Re: [PATCH bpf-next v1 1/2] xdp: recycle Page Pool backed skbs built from XDP frames

[Yunsheng Lin <linyunsheng@huawei.com>]

On 2023/3/8 2:14, Alexander Lobakin wrote:
> From: Yunsheng Lin <linyunsheng@huawei.com>
> Date: Tue, 7 Mar 2023 10:50:34 +0800
> 
>> On 2023/3/6 19:58, Alexander Lobakin wrote:
>>> From: Yunsheng Lin <linyunsheng@huawei.com>
>>> Date: Mon, 6 Mar 2023 09:09:31 +0800
> 
> [...]
> 
>>> Ah, from that perspective. Yes, you're probably right, but would need to
>>> be tested anyway. I don't see any open problems with the PP recycling
>>> right now on the lists, but someone may try to change it one day.
>>> Anyway, this flag is only to do a quick test. We do have
>>> sk_buff::pfmemalloc, but this flag doesn't mean every page from this skb
>>> was pfmemalloced.
>>
>> The point seems to be that sk_buff::pfmemalloc allow false positive, which
>> means skb->pfmemalloc can be set to true while every page from this skb is
>> not pfmemalloced as you mentioned.
>>
>> While skb->pp_recycle can't allow false positive, if that happens, reference
>> counting of the page will not be handled properly if pp and non-pp skb shares
>> the page as the wireless adapter does.
> 
> You mean false-positives in both directions? Because if ->pp_recycle is
> set, the stack can still free non-PP pages. In the opposite case, I mean
> when ->pp_recycle is false and an skb page belongs to a page_pool, yes,
> there'll be issues.

That may depends on what is a PP pages and what is a non-PP pages, it seems
hard to answer now.

For a skb with ->pp_recycle being true and its frag page with page->pp_magic
being PP_SIGNATURE, when calling skb_clone()/pskb_expand_head() or
skb_try_coalesce(), we may call __skb_frag_ref() for the frag page, which
mean a page with page->pp_magic being PP_SIGNATURE can be both PP page
and non-PP page at the same time. So it is important to set the ->pp_recycle
correctly, and it seems hard to get that right from past experience，that's
why a per page marker is suggested.


> But I think the deal is to propagate the flag when you want to attach a
> PP-backed page to the skb? I mean, if someone decides to mix pages with
> different memory models, it's his responsibility to make sure everything
> is fine, because it's not a common/intended way. Isn't it?
> 
>>
>>>
>>>>
>>>>>
>>>>>>
>>>>>> Anyway, I am not sure checking ::pp_magic is correct when a
>>>>>> page will be passing between different subsystem and back to
>>>>>> the network stack eventually, checking ::pp_magic may not be
>>>>>> correct if this happens.
>>>>>>
>>>>>> Another way is to use the bottom two bits in bv_page, see:
>>>>>> https://www.spinics.net/lists/netdev/msg874099.html
> 
> This one is interesting actually. We'd only need one bit -- which is
> 100% free and available in case of page pointers.
> 
>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>>>  
>>>>>>>>>  	/* Allow SKB to reuse area used by xdp_frame */
>>>>>>>>>  	xdp_scrub_frame(xdpf);
> 
> [...]
> 
> Thanks,
> Olek
> 
> .
>