Re: [PATCH v2] nvme-multipath: Early exit if no path is available

[Chao Leng <lengchao@huawei.com>]

On 2021/2/1 18:45, Hannes Reinecke wrote:
> On 2/1/21 10:40 AM, Chao Leng wrote:
>>
>>
>> On 2021/2/1 16:57, Hannes Reinecke wrote:
>>> On 2/1/21 9:47 AM, Chao Leng wrote:
>>>>
>>>>
>>>> On 2021/2/1 15:29, Hannes Reinecke wrote:[ .. ]
>>>>> Urgh. Please, no. That is well impossible to debug.
>>>>> Can you please open-code it to demonstrate where the difference to the current (and my fixed) versions is?
>>>>> I'm still not clear where the problem is once we applied both patches.
>>>> For example assume the list has three path, and all path is not NVME_ANA_OPTIMIZED:
>>>> head->next = ns1;
>>>> ns1->next = ns2;
>>>> ns2->next = head;
>>>> old->next = ns2;
>>>>
>>> And this is where I have issues with.
>>> Where does 'old' come from?
>>> Clearly it was part of the list at one point; so what happened to it?
>> I explained this earlier.
>> In nvme_ns_remove, there is a hole between list_del_rcu and
>> nvme_mpath_clear_current_path. If head->current_path is the "old", and
>> the "old" is removing. The "old" is already removed from the list by
>> list_del_rcu, but head->current_path is not clear to NULL by
>> nvme_mpath_clear_current_path.
>> Find path is race with nvme_ns_remove, use the "old" pass to
>> nvme_round_robin_path to find path.
> 
> Ah. So this should be better:
> 
> @@ -202,10 +202,12 @@ static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
>   static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
>                  struct nvme_ns *ns)
>   {
> -       ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
> -                       siblings);
> -       if (ns)
> -               return ns;
> +       if (ns && !test_bit(NVME_NS_REMOVING, &ns->flags)) {
> +               ns = list_next_or_null_rcu(&head->list, &ns->siblings,
> +                                          struct nvme_ns, siblings);
> +               if (ns)
> +                       return ns;
> +       }
>          return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
>   }
> 
> The 'NVME_NS_REMOVING' bit is set before list_del_rcu() is called, so it should guard against the issue you mentioned.
Looks useless, it is still infinite loop. You can check the workflow for the scenario I mentioned.
> 
> Cheers,
> 
> Hannes