On Wed, 17 May 2023 12:47:42 +0200
Jiri Olsa <olsajiri@xxxxxxxxx> wrote:
> On Wed, May 17, 2023 at 11:45:07AM +0800, Ze Gao wrote:
> > Current implementation calls kprobe related functions before doing
> > ftrace recursion check in fprobe_kprobe_handler, which opens door
> > to kernel crash due to stack recursion if preempt_count_{add, sub}
> > is traceable in kprobe_busy_{begin, end}.
> >
> > Things goes like this without this patch quoted from Steven:
> > "
> > fprobe_kprobe_handler() {
> > kprobe_busy_begin() {
> > preempt_disable() {
> > preempt_count_add() { <-- trace
> > fprobe_kprobe_handler() {
> > [ wash, rinse, repeat, CRASH!!! ]
> > "
> >
> > By refactoring the common part out of fprobe_kprobe_handler and
> > fprobe_handler and call ftrace recursion detection at the very beginning,
> > the whole fprobe_kprobe_handler is free from recursion.
> >
> > Signed-off-by: Ze Gao <zegao@xxxxxxxxxxx>
> > Acked-by: Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
> > Link: https://lore.kernel.org/linux-trace-kernel/20230516071830.8190-3-zegao@xxxxxxxxxxx
> > ---
> > kernel/trace/fprobe.c | 59 ++++++++++++++++++++++++++++++++-----------
> > 1 file changed, 44 insertions(+), 15 deletions(-)
> >
> > diff --git a/kernel/trace/fprobe.c b/kernel/trace/fprobe.c
> > index 9abb3905bc8e..097c740799ba 100644
> > --- a/kernel/trace/fprobe.c
> > +++ b/kernel/trace/fprobe.c
> > @@ -20,30 +20,22 @@ struct fprobe_rethook_node {
> > char data[];
> > };
> >
> > -static void fprobe_handler(unsigned long ip, unsigned long parent_ip,
> > - struct ftrace_ops *ops, struct ftrace_regs *fregs)
> > +static inline void __fprobe_handler(unsigned long ip, unsigned long
> > + parent_ip, struct ftrace_ops *ops, struct ftrace_regs *fregs)
> > {
> > struct fprobe_rethook_node *fpr;
> > struct rethook_node *rh = NULL;
> > struct fprobe *fp;
> > void *entry_data = NULL;
> > - int bit, ret;
> > + int ret;
> >
>
> this change uncovered bug for me introduced by [1]
>
> the bpf's kprobe multi uses either fprobe's entry_handler or exit_handler,
> so the 'ret' value is undefined for return probe path and occasionally we
> won't setup rethook and miss the return probe
Oops, I missed to push my fix.
https://lore.kernel.org/all/168100731160.79534.374827110083836722.stgit@devnote2/
>
> we can either squash this change into your patch or I can make separate
> patch for that.. but given that [1] is quite recent we could just silently
> fix that ;-)
Jiri, I think the above will fix the issue, right?
>
> jirka
>
>
> [1] 39d954200bf6 fprobe: Skip exit_handler if entry_handler returns !0
>
> ---
> diff --git a/kernel/trace/fprobe.c b/kernel/trace/fprobe.c
> index 9abb3905bc8e..293184227394 100644
> --- a/kernel/trace/fprobe.c
> +++ b/kernel/trace/fprobe.c
> @@ -27,7 +27,7 @@ static void fprobe_handler(unsigned long ip, unsigned long parent_ip,
> struct rethook_node *rh = NULL;
> struct fprobe *fp;
> void *entry_data = NULL;
> - int bit, ret;
> + int bit, ret = 0;
>
> fp = container_of(ops, struct fprobe, ops);
> if (fprobe_disabled(fp))
>
>
--
Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
