O_DIRECT + nocow + no snapshots + noatime + lazytime + prefilled writes should be fast!

[socketpair]

Suppose case:

O_DIRECT
nocow set (chattr +C)
no snapshots on the whole FS
Mounted with noatime,lazytime
prefilled file (i.e. dd from urandom)
No reflink copies

App does radom aligned O_DIRECT writes.
For my understanding, these writes should be as fast as writing to raw blockdev. In Ext4 it is.

In BTRFS it gives more than 2 times less iops.

This should be fixed. At least with restrictions (for example swapfile requires some). Yes, making snapshots whil at least one such file is open should be declined. The same with reflink.

Possibly, a new attribute to work with a file as with swapfile.