On Thu, Mar 07, 2013 at 05:47:26PM +0800, Asias He wrote:
> On Thu, Mar 07, 2013 at 09:58:04AM +0100, Stefan Hajnoczi wrote:
> > On Thu, Mar 07, 2013 at 08:26:20AM +0800, Asias He wrote:
> > > On Wed, Mar 06, 2013 at 10:21:09AM +0100, Stefan Hajnoczi wrote:
> > > > On Wed, Mar 06, 2013 at 02:16:30PM +0800, Asias He wrote:
> > > > > +static struct tcm_vhost_evt *tcm_vhost_allocate_evt(struct vhost_scsi *vs,
> > > > > + u32 event, u32 reason)
> > > > > +{
> > > > > + struct tcm_vhost_evt *evt;
> > > > > +
> > > > > + if (atomic_read(&vs->vs_events_nr) > VHOST_SCSI_MAX_EVENT)
> > > > > + return NULL;
> > > > > +
> > > > > + evt = kzalloc(sizeof(*evt), GFP_KERNEL);
> > > > > +
> > > > > + if (evt) {
> > > > > + atomic_inc(&vs->vs_events_nr);
> > > >
> > > > This looks suspicious: checking vs_events_nr > VHOST_SCSI_MAX_EVENT
> > > > first and then incrementing later isn't atomic!
> > >
> > > This does not matter. (1) and (2) are okay. In case (3), the other side
> > > can only decrease the number of event, the limit will not be exceeded.
> > >
> > > (1)
> > > atomic_dec()
> > > atomic_read()
> > > atomic_inc()
> > > (2)
> > > atomic_read()
> > > atomic_inc()
> > > atomic_dec()
> > >
> > > (3)
> > > atomic_read()
> > > atomic_dec()
> > > atomic_inc()
> >
> > The cases you listed are fine but I'm actually concerned about
> > tcm_vhost_allocate_evt() racing with itself. There are 3 callers and
> > I'm not sure which lock prevents them from executing at the same time.
>
> No lock to prevent it. But what is the racing of executing
> tcm_vhost_allocate_evt() at the same time?
atomic_read() <= VHOST_SCSI_MAX_EVENT
atomic_read() <= VHOST_SCSI_MAX_EVENT
atomic_inc()
atomic_inc()
Now vs->vs_events_nr == VHOST_SCSI_MAX_EVENT + 1 which the if statement
was supposed to prevent.
> > > > > +static int tcm_vhost_hotunplug(struct tcm_vhost_tpg *tpg, struct se_lun *lun)
> > > > > +{
> > > > > + struct vhost_scsi *vs = tpg->vhost_scsi;
> > > > > +
> > > > > + mutex_lock(&tpg->tv_tpg_mutex);
> > > > > + vs = tpg->vhost_scsi;
> > > > > + mutex_unlock(&tpg->tv_tpg_mutex);
> > > > > + if (!vs)
> > > > > + return -EOPNOTSUPP;
> > > > > +
> > > > > + if (!tcm_vhost_check_feature(vs, VIRTIO_SCSI_F_HOTPLUG))
> > > > > + return -EOPNOTSUPP;
> > > > > +
> > > > > + return tcm_vhost_send_evt(vs, tpg, lun,
> > > > > + VIRTIO_SCSI_T_TRANSPORT_RESET,
> > > > > + VIRTIO_SCSI_EVT_RESET_REMOVED);
> > > > > +}
> > > >
> > > > tcm_vhost_hotplug() and tcm_vhost_hotunplug() are the same function
> > > > except for VIRTIO_SCSI_EVT_RESET_RESCAN vs
> > > > VIRTIO_SCSI_EVT_RESET_REMOVED. That can be passed in as an argument and
> > > > the code duplication can be eliminated.
> > >
> > > I thought about this also. We can have a tcm_vhost_do_hotplug() helper.
> > >
> > > tcm_vhost_do_hotplug(tpg, lun, plug)
> > >
> > > tcm_vhost_hotplug() {
> > > tcm_vhost_do_hotplug(tpg, lun, true)
> > > }
> > >
> > > tcm_vhost_hotunplug() {
> > > tcm_vhost_do_hotplug(tpg, lun, false)
> > > }
> > >
> > > The reason I did not do that is I do not like the true/false argument
> > > but anyway this could remove duplication. I will do it.
> >
> > true/false makes the calling code hard to read, I suggest passing in
> > VIRTIO_SCSI_EVT_RESET_RESCAN or VIRTIO_SCSI_EVT_RESET_REMOVED as the
> > argument.
>
> Yes. However, I think passing VIRTIO_SCSI_EVT_RESET_* is even worse.
>
> 1) Having VIRTIO_SCSI_EVT_RESET_RESCAN or VIRTIO_SCSI_EVT_RESET_REMOVED
> around VIRTIO_SCSI_T_TRANSPORT_RESET would be nicer.
>
> 2) tcm_vhost_do_hotplug(tpg, lun, VIRTIO_SCSI_EVT_RESET_*)
> doest not make much sense. What the hell is VIRTIO_SCSI_EVT_RESET_* when
> you do hotplug or hotunplug. In contrast, if we have
> tcm_vhost_do_hotplug(tpg, lun, plug), plug means doing hotplug or
> hotunplug.
The VIRTIO_SCSI_EVT_RESET_REMOVED constant is pretty clear ("removed"
means unplug). The VIRTIO_SCSI_EVT_RESET_RESCAN is less clear, but this
code is in drivers/vhost/tcm_vhost.c so you can expect the reader to
know the device specification :).
Anyway, it's not the end of the world if you leave the duplicated code
in, use a boolean parameter, or use the virtio event constant.
Stefan
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
