On 23/10/2024 8:19, Christoph Hellwig wrote:
On Tue, Oct 22, 2024 at 04:23:29PM +0300, Sagi Grimberg wrote:
I agree. NVMe may have a unique need here, but it needs a tag from a
specific hctx while the context requesting it does not map according to
the hctx cpumap. It cannot use any other tag from any other hctx.
The reason is that the connect for a queue must be done from a tag that
belongs to the queue because nvme relies on it when it does resolution back
to
the request to the completion.
I wonder if we should byte the bullet and not use a request for the
connect commands. We've already special cased the AEN command because
it was causing too many problems, and given all the pain so far connect
might also have hit that treshold.
That would be trading one set of problems for another. We had some hard
times in the past to correctly fence AERs against controller states
because it
is not a normal block request.
Plus, unlike AERs, connect has a timeout, which we'll need to take care
of...
Not that it is not possible, I'd just like to avoid this option.
