Hello,

On Saturday 29 August 2009, Andreas Dilger wrote:

quoted

As I understand it, Ted's idea was to avoid problems with the
transactions by doing it after the main work of
unlink()/rmdir()/rename(). I don't know what the better approach is and
don't mind changing it.

This seems reasonable, though in that case it seems it would be much
more efficient to just put the directory inode onto an in-memory list
of directories that have recently emptied a block, and then have a
separate helper thread (or kernel task or whatever) that checks this
list periodically and does freeing of all blocks at the end of the
file that are empty.

But is deleting a huge number of files common enough that the increase in code 
complexity pays off? I'd expect that it doesn't happen often enough to gain 
much. And if it gets more complex, I'd like to have the base functionality 
tested first, so would the patch be acceptable as it is?

Andreas Schlick