niobio / cld2

Automatically exported from code.google.com/p/cld2
0 stars 0 forks source link

Windows build fails: undeclared identifier 'close' #19

Closed GoogleCodeExporter closed 9 years ago

GoogleCodeExporter commented 9 years ago
There appears to be a weird mix of both open() and fopen() (with corresponding 
close() and fclose()) in cld2_dynamic_data_loader.cc, and possibly other places 
in the code. We should consistently use one or the other. To use close() we'd 
also technically need to depend on unistd.h, I think, which we currently don't. 
This is causing some problems for Chromium, though why it has just cropped up 
now I could not say:

https://code.google.com/p/chromium/issues/detail?id=403222

The fix here should be trivial, and I'll take care of it now.

Original issue reported on code.google.com by andrewha...@google.com on 13 Aug 2014 at 9:41

GoogleCodeExporter commented 9 years ago
So it wasn't quite as trivial as hoped, because mmap wants a file descriptor 
(int) instead of a FILE*. Using fopen() and fclose() is the portable way to do 
things, but under the covers they do a bunch of buffering and such that we 
don't really need for our mmap call. On the bright side it seems like using 
fileno() from stdio.h allows us to get the file descriptor, and I can't imagine 
that we will have any problems with mmap'ing the region read-only so long as we 
are also fopen()'ing in "r" (read-only) mode, even if it mmaps the same region 
under the covers. So hopefully all is well.

All unit tests pass in both dynamic and non-dynamic modes with the attached 
patch, so I think everything should be alright.

Original comment by andrewha...@google.com on 13 Aug 2014 at 10:02

Attachments:

GoogleCodeExporter commented 9 years ago
I had a pro/con discussion with another engineer on this, and here is the 
summary: the patch will fix the problem, but is a little sketchy because of the 
mixing of fopen() with mmap(). It would be better to #include unistd.h, and 
continue using open() instead of fopen(); but this is specific to Unix and 
wouldn't work on Windows. That said, the code is *already* broken on Windows 
because of the use of sys/mman.h (for mmap).

So: If we use unistd.h, we deepen the problems for Windows. If we use fopen(), 
we do something questionable for non-Windows (it works, but it isn't really the 
right thing to do).

I think the answer here is to use unistd.h for linux and use a workaround like 
this to fix win32:

// Based on definitions from 
http://sourceforge.net/p/predef/wiki/OperatingSystems/:
#ifdef _WIN32
#include <io.h>
#define OPEN _open
#define CLOSE _close
#else
#include "unistd.h"
#define OPEN open
#define CLOSE close
#endif

We can extend this solution to fix the mmap problem as well, in the near future.

Original comment by andrewha...@google.com on 13 Aug 2014 at 11:12

GoogleCodeExporter commented 9 years ago
I've created issue 20 to track the overall Windows compatibility problem for 
the dynamic data loader.

Original comment by andrewha...@google.com on 13 Aug 2014 at 11:40

GoogleCodeExporter commented 9 years ago
Here is a new patch based on #2 above. As described in issue 20, it disables 
support for win32 in the "file"-based dynamic data apis; only raw pointer mode 
will work on win32 until we have proper mmap support there.

So to be clear, we didn't end up changing the open() to fopen(), because open() 
makes sense for this use case. Updating the bug description accordingly.

Original comment by andrewha...@google.com on 13 Aug 2014 at 12:31

Attachments:

GoogleCodeExporter commented 9 years ago
Should be fixed in r166.

Original comment by andrewha...@google.com on 13 Aug 2014 at 12:35

GoogleCodeExporter commented 9 years ago
Issue 23 has been merged into this issue.

Original comment by andrewha...@google.com on 29 Aug 2014 at 11:33