Repeated checks for camera causes memory leak about python-gphoto2

[jim-easterbrook]

Sorry but I don't have time to look into this now, or for the rest of the weekend. Can you wait a day or two?
PS The list-cameras.py example shows another way to detect a connected camera.

[flyte]

Yep, sure. I was hoping I was doing something obviously wrong though :)

[jim-easterbrook]

Memory leaks in SWIG generated Python interfaces can be a bit hard to track down, especially with a library that has its own reference counting stuff. Is this happening with Python2 or Python3 or both?

[flyte]

It's happening in Python 2 and 3. Here's a far simpler program to reproduce it (run without any cameras plugged in):

import gphoto2 as gp

context = gp.gp_context_new()
while True:
    camera = gp.check_result(gp.gp_camera_new())
    gp.gp_camera_init(camera, context)

It seems to be caused by gp.gp_camera_init(camera, context) as it doesn't leak without that line.

[jim-easterbrook]

When I run your delightfully short script I get a lot of messages swig/python detected a memory leak of type 'Camera *', no destructor found. if I'm using the PYTHON_GPHOTO2_NO_BUILTIN=1 build. With the normal build I don't get the error messages but still have the leak. I probably won't have time to sort this out today, but I'm confident of tracking it down tomorrow.

[jim-easterbrook]

Ignore that - the messages were because I wasn't correctly installing the PYTHON_GPHOTO2_NO_BUILTIN=1 version. When installed correctly the -builtin option makes no difference to the memory leak and generates no error messages.

[jim-easterbrook]

I've tried adding some gp_camera_unref calls to the Python interface to gp_camera_init but it makes no difference. I suspect the leak may be in libgphoto2 itself. I found this discussion which suggests there was a problem in version 2.4. http://gphoto-software.10949.n7.nabble.com/Memory-Leaks-with-libgphoto2-4-5-td8318.html

[flyte]

Ah, well that's interesting. Sorry if I've wasted your time! I suppose the way forward here is to document a workaround in the examples. I'll see if I can come up with something, but as far as I'm aware, whatever you do with a new camera object, it calls gp_camera_init first. I'll try a few things out this morning.

[jim-easterbrook]

No time wasted - I've learnt a bit more about libgphoto2. It appears the leak is caused if you call camera.init without having first called camera.set_port_info to select the camera.
You could try something like this:

context = gp.Context()
camera_list = []
for name, addr in context.camera_autodetect():
    camera_list.append((name, addr))
if camera_list:
    camera = gp.Camera()
    port_info_list = gp.PortInfoList()
    port_info_list.load()
    idx = port_info_list.lookup_path(camera_list[0][1])
    camera.set_port_info(port_info_list[idx])
    camera.init(self.context)

I've not tried this exact code (I adapted it from another project where I allow the user to choose a camera) but it should give you the idea.

[flyte]

So far I've narrowed the leak (if there's only one?) down to gp_port_info_list_load, as can be seen with the following:

import gphoto2 as gp

while True:
    pil = gp.check_result(gp.gp_port_info_list_new())
    gp.gp_port_info_list_load(pil)  # Leaks

and indeed just context.camera_autodetect() itself leaks, presumably because it calls gp_port_info_list_load:

import gphoto2 as gp

context = gp.gp_context_new()
while True:
    context.camera_autodetect()  # Leaks

Going further down the rabbit hole :)

[flyte]

I'm reaching the limit of my ability to track this down now. I'm not sure there's a way we can get around it without the issue being fixed upstream. There's a bug report already at gphoto/libgphoto2#165.

I guess I'll just have to run my code in a bash while loop and exit when a camera is not found/disconnects :(

[jim-easterbrook]

I agree, it's not something I can do anything about in the Python interface. Please reopen this bug if you disagree.

[AntiSol]

Hi Jim, me again. I'm seeing a similar issue to this one.

memleak.py:

import gphoto2 as gp
c = gp.Camera()
while True:
	cfg = c.get_config()

while running this, if you do something like

while true; do ps -eo pid,args,rss | grep "python [Mm]emleak.py" ; sleep 1; done

You'll see rapidly increasing memory usage. Note that you probably don't want to run that python code for very long unless you like seeing your system start swapping like crazy.

I think this is probably an issue with the library that I'm not going to be able to workaround without debugging C (which is beyond my abilities), I think it might be related to gphoto/libgphoto2#165 and gphoto/libgphoto2#208. Would you agree? or is this a different issue you might be able to help me track down?

the device I'm working with has limited memory, so if I can't find a fix or workaround for this it pretty much kills my project. I'd appreciate any suggestions.

[jim-easterbrook]

It's possible this is a bug in the Python interface, as it's not to do with the camera autodetect stuff this bug is about. The whole config tree system is a bit of a nightmare, with nodes allocated in C that may or may not get properly freed when the Python object is deleted. I'll investigate further.

[jim-easterbrook]

I've established that this is a bug, and have opened #67 for it.

[AntiSol]

thanks, I appreciate it.