Repository navigation
Description
- Version: v8.3.0
- Platform:
Linux 19083a5259e3 4.4.0-1022-aws #31-Ubuntu SMP Tue Jun 27 11:27:55 UTC 2017 x86_64 GNU/Linux - Subsystem: Unknown
node appears to leave stdout/stderr in non-blocking mode in certain cases, which can cause errors in other programs which assume these are in blocking mode. I'm not certain if this is a bug or not, but am just trying to understand this behavior better.
I'm having trouble finding authoritative information about non-blocking mode and whether it's "correct" for programs to assume that they'll be given stdout/stderr in blocking mode, but it does appear that a lot of programs aren't able to handle having stdout/stderr in blocking mode (Python being a notable case). It's also strange that node only appears to do this in specific cases (e.g. in Docker without a TTY).
Reproduction
I haven't yet been able to find a way to reproduce this without Docker. I've created a simple test repo that prints the status of the blocking flag before and after running nodejs at https://github.com/chriskuehl/nodejs-nonblocking-repro. It shows that the blocking flag on stdout/stderr changes after nodejs is run.
Alternatively, a reproduction using Docker looks like this:
docker run -i node:latest bashto enter a shell in the container (don't use-t; it doesn't work if using a tty)python -c 'print "x" * 100000'inside the container to show that Python can print a lot of text without failingnodejs -e 'console.log(1+1)'inside the container. After running this command, stdout/stderr have been changed to non-blocking mode.python -c 'print "x" * 100000'inside the container, which should crash after printing some text (you should see a traceback somewhere, it may be at the top or interlaced in the middle of the printed text)
In Python 2, the traceback error is IOError: [Errno 11] Resource temporarily unavailable. In Python 3, it's a little more helpful: BlockingIOError: [Errno 11] write could not complete without blocking.
Activity
maclover7 commentedon Nov 26, 2017
@chriskuehl Does this still reproduce on v9.2.0?
chriskuehl commentedon Nov 26, 2017
@maclover7 yes it does, same reproduction steps. Here's output based on the test repro I linked above:
ivan commentedon Dec 8, 2017
I ran into this today as well and I have a reproducer without docker: https://gist.github.com/ivan/e1eafebc7e5273b6f80d2cc9bdb96a38 (tested on a recent Linux 4.4 and 4.14, nodejs 8.9.2 and 9.2.0)
bnoordhuis commentedon Dec 18, 2017
Potential fix in #17737. If anyone here can confirm it fixes the issue for them?
By the way, a python2 version of the test case:
ivan commentedon Dec 18, 2017
I did not test the fix, but I wonder if it solves the problem when the parent process doesn't wait for the node process to exit? (Is there something that prevents stdout from going nonblocking in this case?)
bnoordhuis commentedon Dec 18, 2017
Depends on what you mean. Whether there's someone waiting doesn't matter to node.js but if another process concurrently tries to read or write to the same file description (note: not file descriptor), it might observe non-blocking I/O.
That's hard to remedy in the UNIX model, though - you can't always reopen stdio to get fresh file descriptions.
gireeshpunathil commentedon May 19, 2018
potential fix in #20592 . @chriskuehl are you in a position to apply the patch , validate and confirm? or you want to wait for a binary?
src: restore stdio on program exit
addaleax commentedon Jun 11, 2018
I guess we have to reopen this :(
(For context, the fix was reverted in #21257)
20 remaining items
Bug 1500436 - Redirect node.js's stderr to a pipe. r=froydnj
Bug 1500436 - Redirect node.js's stderr to a pipe. r=froydnj
Bug 1500436 - Redirect node.js's stderr to a pipe. r=froydnj
Bug 1500436 - Redirect node.js's stderr to a pipe. r=froydnj
Bug 1500436 - Redirect node.js's stderr to a pipe. r=froydnj
Bug 1500436 - Redirect node.js's stderr to a pipe. r=froydnj
Bug 1500436 - Redirect node.js's stderr to a pipe. r=froydnj
Bug 1500436 - Redirect node.js's stderr to a pipe. r=froydnj a=RyanVM
Bug 1500436 - Redirect node.js's stderr to a pipe. r=froydnj
Bug 1500436 - Redirect node.js's stderr to a pipe. r=froydnj a=RyanVM
nh2 commentedon Dec 27, 2019
Quick question:
Given that the merged fix uses
atexit(), I assume the fix does nothing when the node program isSIGKILLed, crashes or is OOM killed?In that case programs that call node as a subprocess must still take extra precautions like Firefox does here by never giving
nodethe actual terminalstdout/stderr, is that right?bnoordhuis commentedon Dec 29, 2019
@nh2 Yes, if node is hard-killed then it won't have a chance to reset stdio. That caveat applies to SIGKILL (because it can't be caught) but not e.g. SIGINT and SIGTERM, their signal handlers restore stdio.
(SIGSEGV and SIGBUS currently don't but that could be added. Then again, those signals mean the program is in an undefined state and trying to clean up might make it worse.)
You can, just reset their file descriptors to blocking I/O and cooked mode afterwards.
Bug 1500436 - Redirect node.js's stderr to a pipe. r=froydnj a=RyanVM
Bug 1500436 - Redirect node.js's stderr to a pipe. r=froydnj
src: restore stdio on program exit