Tuesday, June 14, 2005

PF_KEY in Solaris, or "Dude, Where's My Spec?"

An OpenSolaris IPsec Hello

Hi! Those of you visiting here probably know I'm one of the IPsec guys (actually, I'm the original IPsec guy) here in Solaris-land. Bill may also have some stuff to say about the IPsec source in Solaris.

The kernel source for IPsec (AH, ESP, and the internal databases) lives in usr/src/uts/common/inet/ip/, because we're an integral part of our IP implementation. I should warn you now that there's a mixture of STREAMS boundaries and function calls between the different parts of the IPsec subsystem. It used to be almost all STREAMS, because of broken US Export restrictions (across all political party lines, BTW). We figured we could sell it as exportable to the powers that be more easily if we used a "general-purpose interface" which allowed for easy module perforation for moving our data around. As the restrictions loosened up, we were able to streamline things somewhat. We hope to do even more now that OpenSolaris is available. There are bits and pieces of the actual Solaris IPsec missing from OpenSolaris (especially from ESP) that will show up on OpenSolaris soon as well, now that we're officially open-source. (It's a bit of a chicken & egg problem.)

This entry will be discussing the PF_KEY implementation in Solaris. I assume you know something about how IPsec works, have read RFC 2367, and have a handle on TCP/IP protocol suite principles.

A Brief PF_KEY Synopsis

PF_KEY is analagous to the PF_ROUTE routing socket. See Keith Sklower's Radix-Tree paper at his site for the introduction to routing sockets. Where the routing socket manipulates IP forwarding entries (or routes), the PF_KEY socket manipulates IPsec Security Associations (SAs). A user-space application sends a message to the kernel telling it to ADD, DELETE, or UPDATE SAs, and the kernel sends back a message indicating either success or failure.

The paper makes mention of a message that's little-used in most PF_ROUTE implementations -- RTM_RESOLVE. RTM_RESOLVE allows a user-space application to resolve an address, e.g. a user-space ARP. This inspired PF_KEY's similar message,
SADB_ACQUIRE, which is used to tell a user-space key management (KM) daemon that an outgoing IPsec SA is needed. RFC 2367 has the specification for a PF_KEY socket.

Solaris Changes from RFC 2367

Most, if not all, existing PF_KEY implementations either alter or add to the message types in RFC 2367. Most changes were made because:

  • RFC2367 does not mesh well in to some KM protocols (esp. IKEv1).

  • The UNIX errno space is not sufficient to describe some failures.

  • Some implementors thought PF_KEY would be a suitable place to put their IPsec Security Policy Database (SPD) manipulations.

Solaris addresses the last bullet by introducing a separate PF_POLICY socket for SPD manipulation. The other issues, however, were a problem for us.

All of our changes to PF_KEY were a direct result of implementing The Internet Key Exchange (IKE) as part of our work in Solaris 9. They are summarized below:

  • Extended ACQUIRE - Instead of sending up a message for every IPsec SA that is not present, send up a list of what is needed to protect the packet to the listening KM daemon. This allows a packet that requires AH and ESP to express that protection in one ACQUIRE message.

  • Extended REGISTER - Goes hand-in-hand with Extended ACQUIRE. You tell the kernel that you can handle the Extended ACQUIRE.

  • Inverse ACQUIRE - The closest to policy manipulation we come in PF_KEY... it's a one-time consultation of the IPsec SPD, and you get as an answer an Extended ACQUIRE, just like if a outbound packet was triggering it. This is useful for IKE responders, and for diagnostic listeners on the PF_KEY socket.

  • Diagnostic codes - EINVAL is a frequently occurring value for sadb_msg_errno. Was it a weak DES key? Was it a botched sockaddr structure? The reserved field in struct sadb_msg now contains useful extra data when an EINVAL occurs.

  • typedefs - It's easier to type sadb_ext_t instead of struct sadb_ext.

  • 64-bit alignment - RFC 2367 claims all PF_KEY structures can be aligned on 64-bit boundaries. In Solaris, we force it to happen. That's why net/pfkeyv2.h has a lot of unions in most of its structure definitions.

But Dan... weren't you an author of RFC 2367?!?

Yes I was. Hence the question: Dude, Where's My Spec?

I wasn't allowed (yes, I'm serious; and no, it had nothing to do with any government interference) to work on IPsec or IKE when I first got to Sun, but the RFC was work that was a continuation from my previous job. In hindsight, I think we should've been paying more attention to the customers (authors of KM daemons, of which I'd be one someday). I was wrapped up in non-IPsec work at Sun when I wasn't working on what would become RFC 2367, and I split my attention in a non-optimal fashion.

Enough yapping, let's see some code!

The first place to look is usr/src/uts/common/net/pfkeyv2.h, which gets deposited into /usr/include/net/ on a running system. You'll notice every structure that doesn't have a field of type uint64_t will have a union in it. Here's the base PF_KEY message:

typedef struct sadb_msg {
uint8_t sadb_msg_version; /* Version, currently PF_KEY_V2 */
uint8_t sadb_msg_type; /* ADD, UPDATE, etc. */
uint8_t sadb_msg_errno; /* Error number from UNIX errno space */
uint8_t sadb_msg_satype; /* ESP, AH, etc. */
uint16_t sadb_msg_len; /* Length in 64-bit words. */
uint16_t sadb_msg_reserved; /* must be zero */
* Use the reserved field for extended diagnostic information on errno
* responses.
#define sadb_x_msg_diagnostic sadb_msg_reserved
/* Union is for guaranteeing 64-bit alignment. */
union {
struct {
uint32_t sadb_x_msg_useq; /* Set by originator */
uint32_t sadb_x_msg_upid; /* Set by originator */
} sadb_x_msg_actual;
uint64_t sadb_x_msg_alignment;
} sadb_x_msg_u;
#define sadb_msg_seq sadb_x_msg_u.sadb_x_msg_actual.sadb_x_msg_useq
#define sadb_msg_pid sadb_x_msg_u.sadb_x_msg_actual.sadb_x_msg_upid
} sadb_msg_t;

Notice that every extra field that is not in RFC 2367 uses the _X_ naming convention. In the case above, we took the two uint32_ts and merged them into a union with a uint64_t so that we can force 64-bit alignment on sadb_msg_t. This makes PF_KEY message manipulations 64-bit happy.

If you look at an extension that has a 64-bit type in it already, you'll see that there's no alignment-forcing union inserted into the definition:

typedef struct sadb_lifetime {
uint16_t sadb_lifetime_len;
uint16_t sadb_lifetime_exttype; /* SOFT, HARD, CURRENT */
uint32_t sadb_lifetime_allocations;
uint64_t sadb_lifetime_bytes;
uint64_t sadb_lifetime_addtime; /* These fields are assumed to hold */
uint64_t sadb_lifetime_usetime; /* >= sizeof (time_t). */
} sadb_lifetime_t;

(Yes, PF_KEY is Y2038-ready, as long as the KM application is compiled as 64-bit.)

For people interested in applications that use PF_KEY, I'd suggest investigating the ipseckey(1m) command. Its source can be found in usr/src/cmd/cmd-inet/usr.sbin/ipseckey.c. This program does not use command-line editing yet, but as an example of a PF_KEY consumer, it does the job.

All relevant PF_KEY internal-implementation headers live in usr/src/uts/common/inet. Relevant header files are:

  • ipsec_info.h
    - The structures for M_CTL messages prepended to data that gets passed around between IPsec STREAMS modules. The keysock consumer interface definitions are important here. Note the keysock_in_t, especially ks_in_extv[]. This vector allows easy access to all PF_KEY extension headers. (As we remove STREAMS from IPsec, this file will shrink. If all goes well, all that will remain (in some form) are IPSEC_IN and IPSEC_OUT messages, and that's because you can't enforce policy without some form of packet tagging.)

  • keysock.h
    - The keysock driver implements the PF_KEY socket interface at its most basic. A keysock_t represents an open PF_KEY socket, and a keysock_consumer_t represents a consumer of PF_KEY messages (i.e. AH and ESP).

All relevant PF_KEY internal-implementation source lives one level down from the headers, in the ip/ directory. They are:

  • keysockddi.c
    - DDI and module loading glue.

  • keysock.c
    - STREAMS driver that implements the PF_KEY socket as /dev/keysock. It is also a STREAMS module that sits atop AH and ESP listening for their messages.

  • sadb.c
    - Where IPsec's Security Association Database (SADB) is mostly implemented. You'll notice an ip_sadb.c file, because we want the fast-path lookups to be in IP without going through modstubs.

Most of this entry will be spent in keysock.c. Other portions of the subsystem will either be visited by one of Team IPsec when we have cycles, or if there are enough requests.

keysock either handles messages from a PF_KEY socket, or from the SADB and its consumers. The heavy lifting for user-generated PF_KEY messages is in keysock_parse().

The first thing keysock_parse() does is perform some reality checks. First off, the actual data length should match what's in the sadb_msg_len field. (Note the first of may SADB_nnTOmm() macros, for converting units of 64-bits to units of 8-bits, etc.) The first really interesting part of reality-checking is the "extension vector", or as it's called in the IPsec code, the extv. A PF_KEY message has a base header, followed by one or more extension headers. Let me quote this section from RFC 2367:

There MUST be only one instance of a extension type in a message. (e.g.
Base, Key, Lifetime, Key is forbidden). An EINVAL will be returned if there
are duplicate extensions within a message.

The keysock code takes this to heart in keysock_get_ext(). The extv is a vector of sadb_ext_t pointers, where the specific extension type (SADB_EXT_foo) can be found by merely indexing into the vector by that value. Say you want the sadb_sa_t extension:

sadb_sa_t *sa;

sa = (sadb_sa_t *)extv[SADB_EXT_SA];

The above code snippet shows what you need to do. As we generate the extv, if we see a collision, we return EINVAL (with an appropriate diagnostic). We do not enforce extension ordering inbound or outbound. Once keysock is done with first-pass reality-checks, the extv is sent around (as part of the KEYSOCK_IN M_CTL that is prepended to the data) to all who need it.

Most messages are shuttled off to keysock_passdown() for sending off via STREAMS to either AH or ESP. Unusual inbound messages

SADB_REGISTER sets socket state, as well as informs consumers about the register. If an SADB_REGISTER is sent for a specific SA type (e.g. ESP, AH), then the message is treated like a common-case message, except that when its reply arrive, keysock_t state is altered to indicate a registered socket. If the sadb_msg_satype is set to 0, then the message is an EXTENDED register, and an extended-register extension (sadb_x_ereg_t) is required. keysock converts the 0-terminated list of one-byte values into a bit vector internally. Then it sends the message to consumers like a normal REGISTER.

An inbound SADB_ACQUIRE can be used to signal other KM applications. (If PF_KEY is used to keep keys in the kernel for user-space consumers.) The more common case, however, is a negative ACQUIRE, which means a KM negotiation failed and the internal ACQUIRE record (more on this in a bit) needs to be cancelled.

SADB_FLUSH and SADB_DUMP messages need to lock down the keysock module until their respective operations are finished. FLUSH doesn't take as long, but DUMP needs to keep track of all consumer-originated replies until the consumer indicates it is done.

SADB_X_PROMISC merely changes some keysock state. It never goes to a consumer.

The SADB_X_INVERSE_ACQUIRE handling is a glimpse of things to come for keysock. It does not use the keysock_passdown() method of calling a consumer. It instead calls directly into IPsec (and if we had other in-kernel consumers, it would directly call to those) and returns a message to the user immediately.

The keysock_rput() function handles all messages from consumers. The KEYSOCK_OUT portion of the switch checks for FLUSH and DUMP messages, and releases the clamps on keysock if the final message for a FLUSH or a DUMP has been received. Otherwise, the keysock_passup() does the work.

keysock_passup() is conceptually much simpler than
keysock_parse(). It merely has to make one of three delivery decisions:

  • Do I deliver to all PF_KEY sockets?

  • Do I deliver to all registered PF_KEY sockets?

  • Do I deliver merely to the sender?

The more interesting parts of PF_KEY are handled inside the SADB code (sadb.c) and in the consumers. Those will be the subject of one or more other entries, because of all of the interaction with the IPsec SADB.

This entry was brought to you by the Technorati Tags OpenSolaris and Solaris.