DPDK-dev Archive on lore.kernel.org
 help / color / Atom feed
From: Andrew Rybchenko <arybchenko@solarflare.com>
To: Shahaf Shuler <shahafs@mellanox.com>,
	"pbhagavatula@marvell.com" <pbhagavatula@marvell.com>,
	"jerinj@marvell.com" <jerinj@marvell.com>,
	"ferruh.yigit@intel.com" <ferruh.yigit@intel.com>,
	John McNamara <john.mcnamara@intel.com>,
	Marko Kovacevic <marko.kovacevic@intel.com>,
	Thomas Monjalon <thomas@monjalon.net>
Cc: "dev@dpdk.org" <dev@dpdk.org>
Subject: Re: [dpdk-dev] [PATCH 2/7] ethdev: add mbuf RSS update as a offload
Date: Sun, 18 Aug 2019 10:00:36 +0300
Message-ID: <1a33f43e-6449-dffe-a6c9-c3d5c7118dae@solarflare.com> (raw)
In-Reply-To: <AM0PR0502MB379554AB60F377D7A284F3E5C3A90@AM0PR0502MB3795.eurprd05.prod.outlook.com>

On 8/18/19 9:18 AM, Shahaf Shuler wrote:
> Sunday, August 18, 2019 8:39 AM, Andrew Rybchenko:
>> <marko.kovacevic@intel.com>; Thomas Monjalon <thomas@monjalon.net>
>> Cc: dev@dpdk.org
>> Subject: Re: [dpdk-dev] [PATCH 2/7] ethdev: add mbuf RSS update as a
>> offload
>>
>> On 8/18/19 7:52 AM, Shahaf Shuler wrote:
>>> Friday, August 16, 2019 10:48 AM, Andrew Rybchenko:
>>>> Subject: Re: [dpdk-dev] [PATCH 2/7] ethdev: add mbuf RSS update as a
>>>> offload
>>>>
>>>> On 8/16/19 8:55 AM, pbhagavatula@marvell.com wrote:
>>>>> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
>>>>>
>>>>> Add new Rx offload flag `DEV_RX_OFFLOAD_RSS_HASH` which can be
>> used
>>>> to
>>>>> enable/disable PMDs write to `rte_mbuf::hash::rss`.
>>>> It should be highlighted that presence of the RSS hash is indicated
>>>> by PKT_RX_RSS_HASH flag in mbuf anyway. Now applications have a way
>>>> to check that RSS hash delivery is supported and should enable the
>>>> offload if RSS hash is used. PMD may still provide the hash even if
>>>> the offload is not enabled.
>>> I don't understand how PMDs should act w/ this addition when considering
>> the API breakage to application.
>>
>> There is a deprecation notice for it.
>> I mentioned in my review notes for one of patches in the series that the
>> change should be highlighted in release notes.
>> Yes, it is absolutely required if these patches are accepted.
>>
>>> Currently application don't set this flag, and expect to get the RSS hash
>> result on mbuf.
>>> If PMDs will not set the RSS hash result when flag is not present then
>> applications might break.
>>> If they will always set, then there is no meaning for it.
>>>
>>> as I understand the motivation to save few cycles on the PMD receive path,
>> if we want to include it we should treat it as API breakage and documents it
>> on the release notes.
>>> My option is that some offload should just be usable (OOB) by the fact user
>> enabled them (e.g. RSS). no need to complicate the user by checking and set
>> this field.
>>
>> What I don't understand is why some offloads should just work but another
>> requires action to enable it. Just because it is the current state of things - I
>> don't think it is a good motivation. Sorry.
> Not because it is the current state of things, because it makes user experience much simpler.

If so, it would be simpler to have no controls at all and always have
everything possible.

> You enabled RSS -> you get full RSS behavior

You enable distribution across many queues here.
RSS hash availability is a side effect here.

> You set a flow rule w/ mark -> you get full flow mark behavior
> You set checksum -> you get full csum behavior.
>
>> I think more applications use checksum offloads than RSS hash, but it is still
>> required to enable it. It looks like no single DPDK example uses RSS hash. So,
>> I guess it not widely used by applications as well.
> Well there is at least one called ovs-dpdk, that use the RSS result as the key to access the EMC.
> I know of few more, not upstream, ones.
>
>> Anyway these 2 patches for flow action and RSS hash make all Rx offloads
>> consistent - if you need something, enable it.
> But the user enabled it -
> It enabled RSS by setting ETH_MQ_RX_RSS, why does it need to enable another flag?

Answered above. If you need distribution it does not mean
that you need RSS hash information. There are really many
examples when you don't really need it.

> Same for flow mark.
>
>> And the question is not to save few cycles in the PMD receive path.
>> It makes is possible to not deliver both from NIC to host.
>> 8 bytes (4 RSS hash and 4 flow mark) are more than 10% for the smallest
>> packets.
> There is always the line between how much tight control we want to provide to user (to save cycles/ to save PCI BW) and how much it will be simple for the user to work on top.
> My opinion is that we need to have some basics.

Many thanks, your arguments make sense. I vote for
consistency and more fine grained control which allows
more optimizations and allow to squeeze more performance
from HW and SW. So, my line is a bit lower.  I don't think
that these two patches make user control over-complicated.

>>>>> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
>>>> Reviewed-by: Andrew Rybchenko <arybchenko@solarflare.com>
>>>>
>>>> with above and one note below fixed.
>>>>
>>>>> ---
>>>>>     doc/guides/nics/features.rst   | 2 ++
>>>>>     lib/librte_ethdev/rte_ethdev.h | 1 +
>>>>>     2 files changed, 3 insertions(+)
>>>>>
>>>>> diff --git a/doc/guides/nics/features.rst
>>>>> b/doc/guides/nics/features.rst index d4d55f721..f79b69b38 100644
>>>>> --- a/doc/guides/nics/features.rst
>>>>> +++ b/doc/guides/nics/features.rst
>>>>> @@ -274,6 +274,7 @@ Supports RSS hashing on RX.
>>>>>
>>>>>     * **[uses]     user config**: ``dev_conf.rxmode.mq_mode`` =
>>>> ``ETH_MQ_RX_RSS_FLAG``.
>>>>>     * **[uses]     user config**: ``dev_conf.rx_adv_conf.rss_conf``.
>>>>> +* **[uses]     rte_eth_rxconf,rte_eth_rxmode**:
>>>> ``offloads:DEV_RX_OFFLOAD_RSS_HASH``.
>>>>>     * **[provides] rte_eth_dev_info**: ``flow_type_rss_offloads``.
>>>>>     * **[provides] mbuf**: ``mbuf.ol_flags:PKT_RX_RSS_HASH``,
>> ``mbuf.rss``.
>>>>> @@ -286,6 +287,7 @@ Inner RSS
>>>>>     Supports RX RSS hashing on Inner headers.
>>>>>
>>>>>     * **[uses]    rte_flow_action_rss**: ``level``.
>>>>> +* **[uses]    rte_eth_rxconf,rte_eth_rxmode**:
>>>> ``offloads:DEV_RX_OFFLOAD_RSS_HASH``.
>>>>>     * **[provides] mbuf**: ``mbuf.ol_flags:PKT_RX_RSS_HASH``,
>> ``mbuf.rss``.
>>>>>
>>>>> diff --git a/lib/librte_ethdev/rte_ethdev.h
>>>>> b/lib/librte_ethdev/rte_ethdev.h index f97f0a6e5..889486a11 100644
>>>>> --- a/lib/librte_ethdev/rte_ethdev.h
>>>>> +++ b/lib/librte_ethdev/rte_ethdev.h
>>>>> @@ -1013,6 +1013,7 @@ struct rte_eth_conf {
>>>>>     #define DEV_RX_OFFLOAD_KEEP_CRC		0x00010000
>>>>>     #define DEV_RX_OFFLOAD_SCTP_CKSUM	0x00020000
>>>>>     #define DEV_RX_OFFLOAD_OUTER_UDP_CKSUM  0x00040000
>>>>> +#define DEV_RX_OFFLOAD_RSS_HASH		0x00080000
>>>> Should be added to rte_rx_offload_names in
>>>> lib/librte_ethdev/rte_ethdev.c.


  reply index

Thread overview: 31+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2019-08-16  5:55 [dpdk-dev] [PATCH 0/7] ethdev: add new Rx offload flags pbhagavatula
2019-08-16  5:55 ` [dpdk-dev] [PATCH 1/7] ethdev: add set ptype function pbhagavatula
2019-08-16  8:22   ` Andrew Rybchenko
2019-08-17 16:27     ` [dpdk-dev] [EXT] " Pavan Nikhilesh Bhagavatula
2019-08-16  5:55 ` [dpdk-dev] [PATCH 2/7] ethdev: add mbuf RSS update as a offload pbhagavatula
2019-08-16  7:48   ` Andrew Rybchenko
2019-08-17 11:47     ` Pavan Nikhilesh Bhagavatula
2019-08-18  4:52     ` Shahaf Shuler
2019-08-18  5:38       ` Andrew Rybchenko
2019-08-18  6:18         ` Shahaf Shuler
2019-08-18  7:00           ` Andrew Rybchenko [this message]
2019-08-18 12:11             ` Shahaf Shuler
2019-08-16  5:55 ` [dpdk-dev] [PATCH 3/7] ethdev: add flow action type update as an offload pbhagavatula
2019-08-16  8:05   ` Andrew Rybchenko
2019-08-17 14:23     ` [dpdk-dev] [EXT] " Pavan Nikhilesh Bhagavatula
2019-08-18  9:46       ` Andrew Rybchenko
2019-08-18  4:59     ` [dpdk-dev] " Shahaf Shuler
2019-08-18  5:57       ` Andrew Rybchenko
2019-08-18  6:20         ` Shahaf Shuler
2019-08-18  7:08           ` Andrew Rybchenko
2019-08-16  5:55 ` [dpdk-dev] [PATCH 4/7] net: update Rx RSS hash offload capabilities pbhagavatula
2019-08-16  7:49   ` Andrew Rybchenko
2019-08-17 12:26     ` Pavan Nikhilesh Bhagavatula
2019-08-16  5:55 ` [dpdk-dev] [PATCH 5/7] net: update Rx flow action " pbhagavatula
2019-08-16  7:50   ` Andrew Rybchenko
2019-08-16  5:55 ` [dpdk-dev] [PATCH 6/7] net: add ptype set default functionality pbhagavatula
2019-08-16  8:30   ` Andrew Rybchenko
2019-08-17 17:24     ` Pavan Nikhilesh Bhagavatula
2019-08-16  5:55 ` [dpdk-dev] [PATCH 7/7] examples/eventdev_pipeline: add new Rx RSS hash offload pbhagavatula
2019-08-16  6:01   ` Jerin Jacob Kollanukkaran
2019-08-16  6:02 ` [dpdk-dev] [PATCH 0/7] ethdev: add new Rx offload flags Jerin Jacob Kollanukkaran

Reply instructions:

You may reply publically 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=1a33f43e-6449-dffe-a6c9-c3d5c7118dae@solarflare.com \
    --to=arybchenko@solarflare.com \
    --cc=dev@dpdk.org \
    --cc=ferruh.yigit@intel.com \
    --cc=jerinj@marvell.com \
    --cc=john.mcnamara@intel.com \
    --cc=marko.kovacevic@intel.com \
    --cc=pbhagavatula@marvell.com \
    --cc=shahafs@mellanox.com \
    --cc=thomas@monjalon.net \
    /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

DPDK-dev Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/dpdk-dev/0 dpdk-dev/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 dpdk-dev dpdk-dev/ https://lore.kernel.org/dpdk-dev \
		dev@dpdk.org dpdk-dev@archiver.kernel.org
	public-inbox-index dpdk-dev


Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.dpdk.dev


AGPL code for this site: git clone https://public-inbox.org/ public-inbox