fix(newlib): fix CONFIG_LIBC_OPTIMIZED_MISALIGNED_ACCESS for c2/c3/c6/h2/h21

PMP configurations for load and store addresses may
have different permissions (e.g., "R" vs. "RW").

Due to the timing alignment of internal signals, the address
permission check may be incorrectly applied during the second
part of a misaligned access transaction.

As a workaround, insert two instructions (e.g. ADDI/NOP) between
accessing to different memory regions. This spacing avoids the
false permission check caused by signal timing overlap.
This commit is contained in:
Alexey Lapshin
2025-08-17 11:52:05 +07:00
parent 799c800494
commit 9332d2ccd2
14 changed files with 33 additions and 116 deletions
+1 -1
View File
@@ -148,7 +148,7 @@ if(ESP_TEE_BUILD)
rom_linker_script("newlib-nano")
endif()
rom_linker_script("libc")
if(CONFIG_ESP_ROM_HAS_SUBOPTIMAL_NEWLIB_ON_MISALIGNED_MEMORY)
if(CONFIG_ESP_ROM_HAS_SUBOPTIMAL_NEWLIB_ON_MISALIGNED_MEMORY) # TODO IDF-13852: use optimized memcpy for TEE ?
rom_linker_script("libc-suboptimal_for_misaligned_mem")
endif()
endif()
@@ -18,14 +18,9 @@ if(CONFIG_LIBC_NEWLIB)
endif()
if(CONFIG_LIBC_OPTIMIZED_MISALIGNED_ACCESS)
list(APPEND srcs "test_misaligned_mem_performance.c")
list(APPEND srcs "test_misaligned_access.c")
endif()
idf_component_register(SRCS "${srcs}"
PRIV_REQUIRES unity vfs cmock driver esp_timer spi_flash test_utils pthread esp_psram
WHOLE_ARCHIVE)
if(CONFIG_LIBC_OPTIMIZED_MISALIGNED_ACCESS)
set_source_files_properties(test_misaligned_mem_performance.c
PROPERTIES COMPILE_FLAGS "-Wno-incompatible-pointer-types -Wno-strict-prototypes")
endif()
@@ -130,7 +130,7 @@ TEST_CASE("mem functions in IRAM", "[misaligned_mem]")
TEST_ASSERT_TRUE(fn_in_ram(strncmp));
}
#if CONFIG_ESP_SYSTEM_MEMPROT_PMP
#if CONFIG_IDF_TARGET_ARCH_RISCV
TEST_CASE("access across different PMP regions", "[misaligned_mem]")
{
/*
@@ -1,108 +0,0 @@
/*
* SPDX-FileCopyrightText: 2025 Espressif Systems (Shanghai) CO LTD
*
* SPDX-License-Identifier: Unlicense OR CC0-1.0
*/
#include <stdint.h>
#include <string.h>
#include "esp_heap_caps.h"
#include "hal/cpu_ll.h"
#include "unity.h"
#define MAX_MEMTEST_SIZE 4096
uint32_t test_function_dest_src_size(void (*foo)(), bool pass_size)
{
uint32_t ccount1, ccount2;
char* test_des = heap_caps_aligned_alloc(32, MAX_MEMTEST_SIZE, MALLOC_CAP_INTERNAL | MALLOC_CAP_8BIT);
char* test_src = heap_caps_aligned_alloc(32, MAX_MEMTEST_SIZE, MALLOC_CAP_INTERNAL | MALLOC_CAP_8BIT);
TEST_ASSERT_NOT_NULL(test_des);
TEST_ASSERT_NOT_NULL(test_src);
/* Prepare arrays for X-cmp functions to make the algorithm go through whole buffers. */
memset(test_src, 'a', MAX_MEMTEST_SIZE);
test_src[MAX_MEMTEST_SIZE - 1] = 0;
memset(test_des, 'a', MAX_MEMTEST_SIZE);
test_des[MAX_MEMTEST_SIZE - 1] = 'b';
test_des[MAX_MEMTEST_SIZE - 1] = 0;
ccount1 = esp_cpu_get_cycle_count();
if (pass_size) {
foo(test_des + 1, test_src + 2, MAX_MEMTEST_SIZE - 2);
} else {
foo(test_des + 1, test_src + 2);
}
ccount2 = esp_cpu_get_cycle_count();
heap_caps_free(test_des);
heap_caps_free(test_src);
return ccount2 - ccount1;
}
TEST_CASE("memcpy", "[misaligned_mem]")
{
uint32_t ccount = test_function_dest_src_size(memcpy, true);
/* esp32c2: 4128 cycles instead 28676. */
TEST_ASSERT_LESS_THAN(10000, ccount);
}
TEST_CASE("memcmp", "[misaligned_mem]")
{
uint32_t ccount = test_function_dest_src_size(memcmp, true);
/* esp32c2: 14259 cycles instead 49147. */
TEST_ASSERT_LESS_THAN(20000, ccount);
}
TEST_CASE("memmove", "[misaligned_mem]")
{
uint32_t ccount = test_function_dest_src_size(memmove, true);
/* esp32c2: 8086 cycles instead 33896. */
TEST_ASSERT_LESS_THAN(15000, ccount);
}
TEST_CASE("memmove - overlapping", "[misaligned_mem]")
{
uint32_t ccount1, ccount2;
char* buf = heap_caps_aligned_alloc(32, MAX_MEMTEST_SIZE, MALLOC_CAP_INTERNAL | MALLOC_CAP_8BIT);
TEST_ASSERT_NOT_NULL(buf);
ccount1 = esp_cpu_get_cycle_count();
memmove(buf + 5, buf + 2, MAX_MEMTEST_SIZE - 5);
ccount2 = esp_cpu_get_cycle_count();
heap_caps_free(buf);
/* esp32c2: 11503 cycles instead 45024. */
TEST_ASSERT_LESS_THAN(20000, ccount2 - ccount1);
}
TEST_CASE("strcpy", "[misaligned_mem]")
{
uint32_t ccount = test_function_dest_src_size(strcpy, false);
/* esp32c2: 17313 cycles instead 32771. */
TEST_ASSERT_LESS_THAN(22000, ccount);
}
TEST_CASE("strcmp", "[misaligned_mem]")
{
uint32_t ccount = test_function_dest_src_size(strcmp, false);
/* esp32c2: 13191 cycles instead 32775. */
TEST_ASSERT_LESS_THAN(20000, ccount);
}
TEST_CASE("strncpy", "[misaligned_mem]")
{
uint32_t ccount = test_function_dest_src_size(strncpy, true);
/* esp32c2: 21475 cycles instead 36859. */
TEST_ASSERT_LESS_THAN(25000, ccount);
}
TEST_CASE("strncmp", "[misaligned_mem]")
{
uint32_t ccount = test_function_dest_src_size(strncmp, true);
/* esp32c2: 24369 cycles instead 49141. */
TEST_ASSERT_LESS_THAN(30000, ccount);
}
@@ -263,6 +263,10 @@ config SOC_CPU_IDRAM_SPLIT_USING_PMP
bool
default y
config SOC_CPU_MISALIGNED_ACCESS_ON_PMP_MISMATCH_ISSUE
bool
default y
config SOC_ECC_SUPPORT_POINT_VERIFY_QUIRK
bool
default y
@@ -106,6 +106,8 @@
#define SOC_CPU_WATCHPOINT_MAX_REGION_SIZE 0x80000000 // bytes
#define SOC_CPU_IDRAM_SPLIT_USING_PMP 1
// DIG-694: misaligned access across PMP regions must be spaced at least by two instructions
#define SOC_CPU_MISALIGNED_ACCESS_ON_PMP_MISMATCH_ISSUE 1
/*-------------------------- ECC CAPS --------------------------*/
#define SOC_ECC_SUPPORT_POINT_VERIFY_QUIRK 1 // C2 ECC peripheral has a bug in ECC point verification, if value of K is zero the verification fails
@@ -351,6 +351,10 @@ config SOC_CPU_WATCHPOINT_MAX_REGION_SIZE
hex
default 0x80000000
config SOC_CPU_MISALIGNED_ACCESS_ON_PMP_MISMATCH_ISSUE
bool
default y
config SOC_DS_SIGNATURE_MAX_BIT_LEN
int
default 3072
@@ -137,6 +137,8 @@
#define SOC_CPU_BREAKPOINTS_NUM 8
#define SOC_CPU_WATCHPOINTS_NUM 8
#define SOC_CPU_WATCHPOINT_MAX_REGION_SIZE 0x80000000 // bytes
// DIG-694: misaligned access across PMP regions must be spaced at least by two instructions
#define SOC_CPU_MISALIGNED_ACCESS_ON_PMP_MISMATCH_ISSUE 1
/*-------------------------- DIGITAL SIGNATURE CAPS ----------------------------------------*/
/** The maximum length of a Digital Signature in bits. */
@@ -431,6 +431,10 @@ config SOC_CPU_PMP_REGION_GRANULARITY
int
default 4
config SOC_CPU_MISALIGNED_ACCESS_ON_PMP_MISMATCH_ISSUE
bool
default y
config SOC_DS_SIGNATURE_MAX_BIT_LEN
int
default 3072
@@ -158,6 +158,8 @@
#define SOC_CPU_HAS_PMA 1
#define SOC_CPU_IDRAM_SPLIT_USING_PMP 1
#define SOC_CPU_PMP_REGION_GRANULARITY 4
// DIG-694: misaligned access across PMP regions must be spaced at least by two instructions
#define SOC_CPU_MISALIGNED_ACCESS_ON_PMP_MISMATCH_ISSUE 1
// TODO: IDF-5360 (Copy from esp32c3, need check)
/*-------------------------- DIGITAL SIGNATURE CAPS ----------------------------------------*/
@@ -427,6 +427,10 @@ config SOC_CPU_PMP_REGION_GRANULARITY
int
default 4
config SOC_CPU_MISALIGNED_ACCESS_ON_PMP_MISMATCH_ISSUE
bool
default y
config SOC_MMU_PAGE_SIZE_CONFIGURABLE
bool
default y
@@ -174,6 +174,8 @@
#define SOC_CPU_HAS_PMA 1
#define SOC_CPU_IDRAM_SPLIT_USING_PMP 1
#define SOC_CPU_PMP_REGION_GRANULARITY 4
// DIG-694: misaligned access across PMP regions must be spaced at least by two instructions
#define SOC_CPU_MISALIGNED_ACCESS_ON_PMP_MISMATCH_ISSUE 1
/*-------------------------- MMU CAPS ----------------------------------------*/
#define SOC_MMU_PAGE_SIZE_CONFIGURABLE (1)
@@ -255,6 +255,10 @@ config SOC_CPU_PMP_REGION_GRANULARITY
int
default 4
config SOC_CPU_MISALIGNED_ACCESS_ON_PMP_MISMATCH_ISSUE
bool
default y
config SOC_MMU_PERIPH_NUM
int
default 1
@@ -159,6 +159,8 @@
#define SOC_CPU_HAS_PMA 1
#define SOC_CPU_IDRAM_SPLIT_USING_PMP 1
#define SOC_CPU_PMP_REGION_GRANULARITY 4
// DIG-694: misaligned access across PMP regions must be spaced at least by two instructions
#define SOC_CPU_MISALIGNED_ACCESS_ON_PMP_MISMATCH_ISSUE 1
/*-------------------------- MMU CAPS ----------------------------------------*/
// #define SOC_MMU_PAGE_SIZE_CONFIGURABLE (1)