[PATCH RFC 4/4] bpf,cgroup,perf: extend bpf-cgroup to support tracepoint attachment

[Alexei Starovoitov]

On Thu, Nov 18, 2021 at 03:28:40PM -0500, Kenny Ho wrote:
> @@ -245,6 +256,21 @@ static int compute_effective_progs(struct cgroup *cgrp,
>  	if (!progs)
>  		return -ENOMEM;
>  
> +	if (atype == CGROUP_TRACEPOINT) {
> +		/* TODO: only create event for cgroup that can have process */
> +
> +		attr.config = bpf_attach_subtype;
> +		attr.type = PERF_TYPE_TRACEPOINT;
> +		attr.sample_type = PERF_SAMPLE_RAW;
> +		attr.sample_period = 1;
> +		attr.wakeup_events = 1;
> +
> +		rc = perf_event_create_for_all_cpus(&attr, cgrp,
> +				&cgrp->bpf.per_cg_events);
> +		if (rc)
> +			goto err;
> +	}
...
> +int perf_event_create_for_all_cpus(struct perf_event_attr *attr,
> +				struct cgroup *cgroup,
> +				struct list_head *entries)
> +{
> +	struct perf_event **events;
> +        struct perf_cgroup *perf_cgrp;
> +	int cpu, i = 0;
> +
> +	events = kzalloc(sizeof(struct perf_event *) * num_possible_cpus(),
> +			GFP_KERNEL);
> +
> +	if (!events)
> +		return -ENOMEM;
> +
> +	for_each_possible_cpu(cpu) {
> +		/* allocate first, connect the cgroup later */
> +		events[i] = perf_event_create_kernel_counter(attr, cpu, NULL, NULL, NULL);

This is a very heavy hammer for this task.
There is really no need for perf_event to be created.
Did you consider using raw_tp approach instead?
It doesn't need this heavy stuff.
Also I suspect in follow up you'd be adding tracepoints to GPU code?
Did you consider just leaving few __weak global functions in GPU code
and let bpf progs attach to them as fentry?
I suspect the true hierarchical nature of bpf-cgroup framework isn't necessary.
The bpf program itself can filter for given cgroup.
We have bpf_current_task_under_cgroup() and friends.
I suggest to sprinkle __weak empty funcs in GPU and see what
you can do with it with fentry and bpf_current_task_under_cgroup.
There is also bpf_get_current_ancestor_cgroup_id().