arm64: debug: avoid resetting stepping state machine when TIF_SINGLESTEP

commit 3a402a709500c5a3faca2111668c33d96555e35a upstream.

When TIF_SINGLESTEP is set for a task, the single-step state machine is
enabled and we must take care not to reset it to the active-not-pending
state if it is already in the active-pending state.

Unfortunately, that's exactly what user_enable_single_step does, by
unconditionally setting the SS bit in the SPSR for the current task.
This causes failures in the GDB testsuite, where GDB ends up missing
expected step traps if the instruction being stepped generates another
trap, e.g. PTRACE_EVENT_FORK from an SVC instruction.

This patch fixes the problem by preserving the current state of the
stepping state machine when TIF_SINGLESTEP is set on the current thread.

Cc: <stable@vger.kernel.org>
Reported-by: Yao Qi <yao.qi@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Willy Tarreau <w@1wt.eu>

diff --git a/arch/arm64/kernel/debug-monitors.c b/arch/arm64/kernel/debug-monitors.c
index f03802c63f79992acc5354849ef79adba1e7f1dd..a49bf370bb6af7a293c5a56aa0856cc063cae1a2 100644 (file)
--- a/arch/arm64/kernel/debug-monitors.c
+++ b/arch/arm64/kernel/debug-monitors.c
@@ -459,8 +459,10 @@ int kernel_active_single_step(void)
 /* ptrace API */
 void user_enable_single_step(struct task_struct *task)
 {
-       set_ti_thread_flag(task_thread_info(task), TIF_SINGLESTEP);
-       set_regs_spsr_ss(task_pt_regs(task));
+       struct thread_info *ti = task_thread_info(task);
+
+       if (!test_and_set_ti_thread_flag(ti, TIF_SINGLESTEP))
+               set_regs_spsr_ss(task_pt_regs(task));
 }
 
 void user_disable_single_step(struct task_struct *task)