Re: [PATCH bpf] bpf: Do not grab the bucket spinlock by default on htab batch ops

[Yonghong Song <yhs@fb.com>]

On 2/18/20 7:56 AM, Daniel Borkmann wrote:
> On 2/18/20 4:43 PM, Yonghong Song wrote:
>> On 2/14/20 2:43 PM, Brian Vazquez wrote:
>>> Grabbing the spinlock for every bucket even if it's empty, was causing
>>> significant perfomance cost when traversing htab maps that have only a
>>> few entries. This patch addresses the issue by checking first the
>>> bucket_cnt, if the bucket has some entries then we go and grab the
>>> spinlock and proceed with the batching.
>>>
>>> Tested with a htab of size 50K and different value of populated entries.
>>>
>>> Before:
>>>    Benchmark             Time(ns)        CPU(ns)
>>>    ---------------------------------------------
>>>    BM_DumpHashMap/1       2759655        2752033
>>>    BM_DumpHashMap/10      2933722        2930825
>>>    BM_DumpHashMap/200     3171680        3170265
>>>    BM_DumpHashMap/500     3639607        3635511
>>>    BM_DumpHashMap/1000    4369008        4364981
>>>    BM_DumpHashMap/5k     11171919       11134028
>>>    BM_DumpHashMap/20k    69150080       69033496
>>>    BM_DumpHashMap/39k   190501036      190226162
>>>
>>> After:
>>>    Benchmark             Time(ns)        CPU(ns)
>>>    ---------------------------------------------
>>>    BM_DumpHashMap/1        202707         200109
>>>    BM_DumpHashMap/10       213441         210569
>>>    BM_DumpHashMap/200      478641         472350
>>>    BM_DumpHashMap/500      980061         967102
>>>    BM_DumpHashMap/1000    1863835        1839575
>>>    BM_DumpHashMap/5k      8961836        8902540
>>>    BM_DumpHashMap/20k    69761497       69322756
>>>    BM_DumpHashMap/39k   187437830      186551111
>>>
>>> Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map")
>>> Cc: Yonghong Song <yhs@fb.com>
>>> Signed-off-by: Brian Vazquez <brianvv@google.com>
>>
>> Acked-by: Yonghong Song <yhs@fb.com>
> 
> I must probably be missing something, but how is this safe? Presume we
> traverse in the walk with bucket_cnt = 0. Meanwhile a different CPU added
> entries to this bucket since not locked. Same reader on the other CPU with
> bucket_cnt = 0 then starts to traverse the second
> hlist_nulls_for_each_entry_safe() unlocked e.g. deleting entries?

Thanks for pointing this out. Yes, you are correct. If bucket_cnt is 0
and buck->lock is not held, we should skip the
    hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
       ...
    }
as another cpu may traverse the bucket in parallel by adding/deleting 
the elements.