EBPF Talk: Ringbuf Experience Sharing

I have been using Ringbuf recentlybpf_ringbuf_reserve()I stepped on a hole and recorded it.

Introduction to Ringbuf

answer: have 。

Directly flip throughbpf_ringbuf_reserve()Source code:

// https://github.com/torvalds/linux/blob/5be63fc19fcaa4c236b307420483578a56986a37/kernel/bpf/ringbuf.c#L408static void *__bpf_ringbuf_reserve(struct bpf_ringbuf *rb, u64 size)
{
    // ...    cons_pos = smp_load_acquire(&rb->consumer_pos);    if (in_nmi()) {
        if (!spin_trylock_irqsave(&rb->spinlock, flags))
            return NULL;
    } else {
        spin_lock_irqsave(&rb->spinlock, flags);
    }    pend_pos = rb->pending_pos;
    prod_pos = rb->producer_pos;
    new_prod_pos = prod_pos + len;    // ...    /* pairs with consumer's smp_load_acquire() */
    smp_store_release(&rb->producer_pos, new_prod_pos);    spin_unlock_irqrestore(&rb->spinlock, flags);    return (void *)hdr + BPF_RINGBUF_HDR_SZ;
}BPF_CALL_3(bpf_ringbuf_reserve, struct bpf_map *, map, u64, size, u64, flags)
{
    struct bpf_ringbuf_map *rb_map;    if (unlikely(flags))
        return 0;    rb_map = container_of(map, struct bpf_ringbuf_map, map);
    return (unsigned long)__bpf_ringbuf_reserve(rb_map->rb, size);
}

When writing BPF code, I did not pay attention to the impact of this lock on performance, resulting in poor performance. Example:

static __always_inline void
record_event(struct xdp_md *xdp)
{
    struct event_t *event;    event =…