Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
MIME-Version: 1.0
References: <20190418155652.22181-1-alban@kinvolk.io> <CAH3MdRUMavDst+KssmV=EtwQTudq12ZvXhiP-qiBb+24R8vpGA@mail.gmail.com>
In-Reply-To: <CAH3MdRUMavDst+KssmV=EtwQTudq12ZvXhiP-qiBb+24R8vpGA@mail.gmail.com>
From:   Alban Crequy <alban@kinvolk.io>
Date:   Fri, 19 Apr 2019 13:51:05 +0200
Message-ID: <CADZs7q6XahoWDKWu90xEais_kJ1rKekcXWd2Lh+WyvKiJCtG1w@mail.gmail.com>
Subject: Re: [PATCH bpf-next v2 1/3] bpf: sock ops: add netns ino and dev in
 bpf context
To:     Y Song <ys114321@gmail.com>
Cc:     Alban Crequy <alban.crequy@gmail.com>,
        John Fastabend <john.fastabend@gmail.com>,
        Alexei Starovoitov <ast@kernel.org>,
        Daniel Borkmann <daniel@iogearbox.net>,
        bpf <bpf@vger.kernel.org>, netdev <netdev@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Alban Crequy <alban@kinvolk.io>,
        =?UTF-8?Q?Iago_L=C3=B3pez_Galeiras?= <iago@kinvolk.io>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Apr 18, 2019 at 11:31 PM Y Song <ys114321@gmail.com> wrote:
>
> On Thu, Apr 18, 2019 at 8:58 AM Alban Crequy <alban.crequy@gmail.com> wrote:
> >
> > From: Alban Crequy <alban@kinvolk.io>
> >
> > sockops programs can now access the network namespace inode and device
> > via (struct bpf_sock_ops)->netns_ino and ->netns_dev. This can be useful
> > to apply different policies on different network namespaces.
> >
> > In the unlikely case where network namespaces are not compiled in
> > (CONFIG_NET_NS=n), the verifier will not allow access to ->netns_*.
> >
> > The generated BPF bytecode for netns_ino is loading the correct inode
> > number at the time of execution.
> >
> > However, the generated BPF bytecode for netns_dev is loading an
> > immediate value determined at BPF-load-time by looking at the initial
> > network namespace. In practice, this works because all netns currently
> > use the same virtual device. If this was to change, this code would need
> > to be updated too.
> >
> > Signed-off-by: Alban Crequy <alban@kinvolk.io>
> >
> > ---
> >
> > Changes since v1:
> > - add netns_dev (review from Alexei)
> > ---
> >  include/uapi/linux/bpf.h |  2 ++
> >  net/core/filter.c        | 70 ++++++++++++++++++++++++++++++++++++++++
> >  2 files changed, 72 insertions(+)
> >
> > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> > index eaf2d3284248..f4f841dde42c 100644
> > --- a/include/uapi/linux/bpf.h
> > +++ b/include/uapi/linux/bpf.h
> > @@ -3213,6 +3213,8 @@ struct bpf_sock_ops {
> >         __u32 sk_txhash;
> >         __u64 bytes_received;
> >         __u64 bytes_acked;
> > +       __u64 netns_dev;
> > +       __u64 netns_ino;
>
> Maybe we can define netns_dev as __u32?
>    __u64 netns_ino;
>    __u32 netns_dev;
>
> There is a hole at the end which can be used if the next
> field to be added in the future is a __u32.
>
> From
> static inline u32 new_encode_dev(dev_t dev)
> {
>         unsigned major = MAJOR(dev);
>         unsigned minor = MINOR(dev);
>         return (minor & 0xff) | (major << 8) | ((minor & ~0xff) << 12);
> }
>
> device num is encoded in a u32.

I could do that but there are already two occurrences of "__u64
netns_dev" in bpf.h:
- struct bpf_prog_info
- struct bpf_map_info

Should I keep it a u64 for consistency with the rest of bpf.h, or
change it to u32?

> >  };
> >
> >  /* Definitions for bpf_sock_ops_cb_flags */
> > diff --git a/net/core/filter.c b/net/core/filter.c
> > index 1833926a63fc..93e3429603d7 100644
> > --- a/net/core/filter.c
> > +++ b/net/core/filter.c
> > @@ -75,6 +75,8 @@
> >  #include <net/seg6_local.h>
> >  #include <net/lwtunnel.h>
> >  #include <net/ipv6_stubs.h>
> > +#include <linux/kdev_t.h>
> > +#include <linux/proc_ns.h>
> >
> >  /**
> >   *     sk_filter_trim_cap - run a packet through a socket filter
> > @@ -6774,6 +6776,15 @@ static bool sock_ops_is_valid_access(int off, int size,
> >                 }
> >         } else {
> >                 switch (off) {
> > +               case offsetof(struct bpf_sock_ops, netns_dev):
> > +               case offsetof(struct bpf_sock_ops, netns_ino):
> > +#ifdef CONFIG_NET_NS
> > +                       if (size != sizeof(__u64))
> > +                               return false;
>
> for netns_dev, looks like we have encoding like
>    (minor & 0xff) | (major << 8) | ((minor & ~0xff) << 12)
>
> it is possible user may access like:
>    u16 dev = sk->netns_dev;
> the compiler may translate into a 2-byte load
> or even
>   u8 major = (sk->netns_dev >> 8) & 0xf
> the compiler may translate to a byte load of &sk->netns_dev + 1 or 3
> depends on endianness.
> I suggest we relax the access pattern now to avoid patch later.

Ok, I can do that.

My use case was to compare (struct bpf_sock_ops)->netns_dev/netns_ino
with the value from a map prepared from userspace, i.e. the value of
netns_dev/ino is an opaque buffer for the BPF program, so the problem
is unlikely. But I've got bitten by a similar thing before (struct
sk_msg_md->remote_port/local_port was read in one go as a u64) so we
can improve this a bit.

Thanks for the review, I'll do the rest of the requested changes
probably end of next week.

> > +#else
> > +                       return false;
> > +#endif
> > +                       break;
> >                 case bpf_ctx_range_till(struct bpf_sock_ops, bytes_received,
> >                                         bytes_acked):
> >                         if (size != sizeof(__u64))
> > @@ -7660,6 +7671,11 @@ static u32 sock_addr_convert_ctx_access(enum bpf_access_type type,
> >         return insn - insn_buf;
> >  }
> >
> > +static struct ns_common *sockops_netns_cb(void *private_data)
> > +{
> > +       return &init_net.ns;
> > +}
> > +
> >  static u32 sock_ops_convert_ctx_access(enum bpf_access_type type,
> >                                        const struct bpf_insn *si,
> >                                        struct bpf_insn *insn_buf,
> > @@ -7668,6 +7684,10 @@ static u32 sock_ops_convert_ctx_access(enum bpf_access_type type,
> >  {
> >         struct bpf_insn *insn = insn_buf;
> >         int off;
> > +       struct inode *ns_inode;
> > +       struct path ns_path;
> > +       __u64 netns_dev;
>
> Although there are a few discrepancies,  I suggest to use u64 for
> kernel as __u64 is typically
> used in user space.
>
> > +       void *res;
> >
> >  /* Helper macro for adding read access to tcp_sock or sock fields. */
> >  #define SOCK_OPS_GET_FIELD(BPF_FIELD, OBJ_FIELD, OBJ)                        \
> > @@ -7914,6 +7934,56 @@ static u32 sock_ops_convert_ctx_access(enum bpf_access_type type,
> >                 SOCK_OPS_GET_OR_SET_FIELD(sk_txhash, sk_txhash,
> >                                           struct sock, type);
> >                 break;
> > +
> > +       case offsetof(struct bpf_sock_ops, netns_dev):
> > +#ifdef CONFIG_NET_NS
> > +               /* We get the netns_dev at BPF-load-time and not at
> > +                * BPF-exec-time. We assume that netns_dev is a constant.
> > +                */
> > +               res = ns_get_path_cb(&ns_path, sockops_netns_cb, NULL);
> > +               if (IS_ERR(res)) {
> > +                       netns_dev = 0;
> > +               } else {
> > +                       ns_inode = ns_path.dentry->d_inode;
> > +                       netns_dev = new_encode_dev(ns_inode->i_sb->s_dev);
> > +               }
> > +#else
> > +               netns_dev = 0;
>
> The #else branch is not needed. During is_valid_access check, if CONFIG_NET_NS
> is not defined, the program should have been rejected.
>
> > +#endif
> > +               *insn++ = BPF_MOV64_IMM(si->dst_reg, netns_dev);
> > +               break;
> > +
> > +       case offsetof(struct bpf_sock_ops, netns_ino):
> > +#ifdef CONFIG_NET_NS
> > +               /* Loading: sk_ops->sk->__sk_common.skc_net.net->ns.inum
> > +                * Type: (struct bpf_sock_ops_kern *)
> > +                *       ->(struct sock *)
> > +                *       ->(struct sock_common)
> > +                *       .possible_net_t
> > +                *       .(struct net *)
> > +                *       ->(struct ns_common)
> > +                *       .(unsigned int)
> > +                */
> > +               BUILD_BUG_ON(offsetof(struct sock, __sk_common) != 0);
> > +               BUILD_BUG_ON(offsetof(possible_net_t, net) != 0);
> > +               *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
> > +                                               struct bpf_sock_ops_kern, sk),
> > +                                     si->dst_reg, si->src_reg,
> > +                                     offsetof(struct bpf_sock_ops_kern, sk));
> > +               *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
> > +                                               possible_net_t, net),
> > +                                     si->dst_reg, si->dst_reg,
> > +                                     offsetof(struct sock_common, skc_net));
> > +               *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
> > +                                               struct ns_common, inum),
> > +                                     si->dst_reg, si->dst_reg,
> > +                                     offsetof(struct net, ns) +
> > +                                     offsetof(struct ns_common, inum));
> > +#else
> > +               *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
>
> The same as above, this else branch is not needed.
>
> > +#endif
> > +               break;
> > +
> >         }
> >         return insn - insn_buf;
> >  }
> > --
> > 2.20.1
> >