Setting up watches takes Exponential time

[Shne]

The initial set up of watches takes exponential time in the number of directories to such a degree it would take approximately 30 hours to set up watches on 100,000 directories on my laptop.

I can't test it right now, but I suspect the culprit is line 1855 in pyinotify.py (under python2) (1842 under python3):

if self.get_wd(rpath) is not None:
    # We decide to ignore paths already inserted into
    # the watch manager. Need to be removed with rm_watch()
    # first. Or simply call update_watch() to update it.
    continue

where it checks if a watch descriptor already exists for the directory.

It makes good sense to do when adding watches to dirs that could have been moved from elsewhere in a recursively watched dir, but in the initial set up of watches, if no moving/creating/deleting is being done, it just slows it down immensely. Should be doable in linear time.

So instead of removing this check altogether (if it is the culprit) I suggest a "static" flag or similar to disable the check on the initial watches setup when it is known that the hierarchy won't change during.

[Shne]

So i tested it now. With those two lines outcommented, setting up watches on about 101,000 folders now takes 2 seconds.

[Shne]

According to the man page for inotify_add_watch ( http://www.kernel.org/doc/man-pages/online/pages/man2/inotify_add_watch.2.html ), calling it on an already watched dir doesn't create a new watch descriptor, it just returns the already set wd. With those two lines commented out, the add_watch function's return array will also include wds for dirs that were already watched. I can't see where the problem lies in that.
Is there any problem in returning all wds, even the ones that were already set?

[seb-m]

You're right this is wrong, fixed in 439565e

Thanks for your help Jan.