Re: [PATCH 1/2] bpf: Introduce cpu affinity for sockmap

[Andrii Nakryiko]

On Sun, Nov 3, 2024 at 10:12 PM Jiayuan Chen <mrpre@xxxxxxx> wrote:
>
> On Fri, Nov 01, 2024 at 12:25:51PM -0700, Andrii Nakryiko wrote:
> > On Thu, Oct 31, 2024 at 7:40 PM mrpre <mrpre@xxxxxxx> wrote:
> > >
> > > Why we need cpu affinity:
> > > Mainstream data planes, like Nginx and HAProxy, utilize CPU affinity
> > > by binding user processes to specific CPUs. This avoids interference
> > > between processes and prevents impact from other processes.
> > >
> > > Sockmap, as an optimization to accelerate such proxy programs,
> > > currently lacks the ability to specify CPU affinity. The current
> > > implementation of sockmap handling backlog is based on workqueue,
> > > which operates by calling 'schedule_delayed_work()'. It's current
> > > implementation prefers to schedule on the local CPU, i.e., the CPU
> > > that handled the packet under softirq.
> > >
> > > For extremely high traffic with large numbers of packets,
> > > 'sk_psock_backlog' becomes a large loop.
> > >
> > > For multi-threaded programs with only one map, we expect different
> > > sockets to run on different CPUs. It is important to note that this
> > > feature is not a general performance optimization. Instead, it
> > > provides users with the ability to bind to specific CPU, allowing
> > > them to enhance overall operating system utilization based on their
> > > own system environments.
> > >
> > > Implementation:
> > > 1.When updating the sockmap, support passing a CPU parameter and
> > > save it to the psock.
> > > 2.When scheduling psock, determine which CPU to run on using the
> > > psock's CPU information.
> > > 3.For thoes sockmap without CPU affinity, keep original logic by using
> > > 'schedule_delayed_work()'.
> > >
> > > Performance Testing:
> > > 'client <-> sockmap proxy <-> server'
> > >
> > > Using 'iperf3' tests, with the iperf server bound to CPU5 and the iperf
> > > client bound to CPU6, performance without using CPU affinity is
> > > around 34 Gbits/s, and CPU usage is concentrated on CPU5 and CPU6.
> > > '''
> > > [  5] local 127.0.0.1 port 57144 connected to 127.0.0.1 port 10000
> > > [ ID] Interval           Transfer     Bitrate
> > > [  5]   0.00-1.00   sec  3.95 GBytes  33.9 Gbits/sec
> > > [  5]   1.00-2.00   sec  3.95 GBytes  34.0 Gbits/sec
> > > ......
> > > '''
> > >
> > > With using CPU affinity, the performnce is close to direct connection
> > > (without any proxy).
> > > '''
> > > [  5] local 127.0.0.1 port 56518 connected to 127.0.0.1 port 10000
> > > [ ID] Interval           Transfer     Bitrate
> > > [  5]   0.00-1.00   sec  7.76 GBytes  66.6 Gbits/sec
> > > [  5]   1.00-2.00   sec  7.76 GBytes  66.7 Gbits/sec
> > > ......
> > > '''
> > >
> > > Signed-off-by: Jiayuan Chen <mrpre@xxxxxxx>
> > > ---
> > >  include/linux/bpf.h      |  3 ++-
> > >  include/linux/skmsg.h    |  8 ++++++++
> > >  include/uapi/linux/bpf.h |  4 ++++
> > >  kernel/bpf/syscall.c     | 23 +++++++++++++++++------
> > >  net/core/skmsg.c         | 11 +++++++----
> > >  net/core/sock_map.c      | 12 +++++++-----
> > >  6 files changed, 45 insertions(+), 16 deletions(-)
> > >
> > > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > > index c3ba4d475174..a56028c389e7 100644
> > > --- a/include/linux/bpf.h
> > > +++ b/include/linux/bpf.h
> > > @@ -3080,7 +3080,8 @@ int bpf_prog_test_run_syscall(struct bpf_prog *prog,
> > >
> > >  int sock_map_get_from_fd(const union bpf_attr *attr, struct bpf_prog *prog);
> > >  int sock_map_prog_detach(const union bpf_attr *attr, enum bpf_prog_type ptype);
> > > -int sock_map_update_elem_sys(struct bpf_map *map, void *key, void *value, u64 flags);
> > > +int sock_map_update_elem_sys(struct bpf_map *map, void *key, void *value, u64 flags,
> > > +                            s32 target_cpu);
> > >  int sock_map_bpf_prog_query(const union bpf_attr *attr,
> > >                             union bpf_attr __user *uattr);
> > >  int sock_map_link_create(const union bpf_attr *attr, struct bpf_prog *prog);
> > > diff --git a/include/linux/skmsg.h b/include/linux/skmsg.h
> > > index d9b03e0746e7..919425a92adf 100644
> > > --- a/include/linux/skmsg.h
> > > +++ b/include/linux/skmsg.h
> > > @@ -117,6 +117,7 @@ struct sk_psock {
> > >         struct delayed_work             work;
> > >         struct sock                     *sk_pair;
> > >         struct rcu_work                 rwork;
> > > +       s32                             target_cpu;
> > >  };
> > >
> > >  int sk_msg_alloc(struct sock *sk, struct sk_msg *msg, int len,
> > > @@ -514,6 +515,13 @@ static inline bool sk_psock_strp_enabled(struct sk_psock *psock)
> > >         return !!psock->saved_data_ready;
> > >  }
> > >
> > > +static inline int sk_psock_strp_get_cpu(struct sk_psock *psock)
> > > +{
> > > +       if (psock->target_cpu != -1)
> > > +               return psock->target_cpu;
> > > +       return WORK_CPU_UNBOUND;
> > > +}
> > > +
> > >  #if IS_ENABLED(CONFIG_NET_SOCK_MSG)
> > >
> > >  #define BPF_F_STRPARSER        (1UL << 1)
> > > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> > > index f28b6527e815..2019a87b5d4a 100644
> > > --- a/include/uapi/linux/bpf.h
> > > +++ b/include/uapi/linux/bpf.h
> > > @@ -1509,6 +1509,10 @@ union bpf_attr {
> > >                         __aligned_u64 next_key;
> > >                 };
> > >                 __u64           flags;
> > > +               union {
> > > +                       /* specify the CPU where the sockmap job run on */
> > > +                       __aligned_u64 target_cpu;
> >
> > I have no opinion on the feature itself, I'll leave this to others.
> > But from UAPI perspective:
> >
> > a) why is this a u64 and not, say, int?
> > b) maybe we should just specify this as flags and not have to update
> > all the UAPIs (including libbpf-side)? Just add a new
> > BPF_F_SOCKNMAP_TARGET_CPU flag or something, and specify that highest
> > 32 bits specify the CPU itself?
> >
> > We have similar schema for some other helpers, so not *that* unusual.
> >
> Thank you for your response. I think I should clarify my thoughts:
>
> My idea is to pass a user-space pointer, with the pointer being null
> to indicate that the user has not provided anything.For example, when
> users use the old interface 'bpf_map_update_elem' and pass in u64 of
> 0, it means that the user hasn't specified a CPU. If a u32 or another
> type of value is passed in, when it is 0, it's ambiguous whether this
> indicates target CPU 0 or that the user hasn't provided a value. So
> my design involves passing a user-space pointer.
>
> I also considered using the highest 32 bits of the flag as target_cpu, but
> this approach still encounters the ambiguity mentioned above. Of course
> for programs using libbpf, I can naturally init all the higher 32 bits
> default to 1 to indicate the user hasn't specified a CPU, but this is
> incompatible with programs not using libbpf. Another approach could be
> that a value of 1 for the higher 32 bits indicates CPU 0, and 2 indicates
> CPU 1..., but this seems odd and would require a helper to assist users
> in passing arguments.

See BPF_F_SOCKMAP_TARGET_CPU flag point in my reply. You need an extra
flag that would specify that those 32 bits are specifying a CPU
number. There is no ambiguity. No flag - no CPU, Flag - CPU (even if
zero).

>
> There is another method, like providing an extra 'attr', to replace the
> passed 'target_cpu', which maintains the general nature of
> 'map_update_elem' interface, like:
> '''
> +struct extra_bpf_attr {
> +    u32 target_cpu;
> +};
> struct { /* anonymous struct used by BPF_MAP_*_ELEM commands */
>     __u32   map_fd;
>     __aligned_u64 key;
>     union {
>         __aligned_u64 value;
>         __aligned_u64 next_key;
>     };
>     __u64   flags;
>     +struct extra_bpf_attr extra;
> };
>
> static int bpf_map_update_value(struct bpf_map *map, struct file *map_file,
> -                               void *key, void *value, __u64 flags)
> +                               void *key, void *value, __u64 flags, struct bpf_attr_extra *extra);
> '''
>
> > > +               };
> > >         };
> > >
> > >         struct { /* struct used by BPF_MAP_*_BATCH commands */
> > > diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> > > index 8254b2973157..95f719b9c3f3 100644
> > > --- a/kernel/bpf/syscall.c
> > > +++ b/kernel/bpf/syscall.c
> > > @@ -239,10 +239,9 @@ static int bpf_obj_pin_uptrs(struct btf_record *rec, void *obj)
> > >  }
> > >
> > >  static int bpf_map_update_value(struct bpf_map *map, struct file *map_file,
> > > -                               void *key, void *value, __u64 flags)
> > > +                               void *key, void *value, __u64 flags, s32 target_cpu)
> >
> > yeah, this is what I'm talking about. Think how ridiculous it is for a
> > generic "BPF map update" operation to accept the "target_cpu"
> > parameter.
> >
> > pw-bot: cr
> >
> > >  {
> > >         int err;
> > > -
> >
> > why? don't break whitespace formatting
> >
> > >         /* Need to create a kthread, thus must support schedule */
> > >         if (bpf_map_is_offloaded(map)) {
> > >                 return bpf_map_offload_update_elem(map, key, value, flags);
> > > @@ -252,7 +251,7 @@ static int bpf_map_update_value(struct bpf_map *map, struct file *map_file,
> > >                 return map->ops->map_update_elem(map, key, value, flags);
> > >         } else if (map->map_type == BPF_MAP_TYPE_SOCKHASH ||
> > >                    map->map_type == BPF_MAP_TYPE_SOCKMAP) {
> > > -               return sock_map_update_elem_sys(map, key, value, flags);
> > > +               return sock_map_update_elem_sys(map, key, value, flags, target_cpu);
> > >         } else if (IS_FD_PROG_ARRAY(map)) {
> > >                 return bpf_fd_array_map_update_elem(map, map_file, key, value,
> > >                                                     flags);
> >
> > [...]
>
>