Hi Mimi,
On 8/23/21 1:57 PM, Mimi Zohar wrote:
> On Mon, 2021-08-23 at 08:14 +0000, THOBY Simon wrote:
>> Hi Liqiong,
>>
>> On 8/23/21 10:06 AM, liqiong wrote:
>>> Hi Simon :
>>>
>>> Using a temporary ima_rules variable is not working for "ima_policy_next".
>>>
>>> void *ima_policy_next(struct seq_file *m, void *v, loff_t *pos)
>>> {
>>> struct ima_rule_entry *entry = v;
>>> -
>>> + struct list_head *ima_rules_tmp = rcu_dereference(ima_rules);
>>> rcu_read_lock();
>>> entry = list_entry_rcu(entry->list.next, struct ima_rule_entry, list);
>>> rcu_read_unlock();
>>> (*pos)++;
>>>
>>> - return (&entry->list == ima_rules) ? NULL : entry;
>>> + return (&entry->list == ima_rules_tmp) ? NULL : entry;
>>> }
>>>
>>> It seems no way to fix "ima_rules" change within this function, it will alway
>>> return a entry if "ima_rules" being changed.
>>
>> - I think rcu_dereference() should be called inside the RCU read lock
>> - Maybe we could cheat with:
>> return (&entry->list == &ima_policy_rules || &entry->list == &ima_default_rules) ? NULL : entry;
>> as that's the only two rulesets IMA ever use?
>> Admittedly, this is not as clean as previously, but it should work too.
>>
>> The way I see it, the semaphore solution would not work here either,
>> as ima_policy_next() is called repeatedly as a seq_file
>> (it is set up in ima_fs.c) and we can't control the locking there:
>> we cannot lock across the seq_read() call (that cure could end up be
>> worse than the disease, deadlock-wise), so I fear we cannot protect
>> against a list update while a user is iterating with a lock.
>>
>> So in both cases a cheat like "&entry->list == &ima_policy_rules || &entry->list == &ima_default_rules"
>> maybe need to be considered.
>>
>> What do you think?
>
> Is this an overall suggestion or limited to just ima_policy_next()?
I was thinking only of ima_policy_next(), I don't think (from what I could see in a short glance)
that other functions need such a treatment. The ima_rules_tmp dance is probably safe for the
other uses of ima_rules.
>
> thanks,
>
> Mimi
>
>
Thanks,
Simon
