The setup procedure is fairly simple: create a 64-bit GDT and 64-bit page
tables that include all kernel mappings from the 32-bit address space, but at
the correct 64-bit address, then go through kernel_args and changes all virtual
addresses to 64-bit addresses, and finally switch to long mode and jump to the
kernel.
* platform_allocate_elf_region() is removed, it is implemented in platform-
independent code now (ELF*Class::AllocateRegion). For ELF64 it is now
assumed that 64-bit addresses are mapped in the loader's 32-bit address space
as (address - KERNEL_BASE_64BIT + KERNEL_BASE).
* mapped_delta field from preloaded_*_image removed, now handled compile-time
using the ELF*Class::Map method.
* Also link the kernel with -z max-page-size=0x1000, removes the need for
2MB alignment on the data segment (not going to map the kernel with large
pages for the time being).
The ELF loader now uses a new platform function, platform_allocate_elf_region,
which returns 2 addresses: the real load address and an address where the
region is mapped in the loader's address space. All of the ELF loading code
has been changed to access the load region through the mapped address rather
than the addresses contained in the ELF image. The ELF64 version of
platform_allocate_elf_region on x86 uses the existing MMU code, which maps
everything at 0x80000000, but returns the correct 64-bit address. The long
mode switch code will just set up the 64-bit address space with everything
remapped at the correct address.
After enabling BOOT_SUPPORT_ELF64 on x86 the x86 kernel could no longer be
booted because too many kernel_args allocations were taking place and filling
kernel_args_ranges. This was because for each image load it would check if
the image is ELF64, fail and fall back to ELF32 (each ELF64 check allocated
a preloaded_image). Changed elf_load_image() so that it only tries both when
loading the kernel image, and then for modules it will only try the same ELF
class as the kernel image.
The actual implementation of the ELF loading methods have been put into
an ELFLoader template class that takes a single template parameter, which
is a structure containing all the necessary ELF typedefs. It's a bit
verbose, but I thought it was a neater solution than using a bunch of
standalone functions with a huge number of template parameters. There is
no change to code outside of elf.cpp, the ELF32/ELF64 differences are
handled internally.
* There is now 2 structures, preloaded_elf32_image and preloaded_elf64_image,
which both inherit from preloaded_image.
* For now I've just hardcoded in use of preloaded_elf32_image, but the
bootloader ELF code will shortly be converted to use templates which use
the appropriate structure. The kernel will be changed later when I add
ELF64 support to it.
* All kernel_args data is now compatible between 32-bit and 64-bit kernels.
* Added a FixedWidthPointer template class which uses 64-bit storage to hold
a pointer. This is used in place of raw pointers in kernel_args.
* Added __attribute__((packed)) to kernel_args and all structures contained
within it. This is necessary due to different alignment behaviour for
32-bit and 64-bit compilation with GCC.
* With these changes, kernel_args will now come out the same size for both
the x86_64 kernel and the loader, excluding the preloaded_image structure
which has not yet been changed.
* Tested both an x86 GCC2 and GCC4 build, no problems caused by these changes.
I've tested this change on x86, causing no issues. I've checked over the code
for all other platforms and made the necessary changes and to the best of my
knowledge they should also still work, but I haven't actually built and
tested them. Once I've completed the kernel_args changes the other platforms
will need testing.
Pointers in kernel_args are going to be changed to unconditionally use 64-bit
storage (to make kernel_args compatible with both the x86 and x86_64 kernels).
KMessage stores a pointer to its buffer, however since KMessage is used
outside of the boot code it is undesirable to change it to use 64-bit storage
for the pointer as it may add additional overhead on 32-bit builds. Therefore,
only store the buffer address and size and then construct a KMessage from
those in the kernel.
* set_haiku_revision doesn't currently support ELF64, don't use a
revisioned kernel image on x86_64 for now.
* Don't try to build add-ons for x86_64 yet.
The whole kernel now builds and there are no undefined references when
linking, I just need to fix some strange relocation errors I'm getting
(probably a problem with the linker script) and then I'll have a kernel
image.
The red zone is a 128-byte area below the stack pointer specified by the
AMD64 ABI that can be used by leaf functions for their stack frame without
modifying the stack pointer. It is guaranteed not to be modified by signal
handlers. This cannot be used in kernel mode code, as an interrupt handler
could overwrite it, so stop GCC from generating code that uses it.
Since ICI arguments are used to send addresses in some places, uint32 is
not sufficient on x86_64. addr_t still refers to the same type as uint32
(unsigned long) on other platforms, so this change only really affects
x86_64.
* Most of this is incorrect printf format strings. Changed all strings
causing errors to use the B_PRI* format string definitions, which
means the strings should be correct across all platforms.
* Some other fixes for errors, casts required, etc.
* Based on the x86 one, changes made to match the default GCC x86_64 linker
script and correct kernel load address added.
* Not tested yet as the kernel doesn't fully compile.
* x86_64 is using the existing *_ia32 boot platforms.
* Special flags are required when compiling the loader to get GCC to compile
32-bit code. This adds a new set of rules for compiling boot code rather
than using the kernel rules, which compile using the necessary flags.
* Some x86_64 private headers have been stubbed by #include'ing the x86
versions. These will be replaced later.
This appears to be a problem with GCC's build system: it defaults to having
multilib enabled, but if it is explicitly enabled with --enable-multilib,
the build fails.
Signed-off-by: Ingo Weinhold <[email protected]>
Alex Smith