[PATCH] nvme-rdma: Don't fail the controller if only part of the queues fail to connect

[hch@lst.de (Christoph Hellwig)]

On Thu, Nov 08, 2018@10:20:00AM +0200, Max Gurtovoy wrote:
>
> On 11/7/2018 11:07 AM, Christoph Hellwig wrote:
>> On Tue, Nov 06, 2018@01:10:27PM +0200, Max Gurtovoy wrote:
>>>> This sounds odd.?? Why aren't you concerned that io queues are not
>>>> connecting ?? Are there any log messages hinting at the failures ? any
>>>> way someone looking at the controller knows how many queues were actually
>>>> created ? ? I would assume any failure is significant and should be
>>>> visible, and it's worthwhile knowing whether this is a consistent failure
>>>> or a random failure. and what the failure was.
>>> This may happen (well it happened in the past, and fixed in the block
>>> layer) in case there are offline cpu's or some other reason that some queue
>>> is unmapped.
>>>
>>> I prefer not to relay on the block layer to ensure us 100% mapping and
>>> prefer be safe in our ULP.
>> How do we ensure we ensure any potential new block layer bug returns
>> -EXDEV so that your handling kicks in?
>
> well we can't ensure that. Are you suggesting to do the handling for each 
> failure ?
>
> I guess we'll need this anyway for older kernels that don't have the fix 
> for offline cpu mapping.

No, I'm arguing that adding this just in case code which doesn't
have a good way to actually catch a typical bug is not very useful.