Node 20.3 Crashes all the time when executed inside docker

[ronag]

• Node 20.3.0 crashes on start in non-bullseye docker container
• In bullseye container node-gyp-build (yarn add bufferutil) fails with Text file busy
• In bullseye container + UV_USE_IO_URING=0 everything works

[ronag]

Another possible ref electron/rebuild#1085

[ronag]

@nodejs/libuv

[bnoordhuis]

"Text file busy" means trying to write a shared object or binary that's already in use.

My hunch is that node-gyp has some race condition in reading/writing files that wasn't manifesting (much) when everything still went through the much slower thread pool, whereas io_uring is fast enough to make it much more visible.

[ronag]

Is there a way to disable ioring using a env variable as a temporary workaround when running node-gyp?

[bnoordhuis]

Yes, set UV_USE_IO_URING=0 in the environment. Use at your own risk: not a stable thing, will disappear again the future.

[ronag]

test-srv3.hq.lan:~# docker run -it node:20.3.0 bash
root@6f6c66eb5077:/# UV_USE_IO_URING=0 yarn add bufferutil
node[8]: ../src/node_platform.cc:68:std::unique_ptr<long unsigned int> node::WorkerThreadsTaskRunner::DelayedTaskScheduler::Start(): Assertion `(0) == (uv_thread_create(t.get(), start_thread, this))' failed.
 1: 0xc8e4a0 node::Abort() [node]
 2: 0xc8e51e  [node]
 3: 0xd0a059 node::WorkerThreadsTaskRunner::WorkerThreadsTaskRunner(int) [node]
 4: 0xd0a17c node::NodePlatform::NodePlatform(int, v8::TracingController*, v8::PageAllocator*) [node]
 5: 0xc4bbc4 node::V8Platform::Initialize(int) [node]
 6: 0xc49408  [node]
 7: 0xc497db node::Start(int, char**) [node]
 8: 0x7ff94ec5818a  [/lib/x86_64-linux-gnu/libc.so.6]
 9: 0x7ff94ec58245 __libc_start_main [/lib/x86_64-linux-gnu/libc.so.6]
10: 0xba9ade _start [node]
Aborted

[ronag]

This is actually worse than I thought. Node doesn't run at all with 20.3

[ronag]

test-srv3.hq.lan:~# docker run -it node:20.3.0 node
node[1]: ../src/node_platform.cc:68:std::unique_ptr<long unsigned int> node::WorkerThreadsTaskRunner::DelayedTaskScheduler::Start(): Assertion `(0) == (uv_thread_create(t.get(), start_thread, this))' failed.
 1: 0xc8e4a0 node::Abort() [node]
 2: 0xc8e51e  [node]
 3: 0xd0a059 node::WorkerThreadsTaskRunner::WorkerThreadsTaskRunner(int) [node]
 4: 0xd0a17c node::NodePlatform::NodePlatform(int, v8::TracingController*, v8::PageAllocator*) [node]
 5: 0xc4bbc4 node::V8Platform::Initialize(int) [node]
 6: 0xc49408  [node]
 7: 0xc497db node::Start(int, char**) [node]
 8: 0x7f5c393be18a  [/lib/x86_64-linux-gnu/libc.so.6]
 9: 0x7f5c393be245 __libc_start_main [/lib/x86_64-linux-gnu/libc.so.6]
10: 0xba9ade _start [node]
test-srv3.hq.lan:~# docker run -it node:20.2.0 node
Welcome to Node.js v20.2.0.
Type ".help" for more information.
> 

[mcollina]

I can reproduce this, and this is quite critical.

[mcollina]

cc @nodejs/tsc for visibility

[richardlau]

FWIW on two systems I have access to (a Red Hat owned RHEL 8 machine and test-digitalocean-ubuntu1804-docker-x64-1 from the Build infra) docker run -it node:20.3.0 node is fine:

root@test-digitalocean-ubuntu1804-docker-x64-1:~# docker run -it node:20.3.0 node
Unable to find image 'node:20.3.0' locally
20.3.0: Pulling from library/node
bba7bb10d5ba: Pull complete
ec2b820b8e87: Pull complete
284f2345db05: Pull complete
fea23129f080: Pull complete
9063cd8e3106: Pull complete
4b4424ee38d8: Pull complete
0b4eb4cbb822: Pull complete
43443b026dcf: Pull complete
Digest: sha256:fc738db1cbb81214be1719436605e9d7d84746e5eaf0629762aeba114aa0c28d
Status: Downloaded newer image for node:20.3.0
Welcome to Node.js v20.3.0.
Type ".help" for more information.
>

I can reproduce the assertion failure on an Ubuntu 16.04 host with node:20.3.0 but not with node:20.3.0-bullseye:

root@infra-digitalocean-ubuntu1604-x64-1:~# docker run -it node:20.3.0 node
Unable to find image 'node:20.3.0' locally
20.3.0: Pulling from library/node
bba7bb10d5ba: Pull complete
ec2b820b8e87: Pull complete
284f2345db05: Pull complete
fea23129f080: Pull complete
9063cd8e3106: Pull complete
4b4424ee38d8: Pull complete
0b4eb4cbb822: Pull complete
43443b026dcf: Pull complete
Digest: sha256:fc738db1cbb81214be1719436605e9d7d84746e5eaf0629762aeba114aa0c28d
Status: Downloaded newer image for node:20.3.0
node[1]: ../src/node_platform.cc:68:std::unique_ptr<long unsigned int> node::WorkerThreadsTaskRunner::DelayedTaskScheduler::Start(): Assertion `(0) == (uv_thread_create(t.get(), start_thread, this))' failed.
 1: 0xc8e4a0 node::Abort() [node]
 2: 0xc8e51e  [node]
 3: 0xd0a059 node::WorkerThreadsTaskRunner::WorkerThreadsTaskRunner(int) [node]
 4: 0xd0a17c node::NodePlatform::NodePlatform(int, v8::TracingController*, v8::PageAllocator*) [node]
 5: 0xc4bbc4 node::V8Platform::Initialize(int) [node]
 6: 0xc49408  [node]
 7: 0xc497db node::Start(int, char**) [node]
 8: 0x7f6e8486218a  [/lib/x86_64-linux-gnu/libc.so.6]
 9: 0x7f6e84862245 __libc_start_main [/lib/x86_64-linux-gnu/libc.so.6]
10: 0xba9ade _start [node]
root@infra-digitalocean-ubuntu1604-x64-1:~# docker run -it node:20.3.0-bullseye node
Unable to find image 'node:20.3.0-bullseye' locally
20.3.0-bullseye: Pulling from library/node
93c2d578e421: Already exists
c87e6f3487e1: Already exists
65b4d59f9aba: Already exists
d7edca23d42b: Already exists
25c206b29ffe: Already exists
599134452287: Pull complete
bd8a83c4c2aa: Pull complete
d11f4613ae42: Pull complete
Digest: sha256:ceb28814a32b676bf4f6607e036944adbdb6ba7005214134deb657500b26f0d0
Status: Downloaded newer image for node:20.3.0-bullseye
Welcome to Node.js v20.3.0.
Type ".help" for more information.
>

Our website build is actually broken running apt update with the default Node.js LTS image based on Debian 12 (bookworm) - - we've switched to the Debian 11 (bullseye) based image for now: nodejs/build#3382

[targos]

FWIW I opened an issue about this in the docker-node repo: nodejs/docker-node#1918

TLDR: this is not a problem with Node.js itself, but with the default base OS used by the Docker image, which was upgraded for v20.3.0.

[ronag]

bullseye works for me as well

[ronag]

Now I also get the file busy error:

test-srv3.hq.lan:~# docker run -it node:20.3.0-bullseye bash
root@ed020dd3f80e:/# yarn add bufferutil
yarn add v1.22.19
info No lockfile found.
[1/4] Resolving packages...
[2/4] Fetching packages...
[3/4] Linking dependencies...
[4/4] Building fresh packages...
error /node_modules/bufferutil: Command failed.
Exit code: 126
Command: node-gyp-build
Arguments: 
Directory: /node_modules/bufferutil
Output:
/bin/sh: 1: node-gyp-build: Text file busy

EDIT: Works with UV_USE_IO_URING=0

[ronag]

So to summarize:

• Node 20.3.0 crashes on start in non-bullseye docker container
• In bullseye container install fails with "node-gyp-build: Text file busy"
• In bullseye container + UV_USE_IO_URING=0 everything works

[ronag]

Should I split the uring problem into a separate issue?

[bnoordhuis]

Can someone post the result of strace -o trace.log -yy -f node app.js when it crashes with that uv_thread_create check? I expect to see a failing clone/clone2/clone3 system call but it'd be good to confirm.

[richardlau]

On the Ubuntu 16.04 infra machine I cannot run apt in the bookworm based node:20.3.0 or node:lts containers to install strace in them (it's not there by default).

Another datapoint, adding --security-opt=seccomp:unconfined makes this work on the Ubuntu 16.04 host:

root@infra-digitalocean-ubuntu1604-x64-1:~# docker run --security-opt=seccomp:unconfined -it node:20.3.0 node
Welcome to Node.js v20.3.0.
Type ".help" for more information.
>

[bnoordhuis]

Right, then I can predict with near 100% certainty what the problem is: docker doesn't know about the newish clone3 system call. Its seccomp filter rejects it with some bogus error and node consequently fails when it tries to start a new thread.

This docker seccomp thing is like clockwork, it always pops up when new system calls are starting to see broader use. It's quite possibly fixed in newer versions.

[mcollina]

Updating docker to the latest version fixed it (v24.0.2) for me.

A few notes:

• Ubuntu 22/04 LTS ships with Docker v20.x, which does not support this.
• I did not test any version in between, and I couldn't quickly identify what releases of Docker fixed it. From various comments in issues, it seems runc (a dependency of Docker) fixed it in v1.0.2.

Here is what I think we should do:

• document this error and the UV_USE_IO_URING=0 solution for v20
• disable io_uring when we backport libuv in LTS lines

This seems a future-proof solution while keeping the current functionality available.

[bnoordhuis]

document this error and the UV_USE_IO_URING=0 solution for v20

UV_USE_IO_URING is (intentionally) undocumented and going away again so don't do that.

[mcollina]

@bnoordhuis Would you just document this as "if you are hit by this bug, update docker"?

[ronag]

I think there are two different things here. I'm not sure updating docker will help with the uring problem. Or does it? Please confirm.

[santigimeno]

If I'm reading this correctly there are 2 separate issues here.

• The crash in some instances. This seems to be directly related to a bug with docker and nothing to do with io_uring.
• The Text file busy error, which might or might not be io_uring related but, at least, seems to be exacerbated by the use of io_uring.

I think we should try to understand better the 2nd issue before disabling it.

[santigimeno]

The Text file busy should be fixed in the next release that contains libuv v1.46.0 which should be v20.4.0

[Haringat]

Does this still occur in 20.4.0?

[bnoordhuis]

It shouldn't, libuv now blacklists known-bad kernels.

[bnoordhuis]

I'll go ahead and close this now that it's fixed. To summarize:

• node v20.4.0 and newer contain the fix (well, workaround)

• if you still experience issues while running under docker, upgrade docker

[masterkain]

still no official docker image for 20.4, weird

[raxityo]

Unfortunately the PR for the official build is still open docker-library/official-images#14997

[ekzobrain]

Get the same error with image node:18.17.1 and using "ENV UV_USE_IO_URING=0" does not help.
But 18.17-alpine works even without UV_USE_IO_URING=0

[santigimeno]

18.17.1 comes with libuv 1.44.2 which does not include the io_uring based implementation. Can you open a new issue describing what you're observing?

[mitar]

I observed the same (similar) issue on v12.22.9 and now I upgraded to v20.8.0 and I still have the same issue. When I run node inside a Docker container, I get (this is on v20.8.0):

node[10]: ../src/node_platform.cc:68:std::unique_ptr<long unsigned int> node::WorkerThreadsTaskRunner::DelayedTaskScheduler::Start(): Assertion `(0) == (uv_thread_create(t.get(), start_thread, this))' failed.
 1: 0xc99960 node::Abort() [node]
 2: 0xc999de  [node]
 3: 0xd198a9 node::WorkerThreadsTaskRunner::WorkerThreadsTaskRunner(int) [node]
 4: 0xd199cc node::NodePlatform::NodePlatform(int, v8::TracingController*, v8::PageAllocator*) [node]
 5: 0xc53c23  [node]
 6: 0xc545b4 node::Start(int, char**) [node]
 7: 0x7f6da641bd90  [/lib/x86_64-linux-gnu/libc.so.6]
 8: 0x7f6da641be40 __libc_start_main [/lib/x86_64-linux-gnu/libc.so.6]
 9: 0xbb0c7e _start [node]
Aborted (core dumped)

Docker is pretty old: Docker version 19.03.10, build 9424aeaee9, but sadly I am unable to upgrade it at the moment. I thought that the latest node version should work?

[kamalkech]

i have the same, what is the correct solution ?

[LeoK80]

i have the same, what is the correct solution ?

Try upgrade to latest container runtime (docker, containerd, etc.) to latest. If nothing newer is picked up from your package manager, consider upgrading manually.

[kode54]

Experiencing this with nodejs 22.0.0-1 from official Arch packages, on x86_64, as of kernel 6.9.1. (Also experienced it with -rc7 before that, so I downgraded to 6.8.9 at the time.)