| Age | Commit message (Collapse) | Author |
|
If the CPU speculates into the RAM we haven't initialized, we can fault
when the ECC controller reports bad data. Didn't have that problem when
blasting all of RAM...
We can use the linker to help us figure out the range of these unused
regions. If we don't need to zero anything, the host emits 0 for the
start and end, skipping the zeroing.
I'm also zeroing the 32 byte reservation at the start of ITCM, along
with any padding introduced before the vector table. With this in place,
the rip_through_ram firmware does not report any exceptions.
|
|
Blasting all of RAM with zeros would wipe ourselves away if we were
executing from RAM. It also wipes out the HAB event log, which seems to
reside in the boot ROM reservation.
We now zero individual memory regions. Similar to the copy loops,
we'll skip the loop if the VMA and LMA are equal. It looks silly to
zero things like bss when we know cortex-m's reset handler will do it;
but, the usual loop doesn't use the 64 bit writes we'd need to preload
ciphers in ITCM and OCRAM.
Tested RAM loading with probe-rs, using an 1170EVK w/ ECC RAM active.
|
|
8 byte alignment means we can perform per-section ECC RAM preloading,
usually requiring 64-bit accesses.
|
|
Testing shows that we must blow fuses in order to activate either ECC
engine. Once the fuse is blown, NXP's boot ROM will enable the ECC
engine. So by the time the reset handler is running, ECC may already
be enabled.
The 32-bit copy loops won't work for initializing ITCM and OCRAM.
Since these bus widths are 64-bit and the ECC is computed over 64
bits, these 32 bit accesses result in a read-modify-write (RMW) on the
underlying interface. If these RAM regions aren't already initialized,
we'll fail an ECC check; the read in RMW compares bogus RAM contents
against a bogus ECC. This seems to dramatically slow down
initialization, or stall it (I've seen the CPU spinning in the boot
ROM).
If we listen to NXP and use 64-bit writes to preload ECC-protected RAM
regions, the 32-bit copy loops will work as expected. Note that we can
still issue 64-bit accesses towards DTCM; they're scattered across the
two 32-bit interfaces. There's an NXP whitepaper [1] describing that.
Once TCM ECC is enabled, we need to tell the TCM interfaces to RMW
smaller accesses. Otherwise, we won't generate correct ciphers. We
also might need to retry operations if there was a correction by the
RAM controller. We can perform these with predicated instructions.
Tested on an 1010EVK, where I used the debugger to show the control
registers weren't touched; and tested on the 1170EVK, with the same
approach showing the expected set bits.
[1]: https://www.nxp.com/docs/en/white-paper/CORTEXM7WP.pdf
[2]: https://www.nxp.com/docs/en/application-note/AN13204.pdf
|
|
The boot ROM is allowed to use allocations of OCRAM as scratch space. If
you interact with the boot ROM -- using it to interact with FlexSPI, for
instance -- then you cannot also use these OCRAM regions. See the system
boot chapter in the reference manual for more info.
The boot ROM typically reserves the start of OCRAM. The size varies by
chip. This commit describes a new section, .bootrom_reservation, at the
start of OCRAM. Users are encouraged to place a MaybeUninit buffer
that's sized for your ROM's requirements.
If users don't place anything in this section, then the new directive is
a no-op. On the other hand, the section is defined before other RAM
regions, so that it takes precedence over other sections users might
want to place in OCRAM.
Tested using the runtime's test suite. I'm testing with the Teensy 4
binary, since the test suite can support features in this build. The
updated test suite shows that the linker will place the boot ROM
reservation before placing the stack, suggesting that we're respecting
section priority.
|
|
To form a secondary image, use in_flash to position it in flash, then
override the default setting that disables the boot header.
With proper configuration, NXP's boot ROM can load a backup image. This
commit reconfigures the 1170EVK board for the experiment.
|
|
The vector tables exported by imxrt-ral are smaller than the default
cortex-m-rt vector table. The tests fail once we enable the rt / device
features. With this change, we expect the MCU-dependent vector table sizes.
|
|
The layout, or assignment of FlexRAM banks to memory regions, is more
precise than the count. When a user specifies the count, we still apply
the same layout as we always have. But now, by default, we prefer the
fuse layout for users who have made no choice.
This commit supports RAM loading with probe-rs. After probe-rs resets
the MCU and uses the fuse FlexRAM layout, firmware can perfectly match
that layout without disrupting the data already copied into RAM by the
debugger.
|
|
|
|
Just a proof-of-concept. The new inspect_elf test seems to show that
we're building the correct image. We need another commit in order to
help the program run in FlexRAM.
|
|
|
|
The runtime previously allowed function placement at address 0 in ITCM.
However, if you ever formed a pointer to the function placed there, it
would look like a null pointer. And you would never be able to call that
function if you relied on null pointer optimization. Also, most MCU
reference manuals (RM) recommend against this placement.
This commit reduces the total capacity of ITCM by 32 bytes, the smallest
possible size of a MPU region. Note that this is greater than the RM's
recommendation of a four byte reservation. It affects all supported
MCUs, except the 1180. If you're so inclined, your MPU could disallow
loads, stores, and execution from this reservation.
Revised unit tests should cover this change. Additionally, you can
manually verify that the ITCM region lengths are reduced by 32 bytes by
opening the linker scripts generated by the ELF test suite.
|
|
|
|
Applications linked through this builder can be placed in a flash
reservation. You'll need some other software to launch these programs,
since they lack the boot header required by the NXP boot ROM.
|
|
We added the `.xip` section to ensure that the reset handler and
pre-init functions would be placed in flash. This commit lets users
place other content into that section.
`.xip` is intended for instructions. The runtime builder will place
these instructions into the same load region as `.text`. However,
there's no pre-`main` relocation.
Aligning the `.xip` and the `.text` section produces more predictable
behavior between GNU's ld and LLVM's lld.
|
|
There's no requirement for 16 byte-aligned read-only data, and this is
the only assertion expecting that requirement. Reducing the requirement
lets the test pass when the image is built with a Rust 1.82 toolchain.
|
|
If you define a runtime, you can call `stack_size_env_override` to
define an optional environment variable checked by the runtime builder.
Same goes for the heap. A user can set these environment variables to
override the runtime's stack / heap size. You can use this package's
examples to try it out; see the updated build script.
There's no default environment variable for either memory region. The
package that defines the runtime needs to opt-in to this feature.
|
|
Haven't fully dug into this one, but GNU ld and LLVM lld have slightly
different section positions when we cross from text to data. We can
relax the asserts to show ordering without strictly requiring an offset.
|
|
Depending on the ordering and contents of program headers, the previous
predicate for "is this the program header for this section?" could
select the wrong header. GNU's ld and LLVM's lld produce that different
header ordering and contents, causing select asserts to fail when using
GNU's linker.
This commit changes how we select the program header, approximating the
way GNU objdump figures the value. This new approach needs more
information from the section header, so I'm changing the API to make it
easier to call the section_header method.
The previous approach was influenced by LLVM objdump. Turns out that
LLVM objdump will also compute the wrong LMA for these binaries when
they're linked with GNU ld. GNU objdump always produces the correct
section LMA, no matter the LLVM or GNU linker.
|
|
Users can define their device configuration data (DCD), and place the
data in the .dcd section. If the .dcd section has content, the entry in
the IVT points at the user's DCD. This plays well with imxrt-dcd.
|
|
|