Re: [PATCH bpf-next] RFC: bpf hashmap - improve iteration in the presence of concurrent delete operations

[Martin KaFai Lau]

On Tue, Feb 01, 2022 at 10:35:14AM -0800, Maciej Żenczykowski wrote:
> From: Maciej Żenczykowski <maze@xxxxxxxxxx>
> 
> by resuming from the bucket the key would be in if it still existed
> 
> Note: AFAICT this an API change that would require some sort of guard...
> 
> bpf map iteration was added all the way back in v3.18-rc4-939-g0f8e4bd8a1fc
> but behaviour of find first key with NULL was only done in v4.11-rc7-2042-g8fe45924387b
> (before that you'd get EFAULT)
> 
> this means previously find first key was done by calling with a non existing key
> (AFAICT this was hoping to get lucky, since if your non existing key actually existed,
> it wouldn't actually iterate right, since you couldn't guarantee your magic value was
> the first one)
> 
> ie. do we need a BPF_MAP_GET_NEXT_KEY2 or a sysctl? some other approach?
BPF_MAP_LOOKUP_BATCH has already been added to lookup in batches of buckets,
so the next bpf_map_lookup_batch will start from the beginning of
the next bucket.  Here is the details:

https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@xxxxxxxxxx/#t