hw/xen: Support GSI mapping to PIRQ

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>
Reviewed-by: Paul Durrant <paul@xen.org>

hw/i386/kvm/xen_evtchn.c

@@ -148,6 +148,9 @@ struct XenEvtchnState {
     /* GSI → PIRQ mapping (serialized) */
     uint16_t gsi_pirq[IOAPIC_NUM_PINS];
 
+    /* Per-GSI assertion state (serialized) */
+    uint32_t pirq_gsi_set;
+
     /* Per-PIRQ information (rebuilt on migration) */
     struct pirq_info *pirq;
 };
@@ -246,6 +249,7 @@ static const VMStateDescription xen_evtchn_vmstate = {
         VMSTATE_VARRAY_UINT16_ALLOC(pirq_inuse_bitmap, XenEvtchnState,
                                     nr_pirq_inuse_words, 0,
                                     vmstate_info_uint64, uint64_t),
+        VMSTATE_UINT32(pirq_gsi_set, XenEvtchnState),
         VMSTATE_END_OF_LIST()
     }
 };
@@ -1510,6 +1514,51 @@ 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;
+
+    assert(qemu_mutex_iothread_locked());
+
+    if (!s || gsi < 0 || gsi > IOAPIC_NUM_PINS) {
+        return false;
+    }
+
+    /*
+     * Check that that it *isn't* the event channel GSI, and thus
+     * that we are not recursing and it's safe to take s->port_lock.
+     *
+     * Locking aside, it's perfectly sane to bail out early for that
+     * special case, as it would make no sense for the event channel
+     * GSI to be routed back to event channels, when the delivery
+     * method is to raise the GSI... that recursion wouldn't *just*
+     * be a locking issue.
+     */
+    if (gsi && gsi == s->callback_gsi) {
+        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;
@@ -1621,7 +1670,14 @@ int xen_physdev_eoi_pirq(struct physdev_eoi *eoi)
         return -EINVAL;
     }
 
-    /* XX: Reassert a level IRQ if needed */
+    /* Reassert a level IRQ if needed */
+    if (s->pirq_gsi_set & (1U << gsi)) {
+        int port = s->pirq[pirq].port;
+        if (port) {
+            set_port_pending(s, port);
+        }
+    }
+
     return 0;
 }
 

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.

hw/i386/x86.c

@@ -61,6 +61,11 @@
 #include CONFIG_DEVICES
 #include "kvm/kvm_i386.h"
 
+#ifdef CONFIG_XEN_EMU
+#include "hw/xen/xen.h"
+#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;
 
@@ -610,6 +615,17 @@ 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_mode == XEN_EMULATE && xen_evtchn_set_gsi(n, level)) {
+            break;
+        }
+#endif
         qemu_set_irq(s->ioapic_irq[n], level);
         break;
     case IO_APIC_SECONDARY_IRQBASE