Skip to content

Fix nonblocking response write handling and backpressure #71

Open

Description

@AlexJuca

Context

FKVS uses nonblocking sockets and buffers outgoing responses in client_t.wbuf, but the event loops currently watch read readiness only. If send() returns EAGAIN or only writes part of a response, bytes can remain in wbuf without any write-readiness registration to flush them later.

Relevant code:

  • src/commands/common/command_registry.c (wbuf_append, wbuf_flush, direct send() fallback)
  • src/networking/networking.c (try_process_frames flushes once after reads)
  • src/io/event_dispatcher_epoll.c
  • src/io/event_dispatcher_kqueue.c
  • src/io/event_dispatcher_io_uring.c

Scope

  • Track pending response bytes per client.
  • Register write readiness (EPOLLOUT/EVFILT_WRITE or io_uring send operations) when wbuf_used > 0.
  • Disable write readiness once the buffer is fully flushed.
  • Remove or make safe the direct send() fallback in wbuf_append; all writes must handle partial sends.
  • Decide how to handle clients that stay backpressured too long.

Acceptance criteria

  • Large responses and pipelined command bursts do not stall under receiver backpressure.
  • Partial writes are retried without response corruption or reordering.
  • Tests cover forced partial writes / EAGAIN behavior.
  • No unchecked send() path remains for framed responses.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions