[PATCH 0/2] nvme-fabrics: short-circuit connect retries

[Chao Leng]

On 2021/6/24 16:10, Hannes Reinecke wrote:
> On 6/24/21 9:29 AM, Chao Leng wrote:
>>
>>
>> On 2021/6/24 13:51, Hannes Reinecke wrote:
>>> On 6/23/21 11:38 PM, Sagi Grimberg wrote:
>>>>
>>>>> Hi all,
>>>>>
>>>>> commit f25f8ef70ce2 ("nvme-fc: short-circuit reconnect retries")
>>>>> allowed the fc transport to honour the DNR bit during reconnect
>>>>> retries, allowing to speed up error recovery.
>>>>
>>>> How does this speed up error recovery?
>>>
>>> Well, not exactly error recovery (as there is nothing to recover).
>>> But we won't attempt pointless retries, thereby reducing the noise in
>>> the message log.
>> This conflict with the tcp and rdma target.
>> You may need to delete the improper NVME_SC_DNR at the target.
>> However, this will cause compatibility issues between different versions.
> 
> Which ones?
In many scenarios, the destination sets DNR for abnormal packets,
but each new connection may not have the same error.
> I checked the DNR usage in the target code, and they seem to set it
> correctly (ie the result would not change when the command is retried).
> With the possible exception of ENOSPC handling, as this is arguably
> dynamic and might change with a retry.
The DNR status of the old connection may not be relevant to the re-established connection.
> 
> Cheers,
> 
> Hannes
>