Re: [PATCH bpf-next 4/4] bpf: runqslower: use task local storage

[Yonghong Song]

On 1/11/21 11:14 PM, Andrii Nakryiko wrote:
> On Mon, Jan 11, 2021 at 7:24 PM Yonghong Song <yhs@xxxxxx> wrote:
> > On 1/11/21 2:54 PM, Song Liu wrote:
> > > > On Jan 11, 2021, at 9:49 AM, Yonghong Song <yhs@xxxxxx> wrote:
> > > >
> > > >
> > > >
> > > > On 1/8/21 3:19 PM, Song Liu wrote:
> > > > > Replace hashtab with task local storage in runqslower. This improves the
> > > > > performance of these BPF programs. The following table summarizes average
> > > > > runtime of these programs, in nanoseconds:
> > > > >                             task-local   hash-prealloc   hash-no-prealloc
> > > > > handle__sched_wakeup             125             340               3124
> > > > > handle__sched_wakeup_new        2812            1510               2998
> > > > > handle__sched_switch             151             208                991
> > > > > Note that, task local storage gives better performance than hashtab for
> > > > > handle__sched_wakeup and handle__sched_switch. On the other hand, for
> > > > > handle__sched_wakeup_new, task local storage is slower than hashtab with
> > > > > prealloc. This is because handle__sched_wakeup_new accesses the data for
> > > > > the first time, so it has to allocate the data for task local storage.
> > > > > Once the initial allocation is done, subsequent accesses, as those in
> > > > > handle__sched_wakeup, are much faster with task local storage. If we
> > > > > disable hashtab prealloc, task local storage is much faster for all 3
> > > > > functions.
> > > > > Signed-off-by: Song Liu <songliubraving@xxxxxx>
> > > > > ---
> > > > >    tools/bpf/runqslower/runqslower.bpf.c | 26 +++++++++++++++-----------
> > > > >    1 file changed, 15 insertions(+), 11 deletions(-)
> > > > > diff --git a/tools/bpf/runqslower/runqslower.bpf.c b/tools/bpf/runqslower/runqslower.bpf.c
> > > > > index 1f18a409f0443..c4de4179a0a17 100644
> > > > > --- a/tools/bpf/runqslower/runqslower.bpf.c
> > > > > +++ b/tools/bpf/runqslower/runqslower.bpf.c
> > > > > @@ -11,9 +11,9 @@ const volatile __u64 min_us = 0;
> > > > >    const volatile pid_t targ_pid = 0;
> > > > >      struct {
> > > > > -   __uint(type, BPF_MAP_TYPE_HASH);
> > > > > -   __uint(max_entries, 10240);
> > > > > -   __type(key, u32);
> > > > > +   __uint(type, BPF_MAP_TYPE_TASK_STORAGE);
> > > > > +   __uint(map_flags, BPF_F_NO_PREALLOC);
> > > > > +   __type(key, int);
> > > > >      __type(value, u64);
> > > > >    } start SEC(".maps");
> > > > >    @@ -25,15 +25,19 @@ struct {
> > > > >      /* record enqueue timestamp */
> > > > >    __always_inline
> > > > > -static int trace_enqueue(u32 tgid, u32 pid)
> > > > > +static int trace_enqueue(struct task_struct *t)
> > > > >    {
> > > > > -   u64 ts;
> > > > > +   u32 pid = t->pid;
> > > > > +   u64 ts, *ptr;
> > > > >              if (!pid || (targ_pid && targ_pid != pid))
> > > > >              return 0;
> > > > >              ts = bpf_ktime_get_ns();
> > > > > -   bpf_map_update_elem(&start, &pid, &ts, 0);
> > > > > +   ptr = bpf_task_storage_get(&start, t, 0,
> > > > > +                              BPF_LOCAL_STORAGE_GET_F_CREATE);
> > > > > +   if (ptr)
> > > > > +           *ptr = ts;
> > > > >      return 0;
> > > > >    }
> > > > >    @@ -43,7 +47,7 @@ int handle__sched_wakeup(u64 *ctx)
> > > > >      /* TP_PROTO(struct task_struct *p) */
> > > > >      struct task_struct *p = (void *)ctx[0];
> > > > >    - return trace_enqueue(p->tgid, p->pid);
> > > > > +   return trace_enqueue(p);
> > > > >    }
> > > > >      SEC("tp_btf/sched_wakeup_new")
> > > > > @@ -52,7 +56,7 @@ int handle__sched_wakeup_new(u64 *ctx)
> > > > >      /* TP_PROTO(struct task_struct *p) */
> > > > >      struct task_struct *p = (void *)ctx[0];
> > > > >    - return trace_enqueue(p->tgid, p->pid);
> > > > > +   return trace_enqueue(p);
> > > > >    }
> > > > >      SEC("tp_btf/sched_switch")
> > > > > @@ -70,12 +74,12 @@ int handle__sched_switch(u64 *ctx)
> > > > >              /* ivcsw: treat like an enqueue event and store timestamp */
> > > > >      if (prev->state == TASK_RUNNING)
> > > > > -           trace_enqueue(prev->tgid, prev->pid);
> > > > > +           trace_enqueue(prev);
> > > > >              pid = next->pid;
> > > > >              /* fetch timestamp and calculate delta */
> > > > > -   tsp = bpf_map_lookup_elem(&start, &pid);
> > > > > +   tsp = bpf_task_storage_get(&start, next, 0, 0);
> > > > >      if (!tsp)
> > > > >              return 0;   /* missed enqueue */
> > > >
> > > > Previously, hash table may overflow so we may have missed enqueue.
> > > > Here with task local storage, is it possible to add additional pid
> > > > filtering in the beginning of handle__sched_switch such that
> > > > missed enqueue here can be treated as an error?
> > >
> > > IIUC, hashtab overflow is not the only reason of missed enqueue. If the
> > > wakeup (which calls trace_enqueue) happens before runqslower starts, we
> > > may still get missed enqueue in sched_switch, no?
> >
> > the wakeup won't happen before runqslower starts since runqslower needs
> > to start to do attachment first and then trace_enqueue() can run.
>
> I think Song is right. Given wakeup and sched_switch need to be
> matched, depending at which exact time we attach BPF programs, we can
> end up missing wakeup, but not missing sched_switch, no? So it's not
> an error.

The current approach works fine. What I suggested is to
tighten sched_switch only for target_pid. wakeup (doing queuing) will
be more relaxed than sched_switch to ensure task local storage creation
is always there for target_pid regardless of attachment timing.
I think it should work, but we have to experiment to see actual
results...

> > For the current implementation trace_enqueue() will happen for any non-0
> > pid before setting test_progs tgid, and will happen for any non-0 and
> > test_progs tgid if it is set, so this should be okay if we do filtering
> > in handle__sched_switch. Maybe you can do an experiment to prove whether
> > my point is correct or not.
> >
> > > Thanks,
> > > Song