Bug 198007

(Hint: I found no good bugzilla component for either Fibre Channel or Statistics)
I'm graphing the deltas of some Fibre Channel host statistics like /sys/class/fc_host/host1/statistics/rx_words.
Unfortunately there are odd peaks in the graphs that look like overflows, even though my program (an NRPE plugin for Nagios/Centreon) is expected to handle those. The driver is "Emulex LightPulse Fibre Channel SCSI driver 10.4.8000.0."
To be concrete, the peaks are roughly like this:

rx_frames: 10M
rx_words: 2.4G
tx_frames: 9.4M
tx_words: 2.3G

So the "words" overflows look like 31 bit, but they are a bit too high, and the "frames" overflows look just odd.

I did a little digging in the sources to find that lpfc_hw.h uses 32bit counters (it seems) like this:

        uint32_t xmitByteCnt;
        uint32_t rcvByteCnt;
        uint32_t xmitFrameCnt;
        uint32_t rcvFrameCnt;

But the FC host statistics seem to use 64 bits:
scsi_transport_fc.h:
/* FC Statistics - Following FC HBAAPI v2.0 guidelines */
struct fc_host_statistics {
        /* port statistics */
        u64 seconds_since_last_reset;
        u64 tx_frames;
        u64 tx_words;
        u64 rx_frames;
        u64 rx_words;
        u64 lip_count;
        u64 nos_count;
...

And it seems the 32 bit are assigned to 64 bit:
/usr/src/linux/drivers/scsi/lpfc/lpfc_attr.c:   hs->tx_frames = pmb->un.varRdStatus.xmitFrameCnt;

Unfortunately the counter seems neither 32, nor 64 bit, because I have this reading:
# cat /sys/class/fc_host/host1/statistics/rx_words
0x715d9dd800

I think there is a least one bug somewhere ;-)

The HBA is a HP-branded EMulex Saturn-X (AJ762B) connected to an 8Gbps FC-switch.