RE: [PATCH V4 4/7] Drivers: hv: vmbus: add APIs to register callbacks to process hvsock connection

From: KY Srinivasan
Date: Tue Aug 18 2015 - 01:07:36 EST




> -----Original Message-----
> From: Dexuan Cui
> Sent: Friday, August 7, 2015 3:27 AM
> To: KY Srinivasan <kys@xxxxxxxxxxxxx>; David Miller
> <davem@xxxxxxxxxxxxx>
> Cc: olaf@xxxxxxxxx; gregkh@xxxxxxxxxxxxxxxxxxx; jasowang@xxxxxxxxxx;
> driverdev-devel@xxxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> stephen@xxxxxxxxxxxxxxxxxx; stefanha@xxxxxxxxxx;
> netdev@xxxxxxxxxxxxxxx; apw@xxxxxxxxxxxxx; pebolle@xxxxxxxxxx;
> dan.carpenter@xxxxxxxxxx
> Subject: RE: [PATCH V4 4/7] Drivers: hv: vmbus: add APIs to register callbacks
> to process hvsock connection
>
> > -----Original Message-----
> > From: KY Srinivasan
> > Sent: Friday, August 7, 2015 2:28
> > To: Dexuan Cui <decui@xxxxxxxxxxxxx>; David Miller
> <davem@xxxxxxxxxxxxx>
> > Cc: olaf@xxxxxxxxx; gregkh@xxxxxxxxxxxxxxxxxxx; jasowang@xxxxxxxxxx;
> > driverdev-devel@xxxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> > stephen@xxxxxxxxxxxxxxxxxx; stefanha@xxxxxxxxxx;
> netdev@xxxxxxxxxxxxxxx;
> > apw@xxxxxxxxxxxxx; pebolle@xxxxxxxxxx; dan.carpenter@xxxxxxxxxx
> > Subject: RE: [PATCH V4 4/7] Drivers: hv: vmbus: add APIs to register
> callbacks to
> > process hvsock connection
> >
> > > -----Original Message-----
> > > From: Dexuan Cui
> > > Sent: Wednesday, August 5, 2015 9:54 PM
> > > To: David Miller <davem@xxxxxxxxxxxxx>; KY Srinivasan
> > > <kys@xxxxxxxxxxxxx>
> > > Cc: olaf@xxxxxxxxx; gregkh@xxxxxxxxxxxxxxxxxxx;
> jasowang@xxxxxxxxxx;
> > > driverdev-devel@xxxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> > > stephen@xxxxxxxxxxxxxxxxxx; stefanha@xxxxxxxxxx;
> > > netdev@xxxxxxxxxxxxxxx; apw@xxxxxxxxxxxxx; pebolle@xxxxxxxxxx;
> > > dan.carpenter@xxxxxxxxxx
> > > Subject: RE: [PATCH V4 4/7] Drivers: hv: vmbus: add APIs to register
> callbacks
> > > to process hvsock connection
> > >
> > > > From: devel [mailto:driverdev-devel-bounces@xxxxxxxxxxxxxxxxxxxxxx]
> On
> > > Behalf
> > > > Of Dexuan Cui
> > > > Sent: Thursday, July 30, 2015 18:20
> > > > To: David Miller <davem@xxxxxxxxxxxxx>; KY Srinivasan
> > > <kys@xxxxxxxxxxxxx>
> > > > Cc: olaf@xxxxxxxxx; gregkh@xxxxxxxxxxxxxxxxxxx;
> jasowang@xxxxxxxxxx;
> > > > driverdev-devel@xxxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> > > > stephen@xxxxxxxxxxxxxxxxxx; stefanha@xxxxxxxxxx;
> > > netdev@xxxxxxxxxxxxxxx;
> > > > apw@xxxxxxxxxxxxx; pebolle@xxxxxxxxxx; dan.carpenter@xxxxxxxxxx
> > > > Subject: RE: [PATCH V4 4/7] Drivers: hv: vmbus: add APIs to register
> > > callbacks to
> > > > process hvsock connection
> > > >
> > > > > From: David Miller
> > > > > Sent: Thursday, July 30, 2015 6:27
> > > > >
> > > > > From: Dexuan Cui
> > > > > Date: Tue, 28 Jul 2015 05:35:11 -0700
> > > > >
> > > > > > With the 2 APIs supplied by the VMBus driver, the coming
> net/hvsock
> > > driver
> > > > > > can register 2 callbacks and can know when a new hvsock
> connection is
> > > > > > offered by the host, and when a hvsock connection is being closed
> by
> > > the
> > > > > > host.
> > > > > >
> > > > > This is an extremely terrible interface.
> > > > >
> > > > > It's an opaque hook that allows on registry, and it's solve purpose
> > > > > is to allow a backdoor call into a foreign driver in another module.
> > > > >
> > > > > These are exactly the things we try to avoid.
> > > >
> > > > Hi David,
> > > > Thanks a lot for your reviewing and the suggestion!
> > > >
> > > > > Why not create a real abstraction where clients register an object,
> > > > > that can be contained as a sub-member inside of their own driver
> > > > > private, that provides the callback registry mechanism.
> > >
> > > Hi David,
> > > Can you please have a look at my below questions?
> > >
> > > I like your idea of a real abstraction. Your answer would definitely
> > > help me to implement that correctly.
> > >
> > > > Please pardon me for my inexperience.
> > > > Can you please be a bit more specific?
> > > > I guess maybe you're referencing a common design pattern in the
> driver
> > > > code, so an example in some existing driver would be the best. :-)
> > > >
> > > > "clients register an object " --
> > > > does the "clients" mean the hvsock driver?
> > > > and the "object" means the 2 callbacks?
> > > >
> > > > IMHO, here the vmbus driver has to synchronously pass the 2 events
> > > > to the hvsock driver, so a "backdoor call into the hvsock driver" is
> > > > inevitable anyway?
> > > >
> > > > e.g., in the path vmbus_process_offer() -> hvsock_process_offer(), the
> > > > return value of the latter is important to the former, because on error
> > > > the former needs to clean up some internal states of the vmbus driver
> > > (that
> > > > is, the "goto err_deq_chan").
> > > >
> > > >
> > > > > That way you can register multiple clients, do things like allow
> > > > > AF_PACKET capturing of vmbus traffic, etc.
> > > >
> > > > I thought AF_PACKET can only capture IP packetsor Ethernet frames.
> > > > Can it be used to capture AF_UNIX packet?
> > > > If yes, I suppose we can consider making it work for AF_HYPERV too,
> > > > if people ask for that.
> > > >
> >
> > Dexuan,
> >
> > The notion of a channel on Hyper-V has been mapped to a device on Linux
> and
> > the mechanism we have
> > had of notifying the driver of the creation of the channel was through
> > registering this device with the kernel
> > (vmbus_device_create). The first exception to this was when we
> introduced
> > multi-channel support that broke
> > the assumption of this one to one mapping between the channel and Linux
> > device. In the case of the sub-channels,
> > we handled the driver notification issue via the sub-channel callback that
> the
> > driver registers at the point of
> > opening the channel. Perhaps we could make the sub-channel handling
> > mechanism more generic to handle the case
> > of VMSOCK as well?
> >
> > K. Y
>
> Good suggestion!
> Let me think this over and make a new patch.
>
> Thanks,
> -- Dexuan
>
Looks like VMSOCK handling cannot be accommodated as an extension of sub-channel handling.
However, I still think we should use the standard driver core primitives to handle VMSOCK devices.
What if we extend the vmbus match function to accommodate vmsock devices. Here is one way we
could extend the vmbus match function:

if (the device is not VMSOCK) {
Match the device type ID with the ID table supported by the driver
} else {
Match is based on a special attribute of the driver
}

By doing so, we can use the standard probe call to open the vmsock channel and use the standard mechanism for removing the device.

Regards,

K. Y
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/