[Gluster-devel] help needed with glfsheal

[Emmanuel Dreyfus]

Replying to myself:

The problem happens because of negative long offsets returned by telldir()
that get cast into 64 bit types. The upper higher bits get set to 
0x7fffffff or 0xffffffff depending on signedness (we have signed off_t
and unsigned uint64_t randomly used in the code). The difference on the
most significant bit leads itransform/deinstransform in afr xlator to 
mix offsets between bricks. How unlucky.

This is fixed in http://review.gluster.org/9071

On Wed, Nov 05, 2014 at 02:48:00PM +0000, Emmanuel Dreyfus wrote:
> Hello
> 
> I have been investigating a spurious failure in tests/basic/afr/self-heald.t
> for NetBSD.  It happens in the test that ongoing I/O  is not considered 
> as Pending heal: sometimes I get entries listes by gluster volume heal info
> while they should not.
> 
> Looking at the logs, I see the lovely message "seekdir(...) failed (...)
> Invalid argument (offset reused from another DIR * structure?)" from the 
> index xlator. This is because an offset is reused from a DIR * to another
> one.
> 
> That problem was supposed to be fixed here:
> http://review.gluster.org/8936
> 
> I first rhought about the bug I fixed recently: there is such a spurious
> message when we hit the end of directory. I fixed it in the posix xlator
> but not the index xlator. I added the same fix:
> http://review.gluster.org/9047
> 
> But I still have the same message. Adding more debug messages, I see that
> brick2 complains about the offset reported at EOF for brick1. This means
> that somewhere cli-side  there is a omponent that keeps state between 
> opening of the same directory for different breicks. Anyone has an idea
> where the offending code could be?
> 
> I guess it may impact Linux too, despite the absence of message in the 
> logs, as readdir may start skipping the beginning of the directory because
> offset is not reset to 0 when oving from a brick to another one.
> 
> 
> 
> -- 
> Emmanuel Dreyfus
> manu at netbsd.org
> _______________________________________________
> Gluster-devel mailing list
> Gluster-devel at gluster.org
> http://supercolony.gluster.org/mailman/listinfo/gluster-devel

-- 
Emmanuel Dreyfus
manu at netbsd.org