Re: [PATCH bpf] bpf: Zero-fill non-target CPU slots on BPF_F_CPU insert of new per-CPU element

[Ömer Mete Kaya]

On 9/29/26 08:04, Leon Hwang wrote:
> On 29/9/26 05:47, Ömer Mete Kaya wrote:
>> pcpu_copy_value() with BPF_F_CPU copies the value only to the target
>> CPU's slot and returns immediately, leaving all other CPUs untouched.
>> When inserting a *new* element, the backing per-CPU area may be reused
>> from bpf_mem_cache or a prealloc freelist without being zero-initialized,
>> so non-target CPUs retain stale contents of a previously-deleted
>> element in the same map — a same-map information disclosure.
>>
>> pcpu_init_value() is used to initialize newly allocated elements, while
>> updates to existing elements call pcpu_copy_value() directly; non-target
>> CPU slots intentionally retain their previous values there.
>>
>> The onallcpus=false path in pcpu_init_value() already handles new
>> element initialization correctly by explicitly zeroing non-current CPUs:
>>
>>   if (cpu == current_cpu)
>>       copy_map_value(..., value);
>>   else
>>       zero_map_value(...);
>>
>> Fix by adding a zero-fill pass in pcpu_init_value() after
>> pcpu_copy_value() returns, covering only the BPF_F_CPU new-element
>> path: iterate all CPUs and zero-fill every slot except the target.
>> This leaves pcpu_copy_value() unchanged so the update-existing-element
>> semantics are not affected.
>>
>> Fixes: c6936161fd55 ("bpf: Add BPF_F_CPU and BPF_F_ALL_CPUS flags support for percpu_hash and lru_percpu_hash maps")
>> Signed-off-by: Ömer Mete Kaya <omermetekaya0@xxxxxxxxx>
>> ---
>> Tested with a reproducer:
>> - Before the fix: CPU 1 retained 0xDEADBEEFCAFEBABE from the deleted
>>   element.
>> - After the fix: the non-target CPU read back zero.
> 
> Could you add a selftest to prove the issue?
> 
> You said a reproducer. But, I don't know what it looks like.
> 
>>
>>  kernel/bpf/hashtab.c | 10 ++++++++++
>>  1 file changed, 10 insertions(+)
>>
>> diff --git a/kernel/bpf/hashtab.c b/kernel/bpf/hashtab.c
>> index 8cc07e54daaa..dff4321f2828 100644
>> --- a/kernel/bpf/hashtab.c
>> +++ b/kernel/bpf/hashtab.c
>> @@ -1099,6 +1099,16 @@ static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
>>  		}
>>  	} else {
>>  		pcpu_copy_value(htab, pptr, value, onallcpus, map_flags);
>> +		if (map_flags & BPF_F_CPU) {
>> +			int target = map_flags >> 32;
>> +			int cpu;
>> +
>> +			for_each_possible_cpu(cpu) {
> 
> But wait, the 'map_flags & BPF_F_CPU' has been check in the first 'if
> ()' condition in pcpu_init_value(). See the previous fix:
> https://lore.kernel.org/bpf/20260924102321.2120434-1-donggeunyoo.kernel@xxxxxxxxx/
Ah yes didn't see his fix, and it looks much cleaner.I think mine is no
longer necessary.

Thanks,
Ömer