Patchwork [v3] arm64: prevent regressions in compressed kernel image size when upgrading to binutils 2.27

login
register
mail settings
Submitter Nick Desaulniers
Date Oct. 27, 2017, 4:33 p.m.
Message ID <20171027163341.57550-1-ndesaulniers@google.com>
Download mbox | patch
Permalink /patch/10030321/
State New
Headers show

Comments

Nick Desaulniers - Oct. 27, 2017, 4:33 p.m.
Upon upgrading to binutils 2.27, we found that our lz4 and gzip
compressed kernel images were significantly larger, resulting is 10ms
boot time regressions.

As noted by Rahul:
"aarch64 binaries uses RELA relocations, where each relocation entry
includes an addend value. This is similar to x86_64.  On x86_64, the
addend values are also stored at the relocation offset for relative
relocations. This is an optimization: in the case where code does not
need to be relocated, the loader can simply skip processing relative
relocations.  In binutils-2.25, both bfd and gold linkers did this for
x86_64, but only the gold linker did this for aarch64.  The kernel build
here is using the bfd linker, which stored zeroes at the relocation
offsets for relative relocations.  Since a set of zeroes compresses
better than a set of non-zero addend values, this behavior was resulting
in much better lz4 compression.

The bfd linker in binutils-2.27 is now storing the actual addend values
at the relocation offsets. The behavior is now consistent with what it
does for x86_64 and what gold linker does for both architectures.  The
change happened in this upstream commit:
https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=1f56df9d0d5ad89806c24e71f296576d82344613
Since a bunch of zeroes got replaced by non-zero addend values, we see
the side effect of lz4 compressed image being a bit bigger.

To get the old behavior from the bfd linker, "--no-apply-dynamic-relocs"
flag can be used:
$ LDFLAGS="--no-apply-dynamic-relocs" make
With this flag, the compressed image size is back to what it was with
binutils-2.25.

If the kernel is using ASLR, there aren't additional runtime costs to
--no-apply-dynamic-relocs, as the relocations will need to be applied
again anyway after the kernel is relocated to a random address.

If the kernel is not using ASLR, then presumably the current default
behavior of the linker is better. Since the static linker performed the
dynamic relocs, and the kernel is not moved to a different address at
load time, it can skip applying the relocations all over again."

Some measurements:

$ ld -v
GNU ld (binutils-2.25-f3d35cf6) 2.25.51.20141117
                    ^
$ ls -l vmlinux
-rwxr-x--- 1 ndesaulniers eng 300652760 Oct 26 11:57 vmlinux
$ ls -l Image.lz4-dtb
-rw-r----- 1 ndesaulniers eng 16932627 Oct 26 11:57 Image.lz4-dtb

$ ld -v
GNU ld (binutils-2.27-53dd00a1) 2.27.0.20170315
                    ^
pre patch:
$ ls -l vmlinux
-rwxr-x--- 1 ndesaulniers eng 300376208 Oct 26 11:43 vmlinux
$ ls -l Image.lz4-dtb
-rw-r----- 1 ndesaulniers eng 18159474 Oct 26 11:43 Image.lz4-dtb

post patch:
$ ls -l vmlinux
-rwxr-x--- 1 ndesaulniers eng 300376208 Oct 26 12:06 vmlinux
$ ls -l Image.lz4-dtb
-rw-r----- 1 ndesaulniers eng 16932466 Oct 26 12:06 Image.lz4-dtb

By Siqi's measurement w/ gzip:
binutils 2.27 with this patch (with --no-apply-dynamic-relocs):
Image 41535488
Image.gz 13404067

binutils 2.27 without this patch (without --no-apply-dynamic-relocs):
Image 41535488
Image.gz 14125516

Any compression scheme should be able to get better results from the
longer runs of zeros, not just GZIP and LZ4.

10ms boot time savings isn't anything to get excited about, but users of
arm64+compression+bfd-2.27 should not have to pay a penalty for no
runtime improvement.

Reported-by: Gopinath Elanchezhian <gelanchezhian@google.com>
Reported-by: Sindhuri Pentyala <spentyala@google.com>
Reported-by: Wei Wang <wvw@google.com>
Suggested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Suggested-by: Rahul Chaudhry <rahulchaudhry@google.com>
Suggested-by: Siqi Lin <siqilin@google.com>
Suggested-by: Stephen Hines <srhines@google.com>
Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
---
Changes since v2:
* Combine this this LDFLAGS_vmlinux append to previous block, per Ard.
* Add Siqi's measurements to commit message.
* Change the conditional to check if set rather than if not empty, since
  this is similar to the rest of the Makefile, and more readable IMO,
  though that's subjective. That part can be reverted if we don't want to
  steal blame.

 arch/arm64/Makefile | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
Will Deacon - Oct. 30, 2017, 1:08 p.m.
On Fri, Oct 27, 2017 at 09:33:41AM -0700, Nick Desaulniers wrote:
> Upon upgrading to binutils 2.27, we found that our lz4 and gzip
> compressed kernel images were significantly larger, resulting is 10ms
> boot time regressions.
> 
> As noted by Rahul:
> "aarch64 binaries uses RELA relocations, where each relocation entry
> includes an addend value. This is similar to x86_64.  On x86_64, the
> addend values are also stored at the relocation offset for relative
> relocations. This is an optimization: in the case where code does not
> need to be relocated, the loader can simply skip processing relative
> relocations.  In binutils-2.25, both bfd and gold linkers did this for
> x86_64, but only the gold linker did this for aarch64.  The kernel build
> here is using the bfd linker, which stored zeroes at the relocation
> offsets for relative relocations.  Since a set of zeroes compresses
> better than a set of non-zero addend values, this behavior was resulting
> in much better lz4 compression.
> 
> The bfd linker in binutils-2.27 is now storing the actual addend values
> at the relocation offsets. The behavior is now consistent with what it
> does for x86_64 and what gold linker does for both architectures.  The
> change happened in this upstream commit:
> https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=1f56df9d0d5ad89806c24e71f296576d82344613
> Since a bunch of zeroes got replaced by non-zero addend values, we see
> the side effect of lz4 compressed image being a bit bigger.
> 
> To get the old behavior from the bfd linker, "--no-apply-dynamic-relocs"
> flag can be used:
> $ LDFLAGS="--no-apply-dynamic-relocs" make
> With this flag, the compressed image size is back to what it was with
> binutils-2.25.
> 
> If the kernel is using ASLR, there aren't additional runtime costs to
> --no-apply-dynamic-relocs, as the relocations will need to be applied
> again anyway after the kernel is relocated to a random address.
> 
> If the kernel is not using ASLR, then presumably the current default
> behavior of the linker is better. Since the static linker performed the
> dynamic relocs, and the kernel is not moved to a different address at
> load time, it can skip applying the relocations all over again."

Do you have any numbers booting an uncompressed kernel Image without ASLR
to see if skipping the relocs makes a measurable difference there?

Will
Ard Biesheuvel - Oct. 30, 2017, 1:11 p.m.
On 30 October 2017 at 13:08, Will Deacon <will.deacon@arm.com> wrote:
> On Fri, Oct 27, 2017 at 09:33:41AM -0700, Nick Desaulniers wrote:
>> Upon upgrading to binutils 2.27, we found that our lz4 and gzip
>> compressed kernel images were significantly larger, resulting is 10ms
>> boot time regressions.
>>
>> As noted by Rahul:
>> "aarch64 binaries uses RELA relocations, where each relocation entry
>> includes an addend value. This is similar to x86_64.  On x86_64, the
>> addend values are also stored at the relocation offset for relative
>> relocations. This is an optimization: in the case where code does not
>> need to be relocated, the loader can simply skip processing relative
>> relocations.  In binutils-2.25, both bfd and gold linkers did this for
>> x86_64, but only the gold linker did this for aarch64.  The kernel build
>> here is using the bfd linker, which stored zeroes at the relocation
>> offsets for relative relocations.  Since a set of zeroes compresses
>> better than a set of non-zero addend values, this behavior was resulting
>> in much better lz4 compression.
>>
>> The bfd linker in binutils-2.27 is now storing the actual addend values
>> at the relocation offsets. The behavior is now consistent with what it
>> does for x86_64 and what gold linker does for both architectures.  The
>> change happened in this upstream commit:
>> https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=1f56df9d0d5ad89806c24e71f296576d82344613
>> Since a bunch of zeroes got replaced by non-zero addend values, we see
>> the side effect of lz4 compressed image being a bit bigger.
>>
>> To get the old behavior from the bfd linker, "--no-apply-dynamic-relocs"
>> flag can be used:
>> $ LDFLAGS="--no-apply-dynamic-relocs" make
>> With this flag, the compressed image size is back to what it was with
>> binutils-2.25.
>>
>> If the kernel is using ASLR, there aren't additional runtime costs to
>> --no-apply-dynamic-relocs, as the relocations will need to be applied
>> again anyway after the kernel is relocated to a random address.
>>
>> If the kernel is not using ASLR, then presumably the current default
>> behavior of the linker is better. Since the static linker performed the
>> dynamic relocs, and the kernel is not moved to a different address at
>> load time, it can skip applying the relocations all over again."
>
> Do you have any numbers booting an uncompressed kernel Image without ASLR
> to see if skipping the relocs makes a measurable difference there?
>

Do you mean built with ASLR support but executing at the offset it was
linked at?
Will Deacon - Oct. 30, 2017, 1:12 p.m.
On Mon, Oct 30, 2017 at 01:11:23PM +0000, Ard Biesheuvel wrote:
> On 30 October 2017 at 13:08, Will Deacon <will.deacon@arm.com> wrote:
> > On Fri, Oct 27, 2017 at 09:33:41AM -0700, Nick Desaulniers wrote:
> >> Upon upgrading to binutils 2.27, we found that our lz4 and gzip
> >> compressed kernel images were significantly larger, resulting is 10ms
> >> boot time regressions.
> >>
> >> As noted by Rahul:
> >> "aarch64 binaries uses RELA relocations, where each relocation entry
> >> includes an addend value. This is similar to x86_64.  On x86_64, the
> >> addend values are also stored at the relocation offset for relative
> >> relocations. This is an optimization: in the case where code does not
> >> need to be relocated, the loader can simply skip processing relative
> >> relocations.  In binutils-2.25, both bfd and gold linkers did this for
> >> x86_64, but only the gold linker did this for aarch64.  The kernel build
> >> here is using the bfd linker, which stored zeroes at the relocation
> >> offsets for relative relocations.  Since a set of zeroes compresses
> >> better than a set of non-zero addend values, this behavior was resulting
> >> in much better lz4 compression.
> >>
> >> The bfd linker in binutils-2.27 is now storing the actual addend values
> >> at the relocation offsets. The behavior is now consistent with what it
> >> does for x86_64 and what gold linker does for both architectures.  The
> >> change happened in this upstream commit:
> >> https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=1f56df9d0d5ad89806c24e71f296576d82344613
> >> Since a bunch of zeroes got replaced by non-zero addend values, we see
> >> the side effect of lz4 compressed image being a bit bigger.
> >>
> >> To get the old behavior from the bfd linker, "--no-apply-dynamic-relocs"
> >> flag can be used:
> >> $ LDFLAGS="--no-apply-dynamic-relocs" make
> >> With this flag, the compressed image size is back to what it was with
> >> binutils-2.25.
> >>
> >> If the kernel is using ASLR, there aren't additional runtime costs to
> >> --no-apply-dynamic-relocs, as the relocations will need to be applied
> >> again anyway after the kernel is relocated to a random address.
> >>
> >> If the kernel is not using ASLR, then presumably the current default
> >> behavior of the linker is better. Since the static linker performed the
> >> dynamic relocs, and the kernel is not moved to a different address at
> >> load time, it can skip applying the relocations all over again."
> >
> > Do you have any numbers booting an uncompressed kernel Image without ASLR
> > to see if skipping the relocs makes a measurable difference there?
> >
> 
> Do you mean built with ASLR support but executing at the offset it was
> linked at?

Yeah, sorry for being vague. Basically, the case where the relocs have all
been resolved statically. In other words: what do we lose by disabling this
optimisation?

Will
Ard Biesheuvel - Oct. 30, 2017, 1:35 p.m.
On 30 October 2017 at 13:12, Will Deacon <will.deacon@arm.com> wrote:
> On Mon, Oct 30, 2017 at 01:11:23PM +0000, Ard Biesheuvel wrote:
>> On 30 October 2017 at 13:08, Will Deacon <will.deacon@arm.com> wrote:
>> > On Fri, Oct 27, 2017 at 09:33:41AM -0700, Nick Desaulniers wrote:
>> >> Upon upgrading to binutils 2.27, we found that our lz4 and gzip
>> >> compressed kernel images were significantly larger, resulting is 10ms
>> >> boot time regressions.
>> >>
>> >> As noted by Rahul:
>> >> "aarch64 binaries uses RELA relocations, where each relocation entry
>> >> includes an addend value. This is similar to x86_64.  On x86_64, the
>> >> addend values are also stored at the relocation offset for relative
>> >> relocations. This is an optimization: in the case where code does not
>> >> need to be relocated, the loader can simply skip processing relative
>> >> relocations.  In binutils-2.25, both bfd and gold linkers did this for
>> >> x86_64, but only the gold linker did this for aarch64.  The kernel build
>> >> here is using the bfd linker, which stored zeroes at the relocation
>> >> offsets for relative relocations.  Since a set of zeroes compresses
>> >> better than a set of non-zero addend values, this behavior was resulting
>> >> in much better lz4 compression.
>> >>
>> >> The bfd linker in binutils-2.27 is now storing the actual addend values
>> >> at the relocation offsets. The behavior is now consistent with what it
>> >> does for x86_64 and what gold linker does for both architectures.  The
>> >> change happened in this upstream commit:
>> >> https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=1f56df9d0d5ad89806c24e71f296576d82344613
>> >> Since a bunch of zeroes got replaced by non-zero addend values, we see
>> >> the side effect of lz4 compressed image being a bit bigger.
>> >>
>> >> To get the old behavior from the bfd linker, "--no-apply-dynamic-relocs"
>> >> flag can be used:
>> >> $ LDFLAGS="--no-apply-dynamic-relocs" make
>> >> With this flag, the compressed image size is back to what it was with
>> >> binutils-2.25.
>> >>
>> >> If the kernel is using ASLR, there aren't additional runtime costs to
>> >> --no-apply-dynamic-relocs, as the relocations will need to be applied
>> >> again anyway after the kernel is relocated to a random address.
>> >>
>> >> If the kernel is not using ASLR, then presumably the current default
>> >> behavior of the linker is better. Since the static linker performed the
>> >> dynamic relocs, and the kernel is not moved to a different address at
>> >> load time, it can skip applying the relocations all over again."
>> >
>> > Do you have any numbers booting an uncompressed kernel Image without ASLR
>> > to see if skipping the relocs makes a measurable difference there?
>> >
>>
>> Do you mean built with ASLR support but executing at the offset it was
>> linked at?
>
> Yeah, sorry for being vague. Basically, the case where the relocs have all
> been resolved statically. In other words: what do we lose by disabling this
> optimisation?
>

The code does not deal with that at all, currently: given that this is
new behavior in 2.27, the relocs are processed unconditionally,
regardless of whether the image is loaded at its default base or not.
Will Deacon - Oct. 30, 2017, 2:08 p.m.
On Mon, Oct 30, 2017 at 01:35:49PM +0000, Ard Biesheuvel wrote:
> On 30 October 2017 at 13:12, Will Deacon <will.deacon@arm.com> wrote:
> > On Mon, Oct 30, 2017 at 01:11:23PM +0000, Ard Biesheuvel wrote:
> >> On 30 October 2017 at 13:08, Will Deacon <will.deacon@arm.com> wrote:
> >> > On Fri, Oct 27, 2017 at 09:33:41AM -0700, Nick Desaulniers wrote:
> >> >> Upon upgrading to binutils 2.27, we found that our lz4 and gzip
> >> >> compressed kernel images were significantly larger, resulting is 10ms
> >> >> boot time regressions.
> >> >>
> >> >> As noted by Rahul:
> >> >> "aarch64 binaries uses RELA relocations, where each relocation entry
> >> >> includes an addend value. This is similar to x86_64.  On x86_64, the
> >> >> addend values are also stored at the relocation offset for relative
> >> >> relocations. This is an optimization: in the case where code does not
> >> >> need to be relocated, the loader can simply skip processing relative
> >> >> relocations.  In binutils-2.25, both bfd and gold linkers did this for
> >> >> x86_64, but only the gold linker did this for aarch64.  The kernel build
> >> >> here is using the bfd linker, which stored zeroes at the relocation
> >> >> offsets for relative relocations.  Since a set of zeroes compresses
> >> >> better than a set of non-zero addend values, this behavior was resulting
> >> >> in much better lz4 compression.
> >> >>
> >> >> The bfd linker in binutils-2.27 is now storing the actual addend values
> >> >> at the relocation offsets. The behavior is now consistent with what it
> >> >> does for x86_64 and what gold linker does for both architectures.  The
> >> >> change happened in this upstream commit:
> >> >> https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=1f56df9d0d5ad89806c24e71f296576d82344613
> >> >> Since a bunch of zeroes got replaced by non-zero addend values, we see
> >> >> the side effect of lz4 compressed image being a bit bigger.
> >> >>
> >> >> To get the old behavior from the bfd linker, "--no-apply-dynamic-relocs"
> >> >> flag can be used:
> >> >> $ LDFLAGS="--no-apply-dynamic-relocs" make
> >> >> With this flag, the compressed image size is back to what it was with
> >> >> binutils-2.25.
> >> >>
> >> >> If the kernel is using ASLR, there aren't additional runtime costs to
> >> >> --no-apply-dynamic-relocs, as the relocations will need to be applied
> >> >> again anyway after the kernel is relocated to a random address.
> >> >>
> >> >> If the kernel is not using ASLR, then presumably the current default
> >> >> behavior of the linker is better. Since the static linker performed the
> >> >> dynamic relocs, and the kernel is not moved to a different address at
> >> >> load time, it can skip applying the relocations all over again."
> >> >
> >> > Do you have any numbers booting an uncompressed kernel Image without ASLR
> >> > to see if skipping the relocs makes a measurable difference there?
> >> >
> >>
> >> Do you mean built with ASLR support but executing at the offset it was
> >> linked at?
> >
> > Yeah, sorry for being vague. Basically, the case where the relocs have all
> > been resolved statically. In other words: what do we lose by disabling this
> > optimisation?
> >
> 
> The code does not deal with that at all, currently: given that this is
> new behavior in 2.27, the relocs are processed unconditionally,
> regardless of whether the image is loaded at its default base or not.

Ah yeah, of course. Doesn't that make it impossible to exploit this
optimisation unless you can somehow guarantee that binutils has done
the relocs at link time? I guess you could check for non-zero entries at
the relocation offsets, but you still end up iterating.

Anyway, I'll take the patch but I don't understand how the binutils change
is intended to be used.

Will
Ard Biesheuvel - Oct. 30, 2017, 2:11 p.m.
On 30 October 2017 at 14:08, Will Deacon <will.deacon@arm.com> wrote:
> On Mon, Oct 30, 2017 at 01:35:49PM +0000, Ard Biesheuvel wrote:
>> On 30 October 2017 at 13:12, Will Deacon <will.deacon@arm.com> wrote:
>> > On Mon, Oct 30, 2017 at 01:11:23PM +0000, Ard Biesheuvel wrote:
>> >> On 30 October 2017 at 13:08, Will Deacon <will.deacon@arm.com> wrote:
>> >> > On Fri, Oct 27, 2017 at 09:33:41AM -0700, Nick Desaulniers wrote:
>> >> >> Upon upgrading to binutils 2.27, we found that our lz4 and gzip
>> >> >> compressed kernel images were significantly larger, resulting is 10ms
>> >> >> boot time regressions.
>> >> >>
>> >> >> As noted by Rahul:
>> >> >> "aarch64 binaries uses RELA relocations, where each relocation entry
>> >> >> includes an addend value. This is similar to x86_64.  On x86_64, the
>> >> >> addend values are also stored at the relocation offset for relative
>> >> >> relocations. This is an optimization: in the case where code does not
>> >> >> need to be relocated, the loader can simply skip processing relative
>> >> >> relocations.  In binutils-2.25, both bfd and gold linkers did this for
>> >> >> x86_64, but only the gold linker did this for aarch64.  The kernel build
>> >> >> here is using the bfd linker, which stored zeroes at the relocation
>> >> >> offsets for relative relocations.  Since a set of zeroes compresses
>> >> >> better than a set of non-zero addend values, this behavior was resulting
>> >> >> in much better lz4 compression.
>> >> >>
>> >> >> The bfd linker in binutils-2.27 is now storing the actual addend values
>> >> >> at the relocation offsets. The behavior is now consistent with what it
>> >> >> does for x86_64 and what gold linker does for both architectures.  The
>> >> >> change happened in this upstream commit:
>> >> >> https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=1f56df9d0d5ad89806c24e71f296576d82344613
>> >> >> Since a bunch of zeroes got replaced by non-zero addend values, we see
>> >> >> the side effect of lz4 compressed image being a bit bigger.
>> >> >>
>> >> >> To get the old behavior from the bfd linker, "--no-apply-dynamic-relocs"
>> >> >> flag can be used:
>> >> >> $ LDFLAGS="--no-apply-dynamic-relocs" make
>> >> >> With this flag, the compressed image size is back to what it was with
>> >> >> binutils-2.25.
>> >> >>
>> >> >> If the kernel is using ASLR, there aren't additional runtime costs to
>> >> >> --no-apply-dynamic-relocs, as the relocations will need to be applied
>> >> >> again anyway after the kernel is relocated to a random address.
>> >> >>
>> >> >> If the kernel is not using ASLR, then presumably the current default
>> >> >> behavior of the linker is better. Since the static linker performed the
>> >> >> dynamic relocs, and the kernel is not moved to a different address at
>> >> >> load time, it can skip applying the relocations all over again."
>> >> >
>> >> > Do you have any numbers booting an uncompressed kernel Image without ASLR
>> >> > to see if skipping the relocs makes a measurable difference there?
>> >> >
>> >>
>> >> Do you mean built with ASLR support but executing at the offset it was
>> >> linked at?
>> >
>> > Yeah, sorry for being vague. Basically, the case where the relocs have all
>> > been resolved statically. In other words: what do we lose by disabling this
>> > optimisation?
>> >
>>
>> The code does not deal with that at all, currently: given that this is
>> new behavior in 2.27, the relocs are processed unconditionally,
>> regardless of whether the image is loaded at its default base or not.
>
> Ah yeah, of course. Doesn't that make it impossible to exploit this
> optimisation unless you can somehow guarantee that binutils has done
> the relocs at link time? I guess you could check for non-zero entries at
> the relocation offsets, but you still end up iterating.
>
> Anyway, I'll take the patch but I don't understand how the binutils change
> is intended to be used.
>

This seems to be a 'parity with x86 for the sake of it' thing. I see
how not having to process relocations can be an advantage in some
cases, but it does duplicate information (as this patch proves), and
the cat's already out of the bag anyway.

Patch

diff --git a/arch/arm64/Makefile b/arch/arm64/Makefile
index 939b310913cf..669378099dbe 100644
--- a/arch/arm64/Makefile
+++ b/arch/arm64/Makefile
@@ -14,8 +14,9 @@  LDFLAGS_vmlinux	:=-p --no-undefined -X
 CPPFLAGS_vmlinux.lds = -DTEXT_OFFSET=$(TEXT_OFFSET)
 GZFLAGS		:=-9
 
-ifneq ($(CONFIG_RELOCATABLE),)
-LDFLAGS_vmlinux		+= -pie -shared -Bsymbolic
+ifeq ($(CONFIG_RELOCATABLE), y)
+LDFLAGS_vmlinux		+= -pie -shared -Bsymbolic \
+			$(call ld-option, --no-apply-dynamic-relocs)
 endif
 
 ifeq ($(CONFIG_ARM64_ERRATUM_843419),y)