Re: [PATCH v2] KVM: arm64: nvhe: Fix build with profile optimization

From: Denis Nikitin
Date: Thu Oct 06 2022 - 12:28:48 EST


Hi Mark,

This problem currently blocks the PGO roll on the ChromeOS kernel and
we need some kind of a solution.
Could you please take a look?

Thanks,
Denis

On Thu, Sep 29, 2022 at 9:13 AM Denis Nikitin <denik@xxxxxxxxxxxx> wrote:
>
> Hi Marc,
>
> Please let me know what you think about this approach.
>
> Thanks,
> Denis
>
> On Thu, Sep 22, 2022 at 11:04 PM Manoj Gupta <manojgupta@xxxxxxxxxx> wrote:
> >
> >
> >
> > On Thu, Sep 22, 2022 at 10:01 PM Denis Nikitin <denik@xxxxxxxxxxxx> wrote:
> >>
> >> Hi Mark,
> >>
> >> On Thu, Sep 22, 2022 at 3:38 AM Marc Zyngier <maz@xxxxxxxxxx> wrote:
> >> >
> >> > I was really hoping that you'd just drop the flags from the CFLAGS
> >> > instead of removing the generated section. Something like:
> >> >
> >> > diff --git a/arch/arm64/kvm/hyp/nvhe/Makefile b/arch/arm64/kvm/hyp/nvhe/Makefile
> >> > index b5c5119c7396..e5b2d43925b4 100644
> >> > --- a/arch/arm64/kvm/hyp/nvhe/Makefile
> >> > +++ b/arch/arm64/kvm/hyp/nvhe/Makefile
> >> > @@ -88,7 +88,7 @@ quiet_cmd_hypcopy = HYPCOPY $@
> >> >
> >> > # Remove ftrace, Shadow Call Stack, and CFI CFLAGS.
> >> > # This is equivalent to the 'notrace', '__noscs', and '__nocfi' annotations.
> >> > -KBUILD_CFLAGS := $(filter-out $(CC_FLAGS_FTRACE) $(CC_FLAGS_SCS) $(CC_FLAGS_CFI), $(KBUILD_CFLAGS))
> >> > +KBUILD_CFLAGS := $(filter-out $(CC_FLAGS_FTRACE) $(CC_FLAGS_SCS) $(CC_FLAGS_CFI) -fprofile-sample-use, $(KBUILD_CFLAGS))
> >> >
> >> > # KVM nVHE code is run at a different exception code with a different map, so
> >> > # compiler instrumentation that inserts callbacks or checks into the code may
> >>
> >> Sorry, I moved on with a different approach and didn't explain the rationale.
> >>
> >> Like you mentioned before, the flag `-fprofile-sample-use` does not appear
> >> in the kernel. And it looks confusing when the flag is disabled or filtered out
> >> here. This was the first reason.
> >>
> >> The root cause of the build failure wasn't the compiler profile guided
> >> optimization but the extra metadata in SHT_REL section which llvm injected
> >> into kvm_nvhe.tmp.o for further link optimization.
> >> If we remove the .llvm.call-graph-profile section we fix the build and avoid
> >> potential problems with relocations optimized by the linker. The profile
> >> guided optimization will still be applied by the compiler.
> >>
> >> Let me know what you think about it.
> >>
> >> >
> >> > However, I even failed to reproduce your problem using LLVM 14 as
> >> > packaged by Debian (if that matters, I'm using an arm64 build
> >> > machine). I build the kernel with:
> >> >
> >> > $ make LLVM=1 KCFLAGS=-fprofile-sample-use -j8 vmlinux
> >> >
> >> > and the offending object only contains the following sections:
> >> >
> >
> >
> > Just some comments based on my ChromeOS build experience.
> >
> > fprofile-sample-use needs the profile file name argument to read the pgo data from
> > i.e. -fprofile-sample-use=/path/to/gcov.profile.
> >
> > Since the path to filename can change, it makes filtering out more difficult.
> > It is certainly possible to find and filter the exact argument by some string search of KCFLAGS.
> > But passing -fno-profile-sample-use is easier and less error prone which I believe the previous patch version tried to do.
> >
> >
> >> > arch/arm64/kvm/hyp/nvhe/kvm_nvhe.tmp.o: file format elf64-littleaarch64
> >> >
> >> > Sections:
> >> > Idx Name Size VMA LMA File off Algn
> >> > 0 .hyp.idmap.text 00000ae4 0000000000000000 0000000000000000 00000800 2**11
> >> > CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE
> >> > 1 .hyp.text 0000e988 0000000000000000 0000000000000000 00001800 2**11
> >> > CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE
> >> > 2 .hyp.data..ro_after_init 00000820 0000000000000000 0000000000000000 00010188 2**3
> >> > CONTENTS, ALLOC, LOAD, DATA
> >> > 3 .hyp.rodata 00002e70 0000000000000000 0000000000000000 000109a8 2**3
> >> > CONTENTS, ALLOC, LOAD, RELOC, READONLY, DATA
> >> > 4 .hyp.data..percpu 00001ee0 0000000000000000 0000000000000000 00013820 2**4
> >> > CONTENTS, ALLOC, LOAD, DATA
> >> > 5 .hyp.bss 00001158 0000000000000000 0000000000000000 00015700 2**3
> >> > ALLOC
> >> > 6 .comment 0000001f 0000000000000000 0000000000000000 00017830 2**0
> >> > CONTENTS, READONLY
> >> > 7 .llvm_addrsig 000000b8 0000000000000000 0000000000000000 0001784f 2**0
> >> > CONTENTS, READONLY, EXCLUDE
> >> > 8 .altinstructions 00001284 0000000000000000 0000000000000000 00015700 2**0
> >> > CONTENTS, ALLOC, LOAD, RELOC, READONLY, DATA
> >> > 9 __jump_table 00000960 0000000000000000 0000000000000000 00016988 2**3
> >> > CONTENTS, ALLOC, LOAD, RELOC, DATA
> >> > 10 __bug_table 0000051c 0000000000000000 0000000000000000 000172e8 2**2
> >> > CONTENTS, ALLOC, LOAD, RELOC, DATA
> >> > 11 __kvm_ex_table 00000028 0000000000000000 0000000000000000 00017808 2**3
> >> > CONTENTS, ALLOC, LOAD, RELOC, READONLY, DATA
> >> > 12 .note.GNU-stack 00000000 0000000000000000 0000000000000000 00027370 2**0
> >> > CONTENTS, READONLY
> >> >
> >> > So what am I missing to trigger this issue? Does it rely on something
> >> > like PGO, which is not upstream yet? A bit of handholding would be
> >> > much appreciated.
> >>
> >> Right, it relies on the PGO profile.
> >> On ChromeOS we collect the sample PGO profile from Arm devices with
> >> enabled CoreSight/ETM. You can find more details on ETM at
> >> https://www.kernel.org/doc/Documentation/trace/coresight/coresight.rst.
> >>
> >> https://github.com/Linaro/OpenCSD/blob/master/decoder/tests/auto-fdo/autofdo.md
> >> contains information about the pipeline of collecting, processing, and applying
> >> the profile.
> >>
> >
> > Generally the difficult part is in collecting a good matching profile for the workload.
> > So I think this patch is better than previous since it still keeps the compiler optimization for the hot code paths
> > in the file but removes the problematic section.
> >
> > Thanks,
> > Manoj
> >
> >
> >>
> >> >
> >> > Thanks,
> >> >
> >> > M.
> >> >
> >> > --
> >> > Without deviation from the norm, progress is not possible.
> >>
> >> Thanks,
> >> Denis