4 * Xen models interrupts with abstract event channels. Because each
5 * domain gets 1024 event channels, but NR_IRQ is not that large, we
6 * must dynamically map irqs<->event channels. The event channels
7 * interface with the rest of the kernel by defining a xen interrupt
8 * chip. When an event is recieved, it is mapped to an irq and sent
9 * through the normal interrupt processing path.
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. Hardware interrupts. Not supported at present.
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>
30 #include <asm/ptrace.h>
32 #include <asm/sync_bitops.h>
33 #include <asm/xen/hypercall.h>
34 #include <asm/xen/hypervisor.h>
36 #include <xen/xen-ops.h>
37 #include <xen/events.h>
38 #include <xen/interface/xen.h>
39 #include <xen/interface/event_channel.h>
42 * This lock protects updates to the following mapping and reference-count
43 * arrays. The lock does not need to be acquired to read the mapping tables.
45 static DEFINE_SPINLOCK(irq_mapping_update_lock);
47 /* IRQ <-> VIRQ mapping. */
48 static DEFINE_PER_CPU(int, virq_to_irq[NR_VIRQS]) = {[0 ... NR_VIRQS-1] = -1};
50 /* IRQ <-> IPI mapping */
51 static DEFINE_PER_CPU(int, ipi_to_irq[XEN_NR_IPIS]) = {[0 ... XEN_NR_IPIS-1] = -1};
53 /* Packed IRQ information: binding type, sub-type index, and event channel. */
56 unsigned short evtchn;
61 static struct packed_irq irq_info[NR_IRQS];
72 /* Convenient shorthand for packed representation of an unbound IRQ. */
73 #define IRQ_UNBOUND mk_irq_info(IRQT_UNBOUND, 0, 0)
75 static int evtchn_to_irq[NR_EVENT_CHANNELS] = {
76 [0 ... NR_EVENT_CHANNELS-1] = -1
79 unsigned long bits[NR_EVENT_CHANNELS/BITS_PER_LONG];
81 static struct cpu_evtchn_s *cpu_evtchn_mask_p;
82 static inline unsigned long *cpu_evtchn_mask(int cpu)
84 return cpu_evtchn_mask_p[cpu].bits;
86 static u8 cpu_evtchn[NR_EVENT_CHANNELS];
88 /* Reference counts for bindings to IRQs. */
89 static int irq_bindcount[NR_IRQS];
91 /* Xen will never allocate port zero for any purpose. */
92 #define VALID_EVTCHN(chn) ((chn) != 0)
94 static struct irq_chip xen_dynamic_chip;
96 /* Constructor for packed IRQ information. */
97 static inline struct packed_irq mk_irq_info(u32 type, u32 index, u32 evtchn)
99 return (struct packed_irq) { evtchn, index, type };
103 * Accessors for packed IRQ information.
105 static inline unsigned int evtchn_from_irq(int irq)
107 return irq_info[irq].evtchn;
110 static inline unsigned int index_from_irq(int irq)
112 return irq_info[irq].index;
115 static inline unsigned int type_from_irq(int irq)
117 return irq_info[irq].type;
120 static inline unsigned long active_evtchns(unsigned int cpu,
121 struct shared_info *sh,
124 return (sh->evtchn_pending[idx] &
125 cpu_evtchn_mask(cpu)[idx] &
126 ~sh->evtchn_mask[idx]);
129 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
131 int irq = evtchn_to_irq[chn];
135 cpumask_copy(irq_to_desc(irq)->affinity, cpumask_of(cpu));
138 __clear_bit(chn, cpu_evtchn_mask(cpu_evtchn[chn]));
139 __set_bit(chn, cpu_evtchn_mask(cpu));
141 cpu_evtchn[chn] = cpu;
144 static void init_evtchn_cpu_bindings(void)
147 struct irq_desc *desc;
150 /* By default all event channels notify CPU#0. */
151 for_each_irq_desc(i, desc) {
152 cpumask_copy(desc->affinity, cpumask_of(0));
156 memset(cpu_evtchn, 0, sizeof(cpu_evtchn));
157 memset(cpu_evtchn_mask(0), ~0, sizeof(cpu_evtchn_mask(0)));
160 static inline unsigned int cpu_from_evtchn(unsigned int evtchn)
162 return cpu_evtchn[evtchn];
165 static inline void clear_evtchn(int port)
167 struct shared_info *s = HYPERVISOR_shared_info;
168 sync_clear_bit(port, &s->evtchn_pending[0]);
171 static inline void set_evtchn(int port)
173 struct shared_info *s = HYPERVISOR_shared_info;
174 sync_set_bit(port, &s->evtchn_pending[0]);
177 static inline int test_evtchn(int port)
179 struct shared_info *s = HYPERVISOR_shared_info;
180 return sync_test_bit(port, &s->evtchn_pending[0]);
185 * notify_remote_via_irq - send event to remote end of event channel via irq
186 * @irq: irq of event channel to send event to
188 * Unlike notify_remote_via_evtchn(), this is safe to use across
189 * save/restore. Notifications on a broken connection are silently
192 void notify_remote_via_irq(int irq)
194 int evtchn = evtchn_from_irq(irq);
196 if (VALID_EVTCHN(evtchn))
197 notify_remote_via_evtchn(evtchn);
199 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
201 static void mask_evtchn(int port)
203 struct shared_info *s = HYPERVISOR_shared_info;
204 sync_set_bit(port, &s->evtchn_mask[0]);
207 static void unmask_evtchn(int port)
209 struct shared_info *s = HYPERVISOR_shared_info;
210 unsigned int cpu = get_cpu();
212 BUG_ON(!irqs_disabled());
214 /* Slow path (hypercall) if this is a non-local port. */
215 if (unlikely(cpu != cpu_from_evtchn(port))) {
216 struct evtchn_unmask unmask = { .port = port };
217 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
219 struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
221 sync_clear_bit(port, &s->evtchn_mask[0]);
224 * The following is basically the equivalent of
225 * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
226 * the interrupt edge' if the channel is masked.
228 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
229 !sync_test_and_set_bit(port / BITS_PER_LONG,
230 &vcpu_info->evtchn_pending_sel))
231 vcpu_info->evtchn_upcall_pending = 1;
237 static int find_unbound_irq(void)
240 struct irq_desc *desc;
242 /* Only allocate from dynirq range */
243 for (irq = 0; irq < nr_irqs; irq++)
244 if (irq_bindcount[irq] == 0)
248 panic("No available IRQ to bind to: increase nr_irqs!\n");
250 desc = irq_to_desc_alloc_cpu(irq, 0);
251 if (WARN_ON(desc == NULL))
257 int bind_evtchn_to_irq(unsigned int evtchn)
261 spin_lock(&irq_mapping_update_lock);
263 irq = evtchn_to_irq[evtchn];
266 irq = find_unbound_irq();
268 dynamic_irq_init(irq);
269 set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
270 handle_level_irq, "event");
272 evtchn_to_irq[evtchn] = irq;
273 irq_info[irq] = mk_irq_info(IRQT_EVTCHN, 0, evtchn);
276 irq_bindcount[irq]++;
278 spin_unlock(&irq_mapping_update_lock);
282 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
284 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
286 struct evtchn_bind_ipi bind_ipi;
289 spin_lock(&irq_mapping_update_lock);
291 irq = per_cpu(ipi_to_irq, cpu)[ipi];
293 irq = find_unbound_irq();
297 dynamic_irq_init(irq);
298 set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
299 handle_level_irq, "ipi");
302 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
305 evtchn = bind_ipi.port;
307 evtchn_to_irq[evtchn] = irq;
308 irq_info[irq] = mk_irq_info(IRQT_IPI, ipi, evtchn);
310 per_cpu(ipi_to_irq, cpu)[ipi] = irq;
312 bind_evtchn_to_cpu(evtchn, cpu);
315 irq_bindcount[irq]++;
318 spin_unlock(&irq_mapping_update_lock);
323 static int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
325 struct evtchn_bind_virq bind_virq;
328 spin_lock(&irq_mapping_update_lock);
330 irq = per_cpu(virq_to_irq, cpu)[virq];
333 bind_virq.virq = virq;
334 bind_virq.vcpu = cpu;
335 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
338 evtchn = bind_virq.port;
340 irq = find_unbound_irq();
342 dynamic_irq_init(irq);
343 set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
344 handle_level_irq, "virq");
346 evtchn_to_irq[evtchn] = irq;
347 irq_info[irq] = mk_irq_info(IRQT_VIRQ, virq, evtchn);
349 per_cpu(virq_to_irq, cpu)[virq] = irq;
351 bind_evtchn_to_cpu(evtchn, cpu);
354 irq_bindcount[irq]++;
356 spin_unlock(&irq_mapping_update_lock);
361 static void unbind_from_irq(unsigned int irq)
363 struct evtchn_close close;
364 int evtchn = evtchn_from_irq(irq);
366 spin_lock(&irq_mapping_update_lock);
368 if ((--irq_bindcount[irq] == 0) && VALID_EVTCHN(evtchn)) {
370 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
373 switch (type_from_irq(irq)) {
375 per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
376 [index_from_irq(irq)] = -1;
379 per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
380 [index_from_irq(irq)] = -1;
386 /* Closed ports are implicitly re-bound to VCPU0. */
387 bind_evtchn_to_cpu(evtchn, 0);
389 evtchn_to_irq[evtchn] = -1;
390 irq_info[irq] = IRQ_UNBOUND;
392 dynamic_irq_cleanup(irq);
395 spin_unlock(&irq_mapping_update_lock);
398 int bind_evtchn_to_irqhandler(unsigned int evtchn,
399 irq_handler_t handler,
400 unsigned long irqflags,
401 const char *devname, void *dev_id)
406 irq = bind_evtchn_to_irq(evtchn);
407 retval = request_irq(irq, handler, irqflags, devname, dev_id);
409 unbind_from_irq(irq);
415 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
417 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
418 irq_handler_t handler,
419 unsigned long irqflags, const char *devname, void *dev_id)
424 irq = bind_virq_to_irq(virq, cpu);
425 retval = request_irq(irq, handler, irqflags, devname, dev_id);
427 unbind_from_irq(irq);
433 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
435 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
437 irq_handler_t handler,
438 unsigned long irqflags,
444 irq = bind_ipi_to_irq(ipi, cpu);
448 retval = request_irq(irq, handler, irqflags, devname, dev_id);
450 unbind_from_irq(irq);
457 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
459 free_irq(irq, dev_id);
460 unbind_from_irq(irq);
462 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
464 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
466 int irq = per_cpu(ipi_to_irq, cpu)[vector];
468 notify_remote_via_irq(irq);
471 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
473 struct shared_info *sh = HYPERVISOR_shared_info;
474 int cpu = smp_processor_id();
477 static DEFINE_SPINLOCK(debug_lock);
479 spin_lock_irqsave(&debug_lock, flags);
481 printk("vcpu %d\n ", cpu);
483 for_each_online_cpu(i) {
484 struct vcpu_info *v = per_cpu(xen_vcpu, i);
485 printk("%d: masked=%d pending=%d event_sel %08lx\n ", i,
486 (get_irq_regs() && i == cpu) ? xen_irqs_disabled(get_irq_regs()) : v->evtchn_upcall_mask,
487 v->evtchn_upcall_pending,
488 v->evtchn_pending_sel);
490 printk("pending:\n ");
491 for(i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
492 printk("%08lx%s", sh->evtchn_pending[i],
493 i % 8 == 0 ? "\n " : " ");
494 printk("\nmasks:\n ");
495 for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
496 printk("%08lx%s", sh->evtchn_mask[i],
497 i % 8 == 0 ? "\n " : " ");
499 printk("\nunmasked:\n ");
500 for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
501 printk("%08lx%s", sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
502 i % 8 == 0 ? "\n " : " ");
504 printk("\npending list:\n");
505 for(i = 0; i < NR_EVENT_CHANNELS; i++) {
506 if (sync_test_bit(i, sh->evtchn_pending)) {
507 printk(" %d: event %d -> irq %d\n",
513 spin_unlock_irqrestore(&debug_lock, flags);
520 * Search the CPUs pending events bitmasks. For each one found, map
521 * the event number to an irq, and feed it into do_IRQ() for
524 * Xen uses a two-level bitmap to speed searching. The first level is
525 * a bitset of words which contain pending event bits. The second
526 * level is a bitset of pending events themselves.
528 void xen_evtchn_do_upcall(struct pt_regs *regs)
531 struct shared_info *s = HYPERVISOR_shared_info;
532 struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
533 static DEFINE_PER_CPU(unsigned, nesting_count);
537 unsigned long pending_words;
539 vcpu_info->evtchn_upcall_pending = 0;
541 if (__get_cpu_var(nesting_count)++)
544 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
545 /* Clear master flag /before/ clearing selector flag. */
548 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
549 while (pending_words != 0) {
550 unsigned long pending_bits;
551 int word_idx = __ffs(pending_words);
552 pending_words &= ~(1UL << word_idx);
554 while ((pending_bits = active_evtchns(cpu, s, word_idx)) != 0) {
555 int bit_idx = __ffs(pending_bits);
556 int port = (word_idx * BITS_PER_LONG) + bit_idx;
557 int irq = evtchn_to_irq[port];
560 xen_do_IRQ(irq, regs);
564 BUG_ON(!irqs_disabled());
566 count = __get_cpu_var(nesting_count);
567 __get_cpu_var(nesting_count) = 0;
574 /* Rebind a new event channel to an existing irq. */
575 void rebind_evtchn_irq(int evtchn, int irq)
577 /* Make sure the irq is masked, since the new event channel
578 will also be masked. */
581 spin_lock(&irq_mapping_update_lock);
583 /* After resume the irq<->evtchn mappings are all cleared out */
584 BUG_ON(evtchn_to_irq[evtchn] != -1);
585 /* Expect irq to have been bound before,
586 so the bindcount should be non-0 */
587 BUG_ON(irq_bindcount[irq] == 0);
589 evtchn_to_irq[evtchn] = irq;
590 irq_info[irq] = mk_irq_info(IRQT_EVTCHN, 0, evtchn);
592 spin_unlock(&irq_mapping_update_lock);
594 /* new event channels are always bound to cpu 0 */
595 irq_set_affinity(irq, cpumask_of(0));
597 /* Unmask the event channel. */
601 /* Rebind an evtchn so that it gets delivered to a specific cpu */
602 static void rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
604 struct evtchn_bind_vcpu bind_vcpu;
605 int evtchn = evtchn_from_irq(irq);
607 if (!VALID_EVTCHN(evtchn))
610 /* Send future instances of this interrupt to other vcpu. */
611 bind_vcpu.port = evtchn;
612 bind_vcpu.vcpu = tcpu;
615 * If this fails, it usually just indicates that we're dealing with a
616 * virq or IPI channel, which don't actually need to be rebound. Ignore
617 * it, but don't do the xenlinux-level rebind in that case.
619 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
620 bind_evtchn_to_cpu(evtchn, tcpu);
624 static void set_affinity_irq(unsigned irq, const struct cpumask *dest)
626 unsigned tcpu = cpumask_first(dest);
627 rebind_irq_to_cpu(irq, tcpu);
630 int resend_irq_on_evtchn(unsigned int irq)
632 int masked, evtchn = evtchn_from_irq(irq);
633 struct shared_info *s = HYPERVISOR_shared_info;
635 if (!VALID_EVTCHN(evtchn))
638 masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
639 sync_set_bit(evtchn, s->evtchn_pending);
641 unmask_evtchn(evtchn);
646 static void enable_dynirq(unsigned int irq)
648 int evtchn = evtchn_from_irq(irq);
650 if (VALID_EVTCHN(evtchn))
651 unmask_evtchn(evtchn);
654 static void disable_dynirq(unsigned int irq)
656 int evtchn = evtchn_from_irq(irq);
658 if (VALID_EVTCHN(evtchn))
662 static void ack_dynirq(unsigned int irq)
664 int evtchn = evtchn_from_irq(irq);
666 move_native_irq(irq);
668 if (VALID_EVTCHN(evtchn))
669 clear_evtchn(evtchn);
672 static int retrigger_dynirq(unsigned int irq)
674 int evtchn = evtchn_from_irq(irq);
675 struct shared_info *sh = HYPERVISOR_shared_info;
678 if (VALID_EVTCHN(evtchn)) {
681 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
682 sync_set_bit(evtchn, sh->evtchn_pending);
684 unmask_evtchn(evtchn);
691 static void restore_cpu_virqs(unsigned int cpu)
693 struct evtchn_bind_virq bind_virq;
694 int virq, irq, evtchn;
696 for (virq = 0; virq < NR_VIRQS; virq++) {
697 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
700 BUG_ON(irq_info[irq].type != IRQT_VIRQ);
701 BUG_ON(irq_info[irq].index != virq);
703 /* Get a new binding from Xen. */
704 bind_virq.virq = virq;
705 bind_virq.vcpu = cpu;
706 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
709 evtchn = bind_virq.port;
711 /* Record the new mapping. */
712 evtchn_to_irq[evtchn] = irq;
713 irq_info[irq] = mk_irq_info(IRQT_VIRQ, virq, evtchn);
714 bind_evtchn_to_cpu(evtchn, cpu);
717 unmask_evtchn(evtchn);
721 static void restore_cpu_ipis(unsigned int cpu)
723 struct evtchn_bind_ipi bind_ipi;
724 int ipi, irq, evtchn;
726 for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
727 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
730 BUG_ON(irq_info[irq].type != IRQT_IPI);
731 BUG_ON(irq_info[irq].index != ipi);
733 /* Get a new binding from Xen. */
735 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
738 evtchn = bind_ipi.port;
740 /* Record the new mapping. */
741 evtchn_to_irq[evtchn] = irq;
742 irq_info[irq] = mk_irq_info(IRQT_IPI, ipi, evtchn);
743 bind_evtchn_to_cpu(evtchn, cpu);
746 unmask_evtchn(evtchn);
751 /* Clear an irq's pending state, in preparation for polling on it */
752 void xen_clear_irq_pending(int irq)
754 int evtchn = evtchn_from_irq(irq);
756 if (VALID_EVTCHN(evtchn))
757 clear_evtchn(evtchn);
760 void xen_set_irq_pending(int irq)
762 int evtchn = evtchn_from_irq(irq);
764 if (VALID_EVTCHN(evtchn))
768 bool xen_test_irq_pending(int irq)
770 int evtchn = evtchn_from_irq(irq);
773 if (VALID_EVTCHN(evtchn))
774 ret = test_evtchn(evtchn);
779 /* Poll waiting for an irq to become pending. In the usual case, the
780 irq will be disabled so it won't deliver an interrupt. */
781 void xen_poll_irq(int irq)
783 evtchn_port_t evtchn = evtchn_from_irq(irq);
785 if (VALID_EVTCHN(evtchn)) {
786 struct sched_poll poll;
790 set_xen_guest_handle(poll.ports, &evtchn);
792 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
797 void xen_irq_resume(void)
799 unsigned int cpu, irq, evtchn;
801 init_evtchn_cpu_bindings();
803 /* New event-channel space is not 'live' yet. */
804 for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
807 /* No IRQ <-> event-channel mappings. */
808 for (irq = 0; irq < nr_irqs; irq++)
809 irq_info[irq].evtchn = 0; /* zap event-channel binding */
811 for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
812 evtchn_to_irq[evtchn] = -1;
814 for_each_possible_cpu(cpu) {
815 restore_cpu_virqs(cpu);
816 restore_cpu_ipis(cpu);
820 static struct irq_chip xen_dynamic_chip __read_mostly = {
822 .mask = disable_dynirq,
823 .unmask = enable_dynirq,
825 .set_affinity = set_affinity_irq,
826 .retrigger = retrigger_dynirq,
829 void __init xen_init_IRQ(void)
832 size_t size = nr_cpu_ids * sizeof(struct cpu_evtchn_s);
834 cpu_evtchn_mask_p = kmalloc(size, GFP_KERNEL);
835 BUG_ON(cpu_evtchn_mask == NULL);
837 init_evtchn_cpu_bindings();
839 /* No event channels are 'live' right now. */
840 for (i = 0; i < NR_EVENT_CHANNELS; i++)
843 /* Dynamic IRQ space is currently unbound. Zero the refcnts. */
844 for (i = 0; i < nr_irqs; i++)
845 irq_bindcount[i] = 0;
847 irq_ctx_init(smp_processor_id());