ia64/pv_ops/xen: implement xen pv_irq_ops.
Isaku Yamahata [Fri, 17 Oct 2008 02:18:07 +0000 (11:18 +0900)]
implement xen pv_irq_ops to paravirtualize irq handling
with xen event channel.

Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Signed-off-by: Akio Takebe <takebe_akio@jp.fujitsu.com>
Signed-off-by: Alex Williamson <alex.williamson@hp.com>
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
Signed-off-by: Tony Luck <tony.luck@intel.com>

arch/ia64/xen/Makefile
arch/ia64/xen/irq_xen.c [new file with mode: 0644]
arch/ia64/xen/irq_xen.h [new file with mode: 0644]
arch/ia64/xen/xen_pv_ops.c

index 9b77e8a..01c4289 100644 (file)
@@ -2,7 +2,7 @@
 # Makefile for Xen components
 #
 
-obj-y := hypercall.o xenivt.o xensetup.o xen_pv_ops.o \
+obj-y := hypercall.o xenivt.o xensetup.o xen_pv_ops.o irq_xen.o \
         hypervisor.o xencomm.o xcom_hcall.o grant-table.o
 
 AFLAGS_xenivt.o += -D__IA64_ASM_PARAVIRTUALIZED_XEN
diff --git a/arch/ia64/xen/irq_xen.c b/arch/ia64/xen/irq_xen.c
new file mode 100644 (file)
index 0000000..af93aad
--- /dev/null
@@ -0,0 +1,435 @@
+/******************************************************************************
+ * arch/ia64/xen/irq_xen.c
+ *
+ * Copyright (c) 2008 Isaku Yamahata <yamahata at valinux co jp>
+ *                    VA Linux Systems Japan K.K.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ */
+
+#include <linux/cpu.h>
+
+#include <xen/interface/xen.h>
+#include <xen/interface/callback.h>
+#include <xen/events.h>
+
+#include <asm/xen/privop.h>
+
+#include "irq_xen.h"
+
+/***************************************************************************
+ * pv_irq_ops
+ * irq operations
+ */
+
+static int
+xen_assign_irq_vector(int irq)
+{
+       struct physdev_irq irq_op;
+
+       irq_op.irq = irq;
+       if (HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op))
+               return -ENOSPC;
+
+       return irq_op.vector;
+}
+
+static void
+xen_free_irq_vector(int vector)
+{
+       struct physdev_irq irq_op;
+
+       if (vector < IA64_FIRST_DEVICE_VECTOR ||
+           vector > IA64_LAST_DEVICE_VECTOR)
+               return;
+
+       irq_op.vector = vector;
+       if (HYPERVISOR_physdev_op(PHYSDEVOP_free_irq_vector, &irq_op))
+               printk(KERN_WARNING "%s: xen_free_irq_vecotr fail vector=%d\n",
+                      __func__, vector);
+}
+
+
+static DEFINE_PER_CPU(int, timer_irq) = -1;
+static DEFINE_PER_CPU(int, ipi_irq) = -1;
+static DEFINE_PER_CPU(int, resched_irq) = -1;
+static DEFINE_PER_CPU(int, cmc_irq) = -1;
+static DEFINE_PER_CPU(int, cmcp_irq) = -1;
+static DEFINE_PER_CPU(int, cpep_irq) = -1;
+#define NAME_SIZE      15
+static DEFINE_PER_CPU(char[NAME_SIZE], timer_name);
+static DEFINE_PER_CPU(char[NAME_SIZE], ipi_name);
+static DEFINE_PER_CPU(char[NAME_SIZE], resched_name);
+static DEFINE_PER_CPU(char[NAME_SIZE], cmc_name);
+static DEFINE_PER_CPU(char[NAME_SIZE], cmcp_name);
+static DEFINE_PER_CPU(char[NAME_SIZE], cpep_name);
+#undef NAME_SIZE
+
+struct saved_irq {
+       unsigned int irq;
+       struct irqaction *action;
+};
+/* 16 should be far optimistic value, since only several percpu irqs
+ * are registered early.
+ */
+#define MAX_LATE_IRQ   16
+static struct saved_irq saved_percpu_irqs[MAX_LATE_IRQ];
+static unsigned short late_irq_cnt;
+static unsigned short saved_irq_cnt;
+static int xen_slab_ready;
+
+#ifdef CONFIG_SMP
+/* Dummy stub. Though we may check XEN_RESCHEDULE_VECTOR before __do_IRQ,
+ * it ends up to issue several memory accesses upon percpu data and
+ * thus adds unnecessary traffic to other paths.
+ */
+static irqreturn_t
+xen_dummy_handler(int irq, void *dev_id)
+{
+
+       return IRQ_HANDLED;
+}
+
+static struct irqaction xen_ipi_irqaction = {
+       .handler =      handle_IPI,
+       .flags =        IRQF_DISABLED,
+       .name =         "IPI"
+};
+
+static struct irqaction xen_resched_irqaction = {
+       .handler =      xen_dummy_handler,
+       .flags =        IRQF_DISABLED,
+       .name =         "resched"
+};
+
+static struct irqaction xen_tlb_irqaction = {
+       .handler =      xen_dummy_handler,
+       .flags =        IRQF_DISABLED,
+       .name =         "tlb_flush"
+};
+#endif
+
+/*
+ * This is xen version percpu irq registration, which needs bind
+ * to xen specific evtchn sub-system. One trick here is that xen
+ * evtchn binding interface depends on kmalloc because related
+ * port needs to be freed at device/cpu down. So we cache the
+ * registration on BSP before slab is ready and then deal them
+ * at later point. For rest instances happening after slab ready,
+ * we hook them to xen evtchn immediately.
+ *
+ * FIXME: MCA is not supported by far, and thus "nomca" boot param is
+ * required.
+ */
+static void
+__xen_register_percpu_irq(unsigned int cpu, unsigned int vec,
+                       struct irqaction *action, int save)
+{
+       irq_desc_t *desc;
+       int irq = 0;
+
+       if (xen_slab_ready) {
+               switch (vec) {
+               case IA64_TIMER_VECTOR:
+                       snprintf(per_cpu(timer_name, cpu),
+                                sizeof(per_cpu(timer_name, cpu)),
+                                "%s%d", action->name, cpu);
+                       irq = bind_virq_to_irqhandler(VIRQ_ITC, cpu,
+                               action->handler, action->flags,
+                               per_cpu(timer_name, cpu), action->dev_id);
+                       per_cpu(timer_irq, cpu) = irq;
+                       break;
+               case IA64_IPI_RESCHEDULE:
+                       snprintf(per_cpu(resched_name, cpu),
+                                sizeof(per_cpu(resched_name, cpu)),
+                                "%s%d", action->name, cpu);
+                       irq = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR, cpu,
+                               action->handler, action->flags,
+                               per_cpu(resched_name, cpu), action->dev_id);
+                       per_cpu(resched_irq, cpu) = irq;
+                       break;
+               case IA64_IPI_VECTOR:
+                       snprintf(per_cpu(ipi_name, cpu),
+                                sizeof(per_cpu(ipi_name, cpu)),
+                                "%s%d", action->name, cpu);
+                       irq = bind_ipi_to_irqhandler(XEN_IPI_VECTOR, cpu,
+                               action->handler, action->flags,
+                               per_cpu(ipi_name, cpu), action->dev_id);
+                       per_cpu(ipi_irq, cpu) = irq;
+                       break;
+               case IA64_CMC_VECTOR:
+                       snprintf(per_cpu(cmc_name, cpu),
+                                sizeof(per_cpu(cmc_name, cpu)),
+                                "%s%d", action->name, cpu);
+                       irq = bind_virq_to_irqhandler(VIRQ_MCA_CMC, cpu,
+                                                     action->handler,
+                                                     action->flags,
+                                                     per_cpu(cmc_name, cpu),
+                                                     action->dev_id);
+                       per_cpu(cmc_irq, cpu) = irq;
+                       break;
+               case IA64_CMCP_VECTOR:
+                       snprintf(per_cpu(cmcp_name, cpu),
+                                sizeof(per_cpu(cmcp_name, cpu)),
+                                "%s%d", action->name, cpu);
+                       irq = bind_ipi_to_irqhandler(XEN_CMCP_VECTOR, cpu,
+                                                    action->handler,
+                                                    action->flags,
+                                                    per_cpu(cmcp_name, cpu),
+                                                    action->dev_id);
+                       per_cpu(cmcp_irq, cpu) = irq;
+                       break;
+               case IA64_CPEP_VECTOR:
+                       snprintf(per_cpu(cpep_name, cpu),
+                                sizeof(per_cpu(cpep_name, cpu)),
+                                "%s%d", action->name, cpu);
+                       irq = bind_ipi_to_irqhandler(XEN_CPEP_VECTOR, cpu,
+                                                    action->handler,
+                                                    action->flags,
+                                                    per_cpu(cpep_name, cpu),
+                                                    action->dev_id);
+                       per_cpu(cpep_irq, cpu) = irq;
+                       break;
+               case IA64_CPE_VECTOR:
+               case IA64_MCA_RENDEZ_VECTOR:
+               case IA64_PERFMON_VECTOR:
+               case IA64_MCA_WAKEUP_VECTOR:
+               case IA64_SPURIOUS_INT_VECTOR:
+                       /* No need to complain, these aren't supported. */
+                       break;
+               default:
+                       printk(KERN_WARNING "Percpu irq %d is unsupported "
+                              "by xen!\n", vec);
+                       break;
+               }
+               BUG_ON(irq < 0);
+
+               if (irq > 0) {
+                       /*
+                        * Mark percpu.  Without this, migrate_irqs() will
+                        * mark the interrupt for migrations and trigger it
+                        * on cpu hotplug.
+                        */
+                       desc = irq_desc + irq;
+                       desc->status |= IRQ_PER_CPU;
+               }
+       }
+
+       /* For BSP, we cache registered percpu irqs, and then re-walk
+        * them when initializing APs
+        */
+       if (!cpu && save) {
+               BUG_ON(saved_irq_cnt == MAX_LATE_IRQ);
+               saved_percpu_irqs[saved_irq_cnt].irq = vec;
+               saved_percpu_irqs[saved_irq_cnt].action = action;
+               saved_irq_cnt++;
+               if (!xen_slab_ready)
+                       late_irq_cnt++;
+       }
+}
+
+static void
+xen_register_percpu_irq(ia64_vector vec, struct irqaction *action)
+{
+       __xen_register_percpu_irq(smp_processor_id(), vec, action, 1);
+}
+
+static void
+xen_bind_early_percpu_irq(void)
+{
+       int i;
+
+       xen_slab_ready = 1;
+       /* There's no race when accessing this cached array, since only
+        * BSP will face with such step shortly
+        */
+       for (i = 0; i < late_irq_cnt; i++)
+               __xen_register_percpu_irq(smp_processor_id(),
+                                         saved_percpu_irqs[i].irq,
+                                         saved_percpu_irqs[i].action, 0);
+}
+
+/* FIXME: There's no obvious point to check whether slab is ready. So
+ * a hack is used here by utilizing a late time hook.
+ */
+
+#ifdef CONFIG_HOTPLUG_CPU
+static int __devinit
+unbind_evtchn_callback(struct notifier_block *nfb,
+                      unsigned long action, void *hcpu)
+{
+       unsigned int cpu = (unsigned long)hcpu;
+
+       if (action == CPU_DEAD) {
+               /* Unregister evtchn.  */
+               if (per_cpu(cpep_irq, cpu) >= 0) {
+                       unbind_from_irqhandler(per_cpu(cpep_irq, cpu), NULL);
+                       per_cpu(cpep_irq, cpu) = -1;
+               }
+               if (per_cpu(cmcp_irq, cpu) >= 0) {
+                       unbind_from_irqhandler(per_cpu(cmcp_irq, cpu), NULL);
+                       per_cpu(cmcp_irq, cpu) = -1;
+               }
+               if (per_cpu(cmc_irq, cpu) >= 0) {
+                       unbind_from_irqhandler(per_cpu(cmc_irq, cpu), NULL);
+                       per_cpu(cmc_irq, cpu) = -1;
+               }
+               if (per_cpu(ipi_irq, cpu) >= 0) {
+                       unbind_from_irqhandler(per_cpu(ipi_irq, cpu), NULL);
+                       per_cpu(ipi_irq, cpu) = -1;
+               }
+               if (per_cpu(resched_irq, cpu) >= 0) {
+                       unbind_from_irqhandler(per_cpu(resched_irq, cpu),
+                                               NULL);
+                       per_cpu(resched_irq, cpu) = -1;
+               }
+               if (per_cpu(timer_irq, cpu) >= 0) {
+                       unbind_from_irqhandler(per_cpu(timer_irq, cpu), NULL);
+                       per_cpu(timer_irq, cpu) = -1;
+               }
+       }
+       return NOTIFY_OK;
+}
+
+static struct notifier_block unbind_evtchn_notifier = {
+       .notifier_call = unbind_evtchn_callback,
+       .priority = 0
+};
+#endif
+
+void xen_smp_intr_init_early(unsigned int cpu)
+{
+#ifdef CONFIG_SMP
+       unsigned int i;
+
+       for (i = 0; i < saved_irq_cnt; i++)
+               __xen_register_percpu_irq(cpu, saved_percpu_irqs[i].irq,
+                                         saved_percpu_irqs[i].action, 0);
+#endif
+}
+
+void xen_smp_intr_init(void)
+{
+#ifdef CONFIG_SMP
+       unsigned int cpu = smp_processor_id();
+       struct callback_register event = {
+               .type = CALLBACKTYPE_event,
+               .address = { .ip = (unsigned long)&xen_event_callback },
+       };
+
+       if (cpu == 0) {
+               /* Initialization was already done for boot cpu.  */
+#ifdef CONFIG_HOTPLUG_CPU
+               /* Register the notifier only once.  */
+               register_cpu_notifier(&unbind_evtchn_notifier);
+#endif
+               return;
+       }
+
+       /* This should be piggyback when setup vcpu guest context */
+       BUG_ON(HYPERVISOR_callback_op(CALLBACKOP_register, &event));
+#endif /* CONFIG_SMP */
+}
+
+void __init
+xen_irq_init(void)
+{
+       struct callback_register event = {
+               .type = CALLBACKTYPE_event,
+               .address = { .ip = (unsigned long)&xen_event_callback },
+       };
+
+       xen_init_IRQ();
+       BUG_ON(HYPERVISOR_callback_op(CALLBACKOP_register, &event));
+       late_time_init = xen_bind_early_percpu_irq;
+}
+
+void
+xen_platform_send_ipi(int cpu, int vector, int delivery_mode, int redirect)
+{
+#ifdef CONFIG_SMP
+       /* TODO: we need to call vcpu_up here */
+       if (unlikely(vector == ap_wakeup_vector)) {
+               /* XXX
+                * This should be in __cpu_up(cpu) in ia64 smpboot.c
+                * like x86. But don't want to modify it,
+                * keep it untouched.
+                */
+               xen_smp_intr_init_early(cpu);
+
+               xen_send_ipi(cpu, vector);
+               /* vcpu_prepare_and_up(cpu); */
+               return;
+       }
+#endif
+
+       switch (vector) {
+       case IA64_IPI_VECTOR:
+               xen_send_IPI_one(cpu, XEN_IPI_VECTOR);
+               break;
+       case IA64_IPI_RESCHEDULE:
+               xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
+               break;
+       case IA64_CMCP_VECTOR:
+               xen_send_IPI_one(cpu, XEN_CMCP_VECTOR);
+               break;
+       case IA64_CPEP_VECTOR:
+               xen_send_IPI_one(cpu, XEN_CPEP_VECTOR);
+               break;
+       case IA64_TIMER_VECTOR: {
+               /* this is used only once by check_sal_cache_flush()
+                  at boot time */
+               static int used = 0;
+               if (!used) {
+                       xen_send_ipi(cpu, IA64_TIMER_VECTOR);
+                       used = 1;
+                       break;
+               }
+               /* fallthrough */
+       }
+       default:
+               printk(KERN_WARNING "Unsupported IPI type 0x%x\n",
+                      vector);
+               notify_remote_via_irq(0); /* defaults to 0 irq */
+               break;
+       }
+}
+
+static void __init
+xen_register_ipi(void)
+{
+#ifdef CONFIG_SMP
+       register_percpu_irq(IA64_IPI_VECTOR, &xen_ipi_irqaction);
+       register_percpu_irq(IA64_IPI_RESCHEDULE, &xen_resched_irqaction);
+       register_percpu_irq(IA64_IPI_LOCAL_TLB_FLUSH, &xen_tlb_irqaction);
+#endif
+}
+
+static void
+xen_resend_irq(unsigned int vector)
+{
+       (void)resend_irq_on_evtchn(vector);
+}
+
+const struct pv_irq_ops xen_irq_ops __initdata = {
+       .register_ipi = xen_register_ipi,
+
+       .assign_irq_vector = xen_assign_irq_vector,
+       .free_irq_vector = xen_free_irq_vector,
+       .register_percpu_irq = xen_register_percpu_irq,
+
+       .resend_irq = xen_resend_irq,
+};
diff --git a/arch/ia64/xen/irq_xen.h b/arch/ia64/xen/irq_xen.h
new file mode 100644 (file)
index 0000000..26110f3
--- /dev/null
@@ -0,0 +1,34 @@
+/******************************************************************************
+ * arch/ia64/xen/irq_xen.h
+ *
+ * Copyright (c) 2008 Isaku Yamahata <yamahata at valinux co jp>
+ *                    VA Linux Systems Japan K.K.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ */
+
+#ifndef IRQ_XEN_H
+#define IRQ_XEN_H
+
+extern void (*late_time_init)(void);
+extern char xen_event_callback;
+void __init xen_init_IRQ(void);
+
+extern const struct pv_irq_ops xen_irq_ops __initdata;
+extern void xen_smp_intr_init(void);
+extern void xen_send_ipi(int cpu, int vec);
+
+#endif /* IRQ_XEN_H */
index 41a6cbf..4fe4e62 100644 (file)
@@ -29,6 +29,8 @@
 #include <asm/xen/xencomm.h>
 #include <asm/xen/privop.h>
 
+#include "irq_xen.h"
+
 /***************************************************************************
  * general info
  */
@@ -354,6 +356,7 @@ xen_setup_pv_ops(void)
        pv_init_ops = xen_init_ops;
        pv_cpu_ops = xen_cpu_ops;
        pv_iosapic_ops = xen_iosapic_ops;
+       pv_irq_ops = xen_irq_ops;
 
        paravirt_cpu_asm_init(&xen_cpu_asm_switch);
 }