Calling process_exec too many times in the same running process will error out with EMFILE

[neraid-live]

Context

Odin: dev-2026-02:7b18a08d5
OS: Arch Linux, Linux 6.18.9-arch1-2
CPU: AMD Ryzen 5 8600G w/ Radeon 760M Graphics
RAM: 31167 MiB
Backend: LLVM 21.1.6

Expected Behavior

process_exec should always properly close process handles to avoid EMFILE

Current Behavior

process_exec doesn't clean up process/file handles

Failure Information (for bugs)

Consider the scenario:
You have a program with an infinitely running for loop which will take the stdout of a process_exec call and do parsing on that output and print it to the console every 5s. After a few iterations, process_exec starts erroring out with EMFILE due to the process handles not being properly cleaned up.

Steps to Reproduce

Please provide detailed steps for reproducing the issue.

1. call process_exec a bunch of times

[laytan]

The obvious fix is doing a process_close at the end of the procedure. But I was under the impression that a process_wait which returns that the process exited would clean up the process.

[jasonKercher]

Agreed. This is an oversight since we process_open in process_exec. This should fix it:

diff --git a/core/os/process_linux.odin b/core/os/process_linux.odin
index 7041e16b7..fd385e501 100644
--- a/core/os/process_linux.odin
+++ b/core/os/process_linux.odin
@@ -682,7 +682,7 @@ _reap_terminated :: proc(process: Process) -> (state: Process_State, err: Error)
                state.exit_code = int(info.status)
                state.success = false
        }
-       return
+       return state, _process_close(process)
 }

 _timed_wait_on_handle :: proc(process: Process, timeout: time.Duration) -> (process_state: Process_State, err: Error) {
@@ -767,7 +767,7 @@ _timed_wait_on_handle :: proc(process: Process, timeout: time.Duration) -> (proc
                        process_state.success = false
                }
        }
-       return
+       return process_state, _process_close(process)
 }

 _timed_wait_on_pid :: proc(process: Process, timeout: time.Duration) -> (process_state: Process_State, err: Error) {

I cannot make this change myself though. I'm working on a cruise ship right now, and the ssh stuff for git doesn't seem to work over the ship's connection.

[neraid-live]

Agreed. This is an oversight since we process_open in process_exec. This should fix it:

diff --git a/core/os/process_linux.odin b/core/os/process_linux.odin
index 7041e16b7..fd385e501 100644
--- a/core/os/process_linux.odin
+++ b/core/os/process_linux.odin
@@ -682,7 +682,7 @@ _reap_terminated :: proc(process: Process) -> (state: Process_State, err: Error)
state.exit_code = int(info.status)
state.success = false
}

•   return
  

•   return state, _process_close(process)
  

}

_timed_wait_on_handle :: proc(process: Process, timeout: time.Duration) -> (process_state: Process_State, err: Error) {
@@ -767,7 +767,7 @@ _timed_wait_on_handle :: proc(process: Process, timeout: time.Duration) -> (proc
process_state.success = false
}
}

•   return
  

•   return process_state, _process_close(process)
  

}

_timed_wait_on_pid :: proc(process: Process, timeout: time.Duration) -> (process_state: Process_State, err: Error) {
I cannot make this change myself though. I'm working on a cruise ship right now, and the ssh stuff for git doesn't seem to work over the ship's connection.

Applied the patch on my local build and the issue seems to be fixed. Will anyone make a PR for this? I can make it if that's not a problem

[laytan]

I am looking at this now and am hesitant, with that patch, the following code will close the handle twice:

p, _ := process_start(...)
defer process_close(p)

state, _ := process_wait(p)

It's unfortunate that we pass Process by value, otherwise process_close could zero out the handle, making the 2nd a no-op.

I am unsure what to do here:

1. Always require a process_close
2. Say that code snippet is a misuse of the API and just apply the patch
3. Remove the public process_close procedure (and apply the patch), you always close through process_wait or process_kill

In addition to the patch, we probably want to close the file descriptor when we encounter an error (non-timeout) in process_wait. Not just a successful wait.

[laytan]

After some discussing I ended up implementing 3.

[jasonKercher]

I would consider that snippet a misuse of the API. I wouldn't expect to have to call process_close unless I opened it myself. It might not be a bad idea to change it to take a pointer.

Also, not really sure about closing on error either. I guess there isn't any harm in it if you're already in error territory anyway.