On Tue, Jan 9, 2024 at 10:00 PM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> +
> +SEC("iter/cgroup")
> +int BPF_PROG(cpu_cgroup, struct bpf_iter_meta *meta, struct cgroup *cgrp)
> +{
> + u32 *cpu, nr_running = 0, psi_nr_running = 0, nr_cpus = 0;
> + unsigned int tasks[NR_PSI_TASK_COUNTS];
> + struct psi_group_cpu *groupc;
> + struct bpf_cpumask *mask;
> + struct task_struct *p;
> + struct rq *rq;
> +
> + /* epilogue */
> + if (cgrp == NULL)
> + return 0;
> +
> + mask = bpf_cpumask_create();
> + if (!mask)
> + return 1;
> +
> + p = bpf_task_from_pid(target_pid);
> + if (!p) {
> + bpf_cpumask_release(mask);
> + return 1;
> + }
> +
> + bpf_cpumask_copy(mask, p->cpus_ptr);
> + bpf_for_each(cpumask, cpu, &mask->cpumask) {
> + rq = (struct rq *)bpf_per_cpu_ptr(&runqueues, *cpu);
> + if (!rq)
> + continue;
> + nr_running += rq->nr_running;
> + nr_cpus += 1;
> +
> + groupc = (struct psi_group_cpu *)bpf_per_cpu_ptr(&system_group_pcpu, *cpu);
> + if (!groupc)
> + continue;
> + bpf_probe_read_kernel(&tasks, sizeof(tasks), &groupc->tasks);
> + psi_nr_running += tasks[NR_RUNNING];
> + }
Instead of probe_read_kernel (which is not fast) please use
bpf_rdonly_cast() and access groups->tasks.
array should already be recognized by the verifier, but if not let's
fix the verifier instead of fallback to probe_read.
