mirror of
https://github.com/torvalds/linux
synced 2024-09-20 19:17:24 +00:00
a61a459f58
Testing a network device that has large numbers of bytes/packets may
overflow. Using stats64 when comparing fixes this problem.
I tripped on this while iterating on a qstats patch for mlx5. See below
for confirmation without my added code that this is a bug.
Before this patch (with added debugging output):
$ NETIF=eth0 tools/testing/selftests/drivers/net/stats.py
KTAP version 1
1..4
ok 1 stats.check_pause
ok 2 stats.check_fec
rstat: 481708634 qstat: 666201639514 key: tx-bytes
not ok 3 stats.pkt_byte_sum
ok 4 stats.qstat_by_ifindex
Note the huge delta above ^^^ in the rtnl vs qstats.
After this patch:
$ NETIF=eth0 tools/testing/selftests/drivers/net/stats.py
KTAP version 1
1..4
ok 1 stats.check_pause
ok 2 stats.check_fec
ok 3 stats.pkt_byte_sum
ok 4 stats.qstat_by_ifindex
It looks like rtnl_fill_stats in net/core/rtnetlink.c will attempt to
copy the 64bit stats into a 32bit structure which is probably why this
behavior is occurring.
To show this is happening, you can get the underlying stats that the
stats.py test uses like this:
$ ./cli.py --spec ../../../Documentation/netlink/specs/rt_link.yaml \
--do getlink --json '{"ifi-index": 7}'
And examine the output (heavily snipped to show relevant fields):
'stats': {
'multicast': 3739197,
'rx-bytes': 1201525399,
'rx-packets': 56807158,
'tx-bytes': 492404458,
'tx-packets': 1200285371,
'stats64': {
'multicast': 3739197,
'rx-bytes': 35561263767,
'rx-packets': 56807158,
'tx-bytes': 666212335338,
'tx-packets': 1200285371,
The stats.py test prior to this patch was using the 'stats' structure
above, which matches the failure output on my system.
Comparing side by side, rx-bytes and tx-bytes, and getting ethtool -S
output:
rx-bytes stats: 1201525399
rx-bytes stats64: 35561263767
rx-bytes ethtool: 36203402638
tx-bytes stats: 492404458
tx-bytes stats64: 666212335338
tx-bytes ethtool: 666215360113
Note that the above was taken from a system with an mlx5 NIC, which only
exposes ndo_get_stats64.
Based on the ethtool output and qstat output, it appears that stats.py
should be updated to use the 'stats64' structure for accurate
comparisons when packet/byte counters get very large.
To confirm that this was not related to the qstats code I was iterating
on, I booted a kernel without my driver changes and re-ran the test
which shows the qstats are skipped (as they don't exist for mlx5):
NETIF=eth0 tools/testing/selftests/drivers/net/stats.py
KTAP version 1
1..4
ok 1 stats.check_pause
ok 2 stats.check_fec
ok 3 stats.pkt_byte_sum # SKIP qstats not supported by the device
ok 4 stats.qstat_by_ifindex # SKIP No ifindex supports qstats
But, fetching the stats using the CLI
$ ./cli.py --spec ../../../Documentation/netlink/specs/rt_link.yaml \
--do getlink --json '{"ifi-index": 7}'
Shows the same issue (heavily snipped for relevant fields only):
'stats': {
'multicast': 105489,
'rx-bytes': 530879526,
'rx-packets': 751415,
'tx-bytes': 2510191396,
'tx-packets': 27700323,
'stats64': {
'multicast': 105489,
'rx-bytes': 530879526,
'rx-packets': 751415,
'tx-bytes': 15395093284,
'tx-packets': 27700323,
Comparing side by side with ethtool -S on the unmodified mlx5 driver:
tx-bytes stats: 2510191396
tx-bytes stats64: 15395093284
tx-bytes ethtool: 17718435810
Fixes:
|
||
|---|---|---|
| .. | ||
| accounting | ||
| arch | ||
| bootconfig | ||
| bpf | ||
| build | ||
| certs | ||
| cgroup | ||
| counter | ||
| crypto | ||
| debugging | ||
| firewire | ||
| firmware | ||
| gpio | ||
| hv | ||
| iio | ||
| include | ||
| kvm/kvm_stat | ||
| laptop | ||
| leds | ||
| lib | ||
| memory-model | ||
| mm | ||
| net/ynl | ||
| objtool | ||
| pci | ||
| pcmcia | ||
| perf | ||
| power | ||
| rcu | ||
| scripts | ||
| sound | ||
| spi | ||
| testing | ||
| thermal | ||
| time | ||
| tracing | ||
| usb | ||
| verification | ||
| virtio | ||
| wmi | ||
| workqueue | ||
| Makefile | ||