DPDK patches and discussions
 help / color / mirror / Atom feed
From: Jie Hai <haijie1@huawei.com>
To: Ferruh Yigit <ferruh.yigit@amd.com>,
	huangdengdui <huangdengdui@huawei.com>, <dev@dpdk.org>
Cc: <lihuisong@huawei.com>, <fengchengwen@huawei.com>,
	<liuyonglong@huawei.com>
Subject: Re: [PATCH] net/hns3: fix Rx packet truncation when KEEP CRC enabled
Date: Fri, 8 Mar 2024 19:36:30 +0800	[thread overview]
Message-ID: <eea7caf7-d4b9-c146-9f7a-186baa763e38@huawei.com> (raw)
In-Reply-To: <541e9d04-bed3-47b0-8b11-746f9047e1a0@amd.com>

On 2024/3/1 19:10, Ferruh Yigit wrote:
> On 3/1/2024 6:55 AM, huangdengdui wrote:
>>
>>
>> On 2024/2/29 17:25, Ferruh Yigit wrote:
>>> On 2/29/2024 3:58 AM, huangdengdui wrote:
>>>>
>>>>
>>>> On 2024/2/28 21:07, Ferruh Yigit wrote:
>>>>> On 2/28/2024 2:27 AM, huangdengdui wrote:
>>>>>>
>>>>>>
>>>>>> On 2024/2/27 0:43, Ferruh Yigit wrote:
>>>>>>> On 2/26/2024 3:16 AM, Jie Hai wrote:
>>>>>>>> On 2024/2/23 21:53, Ferruh Yigit wrote:
>>>>>>>>> On 2/20/2024 3:58 AM, Jie Hai wrote:
>>>>>>>>>> Hi, Ferruh,
>>>>>>>>>>
>>>>>>>>>> Thanks for your review.
>>>>>>>>>>
>>>>>>>>>> On 2024/2/7 22:15, Ferruh Yigit wrote:
>>>>>>>>>>> On 2/6/2024 1:10 AM, Jie Hai wrote:
>>>>>>>>>>>> From: Dengdui Huang <huangdengdui@huawei.com>
>>>>>>>>>>>>
>>>>>>>>>>>> When KEEP_CRC offload is enabled, some packets will be truncated and
>>>>>>>>>>>> the CRC is still be stripped in following cases:
>>>>>>>>>>>> 1. For HIP08 hardware, the packet type is TCP and the length
>>>>>>>>>>>>       is less than or equal to 60B.
>>>>>>>>>>>> 2. For other hardwares, the packet type is IP and the length
>>>>>>>>>>>>       is less than or equal to 60B.
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> If a device doesn't support the offload by some packets, it can be
>>>>>>>>>>> option to disable offload for that device, instead of calculating it in
>>>>>>>>>>> software and append it.
>>>>>>>>>>
>>>>>>>>>> The KEEP CRC feature of hns3 is faulty only in the specific packet
>>>>>>>>>> type and small packet(<60B) case.
>>>>>>>>>> What's more, the small ethernet packet is not common.
>>>>>>>>>>
>>>>>>>>>>> Unless you have a specific usecase, or requirement to support the
>>>>>>>>>>> offload.
>>>>>>>>>>
>>>>>>>>>> Yes, some users of hns3 are already using this feature.
>>>>>>>>>> So we cannot drop this offload
>>>>>>>>>>
>>>>>>>>>>> <...>
>>>>>>>>>>>
>>>>>>>>>>>> @@ -2492,10 +2544,16 @@ hns3_recv_pkts_simple(void *rx_queue,
>>>>>>>>>>>>                 goto pkt_err;
>>>>>>>>>>>>               rxm->packet_type = hns3_rx_calc_ptype(rxq, l234_info,
>>>>>>>>>>>> ol_info);
>>>>>>>>>>>> -
>>>>>>>>>>>>             if (rxm->packet_type == RTE_PTYPE_L2_ETHER_TIMESYNC)
>>>>>>>>>>>>                 rxm->ol_flags |= RTE_MBUF_F_RX_IEEE1588_PTP;
>>>>>>>>>>>>     +        if (unlikely(rxq->crc_len > 0)) {
>>>>>>>>>>>> +            if (hns3_need_recalculate_crc(rxq, rxm))
>>>>>>>>>>>> +                hns3_recalculate_crc(rxq, rxm);
>>>>>>>>>>>> +            rxm->pkt_len -= rxq->crc_len;
>>>>>>>>>>>> +            rxm->data_len -= rxq->crc_len;
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> Removing 'crc_len' from 'mbuf->pkt_len' & 'mbuf->data_len' is
>>>>>>>>>>> practically same as stripping CRC.
>>>>>>>>>>>
>>>>>>>>>>> We don't count CRC length in the statistics, but it should be
>>>>>>>>>>> accessible
>>>>>>>>>>> in the payload by the user.
>>>>>>>>>> Our drivers are behaving exactly as you say.
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> If so I missed why mbuf 'pkt_len' and 'data_len' reduced by
>>>>>>>>> 'rxq->crc_len', can you please explain what above lines does?
>>>>>>>>>
>>>>>>>>>
>>>>>>>> @@ -2470,8 +2523,7 @@ hns3_recv_pkts_simple(void *rx_queue,
>>>>>>>>           rxdp->rx.bd_base_info = 0;
>>>>>>>>
>>>>>>>>           rxm->data_off = RTE_PKTMBUF_HEADROOM;
>>>>>>>> -        rxm->pkt_len = (uint16_t)(rte_le_to_cpu_16(rxd.rx.pkt_len)) -
>>>>>>>> -                rxq->crc_len;
>>>>>>>> +        rxm->pkt_len = rte_le_to_cpu_16(rxd.rx.pkt_len);
>>>>>>>>
>>>>>>>> In the previous code above, the 'pkt_len' is set to the length obtained
>>>>>>>> from the BD. the length obtained from the BD already contains CRC length.
>>>>>>>> But as you said above, the DPDK requires that the length of the mbuf
>>>>>>>> does not contain CRC length . So we subtract 'rxq->crc_len' from
>>>>>>>> mbuf'pkt_len' and 'data_len'. This patch doesn't change the logic, it
>>>>>>>> just moves the code around.
>>>>>>>>
>>>>>>>
>>>>>>> Nope, I am not saying mbuf length shouldn't contain CRC length, indeed
>>>>>>> it is other way around and this is our confusion.
>>>>>>>
>>>>>>> CRC length shouldn't be in the statistics, I mean in received bytes stats.
>>>>>>> Assume that received packet is 128 bytes and we know it has the CRC,
>>>>>>> Rx received bytes stat should be 124 (rx_bytes = 128 - CRC = 124)
>>>>>>>
>>>>>>> But mbuf->data_len & mbuf->pkt_len should have full frame length,
>>>>>>> including CRC.
>>>>>>>
>>>>>>> As application explicitly requested to KEEP CRC, it will know last 4
>>>>>>> bytes are CRC.
>>>>>>> Anything after 'mbuf->data_len' in the mbuf buffer is not valid, so if
>>>>>>> you reduce 'mbuf->data_len' by CRC size, application can't know if 4
>>>>>>> bytes after 'mbuf->data_len' is valid CRC or not.
>>>>>>>
>>>>>> I agree with you.
>>>>>>
>>>>>> But the implementation of other PMDs supported KEEP_CRC is like this.
>>>>>> In addition, there are probably many users that are already using it.
>>>>>> If we modify it, it may cause applications incompatible.
>>>>>>
>>>>>> what do you think?
>>>>>>
>>>>> This is documented in the ethdev [1], better to follow the documentation
>>>>> for all PMDs, can you please highlight the relevant driver code, we can
>>>>> discuss it with their maintainers.
>>>>>
>>>>> Alternatively we can document this additionally in the KEEP_CRC feature
>>>>> document if it helps for the applications.
>>>>>
>>>>>
>>>>> [1]
>>>>> https://git.dpdk.org/dpdk/tree/lib/ethdev/rte_ethdev.h?h=v23.11#n257
>>>>
>>>> Currently,this documentation does not describe whether pkt_len and data_len should contain crc_len.
>>>>
>>>
>>> I think it is clear that pkt_len and data_len should contain crc_len, we
>>> can ask for more comments.
>> This patch doesn't change the logic for hns3 PMD and the implementation of
>> other PMDs supported KEEP_CRC is like hns3 PMD. Can we merge this patch first?
>>
> 
> If hns3 behaving against the documented behavior, I don't understand why
> you are pushing for merging this patch, instead of fixing it.
> 
> 
> Other drivers behavior is something else, not directly related to this
> patch, but again if you can provide references we can discuss with their
> maintainers.
> 
> 
>>>
>>>> Do you mean that we add this description in the KEEP_CRC feature document
>>>> and notify all drivers that support KEEP_CRC to follow this documentation?
>>>>
>>>> If so, can you merge this patch first?
>>>> Then we send a RFC to disscuss it with all PMDs maintainer.
>>>>
>>>
>>> Not for drivers, just a suggestion that if we should update feature
>>> documentation with above information for users. So there is no
>>> dependency to features document update.
>>>
>>>
>> Sorry I'm more confused. What should we do next?
> 
> There is already API documentation about KEEP_CRC, I think that is
> already sufficient for driver developers.
> 
> I am just brainstorming if updating './doc/guides/nics/features.rst' can
> help end user, but it is not an action or blocker for this patch.
> 
> Next step is to update this path.
> 


Hi, Ferruh,

Thanks for your attention.
I have the following suggestions for the email discussion.
Please take the time to see if they make sense.

For pkt_len and data_len, there is no clear document indicating
whether the length should include the CRC.
However, according to the usage of the driver and the APP, it
is obvious that almost all drivers do not include the CRC by default.

The issue you raised about pkt_len/data_len supposedly containing CRC
and users not being able to get CRC has been around for a long
time, at least dating back to DPDK 18.11 when there was no hns3
driver. And there is no clear solution to this problem for now.

This issue is not caused by the current patch and is not related
to the problem to be solved by the current patch.
Therefore, it is recommended that the problem corresponding to the current
patch should be fixed first.

The problem that the pkt_len/data_len should contain the CRC
and the user should access the CRC can be discussed later.

I have the following two options:

1. Modify the corresponding document to specify that pkt_len/data_len
should contain CRC, and modify all related drivers. This requires the
participation and discussion of other driver developers.

2. Users can use rte_pktmbuf_dump() to print packet data.
The length of the packet data to be printed can be specified.
However, the API restricts that the length of the data required is
less than data_len. Therefore, users cannot obtain the CRC value.
However, the result of the rte_pktmbuf_tailroom() API can tell how
many bytes after data_len are accessible. Users can use rte_pktmbuf_tailroom
and keep_crc offload to obtain the CRC value of packets.

> .

  reply	other threads:[~2024-03-08 11:36 UTC|newest]

Thread overview: 34+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2024-02-06  1:10 Jie Hai
2024-02-07 14:15 ` Ferruh Yigit
2024-02-20  3:58   ` Jie Hai
2024-02-23 13:53     ` Ferruh Yigit
2024-02-26  3:16       ` Jie Hai
2024-02-26 16:43         ` Ferruh Yigit
2024-02-28  2:27           ` huangdengdui
2024-02-28 13:07             ` Ferruh Yigit
2024-02-29  3:58               ` huangdengdui
2024-02-29  9:25                 ` Ferruh Yigit
2024-03-01  6:55                   ` huangdengdui
2024-03-01 11:10                     ` Ferruh Yigit
2024-03-08 11:36                       ` Jie Hai [this message]
2024-03-22  6:28                         ` Jie Hai
2024-06-03  1:38                       ` Jie Hai
2024-06-03  2:33                         ` Stephen Hemminger
2024-06-03  5:24                           ` Morten Brørup
2024-06-03  7:07                           ` Andrew Rybchenko
2024-07-18 11:48 ` [PATCH v2 0/3] bugfix about KEEP CRC offload Jie Hai
2024-07-18 11:48   ` [PATCH v2 1/3] ethdev: add description for " Jie Hai
2024-07-18 11:57     ` Morten Brørup
2024-07-18 11:48   ` [PATCH v2 2/3] net/hns3: fix packet length do not contain CRC data length Jie Hai
2024-07-18 11:48   ` [PATCH v2 3/3] net/hns3: fix Rx packet without CRC data Jie Hai
2024-07-18 12:35   ` [PATCH v2 0/3] bugfix about KEEP CRC offload lihuisong (C)
2024-07-19  9:04 ` [PATCH v3 " Jie Hai
2024-07-19  9:04   ` [PATCH v3 1/3] ethdev: add description for " Jie Hai
2024-09-05  6:33     ` Andrew Rybchenko
2024-07-19  9:04   ` [PATCH v3 2/3] net/hns3: fix packet length do not contain CRC data length Jie Hai
2024-07-19  9:04   ` [PATCH v3 3/3] net/hns3: fix Rx packet without CRC data Jie Hai
2024-07-19  9:49   ` [PATCH v3 0/3] bugfix about KEEP CRC offload fengchengwen
2024-08-09  9:21   ` Jie Hai
2024-09-05  2:53   ` Jie Hai
2024-10-18  1:39   ` Jie Hai
2024-11-06  2:19   ` Jie Hai

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=eea7caf7-d4b9-c146-9f7a-186baa763e38@huawei.com \
    --to=haijie1@huawei.com \
    --cc=dev@dpdk.org \
    --cc=fengchengwen@huawei.com \
    --cc=ferruh.yigit@amd.com \
    --cc=huangdengdui@huawei.com \
    --cc=lihuisong@huawei.com \
    --cc=liuyonglong@huawei.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).