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.
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.
. 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?
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 .