Development version of BIND 9 - 9.21.10 with meson build system

[Havard Eidnes]

> We just call function uv_available_parallelism() from libuv and
> nothing else. If it returns weird values for NetBSD then it needs fix
> in libuv, I'm afraid.

At least the src/unix/netbsd.c file in libuv has inside

int uv_cpu_info(uv_cpu_info_t** cpu_infos, int* count) {

this section which should do the right thing:

  size = sizeof(numcpus);
  if (sysctlbyname("hw.ncpu", &numcpus, &size, NULL, 0))
    return UV__ERR(errno);
  *count = numcpus;

So at least the code is there.  Whether it works as assumed is of
course another matter:

$ cat t.c
#include <stdio.h>

#include <uv.h>

int
main(int argc, char **argv)
{
        int p;

        p = uv_available_parallelism();
        printf("libuv says ncpus = %d\n", p);
        return 0;
}
$ cc -I /usr/pkg/include t.c -L /usr/pkg/lib -R /usr/pkg/lib -luv
$ ./a.out
libuv says ncpus = 1
$ 

However, pointing the debugger at this reveals that the above
NetBSD-specific uv_cpu_info() doesn't get used in this case, but
instead this code from src/unix/core.c gets to run:

#elif defined(__NetBSD__)
  cpuset_t* set = cpuset_create();
  if (set != NULL) {
    if (0 == sched_getaffinity_np(getpid(), cpuset_size(set), set))
      rc = uv__cpu_count(set);
    cpuset_destroy(set);
  }
#elif ...

which is ... not the same.  This code looks similar to the
__FreeBSD__ section just before it:

#elif defined(__FreeBSD__)
  cpuset_t set;

  memset(&set, 0, sizeof(set));

  if (0 == cpuset_getaffinity(CPU_LEVEL_WHICH, CPU_WHICH_PID, -1, sizeof(set), &set))
    rc = uv__cpu_count(&set);
#elif ...

but is different from what's done on Apple:

#elif defined(__APPLE__)
  int nprocs;
  size_t i;
  size_t len = sizeof(nprocs);
  static const char *mib[] = {
    "hw.activecpu",
    "hw.logicalcpu",
    "hw.ncpu"
  };

  for (i = 0; i < ARRAY_SIZE(mib); i++) {
    if (0 == sysctlbyname(mib[i], &nprocs, &len, NULL, 0) &&
              len == sizeof(nprocs) &&
              nprocs > 0) {
      rc = nprocs;
      break;
    }
  }
#elif ...

Linux does a different and quite a bit more complicated thing,
among what it could end up doing:

  /*
   * Traverse up the cgroup v2 hierarchy, starting from the current cgroup path.
   * At each level, attempt to read the "cpu.max" file, which defines the CPU
   * quota and period.
   *
   * This reflects how Linux applies cgroup limits hierarchically.
   *
   * e.g: given a path like /sys/fs/cgroup/foo/bar/baz, we check:
   *   - /sys/fs/cgroup/foo/bar/baz/cpu.max
   *   - /sys/fs/cgroup/foo/bar/cpu.max
   *   - /sys/fs/cgroup/foo/cpu.max
   *   - /sys/fs/cgroup/cpu.max
   */

If I run gdb correctly, the NetBSD sched_getaffinity_np() returns the
empty set, because no "CPU affinity group" has been configured by the
system administrator.  It would then make sense, I think, to fall back
on the hw.ncpu code, i.e. use what's available.  I'll discuss this
some with co-developers of NetBSD and push to get this fixed.

Regards,

- Håvard