Re: [RFCv2 bpf-next 1/7] bpf: xfrm: Add bpf_xdp_get_xfrm_state() kfunc

[Daniel Xu]

On Wed, Nov 01, 2023 at 06:27:03PM -0700, Alexei Starovoitov wrote:
> On Wed, Nov 1, 2023 at 2:58 PM Daniel Xu <dxu@xxxxxxxxx> wrote:
> >
> > This commit adds an unstable kfunc helper to access internal xfrm_state
> > associated with an SA. This is intended to be used for the upcoming
> > IPsec pcpu work to assign special pcpu SAs to a particular CPU. In other
> > words: for custom software RSS.
> >
> > That being said, the function that this kfunc wraps is fairly generic
> > and used for a lot of xfrm tasks. I'm sure people will find uses
> > elsewhere over time.
> >
> > Co-developed-by: Antony Antony <antony.antony@xxxxxxxxxxx>
> > Signed-off-by: Antony Antony <antony.antony@xxxxxxxxxxx>
> > Signed-off-by: Daniel Xu <dxu@xxxxxxxxx>
> > ---
> >  include/net/xfrm.h        |   9 ++++
> >  net/xfrm/Makefile         |   1 +
> >  net/xfrm/xfrm_policy.c    |   2 +
> >  net/xfrm/xfrm_state_bpf.c | 105 ++++++++++++++++++++++++++++++++++++++
> >  4 files changed, 117 insertions(+)
> >  create mode 100644 net/xfrm/xfrm_state_bpf.c
> >
> > diff --git a/include/net/xfrm.h b/include/net/xfrm.h
> > index c9bb0f892f55..1d107241b901 100644
> > --- a/include/net/xfrm.h
> > +++ b/include/net/xfrm.h
> > @@ -2190,4 +2190,13 @@ static inline int register_xfrm_interface_bpf(void)
> >
> >  #endif
> >
> > +#if IS_ENABLED(CONFIG_DEBUG_INFO_BTF)
> > +int register_xfrm_state_bpf(void);
> > +#else
> > +static inline int register_xfrm_state_bpf(void)
> > +{
> > +       return 0;
> > +}
> > +#endif
> > +
> >  #endif /* _NET_XFRM_H */
> > diff --git a/net/xfrm/Makefile b/net/xfrm/Makefile
> > index cd47f88921f5..547cec77ba03 100644
> > --- a/net/xfrm/Makefile
> > +++ b/net/xfrm/Makefile
> > @@ -21,3 +21,4 @@ obj-$(CONFIG_XFRM_USER_COMPAT) += xfrm_compat.o
> >  obj-$(CONFIG_XFRM_IPCOMP) += xfrm_ipcomp.o
> >  obj-$(CONFIG_XFRM_INTERFACE) += xfrm_interface.o
> >  obj-$(CONFIG_XFRM_ESPINTCP) += espintcp.o
> > +obj-$(CONFIG_DEBUG_INFO_BTF) += xfrm_state_bpf.o
> > diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
> > index c13dc3ef7910..1b7e75159727 100644
> > --- a/net/xfrm/xfrm_policy.c
> > +++ b/net/xfrm/xfrm_policy.c
> > @@ -4218,6 +4218,8 @@ void __init xfrm_init(void)
> >  #ifdef CONFIG_XFRM_ESPINTCP
> >         espintcp_init();
> >  #endif
> > +
> > +       register_xfrm_state_bpf();
> >  }
> >
> >  #ifdef CONFIG_AUDITSYSCALL
> > diff --git a/net/xfrm/xfrm_state_bpf.c b/net/xfrm/xfrm_state_bpf.c
> > new file mode 100644
> > index 000000000000..4aaac134b97a
> > --- /dev/null
> > +++ b/net/xfrm/xfrm_state_bpf.c
> 
> since net/xfrm/xfrm_interface_bpf.c is already there and
> was meant to be use as a file for interface between xfrm and bpf
> may be add new kfuncs there instead of new file?

IIUC xfrm_interface_bpf.c is for "xfrm interfaces" [0]. See
bpf_xfrm_info:if_id.

I could be wrong. But if you or Steffen want it all in one file I don't
really mind either way.

[0]: https://docs.strongswan.org/docs/5.9/features/routeBasedVpn.html#_xfrm_interfaces_on_linux

> 
> 
> > @@ -0,0 +1,105 @@
> > +// SPDX-License-Identifier: GPL-2.0-only
> > +/* Unstable XFRM state BPF helpers.
> > + *
> > + * Note that it is allowed to break compatibility for these functions since the
> > + * interface they are exposed through to BPF programs is explicitly unstable.
> > + */
> > +
> > +#include <linux/bpf.h>
> > +#include <linux/btf_ids.h>
> > +#include <net/xdp.h>
> > +#include <net/xfrm.h>
> > +
> > +/* bpf_xfrm_state_opts - Options for XFRM state lookup helpers
> > + *
> > + * Members:
> > + * @error      - Out parameter, set for any errors encountered
> > + *              Values:
> > + *                -EINVAL - netns_id is less than -1
> > + *                -EINVAL - Passed NULL for opts
> > + *                -EINVAL - opts__sz isn't BPF_XFRM_STATE_OPTS_SZ
> > + *                -ENONET - No network namespace found for netns_id
> > + * @netns_id   - Specify the network namespace for lookup
> > + *              Values:
> > + *                BPF_F_CURRENT_NETNS (-1)
> > + *                  Use namespace associated with ctx
> > + *                [0, S32_MAX]
> > + *                  Network Namespace ID
> > + * @mark       - XFRM mark to match on
> > + * @daddr      - Destination address to match on
> > + * @spi                - Security parameter index to match on
> > + * @proto      - L3 protocol to match on
> > + * @family     - L3 protocol family to match on
> > + */
> > +struct bpf_xfrm_state_opts {
> > +       s32 error;
> > +       s32 netns_id;
> > +       u32 mark;
> > +       xfrm_address_t daddr;
> > +       __be32 spi;
> > +       u8 proto;
> > +       u16 family;
> > +};
> > +
> > +enum {
> > +       BPF_XFRM_STATE_OPTS_SZ = sizeof(struct bpf_xfrm_state_opts),
> > +};
> > +
> > +__diag_push();
> > +__diag_ignore_all("-Wmissing-prototypes",
> > +                 "Global functions as their definitions will be in xfrm_state BTF");
> > +
> > +/* bpf_xdp_get_xfrm_state - Get XFRM state
> > + *
> > + * Parameters:
> > + * @ctx        - Pointer to ctx (xdp_md) in XDP program
> > + *                 Cannot be NULL
> > + * @opts       - Options for lookup (documented above)
> > + *                 Cannot be NULL
> > + * @opts__sz   - Length of the bpf_xfrm_state_opts structure
> > + *                 Must be BPF_XFRM_STATE_OPTS_SZ
> > + */
> > +__bpf_kfunc struct xfrm_state *
> > +bpf_xdp_get_xfrm_state(struct xdp_md *ctx, struct bpf_xfrm_state_opts *opts, u32 opts__sz)
> > +{
> > +       struct xdp_buff *xdp = (struct xdp_buff *)ctx;
> > +       struct net *net = dev_net(xdp->rxq->dev);
> > +
> > +       if (!opts || opts__sz != BPF_XFRM_STATE_OPTS_SZ) {
> > +               opts->error = -EINVAL;
> > +               return NULL;
> > +       }
> > +
> > +       if (unlikely(opts->netns_id < BPF_F_CURRENT_NETNS)) {
> > +               opts->error = -EINVAL;
> > +               return NULL;
> > +       }
> > +
> > +       if (opts->netns_id >= 0) {
> > +               net = get_net_ns_by_id(net, opts->netns_id);
> 
> netns is leaking :(

Argh, will fix.

> 
> > +               if (unlikely(!net)) {
> > +                       opts->error = -ENONET;
> > +                       return NULL;
> > +               }
> > +       }
> > +
> > +       return xfrm_state_lookup(net, opts->mark, &opts->daddr, opts->spi,
> > +                                opts->proto, opts->family);
> 
> After looking into xfrm internals realized that
> refcnt inc/dec and KF_ACQUIRE maybe unnecessary overhead.
> XDP progs run under rcu_read_lock.
> I think you can make a version of __xfrm_state_lookup()
> without xfrm_state_hold_rcu() and avoid two atomics per packet,
> but such xfrm_state may have refcnt==0.
> Since bpf prog will only read from there and won't pass it anywhere
> else it might be ok.
> 
> But considering the rest of ipsec overhead this might be
> a premature optimization and it's better to stay with clean
> acquire/release semantics.

Yeah I think acquire/release will be fine for now. I've been doing a lot
of profiling on tput tests and xdp barely shows up. There are some lower
hanging fruit atm.

> As far as IETF:
> https://datatracker.ietf.org/doc/html/draft-ietf-ipsecme-multi-sa-performance-02
> it's not clear to me why one Child SA (without new pcpu field)
> has to be handled by one cpu.
> 
> Sounds like it's possible to implement differently. At least in SW.
> In HW, I can see how duplicating multiple crypto state and the rest
> in a single queue is difficult.

The child SA without pcpu flag (aka fallback SA) is not pinned to a cpu.
It's shared between all of the cpus. I think Steffen can probably
explain this better if that's not clear.

Thanks,
Daniel