Bug #20718
closedObjects created with Data_Make_Struct and the default free function are not freed
Description
I discovered a memory leak when using the FFI gem prior to version 1.16 and Ruby 3.3 and up.
During debugging I found that this earlier version of FFI uses Data_Make_Struct (https://github.com/ffi/ffi/blob/v1.15.5/ext/ffi_c/Pointer.c#L57) instead of TypedData_Make_Struct and it uses -1 as the free function, which is RUBY_DEFAULT_FREE
When the object goes to get garbage collected, it enters into rb_data_free and it is passed to the RTYPEDDATA_EMBEDDED_P macro even though it is not of RTypedData. Because of that, the conditional is evaluated to false and xfree is never called. This was discovered by using jemalloc leak detection.
I have attached a somewhat minimal replication of the issue. The fix would appear to check the type of the obj before casting it.
Files
Updated by jcalvert (Jonathan Calvert) 5 days ago
ยท Edited
I have added a pull request that should patch the issue. https://github.com/ruby/ruby/pull/11563 - I was able to build Ruby and run it against my example and it does work.
Presumably this would be a candidate for a backport :)
Updated by jcalvert (Jonathan Calvert) 5 days ago
- Status changed from Open to Closed
Applied in changeset git|c1a510a8dffa1c8065e47697cd57edae67126712.
[Bug #20718] Free non-RTypedData
objects
Allow objects that are not of type RTypedData
to use the default
free function, as RTYPEDDATA_EMBEDDED_P
can return a false positive
when casting non-RTypedData
objects.
Updated by byroot (Jean Boussier) 4 days ago
- Backport changed from 3.1: UNKNOWN, 3.2: UNKNOWN, 3.3: UNKNOWN to 3.1: DONTNEED, 3.2: DONTNEED, 3.3: REQUIRED
I think this issue was introduced in 3.3? Let me know if not and I'll update the backport target.
Also for 3.3 the branch maintainer appreciate backport PRs so don't hesitate to open it yourself and tag k0kubun if you wish.