Re: [PATCH] arm64: lds: move .got section out of .text

[Fangrui Song]

On 2023-04-29, Ard Biesheuvel wrote:
> On Fri, 28 Apr 2023 at 23:29, Fangrui Song <maskray@xxxxxxxxxx> wrote:
> > On 2023-04-28, Ard Biesheuvel wrote:
> > >On Fri, 28 Apr 2023 at 22:06, Fangrui Song <maskray@xxxxxxxxxx> wrote:
> > >>
> > >> On 2023-04-28, Ard Biesheuvel wrote:
> > >> >On Fri, 28 Apr 2023 at 19:58, Fangrui Song <maskray@xxxxxxxxxx> wrote:
> > >> >>
> > >> >> On 2023-04-28, Ard Biesheuvel wrote:
> > >> >> >Hello Fangrui,
> > >> >>
> > >> >> Hello Ard, thank you for the rapid response.
> > >> >>
> > >> >> >On Fri, 28 Apr 2023 at 06:05, Fangrui Song <maskray@xxxxxxxxxx> wrote:
> > >> >> >>
> > >> >> >> Currently, the .got section is placed within the output section .text.
> > >> >> >> However, when .got is non-empty, the SHF_WRITE flag is set when linked
> > >> >> >> by lld. GNU ld recognizes .text as a special section and ignores the
> > >> >> >> SHF_WRITE flag. By renaming .text, we can also get the SHF_WRITE flag.
> > >> >> >>
> > >> >> >> Conventionally, the .got section is placed just before .got.plt (which
> > >> >> >> should be empty and omitted in the kernel). Therefore, we move the .got
> > >> >> >> section to a conventional location (between .text and .data) and remove
> > >> >> >> the unneeded `. = ALIGN(16)`.
> > >> >> >>
> > >> >> >> Signed-off-by: Fangrui Song <maskray@xxxxxxxxxx>
> > >> >> >> ---
> > >> >> >>  arch/arm64/kernel/vmlinux.lds.S | 20 ++++++++++----------
> > >> >> >>  1 file changed, 10 insertions(+), 10 deletions(-)
> > >> >> >>
> > >> >> >> diff --git a/arch/arm64/kernel/vmlinux.lds.S b/arch/arm64/kernel/vmlinux.lds.S
> > >> >> >> index b9202c2ee18e..2bcb3b30db41 100644
> > >> >> >> --- a/arch/arm64/kernel/vmlinux.lds.S
> > >> >> >> +++ b/arch/arm64/kernel/vmlinux.lds.S
> > >> >> >> @@ -181,18 +181,8 @@ SECTIONS
> > >> >> >>                         KPROBES_TEXT
> > >> >> >>                         HYPERVISOR_TEXT
> > >> >> >>                         *(.gnu.warning)
> > >> >> >> -               . = ALIGN(16);
> > >> >> >> -               *(.got)                 /* Global offset table          */
> > >> >> >>         }
> > >> >> >>
> > >> >> >> -       /*
> > >> >> >> -        * Make sure that the .got.plt is either completely empty or it
> > >> >> >> -        * contains only the lazy dispatch entries.
> > >> >> >> -        */
> > >> >> >> -       .got.plt : { *(.got.plt) }
> > >> >> >> -       ASSERT(SIZEOF(.got.plt) == 0 || SIZEOF(.got.plt) == 0x18,
> > >> >> >> -              "Unexpected GOT/PLT entries detected!")
> > >> >> >> -
> > >> >> >>         . = ALIGN(SEGMENT_ALIGN);
> > >> >> >>         _etext = .;                     /* End of text section */
> > >> >> >>
> > >> >> >> @@ -247,6 +237,16 @@ SECTIONS
> > >> >> >>
> > >> >> >>         . = ALIGN(SEGMENT_ALIGN);
> > >> >> >>         __inittext_end = .;
> > >> >> >> +
> > >> >> >> +       .got : { *(.got) }
> > >> >> >
> > >> >> >This is the .init region, which gets freed and unmapped after boot. If
> > >> >> >the GOT is non-empty, it needs to remain mapped, so we cannot place it
> > >> >> >here.
> > >> >>
> > >> >> Thanks.  I did not know the constraint.
> > >> >>
> > >> >> >We have the same issue with the .rodata section, which incorporates
> > >> >> >variables marked as __ro_after_init, which are not const qualified. So
> > >> >> >given that .rodata is already emitted as WA, and we cannot do anything
> > >> >> >about that, let's move the GOT in there.
> > >> >>
> > >> >> Yes, writable .data..ro_after_init and __jump_table sections in
> > >> >> include/asm-generic/vmlinux.lds.h (#define RO_DATA(align)) makes the
> > >> >> output section .rodata writable.  Perhaps this is very difficult to fix,
> > >> >> and we will have writable .rodata for a long time.
> > >> >>
> > >> >> What do you think of moving .got/.got.plt immediately before .data?
> > >> >> I want to place .got/.got.plt before the guaranteed-writable sections,
> > >> >> not some sections which are "unfortunately" writable (.rodata, __modver,
> > >> >> .hyp.rodata, .rodata.text, etc).
> > >> >>
> > >> >> For userspace programs, either linked with GNU ld or lld, .got/.got.plt
> > >> >> are usually immediately before .data .
> > >> >>
> > >> >
> > >> >I don't think that would be the right choice.
> > >> >
> > >> >We have five pseudo-segments in the kernel
> > >> >
> > >> >text (RX)
> > >> >rodata (R)
> > >> >inittext (RX)
> > >> >initdata (RW)
> > >> >data (RW)
> > >> >
> > >> >where the init ones disappear entirely when the boot completes.
> > >> >
> > >> >The GOT should not be modifiable, so it should not be in .data. So the
> > >> >only appropriate 'segment' for the GOT is rodata
> > >> >
> > >> >Note that we don't use PIC codegen for the kernel, so all const
> > >> >qualified data structures containing statically initialized global
> > >> >pointer variables are emitted into .rodata as well, and relocated at
> > >> >boot. So having the GOT in rodata too makes sense imho.
> > >>
> > >> arm64's vmlinux is linked with -shared -Bsymbolic and .got has many
> > >> entries that need to be relocated.
> > >
> > >Actually, it is mostly empty, given that we don't use -fpic/-fpie codegen.
> >
> > Yes, .got is usually small, e.g. 0x68 in my `make ARCH=arm64 LLVM=1 -j60 defconfig vmlinux` build.
> >
> > >>  .got is initially writable, and
> > >> becomes read-only after relocation resolving.
> > >
> > >The same applies to .rodata, given that we don't use -fpic/fpie
> > >codegen. And note that I am not referring to .data..ro_after_init here
> > >- I mean .rodata itself, which carries all const qualified global data
> > >structures with statically initialized pointer fields. These are
> > >conceptually the same as GOT entries - the only difference is that
> > >they are not created by the linker and are referenced explicitly from
> > >the C code.
> >
> > I agree that they are very similar, but there is a distinction, that's
> > why we have both .rodata and .data.rel.ro (relro), and the latter is for sections
> > that need relocations.
>
> Yes, but we don't use .data.rel.ro in the kernel because we don't
> compile with -fpic/-fpie
> > >> I think this property is significant enough to make it outside of any rodata
> > >> segment.
> > >>
> > >
> > >I disagree.
> > >
> > >> (In the userspace .got is typically placed in PT_GNU_RELRO by using -z relro.)
> > >>
> > >> If we remove inittext and initdata pseudo-segments which disappear after
> > >> the boot, we have
> > >>
> > >> text (RX)
> > >> rodata (R)
> > >> /// where is .got (RELRO)
> > >> data (RW)
> > >>
> > >> If we consider .got distinct from rodata, I think placing .got immediately
> > >> after rodata or immediately before data is both fine?
> > >>
> > >
> > >We can place it anywhere in the rodata pseudo-segment. What user space
> > >does is not quite relevant here - what is relevant is that we already
> > >have a pseudo-segment with the right properties. Adding a new segment
> > >just for the GOT which is almost empty and is mapped with the same
> > >attributes as rodata seems unnecessary to me.
> > >
> > >Do you have any practical concerns? Or is it just tidiness?
> >
> > Unfortunately, vmlinux has many PT_LOAD program headers. I am trying to
> > find the best position for .got that makes the most sense. When making
> > this decision, I aim to create a layout that looks more normal and is
> > closer to user-space programs. Additionally, I want to reduce the number
> > of PT_LOAD program headers.
>
> Why? The PT_LOAD headers are never consumed by anything. On arm64, we
> never consume the ELF - it is converted to a flat binary during the
> build.

See below.

> > You are arguing that .got should be part of rodata. However, I disagree
> > with this perspective from both practical and linker layout viewpoints.
> >
> > If we place .got below RO_DATA(PAGE_SIZE), it will immediately follow
> > .note or .BTF (see NOTES in include/asm-generic/vmlinux.lds.h).
> > There isn't an option to drop SHF_WRITE from .got, and
> > we will need a separate PT_LOAD program header for .got.
> >
> > Alternatively, if we place .got below .hyp.rodata, we can eliminate one
> > PT_LOAD because the transition from .hyp.rodata to .got does not require
> > a separate PT_LOAD. However, having a half-writable .got before
> > .rodata.text feels strange to me.
> >
> >
> > I feel that the region between .text/.rodata and .data is too messy.
> > Placing .got immediately before .data makes the layout look more normal
> > and closer to user space without causing any pessimization.
>
> But we are not in user space. There is no ELF loader at runtime that
> consumes any of this, The entire layout is self-described by the
> linker script using symbols that the code links to directly.
>
> So frankly, if we could just avoid PT_LOAD headers entirely, or add a
> PHDRS {} section that defines a single region, I'd be ok with that.
> But tweaking the layout to reduce the number of PT_LOAD headers seems
> unnecessary to me.
>
> And moving the GOT into a writable area just to make the ELF metadata
> look tidier is out of the question - that could turn into a security
> bug.
> > ---
> >
> > My linker layout viewpoint is that ideally a program's program headers
> > should looke like
> >
> > https://maskray.me/blog/2020-11-15-explain-gnu-linker-options#layout-related
> >
> > R PT_LOAD
> > RX PT_LOAD
> > RW PT_LOAD (overlaps with PT_GNU_RELRO) .got, .data.rel.ro (if present)
> > RW PT_LOAD
> >
> > For historical reason, I can expect swapped R and RX.
> > Sections such as .got and .data.rel.ro are different from both .rodata
> > and .data.
>
> The kernel is remapped a number of times during the boot - in the
> initial mapping everything is read-only-exec except bss, and the
> physical address is used as the virtual address. When it is remapped
> again using the actual virtual addresses, we have two pseudo-segments,
> where text/rodata/inittext are RX and initdata and data are RW. Then,
> another mapping is created where text and inittext are RX, and rodata,
> initdata and data are RW, until finally inittext and initdata are
> unmapped and freed, and rodata is remapped read-only.
>
> In the context of the above, why does it matter what the PT_LOAD headers say?

Thank you for the explanation.  I may have missed something as I don't
know the boot process at all...  If we place .got immediately before the
_data symbol and the .data section, how is .got considered to be in a
writable area?

This part of your reply may be relevant:

> where text/rodata/inittext are RX and initdata and data are RW.

Do you mean that the regions [__initdata_begin,__initdata_end) and [_data, _end)
are RW?

If so, placing .got between __initdata_end and _data should not make
.got writable.  However, .got now appears to belong to an additional
pseudo-segment, which conceptually is bad.

Therefore, moving .got between .rodata.text and `idmap_pg_dir = .;` will
be a better choice. This makes .got part of a read-only pseudo-segment.

---

If my understanding is correct, at one point .got is indeed writable.
However, the kernel has performed R_AARCH64_RELATIVE resolving very
early, and can then assume that .got is read-only.
This is similar to user-space PT_GNU_RELRO, but the period is so
transient that we can essentially consider .got read-only.