Commit 9af3e04e authored by Martin Schwidefsky's avatar Martin Schwidefsky
Browse files

s390/zcrypt: get rid of ap_poll_requests 



The poll thread of the AP bus is burning CPU while waiting for
crypto requests to complete. We can as well burn a few more cycles
in the poll thread to check if there are pending requests and
remove the atomic operations with the ap_poll_requests.
This improves the code if the machine has adapter interrupts.

Signed-off-by: default avatarHarald Freudenberger <freude@linux.vnet.ibm.com>
Signed-off-by: default avatarMartin Schwidefsky <schwidefsky@de.ibm.com>
parent 0db78559

drivers/s390/crypto/ap_bus.c

+24 −10
Original line number Diff line number Diff line
@@ -91,7 +91,6 @@ static DECLARE_WORK(ap_scan_work, ap_scan_bus); 
 */

static void ap_tasklet_fn(unsigned long);

static DECLARE_TASKLET(ap_tasklet, ap_tasklet_fn, 0);

static atomic_t ap_poll_requests = ATOMIC_INIT(0);

static DECLARE_WAIT_QUEUE_HEAD(ap_poll_wait);

static struct task_struct *ap_poll_kthread = NULL;

static DEFINE_MUTEX(ap_poll_thread_mutex);
@@ -419,7 +418,6 @@ static struct ap_queue_status ap_sm_recv(struct ap_device *ap_dev) 
			 ap_dev->reply->message, ap_dev->reply->length);

	switch (status.response_code) {

	case AP_RESPONSE_NORMAL:

		atomic_dec(&ap_poll_requests);

		ap_dev->queue_count--;

		if (ap_dev->queue_count > 0)

			mod_timer(&ap_dev->timeout,
@@ -436,7 +434,6 @@ static struct ap_queue_status ap_sm_recv(struct ap_device *ap_dev) 
		if (!status.queue_empty || ap_dev->queue_count <= 0)

			break;

		/* The card shouldn't forget requests but who knows. */

		atomic_sub(ap_dev->queue_count, &ap_poll_requests);

		ap_dev->queue_count = 0;

		list_splice_init(&ap_dev->pendingq, &ap_dev->requestq);

		ap_dev->requestq_count += ap_dev->pendingq_count;
@@ -524,7 +521,6 @@ static enum ap_wait ap_sm_write(struct ap_device *ap_dev) 
			   ap_msg->message, ap_msg->length, ap_msg->special);

	switch (status.response_code) {

	case AP_RESPONSE_NORMAL:

		atomic_inc(&ap_poll_requests);

		ap_dev->queue_count++;

		if (ap_dev->queue_count == 1)

			mod_timer(&ap_dev->timeout,
@@ -796,6 +792,27 @@ static void ap_tasklet_fn(unsigned long dummy) 
	ap_sm_wait(wait);

}



static int ap_pending_requests(void)

{

	struct ap_device *ap_dev;

	int id, pending = 0;



	for (id = 0; pending == 0 && id < AP_DEVICES; id++) {

		spin_lock_bh(&ap_device_list_lock);

		list_for_each_entry(ap_dev, &ap_device_list, list) {

			spin_lock_bh(&ap_dev->lock);

			if (ap_dev->queue_count)

				pending = 1;

			spin_unlock_bh(&ap_dev->lock);

			if (pending)

				break;

		}

		spin_unlock_bh(&ap_device_list_lock);

	}



	return pending;

}



/**

 * ap_poll_thread(): Thread that polls for finished requests.

 * @data: Unused pointer
@@ -815,8 +832,7 @@ static int ap_poll_thread(void *data) 
	while (!kthread_should_stop()) {

		add_wait_queue(&ap_poll_wait, &wait);

		set_current_state(TASK_INTERRUPTIBLE);

		if (ap_suspend_flag ||

		    atomic_read(&ap_poll_requests) <= 0) {

		if (ap_suspend_flag || !ap_pending_requests()) {

			schedule();

			try_to_freeze();

		}
@@ -828,7 +844,8 @@ static int ap_poll_thread(void *data) 
			continue;

		}

		ap_tasklet_fn(0);

	} while (!kthread_should_stop());

	}



	return 0;

}
@@ -1267,9 +1284,6 @@ static int ap_device_remove(struct device *dev) 
	spin_unlock_bh(&ap_device_list_lock);

	if (ap_drv->remove)

		ap_drv->remove(ap_dev);

	spin_lock_bh(&ap_dev->lock);

	atomic_sub(ap_dev->queue_count, &ap_poll_requests);

	spin_unlock_bh(&ap_dev->lock);

	return 0;

}

CRA Git | Maintained and supported by SUSTech CRA and CCSE