`istr` causes memory leak about multidict

[thehesiod]

just came to this same conclusion! Here's a diff between when we didn't see it and when we started seeing it: v2.1.6...v3.0.0

[thehesiod]

note: after fixed please add unittest for leak checking, can use resource.getrusage, psutil or something similar

[thehesiod]

I've been trying to find the leak, and with the following code only istr1 + istr2 leak, so it seems to point to the leak being in the bottom part of istr_new

import asyncio
import multidict
from fbn.utils.profiling import MemTracer


def test():
    istr1 = multidict.istr('foobarbaz')
    istr2 = multidict.istr()
    istr3 = multidict.istr(istr2)
    for _ in range(100):
        istr1.title()
        str(istr1)

async def main():
    tracer = MemTracer(5, '/tmp/traces.txt')

    # warmup
    test()

    while True:
        test()
        tracer.tick()


if __name__ == '__main__':
    _loop = asyncio.get_event_loop()
    _loop.run_until_complete(main())

[gpfei]

PyUnicode_Type.tp_new() is called in istr_new(), but PyUnicode_Type.tp_dealloc() is missed in istr_dealloc().

I tried the code below, it seems to work.

void istr_dealloc(istrobject *self)
{   
    Py_XDECREF(self->canonical);
    PyUnicode_Type.tp_dealloc((PyObject*)self);  // <- new added
} 

[thehesiod]

@gpfei hah! I had tried Py_TYPE(self)->tp_free(obj); and it didn't work, nice find! What's interesting is that I went all the way to trying with valgrind and it didn't find any leaks just now ;)

[thehesiod]

@gpfei I created a PR with your fix. I'll add a testcase after returning from the food trucks. I'm kinda anxious for fix because our prod checkers are failing due to this.

[gpfei]

@thehesiod failing due to this fix?

[thehesiod]

No failing due to the leak

[asvetlov]

Fixed by #106