1 // SPDX-License-Identifier: GPL-2.0-only
5 * Xen models interrupts with abstract event channels. Because each
6 * domain gets 1024 event channels, but NR_IRQ is not that large, we
7 * must dynamically map irqs<->event channels. The event channels
8 * interface with the rest of the kernel by defining a xen interrupt
9 * chip. When an event is received, it is mapped to an irq and sent
10 * through the normal interrupt processing path.
12 * There are four kinds of events which can be mapped to an event
15 * 1. Inter-domain notifications. This includes all the virtual
16 * device events, since they're driven by front-ends in another domain
18 * 2. VIRQs, typically used for timers. These are per-cpu events.
20 * 4. PIRQs - Hardware interrupts.
22 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
25 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
27 #include <linux/linkage.h>
28 #include <linux/interrupt.h>
29 #include <linux/irq.h>
30 #include <linux/moduleparam.h>
31 #include <linux/string.h>
32 #include <linux/memblock.h>
33 #include <linux/slab.h>
34 #include <linux/irqnr.h>
35 #include <linux/pci.h>
36 #include <linux/spinlock.h>
37 #include <linux/cpuhotplug.h>
38 #include <linux/atomic.h>
39 #include <linux/ktime.h>
43 #include <asm/ptrace.h>
44 #include <asm/idtentry.h>
46 #include <asm/io_apic.h>
47 #include <asm/i8259.h>
48 #include <asm/xen/pci.h>
50 #include <asm/sync_bitops.h>
51 #include <asm/xen/hypercall.h>
52 #include <asm/xen/hypervisor.h>
57 #include <xen/xen-ops.h>
58 #include <xen/events.h>
59 #include <xen/interface/xen.h>
60 #include <xen/interface/event_channel.h>
61 #include <xen/interface/hvm/hvm_op.h>
62 #include <xen/interface/hvm/params.h>
63 #include <xen/interface/physdev.h>
64 #include <xen/interface/sched.h>
65 #include <xen/interface/vcpu.h>
66 #include <xen/xenbus.h>
67 #include <asm/hw_irq.h>
69 #include "events_internal.h"
71 #undef MODULE_PARAM_PREFIX
72 #define MODULE_PARAM_PREFIX "xen."
74 /* Interrupt types. */
84 * Packed IRQ information:
85 * type - enum xen_irq_type
86 * event channel - irq->event channel mapping
87 * cpu - cpu this event channel is bound to
88 * index - type-specific information:
89 * PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
90 * guest, or GSI (real passthrough IRQ) of the device.
96 struct list_head list;
97 struct list_head eoi_list;
101 short type; /* type: IRQT_* */
102 u8 mask_reason; /* Why is event channel masked */
103 #define EVT_MASK_REASON_EXPLICIT 0x01
104 #define EVT_MASK_REASON_TEMPORARY 0x02
105 #define EVT_MASK_REASON_EOI_PENDING 0x04
106 u8 is_active; /* Is event just being handled? */
108 evtchn_port_t evtchn; /* event channel */
109 unsigned short cpu; /* cpu bound */
110 unsigned short eoi_cpu; /* EOI must happen on this cpu-1 */
111 unsigned int irq_epoch; /* If eoi_cpu valid: irq_epoch of event */
112 u64 eoi_time; /* Time in jiffies when to EOI. */
121 unsigned char vector;
125 struct xenbus_device *interdomain;
129 #define PIRQ_NEEDS_EOI (1 << 0)
130 #define PIRQ_SHAREABLE (1 << 1)
131 #define PIRQ_MSI_GROUP (1 << 2)
133 static uint __read_mostly event_loop_timeout = 2;
134 module_param(event_loop_timeout, uint, 0644);
136 static uint __read_mostly event_eoi_delay = 10;
137 module_param(event_eoi_delay, uint, 0644);
139 const struct evtchn_ops *evtchn_ops;
142 * This lock protects updates to the following mapping and reference-count
143 * arrays. The lock does not need to be acquired to read the mapping tables.
145 static DEFINE_MUTEX(irq_mapping_update_lock);
148 * Lock protecting event handling loop against removing event channels.
149 * Adding of event channels is no issue as the associated IRQ becomes active
150 * only after everything is setup (before request_[threaded_]irq() the handler
151 * can't be entered for an event, as the event channel will be unmasked only
154 static DEFINE_RWLOCK(evtchn_rwlock);
159 * irq_mapping_update_lock
162 * percpu eoi_list_lock
166 static LIST_HEAD(xen_irq_list_head);
168 /* IRQ <-> VIRQ mapping. */
169 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
171 /* IRQ <-> IPI mapping */
172 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
174 /* Event channel distribution data */
175 static atomic_t channels_on_cpu[NR_CPUS];
177 static int **evtchn_to_irq;
179 static unsigned long *pirq_eoi_map;
181 static bool (*pirq_needs_eoi)(unsigned irq);
183 #define EVTCHN_ROW(e) (e / (PAGE_SIZE/sizeof(**evtchn_to_irq)))
184 #define EVTCHN_COL(e) (e % (PAGE_SIZE/sizeof(**evtchn_to_irq)))
185 #define EVTCHN_PER_ROW (PAGE_SIZE / sizeof(**evtchn_to_irq))
187 /* Xen will never allocate port zero for any purpose. */
188 #define VALID_EVTCHN(chn) ((chn) != 0)
190 static struct irq_info *legacy_info_ptrs[NR_IRQS_LEGACY];
192 static struct irq_chip xen_dynamic_chip;
193 static struct irq_chip xen_lateeoi_chip;
194 static struct irq_chip xen_percpu_chip;
195 static struct irq_chip xen_pirq_chip;
196 static void enable_dynirq(struct irq_data *data);
197 static void disable_dynirq(struct irq_data *data);
199 static DEFINE_PER_CPU(unsigned int, irq_epoch);
201 static void clear_evtchn_to_irq_row(int *evtchn_row)
205 for (col = 0; col < EVTCHN_PER_ROW; col++)
206 WRITE_ONCE(evtchn_row[col], -1);
209 static void clear_evtchn_to_irq_all(void)
213 for (row = 0; row < EVTCHN_ROW(xen_evtchn_max_channels()); row++) {
214 if (evtchn_to_irq[row] == NULL)
216 clear_evtchn_to_irq_row(evtchn_to_irq[row]);
220 static int set_evtchn_to_irq(evtchn_port_t evtchn, unsigned int irq)
226 if (evtchn >= xen_evtchn_max_channels())
229 row = EVTCHN_ROW(evtchn);
230 col = EVTCHN_COL(evtchn);
232 if (evtchn_to_irq[row] == NULL) {
233 /* Unallocated irq entries return -1 anyway */
237 evtchn_row = (int *) __get_free_pages(GFP_KERNEL, 0);
238 if (evtchn_row == NULL)
241 clear_evtchn_to_irq_row(evtchn_row);
244 * We've prepared an empty row for the mapping. If a different
245 * thread was faster inserting it, we can drop ours.
247 if (cmpxchg(&evtchn_to_irq[row], NULL, evtchn_row) != NULL)
248 free_page((unsigned long) evtchn_row);
251 WRITE_ONCE(evtchn_to_irq[row][col], irq);
255 int get_evtchn_to_irq(evtchn_port_t evtchn)
257 if (evtchn >= xen_evtchn_max_channels())
259 if (evtchn_to_irq[EVTCHN_ROW(evtchn)] == NULL)
261 return READ_ONCE(evtchn_to_irq[EVTCHN_ROW(evtchn)][EVTCHN_COL(evtchn)]);
264 /* Get info for IRQ */
265 static struct irq_info *info_for_irq(unsigned irq)
267 if (irq < nr_legacy_irqs())
268 return legacy_info_ptrs[irq];
270 return irq_get_chip_data(irq);
273 static void set_info_for_irq(unsigned int irq, struct irq_info *info)
275 if (irq < nr_legacy_irqs())
276 legacy_info_ptrs[irq] = info;
278 irq_set_chip_data(irq, info);
281 /* Per CPU channel accounting */
282 static void channels_on_cpu_dec(struct irq_info *info)
284 if (!info->is_accounted)
287 info->is_accounted = 0;
289 if (WARN_ON_ONCE(info->cpu >= nr_cpu_ids))
292 WARN_ON_ONCE(!atomic_add_unless(&channels_on_cpu[info->cpu], -1 , 0));
295 static void channels_on_cpu_inc(struct irq_info *info)
297 if (WARN_ON_ONCE(info->cpu >= nr_cpu_ids))
300 if (WARN_ON_ONCE(!atomic_add_unless(&channels_on_cpu[info->cpu], 1,
304 info->is_accounted = 1;
307 /* Constructors for packed IRQ information. */
308 static int xen_irq_info_common_setup(struct irq_info *info,
310 enum xen_irq_type type,
311 evtchn_port_t evtchn,
316 BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
320 info->evtchn = evtchn;
322 info->mask_reason = EVT_MASK_REASON_EXPLICIT;
323 raw_spin_lock_init(&info->lock);
325 ret = set_evtchn_to_irq(evtchn, irq);
329 irq_clear_status_flags(irq, IRQ_NOREQUEST|IRQ_NOAUTOEN);
331 return xen_evtchn_port_setup(evtchn);
334 static int xen_irq_info_evtchn_setup(unsigned irq,
335 evtchn_port_t evtchn,
336 struct xenbus_device *dev)
338 struct irq_info *info = info_for_irq(irq);
341 ret = xen_irq_info_common_setup(info, irq, IRQT_EVTCHN, evtchn, 0);
342 info->u.interdomain = dev;
344 atomic_inc(&dev->event_channels);
349 static int xen_irq_info_ipi_setup(unsigned cpu,
351 evtchn_port_t evtchn,
354 struct irq_info *info = info_for_irq(irq);
358 per_cpu(ipi_to_irq, cpu)[ipi] = irq;
360 return xen_irq_info_common_setup(info, irq, IRQT_IPI, evtchn, 0);
363 static int xen_irq_info_virq_setup(unsigned cpu,
365 evtchn_port_t evtchn,
368 struct irq_info *info = info_for_irq(irq);
372 per_cpu(virq_to_irq, cpu)[virq] = irq;
374 return xen_irq_info_common_setup(info, irq, IRQT_VIRQ, evtchn, 0);
377 static int xen_irq_info_pirq_setup(unsigned irq,
378 evtchn_port_t evtchn,
384 struct irq_info *info = info_for_irq(irq);
386 info->u.pirq.pirq = pirq;
387 info->u.pirq.gsi = gsi;
388 info->u.pirq.domid = domid;
389 info->u.pirq.flags = flags;
391 return xen_irq_info_common_setup(info, irq, IRQT_PIRQ, evtchn, 0);
394 static void xen_irq_info_cleanup(struct irq_info *info)
396 set_evtchn_to_irq(info->evtchn, -1);
397 xen_evtchn_port_remove(info->evtchn, info->cpu);
399 channels_on_cpu_dec(info);
403 * Accessors for packed IRQ information.
405 evtchn_port_t evtchn_from_irq(unsigned irq)
407 const struct irq_info *info = NULL;
409 if (likely(irq < nr_irqs))
410 info = info_for_irq(irq);
417 unsigned int irq_from_evtchn(evtchn_port_t evtchn)
419 return get_evtchn_to_irq(evtchn);
421 EXPORT_SYMBOL_GPL(irq_from_evtchn);
423 int irq_from_virq(unsigned int cpu, unsigned int virq)
425 return per_cpu(virq_to_irq, cpu)[virq];
428 static enum ipi_vector ipi_from_irq(unsigned irq)
430 struct irq_info *info = info_for_irq(irq);
432 BUG_ON(info == NULL);
433 BUG_ON(info->type != IRQT_IPI);
438 static unsigned virq_from_irq(unsigned irq)
440 struct irq_info *info = info_for_irq(irq);
442 BUG_ON(info == NULL);
443 BUG_ON(info->type != IRQT_VIRQ);
448 static unsigned pirq_from_irq(unsigned irq)
450 struct irq_info *info = info_for_irq(irq);
452 BUG_ON(info == NULL);
453 BUG_ON(info->type != IRQT_PIRQ);
455 return info->u.pirq.pirq;
458 static enum xen_irq_type type_from_irq(unsigned irq)
460 return info_for_irq(irq)->type;
463 static unsigned cpu_from_irq(unsigned irq)
465 return info_for_irq(irq)->cpu;
468 unsigned int cpu_from_evtchn(evtchn_port_t evtchn)
470 int irq = get_evtchn_to_irq(evtchn);
474 ret = cpu_from_irq(irq);
479 static void do_mask(struct irq_info *info, u8 reason)
483 raw_spin_lock_irqsave(&info->lock, flags);
485 if (!info->mask_reason)
486 mask_evtchn(info->evtchn);
488 info->mask_reason |= reason;
490 raw_spin_unlock_irqrestore(&info->lock, flags);
493 static void do_unmask(struct irq_info *info, u8 reason)
497 raw_spin_lock_irqsave(&info->lock, flags);
499 info->mask_reason &= ~reason;
501 if (!info->mask_reason)
502 unmask_evtchn(info->evtchn);
504 raw_spin_unlock_irqrestore(&info->lock, flags);
508 static bool pirq_check_eoi_map(unsigned irq)
510 return test_bit(pirq_from_irq(irq), pirq_eoi_map);
514 static bool pirq_needs_eoi_flag(unsigned irq)
516 struct irq_info *info = info_for_irq(irq);
517 BUG_ON(info->type != IRQT_PIRQ);
519 return info->u.pirq.flags & PIRQ_NEEDS_EOI;
522 static void bind_evtchn_to_cpu(evtchn_port_t evtchn, unsigned int cpu,
525 int irq = get_evtchn_to_irq(evtchn);
526 struct irq_info *info = info_for_irq(irq);
530 if (IS_ENABLED(CONFIG_SMP) && force_affinity) {
531 cpumask_copy(irq_get_affinity_mask(irq), cpumask_of(cpu));
532 cpumask_copy(irq_get_effective_affinity_mask(irq),
536 xen_evtchn_port_bind_to_cpu(evtchn, cpu, info->cpu);
538 channels_on_cpu_dec(info);
540 channels_on_cpu_inc(info);
544 * notify_remote_via_irq - send event to remote end of event channel via irq
545 * @irq: irq of event channel to send event to
547 * Unlike notify_remote_via_evtchn(), this is safe to use across
548 * save/restore. Notifications on a broken connection are silently
551 void notify_remote_via_irq(int irq)
553 evtchn_port_t evtchn = evtchn_from_irq(irq);
555 if (VALID_EVTCHN(evtchn))
556 notify_remote_via_evtchn(evtchn);
558 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
560 struct lateeoi_work {
561 struct delayed_work delayed;
562 spinlock_t eoi_list_lock;
563 struct list_head eoi_list;
566 static DEFINE_PER_CPU(struct lateeoi_work, lateeoi);
568 static void lateeoi_list_del(struct irq_info *info)
570 struct lateeoi_work *eoi = &per_cpu(lateeoi, info->eoi_cpu);
573 spin_lock_irqsave(&eoi->eoi_list_lock, flags);
574 list_del_init(&info->eoi_list);
575 spin_unlock_irqrestore(&eoi->eoi_list_lock, flags);
578 static void lateeoi_list_add(struct irq_info *info)
580 struct lateeoi_work *eoi = &per_cpu(lateeoi, info->eoi_cpu);
581 struct irq_info *elem;
582 u64 now = get_jiffies_64();
586 if (now < info->eoi_time)
587 delay = info->eoi_time - now;
591 spin_lock_irqsave(&eoi->eoi_list_lock, flags);
593 if (list_empty(&eoi->eoi_list)) {
594 list_add(&info->eoi_list, &eoi->eoi_list);
595 mod_delayed_work_on(info->eoi_cpu, system_wq,
596 &eoi->delayed, delay);
598 list_for_each_entry_reverse(elem, &eoi->eoi_list, eoi_list) {
599 if (elem->eoi_time <= info->eoi_time)
602 list_add(&info->eoi_list, &elem->eoi_list);
605 spin_unlock_irqrestore(&eoi->eoi_list_lock, flags);
608 static void xen_irq_lateeoi_locked(struct irq_info *info, bool spurious)
610 evtchn_port_t evtchn;
612 unsigned int delay = 0;
614 evtchn = info->evtchn;
615 if (!VALID_EVTCHN(evtchn) || !list_empty(&info->eoi_list))
619 struct xenbus_device *dev = info->u.interdomain;
620 unsigned int threshold = 1;
622 if (dev && dev->spurious_threshold)
623 threshold = dev->spurious_threshold;
625 if ((1 << info->spurious_cnt) < (HZ << 2)) {
626 if (info->spurious_cnt != 0xFF)
627 info->spurious_cnt++;
629 if (info->spurious_cnt > threshold) {
630 delay = 1 << (info->spurious_cnt - 1 - threshold);
634 info->eoi_cpu = smp_processor_id();
635 info->eoi_time = get_jiffies_64() + delay;
637 atomic_add(delay, &dev->jiffies_eoi_delayed);
640 atomic_inc(&dev->spurious_events);
642 info->spurious_cnt = 0;
646 if (info->eoi_time &&
647 (info->irq_epoch == per_cpu(irq_epoch, cpu) || delay)) {
648 lateeoi_list_add(info);
654 /* is_active hasn't been reset yet, do it now. */
655 smp_store_release(&info->is_active, 0);
656 do_unmask(info, EVT_MASK_REASON_EOI_PENDING);
659 static void xen_irq_lateeoi_worker(struct work_struct *work)
661 struct lateeoi_work *eoi;
662 struct irq_info *info;
663 u64 now = get_jiffies_64();
666 eoi = container_of(to_delayed_work(work), struct lateeoi_work, delayed);
668 read_lock_irqsave(&evtchn_rwlock, flags);
671 spin_lock(&eoi->eoi_list_lock);
673 info = list_first_entry_or_null(&eoi->eoi_list, struct irq_info,
676 if (info == NULL || now < info->eoi_time) {
677 spin_unlock(&eoi->eoi_list_lock);
681 list_del_init(&info->eoi_list);
683 spin_unlock(&eoi->eoi_list_lock);
687 xen_irq_lateeoi_locked(info, false);
691 mod_delayed_work_on(info->eoi_cpu, system_wq,
692 &eoi->delayed, info->eoi_time - now);
694 read_unlock_irqrestore(&evtchn_rwlock, flags);
697 static void xen_cpu_init_eoi(unsigned int cpu)
699 struct lateeoi_work *eoi = &per_cpu(lateeoi, cpu);
701 INIT_DELAYED_WORK(&eoi->delayed, xen_irq_lateeoi_worker);
702 spin_lock_init(&eoi->eoi_list_lock);
703 INIT_LIST_HEAD(&eoi->eoi_list);
706 void xen_irq_lateeoi(unsigned int irq, unsigned int eoi_flags)
708 struct irq_info *info;
711 read_lock_irqsave(&evtchn_rwlock, flags);
713 info = info_for_irq(irq);
716 xen_irq_lateeoi_locked(info, eoi_flags & XEN_EOI_FLAG_SPURIOUS);
718 read_unlock_irqrestore(&evtchn_rwlock, flags);
720 EXPORT_SYMBOL_GPL(xen_irq_lateeoi);
722 static void xen_irq_init(unsigned irq)
724 struct irq_info *info;
726 info = kzalloc(sizeof(*info), GFP_KERNEL);
728 panic("Unable to allocate metadata for IRQ%d\n", irq);
730 info->type = IRQT_UNBOUND;
733 set_info_for_irq(irq, info);
735 * Interrupt affinity setting can be immediate. No point
736 * in delaying it until an interrupt is handled.
738 irq_set_status_flags(irq, IRQ_MOVE_PCNTXT);
740 INIT_LIST_HEAD(&info->eoi_list);
741 list_add_tail(&info->list, &xen_irq_list_head);
744 static int __must_check xen_allocate_irqs_dynamic(int nvec)
746 int i, irq = irq_alloc_descs(-1, 0, nvec, -1);
749 for (i = 0; i < nvec; i++)
750 xen_irq_init(irq + i);
756 static inline int __must_check xen_allocate_irq_dynamic(void)
759 return xen_allocate_irqs_dynamic(1);
762 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
767 * A PV guest has no concept of a GSI (since it has no ACPI
768 * nor access to/knowledge of the physical APICs). Therefore
769 * all IRQs are dynamically allocated from the entire IRQ
772 if (xen_pv_domain() && !xen_initial_domain())
773 return xen_allocate_irq_dynamic();
775 /* Legacy IRQ descriptors are already allocated by the arch. */
776 if (gsi < nr_legacy_irqs())
779 irq = irq_alloc_desc_at(gsi, -1);
786 static void xen_free_irq(unsigned irq)
788 struct irq_info *info = info_for_irq(irq);
794 write_lock_irqsave(&evtchn_rwlock, flags);
796 if (!list_empty(&info->eoi_list))
797 lateeoi_list_del(info);
799 list_del(&info->list);
801 set_info_for_irq(irq, NULL);
803 WARN_ON(info->refcnt > 0);
805 write_unlock_irqrestore(&evtchn_rwlock, flags);
809 /* Legacy IRQ descriptors are managed by the arch. */
810 if (irq < nr_legacy_irqs())
816 static void xen_evtchn_close(evtchn_port_t port)
818 struct evtchn_close close;
821 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
825 /* Not called for lateeoi events. */
826 static void event_handler_exit(struct irq_info *info)
828 smp_store_release(&info->is_active, 0);
829 clear_evtchn(info->evtchn);
832 static void pirq_query_unmask(int irq)
834 struct physdev_irq_status_query irq_status;
835 struct irq_info *info = info_for_irq(irq);
837 BUG_ON(info->type != IRQT_PIRQ);
839 irq_status.irq = pirq_from_irq(irq);
840 if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
841 irq_status.flags = 0;
843 info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
844 if (irq_status.flags & XENIRQSTAT_needs_eoi)
845 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
848 static void eoi_pirq(struct irq_data *data)
850 struct irq_info *info = info_for_irq(data->irq);
851 evtchn_port_t evtchn = info ? info->evtchn : 0;
852 struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
855 if (!VALID_EVTCHN(evtchn))
858 event_handler_exit(info);
860 if (pirq_needs_eoi(data->irq)) {
861 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
866 static void mask_ack_pirq(struct irq_data *data)
868 disable_dynirq(data);
872 static unsigned int __startup_pirq(unsigned int irq)
874 struct evtchn_bind_pirq bind_pirq;
875 struct irq_info *info = info_for_irq(irq);
876 evtchn_port_t evtchn = evtchn_from_irq(irq);
879 BUG_ON(info->type != IRQT_PIRQ);
881 if (VALID_EVTCHN(evtchn))
884 bind_pirq.pirq = pirq_from_irq(irq);
885 /* NB. We are happy to share unless we are probing. */
886 bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
887 BIND_PIRQ__WILL_SHARE : 0;
888 rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
890 pr_warn("Failed to obtain physical IRQ %d\n", irq);
893 evtchn = bind_pirq.port;
895 pirq_query_unmask(irq);
897 rc = set_evtchn_to_irq(evtchn, irq);
901 info->evtchn = evtchn;
902 bind_evtchn_to_cpu(evtchn, 0, false);
904 rc = xen_evtchn_port_setup(evtchn);
909 do_unmask(info, EVT_MASK_REASON_EXPLICIT);
911 eoi_pirq(irq_get_irq_data(irq));
916 pr_err("irq%d: Failed to set port to irq mapping (%d)\n", irq, rc);
917 xen_evtchn_close(evtchn);
921 static unsigned int startup_pirq(struct irq_data *data)
923 return __startup_pirq(data->irq);
926 static void shutdown_pirq(struct irq_data *data)
928 unsigned int irq = data->irq;
929 struct irq_info *info = info_for_irq(irq);
930 evtchn_port_t evtchn = evtchn_from_irq(irq);
932 BUG_ON(info->type != IRQT_PIRQ);
934 if (!VALID_EVTCHN(evtchn))
937 do_mask(info, EVT_MASK_REASON_EXPLICIT);
938 xen_evtchn_close(evtchn);
939 xen_irq_info_cleanup(info);
942 static void enable_pirq(struct irq_data *data)
947 static void disable_pirq(struct irq_data *data)
949 disable_dynirq(data);
952 int xen_irq_from_gsi(unsigned gsi)
954 struct irq_info *info;
956 list_for_each_entry(info, &xen_irq_list_head, list) {
957 if (info->type != IRQT_PIRQ)
960 if (info->u.pirq.gsi == gsi)
966 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
968 static void __unbind_from_irq(unsigned int irq)
970 evtchn_port_t evtchn = evtchn_from_irq(irq);
971 struct irq_info *info = info_for_irq(irq);
973 if (info->refcnt > 0) {
975 if (info->refcnt != 0)
979 if (VALID_EVTCHN(evtchn)) {
980 unsigned int cpu = cpu_from_irq(irq);
981 struct xenbus_device *dev;
983 xen_evtchn_close(evtchn);
985 switch (type_from_irq(irq)) {
987 per_cpu(virq_to_irq, cpu)[virq_from_irq(irq)] = -1;
990 per_cpu(ipi_to_irq, cpu)[ipi_from_irq(irq)] = -1;
993 dev = info->u.interdomain;
995 atomic_dec(&dev->event_channels);
1001 xen_irq_info_cleanup(info);
1008 * Do not make any assumptions regarding the relationship between the
1009 * IRQ number returned here and the Xen pirq argument.
1011 * Note: We don't assign an event channel until the irq actually started
1012 * up. Return an existing irq if we've already got one for the gsi.
1014 * Shareable implies level triggered, not shareable implies edge
1017 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
1018 unsigned pirq, int shareable, char *name)
1021 struct physdev_irq irq_op;
1024 mutex_lock(&irq_mapping_update_lock);
1026 irq = xen_irq_from_gsi(gsi);
1028 pr_info("%s: returning irq %d for gsi %u\n",
1029 __func__, irq, gsi);
1033 irq = xen_allocate_irq_gsi(gsi);
1040 /* Only the privileged domain can do this. For non-priv, the pcifront
1041 * driver provides a PCI bus that does the call to do exactly
1042 * this in the priv domain. */
1043 if (xen_initial_domain() &&
1044 HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
1050 ret = xen_irq_info_pirq_setup(irq, 0, pirq, gsi, DOMID_SELF,
1051 shareable ? PIRQ_SHAREABLE : 0);
1053 __unbind_from_irq(irq);
1058 pirq_query_unmask(irq);
1059 /* We try to use the handler with the appropriate semantic for the
1060 * type of interrupt: if the interrupt is an edge triggered
1061 * interrupt we use handle_edge_irq.
1063 * On the other hand if the interrupt is level triggered we use
1064 * handle_fasteoi_irq like the native code does for this kind of
1067 * Depending on the Xen version, pirq_needs_eoi might return true
1068 * not only for level triggered interrupts but for edge triggered
1069 * interrupts too. In any case Xen always honors the eoi mechanism,
1070 * not injecting any more pirqs of the same kind if the first one
1071 * hasn't received an eoi yet. Therefore using the fasteoi handler
1072 * is the right choice either way.
1075 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
1076 handle_fasteoi_irq, name);
1078 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
1079 handle_edge_irq, name);
1082 mutex_unlock(&irq_mapping_update_lock);
1087 #ifdef CONFIG_PCI_MSI
1088 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
1091 struct physdev_get_free_pirq op_get_free_pirq;
1093 op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
1094 rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
1096 WARN_ONCE(rc == -ENOSYS,
1097 "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
1099 return rc ? -1 : op_get_free_pirq.pirq;
1102 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
1103 int pirq, int nvec, const char *name, domid_t domid)
1107 mutex_lock(&irq_mapping_update_lock);
1109 irq = xen_allocate_irqs_dynamic(nvec);
1113 for (i = 0; i < nvec; i++) {
1114 irq_set_chip_and_handler_name(irq + i, &xen_pirq_chip, handle_edge_irq, name);
1116 ret = xen_irq_info_pirq_setup(irq + i, 0, pirq + i, 0, domid,
1117 i == 0 ? 0 : PIRQ_MSI_GROUP);
1122 ret = irq_set_msi_desc(irq, msidesc);
1126 mutex_unlock(&irq_mapping_update_lock);
1130 __unbind_from_irq(irq + nvec);
1131 mutex_unlock(&irq_mapping_update_lock);
1136 int xen_destroy_irq(int irq)
1138 struct physdev_unmap_pirq unmap_irq;
1139 struct irq_info *info = info_for_irq(irq);
1142 mutex_lock(&irq_mapping_update_lock);
1145 * If trying to remove a vector in a MSI group different
1146 * than the first one skip the PIRQ unmap unless this vector
1147 * is the first one in the group.
1149 if (xen_initial_domain() && !(info->u.pirq.flags & PIRQ_MSI_GROUP)) {
1150 unmap_irq.pirq = info->u.pirq.pirq;
1151 unmap_irq.domid = info->u.pirq.domid;
1152 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
1153 /* If another domain quits without making the pci_disable_msix
1154 * call, the Xen hypervisor takes care of freeing the PIRQs
1155 * (free_domain_pirqs).
1157 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
1158 pr_info("domain %d does not have %d anymore\n",
1159 info->u.pirq.domid, info->u.pirq.pirq);
1161 pr_warn("unmap irq failed %d\n", rc);
1169 mutex_unlock(&irq_mapping_update_lock);
1173 int xen_irq_from_pirq(unsigned pirq)
1177 struct irq_info *info;
1179 mutex_lock(&irq_mapping_update_lock);
1181 list_for_each_entry(info, &xen_irq_list_head, list) {
1182 if (info->type != IRQT_PIRQ)
1185 if (info->u.pirq.pirq == pirq)
1190 mutex_unlock(&irq_mapping_update_lock);
1196 int xen_pirq_from_irq(unsigned irq)
1198 return pirq_from_irq(irq);
1200 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
1202 static int bind_evtchn_to_irq_chip(evtchn_port_t evtchn, struct irq_chip *chip,
1203 struct xenbus_device *dev)
1208 if (evtchn >= xen_evtchn_max_channels())
1211 mutex_lock(&irq_mapping_update_lock);
1213 irq = get_evtchn_to_irq(evtchn);
1216 irq = xen_allocate_irq_dynamic();
1220 irq_set_chip_and_handler_name(irq, chip,
1221 handle_edge_irq, "event");
1223 ret = xen_irq_info_evtchn_setup(irq, evtchn, dev);
1225 __unbind_from_irq(irq);
1230 * New interdomain events are initially bound to vCPU0 This
1231 * is required to setup the event channel in the first
1232 * place and also important for UP guests because the
1233 * affinity setting is not invoked on them so nothing would
1236 bind_evtchn_to_cpu(evtchn, 0, false);
1238 struct irq_info *info = info_for_irq(irq);
1239 WARN_ON(info == NULL || info->type != IRQT_EVTCHN);
1243 mutex_unlock(&irq_mapping_update_lock);
1248 int bind_evtchn_to_irq(evtchn_port_t evtchn)
1250 return bind_evtchn_to_irq_chip(evtchn, &xen_dynamic_chip, NULL);
1252 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
1254 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
1256 struct evtchn_bind_ipi bind_ipi;
1257 evtchn_port_t evtchn;
1260 mutex_lock(&irq_mapping_update_lock);
1262 irq = per_cpu(ipi_to_irq, cpu)[ipi];
1265 irq = xen_allocate_irq_dynamic();
1269 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
1270 handle_percpu_irq, "ipi");
1272 bind_ipi.vcpu = xen_vcpu_nr(cpu);
1273 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1276 evtchn = bind_ipi.port;
1278 ret = xen_irq_info_ipi_setup(cpu, irq, evtchn, ipi);
1280 __unbind_from_irq(irq);
1285 * Force the affinity mask to the target CPU so proc shows
1286 * the correct target.
1288 bind_evtchn_to_cpu(evtchn, cpu, true);
1290 struct irq_info *info = info_for_irq(irq);
1291 WARN_ON(info == NULL || info->type != IRQT_IPI);
1295 mutex_unlock(&irq_mapping_update_lock);
1299 static int bind_interdomain_evtchn_to_irq_chip(struct xenbus_device *dev,
1300 evtchn_port_t remote_port,
1301 struct irq_chip *chip)
1303 struct evtchn_bind_interdomain bind_interdomain;
1306 bind_interdomain.remote_dom = dev->otherend_id;
1307 bind_interdomain.remote_port = remote_port;
1309 err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
1312 return err ? : bind_evtchn_to_irq_chip(bind_interdomain.local_port,
1316 int bind_interdomain_evtchn_to_irq_lateeoi(struct xenbus_device *dev,
1317 evtchn_port_t remote_port)
1319 return bind_interdomain_evtchn_to_irq_chip(dev, remote_port,
1322 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irq_lateeoi);
1324 static int find_virq(unsigned int virq, unsigned int cpu, evtchn_port_t *evtchn)
1326 struct evtchn_status status;
1330 memset(&status, 0, sizeof(status));
1331 for (port = 0; port < xen_evtchn_max_channels(); port++) {
1332 status.dom = DOMID_SELF;
1334 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
1337 if (status.status != EVTCHNSTAT_virq)
1339 if (status.u.virq == virq && status.vcpu == xen_vcpu_nr(cpu)) {
1348 * xen_evtchn_nr_channels - number of usable event channel ports
1350 * This may be less than the maximum supported by the current
1351 * hypervisor ABI. Use xen_evtchn_max_channels() for the maximum
1354 unsigned xen_evtchn_nr_channels(void)
1356 return evtchn_ops->nr_channels();
1358 EXPORT_SYMBOL_GPL(xen_evtchn_nr_channels);
1360 int bind_virq_to_irq(unsigned int virq, unsigned int cpu, bool percpu)
1362 struct evtchn_bind_virq bind_virq;
1363 evtchn_port_t evtchn = 0;
1366 mutex_lock(&irq_mapping_update_lock);
1368 irq = per_cpu(virq_to_irq, cpu)[virq];
1371 irq = xen_allocate_irq_dynamic();
1376 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
1377 handle_percpu_irq, "virq");
1379 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
1380 handle_edge_irq, "virq");
1382 bind_virq.virq = virq;
1383 bind_virq.vcpu = xen_vcpu_nr(cpu);
1384 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1387 evtchn = bind_virq.port;
1390 ret = find_virq(virq, cpu, &evtchn);
1394 ret = xen_irq_info_virq_setup(cpu, irq, evtchn, virq);
1396 __unbind_from_irq(irq);
1402 * Force the affinity mask for percpu interrupts so proc
1403 * shows the correct target.
1405 bind_evtchn_to_cpu(evtchn, cpu, percpu);
1407 struct irq_info *info = info_for_irq(irq);
1408 WARN_ON(info == NULL || info->type != IRQT_VIRQ);
1412 mutex_unlock(&irq_mapping_update_lock);
1417 static void unbind_from_irq(unsigned int irq)
1419 mutex_lock(&irq_mapping_update_lock);
1420 __unbind_from_irq(irq);
1421 mutex_unlock(&irq_mapping_update_lock);
1424 static int bind_evtchn_to_irqhandler_chip(evtchn_port_t evtchn,
1425 irq_handler_t handler,
1426 unsigned long irqflags,
1427 const char *devname, void *dev_id,
1428 struct irq_chip *chip)
1432 irq = bind_evtchn_to_irq_chip(evtchn, chip, NULL);
1435 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1437 unbind_from_irq(irq);
1444 int bind_evtchn_to_irqhandler(evtchn_port_t evtchn,
1445 irq_handler_t handler,
1446 unsigned long irqflags,
1447 const char *devname, void *dev_id)
1449 return bind_evtchn_to_irqhandler_chip(evtchn, handler, irqflags,
1453 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1455 int bind_evtchn_to_irqhandler_lateeoi(evtchn_port_t evtchn,
1456 irq_handler_t handler,
1457 unsigned long irqflags,
1458 const char *devname, void *dev_id)
1460 return bind_evtchn_to_irqhandler_chip(evtchn, handler, irqflags,
1464 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler_lateeoi);
1466 static int bind_interdomain_evtchn_to_irqhandler_chip(
1467 struct xenbus_device *dev, evtchn_port_t remote_port,
1468 irq_handler_t handler, unsigned long irqflags,
1469 const char *devname, void *dev_id, struct irq_chip *chip)
1473 irq = bind_interdomain_evtchn_to_irq_chip(dev, remote_port, chip);
1477 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1479 unbind_from_irq(irq);
1486 int bind_interdomain_evtchn_to_irqhandler_lateeoi(struct xenbus_device *dev,
1487 evtchn_port_t remote_port,
1488 irq_handler_t handler,
1489 unsigned long irqflags,
1490 const char *devname,
1493 return bind_interdomain_evtchn_to_irqhandler_chip(dev,
1494 remote_port, handler, irqflags, devname,
1495 dev_id, &xen_lateeoi_chip);
1497 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler_lateeoi);
1499 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1500 irq_handler_t handler,
1501 unsigned long irqflags, const char *devname, void *dev_id)
1505 irq = bind_virq_to_irq(virq, cpu, irqflags & IRQF_PERCPU);
1508 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1510 unbind_from_irq(irq);
1516 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1518 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1520 irq_handler_t handler,
1521 unsigned long irqflags,
1522 const char *devname,
1527 irq = bind_ipi_to_irq(ipi, cpu);
1531 irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1532 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1534 unbind_from_irq(irq);
1541 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1543 struct irq_info *info = info_for_irq(irq);
1547 free_irq(irq, dev_id);
1548 unbind_from_irq(irq);
1550 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1553 * xen_set_irq_priority() - set an event channel priority.
1554 * @irq:irq bound to an event channel.
1555 * @priority: priority between XEN_IRQ_PRIORITY_MAX and XEN_IRQ_PRIORITY_MIN.
1557 int xen_set_irq_priority(unsigned irq, unsigned priority)
1559 struct evtchn_set_priority set_priority;
1561 set_priority.port = evtchn_from_irq(irq);
1562 set_priority.priority = priority;
1564 return HYPERVISOR_event_channel_op(EVTCHNOP_set_priority,
1567 EXPORT_SYMBOL_GPL(xen_set_irq_priority);
1569 int evtchn_make_refcounted(evtchn_port_t evtchn)
1571 int irq = get_evtchn_to_irq(evtchn);
1572 struct irq_info *info;
1577 info = info_for_irq(irq);
1582 WARN_ON(info->refcnt != -1);
1588 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1590 int evtchn_get(evtchn_port_t evtchn)
1593 struct irq_info *info;
1596 if (evtchn >= xen_evtchn_max_channels())
1599 mutex_lock(&irq_mapping_update_lock);
1601 irq = get_evtchn_to_irq(evtchn);
1605 info = info_for_irq(irq);
1611 if (info->refcnt <= 0 || info->refcnt == SHRT_MAX)
1617 mutex_unlock(&irq_mapping_update_lock);
1621 EXPORT_SYMBOL_GPL(evtchn_get);
1623 void evtchn_put(evtchn_port_t evtchn)
1625 int irq = get_evtchn_to_irq(evtchn);
1626 if (WARN_ON(irq == -1))
1628 unbind_from_irq(irq);
1630 EXPORT_SYMBOL_GPL(evtchn_put);
1632 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1637 if (unlikely(vector == XEN_NMI_VECTOR)) {
1638 int rc = HYPERVISOR_vcpu_op(VCPUOP_send_nmi, xen_vcpu_nr(cpu),
1641 printk(KERN_WARNING "Sending nmi to CPU%d failed (rc:%d)\n", cpu, rc);
1645 irq = per_cpu(ipi_to_irq, cpu)[vector];
1647 notify_remote_via_irq(irq);
1650 struct evtchn_loop_ctrl {
1656 void handle_irq_for_port(evtchn_port_t port, struct evtchn_loop_ctrl *ctrl)
1659 struct irq_info *info;
1660 struct xenbus_device *dev;
1662 irq = get_evtchn_to_irq(port);
1667 * Check for timeout every 256 events.
1668 * We are setting the timeout value only after the first 256
1669 * events in order to not hurt the common case of few loop
1670 * iterations. The 256 is basically an arbitrary value.
1672 * In case we are hitting the timeout we need to defer all further
1673 * EOIs in order to ensure to leave the event handling loop rather
1674 * sooner than later.
1676 if (!ctrl->defer_eoi && !(++ctrl->count & 0xff)) {
1677 ktime_t kt = ktime_get();
1679 if (!ctrl->timeout) {
1680 kt = ktime_add_ms(kt,
1681 jiffies_to_msecs(event_loop_timeout));
1683 } else if (kt > ctrl->timeout) {
1684 ctrl->defer_eoi = true;
1688 info = info_for_irq(irq);
1689 if (xchg_acquire(&info->is_active, 1))
1692 dev = (info->type == IRQT_EVTCHN) ? info->u.interdomain : NULL;
1694 atomic_inc(&dev->events);
1696 if (ctrl->defer_eoi) {
1697 info->eoi_cpu = smp_processor_id();
1698 info->irq_epoch = __this_cpu_read(irq_epoch);
1699 info->eoi_time = get_jiffies_64() + event_eoi_delay;
1702 generic_handle_irq(irq);
1705 static void __xen_evtchn_do_upcall(void)
1707 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1708 int cpu = smp_processor_id();
1709 struct evtchn_loop_ctrl ctrl = { 0 };
1711 read_lock(&evtchn_rwlock);
1714 vcpu_info->evtchn_upcall_pending = 0;
1716 xen_evtchn_handle_events(cpu, &ctrl);
1718 BUG_ON(!irqs_disabled());
1720 virt_rmb(); /* Hypervisor can set upcall pending. */
1722 } while (vcpu_info->evtchn_upcall_pending);
1724 read_unlock(&evtchn_rwlock);
1727 * Increment irq_epoch only now to defer EOIs only for
1728 * xen_irq_lateeoi() invocations occurring from inside the loop
1731 __this_cpu_inc(irq_epoch);
1734 void xen_evtchn_do_upcall(struct pt_regs *regs)
1736 struct pt_regs *old_regs = set_irq_regs(regs);
1740 __xen_evtchn_do_upcall();
1743 set_irq_regs(old_regs);
1746 void xen_hvm_evtchn_do_upcall(void)
1748 __xen_evtchn_do_upcall();
1750 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1752 /* Rebind a new event channel to an existing irq. */
1753 void rebind_evtchn_irq(evtchn_port_t evtchn, int irq)
1755 struct irq_info *info = info_for_irq(irq);
1760 /* Make sure the irq is masked, since the new event channel
1761 will also be masked. */
1764 mutex_lock(&irq_mapping_update_lock);
1766 /* After resume the irq<->evtchn mappings are all cleared out */
1767 BUG_ON(get_evtchn_to_irq(evtchn) != -1);
1768 /* Expect irq to have been bound before,
1769 so there should be a proper type */
1770 BUG_ON(info->type == IRQT_UNBOUND);
1772 (void)xen_irq_info_evtchn_setup(irq, evtchn, NULL);
1774 mutex_unlock(&irq_mapping_update_lock);
1776 bind_evtchn_to_cpu(evtchn, info->cpu, false);
1778 /* Unmask the event channel. */
1782 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1783 static int xen_rebind_evtchn_to_cpu(struct irq_info *info, unsigned int tcpu)
1785 struct evtchn_bind_vcpu bind_vcpu;
1786 evtchn_port_t evtchn = info ? info->evtchn : 0;
1788 if (!VALID_EVTCHN(evtchn))
1791 if (!xen_support_evtchn_rebind())
1794 /* Send future instances of this interrupt to other vcpu. */
1795 bind_vcpu.port = evtchn;
1796 bind_vcpu.vcpu = xen_vcpu_nr(tcpu);
1799 * Mask the event while changing the VCPU binding to prevent
1800 * it being delivered on an unexpected VCPU.
1802 do_mask(info, EVT_MASK_REASON_TEMPORARY);
1805 * If this fails, it usually just indicates that we're dealing with a
1806 * virq or IPI channel, which don't actually need to be rebound. Ignore
1807 * it, but don't do the xenlinux-level rebind in that case.
1809 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1810 bind_evtchn_to_cpu(evtchn, tcpu, false);
1812 do_unmask(info, EVT_MASK_REASON_TEMPORARY);
1818 * Find the CPU within @dest mask which has the least number of channels
1819 * assigned. This is not precise as the per cpu counts can be modified
1822 static unsigned int select_target_cpu(const struct cpumask *dest)
1824 unsigned int cpu, best_cpu = UINT_MAX, minch = UINT_MAX;
1826 for_each_cpu_and(cpu, dest, cpu_online_mask) {
1827 unsigned int curch = atomic_read(&channels_on_cpu[cpu]);
1829 if (curch < minch) {
1836 * Catch the unlikely case that dest contains no online CPUs. Can't
1839 if (best_cpu == UINT_MAX)
1840 return select_target_cpu(cpu_online_mask);
1845 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1848 unsigned int tcpu = select_target_cpu(dest);
1851 ret = xen_rebind_evtchn_to_cpu(info_for_irq(data->irq), tcpu);
1853 irq_data_update_effective_affinity(data, cpumask_of(tcpu));
1858 static void enable_dynirq(struct irq_data *data)
1860 struct irq_info *info = info_for_irq(data->irq);
1861 evtchn_port_t evtchn = info ? info->evtchn : 0;
1863 if (VALID_EVTCHN(evtchn))
1864 do_unmask(info, EVT_MASK_REASON_EXPLICIT);
1867 static void disable_dynirq(struct irq_data *data)
1869 struct irq_info *info = info_for_irq(data->irq);
1870 evtchn_port_t evtchn = info ? info->evtchn : 0;
1872 if (VALID_EVTCHN(evtchn))
1873 do_mask(info, EVT_MASK_REASON_EXPLICIT);
1876 static void ack_dynirq(struct irq_data *data)
1878 struct irq_info *info = info_for_irq(data->irq);
1879 evtchn_port_t evtchn = info ? info->evtchn : 0;
1881 if (VALID_EVTCHN(evtchn))
1882 event_handler_exit(info);
1885 static void mask_ack_dynirq(struct irq_data *data)
1887 disable_dynirq(data);
1891 static void lateeoi_ack_dynirq(struct irq_data *data)
1893 struct irq_info *info = info_for_irq(data->irq);
1894 evtchn_port_t evtchn = info ? info->evtchn : 0;
1896 if (VALID_EVTCHN(evtchn)) {
1897 do_mask(info, EVT_MASK_REASON_EOI_PENDING);
1899 * Don't call event_handler_exit().
1900 * Need to keep is_active non-zero in order to ignore re-raised
1901 * events after cpu affinity changes while a lateeoi is pending.
1903 clear_evtchn(evtchn);
1907 static void lateeoi_mask_ack_dynirq(struct irq_data *data)
1909 struct irq_info *info = info_for_irq(data->irq);
1910 evtchn_port_t evtchn = info ? info->evtchn : 0;
1912 if (VALID_EVTCHN(evtchn)) {
1913 do_mask(info, EVT_MASK_REASON_EXPLICIT);
1914 event_handler_exit(info);
1918 static int retrigger_dynirq(struct irq_data *data)
1920 struct irq_info *info = info_for_irq(data->irq);
1921 evtchn_port_t evtchn = info ? info->evtchn : 0;
1923 if (!VALID_EVTCHN(evtchn))
1926 do_mask(info, EVT_MASK_REASON_TEMPORARY);
1928 do_unmask(info, EVT_MASK_REASON_TEMPORARY);
1933 static void restore_pirqs(void)
1935 int pirq, rc, irq, gsi;
1936 struct physdev_map_pirq map_irq;
1937 struct irq_info *info;
1939 list_for_each_entry(info, &xen_irq_list_head, list) {
1940 if (info->type != IRQT_PIRQ)
1943 pirq = info->u.pirq.pirq;
1944 gsi = info->u.pirq.gsi;
1947 /* save/restore of PT devices doesn't work, so at this point the
1948 * only devices present are GSI based emulated devices */
1952 map_irq.domid = DOMID_SELF;
1953 map_irq.type = MAP_PIRQ_TYPE_GSI;
1954 map_irq.index = gsi;
1955 map_irq.pirq = pirq;
1957 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1959 pr_warn("xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1960 gsi, irq, pirq, rc);
1965 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1967 __startup_pirq(irq);
1971 static void restore_cpu_virqs(unsigned int cpu)
1973 struct evtchn_bind_virq bind_virq;
1974 evtchn_port_t evtchn;
1977 for (virq = 0; virq < NR_VIRQS; virq++) {
1978 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1981 BUG_ON(virq_from_irq(irq) != virq);
1983 /* Get a new binding from Xen. */
1984 bind_virq.virq = virq;
1985 bind_virq.vcpu = xen_vcpu_nr(cpu);
1986 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1989 evtchn = bind_virq.port;
1991 /* Record the new mapping. */
1992 (void)xen_irq_info_virq_setup(cpu, irq, evtchn, virq);
1993 /* The affinity mask is still valid */
1994 bind_evtchn_to_cpu(evtchn, cpu, false);
1998 static void restore_cpu_ipis(unsigned int cpu)
2000 struct evtchn_bind_ipi bind_ipi;
2001 evtchn_port_t evtchn;
2004 for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
2005 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
2008 BUG_ON(ipi_from_irq(irq) != ipi);
2010 /* Get a new binding from Xen. */
2011 bind_ipi.vcpu = xen_vcpu_nr(cpu);
2012 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
2015 evtchn = bind_ipi.port;
2017 /* Record the new mapping. */
2018 (void)xen_irq_info_ipi_setup(cpu, irq, evtchn, ipi);
2019 /* The affinity mask is still valid */
2020 bind_evtchn_to_cpu(evtchn, cpu, false);
2024 /* Clear an irq's pending state, in preparation for polling on it */
2025 void xen_clear_irq_pending(int irq)
2027 struct irq_info *info = info_for_irq(irq);
2028 evtchn_port_t evtchn = info ? info->evtchn : 0;
2030 if (VALID_EVTCHN(evtchn))
2031 event_handler_exit(info);
2033 EXPORT_SYMBOL(xen_clear_irq_pending);
2034 void xen_set_irq_pending(int irq)
2036 evtchn_port_t evtchn = evtchn_from_irq(irq);
2038 if (VALID_EVTCHN(evtchn))
2042 bool xen_test_irq_pending(int irq)
2044 evtchn_port_t evtchn = evtchn_from_irq(irq);
2047 if (VALID_EVTCHN(evtchn))
2048 ret = test_evtchn(evtchn);
2053 /* Poll waiting for an irq to become pending with timeout. In the usual case,
2054 * the irq will be disabled so it won't deliver an interrupt. */
2055 void xen_poll_irq_timeout(int irq, u64 timeout)
2057 evtchn_port_t evtchn = evtchn_from_irq(irq);
2059 if (VALID_EVTCHN(evtchn)) {
2060 struct sched_poll poll;
2063 poll.timeout = timeout;
2064 set_xen_guest_handle(poll.ports, &evtchn);
2066 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
2070 EXPORT_SYMBOL(xen_poll_irq_timeout);
2071 /* Poll waiting for an irq to become pending. In the usual case, the
2072 * irq will be disabled so it won't deliver an interrupt. */
2073 void xen_poll_irq(int irq)
2075 xen_poll_irq_timeout(irq, 0 /* no timeout */);
2078 /* Check whether the IRQ line is shared with other guests. */
2079 int xen_test_irq_shared(int irq)
2081 struct irq_info *info = info_for_irq(irq);
2082 struct physdev_irq_status_query irq_status;
2087 irq_status.irq = info->u.pirq.pirq;
2089 if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
2091 return !(irq_status.flags & XENIRQSTAT_shared);
2093 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
2095 void xen_irq_resume(void)
2098 struct irq_info *info;
2100 /* New event-channel space is not 'live' yet. */
2101 xen_evtchn_resume();
2103 /* No IRQ <-> event-channel mappings. */
2104 list_for_each_entry(info, &xen_irq_list_head, list) {
2105 /* Zap event-channel binding */
2107 /* Adjust accounting */
2108 channels_on_cpu_dec(info);
2111 clear_evtchn_to_irq_all();
2113 for_each_possible_cpu(cpu) {
2114 restore_cpu_virqs(cpu);
2115 restore_cpu_ipis(cpu);
2121 static struct irq_chip xen_dynamic_chip __read_mostly = {
2124 .irq_disable = disable_dynirq,
2125 .irq_mask = disable_dynirq,
2126 .irq_unmask = enable_dynirq,
2128 .irq_ack = ack_dynirq,
2129 .irq_mask_ack = mask_ack_dynirq,
2131 .irq_set_affinity = set_affinity_irq,
2132 .irq_retrigger = retrigger_dynirq,
2135 static struct irq_chip xen_lateeoi_chip __read_mostly = {
2136 /* The chip name needs to contain "xen-dyn" for irqbalance to work. */
2137 .name = "xen-dyn-lateeoi",
2139 .irq_disable = disable_dynirq,
2140 .irq_mask = disable_dynirq,
2141 .irq_unmask = enable_dynirq,
2143 .irq_ack = lateeoi_ack_dynirq,
2144 .irq_mask_ack = lateeoi_mask_ack_dynirq,
2146 .irq_set_affinity = set_affinity_irq,
2147 .irq_retrigger = retrigger_dynirq,
2150 static struct irq_chip xen_pirq_chip __read_mostly = {
2153 .irq_startup = startup_pirq,
2154 .irq_shutdown = shutdown_pirq,
2155 .irq_enable = enable_pirq,
2156 .irq_disable = disable_pirq,
2158 .irq_mask = disable_dynirq,
2159 .irq_unmask = enable_dynirq,
2161 .irq_ack = eoi_pirq,
2162 .irq_eoi = eoi_pirq,
2163 .irq_mask_ack = mask_ack_pirq,
2165 .irq_set_affinity = set_affinity_irq,
2167 .irq_retrigger = retrigger_dynirq,
2170 static struct irq_chip xen_percpu_chip __read_mostly = {
2171 .name = "xen-percpu",
2173 .irq_disable = disable_dynirq,
2174 .irq_mask = disable_dynirq,
2175 .irq_unmask = enable_dynirq,
2177 .irq_ack = ack_dynirq,
2180 #ifdef CONFIG_XEN_PVHVM
2181 /* Vector callbacks are better than PCI interrupts to receive event
2182 * channel notifications because we can receive vector callbacks on any
2183 * vcpu and we don't need PCI support or APIC interactions. */
2184 void xen_setup_callback_vector(void)
2186 uint64_t callback_via;
2188 if (xen_have_vector_callback) {
2189 callback_via = HVM_CALLBACK_VECTOR(HYPERVISOR_CALLBACK_VECTOR);
2190 if (xen_set_callback_via(callback_via)) {
2191 pr_err("Request for Xen HVM callback vector failed\n");
2192 xen_have_vector_callback = 0;
2197 static __init void xen_alloc_callback_vector(void)
2199 if (!xen_have_vector_callback)
2202 pr_info("Xen HVM callback vector for event delivery is enabled\n");
2203 alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_xen_hvm_callback);
2206 void xen_setup_callback_vector(void) {}
2207 static inline void xen_alloc_callback_vector(void) {}
2210 bool xen_fifo_events = true;
2211 module_param_named(fifo_events, xen_fifo_events, bool, 0);
2213 static int xen_evtchn_cpu_prepare(unsigned int cpu)
2217 xen_cpu_init_eoi(cpu);
2219 if (evtchn_ops->percpu_init)
2220 ret = evtchn_ops->percpu_init(cpu);
2225 static int xen_evtchn_cpu_dead(unsigned int cpu)
2229 if (evtchn_ops->percpu_deinit)
2230 ret = evtchn_ops->percpu_deinit(cpu);
2235 void __init xen_init_IRQ(void)
2238 evtchn_port_t evtchn;
2240 if (xen_fifo_events)
2241 ret = xen_evtchn_fifo_init();
2243 xen_evtchn_2l_init();
2244 xen_fifo_events = false;
2247 xen_cpu_init_eoi(smp_processor_id());
2249 cpuhp_setup_state_nocalls(CPUHP_XEN_EVTCHN_PREPARE,
2250 "xen/evtchn:prepare",
2251 xen_evtchn_cpu_prepare, xen_evtchn_cpu_dead);
2253 evtchn_to_irq = kcalloc(EVTCHN_ROW(xen_evtchn_max_channels()),
2254 sizeof(*evtchn_to_irq), GFP_KERNEL);
2255 BUG_ON(!evtchn_to_irq);
2257 /* No event channels are 'live' right now. */
2258 for (evtchn = 0; evtchn < xen_evtchn_nr_channels(); evtchn++)
2259 mask_evtchn(evtchn);
2261 pirq_needs_eoi = pirq_needs_eoi_flag;
2264 if (xen_pv_domain()) {
2265 if (xen_initial_domain())
2266 pci_xen_initial_domain();
2268 if (xen_feature(XENFEAT_hvm_callback_vector)) {
2269 xen_setup_callback_vector();
2270 xen_alloc_callback_vector();
2273 if (xen_hvm_domain()) {
2275 /* pci_xen_hvm_init must be called after native_init_IRQ so that
2276 * __acpi_register_gsi can point at the right function */
2280 struct physdev_pirq_eoi_gmfn eoi_gmfn;
2282 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
2283 eoi_gmfn.gmfn = virt_to_gfn(pirq_eoi_map);
2284 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
2286 free_page((unsigned long) pirq_eoi_map);
2287 pirq_eoi_map = NULL;
2289 pirq_needs_eoi = pirq_check_eoi_map;