linux: move reset cpu affinity to scheduler

and set it only when there is no exec cpu affinity mask specified.

Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>

src/libcrun/cgroup-setup.c

@@ -506,20 +506,5 @@ enter_cgroup (int cgroup_mode, pid_t pid, pid_t init_pid, const char *path,
       if (UNLIKELY (ret < 0))
         return ret;
     }
-  /* Reset the inherited cpu affinity. Old kernels do that automatically, but
-     new kernels remember the affinity that was set before the cgroup move.
-     This is undesirable, because it inherits the systemd affinity when the container
-     should really move to the container space cpus.
-
-     The sched_setaffinity call will always return an error (EINVAL or ENODEV)
-     when used like this. This is expected and part of the backward compatibility.
-
-     See: https://issues.redhat.com/browse/OCPBUGS-15102   */
-  ret = sched_setaffinity (pid, 0, NULL);
-  if (LIKELY (ret < 0))
-    {
-      if (UNLIKELY (! ((errno == EINVAL) || (errno == ENODEV))))
-        return crun_make_error (err, errno, "failed to reset affinity");
-    }
   return 0;
 }

src/libcrun/cgroup.c

@@ -334,22 +334,6 @@ libcrun_cgroup_enter (struct libcrun_cgroup_args *args, struct libcrun_cgroup_st
             return ret;
         }
     }
-  /* Reset the inherited cpu affinity. Old kernels do that automatically, but
-     new kernels remember the affinity that was set before the cgroup move.
-     This is undesirable, because it inherits the systemd affinity when the container
-     should really move to the container space cpus.
-
-     The sched_setaffinity call will always return an error (EINVAL or ENODEV)
-     when used like this. This is expected and part of the backward compatibility.
-
-     See: https://issues.redhat.com/browse/OCPBUGS-15102   */
-  ret = sched_setaffinity (args->pid, 0, NULL);
-  if (LIKELY (ret < 0))
-    {
-      if (UNLIKELY (! ((errno == EINVAL) || (errno == ENODEV))))
-        return crun_make_error (err, errno, "failed to reset affinity");
-    }
-
 success:
   *out = status;
   status = NULL;

src/libcrun/container.c

@@ -2590,6 +2590,10 @@ libcrun_container_run_internal (libcrun_container_t *container, libcrun_context_
   if (UNLIKELY (ret < 0))
     goto fail;
 
+  ret = libcrun_reset_cpu_affinity_mask (pid, err);
+  if (UNLIKELY (ret < 0))
+    goto fail;
+
   ret = libcrun_set_io_priority (pid, def->process, err);
   if (UNLIKELY (ret < 0))
     goto fail;

src/libcrun/linux.c

@@ -4738,6 +4738,15 @@ handle_pidfd_receiver (pid_t pid, libcrun_container_t *container, libcrun_error_
   return send_fd_to_socket (client_fd, pidfd, err);
 }
 
+static bool
+has_exec_cpu_affinity (runtime_spec_schema_config_schema_process *process)
+{
+  if (process == NULL || process->exec_cpu_affinity == NULL)
+    return false;
+  return (! is_empty_string (process->exec_cpu_affinity->initial))
+         || (! is_empty_string (process->exec_cpu_affinity->final));
+}
+
 pid_t
 libcrun_run_linux_container (libcrun_container_t *container, container_entrypoint_t entrypoint, void *args,
                              int *sync_socket_out, struct libcrun_dirfd_s *cgroup_dirfd, libcrun_error_t *err)
@@ -5060,6 +5069,13 @@ join_process_parent_helper (libcrun_context_t *context,
         return ret;
     }
 
+  if (! has_exec_cpu_affinity (process))
+    {
+      ret = libcrun_reset_cpu_affinity_mask (pid, err);
+      if (UNLIKELY (ret < 0))
+        return ret;
+    }
+
   if (need_move_to_cgroup)
     {
       if (sub_cgroup)

src/libcrun/scheduler.c

@@ -78,6 +78,30 @@ syscall_sched_setattr (pid_t pid, struct sched_attr_s *attr, unsigned int flags)
 #endif
 }
 
+int
+libcrun_reset_cpu_affinity_mask (pid_t pid, libcrun_error_t *err)
+{
+  int ret;
+
+  /* Reset the inherited cpu affinity. Old kernels do that automatically, but
+     new kernels remember the affinity that was set before the cgroup move.
+     This is undesirable, because it inherits the systemd affinity when the container
+     should really move to the container space cpus.
+
+     The sched_setaffinity call will always return an error (EINVAL or ENODEV)
+     when used like this. This is expected and part of the backward compatibility.
+
+     See: https://issues.redhat.com/browse/OCPBUGS-15102   */
+  ret = sched_setaffinity (pid, 0, NULL);
+  if (LIKELY (ret < 0))
+    {
+      if (UNLIKELY (! ((errno == EINVAL) || (errno == ENODEV))))
+        return crun_make_error (err, errno, "failed to reset affinity");
+    }
+
+  return 0;
+}
+
 int
 libcrun_set_scheduler (pid_t pid, runtime_spec_schema_config_schema_process *process, libcrun_error_t *err)
 {

src/libcrun/scheduler.h

@@ -26,4 +26,6 @@ int libcrun_set_scheduler (pid_t pid, runtime_spec_schema_config_schema_process
 
 int libcrun_set_cpu_affinity_from_string (pid_t pid, const char *str, libcrun_error_t *err);
 
+int libcrun_reset_cpu_affinity_mask (pid_t pid, libcrun_error_t *err);
+
 #endif