1 .. SPDX-License-Identifier: GPL-2.0
11 This document describes a set of techniques in the Linux networking stack to
12 take advantage of checksum offload capabilities of various NICs.
14 The following technologies are described:
17 * LCO: Local Checksum Offload
18 * RCO: Remote Checksum Offload
20 Things that should be documented here but aren't yet:
23 * CHECKSUM_UNNECESSARY conversion
29 The interface for offloading a transmit checksum to a device is explained in
30 detail in comments near the top of include/linux/skbuff.h.
32 In brief, it allows to request the device fill in a single ones-complement
33 checksum defined by the sk_buff fields skb->csum_start and skb->csum_offset.
34 The device should compute the 16-bit ones-complement checksum (i.e. the
35 'IP-style' checksum) from csum_start to the end of the packet, and fill in the
36 result at (csum_start + csum_offset).
38 Because csum_offset cannot be negative, this ensures that the previous value of
39 the checksum field is included in the checksum computation, thus it can be used
40 to supply any needed corrections to the checksum (such as the sum of the
41 pseudo-header for UDP or TCP).
43 This interface only allows a single checksum to be offloaded. Where
44 encapsulation is used, the packet may have multiple checksum fields in
45 different header layers, and the rest will have to be handled by another
46 mechanism such as LCO or RCO.
48 CRC32c can also be offloaded using this interface, by means of filling
49 skb->csum_start and skb->csum_offset as described above, and setting
50 skb->csum_not_inet: see skbuff.h comment (section 'D') for more details.
52 No offloading of the IP header checksum is performed; it is always done in
53 software. This is OK because when we build the IP header, we obviously have it
54 in cache, so summing it isn't expensive. It's also rather short.
56 The requirements for GSO are more complicated, because when segmenting an
57 encapsulated packet both the inner and outer checksums may need to be edited or
58 recomputed for each resulting segment. See the skbuff.h comment (section 'E')
61 A driver declares its offload capabilities in netdev->hw_features; see
62 Documentation/networking/netdev-features.rst for more. Note that a device
63 which only advertises NETIF_F_IP[V6]_CSUM must still obey the csum_start and
64 csum_offset given in the SKB; if it tries to deduce these itself in hardware
65 (as some NICs do) the driver should check that the values in the SKB match
66 those which the hardware will deduce, and if not, fall back to checksumming in
67 software instead (with skb_csum_hwoffload_help() or one of the
68 skb_checksum_help() / skb_crc32c_csum_help functions, as mentioned in
69 include/linux/skbuff.h).
71 The stack should, for the most part, assume that checksum offload is supported
72 by the underlying device. The only place that should check is
73 validate_xmit_skb(), and the functions it calls directly or indirectly. That
74 function compares the offload features requested by the SKB (which may include
75 other offloads besides TX Checksum Offload) and, if they are not supported or
76 enabled on the device (determined by netdev->features), performs the
77 corresponding offload in software. In the case of TX Checksum Offload, that
78 means calling skb_csum_hwoffload_help(skb, features).
81 LCO: Local Checksum Offload
82 ===========================
84 LCO is a technique for efficiently computing the outer checksum of an
85 encapsulated datagram when the inner checksum is due to be offloaded.
87 The ones-complement sum of a correctly checksummed TCP or UDP packet is equal
88 to the complement of the sum of the pseudo header, because everything else gets
89 'cancelled out' by the checksum field. This is because the sum was
90 complemented before being written to the checksum field.
92 More generally, this holds in any case where the 'IP-style' ones complement
93 checksum is used, and thus any checksum that TX Checksum Offload supports.
95 That is, if we have set up TX Checksum Offload with a start/offset pair, we
96 know that after the device has filled in that checksum, the ones complement sum
97 from csum_start to the end of the packet will be equal to the complement of
98 whatever value we put in the checksum field beforehand. This allows us to
99 compute the outer checksum without looking at the payload: we simply stop
100 summing when we get to csum_start, then add the complement of the 16-bit word
101 at (csum_start + csum_offset).
103 Then, when the true inner checksum is filled in (either by hardware or by
104 skb_checksum_help()), the outer checksum will become correct by virtue of the
107 LCO is performed by the stack when constructing an outer UDP header for an
108 encapsulation such as VXLAN or GENEVE, in udp_set_csum(). Similarly for the
109 IPv6 equivalents, in udp6_set_csum().
111 It is also performed when constructing an IPv4 GRE header, in
112 net/ipv4/ip_gre.c:build_header(). It is *not* currently performed when
113 constructing an IPv6 GRE header; the GRE checksum is computed over the whole
114 packet in net/ipv6/ip6_gre.c:ip6gre_xmit2(), but it should be possible to use
115 LCO here as IPv6 GRE still uses an IP-style checksum.
117 All of the LCO implementations use a helper function lco_csum(), in
118 include/linux/skbuff.h.
120 LCO can safely be used for nested encapsulations; in this case, the outer
121 encapsulation layer will sum over both its own header and the 'middle' header.
122 This does mean that the 'middle' header will get summed multiple times, but
123 there doesn't seem to be a way to avoid that without incurring bigger costs
127 RCO: Remote Checksum Offload
128 ============================
130 RCO is a technique for eliding the inner checksum of an encapsulated datagram,
131 allowing the outer checksum to be offloaded. It does, however, involve a
132 change to the encapsulation protocols, which the receiver must also support.
133 For this reason, it is disabled by default.
135 RCO is detailed in the following Internet-Drafts:
137 * https://tools.ietf.org/html/draft-herbert-remotecsumoffload-00
138 * https://tools.ietf.org/html/draft-herbert-vxlan-rco-00
140 In Linux, RCO is implemented individually in each encapsulation protocol, and
141 most tunnel types have flags controlling its use. For instance, VXLAN has the
142 flag VXLAN_F_REMCSUM_TX (per struct vxlan_rdst) to indicate that RCO should be
143 used when transmitting to a given remote destination.