[pbs-devel] [PATCH proxmox-backup v2] client: mount: flush output before exiting

[Gabriel Goller]

On 18.06.2024 15:00, Christian Ebner wrote:
>On 6/18/24 14:29, Gabriel Goller wrote:
>>
>>Hmm I think I know what happens...
>>This patch is flaky for me as well, sometimes it works, sometimes it
>>doesn't. The problem is that we move the `pw` file descriptor to
>>`mount_do`, which means that after it returns, the file_descriptor gets
>>dropped.
>>
>>     let (pr, pw) = proxmox_sys::pipe()?;
>>     let pr: OwnedFd = pr.into(); // until next sys bump
>>     let pw: OwnedFd = pw.into();
>>     match unsafe { fork() } {
>>         Ok(ForkResult::Parent { .. }) => {
>>             drop(pw);
>>             // Blocks the parent process until we are ready to go 
>>in the child
>>             let _res = nix::unistd::read(pr.as_raw_fd(), &mut 
>>[0]).unwrap();
>>             Ok(Value::Null)
>>         }
>>         Ok(ForkResult::Child) => {
>>             drop(pr);
>>             nix::unistd::setsid().unwrap();
>>             let result = 
>>proxmox_async::runtime::main(mount_do(param, Some(pw)));
>>                                                                            ^^^
>>pw gets dropped here, which means the fd is closed.
>>
>>             io::stdout().flush()?;
>>             io::stderr().flush()?;
>>             result
>>         }
>>         Err(_) => bail!("failed to daemonize process"),
>>     }
>>
>>
>>In the parent process we read from the fd—which is then closed—meaning
>>that the parent exits at the same time as we do the stderr/stdout flush.
>>
>>And actually this patch was stupid to begin with, because we don't even
>>print the error in the mound_do function, but later in the api-handler
>>:(.
>
>That is actually true, no point in flushing here. So my suggestion for 
>the fix was wrong, as I was primed on the error location :(
>

And I blindly followed :)

>>
>>I think we need to do something like this:
>>
>>     if let Err(e) = nix::unistd::read(pr.as_raw_fd(), &mut [0]) {
>>         // if the fd is closed, or some other error happens, wait so
>>         // that the child can return the error safely, then exit  
>>        nix::sys::wait::wait().unwrap();
>>     }
>>     Ok(Value::Null)

We actually need to also check the Ok branch if the passed buffer is
empty because if the fd get's closed we don't get an error code back.

Will submit a v3 shortly!