[RFC PATCH 3/5] hw/xen: Support GSI mapping to PIRQ

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

If I advertise XENFEAT_hvm_pirqs then a guest now boots successfully as
long as I tell it 'pci=nomsi'.

[root@localhost ~]# cat /proc/interrupts
           CPU0
  0:         52   IO-APIC   2-edge      timer
  1:         16  xen-pirq   1-ioapic-edge  i8042
  4:       1534  xen-pirq   4-ioapic-edge  ttyS0
  8:          1  xen-pirq   8-ioapic-edge  rtc0
  9:          0  xen-pirq   9-ioapic-level  acpi
 11:       5648  xen-pirq  11-ioapic-level  ahci[0000:00:04.0]
 12:        257  xen-pirq  12-ioapic-edge  i8042
...

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
---
 hw/i386/kvm/xen_evtchn.c | 38 +++++++++++++++++++++++++++++++++++++-
 hw/i386/kvm/xen_evtchn.h |  2 ++
 hw/i386/x86.c            | 15 +++++++++++++++
 3 files changed, 54 insertions(+), 1 deletion(-)

diff --git a/hw/i386/kvm/xen_evtchn.c b/hw/i386/kvm/xen_evtchn.c
index 82250daecb..18c88229ab 100644
--- a/hw/i386/kvm/xen_evtchn.c
+++ b/hw/i386/kvm/xen_evtchn.c
@@ -143,6 +143,7 @@ struct XenEvtchnState {
     uint16_t gsi_pirq[GSI_NUM_PINS];
     /* Bitmap of allocated PIRQs (serialized) */
     uint64_t pirq_inuse[DIV_ROUND_UP(MAX_XEN_PIRQ, 64)];
+    uint32_t pirq_gsi_set;
 
     /* Per-PIRQ information (rebuilt on migration) */
     struct pirq_info pirq[MAX_XEN_PIRQ];
@@ -208,6 +209,7 @@ static const VMStateDescription xen_evtchn_vmstate = {
         VMSTATE_UINT16_ARRAY(gsi_pirq, XenEvtchnState, GSI_NUM_PINS),
         VMSTATE_UINT64_ARRAY(pirq_inuse, XenEvtchnState,
                              DIV_ROUND_UP(MAX_XEN_PIRQ, 64)),
+        VMSTATE_UINT32(pirq_gsi_set, XenEvtchnState),
         VMSTATE_END_OF_LIST()
     }
 };
@@ -1425,6 +1427,35 @@ static int allocate_pirq(XenEvtchnState *s, int type, 
int gsi)
     return pirq;
 }
 
+bool xen_evtchn_set_gsi(int gsi, int level)
+{
+    XenEvtchnState *s = xen_evtchn_singleton;
+    int pirq;
+
+    if (!s || gsi < 0 || gsi > GSI_NUM_PINS) {
+        return false;
+    }
+
+    QEMU_LOCK_GUARD(&s->port_lock);
+
+    pirq = s->gsi_pirq[gsi];
+    if (!pirq) {
+        return false;
+    }
+
+    if (level) {
+        int port = s->pirq[pirq].port;
+
+        s->pirq_gsi_set |= (1U << gsi);
+        if (port) {
+            set_port_pending(s, port);
+        }
+    } else {
+        s->pirq_gsi_set &= ~(1U << gsi);
+    }
+    return true;
+}
+
 int xen_physdev_map_pirq(struct physdev_map_pirq *map)
 {
     XenEvtchnState *s = xen_evtchn_singleton;
@@ -1531,8 +1562,13 @@ int xen_physdev_eoi_pirq(struct physdev_eoi *eoi)
     if (gsi < 0) {
         return -EINVAL;
     }
+    if (s->pirq_gsi_set & (1U << gsi)) {
+        int port = s->pirq[pirq].port;
+        if (port) {
+            set_port_pending(s, port);
+        }
+    }
 
-    // XX: Reassert a level IRQ if needed */
     return 0;
 }
 
diff --git a/hw/i386/kvm/xen_evtchn.h b/hw/i386/kvm/xen_evtchn.h
index 2c12506cc2..dba9d6b021 100644
--- a/hw/i386/kvm/xen_evtchn.h
+++ b/hw/i386/kvm/xen_evtchn.h
@@ -24,6 +24,8 @@ void xen_evtchn_set_callback_level(int level);
 
 int xen_evtchn_set_port(uint16_t port);
 
+bool xen_evtchn_set_gsi(int gsi, int level);
+
 /*
  * These functions mirror the libxenevtchn library API, providing the QEMU
  * backend side of "interdomain" event channels.
diff --git a/hw/i386/x86.c b/hw/i386/x86.c
index 78cc131926..201fd5626c 100644
--- a/hw/i386/x86.c
+++ b/hw/i386/x86.c
@@ -61,6 +61,10 @@
 #include CONFIG_DEVICES
 #include "kvm/kvm_i386.h"
 
+#ifdef CONFIG_XEN_EMU
+#include "hw/i386/kvm/xen_evtchn.h"
+#endif
+
 /* Physical Address of PVH entry point read from kernel ELF NOTE */
 static size_t pvh_start_addr;
 
@@ -608,6 +612,17 @@ void gsi_handler(void *opaque, int n, int level)
         }
         /* fall through */
     case ISA_NUM_IRQS ... IOAPIC_NUM_PINS - 1:
+#ifdef CONFIG_XEN_EMU
+        /*
+         * Xen delivers the GSI to the Legacy PIC (not that Legacy PIC
+         * routing actually works properly under Xen). And then to
+         * *either* the PIRQ handling or the I/OAPIC depending on
+         * whether the former wants it.
+         */
+        if (xen_evtchn_set_gsi(n, level)) {
+            break;
+        }
+#endif
         qemu_set_irq(s->ioapic_irq[n], level);
         break;
     case IO_APIC_SECONDARY_IRQBASE
-- 
2.35.3