Simplify locking for protocols.

In an attempt to simplify the protocol implementation, and hopefully
track down a close related race, we've made it so that most protocols
need not worry about locks, and can access the socket lock if they do
need a lock.  They also let the socket manage their workers, for the
most part.  (The req protocol is special, since it needs a top level
work distributor, *and* a resender.)

7 files changed, 210 insertions, 93 deletions

diff --git a/src/core/defs.h b/src/core/defs.h
index 9ae734f7..98f4e661 100644
--- a/src/core/defs.h
+++ b/src/core/defs.h
@@ -18,24 +18,25 @@
 #define NNI_ARG_UNUSED(x)    ((void) x);
 
 // These types are common but have names shared with user space.
-typedef struct nng_socket	nni_sock;
-typedef struct nng_endpoint	nni_ep;
-typedef struct nng_pipe		nni_pipe;
-typedef struct nng_msg		nni_msg;
-typedef struct nng_sockaddr	nni_sockaddr;
+typedef struct nng_socket		nni_sock;
+typedef struct nng_endpoint		nni_ep;
+typedef struct nng_pipe			nni_pipe;
+typedef struct nng_msg			nni_msg;
+typedef struct nng_sockaddr		nni_sockaddr;
 
 // These are our own names.
-typedef struct nni_tran		nni_tran;
-typedef struct nni_tran_ep	nni_tran_ep;
-typedef struct nni_tran_pipe	nni_tran_pipe;
+typedef struct nni_tran			nni_tran;
+typedef struct nni_tran_ep		nni_tran_ep;
+typedef struct nni_tran_pipe		nni_tran_pipe;
 
-typedef struct nni_proto_pipe	nni_proto_pipe;
-typedef struct nni_proto	nni_proto;
+typedef struct nni_proto_sock_ops	nni_proto_sock_ops;
+typedef struct nni_proto_pipe_ops	nni_proto_pipe_ops;
+typedef struct nni_proto		nni_proto;
 
 
-typedef int			nni_signal;     // Turnstile/wakeup channel.
-typedef uint64_t		nni_time;       // Absolute time (usec).
-typedef int64_t			nni_duration;   // Relative time (usec).
+typedef int				nni_signal;     // Wakeup channel.
+typedef uint64_t			nni_time;       // Abs. time (usec).
+typedef int64_t				nni_duration;   // Rel. time (usec).
 
 // Used by transports for scatter gather I/O.
 typedef struct {
diff --git a/src/core/pipe.c b/src/core/pipe.c
index e49ad05a..b3a107ff 100644
--- a/src/core/pipe.c
+++ b/src/core/pipe.c
@@ -77,7 +77,7 @@ nni_pipe_destroy(nni_pipe *p)
 		p->p_tran_ops.pipe_destroy(p->p_tran_data);
 	}
 	if (p->p_proto_data != NULL) {
-		p->p_proto_ops.pipe_fini(p->p_proto_data);
+		p->p_sock->s_pipe_ops.pipe_fini(p->p_proto_data);
 	}
 	NNI_FREE_STRUCT(p);
 }
@@ -88,8 +88,8 @@ nni_pipe_create(nni_pipe **pp, nni_ep *ep)
 {
 	nni_pipe *p;
 	nni_sock *sock = ep->ep_sock;
-	nni_proto *proto = &sock->s_proto;
-	const nni_proto_pipe *pops = proto->proto_pipe;
+	const nni_proto_pipe_ops *ops = &sock->s_pipe_ops;
+	void *pdata;
 	int rv;
 
 	if ((p = NNI_ALLOC_STRUCT(p)) == NULL) {
@@ -105,23 +105,19 @@ nni_pipe_create(nni_pipe **pp, nni_ep *ep)
 	// and we avoid an extra dereference on hot code paths.
 	p->p_tran_ops = *ep->ep_tran->tran_pipe;
 
-	// Make a copy of the protocol ops.  Same rationale.
-	p->p_proto_ops = *pops;
-
-	if ((rv = pops->pipe_init(&p->p_proto_data, p, sock->s_data)) != 0) {
+	if ((rv = ops->pipe_init(&pdata, p, sock->s_data)) != 0) {
 		NNI_FREE_STRUCT(p);
 		return (rv);
 	}
-	rv = nni_thr_init(&p->p_recv_thr, pops->pipe_recv, p->p_proto_data);
-	if (rv != 0) {
-		pops->pipe_fini(&p->p_proto_data);
+	p->p_proto_data = pdata;
+	if ((rv = nni_thr_init(&p->p_recv_thr, ops->pipe_recv, pdata)) != 0) {
+		ops->pipe_fini(&p->p_proto_data);
 		NNI_FREE_STRUCT(p);
 		return (rv);
 	}
-	rv = nni_thr_init(&p->p_send_thr, pops->pipe_send, p->p_proto_data);
-	if (rv != 0) {
+	if ((rv = nni_thr_init(&p->p_send_thr, ops->pipe_send, pdata)) != 0) {
 		nni_thr_fini(&p->p_recv_thr);
-		pops->pipe_fini(&p->p_proto_data);
+		ops->pipe_fini(&p->p_proto_data);
 		NNI_FREE_STRUCT(p);
 		return (rv);
 	}
@@ -155,9 +151,16 @@ nni_pipe_start(nni_pipe *pipe)
 	nni_mtx_lock(&sock->s_mx);
 	if (sock->s_closing) {
 		nni_mtx_unlock(&sock->s_mx);
+		nni_pipe_close(pipe);
 		return (NNG_ECLOSED);
 	}
 
+	if (nni_pipe_peer(pipe) != sock->s_peer) {
+		nni_mtx_unlock(&sock->s_mx);
+		nni_pipe_close(pipe);
+		return (NNG_EPROTO);
+	}
+
 	do {
 		// We generate a new pipe ID, but we make sure it does not
 		// collide with any we already have.  This can only normally
@@ -174,8 +177,7 @@ nni_pipe_start(nni_pipe *pipe)
 		}
 	} while (collide);
 
-	rv = pipe->p_proto_ops.pipe_add(pipe->p_proto_data);
-	if (rv != 0) {
+	if ((rv = sock->s_pipe_ops.pipe_add(pipe->p_proto_data)) != 0) {
 		nni_mtx_unlock(&sock->s_mx);
 		nni_pipe_close(pipe);
 		return (rv);
diff --git a/src/core/pipe.h b/src/core/pipe.h
index 242e1d8c..cc14de86 100644
--- a/src/core/pipe.h
+++ b/src/core/pipe.h
@@ -21,7 +21,6 @@ struct nng_pipe {
 	uint32_t	p_id;
 	nni_tran_pipe	p_tran_ops;
 	void *		p_tran_data;
-	nni_proto_pipe	p_proto_ops;
 	void *		p_proto_data;
 	nni_list_node	p_node;
 	nni_sock *	p_sock;
diff --git a/src/core/protocol.h b/src/core/protocol.h
index c858ad39..000e9c64 100644
--- a/src/core/protocol.h
+++ b/src/core/protocol.h
@@ -1,5 +1,5 @@
 //
-// Copyright 2016 Garrett D'Amore <garrett@damore.org>
+// Copyright 2017 Garrett D'Amore <garrett@damore.org>
 //
 // This software is supplied under the terms of the MIT License, a
 // copy of which should be located in the distribution where this
@@ -21,7 +21,7 @@
 // implementations.
 
 // nni_proto_pipe contains protocol-specific per-pipe operations.
-struct nni_proto_pipe {
+struct nni_proto_pipe_ops {
 	// pipe_init creates the protocol-specific per pipe data structure.
 	// The last argument is the per-socket protocol private data.
 	int	(*pipe_init)(void **, nni_pipe *, void *);
@@ -55,32 +55,52 @@ struct nni_proto_pipe {
 	void	(*pipe_recv)(void *);
 };
 
-struct nni_proto {
-	uint16_t		proto_self;     // our 16-bit protocol ID
-	uint16_t		proto_peer;     // who we peer with (ID)
-	const char *		proto_name;     // string version of our name
-	const nni_proto_pipe *	proto_pipe;     // Per-pipe operations.
+struct nni_proto_sock_ops {
+	// sock_initf creates the protocol instance, which will be stored on
+	// the socket. This is run without the sock lock held, and allocates
+	// storage or other resources for the socket.
+	int		(*sock_init)(void **, nni_sock *);
 
-	// Create protocol instance, which will be stored on the socket.
-	int			(*proto_init)(void **, nni_sock *);
+	// sock_fini destroys the protocol instance.  This is run without the
+	// socket lock held, and is intended to release resources.  It may
+	// block as needed.
+	void		(*sock_fini)(void *);
 
-	// Destroy the protocol instance.
-	void			(*proto_fini)(void *);
+	// Close the protocol instance.  This is run with the lock held,
+	// and intended to initiate closure of the socket.  For example,
+	// it can signal the socket worker threads to exit.
+	void		(*sock_close)(void *);
 
 	// Option manipulation.  These may be NULL.
-	int			(*proto_setopt)(void *, int, const void *,
-	    size_t);
-	int			(*proto_getopt)(void *, int, void *, size_t *);
+	int		(*sock_setopt)(void *, int, const void *, size_t);
+	int		(*sock_getopt)(void *, int, void *, size_t *);
+
+	// sock_send is a send worker.  It can really be anything, but it
+	// is run in a separate thread (if it is non-NULL).
+	void		(*sock_send)(void *);
+
+	// sock_recv is a receive worker.  As with send it can really be
+	// anything, its just a thread that runs for the duration of the
+	// socket.
+	void		(*sock_recv)(void *);
 
 	// Receive filter.  This may be NULL, but if it isn't, then
 	// messages coming into the system are routed here just before being
 	// delivered to the application.  To drop the message, the prtocol
 	// should return NULL, otherwise the message (possibly modified).
-	nni_msg *		(*proto_recv_filter)(void *, nni_msg *);
+	nni_msg *	(*sock_rfilter)(void *, nni_msg *);
 
 	// Send filter.  This may be NULL, but if it isn't, then messages
 	// here are filtered just after they come from the application.
-	nni_msg *		(*proto_send_filter)(void *, nni_msg *);
+	nni_msg *	(*sock_sfilter)(void *, nni_msg *);
+};
+
+struct nni_proto {
+	uint16_t			proto_self;     // our 16-bit D
+	uint16_t			proto_peer;     // who we peer with (ID)
+	const char *			proto_name;     // Our name
+	const nni_proto_sock_ops *	proto_sock_ops; // Per-socket opeations
+	const nni_proto_pipe_ops *	proto_pipe_ops; // Per-pipe operations.
 };
 
 // These functions are not used by protocols, but rather by the socket
diff --git a/src/core/socket.c b/src/core/socket.c
index 4e8eccbf..2ef35d7a 100644
--- a/src/core/socket.c
+++ b/src/core/socket.c
@@ -56,7 +56,7 @@ nni_reaper(void *arg)
 			// Note that if a protocol has rejected the pipe, it
 			// won't have any data.
 			if (pipe->p_active) {
-				pipe->p_proto_ops.pipe_rem(pipe->p_proto_data);
+				sock->s_pipe_ops.pipe_rem(pipe->p_proto_data);
 			}
 			nni_mtx_unlock(&sock->s_mx);
 
@@ -78,6 +78,13 @@ nni_reaper(void *arg)
 }
 
 
+nni_mtx *
+nni_sock_mtx(nni_sock *sock)
+{
+	return (&sock->s_mx);
+}
+
+
 static nni_msg *
 nni_sock_nullfilter(void *arg, nni_msg *mp)
 {
@@ -108,6 +115,22 @@ nni_sock_nullsetopt(void *arg, int num, const void *data, size_t sz)
 }
 
 
+static void
+nni_sock_nullop(void *arg)
+{
+	NNI_ARG_UNUSED(arg);
+}
+
+
+static int
+nni_sock_nulladdpipe(void *arg)
+{
+	NNI_ARG_UNUSED(arg);
+
+	return (0);
+}
+
+
 // nn_sock_open creates the underlying socket.
 int
 nni_sock_open(nni_sock **sockp, uint16_t pnum)
@@ -124,7 +147,8 @@ nni_sock_open(nni_sock **sockp, uint16_t pnum)
 	}
 
 	// We make a copy of the protocol operations.
-	sock->s_proto = *proto;
+	sock->s_protocol = proto->proto_self;
+	sock->s_peer = proto->proto_peer;
 	sock->s_linger = 0;
 	sock->s_sndtimeo = -1;
 	sock->s_rcvtimeo = -1;
@@ -135,6 +159,30 @@ nni_sock_open(nni_sock **sockp, uint16_t pnum)
 	NNI_LIST_INIT(&sock->s_reaps, nni_pipe, p_node);
 	NNI_LIST_INIT(&sock->s_eps, nni_ep, ep_node);
 
+	sock->s_sock_ops = *proto->proto_sock_ops;
+	if (sock->s_sock_ops.sock_sfilter == NULL) {
+		sock->s_sock_ops.sock_sfilter = nni_sock_nullfilter;
+	}
+	if (sock->s_sock_ops.sock_rfilter == NULL) {
+		sock->s_sock_ops.sock_rfilter = nni_sock_nullfilter;
+	}
+	if (sock->s_sock_ops.sock_getopt == NULL) {
+		sock->s_sock_ops.sock_getopt = nni_sock_nullgetopt;
+	}
+	if (sock->s_sock_ops.sock_setopt == NULL) {
+		sock->s_sock_ops.sock_setopt = nni_sock_nullsetopt;
+	}
+	if (sock->s_sock_ops.sock_close == NULL) {
+		sock->s_sock_ops.sock_close = nni_sock_nullop;
+	}
+	sock->s_pipe_ops = *proto->proto_pipe_ops;
+	if (sock->s_pipe_ops.pipe_add == NULL) {
+		sock->s_pipe_ops.pipe_add = nni_sock_nulladdpipe;
+	}
+	if (sock->s_pipe_ops.pipe_rem == NULL) {
+		sock->s_pipe_ops.pipe_rem = nni_sock_nullop;
+	}
+
 	if ((rv = nni_mtx_init(&sock->s_mx)) != 0) {
 		NNI_FREE_STRUCT(sock);
 		return (rv);
@@ -145,6 +193,7 @@ nni_sock_open(nni_sock **sockp, uint16_t pnum)
 		return (rv);
 	}
 
+
 	if ((rv = nni_thr_init(&sock->s_reaper, nni_reaper, sock)) != 0) {
 		nni_cv_fini(&sock->s_cv);
 		nni_mtx_fini(&sock->s_mx);
@@ -161,14 +210,14 @@ nni_sock_open(nni_sock **sockp, uint16_t pnum)
 	}
 	if ((rv = nni_msgq_init(&sock->s_urq, 0)) != 0) {
 		nni_msgq_fini(sock->s_uwq);
-		nni_thr_fini(&sock->s_reaper);
+		nni_thr_fini(&sock->s_recver);
 		nni_cv_fini(&sock->s_cv);
 		nni_mtx_fini(&sock->s_mx);
 		NNI_FREE_STRUCT(sock);
 		return (rv);
 	}
 
-	if ((rv = sock->s_proto.proto_init(&sock->s_data, sock)) != 0) {
+	if ((rv = sock->s_sock_ops.sock_init(&sock->s_data, sock)) != 0) {
 		nni_msgq_fini(sock->s_urq);
 		nni_msgq_fini(sock->s_uwq);
 		nni_thr_fini(&sock->s_reaper);
@@ -177,19 +226,39 @@ nni_sock_open(nni_sock **sockp, uint16_t pnum)
 		NNI_FREE_STRUCT(sock);
 		return (rv);
 	}
-	if (sock->s_proto.proto_send_filter == NULL) {
-		sock->s_proto.proto_send_filter = nni_sock_nullfilter;
-	}
-	if (sock->s_proto.proto_recv_filter == NULL) {
-		sock->s_proto.proto_recv_filter = nni_sock_nullfilter;
-	}
-	if (sock->s_proto.proto_getopt == NULL) {
-		sock->s_proto.proto_getopt = nni_sock_nullgetopt;
+
+	// NB: If these functions are NULL, the thread initialization is
+	// largely a NO-OP.  The system won't actually create the threads.
+	rv = nni_thr_init(&sock->s_sender, sock->s_sock_ops.sock_send,
+		sock->s_data);
+	if (rv != 0) {
+		nni_thr_wait(&sock->s_reaper);
+		sock->s_sock_ops.sock_fini(&sock->s_data);
+		nni_msgq_fini(sock->s_urq);
+		nni_msgq_fini(sock->s_uwq);
+		nni_thr_fini(&sock->s_reaper);
+		nni_cv_fini(&sock->s_cv);
+		nni_mtx_fini(&sock->s_mx);
+		NNI_FREE_STRUCT(sock);
 	}
-	if (sock->s_proto.proto_setopt == NULL) {
-		sock->s_proto.proto_setopt = nni_sock_nullsetopt;
+	rv = nni_thr_init(&sock->s_recver, sock->s_sock_ops.sock_recv,
+		sock->s_data);
+	if (rv != 0) {
+		nni_thr_wait(&sock->s_sender);
+		sock->s_sock_ops.sock_fini(&sock->s_data);
+		nni_msgq_fini(sock->s_urq);
+		nni_msgq_fini(sock->s_uwq);
+		nni_thr_fini(&sock->s_sender);
+		nni_thr_fini(&sock->s_reaper);
+		nni_cv_fini(&sock->s_cv);
+		nni_mtx_fini(&sock->s_mx);
+		NNI_FREE_STRUCT(sock);
 	}
+
+
 	nni_thr_run(&sock->s_reaper);
+	nni_thr_run(&sock->s_recver);
+	nni_thr_run(&sock->s_sender);
 	*sockp = sock;
 	return (0);
 }
@@ -265,10 +334,14 @@ nni_sock_shutdown(nni_sock *sock)
 		nni_mtx_lock(&sock->s_mx);
 	}
 
+	sock->s_sock_ops.sock_close(sock->s_data);
+
 	nni_cv_wake(&sock->s_cv);
 	nni_mtx_unlock(&sock->s_mx);
 
-	// Wait for the reaper to exit.
+	// Wait for the threads to exit.
+	nni_thr_wait(&sock->s_sender);
+	nni_thr_wait(&sock->s_recver);
 	nni_thr_wait(&sock->s_reaper);
 
 	// At this point, there are no threads blocked inside of us
@@ -297,7 +370,7 @@ nni_sock_close(nni_sock *sock)
 	// the results may be tragic.
 
 	// The protocol needs to clean up its state.
-	sock->s_proto.proto_fini(sock->s_data);
+	sock->s_sock_ops.sock_fini(sock->s_data);
 
 	// And we need to clean up *our* state.
 	nni_thr_fini(&sock->s_reaper);
@@ -328,9 +401,10 @@ nni_sock_sendmsg(nni_sock *sock, nni_msg *msg, nni_time expire)
 		return (rv);
 	}
 	besteffort = sock->s_besteffort;
+
+	msg = sock->s_sock_ops.sock_sfilter(sock->s_data, msg);
 	nni_mtx_unlock(&sock->s_mx);
 
-	msg = sock->s_proto.proto_send_filter(sock->s_data, msg);
 	if (msg == NULL) {
 		return (0);
 	}
@@ -372,7 +446,9 @@ nni_sock_recvmsg(nni_sock *sock, nni_msg **msgp, nni_time expire)
 		if (rv != 0) {
 			return (rv);
 		}
-		msg = sock->s_proto.proto_recv_filter(sock->s_data, msg);
+		nni_mtx_lock(&sock->s_mx);
+		msg = sock->s_sock_ops.sock_rfilter(sock->s_data, msg);
+		nni_mtx_unlock(&sock->s_mx);
 		if (msg != NULL) {
 			break;
 		}
@@ -388,14 +464,14 @@ nni_sock_recvmsg(nni_sock *sock, nni_msg **msgp, nni_time expire)
 uint16_t
 nni_sock_proto(nni_sock *sock)
 {
-	return (sock->s_proto.proto_self);
+	return (sock->s_protocol);
 }
 
 
 uint16_t
 nni_sock_peer(nni_sock *sock)
 {
-	return (sock->s_proto.proto_peer);
+	return (sock->s_peer);
 }
 
 
@@ -488,7 +564,7 @@ nni_sock_setopt(nni_sock *sock, int opt, const void *val, size_t size)
 		nni_mtx_unlock(&sock->s_mx);
 		return (NNG_ECLOSED);
 	}
-	rv = sock->s_proto.proto_setopt(sock->s_data, opt, val, size);
+	rv = sock->s_sock_ops.sock_setopt(sock->s_data, opt, val, size);
 	if (rv != NNG_ENOTSUP) {
 		nni_mtx_unlock(&sock->s_mx);
 		return (rv);
@@ -533,7 +609,7 @@ nni_sock_getopt(nni_sock *sock, int opt, void *val, size_t *sizep)
 		nni_mtx_unlock(&sock->s_mx);
 		return (NNG_ECLOSED);
 	}
-	rv = sock->s_proto.proto_getopt(sock->s_data, opt, val, sizep);
+	rv = sock->s_sock_ops.sock_getopt(sock->s_data, opt, val, sizep);
 	if (rv != NNG_ENOTSUP) {
 		nni_mtx_unlock(&sock->s_mx);
 		return (rv);
diff --git a/src/core/socket.h b/src/core/socket.h
index 62347f61..197b5d0d 100644
--- a/src/core/socket.h
+++ b/src/core/socket.h
@@ -14,36 +14,42 @@
 // OUSIDE of the core is STRICTLY VERBOTEN.  NO DIRECT ACCESS BY PROTOCOLS OR
 // TRANSPORTS.
 struct nng_socket {
-	nni_mtx		s_mx;
-	nni_cv		s_cv;
+	nni_mtx			s_mx;
+	nni_cv			s_cv;
 
-	nni_msgq *	s_uwq;          // Upper write queue
-	nni_msgq *	s_urq;          // Upper read queue
+	nni_msgq *		s_uwq;  // Upper write queue
+	nni_msgq *		s_urq;  // Upper read queue
 
-	nni_proto	s_proto;
+	uint16_t		s_protocol;
+	uint16_t		s_peer;
 
-	void *		s_data;         // Protocol private
+	nni_proto_pipe_ops	s_pipe_ops;
+	nni_proto_sock_ops	s_sock_ops;
 
-	// XXX: options
-	nni_duration	s_linger;               // linger time
-	nni_duration	s_sndtimeo;             // send timeout
-	nni_duration	s_rcvtimeo;             // receive timeout
-	nni_duration	s_reconn;               // reconnect time
-	nni_duration	s_reconnmax;            // max reconnect time
-
-	nni_list	s_eps;                  // active endpoints
-	nni_list	s_pipes;                // pipes for this socket
-
-	nni_list	s_reaps;                // pipes to reap
-	nni_thr		s_reaper;
+	void *			s_data; // Protocol private
 
-	int		s_ep_pend;              // EP dial/listen in progress
-	int		s_closing;              // Socket is closing
-	int		s_besteffort;           // Best effort mode delivery
-	int		s_senderr;              // Protocol state machine use
-	int		s_recverr;              // Protocol state machine use
-
-	uint32_t	s_nextid;               // Next Pipe ID.
+	// XXX: options
+	nni_duration		s_linger;       // linger time
+	nni_duration		s_sndtimeo;     // send timeout
+	nni_duration		s_rcvtimeo;     // receive timeout
+	nni_duration		s_reconn;       // reconnect time
+	nni_duration		s_reconnmax;    // max reconnect time
+
+	nni_list		s_eps;          // active endpoints
+	nni_list		s_pipes;        // pipes for this socket
+
+	nni_list		s_reaps;        // pipes to reap
+	nni_thr			s_reaper;
+	nni_thr			s_sender;
+	nni_thr			s_recver;
+
+	int			s_ep_pend;      // EP dial/listen in progress
+	int			s_closing;      // Socket is closing
+	int			s_besteffort;   // Best effort mode delivery
+	int			s_senderr;      // Protocol state machine use
+	int			s_recverr;      // Protocol state machine use
+
+	uint32_t		s_nextid;       // Next Pipe ID.
 };
 
 extern int nni_sock_open(nni_sock **, uint16_t);
@@ -76,5 +82,12 @@ extern nni_msgq *nni_sock_sendq(nni_sock *);
 // inject incoming messages from pipes to it.
 extern nni_msgq *nni_sock_recvq(nni_sock *);
 
+// nni_sock_mtx obtains the socket mutex.  This is for protocols to use
+// from separate threads; they must not hold the lock for extended periods.
+// Additionally, this can only be acquired from separate threads.  The
+// synchronous entry points (excluding the send/recv thread workers) will
+// be called with this lock already held.  We expose the mutex directly
+// here so that protocols can use it to initialize condvars.
+extern nni_mtx *nni_sock_mtx(nni_sock *);
 
 #endif  // CORE_SOCKET_H
diff --git a/src/core/thread.c b/src/core/thread.c
index 7678ad0a..4a871ede 100644
--- a/src/core/thread.c
+++ b/src/core/thread.c
@@ -128,6 +128,10 @@ nni_thr_init(nni_thr *thr, nni_thr_func fn, void *arg)
 		nni_plat_mtx_fini(&thr->mtx);
 		return (rv);
 	}
+	if (fn == NULL) {
+		thr->done = 1;
+		return (0);
+	}
 	if ((rv = nni_plat_thr_init(&thr->thr, nni_thr_wrap, thr)) != 0) {
 		nni_plat_cv_fini(&thr->cv);
 		nni_plat_mtx_fini(&thr->mtx);
@@ -170,7 +174,9 @@ nni_thr_fini(nni_thr *thr)
 		nni_plat_cv_wait(&thr->cv);
 	}
 	nni_plat_mtx_unlock(&thr->mtx);
-	nni_plat_thr_fini(&thr->thr);
+	if (thr->fn != NULL) {
+		nni_plat_thr_fini(&thr->thr);
+	}
 	nni_plat_cv_fini(&thr->cv);
 	nni_plat_mtx_fini(&thr->mtx);
 }