2 RCAS :Allocations in control plane use GPF_ATOMIC because we are working
3 under a spinlock. This is suboptimal. We should define a better locking
4 scheme (most MII/MOI modifications are performed in user context, after
5 a IOCTL). We could define a better locking scheme and revert to GPF_KERNEL
7 * RCAS : Use event management for "external" MOI/MII references: Any
8 protocol, device or module (e.g. tunnel) that holds a reference to a
9 MII/MOI should register for removal notification, and cleanup and put
10 back the reference if they receive a MOI or MII remove event.
12 * RCAS: With the current implementation, &mpls_mii_lock and &mpls_moi_lock
13 are used to synchronize access to the input and output trees. I think
14 that we may need to control access to individual MII/MOI objects too,
15 don't we? : Reader gets a MII pointer locking and
16 unlocking the reader lock that controls the tree, and holds the pointer.
17 Writer locks tree and updates an object. Race? Think bout the locking
18 scheme (per cpu, local irq dis?)
20 * JLEU: bind instructions to name and make them a RTAs in the SHIM_INSTR RTA