Re: [PATCHv4 bpf-next 2/6] uprobe: Do not emulate/sstep original instruction when ip is changed

On Tue, Sep 16, 2025 at 03:28:52PM -0700, Andrii Nakryiko wrote:

On Tue, Sep 16, 2025 at 2:53 PM Jiri Olsa [off-list ref] wrote:

quoted

If uprobe handler changes instruction pointer we still execute single
step) or emulate the original instruction and increment the (new) ip
with its length.

This makes the new instruction pointer bogus and application will
likely crash on illegal instruction execution.

If user decided to take execution elsewhere, it makes little sense
to execute the original instruction, so let's skip it.

Acked-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
 kernel/events/uprobes.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/kernel/events/uprobes.c b/kernel/events/uprobes.c
index 7ca1940607bd..2b32c32bcb77 100644
--- a/kernel/events/uprobes.c
+++ b/kernel/events/uprobes.c

@@ -2741,6 +2741,13 @@ static void handle_swbp(struct pt_regs *regs)

        handler_chain(uprobe, regs);

+       /*
+        * If user decided to take execution elsewhere, it makes little sense
+        * to execute the original instruction, so let's skip it.
+        */
+       if (instruction_pointer(regs) != bp_vaddr)
+               goto out;
+

Peter, Ingo,

Are you guys ok with us routing this through the bpf-next tree? We'll
have a tiny conflict because in perf/core branch there is
arch_uprobe_optimize() call added after handler_chain(), so git merge
will be a bit confused, probably. But it should be trivially
resolvable.

Nah, I suppose that'll be fine. Thanks!