Re: [PATCH v1] s390x/tcg: Fix RISBHG

[Nick Desaulniers via <qemu-devel@nongnu.org>]

On Fri, Jan 8, 2021 at 1:45 AM David Hildenbrand <david@redhat.com> wrote:
>
> On 08.01.21 03:20, Nick Desaulniers wrote:
> > On Thu, Jan 7, 2021 at 3:27 PM David Hildenbrand <dhildenb@redhat.com> wrote:
> >>
> >>
> >>> Am 08.01.2021 um 00:21 schrieb Nick Desaulniers <ndesaulniers@google.com>:
> >>>
> >>> On Thu, Jan 7, 2021 at 3:13 PM David Hildenbrand <david@redhat.com> wrote:
> >>>>
> >>>> RISBHG is broken and currently hinders clang builds of upstream kernels
> >>>> from booting: the kernel crashes early, while decompressing the image.
> >>>>
> >>>>  [...]
> >>>>   Kernel fault: interruption code 0005 ilc:2
> >>>>   Kernel random base: 0000000000000000
> >>>>   PSW : 0000200180000000 0000000000017a1e
> >>>>         R:0 T:0 IO:0 EX:0 Key:0 M:0 W:0 P:0 AS:0 CC:2 PM:0 RI:0 EA:3
> >>>>   GPRS: 0000000000000001 0000000c00000000 00000003fffffff4 00000000fffffff0
> >>>>         0000000000000000 00000000fffffff4 000000000000000c 00000000fffffff0
> >>>>         00000000fffffffc 0000000000000000 00000000fffffff8 00000000008e25a8
> >>>>         0000000000000009 0000000000000002 0000000000000008 000000000000bce0
> >>>>
> >>>> One example of a buggy instruction is:
> >>>>
> >>>>    17dde:       ec 1e 00 9f 20 5d       risbhg  %r1,%r14,0,159,32
> >>>>
> >>>> With %r14 = 0x9 and %r1 = 0x7 should result in %r1 = 0x900000007, however,
> >>>> results in %r1 = 0.
> >>>>
> >>>> Let's interpret values of i3/i4 as documented in the PoP and make
> >>>> computation of "mask" only based on i3 and i4 and use "pmask" only at the
> >>>> very end to make sure wrapping is only applied to the high/low doubleword.
> >>>>
> >>>> With this patch, I can successfully boot a v5.10 kernel built with
> >>>> clang, and gcc builds keep on working.
> >>>>
> >>>> Fixes: 2d6a869833d9 ("target-s390: Implement RISBG")
> >>>> Reported-by: Nick Desaulniers <ndesaulniers@google.com>
> >>>> Cc: Guenter Roeck <linux@roeck-us.net>
> >>>> Cc: Christian Borntraeger <borntraeger@de.ibm.com>
> >>>> Signed-off-by: David Hildenbrand <david@redhat.com>
> >>>> ---
> >>>>
> >>>> This BUG was a nightmare to debug and the code a nightmare to understand.
> >>>>
> >>>> To make clang/gcc builds boot, the following fix is required as well on
> >>>> top of current master: "[PATCH] target/s390x: Fix ALGSI"
> >>>> https://lkml.kernel.org/r/20210107202135.52379-1-david@redhat.com
> >>>
> >>> In that case, a huge thank you!!! for this work! ++beers_owed.
> >>>
> >>
> >> :) a kernel build for z13 should work with the (default) „-cpu qemu“ cpu type.
> >
> > Hmm...so I don't think clang can build a Linux kernel image with
> > CONFIG_MARCH_Z13=y just yet; just defconfig.  Otherwise looks like
> > clang barfs on some of the inline asm constraints.
> >
>
> Ah, right. I overwrote my manual config by a temporary defconfig :)
>
>
> So, I'm on x86-64 F33.
>
> clang version 11.0.0 (Fedora 11.0.0-2.fc33)
> LLVM version 11.0.0
>
> I cannot directly use "LLVM=1" for cross-compilation, as I keep getting
> "error: unknown emulation: elf64_s390" from ld.lld and "error: invalid
> output format: 'elf64-s390'" from llvm-objcopy. I assume that's fixed in
> llvm12?

Right, I suspect that even if ld.lld understood that emulation mode
target, it would still fail due to lack of big endian support.  We've
been building with simply `CC=clang` for s390 linux kernels.
Via: https://www.kernel.org/doc/html/latest/kbuild/llvm.html#llvm-utilities
we usually start with `make CC=clang` then work our way up to `make
LLVM=1`.  So you shouldn't need the below patching, just use
`CC=clang`.

>
> 1. I patch around it (strange, I remember CC= .. used to work, but it no
> longer does)
>
> ---
>
> index e30cf02da8b8..89c57062ed5d 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -427,13 +427,13 @@ KBUILD_HOSTLDLIBS   := $(HOST_LFS_LIBS) $(HOSTLDLIBS)
>  CPP            = $(CC) -E
>  ifneq ($(LLVM),)
>  CC             = clang
> -LD             = ld.lld
> -AR             = llvm-ar
> -NM             = llvm-nm
> -OBJCOPY                = llvm-objcopy
> -OBJDUMP                = llvm-objdump
> -READELF                = llvm-readelf
> -STRIP          = llvm-strip
> +LD             = $(CROSS_COMPILE)ld
> +AR             = $(CROSS_COMPILE)ar
> +NM             = $(CROSS_COMPILE)nm
> +OBJCOPY                = $(CROSS_COMPILE)objcopy
> +OBJDUMP                = $(CROSS_COMPILE)objdump
> +READELF                = $(CROSS_COMPILE)readelf
> +STRIP          = $(CROSS_COMPILE)strip
>  else
>  CC             = $(CROSS_COMPILE)gcc
>  LD             = $(CROSS_COMPILE)ld
>
> ---

Pulling from your github branch, everything looks good; buildroot
support looks good. I'll wire this up to our CI so that we can help
report regressions!
-- 
Thanks,
~Nick Desaulniers