[dm-devel] [PATCH 4/7] dm mpath: remove process_queued_ios()

[Junichi Nomura]

On 02/04/14 17:18, Hannes Reinecke wrote:
>>> @@ -1591,8 +1563,17 @@ static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
>> ...
>>> +		if (m->current_pg && m->pg_init_required)
>>> +			__pg_init_all_paths(m, 0);
>>> +		spin_unlock_irqrestore(&m->lock, flags);
>>> +		dm_table_run_md_queue_async(m->ti->table);
>>> +	}
>>
>> What happens if "!m->current_pg && m->pg_init_required"?
>>
>>From the current logic it means that no valid pg was found, so
> calling pg_init would be pointless.
> We're calling __choose_pgpath() before that, so if that returns
> with current_pg == NULL all paths are down, and calling
> pg_init would be pointless.
> 
> But I think I see to have pg_init_required handling cleared up;
> I'll be doing a patch to unset it at the start of __choose_pgpath(),
> this we we can be sure that it'll be set correctly.

I think it is possible that __choose_pgpath() being called twice
before pg_init_required is checked.

For example,

  multipath_ioctl()
    __choose_pgpath()
      clear pg_init_required
      select a new pg
      __switch_pg()
        set pg_init_required

  map_io()
    __choose_pgpath()
      clear pg_init_required
      select the same pg
      (pg_init_required is not set)
      ...

In this case, pg_init should be submitted to the pg but not.

-- 
Jun'ichi Nomura, NEC Corporation