Re: [net-next V5 PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Daniel Borkmann <daniel@iogearbox.net>]

On 10/09/2017 07:59 PM, Jesper Dangaard Brouer wrote:
[...]
>>> +static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
>>> +{
>>> +	struct bpf_cpu_map_entry *rcpu =
>>> +		__cpu_map_lookup_elem(map, *(u32 *)key);
>>> +
>>> +	return rcpu ? &rcpu->qsize : NULL;
>>
>> I still think from my prior email/comment that we should use per-cpu
>> scratch buffer here. Would be nice to keep the guarantee that noone
>> can modify it, it's just a tiny change.
>
> Well, it's no-longer really needed, right(?), as this patchset update,
> change that bpf-side cannot invoke this.  The userspace-side reading
> this will get a copy.

Ah sorry, you're right, the related change happens in later patch, I
missed that; would be good to avoid a split in future or other option
is to forbid usage initially in check_map_func_compatibility() by
bailing out in BPF_MAP_TYPE_CPUMAP case unconditionally, and then
enabling it for BPF_FUNC_redirect_map in next step, such that should
someone accidentally only backport this patch, we don't allow for
unintended misuse.

Thanks,
Daniel