by Patrick McHardy

[permalink] [raw]

Subject: Re: [GIT] Networking

On Thu, May 02, 2013 at 04:16:25AM -0400, David Miller wrote:
> From: Patrick McHardy <[email protected]>
> Date: Thu, 2 May 2013 09:03:37 +0200
>
> > I'll also have a look at this.
>
> By the mere existence of /sys/devices/${DEV_PATH}/net/${netdev_name}/flags
> we have to preserve the bit layout.
>
> So Linus was right.
>
> So network manager is probably reading that flags sysfs file and
> interpreting it.

Right, that seems plausible.

> I'll fix the layout to how it was before.

I also found one spot in net/core/dev.c which was using an int for the
features. Patch attached.

Attachments:

(No filename) (589.00 B)
01.patch (677.00 B)
Download all attachments

2013-05-02 09:07:40

----- Original Message -----
> From: "Pavel Simerda" <[email protected]>
> To: "Michał Mirosław" <[email protected]>
> Cc: "Bjørn Mork" <[email protected]>, "Ben Hutchings" <[email protected]>, "David Miller" <[email protected]>,
> [email protected], [email protected], [email protected], [email protected],
> [email protected], [email protected], "Dan Williams" <[email protected]>, "Jiri Pirko"
> <[email protected]>
> Sent: Saturday, May 4, 2013 1:35:35 AM
> Subject: Re: [GIT] Networking
>
> ----- Original Message -----
> > From: "Michał Mirosław" <[email protected]>
> > To: "Bjørn Mork" <[email protected]>, "Pavel Šimerda" <[email protected]>
> > Cc: "Ben Hutchings" <[email protected]>, "David Miller"
> > <[email protected]>, [email protected],
> > [email protected], [email protected],
> > [email protected], [email protected],
> > [email protected]
> > Sent: Thursday, May 2, 2013 6:22:47 PM
> > Subject: Re: [GIT] Networking
> >
> > 2013/5/2 Bjørn Mork <[email protected]>:
> > > Ben Hutchings <[email protected]> writes:
> > >> On Thu, 2013-05-02 at 05:17 -0400, David Miller wrote:
> > >>> From: Bjørn Mork <[email protected]>
> > >>> Date: Thu, 02 May 2013 11:06:42 +0200
> > >>>
> > >>> > Adding the new netdev features will make it go from 1 to 2:
> > >>>
> > >>> We already had more than 31 feature bits before Patrick's
> > >>> changes, and I'm pretty sure this was the case when we added
> > >>> that ethtool API.
> > >>
> > >> It wasn't, but this should be OK. Userland is supposed to query the
> > >> number of features using ETHTOOL_GSSET_INFO and then work out the number
> > >> of words/blocks using FEATURE_BITS_TO_BLOCKS().
> > >
> > >
> > > Looking at
> > > http://cgit.freedesktop.org/NetworkManager/NetworkManager/tree/src/platform/nm-linux-platform.c#n1025
> > > there seems to be a couple of bugs in this area. This is certainly
> > > abusing the exported API, but it does mean that NM breaks if you ever
> > > move NETIF_F_VLAN_CHALLENGED (like the 802.1ad patch did):
> > >
> > > ----
> > > #define NETIF_F_VLAN_CHALLENGED (1 << 10)
> > >
> > > static gboolean
> > > link_supports_vlans (NMPlatform *platform, int ifindex)
> > > {
> > > auto_nl_object struct rtnl_link *rtnllink = link_get (platform,
> > > ifindex);
> > > const char *name = nm_platform_link_get_name (ifindex);
> > > struct {
> > > struct ethtool_gfeatures features;
> > > struct ethtool_get_features_block features_block;
> > > } edata = { .features = { .cmd = ETHTOOL_GFEATURES, .size = 1 }
> > > };
> > >
> > > /* Only ARPHRD_ETHER links can possibly support VLANs. */
> > > if (!rtnllink || rtnl_link_get_arptype (rtnllink) !=
> > > ARPHRD_ETHER)
> > > return FALSE;
> > >
> > > if (!name || !ethtool_get (name, &edata))
> > > return FALSE;
> > >
> > > return !(edata.features.features[0].active &
> > > NETIF_F_VLAN_CHALLENGED);
> > > }
> > > ----
> > >
> > >
> > > Not that I see how this particular bug matters unless you need VLAN
> > > support in NM. But there could be similar issues around. I guess
> > > avoiding unnecessary renumbering of the NETIF_F bits can save us some
> > > trouble. Although you can certainly argue that those bits never were
> > > intended to be part of the API, and that using them like this is a user
> > > application bug.
> >
> > This is certainly a bug in NM, and a fresh one: commit
> > b636ea86b1c0a28b77eda311c84d3b2417cad22e from 2013-04-10 14:40:58
> > (GMT). Userspace is expected to use ETHTOOL_GSTRINGS for
> > ETH_SS_FEATURES and find a corresponding bit position by feature name
> > ("vlan-challenged" in this case).
> >
> > Cc: commit's author.
>
> Recorded the bug with NetworkManager bugzilla blocking the upcoming release:
>
> https://bugzilla.gnome.org/show_bug.cgi?id=699649
>
> The code is experimental and is not used by NetworkManager even in the
> 'master' branch except by nm-platform's automated tests. Once we fix it, you
> don't need to worry about it.
>
> Thanks!
>
> Pavel

Fixed in NetworkManager master:

commit 7aefd5b5f4547c294092753b5cc355c95763134a
Author: Dan Winship <[email protected]>
Date: Mon May 6 12:10:01 2013 -0400

platform: fix use of ethtool

The bits in the result of ETHTOOL_GFEATURES are not in any defined
order; you need to use ETHTOOL_GSTRINGS to get the names associated
with each bit to find what each one does. Fix
NMPlatformLinux:link_supports_vlans() to do this.

https://bugzilla.gnome.org/show_bug.cgi?id=699649

diff --git a/src/platform/nm-linux-platform.c b/src/platform/nm-linux-platform.c
index 699c107..eeaeb2e 100644
--- a/src/platform/nm-linux-platform.c
+++ b/src/platform/nm-linux-platform.c
@@ -1024,6 +1024,40 @@ ethtool_get (const char *name, gpointer edata)
return TRUE;
}

+static int
+ethtool_get_stringset_index (const char *ifname, int stringset_id, const char *string)
+{
+ auto_g_free struct ethtool_sset_info *info;
+ auto_g_free struct ethtool_gstrings *strings;
+ guint32 len, i;
+
+ info = g_malloc0 (sizeof (*info) + sizeof (guint32));
+ info->cmd = ETHTOOL_GSSET_INFO;
+ info->reserved = 0;
+ info->sset_mask = 1ULL << stringset_id;
+
+ if (!ethtool_get (ifname, info))
+ return -1;
+ if (!info->sset_mask)
+ return -1;
+
+ len = info->data[0];
+
+ strings = g_malloc0 (sizeof (*strings) + len * ETH_GSTRING_LEN);
+ strings->cmd = ETHTOOL_GSTRINGS;
+ strings->string_set = stringset_id;
+ strings->len = len;
+ if (!ethtool_get (ifname, strings))
+ return -1;
+
+ for (i = 0; i < len; i++) {
+ if (!strcmp ((char *) &strings->data[i * ETH_GSTRING_LEN], string))
+ return i;
+ }
+
+ return -1;
+}
+
static gboolean
link_supports_carrier_detect (NMPlatform *platform, int ifindex)
{
@@ -1041,26 +1075,39 @@ link_supports_carrier_detect (NMPlatform *platform, int ifindex)
return name && ethtool_get (name, &edata);
}

-#define NETIF_F_VLAN_CHALLENGED (1 << 10)
-
static gboolean
link_supports_vlans (NMPlatform *platform, int ifindex)
{
auto_nl_object struct rtnl_link *rtnllink = link_get (platform, ifindex);
const char *name = nm_platform_link_get_name (ifindex);
- struct {
- struct ethtool_gfeatures features;
- struct ethtool_get_features_block features_block;
- } edata = { .features = { .cmd = ETHTOOL_GFEATURES, .size = 1 } };
+ auto_g_free struct ethtool_gfeatures *features;
+ int index, block, bit, size;

/* Only ARPHRD_ETHER links can possibly support VLANs. */
if (!rtnllink || rtnl_link_get_arptype (rtnllink) != ARPHRD_ETHER)
return FALSE;

- if (!name || !ethtool_get (name, &edata))
+ if (!name)
+ return FALSE;
+
+ index = ethtool_get_stringset_index (name, ETH_SS_FEATURES, "vlan-challenged");
+ if (index == -1) {
+ debug ("vlan-challenged ethtool feature does not exist?");
+ return FALSE;
+ }
+
+ block = index / 32;
+ bit = index % 32;
+ size = block + 1;
+
+ features = g_malloc0 (sizeof (*features) + size * sizeof (struct ethtool_get_features_block));
+ features->cmd = ETHTOOL_GFEATURES;
+ features->size = size;
+
+ if (!ethtool_get (name, features))
return FALSE;

- return !(edata.features.features[0].active & NETIF_F_VLAN_CHALLENGED);
+ return !(features->features[block].active & (1 << bit));
}

static gboolean