| msg10134 - (view) |
Author: Greg Chapman (glchapman) |
Date: 2002-04-03 17:10 |
| With the current Win32 build (actually, 2.2.1c1), you get the following behavior when feeding bad filenames to Python's open function: >>> open("hello?.txt") Traceback (most recent call last): File "", line 1, in ? IOError: invalid argument: r Apparently, VC6 maps the Win32 error ERROR_INVALID_NAME to EINVAL, and Python always treats this error as being related to the mode string. It appears (from bug report 476593) that VC6 disagrees with GCC here; I don't know which compiler is right (actually, looking at the VC documentation for _open, it appears GCC is right). At any rate, it appears Python cannot rely on EINVAL referring only to the mode string under Windows. Interestingly, the current Python Win32 build also shows this: >>> open("hello.txt", "x") Traceback (most recent call last): File "", line 1, in ? IOError: [Errno 0] Error: 'hello.txt' When you actually provide a bad mode string, fopen fails, but it does not set errno. (You can see this in the openfile function in _OPEN.C in the VC6 CRT source). It might be possible to use this behavior to produce more accurate error messages under Windows. |
|
|
| msg10135 - (view) |
Author: Michael Hudson (mwh)  |
Date: 2002-04-03 17:15 |
| Logged In: YES user_id=6656 Guess who gets to look at this one. |
|
|
| msg10136 - (view) |
Author: Tim Peters (tim.peters) * |
Date: 2002-04-07 09:00 |
| Logged In: YES user_id=31435 Which version of Windows are you using? Here on Win98SE, I get >>> open("hello?.txt") Traceback (most recent call last): File "", line 1, in ? IOError: [Errno 2] No such file or directory: 'hello?.txt' >>> all the way back through Python 2.0. Note that MS's error translation table (in DOSMAP.C) doesn't contain an entry for ERROR_INVALID_NAME, so the logic in _dosmaperr() would return EINVAL in that case (it can't dream up anything better to return). The errors CreateFile *can* trigger aren't documented, and I expect they vary across Windows flavors (and that Win98SE triggers ERROR_FILE_NOT_FOUND in this case, which maps to ENOENT). Note that MS doesn't catch all senseless mode strings either. For example, >>> open('wtf', 'w hi greg!') <open file 'wtf', mode 'w hi greg!' at 0x00768F30> >>> If you want to create a patch to clean up MS's mess here, be my guest. |
|
|
| msg10137 - (view) |
Author: Greg Chapman (glchapman) |
Date: 2002-04-07 19:09 |
| Logged In: YES user_id=86307 I'm using Windows 2000; I also just tested this on Windows XP Pro and got the same results (as on Windows 2000). As far as a patch, the simplest thing might be to add an ifdef so that EINVAL is not treated as a special case when compiling with VC; this would revert to the behavior before November 2001 (i.e., the patch for bug 476593). If this path is chosen, should there be a special define introduced for invalid use of EINVAL (I don't know if other C libraries have this problem), or should it simply refer to MSVC? The above wouldn't fix the problem with MS's handling of invalid mode strings. I had noticed that you could put anything in a mode string provided the first character was in [rwa]. How do other C libraries handle these kind of errors in mode strings? I have the source to Borland C++ Builder 4. It looks like it checks the entire mode string, but it also fails without setting errno if it finds any invalid characters. Is fopen expected to set errno on failure? (I note that neither the MS nor the Borland documentation for fopen refers to errno.) |
|
|
| msg10138 - (view) |
Author: Tim Peters (tim.peters) * |
Date: 2002-04-07 20:21 |
| Logged In: YES user_id=31435 I'm afraid there are no answers to your questions short of doing an exhaustive study of each implementation of interest, and even then it hits a brick wall because Win32 CreateFile() doesn't document its error conditions and its source code is secret. I'm attaching a patch that should reliably distinguish between mode and filename errors across Windows flavors under MSVC, and that maps EINVALID to ENOENT on Windows (which may not be the best that *could* be done on Win2K, but Windows flavors are inconsistent here and Python has no idea which flavor of Win32 it's running under). If that's good enough for you, I'll check it in,; else I'm inclined to close this as 3rdParty and Won'tFix. |
|
|
| msg10139 - (view) |
Author: Mark Hammond (mhammond) * |
Date: 2002-04-08 00:11 |
| Logged In: YES user_id=14198 Confirming that on Win2k, before the patch I see: >>> open("hello?.txt") Traceback (most recent call last): File "", line 1, in ? IOError: invalid argument: r and after: >>> open("hello?.txt") Traceback (most recent call last): File "", line 1, in ? IOError: [Errno 2] No such file or directory: 'hello?.txt' and a quick scan shows the patch to be reasonable. So +1 on checking it in. |
|
|
| msg10140 - (view) |
Author: Tim Peters (tim.peters) * |
Date: 2002-04-08 04:20 |
| Logged In: YES user_id=31435 Based on Mark's confirmation and reviewlet, I'm closing this report. The patched logic was checked in on the trunk and on the 2.2.1 branch (so will be in 2.2.1 final later this week). |
|
|