Merge branch '2021-06-07-arm-cache-cp15-improvements' into next
To quote the series author, Patrick Delaunay:
On STM32MP15x platform we can use OP-TEE, loaded in DDR in a region
protected by a firewall. This region is reserved in the device with
the "no-map" property as defined in the binding file
doc/device-tree-bindings/reserved-memory/reserved-memory.txt.
Sometime the platform boot failed in U-Boot on a Cortex A7 access to
this region (depending of the binary and the issue can change with compiler
version or with code alignment), then the firewall raise an error,
for example:
After investigation, the forbidden access is a speculative request performed
by the Cortex A7 because all the DDR is mapped as MEMORY with CACHEABLE
property.
The issue is solved only when the region reserved by OP-TEE is no more
mapped in U-Boot as it is already done in Linux kernel.
Tested on DK2 board with OP-TEE 3.12 / TF-A 2.4:
With hard-coded address for OP-TEE reserved memory,
the error doesn't occur.
void dram_bank_mmu_setup(int bank)
{
....
for (i = start >> MMU_SECTION_SHIFT;
i < (start >> MMU_SECTION_SHIFT) + (size >> MMU_SECTION_SHIFT);
i++) {
option = DCACHE_DEFAULT_OPTION;
if (i >= 0xde0)
option = INVALID_ENTRY;
set_section_dcache(i, option);
}
}
Just by modifying the test on 0xde0 to 0xdf0, the OP-TEE memory protected
by firewall is mapped cacheable and the error occurs.
I think that it can be a general issue for ARM architecture: the "no-map" tag
of reserved memory in device should be respected by U-Boot if firewall
is configured before U-Boot execution.
But I don't propose a generic solution in
arm/lib/cache-cp15.c:dram_bank_mmu_setup()
because the device tree parsing done in lmb_init_and_reserve() takes a
long time when it is executed without data cache.