Re: [PATCH bpf-next v2] bpf: Initialize same number of free nodes for each pcpu_freelist

[Xu Kuohai]

On 11/10/2022 7:56 AM, Andrii Nakryiko wrote:
> On Tue, Nov 8, 2022 at 6:05 AM Xu Kuohai <xukuohai@xxxxxxxxxxxxxxx> wrote:
> > From: Xu Kuohai <xukuohai@xxxxxxxxxx>
> >
> > pcpu_freelist_populate() initializes nr_elems / num_possible_cpus() + 1
> > free nodes for some CPUs, and then possibly one CPU with fewer nodes,
> > followed by remaining cpus with 0 nodes. For example, when nr_elems == 256
> > and num_possible_cpus() == 32, if CPU 0 is the current cpu, CPU 0~27
> > each gets 9 free nodes, CPU 28 gets 4 free nodes, CPU 29~31 get 0 free
> > nodes, while in fact each CPU should get 8 nodes equally.
> >
> > This patch initializes nr_elems / num_possible_cpus() free nodes for each
> > CPU firstly, then allocates the remaining free nodes by one for each CPU
> > until no free nodes left.
> >
> > Signed-off-by: Xu Kuohai <xukuohai@xxxxxxxxxx>
> > Acked-by: Yonghong Song <yhs@xxxxxx>
> > ---
> > v2: Update commit message and add Yonghong's ack
> > ---
> >   kernel/bpf/percpu_freelist.c | 9 ++++++---
> >   1 file changed, 6 insertions(+), 3 deletions(-)
> >
> > diff --git a/kernel/bpf/percpu_freelist.c b/kernel/bpf/percpu_freelist.c
> > index b6e7f5c5b9ab..89e84f7381cc 100644
> > --- a/kernel/bpf/percpu_freelist.c
> > +++ b/kernel/bpf/percpu_freelist.c
> > @@ -100,12 +100,15 @@ void pcpu_freelist_populate(struct pcpu_freelist *s, void *buf, u32 elem_size,
> >                              u32 nr_elems)
> >   {
> >          struct pcpu_freelist_head *head;
> > -       int i, cpu, pcpu_entries;
> > +       int i, cpu, pcpu_entries, remain_entries;
> > +
> > +       pcpu_entries = nr_elems / num_possible_cpus();
> > +       remain_entries = nr_elems % num_possible_cpus();
> >
> > -       pcpu_entries = nr_elems / num_possible_cpus() + 1;
> >          i = 0;
> >
> >          for_each_possible_cpu(cpu) {
> > +               int j = i + pcpu_entries + (remain_entries-- > 0 ? 1 : 0);
> >   again:
> >                  head = per_cpu_ptr(s->freelist, cpu);
> >                  /* No locking required as this is not visible yet. */
> > @@ -114,7 +117,7 @@ void pcpu_freelist_populate(struct pcpu_freelist *s, void *buf, u32 elem_size,
> >                  buf += elem_size;
> >                  if (i == nr_elems)
> >                          break;
> > -               if (i % pcpu_entries)
> > +               if (i < j)
> >                          goto again;
> >          }
>
> this loop's logic is quite hard to follow, if we are fixing it, can we
> simplify it maybe? something like:
>
> int cpu, cpu_idx, i, j, n, m;
>
> n = nr_elems / num_possible_cpus();
> m = nr_elems % num_possible_cpus();
>
> for_each_possible_cpu(cpu) {
>      i = n + (cpu_idx < m ? 1 : 0);
>      for (j = 0; j < i; j++) {
>          head = per_cpu_ptr(s->freelist, cpu);
>          pcpu_freelist_push_node(head, buf);
>          buf += elem_size;
>      }
>      cpu_idx++;
> }
>
>
> no gotos, no extra ifs: for each cpu we determine correct number of
> elements to allocate, then just allocate them in a straightforward
> loop

that's great, will update to:

int cpu, cpu_idx, i, j, n, m;

n = nr_elems / num_possible_cpus();
m = nr_elems % num_possible_cpus();

for_each_possible_cpu(cpu) {
      j = min(n + (cpu_idx < m ? 1 : 0), nr_elems);
      for (i = 0; i < j; i++) {
          head = per_cpu_ptr(s->freelist, cpu);
          pcpu_freelist_push_node(head, buf);
          buf += elem_size;
      }
      nr_elems -= j;
      cpu_idx++;
}

> >   }
> > --
> > 2.30.2