Incorrrect latency values reported when using --limitread --direct and --iodepth >=2

[johnpkellner-star]

Here is a command to replicate the result:

elbencho --hosts node1,node2,node3,node4,node5 --port 2611 --infloop --timelimit 300 --nosvcshare --log 2 --cpu --lat --latpercent --liveint 1000 --livecsvex --livecsv 4K.rr.csv --resfile 4K.rr.txt --direct --dropcache --sync --iodepth 2 --rand --randalgo fast -r --limitread 4194304 -b 4K -t 2 /dev/sdb

The latency value recorded in 0us.

If I remove --direct OR change iodepth 1, I begin getting a valid latency value again.

This behavior seems to require all 3 to cause invalid latency values...

1. --limitrread
2. --iodepth >=2
3. --direct