I believe we have worked out the correct conditions for starting/not starting page cleaners / LRU managers in the latest XtraDB source. Specifically, LRU manager has to be started on innodb_read_only too - it evicts in addition to flushing, and so without it, a --innodb-empty-free-list-algorithm=backoff would starve InnoDB of free pages.

I noticed that I had forgotten to update this ticket when I committed the change to MariaDB 10.0.32, 10.1.25. In MariaDB 10.2 there is no buf_flush_lru_manager_thread.

laurynas, thank you for the hint. It looks like I should revert this change, or otherwise if innodb_read_only in combination with innodb_empty_free_list_algorithm=backoff may cause the buffer pool to run out. I will construct and run a test case.

I tried and failed to reproduce the buffer pool starvation. It turns out that when the buf_flush_lru_manager_thread is not running, buf_LRU_get_free_block() will choose the legacy algorithm:

	if (srv_empty_free_list_algorithm == SRV_EMPTY_FREE_LIST_BACKOFF

	    && buf_lru_manager_is_active

	    && (srv_shutdown_state == SRV_SHUTDOWN_NONE

		|| srv_shutdown_state == SRV_SHUTDOWN_CLEANUP)) {

The condition buf_lru_manager_is_active would not hold, because innobase_start_or_create_for_mysql() will not set the variable unless it is starting the buf_flush_lru_manager_thread.
For the record, I used the following test case:

--source include/have_innodb.inc

--source include/not_embedded.inc

SET GLOBAL innodb_file_per_table=on;

create table t(a char(255),b char(255),c char(255), d char(255) not null unique,

primary key(a,b,c))engine=innodb;

begin;

--disable_query_log

let $n=30000;

while ($n)

{

eval insert into t values($n,$n,$n,$n);

dec $n;

}

--enable_query_log

commit;

-- let $restart_parameters=--innodb-read-only --innodb-buffer-pool-size=20m --innodb-buffer-pool-instances=1

-- source include/restart_mysqld.inc

check table t;

By default, innodb_buffer_pool_instances would be 0, which would be replaced with 8 if the innodb_buffer_pool_size is large enough.
It seems that MariaDB has little or no regression test coverage of the innodb_empty_free_list_algorithm=backoff (which is the XtraDB default in 10.0 and 10.1), because the size of a single buffer pool instance would be smaller than 20MiB by default. 1280 pages (20MiB at innodb_page_size=16k) is the limit that is enforced by innodb_empty_free_list_algorithm_allowed().

That is correct, I had forgotten about the added && buf_lru_manager_is_active condition.