Skip to contentSkip to conversation

ibmvscsi: Speed up kexec boot

PATCH In review Anton Blanchard 1 participant 2 messages Dec 23, 2013 – Mar 28, 2014

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ibmvscsi: Speed up kexec boot

During kexec boot we call ibmvscsi_release_crq_queue() to tear down the old CRQ from the previous kernel. ibmvscsi_release_crq_queue() does this by calling H_FREE_CRQ. The hypervisor breaks this work down so as limit the time spent in any one hcall, so we have to loop until complete.

At the moment we delay 0.1 seconds between every H_FREE_CRQ call. I see us calling H_FREE_CRQ 60 times on my box which means we delayed a total of 6 seconds.

This delay is overkill, we are free to handle all busy return codes from the hypervisor as we see fit. Cut the delay down to something more reasonable - 1 ms. This improves my boot time by 6 seconds.

Fix the other places we have arbitrary 100 ms delays.

---

drivers/scsi/ibmvscsi/ibmvscsi.c +4 -4
index fa76440..ed7b52e 100644
@@ -159,7 +159,7 @@ static void ibmvscsi_release_crq_queue(struct crq_queue *queue,
 	tasklet_kill(&hostdata->srp_task);
 	do {
 		if (rc)
-			msleep(100);
+			msleep(1);
 		rc = plpar_hcall_norets(H_FREE_CRQ, vdev->unit_address);
 	} while ((rc == H_BUSY) || (H_IS_LONG_BUSY(rc)));
 	dma_unmap_single(hostdata->dev,
@@ -292,7 +292,7 @@ static int ibmvscsi_reset_crq_queue(struct crq_queue *queue,
 	/* Close the CRQ */
 	do {
 		if (rc)
-			msleep(100);
+			msleep(1);
 		rc = plpar_hcall_norets(H_FREE_CRQ, vdev->unit_address);
 	} while ((rc == H_BUSY) || (H_IS_LONG_BUSY(rc)));
 
@@ -392,7 +392,7 @@ static int ibmvscsi_init_crq_queue(struct crq_queue *queue,
 	rc = 0;
 	do {
 		if (rc)
-			msleep(100);
+			msleep(1);
 		rc = plpar_hcall_norets(H_FREE_CRQ, vdev->unit_address);
 	} while ((rc == H_BUSY) || (H_IS_LONG_BUSY(rc)));
       reg_crq_failed:
@@ -420,7 +420,7 @@ static int ibmvscsi_reenable_crq_queue(struct crq_queue *queue,
 	/* Re-enable the CRQ */
 	do {
 		if (rc)
-			msleep(100);
+			msleep(1);
 		rc = plpar_hcall_norets(H_ENABLE_CRQ, vdev->unit_address);
 	} while ((rc == H_IN_PROGRESS) || (rc == H_BUSY) || (H_IS_LONG_BUSY(rc)));
 
#1 lore raw

Hi,

Quoted 10 lines

During kexec boot we call ibmvscsi_release_crq_queue() to tear down the old CRQ from the previous kernel. ibmvscsi_release_crq_queue() does this by calling H_FREE_CRQ. The hypervisor breaks this work down so as limit the time spent in any one hcall, so we have to loop until complete.

At the moment we delay 0.1 seconds between every H_FREE_CRQ call. I see us calling H_FREE_CRQ 60 times on my box which means we delayed a total of 6 seconds.

Quoted 5 lines

This delay is overkill, we are free to handle all busy return codes from the hypervisor as we see fit. Cut the delay down to something more reasonable - 1 ms. This improves my boot time by 6 seconds.

Fix the other places we have arbitrary 100 ms delays.

Checking in on this patch, 6 seconds is a decent chunk of our boot time on KVM.

Anton

Quoted 43 lines

---

(quoted diff) +0 -0

b/drivers/scsi/ibmvscsi/ibmvscsi.c index fa76440..ed7b52e 100644

drivers/scsi/ibmvscsi/ibmvscsi.c +1 -1
@@ -159,7 +159,7 @@ static void ibmvscsi_release_crq_queue(struct
 crq_queue *queue, tasklet_kill(&hostdata->srp_task);
 	do {
 		if (rc)
-			msleep(100);
+			msleep(1);
 		rc = plpar_hcall_norets(H_FREE_CRQ,
 vdev->unit_address); } while ((rc == H_BUSY) || (H_IS_LONG_BUSY(rc)));
 	dma_unmap_single(hostdata->dev,
@@ -292,7 +292,7 @@ static int ibmvscsi_reset_crq_queue(struct

crq_queue *queue, /* Close the CRQ */

drivers/scsi/ibmvscsi/ibmvscsi.c +1 -1
 	do {
 		if (rc)
-			msleep(100);
+			msleep(1);
 		rc = plpar_hcall_norets(H_FREE_CRQ,

vdev->unit_address); } while ((rc == H_BUSY) || (H_IS_LONG_BUSY(rc)));

drivers/scsi/ibmvscsi/ibmvscsi.c +1 -1
@@ -392,7 +392,7 @@ static int ibmvscsi_init_crq_queue(struct
 crq_queue *queue, rc = 0;
 	do {
 		if (rc)
-			msleep(100);
+			msleep(1);
 		rc = plpar_hcall_norets(H_FREE_CRQ,
 vdev->unit_address); } while ((rc == H_BUSY) || (H_IS_LONG_BUSY(rc)));
       reg_crq_failed:
@@ -420,7 +420,7 @@ static int ibmvscsi_reenable_crq_queue(struct

crq_queue *queue, /* Re-enable the CRQ */

drivers/scsi/ibmvscsi/ibmvscsi.c +1 -1
 	do {
 		if (rc)
-			msleep(100);
+			msleep(1);
 		rc = plpar_hcall_norets(H_ENABLE_CRQ,

vdev->unit_address); } while ((rc == H_IN_PROGRESS) || (rc == H_BUSY) || (H_IS_LONG_BUSY(rc)));