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 received, it is mapped to an irq and sent
9 * through the normal interrupt processing path.
11 * There are four kinds of events which can be mapped to an event
14 * 1. Inter-domain notifications. This includes all the virtual
15 * device events, since they're driven by front-ends in another domain
17 * 2. VIRQs, typically used for timers. These are per-cpu events.
19 * 4. PIRQs - Hardware interrupts.
21 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
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 #include <linux/pci.h>
35 #include <asm/ptrace.h>
38 #include <asm/io_apic.h>
39 #include <asm/sync_bitops.h>
40 #include <asm/xen/pci.h>
41 #include <asm/xen/hypercall.h>
42 #include <asm/xen/hypervisor.h>
46 #include <xen/xen-ops.h>
47 #include <xen/events.h>
48 #include <xen/interface/xen.h>
49 #include <xen/interface/event_channel.h>
50 #include <xen/interface/hvm/hvm_op.h>
51 #include <xen/interface/hvm/params.h>
54 * This lock protects updates to the following mapping and reference-count
55 * arrays. The lock does not need to be acquired to read the mapping tables.
57 static DEFINE_SPINLOCK(irq_mapping_update_lock);
59 static LIST_HEAD(xen_irq_list_head);
61 /* IRQ <-> VIRQ mapping. */
62 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
64 /* IRQ <-> IPI mapping */
65 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
67 /* Interrupt types. */
77 * Packed IRQ information:
78 * type - enum xen_irq_type
79 * event channel - irq->event channel mapping
80 * cpu - cpu this event channel is bound to
81 * index - type-specific information:
82 * PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
83 * guest, or GSI (real passthrough IRQ) of the device.
90 struct list_head list;
91 enum xen_irq_type type; /* type */
93 unsigned short evtchn; /* event channel */
94 unsigned short cpu; /* cpu bound */
102 unsigned char vector;
108 #define PIRQ_NEEDS_EOI (1 << 0)
109 #define PIRQ_SHAREABLE (1 << 1)
111 static int *evtchn_to_irq;
113 static DEFINE_PER_CPU(unsigned long [NR_EVENT_CHANNELS/BITS_PER_LONG],
116 /* Xen will never allocate port zero for any purpose. */
117 #define VALID_EVTCHN(chn) ((chn) != 0)
119 static struct irq_chip xen_dynamic_chip;
120 static struct irq_chip xen_percpu_chip;
121 static struct irq_chip xen_pirq_chip;
122 static void enable_dynirq(struct irq_data *data);
123 static void disable_dynirq(struct irq_data *data);
125 /* Get info for IRQ */
126 static struct irq_info *info_for_irq(unsigned irq)
128 return irq_get_handler_data(irq);
131 /* Constructors for packed IRQ information. */
132 static void xen_irq_info_common_init(struct irq_info *info,
134 enum xen_irq_type type,
135 unsigned short evtchn,
139 BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
143 info->evtchn = evtchn;
146 evtchn_to_irq[evtchn] = irq;
149 static void xen_irq_info_evtchn_init(unsigned irq,
150 unsigned short evtchn)
152 struct irq_info *info = info_for_irq(irq);
154 xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
157 static void xen_irq_info_ipi_init(unsigned cpu,
159 unsigned short evtchn,
162 struct irq_info *info = info_for_irq(irq);
164 xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
168 per_cpu(ipi_to_irq, cpu)[ipi] = irq;
171 static void xen_irq_info_virq_init(unsigned cpu,
173 unsigned short evtchn,
176 struct irq_info *info = info_for_irq(irq);
178 xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
182 per_cpu(virq_to_irq, cpu)[virq] = irq;
185 static void xen_irq_info_pirq_init(unsigned irq,
186 unsigned short evtchn,
189 unsigned short vector,
193 struct irq_info *info = info_for_irq(irq);
195 xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
197 info->u.pirq.pirq = pirq;
198 info->u.pirq.gsi = gsi;
199 info->u.pirq.vector = vector;
200 info->u.pirq.domid = domid;
201 info->u.pirq.flags = flags;
205 * Accessors for packed IRQ information.
207 static unsigned int evtchn_from_irq(unsigned irq)
209 if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
212 return info_for_irq(irq)->evtchn;
215 unsigned irq_from_evtchn(unsigned int evtchn)
217 return evtchn_to_irq[evtchn];
219 EXPORT_SYMBOL_GPL(irq_from_evtchn);
221 static enum ipi_vector ipi_from_irq(unsigned irq)
223 struct irq_info *info = info_for_irq(irq);
225 BUG_ON(info == NULL);
226 BUG_ON(info->type != IRQT_IPI);
231 static unsigned virq_from_irq(unsigned irq)
233 struct irq_info *info = info_for_irq(irq);
235 BUG_ON(info == NULL);
236 BUG_ON(info->type != IRQT_VIRQ);
241 static unsigned pirq_from_irq(unsigned irq)
243 struct irq_info *info = info_for_irq(irq);
245 BUG_ON(info == NULL);
246 BUG_ON(info->type != IRQT_PIRQ);
248 return info->u.pirq.pirq;
251 static enum xen_irq_type type_from_irq(unsigned irq)
253 return info_for_irq(irq)->type;
256 static unsigned cpu_from_irq(unsigned irq)
258 return info_for_irq(irq)->cpu;
261 static unsigned int cpu_from_evtchn(unsigned int evtchn)
263 int irq = evtchn_to_irq[evtchn];
267 ret = cpu_from_irq(irq);
272 static bool pirq_needs_eoi(unsigned irq)
274 struct irq_info *info = info_for_irq(irq);
276 BUG_ON(info->type != IRQT_PIRQ);
278 return info->u.pirq.flags & PIRQ_NEEDS_EOI;
281 static inline unsigned long active_evtchns(unsigned int cpu,
282 struct shared_info *sh,
285 return (sh->evtchn_pending[idx] &
286 per_cpu(cpu_evtchn_mask, cpu)[idx] &
287 ~sh->evtchn_mask[idx]);
290 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
292 int irq = evtchn_to_irq[chn];
296 cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
299 clear_bit(chn, per_cpu(cpu_evtchn_mask, cpu_from_irq(irq)));
300 set_bit(chn, per_cpu(cpu_evtchn_mask, cpu));
302 info_for_irq(irq)->cpu = cpu;
305 static void init_evtchn_cpu_bindings(void)
309 struct irq_info *info;
311 /* By default all event channels notify CPU#0. */
312 list_for_each_entry(info, &xen_irq_list_head, list) {
313 struct irq_desc *desc = irq_to_desc(info->irq);
314 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
318 for_each_possible_cpu(i)
319 memset(per_cpu(cpu_evtchn_mask, i),
320 (i == 0) ? ~0 : 0, sizeof(*per_cpu(cpu_evtchn_mask, i)));
323 static inline void clear_evtchn(int port)
325 struct shared_info *s = HYPERVISOR_shared_info;
326 sync_clear_bit(port, &s->evtchn_pending[0]);
329 static inline void set_evtchn(int port)
331 struct shared_info *s = HYPERVISOR_shared_info;
332 sync_set_bit(port, &s->evtchn_pending[0]);
335 static inline int test_evtchn(int port)
337 struct shared_info *s = HYPERVISOR_shared_info;
338 return sync_test_bit(port, &s->evtchn_pending[0]);
343 * notify_remote_via_irq - send event to remote end of event channel via irq
344 * @irq: irq of event channel to send event to
346 * Unlike notify_remote_via_evtchn(), this is safe to use across
347 * save/restore. Notifications on a broken connection are silently
350 void notify_remote_via_irq(int irq)
352 int evtchn = evtchn_from_irq(irq);
354 if (VALID_EVTCHN(evtchn))
355 notify_remote_via_evtchn(evtchn);
357 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
359 static void mask_evtchn(int port)
361 struct shared_info *s = HYPERVISOR_shared_info;
362 sync_set_bit(port, &s->evtchn_mask[0]);
365 static void unmask_evtchn(int port)
367 struct shared_info *s = HYPERVISOR_shared_info;
368 unsigned int cpu = get_cpu();
370 BUG_ON(!irqs_disabled());
372 /* Slow path (hypercall) if this is a non-local port. */
373 if (unlikely(cpu != cpu_from_evtchn(port))) {
374 struct evtchn_unmask unmask = { .port = port };
375 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
377 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
379 sync_clear_bit(port, &s->evtchn_mask[0]);
382 * The following is basically the equivalent of
383 * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
384 * the interrupt edge' if the channel is masked.
386 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
387 !sync_test_and_set_bit(port / BITS_PER_LONG,
388 &vcpu_info->evtchn_pending_sel))
389 vcpu_info->evtchn_upcall_pending = 1;
395 static void xen_irq_init(unsigned irq)
397 struct irq_info *info;
399 struct irq_desc *desc = irq_to_desc(irq);
401 /* By default all event channels notify CPU#0. */
402 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
405 info = kzalloc(sizeof(*info), GFP_KERNEL);
407 panic("Unable to allocate metadata for IRQ%d\n", irq);
409 info->type = IRQT_UNBOUND;
411 irq_set_handler_data(irq, info);
413 list_add_tail(&info->list, &xen_irq_list_head);
416 static int __must_check xen_allocate_irq_dynamic(void)
421 #ifdef CONFIG_X86_IO_APIC
423 * For an HVM guest or domain 0 which see "real" (emulated or
424 * actual respectively) GSIs we allocate dynamic IRQs
425 * e.g. those corresponding to event channels or MSIs
426 * etc. from the range above those "real" GSIs to avoid
429 if (xen_initial_domain() || xen_hvm_domain())
430 first = get_nr_irqs_gsi();
433 irq = irq_alloc_desc_from(first, -1);
440 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
445 * A PV guest has no concept of a GSI (since it has no ACPI
446 * nor access to/knowledge of the physical APICs). Therefore
447 * all IRQs are dynamically allocated from the entire IRQ
450 if (xen_pv_domain() && !xen_initial_domain())
451 return xen_allocate_irq_dynamic();
453 /* Legacy IRQ descriptors are already allocated by the arch. */
454 if (gsi < NR_IRQS_LEGACY)
457 irq = irq_alloc_desc_at(gsi, -1);
464 static void xen_free_irq(unsigned irq)
466 struct irq_info *info = irq_get_handler_data(irq);
468 list_del(&info->list);
470 irq_set_handler_data(irq, NULL);
474 /* Legacy IRQ descriptors are managed by the arch. */
475 if (irq < NR_IRQS_LEGACY)
481 static void pirq_query_unmask(int irq)
483 struct physdev_irq_status_query irq_status;
484 struct irq_info *info = info_for_irq(irq);
486 BUG_ON(info->type != IRQT_PIRQ);
488 irq_status.irq = pirq_from_irq(irq);
489 if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
490 irq_status.flags = 0;
492 info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
493 if (irq_status.flags & XENIRQSTAT_needs_eoi)
494 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
497 static bool probing_irq(int irq)
499 struct irq_desc *desc = irq_to_desc(irq);
501 return desc && desc->action == NULL;
504 static void eoi_pirq(struct irq_data *data)
506 int evtchn = evtchn_from_irq(data->irq);
507 struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
512 if (VALID_EVTCHN(evtchn))
513 clear_evtchn(evtchn);
515 if (pirq_needs_eoi(data->irq)) {
516 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
521 static void mask_ack_pirq(struct irq_data *data)
523 disable_dynirq(data);
527 static unsigned int __startup_pirq(unsigned int irq)
529 struct evtchn_bind_pirq bind_pirq;
530 struct irq_info *info = info_for_irq(irq);
531 int evtchn = evtchn_from_irq(irq);
534 BUG_ON(info->type != IRQT_PIRQ);
536 if (VALID_EVTCHN(evtchn))
539 bind_pirq.pirq = pirq_from_irq(irq);
540 /* NB. We are happy to share unless we are probing. */
541 bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
542 BIND_PIRQ__WILL_SHARE : 0;
543 rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
545 if (!probing_irq(irq))
546 printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
550 evtchn = bind_pirq.port;
552 pirq_query_unmask(irq);
554 evtchn_to_irq[evtchn] = irq;
555 bind_evtchn_to_cpu(evtchn, 0);
556 info->evtchn = evtchn;
559 unmask_evtchn(evtchn);
560 eoi_pirq(irq_get_irq_data(irq));
565 static unsigned int startup_pirq(struct irq_data *data)
567 return __startup_pirq(data->irq);
570 static void shutdown_pirq(struct irq_data *data)
572 struct evtchn_close close;
573 unsigned int irq = data->irq;
574 struct irq_info *info = info_for_irq(irq);
575 int evtchn = evtchn_from_irq(irq);
577 BUG_ON(info->type != IRQT_PIRQ);
579 if (!VALID_EVTCHN(evtchn))
585 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
588 bind_evtchn_to_cpu(evtchn, 0);
589 evtchn_to_irq[evtchn] = -1;
593 static void enable_pirq(struct irq_data *data)
598 static void disable_pirq(struct irq_data *data)
600 disable_dynirq(data);
603 static int find_irq_by_gsi(unsigned gsi)
605 struct irq_info *info;
607 list_for_each_entry(info, &xen_irq_list_head, list) {
608 if (info->type != IRQT_PIRQ)
611 if (info->u.pirq.gsi == gsi)
619 * Do not make any assumptions regarding the relationship between the
620 * IRQ number returned here and the Xen pirq argument.
622 * Note: We don't assign an event channel until the irq actually started
623 * up. Return an existing irq if we've already got one for the gsi.
625 * Shareable implies level triggered, not shareable implies edge
628 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
629 unsigned pirq, int shareable, char *name)
632 struct physdev_irq irq_op;
634 spin_lock(&irq_mapping_update_lock);
636 irq = find_irq_by_gsi(gsi);
638 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
640 goto out; /* XXX need refcount? */
643 irq = xen_allocate_irq_gsi(gsi);
650 /* Only the privileged domain can do this. For non-priv, the pcifront
651 * driver provides a PCI bus that does the call to do exactly
652 * this in the priv domain. */
653 if (xen_initial_domain() &&
654 HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
660 xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
661 shareable ? PIRQ_SHAREABLE : 0);
663 pirq_query_unmask(irq);
664 /* We try to use the handler with the appropriate semantic for the
665 * type of interrupt: if the interrupt is an edge triggered
666 * interrupt we use handle_edge_irq.
668 * On the other hand if the interrupt is level triggered we use
669 * handle_fasteoi_irq like the native code does for this kind of
672 * Depending on the Xen version, pirq_needs_eoi might return true
673 * not only for level triggered interrupts but for edge triggered
674 * interrupts too. In any case Xen always honors the eoi mechanism,
675 * not injecting any more pirqs of the same kind if the first one
676 * hasn't received an eoi yet. Therefore using the fasteoi handler
677 * is the right choice either way.
680 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
681 handle_fasteoi_irq, name);
683 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
684 handle_edge_irq, name);
687 spin_unlock(&irq_mapping_update_lock);
692 #ifdef CONFIG_PCI_MSI
693 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
696 struct physdev_get_free_pirq op_get_free_pirq;
698 op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
699 rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
701 WARN_ONCE(rc == -ENOSYS,
702 "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
704 return rc ? -1 : op_get_free_pirq.pirq;
707 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
708 int pirq, int vector, const char *name,
713 spin_lock(&irq_mapping_update_lock);
715 irq = xen_allocate_irq_dynamic();
719 irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
722 xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
723 ret = irq_set_msi_desc(irq, msidesc);
727 spin_unlock(&irq_mapping_update_lock);
730 spin_unlock(&irq_mapping_update_lock);
736 int xen_destroy_irq(int irq)
738 struct irq_desc *desc;
739 struct physdev_unmap_pirq unmap_irq;
740 struct irq_info *info = info_for_irq(irq);
743 spin_lock(&irq_mapping_update_lock);
745 desc = irq_to_desc(irq);
749 if (xen_initial_domain()) {
750 unmap_irq.pirq = info->u.pirq.pirq;
751 unmap_irq.domid = info->u.pirq.domid;
752 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
753 /* If another domain quits without making the pci_disable_msix
754 * call, the Xen hypervisor takes care of freeing the PIRQs
755 * (free_domain_pirqs).
757 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
758 printk(KERN_INFO "domain %d does not have %d anymore\n",
759 info->u.pirq.domid, info->u.pirq.pirq);
761 printk(KERN_WARNING "unmap irq failed %d\n", rc);
769 spin_unlock(&irq_mapping_update_lock);
773 int xen_irq_from_pirq(unsigned pirq)
777 struct irq_info *info;
779 spin_lock(&irq_mapping_update_lock);
781 list_for_each_entry(info, &xen_irq_list_head, list) {
782 if (info == NULL || info->type != IRQT_PIRQ)
785 if (info->u.pirq.pirq == pirq)
790 spin_unlock(&irq_mapping_update_lock);
796 int xen_pirq_from_irq(unsigned irq)
798 return pirq_from_irq(irq);
800 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
801 int bind_evtchn_to_irq(unsigned int evtchn)
805 spin_lock(&irq_mapping_update_lock);
807 irq = evtchn_to_irq[evtchn];
810 irq = xen_allocate_irq_dynamic();
814 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
815 handle_edge_irq, "event");
817 xen_irq_info_evtchn_init(irq, evtchn);
821 spin_unlock(&irq_mapping_update_lock);
825 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
827 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
829 struct evtchn_bind_ipi bind_ipi;
832 spin_lock(&irq_mapping_update_lock);
834 irq = per_cpu(ipi_to_irq, cpu)[ipi];
837 irq = xen_allocate_irq_dynamic();
841 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
842 handle_percpu_irq, "ipi");
845 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
848 evtchn = bind_ipi.port;
850 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
852 bind_evtchn_to_cpu(evtchn, cpu);
856 spin_unlock(&irq_mapping_update_lock);
860 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
861 unsigned int remote_port)
863 struct evtchn_bind_interdomain bind_interdomain;
866 bind_interdomain.remote_dom = remote_domain;
867 bind_interdomain.remote_port = remote_port;
869 err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
872 return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
876 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
878 struct evtchn_bind_virq bind_virq;
881 spin_lock(&irq_mapping_update_lock);
883 irq = per_cpu(virq_to_irq, cpu)[virq];
886 irq = xen_allocate_irq_dynamic();
890 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
891 handle_percpu_irq, "virq");
893 bind_virq.virq = virq;
894 bind_virq.vcpu = cpu;
895 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
898 evtchn = bind_virq.port;
900 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
902 bind_evtchn_to_cpu(evtchn, cpu);
906 spin_unlock(&irq_mapping_update_lock);
911 static void unbind_from_irq(unsigned int irq)
913 struct evtchn_close close;
914 int evtchn = evtchn_from_irq(irq);
916 spin_lock(&irq_mapping_update_lock);
918 if (VALID_EVTCHN(evtchn)) {
920 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
923 switch (type_from_irq(irq)) {
925 per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
926 [virq_from_irq(irq)] = -1;
929 per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
930 [ipi_from_irq(irq)] = -1;
936 /* Closed ports are implicitly re-bound to VCPU0. */
937 bind_evtchn_to_cpu(evtchn, 0);
939 evtchn_to_irq[evtchn] = -1;
942 BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
946 spin_unlock(&irq_mapping_update_lock);
949 int bind_evtchn_to_irqhandler(unsigned int evtchn,
950 irq_handler_t handler,
951 unsigned long irqflags,
952 const char *devname, void *dev_id)
956 irq = bind_evtchn_to_irq(evtchn);
959 retval = request_irq(irq, handler, irqflags, devname, dev_id);
961 unbind_from_irq(irq);
967 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
969 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
970 unsigned int remote_port,
971 irq_handler_t handler,
972 unsigned long irqflags,
978 irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
982 retval = request_irq(irq, handler, irqflags, devname, dev_id);
984 unbind_from_irq(irq);
990 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
992 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
993 irq_handler_t handler,
994 unsigned long irqflags, const char *devname, void *dev_id)
998 irq = bind_virq_to_irq(virq, cpu);
1001 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1003 unbind_from_irq(irq);
1009 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1011 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1013 irq_handler_t handler,
1014 unsigned long irqflags,
1015 const char *devname,
1020 irq = bind_ipi_to_irq(ipi, cpu);
1024 irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME;
1025 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1027 unbind_from_irq(irq);
1034 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1036 free_irq(irq, dev_id);
1037 unbind_from_irq(irq);
1039 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1041 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1043 int irq = per_cpu(ipi_to_irq, cpu)[vector];
1045 notify_remote_via_irq(irq);
1048 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1050 struct shared_info *sh = HYPERVISOR_shared_info;
1051 int cpu = smp_processor_id();
1052 unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1054 unsigned long flags;
1055 static DEFINE_SPINLOCK(debug_lock);
1056 struct vcpu_info *v;
1058 spin_lock_irqsave(&debug_lock, flags);
1060 printk("\nvcpu %d\n ", cpu);
1062 for_each_online_cpu(i) {
1064 v = per_cpu(xen_vcpu, i);
1065 pending = (get_irq_regs() && i == cpu)
1066 ? xen_irqs_disabled(get_irq_regs())
1067 : v->evtchn_upcall_mask;
1068 printk("%d: masked=%d pending=%d event_sel %0*lx\n ", i,
1069 pending, v->evtchn_upcall_pending,
1070 (int)(sizeof(v->evtchn_pending_sel)*2),
1071 v->evtchn_pending_sel);
1073 v = per_cpu(xen_vcpu, cpu);
1075 printk("\npending:\n ");
1076 for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1077 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1078 sh->evtchn_pending[i],
1079 i % 8 == 0 ? "\n " : " ");
1080 printk("\nglobal mask:\n ");
1081 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1083 (int)(sizeof(sh->evtchn_mask[0])*2),
1085 i % 8 == 0 ? "\n " : " ");
1087 printk("\nglobally unmasked:\n ");
1088 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1089 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1090 sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1091 i % 8 == 0 ? "\n " : " ");
1093 printk("\nlocal cpu%d mask:\n ", cpu);
1094 for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1095 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1097 i % 8 == 0 ? "\n " : " ");
1099 printk("\nlocally unmasked:\n ");
1100 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1101 unsigned long pending = sh->evtchn_pending[i]
1102 & ~sh->evtchn_mask[i]
1104 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1105 pending, i % 8 == 0 ? "\n " : " ");
1108 printk("\npending list:\n");
1109 for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1110 if (sync_test_bit(i, sh->evtchn_pending)) {
1111 int word_idx = i / BITS_PER_LONG;
1112 printk(" %d: event %d -> irq %d%s%s%s\n",
1113 cpu_from_evtchn(i), i,
1115 sync_test_bit(word_idx, &v->evtchn_pending_sel)
1117 !sync_test_bit(i, sh->evtchn_mask)
1118 ? "" : " globally-masked",
1119 sync_test_bit(i, cpu_evtchn)
1120 ? "" : " locally-masked");
1124 spin_unlock_irqrestore(&debug_lock, flags);
1129 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1130 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1131 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1134 * Mask out the i least significant bits of w
1136 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1139 * Search the CPUs pending events bitmasks. For each one found, map
1140 * the event number to an irq, and feed it into do_IRQ() for
1143 * Xen uses a two-level bitmap to speed searching. The first level is
1144 * a bitset of words which contain pending event bits. The second
1145 * level is a bitset of pending events themselves.
1147 static void __xen_evtchn_do_upcall(void)
1149 int start_word_idx, start_bit_idx;
1150 int word_idx, bit_idx;
1152 int cpu = get_cpu();
1153 struct shared_info *s = HYPERVISOR_shared_info;
1154 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1158 unsigned long pending_words;
1160 vcpu_info->evtchn_upcall_pending = 0;
1162 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1165 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1166 /* Clear master flag /before/ clearing selector flag. */
1169 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1171 start_word_idx = __this_cpu_read(current_word_idx);
1172 start_bit_idx = __this_cpu_read(current_bit_idx);
1174 word_idx = start_word_idx;
1176 for (i = 0; pending_words != 0; i++) {
1177 unsigned long pending_bits;
1178 unsigned long words;
1180 words = MASK_LSBS(pending_words, word_idx);
1183 * If we masked out all events, wrap to beginning.
1190 word_idx = __ffs(words);
1192 pending_bits = active_evtchns(cpu, s, word_idx);
1193 bit_idx = 0; /* usually scan entire word from start */
1194 if (word_idx == start_word_idx) {
1195 /* We scan the starting word in two parts */
1197 /* 1st time: start in the middle */
1198 bit_idx = start_bit_idx;
1200 /* 2nd time: mask bits done already */
1201 bit_idx &= (1UL << start_bit_idx) - 1;
1207 struct irq_desc *desc;
1209 bits = MASK_LSBS(pending_bits, bit_idx);
1211 /* If we masked out all events, move on. */
1215 bit_idx = __ffs(bits);
1218 port = (word_idx * BITS_PER_LONG) + bit_idx;
1219 irq = evtchn_to_irq[port];
1222 desc = irq_to_desc(irq);
1224 generic_handle_irq_desc(irq, desc);
1227 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1229 /* Next caller starts at last processed + 1 */
1230 __this_cpu_write(current_word_idx,
1231 bit_idx ? word_idx :
1232 (word_idx+1) % BITS_PER_LONG);
1233 __this_cpu_write(current_bit_idx, bit_idx);
1234 } while (bit_idx != 0);
1236 /* Scan start_l1i twice; all others once. */
1237 if ((word_idx != start_word_idx) || (i != 0))
1238 pending_words &= ~(1UL << word_idx);
1240 word_idx = (word_idx + 1) % BITS_PER_LONG;
1243 BUG_ON(!irqs_disabled());
1245 count = __this_cpu_read(xed_nesting_count);
1246 __this_cpu_write(xed_nesting_count, 0);
1247 } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1254 void xen_evtchn_do_upcall(struct pt_regs *regs)
1256 struct pt_regs *old_regs = set_irq_regs(regs);
1261 __xen_evtchn_do_upcall();
1264 set_irq_regs(old_regs);
1267 void xen_hvm_evtchn_do_upcall(void)
1269 __xen_evtchn_do_upcall();
1271 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1273 /* Rebind a new event channel to an existing irq. */
1274 void rebind_evtchn_irq(int evtchn, int irq)
1276 struct irq_info *info = info_for_irq(irq);
1278 /* Make sure the irq is masked, since the new event channel
1279 will also be masked. */
1282 spin_lock(&irq_mapping_update_lock);
1284 /* After resume the irq<->evtchn mappings are all cleared out */
1285 BUG_ON(evtchn_to_irq[evtchn] != -1);
1286 /* Expect irq to have been bound before,
1287 so there should be a proper type */
1288 BUG_ON(info->type == IRQT_UNBOUND);
1290 xen_irq_info_evtchn_init(irq, evtchn);
1292 spin_unlock(&irq_mapping_update_lock);
1294 /* new event channels are always bound to cpu 0 */
1295 irq_set_affinity(irq, cpumask_of(0));
1297 /* Unmask the event channel. */
1301 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1302 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1304 struct evtchn_bind_vcpu bind_vcpu;
1305 int evtchn = evtchn_from_irq(irq);
1307 if (!VALID_EVTCHN(evtchn))
1311 * Events delivered via platform PCI interrupts are always
1312 * routed to vcpu 0 and hence cannot be rebound.
1314 if (xen_hvm_domain() && !xen_have_vector_callback)
1317 /* Send future instances of this interrupt to other vcpu. */
1318 bind_vcpu.port = evtchn;
1319 bind_vcpu.vcpu = tcpu;
1322 * If this fails, it usually just indicates that we're dealing with a
1323 * virq or IPI channel, which don't actually need to be rebound. Ignore
1324 * it, but don't do the xenlinux-level rebind in that case.
1326 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1327 bind_evtchn_to_cpu(evtchn, tcpu);
1332 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1335 unsigned tcpu = cpumask_first(dest);
1337 return rebind_irq_to_cpu(data->irq, tcpu);
1340 int resend_irq_on_evtchn(unsigned int irq)
1342 int masked, evtchn = evtchn_from_irq(irq);
1343 struct shared_info *s = HYPERVISOR_shared_info;
1345 if (!VALID_EVTCHN(evtchn))
1348 masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1349 sync_set_bit(evtchn, s->evtchn_pending);
1351 unmask_evtchn(evtchn);
1356 static void enable_dynirq(struct irq_data *data)
1358 int evtchn = evtchn_from_irq(data->irq);
1360 if (VALID_EVTCHN(evtchn))
1361 unmask_evtchn(evtchn);
1364 static void disable_dynirq(struct irq_data *data)
1366 int evtchn = evtchn_from_irq(data->irq);
1368 if (VALID_EVTCHN(evtchn))
1369 mask_evtchn(evtchn);
1372 static void ack_dynirq(struct irq_data *data)
1374 int evtchn = evtchn_from_irq(data->irq);
1378 if (VALID_EVTCHN(evtchn))
1379 clear_evtchn(evtchn);
1382 static void mask_ack_dynirq(struct irq_data *data)
1384 disable_dynirq(data);
1388 static int retrigger_dynirq(struct irq_data *data)
1390 int evtchn = evtchn_from_irq(data->irq);
1391 struct shared_info *sh = HYPERVISOR_shared_info;
1394 if (VALID_EVTCHN(evtchn)) {
1397 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1398 sync_set_bit(evtchn, sh->evtchn_pending);
1400 unmask_evtchn(evtchn);
1407 static void restore_pirqs(void)
1409 int pirq, rc, irq, gsi;
1410 struct physdev_map_pirq map_irq;
1411 struct irq_info *info;
1413 list_for_each_entry(info, &xen_irq_list_head, list) {
1414 if (info->type != IRQT_PIRQ)
1417 pirq = info->u.pirq.pirq;
1418 gsi = info->u.pirq.gsi;
1421 /* save/restore of PT devices doesn't work, so at this point the
1422 * only devices present are GSI based emulated devices */
1426 map_irq.domid = DOMID_SELF;
1427 map_irq.type = MAP_PIRQ_TYPE_GSI;
1428 map_irq.index = gsi;
1429 map_irq.pirq = pirq;
1431 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1433 printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1434 gsi, irq, pirq, rc);
1439 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1441 __startup_pirq(irq);
1445 static void restore_cpu_virqs(unsigned int cpu)
1447 struct evtchn_bind_virq bind_virq;
1448 int virq, irq, evtchn;
1450 for (virq = 0; virq < NR_VIRQS; virq++) {
1451 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1454 BUG_ON(virq_from_irq(irq) != virq);
1456 /* Get a new binding from Xen. */
1457 bind_virq.virq = virq;
1458 bind_virq.vcpu = cpu;
1459 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1462 evtchn = bind_virq.port;
1464 /* Record the new mapping. */
1465 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1466 bind_evtchn_to_cpu(evtchn, cpu);
1470 static void restore_cpu_ipis(unsigned int cpu)
1472 struct evtchn_bind_ipi bind_ipi;
1473 int ipi, irq, evtchn;
1475 for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1476 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1479 BUG_ON(ipi_from_irq(irq) != ipi);
1481 /* Get a new binding from Xen. */
1482 bind_ipi.vcpu = cpu;
1483 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1486 evtchn = bind_ipi.port;
1488 /* Record the new mapping. */
1489 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1490 bind_evtchn_to_cpu(evtchn, cpu);
1494 /* Clear an irq's pending state, in preparation for polling on it */
1495 void xen_clear_irq_pending(int irq)
1497 int evtchn = evtchn_from_irq(irq);
1499 if (VALID_EVTCHN(evtchn))
1500 clear_evtchn(evtchn);
1502 EXPORT_SYMBOL(xen_clear_irq_pending);
1503 void xen_set_irq_pending(int irq)
1505 int evtchn = evtchn_from_irq(irq);
1507 if (VALID_EVTCHN(evtchn))
1511 bool xen_test_irq_pending(int irq)
1513 int evtchn = evtchn_from_irq(irq);
1516 if (VALID_EVTCHN(evtchn))
1517 ret = test_evtchn(evtchn);
1522 /* Poll waiting for an irq to become pending with timeout. In the usual case,
1523 * the irq will be disabled so it won't deliver an interrupt. */
1524 void xen_poll_irq_timeout(int irq, u64 timeout)
1526 evtchn_port_t evtchn = evtchn_from_irq(irq);
1528 if (VALID_EVTCHN(evtchn)) {
1529 struct sched_poll poll;
1532 poll.timeout = timeout;
1533 set_xen_guest_handle(poll.ports, &evtchn);
1535 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1539 EXPORT_SYMBOL(xen_poll_irq_timeout);
1540 /* Poll waiting for an irq to become pending. In the usual case, the
1541 * irq will be disabled so it won't deliver an interrupt. */
1542 void xen_poll_irq(int irq)
1544 xen_poll_irq_timeout(irq, 0 /* no timeout */);
1547 /* Check whether the IRQ line is shared with other guests. */
1548 int xen_test_irq_shared(int irq)
1550 struct irq_info *info = info_for_irq(irq);
1551 struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1553 if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1555 return !(irq_status.flags & XENIRQSTAT_shared);
1557 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1559 void xen_irq_resume(void)
1561 unsigned int cpu, evtchn;
1562 struct irq_info *info;
1564 init_evtchn_cpu_bindings();
1566 /* New event-channel space is not 'live' yet. */
1567 for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1568 mask_evtchn(evtchn);
1570 /* No IRQ <-> event-channel mappings. */
1571 list_for_each_entry(info, &xen_irq_list_head, list)
1572 info->evtchn = 0; /* zap event-channel binding */
1574 for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1575 evtchn_to_irq[evtchn] = -1;
1577 for_each_possible_cpu(cpu) {
1578 restore_cpu_virqs(cpu);
1579 restore_cpu_ipis(cpu);
1585 static struct irq_chip xen_dynamic_chip __read_mostly = {
1588 .irq_disable = disable_dynirq,
1589 .irq_mask = disable_dynirq,
1590 .irq_unmask = enable_dynirq,
1592 .irq_ack = ack_dynirq,
1593 .irq_mask_ack = mask_ack_dynirq,
1595 .irq_set_affinity = set_affinity_irq,
1596 .irq_retrigger = retrigger_dynirq,
1599 static struct irq_chip xen_pirq_chip __read_mostly = {
1602 .irq_startup = startup_pirq,
1603 .irq_shutdown = shutdown_pirq,
1604 .irq_enable = enable_pirq,
1605 .irq_disable = disable_pirq,
1607 .irq_mask = disable_dynirq,
1608 .irq_unmask = enable_dynirq,
1610 .irq_ack = eoi_pirq,
1611 .irq_eoi = eoi_pirq,
1612 .irq_mask_ack = mask_ack_pirq,
1614 .irq_set_affinity = set_affinity_irq,
1616 .irq_retrigger = retrigger_dynirq,
1619 static struct irq_chip xen_percpu_chip __read_mostly = {
1620 .name = "xen-percpu",
1622 .irq_disable = disable_dynirq,
1623 .irq_mask = disable_dynirq,
1624 .irq_unmask = enable_dynirq,
1626 .irq_ack = ack_dynirq,
1629 int xen_set_callback_via(uint64_t via)
1631 struct xen_hvm_param a;
1632 a.domid = DOMID_SELF;
1633 a.index = HVM_PARAM_CALLBACK_IRQ;
1635 return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1637 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1639 #ifdef CONFIG_XEN_PVHVM
1640 /* Vector callbacks are better than PCI interrupts to receive event
1641 * channel notifications because we can receive vector callbacks on any
1642 * vcpu and we don't need PCI support or APIC interactions. */
1643 void xen_callback_vector(void)
1646 uint64_t callback_via;
1647 if (xen_have_vector_callback) {
1648 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1649 rc = xen_set_callback_via(callback_via);
1651 printk(KERN_ERR "Request for Xen HVM callback vector"
1653 xen_have_vector_callback = 0;
1656 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1658 /* in the restore case the vector has already been allocated */
1659 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1660 alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1664 void xen_callback_vector(void) {}
1667 void __init xen_init_IRQ(void)
1671 evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1673 for (i = 0; i < NR_EVENT_CHANNELS; i++)
1674 evtchn_to_irq[i] = -1;
1676 init_evtchn_cpu_bindings();
1678 /* No event channels are 'live' right now. */
1679 for (i = 0; i < NR_EVENT_CHANNELS; i++)
1682 if (xen_hvm_domain()) {
1683 xen_callback_vector();
1685 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1686 * __acpi_register_gsi can point at the right function */
1689 irq_ctx_init(smp_processor_id());
1690 if (xen_initial_domain())
1691 pci_xen_initial_domain();