Hi, all
I have been thinking about eBPF timer APIs, it looks harder than I thought.
The API's themselves are not hard, here is what I have:
int bpf_timer_setup(struct bpf_timer *timer, void *callback_fn,
void *callback_ctx, u64 flags);
int bpf_timer_mod(struct bpf_timer *timer, u64 expires);
int bpf_timer_del(struct bpf_timer *timer);
which is pretty much similar to the current kernel timer API's.
The struct bpf_timer is a bit tricky, we still have to save the kernel timer
there but we do not want eBPF programs to touch it. So I simply save a pointer
to kernel timer inside:
struct bpf_timer {
void *ptr;
};
but we obviously have to prevent eBPF programs from dereferencing it
with the verifier.
The hardest part is on the verifier side, we have to check:
1. Whether a timer is initialized before use. For example:
struct bpf_timer t;
bpf_timer_mod(&t, bpf_jiffies64() + HZ);
2. Whether a timer is still active before exiting. For example:
struct bpf_timer t;
bpf_setup_timer(&t, ....);
bpf_timer_mod(&t, bpf_jiffies64() + HZ);
//end of the eBPF program
I do not see any existing mechanism for checks like these, so potentially need
a lot of work.
And, unlike bpf_for_each_map_elem(), the timer callback is asynchronous, this
makes it harder to check its callback context etc..
Any thoughts and suggestions?
Thanks!
