Re: [PATCH net-next RFC 1/2] page_pool: add page recycling support based on elevated refcnt

[Ilias Apalodimas]

On Thu, Jul 08, 2021 at 08:29:56AM -0700, Alexander Duyck wrote:
> On Thu, Jul 8, 2021 at 8:17 AM Ilias Apalodimas
> <ilias.apalodimas@xxxxxxxxxx> wrote:
> >
> > > > >
> > > > > > > > >
> > > > >
> > > > > [...]
> > > > >
> > > > > > > > > The above expectation is based on that the last user will always
> > > > > > > > > call page_pool_put_full_page() in order to do the recycling or do
> > > > > > > > > the resource cleanup(dma unmaping..etc).
> > > > > > > > >
> > > > > > > > > As the skb_free_head() and skb_release_data() have both checked the
> > > > > > > > > skb->pp_recycle to call the page_pool_put_full_page() if needed, I
> > > > > > > > > think we are safe for most case, the one case I am not so sure above
> > > > > > > > > is the rx zero copy, which seems to also bump up the refcnt before
> > > > > > > > > mapping the page to user space, we might need to ensure rx zero copy
> > > > > > > > > is not the last user of the page or if it is the last user, make sure
> > > > > > > > > it calls page_pool_put_full_page() too.
> > > > > > > >
> > > > > > > > Yes, but the skb->pp_recycle value is per skb, not per page. So my
> > > > > > > > concern is that carrying around that value can be problematic as there
> > > > > > > > are a number of possible cases where the pages might be
> > > > > > > > unintentionally recycled. All it would take is for a packet to get
> > > > > > > > cloned a few times and then somebody starts using pskb_expand_head and
> > > > > > > > you would have multiple cases, possibly simultaneously, of entities
> > > > > > > > trying to free the page. I just worry it opens us up to a number of
> > > > > > > > possible races.
> > > > > > >
> > > > > > > Maybe I missde something, but I thought the cloned SKBs would never trigger
> > > > > > > the recycling path, since they are protected by the atomic dataref check in
> > > > > > > skb_release_data(). What am I missing?
> > > > > >
> > > > > > Are you talking about the head frag? So normally a clone wouldn't
> > > > > > cause an issue because the head isn't changed. In the case of the
> > > > > > head_frag we should be safe since pskb_expand_head will just kmalloc
> > > > > > the new head and clears head_frag so it won't trigger
> > > > > > page_pool_return_skb_page on the head_frag since the dataref just goes
> > > > > > from 2 to 1.
> > > > > >
> > > > > > The problem is that pskb_expand_head memcopies the page frags over and
> > > > > > takes a reference on the pages. At that point you would have two skbs
> > > > > > both pointing to the same set of pages and each one ready to call
> > > > > > page_pool_return_skb_page on the pages at any time and possibly racing
> > > > > > with the other.
> > > > >
> > > > > Ok let me make sure I get the idea properly.
> > > > > When pskb_expand_head is called, the new dataref will be 1, but the
> > > > > head_frag will be set to 0, in which case the recycling code won't be
> > > > > called for that skb.
> > > > > So you are mostly worried about a race within the context of
> > > > > pskb_expand_skb() between copying the frags, releasing the previous head
> > > > > and preparing the new one (on a cloned skb)?
> > > >
> > > > The race is between freeing the two skbs. So the original and the
> > > > clone w/ the expanded head will have separate instances of the page. I
> > > > am pretty certain there is a race if the two of them start trying to
> > > > free the page frags at the same time.
> > > >
> > >
> > > Right, I completely forgot calling __skb_frag_unref() before releasing the
> > > head ...
> > > You are right, this will be a race.  Let me go back to the original mail
> > > thread and see what we can do
> > >
> >
> > What do you think about resetting pp_recycle bit on pskb_expand_head()?
> 
> I assume you mean specifically in the cloned case?
> 

Yes. Even if we do it unconditionally we'll just loose non-cloned buffers from
the recycling.
I'll send a patch later today.

> > If my memory serves me right Eric wanted that from the beginning. Then the
> > cloned/expanded SKB won't trigger the recycling.  If that skb hits the free
> > path first, we'll end up recycling the fragments eventually.  If the
> > original one goes first, we'll just unmap the page(s) and freeing the cloned
> > one will free all the remaining buffers.
> 
> I *think* that should be fine. Effectively what we are doing is making
> it so that if the original skb is freed first the pages are released,
> and if it is released after the clone/expended skb then it can be
> recycled.

Exactly

> 
> The issue is we have to maintain it so that there will be exactly one
> caller of the recycling function for the pages. So any spot where we
> are updating skb->head we will have to see if there is a clone and if
> so we have to clear the pp_recycle flag on our skb so that it doesn't
> try to recycle the page frags as well.

Correct. I'll keep looking around in case there's something less fragile we
can do 

Thanks
/Ilias