SIGSEGV when running tracebacker on loaded workers

[anthonyrisinger]

DETAILS

per discussion on ML, i've found a repro case. this uses the dumbloop, as it matrches the exact problem i'm dealing with (and makes it easy to keep worker loaded).

example simulates a loaded AMQP worker by publishing to the loopback on an artificial memory:// transport (see queue_worker.py for more detail), and then rapidly connects to the tracebacker using a socat busy loop; should segfault in approx 5-10 seconds.

REPRODUCE

1. get uWSGI HEAD
2. add dumbloop to build
3. build uWSGI
4. get Kombu and deps
5. run queue_worker.py
6. die in ~5s :)

COMMANDS

export PIP_REQUIRE_VIRTUALENV=
export PYTHON=$(type -p python2.7 python2.6 python2 python | head -n1)

git clone --depth 1 https://github.com/unbit/uwsgi __bug &&
cd __bug &&
sed -i 's/python/dumbloop,&/' buildconf/default.ini &&
"$PYTHON" uwsgiconfig.py --build &&
"$PYTHON" -m pip install --target eggs kombu &&
wget https://raw.github.com/gist/3934532/c9c1e3b9822e006ae0ce3bcc895fa07be5ef1cf7/queue_worker.py &&
(

    ./uwsgi --master --pp=eggs --loop=dumb --dumbloop-code=queue_worker.py \
        --dumbloop-function=main --enable-threads --py-tracebacker=trace &
    sleep 3
    while socat UNIX-CLIENT:trace1 FILE:/dev/null; do :; done
    kill -INT %%

)

[unbit]

Ok, i have managed to reproduce it, let's see what i will find...

[unbit]

hi, it should be fixed now (and should not leak memory anymore)

[anthonyrisinger]

damn, i was pretty close ... i know i missed the Py_INCREF for sure. anyways, i can confirm fixed by 15 minutes of non-crashing... sooo LGTM!

thanks Roberto!

[anthonyrisinger]

oh btw if you drop a #29 (or really, #ISSUE) github will link issues back to commits... just FYI in-case you didn't know :) makes it nice for reference later on.