[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-Id: <20220617084210.907-1-ignat@cloudflare.com>
Date: Fri, 17 Jun 2022 09:42:10 +0100
From: Ignat Korchagin <ignat@...udflare.com>
To: Herbert Xu <herbert@...dor.apana.org.au>,
"David S . Miller" <davem@...emloft.net>,
linux-crypto@...r.kernel.org, linux-kernel@...r.kernel.org
Cc: kernel-team@...udflare.com, Ignat Korchagin <ignat@...udflare.com>,
kernel test robot <lkp@...el.com>
Subject: [PATCH v2] crypto: rsa - implement Chinese Remainder Theorem for faster private key operations
Changes from v1:
* exported mpi_sub and mpi_mul, otherwise the build fails when RSA is a module
The kernel RSA ASN.1 private key parser already supports only private keys with
additional values to be used with the Chinese Remainder Theorem [1], but these
values are currently not used.
This rudimentary CRT implementation speeds up RSA private key operations for the
following Go benchmark up to ~3x.
This implementation also tries to minimise the allocation of additional MPIs,
so existing MPIs are reused as much as possible (hence the variable names are a
bit weird).
The benchmark used:
```
package keyring_test
import (
"crypto"
"crypto/rand"
"crypto/rsa"
"crypto/x509"
"io"
"syscall"
"testing"
"unsafe"
)
type KeySerial int32
type Keyring int32
const (
KEY_SPEC_PROCESS_KEYRING Keyring = -2
KEYCTL_PKEY_SIGN = 27
)
var (
keyTypeAsym = []byte("asymmetric\x00")
sha256pkcs1 = []byte("enc=pkcs1 hash=sha256\x00")
)
func (keyring Keyring) LoadAsym(desc string, payload []byte) (KeySerial, error) {
cdesc := []byte(desc + "\x00")
serial, _, errno := syscall.Syscall6(syscall.SYS_ADD_KEY, uintptr(unsafe.Pointer(&keyTypeAsym[0])), uintptr(unsafe.Pointer(&cdesc[0])), uintptr(unsafe.Pointer(&payload[0])), uintptr(len(payload)), uintptr(keyring), uintptr(0))
if errno == 0 {
return KeySerial(serial), nil
}
return KeySerial(serial), errno
}
type pkeyParams struct {
key_id KeySerial
in_len uint32
out_or_in2_len uint32
__spare [7]uint32
}
// the output signature buffer is an input parameter here, because we want to
// avoid Go buffer allocation leaking into our benchmarks
func (key KeySerial) Sign(info, digest, out []byte) error {
var params pkeyParams
params.key_id = key
params.in_len = uint32(len(digest))
params.out_or_in2_len = uint32(len(out))
_, _, errno := syscall.Syscall6(syscall.SYS_KEYCTL, KEYCTL_PKEY_SIGN, uintptr(unsafe.Pointer(¶ms)), uintptr(unsafe.Pointer(&info[0])), uintptr(unsafe.Pointer(&digest[0])), uintptr(unsafe.Pointer(&out[0])), uintptr(0))
if errno == 0 {
return nil
}
return errno
}
func BenchmarkSign(b *testing.B) {
priv, err := rsa.GenerateKey(rand.Reader, 2048)
if err != nil {
b.Fatalf("failed to generate private key: %v", err)
}
pkcs8, err := x509.MarshalPKCS8PrivateKey(priv)
if err != nil {
b.Fatalf("failed to serialize the private key to PKCS8 blob: %v", err)
}
serial, err := KEY_SPEC_PROCESS_KEYRING.LoadAsym("test rsa key", pkcs8)
if err != nil {
b.Fatalf("failed to load the private key into the keyring: %v", err)
}
b.Logf("loaded test rsa key: %v", serial)
digest := make([]byte, 32)
_, err = io.ReadFull(rand.Reader, digest)
if err != nil {
b.Fatalf("failed to generate a random digest: %v", err)
}
sig := make([]byte, 256)
for n := 0; n < b.N; n++ {
err = serial.Sign(sha256pkcs1, digest, sig)
if err != nil {
b.Fatalf("failed to sign the digest: %v", err)
}
}
err = rsa.VerifyPKCS1v15(&priv.PublicKey, crypto.SHA256, digest, sig)
if err != nil {
b.Fatalf("failed to verify the signature: %v", err)
}
}
```
[1]: https://en.wikipedia.org/wiki/RSA_(cryptosystem)#Using_the_Chinese_remainder_algorithm
Signed-off-by: Ignat Korchagin <ignat@...udflare.com>
Reported-by: kernel test robot <lkp@...el.com>
---
crypto/rsa.c | 78 ++++++++++++++++++++++++++++++++++++++++++++---
lib/mpi/mpi-add.c | 2 +-
lib/mpi/mpi-mul.c | 1 +
3 files changed, 75 insertions(+), 6 deletions(-)
diff --git a/crypto/rsa.c b/crypto/rsa.c
index 39e04176b04b..0e555ee4addb 100644
--- a/crypto/rsa.c
+++ b/crypto/rsa.c
@@ -17,6 +17,11 @@ struct rsa_mpi_key {
MPI n;
MPI e;
MPI d;
+ MPI p;
+ MPI q;
+ MPI dp;
+ MPI dq;
+ MPI qinv;
};
/*
@@ -35,16 +40,49 @@ static int _rsa_enc(const struct rsa_mpi_key *key, MPI c, MPI m)
/*
* RSADP function [RFC3447 sec 5.1.2]
- * m = c^d mod n;
+ * m_1 = c^dP mod p;
+ * m_2 = c^dQ mod q;
+ * h = (m_1 - m_2) * qInv mod p;
+ * m = m_2 + q * h;
*/
-static int _rsa_dec(const struct rsa_mpi_key *key, MPI m, MPI c)
+static int _rsa_dec_crt(const struct rsa_mpi_key *key, MPI m_or_m1_or_h, MPI c)
{
+ MPI m2, m12_or_qh;
+ int ret = -ENOMEM;
+
/* (1) Validate 0 <= c < n */
if (mpi_cmp_ui(c, 0) < 0 || mpi_cmp(c, key->n) >= 0)
return -EINVAL;
- /* (2) m = c^d mod n */
- return mpi_powm(m, c, key->d, key->n);
+ m2 = mpi_alloc(0);
+ m12_or_qh = mpi_alloc(0);
+ if (!m2 || !m12_or_qh)
+ goto err_free_mpi;
+
+ /* (2i) m_1 = c^dP mod p */
+ ret = mpi_powm(m_or_m1_or_h, c, key->dp, key->p);
+ if (ret)
+ goto err_free_mpi;
+
+ /* (2i) m_2 = c^dQ mod q */
+ ret = mpi_powm(m2, c, key->dq, key->q);
+ if (ret)
+ goto err_free_mpi;
+
+ /* (2iii) h = (m_1 - m_2) * qInv mod p */
+ mpi_sub(m12_or_qh, m_or_m1_or_h, m2);
+ mpi_mulm(m_or_m1_or_h, m12_or_qh, key->qinv, key->p);
+
+ /* (2iv) m = m_2 + q * h */
+ mpi_mul(m12_or_qh, key->q, m_or_m1_or_h);
+ mpi_addm(m_or_m1_or_h, m2, m12_or_qh, key->n);
+
+ ret = 0;
+
+err_free_mpi:
+ mpi_free(m12_or_qh);
+ mpi_free(m2);
+ return ret;
}
static inline struct rsa_mpi_key *rsa_get_key(struct crypto_akcipher *tfm)
@@ -112,7 +150,7 @@ static int rsa_dec(struct akcipher_request *req)
if (!c)
goto err_free_m;
- ret = _rsa_dec(pkey, m, c);
+ ret = _rsa_dec_crt(pkey, m, c);
if (ret)
goto err_free_c;
@@ -134,9 +172,19 @@ static void rsa_free_mpi_key(struct rsa_mpi_key *key)
mpi_free(key->d);
mpi_free(key->e);
mpi_free(key->n);
+ mpi_free(key->p);
+ mpi_free(key->q);
+ mpi_free(key->dp);
+ mpi_free(key->dq);
+ mpi_free(key->qinv);
key->d = NULL;
key->e = NULL;
key->n = NULL;
+ key->p = NULL;
+ key->q = NULL;
+ key->dp = NULL;
+ key->dq = NULL;
+ key->qinv = NULL;
}
static int rsa_check_key_length(unsigned int len)
@@ -217,6 +265,26 @@ static int rsa_set_priv_key(struct crypto_akcipher *tfm, const void *key,
if (!mpi_key->n)
goto err;
+ mpi_key->p = mpi_read_raw_data(raw_key.p, raw_key.p_sz);
+ if (!mpi_key->p)
+ goto err;
+
+ mpi_key->q = mpi_read_raw_data(raw_key.q, raw_key.q_sz);
+ if (!mpi_key->q)
+ goto err;
+
+ mpi_key->dp = mpi_read_raw_data(raw_key.dp, raw_key.dp_sz);
+ if (!mpi_key->dp)
+ goto err;
+
+ mpi_key->dq = mpi_read_raw_data(raw_key.dq, raw_key.dq_sz);
+ if (!mpi_key->dq)
+ goto err;
+
+ mpi_key->qinv = mpi_read_raw_data(raw_key.qinv, raw_key.qinv_sz);
+ if (!mpi_key->qinv)
+ goto err;
+
if (rsa_check_key_length(mpi_get_size(mpi_key->n) << 3)) {
rsa_free_mpi_key(mpi_key);
return -EINVAL;
diff --git a/lib/mpi/mpi-add.c b/lib/mpi/mpi-add.c
index 2cdae54c1bd0..9056fc5167fc 100644
--- a/lib/mpi/mpi-add.c
+++ b/lib/mpi/mpi-add.c
@@ -138,7 +138,7 @@ void mpi_sub(MPI w, MPI u, MPI v)
mpi_add(w, u, vv);
mpi_free(vv);
}
-
+EXPORT_SYMBOL_GPL(mpi_sub);
void mpi_addm(MPI w, MPI u, MPI v, MPI m)
{
diff --git a/lib/mpi/mpi-mul.c b/lib/mpi/mpi-mul.c
index 8f5fa200f297..7f4eda8560dc 100644
--- a/lib/mpi/mpi-mul.c
+++ b/lib/mpi/mpi-mul.c
@@ -82,6 +82,7 @@ void mpi_mul(MPI w, MPI u, MPI v)
if (tmp_limb)
mpi_free_limb_space(tmp_limb);
}
+EXPORT_SYMBOL_GPL(mpi_mul);
void mpi_mulm(MPI w, MPI u, MPI v, MPI m)
{
--
2.36.1
Powered by blists - more mailing lists