[lkml]   [2014]   [Oct]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] kprobes: add kprobe_is_function_probed()
On Tue, 21 Oct 2014, Josh Poimboeuf wrote:

> > This is a rather difficult call actually. I am of course aware of the fact
> > that kernel fucntions can't be uniquely identified by name, but when
> > thinking about this, one has to consider:
> >
> > - ftrace primary userspace interface (set_ftrace_filter) is based on
> > function names
> > - kprobe tracer and uprobe trace primary userspace interface
> > (/sys/kernel/debug/tracing/kprobe_events and uprobe_events respectively)
> > are primarily based on function names
> True, the user space interfaces are done by name, but the kernel
> interfaces aren't necessarily so (see ftrace_set_filter_ip and struct
> kprobe.addr). This is a kernel interface so we can be more precise.

We could probably have both interfaces available (i.e. allowing lookup by
name or by address range, and let the caller decide/handle the
potential duplicates).

> > - the number of conflicts is probably zero, or very close to zero. Try to
> > run this
> >
> > cut -f3 -d' ' /proc/kallsyms | sort | uniq -c
> >
> > So it's questionable whether all the back and forth name->address->name
> > translations all over the place are worth all the trouble.
> On my kernel:
> $ grep ' t \| T ' /proc/kallsyms | awk '{print $3}' |sort |uniq -d |wc -l
> 395
> So there are at least 395 places where this could go wrong...

This is broken anyway ... this will add up a lot of unrelated things
together (umask_show, _iommu_cpumask_show, __uncore_umask_show,
wq_cpumask_show, etc). I believe looking at

cut -f3 -d' ' /proc/kallsyms | sort | uniq -c | sort -nr -k1

gives slightly better picture.

Also keep in mind that a *lot* of the hits are not functions.

> With kpatch, we're setting the ftrace filter by address so we don't
> patch the wrong function. So we already have the address. We also have
> the function length, which is needed for our backtrace safety checks.
> I'm guessing kGraft doesn't have the address + length? I think you
> could call kallsyms_lookup() to get both values.
> Maybe we should see what our unified live patching code ends up looking
> like before deciding what interface(s) we need here?

Yes, that probably makes sense indeed. I am talking to David Miller wrt.
mailinglist creation on as we speak, hopefully it'll
materialize soon.

> > Do we need to be race-free here? If userspace is both instantiating new
> > krpobe and initiating live kernel patching at the "same time", I don't
> > think kernel should try to solve such race ... it's simply there by
> > definition, depending on, for example, scheduling decisions.
> If we're not preventing it, but instead just printing a warning, then
> yeah, that sounds fine to me.
> I think it would also be nice to do the reverse, i.e. have kprobes emit
> a warning if someone tries to probe a replaced function.

Indeed, good idea!


Jiri Kosina

 \ /
  Last update: 2014-10-22 00:01    [W:0.043 / U:3.924 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site