Re: [PATCH net-next v4 2/4] page_pool: Allow drivers to hint on SKB recycling

[Ilias Apalodimas <ilias.apalodimas@linaro.org>]

Hi Eric,

[...]
> > > > +   if (skb->pp_recycle && page_pool_return_skb_page(head))
> > >
> > > This probably should be attempted only in the (skb->head_frag) case ?
> >
> > I think the extra check makes sense.
> 
> What do you mean here ?
> 

I thought you wanted an extra check in the if statement above.  So move the
block under the existing if. Something like

	if (skb->head_frag) {
		#ifdef (CONFIG_PAGE_POOL)
			if (skb->pp_recycle && page_pool_return_skb_page(head))
			return;
		#endif
        skb_free_frag(head);
	} else {
	.....

> >
> > >
> > > Also this patch misses pskb_expand_head()
> >
> > I am not sure I am following. Misses what? pskb_expand_head() will either
> > call skb_release_data() or skb_free_head(), which would either recycle or
> > unmap the buffer for us (depending on the page refcnt)
> 
>  pskb_expand_head() allocates a new skb->head, from slab.
> 
> We should clear skb->pp_recycle for consistency of the skb->head_frag
> clearing we perform there.

Ah right, good catch. I was mostly worried we are not freeing/unmapping
buffers and I completely missed that.  I think nothing bad will happen even
if we don't, since the signature will eventually protect us, but it's
definitely the right thing to do.

> 
> But then, I now realize you use skb->pp_recycle bit for both skb->head
> and fragments,
> and rely on this PP_SIGNATURE thing (I note that patch 1 changelog
> does not describe why a random page will _not_ have this signature by
> bad luck)

Correct.  I've tried to explain in the previous posting as well, but that's
the big difference compared to the initial RFC we sent a few years ago (the
ability to recycle frags as well).

> 
> Please document/describe which struct page fields are aliased with
> page->signature ?
> 

Sure, any preference on this? Right above page_pool_return_skb_page() ?

Keep in mind the current [1/4] patch is wrong, since it will overlap
pp_signature with mapping.  So we'll have interesting results if a page
gets mapped to userspace :).
What Matthew proposed makes sense, we can add something along the lines of: 

+ unsigned long pp_magic;
+ struct page_pool *pp;
+ unsigned long _pp_mapping_pad;
+ unsigned long dma_addr[2];

in struct page. In this case page->mapping aliases to pa->_pp_mapping_pad

The first word (that we'll now be using) is used for a pointer or a
compound_head.  So as long as pp_magic doesn't resemble a pointer and has 
bits 0/1 set to 0 we should be safe.

Thanks!
/Ilias