When IBoE will be merged to upstream?

When IBoE will be merged to upstream?

[Aleksey Senin]

Hi, Roland.
This is actually a continue of the RAW_ET(XXXX)  issue. We want to
make a submition  of the patches to the upstream, but there is not
support for IB transport in Ethernet devices, and the mlx4_en drivers
version  is a bit outdated 1.4.1.1 in upstream and 1.5.1 in the OFED
There is also missing VLAN support that already present in the OFED.
When do you planning to submit changes from OFED to upstream?
Now, concerning new enum type RAW_ETH and exporting it to the kernel.
As you can see there is no special QP types defined in the verbs
userspace, so I think we can't broke userspace applications.This is on
the one hand. On the second hand, we need this new enum becase already
present RAW_ETY actually should be used when formatting non-IBA frames
over IB, and the new one will be used for formatting such pakets over
Ethernet. In additon there are changes in the OFED that using RAW_ETY
type and QP created for such type is different from the RAW_ETH QP.
The first one - RAW_ETY - cannot be created from userspace and there
is no such limitation for newer RAW_ETH type.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Roland Dreier]

 > This is actually a continue of the RAW_ET(XXXX)  issue. We want to
 > make a submition  of the patches to the upstream, but there is not
 > support for IB transport in Ethernet devices, and the mlx4_en drivers
 > version  is a bit outdated 1.4.1.1 in upstream and 1.5.1 in the OFED
 > There is also missing VLAN support that already present in the OFED.
 > When do you planning to submit changes from OFED to upstream?

 - I do not search for more things to merge upstream.  I have enough
   work reviewing things that are sent to me.  So I will never look
   through OFED for changes.

 - I do not handle the mlx4_en driver.  Changes for mlx4_en should go to
   netdev and Dave Miller.

 - I will try to get back to the IBoE changes when I have time, and I
   will admit that my time to spend as RDMA maintainer is nowhere near
   full time and less than it was in the past.

 - I did allocate a fair amount of time to spend on IBoE recently but
   unfortunately the patches were not really in a suitable state to
   merge, and I exhausted that time slice before we reached the end.
   When patch sets sit outside of the upstream kernel and are shipped in
   OFED for months and years, it would probably make upstream merging
   easier if that time was used to fix the patch set.

 - Specifically for the IBoE patches, shouldn't someone have realized
   that having a device-specific interface to do the standard mapping of
   GID to Ethernet address makes no sense?
-- 
Roland Dreier <rolandd-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

Regarding GID to Eth mappings, we discussed using the standard create_ah() Verb for this.
In the kernel, create_ah() will call a generic address resolution function in the cma.
The returned information will be copied back to user-space in a device-specific structure (since address handles are device-specific).

This eliminates adding a new get_eth_l2_addr() ABI call (device-specific or not).
In fact, this approach eliminates adding new ABIs for any kind of address translation...

Does this seem reasonable?
--Liran


-----Original Message-----
From: linux-rdma-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org [mailto:linux-rdma-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org] On Behalf Of Roland Dreier
Sent: Thursday, June 24, 2010 2:24 AM
To: Aleksey Senin
Cc: linux-rdma; monis-smomgflXvOZWk0Htik3J/w@public.gmane.org; alekseys-smomgflXvOZWk0Htik3J/w@public.gmane.org; yiftahs-smomgflXvOZWk0Htik3J/w@public.gmane.org; Tziporet Koren; alexr-smomgflXvOZWk0Htik3J/w@public.gmane.org
Subject: Re: When IBoE will be merged to upstream?

 > This is actually a continue of the RAW_ET(XXXX)  issue. We want to  > make a submition  of the patches to the upstream, but there is not  > support for IB transport in Ethernet devices, and the mlx4_en drivers  > version  is a bit outdated 1.4.1.1 in upstream and 1.5.1 in the OFED  > There is also missing VLAN support that already present in the OFED.
 > When do you planning to submit changes from OFED to upstream?

 - I do not search for more things to merge upstream.  I have enough
   work reviewing things that are sent to me.  So I will never look
   through OFED for changes.

 - I do not handle the mlx4_en driver.  Changes for mlx4_en should go to
   netdev and Dave Miller.

 - I will try to get back to the IBoE changes when I have time, and I
   will admit that my time to spend as RDMA maintainer is nowhere near
   full time and less than it was in the past.

 - I did allocate a fair amount of time to spend on IBoE recently but
   unfortunately the patches were not really in a suitable state to
   merge, and I exhausted that time slice before we reached the end.
   When patch sets sit outside of the upstream kernel and are shipped in
   OFED for months and years, it would probably make upstream merging
   easier if that time was used to fix the patch set.

 - Specifically for the IBoE patches, shouldn't someone have realized
   that having a device-specific interface to do the standard mapping of
   GID to Ethernet address makes no sense?
--
Roland Dreier <rolandd-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Hefty, Sean]

> Regarding GID to Eth mappings, we discussed using the standard create_ah()
> Verb for this.
> In the kernel, create_ah() will call a generic address resolution function
> in the cma.
> The returned information will be copied back to user-space in a device-
> specific structure (since address handles are device-specific).
> 
> This eliminates adding a new get_eth_l2_addr() ABI call (device-specific or
> not).
> In fact, this approach eliminates adding new ABIs for any kind of address
> translation...
> 
> Does this seem reasonable?

The current behavior of ibv_create_ah() requires that the caller provide the L2, and if needed, L3 addressing.  Any translation between the L3 and L2 addressing must be done before the call is made.  E.g. ibv_create_ah does not use the GID to query the SA to obtain LIDs.  Why doesn't IBoE follow this same model?

Callers can use some out of band mechanism for the mapping, call rdma_resolve_addr, or use some standard networking routine.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

S.B.
--Liran 

-----Original Message-----
From: Hefty, Sean [mailto:sean.hefty-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org] 
Sent: Thursday, June 24, 2010 9:06 PM
To: Liran Liss; Roland Dreier; Aleksey Senin
Cc: linux-rdma; monis-smomgflXvOZWk0Htik3J/w@public.gmane.org; alekseys-smomgflXvOZWk0Htik3J/w@public.gmane.org; yiftahs-smomgflXvOZWk0Htik3J/w@public.gmane.org; Tziporet Koren; alexr-smomgflXvOZWk0Htik3J/w@public.gmane.org
Subject: RE: When IBoE will be merged to upstream?

> Regarding GID to Eth mappings, we discussed using the standard 
> create_ah() Verb for this.
> In the kernel, create_ah() will call a generic address resolution 
> function in the cma.
> The returned information will be copied back to user-space in a 
> device- specific structure (since address handles are device-specific).
> 
> This eliminates adding a new get_eth_l2_addr() ABI call 
> (device-specific or not).
> In fact, this approach eliminates adding new ABIs for any kind of 
> address translation...
> 
> Does this seem reasonable?

The current behavior of ibv_create_ah() requires that the caller provide the L2, and if needed, L3 addressing.  Any translation between the L3 and L2 addressing must be done before the call is made.  E.g. ibv_create_ah does not use the GID to query the SA to obtain LIDs.  Why doesn't IBoE follow this same model?

LL: because of the RoCE spec, which states that only GID addressing is used at the Verbs level. The address handle fields are unchanged, and the L2 fields (e.g., lid) are reserved.
Note that in Ethernet, you normally don't specify L2 addresses at the transport level (i.e., sockets).

Callers can use some out of band mechanism for the mapping, call rdma_resolve_addr, or use some standard networking routine.
LL: this would require changes to the Verbs API. rdmacm programs addresses using user-space Verbs, or even just passes the application just address handle attributes...
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Jason Gunthorpe]

> The current behavior of ibv_create_ah() requires that the caller
> provide the L2, and if needed, L3 addressing.  Any translation
> between the L3 and L2 addressing must be done before the call is
> made.  E.g. ibv_create_ah does not use the GID to query the SA to
> obtain LIDs.  Why doesn't IBoE follow this same model?
> 
> LL: because of the RoCE spec, which states that only GID addressing
> is used at the Verbs level. The address handle fields are unchanged,
> and the L2 fields (e.g., lid) are reserved.  Note that in Ethernet,
> you normally don't specify L2 addresses at the transport level
> (i.e., sockets).

We do not have to lavishly follow the IBTA spec in the Linux
implementation, especially if it makes no sense.

I think Sean is on the right track here, the AH should take the L2 as
input just like for IB, and the resolution is done in librdmacm, or
somehow manually.

The verbs layer is not really analogous to sockets anyhow, the
librdmacm is much closer to a socket like interface, and it having
a GID go into rdmacm and a full AH with L2/L3 info come out seems
entirely reasonable.

BTW, what ever was decided about vlans tagging? Is that part of the AH
or do you use seperate RDMA devices per vlan? Seems like a point worth
considering now.

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

VLANs are part of L2 in Ethernet -- when you resolve a destination L3 address to an L2 address, you get the outgoing interface, which also determines the VLAN.
I think this approach has an advantage over an RDMA device per VLAN in that you keep the standard OS VLAN management (vconfig).

I wouldn't judge the RoCE spec so quickly --- it guarantees that rdma application binaries could run on any network.
What do you gain by exposing Eth-specific L2 params in the address handle?
--Liran


-----Original Message-----
From: Jason Gunthorpe [mailto:jgunthorpe-ePGOBjL8dl3ta4EC/59zMFaTQe2KTcn/@public.gmane.org] 
Sent: Thursday, June 24, 2010 11:37 PM
To: Liran Liss
Cc: Hefty, Sean; Roland Dreier; Aleksey Senin; linux-rdma; monis-smomgflXvOZWk0Htik3J/w@public.gmane.org; alekseys-smomgflXvOZWk0Htik3J/w@public.gmane.org; yiftahs-smomgflXvOZWk0Htik3J/w@public.gmane.org; Tziporet Koren; alexr-smomgflXvOZWk0Htik3J/w@public.gmane.org
Subject: Re: When IBoE will be merged to upstream?

> The current behavior of ibv_create_ah() requires that the caller 
> provide the L2, and if needed, L3 addressing.  Any translation between 
> the L3 and L2 addressing must be done before the call is made.  E.g. 
> ibv_create_ah does not use the GID to query the SA to obtain LIDs.  
> Why doesn't IBoE follow this same model?
> 
> LL: because of the RoCE spec, which states that only GID addressing is 
> used at the Verbs level. The address handle fields are unchanged, and 
> the L2 fields (e.g., lid) are reserved.  Note that in Ethernet, you 
> normally don't specify L2 addresses at the transport level (i.e., 
> sockets).

We do not have to lavishly follow the IBTA spec in the Linux implementation, especially if it makes no sense.

I think Sean is on the right track here, the AH should take the L2 as input just like for IB, and the resolution is done in librdmacm, or somehow manually.

The verbs layer is not really analogous to sockets anyhow, the librdmacm is much closer to a socket like interface, and it having a GID go into rdmacm and a full AH with L2/L3 info come out seems entirely reasonable.

BTW, what ever was decided about vlans tagging? Is that part of the AH or do you use seperate RDMA devices per vlan? Seems like a point worth considering now.

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Jason Gunthorpe]

On Fri, Jun 25, 2010 at 11:04:28AM +0300, Liran Liss wrote:

> VLANs are part of L2 in Ethernet -- when you resolve a destination
> L3 address to an L2 address, you get the outgoing interface, which
> also determines the VLAN.  I think this approach has an advantage
> over an RDMA device per VLAN in that you keep the standard OS VLAN
> management (vconfig).

Except that in RoCE all L3 addresses are link local GIDs, which must
be scoped to an interface and cannot be resolved by routing to a
specific interface. vconfig creates child ethernet devices, I think
you have no choice but to do the same for RDMA. The GID, when it is
resolved, must be scoped to the RDMA device it is going to be bound
to, which in turn must be bound to a VLAN.

(BTW, Sean, did AF_IB's sockaddr include a scoping field, and did you
figure out some way to make that work?)

> I wouldn't judge the RoCE spec so quickly --- it guarantees that
> rdma application binaries could run on any network.  What do you
> gain by exposing Eth-specific L2 params in the address handle?

Well, 1) invariably that is how the hardware must work, and verbs is
about exposing that interface to userspace 2) You don't suddenly make
AH setup require network traffic, and potentitally large time
delays 3) it keeps the whole RoCE architecture far more consistent
with IB.

You can pose the same question for IB, why doesn't AH resolution
resolve the GID? There are lots of good answers :)

Also bear in mind that APM is entirely possible over RoCE and doing
that will require a finer touch for managing the data in the AH's.

What do you get by doing all this extra work? I say nothing at
all. Users won't even be able to tell the difference as long as they
use rdmacm to setup the connections.

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Hefty, Sean]

> (BTW, Sean, did AF_IB's sockaddr include a scoping field, and did you
> figure out some way to make that work?)

The af_ib does include a scoping field, but the last set of patches doesn't make use of it.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

Jason,

There are also good reasons why the RoCE standard left the syntax of address handles.
First, it keeps the Verbs unchanged. Even if you are using rdmacm to make connections, you still have to inspect address handles when "connecting" to UD QPs or joining multicast addresses.
In addition, each incoming packet generates a CQE, whose L2 fields also need to be inspected.

Second, making Ethernet L2 fields explicit has implications beyond the address handle and CQE formats. Specifically, a lot of the IBTA defined MADs must be modified as well.
The most evident example is the CM protocol, which has L2 fields in its payloads.

Third, RoCE is not IB; its all about making RDMA user-friendly to Ethernet users.
Most importantly, we don't want to change the way Ethernet networks are managed.
This means that admins configure their normal network interfaces, define VLAN sub-interfaces, assign IP addresses (or use DHCP), and then work with RoCE using IP-mapped addresses, which reference the same IP addresses they use for their Ethernet interfaces.
So, regarding our VLAN discussion:
- RoCE gids are L3 addresses, which are not (necessarily) of link-local scope; people will mostly use IP-mapped gids of global scope.
- These gids will map to an IP address, which then can resolve to an outgoing vlan device exactly as in Ethernet.

We have a specification, we have an implementation, and we have clean way of passing RoCE L2 information to user-space via address handles.
I don't see any substantial reason to change the basic approach.

Regards,
--Liran


-----Original Message-----
From: Jason Gunthorpe [mailto:jgunthorpe-ePGOBjL8dl3ta4EC/59zMFaTQe2KTcn/@public.gmane.org] 
Sent: Friday, June 25, 2010 6:58 PM
To: Liran Liss
Cc: Hefty, Sean; Roland Dreier; Aleksey Senin; linux-rdma; monis-smomgflXvOZWk0Htik3J/w@public.gmane.org; alekseys-smomgflXvOZWk0Htik3J/w@public.gmane.org; yiftahs-smomgflXvOZWk0Htik3J/w@public.gmane.org; Tziporet Koren; alexr-smomgflXvOZWk0Htik3J/w@public.gmane.org
Subject: Re: When IBoE will be merged to upstream?

On Fri, Jun 25, 2010 at 11:04:28AM +0300, Liran Liss wrote:

> VLANs are part of L2 in Ethernet -- when you resolve a destination
> L3 address to an L2 address, you get the outgoing interface, which 
> also determines the VLAN.  I think this approach has an advantage over 
> an RDMA device per VLAN in that you keep the standard OS VLAN 
> management (vconfig).

Except that in RoCE all L3 addresses are link local GIDs, which must be scoped to an interface and cannot be resolved by routing to a specific interface. vconfig creates child ethernet devices, I think you have no choice but to do the same for RDMA. The GID, when it is resolved, must be scoped to the RDMA device it is going to be bound to, which in turn must be bound to a VLAN.

(BTW, Sean, did AF_IB's sockaddr include a scoping field, and did you figure out some way to make that work?)

> I wouldn't judge the RoCE spec so quickly --- it guarantees that rdma 
> application binaries could run on any network.  What do you gain by 
> exposing Eth-specific L2 params in the address handle?

Well, 1) invariably that is how the hardware must work, and verbs is about exposing that interface to userspace 2) You don't suddenly make AH setup require network traffic, and potentitally large time delays 3) it keeps the whole RoCE architecture far more consistent with IB.

You can pose the same question for IB, why doesn't AH resolution resolve the GID? There are lots of good answers :)

Also bear in mind that APM is entirely possible over RoCE and doing that will require a finer touch for managing the data in the AH's.

What do you get by doing all this extra work? I say nothing at all. Users won't even be able to tell the difference as long as they use rdmacm to setup the connections.

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Hefty, Sean]

> There are also good reasons why the RoCE standard left the syntax of
> address handles.
> First, it keeps the Verbs unchanged. Even if you are using rdmacm to make
> connections, you still have to inspect address handles when "connecting" to
> UD QPs or joining multicast addresses.

Verbs isn't an API, and the verbs usage changed anyway - the L2 is no longer provided by the user.  The spec is wrong to make the L3 to L2 mapping device specific, if that was the intent.  ARP or neighbor discovery should not be hidden under create AH or modify QP.

> Second, making Ethernet L2 fields explicit has implications beyond the
> address handle and CQE formats. Specifically, a lot of the IBTA defined
> MADs must be modified as well.

Specifically, which MADs and why?

> The most evident example is the CM protocol, which has L2 fields in its
> payloads.

How does moving the L3 to L2 mapping from outside create AH result in the CM protocol needing to change?  Why is hiding this under modify QP desirable?

> Third, RoCE is not IB; its all about making RDMA user-friendly to Ethernet
> users.

Then why is it part of the IBTA?

> Most importantly, we don't want to change the way Ethernet networks are
> managed.
> This means that admins configure their normal network interfaces, define
> VLAN sub-interfaces, assign IP addresses (or use DHCP), and then work with
> RoCE using IP-mapped addresses, which reference the same IP addresses they
> use for their Ethernet interfaces.
> So, regarding our VLAN discussion:
> - RoCE gids are L3 addresses, which are not (necessarily) of link-local
> scope; people will mostly use IP-mapped gids of global scope.
> - These gids will map to an IP address, which then can resolve to an
> outgoing vlan device exactly as in Ethernet.
> 
> We have a specification, we have an implementation, and we have clean way
> of passing RoCE L2 information to user-space via address handles.
> I don't see any substantial reason to change the basic approach.

All feedback on this has been discarded since the initial patches were submitted, and I still don't see a substantial reason why L3 to L2 mappings are device specific, or why significant network protocols should be hidden under verbs calls.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Jason Gunthorpe]

On Thu, Jul 01, 2010 at 05:20:05PM -0700, Hefty, Sean wrote:

> > The most evident example is the CM protocol, which has L2 fields in its
> > payloads.
> 
> How does moving the L3 to L2 mapping from outside create AH result
> in the CM protocol needing to change?  Why is hiding this under modify
> QP desirable?

I'm not too sure about that either.. The concept of a reduced
functionality TCA does not seem to have been considered as part of
RoCE, there is nothing like APM or path selection, so it is reasonable
to assume that the target in a CM can do the PR lookup equivilent to
get from a GID to the L2 information. The RoCE spec does not specify
that DMAC addresses must be exact match to QP data, so this is not a
problem like it would be in IB.

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Jason Gunthorpe]

On Thu, Jul 01, 2010 at 02:50:55PM +0300, Liran Liss wrote:

> We have a specification, we have an implementation, and we have
> clean way of passing RoCE L2 information to user-space via address
> handles.  I don't see any substantial reason to change the basic
> approach.

Actually, we have a spec that omits how to do the L3 to L2 address
mapping - and there is much disagreement on this point. Since no
standard was reached then agreement will have to be reached at least
with the Linux maintainers before I think anything can realistically
be merged.

The basic approach must be something people can agree on, and it seem
pretty clear to me at least that the current approach is not agreeable
to many people.

All three of your points are all entirely avoidable if you simply
stick to the idea that the L3 address is a GID and not an IP, and rely
on IPoIB like RDMA-CM mechanisms to go from IP to GID to L2 MAC/VLAN.

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Roland Dreier]

 > Third, RoCE is not IB; its all about making RDMA user-friendly to Ethernet users.

This is utter nonsense.  RoCE (or IBoE as I prefer ;) is absolutely
IB-over-Ethernet and it is all about making minimal changes to IB and IB
applications to run on Ethernet.

 > Most importantly, we don't want to change the way Ethernet networks are managed.

That makes sense.  However let's be honest with ourselves -- the
fraction of Ethernet networks using IPv6 as their only or even main
address scheme is pretty small.  Of course having a migration path to
work with IPv6 is important, but for the moment users want to use IPv4
addresses to specify destinations.

 > - RoCE gids are L3 addresses, which are not (necessarily) of link-local
 >   scope; people will mostly use IP-mapped gids of global scope.
 > - These gids will map to an IP address, which then can resolve to an
 >   outgoing vlan device exactly as in Ethernet.

At that level it all makes sense, but the problem is the specifics of
where, when and how the mapping is done.

 > We have a specification, we have an implementation, and we have clean
 > way of passing RoCE L2 information to user-space via address handles.

We may have an implementation but we absolutely don't have a
specification.  Or at least the IBA annex has nothing beyond this:

    A16.5.1 ADDRESS ASSIGNMENT AND RESOLUTION

    Layer 2 local addresses (i.e. SMAC, DMAC), and the methods by which
    those addresses are assigned, are outside the scope of this annex.

    The means for resolving a GID to a local port address (i.e. SMAC or
    DMAC) are outside the scope of this annex. It is assumed that
    standard Ethernet mechanisms, such as ARP or Neighbor Discovery are
    used to maintain an appropriate address cache for RoCE ports.

which was really pretty unfortunate, since it means the exact point
we're talking about is completely unspecified.  Or is there some other
spec you can point to?

(This also means it's pretty important that we get this right, since
every future implementation is going to have a lot of pressure to follow
what Linux does)

 > I don't see any substantial reason to change the basic approach.

I don't really even know what the basic approach is.  For example what's
the plan for handling GIDs that aren't derived from a MAC address?  For
a long time we've assumed that the create_ah verb can't sleep, so where
are you going to do neighbor discovery?

 - R.
-- 
Roland Dreier <rolandd-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

S.B.
--Liran

-----Original Message-----
From: Roland Dreier [mailto:rdreier-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org] 
Sent: Saturday, July 03, 2010 11:34 PM
To: Liran Liss
Cc: Jason Gunthorpe; Hefty, Sean; Aleksey Senin; linux-rdma; monis-smomgflXvOZWk0Htik3J/w@public.gmane.org; alekseys-smomgflXvOZWk0Htik3J/w@public.gmane.org; yiftahs-smomgflXvOZWk0Htik3J/w@public.gmane.org; Tziporet Koren; alexr-smomgflXvOZWk0Htik3J/w@public.gmane.org
Subject: Re: When IBoE will be merged to upstream?

 > Third, RoCE is not IB; its all about making RDMA user-friendly to Ethernet users.

This is utter nonsense.  RoCE (or IBoE as I prefer ;) is absolutely IB-over-Ethernet and it is all about making minimal changes to IB and IB applications to run on Ethernet.
LL: ??? IBoE has a completely different management plane (no SM, no SA), and we want this management plane to be "as Ethernet as possible" - this is what I mean by "not IB".
Otherwise, we are in full agreement: the changes to the IB transport should indeed be minimal.

 > Most importantly, we don't want to change the way Ethernet networks are managed.

That makes sense.  However let's be honest with ourselves -- the fraction of Ethernet networks using IPv6 as their only or even main address scheme is pretty small.  Of course having a migration path to work with IPv6 is important, but for the moment users want to use IPv4 addresses to specify destinations.

LL: I wasn't referring only to IPv6 networks; there are standard ways to represent IPv4 addresses in the IPv6 (and thus, iboe GID) namespace - IPv4 mapped addresses (::ffff<ipv4>).
Any mapped ipv4 address can be resolved according to ipv4.

 > - RoCE gids are L3 addresses, which are not (necessarily) of link-local
 >   scope; people will mostly use IP-mapped gids of global scope.
 > - These gids will map to an IP address, which then can resolve to an
 >   outgoing vlan device exactly as in Ethernet.

At that level it all makes sense, but the problem is the specifics of where, when and how the mapping is done.

 > We have a specification, we have an implementation, and we have clean  > way of passing RoCE L2 information to user-space via address handles.

We may have an implementation but we absolutely don't have a specification.  Or at least the IBA annex has nothing beyond this:

    A16.5.1 ADDRESS ASSIGNMENT AND RESOLUTION

    Layer 2 local addresses (i.e. SMAC, DMAC), and the methods by which
    those addresses are assigned, are outside the scope of this annex.

    The means for resolving a GID to a local port address (i.e. SMAC or
    DMAC) are outside the scope of this annex. It is assumed that
    standard Ethernet mechanisms, such as ARP or Neighbor Discovery are
    used to maintain an appropriate address cache for RoCE ports.

which was really pretty unfortunate, since it means the exact point we're talking about is completely unspecified.  Or is there some other spec you can point to?

(This also means it's pretty important that we get this right, since every future implementation is going to have a lot of pressure to follow what Linux does)

LL: the ibxoe working group has recommended using both IP-mapped and link-local addresses (http://www.t11.org/ftp/t11/pub/fc/study/09-543v0.pdf).
Other than that, there is no comprehensive spec so I am afraid you are right.
It seems natural to base iboe addressing on ipv6 practices:
- map ipv6 addresses in a straight forward manner.
- map ipv4 addresses using ipv4-mapped addresses.

 > I don't see any substantial reason to change the basic approach.

I don't really even know what the basic approach is.  For example what's the plan for handling GIDs that aren't derived from a MAC address?  For a long time we've assumed that the create_ah verb can't sleep, so where are you going to do neighbor discovery?

LL: by basic approach, I mean: without modifying IB L2 fields in address handles, CQEs, or in MAD payloads.
Iboe doesn't need to do discovery on its own; it can inherit the IP addresses, macs and vlans of the eth interface it is associated with.
GIDs that aren't derived from MAC addresses are IP-mapped addresses, which can be resolved according to their associated IP addresses.
So, from an admin's perspective, iboe address resolution matches whatever was configured for the eth interfaces; no new scheme.

Regarding the implementation, there is no inherent issue that prevents create_ah() from sleeping:
- Change a few spinlocks to mutexes in the cma (which sleeps a lot anyway because is modifies QP states)
- Trivial for user-space calls...



 - R.
--
Roland Dreier <rolandd-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Or Gerlitz]

Liran Liss wrote:
> S.B

reading your replies would be much easier if you use strict bottom posting

>> For a long time we've assumed that the create_ah verb can't sleep, 
>> so where are you going to do neighbor discovery?

> re [...] implementation, there is no inherent issue that prevents create_ah() from sleeping:
> - Change a few spinlocks to mutexes in the cma (which sleeps a lot anyway because is 
>   modifies QP states)
> - Trivial for user-space calls...

Documentation/infiniband/core_locking.txt states that "The corresponding functions exported to upper level protocol consumers: ... ib_create_ah ... are therefore safe to call from any context." 

Which is in turn assumed by bunch of components in the kernel IB stack (look for ib_create_ah calls  under drivers/infiniband/core/). The examples you brought here, don't cover them. This way or another, I don't see any reason to break that convension just for the sake of this aspect of the iboe implementation, simply, code with this assumption at hand.

Or.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

> reading your replies would be much easier if you use strict 
> bottom posting

OK :)

> 
> >> For a long time we've assumed that the create_ah verb 
> can't sleep, so 
> >> where are you going to do neighbor discovery?
> 
> > re [...] implementation, there is no inherent issue that 
> prevents create_ah() from sleeping:
> > - Change a few spinlocks to mutexes in the cma (which 
> sleeps a lot anyway because is 
> >   modifies QP states)
> > - Trivial for user-space calls...
> 
> Documentation/infiniband/core_locking.txt states that "The 
> corresponding functions exported to upper level protocol 
> consumers: ... ib_create_ah ... are therefore safe to call 
> from any context." 
> 
> Which is in turn assumed by bunch of components in the kernel 
> IB stack (look for ib_create_ah calls  under 
> drivers/infiniband/core/). The examples you brought here, 
> don't cover them. This way or another, I don't see any reason 
> to break that convension just for the sake of this aspect of 
> the iboe implementation, simply, code with this assumption at hand.
> 
> Or.
> 

I understand this, but keeping ib_create_ah() callable from any context is not a goal by itself.
I am looking for constructive ideas for supporting iboe without breaking Verbs/CQE/CM syntax.

What do you propose?--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Or Gerlitz]

Liran Liss wrote:
> but keeping ib_create_ah() callable from any context is not a goal by itself.

going with your approach, if your proposed design is accepted, I believe that you probably need to patch all the code-chains that makes calls under the current assumption

> I am looking for constructive ideas for supporting iboe without breaking 
> Verbs/CQE/CM syntax. 

I don't agree that exposing the Ethernet L2 related information to the caller is breaking something, the converse, it is a required enhancement. 

I think we need to let resolve through the rdma-cm && get to know at the consumer level, what are the source / destination macs, vlan id and vlan priority used by an IBoE QP, in the exact manner all the IB equivalents (src/dst lid, pkey, sl) are resolved by the rdma-cm and exposed to the consmer app for IB QP.

Or.

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Jason Gunthorpe]

On Wed, Jul 07, 2010 at 09:00:26AM +0300, Or Gerlitz wrote:

> I think we need to let resolve through the rdma-cm && get to know at
> the consumer level, what are the source / destination macs, vlan id
> and vlan priority used by an IBoE QP, in the exact manner all the IB
> equivalents (src/dst lid, pkey, sl) are resolved by the rdma-cm and
> exposed to the consmer app for IB QP.

I agree.

Clearly following the model of IB is the best way to fit this in
without major changing. RDMA-CM is the way to get IP integration, it
uses existing eth devices attached to the master eth device (analogous
to IPoIB devices) and resolves IP to eth device to VLAN header and
neighbour.

If someone needs to do something else special it is pretty easy to do
all the same steps in userspace using netlink, and that could go into
a library, just like PR queries for IB.

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

This discussion has derived into whether we need to expose Eth L2 params across the Verbs interface.
This point has been discussed extensively in the IBTA during the development of the RoCE spec. And the direction chosen by the spec is clear:

"...A verbs consumer using a RoCE network relies strictly on so-called Layer
3 addressing (GIDs); layer 2 addresses (e.g. subnet local identifiers) are not passed across the verbs interface..."

The motivation behind this direction had to do with preserving API transparency for applications.
The goal was to allow existing and future applications to run above RoCE and native IB without changes. 

As opposed to what it may seem at first sight, adding Eth L2 parameters to the address vector, *does not* make RoCE closer to IB.
It actually goes the other way around.
Here is a quick list of what would have to be changed if we were to include Eth L2 address parameters to the Address Vector and other structures/functions that expose L2 params:

Structure changes:
- ibv_wc
- ibv_ah_attr
-- ibv_qp_attr
-- rdma_ud_param
--- rdma_cm_event
- ibv_port_attr

Verb API changes:
- ibv_poll_cq()
- ibv_init_ah_from_wc()
- ibv_create_ah()

- ibv_query_qp()
- ibv_modify_qp()

- ibv_query_port()

- ibv_attach_mcast()
- ibv_detach_mcast()

rdmacm API changes:
- rdma_post_ud_send()
- rdma_get_cm_event()
- rdma_ack_cm_event()
 
As a result of this:
- Existing IB binaries would cease working over RoCE.
- Due to added fields in structures, even just recompiling existing applications from source would be problematic.
- To make future applications work on both ib and RoCE transparently, you would need additional wrappers such as init_ah(), copy_ah(), and ah_is_equal(), and never inspect address handle fields directly.

So why introduce differences between RoCE and IB (for the Application writers!!) when they *aren't* needed? Using rdmacm won't solve this either (UD traffic).
By following the direction set forth by the RoCE spec none of this is required. Existing (rdmacm) application binaries do run over RoCE or IB unchanged.

Granted, the RoCE spec approach introduces 2 *implementation* issues that we need to tackle:
1. Address resolution, which is a generic function, should not be a device-specific call.
In this matter, we already proposed a solution where resolution is done, as required, in generic functions in the kernel.
Specifically, we provide L2 information to user-space drivers via create_ah(), avoiding the need to add a new ABI call altogether, while the resolution would take place in a generic CMA routine.

2. The Kernel currently assumes that create_ah() can execute in atomic context.
One option is to distinguish between the create_ah() calls (in the kernel) that are done for iboe, which are very few, and the rest of the calls that are ib-only.
There are other approaches to solve this as well.

It seems clear that our goal should be to solve these issues inside the kernel, in the cleanest manner as possible, while preserving transparency to the applications.

Comments are welcome.

Liran 

> -----Original Message-----
> From: Or Gerlitz [mailto:ogerlitz-hKgKHo2Ms0FWk0Htik3J/w@public.gmane.org] 
> Sent: Wednesday, July 07, 2010 9:00 AM
> To: Liran Liss
> Cc: Roland Dreier; Jason Gunthorpe; Hefty, Sean; Aleksey 
> Senin; linux-rdma; monis-smomgflXvOZWk0Htik3J/w@public.gmane.org; alekseys-smomgflXvOZWk0Htik3J/w@public.gmane.org; 
> yiftahs-smomgflXvOZWk0Htik3J/w@public.gmane.org; Tziporet Koren; alexr-smomgflXvOZWk0Htik3J/w@public.gmane.org
> Subject: Re: When IBoE will be merged to upstream?
> 
> Liran Liss wrote:
> > but keeping ib_create_ah() callable from any context is not 
> a goal by itself.
> 
> going with your approach, if your proposed design is 
> accepted, I believe that you probably need to patch all the 
> code-chains that makes calls under the current assumption
> 
> > I am looking for constructive ideas for supporting iboe without 
> > breaking Verbs/CQE/CM syntax.
> 
> I don't agree that exposing the Ethernet L2 related 
> information to the caller is breaking something, the 
> converse, it is a required enhancement. 
> 
> I think we need to let resolve through the rdma-cm && get to 
> know at the consumer level, what are the source / destination 
> macs, vlan id and vlan priority used by an IBoE QP, in the 
> exact manner all the IB equivalents (src/dst lid, pkey, sl) 
> are resolved by the rdma-cm and exposed to the consmer app for IB QP.
> 
> Or.
> 
> --
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Jason Gunthorpe]

On Sat, Jul 10, 2010 at 11:55:18AM +0300, Liran Liss wrote:

> "...A verbs consumer using a RoCE network relies strictly on
> so-called Layer 3 addressing (GIDs); layer 2 addresses (e.g. subnet
> local identifiers) are not passed across the verbs interface..."

Ah, hmm, well, I was on that list during this time and I don't think
this statement means what you are saying it does :)
 
> As opposed to what it may seem at first sight, adding Eth L2
> parameters to the address vector, *does not* make RoCE closer to IB.
> It actually goes the other way around.  Here is a quick list of what
> would have to be changed if we were to include Eth L2 address
> parameters to the Address Vector and other structures/functions that
> expose L2 params:

Umh, no.. Stick the L2 address in the GID itself, stick the vlan tag
in either the DLID or the PKey and you are done. No structures get
bigger, nothing really changes.

There are already AH differences between iwarp and IB and generic code
to handle them.

rdmacm can support address resultion only for UD applications.

Still not seeing how this is a big issue? 

You cannot hide destination resolution in create_ah. You just
can't. create_ah does not accept any sort of source address specifier
which is absolutely critical to reoslve a destination ip address in
all cases, ie for instance IPv6 link local addresses, but there are
other cases too.

If IP semantics are going to be used then *ALL* linux IP semantics
must be supported, not just those that are convinent to implement, and
that means you need a source IP address, device, QOS parameters and
destination IP address to make a route determination.

The only subsystem in verbs that has this information is RDMA CM.

To me this is a fundamental problem that completely nixes L3 path 
resolution in create_ah as a possible solution - as you explained
you need to do the L3 path resolution to figure out the vlan tag to
use, to get the desired IP netdevice, to execute the ND query..

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

> > "...A verbs consumer using a RoCE network relies strictly
> on so-called
> > Layer 3 addressing (GIDs); layer 2 addresses (e.g. subnet local
> > identifiers) are not passed across the verbs interface..."
> 
> Ah, hmm, well, I was on that list during this time and I don't think 
> this statement means what you are saying it does :)
>

?? It doesn't get any clearer than this.
  
> > As opposed to what it may seem at first sight, adding Eth L2 
> > parameters to the address vector, *does not* make RoCE closer to IB.
> > It actually goes the other way around.  Here is a quick
> list of what
> > would have to be changed if we were to include Eth L2 address 
> > parameters to the Address Vector and other
> structures/functions that
> > expose L2 params:
> 
> Umh, no.. Stick the L2 address in the GID itself, stick the vlan tag 
> in either the DLID or the PKey and you are done. No structures get 
> bigger, nothing really changes.
> 

Good idea! This is exactly what we do today for addresses that the user explicitly declares as link-local addresses.
But, we can't mandate an overload of the GID in a way that it prevents its use as a true L3 address (eventually routable). 

> There are already AH differences between iwarp and IB and generic code 
> to handle them.

Right, the iwarp code relies on rdmacm only to bind to a device, and then does L3-->L2 resolution on its own.

> 
> rdmacm can support address resultion only for UD applications.
> 
> Still not seeing how this is a big issue? 

UD applications would like to enjoy link-layer transparency as well, you know...

> 
> You cannot hide destination resolution in create_ah. You just can't.

Why?

> create_ah does not accept any sort of source address 
> specifier

You are wrong -- sgid_index specifies it.

> which is absolutely critical to reoslve a 
> destination ip address in all cases, ie for instance IPv6 link local 
> addresses, but there are other cases too.
> 
> If IP semantics are going to be used then *ALL* linux IP semantics 
> must be supported, not just those that are convinent to implement, and 
> that means you need a source IP address, device, QOS parameters and 
> destination IP address to make a route determination.

An address handle has a source address, destination address, flow_label, and traffic class, and is already associated with a specific device.
In fact, it is a superset of every field in the socket address that the application passed rdmacm in the first place!

> 
> The only subsystem in verbs that has this information is RDMA CM.
> 

Right, this is why we will call a generic rdmacm function to resolve the address handle for us :)

> To me this is a fundamental problem that completely nixes L3

Jason, bottom line, I think that we both agree that the rdmacm should do the address resolution.
The difference is that by having the rdmacm initially only bind to the device and complete the resolution later (by a call from create_ah()), we don't change the user API for *all* gid types.
Having addressed your concerns regarding resolution below the Verbs, we continue to believe that this is the best approach.

--Liran
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Jason Gunthorpe]

On Mon, Jul 12, 2010 at 10:58:19AM +0300, Liran Liss wrote:

>> "...A verbs consumer using a RoCE network relies strictly
>> > on so-called
>> > > Layer 3 addressing (GIDs); layer 2 addresses (e.g. subnet local
>> > > identifiers) are not passed across the verbs interface..."
>> > 
>> > Ah, hmm, well, I was on that list during this time and I don't
>> think 
>> > this statement means what you are saying it does  :) 
>> >

> ?? It doesn't get any clearer than this.
  
'subnet local identidifer' == LID

The text is saying that the specification does not use any of the LID
fields in the verbs interface, that is it. It isn't talking about MAC
addresses.

Exactly how and where the MAC address comes about was never decided,
and at least some participants thought it should be a 1:1 algorithmic
mapping from the GID.

Ditto for VLANs, how and where the vlan tag comes about is not part of
the spec.

> Good idea! This is exactly what we do today for addresses that the
> user explicitly declares as link-local addresses.  But, we can't
> mandate an overload of the GID in a way that it prevents its use as
> a true L3 address (eventually routable).

We are very unlikely to see routable IBoE, ever.. But, even if we do
get there some day then we could extend the AH.

BTW, I absolutely hate the mixing of 'Sometimes it is a IPv4,
sometimes it is a GID, and sometimes it is an IPv6' in the same
field. That is just so nasty. The GID is a GID, don't overload it in
an ambiguous way to mean 2 other things!

> > create_ah does not accept any sort of source address 
> > specifier
> 
> You are wrong -- sgid_index specifies it.

So, what do you propose to put in sgid_index? It isn't big enough to
store an IPv6 address. You can't exactly number every IP assigned to
every ethernet interface.

The other fields you mention are not a supserset of socket parameters,
they are only IPv6 parameters, IPv4 uses a different set.

> Jason, bottom line, I think that we both agree that the rdmacm
> should do the address resolution.  The difference is that by having
> the rdmacm initially only bind to the device and complete the
> resolution later (by a call from create_ah()), we don't change the
> user API for *all* gid types.  Having addressed your concerns
> regarding resolution below the Verbs, we continue to believe that
> this is the best approach.

Again, I don't see how what I've outlined changes the API in any
way.

Doing two routing lookups for the same connection is bad design, it is
racey. L2 parameters have to flow from the first routing lookup in
RDMA-CM to everything else.

Liran, I don't think you have at all come close to addressing my
concerns, you still haven't explained how a full route lookup is even
possible in create_ah, for instance. Let alone my other concerns!

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

> >> "...A verbs consumer using a RoCE network relies strictly
> >> > on so-called
> >> > > Layer 3 addressing (GIDs); layer 2 addresses (e.g. subnet local
> >> > > identifiers) are not passed across the verbs interface..."
> >> > 
> >> > Ah, hmm, well, I was on that list during this time and I don't
> >> think
> >> > this statement means what you are saying it does  :)
> >> >
> 
> > ?? It doesn't get any clearer than this.
>   
> 'subnet local identidifer' == LID
> 
> The text is saying that the specification does not use any of the LID
> fields in the verbs interface, that is it. It isn't talking about MAC
> addresses.
> 
> Exactly how and where the MAC address comes about was never decided,
> and at least some participants thought it should be a 1:1 algorithmic
> mapping from the GID.
> 
> Ditto for VLANs, how and where the vlan tag comes about is not part of
> the spec.
> 

You are trying to rewrite history.
Read the spec, address handles fields are fixed.

> > Good idea! This is exactly what we do today for addresses that the
> > user explicitly declares as link-local addresses.  But, we can't
> > mandate an overload of the GID in a way that it prevents its use as
> > a true L3 address (eventually routable).
> 
> We are very unlikely to see routable IBoE, ever..

Says who?

> But, even if we do get there some day then we could extend the AH.

This is unacceptable - we are not going to add another L3 identifier.
 
> BTW, I absolutely hate the mixing of 'Sometimes it is a IPv4,
> sometimes it is a GID, and sometimes it is an IPv6' in the same
> field. That is just so nasty. The GID is a GID, don't overload it in
> an ambiguous way to mean 2 other things!

A GID is a GID indeed -- in a RoCE environment, it's the layer 3 identifier.
All of our intended values are standard ipv6 encapsulations.

> > > create_ah does not accept any sort of source address 
> > > specifier
> > 
> > You are wrong -- sgid_index specifies it.
> 
> So, what do you propose to put in sgid_index? It isn't big enough to
> store an IPv6 address. You can't exactly number every IP assigned to
> every ethernet interface.

An iboe device is associated with a specific Ethernet interface. Thus, its gid table
only needs to map the ip addresses assigned to that interface.

> The other fields you mention are not a supserset of socket parameters,
> they are only IPv6 parameters, IPv4 uses a different set.

Like what?

> > Jason, bottom line, I think that we both agree that the rdmacm
> > should do the address resolution.  The difference is that by having
> > the rdmacm initially only bind to the device and complete the
> > resolution later (by a call from create_ah()), we don't change the
> > user API for *all* gid types.  Having addressed your concerns
> > regarding resolution below the Verbs, we continue to believe that
> > this is the best approach.
> 
> Again, I don't see how what I've outlined changes the API in any
> way.

We currently support link-local gids, but the architecture must not limit the scope.

> 
> Doing two routing lookups for the same connection is bad design, it is
> racey. L2 parameters have to flow from the first routing lookup in
> RDMA-CM to everything else.
> 

So is caching L3-->L2 mappings that change a second later...
So what?

> Liran, I don't think you have at all come close to addressing my
> concerns, you still haven't explained how a full route lookup is even
> possible in create_ah, for instance. Let alone my other concerns!
> 
> Jason
> --
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Jason Gunthorpe]

On Tue, Jul 13, 2010 at 11:26:41AM +0300, Liran Liss wrote:

> > 'subnet local identidifer' == LID
> > 
> > The text is saying that the specification does not use any of the LID
> > fields in the verbs interface, that is it. It isn't talking about MAC
> > addresses.
> > 
> > Exactly how and where the MAC address comes about was never decided,
> > and at least some participants thought it should be a 1:1 algorithmic
> > mapping from the GID.
> > 
> > Ditto for VLANs, how and where the vlan tag comes about is not part of
> > the spec.

> You are trying to rewrite history.
> Read the spec, address handles fields are fixed.

Not really, this was all discussed on this list before the IBxoE
working group was formed, it was discussed in the working group, I
objected to the draft spec leaving this area absent, even. The spec
doesn't say squat about how MAC and VLAN values get into the AH, and
you have already heard how my opinion on this subject differs from
others.

> > But, even if we do get there some day then we could extend the AH.
> 
> This is unacceptable - we are not going to add another L3 identifier.

It wouldn't be adding another L3 itentifier it would be an L2 next hop
MAC address for the router. It would be nice to do this from the start
but if growing the AH is really that scary then it should wait until
someone figures out how to solve the lossless routing problem on ethernet.

> > BTW, I absolutely hate the mixing of 'Sometimes it is a IPv4,
> > sometimes it is a GID, and sometimes it is an IPv6' in the same
> > field. That is just so nasty. The GID is a GID, don't overload it in
> > an ambiguous way to mean 2 other things!
> 
> A GID is a GID indeed -- in a RoCE environment, it's the layer 3 identifier.
> All of our intended values are standard ipv6 encapsulations.

What makes a GID a GID is the fact that it is a seperate addressing
space from IPv6! If it is a GID then you don't overload it, if it is
an IPv6 then you don't get to special case certain things, like link
local!

> > > > create_ah does not accept any sort of source address 
> > > > specifier
> > > 
> > > You are wrong -- sgid_index specifies it.
> > 
> > So, what do you propose to put in sgid_index? It isn't big enough to
> > store an IPv6 address. You can't exactly number every IP assigned to
> > every ethernet interface.
> 
> An iboe device is associated with a specific Ethernet
> interface. Thus, its gid table only needs to map the ip addresses
> assigned to that interface.

A few messages ago you said there was only one RDMA device per
physical ethernet interface, not one per vlan! VLAN interfaces can
have overlapping addreses (ie IPv6 link local) so I really don't see
how creating an GID table helps dis-ambiguate these cases.

> > Doing two routing lookups for the same connection is bad design, it is
> > racey. L2 parameters have to flow from the first routing lookup in
> > RDMA-CM to everything else.
> 
> So is caching L3-->L2 mappings that change a second later...
> So what?

No, it is not the same.

If you do a route lookup you get an atomic result from the routing
table that represents something an admin configured. If you do two
lookups and use information from both then the net result might be a
configuration that was never admin configured - ie you loose the
atomicity of route configuration change.

Normally ND mappings (L3->L2) track updates through the things that
use them. The fact this cannot happen with IBoE is another bug, and
again, a reason why it is unsuitable to treat a GID as an IPv6 address
when you cannot provide the same functionality.

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

> > > The text is saying that the specification does not use any of the 
> > > LID fields in the verbs interface, that is it. It isn't talking 
> > > about MAC addresses.
> > > 
> > > Exactly how and where the MAC address comes about was 
> never decided, 
> > > and at least some participants thought it should be a 1:1 
> > > algorithmic mapping from the GID.
> > > 
> > > Ditto for VLANs, how and where the vlan tag comes about 
> is not part 
> > > of the spec.
> 
> > You are trying to rewrite history.
> > Read the spec, address handles fields are fixed.
> 
> Not really, this was all discussed on this list before the 
> IBxoE working group was formed,

The paragraph above is about the RoCE spec. And *this list* did not write the RoCE spec.

> it was discussed in the 
> working group,

The RoCE spec adopts the verbs defined in the base IB spec and does not add any new input modifiers to the AH verb. You may not agree with it but that does not change the spec.

> I objected to the draft spec leaving this area 
> absent, even.

You should submit a comment on this matter using the IBTA comment tracker database if you intend your concern to be taken into account.

> The spec doesn't say squat about how MAC and 
> VLAN values get into the AH,

True. The spec does not say it because there are no MAC and VLAN input modifiers to the "create AH" verb. The spec assumes the resolution from the L3 address happens below the channel interface.

> and you have already heard how 
> my opinion on this subject differs from others.

I never attempted to misrepresent your opinion. I am just pointing out what the RoCE spec says.

> 
> > > But, even if we do get there some day then we could extend the AH.
> > 
> > This is unacceptable - we are not going to add another L3 
> identifier.
> 
> It wouldn't be adding another L3 itentifier it would be an L2 
> next hop MAC address for the router. It would be nice to do 
> this from the start but if growing the AH is really that 
> scary then it should wait until someone figures out how to 
> solve the lossless routing problem on ethernet.

Augmenting the AH has a significant cost. There is a tradeoff here between preserving the verbs api vs. dealing with the implementation challenges associated with doing address resolution below the verbs. The RoCE spec deliberately chooses one direction. You seem to favor the other one. But in the interest of progress and since we all seem to agree on the way things work when we use link local GIDs, let us move forward with that approach for now. And we can get back to non local GIDs later.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Jason Gunthorpe]

On Thu, Jul 15, 2010 at 01:55:32PM +0300, Liran Liss wrote:

> > I objected to the draft spec leaving this area 
> > absent, even.
> 
> You should submit a comment on this matter using the IBTA comment
> tracker database if you intend your concern to be taken into account.

The position of IBTA is that the L2 layer is not specified as part of
the spec, so of course there is no talk of how to get/create L2
information. The spec is *silent* on the issue of L2 addressing, so,
IMHO, it is compltely wrong to assume it specs one approach over
another, just because it omits L2 addressing related
discussion/fields/etc.

It, unfortunately, becomes implementation defined - and if that means
an implementation chooses to extend the AH, then so be it.

This is the problem with rushing incomplete specs through :)

> > It wouldn't be adding another L3 itentifier it would be an L2 
> > next hop MAC address for the router. It would be nice to do 
> > this from the start but if growing the AH is really that 
> > scary then it should wait until someone figures out how to 
> > solve the lossless routing problem on ethernet.

> Augmenting the AH has a significant cost. There is a tradeoff here
> between preserving the verbs api vs. dealing with the implementation
> challenges associated with doing address resolution below the
> verbs. The RoCE spec deliberately chooses one direction. You seem to
> favor the other one. But in the interest of progress and since we
> all seem to agree on the way things work when we use link local
> GIDs, let us move forward with that approach for now. And we can get
> back to non local GIDs later.

You still have to solve the problem with vlan tags, and either each
vlan interface has a seperate rdma interface or the tag has to flow
into the AH from the RDMA-CM.

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Paul Grun]

> -----Original Message-----
> From: linux-rdma-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org [mailto:linux-rdma-
> owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org] On Behalf Of Jason Gunthorpe
> Sent: Monday, July 12, 2010 9:22 AM
> To: Liran Liss
> Cc: Or Gerlitz; Roland Dreier; Hefty, Sean; Aleksey Senin; linux-rdma;
> monis-smomgflXvOZWk0Htik3J/w@public.gmane.org; alekseys-smomgflXvOZWk0Htik3J/w@public.gmane.org; yiftahs-smomgflXvOZWk0Htik3J/w@public.gmane.org; Tziporet
> Koren; alexr-smomgflXvOZWk0Htik3J/w@public.gmane.org
> Subject: Re: When IBoE will be merged to upstream?
> 
> On Mon, Jul 12, 2010 at 10:58:19AM +0300, Liran Liss wrote:
> 
> >> "...A verbs consumer using a RoCE network relies strictly
> >> > on so-called
> >> > > Layer 3 addressing (GIDs); layer 2 addresses (e.g. subnet local
> >> > > identifiers) are not passed across the verbs interface..."
> >> >
> >> > Ah, hmm, well, I was on that list during this time and I don't
> >> think
> >> > this statement means what you are saying it does  :)
> >> >
> 
> > ?? It doesn't get any clearer than this.
> 
> 'subnet local identidifer' == LID
> 
> The text is saying that the specification does not use any of the LID
> fields in the verbs interface, that is it. It isn't talking about MAC
> addresses.
> 

This is incorrect.  The intent of the text is that a verbs consumer deals
ONLY in layer 3 addresses (GIDs).  It doesn't make any sense to restrict the
interpretation to mean 'LIDs', since LIDs are a NOP for RoCE.  

CA16-17: When accessing the services of a RoCE verbs provider, the
source and destination identifiers contained in the address vector shall
consist of GIDs; the address vector shall not contain layer 2 references
(e.g. local addresses). Layer 2 references include source and destination
local identifiers and LID Path Bits.

"layer 2 references" refers to LIDs, MAC IDs or any other form of layer 2
address.  If the wording of the text is insufficiently clear, please post a
comment on comment tracker on the IBTA website.  Nevertheless, that is the
intent.  

> Exactly how and where the MAC address comes about was never decided,

Correct.  The IBTA IBXoE WG felt that defining the mapping from GID to MAC
ID should be a function of the underlying fabric (Ethernet) and thus was out
of scope for us to define the mapping mechanism.


> and at least some participants thought it should be a 1:1 algorithmic
> mapping from the GID.
> 
> Ditto for VLANs, how and where the vlan tag comes about is not part of
> the spec.
> 
> > Good idea! This is exactly what we do today for addresses that the
> > user explicitly declares as link-local addresses.  But, we can't
> > mandate an overload of the GID in a way that it prevents its use as
> > a true L3 address (eventually routable).
> 
> We are very unlikely to see routable IBoE, ever.. But, even if we do
> get there some day then we could extend the AH.
> 

Please keep in mind that IB routing is not yet defined.  Although the RoCE
spec doesn't explicitly say so, the intent is that routing for RoCE can be
defined once the work on IB routing has been completed. 

> BTW, I absolutely hate the mixing of 'Sometimes it is a IPv4,
> sometimes it is a GID, and sometimes it is an IPv6' in the same
> field. That is just so nasty. The GID is a GID, don't overload it in
> an ambiguous way to mean 2 other things!

I am unaware of any overloading, at least in the RoCE spec.

> 
> > > create_ah does not accept any sort of source address
> > > specifier
> >
> > You are wrong -- sgid_index specifies it.
> 
> So, what do you propose to put in sgid_index? It isn't big enough to
> store an IPv6 address. You can't exactly number every IP assigned to
> every ethernet interface.
> 
> The other fields you mention are not a supserset of socket parameters,
> they are only IPv6 parameters, IPv4 uses a different set.
> 
> > Jason, bottom line, I think that we both agree that the rdmacm
> > should do the address resolution.  The difference is that by having
> > the rdmacm initially only bind to the device and complete the
> > resolution later (by a call from create_ah()), we don't change the
> > user API for *all* gid types.  Having addressed your concerns
> > regarding resolution below the Verbs, we continue to believe that
> > this is the best approach.
> 
> Again, I don't see how what I've outlined changes the API in any
> way.
> 
> Doing two routing lookups for the same connection is bad design, it is
> racey. L2 parameters have to flow from the first routing lookup in
> RDMA-CM to everything else.
> 
> Liran, I don't think you have at all come close to addressing my
> concerns, you still haven't explained how a full route lookup is even
> possible in create_ah, for instance. Let alone my other concerns!
> 
> Jason
> --
> To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
> the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Paul (Chair, IBTA IBXoE WG)

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Jason Gunthorpe]

On Mon, Jul 19, 2010 at 10:28:24PM -0700, Paul Grun wrote:

> This is incorrect.  The intent of the text is that a verbs consumer deals
> ONLY in layer 3 addresses (GIDs).  It doesn't make any sense to restrict the
> interpretation to mean 'LIDs', since LIDs are a NOP for RoCE.  
> 
> CA16-17: When accessing the services of a RoCE verbs provider, the
> source and destination identifiers contained in the address vector shall
> consist of GIDs; the address vector shall not contain layer 2 references
> (e.g. local addresses). Layer 2 references include source and destination
> local identifiers and LID Path Bits.
> 
> "layer 2 references" refers to LIDs, MAC IDs or any other form of layer 2
> address.  If the wording of the text is insufficiently clear, please post a
> comment on comment tracker on the IBTA website.  Nevertheless, that is the
> intent.  

Since it never actually says MAC address it reads like it is just
excluding existing IB L2 addresses from use in ROCEE which makes alot
sense. If MAC address was ment, it should have been listed explicitly!

> > Exactly how and where the MAC address comes about was never decided,
> 
> Correct.  The IBTA IBXoE WG felt that defining the mapping from GID to MAC
> ID should be a function of the underlying fabric (Ethernet) and thus was out
> of scope for us to define the mapping mechanism.

IHMO, it is a bad design to create a architecture that requires the L2
information in the AH, forbid the L2 information from being passed
into the AH APIs, and then not specify how the L2 information is
created.

How is anyone supposed to implement this?

> > BTW, I absolutely hate the mixing of 'Sometimes it is a IPv4,
> > sometimes it is a GID, and sometimes it is an IPv6' in the same
> > field. That is just so nasty. The GID is a GID, don't overload it in
> > an ambiguous way to mean 2 other things!
> 
> I am unaware of any overloading, at least in the RoCE spec.

This is a feature added as part of the proposed patch set that spurned
this whole discussion. It is this feature that made create_ah into a
blocking call ...

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Hefty, Sean]

> > There are already AH differences between iwarp and IB and generic code
> > to handle them.
> 
> Right, the iwarp code relies on rdmacm only to bind to a device, and then
> does L3-->L2 resolution on its own.

iwarp does the resolution as part of connecting, not through a verbs call.

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Roland Dreier]

 > But, we can't mandate an overload of the GID in a way that it
 > prevents its use as a true L3 address (eventually routable).

Actually I'm beginning to think that the only possible way we can use
the GID in IBoE is as a link-local IPv6 addresses containing an Ethernet
address.  Trying to hide neighbour discovery or ARP below the verbs
doesn't seem workable -- being forced to change the locking rules we've
had for the past 5+ years about create_ah is just the beginning.  We get
further problems if a remote address should ever change and I'm probably
missing other issues.

So the best solution I can see is to declare that an IBoE GID must be an
IPv6 address coming from an EUI-64 Ethernet address for the
corresponding port; for MGIDs I guess we use the standard IPv6 mapping
to Ethernet address 33:33:xx:xx:xx:xx.

I'm not sure how we want to handle IPv4 -- presumably unicast ARP can be
done within the RDMA CM, which will then create a DGID with the
appropriate Ethernet address.  However it's not clear to me whether we
need a way to create IPv4 (01:00:5e:xx:xx:xx) multicast addresses.

Also, since there is no way to map a link-local IPv6 address to a
particular interface, then I guess we need a way to pass in the VLAN tag
to be used -- presumably we can steal some other field for the 12 bits.
(The fact that the IBoE annex does not mention VLANs or 802.1q a single
time is just another thing that shows how rushed and incomplete it is)

With all this said, I think it means we do not need to do the mapping
from GID to Ethernet address in the kernel for IBoE user verbs, since it
is so simple -- we can simply add a fairly trivial helper to libibverbs.

 - R.
-- 
Roland Dreier <rolandd-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Jason Gunthorpe]

On Mon, Jul 12, 2010 at 02:20:22PM -0700, Roland Dreier wrote:

> So the best solution I can see is to declare that an IBoE GID must be an
> IPv6 address coming from an EUI-64 Ethernet address for the
> corresponding port; for MGIDs I guess we use the standard IPv6 mapping
> to Ethernet address 33:33:xx:xx:xx:xx.

This is what I have been advocating..

A quibble about multicast - AFAIK this is unsolved. I think some spec
needs to be agreed that documents what sort of multicast snooping
operations switches need to do, ie if IGMP joins imply that IBoE
traffic for the same DMAC is included in the join, or if IBoE requires
a seperate IGMP type process on its own ether-type. That would make it
much clearer what to do with MGIDs.

IPv4 could be handled by mapping a IPv4 multicast address within an
IPv6 mapped address, if necessary.

It would be nice to at least have a plan on how to integrate a
non-link local address, if that is ever necessary in future. An
extended AH with an additional 48 DMAC field seems reasonable to me?

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Roland Dreier]

 > A quibble about multicast - AFAIK this is unsolved. I think some spec
 > needs to be agreed that documents what sort of multicast snooping
 > operations switches need to do, ie if IGMP joins imply that IBoE
 > traffic for the same DMAC is included in the join, or if IBoE requires
 > a seperate IGMP type process on its own ether-type. That would make it
 > much clearer what to do with MGIDs.

I agree -- the current spec is rather broken for multicast.  Choosing a
different ethertype and then saying that all switches will just flood
multicast traffic is half-baked at best.

 > It would be nice to at least have a plan on how to integrate a
 > non-link local address, if that is ever necessary in future. An
 > extended AH with an additional 48 DMAC field seems reasonable to me?

You mean have a next-hop destination + a final destination?  Could be
done I guess.  But I'm not sure how having a routing table where you
have to look up 48-bit Ethernet addresses is all that different from
just having a standard Ethernet forwarding table.

I suppose something based on MAC-in-MAC (a la 802.1ah) could be done but
to be honest the IBoE spec that the IBTA came up with looks rather
broken for routing.

 - R.
-- 
Roland Dreier <rolandd-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

>  > A quibble about multicast - AFAIK this is unsolved. I 
> think some spec  > needs to be agreed that documents what 
> sort of multicast snooping  > operations switches need to do, 
> ie if IGMP joins imply that IBoE  > traffic for the same DMAC 
> is included in the join, or if IBoE requires  > a seperate 
> IGMP type process on its own ether-type. That would make it  
> > much clearer what to do with MGIDs.
>

It would be quite naïve to require *new* snooping functionality in Eth switches. Some switches will gracefully apply to non-ip traffic the filtering information acquired through IGMP snooping. And some will just flood non-ip MC frames within the corresponding VLAN which is benign (e.g. that is the way FIP works). A cleaner solution would be based on MMRP but that, AFAIK, is not very widely deployed so it is less practical at this stage. 
 
> I agree -- the current spec is rather broken for multicast.  
> Choosing a different ethertype and then saying that all 
> switches will just flood multicast traffic is half-baked at best.
>

It is a realistic approach. Do you claim that there are switches that will not forward the packets?
 
>  > It would be nice to at least have a plan on how to 
> integrate a  > non-link local address, if that is ever 
> necessary in future. An  > extended AH with an additional 48 
> DMAC field seems reasonable to me?
> 
> You mean have a next-hop destination + a final destination?  
> Could be done I guess.  But I'm not sure how having a routing 
> table where you have to look up 48-bit Ethernet addresses is 
> all that different from just having a standard Ethernet 
> forwarding table.

I guess Jason suggests regarding the GID as a true L3 address and using a new added L2 field for the next hop L2 address.

> 
> I suppose something based on MAC-in-MAC (a la 802.1ah) could 
> be done but to be honest the IBoE spec that the IBTA came up 
> with looks rather broken for routing.

Routing is out of the scope of the current RoCE spec.
And I do not see how .1ah would be relevant for this purpose.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Roland Dreier]

 > > I agree -- the current spec is rather broken for multicast.  
 > > Choosing a different ethertype and then saying that all 
 > > switches will just flood multicast traffic is half-baked at best.

 > It is a realistic approach. Do you claim that there are switches that will not forward the packets?

No, I think all switches will flood unknown multicast packets.  But
there is a reason that IGMP snooping was invented -- it is inefficient
(to say the least) to flood all multicast traffic.

 - R.
-- 
Roland Dreier <rolandd-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Roland Dreier]

 > No, I think all switches will flood unknown multicast packets.  But
 > there is a reason that IGMP snooping was invented -- it is inefficient
 > (to say the least) to flood all multicast traffic.

And by the way I view the fact that the IBoE spec does not say anything
at all about how to map MGIDs to Ethernet addresses as another serious
shortcoming of the spec.

 - R.
-- 
Roland Dreier <rolandd-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Alex Rosenbaum]

> No, I think all switches will flood unknown multicast packets.  But
> there is a reason that IGMP snooping was invented -- it is inefficient
> (to say the least) to flood all multicast traffic.
>
> - R.


Small correction needed regarding the multicast forwarding.
Since we are talking about IPv6 multicast groups, which translate to
33:33:xx:xx:xx:xx MAC address, the router listener notification protocol
is going to be MLD and not IGMP. Still there are switches which support
MLD forwarding to prevent the network flooding.

Alex

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: When IBoE will be merged to upstream?

[Roland Dreier]

 > Small correction needed regarding the multicast forwarding.
 > Since we are talking about IPv6 multicast groups, which translate to
 > 33:33:xx:xx:xx:xx MAC address, the router listener notification protocol
 > is going to be MLD and not IGMP. Still there are switches which support
 > MLD forwarding to prevent the network flooding.

Well as I said the mapping of IBoE MGID to Ethernet address is not
specified.  However I agree that using the same mapping as IPv6 so we
end up with 33:33:... addresses makes sense.

Yes, you are right that MLD snooping is the mechanism for switches to
discover IPv6 multicast group membership.  However for the IBoE case
there is no requirement that IPv6 multicast group membership corresponds
in any way to the IBoE multicast group membership for the interface (and
indeed as far as I can tell from the IBoE spec, there is no requirement
that any IPv6 interface be configured on an IBoE port).

Furthermore, even if an IBoE interface sends MLD messages for a given
IPv6 group, there is no requirement that a switch use the membership
information for that group to forward multicast packets with a non-IPv6
ethertype.

 - R.
-- 
Roland Dreier <rolandd-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

>  > Small correction needed regarding the multicast forwarding.
>  > Since we are talking about IPv6 multicast groups, which 
> translate to  > 33:33:xx:xx:xx:xx MAC address, the router 
> listener notification protocol  > is going to be MLD and not 
> IGMP. Still there are switches which support  > MLD 
> forwarding to prevent the network flooding.
> 
> Well as I said the mapping of IBoE MGID to Ethernet address 
> is not specified.  However I agree that using the same 
> mapping as IPv6 so we end up with 33:33:... addresses makes sense.

Agreed.

> 
> Yes, you are right that MLD snooping is the mechanism for 
> switches to discover IPv6 multicast group membership.  
> However for the IBoE case there is no requirement that IPv6 
> multicast group membership corresponds in any way to the IBoE 
> multicast group membership for the interface (and indeed as 
> far as I can tell from the IBoE spec, there is no requirement 
> that any IPv6 interface be configured on an IBoE port).
> 
> Furthermore, even if an IBoE interface sends MLD messages for a given
> IPv6 group, there is no requirement that a switch use the 
> membership information for that group to forward multicast 
> packets with a non-IPv6 ethertype.

Right. Initially there can be flooding within the VLAN. In the future we can
evolve to use a group-membership protocol when customers that care about
the efficiency drive their switch vendors to support it.

Are there any other issues that you would like us to address before updating the patches?
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: When IBoE will be merged to upstream?

[Liran Liss]

>  > But, we can't mandate an overload of the GID in a way that 
> it  > prevents its use as a true L3 address (eventually routable).
> 
> Actually I'm beginning to think that the only possible way we 
> can use the GID in IBoE is as a link-local IPv6 addresses 
> containing an Ethernet address.  Trying to hide neighbour 
> discovery or ARP below the verbs doesn't seem workable -- 
> being forced to change the locking rules we've had for the 
> past 5+ years about create_ah is just the beginning.  We get 
> further problems if a remote address should ever change and 
> I'm probably missing other issues.
> 

We believe the problems are workable. But let us stop arguing for a while and
make progress with link local addressees since we all seem to agree with that.
We can get back to non-local GIDs later

> So the best solution I can see is to declare that an IBoE GID 
> must be an
> IPv6 address coming from an EUI-64 Ethernet address for the 
> corresponding port; for MGIDs I guess we use the standard 
> IPv6 mapping to Ethernet address 33:33:xx:xx:xx:xx.
> 
> I'm not sure how we want to handle IPv4 -- presumably unicast 
> ARP can be done within the RDMA CM, which will then create a 
> DGID with the appropriate Ethernet address.  However it's not 
> clear to me whether we need a way to create IPv4 
> (01:00:5e:xx:xx:xx) multicast addresses.
> 
> Also, since there is no way to map a link-local IPv6 address 
> to a particular interface, then I guess we need a way to pass 
> in the VLAN tag to be used -- presumably we can steal some 
> other field for the 12 bits.
> (The fact that the IBoE annex does not mention VLANs or 
> 802.1q a single time is just another thing that shows how 
> rushed and incomplete it is)
> 
> With all this said, I think it means we do not need to do the 
> mapping from GID to Ethernet address in the kernel for IBoE 
> user verbs, since it is so simple -- we can simply add a 
> fairly trivial helper to libibverbs.
> 
>  - R.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html