Hey Maxime,

On Thu, Feb 29, 2024 at 1:25 PM Maxime Coquelin
<maxime.coque...@redhat.com> wrote:
>
> VDUSE_DESTROY_DEVICE ioctl can fail because the device's
> chardev is not released despite close syscall having been
> called. It happens because the events handler thread is
> still polling the file descriptor.
>
> fdset_pipe_notify() is not enough because it does not
> ensure the notification has been handled by the event
> thread, it just returns once the notification is sent.
>
> To fix this, this patch introduces a synchronization
> mechanism based on pthread's condition, so that
> fdset_pipe_notify() only returns once the pipe's read
> callback has been executed.
>
> Fixes: 51d018fdac4e ("vhost: add VDUSE events handler")

This looks to be a generic issue in the fd_man code.
In practice, VDUSE only seems to be affected, so I am ok with this Fixes: tag.


> Cc: sta...@dpdk.org
>
> Signed-off-by: Maxime Coquelin <maxime.coque...@redhat.com>
> ---
>  lib/vhost/fd_man.c | 21 ++++++++++++++++++---
>  lib/vhost/fd_man.h |  5 +++++
>  2 files changed, 23 insertions(+), 3 deletions(-)
>
> diff --git a/lib/vhost/fd_man.c b/lib/vhost/fd_man.c
> index 79a8d2c006..42ce059039 100644
> --- a/lib/vhost/fd_man.c
> +++ b/lib/vhost/fd_man.c
> @@ -309,10 +309,11 @@ fdset_event_dispatch(void *arg)
>  }
>
>  static void
> -fdset_pipe_read_cb(int readfd, void *dat __rte_unused,
> +fdset_pipe_read_cb(int readfd, void *dat,
>                    int *remove __rte_unused)
>  {
>         char charbuf[16];
> +       struct fdset *fdset = dat;
>         int r = read(readfd, charbuf, sizeof(charbuf));
>         /*
>          * Just an optimization, we don't care if read() failed
> @@ -320,6 +321,11 @@ fdset_pipe_read_cb(int readfd, void *dat __rte_unused,
>          * compiler happy
>          */
>         RTE_SET_USED(r);
> +
> +       pthread_mutex_lock(&fdset->sync_mutex);
> +       fdset->sync = true;
> +       pthread_cond_broadcast(&fdset->sync_cond);
> +       pthread_mutex_unlock(&fdset->sync_mutex);
>  }
>
>  void
> @@ -342,7 +348,7 @@ fdset_pipe_init(struct fdset *fdset)
>         }
>
>         ret = fdset_add(fdset, fdset->u.readfd,
> -                       fdset_pipe_read_cb, NULL, NULL);
> +                       fdset_pipe_read_cb, NULL, fdset);
>
>         if (ret < 0) {
>                 VHOST_FDMAN_LOG(ERR,
> @@ -359,7 +365,12 @@ fdset_pipe_init(struct fdset *fdset)
>  void
>  fdset_pipe_notify(struct fdset *fdset)
>  {
> -       int r = write(fdset->u.writefd, "1", 1);
> +       int r;
> +
> +       pthread_mutex_lock(&fdset->sync_mutex);
> +
> +       fdset->sync = false;
> +       r = write(fdset->u.writefd, "1", 1);
>         /*
>          * Just an optimization, we don't care if write() failed
>          * so ignore explicitly its return value to make the
> @@ -367,4 +378,8 @@ fdset_pipe_notify(struct fdset *fdset)
>          */
>         RTE_SET_USED(r);
>
> +       while (!fdset->sync)
> +               pthread_cond_wait(&fdset->sync_cond, &fdset->sync_mutex);
> +
> +       pthread_mutex_unlock(&fdset->sync_mutex);
>  }
> diff --git a/lib/vhost/fd_man.h b/lib/vhost/fd_man.h
> index 6315904c8e..cc19937612 100644
> --- a/lib/vhost/fd_man.h
> +++ b/lib/vhost/fd_man.h
> @@ -6,6 +6,7 @@
>  #define _FD_MAN_H_
>  #include <pthread.h>
>  #include <poll.h>
> +#include <stdbool.h>
>
>  #define MAX_FDS 1024
>
> @@ -35,6 +36,10 @@ struct fdset {
>                         int writefd;
>                 };
>         } u;
> +
> +       pthread_mutex_t sync_mutex;
> +       pthread_cond_t sync_cond;
> +       bool sync;

We should explicitly initialise those in
https://git.dpdk.org/dpdk/tree/lib/vhost/socket.c#n91 and
https://git.dpdk.org/dpdk/tree/lib/vhost/vduse.c#n34.
The rest looks acceptable to me.


-- 
David Marchand

Reply via email to