BPF_MAP_TYPE_PERF_EVENT_ARRAY inconsistency:

[wez470]

For picking the buffer length, https://github.com/isovalent/ebpf-docs/blob/master/docs/linux/map-type/BPF_MAP_TYPE_PERF_EVENT_ARRAY.md states

length can be picked and tuned depending on the use case but should always be 1+2^n memory pages and n should not be 0. So Assuming a 4k page size, valid values would be 12288, 20480, 28672 and so on.

But 28672 can not be achieved through that formula. I'm wondering which is correct here. Should the formula be 1+2*n? Or is 28672 invalid?

[dylandreimerink]

This might need some rephrasing. Key is that the formula is given in "memory pages" as unit, not bytes. And so, if you were to be on a system with pages of 4096 bytes. You get:

• n=1 => 1+2^1 pages => 3 pages => 3 * 4096 bytes => 12288 bytes
• n=2 => 1+2^2 pages => 5 pages => 5 * 4096 bytes => 20480 bytes
• n=3 => 1+2^3 pages => 7 pages => 7 * 4096 bytes => 28672 bytes

The formula for bytes would thus be (1+2^n) * {page size}.

I admit this is confusing and poorly written.

[wez470]

1+2^3 is 9, not 7. So it sounds like you're advocating that formula should be 1+2*n and that 28672 bytes is valid (for 4k page size)

[wez470]

Looking at libs like cilium, it seems they use the power of 2 calculation. So I guess 28672 bytes is probably invalid then and the example should be

• n=3 => 1+2^3 pages => 9 pages => 9 * 4096 bytes => 36864 bytes

[dylandreimerink]

Looking at libs like cilium, it seems they use the power of 2 calculation. So I guess 28672 bytes is probably invalid then and the example should be

* `n=3` => `1+2^3 pages` => `9 pages` => `9 * 4096 bytes` => `36864 bytes`

Yes. Apogees for the confusion, turns out math is hard (for me, sometimes) 😅.

[wez470]

NP! Thanks for clarifying