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.
89 struct list_head list;
90 enum xen_irq_type type; /* type */
92 unsigned short evtchn; /* event channel */
93 unsigned short cpu; /* cpu bound */
101 unsigned char vector;
107 #define PIRQ_NEEDS_EOI (1 << 0)
108 #define PIRQ_SHAREABLE (1 << 1)
110 static int *evtchn_to_irq;
112 static DEFINE_PER_CPU(unsigned long [NR_EVENT_CHANNELS/BITS_PER_LONG],
115 /* Xen will never allocate port zero for any purpose. */
116 #define VALID_EVTCHN(chn) ((chn) != 0)
118 static struct irq_chip xen_dynamic_chip;
119 static struct irq_chip xen_percpu_chip;
120 static struct irq_chip xen_pirq_chip;
121 static void enable_dynirq(struct irq_data *data);
122 static void disable_dynirq(struct irq_data *data);
124 /* Get info for IRQ */
125 static struct irq_info *info_for_irq(unsigned irq)
127 return irq_get_handler_data(irq);
130 /* Constructors for packed IRQ information. */
131 static void xen_irq_info_common_init(struct irq_info *info,
133 enum xen_irq_type type,
134 unsigned short evtchn,
138 BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
142 info->evtchn = evtchn;
145 evtchn_to_irq[evtchn] = irq;
148 static void xen_irq_info_evtchn_init(unsigned irq,
149 unsigned short evtchn)
151 struct irq_info *info = info_for_irq(irq);
153 xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
156 static void xen_irq_info_ipi_init(unsigned cpu,
158 unsigned short evtchn,
161 struct irq_info *info = info_for_irq(irq);
163 xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
167 per_cpu(ipi_to_irq, cpu)[ipi] = irq;
170 static void xen_irq_info_virq_init(unsigned cpu,
172 unsigned short evtchn,
175 struct irq_info *info = info_for_irq(irq);
177 xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
181 per_cpu(virq_to_irq, cpu)[virq] = irq;
184 static void xen_irq_info_pirq_init(unsigned irq,
185 unsigned short evtchn,
188 unsigned short vector,
192 struct irq_info *info = info_for_irq(irq);
194 xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
196 info->u.pirq.pirq = pirq;
197 info->u.pirq.gsi = gsi;
198 info->u.pirq.vector = vector;
199 info->u.pirq.domid = domid;
200 info->u.pirq.flags = flags;
204 * Accessors for packed IRQ information.
206 static unsigned int evtchn_from_irq(unsigned irq)
208 if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
211 return info_for_irq(irq)->evtchn;
214 unsigned irq_from_evtchn(unsigned int evtchn)
216 return evtchn_to_irq[evtchn];
218 EXPORT_SYMBOL_GPL(irq_from_evtchn);
220 static enum ipi_vector ipi_from_irq(unsigned irq)
222 struct irq_info *info = info_for_irq(irq);
224 BUG_ON(info == NULL);
225 BUG_ON(info->type != IRQT_IPI);
230 static unsigned virq_from_irq(unsigned irq)
232 struct irq_info *info = info_for_irq(irq);
234 BUG_ON(info == NULL);
235 BUG_ON(info->type != IRQT_VIRQ);
240 static unsigned pirq_from_irq(unsigned irq)
242 struct irq_info *info = info_for_irq(irq);
244 BUG_ON(info == NULL);
245 BUG_ON(info->type != IRQT_PIRQ);
247 return info->u.pirq.pirq;
250 static enum xen_irq_type type_from_irq(unsigned irq)
252 return info_for_irq(irq)->type;
255 static unsigned cpu_from_irq(unsigned irq)
257 return info_for_irq(irq)->cpu;
260 static unsigned int cpu_from_evtchn(unsigned int evtchn)
262 int irq = evtchn_to_irq[evtchn];
266 ret = cpu_from_irq(irq);
271 static bool pirq_needs_eoi(unsigned irq)
273 struct irq_info *info = info_for_irq(irq);
275 BUG_ON(info->type != IRQT_PIRQ);
277 return info->u.pirq.flags & PIRQ_NEEDS_EOI;
280 static inline unsigned long active_evtchns(unsigned int cpu,
281 struct shared_info *sh,
284 return sh->evtchn_pending[idx] &
285 per_cpu(cpu_evtchn_mask, cpu)[idx] &
286 ~sh->evtchn_mask[idx];
289 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
291 int irq = evtchn_to_irq[chn];
295 cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
298 clear_bit(chn, per_cpu(cpu_evtchn_mask, cpu_from_irq(irq)));
299 set_bit(chn, per_cpu(cpu_evtchn_mask, cpu));
301 info_for_irq(irq)->cpu = cpu;
304 static void init_evtchn_cpu_bindings(void)
308 struct irq_info *info;
310 /* By default all event channels notify CPU#0. */
311 list_for_each_entry(info, &xen_irq_list_head, list) {
312 struct irq_desc *desc = irq_to_desc(info->irq);
313 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
317 for_each_possible_cpu(i)
318 memset(per_cpu(cpu_evtchn_mask, i),
319 (i == 0) ? ~0 : 0, sizeof(*per_cpu(cpu_evtchn_mask, i)));
322 static inline void clear_evtchn(int port)
324 struct shared_info *s = HYPERVISOR_shared_info;
325 sync_clear_bit(port, &s->evtchn_pending[0]);
328 static inline void set_evtchn(int port)
330 struct shared_info *s = HYPERVISOR_shared_info;
331 sync_set_bit(port, &s->evtchn_pending[0]);
334 static inline int test_evtchn(int port)
336 struct shared_info *s = HYPERVISOR_shared_info;
337 return sync_test_bit(port, &s->evtchn_pending[0]);
342 * notify_remote_via_irq - send event to remote end of event channel via irq
343 * @irq: irq of event channel to send event to
345 * Unlike notify_remote_via_evtchn(), this is safe to use across
346 * save/restore. Notifications on a broken connection are silently
349 void notify_remote_via_irq(int irq)
351 int evtchn = evtchn_from_irq(irq);
353 if (VALID_EVTCHN(evtchn))
354 notify_remote_via_evtchn(evtchn);
356 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
358 static void mask_evtchn(int port)
360 struct shared_info *s = HYPERVISOR_shared_info;
361 sync_set_bit(port, &s->evtchn_mask[0]);
364 static void unmask_evtchn(int port)
366 struct shared_info *s = HYPERVISOR_shared_info;
367 unsigned int cpu = get_cpu();
369 BUG_ON(!irqs_disabled());
371 /* Slow path (hypercall) if this is a non-local port. */
372 if (unlikely(cpu != cpu_from_evtchn(port))) {
373 struct evtchn_unmask unmask = { .port = port };
374 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
376 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
378 sync_clear_bit(port, &s->evtchn_mask[0]);
381 * The following is basically the equivalent of
382 * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
383 * the interrupt edge' if the channel is masked.
385 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
386 !sync_test_and_set_bit(port / BITS_PER_LONG,
387 &vcpu_info->evtchn_pending_sel))
388 vcpu_info->evtchn_upcall_pending = 1;
394 static void xen_irq_init(unsigned irq)
396 struct irq_info *info;
398 struct irq_desc *desc = irq_to_desc(irq);
400 /* By default all event channels notify CPU#0. */
401 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
404 info = kzalloc(sizeof(*info), GFP_KERNEL);
406 panic("Unable to allocate metadata for IRQ%d\n", irq);
408 info->type = IRQT_UNBOUND;
410 irq_set_handler_data(irq, info);
412 list_add_tail(&info->list, &xen_irq_list_head);
415 static int __must_check xen_allocate_irq_dynamic(void)
420 #ifdef CONFIG_X86_IO_APIC
422 * For an HVM guest or domain 0 which see "real" (emulated or
423 * actual respectively) GSIs we allocate dynamic IRQs
424 * e.g. those corresponding to event channels or MSIs
425 * etc. from the range above those "real" GSIs to avoid
428 if (xen_initial_domain() || xen_hvm_domain())
429 first = get_nr_irqs_gsi();
432 irq = irq_alloc_desc_from(first, -1);
439 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
444 * A PV guest has no concept of a GSI (since it has no ACPI
445 * nor access to/knowledge of the physical APICs). Therefore
446 * all IRQs are dynamically allocated from the entire IRQ
449 if (xen_pv_domain() && !xen_initial_domain())
450 return xen_allocate_irq_dynamic();
452 /* Legacy IRQ descriptors are already allocated by the arch. */
453 if (gsi < NR_IRQS_LEGACY)
456 irq = irq_alloc_desc_at(gsi, -1);
463 static void xen_free_irq(unsigned irq)
465 struct irq_info *info = irq_get_handler_data(irq);
467 list_del(&info->list);
469 irq_set_handler_data(irq, NULL);
473 /* Legacy IRQ descriptors are managed by the arch. */
474 if (irq < NR_IRQS_LEGACY)
480 static void pirq_query_unmask(int irq)
482 struct physdev_irq_status_query irq_status;
483 struct irq_info *info = info_for_irq(irq);
485 BUG_ON(info->type != IRQT_PIRQ);
487 irq_status.irq = pirq_from_irq(irq);
488 if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
489 irq_status.flags = 0;
491 info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
492 if (irq_status.flags & XENIRQSTAT_needs_eoi)
493 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
496 static bool probing_irq(int irq)
498 struct irq_desc *desc = irq_to_desc(irq);
500 return desc && desc->action == NULL;
503 static void eoi_pirq(struct irq_data *data)
505 int evtchn = evtchn_from_irq(data->irq);
506 struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
511 if (VALID_EVTCHN(evtchn))
512 clear_evtchn(evtchn);
514 if (pirq_needs_eoi(data->irq)) {
515 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
520 static void mask_ack_pirq(struct irq_data *data)
522 disable_dynirq(data);
526 static unsigned int __startup_pirq(unsigned int irq)
528 struct evtchn_bind_pirq bind_pirq;
529 struct irq_info *info = info_for_irq(irq);
530 int evtchn = evtchn_from_irq(irq);
533 BUG_ON(info->type != IRQT_PIRQ);
535 if (VALID_EVTCHN(evtchn))
538 bind_pirq.pirq = pirq_from_irq(irq);
539 /* NB. We are happy to share unless we are probing. */
540 bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
541 BIND_PIRQ__WILL_SHARE : 0;
542 rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
544 if (!probing_irq(irq))
545 printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
549 evtchn = bind_pirq.port;
551 pirq_query_unmask(irq);
553 evtchn_to_irq[evtchn] = irq;
554 bind_evtchn_to_cpu(evtchn, 0);
555 info->evtchn = evtchn;
558 unmask_evtchn(evtchn);
559 eoi_pirq(irq_get_irq_data(irq));
564 static unsigned int startup_pirq(struct irq_data *data)
566 return __startup_pirq(data->irq);
569 static void shutdown_pirq(struct irq_data *data)
571 struct evtchn_close close;
572 unsigned int irq = data->irq;
573 struct irq_info *info = info_for_irq(irq);
574 int evtchn = evtchn_from_irq(irq);
576 BUG_ON(info->type != IRQT_PIRQ);
578 if (!VALID_EVTCHN(evtchn))
584 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
587 bind_evtchn_to_cpu(evtchn, 0);
588 evtchn_to_irq[evtchn] = -1;
592 static void enable_pirq(struct irq_data *data)
597 static void disable_pirq(struct irq_data *data)
599 disable_dynirq(data);
602 static int find_irq_by_gsi(unsigned gsi)
604 struct irq_info *info;
606 list_for_each_entry(info, &xen_irq_list_head, list) {
607 if (info->type != IRQT_PIRQ)
610 if (info->u.pirq.gsi == gsi)
618 * Do not make any assumptions regarding the relationship between the
619 * IRQ number returned here and the Xen pirq argument.
621 * Note: We don't assign an event channel until the irq actually started
622 * up. Return an existing irq if we've already got one for the gsi.
624 * Shareable implies level triggered, not shareable implies edge
627 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
628 unsigned pirq, int shareable, char *name)
631 struct physdev_irq irq_op;
633 spin_lock(&irq_mapping_update_lock);
635 irq = find_irq_by_gsi(gsi);
637 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
639 goto out; /* XXX need refcount? */
642 irq = xen_allocate_irq_gsi(gsi);
649 /* Only the privileged domain can do this. For non-priv, the pcifront
650 * driver provides a PCI bus that does the call to do exactly
651 * this in the priv domain. */
652 if (xen_initial_domain() &&
653 HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
659 xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
660 shareable ? PIRQ_SHAREABLE : 0);
662 pirq_query_unmask(irq);
663 /* We try to use the handler with the appropriate semantic for the
664 * type of interrupt: if the interrupt is an edge triggered
665 * interrupt we use handle_edge_irq.
667 * On the other hand if the interrupt is level triggered we use
668 * handle_fasteoi_irq like the native code does for this kind of
671 * Depending on the Xen version, pirq_needs_eoi might return true
672 * not only for level triggered interrupts but for edge triggered
673 * interrupts too. In any case Xen always honors the eoi mechanism,
674 * not injecting any more pirqs of the same kind if the first one
675 * hasn't received an eoi yet. Therefore using the fasteoi handler
676 * is the right choice either way.
679 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
680 handle_fasteoi_irq, name);
682 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
683 handle_edge_irq, name);
686 spin_unlock(&irq_mapping_update_lock);
691 #ifdef CONFIG_PCI_MSI
692 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
695 struct physdev_get_free_pirq op_get_free_pirq;
697 op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
698 rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
700 WARN_ONCE(rc == -ENOSYS,
701 "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
703 return rc ? -1 : op_get_free_pirq.pirq;
706 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
707 int pirq, int vector, const char *name,
712 spin_lock(&irq_mapping_update_lock);
714 irq = xen_allocate_irq_dynamic();
718 irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
721 xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
722 ret = irq_set_msi_desc(irq, msidesc);
726 spin_unlock(&irq_mapping_update_lock);
729 spin_unlock(&irq_mapping_update_lock);
735 int xen_destroy_irq(int irq)
737 struct irq_desc *desc;
738 struct physdev_unmap_pirq unmap_irq;
739 struct irq_info *info = info_for_irq(irq);
742 spin_lock(&irq_mapping_update_lock);
744 desc = irq_to_desc(irq);
748 if (xen_initial_domain()) {
749 unmap_irq.pirq = info->u.pirq.pirq;
750 unmap_irq.domid = info->u.pirq.domid;
751 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
752 /* If another domain quits without making the pci_disable_msix
753 * call, the Xen hypervisor takes care of freeing the PIRQs
754 * (free_domain_pirqs).
756 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
757 printk(KERN_INFO "domain %d does not have %d anymore\n",
758 info->u.pirq.domid, info->u.pirq.pirq);
760 printk(KERN_WARNING "unmap irq failed %d\n", rc);
768 spin_unlock(&irq_mapping_update_lock);
772 int xen_irq_from_pirq(unsigned pirq)
776 struct irq_info *info;
778 spin_lock(&irq_mapping_update_lock);
780 list_for_each_entry(info, &xen_irq_list_head, list) {
781 if (info == NULL || info->type != IRQT_PIRQ)
784 if (info->u.pirq.pirq == pirq)
789 spin_unlock(&irq_mapping_update_lock);
795 int xen_pirq_from_irq(unsigned irq)
797 return pirq_from_irq(irq);
799 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
800 int bind_evtchn_to_irq(unsigned int evtchn)
804 spin_lock(&irq_mapping_update_lock);
806 irq = evtchn_to_irq[evtchn];
809 irq = xen_allocate_irq_dynamic();
813 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
814 handle_edge_irq, "event");
816 xen_irq_info_evtchn_init(irq, evtchn);
820 spin_unlock(&irq_mapping_update_lock);
824 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
826 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
828 struct evtchn_bind_ipi bind_ipi;
831 spin_lock(&irq_mapping_update_lock);
833 irq = per_cpu(ipi_to_irq, cpu)[ipi];
836 irq = xen_allocate_irq_dynamic();
840 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
841 handle_percpu_irq, "ipi");
844 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
847 evtchn = bind_ipi.port;
849 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
851 bind_evtchn_to_cpu(evtchn, cpu);
855 spin_unlock(&irq_mapping_update_lock);
859 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
860 unsigned int remote_port)
862 struct evtchn_bind_interdomain bind_interdomain;
865 bind_interdomain.remote_dom = remote_domain;
866 bind_interdomain.remote_port = remote_port;
868 err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
871 return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
874 static int find_virq(unsigned int virq, unsigned int cpu)
876 struct evtchn_status status;
877 int port, rc = -ENOENT;
879 memset(&status, 0, sizeof(status));
880 for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
881 status.dom = DOMID_SELF;
883 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
886 if (status.status != EVTCHNSTAT_virq)
888 if (status.u.virq == virq && status.vcpu == cpu) {
896 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
898 struct evtchn_bind_virq bind_virq;
899 int evtchn, irq, ret;
901 spin_lock(&irq_mapping_update_lock);
903 irq = per_cpu(virq_to_irq, cpu)[virq];
906 irq = xen_allocate_irq_dynamic();
910 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
911 handle_percpu_irq, "virq");
913 bind_virq.virq = virq;
914 bind_virq.vcpu = cpu;
915 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
918 evtchn = bind_virq.port;
921 ret = find_virq(virq, cpu);
926 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
928 bind_evtchn_to_cpu(evtchn, cpu);
932 spin_unlock(&irq_mapping_update_lock);
937 static void unbind_from_irq(unsigned int irq)
939 struct evtchn_close close;
940 int evtchn = evtchn_from_irq(irq);
942 spin_lock(&irq_mapping_update_lock);
944 if (VALID_EVTCHN(evtchn)) {
946 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
949 switch (type_from_irq(irq)) {
951 per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
952 [virq_from_irq(irq)] = -1;
955 per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
956 [ipi_from_irq(irq)] = -1;
962 /* Closed ports are implicitly re-bound to VCPU0. */
963 bind_evtchn_to_cpu(evtchn, 0);
965 evtchn_to_irq[evtchn] = -1;
968 BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
972 spin_unlock(&irq_mapping_update_lock);
975 int bind_evtchn_to_irqhandler(unsigned int evtchn,
976 irq_handler_t handler,
977 unsigned long irqflags,
978 const char *devname, void *dev_id)
982 irq = bind_evtchn_to_irq(evtchn);
985 retval = request_irq(irq, handler, irqflags, devname, dev_id);
987 unbind_from_irq(irq);
993 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
995 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
996 unsigned int remote_port,
997 irq_handler_t handler,
998 unsigned long irqflags,
1004 irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1008 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1010 unbind_from_irq(irq);
1016 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1018 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1019 irq_handler_t handler,
1020 unsigned long irqflags, const char *devname, void *dev_id)
1024 irq = bind_virq_to_irq(virq, cpu);
1027 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1029 unbind_from_irq(irq);
1035 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1037 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1039 irq_handler_t handler,
1040 unsigned long irqflags,
1041 const char *devname,
1046 irq = bind_ipi_to_irq(ipi, cpu);
1050 irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME;
1051 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1053 unbind_from_irq(irq);
1060 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1062 free_irq(irq, dev_id);
1063 unbind_from_irq(irq);
1065 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1067 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1069 int irq = per_cpu(ipi_to_irq, cpu)[vector];
1071 notify_remote_via_irq(irq);
1074 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1076 struct shared_info *sh = HYPERVISOR_shared_info;
1077 int cpu = smp_processor_id();
1078 unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1080 unsigned long flags;
1081 static DEFINE_SPINLOCK(debug_lock);
1082 struct vcpu_info *v;
1084 spin_lock_irqsave(&debug_lock, flags);
1086 printk("\nvcpu %d\n ", cpu);
1088 for_each_online_cpu(i) {
1090 v = per_cpu(xen_vcpu, i);
1091 pending = (get_irq_regs() && i == cpu)
1092 ? xen_irqs_disabled(get_irq_regs())
1093 : v->evtchn_upcall_mask;
1094 printk("%d: masked=%d pending=%d event_sel %0*lx\n ", i,
1095 pending, v->evtchn_upcall_pending,
1096 (int)(sizeof(v->evtchn_pending_sel)*2),
1097 v->evtchn_pending_sel);
1099 v = per_cpu(xen_vcpu, cpu);
1101 printk("\npending:\n ");
1102 for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1103 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1104 sh->evtchn_pending[i],
1105 i % 8 == 0 ? "\n " : " ");
1106 printk("\nglobal mask:\n ");
1107 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1109 (int)(sizeof(sh->evtchn_mask[0])*2),
1111 i % 8 == 0 ? "\n " : " ");
1113 printk("\nglobally unmasked:\n ");
1114 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1115 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1116 sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1117 i % 8 == 0 ? "\n " : " ");
1119 printk("\nlocal cpu%d mask:\n ", cpu);
1120 for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1121 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1123 i % 8 == 0 ? "\n " : " ");
1125 printk("\nlocally unmasked:\n ");
1126 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1127 unsigned long pending = sh->evtchn_pending[i]
1128 & ~sh->evtchn_mask[i]
1130 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1131 pending, i % 8 == 0 ? "\n " : " ");
1134 printk("\npending list:\n");
1135 for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1136 if (sync_test_bit(i, sh->evtchn_pending)) {
1137 int word_idx = i / BITS_PER_LONG;
1138 printk(" %d: event %d -> irq %d%s%s%s\n",
1139 cpu_from_evtchn(i), i,
1141 sync_test_bit(word_idx, &v->evtchn_pending_sel)
1143 !sync_test_bit(i, sh->evtchn_mask)
1144 ? "" : " globally-masked",
1145 sync_test_bit(i, cpu_evtchn)
1146 ? "" : " locally-masked");
1150 spin_unlock_irqrestore(&debug_lock, flags);
1155 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1156 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1157 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1160 * Mask out the i least significant bits of w
1162 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1165 * Search the CPUs pending events bitmasks. For each one found, map
1166 * the event number to an irq, and feed it into do_IRQ() for
1169 * Xen uses a two-level bitmap to speed searching. The first level is
1170 * a bitset of words which contain pending event bits. The second
1171 * level is a bitset of pending events themselves.
1173 static void __xen_evtchn_do_upcall(void)
1175 int start_word_idx, start_bit_idx;
1176 int word_idx, bit_idx;
1178 int cpu = get_cpu();
1179 struct shared_info *s = HYPERVISOR_shared_info;
1180 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1184 unsigned long pending_words;
1186 vcpu_info->evtchn_upcall_pending = 0;
1188 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1191 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1192 /* Clear master flag /before/ clearing selector flag. */
1195 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1197 start_word_idx = __this_cpu_read(current_word_idx);
1198 start_bit_idx = __this_cpu_read(current_bit_idx);
1200 word_idx = start_word_idx;
1202 for (i = 0; pending_words != 0; i++) {
1203 unsigned long pending_bits;
1204 unsigned long words;
1206 words = MASK_LSBS(pending_words, word_idx);
1209 * If we masked out all events, wrap to beginning.
1216 word_idx = __ffs(words);
1218 pending_bits = active_evtchns(cpu, s, word_idx);
1219 bit_idx = 0; /* usually scan entire word from start */
1220 if (word_idx == start_word_idx) {
1221 /* We scan the starting word in two parts */
1223 /* 1st time: start in the middle */
1224 bit_idx = start_bit_idx;
1226 /* 2nd time: mask bits done already */
1227 bit_idx &= (1UL << start_bit_idx) - 1;
1233 struct irq_desc *desc;
1235 bits = MASK_LSBS(pending_bits, bit_idx);
1237 /* If we masked out all events, move on. */
1241 bit_idx = __ffs(bits);
1244 port = (word_idx * BITS_PER_LONG) + bit_idx;
1245 irq = evtchn_to_irq[port];
1248 desc = irq_to_desc(irq);
1250 generic_handle_irq_desc(irq, desc);
1253 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1255 /* Next caller starts at last processed + 1 */
1256 __this_cpu_write(current_word_idx,
1257 bit_idx ? word_idx :
1258 (word_idx+1) % BITS_PER_LONG);
1259 __this_cpu_write(current_bit_idx, bit_idx);
1260 } while (bit_idx != 0);
1262 /* Scan start_l1i twice; all others once. */
1263 if ((word_idx != start_word_idx) || (i != 0))
1264 pending_words &= ~(1UL << word_idx);
1266 word_idx = (word_idx + 1) % BITS_PER_LONG;
1269 BUG_ON(!irqs_disabled());
1271 count = __this_cpu_read(xed_nesting_count);
1272 __this_cpu_write(xed_nesting_count, 0);
1273 } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1280 void xen_evtchn_do_upcall(struct pt_regs *regs)
1282 struct pt_regs *old_regs = set_irq_regs(regs);
1287 __xen_evtchn_do_upcall();
1290 set_irq_regs(old_regs);
1293 void xen_hvm_evtchn_do_upcall(void)
1295 __xen_evtchn_do_upcall();
1297 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1299 /* Rebind a new event channel to an existing irq. */
1300 void rebind_evtchn_irq(int evtchn, int irq)
1302 struct irq_info *info = info_for_irq(irq);
1304 /* Make sure the irq is masked, since the new event channel
1305 will also be masked. */
1308 spin_lock(&irq_mapping_update_lock);
1310 /* After resume the irq<->evtchn mappings are all cleared out */
1311 BUG_ON(evtchn_to_irq[evtchn] != -1);
1312 /* Expect irq to have been bound before,
1313 so there should be a proper type */
1314 BUG_ON(info->type == IRQT_UNBOUND);
1316 xen_irq_info_evtchn_init(irq, evtchn);
1318 spin_unlock(&irq_mapping_update_lock);
1320 /* new event channels are always bound to cpu 0 */
1321 irq_set_affinity(irq, cpumask_of(0));
1323 /* Unmask the event channel. */
1327 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1328 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1330 struct evtchn_bind_vcpu bind_vcpu;
1331 int evtchn = evtchn_from_irq(irq);
1333 if (!VALID_EVTCHN(evtchn))
1337 * Events delivered via platform PCI interrupts are always
1338 * routed to vcpu 0 and hence cannot be rebound.
1340 if (xen_hvm_domain() && !xen_have_vector_callback)
1343 /* Send future instances of this interrupt to other vcpu. */
1344 bind_vcpu.port = evtchn;
1345 bind_vcpu.vcpu = tcpu;
1348 * If this fails, it usually just indicates that we're dealing with a
1349 * virq or IPI channel, which don't actually need to be rebound. Ignore
1350 * it, but don't do the xenlinux-level rebind in that case.
1352 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1353 bind_evtchn_to_cpu(evtchn, tcpu);
1358 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1361 unsigned tcpu = cpumask_first(dest);
1363 return rebind_irq_to_cpu(data->irq, tcpu);
1366 int resend_irq_on_evtchn(unsigned int irq)
1368 int masked, evtchn = evtchn_from_irq(irq);
1369 struct shared_info *s = HYPERVISOR_shared_info;
1371 if (!VALID_EVTCHN(evtchn))
1374 masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1375 sync_set_bit(evtchn, s->evtchn_pending);
1377 unmask_evtchn(evtchn);
1382 static void enable_dynirq(struct irq_data *data)
1384 int evtchn = evtchn_from_irq(data->irq);
1386 if (VALID_EVTCHN(evtchn))
1387 unmask_evtchn(evtchn);
1390 static void disable_dynirq(struct irq_data *data)
1392 int evtchn = evtchn_from_irq(data->irq);
1394 if (VALID_EVTCHN(evtchn))
1395 mask_evtchn(evtchn);
1398 static void ack_dynirq(struct irq_data *data)
1400 int evtchn = evtchn_from_irq(data->irq);
1404 if (VALID_EVTCHN(evtchn))
1405 clear_evtchn(evtchn);
1408 static void mask_ack_dynirq(struct irq_data *data)
1410 disable_dynirq(data);
1414 static int retrigger_dynirq(struct irq_data *data)
1416 int evtchn = evtchn_from_irq(data->irq);
1417 struct shared_info *sh = HYPERVISOR_shared_info;
1420 if (VALID_EVTCHN(evtchn)) {
1423 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1424 sync_set_bit(evtchn, sh->evtchn_pending);
1426 unmask_evtchn(evtchn);
1433 static void restore_pirqs(void)
1435 int pirq, rc, irq, gsi;
1436 struct physdev_map_pirq map_irq;
1437 struct irq_info *info;
1439 list_for_each_entry(info, &xen_irq_list_head, list) {
1440 if (info->type != IRQT_PIRQ)
1443 pirq = info->u.pirq.pirq;
1444 gsi = info->u.pirq.gsi;
1447 /* save/restore of PT devices doesn't work, so at this point the
1448 * only devices present are GSI based emulated devices */
1452 map_irq.domid = DOMID_SELF;
1453 map_irq.type = MAP_PIRQ_TYPE_GSI;
1454 map_irq.index = gsi;
1455 map_irq.pirq = pirq;
1457 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1459 printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1460 gsi, irq, pirq, rc);
1465 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1467 __startup_pirq(irq);
1471 static void restore_cpu_virqs(unsigned int cpu)
1473 struct evtchn_bind_virq bind_virq;
1474 int virq, irq, evtchn;
1476 for (virq = 0; virq < NR_VIRQS; virq++) {
1477 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1480 BUG_ON(virq_from_irq(irq) != virq);
1482 /* Get a new binding from Xen. */
1483 bind_virq.virq = virq;
1484 bind_virq.vcpu = cpu;
1485 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1488 evtchn = bind_virq.port;
1490 /* Record the new mapping. */
1491 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1492 bind_evtchn_to_cpu(evtchn, cpu);
1496 static void restore_cpu_ipis(unsigned int cpu)
1498 struct evtchn_bind_ipi bind_ipi;
1499 int ipi, irq, evtchn;
1501 for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1502 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1505 BUG_ON(ipi_from_irq(irq) != ipi);
1507 /* Get a new binding from Xen. */
1508 bind_ipi.vcpu = cpu;
1509 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1512 evtchn = bind_ipi.port;
1514 /* Record the new mapping. */
1515 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1516 bind_evtchn_to_cpu(evtchn, cpu);
1520 /* Clear an irq's pending state, in preparation for polling on it */
1521 void xen_clear_irq_pending(int irq)
1523 int evtchn = evtchn_from_irq(irq);
1525 if (VALID_EVTCHN(evtchn))
1526 clear_evtchn(evtchn);
1528 EXPORT_SYMBOL(xen_clear_irq_pending);
1529 void xen_set_irq_pending(int irq)
1531 int evtchn = evtchn_from_irq(irq);
1533 if (VALID_EVTCHN(evtchn))
1537 bool xen_test_irq_pending(int irq)
1539 int evtchn = evtchn_from_irq(irq);
1542 if (VALID_EVTCHN(evtchn))
1543 ret = test_evtchn(evtchn);
1548 /* Poll waiting for an irq to become pending with timeout. In the usual case,
1549 * the irq will be disabled so it won't deliver an interrupt. */
1550 void xen_poll_irq_timeout(int irq, u64 timeout)
1552 evtchn_port_t evtchn = evtchn_from_irq(irq);
1554 if (VALID_EVTCHN(evtchn)) {
1555 struct sched_poll poll;
1558 poll.timeout = timeout;
1559 set_xen_guest_handle(poll.ports, &evtchn);
1561 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1565 EXPORT_SYMBOL(xen_poll_irq_timeout);
1566 /* Poll waiting for an irq to become pending. In the usual case, the
1567 * irq will be disabled so it won't deliver an interrupt. */
1568 void xen_poll_irq(int irq)
1570 xen_poll_irq_timeout(irq, 0 /* no timeout */);
1573 /* Check whether the IRQ line is shared with other guests. */
1574 int xen_test_irq_shared(int irq)
1576 struct irq_info *info = info_for_irq(irq);
1577 struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1579 if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1581 return !(irq_status.flags & XENIRQSTAT_shared);
1583 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1585 void xen_irq_resume(void)
1587 unsigned int cpu, evtchn;
1588 struct irq_info *info;
1590 init_evtchn_cpu_bindings();
1592 /* New event-channel space is not 'live' yet. */
1593 for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1594 mask_evtchn(evtchn);
1596 /* No IRQ <-> event-channel mappings. */
1597 list_for_each_entry(info, &xen_irq_list_head, list)
1598 info->evtchn = 0; /* zap event-channel binding */
1600 for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1601 evtchn_to_irq[evtchn] = -1;
1603 for_each_possible_cpu(cpu) {
1604 restore_cpu_virqs(cpu);
1605 restore_cpu_ipis(cpu);
1611 static struct irq_chip xen_dynamic_chip __read_mostly = {
1614 .irq_disable = disable_dynirq,
1615 .irq_mask = disable_dynirq,
1616 .irq_unmask = enable_dynirq,
1618 .irq_ack = ack_dynirq,
1619 .irq_mask_ack = mask_ack_dynirq,
1621 .irq_set_affinity = set_affinity_irq,
1622 .irq_retrigger = retrigger_dynirq,
1625 static struct irq_chip xen_pirq_chip __read_mostly = {
1628 .irq_startup = startup_pirq,
1629 .irq_shutdown = shutdown_pirq,
1630 .irq_enable = enable_pirq,
1631 .irq_disable = disable_pirq,
1633 .irq_mask = disable_dynirq,
1634 .irq_unmask = enable_dynirq,
1636 .irq_ack = eoi_pirq,
1637 .irq_eoi = eoi_pirq,
1638 .irq_mask_ack = mask_ack_pirq,
1640 .irq_set_affinity = set_affinity_irq,
1642 .irq_retrigger = retrigger_dynirq,
1645 static struct irq_chip xen_percpu_chip __read_mostly = {
1646 .name = "xen-percpu",
1648 .irq_disable = disable_dynirq,
1649 .irq_mask = disable_dynirq,
1650 .irq_unmask = enable_dynirq,
1652 .irq_ack = ack_dynirq,
1655 int xen_set_callback_via(uint64_t via)
1657 struct xen_hvm_param a;
1658 a.domid = DOMID_SELF;
1659 a.index = HVM_PARAM_CALLBACK_IRQ;
1661 return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1663 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1665 #ifdef CONFIG_XEN_PVHVM
1666 /* Vector callbacks are better than PCI interrupts to receive event
1667 * channel notifications because we can receive vector callbacks on any
1668 * vcpu and we don't need PCI support or APIC interactions. */
1669 void xen_callback_vector(void)
1672 uint64_t callback_via;
1673 if (xen_have_vector_callback) {
1674 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1675 rc = xen_set_callback_via(callback_via);
1677 printk(KERN_ERR "Request for Xen HVM callback vector"
1679 xen_have_vector_callback = 0;
1682 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1684 /* in the restore case the vector has already been allocated */
1685 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1686 alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1690 void xen_callback_vector(void) {}
1693 void __init xen_init_IRQ(void)
1697 evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1699 for (i = 0; i < NR_EVENT_CHANNELS; i++)
1700 evtchn_to_irq[i] = -1;
1702 init_evtchn_cpu_bindings();
1704 /* No event channels are 'live' right now. */
1705 for (i = 0; i < NR_EVENT_CHANNELS; i++)
1708 if (xen_hvm_domain()) {
1709 xen_callback_vector();
1711 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1712 * __acpi_register_gsi can point at the right function */
1715 irq_ctx_init(smp_processor_id());
1716 if (xen_initial_domain())
1717 pci_xen_initial_domain();