scsi: iscsi: Remove iscsi_get_task back_lock requirement

We currently require that the back_lock is held when calling the functions
that manipulate the iscsi_task refcount. The only reason for this is to
handle races where we are handling SCSI-ml EH callbacks and the cmd is
completing at the same time the normal completion path is running, and we
can't return from the EH callback until the driver has stopped accessing
the cmd. Holding the back_lock while also accessing the task->state made it
simple to check that a cmd is completing and also get/put a refcount at the
same time, and at the time we were not as concerned about performance.

The problem is that we don't want to take the back_lock from the xmit path
for normal I/O since it causes contention with the completion path if the
user has chosen to try and split those paths on different CPUs (in this
case abusing the CPUs and ignoring caching improves perf for some uses).

Begins to remove the back_lock requirement for iscsi_get/put_task by
removing the requirement for the get path. Instead of always holding the
back_lock we detect if something has done the last put and is about to call
iscsi_free_task(). A subsequent commit will then allow iSCSI code to do the
last put on a task and only grab the back_lock if the refcount is now zero
and it's going to call iscsi_free_task().

Link: https://lore.kernel.org/r/20220616224557.115234-8-michael.christie@oracle.com
Reviewed-by: Lee Duncan <lduncan@suse.com>
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

1 files changed, 66 insertions, 29 deletions

diff --git a/drivers/scsi/libiscsi.c b/drivers/scsi/libiscsi.c
index fce9f9e5b00b..3a6b827496dd 100644
--- a/drivers/scsi/libiscsi.c
+++ b/drivers/scsi/libiscsi.c
@@ -83,6 +83,8 @@ MODULE_PARM_DESC(debug_libiscsi_eh,
 				"%s " dbg_fmt, __func__, ##arg);	\
 	} while (0);
 
+#define ISCSI_CMD_COMPL_WAIT 5
+
 inline void iscsi_conn_queue_xmit(struct iscsi_conn *conn)
 {
 	struct Scsi_Host *shost = conn->session->host;
@@ -482,11 +484,11 @@ static void iscsi_free_task(struct iscsi_task *task)
 	}
 }
 
-void __iscsi_get_task(struct iscsi_task *task)
+bool iscsi_get_task(struct iscsi_task *task)
 {
-	refcount_inc(&task->refcount);
+	return refcount_inc_not_zero(&task->refcount);
 }
-EXPORT_SYMBOL_GPL(__iscsi_get_task);
+EXPORT_SYMBOL_GPL(iscsi_get_task);
 
 void __iscsi_put_task(struct iscsi_task *task)
 {
@@ -600,20 +602,17 @@ static bool cleanup_queued_task(struct iscsi_task *task)
 }
 
 /*
- * session frwd lock must be held and if not called for a task that is still
- * pending or from the xmit thread, then xmit thread must be suspended
+ * session back and frwd lock must be held and if not called for a task that
+ * is still pending or from the xmit thread, then xmit thread must be suspended
  */
-static void fail_scsi_task(struct iscsi_task *task, int err)
+static void __fail_scsi_task(struct iscsi_task *task, int err)
 {
 	struct iscsi_conn *conn = task->conn;
 	struct scsi_cmnd *sc;
 	int state;
 
-	spin_lock_bh(&conn->session->back_lock);
-	if (cleanup_queued_task(task)) {
-		spin_unlock_bh(&conn->session->back_lock);
+	if (cleanup_queued_task(task))
 		return;
-	}
 
 	if (task->state == ISCSI_TASK_PENDING) {
 		/*
@@ -632,7 +631,15 @@ static void fail_scsi_task(struct iscsi_task *task, int err)
 	sc->result = err << 16;
 	scsi_set_resid(sc, scsi_bufflen(sc));
 	iscsi_complete_task(task, state);
-	spin_unlock_bh(&conn->session->back_lock);
+}
+
+static void fail_scsi_task(struct iscsi_task *task, int err)
+{
+	struct iscsi_session *session = task->conn->session;
+
+	spin_lock_bh(&session->back_lock);
+	__fail_scsi_task(task, err);
+	spin_unlock_bh(&session->back_lock);
 }
 
 static int iscsi_prep_mgmt_task(struct iscsi_conn *conn,
@@ -1450,8 +1457,17 @@ static int iscsi_xmit_task(struct iscsi_conn *conn, struct iscsi_task *task,
 	spin_lock_bh(&conn->session->back_lock);
 
 	if (!conn->task) {
-		/* Take a ref so we can access it after xmit_task() */
-		__iscsi_get_task(task);
+		/*
+		 * Take a ref so we can access it after xmit_task().
+		 *
+		 * This should never fail because the failure paths will have
+		 * stopped the xmit thread.
+		 */
+		if (!iscsi_get_task(task)) {
+			spin_unlock_bh(&conn->session->back_lock);
+			WARN_ON_ONCE(1);
+			return 0;
+		}
 	} else {
 		/* Already have a ref from when we failed to send it last call */
 		conn->task = NULL;
@@ -1493,7 +1509,7 @@ static int iscsi_xmit_task(struct iscsi_conn *conn, struct iscsi_task *task,
 		 * get an extra ref that is released next time we access it
 		 * as conn->task above.
 		 */
-		__iscsi_get_task(task);
+		iscsi_get_task(task);
 		conn->task = task;
 	}
 
@@ -1912,6 +1928,7 @@ static void fail_scsi_tasks(struct iscsi_conn *conn, u64 lun, int error)
 	struct iscsi_task *task;
 	int i;
 
+restart_cmd_loop:
 	spin_lock_bh(&session->back_lock);
 	for (i = 0; i < session->cmds_max; i++) {
 		task = session->cmds[i];
@@ -1920,22 +1937,25 @@ static void fail_scsi_tasks(struct iscsi_conn *conn, u64 lun, int error)
 
 		if (lun != -1 && lun != task->sc->device->lun)
 			continue;
-
-		__iscsi_get_task(task);
-		spin_unlock_bh(&session->back_lock);
+		/*
+		 * The cmd is completing but if this is called from an eh
+		 * callout path then when we return scsi-ml owns the cmd. Wait
+		 * for the completion path to finish freeing the cmd.
+		 */
+		if (!iscsi_get_task(task)) {
+			spin_unlock_bh(&session->back_lock);
+			spin_unlock_bh(&session->frwd_lock);
+			udelay(ISCSI_CMD_COMPL_WAIT);
+			spin_lock_bh(&session->frwd_lock);
+			goto restart_cmd_loop;
+		}
 
 		ISCSI_DBG_SESSION(session,
 				  "failing sc %p itt 0x%x state %d\n",
 				  task->sc, task->itt, task->state);
-		fail_scsi_task(task, error);
-
-		spin_unlock_bh(&session->frwd_lock);
-		iscsi_put_task(task);
-		spin_lock_bh(&session->frwd_lock);
-
-		spin_lock_bh(&session->back_lock);
+		__fail_scsi_task(task, error);
+		__iscsi_put_task(task);
 	}
-
 	spin_unlock_bh(&session->back_lock);
 }
 
@@ -2040,7 +2060,16 @@ enum blk_eh_timer_return iscsi_eh_cmd_timed_out(struct scsi_cmnd *sc)
 		spin_unlock(&session->back_lock);
 		goto done;
 	}
-	__iscsi_get_task(task);
+	if (!iscsi_get_task(task)) {
+		/*
+		 * Racing with the completion path right now, so give it more
+		 * time so that path can complete it like normal.
+		 */
+		rc = BLK_EH_RESET_TIMER;
+		task = NULL;
+		spin_unlock(&session->back_lock);
+		goto done;
+	}
 	spin_unlock(&session->back_lock);
 
 	if (session->state != ISCSI_STATE_LOGGED_IN) {
@@ -2289,6 +2318,7 @@ int iscsi_eh_abort(struct scsi_cmnd *sc)
 
 	ISCSI_DBG_EH(session, "aborting sc %p\n", sc);
 
+completion_check:
 	mutex_lock(&session->eh_mutex);
 	spin_lock_bh(&session->frwd_lock);
 	/*
@@ -2328,13 +2358,20 @@ int iscsi_eh_abort(struct scsi_cmnd *sc)
 		return SUCCESS;
 	}
 
+	if (!iscsi_get_task(task)) {
+		spin_unlock(&session->back_lock);
+		spin_unlock_bh(&session->frwd_lock);
+		mutex_unlock(&session->eh_mutex);
+		/* We are just about to call iscsi_free_task so wait for it. */
+		udelay(ISCSI_CMD_COMPL_WAIT);
+		goto completion_check;
+	}
+
+	ISCSI_DBG_EH(session, "aborting [sc %p itt 0x%x]\n", sc, task->itt);
 	conn = session->leadconn;
 	iscsi_get_conn(conn->cls_conn);
 	conn->eh_abort_cnt++;
 	age = session->age;
-
-	ISCSI_DBG_EH(session, "aborting [sc %p itt 0x%x]\n", sc, task->itt);
-	__iscsi_get_task(task);
 	spin_unlock(&session->back_lock);
 
 	if (task->state == ISCSI_TASK_PENDING) {