blob: aa34e94735101d44fa07bbce6d28faa5e14dd78b [file] [log] [blame]
\def\TITLE{IPv6 Flow Labels}
\input preamble
\Large\bf IPv6 Flow Labels in Linux-2.2.
{ \large Alexey~N.~Kuznetsov } \\
\em Institute for Nuclear Research, Moscow \\
\verb|| \\
\rm April 11, 1999
Every IPv6 packet carries 28 bits of flow information. RFC2460 splits
these bits to two fields: 8 bits of traffic class (or DS field, if you
prefer this term) and 20 bits of flow label. Currently there exist
no well-defined API to manage IPv6 flow information. In this document
I describe an attempt to design the API for Linux-2.2 IPv6 stack.
\vskip 1mm
The API must solve the following tasks:
\item To allow user to set traffic class bits.
\item To allow user to read traffic class bits of received packets.
This feature is not so useful as the first one, however it will be
necessary f.e.\ to implement ECN [RFC2481] for datagram oriented services
or to implement receiver side of SRP or another end-to-end protocol
using traffic class bits.
\item To assign flow labels to packets sent by user.
\item To get flow labels of received packets. I do not know
any applications of this feature, but it is possible that receiver will
want to use flow labels to distinguish sub-flows.
\item To allocate flow labels in the way, compliant to RFC2460. Namely:
Flow labels must be uniformly distributed (pseudo-)random numbers,
so that any subset of 20 bits can be used as hash key.
Flows with coinciding source address and flow label must have identical
destination address and not-fragmentable extensions headers (i.e.\
hop by hop options and all the headers up to and including routing header,
if it is present.)
There is a hole in specs: some hop-by-hop options can be
defined only on per-packet base (f.e.\ jumbo payload option).
Essentially, it means that such options cannot present in packets
with flow labels.
NB notes here and below reflect only my personal opinion,
they should be read with smile or should not be read at all :-).
Flow labels have finite lifetime and source is not allowed to reuse
flow label for another flow within the maximal lifetime has expired,
so that intermediate nodes will be able to invalidate flow state before
the label is taken over by another flow.
Flow state, including lifetime, is propagated along datagram path
by some application specific methods
(f.e.\ in RSVP PATH messages or in some hop-by-hop option).
\section{Sending/receiving flow information.}
It was proposed (Where? I do not remember any explicit statement)
to solve the first four tasks using
\verb|sin6_flowinfo| field added to \verb|struct| \verb|sockaddr_in6|
(see RFC2553).
This method is difficult to consider as reasonable, because it
puts additional overhead to all the services, despite of only
very small subset of them (none, to be more exact) really use it.
It contradicts both to IETF spirit and the letter. Before RFC2553
one justification existed, IPv6 address alignment left 4 byte
hole in \verb|sockaddr_in6| in any case. Now it has no justification.
We have two problems with this method. The first one is common for all OSes:
if \verb|recvmsg()| initializes \verb|sin6_flowinfo| to flow info
of received packet, we loose one very important property of BSD socket API,
namely, we are not allowed to use received address for reply directly
and have to mangle it, even if we are not interested in flowinfo subtleties.
RFC2553 adds new requirement: to clear \verb|sin6_flowinfo|.
Certainly, it is not solution but rather attempt to force applications
to make unnecessary work. Well, as usually, one mistake in design
is followed by attempts to patch the hole and more mistakes...
Another problem is Linux specific. Historically Linux IPv6 did not
initialize \verb|sin6_flowinfo| at all, so that, if kernel does not
support flow labels, this field is not zero, but a random number.
Some applications also did not take care about it.
Following RFC2553 such applications can be considered as broken,
but I still think that they are right: clearing all the address
before filling known fields is robust but stupid solution.
Useless wasting CPU cycles and
memory bandwidth is not a good idea. Such patches are acceptable
as temporary hacks, but not as standard of the future.
By default Linux IPv6 does not read \verb|sin6_flowinfo| field
assuming that common applications are not obliged to initialize it
and are permitted to consider it as pure alignment padding.
In order to tell kernel that application
is aware of this field, it is necessary to set socket option
int on = 1;
setsockopt(sock, SOL_IPV6, IPV6_FLOWINFO_SEND,
(void*)&on, sizeof(on));
Linux kernel never fills \verb|sin6_flowinfo| field, when passing
message to user space, though the kernels which support flow labels
initialize it to zero. If user wants to get received flowinfo, he
will set option \verb|IPV6_FLOWINFO| and after this he will receive
flowinfo as ancillary data object of type \verb|IPV6_FLOWINFO|
(cf.\ RFC2292).
int on = 1;
setsockopt(sock, SOL_IPV6, IPV6_FLOWINFO, (void*)&on, sizeof(on));
Flowinfo received and latched by a connected TCP socket also may be fetched
with \verb|getsockopt()| \verb|IPV6_PKTOPTIONS| together with
another optional information.
Besides that, in the spirit of RFC2292 the option \verb|IPV6_FLOWINFO|
may be used as alternative way to send flowinfo with \verb|sendmsg()| or
to latch it with \verb|IPV6_PKTOPTIONS|.
\paragraph{Note about IPv6 options and destination address.}
\addcontentsline{toc}{subsection}{IPv6 options and destination address}
If \verb|sin6_flowinfo| does contain not zero flow label,
destination address in \verb|sin6_addr| and non-fragmentable
extension headers are ignored. Instead, kernel uses the values
cached at flow setup (see below). However, for connected sockets
kernel prefers the values set at connection time.
After setting socket option \verb|IPV6_FLOWINFO|
flowlabel and DS field are received as ancillary data object
of type \verb|IPV6_FLOWINFO| and level \verb|SOL_IPV6|.
In the cases when it is convenient to use \verb|recvfrom(2)|,
it is possible to replace library variant with your own one,
sort of:
#include <sys/socket.h>
#include <netinet/in6.h>
size_t recvfrom(int fd, char *buf, size_t len, int flags,
struct sockaddr *addr, int *addrlen)
size_t cc;
char cbuf[128];
struct cmsghdr *c;
struct iovec iov = { buf, len };
struct msghdr msg = { addr, *addrlen,
&iov, 1,
cbuf, sizeof(cbuf),
0 };
cc = recvmsg(fd, &msg, flags);
if (cc < 0)
return cc;
((struct sockaddr_in6*)addr)->sin6_flowinfo = 0;
*addrlen = msg.msg_namelen;
for (c=CMSG_FIRSTHDR(&msg); c; c = CMSG_NEXTHDR(&msg, c)) {
if (c->cmsg_level != SOL_IPV6 ||
c->cmsg_type != IPV6_FLOWINFO)
((struct sockaddr_in6*)addr)->sin6_flowinfo = *(__u32*)CMSG_DATA(c);
return cc;
\section{Flow label management.}
Requirements of RFC2460 are pretty tough. Particularly, lifetimes
longer than boot time require to store allocated labels at stable
storage, so that the full implementation necessarily includes user space flow
label manager. There are at least three different approaches:
\item {\bf ``Cooperative''. } We could leave flow label allocation wholly
to user space. When user needs label he requests manager directly. The approach
is valid, but as any ``cooperative'' approach it suffers of security problems.
One idea is to disallow not privileged user to allocate flow
labels, but instead to pass the socket to manager via \verb|SCM_RIGHTS|
control message, so that it will allocate label and assign it to socket
itself. Hmm... the idea is interesting.
\item {\bf ``Indirect''.} Kernel redirects requests to user level daemon
and does not install label until the daemon acknowledged the request.
The approach is the most promising, it is especially pleasant to recognize
parallel with IPsec API [RFC2367,Craig]. Actually, it may share API with
\item {\bf ``Stupid''.} To allocate labels in kernel space. It is the simplest
method, but it suffers of two serious flaws: the first,
we cannot lease labels with lifetimes longer than boot time, the second,
it is sensitive to DoS attacks. Kernel have to remember all the obsolete
labels until their expiration and malicious user may fastly eat all the
flow label space.
Certainly, I choose the most ``stupid'' method. It is the cheapest one
for implementor (i.e.\ me), and taking into account that flow labels
still have no serious applications it is not useful to work on more
advanced API, especially, taking into account that eventually we
will get it for no fee together with IPsec.
Socket option \verb|IPV6_FLOWLABEL_MGR| allows to
request flow label manager to allocate new flow label, to reuse
already allocated one or to delete old flow label.
Its argument is \verb|struct| \verb|in6_flowlabel_req|:
struct in6_flowlabel_req
struct in6_addr flr_dst;
__u32 flr_label;
__u8 flr_action;
__u8 flr_share;
__u16 flr_flags;
__u16 flr_expires;
__u16 flr_linger;
__u32 __flr_reserved;
/* Options in format of IPV6_PKTOPTIONS */
\item \verb|dst| is IPv6 destination address associated with the label.
\item \verb|label| is flow label value in network byte order. If it is zero,
kernel will allocate new pseudo-random number. Otherwise, kernel will try
to lease flow label ordered by user. In this case, it is user task to provide
necessary flow label randomness.
\item \verb|action| is requested operation. Currently, only three operations
are defined:
#define IPV6_FL_A_GET 0 /* Get flow label */
#define IPV6_FL_A_PUT 1 /* Release flow label */
#define IPV6_FL_A_RENEW 2 /* Update expire time */
\item \verb|flags| are optional modifiers. Currently
only \verb|IPV6_FL_A_GET| has modifiers:
#define IPV6_FL_F_CREATE 1 /* Allowed to create new label */
#define IPV6_FL_F_EXCL 2 /* Do not create new label */
\item \verb|share| defines who is allowed to reuse the same flow label.
#define IPV6_FL_S_NONE 0 /* Not defined */
#define IPV6_FL_S_EXCL 1 /* Label is private */
#define IPV6_FL_S_PROCESS 2 /* May be reused by this process */
#define IPV6_FL_S_USER 3 /* May be reused by this user */
#define IPV6_FL_S_ANY 255 /* Anyone may reuse it */
\item \verb|linger| is time in seconds. After the last user releases flow
label, it will not be reused with different destination and options at least
during this time. If \verb|share| is not \verb|IPV6_FL_S_EXCL| the label
still can be shared by another sockets. Current implementation does not allow
unprivileged user to set linger longer than 60 sec.
\item \verb|expires| is time in seconds. Flow label will be kept at least
for this time, but it will not be destroyed before user released it explicitly
or closed all the sockets using it. Current implementation does not allow
unprivileged user to set timeout longer than 60 sec. Proviledged applications
MAY set longer lifetimes, but in this case they MUST save allocated
labels at stable storage and restore them back after reboot before the first
application allocates new flow.
This structure is followed by optional extension headers associated
with this flow label in format of \verb|IPV6_PKTOPTIONS|. Only
\verb|IPV6_HOPOPTS|, \verb|IPV6_RTHDR| and, if \verb|IPV6_RTHDR| presents,
\verb|IPV6_DSTOPTS| are allowed.
The function \verb|get_flow_label| allocates
private flow label.
int get_flow_label(int fd, struct sockaddr_in6 *dst, __u32 fl)
int on = 1;
struct in6_flowlabel_req freq;
memset(&freq, 0, sizeof(freq));
freq.flr_label = htonl(fl);
freq.flr_action = IPV6_FL_A_GET;
freq.flr_flags = IPV6_FL_F_CREATE | IPV6_FL_F_EXCL;
freq.flr_share = IPV6_FL_S_EXCL;
memcpy(&freq.flr_dst, &dst->sin6_addr, 16);
if (setsockopt(fd, SOL_IPV6, IPV6_FLOWLABEL_MGR,
&freq, sizeof(freq)) == -1) {
perror ("can't lease flowlabel");
return -1;
dst->sin6_flowinfo |= freq.flr_label;
if (setsockopt(fd, SOL_IPV6, IPV6_FLOWINFO_SEND,
&on, sizeof(on)) == -1) {
perror ("can't send flowinfo");
freq.flr_action = IPV6_FL_A_PUT;
setsockopt(fd, SOL_IPV6, IPV6_FLOWLABEL_MGR,
&freq, sizeof(freq));
return -1;
return 0;
A bit more complicated example using routing header can be found
in \verb|ping6| utility (\verb|iputils| package). Linux rsvpd backend
contains an example of using operation \verb|IPV6_FL_A_RENEW|.
\paragraph{Listing flow labels.}
\addcontentsline{toc}{subsection}{Listing flow labels}
List of currently allocated
flow labels may be read from \verb|/proc/net/ip6_flowlabel|.
Label S Owner Users Linger Expires Dst Opt
A1BE5 1 0 0 6 3 3ffe2400000000010a0020fffe71fb30 0
\item \verb|Label| is hexadecimal flow label value.
\item \verb|S| is sharing style.
\item \verb|Owner| is ID of creator, it is zero, pid or uid, depending on
sharing style.
\item \verb|Users| is number of applications using the label now.
\item \verb|Linger| is \verb|linger| of this label in seconds.
\item \verb|Expires| is time until expiration of the label in seconds. It may
be negative, if the label is in use.
\item \verb|Dst| is IPv6 destination address.
\item \verb|Opt| is length of options, associated with the label. Option
data are not accessible.
\paragraph{Flow labels and RSVP.}
\addcontentsline{toc}{subsection}{Flow labels and RSVP}
RSVP daemon supports IPv6 flow labels
without any modifications to standard ISI RAPI. Sender must allocate
flow label, fill corresponding sender template and submit it to local rsvp
daemon. rsvpd will check the label and start to announce it in PATH
messages. Rsvpd on sender node will renew the flow label, so that it will not
be reused before path state expires and all the intermediate
routers and receiver purge flow state.
\verb|rtap| utility is modified to parse flow labels. F.e.\ if user allocated
flow label \verb|0xA1234|, he may write:
RTAP> sender 3ffe:2400::1/FL0xA1234 <Tspec>
Receiver makes reservation with command:
RTAP> reserve ff 3ffe:2400::1/FL0xA1234 <Flowspec>