Network Working Group R. R. Stewart INTERNET-DRAFT Cisco Q. Xie L Yarroll Motorola J. Wood K. Poon Sun Microsystems expires in six months March 2, 2001 SCTP Sockets Mapping Status of This Memo This document is an Internet-Draft and is in full conformance with all provisions of Section 10 of [RFC2026]. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF), its areas, and its working groups. Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts can be accessed at http://www.ietf.org/ietf/1id-abstracts.txt The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html. Abstract This document describes a mapping of the Stream Control Transmission Protocol [SCTP] into a sockets API. The benefits of this mapping include compatibility for TCP applications, access to new SCTP features and a consolidated error and event notification scheme. Table of Contents 1. Introduction 2. Conventions 2.1 Data Types 3. UDP-style Interface 3.1 Basic Operation 3.1.1 socket() - UDP Style Syntax 3.1.2 bind() - UDP Style Syntax 3.1.3 sendmsg() and recvmsg() - UDP Style Syntax 3.1.4 close() - UDP Style Syntax 3.2 Implicit Association Setup 3.3 Examples 4. TCP-style Interface 4.1 Basic Operation 4.1.1 socket() - TCP Style Syntax 4.1.2 bind() - TCP Style Syntax 4.1.3 listen() - TCP Style Syntax 4.1.4 accept() - TCP Style Syntax 4.1.5 connect() - TCP Style Syntax 4.1.6 close() - TCP Style Syntax 4.1.7 shutdown() - TCP Style Syntax 4.1.8 sendmsg() and recvmsg() - TCP Style Syntax 4.2 Examples 5. Data Structures 5.1 The msghdr and cmsghdr Structures 5.2 SCTP msg_control Structures 5.2.1 SCTP Initiation Structure 5.2.2 SCTP SNDRCV Structure 5.3 SCTP Notifications 5.3.1 SCTP Notification Structure 5.3.1.1 Communication notifications 5.3.1.2 Interface notifications 5.3.1.3 SCTP Communication error 5.3.1.4 Returned messages 6. Common Operations for Both Styles 6.1 send(), recv(), sendto(), recvfrom() 6.2 setsockopt(), getsockopt() 6.3 read() and write() 7. Socket Options 7.1 Read / Write Options 7.1.1 Retransmission Timeout Parameters (SCTP_RTOINFO) 7.1.2 Association Retransmission Parameter (SCTP_ASSOCRTXINFO) 7.1.3 Path Parameters (SCTP_PATHPARAMS) 7.1.4 Initialization Parameters (SCTP_INITMSG) 7.1.5 Change of Addresses (SCTP_ADD_ADDR/SCTP_DEL_ADDR) 7.1.6 SO_LINGER 7.2 Read-Only Options 7.2.1 Path Information (SCTP_PATHINFO) 7.2.2 Peer Endpoint's Set of Addresses (SCTP_PATHCOUNT, SCTP_ALLPATHS) 7.2.3 Association Status (SCTP_STATUS) 7.3. Ancillary Data Interest Options 8. New Interface 8.1 sctp_bindx() 8.2 Branched-off Association, sctp_peeloff() 9. Security Considerations 10. Authors' Addresses 11. References 1. Introduction The sockets API has provided a standard mapping of the Internet Protocol suite to many operating systems. Both TCP [TCP] and UDP [UDP] have benefited from this standard representation and access method across many diverse platforms. SCTP is a new protocol that provides many of the characteristics of TCP but also incorporates semantics more akin to UDP. This document defines a method to map the existing sockets API for use with SCTP, providing both a base for access to new features and compatibility so that most existing TCP applications can be migrated to SCTP with few (if any) changes. There are three basic design objectives: 1) Maintain consistency with existing sockets APIs: We define a sockets mapping for SCTP that is consistent with other sockets API protocol mappings (for instance, UDP, TCP, IPv4, and IPv6). 2) Support a UDP-style interface This set of semantics is similar to that defined for conntionless protocols, such as UDP. It is more efficient than a TCP-like connection-oriented interface in terms of exploring the new features of SCTP. Note that SCTP is connection-oriented in nature, and it does not support broadcast or multicast communications, as UDP does. 3) Support a TCP-style interface This interface supports the same basic semantics as sockets for connection-oriented protocols, such as TCP. The purpose of defining this interface is to allow existing applications built on connnection-oriented protocols be ported to use SCTP with very little effort, and developers familiar with that semantics can easily adapt to SCTP. Extensions will be added to this mapping to provide mechanisms to exploit new features of SCTP. Goals 2 and 3 are not compatible, so in this document we define two modes of mapping, namely the UDP-style mapping and the TCP-style mapping. These two modes share some common data structures and operations, but will require the use of two different programming models. A mechanism is defined to convert a UDP-style SCTP socket into a TCP-style socket. Some of the SCTP mechanisms cannot be adequately mapped to existing socket interface. In some cases, it is more desirable to have new interface instead of using exisitng socket calls. This document also describes those new interface. 2. Conventions 2.1 Data Types Whenever possible, data types from Draft 6.6 (March 1997) of POSIX 1003.1g are used: uintN_t means an unsigned integer of exactly N bits (e.g., uint16_t). We also assume the argument data types from 1003.1g when possible (e.g., the final argument to setsockopt() is a size_t value). Whenever buffer sizes are specified, the POSIX 1003.1 size_t data type is used. 3. UDP-style Interface The UDP-style interface has the following characteristics: A) Outbound association setup is implicit. B) Messages are delivered in complete messages (with one notable exception). C) New inbound associations are accepted automatically. 3.1 Basic Operation A typical server in this model uses the following socket calls in sequence to prepare an endpoint for servicing requests: 1. socket() 2. bind() 3. setsocketopt() 4. recvmsg() 5. sendmsg() 6. close() A typical client uses the following calls in sequence to setup an association with a server to request services: 1. socket() 2. sendmsg() 3. recvmsg() 4. close() In this model, by default, all the associations connected to the endpoint are represented with a single socket. If the server or client wishes to branch an existing association off to a separate socket, it is required to call sctp_peeloff() and in the parameter specifies one of the transport addresses of the association. The sctp_peeloff() call will return a new socket which can then be used with recv() and send() functions for message passing. See Section 8.2 for more on branched-off associations. Once an association is branched off to a separate socket, it becomes completely separated from the original socket. All subsequent control and data operations to that association must be done through the new socket. For example, the close operation on the original socket will not terminated any association that have been branched off to a different socket. We will discuss the UDP-style socket calls in more details in the following subsections. 3.1.1 socket() - UDP Style Syntax Applications use socket() to create a socket descriptor to represent an SCTP endpoint. The syntax is, sd = socket(PF_INET, SOCK_SEQPACKET, IPPROTO_SCTP); or, sd = socket(PF_INET6, SOCK_SEQPACKET, IPPROTO_SCTP); Here, SOCK_SEQPACKET indicates the creation of a UDP-style socket. The first form creates an endpoint which can use only IPv4 addresses, while, the second form creates an endpoint which can use both IPv6 and IPv4 mapped addresses. 3.1.2 bind() - UDP Style Syntax Applications use bind() to specify which local address the SCTP endpoint should associate itself with as the primary address. An SCTP endpoint can be associated with multiple addresses. To do this, sctp_bindx() is introduced in section 8.1 to help applications do the job of associating multiple addresses. Instead of calling bind(), an application can use sctp_bindx() to associate an SCTP endpoint with multiple addresses. These addresses associated with a socket are the eligible transport addresses for the endpoint to send and receive data. The endpoint will also present these addresses to its peers during the association initialization process, see [SCTP]. After calling bind() or sctp_bindx(), if the endpoint wishes to accept new assocations on the socket, it must enable the SCTP_ASSOC_CHANGE socket option (see section 5.3.1.1). Then the SCTP endpoint will accept all SCTP INIT requests passing the COMMUNICATION_UP notification to the endpoint upon reception of a valid associaition (i.e. the receipt of a valid COOKIE ECHO). The syntax of bind() is, ret = bind(int sd, struct sockaddr *addr, int addrlen); sd - the socket descriptor returned by socket(). addr - the address structure (struct sockaddr_in or struct sockaddr_in6 [RFC 2553]), addrlen - the size of the address structure. If sd is an IPv4 socket, the address passed must be an IPv4 address. If the sd is an IPv6 socket, the address passed can either be an IPv4 or an IPv6 address. Applications cannot call bind() multiple times to associate multiple addresses to an endpoint. After the first call to bind(), all subsequent call will return an error. If addr is specified as INADDR_ANY for an IPv4 or IPv6 socket, or as IN6ADDR_ANY for an IPv6 socket (normally used by server applications), the operating system will associates the endpoint with all the available local interfaces. If a bind() or sctp_bindx() is not called prior to the connect() call, the system picks an ephemeral port and will choose an address set equivalant to binding with INADDR_ANY and IN6ADDR_ANY for IPv4 and IPv6 socket respectively. One of those addresses will be the primary address for the association. This automatically enables the multihoming capability of SCTP. 3.1.3 sendmsg() and recvmsg() - UDP Style Syntax An application uses sendmsg() and recvmsg() call to transmit data to and receive data from its peer. ssize_t sendmsg(int socket, const struct msghdr *message, int flags); ssize_t recvmsg(int socket, struct msghdr *message, int flags); socket - the socket descriptor of the endpoint. message - pointer to the msghdr structure which contains a single user message and possibly some ancillary data. See Section 5 for complete description of the data structures. flags - flags sent or received with the user message, see Section 5 for complete description of the flags. As we will see in Section 5, along with the user data, the ancillary data field is used to carry the sctp_sndrcvinfo and/or the sctp_initmsg structures to perform various SCTP functions including specifying options for sending each user message. Those options, depending on whether sending or receiving, include stream number, stream sequence number, TOS, various flags, context and payload protocol Id, etc. When sending user data with sendmsg(), the msg_name field in msghdr structure will be filled with one of the addresses of the intended receiver. If there is no association existing between the sender and the intended receiver, the sender's SCTP stack will set up a new association and then send the user data (see Section 3.2 for more on implicit association setup). When receiving a user message with recvmsg(), the msg_name field in msghdr structure will be populated with the source IP address of the user data. The caller of recvmsg() can use this address information to determine to which association the received user message belongs. Note, if the socket is a branched-off socket that only represents one association (see Section 3.1), the msg_name field is not used when sending data (i.e., ignored by the SCTP stack). 3.1.4 close() - UDP Style Syntax Applications use close() to perform graceful shutdown (as described in Section 10.1 of [SCTP]) on ALL the associations currently represented by a UDP-style socket. The syntax is ret = close(int sd); sd - the socket descriptor of the associations to be closed. To gracefully shutdown a specific association represented by the UDP-style socket, an application should use the sendmsg() call, passing no user data, but including the appropriate flag in the ancillary data (see Section 5.2.2). If sd in the close() call is a branched-off socket representing only one association, the shutdown is performed on that association only. 3.2 Implicit Association Setup Once all bind() calls are complete on a UDP-style socket, the application can begin sending and receiving data using the sendmsg()/recvmsg() or sendto()/recvfrom() calls, without going through any explicit association setup procedures (i.e., no connect() calls required). Whenever sendmsg() or sendto() is called and the SCTP stack at the sender finds that there is no association existing between the sender and the intended receiver (identified by the address passed either in the msg_name field of msghdr structure in the sendmsg() call or the dest_addr field in the sendto() call), the SCTP stack will automatically setup an association to the intended receiver. Upon the successful association setup a COMMUNICATION_UP notification will be dispatched to the socket at both the sender and receiver side. This notification can be read by the recvmsg() system call (see Section 3.1.3). Note, if the SCTP stack at the sender side supports bundling, the first user message may be bundled with the COOKIE ECHO message [SCTP]. When the SCTP stack sets up a new association implicitly, it first consults the sctp_initmsg structure, which is passed along within the ancillary data in the sendmsg() call (see Section 5.2.1 for details of the data structures), for any special options to be used on the new association. If this information is not present in the sendmsg() call, or if the implicit association setup is triggered by a sendto() call, the default association initialization parameters will be used. These default association parameters may be set with respective setsockopt() calls or be left to the system defaults. Implicit association setup cannot be initiated by send()/recv() calls. 3.3 Examples [ To be filled in later ] 4. TCP-style Interface The goal of this model is to follow as closely as possible the current practice of using the sockets interface for a connection oriented protocol, such as TCP. This model enables existing applications using connection oriented protocols to be ported to SCTP with very little effort. Note that some new SCTP features and some new SCTP socket options can only be utilized through the use of sendmsg() and recvmsg() calls, see Section 4.1.8. 4.1 Basic Operation A typical server in TCP-style model uses the following system call sequence to prepare an SCTP endpoint for servicing requests: 1. socket() 2. bind() 3. listen() 4. accept() The accept() call blocks until a new assocation is set up. It returns with a new socket descriptor. The server then uses the new socket descriptor to communicate with the client, using recv() and send() calls to get requests and send back responses. Then it calls 5. close() to terminate the association. A typical client uses the following system call sequence to setup an association with a server to request services: 1. socket() 2. connect() After returning from connect(), the client uses send() and recv() calls to send out requests and receive responses from the server. The client calls 3. close() to terminate this association when done. 4.1.1 socket() - TCP Style Syntax [Editor's Note: [Should we include return code of these calls in the draft or should it be in a man page of different OSes? We may want to map special error code for SCTP. EMSGSIZE is used below. And if an app chooses not to receive event, we need to map some of those events to an error. We need to figure out the mapping.] Applications calls socket() to create a socket descriptor to represent an SCTP endpoint. The syntax is: sd = socket(PF_INET, SOCK_STREAM, IPPROTO_SCTP); or, sd = socket(PF_INET6, SOCK_STREAM, IPPROTO_SCTP); Here, SOCK_STREAM indicates the creation of a TCP-style socket. The first form creates an endpoint which can use only IPv4 addresses, while the second form creates an endpoint which can use both IPv6 and mapped IPv4 addresses. 4.1.2 bind() - TCP Style Syntax Applications use bind() to pass the primary address assoicated with an SCTP endpoint to the system. An SCTP endpoint can be associated with multiple addresses. To do this, sctp_bindx() is introduced in section 8.1 to help applications do the job of associating multiple addresses. Instead of calling bind(), an application can use sctp_bindx() to associae a SCTP endpoint with multiple addresses. These addresses associated with a socket are the eligible transport addresses for the endpoint to send and receive data. The endpoint will also present these addresses to its peers during the association initialization process, see [SCTP]. The syntax is: ret = bind(int sd, struct sockaddr *addr, int addrlen); sd - the socket descriptor returned by socket() call. addr - the address structure (either struct sockaddr_in or struct sockaddr_in6 defined in [RFC 2553]). addrlen - the size of the address structure. If sd is an IPv4 socket, the address passed must be an IPv4 address. Otherwise, i.e., the sd is an IPv6 socket, the address passed can either be an IPv4 or an IPv6 address. Applications cannot call bind() multiple times to associate multiple addresses to the endpoint. After the first call to bind(), all subsequent calls will return an error. If addr is specified as INADDR_ANY for an IPv4 or IPv6 socket, or as IN6ADDR_ANY for an IPv6 socket (normally used by server applications), the operating system will associate the endpoint with an optimal address set of the available interfaces. The completion of this bind() process does not ready the SCTP endpoint to accept inbound SCTP association requests. Until a listen() system call, described below, is performed on the socket, the SCTP endpoint will promptly reject an inbound SCTP INIT request with an SCTP ABORT and discard data received. 4.1.3 listen() - TCP Style Syntax Applications use listen() to ready the SCTP endpoint for accepting inbound associations. The syntax is: ret = listen(int sd, int backlog); sd - the socket descriptor of the SCTP endpoint. backlog - this specifies the max number of outstanding associations allowed in the socket's accept queue. These are the associations that have finished the four-way initiation handshake (see Section 5 of [SCTP]) and are in the ESTABLISHED state. 4.1.4 accept() - TCP Style Syntax Applications use accept() call to remove an established SCTP assocation from the accept queue of the endpoint. A new socket descriptor will be returned from accept() to represent the newly formed association. The syntax is: new_sd = accept(int sd, struct sockaddr *addr, socklen_t *addrlen); new_sd - the socket descriptor for the newly formed association. sd - the listening socket descriptor. addr - on return, will contain the primary address of the peer endpoint. addrlen - on return, will contain the size of addr. 4.1.5 connect() - TCP Style Syntax Applications use connect() to initiate an association to a peer. The syntax is ret = connect(int sd, const struct sockaddr *addr, int addrlen); sd - the socket descriptor of the endpoint. addr - the peer's address. addrlen - the size of the address. This operation corresponds to the ASSOCIATE primitive described in section 10.1 of [SCTP]. By default, the new association created has only one outbound stream. The SCTP_INITMSG option described in Section 7.1.4 should be used to change the number of outbound streams. If a bind() or sctp_bindx() is not called prior to the connect() call, the system picks an ephemeral port and will choose an address set equivalant to binding with INADDR_ANY and IN6ADDR_ANY for IPv4 and IPv6 socket respectively. One of those addresses will be the primary address for the association. This automatically enables the multihoming capability of SCTP. Note that SCTP allows data exchange, similar to T/TCP [RFC1644], during the association set up phase. If an application wants to do this, it cannot use connect() call. Instead, it should use sendto() or sendmsg() to initiate an assocation. If it uses sendto() and it wants to change initialization behavior, it needs to use the SCTP_INITMSG socket option before calling sendto(). Or it can use SCTP_INIT type sendmsg() to initiate an association without doing the setsockopt(). SCTP does not support half close semantics. This means that unlike T/TCP, MSG_EOF should not be set in the flags parameter when calling sendto() or sendmsg() when the call is used to initiate a connection. MSG_EOF is not an acceptable flag with SCTP socket. [ Editor's note: MSG_EOF can be used to replace the SHUTDOWN flag in the SCTP_SNDRCV message. ] 4.1.6 close() - TCP Style Syntax Applications use close() to gracefully close down an association. The syntax is: ret = close(int sd); sd - the socket descriptor of the association to be closed. This operation corresponds to the SHUTDOWN primitive described in [SCTP] section 10.1. 4.1.7 shutdown() - TCP Style Syntax The socket call shutdown() does not have any meaning with an SCTP socket because SCTP does not have a half closed semantics. Calling shutdown() on an SCTP socket will return an error. To perform the ABORT operation described in [SCTP] section 10.1, an application can use the socket option SO_LINGER. It is described in section 7.1.6. 4.1.8 sendmsg() and recvmsg() - TCP Style Syntax With a TCP-style socket, the application can also use sendmsg() and recvmsg() to transmit data to and receive data from its peer. The semantics is similar to those used in the UDP-style model (section 3.1.3), with the following differences: 1) When sending, the msg_name field in the msghdr is not used to specify the intended receiver, rather it is used to indicate a different peer address if the sender does not want to send the message over the primary address of the receiver. When receiving, if a message is not received from the primary address, the SCTP stack will fill in the msg_name field on return so that the application can retrieve the source address information of the received message. 2) An application must use close() to gracefully shutdown an assocication, or use SO_LINGER option with close() to abort an asssociation. It must not use the ABORT or SHUTDOWN flag in sendmsg(). The system returns an error if an application tries to do so. 4.2 Examples [To be filled in... ] 5. Data Structures We discuss in this section important data structures which are specifc to SCTP and are used with sendmsg() and recvmsg() calls to control SCTP endpoint operations and to access ancillary information. 5.1 The msghdr and cmsghdr Structures The msghdr structure used in the sendmsg() and recvmsg() calls, as well as the ancillary data carried in the structure, is the key for the application to set and get various control information from the SCTP endpoint. The msghdr and the related cmsghdr structures are defined and discussed in details in [RFC2292]. Here we will cite their definitions from [RFC2292]. The msghdr structure: struct msghdr { void *msg_name; /* ptr to socket address structure */ socklen_t msg_namelen; /* size of socket address structure */ struct iovec *msg_iov; /* scatter/gather array */ size_t msg_iovlen; /* # elements in msg_iov */ void *msg_control; /* ancillary data */ socklen_t msg_controllen; /* ancillary data buffer length */ int msg_flags; /* flags on received message */ }; The cmsghdr structure: struct cmsghdr { socklen_t cmsg_len; /* #bytes, including this header */ int cmsg_level; /* originating protocol */ int cmsg_type; /* protocol-specific type */ /* followed by unsigned char cmsg_data[]; */ }; In the msghdr structure, the usage of msg_name has been discussed in previous sections (see Sections 3.1.3 and 4.1.8). The scatter/gather buffers, or I/O vectors (pointed to by the msg_iov field) are treated as a single SCTP data chunk, rather than multiple chunks, for both sendmsg() and recvmsg(). The msg_flags are not used when sending a message with sendmsg(). Upon return from a recvmsg(2) call, the msg_flags normally contains MSG_IS_DATA. Section 5.2.2 defines the ancillary data associated with ordinary data messages. If recvmsg(2) is called with the MSG_ERRQUEUE flag set, msg_flags has MSG_IS_EVENT set to indicate that the cmsg_data field is valid and needs parsing. 5.2 SCTP msg_control Structures A key element of all SCTP-specific socket extensions is the use of ancillary data to specify and access SCTP-specific data via the struct msghdr's msg_control member used in sendmsg() and recvmsg(). Fine-grained control over initialization and sending parameters are handled with ancillary data. Ancillary data also provide critical information for notifications. Each ancillary data item is preceeded by a struct cmsghdr (see Section 5.1), which defines the function and purpose of the data contained in in the cmsg_data[] member. There are two kinds of ancillary data: initialization data, and, header information (SNDRCV). Initialization data sets protocol parameters for new associations. Section 5.2.1 provides more details. Header information can set or report parameters on individual messages in a stream. See section 5.2.2 for how to use SNDRCV ancillary data. By default on a TCP-style socket, SCTP will pass no ancillary data; on a UDP-style socket, SCTP will only pass SCTP_SNDRCV information. Specific ancillary data items can be enabled with socket options defined for SCTP; see section 7.3. Note in particular that for UDP-style sockets, new associations will not be accepted by default. See section 5.2.1 for more information. Note that all ancillary types are fixed length; see section 5.4 for further discussion on this. These data structures use struct sockaddr_storage (defined in [RFC2253]) as a portable, fixed length address format. Other protocols may also provide ancillary data to the socket layer consumer. These ancillary data items from other protocols may intermingle with SCTP data. For example, the IPv6 socket API definitions ([RFC2292] and [RFC2553]) define a number of ancillary data items. If a socket API consumer enables delivery of both SCTP and IPv6 ancillary data, they both may appear in the same msg_control buffer in any order. An application should be prepared to handle other types of ancillary data besides that passed by SCTP. The sockets application must provide a buffer large enough to accomodate all ancillary data provided via recvmsg(). If the buffer is not large enough, the ancillary data will be truncated and the msghdr's msg_flags will include MSG_CTRUNC. This API offers an alternate behaviour with the SCTP_NOTRUC socket option described in section 7.1.7. 5.2.1 SCTP Initiation Structure (SCTP_INIT) This cmsghdr structure provides information for initializing new SCTP associations with sendmsg(). The SCTP_INITMSG socket option uses this same data structure. This structure is not used for recvmsg(). cmsg_level cmsg_type cmsg_data[] ------------ ------------ ---------------------- IPPROTO_SCTP SCTP_INIT struct sctp_initmsg Here is the definition of the sctp_initmsg structure: struct sctp_initmsg { uint16_t sinit_num_ostreams; uint16_t sinit_max_instreams; uint16_t sinit_max_attempts; uint16_t sinit_max_init_timeo; }; sinit_num_ostreams: 16 bits (unsigned integer) This is an integer number representing the number of streams that the application wishes to be able to send to. This number is confirmed in the COMMUNICATION_UP notification and must be verified since it is a negotiated number with the remote endpoint. The default value of 0 indicates to use the endpoint default value. sinit_max_instreams: 16 bits (unsigned integer) This value represents the maximum number of inbound streams the application is prepared to support. This value is bounded by the actual implementation. In other words the user MAY be able to support more streams than the Operating System. In such a case, the Operating System limit overrides the value requested by the user. The default value of 0 indicates to use the endpoint's default value. sinit_max_attempts: 16 bits (unsigned integer) This integer specifies how many attempts the SCTP endpoint should make at resending the INIT. This value overrides the system SCTP 'Max.Init.Retransmits' value. The default value of 0 indicates to use the endpoint's default value. This is normally set to the system's default 'Max.Init.Retransmit' value. sinit_max_init_timeo: 16 bits (unsigned integer) This value represents the largest Time-Out or RTO value to use in attempting a INIT. Normally the 'RTO.Max' is used to limit the doubling of the RTO upon timeout. For the INIT message this value MAY override 'RTO.Max'. This value MUST NOT influence 'RTO.Max' during data transmission and is only used to bound the initial setup time. A default value of 0 indicates to use the endpoint's default value. This is normally set to the system's 'RTO.Max' value (60 seconds). 5.2.2 SCTP Header Information Structure (SCTP_SNDRCV) This cmsghdr structure specifies SCTP options for sendmsg() and describes SCTP header information about a received message through recvmsg(). cmsg_level cmsg_type cmsg_data[] ------------ ------------ ---------------------- IPPROTO_SCTP SCTP_SNDRCV struct sctp_sndrcvinfo Here is the defintion of sctp_sndrcvinfo: struct sctp_sndrcvinfo { uint16_t sinfo_stream; uint16_t sinfo_ssn; uint16_t sinfo_flags; uint32_t sinfo_ppid; uint32_t sinfo_context; uint8_t sinfo_dscp; sctp_assoc_t sinfo_assoc_id; }; sinfo_stream: 16 bits (unsigned integer) For recvmsg() this value contains the message's stream number. For sendmsg() this value holds the stream number that the application wishes to send this message to. If a sender specifies an invalid stream number an error indication is returned and the call fails. sinfo_ssn: 16 bits (unsigned integer) For recvmsg() this value contains the stream sequence number that the remote endpoint placed in the DATA chunk. For fragmented messages this is the same number for all deliveries of the message (if more than one recvmsg() is needed to read the message). The sendmsg() call will ignore this parameter. sinfo_ppid:32 bits (unsigned integer) This value in sendmsg() is an opaque unsigned value that is passed to the remote end in each user message. In recvmsg() this value is the same information that was passed by the upper layer in the peer application. Please note that byte order issues are NOT accounted for and this information is passed opaquely by the SCTP stack from one end to the other. sinfo_context:32 bits (unsigned integer) This value is an opaque 32 bit context datum that is used in the sendmsg() function. This value is passed back to the upper layer if a error occurs on the send of a message and is retrieved with each unsent message (Note: if a endpoint has done multple sends, all of which fail, multiple different sinfo_context values will be returned. One with each user data message). sinfo_flags: 16 bits (unsigned integer) This field may contain any of the following flags and is composed of a bitwise OR of these values. recvmsg() flags: MSG_EOR - This flag is present in the last piece of a message. MSG_UNORDERED - This flag is present when the message was sent non-ordered. sendmsg() flags: MSG_UNORDERED - This flag requests the un-ordered delivery of the message. If this flag is clear the datagram is considered an ordered send. MSG_ABORT - Setting this flag causes the specified association to abort by sending an ABORT message to the peer. MSG_SHUTDOWN - Setting this flag invokes the SCTP graceful shutdown procedures which assure that all data enqueued by both endpoints are successfully transmitted before closing the association. sinfo_dscp: 8 bits (unsigned integer) This field is available to change the DSCP value in the outbound IP packet. The default value of this field is 0. Note only 6 bits of this byte are used, the upper 2 bits are not part of the DS field. Any setting within these upper 2 bits is ignored. sinfo_assoc_id: sizeof (sctp_assoc_t) The association handle field, sinfo_assoc_id, holds the identifier for the association announced in the COMMUNICATION_UP notification. All notifications for a given association have the same identifier. A sctp_sndrcvinfo item always corresponds to the data in msg_iov. 5.3 SCTP Events and Notifications An SCTP application may need to understand and process events and errors that happen on the SCTP stack. These events include network status changes, association startups, remote operational errors and undeliverable messages. All of these can be essential for the application. When an SCTP application layer does a recvmsg(2) the message read is normally a data message from a peer endpoint. If the application wishes to have the SCTP stack deliver notifications of non-data events, it sets the appropriate socket option for the notifications it wants. See section 7.3 for these socket options. When a notification arrives, recvmsg() returns an error code of ENOTIFY. The user must then receive the message by calling recvmsg(2) with the MSG_ERRQUEUE flag set. The notification data is returned in msg_iov. This section details the notification structures. Every notification structure carries some common fields which provides general information. A recvmsg(2) call with the MSG_ERRQUEUE flag set returns AT MOST a single notification. It may return part of a notification if the msg_iov buffer is not large enough. If a single read is not sufficient, msg_flags will have MSG_EOR clear. The user MUST finish reading the notification before subsequent data can arrive. [ Editor's note: Alternative mechanism. Instead of returning an ENOTIFY error, the stack can return the notification in the msg_iov buffer and set the msg_flags to MSG_IS_NOTIFICATION. The app can use code like the following to handle notifications for TCP style socket. while ((n = recvmsg(fd, msg, flags)) > 0) { if (msg->msg_flags & MSG_IS_NOTIFICATION) { handle_event(fd, msg); } else { handle_data(fd, msg); } } ] 5.3.1 SCTP Notification Structure An SCTP application reads notifications by calling recvmsg(2) with the MSG_ERRQUEUE flag set. The notification structure is defined as the union of all notification types. union sctp_notification { uint16_t sn_type; /* Notification type. */ struct sctp_assoc_change; struct sctp_intf_change; struct sctp_remote_error; } sn_type: sizeof (uint16_t) The following table describes the SCTP notification and event types for the field sn_type. sn_type Description --------- --------------------------- SCTP_ASSOC_CHANGE This tag indicates that an association has either been opened or closed. Refer to 5.3.1.1 for details. SCTP_INTF_CHANGE This tag indicates that an address that is part of an existing association has experienced a change of state (e.g. a failure or return to service of the reachability of a endpoint via a specific transport address). Please see 5.3.1.2 for data structure details. SCTP_REMOTE_ERROR The attached error message is an Operational Error received from the remote peer. It includes the complete TLV sent by the remote endpoint. See section 5.3.1.3 for the detailed format. SCTP_SEND_FAILED The attached datagram could not be sent to the remote endpoint. This structure includes the original SCTP_SNDRCVINFO that was used in sending this message i.e. this structure uses the sctp_sndrecvinfo per section 5.3.1.4. 5.3.1.1 SCTP_ASSOC_CHANGE Communication notifications inform the ULP that an SCTP association has either begun or ended. The identifier for the new association resides in the sctp_notification structure in the cmsg_data ancillary data. The notification information has the following format: struct sctp_assoc_change { uint16_t sac_type; uint16_t sac_flags; uint32_t sac_length; sctp_assoc_t sac_assoc_id; uint16_t sac_state; uint16_t sac_error; uint16_t sac_outbound_streams; uint16_t sac_inbound_streams; }; sac_type: It should be SCTP_ASSOC_CHANGE. sac_flags: 16 bits (unsigned integer) This field may contain any of the following flags and is composed of a bitwise OR of these values. MSG_EOR - This flag is present in the last piece of a notification. sac_length: sizeof (uint32_t) This field is the total length of the notification data. The sn_length is the length for the WHOLE notification data, not just the part delivered with this ancillary data. sac_assoc_id: sizeof (sctp_assoc_t) The association id field, sac_assoc_id, holds the identifier for the association. All notifications for a given association have the same association identifier. For TCP style socket, this field is ignored. sac_state: 32 bits (signed integer) This field holds one of a number of values that communicate the event that happened to the association. They include: Event Name Description ---------------- --------------- COMMUNICATION_UP A new association is now ready and data may be exchanged with this peer. COMMUNICATION_LOST The association has failed. The association is now in the closed state. If SEND FAILED notifications are turned on, a COMMUNICATION_LOST is followed by a series of SCTP_SEND_FAILED events, one for each outstanding message. RESTART SCTP has detected that the peer has restarted. SHUTDOWN_COMPLETE The association has gracefully closed. CANT_START_ASSOC The association failed to setup. sac_error: 32 bits (signed integer) If the state was reached due to a error condition (e.g. COMMUNICATION_LOST) any relevant error information is available in this field. This corresponds to the protocol error codes defined in [SCTP]. sac_outbound_streams: 16 bits (unsigned integer) sac_inbound_streams: 16 bits (unsigned integer) The maximum number of streams allowed in each directtion are available in sac_outbound_streams and sac_inbound streams. An application must enable this ancillary data item with setsockopt (see section 7.3) before any new associations will be accepted on a UDP-style socket. This is the mechanism by which a server (or peer application that wishes to accept new associations) informs the SCTP stack to accept new associations on a socket. Clients (i.e. applications on which only active opens are made) can leave this ancillary data item off; they will then be assured that the only associations on the socket will be ones they actively initiated. Server or peer to peer sockets, on the other hand, will always accept new associations, so a well-written application using server UDP-style sockets must be prepared to handle new associations from unwanted peers. 5.3.1.2 SCTP_INTF_CHANGE When a destination address on a multi-homed peer encounters a change in reachability an interface details event is sent. The information has the following structure: struct sctp_intf_change{ uint16_t sic_type; uint16_t sic_flags; uint32_t sic_length; sctp_assoc_t sic_assoc_id; struct sockaddr_storage sic_aaddr; int sic_state; int sic_error; } sic_type: It should be SCTP_INTF_CHANGE. sic_flags: 16 bits (unsigned integer) This field may contain any of the following flags and is composed of a bitwise OR of these values. MSG_EOR - This flag is present in the last piece of a notification. sic_length: sizeof (uint32_t) This field is the total length of the notification data. The sn_length is the length for the WHOLE notification data, not just the part delivered with this ancillary data. sic_assoc_id: sizeof (sctp_assoc_t) The association id field, sic_assoc_id, holds the identifier for the association. All notifications for a given association have the same association identifier. For TCP style socket, this field is ignored. sic_aaddr: sizeof (struct sockaddr_storage) The affected address field, sic_aaddr, holds the remote peer's addresses of the association that is encountering the change of state. state: 32 bits (signed integer) This field holds one of a number of values that communicate the event that happened to the association. They include: Event Name Description ---------------- --------------- ADDRESS_AVAILABLE This address is now reachable. ADDRESS_UNREACHABLE The address specified can no longer be reached. Any data sent to this address is rerouted to an alternate until this address becomes reachable. error: 32 bits (signed integer) If the state was reached due to any error condition (e.g. ADDRESS_UNREACHABLE) any relevant error information is available in this field. 5.3.1.3 SCTP_REMOTE_ERROR A remote peer may send an Operational Error message to its peer. This message indicates a variety of error conditions on an association. Please refer to the SCTP specification [SCTP] section 3.3.10 for a complete list of possible error formats. SCTP error TLVs have the format: struct sctp_remote_error { uint16_t sre_type; uint16_t sre_flags; uint32_t sre_length; sctp_assoc_t sre_assoc_id; uint16_t sre_error; uint16_t sre_len; uint8_t sre_data[0]; }; sre_type: It should be SCTP_REMOTE_ERROR. sre_flags: 16 bits (unsigned integer) This field may contain any of the following flags and is composed of a bitwise OR of these values. MSG_EOR - This flag is present in the last piece of a notification. sre_length: sizeof (uint32_t) This field is the total length of the notification data. The sn_length is the length for the WHOLE notification data, not just the part delivered with this ancillary data. sre_assoc_id: sizeof (sctp_assoc_t) The association id field, sre_assoc_id, holds the identifier for the association. All notifications for a given association have the same association identifier. For TCP style socket, this field is ignored. sre_error: 16 bits (unsigned integer) This value represents one of the Operational Error causes defined in the SCTP specification, in network byte order. sre_len: 16 bits (unsigned integer) This value represents the length of the operational error payload in the msg_iov plus the size of sre_error and sre_len in network byte order. sre_data: variable This contains the payload of the operational error as defined in the SCTP specification [SCTP] section 3.3.10. 5.3.1.4 SCTP_SEND_FAILED If SCTP cannot deliver a message it may return the message as a notification. struct sctp_send_failed { uint16_t sf_type; uint16_t sf_flags; uint32_t sf_length; sctp_assoc_t sf_assoc_id; uint32_t sf_error; struct sctp_sndrcvinfo sf_info; uint8_t sf_data[0]; }; sf_type: It should be SCTP_SEND_FAILED. sf_flags: 16 bits (unsigned integer) This field may contain any of the following flags and is composed of a bitwise OR of these values. MSG_EOR - This flag is present in the last piece of a notification. sf_length: sizeof (uint32_t) This field is the total length of the notification data. The sn_length is the length for the WHOLE notification data, not just the part delivered with this ancillary data. s_assoc_id: sizeof (sctp_assoc_t) The association id field, sf_assoc_id, holds the identifier for the association. All notifications for a given association have the same association identifier. For TCP style socket, this field is ignored. sf_error: 16 bits (unsigned integer) This value represents the reason why the send fails. sf_info: sizeof (struct sctp_sndrcvinfo) The original send information associated with the unsent message. sf_data: variable The unsent message. 5.4 Ancillary Data Considerations and Semantics Programming with ancillary socket data contains some subtleties and pitfalls, which are discussed below. 5.4.1 Multiple Items and Ordering Multiple ancillary data items may be included in any call to sendmsg() or recvmsg(); these may include multiple SCTP or non-SCTP items, or both. The ordering of ancillary data items (either by SCTP or another protocol) is not significant and is implementation-dependant, so applications must not depend on any ordering. The one exception to this is that SCTP_ASSOC_CHANGE events announcing new associations must always preceed any other ancillary data items pertaining to the new assocition. SCTP_SNDRCV items must always correspond to the data in the msghdr's msg_iov member. An implementation may choose to bundle together multiple SCTP ancillary data items (for instance, a SCTP_ASSOC_CHANGE for a new association, followed by a SCTP_SNDRCV info corresponding to data bundled with the association initialization), or the implemantation can choose to deliver these events across multiple calls to recvmsg(). There can be only a single SCTP_SNDRCV info for each sendmsg() or recvmsg() call. Multiple instances of other events may appear in a single call. [ ed. note: should we restrict it to only one SCTP_INIT too? Hmm well one could imagine getting multiple associations at once but if we go with the "event socket" then I think each recv() should be a single notification ... ] 5.4.2 Accessing and Manipulating Ancillary Data Applications can infer the presence of data or ancillary data by examining the msg_iovlen and msg_controllen msghdr members, respectively. Implementations may have different padding requirements for ancillary data, so portable applications should make use of the macros CMSG_FIRSTHDR, CMSG_NXTHDR, CMSG_DATA, CMSG_SPACE, and CMSG_LEN. See [RFC2292] and your SCTP implementation's documentation for more information. Following is an example, from [RFC2292], demonstrating the use of these macros to access ancillary data: struct msghdr msg; struct cmsghdr *cmsgptr; /* fill in msg */ /* call recvmsg() */ for (cmsgptr = CMSG_FIRSTHDR(&msg); cmsgptr != NULL; cmsgptr = CMSG_NXTHDR(&msg, cmsgptr)) { if (cmsgptr->cmsg_level == ... && cmsgptr->cmsg_type == ... ) { u_char *ptr; ptr = CMSG_DATA(cmsgptr); /* process data pointed to by ptr */ } } 5.4.3 Control Message Buffer Sizing The information conveyed via SCTP_SNDRCV and SCTP_ASSOC_CHANGE events will often be fundamental to the correct and sane operation of the sockets application. This is particularly true of the UDP semantics, but also of the TCP semantics. For example, if an application needs to send and receive data on different SCTP streams, SCTP_SNDRCV events are indispensable. Similarly, the only way an application written to the UDP semantics can detect the addition of an assocation is via the SCTP_ASSOC_CHANGE event. Given that some ancillary data is critical, and that multiple ancillary data items may appear in any order, applications should be carefully written to always provide a large enough buffer to contain all possible ancillary data that can be presented by recvmsg(). If the buffer is too small, and crucial data is truncated, it may pose a fatal error condition. Thus it is essential that applications be able to deterministically calculate the maximum required buffer size to pass to recvmsg(). One constraint imposed on this specification that makes this possible is that all ancillary data definitions are of a fixed length. One way to calculate the maximum required buffer size might be to take the sum the sizes of all enabled ancillary data item structures, as calculated by CMSG_SPACE. For example, if we enabled SCTP_INIT, SCTP_SNDRCV_INFO, SCTP_ASSOC_CHANGE, and IPV6_RECVPKTINFO [RFC2292], we would calculate and allocate the buffer size as follows: size_t total; void *buf; total = CMSG_SPACE(sizeof (struct sctp_initmsg)) + CMSG_SPACE(sizeof (struct sctp_sndrcvinfo)) + CMSG_SPACE(sizeof (struct sctp_assoc_change)) + CMSG_SPACE(sizeof (struct in6_pktinfo)); buf = malloc(total); We could then use this buffer for msg_control on each call to recvmsg() and be assured that we would not lose any ancillary data to truncation. 6. Common Operations for Both Styles 6.1 send(), recv(), sendto(), recvfrom() Applications can use send() and sendto() to transmit data to the peer of an SCTP endpoint. recv() and recvfrom() can be used to receive data from the peer. In all calls listed below the socket descriptor passed to these calls must represent a single association. The syntax is: size = send(int sd, connst void *msg, size_t len, int flags); size = sendto(int sd, const void *msg, size_t len, int flags, const struct sockaddr *to, int tolen); size = recv(int sd, void *buf, size_t len, int flags); size = recvfrom(int sd, void *buf, size_t len, int flags, struct sockaddr *from, int *fromlen); sd - the socket descriptor of an SCTP endpoint. msg - the message to be sent. len - the size of the message or the size of buffer. to - one of the peer addresses of the association to be used to send the message. tolen - the size of the address. buf - the buffer to store a received message. from - the buffer to store the peer address used to send the received message. fromlen - the size of the receive buffer. flags - (described below). These calls give access to only basic SCTP protocol features. If either peer in the association uses multiple streams, or sends unordered data or data whose size exceeds its peer's RWND these calls will usually be inadequate, and may deliver the data in unpredictable ways. SCTP has the concept of multiple streams in one association. The above calls do not allow the caller to specify on which stream a message should be sent. The system uses stream 0 as the default stream for send() and sendto(). recv() and recvfrom() return data from any stream, but the caller can not distinguish the different streams. This may result in data seeming to arrive out of order. Similarly, if a data chunk is sent unordered, recv() and recvfrom() provide no indication. If a data chunk exceeds the size of a peer's receive window, the peer must attempt partial delivery of the data. It is possible for data from other chunks to get delivered in between partial deliveries. recvmsg() with the SCTP_DATAIOEVNT option on should always be used in these situations, since it provides the caller with all the information needed to distinguish deliveries of different chunks on various streams. SCTP is message based. The msg buffer above in send() and sendto() is considered to be a single message. This means that if the caller wants to send a message which is composed by several buffers, the caller needs to combine them before calling send() or sendto(). Or the caller can use sendmsg() to do that without combining them. In receiving, if the buffer supplied is not large enough to hold a complete messaage, the receive call returns a EMSGSIZE error. Refer to recvmsg() for a method to receive partial message. The flags parameter is formed by OR'ing one or more of the following: MSG_UNORDERED SCTP has a concept of unordered delivery. When sending, caller can use this flag to tell the system that this message can be delivered unordered. The caller must set this flag in all calls to transmit unorderd messages. Note, the send and recv calls, when used in the UDP-style model, may only be used with high bandwidth socket descriptors (see Section 3.3). 6.2 setsockopt(), getsockopt() Applications use setsockopt() and getsockopt() to set or retrieve socket options. Socket options are used to change the default behavior of sockets calls. They are described in Section 7. The syntax is: ret = getsockopt(int sd, int level, int optname, void *optval, size_t *optlen); ret = setsockopt(int sd, int level, int optname, const void *optval, size_t optlen); sd - the socket descript. level - set to IPPROTO_SCTP for all SCTP options. optname - the option name. optval - the buffer to store the value of the option. optlen - the size of the buffer. 6.3 read() and write() Applications can use read() and write() to send and receive data to and from peer. They have the same semantics as send() and recv() except that the flags parameter cannot be used. Note, these calls, when used in the UDP-style model, may only be used with high bandwidth socket descriptors (see Section 3.3). 7. Socket Options The following sub-section describes various SCTP level socket options that are common to both models. SCTP associations can be mutlihomed. Therefore, certain option parameters include a sockaddr_storage structure to select which peer address the option should be applied to. For the datagram model, an sctp_assoc_t structure (association ID) is used to identify the the association instance that the operation affects. So it must be set when using this model. For the connnection oriented model and high bandwidth datagram sockets (see section 3.3) this association ID arameter can be ignored. In the cases noted below where the parameter is ignored, an application can pass to the system a corresponding option structure similar to those described below but without the association ID parameter, which should be the last field of the option structure. This can make the option setting/getting operation more efficient. If an application does this, it should also specify an appropriate optlen value (i.e. sizeof (option parameter) - sizeof (struct sctp_assoc_t)). Note that socket or IP level options is set or retrieved per socket. This means that for datagram model, those options will be applied to all associations belonging to the socket. And for connection oriented model, those options will be applied to all peer addresses of the association controlled by the socket. Applications should be very careful in setting those options. 7.1 Read / Write Options 7.1.1 Retransmission Timeout Parameters (SCTP_RTOINFO) The protocol parameters used to initialize and bound retransmission timeout (RTO) are tunable. See [SCTP] for more information on how these parameters are used in RTO calculation. The peer address parameter is ignored for TCP style socket. The following structure is used to access and modify these parameters: struct sctp_rtoinfo { uint32_t srto_initial; uint32_t srto_max; uint32_t srto_min; sctp_assoc_t srto_assoc_id; }; srto_initial - This contains the initial RTO value. srto_max and srto_min - These contain the maximum and minumum bounds for all RTOs. srto_assoc_id - (UDP style socket) This is filled in the application, and identifies the association for this query. All parameters are time values, in milliseconds. A value of 0, when modifying the parameters, indicates that the current value should not be changed. To access or modify these parameters, the application should call getsockopt or setsockopt() respectively with the option name SCTP_RTOINFO. 7.1.2 Association Retransmission Parameter (SCTP_ASSOCRTXINFO) The protocol parameter used to set the number of retransmissions sent before an association is considered unreachable is tunable. See [SCTP] for more information on how this parameter is used. The peer address parameter is ignored for TCP style socket. The following structure is used to access and modify this parameters: struct sctp_assocparams { uint16_t sasoc_asocmaxrxt; sctp_assoc_t sasoc_assoc_id; }; sasoc_asocmaxrxt - This contains the maximum retransmission attempts to make for the association. sasoc_assoc_id - (UDP style socket) This is filled in the application, and identifies the association for this query. To access or modify these parameters, the application should call gesockopt or setsockopt() respectively with the option name SCTP_ASSOCRTXINFO. The maximum number of retransmissions before an address is considered unreachable is also tunable, but is address-specific, so it is covered in a seperate option. If an application attempts to set the value of the association maximum retransmission parameter to less than the sum of all maximum retransmission parameters, setsockopt() shall return an error. The reason for this, from [SCTP] section 8.2: Note: When configuring the SCTP endpoint, the user should avoid having the value of 'Association.Max.Retrans' larger than the summation of the 'Path.Max.Retrans' of all the destination addresses for the remote endpoint. Otherwise, all the destination addresses may become inactive while the endpoint still considers the peer endpoint reachable. 7.1.3 Path Parameters (SCTP_PATHPARAMS) Applications can enable or disable heartbeats for any peer address of an association, modify an address's heartbeat interval, and adjust the address's maximum number of retransmissions sent before an address is considered unreachable. The following structure is used to access and modify an address's parameters: struct sctp_pathparams { struct sockaddr_storage spp_address; uint32_t spp_interval; uint16_t spp_pathmaxrxt; sctp_assoc_t spp_assoc_id; }; spp_address - This specifies which address is of interest. spp_interval - This contains the value of the heartbeat interval, in milliseconds. A value of 0, when modifying the parameter, specifies that the heartbeat on this address should be disabled. spp_pathmaxrxt - This contains the maximum number of retransmissions before this address shall be considered unreachable. spath_assoc_id - (UDP style socket) This is filled in the application, and identifies the association for this query. To access or modify these parameters, the application should call gesockopt or setsockopt() respectively with the option name SCTP_PATHPARAMS. 7.1.4 Initialization Parameters (SCTP_INITMSG) Applications can specify protocol parameters for the default association intialization. The structure used to access and modify these parameters is defined in section 3.1.1. The option name argument to setsockopt() and getsockopt() is SCTP_INITMSG. Setting initialization parameters is effective only on an unconnected socket (for the datagram model only future associations are effected by the change). 7.1.6 SO_LINGER An application using the TCP-style socket can use this option to perform the SCTP ABORT primitive. The linger option structure is: struct linger { int l_onoff; /* option on/off */ int l_linger; /* linger time */ }; To enable the option, set l_onoff to 1. If the l_linger value is set to 0, calling close() is the same as the ABORT primitive. If the value is set to a negative value, the setsockopt() call will return an error. If the value is set to a positive value linger_time, the close() can be blocked for at most linger_time ms. If the graceful shutdown phase does not finish during this period, close() will return but the graceful shutdown phase continues in the system. 7.1.7 SCTP_NOTRUC Turn off ancillary data truncation. If an application provides an ancillary data buffer which is too small to hold the necessary ancillary data, recvmsg() will return the ETOOSMALL error instread of truncating the data. Expects an integer boolean flag. [Most of the authors consider this too radical a departure from traditional sockets behaviour and warn that this flag is very likely to go away. The remaining author WILL implement this for a popular operating system...] 7.1.8 SCTP_NODELAY Turn off any Nagle-like algorithm. This means that packets are generally sent as soon as possible and no unnecessary delays are introduced, at the cost of more packets in the network. Expects an integer boolean flag. 7.1.9 SO_RCVBUF Sets receive buffer size. For SCTP TCP-style sockets, this controls the receiver window size. For UDP-style sockets, this controls the receiver window size for all associations bound to the socket descriptor used in the setsockopt() or getsockopt() call. Expects an integer boolean flag. 7.1.10 SO_SNDBUF Sets send buffer size. For SCTP TCP-style sockets, this controls the amount of data SCTP may have waiting in internal buffers to be sent. This option therefore bounds the maximum size of data that can be sent in a single send call. For UDP-style sockets, the effect is the same, except that it applies to all associations bound to the socket descriptor used in the setsockopt() or getsockopt() call. Expects an integer boolean flag. 7.2 Read-Only Options 7.2.1 Path Information (SCTP_PATHINFO) Applications can retrieve information about a specific peer address of an association, including its reachability state, congestion window, and retransmission timer values. This information is read-only, so only getsockopt() operates on this option. Calls to setsockopt() on this option returns an error. The following structure is used to access this information: struct sctp_pathinfo { struct sockaddr_storage spath_address; int32_t spath_state; uint32_t spath_cwnd; uint32_t spath_srtt; uint32_t spath_rto; sctp_assoc_t spath_assoc_id; }; spath_address - This is filled in the application, and contains the peer address of interest. On return from getsockopt(): spath_state - This contains the path's state (either SCTP_ACTIVE or SCTP_INACTIVE). spath_cwnd - This contains the path's current congestion window. spath_srtt - This contains the path's current smoothed round-trip time calculation in milliseconds. spath_rto - This contains the path's current retransmission timeout value in milliseconds. spath_assoc_id - (UDP style socket) This is filled in the application, and identifies the association for this query. To retrieve this information, use getsockopt() with the option name set to SCTP_PATHINFO. 7.2.2 Peer Endpoint's Set of Addresses (SCTP_PATHCOUNT, SCTP_ALLPATHS) Applications can retrieve the set of addresses that correspond to a peer endpoint. Since this set is variable length, two options are needed to retrieve the information: the first, SCTP_PATHCOUNT, takes the following structure as its argument to getsockopt(): struct sctp_pathcnt { uint32_t spthc_numaddrs; sctp_assoc_t spthc_assoc_id; }; spthc_numaddrs - If filled in upon return from this call this indicates the number of addresses associated with the peer. The application can then allocate a buffer large enough to hold all the peer's addresses, and call getsockopt() with SCTP_ALLPATHS. spthc_assoc_id - (UDP style socket) This is filled in the application, and identifies the association for this query. For the datagram model, the first address in the call to SCTP_ALLPATHS MUST be filled in with a valid address that identifies the association. The peer address parameter is ignored for TCP style socket. On return of getsockopt(SCTP_ALLPATHS), each address is represented as a struct sockaddr_storage. So if n is the number of peer addresses, the caller must allocate a buffer of size n * sizeof(struct sockaddr_storage). The application can retrieve information on each address by iterating through the returned list of addresses and calling getsockopt() with the SCTP_PATHINFO option name. This information is read-only. 7.2.3 Association Status (SCTP_STATUS) Applications can retrieve current status information about an association, including association state, peer receiver window size, number of unacked data chunks, and number of data chunks pending receipt. This information is read-only. The following structure is used to access this information: struct sctp_status { int32_t sstat_state; uint32_t sstat_rwnd; uint16_t sstat_unackdata; uint16_t sstat_penddata; struct sctp_pathinfo sstat_primary; sctp_assoc_t sstat_assoc_id; }; sstat_state - This contains the association's current state (states TBD). sstat_rwnd - This contains the association peer's current receiver window size. sstat_unackdata - This is the number of unacked data chunks. sstat_penddata - This is the number of data chunks pending receipt. sstat_primary - This is information on the current primary path. sstat_assoc_id - (UDP style socket) This holds the an identifier for the association. All notifications for a given association have the same association identifier. To access these status values, the application calls getsockopt() with the option name SCTP_STATUS. The sstat_assoc_id parameter is ignored for TCP style socket. 7.3. Ancillary Data Interest Options Applications can receive notifications of certain SCTP events and per-message information as ancillary data with recvmsg(). The following optional information is available to the application: 1. SCTP_RECVDATAIOEVNT: Per-message information (i.e. stream number, TSN, SSN, etc. described in section 5.2.2) 2. SCTP_RECVASSOCEVNT: (described in section 5.3.1) 3. SCTP_RECVPATHEVNT: (described in section 5.3.2) 4. SCTP_RECVSENDFAILEVNT: (described in section 5.3.4) 5. SCTP_RECVPEERERR: (described in section 5.3.3) To receive any ancillary data, first the application registers it's interest by calling setsockopt() to turn on the corresponding flag: int on = 1; setsockopt(fd, IPPROTO_SCTP, SCTP_RECVDATAIOEVNT, &on, sizeof(on)); setsockopt(fd, IPPROTO_SCTP, SCTP_RECVPATHEVNT, &on, sizeof(on)); setsockopt(fd, IPPROTO_SCTP, SCTP_RECVSENDFAILEVNT, &on, sizeof(on)); setsockopt(fd, IPPROTO_SCTP, SCTP_RECVPEERERR, &on, sizeof(on)); Note that for connectionless mode SCTP sockets, the caller of recvmsg() receives ancillary data for ALL associations bound to the file descriptor. For connection-oriented SCTP sockets, the caller receives ancillary data for only the single association bound to the file descriptor. By default the connection oriented socket has all options off. By default the datagram oriented socket has SCTP_REVCVDATAIOEVENT on and all other options off. The format of the data structures for each ancillary data item is given in section 5.2. 8. New Interfaces Depending on the system, the following interface can be implemented as system calls or library funtions. 8.1 sctp_bindx() The syntax of sctp_bindx() is, ret = sctp_bindx(int sd, struct sockaddr_storage *addrs, int addrcnt, int flags); If sd is an IPv4 socket, the addresses passed must be IPv4 addresses. If the sd is an IPv6 socket, the addresses passed can either be IPv4 or IPv6 addresses. A single address may be specified as INADDR_ANY or IN6ADDR_ANY, see section 3.1.2 for this usage. addrs is a pointer to an array of one or more socket addresses. Each address is contained in a struct sockaddr_storage, so each address is fixed length. The caller specifies the number of addresses in the array with addrcnt. On success, sctp_bindx() returns 0. On failure, sctp_bindx() returns -1, and sets errno to the appropriate error code. [ Editor's note: need to fill in all error code? ] For SCTP, the port given in each socket address must be the same, or sctp_bindx() will fail, setting errno to EINVAL . The flags parameter is formed from the bitwise OR of zero or more of the following currently defined flags: SCTP_BINDX_ADD_ADDR SCTP_BINDX_REM_ADDR SCTP_BIND_ADD_ADDR directs SCTP to add the given addresses to the association, and SCTP_BIND_REM_ADDR directs SCTP to remove the given addresses from the association. The two flags are mutually exclusive; if both are given, sctp_bindx() will fail with EINVAL. A caller may not remove all addresses from an association; sctp_bindx() will reject such an attempt with EINVAL. An application can use sctp_bindx(SCTP_BINDX_ADD_ADDR) to associate additional addresses with an endpoint after calling bind(). Or use sctp_bindx(SCTP_BINDX_REM_ADDR) to remove some addresses a listening socket is associated with so that no new association accepted will be associated with those addresses. SCTP_BIND_ADD_ADDR is defined as 0, so that it becomes the default behavior for sctp_bindx() when no flags are given. Adding and removing addresses from a connected association is optional functionality. Implementations that do not support this functionality should return EOPNOTSUPP. [ Editor's note: This does not work well with UDP-style socket because it does not allow changes of address on individual association controlled by a socket. No but I would claim that if you were grouping all the associations in a single fd, then you want a add_address to apply to all associations.. so I don't see it as a issue - R ] 8.2 Branched-off Association After an association is established on a UDP-style socket, the application may wish to branch off the association into a separate socket/file descriptor. This is particularly desirable when, for instance, the application wishes to have a number of sporadic message senders/receivers remain under the original UDP-style socket but branch off those associations carrying high volume data traffic into their own separate socket descriptors. The application uses sctp_peeloff() call to branch off an association into a separate socket (Note the semantics are somewhat changed from the traditional TCP-style accept() call). The syntax is: new_sd = sctp_peeloff(int sd, sctp_assoc_t *assoc_id, int *addrlen) new_sd - the new socket descriptor representing the branched-off association. sd - the original UDP-style socket descriptor returned from the socket() system call (see Section 3.1.1). assoc_id - the specified identifier of the association that is to be branched off to a separate file descriptor (Note, in a traditional TCP-style accept() call, this would be an out parameter, but for the UDP-style call, this is an in parameter). addrlen - an integer pointer to the size of the sockaddr structure addr (in a traditional TCP-style call, this would be a out parameter, but for the UDP-style call this is an in parameter). 9. Security Considerations Many TCP and UDP implementations reserve port numbers below 1024 for privileged users. If the target platform supports privileged users, the SCTP implementation SHOULD restrict the ability to call bind() or sctp_bindx() on these port numbers to privileged users. Similarly unprivelged users should not be able to set protocol parameters which could result in the congestion control algorithm being more agressive than permitted on the public Internet. These paramaters are: struct sctp_rtoinfo [There must be more. I'm digging through the Applicability Statement.] If an unprivileged user inherits a datagram model socket with open associations on a privileged port, it MAY be permitted to accept new associations, but it SHOULD NOT be permitted to open new associations. This could be relevant for the r* family of protocols. [Have we enabled any DoS attacks by making certain parameters visible to upper layers? I need to do a careful analysis on this one... Yes but we will fix it by implementing the agreement that J, K and I reached in San Jose.. i.e. no association automatically until you ask for COMM-UP NOTIFY events, we also need a way to specify backlog if I remember right :)] [Are there other security issues?] 10. Authors' Addresses Randall R. Stewart Tel: +1-815-479-8536 Cisco Systems, Inc. EMail: rrs@cisco.com Crystal Lake, IL 60012 USA Qiaobing Xie Tel: +1-847-632-3028 Motorola, Inc. EMail: qxie1@email.mot.com 1501 W. Shure Drive, Room 2309 Arlington Heights, IL 60004 USA La Monte H.P. Yarroll NIC Handle: LY Motorola, Inc. EMail: piggy@acm.org 1501 W. Shure Drive, IL27-2315 Arlington Heights, IL 60004 USA Jonathan Wood Sun Microsystems, Inc. Email: jonathan.wood@eng.sun.com 901 San Antonio Road Palo Alto, CA 94303 USA Kacheong Poon Sun Microsystems, Inc. Email: kacheong.poon@eng.sun.com 901 San Antonio Road Palo Alto, CA 94303 USA 11. References [RFC1644] Braden, R., "T/TCP -- TCP Extensions for Transactions Functional Specification," RFC 1644, July 1994. [RFC2026] Bradner, S., "The Internet Standards Process -- Revision 3", RFC 2026, October 1996. [RFC2292] W.R. Stevens, M. Thomas, "Advanced Sockets API for IPv6", RFC 2292, February 1998. [RFC2553] R. Gilligan, S. Thomson, J. Bound, W. Stevens. "Basic Socket Interface Extensions for IPv6," RFC 2553, March 1999. [SCTP] R.R. Stewart, Q. Xie, K. Morneault, C. Sharp, H.J. Schwarzbauer, T. Taylor, I. Rytina, M. Kalla, L. Zhang, and, V. Paxson, "Stream Control Transmission Protocol," , July 2000 work in progress. [STEVENS] W.R. Stevens, M. Thomas, E. Nordmark, "Advanced Sockets API for IPv6," , December 1999 (Work in progress)