Re: [RFC PATCH V2 3/3] vhost: access vq metadata through kernel virtual address

[Jason Wang <jasowang@redhat.com>]

On 2018/12/29 上午3:34, David Miller wrote:
> From: Jason Wang <jasowang@redhat.com>
> Date: Fri, 28 Dec 2018 15:55:37 +0800
>
>> +static int vhost_invalidate_vmap(struct vhost_virtqueue *vq,
>> +				 struct vhost_vmap *map,
>> +				 unsigned long uaddr,
>> +				 unsigned long start,
>> +				 unsigned long end,
>> +				 bool blockable)
>> +{
>> +	if (start < uaddr && end >= uaddr) {
>> +		if (!blockable)
>> +			return -EAGAIN;
>> +		mutex_lock(&vq->mutex);
>> +		if (map->addr)
>> +			vunmap(map->unmap_addr);
>> +		map->addr = NULL;
>> +		map->unmap_addr = NULL;
>> +		mutex_unlock(&vq->mutex);
>> +	}
>> +
>> +	return 0;
>> +}
> What are the rules for these invalidate operations?
>
> Can there be partial overlaps?  If so, wouldn't you need some way of
> keeping track of the partially overlapping unmaps so that once all of
> the invalidates covering the range occur you properly cleanup and do
> the vunmap()?


Yes, there can be partial overlap, so the check is buggy. We will remap 
the whole range in vq_meta_prefetch() before datapath path try to use 
them, so there's no need to track partial mapping here.

I spot another bug that the caller will access vq->avail without 
synchronized with vhost ioctl. Since we don't want to hold vq mutex for 
each invalidation, I will tear down MMU notifier during vhost ioctl to 
make sure invalidation request can access them without hold vq mutex.

Thanks