Re: [PATCH v2 1/3] net/bonding: support Tx prepare

[Chas Williams <3chas3@gmail.com>]

On 9/26/22 06:18, Konstantin Ananyev wrote:
> 
> Hi everyone,
> 
> Sorry for late reply.
> 
>>>>>>> The main problem is hard to design a tx_prepare for bonding device:
>>>>>>> 1. as Chas Williams said, there maybe twice hash calc to get target slave
>>>>>>>        devices.
>>>>>>> 2. also more important, if the slave devices have changes(e.g. slave device
>>>>>>>        link down or remove), and if the changes happens between bond-tx-prepare and
>>>>>>>        bond-tx-burst, the output slave will changes, and this may lead to checksum
>>>>>>>        failed. (Note: a bond device with slave devices may from different vendors,
>>>>>>>        and slave devices may have different requirements, e.g. slave-A support calc
>>>>>>>        IPv4 pseudo-head automatic (no need driver pre-calc), but slave-B need driver
>>>>>>>        pre-calc).
>>>>>>>
>>>>>>> Current design cover the above two scenarios by using in-place tx-prepare. and
>>>>>>> in addition, bond devices are not transparent to applications, I think it's a
>>>>>>> practical method to provide tx-prepare support in this way.
>>>>>>>
>>>>>>
>>>>>>
>>>>>> I don't think you need to export an enable/disable routine for the use of
>>>>>> rte_eth_tx_prepare. It's safe to just call that routine, even if it isn't
>>>>>> implemented. You are just trading one branch in DPDK librte_eth_dev for a
>>>>>> branch in drivers/net/bonding.
>>>>>
>>>>> Our first patch was just like yours (just add tx-prepare default), but community
>>>>> is concerned about impacting performance.
>>>>>
>>>>> As a trade-off, I think we can add the enable/disable API.
>>>>
>>>> IMHO, that's a bad idea. If the rte_eth_dev_tx_prepare API affects
>>>> performance adversly, that is not a bonding problem. All applications
>>>> should be calling rte_eth_dev_tx_prepare. There's no defined API
>>>> to determine if rte_eth_dev_tx_prepare should be called. Therefore,
>>>> applications should always call rte_eth_dev_tx_prepare. Regardless,
>>>> as I previously mentioned, you are just trading the location of
>>>> the branch, especially in the bonding case.
>>>>
>>>> If rte_eth_dev_tx_prepare is causing a performance drop, then that API
>>>> should be improved or rewritten. There are PMD that require you to use
>>>> that API. Locally, we had maintained a patch to eliminate the use of
>>>> rte_eth_dev_tx_prepare. However, that has been getting harder and harder
>>>> to maintain. The performance lost by just calling rte_eth_dev_tx_prepare
>>>> was marginal.
>>>>
>>>>>
>>>>>>
>>>>>> I think you missed fixing tx_machine in 802.3ad support. We have been using
>>>>>> the following patch locally which I never got around to submitting.
>>>>>
>>>>> You are right, I will send V3 fix it.
>>>>>
>>>>>>
>>>>>>
>>>>>>    From a458654d68ff5144266807ef136ac3dd2adfcd98 Mon Sep 17 00:00:00 2001
>>>>>> From: "Charles (Chas) Williams" <chwillia@ciena.com>
>>>>>> Date: Tue, 3 May 2022 16:52:37 -0400
>>>>>> Subject: [PATCH] net/bonding: call rte_eth_tx_prepare before rte_eth_tx_burst
>>>>>>
>>>>>> Some PMDs might require a call to rte_eth_tx_prepare before sending the
>>>>>> packets for transmission. Typically, the prepare step handles the VLAN
>>>>>> headers, but it may need to do other things.
>>>>>>
>>>>>> Signed-off-by: Chas Williams <chwillia@ciena.com>
>>>>>
>>>>> ...
>>>>>
>>>>>>                  * ring if transmission fails so the packet isn't lost.
>>>>>> @@ -1322,8 +1350,12 @@ bond_ethdev_tx_burst_broadcast(void *queue, struct rte_mbuf **bufs,
>>>>>>
>>>>>>         /* Transmit burst on each active slave */
>>>>>>         for (i = 0; i < num_of_slaves; i++) {
>>>>>> -        slave_tx_total[i] = rte_eth_tx_burst(slaves[i], bd_tx_q->queue_id,
>>>>>> +        uint16_t nb_prep;
>>>>>> +
>>>>>> +        nb_prep = rte_eth_tx_prepare(slaves[i], bd_tx_q->queue_id,
>>>>>>                         bufs, nb_pkts);
>>>>>> +        slave_tx_total[i] = rte_eth_tx_burst(slaves[i], bd_tx_q->queue_id,
>>>>>> +                    bufs, nb_prep);
>>>>>
>>>>> The tx-prepare may edit packet data, and the broadcast mode will send a packet to all slaves,
>>>>> the packet data is sent and edited at the same time. Is this likely to cause problems ?
>>>>
>>>> This routine is already broken. You can't just increment the refcount
>>>> and send the packet into a PMD's transmit routine. Nothing guarantees
>>>> that a transmit routine will not modify the packet. Many PMDs perform an
>>>> rte_vlan_insert.
>>>
>>> Hmm interesting....
>>> My uderstanding was quite opposite - tx_burst() can't modify packet data and metadata
>>> (except when refcnt==1 and tx_burst() going to free the mbuf and put it back to the mempool).
>>> While tx_prepare() can - actually as I remember that was one of the reasons why a separate routine
>>> was introduced.
>>
>> Is that documented anywhere?
> 
> I looked through, but couldn't find too much except what was already mentioned by Fengcheng:
> rte_eth_tx_prepare() notes:
>      * Since this function can modify packet data, provided mbufs must be safely
>      * writable (e.g. modified data cannot be in shared segment).
> Probably that's not explicit enough, as it doesn't forbid modifying packets in tx_burst clearly.

This certainly seems like one of those gray areas in the DPDK APIs. It
should be made clear what is expected as far as behavior.

> 
>> It's been my experience that the device PMD
>> can do practically anything and you need to protect yourself.  Currently,
>> the af_packet, dpaa2, and vhost driver call rte_vlan_insert. Before 2019,
>> the virtio driver also used to call rte_vlan_insert during its transmit
>> path. Of course, rte_vlan_insert modifies the packet data and the mbuf
>> header.
> Interesting, usually apps that trying to use zero-copy multi-cast TX have packet-header portion
> in a separate segment, so it might even keep working.. But definetly doesn't look right to me:
> if mbuf->refnct > 1,  I think it should be treated as read-only.


rte_vlan_insert might be a problem with broadcast mode. If the refcnt is
> 1, rte_vlan_insert is going to fail. So, the current broadcast mode
implementation probably doesn't work if any PMD uses rte_vlan_insert.

So again, a solution is call tx_pkt_prepare once, then increment the
reference count, and send to the all the members. That works if your
PMD correctly implements tx_pkt_prepare. If it doesn't and call rte_vlan_insert
in the transmit routine, that PMD will need to be fixed to work with
bonding.

  
>   Regardless, it looks like rte_eth_dev_tx_prepare should always be
>> called.
> 
> Again, as I remember, initial agreement was: if any TX offload is enabled,
> tx_prepare() needs to be called (or user has implement similar stuff on his own).
> If no TX offload flags were specified for the packet, tx_prepare() is not necessary.

For the bonding driver, we potentially have a mix of PMDs for the members.
It's difficult to know in advance if your packets will have TX offload flags
or not. If you have a tx_pkt_prepare stub, there's a good chance that your
packets will have some TX offload flags. So, calling tx_pkt_prepare is likely
the "best" intermediate solution.

> 
>   > Handling that correctly in broadcast mode probably means always
>> make a deep copy of the packet, or check to see if all the members are
>> the same PMD type. If so, you can just call prepare once. You could track
>> the mismatched nature during additional/removal of the members. Or just
>> assume people aren't going to mismatch bonding members.
>>
>>
>>>> You should at least perform a clone of the packet so
>>>> that the mbuf headers aren't mangled by each PMD.
> 
> Usually you don't need to clone the whole packet. In many cases it is enough to just attach
> as first segment l2/l3/l4 header portion of the packet.
> At least that's how ip_multicast sample works.

Yes, that's what I meant by deep copy the packet headers. You just copy
enough to modify what you need and keep the bulk of the packet otherwise.

> 
> Just to be safe you
>>>> should perform a partial deep copy of the packet headers in case some
>>>> PMD does an rte_vlan_insert and the other PMDs in the bonding group do
>>>> not need an rte_vlan_insert.
>>>>
>>>> So doing a blind rte_eth_dev_tx_preprare isn't making anything much
>>>> worse.
>>>>
>>>>>
>>>>>>
>>>>>>             if (unlikely(slave_tx_total[i] < nb_pkts))
>>>>>>                 tx_failed_flag = 1;