Skip to content

GitLab

Switch branch/tag
  1. 29 Apr, 2020 1 commit
    • Andre Przywara's avatar
      fdt/wrappers: Replace fdtw_read_cells() implementation · ff4e6c35
      Andre Przywara authored 
      Our fdtw_read_cells() implementation goes to great lengths to
sanity-check every parameter and result, but leaves a big hole open:
The size of the storage the value pointer points at needs to match the
number of cells given. This can't be easily checked at compile time,
since we lose the size information by using a void pointer.
Regardless the current usage of this function is somewhat wrong anyways,
since we use it on single-element, fixed-length properties only, for
which the DT binding specifies the size.
Typically we use those functions dealing with a number of cells in DT
context to deal with *dynamically* sized properties, which depend on
other properties (#size-cells, #clock-cells, ...), to specify the number
of cells needed.

Another problem with the current implementation is the use of
ambiguously sized types (uintptr_t, size_t) together with a certain
expectation about their size. In general there is no relation between
the length of a DT property and the bitness of the code that parses the
DTB: AArch64 code could encounter 32-bit addresses (where the physical
address space is limited to 4GB [1]), while AArch32 code could read
64-bit sized properties (/memory nodes on LPAE systems, [2]).

To make this more clear, fix the potential issues and also align more
with other DT users (Linux and U-Boot), introduce functions to explicitly
read uint32 and uint64 properties. As the other DT consumers, we do this
based on the generic "read array" function.
Convert all users to use either of those two new functions, and make
sure we never use a pointer to anything other than uint32_t or uint64_t
variables directly.

This reveals (and fixes) a bug in plat_spmd_manifest.c, where we write
4 bytes into a uint16_t variable (passed via a void pointer).

Also we change the implementation of the function to better align with
other libfdt users, by using the right types (fdt32_t) and common
variable names (*prop, prop_names).

[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi#n874
[2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm/boot/dts/ecx-2000.dts



Change-Id: I718de960515117ac7a3331a1b177d2ec224a3890
Signed-off-by: default avatarAndre Przywara <andre.przywara@arm.com>
      ff4e6c35
  3. 28 Apr, 2020 2 commits
    • Andre Przywara's avatar
      plat/stm32: Use generic fdt_read_uint32_array() implementation · 52a616b4
      Andre Przywara authored 
      

The device tree parsing code for the STM32 platform is using its own FDT
helper functions, some of them being rather generic.
In particular the existing fdt_read_uint32_array() implementation is now
almost identical to the new generic code in fdt_wrappers.c, so we can
remove the ST specific version and adjust the existing callers.

Compared to the original ST implementation the new version takes a
pointer to the DTB as the first argument, and also swaps the order of
the number of cells and the pointer.

Change-Id: Id06b0f1ba4db1ad1f733be40e82c34f46638551a
Signed-off-by: default avatarAndre Przywara <andre.przywara@arm.com>
      52a616b4
    • Andre Przywara's avatar
      fdt/wrappers: Generalise fdtw_read_array() · 6e3a89f4
      Andre Przywara authored 
      

Currently our fdtw_read_array() implementation requires the length of
the property to exactly match the requested size, which makes it less
flexible for parsing generic device trees.
Also the name is slightly misleading, since we treat the cells of the
array as 32 bit unsigned integers, performing the endianess conversion.

To fix those issues and align the code more with other DT users (Linux
kernel or U-Boot), rename the function to "fdt_read_uint32_array", and
relax the length check to only check if the property covers at least the
number of cells we request.
This also changes the variable names to be more in-line with other DT
users, and switches to the proper data types.

This makes this function more useful in later patches.

Change-Id: Id86f4f588ffcb5106d4476763ecdfe35a735fa6c
Signed-off-by: default avatarAndre Przywara <andre.przywara@arm.com>
      6e3a89f4
  5. 27 Apr, 2020 1 commit
  7. 25 Apr, 2020 1 commit
    • Masahiro Yamada's avatar
      linker_script: move .data section to bl_common.ld.h · caa3e7e0
      Masahiro Yamada authored 
      Move the data section to the common header.

I slightly tweaked some scripts as follows:

[1] bl1.ld.S has ALIGN(16). I added DATA_ALIGN macro, which is 1
    by default, but overridden by bl1.ld.S. Currently, ALIGN(16)
    of the .data section is redundant because commit 41286590


    ("Fix boot failures on some builds linked with ld.lld.") padded
    out the previous section to work around the issue of LLD version
    <= 10.0. This will be fixed in the future release of LLVM, so
    I am keeping the proper way to align LMA.

[2] bl1.ld.S and bl2_el3.ld.S define __DATA_RAM_{START,END}__ instead
    of __DATA_{START,END}__. I put them out of the .data section.

[3] SORT_BY_ALIGNMENT() is missing tsp.ld.S, sp_min.ld.S, and
    mediatek/mt6795/bl31.ld.S. This commit adds SORT_BY_ALIGNMENT()
    for all images, so the symbol order in those three will change,
    but I do not think it is a big deal.

Change-Id: I215bb23c319f045cd88e6f4e8ee2518c67f03692
Signed-off-by: default avatarMasahiro Yamada <yamada.masahiro@socionext.com>
      caa3e7e0
  9. 24 Apr, 2020 3 commits
    • Masahiro Yamada's avatar
      linker_script: move stacks section to bl_common.ld.h · a926a9f6
      Masahiro Yamada authored 
      

The stacks section is the same for all BL linker scripts.

Move it to the common header file.

Change-Id: Ibd253488667ab4f69702d56ff9e9929376704f6c
Signed-off-by: default avatarMasahiro Yamada <yamada.masahiro@socionext.com>
      a926a9f6
    • Masahiro Yamada's avatar
      bl1: remove '.' from stacks section in linker script · b9f7b57d
      Masahiro Yamada authored 
      Only BL1 specifies '.' in the address field of the stacks section.

Commit 4f59d835 ("Make BL1 RO and RW base addresses configurable")
added '.' on purpose but the commit message does not help to understand
why.

This commit gets rid of it in order to factor out the stacks section
into include/common/bl_common.ld.h

I compared the build result for PLAT=qemu.

'aarch64-linux-gnu-nm -n build/qemu/release/bl1/bl1.elf' will change
as follows:

@@ -336,8 +336,8 @@
 000000000e04e0e0 d max_log_level
 000000000e04e0e4 D console_state
 000000000e04e0e5 D __DATA_RAM_END__
-000000000e04e0e5 B __STACKS_START__
 000000000e04e100 b platform_normal_stacks
+000000000e04e100 B __STACKS_START__
 000000000e04f100 b bl1_cpu_context
 000000000e04f100 B __BSS_START__
 000000000e04f100 B __STACKS_END__

After this change, __STACKS_START__ will match to platform_normal_stacks,
and I think it makes more sense.

'aarch64-linux-gnu-objdump -h build/qemu/release/bl1/bl1.elf' will change
as follows:

@@ -9,11 +9,11 @@
                   CONTENTS, ALLOC, LOAD, READONLY, DATA
   2 .data         000000e5  000000000e04e000  0000000000004a60  0001e000  2**4
                   CONTENTS, ALLOC, LOAD, DATA
-  3 stacks        0000101b  000000000e04e0e5  000000000e04e0e5  0001e0e5  2**6
+  3 stacks        00001000  000000000e04e100  0000000000004b45  0001e100  2**6
                   ALLOC
-  4 .bss          000007e0  000000000e04f100  000000000e04f100  0001e0e5  2**5
+  4 .bss          000007e0  000000000e04f100  0000000000004b50  0001f100  2**5
                   ALLOC
-  5 xlat_table    00006000  000000000e050000  000000000e050000  0001e0e5  2**12
+  5 xlat_table    00006000  000000000e050000  0000000000004b45  00020000  2**12
                   ALLOC
   6 coherent_ram  00000000  000000000e056000  000000000e056000  0001f000  2**12
                   CONTENTS

Sandrine pointed me to a useful document [1] to understand why LMAs of
stacks, .bss, and xlat_table section have changed.

Before this patch, they fell into this scenario:
 "If the section has a specific VMA address, then this is used as the
  LMA address as well."

With this commit, the following applies:
 "Otherwise if a memory region can be found that is compatible with the
  current section, and this region contains at least one section, then
  the LMA is set so the difference between the VMA and LMA is the same
  as the difference between the VMA and LMA of the last section in the
  located region."

Anyway, those three sections are not loaded, so the LMA changes will not
be a problem. The size of bl1.bin is still the same.

QEMU still boots successfully with this change.

A good thing is, this fixes the error for the latest LLD. If I use the
mainline LLVM, I see the following error. The alignment check will probably
be included in the LLVM 11 release, so it is better to fix it now.

$ PLAT=qemu CC=clang CROSS_COMPILE=aarch64-linux-gnu-
  [ snip ]
ld.lld: error: address (0xe04e0e5) of section stacks is not a multiple of alignment (64)
make: *** [Makefile:1050: build/qemu/release/bl1/bl1.elf] Error 1

[1]: https://sourceware.org/binutils/docs/ld/Output-Section-LMA.html#Output-Section-LMA



Change-Id: I3d2f3cc2858be8b3ce2eab3812a76d1e0b5f3a32
Signed-off-by: default avatarMasahiro Yamada <yamada.masahiro@socionext.com>
      b9f7b57d
    • Sandrine Bailleux's avatar
      Merge "Provide a hint to power controller for DSU cluster power down" into integration · 1f915222
      Sandrine Bailleux authored
      1f915222
  11. 23 Apr, 2020 4 commits
  13. 22 Apr, 2020 1 commit
  15. 21 Apr, 2020 2 commits
  17. 20 Apr, 2020 2 commits
  19. 17 Apr, 2020 6 commits
  21. 16 Apr, 2020 7 commits
  23. 15 Apr, 2020 6 commits
  25. 14 Apr, 2020 2 commits
  27. 13 Apr, 2020 1 commit
  29. 09 Apr, 2020 1 commit