Created attachment 196212 [details]
Fix locking issues around pf_map_addr.
I came across the following issues with pf_map_addr:
- It operates on unlocked src_nodes and I've seen kernel crashes attributable to this.
- For route-to targets which specify outgoing interface, interface selection happens outside of pf_map_addr and can be taken from different redirection target than was used for IP address selection.
- Hash used for locking and unlocking src_node is calculated multiple times for same src_node.
- Operations on rule->pool are unlocked, there even is a comment in code mentioning circumstances where it can cause kernel crash.
I attached a patch addressing those issues. Please note it is early version, it compiles, it runs, I'm not sure if it is really correct. The patch does the following:
- Struct pf_pool has a lock now. Round-robin algorithm in pf_map_addr locks the pool. Other algorithms don't do it, I don't think they need. Should we also have table operations check this lock?
- pf_find_src_node does not manipulate src_node's state counter. It is a "find" function so why would it manipulate other stuff? Also with new location of sn->states++ removing of unused src_node in pf_create_state will work, I'm pretty sure it was broken before.
- pf_insert_src_node returns locked. The lock is retained across pf_create_state and pf_map_addr.
- pf_insert_src_node returns hash (row) of locked src_node. It is used later for unlocking without recalculating the hash.
- Interface chosen by pf_map_addr is assigned to external variable from within this function. Struct src_node stores the interface.