Quotation in System.Process.system for Windows

[simonpj]

Wagnerdm writes: I had a bit of fun recently tracking down quoting issues with the "system" command in Windows. For the examples below, I'll consistently use "Windows> " as the beginning of some text sent to the Windows command prompt cmd.exe, and use "GHC> " as the beginning of some text sent to a ghci session running in cmd.exe with System.Cmd imported.

The situation is this: I want to hand off a command line which has both a possibly-quoted command name and a (definitely) quoted argument. For concreteness, let's use "more" as the command and "foo.txt" as the argument, so that you can follow along at home on your favorite Windows system.

Windows> echo foo > foo.txt
Windows> "more" "foo.txt"
foo

All good so far. But:

GHC> system "\"more\" \"foo.txt\""
'more" "foo.txt' is not recognized as an internal or external command, operable program or batch file.
ExitFailure 1

After some digging, I discovered that system is shipping out to cmd /c, and indeed:

Windows> cmd /c "more" "foo.txt"
'more" "foo.txt' is not recognized as an internal or external command, operable program or batch file.

I don't know what the *right* fix is. However, after a bit of playing around, I discovered the following:

Windows> cmd /c ""more" "foo.txt""
foo
GHC> system "\"\"more\" \"foo.txt\"\""
foo
ExitSuccess

Wrapping commands with an extra pair of double-quotes this way seemed to give behavior matching the bare cmd.exe for all the examples I could think of, even ones I thought it would break. For example:

GHC> system "\"more foo.txt\""
foo
ExitSuccess

If this turns out to be the right thing to do, it's pretty easy to implement. In the commandToProcess function, at libraries/process/System/Process/Internals.hs:455, the change is just

-   return (cmd, translate cmd ++ "/c " ++ string)
+   return (cmd, translate cmd ++ "/c \"" ++ string ++ "\"")

(And in any case, the examples above should answer this amusing comment, immediately following those lines:

	-- We don't want to put the cmd into a single
	-- argument, because cmd.exe will not try to split it up.  Instead,
	-- we just tack the command on the end of the cmd.exe command line,
	-- which partly works.  There seem to be some quoting issues, but
	-- I don't have the energy to find+fix them right now (ToDo). --SDM
	-- (later) Now I don't know what the above comment means.  sigh.

Later, Brandon comments: What that comment means is that how CMD.EXE handles spaces is Windows-version-dependent. On the other hand, I think it's mostly consistent between XP and Windows 7 --- and I feel sorry for anyone forced to use an older version.

[simonpj]

I think we'd be thrilled if someonen felt able to take up the cudgels and sort this out.

Simon

[simonmar]

The comment worries me a bit, because it implies that when I tried this before it didn't work. Perhaps, as Brandon says, that is because the behaviour changed in Windows at some point (though I find it slightly hard to believe that this was on a Windows version earlier than XP).

Note that the behaviour should be much more reliable if you don't go via cmd, and run the command directly. I believe we do get the quoting right in that case.

If you want to take this forward, I suggest writing a few tests as well as the patch itself.