Make `waiting_fd` behaviour per-IO. #13127

ioquatix · 2025-04-18T01:05:57Z

See #5384 for full context.

Copilot

Pull Request Overview

This PR refactors the “waiting_fd” behavior so that waiting and blocking operations are managed per‑IO rather than via a global list. Key changes include:

Removing the global waiting_fd list initialization and associated memory‐size functions (e.g. in vm_core.h and vm.c).
Replacing the waiting_fd structure and its helper functions in thread.c with per‑IO blocking operation nodes.
Eliminating the thread_fd close test file, consistent with the new per‑IO design.

Reviewed Changes

Copilot reviewed 12 out of 13 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
vm_core.h	Removed initialization of the global waiting_fds list.
vm.c	Removed calls to rb_vm_memsize_waiting_fds in memory size reporting.
thread.c	Commented out and removed waiting_fd structures/functions in favor of per‑IO nodes.
test/-ext-/thread_fd/test_thread_fd_close.rb	Removed outdated test file no longer applicable under the new design.
io.c	Updated blocking call APIs to accept an IO pointer instead of a raw file descriptor.
internal/*	Updated function declarations and struct fields to reflect per‑IO blocking operations.

Files not reviewed (1)

ext/-test-/thread_fd/depend: Language not supported

Comments suppressed due to low confidence (2)

vm.c:3229

Verify that the new per‑IO blocking operations are correctly accounted for in the overall memory size calculations now that the global waiting_fd memory size contribution has been removed.

-        rb_vm_memsize_waiting_fds(&vm->waiting_fds) +

test/-ext-/thread_fd/test_thread_fd_close.rb:1

The removal of the thread_fd close test file reduces test coverage for FD closing behavior; please ensure that the per‑IO waiting/blocking operations introduced in this PR are adequately covered by other tests.

Entire file removed

thread.c

internal/io.h

thread.c

ko1 · 2025-05-02T05:38:03Z

thread.c

-    if (rb_notify_fd_close(fd, &busy)) {
-        rb_notify_fd_close_wait(&busy);
-    }
+    /* No-op */


We found only one practical example in gems:

https://github.com/vitoshalabs/ruby-pcap/blob/master/ext/pcap/Pcap.c#L149

https://gist.github.com/ko1/81aa42d1bee240a8d8f8d431d165e0c0

not sure we need to consider about that.

Thanks for your investigation.

From your linked example, is it safe to call the current implementation of rb_thread_fd_close from within GC?

I think a better solution is to take a reference to an auto close IO instance. It is an easy fix and then the code can be updated to use rb_io_wait etc.

As an interim solution, what about issuing a warning and adding documentation about how to fix it?

Simplified code (as I understand):

io1 = open(...) io2 = IO.for_fd(io1.fileno) # https://github.com/vitoshalabs/ruby-pcap/blob/master/ext/pcap/Pcap.c#L455C94-L455C95 # maybe io2 is used io2.read... rb_notify_fd_close_wait(io1.fileno) # wait for closing interrupting io2

I think using RUBY_IO_MODE_EXTERNAL and rb_io_close is a much better model for us to expose as a public interface:

io1 = open(...) io2 = IO.for_fd(io1.fileno, FMODE_EXTERNAL) # maybe io2 is used io2.read... rb_io_close(io2)

I've updated the documentation to recommend this.

Copilot

Pull Request Overview

This pull request refactors how waiting file descriptors are managed by removing the global waiting_fds list and transitioning to per-IO blocking operation mechanisms. Key changes include eliminating waiting_fds initialization and related memsize computations, overhauling thread and IO blocking logic in thread.c and io.c, and updating tests and deprecated extensions accordingly.

Reviewed Changes

Copilot reviewed 12 out of 14 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
vm_core.h	Removes the initialization of the global waiting_fds list to support a per-IO approach.
vm.c	Eliminates the waiting_fds memsize function and its usage in the overall VM memsize calculation.
thread.c	Removes the waiting_fd structure and associated functions, replacing them with a new blocking operation release.
test/ruby/test_io.rb	Adjusts the range in a loop to properly limit file descriptor usage.
io.c	Updates finalization logic by removing the busy parameter and replacing calls with the new IO close-wait functions.
internal/thread.h & internal/io.h	Remove deprecated APIs and type definitions related to waiting_fd.
test/-ext-/thread_fd/* & ext/-test-/thread_fd/extconf.rb	Remove deprecated thread_fd extension tests and associated build files.

Files not reviewed (2)

common.mk: Language not supported
ext/-test-/thread_fd/depend: Language not supported

Comments suppressed due to low confidence (5)

vm_core.h:1870

The removal of the waiting_fds initialization appears intentional for per-IO handling. Please verify that all parts of the codebase referencing waiting_fds have been updated to use the new per-IO blocking operations.

ccan_list_head_init(&vm->waiting_fds);

vm.c:3196

The removal of the rb_vm_memsize_waiting_fds declaration and its usage in the memsize calculation must be coordinated with the new per-IO mechanism. Confirm that the overall memsize reporting remains accurate after these changes.

size_t rb_vm_memsize_waiting_fds(struct ccan_list_head *waiting_fds); // thread.c

thread.c:151

The waiting_fd structure and its associated functions have been removed in favor of a new blocking operation approach. Please ensure that the new mechanism (using rb_io_blocking_operation_release and related logic) correctly handles concurrency and interrupt processing.

struct waiting_fd {

test/ruby/test_io.rb:3829

Changing the range from (0..fd_setsize+1) to (0...fd_setsize) ensures that file descriptors do not exceed the set size. Confirm that this correction aligns with the intended limits of the fd set.

(0...fd_setsize).map {|i|

io.c:5520

The updated fptr_finalize_flush signature (removing the busy parameter) reflects the new IO close logic. Verify that the new function correctly waits for blocking operations to complete before closing the IO and that all call sites are updated accordingly.

fptr_finalize_flush(rb_io_t *fptr, int noraise, int keepgvl)

ioquatix · 2025-05-13T07:49:31Z

@ko1 requested the following changes:

Fix common.mk dependencies.
Remove verbose warning details.
Update NEWS.
Remove mutex synchronisation around thread_io_close_notify_all.

ntkme · 2025-05-14T03:28:56Z

@ioquatix I think this change is causing segmentation fault in my nightly CI run against ruby head:

https://github.com/sass-contrib/sass-embedded-host-ruby/actions/runs/15010437636/job/42178042427

To summarize the use case: sass-embedded spawns the dart-sass child process and communicates with child process’ stdio. When parent ruby process is forked, the stdio file descriptors of the child process are duplicated. However, because this gem does not support share the child process across multiple parent process, I need to close the stdio file descriptors in the forked ruby process. Right now it crashes when closing the duplicated file descriptors right after fork in the forked ruby process.

ioquatix · 2025-05-14T05:31:39Z

Thanks for your report, I will take a look.

ioquatix · 2025-05-14T08:03:50Z

I was able to reproduce the issue, I'll work on a fix.

ioquatix · 2025-05-14T11:19:52Z

Okay, I found the root cause of the issue.

ioquatix · 2025-05-15T01:06:32Z

Here is the reproduction:

r, w = IO.pipe

thread = Thread.new do
  r.read
end

Thread.pass until thread.status == "sleep"

fork do
  r.close
end

w.close

Process.wait
thread.join

ioquatix · 2025-05-15T02:07:52Z

Proposed fix: #13343

ioquatix requested review from ko1, akr, nobu and Copilot April 18, 2025 01:06

Copilot AI reviewed Apr 18, 2025

View reviewed changes

thread.c Outdated Show resolved Hide resolved

ioquatix marked this pull request as draft April 18, 2025 01:09

ioquatix force-pushed the io-close-isolated branch 3 times, most recently from 9cba6ac to 5d85916 Compare April 18, 2025 01:25

ioquatix mentioned this pull request Apr 18, 2025

Improve IO#close performance and semantics. #5384

Closed

ioquatix force-pushed the io-close-isolated branch 3 times, most recently from aa93db0 to a0aa5d2 Compare April 18, 2025 01:51

This comment has been minimized.

Sign in to view

ioquatix force-pushed the io-close-isolated branch 12 times, most recently from 0ce6a12 to bf8fd7d Compare April 18, 2025 11:05

ioquatix mentioned this pull request Apr 18, 2025

Ensure struct rb_io is passed through to thread.c. #13134

Merged

ioquatix force-pushed the io-close-isolated branch from bf8fd7d to cbc4843 Compare April 19, 2025 01:03

ioquatix mentioned this pull request Apr 19, 2025

Tidy up rb_io_fptr_finalize. #13136

Merged

ioquatix force-pushed the io-close-isolated branch from cbc4843 to 3d742b1 Compare April 19, 2025 10:22

ko1 reviewed May 2, 2025

View reviewed changes

internal/io.h Outdated Show resolved Hide resolved

ko1 reviewed May 2, 2025

View reviewed changes

thread.c Outdated Show resolved Hide resolved

ko1 reviewed May 2, 2025

View reviewed changes

ioquatix force-pushed the io-close-isolated branch from 402685f to cbe7c95 Compare May 3, 2025 07:47

ioquatix requested review from ko1 and Copilot May 3, 2025 10:47

Copilot AI reviewed May 3, 2025

View reviewed changes

ioquatix added 7 commits May 13, 2025 16:33

Make waiting_fd behaviour per-IO.

9a78b14

update io_reopen implementation.

63266c2

Remove dependency on vm_core.h.

73f758f

Rename rb_thread_io_close -> rb_thread_io_close_interrupt.

044128e

Add warning to rb_thread_fd_close.

e7db9a1

Updated documentation.

dec545c

Update dependencies common.mk.

c6d1db9

ioquatix force-pushed the io-close-isolated branch from 345c156 to c6d1db9 Compare May 13, 2025 07:40

ioquatix added 2 commits May 13, 2025 16:44

Unnecessary change.

540c05a

Remove elaborate warning.

ae6feb6

Remove mutex synchronization around thread_io_close_notify_all.

8993485

ioquatix force-pushed the io-close-isolated branch from 6fbf43e to 8993485 Compare May 13, 2025 07:54

ioquatix added 2 commits May 13, 2025 17:07

Add news entry.

05bb3c2

Updatd news entry.

ef61fb4

ioquatix merged commit 425fa0a into ruby:master May 13, 2025
81 checks passed

ioquatix deleted the io-close-isolated branch May 13, 2025 10:02

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Make `waiting_fd` behaviour per-IO. #13127

Make `waiting_fd` behaviour per-IO. #13127

ioquatix commented Apr 18, 2025

Copilot AI left a comment

This comment has been minimized.

ko1 May 2, 2025

ioquatix May 2, 2025

ko1 May 2, 2025

ioquatix May 2, 2025 •

edited

Loading

Copilot AI left a comment

ioquatix commented May 13, 2025 •

edited

Loading

ntkme commented May 14, 2025 •

edited

Loading

ioquatix commented May 14, 2025

ioquatix commented May 14, 2025

ioquatix commented May 14, 2025

ioquatix commented May 15, 2025

ioquatix commented May 15, 2025

Make waiting_fd behaviour per-IO. #13127

Make waiting_fd behaviour per-IO. #13127

Conversation

ioquatix commented Apr 18, 2025

Copilot AI left a comment

Choose a reason for hiding this comment

Pull Request Overview

Reviewed Changes

This comment has been minimized.

ko1 May 2, 2025

Choose a reason for hiding this comment

ioquatix May 2, 2025

Choose a reason for hiding this comment

ko1 May 2, 2025

Choose a reason for hiding this comment

ioquatix May 2, 2025 • edited Loading

Choose a reason for hiding this comment

Copilot AI left a comment

Choose a reason for hiding this comment

Pull Request Overview

Reviewed Changes

ioquatix commented May 13, 2025 • edited Loading

ntkme commented May 14, 2025 • edited Loading

ioquatix commented May 14, 2025

ioquatix commented May 14, 2025

ioquatix commented May 14, 2025

ioquatix commented May 15, 2025

ioquatix commented May 15, 2025

Make `waiting_fd` behaviour per-IO. #13127

Make `waiting_fd` behaviour per-IO. #13127

ioquatix May 2, 2025 •

edited

Loading

ioquatix commented May 13, 2025 •

edited

Loading

ntkme commented May 14, 2025 •

edited

Loading