[PATCH] vhost: fix use-after-free race during cleanup

[PATCH] vhost: fix use-after-free race during cleanup

[Shani Peretz]

This commit fixes a use-after-free that causes the application
to crash on shutdown (detected by ASAN).

The vhost library uses a background event dispatch thread that monitors
fds with epoll. It runs in an infinite loop, waiting for I/O events
and calling callbacks when they occur.

During cleanup, a race condition existed:

  Main Thread:                    Event Dispatch Thread:
  1. Remove fds from fdset        while (1) {
  2. Close file descriptors           epoll_wait() [gets interrupted]
  3. Free fdset memory                [continues loop]
  4. Continue...                      Accesses fdset...   CRASH
                                  }

The main thread would free the fdset memory while the background thread
was still running and using it.

The code had a `destroy` flag that the event dispatch thread checked,
but it was never set during cleanup, and the code never waited for
the thread to actually exit before freeing memory.

This commit implements `fdset_destroy()` that will set the destroy
flag, wait for thread termination, and clean up all resources.
The socket.c is updated to call fdset_destroy() when the last vhost-user
socket is unregistered.

Fixes: 0e38b42bf61c ("vhost: manage FD with epoll")
Cc: stable@dpdk.org

Signed-off-by: Shani Peretz <shperetz@nvidia.com>
---
 lib/vhost/fd_man.c | 38 ++++++++++++++++++++++++++++++++++++++
 lib/vhost/fd_man.h |  1 +
 lib/vhost/socket.c |  7 +++++++
 3 files changed, 46 insertions(+)

diff --git a/lib/vhost/fd_man.c b/lib/vhost/fd_man.c
index f9147edee7..ba1b2ead86 100644
--- a/lib/vhost/fd_man.c
+++ b/lib/vhost/fd_man.c
@@ -393,3 +393,41 @@ fdset_event_dispatch(void *arg)
 
 	return 0;
 }
+
+/**
+ * Destroy the fdset and stop its event dispatch thread.
+ */
+void
+fdset_destroy(struct fdset *pfdset)
+{
+	uint32_t val;
+	int i;
+
+	if (pfdset == NULL)
+		return;
+
+	/* Signal the event dispatch thread to stop */
+	pfdset->destroy = true;
+
+	/* Wait for the event dispatch thread to finish */
+	rte_thread_join(pfdset->tid, &val);
+
+	/* Close the epoll file descriptor */
+	close(pfdset->epfd);
+
+	/* Destroy the mutex */
+	pthread_mutex_destroy(&pfdset->fd_mutex);
+
+	/* Remove from global registry */
+	pthread_mutex_lock(&fdsets_mutex);
+	for (i = 0; i < MAX_FDSETS; i++) {
+		if (fdsets[i] == pfdset) {
+			fdsets[i] = NULL;
+			break;
+		}
+	}
+	pthread_mutex_unlock(&fdsets_mutex);
+
+	/* Free the fdset structure */
+	rte_free(pfdset);
+}
diff --git a/lib/vhost/fd_man.h b/lib/vhost/fd_man.h
index eadcc6fb42..ed2109f3c8 100644
--- a/lib/vhost/fd_man.h
+++ b/lib/vhost/fd_man.h
@@ -21,5 +21,6 @@ int fdset_add(struct fdset *pfdset, int fd,
 
 void fdset_del(struct fdset *pfdset, int fd);
 int fdset_try_del(struct fdset *pfdset, int fd);
+void fdset_destroy(struct fdset *pfdset);
 
 #endif
diff --git a/lib/vhost/socket.c b/lib/vhost/socket.c
index 9b4f332f94..0240da8ff0 100644
--- a/lib/vhost/socket.c
+++ b/lib/vhost/socket.c
@@ -1141,6 +1141,13 @@ rte_vhost_driver_unregister(const char *path)
 		count = --vhost_user.vsocket_cnt;
 		vhost_user.vsockets[i] = vhost_user.vsockets[count];
 		vhost_user.vsockets[count] = NULL;
+
+		/* Check if we need to destroy the vhost fdset */
+		if (vhost_user.vsocket_cnt == 0 && vhost_user.fdset != NULL) {
+			fdset_destroy(vhost_user.fdset);
+			vhost_user.fdset = NULL;
+		}
+
 		pthread_mutex_unlock(&vhost_user.mutex);
 		return 0;
 	}
-- 
2.34.1

Re: [PATCH] vhost: fix use-after-free race during cleanup

[fengchengwen]

On 11/4/2025 4:09 PM, Shani Peretz wrote:
> This commit fixes a use-after-free that causes the application
> to crash on shutdown (detected by ASAN).
> 
> The vhost library uses a background event dispatch thread that monitors
> fds with epoll. It runs in an infinite loop, waiting for I/O events
> and calling callbacks when they occur.
> 
> During cleanup, a race condition existed:
> 
>   Main Thread:                    Event Dispatch Thread:
>   1. Remove fds from fdset        while (1) {
>   2. Close file descriptors           epoll_wait() [gets interrupted]
>   3. Free fdset memory                [continues loop]
>   4. Continue...                      Accesses fdset...   CRASH
>                                   }
> 
> The main thread would free the fdset memory while the background thread
> was still running and using it.

Who will free fdset memory ? I check the lib/vhost/socket.c and found there are no explicit free.

I think it maybe the hugepage free because the fdset use rte_zmalloc(). If it's, please explicit
add it into the commit log.

> 
> The code had a `destroy` flag that the event dispatch thread checked,
> but it was never set during cleanup, and the code never waited for
> the thread to actually exit before freeing memory.
> 
> This commit implements `fdset_destroy()` that will set the destroy
> flag, wait for thread termination, and clean up all resources.
> The socket.c is updated to call fdset_destroy() when the last vhost-user
> socket is unregistered.
> 
> Fixes: 0e38b42bf61c ("vhost: manage FD with epoll")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Shani Peretz <shperetz@nvidia.com>

Re: [PATCH] vhost: fix use-after-free race during cleanup

[Maxime Coquelin]

Hi Shani,

Thanks for the fix, more comments below:

On Tue, Nov 4, 2025 at 10:50 AM fengchengwen <fengchengwen@huawei.com> wrote:
>
> On 11/4/2025 4:09 PM, Shani Peretz wrote:
> > This commit fixes a use-after-free that causes the application
> > to crash on shutdown (detected by ASAN).
> >
> > The vhost library uses a background event dispatch thread that monitors
> > fds with epoll. It runs in an infinite loop, waiting for I/O events
> > and calling callbacks when they occur.
> >
> > During cleanup, a race condition existed:
> >
> >   Main Thread:                    Event Dispatch Thread:
> >   1. Remove fds from fdset        while (1) {
> >   2. Close file descriptors           epoll_wait() [gets interrupted]
> >   3. Free fdset memory                [continues loop]
> >   4. Continue...                      Accesses fdset...   CRASH
> >                                   }
> >
> > The main thread would free the fdset memory while the background thread
> > was still running and using it.
>
> Who will free fdset memory ? I check the lib/vhost/socket.c and found there are no explicit free.
>
> I think it maybe the hugepage free because the fdset use rte_zmalloc(). If it's, please explicit
> add it into the commit log.

I agree with Feng, it would be good to provide more information on who
is freeing the memory.

> >
> > The code had a `destroy` flag that the event dispatch thread checked,
> > but it was never set during cleanup, and the code never waited for
> > the thread to actually exit before freeing memory.
> >
> > This commit implements `fdset_destroy()` that will set the destroy
> > flag, wait for thread termination, and clean up all resources.
> > The socket.c is updated to call fdset_destroy() when the last vhost-user
> > socket is unregistered.
> >
> > Fixes: 0e38b42bf61c ("vhost: manage FD with epoll")
> > Cc: stable@dpdk.org
> >
> > Signed-off-by: Shani Peretz <shperetz@nvidia.com>
>
>

We also need to call fdset_destroy in vduse_device_destroy() if it is
destorying the last VDUSE device.
We might need to add a counter to struct vduse to know whether this is
the last device.

Other than that, the patch looks good to me.

Thanks,
Maxime

RE: [PATCH] vhost: fix use-after-free race during cleanup

[Shani Peretz]

> -----Original Message-----
> From: fengchengwen <fengchengwen@huawei.com>
> Sent: Tuesday, 4 November 2025 11:33
> To: Shani Peretz <shperetz@nvidia.com>; dev@dpdk.org
> Cc: stable@dpdk.org; Maxime Coquelin <maxime.coquelin@redhat.com>;
> Chenbo Xia <chenbox@nvidia.com>; David Marchand
> <david.marchand@redhat.com>
> Subject: Re: [PATCH] vhost: fix use-after-free race during cleanup
> 
> External email: Use caution opening links or attachments
> 
> 
> On 11/4/2025 4:09 PM, Shani Peretz wrote:
> > This commit fixes a use-after-free that causes the application to
> > crash on shutdown (detected by ASAN).
> >
> > The vhost library uses a background event dispatch thread that
> > monitors fds with epoll. It runs in an infinite loop, waiting for I/O
> > events and calling callbacks when they occur.
> >
> > During cleanup, a race condition existed:
> >
> >   Main Thread:                    Event Dispatch Thread:
> >   1. Remove fds from fdset        while (1) {
> >   2. Close file descriptors           epoll_wait() [gets interrupted]
> >   3. Free fdset memory                [continues loop]
> >   4. Continue...                      Accesses fdset...   CRASH
> >                                   }
> >
> > The main thread would free the fdset memory while the background
> > thread was still running and using it.
> 
> Who will free fdset memory ? I check the lib/vhost/socket.c and found there
> are no explicit free.
> 
> I think it maybe the hugepage free because the fdset use rte_zmalloc(). If it's,
> please explicit add it into the commit log.

Yes you're right I double checked with a debugger and indeed the fdset memory is freed when hugepage free, I'll update the commit log.

> 
> >
> > The code had a `destroy` flag that the event dispatch thread checked,
> > but it was never set during cleanup, and the code never waited for the
> > thread to actually exit before freeing memory.
> >
> > This commit implements `fdset_destroy()` that will set the destroy
> > flag, wait for thread termination, and clean up all resources.
> > The socket.c is updated to call fdset_destroy() when the last
> > vhost-user socket is unregistered.
> >
> > Fixes: 0e38b42bf61c ("vhost: manage FD with epoll")
> > Cc: stable@dpdk.org
> >
> > Signed-off-by: Shani Peretz <shperetz@nvidia.com>
> 

[PATCH v2] vhost: fix use-after-free race during cleanup

[Shani Peretz]

This commit fixes a use-after-free that causes the application to crash
on shutdown (detected by ASAN).

The vhost library uses a background event dispatch thread that monitors
fds with epoll. It runs in an infinite loop, waiting for I/O events
and calling callbacks when they occur.

During cleanup, a race condition existed:

  Main Thread:                    Event Dispatch Thread:
  1. Remove fds from fdset        while (1) {
  2. Close file descriptors           epoll_wait() [gets interrupted]
  3. Free fdset memory                [continues loop]
  4. Continue...                      Accesses fdset...   CRASH
                                  }

There isn't explicit cleanup of the fdset structure.
The fdset structue is allocated with rte_zmalloc() and the memory would
only be reclaimed at application shutdown when rte_eal_cleanup() is called,
which invokes rte_eal_memory_detach() to unmap all the hugepage memory.
Meanwhile, the event dispatch thread could still be running and accessing
the fdset.

The code had a `destroy` flag that the event dispatch thread checked,
but it was never set during cleanup, and the code never waited for
the thread to actually exit before freeing memory.

To fix this, the commit implements `fdset_destroy()` that will set the
destroy flag, wait for thread termination, and clean up all resources.

The socket.c and vduse.c are updated to call fdset_destroy() when
the last socket/device is unregistered.

For vduse, reference counting was added to track the number of devices
using the fdset. The fdset is destroyed when the last device is removed.

Fixes: 0e38b42bf61c ("vhost: manage FD with epoll")
Cc: stable@dpdk.org

Signed-off-by: Shani Peretz <shperetz@nvidia.com>



----------
v2:
- Extended the fix to vduse.c, added reference counting and mutex to vduse
  structure to track the number of devices using the fdset
- Call fdset_destroy() when last device is removed in vduse_device_destroy()
  and in error paths of vduse_device_create()
- Added mutex protection when checking/setting the destroy flag to prevent
  race conditions in both fdset_event_dispatch() and fdset_destroy()

---
 lib/vhost/fd_man.c | 45 ++++++++++++++++++++++++++++++++++++-
 lib/vhost/fd_man.h |  1 +
 lib/vhost/socket.c |  7 ++++++
 lib/vhost/vduse.c  | 56 +++++++++++++++++++++++++++++++++++++---------
 4 files changed, 98 insertions(+), 11 deletions(-)

diff --git a/lib/vhost/fd_man.c b/lib/vhost/fd_man.c
index f9147edee7..b4597dec75 100644
--- a/lib/vhost/fd_man.c
+++ b/lib/vhost/fd_man.c
@@ -387,9 +387,52 @@ fdset_event_dispatch(void *arg)
 			}
 		}

-		if (pfdset->destroy)
+		pthread_mutex_lock(&pfdset->fd_mutex);
+		bool should_destroy = pfdset->destroy;
+		pthread_mutex_unlock(&pfdset->fd_mutex);
+		if (should_destroy)
 			break;
 	}

 	return 0;
 }
+
+/**
+ * Destroy the fdset and stop its event dispatch thread.
+ */
+void
+fdset_destroy(struct fdset *pfdset)
+{
+	uint32_t val;
+	int i;
+
+	if (pfdset == NULL)
+		return;
+
+	/* Signal the event dispatch thread to stop */
+	pthread_mutex_lock(&pfdset->fd_mutex);
+	pfdset->destroy = true;
+	pthread_mutex_unlock(&pfdset->fd_mutex);
+
+	/* Wait for the event dispatch thread to finish */
+	rte_thread_join(pfdset->tid, &val);
+
+	/* Close the epoll file descriptor */
+	close(pfdset->epfd);
+
+	/* Destroy the mutex */
+	pthread_mutex_destroy(&pfdset->fd_mutex);
+
+	/* Remove from global registry */
+	pthread_mutex_lock(&fdsets_mutex);
+	for (i = 0; i < MAX_FDSETS; i++) {
+		if (fdsets[i] == pfdset) {
+			fdsets[i] = NULL;
+			break;
+		}
+	}
+	pthread_mutex_unlock(&fdsets_mutex);
+
+	/* Free the fdset structure */
+	rte_free(pfdset);
+}
diff --git a/lib/vhost/fd_man.h b/lib/vhost/fd_man.h
index eadcc6fb42..ed2109f3c8 100644
--- a/lib/vhost/fd_man.h
+++ b/lib/vhost/fd_man.h
@@ -21,5 +21,6 @@ int fdset_add(struct fdset *pfdset, int fd,

 void fdset_del(struct fdset *pfdset, int fd);
 int fdset_try_del(struct fdset *pfdset, int fd);
+void fdset_destroy(struct fdset *pfdset);

 #endif
diff --git a/lib/vhost/socket.c b/lib/vhost/socket.c
index 9b4f332f94..0240da8ff0 100644
--- a/lib/vhost/socket.c
+++ b/lib/vhost/socket.c
@@ -1141,6 +1141,13 @@ rte_vhost_driver_unregister(const char *path)
 		count = --vhost_user.vsocket_cnt;
 		vhost_user.vsockets[i] = vhost_user.vsockets[count];
 		vhost_user.vsockets[count] = NULL;
+
+		/* Check if we need to destroy the vhost fdset */
+		if (vhost_user.vsocket_cnt == 0 && vhost_user.fdset != NULL) {
+			fdset_destroy(vhost_user.fdset);
+			vhost_user.fdset = NULL;
+		}
+
 		pthread_mutex_unlock(&vhost_user.mutex);
 		return 0;
 	}
diff --git a/lib/vhost/vduse.c b/lib/vhost/vduse.c
index 68e56843fd..f255f717ab 100644
--- a/lib/vhost/vduse.c
+++ b/lib/vhost/vduse.c
@@ -30,9 +30,15 @@

 struct vduse {
 	struct fdset *fdset;
+	int device_cnt;
+	pthread_mutex_t mutex;
 };

-static struct vduse vduse;
+static struct vduse vduse = {
+	.fdset = NULL,
+	.device_cnt = 0,
+	.mutex = PTHREAD_MUTEX_INITIALIZER,
+};

 static const char * const vduse_reqs_str[] = {
 	"VDUSE_GET_VQ_STATE",
@@ -683,19 +689,16 @@ vduse_device_create(const char *path, bool compliant_ol_flags)
 	const char *name = path + strlen("/dev/vduse/");
 	bool reconnect = false;

-	if (vduse.fdset == NULL) {
-		vduse.fdset = fdset_init("vduse-evt");
-		if (vduse.fdset == NULL) {
-			VHOST_CONFIG_LOG(path, ERR, "failed to init VDUSE fdset");
-			return -1;
-		}
-	}
+	pthread_mutex_lock(&vduse.mutex);
+	vduse.device_cnt++;
+	pthread_mutex_unlock(&vduse.mutex);

 	control_fd = open(VDUSE_CTRL_PATH, O_RDWR);
 	if (control_fd < 0) {
 		VHOST_CONFIG_LOG(name, ERR, "Failed to open %s: %s",
 				VDUSE_CTRL_PATH, strerror(errno));
-		return -1;
+		ret = -1;
+		goto out_dec_cnt;
 	}

 	if (ioctl(control_fd, VDUSE_SET_API_VERSION, &ver)) {
@@ -845,6 +848,19 @@ vduse_device_create(const char *path, bool compliant_ol_flags)

 	dev->cvq = dev->virtqueue[max_queue_pairs * 2];

+	/* Only allocate when we know device creation will succeed */
+	pthread_mutex_lock(&vduse.mutex);
+	if (vduse.fdset == NULL) {
+		vduse.fdset = fdset_init("vduse-evt");
+		if (vduse.fdset == NULL) {
+			VHOST_CONFIG_LOG(path, ERR, "failed to init VDUSE fdset");
+			pthread_mutex_unlock(&vduse.mutex);
+			ret = -1;
+			goto out_log_unmap;
+		}
+	}
+	pthread_mutex_unlock(&vduse.mutex);
+
 	ret = fdset_add(vduse.fdset, dev->vduse_dev_fd, vduse_events_handler, NULL, dev);
 	if (ret) {
 		VHOST_CONFIG_LOG(name, ERR, "Failed to add fd %d to vduse fdset",
@@ -861,6 +877,8 @@ vduse_device_create(const char *path, bool compliant_ol_flags)
 	return 0;

 out_log_unmap:
+	if (vduse.fdset != NULL)
+		fdset_del(vduse.fdset, dev->vduse_dev_fd);
 	munmap(dev->reconnect_log, sizeof(*dev->reconnect_log));
 out_dev_destroy:
 	vhost_destroy_device(vid);
@@ -870,6 +888,14 @@ vduse_device_create(const char *path, bool compliant_ol_flags)
 	ioctl(control_fd, VDUSE_DESTROY_DEV, name);
 out_ctrl_close:
 	close(control_fd);
+out_dec_cnt:
+	pthread_mutex_lock(&vduse.mutex);
+	vduse.device_cnt--;
+	if (vduse.device_cnt == 0 && vduse.fdset != NULL) {
+		fdset_destroy(vduse.fdset);
+		vduse.fdset = NULL;
+	}
+	pthread_mutex_unlock(&vduse.mutex);

 	return ret;
 }
@@ -899,7 +925,17 @@ vduse_device_destroy(const char *path)

 	vduse_device_stop(dev);

-	fdset_del(vduse.fdset, dev->vduse_dev_fd);
+	if (vduse.fdset != NULL)
+		fdset_del(vduse.fdset, dev->vduse_dev_fd);
+
+	/* Check if we need to destroy the vduse fdset */
+	pthread_mutex_lock(&vduse.mutex);
+	vduse.device_cnt--;
+	if (vduse.device_cnt == 0 && vduse.fdset != NULL) {
+		fdset_destroy(vduse.fdset);
+		vduse.fdset = NULL;
+	}
+	pthread_mutex_unlock(&vduse.mutex);

 	if (dev->vduse_dev_fd >= 0) {
 		close(dev->vduse_dev_fd);
--
2.34.1