Re: fstat st_size on open files on Parallels filesystem is wrong

[lennox@cs.columbia.edu]

On Tuesday, April 22 2014, "Corinna Vinschen" wrote to "cygwin at cygwin.com" saying:

> On Apr 21 14:46, lennox at cs.columbia.edu wrote:
> > On Monday, April 21 2014, "Andrey Repin" wrote to "lennox at
> > cs.columbia.edu, cygwin at cygwin.com" saying:
> > 
> > > Greetings, lennox at cs.columbia.edu!
> > > 
> > > > I’m running cygwin64 1.7.29 in a Windows 8.1 Pro virtual machine, running in
> > > > Parallels Desktop 9.0.24229 on Mac OS X 10.9.2.
> > > 
> > > > Parallels Desktop automatically mounts my Mac OS X home directory as a Z:
> > > > drive in Windows.  Cygwin mount reports this drive as being type "prlsf".
> > > 
> > > > Unfortunately, I've discovered that if I have an open file on this
> > > > filesystem which has been written to, the size returned by Cygwin fstat() on
> > > > the open file is wrong.  A stat() of the file after it's been closed is
> > > > correct.
> > > 
> > > > This has the consequence that emacs always thinks saved files have been
> > > > modified externally, since emacs looks at files' sizes (as well as their
> > > > modification times) to detect external changes.  This makes emacs
> > > > near-unusable.
> > > 
> > > > This problem does not occur for files in my Cygwin home directory, or other
> > > > locations mounted on my Windows C: drive.
> > > 
> > > > I've attached a simple unit test program that illustrates the problem.
> > > > I've also attached my cygcheck -s -v -r output.
> > > 
> > > > Any ideas?  Is this a Cygwin bug, a Parallels bug, or something else?
> > > > Glancing over the Cygwin code, I see that there are a few cases where fstat
> > > > has special cases for certain filesystem types.
> > > 
> > > You never flushing the buffer in your test code, or I'm reading it wrong?
> > 
> > This is using Posix APIs -- open() / write() -- not C APIs, fopen() /
> > fwrite(), so there shouldn't be a buffer?  Notice that the test behaves as I
> > expect for a file on NTFS.
> > 
> > Adding a call to fsync() prior to the fstat() call doesn't change anything.
> 
> This is actually a bad sign.  The problem you're describing occurs on
> NFS, too.  If you write to the file, a subsequent call to fetch stat
> attributes does not return the actual size of the file, but the size at
> the time the handle has been opened.
> 
> However, on NFS, a call to FlushFileBuffers helps to kick stat back into
> shape.  That's the Win32 function called from fsync as well.  What is
> Cygwin supposed to do if that doesn't work?

It's certainly possible that the file system driver in the Parallels Tools
has a bug, and if so I'm happy to report it to Parallels.

I'll see if I can reproduce the test case using just Win32 APIs.  Are there
any particular gotchas I should watch out for, or should just looking at
what Win32 functions are called in fhandler_disk_file.cc be sufficient?


-- 
Jonathan Lennox
lennox at cs.columbia.edu

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple