| Message ID | 309B89C4C689E141A5FF6A0C5FB2118B8C5EF4F7@ORSMSX101.amr.corp.intel.com |
|---|---|
| State | Changes Requested |
| Headers | show |
On Tue, 2017-04-25 at 07:10 +0000, Brown, Aaron F wrote: > > From: Intel-wired-lan [mailto:intel-wired-lan-bounces@lists.osuosl. > > org] On > > Behalf Of Benjamin Poirier > > Sent: Monday, April 24, 2017 12:10 PM > > To: Neftin, Sasha <sasha.neftin@intel.com> > > Cc: Kirsher@f1.synalogic.ca; Stefan Priebe <s.priebe@profihost.ag>; > > netdev@vger.kernel.org; intel-wired-lan@lists.osuosl.org > > Subject: Re: [Intel-wired-lan] [PATCH 1/2] e1000e: Don't return > > uninitialized > > stats > > > > Sasha, please use reply-all to keep everyone in cc (including > > me...). > > > > On 2017/04/24 11:17, Neftin, Sasha wrote: > > > On 4/23/2017 15:53, Neftin, Sasha wrote: > > > > -----Original Message----- > > > > From: Intel-wired-lan [mailto:intel-wired-lan-bounces@lists.osu > > > > osl.org] > > > > On Behalf Of Benjamin Poirier > > > > Sent: Saturday, April 22, 2017 00:20 > > > > To: Kirsher, Jeffrey T <jeffrey.t.kirsher@intel.com> > > > > Cc: netdev@vger.kernel.org; intel-wired-lan@lists.osuosl.org; > > > > Stefan > > > > Priebe <s.priebe@profihost.ag> > > > > Subject: [Intel-wired-lan] [PATCH 1/2] e1000e: Don't return > > > > uninitialized > > > > stats > > > > > > > > Some statistics passed to ethtool are garbage because > > > > e1000e_get_stats64() doesn't write them, for example: > > tx_heartbeat_errors. > > This leaks kernel memory to userspace and confuses users. > > > > > > > > Do like ixgbe and use dev_get_stats() which first zeroes out > > > > rtnl_link_stats64. > > > > > > > > Reported-by: Stefan Priebe <s.priebe@profihost.ag> > > > > Signed-off-by: Benjamin Poirier <bpoirier@suse.com> > > > > --- > > > > drivers/net/ethernet/intel/e1000e/ethtool.c | 2 +- > > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > > > diff --git a/drivers/net/ethernet/intel/e1000e/ethtool.c > > > > b/drivers/net/ethernet/intel/e1000e/ethtool.c > > > > index 7aff68a4a4df..f117b90cdc2f 100644 > > > > --- a/drivers/net/ethernet/intel/e1000e/ethtool.c > > > > +++ b/drivers/net/ethernet/intel/e1000e/ethtool.c > > > > @@ -2063,7 +2063,7 @@ static void > > > > e1000_get_ethtool_stats(struct > > > > net_device *netdev, > > > > pm_runtime_get_sync(netdev->dev.parent); > > > > - e1000e_get_stats64(netdev, &net_stats); > > > > + dev_get_stats(netdev, &net_stats); > > > > pm_runtime_put_sync(netdev->dev.parent); > > > > -- > > > > 2.12.2 > > > > > > > > _______________________________________________ > > > > Intel-wired-lan mailing list > > > > Intel-wired-lan@lists.osuosl.org > > > > http://lists.osuosl.org/mailman/listinfo/intel-wired-lan > > > > > > Hello, > > > > > > We would like to not accept this patch. Suggested generic method > > > '*dev_get_stats' (net/core/dev.c) calls 'ops->ndo_get_stats64' > > > method > > > > which > > > eventually calls e1000e_get_stats64 (netdev.c) - so there is same > > > functionality. Also, see that 'e1000e_get_stats64' method in > > > netdev.c (line > > > > No, it's not the same functionality because dev_get_stats() does a > > memset on the rtnl_link_stats64 struct. > > > > > 5928) calls 'memset' with 0's before update statistics. Local > > > sanity check > > > > I don't see any memset in e1000e_get_stats64(). What kernel version > > are > > you looking at? > > The call to memset was removed from the upstream kernel with: > ------------------------------------------------------------------- > ----------------- > commit 5944701df90d9577658e2354cc27c4ceaeca30fe > Author: stephen hemminger <stephen@networkplumber.org> > Date: Fri Jan 6 19:12:53 2017 -0800 > > net: remove useless memset's in drivers get_stats64 > > In dev_get_stats() the statistic structure storage has already > been > zeroed. Therefore network drivers do not need to call memset() > again. > ... > < changes to other drivers snipped out > > ... > diff --git a/drivers/net/ethernet/intel/e1000e/netdev.c > b/drivers/net/ethernet/int > index 723025b..79651eb 100644 > --- a/drivers/net/ethernet/intel/e1000e/netdev.c > +++ b/drivers/net/ethernet/intel/e1000e/netdev.c > @@ -5925,7 +5925,6 @@ void e1000e_get_stats64(struct net_device > *netdev, > { > struct e1000_adapter *adapter = netdev_priv(netdev); > > - memset(stats, 0, sizeof(struct rtnl_link_stats64)); > spin_lock(&adapter->stats64_lock); > e1000e_update_stats(adapter); > /* Fill out the OS statistics structure */ > ------------------------------------------------------------------- > ----------------- > > This also is where the bad counters start to show up for e1000e for > my test systems. From this driver on I see (very) large values for > tx_dropped, rx_over_errors and tx_fifo_errors on driver load (even > before bringing the interface up. It seems the memset is not so > useless for this driver after all. Would simply reverting the e1000e > portion of this patch resolve the issue? Looks like Aaron beat me to the punch on pointing out that we had this very code in there before. It appears that Stephen's assertion/assumption was incorrect about the stats structure being zero'd out, which is why we are seeing the issue. I have no issue reverting Stephen's earlier patch, or do we want to pursue why the stats structure is not zero'd out and resolve that instead. Either way, just want to make sure we are all on the same page as to the right solution so that we do not end up repeating this in the future.
On 2017/04/25 02:07, Jeff Kirsher wrote: [...] > > > > > > I don't see any memset in e1000e_get_stats64(). What kernel version > > > are > > > you looking at? > > > > The call to memset was removed from the upstream kernel with: > > ------------------------------------------------------------------- > > ----------------- > > commit 5944701df90d9577658e2354cc27c4ceaeca30fe > > Author: stephen hemminger <stephen@networkplumber.org> > > Date: Fri Jan 6 19:12:53 2017 -0800 > > > > net: remove useless memset's in drivers get_stats64 > > > > In dev_get_stats() the statistic structure storage has already > > been > > zeroed. Therefore network drivers do not need to call memset() > > again. > > ... > > < changes to other drivers snipped out > > > ... > > diff --git a/drivers/net/ethernet/intel/e1000e/netdev.c > > b/drivers/net/ethernet/int > > index 723025b..79651eb 100644 > > --- a/drivers/net/ethernet/intel/e1000e/netdev.c > > +++ b/drivers/net/ethernet/intel/e1000e/netdev.c > > @@ -5925,7 +5925,6 @@ void e1000e_get_stats64(struct net_device > > *netdev, > > { > > struct e1000_adapter *adapter = netdev_priv(netdev); > > > > - memset(stats, 0, sizeof(struct rtnl_link_stats64)); > > spin_lock(&adapter->stats64_lock); > > e1000e_update_stats(adapter); > > /* Fill out the OS statistics structure */ > > ------------------------------------------------------------------- > > ----------------- > > > > This also is where the bad counters start to show up for e1000e for > > my test systems. From this driver on I see (very) large values for > > tx_dropped, rx_over_errors and tx_fifo_errors on driver load (even > > before bringing the interface up. It seems the memset is not so > > useless for this driver after all. Would simply reverting the e1000e > > portion of this patch resolve the issue? > > Looks like Aaron beat me to the punch on pointing out that we had this > very code in there before. It appears that Stephen's > assertion/assumption was incorrect about the stats structure being > zero'd out, which is why we are seeing the issue. > > I have no issue reverting Stephen's earlier patch, or do we want to > pursue why the stats structure is not zero'd out and resolve that > instead. Either way, just want to make sure we are all on the same > page as to the right solution so that we do not end up repeating this > in the future. If you revert the e1000e part of 5944701df90d ("net: remove useless memset's in drivers get_stats64", v4.11-rc1) it will fix the issue with ethtool but memset will be done twice for code paths that call dev_get_stats() (sysfs, rtnl, ...). Not a big deal but this is not a problem in the approach I initially suggested. Alternatively, we could put a memset in e1000_get_ethtool_stats().
On 2017/04/25 10:54, Stephen Hemminger wrote: [...] > > > The call to memset was removed from the upstream kernel with: > > > ------------------------------------------------------------------- > > > ----------------- > > > commit 5944701df90d9577658e2354cc27c4ceaeca30fe > > > Author: stephen hemminger <stephen@networkplumber.org> > > > Date: Fri Jan 6 19:12:53 2017 -0800 > > > > > > net: remove useless memset's in drivers get_stats64 > > > > > > In dev_get_stats() the statistic structure storage has already > > > been > > > zeroed. Therefore network drivers do not need to call memset() > > > again. > > > ... > > > < changes to other drivers snipped out > > > > ... > > > diff --git a/drivers/net/ethernet/intel/e1000e/netdev.c > > > b/drivers/net/ethernet/int > > > index 723025b..79651eb 100644 > > > --- a/drivers/net/ethernet/intel/e1000e/netdev.c > > > +++ b/drivers/net/ethernet/intel/e1000e/netdev.c > > > @@ -5925,7 +5925,6 @@ void e1000e_get_stats64(struct net_device > > > *netdev, > > > { > > > struct e1000_adapter *adapter = netdev_priv(netdev); > > > > > > - memset(stats, 0, sizeof(struct rtnl_link_stats64)); > > > spin_lock(&adapter->stats64_lock); > > > e1000e_update_stats(adapter); > > > /* Fill out the OS statistics structure */ > > > ------------------------------------------------------------------- > > > ----------------- > > > > > > This also is where the bad counters start to show up for e1000e for > > > my test systems. From this driver on I see (very) large values for > > > tx_dropped, rx_over_errors and tx_fifo_errors on driver load (even > > > before bringing the interface up. It seems the memset is not so > > > useless for this driver after all. Would simply reverting the e1000e > > > portion of this patch resolve the issue? > > > > Looks like Aaron beat me to the punch on pointing out that we had this > > very code in there before. It appears that Stephen's > > assertion/assumption was incorrect about the stats structure being > > zero'd out, which is why we are seeing the issue. > > > > I have no issue reverting Stephen's earlier patch, or do we want to > > pursue why the stats structure is not zero'd out and resolve that > > instead. Either way, just want to make sure we are all on the same > > page as to the right solution so that we do not end up repeating this > > in the future. > > Lets's fix this in the base code. > > From: Stephen Hemminger <sthemmin@microsoft.com> > Date: Tue, 25 Apr 2017 10:50:19 -0700 > Subject: [PATCH net] net: always zero statistics > > Drivers with 32 bit statistics API also should get zeroed statistics. > > Fixes: 5944701df90d ("net: remove useless memset's in drivers get_stats64") This is probably a good change to do but it doesn't fix anything in 5944701df90d, especially not the problem with e1000e.
diff --git a/drivers/net/ethernet/intel/e1000e/netdev.c b/drivers/net/ethernet/int index 723025b..79651eb 100644 --- a/drivers/net/ethernet/intel/e1000e/netdev.c +++ b/drivers/net/ethernet/intel/e1000e/netdev.c @@ -5925,7 +5925,6 @@ void e1000e_get_stats64(struct net_device *netdev, { struct e1000_adapter *adapter = netdev_priv(netdev); - memset(stats, 0, sizeof(struct rtnl_link_stats64)); spin_lock(&adapter->stats64_lock); e1000e_update_stats(adapter); /* Fill out the OS statistics structure */