On 10/15/17 9:59 AM, Ido Schimmel wrote:
> On Sun, Oct 15, 2017 at 09:24:07AM -0600, David Ahern wrote:
>> On 10/15/17 1:50 AM, Ido Schimmel wrote:
>>> On Fri, Oct 13, 2017 at 04:02:09PM -0700, David Ahern wrote:
>>>> ipv6_add_addr is called in process context with rtnl lock held
>>>> (e.g., manual config of an address) or during softirq processing
>>>> (e.g., autoconf and address from a router advertisement).
>>>>
>>>> Currently, ipv6_add_addr calls rcu_read_lock_bh shortly after entry
>>>> and does not call unlock until exit, minus the call around the address
>>>> validator notifier. Similarly, addrconf_hash_lock is taken after the
>>>> validator notifier and held until exit. This forces the allocation of
>>>> inet6_ifaddr to always be atomic.
>>>>
>>>> Refactor ipv6_add_addr as follows:
>>>> 1. add an input boolean to discriminate the call path (process context
>>>> or softirq). This new flag controls whether the alloc can be done
>>>> with GFP_KERNEL or GFP_ATOMIC.
>>>>
>>>> 2. Move the rcu_read_lock_bh and unlock calls only around functions that
>>>> do rcu updates.
>>>>
>>>> 3. Remove the in6_dev_hold and put added by 3ad7d2468f79f ("Ipvlan should
>>>> return an error when an address is already in use."). This was done
>>>> presumably because rcu_read_unlock_bh needs to be called before calling
>>>> the validator. Since rcu_read_lock is not needed before the validator
>>>> runs revert the hold and put added by 3ad7d2468f79f and only do the
>>>> hold when setting ifp->idev.
>>>>
>>>> 4. move duplicate address check and insertion of new address in the global
>>>> address hash into a helper. The helper is called after an ifa is
>>>> allocated and filled in.
>>>>
>>>> This allows the ifa for manually configured addresses to be done with
>>>> GFP_KERNEL and reduces the overall amount of time with rcu_read_lock held
>>>> and hash table spinlock held.
>>>>
>>>> Signed-off-by: David Ahern <dsah...@gmail.com>
>>>
>>> [...]
>>>
>>>> @@ -1073,21 +1085,19 @@ ipv6_add_addr(struct inet6_dev *idev, const struct
>>>> in6_addr *addr,
>>>>
>>>> in6_ifa_hold(ifa);
>>>> write_unlock(&idev->lock);
>>>> -out2:
>>>> +
>>>> rcu_read_unlock_bh();
>>>>
>>>> - if (likely(err == 0))
>>>> - inet6addr_notifier_call_chain(NETDEV_UP, ifa);
>>>> - else {
>>>> + inet6addr_notifier_call_chain(NETDEV_UP, ifa);
>>>> +out:
>>>> + if (unlikely(err < 0)) {
>>>> + if (rt)
>>>> + ip6_rt_put(rt);
>>>
>>> I believe 'rt' needs to be set to NULL after addrconf_dst_alloc()
>>> fails.
>>
>> The above frees rt and the line below frees the ifa and resets the value
>> to an error, so after the line above rt is no longer referenced.
>
> Earlier in the code we have:
>
> rt = addrconf_dst_alloc(idev, addr, false);
> if (IS_ERR(rt)) {
> err = PTR_ERR(rt);
> goto out;
> }
>
> So we end up calling ip6_rt_put() with an error value. I believe it
> should be:
>
> rt = addrconf_dst_alloc(idev, addr, false);
> if (IS_ERR(rt)) {
> err = PTR_ERR(rt);
> rt = NULL;
> goto out;
> }
>
gotcha. Will fix.
