Message-ID: <BE9C38F5-42FB-4EFA-ADFB-3D13E6ABC1EE@intel.com>
Date:   Tue, 4 Jun 2019 19:47:54 +0000
From:   "Patel, Vedang" <vedang.patel@...el.com>
To:     Murali Karicheri <m-karicheri2@...com>
CC:     "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        "Kirsher, Jeffrey T" <jeffrey.t.kirsher@...el.com>,
        "David S . Miller" <davem@...emloft.net>,
        Jamal Hadi Salim <jhs@...atatu.com>,
        Cong Wang <xiyou.wangcong@...il.com>,
        Jiri Pirko <jiri@...nulli.us>,
        "intel-wired-lan@...ts.osuosl.org" <intel-wired-lan@...ts.osuosl.org>,
        "Gomes, Vinicius" <vinicius.gomes@...el.com>,
        "l@...ileo.org" <l@...ileo.org>
Subject: Re: [PATCH net-next v1 0/7] net/sched: Add txtime assist support
 for taprio

Hi Murali,

> On Jun 3, 2019, at 6:50 AM, Murali Karicheri <m-karicheri2@...com> wrote:
> 
> Hi Vedang,
> 
> On 05/28/2019 01:46 PM, Vedang Patel wrote:
>> Currently, we are seeing packets being transmitted outside their timeslices. We
>> can confirm that the packets are being dequeued at the right time. So, the
>> delay is induced after the packet is dequeued, because taprio, without any
>> offloading, has no control of when a packet is actually transmitted.
>> In order to solve this, we are making use of the txtime feature provided by ETF
>> qdisc. Hardware offloading needs to be supported by the ETF qdisc in order to
>> take advantage of this feature. The taprio qdisc will assign txtime (in
>> skb->tstamp) for all the packets which do not have the txtime allocated via the
>> SO_TXTIME socket option. For the packets which already have SO_TXTIME set,
>> taprio will validate whether the packet will be transmitted in the correct
>> interval.
>> In order to support this, the following parameters have been added:
>> - offload (taprio): This is added in order to support different offloading
>>   modes which will be added in the future.
>> - txtime-delay (taprio): this is the time the packet takes to traverse through
>>   the kernel to adapter card.
> 
> I am very new to this TC and QoS handling of Linux kernel and TSN. So sorry some of the  questions below are very basic in nature. I would soon would be working to add this support in our platform based on this.
> So please answer.
No problems. This is my first contribution to networking subsystem as well. :)
> 
> Is txtime-delay from the instance qdisc de-queue the packet to the time
> packets get onto the wire? I am wondering if this time is deterministic
> or we have some way to ensure this can be tuned to meet a max value?
> Also how would one calculate this value or have to measure it?
> 
So, consider a scenario where we just received a packet to send and the gate corresponding to that packet is open. But, we are not sure whether we can transmit the packet fully before the gate closes. Txtime-delay will be useful parameter in such situations. So, it is basically the time between start of taprio_enqueue() and packet hitting the wire.

It’s probably going to be a little tricky to measure because the packet changes hands. But, you could get reasonable accuracy  if you print out the CLOCK_REALTIME at the beginning of the packet and substrat it from the Hardware transmit timestamp of the packet. You will have to run phc2sys at least to ensure that the PTP clock in network card and CLOCK_REALTIME are synchronized. 
>> - skip_sock_check (etf): etf qdisc currently drops any packet which does not
>>   have the SO_TXTIME socket option set. This check can be skipped by specifying
>>   this option.
>> Following is an example configuration:
>> $ tc qdisc replace dev $IFACE parent root handle 100 taprio \\
>>       num_tc 3 \\
>>       map 2 2 1 0 2 2 2 2 2 2 2 2 2 2 2 2 \\
>>       queues 1@0 1@0 1@0 \\
>>       base-time $BASE_TIME \\
>>       sched-entry S 01 300000 \\
>>       sched-entry S 02 300000 \\
>>       sched-entry S 04 400000 \\
>>       offload 2 \\
>>       txtime-delay 40000 \\
>>       clockid CLOCK_TAI
> 
> Could you tell me what is CLOCK_TAI?? I see below in the source code for
> the definition in include/uapi/linux/time.h
It is International atomic time. It is currently exactly 37 seconds ahead of UTC (or CLOCK_REALTIME). 

More info: https://en.wikipedia.org/wiki/International_Atomic_Time
> 
> /*
> * The driver implementing this got removed. The clock ID is kept as a
> * place holder. Do not reuse!
> */
> #define CLOCK_SGI_CYCLE			10
> #define CLOCK_TAI			11
> 
> So why do I see such defines being used in the example that is being
> removed?
> 
I believe the comment is just referring to CLOCK_SGI_CYCLE. CLOCK_TAI is still being supported.
> AFAIK, From the usage point of view, TSN requires the network being
> synchronized through linux PTP. For synchronization phc2sys is used to
> synchronize system time to the PHC. So why don't one use system time for
> this?
> 
> So application will use clock_gettime() to know current time and
> schedule the packet for transmission as well as user would use scripts
> or such to check current system time and issue tc command above. So the
> command should use CLOCK_REALTIME in that case. So all h/w and software
> are aligned to the same time source. Or Have I understood it wrong? I am
> assuming that for the offloaded case, h/w schedule Gate open, send
> packet etc are synchronized to the PHC or use a translated time w.r.t the common time (network time. Assuming PHC tracks this).
> 
CLOCK_REALTIME is also supported by the driver. So, you can definitely use that. CLOCK_TAI is basically CLOCK_REALTIME + UTC_Offset. So, you can use either.
> Thanks in advance for clarifying this.
> 
>> $ tc qdisc replace dev $IFACE parent 100:1 etf \\
>>       offload delta 200000 clockid CLOCK_TAI skip_sock_check
>> Here, the "offload" parameter is indicating that the TXTIME_OFFLOAD mode is
>> enabled. Also, that all the traffic classes have been assigned the same queue.
>> This is to prevent the traffic classes in the lower priority queues from
>> getting starved. Note that this configuration is specific to the i210 ethernet
>> card. Other network cards where the hardware queues are given the same
>> priority, might be able to utilize more than one queue.
>> Following are some of the other highlights of the series:
>> - Fix a bug where hardware timestamping and SO_TXTIME options cannot be used
>>   together. (Patch 1)
>> - Introduce hardware offloading. This patch introduces offload parameter which
>>   can be used to enable the txtime offload mode. It will also support enabling
>>   full offload when the support is available in drivers. (Patch 3)
>> - Make TxTime assist mode work with TCP packets. (Patch 6 & 7)
>> The following changes are recommended to be done in order to get the best
>> performance from taprio in this mode:
>> # ip link set dev enp1s0 mtu 1514
> 
> May I know why MTU is changed to 1514 to include the Ethernet header?
> 
For TSN, in general, Jumbo frames are not allowed. Which is why this setting is used.
>> # ethtool -K eth0 gso off
>> # ethtool -K eth0 tso off
Let’s consider a scenario where an application is trying to send a lot of small packets. By default, these packets will be combined into a single packet and this packet will be segmented in the hardware. But, we are not sure whether these packets can be transmitted before the gate for this application closes. So, these options are used to prevent this from happening. 
>> # ethtool --set-eee eth0 eee off
> 
This option just disables energy efficient ethernet so that we do not suffer from latency spikes due to the NIC trying to wake up when the packet is supposed to be sent.

Thanks,
Vedang
> Could you please explain why these are needed?
> 
> Thanks
> 
> Murali
>> Thanks,
>> Vedang Patel
>> Vedang Patel (6):
>>   igb: clear out tstamp after sending the packet.
>>   etf: Add skip_sock_check
>>   taprio: calculate cycle_time when schedule is installed
>>   taprio: Add support for txtime offload mode.
>>   taprio: make clock reference conversions easier
>>   taprio: Adjust timestamps for TCP packets.
>> Vinicius Costa Gomes (1):
>>   taprio: Add the skeleton to enable hardware offloading
>>  drivers/net/ethernet/intel/igb/igb_main.c |   1 +
>>  include/uapi/linux/pkt_sched.h            |   6 +
>>  net/sched/sch_etf.c                       |  10 +
>>  net/sched/sch_taprio.c                    | 548 ++++++++++++++++++++--
>>  4 files changed, 532 insertions(+), 33 deletions(-)

Powered by blists - more mailing lists