[dpdk-dev] [RFC 17/24] vhost: add virtio-vhost-user transport

Stefan Hajnoczi stefanha at redhat.com
Fri Jan 19 14:44:37 CET 2018


This patch adds a new transport to librte_vhost for the
virtio-vhost-user device.  This device replaces the AF_UNIX socket used
by traditional vhost-user with a virtio device that tunnels vhost-user
protocol messages.  This allows a guest to act as a vhost device backend
for other guests.

The intended use case is for running DPDK inside a guest.  Other guests
can communicate via DPDK's "vhost" vdev driver.

For more information on virtio-vhost-user, see
https://wiki.qemu.org/Features/VirtioVhostUser.

Signed-off-by: Stefan Hajnoczi <stefanha at redhat.com>
---
 drivers/librte_vhost/Makefile                  |    1 +
 drivers/librte_vhost/vhost.h                   |    3 +
 drivers/librte_vhost/virtio_vhost_user.h       |   18 +
 drivers/librte_vhost/trans_virtio_vhost_user.c | 1050 ++++++++++++++++++++++++
 4 files changed, 1072 insertions(+)
 create mode 100644 drivers/librte_vhost/virtio_vhost_user.h
 create mode 100644 drivers/librte_vhost/trans_virtio_vhost_user.c

diff --git a/drivers/librte_vhost/Makefile b/drivers/librte_vhost/Makefile
index 8a56c32af..c8cce6fac 100644
--- a/drivers/librte_vhost/Makefile
+++ b/drivers/librte_vhost/Makefile
@@ -23,6 +23,7 @@ LDLIBS += -lrte_eal -lrte_mempool -lrte_mbuf -lrte_ethdev -lrte_net
 SRCS-$(CONFIG_RTE_LIBRTE_VHOST) := fd_man.c iotlb.c socket.c vhost.c \
 					vhost_user.c virtio_net.c \
 					trans_af_unix.c \
+					trans_virtio_vhost_user.c \
 					virtio_pci.c
 
 # install includes
diff --git a/drivers/librte_vhost/vhost.h b/drivers/librte_vhost/vhost.h
index 08ad874ef..1d4d1d139 100644
--- a/drivers/librte_vhost/vhost.h
+++ b/drivers/librte_vhost/vhost.h
@@ -338,6 +338,9 @@ struct vhost_transport_ops {
 /** The traditional AF_UNIX vhost-user protocol transport. */
 extern const struct vhost_transport_ops af_unix_trans_ops;
 
+/** The virtio-vhost-user PCI vhost-user protocol transport. */
+extern const struct vhost_transport_ops virtio_vhost_user_trans_ops;
+
 /**
  * Device structure contains all configuration information relating
  * to the device.
diff --git a/drivers/librte_vhost/virtio_vhost_user.h b/drivers/librte_vhost/virtio_vhost_user.h
new file mode 100644
index 000000000..baeaa7494
--- /dev/null
+++ b/drivers/librte_vhost/virtio_vhost_user.h
@@ -0,0 +1,18 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright (C) 2018 Red Hat, Inc.
+ */
+
+#ifndef _LINUX_VIRTIO_VHOST_USER_H
+#define _LINUX_VIRTIO_VHOST_USER_H
+
+#include <stdint.h>
+
+struct virtio_vhost_user_config {
+    uint32_t status;
+#define VIRTIO_VHOST_USER_STATUS_SLAVE_UP 0
+#define VIRTIO_VHOST_USER_STATUS_MASTER_UP 1
+    uint32_t max_vhost_queues;
+    uint8_t uuid[16];
+};
+
+#endif /* _LINUX_VIRTIO_VHOST_USER_H */
diff --git a/drivers/librte_vhost/trans_virtio_vhost_user.c b/drivers/librte_vhost/trans_virtio_vhost_user.c
new file mode 100644
index 000000000..df654eb71
--- /dev/null
+++ b/drivers/librte_vhost/trans_virtio_vhost_user.c
@@ -0,0 +1,1050 @@
+/* SPDX-License-Idenitifier: BSD-3-Clause
+ * Copyright (C) 2018 Red Hat, Inc.
+ */
+
+/*
+ * @file
+ * virtio-vhost-user PCI transport driver
+ *
+ * This vhost-user transport communicates with the vhost-user master process
+ * over the virtio-vhost-user PCI device.
+ *
+ * Interrupts are used since this is the control path, not the data path.  This
+ * way the vhost-user command processing doesn't interfere with packet
+ * processing.  This is similar to the AF_UNIX transport's fdman thread that
+ * processes socket I/O separately.
+ *
+ * This transport replaces the usual vhost-user file descriptor passing with a
+ * PCI BAR that contains doorbell registers for callfd and logfd, and shared
+ * memory for the memory table regions.
+ *
+ * VIRTIO device specification:
+ * https://stefanha.github.io/virtio/vhost-user-slave.html#x1-2830007
+ */
+
+#include <rte_log.h>
+#include <rte_malloc.h>
+#include <rte_bus_pci.h>
+#include <rte_io.h>
+
+#include "vhost.h"
+#include "virtio_pci.h"
+#include "virtqueue.h"
+#include "virtio_vhost_user.h"
+#include "vhost_user.h"
+
+/*
+ * Data structures:
+ *
+ * Successfully probed virtio-vhost-user PCI adapters are added to
+ * vvu_pci_device_list as struct vvu_pci_device elements.
+ *
+ * When rte_vhost_driver_register() is called, a struct vvu_socket is created
+ * as the endpoint for future vhost-user connections.  The struct vvu_socket is
+ * associated with the struct vvu_pci_device that will be used for
+ * communication.
+ *
+ * When a vhost-user protocol connection is established, a struct
+ * vvu_connection is created and the application's new_device(int vid) callback
+ * is invoked.
+ */
+
+/** Probed PCI devices for lookup by rte_vhost_driver_register() */
+TAILQ_HEAD(, vvu_pci_device) vvu_pci_device_list =
+	TAILQ_HEAD_INITIALIZER(vvu_pci_device_list);
+
+struct vvu_socket;
+struct vvu_connection;
+
+/** A virtio-vhost-vsock PCI adapter */
+struct vvu_pci_device {
+	struct virtio_hw hw;
+	struct rte_pci_device *pci_dev;
+	struct vvu_socket *s;
+	TAILQ_ENTRY(vvu_pci_device) next;
+};
+
+/** A vhost-user endpoint (aka per-path state) */
+struct vvu_socket {
+	struct vhost_user_socket socket; /* must be first field! */
+	struct vvu_pci_device *pdev;
+	struct vvu_connection *conn;
+
+	/** Doorbell registers */
+	uint16_t *doorbells;
+
+	/** This struct virtio_vhost_user_config field determines the number of
+	 * doorbells available so we keep it saved.
+	 */
+	uint32_t max_vhost_queues;
+
+	/** Receive buffers */
+	const struct rte_memzone *rxbuf_mz;
+
+	/** Transmit buffers.  It is assumed that the device completes them
+	 * in-order so a single wrapping index can be used to select the next
+	 * free buffer.
+	 */
+	const struct rte_memzone *txbuf_mz;
+	unsigned int txbuf_idx;
+};
+
+/** A vhost-user protocol session (aka per-vid state) */
+struct vvu_connection {
+	struct virtio_net device; /* must be first field! */
+	struct vvu_socket *s;
+};
+
+/** Virtio feature bits that we support */
+#define VVU_VIRTIO_FEATURES ((1ULL << VIRTIO_F_NOTIFY_ON_EMPTY) | \
+			     (1ULL << VIRTIO_F_ANY_LAYOUT) | \
+			     (1ULL << VIRTIO_F_VERSION_1) | \
+			     (1ULL << VIRTIO_F_IOMMU_PLATFORM))
+
+/** Virtqueue indices */
+enum {
+	VVU_VQ_RX,
+	VVU_VQ_TX,
+	VVU_VQ_MAX,
+};
+
+enum {
+	/** Receive buffer size, in bytes */
+	VVU_RXBUF_SIZE = 1024,
+
+	/** Transmit buffer size, in bytes */
+	VVU_TXBUF_SIZE = 1024,
+};
+
+/** Look up a struct vvu_pci_device from a DomBDF string */
+static struct vvu_pci_device *
+vvu_pci_by_name(const char *name)
+{
+	struct vvu_pci_device *pdev;
+
+	TAILQ_FOREACH(pdev, &vvu_pci_device_list, next) {
+		if (!strcmp(pdev->pci_dev->device.name, name))
+			return pdev;
+	}
+	return NULL;
+}
+
+/** Start connection establishment */
+static void
+vvu_connect(struct vvu_socket *s)
+{
+	struct virtio_hw *hw = &s->pdev->hw;
+	uint32_t status;
+
+	virtio_pci_read_dev_config(hw,
+			offsetof(struct virtio_vhost_user_config, status),
+			&status, sizeof(status));
+	status |= RTE_LE32(1u << VIRTIO_VHOST_USER_STATUS_SLAVE_UP);
+	virtio_pci_write_dev_config(hw,
+			offsetof(struct virtio_vhost_user_config, status),
+			&status, sizeof(status));
+}
+
+static void
+vvu_disconnect(struct vvu_socket *s)
+{
+	struct vhost_user_socket *vsocket = &s->socket;
+	struct vvu_connection *conn = s->conn;
+	struct virtio_hw *hw = &s->pdev->hw;
+	uint32_t status;
+
+	if (conn) {
+		if (vsocket->notify_ops->destroy_connection)
+			vsocket->notify_ops->destroy_connection(conn->device.vid);
+
+		vhost_destroy_device(conn->device.vid);
+	}
+
+	/* Make sure we're disconnected */
+	virtio_pci_read_dev_config(hw,
+			offsetof(struct virtio_vhost_user_config, status),
+			&status, sizeof(status));
+	status &= ~RTE_LE32(1u << VIRTIO_VHOST_USER_STATUS_SLAVE_UP);
+	virtio_pci_write_dev_config(hw,
+			offsetof(struct virtio_vhost_user_config, status),
+			&status, sizeof(status));
+}
+
+static void
+vvu_reconnect(struct vvu_socket *s)
+{
+	vvu_disconnect(s);
+	vvu_connect(s);
+}
+
+static void vvu_process_rxq(struct vvu_socket *s);
+
+static void
+vvu_cleanup_device(struct virtio_net *dev, int destroy __rte_unused)
+{
+	struct vvu_connection *conn =
+		container_of(dev, struct vvu_connection, device);
+	struct vvu_socket *s = conn->s;
+
+	s->conn = NULL;
+	vvu_process_rxq(s); /* discard old replies from master */
+	vvu_reconnect(s);
+}
+
+static int
+vvu_vring_call(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
+	struct vvu_connection *conn =
+		container_of(dev, struct vvu_connection, device);
+	struct vvu_socket *s = conn->s;
+	uint16_t vq_idx = vq->vring_idx;
+
+	RTE_LOG(DEBUG, VHOST_CONFIG, "%s vq_idx %u\n", __func__, vq_idx);
+
+	rte_write16(rte_cpu_to_le_16(vq_idx), &s->doorbells[vq_idx]);
+	return 0;
+}
+
+static int
+vvu_send_reply(struct virtio_net *dev, struct VhostUserMsg *reply)
+{
+	struct vvu_connection *conn =
+		container_of(dev, struct vvu_connection, device);
+	struct vvu_socket *s = conn->s;
+	struct virtqueue *vq = s->pdev->hw.vqs[VVU_VQ_TX];
+	struct vring_desc *desc;
+	struct vq_desc_extra *descx;
+	unsigned int i;
+	void *buf;
+	size_t len;
+
+	RTE_LOG(DEBUG, VHOST_CONFIG,
+		"%s request %u flags %#x size %u\n",
+		__func__, reply->request.master,
+		reply->flags, reply->size);
+
+	/* TODO convert reply to little-endian */
+
+	if (virtqueue_full(vq)) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Out of tx buffers\n");
+		return -1;
+	}
+
+	i = s->txbuf_idx;
+	len = VHOST_USER_HDR_SIZE + reply->size;
+	buf = (uint8_t *)s->txbuf_mz->addr + i * VVU_TXBUF_SIZE;
+
+	memcpy(buf, reply, len);
+
+	desc = &vq->vq_ring.desc[i];
+	descx = &vq->vq_descx[i];
+
+	desc->addr = rte_cpu_to_le_64(s->txbuf_mz->iova + i * VVU_TXBUF_SIZE);
+	desc->len = rte_cpu_to_le_32(len);
+	desc->flags = 0;
+
+	descx->cookie = buf;
+	descx->ndescs = 1;
+
+	vq->vq_free_cnt--;
+	s->txbuf_idx = (s->txbuf_idx + 1) & (vq->vq_nentries - 1);
+
+	vq_update_avail_ring(vq, i);
+	vq_update_avail_idx(vq);
+
+	if (virtqueue_kick_prepare(vq))
+		virtqueue_notify(vq);
+
+	return 0;
+}
+
+static int
+vvu_map_mem_regions(struct virtio_net *dev)
+{
+	struct vvu_connection *conn =
+		container_of(dev, struct vvu_connection, device);
+	struct vvu_socket *s = conn->s;
+	struct rte_pci_device *pci_dev = s->pdev->pci_dev;
+	uint8_t *mmap_addr;
+	uint32_t i;
+
+	/* Memory regions start after the doorbell registers */
+	mmap_addr = (uint8_t *)pci_dev->mem_resource[2].addr +
+		    RTE_ALIGN_CEIL((s->max_vhost_queues + 1 /* log fd */) *
+				   sizeof(uint16_t), 4096);
+
+	for (i = 0; i < dev->mem->nregions; i++) {
+		struct rte_vhost_mem_region *reg = &dev->mem->regions[i];
+
+		reg->mmap_addr = mmap_addr;
+		reg->host_user_addr = (uint64_t)(uintptr_t)reg->mmap_addr +
+				      reg->mmap_size - reg->size;
+
+		mmap_addr += reg->mmap_size;
+	}
+
+	return 0;
+}
+
+static void
+vvu_unmap_mem_regions(struct virtio_net *dev)
+{
+	uint32_t i;
+
+	for (i = 0; i < dev->mem->nregions; i++) {
+		struct rte_vhost_mem_region *reg = &dev->mem->regions[i];
+
+		/* Just clear the pointers, the PCI BAR stays there */
+		reg->mmap_addr = NULL;
+		reg->host_user_addr = 0;
+	}
+}
+
+static void vvu_process_new_connection(struct vvu_socket *s)
+{
+	struct vhost_user_socket *vsocket = &s->socket;
+	struct vvu_connection *conn;
+	struct virtio_net *dev;
+	size_t size;
+
+	dev = vhost_new_device(vsocket->trans_ops);
+	if (!dev) {
+		vvu_reconnect(s);
+		return;
+	}
+
+	conn = container_of(dev, struct vvu_connection, device);
+	conn->s = s;
+
+	size = strnlen(vsocket->path, PATH_MAX);
+	vhost_set_ifname(dev->vid, vsocket->path, size);
+
+	RTE_LOG(INFO, VHOST_CONFIG, "new device, handle is %d\n", dev->vid);
+
+	if (vsocket->notify_ops->new_connection) {
+		int ret = vsocket->notify_ops->new_connection(dev->vid);
+		if (ret < 0) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+				"failed to add vhost user connection\n");
+			vhost_destroy_device(dev->vid);
+			vvu_reconnect(s);
+			return;
+		}
+	}
+
+	s->conn = conn;
+	return;
+}
+
+static void vvu_process_status_change(struct vvu_socket *s, bool slave_up,
+				      bool master_up)
+{
+	RTE_LOG(DEBUG, VHOST_CONFIG, "%s slave_up %d master_up %d\n",
+		__func__, slave_up, master_up);
+
+	/* Disconnected from the master, try reconnecting */
+	if (!slave_up) {
+		vvu_reconnect(s);
+		return;
+	}
+
+	if (master_up && !s->conn) {
+		vvu_process_new_connection(s);
+		return;
+	}
+}
+
+static void
+vvu_process_txq(struct vvu_socket *s)
+{
+	struct virtio_hw *hw = &s->pdev->hw;
+	struct virtqueue *vq = hw->vqs[VVU_VQ_TX];
+	uint16_t n = VIRTQUEUE_NUSED(vq);
+
+	virtio_rmb();
+
+	/* Just mark the buffers complete */
+	vq->vq_used_cons_idx += n;
+	vq->vq_free_cnt += n;
+}
+
+static void
+vvu_process_rxq(struct vvu_socket *s)
+{
+	struct virtio_hw *hw = &s->pdev->hw;
+	struct virtqueue *vq = hw->vqs[VVU_VQ_RX];
+	bool refilled = false;
+
+	while (VIRTQUEUE_NUSED(vq)) {
+		struct vring_used_elem *uep;
+		VhostUserMsg *msg;
+		uint32_t len;
+		uint32_t desc_idx;
+		uint16_t used_idx;
+		size_t i;
+
+		virtio_rmb();
+
+		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
+		uep = &vq->vq_ring.used->ring[used_idx];
+		desc_idx = rte_le_to_cpu_32(uep->id);
+
+		msg = vq->vq_descx[desc_idx].cookie;
+		len = rte_le_to_cpu_32(uep->len);
+
+		if (msg->size > sizeof(VhostUserMsg) ||
+		    len != VHOST_USER_HDR_SIZE + msg->size) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+				"Invalid vhost-user message size %u, got %u bytes\n",
+				msg->size, len);
+			/* TODO reconnect */
+			abort();
+		}
+
+		RTE_LOG(DEBUG, VHOST_CONFIG,
+			"%s request %u flags %#x size %u\n",
+			__func__, msg->request.master,
+			msg->flags, msg->size);
+
+		/* Mark file descriptors invalid */
+		for (i = 0; i < RTE_DIM(msg->fds); i++)
+			msg->fds[i] = VIRTIO_INVALID_EVENTFD;
+
+		/* Only process messages while connected */
+		if (s->conn) {
+			if (vhost_user_msg_handler(s->conn->device.vid,
+						   msg) < 0) {
+				/* TODO reconnect */
+				abort();
+			}
+		}
+
+		vq->vq_used_cons_idx++;
+
+		/* Refill rxq */
+		vq_update_avail_ring(vq, desc_idx);
+		vq_update_avail_idx(vq);
+		refilled = true;
+	}
+
+	if (!refilled)
+		return;
+	if (virtqueue_kick_prepare(vq))
+		virtqueue_notify(vq);
+}
+
+/* TODO Audit thread safety.  There are 3 threads involved:
+ * 1. The main process thread that calls librte_vhost APIs during startup.
+ * 2. The interrupt thread that calls vvu_interrupt_handler().
+ * 3. Packet processing threads (lcores) calling librte_vhost APIs.
+ *
+ * It may be necessary to use locks if any of these code paths can race.  The
+ * librte_vhost API entry points already do some locking but this needs to be
+ * checked.
+ */
+static void
+vvu_interrupt_handler(void *cb_arg)
+{
+	struct vvu_socket *s = cb_arg;
+	struct virtio_hw *hw = &s->pdev->hw;
+	struct rte_intr_handle *intr_handle = &s->pdev->pci_dev->intr_handle;
+	uint8_t isr;
+
+	/* Read Interrupt Status Register (which also clears it) */
+	isr = VTPCI_OPS(hw)->get_isr(hw);
+
+	if (isr & VIRTIO_PCI_ISR_CONFIG) {
+		uint32_t status;
+		bool slave_up;
+		bool master_up;
+
+		virtio_pci_read_dev_config(hw,
+				offsetof(struct virtio_vhost_user_config, status),
+				&status, sizeof(status));
+		status = rte_le_to_cpu_32(status);
+
+		RTE_LOG(DEBUG, VHOST_CONFIG, "%s isr %#x status %#x\n", __func__, isr, status);
+
+		slave_up = status & (1u << VIRTIO_VHOST_USER_STATUS_SLAVE_UP);
+		master_up = status & (1u << VIRTIO_VHOST_USER_STATUS_MASTER_UP);
+		vvu_process_status_change(s, slave_up, master_up);
+	} else
+		RTE_LOG(DEBUG, VHOST_CONFIG, "%s isr %#x\n", __func__, isr);
+
+	/* Re-arm before processing virtqueues so no interrupts are lost */
+	rte_intr_enable(intr_handle);
+
+	vvu_process_txq(s);
+	vvu_process_rxq(s);
+}
+
+static int
+vvu_virtio_pci_init_rxq(struct vvu_socket *s)
+{
+	char name[sizeof("0000:00:00.00 vq 0 rxbufs")];
+	struct virtqueue *vq;
+	size_t size;
+	size_t align;
+	int i;
+
+	vq = s->pdev->hw.vqs[VVU_VQ_RX];
+
+	snprintf(name, sizeof(name), "%s vq %u rxbufs",
+		 s->pdev->pci_dev->device.name, VVU_VQ_RX);
+
+	/* Allocate more than sizeof(VhostUserMsg) so there is room to grow */
+	size = vq->vq_nentries * VVU_RXBUF_SIZE;
+	align = 1024;
+	s->rxbuf_mz = rte_memzone_reserve_aligned(name, size, SOCKET_ID_ANY,
+						  0, align);
+	if (!s->rxbuf_mz) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Failed to allocate rxbuf memzone\n");
+		return -1;
+	}
+
+	for (i = 0; i < vq->vq_nentries; i++) {
+		struct vring_desc *desc = &vq->vq_ring.desc[i];
+		struct vq_desc_extra *descx = &vq->vq_descx[i];
+
+		desc->addr = rte_cpu_to_le_64(s->rxbuf_mz->iova +
+				              i * VVU_RXBUF_SIZE);
+		desc->len = RTE_LE32(VVU_RXBUF_SIZE);
+		desc->flags = RTE_LE16(VRING_DESC_F_WRITE);
+
+		descx->cookie = (uint8_t *)s->rxbuf_mz->addr + i * VVU_RXBUF_SIZE;
+		descx->ndescs = 1;
+
+		vq_update_avail_ring(vq, i);
+		vq->vq_free_cnt--;
+	}
+
+	vq_update_avail_idx(vq);
+	virtqueue_notify(vq);
+	return 0;
+}
+
+static int
+vvu_virtio_pci_init_txq(struct vvu_socket *s)
+{
+	char name[sizeof("0000:00:00.00 vq 0 txbufs")];
+	struct virtqueue *vq;
+	size_t size;
+	size_t align;
+
+	vq = s->pdev->hw.vqs[VVU_VQ_TX];
+
+	snprintf(name, sizeof(name), "%s vq %u txbufs",
+		 s->pdev->pci_dev->device.name, VVU_VQ_TX);
+
+	/* Allocate more than sizeof(VhostUserMsg) so there is room to grow */
+	size = vq->vq_nentries * VVU_TXBUF_SIZE;
+	align = 1024;
+	s->txbuf_mz = rte_memzone_reserve_aligned(name, size, SOCKET_ID_ANY,
+						  0, align);
+	if (!s->txbuf_mz) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Failed to allocate txbuf memzone\n");
+		return -1;
+	}
+
+	s->txbuf_idx = 0;
+	return 0;
+}
+
+static void
+virtio_init_vring(struct virtqueue *vq)
+{
+	int size = vq->vq_nentries;
+	struct vring *vr = &vq->vq_ring;
+	uint8_t *ring_mem = vq->vq_ring_virt_mem;
+
+	memset(ring_mem, 0, vq->vq_ring_size);
+	vring_init(vr, size, ring_mem, VIRTIO_PCI_VRING_ALIGN);
+	vq->vq_used_cons_idx = 0;
+	vq->vq_desc_head_idx = 0;
+	vq->vq_avail_idx = 0;
+	vq->vq_desc_tail_idx = (uint16_t)(vq->vq_nentries - 1);
+	vq->vq_free_cnt = vq->vq_nentries;
+	memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
+
+	vring_desc_init(vr->desc, size);
+	virtqueue_enable_intr(vq);
+}
+
+static int
+vvu_virtio_pci_init_vq(struct vvu_socket *s, int vq_idx)
+{
+	char vq_name[sizeof("0000:00:00.00 vq 0")];
+	struct virtio_hw *hw = &s->pdev->hw;
+	const struct rte_memzone *mz;
+	struct virtqueue *vq;
+	uint16_t q_num;
+	size_t size;
+
+	q_num = VTPCI_OPS(hw)->get_queue_num(hw, vq_idx);
+	RTE_LOG(DEBUG, VHOST_CONFIG, "vq %d q_num: %u\n", vq_idx, q_num);
+	if (q_num == 0) {
+		RTE_LOG(ERR, VHOST_CONFIG, "virtqueue %d does not exist\n",
+			vq_idx);
+		return -1;
+	}
+
+	if (!rte_is_power_of_2(q_num)) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"virtqueue %d has non-power of 2 size (%u)\n",
+			vq_idx, q_num);
+		return -1;
+	}
+
+	snprintf(vq_name, sizeof(vq_name), "%s vq %u",
+		 s->pdev->pci_dev->device.name, vq_idx);
+
+	size = RTE_ALIGN_CEIL(sizeof(*vq) +
+			      q_num * sizeof(struct vq_desc_extra),
+			      RTE_CACHE_LINE_SIZE);
+	vq = rte_zmalloc(vq_name, size, RTE_CACHE_LINE_SIZE);
+	if (!vq) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Failed to allocated virtqueue %d\n", vq_idx);
+		return -1;
+	}
+	hw->vqs[vq_idx] = vq;
+
+	vq->hw = hw;
+	vq->vq_queue_index = vq_idx;
+	vq->vq_nentries = q_num;
+
+	size = vring_size(q_num, VIRTIO_PCI_VRING_ALIGN);
+	vq->vq_ring_size = RTE_ALIGN_CEIL(size, VIRTIO_PCI_VRING_ALIGN);
+
+	mz = rte_memzone_reserve_aligned(vq_name, vq->vq_ring_size,
+					 SOCKET_ID_ANY, 0,
+					 VIRTIO_PCI_VRING_ALIGN);
+	if (mz == NULL) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Failed to reserve memzone for virtqueue %d\n",
+			vq_idx);
+		goto err_vq;
+	}
+
+	memset(mz->addr, 0, mz->len);
+
+	vq->mz = mz;
+	vq->vq_ring_mem = mz->iova;
+	vq->vq_ring_virt_mem = mz->addr;
+	virtio_init_vring(vq);
+
+	if (VTPCI_OPS(hw)->setup_queue(hw, vq) < 0)
+		goto err_mz;
+
+	return 0;
+
+err_mz:
+	rte_memzone_free(mz);
+
+err_vq:
+	hw->vqs[vq_idx] = NULL;
+	rte_free(vq);
+	return -1;
+}
+
+static void
+vvu_virtio_pci_free_virtqueues(struct vvu_socket *s)
+{
+	struct virtio_hw *hw = &s->pdev->hw;
+	int i;
+
+	if (s->rxbuf_mz) {
+		rte_memzone_free(s->rxbuf_mz);
+		s->rxbuf_mz = NULL;
+	}
+	if (s->txbuf_mz) {
+		rte_memzone_free(s->txbuf_mz);
+		s->txbuf_mz = NULL;
+	}
+
+	for (i = 0; i < VVU_VQ_MAX; i++) {
+		struct virtqueue *vq = hw->vqs[i];
+
+		if (!vq)
+			continue;
+
+		rte_memzone_free(vq->mz);
+		rte_free(vq);
+		hw->vqs[i] = NULL;
+	}
+
+	rte_free(hw->vqs);
+	hw->vqs = NULL;
+}
+
+static void
+vvu_virtio_pci_intr_cleanup(struct vvu_socket *s)
+{
+	struct virtio_hw *hw = &s->pdev->hw;
+	struct rte_intr_handle *intr_handle = &s->pdev->pci_dev->intr_handle;
+	int i;
+
+	for (i = 0; i < VVU_VQ_MAX; i++)
+		VTPCI_OPS(hw)->set_queue_irq(hw, hw->vqs[i],
+					     VIRTIO_MSI_NO_VECTOR);
+	VTPCI_OPS(hw)->set_config_irq(hw, VIRTIO_MSI_NO_VECTOR);
+	rte_intr_disable(intr_handle);
+	rte_intr_callback_unregister(intr_handle, vvu_interrupt_handler, s);
+	rte_intr_efd_disable(intr_handle);
+}
+
+static int
+vvu_virtio_pci_init_intr(struct vvu_socket *s)
+{
+	struct virtio_hw *hw = &s->pdev->hw;
+	struct rte_intr_handle *intr_handle = &s->pdev->pci_dev->intr_handle;
+	int i;
+
+	if (!rte_intr_cap_multiple(intr_handle)) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Multiple intr vector not supported\n");
+		return -1;
+	}
+
+	if (rte_intr_efd_enable(intr_handle, VVU_VQ_MAX) < 0) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Failed to create eventfds\n");
+		return -1;
+	}
+
+	if (rte_intr_callback_register(intr_handle, vvu_interrupt_handler, s) < 0) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Failed to register interrupt callback\n");
+		goto err_efd;
+	}
+
+	if (rte_intr_enable(intr_handle) < 0)
+		goto err_callback;
+
+	if (VTPCI_OPS(hw)->set_config_irq(hw, 0) == VIRTIO_MSI_NO_VECTOR) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Failed to set config MSI-X vector\n");
+		goto err_enable;
+	}
+
+	/* TODO use separate vectors and interrupt handler functions.  It seems
+	 * <rte_interrupts.h> doesn't allow efds to have interrupt_handler
+	 * functions and it just clears efds when they are raised.  As a
+	 * workaround we use the configuration change interrupt for virtqueue
+	 * interrupts!
+	 */
+	for (i = 0; i < VVU_VQ_MAX; i++) {
+		if (VTPCI_OPS(hw)->set_queue_irq(hw, hw->vqs[i], 0) ==
+				VIRTIO_MSI_NO_VECTOR) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+				"Failed to set virtqueue MSI-X vector\n");
+			goto err_vq;
+		}
+	}
+
+	return 0;
+
+err_vq:
+	for (i = 0; i < VVU_VQ_MAX; i++)
+		VTPCI_OPS(hw)->set_queue_irq(hw, hw->vqs[i],
+					     VIRTIO_MSI_NO_VECTOR);
+	VTPCI_OPS(hw)->set_config_irq(hw, VIRTIO_MSI_NO_VECTOR);
+err_enable:
+	rte_intr_disable(intr_handle);
+err_callback:
+	rte_intr_callback_unregister(intr_handle, vvu_interrupt_handler, s);
+err_efd:
+	rte_intr_efd_disable(intr_handle);
+	return -1;
+}
+
+static int
+vvu_virtio_pci_init_bar(struct vvu_socket *s)
+{
+	struct rte_pci_device *pci_dev = s->pdev->pci_dev;
+	struct virtio_net *dev = NULL; /* just for sizeof() */
+
+	s->doorbells = pci_dev->mem_resource[2].addr;
+	if (!s->doorbells) {
+		RTE_LOG(ERR, VHOST_CONFIG, "BAR 2 not availabled\n");
+		return -1;
+	}
+
+	/* The number of doorbells is max_vhost_queues + 1 */
+	virtio_pci_read_dev_config(&s->pdev->hw,
+			offsetof(struct virtio_vhost_user_config,
+				 max_vhost_queues),
+			&s->max_vhost_queues,
+			sizeof(s->max_vhost_queues));
+	s->max_vhost_queues = rte_le_to_cpu_32(s->max_vhost_queues);
+	if (s->max_vhost_queues < RTE_DIM(dev->virtqueue)) {
+		/* We could support devices with a smaller max number of
+		 * virtqueues than dev->virtqueue[] in the future.  Fail early
+		 * for now since the current assumption is that all of
+		 * dev->virtqueue[] can be used.
+		 */
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Device supports fewer virtqueues than driver!\n");
+		return -1;
+	}
+
+	return 0;
+}
+
+static int
+vvu_virtio_pci_init(struct vvu_socket *s)
+{
+	uint64_t host_features;
+	struct virtio_hw *hw = &s->pdev->hw;
+	int i;
+
+	virtio_pci_set_status(hw, VIRTIO_CONFIG_STATUS_ACK);
+	virtio_pci_set_status(hw, VIRTIO_CONFIG_STATUS_DRIVER);
+
+	hw->guest_features = VVU_VIRTIO_FEATURES;
+	host_features = VTPCI_OPS(hw)->get_features(hw);
+	hw->guest_features = virtio_pci_negotiate_features(hw, host_features);
+
+	if (!virtio_pci_with_feature(hw, VIRTIO_F_VERSION_1)) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Missing VIRTIO 1 feature bit\n");
+		goto err;
+	}
+
+	virtio_pci_set_status(hw, VIRTIO_CONFIG_STATUS_FEATURES_OK);
+	if (!(virtio_pci_get_status(hw) & VIRTIO_CONFIG_STATUS_FEATURES_OK)) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Failed to set FEATURES_OK\n");
+		goto err;
+	}
+
+	if (vvu_virtio_pci_init_bar(s) < 0)
+		goto err;
+
+	hw->vqs = rte_zmalloc(NULL, sizeof(struct virtqueue *) * VVU_VQ_MAX, 0);
+	if (!hw->vqs)
+		goto err;
+
+	for (i = 0; i < VVU_VQ_MAX; i++) {
+		if (vvu_virtio_pci_init_vq(s, i) < 0) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+				"virtqueue %u init failed\n", i);
+			goto err_init_vq;
+		}
+	}
+
+	if (vvu_virtio_pci_init_rxq(s) < 0)
+		goto err_init_vq;
+
+	if (vvu_virtio_pci_init_txq(s) < 0)
+		goto err_init_vq;
+
+	if (vvu_virtio_pci_init_intr(s) < 0)
+		goto err_init_vq;
+
+	virtio_pci_set_status(hw, VIRTIO_CONFIG_STATUS_DRIVER_OK);
+
+	return 0;
+
+err_init_vq:
+	vvu_virtio_pci_free_virtqueues(s);
+
+err:
+	virtio_pci_reset(hw);
+	RTE_LOG(DEBUG, VHOST_CONFIG, "%s failed\n", __func__);
+	return -1;
+}
+
+static int
+vvu_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
+	      struct rte_pci_device *pci_dev)
+{
+	struct vvu_pci_device *pdev;
+
+	/* TODO support multi-process applications */
+	if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"virtio-vhost-pci does not support multi-process "
+			"applications\n");
+		return -1;
+	}
+
+	pdev = rte_zmalloc_socket(pci_dev->device.name, sizeof(*pdev),
+				  RTE_CACHE_LINE_SIZE,
+				  pci_dev->device.numa_node);
+	if (!pdev)
+		return -1;
+
+	pdev->pci_dev = pci_dev;
+
+	if (virtio_pci_init(pci_dev, &pdev->hw) != 0) {
+		rte_free(pdev);
+		return -1;
+	}
+
+	/* Reset the device now, the rest is done in vvu_socket_init() */
+	virtio_pci_reset(&pdev->hw);
+
+	if (pdev->hw.use_msix == VIRTIO_MSIX_NONE) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"MSI-X is required for PCI device at %s\n",
+			pci_dev->device.name);
+		rte_free(pdev);
+		rte_pci_unmap_device(pci_dev);
+		return -1;
+	}
+
+	TAILQ_INSERT_TAIL(&vvu_pci_device_list, pdev, next);
+
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"Added virtio-vhost-user device at %s\n",
+		pci_dev->device.name);
+
+	return 0;
+}
+
+static int
+vvu_pci_remove(struct rte_pci_device *pci_dev)
+{
+	struct vvu_pci_device *pdev;
+
+	TAILQ_FOREACH(pdev, &vvu_pci_device_list, next)
+		if (pdev->pci_dev == pci_dev)
+			break;
+	if (!pdev)
+		return -1;
+
+	if (pdev->s) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Cannot remove PCI device at %s with vhost still attached\n",
+			pci_dev->device.name);
+		return -1;
+	}
+
+	TAILQ_REMOVE(&vvu_pci_device_list, pdev, next);
+	rte_free(pdev);
+	rte_pci_unmap_device(pci_dev);
+	return 0;
+}
+
+static const struct rte_pci_id pci_id_vvu_map[] = {
+	{ RTE_PCI_DEVICE(VIRTIO_PCI_VENDORID,
+			 VIRTIO_PCI_LEGACY_DEVICEID_VHOST_USER) },
+	{ RTE_PCI_DEVICE(VIRTIO_PCI_VENDORID,
+			 VIRTIO_PCI_MODERN_DEVICEID_VHOST_USER) },
+	{ .vendor_id = 0, /* sentinel */ },
+};
+
+static struct rte_pci_driver vvu_pci_driver = {
+	.driver = {
+		.name = "virtio_vhost_user",
+	},
+	.id_table = pci_id_vvu_map,
+	.drv_flags = 0,
+	.probe = vvu_pci_probe,
+	.remove = vvu_pci_remove,
+};
+
+RTE_INIT(vvu_pci_init);
+static void
+vvu_pci_init(void)
+{
+	if (rte_eal_iopl_init() != 0) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"IOPL call failed - cannot use virtio-vhost-user\n");
+		return;
+	}
+
+	rte_pci_register(&vvu_pci_driver);
+}
+
+static int
+vvu_socket_init(struct vhost_user_socket *vsocket, uint64_t flags)
+{
+	struct vvu_socket *s =
+		container_of(vsocket, struct vvu_socket, socket);
+	struct vvu_pci_device *pdev;
+
+	if (flags & RTE_VHOST_USER_NO_RECONNECT) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"error: reconnect cannot be disabled for virtio-vhost-user\n");
+		return -1;
+	}
+	if (flags & RTE_VHOST_USER_CLIENT) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"error: virtio-vhost-user does not support client mode\n");
+		return -1;
+	}
+	if (flags & RTE_VHOST_USER_DEQUEUE_ZERO_COPY) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"error: virtio-vhost-user does not support dequeue-zero-copy\n");
+		return -1;
+	}
+
+	pdev = vvu_pci_by_name(vsocket->path);
+	if (!pdev) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Cannot find virtio-vhost-user PCI device at %s\n",
+			vsocket->path);
+		return -1;
+	}
+
+	if (pdev->s) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Device at %s is already in use\n",
+			vsocket->path);
+		return -1;
+	}
+
+	s->pdev = pdev;
+	pdev->s = s;
+
+	if (vvu_virtio_pci_init(s) < 0) {
+		s->pdev = NULL;
+		pdev->s = NULL;
+		return -1;
+	}
+
+	RTE_LOG(INFO, VHOST_CONFIG, "%s at %s\n", __func__, vsocket->path);
+	return 0;
+}
+
+static void
+vvu_socket_cleanup(struct vhost_user_socket *vsocket)
+{
+	struct vvu_socket *s =
+		container_of(vsocket, struct vvu_socket, socket);
+
+	if (s->conn)
+		vhost_destroy_device(s->conn->device.vid);
+
+	vvu_virtio_pci_intr_cleanup(s);
+	virtio_pci_reset(&s->pdev->hw);
+	vvu_virtio_pci_free_virtqueues(s);
+
+	s->pdev->s = NULL;
+	s->pdev = NULL;
+}
+
+static int
+vvu_socket_start(struct vhost_user_socket *vsocket)
+{
+	struct vvu_socket *s =
+		container_of(vsocket, struct vvu_socket, socket);
+
+	vvu_connect(s);
+	return 0;
+}
+
+const struct vhost_transport_ops virtio_vhost_user_trans_ops = {
+	.socket_size = sizeof(struct vvu_socket),
+	.device_size = sizeof(struct vvu_connection),
+	.socket_init = vvu_socket_init,
+	.socket_cleanup = vvu_socket_cleanup,
+	.socket_start = vvu_socket_start,
+	.cleanup_device = vvu_cleanup_device,
+	.vring_call = vvu_vring_call,
+	.send_reply = vvu_send_reply,
+	.map_mem_regions = vvu_map_mem_regions,
+	.unmap_mem_regions = vvu_unmap_mem_regions,
+};
-- 
2.14.3



More information about the dev mailing list