On Sat, 2015-08-29 at 13:45 -0700, David Miller wrote:

While reviewing this I noticed that the gro cells code goes through all
of the trouble of allocating full per-cpu objects to manage the packet
queuing and processing, but then it uses a full spinlock for protection.

It ought to be sufficient to just block out BH processing or, at worst,
local cpu interrupts, to protect the individual per-cpu cells.

Original commit envisioned a possible improvement of gro_cells in case a
NIC receives all GRE traffic on a single RX queue.

(If for example Toeplitz hash RSS looks at outer IP header, and all
traffic is encapsulated with a constant tuple)

Idea was to add a mode where we could select a gro cell not based on
current cpu number but a hash based on flow software flow dissection.

If we do not plan to add this feature, we certainly can remove the
spinlocks now we have per cpu cells.