[DISPATCH-1962] Make LSan suppressions more targeted and specific (within reason)

• Type: Test
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.15.0
• Fix Version/s: 1.16.0
• Assignee: Jiri Daněk
• Reporter: Jiri Daněk
• Created: 14/Feb/21 12:40
• Updated: 14/May/21 14:09
• Resolved: 27/Apr/21 06:55

Description

There are some unfortunate suppressions in tests/lsan.supp at this moment:

leak:*libwebsockets*

This is way too broad. It suppresses leaks in Dispatch files, since it matches e.g. src/http-libwebsockets.c.

The stars at the beginning and end are actually assumed implicitly. If you do not want substring match, you have to do ^foo$. LSan suppression format is simplistic, very unlike Valgrind's.

leak:run_unit_tests.c

Same thing, any leaks revealed by running unit_tests get suppressed. This suppression suppresses all leak traces that include run_unit_tests.c anywhere in the stack. What't the point of running such tests under leak detector, then?

leak:run_unit_tests.c
leak:^libqpid-proton.so$

Same thing. The patterns suppress all leaks that include Python (or Proton) anywhere in the stacktrace. That means there are huge blind spots where dispatch leaks can hide. This is a weakness of the lsan.supp syntax (Valgrind suppressions can be much more targeted and discerning).

Python leaks

Leaks are known and there is ongoing effort to fight them: https://bugs.python.org/issue1635741 (https://bugs.python.org/issue25302) and https://www.python.org/dev/peps/pep-3121

Here's valgrind suppression file from somebody who actually investigated the Python leaks and identified the harmless ones: https://github.com/libgit2/pygit2/blob/master/misc/valgrind-python.supp

One example of a hidden leak in dispatch, which is revealed by making Python suppressions more targetted:

9: Direct leak of 56 byte(s) in 1 object(s) allocated from:
9:     #0 0x7f78a3606e8f in __interceptor_malloc (/nix/store/g40sl3zh3nv52vj0mrl4iki5iphh5ika-gcc-10.2.0-lib/lib/libasan.so.6+0xace8f)
9:     #1 0x7f78a2d64afb in qd_malloc ../include/qpid/dispatch/ctools.h:229
9:     #2 0x7f78a2d657da in qdr_core_subscribe ../src/router_core/route_tables.c:149
9:     #3 0x7f78a2c83072 in IoAdapter_init ../src/python_embedded.c:711
9:     #4 0x7f78a2353a6c in type_call (/nix/store/r85nxfnwiv45nbmf5yb60jj8ajim4m7w-python3-3.8.5/lib/libpython3.8.so.1.0+0x165a6c)

The problem is in

class Agent:
    ...
    def activate(self, address):
        ...
        self.io = IoAdapter(self.receive, address, 'L', '0', TREATMENT_ANYCAST_CLOSEST)

IoAdapter refers to Agent (through the bound method reference self.receive) and Agent refers to IoAdapter (through property self.io). Since IoAdapter is implemented in C and does not implement support for Python's cyclic GC, there is no way to break the cycle.

Heap dump in attachment. The bound method is at the top of the picture. (Ignore the Mock objects, I was trying to simplify the picture while not getting crashes due to too much meddling).

Random observations

It is possible to build special Debug build of Python, which has tools to detect leaks, asserts to prevent negative refcounts, etc. https://pythonextensionpatterns.readthedocs.io/en/latest/debugging/debug_python.html#debug-version-of-python-memory-alloc-label

Use the following to detect python leaks (instead of valgrind)
https://docs.python.org/3/library/tracemalloc.html

Use https://pypi.org/project/objgraph (with graphviz) to view heap object trees. The following renders the picture as a png under /tmp and prints the path to stdout.

int ret = PyRun_SimpleString("import objgraph; objgraph.show_backrefs(config.global_agent, max_depth=10)\n\n");
PyErr_PrintEx(0);
assert(ret == 0);

Tools

CPyChecker

https://dmalcolm.fedorapeople.org/presentations/PyCon-US-2013/PyCon-US-2013-dmalcolm-StaticAnalysis.html#(38)
https://emptysqua.re/blog/analyzing-python-c-extensions-with-cpychecker/
https://nedbatchelder.com/blog/201502/cpychecker.html

Pungi

http://www.cse.psu.edu/~gxt29/papers/refcount.pdf

Use CFFI or similar, instead of writing a C module.

Issue Links

• causes: DISPATCH-2098 Crash when ctrl-c on router that has a libwebsockets listener connected to a console (2)
• causes: DISPATCH-2109 Shutdown crash: member access within null pointer of type 'struct qdr_node_t'
• contains: DISPATCH-848 Direct leak of 48 byte(s) in 1 object(s) allocated from qdr_core_subscribe in router_core/route_tables.c:149
• contains: DISPATCH-1934 Dispatch should not ignore all Proton memory leak reports wholesale
• links to: GitHub PR DISPATCH-1962 Make LSan suppressions more targeted and specific #1048