RFC 3542
Advanced Sockets Application Program Interface (API) for IPv6
3. IPv6 Raw Sockets
Raw sockets bypass the transport layer (TCP or UDP). With IPv4, raw sockets are used to access ICMPv4, IGMPv4, and to read and write IPv4 datagrams containing a protocol field that the kernel does not process. An example of the latter is a routing daemon for OSPF, since it uses IPv4 protocol field 89. With IPv6 raw sockets will be used for ICMPv6 and to read and write IPv6 datagrams containing a Next Header field that the kernel does not process. Examples of the latter are a routing daemon for OSPF for IPv6 and RSVP (protocol field 46). All data sent via raw sockets must be in network byte order and all data received via raw sockets will be in network byte order. This differs from the IPv4 raw sockets, which did not specify a byte ordering and used the host's byte order for certain IP header fields. Another difference from IPv4 raw sockets is that complete packets (that is, IPv6 packets with extension headers) cannot be sent or received using the IPv6 raw sockets API. Instead, ancillary data objects are used to transfer the extension headers and hoplimit information, as described in Section 6. Should an application need access to the complete IPv6 packet, some other technique, such as the datalink interfaces BPF or DLPI, must be used. All fields except the flow label in the IPv6 header that an application might want to change (i.e., everything other than the version number) can be modified using ancillary data and/or socket options by the application for output. All fields except the flow label in a received IPv6 header (other than the version number and Next Header fields) and all extension headers that an application might want to know are also made available to the application as ancillary data on input. Hence there is no need for a socket option
similar to the IPv4 IP_HDRINCL socket option and on receipt the application will only receive the payload i.e., the data after the IPv6 header and all the extension headers. This API does not define access to the flow label field, because today there is no standard usage of the field. When writing to a raw socket the kernel will automatically fragment the packet if its size exceeds the path MTU, inserting the required fragment headers. On input the kernel reassembles received fragments, so the reader of a raw socket never sees any fragment headers. When we say "an ICMPv6 raw socket" we mean a socket created by calling the socket function with the three arguments AF_INET6, SOCK_RAW, and IPPROTO_ICMPV6. Most IPv4 implementations give special treatment to a raw socket created with a third argument to socket() of IPPROTO_RAW, whose value is normally 255, to have it mean that the application will send down complete packets including the IPv4 header. (Note: This feature was added to IPv4 in 1988 by Van Jacobson to support traceroute, allowing a complete IP header to be passed by the application, before the IP_HDRINCL socket option was added.) We note that IPPROTO_RAW has no special meaning to an IPv6 raw socket (and the IANA currently reserves the value of 255 when used as a next-header field).3.1. Checksums
The kernel will calculate and insert the ICMPv6 checksum for ICMPv6 raw sockets, since this checksum is mandatory. For other raw IPv6 sockets (that is, for raw IPv6 sockets created with a third argument other than IPPROTO_ICMPV6), the application must set the new IPV6_CHECKSUM socket option to have the kernel (1) compute and store a checksum for output, and (2) verify the received checksum on input, discarding the packet if the checksum is in error. This option prevents applications from having to perform source address selection on the packets they send. The checksum will incorporate the IPv6 pseudo-header, defined in Section 8.1 of [RFC- 2460]. This new socket option also specifies an integer offset into the user data of where the checksum is located. int offset = 2; setsockopt(fd, IPPROTO_IPV6, IPV6_CHECKSUM, &offset, sizeof(offset));
By default, this socket option is disabled. Setting the offset to -1 also disables the option. By disabled we mean (1) the kernel will not calculate and store a checksum for outgoing packets, and (2) the kernel will not verify a checksum for received packets. This option assumes the use of the 16-bit one's complement of the one's complement sum as the checksum algorithm and that the checksum field is aligned on a 16-bit boundary. Thus, specifying a positive odd value as offset is invalid, and setsockopt() will fail for such offset values. An attempt to set IPV6_CHECKSUM for an ICMPv6 socket will fail. Also, an attempt to set or get IPV6_CHECKSUM for a non-raw IPv6 socket will fail. (Note: Since the checksum is always calculated by the kernel for an ICMPv6 socket, applications are not able to generate ICMPv6 packets with incorrect checksums (presumably for testing purposes) using this API.)3.2. ICMPv6 Type Filtering
ICMPv4 raw sockets receive most ICMPv4 messages received by the kernel. (We say "most" and not "all" because Berkeley-derived kernels never pass echo requests, timestamp requests, or address mask requests to a raw socket. Instead these three messages are processed entirely by the kernel.) But ICMPv6 is a superset of ICMPv4, also including the functionality of IGMPv4 and ARPv4. This means that an ICMPv6 raw socket can potentially receive many more messages than would be received with an ICMPv4 raw socket: ICMP messages similar to ICMPv4, along with neighbor solicitations, neighbor advertisements, and the three multicast listener discovery messages. Most applications using an ICMPv6 raw socket care about only a small subset of the ICMPv6 message types. To transfer extraneous ICMPv6 messages from the kernel to user can incur a significant overhead. Therefore this API includes a method of filtering ICMPv6 messages by the ICMPv6 type field. Each ICMPv6 raw socket has an associated filter whose datatype is defined as struct icmp6_filter; This structure, along with the macros and constants defined later in this section, are defined as a result of including the <netinet/icmp6.h>.
The current filter is fetched and stored using getsockopt() and
setsockopt() with a level of IPPROTO_ICMPV6 and an option name of
ICMP6_FILTER.
Six macros operate on an icmp6_filter structure:
void ICMP6_FILTER_SETPASSALL (struct icmp6_filter *);
void ICMP6_FILTER_SETBLOCKALL(struct icmp6_filter *);
void ICMP6_FILTER_SETPASS ( int, struct icmp6_filter *);
void ICMP6_FILTER_SETBLOCK( int, struct icmp6_filter *);
int ICMP6_FILTER_WILLPASS (int,
const struct icmp6_filter *);
int ICMP6_FILTER_WILLBLOCK(int,
const struct icmp6_filter *);
The first argument to the last four macros (an integer) is an ICMPv6
message type, between 0 and 255. The pointer argument to all six
macros is a pointer to a filter that is modified by the first four
macros and is examined by the last two macros.
The first two macros, SETPASSALL and SETBLOCKALL, let us specify that
all ICMPv6 messages are passed to the application or that all ICMPv6
messages are blocked from being passed to the application.
The next two macros, SETPASS and SETBLOCK, let us specify that
messages of a given ICMPv6 type should be passed to the application
or not passed to the application (blocked).
The final two macros, WILLPASS and WILLBLOCK, return true or false
depending whether the specified message type is passed to the
application or blocked from being passed to the application by the
filter pointed to by the second argument.
When an ICMPv6 raw socket is created, it will by default pass all
ICMPv6 message types to the application.
As an example, a program that wants to receive only router
advertisements could execute the following:
struct icmp6_filter myfilt;
fd = socket(AF_INET6, SOCK_RAW, IPPROTO_ICMPV6);
ICMP6_FILTER_SETBLOCKALL(&myfilt);
ICMP6_FILTER_SETPASS(ND_ROUTER_ADVERT, &myfilt);
setsockopt(fd, IPPROTO_ICMPV6, ICMP6_FILTER, &myfilt,
sizeof(myfilt));
The filter structure is declared and then initialized to block all
messages types. The filter structure is then changed to allow router
advertisement messages to be passed to the application and the filter
is installed using setsockopt().
In order to clear an installed filter the application can issue a
setsockopt for ICMP6_FILTER with a zero length. When no such filter
has been installed, getsockopt() will return the kernel default
filter.
The icmp6_filter structure is similar to the fd_set datatype used
with the select() function in the sockets API. The icmp6_filter
structure is an opaque datatype and the application should not care
how it is implemented. All the application does with this datatype
is allocate a variable of this type, pass a pointer to a variable of
this type to getsockopt() and setsockopt(), and operate on a variable
of this type using the six macros that we just defined.
Nevertheless, it is worth showing a simple implementation of this
datatype and the six macros.
struct icmp6_filter {
uint32_t icmp6_filt[8]; /* 8*32 = 256 bits */
};
#define ICMP6_FILTER_WILLPASS(type, filterp) \
((((filterp)->icmp6_filt[(type) >> 5]) & \
(1 << ((type) & 31))) != 0)
#define ICMP6_FILTER_WILLBLOCK(type, filterp) \
((((filterp)->icmp6_filt[(type) >> 5]) & \
(1 << ((type) & 31))) == 0)
#define ICMP6_FILTER_SETPASS(type, filterp) \
((((filterp)->icmp6_filt[(type) >> 5]) |= \
(1 << ((type) & 31))))
#define ICMP6_FILTER_SETBLOCK(type, filterp) \
((((filterp)->icmp6_filt[(type) >> 5]) &= \
~(1 << ((type) & 31))))
#define ICMP6_FILTER_SETPASSALL(filterp) \
memset((filterp), 0xFF, sizeof(struct icmp6_filter))
#define ICMP6_FILTER_SETBLOCKALL(filterp) \
memset((filterp), 0, sizeof(struct icmp6_filter))
(Note: These sample definitions have two limitations that an
implementation may want to change. The first four macros evaluate
their first argument two times. The second two macros require the
inclusion of the <string.h> header for the memset() function.)
3.3. ICMPv6 Verification of Received Packets
The protocol stack will verify the ICMPv6 checksum and discard any packets with invalid checksums. An implementation might perform additional validity checks on the ICMPv6 message content and discard malformed packets. However, a portable application must not assume that such validity checks have been performed. The protocol stack should not automatically discard packets if the ICMP type is unknown to the stack. For extensibility reasons received ICMP packets with any type (informational or error) must be passed to the applications (subject to ICMP6_FILTER filtering on the type value and the checksum verification).4. Access to IPv6 and Extension Headers
Applications need to be able to control IPv6 header and extension header content when sending as well as being able to receive the content of these headers. This is done by defining socket option types which can be used both with setsockopt and with ancillary data. Ancillary data is discussed in Appendix A. The following optional information can be exchanged between the application and the kernel: 1. The send/receive interface and source/destination address, 2. The hop limit, 3. Next hop address, 4. The traffic class, 5. Routing header, 6. Hop-by-Hop options header, and 7. Destination options header. First, to receive any of this optional information (other than the next hop address, which can only be set) on a UDP or raw socket, the application must call setsockopt() to turn on the corresponding flag: int on = 1; setsockopt(fd, IPPROTO_IPV6, IPV6_RECVPKTINFO, &on, sizeof(on)); setsockopt(fd, IPPROTO_IPV6, IPV6_RECVHOPLIMIT, &on, sizeof(on)); setsockopt(fd, IPPROTO_IPV6, IPV6_RECVRTHDR, &on, sizeof(on)); setsockopt(fd, IPPROTO_IPV6, IPV6_RECVHOPOPTS, &on, sizeof(on)); setsockopt(fd, IPPROTO_IPV6, IPV6_RECVDSTOPTS, &on, sizeof(on)); setsockopt(fd, IPPROTO_IPV6, IPV6_RECVTCLASS, &on, sizeof(on));
When any of these options are enabled, the corresponding data is returned as control information by recvmsg(), as one or more ancillary data objects. This document does not define how to receive the optional information on a TCP socket. See Section 4.1 for more details. Two different mechanisms exist for sending this optional information: 1. Using setsockopt to specify the option content for a socket. These are known "sticky" options since they affect all transmitted packets on the socket until either a new setsockopt is done or the options are overridden using ancillary data. 2. Using ancillary data to specify the option content for a single datagram. This only applies to datagram and raw sockets; not to TCP sockets. The three socket option parameters and the three cmsghdr fields that describe the options/ancillary data objects are summarized as: opt level/ optname/ optval/ cmsg_level cmsg_type cmsg_data[] ------------ ------------ ------------------------ IPPROTO_IPV6 IPV6_PKTINFO in6_pktinfo structure IPPROTO_IPV6 IPV6_HOPLIMIT int IPPROTO_IPV6 IPV6_NEXTHOP socket address structure IPPROTO_IPV6 IPV6_RTHDR ip6_rthdr structure IPPROTO_IPV6 IPV6_HOPOPTS ip6_hbh structure IPPROTO_IPV6 IPV6_DSTOPTS ip6_dest structure IPPROTO_IPV6 IPV6_RTHDRDSTOPTS ip6_dest structure IPPROTO_IPV6 IPV6_TCLASS int (Note: IPV6_HOPLIMIT can be used as ancillary data items only) All these options are described in detail in Section 6, 7, 8 and 9. All the constants beginning with IPV6_ are defined as a result of including <netinet/in.h>. Note: We intentionally use the same constant for the cmsg_level member as is used as the second argument to getsockopt() and setsockopt() (what is called the "level"), and the same constant for the cmsg_type member as is used as the third argument to getsockopt() and setsockopt() (what is called the "option name").
Issuing getsockopt() for the above options will return the sticky
option value i.e., the value set with setsockopt(). If no sticky
option value has been set getsockopt() will return the following
values:
- For the IPV6_PKTINFO option, it will return an in6_pktinfo
structure with ipi6_addr being in6addr_any and ipi6_ifindex being
zero.
- For the IPV6_TCLASS option, it will return the kernel default
value.
- For other options, it will indicate the lack of the option value
with optlen being zero.
The application does not explicitly need to access the data
structures for the Routing header, Hop-by-Hop options header, and
Destination options header, since the API to these features is
through a set of inet6_rth_XXX() and inet6_opt_XXX() functions that
we define in Section 7 and Section 10. Those functions simplify the
interface to these features instead of requiring the application to
know the intimate details of the extension header formats.
When specifying extension headers, this API assumes the header
ordering and the number of occurrences of each header as described in
[RFC-2460]. More details about the ordering issue will be discussed
in Section 12.
4.1. TCP Implications
It is not possible to use ancillary data to transmit the above
options for TCP since there is not a one-to-one mapping between send
operations and the TCP segments being transmitted. Instead an
application can use setsockopt to specify them as sticky options.
When the application uses setsockopt to specify the above options it
is expected that TCP will start using the new information when
sending segments. However, TCP may or may not use the new
information when retransmitting segments that were originally sent
when the old sticky options were in effect.
It is unclear how a TCP application can use received information
(such as extension headers) due to the lack of mapping between
received TCP segments and receive operations. In particular, the
received information could not be used for access control purposes
like on UDP and raw sockets.
This specification therefore does not define how to get the received information on TCP sockets. The result of the IPV6_RECVxxx options on a TCP socket is undefined as well.4.2. UDP and Raw Socket Implications
The receive behavior for UDP and raw sockets is quite straightforward. After the application has enabled an IPV6_RECVxxx socket option it will receive ancillary data items for every recvmsg() call containing the requested information. However, if the information is not present in the packet the ancillary data item will not be included. For example, if the application enables IPV6_RECVRTHDR and a received datagram does not contain a Routing header there will not be an IPV6_RTHDR ancillary data item. Note that due to buffering in the socket implementation there might be some packets queued when an IPV6_RECVxxx option is enabled and they might not have the ancillary data information. For sending the application has the choice between using sticky options and ancillary data. The application can also use both having the sticky options specify the "default" and using ancillary data to override the default options. When an ancillary data item is specified in a call to sendmsg(), the item will override an existing sticky option of the same name (if previously specified). For example, if the application has set IPV6_RTHDR using a sticky option and later passes IPV6_RTHDR as ancillary data this will override the IPV6_RTHDR sticky option and the routing header of the outgoing packet will be from the ancillary data item, not from the sticky option. Note, however, that other sticky options than IPV6_RTHDR will not be affected by the IPV6_RTHDR ancillary data item; the overriding mechanism only works for the same type of sticky options and ancillary data items. (Note: the overriding rule is different from the one in RFC 2292. In RFC 2292, an ancillary data item overrode all sticky options previously defined. This was reasonable, because sticky options could only be specified as a set by a single socket option. However, in this API, each option is separated so that it can be specified as a single sticky option. Additionally, there are much more ancillary data items and sticky options than in RFC 2292, including ancillary- only one. Thus, it should be natural for application programmers to separate the overriding rule as well.) An application can also temporarily disable a particular sticky option by specifying a corresponding ancillary data item that could disable the sticky option when being used as an argument for a socket option. For example, if the application has set IPV6_HOPOPTS as a
sticky option and later passes IPV6_HOPOPTS with a zero length as an ancillary data item, the packet will not have a Hop-by-Hop options header.5. Extensions to Socket Ancillary Data
This specification uses ancillary data as defined in Posix with some compatible extensions, which are described in the following subsections. Section 20 will provide a detailed overview of ancillary data and related structures and macros, including the extensions.5.1. CMSG_NXTHDR
struct cmsghdr *CMSG_NXTHDR(const struct msghdr *mhdr, const struct cmsghdr *cmsg); CMSG_NXTHDR() returns a pointer to the cmsghdr structure describing the next ancillary data object. Mhdr is a pointer to a msghdr structure and cmsg is a pointer to a cmsghdr structure. If there is not another ancillary data object, the return value is NULL. The following behavior of this macro is new to this API: if the value of the cmsg pointer is NULL, a pointer to the cmsghdr structure describing the first ancillary data object is returned. That is, CMSG_NXTHDR(mhdr, NULL) is equivalent to CMSG_FIRSTHDR(mhdr). If there are no ancillary data objects, the return value is NULL.5.2. CMSG_SPACE
socklen_t CMSG_SPACE(socklen_t length); This macro is new with this API. Given the length of an ancillary data object, CMSG_SPACE() returns an upper bound on the space required by the object and its cmsghdr structure, including any padding needed to satisfy alignment requirements. This macro can be used, for example, when allocating space dynamically for the ancillary data. This macro should not be used to initialize the cmsg_len member of a cmsghdr structure; instead use the CMSG_LEN() macro.
5.3. CMSG_LEN
socklen_t CMSG_LEN(socklen_t length); This macro is new with this API. Given the length of an ancillary data object, CMSG_LEN() returns the value to store in the cmsg_len member of the cmsghdr structure, taking into account any padding needed to satisfy alignment requirements. Note the difference between CMSG_SPACE() and CMSG_LEN(), shown also in the figure in Section 20.2: the former accounts for any required padding at the end of the ancillary data object and the latter is the actual length to store in the cmsg_len member of the ancillary data object.6. Packet Information
There are five pieces of information that an application can specify for an outgoing packet using ancillary data: 1. the source IPv6 address, 2. the outgoing interface index, 3. the outgoing hop limit, 4. the next hop address, and 5. the outgoing traffic class value. Four similar pieces of information can be returned for a received packet as ancillary data: 1. the destination IPv6 address, 2. the arriving interface index, 3. the arriving hop limit, and 4. the arriving traffic class value. The first two pieces of information are contained in an in6_pktinfo structure that is set with setsockopt() or sent as ancillary data with sendmsg() and received as ancillary data with recvmsg(). This structure is defined as a result of including <netinet/in.h>. struct in6_pktinfo { struct in6_addr ipi6_addr; /* src/dst IPv6 address */ unsigned int ipi6_ifindex; /* send/recv interface index */ }; In the socket option and cmsghdr level will be IPPROTO_IPV6, the type will be IPV6_PKTINFO, and the first byte of the option value and cmsg_data[] will be the first byte of the in6_pktinfo structure. An application can clear any sticky IPV6_PKTINFO option by doing a
"regular" setsockopt with ipi6_addr being in6addr_any and
ipi6_ifindex being zero.
This information is returned as ancillary data by recvmsg() only if
the application has enabled the IPV6_RECVPKTINFO socket option:
int on = 1;
setsockopt(fd, IPPROTO_IPV6, IPV6_RECVPKTINFO, &on, sizeof(on));
(Note: The hop limit is not contained in the in6_pktinfo structure
for the following reason. Some UDP servers want to respond to client
requests by sending their reply out the same interface on which the
request was received and with the source IPv6 address of the reply
equal to the destination IPv6 address of the request. To do this the
application can enable just the IPV6_RECVPKTINFO socket option and
then use the received control information from recvmsg() as the
outgoing control information for sendmsg(). The application need not
examine or modify the in6_pktinfo structure at all. But if the hop
limit were contained in this structure, the application would have to
parse the received control information and change the hop limit
member, since the received hop limit is not the desired value for an
outgoing packet.)
6.1. Specifying/Receiving the Interface
Interfaces on an IPv6 node are identified by a small positive
integer, as described in Section 4 of [RFC-3493]. That document also
describes a function to map an interface name to its interface index,
a function to map an interface index to its interface name, and a
function to return all the interface names and indexes. Notice from
this document that no interface is ever assigned an index of 0.
When specifying the outgoing interface, if the ipi6_ifindex value is
0, the kernel will choose the outgoing interface.
The ordering among various options that can specify the outgoing
interface, including IPV6_PKTINFO, is defined in Section 6.7.
When the IPV6_RECVPKTINFO socket option is enabled, the received
interface index is always returned as the ipi6_ifindex member of the
in6_pktinfo structure.
6.2. Specifying/Receiving Source/Destination Address
The source IPv6 address can be specified by calling bind() before each output operation, but supplying the source address together with the data requires less overhead (i.e., fewer system calls) and requires less state to be stored and protected in a multithreaded application. When specifying the source IPv6 address as ancillary data, if the ipi6_addr member of the in6_pktinfo structure is the unspecified address (IN6ADDR_ANY_INIT or in6addr_any), then (a) if an address is currently bound to the socket, it is used as the source address, or (b) if no address is currently bound to the socket, the kernel will choose the source address. If the ipi6_addr member is not the unspecified address, but the socket has already bound a source address, then the ipi6_addr value overrides the already-bound source address for this output operation only. The kernel must verify that the requested source address is indeed a unicast address assigned to the node. When the address is a scoped one, there may be ambiguity about its scope zone. This is particularly the case for link-local addresses. In such a case, the kernel must first determine the appropriate scope zone based on the zone of the destination address or the outgoing interface (if known), then qualify the address. This also means that it is not feasible to specify the source address for a non-binding socket by the IPV6_PKTINFO sticky option, unless the outgoing interface is also specified. The application should simply use bind() for such purposes. IPV6_PKTINFO can also be used as a sticky option for specifying the socket's default source address. However, the ipi6_addr member must be the unspecified address for TCP sockets, because it is not possible to dynamically change the source address of a TCP connection. When the IPV6_PKTINFO option is specified for a TCP socket with a non-unspecified address, the call will fail. This restriction should be applied even before the socket binds a specific address. When the in6_pktinfo structure is returned as ancillary data by recvmsg(), the ipi6_addr member contains the destination IPv6 address from the received packet.6.3. Specifying/Receiving the Hop Limit
The outgoing hop limit is normally specified with either the IPV6_UNICAST_HOPS socket option or the IPV6_MULTICAST_HOPS socket option, both of which are described in [RFC-3493]. Specifying the
hop limit as ancillary data lets the application override either the
kernel's default or a previously specified value, for either a
unicast destination or a multicast destination, for a single output
operation. Returning the received hop limit is useful for IPv6
applications that need to verify that the received hop limit is 255
(e.g., that the packet has not been forwarded).
The received hop limit is returned as ancillary data by recvmsg()
only if the application has enabled the IPV6_RECVHOPLIMIT socket
option:
int on = 1;
setsockopt(fd, IPPROTO_IPV6, IPV6_RECVHOPLIMIT, &on, sizeof(on));
In the cmsghdr structure containing this ancillary data, the
cmsg_level member will be IPPROTO_IPV6, the cmsg_type member will be
IPV6_HOPLIMIT, and the first byte of cmsg_data[] will be the first
byte of the integer hop limit.
Nothing special need be done to specify the outgoing hop limit: just
specify the control information as ancillary data for sendmsg(). As
specified in [RFC-3493], the interpretation of the integer hop limit
value is
x < -1: return an error of EINVAL
x == -1: use kernel default
0 <= x <= 255: use x
x >= 256: return an error of EINVAL
This API defines IPV6_HOPLIMIT as an ancillary-only option, that is,
the option name cannot be used as a socket option. This is because
[RFC-3493] has more fine-grained socket options; IPV6_UNICAST_HOPS
and IPV6_MULTICAST_HOPS.
6.4. Specifying the Next Hop Address
The IPV6_NEXTHOP ancillary data object specifies the next hop for the
datagram as a socket address structure. In the cmsghdr structure
containing this ancillary data, the cmsg_level member will be
IPPROTO_IPV6, the cmsg_type member will be IPV6_NEXTHOP, and the
first byte of cmsg_data[] will be the first byte of the socket
address structure.
This is a privileged option. (Note: It is implementation defined and
beyond the scope of this document to define what "privileged" means.
Unix systems use this term to mean the process must have an effective
user ID of 0.)
This API only defines the case where the socket address contains an IPv6 address (i.e., the sa_family member is AF_INET6). And, in this case, the node identified by that address must be a neighbor of the sending host. If that address equals the destination IPv6 address of the datagram, then this is equivalent to the existing SO_DONTROUTE socket option. This option does not have any meaning for multicast destinations. In such a case, the specified next hop will be ignored. When the outgoing interface is specified by IPV6_PKTINFO as well, the next hop specified by this option must be reachable via the specified interface. In order to clear a sticky IPV6_NEXTHOP option the application must issue a setsockopt for IPV6_NEXTHOP with a zero length.6.5. Specifying/Receiving the Traffic Class value
The outgoing traffic class is normally set to 0. Specifying the traffic class as ancillary data lets the application override either the kernel's default or a previously specified value, for either a unicast destination or a multicast destination, for a single output operation. Returning the received traffic class is useful for programs such as a diffserv debugging tool and for user level ECN (explicit congestion notification) implementation. The received traffic class is returned as ancillary data by recvmsg() only if the application has enabled the IPV6_RECVTCLASS socket option: int on = 1; setsockopt(fd, IPPROTO_IPV6, IPV6_RECVTCLASS, &on, sizeof(on)); In the cmsghdr structure containing this ancillary data, the cmsg_level member will be IPPROTO_IPV6, the cmsg_type member will be IPV6_TCLASS, and the first byte of cmsg_data[] will be the first byte of the integer traffic class. To specify the outgoing traffic class value, just specify the control information as ancillary data for sendmsg() or using setsockopt(). Just like the hop limit value, the interpretation of the integer traffic class value is x < -1: return an error of EINVAL x == -1: use kernel default 0 <= x <= 255: use x x >= 256: return an error of EINVAL
In order to clear a sticky IPV6_TCLASS option the application can specify -1 as the value. There are cases where the kernel needs to control the traffic class value and conflicts with the user-specified value on the outgoing traffic. An example is an implementation of ECN in the kernel, setting 2 bits of the traffic class value. In such cases, the kernel should override the user-specified value. On the incoming traffic, the kernel may mask some of the bits in the traffic class field.6.6. Additional Errors with sendmsg() and setsockopt()
With the IPV6_PKTINFO socket option there are no additional errors possible with the call to recvmsg(). But when specifying the outgoing interface or the source address, additional errors are possible from sendmsg() or setsockopt(). Note that some implementations might only be able to return this type of errors for setsockopt(). The following are examples, but some of these may not be provided by some implementations, and some implementations may define additional errors: ENXIO The interface specified by ipi6_ifindex does not exist. ENETDOWN The interface specified by ipi6_ifindex is not enabled for IPv6 use. EADDRNOTAVAIL ipi6_ifindex specifies an interface but the address ipi6_addr is not available for use on that interface. EHOSTUNREACH No route to the destination exists over the interface specified by ipi6_ifindex.6.7. Summary of Outgoing Interface Selection
This document and [RFC-3493] specify various methods that affect the selection of the packet's outgoing interface. This subsection summarizes the ordering among those in order to ensure deterministic behavior. For a given outgoing packet on a given socket, the outgoing interface is determined in the following order: 1. if an interface is specified in an IPV6_PKTINFO ancillary data item, the interface is used. 2. otherwise, if an interface is specified in an IPV6_PKTINFO sticky option, the interface is used.
3. otherwise, if the destination address is a multicast address and
the IPV6_MULTICAST_IF socket option is specified for the socket,
the interface is used.
4. otherwise, if an IPV6_NEXTHOP ancillary data item is specified,
the interface to the next hop is used.
5. otherwise, if an IPV6_NEXTHOP sticky option is specified, the
interface to the next hop is used.
6. otherwise, the outgoing interface should be determined in an
implementation dependent manner.
The ordering above particularly means if the application specifies an
interface by the IPV6_MULTICAST_IF socket option (described in [RFC-
3493]) as well as specifying a different interface by the
IPV6_PKTINFO sticky option, the latter will override the former for
every multicast packet on the corresponding socket. The reason for
the ordering comes from expectation that the source address is
specified as well and that the pair of the address and the outgoing
interface should be preferred.
In any case, the kernel must also verify that the source and
destination addresses do not break their scope zones with regard to
the outgoing interface.
7. Routing Header Option
Source routing in IPv6 is accomplished by specifying a Routing header
as an extension header. There can be different types of Routing
headers, but IPv6 currently defines only the Type 0 Routing header
[RFC-2460]. This type supports up to 127 intermediate nodes (limited
by the length field in the extension header). With this maximum
number of intermediate nodes, a source, and a destination, there are
128 hops.
Source routing with the IPv4 sockets API (the IP_OPTIONS socket
option) requires the application to build the source route in the
format that appears as the IPv4 header option, requiring intimate
knowledge of the IPv4 options format. This IPv6 API, however,
defines six functions that the application calls to build and examine
a Routing header, and the ability to use sticky options or ancillary
data to communicate this information between the application and the
kernel using the IPV6_RTHDR option.
Three functions build a Routing header:
inet6_rth_space() - return #bytes required for Routing header
inet6_rth_init() - initialize buffer data for Routing header
inet6_rth_add() - add one IPv6 address to the Routing header
Three functions deal with a returned Routing header:
inet6_rth_reverse() - reverse a Routing header
inet6_rth_segments() - return #segments in a Routing header
inet6_rth_getaddr() - fetch one address from a Routing header
The function prototypes for these functions are defined as a result
of including <netinet/in.h>.
To receive a Routing header the application must enable the
IPV6_RECVRTHDR socket option:
int on = 1;
setsockopt(fd, IPPROTO_IPV6, IPV6_RECVRTHDR, &on, sizeof(on));
Each received Routing header is returned as one ancillary data object
described by a cmsghdr structure with cmsg_type set to IPV6_RTHDR.
When multiple Routing headers are received, multiple ancillary data
objects (with cmsg_type set to IPV6_RTHDR) will be returned to the
application.
To send a Routing header the application specifies it either as
ancillary data in a call to sendmsg() or using setsockopt(). For the
sending side, this API assumes the number of occurrences of the
Routing header as described in [RFC-2460]. That is, applications can
only specify at most one outgoing Routing header.
The application can remove any sticky Routing header by calling
setsockopt() for IPV6_RTHDR with a zero option length.
When using ancillary data a Routing header is passed between the
application and the kernel as follows: The cmsg_level member has a
value of IPPROTO_IPV6 and the cmsg_type member has a value of
IPV6_RTHDR. The contents of the cmsg_data[] member is implementation
dependent and should not be accessed directly by the application, but
should be accessed using the six functions that we are about to
describe.
The following constant is defined as a result of including the
<netinet/in.h>:
#define IPV6_RTHDR_TYPE_0 0 /* IPv6 Routing header type 0 */
When a Routing header is specified, the destination address specified for connect(), sendto(), or sendmsg() is the final destination address of the datagram. The Routing header then contains the addresses of all the intermediate nodes.7.1. inet6_rth_space
socklen_t inet6_rth_space(int type, int segments); This function returns the number of bytes required to hold a Routing header of the specified type containing the specified number of segments (addresses). For an IPv6 Type 0 Routing header, the number of segments must be between 0 and 127, inclusive. The return value is just the space for the Routing header. When the application uses ancillary data it must pass the returned length to CMSG_SPACE() to determine how much memory is needed for the ancillary data object (including the cmsghdr structure). If the return value is 0, then either the type of the Routing header is not supported by this implementation or the number of segments is invalid for this type of Routing header. (Note: This function returns the size but does not allocate the space required for the ancillary data. This allows an application to allocate a larger buffer, if other ancillary data objects are desired, since all the ancillary data objects must be specified to sendmsg() as a single msg_control buffer.)7.2. inet6_rth_init
void *inet6_rth_init(void *bp, socklen_t bp_len, int type, int segments); This function initializes the buffer pointed to by bp to contain a Routing header of the specified type and sets ip6r_len based on the segments parameter. bp_len is only used to verify that the buffer is large enough. The ip6r_segleft field is set to zero; inet6_rth_add() will increment it. When the application uses ancillary data the application must initialize any cmsghdr fields. The caller must allocate the buffer and its size can be determined by calling inet6_rth_space(). Upon success the return value is the pointer to the buffer (bp), and this is then used as the first argument to the inet6_rth_add() function. Upon an error the return value is NULL.
7.3. inet6_rth_add
int inet6_rth_add(void *bp, const struct in6_addr *addr); This function adds the IPv6 address pointed to by addr to the end of the Routing header being constructed. If successful, the segleft member of the Routing Header is updated to account for the new address in the Routing header and the return value of the function is 0. Upon an error the return value of the function is -1.7.4. inet6_rth_reverse
int inet6_rth_reverse(const void *in, void *out); This function takes a Routing header extension header (pointed to by the first argument) and writes a new Routing header that sends datagrams along the reverse of that route. The function reverses the order of the addresses and sets the segleft member in the new Routing header to the number of segments. Both arguments are allowed to point to the same buffer (that is, the reversal can occur in place). The return value of the function is 0 on success, or -1 upon an error.7.5. inet6_rth_segments
int inet6_rth_segments(const void *bp); This function returns the number of segments (addresses) contained in the Routing header described by bp. On success the return value is zero or greater. The return value of the function is -1 upon an error.7.6. inet6_rth_getaddr
struct in6_addr *inet6_rth_getaddr(const void *bp, int index); This function returns a pointer to the IPv6 address specified by index (which must have a value between 0 and one less than the value returned by inet6_rth_segments()) in the Routing header described by bp. An application should first call inet6_rth_segments() to obtain the number of segments in the Routing header. Upon an error the return value of the function is NULL.
8. Hop-By-Hop Options
A variable number of Hop-by-Hop options can appear in a single Hop- by-Hop options header. Each option in the header is TLV-encoded with a type, length, and value. This IPv6 API defines seven functions that the application calls to build and examine a Hop-by_Hop options header, and the ability to use sticky options or ancillary data to communicate this information between the application and the kernel. This uses the IPV6_HOPOPTS for a Hop-by-Hop options header. Today several Hop-by-Hop options are defined for IPv6. Two pad options, Pad1 and PadN, are for alignment purposes and are automatically inserted by the inet6_opt_XXX() routines and ignored by the inet6_opt_XXX() routines on the receive side. This section of the API is therefore defined for other (and future) Hop-by-Hop options that an application may need to specify and receive. Four functions build an options header: inet6_opt_init() - initialize buffer data for options header inet6_opt_append() - add one TLV option to the options header inet6_opt_finish() - finish adding TLV options to the options header inet6_opt_set_val() - add one component of the option content to the option Three functions deal with a returned options header: inet6_opt_next() - extract the next option from the options header inet6_opt_find() - extract an option of a specified type from the header inet6_opt_get_val() - retrieve one component of the option content Individual Hop-by-Hop options (and Destination options, which are described in Section 9 and are very similar to the Hop-by-Hop options) may have specific alignment requirements. For example, the 4-byte Jumbo Payload length should appear on a 4-byte boundary, and IPv6 addresses are normally aligned on an 8-byte boundary. These requirements and the terminology used with these options are discussed in Section 4.2 and Appendix B of [RFC-2460]. The alignment of first byte of each option is specified by two values, called x and y, written as "xn + y". This states that the option must appear at an integer multiple of x bytes from the beginning of the options header (x can have the values 1, 2, 4, or 8), plus y bytes (y can have a value between 0 and 7, inclusive). The Pad1 and PadN options are inserted as needed to maintain the required alignment. The
functions below need to know the alignment of the end of the option (which is always in the form "xn," where x can have the values 1, 2, 4, or 8) and the total size of the data portion of the option. These are passed as the "align" and "len" arguments to inet6_opt_append(). Multiple Hop-by-Hop options must be specified by the application by placing them in a single extension header. Finally, we note that use of some Hop-by-Hop options or some Destination options, might require special privilege. That is, normal applications (without special privilege) might be forbidden from setting certain options in outgoing packets, and might never see certain options in received packets.8.1. Receiving Hop-by-Hop Options
To receive a Hop-by-Hop options header the application must enable the IPV6_RECVHOPOPTS socket option: int on = 1; setsockopt(fd, IPPROTO_IPV6, IPV6_RECVHOPOPTS, &on, sizeof(on)); When using ancillary data a Hop-by-hop options header is passed between the application and the kernel as follows: The cmsg_level member will be IPPROTO_IPV6 and the cmsg_type member will be IPV6_HOPOPTS. These options are then processed by calling the inet6_opt_next(), inet6_opt_find(), and inet6_opt_get_val() functions, described in Section 10.8.2. Sending Hop-by-Hop Options
To send a Hop-by-Hop options header, the application specifies the header either as ancillary data in a call to sendmsg() or using setsockopt(). The application can remove any sticky Hop-by-Hop options header by calling setsockopt() for IPV6_HOPOPTS with a zero option length. All the Hop-by-Hop options must be specified by a single ancillary data object. The cmsg_level member is set to IPPROTO_IPV6 and the cmsg_type member is set to IPV6_HOPOPTS. The option is normally constructed using the inet6_opt_init(), inet6_opt_append(), inet6_opt_finish(), and inet6_opt_set_val() functions, described in Section 10. Additional errors may be possible from sendmsg() and setsockopt() if the specified option is in error.
9. Destination Options
A variable number of Destination options can appear in one or more Destination options headers. As defined in [RFC-2460], a Destination options header appearing before a Routing header is processed by the first destination plus any subsequent destinations specified in the Routing header, while a Destination options header that is not followed by a Routing header is processed only by the final destination. As with the Hop-by-Hop options, each option in a Destination options header is TLV-encoded with a type, length, and value.9.1. Receiving Destination Options
To receive Destination options header the application must enable the IPV6_RECVDSTOPTS socket option: int on = 1; setsockopt(fd, IPPROTO_IPV6, IPV6_RECVDSTOPTS, &on, sizeof(on)); Each Destination options header is returned as one ancillary data object described by a cmsghdr structure with cmsg_level set to IPPROTO_IPV6 and cmsg_type set to IPV6_DSTOPTS. These options are then processed by calling the inet6_opt_next(), inet6_opt_find(), and inet6_opt_get_value() functions.9.2. Sending Destination Options
To send a Destination options header, the application specifies it either as ancillary data in a call to sendmsg() or using setsockopt(). The application can remove any sticky Destination options header by calling setsockopt() for IPV6_RTHDRDSTOPTS/IPV6_DSTOPTS with a zero option length. This API assumes the ordering about extension headers as described in [RFC-2460]. Thus, one set of Destination options can only appear before a Routing header, and one set can only appear after a Routing header (or in a packet with no Routing header). Each set can consist of one or more options but each set is a single extension header. Today all destination options that an application may want to specify can be put after (or without) a Routing header. Thus, applications should usually need IPV6_DSTOPTS only and should avoid using IPV6_RTHDRDSTOPTS whenever possible.
When using ancillary data a Destination options header is passed between the application and the kernel as follows: The set preceding a Routing header are specified with the cmsg_level member set to IPPROTO_IPV6 and the cmsg_type member set to IPV6_RTHDRDSTOPTS. Any setsockopt or ancillary data for IPV6_RTHDRDSTOPTS is silently ignored when sending packets unless a Routing header is also specified. Note that the "Routing header" here means the one specified by this API. Even when the kernel inserts a routing header in its internal routine (e.g., in a mobile IPv6 stack), the Destination options header specified by IPV6_RTHDRDSTOPTS will still be ignored unless the application explicitly specifies its own Routing header. The set of Destination options after a Routing header, which are also used when no Routing header is present, are specified with the cmsg_level member is set to IPPROTO_IPV6 and the cmsg_type member is set to IPV6_DSTOPTS. The Destination options are normally constructed using the inet6_opt_init(), inet6_opt_append(), inet6_opt_finish(), and inet6_opt_set_val() functions, described in Section 10. Additional errors may be possible from sendmsg() and setsockopt() if the specified option is in error.10. Hop-by-Hop and Destination Options Processing
Building and parsing the Hop-by-Hop and Destination options is complicated for the reasons given earlier. We therefore define a set of functions to help the application. These functions assume the formatting rules specified in Appendix B in [RFC-2460] i.e., that the largest field is placed last in the option. The function prototypes for these functions are defined as a result of including <netinet/in.h>. The first 3 functions (init, append, and finish) are used both to calculate the needed buffer size for the options, and to actually encode the options once the application has allocated a buffer for the header. In order to only calculate the size the application must pass a NULL extbuf and a zero extlen to those functions.
10.1. inet6_opt_init
int inet6_opt_init(void *extbuf, socklen_t extlen); This function returns the number of bytes needed for the empty extension header i.e., without any options. If extbuf is not NULL it also initializes the extension header to have the correct length field. In that case if the extlen value is not a positive (i.e., non-zero) multiple of 8 the function fails and returns -1. (Note: since the return value on success is based on a "constant" parameter, i.e., the empty extension header, an implementation may return a constant value. However, this specification does not require the value be constant, and leaves it as implementation dependent. The application should not assume a particular constant value as a successful return value of this function.)10.2. inet6_opt_append
int inet6_opt_append(void *extbuf, socklen_t extlen, int offset, uint8_t type, socklen_t len, uint_t align, void **databufp); Offset should be the length returned by inet6_opt_init() or a previous inet6_opt_append(). This function returns the updated total length taking into account adding an option with length 'len' and alignment 'align'. If extbuf is not NULL then, in addition to returning the length, the function inserts any needed pad option, initializes the option (setting the type and length fields) and returns a pointer to the location for the option content in databufp. If the option does not fit in the extension header buffer the function returns -1. Type is the 8-bit option type. Len is the length of the option data (i.e., excluding the option type and option length fields). Once inet6_opt_append() has been called the application can use the databuf directly, or use inet6_opt_set_val() to specify the content of the option. The option type must have a value from 2 to 255, inclusive. (0 and 1 are reserved for the Pad1 and PadN options, respectively.) The option data length must have a value between 0 and 255, inclusive, and is the length of the option data that follows. The align parameter must have a value of 1, 2, 4, or 8. The align value can not exceed the value of len.
10.3. inet6_opt_finish
int inet6_opt_finish(void *extbuf, socklen_t extlen, int offset); Offset should be the length returned by inet6_opt_init() or inet6_opt_append(). This function returns the updated total length taking into account the final padding of the extension header to make it a multiple of 8 bytes. If extbuf is not NULL the function also initializes the option by inserting a Pad1 or PadN option of the proper length. If the necessary pad does not fit in the extension header buffer the function returns -1.10.4. inet6_opt_set_val
int inet6_opt_set_val(void *databuf, int offset, void *val, socklen_t vallen); Databuf should be a pointer returned by inet6_opt_append(). This function inserts data items of various sizes in the data portion of the option. Val should point to the data to be inserted. Offset specifies where in the data portion of the option the value should be inserted; the first byte after the option type and length is accessed by specifying an offset of zero. The caller should ensure that each field is aligned on its natural boundaries as described in Appendix B of [RFC-2460], but the function must not rely on the caller's behavior. Even when the alignment requirement is not satisfied, inet6_opt_set_val should just copy the data as required. The function returns the offset for the next field (i.e., offset + vallen) which can be used when composing option content with multiple fields.10.5. inet6_opt_next
int inet6_opt_next(void *extbuf, socklen_t extlen, int offset, uint8_t *typep, socklen_t *lenp, void **databufp); This function parses received option extension headers returning the next option. Extbuf and extlen specifies the extension header. Offset should either be zero (for the first option) or the length returned by a previous call to inet6_opt_next() or inet6_opt_find(). It specifies the position where to continue scanning the extension buffer. The next option is returned by updating typep, lenp, and
databufp. Typep stores the option type, lenp stores the length of the option data (i.e., excluding the option type and option length fields), and databufp points the data field of the option. This function returns the updated "previous" length computed by advancing past the option that was returned. This returned "previous" length can then be passed to subsequent calls to inet6_opt_next(). This function does not return any PAD1 or PADN options. When there are no more options or if the option extension header is malformed the return value is -1.10.6. inet6_opt_find
int inet6_opt_find(void *extbuf, socklen_t extlen, int offset, uint8_t type, socklen_t *lenp, void **databufp); This function is similar to the previously described inet6_opt_next() function, except this function lets the caller specify the option type to be searched for, instead of always returning the next option in the extension header. If an option of the specified type is located, the function returns the updated "previous" total length computed by advancing past the option that was returned and past any options that didn't match the type. This returned "previous" length can then be passed to subsequent calls to inet6_opt_find() for finding the next occurrence of the same option type. If an option of the specified type is not located, the return value is -1. If the option extension header is malformed, the return value is -1.10.7. inet6_opt_get_val
int inet6_opt_get_val(void *databuf, int offset, void *val, socklen_t vallen); Databuf should be a pointer returned by inet6_opt_next() or inet6_opt_find(). This function extracts data items of various sizes in the data portion of the option. Val should point to the destination for the extracted data. Offset specifies from where in the data portion of the option the value should be extracted; the first byte after the option type and length is accessed by specifying an offset of zero. It is expected that each field is aligned on its natural boundaries as described in Appendix B of [RFC-2460], but the function must not rely on the alignment.
The function returns the offset for the next field (i.e., offset + vallen) which can be used when extracting option content with multiple fields.11. Additional Advanced API Functions
11.1. Sending with the Minimum MTU
Unicast applications should usually let the kernel perform path MTU discovery [RFC-1981], as long as the kernel supports it, and should not care about the path MTU. Some applications, however, might not want to incur the overhead of path MTU discovery, especially if the applications only send a single datagram to a destination. A potential example is a DNS server. [RFC-1981] describes how path MTU discovery works for multicast destinations. From practice in using IPv4 multicast, however, many careless applications that send large multicast packets on the wire have caused implosion of ICMPv4 error messages. The situation can be worse when there is a filtering node that blocks the ICMPv4 messages. Though the filtering issue applies to unicast as well, the impact is much larger in the multicast cases. Thus, applications sending multicast traffic should explicitly enable path MTU discovery only when they understand that the benefit of possibly larger MTU usage outweighs the possible impact of MTU discovery for active sources across the delivery tree(s). This default behavior is based on the today's practice with IPv4 multicast and path MTU discovery. The behavior may change in the future once it is found that path MTU discovery effectively works with actual multicast applications and network configurations. This specification defines a mechanism to avoid path MTU discovery by sending at the minimum IPv6 MTU [RFC-2460]. If the packet is larger than the minimum MTU and this feature has been enabled the IP layer will fragment to the minimum MTU. To control the policy about path MTU discovery, applications can use the IPV6_USE_MIN_MTU socket option. As described above, the default policy should depend on whether the destination is unicast or multicast. For unicast destinations path MTU discovery should be performed by default. For multicast destinations path MTU discovery should be disabled by default. This option thus takes the following three types of integer arguments: -1: perform path MTU discovery for unicast destinations but do not perform it for multicast destinations. Packets to multicast destinations are therefore sent with the minimum MTU.
0: always perform path MTU discovery.
1: always disable path MTU discovery and send packets at the minimum
MTU.
The default value of this option is -1. Values other than -1, 0, and
1 are invalid, and an error EINVAL will be returned for those values.
As an example, if a unicast application intentionally wants to
disable path MTU discovery, it will add the following lines:
int on = 1;
setsockopt(fd, IPPROTO_IPV6, IPV6_USE_MIN_MTU, &on, sizeof(on));
Note that this API intentionally excludes the case where the
application wants to perform path MTU discovery for multicast but to
disable it for unicast. This is because such usage is not feasible
considering a scale of performance issues around whether to do path
MTU discovery or not. When path MTU discovery makes sense to a
destination but not to a different destination, regardless of whether
the destination is unicast or multicast, applications either need to
toggle the option between sending such packets on the same socket, or
use different sockets for the two classes of destinations.
This option can also be sent as ancillary data. In the cmsghdr
structure containing this ancillary data, the cmsg_level member will
be IPPROTO_IPV6, the cmsg_type member will be IPV6_USE_MIN_MTU, and
the first byte of cmsg_data[] will be the first byte of the integer.
11.2. Sending without Fragmentation
In order to provide for easy porting of existing UDP and raw socket
applications IPv6 implementations will, when originating packets,
automatically insert a fragment header in the packet if the packet is
too big for the path MTU.
Some applications might not want this behavior. An example is
traceroute which might want to discover the actual path MTU.
This specification defines a mechanism to turn off the automatic
inserting of a fragment header for UDP and raw sockets. This can be
enabled using the IPV6_DONTFRAG socket option.
int on = 1;
setsockopt(fd, IPPROTO_IPV6, IPV6_DONTFRAG, &on, sizeof(on));
By default, this socket option is disabled. Setting the value to 0 also disables the option i.e., reverts to the default behavior of automatic inserting. This option can also be sent as ancillary data. In the cmsghdr structure containing this ancillary data, the cmsg_level member will be IPPROTO_IPV6, the cmsg_type member will be IPV6_DONTFRAG, and the first byte of cmsg_data[] will be the first byte of the integer. This API only specifies the use of this option for UDP and raw sockets, and does not define the usage for TCP sockets. When the data size is larger than the MTU of the outgoing interface, the packet will be discarded. Applications can know the result by enabling the IPV6_RECVPATHMTU option described below and receiving the corresponding ancillary data items. An additional error EMSGSIZE may also be returned in some implementations. Note, however, that some other implementations might not be able to return this additional error when sending a message.11.3. Path MTU Discovery and UDP
UDP and raw socket applications need to be able to determine the "maximum send transport-message size" (Section 5.1 of [RFC-1981]) to a given destination so that those applications can participate in path MTU discovery. This lets those applications send smaller datagrams to the destination, avoiding fragmentation. This is accomplished using a new ancillary data item (IPV6_PATHMTU) which is delivered to recvmsg() without any actual data. The application can enable the receipt of IPV6_PATHMTU ancillary data items by setting the IPV6_RECVPATHMTU socket option. int on = 1; setsockopt(fd, IPPROTO_IPV6, IPV6_RECVPATHMTU, &on, sizeof(on)); By default, this socket option is disabled. Setting the value to 0 also disables the option. This API only specifies the use of this option for UDP and raw sockets, and does not define the usage for TCP sockets. When the application is sending packets too big for the path MTU recvmsg() will return zero (indicating no data) but there will be a cmsghdr with cmsg_type set to IPV6_PATHMTU, and cmsg_len will indicate that cmsg_data is sizeof(struct ip6_mtuinfo) bytes long. This can happen when the sending node receives a corresponding ICMPv6 packet too big error, or when the packet is sent from a socket with the IPV6_DONTFRAG option being on and the packet size is larger than the MTU of the outgoing interface. This indication is considered as an ancillary data item for a separate (empty) message. Thus, when
there are buffered messages (i.e., messages that the application has
not received yet) on the socket the application will first receive
the buffered messages and then receive the indication.
The first byte of cmsg_data[] will point to a struct ip6_mtuinfo
carrying the path MTU to use together with the IPv6 destination
address.
struct ip6_mtuinfo {
struct sockaddr_in6 ip6m_addr; /* dst address including
zone ID */
uint32_t ip6m_mtu; /* path MTU in host byte order */
};
This cmsghdr will be passed to every socket that sets the
IPV6_RECVPATHMTU socket option, even if the socket is non-connected.
Note that this also means an application that sets the option may
receive an IPV6_MTU ancillary data item for each ICMP too big error
the node receives, including such ICMP errors caused by other
applications on the node. Thus, an application that wants to perform
the path MTU discovery by itself needs to keep history of
destinations that it has actually sent to and to compare the address
returned in the ip6_mtuinfo structure to the history. An
implementation may choose not to delivery data to a connected socket
that has a foreign address that is different than the address
specified in the ip6m_addr structure.
When an application sends a packet with a routing header, the final
destination stored in the ip6m_addr member does not necessarily
contain complete information of the entire path.
11.4. Determining the Current Path MTU
Some applications might need to determine the current path MTU e.g.,
applications using IPV6_RECVPATHMTU might want to pick a good
starting value.
This specification defines a get-only socket option to retrieve the
current path MTU value for the destination of a given connected
socket. If the IP layer does not have a cached path MTU value it
will return the interface MTU for the interface that will be used
when sending to the destination address.
This information is retrieved using the IPV6_PATHMTU socket option.
This option takes a pointer to the ip6_mtuinfo structure as the
fourth argument, and the size of the structure should be passed as a
value-result parameter in the fifth argument.
struct ip6_mtuinfo mtuinfo;
socklen_t infolen = sizeof(mtuinfo);
getsockopt(fd, IPPROTO_IPV6, IPV6_PATHMTU, &mtuinfo, &infolen);
When the call succeeds, the path MTU value is stored in the ip6m_mtu
member of the ip6_mtuinfo structure. Since the socket is connected,
the ip6m_addr member is meaningless and should not be referred to by
the application.
This option can only be used for a connected socket, because a non-
connected socket does not have the information of the destination and
there is no way to pass the destination via getsockopt(). When
getsockopt() for this option is issued on a non-connected socket, the
call will fail. Despite this limitation, this option is still useful
from a practical point of view, because applications that care about
the path MTU tend to send a lot of packets to a single destination
and to connect the socket to the destination for performance reasons.
If the application needs to get the MTU value in a more generic way,
it should use a more generic interface, such as routing sockets
[TCPIPILLUST].
(page 48 continued on part 3)
