[dpdk-dev] [PATCH] vhost: fix vhost_user_set_mem_table error

[Tiwei Bie]

Hi Yi,

On Wed, Sep 20, 2017 at 04:32:03PM +0800, Yi Yang wrote:
> Usually vhost_user message VHOST_USER_SET_MEM_TABLE
> is only sent out during initialization and only sent
> once, but it isn't so for memory hotplug and hotunplug
> , for that case, vhost_user message VHOST_USER_SET_MEM_TABLE
> will be resent with the old memory regions (not hotunplugged)
> and the new memory regions (hotplugged), so current
> vhost_user_set_mem_table implementation is wrong for
> memory hotplug and hotunplug case, it will free current
> memory regions and unmap hugepages no matter if they
> are be using or not and if they are memory regions which
> have been initialized and mapped in the previous vhost_user
> message VHOST_USER_SET_MEM_TABLE or not.
> 
> This commit fixed this issue very well, it will keep them
> intact for those old memory region it will unmap those
> memroy regions if they have been hotunplugged, it will
> initialize the new hotplugged memory regions and map
> hugepages if they are hotplugged.
> 
> vhost_user message VHOST_USER_SET_MEM_TABLE will include
> all the current effective memory regions, the hotunplugged
> memory regions won't be included in VHOST_USER_SET_MEM_TABLE,
> the hotplugged memroy regions will be included in
> VHOST_USER_SET_MEM_TABLE.
> 
> This has been verified in OVS DPDK by memory hotplug and
> hotunplug in qemu.
> 
> Signed-off-by: Yi Yang <yi.y.yang at intel.com>
> ---
>  lib/librte_vhost/vhost_user.c | 72 ++++++++++++++++++++++++++++++++++++++-----
>  1 file changed, 65 insertions(+), 7 deletions(-)
> 
> diff --git a/lib/librte_vhost/vhost_user.c b/lib/librte_vhost/vhost_user.c
> index ad2e8d3..1c475d1 100644
> --- a/lib/librte_vhost/vhost_user.c
> +++ b/lib/librte_vhost/vhost_user.c
[...]
> @@ -532,7 +558,7 @@ vhost_user_set_mem_table(struct virtio_net *dev, struct VhostUserMsg *pmsg)
>  		}
>  	}
>  
> -	dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
> +	dev->mem = rte_realloc(dev->mem, sizeof(struct rte_vhost_memory) +

The rte_realloc() will free the memory pointed by the dev->mem.
But it's possible that some other threads are using it (e.g. by
calling rte_vhost_gpa_to_vva()). It's an issue that needs to be
fixed in this patch.

Best regards,
Tiwei Bie