[PATCH 2/2] linux/x86: use shared page indicating the need for an EOI notification

["Jan Beulich" <jbeulich@novell.com>]

As usual, written and tested on 2.6.27.7 and made apply to the 2.6.18
tree without further testing.

Signed-off-by: Jan Beulich <jbeulich@novell.com>

Index: head-2008-11-25/drivers/xen/core/evtchn.c
===================================================================
--- head-2008-11-25.orig/drivers/xen/core/evtchn.c	2008-11-26 15:59:04.000000000 +0100
+++ head-2008-11-25/drivers/xen/core/evtchn.c	2008-11-26 16:00:52.000000000 +0100
@@ -130,9 +130,6 @@ DEFINE_PER_CPU(int, ipi_to_irq[NR_IPIS])
 /* Reference counts for bindings to IRQs. */
 static int irq_bindcount[NR_IRQS];
 
-/* Bitmap indicating which PIRQs require Xen to be notified on unmask. */
-static DECLARE_BITMAP(pirq_needs_eoi, NR_PIRQS);
-
 #ifdef CONFIG_SMP
 
 static u8 cpu_evtchn[NR_EVENT_CHANNELS];
@@ -786,16 +783,48 @@ static struct irq_chip dynirq_chip = {
 	.retrigger = resend_irq_on_evtchn,
 };
 
-static inline void pirq_unmask_notify(int irq)
+/* Bitmap indicating which PIRQs require Xen to be notified on unmask. */
+static bool pirq_eoi_does_unmask;
+static DECLARE_BITMAP(pirq_needs_eoi, ALIGN(NR_PIRQS, PAGE_SIZE * 8))
+	__attribute__ ((__section__(".bss.page_aligned"), __aligned__(PAGE_SIZE)));
+
+static void pirq_unmask_and_notify(unsigned int evtchn, unsigned int irq)
 {
 	struct physdev_eoi eoi = { .irq = evtchn_get_xen_pirq(irq) };
-	if (unlikely(test_bit(irq - PIRQ_BASE, pirq_needs_eoi)))
-		VOID(HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi));
+
+	if (pirq_eoi_does_unmask) {
+		if (test_bit(eoi.irq, pirq_needs_eoi))
+			VOID(HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi));
+		else
+			unmask_evtchn(evtchn);
+	} else if (test_bit(irq - PIRQ_BASE, pirq_needs_eoi)) {
+		if (smp_processor_id() != cpu_from_evtchn(evtchn)) {
+			struct evtchn_unmask unmask = { .port = evtchn };
+			struct multicall_entry mcl[2];
+
+			mcl[0].op = __HYPERVISOR_event_channel_op;
+			mcl[0].args[0] = EVTCHNOP_unmask;
+			mcl[0].args[1] = (unsigned long)&unmask;
+			mcl[1].op = __HYPERVISOR_physdev_op;
+			mcl[1].args[0] = PHYSDEVOP_eoi;
+			mcl[1].args[1] = (unsigned long)&eoi;
+
+			if (HYPERVISOR_multicall(mcl, 2))
+				BUG();
+		} else {
+			unmask_evtchn(evtchn);
+			VOID(HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi));
+		}
+	} else
+		unmask_evtchn(evtchn);
 }
 
 static inline void pirq_query_unmask(int irq)
 {
 	struct physdev_irq_status_query irq_status;
+
+	if (pirq_eoi_does_unmask)
+		return;
 	irq_status.irq = evtchn_get_xen_pirq(irq);
 	if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 		irq_status.flags = 0;
@@ -836,8 +864,7 @@ static unsigned int startup_pirq(unsigne
 	irq_info[irq] = mk_irq_info(IRQT_PIRQ, bind_pirq.pirq, evtchn);
 
  out:
-	unmask_evtchn(evtchn);
-	pirq_unmask_notify(irq);
+	pirq_unmask_and_notify(evtchn, irq);
 
 	return 0;
 }
@@ -889,10 +916,8 @@ static void end_pirq(unsigned int irq)
 	if ((irq_desc[irq].status & (IRQ_DISABLED|IRQ_PENDING)) ==
 	    (IRQ_DISABLED|IRQ_PENDING)) {
 		shutdown_pirq(irq);
-	} else if (VALID_EVTCHN(evtchn)) {
-		unmask_evtchn(evtchn);
-		pirq_unmask_notify(irq);
-	}
+	} else if (VALID_EVTCHN(evtchn))
+		pirq_unmask_and_notify(evtchn, irq);
 }
 
 static struct hw_interrupt_type pirq_type = {
@@ -1045,6 +1070,14 @@ static int evtchn_resume(struct sys_devi
 
 	init_evtchn_cpu_bindings();
 
+	if (pirq_eoi_does_unmask) {
+		struct physdev_pirq_eoi_mfn eoi_mfn;
+
+		eoi_mfn.mfn = virt_to_bus(pirq_needs_eoi) >> PAGE_SHIFT;
+		if (HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_mfn, &eoi_mfn))
+			BUG();
+	}
+
 	/* New event-channel space is not 'live' yet. */
 	for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
 		mask_evtchn(evtchn);
@@ -1167,6 +1200,7 @@ int evtchn_get_xen_pirq(int irq)
 void __init xen_init_IRQ(void)
 {
 	unsigned int i;
+	struct physdev_pirq_eoi_mfn eoi_mfn;
 
 	init_evtchn_cpu_bindings();
 
@@ -1179,6 +1213,11 @@ void __init xen_init_IRQ(void)
 
 	init_evtchn_cpu_bindings();
 
+	BUG_ON(!bitmap_empty(pirq_needs_eoi, PAGE_SIZE * 8));
+	eoi_mfn.mfn = virt_to_bus(pirq_needs_eoi) >> PAGE_SHIFT;
+	if (HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_mfn, &eoi_mfn) == 0)
+		pirq_eoi_does_unmask = true;
+
 	/* No event channels are 'live' right now. */
 	for (i = 0; i < NR_EVENT_CHANNELS; i++)
 		mask_evtchn(i);
Index: head-2008-11-25/include/xen/interface/physdev.h
===================================================================
--- head-2008-11-25.orig/include/xen/interface/physdev.h	2008-11-26 15:59:04.000000000 +0100
+++ head-2008-11-25/include/xen/interface/physdev.h	2008-11-24 15:16:10.000000000 +0100
@@ -41,6 +41,21 @@ typedef struct physdev_eoi physdev_eoi_t
 DEFINE_XEN_GUEST_HANDLE(physdev_eoi_t);
 
 /*
+ * Register a shared page for the hypervisor to indicate whether the guest
+ * must issue PHYSDEVOP_eoi. The semantics of PHYSDEVOP_eoi change slightly
+ * once the guest used this function in that the associated event channel
+ * will automatically get unmasked. The page registered is used as a bit
+ * array indexed by Xen's PIRQ value.
+ */
+#define PHYSDEVOP_pirq_eoi_mfn          17
+struct physdev_pirq_eoi_mfn {
+    /* IN */
+    xen_pfn_t mfn;
+};
+typedef struct physdev_pirq_eoi_mfn physdev_pirq_eoi_mfn_t;
+DEFINE_XEN_GUEST_HANDLE(physdev_pirq_eoi_mfn_t);
+
+/*
  * Query the status of an IRQ line.
  * @arg == pointer to physdev_irq_status_query structure.
  */