Re: Re: [PATCH net-next v2 3/3] skbuff: keep track of pp page when __skb_frag_ref() is called

[Yunsheng Lin]

On 2021/9/16 19:04, Yunsheng Lin wrote:
> On 2021/9/16 18:38, Ilias Apalodimas wrote:
>> On Thu, Sep 16, 2021 at 05:33:39PM +0800, Yunsheng Lin wrote:
>>> On 2021/9/16 16:44, Ilias Apalodimas wrote:
>>>>>>> appear if we try to pull in your patches on using page pool and recycling
>>>>
>>>> [...]
>>>>
>>>>>>> for Tx where TSO and skb_split are used?
>>>>>
>>>>> As my understanding, the problem might exists without tx recycling, because a
>>>>> skb from wire would be passed down to the tcp stack and retransmited back to
>>>>> the wire theoretically. As I am not able to setup a configuration to verify
>>>>> and test it and the handling seems tricky, so I am targetting net-next branch
>>>>> instead of net branch.
>>>>>
>>>>>>>
>>>>>>> I'll be honest, when I came up with the recycling idea for page pool, I
>>>>>>> never intended to support Tx.  I agree with Alexander here,  If people want
>>>>>>> to use it on Tx and think there's value,  we might need to go back to the
>>>>>>> drawing board and see what I've missed.  It's still early and there's a
>>>>>>> handful of drivers using it,  so it will less painful now.
>>>>>
>>>>> Yes, we also need to prototype it to see if there is something missing in the
>>>>> drawing board and how much improvement we get from that:)
>>>>>
>>>>>>
>>>>>> I agree, page_pool is NOT designed or intended for TX support.
>>>>>> E.g. it doesn't make sense to allocate a page_pool instance per socket, as the backing memory structures for page_pool are too much.
>>>>>> As the number RX-queues are more limited it was deemed okay that we use page_pool per RX-queue, which sacrifice some memory to gain speed.
>>>>>
>>>>> As memtioned before, Tx recycling is based on page_pool instance per socket.
>>>>> it shares the page_pool instance with rx.
>>>>>
>>>>> Anyway, based on feedback from edumazet and dsahern, I am still trying to
>>>>> see if the page pool is meaningful for tx.
>>>>>
>>>>>>
>>>>>>
>>>>>>> The pp_recycle_bit was introduced to make the checking faster, instead of
>>>>>>> getting stuff into cache and check the page signature.  If that ends up
>>>>>>> being counterproductive, we could just replace the entire logic with the
>>>>>>> frag count and the page signature, couldn't we?  In that case we should be
>>>>>>> very cautious and measure potential regression on the standard path.
>>>>>>
>>>>>> +1
>>>>>
>>>>> I am not sure "pp_recycle_bit was introduced to make the checking faster" is a
>>>>> valid. The size of "struct page" is only about 9 words(36/72 bytes), which is
>>>>> mostly to be in the same cache line, and both standard path and recycle path have
>>>>> been touching the "struct page", so it seems the overhead for checking signature
>>>>> seems minimal.
>>>>>
>>>>> I agree that we need to be cautious and measure potential regression on the
>>>>> standard path.
>>>>
>>>> well pp_recycle is on the same cache line boundary with the head_frag we
>>>> need to decide on recycling. After that we start checking page signatures
>>>> etc,  which means the default release path remains mostly unaffected.  
>>>>
>>>> I guess what you are saying here, is that 'struct page' is going to be
>>>> accessed eventually by the default network path,  so there won't be any 
>>>> noticeable performance hit?  What about the other usecases we have
>>>
>>> Yes.
>>
>> In that case you'd need to call virt_to_head_page() early though, get it
>> and then compare the signature.   I guess that's avoidable by using 
>> frag->bv_page for the fragments?
> 
> If a page of a skb frag is from page pool, It seems frag->bv_page is
> always point to head_page of a compound page, so the calling of
> virt_to_head_page() does not seems necessary.
> 
> bit 0 of frag->bv_page is different way of indicatior for a pp page,
> it is better we do not confuse with the page signature way. Using
> a bit 0 may give us a free word in 'struct page' if we manage to
> use skb->pp_recycle to indicate a head page of the skb uniquely, meaning
> page->pp_magic can be used for future feature.
> 
> 
>>
>>>
>>>> for pp_recycle right now?  __skb_frag_unref() in skb_shift() or
>>>> skb_try_coalesce() (the latter can probably be removed tbh).
>>>
>>> If we decide to go with accurate indicator of a pp page, we just need
>>> to make sure network stack use __skb_frag_unref() and __skb_frag_ref()
>>> to put and get a page frag, the indicator checking need only done in
>>> __skb_frag_unref() and __skb_frag_ref(), so the skb_shift() and
>>> skb_try_coalesce() should be fine too.
>>>
>>>>
>>>>>
>>>>> Another way is to use the bit 0 of frag->bv_page ptr to indicate if a frag
>>>>> page is from page pool.
>>>>
>>>> Instead of the 'struct page' signature?  And the pp_recycle bit will
>>>> continue to exist?  
>>>
>>> pp_recycle bit might only exist or is only used for the head page for the skb.
>>> The bit 0 of frag->bv_page ptr can be used to indicate a frag page uniquely.
>>> Doing a memcpying of shinfo or "*fragto = *fragfrom" automatically pass the
>>> indicator to the new shinfo before doing a __skb_frag_ref(), and __skb_frag_ref()
>>> will increment the _refcount or pp_frag_count according to the bit 0 of
>>> frag->bv_page.
>>>
>>> By the way, I also prototype the above idea, and it seems to work well too.
>>>
>>
>> As long as no one else touches this, it's just another way of identifying a
>> page_pool allocated page.  But are we gaining by that?  Not using
>> virt_to_head_page() as stated above? But in that case you still need to
>> keep pp_recycle around. 
> 
> No, we do not need the pp_recycle, as long as the we make sure __skb_frag_ref()
> is called after memcpying the shinfo or doing "*fragto = *fragfrom".

Acctually it seems we do not need to ensure __skb_frag_ref() is called after
memcpying the shinfo or doing "*fragto = *fragfrom".

Just make sure the bit 0 of frag->bv_page is passed to the new frag->bv_page(
by memcpying the shinfo or doing "*fragto = *fragfrom"), __skb_frag_ref() and
__skb_frag_unref() will check the bit 0 of frag->bv_page to update the _refcount
or pp_frag_count accordingly.

> 
>>
>>>> .
>>>> Right now the 'naive' explanation on the recycling decision is something like:
>>>>
>>>> if (pp_recycle) <--- recycling bit is set
>>>>     (check page signature) <--- signature matches page pool
>>>> 		(check fragment refcnt) <--- If frags are enabled and is the last consumer
>>>> 			recycle
>>>>
>>>> If we can proove the performance is unaffected when we eliminate the first if,
>>>> then obviously we should remove it.  I'll try running that test here and see,
>>>> but keep in mind I am only testing on an 1GB interface.  Any chance we can get 
>>>> measurements on a beefier hardware using hns3 ?
>>>
>>> Sure, I will try it.
>>> As the kind of performance overhead is small, any performance testcase in mind?
>>>
>>
>> 'eliminate the first if' wasn't accurate.  I meant switch the first if and
>> check the struct page signature instead.  That would be the best solution
>> imho.  We effectively have a single rule to check if a packet comes from
>> page_pool or not.
> 
> I am not sure what does "switch " means here, if the page signature can
> indicate a pp page uniquely, the "if (pp_recycle)" checking can be removed.
> 
>>
>> You can start by sending a lot of packets and dropping those immediately.
>> That should put enough stress on the receive path and the allocators and it
>> should give us a rough idea. 
>>
>>>>
>>>>>
>>>>>>
>>>>>>> But in general,  I'd be happier if we only had a simple logic in our
>>>>>>> testing for the pages we have to recycle.  Debugging and understanding this
>>>>>>> otherwise will end up being a mess.
>>>>>>
>>>>>>
>>>>
>>>> [...]
>>>>
>>>> Regards
>>>> /Ilias
>>>> .
>>>>
>>
>> Regards
>> /Ilias
>> .
>>
> .
>