On 27/02/2017 11:18, Marc-André Lureau wrote: > Since commit b0a335e351103bf92f3f9d0bd5759311be8156ac, a socket write > may trigger a disconnect events, calling vhost_user_stop() and clearing > all the vhost_dev strutures holding data that vhost.c functions expect > to remain valid. Delay the cleanup to keep the vhost_dev structure > valid during the vhost.c functions. > > Signed-off-by: Marc-André Lureau <[email protected]> > --- > net/vhost-user.c | 56 > ++++++++++++++++++++++++++++++++++++++++++++++++++------ > 1 file changed, 50 insertions(+), 6 deletions(-) > > diff --git a/net/vhost-user.c b/net/vhost-user.c > index 77b8110f8c..028bf0cf5d 100644 > --- a/net/vhost-user.c > +++ b/net/vhost-user.c > @@ -25,6 +25,7 @@ typedef struct VhostUserState { > guint watch; > uint64_t acked_features; > bool started; > + QEMUBH *chr_closed_bh; > } VhostUserState; > > VHostNetState *vhost_user_get_vhost_net(NetClientState *nc) > @@ -190,9 +191,45 @@ static gboolean net_vhost_user_watch(GIOChannel *chan, > GIOCondition cond, > > qemu_chr_fe_disconnect(&s->chr); > > + s->watch = 0; > return FALSE; > } > > +static void net_vhost_user_event(void *opaque, int event); > + > +static void chr_closed_bh(void *opaque) > +{ > + const char *name = opaque; > + NetClientState *ncs[MAX_QUEUE_NUM]; > + VhostUserState *s; > + Error *err = NULL; > + int queues; > + > + queues = qemu_find_net_clients_except(name, ncs, > + NET_CLIENT_DRIVER_NIC, > + MAX_QUEUE_NUM); > + assert(queues < MAX_QUEUE_NUM); > + > + s = DO_UPCAST(VhostUserState, nc, ncs[0]); > + > + qmp_set_link(name, false, &err); > + vhost_user_stop(queues, ncs); > + if (s->watch) { > + g_source_remove(s->watch); > + } > + s->watch = 0; > + > + qemu_bh_delete(s->chr_closed_bh); > + s->chr_closed_bh = NULL; > + > + qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, net_vhost_user_event, > + opaque, NULL, true); > + > + if (err) { > + error_report_err(err); > + } > +} > + > static void net_vhost_user_event(void *opaque, int event) > { > const char *name = opaque; > @@ -212,20 +249,27 @@ static void net_vhost_user_event(void *opaque, int > event) > trace_vhost_user_event(chr->label, event); > switch (event) { > case CHR_EVENT_OPENED: > - s->watch = qemu_chr_fe_add_watch(&s->chr, G_IO_HUP, > - net_vhost_user_watch, s); > if (vhost_user_start(queues, ncs, &s->chr) < 0) { > qemu_chr_fe_disconnect(&s->chr); > return; > } > + s->watch = qemu_chr_fe_add_watch(&s->chr, G_IO_HUP, > + net_vhost_user_watch, s); > qmp_set_link(name, true, &err); > + s->chr_closed_bh = qemu_bh_new(chr_closed_bh, opaque); > s->started = true; > break; > case CHR_EVENT_CLOSED: > - qmp_set_link(name, false, &err); > - vhost_user_stop(queues, ncs); > - g_source_remove(s->watch); > - s->watch = 0; > + /* a close event may happen during a read/write, but vhost > + * code assumes the vhost_dev remains setup, so delay the > + * stop & clear to idle. > + * FIXME: better handle failure in vhost code, remove bh > + */ > + if (s->chr_closed_bh) { > + qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, NULL, > + NULL, NULL, false); > + qemu_bh_schedule(s->chr_closed_bh); > + }
The bottom half adds a small overhead to the event loop all the time, even if it is not scheduled. Would it be possible to create it here instead? You can have a s->state enum (OPENED, CLOSING, CLOSED for example). You can even use aio_bh_schedule_oneshot to avoid having to store the QEMUBH pointer. Paolo > break; > } > >
