> -----Original Message----- > From: Tetsuya Mukawa [mailto:mukawa at igel.co.jp] > Sent: Sunday, November 09, 2014 10:13 PM > To: Xie, Huawei; dev at dpdk.org > Cc: nakajima.yoshihiro at lab.ntt.co.jp; masutani.hitoshi at lab.ntt.co.jp > Subject: Re: [dpdk-dev] [RFC PATCH 3/7] lib/librte_vhost: Add an abstraction > layer tointerpret messages > > Hi Xie, > > (2014/11/08 5:43), Xie, Huawei wrote: > >> -struct vhost_net_device_ops const *get_virtio_net_callbacks(void); > >> +struct vhost_net_device_ops const *get_virtio_net_callbacks( > >> + vhost_driver_type_t type); > > Tetsuya: > > I feel currently it is better we still keep the common > > get_virtio_net_callbacks(). > > For the message flow from control layer 1 (cuse ioctl or user sock message > recv/xmit)---> cuse/user local message handling layer 2-> common virtio > message handling layer 3 > > Layer 1 and layer 2 belong to one module. It is that module's choice > > whether > to implement callbacks between internal layer1 and layer2. We don't need to > force that. > > Besides, even that module wants to define the ops between layer 1 and > > layer2, > the interface could be different between cuse/user. > > Refer to the following code for user: > > > > vhost-user-server.c: > > case VHOST_USER_SET_MEM_TABLE: > > user_set_mem_table(ctx, &msg) > > > > virtio-net-user.c: > > user_set_mem_table(struct vhost_device_ctx ctx, struct VhostUserMsg *pmsg) > > { > > > > .... > > > > ops->set_mem_table(ctx, regions, memory.nregions); > > } > > > > > I may misunderstand what you say, please let me know in the case. > I guess it's difficult to remove 'vhost_driver_type_t' from > 'get_virtio_net_callbacks()'. > In original vhost example code, there are 2 layers related with > initialization as you mentioned. > + Layer1: cuse ioctl handling layer. > + Layer2: vhost-cuse( = vhost-net) message handling layer. > > Layer1 needs function pointers to call Layer2 functions. > 'get_virtio_net_callbacks()' is used for that purpose. > > My RFC is based on above, but Layer1/2 are abstracted to hide vhost-cuse > and vhost-user. > + Layer1: device control abstraction layer. > -- Layer1-a: cuse ioctl handling layer. > -- Layer1-b: unix domain socket handling layer. > + Layer2: message handling abstraction layer. > -- Layer2-a: vhost-cuse(vhost-net) message handling layer. > -- Layer2-b: vhost-user message handling layer. > > Still Layer1 needs function pointers of Layer2. > So, anyway, we still need to implement 'get_virtio_net_callbacks()'. > > Also, as you mentioned, function definition and behavior are different > between Layer2-a and Lanyer2-b like 'user_set_mem_table()'. > Because of this, 'get_virtio_net_callbacks()' need to return collect > function pointers to Layer1. > So I guess 'get_virtio_net_callbacks()' needs 'vhost_driver_type_t' to > know which function pointers are needed by Layer1.
Here all layer 2 implementations are required to return same type of vhost_net_device_ops function pointers to layer 1, so layer 1 need to do some kind of preprocessing of its message or wrap some private message ctx in like vhost_device_ctx, and then pass the message to layer2. But as we have a more common layer 3, virtio-net layer, how about we put common message handler in virtio net layer as much as possible, and different layer 2 only do the local message preprocessing, and then pass common message format to layer 3? I think we at least need to define functional pointers between layer 2 and layer 3. Layer 1 and layer 2 actually are sub layers of the same layer. It is that layer(cuse/user) implementation's choice whether to provide an interface between them, and the interface could be different in terms of function prototype. Let us say we are to implement a new vhost, I only care the common interface provided by layer 3. I don't want to register another callbacks for my driver which are used by myself only. Let us think more about this. > > If someone wants to implement new vhost-backend, of course they can > implement Layer2 implementation and Layer1 together. > In the case, they doesn't need to call 'get_virtio_net_callbacks()'. > Also they can reuse existing Layer2 implementation by calling > 'get_virtio_net_callbacks()' with existing driver type, or they can > implement a new Layer2 implementation for new vhost-backend. > > BTW, the name of 'vhost_driver_type_t' is redundant, I will change the name. > > Tetsuya
