2016-09-26 16:24, Iremonger, Bernard:
> Hi Bruce, Thomas,
>
> <snip>
>
> > Subject: Re: [dpdk-dev] [PATCH] net/vhost: Add function to retreive the
> > 'vid'
> > for a given port id
> >
> > On Mon, Sep 26, 2016 at 04:26:27PM +0200, Thomas Monjalon wrote:
> > > 2016-09-26 14:18, Bruce Richardson:
> > > > On Mon, Sep 26, 2016 at 03:12:01PM +0200, Thomas Monjalon wrote:
> > > > > 2016-09-23 21:23, Wiles, Keith:
> > > > > > On Sep 23, 2016, at 12:26 AM, Yuanhan Liu
> > <yuanhan.liu at linux.intel.com> wrote:
> > > > > > > On Thu, Sep 22, 2016 at 06:43:55PM +0200, Thomas Monjalon wrote:
> > > > > > >>>>>>>> There could be a similar need in other PMD.
> > > > > > >>>>>>>> If we can get an opaque identifier of the device which
> > > > > > >>>>>>>> is not the port id, we could call some specific
> > > > > > >>>>>>>> functions of the driver not implemented in the generic
> > ethdev API.
> > > > > > >>>>>>>
> > > > > > >>>>>>> That means you have to add/export the PMD API first.
> > > > > > >>>>>>> Isn't it against what you are proposing -- "I think we
> > > > > > >>>>>>> should not add any API to the PMDs" ;)
> > > > > > >>>>>>
> > > > > > >>>>>> Yes you are totally right :) Except that in vhost case,
> > > > > > >>>>>> we would not have any API in the PMD.
> > > > > > >>>>>> But it would allow to have some specific API in other
> > > > > > >>>>>> PMDs for the features which do not fit in a generic API.
> > > > > > >>>>>
> > > > > > >>>>> So, does that mean you are okay with this patch now? I
> > > > > > >>>>> mean, okay to introduce a vhost PMD API?
> > > > > > >>>>
> > > > > > >>>> It means I would be in favor of introducing API in drivers
> > > > > > >>>> for very specific features.
> > > > > > >>>> In this case, I am not sure that retrieving an internal id is
> > > > > > >>>> very
> > specific.
> > > > > > >>>
> > > > > > >>> It's not, instead, it's very generic. The "internal id" is
> > > > > > >>> actually the public interface to vhost-user application, like
> > > > > > >>> "fd" to
> > file APIs.
> > > > > > >>>
> > > > > > >>> Instead of introducing a few specific wrappers/APIs, I'd
> > > > > > >>> prefer to introduce a generic one to get the handle, and let
> > > > > > >>> the application to call other vhost APIs.
> > > > > > >>
> > > > > > >> Yes it makes sense.
> > > > > > >> I was thinking of introducing a function to get an internal
> > > > > > >> id from ethdev, in order to use it with any driver or underlying
> > library.
> > > > > > >> But it would be an opaque pointer and you need an int.
> > > > > > >> Note that we can cast an int into a pointer, so I am not sure
> > > > > > >> what is
> > best.
> > > > > > >
> > > > > > > Yes, that should work. But I just doubt what the "opaque
> > > > > > > pointer" could be for other PMD drivers, and what the
> > > > > > > application could do with it. For a typical nic PMD driver, I
> > > > > > > can think of nothing is valuable to export to user applications.
> > > > > > >
> > > > > > > But maybe it's valuable to other virtual PMD drives as well,
> > > > > > > like the TAP pmd from Keith?
> > > > > >
> > > > > > I do not see a need in the TAP PMD other then returning the FD for
> > TUN/TAP device. Not sure what any application would need with the FD
> > here, as it could cause some problems.
> > > > > >
> > > > > > This feels like we are talking about a IOCTL like generic interface
> > > > > > into
> > the PMD. Then we can add new one types and reject types in the PMD that
> > are not supported. Would this not be a better method for all future PMD
> > APIs?
> > > > > >
> > > > > > Here is just a thought as to how to solve this problem without a PMD
> > specific API. A number of current ethdev APIs could be removed to use the
> > API below. The APIs would be removed from ethdev structure and have the
> > current APIs use the API below. I know some are not happy with number of
> > APIs in the ethdev structure.
> > > > > >
> > > > > > The API could be something like this:
> > > > > > struct rte_tlv { /* Type/Length/Value like structure */
> > > > > > uint16_t type; /* Type of command */
> > > > > > uint16_t len; /* Length of data section on input and on
> > > > > > output
> > */
> > > > > > uint16_t tlen; /* Total or max length of data buffer */
> > > > > > uint8_t data[0];
> > > > > > };
> > > > > >
> > > > > > int rte_eth_dev_ioctl(int pid, int qid, struct rte_tlv *tlv);
> > > > >
> > > > > Yes we are talking about having some specific functions per driver
> > > > > which are not defined in the generic ethdev layer.
> > > > > We need only one function in ethdev to give access to driver-specific
> > API.
> > > > > My idea is to convert the port id into an opaque handler.
> > > > > Your idea is to use the port id in an ioctl like function.
> > > > >
> > > > > About the implementation, these are the 2 differences between my
> > > > > proposal and yours:
> > > > > - You use the well known port id, whereas I need another handler
> > > > > which is understood by the driver.
> > > > > - You need to build a message string which will be decoded by the
> > driver.
> > > > > I propose to directly offer some specific functions in the drivers
> > > > > which are more convenient to use and easier for code review/debug.
> > > > >
> > > > > No conclusion here. I just want to make sure that we are on the
> > > > > same page, and would like to have feedback from others. Thanks
> > > >
> > > > I personally don't like the idea of having a generic IOCTL in
> > > > ethdev. If you want to have NIC-specific functions provided by a
> > > > driver, that is fine, but any app using those is going to be limited to
> > working only with that driver.
> > > >
> > > > In that case, since the driver in question is known, I don't see any
> > > > reason to go through the ethdev layer. I think it would be much
> > > > clearer to have the app instead include the driver's header file and
> > > > call the driver function directly. The #include at the top of the
> > > > file makes the dependency very clear, and having a function name
> > > > instead of IOCTL with magic command numbers allows the action take by
> > the function to be clearer too.
> > >
> > > So you are against an IOCTL API. Me too.
> > > You agree that an application can be NIC-specific and include an
> > > header file given by the driver to offer very specific features. Me too.
> > >
> > > My proposal was to convert the port id to an opaque pointer as handler
> > > of these driver APIs. After an offline discussion, we agreed that it
> > > is not necessary because drivers manage rte_eth_dev struct and port_id
> > > through
> > > lib/librte_ether/rte_ethdev.h: extern struct rte_eth_dev
> > > rte_eth_devices[];
> > >
> >
> > +1. I agree with your proposal, and I also agree that no ethdev changes
> > +are
> > necessary to support drivers having their own private functions.
>
> I am not sure what has been agreed here.
>
> Looking at the code in struct rte_eth_dev{}
>
> struct rte_eth_dev{
> ...
> const struct eth_dev_ops *dev_ops; /**< Functions exported by PMD */
> ...
> }
>
> The driver functions are only accessible if they are in struct eth_dev_ops.
> I thought the issue here was that driver functions should not be added to
> the struct eth_dev_ops.
> Hence the need for an eth_dev API to return a pointer to a driver dev_ops
> structure containing the driver functions.
An example from an application point of view:
#include <rte_pmd_ixgbe.h>
rte_pmd_ixgbe_vf_ping(port_id, vf_id);
