Kernel#exec should raise Errno::ENOENT if the program to execute doesn't exist

Not enough time...Bumping

I tried the provided testcase on my Windows XP SP2 box with current JRuby trunk and it passes, so I can not reproduce it. Is this issue already fixed? On what environment have you seen that bug?

Indeed, the test passes on Windows, but fails on Linux.

Depending on how the external command is being executed, the error is either there or not. If programm is launched via shell, then there will be no error. if program is launched directly, the error will be correctly generated.

The cause of this problem is that unlike MRI, jruby does not search the PATH before attempting to execute the arguments to #exec or #system. If this is implemented in jruby, it will resolve JRUBY-2235 as well, which relates to the error output produced when one of those untested executions fail.

Port of MRI's executable path search – verifies that executables can be found before attempting to execute them. MRI's functionality has some quirks, e.g. doesn't search if shell metacharacters, broadly defined, are found in the argument to exec(). This carries over to the jruby implementation. However, the jruby implementation handles tilde-paths better: MRI currently punts in the case that the tilde is followed by a name. I don't have access to a Windows machine at the moment, so I was unable to test there – but I did try to carry over the relevant support for that platform. This should also fix JRUBY-2235.

Riley: The patch looks really excellent, and I'm sorry we didn't see it sooner. But it does seem to cause a failure in test/testEnv.rb. That may be a bad test or a flaw in the patch. Can you look into it? I'd really like to have this path-searching logic in 1.1.7.

The testEnv failure:

◆ jruby test/testEnv.rb 
.........................
--------------------------------------------------------------------------------
--------------------------------------------------------------------------------
Tests: 25. (Ok: 25; Failed: 0)
<script>:1: No such file or directory - echo $__JRUBY_T1 (Errno::ENOENT)

The testEnv failure is due to a bug in the path-search. I'm looking into it.

I forgot to tokenize the string on whitespace – I was passing the whole command string for lookup, rather than just the first whitespace-delimited argument. The patch is updated, with some additional tests.

The patch needs (at least) one more revision: MRI punts on doing the path search if there are any shell metacharacters. JRuby should do the same thing, or invocations like `LC_ALL=C date` will try to look up "LC_ALL=C".

Well, we're getting there I have a different failure now. Are you able to run "ant test" to completion on your system?

    [junit] Testcase: test_backquote(org.jruby.test.TestUnitTestSuite$ScriptTest):	FAILED
    [junit] Faults encountered running test/test_backquote, complete output follows:
    [junit] Error:
    [junit] test_empty_backquotes(TestBackquote):
    [junit] Errno::ENOENT: No such file or directory - 
    [junit]     <script>:1:in `test_empty_backquotes'

Mea culpa! I didn't run the whole suite, but now that I do, I get the same result that you do. Looks like test/test_load is also broken, but that's due to the shell metacharacter problem. Working on a fix for both issues.

Updated patch:

• Added support for shell metacharacter detection – system() and exec() don't attempt path executable search when a shell metacharacter (as defined by MRI) is present in the commandline.
• Backquote never attempts path executable search.

I think it's passing all the tests now – but let me know if you find different!

OK, it's time to fix this, once and for all.

Riley, thank you very much for the patch! It was a great staring point I used to improve upon.

Lots of new tests been added, all of them pass now on Windows and Linux, ShellLauncher got a major refactoring along the way.

See cc4b619 and 138e726 on master branch.

Tests and specs pass too.