esp-idf/components/mbedtls/port/esp32/bignum.c
Mahavir Jain 84b0254fbf
esp32: mpi: add workaround for data corruption issue observed with IDF 5.x toolchain
This fix adds a workaround to disable compiler optimization flag "-ftree-loop-distribute-patterns"
for `mpi_to_mem_block` routine. It was observed that compiler with release configuration was falling
back to `memset` call from ROM library causing an issue in correctly zero initializing MPI peripheral
block.

Please see following linked issue for more discussion and context on this issue.

Closes https://github.com/espressif/esp-idf/issues/8710
Closes https://github.com/espressif/esp-idf/issues/9371
Closes https://github.com/espressif/esp-idf/issues/9256
Closes IDFGH-7102
Closes IDFGH-7842
Closes IDFGH-7714
Closes IDFCI-1452
Closes IDF-6029
2022-10-27 09:54:26 +05:30

302 lines
9.3 KiB
C

/*
* Multi-precision integer library
* ESP32 hardware accelerated parts based on mbedTLS implementation
*
* SPDX-FileCopyrightText: The Mbed TLS Contributors
*
* SPDX-License-Identifier: Apache-2.0
*
* SPDX-FileContributor: 2016-2022 Espressif Systems (Shanghai) CO LTD
*/
#include "soc/hwcrypto_periph.h"
#include "soc/dport_reg.h"
#include "esp_private/periph_ctrl.h"
#include <mbedtls/bignum.h>
#include "bignum_impl.h"
#include <sys/param.h>
#include <sys/lock.h>
static _lock_t mpi_lock;
/* Round up number of words to nearest
512 bit (16 word) block count.
*/
size_t esp_mpi_hardware_words(size_t words)
{
return (words + 0xF) & ~0xF;
}
void esp_mpi_enable_hardware_hw_op( void )
{
/* newlib locks lazy initialize on ESP-IDF */
_lock_acquire(&mpi_lock);
/* Enable RSA hardware */
periph_module_enable(PERIPH_RSA_MODULE);
DPORT_REG_CLR_BIT(DPORT_RSA_PD_CTRL_REG, DPORT_RSA_PD);
while (DPORT_REG_READ(RSA_CLEAN_REG) != 1)
{ }
// Note: from enabling RSA clock to here takes about 1.3us
}
void esp_mpi_disable_hardware_hw_op( void )
{
DPORT_REG_SET_BIT(DPORT_RSA_PD_CTRL_REG, DPORT_RSA_PD);
/* Disable RSA hardware */
periph_module_disable(PERIPH_RSA_MODULE);
_lock_release(&mpi_lock);
}
void esp_mpi_interrupt_enable( bool enable )
{
DPORT_REG_WRITE(RSA_INTERRUPT_REG, enable);
}
void esp_mpi_interrupt_clear( void )
{
DPORT_REG_WRITE(RSA_CLEAR_INTERRUPT_REG, 1);
}
/* Copy mbedTLS MPI bignum 'mpi' to hardware memory block at 'mem_base'.
If hw_words is higher than the number of words in the bignum then
these additional words will be zeroed in the memory buffer.
*/
/* Please see detailed note inside the function body below.
* Relevant: https://github.com/espressif/esp-idf/issues/8710 and IDF-6029
*/
static inline void __attribute__((optimize("-fno-tree-loop-distribute-patterns")))
mpi_to_mem_block(uint32_t mem_base, const mbedtls_mpi *mpi, size_t hw_words)
{
uint32_t *pbase = (uint32_t *)mem_base;
uint32_t copy_words = MIN(hw_words, mpi->MBEDTLS_PRIVATE(n));
/* Copy MPI data to memory block registers */
for (uint32_t i = 0; i < copy_words; i++) {
pbase[i] = mpi->MBEDTLS_PRIVATE(p[i]);
}
/* Zero any remaining memory block data */
for (uint32_t i = copy_words; i < hw_words; i++) {
pbase[i] = 0;
}
#if _INTERNAL_DEBUG_PURPOSE
/*
* With Xtensa GCC 11.2.0 (from ESP-IDF v5.x), it was observed that above zero initialization
* loop gets optimized to `memset` call from the ROM library. This was causing an issue that
* specific write (store) operation to the MPI peripheral block was getting lost erroneously.
* Following data re-verify loop could catch it during runtime.
*
* As a workaround, we are disabling loop distribute patterns for this function and hence
* compiler does not enforce usage of `memset` (or `memcpy`) calls for this routine. It
* appears that `-ftree-loop-distribute-patterns` was enabled with O2/Os starting from
* GCC-10.x. It is quite possible that there is some issue with DPORT write with sequence of
* store instructions as generated by `memset` call, but for now this should serve as good
* interim workaround without any impact on the performance.
*
* Please see IDF-6029 for more details.
*/
//for (uint32_t i = copy_words; i < hw_words; i++) { assert(pbase[i] == 0); }
#endif
}
/* Read mbedTLS MPI bignum back from hardware memory block.
Reads num_words words from block.
Bignum 'x' should already be grown to at least num_words by caller (can be done while
calculation is in progress, to save some cycles)
*/
static inline void mem_block_to_mpi(mbedtls_mpi *x, uint32_t mem_base, size_t num_words)
{
assert(x->MBEDTLS_PRIVATE(n) >= num_words);
/* Copy data from memory block registers */
esp_dport_access_read_buffer(x->MBEDTLS_PRIVATE(p), mem_base, num_words);
/* Zero any remaining limbs in the bignum, if the buffer is bigger
than num_words */
for (size_t i = num_words; i < x->MBEDTLS_PRIVATE(n); i++) {
x->MBEDTLS_PRIVATE(p[i]) = 0;
}
}
/* Begin an RSA operation. op_reg specifies which 'START' register
to write to.
*/
static inline void start_op(uint32_t op_reg)
{
/* Clear interrupt status */
DPORT_REG_WRITE(RSA_INTERRUPT_REG, 1);
/* Note: above REG_WRITE includes a memw, so we know any writes
to the memory blocks are also complete. */
DPORT_REG_WRITE(op_reg, 1);
}
/* Wait for an RSA operation to complete.
*/
static inline void wait_op_complete(void)
{
while (DPORT_REG_READ(RSA_INTERRUPT_REG) != 1)
{ }
/* clear the interrupt */
DPORT_REG_WRITE(RSA_INTERRUPT_REG, 1);
}
/* Read result from last MPI operation */
void esp_mpi_read_result_hw_op(mbedtls_mpi *Z, size_t z_words)
{
wait_op_complete();
mem_block_to_mpi(Z, RSA_MEM_Z_BLOCK_BASE, z_words);
}
/* Z = (X * Y) mod M */
void esp_mpi_mul_mpi_mod_hw_op(const mbedtls_mpi *X, const mbedtls_mpi *Y, const mbedtls_mpi *M, const mbedtls_mpi *Rinv, mbedtls_mpi_uint Mprime, size_t hw_words)
{
/* Load M, X, Rinv, Mprime (Mprime is mod 2^32) */
mpi_to_mem_block(RSA_MEM_M_BLOCK_BASE, M, hw_words);
mpi_to_mem_block(RSA_MEM_X_BLOCK_BASE, X, hw_words);
mpi_to_mem_block(RSA_MEM_RB_BLOCK_BASE, Rinv, hw_words);
DPORT_REG_WRITE(RSA_M_DASH_REG, (uint32_t)Mprime);
/* "mode" register loaded with number of 512-bit blocks, minus 1 */
DPORT_REG_WRITE(RSA_MULT_MODE_REG, (hw_words / 16) - 1);
/* Execute first stage montgomery multiplication */
start_op(RSA_MULT_START_REG);
wait_op_complete();
/* execute second stage */
/* Load Y to X input memory block, rerun */
mpi_to_mem_block(RSA_MEM_X_BLOCK_BASE, Y, hw_words);
start_op(RSA_MULT_START_REG);
}
/* Z = X * Y */
void esp_mpi_mul_mpi_hw_op(const mbedtls_mpi *X, const mbedtls_mpi *Y, size_t hw_words)
{
/* Copy X (right-extended) & Y (left-extended) to memory block */
mpi_to_mem_block(RSA_MEM_X_BLOCK_BASE, X, hw_words);
mpi_to_mem_block(RSA_MEM_Z_BLOCK_BASE + hw_words * 4, Y, hw_words);
/* NB: as Y is left-extended, we don't zero the bottom words_mult words of Y block.
This is OK for now because zeroing is done by hardware when we do esp_mpi_acquire_hardware().
*/
DPORT_REG_WRITE(RSA_M_DASH_REG, 0);
/* "mode" register loaded with number of 512-bit blocks in result,
plus 7 (for range 9-12). (this is ((N~ / 32) - 1) + 8))
*/
DPORT_REG_WRITE(RSA_MULT_MODE_REG, ((hw_words * 2) / 16) + 7);
start_op(RSA_MULT_START_REG);
}
int esp_mont_hw_op(mbedtls_mpi *Z, const mbedtls_mpi *X, const mbedtls_mpi *Y, const mbedtls_mpi *M,
mbedtls_mpi_uint Mprime,
size_t hw_words,
bool again)
{
// Note Z may be the same pointer as X or Y
int ret = 0;
// montgomery mult prepare
if (again == false) {
mpi_to_mem_block(RSA_MEM_M_BLOCK_BASE, M, hw_words);
DPORT_REG_WRITE(RSA_M_DASH_REG, Mprime);
DPORT_REG_WRITE(RSA_MULT_MODE_REG, hw_words / 16 - 1);
}
mpi_to_mem_block(RSA_MEM_X_BLOCK_BASE, X, hw_words);
mpi_to_mem_block(RSA_MEM_RB_BLOCK_BASE, Y, hw_words);
start_op(RSA_MULT_START_REG);
Z->MBEDTLS_PRIVATE(s) = 1; // The sign of Z will be = M->s (but M->s is always 1)
MBEDTLS_MPI_CHK( mbedtls_mpi_grow(Z, hw_words) );
wait_op_complete();
/* Read back the result */
mem_block_to_mpi(Z, RSA_MEM_Z_BLOCK_BASE, hw_words);
/* from HAC 14.36 - 3. If Z >= M then Z = Z - M */
if (mbedtls_mpi_cmp_mpi(Z, M) >= 0) {
MBEDTLS_MPI_CHK(mbedtls_mpi_sub_mpi(Z, Z, M));
}
cleanup:
return ret;
}
/* Special-case of mbedtls_mpi_mult_mpi(), where we use hardware montgomery mod
multiplication to calculate an mbedtls_mpi_mult_mpi result where either
A or B are >2048 bits so can't use the standard multiplication method.
Result (z_words, based on A bits + B bits) must still be less than 4096 bits.
This case is simpler than the general case modulo multiply of
esp_mpi_mul_mpi_mod() because we can control the other arguments:
* Modulus is chosen with M=(2^num_bits - 1) (ie M=R-1), so output
isn't actually modulo anything.
* Mprime and Rinv are therefore predictable as follows:
Mprime = 1
Rinv = 1
(See RSA Accelerator section in Technical Reference for more about Mprime, Rinv)
*/
void esp_mpi_mult_mpi_failover_mod_mult_hw_op(const mbedtls_mpi *X, const mbedtls_mpi *Y, size_t num_words)
{
size_t hw_words = num_words;
/* M = 2^num_words - 1, so block is entirely FF */
for (size_t i = 0; i < hw_words; i++) {
DPORT_REG_WRITE(RSA_MEM_M_BLOCK_BASE + i * 4, UINT32_MAX);
}
/* Mprime = 1 */
DPORT_REG_WRITE(RSA_M_DASH_REG, 1);
/* "mode" register loaded with number of 512-bit blocks, minus 1 */
DPORT_REG_WRITE(RSA_MULT_MODE_REG, (hw_words / 16) - 1);
/* Load X */
mpi_to_mem_block(RSA_MEM_X_BLOCK_BASE, X, hw_words);
/* Rinv = 1, write first word */
DPORT_REG_WRITE(RSA_MEM_RB_BLOCK_BASE, 1);
/* Zero out rest of the Rinv words */
for (size_t i = 1; i < hw_words; i++) {
DPORT_REG_WRITE(RSA_MEM_RB_BLOCK_BASE + i * 4, 0);
}
start_op(RSA_MULT_START_REG);
wait_op_complete();
/* finish the modular multiplication */
/* Load Y to X input memory block, rerun */
mpi_to_mem_block(RSA_MEM_X_BLOCK_BASE, Y, hw_words);
start_op(RSA_MULT_START_REG);
}