To use qgroup you need to enable quota first using btrfs quota enable command.
Qgroup is not stable yet and will impact performance in current mainline kernel (v4.14).
The limits are separated between shared and exclusive and reflect the extent ownership. For example a fresh snapshot shares almost all the blocks with the original subvolume, new writes to either subvolume will raise towards the exclusive limit.
The qgroup identifiers conform to level/id where level 0 is reserved to the qgroups associated with subvolumes. Such qgroups are created automatically.
The qgroup hierarchy is built by commands create and assign.
If the qgroup of a subvolume is destroyed, quota about the subvolume will not be functional until qgroup 0/<subvolume id> is created again.
create <qgroupid> <path>
For the 0/<subvolume id> qgroup, a qgroup can be created even before the subvolume is created.
destroy <qgroupid> <path>
If a qgroup is not isolated, meaning it is a parent or child qgroup, then it can only be destroyed after the relationship is removed.
limit [options] <size>|none [<qgroupid>] <path>
If <qgroupid> is not given, qgroup of the subvolume identified by <path> is used if possible.
remove <src> <dst> <path>
The same as assign subcommand.
show [options] <path>
<attr> can be one or more of qgroupid,rfer,excl,max_rfer,max_excl.
Prefix '+' means ascending order and '-' means descending order of <attr>. If no prefix is given, use ascending order by default.
If multiple <attr>s is given, use comma to separate.
The information consists of bytes owned exclusively (excl) or shared/referred to (rfer). There’s no explicit information about which extents are shared or owned exclusively. This means when qgroup relationship changes, extent owners change and qgroup numbers are no longer consistent unless we do a full rescan.
However there are cases where we can avoid a full rescan, if a subvolume whose rfer number equals its excl number, which means all bytes are exclusively owned, then assigning/removing this subvolume only needs to add/subtract rfer number from its parent qgroup. This can speed up the rescan.