Re: [Bug-tar] listed-incremental broken in 1.25 on Solaris 10

[Markus Duft]

On 02/04/2011 01:51 AM, Paul Eggert wrote:
> On 02/02/11 23:48, Markus Duft wrote:
>> which filesystem is _not_ buggy in this sense?
> 
> I haven't run into the problem myself.  I normally
> use RHEL 5.5 + NFS, or Solaris 10 + NFS, or Ubuntu 10.10
> + ext4.  But I haven't been looking for the problem
> either.
> 
> Do you have a small test case that
> reproduces the problem?  We should add it to the
> GNU tar test cases.  That will help us find out.

the case where this happens on interix is simple - a single file in an archive, 
which gets the wrong date when extracting. On linux, it seems that a very large 
archive, with very many files could trigger the problem, but not sure of that. 
i now know that on linux, it happens on NFS... but badly reproducible.

i yesterday discussed the issue with some other developers from my company; we 
came to the conclusion, that most likely a close would not write timestamps as 
long as there is no data to write (thats why the fsync() solution works), but 
that will only be the case when the OS/FS is fast enough with flushing data. i 
know that windows is sooooo slow, which explains why i hit the issue all the 
time (and why the issue goes away if i wait a sec in a breakpoint before 
setting timestamps). linux/solaris/whatever will likely hit the issue only when 
under high load - is this possible? just my thoughts ;)

markus