Re: [PATCH] x86/vdso: include generic __lshrdi3 in 32-bit vDSO

[hpa]

On March 15, 2019 3:29:06 PM PDT, Matthias Kaehlcke <mka@xxxxxxxxxxxx> wrote:
>Hi Nick,
>
>On Fri, Mar 15, 2019 at 02:31:09PM -0700, 'Nick Desaulniers' via Clang
>Built Linux wrote:
>> On Fri, Mar 15, 2019 at 12:54 PM Matthias Kaehlcke <mka@xxxxxxxxxxxx>
>wrote:
>> >
>> > Building the 32-bit vDSO with a recent clang version fails due
>> > to undefined symbols:
>> >
>> > arch/x86/entry/vdso/vdso32.so.dbg: undefined symbols found
>> >
>> > The undefined symbol in this case is __lshrdi3, which is part of
>> > the compiler runtime library, however the vDSO isn't linked against
>> > this library.
>> >
>> > Include the kernel version of __lshrdi3 in the 32-bit vDSO build.
>> 
>> __lshrdi3 is used for "logical shift right double-word by int" (best
>> guess), so anywhere there's a right shift of a u64.  Looks like
>> there's a few of these in arch/x86/entry/vdso/, so it's legal for the
>> compiler to emit this libcall.  Do you know which function
>> specifically in the .so has a relocation referencing __lshrdi3
>> specifically?
>
>It's the right shifts in do_realtime() and do_monotonic().
>
>> Is there a config I can set to reproduce this, in order to help
>> test?
>
>I encountered it with a Chrome OS specific configuration, but a
>defconfig should do. Note that you probably need a development version
>of clang to reproduce this.
>
>> >
>> > Signed-off-by: Matthias Kaehlcke <mka@xxxxxxxxxxxx>
>> > ---
>> >  arch/x86/entry/vdso/Makefile | 7 ++++++-
>> >  lib/lshrdi3.c                | 4 +++-
>> >  2 files changed, 9 insertions(+), 2 deletions(-)
>> >
>> > diff --git a/arch/x86/entry/vdso/Makefile
>b/arch/x86/entry/vdso/Makefile
>> > index 5bfe2243a08f..7517cd87e10b 100644
>> > --- a/arch/x86/entry/vdso/Makefile
>> > +++ b/arch/x86/entry/vdso/Makefile
>> > @@ -144,6 +144,7 @@ KBUILD_CFLAGS_32 += $(call cc-option,
>-fno-stack-protector)
>> >  KBUILD_CFLAGS_32 += $(call cc-option, -foptimize-sibling-calls)
>> >  KBUILD_CFLAGS_32 += -fno-omit-frame-pointer
>> >  KBUILD_CFLAGS_32 += -DDISABLE_BRANCH_PROFILING
>> > +KBUILD_CFLAGS_32 += -DBUILD_VDSO
>> >
>> >  ifdef CONFIG_RETPOLINE
>> >  ifneq ($(RETPOLINE_VDSO_CFLAGS),)
>> > @@ -153,12 +154,16 @@ endif
>> >
>> >  $(obj)/vdso32.so.dbg: KBUILD_CFLAGS = $(KBUILD_CFLAGS_32)
>> >
>> > +$(obj)/vdso32/lshrdi3.o: $(srctree)/lib/lshrdi3.c FORCE
>> > +       $(call if_changed_rule,cc_o_c)
>> 
>> + Masahiro to help look at this part (I don't understand this part
>> of kbuild).
>
>I bluntly stole that from arch/x86/purgatory/Makefile , which does
>something similar.
>
>> 
>> > +
>> >  $(obj)/vdso32.so.dbg: FORCE \
>> >                       $(obj)/vdso32/vdso32.lds \
>> >                       $(obj)/vdso32/vclock_gettime.o \
>> >                       $(obj)/vdso32/note.o \
>> >                       $(obj)/vdso32/system_call.o \
>> > -                     $(obj)/vdso32/sigreturn.o
>> > +                     $(obj)/vdso32/sigreturn.o \
>> > +                     $(obj)/vdso32/lshrdi3.o
>> >         $(call if_changed,vdso)
>> >
>> >  #
>> > diff --git a/lib/lshrdi3.c b/lib/lshrdi3.c
>> > index 99cfa5721f2d..8a4fc6bcf3a4 100644
>> > --- a/lib/lshrdi3.c
>> > +++ b/lib/lshrdi3.c
>> > @@ -16,7 +16,7 @@
>> >   * to the Free Software Foundation, Inc.
>> >   */
>> >
>> > -#include <linux/module.h>
>> > +#include <linux/export.h>
>> 
>> Is this a simple cleanup, or?
>
>The vDSO build is unhappy when modules.h draws in a whole bunch of
>other kernel headers and export.h is all that's need. It seemed
>reasonable to do the 'cleanup' in this patch since we touch it anyway
>to place EXPORT_SYMBOL within an #ifdef.
>
>> >  #include <linux/libgcc.h>
>> >
>> >  long long notrace __lshrdi3(long long u, word_type b)
>> > @@ -42,4 +42,6 @@ long long notrace __lshrdi3(long long u,
>word_type b)
>> >
>> >         return w.ll;
>> >  }
>> > +#ifndef BUILD_VDSO
>> >  EXPORT_SYMBOL(__lshrdi3);
>> > +#endif
>> 
>> Compilers (GCC and Clang) will always assume their runtime has these
>> helper functions; whether or not they emit libcalls vs inline
>routines
>> is implementation defined.  So I agree with this patch; I just would
>> like to help confirm/test it.
>
>Thanks for your help!
>
>Matthias

Note: it is also probably no reason to use -Os/-Oz for the vdso.
-- 
Sent from my Android device with K-9 Mail. Please excuse my brevity.