Re: [coreutils] Re: Bug#609049: du: process_file: Assertion `level == prev_level - 1' failed

[Pádraig Brady]

On 09/01/11 22:05, Jim Meyering wrote:
> diff --git a/tests/du/move-dir-while-traversing 
> b/tests/du/move-dir-while-traversing

> +# We use a python-inotify script, so...
> +python -m pyinotify -h > /dev/null \
> +  || skip_ 'python-inotify package not installed'

A small point is that error is a bit fedora specific.
The package is python-pyinotify on debian/ubuntu for example.

> +
> +# Move a directory "up" while du is processing its sub-directories.
> +# While du is processing a hierarchy .../B/C/D/... this script
> +# detects when du opens D/, and then moves C/ "up" one level
> +# so that it is a sibling of B/.
> +# Given the inherent race condition, we have to add enough "weight"
> +# under D/ so that in most cases, the monitor performs the single
> +# rename syscall before du finishes processing the subtree under D/.
> +
> +cat <<'EOF' > inotify-watch-for-dir-access.py
> +#!/usr/bin/env python
> +from pyinotify import *

I generally don't include everything from a module,
to keep the namespace clean.
I'd do instead:

import pyinotify as pn
import os,sys
...


> +dir = sys.argv[1]
> +dest_parent = os.path.dirname(os.path.dirname(dir))
> +dest = os.path.join(dest_parent, os.path.basename(dir))
> +
> +class ProcessDir(ProcessEvent):
> +
> +    def process_IN_OPEN(self, event):
> +        os.rename(dir, dest)
> +        sys.exit(0)
> +
> +    def process_default(self, event):
> +        pass
> +
> +wm = WatchManager()
> +notifier = Notifier(wm)
> +wm.watch_transient_file(dir, IN_OPEN, ProcessDir)
> +print 'started'

The above print is buffered by default.
As we're using it for synchronization I'd

sys.stdout.write('started\n')
sys.stdout.flush()

> +notifier.loop()
> +EOF
> +chmod a+x inotify-watch-for-dir-access.py
> +
> +long=d/e/f/g/h/i/j/k/l/m/n/o/p/q/r/s/t/u/v/w/x/y/z
> +t=T/U
> +mkdir -p $t/3/a/b/c/$long d2 || framework_failure
> +timeout 6 ./inotify-watch-for-dir-access.py $t/3/a/b > start-msg &
> +
> +# Wait for the watcher to start...
> +nonempty() { test -s start-msg && return 0; sleep $1; }
> +retry_delay_ nonempty .1 5

I think the above may iterate only once?
How about:

nonempty() {
  test -s start-msg || { sleep $1; return 1; }
}
retry_delay_ nonempty .1 5 || framework_failure

> +
> +# The above delay is insufficient in ~50% of my trials.
> +# Sometimes, when under heavy load, a parallel "make check" would
> +# fail this test when sleeping just 0.1 seconds here.
> +sleep 1

Hopefully this extra sleep is not required now

Nice fix BTW!

cheers,
Pádraig.