erdma: Elastic RDMA Adatper (ERDMA) userspace provider driver

[hz-cheng]

Hello all,

This PR introduces the Elastic RDMA Adapter (ERDMA) userspace provider driver, and the patchset for review purpose was sent to the linux-rdma mail list already [1]. The kernel driver of ERDMA can refer this link [2].

The main feature of ERDMA userspace provider includes: supports RC QP, supports RDMA Write/Send/RDMA Read/Immediate opcode in post_send, supports post_recv, and supports CQs with polling mode and event mode. Now we does not support SRQ yet.

Besides, this PR has already issued the review suggestions, including:

• Fix coding style issues
• Remove unnecessary memset().
• Remove some magic number in the code.

Thanks,
Cheng Xu

[1] https://lore.kernel.org/all/20211224065522.29734-1-chengyou@linux.alibaba.com/
[2] https://lore.kernel.org/all/20211221024858.25938-1-chengyou@linux.alibaba.com/

[hz-cheng]

Hi,

I just updated the PR, changes are:

1. Rebase the code to latest rdma-core code.
2. Format the code using clang-format, because the userspace provider has the same format issues with the former kernel patch. Now both of them are fixed.
3. Fix a double-free bug in erdma_create_cq.
4. Memset the cq buffer to zero in erdma_create_cq.

Thanks,
Cheng Xu

[hz-cheng]

I just updated the PR, changes including:

1. Sync kernel headers by kernel-headers/update script.
2. Sort function implementations in erdma_context_ops.
3. Let the initialize statements in functions be in inverted triangle order.
4. Fix bugs in mmap's return value check
5. Refactor the doorbell record allocation (because many new API can be applied, link list, bitmap, etc.,)
6. Refactor erdma_poll_cq in the similar way to the kernel's poll_cq code.
7. Remove PCIe device ID 5007 support.
8. Use the proper type (int -> uint32_t, int -> size_t and so on) in some definitions.
9. Some small changes, following the review suggestions for kernel code.

The changes with patch format can be get at here.

[hz-cheng]

Some more issues need to be fixed, so I pushed again. Changes are:

1. Put verbs_provider_erdma at the right place in libibverbs/verbs.h
2. Fix a memory leak issue in erdma_destroy_qp
3. Make some statements in erdma_create_qp shorter.
4. Update the erdma WC mapping table, and the related code in erdma_poll_cq

The changes with patch format can be get at here.

[hz-cheng]

I reviewed changed files in mana_ib, and found that erdma missed the modification need of RDMA_STATIC_PREFIX,
so I push this version.

Also add necessary information in debian pkg build.

The changes with patch format can be get at here.

[hz-cheng]

I saw that Leon had synced the kernel header to master. To avoid potential conflicts (through the CI does not reported), I rebase the code to the latest master code. This push also includes a change: using struct erdma_sge sgl[] instead of struct erdma_sge sgl[0] in definitions in erdma_hw.h.