[PATCH v3 3/3] nvme-rdma: Handle number of queue changes

[Chao Leng]

On 2022/8/30 16:21, Sagi Grimberg wrote:
> 
>>> On Mon, Aug 29, 2022 at 05:22:50PM +0800, Chao Leng wrote:
>>>>> +    ret = nvme_rdma_start_io_queues(ctrl, nr_queues,
>>>>> +                    ctrl->tag_set.nr_hw_queues + 1);
>>>>> +    if (ret)
>>>>> +        goto out_wait_freeze_timed_out;
>>>> nvme_rdma_start_io_queues(ctrl, 1, nr_queues) need to clean.
>>>
>>> the teardown path in nvme_rdma_stop_io_queus() is using
>>>
>>>     test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags)
>>>
>>> when iterating over all nr_hw_queues.
>> I mean that we need realloc the ctrl->queues for
>> nvme_rdma_start_io_queues(ctrl, 1, nr_queues).
>> Otherwise, unallocated memory will be accessed when
>> the new queue_count is bigger than the old one.
> 
> That can't happen. ctrl->queues is allocated at the start
> of the controller lifetime and will never exceed this queue
> count.
I don't understand why it can't happen.
nvme_rdma_start_io_queues(ctrl, nr_queues, ctrl->tag_set.nr_hw_queues + 1)
is designed for dealing with the scenario that the new queue_count is
bigger than the old one. it will access the unallocated memory.
ctrl->queues is just allocated the first queue count buffers at the start
of the controller lifetime, if the new queue count is bigger than the
first queue count, ctrl-queues will be insufficient.
> .