On 03.05.2017 00:13, Eric W. Biederman wrote:
> Kirill Tkhai <ktkhai@xxxxxxxxxxxxx> writes:
>
>> On 02.05.2017 19:33, Oleg Nesterov wrote:
>>> sorry for delay, vacation...
>>>
>>> On 04/28, Kirill Tkhai wrote:
>>>>
>>>> On 27.04.2017 19:22, Oleg Nesterov wrote:
>>>>>
>>>>> Ah, OK, I didn't notice the ns->child_reaper check in pidns_for_children_get().
>>>>>
>>>>> But note that it doesn't need tasklist_lock too.
>>>>
>>>> Hm, are there possible strange situations with memory ordering, when we see
>>>> ns->child_reaper of already died ns, which was placed in the same memory?
>>>> Do we have to use some memory barriers here?
>>>
>>> Could you spell please? I don't understand your concerns...
>>>
>>> I don't see how, say,
>>>
>>> static struct ns_common *pidns_for_children_get(struct task_struct *task)
>>> {
>>> struct ns_common *ns = NULL;
>>> struct pid_namespace *pid_ns;
>>>
>>> task_lock(task);
>>> if (task->nsproxy) {
>>> pid_ns = task->nsproxy->pid_ns_for_children;
>>> if (pid_ns->child_reaper) {
> ^^^^^^^^^^^^^^^^^^^^
> Oleg my apologies I missed this line earlier.
> This does look like a valid way to skip read_lock(&tasklist_lock);
>>> ns = &pid_ns->ns;
>>> get_pid_ns(ns);
> ^^^^^^^^^^^^^ This needs to be:
> get_pid_ns(pid_ns);
>
>>> }
>>> }
>>> task_unlock(task);
>>>
>>> return ns;
>>> }
>>>
>>> can be wrong. It also looks more clean to me.
>>>
>>> ->child_reaper is not stable without tasklist, it can be dead/etc, but
>>> we do not care?
>>
>> I mean the following. We had a pid_ns1 with a child_reaper set. Then
>> it became dead, and a new pid_ns2 were allocated in the same memory.
>
> task->nsproxy->pid_ns_for_children is always changed with
> task_lock(task) held. See switch_task_namespaces (used by unshare and
> setns). This also gives us the guarantee that the pid_ns reference
> won't be freed/reused in any for until task_lock(task) is dropped.
Now I've checked kmem_cache_zalloc() and it looks like it zeroes cache memory
content synchronous on allocation (it seems there is no pre-zeroed memory
for GFP_ZERO cases).
So, the zeroing happens before switch_task_namespaces() (and task_unlock())
and we're really safe after task_lock() in pidns_for_children_get().
Ok, I'll send new version of the patchset.
