MDEV-36345: Memleak on shutdown in acl_load_mutex test

[kjarir]

Problem

In plugin_add(), init_alloc_root() was called unconditionally after plugin_insert_or_reuse(), even when my_hash_insert() failed and the plugin state was set to PLUGIN_IS_FREED.

This mem_root was never freed because plugin_del() only runs free_root() on plugins that go through the normal deletion path. Plugins immediately marked PLUGIN_IS_FREED skip this path entirely, causing the leak detected by LeakSanitizer during the roles.acl_load_mutex test.

Fix

Guard init_alloc_root() inside the else branch of my_hash_insert() so it is only called when plugin registration succeeds.

Testing

Verified with ASAN+UBSAN build (clang, Debug):

• roles.acl_load_mutex-5170 test passes
• No LeakSanitizer report after fix

[gkodinov]

Thank you for your contribution! This is a preliminary review.

This is an interesting analysis! I have one question:

How come my_hash_insert() can fail at that point?

The code does the following:

• takes a lock guarding the plugin hash
• looks for the plugin into plugin hash, fails if found
• constructs a new plugin hash entry
• tries to insert it while holding the lock

Looking at my_hash_insert() Inserting that new hash entry can fails because:

• the hash has an entry with the same name
• there's no more memory

I fail to understand how my_hash_insert() can fail due to a duplicate entry: it's just been checked under the same lock hold.
Yes, it can fail due to lack of memory. But then you'd have bigger issues.

And now the formal reason: I've tried running the test mentioned under asan/msan, but I do not get any issues. Neither does buildbot. Can you please think of a test case that would expose this problem?
Frankly, I can't at this point. Hence my question above.

[kjarir]

Thank you for the detailed review @gkodinov These are valid points. My ASAN trace was on macOS (Apple Clang, ARM64) which may behave differently from Linux. I'm investigating further to find a proper reproducer on Linux and will update the PR once I have a clearer picture.