Received-SPF: pass (google.com: domain of linux-wireless-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20201125173436.1894624-1-elver@google.com> <20201125124313.593fc2b5@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
 <CANpmjNP_=Awx0-eZisMXzgXxKqf7hcrZYCYzFXuebPcwZtkoLw@mail.gmail.com>
 <CAF=yD-JtRUjmy+12kTL=YY8Cfi_c92GVbHZ647smWmasLYiNMg@mail.gmail.com> <CANpmjNO8H9OJDTcKhg4PRVEV04Gxnb56mJY2cB9j4cH+4nznhQ@mail.gmail.com>
In-Reply-To: <CANpmjNO8H9OJDTcKhg4PRVEV04Gxnb56mJY2cB9j4cH+4nznhQ@mail.gmail.com>
From:   Willem de Bruijn <willemdebruijn.kernel@gmail.com>
Date:   Fri, 27 Nov 2020 11:50:34 -0500
Message-ID: <CA+FuTSfCSZFC2Bz5WpnaoU__jrd8sSwsDqN1TNar3yeGNbVeQQ@mail.gmail.com>
Subject: Re: [PATCH net-next] net: switch to storing KCOV handle directly in sk_buff
To:     Marco Elver <elver@google.com>
Cc:     Willem de Bruijn <willemdebruijn.kernel@gmail.com>,
        Jakub Kicinski <kuba@kernel.org>,
        "David S. Miller" <davem@davemloft.net>,
        Johannes Berg <johannes@sipsolutions.net>,
        Aleksandr Nogikh <a.nogikh@gmail.com>,
        Andrey Konovalov <andreyknvl@google.com>,
        Dmitry Vyukov <dvyukov@google.com>,
        LKML <linux-kernel@vger.kernel.org>,
        Netdev <netdev@vger.kernel.org>,
        linux-wireless <linux-wireless@vger.kernel.org>,
        Ido Schimmel <idosch@idosch.org>,
        Florian Westphal <fw@strlen.de>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Fri, Nov 27, 2020 at 7:26 AM Marco Elver <elver@google.com> wrote:
>
> On Thu, 26 Nov 2020 at 17:35, Willem de Bruijn
> <willemdebruijn.kernel@gmail.com> wrote:
> > On Thu, Nov 26, 2020 at 3:19 AM Marco Elver <elver@google.com> wrote:
> [...]
> > > Will send v2.
> >
> > Does it make more sense to revert the patch that added the extensions
> > and the follow-on fixes and add a separate new patch instead?
>
> That doesn't work, because then we'll end up with a build-broken
> commit in between the reverts and the new version, because mac80211
> uses skb_get_kcov_handle().
>
> > If adding a new field to the skb, even if only in debug builds,
> > please check with pahole how it affects struct layout if you
> > haven't yet.
>
> Without KCOV:
>
>         /* size: 224, cachelines: 4, members: 72 */
>         /* sum members: 217, holes: 1, sum holes: 2 */
>         /* sum bitfield members: 36 bits, bit holes: 2, sum bit holes: 4 bits */
>         /* forced alignments: 2 */
>         /* last cacheline: 32 bytes */
>
> With KCOV:
>
>         /* size: 232, cachelines: 4, members: 73 */
>         /* sum members: 225, holes: 1, sum holes: 2 */
>         /* sum bitfield members: 36 bits, bit holes: 2, sum bit holes: 4 bits */
>         /* forced alignments: 2 */
>         /* last cacheline: 40 bytes */

Thanks. defconfig leaves some symbols disabled, but manually enabling
them just fills a hole, so 232 is indeed the worst case allocation.

I recall a firm edict against growing skb, but I don't know of a
hard limit at exactly 224.

There is a limit at 2048 - sizeof(struct skb_shared_data) == 1728B
when using pages for two ETH_FRAME_LEN (1514) allocations.

This would leave 1728 - 1514 == 214B if also squeezing the skb itself
in with the same allocation.

But I have no idea if this is used anywhere. Certainly have no example
ready. And as you show, the previous default already is at 224.

If no one else knows of a hard limit at 224 or below, I suppose the
next technical limit is just 256 for kmem cache purposes.

My understanding was that skb_extensions was supposed to solve this
problem of extending the skb without growing the main structure. Not
for this patch, but I wonder if we can resolve the issues exposed here
and make usable in more conditions.