Re: [PATCH RFC 0/3] Drivers: hv: utils: re-implement the kernel/userspace communication layer

[Vitaly Kuznetsov <vkuznets@redhat.com>]

Radim Krčmář <rkrcmar@redhat.com> writes:

> 2015-02-27 17:14+0100, Vitaly Kuznetsov:
>> This series converts kvp/vss daemons to use misc char devices instead of
>> netlink for userspace/kernel communication and then updates fcopy to be
>> consistent with kvp/vss.
>> 
>> Userspace/kernel communication via netlink has a number of issues:
>> - It is hard for userspace to figure out if the kernel part was loaded or not
>>   and this fact can change as there is a way to enable/disable the service from
>>   host side.
>
> (Hm, this should be just a message to the userspace daemon, but netlink
>  probably makes it complicated anyway.)
>
>>              Racy daemon startup is also a problem.
>
> (Is it significantly worse than what we need to protect devices?)
>
>> - When the userspace daemon restarts/dies kernel part doesn't receive a
>>   notification.
>
> (True, we could use a other-side-closed callback.)

With normal devices we can use e.g. udev/systemd machinery to start/stop
service on device hotplug/hotunplug (and these devices are actually
pluggable/unpluggable from host side) without any special code in
kernel/userspace parts and I'd like to use that.

>
>> - Netlink communication is not stable under heavy load.
>
> (The message order changes?)
>

It is a disaster if it does (the whole transaction will get lost). Same
if any of these messages gets lost.

>> RFC: I'm a bit puzzled on how to split commits 1 and 2 avoiding breakages.
>
> Split the userspace part -- it won't break bisects.
>

Sure if it simplifies the review.

> And then, you could refactor drivers first ... the way we communicate
> with userspace should have little impact on what the rest does (or how).
> At first sight, there are three units, apart from glue,
>  1) communication with host
>  2) communication with userspace
>  3) repacking of data between first two
>
> With an API for userspace communication, the amount of code to replace
> netlink could be lower and resulting patches definitely easier to
> review.  (And with extra work, both ABIs could even live side-by-side ;)

Ok, thanks!

-- 
  Vitaly