From mboxrd@z Thu Jan 1 00:00:00 1970 From: Adrien Mazarguil Subject: Re: [PATCH 01/12] ethdev: add API to query what/if packet type is set Date: Wed, 6 Jan 2016 18:22:48 +0100 Message-ID: <20160106172248.GT12095@6wind.com> References: <1451544799-70776-1-git-send-email-jianfeng.tan@intel.com> <1451544799-70776-2-git-send-email-jianfeng.tan@intel.com> <20160104113814.GT3806@6wind.com> <2601191342CEEE43887BDE71AB97725836AE1002@irsmsx105.ger.corp.intel.com> <20160105161423.GE4712@autoinstall.dev.6wind.com> <2601191342CEEE43887BDE71AB97725836AE18E3@irsmsx105.ger.corp.intel.com> <20160106100053.GJ12095@6wind.com> <2601191342CEEE43887BDE71AB97725836AE1B46@irsmsx105.ger.corp.intel.com> <20160106154438.GP12095@6wind.com> <2601191342CEEE43887BDE71AB97725836AE2DBC@irsmsx105.ger.corp.intel.com> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: quoted-printable Cc: "dev@dpdk.org" To: "Ananyev, Konstantin" Return-path: Received: from mail-wm0-f53.google.com (mail-wm0-f53.google.com [74.125.82.53]) by dpdk.org (Postfix) with ESMTP id DB29095C2 for ; Wed, 6 Jan 2016 18:23:07 +0100 (CET) Received: by mail-wm0-f53.google.com with SMTP id b14so85561647wmb.1 for ; Wed, 06 Jan 2016 09:23:07 -0800 (PST) Content-Disposition: inline In-Reply-To: <2601191342CEEE43887BDE71AB97725836AE2DBC@irsmsx105.ger.corp.intel.com> List-Id: patches and discussions about DPDK List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org Sender: "dev" On Wed, Jan 06, 2016 at 04:44:43PM +0000, Ananyev, Konstantin wrote: >=20 >=20 > > -----Original Message----- > > From: Adrien Mazarguil [mailto:adrien.mazarguil@6wind.com] > > Sent: Wednesday, January 06, 2016 3:45 PM > > To: Ananyev, Konstantin > > Cc: N=C3=A9lio Laranjeiro; Tan, Jianfeng; dev@dpdk.org > > Subject: Re: [dpdk-dev] [PATCH 01/12] ethdev: add API to query what/i= f packet type is set > >=20 > > On Wed, Jan 06, 2016 at 02:29:07PM +0000, Ananyev, Konstantin wrote: > > > > > > > > > > -----Original Message----- > > > > From: Adrien Mazarguil [mailto:adrien.mazarguil@6wind.com] > > > > Sent: Wednesday, January 06, 2016 10:01 AM > > > > To: Ananyev, Konstantin > > > > Cc: N=C3=A9lio Laranjeiro; Tan, Jianfeng; dev@dpdk.org > > > > Subject: Re: [dpdk-dev] [PATCH 01/12] ethdev: add API to query wh= at/if packet type is set > > > > > > > > On Tue, Jan 05, 2016 at 04:50:31PM +0000, Ananyev, Konstantin wro= te: > > > > [...] > > > > > > -----Original Message----- > > > > > > From: N=C3=A9lio Laranjeiro [mailto:nelio.laranjeiro@6wind.co= m] > > > > [...] > > > > > > I think we miss a comment here in how those 2/6/4 values are = chosen > > > > > > because, according to the mask, I expect 16 possibilities but= I get > > > > > > less. It will help a lot anyone who needs to add a new type. > > > > > > > > > > > > Extending the snprintf behavior above, it is best to remove t= he mask > > > > > > argument altogether and have rte_eth_dev_get_ptype_info() ret= urn the > > > > > > entire list every time. Applications need to iterate on the = result in > > > > > > any case. > > > > > > > > > > I think we'd better keep mask argument. > > > > > In many cases upper layer only interested in some particular s= ubset of > > > > > all packet types that HW can recognise. > > > > > Let say l3fwd only cares about RTE_PTYPE_L3_MASK, it is not in= terested in L4, > > > > > tunnelling packet types, etc. > > > > > If caller needs to know all recognised ptypes, he can set mask=3D= =3D-1, > > > > > In that case all supported packet types will be returned. > > > > > > > > There are other drawbacks to the mask argument in my opinion. The= API will > > > > have to be updated again as soon as 32 bits aren't enough to repr= esent all > > > > possible masks. We can't predict it will be large enough forever = but on the > > > > other hand, using uint64_t seems overkill at this point. > > > > > > Inside rte_mbuf packet_type itself is a 32 bit value. > > > These 32 bits are divided into several fields to mark packet types, > > > i.e: bits [0-3] are for all possible L2 types, bits [4-7] for L3 ty= pes, etc. > > > As long as packet_type itself is 32bits, 32bit mask is sufficient. > > > If we'll ever run out of 32 bits in packet_type itself, it will be = ABI change anyway. > >=20 > > Sure, however why not do it now this issue has been raised so this fu= nction > > doesn't need updating the day it breaks? I know there's a million oth= er > > places with a similar problem but I'm all for making new code future = proof. >=20 > If rte_mbuf packet_type will have to be increased to 64bit long, then > this function will have to change anyway (with or without mask paramete= r). > It will have to become: >=20 > rte_eth_dev_get_ptype_info(uint8_t portid, uint64_t ptypes[], ...) >=20 > So I think we don't have to worry about mask parameter itself. Well, yes, besides I overlooked ptypes[] itself is 32 bit, working around the type width of the mask wouldn't help much. > > Perhaps in this particular case there is no way to hit the limit (alt= hough > > there are only four unused bits left to extend RTE_PTYPE masks) but w= e've > > seen this happen too many times with subsequent ABI breakage. >=20 > When ptype was introduced we tried to reserve some free space for each = layer (L2/L3/L4/...), > so it wouldn't be overrun immediately. > But of course if there would be a new HW that can recognise dozen new p= acket types - it is possible. > Do you have any particular use-case in mind?=20 No, that was just to illustrate my point. > > > > I think this use for masks should be avoided when performance doe= s not > > > > matter much, as in this case, user application cannot know the nu= mber of > > > > entries in advance and must rely on the returned value to iterate= . > > > > > > User doesn't know numbers of entries in advance anyway (with and wi= thout the mask). > > > That's why this function was introduced at first place. > > > > > > With mask it just a bit more handy, in case user cares only about p= articular subset of supported > > > packet types (only L2 let say). > >=20 > > OK, so we definitely need something to let applications know the laye= r a > > given packet type belongs to, I'm sure it can be done in a convenient= way > > that won't be limited to the underlying type of the mask. > >=20 > > > > A helper function can be added to convert a RTE_PTYPE_* value to = the layer > > > > it belongs to (using enum to define possible values). > > > > > > Not sure what for? > >=20 > > This is assuming rte_eth_dev_get_ptype_info() doesn't filter anything= (no > > "mask" argument). In that case a separate function must be added to c= onvert > > RTE_PTYPE_* values to a layer, so applications can look for interesti= ng > > packet types while parsing plist[] on their own. >=20 > Honestly, I don't see why do you need that. > You already do know that let say RTE_PTYPE_L3_IPV4 belongs to L3. > Why do you need some extra enum here? > From my thought - the only purpose of mask parameter was to limit numbe= r of elements in the ptypes[] at return. > So let say user would need to iterate over 10 elements, instead of 100 = to find > the ones he is interested in. Since this is already a slow manner for retrieving types, 10 or 100 doesn= 't make much difference. Such a function shouldn't be used in the data path directly. My point is, since we're dealing with a slow function, let's keep its API= as simple as possible. By having a mask to match, a large number of checks a= re added in all PMDs while they could just fill the array without bothering. The filtering logic is an application requirement that could b= e useful in its own function as well (converting any random value to its related layer or mask). > > This layer information could be defined as an enum, i.e.: > >=20 > > enum rte_ptype_info { > > RTE_PTYPE_UNKNOWN, > > RTE_PTYPE_L2, > > RTE_PTYPE_L3, > > ... > > }; > >=20 > > Or even an int value (2 standing for for "layer 2" etc. Tunnel encaps= ulation > > wouldn't be described easily that way though). It's just an idea. > >=20 > > > > If we absolutely want a mean to filter returned values, I suggest= we use > > > > this enum instead of the mask argument. > > > > Since it won't be a mask, it won't > > > > have to be updated every time a new protocol requires extending o= ne. > > > > > > Number of bits PTYPE_L2/L3/L4,... layers are already defined. > > > So let say RTE_PTYPE_L2_MASK shouldn't change if you'll add new L2 = ptype - > > > there are few reserved values right now. > > > if one day we'll run out bits in let say RTE_PTYPE_L2_MASK and wil= l have to increase its size - > > > it would mean change of the packet_type layout and possible ABI bre= akage anyway. > >=20 > > I'm aware of this, only pointing out we tend to rely on masks and typ= e > > boundaries a bit too much when there are other methods that are as (i= f not > > more) convenient. >=20 > Yes, we do rely on masks in ptype. > That's how ptype was defined. > Let say to check that incoming packet is Ether/IPv4(no extentions)/UDP, > you probably would do: >=20 > if (mbuf->packet_type & (RTE_PTYPE_L2_MASK | RTE_PTYPE_L3_MASK | RTE_PT= YPE_L4_MASK) =3D=3D > (RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_UDP)) {...} All right, let's not use a different method to filter packet types. > > Perhaps some sort of tunneled packet types beyond inner L4 consuming = the > > four remaining bits will be added? That could happen soon. >=20 > As I said above: do you have particular scenario in mind when 32bits fo= r packet_type > might be not enough? > If yes, then it is probably a good idea to submit an RFC for extending = it to 64 bit, > or introduce packet_type2, or whatever would be your preferred way to d= eal with it. No, really, I guess we'll extend ptype to 64 bit when necessary. My point= on filtering separately still stands. > Konstantin >=20 --=20 Adrien Mazarguil 6WIND