1 Netdev features mess and how to get out from it alive
2 =====================================================
5 Michał Mirosław <mirq-linux@rere.qmqm.pl>
10 ======================
12 Long gone are the days when a network card would just take and give packets
13 verbatim. Today's devices add multiple features and bugs (read: offloads)
14 that relieve an OS of various tasks like generating and checking checksums,
15 splitting packets, classifying them. Those capabilities and their state
16 are commonly referred to as netdev features in Linux kernel world.
18 There are currently three sets of features relevant to the driver, and
19 one used internally by network core:
21 1. netdev->hw_features set contains features whose state may possibly
22 be changed (enabled or disabled) for a particular device by user's
23 request. This set should be initialized in ndo_init callback and not
26 2. netdev->features set contains features which are currently enabled
27 for a device. This should be changed only by network core or in
28 error paths of ndo_set_features callback.
30 3. netdev->vlan_features set contains features whose state is inherited
31 by child VLAN devices (limits netdev->features set). This is currently
32 used for all VLAN devices whether tags are stripped or inserted in
35 4. netdev->wanted_features set contains feature set requested by user.
36 This set is filtered by ndo_fix_features callback whenever it or
37 some device-specific conditions change. This set is internal to
38 networking core and should not be referenced in drivers.
42 Part II: Controlling enabled features
43 =======================================
45 When current feature set (netdev->features) is to be changed, new set
46 is calculated and filtered by calling ndo_fix_features callback
47 and netdev_fix_features(). If the resulting set differs from current
48 set, it is passed to ndo_set_features callback and (if the callback
49 returns success) replaces value stored in netdev->features.
50 NETDEV_FEAT_CHANGE notification is issued after that whenever current
51 set might have changed.
53 The following events trigger recalculation:
54 1. device's registration, after ndo_init returned success
55 2. user requested changes in features state
56 3. netdev_update_features() is called
58 ndo_*_features callbacks are called with rtnl_lock held. Missing callbacks
59 are treated as always returning success.
61 A driver that wants to trigger recalculation must do so by calling
62 netdev_update_features() while holding rtnl_lock. This should not be done
63 from ndo_*_features callbacks. netdev->features should not be modified by
64 driver except by means of ndo_fix_features callback.
68 Part III: Implementation hints
69 ================================
73 All dependencies between features should be resolved here. The resulting
74 set can be reduced further by networking core imposed limitations (as coded
75 in netdev_fix_features()). For this reason it is safer to disable a feature
76 when its dependencies are not met instead of forcing the dependency on.
78 This callback should not modify hardware nor driver state (should be
79 stateless). It can be called multiple times between successive
80 ndo_set_features calls.
82 Callback must not alter features contained in NETIF_F_SOFT_FEATURES or
83 NETIF_F_NEVER_CHANGE sets. The exception is NETIF_F_VLAN_CHALLENGED but
84 care must be taken as the change won't affect already configured VLANs.
88 Hardware should be reconfigured to match passed feature set. The set
89 should not be altered unless some error condition happens that can't
90 be reliably detected in ndo_fix_features. In this case, the callback
91 should update netdev->features to match resulting hardware state.
92 Errors returned are not (and cannot be) propagated anywhere except dmesg.
93 (Note: successful return is zero, >0 means silent error.)
100 For current list of features, see include/linux/netdev_features.h.
101 This section describes semantics of some of them.
103 * Transmit checksumming
105 For complete description, see comments near the top of include/linux/skbuff.h.
107 Note: NETIF_F_HW_CSUM is a superset of NETIF_F_IP_CSUM + NETIF_F_IPV6_CSUM.
108 It means that device can fill TCP/UDP-like checksum anywhere in the packets
109 whatever headers there might be.
111 * Transmit TCP segmentation offload
113 NETIF_F_TSO_ECN means that hardware can properly split packets with CWR bit
114 set, be it TCPv4 (when NETIF_F_TSO is enabled) or TCPv6 (NETIF_F_TSO6).
116 * Transmit DMA from high memory
118 On platforms where this is relevant, NETIF_F_HIGHDMA signals that
119 ndo_start_xmit can handle skbs with frags in high memory.
121 * Transmit scatter-gather
123 Those features say that ndo_start_xmit can handle fragmented skbs:
124 NETIF_F_SG --- paged skbs (skb_shinfo()->frags), NETIF_F_FRAGLIST ---
125 chained skbs (skb->next/prev list).
129 Features contained in NETIF_F_SOFT_FEATURES are features of networking
130 stack. Driver should not change behaviour based on them.
132 * LLTX driver (deprecated for hardware drivers)
134 NETIF_F_LLTX should be set in drivers that implement their own locking in
135 transmit path or don't need locking at all (e.g. software tunnels).
136 In ndo_start_xmit, it is recommended to use a try_lock and return
137 NETDEV_TX_LOCKED when the spin lock fails. The locking should also properly
138 protect against other callbacks (the rules you need to find out).
140 Don't use it for new drivers.
144 NETIF_F_NETNS_LOCAL is set for devices that are not allowed to move between
145 network namespaces (e.g. loopback).
147 Don't use it in drivers.
151 NETIF_F_VLAN_CHALLENGED should be set for devices which can't cope with VLAN
152 headers. Some drivers set this because the cards can't handle the bigger MTU.
153 [FIXME: Those cases could be fixed in VLAN code by allowing only reduced-MTU
154 VLANs. This may be not useful, though.]
158 This requests that the NIC append the Ethernet Frame Checksum (FCS)
159 to the end of the skb data. This allows sniffers and other tools to
160 read the CRC recorded by the NIC on receipt of the packet.
164 This requests that the NIC receive all possible frames, including errored
165 frames (such as bad FCS, etc). This can be helpful when sniffing a link with
166 bad packets on it. Some NICs may receive more packets if also put into normal