Hello, all.
In preparation for making mantis and wiki more publicly accessible,
I am getting SSL certs sorted out for those services. Due to the
fact that the servers are running on an Marvell Kirkwood with an
asynchronous crypto offload engine, I would really rather like
to use that, because with it crypto is effectively free, and
without it it's expensive.
So last night I rebuilt OpenSSL with -DHAVE_CRYPTODEV
(-DUSE_CRYPTODEV_DIGESTS seems to cause various things to
complain and break, so I skipped that for now), and everything
works OK, right up to the point where something tries to use an
algorithm that can be offloaded (such aes-128-cbc), at which
point things fail almost completely silently. For example, sshd
with verbose and debug logging levels only reports one single
line regarding the problem:
fatal: evp_crypt: EVP_Cipher failed
Specifying any non-offloadable algorithm works fine (albeit
as slowly as usual).
That error appears to be getting emitted by OpenSSH sshd
rather than OpenSSL or Cryptodev. The relevant bit of code
is in OpenSSH's cipher.c file:
void
cipher_crypt(CipherContext *cc, u_char *dest, const u_char *src, u_int
len)
{
if (len % cc->cipher->block_size)
fatal("cipher_encrypt: bad plaintext length %d", len);
if (EVP_Cipher(&cc->evp, dest, (u_char *)src, len) == 0)
fatal("evp_crypt: EVP_Cipher failed");
}
EVP_Cipher function is part of OpenSSL, declared in:
crypto/evp/evp.h:
int EVP_Cipher(EVP_CIPHER_CTX *c,
unsigned char *out,
const unsigned char *in,
unsigned int inl);
and defined in:
crypto/evp/evp_lib.c:
int EVP_Cipher(EVP_CIPHER_CTX *ctx, unsigned char *out, const unsigned
char *in, unsigned int inl)
{
#ifdef OPENSSL_FIPS
FIPS_selftest_check();
#endif
return ctx->cipher->do_cipher(ctx,out,in,inl);
}
So it would appear that
ctx->cipher->do_cipher(ctx,out,in,inl);
returns 0.
The most annoying part is that doing this in two separate sessions
works:
Server session:
openssl s_server -accept 1234 -nocert -cipher PSK-AES128-CBC-SHA -psk
12345
Client session:
openssl s_client -connect localhost:1234 -cipher PSK-AES128-CBC-SHA -psk
12345
What comes in on the client, comes out on the server, and the crypto
engine interrupt count (grep mv_crypto /proc/interrupts) ticks up as
expected.
So the the basic crypto offload part seems to work just fine.
At that point I ran out of functioning brain cells last night.
I have not yet tried it with mod_ssl, which may yet turn out to work
fine, but ssh is just as valid a use case and was easier to test.
Has anyone else here fought cyptodev with any success and chased
things further toward something resembling a conclusion?
Gordan
_______________________________________________
users mailing list
[email protected]
http://lists.redsleeve.org/mailman/listinfo/users
