[RFC PATCH] ARM: Use generic BUG() handler

[Dave Martin]

On Thu, Mar 3, 2011 at 2:11 PM, Russell King - ARM Linux
<linux at arm.linux.org.uk> wrote:
> On Wed, Mar 02, 2011 at 10:59:50AM +0000, Dave Martin wrote:
>> +     asm volatile("1:\t" __bug_instr "\n"            q       \
>> +                  ".pushsection .rodata.str, \"a\"\n"        \
>> +                  "2:\t.asciz \"" #__file "\"\n"             \
>> +                  ".popsection\n"                            \
>
> Doesn't this mean we end up with multiple file names?

Hmmm, yes, you're right.

After a bit of digging in the documentation, I find that ELF supports
mergable string sections though, so

.pushsection .rodata.str, "aMS", 1
    .asciz __FILE__
.popsection

...actually looks like it ought to do the right thing: duplicate
strings in the section get merged during linking, even if there are
duplicates from a single object.

Simple experiments suggest that the linker does this right, even
merging distinct strings which share a common suffix.  Do we use this
elsewhere in the kernel?  It it's not already used, it could be a win
for any large, static string tables.

>
>> +                  ".pushsection __bug_table,\"a\"\n"         \
>> +                  "3:\t.word 1b, 2b\n"                       \
>> +                  "\t.hword " #__line ", 0\n"                \
>> +                  ".popsection"                              \
>> +     unreachable();                                          \
>> +} while (0)
>
> Second problem is that the above produces this:
>
>        1:      .word 0xec000000
> .pushsection .rodata.str, "a"
> 2:      .asciz "__FILE__"
> .popsection
> .pushsection __bug_table,"a"
> 3:      .word 1b, 2b
>        .hword __LINE__, 0
> .popsection
>
> which is clearly not what we want.

Indeed.  I did say I hadn't tested it :)

> Adding another level of indirection
> starts to get closer to what we desire:
>
>        1:      .word 0xec000000
> .pushsection .rodata.str, "a"
> 2:      .asciz ""t.c""
> .popsection
> .pushsection __bug_table,"a"
> 3:      .word 1b, 2b
>        .hword 19, 0
> .popsection
>
> but we're still ending up with multiple strings containing the filename,
> which is going to excessively bloat the kernel no end.
>

We'd need to remove the duplicate quotes, too.

For the .org <blah> + sizeof (struct bug_entry) problem, Will pointed
out that there's a BUILD_BUG_ON() macro which we could used to avoid
silently producing a bad build if the result of that sizeof isn't what
we expect.

So I still think it could work ... but it's a bit of an ugly hack I confess.

Cheers
---Dave