On Thu, Nov 06, 2014 at 03:03:18PM +0100, Gustaf Neumann wrote: > The seemingly same write() operation, which refuses to write to > the access.log in driver.c works fine when called directly > in nslog.c.
> https://bitbucket.org/naviserver/naviserver/commits/4cb76f20a8464fde98ac3147c6184a90442a2808 Wow, weird. I tried turning your new fix on/off, and confirm the behavior you see. Yet as you said, the new code you added to NsAsyncWrite() is the same as the old/normal code in NsAsyncWrite()! So what could the difference possibly be? Since we're not actually using the async writer thread feature at all, it's even the same thread calling _write() in both cases. LogTrace() is in nslog.dll while NsAsyncWrite() is in libnsd.dll, but I don't see how/why that could possibly matter. Microsoft's docs for _write() and _open() don't mention anything suspicious: http://msdn.microsoft.com/en-us/library/1570wh78%28v=vs.100%29.aspx http://msdn.microsoft.com/en-us/library/z0kc8e3z%28v=vs.100%29.aspx WinDbg thinks the source file used for Microsoft's _write() was: "f:\dd\vctools\crt_bld\self_x86\crt\src\write.c" Naturally I don't have that exact path, but I do have this, which seems to be the same thing: "c:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\crt\src\write.c" For binary files, _write() is a reasonably simple wrapper around WriteFile(), but it's surprisingly complicated for text-mode files. My access log seems to always be file descriptor 3, so in WinDbg I set a breakpoint like this: bu NsAsyncWrite ".if (fd >= 0n3) {} .else {gc}" Naviserver opens its access log in text mode; that's the default, and by stepping through in WinDbg I do see it enter some of the text-mode sections in _write_nolock(). As a little experiment, in LogOpen() I added a _O_BINARY flag, so it opened the access log like this: fd = open(logPtr->file, O_APPEND|O_WRONLY|O_CREAT|_O_BINARY, 0644); That made no difference to the behavior. (I did NOT run that under WinDbg to confirm, but presumably it was in binary mode like I asked.) So this lost writes bug doesn't seem to depend on text vs. binary mode, it's the same for both. So, I have no idea. The only ways I can imagine to try further to track this down would be: 1. See if the bug is replicable in a standalone, non-Naviserver C program. Maybe with two different DLLs involved. 2. Or, in NsAsyncWrite, do not use _write(), instead call WriteFile() directly and see what happens. That sounds like a lot of work though, much more than this really warrants. I'm really glad you found a work-around, Gustaf! I suggest we just leave it as is, but add some comments about why it's there. -- Andrew Piskorski <a...@piskorski.com> ------------------------------------------------------------------------------ _______________________________________________ naviserver-devel mailing list naviserver-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/naviserver-devel
