bar the commit has following:

--echo # (duplicate) MDEV-20380 Server crash during update

CREATE TABLE gafld (

  nuigafld INTEGER NOT NULL,

  ucrgafld VARCHAR(30) COLLATE UTF8_BIN NOT NULL

           DEFAULT SUBSTRING_INDEX(USER(),'@',1)

);

EXPLAIN UPDATE gafld SET nuigafld = 0 WHERE nuigafld = 10;

EXPLAIN UPDATE gafld SET nuigafld = 0 WHERE nuigafld = 10;

DROP TABLE gafld;

it uses DEFAULT expression

and another one has DEFAULT (REVERSE(SUBSTRING_INDEX(REVERSE(user()), '@', 1)))

bar I have managed to make the leak fix, it's here https://github.com/MariaDB/server/commit/6621be137f1807303c967329f536231a510ee3c0

IMHO all approach to set arena before fixing (i.e. prepare phase is incorrect)

All works like this:

1. set arena to some permanent (statement/table etc.)
2. make parse/allocate
3. reset arena to run-time arena
4. make prepare
5. make execute and so on...

So here it should be the same.

When items during prepare need to make some permanent changes they refer thd->stmt_arena, maybe you should assign it, but also there are other tests (leke if it is prepared statements) so it is better to check all such cases.

Also I think Field::set_default() arena switching is a memory leak.

OK to push

nikitamalyavin you made expr_arena usage in fix_all_session_vcol_exprs(). Did you care that it is memory leak until table flush? I'm just preoccupied with similar problem of MDEV-25672 but I don't like continuous allocations on TABLE::mem_root.

The proper fix is to make expr_arena a separate mem_root and reallocate it each time of expr refresh. But this is a lot of various problems like delayed insert, items clone (with String::copy() for all properties), cached_table cleanup, proper name resolution context and what else (no one can tell the complete list).