This is the mail archive of the cygwin mailing list for the Cygwin project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: untarring symlinks with ../ fails randomly, silghtly OT


On Jul  4 12:46, Corinna Vinschen wrote:
> On Jul  4 11:15, Wolf Geldmacher wrote:
> > As an aside:
> > 	I also used to have some trouble with "rm -rf" of a directory
> > 	hierarchy failing more or less reproducibly (like: 80% of the
> > 	time) because files were presumably still "in use". Repeating
> > 	the command several times would succeed, though.
> > 
> > 	Downgrading from cygwin1.dll/1.7.9.1 to cygwin1.dll/1.7.8.1
> > 	seems to have solved that issue as well - still have to see
> > 	the first "retry to delete".
> > 
> > This may or may not be related to the original report, as it also reeks
> > of a race condition during file/directory operations.
> 
> I can neither reproduce the tar problem, nor can I reprocude the rm
> problem.  I tried this under 2008R2 which is basically the same as your
> W7-64 bit.  I used local and remote drives to test the issue but to no
> avail.

Finally I managed to reproduce the problem and now I see what happens.

Windows does not write back the file change timestamp unless the file
buffers are flushed.  This usually occurs at close time.  In contrast to
POSIX specifications the timestamps are *not* automatically updated when
a call to fetch file metadata is performed.

Here's what tar does when creating the symlink:

  1. create file with 000 permissions
  2. fstat
  3. close file
  [...]
  4. stat file
  5. if fstat.st_ctime != stat.st_ctime ==> symlink placeholder has been
     overwritten.

The problem is that the call to fstat on the opened handle gets some
value of the change time timestamp, but the subsequent close changes
the timestamp again.

Speculation: It seems that the timestamp fstat sees is the timestamp
created at the time NtCreateFile is called, while the timestamp from the
call to NtSetSecurityFile to change the DACL is cached and only updated
when calling NtClose.

This also explains why this doesn't occur in 1.7.8.  In 1.7.8, the DACL
has been written using another file handle, because the original handle
didn't have the right to change the DACL.  By adding the WRITE_DAC flag,
I allowed Cygwin to use the original file handle to write the DACL.  The
difference is:

1.7.8:

  - create file
  -   open file for writing the DACL
  -   write DACL
  -   close
  - do whatever the orignal handle was opened for
  - close

1.7.9:

  - create file
  - write DACL
  - do whatever the orignal handle was opened for
  - close

So, with 1.7.9 the close call after writing the DACL is missing, which
accounts for the missing flushing of the file metadata.

By calling FlushFileBuffers in fstat before calling NtQueryInformationFile
I can fix the problem.  Unfortunately that slows down applications like tar,
which use fstat a lot, a lot.

There are two solutions, one is reverting to the 1.7.8 state, which
means, writing the DACL requires to open the file again, or calling
FlushFileBuffers in fstat.
I compared both solutions.  On my hardware, calling tar xzf on your file
is 500% slower if fstat calls FlushFileBuffers compared to just dropping
the WRITE_DAC flag from the open call.  Wow!  Imagine that I added the
WRITE_DAC flag to gain performance...

So I guess this all boils down to the fact that adding WRITE_DAC was
not really a good move.  It's a shame that Windows punishes every try
to speed up file operations with a raise in non-POSIXy behaviour :-(((

I changed that in CVS and right now I'm generating a new developer
snapshot on http://cygwn.com/snapshots/.  Give it a try, please.


Thanks,
Corinna

-- 
Corinna Vinschen                  Please, send mails regarding Cygwin to
Cygwin Project Co-Leader          cygwin AT cygwin DOT com
Red Hat

--
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


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]