diff --git a/lib/core/connect.c b/lib/core/connect.c
index f1c6bf4b8..1bcb50f01 100644
--- a/lib/core/connect.c
+++ b/lib/core/connect.c
@@ -45,6 +45,9 @@ lws_client_connect_via_info(const struct lws_client_connect_info *i)
 	struct lws *wsi;
 	const struct lws_protocols *p;
 	const char *local = i->protocol;
+#if LWS_MAX_SMP > 1
+	int n, tid;
+#endif
 
 	if (i->context->requested_kill)
 		return NULL;
@@ -67,27 +70,6 @@ lws_client_connect_via_info(const struct lws_client_connect_info *i)
 	wsi->context = i->context;
 	wsi->desc.sockfd = LWS_SOCK_INVALID;
 
-	/*
-	 * PHASE 2: Choose an initial role for the wsi and do role-specific init
-	 *
-	 * Note the initial role may not reflect the final role, eg,
-	 * we may want ws, but first we have to go through h1 to get that
-	 */
-
-	lws_role_call_client_bind(wsi, i);
-
-	/*
-	 * PHASE 3: fill up the wsi with stuff from the connect_info as far as
-	 * it can go.  It's uncertain because not only is our connection
-	 * going to complete asynchronously, we might have bound to h1 and not
-	 * even be able to get ahold of an ah immediately.
-	 */
-
-	wsi->user_space = NULL;
-	wsi->pending_timeout = NO_PENDING_TIMEOUT;
-	wsi->position_in_fds_table = LWS_NO_FDS_POS;
-	wsi->c_port = i->port;
-
 	wsi->vhost = NULL;
 	if (!i->vhost)
 		lws_vhost_bind_wsi(i->context->vhost_list, wsi);
@@ -100,6 +82,54 @@ lws_client_connect_via_info(const struct lws_client_connect_info *i)
 		goto bail;
 	}
 
+	/*
+	 * PHASE 2: if SMP, bind the client to whatever tsi the current thread
+	 * represents
+	 */
+
+#if LWS_MAX_SMP > 1
+	tid = wsi->vhost->protocols[0].callback(wsi,
+				     LWS_CALLBACK_GET_THREAD_ID, NULL, NULL, 0);
+
+	lws_context_lock(i->context, "client find tsi");
+
+	for (n = 0; n < i->context->count_threads; n++)
+		if (i->context->pt[n].service_tid == tid) {
+			lwsl_info("%s: client binds to caller tsi %d\n",
+				  __func__, n);
+			wsi->tsi = n;
+			break;
+		}
+
+	/*
+	 * this binding is sort of provisional, since when we try to insert
+	 * into the pt fds, there may be no space and it will fail
+	 */
+
+	lws_context_unlock(i->context);
+#endif
+
+	/*
+	 * PHASE 3: Choose an initial role for the wsi and do role-specific init
+	 *
+	 * Note the initial role may not reflect the final role, eg,
+	 * we may want ws, but first we have to go through h1 to get that
+	 */
+
+	lws_role_call_client_bind(wsi, i);
+
+	/*
+	 * PHASE 4: fill up the wsi with stuff from the connect_info as far as
+	 * it can go.  It's uncertain because not only is our connection
+	 * going to complete asynchronously, we might have bound to h1 and not
+	 * even be able to get ahold of an ah immediately.
+	 */
+
+	wsi->user_space = NULL;
+	wsi->pending_timeout = NO_PENDING_TIMEOUT;
+	wsi->position_in_fds_table = LWS_NO_FDS_POS;
+	wsi->c_port = i->port;
+
 	wsi->protocol = &wsi->vhost->protocols[0];
 	wsi->client_pipeline = !!(i->ssl_connection & LCCSCF_PIPELINE);
 
@@ -111,7 +141,7 @@ lws_client_connect_via_info(const struct lws_client_connect_info *i)
 	}
 
 	/*
-	 * PHASE 4: handle external user_space now, generic alloc is done in
+	 * PHASE 5: handle external user_space now, generic alloc is done in
 	 * role finalization
 	 */
 
@@ -130,7 +160,7 @@ lws_client_connect_via_info(const struct lws_client_connect_info *i)
 #endif
 
 	/*
-	 * PHASE 5: stash the things from connect_info that we can't process
+	 * PHASE 6: stash the things from connect_info that we can't process
 	 * right now, eg, if http binding, without an ah.  If h1 and no ah, we
 	 * will go on the ah waiting list and process those things later (after
 	 * the connect_info and maybe the things pointed to have gone out of
@@ -180,7 +210,7 @@ lws_client_connect_via_info(const struct lws_client_connect_info *i)
 	}
 
 	/*
-	 * PHASE 6: Do any role-specific finalization processing.  We can still
+	 * PHASE 7: Do any role-specific finalization processing.  We can still
 	 * see important info things via wsi->stash
 	 */
 
diff --git a/lib/core/context.c b/lib/core/context.c
index ebc1fb84c..5557b814f 100644
--- a/lib/core/context.c
+++ b/lib/core/context.c
@@ -667,6 +667,11 @@ lws_callback_http_dummy(struct lws *wsi, enum lws_callback_reasons reason,
 			    si->where, si->ret);
 		break;
 
+#if LWS_MAX_SMP > 1
+	case LWS_CALLBACK_GET_THREAD_ID:
+		return (int)(unsigned long long)pthread_self();
+#endif
+
 	default:
 		break;
 	}
diff --git a/lib/core/pollfd.c b/lib/core/pollfd.c
index 24f7c7ab7..514ea255e 100644
--- a/lib/core/pollfd.c
+++ b/lib/core/pollfd.c
@@ -182,7 +182,7 @@ _lws_change_pollfd(struct lws *wsi, int _and, int _or, struct lws_pollargs *pa)
 			ret = -1;
 			goto bail;
 		}
-		sampled_tid = context->service_tid;
+		sampled_tid = pt->service_tid;
 		if (sampled_tid && wsi->vhost) {
 			tid = wsi->vhost->protocols[0].callback(wsi,
 				     LWS_CALLBACK_GET_THREAD_ID, NULL, NULL, 0);
@@ -335,6 +335,7 @@ __remove_wsi_socket_from_fds(struct lws *wsi)
 		pt->fds[m] = pt->fds[pt->fds_count - 1];
 		/* this decrements pt->fds_count */
 		lws_plat_delete_socket_from_fds(context, wsi, m);
+		pt->count_conns--;
 		v = (int) pt->fds[m].fd;
 		/* end guy's "position in fds table" is now the deletion guy's old one */
 		end_wsi = wsi_from_fd(context, v);
diff --git a/lib/core/private.h b/lib/core/private.h
index 28b310c15..de78bf27a 100644
--- a/lib/core/private.h
+++ b/lib/core/private.h
@@ -394,6 +394,15 @@ struct lws_context_per_thread {
 	unsigned long count_conns;
 	unsigned int fds_count;
 
+	/*
+	 * set to the Thread ID that's doing the service loop just before entry
+	 * to poll indicates service thread likely idling in poll()
+	 * volatile because other threads may check it as part of processing
+	 * for pollfd event change.
+	 */
+	volatile int service_tid;
+	int service_tid_detected;
+
 	volatile unsigned char inside_poll;
 	volatile unsigned char foreign_spinlock;
 
@@ -691,14 +700,6 @@ struct lws_context {
 	unsigned int doing_protocol_init:1;
 	unsigned int done_protocol_destroy_cb:1;
 	unsigned int finalize_destroy_after_internal_loops_stopped:1;
-	/*
-	 * set to the Thread ID that's doing the service loop just before entry
-	 * to poll indicates service thread likely idling in poll()
-	 * volatile because other threads may check it as part of processing
-	 * for pollfd event change.
-	 */
-	volatile int service_tid;
-	int service_tid_detected;
 
 	short count_threads;
 	short plugin_protocol_count;
diff --git a/lib/libwebsockets.h b/lib/libwebsockets.h
index ad074ec30..273f835e6 100644
--- a/lib/libwebsockets.h
+++ b/lib/libwebsockets.h
@@ -3491,6 +3491,12 @@ struct lws_client_connect_info {
 	 *   non-NULL: this protocol name is used to bind the connection to
 	 *             the local protocol handler.  .protocol is used for the
 	 *             list of remote ws protocols we could accept */
+	const char *alpn;
+	/**< NULL: allow lws default ALPN list, from vhost if present or from
+	 *       list of roles built into lws
+	 * non-NULL: require one from provided comma-separated list of alpn
+	 *           tokens
+	 */
 
 	/* Add new things just above here ---^
 	 * This is part of the ABI, don't needlessly break compatibility
@@ -3499,12 +3505,6 @@ struct lws_client_connect_info {
 	 * members added above will see 0 (default) even if the app
 	 * was not built against the newer headers.
 	 */
-	const char *alpn;
-	/* NULL: allow lws default ALPN list, from vhost if present or from
-	 *       list of roles built into lws
-	 * non-NULL: require one from provided comma-separated list of alpn
-	 *           tokens
-	 */
 
 	void *_unused[4]; /**< dummy */
 };
diff --git a/lib/plat/esp32/esp32-service.c b/lib/plat/esp32/esp32-service.c
index 3549d9d49..fcf3488eb 100644
--- a/lib/plat/esp32/esp32-service.c
+++ b/lib/plat/esp32/esp32-service.c
@@ -76,16 +76,14 @@ _lws_plat_service_tsi(struct lws_context *context, int timeout_ms, int tsi)
 	if (timeout_ms < 0)
 		goto faked_service;
 
-	if (!context->service_tid_detected) {
+	if (!pt->service_tid_detected) {
 		struct lws *_lws = lws_zalloc(sizeof(*_lws), "tid probe");
 
 		_lws->context = context;
 
-		context->service_tid_detected =
-			context->vhost_list->protocols[0].callback(
+		pt->service_tid = context->vhost_list->protocols[0].callback(
 			_lws, LWS_CALLBACK_GET_THREAD_ID, NULL, NULL, 0);
-		context->service_tid = context->service_tid_detected;
-		context->service_tid_detected = 1;
+		pt->service_tid_detected = 1;
 		lws_free(_lws);
 	}
 
diff --git a/lib/plat/optee/lws-plat-optee.c b/lib/plat/optee/lws-plat-optee.c
index d5ba9017f..9803dad49 100644
--- a/lib/plat/optee/lws-plat-optee.c
+++ b/lib/plat/optee/lws-plat-optee.c
@@ -110,17 +110,15 @@ _lws_plat_service_tsi(struct lws_context *context, int timeout_ms, int tsi)
 	if (timeout_ms < 0)
 		goto faked_service;
 
-	if (!context->service_tid_detected) {
+	if (!pt->service_tid_detected) {
 		struct lws _lws;
 
 		memset(&_lws, 0, sizeof(_lws));
 		_lws.context = context;
 
-		context->service_tid_detected =
-			context->vhost_list->protocols[0].callback(
+		pt->service_tid = context->vhost_list->protocols[0].callback(
 			&_lws, LWS_CALLBACK_GET_THREAD_ID, NULL, NULL, 0);
-		context->service_tid = context->service_tid_detected;
-		context->service_tid_detected = 1;
+		pt->service_tid_detected = 1;
 	}
 
 	/*
diff --git a/lib/plat/unix/unix-service.c b/lib/plat/unix/unix-service.c
index 9a7b84f09..00c1fc926 100644
--- a/lib/plat/unix/unix-service.c
+++ b/lib/plat/unix/unix-service.c
@@ -52,17 +52,16 @@ _lws_plat_service_tsi(struct lws_context *context, int timeout_ms, int tsi)
 	if (context->event_loop_ops->run_pt)
 		context->event_loop_ops->run_pt(context, tsi);
 
-	if (!context->service_tid_detected) {
+	if (!pt->service_tid_detected) {
 		struct lws _lws;
 
 		memset(&_lws, 0, sizeof(_lws));
 		_lws.context = context;
 
-		context->service_tid_detected =
+		pt->service_tid  =
 			context->vhost_list->protocols[0].callback(
 			&_lws, LWS_CALLBACK_GET_THREAD_ID, NULL, NULL, 0);
-		context->service_tid = context->service_tid_detected;
-		context->service_tid_detected = 1;
+		pt->service_tid_detected = 1;
 	}
 
 	/*
diff --git a/lib/plat/windows/windows-service.c b/lib/plat/windows/windows-service.c
index 9351da031..b027adddd 100644
--- a/lib/plat/windows/windows-service.c
+++ b/lib/plat/windows/windows-service.c
@@ -42,17 +42,16 @@ _lws_plat_service_tsi(struct lws_context *context, int timeout_ms, int tsi)
 
 	pt = &context->pt[tsi];
 
-	if (!context->service_tid_detected) {
+	if (!pt->service_tid_detected) {
 		struct lws _lws;
 
 		memset(&_lws, 0, sizeof(_lws));
 		_lws.context = context;
 
-		context->service_tid_detected = context->vhost_list->
+		pt->service_tid = context->vhost_list->
 			protocols[0].callback(&_lws, LWS_CALLBACK_GET_THREAD_ID,
 						  NULL, NULL, 0);
-		context->service_tid = context->service_tid_detected;
-		context->service_tid_detected = 1;
+		pt->service_tid_detected = 1;
 	}
 
 	if (timeout_ms < 0) {
@@ -173,8 +172,6 @@ _lws_plat_service_tsi(struct lws_context *context, int timeout_ms, int tsi)
 		}
 	}
 
-	context->service_tid = 0;
-
 	if (ev == WSA_WAIT_TIMEOUT)
 		lws_service_fd(context, NULL);