[SCSI] bnx2i: Added the use of kthreads to handle SCSI cmd completion

This patch breaks the SCSI cmd completion into two parts:
1. The bh will allocate and queued work to the cmd specific CPU IO
completion kthread.  The CPU for the cmd is from the sc->request->cpu.

2. The CPU specific IO completion kthread will call the scsi_cmd_resp
routine to do the actual cmd completion.

In the normal case, these IO completion kthreads should complete before
the blk IO times out at 60s.  However, in the case when these kthreads
are blocked for whatever reason and exceeded the timeout, the call
to conn_destroy will have to iterate and exhaust all related work in the
percpu work list for all online CPUs.  This will guarantee the protection
of the work->session and conn pointers before they get freed.

Also modified the event coalescing formula to have at least the
event_coal_min outstanding cmds in the pipeline so the SCSI producer
would not get underrun.

Also changed the following SCSI parameters:
- can_queue from 1024 to 2048
- cmds_per_lun from 24 to 128

Signed-off-by: Eddie Wai <eddie.wai@broadcom.com>
Acked-by: Benjamin Li <benli@broadcom.com>
Acked-by: Michael Chan <mchan@broadcom.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>

4 files changed, 333 insertions, 35 deletions

diff --git a/drivers/scsi/bnx2i/bnx2i.h b/drivers/scsi/bnx2i/bnx2i.h
index 6bdd25a..239bc4e 100644
--- a/drivers/scsi/bnx2i/bnx2i.h
+++ b/drivers/scsi/bnx2i/bnx2i.h
@@ -22,11 +22,14 @@
 #include <linux/pci.h>
 #include <linux/spinlock.h>
 #include <linux/interrupt.h>
+#include <linux/delay.h>
 #include <linux/sched.h>
 #include <linux/in.h>
 #include <linux/kfifo.h>
 #include <linux/netdevice.h>
 #include <linux/completion.h>
+#include <linux/kthread.h>
+#include <linux/cpu.h>
 
 #include <scsi/scsi_cmnd.h>
 #include <scsi/scsi_device.h>
@@ -202,10 +205,13 @@ struct io_bdt {
 /**
  * bnx2i_cmd - iscsi command structure
  *
+ * @hdr:                iSCSI header
+ * @conn:               iscsi_conn pointer
  * @scsi_cmd:           SCSI-ML task pointer corresponding to this iscsi cmd
  * @sg:                 SG list
  * @io_tbl:             buffer descriptor (BD) table
  * @bd_tbl_dma:         buffer descriptor (BD) table's dma address
+ * @req:                bnx2i specific command request struct
  */
 struct bnx2i_cmd {
 	struct iscsi_hdr hdr;
@@ -229,6 +235,7 @@ struct bnx2i_cmd {
  * @gen_pdu:               login/nopout/logout pdu resources
  * @violation_notified:    bit mask used to track iscsi error/warning messages
  *                         already printed out
+ * @work_cnt:              keeps track of the number of outstanding work
  *
  * iSCSI connection structure
  */
@@ -252,6 +259,8 @@ struct bnx2i_conn {
 	 */
 	struct generic_pdu_resc gen_pdu;
 	u64 violation_notified;
+
+	atomic_t work_cnt;
 };
 
 
@@ -661,7 +670,6 @@ enum {
  * @hba:                adapter to which this connection belongs
  * @conn:               iscsi connection this EP is linked to
  * @cls_ep:             associated iSCSI endpoint pointer
- * @sess:               iscsi session this EP is linked to
  * @cm_sk:              cnic sock struct
  * @hba_age:            age to detect if 'iscsid' issues ep_disconnect()
  *                      after HBA reset is completed by bnx2i/cnic/bnx2
@@ -687,7 +695,7 @@ struct bnx2i_endpoint {
 	u32 hba_age;
 	u32 state;
 	unsigned long timestamp;
-	int num_active_cmds;
+	atomic_t num_active_cmds;
 	u32 ec_shift;
 
 	struct qp_info qp;
@@ -700,6 +708,19 @@ struct bnx2i_endpoint {
 };
 
 
+struct bnx2i_work {
+	struct list_head list;
+	struct iscsi_session *session;
+	struct bnx2i_conn *bnx2i_conn;
+	struct cqe cqe;
+};
+
+struct bnx2i_percpu_s {
+	struct task_struct *iothread;
+	struct list_head work_list;
+	spinlock_t p_work_lock;
+};
+
 
 /* Global variables */
 extern unsigned int error_mask1, error_mask2;
@@ -783,7 +804,7 @@ extern struct bnx2i_endpoint *bnx2i_find_ep_in_destroy_list(
 		struct bnx2i_hba *hba, u32 iscsi_cid);
 
 extern int bnx2i_map_ep_dbell_regs(struct bnx2i_endpoint *ep);
-extern void bnx2i_arm_cq_event_coalescing(struct bnx2i_endpoint *ep, u8 action);
+extern int bnx2i_arm_cq_event_coalescing(struct bnx2i_endpoint *ep, u8 action);
 
 extern int bnx2i_hw_ep_disconnect(struct bnx2i_endpoint *bnx2i_ep);
 
@@ -793,4 +814,8 @@ extern void bnx2i_print_active_cmd_queue(struct bnx2i_conn *conn);
 extern void bnx2i_print_xmit_pdu_queue(struct bnx2i_conn *conn);
 extern void bnx2i_print_recv_state(struct bnx2i_conn *conn);
 
+extern int bnx2i_percpu_io_thread(void *arg);
+extern int bnx2i_process_scsi_cmd_resp(struct iscsi_session *session,
+				       struct bnx2i_conn *bnx2i_conn,
+				       struct cqe *cqe);
 #endif
diff --git a/drivers/scsi/bnx2i/bnx2i_hwi.c b/drivers/scsi/bnx2i/bnx2i_hwi.c
index 550e6c4..a501a72 100644
--- a/drivers/scsi/bnx2i/bnx2i_hwi.c
+++ b/drivers/scsi/bnx2i/bnx2i_hwi.c
@@ -17,6 +17,8 @@
 #include <scsi/libiscsi.h>
 #include "bnx2i.h"
 
+DECLARE_PER_CPU(struct bnx2i_percpu_s, bnx2i_percpu);
+
 /**
  * bnx2i_get_cid_num - get cid from ep
  * @ep: 	endpoint pointer
@@ -131,16 +133,16 @@ static void bnx2i_iscsi_license_error(struct bnx2i_hba *hba, u32 error_code)
  *	the driver. EQ event is generated CQ index is hit or at least 1 CQ is
  *	outstanding and on chip timer expires
  */
-void bnx2i_arm_cq_event_coalescing(struct bnx2i_endpoint *ep, u8 action)
+int bnx2i_arm_cq_event_coalescing(struct bnx2i_endpoint *ep, u8 action)
 {
 	struct bnx2i_5771x_cq_db *cq_db;
 	u16 cq_index;
-	u16 next_index;
+	u16 next_index = 0;
 	u32 num_active_cmds;
 
 	/* Coalesce CQ entries only on 10G devices */
 	if (!test_bit(BNX2I_NX2_DEV_57710, &ep->hba->cnic_dev_type))
-		return;
+		return 0;
 
 	/* Do not update CQ DB multiple times before firmware writes
 	 * '0xFFFF' to CQDB->SQN field. Deviation may cause spurious
@@ -150,16 +152,17 @@ void bnx2i_arm_cq_event_coalescing(struct bnx2i_endpoint *ep, u8 action)
 
 	if (action != CNIC_ARM_CQE_FP)
 		if (cq_db->sqn[0] && cq_db->sqn[0] != 0xFFFF)
-			return;
+			return 0;
 
 	if (action == CNIC_ARM_CQE || action == CNIC_ARM_CQE_FP) {
-		num_active_cmds = ep->num_active_cmds;
+		num_active_cmds = atomic_read(&ep->num_active_cmds);
 		if (num_active_cmds <= event_coal_min)
 			next_index = 1;
-		else
-			next_index = event_coal_min +
-				     ((num_active_cmds - event_coal_min) >>
-				     ep->ec_shift);
+		else {
+			next_index = num_active_cmds >> ep->ec_shift;
+			if (next_index > num_active_cmds - event_coal_min)
+				next_index = num_active_cmds - event_coal_min;
+		}
 		if (!next_index)
 			next_index = 1;
 		cq_index = ep->qp.cqe_exp_seq_sn + next_index - 1;
@@ -170,6 +173,7 @@ void bnx2i_arm_cq_event_coalescing(struct bnx2i_endpoint *ep, u8 action)
 
 		cq_db->sqn[0] = cq_index;
 	}
+	return next_index;
 }
 
 
@@ -265,7 +269,7 @@ static void bnx2i_ring_sq_dbell(struct bnx2i_conn *bnx2i_conn, int count)
 	struct bnx2i_5771x_sq_rq_db *sq_db;
 	struct bnx2i_endpoint *ep = bnx2i_conn->ep;
 
-	ep->num_active_cmds++;
+	atomic_inc(&ep->num_active_cmds);
 	wmb();	/* flush SQ WQE memory before the doorbell is rung */
 	if (test_bit(BNX2I_NX2_DEV_57710, &ep->hba->cnic_dev_type)) {
 		sq_db = (struct bnx2i_5771x_sq_rq_db *) ep->qp.sq_pgtbl_virt;
@@ -1331,14 +1335,15 @@ int bnx2i_send_fw_iscsi_init_msg(struct bnx2i_hba *hba)
 
 /**
  * bnx2i_process_scsi_cmd_resp - this function handles scsi cmd completion.
- * @conn:	iscsi connection
+ * @session:	iscsi session
+ * @bnx2i_conn:	bnx2i connection
  * @cqe:	pointer to newly DMA'ed CQE entry for processing
  *
  * process SCSI CMD Response CQE & complete the request to SCSI-ML
  */
-static int bnx2i_process_scsi_cmd_resp(struct iscsi_session *session,
-				       struct bnx2i_conn *bnx2i_conn,
-				       struct cqe *cqe)
+int bnx2i_process_scsi_cmd_resp(struct iscsi_session *session,
+				struct bnx2i_conn *bnx2i_conn,
+				struct cqe *cqe)
 {
 	struct iscsi_conn *conn = bnx2i_conn->cls_conn->dd_data;
 	struct bnx2i_cmd_response *resp_cqe;
@@ -1348,7 +1353,7 @@ static int bnx2i_process_scsi_cmd_resp(struct iscsi_session *session,
 	u32 datalen = 0;
 
 	resp_cqe = (struct bnx2i_cmd_response *)cqe;
-	spin_lock(&session->lock);
+	spin_lock_bh(&session->lock);
 	task = iscsi_itt_to_task(conn,
 				 resp_cqe->itt & ISCSI_CMD_RESPONSE_INDEX);
 	if (!task)
@@ -1409,7 +1414,7 @@ done:
 	__iscsi_complete_pdu(conn, (struct iscsi_hdr *)hdr,
 			     conn->data, datalen);
 fail:
-	spin_unlock(&session->lock);
+	spin_unlock_bh(&session->lock);
 	return 0;
 }
 
@@ -1836,21 +1841,130 @@ static void bnx2i_process_cmd_cleanup_resp(struct iscsi_session *session,
 }
 
 
+/**
+ * bnx2i_percpu_io_thread - thread per cpu for ios
+ *
+ * @arg:	ptr to bnx2i_percpu_info structure
+ */
+int bnx2i_percpu_io_thread(void *arg)
+{
+	struct bnx2i_percpu_s *p = arg;
+	struct bnx2i_work *work, *tmp;
+	LIST_HEAD(work_list);
+
+	set_user_nice(current, -20);
+
+	while (!kthread_should_stop()) {
+		spin_lock_bh(&p->p_work_lock);
+		while (!list_empty(&p->work_list)) {
+			list_splice_init(&p->work_list, &work_list);
+			spin_unlock_bh(&p->p_work_lock);
+
+			list_for_each_entry_safe(work, tmp, &work_list, list) {
+				list_del_init(&work->list);
+				/* work allocated in the bh, freed here */
+				bnx2i_process_scsi_cmd_resp(work->session,
+							    work->bnx2i_conn,
+							    &work->cqe);
+				atomic_dec(&work->bnx2i_conn->work_cnt);
+				kfree(work);
+			}
+			spin_lock_bh(&p->p_work_lock);
+		}
+		set_current_state(TASK_INTERRUPTIBLE);
+		spin_unlock_bh(&p->p_work_lock);
+		schedule();
+	}
+	__set_current_state(TASK_RUNNING);
+
+	return 0;
+}
+
+
+/**
+ * bnx2i_queue_scsi_cmd_resp - queue cmd completion to the percpu thread
+ * @bnx2i_conn:		bnx2i connection
+ *
+ * this function is called by generic KCQ handler to queue all pending cmd
+ * completion CQEs
+ *
+ * The implementation is to queue the cmd response based on the
+ * last recorded command for the given connection.  The
+ * cpu_id gets recorded upon task_xmit.  No out-of-order completion!
+ */
+static int bnx2i_queue_scsi_cmd_resp(struct iscsi_session *session,
+				     struct bnx2i_conn *bnx2i_conn,
+				     struct bnx2i_nop_in_msg *cqe)
+{
+	struct bnx2i_work *bnx2i_work = NULL;
+	struct bnx2i_percpu_s *p = NULL;
+	struct iscsi_task *task;
+	struct scsi_cmnd *sc;
+	int rc = 0;
+
+	spin_lock(&session->lock);
+	task = iscsi_itt_to_task(bnx2i_conn->cls_conn->dd_data,
+				 cqe->itt & ISCSI_CMD_RESPONSE_INDEX);
+	if (!task) {
+		spin_unlock(&session->lock);
+		return -EINVAL;
+	}
+	sc = task->sc;
+	spin_unlock(&session->lock);
+
+	p = &per_cpu(bnx2i_percpu, sc->request->cpu);
+	spin_lock(&p->p_work_lock);
+	if (unlikely(!p->iothread)) {
+		rc = -EINVAL;
+		goto err;
+	}
+	/* Alloc and copy to the cqe */
+	bnx2i_work = kzalloc(sizeof(struct bnx2i_work), GFP_ATOMIC);
+	if (bnx2i_work) {
+		INIT_LIST_HEAD(&bnx2i_work->list);
+		bnx2i_work->session = session;
+		bnx2i_work->bnx2i_conn = bnx2i_conn;
+		memcpy(&bnx2i_work->cqe, cqe, sizeof(struct cqe));
+		list_add_tail(&bnx2i_work->list, &p->work_list);
+		atomic_inc(&bnx2i_conn->work_cnt);
+		wake_up_process(p->iothread);
+		spin_unlock(&p->p_work_lock);
+		goto done;
+	} else
+		rc = -ENOMEM;
+err:
+	spin_unlock(&p->p_work_lock);
+	bnx2i_process_scsi_cmd_resp(session, bnx2i_conn, (struct cqe *)cqe);
+done:
+	return rc;
+}
+
 
 /**
  * bnx2i_process_new_cqes - process newly DMA'ed CQE's
- * @bnx2i_conn:		iscsi connection
+ * @bnx2i_conn:		bnx2i connection
  *
  * this function is called by generic KCQ handler to process all pending CQE's
  */
-static void bnx2i_process_new_cqes(struct bnx2i_conn *bnx2i_conn)
+static int bnx2i_process_new_cqes(struct bnx2i_conn *bnx2i_conn)
 {
 	struct iscsi_conn *conn = bnx2i_conn->cls_conn->dd_data;
 	struct iscsi_session *session = conn->session;
-	struct qp_info *qp = &bnx2i_conn->ep->qp;
+	struct qp_info *qp;
 	struct bnx2i_nop_in_msg *nopin;
 	int tgt_async_msg;
+	int cqe_cnt = 0;
 
+	if (bnx2i_conn->ep == NULL)
+		return 0;
+
+	qp = &bnx2i_conn->ep->qp;
+
+	if (!qp->cq_virt) {
+		printk(KERN_ALERT "bnx2i (%s): cq resr freed in bh execution!",
+			bnx2i_conn->hba->netdev->name);
+		goto out;
+	}
 	while (1) {
 		nopin = (struct bnx2i_nop_in_msg *) qp->cq_cons_qe;
 		if (nopin->cq_req_sn != qp->cqe_exp_seq_sn)
@@ -1873,8 +1987,9 @@ static void bnx2i_process_new_cqes(struct bnx2i_conn *bnx2i_conn)
 		switch (nopin->op_code) {
 		case ISCSI_OP_SCSI_CMD_RSP:
 		case ISCSI_OP_SCSI_DATA_IN:
-			bnx2i_process_scsi_cmd_resp(session, bnx2i_conn,
-						    qp->cq_cons_qe);
+			/* Run the kthread engine only for data cmds
+			   All other cmds will be completed in this bh! */
+			bnx2i_queue_scsi_cmd_resp(session, bnx2i_conn, nopin);
 			break;
 		case ISCSI_OP_LOGIN_RSP:
 			bnx2i_process_login_resp(session, bnx2i_conn,
@@ -1918,13 +2033,21 @@ static void bnx2i_process_new_cqes(struct bnx2i_conn *bnx2i_conn)
 			printk(KERN_ALERT "bnx2i: unknown opcode 0x%x\n",
 					  nopin->op_code);
 		}
-		if (!tgt_async_msg)
-			bnx2i_conn->ep->num_active_cmds--;
+		if (!tgt_async_msg) {
+			if (!atomic_read(&bnx2i_conn->ep->num_active_cmds))
+				printk(KERN_ALERT "bnx2i (%s): no active cmd! "
+				       "op 0x%x\n",
+				       bnx2i_conn->hba->netdev->name,
+				       nopin->op_code);
+			else
+				atomic_dec(&bnx2i_conn->ep->num_active_cmds);
+		}
 cqe_out:
 		/* clear out in production version only, till beta keep opcode
 		 * field intact, will be helpful in debugging (context dump)
 		 * nopin->op_code = 0;
 		 */
+		cqe_cnt++;
 		qp->cqe_exp_seq_sn++;
 		if (qp->cqe_exp_seq_sn == (qp->cqe_size * 2 + 1))
 			qp->cqe_exp_seq_sn = ISCSI_INITIAL_SN;
@@ -1937,6 +2060,8 @@ cqe_out:
 			qp->cq_cons_idx++;
 		}
 	}
+out:
+	return cqe_cnt;
 }
 
 /**
@@ -1952,6 +2077,7 @@ static void bnx2i_fastpath_notification(struct bnx2i_hba *hba,
 {
 	struct bnx2i_conn *bnx2i_conn;
 	u32 iscsi_cid;
+	int nxt_idx;
 
 	iscsi_cid = new_cqe_kcqe->iscsi_conn_id;
 	bnx2i_conn = bnx2i_get_conn_from_id(hba, iscsi_cid);
@@ -1964,9 +2090,12 @@ static void bnx2i_fastpath_notification(struct bnx2i_hba *hba,
 		printk(KERN_ALERT "cid #%x - ep not bound\n", iscsi_cid);
 		return;
 	}
+
 	bnx2i_process_new_cqes(bnx2i_conn);
-	bnx2i_arm_cq_event_coalescing(bnx2i_conn->ep, CNIC_ARM_CQE_FP);
-	bnx2i_process_new_cqes(bnx2i_conn);
+	nxt_idx = bnx2i_arm_cq_event_coalescing(bnx2i_conn->ep,
+						CNIC_ARM_CQE_FP);
+	if (nxt_idx && nxt_idx == bnx2i_process_new_cqes(bnx2i_conn))
+		bnx2i_arm_cq_event_coalescing(bnx2i_conn->ep, CNIC_ARM_CQE_FP);
 }
 
 
@@ -2312,7 +2441,7 @@ static void bnx2i_process_ofld_cmpl(struct bnx2i_hba *hba,
 			printk(KERN_ALERT "bnx2i (%s): ofld1 cmpl - invalid "
 				"opcode\n", hba->netdev->name);
 		else if (ofld_kcqe->completion_status ==
-			ISCSI_KCQE_COMPLETION_STATUS_CID_BUSY)
+			 ISCSI_KCQE_COMPLETION_STATUS_CID_BUSY)
 			/* error status code valid only for 5771x chipset */
 			ep->state = EP_STATE_OFLD_FAILED_CID_BUSY;
 		else
@@ -2511,7 +2640,7 @@ static void bnx2i_cm_remote_abort(struct cnic_sock *cm_sk)
 
 
 static int bnx2i_send_nl_mesg(void *context, u32 msg_type,
-			       char *buf, u16 buflen)
+			      char *buf, u16 buflen)
 {
 	struct bnx2i_hba *hba = context;
 	int rc;
diff --git a/drivers/scsi/bnx2i/bnx2i_init.c b/drivers/scsi/bnx2i/bnx2i_init.c
index 6adbdc3..0f7fb14 100644
--- a/drivers/scsi/bnx2i/bnx2i_init.c
+++ b/drivers/scsi/bnx2i/bnx2i_init.c
@@ -40,7 +40,7 @@ unsigned int event_coal_min = 24;
 module_param(event_coal_min, int, 0664);
 MODULE_PARM_DESC(event_coal_min, "Event Coalescing Minimum Commands");
 
-unsigned int event_coal_div = 1;
+unsigned int event_coal_div = 2;
 module_param(event_coal_div, int, 0664);
 MODULE_PARM_DESC(event_coal_div, "Event Coalescing Divide Factor");
 
@@ -66,6 +66,15 @@ MODULE_PARM_DESC(rq_size, "Configure RQ size");
 
 u64 iscsi_error_mask = 0x00;
 
+DEFINE_PER_CPU(struct bnx2i_percpu_s, bnx2i_percpu);
+
+static int bnx2i_cpu_callback(struct notifier_block *nfb,
+			      unsigned long action, void *hcpu);
+/* notification function for CPU hotplug events */
+static struct notifier_block bnx2i_cpu_notifier = {
+	.notifier_call = bnx2i_cpu_callback,
+};
+
 
 /**
  * bnx2i_identify_device - identifies NetXtreme II device type
@@ -362,6 +371,91 @@ void bnx2i_ulp_exit(struct cnic_dev *dev)
 
 
 /**
+ * bnx2i_percpu_thread_create - Create a receive thread for an
+ *				online CPU
+ *
+ * @cpu:	cpu index for the online cpu
+ */
+static void bnx2i_percpu_thread_create(unsigned int cpu)
+{
+	struct bnx2i_percpu_s *p;
+	struct task_struct *thread;
+
+	p = &per_cpu(bnx2i_percpu, cpu);
+
+	thread = kthread_create(bnx2i_percpu_io_thread, (void *)p,
+				"bnx2i_thread/%d", cpu);
+	/* bind thread to the cpu */
+	if (likely(!IS_ERR(thread))) {
+		kthread_bind(thread, cpu);
+		p->iothread = thread;
+		wake_up_process(thread);
+	}
+}
+
+
+static void bnx2i_percpu_thread_destroy(unsigned int cpu)
+{
+	struct bnx2i_percpu_s *p;
+	struct task_struct *thread;
+	struct bnx2i_work *work, *tmp;
+
+	/* Prevent any new work from being queued for this CPU */
+	p = &per_cpu(bnx2i_percpu, cpu);
+	spin_lock_bh(&p->p_work_lock);
+	thread = p->iothread;
+	p->iothread = NULL;
+
+	/* Free all work in the list */
+	list_for_each_entry_safe(work, tmp, &p->work_list, list) {
+		list_del_init(&work->list);
+		bnx2i_process_scsi_cmd_resp(work->session,
+					    work->bnx2i_conn, &work->cqe);
+		kfree(work);
+	}
+
+	spin_unlock_bh(&p->p_work_lock);
+	if (thread)
+		kthread_stop(thread);
+}
+
+
+/**
+ * bnx2i_cpu_callback - Handler for CPU hotplug events
+ *
+ * @nfb:	The callback data block
+ * @action:	The event triggering the callback
+ * @hcpu:	The index of the CPU that the event is for
+ *
+ * This creates or destroys per-CPU data for iSCSI
+ *
+ * Returns NOTIFY_OK always.
+ */
+static int bnx2i_cpu_callback(struct notifier_block *nfb,
+			      unsigned long action, void *hcpu)
+{
+	unsigned cpu = (unsigned long)hcpu;
+
+	switch (action) {
+	case CPU_ONLINE:
+	case CPU_ONLINE_FROZEN:
+		printk(KERN_INFO "bnx2i: CPU %x online: Create Rx thread\n",
+			cpu);
+		bnx2i_percpu_thread_create(cpu);
+		break;
+	case CPU_DEAD:
+	case CPU_DEAD_FROZEN:
+		printk(KERN_INFO "CPU %x offline: Remove Rx thread\n", cpu);
+		bnx2i_percpu_thread_destroy(cpu);
+		break;
+	default:
+		break;
+	}
+	return NOTIFY_OK;
+}
+
+
+/**
  * bnx2i_mod_init - module init entry point
  *
  * initialize any driver wide global data structures such as endpoint pool,
@@ -371,6 +465,8 @@ void bnx2i_ulp_exit(struct cnic_dev *dev)
 static int __init bnx2i_mod_init(void)
 {
 	int err;
+	unsigned cpu = 0;
+	struct bnx2i_percpu_s *p;
 
 	printk(KERN_INFO "%s", version);
 
@@ -393,6 +489,20 @@ static int __init bnx2i_mod_init(void)
 		goto unreg_xport;
 	}
 
+	/* Create percpu kernel threads to handle iSCSI I/O completions */
+	for_each_possible_cpu(cpu) {
+		p = &per_cpu(bnx2i_percpu, cpu);
+		INIT_LIST_HEAD(&p->work_list);
+		spin_lock_init(&p->p_work_lock);
+		p->iothread = NULL;
+	}
+
+	for_each_online_cpu(cpu)
+		bnx2i_percpu_thread_create(cpu);
+
+	/* Initialize per CPU interrupt thread */
+	register_hotcpu_notifier(&bnx2i_cpu_notifier);
+
 	return 0;
 
 unreg_xport:
@@ -413,6 +523,7 @@ out:
 static void __exit bnx2i_mod_exit(void)
 {
 	struct bnx2i_hba *hba;
+	unsigned cpu = 0;
 
 	mutex_lock(&bnx2i_dev_lock);
 	while (!list_empty(&adapter_list)) {
@@ -430,6 +541,11 @@ static void __exit bnx2i_mod_exit(void)
 	}
 	mutex_unlock(&bnx2i_dev_lock);
 
+	unregister_hotcpu_notifier(&bnx2i_cpu_notifier);
+
+	for_each_online_cpu(cpu)
+		bnx2i_percpu_thread_destroy(cpu);
+
 	iscsi_unregister_transport(&bnx2i_iscsi_transport);
 	cnic_unregister_driver(CNIC_ULP_ISCSI);
 }
diff --git a/drivers/scsi/bnx2i/bnx2i_iscsi.c b/drivers/scsi/bnx2i/bnx2i_iscsi.c
index 041928b..9d40f32 100644
--- a/drivers/scsi/bnx2i/bnx2i_iscsi.c
+++ b/drivers/scsi/bnx2i/bnx2i_iscsi.c
@@ -27,6 +27,7 @@ static struct scsi_host_template bnx2i_host_template;
  */
 static DEFINE_SPINLOCK(bnx2i_resc_lock); /* protects global resources */
 
+DECLARE_PER_CPU(struct bnx2i_percpu_s, bnx2i_percpu);
 
 static int bnx2i_adapter_ready(struct bnx2i_hba *hba)
 {
@@ -1214,7 +1215,8 @@ static int bnx2i_task_xmit(struct iscsi_task *task)
 	struct bnx2i_cmd *cmd = task->dd_data;
 	struct iscsi_cmd *hdr = (struct iscsi_cmd *) task->hdr;
 
-	if (bnx2i_conn->ep->num_active_cmds + 1 > hba->max_sqes)
+	if (atomic_read(&bnx2i_conn->ep->num_active_cmds) + 1  >
+	    hba->max_sqes)
 		return -ENOMEM;
 
 	/*
@@ -1354,6 +1356,9 @@ bnx2i_conn_create(struct iscsi_cls_session *cls_session, uint32_t cid)
 	bnx2i_conn = conn->dd_data;
 	bnx2i_conn->cls_conn = cls_conn;
 	bnx2i_conn->hba = hba;
+
+	atomic_set(&bnx2i_conn->work_cnt, 0);
+
 	/* 'ep' ptr will be assigned in bind() call */
 	bnx2i_conn->ep = NULL;
 	init_completion(&bnx2i_conn->cmd_cleanup_cmpl);
@@ -1457,11 +1462,34 @@ static void bnx2i_conn_destroy(struct iscsi_cls_conn *cls_conn)
 	struct bnx2i_conn *bnx2i_conn = conn->dd_data;
 	struct Scsi_Host *shost;
 	struct bnx2i_hba *hba;
+	struct bnx2i_work *work, *tmp;
+	unsigned cpu = 0;
+	struct bnx2i_percpu_s *p;
 
 	shost = iscsi_session_to_shost(iscsi_conn_to_session(cls_conn));
 	hba = iscsi_host_priv(shost);
 
 	bnx2i_conn_free_login_resources(hba, bnx2i_conn);
+
+	if (atomic_read(&bnx2i_conn->work_cnt)) {
+		for_each_online_cpu(cpu) {
+			p = &per_cpu(bnx2i_percpu, cpu);
+			spin_lock_bh(&p->p_work_lock);
+			list_for_each_entry_safe(work, tmp,
+						 &p->work_list, list) {
+				if (work->session == conn->session &&
+				    work->bnx2i_conn == bnx2i_conn) {
+					list_del_init(&work->list);
+					kfree(work);
+					if (!atomic_dec_and_test(
+							&bnx2i_conn->work_cnt))
+						break;
+				}
+			}
+			spin_unlock_bh(&p->p_work_lock);
+		}
+	}
+
 	iscsi_conn_teardown(cls_conn);
 }
 
@@ -1769,7 +1797,7 @@ static struct iscsi_endpoint *bnx2i_ep_connect(struct Scsi_Host *shost,
 	}
 	bnx2i_ep = ep->dd_data;
 
-	bnx2i_ep->num_active_cmds = 0;
+	atomic_set(&bnx2i_ep->num_active_cmds, 0);
 	iscsi_cid = bnx2i_alloc_iscsi_cid(hba);
 	if (iscsi_cid == -1) {
 		printk(KERN_ALERT "bnx2i (%s): alloc_ep - unable to allocate "
@@ -2163,9 +2191,9 @@ static struct scsi_host_template bnx2i_host_template = {
 	.eh_device_reset_handler = iscsi_eh_device_reset,
 	.eh_target_reset_handler = iscsi_eh_recover_target,
 	.change_queue_depth	= iscsi_change_queue_depth,
-	.can_queue		= 1024,
+	.can_queue		= 2048,
 	.max_sectors		= 127,
-	.cmd_per_lun		= 24,
+	.cmd_per_lun		= 128,
 	.this_id		= -1,
 	.use_clustering		= ENABLE_CLUSTERING,
 	.sg_tablesize		= ISCSI_MAX_BDS_PER_CMD,