Re: [PATCH bpf-next 2/2] selftests/bpf: Add test for struct_ops map release

[Martin KaFai Lau <martin.lau@linux.dev>]

On 11/8/24 12:26 AM, Xu Kuohai wrote:
> -static void bpf_testmod_test_2(int a, int b)
> +static void bpf_dummy_unreg(void *kdata, struct bpf_link *link)
>   {
> +	WRITE_ONCE(__bpf_dummy_ops, &__bpf_testmod_ops);
>   }

[ ... ]

> +static int run_struct_ops(const char *val, const struct kernel_param *kp)
> +{
> +	int ret;
> +	unsigned int repeat;
> +	struct bpf_testmod_ops *ops;
> +
> +	ret = kstrtouint(val, 10, &repeat);
> +	if (ret)
> +		return ret;
> +
> +	if (repeat > 10000)
> +		return -ERANGE;
> +
> +	while (repeat-- > 0) {
> +		ops = READ_ONCE(__bpf_dummy_ops);

I don't think it is the usual bpf_struct_ops implementation which only uses 
READ_ONCE and WRITE_ONCE to protect the registered ops. tcp-cc uses a 
refcnt+rcu. It seems hid uses synchronize_srcu(). sched_ext seems to also use 
kthread_flush_work() to wait for all ops calling finished. Meaning I don't think 
the current bpf_struct_ops unreg implementation will run into this issue for 
sleepable ops.

The current synchronize_rcu_mult(call_rcu, call_rcu_tasks) is only needed for 
the tcp-cc because a tcp-cc's ops (which uses refcnt+rcu) can decrement its own 
refcnt. Looking back, this was a mistake (mine). A new tcp-cc ops should have 
been introduced instead to return a new tcp-cc-ops to be used.

> +		if (ops->test_1)
> +			ops->test_1();
> +		if (ops->test_2)
> +			ops->test_2(0, 0);
> +	}
> +
> +	return 0;
> +}