David Stevens a écrit :

I did some testing with this and see at least a 20% improvement
without drop.

I agree with Peter's recommended changes (esp. sentinel vs null),
and also the trivial brace indentation  in softirq_delay_exec(),
but otherwise looks  good to me. Nice work.

                                        +-DLS

Still I dont like very much all softirq.c changes. I feel very
uncomfortable to justify one extra call in do_softirq(), and
not very clean interface (stuff about locking, barriers...)

Easy way could be to add a SOFTIRQ but its not very wise.

I was wondering if we could use the infrastructure added in commit
54514a70adefe356afe854e2d3912d46668068e6
(softirq: Add support for triggering softirq work on softirqs.)
But I dont understand how it can works...
(softirq_work_list is feeded, but never processed)

Alternatively, we could use a framework dedicated to
network use, with well defined semantic :

Calling softirq_delay_exec() from net_rx_action(),
from this function, we know if time_squeeze was incremented,
or all netdev_budget consumed, and in this stress case 
try to give the wakeups job to another cpu.