Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Jens Axboe <axboe@kernel.dk>
To:     linux-kernel@vger.kernel.org, io-uring@vger.kernel.org
Cc:     peterz@infradead.org, oleg@redhat.com, tglx@linutronix.de,
        Jens Axboe <axboe@kernel.dk>,
        Roman Gershman <romger@amazon.com>
Subject: [PATCH 4/4] task_work: use TIF_NOTIFY_SIGNAL if available
Date:   Thu,  8 Oct 2020 09:27:52 -0600
Message-Id: <20201008152752.218889-5-axboe@kernel.dk>
In-Reply-To: <20201008152752.218889-1-axboe@kernel.dk>
References: <20201008152752.218889-1-axboe@kernel.dk>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

If the arch supports TIF_NOTIFY_SIGNAL, then use that for TWA_SIGNAL as
it's more efficient than using the signal delivery method. This is
especially true on threaded applications, where ->sighand is shared
across threads, but it's also lighter weight on non-shared cases.

io_uring is a heavy consumer of TWA_SIGNAL based task_work. On my test
box, even just using 16 threads shows a nice improvement running an
io_uring based echo server.

stock kernel:
0.01% <= 0.1 milliseconds
95.86% <= 0.2 milliseconds
98.27% <= 0.3 milliseconds
99.71% <= 0.4 milliseconds
100.00% <= 0.5 milliseconds
100.00% <= 0.6 milliseconds
100.00% <= 0.7 milliseconds
100.00% <= 0.8 milliseconds
100.00% <= 0.9 milliseconds
100.00% <= 1.0 milliseconds
100.00% <= 1.1 milliseconds
100.00% <= 2 milliseconds
100.00% <= 3 milliseconds
100.00% <= 3 milliseconds
1378930.00 requests per second
~1600% CPU

1.38M requests/second, and all 16 CPUs are maxed out.

patched kernel:
0.01% <= 0.1 milliseconds
98.24% <= 0.2 milliseconds
99.47% <= 0.3 milliseconds
99.99% <= 0.4 milliseconds
100.00% <= 0.5 milliseconds
100.00% <= 0.6 milliseconds
100.00% <= 0.7 milliseconds
100.00% <= 0.8 milliseconds
100.00% <= 0.9 milliseconds
100.00% <= 1.2 milliseconds
1666111.38 requests per second
~1450% CPU

1.67M requests/second, and we're no longer just hammering on the sighand
lock. The original reporter states:

"For 5.7.15 my benchmark achieves 1.6M qps and system cpu is at ~80%.
 for 5.7.16 or later it achieves only 1M qps and the system cpu is is
 at ~100%"

with the only difference there being that TWA_SIGNAL is used
unconditionally in 5.7.16, since we need it to be able to solve an
inability to run task_work if the application is waiting in the kernel
already on an event that needs task_work run to be satisfied. Also
see commit 0ba9c9edcd15.

Reported-by: Roman Gershman <romger@amazon.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 kernel/task_work.c | 41 +++++++++++++++++++++++++++++------------
 1 file changed, 29 insertions(+), 12 deletions(-)

diff --git a/kernel/task_work.c b/kernel/task_work.c
index 613b2d634af8..95604e57af46 100644
--- a/kernel/task_work.c
+++ b/kernel/task_work.c
@@ -5,6 +5,34 @@
 
 static struct callback_head work_exited; /* all we need is ->next == NULL */
 
+/*
+ * TWA_SIGNAL signaling - use TIF_NOTIFY_SIGNAL, if available, as it's faster
+ * than TIF_SIGPENDING as there's no dependency on ->sighand. The latter is
+ * shared for threads, and can cause contention on sighand->lock. Even for
+ * the non-threaded case TIF_NOTIFY_SIGNAL is more efficient, as no locking
+ * or IRQ disabling is involved for notification (or running) purposes.
+ */
+static void task_work_notify_signal(struct task_struct *task)
+{
+#ifdef TIF_NOTIFY_SIGNAL
+	set_notify_signal(task);
+#else
+	unsigned long flags;
+
+	/*
+	 * Only grab the sighand lock if we don't already have some
+	 * task_work pending. This pairs with the smp_store_mb()
+	 * in get_signal(), see comment there.
+	 */
+	if (!(READ_ONCE(task->jobctl) & JOBCTL_TASK_WORK) &&
+	    lock_task_sighand(task, &flags)) {
+		task->jobctl |= JOBCTL_TASK_WORK;
+		signal_wake_up(task, 0);
+		unlock_task_sighand(task, &flags);
+	}
+#endif
+}
+
 /**
  * task_work_add - ask the @task to execute @work->func()
  * @task: the task which should run the callback
@@ -28,7 +56,6 @@ int
 task_work_add(struct task_struct *task, struct callback_head *work, int notify)
 {
 	struct callback_head *head;
-	unsigned long flags;
 
 	do {
 		head = READ_ONCE(task->task_works);
@@ -42,17 +69,7 @@ task_work_add(struct task_struct *task, struct callback_head *work, int notify)
 		set_notify_resume(task);
 		break;
 	case TWA_SIGNAL:
-		/*
-		 * Only grab the sighand lock if we don't already have some
-		 * task_work pending. This pairs with the smp_store_mb()
-		 * in get_signal(), see comment there.
-		 */
-		if (!(READ_ONCE(task->jobctl) & JOBCTL_TASK_WORK) &&
-		    lock_task_sighand(task, &flags)) {
-			task->jobctl |= JOBCTL_TASK_WORK;
-			signal_wake_up(task, 0);
-			unlock_task_sighand(task, &flags);
-		}
+		task_work_notify_signal(task);
 		break;
 	}
 
-- 
2.28.0