Re: [RFC PATCH 6/7] x86/kexec: Debugging support: Dump registers on exception

[H. Peter Anvin]

On 11/5/24 12:38, Woodhouse, David wrote:
> On Sun, 2024-11-03 at 05:35 +0000, David Woodhouse wrote:
> > +
> > +/* Print the byte in %bl, clobber %rax */
> > +SYM_CODE_START_LOCAL_NOALIGN(pr_byte)
> > +       movb    %bl, %al
> > +       nop
> > +       andb    $0x0f, %al
> > +       addb    $0x30, %al
> > +       cmpb    $0x3a, %al
> > +       jb      1f
> > +       addb    $('a' - '0' - 10), %al
> > +1:     pr_char
> > +       ANNOTATE_UNRET_SAFE
> > +       ret
> > +SYM_CODE_END(pr_byte)
> > +
>
> Obviously that function name (and comment) are wrong; fixed in my tree.
> at
> https://git.infradead.org/users/dwmw2/linux.git/shortlog/refs/heads/kexec-debug
>
> This function (and also pr_qword) are also what objtool is complaining
> about:
>
> vmlinux.o: warning: objtool: relocate_range+0x2f6: unreachable instruction
> vmlinux.o: warning: objtool: relocate_range+0x305: unreachable instruction
>
> I don't quite see why, because pr_qword() quite blatantly calls
> pr_nyblle(), as it's now named. And exc_handler() repeatedly calls
> pr_qword().
>
> But most of the objtool annotations I've added here were just to make
> it shut up and build, without much though. Peter, Josh, any chance you
> can help me fix it up please?
>
> It would also be really useful if objtool would let me have data inside
> a "code" segment, without complaining that it can't decode it as
> instructions — and without also failing to decode the first instruction
> of the *subsequent* function. I've put the GDT at the end to work
> around that, but it's a bit nasty.

Looking at your code, you have a much bigger problem here:

+/*
+ * This allows other types of serial ports to be used.
+ *  - %al: Character to be printed (no clobber %rax)
+ *  - %rdx: MMIO address or port.
+ */
+.macro pr_char
+       outb    %al, %dx
+.endm
+

This will overflow your UART buffer very quickly since you are now 
dumping a whole bunch of data. The URT buffer -- if you even have one 
and it is enabled -- is only 16 bytes in a standard 16550A UART. In 
older UARTs (or emulated older UARTs) you might not have a buffer at 
all. To print more than a handful of bytes, you need to poll for the 
THRE bit=1 (bit 5 of register 5).

What is the point of writing this code in assembly in the first place? A 
much more logical thing to do is to just push the registers you haven't 
pushed already onto the stack and call a C function to do the actual 
dumping? It isn't like it is in any shape, way or form performance critical.

	-hpa