Re: [PATCH net-next v5 12/19] xdp: add generic xdp_build_skb_from_buff()

[Alexander Lobakin <aleksander.lobakin@intel.com>]

From: Amit Cohen <amcohen@nvidia.com>
Date: Sun, 17 Nov 2024 12:42:11 +0000

> 
> 
>> -----Original Message-----
>> From: Alexander Lobakin <aleksander.lobakin@intel.com>
>> Sent: Friday, 15 November 2024 16:35
>> To: Ido Schimmel <idosch@idosch.org>

[...]

>> Regarding your usecase -- after calling this function, you are free to
>> overwrite any skb fields as this helper doesn't pass it up the stack.
>> For example, in ice driver we have port reps and sometimes we need to
>> pass a different net_device, not the one saved in rxq_info. So when
>> switching to this function, we'll do eth_type_trans() once again (it's
>> either way under unlikely() in our code as it's swichdev slowpath).
>> Same for the queue number in rxq_info.
> 
> With this series, maintaining 'struct xdp_mem_allocator' in hash-table looks unnecessary.
> If so, xdp_reg_mem_model() does not need 'allocator' when mem_type is Page-Pool.
> 
> Is there a reason for not removing 'mem_id_ht'? With this patch, the nodes are no longer used.

They actually are. xdp_unreg_mem_model() performs lookup and calls
page_pool_destroy() basing on what id you have in rxq_info.mem.
__xdp_reg_mem_model() calls page_pool_use_xdp_mem() which increments
pool's refcount, so that the pool can't be destroyed until the
xdp_rxq_info it's connected to is unregistered.
xdp_rxq_info is 64 bytes on x86_64, meaning replacing xdp_mem_info there
with direct PP pointer will blow it up to 128 bytes (64-byte CL) (don't
forget that xdp_rxq_info still needs to have mem.type set).

> 
>>
>>>
>>>>
>>>> To be clear, I understand it is not a common use case.
>>>>
>>>> Thanks
>>
>> Thanks,
>> Olek

Thanks,
Olek