stdio regression: process.stdin getter drops pending stdin data

Version: v15.3.0
Platform: macOS Darwin
Subsystem: process (or maybe streams?)

What steps will reproduce the bug?

In v15, if process.stdin is referenced, and is not a TTY or Pipe, then its data will be not be seen when passed to a child process. It's almost like it's already consumed or something.

$ node -v
v15.3.0

$ cat shell.js
const { spawn } = require('child_process')
// just reference the object, don't do anything with it
if (process.env.LOAD_STDIN === '1')
  process.stdin
const child = spawn(process.env.SHELL, [], { stdio: 'inherit' })
child.on('close', (code, signal) => console.error({ code, signal }))

$ cat cmd.sh
#!/bin/bash
echo "hello from node's child shell"

$ node shell.js < cmd.sh
hello from node's child shell
{ code: 0, signal: null }

## This is the error:

$ LOAD_STDIN=1 node shell.js < cmd.sh
{ code: 0, signal: null }

## but this works, which is interesting?

$ cat cmd.sh | LOAD_STDIN=1 node shell.js
hello from node's child shell
{ code: 0, signal: null }

$ nave use 14

$ node -v
v14.15.1

$ node shell.js < cmd.sh
hello from node's child shell
{ code: 0, signal: null }

$ LOAD_STDIN=1 node shell.js < cmd.sh
hello from node's child shell
{ code: 0, signal: null }

$ cat cmd.sh | LOAD_STDIN=1 node shell.js
hello from node's child shell
{ code: 0, signal: null }

How often does it reproduce? Is there a required condition?

100% of the time on node 15. Not in any prior version tested (14, 12, and 10).

What is the expected behavior?

Stdin data should not be consumed when process.stdin is merely referenced. This causes problems when a script checks process.stdin.isTTY to print a helpful message prior to opening a shell in a child process.

What do you see instead?

Stdin data is not passed to child process if stdin is referenced, and not a Pipe or TTY.

Additional information

This kind of is a pain for npm exec. It means you can't do something like npx some-interactive-thing < list-of-commands.

@isaacs

@isaacs, there is a relevent discussion here #35997
The creation of the implicit stdio ReadStream is in the process.stdio getter

able to recreate with the said version, but not consistently.

If I run ~10 times, ~2 runs go fine.
If I run it under dtruss, it passes consistently

so I believe there is a race condition happening here.

@gireeshpunathil

@gireeshpunathil Yes there is. The ReadingStream starts reading immediately as well as the clone/exec path. This is the race condition. If libuv was able to read from the fd before the clone/exec, that the data is lost.
On Linux, at least on my machine, strace allows you to observe both situations.
@joyeecheung maybe create the ReadStream in stopped state?

@mmomtchev

@mmomtchev - thanks, I too got a linux box which recreates pretty easily.

passing case:

[pid 661072] execve("/bin/bash", ["/bin/bash"], 0x55dd87501770 /* 22 vars */ <unfinished ...>
...
[pid 661072] read(0, "#!/bin/bash\necho \"hello from nod"..., 49) = 49
[pid 661072] fstat(1, {st_mode=S_IFCHR|0620, st_rdev=makedev(0x88, 0), ...}) = 0
[pid 661072] write(1, "hello from node's child shell\n", 30) = 30
[pid 661072] read(0, "", 49)            = 0
[pid 661072] rt_sigprocmask(SIG_BLOCK, [CHLD], [], 8) = 0
[pid 661072] rt_sigprocmask(SIG_SETMASK, [], NULL, 8) = 0
[pid 661072] exit_group(0)              = ?

the process 661072 is exec'ing into bash and reads from 0.

failing case:

489 [pid 661087] execve("/bin/bash", ["/bin/bash"], 0x557edd43a770 /* 22 vars */ <unfinished ...>
...
615 [pid 661088] <... read resumed>"#!/bin/bash\necho \"hello from nod"..., 65536) = 49
..
701 [pid 661087] read(0, "", 49)            = 0
702 [pid 661087] rt_sigprocmask(SIG_BLOCK, [CHLD], [], 8) = 0
703 [pid 661087] rt_sigprocmask(SIG_SETMASK, [], NULL, 8) = 0
704 [pid 661087] exit_group(0)              = ?

The process 661087 is exec'ing into bash but does ot get any data, due to the previous read by 661088 - but what it is? (it is cloned from 661080 which is the parent process)

@gireeshpunathil

@gireeshpunathil @isaacs I solved it by implementing a manualStart for ReadStream then using it in the process.stdin getter, just a sec to clean my debug printfs and if the unit tests pass, I will submit the PR

@gireeshpunathil

@gireeshpunathil All other stdio ReadStream start in paused mode - except file - which didn't have a manualStart option

@isaacs

@isaacs @gireeshpunathil This will be a platform-dependent unit test, exclusive to darwin and linux unless you have an idea?

stdio: implement manual start for ReadStream

The unit test is awful, I admit, if you have any ideas, fell free to comment

stdio: implement manual start for ReadStream

stream: lazy read ReadStream

Navigation Menu

Repository navigation

stdio regression: process.stdin getter drops pending stdin data #36251

What steps will reproduce the bug?

How often does it reproduce? Is there a required condition?

What is the expected behavior?

What do you see instead?

Additional information

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

stdio regression: process.stdin getter drops pending stdin data #36251

Description

What steps will reproduce the bug?

How often does it reproduce? Is there a required condition?

What is the expected behavior?

What do you see instead?

Additional information

Activity

mmomtchev commented on Nov 26, 2020

gireeshpunathil commented on Nov 26, 2020

mmomtchev commented on Nov 26, 2020

gireeshpunathil commented on Nov 26, 2020

mmomtchev commented on Nov 26, 2020

mmomtchev commented on Nov 26, 2020

mmomtchev commented on Nov 26, 2020

mmomtchev commented on Nov 26, 2020

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions