With min_bytes_to_use_mmap_io=1 memory regions for mmap() will not be
accounted. But this creates issues, since if you will run simple
"SELECT * FROM big_enough_table" it will mmap lot's of files and the
process RSS will grow at some point, and before the kernel will
understand that it needs to unmap physical pages the clickhouse-server
will sync total_memory_tracker with RSS of the process and the
allocations will fail by per-server memory limits.
Also note, that this does not address all the issues with mmap()
regions.
v2: move track/untrack into the MMapReadBufferFromFileWithCache to avoid
data-race [1]
[1]: https://clickhouse-test-reports.s3.yandex.net/23211/6acf6cb9b22cee7db8d1aa523a6208f63e792f7a/functional_stateful_tests_(thread).html#fail1