Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

mlx5: Introduce data direct placement (DDP) over the DV API #1494

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

mlx5: Introduce data direct placement (DDP) over the DV API #1494

wants to merge 4 commits into from

Commits on Sep 3, 2024

  1. Update kernel headers 

    To commit: 06ac609a5f35 ("RDMA/mlx5: Support OOO RX WQE consumption").

Signed-off-by: Yishai Hadas <[email protected]>
    Yishai Hadas committed Sep 3, 2024
    Configuration menu
    Copy the full SHA
    fc196ad View commit details
    Browse the repository at this point in the history

  2. mlx5: Handle OOO WQE consumption in CQE generation 

    Based on IB specification, the current code assumes that WQE buffers are
consumed and CQEs are generated in-order.
The WQE in-order consumption is not guaranteed when HW has to handle
out-of-order (OOO) packets. HW may consume buffers OOO but generate CQEs
in-order.

When scatter2cqe is enabled, we must scatter the data to the correct WQE
buffer. This also applies to WR IDs. Assuming incremental WQE indexes
leads to incorrect WR IDs being returned to users.

Therefore, we need to use WQE index field from CQE to access the WQE
instead of assuming the WQE's order is the same as the CQE's and access
the next WQE in the WQ in an incremental way.

So, this is a preparation patch to support the WQE's OOO mode as will be
introduced in the next patches from the series.

Signed-off-by: Edward Srouji <[email protected]>
Signed-off-by: Yishai Hadas <[email protected]>
    @EdwardSro
    EdwardSro authored and Yishai Hadas committed Sep 3, 2024
    Configuration menu
    Copy the full SHA
    88a2a38 View commit details
    Browse the repository at this point in the history

  3. mlx5: Remove max_post assignment in create_qp() 

    In create_qp(), RQ max_post value was redundantly overwritten after
being calculated already as part of mlx5_calc_rq_size().

This assumption that max_post is equal to wqe_cnt is not necessarily
true in all cases, as will be changed in upcoming patches where the OOO
RX support feature is introduced.

The line was removed, and the calculation is now handled by
mlx5_calc_rq_size() without overwriting its calculation.

Signed-off-by: Edward Srouji <[email protected]>
Signed-off-by: Yishai Hadas <[email protected]>
    @EdwardSro
    EdwardSro authored and Yishai Hadas committed Sep 3, 2024
    Configuration menu
    Copy the full SHA
    7035809 View commit details
    Browse the repository at this point in the history

  4. mlx5: Support OOO RX WQE consumption 

    Add a new MLX5DV_QP_CREATE_OOO_DP flag, which allows WRs on the receiver
side of a QP to be consumed out-of-order (OOO).
Additionally, it permits the sender side to transmit messages without
guaranteeing arrival order on the receiver side.

When enabled, the flag ensures that the completion ordering of WRs
remains unchanged, regardless of the consumption order of Receive WRs.
RDMA Read and RDMA Atomic operations on the responder side continue to
be executed in order, while the ordering of data placement for RDMA
Write and Send operations is not guaranteed.

The MLX5DV_QP_CREATE_OOO_DP flag must be set on both the sender and
receiver sides of a QP, such as DCT and DCI, to allow the sender side to
transmit messages without guaranteeing any arrival ordering on the
receiver side.

It is optional, and its availability must be queried via the application
using mlx5dv_query_device() with a newly added
MLX5DV_CONTEXT_MASK_OOO_RECV_WRS flag.

Although enabling OOO on the QP is relevant in Init to RTR modification
stage, the relevant flag is passed by the user on QP creation.
This should be done because internally, when enabled on a QP, its RQ
buffer size is double the user requested size if it's a cyclic
implemented buffer (e.g. RC, UC, UD, etc.). This is to prevent a
scenario where WQE overwrites may happen when WQEs are consumed OOO.

If the Kernel or device does not support this feature, creating the QP
with this flag will fail.

Signed-off-by: Edward Srouji <[email protected]>
Signed-off-by: Yishai Hadas <[email protected]>
    @EdwardSro
    EdwardSro authored and Yishai Hadas committed Sep 3, 2024
    Configuration menu
    Copy the full SHA
    89369e2 View commit details
    Browse the repository at this point in the history