Back when we were in the process of moving from the sophisticated combination of locks to protect the inpcb database to the epoch there was a serious concern. The concern was that the epoch is great for stuff that doesn't change all the time - interfaces, addresses, etc, but if we use epoch for connections, then its garbage collector potentially can grow really long purge lists. Back then myself and Jeff came with a plan and a promise. Jeff was already sketching the SMR and promised it to be ready soon. And I promised to use SMR for the inpcbs. As part of the work the KPI between the pcb database and rest of the stack is reviewed and made more solid.
- The database is already build on CK lists.
- SMR section makes it safe to access a pcb after it went through uma_zfree(). We always want a particular pcb locked, so after successful lookup and lock we just check INP_FREED.
- Transitioning from SMR section to an rwlock is sophisticated, but this all is hidden from normal KPI users.
- List traversals (e.g. netstat or raw sockets) are now done with an iterator-like KPI, that also hides all SMR complexity.
Despite added complexity in SMR-to-rwlock transitions, the overall change makes the code simpler to understand.
Note. This change requires PCBGROUP to be retired.
The full branch, that has some preparatory changes, removal of PCBGROUP and rewrite of TCP_HPTS referencing model of inpcbs can be found here: https://github.com/glebius/FreeBSD/commits/inpcb