From: Don Dutile <ddutile@redhat.com>
Date: Fri, 11 Jan 2008 16:22:34 -0500
Subject: [xen] hvm: evtchn to fake pci interrupt propagation
Message-id: 4787DE1A.6050501@redhat.com
O-Subject: [RHEL5.2 PATCH] [Xen][2/4]: High Intr and Softintr rate in HVM's with pv-on-hvm drivers
Bugzilla: 412721

Patch 2/4: hvm-evtchn-to-fake-pci-intr-3-1.patch

This is the core/major portion of the fix.

It is the xen-3.1-testing upstream patch verbatim, cset 15991:
   http://xenbits.xensource.com/staging/xen-3.1-testing.hg?rev/5a64f1bf9460

It is a backport of the xen-unstable evtchn-to-fake pci interrupt patch,
which resolved the problem for xen-3.2.0 hypervisor.

Please review & ACK.

- Don

# HG changeset patch
# User Keir Fraser <keir.fraser@citrix.com>
# Date 1199965759 0
# Node ID 48c800d5a3718a4a8115c02fc0b76d16f85d35f8
# Parent 337dc1a3f7ca3663d7c057576acb293b08e7a70f
hvm: Fix evtchn-to-fake-pci interrupt propagation.

Previously the evtchn_upcall_pending flag would only ever be sampled
on VCPU0, possibly leading to long delays in deasserting the
fake-pci-device INTx line if the interrupt is actually delivered to
other than VCPU0.

Diagnosed by Ian Jackson <ian.jackson@eu.citrix.com>

Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
xen-unstable changeset:   16692:9865d5e82802fda2ac842c0c28a0833f2126551f
xen-unstable date:        Tue Jan 08 15:55:29 2008 +0000

Acked-by: Bill Burns <bburns@redhat.com>
Acked-by: "Stephen C. Tweedie" <sct@redhat.com>

diff --git a/arch/x86/hvm/irq.c b/arch/x86/hvm/irq.c
index 6b5f387..25ada54 100644
--- a/arch/x86/hvm/irq.c
+++ b/arch/x86/hvm/irq.c
@@ -125,17 +125,13 @@ void hvm_isa_irq_deassert(
     spin_unlock(&d->arch.hvm_domain.irq_lock);
 }
 
-void hvm_set_callback_irq_level(void)
+static void hvm_set_callback_irq_level(struct vcpu *v)
 {
-    struct vcpu *v = current;
     struct domain *d = v->domain;
     struct hvm_irq *hvm_irq = &d->arch.hvm_domain.irq;
     unsigned int gsi, pdev, pintx, asserted;
 
-    /* Fast lock-free tests. */
-    if ( (v->vcpu_id != 0) ||
-         (hvm_irq->callback_via_type == HVMIRQ_callback_none) )
-        return;
+    ASSERT(v->vcpu_id == 0);
 
     spin_lock(&d->arch.hvm_domain.irq_lock);
 
@@ -177,6 +173,22 @@ void hvm_set_callback_irq_level(void)
     spin_unlock(&d->arch.hvm_domain.irq_lock);
 }
 
+void hvm_maybe_deassert_evtchn_irq(void)
+{
+    struct domain *d = current->domain;
+    struct hvm_irq *hvm_irq = &d->arch.hvm_domain.irq;
+
+    if ( hvm_irq->callback_via_asserted &&
+         !vcpu_info(d->vcpu[0], evtchn_upcall_pending) )
+        hvm_set_callback_irq_level(d->vcpu[0]);
+}
+
+void hvm_assert_evtchn_irq(struct vcpu *v)
+{
+    if ( v->vcpu_id == 0 )
+        hvm_set_callback_irq_level(v);
+}
+
 void hvm_set_pci_link_route(struct domain *d, u8 link, u8 isa_irq)
 {
     struct hvm_irq *hvm_irq = &d->arch.hvm_domain.irq;
@@ -337,15 +349,10 @@ int is_isa_irq_masked(struct vcpu *v, int isa_irq)
             domain_vioapic(v->domain)->redirtbl[gsi].fields.mask);
 }
 
-/*
- * TODO: 1. Should not need special treatment of event-channel events.
- *       2. Should take notice of interrupt shadows (or clear them).
- */
 int hvm_local_events_need_delivery(struct vcpu *v)
 {
-    int pending;
+    int pending = cpu_has_pending_irq(v);
 
-    pending = (vcpu_info(v, evtchn_upcall_pending) || cpu_has_pending_irq(v));
     if ( unlikely(pending) )
         pending = hvm_interrupts_enabled(v); 
 
diff --git a/arch/x86/hvm/svm/intr.c b/arch/x86/hvm/svm/intr.c
index e9325e8..4965376 100644
--- a/arch/x86/hvm/svm/intr.c
+++ b/arch/x86/hvm/svm/intr.c
@@ -124,7 +124,7 @@ asmlinkage void svm_intr_assist(void)
 
     /* Crank the handle on interrupt state and check for new interrrupts. */
     pt_update_irq(v);
-    hvm_set_callback_irq_level();
+    hvm_maybe_deassert_evtchn_irq();
     if ( !cpu_has_pending_irq(v) )
         goto out;
 
diff --git a/arch/x86/hvm/vmx/intr.c b/arch/x86/hvm/vmx/intr.c
index 92e0f9c..0f6a0e2 100644
--- a/arch/x86/hvm/vmx/intr.c
+++ b/arch/x86/hvm/vmx/intr.c
@@ -115,7 +115,7 @@ asmlinkage void vmx_intr_assist(void)
 
     pt_update_irq(v);
 
-    hvm_set_callback_irq_level();
+    hvm_maybe_deassert_evtchn_irq();
 
     update_tpr_threshold(vcpu_vlapic(v));
 
diff --git a/include/asm-x86/event.h b/include/asm-x86/event.h
index 32b157b..340ec44 100644
--- a/include/asm-x86/event.h
+++ b/include/asm-x86/event.h
@@ -31,7 +31,12 @@ static inline void vcpu_kick(struct vcpu *v)
 
 static inline void vcpu_mark_events_pending(struct vcpu *v)
 {
-    if ( !test_and_set_bit(0, &vcpu_info(v, evtchn_upcall_pending)) )
+    if ( test_and_set_bit(0, &vcpu_info(v, evtchn_upcall_pending)) )
+        return;
+
+    if ( is_hvm_vcpu(v) )
+        hvm_assert_evtchn_irq(v);
+    else
         vcpu_kick(v);
 }
 
diff --git a/include/asm-x86/hvm/irq.h b/include/asm-x86/hvm/irq.h
index 1095e86..e46bdb4 100644
--- a/include/asm-x86/hvm/irq.h
+++ b/include/asm-x86/hvm/irq.h
@@ -112,7 +112,8 @@ void hvm_isa_irq_deassert(
 
 void hvm_set_pci_link_route(struct domain *d, u8 link, u8 isa_irq);
 
-void hvm_set_callback_irq_level(void);
+void hvm_maybe_deassert_evtchn_irq(void);
+void hvm_assert_evtchn_irq(struct vcpu *v);
 void hvm_set_callback_via(struct domain *d, uint64_t via);
 
 int cpu_get_interrupt(struct vcpu *v, int *type);