From: Eric Dumazet <eduma...@google.com>
Date: Fri, 1 Mar 2024 14:25:37 +0100
> On Fri, Mar 1, 2024 at 1:59 PM Alexander Lobakin
> <aleksander.loba...@intel.com> wrote:
>>
>> From: Eric Dumazet <eduma...@google.com>
>> Date: Fri, 1 Mar 2024 09:03:55 +0100
>>
>>> On Fri, Mar 1, 2024 at 7:59 AM Jakub Kicinski <k...@kernel.org> wrote:
>>>>
>>>> On Thu, 29 Feb 2024 13:30:22 -0800 Kees Cook wrote:
>>
>> Re WARN_ONCE() in netdev_priv(): netdev_priv() is VERY hot, I'm not sure
>> we want to add checks there. Maybe under CONFIG_DEBUG_NET?
>>
>>>
>>>>> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
>>>>> index 118c40258d07..b476809d0bae 100644
>>>>> --- a/include/linux/netdevice.h
>>>>> +++ b/include/linux/netdevice.h
>>>>> @@ -1815,6 +1815,8 @@ enum netdev_stat_type {
>>>>> NETDEV_PCPU_STAT_DSTATS, /* struct pcpu_dstats */
>>>>> };
>>>>>
>>>>> +#define NETDEV_ALIGN 32
>>>>
>>>> Unless someone knows what this is for it should go.
>>>> Align priv to cacheline size.
>>>
>>> +2
>>>
>>
>> Maybe
>>
>>> #define NETDEV_ALIGN L1_CACHE_BYTES
>>
>> #define NETDEV_ALIGN max(SMP_CACHE_BYTES, 32)
>
> Why would we care if some arches have a very small SMP_CACHE_BYTES ?
Oh sorry, I thought %SMP_CACHE_BYTES is 1 when !SMP.
We can then just add ____cacheline_aligned to both struct net_device and
its ::priv flex array and that's it.
I like the idea of declaring priv explicitly rather than doing size +
ptr magic. But maybe we could just add this flex array to struct
net_device and avoid introducing a new structure.
> Bet it !
>
> IMO nothing in networking mandates this minimal 32 byte alignment.
>
>>
>> ?
>>
>> (or even max(1 << INTERNODE_CACHE_SHIFT, 32))
>
> I do not think so.
>
> INTERNODE_CACHE_SHIFT is a bit extreme on allyesconfig on x86 :/
> (with CONFIG_X86_VSMP=y)
>
>
>>
>>>
>>> or a general replacement of NETDEV_ALIGN....
>>>
>>>
>>
>> + I'd align both struct net_device AND its private space to
>> %NETDEV_ALIGN and remove this weird PTR_ALIGN. {k,v}malloc ensures
>> natural alignment of allocations for at least a couple years already
>> (IOW if struct net_device is aligned to 64, {k,v}malloc will *always*
>> return a 64-byte aligned address).
>
> I think that with SLAB or SLOB in the past with some DEBUG options
> there was no such guarantee.
>
> But this is probably no longer the case, and heavy DEBUG options these
> days (KASAN, KFENCE...)
> do not expect fast networking anyway.
Thanks,
Olek
