Consider a test case:

heap_t heap;

my_class_t *c = NULL;

{

  c = new (heap) my_class_t();

  c->use();

  c->cleanup();

  // delete (heap) c;

}

c->use();

ASAN should bark on second c->use() but in general it won't because heap memory won't freed. It's partially solved with Sql_alloc but for classes only and not for stuff like THD::make_clex_string(). And it doesn't work at all for mem_heap_t.

We may store size_t with allocation length right before allocated chunk. This will serve as a redzone too.

dealloc_root() and mem_heap_dealloc() can be added. They will read allocated size from redzone and poison freed chunk. No real memory freeing.

This is simple to add but to make it really work those functions should be added to a lot of places in code. Like manual allocation pairs malloc()/free(). But I think one goal was to not call free() on every allocated chunk. Anyway, this is a C style.

C++ style is to create (in terms of C++17) a memory resources with will actually own memory. Examples are bulk allocator, thread-safe bulk allocator, static storage + heap. And allocator is a handler (pointer) to that resource. In that case poisoning of deallocated memory will happen in allocator_t::deallocate().

kevg, is this still applicable after MDEV-17797 got fixed?

marko, yes. I even have a PR for this: https://github.com/MariaDB/server/pull/954
One more patch is needed for InnoDB.