On Mon, Apr 13, 2020 at 02:05:00PM +0200, Son Luong Ngoc wrote:
> I am trying to write a simple git trace2 event collector and I notice
> that when git doing git pull with trace events being sent to a unix
> stream socket, the entire operation halted.
>
> Reproduce as follow:
> ```
> cd git/git
> git config trace2.eventTarget af_unix:stream:/tmp/git_trace.sock
> git config trace2.eventBrief false
> (rm /tmp/git_trace.sock | ) && nc -lkU /tmp/git_trace.sock
>
> # In a different terminal
> git pull # Pull stuck and never complete
> ```
I think the issue is the use of netcat as the server side.
Your git-pull involves multiple simultaneously-running Git processes.
But "nc -k" will only accept() a new client once the old one has
disconnected. So we'd deadlock any time we have this situation:
- process A opens a stream to the socket, and keeps it open
- process A spawns process B and waits for it to finish
- process B tries to open a stream to the socket, which will block
waiting for netcat to accept()
Now A cannot make forward progress until B finishes, but B will not make
forward progress until A closes the socket.
I was able to reproduce the issue locally, and process "A" was git-pull
and process "B" was git-merge.
Switching to using datagrams ("nc -u" and "af_unix:dgram") makes the
problem go away, because the connections are able to interleave their
packets.
> This does not happen when you set eventBrief to true
> ```
> git config trace2.eventBrief true
> ```
I don't know why this would matter (and it didn't seem to in my tests).
> Worth to note that if eventTarget is a file instead of a socket,
> everything works fine.
That makes sense, since their file writes can be interleaved.
-Peff
