NAME

nsp — NBMK/CyberGuard/NetOctave NSP2000 crypto accelerator

SYNOPSIS

nsp* at pci? dev ? function ?

DESCRIPTION

The nsp driver supports cards using the NSP2000 cryptographic processor, originally manufactured and sold by NetOctave, then CyberGuard, and presently (late 2008) by NBMK Encryption Technologies.

The NSP2000 is a cryptographic “packet processor” or “macro processor” featuring extensive support for protocol handshake acceleration and protocol record operations (e.g. single-pass pad-encrypt-and-hash for SSL or ESP messages). It also provides various cryptographic and mathematical primitives such as random number generation, encryption/decryption (DES, 3DES, and RC4), hash computation (MD5, SHA1, and HMAC), and an extensive set of operations for arbitrary precision arithmetic. It contains a tamper-resistant write-only memory region for storage of cryptographic keys.

The nsp driver registers support for the following operations with opencrypto(9):

CRYPTO_DES_CBC

DES in CBC mode.

CRYPTO_3DES_CBC

Triple-DES in CBC mode.

CRYPTO_MD5

The MD5 hash algorithm.

CRYPTO_SHA1

The SHA-1 hash algorithm.

CRYPTO_SHA1_HMAC

The HMAC message authentication code using SHA-1 as the hash function.

CRYPTO_MD5_HMAC

The HMAC message authentication code using MD5 as the hash function.

CRK_MOD

Compute x modulo y.

CRK_MOD_ADD

Modular addition.

CRK_MOD_ADDINV

Modular additive inversion.

CRK_MOD_SUB

Modular subtraction.

CRK_MOD_MULT

Modular multiplication.

CRK_MOD_MULTINV

Modular multiplicative inversion.

CRK_MOD_EXP

Modular exponentiation.

CRK_DSA_SIGN

DSA signature creation.

CRK_DSA_VERIFY

DSA signature verification.

CRK_DH_COMPUTE_KEY

Diffie-Hellman key computation.

PERFORMANCE

The nsp driver can perform several hundred 1024-bit RSA operations per second, and can encrypt and hash about 200Mbit/sec of data with symmetric operations. Each figure is approximately 1/3 the rated throughput for the device.

Several restrictions limit the performance of this driver:

1. The CRK_MOD_EXP_CRT operation (modular exponentiation with operands in Chinese Remainder Theorem form) is unfortunately not supported because the opencrypto(9) interface specifies this operation in a way which may only be compatible with the ubsec(4) accelerator.
2. The handshake operations and record transforms are not supported as they are a poor fit for the current opencrypto(9) API. Support for either would require a method of passing record-transform contexts between layers of the framework, likely in both directions across the user-kernel boundary. Without record operations, the host CPU will almost always perform RC4 faster than the NSP2000, so RC4 support is disabled in the nsp driver.
3. The on-board key memory is not supported. It would be relatively easy to add support for this feature to opencrypto(9), but the interface for supporting this functionality in OpenSSL in OpenSSL is complex and poorly documented, which makes kernel support useless.
4. The OpenSSL “engine” for crypto(4) does not yet support the HMAC forms of the hash operations, which roughly halves performance for many workloads.

On a more positive note, the NSP2000 and nsp driver offer excellent performance for small modular arithmetic operations, achieving 75,000 or more such operations per second.

SEE ALSO

crypto(4), fast_ipsec(4), intro(4), rnd(4), opencrypto(9)

HISTORY

The nsp device driver is descended from the NetOctave SDK for FreeBSD 4.11, where it was called “noct”. It is unrelated to the driver of that name which appeared in OpenBSD 3.2, which does not support the public-key (or other bignum) functions of the device. The nsp driver was ported to NetBSD 5.0 by Coyote Point Systems, Inc and generously made available under a BSD-style license by NBMK Encryption Technologies, Inc, the corporate successor of NetOctave.

The nsp device driver does not currently support the device node interface provided by the original NetOctave “noct” driver (which offers handshake acceleration, record operations, memory-mapped handling of packet payloads, and several other useful features) but most of the code to do so is still present, albeit in untested form.

BUGS

Support for limitations of the NSP2000 PCI interface (broken burst-mode operation, lack of scatter-gather support) is present but tested only on a fairly small range of host systems.

It appears that most if not all NSP2000 cards ever manufactured were designed to carry either one or two accelerator chips, which suggests that cards exist with both chips populated. The nsp driver has never been tested with more than one instance present at a time.