Create a common verbs transport library

Hi All,

One of the conditions to move the hfi1 driver from staging into the normal drivers/infiniband/hw directory is to handle the code duplication in our verbs layer. This is going to be done by creating a new kmod which we will call rdmavt, for RDMA verbs transport. This will eventually live in the existing drivers/infiniband tree in a new sw directory:
drivers/infiniband/sw/vt. This new directory can serve as a home for soft roce when its ready as well.

The verbs library will start out life in drivers/staging/rdma/vt alongside hfi1. We (Intel) will push incremental patches to keep the community apprised of the development and allow for early and more continuous feedback. Once complete the plan would be to move out of staging along with hfi1.

The current verbs support in the IB core should not need to be modified, rdmavt is just another verbs provider. Drivers will not use rdmavt directly. Rather, rdmavt will use the drivers to abstract away the hardware differences. Here is a diagram of what this will look like.

  +-------+
  |Ib Core|
  +-------+
      +
      |
+------v--------+
|Verbs Transport|
+-+----------+--+
 |          |
 |          |
+-v--+     +-v--+
|qib |     |hfi1|
+----+     +----+

-Denny
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

We initially thought to implement a shared library that contains the transport logic.

However, it seems that a SW Verbs transport driver would allow better code sharing.

In fact, the VT driver would need only a single user-space driver for all "backends". Any direct HW access from user-space should be exposed by the corresponding backend driver and accessed by a different library (e.g., psm).

At a high-level, it seems that we should do as follows:

- Decide on an initial code base for VT (rxe/hfi/qib), clone it, and rename to VT

- Split the code to VT and backend and create the initial backend APIs, e.g.:

-- Send packet

-- Deliver packet (receive)

-- Attach multicast

-- Packet buffer allocation

-- Notify when more send space is available

- In parallel, prepare the backends of other drivers while enhancing VT as needed.

Do you have any preferences to the initial code base?

Do you already have some code that we can look at?

I assume by user-space driver we are talking about libverbs? We have separate libraries for ipath/qib and hfi. We should probably coalesce these into a single library but that is a separate issue. PSM is also unrelated to the work here since PSM is not verbs.

Quoted 7 lines

At a high-level, it seems that we should do as follows:

- Decide on an initial code base for VT (rxe/hfi/qib), clone it, and rename to VT

- Split the code to VT and backend and create the initial backend APIs, e.g.:

We have been planning a bit of a different approach. My thoughts are we make VT a completely new kmod. It will start out life lettings verbs calls from the core go into the drivers to do their thing, but will contain a bunch of the duplicated code that we have in hfi1/qib/ipath. The next step is to move piece by piece the rest of the verbs code.

Quoted 8 lines

-- Send packet

-- Deliver packet (receive)

-- Attach multicast

-- Packet buffer allocation

Quoted 4 lines

-- Notify when more send space is available

- In parallel, prepare the backends of other drivers while enhancing VT as needed.

Yes, we need to come up with an API, I'm not fully sure what that should look like yet, it is a work in progress.

Quoted 4 lines

Do you have any preferences to the initial code base?

Do you already have some code that we can look at?

We'll be starting out with making changes to hfi1 and qib to follow shortly behind. No code just yet, but I should have something to post as an RFC very soon (in the next two weeks).

Thanks

-Denny

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

Quoted 5 lines

I assume by user-space driver we are talking about libverbs? We have separate libraries for ipath/qib and hfi. We should probably coalesce these into a single library but that is a separate issue. PSM is also unrelated to the work here since PSM is not verbs.

Not libibverbs but lib<provider> (e.g. libmlx4, libqib,...), like for any other kernel verbs provider.
In other wordsm if the name of the SW verbs provider is called VT (verbs transport) than the library we need to develop will be libVT. However, the challenge here is larger than for other providers due to different style of backends.
Take for example the ibv_post_send() verb. For some totally incapable kernel bypass backends this requires that this verb go down to the kernel (Soft RoCE for example) but for others it needs to put the WQE directly on the HW in some HW specific manner.
This needs some architectural effort.

Quoted 9 lines

Quoted 7 lines

At a high-level, it seems that we should do as follows:

- Decide on an initial code base for VT (rxe/hfi/qib), clone it, and rename to VT

- Split the code to VT and backend and create the initial backend APIs, e.g.:

Quoted 6 lines

We have been planning a bit of a different approach. My thoughts are we make VT a completely new kmod. It will start out life lettings verbs calls from the core go into the drivers to do their thing, but will contain a bunch of the duplicated code that we have in hfi1/qib/ipath. The next step is to move piece by piece the rest of the verbs code.

We agree with the kmod approach. However
1. Until we get rid completely from code duplication I don't think we can get out from being in staging phase so the initial work better be complete. Don't you think so?

2. Soft RoCE is a candidate driver (backend) for this framework and we would like to make sure that the new VT provider meets out needs.

Quoted 10 lines

Quoted 10 lines

-- Send packet

-- Deliver packet (receive)

-- Attach multicast

-- Packet buffer allocation

-- Notify when more send space is available

Quoted 6 lines

Quoted 2 lines

- In parallel, prepare the backends of other drivers while enhancing VT as needed.

Yes, we need to come up with an API, I'm not fully sure what that should look like yet, it is a work in progress.

I think that this will be a good place to start before implementation. Can you please share API between backends and VT so we can comment on?

Quoted 5 lines

Quoted 3 lines

Do you have any preferences to the initial code base?

Do you already have some code that we can look at?

Quoted 3 lines

We'll be starting out with making changes to hfi1 and qib to follow shortly behind. No code just yet, but I should have something to post as an RFC very soon (in the next two weeks).

I suggest that implementation will start after we close the arch. Like I said we want to make sure that Soft RoCE requirements are met in the design.
I strongly recommend that we figure out a mechanism to allows us to cooperate. I hope that you see it this way too,

Quoted 8 lines · Moni Shoua

Quoted 8 lines

I assume by user-space driver we are talking about libverbs? We have separate libraries for ipath/qib and hfi. We should probably coalesce these
into a single library but that is a separate issue. PSM is also unrelated to
the work here since PSM is not verbs.

Quoted 11 lines

Not libibverbs but lib<provider> (e.g. libmlx4, libqib,...), like for any other kernel verbs provider.
In other wordsm if the name of the SW verbs provider is called VT (verbs transport) than the library we need to develop will be libVT. However, the challenge here is larger than for other providers due to different style of backends.
Take for example the ibv_post_send() verb. For some totally incapable kernel bypass backends this requires that this verb go down to the kernel (Soft RoCE for example) but for others it needs to put the WQE directly on the HW in some HW specific manner.
This needs some architectural effort.

Ok, I'm with you now. Yes, this is an issue and agree that it will take some architectural work.

Quoted 20 lines

Quoted 10 lines

Quoted 7 lines

At a high-level, it seems that we should do as follows:

- Decide on an initial code base for VT (rxe/hfi/qib), clone it, and rename to VT

- Split the code to VT and backend and create the initial backend APIs, e.g.:

Quoted 10 lines

We have been planning a bit of a different approach. My thoughts are we make
VT a completely new kmod. It will start out life lettings verbs calls from
the core go into the drivers to do their thing, but will contain a bunch of
the duplicated code that we have in hfi1/qib/ipath. The next step is to move
piece by piece the rest of the verbs code.

Quoted 4 lines

We agree with the kmod approach. However
1. Until we get rid completely from code duplication I don't think we can get out from being in staging phase so the initial work better be complete. Don't you think so?

Yes, I completely agree. Code duplication issue should be solved before we move out of staging.
I could have been more clear on my intentions here. The initial set of patches that we will submit for RFC will not be complete. That is not meant to get us out of staging and satisfy the hfi1 TODO item. It is a stepping stone to get things moving and start getting feedback from the community. We want to engage early on. So the sooner we have something tangible that we can all look at the better.

Quoted 2 lines

2. Soft RoCE is a candidate driver (backend) for this framework and we would like to make sure that the new VT provider meets out needs.

Yes, while my focus is of course hfi1 and qib, we certainly realize there is a larger community here that has a vested interest, and this is why we want to engage early.

Quoted 18 lines

Quoted 17 lines

Quoted 4 lines

-- Send packet

-- Deliver packet (receive)

Quoted 8 lines

-- Attach multicast

-- Packet buffer allocation

-- Notify when more send space is available

- In parallel, prepare the backends of other drivers while enhancing VT as needed.

Yes, we need to come up with an API, I'm not fully sure what that should look like yet, it is a work in progress.

Quoted 2 lines

I think that this will be a good place to start before implementation. Can you please share API between backends and VT so we can comment on?

This will be coming very soon, and we anticipate needing to make changes so that all drivers are best supported. Don't think of it as, "this is what we are going to do". Think of it as, "this is what we are thinking, does it work for you?".

Quoted 12 lines

Quoted 11 lines

Quoted 3 lines

Do you have any preferences to the initial code base?

Do you already have some code that we can look at?

We'll be starting out with making changes to hfi1 and qib to follow shortly
behind. No code just yet, but I should have something to post as an RFC very
soon (in the next two weeks).

Quoted 6 lines

I suggest that implementation will start after we close the arch. Like I said we want to make sure that Soft RoCE requirements are met in the design.
I strongly recommend that we figure out a mechanism to allows us to cooperate.
I hope that you see it this way too,

Yes, I think by doing the development in the open, and sharing code early is our best way to do this. We can work together to make sure everyone's needs are met.

I'm thinking perhaps we'll post some stuff to GitHub and you folks can take a look. We can do some back and forth before even posting patches to the list. I'll check into the logistics and share the details soon?

From my experience this can be very useful to align all on the

Quoted 3 lines

I'm thinking perhaps we'll post some stuff to GitHub and you folks can take a look. We can do some back and forth before even posting patches to the list. I'll check into the logistics and share the details soon?

This will be great. Just send us the ref.
If you'd like we can have an online meeting to have an initial talk. From my experience this can be very useful to align all on the subject.

Quoted 11 lines · Moni Shoua

Quoted 5 lines

I'm thinking perhaps we'll post some stuff to GitHub and you folks can take
a look. We can do some back and forth before even posting patches to the list. I'll check into the logistics and share the details soon?

This will be great. Just send us the ref.
If you'd like we can have an online meeting to have an initial talk. From my experience this can be very useful to align all on the subject.

I have created a branch on a GitHub repo. You can find it at:
https://github.com/weiny2/linux-kernel/tree/rdmavt

This is a starting point to get the ball rolling. I have a few more patches to further refine the skeleton rdmavt module. I should be able to post those early next week. At that point we will have a more clear idea of what the API between the drivers and rdmavt will need to look like for hfi1 and qib. We will look for your feedback on how soft roce is going to fit in as well.

-Denny

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html