[dpdk-dev] [PATCH] ethdev: extend flow metadata

[Yigit, Ferruh]

On 7/29/2019 4:06 PM, Adrien Mazarguil wrote:
> On Sun, Jul 14, 2019 at 02:46:58PM +0300, Andrew Rybchenko wrote:
>> On 11.07.2019 10:44, Adrien Mazarguil wrote:
>>> On Wed, Jul 10, 2019 at 04:37:46PM +0000, Yongseok Koh wrote:
>>>>> On Jul 10, 2019, at 5:26 AM, Thomas Monjalon <thomas at monjalon.net> wrote:
>>>>>
>>>>> 10/07/2019 14:01, Bruce Richardson:
>>>>>> On Wed, Jul 10, 2019 at 12:07:43PM +0200, Olivier Matz wrote:
>>>>>>> On Wed, Jul 10, 2019 at 10:55:34AM +0100, Bruce Richardson wrote:
>>>>>>>> On Wed, Jul 10, 2019 at 11:31:56AM +0200, Olivier Matz wrote:
>>>>>>>>> On Thu, Jul 04, 2019 at 04:21:22PM -0700, Yongseok Koh wrote:
>>>>>>>>>> Currently, metadata can be set on egress path via mbuf tx_meatadata field
>>>>>>>>>> with PKT_TX_METADATA flag and RTE_FLOW_ITEM_TYPE_RX_META matches metadata.
>>>>>>>>>>
>>>>>>>>>> This patch extends the usability.
>>>>>>>>>>
>>>>>>>>>> 1) RTE_FLOW_ACTION_TYPE_SET_META
>>>>>>>>>>
>>>>>>>>>> When supporting multiple tables, Tx metadata can also be set by a rule and
>>>>>>>>>> matched by another rule. This new action allows metadata to be set as a
>>>>>>>>>> result of flow match.
>>>>>>>>>>
>>>>>>>>>> 2) Metadata on ingress
>>>>>>>>>>
>>>>>>>>>> There's also need to support metadata on packet Rx. Metadata can be set by
>>>>>>>>>> SET_META action and matched by META item like Tx. The final value set by
>>>>>>>>>> the action will be delivered to application via mbuf metadata field with
>>>>>>>>>> PKT_RX_METADATA ol_flag.
>>>>>>>>>>
>>>>>>>>>> For this purpose, mbuf->tx_metadata is moved as a separate new field and
>>>>>>>>>> renamed to 'metadata' to support both Rx and Tx metadata.
>>>>>>>>>>
>>>>>>>>>> For loopback/hairpin packet, metadata set on Rx/Tx may or may not be
>>>>>>>>>> propagated to the other path depending on HW capability.
>>>>>>>>>>
>>>>>>>>>> Signed-off-by: Yongseok Koh <yskoh at mellanox.com>
>>>>>>>>>> --- a/lib/librte_mbuf/rte_mbuf.h
>>>>>>>>>> +++ b/lib/librte_mbuf/rte_mbuf.h
>>>>>>>>>> @@ -648,17 +653,6 @@ struct rte_mbuf {
>>>>>>>>>> 			/**< User defined tags. See rte_distributor_process() */
>>>>>>>>>> 			uint32_t usr;
>>>>>>>>>> 		} hash;                   /**< hash information */
>>>>>>>>>> -		struct {
>>>>>>>>>> -			/**
>>>>>>>>>> -			 * Application specific metadata value
>>>>>>>>>> -			 * for egress flow rule match.
>>>>>>>>>> -			 * Valid if PKT_TX_METADATA is set.
>>>>>>>>>> -			 * Located here to allow conjunct use
>>>>>>>>>> -			 * with hash.sched.hi.
>>>>>>>>>> -			 */
>>>>>>>>>> -			uint32_t tx_metadata;
>>>>>>>>>> -			uint32_t reserved;
>>>>>>>>>> -		};
>>>>>>>>>> 	};
>>>>>>>>>>
>>>>>>>>>> 	/** Outer VLAN TCI (CPU order), valid if PKT_RX_QINQ is set. */
>>>>>>>>>> @@ -727,6 +721,11 @@ struct rte_mbuf {
>>>>>>>>>> 	 */
>>>>>>>>>> 	struct rte_mbuf_ext_shared_info *shinfo;
>>>>>>>>>>
>>>>>>>>>> +	/** Application specific metadata value for flow rule match.
>>>>>>>>>> +	 * Valid if PKT_RX_METADATA or PKT_TX_METADATA is set.
>>>>>>>>>> +	 */
>>>>>>>>>> +	uint32_t metadata;
>>>>>>>>>> +
>>>>>>>>>> } __rte_cache_aligned;
>>>>>>>>> This will break the ABI, so we cannot put it in 19.08, and we need a
>>>>>>>>> deprecation notice.
>>>>>>>>>
>>>>>>>> Does it actually break the ABI? Adding a new field to the mbuf should only
>>>>>>>> break the ABI if it either causes new fields to move or changes the
>>>>>>>> structure size. Since this is at the end, it's not going to move any older
>>>>>>>> fields, and since everything is cache-aligned I don't think the structure
>>>>>>>> size changes either.
>>>>>>> I think it does break the ABI: in previous version, when the PKT_TX_METADATA
>>>>>>> flag is set, the associated value is put in m->tx_metadata (offset 44 on
>>>>>>> x86-64), and in the next version, it will be in m->metadata (offset 112). So,
>>>>>>> these 2 versions are not binary compatible.
>>>>>>>
>>>>>>> Anyway, at least it breaks the API.
>>>>>> Ok, I misunderstood. I thought it was the structure change itself you were
>>>>>> saying broke the ABI. Yes, putting the data in a different place is indeed
>>>>>> an ABI break.
>>>>> We could add the new field and keep the old one unused,
>>>>> so it does not break the ABI.
>>>> Still breaks ABI if PKT_TX_METADATA is set. :-) In order not to break it, I can
>>>> keep the current union'd field (tx_metadata) as is with PKT_TX_METADATA, add
>>>> the new one at the end and make it used with the new PKT_RX_METADATA.
>>>>
>>>>> However I suppose everybody will prefer a version using dynamic fields.
>>>>> Is someone against using dynamic field for such usage?
>>>> However, given that the amazing dynamic fields is coming soon (thanks for your
>>>> effort, Olivier and Thomas!), I'd be honored to be the first user of it.
>>>>
>>>> Olivier, I'll take a look at your RFC.
>>> Just got a crazy idea while reading this thread... How about repurposing
>>> that "reserved" field as "rx_metadata" in the meantime?
>>
>> It overlaps with hash.fdir.hi which has RSS hash.
> 
> While it does overlap with hash.fdir.hi, isn't the RSS hash stored in the
> "rss" field overlapping with hash.fdir.lo? (see struct rte_flow_action_rss)
> 
> hash.fdir.hi was originally used by FDIR and later repurposed by rte_flow
> for its MARK action, which neatly qualifies as Rx metadata so renaming
> "reserved" as "rx_metadata" could already make sense.
> 
> That is, assuming users do not need two different kinds of Rx metadata
> returned simultaneously with their packets. I think it's safe.
> 
>>> I know reserved fields are cursed and no one's ever supposed to touch them
>>> but this risk is mitigated by having the end user explicitly request its
>>> use, so the patch author (and his relatives) should be safe from the
>>> resulting bad juju.
>>>
>>> Joke aside, while I like the idea of Tx/Rx META, I think the similarities
>>> with MARK (and TAG eventually) is a problem. I wasn't available and couldn't
>>> comment when META was originally added to the Tx path, but there's a lot of
>>> overlap between these items/actions, without anything explaining to the end
>>> user how and why they should pick one over the other, if they can be
>>> combined at all and what happens in that case.
>>>
>>> All this must be documented, then we should think about unifying their
>>> respective features and deprecate the less capable items/actions. In my
>>> opinion, users need exactly one method to mark/match some mark while
>>> processing Rx/Tx traffic and *optionally* have that mark read from/written
>>> to the mbuf, which may or may not be possible depending on HW features.
> 
> Thoughts regarding this suggestion? From a user perspective I think all
> these actions should be unified but maybe there are good reasons to keep
> them separate?
> 

I think more recent plan is introducing dynamic fields for the remaining 16
bytes in the second cacheline.

I will update the patch as rejected, is there any objection?