RE: [PATCH] Drivers: hv: vmbus: Add cpu read lock

[Haiyang Zhang]

> -----Original Message-----
> From: Michael Kelley (LINUX) <mikelley@xxxxxxxxxxxxx>
> Sent: Thursday, June 9, 2022 9:51 AM
> To: Saurabh Sengar <ssengar@xxxxxxxxxxxxxxxxxxx>; KY Srinivasan
> <kys@xxxxxxxxxxxxx>; Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>; Stephen
> Hemminger <sthemmin@xxxxxxxxxxxxx>; wei.liu@xxxxxxxxxx; Dexuan Cui
> <decui@xxxxxxxxxxxxx>; linux-hyperv@xxxxxxxxxxxxxxx; linux-
> kernel@xxxxxxxxxxxxxxx; Saurabh Singh Sengar <ssengar@xxxxxxxxxxxxx>
> Subject: RE: [PATCH] Drivers: hv: vmbus: Add cpu read lock
> 
> From: Saurabh Sengar <ssengar@xxxxxxxxxxxxxxxxxxx> Sent: Wednesday, June
> 8, 2022 10:27 PM
> >
> > Add cpus_read_lock to prevent CPUs from going offline between query and
> > actual use of cpumask. cpumask_of_node is first queried, and based on it
> > used later, in case any CPU goes offline between these two events, it can
> > potentially cause an infinite loop of retries.
> >
> > Signed-off-by: Saurabh Sengar <ssengar@xxxxxxxxxxxxxxxxxxx>
> > ---
> >  drivers/hv/channel_mgmt.c | 4 ++++
> >  1 file changed, 4 insertions(+)
> >
> > diff --git a/drivers/hv/channel_mgmt.c b/drivers/hv/channel_mgmt.c
> > index 85a2142..6a88b7e 100644
> > --- a/drivers/hv/channel_mgmt.c
> > +++ b/drivers/hv/channel_mgmt.c
> > @@ -749,6 +749,9 @@ static void init_vp_index(struct vmbus_channel
> *channel)
> >  		return;
> >  	}
> >
> > +	/* No CPUs should come up or down during this. */
> > +	cpus_read_lock();
> > +
> >  	for (i = 1; i <= ncpu + 1; i++) {
> >  		while (true) {
> >  			numa_node = next_numa_node_id++;
> > @@ -781,6 +784,7 @@ static void init_vp_index(struct vmbus_channel
> *channel)
> >  			break;
> >  	}
> >
> > +	cpus_read_unlock();
> >  	channel->target_cpu = target_cpu;
> >
> >  	free_cpumask_var(available_mask);
> > --
> > 1.8.3.1
> 
> This patch was motivated because I suggested a potential issue here during
> a separate conversation with Saurabh, but it turns out I was wrong. :-(
> 
> init_vp_index() is only called from vmbus_process_offer(), and the
> cpus_read_lock() is already held when init_vp_index() is called.  So the
> issue doesn't exist, and this patch isn't needed.
> 
> However, looking at vmbus_process_offer(), there appears to be a
> different problem in that cpus_read_unlock() is not called when taking
> the error return because the sub_channel_index is zero.
> 
> Michael
> 

        } else {
                /*
                 * Check to see if this is a valid sub-channel.
                 */
                if (newchannel->offermsg.offer.sub_channel_index == 0) {
                        mutex_unlock(&vmbus_connection.channel_mutex);
                        /*
                         * Don't call free_channel(), because newchannel->kobj
                         * is not initialized yet.
                         */
                        kfree(newchannel);
                        WARN_ON_ONCE(1);
                        return;
                }

If this happens, it should be a host bug. Yes, I also think the cpus_read_unlock()
is missing in this error path.

Thanks,
- Haiyang