Received-SPF: pass (google.com: domain of linux-kernel+bounces-165006-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45e3:2400::1 as permitted sender) client-ip=2604:1380:45e3:2400::1;
Precedence: bulk
MIME-Version: 1.0
References: <20240424165646.1625690-2-dtatulea@nvidia.com> <4ba023709249e11d97c78a98ac7db3b37f419960.camel@nvidia.com>
 <CAHS8izMbAJHatnM6SvsZVLPY+N7LgGJg03pSdNfSRFCufGh9Zg@mail.gmail.com>
 <4c20b500c2ed615aba424c0f3c7a79f5f5a04171.camel@nvidia.com>
 <20240426160557.51de91f9@kernel.org> <c307a3086d255d1dfed22284f500aa9fb70f11a3.camel@nvidia.com>
 <7a5a1d74040052afc8cc6cc5c2700fdf2e836b0c.camel@nvidia.com>
In-Reply-To: <7a5a1d74040052afc8cc6cc5c2700fdf2e836b0c.camel@nvidia.com>
From: Mina Almasry <almasrymina@google.com>
Date: Wed, 1 May 2024 00:48:43 -0700
Message-ID: <CAHS8izOsFqiSiS4Z-E-jfD70aogNp5Bcyt7Rk8xFoR2TcDjz=g@mail.gmail.com>
Subject: Re: [RFC PATCH] net: Fix one page_pool page leak from skb_frag_unref
To: Dragos Tatulea <dtatulea@nvidia.com>
Cc: "kuba@kernel.org" <kuba@kernel.org>, "davem@davemloft.net" <davem@davemloft.net>, 
	"ilias.apalodimas@linaro.org" <ilias.apalodimas@linaro.org>, 
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>, 
	"jacob.e.keller@intel.com" <jacob.e.keller@intel.com>, "pabeni@redhat.com" <pabeni@redhat.com>, 
	Jianbo Liu <jianbol@nvidia.com>, "netdev@vger.kernel.org" <netdev@vger.kernel.org>, 
	"edumazet@google.com" <edumazet@google.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Tue, Apr 30, 2024 at 11:20=E2=80=AFPM Dragos Tatulea <dtatulea@nvidia.co=
m> wrote:
>
> On Mon, 2024-04-29 at 09:39 +0200, Dragos Tatulea wrote:
> > On Fri, 2024-04-26 at 16:05 -0700, Jakub Kicinski wrote:
> > > On Thu, 25 Apr 2024 08:17:28 +0000 Dragos Tatulea wrote:
> > > > >  The unref path always dropped a regular page
> > > > > ref, thanks to this commit as you point out:
> > > > >
> > > > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.gi=
t/commit/?id=3D2cc3aeb5ecccec0d266813172fcd82b4b5fa5803
> > > > >
> > > > > AFAICT the correct fix is to actually revert commit 2cc3aeb5eccc
> > > > > ("skbuff: Fix a potential race while recycling page_pool packets"=
).
> > > > > The reason is that now that skb_frag_ref() can grab page-pool ref=
s, we
> > > > > don't need to make sure there is only 1 SKB that triggers the rec=
ycle
> > > > > path anymore. All the skb and its clones can obtain page-pool ref=
s,
> > > > > and in the unref path we drop the page-pool refs. page_pool_put_p=
age()
> > > > > detects correctly that the last page-pool ref is put and recycles=
 the
> > > > > page only then.
> > > > >
> > > > I don't think this is a good way forward. For example, skb->pp_recy=
cle is used
> > > > as a hint in skb_gro_receive to avoid coalescing skbs with differen=
t pp_recycle
> > > > flag states. This could interfere with that.
> > >
> > > That's a bit speculative, right? The simple invariant we are trying t=
o
> > > hold is that if skb->pp_recycle && skb_frag_is_pp(skb, i) then the
> > > reference skb is holding on that frag is a pp reference, not page
> > > reference.
> > >
> > Yes, it was a speculative statement. After re-reading it and the code o=
f
> > skb_gro_receive() it makes less sense now.
> >
> > Mina's suggestion to revert commit 2cc3aeb5eccc ("skbuff: Fix a potenti=
al race
> > while recycling page_pool packets") seems less scary now. I just hope w=
e don't
> > bump into too many scenarios similar to the ipsec one...
> >
> > > skb_gro_receive() needs to maintain that invariant, if it doesn't
> > > we need to fix it..
> > >
> >
> Gentle ping. Not sure how to proceed with this:
>
> 1) Revert commit 2cc3aeb5eccc
> ("skbuff: Fix a potential race while recycling page_pool packets"). I tes=
ted
> this btw and it works (for this specific scenario).
>
> 2) Revert Mina's commit a580ea994fd3 ("net: mirror skb frag ref/unref hel=
pers")
> for now.
>

I vote for #1, and IIUC Jakub's feedback, he seems to prefer this as
well. If we continue to run into edge cases after the revert of #1, I
think we may want to do #2 and I can look to reland it with the kunit
tests that Jakub suggested that reproduce these edge cases.

I can upload #1 in the morning if there are no objections. I don't see
any regressions with #1 but I was never able to repo this issue.

--=20
Thanks,
Mina