Merge tag 'core-core-2022-03-21' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip

Pull core process handling RT latency updates from Thomas Gleixner:

 - Reduce the amount of work to release a task stack in context switch.
   There is no real reason to do cgroup accounting and memory freeing in
   this performance sensitive context.

   Aside of this the invoked functions cannot be called from this
   preemption disabled context on PREEMPT_RT enabled kernels. Solve this
   by moving the accounting into do_exit() and delaying the freeing of
   the stack unless the vmap stack can be cached.

 - Provide a mechanism to delay raising signals from atomic context on
   PREEMPT_RT enabled kernels as sighand::lock cannot be acquired. Store
   the information in the task struct and raise it in the exit path.

* tag 'core-core-2022-03-21' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  signal, x86: Delay calling signals in atomic on RT enabled kernels
  fork: Use IS_ENABLED() in account_kernel_stack()
  fork: Only cache the VMAP stack in finish_task_switch()
  fork: Move task stack accounting to do_exit()
  fork: Move memcg_charge_kernel_stack() into CONFIG_VMAP_STACK
  fork: Don't assign the stack pointer in dup_task_struct()
  fork, IA64: Provide alloc_thread_stack_node() for IA64
  fork: Duplicate task_struct before stack allocation
  fork: Redo ifdefs around task stack handling

arch/ia64/include/asm/thread_info.h

@@ -55,15 +55,15 @@ struct thread_info {
 #ifndef ASM_OFFSETS_C
 /* how to get the thread information struct from C */
 #define current_thread_info()	((struct thread_info *) ((char *) current + IA64_TASK_SIZE))
-#define alloc_thread_stack_node(tsk, node)	\
+#define arch_alloc_thread_stack_node(tsk, node)	\
 		((unsigned long *) ((char *) (tsk) + IA64_TASK_SIZE))
 #define task_thread_info(tsk)	((struct thread_info *) ((char *) (tsk) + IA64_TASK_SIZE))
 #else
 #define current_thread_info()	((struct thread_info *) 0)
-#define alloc_thread_stack_node(tsk, node)	((unsigned long *) 0)
+#define arch_alloc_thread_stack_node(tsk, node)	((unsigned long *) 0)
 #define task_thread_info(tsk)	((struct thread_info *) 0)
 #endif
-#define free_thread_stack(tsk)	/* nothing */
+#define arch_free_thread_stack(tsk)	/* nothing */
 #define task_stack_page(tsk)	((void *)(tsk))
 
 #define __HAVE_THREAD_FUNCTIONS

arch/x86/Kconfig

@@ -120,6 +120,7 @@ config X86
 	select ARCH_WANTS_NO_INSTR
 	select ARCH_WANT_HUGE_PMD_SHARE
 	select ARCH_WANT_LD_ORPHAN_WARN
+	select ARCH_WANTS_RT_DELAYED_SIGNALS
 	select ARCH_WANTS_THP_SWAP		if X86_64
 	select ARCH_HAS_PARANOID_L1D_FLUSH
 	select BUILDTIME_TABLE_SORT

include/linux/sched.h

@@ -1090,6 +1090,9 @@ struct task_struct {
 	/* Restored if set_restore_sigmask() was used: */
 	sigset_t			saved_sigmask;
 	struct sigpending		pending;
+#ifdef CONFIG_RT_DELAYED_SIGNALS
+	struct kernel_siginfo		forced_info;
+#endif
 	unsigned long			sas_ss_sp;
 	size_t				sas_ss_size;
 	unsigned int			sas_ss_flags;

include/linux/sched/task_stack.h

@@ -79,6 +79,8 @@ static inline void *try_get_task_stack(struct task_struct *tsk)
 static inline void put_task_stack(struct task_struct *tsk) {}
 #endif
 
+void exit_task_stack_account(struct task_struct *tsk);
+
 #define task_stack_end_corrupted(task) \
 		(*(end_of_stack(task)) != STACK_END_MAGIC)
 

kernel/Kconfig.preempt

@@ -132,4 +132,14 @@ config SCHED_CORE
 	  which is the likely usage by Linux distributions, there should
 	  be no measurable impact on performance.
 
-
+config ARCH_WANTS_RT_DELAYED_SIGNALS
+	bool
+	help
+	  This option is selected by architectures where raising signals
+	  can happen in atomic contexts on PREEMPT_RT enabled kernels. This
+	  option delays raising the signal until the return to user space
+	  loop where it is also delivered. X86 requires this to deliver
+	  signals from trap handlers which run on IST stacks.
+
+config RT_DELAYED_SIGNALS
+	def_bool PREEMPT_RT && ARCH_WANTS_RT_DELAYED_SIGNALS

kernel/entry/common.c

@@ -148,6 +148,18 @@ static void handle_signal_work(struct pt_regs *regs, unsigned long ti_work)
 	arch_do_signal_or_restart(regs, ti_work & _TIF_SIGPENDING);
 }
 
+#ifdef CONFIG_RT_DELAYED_SIGNALS
+static inline void raise_delayed_signal(void)
+{
+	if (unlikely(current->forced_info.si_signo)) {
+		force_sig_info(&current->forced_info);
+		current->forced_info.si_signo = 0;
+	}
+}
+#else
+static inline void raise_delayed_signal(void) { }
+#endif
+
 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
 					    unsigned long ti_work)
 {
@@ -162,6 +174,8 @@ static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
 		if (ti_work & _TIF_NEED_RESCHED)
 			schedule();
 
+		raise_delayed_signal();
+
 		if (ti_work & _TIF_UPROBE)
 			uprobe_notify_resume(regs);
 

kernel/exit.c

@@ -845,6 +845,7 @@ void __noreturn do_exit(long code)
 		put_page(tsk->task_frag.page);
 
 	validate_creds_for_do_exit(tsk);
+	exit_task_stack_account(tsk);
 
 	check_stack_usage();
 	preempt_disable();

kernel/signal.c

@@ -1307,6 +1307,43 @@ enum sig_handler {
 	HANDLER_EXIT,	 /* Only visible as the process exit code */
 };
 
+/*
+ * On some archictectures, PREEMPT_RT has to delay sending a signal from a
+ * trap since it cannot 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 CONFIG_RT_DELAYED_SIGNALS
+static inline bool force_sig_delayed(struct kernel_siginfo *info,
+				     struct task_struct *t)
+{
+	if (!in_atomic())
+		return false;
+
+	if (WARN_ON_ONCE(t->forced_info.si_signo))
+		return true;
+
+	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 true;
+}
+#else
+static inline bool force_sig_delayed(struct kernel_siginfo *info,
+				     struct task_struct *t)
+{
+	return false;
+}
+#endif
+
 /*
  * Force a signal that the process can't ignore: if necessary
  * we unblock the signal and change any SIG_IGN to SIG_DFL.
@@ -1327,6 +1364,9 @@ force_sig_info_to_task(struct kernel_siginfo *info, struct task_struct *t,
 	struct k_sigaction *action;
 	int sig = info->si_signo;
 
+	if (force_sig_delayed(info, t))
+		return 0;
+
 	spin_lock_irqsave(&t->sighand->siglock, flags);
 	action = &t->sighand->action[sig-1];
 	ignored = action->sa.sa_handler == SIG_IGN;