numam-spdk/lib/vhost/vhost.c

/*-
 *   BSD LICENSE
 *
 *   Copyright(c) Intel Corporation. All rights reserved.
 *   All rights reserved.
 *
 *   Redistribution and use in source and binary forms, with or without
 *   modification, are permitted provided that the following conditions
 *   are met:
 *
 *     * Redistributions of source code must retain the above copyright
 *       notice, this list of conditions and the following disclaimer.
 *     * Redistributions in binary form must reproduce the above copyright
 *       notice, this list of conditions and the following disclaimer in
 *       the documentation and/or other materials provided with the
 *       distribution.
 *     * Neither the name of Intel Corporation nor the names of its
 *       contributors may be used to endorse or promote products derived
 *       from this software without specific prior written permission.
 *
 *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
 *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
 *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
 *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
 *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

#include "spdk/stdinc.h"

#include "spdk/env.h"
#include "spdk/likely.h"
#include "spdk/string.h"
#include "spdk/util.h"
#include "spdk/barrier.h"
#include "spdk/vhost.h"
#include "vhost_internal.h"

#include "spdk_internal/memory.h"

static TAILQ_HEAD(, vhost_poll_group) g_poll_groups = TAILQ_HEAD_INITIALIZER(g_poll_groups);

/* Temporary cpuset for poll group assignment */
static struct spdk_cpuset *g_tmp_cpuset;

/* Path to folder where character device will be created. Can be set by user. */
static char dev_dirname[PATH_MAX] = "";

/* Thread performing all vhost management operations */
static struct spdk_thread *g_vhost_init_thread;

static spdk_vhost_fini_cb g_fini_cpl_cb;

/**
 * DPDK calls our callbacks synchronously but the work those callbacks
 * perform needs to be async. Luckily, all DPDK callbacks are called on
 * a DPDK-internal pthread, so we'll just wait on a semaphore in there.
 */
static sem_t g_dpdk_sem;

/** Return code for the current DPDK callback */
static int g_dpdk_response;

struct vhost_session_fn_ctx {
	/** Device pointer obtained before enqueuing the event */
	struct spdk_vhost_dev *vdev;

	/** ID of the session to send event to. */
	uint32_t vsession_id;

	/** User provided function to be executed on session's thread. */
	spdk_vhost_session_fn cb_fn;

	/**
	 * User provided function to be called on the init thread
	 * after iterating through all sessions.
	 */
	spdk_vhost_dev_fn cpl_fn;

	/** Custom user context */
	void *user_ctx;
};

static int new_connection(int vid);
static int start_device(int vid);
static void stop_device(int vid);
static void destroy_connection(int vid);

#ifdef SPDK_CONFIG_VHOST_INTERNAL_LIB
static int get_config(int vid, uint8_t *config, uint32_t len);
static int set_config(int vid, uint8_t *config, uint32_t offset,
		      uint32_t size, uint32_t flags);
#endif

const struct vhost_device_ops g_spdk_vhost_ops = {
	.new_device =  start_device,
	.destroy_device = stop_device,
	.new_connection = new_connection,
	.destroy_connection = destroy_connection,
#ifdef SPDK_CONFIG_VHOST_INTERNAL_LIB
	.get_config = get_config,
	.set_config = set_config,
	.vhost_nvme_admin_passthrough = vhost_nvme_admin_passthrough,
	.vhost_nvme_set_cq_call = vhost_nvme_set_cq_call,
	.vhost_nvme_get_cap = vhost_nvme_get_cap,
	.vhost_nvme_set_bar_mr = vhost_nvme_set_bar_mr,
#endif
};

static TAILQ_HEAD(, spdk_vhost_dev) g_vhost_devices = TAILQ_HEAD_INITIALIZER(
			g_vhost_devices);
static pthread_mutex_t g_vhost_mutex = PTHREAD_MUTEX_INITIALIZER;

void *vhost_gpa_to_vva(struct spdk_vhost_session *vsession, uint64_t addr, uint64_t len)
{
	void *vva;
	uint64_t newlen;

	newlen = len;
	vva = (void *)rte_vhost_va_from_guest_pa(vsession->mem, addr, &newlen);
	if (newlen != len) {
		return NULL;
	}

	return vva;

}

static void
vhost_log_req_desc(struct spdk_vhost_session *vsession, struct spdk_vhost_virtqueue *virtqueue,
		   uint16_t req_id)
{
	struct vring_desc *desc, *desc_table;
	uint32_t desc_table_size;
	int rc;

	if (spdk_likely(!vhost_dev_has_feature(vsession, VHOST_F_LOG_ALL))) {
		return;
	}

	rc = vhost_vq_get_desc(vsession, virtqueue, req_id, &desc, &desc_table, &desc_table_size);
	if (spdk_unlikely(rc != 0)) {
		SPDK_ERRLOG("Can't log used ring descriptors!\n");
		return;
	}

	do {
		if (vhost_vring_desc_is_wr(desc)) {
			/* To be honest, only pages realy touched should be logged, but
			 * doing so would require tracking those changes in each backed.
			 * Also backend most likely will touch all/most of those pages so
			 * for lets assume we touched all pages passed to as writeable buffers. */
			rte_vhost_log_write(vsession->vid, desc->addr, desc->len);
		}
		vhost_vring_desc_get_next(&desc, desc_table, desc_table_size);
	} while (desc);
}

static void
vhost_log_used_vring_elem(struct spdk_vhost_session *vsession,
			  struct spdk_vhost_virtqueue *virtqueue,
			  uint16_t idx)
{
	uint64_t offset, len;
	uint16_t vq_idx;

	if (spdk_likely(!vhost_dev_has_feature(vsession, VHOST_F_LOG_ALL))) {
		return;
	}

	offset = offsetof(struct vring_used, ring[idx]);
	len = sizeof(virtqueue->vring.used->ring[idx]);
	vq_idx = virtqueue - vsession->virtqueue;

	rte_vhost_log_used_vring(vsession->vid, vq_idx, offset, len);
}

static void
vhost_log_used_vring_idx(struct spdk_vhost_session *vsession,
			 struct spdk_vhost_virtqueue *virtqueue)
{
	uint64_t offset, len;
	uint16_t vq_idx;

	if (spdk_likely(!vhost_dev_has_feature(vsession, VHOST_F_LOG_ALL))) {
		return;
	}

	offset = offsetof(struct vring_used, idx);
	len = sizeof(virtqueue->vring.used->idx);
	vq_idx = virtqueue - vsession->virtqueue;

	rte_vhost_log_used_vring(vsession->vid, vq_idx, offset, len);
}

/*
 * Get available requests from avail ring.
 */
uint16_t
vhost_vq_avail_ring_get(struct spdk_vhost_virtqueue *virtqueue, uint16_t *reqs,
			uint16_t reqs_len)
{
	struct rte_vhost_vring *vring = &virtqueue->vring;
	struct vring_avail *avail = vring->avail;
	uint16_t size_mask = vring->size - 1;
	uint16_t last_idx = virtqueue->last_avail_idx, avail_idx = avail->idx;
	uint16_t count, i;

	count = avail_idx - last_idx;
	if (spdk_likely(count == 0)) {
		return 0;
	}

	if (spdk_unlikely(count > vring->size)) {
		/* TODO: the queue is unrecoverably broken and should be marked so.
		 * For now we will fail silently and report there are no new avail entries.
		 */
		return 0;
	}

	count = spdk_min(count, reqs_len);
	virtqueue->last_avail_idx += count;
	for (i = 0; i < count; i++) {
		reqs[i] = vring->avail->ring[(last_idx + i) & size_mask];
	}

	SPDK_DEBUGLOG(SPDK_LOG_VHOST_RING,
		      "AVAIL: last_idx=%"PRIu16" avail_idx=%"PRIu16" count=%"PRIu16"\n",
		      last_idx, avail_idx, count);

	return count;
}

static bool
vhost_vring_desc_is_indirect(struct vring_desc *cur_desc)
{
	return !!(cur_desc->flags & VRING_DESC_F_INDIRECT);
}

int
vhost_vq_get_desc(struct spdk_vhost_session *vsession, struct spdk_vhost_virtqueue *virtqueue,
		  uint16_t req_idx, struct vring_desc **desc, struct vring_desc **desc_table,
		  uint32_t *desc_table_size)
{
	if (spdk_unlikely(req_idx >= virtqueue->vring.size)) {
		return -1;
	}

	*desc = &virtqueue->vring.desc[req_idx];

	if (vhost_vring_desc_is_indirect(*desc)) {
		*desc_table_size = (*desc)->len / sizeof(**desc);
		*desc_table = vhost_gpa_to_vva(vsession, (*desc)->addr,
					       sizeof(**desc) * *desc_table_size);
		*desc = *desc_table;
		if (*desc == NULL) {
			return -1;
		}

		return 0;
	}

	*desc_table = virtqueue->vring.desc;
	*desc_table_size = virtqueue->vring.size;

	return 0;
}

int
vhost_vq_used_signal(struct spdk_vhost_session *vsession,
		     struct spdk_vhost_virtqueue *virtqueue)
{
	if (virtqueue->used_req_cnt == 0) {
		return 0;
	}

	virtqueue->req_cnt += virtqueue->used_req_cnt;
	virtqueue->used_req_cnt = 0;

	SPDK_DEBUGLOG(SPDK_LOG_VHOST_RING,
		      "Queue %td - USED RING: sending IRQ: last used %"PRIu16"\n",
		      virtqueue - vsession->virtqueue, virtqueue->last_used_idx);

	if (rte_vhost_vring_call(vsession->vid, virtqueue->vring_idx) == 0) {
		/* interrupt signalled */
		return 1;
	} else {
		/* interrupt not signalled */
		return 0;
	}
}


static void
check_session_io_stats(struct spdk_vhost_session *vsession, uint64_t now)
{
	struct spdk_vhost_virtqueue *virtqueue;
	uint32_t irq_delay_base = vsession->coalescing_delay_time_base;
	uint32_t io_threshold = vsession->coalescing_io_rate_threshold;
	int32_t irq_delay;
	uint32_t req_cnt;
	uint16_t q_idx;

	if (now < vsession->next_stats_check_time) {
		return;
	}

	vsession->next_stats_check_time = now + vsession->stats_check_interval;
	for (q_idx = 0; q_idx < vsession->max_queues; q_idx++) {
		virtqueue = &vsession->virtqueue[q_idx];

		req_cnt = virtqueue->req_cnt + virtqueue->used_req_cnt;
		if (req_cnt <= io_threshold) {
			continue;
		}

		irq_delay = (irq_delay_base * (req_cnt - io_threshold)) / io_threshold;
		virtqueue->irq_delay_time = (uint32_t) spdk_max(0, irq_delay);

		virtqueue->req_cnt = 0;
		virtqueue->next_event_time = now;
	}
}

void
vhost_session_used_signal(struct spdk_vhost_session *vsession)
{
	struct spdk_vhost_virtqueue *virtqueue;
	uint64_t now;
	uint16_t q_idx;

	if (vsession->coalescing_delay_time_base == 0) {
		for (q_idx = 0; q_idx < vsession->max_queues; q_idx++) {
			virtqueue = &vsession->virtqueue[q_idx];

			if (virtqueue->vring.desc == NULL ||
			    (virtqueue->vring.avail->flags & VRING_AVAIL_F_NO_INTERRUPT)) {
				continue;
			}

			vhost_vq_used_signal(vsession, virtqueue);
		}
	} else {
		now = spdk_get_ticks();
		check_session_io_stats(vsession, now);

		for (q_idx = 0; q_idx < vsession->max_queues; q_idx++) {
			virtqueue = &vsession->virtqueue[q_idx];

			/* No need for event right now */
			if (now < virtqueue->next_event_time ||
			    (virtqueue->vring.avail->flags & VRING_AVAIL_F_NO_INTERRUPT)) {
				continue;
			}

			if (!vhost_vq_used_signal(vsession, virtqueue)) {
				continue;
			}

			/* Syscall is quite long so update time */
			now = spdk_get_ticks();
			virtqueue->next_event_time = now + virtqueue->irq_delay_time;
		}
	}
}

static int
vhost_session_set_coalescing(struct spdk_vhost_dev *vdev,
			     struct spdk_vhost_session *vsession, void *ctx)
{
	vsession->coalescing_delay_time_base =
		vdev->coalescing_delay_us * spdk_get_ticks_hz() / 1000000ULL;
	vsession->coalescing_io_rate_threshold =
		vdev->coalescing_iops_threshold * SPDK_VHOST_STATS_CHECK_INTERVAL_MS / 1000U;
	return 0;
}

static int
vhost_dev_set_coalescing(struct spdk_vhost_dev *vdev, uint32_t delay_base_us,
			 uint32_t iops_threshold)
{
	uint64_t delay_time_base = delay_base_us * spdk_get_ticks_hz() / 1000000ULL;
	uint32_t io_rate = iops_threshold * SPDK_VHOST_STATS_CHECK_INTERVAL_MS / 1000U;

	if (delay_time_base >= UINT32_MAX) {
		SPDK_ERRLOG("Delay time of %"PRIu32" is to big\n", delay_base_us);
		return -EINVAL;
	} else if (io_rate == 0) {
		SPDK_ERRLOG("IOPS rate of %"PRIu32" is too low. Min is %u\n", io_rate,
			    1000U / SPDK_VHOST_STATS_CHECK_INTERVAL_MS);
		return -EINVAL;
	}

	vdev->coalescing_delay_us = delay_base_us;
	vdev->coalescing_iops_threshold = iops_threshold;
	return 0;
}

int
spdk_vhost_set_coalescing(struct spdk_vhost_dev *vdev, uint32_t delay_base_us,
			  uint32_t iops_threshold)
{
	int rc;

	rc = vhost_dev_set_coalescing(vdev, delay_base_us, iops_threshold);
	if (rc != 0) {
		return rc;
	}

	vhost_dev_foreach_session(vdev, vhost_session_set_coalescing, NULL, NULL);
	return 0;
}

void
spdk_vhost_get_coalescing(struct spdk_vhost_dev *vdev, uint32_t *delay_base_us,
			  uint32_t *iops_threshold)
{
	if (delay_base_us) {
		*delay_base_us = vdev->coalescing_delay_us;
	}

	if (iops_threshold) {
		*iops_threshold = vdev->coalescing_iops_threshold;
	}
}

/*
 * Enqueue id and len to used ring.
 */
void
vhost_vq_used_ring_enqueue(struct spdk_vhost_session *vsession,
			   struct spdk_vhost_virtqueue *virtqueue,
			   uint16_t id, uint32_t len)
{
	struct rte_vhost_vring *vring = &virtqueue->vring;
	struct vring_used *used = vring->used;
	uint16_t last_idx = virtqueue->last_used_idx & (vring->size - 1);

	SPDK_DEBUGLOG(SPDK_LOG_VHOST_RING,
		      "Queue %td - USED RING: last_idx=%"PRIu16" req id=%"PRIu16" len=%"PRIu32"\n",
		      virtqueue - vsession->virtqueue, virtqueue->last_used_idx, id, len);

	vhost_log_req_desc(vsession, virtqueue, id);

	virtqueue->last_used_idx++;
	used->ring[last_idx].id = id;
	used->ring[last_idx].len = len;

	/* Ensure the used ring is updated before we log it or increment used->idx. */
	spdk_smp_wmb();

	vhost_log_used_vring_elem(vsession, virtqueue, last_idx);
	* (volatile uint16_t *) &used->idx = virtqueue->last_used_idx;
	vhost_log_used_vring_idx(vsession, virtqueue);

	virtqueue->used_req_cnt++;
}

int
vhost_vring_desc_get_next(struct vring_desc **desc,
			  struct vring_desc *desc_table, uint32_t desc_table_size)
{
	struct vring_desc *old_desc = *desc;
	uint16_t next_idx;

	if ((old_desc->flags & VRING_DESC_F_NEXT) == 0) {
		*desc = NULL;
		return 0;
	}

	next_idx = old_desc->next;
	if (spdk_unlikely(next_idx >= desc_table_size)) {
		*desc = NULL;
		return -1;
	}

	*desc = &desc_table[next_idx];
	return 0;
}

bool
vhost_vring_desc_is_wr(struct vring_desc *cur_desc)
{
	return !!(cur_desc->flags & VRING_DESC_F_WRITE);
}

int
vhost_vring_desc_to_iov(struct spdk_vhost_session *vsession, struct iovec *iov,
			uint16_t *iov_index, const struct vring_desc *desc)
{
	uint64_t len;
	uint64_t remaining = desc->len;
	uintptr_t payload = desc->addr;
	uintptr_t vva;

	do {
		if (*iov_index >= SPDK_VHOST_IOVS_MAX) {
			SPDK_ERRLOG("SPDK_VHOST_IOVS_MAX(%d) reached\n", SPDK_VHOST_IOVS_MAX);
			return -1;
		}
		len = remaining;
		vva = (uintptr_t)rte_vhost_va_from_guest_pa(vsession->mem, payload, &len);
		if (vva == 0 || len == 0) {
			SPDK_ERRLOG("gpa_to_vva(%p) == NULL\n", (void *)payload);
			return -1;
		}
		iov[*iov_index].iov_base = (void *)vva;
		iov[*iov_index].iov_len = len;
		remaining -= len;
		payload += len;
		(*iov_index)++;
	} while (remaining);

	return 0;
}

static struct spdk_vhost_session *
vhost_session_find_by_id(struct spdk_vhost_dev *vdev, unsigned id)
{
	struct spdk_vhost_session *vsession;

	TAILQ_FOREACH(vsession, &vdev->vsessions, tailq) {
		if (vsession->id == id) {
			return vsession;
		}
	}

	return NULL;
}

struct spdk_vhost_session *
vhost_session_find_by_vid(int vid)
{
	struct spdk_vhost_dev *vdev;
	struct spdk_vhost_session *vsession;

	TAILQ_FOREACH(vdev, &g_vhost_devices, tailq) {
		TAILQ_FOREACH(vsession, &vdev->vsessions, tailq) {
			if (vsession->vid == vid) {
				return vsession;
			}
		}
	}

	return NULL;
}

#define SHIFT_2MB	21
#define SIZE_2MB	(1ULL << SHIFT_2MB)
#define FLOOR_2MB(x)	(((uintptr_t)x) / SIZE_2MB) << SHIFT_2MB
#define CEIL_2MB(x)	((((uintptr_t)x) + SIZE_2MB - 1) / SIZE_2MB) << SHIFT_2MB

static void
vhost_session_mem_register(struct spdk_vhost_session *vsession)
{
	struct rte_vhost_mem_region *region;
	uint32_t i;
	uint64_t previous_start = UINT64_MAX;

	for (i = 0; i < vsession->mem->nregions; i++) {
		uint64_t start, end, len;
		region = &vsession->mem->regions[i];
		start = FLOOR_2MB(region->mmap_addr);
		end = CEIL_2MB(region->mmap_addr + region->mmap_size);
		if (start == previous_start) {
			start += (size_t) SIZE_2MB;
		}
		previous_start = start;
		len = end - start;
		SPDK_INFOLOG(SPDK_LOG_VHOST, "Registering VM memory for vtophys translation - 0x%jx len:0x%jx\n",
			     start, len);

		if (spdk_mem_register((void *)start, len) != 0) {
			SPDK_WARNLOG("Failed to register memory region %"PRIu32". Future vtophys translation might fail.\n",
				     i);
			continue;
		}
	}
}

static void
vhost_session_mem_unregister(struct spdk_vhost_session *vsession)
{
	struct rte_vhost_mem_region *region;
	uint32_t i;
	uint64_t previous_start = UINT64_MAX;

	for (i = 0; i < vsession->mem->nregions; i++) {
		uint64_t start, end, len;
		region = &vsession->mem->regions[i];
		start = FLOOR_2MB(region->mmap_addr);
		end = CEIL_2MB(region->mmap_addr + region->mmap_size);
		if (start == previous_start) {
			start += (size_t) SIZE_2MB;
		}
		previous_start = start;
		len = end - start;

		if (spdk_vtophys((void *) start, NULL) == SPDK_VTOPHYS_ERROR) {
			continue; /* region has not been registered */
		}

		if (spdk_mem_unregister((void *)start, len) != 0) {
			assert(false);
		}
	}

}

struct spdk_vhost_dev *
spdk_vhost_dev_next(struct spdk_vhost_dev *vdev)
{
	if (vdev == NULL) {
		return TAILQ_FIRST(&g_vhost_devices);
	}

	return TAILQ_NEXT(vdev, tailq);
}

struct spdk_vhost_dev *
spdk_vhost_dev_find(const char *ctrlr_name)
{
	struct spdk_vhost_dev *vdev;
	size_t dev_dirname_len = strlen(dev_dirname);

	if (strncmp(ctrlr_name, dev_dirname, dev_dirname_len) == 0) {
		ctrlr_name += dev_dirname_len;
	}

	TAILQ_FOREACH(vdev, &g_vhost_devices, tailq) {
		if (strcmp(vdev->name, ctrlr_name) == 0) {
			return vdev;
		}
	}

	return NULL;
}

static int
vhost_parse_core_mask(const char *mask, struct spdk_cpuset *cpumask)
{
	int rc;

	if (cpumask == NULL) {
		return -1;
	}

	if (mask == NULL) {
		spdk_cpuset_copy(cpumask, spdk_app_get_core_mask());
		return 0;
	}

	rc = spdk_app_parse_core_mask(mask, cpumask);
	if (rc < 0) {
		SPDK_ERRLOG("invalid cpumask %s\n", mask);
		return -1;
	}

	if (spdk_cpuset_count(cpumask) == 0) {
		SPDK_ERRLOG("no cpu is selected among reactor mask(=%s)\n",
			    spdk_cpuset_fmt(spdk_app_get_core_mask()));
		return -1;
	}

	return 0;
}

static void *
_start_rte_driver(void *arg)
{
	char *path = arg;

	if (rte_vhost_driver_start(path) != 0) {
		return NULL;
	}

	return path;
}

int
vhost_dev_register(struct spdk_vhost_dev *vdev, const char *name, const char *mask_str,
		   const struct spdk_vhost_dev_backend *backend)
{
	char path[PATH_MAX];
	struct stat file_stat;
	struct spdk_cpuset *cpumask;
	int rc;

	assert(vdev);
	if (name == NULL) {
		SPDK_ERRLOG("Can't register controller with no name\n");
		return -EINVAL;
	}

	cpumask = spdk_cpuset_alloc();
	if (!cpumask) {
		SPDK_ERRLOG("spdk_cpuset_alloc failed\n");
		return -ENOMEM;
	}

	if (vhost_parse_core_mask(mask_str, cpumask) != 0) {
		SPDK_ERRLOG("cpumask %s is invalid (app mask is 0x%s)\n",
			    mask_str, spdk_cpuset_fmt(spdk_app_get_core_mask()));
		rc = -EINVAL;
		goto out;
	}

	if (spdk_vhost_dev_find(name)) {
		SPDK_ERRLOG("vhost controller %s already exists.\n", name);
		rc = -EEXIST;
		goto out;
	}

	if (snprintf(path, sizeof(path), "%s%s", dev_dirname, name) >= (int)sizeof(path)) {
		SPDK_ERRLOG("Resulting socket path for controller %s is too long: %s%s\n", name, dev_dirname,
			    name);
		rc = -EINVAL;
		goto out;
	}

	/* Register vhost driver to handle vhost messages. */
	if (stat(path, &file_stat) != -1) {
		if (!S_ISSOCK(file_stat.st_mode)) {
			SPDK_ERRLOG("Cannot create a domain socket at path \"%s\": "
				    "The file already exists and is not a socket.\n",
				    path);
			rc = -EIO;
			goto out;
		} else if (unlink(path) != 0) {
			SPDK_ERRLOG("Cannot create a domain socket at path \"%s\": "
				    "The socket already exists and failed to unlink.\n",
				    path);
			rc = -EIO;
			goto out;
		}
	}

	if (rte_vhost_driver_register(path, 0) != 0) {
		SPDK_ERRLOG("Could not register controller %s with vhost library\n", name);
		SPDK_ERRLOG("Check if domain socket %s already exists\n", path);
		rc = -EIO;
		goto out;
	}
	if (rte_vhost_driver_set_features(path, backend->virtio_features) ||
	    rte_vhost_driver_disable_features(path, backend->disabled_features)) {
		SPDK_ERRLOG("Couldn't set vhost features for controller %s\n", name);

		rte_vhost_driver_unregister(path);
		rc = -EIO;
		goto out;
	}

	if (rte_vhost_driver_callback_register(path, &g_spdk_vhost_ops) != 0) {
		rte_vhost_driver_unregister(path);
		SPDK_ERRLOG("Couldn't register callbacks for controller %s\n", name);
		rc = -EIO;
		goto out;
	}

	vdev->name = strdup(name);
	vdev->path = strdup(path);
	if (vdev->name == NULL || vdev->path == NULL) {
		free(vdev->name);
		free(vdev->path);
		rte_vhost_driver_unregister(path);
		rc = -EIO;
		goto out;
	}

	vdev->cpumask = cpumask;
	vdev->registered = true;
	vdev->backend = backend;
	TAILQ_INIT(&vdev->vsessions);
	TAILQ_INSERT_TAIL(&g_vhost_devices, vdev, tailq);

	vhost_dev_set_coalescing(vdev, SPDK_VHOST_COALESCING_DELAY_BASE_US,
				 SPDK_VHOST_VQ_IOPS_COALESCING_THRESHOLD);

	vhost_dev_install_rte_compat_hooks(vdev);

	/* The following might start a POSIX thread that polls for incoming
	 * socket connections and calls backend->start/stop_device. These backend
	 * callbacks are also protected by the global SPDK vhost mutex, so we're
	 * safe with not initializing the vdev just yet.
	 */
	if (spdk_call_unaffinitized(_start_rte_driver, path) == NULL) {
		SPDK_ERRLOG("Failed to start vhost driver for controller %s (%d): %s\n",
			    name, errno, spdk_strerror(errno));
		rte_vhost_driver_unregister(path);
		TAILQ_REMOVE(&g_vhost_devices, vdev, tailq);
		free(vdev->name);
		free(vdev->path);
		rc = -EIO;
		goto out;
	}

	SPDK_INFOLOG(SPDK_LOG_VHOST, "Controller %s: new controller added\n", vdev->name);
	return 0;

out:
	spdk_cpuset_free(cpumask);
	return rc;
}

int
vhost_dev_unregister(struct spdk_vhost_dev *vdev)
{
	if (!TAILQ_EMPTY(&vdev->vsessions)) {
		SPDK_ERRLOG("Controller %s has still valid connection.\n", vdev->name);
		return -EBUSY;
	}

	if (vdev->registered && rte_vhost_driver_unregister(vdev->path) != 0) {
		SPDK_ERRLOG("Could not unregister controller %s with vhost library\n"
			    "Check if domain socket %s still exists\n",
			    vdev->name, vdev->path);
		return -EIO;
	}

	SPDK_INFOLOG(SPDK_LOG_VHOST, "Controller %s: removed\n", vdev->name);

	free(vdev->name);
	free(vdev->path);
	spdk_cpuset_free(vdev->cpumask);
	TAILQ_REMOVE(&g_vhost_devices, vdev, tailq);
	return 0;
}

static struct spdk_vhost_session *
vhost_session_next(struct spdk_vhost_dev *vdev, unsigned prev_id)
{
	struct spdk_vhost_session *vsession;

	TAILQ_FOREACH(vsession, &vdev->vsessions, tailq) {
		if (vsession->id > prev_id) {
			return vsession;
		}
	}

	return NULL;
}

const char *
spdk_vhost_dev_get_name(struct spdk_vhost_dev *vdev)
{
	assert(vdev != NULL);
	return vdev->name;
}

const struct spdk_cpuset *
spdk_vhost_dev_get_cpumask(struct spdk_vhost_dev *vdev)
{
	assert(vdev != NULL);
	return vdev->cpumask;
}

struct vhost_poll_group *
vhost_get_poll_group(struct spdk_cpuset *cpumask)
{
	struct vhost_poll_group *pg, *selected_pg;
	uint32_t min_ctrlrs;

	min_ctrlrs = INT_MAX;
	selected_pg = TAILQ_FIRST(&g_poll_groups);

	TAILQ_FOREACH(pg, &g_poll_groups, tailq) {
		spdk_cpuset_copy(g_tmp_cpuset, cpumask);
		spdk_cpuset_and(g_tmp_cpuset, spdk_thread_get_cpumask(pg->thread));

		/* ignore threads which could be relocated to a non-masked cpu. */
		if (!spdk_cpuset_equal(g_tmp_cpuset, spdk_thread_get_cpumask(pg->thread))) {
			continue;
		}

		if (pg->ref < min_ctrlrs) {
			selected_pg = pg;
			min_ctrlrs = pg->ref;
		}
	}

	assert(selected_pg != NULL);
	return selected_pg;
}

static struct vhost_poll_group *
_get_current_poll_group(void)
{
	struct vhost_poll_group *pg;
	struct spdk_thread *cur_thread = spdk_get_thread();

	TAILQ_FOREACH(pg, &g_poll_groups, tailq) {
		if (pg->thread == cur_thread) {
			return pg;
		}
	}

	return NULL;
}

static void
wait_for_semaphore(int timeout_sec, const char *errmsg)
{
	struct timespec timeout;
	int rc;

	clock_gettime(CLOCK_REALTIME, &timeout);
	timeout.tv_sec += timeout_sec;
	rc = sem_timedwait(&g_dpdk_sem, &timeout);
	if (rc != 0) {
		SPDK_ERRLOG("Timeout waiting for event: %s.\n", errmsg);
		sem_wait(&g_dpdk_sem);
	}
}

static void
vhost_session_cb_done(int rc)
{
	g_dpdk_response = rc;
	sem_post(&g_dpdk_sem);
}

void
vhost_session_start_done(struct spdk_vhost_session *vsession, int response)
{
	if (response == 0) {
		vsession->started = true;
		vsession->poll_group = _get_current_poll_group();
		assert(vsession->poll_group != NULL);
		assert(vsession->poll_group->ref < UINT_MAX);
		vsession->poll_group->ref++;

		assert(vsession->vdev->active_session_num < UINT32_MAX);
		vsession->vdev->active_session_num++;
	}

	vhost_session_cb_done(response);
}

void
vhost_session_stop_done(struct spdk_vhost_session *vsession, int response)
{
	if (response == 0) {
		vsession->started = false;
		assert(vsession->poll_group != NULL);
		assert(vsession->poll_group->ref > 0);
		vsession->poll_group->ref--;
		vsession->poll_group = NULL;

		assert(vsession->vdev->active_session_num > 0);
		vsession->vdev->active_session_num--;
	}

	vhost_session_cb_done(response);
}

static void
vhost_event_cb(void *arg1)
{
	struct vhost_session_fn_ctx *ctx = arg1;
	struct spdk_vhost_session *vsession;

	if (pthread_mutex_trylock(&g_vhost_mutex) != 0) {
		spdk_thread_send_msg(spdk_get_thread(), vhost_event_cb, arg1);
		return;
	}

	vsession = vhost_session_find_by_id(ctx->vdev, ctx->vsession_id);
	ctx->cb_fn(ctx->vdev, vsession, NULL);
	pthread_mutex_unlock(&g_vhost_mutex);
}

int
vhost_session_send_event(struct vhost_poll_group *pg,
			 struct spdk_vhost_session *vsession,
			 spdk_vhost_session_fn cb_fn, unsigned timeout_sec,
			 const char *errmsg)
{
	struct vhost_session_fn_ctx ev_ctx = {0};

	ev_ctx.vdev = vsession->vdev;
	ev_ctx.vsession_id = vsession->id;
	ev_ctx.cb_fn = cb_fn;

	spdk_thread_send_msg(pg->thread, vhost_event_cb, &ev_ctx);

	pthread_mutex_unlock(&g_vhost_mutex);
	wait_for_semaphore(timeout_sec, errmsg);
	pthread_mutex_lock(&g_vhost_mutex);

	return g_dpdk_response;
}

static void foreach_session_continue(struct vhost_session_fn_ctx *ev_ctx,
				     struct spdk_vhost_session *vsession);

static void
foreach_session_finish_cb(void *arg1)
{
	struct vhost_session_fn_ctx *ctx = arg1;
	struct spdk_vhost_dev *vdev = ctx->vdev;

	if (pthread_mutex_trylock(&g_vhost_mutex) != 0) {
		spdk_thread_send_msg(spdk_get_thread(),
				     foreach_session_finish_cb, arg1);
		return;
	}

	assert(vdev->pending_async_op_num > 0);
	vdev->pending_async_op_num--;
	if (ctx->cpl_fn != NULL) {
		ctx->cpl_fn(vdev, ctx->user_ctx);
	}

	pthread_mutex_unlock(&g_vhost_mutex);
	free(ctx);
}

static void
foreach_session_continue_cb(void *arg1)
{
	struct vhost_session_fn_ctx *ctx = arg1;
	struct spdk_vhost_session *vsession = NULL;
	struct spdk_vhost_dev *vdev = ctx->vdev;
	int rc;

	if (pthread_mutex_trylock(&g_vhost_mutex) != 0) {
		spdk_thread_send_msg(spdk_get_thread(),
				     foreach_session_continue_cb, arg1);
		return;
	}

	vsession = vhost_session_find_by_id(vdev, ctx->vsession_id);
	if (vsession == NULL || !vsession->initialized) {
		/* The session must have been removed in the meantime, so we
		 * just skip it in our foreach chain
		 */
		goto out_unlock_continue;
	}

	if (vsession->started && vsession->poll_group->thread != spdk_get_thread()) {
		/* if session has been relocated to other thread, it is no longer thread-safe
		 * to access its contents here. Even though we're running under the global
		 * vhost mutex, the session itself (and its pollers) are not. We need to chase
		 * the session thread as many times as necessary.
		 */
		spdk_thread_send_msg(vsession->poll_group->thread,
				     foreach_session_continue_cb, arg1);
		pthread_mutex_unlock(&g_vhost_mutex);
		return;
	}

	rc = ctx->cb_fn(vdev, vsession, ctx->user_ctx);
	if (rc < 0) {
		pthread_mutex_unlock(&g_vhost_mutex);
		free(ctx);
		return;
	}

out_unlock_continue:
	vsession = vhost_session_next(vdev, ctx->vsession_id);
	foreach_session_continue(ctx, vsession);
	pthread_mutex_unlock(&g_vhost_mutex);
}

static void
foreach_session_continue(struct vhost_session_fn_ctx *ev_ctx,
			 struct spdk_vhost_session *vsession)
{
	struct spdk_vhost_dev *vdev = ev_ctx->vdev;
	int rc;

	while (vsession != NULL && !vsession->started) {
		if (vsession->initialized) {
			rc = ev_ctx->cb_fn(vdev, vsession, ev_ctx->user_ctx);
			if (rc < 0) {
				return;
			}
		}

		vsession = vhost_session_next(vdev, vsession->id);
	}

	if (vsession != NULL) {
		ev_ctx->vsession_id = vsession->id;
		spdk_thread_send_msg(vsession->poll_group->thread,
				     foreach_session_continue_cb, ev_ctx);
	} else {
		ev_ctx->vsession_id = UINT32_MAX;
		spdk_thread_send_msg(g_vhost_init_thread,
				     foreach_session_finish_cb, ev_ctx);
	}
}

void
vhost_dev_foreach_session(struct spdk_vhost_dev *vdev,
			  spdk_vhost_session_fn fn,
			  spdk_vhost_dev_fn cpl_fn,
			  void *arg)
{
	struct spdk_vhost_session *vsession = TAILQ_FIRST(&vdev->vsessions);
	struct vhost_session_fn_ctx *ev_ctx;

	ev_ctx = calloc(1, sizeof(*ev_ctx));
	if (ev_ctx == NULL) {
		SPDK_ERRLOG("Failed to alloc vhost event.\n");
		assert(false);
		return;
	}

	ev_ctx->vdev = vdev;
	ev_ctx->cb_fn = fn;
	ev_ctx->cpl_fn = cpl_fn;
	ev_ctx->user_ctx = arg;

	assert(vdev->pending_async_op_num < UINT32_MAX);
	vdev->pending_async_op_num++;
	foreach_session_continue(ev_ctx, vsession);
}

static void
_stop_session(struct spdk_vhost_session *vsession)
{
	struct spdk_vhost_dev *vdev = vsession->vdev;
	struct spdk_vhost_virtqueue *q;
	int rc;
	uint16_t i;

	rc = vdev->backend->stop_session(vsession);
	if (rc != 0) {
		SPDK_ERRLOG("Couldn't stop device with vid %d.\n", vsession->vid);
		pthread_mutex_unlock(&g_vhost_mutex);
		return;
	}

	for (i = 0; i < vsession->max_queues; i++) {
		q = &vsession->virtqueue[i];
		if (q->vring.desc == NULL) {
			continue;
		}
		rte_vhost_set_vring_base(vsession->vid, i, q->last_avail_idx, q->last_used_idx);
	}

	vhost_session_mem_unregister(vsession);
	free(vsession->mem);
}

static void
stop_device(int vid)
{
	struct spdk_vhost_session *vsession;

	pthread_mutex_lock(&g_vhost_mutex);
	vsession = vhost_session_find_by_vid(vid);
	if (vsession == NULL) {
		SPDK_ERRLOG("Couldn't find session with vid %d.\n", vid);
		pthread_mutex_unlock(&g_vhost_mutex);
		return;
	}

	if (!vsession->started) {
		/* already stopped, nothing to do */
		pthread_mutex_unlock(&g_vhost_mutex);
		return;
	}

	_stop_session(vsession);
	pthread_mutex_unlock(&g_vhost_mutex);
}

static int
start_device(int vid)
{
	struct spdk_vhost_dev *vdev;
	struct spdk_vhost_session *vsession;
	int rc = -1;
	uint16_t i;

	pthread_mutex_lock(&g_vhost_mutex);

	vsession = vhost_session_find_by_vid(vid);
	if (vsession == NULL) {
		SPDK_ERRLOG("Couldn't find session with vid %d.\n", vid);
		goto out;
	}

	vdev = vsession->vdev;
	if (vsession->started) {
		/* already started, nothing to do */
		rc = 0;
		goto out;
	}

	vsession->max_queues = 0;
	memset(vsession->virtqueue, 0, sizeof(vsession->virtqueue));
	for (i = 0; i < SPDK_VHOST_MAX_VQUEUES; i++) {
		struct spdk_vhost_virtqueue *q = &vsession->virtqueue[i];

		q->vring_idx = -1;
		if (rte_vhost_get_vhost_vring(vid, i, &q->vring)) {
			continue;
		}
		q->vring_idx = i;

		if (q->vring.desc == NULL || q->vring.size == 0) {
			continue;
		}

		if (rte_vhost_get_vring_base(vsession->vid, i, &q->last_avail_idx, &q->last_used_idx)) {
			q->vring.desc = NULL;
			continue;
		}

		/* Disable I/O submission notifications, we'll be polling. */
		q->vring.used->flags = VRING_USED_F_NO_NOTIFY;
		vsession->max_queues = i + 1;
	}

	if (rte_vhost_get_negotiated_features(vid, &vsession->negotiated_features) != 0) {
		SPDK_ERRLOG("vhost device %d: Failed to get negotiated driver features\n", vid);
		goto out;
	}

	if (rte_vhost_get_mem_table(vid, &vsession->mem) != 0) {
		SPDK_ERRLOG("vhost device %d: Failed to get guest memory table\n", vid);
		goto out;
	}

	/*
	 * Not sure right now but this look like some kind of QEMU bug and guest IO
	 * might be frozed without kicking all queues after live-migration. This look like
	 * the previous vhost instance failed to effectively deliver all interrupts before
	 * the GET_VRING_BASE message. This shouldn't harm guest since spurious interrupts
	 * should be ignored by guest virtio driver.
	 *
	 * Tested on QEMU 2.10.91 and 2.11.50.
	 */
	for (i = 0; i < vsession->max_queues; i++) {
		struct spdk_vhost_virtqueue *q = &vsession->virtqueue[i];

		if (q->vring.desc != NULL && q->vring.size > 0) {
			rte_vhost_vring_call(vsession->vid, q->vring_idx);
		}
	}

	vhost_session_set_coalescing(vdev, vsession, NULL);
	vhost_session_mem_register(vsession);
	vsession->initialized = true;
	rc = vdev->backend->start_session(vsession);
	if (rc != 0) {
		vhost_session_mem_unregister(vsession);
		free(vsession->mem);
		goto out;
	}

out:
	pthread_mutex_unlock(&g_vhost_mutex);
	return rc;
}

#ifdef SPDK_CONFIG_VHOST_INTERNAL_LIB
static int
get_config(int vid, uint8_t *config, uint32_t len)
{
	struct spdk_vhost_session *vsession;
	struct spdk_vhost_dev *vdev;
	int rc = -1;

	pthread_mutex_lock(&g_vhost_mutex);
	vsession = vhost_session_find_by_vid(vid);
	if (vsession == NULL) {
		SPDK_ERRLOG("Couldn't find session with vid %d.\n", vid);
		goto out;
	}

	vdev = vsession->vdev;
	if (vdev->backend->vhost_get_config) {
		rc = vdev->backend->vhost_get_config(vdev, config, len);
	}

out:
	pthread_mutex_unlock(&g_vhost_mutex);
	return rc;
}

static int
set_config(int vid, uint8_t *config, uint32_t offset, uint32_t size, uint32_t flags)
{
	struct spdk_vhost_session *vsession;
	struct spdk_vhost_dev *vdev;
	int rc = -1;

	pthread_mutex_lock(&g_vhost_mutex);
	vsession = vhost_session_find_by_vid(vid);
	if (vsession == NULL) {
		SPDK_ERRLOG("Couldn't find session with vid %d.\n", vid);
		goto out;
	}

	vdev = vsession->vdev;
	if (vdev->backend->vhost_set_config) {
		rc = vdev->backend->vhost_set_config(vdev, config, offset, size, flags);
	}

out:
	pthread_mutex_unlock(&g_vhost_mutex);
	return rc;
}
#endif

int
spdk_vhost_set_socket_path(const char *basename)
{
	int ret;

	if (basename && strlen(basename) > 0) {
		ret = snprintf(dev_dirname, sizeof(dev_dirname) - 2, "%s", basename);
		if (ret <= 0) {
			return -EINVAL;
		}
		if ((size_t)ret >= sizeof(dev_dirname) - 2) {
			SPDK_ERRLOG("Char dev dir path length %d is too long\n", ret);
			return -EINVAL;
		}

		if (dev_dirname[ret - 1] != '/') {
			dev_dirname[ret] = '/';
			dev_dirname[ret + 1]  = '\0';
		}
	}

	return 0;
}

void
vhost_dump_info_json(struct spdk_vhost_dev *vdev, struct spdk_json_write_ctx *w)
{
	assert(vdev->backend->dump_info_json != NULL);
	vdev->backend->dump_info_json(vdev, w);
}

int
spdk_vhost_dev_remove(struct spdk_vhost_dev *vdev)
{
	if (vdev->pending_async_op_num) {
		return -EBUSY;
	}

	return vdev->backend->remove_device(vdev);
}

static int
new_connection(int vid)
{
	struct spdk_vhost_dev *vdev;
	struct spdk_vhost_session *vsession;
	char ifname[PATH_MAX];

	pthread_mutex_lock(&g_vhost_mutex);

	if (rte_vhost_get_ifname(vid, ifname, PATH_MAX) < 0) {
		SPDK_ERRLOG("Couldn't get a valid ifname for device with vid %d\n", vid);
		pthread_mutex_unlock(&g_vhost_mutex);
		return -1;
	}

	vdev = spdk_vhost_dev_find(ifname);
	if (vdev == NULL) {
		SPDK_ERRLOG("Couldn't find device with vid %d to create connection for.\n", vid);
		pthread_mutex_unlock(&g_vhost_mutex);
		return -1;
	}

	/* We expect sessions inside vdev->vsessions to be sorted in ascending
	 * order in regard of vsession->id. For now we always set id = vsessions_cnt++
	 * and append each session to the very end of the vsessions list.
	 * This is required for spdk_vhost_dev_foreach_session() to work.
	 */
	if (vdev->vsessions_num == UINT_MAX) {
		assert(false);
		return -EINVAL;
	}

	if (posix_memalign((void **)&vsession, SPDK_CACHE_LINE_SIZE, sizeof(*vsession) +
			   vdev->backend->session_ctx_size)) {
		SPDK_ERRLOG("vsession alloc failed\n");
		pthread_mutex_unlock(&g_vhost_mutex);
		return -1;
	}
	memset(vsession, 0, sizeof(*vsession) + vdev->backend->session_ctx_size);

	vsession->vdev = vdev;
	vsession->vid = vid;
	vsession->id = vdev->vsessions_num++;
	vsession->name = spdk_sprintf_alloc("%ss%u", vdev->name, vsession->vid);
	if (vsession->name == NULL) {
		SPDK_ERRLOG("vsession alloc failed\n");
		pthread_mutex_unlock(&g_vhost_mutex);
		free(vsession);
		return -1;
	}
	vsession->poll_group = NULL;
	vsession->started = false;
	vsession->initialized = false;
	vsession->next_stats_check_time = 0;
	vsession->stats_check_interval = SPDK_VHOST_STATS_CHECK_INTERVAL_MS *
					 spdk_get_ticks_hz() / 1000UL;
	TAILQ_INSERT_TAIL(&vdev->vsessions, vsession, tailq);

	vhost_session_install_rte_compat_hooks(vsession);
	pthread_mutex_unlock(&g_vhost_mutex);
	return 0;
}

static void
destroy_connection(int vid)
{
	struct spdk_vhost_session *vsession;

	pthread_mutex_lock(&g_vhost_mutex);
	vsession = vhost_session_find_by_vid(vid);
	if (vsession == NULL) {
		SPDK_ERRLOG("Couldn't find session with vid %d.\n", vid);
		pthread_mutex_unlock(&g_vhost_mutex);
		return;
	}

	if (vsession->started) {
		_stop_session(vsession);
	}

	TAILQ_REMOVE(&vsession->vdev->vsessions, vsession, tailq);
	free(vsession->name);
	free(vsession);
	pthread_mutex_unlock(&g_vhost_mutex);
}

void
spdk_vhost_lock(void)
{
	pthread_mutex_lock(&g_vhost_mutex);
}

int
spdk_vhost_trylock(void)
{
	return -pthread_mutex_trylock(&g_vhost_mutex);
}

void
spdk_vhost_unlock(void)
{
	pthread_mutex_unlock(&g_vhost_mutex);
}

static void
vhost_create_poll_group_done(void *ctx)
{
	spdk_vhost_init_cb init_cb = ctx;
	int ret;

	if (TAILQ_EMPTY(&g_poll_groups)) {
		/* No threads? Iteration failed? */
		init_cb(-ECHILD);
		return;
	}

	ret = vhost_scsi_controller_construct();
	if (ret != 0) {
		SPDK_ERRLOG("Cannot construct vhost controllers\n");
		goto out;
	}

	ret = vhost_blk_controller_construct();
	if (ret != 0) {
		SPDK_ERRLOG("Cannot construct vhost block controllers\n");
		goto out;
	}

#ifdef SPDK_CONFIG_VHOST_INTERNAL_LIB
	ret = vhost_nvme_controller_construct();
	if (ret != 0) {
		SPDK_ERRLOG("Cannot construct vhost NVMe controllers\n");
		goto out;
	}
#endif

out:
	init_cb(ret);
}

static void
vhost_create_poll_group(void *ctx)
{
	struct vhost_poll_group *pg;

	pg = calloc(1, sizeof(*pg));
	if (!pg) {
		SPDK_ERRLOG("Not enough memory to allocate poll groups\n");
		spdk_app_stop(-ENOMEM);
		return;
	}

	pg->thread = spdk_get_thread();
	TAILQ_INSERT_TAIL(&g_poll_groups, pg, tailq);
}

void
spdk_vhost_init(spdk_vhost_init_cb init_cb)
{
	size_t len;
	int ret;

	g_vhost_init_thread = spdk_get_thread();
	assert(g_vhost_init_thread != NULL);

	if (dev_dirname[0] == '\0') {
		if (getcwd(dev_dirname, sizeof(dev_dirname) - 1) == NULL) {
			SPDK_ERRLOG("getcwd failed (%d): %s\n", errno, spdk_strerror(errno));
			ret = -1;
			goto err_out;
		}

		len = strlen(dev_dirname);
		if (dev_dirname[len - 1] != '/') {
			dev_dirname[len] = '/';
			dev_dirname[len + 1] = '\0';
		}
	}

	g_tmp_cpuset = spdk_cpuset_alloc();
	if (g_tmp_cpuset == NULL) {
		ret = -1;
		goto err_out;
	}

	ret = sem_init(&g_dpdk_sem, 0, 0);
	if (ret != 0) {
		SPDK_ERRLOG("Failed to initialize semaphore for rte_vhost pthread.\n");
		spdk_cpuset_free(g_tmp_cpuset);
		ret = -1;
		goto err_out;
	}

	spdk_for_each_thread(vhost_create_poll_group,
			     init_cb,
			     vhost_create_poll_group_done);
	return;
err_out:
	init_cb(ret);
}

static void
_spdk_vhost_fini(void *arg1)
{
	struct spdk_vhost_dev *vdev, *tmp;
	struct vhost_poll_group *pg, *tpg;

	spdk_vhost_lock();
	vdev = spdk_vhost_dev_next(NULL);
	while (vdev != NULL) {
		tmp = spdk_vhost_dev_next(vdev);
		spdk_vhost_dev_remove(vdev);
		/* don't care if it fails, there's nothing we can do for now */
		vdev = tmp;
	}
	spdk_vhost_unlock();

	/* All devices are removed now. */
	sem_destroy(&g_dpdk_sem);
	spdk_cpuset_free(g_tmp_cpuset);
	TAILQ_FOREACH_SAFE(pg, &g_poll_groups, tailq, tpg) {
		TAILQ_REMOVE(&g_poll_groups, pg, tailq);
		free(pg);
	}
	g_fini_cpl_cb();
}

static void *
session_shutdown(void *arg)
{
	struct spdk_vhost_dev *vdev = NULL;

	TAILQ_FOREACH(vdev, &g_vhost_devices, tailq) {
		rte_vhost_driver_unregister(vdev->path);
		vdev->registered = false;
	}

	SPDK_INFOLOG(SPDK_LOG_VHOST, "Exiting\n");
	spdk_thread_send_msg(g_vhost_init_thread, _spdk_vhost_fini, NULL);
	return NULL;
}

void
spdk_vhost_fini(spdk_vhost_fini_cb fini_cb)
{
	pthread_t tid;
	int rc;

	assert(spdk_get_thread() == g_vhost_init_thread);
	g_fini_cpl_cb = fini_cb;

	/* rte_vhost API for removing sockets is not asynchronous. Since it may call SPDK
	 * ops for stopping a device or removing a connection, we need to call it from
	 * a separate thread to avoid deadlock.
	 */
	rc = pthread_create(&tid, NULL, &session_shutdown, NULL);
	if (rc < 0) {
		SPDK_ERRLOG("Failed to start session shutdown thread (%d): %s\n", rc, spdk_strerror(rc));
		abort();
	}
	pthread_detach(tid);
}

void
spdk_vhost_config_json(struct spdk_json_write_ctx *w)
{
	struct spdk_vhost_dev *vdev;
	uint32_t delay_base_us;
	uint32_t iops_threshold;

	spdk_json_write_array_begin(w);

	spdk_vhost_lock();
	vdev = spdk_vhost_dev_next(NULL);
	while (vdev != NULL) {
		vdev->backend->write_config_json(vdev, w);

		spdk_vhost_get_coalescing(vdev, &delay_base_us, &iops_threshold);
		if (delay_base_us) {
			spdk_json_write_object_begin(w);
			spdk_json_write_named_string(w, "method", "vhost_controller_set_coalescing");

			spdk_json_write_named_object_begin(w, "params");
			spdk_json_write_named_string(w, "ctrlr", vdev->name);
			spdk_json_write_named_uint32(w, "delay_base_us", delay_base_us);
			spdk_json_write_named_uint32(w, "iops_threshold", iops_threshold);
			spdk_json_write_object_end(w);

			spdk_json_write_object_end(w);
		}
		vdev = spdk_vhost_dev_next(vdev);
	}
	spdk_vhost_unlock();

	spdk_json_write_array_end(w);
}

SPDK_LOG_REGISTER_COMPONENT("vhost", SPDK_LOG_VHOST)
SPDK_LOG_REGISTER_COMPONENT("vhost_ring", SPDK_LOG_VHOST_RING)