[syzbot] riscv/fixes boot error: can't ssh into the instance

[Aleksandr Nogikh]

On Wed, Feb 16, 2022 at 12:56 PM Dmitry Vyukov <dvyukov at google.com> wrote:
>
> On Wed, 16 Feb 2022 at 12:47, Aleksandr Nogikh <nogikh at google.com> wrote:
> >
> > On Wed, Feb 16, 2022 at 11:37 AM Aleksandr Nogikh <nogikh at google.com> wrote:
> > >
> > > Hi Alex,
> > >
> > > On Wed, Feb 16, 2022 at 5:14 AM Alexandre Ghiti <alex at ghiti.fr> wrote:
> > > >
> > > > Hi Dmitry,
> > > >
> > > > On 2/15/22 18:12, Dmitry Vyukov wrote:
> > > > > On Wed, 2 Feb 2022 at 14:18, Alexandre Ghiti
> > > > > <alexandre.ghiti at canonical.com> wrote:
> > > > >> Hi Aleksandr,
> > > > >>
> > > > >> On Wed, Feb 2, 2022 at 12:08 PM Aleksandr Nogikh <nogikh at google.com> wrote:
> > > > >>> Hello,
> > > > >>>
> > > > >>> syzbot has already not been able to fuzz its RISC-V instance for 97
> > > > >> That's a longtime, I'll take a look more regularly.
> > > > >>
> > > > >>> days now because the compiled kernel cannot boot. I bisected the issue
> > > > >>> to the following commit:
> > > > >>>
> > > > >>> commit 54c5639d8f507ebefa814f574cb6f763033a72a5
> > > > >>> Author: Alexandre Ghiti <alexandre.ghiti at canonical.com>
> > > > >>> Date:   Fri Oct 29 06:59:27 2021 +0200
> > > > >>>
> > > > >>>      riscv: Fix asan-stack clang build
> > > > >>>
> > > > >>> Apparently, the problem appears on GCC-built RISC-V kernels with KASAN
> > > > >>> enabled. In the previous message syzbot mentions
> > > > >>> "riscv64-linux-gnu-gcc (Debian 10.2.1-6) 10.2.1 20210110, GNU ld (GNU
> > > > >>> Binutils for Debian) 2.35.2", but the issue also reproduces finely on
> > > > >>> a newer GCC compiler: "riscv64-linux-gnu-gcc (Debian 11.2.0-10)
> > > > >>> 11.2.0, GNU ld (GNU Binutils for Debian) 2.37".
> > > > >>> For convenience, I also duplicate the .config file from the bot's
> > > > >>> message: https://syzkaller.appspot.com/x/.config?x=522544a2e0ef2a7d
> > > > >>>
> > > > >>> Can someone with KASAN and RISC-V expertise please take a look?
> > > > >> I'll take a look at that today.
> > > > >>
> > > > >> Thanks for reporting the issue,
> > > > >
> > > >
> > > > I took a quick look, not enough to fix it but I know the issue comes
> > > > from the inline instrumentation, I have no problem with the outline
> > > > instrumentation. I need to find some cycles to work on this, my goal is
> > > > to fix this for 5.17.
> > >
> > > Thanks for the update!
> > >
> > > Can you please share the .config with which you tested the outline
> > > instrumentation?
> > > I updated the syzbot config to use KASAN_OUTLINE instead of KASAN_INLINE,
> > > but it still does not boot :(
> > >
> > > Here's what I used:
> > > https://gist.github.com/a-nogikh/279c85c2d24f47efcc3e865c08844138
> >
> > Update: it doesn't boot with that big config, but boots if I generate
> > a simple one with KASAN_OUTLINE:
> >
> > make defconfig ARCH=riscv CROSS_COMPILE=riscv64-linux-gnu-
> > ./scripts/config -e KASAN -e KASAN_OUTLINE
> > make olddefconfig ARCH=riscv CROSS_COMPILE=riscv64-linux-gnu-
> >
> > And it indeed doesn't work if I use KASAN_INLINE.
>
> It may be an issue with code size. Full syzbot config + KASAN + KCOV
> produce hugely massive .text. It may be hitting some limitation in the
> bootloader/kernel bootstrap code.

I bisected the difference between the config we use on syzbot and the
simple one that was generated like I described above.
Turns out that it's the DEBUG_VIRTUAL config that makes the difference.

make defconfig ARCH=riscv CROSS_COMPILE=riscv64-linux-gnu-
./scripts/config -e KASAN -e KASAN_OUTLINE -e DEBUG_VIRTUAL
make olddefconfig ARCH=riscv CROSS_COMPILE=riscv64-linux-gnu-

And the resulting kernel does not boot.
My env: the `riscv/fixes` branch, commit
6df2a016c0c8a3d0933ef33dd192ea6606b115e3, qemu 6.2.0.