arm-trusted-firmware/lib/xlat_tables_v2/xlat_tables_context.c

271 lines
7.2 KiB
C
Raw Normal View History

/*
Read-only xlat tables for BL31 memory This patch introduces a build flag which allows the xlat tables to be mapped in a read-only region within BL31 memory. It makes it much harder for someone who has acquired the ability to write to arbitrary secure memory addresses to gain control of the translation tables. The memory attributes of the descriptors describing the tables themselves are changed to read-only secure data. This change happens at the end of BL31 runtime setup. Until this point, the tables have read-write permissions. This gives a window of opportunity for changes to be made to the tables with the MMU on (e.g. reclaiming init code). No changes can be made to the tables with the MMU turned on from this point onwards. This change is also enabled for sp_min and tspd. To make all this possible, the base table was moved to .rodata. The penalty we pay is that now .rodata must be aligned to the size of the base table (512B alignment). Still, this is better than putting the base table with the higher level tables in the xlat_table section, as that would cost us a full 4KB page. Changing the tables from read-write to read-only cannot be done with the MMU on, as the break-before-make sequence would invalidate the descriptor which resolves the level 3 page table where that very descriptor is located. This would make the translation required for writing the changes impossible, generating an MMU fault. The caches are also flushed. Signed-off-by: Petre-Ionut Tudor <petre-ionut.tudor@arm.com> Change-Id: Ibe5de307e6dc94c67d6186139ac3973516430466
2019-11-07 15:18:03 +00:00
* Copyright (c) 2017-2020, ARM Limited and Contributors. All rights reserved.
*
* SPDX-License-Identifier: BSD-3-Clause
*/
Read-only xlat tables for BL31 memory This patch introduces a build flag which allows the xlat tables to be mapped in a read-only region within BL31 memory. It makes it much harder for someone who has acquired the ability to write to arbitrary secure memory addresses to gain control of the translation tables. The memory attributes of the descriptors describing the tables themselves are changed to read-only secure data. This change happens at the end of BL31 runtime setup. Until this point, the tables have read-write permissions. This gives a window of opportunity for changes to be made to the tables with the MMU on (e.g. reclaiming init code). No changes can be made to the tables with the MMU turned on from this point onwards. This change is also enabled for sp_min and tspd. To make all this possible, the base table was moved to .rodata. The penalty we pay is that now .rodata must be aligned to the size of the base table (512B alignment). Still, this is better than putting the base table with the higher level tables in the xlat_table section, as that would cost us a full 4KB page. Changing the tables from read-write to read-only cannot be done with the MMU on, as the break-before-make sequence would invalidate the descriptor which resolves the level 3 page table where that very descriptor is located. This would make the translation required for writing the changes impossible, generating an MMU fault. The caches are also flushed. Signed-off-by: Petre-Ionut Tudor <petre-ionut.tudor@arm.com> Change-Id: Ibe5de307e6dc94c67d6186139ac3973516430466
2019-11-07 15:18:03 +00:00
#include <arch_helpers.h>
#include <assert.h>
#include <platform_def.h>
#include <common/debug.h>
#include <lib/xlat_tables/xlat_tables_defs.h>
#include <lib/xlat_tables/xlat_tables_v2.h>
#include "xlat_tables_private.h"
/*
* MMU configuration register values for the active translation context. Used
* from the MMU assembly helpers.
*/
uint64_t mmu_cfg_params[MMU_CFG_PARAM_MAX];
/*
* Allocate and initialise the default translation context for the BL image
* currently executing.
*/
REGISTER_XLAT_CONTEXT(tf, MAX_MMAP_REGIONS, MAX_XLAT_TABLES,
PLAT_VIRT_ADDR_SPACE_SIZE, PLAT_PHY_ADDR_SPACE_SIZE);
void mmap_add_region(unsigned long long base_pa, uintptr_t base_va, size_t size,
unsigned int attr)
{
mmap_region_t mm = MAP_REGION(base_pa, base_va, size, attr);
mmap_add_region_ctx(&tf_xlat_ctx, &mm);
}
void mmap_add(const mmap_region_t *mm)
{
mmap_add_ctx(&tf_xlat_ctx, mm);
}
void mmap_add_region_alloc_va(unsigned long long base_pa, uintptr_t *base_va,
size_t size, unsigned int attr)
{
mmap_region_t mm = MAP_REGION_ALLOC_VA(base_pa, size, attr);
mmap_add_region_alloc_va_ctx(&tf_xlat_ctx, &mm);
*base_va = mm.base_va;
}
void mmap_add_alloc_va(mmap_region_t *mm)
{
while (mm->granularity != 0U) {
assert(mm->base_va == 0U);
mmap_add_region_alloc_va_ctx(&tf_xlat_ctx, mm);
mm++;
}
}
#if PLAT_XLAT_TABLES_DYNAMIC
int mmap_add_dynamic_region(unsigned long long base_pa, uintptr_t base_va,
size_t size, unsigned int attr)
{
mmap_region_t mm = MAP_REGION(base_pa, base_va, size, attr);
return mmap_add_dynamic_region_ctx(&tf_xlat_ctx, &mm);
}
int mmap_add_dynamic_region_alloc_va(unsigned long long base_pa,
uintptr_t *base_va, size_t size,
unsigned int attr)
{
mmap_region_t mm = MAP_REGION_ALLOC_VA(base_pa, size, attr);
int rc = mmap_add_dynamic_region_alloc_va_ctx(&tf_xlat_ctx, &mm);
*base_va = mm.base_va;
return rc;
}
int mmap_remove_dynamic_region(uintptr_t base_va, size_t size)
{
return mmap_remove_dynamic_region_ctx(&tf_xlat_ctx,
base_va, size);
}
#endif /* PLAT_XLAT_TABLES_DYNAMIC */
void __init init_xlat_tables(void)
{
assert(tf_xlat_ctx.xlat_regime == EL_REGIME_INVALID);
unsigned int current_el = xlat_arch_current_el();
if (current_el == 1U) {
tf_xlat_ctx.xlat_regime = EL1_EL0_REGIME;
} else if (current_el == 2U) {
tf_xlat_ctx.xlat_regime = EL2_REGIME;
} else {
assert(current_el == 3U);
tf_xlat_ctx.xlat_regime = EL3_REGIME;
}
init_xlat_tables_ctx(&tf_xlat_ctx);
}
int xlat_get_mem_attributes(uintptr_t base_va, uint32_t *attr)
{
return xlat_get_mem_attributes_ctx(&tf_xlat_ctx, base_va, attr);
}
int xlat_change_mem_attributes(uintptr_t base_va, size_t size, uint32_t attr)
{
return xlat_change_mem_attributes_ctx(&tf_xlat_ctx, base_va, size, attr);
}
Read-only xlat tables for BL31 memory This patch introduces a build flag which allows the xlat tables to be mapped in a read-only region within BL31 memory. It makes it much harder for someone who has acquired the ability to write to arbitrary secure memory addresses to gain control of the translation tables. The memory attributes of the descriptors describing the tables themselves are changed to read-only secure data. This change happens at the end of BL31 runtime setup. Until this point, the tables have read-write permissions. This gives a window of opportunity for changes to be made to the tables with the MMU on (e.g. reclaiming init code). No changes can be made to the tables with the MMU turned on from this point onwards. This change is also enabled for sp_min and tspd. To make all this possible, the base table was moved to .rodata. The penalty we pay is that now .rodata must be aligned to the size of the base table (512B alignment). Still, this is better than putting the base table with the higher level tables in the xlat_table section, as that would cost us a full 4KB page. Changing the tables from read-write to read-only cannot be done with the MMU on, as the break-before-make sequence would invalidate the descriptor which resolves the level 3 page table where that very descriptor is located. This would make the translation required for writing the changes impossible, generating an MMU fault. The caches are also flushed. Signed-off-by: Petre-Ionut Tudor <petre-ionut.tudor@arm.com> Change-Id: Ibe5de307e6dc94c67d6186139ac3973516430466
2019-11-07 15:18:03 +00:00
#if PLAT_RO_XLAT_TABLES
/* Change the memory attributes of the descriptors which resolve the address
* range that belongs to the translation tables themselves, which are by default
* mapped as part of read-write data in the BL image's memory.
*
* Since the translation tables map themselves via these level 3 (page)
* descriptors, any change applied to them with the MMU on would introduce a
* chicken and egg problem because of the break-before-make sequence.
* Eventually, it would reach the descriptor that resolves the very table it
* belongs to and the invalidation (break step) would cause the subsequent write
* (make step) to it to generate an MMU fault. Therefore, the MMU is disabled
* before making the change.
*
* No assumption is made about what data this function needs, therefore all the
* caches are flushed in order to ensure coherency. A future optimization would
* be to only flush the required data to main memory.
*/
int xlat_make_tables_readonly(void)
{
assert(tf_xlat_ctx.initialized == true);
#ifdef __aarch64__
if (tf_xlat_ctx.xlat_regime == EL1_EL0_REGIME) {
disable_mmu_el1();
} else if (tf_xlat_ctx.xlat_regime == EL3_REGIME) {
disable_mmu_el3();
} else {
assert(tf_xlat_ctx.xlat_regime == EL2_REGIME);
return -1;
}
/* Flush all caches. */
dcsw_op_all(DCCISW);
#else /* !__aarch64__ */
assert(tf_xlat_ctx.xlat_regime == EL1_EL0_REGIME);
/* On AArch32, we flush the caches before disabling the MMU. The reason
* for this is that the dcsw_op_all AArch32 function pushes some
* registers onto the stack under the assumption that it is writing to
* cache, which is not true with the MMU off. This would result in the
* stack becoming corrupted and a wrong/junk value for the LR being
* restored at the end of the routine.
*/
dcsw_op_all(DC_OP_CISW);
disable_mmu_secure();
#endif
int rc = xlat_change_mem_attributes_ctx(&tf_xlat_ctx,
(uintptr_t)tf_xlat_ctx.tables,
tf_xlat_ctx.tables_num * XLAT_TABLE_SIZE,
MT_RO_DATA | MT_SECURE);
#ifdef __aarch64__
if (tf_xlat_ctx.xlat_regime == EL1_EL0_REGIME) {
enable_mmu_el1(0U);
} else {
assert(tf_xlat_ctx.xlat_regime == EL3_REGIME);
enable_mmu_el3(0U);
}
#else /* !__aarch64__ */
enable_mmu_svc_mon(0U);
#endif
if (rc == 0) {
tf_xlat_ctx.readonly_tables = true;
}
return rc;
}
#endif /* PLAT_RO_XLAT_TABLES */
/*
* If dynamic allocation of new regions is disabled then by the time we call the
* function enabling the MMU, we'll have registered all the memory regions to
* map for the system's lifetime. Therefore, at this point we know the maximum
* physical address that will ever be mapped.
*
* If dynamic allocation is enabled then we can't make any such assumption
* because the maximum physical address could get pushed while adding a new
* region. Therefore, in this case we have to assume that the whole address
* space size might be mapped.
*/
#ifdef PLAT_XLAT_TABLES_DYNAMIC
#define MAX_PHYS_ADDR tf_xlat_ctx.pa_max_address
#else
#define MAX_PHYS_ADDR tf_xlat_ctx.max_pa
#endif
#ifdef __aarch64__
void enable_mmu_el1(unsigned int flags)
{
setup_mmu_cfg((uint64_t *)&mmu_cfg_params, flags,
tf_xlat_ctx.base_table, MAX_PHYS_ADDR,
tf_xlat_ctx.va_max_address, EL1_EL0_REGIME);
enable_mmu_direct_el1(flags);
}
void enable_mmu_el2(unsigned int flags)
{
setup_mmu_cfg((uint64_t *)&mmu_cfg_params, flags,
tf_xlat_ctx.base_table, MAX_PHYS_ADDR,
tf_xlat_ctx.va_max_address, EL2_REGIME);
enable_mmu_direct_el2(flags);
}
void enable_mmu_el3(unsigned int flags)
{
setup_mmu_cfg((uint64_t *)&mmu_cfg_params, flags,
tf_xlat_ctx.base_table, MAX_PHYS_ADDR,
tf_xlat_ctx.va_max_address, EL3_REGIME);
enable_mmu_direct_el3(flags);
}
void enable_mmu(unsigned int flags)
{
switch (get_current_el_maybe_constant()) {
case 1:
enable_mmu_el1(flags);
break;
case 2:
enable_mmu_el2(flags);
break;
case 3:
enable_mmu_el3(flags);
break;
default:
panic();
}
}
#else /* !__aarch64__ */
void enable_mmu_svc_mon(unsigned int flags)
{
setup_mmu_cfg((uint64_t *)&mmu_cfg_params, flags,
tf_xlat_ctx.base_table, MAX_PHYS_ADDR,
tf_xlat_ctx.va_max_address, EL1_EL0_REGIME);
enable_mmu_direct_svc_mon(flags);
}
void enable_mmu_hyp(unsigned int flags)
{
setup_mmu_cfg((uint64_t *)&mmu_cfg_params, flags,
tf_xlat_ctx.base_table, MAX_PHYS_ADDR,
tf_xlat_ctx.va_max_address, EL2_REGIME);
enable_mmu_direct_hyp(flags);
}
#endif /* __aarch64__ */