Message-ID: <AANLkTi=sLkXreC3LNsDuFDSeNMebr+ScMc12=gngd2cc@mail.gmail.com>
Date:	Thu, 17 Mar 2011 13:16:15 -0700
From:	Jerry Chu <hkchu@...gle.com>
To:	Eric Dumazet <eric.dumazet@...il.com>
Cc:	netdev@...r.kernel.org
Subject: Re: [PATCH] Add useful per-connection TCP stats for diagnosis purpose.

Eric, thanks for the prompt feedback.

On Thu, Mar 17, 2011 at 1:42 AM, Eric Dumazet <eric.dumazet@...il.com> wrote:
> Le jeudi 17 mars 2011 à 01:06 -0700, H.K. Jerry Chu a écrit :
>> From: Jerry Chu <hkchu@...gle.com>
>>
>> This patch add a number of very useful counters/stats (defined in
>> tcp_stats.h) to help diagnosing TCP related problems.
>>
>> create_time     - when the connection was created (in jiffies)
>> total_inbytes   - total inbytes as consumed by the receiving apps.
>> total_outbytes  - total outbytes sent down from the transmitting apps.
>>
>> total_outdatasegs - total data carrying segments sent so far, including
>>               retransmitted ones.
>>
>> total_xmit      - total accumulated time (usecs) when the connection
>>               has something to send.
>>
>> total_retrans_time - total time (usecs, accumulated) the connection
>>               spends trying to recover lost packets. For each
>>               loss event the time is measured from the lost packet
>>               was first sent till the retransmitted packet was
>>               eventually ack'ed.
>>
>> total_cwnd_limit - total time (usecs, excluding time spent on loss
>>               recovery) the xmit is stopped due to cwnd limited
>>
>> total_swnd_limit - total time (usecs) theconnection is swnd limited
>>
>> The following two counters are for listeners only:
>>
>> accepted_reqs   - total # of accepted connection requests.
>> listen_drops    - total # of dropped SYN reqs (SYN cookies excluded) due
>>               to listener's queue overflow.
>>
>> total_retrans_time/total_retrans ratio gives a rough picture of how
>> quickly in average the connection can recover from a pkt loss. E.g.,
>> when the network is more congested, or the traffic contains mainly
>> smaller RPC where tail drop often requires RTO to recover,
>> the total_retrans_time/total_retrans ratio tends to be higher.
>>
>> Currently the new counters/stats are exported through /proc/net/tcp.
>
> Please dont. Use iproute2 instead.
>
>> Some simple, abbreviated field names have been added to the output of
>> /proc/net/tcp in order to allow backward/forward compatibility in the
>> future. Obviously the new counters/stats can also be easily exported
>> through other APIs.
>>
>
> /proc/net/tcp is legacy. You should touch it eventually, but after
> "other APIS" are done. It was the old way (quick but a bit ugly)

Understood. /proc/net/tcp is a much more expedient way of exporting these
counters because it doesn't requires any additional, special tool to read it,
unless other APIs (e.g., netlink). Note that backward compatibility to
/proc/net/tcp has been ensured by adding field names in the heading.

>
>
>
>> Signed-off-by: H.K. Jerry Chu <hkchu@...gle.com>
>> ---
>>  include/linux/ktime.h    |    3 ++
>>  include/linux/tcp.h      |    1 +
>>  include/net/tcp_stats.h  |   65 ++++++++++++++++++++++++++++++++++++++++++++++
>>  net/ipv4/tcp.c           |   30 ++++++++++++++++++---
>>  net/ipv4/tcp_input.c     |   13 +++++++++
>>  net/ipv4/tcp_ipv4.c      |   41 ++++++++++++++++++++++++++---
>>  net/ipv4/tcp_minisocks.c |    9 ++++++
>>  net/ipv4/tcp_output.c    |   47 +++++++++++++++++++++++++++++++--
>>  net/ipv6/tcp_ipv6.c      |    8 +++++
>>  9 files changed, 206 insertions(+), 11 deletions(-)
>>  create mode 100644 include/net/tcp_stats.h
>>
>> diff --git a/include/linux/ktime.h b/include/linux/ktime.h
>> index e1ceaa9..e60e758 100644
>> --- a/include/linux/ktime.h
>> +++ b/include/linux/ktime.h
>> @@ -333,6 +333,9 @@ extern void ktime_get_ts(struct timespec *ts);
>>  /* Get the real (wall-) time in timespec format: */
>>  #define ktime_get_real_ts(ts)        getnstimeofday(ts)
>
> Hmm, this kind of changes are out of netdev scope and should be avoided

Ok. (It was moved out of tcp_stats.h only at the last minute.)

>
>>
>> +#define ktime_since(a)               ktime_to_us(ktime_sub(ktime_get(), (a)))
>
> us are implied in ktime_since() ? thats strange.

Ok.

>
>> +#define ktime_zero(a)                ktime_equal((a), ktime_set(0, 0))
>
> ktime_zero() sounds like : "give me zero time" or "clear the ktime
> field".

Yes I actually have been flip-flopping on the name...

>
>> +
>>  static inline ktime_t ns_to_ktime(u64 ns)
>>  {
>>       static const ktime_t ktime_zero = { .tv64 = 0 };
>> diff --git a/include/linux/tcp.h b/include/linux/tcp.h
>> index e64f4c6..ea5cb5d 100644
>> --- a/include/linux/tcp.h
>> +++ b/include/linux/tcp.h
>> @@ -460,6 +460,7 @@ struct tcp_sock {
>>        * contains related tcp_cookie_transactions fields.
>>        */
>>       struct tcp_cookie_values  *cookie_values;
>> +     struct tcp_stats        *conn_stats;
>>  };
>
> Really, using separate cache lines to store some stats is expensive.
> You should add counters in existing structure, to avoid additional cache
> line dirties. Carefully placing stats in already dirtied cache lines.

This was how it was done initially but then we wanted to allow future
extension to include possibly a lot more counters, something like Web100
(RFC4898). For the latter the memory/performance hit will likely require
a config option, and a separate structure will make this easier. Does it
make sense?

>
> You also should use native ktime_t infrastructure, to make the maths
> really fast in fast path.
>
> Only when stats are to be returned to user, you'll have to convert the
> native timestamps to user exportable ones.

Good point! Will do. (I mistakenly thought ktime_t is a larger structure.)

>
> Quite frankly, using u64 fields allow nanosec resolution.

I wish to use less bits because the final report only needs ms or even
sec resolution but the intermediate computation needs to capture usec
resolution.

>
> BTW, we probably could 'export' sk->sk_drops for TCP, like we do for
> UDP.

There are many other potentially useful counters/stats (spurious_retrans,
min_rtt, total_rto,...) but there is a tradeoff against memory/performance hit
so for the first round I'm focusing on what i feel is the most useful set.

Thanks,

Jerry

>
>
>
>
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists