[PATCH 1/5] blk-mq: move srcu from blk_mq_hw_ctx to request_queue

[Sagi Grimberg]

>>>>> +	bool			alloc_srcu;
>>>>
>>>> I found the following statement multiple times in this patch:
>>>>
>>>> WARN_ON_ONCE(q->alloc_srcu != !!(q->tag_set->flags & BLK_MQ_F_BLOCKING));
>>>>
>>>> Does this mean that the new q->alloc_srcu member variable can be left out
>>>> and that it can be replaced with the following test?
>>>>
>>>> q->tag_set->flags & BLK_MQ_F_BLOCKING
>>>
>>> q->tag_set can't be used anymore after blk_cleanup_queue() returns,
>>> and we need the flag for freeing request_queue instance.
>>
>> Why not just look at the queue->srcu pointer? it is allocated only
>> for BLK_MQ_F_BLOCKING no?
> 
> Yeah, we can add one extra srcu pointer to request queue, but this way
> needs one extra fetch to q->srcu in fast path compared with current
> code base, so io_uring workload may be affected a bit.

Yea you're right. We should at some point make has_srcu and
mq_sysfs_init_done bits in a flags member, but we have like 6 more
before we need to do it...