Repository navigation
http: emit 'drain' after flushing coalesced chunks - #66782
Merged
Merged
Conversation
Since chunked writes are coalesced while the message is corked, write() decides backpressure from the byte length of the buffered chunks and sets kNeedDrain. When uncork() flushes them, net.Socket accounts string chunks in UTF-16 code units, so with multi-byte text it can stay below its own highWaterMark. It then never needs to drain and never emits 'drain'. If the flushed writes complete asynchronously, the check at the end of uncork() also runs while bytes are still pending, so the response stalls with writableNeedDrain === true and writableLength === 0. Re-check the drain condition once the flushed data has been written, unless the message or its socket has been destroyed in the meantime. Fixes: nodejs#66781 Refs: nodejs#64987 Assisted-by: AI coding agent Signed-off-by: junyup0503 <97156898+junyup0503@users.noreply.github.com>
Collaborator
|
Review requested:
|
nodejs-github-bot
added
http
Contributor
|
Welcome to Node.js, and thank you for your first contribution! Before review, please take a moment to read:
Please make sure every commit is signed off. For a first pull request, GitHub Actions require collaborator approval and Jenkins CI must be started by a collaborator or triager, so an initial wait is normal. |
The socket reference is captured when uncork() runs, so it is null if the message had no socket yet and the write was buffered in outputData. Re-reading this[kSocket] in the callback checks the socket the data was actually flushed to. In practice _flush() usually clears kNeedDrain before this callback runs, so this is a defensive guard rather than a fix for an observed crash. |
par333k
reviewed
Oct 16, 2026
Comment on lines
+382
to
+384
| // Re-check once the flushed data has been written. | ||
| if (!err && this.writableNeedDrain && !socket.destroyed && | ||
| this.writableLength === 0) { |
There was a problem hiding this comment.
Suggested change
| // Re-check once the flushed data has been written. | |
| if (!err && this.writableNeedDrain && !socket.destroyed && | |
| this.writableLength === 0) { | |
| const currentSocket = this[kSocket]; | |
| if (!err && this.writableNeedDrain && currentSocket && | |
| !currentSocket.destroyed && this.writableLength === 0) { |
Member
There was a problem hiding this comment.
+1, I think without this there is likely a reachable crash here, good catch @par333k
Contributor
Author
This was referenced Oct 22, 2026
7 tasks done
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #66782 +/- ##
==========================================
+ Coverage 90.49% 90.70% +0.20%
==========================================
Files 797 797
Lines 277598 278906 +1308
Branches 53336 53630 +294
==========================================
+ Hits 251220 252978 +1758
+ Misses 16790 16297 -493
- Partials 9588 9631 +43
🚀 New features to boost your workflow:
|
pimterry
requested changes
Oct 28, 2026
pimterry
left a comment
pimterry
left a comment
Member
There was a problem hiding this comment.
The comment here needs resolving, but otherwise I think this is good to go and fixes a real regression. Thanks for the fix @junyup0503!
Comment on lines
+382
to
+384
| // Re-check once the flushed data has been written. | ||
| if (!err && this.writableNeedDrain && !socket.destroyed && | ||
| this.writableLength === 0) { |
Member
There was a problem hiding this comment.
+1, I think without this there is likely a reachable crash here, good catch @par333k
Read the current socket after the coalesced write callbacks run, since uncork() may initially buffer data before a socket is assigned. Check that the current socket exists and is not destroyed before emitting drain. Refs: nodejs#66782 (comment) Assisted-by: AI coding agent Signed-off-by: junyup0503 <97156898+junyup0503@users.noreply.github.com>
pimterry
approved these changes
Oct 30, 2026
pimterry
added
the
request-ci
github-actions
Bot
removed
the
request-ci
Collaborator
ronag
approved these changes
Oct 30, 2026
Contributor
Author
|
I applied @par333k’s suggestion locally but missed the co-author attribution. Could we include them as a co-author in the final commit when landing this PR? |
Collaborator
inoway46
added
the
resume-ci
github-actions
Bot
added
resume-ci-failed
Contributor
Failed to resume CI
Full Auto Start CI output |
ronag
reviewed
Nov 1, 2026
Comment on lines
+384
to
+385
| if (!err && this.writableNeedDrain && currentSocket && | ||
| !currentSocket.destroyed && this.writableLength === 0) { |
Member
There was a problem hiding this comment.
Suggested change
| if (!err && this.writableNeedDrain && currentSocket && | |
| !currentSocket.destroyed && this.writableLength === 0) { | |
| if (!err && this.writableNeedDrain && !this.destroyed && this.writableLength === 0) { |
Maybe?
Contributor
Author
There was a problem hiding this comment.
Applied in 3f263a8, thanks. CI needs a fresh run — the auto-resume failed with an infra error.
Rely on the message's own destroyed state instead of the socket's when re-checking the drain condition after the coalesced chunks are flushed. writableNeedDrain already excludes destroyed and finished messages. Refs: nodejs#66782 (comment) Co-authored-by: Robert Nagy <ronagy@icloud.com> Co-authored-by: Park Jin Woong <35551417+par333k@users.noreply.github.com> Assisted-by: AI coding agent Signed-off-by: junyup0503 <97156898+junyup0503@users.noreply.github.com>
Collaborator
Member
|
The AIX test here is an issue while will not occur on a re-run, so once the existing PR run is complete a re-run should allow it to pass |
Contributor
|
Related Windows failure: |
The default stream highWaterMark is 16 KiB on Windows and 64 KiB elsewhere. With the smaller value the fake socket exceeds its own highWaterMark in UTF-16 code units and emits 'drain' by itself before the write callback destroys the response, which trips mustNotCall(). Pin the highWaterMark so the payload sizes stay on the intended side of it on every platform, and assert that the socket never drains on its own in the destroy scenario. Refs: nodejs#66782 (comment) Assisted-by: AI coding agent Signed-off-by: junyup0503 <97156898+junyup0503@users.noreply.github.com>
Contributor
Author
|
Thanks. That was a test assumption: the default highWaterMark is 16 KiB on Windows, so the fake socket emitted 'drain' on its own before the write callback destroyed the response. Fixed in 0ecaa63 by pinning the socket's highWaterMark; reproduced and verified locally with the 16 KiB default. A fresh CI run is needed. |
ronag
approved these changes
Nov 4, 2026
ronag
added
request-ci
github-actions
Bot
removed
the
request-ci
Collaborator
Contributor
Author
|
Windows failed again in 78822, but ci.nodejs.org currently requires login so I can't see which test failed. Could someone share the failing test output from node-test-commit-windows-fanned/81639? Locally the updated test passes with the 16 KiB default on every run, so I'd like to confirm whether it's this test or an unrelated Windows failure. |
Collaborator
ronag
approved these changes
Nov 7, 2026
ronag
added
the
request-ci
github-actions
Bot
added
request-ci-failed
Contributor
Failed to start CIFull Auto Start CI output |
mcollina
added
the
commit-queue
nodejs-github-bot
added
commit-queue-failed
Collaborator
Commit Queue failedThis pull request has multiple commits, but no landing policy was selected. Add
commit-queue-squash
The pull request was removed from the Commit Queue and labeled
commit-queue-failed
Full Commit Queue output |
ronag
added
author ready
Collaborator
|
Landed in 66c5a02 |
nodejs-github-bot
removed
the
commit-queue
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Since #64987, chunked writes made while the response is auto-corked are coalesced, and
write()returnsfalsebased on the byte length of the buffered chunks. Whenuncork()flushes them,net.Socketcounts string chunks in UTF-16 code units, so with multi-byte text the socket can stay below its ownhighWaterMarkand never emits'drain'. If the flushed writes complete asynchronously, the check at the end ofuncork()also runs while bytes are still pending, so the response stalls withwritableNeedDrain === trueandwritableLength === 0.This re-checks the drain condition in the write callback of the flushed data, unless the message or its socket has been destroyed. It does not change the accounting itself (e.g. encoding the coalesced strings to
Buffers), to avoid adding cost to the write path that #64987 optimized.The new test uses a
Duplexthat counts strings likenet.Socketand completes writes asynchronously, so it does not depend on OS socket buffer sizes. It hangs onmainwithout the change and passes with it. It also checks that'drain'is not emitted after the response is destroyed from a write callback.parallel/test-http*,parallel/test-stream*,parallel/test-net*,sequential/test-http*andmake lint-jspass locally.Fixes: #66781
Refs: #64987
I used an AI coding agent to help investigate and write this change. I reproduced the issue and verified the fix and the tests myself.