On 1/30/20 11:34 PM, Christoph Hellwig wrote:

Can you explain this check?  This looks weird to me  I think we need
a generalized check if a make_request based driver supports REQ_NOWAIT
instead (and as a separate patch / patchset).

Original check used to reject polled IO for stackable block devices as 
"not supported". To solve that situation I introduced additional check 
to reject all non REQ_HIPRI requests. That check is not intended to 
generalize, like you indicated, but to conservatively select which 
requests to accept.
Perhaps there is better way to do that. Any suggestions?