Skip to content

feature: io-uring zerocopy send support #602

Description

@AsakuraMizu

References

Notes

  1. This may also be related to Tracking: multishot opcode support #104, because these opcodes "might result in two completion queue entries", which may also require changes to the semantics of the current OpCode and related stuffs.
  2. There is also zero copy rx but its support appears to be very limited.

Activity

  1. changed the title [-]feature request: io-uring zerocopy send support[/-] [+]feature: io-uring zerocopy send support[/+] on Jan 1, 2026
  2. Berrysoft commented on Jan 5, 2026

    @Berrysoft
    Member

    I think that it meets the same problem as #104 . It's easier to design the high-level APIs, but hard to write the driver.

    For send_zc, I think it's necessary to wait for the second cqe before return.

  3. George-Miao commented on Jan 5, 2026

    @George-Miao
    Member

    We need to change how the driver work. It need to be able to take n CQE per SQE instead of only 1. This will also solve multishot.

  4. AsakuraMizu commented on Jan 5, 2026

    @AsakuraMizu
    MemberAuthor

    It's easier to design the high-level APIs

    Well, not actually; I'm thinking of making something that returns a BufResult<usize, impl Future<Output = T>> 🤣

  5. Berrysoft commented on Jan 5, 2026

    @Berrysoft
    Member

    I have a similar idea.

    But eventually you'll find that such a safe API is not that convenient as the unsafe one. If you use SendZc in C, you can carefully deal with part of the buffer that is not consumed by the kernel. However, you cannot do so in safe Rust.

  6. self-assigned this
    on Mar 11, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions