[GSoC] Ways to make the x86_64 port work with UEFI

Tue May 29 16:26:08 UTC 2018

On Tue, May 29, 2018 at 9:28 PM, Amaan Cheval <amaan.cheval at gmail.com> wrote:
> Noted, thanks a ton for the details! Unrelated to the topic at hand,
> but out of interest, is this the only reading material for further
> details? http://ceur-ws.org/Vol-1697/EWiLi16_12.pdf
>
> In brief: My tests for keeping a libfake.a (compiled without -fpic)
> and a loader.so with a user-appXYZ.c have been successful, but I'm not
> sure if my assumptions hold for all cases. See the details below for
> more.
>
> Next actions:
> - Read more of Linkers and Loaders since it seems to be the only
> detailed resource I've found at this point
> - Experiment with actually using the existing librtems*.a I've got and
> making them boot as a PE UEFI application image
>
> --------------------------------------------------------------------------------
>
> In more detail:
>
> The problem:
> That UEFI needs a relocatable PE file, i.e. one that can function
> regardless of the physical address it's loaded at (no virtual
> addresses that early).
> To build an ELF of that kind, the resources I've seen all build their
> source with -fpic, and then use objcopy to convert the ELF into a
> relocatable PE, with an embedded runtime self-relocator (akin to
> load-time relocation, if I'm understanding correctly).
>
> What Joel suggested seems to be the simplest option - see if not using
> -fpic for _all_ of RTEMS' build system is fine. I think it might be
> from some testing, but I'm not sure if this is conclusive since I need
> to understand the specifics of the entire development process better.
>
> So here's my understanding of the situation at the moment:
>
> - librtems*.a is made up of object files, compiled without -fpic, and
> that should be fine because I believe object files will use RIP
> relative addressing code by default on x64 where it can, and leave the
> rest for link-time relocations to handle. IF this is true, this works
> perfectly for us because all memory accesses and jumps/calls are
> relative.
>
> - We can have a loader.c which acts as the core with the efi_main
> function - compile it with -fpic into loader.so, and then link
> loader.so, librtems*.a, and user-appXYZ.c together to form a
> relocatable ELF, then convert it into a PE using objcopy. Note that
> from what I can tell, the ELF generated from this still has type EXEC,
> not DYN, according to readelf.

Correction: This was a leftover file that I'd forgotten to take out
after renaming a target. Sorry about the confusion. Just using the
"-shared" flag does cause the resulting ELF to be of type DYN.

>
> The concerns I have are about my assumptions; if GCC generates any
> code that uses absolute addressing and that is resolved as a link-time
> relocation, that could be problematic because the final relocatable PE
> may not match up with the resolved absolute address.
>
> My tests with a fake static archive library, and creating a PE have
> been successful, but I'm unsure of how to trigger the relocation
> behavior by the UEFI firmware (i.e. loading the UEFI image at an
> address other than it's preferred one). One idea is to have a UEFI
> application image that loads this test UEFI application image through
> the "LoadImage" function UEFI provides as a service and then to use
> QEMU's monitor / gdb inspection capabilities to see if the address the
> image is loaded at genuinely changes.
>
> If any of you have any resources, that'd be highly appreciated. Some
> resources I'm using so far are:
> - https://eli.thegreenplace.net/2011/11/03/position-independent-code-pic-in-shared-libraries/
> - https://eli.thegreenplace.net/2011/08/25/load-time-relocation-of-shared-libraries/
> - https://eli.thegreenplace.net/2012/01/03/understanding-the-x64-code-models
>
> Sorry about the length of the email!
>
> On Fri, May 25, 2018 at 10:51 PM, Joel Sherrill <joel at rtems.org> wrote:
>>
>>
>> On Fri, May 25, 2018, 12:11 PM Amaan Cheval <amaan.cheval at gmail.com> wrote:
>>>
>>> Hey! Could you link me to some code that you used for the Deos setup
>>> you mentioned?
>>> My understanding is that the -shared option can link static archives
>>> to create a "shared" library in the sense that it doesn't include the
>>> usual crt0 runtime environment and whatnot, but the code within is
>>> still position-dependent. Given that the PE image that EFI needs is
>>> one that needs to be truly relocatable, this may not work - BUT, I've
>>> only just noticed the ./gnuefi/reloc_x86_64.c file which seems to
>>> handle some kinds of runtime relocations encoded within the converted
>>> PE file, so maybe this will work after all. I'll continue to
>>> investigate and let you know how it goes!
>>
>>
>> Deos isn't a good example except that you can compile with -fPIC and put
>> that code into a static library. Deos is a closed source Level A (man rated
>> flight) ARINC 653 RTOS. It's boot process reads configuration information
>> about each partition and associates .so's with each address space per the
>> configuration. It can't change after that.
>>
>> The RTEMS exe is mostly linked as normal except to use some arguments to say
>> some symbols are from a shared library.
>>
>> The base address of the exe is that of the provided virtual address space
>> with .data and .bss in their respective spaces.
>>
>> And our entry point is in C so there is no asm before that. Great
>> simplification.
>>
>>>
>>> Regarding how TLS differs with PIC - could you elaborate? Is it
>>> something we'll need to solve for if we go with the -fPIC option, or
>>> is it something we need to keep in mind as a limitation, but isn't
>>> really a blocker?
>>
>>
>> I don't think PIC changes the TLS mechanism on arm or PowerPC but on i386,
>> when not PIC the TLS base is in %gs and it's a subroutine call when PIC.
>> Just as well for Deos since they assume an application won't change the
>> segment register values.
>>
>> Other than this one TLS difference, it is a normal exe to me. They just
>> magically provide their .so's before we run.
>>>
>>>
>>> On Fri, May 25, 2018 at 10:13 PM, Joel Sherrill <joel at rtems.org> wrote:
>>> >
>>> >
>>> > On Fri, May 25, 2018, 11:15 AM Amaan Cheval <amaan.cheval at gmail.com>
>>> > wrote:
>>> >>
>>> >> Hey!
>>> >>
>>> >> Skippable details about how FreeBSD handles the UEFI boot process!
>>> >>
>>> >>
>>> >> --------------------------------------------------------------------------------
>>> >>
>>> >> Having looked into it a bit more, my understanding of how FreeBSD
>>> >> handles this process is:
>>> >> - They build a two-stage bootloader for EFI, called boot1.efi and
>>> >> loader.efi[1]
>>> >> - loader.efi is an interactive prompt which may autoboot, or a "boot
>>> >> kernelImg" command can be used to load the actual kernel
>>> >> - The kernel is loaded as an ELF through helper functions. The
>>> >> command_boot[2] function drives this:
>>> >>   - In brief, through calls go through:
>>> >>     command_boot -> mod_loadkld -> file_load ->
>>> >> file_formats[i]->l_load (actually the loadfile function in
>>> >> load_elf.c[3])
>>> >>   - The loadfile function parses the program and section headers of
>>> >> the ELF file (through more function detours that are not really
>>> >> important)
>>> >>   - Once the ELF has been loaded at the correct entry_addr that it
>>> >> expects to be loaded at in memory, the l_exec[4] function is called,
>>> >> which is actually elf64_exec in elf64_freebsd.c[5], at which hopefully
>>> >> through trampolining magic, the control flow will transfer to the
>>> >> kernel or ELF module
>>> >>
>>> >>
>>> >>
>>> >> --------------------------------------------------------------------------------
>>> >>
>>> >> What this means for RTEMS if we go with gnu-efi is essentially 2
>>> >> options, given that the objcopy method of converting from ELF->PE
>>> >> requires the ELF to be a position-independent shared library:
>>> >>
>>> >> - Using -fPIC to compile all of RTEMS, including the RTEMS user's
>>> >> application code. This way we'd have librtemsbso.so, librtemscpu.so,
>>> >> etc. which would then be linked into user_app.c through -fPIC and
>>> >> -shared flags still, creating one singular hello.so, which can then
>>> >> finally be converted into hello.efi and put on a FAT filesystem and
>>> >> booted. This seems doable, but I'm fairly concerned about it further
>>> >> complicating our build system and likely being quite singular in its
>>> >> focus on EFI.
>>> >
>>> >
>>> > I'm using PIC on the Deos BSP. RTEMS is still a .a and exes are linked
>>> > with
>>> > our static libraries and Deos .so.
>>> >
>>> > Unless the loader forces something, you can use PIC with no build system
>>> > changes.
>>> >
>>> > note that thread local storage is different on i386 with and without
>>> > PIC.
>>> >>
>>> >>
>>> >> - The FreeBSD way of a (loader.efi) and a hello.exe (ELF64) put on
>>> >> possibly the same partition on the FAT filesystem required for UEFI
>>> >> application images anyway. The loader.efi can find the hello.exe file
>>> >> through perhaps a config file it can read or by having a magic-name
>>> >> like rtems.exe or something. This effectively means we need an ELF
>>> >> dynamic linker / loader (akin to ld.so) within RTEMS' source. I think
>>> >> using FreeBSD's code for this should be fine. One added benefit of
>>> >> this method is that librtems* and user applications remain as ELF64s,
>>> >> which in the future could also be used with Multiboot with a slightly
>>> >> modified "loader" (i.e. one which generates the apt Multiboot magic
>>> >> header, and boots the PC from 32-bit protected mode to 64-bit long
>>> >> mode).
>>> >>
>>> >> I prefer the latter approach personally. If both of these seem too
>>> >> complicated, we can of course go back to considering generating the PE
>>> >> header format in ASM the way Linux distros use EFISTUB and the code
>>> >> Chris shared (as I mentioned in my original blog post) for wimboot.
>>> >> Those approaches may be significantly simpler in a sense, but may
>>> >> limit how we use UEFI Services - I'm not sure about the details of
>>> >> this yet - I can investigate if y'all aren't fond of the option I laid
>>> >> down above.
>>> >>
>>> >> Let me know!
>>> >>
>>> >> [1]
>>> >>
>>> >> https://www.freebsd.org/cgi/man.cgi?query=loader&apropos=0&sektion=8&manpath=FreeBSD+11.1-RELEASE+and+Ports&arch=default&format=html
>>> >> [2]
>>> >>
>>> >> https://github.com/freebsd/freebsd/blob/433bd38e3a0349f9f89f9d54594172c75b002b74/stand/common/boot.c#L53
>>> >> [3]
>>> >>
>>> >> https://github.com/freebsd/freebsd/blob/d8596f6f687a64b994b065f3058155405dfc39db/stand/common/load_elf.c#L150
>>> >> [4]
>>> >>
>>> >> https://github.com/freebsd/freebsd/blob/433bd38e3a0349f9f89f9d54594172c75b002b74/stand/common/boot.c#L107
>>> >> [5]
>>> >>
>>> >> https://github.com/freebsd/freebsd/blob/d8596f6f687a64b994b065f3058155405dfc39db/stand/efi/loader/arch/amd64/elf64_freebsd.c#L93
>>> >>
>>> >> On Sun, May 20, 2018 at 10:52 PM, Joel Sherrill <joel at rtems.org> wrote:
>>> >> >
>>> >> >
>>> >> > On Sun, May 20, 2018, 12:10 PM Amaan Cheval <amaan.cheval at gmail.com>
>>> >> > wrote:
>>> >> >>
>>> >> >> On Sat, May 19, 2018 at 6:51 PM, Gedare Bloom <gedare at rtems.org>
>>> >> >> wrote:
>>> >> >> > On Fri, May 18, 2018 at 5:53 PM, Joel Sherrill <joel at rtems.org>
>>> >> >> > wrote:
>>> >> >> >>
>>> >> >> >>
>>> >> >> >> On Fri, May 18, 2018 at 3:24 PM, Amaan Cheval
>>> >> >> >> <amaan.cheval at gmail.com>
>>> >> >> >> wrote:
>>> >> >> >>>
>>> >> >> >>> Hi everyone!
>>> >> >> >>>
>>> >> >> >>> I've written a quick blog post summarizing the options I've
>>> >> >> >>> considered
>>> >> >> >>> to make the x86_64 port work with UEFI firmware - the primary
>>> >> >> >>> winner
>>> >> >> >>> seems to be in my eyes to use "gnu-efi" and to add support for
>>> >> >> >>> the
>>> >> >> >>> target "pei-x86-64" (aliased to "efi-app-x86_64") to
>>> >> >> >>> "x86_64-rtems5-objcopy" in binutils. I've submitted a patch for
>>> >> >> >>> this
>>> >> >> >>> here[1].
>>> >> >> >>
>>> >> >> >>
>>> >> >> >> That patch is quite simple so shouldn't be a problem if this is
>>> >> >> >> the
>>> >> >> >> direction
>>> >> >> >> that gets consensus.
>>> >> >> >>>
>>> >> >> >>>
>>> >> >> >>> The blog post is here:
>>> >> >> >>> https://blog.whatthedude.com/post/uefi-app-options/
>>> >> >> >>>
>>> >> >> >>> I'd appreciate all feedback (and please do let me know if I
>>> >> >> >>> haven't
>>> >> >> >>> provided enough context)!
>>> >> >> >>>
>>> >> >> >>> Specifically, some concerns I'd like to discuss are:
>>> >> >> >>>
>>> >> >> >>> - Does everyone agree with me on choosing gnu-efi + objcopy as
>>> >> >> >>> our
>>> >> >> >>> method of choice?
>>> >> >> >>
>>> >> >> >>
>>> >> >> >> Does using gnu-efi add code that runs on the target? Can you
>>> >> >> >> point
>>> >> >> >> us to the files, if so.
>>> >> >>
>>> >> >> Sure. The files would run on the target, yes. These are the ones
>>> >> >> listed here (as linked to in my blog post, perhaps without
>>> >> >> sufficient
>>> >> >> emphasis):
>>> >> >> https://wiki.osdev.org/UEFI#Developing_with_GNU-EFI
>>> >> >>
>>> >> >> >>
>>> >> >> >> Can you tell which approach FreeBSD takes?
>>> >> >>
>>> >> >> FreeBSD takes the gnu-efi approach I see as the "winner" here (also
>>> >> >> a
>>> >> >> link in the post):
>>> >> >>
>>> >> >>
>>> >> >>
>>> >> >> https://github.com/freebsd/freebsd/blob/996b0b6d81cf31cd8d58af5d8b45f0b4945d960d/stand/efi/loader/Makefile#L98-L1
>>> >> >
>>> >> >
>>> >> > This is (no surprise) appropriately licensed and IMO the winning
>>> >> > solution.
>>> >> > Knowing it is what FreeBSD does makes it an easy choice.
>>> >> >
>>> >> > A comment in the readme mentions there is a i386 version of this code
>>> >> > so
>>> >> > that could be used to let pc386 boot from UEFI.
>>> >> >
>>> >> >>
>>> >> >>
>>> >> >> >>
>>> >> >> >>>
>>> >> >> >>> - How do we integrate gnu-efi into our build process? A part of
>>> >> >> >>> the
>>> >> >> >>> RSB, making sure the path to the libraries are in an exported
>>> >> >> >>> variable? Or perhaps a part of the RTEMS kernel itself if the
>>> >> >> >>> licenses
>>> >> >> >>> are compatible (I don't see any on the project[2], only
>>> >> >> >>> copyright
>>> >> >> >>> notices within the source files of the release versions).
>>> >> >> >>
>>> >> >> >>
>>> >> >> >> GNU-efi would be built like qemu or the device tree compiler
>>> >> >> >> would
>>> >> >> >> be my guess and x86_64-rtems toolset might add that to the
>>> >> >> >> standard
>>> >> >> >> set of tools. License on host tools being GPL isn't an issue.
>>> >> >> >>
>>> >> >> >
>>> >> >> > It appears to be a standard 2-clause BSD released by Intel as
>>> >> >> > specified in the README file of gnu-efi.
>>> >> >> >
>>> >> >> >>
>>> >> >> >>>
>>> >> >> >>> - Regardless of how we manage UEFI, do we require Multiboot
>>> >> >> >>> support
>>> >> >> >>> too? Multiboot drops us in a 32-bit protected mode environment,
>>> >> >> >>> whereas 64-bit UEFI firmware will boot us into 64-bit long mode
>>> >> >> >>> -
>>> >> >> >>> this
>>> >> >> >>> would mean the kernel would need to support separate code-paths
>>> >> >> >>> for
>>> >> >> >>> the 2 if we want to support both methods.
>>> >> >> >>
>>> >> >> >>
>>> >> >> >> That's a good question. For GSoC, I think UEFI is fine and
>>> >> >> >> perhaps a
>>> >> >> >> ticket
>>> >> >> >> under the general "modern PC support" ticket for multiboot
>>> >> >> >> support.
>>> >> >> >> Unless
>>> >> >> >> that eliminates a LOT of PCs.
>>> >> >> >>
>>> >> >> >> I don't want you to spend all summer getting an image to boot
>>> >> >> >> both
>>> >> >> >> ways. Personally, I want you to have a working BSP one way. :)
>>> >> >> > +1
>>> >> >> >
>>> >> >>
>>> >> >> Noted, thanks!
>>> >> >>
>>> >> >> >>>
>>> >> >> >>>
>>> >> >> >>> [1] https://www.sourceware.org/ml/binutils/2018-05/msg00197.html
>>> >> >> >>> [2] https://sourceforge.net/projects/gnu-efi/
>>> >> >> >>
>>> >> >> >>
>>> >> >> >> --joel
>>> >> >> >>
>>> >> >> >> _______________________________________________
>>> >> >> >> devel mailing list
>>> >> >> >> devel at rtems.org
>>> >> >> >> http://lists.rtems.org/mailman/listinfo/devel