[dpdk-dev] [PATCH] [v1]vhost: support inflight share memory protocol feature

[dpdk-dev] [PATCH] [v1]vhost: support inflight share memory protocol feature

[JinYu]

This patch introduces two new messages VHOST_USER_GET_INFLIGHT_FD
and VHOST_USER_SET_INFLIGHT_FD to support transferring a shared
buffer between qemu and backend.

Firstly, qemu uses VHOST_USER_GET_INFLIGHT_FD to get the
shared buffer from backend. Then qemu should send it back
through VHOST_USER_SET_INFLIGHT_FD each time we start vhost-user.

This shared buffer is used to process inflight I/O when backend
reconnect.

Signed-off-by: LinLi <lilin24@baidu.com>
Signed-off-by: XunNi <nixun@baidu.com>
Signed-off-by: YuZhang <zhangyu31@baidu.com>
Signed-off-by: JinYu <jin.yu@intel.com>
---
V1 - specify the APIs are split-ring only
---
 lib/librte_vhost/rte_vhost.h           |  99 +++++++++
 lib/librte_vhost/rte_vhost_version.map |   4 +
 lib/librte_vhost/vhost.c               | 135 ++++++++++++
 lib/librte_vhost/vhost.h               |  12 +
 lib/librte_vhost/vhost_user.c          | 292 +++++++++++++++++++++++++
 lib/librte_vhost/vhost_user.h          |  13 +-
 6 files changed, 554 insertions(+), 1 deletion(-)

diff --git a/lib/librte_vhost/rte_vhost.h b/lib/librte_vhost/rte_vhost.h
index 0226b3eff..c9959c15a 100644
--- a/lib/librte_vhost/rte_vhost.h
+++ b/lib/librte_vhost/rte_vhost.h
@@ -71,6 +71,10 @@ extern "C" {
 #define VHOST_USER_PROTOCOL_F_HOST_NOTIFIER 11
 #endif
 
+#ifndef VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD
+#define VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD 12
+#endif
+
 /** Indicate whether protocol features negotiation is supported. */
 #ifndef VHOST_USER_F_PROTOCOL_FEATURES
 #define VHOST_USER_F_PROTOCOL_FEATURES	30
@@ -98,12 +102,41 @@ struct rte_vhost_memory {
 	struct rte_vhost_mem_region regions[];
 };
 
+struct inflight_desc_split {
+	uint8_t		inflight;
+	uint8_t		padding[5];
+	uint16_t	next;
+	uint64_t	counter;
+};
+
+struct inflight_info_split {
+	uint64_t		features;
+	uint16_t		version;
+	uint16_t		desc_num;
+	uint16_t		last_inflight_io;
+	uint16_t		used_idx;
+	struct inflight_desc_split desc[0];
+};
+
+struct resubmit_desc {
+	uint16_t index;
+	uint64_t counter;
+};
+
+struct resubmit_info {
+	struct resubmit_desc	*resubmit_list;
+	uint16_t		resubmit_num;
+};
+
 struct rte_vhost_vring {
 	struct vring_desc	*desc;
 	struct vring_avail	*avail;
 	struct vring_used	*used;
 	uint64_t		log_guest_addr;
 
+	struct inflight_info_split *inflight;
+	struct resubmit_info	*resubmit_inflight;
+
 	/** Deprecated, use rte_vhost_vring_call() instead. */
 	int			callfd;
 
@@ -603,6 +636,22 @@ uint16_t rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
  */
 int rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem);
 
+/**
+ * Get guest vring info, including the vring address, vring size, inflight, etc.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param vring
+ *  the structure to hold the requested vring info
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_get_vhost_vring_with_inflight_split(int vid, uint16_t vring_idx,
+			      struct rte_vhost_vring *vring);
+
 /**
  * Get guest vring info, including the vring address, vring size, etc.
  *
@@ -631,6 +680,56 @@ int rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
  */
 int rte_vhost_vring_call(int vid, uint16_t vring_idx);
 
+/**
+ * set inflight flag for a desc.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
+		uint16_t idx);
+
+/**
+ * clear inflight flag for a desc.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param last_used_idx
+ *  next free used_idx
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
+		uint16_t last_used_idx, uint16_t idx);
+
+/**
+ * set last inflight io index.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
+		uint16_t idx);
+
 /**
  * Get vhost RX queue avail count.
  *
diff --git a/lib/librte_vhost/rte_vhost_version.map b/lib/librte_vhost/rte_vhost_version.map
index 5f1d4a75c..709593701 100644
--- a/lib/librte_vhost/rte_vhost_version.map
+++ b/lib/librte_vhost/rte_vhost_version.map
@@ -87,4 +87,8 @@ EXPERIMENTAL {
 	rte_vdpa_relay_vring_used;
 	rte_vhost_extern_callback_register;
 	rte_vhost_driver_set_protocol_features;
+	rte_vhost_set_inflight_desc_split;
+	rte_vhost_clr_inflight_desc_split;
+	rte_vhost_set_last_inflight_io_split;
+	rte_vhost_get_vhost_vring_with_inflight_split;
 };
diff --git a/lib/librte_vhost/vhost.c b/lib/librte_vhost/vhost.c
index 981837b5d..ed40b5b58 100644
--- a/lib/librte_vhost/vhost.c
+++ b/lib/librte_vhost/vhost.c
@@ -240,6 +240,16 @@ cleanup_vq(struct vhost_virtqueue *vq, int destroy)
 		close(vq->callfd);
 	if (vq->kickfd >= 0)
 		close(vq->kickfd);
+	if (vq->inflight)
+		vq->inflight = NULL;
+	if (vq->resubmit_inflight->resubmit_list) {
+		free(vq->resubmit_inflight->resubmit_list);
+		vq->resubmit_inflight->resubmit_list = NULL;
+	}
+	if (vq->resubmit_inflight) {
+		free(vq->resubmit_inflight);
+		vq->resubmit_inflight = NULL;
+	}
 }
 
 /*
@@ -726,6 +736,39 @@ rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
 	return 0;
 }
 
+int __rte_experimental
+rte_vhost_get_vhost_vring_with_inflight_split(int vid, uint16_t vring_idx,
+			  struct rte_vhost_vring *vring)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (dev == NULL || vring == NULL)
+		return -1;
+
+	if (vring_idx >= VHOST_MAX_VRING)
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (!vq)
+		return -1;
+
+	vring->desc  = vq->desc;
+	vring->avail = vq->avail;
+	vring->used  = vq->used;
+	vring->log_guest_addr  = vq->log_guest_addr;
+
+	vring->callfd  = vq->callfd;
+	vring->kickfd  = vq->kickfd;
+	vring->size    = vq->size;
+
+	vring->inflight = vq->inflight;
+	vring->resubmit_inflight = vq->resubmit_inflight;
+
+	return 0;
+}
+
 int
 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
 			  struct rte_vhost_vring *vring)
@@ -781,6 +824,98 @@ rte_vhost_vring_call(int vid, uint16_t vring_idx)
 	return 0;
 }
 
+int
+rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (unlikely(!vq))
+		return -1;
+
+	if (unlikely(!vq->inflight))
+		return -1;
+
+	vq->inflight->desc[idx].counter = vq->counter++;
+	vq->inflight->desc[idx].inflight = 1;
+	return 0;
+}
+
+int
+rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
+	uint16_t last_used_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (unlikely(!vq))
+		return -1;
+
+	if (unlikely(!vq->inflight))
+		return -1;
+
+	rte_compiler_barrier();
+
+	vq->inflight->desc[idx].inflight = 0;
+
+	rte_compiler_barrier();
+
+	vq->inflight->used_idx = last_used_idx;
+	return 0;
+}
+
+int
+rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (!vq)
+		return -1;
+
+	if (unlikely(!vq->inflight))
+		return -1;
+
+	vq->inflight->last_inflight_io = idx;
+	return 0;
+}
+
 uint16_t
 rte_vhost_avail_entries(int vid, uint16_t queue_id)
 {
diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
index 884befa85..c18356c15 100644
--- a/lib/librte_vhost/vhost.h
+++ b/lib/librte_vhost/vhost.h
@@ -128,6 +128,11 @@ struct vhost_virtqueue {
 	/* Physical address of used ring, for logging */
 	uint64_t		log_guest_addr;
 
+	/* Inflight share memory info */
+	struct inflight_info_split *inflight;
+	struct resubmit_info	*resubmit_inflight;
+	uint64_t		counter;
+
 	uint16_t		nr_zmbuf;
 	uint16_t		zmbuf_size;
 	uint16_t		last_zmbuf_idx;
@@ -286,6 +291,12 @@ struct guest_page {
 	uint64_t size;
 };
 
+struct inflight_mem_info {
+	int		fd;
+	void		*addr;
+	uint64_t	size;
+};
+
 /**
  * Device structure contains all configuration information relating
  * to the device.
@@ -303,6 +314,7 @@ struct virtio_net {
 	uint32_t		nr_vring;
 	int			dequeue_zero_copy;
 	struct vhost_virtqueue	*virtqueue[VHOST_MAX_QUEUE_PAIRS * 2];
+	struct inflight_mem_info inflight_info;
 #define IF_NAME_SZ (PATH_MAX > IFNAMSIZ ? PATH_MAX : IFNAMSIZ)
 	char			ifname[IF_NAME_SZ];
 	uint64_t		log_size;
diff --git a/lib/librte_vhost/vhost_user.c b/lib/librte_vhost/vhost_user.c
index c9e29ece8..41dfeb3b2 100644
--- a/lib/librte_vhost/vhost_user.c
+++ b/lib/librte_vhost/vhost_user.c
@@ -31,6 +31,8 @@
 #include <sys/stat.h>
 #include <sys/syscall.h>
 #include <assert.h>
+#include <sys/syscall.h>
+#include <asm/unistd.h>
 #ifdef RTE_LIBRTE_VHOST_NUMA
 #include <numaif.h>
 #endif
@@ -49,6 +51,15 @@
 #define VIRTIO_MIN_MTU 68
 #define VIRTIO_MAX_MTU 65535
 
+#define INFLIGHT_ALIGNMENT	64
+#define INFLIGHT_VERSION	0xabcd
+#define VIRTQUEUE_MAX_SIZE	1024
+
+#define CLOEXEC	0x0001U
+
+#define ALIGN_DOWN(n, m) ((n) / (m) * (m))
+#define ALIGN_UP(n, m) ALIGN_DOWN((n) + (m) - 1, (m))
+
 static const char *vhost_message_str[VHOST_USER_MAX] = {
 	[VHOST_USER_NONE] = "VHOST_USER_NONE",
 	[VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
@@ -78,6 +89,8 @@ static const char *vhost_message_str[VHOST_USER_MAX] = {
 	[VHOST_USER_POSTCOPY_ADVISE]  = "VHOST_USER_POSTCOPY_ADVISE",
 	[VHOST_USER_POSTCOPY_LISTEN]  = "VHOST_USER_POSTCOPY_LISTEN",
 	[VHOST_USER_POSTCOPY_END]  = "VHOST_USER_POSTCOPY_END",
+	[VHOST_USER_GET_INFLIGHT_FD] = "VHOST_USER_GET_INFLIGHT_FD",
+	[VHOST_USER_SET_INFLIGHT_FD] = "VHOST_USER_SET_INFLIGHT_FD",
 };
 
 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
@@ -160,6 +173,16 @@ vhost_backend_cleanup(struct virtio_net *dev)
 		dev->log_addr = 0;
 	}
 
+	if (dev->inflight_info.addr) {
+		munmap(dev->inflight_info.addr, dev->inflight_info.size);
+		dev->inflight_info.addr = NULL;
+	}
+
+	if (dev->inflight_info.fd > 0) {
+		close(dev->inflight_info.fd);
+		dev->inflight_info.fd = -1;
+	}
+
 	if (dev->slave_req_fd >= 0) {
 		close(dev->slave_req_fd);
 		dev->slave_req_fd = -1;
@@ -1165,6 +1188,184 @@ virtio_is_ready(struct virtio_net *dev)
 	return 1;
 }
 
+static int mem_create(const char *name, unsigned int flags)
+{
+#ifdef __NR_memfd_create
+	return syscall(__NR_memfd_create, name, flags);
+#else
+	return -1;
+#endif
+}
+
+void *inflight_mem_alloc(const char *name, size_t size, int *fd)
+{
+	void *ptr;
+	int mfd = -1;
+	char fname[20] = "/tmp/memfd-XXXXXX";
+
+	*fd = -1;
+	mfd = mem_create(name, CLOEXEC);
+	if (mfd != -1) {
+		if (ftruncate(mfd, size) == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"ftruncate fail for alloc inflight buffer\n");
+			close(mfd);
+			return NULL;
+		}
+	} else {
+		mfd = mkstemp(fname);
+		unlink(fname);
+
+		if (mfd == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"mkstemp fail for alloc inflight buffer\n");
+			return NULL;
+		}
+
+		if (ftruncate(mfd, size) == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"ftruncate fail for alloc inflight buffer\n");
+			close(mfd);
+			return NULL;
+		}
+	}
+
+	ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
+	if (ptr == MAP_FAILED) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+				"mmap fail for alloc inflight buffer\n");
+		close(mfd);
+		return NULL;
+	}
+
+	*fd = mfd;
+	return ptr;
+}
+
+static uint32_t get_pervq_shm_size_split(uint16_t queue_size)
+{
+	return ALIGN_UP(sizeof(struct inflight_desc_split) * queue_size +
+		sizeof(uint64_t) * 1 + sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
+}
+
+static int
+vhost_user_get_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
+		int main_fd __rte_unused)
+{
+	int fd;
+	uint64_t mmap_size;
+	void *addr;
+	uint16_t num_queues, queue_size;
+	struct virtio_net *dev = *pdev;
+
+	if (msg->size != sizeof(msg->payload.inflight)) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Invalid get_inflight_fd message size is %d",
+			msg->size);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	num_queues = msg->payload.inflight.num_queues;
+	queue_size = msg->payload.inflight.queue_size;
+
+	RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd num_queues: %u\n",
+			msg->payload.inflight.num_queues);
+	RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd queue_size: %u\n",
+			msg->payload.inflight.queue_size);
+	mmap_size = num_queues * get_pervq_shm_size_split(queue_size);
+
+	addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
+	if (!addr) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Failed to alloc vhost inflight area");
+			msg->payload.inflight.mmap_size = 0;
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+	memset(addr, 0, mmap_size);
+
+	dev->inflight_info.addr = addr;
+	dev->inflight_info.size = msg->payload.inflight.mmap_size = mmap_size;
+	dev->inflight_info.fd = msg->fds[0] = fd;
+	msg->payload.inflight.mmap_offset = 0;
+	msg->fd_num = 1;
+
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight mmap_size: %lu\n",
+			msg->payload.inflight.mmap_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight mmap_offset: %lu\n",
+			msg->payload.inflight.mmap_offset);
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight fd: %d\n", msg->fds[0]);
+
+	return RTE_VHOST_MSG_RESULT_REPLY;
+}
+
+static int
+vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
+		int main_fd __rte_unused)
+{
+	int fd, i;
+	uint64_t mmap_size, mmap_offset;
+	uint16_t num_queues, queue_size;
+	uint32_t pervq_inflight_size;
+	void *addr;
+	struct vhost_virtqueue *vq;
+	struct virtio_net *dev = *pdev;
+
+	fd = msg->fds[0];
+	if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Invalid set_inflight_fd message size is %d,fd is %d\n",
+			msg->size, fd);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	mmap_size = msg->payload.inflight.mmap_size;
+	mmap_offset = msg->payload.inflight.mmap_offset;
+	num_queues = msg->payload.inflight.num_queues;
+	queue_size = msg->payload.inflight.queue_size;
+	pervq_inflight_size = get_pervq_shm_size_split(queue_size);
+
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd mmap_size: %lu\n", mmap_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd mmap_offset: %lu\n", mmap_offset);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd num_queues: %u\n", num_queues);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd queue_size: %u\n", queue_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd fd: %d\n", fd);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd pervq_inflight_size: %d\n",
+		pervq_inflight_size);
+
+	if (dev->inflight_info.addr)
+		munmap(dev->inflight_info.addr, dev->inflight_info.size);
+
+	addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
+			fd, mmap_offset);
+	if (addr == MAP_FAILED) {
+		RTE_LOG(ERR, VHOST_CONFIG, "failed to mmap share memory.\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	if (dev->inflight_info.fd)
+		close(dev->inflight_info.fd);
+
+	dev->inflight_info.fd = fd;
+	dev->inflight_info.addr = addr;
+	dev->inflight_info.size = mmap_size;
+
+	for (i = 0; i < num_queues; i++) {
+		vq = dev->virtqueue[i];
+		vq->inflight = (struct inflight_info_split *)addr;
+		vq->inflight->desc_num = queue_size;
+		addr = (void *)((char *)addr + pervq_inflight_size);
+	}
+
+	return RTE_VHOST_MSG_RESULT_OK;
+}
+
 static int
 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
 			int main_fd __rte_unused)
@@ -1201,6 +1402,89 @@ static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
 	return RTE_VHOST_MSG_RESULT_OK;
 }
 
+static int
+resubmit_desc_compare(const void *a, const void *b)
+{
+	const struct resubmit_desc *desc0 = (const struct resubmit_desc *)a;
+	const struct resubmit_desc *desc1 = (const struct resubmit_desc *)b;
+
+	if (desc1->counter > desc0->counter &&
+		(desc1->counter - desc0->counter) < VIRTQUEUE_MAX_SIZE * 2)
+		return 1;
+
+	return -1;
+}
+
+static int
+vhost_check_queue_inflights_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
+	struct vring_used *used = vq->used;
+	uint16_t i = 0;
+	uint16_t resubmit_num = 0;
+	struct resubmit_info *resubmit = NULL;
+
+	if (!(dev->features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
+		return RTE_VHOST_MSG_RESULT_OK;
+
+	if ((!vq->inflight))
+		return RTE_VHOST_MSG_RESULT_ERR;
+
+	if (!vq->inflight->version) {
+		vq->inflight->version = INFLIGHT_VERSION;
+		return RTE_VHOST_MSG_RESULT_OK;
+	}
+
+	vq->resubmit_inflight = NULL;
+	vq->counter = 0;
+
+	if (vq->inflight->used_idx != used->idx) {
+		vq->inflight->desc[vq->inflight->last_inflight_io].inflight = 0;
+		rte_compiler_barrier();
+		vq->inflight->used_idx = used->idx;
+	}
+
+	for (i = 0; i < vq->inflight->desc_num; i++) {
+		if (vq->inflight->desc[i].inflight == 1)
+			resubmit_num++;
+	}
+
+	vq->last_avail_idx += resubmit_num;
+
+	if (resubmit_num) {
+		resubmit  = calloc(1, sizeof(struct resubmit_info));
+		if (!resubmit) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Failed to allocate memory for resubmit info.\n");
+			return RTE_VHOST_MSG_RESULT_ERR;
+		}
+
+		resubmit->resubmit_list = calloc(resubmit_num,
+			sizeof(struct resubmit_desc));
+		if (!resubmit->resubmit_list) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Failed to allocate memory for inflight desc.\n");
+			return RTE_VHOST_MSG_RESULT_ERR;
+		}
+
+		for (i = 0; i < vq->inflight->desc_num; i++) {
+			if (vq->inflight->desc[i].inflight == 1) {
+				resubmit->resubmit_list[resubmit->resubmit_num].index = i;
+				resubmit->resubmit_list[resubmit->resubmit_num].counter =
+					vq->inflight->desc[i].counter;
+				resubmit->resubmit_num++;
+			}
+		}
+
+		if (resubmit->resubmit_num > 1)
+			qsort(resubmit->resubmit_list, resubmit->resubmit_num,
+				sizeof(struct resubmit_desc), resubmit_desc_compare);
+
+		vq->counter = resubmit->resubmit_list[0].counter + 1;
+		vq->resubmit_inflight = resubmit;
+	}
+
+	return RTE_VHOST_MSG_RESULT_OK;
+}
+
 static int
 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
 			int main_fd __rte_unused)
@@ -1242,6 +1526,12 @@ vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
 		close(vq->kickfd);
 	vq->kickfd = file.fd;
 
+	if (vhost_check_queue_inflights_split(dev, vq)) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Failed to inflights for vq: %d\n", file.index);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
 	return RTE_VHOST_MSG_RESULT_OK;
 }
 
@@ -1762,6 +2052,8 @@ static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
 	[VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
 	[VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
 	[VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
+	[VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
+	[VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
 };
 
 
diff --git a/lib/librte_vhost/vhost_user.h b/lib/librte_vhost/vhost_user.h
index 2a650fe4b..99a773910 100644
--- a/lib/librte_vhost/vhost_user.h
+++ b/lib/librte_vhost/vhost_user.h
@@ -54,7 +54,9 @@ typedef enum VhostUserRequest {
 	VHOST_USER_POSTCOPY_ADVISE = 28,
 	VHOST_USER_POSTCOPY_LISTEN = 29,
 	VHOST_USER_POSTCOPY_END = 30,
-	VHOST_USER_MAX = 31
+	VHOST_USER_GET_INFLIGHT_FD = 31,
+	VHOST_USER_SET_INFLIGHT_FD = 32,
+	VHOST_USER_MAX = 33
 } VhostUserRequest;
 
 typedef enum VhostUserSlaveRequest {
@@ -112,6 +114,13 @@ typedef struct VhostUserVringArea {
 	uint64_t offset;
 } VhostUserVringArea;
 
+typedef struct VhostUserInflight {
+	uint64_t mmap_size;
+	uint64_t mmap_offset;
+	uint16_t num_queues;
+	uint16_t queue_size;
+} VhostUserInflight;
+
 typedef struct VhostUserMsg {
 	union {
 		uint32_t master; /* a VhostUserRequest value */
@@ -131,6 +140,7 @@ typedef struct VhostUserMsg {
 		struct vhost_vring_addr addr;
 		VhostUserMemory memory;
 		VhostUserLog    log;
+		VhostUserInflight inflight;
 		struct vhost_iotlb_msg iotlb;
 		VhostUserCryptoSessionParam crypto_session;
 		VhostUserVringArea area;
@@ -148,6 +158,7 @@ typedef struct VhostUserMsg {
 /* vhost_user.c */
 int vhost_user_msg_handler(int vid, int fd);
 int vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm);
+void *inflight_mem_alloc(const char *name, size_t size, int *fd);
 
 /* socket.c */
 int read_fd_message(int sockfd, char *buf, int buflen, int *fds, int max_fds,
-- 
2.17.2

Re: [dpdk-dev] [PATCH] [v1]vhost: support inflight share memory protocol feature

[Maxime Coquelin]

Hi Jin,

On 6/26/19 6:38 PM, JinYu wrote:
> This patch introduces two new messages VHOST_USER_GET_INFLIGHT_FD
> and VHOST_USER_SET_INFLIGHT_FD to support transferring a shared
> buffer between qemu and backend.
> 
> Firstly, qemu uses VHOST_USER_GET_INFLIGHT_FD to get the
> shared buffer from backend. Then qemu should send it back
> through VHOST_USER_SET_INFLIGHT_FD each time we start vhost-user.
> 
> This shared buffer is used to process inflight I/O when backend
> reconnect.
> 
> Signed-off-by: LinLi <lilin24@baidu.com>
> Signed-off-by: XunNi <nixun@baidu.com>
> Signed-off-by: YuZhang <zhangyu31@baidu.com>
> Signed-off-by: JinYu <jin.yu@intel.com>
> ---
> V1 - specify the APIs are split-ring only
> ---
>   lib/librte_vhost/rte_vhost.h           |  99 +++++++++
>   lib/librte_vhost/rte_vhost_version.map |   4 +
>   lib/librte_vhost/vhost.c               | 135 ++++++++++++
>   lib/librte_vhost/vhost.h               |  12 +
>   lib/librte_vhost/vhost_user.c          | 292 +++++++++++++++++++++++++
>   lib/librte_vhost/vhost_user.h          |  13 +-
>   6 files changed, 554 insertions(+), 1 deletion(-)
> 
> diff --git a/lib/librte_vhost/rte_vhost.h b/lib/librte_vhost/rte_vhost.h
> index 0226b3eff..c9959c15a 100644
> --- a/lib/librte_vhost/rte_vhost.h
> +++ b/lib/librte_vhost/rte_vhost.h
> @@ -71,6 +71,10 @@ extern "C" {
>   #define VHOST_USER_PROTOCOL_F_HOST_NOTIFIER 11
>   #endif
>   
> +#ifndef VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD
> +#define VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD 12
> +#endif
> +
>   /** Indicate whether protocol features negotiation is supported. */
>   #ifndef VHOST_USER_F_PROTOCOL_FEATURES
>   #define VHOST_USER_F_PROTOCOL_FEATURES	30
> @@ -98,12 +102,41 @@ struct rte_vhost_memory {
>   	struct rte_vhost_mem_region regions[];
>   };
>   
> +struct inflight_desc_split {
> +	uint8_t		inflight;
> +	uint8_t		padding[5];
> +	uint16_t	next;
> +	uint64_t	counter;
> +};
> +
> +struct inflight_info_split {
> +	uint64_t		features;
> +	uint16_t		version;
> +	uint16_t		desc_num;
> +	uint16_t		last_inflight_io;
> +	uint16_t		used_idx;
> +	struct inflight_desc_split desc[0];
> +};
> +
> +struct resubmit_desc {
> +	uint16_t index;
> +	uint64_t counter;
> +};
> +
> +struct resubmit_info {
> +	struct resubmit_desc	*resubmit_list;
> +	uint16_t		resubmit_num;
> +};
> +
>   struct rte_vhost_vring {
>   	struct vring_desc	*desc;
>   	struct vring_avail	*avail;
>   	struct vring_used	*used;
>   	uint64_t		log_guest_addr;
>   
> +	struct inflight_info_split *inflight;
> +	struct resubmit_info	*resubmit_inflight;
> +

This is breaking the ABI, and discussing with Tiwei, we think it could
be avoided by defining its own structure of inflight and inflight
resubmit pointers. More below:

>   	/** Deprecated, use rte_vhost_vring_call() instead. */
>   	int			callfd;
>   
> @@ -603,6 +636,22 @@ uint16_t rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
>    */
>   int rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem);
>   
> +/**
> + * Get guest vring info, including the vring address, vring size, inflight, etc.
> + *
> + * @param vid
> + *  vhost device ID
> + * @param vring_idx
> + *  vring index
> + * @param vring
> + *  the structure to hold the requested vring info
> + * @return
> + *  0 on success, -1 on failure
> + */
> +int __rte_experimental
> +rte_vhost_get_vhost_vring_with_inflight_split(int vid, uint16_t vring_idx,
> +			      struct rte_vhost_vring *vring);

Here a new API is introduced to get vting info with inflight.
It could be just changed to get only the inflight info. It means the
caller would first call rte_vhost_get_vhost_vring() and then the new
function, which I guess may be renamed to:

rte_vhost_get_vhost_vring_inflight_split

Other than that, the patch looks good to me.

(Next time, please run get_maintainer.sh script so that all the
maintainers are in Cc:).

Thanks,
Maxime

Re: [dpdk-dev] [PATCH] [v1]vhost: support inflight share memory protocol feature

[Yu, Jin]

> From: Maxime Coquelin [mailto:maxime.coquelin@redhat.com]
> Sent: Friday, July 5, 2019 4:47 PM
> To: Yu, Jin <jin.yu@intel.com>; dev@dpdk.org
> Cc: Liu, Changpeng <changpeng.liu@intel.com>; LinLi <lilin24@baidu.com>;
> XunNi <nixun@baidu.com>; YuZhang <zhangyu31@baidu.com>; Bie, Tiwei
> <tiwei.bie@intel.com>
> Subject: Re: [dpdk-dev] [PATCH] [v1]vhost: support inflight share memory
> protocol feature
> 
> Hi Jin,
> 
> On 6/26/19 6:38 PM, JinYu wrote:
> > This patch introduces two new messages VHOST_USER_GET_INFLIGHT_FD
> and
> > VHOST_USER_SET_INFLIGHT_FD to support transferring a shared buffer
> > between qemu and backend.
> >
> > Firstly, qemu uses VHOST_USER_GET_INFLIGHT_FD to get the shared buffer
> > from backend. Then qemu should send it back through
> > VHOST_USER_SET_INFLIGHT_FD each time we start vhost-user.
> >
> > This shared buffer is used to process inflight I/O when backend
> > reconnect.
> >
> > Signed-off-by: LinLi <lilin24@baidu.com>
> > Signed-off-by: XunNi <nixun@baidu.com>
> > Signed-off-by: YuZhang <zhangyu31@baidu.com>
> > Signed-off-by: JinYu <jin.yu@intel.com>
> > ---
> > V1 - specify the APIs are split-ring only
> > ---
> >   lib/librte_vhost/rte_vhost.h           |  99 +++++++++
> >   lib/librte_vhost/rte_vhost_version.map |   4 +
> >   lib/librte_vhost/vhost.c               | 135 ++++++++++++
> >   lib/librte_vhost/vhost.h               |  12 +
> >   lib/librte_vhost/vhost_user.c          | 292 +++++++++++++++++++++++++
> >   lib/librte_vhost/vhost_user.h          |  13 +-
> >   6 files changed, 554 insertions(+), 1 deletion(-)
> >
> > diff --git a/lib/librte_vhost/rte_vhost.h
> > b/lib/librte_vhost/rte_vhost.h index 0226b3eff..c9959c15a 100644
> > --- a/lib/librte_vhost/rte_vhost.h
> > +++ b/lib/librte_vhost/rte_vhost.h
> > @@ -71,6 +71,10 @@ extern "C" {
> >   #define VHOST_USER_PROTOCOL_F_HOST_NOTIFIER 11
> >   #endif
> >
> > +#ifndef VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD
> > +#define VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD 12 #endif
> > +
> >   /** Indicate whether protocol features negotiation is supported. */
> >   #ifndef VHOST_USER_F_PROTOCOL_FEATURES
> >   #define VHOST_USER_F_PROTOCOL_FEATURES	30
> > @@ -98,12 +102,41 @@ struct rte_vhost_memory {
> >   	struct rte_vhost_mem_region regions[];
> >   };
> >
> > +struct inflight_desc_split {
> > +	uint8_t		inflight;
> > +	uint8_t		padding[5];
> > +	uint16_t	next;
> > +	uint64_t	counter;
> > +};
> > +
> > +struct inflight_info_split {
> > +	uint64_t		features;
> > +	uint16_t		version;
> > +	uint16_t		desc_num;
> > +	uint16_t		last_inflight_io;
> > +	uint16_t		used_idx;
> > +	struct inflight_desc_split desc[0];
> > +};
> > +
> > +struct resubmit_desc {
> > +	uint16_t index;
> > +	uint64_t counter;
> > +};
> > +
> > +struct resubmit_info {
> > +	struct resubmit_desc	*resubmit_list;
> > +	uint16_t		resubmit_num;
> > +};
> > +
> >   struct rte_vhost_vring {
> >   	struct vring_desc	*desc;
> >   	struct vring_avail	*avail;
> >   	struct vring_used	*used;
> >   	uint64_t		log_guest_addr;
> >
> > +	struct inflight_info_split *inflight;
> > +	struct resubmit_info	*resubmit_inflight;
> > +
> 
> This is breaking the ABI, and discussing with Tiwei, we think it could be
> avoided by defining its own structure of inflight and inflight resubmit
> pointers. More below:
So you means that define a new struct such as struct rte_vhost_inflight_ring?
> 
> >   	/** Deprecated, use rte_vhost_vring_call() instead. */
> >   	int			callfd;
> >
> > @@ -603,6 +636,22 @@ uint16_t rte_vhost_dequeue_burst(int vid,
> uint16_t queue_id,
> >    */
> >   int rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem);
> >
> > +/**
> > + * Get guest vring info, including the vring address, vring size, inflight, etc.
> > + *
> > + * @param vid
> > + *  vhost device ID
> > + * @param vring_idx
> > + *  vring index
> > + * @param vring
> > + *  the structure to hold the requested vring info
> > + * @return
> > + *  0 on success, -1 on failure
> > + */
> > +int __rte_experimental
> > +rte_vhost_get_vhost_vring_with_inflight_split(int vid, uint16_t vring_idx,
> > +			      struct rte_vhost_vring *vring);
> 
> Here a new API is introduced to get vting info with inflight.
> It could be just changed to get only the inflight info. It means the caller would
> first call rte_vhost_get_vhost_vring() and then the new function, which I
> guess may be renamed to:
> 
> rte_vhost_get_vhost_vring_inflight_split
Got It.
> 
> Other than that, the patch looks good to me.
> 
> (Next time, please run get_maintainer.sh script so that all the maintainers
> are in Cc:).
> 
> Thanks,
> Maxime
Thanks Maxime, also I will submit an example code of vhost_user_blk  which support the inflight feature.
It can show how these APIs work.

Thanks
Jin

[dpdk-dev] [PATCH] [v2, 1/2]vhost: support inflight share memory protocol feature

[JinYu]

This patch introduces two new messages VHOST_USER_GET_INFLIGHT_FD
and VHOST_USER_SET_INFLIGHT_FD to support transferring a shared
buffer between qemu and backend.

Firstly, qemu uses VHOST_USER_GET_INFLIGHT_FD to get the
shared buffer from backend. Then qemu should send it back
through VHOST_USER_SET_INFLIGHT_FD each time we start vhost-user.

This shared buffer is used to process inflight I/O when backend
reconnect.

Signed-off-by: LinLi <lilin24@baidu.com>
Signed-off-by: XunNi <nixun@baidu.com>
Signed-off-by: YuZhang <zhangyu31@baidu.com>
Signed-off-by: JinYu <jin.yu@intel.com>
---
v1 - specify the APIs are split-ring only
v2 - fix APIs and judge split or packed
---
 lib/librte_vhost/rte_vhost.h           | 105 +++++++++
 lib/librte_vhost/rte_vhost_version.map |   4 +
 lib/librte_vhost/vhost.c               | 158 ++++++++++++-
 lib/librte_vhost/vhost.h               |  16 ++
 lib/librte_vhost/vhost_user.c          | 313 +++++++++++++++++++++++++
 lib/librte_vhost/vhost_user.h          |  13 +-
 6 files changed, 607 insertions(+), 2 deletions(-)

diff --git a/lib/librte_vhost/rte_vhost.h b/lib/librte_vhost/rte_vhost.h
index 0226b3eff..a17aff876 100644
--- a/lib/librte_vhost/rte_vhost.h
+++ b/lib/librte_vhost/rte_vhost.h
@@ -71,6 +71,10 @@ extern "C" {
 #define VHOST_USER_PROTOCOL_F_HOST_NOTIFIER 11
 #endif
 
+#ifndef VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD
+#define VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD 12
+#endif
+
 /** Indicate whether protocol features negotiation is supported. */
 #ifndef VHOST_USER_F_PROTOCOL_FEATURES
 #define VHOST_USER_F_PROTOCOL_FEATURES	30
@@ -98,6 +102,41 @@ struct rte_vhost_memory {
 	struct rte_vhost_mem_region regions[];
 };
 
+struct inflight_desc_split {
+	uint8_t		inflight;
+	uint8_t		padding[5];
+	uint16_t	next;
+	uint64_t	counter;
+};
+
+struct inflight_info_split {
+	uint64_t		features;
+	uint16_t		version;
+	uint16_t		desc_num;
+	uint16_t		last_inflight_io;
+	uint16_t		used_idx;
+	struct inflight_desc_split desc[0];
+};
+
+struct resubmit_desc {
+	uint16_t index;
+	uint64_t counter;
+};
+
+struct resubmit_info {
+	struct resubmit_desc	*resubmit_list;
+	uint16_t		resubmit_num;
+};
+
+struct rte_vhost_ring_inflight_split {
+	union {
+		struct inflight_info_split *inflight_split;
+		/* TODO */
+	};
+
+	struct resubmit_info *resubmit_inflight_split;
+};
+
 struct rte_vhost_vring {
 	struct vring_desc	*desc;
 	struct vring_avail	*avail;
@@ -603,6 +642,22 @@ uint16_t rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
  */
 int rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem);
 
+/**
+ * Get guest inflight vring info, including inflight ring and resubmit list.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param vring
+ *  the structure to hold the requested inflight vring info
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_get_vhost_ring_inflight_split(int vid, uint16_t vring_idx,
+			      struct rte_vhost_ring_inflight_split *vring);
+
 /**
  * Get guest vring info, including the vring address, vring size, etc.
  *
@@ -631,6 +686,56 @@ int rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
  */
 int rte_vhost_vring_call(int vid, uint16_t vring_idx);
 
+/**
+ * set inflight flag for a desc.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
+		uint16_t idx);
+
+/**
+ * clear inflight flag for a desc.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param last_used_idx
+ *  next free used_idx
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
+		uint16_t last_used_idx, uint16_t idx);
+
+/**
+ * set last inflight io index.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
+		uint16_t idx);
+
 /**
  * Get vhost RX queue avail count.
  *
diff --git a/lib/librte_vhost/rte_vhost_version.map b/lib/librte_vhost/rte_vhost_version.map
index 5f1d4a75c..709593701 100644
--- a/lib/librte_vhost/rte_vhost_version.map
+++ b/lib/librte_vhost/rte_vhost_version.map
@@ -87,4 +87,8 @@ EXPERIMENTAL {
 	rte_vdpa_relay_vring_used;
 	rte_vhost_extern_callback_register;
 	rte_vhost_driver_set_protocol_features;
+	rte_vhost_set_inflight_desc_split;
+	rte_vhost_clr_inflight_desc_split;
+	rte_vhost_set_last_inflight_io_split;
+	rte_vhost_get_vhost_vring_with_inflight_split;
 };
diff --git a/lib/librte_vhost/vhost.c b/lib/librte_vhost/vhost.c
index 981837b5d..ecd580cc5 100644
--- a/lib/librte_vhost/vhost.c
+++ b/lib/librte_vhost/vhost.c
@@ -242,6 +242,30 @@ cleanup_vq(struct vhost_virtqueue *vq, int destroy)
 		close(vq->kickfd);
 }
 
+void
+cleanup_vq_inflight_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
+	if (!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
+		return;
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG, "function only be used for split inflight.\n");
+	} else {
+		if (vq->inflight_split)
+			vq->inflight_split = NULL;
+		if (vq->resubmit_inflight_split) {
+			if (vq->resubmit_inflight_split->resubmit_list) {
+				free(vq->resubmit_inflight_split->resubmit_list);
+				vq->resubmit_inflight_split->resubmit_list = NULL;
+			}
+			free(vq->resubmit_inflight_split);
+			vq->resubmit_inflight_split = NULL;
+		}
+	}
+}
+
 /*
  * Unmap any memory, close any file descriptors and
  * free any memory owned by a device.
@@ -253,8 +277,10 @@ cleanup_device(struct virtio_net *dev, int destroy)
 
 	vhost_backend_cleanup(dev);
 
-	for (i = 0; i < dev->nr_vring; i++)
+	for (i = 0; i < dev->nr_vring; i++) {
 		cleanup_vq(dev->virtqueue[i], destroy);
+		cleanup_vq_inflight_split(dev, dev->virtqueue[i]);
+	}
 }
 
 void
@@ -726,6 +752,35 @@ rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
 	return 0;
 }
 
+int __rte_experimental
+rte_vhost_get_vhost_ring_inflight_split(int vid, uint16_t vring_idx,
+			  struct rte_vhost_ring_inflight_split *vring)
+{
+	struct virtio_net *dev;
+	struct inflight_info_split *inflight_vring;
+
+	dev = get_device(vid);
+	if (dev == NULL || vring == NULL)
+		return -1;
+
+	if (vring_idx >= VHOST_MAX_VRING)
+		return -1;
+
+	inflight_vring = dev->virtqueue[vring_idx]->inflight_split;
+	if (!inflight_vring)
+		return -1;
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG, "Can't be packed inflight ring.\n");
+	} else {
+		vring->inflight_split = inflight_vring;
+		vring->resubmit_inflight_split = dev->virtqueue[vring_idx]->resubmit_inflight_split;
+	}
+
+	return 0;
+}
+
 int
 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
 			  struct rte_vhost_vring *vring)
@@ -781,6 +836,107 @@ rte_vhost_vring_call(int vid, uint16_t vring_idx)
 	return 0;
 }
 
+int
+rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vq_is_packed(dev)))
+		return -1;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (unlikely(!vq))
+		return -1;
+
+	if (unlikely(!vq->inflight_split))
+		return -1;
+
+	vq->inflight_split->desc[idx].counter = vq->global_counter++;
+	vq->inflight_split->desc[idx].inflight = 1;
+	return 0;
+}
+
+int
+rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
+	uint16_t last_used_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vq_is_packed(dev)))
+		return -1;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (unlikely(!vq))
+		return -1;
+
+	if (unlikely(!vq->inflight_split))
+		return -1;
+
+	rte_compiler_barrier();
+
+	vq->inflight_split->desc[idx].inflight = 0;
+
+	rte_compiler_barrier();
+
+	vq->inflight_split->used_idx = last_used_idx;
+	return 0;
+}
+
+int
+rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vq_is_packed(dev)))
+		return -1;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (!vq)
+		return -1;
+
+	if (unlikely(!vq->inflight_split))
+		return -1;
+
+	vq->inflight_split->last_inflight_io = idx;
+	return 0;
+}
+
 uint16_t
 rte_vhost_avail_entries(int vid, uint16_t queue_id)
 {
diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
index 884befa85..22925586d 100644
--- a/lib/librte_vhost/vhost.h
+++ b/lib/librte_vhost/vhost.h
@@ -128,6 +128,14 @@ struct vhost_virtqueue {
 	/* Physical address of used ring, for logging */
 	uint64_t		log_guest_addr;
 
+	/* inflight share memory info */
+	union {
+		struct inflight_info_split *inflight_split;
+		/* TODO */
+	};
+	struct resubmit_info *resubmit_inflight_split;
+	uint64_t		global_counter;
+
 	uint16_t		nr_zmbuf;
 	uint16_t		zmbuf_size;
 	uint16_t		last_zmbuf_idx;
@@ -286,6 +294,12 @@ struct guest_page {
 	uint64_t size;
 };
 
+struct inflight_mem_info {
+	int		fd;
+	void		*addr;
+	uint64_t	size;
+};
+
 /**
  * Device structure contains all configuration information relating
  * to the device.
@@ -303,6 +317,7 @@ struct virtio_net {
 	uint32_t		nr_vring;
 	int			dequeue_zero_copy;
 	struct vhost_virtqueue	*virtqueue[VHOST_MAX_QUEUE_PAIRS * 2];
+	struct inflight_mem_info inflight_info;
 #define IF_NAME_SZ (PATH_MAX > IFNAMSIZ ? PATH_MAX : IFNAMSIZ)
 	char			ifname[IF_NAME_SZ];
 	uint64_t		log_size;
@@ -467,6 +482,7 @@ void vhost_destroy_device(int);
 void vhost_destroy_device_notify(struct virtio_net *dev);
 
 void cleanup_vq(struct vhost_virtqueue *vq, int destroy);
+void cleanup_vq_inflight_split(struct virtio_net *dev, struct vhost_virtqueue *vq);
 void free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq);
 
 int alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx);
diff --git a/lib/librte_vhost/vhost_user.c b/lib/librte_vhost/vhost_user.c
index c9e29ece8..feb642ad7 100644
--- a/lib/librte_vhost/vhost_user.c
+++ b/lib/librte_vhost/vhost_user.c
@@ -31,6 +31,8 @@
 #include <sys/stat.h>
 #include <sys/syscall.h>
 #include <assert.h>
+#include <sys/syscall.h>
+#include <asm/unistd.h>
 #ifdef RTE_LIBRTE_VHOST_NUMA
 #include <numaif.h>
 #endif
@@ -49,6 +51,15 @@
 #define VIRTIO_MIN_MTU 68
 #define VIRTIO_MAX_MTU 65535
 
+#define INFLIGHT_ALIGNMENT	64
+#define INFLIGHT_VERSION	0xabcd
+#define VIRTQUEUE_MAX_SIZE	1024
+
+#define CLOEXEC	0x0001U
+
+#define ALIGN_DOWN(n, m) ((n) / (m) * (m))
+#define ALIGN_UP(n, m) ALIGN_DOWN((n) + (m) - 1, (m))
+
 static const char *vhost_message_str[VHOST_USER_MAX] = {
 	[VHOST_USER_NONE] = "VHOST_USER_NONE",
 	[VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
@@ -78,6 +89,8 @@ static const char *vhost_message_str[VHOST_USER_MAX] = {
 	[VHOST_USER_POSTCOPY_ADVISE]  = "VHOST_USER_POSTCOPY_ADVISE",
 	[VHOST_USER_POSTCOPY_LISTEN]  = "VHOST_USER_POSTCOPY_LISTEN",
 	[VHOST_USER_POSTCOPY_END]  = "VHOST_USER_POSTCOPY_END",
+	[VHOST_USER_GET_INFLIGHT_FD] = "VHOST_USER_GET_INFLIGHT_FD",
+	[VHOST_USER_SET_INFLIGHT_FD] = "VHOST_USER_SET_INFLIGHT_FD",
 };
 
 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
@@ -160,6 +173,16 @@ vhost_backend_cleanup(struct virtio_net *dev)
 		dev->log_addr = 0;
 	}
 
+	if (dev->inflight_info.addr) {
+		munmap(dev->inflight_info.addr, dev->inflight_info.size);
+		dev->inflight_info.addr = NULL;
+	}
+
+	if (dev->inflight_info.fd > 0) {
+		close(dev->inflight_info.fd);
+		dev->inflight_info.fd = -1;
+	}
+
 	if (dev->slave_req_fd >= 0) {
 		close(dev->slave_req_fd);
 		dev->slave_req_fd = -1;
@@ -306,6 +329,7 @@ vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 
 			dev->virtqueue[dev->nr_vring] = NULL;
 			cleanup_vq(vq, 1);
+			cleanup_vq_inflight_split(dev, vq);
 			free_vq(dev, vq);
 		}
 	}
@@ -1165,6 +1189,204 @@ virtio_is_ready(struct virtio_net *dev)
 	return 1;
 }
 
+static int mem_create(const char *name, unsigned int flags)
+{
+#ifdef __NR_memfd_create
+	return syscall(__NR_memfd_create, name, flags);
+#else
+	return -1;
+#endif
+}
+
+void *inflight_mem_alloc(const char *name, size_t size, int *fd)
+{
+	void *ptr;
+	int mfd = -1;
+	char fname[20] = "/tmp/memfd-XXXXXX";
+
+	*fd = -1;
+	mfd = mem_create(name, CLOEXEC);
+	if (mfd != -1) {
+		if (ftruncate(mfd, size) == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"ftruncate fail for alloc inflight buffer\n");
+			close(mfd);
+			return NULL;
+		}
+	} else {
+		mfd = mkstemp(fname);
+		unlink(fname);
+
+		if (mfd == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"mkstemp fail for alloc inflight buffer\n");
+			return NULL;
+		}
+
+		if (ftruncate(mfd, size) == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"ftruncate fail for alloc inflight buffer\n");
+			close(mfd);
+			return NULL;
+		}
+	}
+
+	ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
+	if (ptr == MAP_FAILED) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+				"mmap fail for alloc inflight buffer\n");
+		close(mfd);
+		return NULL;
+	}
+
+	*fd = mfd;
+	return ptr;
+}
+
+static uint32_t get_pervq_shm_size_split(uint16_t queue_size)
+{
+	return ALIGN_UP(sizeof(struct inflight_desc_split) * queue_size +
+		sizeof(uint64_t) * 1 + sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
+}
+
+static int
+vhost_user_get_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
+		int main_fd __rte_unused)
+{
+	int fd;
+	uint64_t mmap_size;
+	void *addr;
+	uint16_t num_queues, queue_size;
+	struct virtio_net *dev = *pdev;
+
+	if (msg->size != sizeof(msg->payload.inflight)) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Invalid get_inflight_fd message size is %d",
+			msg->size);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	num_queues = msg->payload.inflight.num_queues;
+	queue_size = msg->payload.inflight.queue_size;
+
+	RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd num_queues: %u\n",
+			msg->payload.inflight.num_queues);
+	RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd queue_size: %u\n",
+			msg->payload.inflight.queue_size);
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Don't support the packed ring\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	} else {
+		mmap_size = num_queues * get_pervq_shm_size_split(queue_size);
+	}
+
+	addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
+	if (!addr) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Failed to alloc vhost inflight area");
+			msg->payload.inflight.mmap_size = 0;
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+	memset(addr, 0, mmap_size);
+
+	dev->inflight_info.addr = addr;
+	dev->inflight_info.size = msg->payload.inflight.mmap_size = mmap_size;
+	dev->inflight_info.fd = msg->fds[0] = fd;
+	msg->payload.inflight.mmap_offset = 0;
+	msg->fd_num = 1;
+
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight mmap_size: %lu\n",
+			msg->payload.inflight.mmap_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight mmap_offset: %lu\n",
+			msg->payload.inflight.mmap_offset);
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight fd: %d\n", msg->fds[0]);
+
+	return RTE_VHOST_MSG_RESULT_REPLY;
+}
+
+static int
+vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
+		int main_fd __rte_unused)
+{
+	int fd, i;
+	uint64_t mmap_size, mmap_offset;
+	uint16_t num_queues, queue_size;
+	uint32_t pervq_inflight_size;
+	void *addr;
+	struct vhost_virtqueue *vq;
+	struct virtio_net *dev = *pdev;
+
+	fd = msg->fds[0];
+	if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Invalid set_inflight_fd message size is %d,fd is %d\n",
+			msg->size, fd);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	mmap_size = msg->payload.inflight.mmap_size;
+	mmap_offset = msg->payload.inflight.mmap_offset;
+	num_queues = msg->payload.inflight.num_queues;
+	queue_size = msg->payload.inflight.queue_size;
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Don't support the packed ring\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	} else {
+		pervq_inflight_size = get_pervq_shm_size_split(queue_size);
+	}
+
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd mmap_size: %lu\n", mmap_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd mmap_offset: %lu\n", mmap_offset);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd num_queues: %u\n", num_queues);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd queue_size: %u\n", queue_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd fd: %d\n", fd);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd pervq_inflight_size: %d\n",
+		pervq_inflight_size);
+
+	if (dev->inflight_info.addr)
+		munmap(dev->inflight_info.addr, dev->inflight_info.size);
+
+	addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
+			fd, mmap_offset);
+	if (addr == MAP_FAILED) {
+		RTE_LOG(ERR, VHOST_CONFIG, "failed to mmap share memory.\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	if (dev->inflight_info.fd)
+		close(dev->inflight_info.fd);
+
+	dev->inflight_info.fd = fd;
+	dev->inflight_info.addr = addr;
+	dev->inflight_info.size = mmap_size;
+
+	for (i = 0; i < num_queues; i++) {
+		vq = dev->virtqueue[i];
+		if (vq_is_packed(dev)) {
+			/* TODO */
+		} else {
+			vq->inflight_split = (struct inflight_info_split *)addr;
+			vq->inflight_split->desc_num = queue_size;
+		}
+		addr = (void *)((char *)addr + pervq_inflight_size);
+	}
+
+	return RTE_VHOST_MSG_RESULT_OK;
+}
+
 static int
 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
 			int main_fd __rte_unused)
@@ -1201,6 +1423,89 @@ static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
 	return RTE_VHOST_MSG_RESULT_OK;
 }
 
+static int
+resubmit_desc_compare(const void *a, const void *b)
+{
+	const struct resubmit_desc *desc0 = (const struct resubmit_desc *)a;
+	const struct resubmit_desc *desc1 = (const struct resubmit_desc *)b;
+
+	if (desc1->counter > desc0->counter &&
+		(desc1->counter - desc0->counter) < VIRTQUEUE_MAX_SIZE * 2)
+		return 1;
+
+	return -1;
+}
+
+static int
+vhost_check_queue_inflights_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
+	struct vring_used *used = vq->used;
+	uint16_t i = 0;
+	uint16_t resubmit_num = 0;
+	struct resubmit_info *resubmit = NULL;
+
+	if (!(dev->features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
+		return RTE_VHOST_MSG_RESULT_OK;
+
+	if ((!vq->inflight_split))
+		return RTE_VHOST_MSG_RESULT_ERR;
+
+	if (!vq->inflight_split->version) {
+		vq->inflight_split->version = INFLIGHT_VERSION;
+		return RTE_VHOST_MSG_RESULT_OK;
+	}
+
+	vq->resubmit_inflight_split = NULL;
+	vq->global_counter = 0;
+
+	if (vq->inflight_split->used_idx != used->idx) {
+		vq->inflight_split->desc[vq->inflight_split->last_inflight_io].inflight = 0;
+		rte_compiler_barrier();
+		vq->inflight_split->used_idx = used->idx;
+	}
+
+	for (i = 0; i < vq->inflight_split->desc_num; i++) {
+		if (vq->inflight_split->desc[i].inflight == 1)
+			resubmit_num++;
+	}
+
+	vq->last_avail_idx += resubmit_num;
+
+	if (resubmit_num) {
+		resubmit  = calloc(1, sizeof(struct resubmit_info));
+		if (!resubmit) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Failed to allocate memory for resubmit info.\n");
+			return RTE_VHOST_MSG_RESULT_ERR;
+		}
+
+		resubmit->resubmit_list = calloc(resubmit_num,
+			sizeof(struct resubmit_desc));
+		if (!resubmit->resubmit_list) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Failed to allocate memory for inflight desc.\n");
+			return RTE_VHOST_MSG_RESULT_ERR;
+		}
+
+		for (i = 0; i < vq->inflight_split->desc_num; i++) {
+			if (vq->inflight_split->desc[i].inflight == 1) {
+				resubmit->resubmit_list[resubmit->resubmit_num].index = i;
+				resubmit->resubmit_list[resubmit->resubmit_num].counter =
+					vq->inflight_split->desc[i].counter;
+				resubmit->resubmit_num++;
+			}
+		}
+
+		if (resubmit->resubmit_num > 1)
+			qsort(resubmit->resubmit_list, resubmit->resubmit_num,
+				sizeof(struct resubmit_desc), resubmit_desc_compare);
+
+		vq->global_counter = resubmit->resubmit_list[0].counter + 1;
+		vq->resubmit_inflight_split = resubmit;
+	}
+
+	return RTE_VHOST_MSG_RESULT_OK;
+}
+
 static int
 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
 			int main_fd __rte_unused)
@@ -1242,6 +1547,12 @@ vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
 		close(vq->kickfd);
 	vq->kickfd = file.fd;
 
+	if (vhost_check_queue_inflights_split(dev, vq)) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Failed to inflights for vq: %d\n", file.index);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
 	return RTE_VHOST_MSG_RESULT_OK;
 }
 
@@ -1762,6 +2073,8 @@ static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
 	[VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
 	[VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
 	[VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
+	[VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
+	[VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
 };
 
 
diff --git a/lib/librte_vhost/vhost_user.h b/lib/librte_vhost/vhost_user.h
index 2a650fe4b..99a773910 100644
--- a/lib/librte_vhost/vhost_user.h
+++ b/lib/librte_vhost/vhost_user.h
@@ -54,7 +54,9 @@ typedef enum VhostUserRequest {
 	VHOST_USER_POSTCOPY_ADVISE = 28,
 	VHOST_USER_POSTCOPY_LISTEN = 29,
 	VHOST_USER_POSTCOPY_END = 30,
-	VHOST_USER_MAX = 31
+	VHOST_USER_GET_INFLIGHT_FD = 31,
+	VHOST_USER_SET_INFLIGHT_FD = 32,
+	VHOST_USER_MAX = 33
 } VhostUserRequest;
 
 typedef enum VhostUserSlaveRequest {
@@ -112,6 +114,13 @@ typedef struct VhostUserVringArea {
 	uint64_t offset;
 } VhostUserVringArea;
 
+typedef struct VhostUserInflight {
+	uint64_t mmap_size;
+	uint64_t mmap_offset;
+	uint16_t num_queues;
+	uint16_t queue_size;
+} VhostUserInflight;
+
 typedef struct VhostUserMsg {
 	union {
 		uint32_t master; /* a VhostUserRequest value */
@@ -131,6 +140,7 @@ typedef struct VhostUserMsg {
 		struct vhost_vring_addr addr;
 		VhostUserMemory memory;
 		VhostUserLog    log;
+		VhostUserInflight inflight;
 		struct vhost_iotlb_msg iotlb;
 		VhostUserCryptoSessionParam crypto_session;
 		VhostUserVringArea area;
@@ -148,6 +158,7 @@ typedef struct VhostUserMsg {
 /* vhost_user.c */
 int vhost_user_msg_handler(int vid, int fd);
 int vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm);
+void *inflight_mem_alloc(const char *name, size_t size, int *fd);
 
 /* socket.c */
 int read_fd_message(int sockfd, char *buf, int buflen, int *fds, int max_fds,
-- 
2.17.2

[dpdk-dev] [PATCH] [v2, 1/2]vhost: support inflight share memory protocol feature

[JinYu]

This patch introduces two new messages VHOST_USER_GET_INFLIGHT_FD
and VHOST_USER_SET_INFLIGHT_FD to support transferring a shared
buffer between qemu and backend.

Firstly, qemu uses VHOST_USER_GET_INFLIGHT_FD to get the
shared buffer from backend. Then qemu should send it back
through VHOST_USER_SET_INFLIGHT_FD each time we start vhost-user.

This shared buffer is used to process inflight I/O when backend
reconnect.

Signed-off-by: LinLi <lilin24@baidu.com>
Signed-off-by: XunNi <nixun@baidu.com>
Signed-off-by: YuZhang <zhangyu31@baidu.com>
Signed-off-by: JinYu <jin.yu@intel.com>
---
v1 - specify the APIs are split-ring only
v2 - fix APIs and judge split or packed
---
 lib/librte_vhost/rte_vhost.h           | 105 +++++++++
 lib/librte_vhost/rte_vhost_version.map |   4 +
 lib/librte_vhost/vhost.c               | 158 ++++++++++++-
 lib/librte_vhost/vhost.h               |  16 ++
 lib/librte_vhost/vhost_user.c          | 313 +++++++++++++++++++++++++
 lib/librte_vhost/vhost_user.h          |  13 +-
 6 files changed, 607 insertions(+), 2 deletions(-)

diff --git a/lib/librte_vhost/rte_vhost.h b/lib/librte_vhost/rte_vhost.h
index 0226b3eff..a17aff876 100644
--- a/lib/librte_vhost/rte_vhost.h
+++ b/lib/librte_vhost/rte_vhost.h
@@ -71,6 +71,10 @@ extern "C" {
 #define VHOST_USER_PROTOCOL_F_HOST_NOTIFIER 11
 #endif
 
+#ifndef VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD
+#define VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD 12
+#endif
+
 /** Indicate whether protocol features negotiation is supported. */
 #ifndef VHOST_USER_F_PROTOCOL_FEATURES
 #define VHOST_USER_F_PROTOCOL_FEATURES	30
@@ -98,6 +102,41 @@ struct rte_vhost_memory {
 	struct rte_vhost_mem_region regions[];
 };
 
+struct inflight_desc_split {
+	uint8_t		inflight;
+	uint8_t		padding[5];
+	uint16_t	next;
+	uint64_t	counter;
+};
+
+struct inflight_info_split {
+	uint64_t		features;
+	uint16_t		version;
+	uint16_t		desc_num;
+	uint16_t		last_inflight_io;
+	uint16_t		used_idx;
+	struct inflight_desc_split desc[0];
+};
+
+struct resubmit_desc {
+	uint16_t index;
+	uint64_t counter;
+};
+
+struct resubmit_info {
+	struct resubmit_desc	*resubmit_list;
+	uint16_t		resubmit_num;
+};
+
+struct rte_vhost_ring_inflight_split {
+	union {
+		struct inflight_info_split *inflight_split;
+		/* TODO */
+	};
+
+	struct resubmit_info *resubmit_inflight_split;
+};
+
 struct rte_vhost_vring {
 	struct vring_desc	*desc;
 	struct vring_avail	*avail;
@@ -603,6 +642,22 @@ uint16_t rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
  */
 int rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem);
 
+/**
+ * Get guest inflight vring info, including inflight ring and resubmit list.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param vring
+ *  the structure to hold the requested inflight vring info
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_get_vhost_ring_inflight_split(int vid, uint16_t vring_idx,
+			      struct rte_vhost_ring_inflight_split *vring);
+
 /**
  * Get guest vring info, including the vring address, vring size, etc.
  *
@@ -631,6 +686,56 @@ int rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
  */
 int rte_vhost_vring_call(int vid, uint16_t vring_idx);
 
+/**
+ * set inflight flag for a desc.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
+		uint16_t idx);
+
+/**
+ * clear inflight flag for a desc.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param last_used_idx
+ *  next free used_idx
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
+		uint16_t last_used_idx, uint16_t idx);
+
+/**
+ * set last inflight io index.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
+		uint16_t idx);
+
 /**
  * Get vhost RX queue avail count.
  *
diff --git a/lib/librte_vhost/rte_vhost_version.map b/lib/librte_vhost/rte_vhost_version.map
index 5f1d4a75c..709593701 100644
--- a/lib/librte_vhost/rte_vhost_version.map
+++ b/lib/librte_vhost/rte_vhost_version.map
@@ -87,4 +87,8 @@ EXPERIMENTAL {
 	rte_vdpa_relay_vring_used;
 	rte_vhost_extern_callback_register;
 	rte_vhost_driver_set_protocol_features;
+	rte_vhost_set_inflight_desc_split;
+	rte_vhost_clr_inflight_desc_split;
+	rte_vhost_set_last_inflight_io_split;
+	rte_vhost_get_vhost_vring_with_inflight_split;
 };
diff --git a/lib/librte_vhost/vhost.c b/lib/librte_vhost/vhost.c
index 981837b5d..ecd580cc5 100644
--- a/lib/librte_vhost/vhost.c
+++ b/lib/librte_vhost/vhost.c
@@ -242,6 +242,30 @@ cleanup_vq(struct vhost_virtqueue *vq, int destroy)
 		close(vq->kickfd);
 }
 
+void
+cleanup_vq_inflight_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
+	if (!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
+		return;
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG, "function only be used for split inflight.\n");
+	} else {
+		if (vq->inflight_split)
+			vq->inflight_split = NULL;
+		if (vq->resubmit_inflight_split) {
+			if (vq->resubmit_inflight_split->resubmit_list) {
+				free(vq->resubmit_inflight_split->resubmit_list);
+				vq->resubmit_inflight_split->resubmit_list = NULL;
+			}
+			free(vq->resubmit_inflight_split);
+			vq->resubmit_inflight_split = NULL;
+		}
+	}
+}
+
 /*
  * Unmap any memory, close any file descriptors and
  * free any memory owned by a device.
@@ -253,8 +277,10 @@ cleanup_device(struct virtio_net *dev, int destroy)
 
 	vhost_backend_cleanup(dev);
 
-	for (i = 0; i < dev->nr_vring; i++)
+	for (i = 0; i < dev->nr_vring; i++) {
 		cleanup_vq(dev->virtqueue[i], destroy);
+		cleanup_vq_inflight_split(dev, dev->virtqueue[i]);
+	}
 }
 
 void
@@ -726,6 +752,35 @@ rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
 	return 0;
 }
 
+int __rte_experimental
+rte_vhost_get_vhost_ring_inflight_split(int vid, uint16_t vring_idx,
+			  struct rte_vhost_ring_inflight_split *vring)
+{
+	struct virtio_net *dev;
+	struct inflight_info_split *inflight_vring;
+
+	dev = get_device(vid);
+	if (dev == NULL || vring == NULL)
+		return -1;
+
+	if (vring_idx >= VHOST_MAX_VRING)
+		return -1;
+
+	inflight_vring = dev->virtqueue[vring_idx]->inflight_split;
+	if (!inflight_vring)
+		return -1;
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG, "Can't be packed inflight ring.\n");
+	} else {
+		vring->inflight_split = inflight_vring;
+		vring->resubmit_inflight_split = dev->virtqueue[vring_idx]->resubmit_inflight_split;
+	}
+
+	return 0;
+}
+
 int
 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
 			  struct rte_vhost_vring *vring)
@@ -781,6 +836,107 @@ rte_vhost_vring_call(int vid, uint16_t vring_idx)
 	return 0;
 }
 
+int
+rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vq_is_packed(dev)))
+		return -1;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (unlikely(!vq))
+		return -1;
+
+	if (unlikely(!vq->inflight_split))
+		return -1;
+
+	vq->inflight_split->desc[idx].counter = vq->global_counter++;
+	vq->inflight_split->desc[idx].inflight = 1;
+	return 0;
+}
+
+int
+rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
+	uint16_t last_used_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vq_is_packed(dev)))
+		return -1;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (unlikely(!vq))
+		return -1;
+
+	if (unlikely(!vq->inflight_split))
+		return -1;
+
+	rte_compiler_barrier();
+
+	vq->inflight_split->desc[idx].inflight = 0;
+
+	rte_compiler_barrier();
+
+	vq->inflight_split->used_idx = last_used_idx;
+	return 0;
+}
+
+int
+rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vq_is_packed(dev)))
+		return -1;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (!vq)
+		return -1;
+
+	if (unlikely(!vq->inflight_split))
+		return -1;
+
+	vq->inflight_split->last_inflight_io = idx;
+	return 0;
+}
+
 uint16_t
 rte_vhost_avail_entries(int vid, uint16_t queue_id)
 {
diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
index 884befa85..22925586d 100644
--- a/lib/librte_vhost/vhost.h
+++ b/lib/librte_vhost/vhost.h
@@ -128,6 +128,14 @@ struct vhost_virtqueue {
 	/* Physical address of used ring, for logging */
 	uint64_t		log_guest_addr;
 
+	/* inflight share memory info */
+	union {
+		struct inflight_info_split *inflight_split;
+		/* TODO */
+	};
+	struct resubmit_info *resubmit_inflight_split;
+	uint64_t		global_counter;
+
 	uint16_t		nr_zmbuf;
 	uint16_t		zmbuf_size;
 	uint16_t		last_zmbuf_idx;
@@ -286,6 +294,12 @@ struct guest_page {
 	uint64_t size;
 };
 
+struct inflight_mem_info {
+	int		fd;
+	void		*addr;
+	uint64_t	size;
+};
+
 /**
  * Device structure contains all configuration information relating
  * to the device.
@@ -303,6 +317,7 @@ struct virtio_net {
 	uint32_t		nr_vring;
 	int			dequeue_zero_copy;
 	struct vhost_virtqueue	*virtqueue[VHOST_MAX_QUEUE_PAIRS * 2];
+	struct inflight_mem_info inflight_info;
 #define IF_NAME_SZ (PATH_MAX > IFNAMSIZ ? PATH_MAX : IFNAMSIZ)
 	char			ifname[IF_NAME_SZ];
 	uint64_t		log_size;
@@ -467,6 +482,7 @@ void vhost_destroy_device(int);
 void vhost_destroy_device_notify(struct virtio_net *dev);
 
 void cleanup_vq(struct vhost_virtqueue *vq, int destroy);
+void cleanup_vq_inflight_split(struct virtio_net *dev, struct vhost_virtqueue *vq);
 void free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq);
 
 int alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx);
diff --git a/lib/librte_vhost/vhost_user.c b/lib/librte_vhost/vhost_user.c
index c9e29ece8..feb642ad7 100644
--- a/lib/librte_vhost/vhost_user.c
+++ b/lib/librte_vhost/vhost_user.c
@@ -31,6 +31,8 @@
 #include <sys/stat.h>
 #include <sys/syscall.h>
 #include <assert.h>
+#include <sys/syscall.h>
+#include <asm/unistd.h>
 #ifdef RTE_LIBRTE_VHOST_NUMA
 #include <numaif.h>
 #endif
@@ -49,6 +51,15 @@
 #define VIRTIO_MIN_MTU 68
 #define VIRTIO_MAX_MTU 65535
 
+#define INFLIGHT_ALIGNMENT	64
+#define INFLIGHT_VERSION	0xabcd
+#define VIRTQUEUE_MAX_SIZE	1024
+
+#define CLOEXEC	0x0001U
+
+#define ALIGN_DOWN(n, m) ((n) / (m) * (m))
+#define ALIGN_UP(n, m) ALIGN_DOWN((n) + (m) - 1, (m))
+
 static const char *vhost_message_str[VHOST_USER_MAX] = {
 	[VHOST_USER_NONE] = "VHOST_USER_NONE",
 	[VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
@@ -78,6 +89,8 @@ static const char *vhost_message_str[VHOST_USER_MAX] = {
 	[VHOST_USER_POSTCOPY_ADVISE]  = "VHOST_USER_POSTCOPY_ADVISE",
 	[VHOST_USER_POSTCOPY_LISTEN]  = "VHOST_USER_POSTCOPY_LISTEN",
 	[VHOST_USER_POSTCOPY_END]  = "VHOST_USER_POSTCOPY_END",
+	[VHOST_USER_GET_INFLIGHT_FD] = "VHOST_USER_GET_INFLIGHT_FD",
+	[VHOST_USER_SET_INFLIGHT_FD] = "VHOST_USER_SET_INFLIGHT_FD",
 };
 
 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
@@ -160,6 +173,16 @@ vhost_backend_cleanup(struct virtio_net *dev)
 		dev->log_addr = 0;
 	}
 
+	if (dev->inflight_info.addr) {
+		munmap(dev->inflight_info.addr, dev->inflight_info.size);
+		dev->inflight_info.addr = NULL;
+	}
+
+	if (dev->inflight_info.fd > 0) {
+		close(dev->inflight_info.fd);
+		dev->inflight_info.fd = -1;
+	}
+
 	if (dev->slave_req_fd >= 0) {
 		close(dev->slave_req_fd);
 		dev->slave_req_fd = -1;
@@ -306,6 +329,7 @@ vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 
 			dev->virtqueue[dev->nr_vring] = NULL;
 			cleanup_vq(vq, 1);
+			cleanup_vq_inflight_split(dev, vq);
 			free_vq(dev, vq);
 		}
 	}
@@ -1165,6 +1189,204 @@ virtio_is_ready(struct virtio_net *dev)
 	return 1;
 }
 
+static int mem_create(const char *name, unsigned int flags)
+{
+#ifdef __NR_memfd_create
+	return syscall(__NR_memfd_create, name, flags);
+#else
+	return -1;
+#endif
+}
+
+void *inflight_mem_alloc(const char *name, size_t size, int *fd)
+{
+	void *ptr;
+	int mfd = -1;
+	char fname[20] = "/tmp/memfd-XXXXXX";
+
+	*fd = -1;
+	mfd = mem_create(name, CLOEXEC);
+	if (mfd != -1) {
+		if (ftruncate(mfd, size) == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"ftruncate fail for alloc inflight buffer\n");
+			close(mfd);
+			return NULL;
+		}
+	} else {
+		mfd = mkstemp(fname);
+		unlink(fname);
+
+		if (mfd == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"mkstemp fail for alloc inflight buffer\n");
+			return NULL;
+		}
+
+		if (ftruncate(mfd, size) == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"ftruncate fail for alloc inflight buffer\n");
+			close(mfd);
+			return NULL;
+		}
+	}
+
+	ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
+	if (ptr == MAP_FAILED) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+				"mmap fail for alloc inflight buffer\n");
+		close(mfd);
+		return NULL;
+	}
+
+	*fd = mfd;
+	return ptr;
+}
+
+static uint32_t get_pervq_shm_size_split(uint16_t queue_size)
+{
+	return ALIGN_UP(sizeof(struct inflight_desc_split) * queue_size +
+		sizeof(uint64_t) * 1 + sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
+}
+
+static int
+vhost_user_get_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
+		int main_fd __rte_unused)
+{
+	int fd;
+	uint64_t mmap_size;
+	void *addr;
+	uint16_t num_queues, queue_size;
+	struct virtio_net *dev = *pdev;
+
+	if (msg->size != sizeof(msg->payload.inflight)) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Invalid get_inflight_fd message size is %d",
+			msg->size);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	num_queues = msg->payload.inflight.num_queues;
+	queue_size = msg->payload.inflight.queue_size;
+
+	RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd num_queues: %u\n",
+			msg->payload.inflight.num_queues);
+	RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd queue_size: %u\n",
+			msg->payload.inflight.queue_size);
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Don't support the packed ring\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	} else {
+		mmap_size = num_queues * get_pervq_shm_size_split(queue_size);
+	}
+
+	addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
+	if (!addr) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Failed to alloc vhost inflight area");
+			msg->payload.inflight.mmap_size = 0;
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+	memset(addr, 0, mmap_size);
+
+	dev->inflight_info.addr = addr;
+	dev->inflight_info.size = msg->payload.inflight.mmap_size = mmap_size;
+	dev->inflight_info.fd = msg->fds[0] = fd;
+	msg->payload.inflight.mmap_offset = 0;
+	msg->fd_num = 1;
+
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight mmap_size: %lu\n",
+			msg->payload.inflight.mmap_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight mmap_offset: %lu\n",
+			msg->payload.inflight.mmap_offset);
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight fd: %d\n", msg->fds[0]);
+
+	return RTE_VHOST_MSG_RESULT_REPLY;
+}
+
+static int
+vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
+		int main_fd __rte_unused)
+{
+	int fd, i;
+	uint64_t mmap_size, mmap_offset;
+	uint16_t num_queues, queue_size;
+	uint32_t pervq_inflight_size;
+	void *addr;
+	struct vhost_virtqueue *vq;
+	struct virtio_net *dev = *pdev;
+
+	fd = msg->fds[0];
+	if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Invalid set_inflight_fd message size is %d,fd is %d\n",
+			msg->size, fd);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	mmap_size = msg->payload.inflight.mmap_size;
+	mmap_offset = msg->payload.inflight.mmap_offset;
+	num_queues = msg->payload.inflight.num_queues;
+	queue_size = msg->payload.inflight.queue_size;
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Don't support the packed ring\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	} else {
+		pervq_inflight_size = get_pervq_shm_size_split(queue_size);
+	}
+
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd mmap_size: %lu\n", mmap_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd mmap_offset: %lu\n", mmap_offset);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd num_queues: %u\n", num_queues);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd queue_size: %u\n", queue_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd fd: %d\n", fd);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd pervq_inflight_size: %d\n",
+		pervq_inflight_size);
+
+	if (dev->inflight_info.addr)
+		munmap(dev->inflight_info.addr, dev->inflight_info.size);
+
+	addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
+			fd, mmap_offset);
+	if (addr == MAP_FAILED) {
+		RTE_LOG(ERR, VHOST_CONFIG, "failed to mmap share memory.\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	if (dev->inflight_info.fd)
+		close(dev->inflight_info.fd);
+
+	dev->inflight_info.fd = fd;
+	dev->inflight_info.addr = addr;
+	dev->inflight_info.size = mmap_size;
+
+	for (i = 0; i < num_queues; i++) {
+		vq = dev->virtqueue[i];
+		if (vq_is_packed(dev)) {
+			/* TODO */
+		} else {
+			vq->inflight_split = (struct inflight_info_split *)addr;
+			vq->inflight_split->desc_num = queue_size;
+		}
+		addr = (void *)((char *)addr + pervq_inflight_size);
+	}
+
+	return RTE_VHOST_MSG_RESULT_OK;
+}
+
 static int
 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
 			int main_fd __rte_unused)
@@ -1201,6 +1423,89 @@ static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
 	return RTE_VHOST_MSG_RESULT_OK;
 }
 
+static int
+resubmit_desc_compare(const void *a, const void *b)
+{
+	const struct resubmit_desc *desc0 = (const struct resubmit_desc *)a;
+	const struct resubmit_desc *desc1 = (const struct resubmit_desc *)b;
+
+	if (desc1->counter > desc0->counter &&
+		(desc1->counter - desc0->counter) < VIRTQUEUE_MAX_SIZE * 2)
+		return 1;
+
+	return -1;
+}
+
+static int
+vhost_check_queue_inflights_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
+	struct vring_used *used = vq->used;
+	uint16_t i = 0;
+	uint16_t resubmit_num = 0;
+	struct resubmit_info *resubmit = NULL;
+
+	if (!(dev->features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
+		return RTE_VHOST_MSG_RESULT_OK;
+
+	if ((!vq->inflight_split))
+		return RTE_VHOST_MSG_RESULT_ERR;
+
+	if (!vq->inflight_split->version) {
+		vq->inflight_split->version = INFLIGHT_VERSION;
+		return RTE_VHOST_MSG_RESULT_OK;
+	}
+
+	vq->resubmit_inflight_split = NULL;
+	vq->global_counter = 0;
+
+	if (vq->inflight_split->used_idx != used->idx) {
+		vq->inflight_split->desc[vq->inflight_split->last_inflight_io].inflight = 0;
+		rte_compiler_barrier();
+		vq->inflight_split->used_idx = used->idx;
+	}
+
+	for (i = 0; i < vq->inflight_split->desc_num; i++) {
+		if (vq->inflight_split->desc[i].inflight == 1)
+			resubmit_num++;
+	}
+
+	vq->last_avail_idx += resubmit_num;
+
+	if (resubmit_num) {
+		resubmit  = calloc(1, sizeof(struct resubmit_info));
+		if (!resubmit) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Failed to allocate memory for resubmit info.\n");
+			return RTE_VHOST_MSG_RESULT_ERR;
+		}
+
+		resubmit->resubmit_list = calloc(resubmit_num,
+			sizeof(struct resubmit_desc));
+		if (!resubmit->resubmit_list) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Failed to allocate memory for inflight desc.\n");
+			return RTE_VHOST_MSG_RESULT_ERR;
+		}
+
+		for (i = 0; i < vq->inflight_split->desc_num; i++) {
+			if (vq->inflight_split->desc[i].inflight == 1) {
+				resubmit->resubmit_list[resubmit->resubmit_num].index = i;
+				resubmit->resubmit_list[resubmit->resubmit_num].counter =
+					vq->inflight_split->desc[i].counter;
+				resubmit->resubmit_num++;
+			}
+		}
+
+		if (resubmit->resubmit_num > 1)
+			qsort(resubmit->resubmit_list, resubmit->resubmit_num,
+				sizeof(struct resubmit_desc), resubmit_desc_compare);
+
+		vq->global_counter = resubmit->resubmit_list[0].counter + 1;
+		vq->resubmit_inflight_split = resubmit;
+	}
+
+	return RTE_VHOST_MSG_RESULT_OK;
+}
+
 static int
 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
 			int main_fd __rte_unused)
@@ -1242,6 +1547,12 @@ vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
 		close(vq->kickfd);
 	vq->kickfd = file.fd;
 
+	if (vhost_check_queue_inflights_split(dev, vq)) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Failed to inflights for vq: %d\n", file.index);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
 	return RTE_VHOST_MSG_RESULT_OK;
 }
 
@@ -1762,6 +2073,8 @@ static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
 	[VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
 	[VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
 	[VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
+	[VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
+	[VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
 };
 
 
diff --git a/lib/librte_vhost/vhost_user.h b/lib/librte_vhost/vhost_user.h
index 2a650fe4b..99a773910 100644
--- a/lib/librte_vhost/vhost_user.h
+++ b/lib/librte_vhost/vhost_user.h
@@ -54,7 +54,9 @@ typedef enum VhostUserRequest {
 	VHOST_USER_POSTCOPY_ADVISE = 28,
 	VHOST_USER_POSTCOPY_LISTEN = 29,
 	VHOST_USER_POSTCOPY_END = 30,
-	VHOST_USER_MAX = 31
+	VHOST_USER_GET_INFLIGHT_FD = 31,
+	VHOST_USER_SET_INFLIGHT_FD = 32,
+	VHOST_USER_MAX = 33
 } VhostUserRequest;
 
 typedef enum VhostUserSlaveRequest {
@@ -112,6 +114,13 @@ typedef struct VhostUserVringArea {
 	uint64_t offset;
 } VhostUserVringArea;
 
+typedef struct VhostUserInflight {
+	uint64_t mmap_size;
+	uint64_t mmap_offset;
+	uint16_t num_queues;
+	uint16_t queue_size;
+} VhostUserInflight;
+
 typedef struct VhostUserMsg {
 	union {
 		uint32_t master; /* a VhostUserRequest value */
@@ -131,6 +140,7 @@ typedef struct VhostUserMsg {
 		struct vhost_vring_addr addr;
 		VhostUserMemory memory;
 		VhostUserLog    log;
+		VhostUserInflight inflight;
 		struct vhost_iotlb_msg iotlb;
 		VhostUserCryptoSessionParam crypto_session;
 		VhostUserVringArea area;
@@ -148,6 +158,7 @@ typedef struct VhostUserMsg {
 /* vhost_user.c */
 int vhost_user_msg_handler(int vid, int fd);
 int vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm);
+void *inflight_mem_alloc(const char *name, size_t size, int *fd);
 
 /* socket.c */
 int read_fd_message(int sockfd, char *buf, int buflen, int *fds, int max_fds,
-- 
2.17.2

[dpdk-dev] [PATCH v2] [v2, 1/2]vhost: support inflight share memory protocol feature

[JinYu]

This patch introduces two new messages VHOST_USER_GET_INFLIGHT_FD
and VHOST_USER_SET_INFLIGHT_FD to support transferring a shared
buffer between qemu and backend.

Firstly, qemu uses VHOST_USER_GET_INFLIGHT_FD to get the
shared buffer from backend. Then qemu should send it back
through VHOST_USER_SET_INFLIGHT_FD each time we start vhost-user.

This shared buffer is used to process inflight I/O when backend
reconnect.

Signed-off-by: LinLi <lilin24@baidu.com>
Signed-off-by: XunNi <nixun@baidu.com>
Signed-off-by: YuZhang <zhangyu31@baidu.com>
Signed-off-by: JinYu <jin.yu@intel.com>
---
v1 - specify the APIs are split-ring only
v2 - fix APIs and judge split or packed
---
 lib/librte_vhost/rte_vhost.h           | 105 +++++++++
 lib/librte_vhost/rte_vhost_version.map |   4 +
 lib/librte_vhost/vhost.c               | 158 ++++++++++++-
 lib/librte_vhost/vhost.h               |  16 ++
 lib/librte_vhost/vhost_user.c          | 313 +++++++++++++++++++++++++
 lib/librte_vhost/vhost_user.h          |  13 +-
 6 files changed, 607 insertions(+), 2 deletions(-)

diff --git a/lib/librte_vhost/rte_vhost.h b/lib/librte_vhost/rte_vhost.h
index 0226b3eff..a17aff876 100644
--- a/lib/librte_vhost/rte_vhost.h
+++ b/lib/librte_vhost/rte_vhost.h
@@ -71,6 +71,10 @@ extern "C" {
 #define VHOST_USER_PROTOCOL_F_HOST_NOTIFIER 11
 #endif
 
+#ifndef VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD
+#define VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD 12
+#endif
+
 /** Indicate whether protocol features negotiation is supported. */
 #ifndef VHOST_USER_F_PROTOCOL_FEATURES
 #define VHOST_USER_F_PROTOCOL_FEATURES	30
@@ -98,6 +102,41 @@ struct rte_vhost_memory {
 	struct rte_vhost_mem_region regions[];
 };
 
+struct inflight_desc_split {
+	uint8_t		inflight;
+	uint8_t		padding[5];
+	uint16_t	next;
+	uint64_t	counter;
+};
+
+struct inflight_info_split {
+	uint64_t		features;
+	uint16_t		version;
+	uint16_t		desc_num;
+	uint16_t		last_inflight_io;
+	uint16_t		used_idx;
+	struct inflight_desc_split desc[0];
+};
+
+struct resubmit_desc {
+	uint16_t index;
+	uint64_t counter;
+};
+
+struct resubmit_info {
+	struct resubmit_desc	*resubmit_list;
+	uint16_t		resubmit_num;
+};
+
+struct rte_vhost_ring_inflight_split {
+	union {
+		struct inflight_info_split *inflight_split;
+		/* TODO */
+	};
+
+	struct resubmit_info *resubmit_inflight_split;
+};
+
 struct rte_vhost_vring {
 	struct vring_desc	*desc;
 	struct vring_avail	*avail;
@@ -603,6 +642,22 @@ uint16_t rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
  */
 int rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem);
 
+/**
+ * Get guest inflight vring info, including inflight ring and resubmit list.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param vring
+ *  the structure to hold the requested inflight vring info
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_get_vhost_ring_inflight_split(int vid, uint16_t vring_idx,
+			      struct rte_vhost_ring_inflight_split *vring);
+
 /**
  * Get guest vring info, including the vring address, vring size, etc.
  *
@@ -631,6 +686,56 @@ int rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
  */
 int rte_vhost_vring_call(int vid, uint16_t vring_idx);
 
+/**
+ * set inflight flag for a desc.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
+		uint16_t idx);
+
+/**
+ * clear inflight flag for a desc.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param last_used_idx
+ *  next free used_idx
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
+		uint16_t last_used_idx, uint16_t idx);
+
+/**
+ * set last inflight io index.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
+		uint16_t idx);
+
 /**
  * Get vhost RX queue avail count.
  *
diff --git a/lib/librte_vhost/rte_vhost_version.map b/lib/librte_vhost/rte_vhost_version.map
index 5f1d4a75c..709593701 100644
--- a/lib/librte_vhost/rte_vhost_version.map
+++ b/lib/librte_vhost/rte_vhost_version.map
@@ -87,4 +87,8 @@ EXPERIMENTAL {
 	rte_vdpa_relay_vring_used;
 	rte_vhost_extern_callback_register;
 	rte_vhost_driver_set_protocol_features;
+	rte_vhost_set_inflight_desc_split;
+	rte_vhost_clr_inflight_desc_split;
+	rte_vhost_set_last_inflight_io_split;
+	rte_vhost_get_vhost_vring_with_inflight_split;
 };
diff --git a/lib/librte_vhost/vhost.c b/lib/librte_vhost/vhost.c
index 981837b5d..ecd580cc5 100644
--- a/lib/librte_vhost/vhost.c
+++ b/lib/librte_vhost/vhost.c
@@ -242,6 +242,30 @@ cleanup_vq(struct vhost_virtqueue *vq, int destroy)
 		close(vq->kickfd);
 }
 
+void
+cleanup_vq_inflight_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
+	if (!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
+		return;
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG, "function only be used for split inflight.\n");
+	} else {
+		if (vq->inflight_split)
+			vq->inflight_split = NULL;
+		if (vq->resubmit_inflight_split) {
+			if (vq->resubmit_inflight_split->resubmit_list) {
+				free(vq->resubmit_inflight_split->resubmit_list);
+				vq->resubmit_inflight_split->resubmit_list = NULL;
+			}
+			free(vq->resubmit_inflight_split);
+			vq->resubmit_inflight_split = NULL;
+		}
+	}
+}
+
 /*
  * Unmap any memory, close any file descriptors and
  * free any memory owned by a device.
@@ -253,8 +277,10 @@ cleanup_device(struct virtio_net *dev, int destroy)
 
 	vhost_backend_cleanup(dev);
 
-	for (i = 0; i < dev->nr_vring; i++)
+	for (i = 0; i < dev->nr_vring; i++) {
 		cleanup_vq(dev->virtqueue[i], destroy);
+		cleanup_vq_inflight_split(dev, dev->virtqueue[i]);
+	}
 }
 
 void
@@ -726,6 +752,35 @@ rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
 	return 0;
 }
 
+int __rte_experimental
+rte_vhost_get_vhost_ring_inflight_split(int vid, uint16_t vring_idx,
+			  struct rte_vhost_ring_inflight_split *vring)
+{
+	struct virtio_net *dev;
+	struct inflight_info_split *inflight_vring;
+
+	dev = get_device(vid);
+	if (dev == NULL || vring == NULL)
+		return -1;
+
+	if (vring_idx >= VHOST_MAX_VRING)
+		return -1;
+
+	inflight_vring = dev->virtqueue[vring_idx]->inflight_split;
+	if (!inflight_vring)
+		return -1;
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG, "Can't be packed inflight ring.\n");
+	} else {
+		vring->inflight_split = inflight_vring;
+		vring->resubmit_inflight_split = dev->virtqueue[vring_idx]->resubmit_inflight_split;
+	}
+
+	return 0;
+}
+
 int
 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
 			  struct rte_vhost_vring *vring)
@@ -781,6 +836,107 @@ rte_vhost_vring_call(int vid, uint16_t vring_idx)
 	return 0;
 }
 
+int
+rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vq_is_packed(dev)))
+		return -1;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (unlikely(!vq))
+		return -1;
+
+	if (unlikely(!vq->inflight_split))
+		return -1;
+
+	vq->inflight_split->desc[idx].counter = vq->global_counter++;
+	vq->inflight_split->desc[idx].inflight = 1;
+	return 0;
+}
+
+int
+rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
+	uint16_t last_used_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vq_is_packed(dev)))
+		return -1;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (unlikely(!vq))
+		return -1;
+
+	if (unlikely(!vq->inflight_split))
+		return -1;
+
+	rte_compiler_barrier();
+
+	vq->inflight_split->desc[idx].inflight = 0;
+
+	rte_compiler_barrier();
+
+	vq->inflight_split->used_idx = last_used_idx;
+	return 0;
+}
+
+int
+rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vq_is_packed(dev)))
+		return -1;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (!vq)
+		return -1;
+
+	if (unlikely(!vq->inflight_split))
+		return -1;
+
+	vq->inflight_split->last_inflight_io = idx;
+	return 0;
+}
+
 uint16_t
 rte_vhost_avail_entries(int vid, uint16_t queue_id)
 {
diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
index 884befa85..22925586d 100644
--- a/lib/librte_vhost/vhost.h
+++ b/lib/librte_vhost/vhost.h
@@ -128,6 +128,14 @@ struct vhost_virtqueue {
 	/* Physical address of used ring, for logging */
 	uint64_t		log_guest_addr;
 
+	/* inflight share memory info */
+	union {
+		struct inflight_info_split *inflight_split;
+		/* TODO */
+	};
+	struct resubmit_info *resubmit_inflight_split;
+	uint64_t		global_counter;
+
 	uint16_t		nr_zmbuf;
 	uint16_t		zmbuf_size;
 	uint16_t		last_zmbuf_idx;
@@ -286,6 +294,12 @@ struct guest_page {
 	uint64_t size;
 };
 
+struct inflight_mem_info {
+	int		fd;
+	void		*addr;
+	uint64_t	size;
+};
+
 /**
  * Device structure contains all configuration information relating
  * to the device.
@@ -303,6 +317,7 @@ struct virtio_net {
 	uint32_t		nr_vring;
 	int			dequeue_zero_copy;
 	struct vhost_virtqueue	*virtqueue[VHOST_MAX_QUEUE_PAIRS * 2];
+	struct inflight_mem_info inflight_info;
 #define IF_NAME_SZ (PATH_MAX > IFNAMSIZ ? PATH_MAX : IFNAMSIZ)
 	char			ifname[IF_NAME_SZ];
 	uint64_t		log_size;
@@ -467,6 +482,7 @@ void vhost_destroy_device(int);
 void vhost_destroy_device_notify(struct virtio_net *dev);
 
 void cleanup_vq(struct vhost_virtqueue *vq, int destroy);
+void cleanup_vq_inflight_split(struct virtio_net *dev, struct vhost_virtqueue *vq);
 void free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq);
 
 int alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx);
diff --git a/lib/librte_vhost/vhost_user.c b/lib/librte_vhost/vhost_user.c
index c9e29ece8..feb642ad7 100644
--- a/lib/librte_vhost/vhost_user.c
+++ b/lib/librte_vhost/vhost_user.c
@@ -31,6 +31,8 @@
 #include <sys/stat.h>
 #include <sys/syscall.h>
 #include <assert.h>
+#include <sys/syscall.h>
+#include <asm/unistd.h>
 #ifdef RTE_LIBRTE_VHOST_NUMA
 #include <numaif.h>
 #endif
@@ -49,6 +51,15 @@
 #define VIRTIO_MIN_MTU 68
 #define VIRTIO_MAX_MTU 65535
 
+#define INFLIGHT_ALIGNMENT	64
+#define INFLIGHT_VERSION	0xabcd
+#define VIRTQUEUE_MAX_SIZE	1024
+
+#define CLOEXEC	0x0001U
+
+#define ALIGN_DOWN(n, m) ((n) / (m) * (m))
+#define ALIGN_UP(n, m) ALIGN_DOWN((n) + (m) - 1, (m))
+
 static const char *vhost_message_str[VHOST_USER_MAX] = {
 	[VHOST_USER_NONE] = "VHOST_USER_NONE",
 	[VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
@@ -78,6 +89,8 @@ static const char *vhost_message_str[VHOST_USER_MAX] = {
 	[VHOST_USER_POSTCOPY_ADVISE]  = "VHOST_USER_POSTCOPY_ADVISE",
 	[VHOST_USER_POSTCOPY_LISTEN]  = "VHOST_USER_POSTCOPY_LISTEN",
 	[VHOST_USER_POSTCOPY_END]  = "VHOST_USER_POSTCOPY_END",
+	[VHOST_USER_GET_INFLIGHT_FD] = "VHOST_USER_GET_INFLIGHT_FD",
+	[VHOST_USER_SET_INFLIGHT_FD] = "VHOST_USER_SET_INFLIGHT_FD",
 };
 
 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
@@ -160,6 +173,16 @@ vhost_backend_cleanup(struct virtio_net *dev)
 		dev->log_addr = 0;
 	}
 
+	if (dev->inflight_info.addr) {
+		munmap(dev->inflight_info.addr, dev->inflight_info.size);
+		dev->inflight_info.addr = NULL;
+	}
+
+	if (dev->inflight_info.fd > 0) {
+		close(dev->inflight_info.fd);
+		dev->inflight_info.fd = -1;
+	}
+
 	if (dev->slave_req_fd >= 0) {
 		close(dev->slave_req_fd);
 		dev->slave_req_fd = -1;
@@ -306,6 +329,7 @@ vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 
 			dev->virtqueue[dev->nr_vring] = NULL;
 			cleanup_vq(vq, 1);
+			cleanup_vq_inflight_split(dev, vq);
 			free_vq(dev, vq);
 		}
 	}
@@ -1165,6 +1189,204 @@ virtio_is_ready(struct virtio_net *dev)
 	return 1;
 }
 
+static int mem_create(const char *name, unsigned int flags)
+{
+#ifdef __NR_memfd_create
+	return syscall(__NR_memfd_create, name, flags);
+#else
+	return -1;
+#endif
+}
+
+void *inflight_mem_alloc(const char *name, size_t size, int *fd)
+{
+	void *ptr;
+	int mfd = -1;
+	char fname[20] = "/tmp/memfd-XXXXXX";
+
+	*fd = -1;
+	mfd = mem_create(name, CLOEXEC);
+	if (mfd != -1) {
+		if (ftruncate(mfd, size) == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"ftruncate fail for alloc inflight buffer\n");
+			close(mfd);
+			return NULL;
+		}
+	} else {
+		mfd = mkstemp(fname);
+		unlink(fname);
+
+		if (mfd == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"mkstemp fail for alloc inflight buffer\n");
+			return NULL;
+		}
+
+		if (ftruncate(mfd, size) == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"ftruncate fail for alloc inflight buffer\n");
+			close(mfd);
+			return NULL;
+		}
+	}
+
+	ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
+	if (ptr == MAP_FAILED) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+				"mmap fail for alloc inflight buffer\n");
+		close(mfd);
+		return NULL;
+	}
+
+	*fd = mfd;
+	return ptr;
+}
+
+static uint32_t get_pervq_shm_size_split(uint16_t queue_size)
+{
+	return ALIGN_UP(sizeof(struct inflight_desc_split) * queue_size +
+		sizeof(uint64_t) * 1 + sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
+}
+
+static int
+vhost_user_get_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
+		int main_fd __rte_unused)
+{
+	int fd;
+	uint64_t mmap_size;
+	void *addr;
+	uint16_t num_queues, queue_size;
+	struct virtio_net *dev = *pdev;
+
+	if (msg->size != sizeof(msg->payload.inflight)) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Invalid get_inflight_fd message size is %d",
+			msg->size);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	num_queues = msg->payload.inflight.num_queues;
+	queue_size = msg->payload.inflight.queue_size;
+
+	RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd num_queues: %u\n",
+			msg->payload.inflight.num_queues);
+	RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd queue_size: %u\n",
+			msg->payload.inflight.queue_size);
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Don't support the packed ring\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	} else {
+		mmap_size = num_queues * get_pervq_shm_size_split(queue_size);
+	}
+
+	addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
+	if (!addr) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Failed to alloc vhost inflight area");
+			msg->payload.inflight.mmap_size = 0;
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+	memset(addr, 0, mmap_size);
+
+	dev->inflight_info.addr = addr;
+	dev->inflight_info.size = msg->payload.inflight.mmap_size = mmap_size;
+	dev->inflight_info.fd = msg->fds[0] = fd;
+	msg->payload.inflight.mmap_offset = 0;
+	msg->fd_num = 1;
+
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight mmap_size: %lu\n",
+			msg->payload.inflight.mmap_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight mmap_offset: %lu\n",
+			msg->payload.inflight.mmap_offset);
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight fd: %d\n", msg->fds[0]);
+
+	return RTE_VHOST_MSG_RESULT_REPLY;
+}
+
+static int
+vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
+		int main_fd __rte_unused)
+{
+	int fd, i;
+	uint64_t mmap_size, mmap_offset;
+	uint16_t num_queues, queue_size;
+	uint32_t pervq_inflight_size;
+	void *addr;
+	struct vhost_virtqueue *vq;
+	struct virtio_net *dev = *pdev;
+
+	fd = msg->fds[0];
+	if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Invalid set_inflight_fd message size is %d,fd is %d\n",
+			msg->size, fd);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	mmap_size = msg->payload.inflight.mmap_size;
+	mmap_offset = msg->payload.inflight.mmap_offset;
+	num_queues = msg->payload.inflight.num_queues;
+	queue_size = msg->payload.inflight.queue_size;
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Don't support the packed ring\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	} else {
+		pervq_inflight_size = get_pervq_shm_size_split(queue_size);
+	}
+
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd mmap_size: %lu\n", mmap_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd mmap_offset: %lu\n", mmap_offset);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd num_queues: %u\n", num_queues);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd queue_size: %u\n", queue_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd fd: %d\n", fd);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd pervq_inflight_size: %d\n",
+		pervq_inflight_size);
+
+	if (dev->inflight_info.addr)
+		munmap(dev->inflight_info.addr, dev->inflight_info.size);
+
+	addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
+			fd, mmap_offset);
+	if (addr == MAP_FAILED) {
+		RTE_LOG(ERR, VHOST_CONFIG, "failed to mmap share memory.\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	if (dev->inflight_info.fd)
+		close(dev->inflight_info.fd);
+
+	dev->inflight_info.fd = fd;
+	dev->inflight_info.addr = addr;
+	dev->inflight_info.size = mmap_size;
+
+	for (i = 0; i < num_queues; i++) {
+		vq = dev->virtqueue[i];
+		if (vq_is_packed(dev)) {
+			/* TODO */
+		} else {
+			vq->inflight_split = (struct inflight_info_split *)addr;
+			vq->inflight_split->desc_num = queue_size;
+		}
+		addr = (void *)((char *)addr + pervq_inflight_size);
+	}
+
+	return RTE_VHOST_MSG_RESULT_OK;
+}
+
 static int
 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
 			int main_fd __rte_unused)
@@ -1201,6 +1423,89 @@ static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
 	return RTE_VHOST_MSG_RESULT_OK;
 }
 
+static int
+resubmit_desc_compare(const void *a, const void *b)
+{
+	const struct resubmit_desc *desc0 = (const struct resubmit_desc *)a;
+	const struct resubmit_desc *desc1 = (const struct resubmit_desc *)b;
+
+	if (desc1->counter > desc0->counter &&
+		(desc1->counter - desc0->counter) < VIRTQUEUE_MAX_SIZE * 2)
+		return 1;
+
+	return -1;
+}
+
+static int
+vhost_check_queue_inflights_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
+	struct vring_used *used = vq->used;
+	uint16_t i = 0;
+	uint16_t resubmit_num = 0;
+	struct resubmit_info *resubmit = NULL;
+
+	if (!(dev->features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
+		return RTE_VHOST_MSG_RESULT_OK;
+
+	if ((!vq->inflight_split))
+		return RTE_VHOST_MSG_RESULT_ERR;
+
+	if (!vq->inflight_split->version) {
+		vq->inflight_split->version = INFLIGHT_VERSION;
+		return RTE_VHOST_MSG_RESULT_OK;
+	}
+
+	vq->resubmit_inflight_split = NULL;
+	vq->global_counter = 0;
+
+	if (vq->inflight_split->used_idx != used->idx) {
+		vq->inflight_split->desc[vq->inflight_split->last_inflight_io].inflight = 0;
+		rte_compiler_barrier();
+		vq->inflight_split->used_idx = used->idx;
+	}
+
+	for (i = 0; i < vq->inflight_split->desc_num; i++) {
+		if (vq->inflight_split->desc[i].inflight == 1)
+			resubmit_num++;
+	}
+
+	vq->last_avail_idx += resubmit_num;
+
+	if (resubmit_num) {
+		resubmit  = calloc(1, sizeof(struct resubmit_info));
+		if (!resubmit) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Failed to allocate memory for resubmit info.\n");
+			return RTE_VHOST_MSG_RESULT_ERR;
+		}
+
+		resubmit->resubmit_list = calloc(resubmit_num,
+			sizeof(struct resubmit_desc));
+		if (!resubmit->resubmit_list) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Failed to allocate memory for inflight desc.\n");
+			return RTE_VHOST_MSG_RESULT_ERR;
+		}
+
+		for (i = 0; i < vq->inflight_split->desc_num; i++) {
+			if (vq->inflight_split->desc[i].inflight == 1) {
+				resubmit->resubmit_list[resubmit->resubmit_num].index = i;
+				resubmit->resubmit_list[resubmit->resubmit_num].counter =
+					vq->inflight_split->desc[i].counter;
+				resubmit->resubmit_num++;
+			}
+		}
+
+		if (resubmit->resubmit_num > 1)
+			qsort(resubmit->resubmit_list, resubmit->resubmit_num,
+				sizeof(struct resubmit_desc), resubmit_desc_compare);
+
+		vq->global_counter = resubmit->resubmit_list[0].counter + 1;
+		vq->resubmit_inflight_split = resubmit;
+	}
+
+	return RTE_VHOST_MSG_RESULT_OK;
+}
+
 static int
 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
 			int main_fd __rte_unused)
@@ -1242,6 +1547,12 @@ vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
 		close(vq->kickfd);
 	vq->kickfd = file.fd;
 
+	if (vhost_check_queue_inflights_split(dev, vq)) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Failed to inflights for vq: %d\n", file.index);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
 	return RTE_VHOST_MSG_RESULT_OK;
 }
 
@@ -1762,6 +2073,8 @@ static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
 	[VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
 	[VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
 	[VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
+	[VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
+	[VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
 };
 
 
diff --git a/lib/librte_vhost/vhost_user.h b/lib/librte_vhost/vhost_user.h
index 2a650fe4b..99a773910 100644
--- a/lib/librte_vhost/vhost_user.h
+++ b/lib/librte_vhost/vhost_user.h
@@ -54,7 +54,9 @@ typedef enum VhostUserRequest {
 	VHOST_USER_POSTCOPY_ADVISE = 28,
 	VHOST_USER_POSTCOPY_LISTEN = 29,
 	VHOST_USER_POSTCOPY_END = 30,
-	VHOST_USER_MAX = 31
+	VHOST_USER_GET_INFLIGHT_FD = 31,
+	VHOST_USER_SET_INFLIGHT_FD = 32,
+	VHOST_USER_MAX = 33
 } VhostUserRequest;
 
 typedef enum VhostUserSlaveRequest {
@@ -112,6 +114,13 @@ typedef struct VhostUserVringArea {
 	uint64_t offset;
 } VhostUserVringArea;
 
+typedef struct VhostUserInflight {
+	uint64_t mmap_size;
+	uint64_t mmap_offset;
+	uint16_t num_queues;
+	uint16_t queue_size;
+} VhostUserInflight;
+
 typedef struct VhostUserMsg {
 	union {
 		uint32_t master; /* a VhostUserRequest value */
@@ -131,6 +140,7 @@ typedef struct VhostUserMsg {
 		struct vhost_vring_addr addr;
 		VhostUserMemory memory;
 		VhostUserLog    log;
+		VhostUserInflight inflight;
 		struct vhost_iotlb_msg iotlb;
 		VhostUserCryptoSessionParam crypto_session;
 		VhostUserVringArea area;
@@ -148,6 +158,7 @@ typedef struct VhostUserMsg {
 /* vhost_user.c */
 int vhost_user_msg_handler(int vid, int fd);
 int vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm);
+void *inflight_mem_alloc(const char *name, size_t size, int *fd);
 
 /* socket.c */
 int read_fd_message(int sockfd, char *buf, int buflen, int *fds, int max_fds,
-- 
2.17.2

[dpdk-dev] [PATCH v2] [1/2]vhost: support inflight share memory protocol feature

[JinYu]

This patch introduces two new messages VHOST_USER_GET_INFLIGHT_FD
and VHOST_USER_SET_INFLIGHT_FD to support transferring a shared
buffer between qemu and backend.

Firstly, qemu uses VHOST_USER_GET_INFLIGHT_FD to get the
shared buffer from backend. Then qemu should send it back
through VHOST_USER_SET_INFLIGHT_FD each time we start vhost-user.

This shared buffer is used to process inflight I/O when backend
reconnect.

Signed-off-by: LinLi <lilin24@baidu.com>
Signed-off-by: XunNi <nixun@baidu.com>
Signed-off-by: YuZhang <zhangyu31@baidu.com>
Signed-off-by: JinYu <jin.yu@intel.com>
---
v1 - specify the APIs are split-ring only
v2 - fix APIs and judge split or packed
---
 lib/librte_vhost/rte_vhost.h           | 105 +++++++++
 lib/librte_vhost/rte_vhost_version.map |   4 +
 lib/librte_vhost/vhost.c               | 158 ++++++++++++-
 lib/librte_vhost/vhost.h               |  16 ++
 lib/librte_vhost/vhost_user.c          | 313 +++++++++++++++++++++++++
 lib/librte_vhost/vhost_user.h          |  13 +-
 6 files changed, 607 insertions(+), 2 deletions(-)

diff --git a/lib/librte_vhost/rte_vhost.h b/lib/librte_vhost/rte_vhost.h
index 0226b3eff..a17aff876 100644
--- a/lib/librte_vhost/rte_vhost.h
+++ b/lib/librte_vhost/rte_vhost.h
@@ -71,6 +71,10 @@ extern "C" {
 #define VHOST_USER_PROTOCOL_F_HOST_NOTIFIER 11
 #endif
 
+#ifndef VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD
+#define VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD 12
+#endif
+
 /** Indicate whether protocol features negotiation is supported. */
 #ifndef VHOST_USER_F_PROTOCOL_FEATURES
 #define VHOST_USER_F_PROTOCOL_FEATURES	30
@@ -98,6 +102,41 @@ struct rte_vhost_memory {
 	struct rte_vhost_mem_region regions[];
 };
 
+struct inflight_desc_split {
+	uint8_t		inflight;
+	uint8_t		padding[5];
+	uint16_t	next;
+	uint64_t	counter;
+};
+
+struct inflight_info_split {
+	uint64_t		features;
+	uint16_t		version;
+	uint16_t		desc_num;
+	uint16_t		last_inflight_io;
+	uint16_t		used_idx;
+	struct inflight_desc_split desc[0];
+};
+
+struct resubmit_desc {
+	uint16_t index;
+	uint64_t counter;
+};
+
+struct resubmit_info {
+	struct resubmit_desc	*resubmit_list;
+	uint16_t		resubmit_num;
+};
+
+struct rte_vhost_ring_inflight_split {
+	union {
+		struct inflight_info_split *inflight_split;
+		/* TODO */
+	};
+
+	struct resubmit_info *resubmit_inflight_split;
+};
+
 struct rte_vhost_vring {
 	struct vring_desc	*desc;
 	struct vring_avail	*avail;
@@ -603,6 +642,22 @@ uint16_t rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
  */
 int rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem);
 
+/**
+ * Get guest inflight vring info, including inflight ring and resubmit list.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param vring
+ *  the structure to hold the requested inflight vring info
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_get_vhost_ring_inflight_split(int vid, uint16_t vring_idx,
+			      struct rte_vhost_ring_inflight_split *vring);
+
 /**
  * Get guest vring info, including the vring address, vring size, etc.
  *
@@ -631,6 +686,56 @@ int rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
  */
 int rte_vhost_vring_call(int vid, uint16_t vring_idx);
 
+/**
+ * set inflight flag for a desc.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
+		uint16_t idx);
+
+/**
+ * clear inflight flag for a desc.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param last_used_idx
+ *  next free used_idx
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
+		uint16_t last_used_idx, uint16_t idx);
+
+/**
+ * set last inflight io index.
+ *
+ * @param vid
+ *  vhost device ID
+ * @param vring_idx
+ *  vring index
+ * @param idx
+ *  inflight entry index
+ * @return
+ *  0 on success, -1 on failure
+ */
+int __rte_experimental
+rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
+		uint16_t idx);
+
 /**
  * Get vhost RX queue avail count.
  *
diff --git a/lib/librte_vhost/rte_vhost_version.map b/lib/librte_vhost/rte_vhost_version.map
index 5f1d4a75c..709593701 100644
--- a/lib/librte_vhost/rte_vhost_version.map
+++ b/lib/librte_vhost/rte_vhost_version.map
@@ -87,4 +87,8 @@ EXPERIMENTAL {
 	rte_vdpa_relay_vring_used;
 	rte_vhost_extern_callback_register;
 	rte_vhost_driver_set_protocol_features;
+	rte_vhost_set_inflight_desc_split;
+	rte_vhost_clr_inflight_desc_split;
+	rte_vhost_set_last_inflight_io_split;
+	rte_vhost_get_vhost_vring_with_inflight_split;
 };
diff --git a/lib/librte_vhost/vhost.c b/lib/librte_vhost/vhost.c
index 981837b5d..ecd580cc5 100644
--- a/lib/librte_vhost/vhost.c
+++ b/lib/librte_vhost/vhost.c
@@ -242,6 +242,30 @@ cleanup_vq(struct vhost_virtqueue *vq, int destroy)
 		close(vq->kickfd);
 }
 
+void
+cleanup_vq_inflight_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
+	if (!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
+		return;
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG, "function only be used for split inflight.\n");
+	} else {
+		if (vq->inflight_split)
+			vq->inflight_split = NULL;
+		if (vq->resubmit_inflight_split) {
+			if (vq->resubmit_inflight_split->resubmit_list) {
+				free(vq->resubmit_inflight_split->resubmit_list);
+				vq->resubmit_inflight_split->resubmit_list = NULL;
+			}
+			free(vq->resubmit_inflight_split);
+			vq->resubmit_inflight_split = NULL;
+		}
+	}
+}
+
 /*
  * Unmap any memory, close any file descriptors and
  * free any memory owned by a device.
@@ -253,8 +277,10 @@ cleanup_device(struct virtio_net *dev, int destroy)
 
 	vhost_backend_cleanup(dev);
 
-	for (i = 0; i < dev->nr_vring; i++)
+	for (i = 0; i < dev->nr_vring; i++) {
 		cleanup_vq(dev->virtqueue[i], destroy);
+		cleanup_vq_inflight_split(dev, dev->virtqueue[i]);
+	}
 }
 
 void
@@ -726,6 +752,35 @@ rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
 	return 0;
 }
 
+int __rte_experimental
+rte_vhost_get_vhost_ring_inflight_split(int vid, uint16_t vring_idx,
+			  struct rte_vhost_ring_inflight_split *vring)
+{
+	struct virtio_net *dev;
+	struct inflight_info_split *inflight_vring;
+
+	dev = get_device(vid);
+	if (dev == NULL || vring == NULL)
+		return -1;
+
+	if (vring_idx >= VHOST_MAX_VRING)
+		return -1;
+
+	inflight_vring = dev->virtqueue[vring_idx]->inflight_split;
+	if (!inflight_vring)
+		return -1;
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG, "Can't be packed inflight ring.\n");
+	} else {
+		vring->inflight_split = inflight_vring;
+		vring->resubmit_inflight_split = dev->virtqueue[vring_idx]->resubmit_inflight_split;
+	}
+
+	return 0;
+}
+
 int
 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
 			  struct rte_vhost_vring *vring)
@@ -781,6 +836,107 @@ rte_vhost_vring_call(int vid, uint16_t vring_idx)
 	return 0;
 }
 
+int
+rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vq_is_packed(dev)))
+		return -1;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (unlikely(!vq))
+		return -1;
+
+	if (unlikely(!vq->inflight_split))
+		return -1;
+
+	vq->inflight_split->desc[idx].counter = vq->global_counter++;
+	vq->inflight_split->desc[idx].inflight = 1;
+	return 0;
+}
+
+int
+rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
+	uint16_t last_used_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vq_is_packed(dev)))
+		return -1;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (unlikely(!vq))
+		return -1;
+
+	if (unlikely(!vq->inflight_split))
+		return -1;
+
+	rte_compiler_barrier();
+
+	vq->inflight_split->desc[idx].inflight = 0;
+
+	rte_compiler_barrier();
+
+	vq->inflight_split->used_idx = last_used_idx;
+	return 0;
+}
+
+int
+rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx, uint16_t idx)
+{
+	struct virtio_net *dev;
+	struct vhost_virtqueue *vq;
+
+	dev = get_device(vid);
+	if (unlikely(!dev))
+		return -1;
+
+	if (unlikely(!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
+		return 0;
+
+	if (unlikely(vq_is_packed(dev)))
+		return -1;
+
+	if (unlikely(vring_idx >= VHOST_MAX_VRING))
+		return -1;
+
+	vq = dev->virtqueue[vring_idx];
+	if (!vq)
+		return -1;
+
+	if (unlikely(!vq->inflight_split))
+		return -1;
+
+	vq->inflight_split->last_inflight_io = idx;
+	return 0;
+}
+
 uint16_t
 rte_vhost_avail_entries(int vid, uint16_t queue_id)
 {
diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
index 884befa85..22925586d 100644
--- a/lib/librte_vhost/vhost.h
+++ b/lib/librte_vhost/vhost.h
@@ -128,6 +128,14 @@ struct vhost_virtqueue {
 	/* Physical address of used ring, for logging */
 	uint64_t		log_guest_addr;
 
+	/* inflight share memory info */
+	union {
+		struct inflight_info_split *inflight_split;
+		/* TODO */
+	};
+	struct resubmit_info *resubmit_inflight_split;
+	uint64_t		global_counter;
+
 	uint16_t		nr_zmbuf;
 	uint16_t		zmbuf_size;
 	uint16_t		last_zmbuf_idx;
@@ -286,6 +294,12 @@ struct guest_page {
 	uint64_t size;
 };
 
+struct inflight_mem_info {
+	int		fd;
+	void		*addr;
+	uint64_t	size;
+};
+
 /**
  * Device structure contains all configuration information relating
  * to the device.
@@ -303,6 +317,7 @@ struct virtio_net {
 	uint32_t		nr_vring;
 	int			dequeue_zero_copy;
 	struct vhost_virtqueue	*virtqueue[VHOST_MAX_QUEUE_PAIRS * 2];
+	struct inflight_mem_info inflight_info;
 #define IF_NAME_SZ (PATH_MAX > IFNAMSIZ ? PATH_MAX : IFNAMSIZ)
 	char			ifname[IF_NAME_SZ];
 	uint64_t		log_size;
@@ -467,6 +482,7 @@ void vhost_destroy_device(int);
 void vhost_destroy_device_notify(struct virtio_net *dev);
 
 void cleanup_vq(struct vhost_virtqueue *vq, int destroy);
+void cleanup_vq_inflight_split(struct virtio_net *dev, struct vhost_virtqueue *vq);
 void free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq);
 
 int alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx);
diff --git a/lib/librte_vhost/vhost_user.c b/lib/librte_vhost/vhost_user.c
index c9e29ece8..32cfcd258 100644
--- a/lib/librte_vhost/vhost_user.c
+++ b/lib/librte_vhost/vhost_user.c
@@ -31,6 +31,8 @@
 #include <sys/stat.h>
 #include <sys/syscall.h>
 #include <assert.h>
+#include <sys/syscall.h>
+#include <asm/unistd.h>
 #ifdef RTE_LIBRTE_VHOST_NUMA
 #include <numaif.h>
 #endif
@@ -49,6 +51,15 @@
 #define VIRTIO_MIN_MTU 68
 #define VIRTIO_MAX_MTU 65535
 
+#define INFLIGHT_ALIGNMENT	64
+#define INFLIGHT_VERSION	0xabcd
+#define VIRTQUEUE_MAX_SIZE	1024
+
+#define CLOEXEC	0x0001U
+
+#define ALIGN_DOWN(n, m) ((n) / (m) * (m))
+#define ALIGN_UP(n, m) ALIGN_DOWN((n) + (m) - 1, (m))
+
 static const char *vhost_message_str[VHOST_USER_MAX] = {
 	[VHOST_USER_NONE] = "VHOST_USER_NONE",
 	[VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
@@ -78,6 +89,8 @@ static const char *vhost_message_str[VHOST_USER_MAX] = {
 	[VHOST_USER_POSTCOPY_ADVISE]  = "VHOST_USER_POSTCOPY_ADVISE",
 	[VHOST_USER_POSTCOPY_LISTEN]  = "VHOST_USER_POSTCOPY_LISTEN",
 	[VHOST_USER_POSTCOPY_END]  = "VHOST_USER_POSTCOPY_END",
+	[VHOST_USER_GET_INFLIGHT_FD] = "VHOST_USER_GET_INFLIGHT_FD",
+	[VHOST_USER_SET_INFLIGHT_FD] = "VHOST_USER_SET_INFLIGHT_FD",
 };
 
 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
@@ -160,6 +173,16 @@ vhost_backend_cleanup(struct virtio_net *dev)
 		dev->log_addr = 0;
 	}
 
+	if (dev->inflight_info.addr) {
+		munmap(dev->inflight_info.addr, dev->inflight_info.size);
+		dev->inflight_info.addr = NULL;
+	}
+
+	if (dev->inflight_info.fd > 0) {
+		close(dev->inflight_info.fd);
+		dev->inflight_info.fd = -1;
+	}
+
 	if (dev->slave_req_fd >= 0) {
 		close(dev->slave_req_fd);
 		dev->slave_req_fd = -1;
@@ -306,6 +329,7 @@ vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 
 			dev->virtqueue[dev->nr_vring] = NULL;
 			cleanup_vq(vq, 1);
+			cleanup_vq_inflight_split(dev, vq);
 			free_vq(dev, vq);
 		}
 	}
@@ -1165,6 +1189,204 @@ virtio_is_ready(struct virtio_net *dev)
 	return 1;
 }
 
+static int mem_create(const char *name, unsigned int flags)
+{
+#ifdef __NR_memfd_create
+	return syscall(__NR_memfd_create, name, flags);
+#else
+	return -1;
+#endif
+}
+
+void *inflight_mem_alloc(const char *name, size_t size, int *fd)
+{
+	void *ptr;
+	int mfd = -1;
+	char fname[20] = "/tmp/memfd-XXXXXX";
+
+	*fd = -1;
+	mfd = mem_create(name, CLOEXEC);
+	if (mfd != -1) {
+		if (ftruncate(mfd, size) == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"ftruncate fail for alloc inflight buffer\n");
+			close(mfd);
+			return NULL;
+		}
+	} else {
+		mfd = mkstemp(fname);
+		unlink(fname);
+
+		if (mfd == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"mkstemp fail for alloc inflight buffer\n");
+			return NULL;
+		}
+
+		if (ftruncate(mfd, size) == -1) {
+			RTE_LOG(ERR, VHOST_CONFIG,
+					"ftruncate fail for alloc inflight buffer\n");
+			close(mfd);
+			return NULL;
+		}
+	}
+
+	ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
+	if (ptr == MAP_FAILED) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+				"mmap fail for alloc inflight buffer\n");
+		close(mfd);
+		return NULL;
+	}
+
+	*fd = mfd;
+	return ptr;
+}
+
+static uint32_t get_pervq_shm_size_split(uint16_t queue_size)
+{
+	return ALIGN_UP(sizeof(struct inflight_desc_split) * queue_size +
+		sizeof(uint64_t) * 1 + sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
+}
+
+static int
+vhost_user_get_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
+		int main_fd __rte_unused)
+{
+	int fd;
+	uint64_t mmap_size;
+	void *addr;
+	uint16_t num_queues, queue_size;
+	struct virtio_net *dev = *pdev;
+
+	if (msg->size != sizeof(msg->payload.inflight)) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Invalid get_inflight_fd message size is %d",
+			msg->size);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	num_queues = msg->payload.inflight.num_queues;
+	queue_size = msg->payload.inflight.queue_size;
+
+	RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd num_queues: %u\n",
+			msg->payload.inflight.num_queues);
+	RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd queue_size: %u\n",
+			msg->payload.inflight.queue_size);
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Don't support the packed ring\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	} else {
+		mmap_size = num_queues * get_pervq_shm_size_split(queue_size);
+	}
+
+	addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
+	if (!addr) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Failed to alloc vhost inflight area");
+			msg->payload.inflight.mmap_size = 0;
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+	memset(addr, 0, mmap_size);
+
+	dev->inflight_info.addr = addr;
+	dev->inflight_info.size = msg->payload.inflight.mmap_size = mmap_size;
+	dev->inflight_info.fd = msg->fds[0] = fd;
+	msg->payload.inflight.mmap_offset = 0;
+	msg->fd_num = 1;
+
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight mmap_size: %lu\n",
+			msg->payload.inflight.mmap_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight mmap_offset: %lu\n",
+			msg->payload.inflight.mmap_offset);
+	RTE_LOG(INFO, VHOST_CONFIG,
+			"send inflight fd: %d\n", msg->fds[0]);
+
+	return RTE_VHOST_MSG_RESULT_REPLY;
+}
+
+static int
+vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
+		int main_fd __rte_unused)
+{
+	int fd, i;
+	uint64_t mmap_size, mmap_offset;
+	uint16_t num_queues, queue_size;
+	uint32_t pervq_inflight_size;
+	void *addr;
+	struct vhost_virtqueue *vq;
+	struct virtio_net *dev = *pdev;
+
+	fd = msg->fds[0];
+	if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
+		RTE_LOG(ERR, VHOST_CONFIG, "Invalid set_inflight_fd message size is %d,fd is %d\n",
+			msg->size, fd);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	mmap_size = msg->payload.inflight.mmap_size;
+	mmap_offset = msg->payload.inflight.mmap_offset;
+	num_queues = msg->payload.inflight.num_queues;
+	queue_size = msg->payload.inflight.queue_size;
+
+	if (vq_is_packed(dev)) {
+		/* TODO */
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Don't support the packed ring\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	} else {
+		pervq_inflight_size = get_pervq_shm_size_split(queue_size);
+	}
+
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd mmap_size: %lu\n", mmap_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd mmap_offset: %lu\n", mmap_offset);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd num_queues: %u\n", num_queues);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd queue_size: %u\n", queue_size);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd fd: %d\n", fd);
+	RTE_LOG(INFO, VHOST_CONFIG,
+		"set_inflight_fd pervq_inflight_size: %d\n",
+		pervq_inflight_size);
+
+	if (dev->inflight_info.addr)
+		munmap(dev->inflight_info.addr, dev->inflight_info.size);
+
+	addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
+			fd, mmap_offset);
+	if (addr == MAP_FAILED) {
+		RTE_LOG(ERR, VHOST_CONFIG, "failed to mmap share memory.\n");
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
+	if (dev->inflight_info.fd)
+		close(dev->inflight_info.fd);
+
+	dev->inflight_info.fd = fd;
+	dev->inflight_info.addr = addr;
+	dev->inflight_info.size = mmap_size;
+
+	for (i = 0; i < num_queues; i++) {
+		vq = dev->virtqueue[i];
+		if (vq_is_packed(dev)) {
+			/* TODO */
+		} else {
+			vq->inflight_split = (struct inflight_info_split *)addr;
+			vq->inflight_split->desc_num = queue_size;
+		}
+		addr = (void *)((char *)addr + pervq_inflight_size);
+	}
+
+	return RTE_VHOST_MSG_RESULT_OK;
+}
+
 static int
 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
 			int main_fd __rte_unused)
@@ -1201,6 +1423,89 @@ static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
 	return RTE_VHOST_MSG_RESULT_OK;
 }
 
+static int
+resubmit_desc_compare(const void *a, const void *b)
+{
+	const struct resubmit_desc *desc0 = (const struct resubmit_desc *)a;
+	const struct resubmit_desc *desc1 = (const struct resubmit_desc *)b;
+
+	if (desc1->counter > desc0->counter &&
+		(desc1->counter - desc0->counter) < VIRTQUEUE_MAX_SIZE * 2)
+		return 1;
+
+	return -1;
+}
+
+static int
+vhost_check_queue_inflights_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
+	struct vring_used *used = vq->used;
+	uint16_t i = 0;
+	uint16_t resubmit_num = 0;
+	struct resubmit_info *resubmit = NULL;
+
+	if (!(dev->protocol_features &
+		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
+		return RTE_VHOST_MSG_RESULT_OK;
+
+	if ((!vq->inflight_split))
+		return RTE_VHOST_MSG_RESULT_ERR;
+
+	if (!vq->inflight_split->version) {
+		vq->inflight_split->version = INFLIGHT_VERSION;
+		return RTE_VHOST_MSG_RESULT_OK;
+	}
+
+	vq->resubmit_inflight_split = NULL;
+	vq->global_counter = 0;
+
+	if (vq->inflight_split->used_idx != used->idx) {
+		vq->inflight_split->desc[vq->inflight_split->last_inflight_io].inflight = 0;
+		rte_compiler_barrier();
+		vq->inflight_split->used_idx = used->idx;
+	}
+
+	for (i = 0; i < vq->inflight_split->desc_num; i++) {
+		if (vq->inflight_split->desc[i].inflight == 1)
+			resubmit_num++;
+	}
+
+	vq->last_avail_idx += resubmit_num;
+
+	if (resubmit_num) {
+		resubmit  = calloc(1, sizeof(struct resubmit_info));
+		if (!resubmit) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Failed to allocate memory for resubmit info.\n");
+			return RTE_VHOST_MSG_RESULT_ERR;
+		}
+
+		resubmit->resubmit_list = calloc(resubmit_num,
+			sizeof(struct resubmit_desc));
+		if (!resubmit->resubmit_list) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Failed to allocate memory for inflight desc.\n");
+			return RTE_VHOST_MSG_RESULT_ERR;
+		}
+
+		for (i = 0; i < vq->inflight_split->desc_num; i++) {
+			if (vq->inflight_split->desc[i].inflight == 1) {
+				resubmit->resubmit_list[resubmit->resubmit_num].index = i;
+				resubmit->resubmit_list[resubmit->resubmit_num].counter =
+					vq->inflight_split->desc[i].counter;
+				resubmit->resubmit_num++;
+			}
+		}
+
+		if (resubmit->resubmit_num > 1)
+			qsort(resubmit->resubmit_list, resubmit->resubmit_num,
+				sizeof(struct resubmit_desc), resubmit_desc_compare);
+
+		vq->global_counter = resubmit->resubmit_list[0].counter + 1;
+		vq->resubmit_inflight_split = resubmit;
+	}
+
+	return RTE_VHOST_MSG_RESULT_OK;
+}
+
 static int
 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
 			int main_fd __rte_unused)
@@ -1242,6 +1547,12 @@ vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
 		close(vq->kickfd);
 	vq->kickfd = file.fd;
 
+	if (vhost_check_queue_inflights_split(dev, vq)) {
+		RTE_LOG(ERR, VHOST_CONFIG,
+			"Failed to inflights for vq: %d\n", file.index);
+		return RTE_VHOST_MSG_RESULT_ERR;
+	}
+
 	return RTE_VHOST_MSG_RESULT_OK;
 }
 
@@ -1762,6 +2073,8 @@ static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
 	[VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
 	[VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
 	[VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
+	[VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
+	[VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
 };
 
 
diff --git a/lib/librte_vhost/vhost_user.h b/lib/librte_vhost/vhost_user.h
index 2a650fe4b..99a773910 100644
--- a/lib/librte_vhost/vhost_user.h
+++ b/lib/librte_vhost/vhost_user.h
@@ -54,7 +54,9 @@ typedef enum VhostUserRequest {
 	VHOST_USER_POSTCOPY_ADVISE = 28,
 	VHOST_USER_POSTCOPY_LISTEN = 29,
 	VHOST_USER_POSTCOPY_END = 30,
-	VHOST_USER_MAX = 31
+	VHOST_USER_GET_INFLIGHT_FD = 31,
+	VHOST_USER_SET_INFLIGHT_FD = 32,
+	VHOST_USER_MAX = 33
 } VhostUserRequest;
 
 typedef enum VhostUserSlaveRequest {
@@ -112,6 +114,13 @@ typedef struct VhostUserVringArea {
 	uint64_t offset;
 } VhostUserVringArea;
 
+typedef struct VhostUserInflight {
+	uint64_t mmap_size;
+	uint64_t mmap_offset;
+	uint16_t num_queues;
+	uint16_t queue_size;
+} VhostUserInflight;
+
 typedef struct VhostUserMsg {
 	union {
 		uint32_t master; /* a VhostUserRequest value */
@@ -131,6 +140,7 @@ typedef struct VhostUserMsg {
 		struct vhost_vring_addr addr;
 		VhostUserMemory memory;
 		VhostUserLog    log;
+		VhostUserInflight inflight;
 		struct vhost_iotlb_msg iotlb;
 		VhostUserCryptoSessionParam crypto_session;
 		VhostUserVringArea area;
@@ -148,6 +158,7 @@ typedef struct VhostUserMsg {
 /* vhost_user.c */
 int vhost_user_msg_handler(int vid, int fd);
 int vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm);
+void *inflight_mem_alloc(const char *name, size_t size, int *fd);
 
 /* socket.c */
 int read_fd_message(int sockfd, char *buf, int buflen, int *fds, int max_fds,
-- 
2.17.2