Memory leak on compiling method call with kwargs
The following code produces a memory leak:
It comes from the allocation in
struct rb_callinfo_kwarg *kw_arg = rb_xmalloc_mul_add(len, sizeof(VALUE), sizeof(struct rb_callinfo_kwarg));
Later it's packed into
struct rb_callinfo, but
imemo_callinfo is a GC-managed object that has no
free calls in
obj_free function in gc.c.
I've tried to fix it by calling
callinfo->kwarg and it fixed leak in
./miniruby -e 'p(foo: 1)', but it breaks
make install. My addition causes a double-free error, looks like either
callinfo->kwarg is shared between multiple objects.
kwarg field is a
const pointer, so that's somewhat expected (I was under impression that
const qualifier is redundant) :)
I'm pretty sure old versions of Ruby are also affected by this memory leak.
Updated by jeremyevans0 (Jeremy Evans) about 1 year ago
- Assignee set to ko1 (Koichi Sasada)
I did some testing and can confirm this is still a bug in the master branch. From my testing, this isn't related to not freeing the
struct rb_callinfo_kwarg, it's because compiling a keyword argument method call causes two imemo_callinfo objects to be created, and neither of these is collected unless the method they are calling is removed. Maybe there is a separate bug in regards to not freeing
Here's an example reproducer. This compiles and runs code with 2 keyword argument method calls in a loop with 30,000 iterations. Each 10,000 iterations, it prints the number of T_IMEMO objects after a full garbage collection.
def p(**) end s = 'p(foo: 1); p(foo: 1)' 30000.times do |i| eval(s) if i % 10000 == 0 GC.start puts(ObjectSpace.count_objects[:T_IMEMO].to_s) end end GC.start puts(ObjectSpace.count_objects[:T_IMEMO].to_s) Object.send(:remove_method, :p) GC.start puts(ObjectSpace.count_objects[:T_IMEMO].to_s)
Here's the output, showing the
T_IMEMO objects increasing at ~40,000 per print (2 for each of the 2 keyword method calls, 10000 times), and being retained until the related method is removed:
5003 45007 85007 124988 4993
My uneducated approach to fix this would be to tie imemo_callinfo to the containing iseq and not the method definition. That way, if the instruction sequence is collected, any imemo_callinfo inside it are collected with it. Theoretically, imemo_callinfo is caller information, not callee information, and should be tied to callers (the iseq), not callees (the method). However, I don't know whether that approach is viable, and haven't attempted it yet.
I was also able to determine that this imemo_callinfo issue was introduced in Ruby 3.0, since running the above example on Ruby 2.7 shows that the number of T_IMEMO objects remained roughly the same. I bisected and this bug was introduced in b9007b6c548f91e88fd3f2ffa23de740431fa969 (
Introduce disposable call-cache.).