signal/x86: Delay calling signals in atomic

On x86_64 we must disable preemption before we enable interrupts
for stack faults, int3 and debugging, because the current task is using
a per CPU debug stack defined by the IST. If we schedule out, another task
can come in and use the same stack and cause the stack to be corrupted
and crash the kernel on return.

When CONFIG_PREEMPT_RT is enabled, spin_locks become mutexes, and
one of these is the spin lock used in signal handling.

Some of the debug code (int3) causes do_trap() to send a signal.
This function calls a spin lock that has been converted to a mutex
and has the possibility to sleep. If this happens, the above issues with
the corrupted stack is possible.

Instead of calling the signal right away, for PREEMPT_RT and x86_64,
the signal information is stored on the stacks task_struct and
TIF_NOTIFY_RESUME is set. Then on exit of the trap, the signal resume
code will send the signal when preemption is enabled.

[ rostedt: Switched from #ifdef CONFIG_PREEMPT_RT to
  ARCH_RT_DELAYS_SIGNAL_SEND and added comments to the code. ]


Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
[bigeasy: also needed on 32bit as per Yang Shi <yang.shi@linaro.org>]
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>

arch/x86/include/asm/signal.h

@@ -28,6 +28,19 @@ typedef struct {
 #define SA_IA32_ABI	0x02000000u
 #define SA_X32_ABI	0x01000000u
 
+/*
+ * Because some traps use the IST stack, we must keep preemption
+ * disabled while calling do_trap(), but do_trap() may call
+ * force_sig_info() which will grab the signal spin_locks for the
+ * task, which in PREEMPT_RT are mutexes.  By defining
+ * ARCH_RT_DELAYS_SIGNAL_SEND the force_sig_info() will set
+ * TIF_NOTIFY_RESUME and set up the signal to be sent on exit of the
+ * trap.
+ */
+#if defined(CONFIG_PREEMPT_RT)
+#define ARCH_RT_DELAYS_SIGNAL_SEND
+#endif
+
 #ifndef CONFIG_COMPAT
 #define compat_sigset_t compat_sigset_t
 typedef sigset_t compat_sigset_t;

include/linux/sched.h

@@ -1080,6 +1080,10 @@ struct task_struct {
 	/* Restored if set_restore_sigmask() was used: */
 	sigset_t			saved_sigmask;
 	struct sigpending		pending;
+#ifdef CONFIG_PREEMPT_RT
+	/* TODO: move me into ->restart_block ? */
+	struct				kernel_siginfo forced_info;
+#endif
 	unsigned long			sas_ss_sp;
 	size_t				sas_ss_size;
 	unsigned int			sas_ss_flags;

kernel/entry/common.c

@@ -162,6 +162,14 @@ static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
 		if (ti_work & _TIF_NEED_RESCHED)
 			schedule();
 
+#ifdef ARCH_RT_DELAYS_SIGNAL_SEND
+		if (unlikely(current->forced_info.si_signo)) {
+			struct task_struct *t = current;
+			force_sig_info(&t->forced_info);
+			t->forced_info.si_signo = 0;
+		}
+#endif
+
 		if (ti_work & _TIF_UPROBE)
 			uprobe_notify_resume(regs);
 

kernel/signal.c

@@ -1330,6 +1330,34 @@ force_sig_info_to_task(struct kernel_siginfo *info, struct task_struct *t, bool
 	struct k_sigaction *action;
 	int sig = info->si_signo;
 
+	/*
+	 * On some archs, PREEMPT_RT has to delay sending a signal from a trap
+	 * since it can not enable preemption, and the signal code's spin_locks
+	 * turn into mutexes. Instead, it must set TIF_NOTIFY_RESUME which will
+	 * send the signal on exit of the trap.
+	 */
+#ifdef ARCH_RT_DELAYS_SIGNAL_SEND
+	if (in_atomic()) {
+		struct task_struct *t = current;
+
+		if (WARN_ON_ONCE(t->forced_info.si_signo))
+			return 0;
+
+		if (is_si_special(info)) {
+			WARN_ON_ONCE(info != SEND_SIG_PRIV);
+			t->forced_info.si_signo = info->si_signo;
+			t->forced_info.si_errno = 0;
+			t->forced_info.si_code = SI_KERNEL;
+			t->forced_info.si_pid = 0;
+			t->forced_info.si_uid = 0;
+		} else {
+			t->forced_info = *info;
+		}
+
+		set_tsk_thread_flag(t, TIF_NOTIFY_RESUME);
+		return 0;
+	}
+#endif
 	spin_lock_irqsave(&t->sighand->siglock, flags);
 	action = &t->sighand->action[sig-1];
 	ignored = action->sa.sa_handler == SIG_IGN;