Re: [RFC] net: esp: fix bad handling of pages from page_pool

[Mina Almasry]

On Wed, Mar 6, 2024 at 10:41 AM Mina Almasry <almasrymina@xxxxxxxxxx> wrote:
>
> On Wed, Mar 6, 2024 at 9:41 AM Jakub Kicinski <kuba@xxxxxxxxxx> wrote:
> >
> > On Wed, 6 Mar 2024 17:09:57 +0000 Dragos Tatulea wrote:
> > > > Does the caller need to check skb->pp_recycle? pp_recycle seems like a
> > > > redundant bit. We can tell whether the page is pp by checking
> > > > is_pp_page(page). the pages in the frag must be pp pages when
> > > > skb->pp_recycle is set and must be non pp pages when the
> > > > skb->pp_recycle is not set, so it all seems redundant to me.
> > > >
> > > AFAIU we don't have to check for pp_recycle, at least not in this specific case.
> >
> > Definitely not something we assuming in a fix that needs to go
> > to stable.
> >
> > So far, AFAIU, it's legal to have an skb without skb->pp_recycle
> > set, which holds full page refs to a PP page.
>
> Interesting. I apologize then I did not realize this combination is
> legal. But I have a follow up question: what is the ref code supposed
> to do in this combination? AFAIU:
>
> - skb->pp_recycle && is_pp_page()
> ref via page_pool_ref_page()
> unref via page_pool_unref_page()
>
> - !skb->pp_recycle && !is_pp_page()
> ref via get_page()
> unref via put_page()
>
> - !skb->pp_recycle && is_pp_page()
> ref via ?
> unref via?
>
> Also is this combination also legal you think? and if so what to do?
> - skb->pp_recycle && !is_pp_page()
> ref via ?
> unref via?
>
> I'm asking because I'm starting to wonder if this patch has some issue in it:
> https://patchwork.kernel.org/project/netdevbpf/patch/20231215033011.12107-4-liangchen.linux@xxxxxxxxx/
>
> Because in this patch, if we have a !skb->pp_recycle & is_pp_page()
> combination we end up doing in skb_try_coalesce():
> ref via page_pool_ref_page()
> unref via put_page() via eventual napi_frag_unref()
>
> which seems like an issue, no?
>

Gah, nevermind, skb_pp_frag_ref() actually returns -EINVAL if
!skb->pp_recycle, and in the call site we do a skb_frag_ref() on this
error, so all in all we end up doing a get_page/put_page pair. Sorry
for the noise.

So we're supposed to:
- !skb->pp_recycle && is_pp_page()
ref via get_page
unref via put_page

Very subtle stuff (for me at least). I'll try to propose some cleanup
to make this a bit simpler using helpers that handle all these subtle
details internally so that the call sites don't have to do this
special handling.


-- 
Thanks,
Mina