Date:   Thu, 10 Aug 2017 19:32:10 -0700
From:   Jerry Chu <hkchu@...gle.com>
To:     Rao Shoaib <rao.shoaib@...cle.com>
Cc:     David Miller <davem@...emloft.net>, codesoldier1@...il.com,
        Yuchung Cheng <ycheng@...gle.com>,
        Alexey Kuznetsov <kuznet@....inr.ac.ru>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>
Subject: Re: [PATCH v1 net] TCP_USER_TIMEOUT and tcp_keepalive should conform
 to RFC5482

On Thu, Aug 10, 2017 at 2:05 PM, Rao Shoaib <rao.shoaib@...cle.com> wrote:
>
>
> On 08/09/2017 09:59 PM, Jerry Chu wrote:
>>
>> On Wed, Aug 9, 2017 at 8:32 PM, Jerry Chu <hkchu@...gle.com> wrote:
>>>
>>> On Wed, Aug 9, 2017 at 5:47 PM, Rao Shoaib <rao.shoaib@...cle.com> wrote:
>>>>
>>>>
>>>> On 08/09/2017 05:30 PM, David Miller wrote:
>>>>>
>>>>> From: Joe Smith <codesoldier1@...il.com>
>>>>> Date: Wed, 9 Aug 2017 17:20:32 -0700
>>>>>
>>>>>> Making Linux conform to standards and behavior that is logical seems
>>>>>> like a good enough reason.
>>>>>
>>>>> That's an awesome attitude to have when we're implementing something
>>>>> new and don't have the facility already.
>>>>>
>>>>> But when we have something already the only important consideration is
>>>>> not breaking existing apps which rely on that behavior.
>>>>>
>>>>> That is much, much, more important than standards compliance.
>>>>>
>>>>> If users are confused, just fix the documentation.
>>>>
>>>> David,
>>>>
>>>> If it was just confusion than sure fixing the documentation is fine.
>>>> What if
>>>> the logic is incorrect, does not conform to the standard that is says it
>>>> is
>>>
>>> Not sure what part of logic is "incorrect" when it was a homegrown Linux
>>> API
>>> with no need to conform with any "standard"? Note that the new API was
>>> invented
>>> 7 years ago not out of need for RFC5482. In fact I initially call the
>>> option
>>> TCP_FAILFAST and did not even know the existence of RFC5482 until someone
>>> around the same time proposed a UTO option specifically for RFC5482 and I
>>> thought the two can be combined. (This is roughly the memory I can
>>> recollect so far.)
>>>
>>> So you see my focus back then was to devise a "failfast" option whereas
>>> RFC5482
>>> was meant for a "failslow" case. I think that explains why I let the
>>> option override
>>> keepalive so a TCP connection can "fail fast" while RFC5482 4.2 tries to
>>> prevent
>>> keepalive failure ahead of UTO, favoring "fail slow".
>>>
>>> If we start from a clean slate then perhaps one can argue the semantic
>>> either way
>>> but we do not have a clean slate. For that I still slightly favor not
>>> changing the code
>>> because the risk of breakage is definitely non-zero and the issue you're
>>> having
>>> seem to be only related to documentation.
>
> We all make mistakes and over look things, that seems to be the case here.

First of all, I do not understand what "mistake" has been committed? Like I
said, the API was not invented for RFC5482 specifically. It was really meant for
a "fail fast" case that was requested of me fir our internal use case
for a product
that back then had only 100 million users (but has mostly like grown to many
billions now).

> If this was so important than I am sure there was a use case. None has been
> presented. Without a use case I do not understand why we have to live with

I don't understand this one either. If you want a use case, you can
ask politely,
no need to flame or fume. Also why would I waste my energy on something
no one would use, then even more energy arguing now for maintaining its
semantic?

This API was requested of me by our Android group and has since been
deployed on all our Android servers. I initially pushed back and asked them to
just program a short KEEPALIVE time for all the Android connections but they
can't use it because it would wake up clients too often, unnecessarily
waste a lot
of battery power.

Its uses inside Google have proliferated since. Most recently (a few months ago)
I got an email from our Nest division saying they hit a bug related to
TCP_USER_TIMEOUT and even pointed me to
https://bugzilla.redhat.com/show_bug.cgi?id=1189241 (but i haven't got the time
to look into it and in the meantime they found a workaround)

So what's your use case then that really requires changing the implementation?

> broken logic when we have a chance to fix it and make the code better.
>
> If this change does break something (very very unlikely) we will understand
> the use case and provide an appropriate solution.
>
>> One more thing - the proposed patch compares TCP_KEEPIDLE against
>> TCP_USER_TIMEOUT. But I don't think TCP_KEEPIDLE is what the
>> "keep-alive
>> timer" referred to in RFC5482. Linux keepalive implementation seems to use
>> #
>> of retries (TCP_KEEPCNT) rather than time duration (keep-alive time) to
>> determine when to quit. If that is the case then your proposed change is
>> not
>> fully "compliant" either and the best is probably just don't change.
>
> Did you look at the patch and what it changes ?
>
> Take a look at the TCP_KEEPIDLE socket option and see what it does or just
> look at the man page of tcp(7)
>
> TCP_KEEPIDLE (since Linux 2.4) The time (in seconds) the connection needs to
> remain idle before TCP starts sending keepalive probes, if the socket option
> SO_KEEPALIVE has been set on this socket. This option should not be used in
> code intended to be portable.

Yes precisely my point above - the intent of section 4.2 in RFC5482, if i guess
correctly, is to avoid the case in some implementations where keep-alive code
uses a different timer than the regular retrans timer where TCP_USER_TIMEOUT
logic is allied, hence not subject to TCP_USER_TIMEOUT. E.g., if the
app programs
TCP_USER_TIMEOUT to be one hour, but a keepalive timeout to be 10 minutes,
RFC5482 does not want a connection to be aborted after 10 minutes due to
keepalive probe failure. This can be readily translated in some OSes if the
keepalive API is time/duration based. E.g., one can use the same
TCP_USER_TIMEOUT value to program the keepalive API.

But Linux's keepalive API has a TCP_KEEPIDLE period plus TCP_KEEPCNT of
retries plus TCP_KEEPINTVL. So depending on TCP_KEEPCNT and
TCP_KEEPINTVL, the total duration for keepalive may be >> TCP_KEEPIDLE
hence my comment above.

Again my main concern is backward compatibility. Our existing use cases most
likely have TCP_USER_TIMEOUT value < TCP_KEEPIDLE value so the risk
is mainly in removing the icsk_user_timeout check inside tcp_keepalive_timer().

Jerry

>
> Shoaib.
>
>>
>>> Jerry
>>>
>>>> implementing and easy to fix with little or no risk of breakage.
>>>>
>>>> The proposed patch changes a feature that no one uses. It also imposes
>>>> the
>>>> relation ship between keepalive and timeout values that is required by
>>>> the
>>>> RFC and make sense.
>>>>
>>>> You are the final authority, if you say we should just fix the
>>>> documentation
>>>> than that is fine.
>>>>
>>>> Shoaib
>>>>
>

Powered by blists - more mailing lists