freelist: separate out metadata from user data

[rauteric]

Updates the interface of freelist to return an elem struct that contains
a buffer pointer and (optionally) memory registration info. Also updates
sendrecv, rdma, and scheduler codes to use the new interface.

This separation paves the way for storing freelist user data in GPU
memory.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[bwbarrett]

I think having two different interfaces is a mistake. We should have one way to allocate / free freelist entries, whether or not they're registered.

[rauteric]

I think having two different interfaces is a mistake. We should have one way to allocate / free freelist entries, whether or not they're registered.

Ok. My initial plan was to have entry_alloc() return both the buffer and an optional memory registration, but I couldn't make that work because we need to keep track of the entry structures, which are allocated separately.

So basically, we're going to end up with the "mr" interface variant (that returns an elem_t) in this PR being used for both the simple and complex freelists, That is,

static inline struct nccl_ofi_freelist_elem_t *nccl_ofi_freelist_entry_alloc
					       (nccl_ofi_freelist_t *freelist)

and

static inline void nccl_ofi_freelist_entry_free(nccl_ofi_freelist_t *freelist,
						   struct nccl_ofi_freelist_elem_t *entry)

We'll need to update both SENDRECV and RDMA protocols to use this interface.

[rauteric]

Rebased on master

[rauteric]

Addressed feedback from Brian

[rauteric]

Rebase on master

[rauteric]

Rebase on master

[rauteric]

bot:aws:retest

From CI:

Cloning into 'aws-ofi-nccl'...
fatal: unable to access 'https://github.com/aws/aws-ofi-nccl.git/': The requested URL returned error: 502

Not sure what happened there; trying again.

Changes made

[bwbarrett]

In the end, I really hate this interface and think I was wrong about having the same interface for both use cases. But I'd rather do that in C++ where we just use a builtin PMR interface for freelists, rather than continuing to iterate here. So let's ship it and move on.

[aws-nslick]

did not test myself, but looks good to me.