Every now and then, especially in IMPORT TABLESPACE tests, we encounter an assertion failure in log_checkpoint():

#7  0x000055cadaf300cb in log_checkpoint ()

    at /mariadb/10.5-MDEV-12353/storage/innobase/log/log0log.cc:1367

1367		ut_ad(oldest_lsn >= log_sys.last_checkpoint_lsn);

(gdb) i lo

oldest_lsn = 93891

flush_lsn = <optimized out>

end_lsn = <optimized out>

do_write = <optimized out>

(gdb) p log_sys.last_checkpoint_lsn
$1 = 93906
(gdb) p log_sys.next_checkpoint_lsn
$2 = 93906
(gdb) p log_sys.last_checkpoint_lsn-oldest_lsn
$3 = 15
(gdb) p log_sys.lsn
$4 = {<std::__atomic_base<unsigned long>> = {_M_i = 121116}, <No data fields>}

The difference is more than the size of FILE_CHECKPOINT record, which is 12 bytes.

If I understand it correctly, the the assertion is reporting that the previous checkpoint was too long ago, and we were not actually crash-safe. This is an indication that a log_free_check() call is missing. I do not see any call to in row0import.cc. The function row_import_for_mysql() is writing a small amount of redo log in trx_undo_assign().

I have the feeling that this is related to either MDEV-12353 changing the redo log format, or MDEV-21534 improving the log write speed.

One affected test is encryption.innodb-checksum-algorithm, but there probably are others.