[External] Re: [PATCH 3/3] efistub: fix missed the initialization of gp
Ard Biesheuvel
ardb at kernel.org
Thu Mar 7 08:48:49 PST 2024
On Thu, 7 Mar 2024 at 04:19, yunhui cui <cuiyunhui at bytedance.com> wrote:
>
> Hi Ard,
>
> On Thu, Mar 7, 2024 at 12:15 AM Ard Biesheuvel <ardb at kernel.org> wrote:
> >
> > On Wed, 6 Mar 2024 at 16:44, Ard Biesheuvel <ardb at kernel.org> wrote:
> > >
> > > On Wed, 6 Mar 2024 at 16:21, Palmer Dabbelt <palmer at dabbelt.com> wrote:
> > > >
> > > > On Wed, 06 Mar 2024 05:09:07 PST (-0800), Ard Biesheuvel wrote:
> > > > > On Wed, 6 Mar 2024 at 14:02, Ard Biesheuvel <ardb at kernel.org> wrote:
> > > > >>
> > > > >> On Wed, 6 Mar 2024 at 13:34, yunhui cui <cuiyunhui at bytedance.com> wrote:
> > > > >> >
> > > > >> > Hi Ard,
> > > > >> >
> > > > >> > On Wed, Mar 6, 2024 at 5:36 PM Ard Biesheuvel <ardb at kernel.org> wrote:
> > > > >> > >
> > > > >> > > On Wed, 6 Mar 2024 at 09:56, Yunhui Cui <cuiyunhui at bytedance.com> wrote:
> > > > >> > > >
> > > > >> > > > Compared with gcc version 12, gcc version 13 uses the gp
> > > > >> > > > register for compilation optimization, but the efistub module
> > > > >> > > > does not initialize gp.
> > > > >> > > >
> > > > >> > > > Signed-off-by: Yunhui Cui <cuiyunhui at bytedance.com>
> > > > >> > > > Co-Developed-by: Zhipeng Xu <xuzhipeng.1973 at bytedance.com>
> > > > >> > >
> > > > >> > > This needs a sign-off, and your signoff needs to come after.
> > > > >> > >
> > > > >> > > > ---
> > > > >> > > > arch/riscv/kernel/efi-header.S | 11 ++++++++++-
> > > > >> > > > 1 file changed, 10 insertions(+), 1 deletion(-)
> > > > >> > > >
> > > > >> > > > diff --git a/arch/riscv/kernel/efi-header.S b/arch/riscv/kernel/efi-header.S
> > > > >> > > > index 515b2dfbca75..fa17c08c092a 100644
> > > > >> > > > --- a/arch/riscv/kernel/efi-header.S
> > > > >> > > > +++ b/arch/riscv/kernel/efi-header.S
> > > > >> > > > @@ -40,7 +40,7 @@ optional_header:
> > > > >> > > > .long __pecoff_data_virt_end - __pecoff_text_end // SizeOfInitializedData
> > > > >> > > > #endif
> > > > >> > > > .long 0 // SizeOfUninitializedData
> > > > >> > > > - .long __efistub_efi_pe_entry - _start // AddressOfEntryPoint
> > > > >> > > > + .long _efistub_entry - _start // AddressOfEntryPoint
> > > > >> > > > .long efi_header_end - _start // BaseOfCode
> > > > >> > > > #ifdef CONFIG_32BIT
> > > > >> > > > .long __pecoff_text_end - _start // BaseOfData
> > > > >> > > > @@ -121,4 +121,13 @@ section_table:
> > > > >> > > >
> > > > >> > > > .balign 0x1000
> > > > >> > > > efi_header_end:
> > > > >> > > > +
> > > > >> > > > + .global _efistub_entry
> > > > >> > > > +_efistub_entry:
> > > > >> > >
> > > > >> > > This should go into .text or .init.text, not the header.
> > > > >> > >
> > > > >> > > > + /* Reload the global pointer */
> > > > >> > > > + load_global_pointer
> > > > >> > > > +
> > > > >> > >
> > > > >> > > What is supposed to happen here if CONFIG_SHADOW_CALL_STACK=y? The EFI
> > > > >> > > stub Makefile removes the SCS CFLAGS, so the stub will be built
> > > > >> > > without shadow call stack support, which I guess means that it might
> > > > >> > > use GP as a global pointer as usual?
> > > > >> > >
> > > > >> > > > + call __efistub_efi_pe_entry
> > > > >> > > > + ret
> > > > >> > > > +
> > > > >> > >
> > > > >> > > You are returning to the firmware here, but after modifying the GP
> > > > >> > > register. Shouldn't you restore it to its old value?
> > > > >> > There is no need to restore the value of the gp register. Where gp is
> > > > >> > needed, the gp register must first be initialized. And here is the
> > > > >> > entry.
> > > > >> >
> > > > >>
> > > > >> But how should the firmware know that GP was corrupted after calling
> > > > >> the kernel's EFI entrypoint? The EFI stub can return to the firmware
> > > > >> if it encounters any errors while still running in the EFI boot
> > > > >> services.
> > > > >>
> > > > >
> > > > > Actually, I wonder if GP can be modified at all before
> > > > > ExitBootServices(). The EFI timer interrupt is still live at this
> > > > > point, and so the firmware is being called behind your back, and might
> > > > > rely on GP retaining its original value.
> > > >
> > > > [A few of us are talking on IRC as I'm writing this...]
> > > >
> > > > The UEFI spec says "UEFI firmware must neither trust the
> > > > values of tp and gp nor make an assumption of owning the write access to
> > > > these register in any circumstances". It's kind of vague what "UEFI
> > > > firmware" means here, but I think it's reasonable to assume that the
> > > > kernel (and thus the EFI stub) is not included there.
> > > >
> > > > So under that interpretation, the kernel (including the EFI stub) would
> > > > be allowed to overwrite GP with whatever it wants.
> > > >
> > >
> > > OK, so even if the UEFI spec seems to suggest that using GP in EFI
> > > applications such as the Linux EFI stub should be safe, I'd still like
> > > to understand why this change is necessary. The patches you are
> > > reverting are supposed to ensure that a) the compiler does not
> > > generate references that can be relaxed to GP based ones, and b) no
> > > R_RISCV_RELAX relocations are present in any of the code that runs in
> > > the context of the EFI firmware.
> > >
> > > Are you still seeing GP based symbol references? Is there C code that
> > > gets pulled into the EFI stub that uses GP based relocations perhaps?
> > > (see list below). If any of those are implemented in C, they should
> > > not be used by the EFI stub directly unless they are guaranteed to be
> > > uninstrumented and callable at arbitrary offsets other than the one
> > > they were linked to run at.
> > >
> > >
> > > __efistub_memcmp = memcmp;
> > > __efistub_memchr = memchr;
> > > __efistub_memcpy = memcpy;
> > > __efistub_memmove = memmove;
> > > __efistub_memset = memset;
> > > __efistub_strlen = strlen;
> > > __efistub_strnlen = strnlen;
> > > __efistub_strcmp = strcmp;
> > > __efistub_strncmp = strncmp;
> > > __efistub_strrchr = strrchr;
> > > __efistub___memcpy = memcpy;
> > > __efistub___memmove = memmove;
> > > __efistub___memset = memset;
> > > __efistub__start = _start;
> > > __efistub__start_kernel = _start_kernel;
> > >
> > > (from arch/riscv/kernel/image-vars.h)
> >
> > Uhm never mind - these are all gone now, I was looking at a v6.1
> > kernel source tree.
> >
> > So that means that, as far as I can tell, the only kernel C code that
> > executes in the context of the EFI firmware is built with -mno-relax
> > and is checked for the absence of R_RISCV_RELAX relocations. So I fail
> > to see why these changes are needed.
> >
> > Yunhui, could you please explain the reason for this series?
>
> From the logic of binutils, if "__global_pointer$" exists, it is
> possible to use GP for optimization. For RISC-V, "__global_pointer$"
> was introduced in commit "fbe934d69eb7e". Therefore, for the system as
> a whole, we should keep using GP uniformly.
There is no 'system as a whole' that can use GP 'uniformly'
The EFI stub is a separate executable that runs from a different
mapping of memory, in an execution context managed by the firmware. It
happens to be linked into the same executable as the vmlinux kernel.
> The root cause of this
> problem is that GP is not loaded, rather than "On RISC-V, we also
> avoid GP based relocations..." as commit "d2baf8cc82c17" said.
GP is not loaded because in the EFI firmware context, there is no safe
way to rely on it.
> We need
> to address problems head-on, rather than avoid them.
>
So what solution are you proposing for the potential GP conflicts
between the boot loader, the Linux EFI stub and the firmware?
More information about the linux-riscv
mailing list