[openssl-commits] [openssl] OpenSSL_1_0_2-stable update
Matt Caswell
matt at openssl.org
Thu Dec 6 11:23:25 UTC 2018
The branch OpenSSL_1_0_2-stable has been updated
via 110ef88b99f1acc6b976f2e49153734924181db2 (commit)
via b29b91bc7ea8dfe47d873d7953175c57556a4488 (commit)
via 47f8fba64353a637cacdd8751cab25a9f3be3715 (commit)
via 3b776fd785efb1c0af4b28ad0b8036b23071904d (commit)
via 8db50d6dafc67fbaa0380420fc1f74f977d16606 (commit)
from 02354431e2a09d705082f986babf14c056f47b78 (commit)
- Log -----------------------------------------------------------------
commit 110ef88b99f1acc6b976f2e49153734924181db2
Author: Andy Polyakov <appro at openssl.org>
Date: Fri Sep 14 17:24:13 2018 +0200
rsa/rsa_ssl.c: make RSA_padding_check_SSLv23 constant-time.
Copy of RSA_padding_check_PKCS1_type_2 with a twist that rejects padding
if nul delimiter is preceded by 8 consecutive 0x03 bytes.
Reviewed-by: Richard Levitte <levitte at openssl.org>
Reviewed-by: Matt Caswell <matt at openssl.org>
(cherry picked from commit 603221407ddc6404f8c417c6beadebf84449074c)
Resolved conflicts:
crypto/rsa/rsa_ssl.c
(Merged from https://github.com/openssl/openssl/pull/7737)
commit b29b91bc7ea8dfe47d873d7953175c57556a4488
Author: Andy Polyakov <appro at openssl.org>
Date: Thu Sep 6 21:54:23 2018 +0200
rsa/rsa_oaep.c: remove memcpy calls from RSA_padding_check_PKCS1_OAEP.
And make RSAErr call unconditional.
Reviewed-by: Richard Levitte <levitte at openssl.org>
Reviewed-by: Matt Caswell <matt at openssl.org>
(cherry picked from commit 75f5e944be97f28867e7c489823c889d89d0bd06)
Resolved conflicts:
crypto/rsa/rsa_oaep.c
(Merged from https://github.com/openssl/openssl/pull/7737)
commit 47f8fba64353a637cacdd8751cab25a9f3be3715
Author: Andy Polyakov <appro at openssl.org>
Date: Sat Sep 1 12:00:33 2018 +0200
rsa/rsa_pk1.c: remove memcpy calls from RSA_padding_check_PKCS1_type_2.
And make RSAErr call unconditional.
Reviewed-by: Richard Levitte <levitte at openssl.org>
Reviewed-by: Matt Caswell <matt at openssl.org>
(cherry picked from commit e875b0cf2f10bf2adf73e0c2ec81428290f4660c)
Resolved conflicts:
crypto/rsa/rsa_pk1.c
(Merged from https://github.com/openssl/openssl/pull/7737)
commit 3b776fd785efb1c0af4b28ad0b8036b23071904d
Author: Andy Polyakov <appro at openssl.org>
Date: Fri Nov 30 21:07:18 2018 +0100
rsa/rsa_eay.c: make RSAerr call in rsa_ossl_private_decrypt unconditional.
Reviewed-by: Richard Levitte <levitte at openssl.org>
Reviewed-by: Matt Caswell <matt at openssl.org>
(cherry picked from commit 89072e0c2a483f2ad678e723e112712567b0ceb1)
(Merged from https://github.com/openssl/openssl/pull/7737)
commit 8db50d6dafc67fbaa0380420fc1f74f977d16606
Author: Andy Polyakov <appro at openssl.org>
Date: Sat Sep 1 12:19:30 2018 +0200
err/err.c: add err_clear_last_constant_time.
Expected usage pattern is to unconditionally set error and then
wipe it if there was no actual error.
Reviewed-by: Richard Levitte <levitte at openssl.org>
Reviewed-by: Matt Caswell <matt at openssl.org>
(cherry picked from commit f658a3b64d8750642f4975090740865f770c2a1b)
Resolved conflicts:
crypto/err/err.c
crypto/constant_time_locl.h
(Merged from https://github.com/openssl/openssl/pull/7737)
-----------------------------------------------------------------------
Summary of changes:
crypto/constant_time_locl.h | 6 ++
crypto/err/err.c | 38 ++++++++
crypto/rsa/rsa_eay.c | 5 +-
crypto/rsa/rsa_oaep.c | 96 +++++++++++---------
crypto/rsa/rsa_pk1.c | 98 ++++++++++----------
crypto/rsa/rsa_ssl.c | 134 ++++++++++++++++++++--------
doc/crypto/RSA_padding_add_PKCS1_type_1.pod | 7 +-
7 files changed, 258 insertions(+), 126 deletions(-)
diff --git a/crypto/constant_time_locl.h b/crypto/constant_time_locl.h
index c786aea..a5734f2 100644
--- a/crypto/constant_time_locl.h
+++ b/crypto/constant_time_locl.h
@@ -204,6 +204,12 @@ static inline int constant_time_select_int(unsigned int mask, int a, int b)
return (int)(constant_time_select(mask, (unsigned)(a), (unsigned)(b)));
}
+/*
+ * Expected usage pattern is to unconditionally set error and then
+ * wipe it if there was no actual error. |clear| is 1 or 0.
+ */
+void err_clear_last_constant_time(int clear);
+
#ifdef __cplusplus
}
#endif
diff --git a/crypto/err/err.c b/crypto/err/err.c
index e9ef215..5ce774a 100644
--- a/crypto/err/err.c
+++ b/crypto/err/err.c
@@ -118,6 +118,7 @@
#include <openssl/buffer.h>
#include <openssl/bio.h>
#include <openssl/err.h>
+#include "constant_time_locl.h"
DECLARE_LHASH_OF(ERR_STRING_DATA);
DECLARE_LHASH_OF(ERR_STATE);
@@ -1156,3 +1157,40 @@ int ERR_pop_to_mark(void)
es->err_flags[es->top] &= ~ERR_FLAG_MARK;
return 1;
}
+
+#ifdef UINTPTR_T
+# undef UINTPTR_T
+#endif
+/*
+ * uintptr_t is the answer, but unformtunately we can't assume that all
+ * compilers supported by 1.0.2 have it :-(
+ */
+#if defined(OPENSSL_SYS_VMS) && __INITIAL_POINTER_SIZE==64
+/*
+ * But we can't use size_t on VMS, because it adheres to sizeof(size_t)==4
+ * even in 64-bit builds, which means that it won't work as mask.
+ */
+# define UINTPTR_T unsigned long long
+#else
+# define UINTPTR_T size_t
+#endif
+
+void err_clear_last_constant_time(int clear)
+{
+ ERR_STATE *es;
+ int top;
+
+ es = ERR_get_state();
+ if (es == NULL)
+ return;
+
+ top = es->top;
+
+ es->err_flags[top] &= ~(0 - clear);
+ es->err_buffer[top] &= ~(0UL - clear);
+ es->err_file[top] = (const char *)((UINTPTR_T)es->err_file[top] &
+ ~((UINTPTR_T)0 - clear));
+ es->err_line[top] |= 0 - clear;
+
+ es->top = (top + ERR_NUM_ERRORS - clear) % ERR_NUM_ERRORS;
+}
diff --git a/crypto/rsa/rsa_eay.c b/crypto/rsa/rsa_eay.c
index 1155583..7f20fd6 100644
--- a/crypto/rsa/rsa_eay.c
+++ b/crypto/rsa/rsa_eay.c
@@ -115,6 +115,7 @@
#include <openssl/rsa.h>
#include <openssl/rand.h>
#include "bn_int.h"
+#include "constant_time_locl.h"
#ifndef RSA_NULL
@@ -587,8 +588,8 @@ static int RSA_eay_private_decrypt(int flen, const unsigned char *from,
RSAerr(RSA_F_RSA_EAY_PRIVATE_DECRYPT, RSA_R_UNKNOWN_PADDING_TYPE);
goto err;
}
- if (r < 0)
- RSAerr(RSA_F_RSA_EAY_PRIVATE_DECRYPT, RSA_R_PADDING_CHECK_FAILED);
+ RSAerr(RSA_F_RSA_EAY_PRIVATE_DECRYPT, RSA_R_PADDING_CHECK_FAILED);
+ err_clear_last_constant_time(r >= 0);
err:
if (ctx != NULL) {
diff --git a/crypto/rsa/rsa_oaep.c b/crypto/rsa/rsa_oaep.c
index 3fb8f6b..033ea5a 100644
--- a/crypto/rsa/rsa_oaep.c
+++ b/crypto/rsa/rsa_oaep.c
@@ -121,7 +121,7 @@ int RSA_padding_check_PKCS1_OAEP_mgf1(unsigned char *to, int tlen,
const EVP_MD *mgf1md)
{
int i, dblen = 0, mlen = -1, one_index = 0, msg_index;
- unsigned int good, found_one_byte;
+ unsigned int good = 0, found_one_byte, mask;
const unsigned char *maskedseed, *maskeddb;
/*
* |em| is the encoded message, zero-padded to exactly |num| bytes: em =
@@ -148,8 +148,11 @@ int RSA_padding_check_PKCS1_OAEP_mgf1(unsigned char *to, int tlen,
* the ciphertext, see PKCS #1 v2.2, section 7.1.2.
* This does not leak any side-channel information.
*/
- if (num < flen || num < 2 * mdlen + 2)
- goto decoding_err;
+ if (num < flen || num < 2 * mdlen + 2) {
+ RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_OAEP_MGF1,
+ RSA_R_OAEP_DECODING_ERROR);
+ return -1;
+ }
dblen = num - mdlen - 1;
db = OPENSSL_malloc(dblen);
@@ -158,26 +161,26 @@ int RSA_padding_check_PKCS1_OAEP_mgf1(unsigned char *to, int tlen,
goto cleanup;
}
- if (flen != num) {
- em = OPENSSL_malloc(num);
- if (em == NULL) {
- RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_OAEP_MGF1,
- ERR_R_MALLOC_FAILURE);
- goto cleanup;
- }
+ em = OPENSSL_malloc(num);
+ if (em == NULL) {
+ RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_OAEP_MGF1,
+ ERR_R_MALLOC_FAILURE);
+ goto cleanup;
+ }
- /*
- * Caller is encouraged to pass zero-padded message created with
- * BN_bn2binpad, but if it doesn't, we do this zero-padding copy
- * to avoid leaking that information. The copy still leaks some
- * side-channel information, but it's impossible to have a fixed
- * memory access pattern since we can't read out of the bounds of
- * |from|.
- */
- memset(em, 0, num);
- memcpy(em + num - flen, from, flen);
- from = em;
+ /*
+ * Caller is encouraged to pass zero-padded message created with
+ * BN_bn2binpad. Trouble is that since we can't read out of |from|'s
+ * bounds, it's impossible to have an invariant memory access pattern
+ * in case |from| was not zero-padded in advance.
+ */
+ for (from += flen, em += num, i = 0; i < num; i++) {
+ mask = ~constant_time_is_zero(flen);
+ flen -= 1 & mask;
+ from -= 1 & mask;
+ *--em = *from & mask;
}
+ from = em;
/*
* The first byte must be zero, however we must not leak if this is
@@ -224,37 +227,50 @@ int RSA_padding_check_PKCS1_OAEP_mgf1(unsigned char *to, int tlen,
* so plaintext-awareness ensures timing side-channels are no longer a
* concern.
*/
- if (!good)
- goto decoding_err;
-
msg_index = one_index + 1;
mlen = dblen - msg_index;
- if (tlen < mlen) {
- RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_OAEP_MGF1, RSA_R_DATA_TOO_LARGE);
- mlen = -1;
- } else {
- memcpy(to, db + msg_index, mlen);
- goto cleanup;
+ /*
+ * For good measure, do this check in constant tine as well.
+ */
+ good &= constant_time_ge(tlen, mlen);
+
+ /*
+ * Even though we can't fake result's length, we can pretend copying
+ * |tlen| bytes where |mlen| bytes would be real. Last |tlen| of |dblen|
+ * bytes are viewed as circular buffer with start at |tlen|-|mlen'|,
+ * where |mlen'| is "saturated" |mlen| value. Deducing information
+ * about failure or |mlen| would take attacker's ability to observe
+ * memory access pattern with byte granularity *as it occurs*. It
+ * should be noted that failure is indistinguishable from normal
+ * operation if |tlen| is fixed by protocol.
+ */
+ tlen = constant_time_select_int(constant_time_lt(dblen, tlen), dblen, tlen);
+ msg_index = constant_time_select_int(good, msg_index, dblen - tlen);
+ mlen = dblen - msg_index;
+ for (from = db + msg_index, mask = good, i = 0; i < tlen; i++) {
+ unsigned int equals = constant_time_eq(i, mlen);
+
+ from -= dblen & equals; /* if (i == dblen) rewind */
+ mask &= mask ^ equals; /* if (i == dblen) mask = 0 */
+ to[i] = constant_time_select_8(mask, from[i], to[i]);
}
- decoding_err:
/*
* To avoid chosen ciphertext attacks, the error message should not
* reveal which kind of decoding error happened.
*/
RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_OAEP_MGF1,
RSA_R_OAEP_DECODING_ERROR);
+ err_clear_last_constant_time(1 & good);
cleanup:
- if (db != NULL) {
- OPENSSL_cleanse(db, dblen);
- OPENSSL_free(db);
- }
- if (em != NULL) {
- OPENSSL_cleanse(em, num);
- OPENSSL_free(em);
- }
- return mlen;
+ OPENSSL_cleanse(seed, sizeof(seed));
+ OPENSSL_cleanse(db, dblen);
+ OPENSSL_free(db);
+ OPENSSL_cleanse(em, num);
+ OPENSSL_free(em);
+
+ return constant_time_select_int(good, mlen, -1);
}
int PKCS1_MGF1(unsigned char *mask, long len,
diff --git a/crypto/rsa/rsa_pk1.c b/crypto/rsa/rsa_pk1.c
index 5d7882a..074bc0a 100644
--- a/crypto/rsa/rsa_pk1.c
+++ b/crypto/rsa/rsa_pk1.c
@@ -207,7 +207,7 @@ int RSA_padding_check_PKCS1_type_2(unsigned char *to, int tlen,
int i;
/* |em| is the encoded message, zero-padded to exactly |num| bytes */
unsigned char *em = NULL;
- unsigned int good, found_zero_byte;
+ unsigned int good, found_zero_byte, mask;
int zero_index = 0, msg_index, mlen = -1;
if (tlen < 0 || flen < 0)
@@ -218,40 +218,41 @@ int RSA_padding_check_PKCS1_type_2(unsigned char *to, int tlen,
* section 7.2.2.
*/
- if (flen > num)
- goto err;
-
- if (num < 11)
- goto err;
+ if (flen > num || num < 11) {
+ RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2,
+ RSA_R_PKCS_DECODING_ERROR);
+ return -1;
+ }
- if (flen != num) {
- em = OPENSSL_malloc(num);
- if (em == NULL) {
- RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2, ERR_R_MALLOC_FAILURE);
- return -1;
- }
- /*
- * Caller is encouraged to pass zero-padded message created with
- * BN_bn2binpad, but if it doesn't, we do this zero-padding copy
- * to avoid leaking that information. The copy still leaks some
- * side-channel information, but it's impossible to have a fixed
- * memory access pattern since we can't read out of the bounds of
- * |from|.
- */
- memset(em, 0, num);
- memcpy(em + num - flen, from, flen);
- from = em;
+ em = OPENSSL_malloc(num);
+ if (em == NULL) {
+ RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2, ERR_R_MALLOC_FAILURE);
+ return -1;
}
+ /*
+ * Caller is encouraged to pass zero-padded message created with
+ * BN_bn2binpad. Trouble is that since we can't read out of |from|'s
+ * bounds, it's impossible to have an invariant memory access pattern
+ * in case |from| was not zero-padded in advance.
+ */
+ for (from += flen, em += num, i = 0; i < num; i++) {
+ mask = ~constant_time_is_zero(flen);
+ flen -= 1 & mask;
+ from -= 1 & mask;
+ *--em = *from & mask;
+ }
+ from = em;
good = constant_time_is_zero(from[0]);
good &= constant_time_eq(from[1], 2);
+ /* scan over padding data */
found_zero_byte = 0;
for (i = 2; i < num; i++) {
unsigned int equals0 = constant_time_is_zero(from[i]);
- zero_index =
- constant_time_select_int(~found_zero_byte & equals0, i,
- zero_index);
+
+ zero_index = constant_time_select_int(~found_zero_byte & equals0,
+ i, zero_index);
found_zero_byte |= equals0;
}
@@ -260,7 +261,7 @@ int RSA_padding_check_PKCS1_type_2(unsigned char *to, int tlen,
* If we never found a 0-byte, then |zero_index| is 0 and the check
* also fails.
*/
- good &= constant_time_ge((unsigned int)(zero_index), 2 + 8);
+ good &= constant_time_ge(zero_index, 2 + 8);
/*
* Skip the zero byte. This is incorrect if we never found a zero-byte
@@ -270,30 +271,35 @@ int RSA_padding_check_PKCS1_type_2(unsigned char *to, int tlen,
mlen = num - msg_index;
/*
- * For good measure, do this check in constant time as well; it could
- * leak something if |tlen| was assuming valid padding.
+ * For good measure, do this check in constant time as well.
*/
- good &= constant_time_ge((unsigned int)(tlen), (unsigned int)(mlen));
+ good &= constant_time_ge(tlen, mlen);
/*
- * We can't continue in constant-time because we need to copy the result
- * and we cannot fake its length. This unavoidably leaks timing
- * information at the API boundary.
+ * Even though we can't fake result's length, we can pretend copying
+ * |tlen| bytes where |mlen| bytes would be real. Last |tlen| of |num|
+ * bytes are viewed as circular buffer with start at |tlen|-|mlen'|,
+ * where |mlen'| is "saturated" |mlen| value. Deducing information
+ * about failure or |mlen| would take attacker's ability to observe
+ * memory access pattern with byte granularity *as it occurs*. It
+ * should be noted that failure is indistinguishable from normal
+ * operation if |tlen| is fixed by protocol.
*/
- if (!good) {
- mlen = -1;
- goto err;
+ tlen = constant_time_select_int(constant_time_lt(num, tlen), num, tlen);
+ msg_index = constant_time_select_int(good, msg_index, num - tlen);
+ mlen = num - msg_index;
+ for (from += msg_index, mask = good, i = 0; i < tlen; i++) {
+ unsigned int equals = constant_time_eq(i, mlen);
+
+ from -= tlen & equals; /* if (i == mlen) rewind */
+ mask &= mask ^ equals; /* if (i == mlen) mask = 0 */
+ to[i] = constant_time_select_8(mask, from[i], to[i]);
}
- memcpy(to, from + msg_index, mlen);
+ OPENSSL_cleanse(em, num);
+ OPENSSL_free(em);
+ RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2, RSA_R_PKCS_DECODING_ERROR);
+ err_clear_last_constant_time(1 & good);
- err:
- if (em != NULL) {
- OPENSSL_cleanse(em, num);
- OPENSSL_free(em);
- }
- if (mlen == -1)
- RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2,
- RSA_R_PKCS_DECODING_ERROR);
- return mlen;
+ return constant_time_select_int(good, mlen, -1);
}
diff --git a/crypto/rsa/rsa_ssl.c b/crypto/rsa/rsa_ssl.c
index 831f75a..e9a5fe2 100644
--- a/crypto/rsa/rsa_ssl.c
+++ b/crypto/rsa/rsa_ssl.c
@@ -61,6 +61,7 @@
#include <openssl/bn.h>
#include <openssl/rsa.h>
#include <openssl/rand.h>
+#include "constant_time_locl.h"
int RSA_padding_add_SSLv23(unsigned char *to, int tlen,
const unsigned char *from, int flen)
@@ -101,57 +102,116 @@ int RSA_padding_add_SSLv23(unsigned char *to, int tlen,
return (1);
}
+/*
+ * Copy of RSA_padding_check_PKCS1_type_2 with a twist that rejects padding
+ * if nul delimiter is preceded by 8 consecutive 0x03 bytes. It also
+ * preserves error code reporting for backward compatibility.
+ */
int RSA_padding_check_SSLv23(unsigned char *to, int tlen,
const unsigned char *from, int flen, int num)
{
- int i, j, k;
- const unsigned char *p;
+ int i;
+ /* |em| is the encoded message, zero-padded to exactly |num| bytes */
+ unsigned char *em = NULL;
+ unsigned int good, found_zero_byte, mask, threes_in_row;
+ int zero_index = 0, msg_index, mlen = -1, err;
- p = from;
if (flen < 10) {
RSAerr(RSA_F_RSA_PADDING_CHECK_SSLV23, RSA_R_DATA_TOO_SMALL);
return (-1);
}
- /* Accept even zero-padded input */
- if (flen == num) {
- if (*(p++) != 0) {
- RSAerr(RSA_F_RSA_PADDING_CHECK_SSLV23, RSA_R_BLOCK_TYPE_IS_NOT_02);
- return -1;
- }
- flen--;
+
+ em = OPENSSL_malloc(num);
+ if (em == NULL) {
+ RSAerr(RSA_F_RSA_PADDING_CHECK_SSLV23, ERR_R_MALLOC_FAILURE);
+ return -1;
}
- if ((num != (flen + 1)) || (*(p++) != 02)) {
- RSAerr(RSA_F_RSA_PADDING_CHECK_SSLV23, RSA_R_BLOCK_TYPE_IS_NOT_02);
- return (-1);
+ /*
+ * Caller is encouraged to pass zero-padded message created with
+ * BN_bn2binpad. Trouble is that since we can't read out of |from|'s
+ * bounds, it's impossible to have an invariant memory access pattern
+ * in case |from| was not zero-padded in advance.
+ */
+ for (from += flen, em += num, i = 0; i < num; i++) {
+ mask = ~constant_time_is_zero(flen);
+ flen -= 1 & mask;
+ from -= 1 & mask;
+ *--em = *from & mask;
}
+ from = em;
+
+ good = constant_time_is_zero(from[0]);
+ good &= constant_time_eq(from[1], 2);
+ err = constant_time_select_int(good, 0, RSA_R_BLOCK_TYPE_IS_NOT_02);
+ mask = ~good;
/* scan over padding data */
- j = flen - 1; /* one for type */
- for (i = 0; i < j; i++)
- if (*(p++) == 0)
- break;
-
- if ((i == j) || (i < 8)) {
- RSAerr(RSA_F_RSA_PADDING_CHECK_SSLV23,
- RSA_R_NULL_BEFORE_BLOCK_MISSING);
- return (-1);
- }
- for (k = -9; k < -1; k++) {
- if (p[k] != 0x03)
- break;
- }
- if (k == -1) {
- RSAerr(RSA_F_RSA_PADDING_CHECK_SSLV23, RSA_R_SSLV3_ROLLBACK_ATTACK);
- return (-1);
+ found_zero_byte = 0;
+ threes_in_row = 0;
+ for (i = 2; i < num; i++) {
+ unsigned int equals0 = constant_time_is_zero(from[i]);
+
+ zero_index = constant_time_select_int(~found_zero_byte & equals0,
+ i, zero_index);
+ found_zero_byte |= equals0;
+
+ threes_in_row += 1 & ~found_zero_byte;
+ threes_in_row &= found_zero_byte | constant_time_eq(from[i], 3);
}
- i++; /* Skip over the '\0' */
- j -= i;
- if (j > tlen) {
- RSAerr(RSA_F_RSA_PADDING_CHECK_SSLV23, RSA_R_DATA_TOO_LARGE);
- return (-1);
+ /*
+ * PS must be at least 8 bytes long, and it starts two bytes into |from|.
+ * If we never found a 0-byte, then |zero_index| is 0 and the check
+ * also fails.
+ */
+ good &= constant_time_ge(zero_index, 2 + 8);
+ err = constant_time_select_int(mask | good, err,
+ RSA_R_NULL_BEFORE_BLOCK_MISSING);
+ mask = ~good;
+
+ good &= constant_time_lt(threes_in_row, 8);
+ err = constant_time_select_int(mask | good, err,
+ RSA_R_SSLV3_ROLLBACK_ATTACK);
+ mask = ~good;
+
+ /*
+ * Skip the zero byte. This is incorrect if we never found a zero-byte
+ * but in this case we also do not copy the message out.
+ */
+ msg_index = zero_index + 1;
+ mlen = num - msg_index;
+
+ /*
+ * For good measure, do this check in constant time as well.
+ */
+ good &= constant_time_ge(tlen, mlen);
+ err = constant_time_select_int(mask | good, err, RSA_R_DATA_TOO_LARGE);
+
+ /*
+ * Even though we can't fake result's length, we can pretend copying
+ * |tlen| bytes where |mlen| bytes would be real. Last |tlen| of |num|
+ * bytes are viewed as circular buffer with start at |tlen|-|mlen'|,
+ * where |mlen'| is "saturated" |mlen| value. Deducing information
+ * about failure or |mlen| would take attacker's ability to observe
+ * memory access pattern with byte granularity *as it occurs*. It
+ * should be noted that failure is indistinguishable from normal
+ * operation if |tlen| is fixed by protocol.
+ */
+ tlen = constant_time_select_int(constant_time_lt(num, tlen), num, tlen);
+ msg_index = constant_time_select_int(good, msg_index, num - tlen);
+ mlen = num - msg_index;
+ for (from += msg_index, mask = good, i = 0; i < tlen; i++) {
+ unsigned int equals = constant_time_eq(i, mlen);
+
+ from -= tlen & equals; /* if (i == mlen) rewind */
+ mask &= mask ^ equals; /* if (i == mlen) mask = 0 */
+ to[i] = constant_time_select_8(mask, from[i], to[i]);
}
- memcpy(to, p, (unsigned int)j);
- return (j);
+ OPENSSL_cleanse(em, num);
+ OPENSSL_free(em);
+ RSAerr(RSA_F_RSA_PADDING_CHECK_SSLV23, err);
+ err_clear_last_constant_time(1 & good);
+
+ return constant_time_select_int(good, mlen, -1);
}
diff --git a/doc/crypto/RSA_padding_add_PKCS1_type_1.pod b/doc/crypto/RSA_padding_add_PKCS1_type_1.pod
index f20f815..800e777 100644
--- a/doc/crypto/RSA_padding_add_PKCS1_type_1.pod
+++ b/doc/crypto/RSA_padding_add_PKCS1_type_1.pod
@@ -109,7 +109,12 @@ L<ERR_get_error(3)|ERR_get_error(3)>.
The RSA_padding_check_PKCS1_type_2() padding check leaks timing
information which can potentially be used to mount a Bleichenbacher
padding oracle attack. This is an inherent weakness in the PKCS #1
-v1.5 padding design. Prefer PKCS1_OAEP padding.
+v1.5 padding design. Prefer PKCS1_OAEP padding. Otherwise it can
+be recommended to pass zero-padded B<f>, so that B<fl> equals to
+B<rsa_len>, and if fixed by protocol, B<tlen> being set to the
+expected length. In such case leakage would be minimal, it would
+take attacker's ability to observe memory access pattern with byte
+granilarity as it occurs, post-factum timing analysis won't do.
=head1 SEE ALSO
More information about the openssl-commits
mailing list