[dpdk-dev] [PATCH] ethdev: add flow tag

[Adrien Mazarguil]

On Fri, Jul 05, 2019 at 06:05:50PM +0000, Yongseok Koh wrote:
> > On Jul 5, 2019, at 6:54 AM, Adrien Mazarguil <adrien.mazarguil at 6wind.com> wrote:
> > 
> > On Thu, Jul 04, 2019 at 04:23:02PM -0700, Yongseok Koh wrote:
> >> A tag is a transient data which can be used during flow match. This can be
> >> used to store match result from a previous table so that the same pattern
> >> need not be matched again on the next table. Even if outer header is
> >> decapsulated on the previous match, the match result can be kept.
> >> 
> >> Some device expose internal registers of its flow processing pipeline and
> >> those registers are quite useful for stateful connection tracking as it
> >> keeps status of flow matching. Multiple tags are supported by specifying
> >> index.
> >> 
> >> Example testpmd commands are:
> >> 
> >>  flow create 0 ingress pattern ... / end
> >>    actions set_tag index 2 value 0xaa00bb mask 0xffff00ff /
> >>            set_tag index 3 value 0x123456 mask 0xffffff /
> >>            vxlan_decap / jump group 1 / end
> >> 
> >>  flow create 0 ingress pattern ... / end
> >>    actions set_tag index 2 value 0xcc00 mask 0xff00 /
> >>            set_tag index 3 value 0x123456 mask 0xffffff /
> >>            vxlan_decap / jump group 1 / end
> >> 
> >>  flow create 0 ingress group 1
> >>    pattern tag index is 2 value spec 0xaa00bb value mask 0xffff00ff /
> >>            eth ... / end
> >>    actions ... jump group 2 / end
> >> 
> >>  flow create 0 ingress group 1
> >>    pattern tag index is 2 value spec 0xcc00 value mask 0xff00 /
> >>            tag index is 3 value spec 0x123456 value mask 0xffffff /
> >>            eth ... / end
> >>    actions ... / end
> >> 
> >>  flow create 0 ingress group 2
> >>    pattern tag index is 3 value spec 0x123456 value mask 0xffffff /
> >>            eth ... / end
> >>    actions ... / end
> >> 
> >> Signed-off-by: Yongseok Koh <yskoh at mellanox.com>
> > 
> > Hi Yongseok,
> > 
> > Only high level questions for now, while it unquestionably looks useful,
> > from a user standpoint exposing the separate index seems redundant and not
> > necessarily convenient. Using the following example to illustrate:
> > 
> > actions set_tag index 3 value 0x123456 mask 0xfffff
> > 
> > pattern tag index is 3 value spec 0x123456 value mask 0xffffff
> > 
> > I might be missing something, but why isn't this enough:
> > 
> > pattern tag index is 3 # match whatever is stored at index 3
> > 
> > Assuming it can work, then why bother with providing value spec/mask on
> > set_tag? A flow rule pattern matches something, sets some arbitrary tag to
> > be matched by a subsequent flow rule and that's it. It even seems like
> > relying on the index only on both occasions is enough for identification.
> > 
> > Same question for the opposite approach; relying on the value, never
> > mentioning the index.
> > 
> > I'm under the impression that the index is a hardware-specific constraint
> > that shouldn't be exposed (especially since it's an 8-bit field). If so, a
> > PMD could keep track of used indices without having them exposed through the
> > public API.
> 
> 
> Thank you for review, Adrien.
> Hope you are doing well. It's been long since we talked each other. :-)

Yeah clearly! Hope you're doing well too. I'm somewhat busy hence slow to
answer these days...

 <dev at dpdk.org> hey!
 <dev at dpdk.org> no private talks!

Back to the topic:

> Your approach will work too in general but we have a request from customer that
> they want to partition this limited tag storage. Assuming that HW exposes 32bit
> tags (those are 'registers' in HW pipeline in mlx5 HW). Then, customers want to
> store multiple data even in a 32-bit storage. For example, 16bit vlan tag, 8bit
> table id and 8bit flow id. As they want to split one 32bit storage, I thought it
> is better to provide mask when setting/matching the value. Even some customer
> wants to store multiple flags bit by bit like ol_flags. They do want to alter
> only partial bits.
> 
> And for the index, it is to reference an entry of tags array as HW can provide
> larger registers than 32-bit. For example, mlx5 HW would provide 4 of 32b
> storage which users can use for their own sake.
> 	tag[0], tag[1], tag[2], tag[3]

OK, looks like I missed the point then. I initially took it for a funky
alternative to RTE_FLOW_ITEM_TYPE_META & RTE_FLOW_ACTION_TYPE_SET_META
(ingress extended [1]) but while it could be used like that, it's more of a
way to temporarily store and retrieve a small amount of data, correct?

Out of curiosity, are these registers independent from META and other
items/actions in mlx5, otherwise what happens if they are combined?

Are there other uses for these registers? Say, referencing their contents
from other places in a flow rule so they don't have to be hard-coded?

Right now I'm still uncomfortable with such a feature in the public API
because compared to META [1], this approach looks very hardware-specific and
seemingly difficult to map on different HW architectures.

However, the main problem is that as described, its end purpose seems
redundant with META, which I think can cover the use cases you gave. So what
can an application do with this that couldn't be done in a more generic
fashion through META?

I may still be missing something and I'm open to ideas, but assuming it
doesn't make it into the public rte_flow API, it remains an interesting
feature on its own merit which could be added to DPDK as PMD-specific
pattern items/actions [2]. mlx5 doesn't have any yet, but it's pretty common
for PMDs to expose a public header that dedicated applications can include
to use this kind of features (look for rte_pmd_*.h, e.g. rte_pmd_ixgbe.h).
No problem with that.

[1] "[PATCH] ethdev: extend flow metadata"
    https://mails.dpdk.org/archives/dev/2019-July/137305.html

[2] "Negative types"
    https://doc.dpdk.org/guides/prog_guide/rte_flow.html#negative-types

-- 
Adrien Mazarguil
6WIND