Re: [PATCH bpf-next v2 5/7] mm: introduce BPF kfunc to access memory events

[Roman Gushchin]

bot+bpf-ci@xxxxxxxxxx writes:

>> diff --git a/mm/bpf_memcontrol.c b/mm/bpf_memcontrol.c
>> index d84fe6f3ed43..858eb43766ce 100644
>> --- a/mm/bpf_memcontrol.c
>> +++ b/mm/bpf_memcontrol.c
>> @@ -103,6 +103,22 @@ __bpf_kfunc unsigned long bpf_mem_cgroup_usage(struct mem_cgroup *memcg)
>>  	return mem_cgroup_usage(memcg, false) * PAGE_SIZE;
>>  }
>>
>> +/**
>> + * bpf_mem_cgroup_memory_events - Read memory cgroup's memory event value
>> + * @memcg: memory cgroup
>> + * @event: memory event id
>> + *
>> + * Returns current memory event count.
>> + */
>> +__bpf_kfunc unsigned long bpf_mem_cgroup_memory_events(struct mem_cgroup *memcg,
>> +						enum memcg_memory_event event)
>> +{
>> +	if (event >= MEMCG_NR_MEMORY_EVENTS)
>> +		return (unsigned long)-1;
>
> Should this also check for negative values?  The enum type is typically
> signed, so a BPF program passing a negative event value would pass this
> check but cause an out-of-bounds read from memcg->memory_events[].
>
> The similar bpf_mem_cgroup_page_state() in this same file uses:
>
>     if (idx < 0 || idx >= MEMCG_NR_STAT)
>
> which checks both bounds.  Could the check here be changed to:
>
>     if (event < 0 || event >= MEMCG_NR_MEMORY_EVENTS)
>
> to match that pattern?

No, the bpf verifier enforces event > 0.
It's a false positive.