diff mbox series

[PULL,019/115] vmbus: vmbus implementation

Message ID 20200611194449.31468-20-pbonzini@redhat.com
State New
Headers show
Series [PULL,001/115] docker.py/build: support -t and -f arguments | expand

Commit Message

Paolo Bonzini June 11, 2020, 7:43 p.m. UTC
From: Jon Doron <arilou@gmail.com>

Add the VMBus infrastructure -- bus, devices, root bridge, vmbus state
machine, vmbus channel interactions, etc.

VMBus is a collection of technologies.  At its lowest layer, it's a message
passing and signaling mechanism, allowing efficient passing of messages to and
from guest VMs.  A layer higher, it's a mechanism for defining channels of
communication, where each channel is tagged with a type (which implies a
protocol) and a instance ID.  A layer higher than that, it's a bus driver,
serving as the basis of device enumeration within a VM, where a channel can
optionally be exposed as a paravirtual device.  When a server-side (paravirtual
back-end) component wishes to offer a channel to a guest VM, it does so by
specifying a channel type, a mode, and an instance ID.  VMBus then exposes this
in the guest.

More information about VMBus can be found in the file
vmbuskernelmodeclientlibapi.h in Microsoft's WDK.

TODO:
 - split into smaller palatable pieces
 - more comments
 - check and handle corner cases

Kudos to Evgeny Yakovlev (formerly eyakovlev@virtuozzo.com) and Andrey
Smetatin (formerly asmetanin@virtuozzo.com) for research and
prototyping.

Signed-off-by: Roman Kagan <rkagan@virtuozzo.com>
Signed-off-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Signed-off-by: Jon Doron <arilou@gmail.com>
Message-Id: <20200424123444.3481728-4-arilou@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 Makefile.objs                    |    1 +
 hw/hyperv/Kconfig                |    5 +
 hw/hyperv/Makefile.objs          |    1 +
 hw/hyperv/trace-events           |   18 +
 hw/hyperv/vmbus.c                | 2672 ++++++++++++++++++++++++++++++
 include/hw/hyperv/vmbus-bridge.h |   32 +
 include/hw/hyperv/vmbus.h        |  227 +++
 7 files changed, 2956 insertions(+)
 create mode 100644 hw/hyperv/trace-events
 create mode 100644 hw/hyperv/vmbus.c
 create mode 100644 include/hw/hyperv/vmbus-bridge.h
 create mode 100644 include/hw/hyperv/vmbus.h
diff mbox series

Patch

diff --git a/Makefile.objs b/Makefile.objs
index 99774cfd25..c09d95dfe3 100644
--- a/Makefile.objs
+++ b/Makefile.objs
@@ -150,6 +150,7 @@  trace-events-subdirs += hw/block/dataplane
 trace-events-subdirs += hw/char
 trace-events-subdirs += hw/dma
 trace-events-subdirs += hw/hppa
+trace-events-subdirs += hw/hyperv
 trace-events-subdirs += hw/i2c
 trace-events-subdirs += hw/i386
 trace-events-subdirs += hw/i386/xen
diff --git a/hw/hyperv/Kconfig b/hw/hyperv/Kconfig
index a1fa8ff9be..3fbfe41c9e 100644
--- a/hw/hyperv/Kconfig
+++ b/hw/hyperv/Kconfig
@@ -6,3 +6,8 @@  config HYPERV_TESTDEV
     bool
     default y if TEST_DEVICES
     depends on HYPERV
+
+config VMBUS
+    bool
+    default y
+    depends on HYPERV
diff --git a/hw/hyperv/Makefile.objs b/hw/hyperv/Makefile.objs
index edaca2f763..5b614e040c 100644
--- a/hw/hyperv/Makefile.objs
+++ b/hw/hyperv/Makefile.objs
@@ -1,2 +1,3 @@ 
 obj-y += hyperv.o
 obj-$(CONFIG_HYPERV_TESTDEV) += hyperv_testdev.o
+obj-$(CONFIG_VMBUS) += vmbus.o
diff --git a/hw/hyperv/trace-events b/hw/hyperv/trace-events
new file mode 100644
index 0000000000..ba5bd62d61
--- /dev/null
+++ b/hw/hyperv/trace-events
@@ -0,0 +1,18 @@ 
+# vmbus
+vmbus_recv_message(uint32_t type, uint32_t size) "type %d size %d"
+vmbus_signal_event(void) ""
+vmbus_channel_notify_guest(uint32_t chan_id) "channel #%d"
+vmbus_post_msg(uint32_t type, uint32_t size) "type %d size %d"
+vmbus_msg_cb(int status) "message status %d"
+vmbus_process_incoming_message(uint32_t message_type) "type %d"
+vmbus_initiate_contact(uint16_t major, uint16_t minor, uint32_t vcpu, uint64_t monitor_page1, uint64_t monitor_page2, uint64_t interrupt_page) "version %d.%d target vp %d mon pages 0x%"PRIx64",0x%"PRIx64" int page 0x%"PRIx64
+vmbus_send_offer(uint32_t chan_id, void *dev) "channel #%d dev %p"
+vmbus_terminate_offers(void) ""
+vmbus_gpadl_header(uint32_t gpadl_id, uint16_t num_gfns) "gpadl #%d gfns %d"
+vmbus_gpadl_body(uint32_t gpadl_id) "gpadl #%d"
+vmbus_gpadl_created(uint32_t gpadl_id) "gpadl #%d"
+vmbus_gpadl_teardown(uint32_t gpadl_id) "gpadl #%d"
+vmbus_gpadl_torndown(uint32_t gpadl_id) "gpadl #%d"
+vmbus_open_channel(uint32_t chan_id, uint32_t gpadl_id, uint32_t target_vp) "channel #%d gpadl #%d target vp %d"
+vmbus_channel_open(uint32_t chan_id, uint32_t status) "channel #%d status %d"
+vmbus_close_channel(uint32_t chan_id) "channel #%d"
diff --git a/hw/hyperv/vmbus.c b/hw/hyperv/vmbus.c
new file mode 100644
index 0000000000..802bbc8c96
--- /dev/null
+++ b/hw/hyperv/vmbus.c
@@ -0,0 +1,2672 @@ 
+/*
+ * QEMU Hyper-V VMBus
+ *
+ * Copyright (c) 2017-2018 Virtuozzo International GmbH.
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory.
+ */
+
+#include "qemu/osdep.h"
+#include "qemu/error-report.h"
+#include "qemu/main-loop.h"
+#include "qapi/error.h"
+#include "migration/vmstate.h"
+#include "hw/qdev-properties.h"
+#include "hw/hyperv/hyperv.h"
+#include "hw/hyperv/vmbus.h"
+#include "hw/hyperv/vmbus-bridge.h"
+#include "hw/sysbus.h"
+#include "cpu.h"
+#include "trace.h"
+
+#define TYPE_VMBUS "vmbus"
+#define VMBUS(obj) OBJECT_CHECK(VMBus, (obj), TYPE_VMBUS)
+
+enum {
+    VMGPADL_INIT,
+    VMGPADL_ALIVE,
+    VMGPADL_TEARINGDOWN,
+    VMGPADL_TORNDOWN,
+};
+
+struct VMBusGpadl {
+    /* GPADL id */
+    uint32_t id;
+    /* associated channel id (rudimentary?) */
+    uint32_t child_relid;
+
+    /* number of pages in the GPADL as declared in GPADL_HEADER message */
+    uint32_t num_gfns;
+    /*
+     * Due to limited message size, GPADL may not fit fully in a single
+     * GPADL_HEADER message, and is further popluated using GPADL_BODY
+     * messages.  @seen_gfns is the number of pages seen so far; once it
+     * reaches @num_gfns, the GPADL is ready to use.
+     */
+    uint32_t seen_gfns;
+    /* array of GFNs (of size @num_gfns once allocated) */
+    uint64_t *gfns;
+
+    uint8_t state;
+
+    QTAILQ_ENTRY(VMBusGpadl) link;
+    VMBus *vmbus;
+    unsigned refcount;
+};
+
+/*
+ * Wrap sequential read from / write to GPADL.
+ */
+typedef struct GpadlIter {
+    VMBusGpadl *gpadl;
+    AddressSpace *as;
+    DMADirection dir;
+    /* offset into GPADL where the next i/o will be performed */
+    uint32_t off;
+    /*
+     * Cached mapping of the currently accessed page, up to page boundary.
+     * Updated lazily on i/o.
+     * Note: MemoryRegionCache can not be used here because pages in the GPADL
+     * are non-contiguous and may belong to different memory regions.
+     */
+    void *map;
+    /* offset after last i/o (i.e. not affected by seek) */
+    uint32_t last_off;
+    /*
+     * Indicator that the iterator is active and may have a cached mapping.
+     * Allows to enforce bracketing of all i/o (which may create cached
+     * mappings) and thus exclude mapping leaks.
+     */
+    bool active;
+} GpadlIter;
+
+/*
+ * Ring buffer.  There are two of them, sitting in the same GPADL, for each
+ * channel.
+ * Each ring buffer consists of a set of pages, with the first page containing
+ * the ring buffer header, and the remaining pages being for data packets.
+ */
+typedef struct VMBusRingBufCommon {
+    AddressSpace *as;
+    /* GPA of the ring buffer header */
+    dma_addr_t rb_addr;
+    /* start and length of the ring buffer data area within GPADL */
+    uint32_t base;
+    uint32_t len;
+
+    GpadlIter iter;
+} VMBusRingBufCommon;
+
+typedef struct VMBusSendRingBuf {
+    VMBusRingBufCommon common;
+    /* current write index, to be committed at the end of send */
+    uint32_t wr_idx;
+    /* write index at the start of send */
+    uint32_t last_wr_idx;
+    /* space to be requested from the guest */
+    uint32_t wanted;
+    /* space reserved for planned sends */
+    uint32_t reserved;
+    /* last seen read index */
+    uint32_t last_seen_rd_idx;
+} VMBusSendRingBuf;
+
+typedef struct VMBusRecvRingBuf {
+    VMBusRingBufCommon common;
+    /* current read index, to be committed at the end of receive */
+    uint32_t rd_idx;
+    /* read index at the start of receive */
+    uint32_t last_rd_idx;
+    /* last seen write index */
+    uint32_t last_seen_wr_idx;
+} VMBusRecvRingBuf;
+
+
+enum {
+    VMOFFER_INIT,
+    VMOFFER_SENDING,
+    VMOFFER_SENT,
+};
+
+enum {
+    VMCHAN_INIT,
+    VMCHAN_OPENING,
+    VMCHAN_OPEN,
+};
+
+struct VMBusChannel {
+    VMBusDevice *dev;
+
+    /* channel id */
+    uint32_t id;
+    /*
+     * subchannel index within the device; subchannel #0 is "primary" and
+     * always exists
+     */
+    uint16_t subchan_idx;
+    uint32_t open_id;
+    /* VP_INDEX of the vCPU to notify with (synthetic) interrupts */
+    uint32_t target_vp;
+    /* GPADL id to use for the ring buffers */
+    uint32_t ringbuf_gpadl;
+    /* start (in pages) of the send ring buffer within @ringbuf_gpadl */
+    uint32_t ringbuf_send_offset;
+
+    uint8_t offer_state;
+    uint8_t state;
+    bool is_open;
+
+    /* main device worker; copied from the device class */
+    VMBusChannelNotifyCb notify_cb;
+    /*
+     * guest->host notifications, either sent directly or dispatched via
+     * interrupt page (older VMBus)
+     */
+    EventNotifier notifier;
+
+    VMBus *vmbus;
+    /*
+     * SINT route to signal with host->guest notifications; may be shared with
+     * the main VMBus SINT route
+     */
+    HvSintRoute *notify_route;
+    VMBusGpadl *gpadl;
+
+    VMBusSendRingBuf send_ringbuf;
+    VMBusRecvRingBuf recv_ringbuf;
+
+    QTAILQ_ENTRY(VMBusChannel) link;
+};
+
+/*
+ * Hyper-V spec mandates that every message port has 16 buffers, which means
+ * that the guest can post up to this many messages without blocking.
+ * Therefore a queue for incoming messages has to be provided.
+ * For outgoing (i.e. host->guest) messages there's no queue; the VMBus just
+ * doesn't transition to a new state until the message is known to have been
+ * successfully delivered to the respective SynIC message slot.
+ */
+#define HV_MSG_QUEUE_LEN     16
+
+/* Hyper-V devices never use channel #0.  Must be something special. */
+#define VMBUS_FIRST_CHANID      1
+/* Each channel occupies one bit within a single event page sint slot. */
+#define VMBUS_CHANID_COUNT      (HV_EVENT_FLAGS_COUNT - VMBUS_FIRST_CHANID)
+/* Leave a few connection numbers for other purposes. */
+#define VMBUS_CHAN_CONNECTION_OFFSET     16
+
+/*
+ * Since the success or failure of sending a message is reported
+ * asynchronously, the VMBus state machine has effectively two entry points:
+ * vmbus_run and vmbus_msg_cb (the latter is called when the host->guest
+ * message delivery status becomes known).  Both are run as oneshot BHs on the
+ * main aio context, ensuring serialization.
+ */
+enum {
+    VMBUS_LISTEN,
+    VMBUS_HANDSHAKE,
+    VMBUS_OFFER,
+    VMBUS_CREATE_GPADL,
+    VMBUS_TEARDOWN_GPADL,
+    VMBUS_OPEN_CHANNEL,
+    VMBUS_UNLOAD,
+    VMBUS_STATE_MAX
+};
+
+struct VMBus {
+    BusState parent;
+
+    uint8_t state;
+    /* protection against recursive aio_poll (see vmbus_run) */
+    bool in_progress;
+    /* whether there's a message being delivered to the guest */
+    bool msg_in_progress;
+    uint32_t version;
+    /* VP_INDEX of the vCPU to send messages and interrupts to */
+    uint32_t target_vp;
+    HvSintRoute *sint_route;
+    /*
+     * interrupt page for older protocol versions; newer ones use SynIC event
+     * flags directly
+     */
+    hwaddr int_page_gpa;
+
+    DECLARE_BITMAP(chanid_bitmap, VMBUS_CHANID_COUNT);
+
+    /* incoming message queue */
+    struct hyperv_post_message_input rx_queue[HV_MSG_QUEUE_LEN];
+    uint8_t rx_queue_head;
+    uint8_t rx_queue_size;
+    QemuMutex rx_queue_lock;
+
+    QTAILQ_HEAD(, VMBusGpadl) gpadl_list;
+    QTAILQ_HEAD(, VMBusChannel) channel_list;
+
+    /*
+     * guest->host notifications for older VMBus, to be dispatched via
+     * interrupt page
+     */
+    EventNotifier notifier;
+};
+
+static bool gpadl_full(VMBusGpadl *gpadl)
+{
+    return gpadl->seen_gfns == gpadl->num_gfns;
+}
+
+static VMBusGpadl *create_gpadl(VMBus *vmbus, uint32_t id,
+                                uint32_t child_relid, uint32_t num_gfns)
+{
+    VMBusGpadl *gpadl = g_new0(VMBusGpadl, 1);
+
+    gpadl->id = id;
+    gpadl->child_relid = child_relid;
+    gpadl->num_gfns = num_gfns;
+    gpadl->gfns = g_new(uint64_t, num_gfns);
+    QTAILQ_INSERT_HEAD(&vmbus->gpadl_list, gpadl, link);
+    gpadl->vmbus = vmbus;
+    gpadl->refcount = 1;
+    return gpadl;
+}
+
+static void free_gpadl(VMBusGpadl *gpadl)
+{
+    QTAILQ_REMOVE(&gpadl->vmbus->gpadl_list, gpadl, link);
+    g_free(gpadl->gfns);
+    g_free(gpadl);
+}
+
+static VMBusGpadl *find_gpadl(VMBus *vmbus, uint32_t gpadl_id)
+{
+    VMBusGpadl *gpadl;
+    QTAILQ_FOREACH(gpadl, &vmbus->gpadl_list, link) {
+        if (gpadl->id == gpadl_id) {
+            return gpadl;
+        }
+    }
+    return NULL;
+}
+
+VMBusGpadl *vmbus_get_gpadl(VMBusChannel *chan, uint32_t gpadl_id)
+{
+    VMBusGpadl *gpadl = find_gpadl(chan->vmbus, gpadl_id);
+    if (!gpadl || !gpadl_full(gpadl)) {
+        return NULL;
+    }
+    gpadl->refcount++;
+    return gpadl;
+}
+
+void vmbus_put_gpadl(VMBusGpadl *gpadl)
+{
+    if (!gpadl) {
+        return;
+    }
+    if (--gpadl->refcount) {
+        return;
+    }
+    free_gpadl(gpadl);
+}
+
+uint32_t vmbus_gpadl_len(VMBusGpadl *gpadl)
+{
+    return gpadl->num_gfns * TARGET_PAGE_SIZE;
+}
+
+static void gpadl_iter_init(GpadlIter *iter, VMBusGpadl *gpadl,
+                            AddressSpace *as, DMADirection dir)
+{
+    iter->gpadl = gpadl;
+    iter->as = as;
+    iter->dir = dir;
+    iter->active = false;
+}
+
+static inline void gpadl_iter_cache_unmap(GpadlIter *iter)
+{
+    uint32_t map_start_in_page = (uintptr_t)iter->map & ~TARGET_PAGE_MASK;
+    uint32_t io_end_in_page = ((iter->last_off - 1) & ~TARGET_PAGE_MASK) + 1;
+
+    /* mapping is only done to do non-zero amount of i/o */
+    assert(iter->last_off > 0);
+    assert(map_start_in_page < io_end_in_page);
+
+    dma_memory_unmap(iter->as, iter->map, TARGET_PAGE_SIZE - map_start_in_page,
+                     iter->dir, io_end_in_page - map_start_in_page);
+}
+
+/*
+ * Copy exactly @len bytes between the GPADL pointed to by @iter and @buf.
+ * The direction of the copy is determined by @iter->dir.
+ * The caller must ensure the operation overflows neither @buf nor the GPADL
+ * (there's an assert for the latter).
+ * Reuse the currently mapped page in the GPADL if possible.
+ */
+static ssize_t gpadl_iter_io(GpadlIter *iter, void *buf, uint32_t len)
+{
+    ssize_t ret = len;
+
+    assert(iter->active);
+
+    while (len) {
+        uint32_t off_in_page = iter->off & ~TARGET_PAGE_MASK;
+        uint32_t pgleft = TARGET_PAGE_SIZE - off_in_page;
+        uint32_t cplen = MIN(pgleft, len);
+        void *p;
+
+        /* try to reuse the cached mapping */
+        if (iter->map) {
+            uint32_t map_start_in_page =
+                (uintptr_t)iter->map & ~TARGET_PAGE_MASK;
+            uint32_t off_base = iter->off & ~TARGET_PAGE_MASK;
+            uint32_t mapped_base = (iter->last_off - 1) & ~TARGET_PAGE_MASK;
+            if (off_base != mapped_base || off_in_page < map_start_in_page) {
+                gpadl_iter_cache_unmap(iter);
+                iter->map = NULL;
+            }
+        }
+
+        if (!iter->map) {
+            dma_addr_t maddr;
+            dma_addr_t mlen = pgleft;
+            uint32_t idx = iter->off >> TARGET_PAGE_BITS;
+            assert(idx < iter->gpadl->num_gfns);
+
+            maddr = (iter->gpadl->gfns[idx] << TARGET_PAGE_BITS) | off_in_page;
+
+            iter->map = dma_memory_map(iter->as, maddr, &mlen, iter->dir);
+            if (mlen != pgleft) {
+                dma_memory_unmap(iter->as, iter->map, mlen, iter->dir, 0);
+                iter->map = NULL;
+                return -EFAULT;
+            }
+        }
+
+        p = (void *)(((uintptr_t)iter->map & TARGET_PAGE_MASK) | off_in_page);
+        if (iter->dir == DMA_DIRECTION_FROM_DEVICE) {
+            memcpy(p, buf, cplen);
+        } else {
+            memcpy(buf, p, cplen);
+        }
+
+        buf += cplen;
+        len -= cplen;
+        iter->off += cplen;
+        iter->last_off = iter->off;
+    }
+
+    return ret;
+}
+
+/*
+ * Position the iterator @iter at new offset @new_off.
+ * If this results in the cached mapping being unusable with the new offset,
+ * unmap it.
+ */
+static inline void gpadl_iter_seek(GpadlIter *iter, uint32_t new_off)
+{
+    assert(iter->active);
+    iter->off = new_off;
+}
+
+/*
+ * Start a series of i/o on the GPADL.
+ * After this i/o and seek operations on @iter become legal.
+ */
+static inline void gpadl_iter_start_io(GpadlIter *iter)
+{
+    assert(!iter->active);
+    /* mapping is cached lazily on i/o */
+    iter->map = NULL;
+    iter->active = true;
+}
+
+/*
+ * End the eariler started series of i/o on the GPADL and release the cached
+ * mapping if any.
+ */
+static inline void gpadl_iter_end_io(GpadlIter *iter)
+{
+    assert(iter->active);
+
+    if (iter->map) {
+        gpadl_iter_cache_unmap(iter);
+    }
+
+    iter->active = false;
+}
+
+static void vmbus_resched(VMBus *vmbus);
+static void vmbus_msg_cb(void *data, int status);
+
+ssize_t vmbus_iov_to_gpadl(VMBusChannel *chan, VMBusGpadl *gpadl, uint32_t off,
+                           const struct iovec *iov, size_t iov_cnt)
+{
+    GpadlIter iter;
+    size_t i;
+    ssize_t ret = 0;
+
+    gpadl_iter_init(&iter, gpadl, chan->dev->dma_as,
+                    DMA_DIRECTION_FROM_DEVICE);
+    gpadl_iter_start_io(&iter);
+    gpadl_iter_seek(&iter, off);
+    for (i = 0; i < iov_cnt; i++) {
+        ret = gpadl_iter_io(&iter, iov[i].iov_base, iov[i].iov_len);
+        if (ret < 0) {
+            goto out;
+        }
+    }
+out:
+    gpadl_iter_end_io(&iter);
+    return ret;
+}
+
+int vmbus_map_sgl(VMBusChanReq *req, DMADirection dir, struct iovec *iov,
+                  unsigned iov_cnt, size_t len, size_t off)
+{
+    int ret_cnt = 0, ret;
+    unsigned i;
+    QEMUSGList *sgl = &req->sgl;
+    ScatterGatherEntry *sg = sgl->sg;
+
+    for (i = 0; i < sgl->nsg; i++) {
+        if (sg[i].len > off) {
+            break;
+        }
+        off -= sg[i].len;
+    }
+    for (; len && i < sgl->nsg; i++) {
+        dma_addr_t mlen = MIN(sg[i].len - off, len);
+        dma_addr_t addr = sg[i].base + off;
+        len -= mlen;
+        off = 0;
+
+        for (; mlen; ret_cnt++) {
+            dma_addr_t l = mlen;
+            dma_addr_t a = addr;
+
+            if (ret_cnt == iov_cnt) {
+                ret = -ENOBUFS;
+                goto err;
+            }
+
+            iov[ret_cnt].iov_base = dma_memory_map(sgl->as, a, &l, dir);
+            if (!l) {
+                ret = -EFAULT;
+                goto err;
+            }
+            iov[ret_cnt].iov_len = l;
+            addr += l;
+            mlen -= l;
+        }
+    }
+
+    return ret_cnt;
+err:
+    vmbus_unmap_sgl(req, dir, iov, ret_cnt, 0);
+    return ret;
+}
+
+void vmbus_unmap_sgl(VMBusChanReq *req, DMADirection dir, struct iovec *iov,
+                     unsigned iov_cnt, size_t accessed)
+{
+    QEMUSGList *sgl = &req->sgl;
+    unsigned i;
+
+    for (i = 0; i < iov_cnt; i++) {
+        size_t acsd = MIN(accessed, iov[i].iov_len);
+        dma_memory_unmap(sgl->as, iov[i].iov_base, iov[i].iov_len, dir, acsd);
+        accessed -= acsd;
+    }
+}
+
+static const VMStateDescription vmstate_gpadl = {
+    .name = "vmbus/gpadl",
+    .version_id = 0,
+    .minimum_version_id = 0,
+    .fields = (VMStateField[]) {
+        VMSTATE_UINT32(id, VMBusGpadl),
+        VMSTATE_UINT32(child_relid, VMBusGpadl),
+        VMSTATE_UINT32(num_gfns, VMBusGpadl),
+        VMSTATE_UINT32(seen_gfns, VMBusGpadl),
+        VMSTATE_VARRAY_UINT32_ALLOC(gfns, VMBusGpadl, num_gfns, 0,
+                                    vmstate_info_uint64, uint64_t),
+        VMSTATE_UINT8(state, VMBusGpadl),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
+/*
+ * Wrap the index into a ring buffer of @len bytes.
+ * @idx is assumed not to exceed twice the size of the ringbuffer, so only
+ * single wraparound is considered.
+ */
+static inline uint32_t rb_idx_wrap(uint32_t idx, uint32_t len)
+{
+    if (idx >= len) {
+        idx -= len;
+    }
+    return idx;
+}
+
+/*
+ * Circular difference between two indices into a ring buffer of @len bytes.
+ * @allow_catchup - whether @idx1 may catch up @idx2; e.g. read index may catch
+ * up write index but not vice versa.
+ */
+static inline uint32_t rb_idx_delta(uint32_t idx1, uint32_t idx2, uint32_t len,
+                                    bool allow_catchup)
+{
+    return rb_idx_wrap(idx2 + len - idx1 - !allow_catchup, len);
+}
+
+static vmbus_ring_buffer *ringbuf_map_hdr(VMBusRingBufCommon *ringbuf)
+{
+    vmbus_ring_buffer *rb;
+    dma_addr_t mlen = sizeof(*rb);
+
+    rb = dma_memory_map(ringbuf->as, ringbuf->rb_addr, &mlen,
+                        DMA_DIRECTION_FROM_DEVICE);
+    if (mlen != sizeof(*rb)) {
+        dma_memory_unmap(ringbuf->as, rb, mlen,
+                         DMA_DIRECTION_FROM_DEVICE, 0);
+        return NULL;
+    }
+    return rb;
+}
+
+static void ringbuf_unmap_hdr(VMBusRingBufCommon *ringbuf,
+                              vmbus_ring_buffer *rb, bool dirty)
+{
+    assert(rb);
+
+    dma_memory_unmap(ringbuf->as, rb, sizeof(*rb), DMA_DIRECTION_FROM_DEVICE,
+                     dirty ? sizeof(*rb) : 0);
+}
+
+static void ringbuf_init_common(VMBusRingBufCommon *ringbuf, VMBusGpadl *gpadl,
+                                AddressSpace *as, DMADirection dir,
+                                uint32_t begin, uint32_t end)
+{
+    ringbuf->as = as;
+    ringbuf->rb_addr = gpadl->gfns[begin] << TARGET_PAGE_BITS;
+    ringbuf->base = (begin + 1) << TARGET_PAGE_BITS;
+    ringbuf->len = (end - begin - 1) << TARGET_PAGE_BITS;
+    gpadl_iter_init(&ringbuf->iter, gpadl, as, dir);
+}
+
+static int ringbufs_init(VMBusChannel *chan)
+{
+    vmbus_ring_buffer *rb;
+    VMBusSendRingBuf *send_ringbuf = &chan->send_ringbuf;
+    VMBusRecvRingBuf *recv_ringbuf = &chan->recv_ringbuf;
+
+    if (chan->ringbuf_send_offset <= 1 ||
+        chan->gpadl->num_gfns <= chan->ringbuf_send_offset + 1) {
+        return -EINVAL;
+    }
+
+    ringbuf_init_common(&recv_ringbuf->common, chan->gpadl, chan->dev->dma_as,
+                        DMA_DIRECTION_TO_DEVICE, 0, chan->ringbuf_send_offset);
+    ringbuf_init_common(&send_ringbuf->common, chan->gpadl, chan->dev->dma_as,
+                        DMA_DIRECTION_FROM_DEVICE, chan->ringbuf_send_offset,
+                        chan->gpadl->num_gfns);
+    send_ringbuf->wanted = 0;
+    send_ringbuf->reserved = 0;
+
+    rb = ringbuf_map_hdr(&recv_ringbuf->common);
+    if (!rb) {
+        return -EFAULT;
+    }
+    recv_ringbuf->rd_idx = recv_ringbuf->last_rd_idx = rb->read_index;
+    ringbuf_unmap_hdr(&recv_ringbuf->common, rb, false);
+
+    rb = ringbuf_map_hdr(&send_ringbuf->common);
+    if (!rb) {
+        return -EFAULT;
+    }
+    send_ringbuf->wr_idx = send_ringbuf->last_wr_idx = rb->write_index;
+    send_ringbuf->last_seen_rd_idx = rb->read_index;
+    rb->feature_bits |= VMBUS_RING_BUFFER_FEAT_PENDING_SZ;
+    ringbuf_unmap_hdr(&send_ringbuf->common, rb, true);
+
+    if (recv_ringbuf->rd_idx >= recv_ringbuf->common.len ||
+        send_ringbuf->wr_idx >= send_ringbuf->common.len) {
+        return -EOVERFLOW;
+    }
+
+    return 0;
+}
+
+/*
+ * Perform io between the GPADL-backed ringbuffer @ringbuf and @buf, wrapping
+ * around if needed.
+ * @len is assumed not to exceed the size of the ringbuffer, so only single
+ * wraparound is considered.
+ */
+static ssize_t ringbuf_io(VMBusRingBufCommon *ringbuf, void *buf, uint32_t len)
+{
+    ssize_t ret1 = 0, ret2 = 0;
+    uint32_t remain = ringbuf->len + ringbuf->base - ringbuf->iter.off;
+
+    if (len >= remain) {
+        ret1 = gpadl_iter_io(&ringbuf->iter, buf, remain);
+        if (ret1 < 0) {
+            return ret1;
+        }
+        gpadl_iter_seek(&ringbuf->iter, ringbuf->base);
+        buf += remain;
+        len -= remain;
+    }
+    ret2 = gpadl_iter_io(&ringbuf->iter, buf, len);
+    if (ret2 < 0) {
+        return ret2;
+    }
+    return ret1 + ret2;
+}
+
+/*
+ * Position the circular iterator within @ringbuf to offset @new_off, wrapping
+ * around if needed.
+ * @new_off is assumed not to exceed twice the size of the ringbuffer, so only
+ * single wraparound is considered.
+ */
+static inline void ringbuf_seek(VMBusRingBufCommon *ringbuf, uint32_t new_off)
+{
+    gpadl_iter_seek(&ringbuf->iter,
+                    ringbuf->base + rb_idx_wrap(new_off, ringbuf->len));
+}
+
+static inline uint32_t ringbuf_tell(VMBusRingBufCommon *ringbuf)
+{
+    return ringbuf->iter.off - ringbuf->base;
+}
+
+static inline void ringbuf_start_io(VMBusRingBufCommon *ringbuf)
+{
+    gpadl_iter_start_io(&ringbuf->iter);
+}
+
+static inline void ringbuf_end_io(VMBusRingBufCommon *ringbuf)
+{
+    gpadl_iter_end_io(&ringbuf->iter);
+}
+
+VMBusDevice *vmbus_channel_device(VMBusChannel *chan)
+{
+    return chan->dev;
+}
+
+VMBusChannel *vmbus_device_channel(VMBusDevice *dev, uint32_t chan_idx)
+{
+    if (chan_idx >= dev->num_channels) {
+        return NULL;
+    }
+    return &dev->channels[chan_idx];
+}
+
+uint32_t vmbus_channel_idx(VMBusChannel *chan)
+{
+    return chan - chan->dev->channels;
+}
+
+void vmbus_channel_notify_host(VMBusChannel *chan)
+{
+    event_notifier_set(&chan->notifier);
+}
+
+bool vmbus_channel_is_open(VMBusChannel *chan)
+{
+    return chan->is_open;
+}
+
+/*
+ * Notify the guest side about the data to work on in the channel ring buffer.
+ * The notification is done by signaling a dedicated per-channel SynIC event
+ * flag (more recent guests) or setting a bit in the interrupt page and firing
+ * the VMBus SINT (older guests).
+ */
+static int vmbus_channel_notify_guest(VMBusChannel *chan)
+{
+    int res = 0;
+    unsigned long *int_map, mask;
+    unsigned idx;
+    hwaddr addr = chan->vmbus->int_page_gpa;
+    hwaddr len = TARGET_PAGE_SIZE / 2, dirty = 0;
+
+    trace_vmbus_channel_notify_guest(chan->id);
+
+    if (!addr) {
+        return hyperv_set_event_flag(chan->notify_route, chan->id);
+    }
+
+    int_map = cpu_physical_memory_map(addr, &len, 1);
+    if (len != TARGET_PAGE_SIZE / 2) {
+        res = -ENXIO;
+        goto unmap;
+    }
+
+    idx = BIT_WORD(chan->id);
+    mask = BIT_MASK(chan->id);
+    if ((atomic_fetch_or(&int_map[idx], mask) & mask) != mask) {
+        res = hyperv_sint_route_set_sint(chan->notify_route);
+        dirty = len;
+    }
+
+unmap:
+    cpu_physical_memory_unmap(int_map, len, 1, dirty);
+    return res;
+}
+
+#define VMBUS_PKT_TRAILER      sizeof(uint64_t)
+
+static uint32_t vmbus_pkt_hdr_set_offsets(vmbus_packet_hdr *hdr,
+                                          uint32_t desclen, uint32_t msglen)
+{
+    hdr->offset_qwords = sizeof(*hdr) / sizeof(uint64_t) +
+        DIV_ROUND_UP(desclen, sizeof(uint64_t));
+    hdr->len_qwords = hdr->offset_qwords +
+        DIV_ROUND_UP(msglen, sizeof(uint64_t));
+    return hdr->len_qwords * sizeof(uint64_t) + VMBUS_PKT_TRAILER;
+}
+
+/*
+ * Simplified ring buffer operation with paired barriers annotations in the
+ * producer and consumer loops:
+ *
+ * producer                           * consumer
+ * ~~~~~~~~                           * ~~~~~~~~
+ * write pending_send_sz              * read write_index
+ * smp_mb                       [A]   * smp_mb                       [C]
+ * read read_index                    * read packet
+ * smp_mb                       [B]   * read/write out-of-band data
+ * read/write out-of-band data        * smp_mb                       [B]
+ * write packet                       * write read_index
+ * smp_mb                       [C]   * smp_mb                       [A]
+ * write write_index                  * read pending_send_sz
+ * smp_wmb                      [D]   * smp_rmb                      [D]
+ * write pending_send_sz              * read write_index
+ * ...                                * ...
+ */
+
+static inline uint32_t ringbuf_send_avail(VMBusSendRingBuf *ringbuf)
+{
+    /* don't trust guest data */
+    if (ringbuf->last_seen_rd_idx >= ringbuf->common.len) {
+        return 0;
+    }
+    return rb_idx_delta(ringbuf->wr_idx, ringbuf->last_seen_rd_idx,
+                        ringbuf->common.len, false);
+}
+
+static ssize_t ringbuf_send_update_idx(VMBusChannel *chan)
+{
+    VMBusSendRingBuf *ringbuf = &chan->send_ringbuf;
+    vmbus_ring_buffer *rb;
+    uint32_t written;
+
+    written = rb_idx_delta(ringbuf->last_wr_idx, ringbuf->wr_idx,
+                           ringbuf->common.len, true);
+    if (!written) {
+        return 0;
+    }
+
+    rb = ringbuf_map_hdr(&ringbuf->common);
+    if (!rb) {
+        return -EFAULT;
+    }
+
+    ringbuf->reserved -= written;
+
+    /* prevent reorder with the data operation and packet write */
+    smp_mb();                   /* barrier pair [C] */
+    rb->write_index = ringbuf->wr_idx;
+
+    /*
+     * If the producer earlier indicated that it wants to be notified when the
+     * consumer frees certain amount of space in the ring buffer, that amount
+     * is reduced by the size of the completed write.
+     */
+    if (ringbuf->wanted) {
+        /* otherwise reservation would fail */
+        assert(ringbuf->wanted < written);
+        ringbuf->wanted -= written;
+        /* prevent reorder with write_index write */
+        smp_wmb();              /* barrier pair [D] */
+        rb->pending_send_sz = ringbuf->wanted;
+    }
+
+    /* prevent reorder with write_index or pending_send_sz write */
+    smp_mb();                   /* barrier pair [A] */
+    ringbuf->last_seen_rd_idx = rb->read_index;
+
+    /*
+     * The consumer may have missed the reduction of pending_send_sz and skip
+     * notification, so re-check the blocking condition, and, if it's no longer
+     * true, ensure processing another iteration by simulating consumer's
+     * notification.
+     */
+    if (ringbuf_send_avail(ringbuf) >= ringbuf->wanted) {
+        vmbus_channel_notify_host(chan);
+    }
+
+    /* skip notification by consumer's request */
+    if (rb->interrupt_mask) {
+        goto out;
+    }
+
+    /*
+     * The consumer hasn't caught up with the producer's previous state so it's
+     * not blocked.
+     * (last_seen_rd_idx comes from the guest but it's safe to use w/o
+     * validation here as it only affects notification.)
+     */
+    if (rb_idx_delta(ringbuf->last_seen_rd_idx, ringbuf->wr_idx,
+                     ringbuf->common.len, true) > written) {
+        goto out;
+    }
+
+    vmbus_channel_notify_guest(chan);
+out:
+    ringbuf_unmap_hdr(&ringbuf->common, rb, true);
+    ringbuf->last_wr_idx = ringbuf->wr_idx;
+    return written;
+}
+
+int vmbus_channel_reserve(VMBusChannel *chan,
+                          uint32_t desclen, uint32_t msglen)
+{
+    VMBusSendRingBuf *ringbuf = &chan->send_ringbuf;
+    vmbus_ring_buffer *rb = NULL;
+    vmbus_packet_hdr hdr;
+    uint32_t needed = ringbuf->reserved +
+        vmbus_pkt_hdr_set_offsets(&hdr, desclen, msglen);
+
+    /* avoid touching the guest memory if possible */
+    if (likely(needed <= ringbuf_send_avail(ringbuf))) {
+        goto success;
+    }
+
+    rb = ringbuf_map_hdr(&ringbuf->common);
+    if (!rb) {
+        return -EFAULT;
+    }
+
+    /* fetch read index from guest memory and try again */
+    ringbuf->last_seen_rd_idx = rb->read_index;
+
+    if (likely(needed <= ringbuf_send_avail(ringbuf))) {
+        goto success;
+    }
+
+    rb->pending_send_sz = needed;
+
+    /*
+     * The consumer may have made progress and freed up some space before
+     * seeing updated pending_send_sz, so re-read read_index (preventing
+     * reorder with the pending_send_sz write) and try again.
+     */
+    smp_mb();                   /* barrier pair [A] */
+    ringbuf->last_seen_rd_idx = rb->read_index;
+
+    if (needed > ringbuf_send_avail(ringbuf)) {
+        goto out;
+    }
+
+success:
+    ringbuf->reserved = needed;
+    needed = 0;
+
+    /* clear pending_send_sz if it was set */
+    if (ringbuf->wanted) {
+        if (!rb) {
+            rb = ringbuf_map_hdr(&ringbuf->common);
+            if (!rb) {
+                /* failure to clear pending_send_sz is non-fatal */
+                goto out;
+            }
+        }
+
+        rb->pending_send_sz = 0;
+    }
+
+    /* prevent reorder of the following data operation with read_index read */
+    smp_mb();                   /* barrier pair [B] */
+
+out:
+    if (rb) {
+        ringbuf_unmap_hdr(&ringbuf->common, rb, ringbuf->wanted == needed);
+    }
+    ringbuf->wanted = needed;
+    return needed ? -ENOSPC : 0;
+}
+
+ssize_t vmbus_channel_send(VMBusChannel *chan, uint16_t pkt_type,
+                           void *desc, uint32_t desclen,
+                           void *msg, uint32_t msglen,
+                           bool need_comp, uint64_t transaction_id)
+{
+    ssize_t ret = 0;
+    vmbus_packet_hdr hdr;
+    uint32_t totlen;
+    VMBusSendRingBuf *ringbuf = &chan->send_ringbuf;
+
+    if (!vmbus_channel_is_open(chan)) {
+        return -EINVAL;
+    }
+
+    totlen = vmbus_pkt_hdr_set_offsets(&hdr, desclen, msglen);
+    hdr.type = pkt_type;
+    hdr.flags = need_comp ? VMBUS_PACKET_FLAG_REQUEST_COMPLETION : 0;
+    hdr.transaction_id = transaction_id;
+
+    assert(totlen <= ringbuf->reserved);
+
+    ringbuf_start_io(&ringbuf->common);
+    ringbuf_seek(&ringbuf->common, ringbuf->wr_idx);
+    ret = ringbuf_io(&ringbuf->common, &hdr, sizeof(hdr));
+    if (ret < 0) {
+        goto out;
+    }
+    if (desclen) {
+        assert(desc);
+        ret = ringbuf_io(&ringbuf->common, desc, desclen);
+        if (ret < 0) {
+            goto out;
+        }
+        ringbuf_seek(&ringbuf->common,
+                     ringbuf->wr_idx + hdr.offset_qwords * sizeof(uint64_t));
+    }
+    ret = ringbuf_io(&ringbuf->common, msg, msglen);
+    if (ret < 0) {
+        goto out;
+    }
+    ringbuf_seek(&ringbuf->common, ringbuf->wr_idx + totlen);
+    ringbuf->wr_idx = ringbuf_tell(&ringbuf->common);
+    ret = 0;
+out:
+    ringbuf_end_io(&ringbuf->common);
+    if (ret) {
+        return ret;
+    }
+    return ringbuf_send_update_idx(chan);
+}
+
+ssize_t vmbus_channel_send_completion(VMBusChanReq *req,
+                                      void *msg, uint32_t msglen)
+{
+    assert(req->need_comp);
+    return vmbus_channel_send(req->chan, VMBUS_PACKET_COMP, NULL, 0,
+                              msg, msglen, false, req->transaction_id);
+}
+
+static int sgl_from_gpa_ranges(QEMUSGList *sgl, VMBusDevice *dev,
+                               VMBusRingBufCommon *ringbuf, uint32_t len)
+{
+    int ret;
+    vmbus_pkt_gpa_direct hdr;
+    hwaddr curaddr = 0;
+    hwaddr curlen = 0;
+    int num;
+
+    if (len < sizeof(hdr)) {
+        return -EIO;
+    }
+    ret = ringbuf_io(ringbuf, &hdr, sizeof(hdr));
+    if (ret < 0) {
+        return ret;
+    }
+    len -= sizeof(hdr);
+
+    num = (len - hdr.rangecount * sizeof(vmbus_gpa_range)) / sizeof(uint64_t);
+    if (num < 0) {
+        return -EIO;
+    }
+    qemu_sglist_init(sgl, DEVICE(dev), num, ringbuf->as);
+
+    for (; hdr.rangecount; hdr.rangecount--) {
+        vmbus_gpa_range range;
+
+        if (len < sizeof(range)) {
+            goto eio;
+        }
+        ret = ringbuf_io(ringbuf, &range, sizeof(range));
+        if (ret < 0) {
+            goto err;
+        }
+        len -= sizeof(range);
+
+        if (range.byte_offset & TARGET_PAGE_MASK) {
+            goto eio;
+        }
+
+        for (; range.byte_count; range.byte_offset = 0) {
+            uint64_t paddr;
+            uint32_t plen = MIN(range.byte_count,
+                                TARGET_PAGE_SIZE - range.byte_offset);
+
+            if (len < sizeof(uint64_t)) {
+                goto eio;
+            }
+            ret = ringbuf_io(ringbuf, &paddr, sizeof(paddr));
+            if (ret < 0) {
+                goto err;
+            }
+            len -= sizeof(uint64_t);
+            paddr <<= TARGET_PAGE_BITS;
+            paddr |= range.byte_offset;
+            range.byte_count -= plen;
+
+            if (curaddr + curlen == paddr) {
+                /* consecutive fragments - join */
+                curlen += plen;
+            } else {
+                if (curlen) {
+                    qemu_sglist_add(sgl, curaddr, curlen);
+                }
+
+                curaddr = paddr;
+                curlen = plen;
+            }
+        }
+    }
+
+    if (curlen) {
+        qemu_sglist_add(sgl, curaddr, curlen);
+    }
+
+    return 0;
+eio:
+    ret = -EIO;
+err:
+    qemu_sglist_destroy(sgl);
+    return ret;
+}
+
+static VMBusChanReq *vmbus_alloc_req(VMBusChannel *chan,
+                                     uint32_t size, uint16_t pkt_type,
+                                     uint32_t msglen, uint64_t transaction_id,
+                                     bool need_comp)
+{
+    VMBusChanReq *req;
+    uint32_t msgoff = QEMU_ALIGN_UP(size, __alignof__(*req->msg));
+    uint32_t totlen = msgoff + msglen;
+
+    req = g_malloc0(totlen);
+    req->chan = chan;
+    req->pkt_type = pkt_type;
+    req->msg = (void *)req + msgoff;
+    req->msglen = msglen;
+    req->transaction_id = transaction_id;
+    req->need_comp = need_comp;
+    return req;
+}
+
+int vmbus_channel_recv_start(VMBusChannel *chan)
+{
+    VMBusRecvRingBuf *ringbuf = &chan->recv_ringbuf;
+    vmbus_ring_buffer *rb;
+
+    rb = ringbuf_map_hdr(&ringbuf->common);
+    if (!rb) {
+        return -EFAULT;
+    }
+    ringbuf->last_seen_wr_idx = rb->write_index;
+    ringbuf_unmap_hdr(&ringbuf->common, rb, false);
+
+    if (ringbuf->last_seen_wr_idx >= ringbuf->common.len) {
+        return -EOVERFLOW;
+    }
+
+    /* prevent reorder of the following data operation with write_index read */
+    smp_mb();                   /* barrier pair [C] */
+    return 0;
+}
+
+void *vmbus_channel_recv_peek(VMBusChannel *chan, uint32_t size)
+{
+    VMBusRecvRingBuf *ringbuf = &chan->recv_ringbuf;
+    vmbus_packet_hdr hdr = {};
+    VMBusChanReq *req;
+    uint32_t avail;
+    uint32_t totlen, pktlen, msglen, msgoff, desclen;
+
+    assert(size >= sizeof(*req));
+
+    /* safe as last_seen_wr_idx is validated in vmbus_channel_recv_start */
+    avail = rb_idx_delta(ringbuf->rd_idx, ringbuf->last_seen_wr_idx,
+                         ringbuf->common.len, true);
+    if (avail < sizeof(hdr)) {
+        return NULL;
+    }
+
+    ringbuf_seek(&ringbuf->common, ringbuf->rd_idx);
+    if (ringbuf_io(&ringbuf->common, &hdr, sizeof(hdr)) < 0) {
+        return NULL;
+    }
+
+    pktlen = hdr.len_qwords * sizeof(uint64_t);
+    totlen = pktlen + VMBUS_PKT_TRAILER;
+    if (totlen > avail) {
+        return NULL;
+    }
+
+    msgoff = hdr.offset_qwords * sizeof(uint64_t);
+    if (msgoff > pktlen || msgoff < sizeof(hdr)) {
+        error_report("%s: malformed packet: %u %u", __func__, msgoff, pktlen);
+        return NULL;
+    }
+
+    msglen = pktlen - msgoff;
+
+    req = vmbus_alloc_req(chan, size, hdr.type, msglen, hdr.transaction_id,
+                          hdr.flags & VMBUS_PACKET_FLAG_REQUEST_COMPLETION);
+
+    switch (hdr.type) {
+    case VMBUS_PACKET_DATA_USING_GPA_DIRECT:
+        desclen = msgoff - sizeof(hdr);
+        if (sgl_from_gpa_ranges(&req->sgl, chan->dev, &ringbuf->common,
+                                desclen) < 0) {
+            error_report("%s: failed to convert GPA ranges to SGL", __func__);
+            goto free_req;
+        }
+        break;
+    case VMBUS_PACKET_DATA_INBAND:
+    case VMBUS_PACKET_COMP:
+        break;
+    default:
+        error_report("%s: unexpected msg type: %x", __func__, hdr.type);
+        goto free_req;
+    }
+
+    ringbuf_seek(&ringbuf->common, ringbuf->rd_idx + msgoff);
+    if (ringbuf_io(&ringbuf->common, req->msg, msglen) < 0) {
+        goto free_req;
+    }
+    ringbuf_seek(&ringbuf->common, ringbuf->rd_idx + totlen);
+
+    return req;
+free_req:
+    vmbus_free_req(req);
+    return NULL;
+}
+
+void vmbus_channel_recv_pop(VMBusChannel *chan)
+{
+    VMBusRecvRingBuf *ringbuf = &chan->recv_ringbuf;
+    ringbuf->rd_idx = ringbuf_tell(&ringbuf->common);
+}
+
+ssize_t vmbus_channel_recv_done(VMBusChannel *chan)
+{
+    VMBusRecvRingBuf *ringbuf = &chan->recv_ringbuf;
+    vmbus_ring_buffer *rb;
+    uint32_t read;
+
+    read = rb_idx_delta(ringbuf->last_rd_idx, ringbuf->rd_idx,
+                        ringbuf->common.len, true);
+    if (!read) {
+        return 0;
+    }
+
+    rb = ringbuf_map_hdr(&ringbuf->common);
+    if (!rb) {
+        return -EFAULT;
+    }
+
+    /* prevent reorder with the data operation and packet read */
+    smp_mb();                   /* barrier pair [B] */
+    rb->read_index = ringbuf->rd_idx;
+
+    /* prevent reorder of the following pending_send_sz read */
+    smp_mb();                   /* barrier pair [A] */
+
+    if (rb->interrupt_mask) {
+        goto out;
+    }
+
+    if (rb->feature_bits & VMBUS_RING_BUFFER_FEAT_PENDING_SZ) {
+        uint32_t wr_idx, wr_avail;
+        uint32_t wanted = rb->pending_send_sz;
+
+        if (!wanted) {
+            goto out;
+        }
+
+        /* prevent reorder with pending_send_sz read */
+        smp_rmb();              /* barrier pair [D] */
+        wr_idx = rb->write_index;
+
+        wr_avail = rb_idx_delta(wr_idx, ringbuf->rd_idx, ringbuf->common.len,
+                                true);
+
+        /* the producer wasn't blocked on the consumer state */
+        if (wr_avail >= read + wanted) {
+            goto out;
+        }
+        /* there's not enough space for the producer to make progress */
+        if (wr_avail < wanted) {
+            goto out;
+        }
+    }
+
+    vmbus_channel_notify_guest(chan);
+out:
+    ringbuf_unmap_hdr(&ringbuf->common, rb, true);
+    ringbuf->last_rd_idx = ringbuf->rd_idx;
+    return read;
+}
+
+void vmbus_free_req(void *req)
+{
+    VMBusChanReq *r = req;
+
+    if (!req) {
+        return;
+    }
+
+    if (r->sgl.dev) {
+        qemu_sglist_destroy(&r->sgl);
+    }
+    g_free(req);
+}
+
+static void channel_event_cb(EventNotifier *e)
+{
+    VMBusChannel *chan = container_of(e, VMBusChannel, notifier);
+    if (event_notifier_test_and_clear(e)) {
+        /*
+         * All receives are supposed to happen within the device worker, so
+         * bracket it with ringbuf_start/end_io on the receive ringbuffer, and
+         * potentially reuse the cached mapping throughout the worker.
+         * Can't do this for sends as they may happen outside the device
+         * worker.
+         */
+        VMBusRecvRingBuf *ringbuf = &chan->recv_ringbuf;
+        ringbuf_start_io(&ringbuf->common);
+        chan->notify_cb(chan);
+        ringbuf_end_io(&ringbuf->common);
+
+    }
+}
+
+static int alloc_chan_id(VMBus *vmbus)
+{
+    int ret;
+
+    ret = find_next_zero_bit(vmbus->chanid_bitmap, VMBUS_CHANID_COUNT, 0);
+    if (ret == VMBUS_CHANID_COUNT) {
+        return -ENOMEM;
+    }
+    return ret + VMBUS_FIRST_CHANID;
+}
+
+static int register_chan_id(VMBusChannel *chan)
+{
+    return test_and_set_bit(chan->id - VMBUS_FIRST_CHANID,
+                            chan->vmbus->chanid_bitmap) ? -EEXIST : 0;
+}
+
+static void unregister_chan_id(VMBusChannel *chan)
+{
+    clear_bit(chan->id - VMBUS_FIRST_CHANID, chan->vmbus->chanid_bitmap);
+}
+
+static uint32_t chan_connection_id(VMBusChannel *chan)
+{
+    return VMBUS_CHAN_CONNECTION_OFFSET + chan->id;
+}
+
+static void init_channel(VMBus *vmbus, VMBusDevice *dev, VMBusDeviceClass *vdc,
+                         VMBusChannel *chan, uint16_t idx, Error **errp)
+{
+    int res;
+
+    chan->dev = dev;
+    chan->notify_cb = vdc->chan_notify_cb;
+    chan->subchan_idx = idx;
+    chan->vmbus = vmbus;
+
+    res = alloc_chan_id(vmbus);
+    if (res < 0) {
+        error_setg(errp, "no spare channel id");
+        return;
+    }
+    chan->id = res;
+    register_chan_id(chan);
+
+    /*
+     * The guest drivers depend on the device subchannels (idx #1+) to be
+     * offered after the primary channel (idx #0) of that device.  To ensure
+     * that, record the channels on the channel list in the order they appear
+     * within the device.
+     */
+    QTAILQ_INSERT_TAIL(&vmbus->channel_list, chan, link);
+}
+
+static void deinit_channel(VMBusChannel *chan)
+{
+    assert(chan->state == VMCHAN_INIT);
+    QTAILQ_REMOVE(&chan->vmbus->channel_list, chan, link);
+    unregister_chan_id(chan);
+}
+
+static void create_channels(VMBus *vmbus, VMBusDevice *dev, Error **errp)
+{
+    uint16_t i;
+    VMBusDeviceClass *vdc = VMBUS_DEVICE_GET_CLASS(dev);
+    Error *err = NULL;
+
+    dev->num_channels = vdc->num_channels ? vdc->num_channels(dev) : 1;
+    if (dev->num_channels < 1) {
+        error_setg(&err, "invalid #channels: %u", dev->num_channels);
+        goto error_out;
+    }
+
+    dev->channels = g_new0(VMBusChannel, dev->num_channels);
+    for (i = 0; i < dev->num_channels; i++) {
+        init_channel(vmbus, dev, vdc, &dev->channels[i], i, &err);
+        if (err) {
+            goto err_init;
+        }
+    }
+
+    return;
+
+err_init:
+    while (i--) {
+        deinit_channel(&dev->channels[i]);
+    }
+error_out:
+    error_propagate(errp, err);
+}
+
+static void free_channels(VMBusDevice *dev)
+{
+    uint16_t i;
+    for (i = 0; i < dev->num_channels; i++) {
+        deinit_channel(&dev->channels[i]);
+    }
+    g_free(dev->channels);
+}
+
+static HvSintRoute *make_sint_route(VMBus *vmbus, uint32_t vp_index)
+{
+    VMBusChannel *chan;
+
+    if (vp_index == vmbus->target_vp) {
+        hyperv_sint_route_ref(vmbus->sint_route);
+        return vmbus->sint_route;
+    }
+
+    QTAILQ_FOREACH(chan, &vmbus->channel_list, link) {
+        if (chan->target_vp == vp_index && vmbus_channel_is_open(chan)) {
+            hyperv_sint_route_ref(chan->notify_route);
+            return chan->notify_route;
+        }
+    }
+
+    return hyperv_sint_route_new(vp_index, VMBUS_SINT, NULL, NULL);
+}
+
+static void open_channel(VMBusChannel *chan)
+{
+    VMBusDeviceClass *vdc = VMBUS_DEVICE_GET_CLASS(chan->dev);
+
+    chan->gpadl = vmbus_get_gpadl(chan, chan->ringbuf_gpadl);
+    if (!chan->gpadl) {
+        return;
+    }
+
+    if (ringbufs_init(chan)) {
+        goto put_gpadl;
+    }
+
+    if (event_notifier_init(&chan->notifier, 0)) {
+        goto put_gpadl;
+    }
+
+    event_notifier_set_handler(&chan->notifier, channel_event_cb);
+
+    if (hyperv_set_event_flag_handler(chan_connection_id(chan),
+                                      &chan->notifier)) {
+        goto cleanup_notifier;
+    }
+
+    chan->notify_route = make_sint_route(chan->vmbus, chan->target_vp);
+    if (!chan->notify_route) {
+        goto clear_event_flag_handler;
+    }
+
+    if (vdc->open_channel && vdc->open_channel(chan)) {
+        goto unref_sint_route;
+    }
+
+    chan->is_open = true;
+    return;
+
+unref_sint_route:
+    hyperv_sint_route_unref(chan->notify_route);
+clear_event_flag_handler:
+    hyperv_set_event_flag_handler(chan_connection_id(chan), NULL);
+cleanup_notifier:
+    event_notifier_set_handler(&chan->notifier, NULL);
+    event_notifier_cleanup(&chan->notifier);
+put_gpadl:
+    vmbus_put_gpadl(chan->gpadl);
+}
+
+static void close_channel(VMBusChannel *chan)
+{
+    VMBusDeviceClass *vdc = VMBUS_DEVICE_GET_CLASS(chan->dev);
+
+    if (!chan->is_open) {
+        return;
+    }
+
+    if (vdc->close_channel) {
+        vdc->close_channel(chan);
+    }
+
+    hyperv_sint_route_unref(chan->notify_route);
+    hyperv_set_event_flag_handler(chan_connection_id(chan), NULL);
+    event_notifier_set_handler(&chan->notifier, NULL);
+    event_notifier_cleanup(&chan->notifier);
+    vmbus_put_gpadl(chan->gpadl);
+    chan->is_open = false;
+}
+
+static int channel_post_load(void *opaque, int version_id)
+{
+    VMBusChannel *chan = opaque;
+
+    return register_chan_id(chan);
+}
+
+static const VMStateDescription vmstate_channel = {
+    .name = "vmbus/channel",
+    .version_id = 0,
+    .minimum_version_id = 0,
+    .post_load = channel_post_load,
+    .fields = (VMStateField[]) {
+        VMSTATE_UINT32(id, VMBusChannel),
+        VMSTATE_UINT16(subchan_idx, VMBusChannel),
+        VMSTATE_UINT32(open_id, VMBusChannel),
+        VMSTATE_UINT32(target_vp, VMBusChannel),
+        VMSTATE_UINT32(ringbuf_gpadl, VMBusChannel),
+        VMSTATE_UINT32(ringbuf_send_offset, VMBusChannel),
+        VMSTATE_UINT8(offer_state, VMBusChannel),
+        VMSTATE_UINT8(state, VMBusChannel),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
+static VMBusChannel *find_channel(VMBus *vmbus, uint32_t id)
+{
+    VMBusChannel *chan;
+    QTAILQ_FOREACH(chan, &vmbus->channel_list, link) {
+        if (chan->id == id) {
+            return chan;
+        }
+    }
+    return NULL;
+}
+
+static int enqueue_incoming_message(VMBus *vmbus,
+                                    const struct hyperv_post_message_input *msg)
+{
+    int ret = 0;
+    uint8_t idx, prev_size;
+
+    qemu_mutex_lock(&vmbus->rx_queue_lock);
+
+    if (vmbus->rx_queue_size == HV_MSG_QUEUE_LEN) {
+        ret = -ENOBUFS;
+        goto out;
+    }
+
+    prev_size = vmbus->rx_queue_size;
+    idx = (vmbus->rx_queue_head + vmbus->rx_queue_size) % HV_MSG_QUEUE_LEN;
+    memcpy(&vmbus->rx_queue[idx], msg, sizeof(*msg));
+    vmbus->rx_queue_size++;
+
+    /* only need to resched if the queue was empty before */
+    if (!prev_size) {
+        vmbus_resched(vmbus);
+    }
+out:
+    qemu_mutex_unlock(&vmbus->rx_queue_lock);
+    return ret;
+}
+
+static uint16_t vmbus_recv_message(const struct hyperv_post_message_input *msg,
+                                   void *data)
+{
+    VMBus *vmbus = data;
+    struct vmbus_message_header *vmbus_msg;
+
+    if (msg->message_type != HV_MESSAGE_VMBUS) {
+        return HV_STATUS_INVALID_HYPERCALL_INPUT;
+    }
+
+    if (msg->payload_size < sizeof(struct vmbus_message_header)) {
+        return HV_STATUS_INVALID_HYPERCALL_INPUT;
+    }
+
+    vmbus_msg = (struct vmbus_message_header *)msg->payload;
+
+    trace_vmbus_recv_message(vmbus_msg->message_type, msg->payload_size);
+
+    if (vmbus_msg->message_type == VMBUS_MSG_INVALID ||
+        vmbus_msg->message_type >= VMBUS_MSG_COUNT) {
+        error_report("vmbus: unknown message type %#x",
+                     vmbus_msg->message_type);
+        return HV_STATUS_INVALID_HYPERCALL_INPUT;
+    }
+
+    if (enqueue_incoming_message(vmbus, msg)) {
+        return HV_STATUS_INSUFFICIENT_BUFFERS;
+    }
+    return HV_STATUS_SUCCESS;
+}
+
+static bool vmbus_initialized(VMBus *vmbus)
+{
+    return vmbus->version > 0 && vmbus->version <= VMBUS_VERSION_CURRENT;
+}
+
+static void vmbus_reset_all(VMBus *vmbus)
+{
+    qbus_reset_all(BUS(vmbus));
+}
+
+static void post_msg(VMBus *vmbus, void *msgdata, uint32_t msglen)
+{
+    int ret;
+    struct hyperv_message msg = {
+        .header.message_type = HV_MESSAGE_VMBUS,
+    };
+
+    assert(!vmbus->msg_in_progress);
+    assert(msglen <= sizeof(msg.payload));
+    assert(msglen >= sizeof(struct vmbus_message_header));
+
+    vmbus->msg_in_progress = true;
+
+    trace_vmbus_post_msg(((struct vmbus_message_header *)msgdata)->message_type,
+                         msglen);
+
+    memcpy(msg.payload, msgdata, msglen);
+    msg.header.payload_size = ROUND_UP(msglen, VMBUS_MESSAGE_SIZE_ALIGN);
+
+    ret = hyperv_post_msg(vmbus->sint_route, &msg);
+    if (ret == 0 || ret == -EAGAIN) {
+        return;
+    }
+
+    error_report("message delivery fatal failure: %d; aborting vmbus", ret);
+    vmbus_reset_all(vmbus);
+}
+
+static int vmbus_init(VMBus *vmbus)
+{
+    if (vmbus->target_vp != (uint32_t)-1) {
+        vmbus->sint_route = hyperv_sint_route_new(vmbus->target_vp, VMBUS_SINT,
+                                                  vmbus_msg_cb, vmbus);
+        if (!vmbus->sint_route) {
+            error_report("failed to set up SINT route");
+            return -ENOMEM;
+        }
+    }
+    return 0;
+}
+
+static void vmbus_deinit(VMBus *vmbus)
+{
+    VMBusGpadl *gpadl, *tmp_gpadl;
+    VMBusChannel *chan;
+
+    QTAILQ_FOREACH_SAFE(gpadl, &vmbus->gpadl_list, link, tmp_gpadl) {
+        if (gpadl->state == VMGPADL_TORNDOWN) {
+            continue;
+        }
+        vmbus_put_gpadl(gpadl);
+    }
+
+    QTAILQ_FOREACH(chan, &vmbus->channel_list, link) {
+        chan->offer_state = VMOFFER_INIT;
+    }
+
+    hyperv_sint_route_unref(vmbus->sint_route);
+    vmbus->sint_route = NULL;
+    vmbus->int_page_gpa = 0;
+    vmbus->target_vp = (uint32_t)-1;
+    vmbus->version = 0;
+    vmbus->state = VMBUS_LISTEN;
+    vmbus->msg_in_progress = false;
+}
+
+static void handle_initiate_contact(VMBus *vmbus,
+                                    vmbus_message_initiate_contact *msg,
+                                    uint32_t msglen)
+{
+    if (msglen < sizeof(*msg)) {
+        return;
+    }
+
+    trace_vmbus_initiate_contact(msg->version_requested >> 16,
+                                 msg->version_requested & 0xffff,
+                                 msg->target_vcpu, msg->monitor_page1,
+                                 msg->monitor_page2, msg->interrupt_page);
+
+    /*
+     * Reset vmbus on INITIATE_CONTACT regardless of its previous state.
+     * Useful, in particular, with vmbus-aware BIOS which can't shut vmbus down
+     * before handing over to OS loader.
+     */
+    vmbus_reset_all(vmbus);
+
+    vmbus->target_vp = msg->target_vcpu;
+    vmbus->version = msg->version_requested;
+    if (vmbus->version < VMBUS_VERSION_WIN8) {
+        /* linux passes interrupt page even when it doesn't need it */
+        vmbus->int_page_gpa = msg->interrupt_page;
+    }
+    vmbus->state = VMBUS_HANDSHAKE;
+
+    if (vmbus_init(vmbus)) {
+        error_report("failed to init vmbus; aborting");
+        vmbus_deinit(vmbus);
+        return;
+    }
+}
+
+static void send_handshake(VMBus *vmbus)
+{
+    struct vmbus_message_version_response msg = {
+        .header.message_type = VMBUS_MSG_VERSION_RESPONSE,
+        .version_supported = vmbus_initialized(vmbus),
+    };
+
+    post_msg(vmbus, &msg, sizeof(msg));
+}
+
+static void handle_request_offers(VMBus *vmbus, void *msgdata, uint32_t msglen)
+{
+    VMBusChannel *chan;
+
+    if (!vmbus_initialized(vmbus)) {
+        return;
+    }
+
+    QTAILQ_FOREACH(chan, &vmbus->channel_list, link) {
+        if (chan->offer_state == VMOFFER_INIT) {
+            chan->offer_state = VMOFFER_SENDING;
+            break;
+        }
+    }
+
+    vmbus->state = VMBUS_OFFER;
+}
+
+static void send_offer(VMBus *vmbus)
+{
+    VMBusChannel *chan;
+    struct vmbus_message_header alloffers_msg = {
+        .message_type = VMBUS_MSG_ALLOFFERS_DELIVERED,
+    };
+
+    QTAILQ_FOREACH(chan, &vmbus->channel_list, link) {
+        if (chan->offer_state == VMOFFER_SENDING) {
+            VMBusDeviceClass *vdc = VMBUS_DEVICE_GET_CLASS(chan->dev);
+            /* Hyper-V wants LE GUIDs */
+            QemuUUID classid = qemu_uuid_bswap(vdc->classid);
+            QemuUUID instanceid = qemu_uuid_bswap(chan->dev->instanceid);
+            struct vmbus_message_offer_channel msg = {
+                .header.message_type = VMBUS_MSG_OFFERCHANNEL,
+                .child_relid = chan->id,
+                .connection_id = chan_connection_id(chan),
+                .channel_flags = vdc->channel_flags,
+                .mmio_size_mb = vdc->mmio_size_mb,
+                .sub_channel_index = vmbus_channel_idx(chan),
+                .interrupt_flags = VMBUS_OFFER_INTERRUPT_DEDICATED,
+            };
+
+            memcpy(msg.type_uuid, &classid, sizeof(classid));
+            memcpy(msg.instance_uuid, &instanceid, sizeof(instanceid));
+
+            trace_vmbus_send_offer(chan->id, chan->dev);
+
+            post_msg(vmbus, &msg, sizeof(msg));
+            return;
+        }
+    }
+
+    /* no more offers, send terminator message */
+    trace_vmbus_terminate_offers();
+    post_msg(vmbus, &alloffers_msg, sizeof(alloffers_msg));
+}
+
+static bool complete_offer(VMBus *vmbus)
+{
+    VMBusChannel *chan;
+
+    QTAILQ_FOREACH(chan, &vmbus->channel_list, link) {
+        if (chan->offer_state == VMOFFER_SENDING) {
+            chan->offer_state = VMOFFER_SENT;
+            goto next_offer;
+        }
+    }
+    /*
+     * no transitioning channels found so this is completing the terminator
+     * message, and vmbus can move to the next state
+     */
+    return true;
+
+next_offer:
+    /* try to mark another channel for offering */
+    QTAILQ_FOREACH(chan, &vmbus->channel_list, link) {
+        if (chan->offer_state == VMOFFER_INIT) {
+            chan->offer_state = VMOFFER_SENDING;
+            break;
+        }
+    }
+    /*
+     * if an offer has been sent there are more offers or the terminator yet to
+     * send, so no state transition for vmbus
+     */
+    return false;
+}
+
+
+static void handle_gpadl_header(VMBus *vmbus, vmbus_message_gpadl_header *msg,
+                                uint32_t msglen)
+{
+    VMBusGpadl *gpadl;
+    uint32_t num_gfns, i;
+
+    /* must include at least one gpa range */
+    if (msglen < sizeof(*msg) + sizeof(msg->range[0]) ||
+        !vmbus_initialized(vmbus)) {
+        return;
+    }
+
+    num_gfns = (msg->range_buflen - msg->rangecount * sizeof(msg->range[0])) /
+               sizeof(msg->range[0].pfn_array[0]);
+
+    trace_vmbus_gpadl_header(msg->gpadl_id, num_gfns);
+
+    /*
+     * In theory the GPADL_HEADER message can define a GPADL with multiple GPA
+     * ranges each with arbitrary size and alignment.  However in practice only
+     * single-range page-aligned GPADLs have been observed so just ignore
+     * anything else and simplify things greatly.
+     */
+    if (msg->rangecount != 1 || msg->range[0].byte_offset ||
+        (msg->range[0].byte_count != (num_gfns << TARGET_PAGE_BITS))) {
+        return;
+    }
+
+    /* ignore requests to create already existing GPADLs */
+    if (find_gpadl(vmbus, msg->gpadl_id)) {
+        return;
+    }
+
+    gpadl = create_gpadl(vmbus, msg->gpadl_id, msg->child_relid, num_gfns);
+
+    for (i = 0; i < num_gfns &&
+         (void *)&msg->range[0].pfn_array[i + 1] <= (void *)msg + msglen;
+         i++) {
+        gpadl->gfns[gpadl->seen_gfns++] = msg->range[0].pfn_array[i];
+    }
+
+    if (gpadl_full(gpadl)) {
+        vmbus->state = VMBUS_CREATE_GPADL;
+    }
+}
+
+static void handle_gpadl_body(VMBus *vmbus, vmbus_message_gpadl_body *msg,
+                              uint32_t msglen)
+{
+    VMBusGpadl *gpadl;
+    uint32_t num_gfns_left, i;
+
+    if (msglen < sizeof(*msg) || !vmbus_initialized(vmbus)) {
+        return;
+    }
+
+    trace_vmbus_gpadl_body(msg->gpadl_id);
+
+    gpadl = find_gpadl(vmbus, msg->gpadl_id);
+    if (!gpadl) {
+        return;
+    }
+
+    num_gfns_left = gpadl->num_gfns - gpadl->seen_gfns;
+    assert(num_gfns_left);
+
+    for (i = 0; i < num_gfns_left &&
+         (void *)&msg->pfn_array[i + 1] <= (void *)msg + msglen; i++) {
+        gpadl->gfns[gpadl->seen_gfns++] = msg->pfn_array[i];
+    }
+
+    if (gpadl_full(gpadl)) {
+        vmbus->state = VMBUS_CREATE_GPADL;
+    }
+}
+
+static void send_create_gpadl(VMBus *vmbus)
+{
+    VMBusGpadl *gpadl;
+
+    QTAILQ_FOREACH(gpadl, &vmbus->gpadl_list, link) {
+        if (gpadl_full(gpadl) && gpadl->state == VMGPADL_INIT) {
+            struct vmbus_message_gpadl_created msg = {
+                .header.message_type = VMBUS_MSG_GPADL_CREATED,
+                .gpadl_id = gpadl->id,
+                .child_relid = gpadl->child_relid,
+            };
+
+            trace_vmbus_gpadl_created(gpadl->id);
+            post_msg(vmbus, &msg, sizeof(msg));
+            return;
+        }
+    }
+
+    assert(false);
+}
+
+static bool complete_create_gpadl(VMBus *vmbus)
+{
+    VMBusGpadl *gpadl;
+
+    QTAILQ_FOREACH(gpadl, &vmbus->gpadl_list, link) {
+        if (gpadl_full(gpadl) && gpadl->state == VMGPADL_INIT) {
+            gpadl->state = VMGPADL_ALIVE;
+
+            return true;
+        }
+    }
+
+    assert(false);
+    return false;
+}
+
+static void handle_gpadl_teardown(VMBus *vmbus,
+                                  vmbus_message_gpadl_teardown *msg,
+                                  uint32_t msglen)
+{
+    VMBusGpadl *gpadl;
+
+    if (msglen < sizeof(*msg) || !vmbus_initialized(vmbus)) {
+        return;
+    }
+
+    trace_vmbus_gpadl_teardown(msg->gpadl_id);
+
+    gpadl = find_gpadl(vmbus, msg->gpadl_id);
+    if (!gpadl || gpadl->state == VMGPADL_TORNDOWN) {
+        return;
+    }
+
+    gpadl->state = VMGPADL_TEARINGDOWN;
+    vmbus->state = VMBUS_TEARDOWN_GPADL;
+}
+
+static void send_teardown_gpadl(VMBus *vmbus)
+{
+    VMBusGpadl *gpadl;
+
+    QTAILQ_FOREACH(gpadl, &vmbus->gpadl_list, link) {
+        if (gpadl->state == VMGPADL_TEARINGDOWN) {
+            struct vmbus_message_gpadl_torndown msg = {
+                .header.message_type = VMBUS_MSG_GPADL_TORNDOWN,
+                .gpadl_id = gpadl->id,
+            };
+
+            trace_vmbus_gpadl_torndown(gpadl->id);
+            post_msg(vmbus, &msg, sizeof(msg));
+            return;
+        }
+    }
+
+    assert(false);
+}
+
+static bool complete_teardown_gpadl(VMBus *vmbus)
+{
+    VMBusGpadl *gpadl;
+
+    QTAILQ_FOREACH(gpadl, &vmbus->gpadl_list, link) {
+        if (gpadl->state == VMGPADL_TEARINGDOWN) {
+            gpadl->state = VMGPADL_TORNDOWN;
+            vmbus_put_gpadl(gpadl);
+            return true;
+        }
+    }
+
+    assert(false);
+    return false;
+}
+
+static void handle_open_channel(VMBus *vmbus, vmbus_message_open_channel *msg,
+                                uint32_t msglen)
+{
+    VMBusChannel *chan;
+
+    if (msglen < sizeof(*msg) || !vmbus_initialized(vmbus)) {
+        return;
+    }
+
+    trace_vmbus_open_channel(msg->child_relid, msg->ring_buffer_gpadl_id,
+                             msg->target_vp);
+    chan = find_channel(vmbus, msg->child_relid);
+    if (!chan || chan->state != VMCHAN_INIT) {
+        return;
+    }
+
+    chan->ringbuf_gpadl = msg->ring_buffer_gpadl_id;
+    chan->ringbuf_send_offset = msg->ring_buffer_offset;
+    chan->target_vp = msg->target_vp;
+    chan->open_id = msg->open_id;
+
+    open_channel(chan);
+
+    chan->state = VMCHAN_OPENING;
+    vmbus->state = VMBUS_OPEN_CHANNEL;
+}
+
+static void send_open_channel(VMBus *vmbus)
+{
+    VMBusChannel *chan;
+
+    QTAILQ_FOREACH(chan, &vmbus->channel_list, link) {
+        if (chan->state == VMCHAN_OPENING) {
+            struct vmbus_message_open_result msg = {
+                .header.message_type = VMBUS_MSG_OPENCHANNEL_RESULT,
+                .child_relid = chan->id,
+                .open_id = chan->open_id,
+                .status = !vmbus_channel_is_open(chan),
+            };
+
+            trace_vmbus_channel_open(chan->id, msg.status);
+            post_msg(vmbus, &msg, sizeof(msg));
+            return;
+        }
+    }
+
+    assert(false);
+}
+
+static bool complete_open_channel(VMBus *vmbus)
+{
+    VMBusChannel *chan;
+
+    QTAILQ_FOREACH(chan, &vmbus->channel_list, link) {
+        if (chan->state == VMCHAN_OPENING) {
+            if (vmbus_channel_is_open(chan)) {
+                chan->state = VMCHAN_OPEN;
+                /*
+                 * simulate guest notification of ringbuffer space made
+                 * available, for the channel protocols where the host
+                 * initiates the communication
+                 */
+                vmbus_channel_notify_host(chan);
+            } else {
+                chan->state = VMCHAN_INIT;
+            }
+            return true;
+        }
+    }
+
+    assert(false);
+    return false;
+}
+
+static void vdev_reset_on_close(VMBusDevice *vdev)
+{
+    uint16_t i;
+
+    for (i = 0; i < vdev->num_channels; i++) {
+        if (vmbus_channel_is_open(&vdev->channels[i])) {
+            return;
+        }
+    }
+
+    /* all channels closed -- reset device */
+    qdev_reset_all(DEVICE(vdev));
+}
+
+static void handle_close_channel(VMBus *vmbus, vmbus_message_close_channel *msg,
+                                 uint32_t msglen)
+{
+    VMBusChannel *chan;
+
+    if (msglen < sizeof(*msg) || !vmbus_initialized(vmbus)) {
+        return;
+    }
+
+    trace_vmbus_close_channel(msg->child_relid);
+
+    chan = find_channel(vmbus, msg->child_relid);
+    if (!chan) {
+        return;
+    }
+
+    close_channel(chan);
+    chan->state = VMCHAN_INIT;
+
+    vdev_reset_on_close(chan->dev);
+}
+
+static void handle_unload(VMBus *vmbus, void *msg, uint32_t msglen)
+{
+    vmbus->state = VMBUS_UNLOAD;
+}
+
+static void send_unload(VMBus *vmbus)
+{
+    vmbus_message_header msg = {
+        .message_type = VMBUS_MSG_UNLOAD_RESPONSE,
+    };
+
+    qemu_mutex_lock(&vmbus->rx_queue_lock);
+    vmbus->rx_queue_size = 0;
+    qemu_mutex_unlock(&vmbus->rx_queue_lock);
+
+    post_msg(vmbus, &msg, sizeof(msg));
+    return;
+}
+
+static bool complete_unload(VMBus *vmbus)
+{
+    vmbus_reset_all(vmbus);
+    return true;
+}
+
+static void process_message(VMBus *vmbus)
+{
+    struct hyperv_post_message_input *hv_msg;
+    struct vmbus_message_header *msg;
+    void *msgdata;
+    uint32_t msglen;
+
+    qemu_mutex_lock(&vmbus->rx_queue_lock);
+
+    if (!vmbus->rx_queue_size) {
+        goto unlock;
+    }
+
+    hv_msg = &vmbus->rx_queue[vmbus->rx_queue_head];
+    msglen =  hv_msg->payload_size;
+    if (msglen < sizeof(*msg)) {
+        goto out;
+    }
+    msgdata = hv_msg->payload;
+    msg = (struct vmbus_message_header *)msgdata;
+
+    trace_vmbus_process_incoming_message(msg->message_type);
+
+    switch (msg->message_type) {
+    case VMBUS_MSG_INITIATE_CONTACT:
+        handle_initiate_contact(vmbus, msgdata, msglen);
+        break;
+    case VMBUS_MSG_REQUESTOFFERS:
+        handle_request_offers(vmbus, msgdata, msglen);
+        break;
+    case VMBUS_MSG_GPADL_HEADER:
+        handle_gpadl_header(vmbus, msgdata, msglen);
+        break;
+    case VMBUS_MSG_GPADL_BODY:
+        handle_gpadl_body(vmbus, msgdata, msglen);
+        break;
+    case VMBUS_MSG_GPADL_TEARDOWN:
+        handle_gpadl_teardown(vmbus, msgdata, msglen);
+        break;
+    case VMBUS_MSG_OPENCHANNEL:
+        handle_open_channel(vmbus, msgdata, msglen);
+        break;
+    case VMBUS_MSG_CLOSECHANNEL:
+        handle_close_channel(vmbus, msgdata, msglen);
+        break;
+    case VMBUS_MSG_UNLOAD:
+        handle_unload(vmbus, msgdata, msglen);
+        break;
+    default:
+        error_report("unknown message type %#x", msg->message_type);
+        break;
+    }
+
+out:
+    vmbus->rx_queue_size--;
+    vmbus->rx_queue_head++;
+    vmbus->rx_queue_head %= HV_MSG_QUEUE_LEN;
+
+    vmbus_resched(vmbus);
+unlock:
+    qemu_mutex_unlock(&vmbus->rx_queue_lock);
+}
+
+static const struct {
+    void (*run)(VMBus *vmbus);
+    bool (*complete)(VMBus *vmbus);
+} state_runner[] = {
+    [VMBUS_LISTEN]         = {process_message,     NULL},
+    [VMBUS_HANDSHAKE]      = {send_handshake,      NULL},
+    [VMBUS_OFFER]          = {send_offer,          complete_offer},
+    [VMBUS_CREATE_GPADL]   = {send_create_gpadl,   complete_create_gpadl},
+    [VMBUS_TEARDOWN_GPADL] = {send_teardown_gpadl, complete_teardown_gpadl},
+    [VMBUS_OPEN_CHANNEL]   = {send_open_channel,   complete_open_channel},
+    [VMBUS_UNLOAD]         = {send_unload,         complete_unload},
+};
+
+static void vmbus_do_run(VMBus *vmbus)
+{
+    if (vmbus->msg_in_progress) {
+        return;
+    }
+
+    assert(vmbus->state < VMBUS_STATE_MAX);
+    assert(state_runner[vmbus->state].run);
+    state_runner[vmbus->state].run(vmbus);
+}
+
+static void vmbus_run(void *opaque)
+{
+    VMBus *vmbus = opaque;
+
+    /* make sure no recursion happens (e.g. due to recursive aio_poll()) */
+    if (vmbus->in_progress) {
+        return;
+    }
+
+    vmbus->in_progress = true;
+    /*
+     * FIXME: if vmbus_resched() is called from within vmbus_do_run(), it
+     * should go *after* the code that can result in aio_poll; otherwise
+     * reschedules can be missed.  No idea how to enforce that.
+     */
+    vmbus_do_run(vmbus);
+    vmbus->in_progress = false;
+}
+
+static void vmbus_msg_cb(void *data, int status)
+{
+    VMBus *vmbus = data;
+    bool (*complete)(VMBus *vmbus);
+
+    assert(vmbus->msg_in_progress);
+
+    trace_vmbus_msg_cb(status);
+
+    if (status == -EAGAIN) {
+        goto out;
+    }
+    if (status) {
+        error_report("message delivery fatal failure: %d; aborting vmbus",
+                     status);
+        vmbus_reset_all(vmbus);
+        return;
+    }
+
+    assert(vmbus->state < VMBUS_STATE_MAX);
+    complete = state_runner[vmbus->state].complete;
+    if (!complete || complete(vmbus)) {
+        vmbus->state = VMBUS_LISTEN;
+    }
+out:
+    vmbus->msg_in_progress = false;
+    vmbus_resched(vmbus);
+}
+
+static void vmbus_resched(VMBus *vmbus)
+{
+    aio_bh_schedule_oneshot(qemu_get_aio_context(), vmbus_run, vmbus);
+}
+
+static void vmbus_signal_event(EventNotifier *e)
+{
+    VMBusChannel *chan;
+    VMBus *vmbus = container_of(e, VMBus, notifier);
+    unsigned long *int_map;
+    hwaddr addr, len;
+    bool is_dirty = false;
+
+    if (!event_notifier_test_and_clear(e)) {
+        return;
+    }
+
+    trace_vmbus_signal_event();
+
+    if (!vmbus->int_page_gpa) {
+        return;
+    }
+
+    addr = vmbus->int_page_gpa + TARGET_PAGE_SIZE / 2;
+    len = TARGET_PAGE_SIZE / 2;
+    int_map = cpu_physical_memory_map(addr, &len, 1);
+    if (len != TARGET_PAGE_SIZE / 2) {
+        goto unmap;
+    }
+
+    QTAILQ_FOREACH(chan, &vmbus->channel_list, link) {
+        if (bitmap_test_and_clear_atomic(int_map, chan->id, 1)) {
+            if (!vmbus_channel_is_open(chan)) {
+                continue;
+            }
+            vmbus_channel_notify_host(chan);
+            is_dirty = true;
+        }
+    }
+
+unmap:
+    cpu_physical_memory_unmap(int_map, len, 1, is_dirty);
+}
+
+static void vmbus_dev_realize(DeviceState *dev, Error **errp)
+{
+    VMBusDevice *vdev = VMBUS_DEVICE(dev);
+    VMBusDeviceClass *vdc = VMBUS_DEVICE_GET_CLASS(vdev);
+    VMBus *vmbus = VMBUS(qdev_get_parent_bus(dev));
+    BusChild *child;
+    Error *err = NULL;
+    char idstr[UUID_FMT_LEN + 1];
+
+    assert(!qemu_uuid_is_null(&vdev->instanceid));
+
+    /* Check for instance id collision for this class id */
+    QTAILQ_FOREACH(child, &BUS(vmbus)->children, sibling) {
+        VMBusDevice *child_dev = VMBUS_DEVICE(child->child);
+
+        if (child_dev == vdev) {
+            continue;
+        }
+
+        if (qemu_uuid_is_equal(&child_dev->instanceid, &vdev->instanceid)) {
+            qemu_uuid_unparse(&vdev->instanceid, idstr);
+            error_setg(&err, "duplicate vmbus device instance id %s", idstr);
+            goto error_out;
+        }
+    }
+
+    vdev->dma_as = &address_space_memory;
+
+    create_channels(vmbus, vdev, &err);
+    if (err) {
+        goto error_out;
+    }
+
+    if (vdc->vmdev_realize) {
+        vdc->vmdev_realize(vdev, &err);
+        if (err) {
+            goto err_vdc_realize;
+        }
+    }
+    return;
+
+err_vdc_realize:
+    free_channels(vdev);
+error_out:
+    error_propagate(errp, err);
+}
+
+static void vmbus_dev_reset(DeviceState *dev)
+{
+    uint16_t i;
+    VMBusDevice *vdev = VMBUS_DEVICE(dev);
+    VMBusDeviceClass *vdc = VMBUS_DEVICE_GET_CLASS(vdev);
+
+    if (vdev->channels) {
+        for (i = 0; i < vdev->num_channels; i++) {
+            VMBusChannel *chan = &vdev->channels[i];
+            close_channel(chan);
+            chan->state = VMCHAN_INIT;
+        }
+    }
+
+    if (vdc->vmdev_reset) {
+        vdc->vmdev_reset(vdev);
+    }
+}
+
+static void vmbus_dev_unrealize(DeviceState *dev)
+{
+    VMBusDevice *vdev = VMBUS_DEVICE(dev);
+    VMBusDeviceClass *vdc = VMBUS_DEVICE_GET_CLASS(vdev);
+
+    if (vdc->vmdev_unrealize) {
+        vdc->vmdev_unrealize(vdev);
+    }
+    free_channels(vdev);
+}
+
+static void vmbus_dev_class_init(ObjectClass *klass, void *data)
+{
+    DeviceClass *kdev = DEVICE_CLASS(klass);
+    kdev->bus_type = TYPE_VMBUS;
+    kdev->realize = vmbus_dev_realize;
+    kdev->unrealize = vmbus_dev_unrealize;
+    kdev->reset = vmbus_dev_reset;
+}
+
+static Property vmbus_dev_instanceid =
+                        DEFINE_PROP_UUID("instanceid", VMBusDevice, instanceid);
+
+static void vmbus_dev_instance_init(Object *obj)
+{
+    VMBusDevice *vdev = VMBUS_DEVICE(obj);
+    VMBusDeviceClass *vdc = VMBUS_DEVICE_GET_CLASS(vdev);
+
+    if (!qemu_uuid_is_null(&vdc->instanceid)) {
+        /* Class wants to only have a single instance with a fixed UUID */
+        vdev->instanceid = vdc->instanceid;
+    } else {
+        qdev_property_add_static(DEVICE(vdev), &vmbus_dev_instanceid);
+    }
+}
+
+const VMStateDescription vmstate_vmbus_dev = {
+    .name = TYPE_VMBUS_DEVICE,
+    .version_id = 0,
+    .minimum_version_id = 0,
+    .fields = (VMStateField[]) {
+        VMSTATE_UINT8_ARRAY(instanceid.data, VMBusDevice, 16),
+        VMSTATE_UINT16(num_channels, VMBusDevice),
+        VMSTATE_STRUCT_VARRAY_POINTER_UINT16(channels, VMBusDevice,
+                                             num_channels, vmstate_channel,
+                                             VMBusChannel),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
+/* vmbus generic device base */
+static const TypeInfo vmbus_dev_type_info = {
+    .name = TYPE_VMBUS_DEVICE,
+    .parent = TYPE_DEVICE,
+    .abstract = true,
+    .instance_size = sizeof(VMBusDevice),
+    .class_size = sizeof(VMBusDeviceClass),
+    .class_init = vmbus_dev_class_init,
+    .instance_init = vmbus_dev_instance_init,
+};
+
+static void vmbus_realize(BusState *bus, Error **errp)
+{
+    int ret = 0;
+    Error *local_err = NULL;
+    VMBus *vmbus = VMBUS(bus);
+
+    qemu_mutex_init(&vmbus->rx_queue_lock);
+
+    QTAILQ_INIT(&vmbus->gpadl_list);
+    QTAILQ_INIT(&vmbus->channel_list);
+
+    ret = hyperv_set_msg_handler(VMBUS_MESSAGE_CONNECTION_ID,
+                                 vmbus_recv_message, vmbus);
+    if (ret != 0) {
+        error_setg(&local_err, "hyperv set message handler failed: %d", ret);
+        goto error_out;
+    }
+
+    ret = event_notifier_init(&vmbus->notifier, 0);
+    if (ret != 0) {
+        error_setg(&local_err, "event notifier failed to init with %d", ret);
+        goto remove_msg_handler;
+    }
+
+    event_notifier_set_handler(&vmbus->notifier, vmbus_signal_event);
+    ret = hyperv_set_event_flag_handler(VMBUS_EVENT_CONNECTION_ID,
+                                        &vmbus->notifier);
+    if (ret != 0) {
+        error_setg(&local_err, "hyperv set event handler failed with %d", ret);
+        goto clear_event_notifier;
+    }
+
+    return;
+
+clear_event_notifier:
+    event_notifier_cleanup(&vmbus->notifier);
+remove_msg_handler:
+    hyperv_set_msg_handler(VMBUS_MESSAGE_CONNECTION_ID, NULL, NULL);
+error_out:
+    qemu_mutex_destroy(&vmbus->rx_queue_lock);
+    error_propagate(errp, local_err);
+}
+
+static void vmbus_unrealize(BusState *bus)
+{
+    VMBus *vmbus = VMBUS(bus);
+
+    hyperv_set_msg_handler(VMBUS_MESSAGE_CONNECTION_ID, NULL, NULL);
+    hyperv_set_event_flag_handler(VMBUS_EVENT_CONNECTION_ID, NULL);
+    event_notifier_cleanup(&vmbus->notifier);
+
+    qemu_mutex_destroy(&vmbus->rx_queue_lock);
+}
+
+static void vmbus_reset(BusState *bus)
+{
+    vmbus_deinit(VMBUS(bus));
+}
+
+static char *vmbus_get_dev_path(DeviceState *dev)
+{
+    BusState *bus = qdev_get_parent_bus(dev);
+    return qdev_get_dev_path(bus->parent);
+}
+
+static char *vmbus_get_fw_dev_path(DeviceState *dev)
+{
+    VMBusDevice *vdev = VMBUS_DEVICE(dev);
+    char uuid[UUID_FMT_LEN + 1];
+
+    qemu_uuid_unparse(&vdev->instanceid, uuid);
+    return g_strdup_printf("%s@%s", qdev_fw_name(dev), uuid);
+}
+
+static void vmbus_class_init(ObjectClass *klass, void *data)
+{
+    BusClass *k = BUS_CLASS(klass);
+
+    k->get_dev_path = vmbus_get_dev_path;
+    k->get_fw_dev_path = vmbus_get_fw_dev_path;
+    k->realize = vmbus_realize;
+    k->unrealize = vmbus_unrealize;
+    k->reset = vmbus_reset;
+}
+
+static int vmbus_pre_load(void *opaque)
+{
+    VMBusChannel *chan;
+    VMBus *vmbus = VMBUS(opaque);
+
+    /*
+     * channel IDs allocated by the source will come in the migration stream
+     * for each channel, so clean up the ones allocated at realize
+     */
+    QTAILQ_FOREACH(chan, &vmbus->channel_list, link) {
+        unregister_chan_id(chan);
+    }
+
+    return 0;
+}
+static int vmbus_post_load(void *opaque, int version_id)
+{
+    int ret;
+    VMBus *vmbus = VMBUS(opaque);
+    VMBusGpadl *gpadl;
+    VMBusChannel *chan;
+
+    ret = vmbus_init(vmbus);
+    if (ret) {
+        return ret;
+    }
+
+    QTAILQ_FOREACH(gpadl, &vmbus->gpadl_list, link) {
+        gpadl->vmbus = vmbus;
+        gpadl->refcount = 1;
+    }
+
+    /*
+     * reopening channels depends on initialized vmbus so it's done here
+     * instead of channel_post_load()
+     */
+    QTAILQ_FOREACH(chan, &vmbus->channel_list, link) {
+
+        if (chan->state == VMCHAN_OPENING || chan->state == VMCHAN_OPEN) {
+            open_channel(chan);
+        }
+
+        if (chan->state != VMCHAN_OPEN) {
+            continue;
+        }
+
+        if (!vmbus_channel_is_open(chan)) {
+            /* reopen failed, abort loading */
+            return -1;
+        }
+
+        /* resume processing on the guest side if it missed the notification */
+        hyperv_sint_route_set_sint(chan->notify_route);
+        /* ditto on the host side */
+        vmbus_channel_notify_host(chan);
+    }
+
+    vmbus_resched(vmbus);
+    return 0;
+}
+
+static const VMStateDescription vmstate_post_message_input = {
+    .name = "vmbus/hyperv_post_message_input",
+    .version_id = 0,
+    .minimum_version_id = 0,
+    .fields = (VMStateField[]) {
+        /*
+         * skip connection_id and message_type as they are validated before
+         * queueing and ignored on dequeueing
+         */
+        VMSTATE_UINT32(payload_size, struct hyperv_post_message_input),
+        VMSTATE_UINT8_ARRAY(payload, struct hyperv_post_message_input,
+                            HV_MESSAGE_PAYLOAD_SIZE),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
+static bool vmbus_rx_queue_needed(void *opaque)
+{
+    VMBus *vmbus = VMBUS(opaque);
+    return vmbus->rx_queue_size;
+}
+
+static const VMStateDescription vmstate_rx_queue = {
+    .name = "vmbus/rx_queue",
+    .version_id = 0,
+    .minimum_version_id = 0,
+    .needed = vmbus_rx_queue_needed,
+    .fields = (VMStateField[]) {
+        VMSTATE_UINT8(rx_queue_head, VMBus),
+        VMSTATE_UINT8(rx_queue_size, VMBus),
+        VMSTATE_STRUCT_ARRAY(rx_queue, VMBus,
+                             HV_MSG_QUEUE_LEN, 0,
+                             vmstate_post_message_input,
+                             struct hyperv_post_message_input),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
+static const VMStateDescription vmstate_vmbus = {
+    .name = TYPE_VMBUS,
+    .version_id = 0,
+    .minimum_version_id = 0,
+    .pre_load = vmbus_pre_load,
+    .post_load = vmbus_post_load,
+    .fields = (VMStateField[]) {
+        VMSTATE_UINT8(state, VMBus),
+        VMSTATE_UINT32(version, VMBus),
+        VMSTATE_UINT32(target_vp, VMBus),
+        VMSTATE_UINT64(int_page_gpa, VMBus),
+        VMSTATE_QTAILQ_V(gpadl_list, VMBus, 0,
+                         vmstate_gpadl, VMBusGpadl, link),
+        VMSTATE_END_OF_LIST()
+    },
+    .subsections = (const VMStateDescription * []) {
+        &vmstate_rx_queue,
+        NULL
+    }
+};
+
+static const TypeInfo vmbus_type_info = {
+    .name = TYPE_VMBUS,
+    .parent = TYPE_BUS,
+    .instance_size = sizeof(VMBus),
+    .class_init = vmbus_class_init,
+};
+
+static void vmbus_bridge_realize(DeviceState *dev, Error **errp)
+{
+    VMBusBridge *bridge = VMBUS_BRIDGE(dev);
+
+    /*
+     * here there's at least one vmbus bridge that is being realized, so
+     * vmbus_bridge_find can only return NULL if it's not unique
+     */
+    if (!vmbus_bridge_find()) {
+        error_setg(errp, "there can be at most one %s in the system",
+                   TYPE_VMBUS_BRIDGE);
+        return;
+    }
+
+    if (!hyperv_is_synic_enabled()) {
+        error_report("VMBus requires usable Hyper-V SynIC and VP_INDEX");
+        return;
+    }
+
+    bridge->bus = VMBUS(qbus_create(TYPE_VMBUS, dev, "vmbus"));
+}
+
+static char *vmbus_bridge_ofw_unit_address(const SysBusDevice *dev)
+{
+    /* there can be only one VMBus */
+    return g_strdup("0");
+}
+
+static const VMStateDescription vmstate_vmbus_bridge = {
+    .name = TYPE_VMBUS_BRIDGE,
+    .version_id = 0,
+    .minimum_version_id = 0,
+    .fields = (VMStateField[]) {
+        VMSTATE_STRUCT_POINTER(bus, VMBusBridge, vmstate_vmbus, VMBus),
+        VMSTATE_END_OF_LIST()
+    },
+};
+
+static void vmbus_bridge_class_init(ObjectClass *klass, void *data)
+{
+    DeviceClass *k = DEVICE_CLASS(klass);
+    SysBusDeviceClass *sk = SYS_BUS_DEVICE_CLASS(klass);
+
+    k->realize = vmbus_bridge_realize;
+    k->fw_name = "vmbus";
+    sk->explicit_ofw_unit_address = vmbus_bridge_ofw_unit_address;
+    set_bit(DEVICE_CATEGORY_BRIDGE, k->categories);
+    k->vmsd = &vmstate_vmbus_bridge;
+    /* override SysBusDevice's default */
+    k->user_creatable = true;
+}
+
+static const TypeInfo vmbus_bridge_type_info = {
+    .name = TYPE_VMBUS_BRIDGE,
+    .parent = TYPE_SYS_BUS_DEVICE,
+    .instance_size = sizeof(VMBusBridge),
+    .class_init = vmbus_bridge_class_init,
+};
+
+static void vmbus_register_types(void)
+{
+    type_register_static(&vmbus_bridge_type_info);
+    type_register_static(&vmbus_dev_type_info);
+    type_register_static(&vmbus_type_info);
+}
+
+type_init(vmbus_register_types)
diff --git a/include/hw/hyperv/vmbus-bridge.h b/include/hw/hyperv/vmbus-bridge.h
new file mode 100644
index 0000000000..9cc8f780de
--- /dev/null
+++ b/include/hw/hyperv/vmbus-bridge.h
@@ -0,0 +1,32 @@ 
+/*
+ * QEMU Hyper-V VMBus root bridge
+ *
+ * Copyright (c) 2017-2018 Virtuozzo International GmbH.
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory.
+ */
+
+#ifndef HW_HYPERV_VMBUS_BRIDGE_H
+#define HW_HYPERV_VMBUS_BRIDGE_H
+
+#include "hw/sysbus.h"
+
+#define TYPE_VMBUS_BRIDGE "vmbus-bridge"
+
+typedef struct VMBus VMBus;
+
+typedef struct VMBusBridge {
+    SysBusDevice parent_obj;
+
+    VMBus *bus;
+} VMBusBridge;
+
+#define VMBUS_BRIDGE(obj) OBJECT_CHECK(VMBusBridge, (obj), TYPE_VMBUS_BRIDGE)
+
+static inline VMBusBridge *vmbus_bridge_find(void)
+{
+    return VMBUS_BRIDGE(object_resolve_path_type("", TYPE_VMBUS_BRIDGE, NULL));
+}
+
+#endif
diff --git a/include/hw/hyperv/vmbus.h b/include/hw/hyperv/vmbus.h
new file mode 100644
index 0000000000..99b647e1d6
--- /dev/null
+++ b/include/hw/hyperv/vmbus.h
@@ -0,0 +1,227 @@ 
+/*
+ * QEMU Hyper-V VMBus
+ *
+ * Copyright (c) 2017-2018 Virtuozzo International GmbH.
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory.
+ */
+
+#ifndef HW_HYPERV_VMBUS_H
+#define HW_HYPERV_VMBUS_H
+
+#include "sysemu/sysemu.h"
+#include "sysemu/dma.h"
+#include "hw/qdev-core.h"
+#include "migration/vmstate.h"
+#include "hw/hyperv/vmbus-proto.h"
+#include "qemu/uuid.h"
+
+#define TYPE_VMBUS_DEVICE "vmbus-dev"
+
+#define VMBUS_DEVICE(obj) \
+    OBJECT_CHECK(VMBusDevice, (obj), TYPE_VMBUS_DEVICE)
+#define VMBUS_DEVICE_CLASS(klass) \
+    OBJECT_CLASS_CHECK(VMBusDeviceClass, (klass), TYPE_VMBUS_DEVICE)
+#define VMBUS_DEVICE_GET_CLASS(obj) \
+    OBJECT_GET_CLASS(VMBusDeviceClass, (obj), TYPE_VMBUS_DEVICE)
+
+/*
+ * Object wrapping a GPADL -- GPA Descriptor List -- an array of guest physical
+ * pages, to be used for various buffers shared between the host and the guest.
+ */
+typedef struct VMBusGpadl VMBusGpadl;
+/*
+ * VMBus channel -- a pair of ring buffers for either direction, placed within
+ * one GPADL, and the associated notification means.
+ */
+typedef struct VMBusChannel VMBusChannel;
+/*
+ * Base class for VMBus devices.  Includes one or more channels.  Identified by
+ * class GUID and instance GUID.
+ */
+typedef struct VMBusDevice VMBusDevice;
+
+typedef void(*VMBusChannelNotifyCb)(struct VMBusChannel *chan);
+
+typedef struct VMBusDeviceClass {
+    DeviceClass parent;
+
+    QemuUUID classid;
+    QemuUUID instanceid;     /* Fixed UUID for singleton devices */
+    uint16_t channel_flags;
+    uint16_t mmio_size_mb;
+
+    /* Extentions to standard device callbacks */
+    void (*vmdev_realize)(VMBusDevice *vdev, Error **errp);
+    void (*vmdev_unrealize)(VMBusDevice *vdev);
+    void (*vmdev_reset)(VMBusDevice *vdev);
+    /*
+     * Calculate the number of channels based on the device properties.  Called
+     * at realize time.
+     **/
+    uint16_t (*num_channels)(VMBusDevice *vdev);
+    /*
+     * Device-specific actions to complete the otherwise successful process of
+     * opening a channel.
+     * Return 0 on success, -errno on failure.
+     */
+    int (*open_channel)(VMBusChannel *chan);
+    /*
+     * Device-specific actions to perform before closing a channel.
+     */
+    void (*close_channel)(VMBusChannel *chan);
+    /*
+     * Main device worker; invoked in response to notifications from either
+     * side, when there's work to do with the data in the channel ring buffers.
+     */
+    VMBusChannelNotifyCb chan_notify_cb;
+} VMBusDeviceClass;
+
+struct VMBusDevice {
+    DeviceState parent;
+    QemuUUID instanceid;
+    uint16_t num_channels;
+    VMBusChannel *channels;
+    AddressSpace *dma_as;
+};
+
+extern const VMStateDescription vmstate_vmbus_dev;
+
+/*
+ * A unit of work parsed out of a message in the receive (i.e. guest->host)
+ * ring buffer of a channel.  It's supposed to be subclassed (through
+ * embedding) by the specific devices.
+ */
+typedef struct VMBusChanReq {
+    VMBusChannel *chan;
+    uint16_t pkt_type;
+    uint32_t msglen;
+    void *msg;
+    uint64_t transaction_id;
+    bool need_comp;
+    QEMUSGList sgl;
+} VMBusChanReq;
+
+VMBusDevice *vmbus_channel_device(VMBusChannel *chan);
+VMBusChannel *vmbus_device_channel(VMBusDevice *dev, uint32_t chan_idx);
+uint32_t vmbus_channel_idx(VMBusChannel *chan);
+bool vmbus_channel_is_open(VMBusChannel *chan);
+
+/*
+ * Notify (on guest's behalf) the host side of the channel that there's data in
+ * the ringbuffer to process.
+ */
+void vmbus_channel_notify_host(VMBusChannel *chan);
+
+/*
+ * Reserve space for a packet in the send (i.e. host->guest) ringbuffer.  If
+ * there isn't enough room, indicate that to the guest, to be notified when it
+ * becomes available.
+ * Return 0 on success, negative errno on failure.
+ * The ringbuffer indices are NOT updated, the requested space indicator may.
+ */
+int vmbus_channel_reserve(VMBusChannel *chan,
+                          uint32_t desclen, uint32_t msglen);
+
+/*
+ * Send a packet to the guest.  The space for the packet MUST be reserved
+ * first.
+ * Return total number of bytes placed in the send ringbuffer on success,
+ * negative errno on failure.
+ * The ringbuffer indices are updated on success, and the guest is signaled if
+ * needed.
+ */
+ssize_t vmbus_channel_send(VMBusChannel *chan, uint16_t pkt_type,
+                           void *desc, uint32_t desclen,
+                           void *msg, uint32_t msglen,
+                           bool need_comp, uint64_t transaction_id);
+
+/*
+ * Prepare to fetch a batch of packets from the receive ring buffer.
+ * Return 0 on success, negative errno on failure.
+ */
+int vmbus_channel_recv_start(VMBusChannel *chan);
+
+/*
+ * Shortcut for a common case of sending a simple completion packet with no
+ * auxiliary descriptors.
+ */
+ssize_t vmbus_channel_send_completion(VMBusChanReq *req,
+                                      void *msg, uint32_t msglen);
+
+/*
+ * Peek at the receive (i.e. guest->host) ring buffer and extract a unit of
+ * work (a device-specific subclass of VMBusChanReq) from a packet if there's
+ * one.
+ * Return an allocated buffer, containing the request of @size with filled
+ * VMBusChanReq at the beginning, followed by the message payload, or NULL on
+ * failure.
+ * The ringbuffer indices are NOT updated, nor is the private copy of the read
+ * index.
+ */
+void *vmbus_channel_recv_peek(VMBusChannel *chan, uint32_t size);
+
+/*
+ * Update the private copy of the read index once the preceding peek is deemed
+ * successful.
+ * The ringbuffer indices are NOT updated.
+ */
+void vmbus_channel_recv_pop(VMBusChannel *chan);
+
+/*
+ * Propagate the private copy of the read index into the receive ring buffer,
+ * and thus complete the reception of a series of packets.  Notify guest if
+ * needed.
+ * Return the number of bytes popped off the receive ring buffer by the
+ * preceding recv_peek/recv_pop calls on success, negative errno on failure.
+ */
+ssize_t vmbus_channel_recv_done(VMBusChannel *chan);
+
+/*
+ * Free the request allocated by vmbus_channel_recv_peek, together with its
+ * fields.
+ */
+void vmbus_free_req(void *req);
+
+/*
+ * Find and reference a GPADL by @gpadl_id.
+ * If not found return NULL.
+ */
+VMBusGpadl *vmbus_get_gpadl(VMBusChannel *chan, uint32_t gpadl_id);
+
+/*
+ * Unreference @gpadl.  If the reference count drops to zero, free it.
+ * @gpadl may be NULL, in which case nothing is done.
+ */
+void vmbus_put_gpadl(VMBusGpadl *gpadl);
+
+/*
+ * Calculate total length in bytes of @gpadl.
+ * @gpadl must be valid.
+ */
+uint32_t vmbus_gpadl_len(VMBusGpadl *gpadl);
+
+/*
+ * Copy data from @iov to @gpadl at offset @off.
+ * Return the number of bytes copied, or a negative status on failure.
+ */
+ssize_t vmbus_iov_to_gpadl(VMBusChannel *chan, VMBusGpadl *gpadl, uint32_t off,
+                           const struct iovec *iov, size_t iov_cnt);
+
+/*
+ * Map SGList contained in the request @req, at offset @off and no more than
+ * @len bytes, for io in direction @dir, and populate @iov with the mapped
+ * iovecs.
+ * Return the number of iovecs mapped, or negative status on failure.
+ */
+int vmbus_map_sgl(VMBusChanReq *req, DMADirection dir, struct iovec *iov,
+                  unsigned iov_cnt, size_t len, size_t off);
+
+/*
+ * Unmap *iov mapped with vmbus_map_sgl, marking the number of bytes @accessed.
+ */
+void vmbus_unmap_sgl(VMBusChanReq *req, DMADirection dir, struct iovec *iov,
+                     unsigned iov_cnt, size_t accessed);
+
+#endif