Re: riscv+KASAN does not boot

[Dmitry Vyukov <dvyukov@google.com>]

On Tue, Feb 16, 2021 at 12:25 PM Dmitry Vyukov <dvyukov@google.com> wrote:
>
> On Tue, Feb 16, 2021 at 12:17 PM Dmitry Vyukov <dvyukov@google.com> wrote:
> >
> > On Fri, Jan 29, 2021 at 9:11 AM Dmitry Vyukov <dvyukov@google.com> wrote:
> > > > I was fixing KASAN support for my sv48 patchset so I took a look at your
> > > > issue: I built a kernel on top of the branch riscv/fixes using
> > > > https://github.com/google/syzkaller/blob/269d24e857a757d09a898086a2fa6fa5d827c3e1/dashboard/config/linux/upstream-riscv64-kasan.config
> > > > and Buildroot 2020.11. I have the warnings regarding the use of
> > > > __virt_to_phys on wrong addresses (but that's normal since this function
> > > > is used in virt_addr_valid) but not the segfaults you describe.
> > >
> > > Hi Alex,
> > >
> > > Let me try to rebuild buildroot image. Maybe there was something wrong
> > > with my build, though, I did 'make clean' before doing. But at the
> > > same time it worked back in June...
> > >
> > > Re WARNINGs, they indicate kernel bugs. I am working on setting up a
> > > syzbot instance on riscv. If there a WARNING during boot then the
> > > kernel will be marked as broken. No further testing will happen.
> > > Is it a mis-use of WARN_ON? If so, could anybody please remove it or
> > > replace it with pr_err.
> >
> >
> > Hi,
> >
> > I've localized one issue with riscv/KASAN:
> > KASAN breaks VDSO and that's I think the root cause of weird faults I
> > saw earlier. The following patch fixes it.
> > Could somebody please upstream this fix? I don't know how to add/run
> > tests for this.
> > Thanks
> >
> > diff --git a/arch/riscv/kernel/vdso/Makefile b/arch/riscv/kernel/vdso/Makefile
> > index 0cfd6da784f84..cf3a383c1799d 100644
> > --- a/arch/riscv/kernel/vdso/Makefile
> > +++ b/arch/riscv/kernel/vdso/Makefile
> > @@ -35,6 +35,7 @@ CFLAGS_REMOVE_vgettimeofday.o = $(CC_FLAGS_FTRACE) -Os
> >  # Disable gcov profiling for VDSO code
> >  GCOV_PROFILE := n
> >  KCOV_INSTRUMENT := n
> > +KASAN_SANITIZE := n
> >
> >  # Force dependency
> >  $(obj)/vdso.o: $(obj)/vdso.so
>
>
>
> Second issue I am seeing seems to be related to text segment size.
> I check out v5.11 and use this config:
> https://gist.github.com/dvyukov/6af25474d455437577a84213b0cc9178
>
> Then trying to boot it using:
> QEMU emulator version 5.2.0 (Debian 1:5.2+dfsg-3)
> $ qemu-system-riscv64 -machine virt -smp 2 -m 4G ...
>
> It shows no output from the kernel whatsoever, even though I have
> earlycon and output shows very early with other configs.
> Kernel boots fine with defconfig and other smaller configs.
>
> If I enable KASAN_OUTLINE and CC_OPTIMIZE_FOR_SIZE, then this config
> also boots fine. Both of these options significantly reduce kernel
> size. However, I can also boot the kernel without these 2 configs, if
> I disable a whole lot of subsystem configs. This makes me think that
> there is an issue related to kernel size somewhere in
> qemu/bootloader/kernel bootstrap code.
> Does it make sense to you? Can somebody reproduce what I am seeing?



I am debugging the next issue with VDSO. clock_gettime is broken in
some weird way.
syzkaller has this function:

static uint64 current_time_ms(void)
{
        struct timespec ts;
        if (clock_gettime(CLOCK_MONOTONIC, &ts))
        //if (syscall(SYS_clock_gettime, CLOCK_MONOTONIC, &ts))
                fail("clock_gettime failed");
        return (uint64)ts.tv_sec * 1000 + (uint64)ts.tv_nsec / 1000000;
}

When using clock_gettime it producer some nonsense that breaks all
timeouts (in particular monotonic time goes backwards):
pid=4343 now=836038064151457975
pid=4343 now=836038064151457975
pid=4343 now=836038064151457970
pid=4343 now=836038064151457971

When I tested it calling real syscall, it works as expected:
pid=4876 now=2493379
pid=4876 now=2493392
pid=4876 now=2493395
pid=4876 now=2493409
pid=4876 now=2493414

Is it a known issue? Any ideas?