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/cpuid.h>
49 #include <asm/xen/pci.h>
51 #include <asm/sync_bitops.h>
52 #include <asm/xen/hypercall.h>
53 #include <asm/xen/hypervisor.h>
58 #include <xen/xen-ops.h>
59 #include <xen/events.h>
60 #include <xen/interface/xen.h>
61 #include <xen/interface/event_channel.h>
62 #include <xen/interface/hvm/hvm_op.h>
63 #include <xen/interface/hvm/params.h>
64 #include <xen/interface/physdev.h>
65 #include <xen/interface/sched.h>
66 #include <xen/interface/vcpu.h>
67 #include <xen/xenbus.h>
68 #include <asm/hw_irq.h>
70 #include "events_internal.h"
72 #undef MODULE_PARAM_PREFIX
73 #define MODULE_PARAM_PREFIX "xen."
75 /* Interrupt types. */
85 * Packed IRQ information:
86 * type - enum xen_irq_type
87 * event channel - irq->event channel mapping
88 * cpu - cpu this event channel is bound to
89 * index - type-specific information:
90 * PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
91 * guest, or GSI (real passthrough IRQ) of the device.
97 struct list_head list;
98 struct list_head eoi_list;
102 short type; /* type: IRQT_* */
103 u8 mask_reason; /* Why is event channel masked */
104 #define EVT_MASK_REASON_EXPLICIT 0x01
105 #define EVT_MASK_REASON_TEMPORARY 0x02
106 #define EVT_MASK_REASON_EOI_PENDING 0x04
107 u8 is_active; /* Is event just being handled? */
109 evtchn_port_t evtchn; /* event channel */
110 unsigned short cpu; /* cpu bound */
111 unsigned short eoi_cpu; /* EOI must happen on this cpu-1 */
112 unsigned int irq_epoch; /* If eoi_cpu valid: irq_epoch of event */
113 u64 eoi_time; /* Time in jiffies when to EOI. */
122 unsigned char vector;
126 struct xenbus_device *interdomain;
130 #define PIRQ_NEEDS_EOI (1 << 0)
131 #define PIRQ_SHAREABLE (1 << 1)
132 #define PIRQ_MSI_GROUP (1 << 2)
134 static uint __read_mostly event_loop_timeout = 2;
135 module_param(event_loop_timeout, uint, 0644);
137 static uint __read_mostly event_eoi_delay = 10;
138 module_param(event_eoi_delay, uint, 0644);
140 const struct evtchn_ops *evtchn_ops;
143 * This lock protects updates to the following mapping and reference-count
144 * arrays. The lock does not need to be acquired to read the mapping tables.
146 static DEFINE_MUTEX(irq_mapping_update_lock);
149 * Lock protecting event handling loop against removing event channels.
150 * Adding of event channels is no issue as the associated IRQ becomes active
151 * only after everything is setup (before request_[threaded_]irq() the handler
152 * can't be entered for an event, as the event channel will be unmasked only
155 static DEFINE_RWLOCK(evtchn_rwlock);
160 * irq_mapping_update_lock
163 * percpu eoi_list_lock
167 static LIST_HEAD(xen_irq_list_head);
169 /* IRQ <-> VIRQ mapping. */
170 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
172 /* IRQ <-> IPI mapping */
173 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
175 /* Event channel distribution data */
176 static atomic_t channels_on_cpu[NR_CPUS];
178 static int **evtchn_to_irq;
180 static unsigned long *pirq_eoi_map;
182 static bool (*pirq_needs_eoi)(unsigned irq);
184 #define EVTCHN_ROW(e) (e / (PAGE_SIZE/sizeof(**evtchn_to_irq)))
185 #define EVTCHN_COL(e) (e % (PAGE_SIZE/sizeof(**evtchn_to_irq)))
186 #define EVTCHN_PER_ROW (PAGE_SIZE / sizeof(**evtchn_to_irq))
188 /* Xen will never allocate port zero for any purpose. */
189 #define VALID_EVTCHN(chn) ((chn) != 0)
191 static struct irq_info *legacy_info_ptrs[NR_IRQS_LEGACY];
193 static struct irq_chip xen_dynamic_chip;
194 static struct irq_chip xen_lateeoi_chip;
195 static struct irq_chip xen_percpu_chip;
196 static struct irq_chip xen_pirq_chip;
197 static void enable_dynirq(struct irq_data *data);
198 static void disable_dynirq(struct irq_data *data);
200 static DEFINE_PER_CPU(unsigned int, irq_epoch);
202 static void clear_evtchn_to_irq_row(int *evtchn_row)
206 for (col = 0; col < EVTCHN_PER_ROW; col++)
207 WRITE_ONCE(evtchn_row[col], -1);
210 static void clear_evtchn_to_irq_all(void)
214 for (row = 0; row < EVTCHN_ROW(xen_evtchn_max_channels()); row++) {
215 if (evtchn_to_irq[row] == NULL)
217 clear_evtchn_to_irq_row(evtchn_to_irq[row]);
221 static int set_evtchn_to_irq(evtchn_port_t evtchn, unsigned int irq)
227 if (evtchn >= xen_evtchn_max_channels())
230 row = EVTCHN_ROW(evtchn);
231 col = EVTCHN_COL(evtchn);
233 if (evtchn_to_irq[row] == NULL) {
234 /* Unallocated irq entries return -1 anyway */
238 evtchn_row = (int *) __get_free_pages(GFP_KERNEL, 0);
239 if (evtchn_row == NULL)
242 clear_evtchn_to_irq_row(evtchn_row);
245 * We've prepared an empty row for the mapping. If a different
246 * thread was faster inserting it, we can drop ours.
248 if (cmpxchg(&evtchn_to_irq[row], NULL, evtchn_row) != NULL)
249 free_page((unsigned long) evtchn_row);
252 WRITE_ONCE(evtchn_to_irq[row][col], irq);
256 int get_evtchn_to_irq(evtchn_port_t evtchn)
258 if (evtchn >= xen_evtchn_max_channels())
260 if (evtchn_to_irq[EVTCHN_ROW(evtchn)] == NULL)
262 return READ_ONCE(evtchn_to_irq[EVTCHN_ROW(evtchn)][EVTCHN_COL(evtchn)]);
265 /* Get info for IRQ */
266 static struct irq_info *info_for_irq(unsigned irq)
268 if (irq < nr_legacy_irqs())
269 return legacy_info_ptrs[irq];
271 return irq_get_chip_data(irq);
274 static void set_info_for_irq(unsigned int irq, struct irq_info *info)
276 if (irq < nr_legacy_irqs())
277 legacy_info_ptrs[irq] = info;
279 irq_set_chip_data(irq, info);
282 /* Per CPU channel accounting */
283 static void channels_on_cpu_dec(struct irq_info *info)
285 if (!info->is_accounted)
288 info->is_accounted = 0;
290 if (WARN_ON_ONCE(info->cpu >= nr_cpu_ids))
293 WARN_ON_ONCE(!atomic_add_unless(&channels_on_cpu[info->cpu], -1 , 0));
296 static void channels_on_cpu_inc(struct irq_info *info)
298 if (WARN_ON_ONCE(info->cpu >= nr_cpu_ids))
301 if (WARN_ON_ONCE(!atomic_add_unless(&channels_on_cpu[info->cpu], 1,
305 info->is_accounted = 1;
308 /* Constructors for packed IRQ information. */
309 static int xen_irq_info_common_setup(struct irq_info *info,
311 enum xen_irq_type type,
312 evtchn_port_t evtchn,
317 BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
321 info->evtchn = evtchn;
323 info->mask_reason = EVT_MASK_REASON_EXPLICIT;
324 raw_spin_lock_init(&info->lock);
326 ret = set_evtchn_to_irq(evtchn, irq);
330 irq_clear_status_flags(irq, IRQ_NOREQUEST|IRQ_NOAUTOEN);
332 return xen_evtchn_port_setup(evtchn);
335 static int xen_irq_info_evtchn_setup(unsigned irq,
336 evtchn_port_t evtchn,
337 struct xenbus_device *dev)
339 struct irq_info *info = info_for_irq(irq);
342 ret = xen_irq_info_common_setup(info, irq, IRQT_EVTCHN, evtchn, 0);
343 info->u.interdomain = dev;
345 atomic_inc(&dev->event_channels);
350 static int xen_irq_info_ipi_setup(unsigned cpu,
352 evtchn_port_t evtchn,
355 struct irq_info *info = info_for_irq(irq);
359 per_cpu(ipi_to_irq, cpu)[ipi] = irq;
361 return xen_irq_info_common_setup(info, irq, IRQT_IPI, evtchn, 0);
364 static int xen_irq_info_virq_setup(unsigned cpu,
366 evtchn_port_t evtchn,
369 struct irq_info *info = info_for_irq(irq);
373 per_cpu(virq_to_irq, cpu)[virq] = irq;
375 return xen_irq_info_common_setup(info, irq, IRQT_VIRQ, evtchn, 0);
378 static int xen_irq_info_pirq_setup(unsigned irq,
379 evtchn_port_t evtchn,
385 struct irq_info *info = info_for_irq(irq);
387 info->u.pirq.pirq = pirq;
388 info->u.pirq.gsi = gsi;
389 info->u.pirq.domid = domid;
390 info->u.pirq.flags = flags;
392 return xen_irq_info_common_setup(info, irq, IRQT_PIRQ, evtchn, 0);
395 static void xen_irq_info_cleanup(struct irq_info *info)
397 set_evtchn_to_irq(info->evtchn, -1);
398 xen_evtchn_port_remove(info->evtchn, info->cpu);
400 channels_on_cpu_dec(info);
404 * Accessors for packed IRQ information.
406 evtchn_port_t evtchn_from_irq(unsigned irq)
408 const struct irq_info *info = NULL;
410 if (likely(irq < nr_irqs))
411 info = info_for_irq(irq);
418 unsigned int irq_from_evtchn(evtchn_port_t evtchn)
420 return get_evtchn_to_irq(evtchn);
422 EXPORT_SYMBOL_GPL(irq_from_evtchn);
424 int irq_from_virq(unsigned int cpu, unsigned int virq)
426 return per_cpu(virq_to_irq, cpu)[virq];
429 static enum ipi_vector ipi_from_irq(unsigned irq)
431 struct irq_info *info = info_for_irq(irq);
433 BUG_ON(info == NULL);
434 BUG_ON(info->type != IRQT_IPI);
439 static unsigned virq_from_irq(unsigned irq)
441 struct irq_info *info = info_for_irq(irq);
443 BUG_ON(info == NULL);
444 BUG_ON(info->type != IRQT_VIRQ);
449 static unsigned pirq_from_irq(unsigned irq)
451 struct irq_info *info = info_for_irq(irq);
453 BUG_ON(info == NULL);
454 BUG_ON(info->type != IRQT_PIRQ);
456 return info->u.pirq.pirq;
459 static enum xen_irq_type type_from_irq(unsigned irq)
461 return info_for_irq(irq)->type;
464 static unsigned cpu_from_irq(unsigned irq)
466 return info_for_irq(irq)->cpu;
469 unsigned int cpu_from_evtchn(evtchn_port_t evtchn)
471 int irq = get_evtchn_to_irq(evtchn);
475 ret = cpu_from_irq(irq);
480 static void do_mask(struct irq_info *info, u8 reason)
484 raw_spin_lock_irqsave(&info->lock, flags);
486 if (!info->mask_reason)
487 mask_evtchn(info->evtchn);
489 info->mask_reason |= reason;
491 raw_spin_unlock_irqrestore(&info->lock, flags);
494 static void do_unmask(struct irq_info *info, u8 reason)
498 raw_spin_lock_irqsave(&info->lock, flags);
500 info->mask_reason &= ~reason;
502 if (!info->mask_reason)
503 unmask_evtchn(info->evtchn);
505 raw_spin_unlock_irqrestore(&info->lock, flags);
509 static bool pirq_check_eoi_map(unsigned irq)
511 return test_bit(pirq_from_irq(irq), pirq_eoi_map);
515 static bool pirq_needs_eoi_flag(unsigned irq)
517 struct irq_info *info = info_for_irq(irq);
518 BUG_ON(info->type != IRQT_PIRQ);
520 return info->u.pirq.flags & PIRQ_NEEDS_EOI;
523 static void bind_evtchn_to_cpu(evtchn_port_t evtchn, unsigned int cpu,
526 int irq = get_evtchn_to_irq(evtchn);
527 struct irq_info *info = info_for_irq(irq);
531 if (IS_ENABLED(CONFIG_SMP) && force_affinity) {
532 cpumask_copy(irq_get_affinity_mask(irq), cpumask_of(cpu));
533 cpumask_copy(irq_get_effective_affinity_mask(irq),
537 xen_evtchn_port_bind_to_cpu(evtchn, cpu, info->cpu);
539 channels_on_cpu_dec(info);
541 channels_on_cpu_inc(info);
545 * notify_remote_via_irq - send event to remote end of event channel via irq
546 * @irq: irq of event channel to send event to
548 * Unlike notify_remote_via_evtchn(), this is safe to use across
549 * save/restore. Notifications on a broken connection are silently
552 void notify_remote_via_irq(int irq)
554 evtchn_port_t evtchn = evtchn_from_irq(irq);
556 if (VALID_EVTCHN(evtchn))
557 notify_remote_via_evtchn(evtchn);
559 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
561 struct lateeoi_work {
562 struct delayed_work delayed;
563 spinlock_t eoi_list_lock;
564 struct list_head eoi_list;
567 static DEFINE_PER_CPU(struct lateeoi_work, lateeoi);
569 static void lateeoi_list_del(struct irq_info *info)
571 struct lateeoi_work *eoi = &per_cpu(lateeoi, info->eoi_cpu);
574 spin_lock_irqsave(&eoi->eoi_list_lock, flags);
575 list_del_init(&info->eoi_list);
576 spin_unlock_irqrestore(&eoi->eoi_list_lock, flags);
579 static void lateeoi_list_add(struct irq_info *info)
581 struct lateeoi_work *eoi = &per_cpu(lateeoi, info->eoi_cpu);
582 struct irq_info *elem;
583 u64 now = get_jiffies_64();
587 if (now < info->eoi_time)
588 delay = info->eoi_time - now;
592 spin_lock_irqsave(&eoi->eoi_list_lock, flags);
594 if (list_empty(&eoi->eoi_list)) {
595 list_add(&info->eoi_list, &eoi->eoi_list);
596 mod_delayed_work_on(info->eoi_cpu, system_wq,
597 &eoi->delayed, delay);
599 list_for_each_entry_reverse(elem, &eoi->eoi_list, eoi_list) {
600 if (elem->eoi_time <= info->eoi_time)
603 list_add(&info->eoi_list, &elem->eoi_list);
606 spin_unlock_irqrestore(&eoi->eoi_list_lock, flags);
609 static void xen_irq_lateeoi_locked(struct irq_info *info, bool spurious)
611 evtchn_port_t evtchn;
613 unsigned int delay = 0;
615 evtchn = info->evtchn;
616 if (!VALID_EVTCHN(evtchn) || !list_empty(&info->eoi_list))
620 struct xenbus_device *dev = info->u.interdomain;
621 unsigned int threshold = 1;
623 if (dev && dev->spurious_threshold)
624 threshold = dev->spurious_threshold;
626 if ((1 << info->spurious_cnt) < (HZ << 2)) {
627 if (info->spurious_cnt != 0xFF)
628 info->spurious_cnt++;
630 if (info->spurious_cnt > threshold) {
631 delay = 1 << (info->spurious_cnt - 1 - threshold);
635 info->eoi_cpu = smp_processor_id();
636 info->eoi_time = get_jiffies_64() + delay;
638 atomic_add(delay, &dev->jiffies_eoi_delayed);
641 atomic_inc(&dev->spurious_events);
643 info->spurious_cnt = 0;
647 if (info->eoi_time &&
648 (info->irq_epoch == per_cpu(irq_epoch, cpu) || delay)) {
649 lateeoi_list_add(info);
655 /* is_active hasn't been reset yet, do it now. */
656 smp_store_release(&info->is_active, 0);
657 do_unmask(info, EVT_MASK_REASON_EOI_PENDING);
660 static void xen_irq_lateeoi_worker(struct work_struct *work)
662 struct lateeoi_work *eoi;
663 struct irq_info *info;
664 u64 now = get_jiffies_64();
667 eoi = container_of(to_delayed_work(work), struct lateeoi_work, delayed);
669 read_lock_irqsave(&evtchn_rwlock, flags);
672 spin_lock(&eoi->eoi_list_lock);
674 info = list_first_entry_or_null(&eoi->eoi_list, struct irq_info,
677 if (info == NULL || now < info->eoi_time) {
678 spin_unlock(&eoi->eoi_list_lock);
682 list_del_init(&info->eoi_list);
684 spin_unlock(&eoi->eoi_list_lock);
688 xen_irq_lateeoi_locked(info, false);
692 mod_delayed_work_on(info->eoi_cpu, system_wq,
693 &eoi->delayed, info->eoi_time - now);
695 read_unlock_irqrestore(&evtchn_rwlock, flags);
698 static void xen_cpu_init_eoi(unsigned int cpu)
700 struct lateeoi_work *eoi = &per_cpu(lateeoi, cpu);
702 INIT_DELAYED_WORK(&eoi->delayed, xen_irq_lateeoi_worker);
703 spin_lock_init(&eoi->eoi_list_lock);
704 INIT_LIST_HEAD(&eoi->eoi_list);
707 void xen_irq_lateeoi(unsigned int irq, unsigned int eoi_flags)
709 struct irq_info *info;
712 read_lock_irqsave(&evtchn_rwlock, flags);
714 info = info_for_irq(irq);
717 xen_irq_lateeoi_locked(info, eoi_flags & XEN_EOI_FLAG_SPURIOUS);
719 read_unlock_irqrestore(&evtchn_rwlock, flags);
721 EXPORT_SYMBOL_GPL(xen_irq_lateeoi);
723 static void xen_irq_init(unsigned irq)
725 struct irq_info *info;
727 info = kzalloc(sizeof(*info), GFP_KERNEL);
729 panic("Unable to allocate metadata for IRQ%d\n", irq);
731 info->type = IRQT_UNBOUND;
734 set_info_for_irq(irq, info);
736 * Interrupt affinity setting can be immediate. No point
737 * in delaying it until an interrupt is handled.
739 irq_set_status_flags(irq, IRQ_MOVE_PCNTXT);
741 INIT_LIST_HEAD(&info->eoi_list);
742 list_add_tail(&info->list, &xen_irq_list_head);
745 static int __must_check xen_allocate_irqs_dynamic(int nvec)
747 int i, irq = irq_alloc_descs(-1, 0, nvec, -1);
750 for (i = 0; i < nvec; i++)
751 xen_irq_init(irq + i);
757 static inline int __must_check xen_allocate_irq_dynamic(void)
760 return xen_allocate_irqs_dynamic(1);
763 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
768 * A PV guest has no concept of a GSI (since it has no ACPI
769 * nor access to/knowledge of the physical APICs). Therefore
770 * all IRQs are dynamically allocated from the entire IRQ
773 if (xen_pv_domain() && !xen_initial_domain())
774 return xen_allocate_irq_dynamic();
776 /* Legacy IRQ descriptors are already allocated by the arch. */
777 if (gsi < nr_legacy_irqs())
780 irq = irq_alloc_desc_at(gsi, -1);
787 static void xen_free_irq(unsigned irq)
789 struct irq_info *info = info_for_irq(irq);
795 write_lock_irqsave(&evtchn_rwlock, flags);
797 if (!list_empty(&info->eoi_list))
798 lateeoi_list_del(info);
800 list_del(&info->list);
802 set_info_for_irq(irq, NULL);
804 WARN_ON(info->refcnt > 0);
806 write_unlock_irqrestore(&evtchn_rwlock, flags);
810 /* Legacy IRQ descriptors are managed by the arch. */
811 if (irq < nr_legacy_irqs())
817 static void xen_evtchn_close(evtchn_port_t port)
819 struct evtchn_close close;
822 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
826 /* Not called for lateeoi events. */
827 static void event_handler_exit(struct irq_info *info)
829 smp_store_release(&info->is_active, 0);
830 clear_evtchn(info->evtchn);
833 static void pirq_query_unmask(int irq)
835 struct physdev_irq_status_query irq_status;
836 struct irq_info *info = info_for_irq(irq);
838 BUG_ON(info->type != IRQT_PIRQ);
840 irq_status.irq = pirq_from_irq(irq);
841 if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
842 irq_status.flags = 0;
844 info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
845 if (irq_status.flags & XENIRQSTAT_needs_eoi)
846 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
849 static void eoi_pirq(struct irq_data *data)
851 struct irq_info *info = info_for_irq(data->irq);
852 evtchn_port_t evtchn = info ? info->evtchn : 0;
853 struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
856 if (!VALID_EVTCHN(evtchn))
859 event_handler_exit(info);
861 if (pirq_needs_eoi(data->irq)) {
862 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
867 static void mask_ack_pirq(struct irq_data *data)
869 disable_dynirq(data);
873 static unsigned int __startup_pirq(unsigned int irq)
875 struct evtchn_bind_pirq bind_pirq;
876 struct irq_info *info = info_for_irq(irq);
877 evtchn_port_t evtchn = evtchn_from_irq(irq);
880 BUG_ON(info->type != IRQT_PIRQ);
882 if (VALID_EVTCHN(evtchn))
885 bind_pirq.pirq = pirq_from_irq(irq);
886 /* NB. We are happy to share unless we are probing. */
887 bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
888 BIND_PIRQ__WILL_SHARE : 0;
889 rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
891 pr_warn("Failed to obtain physical IRQ %d\n", irq);
894 evtchn = bind_pirq.port;
896 pirq_query_unmask(irq);
898 rc = set_evtchn_to_irq(evtchn, irq);
902 info->evtchn = evtchn;
903 bind_evtchn_to_cpu(evtchn, 0, false);
905 rc = xen_evtchn_port_setup(evtchn);
910 do_unmask(info, EVT_MASK_REASON_EXPLICIT);
912 eoi_pirq(irq_get_irq_data(irq));
917 pr_err("irq%d: Failed to set port to irq mapping (%d)\n", irq, rc);
918 xen_evtchn_close(evtchn);
922 static unsigned int startup_pirq(struct irq_data *data)
924 return __startup_pirq(data->irq);
927 static void shutdown_pirq(struct irq_data *data)
929 unsigned int irq = data->irq;
930 struct irq_info *info = info_for_irq(irq);
931 evtchn_port_t evtchn = evtchn_from_irq(irq);
933 BUG_ON(info->type != IRQT_PIRQ);
935 if (!VALID_EVTCHN(evtchn))
938 do_mask(info, EVT_MASK_REASON_EXPLICIT);
939 xen_evtchn_close(evtchn);
940 xen_irq_info_cleanup(info);
943 static void enable_pirq(struct irq_data *data)
948 static void disable_pirq(struct irq_data *data)
950 disable_dynirq(data);
953 int xen_irq_from_gsi(unsigned gsi)
955 struct irq_info *info;
957 list_for_each_entry(info, &xen_irq_list_head, list) {
958 if (info->type != IRQT_PIRQ)
961 if (info->u.pirq.gsi == gsi)
967 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
969 static void __unbind_from_irq(unsigned int irq)
971 evtchn_port_t evtchn = evtchn_from_irq(irq);
972 struct irq_info *info = info_for_irq(irq);
974 if (info->refcnt > 0) {
976 if (info->refcnt != 0)
980 if (VALID_EVTCHN(evtchn)) {
981 unsigned int cpu = cpu_from_irq(irq);
982 struct xenbus_device *dev;
984 xen_evtchn_close(evtchn);
986 switch (type_from_irq(irq)) {
988 per_cpu(virq_to_irq, cpu)[virq_from_irq(irq)] = -1;
991 per_cpu(ipi_to_irq, cpu)[ipi_from_irq(irq)] = -1;
994 dev = info->u.interdomain;
996 atomic_dec(&dev->event_channels);
1002 xen_irq_info_cleanup(info);
1009 * Do not make any assumptions regarding the relationship between the
1010 * IRQ number returned here and the Xen pirq argument.
1012 * Note: We don't assign an event channel until the irq actually started
1013 * up. Return an existing irq if we've already got one for the gsi.
1015 * Shareable implies level triggered, not shareable implies edge
1018 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
1019 unsigned pirq, int shareable, char *name)
1022 struct physdev_irq irq_op;
1025 mutex_lock(&irq_mapping_update_lock);
1027 irq = xen_irq_from_gsi(gsi);
1029 pr_info("%s: returning irq %d for gsi %u\n",
1030 __func__, irq, gsi);
1034 irq = xen_allocate_irq_gsi(gsi);
1041 /* Only the privileged domain can do this. For non-priv, the pcifront
1042 * driver provides a PCI bus that does the call to do exactly
1043 * this in the priv domain. */
1044 if (xen_initial_domain() &&
1045 HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
1051 ret = xen_irq_info_pirq_setup(irq, 0, pirq, gsi, DOMID_SELF,
1052 shareable ? PIRQ_SHAREABLE : 0);
1054 __unbind_from_irq(irq);
1059 pirq_query_unmask(irq);
1060 /* We try to use the handler with the appropriate semantic for the
1061 * type of interrupt: if the interrupt is an edge triggered
1062 * interrupt we use handle_edge_irq.
1064 * On the other hand if the interrupt is level triggered we use
1065 * handle_fasteoi_irq like the native code does for this kind of
1068 * Depending on the Xen version, pirq_needs_eoi might return true
1069 * not only for level triggered interrupts but for edge triggered
1070 * interrupts too. In any case Xen always honors the eoi mechanism,
1071 * not injecting any more pirqs of the same kind if the first one
1072 * hasn't received an eoi yet. Therefore using the fasteoi handler
1073 * is the right choice either way.
1076 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
1077 handle_fasteoi_irq, name);
1079 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
1080 handle_edge_irq, name);
1083 mutex_unlock(&irq_mapping_update_lock);
1088 #ifdef CONFIG_PCI_MSI
1089 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
1092 struct physdev_get_free_pirq op_get_free_pirq;
1094 op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
1095 rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
1097 WARN_ONCE(rc == -ENOSYS,
1098 "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
1100 return rc ? -1 : op_get_free_pirq.pirq;
1103 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
1104 int pirq, int nvec, const char *name, domid_t domid)
1108 mutex_lock(&irq_mapping_update_lock);
1110 irq = xen_allocate_irqs_dynamic(nvec);
1114 for (i = 0; i < nvec; i++) {
1115 irq_set_chip_and_handler_name(irq + i, &xen_pirq_chip, handle_edge_irq, name);
1117 ret = xen_irq_info_pirq_setup(irq + i, 0, pirq + i, 0, domid,
1118 i == 0 ? 0 : PIRQ_MSI_GROUP);
1123 ret = irq_set_msi_desc(irq, msidesc);
1127 mutex_unlock(&irq_mapping_update_lock);
1131 __unbind_from_irq(irq + nvec);
1132 mutex_unlock(&irq_mapping_update_lock);
1137 int xen_destroy_irq(int irq)
1139 struct physdev_unmap_pirq unmap_irq;
1140 struct irq_info *info = info_for_irq(irq);
1143 mutex_lock(&irq_mapping_update_lock);
1146 * If trying to remove a vector in a MSI group different
1147 * than the first one skip the PIRQ unmap unless this vector
1148 * is the first one in the group.
1150 if (xen_initial_domain() && !(info->u.pirq.flags & PIRQ_MSI_GROUP)) {
1151 unmap_irq.pirq = info->u.pirq.pirq;
1152 unmap_irq.domid = info->u.pirq.domid;
1153 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
1154 /* If another domain quits without making the pci_disable_msix
1155 * call, the Xen hypervisor takes care of freeing the PIRQs
1156 * (free_domain_pirqs).
1158 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
1159 pr_info("domain %d does not have %d anymore\n",
1160 info->u.pirq.domid, info->u.pirq.pirq);
1162 pr_warn("unmap irq failed %d\n", rc);
1170 mutex_unlock(&irq_mapping_update_lock);
1174 int xen_irq_from_pirq(unsigned pirq)
1178 struct irq_info *info;
1180 mutex_lock(&irq_mapping_update_lock);
1182 list_for_each_entry(info, &xen_irq_list_head, list) {
1183 if (info->type != IRQT_PIRQ)
1186 if (info->u.pirq.pirq == pirq)
1191 mutex_unlock(&irq_mapping_update_lock);
1197 int xen_pirq_from_irq(unsigned irq)
1199 return pirq_from_irq(irq);
1201 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
1203 static int bind_evtchn_to_irq_chip(evtchn_port_t evtchn, struct irq_chip *chip,
1204 struct xenbus_device *dev)
1209 if (evtchn >= xen_evtchn_max_channels())
1212 mutex_lock(&irq_mapping_update_lock);
1214 irq = get_evtchn_to_irq(evtchn);
1217 irq = xen_allocate_irq_dynamic();
1221 irq_set_chip_and_handler_name(irq, chip,
1222 handle_edge_irq, "event");
1224 ret = xen_irq_info_evtchn_setup(irq, evtchn, dev);
1226 __unbind_from_irq(irq);
1231 * New interdomain events are initially bound to vCPU0 This
1232 * is required to setup the event channel in the first
1233 * place and also important for UP guests because the
1234 * affinity setting is not invoked on them so nothing would
1237 bind_evtchn_to_cpu(evtchn, 0, false);
1239 struct irq_info *info = info_for_irq(irq);
1240 WARN_ON(info == NULL || info->type != IRQT_EVTCHN);
1244 mutex_unlock(&irq_mapping_update_lock);
1249 int bind_evtchn_to_irq(evtchn_port_t evtchn)
1251 return bind_evtchn_to_irq_chip(evtchn, &xen_dynamic_chip, NULL);
1253 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
1255 int bind_evtchn_to_irq_lateeoi(evtchn_port_t evtchn)
1257 return bind_evtchn_to_irq_chip(evtchn, &xen_lateeoi_chip, NULL);
1259 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq_lateeoi);
1261 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
1263 struct evtchn_bind_ipi bind_ipi;
1264 evtchn_port_t evtchn;
1267 mutex_lock(&irq_mapping_update_lock);
1269 irq = per_cpu(ipi_to_irq, cpu)[ipi];
1272 irq = xen_allocate_irq_dynamic();
1276 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
1277 handle_percpu_irq, "ipi");
1279 bind_ipi.vcpu = xen_vcpu_nr(cpu);
1280 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1283 evtchn = bind_ipi.port;
1285 ret = xen_irq_info_ipi_setup(cpu, irq, evtchn, ipi);
1287 __unbind_from_irq(irq);
1292 * Force the affinity mask to the target CPU so proc shows
1293 * the correct target.
1295 bind_evtchn_to_cpu(evtchn, cpu, true);
1297 struct irq_info *info = info_for_irq(irq);
1298 WARN_ON(info == NULL || info->type != IRQT_IPI);
1302 mutex_unlock(&irq_mapping_update_lock);
1306 static int bind_interdomain_evtchn_to_irq_chip(struct xenbus_device *dev,
1307 evtchn_port_t remote_port,
1308 struct irq_chip *chip)
1310 struct evtchn_bind_interdomain bind_interdomain;
1313 bind_interdomain.remote_dom = dev->otherend_id;
1314 bind_interdomain.remote_port = remote_port;
1316 err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
1319 return err ? : bind_evtchn_to_irq_chip(bind_interdomain.local_port,
1323 int bind_interdomain_evtchn_to_irq_lateeoi(struct xenbus_device *dev,
1324 evtchn_port_t remote_port)
1326 return bind_interdomain_evtchn_to_irq_chip(dev, remote_port,
1329 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irq_lateeoi);
1331 static int find_virq(unsigned int virq, unsigned int cpu, evtchn_port_t *evtchn)
1333 struct evtchn_status status;
1337 memset(&status, 0, sizeof(status));
1338 for (port = 0; port < xen_evtchn_max_channels(); port++) {
1339 status.dom = DOMID_SELF;
1341 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
1344 if (status.status != EVTCHNSTAT_virq)
1346 if (status.u.virq == virq && status.vcpu == xen_vcpu_nr(cpu)) {
1355 * xen_evtchn_nr_channels - number of usable event channel ports
1357 * This may be less than the maximum supported by the current
1358 * hypervisor ABI. Use xen_evtchn_max_channels() for the maximum
1361 unsigned xen_evtchn_nr_channels(void)
1363 return evtchn_ops->nr_channels();
1365 EXPORT_SYMBOL_GPL(xen_evtchn_nr_channels);
1367 int bind_virq_to_irq(unsigned int virq, unsigned int cpu, bool percpu)
1369 struct evtchn_bind_virq bind_virq;
1370 evtchn_port_t evtchn = 0;
1373 mutex_lock(&irq_mapping_update_lock);
1375 irq = per_cpu(virq_to_irq, cpu)[virq];
1378 irq = xen_allocate_irq_dynamic();
1383 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
1384 handle_percpu_irq, "virq");
1386 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
1387 handle_edge_irq, "virq");
1389 bind_virq.virq = virq;
1390 bind_virq.vcpu = xen_vcpu_nr(cpu);
1391 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1394 evtchn = bind_virq.port;
1397 ret = find_virq(virq, cpu, &evtchn);
1401 ret = xen_irq_info_virq_setup(cpu, irq, evtchn, virq);
1403 __unbind_from_irq(irq);
1409 * Force the affinity mask for percpu interrupts so proc
1410 * shows the correct target.
1412 bind_evtchn_to_cpu(evtchn, cpu, percpu);
1414 struct irq_info *info = info_for_irq(irq);
1415 WARN_ON(info == NULL || info->type != IRQT_VIRQ);
1419 mutex_unlock(&irq_mapping_update_lock);
1424 static void unbind_from_irq(unsigned int irq)
1426 mutex_lock(&irq_mapping_update_lock);
1427 __unbind_from_irq(irq);
1428 mutex_unlock(&irq_mapping_update_lock);
1431 static int bind_evtchn_to_irqhandler_chip(evtchn_port_t evtchn,
1432 irq_handler_t handler,
1433 unsigned long irqflags,
1434 const char *devname, void *dev_id,
1435 struct irq_chip *chip)
1439 irq = bind_evtchn_to_irq_chip(evtchn, chip, NULL);
1442 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1444 unbind_from_irq(irq);
1451 int bind_evtchn_to_irqhandler(evtchn_port_t evtchn,
1452 irq_handler_t handler,
1453 unsigned long irqflags,
1454 const char *devname, void *dev_id)
1456 return bind_evtchn_to_irqhandler_chip(evtchn, handler, irqflags,
1460 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1462 int bind_evtchn_to_irqhandler_lateeoi(evtchn_port_t evtchn,
1463 irq_handler_t handler,
1464 unsigned long irqflags,
1465 const char *devname, void *dev_id)
1467 return bind_evtchn_to_irqhandler_chip(evtchn, handler, irqflags,
1471 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler_lateeoi);
1473 static int bind_interdomain_evtchn_to_irqhandler_chip(
1474 struct xenbus_device *dev, evtchn_port_t remote_port,
1475 irq_handler_t handler, unsigned long irqflags,
1476 const char *devname, void *dev_id, struct irq_chip *chip)
1480 irq = bind_interdomain_evtchn_to_irq_chip(dev, remote_port, chip);
1484 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1486 unbind_from_irq(irq);
1493 int bind_interdomain_evtchn_to_irqhandler_lateeoi(struct xenbus_device *dev,
1494 evtchn_port_t remote_port,
1495 irq_handler_t handler,
1496 unsigned long irqflags,
1497 const char *devname,
1500 return bind_interdomain_evtchn_to_irqhandler_chip(dev,
1501 remote_port, handler, irqflags, devname,
1502 dev_id, &xen_lateeoi_chip);
1504 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler_lateeoi);
1506 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1507 irq_handler_t handler,
1508 unsigned long irqflags, const char *devname, void *dev_id)
1512 irq = bind_virq_to_irq(virq, cpu, irqflags & IRQF_PERCPU);
1515 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1517 unbind_from_irq(irq);
1523 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1525 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1527 irq_handler_t handler,
1528 unsigned long irqflags,
1529 const char *devname,
1534 irq = bind_ipi_to_irq(ipi, cpu);
1538 irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1539 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1541 unbind_from_irq(irq);
1548 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1550 struct irq_info *info = info_for_irq(irq);
1554 free_irq(irq, dev_id);
1555 unbind_from_irq(irq);
1557 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1560 * xen_set_irq_priority() - set an event channel priority.
1561 * @irq:irq bound to an event channel.
1562 * @priority: priority between XEN_IRQ_PRIORITY_MAX and XEN_IRQ_PRIORITY_MIN.
1564 int xen_set_irq_priority(unsigned irq, unsigned priority)
1566 struct evtchn_set_priority set_priority;
1568 set_priority.port = evtchn_from_irq(irq);
1569 set_priority.priority = priority;
1571 return HYPERVISOR_event_channel_op(EVTCHNOP_set_priority,
1574 EXPORT_SYMBOL_GPL(xen_set_irq_priority);
1576 int evtchn_make_refcounted(evtchn_port_t evtchn)
1578 int irq = get_evtchn_to_irq(evtchn);
1579 struct irq_info *info;
1584 info = info_for_irq(irq);
1589 WARN_ON(info->refcnt != -1);
1595 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1597 int evtchn_get(evtchn_port_t evtchn)
1600 struct irq_info *info;
1603 if (evtchn >= xen_evtchn_max_channels())
1606 mutex_lock(&irq_mapping_update_lock);
1608 irq = get_evtchn_to_irq(evtchn);
1612 info = info_for_irq(irq);
1618 if (info->refcnt <= 0 || info->refcnt == SHRT_MAX)
1624 mutex_unlock(&irq_mapping_update_lock);
1628 EXPORT_SYMBOL_GPL(evtchn_get);
1630 void evtchn_put(evtchn_port_t evtchn)
1632 int irq = get_evtchn_to_irq(evtchn);
1633 if (WARN_ON(irq == -1))
1635 unbind_from_irq(irq);
1637 EXPORT_SYMBOL_GPL(evtchn_put);
1639 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1644 if (unlikely(vector == XEN_NMI_VECTOR)) {
1645 int rc = HYPERVISOR_vcpu_op(VCPUOP_send_nmi, xen_vcpu_nr(cpu),
1648 printk(KERN_WARNING "Sending nmi to CPU%d failed (rc:%d)\n", cpu, rc);
1652 irq = per_cpu(ipi_to_irq, cpu)[vector];
1654 notify_remote_via_irq(irq);
1657 struct evtchn_loop_ctrl {
1663 void handle_irq_for_port(evtchn_port_t port, struct evtchn_loop_ctrl *ctrl)
1666 struct irq_info *info;
1667 struct xenbus_device *dev;
1669 irq = get_evtchn_to_irq(port);
1674 * Check for timeout every 256 events.
1675 * We are setting the timeout value only after the first 256
1676 * events in order to not hurt the common case of few loop
1677 * iterations. The 256 is basically an arbitrary value.
1679 * In case we are hitting the timeout we need to defer all further
1680 * EOIs in order to ensure to leave the event handling loop rather
1681 * sooner than later.
1683 if (!ctrl->defer_eoi && !(++ctrl->count & 0xff)) {
1684 ktime_t kt = ktime_get();
1686 if (!ctrl->timeout) {
1687 kt = ktime_add_ms(kt,
1688 jiffies_to_msecs(event_loop_timeout));
1690 } else if (kt > ctrl->timeout) {
1691 ctrl->defer_eoi = true;
1695 info = info_for_irq(irq);
1696 if (xchg_acquire(&info->is_active, 1))
1699 dev = (info->type == IRQT_EVTCHN) ? info->u.interdomain : NULL;
1701 atomic_inc(&dev->events);
1703 if (ctrl->defer_eoi) {
1704 info->eoi_cpu = smp_processor_id();
1705 info->irq_epoch = __this_cpu_read(irq_epoch);
1706 info->eoi_time = get_jiffies_64() + event_eoi_delay;
1709 generic_handle_irq(irq);
1712 static void __xen_evtchn_do_upcall(void)
1714 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1715 int cpu = smp_processor_id();
1716 struct evtchn_loop_ctrl ctrl = { 0 };
1718 read_lock(&evtchn_rwlock);
1721 vcpu_info->evtchn_upcall_pending = 0;
1723 xen_evtchn_handle_events(cpu, &ctrl);
1725 BUG_ON(!irqs_disabled());
1727 virt_rmb(); /* Hypervisor can set upcall pending. */
1729 } while (vcpu_info->evtchn_upcall_pending);
1731 read_unlock(&evtchn_rwlock);
1734 * Increment irq_epoch only now to defer EOIs only for
1735 * xen_irq_lateeoi() invocations occurring from inside the loop
1738 __this_cpu_inc(irq_epoch);
1741 void xen_evtchn_do_upcall(struct pt_regs *regs)
1743 struct pt_regs *old_regs = set_irq_regs(regs);
1747 __xen_evtchn_do_upcall();
1750 set_irq_regs(old_regs);
1753 void xen_hvm_evtchn_do_upcall(void)
1755 __xen_evtchn_do_upcall();
1757 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1759 /* Rebind a new event channel to an existing irq. */
1760 void rebind_evtchn_irq(evtchn_port_t evtchn, int irq)
1762 struct irq_info *info = info_for_irq(irq);
1767 /* Make sure the irq is masked, since the new event channel
1768 will also be masked. */
1771 mutex_lock(&irq_mapping_update_lock);
1773 /* After resume the irq<->evtchn mappings are all cleared out */
1774 BUG_ON(get_evtchn_to_irq(evtchn) != -1);
1775 /* Expect irq to have been bound before,
1776 so there should be a proper type */
1777 BUG_ON(info->type == IRQT_UNBOUND);
1779 (void)xen_irq_info_evtchn_setup(irq, evtchn, NULL);
1781 mutex_unlock(&irq_mapping_update_lock);
1783 bind_evtchn_to_cpu(evtchn, info->cpu, false);
1785 /* Unmask the event channel. */
1789 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1790 static int xen_rebind_evtchn_to_cpu(struct irq_info *info, unsigned int tcpu)
1792 struct evtchn_bind_vcpu bind_vcpu;
1793 evtchn_port_t evtchn = info ? info->evtchn : 0;
1795 if (!VALID_EVTCHN(evtchn))
1798 if (!xen_support_evtchn_rebind())
1801 /* Send future instances of this interrupt to other vcpu. */
1802 bind_vcpu.port = evtchn;
1803 bind_vcpu.vcpu = xen_vcpu_nr(tcpu);
1806 * Mask the event while changing the VCPU binding to prevent
1807 * it being delivered on an unexpected VCPU.
1809 do_mask(info, EVT_MASK_REASON_TEMPORARY);
1812 * If this fails, it usually just indicates that we're dealing with a
1813 * virq or IPI channel, which don't actually need to be rebound. Ignore
1814 * it, but don't do the xenlinux-level rebind in that case.
1816 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1817 bind_evtchn_to_cpu(evtchn, tcpu, false);
1819 do_unmask(info, EVT_MASK_REASON_TEMPORARY);
1825 * Find the CPU within @dest mask which has the least number of channels
1826 * assigned. This is not precise as the per cpu counts can be modified
1829 static unsigned int select_target_cpu(const struct cpumask *dest)
1831 unsigned int cpu, best_cpu = UINT_MAX, minch = UINT_MAX;
1833 for_each_cpu_and(cpu, dest, cpu_online_mask) {
1834 unsigned int curch = atomic_read(&channels_on_cpu[cpu]);
1836 if (curch < minch) {
1843 * Catch the unlikely case that dest contains no online CPUs. Can't
1846 if (best_cpu == UINT_MAX)
1847 return select_target_cpu(cpu_online_mask);
1852 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1855 unsigned int tcpu = select_target_cpu(dest);
1858 ret = xen_rebind_evtchn_to_cpu(info_for_irq(data->irq), tcpu);
1860 irq_data_update_effective_affinity(data, cpumask_of(tcpu));
1865 static void enable_dynirq(struct irq_data *data)
1867 struct irq_info *info = info_for_irq(data->irq);
1868 evtchn_port_t evtchn = info ? info->evtchn : 0;
1870 if (VALID_EVTCHN(evtchn))
1871 do_unmask(info, EVT_MASK_REASON_EXPLICIT);
1874 static void disable_dynirq(struct irq_data *data)
1876 struct irq_info *info = info_for_irq(data->irq);
1877 evtchn_port_t evtchn = info ? info->evtchn : 0;
1879 if (VALID_EVTCHN(evtchn))
1880 do_mask(info, EVT_MASK_REASON_EXPLICIT);
1883 static void ack_dynirq(struct irq_data *data)
1885 struct irq_info *info = info_for_irq(data->irq);
1886 evtchn_port_t evtchn = info ? info->evtchn : 0;
1888 if (VALID_EVTCHN(evtchn))
1889 event_handler_exit(info);
1892 static void mask_ack_dynirq(struct irq_data *data)
1894 disable_dynirq(data);
1898 static void lateeoi_ack_dynirq(struct irq_data *data)
1900 struct irq_info *info = info_for_irq(data->irq);
1901 evtchn_port_t evtchn = info ? info->evtchn : 0;
1903 if (VALID_EVTCHN(evtchn)) {
1904 do_mask(info, EVT_MASK_REASON_EOI_PENDING);
1906 * Don't call event_handler_exit().
1907 * Need to keep is_active non-zero in order to ignore re-raised
1908 * events after cpu affinity changes while a lateeoi is pending.
1910 clear_evtchn(evtchn);
1914 static void lateeoi_mask_ack_dynirq(struct irq_data *data)
1916 struct irq_info *info = info_for_irq(data->irq);
1917 evtchn_port_t evtchn = info ? info->evtchn : 0;
1919 if (VALID_EVTCHN(evtchn)) {
1920 do_mask(info, EVT_MASK_REASON_EXPLICIT);
1921 event_handler_exit(info);
1925 static int retrigger_dynirq(struct irq_data *data)
1927 struct irq_info *info = info_for_irq(data->irq);
1928 evtchn_port_t evtchn = info ? info->evtchn : 0;
1930 if (!VALID_EVTCHN(evtchn))
1933 do_mask(info, EVT_MASK_REASON_TEMPORARY);
1935 do_unmask(info, EVT_MASK_REASON_TEMPORARY);
1940 static void restore_pirqs(void)
1942 int pirq, rc, irq, gsi;
1943 struct physdev_map_pirq map_irq;
1944 struct irq_info *info;
1946 list_for_each_entry(info, &xen_irq_list_head, list) {
1947 if (info->type != IRQT_PIRQ)
1950 pirq = info->u.pirq.pirq;
1951 gsi = info->u.pirq.gsi;
1954 /* save/restore of PT devices doesn't work, so at this point the
1955 * only devices present are GSI based emulated devices */
1959 map_irq.domid = DOMID_SELF;
1960 map_irq.type = MAP_PIRQ_TYPE_GSI;
1961 map_irq.index = gsi;
1962 map_irq.pirq = pirq;
1964 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1966 pr_warn("xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1967 gsi, irq, pirq, rc);
1972 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1974 __startup_pirq(irq);
1978 static void restore_cpu_virqs(unsigned int cpu)
1980 struct evtchn_bind_virq bind_virq;
1981 evtchn_port_t evtchn;
1984 for (virq = 0; virq < NR_VIRQS; virq++) {
1985 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1988 BUG_ON(virq_from_irq(irq) != virq);
1990 /* Get a new binding from Xen. */
1991 bind_virq.virq = virq;
1992 bind_virq.vcpu = xen_vcpu_nr(cpu);
1993 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1996 evtchn = bind_virq.port;
1998 /* Record the new mapping. */
1999 (void)xen_irq_info_virq_setup(cpu, irq, evtchn, virq);
2000 /* The affinity mask is still valid */
2001 bind_evtchn_to_cpu(evtchn, cpu, false);
2005 static void restore_cpu_ipis(unsigned int cpu)
2007 struct evtchn_bind_ipi bind_ipi;
2008 evtchn_port_t evtchn;
2011 for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
2012 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
2015 BUG_ON(ipi_from_irq(irq) != ipi);
2017 /* Get a new binding from Xen. */
2018 bind_ipi.vcpu = xen_vcpu_nr(cpu);
2019 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
2022 evtchn = bind_ipi.port;
2024 /* Record the new mapping. */
2025 (void)xen_irq_info_ipi_setup(cpu, irq, evtchn, ipi);
2026 /* The affinity mask is still valid */
2027 bind_evtchn_to_cpu(evtchn, cpu, false);
2031 /* Clear an irq's pending state, in preparation for polling on it */
2032 void xen_clear_irq_pending(int irq)
2034 struct irq_info *info = info_for_irq(irq);
2035 evtchn_port_t evtchn = info ? info->evtchn : 0;
2037 if (VALID_EVTCHN(evtchn))
2038 event_handler_exit(info);
2040 EXPORT_SYMBOL(xen_clear_irq_pending);
2041 void xen_set_irq_pending(int irq)
2043 evtchn_port_t evtchn = evtchn_from_irq(irq);
2045 if (VALID_EVTCHN(evtchn))
2049 bool xen_test_irq_pending(int irq)
2051 evtchn_port_t evtchn = evtchn_from_irq(irq);
2054 if (VALID_EVTCHN(evtchn))
2055 ret = test_evtchn(evtchn);
2060 /* Poll waiting for an irq to become pending with timeout. In the usual case,
2061 * the irq will be disabled so it won't deliver an interrupt. */
2062 void xen_poll_irq_timeout(int irq, u64 timeout)
2064 evtchn_port_t evtchn = evtchn_from_irq(irq);
2066 if (VALID_EVTCHN(evtchn)) {
2067 struct sched_poll poll;
2070 poll.timeout = timeout;
2071 set_xen_guest_handle(poll.ports, &evtchn);
2073 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
2077 EXPORT_SYMBOL(xen_poll_irq_timeout);
2078 /* Poll waiting for an irq to become pending. In the usual case, the
2079 * irq will be disabled so it won't deliver an interrupt. */
2080 void xen_poll_irq(int irq)
2082 xen_poll_irq_timeout(irq, 0 /* no timeout */);
2085 /* Check whether the IRQ line is shared with other guests. */
2086 int xen_test_irq_shared(int irq)
2088 struct irq_info *info = info_for_irq(irq);
2089 struct physdev_irq_status_query irq_status;
2094 irq_status.irq = info->u.pirq.pirq;
2096 if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
2098 return !(irq_status.flags & XENIRQSTAT_shared);
2100 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
2102 void xen_irq_resume(void)
2105 struct irq_info *info;
2107 /* New event-channel space is not 'live' yet. */
2108 xen_evtchn_resume();
2110 /* No IRQ <-> event-channel mappings. */
2111 list_for_each_entry(info, &xen_irq_list_head, list) {
2112 /* Zap event-channel binding */
2114 /* Adjust accounting */
2115 channels_on_cpu_dec(info);
2118 clear_evtchn_to_irq_all();
2120 for_each_possible_cpu(cpu) {
2121 restore_cpu_virqs(cpu);
2122 restore_cpu_ipis(cpu);
2128 static struct irq_chip xen_dynamic_chip __read_mostly = {
2131 .irq_disable = disable_dynirq,
2132 .irq_mask = disable_dynirq,
2133 .irq_unmask = enable_dynirq,
2135 .irq_ack = ack_dynirq,
2136 .irq_mask_ack = mask_ack_dynirq,
2138 .irq_set_affinity = set_affinity_irq,
2139 .irq_retrigger = retrigger_dynirq,
2142 static struct irq_chip xen_lateeoi_chip __read_mostly = {
2143 /* The chip name needs to contain "xen-dyn" for irqbalance to work. */
2144 .name = "xen-dyn-lateeoi",
2146 .irq_disable = disable_dynirq,
2147 .irq_mask = disable_dynirq,
2148 .irq_unmask = enable_dynirq,
2150 .irq_ack = lateeoi_ack_dynirq,
2151 .irq_mask_ack = lateeoi_mask_ack_dynirq,
2153 .irq_set_affinity = set_affinity_irq,
2154 .irq_retrigger = retrigger_dynirq,
2157 static struct irq_chip xen_pirq_chip __read_mostly = {
2160 .irq_startup = startup_pirq,
2161 .irq_shutdown = shutdown_pirq,
2162 .irq_enable = enable_pirq,
2163 .irq_disable = disable_pirq,
2165 .irq_mask = disable_dynirq,
2166 .irq_unmask = enable_dynirq,
2168 .irq_ack = eoi_pirq,
2169 .irq_eoi = eoi_pirq,
2170 .irq_mask_ack = mask_ack_pirq,
2172 .irq_set_affinity = set_affinity_irq,
2174 .irq_retrigger = retrigger_dynirq,
2177 static struct irq_chip xen_percpu_chip __read_mostly = {
2178 .name = "xen-percpu",
2180 .irq_disable = disable_dynirq,
2181 .irq_mask = disable_dynirq,
2182 .irq_unmask = enable_dynirq,
2184 .irq_ack = ack_dynirq,
2188 #ifdef CONFIG_XEN_PVHVM
2189 /* Vector callbacks are better than PCI interrupts to receive event
2190 * channel notifications because we can receive vector callbacks on any
2191 * vcpu and we don't need PCI support or APIC interactions. */
2192 void xen_setup_callback_vector(void)
2194 uint64_t callback_via;
2196 if (xen_have_vector_callback) {
2197 callback_via = HVM_CALLBACK_VECTOR(HYPERVISOR_CALLBACK_VECTOR);
2198 if (xen_set_callback_via(callback_via)) {
2199 pr_err("Request for Xen HVM callback vector failed\n");
2200 xen_have_vector_callback = false;
2206 * Setup per-vCPU vector-type callbacks. If this setup is unavailable,
2207 * fallback to the global vector-type callback.
2209 static __init void xen_init_setup_upcall_vector(void)
2211 if (!xen_have_vector_callback)
2214 if ((cpuid_eax(xen_cpuid_base() + 4) & XEN_HVM_CPUID_UPCALL_VECTOR) &&
2215 !xen_set_upcall_vector(0))
2216 xen_percpu_upcall = true;
2217 else if (xen_feature(XENFEAT_hvm_callback_vector))
2218 xen_setup_callback_vector();
2220 xen_have_vector_callback = false;
2223 int xen_set_upcall_vector(unsigned int cpu)
2226 xen_hvm_evtchn_upcall_vector_t op = {
2227 .vector = HYPERVISOR_CALLBACK_VECTOR,
2228 .vcpu = per_cpu(xen_vcpu_id, cpu),
2231 rc = HYPERVISOR_hvm_op(HVMOP_set_evtchn_upcall_vector, &op);
2235 /* Trick toolstack to think we are enlightened. */
2237 rc = xen_set_callback_via(1);
2242 static __init void xen_alloc_callback_vector(void)
2244 if (!xen_have_vector_callback)
2247 pr_info("Xen HVM callback vector for event delivery is enabled\n");
2248 alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_xen_hvm_callback);
2251 void xen_setup_callback_vector(void) {}
2252 static inline void xen_init_setup_upcall_vector(void) {}
2253 int xen_set_upcall_vector(unsigned int cpu) {}
2254 static inline void xen_alloc_callback_vector(void) {}
2255 #endif /* CONFIG_XEN_PVHVM */
2256 #endif /* CONFIG_X86 */
2258 bool xen_fifo_events = true;
2259 module_param_named(fifo_events, xen_fifo_events, bool, 0);
2261 static int xen_evtchn_cpu_prepare(unsigned int cpu)
2265 xen_cpu_init_eoi(cpu);
2267 if (evtchn_ops->percpu_init)
2268 ret = evtchn_ops->percpu_init(cpu);
2273 static int xen_evtchn_cpu_dead(unsigned int cpu)
2277 if (evtchn_ops->percpu_deinit)
2278 ret = evtchn_ops->percpu_deinit(cpu);
2283 void __init xen_init_IRQ(void)
2286 evtchn_port_t evtchn;
2288 if (xen_fifo_events)
2289 ret = xen_evtchn_fifo_init();
2291 xen_evtchn_2l_init();
2292 xen_fifo_events = false;
2295 xen_cpu_init_eoi(smp_processor_id());
2297 cpuhp_setup_state_nocalls(CPUHP_XEN_EVTCHN_PREPARE,
2298 "xen/evtchn:prepare",
2299 xen_evtchn_cpu_prepare, xen_evtchn_cpu_dead);
2301 evtchn_to_irq = kcalloc(EVTCHN_ROW(xen_evtchn_max_channels()),
2302 sizeof(*evtchn_to_irq), GFP_KERNEL);
2303 BUG_ON(!evtchn_to_irq);
2305 /* No event channels are 'live' right now. */
2306 for (evtchn = 0; evtchn < xen_evtchn_nr_channels(); evtchn++)
2307 mask_evtchn(evtchn);
2309 pirq_needs_eoi = pirq_needs_eoi_flag;
2312 if (xen_pv_domain()) {
2313 if (xen_initial_domain())
2314 pci_xen_initial_domain();
2316 xen_init_setup_upcall_vector();
2317 xen_alloc_callback_vector();
2320 if (xen_hvm_domain()) {
2322 /* pci_xen_hvm_init must be called after native_init_IRQ so that
2323 * __acpi_register_gsi can point at the right function */
2327 struct physdev_pirq_eoi_gmfn eoi_gmfn;
2329 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
2330 eoi_gmfn.gmfn = virt_to_gfn(pirq_eoi_map);
2331 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
2333 free_page((unsigned long) pirq_eoi_map);
2334 pirq_eoi_map = NULL;
2336 pirq_needs_eoi = pirq_check_eoi_map;