On 2014/6/9 17:13, David Rientjes wrote: > On Mon, 9 Jun 2014, Gu Zheng wrote: > >>> I think your patch addresses the problem that you're reporting but misses >>> the larger problem with cpuset.mems rebinding on fork(). When the >>> forker's task_struct is duplicated (which includes ->mems_allowed) and it >>> races with an update to cpuset_being_rebound in update_tasks_nodemask() >>> then the task's mems_allowed doesn't get updated. >> >> Yes, you are right, this patch just wants to address the bug reported above. >> The race condition you mentioned above inherently exists there, but it is yet >> another issue, the rcu lock here makes no sense to it, and I think we need >> additional sync-mechanisms if want to fix it. > > Yes, the rcu lock is not providing protection for any critical section > here that requires (1) the forker's cpuset to be stored in > cpuset_being_rebound or (2) the forked thread's cpuset to be rebound by > the cpuset nodemask update, and no race involving the two. > Yes, this is a long-standing issue. Besides the race you described, the child task's mems_allowed can be wrong if the cpuset's nodemask changes before the child has been added to the cgroup's tasklist. I remember Tejun once said he wanted to disallow task migration between cgroups during fork, and that should fix this problem. >> But thinking more, though the current implementation has flaw, but I worry >> about the negative effect if we really want to fix it. Or maybe the fear >> is unnecessary.:) >> > > It needs to be slightly rewritten to work properly without negatively > impacting the latency of fork(). Do you have the cycles to do it? > Sounds you have other idea? -- To unsubscribe from this list: send the line "unsubscribe stable" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html