wait_for_read in buf_page_get_low hurts performance

I'm benchmarking 10.4 vs 10.11, using sysbench, and oltp_update_index.
It is a rather large benchmark, with about 23GB data , which is run with

sysbench --db-driver=mysql --range-size=100 --table-size=5000000 --tables=8 --threads=16 --events=0 --warmup-time=5 --time=600 oltp_update_index --rand-type=uniform

I run it "cold", i,e with slow shutdown after the load, and without bufferpool loading, i.e

innodb_fast_shutdown=0

innodb_buffer_pool_load_at_startup=OFF

innodb_buffer_pool_dump_at_shutdown=OFF

One thing I noticed, is a very slow warmup
compared 10.4, where it starts at about 40K tps, 10.11 starts at 2K tps, and reaches normal speed ~40-50K tps only after about 140 seconds. mysqld reads from SSD at 80MB/s in 10.4 at that time, and 8-10 times less in 10.11

Note, this is still mostly in-memory benchmark, the buffer pool is large enough.

I noticed a particular stack, that comes often during warmup, but not after, which I would attribute to that slowness.

NtDelayExecution

RtlDelayExecution

SleepEx

std::this_thread::sleep_until<std::chrono::steady_clock,std::chrono::duration<__int64,std::ratio<1,1000000000> > >

std::this_thread::sleep_for

buf_page_get_low

btr_latch_prev

btr_cur_t::search_leaf row_ins_sec_index_entry_low

row_ins_sec_index_entry

row_upd_sec_index_entry

row_upd_sec_step row_upd row_upd_step

row_update_for_mysql

ha_innobase::update_row

handler::ha_update_row

mysql_update

mysql_execute_command

Prepared_statement::execute

Prepared_statement::execute_loop

mysql_stmt_execute_common

mysqld_stmt_execute

dispatch_command

do_command

....

Note, that Windows' Sleep can't be less that 1ms, and by default, they would still amount to 14ms or something like that.

According to Marko, this sleeping is due to MDEV-33543, something change buffer, and it is not needed in 11.x anymore.
UPD: If I revert 90b95c6149c72f43aa2324353a76f370d018a5ac (MDEV-33543), there is no warmup, performance is much better