Fix unix socket half-close handling in Sockets#60961
Fix unix socket half-close handling in Sockets#60961abslock128w wants to merge 4 commits intoJuliaLang:masterfrom
Conversation
Added tests for Unix domain socket half-close behavior.
Removed unnecessary checks for TTY and PipeEndpoint in EOF handling.
Handle stream closure for PipeEndpoint in UV_EOF case.
|
Ref: #60959 @abslock128w apologies if this is unfounded, but I get the sense that LLMs may have been used in formulating this PR. If this is the case, we would appreciate it if you would declare this. |
yeah I did use an LLM a bit while working on this, used it to understand the already writtne code |
| stream_unknown_type = @handle_as handle LibuvStream | ||
| nrequested = ccall(:jl_uv_buf_len, Csize_t, (Ptr{Cvoid},), buf) | ||
|
|
||
| function readcb_specialized(stream::LibuvStream, nread::Int, nrequested::UInt) |
There was a problem hiding this comment.
Enabling "Hide whitespace" in the GitHub diff viewer, it seems to me that the changes to this function amount to:
- wrapping most of the code into a
try/finally/endwithunlockin the finalize block -- that seems plausible; - inserting some empty lines, wrapping an overly long line -- I am neutral on this;
- removing or reducing a lot of comments -- that part seems less plausible: why do we want to remove comments or reduce them? Were they incorrect?
There was a problem hiding this comment.
Hey there,
Thank you for pointing out the mistake!
I am new to this whole open source contribution thing...
And about the comment part I never really use comments that's wrong on my part ( i know its a good practice to do so,
will try to keep that in mind as I move forward! )
| function reseteof(x::TTY) | ||
| iolock_begin() | ||
| if x.status == StatusEOF | ||
| x.status = StatusOpen | ||
| nothing | ||
| end | ||
| iolock_end() |
There was a problem hiding this comment.
Why is reseteof(x::TTY) being removed?
There was a problem hiding this comment.
I mistakenly removed the TTY fucntion
This will probably brake the interactive input...
Will add in the next commit.
Again thank you for pointing it out!
5 participants
Summary
Fix incorrect handling of unix domain socket half-close where a
PipeEndpointwas fully closed onUV_EOF, even when the write sidewas still usable.
Problem
When libuv reports
UV_EOF, Julia currently finalizes certainLibuvStreams (notablyPipeEndpoint) too aggressively.This breaks valid half-close semantics, where the read side reaches EOF but the socket remains writable.
This behavior causes issues in clients that rely on
half-closed unix sockets, forcing workarounds such as splitting stdio
into multiple sockets.
Solution
Update
uv_readcbto preserve the stream when the underlying handleremains writable, allowing correct half-close behavior while still
closing truly unusable streams.
The change aligns Julia’s behavior with libuv semantics and POSIX
expectations.