Original bug ID: 4541
Reporter: jsk
Status: closed (set by @damiendoligez on 2010-04-20T15:46:35Z)
Resolution: fixed
Priority: normal
Severity: crash
Platform: x86
OS: Ubuntu Linux
OS Version: 7.10
Version: 3.10.2
Fixed in version: 3.12.0+dev
Category: ~DO NOT USE (was: OCaml general)
Monitored by: bacam

Bug description

When debugging a bytecode program that calls Unix.fork(), ocamlrun
and ocamldebug usually crash immediately after the fork operation.

1. Create a program called "test.ml", containing the following:

   let _ =
       Printf.printf "process %n starting\n" (Unix.getpid ());
       flush stdout;
       let _ = Unix.fork () in
       Printf.printf "process %n ending\n" (Unix.getpid ());
   

2. Compile the program:

   "ocamlc -custom -g unix.cma test.ml -o test.exe".  
   

3. Start the debugger, specifying the location of "unix.ml" and an
   explicit socket:

   "ocamldebug -I <path-to-unix> -s <socket-name> test.exe"
   

4. Issue the following commands to ocamldebug:

   "set loadingmode manual"
   "goto 0"
   

5. Manually load the program (possibly on another machine):

   "CAML_DEBUG_SOCKET=<socket-name> ./test.exe"
   

6. Repeatedly issue the "step" command to ocamldebug.

At the point just after the call to Unix.fork(), ocamldebug usually
terminates with an error message of the form:

(ocd) Garbage data from process <n>
>> Fatal error: Debugcom.do_go
Uncaught exception: Misc.Fatal_error

Ocamlrun usually also terminates itself at this point.

This appears to be happening because ocamldebug is receiving unexpected
data from ocamlrun, presumably because the forked ocamlrun processes are
competing with one another to send data along the same connection to
ocamldebug.

Occasionally, though, ocamldebug doesn't crash, but continues to report
data from both parent and child ocamlrun processes. Issuing further step
commands to ocamldebug causes the forked processes to repond in a round-
robin style. Presumably this is because the forked ocamlrun processes
are still competing with one another to read data from the same shared
connection to ocamlbug, and are picking up commands alternately, one
after the other.

The current behaviour is making it difficult for us to debug programs
that launch other processes using the traditional "fork-exec" sequence,
since debugged programs have a very good chance of crashing immediately
after they call Unix.fork().

I'm wondering, is it possible to make ocamlrun aware of Unix.fork(),
so that a child ocamlrun process doesn't attempt to communicate with
ocamldebug through the same socket as its parent process?

Thanks for your help

Jonathan

Jonathan Knowles
Citrix Systems Research & Development

File attachments

• fork.patch