Hi Simon,
On 2021/8/20 21:23, THOBY Simon wrote:
Hi Liqiong,
On 8/20/21 12:15 PM, 李力琼 wrote:
Hi, Simon:
This solution is better then rwsem, a temp "ima_rules" variable should
can fix. I also have a another idea, with a little trick, default list
can traverse to the new list, so we don't need care about the read side.
here is the patch:
@@ -918,8 +918,21 @@ void ima_update_policy(void)
list_splice_tail_init_rcu(&ima_temp_rules, policy, synchronize_rcu);
if (ima_rules != policy) {
+ struct list_head *prev_rules = ima_rules;
+ struct list_head *first = ima_rules->next;
ima_policy_flag = 0;
+
+ /*
+ * Make the previous list can traverse to new list,
+ * that is tricky, or there is a deadly loop whithin
+ * "list_for_each_entry_rcu(entry, ima_rules, list)"
+ *
+ * After update "ima_rules", restore the previous list.
+ */
I think this could be rephrased to be a tad clearer, I am not quite sure
how I must interpret the first sentence of the comment.
I got it, how about this:
/*
* The previous list has to traverse to new list,
* Or there may be a deadly loop within
* "list_for_each_entry_rcu(entry, ima_rules, list)"
*
* That is tricky, after updated "ima_rules", restore the previous list.
*/
+ prev_rules->next = policy->next;
ima_rules = policy;
+ syncchronize_rcu();
I'm a bit puzzled as you seem to imply in the mail this patch was tested,
but there is no 'syncchronize_rcu' (with two 'c') symbol in the kernel.
Was that a copy/paste error? Or maybe you forgot the 'not' in "This
patch has been tested"? These errors happen, and I am myself quite an
expert in doing them :)
Sorry for the mistake, I copy/paste the patch and delete/edit some lines,
have reviewed before sending, but not found. I have made a case to reproduce
the error, dumping "ima_rules" and every item address of list in the error
situaiton, I can watchthe "ima_rules" change, old list traversing to the
new list.
And I have been doing a reboot test which found this bug. This patch
seems to work fine.
+ prev_rules->next = first;
The side effect is the "ima_default_rules" will be changed a little while.
But it make sense, the process should be checked again by the new policy.
This patch has been tested, if will do, I can resubmit this patch.>
How about this ?
Correct me if I'm wrong, here is how I think I understand you patch.
We start with a situation like that (step 0):
ima_rules --> List entry 0 (head node) = ima_default_rules <-> List entry 1 <-> List entry 2 <-> ... <-> List entry 0
Then we decide to update the policy for the first time, so
'ima_rules [&ima_default_rules] != policy [&ima_policy_rules]'.
We enter the condition.
First we copy the current value of ima_rules (&ima_default_rules)
to a temporary variable 'prev_rules'. We also create a pointer dubbed
'first' to the entry 1 in the default list (step 1):
prev_rules -------------
\/
ima_rules --> List entry 0 (head node) = ima_default_rules <-> List entry 1 <-> List entry 2 <-> ... <-> List entry 0
/\
first --------------------------------------------------------------
Then we update prev_rules->next to point to policy->next (step 2):
List entry 1 <-> List entry 2 <-> ... -> List entry 0
/\
first
(notice that list entry 0 no longer points backwards to 'list entry 1',
but I don't think there is any reverse iteration in IMA, so it should be
safe)
prev_rules -------------
\/
ima_rules --> List entry 0 (head node) = ima_default_rules
|
|
-------------------------------------------
\/
policy --> policy entry 0' (head node) = ima_policy_rules <-> policy entry 1' <-> policy entry 2' <-> .... <-> policy entry 0'
We then update ima_rules to point to ima_policy_rules (step 3):
List entry 1 <-> List entry 2 <-> ... -> List entry 0
/\
first
prev_rules -------------
\/
ima_rules List entry 0 (head node) = ima_default_rules
| |
| |
| ------------------------------------------
--------------- |
\/ \/
policy --> policy entry 0' (head node) = ima_policy_rules <-> policy entry 1' <-> policy entry 2' <-> .... <-> policy entry 0'
/\
first --------------------------------------------------------------
Then we run synchronize_rcu() to wait for any RCU reader to exit their loops (step 4).
Finally we update prev_rules->next to point back to the ima policy and fix the loop (step 5):
List entry 1 <-> List entry 2 <-> ... -> List entry 0
/\
first
prev_rules ---> List entry 0 (head node) = ima_default_rules <-> List entry 1 <-> List entry 2 <-> ... <-> List entry 0
/\
first (now useless)
ima_rules
|
|
|
---------------
\/
policy --> policy entry 0' (head node) = ima_policy_rules <-> policy entry 1' <-> policy entry 2' <-> .... <-> policy entry 0'
The goal is that readers should still be able to loop
(forward, as we saw that backward looping is temporarily broken)
while in steps 0-4.
Yes, It's the workflow.
I'm not completely sure what would happen to a client that started iterating
over ima_rules right after step 2.
Wouldn't they be able to start looping through the new policy
as 'List entry 0 (head node) = ima_default_rules' points to ima_policy_rules?
And if they, wouldn't they loop until the write to 'ima_rule' at step 3 (admittedly
very shortly thereafter) completed?
And would the compiler be allowed to optimize the read to 'ima_rules' in the
list_for_each_entry() loop, thereby never reloading the new value for
'ima_rules', and thus looping forever, just what we are trying to avoid?
Yes, "ima_rules" cache not update in time, It's a risk. I am not sure
if "WRITE_ONCE"
can do this trick. How about:
WRITE_ONCE(prev_rules->next, policy->next);
WRITE_ONCE(ima_rules, policy);
If can't fix the cache issue, maybe the "ima_rules_tmp" solution is the
best way.
I will test it.
Overall, I'm tempted to say this is perhaps a bit too complex (at least,
my head tells me it is, but that may very well be because I'm terrible
at concurrency issues).
Honestly, in this case I think awaiting input from more experienced
kernel devs than I is the best path forward :-)
----------
Regards,
liqiong
Thanks,
Simon
