Message-ID: <79bc446c-72a0-9209-98bc-e1d85a3a360a@pensando.io>
Date:   Mon, 27 Jan 2020 09:57:23 -0800
From:   Shannon Nelson <snelson@...sando.io>
To:     Leon Romanovsky <leon@...nel.org>, Jakub Kicinski <kuba@...nel.org>
Cc:     "David S . Miller" <davem@...emloft.net>,
        Michal Kalderon <michal.kalderon@...vell.com>,
        linux-netdev <netdev@...r.kernel.org>,
        RDMA mailing list <linux-rdma@...r.kernel.org>
Subject: Re: [PATCH net-next] net/core: Replace driver version to be kernel
 version

On 1/26/20 9:49 PM, Leon Romanovsky wrote:
> On Sun, Jan 26, 2020 at 01:33:53PM -0800, Jakub Kicinski wrote:
>> On Sun, 26 Jan 2020 23:08:50 +0200, Leon Romanovsky wrote:
>>> On Sun, Jan 26, 2020 at 12:49:57PM -0800, Jakub Kicinski wrote:
>>>> On Sun, 26 Jan 2020 21:41:10 +0200, Leon Romanovsky wrote:
>>>>>> This will end up affecting out-of-tree drivers as well, where it is useful
>>>>>> to know what the version number is, most especially since it is different
>>>>>> from what the kernel provided driver is.  How else are we to get this
>>>>>> information out to the user?  If this feature gets squashed, we'll end up
>>>>>> having to abuse some other mechanism so we can get the live information from
>>>>>> the driver, and probably each vendor will find a different way to sneak it
>>>>>> out, giving us more chaos than where we started.  At least the ethtool
>>>>>> version field is a known and consistent place for the version info.
>>>> Shannon does have a point that out of tree drivers still make use of
>>>> this field. Perhaps it would be a more suitable first step to print the
>>>> kernel version as default and add a comment saying upstream modules
>>>> shouldn't overwrite it (perhaps one day CI can catch new violators).
>>> Shannon proposed to remove this field and it was me who said no :)
>> Obviously, we can't remove fields from UAPI structs.
>>
>>> My plan is to overwrite ->version, delete all users and add
>>> WARN_ONEC(strcpy(..->version_)...) inside net/ethtool/ to catch
>>> abusers.
>> What I was thinking just now was: initialize ->version to utsname
>> before drivers are called, delete all upstream users, add a coccicheck
>> for upstream drivers which try to report the version.
>>
>>>> The NFP reports the git hash of the driver source plus the string
>>>> "(oot)" for out-of-tree:
>>>>
>>>> https://github.com/Netronome/nfp-drv-kmods/blob/master/src/Kbuild#L297
>>>> https://github.com/Netronome/nfp-drv-kmods/blob/master/src/Kbuild#L315
>>> I was inspired by upstream code.
>>> https://elixir.bootlin.com/linux/v5.5-rc7/source/drivers/net/ethernet/netronome/nfp/nfp_net_ethtool.c#L184
>> Right, upstream nfp reports kernel version (both in modinfo and ethtool)
>> GitHub/compat/backport/out-of-tree reports kernel version in which the
>> code was expected to appear in modinfo:
>>
>> https://github.com/Netronome/nfp-drv-kmods/commit/7ec15c47caf5dbdf1f9806410535ad5b7373ec34#diff-492d7fa4004d885a38cfa889ed1adbe7L1284
>>
>> And git hash of the driver source plus out of tree marker in ethtool.
>>
>> That means it's out-of-tree driver which has to carry the extra code
>> and require extra feeding. As backport should IMHO.
>>
>>>>> Leaving to deal with driver version to vendors is not an option too,
>>>>> because they prove for more than once that they are not capable to
>>>>> define user visible interfaces. It comes due to their natural believe
>>>>> that their company is alone in the world and user visible interface
>>>>> should be suitable for them only.
>>>>>
>>>>> It is already impossible for users to distinguish properly versions
>>>>> of different vendors, because they use arbitrary strings with some
>>>>> numbers.
>>>> That is true. But reporting the kernel version without even as much as
>>>> in-tree/out-of-tree indication makes the field entirely meaningless.
>>> The long-standing policy in kernel that we don't really care about
>>> out-of-tree code.
>> Yeah... we all know it's not that simple :)
> It is simple, unfortunately netdev people like to complicate things
> by declaring ABI in very vague way which sometimes goes so far that
> it ends more strict than anyone would imagine.
>
> We, RDMA and many other subsystems mentioned in that ksummit thread,
> removed MODULE_VERSION() a long time ago and got zero complains from
> the real users.
>
>> The in-tree driver versions are meaningless and cause annoying churn
>> when people arbitrarily bump them. If we can get people to stop doing
>> that we'll be happy, that's all there is to it.
>>
>> Out of tree the field is useful, so we don't have to take it away just
>> as a matter of principle. If we can't convince people to stop bringing
>> the versions into the tree that'll be another story...
> As Shannon pointed, even experienced people will try to sneak those
> changes. I assume that it is mainly because they are pushed to do it
> by the people who doesn't understand Linux kernel process.
>
I don't think that the Intel Networking folks were trying to "sneak" 
something through, I think they have simply been continuing the process 
that they've been following for years.  When we have groups such as 
them, with a long history of contributions to drivers and stack, not 
following the new rules, perhaps we need to take a look at how we're 
publicizing these changes.

sln

Powered by blists - more mailing lists