Spurious network metric test failure in qa/635

[Nathan Scott <nathans@xxxxxxxxxx>]

Hi Ken,

Out of the blue I had a reproducible failure in qa/635 today.
This looks like one of yours (kenj-pc reference in there :) &
it uses paste(1) - gotta be you ;) ... it's a netstat vs pcp
network.* metric comparison script.

$ diff 635.out 635.out.bad
1a2
> network.interface.out.drops["vnet0"] value 0 not in range 1..1

The test script contains this comment:
# Note: drops metric is really ?-DRP + ?-OVR ... visible in netstat, but
#       not /proc/net/dev that the Linux PMDA uses

The failure is to do with the TX-OVR field in netstat output.
The script sums TX-OVR and TX-DRP at line 95 for comparison to
network.interface.out.drops.  However, its not clear what OVR
is in this context - it *seems* to map to the "fifo" column in
/proc/net/dev ... ?  (no mention on netstat man page, googling
finds some articles suggesting it means "couldn't send or recv"
or similarly generic words).  -- i.e. network.interface.*.fifo

I've attached "netstat -ai" and /proc/net/dev contents and can
confirm that network.interface.out.fifo does indeed return 1
for this interface.

Do you know what the right behaviour for the test here is?  It
looks like it should be treating drops and fifo distinctly (and
comparing the OVR columns to the fifo metrics), maybe?

thanks!

--
Nathan

netstat_ai.txt
Description: Text document

proc_net_dev.txt
Description: Text document