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/rcupdate.h>
37 #include <linux/spinlock.h>
38 #include <linux/cpuhotplug.h>
39 #include <linux/atomic.h>
40 #include <linux/ktime.h>
44 #include <asm/ptrace.h>
45 #include <asm/idtentry.h>
47 #include <asm/io_apic.h>
48 #include <asm/i8259.h>
49 #include <asm/xen/cpuid.h>
50 #include <asm/xen/pci.h>
52 #include <asm/sync_bitops.h>
53 #include <asm/xen/hypercall.h>
54 #include <asm/xen/hypervisor.h>
59 #include <xen/xen-ops.h>
60 #include <xen/events.h>
61 #include <xen/interface/xen.h>
62 #include <xen/interface/event_channel.h>
63 #include <xen/interface/hvm/hvm_op.h>
64 #include <xen/interface/hvm/params.h>
65 #include <xen/interface/physdev.h>
66 #include <xen/interface/sched.h>
67 #include <xen/interface/vcpu.h>
68 #include <xen/xenbus.h>
69 #include <asm/hw_irq.h>
71 #include "events_internal.h"
73 #undef MODULE_PARAM_PREFIX
74 #define MODULE_PARAM_PREFIX "xen."
76 /* Interrupt types. */
86 * Packed IRQ information:
87 * type - enum xen_irq_type
88 * event channel - irq->event channel mapping
89 * cpu - cpu this event channel is bound to
90 * index - type-specific information:
91 * PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
92 * guest, or GSI (real passthrough IRQ) of the device.
98 struct list_head list;
99 struct list_head eoi_list;
100 struct rcu_work rwork;
104 short type; /* type: IRQT_* */
105 u8 mask_reason; /* Why is event channel masked */
106 #define EVT_MASK_REASON_EXPLICIT 0x01
107 #define EVT_MASK_REASON_TEMPORARY 0x02
108 #define EVT_MASK_REASON_EOI_PENDING 0x04
109 u8 is_active; /* Is event just being handled? */
111 evtchn_port_t evtchn; /* event channel */
112 unsigned short cpu; /* cpu bound */
113 unsigned short eoi_cpu; /* EOI must happen on this cpu-1 */
114 unsigned int irq_epoch; /* If eoi_cpu valid: irq_epoch of event */
115 u64 eoi_time; /* Time in jiffies when to EOI. */
117 bool is_static; /* Is event channel static */
125 unsigned char vector;
129 struct xenbus_device *interdomain;
133 #define PIRQ_NEEDS_EOI (1 << 0)
134 #define PIRQ_SHAREABLE (1 << 1)
135 #define PIRQ_MSI_GROUP (1 << 2)
137 static uint __read_mostly event_loop_timeout = 2;
138 module_param(event_loop_timeout, uint, 0644);
140 static uint __read_mostly event_eoi_delay = 10;
141 module_param(event_eoi_delay, uint, 0644);
143 const struct evtchn_ops *evtchn_ops;
146 * This lock protects updates to the following mapping and reference-count
147 * arrays. The lock does not need to be acquired to read the mapping tables.
149 static DEFINE_MUTEX(irq_mapping_update_lock);
154 * irq_mapping_update_lock
156 * percpu eoi_list_lock
160 static LIST_HEAD(xen_irq_list_head);
162 /* IRQ <-> VIRQ mapping. */
163 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
165 /* IRQ <-> IPI mapping */
166 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
167 /* Cache for IPI event channels - needed for hot cpu unplug (avoid RCU usage). */
168 static DEFINE_PER_CPU(evtchn_port_t [XEN_NR_IPIS], ipi_to_evtchn) = {[0 ... XEN_NR_IPIS-1] = 0};
170 /* Event channel distribution data */
171 static atomic_t channels_on_cpu[NR_CPUS];
173 static int **evtchn_to_irq;
175 static unsigned long *pirq_eoi_map;
177 static bool (*pirq_needs_eoi)(struct irq_info *info);
179 #define EVTCHN_ROW(e) (e / (PAGE_SIZE/sizeof(**evtchn_to_irq)))
180 #define EVTCHN_COL(e) (e % (PAGE_SIZE/sizeof(**evtchn_to_irq)))
181 #define EVTCHN_PER_ROW (PAGE_SIZE / sizeof(**evtchn_to_irq))
183 /* Xen will never allocate port zero for any purpose. */
184 #define VALID_EVTCHN(chn) ((chn) != 0)
186 static struct irq_info *legacy_info_ptrs[NR_IRQS_LEGACY];
188 static struct irq_chip xen_dynamic_chip;
189 static struct irq_chip xen_lateeoi_chip;
190 static struct irq_chip xen_percpu_chip;
191 static struct irq_chip xen_pirq_chip;
192 static void enable_dynirq(struct irq_data *data);
194 static DEFINE_PER_CPU(unsigned int, irq_epoch);
196 static void clear_evtchn_to_irq_row(int *evtchn_row)
200 for (col = 0; col < EVTCHN_PER_ROW; col++)
201 WRITE_ONCE(evtchn_row[col], -1);
204 static void clear_evtchn_to_irq_all(void)
208 for (row = 0; row < EVTCHN_ROW(xen_evtchn_max_channels()); row++) {
209 if (evtchn_to_irq[row] == NULL)
211 clear_evtchn_to_irq_row(evtchn_to_irq[row]);
215 static int set_evtchn_to_irq(evtchn_port_t evtchn, unsigned int irq)
221 if (evtchn >= xen_evtchn_max_channels())
224 row = EVTCHN_ROW(evtchn);
225 col = EVTCHN_COL(evtchn);
227 if (evtchn_to_irq[row] == NULL) {
228 /* Unallocated irq entries return -1 anyway */
232 evtchn_row = (int *) __get_free_pages(GFP_KERNEL, 0);
233 if (evtchn_row == NULL)
236 clear_evtchn_to_irq_row(evtchn_row);
239 * We've prepared an empty row for the mapping. If a different
240 * thread was faster inserting it, we can drop ours.
242 if (cmpxchg(&evtchn_to_irq[row], NULL, evtchn_row) != NULL)
243 free_page((unsigned long) evtchn_row);
246 WRITE_ONCE(evtchn_to_irq[row][col], irq);
250 /* Get info for IRQ */
251 static struct irq_info *info_for_irq(unsigned irq)
253 if (irq < nr_legacy_irqs())
254 return legacy_info_ptrs[irq];
256 return irq_get_chip_data(irq);
259 static void set_info_for_irq(unsigned int irq, struct irq_info *info)
261 if (irq < nr_legacy_irqs())
262 legacy_info_ptrs[irq] = info;
264 irq_set_chip_data(irq, info);
267 static struct irq_info *evtchn_to_info(evtchn_port_t evtchn)
271 if (evtchn >= xen_evtchn_max_channels())
273 if (evtchn_to_irq[EVTCHN_ROW(evtchn)] == NULL)
275 irq = READ_ONCE(evtchn_to_irq[EVTCHN_ROW(evtchn)][EVTCHN_COL(evtchn)]);
277 return (irq < 0) ? NULL : info_for_irq(irq);
280 /* Per CPU channel accounting */
281 static void channels_on_cpu_dec(struct irq_info *info)
283 if (!info->is_accounted)
286 info->is_accounted = 0;
288 if (WARN_ON_ONCE(info->cpu >= nr_cpu_ids))
291 WARN_ON_ONCE(!atomic_add_unless(&channels_on_cpu[info->cpu], -1 , 0));
294 static void channels_on_cpu_inc(struct irq_info *info)
296 if (WARN_ON_ONCE(info->cpu >= nr_cpu_ids))
299 if (WARN_ON_ONCE(!atomic_add_unless(&channels_on_cpu[info->cpu], 1,
303 info->is_accounted = 1;
306 static void xen_irq_free_desc(unsigned int irq)
308 /* Legacy IRQ descriptors are managed by the arch. */
309 if (irq >= nr_legacy_irqs())
313 static void delayed_free_irq(struct work_struct *work)
315 struct irq_info *info = container_of(to_rcu_work(work), struct irq_info,
317 unsigned int irq = info->irq;
319 /* Remove the info pointer only now, with no potential users left. */
320 set_info_for_irq(irq, NULL);
324 xen_irq_free_desc(irq);
327 /* Constructors for packed IRQ information. */
328 static int xen_irq_info_common_setup(struct irq_info *info,
329 enum xen_irq_type type,
330 evtchn_port_t evtchn,
335 BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
338 info->evtchn = evtchn;
340 info->mask_reason = EVT_MASK_REASON_EXPLICIT;
341 raw_spin_lock_init(&info->lock);
343 ret = set_evtchn_to_irq(evtchn, info->irq);
347 irq_clear_status_flags(info->irq, IRQ_NOREQUEST | IRQ_NOAUTOEN);
349 return xen_evtchn_port_setup(evtchn);
352 static int xen_irq_info_evtchn_setup(struct irq_info *info,
353 evtchn_port_t evtchn,
354 struct xenbus_device *dev)
358 ret = xen_irq_info_common_setup(info, IRQT_EVTCHN, evtchn, 0);
359 info->u.interdomain = dev;
361 atomic_inc(&dev->event_channels);
366 static int xen_irq_info_ipi_setup(struct irq_info *info, unsigned int cpu,
367 evtchn_port_t evtchn, enum ipi_vector ipi)
371 per_cpu(ipi_to_irq, cpu)[ipi] = info->irq;
372 per_cpu(ipi_to_evtchn, cpu)[ipi] = evtchn;
374 return xen_irq_info_common_setup(info, IRQT_IPI, evtchn, 0);
377 static int xen_irq_info_virq_setup(struct irq_info *info, unsigned int cpu,
378 evtchn_port_t evtchn, unsigned int virq)
382 per_cpu(virq_to_irq, cpu)[virq] = info->irq;
384 return xen_irq_info_common_setup(info, IRQT_VIRQ, evtchn, 0);
387 static int xen_irq_info_pirq_setup(struct irq_info *info, evtchn_port_t evtchn,
388 unsigned int pirq, unsigned int gsi,
389 uint16_t domid, unsigned char flags)
391 info->u.pirq.pirq = pirq;
392 info->u.pirq.gsi = gsi;
393 info->u.pirq.domid = domid;
394 info->u.pirq.flags = flags;
396 return xen_irq_info_common_setup(info, IRQT_PIRQ, evtchn, 0);
399 static void xen_irq_info_cleanup(struct irq_info *info)
401 set_evtchn_to_irq(info->evtchn, -1);
402 xen_evtchn_port_remove(info->evtchn, info->cpu);
404 channels_on_cpu_dec(info);
408 * Accessors for packed IRQ information.
410 static evtchn_port_t evtchn_from_irq(unsigned int irq)
412 const struct irq_info *info = NULL;
414 if (likely(irq < nr_irqs))
415 info = info_for_irq(irq);
422 unsigned int irq_from_evtchn(evtchn_port_t evtchn)
424 struct irq_info *info = evtchn_to_info(evtchn);
426 return info ? info->irq : -1;
428 EXPORT_SYMBOL_GPL(irq_from_evtchn);
430 int irq_evtchn_from_virq(unsigned int cpu, unsigned int virq,
431 evtchn_port_t *evtchn)
433 int irq = per_cpu(virq_to_irq, cpu)[virq];
435 *evtchn = evtchn_from_irq(irq);
440 static enum ipi_vector ipi_from_irq(struct irq_info *info)
442 BUG_ON(info == NULL);
443 BUG_ON(info->type != IRQT_IPI);
448 static unsigned int virq_from_irq(struct irq_info *info)
450 BUG_ON(info == NULL);
451 BUG_ON(info->type != IRQT_VIRQ);
456 static unsigned int pirq_from_irq(struct irq_info *info)
458 BUG_ON(info == NULL);
459 BUG_ON(info->type != IRQT_PIRQ);
461 return info->u.pirq.pirq;
464 unsigned int cpu_from_evtchn(evtchn_port_t evtchn)
466 struct irq_info *info = evtchn_to_info(evtchn);
468 return info ? info->cpu : 0;
471 static void do_mask(struct irq_info *info, u8 reason)
475 raw_spin_lock_irqsave(&info->lock, flags);
477 if (!info->mask_reason)
478 mask_evtchn(info->evtchn);
480 info->mask_reason |= reason;
482 raw_spin_unlock_irqrestore(&info->lock, flags);
485 static void do_unmask(struct irq_info *info, u8 reason)
489 raw_spin_lock_irqsave(&info->lock, flags);
491 info->mask_reason &= ~reason;
493 if (!info->mask_reason)
494 unmask_evtchn(info->evtchn);
496 raw_spin_unlock_irqrestore(&info->lock, flags);
500 static bool pirq_check_eoi_map(struct irq_info *info)
502 return test_bit(pirq_from_irq(info), pirq_eoi_map);
506 static bool pirq_needs_eoi_flag(struct irq_info *info)
508 BUG_ON(info->type != IRQT_PIRQ);
510 return info->u.pirq.flags & PIRQ_NEEDS_EOI;
513 static void bind_evtchn_to_cpu(struct irq_info *info, unsigned int cpu,
516 if (IS_ENABLED(CONFIG_SMP) && force_affinity) {
517 struct irq_data *data = irq_get_irq_data(info->irq);
519 irq_data_update_affinity(data, cpumask_of(cpu));
520 irq_data_update_effective_affinity(data, cpumask_of(cpu));
523 xen_evtchn_port_bind_to_cpu(info->evtchn, cpu, info->cpu);
525 channels_on_cpu_dec(info);
527 channels_on_cpu_inc(info);
531 * notify_remote_via_irq - send event to remote end of event channel via irq
532 * @irq: irq of event channel to send event to
534 * Unlike notify_remote_via_evtchn(), this is safe to use across
535 * save/restore. Notifications on a broken connection are silently
538 void notify_remote_via_irq(int irq)
540 evtchn_port_t evtchn = evtchn_from_irq(irq);
542 if (VALID_EVTCHN(evtchn))
543 notify_remote_via_evtchn(evtchn);
545 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
547 struct lateeoi_work {
548 struct delayed_work delayed;
549 spinlock_t eoi_list_lock;
550 struct list_head eoi_list;
553 static DEFINE_PER_CPU(struct lateeoi_work, lateeoi);
555 static void lateeoi_list_del(struct irq_info *info)
557 struct lateeoi_work *eoi = &per_cpu(lateeoi, info->eoi_cpu);
560 spin_lock_irqsave(&eoi->eoi_list_lock, flags);
561 list_del_init(&info->eoi_list);
562 spin_unlock_irqrestore(&eoi->eoi_list_lock, flags);
565 static void lateeoi_list_add(struct irq_info *info)
567 struct lateeoi_work *eoi = &per_cpu(lateeoi, info->eoi_cpu);
568 struct irq_info *elem;
569 u64 now = get_jiffies_64();
573 if (now < info->eoi_time)
574 delay = info->eoi_time - now;
578 spin_lock_irqsave(&eoi->eoi_list_lock, flags);
580 elem = list_first_entry_or_null(&eoi->eoi_list, struct irq_info,
582 if (!elem || info->eoi_time < elem->eoi_time) {
583 list_add(&info->eoi_list, &eoi->eoi_list);
584 mod_delayed_work_on(info->eoi_cpu, system_wq,
585 &eoi->delayed, delay);
587 list_for_each_entry_reverse(elem, &eoi->eoi_list, eoi_list) {
588 if (elem->eoi_time <= info->eoi_time)
591 list_add(&info->eoi_list, &elem->eoi_list);
594 spin_unlock_irqrestore(&eoi->eoi_list_lock, flags);
597 static void xen_irq_lateeoi_locked(struct irq_info *info, bool spurious)
599 evtchn_port_t evtchn;
601 unsigned int delay = 0;
603 evtchn = info->evtchn;
604 if (!VALID_EVTCHN(evtchn) || !list_empty(&info->eoi_list))
608 struct xenbus_device *dev = info->u.interdomain;
609 unsigned int threshold = 1;
611 if (dev && dev->spurious_threshold)
612 threshold = dev->spurious_threshold;
614 if ((1 << info->spurious_cnt) < (HZ << 2)) {
615 if (info->spurious_cnt != 0xFF)
616 info->spurious_cnt++;
618 if (info->spurious_cnt > threshold) {
619 delay = 1 << (info->spurious_cnt - 1 - threshold);
623 info->eoi_cpu = smp_processor_id();
624 info->eoi_time = get_jiffies_64() + delay;
626 atomic_add(delay, &dev->jiffies_eoi_delayed);
629 atomic_inc(&dev->spurious_events);
631 info->spurious_cnt = 0;
635 if (info->eoi_time &&
636 (info->irq_epoch == per_cpu(irq_epoch, cpu) || delay)) {
637 lateeoi_list_add(info);
643 /* is_active hasn't been reset yet, do it now. */
644 smp_store_release(&info->is_active, 0);
645 do_unmask(info, EVT_MASK_REASON_EOI_PENDING);
648 static void xen_irq_lateeoi_worker(struct work_struct *work)
650 struct lateeoi_work *eoi;
651 struct irq_info *info;
652 u64 now = get_jiffies_64();
655 eoi = container_of(to_delayed_work(work), struct lateeoi_work, delayed);
660 spin_lock_irqsave(&eoi->eoi_list_lock, flags);
662 info = list_first_entry_or_null(&eoi->eoi_list, struct irq_info,
668 if (now < info->eoi_time) {
669 mod_delayed_work_on(info->eoi_cpu, system_wq,
671 info->eoi_time - now);
675 list_del_init(&info->eoi_list);
677 spin_unlock_irqrestore(&eoi->eoi_list_lock, flags);
681 xen_irq_lateeoi_locked(info, false);
684 spin_unlock_irqrestore(&eoi->eoi_list_lock, flags);
689 static void xen_cpu_init_eoi(unsigned int cpu)
691 struct lateeoi_work *eoi = &per_cpu(lateeoi, cpu);
693 INIT_DELAYED_WORK(&eoi->delayed, xen_irq_lateeoi_worker);
694 spin_lock_init(&eoi->eoi_list_lock);
695 INIT_LIST_HEAD(&eoi->eoi_list);
698 void xen_irq_lateeoi(unsigned int irq, unsigned int eoi_flags)
700 struct irq_info *info;
704 info = info_for_irq(irq);
707 xen_irq_lateeoi_locked(info, eoi_flags & XEN_EOI_FLAG_SPURIOUS);
711 EXPORT_SYMBOL_GPL(xen_irq_lateeoi);
713 static struct irq_info *xen_irq_init(unsigned int irq)
715 struct irq_info *info;
717 info = kzalloc(sizeof(*info), GFP_KERNEL);
720 info->type = IRQT_UNBOUND;
722 INIT_RCU_WORK(&info->rwork, delayed_free_irq);
724 set_info_for_irq(irq, info);
726 * Interrupt affinity setting can be immediate. No point
727 * in delaying it until an interrupt is handled.
729 irq_set_status_flags(irq, IRQ_MOVE_PCNTXT);
731 INIT_LIST_HEAD(&info->eoi_list);
732 list_add_tail(&info->list, &xen_irq_list_head);
738 static struct irq_info *xen_allocate_irq_dynamic(void)
740 int irq = irq_alloc_desc_from(0, -1);
741 struct irq_info *info = NULL;
744 info = xen_irq_init(irq);
746 xen_irq_free_desc(irq);
752 static struct irq_info *xen_allocate_irq_gsi(unsigned int gsi)
755 struct irq_info *info;
758 * A PV guest has no concept of a GSI (since it has no ACPI
759 * nor access to/knowledge of the physical APICs). Therefore
760 * all IRQs are dynamically allocated from the entire IRQ
763 if (xen_pv_domain() && !xen_initial_domain())
764 return xen_allocate_irq_dynamic();
766 /* Legacy IRQ descriptors are already allocated by the arch. */
767 if (gsi < nr_legacy_irqs())
770 irq = irq_alloc_desc_at(gsi, -1);
772 info = xen_irq_init(irq);
774 xen_irq_free_desc(irq);
779 static void xen_free_irq(struct irq_info *info)
784 if (!list_empty(&info->eoi_list))
785 lateeoi_list_del(info);
787 list_del(&info->list);
789 WARN_ON(info->refcnt > 0);
791 queue_rcu_work(system_wq, &info->rwork);
794 /* Not called for lateeoi events. */
795 static void event_handler_exit(struct irq_info *info)
797 smp_store_release(&info->is_active, 0);
798 clear_evtchn(info->evtchn);
801 static void pirq_query_unmask(struct irq_info *info)
803 struct physdev_irq_status_query irq_status;
805 irq_status.irq = pirq_from_irq(info);
806 if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
807 irq_status.flags = 0;
809 info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
810 if (irq_status.flags & XENIRQSTAT_needs_eoi)
811 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
814 static void do_eoi_pirq(struct irq_info *info)
816 struct physdev_eoi eoi = { .irq = pirq_from_irq(info) };
819 if (!VALID_EVTCHN(info->evtchn))
822 event_handler_exit(info);
824 if (pirq_needs_eoi(info)) {
825 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
830 static void eoi_pirq(struct irq_data *data)
832 struct irq_info *info = info_for_irq(data->irq);
837 static void do_disable_dynirq(struct irq_info *info)
839 if (VALID_EVTCHN(info->evtchn))
840 do_mask(info, EVT_MASK_REASON_EXPLICIT);
843 static void disable_dynirq(struct irq_data *data)
845 struct irq_info *info = info_for_irq(data->irq);
848 do_disable_dynirq(info);
851 static void mask_ack_pirq(struct irq_data *data)
853 struct irq_info *info = info_for_irq(data->irq);
856 do_disable_dynirq(info);
861 static unsigned int __startup_pirq(struct irq_info *info)
863 struct evtchn_bind_pirq bind_pirq;
864 evtchn_port_t evtchn = info->evtchn;
867 if (VALID_EVTCHN(evtchn))
870 bind_pirq.pirq = pirq_from_irq(info);
871 /* NB. We are happy to share unless we are probing. */
872 bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
873 BIND_PIRQ__WILL_SHARE : 0;
874 rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
876 pr_warn("Failed to obtain physical IRQ %d\n", info->irq);
879 evtchn = bind_pirq.port;
881 pirq_query_unmask(info);
883 rc = set_evtchn_to_irq(evtchn, info->irq);
887 info->evtchn = evtchn;
888 bind_evtchn_to_cpu(info, 0, false);
890 rc = xen_evtchn_port_setup(evtchn);
895 do_unmask(info, EVT_MASK_REASON_EXPLICIT);
902 pr_err("irq%d: Failed to set port to irq mapping (%d)\n", info->irq,
904 xen_evtchn_close(evtchn);
908 static unsigned int startup_pirq(struct irq_data *data)
910 struct irq_info *info = info_for_irq(data->irq);
912 return __startup_pirq(info);
915 static void shutdown_pirq(struct irq_data *data)
917 struct irq_info *info = info_for_irq(data->irq);
918 evtchn_port_t evtchn = info->evtchn;
920 BUG_ON(info->type != IRQT_PIRQ);
922 if (!VALID_EVTCHN(evtchn))
925 do_mask(info, EVT_MASK_REASON_EXPLICIT);
926 xen_irq_info_cleanup(info);
927 xen_evtchn_close(evtchn);
930 static void enable_pirq(struct irq_data *data)
935 static void disable_pirq(struct irq_data *data)
937 disable_dynirq(data);
940 int xen_irq_from_gsi(unsigned gsi)
942 struct irq_info *info;
944 list_for_each_entry(info, &xen_irq_list_head, list) {
945 if (info->type != IRQT_PIRQ)
948 if (info->u.pirq.gsi == gsi)
954 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
956 static void __unbind_from_irq(struct irq_info *info, unsigned int irq)
958 evtchn_port_t evtchn;
959 bool close_evtchn = false;
962 xen_irq_free_desc(irq);
966 if (info->refcnt > 0) {
968 if (info->refcnt != 0)
972 evtchn = info->evtchn;
974 if (VALID_EVTCHN(evtchn)) {
975 unsigned int cpu = info->cpu;
976 struct xenbus_device *dev;
978 if (!info->is_static)
981 switch (info->type) {
983 per_cpu(virq_to_irq, cpu)[virq_from_irq(info)] = -1;
986 per_cpu(ipi_to_irq, cpu)[ipi_from_irq(info)] = -1;
987 per_cpu(ipi_to_evtchn, cpu)[ipi_from_irq(info)] = 0;
990 dev = info->u.interdomain;
992 atomic_dec(&dev->event_channels);
998 xen_irq_info_cleanup(info);
1001 xen_evtchn_close(evtchn);
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)
1020 struct irq_info *info;
1021 struct physdev_irq irq_op;
1024 mutex_lock(&irq_mapping_update_lock);
1026 ret = xen_irq_from_gsi(gsi);
1028 pr_info("%s: returning irq %d for gsi %u\n",
1029 __func__, ret, gsi);
1033 info = xen_allocate_irq_gsi(gsi);
1037 irq_op.irq = info->irq;
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(info, 0, pirq, gsi, DOMID_SELF,
1051 shareable ? PIRQ_SHAREABLE : 0);
1053 __unbind_from_irq(info, info->irq);
1057 pirq_query_unmask(info);
1058 /* We try to use the handler with the appropriate semantic for the
1059 * type of interrupt: if the interrupt is an edge triggered
1060 * interrupt we use handle_edge_irq.
1062 * On the other hand if the interrupt is level triggered we use
1063 * handle_fasteoi_irq like the native code does for this kind of
1066 * Depending on the Xen version, pirq_needs_eoi might return true
1067 * not only for level triggered interrupts but for edge triggered
1068 * interrupts too. In any case Xen always honors the eoi mechanism,
1069 * not injecting any more pirqs of the same kind if the first one
1070 * hasn't received an eoi yet. Therefore using the fasteoi handler
1071 * is the right choice either way.
1074 irq_set_chip_and_handler_name(info->irq, &xen_pirq_chip,
1075 handle_fasteoi_irq, name);
1077 irq_set_chip_and_handler_name(info->irq, &xen_pirq_chip,
1078 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)
1107 struct irq_info *info;
1109 mutex_lock(&irq_mapping_update_lock);
1111 irq = irq_alloc_descs(-1, 0, nvec, -1);
1115 for (i = 0; i < nvec; i++) {
1116 info = xen_irq_init(irq + i);
1122 irq_set_chip_and_handler_name(irq + i, &xen_pirq_chip, handle_edge_irq, name);
1124 ret = xen_irq_info_pirq_setup(info, 0, pirq + i, 0, domid,
1125 i == 0 ? 0 : PIRQ_MSI_GROUP);
1130 ret = irq_set_msi_desc(irq, msidesc);
1134 mutex_unlock(&irq_mapping_update_lock);
1139 info = info_for_irq(irq + nvec);
1140 __unbind_from_irq(info, irq + nvec);
1142 mutex_unlock(&irq_mapping_update_lock);
1147 int xen_destroy_irq(int irq)
1149 struct physdev_unmap_pirq unmap_irq;
1150 struct irq_info *info = info_for_irq(irq);
1153 mutex_lock(&irq_mapping_update_lock);
1156 * If trying to remove a vector in a MSI group different
1157 * than the first one skip the PIRQ unmap unless this vector
1158 * is the first one in the group.
1160 if (xen_initial_domain() && !(info->u.pirq.flags & PIRQ_MSI_GROUP)) {
1161 unmap_irq.pirq = info->u.pirq.pirq;
1162 unmap_irq.domid = info->u.pirq.domid;
1163 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
1164 /* If another domain quits without making the pci_disable_msix
1165 * call, the Xen hypervisor takes care of freeing the PIRQs
1166 * (free_domain_pirqs).
1168 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
1169 pr_info("domain %d does not have %d anymore\n",
1170 info->u.pirq.domid, info->u.pirq.pirq);
1172 pr_warn("unmap irq failed %d\n", rc);
1180 mutex_unlock(&irq_mapping_update_lock);
1184 int xen_pirq_from_irq(unsigned irq)
1186 struct irq_info *info = info_for_irq(irq);
1188 return pirq_from_irq(info);
1190 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
1192 static int bind_evtchn_to_irq_chip(evtchn_port_t evtchn, struct irq_chip *chip,
1193 struct xenbus_device *dev, bool shared)
1196 struct irq_info *info;
1198 if (evtchn >= xen_evtchn_max_channels())
1201 mutex_lock(&irq_mapping_update_lock);
1203 info = evtchn_to_info(evtchn);
1206 info = xen_allocate_irq_dynamic();
1210 irq_set_chip_and_handler_name(info->irq, chip,
1211 handle_edge_irq, "event");
1213 ret = xen_irq_info_evtchn_setup(info, evtchn, dev);
1215 __unbind_from_irq(info, info->irq);
1219 * New interdomain events are initially bound to vCPU0 This
1220 * is required to setup the event channel in the first
1221 * place and also important for UP guests because the
1222 * affinity setting is not invoked on them so nothing would
1225 bind_evtchn_to_cpu(info, 0, false);
1226 } else if (!WARN_ON(info->type != IRQT_EVTCHN)) {
1227 if (shared && !WARN_ON(info->refcnt < 0))
1234 mutex_unlock(&irq_mapping_update_lock);
1239 int bind_evtchn_to_irq(evtchn_port_t evtchn)
1241 return bind_evtchn_to_irq_chip(evtchn, &xen_dynamic_chip, NULL, false);
1243 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
1245 int bind_evtchn_to_irq_lateeoi(evtchn_port_t evtchn)
1247 return bind_evtchn_to_irq_chip(evtchn, &xen_lateeoi_chip, NULL, false);
1249 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq_lateeoi);
1251 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
1253 struct evtchn_bind_ipi bind_ipi;
1254 evtchn_port_t evtchn;
1255 struct irq_info *info;
1258 mutex_lock(&irq_mapping_update_lock);
1260 ret = per_cpu(ipi_to_irq, cpu)[ipi];
1263 info = xen_allocate_irq_dynamic();
1267 irq_set_chip_and_handler_name(info->irq, &xen_percpu_chip,
1268 handle_percpu_irq, "ipi");
1270 bind_ipi.vcpu = xen_vcpu_nr(cpu);
1271 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1274 evtchn = bind_ipi.port;
1276 ret = xen_irq_info_ipi_setup(info, cpu, evtchn, ipi);
1278 __unbind_from_irq(info, info->irq);
1282 * Force the affinity mask to the target CPU so proc shows
1283 * the correct target.
1285 bind_evtchn_to_cpu(info, cpu, true);
1288 info = info_for_irq(ret);
1289 WARN_ON(info == NULL || info->type != IRQT_IPI);
1293 mutex_unlock(&irq_mapping_update_lock);
1297 static int bind_interdomain_evtchn_to_irq_chip(struct xenbus_device *dev,
1298 evtchn_port_t remote_port,
1299 struct irq_chip *chip,
1302 struct evtchn_bind_interdomain bind_interdomain;
1305 bind_interdomain.remote_dom = dev->otherend_id;
1306 bind_interdomain.remote_port = remote_port;
1308 err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
1311 return err ? : bind_evtchn_to_irq_chip(bind_interdomain.local_port,
1315 int bind_interdomain_evtchn_to_irq_lateeoi(struct xenbus_device *dev,
1316 evtchn_port_t remote_port)
1318 return bind_interdomain_evtchn_to_irq_chip(dev, remote_port,
1319 &xen_lateeoi_chip, false);
1321 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irq_lateeoi);
1323 static int find_virq(unsigned int virq, unsigned int cpu, evtchn_port_t *evtchn)
1325 struct evtchn_status status;
1329 memset(&status, 0, sizeof(status));
1330 for (port = 0; port < xen_evtchn_max_channels(); port++) {
1331 status.dom = DOMID_SELF;
1333 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
1336 if (status.status != EVTCHNSTAT_virq)
1338 if (status.u.virq == virq && status.vcpu == xen_vcpu_nr(cpu)) {
1347 * xen_evtchn_nr_channels - number of usable event channel ports
1349 * This may be less than the maximum supported by the current
1350 * hypervisor ABI. Use xen_evtchn_max_channels() for the maximum
1353 unsigned xen_evtchn_nr_channels(void)
1355 return evtchn_ops->nr_channels();
1357 EXPORT_SYMBOL_GPL(xen_evtchn_nr_channels);
1359 int bind_virq_to_irq(unsigned int virq, unsigned int cpu, bool percpu)
1361 struct evtchn_bind_virq bind_virq;
1362 evtchn_port_t evtchn = 0;
1363 struct irq_info *info;
1366 mutex_lock(&irq_mapping_update_lock);
1368 ret = per_cpu(virq_to_irq, cpu)[virq];
1371 info = xen_allocate_irq_dynamic();
1376 irq_set_chip_and_handler_name(info->irq, &xen_percpu_chip,
1377 handle_percpu_irq, "virq");
1379 irq_set_chip_and_handler_name(info->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(info, cpu, evtchn, virq);
1396 __unbind_from_irq(info, info->irq);
1401 * Force the affinity mask for percpu interrupts so proc
1402 * shows the correct target.
1404 bind_evtchn_to_cpu(info, cpu, percpu);
1407 info = info_for_irq(ret);
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 struct irq_info *info;
1421 mutex_lock(&irq_mapping_update_lock);
1422 info = info_for_irq(irq);
1423 __unbind_from_irq(info, irq);
1424 mutex_unlock(&irq_mapping_update_lock);
1427 static int bind_evtchn_to_irqhandler_chip(evtchn_port_t evtchn,
1428 irq_handler_t handler,
1429 unsigned long irqflags,
1430 const char *devname, void *dev_id,
1431 struct irq_chip *chip)
1435 irq = bind_evtchn_to_irq_chip(evtchn, chip, NULL,
1436 irqflags & IRQF_SHARED);
1439 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1441 unbind_from_irq(irq);
1448 int bind_evtchn_to_irqhandler(evtchn_port_t evtchn,
1449 irq_handler_t handler,
1450 unsigned long irqflags,
1451 const char *devname, void *dev_id)
1453 return bind_evtchn_to_irqhandler_chip(evtchn, handler, irqflags,
1457 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1459 int bind_evtchn_to_irqhandler_lateeoi(evtchn_port_t evtchn,
1460 irq_handler_t handler,
1461 unsigned long irqflags,
1462 const char *devname, void *dev_id)
1464 return bind_evtchn_to_irqhandler_chip(evtchn, handler, irqflags,
1468 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler_lateeoi);
1470 static int bind_interdomain_evtchn_to_irqhandler_chip(
1471 struct xenbus_device *dev, evtchn_port_t remote_port,
1472 irq_handler_t handler, unsigned long irqflags,
1473 const char *devname, void *dev_id, struct irq_chip *chip)
1477 irq = bind_interdomain_evtchn_to_irq_chip(dev, remote_port, chip,
1478 irqflags & IRQF_SHARED);
1482 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1484 unbind_from_irq(irq);
1491 int bind_interdomain_evtchn_to_irqhandler_lateeoi(struct xenbus_device *dev,
1492 evtchn_port_t remote_port,
1493 irq_handler_t handler,
1494 unsigned long irqflags,
1495 const char *devname,
1498 return bind_interdomain_evtchn_to_irqhandler_chip(dev,
1499 remote_port, handler, irqflags, devname,
1500 dev_id, &xen_lateeoi_chip);
1502 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler_lateeoi);
1504 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1505 irq_handler_t handler,
1506 unsigned long irqflags, const char *devname, void *dev_id)
1510 irq = bind_virq_to_irq(virq, cpu, irqflags & IRQF_PERCPU);
1513 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1515 unbind_from_irq(irq);
1521 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1523 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1525 irq_handler_t handler,
1526 unsigned long irqflags,
1527 const char *devname,
1532 irq = bind_ipi_to_irq(ipi, cpu);
1536 irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1537 retval = request_irq(irq, handler, irqflags, devname, dev_id);
1539 unbind_from_irq(irq);
1546 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1548 struct irq_info *info = info_for_irq(irq);
1552 free_irq(irq, dev_id);
1553 unbind_from_irq(irq);
1555 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1558 * xen_set_irq_priority() - set an event channel priority.
1559 * @irq:irq bound to an event channel.
1560 * @priority: priority between XEN_IRQ_PRIORITY_MAX and XEN_IRQ_PRIORITY_MIN.
1562 int xen_set_irq_priority(unsigned irq, unsigned priority)
1564 struct evtchn_set_priority set_priority;
1566 set_priority.port = evtchn_from_irq(irq);
1567 set_priority.priority = priority;
1569 return HYPERVISOR_event_channel_op(EVTCHNOP_set_priority,
1572 EXPORT_SYMBOL_GPL(xen_set_irq_priority);
1574 int evtchn_make_refcounted(evtchn_port_t evtchn, bool is_static)
1576 struct irq_info *info = evtchn_to_info(evtchn);
1581 WARN_ON(info->refcnt != -1);
1584 info->is_static = is_static;
1588 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1590 int evtchn_get(evtchn_port_t evtchn)
1592 struct irq_info *info;
1595 if (evtchn >= xen_evtchn_max_channels())
1598 mutex_lock(&irq_mapping_update_lock);
1600 info = evtchn_to_info(evtchn);
1606 if (info->refcnt <= 0 || info->refcnt == SHRT_MAX)
1612 mutex_unlock(&irq_mapping_update_lock);
1616 EXPORT_SYMBOL_GPL(evtchn_get);
1618 void evtchn_put(evtchn_port_t evtchn)
1620 struct irq_info *info = evtchn_to_info(evtchn);
1624 unbind_from_irq(info->irq);
1626 EXPORT_SYMBOL_GPL(evtchn_put);
1628 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1630 evtchn_port_t evtchn;
1633 if (unlikely(vector == XEN_NMI_VECTOR)) {
1634 int rc = HYPERVISOR_vcpu_op(VCPUOP_send_nmi, xen_vcpu_nr(cpu),
1637 printk(KERN_WARNING "Sending nmi to CPU%d failed (rc:%d)\n", cpu, rc);
1641 evtchn = per_cpu(ipi_to_evtchn, cpu)[vector];
1642 BUG_ON(evtchn == 0);
1643 notify_remote_via_evtchn(evtchn);
1646 struct evtchn_loop_ctrl {
1652 void handle_irq_for_port(evtchn_port_t port, struct evtchn_loop_ctrl *ctrl)
1654 struct irq_info *info = evtchn_to_info(port);
1655 struct xenbus_device *dev;
1661 * Check for timeout every 256 events.
1662 * We are setting the timeout value only after the first 256
1663 * events in order to not hurt the common case of few loop
1664 * iterations. The 256 is basically an arbitrary value.
1666 * In case we are hitting the timeout we need to defer all further
1667 * EOIs in order to ensure to leave the event handling loop rather
1668 * sooner than later.
1670 if (!ctrl->defer_eoi && !(++ctrl->count & 0xff)) {
1671 ktime_t kt = ktime_get();
1673 if (!ctrl->timeout) {
1674 kt = ktime_add_ms(kt,
1675 jiffies_to_msecs(event_loop_timeout));
1677 } else if (kt > ctrl->timeout) {
1678 ctrl->defer_eoi = true;
1682 if (xchg_acquire(&info->is_active, 1))
1685 dev = (info->type == IRQT_EVTCHN) ? info->u.interdomain : NULL;
1687 atomic_inc(&dev->events);
1689 if (ctrl->defer_eoi) {
1690 info->eoi_cpu = smp_processor_id();
1691 info->irq_epoch = __this_cpu_read(irq_epoch);
1692 info->eoi_time = get_jiffies_64() + event_eoi_delay;
1695 generic_handle_irq(info->irq);
1698 int xen_evtchn_do_upcall(void)
1700 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1701 int ret = vcpu_info->evtchn_upcall_pending ? IRQ_HANDLED : IRQ_NONE;
1702 int cpu = smp_processor_id();
1703 struct evtchn_loop_ctrl ctrl = { 0 };
1706 * When closing an event channel the associated IRQ must not be freed
1707 * until all cpus have left the event handling loop. This is ensured
1708 * by taking the rcu_read_lock() while handling events, as freeing of
1709 * the IRQ is handled via queue_rcu_work() _after_ closing the event
1715 vcpu_info->evtchn_upcall_pending = 0;
1717 xen_evtchn_handle_events(cpu, &ctrl);
1719 BUG_ON(!irqs_disabled());
1721 virt_rmb(); /* Hypervisor can set upcall pending. */
1723 } while (vcpu_info->evtchn_upcall_pending);
1728 * Increment irq_epoch only now to defer EOIs only for
1729 * xen_irq_lateeoi() invocations occurring from inside the loop
1732 __this_cpu_inc(irq_epoch);
1736 EXPORT_SYMBOL_GPL(xen_evtchn_do_upcall);
1738 /* Rebind a new event channel to an existing irq. */
1739 void rebind_evtchn_irq(evtchn_port_t evtchn, int irq)
1741 struct irq_info *info = info_for_irq(irq);
1746 /* Make sure the irq is masked, since the new event channel
1747 will also be masked. */
1750 mutex_lock(&irq_mapping_update_lock);
1752 /* After resume the irq<->evtchn mappings are all cleared out */
1753 BUG_ON(evtchn_to_info(evtchn));
1754 /* Expect irq to have been bound before,
1755 so there should be a proper type */
1756 BUG_ON(info->type == IRQT_UNBOUND);
1759 (void)xen_irq_info_evtchn_setup(info, evtchn, NULL);
1761 mutex_unlock(&irq_mapping_update_lock);
1763 bind_evtchn_to_cpu(info, info->cpu, false);
1765 /* Unmask the event channel. */
1769 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1770 static int xen_rebind_evtchn_to_cpu(struct irq_info *info, unsigned int tcpu)
1772 struct evtchn_bind_vcpu bind_vcpu;
1773 evtchn_port_t evtchn = info ? info->evtchn : 0;
1775 if (!VALID_EVTCHN(evtchn))
1778 if (!xen_support_evtchn_rebind())
1781 /* Send future instances of this interrupt to other vcpu. */
1782 bind_vcpu.port = evtchn;
1783 bind_vcpu.vcpu = xen_vcpu_nr(tcpu);
1786 * Mask the event while changing the VCPU binding to prevent
1787 * it being delivered on an unexpected VCPU.
1789 do_mask(info, EVT_MASK_REASON_TEMPORARY);
1792 * If this fails, it usually just indicates that we're dealing with a
1793 * virq or IPI channel, which don't actually need to be rebound. Ignore
1794 * it, but don't do the xenlinux-level rebind in that case.
1796 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1797 bind_evtchn_to_cpu(info, tcpu, false);
1799 do_unmask(info, EVT_MASK_REASON_TEMPORARY);
1805 * Find the CPU within @dest mask which has the least number of channels
1806 * assigned. This is not precise as the per cpu counts can be modified
1809 static unsigned int select_target_cpu(const struct cpumask *dest)
1811 unsigned int cpu, best_cpu = UINT_MAX, minch = UINT_MAX;
1813 for_each_cpu_and(cpu, dest, cpu_online_mask) {
1814 unsigned int curch = atomic_read(&channels_on_cpu[cpu]);
1816 if (curch < minch) {
1823 * Catch the unlikely case that dest contains no online CPUs. Can't
1826 if (best_cpu == UINT_MAX)
1827 return select_target_cpu(cpu_online_mask);
1832 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1835 unsigned int tcpu = select_target_cpu(dest);
1838 ret = xen_rebind_evtchn_to_cpu(info_for_irq(data->irq), tcpu);
1840 irq_data_update_effective_affinity(data, cpumask_of(tcpu));
1845 static void enable_dynirq(struct irq_data *data)
1847 struct irq_info *info = info_for_irq(data->irq);
1848 evtchn_port_t evtchn = info ? info->evtchn : 0;
1850 if (VALID_EVTCHN(evtchn))
1851 do_unmask(info, EVT_MASK_REASON_EXPLICIT);
1854 static void do_ack_dynirq(struct irq_info *info)
1856 evtchn_port_t evtchn = info->evtchn;
1858 if (VALID_EVTCHN(evtchn))
1859 event_handler_exit(info);
1862 static void ack_dynirq(struct irq_data *data)
1864 struct irq_info *info = info_for_irq(data->irq);
1867 do_ack_dynirq(info);
1870 static void mask_ack_dynirq(struct irq_data *data)
1872 struct irq_info *info = info_for_irq(data->irq);
1875 do_disable_dynirq(info);
1876 do_ack_dynirq(info);
1880 static void lateeoi_ack_dynirq(struct irq_data *data)
1882 struct irq_info *info = info_for_irq(data->irq);
1883 evtchn_port_t evtchn = info ? info->evtchn : 0;
1885 if (VALID_EVTCHN(evtchn)) {
1886 do_mask(info, EVT_MASK_REASON_EOI_PENDING);
1888 * Don't call event_handler_exit().
1889 * Need to keep is_active non-zero in order to ignore re-raised
1890 * events after cpu affinity changes while a lateeoi is pending.
1892 clear_evtchn(evtchn);
1896 static void lateeoi_mask_ack_dynirq(struct irq_data *data)
1898 struct irq_info *info = info_for_irq(data->irq);
1899 evtchn_port_t evtchn = info ? info->evtchn : 0;
1901 if (VALID_EVTCHN(evtchn)) {
1902 do_mask(info, EVT_MASK_REASON_EXPLICIT);
1903 event_handler_exit(info);
1907 static int retrigger_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))
1915 do_mask(info, EVT_MASK_REASON_TEMPORARY);
1917 do_unmask(info, EVT_MASK_REASON_TEMPORARY);
1922 static void restore_pirqs(void)
1924 int pirq, rc, irq, gsi;
1925 struct physdev_map_pirq map_irq;
1926 struct irq_info *info;
1928 list_for_each_entry(info, &xen_irq_list_head, list) {
1929 if (info->type != IRQT_PIRQ)
1932 pirq = info->u.pirq.pirq;
1933 gsi = info->u.pirq.gsi;
1936 /* save/restore of PT devices doesn't work, so at this point the
1937 * only devices present are GSI based emulated devices */
1941 map_irq.domid = DOMID_SELF;
1942 map_irq.type = MAP_PIRQ_TYPE_GSI;
1943 map_irq.index = gsi;
1944 map_irq.pirq = pirq;
1946 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1948 pr_warn("xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1949 gsi, irq, pirq, rc);
1954 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1956 __startup_pirq(info);
1960 static void restore_cpu_virqs(unsigned int cpu)
1962 struct evtchn_bind_virq bind_virq;
1963 evtchn_port_t evtchn;
1964 struct irq_info *info;
1967 for (virq = 0; virq < NR_VIRQS; virq++) {
1968 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1970 info = info_for_irq(irq);
1972 BUG_ON(virq_from_irq(info) != virq);
1974 /* Get a new binding from Xen. */
1975 bind_virq.virq = virq;
1976 bind_virq.vcpu = xen_vcpu_nr(cpu);
1977 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1980 evtchn = bind_virq.port;
1982 /* Record the new mapping. */
1983 xen_irq_info_virq_setup(info, cpu, evtchn, virq);
1984 /* The affinity mask is still valid */
1985 bind_evtchn_to_cpu(info, cpu, false);
1989 static void restore_cpu_ipis(unsigned int cpu)
1991 struct evtchn_bind_ipi bind_ipi;
1992 evtchn_port_t evtchn;
1993 struct irq_info *info;
1996 for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1997 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1999 info = info_for_irq(irq);
2001 BUG_ON(ipi_from_irq(info) != ipi);
2003 /* Get a new binding from Xen. */
2004 bind_ipi.vcpu = xen_vcpu_nr(cpu);
2005 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
2008 evtchn = bind_ipi.port;
2010 /* Record the new mapping. */
2011 xen_irq_info_ipi_setup(info, cpu, evtchn, ipi);
2012 /* The affinity mask is still valid */
2013 bind_evtchn_to_cpu(info, cpu, false);
2017 /* Clear an irq's pending state, in preparation for polling on it */
2018 void xen_clear_irq_pending(int irq)
2020 struct irq_info *info = info_for_irq(irq);
2021 evtchn_port_t evtchn = info ? info->evtchn : 0;
2023 if (VALID_EVTCHN(evtchn))
2024 event_handler_exit(info);
2026 EXPORT_SYMBOL(xen_clear_irq_pending);
2028 bool xen_test_irq_pending(int irq)
2030 evtchn_port_t evtchn = evtchn_from_irq(irq);
2033 if (VALID_EVTCHN(evtchn))
2034 ret = test_evtchn(evtchn);
2039 /* Poll waiting for an irq to become pending with timeout. In the usual case,
2040 * the irq will be disabled so it won't deliver an interrupt. */
2041 void xen_poll_irq_timeout(int irq, u64 timeout)
2043 evtchn_port_t evtchn = evtchn_from_irq(irq);
2045 if (VALID_EVTCHN(evtchn)) {
2046 struct sched_poll poll;
2049 poll.timeout = timeout;
2050 set_xen_guest_handle(poll.ports, &evtchn);
2052 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
2056 EXPORT_SYMBOL(xen_poll_irq_timeout);
2057 /* Poll waiting for an irq to become pending. In the usual case, the
2058 * irq will be disabled so it won't deliver an interrupt. */
2059 void xen_poll_irq(int irq)
2061 xen_poll_irq_timeout(irq, 0 /* no timeout */);
2064 /* Check whether the IRQ line is shared with other guests. */
2065 int xen_test_irq_shared(int irq)
2067 struct irq_info *info = info_for_irq(irq);
2068 struct physdev_irq_status_query irq_status;
2073 irq_status.irq = info->u.pirq.pirq;
2075 if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
2077 return !(irq_status.flags & XENIRQSTAT_shared);
2079 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
2081 void xen_irq_resume(void)
2084 struct irq_info *info;
2086 /* New event-channel space is not 'live' yet. */
2087 xen_evtchn_resume();
2089 /* No IRQ <-> event-channel mappings. */
2090 list_for_each_entry(info, &xen_irq_list_head, list) {
2091 /* Zap event-channel binding */
2093 /* Adjust accounting */
2094 channels_on_cpu_dec(info);
2097 clear_evtchn_to_irq_all();
2099 for_each_possible_cpu(cpu) {
2100 restore_cpu_virqs(cpu);
2101 restore_cpu_ipis(cpu);
2107 static struct irq_chip xen_dynamic_chip __read_mostly = {
2110 .irq_disable = disable_dynirq,
2111 .irq_mask = disable_dynirq,
2112 .irq_unmask = enable_dynirq,
2114 .irq_ack = ack_dynirq,
2115 .irq_mask_ack = mask_ack_dynirq,
2117 .irq_set_affinity = set_affinity_irq,
2118 .irq_retrigger = retrigger_dynirq,
2121 static struct irq_chip xen_lateeoi_chip __read_mostly = {
2122 /* The chip name needs to contain "xen-dyn" for irqbalance to work. */
2123 .name = "xen-dyn-lateeoi",
2125 .irq_disable = disable_dynirq,
2126 .irq_mask = disable_dynirq,
2127 .irq_unmask = enable_dynirq,
2129 .irq_ack = lateeoi_ack_dynirq,
2130 .irq_mask_ack = lateeoi_mask_ack_dynirq,
2132 .irq_set_affinity = set_affinity_irq,
2133 .irq_retrigger = retrigger_dynirq,
2136 static struct irq_chip xen_pirq_chip __read_mostly = {
2139 .irq_startup = startup_pirq,
2140 .irq_shutdown = shutdown_pirq,
2141 .irq_enable = enable_pirq,
2142 .irq_disable = disable_pirq,
2144 .irq_mask = disable_dynirq,
2145 .irq_unmask = enable_dynirq,
2147 .irq_ack = eoi_pirq,
2148 .irq_eoi = eoi_pirq,
2149 .irq_mask_ack = mask_ack_pirq,
2151 .irq_set_affinity = set_affinity_irq,
2153 .irq_retrigger = retrigger_dynirq,
2156 static struct irq_chip xen_percpu_chip __read_mostly = {
2157 .name = "xen-percpu",
2159 .irq_disable = disable_dynirq,
2160 .irq_mask = disable_dynirq,
2161 .irq_unmask = enable_dynirq,
2163 .irq_ack = ack_dynirq,
2167 #ifdef CONFIG_XEN_PVHVM
2168 /* Vector callbacks are better than PCI interrupts to receive event
2169 * channel notifications because we can receive vector callbacks on any
2170 * vcpu and we don't need PCI support or APIC interactions. */
2171 void xen_setup_callback_vector(void)
2173 uint64_t callback_via;
2175 if (xen_have_vector_callback) {
2176 callback_via = HVM_CALLBACK_VECTOR(HYPERVISOR_CALLBACK_VECTOR);
2177 if (xen_set_callback_via(callback_via)) {
2178 pr_err("Request for Xen HVM callback vector failed\n");
2179 xen_have_vector_callback = false;
2185 * Setup per-vCPU vector-type callbacks. If this setup is unavailable,
2186 * fallback to the global vector-type callback.
2188 static __init void xen_init_setup_upcall_vector(void)
2190 if (!xen_have_vector_callback)
2193 if ((cpuid_eax(xen_cpuid_base() + 4) & XEN_HVM_CPUID_UPCALL_VECTOR) &&
2194 !xen_set_upcall_vector(0))
2195 xen_percpu_upcall = true;
2196 else if (xen_feature(XENFEAT_hvm_callback_vector))
2197 xen_setup_callback_vector();
2199 xen_have_vector_callback = false;
2202 int xen_set_upcall_vector(unsigned int cpu)
2205 xen_hvm_evtchn_upcall_vector_t op = {
2206 .vector = HYPERVISOR_CALLBACK_VECTOR,
2207 .vcpu = per_cpu(xen_vcpu_id, cpu),
2210 rc = HYPERVISOR_hvm_op(HVMOP_set_evtchn_upcall_vector, &op);
2214 /* Trick toolstack to think we are enlightened. */
2216 rc = xen_set_callback_via(1);
2221 static __init void xen_alloc_callback_vector(void)
2223 if (!xen_have_vector_callback)
2226 pr_info("Xen HVM callback vector for event delivery is enabled\n");
2227 sysvec_install(HYPERVISOR_CALLBACK_VECTOR, sysvec_xen_hvm_callback);
2230 void xen_setup_callback_vector(void) {}
2231 static inline void xen_init_setup_upcall_vector(void) {}
2232 int xen_set_upcall_vector(unsigned int cpu) {}
2233 static inline void xen_alloc_callback_vector(void) {}
2234 #endif /* CONFIG_XEN_PVHVM */
2235 #endif /* CONFIG_X86 */
2237 bool xen_fifo_events = true;
2238 module_param_named(fifo_events, xen_fifo_events, bool, 0);
2240 static int xen_evtchn_cpu_prepare(unsigned int cpu)
2244 xen_cpu_init_eoi(cpu);
2246 if (evtchn_ops->percpu_init)
2247 ret = evtchn_ops->percpu_init(cpu);
2252 static int xen_evtchn_cpu_dead(unsigned int cpu)
2256 if (evtchn_ops->percpu_deinit)
2257 ret = evtchn_ops->percpu_deinit(cpu);
2262 void __init xen_init_IRQ(void)
2265 evtchn_port_t evtchn;
2267 if (xen_fifo_events)
2268 ret = xen_evtchn_fifo_init();
2270 xen_evtchn_2l_init();
2271 xen_fifo_events = false;
2274 xen_cpu_init_eoi(smp_processor_id());
2276 cpuhp_setup_state_nocalls(CPUHP_XEN_EVTCHN_PREPARE,
2277 "xen/evtchn:prepare",
2278 xen_evtchn_cpu_prepare, xen_evtchn_cpu_dead);
2280 evtchn_to_irq = kcalloc(EVTCHN_ROW(xen_evtchn_max_channels()),
2281 sizeof(*evtchn_to_irq), GFP_KERNEL);
2282 BUG_ON(!evtchn_to_irq);
2284 /* No event channels are 'live' right now. */
2285 for (evtchn = 0; evtchn < xen_evtchn_nr_channels(); evtchn++)
2286 mask_evtchn(evtchn);
2288 pirq_needs_eoi = pirq_needs_eoi_flag;
2291 if (xen_pv_domain()) {
2292 if (xen_initial_domain())
2293 pci_xen_initial_domain();
2295 xen_init_setup_upcall_vector();
2296 xen_alloc_callback_vector();
2299 if (xen_hvm_domain()) {
2301 /* pci_xen_hvm_init must be called after native_init_IRQ so that
2302 * __acpi_register_gsi can point at the right function */
2306 struct physdev_pirq_eoi_gmfn eoi_gmfn;
2308 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
2309 eoi_gmfn.gmfn = virt_to_gfn(pirq_eoi_map);
2310 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
2312 free_page((unsigned long) pirq_eoi_map);
2313 pirq_eoi_map = NULL;
2315 pirq_needs_eoi = pirq_check_eoi_map;