hv_sock: implements Hyper-V transport for Virtual Sockets (AF_VSOCK)
authorDexuan Cui <decui@microsoft.com>
Sat, 26 Aug 2017 04:52:43 +0000 (04:52 +0000)
committerDavid S. Miller <davem@davemloft.net>
Mon, 28 Aug 2017 22:38:18 +0000 (15:38 -0700)
Hyper-V Sockets (hv_sock) supplies a byte-stream based communication
mechanism between the host and the guest. It uses VMBus ringbuffer as the
transportation layer.

With hv_sock, applications between the host (Windows 10, Windows Server
2016 or newer) and the guest can talk with each other using the traditional
socket APIs.

More info about Hyper-V Sockets is available here:

"Make your own integration services":
https://docs.microsoft.com/en-us/virtualization/hyper-v-on-windows/user-guide/make-integration-service

The patch implements the necessary support in Linux guest by introducing a new
vsock transport for AF_VSOCK.

Signed-off-by: Dexuan Cui <decui@microsoft.com>
Cc: K. Y. Srinivasan <kys@microsoft.com>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: Stephen Hemminger <sthemmin@microsoft.com>
Cc: Andy King <acking@vmware.com>
Cc: Dmitry Torokhov <dtor@vmware.com>
Cc: George Zhang <georgezhang@vmware.com>
Cc: Jorgen Hansen <jhansen@vmware.com>
Cc: Reilly Grant <grantr@vmware.com>
Cc: Asias He <asias@redhat.com>
Cc: Stefan Hajnoczi <stefanha@redhat.com>
Cc: Vitaly Kuznetsov <vkuznets@redhat.com>
Cc: Cathy Avery <cavery@redhat.com>
Cc: Rolf Neugebauer <rolf.neugebauer@docker.com>
Cc: Marcelo Cerri <marcelo.cerri@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
MAINTAINERS
net/vmw_vsock/Kconfig
net/vmw_vsock/Makefile
net/vmw_vsock/hyperv_transport.c [new file with mode: 0644]

index 9d47ed980fffdd2d063c6571c1b7aa38dc66d8ef..6fdfe2685eededc020fd5a4064268ae3eea6581e 100644 (file)
@@ -6286,6 +6286,7 @@ F:        drivers/net/hyperv/
 F:     drivers/scsi/storvsc_drv.c
 F:     drivers/uio/uio_hv_generic.c
 F:     drivers/video/fbdev/hyperv_fb.c
+F:     net/vmw_vsock/hyperv_transport.c
 F:     include/linux/hyperv.h
 F:     tools/hv/
 F:     Documentation/ABI/stable/sysfs-bus-vmbus
index 8831e7c421672ec5775377748b940793f4f879b1..a24369d175fd6564fadcc56fd3d21e740c9eabc1 100644 (file)
@@ -46,3 +46,15 @@ config VIRTIO_VSOCKETS_COMMON
          This option is selected by any driver which needs to access
          the virtio_vsock.  The module will be called
          vmw_vsock_virtio_transport_common.
+
+config HYPERV_VSOCKETS
+       tristate "Hyper-V transport for Virtual Sockets"
+       depends on VSOCKETS && HYPERV
+       help
+         This module implements a Hyper-V transport for Virtual Sockets.
+
+         Enable this transport if your Virtual Machine host supports Virtual
+         Sockets over Hyper-V VMBus.
+
+         To compile this driver as a module, choose M here: the module will be
+         called hv_sock. If unsure, say N.
index 09fc2eb29dc88898f631bfdf3e747bf16fbe4462..e63d574234a98974be767ae9b3a081cf4949d7ff 100644 (file)
@@ -2,6 +2,7 @@ obj-$(CONFIG_VSOCKETS) += vsock.o
 obj-$(CONFIG_VMWARE_VMCI_VSOCKETS) += vmw_vsock_vmci_transport.o
 obj-$(CONFIG_VIRTIO_VSOCKETS) += vmw_vsock_virtio_transport.o
 obj-$(CONFIG_VIRTIO_VSOCKETS_COMMON) += vmw_vsock_virtio_transport_common.o
+obj-$(CONFIG_HYPERV_VSOCKETS) += hv_sock.o
 
 vsock-y += af_vsock.o af_vsock_tap.o vsock_addr.o
 
@@ -11,3 +12,5 @@ vmw_vsock_vmci_transport-y += vmci_transport.o vmci_transport_notify.o \
 vmw_vsock_virtio_transport-y += virtio_transport.o
 
 vmw_vsock_virtio_transport_common-y += virtio_transport_common.o
+
+hv_sock-y += hyperv_transport.o
diff --git a/net/vmw_vsock/hyperv_transport.c b/net/vmw_vsock/hyperv_transport.c
new file mode 100644 (file)
index 0000000..14ed5a3
--- /dev/null
@@ -0,0 +1,904 @@
+/*
+ * Hyper-V transport for vsock
+ *
+ * Hyper-V Sockets supplies a byte-stream based communication mechanism
+ * between the host and the VM. This driver implements the necessary
+ * support in the VM by introducing the new vsock transport.
+ *
+ * Copyright (c) 2017, Microsoft Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ */
+#include <linux/module.h>
+#include <linux/vmalloc.h>
+#include <linux/hyperv.h>
+#include <net/sock.h>
+#include <net/af_vsock.h>
+
+/* The host side's design of the feature requires 6 exact 4KB pages for
+ * recv/send rings respectively -- this is suboptimal considering memory
+ * consumption, however unluckily we have to live with it, before the
+ * host comes up with a better design in the future.
+ */
+#define PAGE_SIZE_4K           4096
+#define RINGBUFFER_HVS_RCV_SIZE (PAGE_SIZE_4K * 6)
+#define RINGBUFFER_HVS_SND_SIZE (PAGE_SIZE_4K * 6)
+
+/* The MTU is 16KB per the host side's design */
+#define HVS_MTU_SIZE           (1024 * 16)
+
+struct vmpipe_proto_header {
+       u32 pkt_type;
+       u32 data_size;
+};
+
+/* For recv, we use the VMBus in-place packet iterator APIs to directly copy
+ * data from the ringbuffer into the userspace buffer.
+ */
+struct hvs_recv_buf {
+       /* The header before the payload data */
+       struct vmpipe_proto_header hdr;
+
+       /* The payload */
+       u8 data[HVS_MTU_SIZE];
+};
+
+/* We can send up to HVS_MTU_SIZE bytes of payload to the host, but let's use
+ * a small size, i.e. HVS_SEND_BUF_SIZE, to minimize the dynamically-allocated
+ * buffer, because tests show there is no significant performance difference.
+ *
+ * Note: the buffer can be eliminated in the future when we add new VMBus
+ * ringbuffer APIs that allow us to directly copy data from userspace buffer
+ * to VMBus ringbuffer.
+ */
+#define HVS_SEND_BUF_SIZE (PAGE_SIZE_4K - sizeof(struct vmpipe_proto_header))
+
+struct hvs_send_buf {
+       /* The header before the payload data */
+       struct vmpipe_proto_header hdr;
+
+       /* The payload */
+       u8 data[HVS_SEND_BUF_SIZE];
+};
+
+#define HVS_HEADER_LEN (sizeof(struct vmpacket_descriptor) + \
+                        sizeof(struct vmpipe_proto_header))
+
+/* See 'prev_indices' in hv_ringbuffer_read(), hv_ringbuffer_write(), and
+ * __hv_pkt_iter_next().
+ */
+#define VMBUS_PKT_TRAILER_SIZE (sizeof(u64))
+
+#define HVS_PKT_LEN(payload_len)       (HVS_HEADER_LEN + \
+                                        ALIGN((payload_len), 8) + \
+                                        VMBUS_PKT_TRAILER_SIZE)
+
+union hvs_service_id {
+       uuid_le srv_id;
+
+       struct {
+               unsigned int svm_port;
+               unsigned char b[sizeof(uuid_le) - sizeof(unsigned int)];
+       };
+};
+
+/* Per-socket state (accessed via vsk->trans) */
+struct hvsock {
+       struct vsock_sock *vsk;
+
+       uuid_le vm_srv_id;
+       uuid_le host_srv_id;
+
+       struct vmbus_channel *chan;
+       struct vmpacket_descriptor *recv_desc;
+
+       /* The length of the payload not delivered to userland yet */
+       u32 recv_data_len;
+       /* The offset of the payload */
+       u32 recv_data_off;
+
+       /* Have we sent the zero-length packet (FIN)? */
+       bool fin_sent;
+};
+
+/* In the VM, we support Hyper-V Sockets with AF_VSOCK, and the endpoint is
+ * <cid, port> (see struct sockaddr_vm). Note: cid is not really used here:
+ * when we write apps to connect to the host, we can only use VMADDR_CID_ANY
+ * or VMADDR_CID_HOST (both are equivalent) as the remote cid, and when we
+ * write apps to bind() & listen() in the VM, we can only use VMADDR_CID_ANY
+ * as the local cid.
+ *
+ * On the host, Hyper-V Sockets are supported by Winsock AF_HYPERV:
+ * https://docs.microsoft.com/en-us/virtualization/hyper-v-on-windows/user-
+ * guide/make-integration-service, and the endpoint is <VmID, ServiceId> with
+ * the below sockaddr:
+ *
+ * struct SOCKADDR_HV
+ * {
+ *    ADDRESS_FAMILY Family;
+ *    USHORT Reserved;
+ *    GUID VmId;
+ *    GUID ServiceId;
+ * };
+ * Note: VmID is not used by Linux VM and actually it isn't transmitted via
+ * VMBus, because here it's obvious the host and the VM can easily identify
+ * each other. Though the VmID is useful on the host, especially in the case
+ * of Windows container, Linux VM doesn't need it at all.
+ *
+ * To make use of the AF_VSOCK infrastructure in Linux VM, we have to limit
+ * the available GUID space of SOCKADDR_HV so that we can create a mapping
+ * between AF_VSOCK port and SOCKADDR_HV Service GUID. The rule of writing
+ * Hyper-V Sockets apps on the host and in Linux VM is:
+ *
+ ****************************************************************************
+ * The only valid Service GUIDs, from the perspectives of both the host and *
+ * Linux VM, that can be connected by the other end, must conform to this   *
+ * format: <port>-facb-11e6-bd58-64006a7986d3, and the "port" must be in    *
+ * this range [0, 0x7FFFFFFF].                                              *
+ ****************************************************************************
+ *
+ * When we write apps on the host to connect(), the GUID ServiceID is used.
+ * When we write apps in Linux VM to connect(), we only need to specify the
+ * port and the driver will form the GUID and use that to request the host.
+ *
+ * From the perspective of Linux VM:
+ * 1. the local ephemeral port (i.e. the local auto-bound port when we call
+ * connect() without explicit bind()) is generated by __vsock_bind_stream(),
+ * and the range is [1024, 0xFFFFFFFF).
+ * 2. the remote ephemeral port (i.e. the auto-generated remote port for
+ * a connect request initiated by the host's connect()) is generated by
+ * hvs_remote_addr_init() and the range is [0x80000000, 0xFFFFFFFF).
+ */
+
+#define MAX_LISTEN_PORT                        ((u32)0x7FFFFFFF)
+#define MAX_VM_LISTEN_PORT             MAX_LISTEN_PORT
+#define MAX_HOST_LISTEN_PORT           MAX_LISTEN_PORT
+#define MIN_HOST_EPHEMERAL_PORT                (MAX_HOST_LISTEN_PORT + 1)
+
+/* 00000000-facb-11e6-bd58-64006a7986d3 */
+static const uuid_le srv_id_template =
+       UUID_LE(0x00000000, 0xfacb, 0x11e6, 0xbd, 0x58,
+               0x64, 0x00, 0x6a, 0x79, 0x86, 0xd3);
+
+static bool is_valid_srv_id(const uuid_le *id)
+{
+       return !memcmp(&id->b[4], &srv_id_template.b[4], sizeof(uuid_le) - 4);
+}
+
+static unsigned int get_port_by_srv_id(const uuid_le *svr_id)
+{
+       return *((unsigned int *)svr_id);
+}
+
+static void hvs_addr_init(struct sockaddr_vm *addr, const uuid_le *svr_id)
+{
+       unsigned int port = get_port_by_srv_id(svr_id);
+
+       vsock_addr_init(addr, VMADDR_CID_ANY, port);
+}
+
+static void hvs_remote_addr_init(struct sockaddr_vm *remote,
+                                struct sockaddr_vm *local)
+{
+       static u32 host_ephemeral_port = MIN_HOST_EPHEMERAL_PORT;
+       struct sock *sk;
+
+       vsock_addr_init(remote, VMADDR_CID_ANY, VMADDR_PORT_ANY);
+
+       while (1) {
+               /* Wrap around ? */
+               if (host_ephemeral_port < MIN_HOST_EPHEMERAL_PORT ||
+                   host_ephemeral_port == VMADDR_PORT_ANY)
+                       host_ephemeral_port = MIN_HOST_EPHEMERAL_PORT;
+
+               remote->svm_port = host_ephemeral_port++;
+
+               sk = vsock_find_connected_socket(remote, local);
+               if (!sk) {
+                       /* Found an available ephemeral port */
+                       return;
+               }
+
+               /* Release refcnt got in vsock_find_connected_socket */
+               sock_put(sk);
+       }
+}
+
+static void hvs_set_channel_pending_send_size(struct vmbus_channel *chan)
+{
+       set_channel_pending_send_size(chan,
+                                     HVS_PKT_LEN(HVS_SEND_BUF_SIZE));
+
+       /* See hvs_stream_has_space(): we must make sure the host has seen
+        * the new pending send size, before we can re-check the writable
+        * bytes.
+        */
+       virt_mb();
+}
+
+static void hvs_clear_channel_pending_send_size(struct vmbus_channel *chan)
+{
+       set_channel_pending_send_size(chan, 0);
+
+       /* Ditto */
+       virt_mb();
+}
+
+static bool hvs_channel_readable(struct vmbus_channel *chan)
+{
+       u32 readable = hv_get_bytes_to_read(&chan->inbound);
+
+       /* 0-size payload means FIN */
+       return readable >= HVS_PKT_LEN(0);
+}
+
+static int hvs_channel_readable_payload(struct vmbus_channel *chan)
+{
+       u32 readable = hv_get_bytes_to_read(&chan->inbound);
+
+       if (readable > HVS_PKT_LEN(0)) {
+               /* At least we have 1 byte to read. We don't need to return
+                * the exact readable bytes: see vsock_stream_recvmsg() ->
+                * vsock_stream_has_data().
+                */
+               return 1;
+       }
+
+       if (readable == HVS_PKT_LEN(0)) {
+               /* 0-size payload means FIN */
+               return 0;
+       }
+
+       /* No payload or FIN */
+       return -1;
+}
+
+static size_t hvs_channel_writable_bytes(struct vmbus_channel *chan)
+{
+       u32 writeable = hv_get_bytes_to_write(&chan->outbound);
+       size_t ret;
+
+       /* The ringbuffer mustn't be 100% full, and we should reserve a
+        * zero-length-payload packet for the FIN: see hv_ringbuffer_write()
+        * and hvs_shutdown().
+        */
+       if (writeable <= HVS_PKT_LEN(1) + HVS_PKT_LEN(0))
+               return 0;
+
+       ret = writeable - HVS_PKT_LEN(1) - HVS_PKT_LEN(0);
+
+       return round_down(ret, 8);
+}
+
+static int hvs_send_data(struct vmbus_channel *chan,
+                        struct hvs_send_buf *send_buf, size_t to_write)
+{
+       send_buf->hdr.pkt_type = 1;
+       send_buf->hdr.data_size = to_write;
+       return vmbus_sendpacket(chan, &send_buf->hdr,
+                               sizeof(send_buf->hdr) + to_write,
+                               0, VM_PKT_DATA_INBAND, 0);
+}
+
+static void hvs_channel_cb(void *ctx)
+{
+       struct sock *sk = (struct sock *)ctx;
+       struct vsock_sock *vsk = vsock_sk(sk);
+       struct hvsock *hvs = vsk->trans;
+       struct vmbus_channel *chan = hvs->chan;
+
+       if (hvs_channel_readable(chan))
+               sk->sk_data_ready(sk);
+
+       /* See hvs_stream_has_space(): when we reach here, the writable bytes
+        * may be already less than HVS_PKT_LEN(HVS_SEND_BUF_SIZE).
+        */
+       if (hv_get_bytes_to_write(&chan->outbound) > 0)
+               sk->sk_write_space(sk);
+}
+
+static void hvs_close_connection(struct vmbus_channel *chan)
+{
+       struct sock *sk = get_per_channel_state(chan);
+       struct vsock_sock *vsk = vsock_sk(sk);
+
+       sk->sk_state = SS_UNCONNECTED;
+       sock_set_flag(sk, SOCK_DONE);
+       vsk->peer_shutdown |= SEND_SHUTDOWN | RCV_SHUTDOWN;
+
+       sk->sk_state_change(sk);
+}
+
+static void hvs_open_connection(struct vmbus_channel *chan)
+{
+       uuid_le *if_instance, *if_type;
+       unsigned char conn_from_host;
+
+       struct sockaddr_vm addr;
+       struct sock *sk, *new = NULL;
+       struct vsock_sock *vnew;
+       struct hvsock *hvs, *hvs_new;
+       int ret;
+
+       if_type = &chan->offermsg.offer.if_type;
+       if_instance = &chan->offermsg.offer.if_instance;
+       conn_from_host = chan->offermsg.offer.u.pipe.user_def[0];
+
+       /* The host or the VM should only listen on a port in
+        * [0, MAX_LISTEN_PORT]
+        */
+       if (!is_valid_srv_id(if_type) ||
+           get_port_by_srv_id(if_type) > MAX_LISTEN_PORT)
+               return;
+
+       hvs_addr_init(&addr, conn_from_host ? if_type : if_instance);
+       sk = vsock_find_bound_socket(&addr);
+       if (!sk)
+               return;
+
+       if ((conn_from_host && sk->sk_state != VSOCK_SS_LISTEN) ||
+           (!conn_from_host && sk->sk_state != SS_CONNECTING))
+               goto out;
+
+       if (conn_from_host) {
+               if (sk->sk_ack_backlog >= sk->sk_max_ack_backlog)
+                       goto out;
+
+               new = __vsock_create(sock_net(sk), NULL, sk, GFP_KERNEL,
+                                    sk->sk_type, 0);
+               if (!new)
+                       goto out;
+
+               new->sk_state = SS_CONNECTING;
+               vnew = vsock_sk(new);
+               hvs_new = vnew->trans;
+               hvs_new->chan = chan;
+       } else {
+               hvs = vsock_sk(sk)->trans;
+               hvs->chan = chan;
+       }
+
+       set_channel_read_mode(chan, HV_CALL_DIRECT);
+       ret = vmbus_open(chan, RINGBUFFER_HVS_SND_SIZE,
+                        RINGBUFFER_HVS_RCV_SIZE, NULL, 0,
+                        hvs_channel_cb, conn_from_host ? new : sk);
+       if (ret != 0) {
+               if (conn_from_host) {
+                       hvs_new->chan = NULL;
+                       sock_put(new);
+               } else {
+                       hvs->chan = NULL;
+               }
+               goto out;
+       }
+
+       set_per_channel_state(chan, conn_from_host ? new : sk);
+       vmbus_set_chn_rescind_callback(chan, hvs_close_connection);
+
+       if (conn_from_host) {
+               new->sk_state = SS_CONNECTED;
+               sk->sk_ack_backlog++;
+
+               hvs_addr_init(&vnew->local_addr, if_type);
+               hvs_remote_addr_init(&vnew->remote_addr, &vnew->local_addr);
+
+               hvs_new->vm_srv_id = *if_type;
+               hvs_new->host_srv_id = *if_instance;
+
+               vsock_insert_connected(vnew);
+
+               lock_sock(sk);
+               vsock_enqueue_accept(sk, new);
+               release_sock(sk);
+       } else {
+               sk->sk_state = SS_CONNECTED;
+               sk->sk_socket->state = SS_CONNECTED;
+
+               vsock_insert_connected(vsock_sk(sk));
+       }
+
+       sk->sk_state_change(sk);
+
+out:
+       /* Release refcnt obtained when we called vsock_find_bound_socket() */
+       sock_put(sk);
+}
+
+static u32 hvs_get_local_cid(void)
+{
+       return VMADDR_CID_ANY;
+}
+
+static int hvs_sock_init(struct vsock_sock *vsk, struct vsock_sock *psk)
+{
+       struct hvsock *hvs;
+
+       hvs = kzalloc(sizeof(*hvs), GFP_KERNEL);
+       if (!hvs)
+               return -ENOMEM;
+
+       vsk->trans = hvs;
+       hvs->vsk = vsk;
+
+       return 0;
+}
+
+static int hvs_connect(struct vsock_sock *vsk)
+{
+       union hvs_service_id vm, host;
+       struct hvsock *h = vsk->trans;
+
+       vm.srv_id = srv_id_template;
+       vm.svm_port = vsk->local_addr.svm_port;
+       h->vm_srv_id = vm.srv_id;
+
+       host.srv_id = srv_id_template;
+       host.svm_port = vsk->remote_addr.svm_port;
+       h->host_srv_id = host.srv_id;
+
+       return vmbus_send_tl_connect_request(&h->vm_srv_id, &h->host_srv_id);
+}
+
+static int hvs_shutdown(struct vsock_sock *vsk, int mode)
+{
+       struct sock *sk = sk_vsock(vsk);
+       struct vmpipe_proto_header hdr;
+       struct hvs_send_buf *send_buf;
+       struct hvsock *hvs;
+
+       if (!(mode & SEND_SHUTDOWN))
+               return 0;
+
+       lock_sock(sk);
+
+       hvs = vsk->trans;
+       if (hvs->fin_sent)
+               goto out;
+
+       send_buf = (struct hvs_send_buf *)&hdr;
+
+       /* It can't fail: see hvs_channel_writable_bytes(). */
+       (void)hvs_send_data(hvs->chan, send_buf, 0);
+
+       hvs->fin_sent = true;
+out:
+       release_sock(sk);
+       return 0;
+}
+
+static void hvs_release(struct vsock_sock *vsk)
+{
+       struct hvsock *hvs = vsk->trans;
+       struct vmbus_channel *chan = hvs->chan;
+
+       if (chan)
+               hvs_shutdown(vsk, RCV_SHUTDOWN | SEND_SHUTDOWN);
+
+       vsock_remove_sock(vsk);
+}
+
+static void hvs_destruct(struct vsock_sock *vsk)
+{
+       struct hvsock *hvs = vsk->trans;
+       struct vmbus_channel *chan = hvs->chan;
+
+       if (chan)
+               vmbus_hvsock_device_unregister(chan);
+
+       kfree(hvs);
+}
+
+static int hvs_dgram_bind(struct vsock_sock *vsk, struct sockaddr_vm *addr)
+{
+       return -EOPNOTSUPP;
+}
+
+static int hvs_dgram_dequeue(struct vsock_sock *vsk, struct msghdr *msg,
+                            size_t len, int flags)
+{
+       return -EOPNOTSUPP;
+}
+
+static int hvs_dgram_enqueue(struct vsock_sock *vsk,
+                            struct sockaddr_vm *remote, struct msghdr *msg,
+                            size_t dgram_len)
+{
+       return -EOPNOTSUPP;
+}
+
+static bool hvs_dgram_allow(u32 cid, u32 port)
+{
+       return false;
+}
+
+static int hvs_update_recv_data(struct hvsock *hvs)
+{
+       struct hvs_recv_buf *recv_buf;
+       u32 payload_len;
+
+       recv_buf = (struct hvs_recv_buf *)(hvs->recv_desc + 1);
+       payload_len = recv_buf->hdr.data_size;
+
+       if (payload_len > HVS_MTU_SIZE)
+               return -EIO;
+
+       if (payload_len == 0)
+               hvs->vsk->peer_shutdown |= SEND_SHUTDOWN;
+
+       hvs->recv_data_len = payload_len;
+       hvs->recv_data_off = 0;
+
+       return 0;
+}
+
+static ssize_t hvs_stream_dequeue(struct vsock_sock *vsk, struct msghdr *msg,
+                                 size_t len, int flags)
+{
+       struct hvsock *hvs = vsk->trans;
+       bool need_refill = !hvs->recv_desc;
+       struct hvs_recv_buf *recv_buf;
+       u32 to_read;
+       int ret;
+
+       if (flags & MSG_PEEK)
+               return -EOPNOTSUPP;
+
+       if (need_refill) {
+               hvs->recv_desc = hv_pkt_iter_first(hvs->chan);
+               ret = hvs_update_recv_data(hvs);
+               if (ret)
+                       return ret;
+       }
+
+       recv_buf = (struct hvs_recv_buf *)(hvs->recv_desc + 1);
+       to_read = min_t(u32, len, hvs->recv_data_len);
+       ret = memcpy_to_msg(msg, recv_buf->data + hvs->recv_data_off, to_read);
+       if (ret != 0)
+               return ret;
+
+       hvs->recv_data_len -= to_read;
+       if (hvs->recv_data_len == 0) {
+               hvs->recv_desc = hv_pkt_iter_next(hvs->chan, hvs->recv_desc);
+               if (hvs->recv_desc) {
+                       ret = hvs_update_recv_data(hvs);
+                       if (ret)
+                               return ret;
+               }
+       } else {
+               hvs->recv_data_off += to_read;
+       }
+
+       return to_read;
+}
+
+static ssize_t hvs_stream_enqueue(struct vsock_sock *vsk, struct msghdr *msg,
+                                 size_t len)
+{
+       struct hvsock *hvs = vsk->trans;
+       struct vmbus_channel *chan = hvs->chan;
+       struct hvs_send_buf *send_buf;
+       ssize_t to_write, max_writable, ret;
+
+       BUILD_BUG_ON(sizeof(*send_buf) != PAGE_SIZE_4K);
+
+       send_buf = kmalloc(sizeof(*send_buf), GFP_KERNEL);
+       if (!send_buf)
+               return -ENOMEM;
+
+       max_writable = hvs_channel_writable_bytes(chan);
+       to_write = min_t(ssize_t, len, max_writable);
+       to_write = min_t(ssize_t, to_write, HVS_SEND_BUF_SIZE);
+
+       ret = memcpy_from_msg(send_buf->data, msg, to_write);
+       if (ret < 0)
+               goto out;
+
+       ret = hvs_send_data(hvs->chan, send_buf, to_write);
+       if (ret < 0)
+               goto out;
+
+       ret = to_write;
+out:
+       kfree(send_buf);
+       return ret;
+}
+
+static s64 hvs_stream_has_data(struct vsock_sock *vsk)
+{
+       struct hvsock *hvs = vsk->trans;
+       s64 ret;
+
+       if (hvs->recv_data_len > 0)
+               return 1;
+
+       switch (hvs_channel_readable_payload(hvs->chan)) {
+       case 1:
+               ret = 1;
+               break;
+       case 0:
+               vsk->peer_shutdown |= SEND_SHUTDOWN;
+               ret = 0;
+               break;
+       default: /* -1 */
+               ret = 0;
+               break;
+       }
+
+       return ret;
+}
+
+static s64 hvs_stream_has_space(struct vsock_sock *vsk)
+{
+       struct hvsock *hvs = vsk->trans;
+       struct vmbus_channel *chan = hvs->chan;
+       s64 ret;
+
+       ret = hvs_channel_writable_bytes(chan);
+       if (ret > 0)  {
+               hvs_clear_channel_pending_send_size(chan);
+       } else {
+               /* See hvs_channel_cb() */
+               hvs_set_channel_pending_send_size(chan);
+
+               /* Re-check the writable bytes to avoid race */
+               ret = hvs_channel_writable_bytes(chan);
+               if (ret > 0)
+                       hvs_clear_channel_pending_send_size(chan);
+       }
+
+       return ret;
+}
+
+static u64 hvs_stream_rcvhiwat(struct vsock_sock *vsk)
+{
+       return HVS_MTU_SIZE + 1;
+}
+
+static bool hvs_stream_is_active(struct vsock_sock *vsk)
+{
+       struct hvsock *hvs = vsk->trans;
+
+       return hvs->chan != NULL;
+}
+
+static bool hvs_stream_allow(u32 cid, u32 port)
+{
+       /* The host's port range [MIN_HOST_EPHEMERAL_PORT, 0xFFFFFFFF) is
+        * reserved as ephemeral ports, which are used as the host's ports
+        * when the host initiates connections.
+        *
+        * Perform this check in the guest so an immediate error is produced
+        * instead of a timeout.
+        */
+       if (port > MAX_HOST_LISTEN_PORT)
+               return false;
+
+       if (cid == VMADDR_CID_HOST)
+               return true;
+
+       return false;
+}
+
+static
+int hvs_notify_poll_in(struct vsock_sock *vsk, size_t target, bool *readable)
+{
+       struct hvsock *hvs = vsk->trans;
+
+       *readable = hvs_channel_readable(hvs->chan);
+       return 0;
+}
+
+static
+int hvs_notify_poll_out(struct vsock_sock *vsk, size_t target, bool *writable)
+{
+       *writable = hvs_stream_has_space(vsk) > 0;
+
+       return 0;
+}
+
+static
+int hvs_notify_recv_init(struct vsock_sock *vsk, size_t target,
+                        struct vsock_transport_recv_notify_data *d)
+{
+       return 0;
+}
+
+static
+int hvs_notify_recv_pre_block(struct vsock_sock *vsk, size_t target,
+                             struct vsock_transport_recv_notify_data *d)
+{
+       return 0;
+}
+
+static
+int hvs_notify_recv_pre_dequeue(struct vsock_sock *vsk, size_t target,
+                               struct vsock_transport_recv_notify_data *d)
+{
+       return 0;
+}
+
+static
+int hvs_notify_recv_post_dequeue(struct vsock_sock *vsk, size_t target,
+                                ssize_t copied, bool data_read,
+                                struct vsock_transport_recv_notify_data *d)
+{
+       return 0;
+}
+
+static
+int hvs_notify_send_init(struct vsock_sock *vsk,
+                        struct vsock_transport_send_notify_data *d)
+{
+       return 0;
+}
+
+static
+int hvs_notify_send_pre_block(struct vsock_sock *vsk,
+                             struct vsock_transport_send_notify_data *d)
+{
+       return 0;
+}
+
+static
+int hvs_notify_send_pre_enqueue(struct vsock_sock *vsk,
+                               struct vsock_transport_send_notify_data *d)
+{
+       return 0;
+}
+
+static
+int hvs_notify_send_post_enqueue(struct vsock_sock *vsk, ssize_t written,
+                                struct vsock_transport_send_notify_data *d)
+{
+       return 0;
+}
+
+static void hvs_set_buffer_size(struct vsock_sock *vsk, u64 val)
+{
+       /* Ignored. */
+}
+
+static void hvs_set_min_buffer_size(struct vsock_sock *vsk, u64 val)
+{
+       /* Ignored. */
+}
+
+static void hvs_set_max_buffer_size(struct vsock_sock *vsk, u64 val)
+{
+       /* Ignored. */
+}
+
+static u64 hvs_get_buffer_size(struct vsock_sock *vsk)
+{
+       return -ENOPROTOOPT;
+}
+
+static u64 hvs_get_min_buffer_size(struct vsock_sock *vsk)
+{
+       return -ENOPROTOOPT;
+}
+
+static u64 hvs_get_max_buffer_size(struct vsock_sock *vsk)
+{
+       return -ENOPROTOOPT;
+}
+
+static struct vsock_transport hvs_transport = {
+       .get_local_cid            = hvs_get_local_cid,
+
+       .init                     = hvs_sock_init,
+       .destruct                 = hvs_destruct,
+       .release                  = hvs_release,
+       .connect                  = hvs_connect,
+       .shutdown                 = hvs_shutdown,
+
+       .dgram_bind               = hvs_dgram_bind,
+       .dgram_dequeue            = hvs_dgram_dequeue,
+       .dgram_enqueue            = hvs_dgram_enqueue,
+       .dgram_allow              = hvs_dgram_allow,
+
+       .stream_dequeue           = hvs_stream_dequeue,
+       .stream_enqueue           = hvs_stream_enqueue,
+       .stream_has_data          = hvs_stream_has_data,
+       .stream_has_space         = hvs_stream_has_space,
+       .stream_rcvhiwat          = hvs_stream_rcvhiwat,
+       .stream_is_active         = hvs_stream_is_active,
+       .stream_allow             = hvs_stream_allow,
+
+       .notify_poll_in           = hvs_notify_poll_in,
+       .notify_poll_out          = hvs_notify_poll_out,
+       .notify_recv_init         = hvs_notify_recv_init,
+       .notify_recv_pre_block    = hvs_notify_recv_pre_block,
+       .notify_recv_pre_dequeue  = hvs_notify_recv_pre_dequeue,
+       .notify_recv_post_dequeue = hvs_notify_recv_post_dequeue,
+       .notify_send_init         = hvs_notify_send_init,
+       .notify_send_pre_block    = hvs_notify_send_pre_block,
+       .notify_send_pre_enqueue  = hvs_notify_send_pre_enqueue,
+       .notify_send_post_enqueue = hvs_notify_send_post_enqueue,
+
+       .set_buffer_size          = hvs_set_buffer_size,
+       .set_min_buffer_size      = hvs_set_min_buffer_size,
+       .set_max_buffer_size      = hvs_set_max_buffer_size,
+       .get_buffer_size          = hvs_get_buffer_size,
+       .get_min_buffer_size      = hvs_get_min_buffer_size,
+       .get_max_buffer_size      = hvs_get_max_buffer_size,
+};
+
+static int hvs_probe(struct hv_device *hdev,
+                    const struct hv_vmbus_device_id *dev_id)
+{
+       struct vmbus_channel *chan = hdev->channel;
+
+       hvs_open_connection(chan);
+
+       /* Always return success to suppress the unnecessary error message
+        * in vmbus_probe(): on error the host will rescind the device in
+        * 30 seconds and we can do cleanup at that time in
+        * vmbus_onoffer_rescind().
+        */
+       return 0;
+}
+
+static int hvs_remove(struct hv_device *hdev)
+{
+       struct vmbus_channel *chan = hdev->channel;
+
+       vmbus_close(chan);
+
+       return 0;
+}
+
+/* This isn't really used. See vmbus_match() and vmbus_probe() */
+static const struct hv_vmbus_device_id id_table[] = {
+       {},
+};
+
+static struct hv_driver hvs_drv = {
+       .name           = "hv_sock",
+       .hvsock         = true,
+       .id_table       = id_table,
+       .probe          = hvs_probe,
+       .remove         = hvs_remove,
+};
+
+static int __init hvs_init(void)
+{
+       int ret;
+
+       if (vmbus_proto_version < VERSION_WIN10)
+               return -ENODEV;
+
+       ret = vmbus_driver_register(&hvs_drv);
+       if (ret != 0)
+               return ret;
+
+       ret = vsock_core_init(&hvs_transport);
+       if (ret) {
+               vmbus_driver_unregister(&hvs_drv);
+               return ret;
+       }
+
+       return 0;
+}
+
+static void __exit hvs_exit(void)
+{
+       vsock_core_exit();
+       vmbus_driver_unregister(&hvs_drv);
+}
+
+module_init(hvs_init);
+module_exit(hvs_exit);
+
+MODULE_DESCRIPTION("Hyper-V Sockets");
+MODULE_VERSION("1.0.0");
+MODULE_LICENSE("GPL");
+MODULE_ALIAS_NETPROTO(PF_VSOCK);