> Sure this works? It does. > nvme_req_needs_retry() checks blk_noretry_request(): > The original idea here was to _set_ these bits on multipath path devices > so that they won't attempt any retry, but rather forward the I/O error > to the multipath device itself for failover. > So if these bits are set (as they should be for multipathed devices) > we'll never attempt any failover... While that might have been the "original" idea, it isn't what this code does. We never set any of the REQ_FAILFAST_* bits in nvme_ns_head_make_request. In NVMe there aren't really any device equivalents of REQ_FAILFAST_ that make sense for multipath. The only one that we map to is the limited retry bit, and that is media centric, so a failver would not help.