Re: [PATCH v8 0/6] add support for relative references in special sections

[Ard Biesheuvel]

On 11 March 2018 at 12:38, Ard Biesheuvel  wrote:
> This adds support for emitting special sections such as initcall arrays,
> PCI fixups and tracepoints as relative references rather than absolute
> references. This reduces the size by 50% on 64-bit architectures, but
> more importantly, it removes the need for carrying relocation metadata
> for these sections in relocatable kernels (e.g., for KASLR) that needs
> to be fixed up at boot time. On arm64, this reduces the vmlinux footprint
> of such a reference by 8x (8 byte absolute reference + 24 byte RELA entry
> vs 4 byte relative reference)
>
> Patch #3 was sent out before as a single patch. This series supersedes
> the previous submission. This version makes relative ksymtab entries
> dependent on the new Kconfig symbol HAVE_ARCH_PREL32_RELOCATIONS rather
> than trying to infer from kbuild test robot replies for which architectures
> it should be blacklisted.
>
> Patch #1 introduces the new Kconfig symbol HAVE_ARCH_PREL32_RELOCATIONS,
> and sets it for the main architectures that are expected to benefit the
> most from this feature, i.e., 64-bit architectures or ones that use
> runtime relocations.
>
> Patch #2 add support for #define'ing __DISABLE_EXPORTS to get rid of
> ksymtab/kcrctab sections in decompressor and EFI stub objects when
> rebuilding existing C files to run in a different context.
>
> Patches #4 - #6 implement relative references for initcalls, PCI fixups
> and tracepoints, respectively, all of which produce sections with order
> ~1000 entries on an arm64 defconfig kernel with tracing enabled. This
> means we save about 28 KB of vmlinux space for each of these patches.
>
> [From the v7 series blurb, which included the jump_label patches as well]:
>   For the arm64 kernel, all patches combined reduce the memory footprint of
>   vmlinux by about 1.3 MB (using a config copied from Ubuntu that has KASLR
>   enabled), of which ~1 MB is the size reduction of the RELA section in .init,
>   and the remaining 300 KB is reduction of .text/.data.
>
> Branch:
> git://git.kernel.org/pub/scm/linux/kernel/git/ardb/linux.git 
> relative-special-sections-v8
>

All,

I think this has converged to a mergeable state now (taking the
mini-tweak for patch #3 into account I provided in a reply)

Are there any objections to this being merged?

Andrew,

If nobody objects, is this something you would care to pick up for the
next cycle? Do I need to respin to incorporate the mini-tweak and
apply Nico's ack to patch #2?

Thanks,
Ard.


> Changes since v7:
> - dropped the jump_label patches, these will be revisited in a separate series
> - reorder __DISABLE_EXPORTS with __KSYM_DEPS__ check in #2
> - use offset_to_ptr() helper function to abstract the relative pointer
>   conversion [int *off -> (ulong)off + *off] (#3 - #6)
> - rebase onto v4.16-rc3
>
> Changes since v6:
> - drop S390 from patch #1 introducing HAVE_ARCH_PREL32_RELOCATIONS: kbuild
>   robot threw me some s390 curveballs, and given that s390 does not define
>   CONFIG_RELOCATABLE in the first place, it does not benefit as much from
>   relative references as arm64, x86 and power do
> - add patch to allow symbol exports to be disabled at compilation unit
>   granularity (#2)
> - get rid of arm64 vmlinux.lds.S hunk to ensure code generated by 
> __ADDRESSABLE
>   gets discarded from the EFI stub - it is no longer needed after adding #2 
> (#1)
> - change _ADDRESSABLE() to emit a data reference, not a code reference - this
>   is another simplification made possible by patch #2 (#3)
> - add Steven's ack to #6
> - split x86 jump_label patch into two (#9, #10)
>
> Changes since v5:
> - add missing jump_label prototypes to s390 jump_label.h (#6)
> - fix inverted condition in call to jump_entry_is_module_init() (#6)
>
> Changes since v4:
> - add patches to convert x86 and arm64 to use relative references for jump
>   tables (#6 - #8)
> - rename PCI patch and add Bjorn's ack (#4)
> - rebase onto v4.15-rc5
>
> Changes since v3:
> - fix module unload issue in patch #5 reported by Jessica, by reusing the
>   updated routine for_each_tracepoint_range() for the quiescent check at
>   module unload time; this requires this routine to be moved before
>   tracepoint_module_going() in kernel/tracepoint.c
> - add Jessica's ack to #2
> - rebase onto v4.14-rc1
>
> Changes since v2:
> - Revert my slightly misguided attempt to appease checkpatch, which resulted
>   in needless churn and worse code. This v3 is based on v1 with a few tweaks
>   that were actually reasonable checkpatch warnings: unnecessary braces (as
>   pointed out by Ingo) and other minor whitespace misdemeanors.
>
> Changes since v1:
> - Remove checkpatch errors to the extent feasible: in some cases, this
>   involves moving extern declarations into C files, and switching to
>   struct definitions rather than typedefs. Some errors are impossible
>   to fix: please find the remaining ones after the diffstat.
> - Used 'int' 

Re: [PATCH v8 0/6] add support for relative references in special sections

[Kees Cook]

On Sun, Mar 11, 2018 at 5:44 AM, Ard Biesheuvel
 wrote:
> On 11 March 2018 at 12:38, Ard Biesheuvel  wrote:
>> Cc: James Morris 
>
> Note when replying: this ^^^ email address no longer works.

James updated MAINTAINERS to: James Morris 

-- 
Kees Cook
Pixel Security

Re: [PATCH v8 0/6] add support for relative references in special sections

[Ard Biesheuvel]

On 11 March 2018 at 12:38, Ard Biesheuvel  wrote:
...
> Cc: Arnd Bergmann 
> Cc: Kees Cook 
> Cc: Will Deacon 
> Cc: Michael Ellerman 
> Cc: Thomas Garnier 
> Cc: Thomas Gleixner 
> Cc: "Serge E. Hallyn" 
> Cc: Bjorn Helgaas 
> Cc: Benjamin Herrenschmidt 
> Cc: Russell King 
> Cc: Paul Mackerras 
> Cc: Catalin Marinas 
> Cc: Petr Mladek 
> Cc: Ingo Molnar 
> Cc: James Morris 

Note when replying: this ^^^ email address no longer works.

> Cc: Andrew Morton 
> Cc: Nicolas Pitre 
> Cc: Josh Poimboeuf 
> Cc: Steven Rostedt 
> Cc: Sergey Senozhatsky 
> Cc: Linus Torvalds 
> Cc: Jessica Yu 
>
> Cc: linux-arm-ker...@lists.infradead.org
> Cc: linux-ker...@vger.kernel.org
> Cc: linuxppc-dev@lists.ozlabs.org
> Cc: x...@kernel.org
>
...

[PATCH v8 0/6] add support for relative references in special sections

[Ard Biesheuvel]

This adds support for emitting special sections such as initcall arrays,
PCI fixups and tracepoints as relative references rather than absolute
references. This reduces the size by 50% on 64-bit architectures, but
more importantly, it removes the need for carrying relocation metadata
for these sections in relocatable kernels (e.g., for KASLR) that needs
to be fixed up at boot time. On arm64, this reduces the vmlinux footprint
of such a reference by 8x (8 byte absolute reference + 24 byte RELA entry
vs 4 byte relative reference)

Patch #3 was sent out before as a single patch. This series supersedes
the previous submission. This version makes relative ksymtab entries
dependent on the new Kconfig symbol HAVE_ARCH_PREL32_RELOCATIONS rather
than trying to infer from kbuild test robot replies for which architectures
it should be blacklisted.

Patch #1 introduces the new Kconfig symbol HAVE_ARCH_PREL32_RELOCATIONS,
and sets it for the main architectures that are expected to benefit the
most from this feature, i.e., 64-bit architectures or ones that use
runtime relocations.

Patch #2 add support for #define'ing __DISABLE_EXPORTS to get rid of
ksymtab/kcrctab sections in decompressor and EFI stub objects when
rebuilding existing C files to run in a different context.

Patches #4 - #6 implement relative references for initcalls, PCI fixups
and tracepoints, respectively, all of which produce sections with order
~1000 entries on an arm64 defconfig kernel with tracing enabled. This
means we save about 28 KB of vmlinux space for each of these patches.

[From the v7 series blurb, which included the jump_label patches as well]:
  For the arm64 kernel, all patches combined reduce the memory footprint of
  vmlinux by about 1.3 MB (using a config copied from Ubuntu that has KASLR
  enabled), of which ~1 MB is the size reduction of the RELA section in .init,
  and the remaining 300 KB is reduction of .text/.data.

Branch:
git://git.kernel.org/pub/scm/linux/kernel/git/ardb/linux.git 
relative-special-sections-v8

Changes since v7:
- dropped the jump_label patches, these will be revisited in a separate series
- reorder __DISABLE_EXPORTS with __KSYM_DEPS__ check in #2
- use offset_to_ptr() helper function to abstract the relative pointer
  conversion [int *off -> (ulong)off + *off] (#3 - #6)
- rebase onto v4.16-rc3

Changes since v6:
- drop S390 from patch #1 introducing HAVE_ARCH_PREL32_RELOCATIONS: kbuild
  robot threw me some s390 curveballs, and given that s390 does not define
  CONFIG_RELOCATABLE in the first place, it does not benefit as much from
  relative references as arm64, x86 and power do
- add patch to allow symbol exports to be disabled at compilation unit
  granularity (#2)
- get rid of arm64 vmlinux.lds.S hunk to ensure code generated by __ADDRESSABLE
  gets discarded from the EFI stub - it is no longer needed after adding #2 (#1)
- change _ADDRESSABLE() to emit a data reference, not a code reference - this
  is another simplification made possible by patch #2 (#3)
- add Steven's ack to #6
- split x86 jump_label patch into two (#9, #10)

Changes since v5:
- add missing jump_label prototypes to s390 jump_label.h (#6)
- fix inverted condition in call to jump_entry_is_module_init() (#6)

Changes since v4:
- add patches to convert x86 and arm64 to use relative references for jump
  tables (#6 - #8)
- rename PCI patch and add Bjorn's ack (#4)
- rebase onto v4.15-rc5

Changes since v3:
- fix module unload issue in patch #5 reported by Jessica, by reusing the
  updated routine for_each_tracepoint_range() for the quiescent check at
  module unload time; this requires this routine to be moved before
  tracepoint_module_going() in kernel/tracepoint.c
- add Jessica's ack to #2
- rebase onto v4.14-rc1

Changes since v2:
- Revert my slightly misguided attempt to appease checkpatch, which resulted
  in needless churn and worse code. This v3 is based on v1 with a few tweaks
  that were actually reasonable checkpatch warnings: unnecessary braces (as
  pointed out by Ingo) and other minor whitespace misdemeanors.

Changes since v1:
- Remove checkpatch errors to the extent feasible: in some cases, this
  involves moving extern declarations into C files, and switching to
  struct definitions rather than typedefs. Some errors are impossible
  to fix: please find the remaining ones after the diffstat.
- Used 'int' instead if 'signed int' for the various offset fields: there
  is no ambiguity between architectures regarding its signedness (unlike
  'char')
- Refactor the different patches to be more uniform in the way they define
  the section entry type and accessors in the .h file, and avoid the need to
  add #ifdefs to the C code.

Cc: Arnd Bergmann 
Cc: Kees Cook 
Cc: Will Deacon 
Cc: Michael Ellerman 
Cc: Thomas Garnier 
Cc: Thomas Gleixner 
Cc: "Serge E. Hallyn" 
Cc: Bjorn Helgaas