Re: [RFC PATCH] net: Fix one page_pool page leak from skb_frag_unref

[Dragos Tatulea]

On Thu, 2024-04-25 at 12:20 -0700, Mina Almasry wrote:
> On Thu, Apr 25, 2024 at 1:17 AM Dragos Tatulea <dtatulea@xxxxxxxxxx> wrote:
> > 
> > On Wed, 2024-04-24 at 15:08 -0700, Mina Almasry wrote:
> > >  If that doesn't work, I think I prefer
> > > reverting a580ea994fd3 ("net: mirror skb frag ref/unref helpers")
> > > rather than merging this fix to make sure we removed the underlying
> > > cause of the issue.
> > This is the safest bet.
> > 
> > So, to recap, I see 2 possibilities:
> > 
> > 1) Revert a580ea994fd3 ("net: mirror skb frag ref/unref helpers"): safe, but it
> > will probably have to come back in one way or another.
> > 2) Drop the recycle checks from skb_frag_ref/unref: this enforces the rule of
> > always referencing/dereferencing pages based on their type (page_pool or
> > normal).
> > 
> 
> If this works, I would be very happy. I personally think ref/unref
> should be done based on the page type. For me the net stack using the
> regular {get|put}_page on a pp page isn't great. It requires special
> handling to make sure the ref + unref are in sync. Also if the last pp
> ref is dropped while there are pending regular refs,
> __page_pool_page_can_be_recycled() check will fail and the page will
> not be recycled.
> 
> On the other hand, since 0a149ab78ee2 ("page_pool: transition to
> reference count management after page draining") I'm not sure there is
> any reason to continue to use get/put_page on pp-pages, we can use the
> new pp-ref instead.
> 
> I don't see any regressions with this diff (needs cleanup), but your
> test setup seems much much better than mine (I think this is the
> second reffing issue you manage to repro):
> 
> diff --git a/include/linux/skbuff_ref.h b/include/linux/skbuff_ref.h
> index 4dcdbe9fbc5f..4c72227dce1b 100644
> --- a/include/linux/skbuff_ref.h
> +++ b/include/linux/skbuff_ref.h
> @@ -31,7 +31,7 @@ static inline bool napi_pp_get_page(struct page *page)
>  static inline void skb_page_ref(struct page *page, bool recycle)
>  {
>  #ifdef CONFIG_PAGE_POOL
> -       if (recycle && napi_pp_get_page(page))
> +       if (napi_pp_get_page(page))
>                 return;
>  #endif
>         get_page(page);
> @@ -69,7 +69,7 @@ static inline void
>  skb_page_unref(struct page *page, bool recycle)
>  {
>  #ifdef CONFIG_PAGE_POOL
> -       if (recycle && napi_pp_put_page(page))
> +       if (napi_pp_put_page(page))
>                 return;
>  #endif
>         put_page(page);
> 
> 
This is option 2. I thought this would fix everything. But I just tested and
it's not the case: we are now reaching a negative pp_ref_count. So probably
somewhere a regular page reference is still being taken...

Thanks,
Dragos