Re: [PATCH V2 3/4] vmbus: add per-channel sysfs info

[Stephen Hemminger]

On Thu, 18 Oct 2018 15:32:35 +0000
Michael Kelley <mikelley@xxxxxxxxxxxxx> wrote:

> From Olaf Hering  Sent: Thursday, October 18, 2018 8:20 AM
> >  
> > > This extends existing vmbus related sysfs structure to provide per-channel
> > > state information. This is useful when diagnosing issues with multiple
> > > queues in networking and storage.  
> >   
> > > +++ b/drivers/hv/vmbus_drv.c
> > > +static ssize_t write_avail_show(const struct vmbus_channel *channel, char *buf)
> > > +{
> > > +	const struct hv_ring_buffer_info *rbi = &channel->outbound;
> > > +
> > > +	return sprintf(buf, "%u\n", hv_get_bytes_to_write(rbi));
> > > +}
> > > +VMBUS_CHAN_ATTR_RO(write_avail);  
> > 
> > This is upstream since a year.
> > 
> > But I wonder how this can work if vmbus_device_register is called,
> > and then something reads the populated sysfs files before vmbus_open returns.
> > Nothing protects rbi->ring_buffer in this case, which remains NULL
> > until vmbus_open populates it.
> > 
> > A simple reproduce, with a modular kernel, is to boot with init=/bin/bash
> > head /sys/bus/vmbus/devices/*/channels/*/*
> >   
> 
> There are multiple race conditions with this and other VMbus sysfs information.
> There's a race on the close path as well.  I've got an action on my list to get it
> cleaned up.
> 
> Michael
> 

There is also a bunch of issues with code like:

static ssize_t id_show(struct device *dev, struct device_attribute *dev_attr,
		       char *buf)
{
	struct hv_device *hv_dev = device_to_hv_device(dev);

	if (!hv_dev->channel)
		return -ENODEV;
	return sprintf(buf, "%d\n", hv_dev->channel->offermsg.child_relid);
}

Which should be using ACCESS_ONCE on hv_dev->channel or doing proper RCU.