pyport.h, Wince and errno getter/setter

[a77a3d74-e722-44ae-97d2-39445fc24e84]

BPO	505846
Nosy	@tim-one, @loewis, @jackjansen
Files	pyport.diff: pyport.h context diff #3

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields: ```python assignee = None closed_at = created_at = labels = ['interpreter-core'] title = 'pyport.h, Wince and errno getter/setter' updated_at = user = 'https://bugs.python.org/bkc' ``` bugs.python.org fields: ```python activity = actor = 'nnorwitz' assignee = 'nnorwitz' closed = True closed_date = None closer = None components = ['Interpreter Core'] creation = creator = 'bkc' dependencies = [] files = ['3922'] hgrepos = [] issue_num = 505846 keywords = ['patch'] message_count = 19.0 messages = ['38764', '38765', '38766', '38767', '38768', '38769', '38770', '38771', '38772', '38773', '38774', '38775', '38776', '38777', '38778', '38779', '38780', '38781', '38782'] nosy_count = 5.0 nosy_names = ['tim.peters', 'loewis', 'nnorwitz', 'jackjansen', 'bkc'] pr_nums = [] priority = 'normal' resolution = 'rejected' stage = None status = 'closed' superseder = None type = None url = 'https://bugs.python.org/issue505846' versions = ['Python 2.3'] ```

[a77a3d74-e722-44ae-97d2-39445fc24e84]

Most of the remaining Windows CE diffs are due to the lack of errno on Windows CE. There are other OS's that do not have errno (but they have a workalike method).

At first I had simply commented out all references in the code to errno, but that quickly became unworkable.

Wince and NetWare use a function to set the per- thread "errno" value. Although errno #defines (like ERANGE) are not defined for Wince, they are defined for NetWare. Removing references to errno would artificially hobble the NetWare port.

These platforms also use a function to retrieve the current errno value.

The attached diff for pyport.h attempts to standardize the access method for errno (or it's work-alike) by providing SetErrno(), ClearErrno() and GetErrno() macros.

ClearErrno() is SetErrno(0)

I've found and changed all direct references to errno to use these macros. This patch must be submitted before the patches for other modules.

--

I see two negatives with this approach:

1. It will be a pain to think GetErrno() instead of "errno" when modifying/writing new code.

2. Extension modules will need access to pyport.h for the target platform.

In the worst case, directly referencing errno instead of using the macros will break only those platforms for which the macros do something different. That is, Wince and NetWare.

--

An alternative spelling/capitalization of these macros might make them more appealing. Feel free to make a suggestion.

--

It's probably incorrect for me to use SetErrno() as a function, such as

SetErrno(1);

I think the semi-colon is not needed, but wasn't entirely certain. On better advice, I will fix these statements in the remaining source files if this patch is accepted.

[d21744ff-f396-4c71-955e-7dbd2e886779]

Logged In: YES user_id=33168

Typically, the semi-colon problem is dealt with as in Py_SET_ERANGE_IF_OVERFLOW.

So,

#define SetErrno(X) do { SetLastError(X); } while (0)

I don't think (but can't remember if) there is any problem for single statements like you have. You could probably do:

#ifndef MS_WINCE
#define SetErrno(X) errno = (X)     /* note no ; */
#else
#define SetErrno(X) SetLastError(X) /* note no ; */
#endif

[tim-one]

Logged In: YES user_id=31435

All identifiers defined in pyport.h must begin with "Py_".
pyport.h is (and must be) #include'd by extension modules, and we need the prefix to avoid stomping on their namespace, and to make clear (to them and to us) that the gimmicks are part of Python's portability layer. A name like "SetErrno" is certain to conflict with some other package's attempt to worm around errno problems; Py_SetErrno () is not. Agree with Neal's final suggestion about dealing with semicolons.

[a77a3d74-e722-44ae-97d2-39445fc24e84]

Logged In: YES user_id=4631

Here is an amended diff with the suggested changes. I've tested the semi-colon handling on EVT, it works as suggested.

--

Question: What is the prefered style, #ifdef xyz or #if defined(xyz) ?

I try to use #ifdef xyz, but sometimes there's multiple possibilities and #if defined(x) || defined(y) is needed. Is that okay?

--

Upcoming issue (hoping you address in your reply). There are many "goto finally" statements in various modules. Unfortunately EVT treats "finally" as a reserved word, even when compiling in non C++ mode. Also, Metrowerks does the same.

I've changed all of these to "goto my_finally" as a quick work-around. I know "my_finally" sounds yucky, what's your recommendation for this?

[d21744ff-f396-4c71-955e-7dbd2e886779]

Logged In: YES user_id=33168

Need to change the // comment to / \/. gcc accepts this for C, but it's non-standard (at least it was, it may have changed in C99).

You can have 1 Py_SET_ERANGE_IF_OVERFLOW for both platforms if you do this:

#ifndef ERANGE
#define ERANGE 1
#endif

#define Py_SET_ERANGE_IF_OVERFLOW(X) \
    do { \
        if (Py_GetErrno() == 0 && ((X) == Py_HUGE_VAL || \
                                   (X) == -Py_HUGE_VAL))  \
             Py_SetErrno(ERANGE); \
    } while(0)
I'm not sure of the usefulness of Py_ClearErrno(), since
it's the same on all platforms.  If errno might be set to
something other than 0 in the future, it would be good to
make the change now.

I would suggest changing finally to cleanup.

[a77a3d74-e722-44ae-97d2-39445fc24e84]

Logged In: YES user_id=4631

I can post a new diff for the // or would you be willing to just change the patch you have?

I cannot use the same macros for Py_SET_ERANGE_IF_OVERFLOW (X) because Wince doesn't have ERANGE. You'll note the use of Py_SetErrno(1) which is frankly bogus. This is related to your comment on Py_ClearErrno()

Using (errno == 0) as meaning "no error" seems to me to be a python source "convention" forced on it by (mostly) floating point side effects. Because the math routines are indicating overflow errors through the side effect of setting errno (rather than returning an explicit NaN that works on all platforms), we must set errno = 0 before calling these math functions.

I suppose it's possible that on some platform "clearing the last error value" wouldn't be done this way, but rather might be an explicit function call. Since I was going through the source looking for all errno's, I felt it was clearer to say Py_ClearErrno() rather than Py_SetErrno(0), even though in the end they do the same thing on currently supported platforms.

I'm easy, if you want to replace Py_ClearErrno() with Py_SetErrno(0) I can do that too.

--

Regarding goto targets.. is it likely that "cleanup" might also collide with local variables? would _cleanup or __cleanup work for you?

[tim-one]

Logged In: YES user_id=31435

Brad, errno is required by ANSI C, which also defines the semantics of a 0 value. Setting errno to 0, and taking errno==0 as meaning "no error", are 100% portable across platforms with a standard-conforming C implementation. If this platform doesn't support standard C, I have to question whether the core should even try to cater to it:
the changes needed make no sense to C programmers, so may become a maintenance nightmare.

I don't think putting a layer around errno is going to be hard to live with, provided that it merely tries to emulate standard behavior. For that reason, setting errno to 0 is correct, but inventing a new ClearErrno concept is wrong (the latter makes no sense to anyone except its inventor \<wink>).

[a77a3d74-e722-44ae-97d2-39445fc24e84]

Logged In: YES user_id=4631

I've eliminated Py_ClearErrno() and updated all the source to use Py_SetErrno(0). Attached is an updated diff for pyport.h

[a77a3d74-e722-44ae-97d2-39445fc24e84]

Logged In: YES user_id=4631

Hi folks, just wondering if this patch is going to be rejected, or if you're all too busy and I have to be more patient ;-)

I have a passle of Python-CE folks waiting on me to finish checking in patches. This is the worst one, I promise!

Let me know what you want me to do, when you get a chance. Thanks

[d21744ff-f396-4c71-955e-7dbd2e886779]

Logged In: YES user_id=33168

Tim, I can check in or do whatever else needs to be done to check this in and move this forward. How do you want to procede?

Brad, I think most people are pretty busy right now.

[a77a3d74-e722-44ae-97d2-39445fc24e84]

Logged In: YES user_id=4631

Hi folks,

I need to proceed with the port to NetWare so I have something to demo at Brainshare in March. Unfortunately future patches from me will include both WINCE and NetWare specific patches, though hopefully there won't be much other than config.h and this patch (which is required for NetWare).

Is there anything I can do to make this patch more acceptable? Send a bottle of wine, perhaps? ;-)

[jackjansen]

Logged In: YES user_id=45365

Brad, I think this patch might be asking for too much. You're asking that all accesses to errno be replaced by GetErrno() or SetErrno() calls, really...

And for many cases there is a workaround, where you don't have to change user code (i.e. the normal C code still uses what it thinks is an errno variable). On my system errno is
#define errno (*__error())
and the __error() routine returns a pointer to the errno-variable for the current thread. For the GetErrno function this would be good enough, and with a bit of effort you could probably get it to work for the Set function too (possibly by doing the actual Set work in the next Get call).

[61337411-43fc-4a9c-b8d5-4060aede66d0]

Logged In: YES user_id=21627

The patch requires further surgery:

What is DONT_HAVE_TIME_H? If you want to test for presence of \<time.h>, you need to add HAVE_TIME_H to the autoconf machinery, and all manually-maintained copies of pyconfig.h. Including just the configure.in changes is fine; no need to include changes to generated files.

[61337411-43fc-4a9c-b8d5-4060aede66d0]

Logged In: YES user_id=21627

Any chances that updates to this patch are forthcoming?

If not, it will be rejected by October 1.

[a77a3d74-e722-44ae-97d2-39445fc24e84]

Logged In: YES user_id=4631

I would very much like to move this forward.

Is all you need is a refreshed diff without pyconfig.h diffs?

I'll have to check why DONT_HAVE_TIME_H is in there. I think perhaps because the CE portion is using the Win32 hand-made config, and only differs by a little bit.

What about jackjansen's post from 4-19? I cannot use the macro trick he suggests because there are two different functions for accessing errno, one for get and one for set.

Regarding his comment about changing all errno accesses: I'm still committed to submitting appropriate diffs for the core and any other module ported to CE or NetWare. In fact, it's time to refresh my CVS copy.

What do you suggest I check out? Head, or a specific revision?

Thanks

[d21744ff-f396-4c71-955e-7dbd2e886779]

Logged In: YES user_id=33168

You should work off CVS head.

[jackjansen]

Logged In: YES user_id=45365

Brad, if there's no other way to do your errno magic than by replacing all errno accesses with macros then so be it. *I* definitely don't want to hold off your patch because of that.

[a77a3d74-e722-44ae-97d2-39445fc24e84]

Logged In: YES user_id=4631

Based on recent python-dev discussion, I wish to retract this patch.

[d21744ff-f396-4c71-955e-7dbd2e886779]

Logged In: YES user_id=33168

Closing, since Brad retracted. Brad, I hope you continue your work. It seems the last part of the discussion was that Py_SerErrno() would be acceptable for the few places it is necessary. But continue to use errno in the code and for Win CE, #define errno GetErrno().