On 2012/9/13 14:57, Daniel Wagner wrote:

Hi Li,

On 13.09.2012 08:41, Li Zefan wrote:

quoted

quoted

@@ -1321,11 +1321,13 @@ static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
       * take duplicate reference counts on a subsystem that's already used,
       * but rebind_subsystems handles this case.
       */
-    for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
+    for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
          unsigned long bit = 1UL << i;

          if (!(bit & opts->subsys_mask))
              continue;
+        if (!subsys[i]->module)
+            continue;

This check is not necessary. If it's builtin, try_module_get() will just return 1, and
we're fine.

Yes, I didn't see the try_module_get. Although I think with leaving the test away it would change the behavior, e.g.

        if (!subsys[i]->module)
            continue;
        if (!try_module_get(subsys[i]->module)) {
            module_pin_failed = true;
            break;
        }

module_pin_failed would be set then and we would jump into the error code later.

no behavioral change. For a builtin subsys, we won't run into the if block and have module_pin_failed be set.

This tests looks a bit ugly though I think leaving it away and relying on try_module_get() is not correct.

I don't think this is bad. The block layer code does the similar thing in elevator_get().

And we call module_put() unconditionally in rebind_subsys().

quoted

quoted

@@ -1437,6 +1443,7 @@ static void init_cgroup_housekeeping(struct cgroup *cgrp)
      INIT_LIST_HEAD(&cgrp->event_list);
      spin_lock_init(&cgrp->event_list_lock);
      simple_xattrs_init(&cgrp->xattrs);
+    memset(cgrp->subsys, 0, sizeof(cgrp->subsys));

This seems an unrelated change, and is redundant. Am I missing something?

The reason why it is necessary to NULL all the entries in the array, is that task_cls_classid() and task_netprioidx() check the return pointer from task_subsys_state(). If it is NULL those function know that the subsystem is not ready to be used. Should I move this change to the next patch then?

It's already guaranteed the passing @cgrp is zeored. that's why cgrp->flags is not explicitly initialized here.