[PATCH] exit: combine work under lock in synchronize_group_exit() and coredump_task_exit()

[PATCH] exit: combine work under lock in synchronize_group_exit() and coredump_task_exit()

[Mateusz Guzik]

This reduces single-threaded overhead as it avoids one lock+irq trip on
exit.

It also improves scalability of spawning and killing threads within one
process (just shy of 5% when doing it on 24 cores on my test jig).

This happens to test to for coredumping prior to handling kcov and
kmsan, which afaics is harmless.

I however was not comfortable lifting dumping prior to it, so it still
happens after.

Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
---
 kernel/exit.c | 67 +++++++++++++++++++++++++++++----------------------
 1 file changed, 38 insertions(+), 29 deletions(-)

diff --git a/kernel/exit.c b/kernel/exit.c
index f97a2bbc9db9..f799c32dd8f5 100644
--- a/kernel/exit.c
+++ b/kernel/exit.c
@@ -412,9 +412,9 @@ kill_orphaned_pgrp(struct task_struct *tsk, struct task_struct *parent)
 	}
 }
 
-static void coredump_task_exit(struct task_struct *tsk)
+static struct core_state *coredump_task_exit_prep(struct task_struct *tsk)
 {
-	struct core_state *core_state;
+	lockdep_assert_held(&tsk->sighand->siglock);
 
 	/*
 	 * Serialize with any possible pending coredump.
@@ -423,33 +423,37 @@ static void coredump_task_exit(struct task_struct *tsk)
 	 * will increment ->nr_threads for each thread in the
 	 * group without PF_POSTCOREDUMP set.
 	 */
-	spin_lock_irq(&tsk->sighand->siglock);
 	tsk->flags |= PF_POSTCOREDUMP;
-	core_state = tsk->signal->core_state;
-	spin_unlock_irq(&tsk->sighand->siglock);
-	if (core_state) {
-		struct core_thread self;
-
-		self.task = current;
-		if (self.task->flags & PF_SIGNALED)
-			self.next = xchg(&core_state->dumper.next, &self);
-		else
-			self.task = NULL;
-		/*
-		 * Implies mb(), the result of xchg() must be visible
-		 * to core_state->dumper.
-		 */
-		if (atomic_dec_and_test(&core_state->nr_threads))
-			complete(&core_state->startup);
+	return tsk->signal->core_state;
+}
 
-		for (;;) {
-			set_current_state(TASK_IDLE|TASK_FREEZABLE);
-			if (!self.task) /* see coredump_finish() */
-				break;
-			schedule();
-		}
-		__set_current_state(TASK_RUNNING);
+static void coredump_task_exit_finish(struct task_struct *tsk,
+				      struct core_state *core_state)
+{
+	struct core_thread self;
+
+	if (likely(!core_state))
+		return;
+
+	self.task = current;
+	if (self.task->flags & PF_SIGNALED)
+		self.next = xchg(&core_state->dumper.next, &self);
+	else
+		self.task = NULL;
+	/*
+	 * Implies mb(), the result of xchg() must be visible
+	 * to core_state->dumper.
+	 */
+	if (atomic_dec_and_test(&core_state->nr_threads))
+		complete(&core_state->startup);
+
+	for (;;) {
+		set_current_state(TASK_IDLE|TASK_FREEZABLE);
+		if (!self.task) /* see coredump_finish() */
+			break;
+		schedule();
 	}
+	__set_current_state(TASK_RUNNING);
 }
 
 #ifdef CONFIG_MEMCG
@@ -878,7 +882,8 @@ static void synchronize_group_exit(struct task_struct *tsk, long code)
 	struct sighand_struct *sighand = tsk->sighand;
 	struct signal_struct *signal = tsk->signal;
 
-	spin_lock_irq(&sighand->siglock);
+	lockdep_assert_held(&sighand->siglock);
+
 	signal->quick_threads--;
 	if ((signal->quick_threads == 0) &&
 	    !(signal->flags & SIGNAL_GROUP_EXIT)) {
@@ -886,24 +891,28 @@ static void synchronize_group_exit(struct task_struct *tsk, long code)
 		signal->group_exit_code = code;
 		signal->group_stop_count = 0;
 	}
-	spin_unlock_irq(&sighand->siglock);
 }
 
 void __noreturn do_exit(long code)
 {
 	struct task_struct *tsk = current;
+	struct sighand_struct *sighand = tsk->sighand;
+	struct core_state *core_state;
 	int group_dead;
 
 	WARN_ON(irqs_disabled());
 
+	spin_lock_irq(&sighand->siglock);
 	synchronize_group_exit(tsk, code);
+	core_state = coredump_task_exit_prep(tsk);
+	spin_unlock_irq(&sighand->siglock);
 
 	WARN_ON(tsk->plug);
 
 	kcov_task_exit(tsk);
 	kmsan_task_exit(tsk);
 
-	coredump_task_exit(tsk);
+	coredump_task_exit_finish(tsk, core_state);
 	ptrace_event(PTRACE_EVENT_EXIT, code);
 	user_events_exit(tsk);
 
-- 
2.43.0

Re: [PATCH] exit: combine work under lock in synchronize_group_exit() and coredump_task_exit()

[Oleg Nesterov]

On 03/19, Mateusz Guzik wrote:
>
> This happens to test to for coredumping prior to handling kcov and
> kmsan, which afaics is harmless.
>
> I however was not comfortable lifting dumping prior to it, so it still
> happens after.

Well, I'd really prefer to move synchronize_group_exit(tsk, code) down,
after kmsan_task_exit() for the start.

Then the patch and resulting code will be simpler, coredump_task_exit()
can be even folded into synchronize_group_exit().

Oleg.