On 07/11/2013 03:09 AM, Jiri Kosina wrote:
>>
>> I'm wondering if it would be easier/more general to just return to the
>> instruction. The "more general" bit would allow this to be used for
>> other things, like alternatives,
>
> As Boris already pointed out, this is not really that interesting, as it's
> being done through text_poke_early(), which is rather a different story
> anyway.
>
>> and perhaps eventually dynamic call patching.
>
> Umm ... could you please elaborate either what exactly do you mean by
> that, or why it can't be used currently as-is?
Dynamic call patching would be changing a CALL instruction *emitted by
the compiler* (and therefore lacking any metadata annotation) from one
target function to another. Because it lacks metadata annotations, we
can't do this as a "big bang" (all at the same time) but rather would
have to do it on demand (the original CALL would point to a "patch me"
routine.) This means a lot of patching cycles; stop_machine() is a
total nonstarter, even IPIs might be too expensive.
There is an alternative, which is postprocessing the executable to
generate metadata, but that has its own trickiness.
>> Returning to the instruction will, in effect, be a busy-wait for the
>> faulted CPU until the patch is complete; more or less what stop_machine
>> would do, but only for a CPU which actually strays into the affected
>> region.
>
> To be honest, I fail to see a clear advantage ... we don't avoid any extra
> IPI by it, and wrt. "correctness", the end result is the same.
>
The current code assumes that one of the two code sequences is a NOP,
and therefore that jumping over the region is legal. This does not
allow for transitioning one active code sequence to another.
-hpa
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
