[PATCH] examples/vhost: fix hang on exit

[PATCH] examples/vhost: fix hang on exit

[Stephen Hemminger]

destroy_device() waits for the worker lcore to ack device removal.
Since the signal handler rework, the workers exit before the vhost
drivers are unregistered, so the destroy callback spins forever.

Unregister the drivers while the workers are still polling, then
stop and join them.

Bugzilla ID: 2071
Fixes: 5ffa8fafcbd3 ("examples/vhost: make signal handler safe")
Cc: stable@dpdk.org

Reported-by: Yu Jiang <yux.jiang@intel.com>
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
---
 examples/vhost/main.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/examples/vhost/main.c b/examples/vhost/main.c
index 523195f5c4..16c9387f8f 100644
--- a/examples/vhost/main.c
+++ b/examples/vhost/main.c
@@ -69,6 +69,8 @@ static uint32_t enabled_port_mask = 0;
 /* Set by the SIGINT handler to stop the worker loops. */
 static volatile sig_atomic_t quit;
 
+static RTE_ATOMIC(bool) stop_workers;
+
 /* Promiscuous mode */
 static uint32_t promiscuous;
 
@@ -1242,7 +1244,7 @@ switch_worker(void *arg)
 		}
 	}
 
-	while (!quit) {
+	while (!rte_atomic_load_explicit(&stop_workers, rte_memory_order_relaxed)) {
 		drain_mbuf_table(tx_q);
 		drain_vhost_table();
 		/*
-- 
2.53.0

Re: [PATCH] examples/vhost: fix hang on exit

[Maxime Coquelin]

On 10/29/26 3:47 PM, Stephen Hemminger wrote:
> destroy_device() waits for the worker lcore to ack device removal.
> Since the signal handler rework, the workers exit before the vhost
> drivers are unregistered, so the destroy callback spins forever.
>
> Unregister the drivers while the workers are still polling, then
> stop and join them.
>
> Bugzilla ID: 2071
> Fixes: 5ffa8fafcbd3 ("examples/vhost: make signal handler safe")
> Cc: stable@dpdk.org
>
> Reported-by: Yu Jiang <yux.jiang@intel.com>
> Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
> ---
>   examples/vhost/main.c | 4 +++-
>   1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/examples/vhost/main.c b/examples/vhost/main.c
> index 523195f5c4..16c9387f8f 100644
> --- a/examples/vhost/main.c
> +++ b/examples/vhost/main.c
> @@ -69,6 +69,8 @@ static uint32_t enabled_port_mask = 0;
>   /* Set by the SIGINT handler to stop the worker loops. */
>   static volatile sig_atomic_t quit;
>   
> +static RTE_ATOMIC(bool) stop_workers;
> +
>   /* Promiscuous mode */
>   static uint32_t promiscuous;
>   
> @@ -1242,7 +1244,7 @@ switch_worker(void *arg)
>   		}
>   	}
>   
> -	while (!quit) {
> +	while (!rte_atomic_load_explicit(&stop_workers, rte_memory_order_relaxed)) {
>   		drain_mbuf_table(tx_q);
>   		drain_vhost_table();
>   		/*

Are you sure you committed all the changes? I don't see a store of true 
to stop_workers.

[PATCH v2] examples/vhost: fix hang on exit

[Stephen Hemminger]

destroy_device() waits for the worker lcore to ack device removal.
Since the signal handler rework, the workers exit before the vhost
drivers are unregistered, so the destroy callback spins forever.

Maing thread now blocks SIGINT and SIGTERM and uses sigwait
to wait for one of these signals. Then unregisters and tells
worker threads to stop.

Bugzilla ID: 2071
Fixes: 5ffa8fafcbd3 ("examples/vhost: make signal handler safe")
Cc: stable@dpdk.org

Reported-by: Yu Jiang <yux.jiang@intel.com>
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
---
 examples/vhost/main.c | 27 +++++++++++++++------------
 1 file changed, 15 insertions(+), 12 deletions(-)

diff --git a/examples/vhost/main.c b/examples/vhost/main.c
index 523195f5c4..e7434fd189 100644
--- a/examples/vhost/main.c
+++ b/examples/vhost/main.c
@@ -66,8 +66,7 @@ static int total_num_mbufs = NUM_MBUFS_DEFAULT;
 /* mask of enabled ports */
 static uint32_t enabled_port_mask = 0;
 
-/* Set by the SIGINT handler to stop the worker loops. */
-static volatile sig_atomic_t quit;
+static RTE_ATOMIC(bool) stop_workers;
 
 /* Promiscuous mode */
 static uint32_t promiscuous;
@@ -1242,7 +1241,7 @@ switch_worker(void *arg)
 		}
 	}
 
-	while (!quit) {
+	while (!rte_atomic_load_explicit(&stop_workers, rte_memory_order_relaxed)) {
 		drain_mbuf_table(tx_q);
 		drain_vhost_table();
 		/*
@@ -1494,12 +1493,6 @@ unregister_drivers(int socket_num)
 	}
 }
 
-static void
-sigint_handler(__rte_unused int signum)
-{
-	quit = 1;
-}
-
 /*
  * Main function, does initialisation and calls the per-lcore functions.
  */
@@ -1513,8 +1506,13 @@ main(int argc, char *argv[])
 	uint16_t portid;
 	rte_thread_t tid;
 	uint64_t flags = RTE_VHOST_USER_NET_COMPLIANT_OL_FLAGS;
+	sigset_t set;
+	int sig;
 
-	signal(SIGINT, sigint_handler);
+	sigemptyset(&set);
+	sigaddset(&set, SIGINT);
+	sigaddset(&set, SIGTERM);
+	pthread_sigmask(SIG_BLOCK, &set, NULL);
 
 	/* init EAL */
 	ret = rte_eal_init(argc, argv);
@@ -1654,12 +1652,17 @@ main(int argc, char *argv[])
 		}
 	}
 
-	RTE_LCORE_FOREACH_WORKER(lcore_id)
-		rte_eal_wait_lcore(lcore_id);
+	/* Wait until SIGINT or SIGTERM */
+	sigwait(&set, &sig);
+	pthread_sigmask(SIG_UNBLOCK, &set, NULL);
 
 	/* Unregister vhost driver. */
 	unregister_drivers(nb_sockets);
 
+	/* Tell worker threads to exit */
+	rte_atomic_store_explicit(&stop_workers, true, rte_memory_order_relaxed);
+	rte_eal_mp_wait_lcore();
+
 	/* clean up the EAL */
 	rte_eal_cleanup();
 
-- 
2.53.0

Re: [PATCH v2] examples/vhost: fix hang on exit

[Stephen Hemminger]

On Thu, 29 Oct 2026 09:57:29 -0700
Stephen Hemminger <stephen@networkplumber.org> wrote:

> destroy_device() waits for the worker lcore to ack device removal.
> Since the signal handler rework, the workers exit before the vhost
> drivers are unregistered, so the destroy callback spins forever.
> 
> Maing thread now blocks SIGINT and SIGTERM and uses sigwait
> to wait for one of these signals. Then unregisters and tells
> worker threads to stop.
> 
> Bugzilla ID: 2071
> Fixes: 5ffa8fafcbd3 ("examples/vhost: make signal handler safe")
> Cc: stable@dpdk.org
> 
> Reported-by: Yu Jiang <yux.jiang@intel.com>
> Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>

This is compile test only, and needs ack from Maxime or submitter.