xen: event channels
authorJeremy Fitzhardinge <jeremy@xensource.com>
Wed, 18 Jul 2007 01:37:05 +0000 (18:37 -0700)
committerJeremy Fitzhardinge <jeremy@goop.org>
Wed, 18 Jul 2007 15:47:42 +0000 (08:47 -0700)
Xen implements interrupts in terms of event channels.  Each guest
domain gets 1024 event channels which can be used for a variety of
purposes, such as Xen timer events, inter-domain events,
inter-processor events (IPI) or for real hardware IRQs.

Within the kernel, we map the event channels to IRQs, and implement
the whole interrupt handling using a Xen irq_chip.

Rather than setting NR_IRQ to 1024 under PARAVIRT in order to
accomodate Xen, we create a dynamic mapping between event channels and
IRQs.  Ideally, Linux will eventually move towards dynamically
allocating per-irq structures, and we can use a 1:1 mapping between
event channels and irqs.

Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Eric W. Biederman <ebiederm@xmission.com>
arch/i386/xen/Makefile
arch/i386/xen/enlighten.c
arch/i386/xen/events.c [new file with mode: 0644]
include/xen/events.h [new file with mode: 0644]

index 803c1ee2b768472e7281c4702847a7767fdfbd00..7a78f27bfb1600a108c7a8892d0a2f39cf58a76c 100644 (file)
@@ -1 +1,2 @@
-obj-y          := enlighten.o setup.o features.o multicalls.o mmu.o
+obj-y          := enlighten.o setup.o features.o multicalls.o mmu.o \
+                       events.o
index c0b0aa7af1450db870b8a8ba3f58231ae2855846..6417dfdccb4cf53b125b7a69333cf32a334705e3 100644 (file)
@@ -607,6 +607,7 @@ static const struct paravirt_ops xen_paravirt_ops __initdata = {
 
        .memory_setup = xen_memory_setup,
        .arch_setup = xen_arch_setup,
+       .init_IRQ = xen_init_IRQ,
 
        .cpuid = xen_cpuid,
 
diff --git a/arch/i386/xen/events.c b/arch/i386/xen/events.c
new file mode 100644 (file)
index 0000000..e7c5d00
--- /dev/null
@@ -0,0 +1,511 @@
+/*
+ * Xen event channels
+ *
+ * Xen models interrupts with abstract event channels.  Because each
+ * domain gets 1024 event channels, but NR_IRQ is not that large, we
+ * must dynamically map irqs<->event channels.  The event channels
+ * interface with the rest of the kernel by defining a xen interrupt
+ * chip.  When an event is recieved, it is mapped to an irq and sent
+ * through the normal interrupt processing path.
+ *
+ * There are four kinds of events which can be mapped to an event
+ * channel:
+ *
+ * 1. Inter-domain notifications.  This includes all the virtual
+ *    device events, since they're driven by front-ends in another domain
+ *    (typically dom0).
+ * 2. VIRQs, typically used for timers.  These are per-cpu events.
+ * 3. IPIs.
+ * 4. Hardware interrupts. Not supported at present.
+ *
+ * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
+ */
+
+#include <linux/linkage.h>
+#include <linux/interrupt.h>
+#include <linux/irq.h>
+#include <linux/module.h>
+#include <linux/string.h>
+
+#include <asm/ptrace.h>
+#include <asm/irq.h>
+#include <asm/sync_bitops.h>
+#include <asm/xen/hypercall.h>
+
+#include <xen/events.h>
+#include <xen/interface/xen.h>
+#include <xen/interface/event_channel.h>
+
+#include "xen-ops.h"
+
+/*
+ * This lock protects updates to the following mapping and reference-count
+ * arrays. The lock does not need to be acquired to read the mapping tables.
+ */
+static DEFINE_SPINLOCK(irq_mapping_update_lock);
+
+/* IRQ <-> VIRQ mapping. */
+static DEFINE_PER_CPU(int, virq_to_irq[NR_VIRQS]) = {[0 ... NR_VIRQS-1] = -1};
+
+/* Packed IRQ information: binding type, sub-type index, and event channel. */
+struct packed_irq
+{
+       unsigned short evtchn;
+       unsigned char index;
+       unsigned char type;
+};
+
+static struct packed_irq irq_info[NR_IRQS];
+
+/* Binding types. */
+enum { IRQT_UNBOUND, IRQT_PIRQ, IRQT_VIRQ, IRQT_IPI, IRQT_EVTCHN };
+
+/* Convenient shorthand for packed representation of an unbound IRQ. */
+#define IRQ_UNBOUND    mk_irq_info(IRQT_UNBOUND, 0, 0)
+
+static int evtchn_to_irq[NR_EVENT_CHANNELS] = {
+       [0 ... NR_EVENT_CHANNELS-1] = -1
+};
+static unsigned long cpu_evtchn_mask[NR_CPUS][NR_EVENT_CHANNELS/BITS_PER_LONG];
+static u8 cpu_evtchn[NR_EVENT_CHANNELS];
+
+/* Reference counts for bindings to IRQs. */
+static int irq_bindcount[NR_IRQS];
+
+/* Xen will never allocate port zero for any purpose. */
+#define VALID_EVTCHN(chn)      ((chn) != 0)
+
+/*
+ * Force a proper event-channel callback from Xen after clearing the
+ * callback mask. We do this in a very simple manner, by making a call
+ * down into Xen. The pending flag will be checked by Xen on return.
+ */
+void force_evtchn_callback(void)
+{
+       (void)HYPERVISOR_xen_version(0, NULL);
+}
+EXPORT_SYMBOL_GPL(force_evtchn_callback);
+
+static struct irq_chip xen_dynamic_chip;
+
+/* Constructor for packed IRQ information. */
+static inline struct packed_irq mk_irq_info(u32 type, u32 index, u32 evtchn)
+{
+       return (struct packed_irq) { evtchn, index, type };
+}
+
+/*
+ * Accessors for packed IRQ information.
+ */
+static inline unsigned int evtchn_from_irq(int irq)
+{
+       return irq_info[irq].evtchn;
+}
+
+static inline unsigned int index_from_irq(int irq)
+{
+       return irq_info[irq].index;
+}
+
+static inline unsigned int type_from_irq(int irq)
+{
+       return irq_info[irq].type;
+}
+
+static inline unsigned long active_evtchns(unsigned int cpu,
+                                          struct shared_info *sh,
+                                          unsigned int idx)
+{
+       return (sh->evtchn_pending[idx] &
+               cpu_evtchn_mask[cpu][idx] &
+               ~sh->evtchn_mask[idx]);
+}
+
+static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
+{
+       int irq = evtchn_to_irq[chn];
+
+       BUG_ON(irq == -1);
+#ifdef CONFIG_SMP
+       irq_desc[irq].affinity = cpumask_of_cpu(cpu);
+#endif
+
+       __clear_bit(chn, cpu_evtchn_mask[cpu_evtchn[chn]]);
+       __set_bit(chn, cpu_evtchn_mask[cpu]);
+
+       cpu_evtchn[chn] = cpu;
+}
+
+static void init_evtchn_cpu_bindings(void)
+{
+#ifdef CONFIG_SMP
+       int i;
+       /* By default all event channels notify CPU#0. */
+       for (i = 0; i < NR_IRQS; i++)
+               irq_desc[i].affinity = cpumask_of_cpu(0);
+#endif
+
+       memset(cpu_evtchn, 0, sizeof(cpu_evtchn));
+       memset(cpu_evtchn_mask[0], ~0, sizeof(cpu_evtchn_mask[0]));
+}
+
+static inline unsigned int cpu_from_evtchn(unsigned int evtchn)
+{
+       return cpu_evtchn[evtchn];
+}
+
+static inline void clear_evtchn(int port)
+{
+       struct shared_info *s = HYPERVISOR_shared_info;
+       sync_clear_bit(port, &s->evtchn_pending[0]);
+}
+
+static inline void set_evtchn(int port)
+{
+       struct shared_info *s = HYPERVISOR_shared_info;
+       sync_set_bit(port, &s->evtchn_pending[0]);
+}
+
+
+/**
+ * notify_remote_via_irq - send event to remote end of event channel via irq
+ * @irq: irq of event channel to send event to
+ *
+ * Unlike notify_remote_via_evtchn(), this is safe to use across
+ * save/restore. Notifications on a broken connection are silently
+ * dropped.
+ */
+void notify_remote_via_irq(int irq)
+{
+       int evtchn = evtchn_from_irq(irq);
+
+       if (VALID_EVTCHN(evtchn))
+               notify_remote_via_evtchn(evtchn);
+}
+EXPORT_SYMBOL_GPL(notify_remote_via_irq);
+
+static void mask_evtchn(int port)
+{
+       struct shared_info *s = HYPERVISOR_shared_info;
+       sync_set_bit(port, &s->evtchn_mask[0]);
+}
+
+static void unmask_evtchn(int port)
+{
+       struct shared_info *s = HYPERVISOR_shared_info;
+       unsigned int cpu = get_cpu();
+
+       BUG_ON(!irqs_disabled());
+
+       /* Slow path (hypercall) if this is a non-local port. */
+       if (unlikely(cpu != cpu_from_evtchn(port))) {
+               struct evtchn_unmask unmask = { .port = port };
+               (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
+       } else {
+               struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
+
+               sync_clear_bit(port, &s->evtchn_mask[0]);
+
+               /*
+                * The following is basically the equivalent of
+                * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
+                * the interrupt edge' if the channel is masked.
+                */
+               if (sync_test_bit(port, &s->evtchn_pending[0]) &&
+                   !sync_test_and_set_bit(port / BITS_PER_LONG,
+                                          &vcpu_info->evtchn_pending_sel))
+                       vcpu_info->evtchn_upcall_pending = 1;
+       }
+
+       put_cpu();
+}
+
+static int find_unbound_irq(void)
+{
+       int irq;
+
+       /* Only allocate from dynirq range */
+       for (irq = 0; irq < NR_IRQS; irq++)
+               if (irq_bindcount[irq] == 0)
+                       break;
+
+       if (irq == NR_IRQS)
+               panic("No available IRQ to bind to: increase NR_IRQS!\n");
+
+       return irq;
+}
+
+static int bind_evtchn_to_irq(unsigned int evtchn)
+{
+       int irq;
+
+       spin_lock(&irq_mapping_update_lock);
+
+       irq = evtchn_to_irq[evtchn];
+
+       if (irq == -1) {
+               irq = find_unbound_irq();
+
+               dynamic_irq_init(irq);
+               set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
+                                             handle_level_irq, "event");
+
+               evtchn_to_irq[evtchn] = irq;
+               irq_info[irq] = mk_irq_info(IRQT_EVTCHN, 0, evtchn);
+       }
+
+       irq_bindcount[irq]++;
+
+       spin_unlock(&irq_mapping_update_lock);
+
+       return irq;
+}
+
+static int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
+{
+       struct evtchn_bind_virq bind_virq;
+       int evtchn, irq;
+
+       spin_lock(&irq_mapping_update_lock);
+
+       irq = per_cpu(virq_to_irq, cpu)[virq];
+
+       if (irq == -1) {
+               bind_virq.virq = virq;
+               bind_virq.vcpu = cpu;
+               if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
+                                               &bind_virq) != 0)
+                       BUG();
+               evtchn = bind_virq.port;
+
+               irq = find_unbound_irq();
+
+               dynamic_irq_init(irq);
+               set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
+                                             handle_level_irq, "virq");
+
+               evtchn_to_irq[evtchn] = irq;
+               irq_info[irq] = mk_irq_info(IRQT_VIRQ, virq, evtchn);
+
+               per_cpu(virq_to_irq, cpu)[virq] = irq;
+
+               bind_evtchn_to_cpu(evtchn, cpu);
+       }
+
+       irq_bindcount[irq]++;
+
+       spin_unlock(&irq_mapping_update_lock);
+
+       return irq;
+}
+
+static void unbind_from_irq(unsigned int irq)
+{
+       struct evtchn_close close;
+       int evtchn = evtchn_from_irq(irq);
+
+       spin_lock(&irq_mapping_update_lock);
+
+       if (VALID_EVTCHN(evtchn) && (--irq_bindcount[irq] == 0)) {
+               close.port = evtchn;
+               if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
+                       BUG();
+
+               switch (type_from_irq(irq)) {
+               case IRQT_VIRQ:
+                       per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
+                               [index_from_irq(irq)] = -1;
+                       break;
+               default:
+                       break;
+               }
+
+               /* Closed ports are implicitly re-bound to VCPU0. */
+               bind_evtchn_to_cpu(evtchn, 0);
+
+               evtchn_to_irq[evtchn] = -1;
+               irq_info[irq] = IRQ_UNBOUND;
+
+               dynamic_irq_init(irq);
+       }
+
+       spin_unlock(&irq_mapping_update_lock);
+}
+
+int bind_evtchn_to_irqhandler(unsigned int evtchn,
+                             irqreturn_t (*handler)(int, void *),
+                             unsigned long irqflags,
+                             const char *devname, void *dev_id)
+{
+       unsigned int irq;
+       int retval;
+
+       irq = bind_evtchn_to_irq(evtchn);
+       retval = request_irq(irq, handler, irqflags, devname, dev_id);
+       if (retval != 0) {
+               unbind_from_irq(irq);
+               return retval;
+       }
+
+       return irq;
+}
+EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
+
+int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
+                           irqreturn_t (*handler)(int, void *),
+                           unsigned long irqflags, const char *devname, void *dev_id)
+{
+       unsigned int irq;
+       int retval;
+
+       irq = bind_virq_to_irq(virq, cpu);
+       retval = request_irq(irq, handler, irqflags, devname, dev_id);
+       if (retval != 0) {
+               unbind_from_irq(irq);
+               return retval;
+       }
+
+       return irq;
+}
+EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
+
+void unbind_from_irqhandler(unsigned int irq, void *dev_id)
+{
+       free_irq(irq, dev_id);
+       unbind_from_irq(irq);
+}
+EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
+
+/*
+ * Search the CPUs pending events bitmasks.  For each one found, map
+ * the event number to an irq, and feed it into do_IRQ() for
+ * handling.
+ *
+ * Xen uses a two-level bitmap to speed searching.  The first level is
+ * a bitset of words which contain pending event bits.  The second
+ * level is a bitset of pending events themselves.
+ */
+fastcall void xen_evtchn_do_upcall(struct pt_regs *regs)
+{
+       int cpu = get_cpu();
+       struct shared_info *s = HYPERVISOR_shared_info;
+       struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
+       unsigned long pending_words;
+
+       vcpu_info->evtchn_upcall_pending = 0;
+
+       /* NB. No need for a barrier here -- XCHG is a barrier on x86. */
+       pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
+       while (pending_words != 0) {
+               unsigned long pending_bits;
+               int word_idx = __ffs(pending_words);
+               pending_words &= ~(1UL << word_idx);
+
+               while ((pending_bits = active_evtchns(cpu, s, word_idx)) != 0) {
+                       int bit_idx = __ffs(pending_bits);
+                       int port = (word_idx * BITS_PER_LONG) + bit_idx;
+                       int irq = evtchn_to_irq[port];
+
+                       if (irq != -1) {
+                               regs->orig_eax = ~irq;
+                               do_IRQ(regs);
+                       }
+               }
+       }
+
+       put_cpu();
+}
+
+/* Rebind an evtchn so that it gets delivered to a specific cpu */
+static void rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
+{
+       struct evtchn_bind_vcpu bind_vcpu;
+       int evtchn = evtchn_from_irq(irq);
+
+       if (!VALID_EVTCHN(evtchn))
+               return;
+
+       /* Send future instances of this interrupt to other vcpu. */
+       bind_vcpu.port = evtchn;
+       bind_vcpu.vcpu = tcpu;
+
+       /*
+        * If this fails, it usually just indicates that we're dealing with a
+        * virq or IPI channel, which don't actually need to be rebound. Ignore
+        * it, but don't do the xenlinux-level rebind in that case.
+        */
+       if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
+               bind_evtchn_to_cpu(evtchn, tcpu);
+}
+
+
+static void set_affinity_irq(unsigned irq, cpumask_t dest)
+{
+       unsigned tcpu = first_cpu(dest);
+       rebind_irq_to_cpu(irq, tcpu);
+}
+
+static void enable_dynirq(unsigned int irq)
+{
+       int evtchn = evtchn_from_irq(irq);
+
+       if (VALID_EVTCHN(evtchn))
+               unmask_evtchn(evtchn);
+}
+
+static void disable_dynirq(unsigned int irq)
+{
+       int evtchn = evtchn_from_irq(irq);
+
+       if (VALID_EVTCHN(evtchn))
+               mask_evtchn(evtchn);
+}
+
+static void ack_dynirq(unsigned int irq)
+{
+       int evtchn = evtchn_from_irq(irq);
+
+       move_native_irq(irq);
+
+       if (VALID_EVTCHN(evtchn))
+               clear_evtchn(evtchn);
+}
+
+static int retrigger_dynirq(unsigned int irq)
+{
+       int evtchn = evtchn_from_irq(irq);
+       int ret = 0;
+
+       if (VALID_EVTCHN(evtchn)) {
+               set_evtchn(evtchn);
+               ret = 1;
+       }
+
+       return ret;
+}
+
+static struct irq_chip xen_dynamic_chip __read_mostly = {
+       .name           = "xen-dyn",
+       .mask           = disable_dynirq,
+       .unmask         = enable_dynirq,
+       .ack            = ack_dynirq,
+       .set_affinity   = set_affinity_irq,
+       .retrigger      = retrigger_dynirq,
+};
+
+void __init xen_init_IRQ(void)
+{
+       int i;
+
+       init_evtchn_cpu_bindings();
+
+       /* No event channels are 'live' right now. */
+       for (i = 0; i < NR_EVENT_CHANNELS; i++)
+               mask_evtchn(i);
+
+       /* Dynamic IRQ space is currently unbound. Zero the refcnts. */
+       for (i = 0; i < NR_IRQS; i++)
+               irq_bindcount[i] = 0;
+
+       irq_ctx_init(smp_processor_id());
+}
diff --git a/include/xen/events.h b/include/xen/events.h
new file mode 100644 (file)
index 0000000..77f71c9
--- /dev/null
@@ -0,0 +1,28 @@
+#ifndef _XEN_EVENTS_H
+#define _XEN_EVENTS_H
+
+#include <linux/irq.h>
+
+int bind_evtchn_to_irqhandler(unsigned int evtchn,
+                             irqreturn_t (*handler)(int, void *),
+                             unsigned long irqflags, const char *devname,
+                             void *dev_id);
+int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
+                           irqreturn_t (*handler)(int, void *),
+                           unsigned long irqflags, const char *devname, void *dev_id);
+
+/*
+ * Common unbind function for all event sources. Takes IRQ to unbind from.
+ * Automatically closes the underlying event channel (even for bindings
+ * made with bind_evtchn_to_irqhandler()).
+ */
+void unbind_from_irqhandler(unsigned int irq, void *dev_id);
+
+static inline void notify_remote_via_evtchn(int port)
+{
+       struct evtchn_send send = { .port = port };
+       (void)HYPERVISOR_event_channel_op(EVTCHNOP_send, &send);
+}
+
+extern void notify_remote_via_irq(int irq);
+#endif /* _XEN_EVENTS_H */