On 10/07/2013 05:17 PM, Daniel Borkmann wrote:

On 10/07/2013 05:07 AM, Gao feng wrote:

quoted

On 10/05/2013 02:20 AM, Daniel Borkmann wrote:

quoted

+static void cgroup_attach(struct cgroup_subsys_state *css,
+              struct cgroup_taskset *tset)
+{
+    struct task_struct *p;
+    void *v;
+
+    cgroup_taskset_for_each(p, css, tset) {
+        task_lock(p);
+        v = (void *)(unsigned long) task_fwid(p);

Shouldn't v be css_nf_state(css)->fwid?

Nope, this is in line with net_cls and net_prio; the task has been
moved there via cgroup backend already through cgroup_attach_task(),

Yes, these tasks have already been migrated to this cgroup.

so we only need to update each of it's socket sk_cgrp_fwid parts.

Sorry, I still don't know in which situation that css_nf_state(css)->fwid
isn't equal to task_fwid(p).

two threads write the same pid to different cgroup at the same time?
it seems can not happen since we have cgroup_mutex protected.

css is not strictly for net_filter. See also: 6a328d8c6f (cgroup:
net_cls: Rework update socket logic)

quoted

quoted

+        iterate_fd(p->files, 0, cgroup_fwid_update, v);
+        task_unlock(p);
+    }
+}