Re: [PATCH] module interface improvement for kprobes

From: Christoph Hellwig
Date: Sat Aug 05 2006 - 07:33:23 EST


On Fri, Aug 04, 2006 at 01:30:39PM -0500, David Smith wrote:
> Why shouldn't I put a probe into a module other than at a symbol I can
> find with kallsyms? For example, I'm interested when a particular
> module hits an error condition that occurs. I don't want to probe how
> many times the function gets called - just when the error condition
> occurs.

How do you find that offset? You'll probably mention the S-Word but
we really want something that works with the latest kernel, not just
the vendor trees.

> With the existing interface, if I use kallsysms to find the value of a
> symbol, the module can be unloaded between the time I use kallsyms and
> register the kprobe. The patch I included fixes that race condition by
> incrementing the module reference count.

Yes, and that's a good thing. But the interface for doing it is wrong.
You don't really want the users to do all that by itself. For the typical
case of putting a probe at the usual points you want an interface that
puts in the probe given a name and does the right thing for you. For example
the interface I proposed in my last mail. Adding another field to struct
kprobe to specify an offset into the symbol would be the logical extension
of that.

> Your example works for a very small number of symbols, but with a large
> number it could take a long time to register the kprobes. Plus, that
> would need to be done every time the kprobe was registered. With my
> patch, the symbol lookup can be done once, then all those symbols can be
> turned into offsets from the base address of the module.

Registering a kprobe is everything but a fastpath, and you definitly should
not have a lot of probes anyway. It's far more worthwhile to have a sane
interface that the user can't get wrong then a small speedup in something
that's not a fastpath. I think Rusty even has a paper or talk about why
this is absolutely nessecary :)
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/