A backlog queue is used to buffer incoming traffic and its length is controlled by net.core.netdev_max_backlog. On fast connections (10Gb/s or higher) the default backlog length of 1024 can be insufficient; here the backlog length is increased if 1/16 of current backlog size in the last minute is dropped (drops occur when the backlog limit is reached). In addition, backlog drops can avoid small flows; the tunable net.core.flow_limit_cpu_bitmap can be used to set this on a per-cpu basis; when we see sufficient drops on a CPU, the appropriate bit is set in the CPU bitmask to prioritize small flows for drop avoidance.

When NAPI polls to handle multiple packets, the number of packets is limited by net.core.netdev_budget while the time is limited by net.core.netdev_budget_usecs. If we hit the limit of number of packets processed without using the usecs budget the time_squeezed softnet stat is bumped; if we see increases in time_squeezed, bump netdev_budget/netdev_budget_usecs.

However, we want to limit such increases if they lead to longer task scheduling wait times, so we monitor the ratio of time tasks spend waiting versus running across all processors, and if we see correlations between increases in netdev budget and wait/run ratio increases, netdev budget is tuned down.

Tunables:

bpf(2), bpftune(8),