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)
618 int xen_allocate_pirq_gsi(unsigned gsi)
624 * Do not make any assumptions regarding the relationship between the
625 * IRQ number returned here and the Xen pirq argument.
627 * Note: We don't assign an event channel until the irq actually started
628 * up. Return an existing irq if we've already got one for the gsi.
630 * Shareable implies level triggered, not shareable implies edge
633 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
634 unsigned pirq, int shareable, char *name)
637 struct physdev_irq irq_op;
639 spin_lock(&irq_mapping_update_lock);
641 irq = find_irq_by_gsi(gsi);
643 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
645 goto out; /* XXX need refcount? */
648 irq = xen_allocate_irq_gsi(gsi);
655 /* Only the privileged domain can do this. For non-priv, the pcifront
656 * driver provides a PCI bus that does the call to do exactly
657 * this in the priv domain. */
658 if (xen_initial_domain() &&
659 HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
665 xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
666 shareable ? PIRQ_SHAREABLE : 0);
668 pirq_query_unmask(irq);
669 /* We try to use the handler with the appropriate semantic for the
670 * type of interrupt: if the interrupt is an edge triggered
671 * interrupt we use handle_edge_irq.
673 * On the other hand if the interrupt is level triggered we use
674 * handle_fasteoi_irq like the native code does for this kind of
677 * Depending on the Xen version, pirq_needs_eoi might return true
678 * not only for level triggered interrupts but for edge triggered
679 * interrupts too. In any case Xen always honors the eoi mechanism,
680 * not injecting any more pirqs of the same kind if the first one
681 * hasn't received an eoi yet. Therefore using the fasteoi handler
682 * is the right choice either way.
685 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
686 handle_fasteoi_irq, name);
688 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
689 handle_edge_irq, name);
692 spin_unlock(&irq_mapping_update_lock);
697 #ifdef CONFIG_PCI_MSI
698 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
701 struct physdev_get_free_pirq op_get_free_pirq;
703 op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
704 rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
706 WARN_ONCE(rc == -ENOSYS,
707 "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
709 return rc ? -1 : op_get_free_pirq.pirq;
712 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
713 int pirq, int vector, const char *name,
718 spin_lock(&irq_mapping_update_lock);
720 irq = xen_allocate_irq_dynamic();
724 irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
727 xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
728 ret = irq_set_msi_desc(irq, msidesc);
732 spin_unlock(&irq_mapping_update_lock);
735 spin_unlock(&irq_mapping_update_lock);
741 int xen_destroy_irq(int irq)
743 struct irq_desc *desc;
744 struct physdev_unmap_pirq unmap_irq;
745 struct irq_info *info = info_for_irq(irq);
748 spin_lock(&irq_mapping_update_lock);
750 desc = irq_to_desc(irq);
754 if (xen_initial_domain()) {
755 unmap_irq.pirq = info->u.pirq.pirq;
756 unmap_irq.domid = info->u.pirq.domid;
757 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
758 /* If another domain quits without making the pci_disable_msix
759 * call, the Xen hypervisor takes care of freeing the PIRQs
760 * (free_domain_pirqs).
762 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
763 printk(KERN_INFO "domain %d does not have %d anymore\n",
764 info->u.pirq.domid, info->u.pirq.pirq);
766 printk(KERN_WARNING "unmap irq failed %d\n", rc);
774 spin_unlock(&irq_mapping_update_lock);
778 int xen_irq_from_pirq(unsigned pirq)
782 struct irq_info *info;
784 spin_lock(&irq_mapping_update_lock);
786 list_for_each_entry(info, &xen_irq_list_head, list) {
787 if (info == NULL || info->type != IRQT_PIRQ)
790 if (info->u.pirq.pirq == pirq)
795 spin_unlock(&irq_mapping_update_lock);
801 int xen_pirq_from_irq(unsigned irq)
803 return pirq_from_irq(irq);
805 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
806 int bind_evtchn_to_irq(unsigned int evtchn)
810 spin_lock(&irq_mapping_update_lock);
812 irq = evtchn_to_irq[evtchn];
815 irq = xen_allocate_irq_dynamic();
819 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
820 handle_edge_irq, "event");
822 xen_irq_info_evtchn_init(irq, evtchn);
826 spin_unlock(&irq_mapping_update_lock);
830 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
832 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
834 struct evtchn_bind_ipi bind_ipi;
837 spin_lock(&irq_mapping_update_lock);
839 irq = per_cpu(ipi_to_irq, cpu)[ipi];
842 irq = xen_allocate_irq_dynamic();
846 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
847 handle_percpu_irq, "ipi");
850 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
853 evtchn = bind_ipi.port;
855 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
857 bind_evtchn_to_cpu(evtchn, cpu);
861 spin_unlock(&irq_mapping_update_lock);
865 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
866 unsigned int remote_port)
868 struct evtchn_bind_interdomain bind_interdomain;
871 bind_interdomain.remote_dom = remote_domain;
872 bind_interdomain.remote_port = remote_port;
874 err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
877 return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
880 static int find_virq(unsigned int virq, unsigned int cpu)
882 struct evtchn_status status;
883 int port, rc = -ENOENT;
885 memset(&status, 0, sizeof(status));
886 for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
887 status.dom = DOMID_SELF;
889 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
892 if (status.status != EVTCHNSTAT_virq)
894 if (status.u.virq == virq && status.vcpu == cpu) {
902 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
904 struct evtchn_bind_virq bind_virq;
905 int evtchn, irq, ret;
907 spin_lock(&irq_mapping_update_lock);
909 irq = per_cpu(virq_to_irq, cpu)[virq];
912 irq = xen_allocate_irq_dynamic();
916 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
917 handle_percpu_irq, "virq");
919 bind_virq.virq = virq;
920 bind_virq.vcpu = cpu;
921 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
924 evtchn = bind_virq.port;
927 ret = find_virq(virq, cpu);
932 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
934 bind_evtchn_to_cpu(evtchn, cpu);
938 spin_unlock(&irq_mapping_update_lock);
943 static void unbind_from_irq(unsigned int irq)
945 struct evtchn_close close;
946 int evtchn = evtchn_from_irq(irq);
948 spin_lock(&irq_mapping_update_lock);
950 if (VALID_EVTCHN(evtchn)) {
952 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
955 switch (type_from_irq(irq)) {
957 per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
958 [virq_from_irq(irq)] = -1;
961 per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
962 [ipi_from_irq(irq)] = -1;
968 /* Closed ports are implicitly re-bound to VCPU0. */
969 bind_evtchn_to_cpu(evtchn, 0);
971 evtchn_to_irq[evtchn] = -1;
974 BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
978 spin_unlock(&irq_mapping_update_lock);
981 int bind_evtchn_to_irqhandler(unsigned int evtchn,
982 irq_handler_t handler,
983 unsigned long irqflags,
984 const char *devname, void *dev_id)
988 irq = bind_evtchn_to_irq(evtchn);
991 retval = request_irq(irq, handler, irqflags, devname, dev_id);
993 unbind_from_irq(irq);
999 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1001 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1002 unsigned int remote_port,
1003 irq_handler_t handler,
1004 unsigned long irqflags,
1005 const char *devname,
1010 irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1014 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1016 unbind_from_irq(irq);
1022 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1024 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1025 irq_handler_t handler,
1026 unsigned long irqflags, const char *devname, void *dev_id)
1030 irq = bind_virq_to_irq(virq, cpu);
1033 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1035 unbind_from_irq(irq);
1041 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1043 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1045 irq_handler_t handler,
1046 unsigned long irqflags,
1047 const char *devname,
1052 irq = bind_ipi_to_irq(ipi, cpu);
1056 irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME;
1057 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1059 unbind_from_irq(irq);
1066 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1068 free_irq(irq, dev_id);
1069 unbind_from_irq(irq);
1071 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1073 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1075 int irq = per_cpu(ipi_to_irq, cpu)[vector];
1077 notify_remote_via_irq(irq);
1080 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1082 struct shared_info *sh = HYPERVISOR_shared_info;
1083 int cpu = smp_processor_id();
1084 unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1086 unsigned long flags;
1087 static DEFINE_SPINLOCK(debug_lock);
1088 struct vcpu_info *v;
1090 spin_lock_irqsave(&debug_lock, flags);
1092 printk("\nvcpu %d\n ", cpu);
1094 for_each_online_cpu(i) {
1096 v = per_cpu(xen_vcpu, i);
1097 pending = (get_irq_regs() && i == cpu)
1098 ? xen_irqs_disabled(get_irq_regs())
1099 : v->evtchn_upcall_mask;
1100 printk("%d: masked=%d pending=%d event_sel %0*lx\n ", i,
1101 pending, v->evtchn_upcall_pending,
1102 (int)(sizeof(v->evtchn_pending_sel)*2),
1103 v->evtchn_pending_sel);
1105 v = per_cpu(xen_vcpu, cpu);
1107 printk("\npending:\n ");
1108 for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1109 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1110 sh->evtchn_pending[i],
1111 i % 8 == 0 ? "\n " : " ");
1112 printk("\nglobal mask:\n ");
1113 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1115 (int)(sizeof(sh->evtchn_mask[0])*2),
1117 i % 8 == 0 ? "\n " : " ");
1119 printk("\nglobally unmasked:\n ");
1120 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1121 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1122 sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1123 i % 8 == 0 ? "\n " : " ");
1125 printk("\nlocal cpu%d mask:\n ", cpu);
1126 for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1127 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1129 i % 8 == 0 ? "\n " : " ");
1131 printk("\nlocally unmasked:\n ");
1132 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1133 unsigned long pending = sh->evtchn_pending[i]
1134 & ~sh->evtchn_mask[i]
1136 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1137 pending, i % 8 == 0 ? "\n " : " ");
1140 printk("\npending list:\n");
1141 for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1142 if (sync_test_bit(i, sh->evtchn_pending)) {
1143 int word_idx = i / BITS_PER_LONG;
1144 printk(" %d: event %d -> irq %d%s%s%s\n",
1145 cpu_from_evtchn(i), i,
1147 sync_test_bit(word_idx, &v->evtchn_pending_sel)
1149 !sync_test_bit(i, sh->evtchn_mask)
1150 ? "" : " globally-masked",
1151 sync_test_bit(i, cpu_evtchn)
1152 ? "" : " locally-masked");
1156 spin_unlock_irqrestore(&debug_lock, flags);
1161 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1162 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1163 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1166 * Mask out the i least significant bits of w
1168 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1171 * Search the CPUs pending events bitmasks. For each one found, map
1172 * the event number to an irq, and feed it into do_IRQ() for
1175 * Xen uses a two-level bitmap to speed searching. The first level is
1176 * a bitset of words which contain pending event bits. The second
1177 * level is a bitset of pending events themselves.
1179 static void __xen_evtchn_do_upcall(void)
1181 int start_word_idx, start_bit_idx;
1182 int word_idx, bit_idx;
1184 int cpu = get_cpu();
1185 struct shared_info *s = HYPERVISOR_shared_info;
1186 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1190 unsigned long pending_words;
1192 vcpu_info->evtchn_upcall_pending = 0;
1194 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1197 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1198 /* Clear master flag /before/ clearing selector flag. */
1201 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1203 start_word_idx = __this_cpu_read(current_word_idx);
1204 start_bit_idx = __this_cpu_read(current_bit_idx);
1206 word_idx = start_word_idx;
1208 for (i = 0; pending_words != 0; i++) {
1209 unsigned long pending_bits;
1210 unsigned long words;
1212 words = MASK_LSBS(pending_words, word_idx);
1215 * If we masked out all events, wrap to beginning.
1222 word_idx = __ffs(words);
1224 pending_bits = active_evtchns(cpu, s, word_idx);
1225 bit_idx = 0; /* usually scan entire word from start */
1226 if (word_idx == start_word_idx) {
1227 /* We scan the starting word in two parts */
1229 /* 1st time: start in the middle */
1230 bit_idx = start_bit_idx;
1232 /* 2nd time: mask bits done already */
1233 bit_idx &= (1UL << start_bit_idx) - 1;
1239 struct irq_desc *desc;
1241 bits = MASK_LSBS(pending_bits, bit_idx);
1243 /* If we masked out all events, move on. */
1247 bit_idx = __ffs(bits);
1250 port = (word_idx * BITS_PER_LONG) + bit_idx;
1251 irq = evtchn_to_irq[port];
1254 desc = irq_to_desc(irq);
1256 generic_handle_irq_desc(irq, desc);
1259 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1261 /* Next caller starts at last processed + 1 */
1262 __this_cpu_write(current_word_idx,
1263 bit_idx ? word_idx :
1264 (word_idx+1) % BITS_PER_LONG);
1265 __this_cpu_write(current_bit_idx, bit_idx);
1266 } while (bit_idx != 0);
1268 /* Scan start_l1i twice; all others once. */
1269 if ((word_idx != start_word_idx) || (i != 0))
1270 pending_words &= ~(1UL << word_idx);
1272 word_idx = (word_idx + 1) % BITS_PER_LONG;
1275 BUG_ON(!irqs_disabled());
1277 count = __this_cpu_read(xed_nesting_count);
1278 __this_cpu_write(xed_nesting_count, 0);
1279 } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1286 void xen_evtchn_do_upcall(struct pt_regs *regs)
1288 struct pt_regs *old_regs = set_irq_regs(regs);
1293 __xen_evtchn_do_upcall();
1296 set_irq_regs(old_regs);
1299 void xen_hvm_evtchn_do_upcall(void)
1301 __xen_evtchn_do_upcall();
1303 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1305 /* Rebind a new event channel to an existing irq. */
1306 void rebind_evtchn_irq(int evtchn, int irq)
1308 struct irq_info *info = info_for_irq(irq);
1310 /* Make sure the irq is masked, since the new event channel
1311 will also be masked. */
1314 spin_lock(&irq_mapping_update_lock);
1316 /* After resume the irq<->evtchn mappings are all cleared out */
1317 BUG_ON(evtchn_to_irq[evtchn] != -1);
1318 /* Expect irq to have been bound before,
1319 so there should be a proper type */
1320 BUG_ON(info->type == IRQT_UNBOUND);
1322 xen_irq_info_evtchn_init(irq, evtchn);
1324 spin_unlock(&irq_mapping_update_lock);
1326 /* new event channels are always bound to cpu 0 */
1327 irq_set_affinity(irq, cpumask_of(0));
1329 /* Unmask the event channel. */
1333 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1334 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1336 struct evtchn_bind_vcpu bind_vcpu;
1337 int evtchn = evtchn_from_irq(irq);
1339 if (!VALID_EVTCHN(evtchn))
1343 * Events delivered via platform PCI interrupts are always
1344 * routed to vcpu 0 and hence cannot be rebound.
1346 if (xen_hvm_domain() && !xen_have_vector_callback)
1349 /* Send future instances of this interrupt to other vcpu. */
1350 bind_vcpu.port = evtchn;
1351 bind_vcpu.vcpu = tcpu;
1354 * If this fails, it usually just indicates that we're dealing with a
1355 * virq or IPI channel, which don't actually need to be rebound. Ignore
1356 * it, but don't do the xenlinux-level rebind in that case.
1358 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1359 bind_evtchn_to_cpu(evtchn, tcpu);
1364 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1367 unsigned tcpu = cpumask_first(dest);
1369 return rebind_irq_to_cpu(data->irq, tcpu);
1372 int resend_irq_on_evtchn(unsigned int irq)
1374 int masked, evtchn = evtchn_from_irq(irq);
1375 struct shared_info *s = HYPERVISOR_shared_info;
1377 if (!VALID_EVTCHN(evtchn))
1380 masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1381 sync_set_bit(evtchn, s->evtchn_pending);
1383 unmask_evtchn(evtchn);
1388 static void enable_dynirq(struct irq_data *data)
1390 int evtchn = evtchn_from_irq(data->irq);
1392 if (VALID_EVTCHN(evtchn))
1393 unmask_evtchn(evtchn);
1396 static void disable_dynirq(struct irq_data *data)
1398 int evtchn = evtchn_from_irq(data->irq);
1400 if (VALID_EVTCHN(evtchn))
1401 mask_evtchn(evtchn);
1404 static void ack_dynirq(struct irq_data *data)
1406 int evtchn = evtchn_from_irq(data->irq);
1410 if (VALID_EVTCHN(evtchn))
1411 clear_evtchn(evtchn);
1414 static void mask_ack_dynirq(struct irq_data *data)
1416 disable_dynirq(data);
1420 static int retrigger_dynirq(struct irq_data *data)
1422 int evtchn = evtchn_from_irq(data->irq);
1423 struct shared_info *sh = HYPERVISOR_shared_info;
1426 if (VALID_EVTCHN(evtchn)) {
1429 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1430 sync_set_bit(evtchn, sh->evtchn_pending);
1432 unmask_evtchn(evtchn);
1439 static void restore_pirqs(void)
1441 int pirq, rc, irq, gsi;
1442 struct physdev_map_pirq map_irq;
1443 struct irq_info *info;
1445 list_for_each_entry(info, &xen_irq_list_head, list) {
1446 if (info->type != IRQT_PIRQ)
1449 pirq = info->u.pirq.pirq;
1450 gsi = info->u.pirq.gsi;
1453 /* save/restore of PT devices doesn't work, so at this point the
1454 * only devices present are GSI based emulated devices */
1458 map_irq.domid = DOMID_SELF;
1459 map_irq.type = MAP_PIRQ_TYPE_GSI;
1460 map_irq.index = gsi;
1461 map_irq.pirq = pirq;
1463 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1465 printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1466 gsi, irq, pirq, rc);
1471 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1473 __startup_pirq(irq);
1477 static void restore_cpu_virqs(unsigned int cpu)
1479 struct evtchn_bind_virq bind_virq;
1480 int virq, irq, evtchn;
1482 for (virq = 0; virq < NR_VIRQS; virq++) {
1483 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1486 BUG_ON(virq_from_irq(irq) != virq);
1488 /* Get a new binding from Xen. */
1489 bind_virq.virq = virq;
1490 bind_virq.vcpu = cpu;
1491 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1494 evtchn = bind_virq.port;
1496 /* Record the new mapping. */
1497 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1498 bind_evtchn_to_cpu(evtchn, cpu);
1502 static void restore_cpu_ipis(unsigned int cpu)
1504 struct evtchn_bind_ipi bind_ipi;
1505 int ipi, irq, evtchn;
1507 for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1508 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1511 BUG_ON(ipi_from_irq(irq) != ipi);
1513 /* Get a new binding from Xen. */
1514 bind_ipi.vcpu = cpu;
1515 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1518 evtchn = bind_ipi.port;
1520 /* Record the new mapping. */
1521 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1522 bind_evtchn_to_cpu(evtchn, cpu);
1526 /* Clear an irq's pending state, in preparation for polling on it */
1527 void xen_clear_irq_pending(int irq)
1529 int evtchn = evtchn_from_irq(irq);
1531 if (VALID_EVTCHN(evtchn))
1532 clear_evtchn(evtchn);
1534 EXPORT_SYMBOL(xen_clear_irq_pending);
1535 void xen_set_irq_pending(int irq)
1537 int evtchn = evtchn_from_irq(irq);
1539 if (VALID_EVTCHN(evtchn))
1543 bool xen_test_irq_pending(int irq)
1545 int evtchn = evtchn_from_irq(irq);
1548 if (VALID_EVTCHN(evtchn))
1549 ret = test_evtchn(evtchn);
1554 /* Poll waiting for an irq to become pending with timeout. In the usual case,
1555 * the irq will be disabled so it won't deliver an interrupt. */
1556 void xen_poll_irq_timeout(int irq, u64 timeout)
1558 evtchn_port_t evtchn = evtchn_from_irq(irq);
1560 if (VALID_EVTCHN(evtchn)) {
1561 struct sched_poll poll;
1564 poll.timeout = timeout;
1565 set_xen_guest_handle(poll.ports, &evtchn);
1567 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1571 EXPORT_SYMBOL(xen_poll_irq_timeout);
1572 /* Poll waiting for an irq to become pending. In the usual case, the
1573 * irq will be disabled so it won't deliver an interrupt. */
1574 void xen_poll_irq(int irq)
1576 xen_poll_irq_timeout(irq, 0 /* no timeout */);
1579 /* Check whether the IRQ line is shared with other guests. */
1580 int xen_test_irq_shared(int irq)
1582 struct irq_info *info = info_for_irq(irq);
1583 struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1585 if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1587 return !(irq_status.flags & XENIRQSTAT_shared);
1589 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1591 void xen_irq_resume(void)
1593 unsigned int cpu, evtchn;
1594 struct irq_info *info;
1596 init_evtchn_cpu_bindings();
1598 /* New event-channel space is not 'live' yet. */
1599 for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1600 mask_evtchn(evtchn);
1602 /* No IRQ <-> event-channel mappings. */
1603 list_for_each_entry(info, &xen_irq_list_head, list)
1604 info->evtchn = 0; /* zap event-channel binding */
1606 for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1607 evtchn_to_irq[evtchn] = -1;
1609 for_each_possible_cpu(cpu) {
1610 restore_cpu_virqs(cpu);
1611 restore_cpu_ipis(cpu);
1617 static struct irq_chip xen_dynamic_chip __read_mostly = {
1620 .irq_disable = disable_dynirq,
1621 .irq_mask = disable_dynirq,
1622 .irq_unmask = enable_dynirq,
1624 .irq_ack = ack_dynirq,
1625 .irq_mask_ack = mask_ack_dynirq,
1627 .irq_set_affinity = set_affinity_irq,
1628 .irq_retrigger = retrigger_dynirq,
1631 static struct irq_chip xen_pirq_chip __read_mostly = {
1634 .irq_startup = startup_pirq,
1635 .irq_shutdown = shutdown_pirq,
1636 .irq_enable = enable_pirq,
1637 .irq_disable = disable_pirq,
1639 .irq_mask = disable_dynirq,
1640 .irq_unmask = enable_dynirq,
1642 .irq_ack = eoi_pirq,
1643 .irq_eoi = eoi_pirq,
1644 .irq_mask_ack = mask_ack_pirq,
1646 .irq_set_affinity = set_affinity_irq,
1648 .irq_retrigger = retrigger_dynirq,
1651 static struct irq_chip xen_percpu_chip __read_mostly = {
1652 .name = "xen-percpu",
1654 .irq_disable = disable_dynirq,
1655 .irq_mask = disable_dynirq,
1656 .irq_unmask = enable_dynirq,
1658 .irq_ack = ack_dynirq,
1661 int xen_set_callback_via(uint64_t via)
1663 struct xen_hvm_param a;
1664 a.domid = DOMID_SELF;
1665 a.index = HVM_PARAM_CALLBACK_IRQ;
1667 return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1669 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1671 #ifdef CONFIG_XEN_PVHVM
1672 /* Vector callbacks are better than PCI interrupts to receive event
1673 * channel notifications because we can receive vector callbacks on any
1674 * vcpu and we don't need PCI support or APIC interactions. */
1675 void xen_callback_vector(void)
1678 uint64_t callback_via;
1679 if (xen_have_vector_callback) {
1680 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1681 rc = xen_set_callback_via(callback_via);
1683 printk(KERN_ERR "Request for Xen HVM callback vector"
1685 xen_have_vector_callback = 0;
1688 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1690 /* in the restore case the vector has already been allocated */
1691 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1692 alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1696 void xen_callback_vector(void) {}
1699 void __init xen_init_IRQ(void)
1703 evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1705 for (i = 0; i < NR_EVENT_CHANNELS; i++)
1706 evtchn_to_irq[i] = -1;
1708 init_evtchn_cpu_bindings();
1710 /* No event channels are 'live' right now. */
1711 for (i = 0; i < NR_EVENT_CHANNELS; i++)
1714 if (xen_hvm_domain()) {
1715 xen_callback_vector();
1717 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1718 * __acpi_register_gsi can point at the right function */
1721 irq_ctx_init(smp_processor_id());
1722 if (xen_initial_domain())