[S390] zcrypt: fix request timeout handling
Under very high load zcrypt requests may timeout while waiting on the
request queue. Modify zcrypt that timeouts are based on crypto adapter
responses. A timeout occurs only if a crypto adapter does not respond
within a given time frame to sumitted requests.
Signed-off-by: Ralph Wuerthner <rwuerthn@de.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
diff --git a/drivers/s390/crypto/ap_bus.c b/drivers/s390/crypto/ap_bus.c
index 5aac0ec..90bd220 100644
--- a/drivers/s390/crypto/ap_bus.c
+++ b/drivers/s390/crypto/ap_bus.c
@@ -43,6 +43,7 @@
static void ap_poll_timeout(unsigned long);
static int ap_poll_thread_start(void);
static void ap_poll_thread_stop(void);
+static void ap_request_timeout(unsigned long);
/**
* Module description.
@@ -189,6 +190,7 @@
case AP_RESPONSE_NORMAL:
return 0;
case AP_RESPONSE_Q_FULL:
+ case AP_RESPONSE_RESET_IN_PROGRESS:
return -EBUSY;
default: /* Device is gone. */
return -ENODEV;
@@ -252,6 +254,8 @@
if (status.queue_empty)
return -ENOENT;
return -EBUSY;
+ case AP_RESPONSE_RESET_IN_PROGRESS:
+ return -EBUSY;
default:
return -ENODEV;
}
@@ -326,11 +330,12 @@
i = AP_MAX_RESET; /* return with -ENODEV */
break;
case AP_RESPONSE_RESET_IN_PROGRESS:
+ rc = -EBUSY;
case AP_RESPONSE_BUSY:
default:
break;
}
- if (rc != -ENODEV)
+ if (rc != -ENODEV && rc != -EBUSY)
break;
if (i < AP_MAX_RESET - 1) {
udelay(5);
@@ -341,6 +346,40 @@
}
/**
+ * Arm request timeout if a AP device was idle and a new request is submitted.
+ */
+static void ap_increase_queue_count(struct ap_device *ap_dev)
+{
+ int timeout = ap_dev->drv->request_timeout;
+
+ ap_dev->queue_count++;
+ if (ap_dev->queue_count == 1) {
+ mod_timer(&ap_dev->timeout, jiffies + timeout);
+ ap_dev->reset = AP_RESET_ARMED;
+ }
+}
+
+/**
+ * AP device is still alive, re-schedule request timeout if there are still
+ * pending requests.
+ */
+static void ap_decrease_queue_count(struct ap_device *ap_dev)
+{
+ int timeout = ap_dev->drv->request_timeout;
+
+ ap_dev->queue_count--;
+ if (ap_dev->queue_count > 0)
+ mod_timer(&ap_dev->timeout, jiffies + timeout);
+ else
+ /**
+ * The timeout timer should to be disabled now - since
+ * del_timer_sync() is very expensive, we just tell via the
+ * reset flag to ignore the pending timeout timer.
+ */
+ ap_dev->reset = AP_RESET_IGNORE;
+}
+
+/**
* AP device related attributes.
*/
static ssize_t ap_hwtype_show(struct device *dev,
@@ -498,6 +537,7 @@
struct ap_driver *ap_drv = ap_dev->drv;
ap_flush_queue(ap_dev);
+ del_timer_sync(&ap_dev->timeout);
if (ap_drv->remove)
ap_drv->remove(ap_dev);
spin_lock_bh(&ap_device_lock);
@@ -759,17 +799,21 @@
__ap_scan_bus);
rc = ap_query_queue(qid, &queue_depth, &device_type);
if (dev) {
+ if (rc == -EBUSY) {
+ set_current_state(TASK_UNINTERRUPTIBLE);
+ schedule_timeout(AP_RESET_TIMEOUT);
+ rc = ap_query_queue(qid, &queue_depth,
+ &device_type);
+ }
ap_dev = to_ap_dev(dev);
spin_lock_bh(&ap_dev->lock);
if (rc || ap_dev->unregistered) {
spin_unlock_bh(&ap_dev->lock);
- put_device(dev);
device_unregister(dev);
+ put_device(dev);
continue;
- } else
- spin_unlock_bh(&ap_dev->lock);
- }
- if (dev) {
+ }
+ spin_unlock_bh(&ap_dev->lock);
put_device(dev);
continue;
}
@@ -788,6 +832,8 @@
INIT_LIST_HEAD(&ap_dev->pendingq);
INIT_LIST_HEAD(&ap_dev->requestq);
INIT_LIST_HEAD(&ap_dev->list);
+ setup_timer(&ap_dev->timeout, ap_request_timeout,
+ (unsigned long) ap_dev);
if (device_type == 0)
ap_probe_device_type(ap_dev);
else
@@ -853,7 +899,7 @@
switch (status.response_code) {
case AP_RESPONSE_NORMAL:
atomic_dec(&ap_poll_requests);
- ap_dev->queue_count--;
+ ap_decrease_queue_count(ap_dev);
list_for_each_entry(ap_msg, &ap_dev->pendingq, list) {
if (ap_msg->psmid != ap_dev->reply->psmid)
continue;
@@ -904,7 +950,7 @@
switch (status.response_code) {
case AP_RESPONSE_NORMAL:
atomic_inc(&ap_poll_requests);
- ap_dev->queue_count++;
+ ap_increase_queue_count(ap_dev);
list_move_tail(&ap_msg->list, &ap_dev->pendingq);
ap_dev->requestq_count--;
ap_dev->pendingq_count++;
@@ -914,6 +960,7 @@
*flags |= 2;
break;
case AP_RESPONSE_Q_FULL:
+ case AP_RESPONSE_RESET_IN_PROGRESS:
*flags |= 2;
break;
case AP_RESPONSE_MESSAGE_TOO_BIG:
@@ -960,10 +1007,11 @@
list_add_tail(&ap_msg->list, &ap_dev->pendingq);
atomic_inc(&ap_poll_requests);
ap_dev->pendingq_count++;
- ap_dev->queue_count++;
+ ap_increase_queue_count(ap_dev);
ap_dev->total_request_count++;
break;
case AP_RESPONSE_Q_FULL:
+ case AP_RESPONSE_RESET_IN_PROGRESS:
list_add_tail(&ap_msg->list, &ap_dev->requestq);
ap_dev->requestq_count++;
ap_dev->total_request_count++;
@@ -1046,6 +1094,25 @@
}
/**
+ * Reset a not responding AP device and move all requests from the
+ * pending queue to the request queue.
+ */
+static void ap_reset(struct ap_device *ap_dev)
+{
+ int rc;
+
+ ap_dev->reset = AP_RESET_IGNORE;
+ 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;
+ ap_dev->pendingq_count = 0;
+ rc = ap_init_queue(ap_dev->qid);
+ if (rc == -ENODEV)
+ ap_dev->unregistered = 1;
+}
+
+/**
* Poll all AP devices on the bus in a round robin fashion. Continue
* polling until bit 2^0 of the control flags is not set. If bit 2^1
* of the control flags has been set arm the poll timer.
@@ -1056,6 +1123,8 @@
if (!ap_dev->unregistered) {
if (ap_poll_queue(ap_dev, flags))
ap_dev->unregistered = 1;
+ if (ap_dev->reset == AP_RESET_DO)
+ ap_reset(ap_dev);
}
spin_unlock(&ap_dev->lock);
return 0;
@@ -1147,6 +1216,17 @@
mutex_unlock(&ap_poll_thread_mutex);
}
+/**
+ * Handling of request timeouts
+ */
+static void ap_request_timeout(unsigned long data)
+{
+ struct ap_device *ap_dev = (struct ap_device *) data;
+
+ if (ap_dev->reset == AP_RESET_ARMED)
+ ap_dev->reset = AP_RESET_DO;
+}
+
static void ap_reset_domain(void)
{
int i;