Re: [PATCH v1 1/4] xen: introduce CONFIG_GENERIC_BUG_FRAME

[Oleksii <oleksii.kurochko@xxxxxxxxx>]

On Thu, 2023-02-16 at 12:44 +0200, Oleksii wrote:
> On Thu, 2023-02-16 at 08:31 +0100, Jan Beulich wrote:
> > On 15.02.2023 18:59, Oleksii wrote:
> > > Hello Jan and community,
> > > 
> > > I experimented and switched RISC-V to x86 implementation. All
> > > that
> > > I
> > > changed in x86 implementation for RISC-V was _ASM_BUGFRAME_TEXT.
> > > Other
> > > things are the same as for x86.
> > > 
> > > For RISC-V it is fine to skip '%c' modifier so _ASM_BUGFRAME_TEXT
> > > will
> > > look like:
> > > 
> > > #define _ASM_BUGFRAME_TEXT(second_frame) \
> > >     ".Lbug%=: ebreak\n"   
> > >     ".pushsection .bug_frames.%[bf_type], \"a\", @progbits\n"
> > >     ".p2align 2\n"
> > >     ".Lfrm%=:\n"
> > >     ".long (.Lbug%= - .Lfrm%=) + %[bf_line_hi]\n"
> > >     ".long (%[bf_ptr] - .Lfrm%=) + %[bf_line_lo]\n"
> > >     ".if " #second_frame "\n"
> > >     ".long 0, %[bf_msg] - .Lfrm%=\n"
> > >     ".endif\n"
> > >     ".popsection\n"
> > 
> > I expect this could be further abstracted such that only the actual
> > instruction is arch-specific.
> Generally, the only thing that can't be abstracted ( as you mentioned
> is an instruction ).
> 
> So we can make additional defines:
>   1. #define MODIFIER "" or "c" (depends on architecture) and use it
>  
> the following way:
>    ".pushsection .bug_frames.%"MODIFIER"[bf_type], \"a\",
> @progbits\n"
>    ...
>   2. #define BUG_INSTR "ebreak" | "ud2" | "..." and use it in the
> following way:
>    ".Lbug%=: "BUG_ISNTR"\n"
>    ...
> Except for these defines which should be specified for each
> architecture
> all other code will be the same for all architectures.
> 
> But as I understand with modifier 'c' can be issues that not all
> compilers support but for ARM and  x86 before immediate is present
> punctuation # or $ which are removed by modifier 'c'. And I am
> wondering what other ways to remove punctuation before immediate
> exist.
> 
> Do you think we should consider that as an issue?
> 
> > 
> > > The only thing I am worried about is:
> > > 
> > > #define _ASM_BUGFRAME_INFO(type, line, ptr, msg) \
> > >   [bf_type] "i" (type), ...
> > > because as I understand it can be an issue with 'i' modifier in
> > > case of
> > > PIE. I am not sure that Xen enables PIE somewhere but still...
> > > If it is not an issue then we can use x86 implementation as a
> > > generic
> > > one.
> > 
> > "i" is not generally an issue with PIE, it only is when the value
> > is
> > the
> > address of a symbol. Here "type" is a constant in all cases. (Or
> > else
> > how would you express an immediate operand of an instruction in an
> > asm()?)
> Hmm. I don't know other ways to express an immediate operand except
> 'i'.
> 
> It looks like type, line, msg are always 'constants' 
> (
> they possibly can be passed without 'i' and used directly inside
> asm():
>    ...
>    ".pushsection .bug_frames." __stringify(type) ", \"a\",
> %progbits\n"\
>    ...
> ) but the issue will be with 'ptr' for which we have to use 'i'.
> 
> And I am not sure about all 'constants'. For example, I think that it
> can be an issue with 'line' = '((line & ((1 << BUG_LINE_LO_WIDTH) -
> 1))
> << BUG_DISP_WIDTH)' if we will use that directly inside asm(...).
> 
I think we can avoid 'i' by using 'r' + some kind of mov instruction to
write a register value to memory. The code below is pseudo-code:
#define _ASM_BUGFRAME_TEXT(second_frame)
    ...
    "loc_disp:\n"
    "    .long 0x0"
    "mov [loc_disp], some_register"
    ...
And the we have to define mov for each architecture.
                                            \
> > 
> > Jan
> 
> ~ Oleksii