hgst / libnvme Goto Github PK

When the request is larger than the maximum IO size OR striping is enabled, _nvme_ns_split_request will be used to break the request up into multiple ones: https://github.com/hgst/libnvme/blob/master/lib/nvme/nvme_ns.c#L231

However, it just passes the cb_fn to each of these. So any consumer of this API will have the cb_fn called more than once, and it cannot be reasonably expected to keep track of how many sub-requests there are and when the last one has completed.

on one nvme 4k rand write, queue=128
using spdk get about 360,000 iops;
while using libnvme got about 71,000 iops.

for read testing has no this problem.