Message-ID: <778d990f-5e07-b7af-498e-61f69c26f040@cambridgegreys.com>
Date:   Tue, 10 Dec 2019 07:08:40 +0000
From:   Anton Ivanov <anton.ivanov@...bridgegreys.com>
To:     Brendan Higgins <brendanhiggins@...gle.com>
Cc:     johannes.berg@...el.com, Richard Weinberger <richard@....at>,
        Jeff Dike <jdike@...toit.com>, linux-um@...ts.infradead.org,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        David Gow <davidgow@...gle.com>
Subject: Re: [RFC v1 1/2] um: drivers: remove support for UML_NET_PCAP

On 09/12/2019 23:40, Brendan Higgins wrote:
> On Sat, Dec 7, 2019 at 1:15 AM Anton Ivanov
> <anton.ivanov@...bridgegreys.com> wrote:
>> On 07/12/2019 01:21, Brendan Higgins wrote:
>>> On Fri, Dec 06, 2019 at 04:32:34PM -0800, Brendan Higgins wrote:
>>>> On Thu, Dec 5, 2019 at 11:23 PM Anton Ivanov
>>>> <anton.ivanov@...bridgegreys.com> wrote:
>>>> [...]
>>>>> 1. There is a proposed patch for the build system to fix it.
>>> So I just tried the patch you linked on the cover letter[1], and I am
>>> still getting the build error described above:
>>>
>>> arch/um/drivers/pcap_user.c:35:12: error: conflicting types for ‘pcap_open’
>>>    static int pcap_open(void *data)
>>>               ^~~~~~~~~
>>> In file included from /usr/include/pcap.h:43,
>>>                    from arch/um/drivers/pcap_user.c:7:
>>> /usr/include/pcap/pcap.h:859:18: note: previous declaration of ‘pcap_open’ was here
>>>    PCAP_API pcap_t *pcap_open(const char *source, int snaplen, int flags,
>>>
>>> Looking at the patch, I wouldn't expect it to solve this problem.
>>>
>>> Are there maybe different conflicting libpcap-dev libraries and I have
>>> the wrong one? Or is this just still broken?
>>>
>>>>> 2. We should be removing all old drivers and replacing them with the
>>>>> vector ones.
>>>> Hmm...does this mean you would entertain a patch removing all the
>>>> non-vector UML network drivers? I would be happy to see VDE go as
>>>> well.
>>>>
>>>> In any event, it sounds like I should probably drop this patch as it
>>>> is currently.
>>>>
>>>> Thanks!
>>> [1] https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=938962#79
>>>
>>> _______________________________________________
>>> linux-um mailing list
>>> linux-um@...ts.infradead.org
>>> http://lists.infradead.org/mailman/listinfo/linux-um
>>>
>> OK, looks like the pcap.h differs now as well.
>>
>> I will fix that too. It looks like you need both a pcap fix and a
>> library linking fix for this to work.
>>
>> The patch fixes the issue with the build system which no longer provides
>> the means for UML to specify extra libraries (I probably had an older
>> pcap version on the machine where I tested this).
>>
>> IMHO frankly it is no longer necessary.
>>
>> 5.5-rc1 vector raw now has the facility to add/remove (including at
>> runtime) filters compiled with pcap outside UML. It was merged this week.
>>
>> We now have the following line-up for vector drivers - EoGRE, EoL2TPv3,
>> RAW (+/- BPF), TAP and BESS. Speeds are 2.5 to 9Gbit on my machine
>> (mid-range Ryzen desktop).
>>
>> If I figure out a way to get hold of the underlying tap raw sockets the
>> same way vhost does, TAP can probably go to 12Gbit or thereabouts. Same
>> applies to getting zerocopy working with raw.
>>
>> As a basis for comparison I get 18Gbit on the same machine using vEth
>> and containers. 50% of that is actually a very decent number.
>>
>> While vector drivers are not 1:1 replacements for the existing drivers,
>> you can achieve the same topologies and the same connectivity at much
>> higher performance - the old drivers test out in the 500Mbit range on
>> the same hardware.
>>
>> IMHO we should at least mark them as "obsolete" and start preparing to
>> remove them.
> Alright, I will send a patch out which marks the other network drivers
> as "obsolete".
>
> Clarification: Should I mark all UML network devices as "obsolete"
> except for NET_VECTOR? Daemon and MCAST looked to me (I am not a
> networking expert), like they might not be covered by vector.
>
They are not 1:1 replacements unfortunately.

However, in order to fix daemon I have to rewrite the switch from 
uml-utilities too. It is beyond obsolete.

MCAST for 2 UML instances can be replaced by either GRE or L2TPv3, for 
more than 2 you are better off introducing a proper switch.

I am OK if all old drivers are marked as obsolete at this point, so 
please proceed with the patch.

Brgds,

-- 
Anton R. Ivanov
Cambridgegreys Limited. Registered in England. Company Number 10273661
https://www.cambridgegreys.com/

Powered by blists - more mailing lists