Make Abort() interrupt the libcurl and NSURLSession transports by bkaradzic-microsoft · Pull Request #34 · BabylonJS/UrlLib

bkaradzic-microsoft · 2026-06-16T23:53:36Z

Makes UrlRequest::Abort() actually interrupt the libcurl and NSURLSession transports. Today Abort() cancels m_cancellationSource, but only the Windows backend observes it (its WinRT continuations are guarded by .then(m_cancellationSource)); the curl and NSURLSession backends ignore it, so an in-flight request blocks until the transport's own timeout and a consumer's cancellation can't actually stop the work. Flagged as a follow-up in #31.

It's the transport-side prerequisite for fetch AbortSignal support in JsRuntimeHost (BabylonJS/JsRuntimeHost#196): the polyfill there wires init.signal to UrlRequest::Abort(), which currently no-ops on Linux/Apple.

Changes

libcurl: drive the transfer through the multi interface (curl_multi_perform + curl_multi_poll with a bounded 100 ms timeout) instead of curl_easy_perform, re-checking m_cancellationSource each iteration. curl_easy_perform blocks in an internal poll that, against a peer which accepts but never responds, can wait indefinitely without invoking any callback; the bounded poll bounds abort latency to ~100 ms regardless of peer activity. A cancelled transfer reports CURLE_ABORTED_BY_CALLBACK.
NSURLSession: register a cancellation listener that [task cancel]s the in-flight data task; its completion handler then fires with NSURLErrorCancelled. The task is captured weakly (no retention past completion; a late Abort() no-ops) and the ticket is emplace()d because arcana::cancellation::ticket is a move-only final_action.
Test: AbortInterruptsInFlightRequest, backed by an offline-deterministic HangingServer (a loopback listener that accepts but never responds). Skipped on the Windows backend, which observes Abort() but doesn't populate the transport-error accessors.

Validation

Built and ran UrlLibTests on Win32 / macOS / Linux (gcc + clang) via a fork CI mirror (this branch combined with #32's workflows) — all four jobs green. The abort test runs and passes in 301 ms on Ubuntu (curl) and exercises NSURLErrorCancelled on macOS (NSURLSession):

[ RUN      ] UrlRequestErrorReporting.AbortInterruptsInFlightRequest
[       OK ] UrlRequestErrorReporting.AbortInterruptsInFlightRequest (301 ms)
[  PASSED  ] 7 tests.

This validation caught and fixed two bugs that Win32-only building had hidden: a macOS-only compile error (std::optional<ticket>::operator= on a move-only final_action → emplace()), and a Linux hang (the original progress-callback abort blocked against an idle peer → reworked to the multi interface; the HangingServer fixture also deadlocked its own teardown because close() doesn't wake a blocked accept() on Linux → shutdown() first).

UrlLib main has no CI yet (the workflows are in #32, still open), so this PR doesn't get automated in-repo runs — the fork mirror is the evidence until #32 lands.

Scope / follow-ups

AbortSignal.timeout() and the JSC/JSI rejection trackers are unrelated JsRuntimeHost follow-ups.

Related: #31, #32, BabylonJS/JsRuntimeHost#196.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

Copilot

Pull request overview

This PR makes UrlRequest::Abort() actually interrupt in-flight requests on the libcurl (Unix) and NSURLSession (Apple) backends, aligning abort behavior with the existing Windows backend and enabling upstream AbortSignal plumbing to work cross-platform.

Changes:

Unix/libcurl: add a CURLOPT_XFERINFOFUNCTION progress callback that aborts curl_easy_perform() when the request cancellation source is cancelled.
Apple/NSURLSession: register a cancellation listener that cancels the active NSURLSessionDataTask.
Tests: add an offline-deterministic HangingServer and an AbortInterruptsInFlightRequest test asserting prompt interruption and transport error details.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
Tests/UrlRequestErrorReporting.cpp	Adds `HangingServer` fixture and a new abort-interrupts-in-flight test case.
Source/UrlRequest_Unix.cpp	Wires `Abort()` into libcurl via the transfer progress callback and adds the corresponding stable error symbol.
Source/UrlRequest_Apple.mm	Wires `Abort()` into `NSURLSession` by cancelling the in-flight data task via a cancellation listener.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

bkaradzic-microsoft · 2026-06-18T17:20:50Z

+            // Observe Abort(): NSURLSession runs the request asynchronously and does not watch
+            // m_cancellationSource on its own. Cancelling the task makes its completion handler fire
+            // with NSURLErrorCancelled (recorded as the transport error). The listener fires
+            // synchronously if the request was already aborted. The ticket is reset on each send and
+            // released before m_cancellationSource (a base member) is destroyed.
+            m_cancellationTicket = m_cancellationSource.add_listener([task]() {
+                [task cancel];
+            });


Good catch — fixed in ace3224. The listener now captures the task __weak and loads a strong ref before -cancel: once the request completes NSURLSession releases the task, the weak ref reads nil, and the cancel is a safe no-op, so there's no retention past completion and a late Abort() won't message a finished task.

bghgary

[Reviewed by Copilot on behalf of @bghgary]

Have the curl and NSURLSession paths been exercised locally? No CI covers them yet (#32 is still open), and the description notes only Win32 was built — where the abort test skips.

bghgary · 2026-06-18T22:02:12Z

+                // worker thread while Abort() is called from another.
+                curl_check(curl_easy_setopt(m_curl, CURLOPT_NOPROGRESS, 0L));
+                curl_check(curl_easy_setopt(m_curl, CURLOPT_XFERINFODATA, &m_cancellationSource));
+                curl_check(curl_easy_setopt(m_curl, CURLOPT_XFERINFOFUNCTION,


This idle-tick abort relies on the easy interface; the multi interface does not call the progress fn during idle periods, so a future migration would silently break abort against a hanging/idle peer. Worth noting that dependency in the comment.

Switched to the libcurl multi interface to remove this dependency entirely. The transfer now runs through curl_multi_perform + curl_multi_poll with a bounded 100ms timeout, re-checking cancelled() each iteration, so abort latency is ~100ms regardless of whether the peer is idle -- no reliance on the easy interface ticking a progress callback. (The progress-callback + socket-shutdown approaches both empirically hung against an idle peer in CI; the multi loop is deterministic.)

UrlRequest::Abort() cancels m_cancellationSource, but only the Windows backend observed it (its WinRT continuations are guarded by .then(m_cancellationSource)). The libcurl and NSURLSession backends ignored it, so an in-flight request blocked until the transport's own timeout -- and a consumer's cancellation (e.g. fetch's AbortSignal in JsRuntimeHost) could not actually stop the work. Noted as a follow-up in BabylonJS#31. - libcurl: drive the transfer through the multi interface (curl_multi_perform + curl_multi_poll with a bounded 100ms timeout) instead of curl_easy_perform, re-checking m_cancellationSource each iteration. curl_easy_perform blocks in an internal poll that, against a peer which accepts but never responds, can wait indefinitely without invoking any callback; the bounded poll bounds abort latency to ~100ms regardless of peer activity. A cancelled transfer reports CURLE_ABORTED_BY_CALLBACK (added to the stable-symbol map). - NSURLSession: register a cancellation listener that [task cancel]s the in-flight data task; its completion handler then fires with NSURLErrorCancelled. The task is captured weakly so the listener does not keep a finished task alive (and a late Abort() no-ops), and the ticket is emplace()d because arcana::cancellation::ticket is a move-only final_action whose assignment operators are deleted. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Verifies UrlRequest::Abort() promptly interrupts a request that is connected but awaiting a response, rather than waiting for the transport's own timeout. Backed by a loopback HangingServer fixture that accepts connections but never responds; the test sends a request, aborts mid-flight, and asserts the promise settles quickly (~300ms) with a transport error (NSURLErrorCancelled on Apple, a curl: error on Linux). Skipped on the Windows backend, which observes Abort() but does not populate the transport-error accessors. The fixture shutdown()s its listening socket before close() in teardown: on Linux close() in another thread does not interrupt a blocked accept(), so the accept thread would never exit and ~HangingServer would deadlock in join(). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

bkaradzic-microsoft · 2026-06-19T00:06:27Z

Good call to push on this — they hadn't been, and validating them caught two real bugs. I ran the curl and NSURLSession paths on a fork CI mirror (this branch combined with #32's workflows) so the Linux and macOS jobs actually build and run UrlLibTests. All four jobs green (Win32 / macOS / Ubuntu gcc+clang).

What the validation caught and fixed:

macOS compile error: std::optional<arcana::cancellation::ticket>::operator= doesn't compile (the ticket is a move-only final_action); now uses emplace(). UrlRequest_Apple.mm isn't compiled on Win32, so this only surfaced on a real macOS build.
Linux hang: the original progress-callback abort (and a socket-shutdown variant) hung against a peer that accepts but never responds — curl_easy_perform's internal poll blocks without ticking the callback. Reworked to drive the easy handle through the multi interface with a bounded 100ms poll, re-checking cancelled() each iteration (your earlier comment's direction). Also fixed the test's HangingServer fixture, which deadlocked its own teardown on Linux (close() doesn't wake a blocked accept() there; shutdown() does).

Evidence — the abort test now runs and passes in 301ms on Ubuntu (it previously ran to the job timeout):

[ RUN      ] UrlRequestErrorReporting.AbortInterruptsInFlightRequest
[       OK ] UrlRequestErrorReporting.AbortInterruptsInFlightRequest (301 ms)
[  PASSED  ] 7 tests.

macOS exercises the NSURLSession path (NSURLErrorCancelled) and is green as well.

I've also tidied the branch to two coherent commits (transport change + test). This PR still depends on #32 landing for in-repo CI; until then the fork mirror is the evidence.

bkaradzic-microsoft requested a review from bghgary June 16, 2026 23:53

bkaradzic-microsoft marked this pull request as ready for review June 18, 2026 15:12

Copilot AI review requested due to automatic review settings June 18, 2026 15:12

Copilot started reviewing on behalf of bkaradzic-microsoft June 18, 2026 15:12 View session

Copilot AI reviewed Jun 18, 2026

View reviewed changes

bghgary reviewed Jun 18, 2026

View reviewed changes

Copilot AI added 2 commits June 18, 2026 17:05

bkaradzic-microsoft force-pushed the urllib-abort-backends branch from 4abf687 to fa80a75 Compare June 19, 2026 00:05

bkaradzic-microsoft requested a review from bghgary June 19, 2026 00:15

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Make Abort() interrupt the libcurl and NSURLSession transports#34

Make Abort() interrupt the libcurl and NSURLSession transports#34
bkaradzic-microsoft wants to merge 2 commits into
BabylonJS:mainfrom
bkaradzic-microsoft:urllib-abort-backends

bkaradzic-microsoft commented Jun 16, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

bkaradzic-microsoft Jun 18, 2026

Uh oh!

bghgary left a comment

Uh oh!

bghgary Jun 18, 2026

Uh oh!

bkaradzic-microsoft Jun 19, 2026

Uh oh!

bkaradzic-microsoft commented Jun 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

bkaradzic-microsoft commented Jun 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes

Validation

Scope / follow-ups

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

bkaradzic-microsoft Jun 18, 2026

Choose a reason for hiding this comment

Uh oh!

bghgary left a comment

Choose a reason for hiding this comment

Uh oh!

bghgary Jun 18, 2026

Choose a reason for hiding this comment

Uh oh!

bkaradzic-microsoft Jun 19, 2026

Choose a reason for hiding this comment

Uh oh!

bkaradzic-microsoft commented Jun 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

bkaradzic-microsoft commented Jun 16, 2026 •

edited

Loading