}
}
+#define CPUMAP_BATCH 8
+
static int cpu_map_kthread_run(void *data)
{
struct bpf_cpu_map_entry *rcpu = data;
* kthread_stop signal until queue is empty.
*/
while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
- unsigned int processed = 0, drops = 0, sched = 0;
- struct xdp_frame *xdpf;
+ unsigned int drops = 0, sched = 0;
+ void *frames[CPUMAP_BATCH];
+ int i, n;
/* Release CPU reschedule checks */
if (__ptr_ring_empty(rcpu->queue)) {
sched = cond_resched();
}
- /* Process packets in rcpu->queue */
- local_bh_disable();
/*
* The bpf_cpu_map_entry is single consumer, with this
* kthread CPU pinned. Lockless access to ptr_ring
* consume side valid as no-resize allowed of queue.
*/
- while ((xdpf = __ptr_ring_consume(rcpu->queue))) {
+ n = ptr_ring_consume_batched(rcpu->queue, frames, CPUMAP_BATCH);
+
+ local_bh_disable();
+ for (i = 0; i < n; i++) {
+ struct xdp_frame *xdpf = frames[i];
struct sk_buff *skb;
int ret;
ret = netif_receive_skb_core(skb);
if (ret == NET_RX_DROP)
drops++;
-
- /* Limit BH-disable period */
- if (++processed == 8)
- break;
}
/* Feedback loop via tracepoint */
- trace_xdp_cpumap_kthread(rcpu->map_id, processed, drops, sched);
+ trace_xdp_cpumap_kthread(rcpu->map_id, n, drops, sched);
local_bh_enable(); /* resched point, may call do_softirq() */
}