[SCSI] ipr: Reduce lock contention

This patch reduces lock contention while implementing distributed
completion processing.

Signed-off-by: Wen Xiong <wenxiong@linux.vnet.ibm.com>
Acked-by: Brian King <brking@linux.vnet.ibm.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>

1 files changed, 228 insertions, 95 deletions

diff --git a/drivers/scsi/ipr.c b/drivers/scsi/ipr.c
index b8d759181611..f1eea5df54c6 100644
--- a/drivers/scsi/ipr.c
+++ b/drivers/scsi/ipr.c
@@ -552,7 +552,8 @@ static void ipr_trc_hook(struct ipr_cmnd *ipr_cmd,
 	struct ipr_trace_entry *trace_entry;
 	struct ipr_ioa_cfg *ioa_cfg = ipr_cmd->ioa_cfg;
 
-	trace_entry = &ioa_cfg->trace[ioa_cfg->trace_index++];
+	trace_entry = &ioa_cfg->trace[atomic_add_return
+			(1, &ioa_cfg->trace_index)%IPR_NUM_TRACE_ENTRIES];
 	trace_entry->time = jiffies;
 	trace_entry->op_code = ipr_cmd->ioarcb.cmd_pkt.cdb[0];
 	trace_entry->type = type;
@@ -563,6 +564,7 @@ static void ipr_trc_hook(struct ipr_cmnd *ipr_cmd,
 	trace_entry->cmd_index = ipr_cmd->cmd_index & 0xff;
 	trace_entry->res_handle = ipr_cmd->ioarcb.res_handle;
 	trace_entry->u.add_data = add_data;
+	wmb();
 }
 #else
 #define ipr_trc_hook(ipr_cmd, type, add_data) do { } while (0)
@@ -697,9 +699,15 @@ static void ipr_mask_and_clear_interrupts(struct ipr_ioa_cfg *ioa_cfg,
 					  u32 clr_ints)
 {
 	volatile u32 int_reg;
+	int i;
 
 	/* Stop new interrupts */
-	ioa_cfg->allow_interrupts = 0;
+	for (i = 0; i < ioa_cfg->hrrq_num; i++) {
+		spin_lock(&ioa_cfg->hrrq[i]._lock);
+		ioa_cfg->hrrq[i].allow_interrupts = 0;
+		spin_unlock(&ioa_cfg->hrrq[i]._lock);
+	}
+	wmb();
 
 	/* Set interrupt mask to stop all new interrupts */
 	if (ioa_cfg->sis64)
@@ -818,6 +826,7 @@ static void ipr_fail_all_ops(struct ipr_ioa_cfg *ioa_cfg)
 
 	ENTER;
 	for_each_hrrq(hrrq, ioa_cfg) {
+		spin_lock(&hrrq->_lock);
 		list_for_each_entry_safe(ipr_cmd,
 					temp, &hrrq->hrrq_pending_q, queue) {
 			list_del(&ipr_cmd->queue);
@@ -837,6 +846,7 @@ static void ipr_fail_all_ops(struct ipr_ioa_cfg *ioa_cfg)
 			del_timer(&ipr_cmd->timer);
 			ipr_cmd->done(ipr_cmd);
 		}
+		spin_unlock(&hrrq->_lock);
 	}
 	LEAVE;
 }
@@ -991,12 +1001,9 @@ static void ipr_send_blocking_cmd(struct ipr_cmnd *ipr_cmd,
 static int ipr_get_hrrq_index(struct ipr_ioa_cfg *ioa_cfg)
 {
 	if (ioa_cfg->hrrq_num == 1)
-		ioa_cfg->hrrq_index = 0;
-	else {
-		if (++ioa_cfg->hrrq_index >= ioa_cfg->hrrq_num)
-			ioa_cfg->hrrq_index = 1;
-	}
-	return ioa_cfg->hrrq_index;
+		return 0;
+	else
+		return (atomic_add_return(1, &ioa_cfg->hrrq_index) % (ioa_cfg->hrrq_num - 1)) + 1;
 }
 
 /**
@@ -1018,7 +1025,7 @@ static void ipr_send_hcam(struct ipr_ioa_cfg *ioa_cfg, u8 type,
 	struct ipr_cmnd *ipr_cmd;
 	struct ipr_ioarcb *ioarcb;
 
-	if (ioa_cfg->allow_cmds) {
+	if (ioa_cfg->hrrq[IPR_INIT_HRRQ].allow_cmds) {
 		ipr_cmd = ipr_get_free_ipr_cmnd(ioa_cfg);
 		list_add_tail(&ipr_cmd->queue, &ipr_cmd->hrrq->hrrq_pending_q);
 		list_add_tail(&hostrcb->queue, &ioa_cfg->hostrcb_pending_q);
@@ -2564,7 +2571,7 @@ static int ipr_reset_reload(struct ipr_ioa_cfg *ioa_cfg,
 
 	/* If we got hit with a host reset while we were already resetting
 	 the adapter for some reason, and the reset failed. */
-	if (ioa_cfg->ioa_is_dead) {
+	if (ioa_cfg->hrrq[IPR_INIT_HRRQ].ioa_is_dead) {
 		ipr_trace;
 		return FAILED;
 	}
@@ -3205,7 +3212,8 @@ static void ipr_worker_thread(struct work_struct *work)
 restart:
 	do {
 		did_work = 0;
-		if (!ioa_cfg->allow_cmds || !ioa_cfg->allow_ml_add_del) {
+		if (!ioa_cfg->hrrq[IPR_INIT_HRRQ].allow_cmds ||
+		    !ioa_cfg->allow_ml_add_del) {
 			spin_unlock_irqrestore(ioa_cfg->host->host_lock, lock_flags);
 			return;
 		}
@@ -3453,7 +3461,7 @@ static ssize_t ipr_show_adapter_state(struct device *dev,
 	int len;
 
 	spin_lock_irqsave(ioa_cfg->host->host_lock, lock_flags);
-	if (ioa_cfg->ioa_is_dead)
+	if (ioa_cfg->hrrq[IPR_INIT_HRRQ].ioa_is_dead)
 		len = snprintf(buf, PAGE_SIZE, "offline\n");
 	else
 		len = snprintf(buf, PAGE_SIZE, "online\n");
@@ -3479,14 +3487,20 @@ static ssize_t ipr_store_adapter_state(struct device *dev,
 	struct Scsi_Host *shost = class_to_shost(dev);
 	struct ipr_ioa_cfg *ioa_cfg = (struct ipr_ioa_cfg *)shost->hostdata;
 	unsigned long lock_flags;
-	int result = count;
+	int result = count, i;
 
 	if (!capable(CAP_SYS_ADMIN))
 		return -EACCES;
 
 	spin_lock_irqsave(ioa_cfg->host->host_lock, lock_flags);
-	if (ioa_cfg->ioa_is_dead && !strncmp(buf, "online", 6)) {
-		ioa_cfg->ioa_is_dead = 0;
+	if (ioa_cfg->hrrq[IPR_INIT_HRRQ].ioa_is_dead &&
+	    !strncmp(buf, "online", 6)) {
+		for (i = 0; i < ioa_cfg->hrrq_num; i++) {
+			spin_lock(&ioa_cfg->hrrq[i]._lock);
+			ioa_cfg->hrrq[i].ioa_is_dead = 0;
+			spin_unlock(&ioa_cfg->hrrq[i]._lock);
+		}
+		wmb();
 		ioa_cfg->reset_retries = 0;
 		ioa_cfg->in_ioa_bringdown = 0;
 		ipr_initiate_ioa_reset(ioa_cfg, IPR_SHUTDOWN_NONE);
@@ -4066,7 +4080,7 @@ static int ipr_alloc_dump(struct ipr_ioa_cfg *ioa_cfg)
 
 	ioa_cfg->dump = dump;
 	ioa_cfg->sdt_state = WAIT_FOR_DUMP;
-	if (ioa_cfg->ioa_is_dead && !ioa_cfg->dump_taken) {
+	if (ioa_cfg->hrrq[IPR_INIT_HRRQ].ioa_is_dead && !ioa_cfg->dump_taken) {
 		ioa_cfg->dump_taken = 1;
 		schedule_work(&ioa_cfg->work_q);
 	}
@@ -4861,10 +4875,11 @@ static int __ipr_eh_dev_reset(struct scsi_cmnd *scsi_cmd)
 	 */
 	if (ioa_cfg->in_reset_reload)
 		return FAILED;
-	if (ioa_cfg->ioa_is_dead)
+	if (ioa_cfg->hrrq[IPR_INIT_HRRQ].ioa_is_dead)
 		return FAILED;
 
 	for_each_hrrq(hrrq, ioa_cfg) {
+		spin_lock(&hrrq->_lock);
 		list_for_each_entry(ipr_cmd, &hrrq->hrrq_pending_q, queue) {
 			if (ipr_cmd->ioarcb.res_handle == res->res_handle) {
 				if (ipr_cmd->scsi_cmd)
@@ -4878,6 +4893,7 @@ static int __ipr_eh_dev_reset(struct scsi_cmnd *scsi_cmd)
 				}
 			}
 		}
+		spin_unlock(&hrrq->_lock);
 	}
 	res->resetting_device = 1;
 	scmd_printk(KERN_ERR, scsi_cmd, "Resetting device\n");
@@ -4889,6 +4905,7 @@ static int __ipr_eh_dev_reset(struct scsi_cmnd *scsi_cmd)
 		spin_lock_irq(scsi_cmd->device->host->host_lock);
 
 		for_each_hrrq(hrrq, ioa_cfg) {
+			spin_lock(&hrrq->_lock);
 			list_for_each_entry(ipr_cmd,
 					    &hrrq->hrrq_pending_q, queue) {
 				if (ipr_cmd->ioarcb.res_handle ==
@@ -4897,6 +4914,7 @@ static int __ipr_eh_dev_reset(struct scsi_cmnd *scsi_cmd)
 					break;
 				}
 			}
+			spin_unlock(&hrrq->_lock);
 		}
 	} else
 		rc = ipr_device_reset(ioa_cfg, res);
@@ -5020,7 +5038,8 @@ static int ipr_cancel_op(struct scsi_cmnd *scsi_cmd)
 	 * This will force the mid-layer to call ipr_eh_host_reset,
 	 * which will then go to sleep and wait for the reset to complete
 	 */
-	if (ioa_cfg->in_reset_reload || ioa_cfg->ioa_is_dead)
+	if (ioa_cfg->in_reset_reload ||
+	    ioa_cfg->hrrq[IPR_INIT_HRRQ].ioa_is_dead)
 		return FAILED;
 	if (!res)
 		return FAILED;
@@ -5036,6 +5055,7 @@ static int ipr_cancel_op(struct scsi_cmnd *scsi_cmd)
 		return FAILED;
 
 	for_each_hrrq(hrrq, ioa_cfg) {
+		spin_lock(&hrrq->_lock);
 		list_for_each_entry(ipr_cmd, &hrrq->hrrq_pending_q, queue) {
 			if (ipr_cmd->scsi_cmd == scsi_cmd) {
 				ipr_cmd->done = ipr_scsi_eh_done;
@@ -5043,6 +5063,7 @@ static int ipr_cancel_op(struct scsi_cmnd *scsi_cmd)
 				break;
 			}
 		}
+		spin_unlock(&hrrq->_lock);
 	}
 
 	if (!op_found)
@@ -5112,6 +5133,7 @@ static irqreturn_t ipr_handle_other_interrupt(struct ipr_ioa_cfg *ioa_cfg,
 {
 	irqreturn_t rc = IRQ_HANDLED;
 	u32 int_mask_reg;
+
 	int_mask_reg = readl(ioa_cfg->regs.sense_interrupt_mask_reg32);
 	int_reg &= ~int_mask_reg;
 
@@ -5173,6 +5195,7 @@ static irqreturn_t ipr_handle_other_interrupt(struct ipr_ioa_cfg *ioa_cfg,
 		ipr_mask_and_clear_interrupts(ioa_cfg, ~0);
 		ipr_initiate_ioa_reset(ioa_cfg, IPR_SHUTDOWN_NONE);
 	}
+
 	return rc;
 }
 
@@ -5205,7 +5228,7 @@ static int __ipr_process_hrrq(struct ipr_hrr_queue *hrr_queue,
 	int num_hrrq = 0;
 
 	/* If interrupts are disabled, ignore the interrupt */
-	if (!ioa_cfg->allow_interrupts)
+	if (!hrr_queue->allow_interrupts)
 		return 0;
 
 	while ((be32_to_cpu(*hrr_queue->hrrq_curr) & IPR_HRRQ_TOGGLE_BIT) ==
@@ -5252,7 +5275,7 @@ static irqreturn_t ipr_isr(int irq, void *devp)
 {
 	struct ipr_hrr_queue *hrrq = (struct ipr_hrr_queue *)devp;
 	struct ipr_ioa_cfg *ioa_cfg = hrrq->ioa_cfg;
-	unsigned long lock_flags = 0;
+	unsigned long hrrq_flags = 0;
 	u32 int_reg = 0;
 	u32 ioasc;
 	u16 cmd_index;
@@ -5262,10 +5285,10 @@ static irqreturn_t ipr_isr(int irq, void *devp)
 	irqreturn_t rc = IRQ_NONE;
 	LIST_HEAD(doneq);
 
-	spin_lock_irqsave(ioa_cfg->host->host_lock, lock_flags);
+	spin_lock_irqsave(hrrq->lock, hrrq_flags);
 	/* If interrupts are disabled, ignore the interrupt */
-	if (!ioa_cfg->allow_interrupts) {
-		spin_unlock_irqrestore(ioa_cfg->host->host_lock, lock_flags);
+	if (!hrrq->allow_interrupts) {
+		spin_unlock_irqrestore(hrrq->lock, hrrq_flags);
 		return IRQ_NONE;
 	}
 
@@ -5332,7 +5355,7 @@ static irqreturn_t ipr_isr(int irq, void *devp)
 		rc = ipr_handle_other_interrupt(ioa_cfg, int_reg);
 
 unlock_out:
-	spin_unlock_irqrestore(ioa_cfg->host->host_lock, lock_flags);
+	spin_unlock_irqrestore(hrrq->lock, hrrq_flags);
 	list_for_each_entry_safe(ipr_cmd, temp, &doneq, queue) {
 		list_del(&ipr_cmd->queue);
 		del_timer(&ipr_cmd->timer);
@@ -5352,17 +5375,16 @@ unlock_out:
 static irqreturn_t ipr_isr_mhrrq(int irq, void *devp)
 {
 	struct ipr_hrr_queue *hrrq = (struct ipr_hrr_queue *)devp;
-	struct ipr_ioa_cfg *ioa_cfg = hrrq->ioa_cfg;
-	unsigned long lock_flags = 0;
+	unsigned long hrrq_flags = 0;
 	struct ipr_cmnd *ipr_cmd, *temp;
 	irqreturn_t rc = IRQ_NONE;
 	LIST_HEAD(doneq);
 
-	spin_lock_irqsave(ioa_cfg->host->host_lock, lock_flags);
+	spin_lock_irqsave(hrrq->lock, hrrq_flags);
 
 	/* If interrupts are disabled, ignore the interrupt */
-	if (!ioa_cfg->allow_interrupts) {
-		spin_unlock_irqrestore(ioa_cfg->host->host_lock, lock_flags);
+	if (!hrrq->allow_interrupts) {
+		spin_unlock_irqrestore(hrrq->lock, hrrq_flags);
 		return IRQ_NONE;
 	}
 
@@ -5372,7 +5394,7 @@ static irqreturn_t ipr_isr_mhrrq(int irq, void *devp)
 		if (__ipr_process_hrrq(hrrq, &doneq))
 			rc =  IRQ_HANDLED;
 
-	spin_unlock_irqrestore(ioa_cfg->host->host_lock, lock_flags);
+	spin_unlock_irqrestore(hrrq->lock, hrrq_flags);
 
 	list_for_each_entry_safe(ipr_cmd, temp, &doneq, queue) {
 		list_del(&ipr_cmd->queue);
@@ -5965,14 +5987,14 @@ static void ipr_scsi_done(struct ipr_cmnd *ipr_cmd)
 	if (likely(IPR_IOASC_SENSE_KEY(ioasc) == 0)) {
 		scsi_dma_unmap(scsi_cmd);
 
-		spin_lock_irqsave(ioa_cfg->host->host_lock, hrrq_flags);
+		spin_lock_irqsave(ipr_cmd->hrrq->lock, hrrq_flags);
 		list_add_tail(&ipr_cmd->queue, &ipr_cmd->hrrq->hrrq_free_q);
 		scsi_cmd->scsi_done(scsi_cmd);
-		spin_unlock_irqrestore(ioa_cfg->host->host_lock, hrrq_flags);
+		spin_unlock_irqrestore(ipr_cmd->hrrq->lock, hrrq_flags);
 	} else {
-		spin_lock_irqsave(ioa_cfg->host->host_lock, hrrq_flags);
+		spin_lock_irqsave(ipr_cmd->hrrq->lock, hrrq_flags);
 		ipr_erp_start(ioa_cfg, ipr_cmd);
-		spin_unlock_irqrestore(ioa_cfg->host->host_lock, hrrq_flags);
+		spin_unlock_irqrestore(ipr_cmd->hrrq->lock, hrrq_flags);
 	}
 }
 
@@ -5995,26 +6017,34 @@ static int ipr_queuecommand(struct Scsi_Host *shost,
 	struct ipr_resource_entry *res;
 	struct ipr_ioarcb *ioarcb;
 	struct ipr_cmnd *ipr_cmd;
-	unsigned long lock_flags;
+	unsigned long hrrq_flags, lock_flags;
 	int rc;
 	struct ipr_hrr_queue *hrrq;
 	int hrrq_id;
 
 	ioa_cfg = (struct ipr_ioa_cfg *)shost->hostdata;
 
-	spin_lock_irqsave(shost->host_lock, lock_flags);
 	scsi_cmd->result = (DID_OK << 16);
 	res = scsi_cmd->device->hostdata;
+
+	if (ipr_is_gata(res) && res->sata_port) {
+		spin_lock_irqsave(ioa_cfg->host->host_lock, lock_flags);
+		rc = ata_sas_queuecmd(scsi_cmd, res->sata_port->ap);
+		spin_unlock_irqrestore(ioa_cfg->host->host_lock, lock_flags);
+		return rc;
+	}
+
 	hrrq_id = ipr_get_hrrq_index(ioa_cfg);
 	hrrq = &ioa_cfg->hrrq[hrrq_id];
 
+	spin_lock_irqsave(hrrq->lock, hrrq_flags);
 	/*
 	 * We are currently blocking all devices due to a host reset
 	 * We have told the host to stop giving us new requests, but
 	 * ERP ops don't count. FIXME
 	 */
-	if (unlikely(!ioa_cfg->allow_cmds && !ioa_cfg->ioa_is_dead)) {
-		spin_unlock_irqrestore(shost->host_lock, lock_flags);
+	if (unlikely(!hrrq->allow_cmds && !hrrq->ioa_is_dead)) {
+		spin_unlock_irqrestore(hrrq->lock, hrrq_flags);
 		return SCSI_MLQUEUE_HOST_BUSY;
 	}
 
@@ -6022,23 +6052,17 @@ static int ipr_queuecommand(struct Scsi_Host *shost,
 	 * FIXME - Create scsi_set_host_offline interface
 	 *  and the ioa_is_dead check can be removed
 	 */
-	if (unlikely(ioa_cfg->ioa_is_dead || !res)) {
-		spin_unlock_irqrestore(shost->host_lock, lock_flags);
+	if (unlikely(hrrq->ioa_is_dead || !res)) {
+		spin_unlock_irqrestore(hrrq->lock, hrrq_flags);
 		goto err_nodev;
 	}
 
-	if (ipr_is_gata(res) && res->sata_port) {
-		rc = ata_sas_queuecmd(scsi_cmd, res->sata_port->ap);
-		spin_unlock_irqrestore(shost->host_lock, lock_flags);
-		return rc;
-	}
-
 	ipr_cmd = __ipr_get_free_ipr_cmnd(hrrq);
 	if (ipr_cmd == NULL) {
-		spin_unlock_irqrestore(shost->host_lock, lock_flags);
+		spin_unlock_irqrestore(hrrq->lock, hrrq_flags);
 		return SCSI_MLQUEUE_HOST_BUSY;
 	}
-	spin_unlock_irqrestore(shost->host_lock, lock_flags);
+	spin_unlock_irqrestore(hrrq->lock, hrrq_flags);
 
 	ipr_init_ipr_cmnd(ipr_cmd, ipr_scsi_done);
 	ioarcb = &ipr_cmd->ioarcb;
@@ -6068,18 +6092,18 @@ static int ipr_queuecommand(struct Scsi_Host *shost,
 	else
 		rc = ipr_build_ioadl(ioa_cfg, ipr_cmd);
 
-	spin_lock_irqsave(shost->host_lock, lock_flags);
-	if (unlikely(rc || (!ioa_cfg->allow_cmds && !ioa_cfg->ioa_is_dead))) {
+	spin_lock_irqsave(hrrq->lock, hrrq_flags);
+	if (unlikely(rc || (!hrrq->allow_cmds && !hrrq->ioa_is_dead))) {
 		list_add_tail(&ipr_cmd->queue, &hrrq->hrrq_free_q);
-		spin_unlock_irqrestore(shost->host_lock, lock_flags);
+		spin_unlock_irqrestore(hrrq->lock, hrrq_flags);
 		if (!rc)
 			scsi_dma_unmap(scsi_cmd);
 		return SCSI_MLQUEUE_HOST_BUSY;
 	}
 
-	if (unlikely(ioa_cfg->ioa_is_dead)) {
+	if (unlikely(hrrq->ioa_is_dead)) {
 		list_add_tail(&ipr_cmd->queue, &hrrq->hrrq_free_q);
-		spin_unlock_irqrestore(shost->host_lock, lock_flags);
+		spin_unlock_irqrestore(hrrq->lock, hrrq_flags);
 		scsi_dma_unmap(scsi_cmd);
 		goto err_nodev;
 	}
@@ -6092,15 +6116,15 @@ static int ipr_queuecommand(struct Scsi_Host *shost,
 	list_add_tail(&ipr_cmd->queue, &hrrq->hrrq_pending_q);
 	ipr_trc_hook(ipr_cmd, IPR_TRACE_START, IPR_GET_RES_PHYS_LOC(res));
 	ipr_send_command(ipr_cmd);
-	spin_unlock_irqrestore(shost->host_lock, lock_flags);
+	spin_unlock_irqrestore(hrrq->lock, hrrq_flags);
 	return 0;
 
 err_nodev:
-	spin_lock_irqsave(shost->host_lock, lock_flags);
+	spin_lock_irqsave(hrrq->lock, hrrq_flags);
 	memset(scsi_cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
 	scsi_cmd->result = (DID_NO_CONNECT << 16);
 	scsi_cmd->scsi_done(scsi_cmd);
-	spin_unlock_irqrestore(shost->host_lock, lock_flags);
+	spin_unlock_irqrestore(hrrq->lock, hrrq_flags);
 	return 0;
 }
 
@@ -6198,7 +6222,7 @@ static void ipr_ata_phy_reset(struct ata_port *ap)
 		spin_lock_irqsave(ioa_cfg->host->host_lock, flags);
 	}
 
-	if (!ioa_cfg->allow_cmds)
+	if (!ioa_cfg->hrrq[IPR_INIT_HRRQ].allow_cmds)
 		goto out_unlock;
 
 	rc = ipr_device_reset(ioa_cfg, res);
@@ -6240,12 +6264,14 @@ static void ipr_ata_post_internal(struct ata_queued_cmd *qc)
 	}
 
 	for_each_hrrq(hrrq, ioa_cfg) {
+		spin_lock(&hrrq->_lock);
 		list_for_each_entry(ipr_cmd, &hrrq->hrrq_pending_q, queue) {
 			if (ipr_cmd->qc == qc) {
 				ipr_device_reset(ioa_cfg, sata_port->res);
 				break;
 			}
 		}
+		spin_unlock(&hrrq->_lock);
 	}
 	spin_unlock_irqrestore(ioa_cfg->host->host_lock, flags);
 }
@@ -6294,6 +6320,7 @@ static void ipr_sata_done(struct ipr_cmnd *ipr_cmd)
 	struct ipr_resource_entry *res = sata_port->res;
 	u32 ioasc = be32_to_cpu(ipr_cmd->s.ioasa.hdr.ioasc);
 
+	spin_lock(&ipr_cmd->hrrq->_lock);
 	if (ipr_cmd->ioa_cfg->sis64)
 		memcpy(&sata_port->ioasa, &ipr_cmd->s.ioasa64.u.gata,
 		       sizeof(struct ipr_ioasa_gata));
@@ -6310,6 +6337,7 @@ static void ipr_sata_done(struct ipr_cmnd *ipr_cmd)
 	else
 		qc->err_mask |= ac_err_mask(sata_port->ioasa.status);
 	list_add_tail(&ipr_cmd->queue, &ipr_cmd->hrrq->hrrq_free_q);
+	spin_unlock(&ipr_cmd->hrrq->_lock);
 	ata_qc_complete(qc);
 }
 
@@ -6405,6 +6433,48 @@ static void ipr_build_ata_ioadl(struct ipr_cmnd *ipr_cmd,
 }
 
 /**
+ * ipr_qc_defer - Get a free ipr_cmd
+ * @qc:	queued command
+ *
+ * Return value:
+ *	0 if success
+ **/
+static int ipr_qc_defer(struct ata_queued_cmd *qc)
+{
+	struct ata_port *ap = qc->ap;
+	struct ipr_sata_port *sata_port = ap->private_data;
+	struct ipr_ioa_cfg *ioa_cfg = sata_port->ioa_cfg;
+	struct ipr_cmnd *ipr_cmd;
+	struct ipr_hrr_queue *hrrq;
+	int hrrq_id;
+
+	hrrq_id = ipr_get_hrrq_index(ioa_cfg);
+	hrrq = &ioa_cfg->hrrq[hrrq_id];
+
+	qc->lldd_task = NULL;
+	spin_lock(&hrrq->_lock);
+	if (unlikely(hrrq->ioa_is_dead)) {
+		spin_unlock(&hrrq->_lock);
+		return 0;
+	}
+
+	if (unlikely(!hrrq->allow_cmds)) {
+		spin_unlock(&hrrq->_lock);
+		return ATA_DEFER_LINK;
+	}
+
+	ipr_cmd = __ipr_get_free_ipr_cmnd(hrrq);
+	if (ipr_cmd == NULL) {
+		spin_unlock(&hrrq->_lock);
+		return ATA_DEFER_LINK;
+	}
+
+	qc->lldd_task = ipr_cmd;
+	spin_unlock(&hrrq->_lock);
+	return 0;
+}
+
+/**
  * ipr_qc_issue - Issue a SATA qc to a device
  * @qc:	queued command
  *
@@ -6420,15 +6490,23 @@ static unsigned int ipr_qc_issue(struct ata_queued_cmd *qc)
 	struct ipr_cmnd *ipr_cmd;
 	struct ipr_ioarcb *ioarcb;
 	struct ipr_ioarcb_ata_regs *regs;
-	struct ipr_hrr_queue *hrrq;
-	int hrrq_id;
 
-	if (unlikely(!ioa_cfg->allow_cmds || ioa_cfg->ioa_is_dead))
+	if (qc->lldd_task == NULL)
+		ipr_qc_defer(qc);
+
+	ipr_cmd = qc->lldd_task;
+	if (ipr_cmd == NULL)
 		return AC_ERR_SYSTEM;
 
-	hrrq_id = ipr_get_hrrq_index(ioa_cfg);
-	hrrq = &ioa_cfg->hrrq[hrrq_id];
-	ipr_cmd = __ipr_get_free_ipr_cmnd(hrrq);
+	qc->lldd_task = NULL;
+	spin_lock(&ipr_cmd->hrrq->_lock);
+	if (unlikely(!ipr_cmd->hrrq->allow_cmds ||
+			ipr_cmd->hrrq->ioa_is_dead)) {
+		list_add_tail(&ipr_cmd->queue, &ipr_cmd->hrrq->hrrq_free_q);
+		spin_unlock(&ipr_cmd->hrrq->_lock);
+		return AC_ERR_SYSTEM;
+	}
+
 	ipr_init_ipr_cmnd(ipr_cmd, ipr_lock_and_done);
 	ioarcb = &ipr_cmd->ioarcb;
 
@@ -6441,7 +6519,7 @@ static unsigned int ipr_qc_issue(struct ata_queued_cmd *qc)
 	memset(regs, 0, sizeof(*regs));
 	ioarcb->add_cmd_parms_len = cpu_to_be16(sizeof(*regs));
 
-	list_add_tail(&ipr_cmd->queue, &hrrq->hrrq_pending_q);
+	list_add_tail(&ipr_cmd->queue, &ipr_cmd->hrrq->hrrq_pending_q);
 	ipr_cmd->qc = qc;
 	ipr_cmd->done = ipr_sata_done;
 	ipr_cmd->ioarcb.res_handle = res->res_handle;
@@ -6481,10 +6559,12 @@ static unsigned int ipr_qc_issue(struct ata_queued_cmd *qc)
 
 	default:
 		WARN_ON(1);
+		spin_unlock(&ipr_cmd->hrrq->_lock);
 		return AC_ERR_INVALID;
 	}
 
 	ipr_send_command(ipr_cmd);
+	spin_unlock(&ipr_cmd->hrrq->_lock);
 
 	return 0;
 }
@@ -6523,6 +6603,7 @@ static struct ata_port_operations ipr_sata_ops = {
 	.hardreset = ipr_sata_reset,
 	.post_internal_cmd = ipr_ata_post_internal,
 	.qc_prep = ata_noop_qc_prep,
+	.qc_defer = ipr_qc_defer,
 	.qc_issue = ipr_qc_issue,
 	.qc_fill_rtf = ipr_qc_fill_rtf,
 	.port_start = ata_sas_port_start,
@@ -6620,11 +6701,16 @@ static int ipr_ioa_reset_done(struct ipr_cmnd *ipr_cmd)
 	struct ipr_ioa_cfg *ioa_cfg = ipr_cmd->ioa_cfg;
 	struct ipr_resource_entry *res;
 	struct ipr_hostrcb *hostrcb, *temp;
-	int i = 0;
+	int i = 0, j;
 
 	ENTER;
 	ioa_cfg->in_reset_reload = 0;
-	ioa_cfg->allow_cmds = 1;
+	for (j = 0; j < ioa_cfg->hrrq_num; j++) {
+		spin_lock(&ioa_cfg->hrrq[j]._lock);
+		ioa_cfg->hrrq[j].allow_cmds = 1;
+		spin_unlock(&ioa_cfg->hrrq[j]._lock);
+	}
+	wmb();
 	ioa_cfg->reset_cmd = NULL;
 	ioa_cfg->doorbell |= IPR_RUNTIME_RESET;
 
@@ -6655,7 +6741,7 @@ static int ipr_ioa_reset_done(struct ipr_cmnd *ipr_cmd)
 	scsi_unblock_requests(ioa_cfg->host);
 	spin_lock(ioa_cfg->host->host_lock);
 
-	if (!ioa_cfg->allow_cmds)
+	if (!ioa_cfg->hrrq[IPR_INIT_HRRQ].allow_cmds)
 		scsi_block_requests(ioa_cfg->host);
 
 	LEAVE;
@@ -7452,8 +7538,8 @@ static int ipr_ioafp_identify_hrrq(struct ipr_cmnd *ipr_cmd)
 	ipr_cmd->job_step = ipr_ioafp_std_inquiry;
 	dev_info(&ioa_cfg->pdev->dev, "Starting IOA initialization sequence.\n");
 
-	if (ioa_cfg->hrrq_index < ioa_cfg->hrrq_num) {
-		hrrq = &ioa_cfg->hrrq[ioa_cfg->hrrq_index];
+	if (ioa_cfg->identify_hrrq_index < ioa_cfg->hrrq_num) {
+		hrrq = &ioa_cfg->hrrq[ioa_cfg->identify_hrrq_index];
 
 		ioarcb->cmd_pkt.cdb[0] = IPR_ID_HOST_RR_Q;
 		ioarcb->res_handle = cpu_to_be32(IPR_IOA_RES_HANDLE);
@@ -7481,7 +7567,8 @@ static int ipr_ioafp_identify_hrrq(struct ipr_cmnd *ipr_cmd)
 			(sizeof(u32) * hrrq->size) & 0xff;
 
 		if (ioarcb->cmd_pkt.cdb[1] & IPR_ID_HRRQ_SELE_ENABLE)
-			ioarcb->cmd_pkt.cdb[9] = ioa_cfg->hrrq_index;
+			ioarcb->cmd_pkt.cdb[9] =
+					ioa_cfg->identify_hrrq_index;
 
 		if (ioa_cfg->sis64) {
 			ioarcb->cmd_pkt.cdb[10] =
@@ -7495,24 +7582,19 @@ static int ipr_ioafp_identify_hrrq(struct ipr_cmnd *ipr_cmd)
 		}
 
 		if (ioarcb->cmd_pkt.cdb[1] & IPR_ID_HRRQ_SELE_ENABLE)
-			ioarcb->cmd_pkt.cdb[14] = ioa_cfg->hrrq_index;
+			ioarcb->cmd_pkt.cdb[14] =
+					ioa_cfg->identify_hrrq_index;
 
 		ipr_do_req(ipr_cmd, ipr_reset_ioa_job, ipr_timeout,
 			   IPR_INTERNAL_TIMEOUT);
 
-		if (++ioa_cfg->hrrq_index < ioa_cfg->hrrq_num)
-				ipr_cmd->job_step = ipr_ioafp_identify_hrrq;
+		if (++ioa_cfg->identify_hrrq_index < ioa_cfg->hrrq_num)
+			ipr_cmd->job_step = ipr_ioafp_identify_hrrq;
 
 		LEAVE;
 		return IPR_RC_JOB_RETURN;
-
 	}
 
-	if (ioa_cfg->hrrq_num == 1)
-		ioa_cfg->hrrq_index = 0;
-	else
-		ioa_cfg->hrrq_index = 1;
-
 	LEAVE;
 	return IPR_RC_JOB_CONTINUE;
 }
@@ -7571,7 +7653,6 @@ static void ipr_reset_start_timer(struct ipr_cmnd *ipr_cmd,
 	ipr_cmd->timer.expires = jiffies + timeout;
 	ipr_cmd->timer.function = (void (*)(unsigned long))ipr_reset_timer_done;
 	add_timer(&ipr_cmd->timer);
-	LEAVE;
 }
 
 /**
@@ -7586,6 +7667,7 @@ static void ipr_init_ioa_mem(struct ipr_ioa_cfg *ioa_cfg)
 	struct ipr_hrr_queue *hrrq;
 
 	for_each_hrrq(hrrq, ioa_cfg) {
+		spin_lock(&hrrq->_lock);
 		memset(hrrq->host_rrq, 0, sizeof(u32) * hrrq->size);
 
 		/* Initialize Host RRQ pointers */
@@ -7593,9 +7675,15 @@ static void ipr_init_ioa_mem(struct ipr_ioa_cfg *ioa_cfg)
 		hrrq->hrrq_end = &hrrq->host_rrq[hrrq->size - 1];
 		hrrq->hrrq_curr = hrrq->hrrq_start;
 		hrrq->toggle_bit = 1;
+		spin_unlock(&hrrq->_lock);
 	}
+	wmb();
 
-	ioa_cfg->hrrq_index = 0;
+	ioa_cfg->identify_hrrq_index = 0;
+	if (ioa_cfg->hrrq_num == 1)
+		atomic_set(&ioa_cfg->hrrq_index, 0);
+	else
+		atomic_set(&ioa_cfg->hrrq_index, 1);
 
 	/* Zero out config table */
 	memset(ioa_cfg->u.cfg_table, 0, ioa_cfg->cfg_table_size);
@@ -7673,12 +7761,18 @@ static int ipr_reset_enable_ioa(struct ipr_cmnd *ipr_cmd)
 	struct ipr_ioa_cfg *ioa_cfg = ipr_cmd->ioa_cfg;
 	volatile u32 int_reg;
 	volatile u64 maskval;
+	int i;
 
 	ENTER;
 	ipr_cmd->job_step = ipr_ioafp_identify_hrrq;
 	ipr_init_ioa_mem(ioa_cfg);
 
-	ioa_cfg->allow_interrupts = 1;
+	for (i = 0; i < ioa_cfg->hrrq_num; i++) {
+		spin_lock(&ioa_cfg->hrrq[i]._lock);
+		ioa_cfg->hrrq[i].allow_interrupts = 1;
+		spin_unlock(&ioa_cfg->hrrq[i]._lock);
+	}
+	wmb();
 	if (ioa_cfg->sis64) {
 		/* Set the adapter to the correct endian mode. */
 		writel(IPR_ENDIAN_SWAP_KEY, ioa_cfg->regs.endian_swap_reg);
@@ -8237,7 +8331,8 @@ static int ipr_reset_shutdown_ioa(struct ipr_cmnd *ipr_cmd)
 	int rc = IPR_RC_JOB_CONTINUE;
 
 	ENTER;
-	if (shutdown_type != IPR_SHUTDOWN_NONE && !ioa_cfg->ioa_is_dead) {
+	if (shutdown_type != IPR_SHUTDOWN_NONE &&
+			!ioa_cfg->hrrq[IPR_INIT_HRRQ].ioa_is_dead) {
 		ipr_cmd->ioarcb.res_handle = cpu_to_be32(IPR_IOA_RES_HANDLE);
 		ipr_cmd->ioarcb.cmd_pkt.request_type = IPR_RQTYPE_IOACMD;
 		ipr_cmd->ioarcb.cmd_pkt.cdb[0] = IPR_IOA_SHUTDOWN;
@@ -8321,9 +8416,15 @@ static void _ipr_initiate_ioa_reset(struct ipr_ioa_cfg *ioa_cfg,
 				    enum ipr_shutdown_type shutdown_type)
 {
 	struct ipr_cmnd *ipr_cmd;
+	int i;
 
 	ioa_cfg->in_reset_reload = 1;
-	ioa_cfg->allow_cmds = 0;
+	for (i = 0; i < ioa_cfg->hrrq_num; i++) {
+		spin_lock(&ioa_cfg->hrrq[i]._lock);
+		ioa_cfg->hrrq[i].allow_cmds = 0;
+		spin_unlock(&ioa_cfg->hrrq[i]._lock);
+	}
+	wmb();
 	scsi_block_requests(ioa_cfg->host);
 
 	ipr_cmd = ipr_get_free_ipr_cmnd(ioa_cfg);
@@ -8349,7 +8450,9 @@ static void _ipr_initiate_ioa_reset(struct ipr_ioa_cfg *ioa_cfg,
 static void ipr_initiate_ioa_reset(struct ipr_ioa_cfg *ioa_cfg,
 				   enum ipr_shutdown_type shutdown_type)
 {
-	if (ioa_cfg->ioa_is_dead)
+	int i;
+
+	if (ioa_cfg->hrrq[IPR_INIT_HRRQ].ioa_is_dead)
 		return;
 
 	if (ioa_cfg->in_reset_reload) {
@@ -8364,7 +8467,12 @@ static void ipr_initiate_ioa_reset(struct ipr_ioa_cfg *ioa_cfg,
 			"IOA taken offline - error recovery failed\n");
 
 		ioa_cfg->reset_retries = 0;
-		ioa_cfg->ioa_is_dead = 1;
+		for (i = 0; i < ioa_cfg->hrrq_num; i++) {
+			spin_lock(&ioa_cfg->hrrq[i]._lock);
+			ioa_cfg->hrrq[i].ioa_is_dead = 1;
+			spin_unlock(&ioa_cfg->hrrq[i]._lock);
+		}
+		wmb();
 
 		if (ioa_cfg->in_ioa_bringdown) {
 			ioa_cfg->reset_cmd = NULL;
@@ -8396,8 +8504,16 @@ static void ipr_initiate_ioa_reset(struct ipr_ioa_cfg *ioa_cfg,
  */
 static int ipr_reset_freeze(struct ipr_cmnd *ipr_cmd)
 {
+	struct ipr_ioa_cfg *ioa_cfg = ipr_cmd->ioa_cfg;
+	int i;
+
 	/* Disallow new interrupts, avoid loop */
-	ipr_cmd->ioa_cfg->allow_interrupts = 0;
+	for (i = 0; i < ioa_cfg->hrrq_num; i++) {
+		spin_lock(&ioa_cfg->hrrq[i]._lock);
+		ioa_cfg->hrrq[i].allow_interrupts = 0;
+		spin_unlock(&ioa_cfg->hrrq[i]._lock);
+	}
+	wmb();
 	list_add_tail(&ipr_cmd->queue, &ipr_cmd->hrrq->hrrq_pending_q);
 	ipr_cmd->done = ipr_reset_ioa_job;
 	return IPR_RC_JOB_RETURN;
@@ -8455,13 +8571,19 @@ static void ipr_pci_perm_failure(struct pci_dev *pdev)
 {
 	unsigned long flags = 0;
 	struct ipr_ioa_cfg *ioa_cfg = pci_get_drvdata(pdev);
+	int i;
 
 	spin_lock_irqsave(ioa_cfg->host->host_lock, flags);
 	if (ioa_cfg->sdt_state == WAIT_FOR_DUMP)
 		ioa_cfg->sdt_state = ABORT_DUMP;
 	ioa_cfg->reset_retries = IPR_NUM_RESET_RELOAD_RETRIES;
 	ioa_cfg->in_ioa_bringdown = 1;
-	ioa_cfg->allow_cmds = 0;
+	for (i = 0; i < ioa_cfg->hrrq_num; i++) {
+		spin_lock(&ioa_cfg->hrrq[i]._lock);
+		ioa_cfg->hrrq[i].allow_cmds = 0;
+		spin_unlock(&ioa_cfg->hrrq[i]._lock);
+	}
+	wmb();
 	ipr_initiate_ioa_reset(ioa_cfg, IPR_SHUTDOWN_NONE);
 	spin_unlock_irqrestore(ioa_cfg->host->host_lock, flags);
 }
@@ -8522,7 +8644,7 @@ static int ipr_probe_ioa_part2(struct ipr_ioa_cfg *ioa_cfg)
 	wait_event(ioa_cfg->reset_wait_q, !ioa_cfg->in_reset_reload);
 	spin_lock_irqsave(ioa_cfg->host->host_lock, host_lock_flags);
 
-	if (ioa_cfg->ioa_is_dead) {
+	if (ioa_cfg->hrrq[IPR_INIT_HRRQ].ioa_is_dead) {
 		rc = -EIO;
 	} else if (ipr_invalid_adapter(ioa_cfg)) {
 		if (!ipr_testmode)
@@ -8629,10 +8751,13 @@ static void ipr_free_all_resources(struct ipr_ioa_cfg *ioa_cfg)
 	} else
 		free_irq(pdev->irq, &ioa_cfg->hrrq[0]);
 
-	if (ioa_cfg->intr_flag == IPR_USE_MSI)
+	if (ioa_cfg->intr_flag == IPR_USE_MSI) {
 		pci_disable_msi(pdev);
-	else if (ioa_cfg->intr_flag == IPR_USE_MSIX)
+		ioa_cfg->intr_flag &= ~IPR_USE_MSI;
+	} else if (ioa_cfg->intr_flag == IPR_USE_MSIX) {
 		pci_disable_msix(pdev);
+		ioa_cfg->intr_flag &= ~IPR_USE_MSIX;
+	}
 
 	iounmap(ioa_cfg->hdw_dma_regs);
 	pci_release_regions(pdev);
@@ -8800,6 +8925,11 @@ static int ipr_alloc_mem(struct ipr_ioa_cfg *ioa_cfg)
 	for (i = 0; i < ioa_cfg->hrrq_num; i++) {
 		INIT_LIST_HEAD(&ioa_cfg->hrrq[i].hrrq_free_q);
 		INIT_LIST_HEAD(&ioa_cfg->hrrq[i].hrrq_pending_q);
+		spin_lock_init(&ioa_cfg->hrrq[i]._lock);
+		if (i == 0)
+			ioa_cfg->hrrq[i].lock = ioa_cfg->host->host_lock;
+		else
+			ioa_cfg->hrrq[i].lock = &ioa_cfg->hrrq[i]._lock;
 	}
 
 	if (ipr_alloc_cmd_blks(ioa_cfg))
@@ -9154,9 +9284,9 @@ static int ipr_test_msi(struct ipr_ioa_cfg *ioa_cfg, struct pci_dev *pdev)
 	writel(IPR_PCII_IO_DEBUG_ACKNOWLEDGE, ioa_cfg->regs.sense_interrupt_reg32);
 	int_reg = readl(ioa_cfg->regs.sense_interrupt_reg);
 	wait_event_timeout(ioa_cfg->msi_wait_q, ioa_cfg->msi_received, HZ);
+	spin_lock_irqsave(ioa_cfg->host->host_lock, lock_flags);
 	ipr_mask_and_clear_interrupts(ioa_cfg, ~IPR_PCII_IOA_TRANS_TO_OPER);
 
-	spin_lock_irqsave(ioa_cfg->host->host_lock, lock_flags);
 	if (!ioa_cfg->msi_received) {
 		/* MSI test failed */
 		dev_info(&pdev->dev, "MSI test failed.  Falling back to LSI.\n");
@@ -9189,6 +9319,7 @@ static int ipr_probe_ioa(struct pci_dev *pdev,
 	void __iomem *ipr_regs;
 	int rc = PCIBIOS_SUCCESSFUL;
 	volatile u32 mask, uproc, interrupts;
+	unsigned long lock_flags;
 
 	ENTER;
 
@@ -9291,10 +9422,10 @@ static int ipr_probe_ioa(struct pci_dev *pdev,
 	}
 
 	if (ioa_cfg->ipr_chip->intr_type == IPR_USE_MSI &&
-		ipr_enable_msix(ioa_cfg) == 0)
+			ipr_enable_msix(ioa_cfg) == 0)
 		ioa_cfg->intr_flag = IPR_USE_MSIX;
 	else if (ioa_cfg->ipr_chip->intr_type == IPR_USE_MSI &&
-		ipr_enable_msi(ioa_cfg) == 0)
+			ipr_enable_msi(ioa_cfg) == 0)
 		ioa_cfg->intr_flag = IPR_USE_MSI;
 	else {
 		ioa_cfg->intr_flag = IPR_USE_LSI;
@@ -9380,7 +9511,9 @@ static int ipr_probe_ioa(struct pci_dev *pdev,
 	if (interrupts & IPR_PCII_IOA_UNIT_CHECKED)
 		ioa_cfg->ioa_unit_checked = 1;
 
+	spin_lock_irqsave(ioa_cfg->host->host_lock, lock_flags);
 	ipr_mask_and_clear_interrupts(ioa_cfg, ~IPR_PCII_IOA_TRANS_TO_OPER);
+	spin_unlock_irqrestore(ioa_cfg->host->host_lock, lock_flags);
 
 	if (ioa_cfg->intr_flag == IPR_USE_MSI
 			|| ioa_cfg->intr_flag == IPR_USE_MSIX) {
@@ -9767,7 +9900,7 @@ static int ipr_halt(struct notifier_block *nb, ulong event, void *buf)
 
 	list_for_each_entry(ioa_cfg, &ipr_ioa_head, queue) {
 		spin_lock_irqsave(ioa_cfg->host->host_lock, flags);
-		if (!ioa_cfg->allow_cmds) {
+		if (!ioa_cfg->hrrq[IPR_INIT_HRRQ].allow_cmds) {
 			spin_unlock_irqrestore(ioa_cfg->host->host_lock, flags);
 			continue;
 		}