Re: Fwd: [PATCH] arm: ensure symbol is a thumb symbol in new binutils

[Russell King - ARM Linux <linux at armlinux dot org dot uk>]

On Thu, Nov 23, 2017 at 11:47:56AM +0100, Jason A. Donenfeld wrote:
> Hello Nick & Binutils developers,
> 
> It would appear that recent changes in the binutils assembler (perhaps
> 52a86f843b6dee1de9977293da9786649b146b05? perhaps not?) have broken
> the kernel when compiled in thumb2 mode. We currently do not have a
> way of working around your breaking changes without adding additional
> runtime instructions, which isn't acceptable for us. Details are
> included in the thread below.

Hi Nick,

I have to ask why it was thought a good idea to change the behaviour
of the assembler's "adr" pseudo-instruction after it's had the existing
behaviour for such a long time.

I see in the bug report https://sourceware.org/bugzilla/show_bug.cgi?id=21458
that guidance was sort from ARM Ltd, but none was forthcoming - which
implies that there was little technical input to make the change.
Given that the change has been known to break at least three programs
(the Linux kernel, libavcodec and openssl), that there's no technical
reason to make the change, and that it introduces inconsistency in
the assembler, is there enough reason to get the change at least
partially removed - restoring the long-standing "adr" behaviour?

The inconsistency is introduced in that "adr" now behaves differently
depending on whether the symbol is marked as a thumb function (which
requires an explicit pseudo-op) or not.  This means that either we
need additional labels that are not marked as a thumb function for use
with "adr" or we need to mark _every_ symbol in thumb code that is used
with "adr" with a .thumb_func annotation, including the numeric local
labels.

This makes it much harder to review patches for projects and ensure
that they are correct as it means that reviewers now need to do detailed
in-depth analysis of the code after /any/ patch that uses "adr" has been
applied.  Given that the failure is will only be seen at runtime, this
is particularly bad.

> 
> Thanks,
> Jason
> 
> Forwarded conversation
> Subject: [PATCH] arm: ensure symbol is a thumb symbol in new binutils
> ------------------------
> 
> From: Jason A. Donenfeld <Jason@zx2c4.com>
> Date: Tue, Nov 21, 2017 at 6:27 PM
> To: linux@armlinux.org.uk, linux-arm-kernel@lists.infradead.org
> Cc: "Jason A. Donenfeld" <Jason@zx2c4.com>, stable@vger.kernel.org
> 
> 
> On older versions of binutils, \sym points to an aligned address. On
> newer versions of binutils, \sym sometimes points to the unaligned thumb
> address in mysterious and buggy circumstances. In order to homogenize
> this behavior, rather than adding 1, we simply OR in 1, so that already
> unaligned instructions don't change. This fix is required for a
> pedestrian THUMB2_KERNEL to boot without crashing when built with
> non-old binutils.
> 
> While it works, the downside is that we have to add an `orr` instruction
> to a fast path. The assembler can't do this at assemble time via "|1"
> because "invalid operands (.text and *ABS* sections) for `|'", so we're
> forced to do this. A better solution would be to have consistent
> binutils behavior, or to have some kind of \sym feature detection that
> won't turn into a maze of version comparisons. However, it's at the
> moment unclear how to achieve this.
> 
> The rest of this commit message contains all of the relevant
> information.
> 
> My tests concerned these versions:
>     broken: GNU ld (Gentoo 2.29.1 p3) 2.29.1
>     working: GNU ld (GNU Binutils for Ubuntu) 2.26.1
> 
> These produced the following code:
> --- broken      2017-11-21 17:44:14.523416082 +0100
> +++ working     2017-11-21 17:44:44.548461234 +0100
> @@ -133,7 +133,7 @@
>   160:  f01a 0ff0       tst.w   sl, #240        ; 0xf0
>   164:  d111            bne.n   18a <__sys_trace>
>   166:  f5b7 7fc8       cmp.w   r7, #400        ; 0x190
> - 16a:  f2af 1e6a       subw    lr, pc, #362    ; 0x16a
> + 16a:  f2af 1e6b       subw    lr, pc, #363    ; 0x16b
>   16e:  bf38            it      cc
>   170:  f858 f027       ldrcc.w pc, [r8, r7, lsl #2]
>   174:  a902            add     r1, sp, #8
> 
> The differing instruction corresponds with this actual line in
> arch/arm/kernel/entry-common.S:
>       badr    lr, ret_fast_syscall            @ return address
> 
> Running the broken kernel results in a runtime OOPS with:
>     PC is at ret_fast_syscall+0x4/0x52
>     LR is at ret_fast_syscall+0x2/0x52
> 
> The disassembly of that function for the crashing kernel is:
> .text:00000000 ret_fast_syscall                        ; CODE XREF:
> sys_syscall+1C↓j
> .text:00000000                 CPSID           I       ; jumptable
> 00000840 cases 15,18-376
> .text:00000002
> .text:00000002 loc_2                                   ; DATA XREF:
> sys_syscall-6BA↓o
> .text:00000002                 LDR.W           R2, [R9,#8]
> .text:00000006                 CMP.W           R2, #0xBF000000
> 
> Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
> Cc: stable@vger.kernel.org
> ---
>  arch/arm/include/asm/assembler.h | 5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/arm/include/asm/assembler.h b/arch/arm/include/asm/assembler.h
> index ad301f107dd2..c62a3b6b0a3e 100644
> --- a/arch/arm/include/asm/assembler.h
> +++ b/arch/arm/include/asm/assembler.h
> @@ -194,10 +194,9 @@
>   */
>         .irp    c,,eq,ne,cs,cc,mi,pl,vs,vc,hi,ls,ge,lt,gt,le,hs,lo
>         .macro  badr\c, rd, sym
> -#ifdef CONFIG_THUMB2_KERNEL
> -       adr\c   \rd, \sym + 1
> -#else
>         adr\c   \rd, \sym
> +#ifdef CONFIG_THUMB2_KERNEL
> +       orr\c   \rd, \rd, 1
>  #endif
>         .endm
>         .endr
> --
> 2.15.0
> 
> 
> ----------
> From: Russell King - ARM Linux <linux@armlinux.org.uk>
> Date: Tue, Nov 21, 2017 at 6:38 PM
> To: "Jason A. Donenfeld" <Jason@zx2c4.com>, Will Deacon <will.deacon@arm.com>
> Cc: linux-arm-kernel@lists.infradead.org, lkml@vger.kernel.org,
> stable@vger.kernel.org
> 
> 
> As it just seems to be a limited range of binutils versions that are
> affected, I'd rather not impact the kernel fast-paths with extra
> cycles just because binutils decided to change behaviour.  I'd prefer
> to inform people about the problem and get them to change to a non-
> buggy binutils.
> 
> This seems to be the second binutils bug that's biting us within the
> last month... what's going on with binutils QA?
> 
>  arch/arm/Makefile        |  7 +++++--
>  arch/arm/tools/Makefile  |  5 ++++-
>  arch/arm/tools/toolcheck | 24 ++++++++++++++++++++++++
>  3 files changed, 33 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/arm/Makefile b/arch/arm/Makefile
> index 1cfac5119545..9e70d0435121 100644
> --- a/arch/arm/Makefile
> +++ b/arch/arm/Makefile
> @@ -319,16 +319,19 @@ all:      $(notdir $(KBUILD_IMAGE)) $(KBUILD_DTBS)
>  archheaders:
>         $(Q)$(MAKE) $(build)=arch/arm/tools uapi
> 
> -archprepare:
> +archprepare: toolcheck
>         $(Q)$(MAKE) $(build)=arch/arm/tools kapi
> 
> +toolcheck:
> +       $(Q)$(MAKE) $(build)=arch/arm/tools $@
> +
>  # Convert bzImage to zImage
>  bzImage: zImage
> 
>  BOOT_TARGETS   = zImage Image xipImage bootpImage uImage
>  INSTALL_TARGETS        = zinstall uinstall install
> 
> -PHONY += bzImage $(BOOT_TARGETS) $(INSTALL_TARGETS)
> +PHONY += bzImage $(BOOT_TARGETS) $(INSTALL_TARGETS) toolcheck
> 
>  bootpImage uImage: zImage
>  zImage: Image
> diff --git a/arch/arm/tools/Makefile b/arch/arm/tools/Makefile
> index ddb89a7db36f..fa77351ccefd 100644
> --- a/arch/arm/tools/Makefile
> +++ b/arch/arm/tools/Makefile
> @@ -23,12 +23,15 @@ uapi-hdrs-y += $(uapi)/unistd-eabi.h
> 
>  targets += $(addprefix ../../../,$(gen-y) $(kapi-hdrs-y) $(uapi-hdrs-y))
> 
> -PHONY += kapi uapi
> +PHONY += kapi uapi toolcheck
> 
>  kapi:  $(kapi-hdrs-y) $(gen-y)
> 
>  uapi:  $(uapi-hdrs-y)
> 
> +toolcheck:
> +       @$(CONFIG_SHELL) '$(srctree)/$(src)/toolcheck'
> +
>  # Create output directory if not already present
>  _dummy := $(shell [ -d '$(kapi)' ] || mkdir -p '$(kapi)') \
>            $(shell [ -d '$(uapi)' ] || mkdir -p '$(uapi)')
> diff --git a/arch/arm/tools/toolcheck b/arch/arm/tools/toolcheck
> index e69de29bb2d1..97bbeeb691da 100644
> --- a/arch/arm/tools/toolcheck
> +++ b/arch/arm/tools/toolcheck
> @@ -0,0 +1,24 @@
> +#!/bin/sh -ex
> +if grep -q 'CONFIG_THUMB2_KERNEL=y' .config; then
> +   tmp=$(mktemp -d /tmp/binutils-test.XXXXXXXXXX)
> +   cat <<EOF | $AS $ASFLAGS -o $tmp/test.o
> +       .syntax unified
> +       .thumb
> +       .macro  badr, reg, sym
> +       adr     \reg, \sym + 1
> +       .endm
> +
> +test:
> +       mov     r0, #0
> +       badr    lr, test
> +EOF
> +   if ! $OBJDUMP -d $tmp/test.o | grep -q '4:\s*f2af 0e07'; then
> +      echo "Error: your assembler version produces buggy kernels" >&2
> +      $AS --version | head -n1 >&2
> +      rm $tmp/*.o
> +      rmdir $tmp
> +      exit 1
> +   fi
> +   rm $tmp/*.o
> +   rmdir $tmp
> +fi
> 
> ----------
> From: Jason A. Donenfeld <Jason@zx2c4.com>
> Date: Tue, Nov 21, 2017 at 6:46 PM
> To: Russell King - ARM Linux <linux@armlinux.org.uk>
> Cc: Will Deacon <will.deacon@arm.com>,
> linux-arm-kernel@lists.infradead.org, lkml@vger.kernel.org,
> stable@vger.kernel.org
> 
> 
> Hi Russell,
> 
> This doesn't actually catch the issues. In the buggy binutils, it
> appears that sometimes adr grabs the right symbol and sometimes it
> doesn't. I'm not yet able to figure out a minimal condition for it
> going one way or the other.
> 
> Jason
> 
> ----------
> From: Russell King - ARM Linux <linux@armlinux.org.uk>
> Date: Thu, Nov 23, 2017 at 11:35 AM
> To: "Jason A. Donenfeld" <Jason@zx2c4.com>
> Cc: Will Deacon <will.deacon@arm.com>,
> linux-arm-kernel@lists.infradead.org, lkml@vger.kernel.org,
> stable@vger.kernel.org
> 
> 
> On Thu, Nov 23, 2017 at 12:34:08AM +0100, Jason A. Donenfeld wrote:
> > On Tue, Nov 21, 2017 at 6:49 PM, Russell King - ARM Linux
> > <linux@armlinux.org.uk> wrote:
> > > What if we locate the "badr" instruction to the same offset - does
> > > that trigger the binutils bug?  Note that the grep expression will
> > > need updating...
> >
> > Nope, this too does not reproduce it. I'm having a bit of trouble
> > making a minimal test case to reproduce it. But I can reproduce it
> > everytime by simply assembling the file in question using that
> > binutils.
> 
> If it's that hard to reproduce it, it makes me wonder if it's being
> caused by some memory allocation being re-used without full
> initialisation, and it's reading stale data.
> 
> We can't easily build entry-common.o and check it for the problem as the
> position of the "local_restart" code depends on various configuration
> options.
> 
> I found this URL:
> 
> https://fossies.org/diffs/binutils/2.28_vs_2.29/gas/config/tc-arm.c-diff.html
> 
> and if you search down to around "line 8358", which is in do_adr(),
> there is this new code added:
> 
>       if (inst.reloc.exp.X_op == O_symbol
>           && inst.reloc.exp.X_add_symbol != NULL
>           && S_IS_DEFINED (inst.reloc.exp.X_add_symbol)
>           && THUMB_IS_FUNC (inst.reloc.exp.X_add_symbol))
>         inst.reloc.exp.X_add_number += 1;
> 
> which would account for the issue you're seeing.
> 
> Given that this change breaks openssl, and breaks the kernel, and
> this behaviour is something that we've come to rely upon in the
> kernel since T2 was introduced, and there's no way around it without
> adding additional instructions, I have to ask what the hell binutils
> people think they are doing changing the behaviour of the assembler
> in such a gratuitous way, and how they think that users of their
> crapware are going to be able to write code that assembles correctly
> on both pre-2.29 assemblers and post-2.29 assemblers.
> 
> Sorry, but gratuitous changes like this in the toolchain really
> annoy me, and just give me more reason to stick with my old known
> working versions (binutils 2.25, gcc 4.7.4!) rather than move
> forward and then not know whether bugs are due to crappy toolchains
> or really something wrong in the program.
> 
> binutils people need to realise that what they offer is an interface
> for converting assembly code into opcodes and if they change the
> translation of that in a visible way, people are going to get
> annoyed - just like if we in the kernel change the kernel's user
> visible API.
> 
> IMHO, binutils should have exactly the same rules - if a change causes
> a regression for a user, the change is wrong and should be reverted.

-- 
RMK's Patch system: http://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 8.8Mbps down 630kbps up
According to speedtest.net: 8.21Mbps down 510kbps up