Segmentation Fault When Creating Cyclic Embedded Messages

[steventang2013]

When adding an object to a repeated field, pyrobuf tosses a seg fault.

Environment:

Python 2.7.10
pyrobuf 0.5.11
libprotoc 2.6.1

Proto definition file:

message AResponse{
        repeated A a = 1;
}

message B {
        required A a = 1;
}

message A {
        required string text = 1;
        optional B b = 2;
}

Output error:

In [8]: a_response = AResponse()
In [9]: a = a_response.a.add()

Segmentation fault: 11

When I compile the exact same definition file with protoc instead of pyrobuf, it runs successfully.

[tburmeister]

Ran under gdb and quit out when the backtrace hit frame 12000 or something...

The underlying issue is that we create empty sub messages on message initialization, even for optional sub messages, as opposed to setting them as None. In this instance, we continue recursively initializing sub messages until we finally explode. I have to think a little about the cleanest solution.

[tburmeister]

Just put up and attached a proposed solution for this problem, which is to lazy init sub messages so we only instantiate them the first time we attempt to access them. I can now do the following:

In [1]: from issue_69_proto import *

In [2]: a = A()

In [3]: a.b
Out[3]: <issue_69_proto.B at 0x104430f50>

In [4]: a.b.a
Out[4]: <issue_69_proto.A at 0x104195bf0>

In [5]: a.b.a.b.a
Out[5]: <issue_69_proto.A at 0x1041958f0>

In [6]: b = B()

In [7]: b.a.b.a
Out[7]: <issue_69_proto.A at 0x1041956b0>

[steventang2013]

Thanks for applying the fix!

Originally, I thought that protobuf lacked cycle support, but that wasn't the case when I compiled it with protoc. Any idea when this fix will be merged?

[tburmeister]

Late response here, but just merged in the fix.