I'm sorry I couldn't contribute to the discussion in time. The fix of the fstat() Win32 port looks good to me. I agree that there's a need for multiple fseek() ports to address the shortcomings of the MSVC functionality.
The documentation event states that "on devices incapable of seeking, the return value is undefined". A simple wrapper using GetFileType() or the new fstat(), to filter non-seekable devices before delegation, will probably suffice.
--
Daniel
On Mon, Mar 13, 2023 at 05:49:41PM +0100, Juan José Santamaría Flecha wrote:
> WFM, making fseek() behaviour more resilient seems like a good improvement
> overall.
I have not looked in details, but my guess would be to add a
win32seek.c similar to win32stat.c with a port of fseek() that's more
resilient to the definitions in POSIX.
> Should we open a new thread to make that part more visible?
Yes, perhaps it makes sense to do so to attract the correct audience,
There may be a few things we are missing.
When it comes to pg_dump, both fixes are required, still it seems to
me that adjusting the fstat() port and the fseek() ports are two
different bugs, as they influence different parts of the code base
when taken individually (aka this fseek() port for WIN32 would need
fstat() to properly detect a pipe, as far as I understand).
Meanwhile, I'll go apply and backpatch 0001 to fix the first bug at
hand with the fstat() port, if there are no objections.
--
Michael