Re: bpf timer design

[Cong Wang <xiyou.wangcong@xxxxxxxxx>]

On Sat, Mar 13, 2021 at 11:19 AM Cong Wang <xiyou.wangcong@xxxxxxxxx> wrote:
>
> Please let me know what you think about introducing a timer
> map, something like below:
>
> struct {
>      __uint(type, BPF_MAP_TYPE_TIMER);
> } map SEC(".maps");
>
> struct bpf_timer t;

After some thoughts, I think the following solution is much better.

We still need a timer map:

struct {
     __uint(type, BPF_MAP_TYPE_TIMER);
} map SEC(".maps");

However, its key is not a pointer to timer, it is a timer ID allocated with

u32 bpf_timer_create(void *callback, void *arg, u64 flags);

which returns a globally unique ID. So, we end up having code like
this:

u32 timer_id;

static int timer_cb(void *arg)
{
  // show how to rearm a timer
  u64 new_expires = ...;
  bpf_map_update_elem(&map, &timer_id, &new_expires, 0);
}

int bpf_timer_test(...)
{
  u64 expires = ...;

  timer_id = bpf_timer_create(timer_cb, arg, 0);
  bpf_map_update_elem(&map, &timer_id, &expires, 0);

  // wait for timer deletion synchronously
  bpf_map_delete_elem(&map, &timer_id);
}

In kernel, we can use an IDR to allocate these ID's and save a kernel
timer pointer for each ID there.

With this solution, we don't need to change much in the verifier,
probably only verifying the callback arg pointer for bpf_timer_create().

Any thoughts on this proposal?

Thanks!