-
Improvement
-
Resolution: Unresolved
-
Medium
-
None
-
None
-
None
-
3
-
9223372036854775807
on new kernels ext4 doesn't maintain a list committed transactions, so a new mechanism was introduced in LU-19142 - it uses an rbtree to maintain a collection of transactions with corresponding list of callbacks.
I think this approach has few drawbacks:
1) there are only two transactions at any point in ext4: running and committing, using rbtree for 2 items seems a bit overuse.
2) each osd_thandle now has 24 bytes rb_node struct, but only one per ext4 transaction is used while we have hundreds or thousands of osd_thandle per ext4 transaction
3) a single spinlock is used to register a callback and to run a callback - we actually don't need spinlock to run callbacks
- is related to
-
LU-19142 Restore ldiskfs transaction handling
-
- Resolved
-