_to_cpu(irq, tcpu);
1046 }
1047
1048 int resend_irq_on_evtchn(unsigned int irq)
1049 {
1050         int masked, evtchn = evtchn_from_irq(irq);
1051         struct shared_info *s = HYPERVISOR_shared_info;
1052
1053         if (!VALID_EVTCHN(evtchn))
1054                 return 1;
1055
1056         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1057         sync_set_bit(evtchn, s->evtchn_pending);
1058         if (!masked)
1059                 unmask_evtchn(evtchn);
1060
1061         return 1;
1062 }
1063
1064 static void enable_dynirq(unsigned int irq)
1065 {
1066         int evtchn = evtchn_from_irq(irq);
1067
1068         if (VALID_EVTCHN(evtchn))
1069                 unmask_evtchn(evtchn);
1070 }
1071
1072 static void disable_dynirq(unsigned int irq)
1073 {
1074         int evtchn = evtchn_from_irq(irq);
1075
1076         if (VALID_EVTCHN(evtchn))
1077                 mask_evtchn(evtchn);
1078 }
1079
1080 static void ack_dynirq(unsigned int irq)
1081 {
1082         int evtchn = evtchn_from_irq(irq);
1083
1084         move_native_irq(irq);
1085
1086         if (VALID_EVTCHN(evtchn))
1087                 clear_evtchn(evtchn);
1088 }
1089
1090 static int retrigger_dynirq(unsigned int irq)
1091 {
1092         int evtchn = evtchn_from_irq(irq);
1093         struct shared_info *sh = HYPERVISOR_shared_info;
1094         int ret = 0;
1095
1096         if (VALID_EVTCHN(evtchn)) {
1097                 int masked;
1098
1099                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1100                 sync_set_bit(evtchn, sh->evtchn_pending);
1101                 if (!masked)
1102                         unmask_evtchn(evtchn);
1103                 ret = 1;
1104         }
1105
1106         return ret;
1107 }
1108
1109 static void restore_cpu_virqs(unsigned int cpu)
1110 {
1111         struct evtchn_bind_virq bind_virq;
1112         int virq, irq, evtchn;
1113
1114         for (virq = 0; virq < NR_VIRQS; virq++) {
1115                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1116                         continue;
1117
1118                 BUG_ON(virq_from_irq(irq) != virq);
1119
1120                 /* Get a new binding from Xen. */
1121                 bind_virq.virq = virq;
1122                 bind_virq.vcpu = cpu;
1123                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1124                                                 &bind_virq) != 0)
1125                         BUG();
1126                 evtchn = bind_virq.port;
1127
1128                 /* Record the new mapping. */
1129                 evtchn_to_irq[evtchn] = irq;
1130                 irq_info[irq] = mk_virq_info(evtchn, virq);
1131                 bind_evtchn_to_cpu(evtchn, cpu);
1132
1133                 /* Ready for use. */
1134                 unmask_evtchn(evtchn);
1135         }
1136 }
1137
1138 static void restore_cpu_ipis(unsigned int cpu)
1139 {
1140         struct evtchn_bind_ipi bind_ipi;
1141         int ipi, irq, evtchn;
1142
1143         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1144                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1145                         continue;
1146
1147                 BUG_ON(ipi_from_irq(irq) != ipi);
1148
1149                 /* Get a new binding from Xen. */
1150                 bind_ipi.vcpu = cpu;
1151                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1152                                                 &bind_ipi) != 0)
1153                         BUG();
1154                 evtchn = bind_ipi.port;
1155
1156                 /* Record the new mapping. */
1157                 evtchn_to_irq[evtchn] = irq;
1158                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
1159                 bind_evtchn_to_cpu(evtchn, cpu);
1160
1161                 /* Ready for use. */
1162                 unmask_evtchn(evtchn);
1163
1164         }
1165 }
1166
1167 /* Clear an irq's pending state, in preparation for polling on it */
1168 void xen_clear_irq_pending(int irq)
1169 {
1170         int evtchn = evtchn_from_irq(irq);
1171
1172         if (VALID_EVTCHN(evtchn))
1173                 clear_evtchn(evtchn);
1174 }
1175 EXPORT_SYMBOL(xen_clear_irq_pending);
1176 void xen_set_irq_pending(int irq)
1177 {
1178         int evtchn = evtchn_from_irq(irq);
1179
1180         if (VALID_EVTCHN(evtchn))
1181                 set_evtchn(evtchn);
1182 }
1183
1184 bool xen_test_irq_pending(int irq)
1185 {
1186         int evtchn = evtchn_from_irq(irq);
1187         bool ret = false;
1188
1189         if (VALID_EVTCHN(evtchn))
1190                 ret = test_evtchn(evtchn);
1191
1192         return ret;
1193 }
1194
1195 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1196  * the irq will be disabled so it won't deliver an interrupt. */
1197 void xen_poll_irq_timeout(int irq, u64 timeout)
1198 {
1199         evtchn_port_t evtchn = evtchn_from_irq(irq);
1200
1201         if (VALID_EVTCHN(evtchn)) {
1202                 struct sched_poll poll;
1203
1204                 poll.nr_ports = 1;
1205                 poll.timeout = timeout;
1206                 set_xen_guest_handle(poll.ports, &evtchn);
1207
1208                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1209                         BUG();
1210         }
1211 }
1212 EXPORT_SYMBOL(xen_poll_irq_timeout);
1213 /* Poll waiting for an irq to become pending.  In the usual case, the
1214  * irq will be disabled so it won't deliver an interrupt. */
1215 void xen_poll_irq(int irq)
1216 {
1217         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1218 }
1219
1220 void xen_irq_resume(void)
1221 {
1222         unsigned int cpu, irq, evtchn;
1223
1224         init_evtchn_cpu_bindings();
1225
1226         /* New event-channel space is not 'live' yet. */
1227         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1228                 mask_evtchn(evtchn);
1229
1230         /* No IRQ <-> event-channel mappings. */
1231         for (irq = 0; irq < nr_irqs; irq++)
1232                 irq_info[irq].evtchn = 0; /* zap event-channel binding */
1233
1234         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1235                 evtchn_to_irq[evtchn] = -1;
1236
1237         for_each_possible_cpu(cpu) {
1238                 restore_cpu_virqs(cpu);
1239                 restore_cpu_ipis(cpu);
1240         }
1241 }
1242
1243 static struct irq_chip xen_dynamic_chip __read_mostly = {
1244         .name           = "xen-dyn",
1245
1246         .disable        = disable_dynirq,
1247         .mask           = disable_dynirq,
1248         .unmask         = enable_dynirq,
1249
1250         .ack            = ack_dynirq,
1251         .set_affinity   = set_affinity_irq,
1252         .retrigger      = retrigger_dynirq,
1253 };
1254
1255 static struct irq_chip xen_pirq_chip __read_mostly = {
1256         .name           = "xen-pirq",
1257
1258         .startup        = startup_pirq,
1259         .shutdown       = shutdown_pirq,
1260
1261         .enable         = enable_pirq,
1262         .unmask         = enable_pirq,
1263
1264         .disable        = disable_pirq,
1265         .mask           = disable_pirq,
1266
1267         .ack            = ack_pirq,
1268         .end            = end_pirq,
1269
1270         .set_affinity   = set_affinity_irq,
1271
1272         .retrigger      = retrigger_dynirq,
1273 };
1274
1275 static struct irq_chip xen_percpu_chip __read_mostly = {
1276         .name           = "xen-percpu",
1277
1278         .disable        = disable_dynirq,
1279         .mask           = disable_dynirq,
1280         .unmask         = enable_dynirq,
1281
1282         .ack            = ack_dynirq,
1283 };
1284
1285 int xen_set_callback_via(uint64_t via)
1286 {
1287         struct xen_hvm_param a;
1288         a.domid = DOMID_SELF;
1289         a.index = HVM_PARAM_CALLBACK_IRQ;
1290         a.value = via;
1291         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1292 }
1293 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1294
1295 #ifdef CONFIG_XEN_PVHVM
1296 /* Vector callbacks are better than PCI interrupts to receive event
1297  * channel notifications because we can receive vector callbacks on any
1298  * vcpu and we don't need PCI support or APIC interactions. */
1299 void xen_callback_vector(void)
1300 {
1301         int rc;
1302         uint64_t callback_via;
1303         if (xen_have_vector_callback) {
1304                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1305                 rc = xen_set_callback_via(callback_via);
1306                 if (rc) {
1307                         printk(KERN_ERR "Request for Xen HVM callback vector"
1308                                         " failed.\n");
1309                         xen_have_vector_callback = 0;
1310                         return;
1311                 }
1312                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1313                                 "enabled\n");
1314                 /* in the restore case the vector has already been allocated */
1315                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1316                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1317         }
1318 }
1319 #else
1320 void xen_callback_vector(void) {}
1321 #endif
1322
1323 void __init xen_init_IRQ(void)
1324 {
1325         int i;
1326
1327         cpu_evtchn_mask_p = kcalloc(nr_cpu_ids, sizeof(struct cpu_evtchn_s),
1328                                     GFP_KERNEL);
1329         irq_info = kcalloc(nr_irqs, sizeof(*irq_info), GFP_KERNEL);
1330
1331         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1332                                     GFP_KERNEL);
1333         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1334                 evtchn_to_irq[i] = -1;
1335
1336         init_evtchn_cpu_bindings();
1337
1338         /* No event channels are 'live' right now. */
1339         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1340                 mask_evtchn(i);
1341
1342         if (xen_hvm_domain()) {
1343                 xen_callback_vector();
1344                 native_init_IRQ();
1345         } else {
1346                 irq_ctx_init(smp_processor_id());
1347         }
1348 }