is remote_host->metrics[] update thread-safe?

[andywu106]

Potential Thread Safety Issue in conn_tunner_sockops()​​

The conn_tunner_sockops() function updates the metric_value and metric_cnt in remote_host->metrics[] ​​without lock protection​​. Since conn_tunner_sockops() may execute concurrently across multiple CPUs, could this lead to ​​thread safety issues​​？such as:

​​Data races​​ between concurrent writes to metric_value/metric_cnt
​​Inconsistent metrics​​ due to non-atomic updates

[alan-maguire]

it could; potentlially adding bpf spin_lock()/unlock() here would help. per-cpu maps won't help here as the whole point is we want to aggregate across threads. looking at https://docs.ebpf.io/linux/helper-function/bpf_spin_lock/ looks like bpf_spin_lock was added in the 5.1 kernel and is supported in sock ops progs so doable without regressing the set of kernels that bpftune supports (we start from 5.4).