A live partition migration (LPM) results in a CRQ disconnect similar to
a hard reset. In this LPM case the hypervisor moslty perserves the CRQ
transport such that it simply needs to be reenabled. However, the
capabilities may have changed such as fewer channels, or no channels at
all. Further, its possible that there may be sub-CRQ support, but no
channel support. The CRQ reenable path currently doesn't take any of
this into consideration.
For simpilicty release and reinitialize sub-CRQs during reenable, and
set do_enquiry and using_channels with the appropriate values to trigger
channel renegotiation.
Signed-off-by: Tyrel Datwyler <[email protected]>
---
drivers/scsi/ibmvscsi/ibmvfc.c | 13 +++++++++++++
1 file changed, 13 insertions(+)
diff --git a/drivers/scsi/ibmvscsi/ibmvfc.c b/drivers/scsi/ibmvscsi/ibmvfc.c
index 4ac2c442e1e2..9ae6be56e375 100644
--- a/drivers/scsi/ibmvscsi/ibmvfc.c
+++ b/drivers/scsi/ibmvscsi/ibmvfc.c
@@ -903,6 +903,9 @@ static int ibmvfc_reenable_crq_queue(struct ibmvfc_host *vhost)
{
int rc = 0;
struct vio_dev *vdev = to_vio_dev(vhost->dev);
+ unsigned long flags;
+
+ ibmvfc_release_sub_crqs(vhost);
/* Re-enable the CRQ */
do {
@@ -914,6 +917,16 @@ static int ibmvfc_reenable_crq_queue(struct ibmvfc_host *vhost)
if (rc)
dev_err(vhost->dev, "Error enabling adapter (rc=%d)\n", rc);
+ spin_lock_irqsave(vhost->host->host_lock, flags);
+ spin_lock(vhost->crq.q_lock);
+ vhost->do_enquiry = 1;
+ vhost->using_channels = 0;
+
+ ibmvfc_init_sub_crqs(vhost);
+
+ spin_unlock(vhost->crq.q_lock);
+ spin_unlock_irqrestore(vhost->host->host_lock, flags);
+
return rc;
}
--
2.27.0
On 2/25/21 12:48 PM, Tyrel Datwyler wrote:
> A live partition migration (LPM) results in a CRQ disconnect similar to
> a hard reset. In this LPM case the hypervisor moslty perserves the CRQ
> transport such that it simply needs to be reenabled. However, the
> capabilities may have changed such as fewer channels, or no channels at
> all. Further, its possible that there may be sub-CRQ support, but no
> channel support. The CRQ reenable path currently doesn't take any of
> this into consideration.
>
> For simpilicty release and reinitialize sub-CRQs during reenable, and
> set do_enquiry and using_channels with the appropriate values to trigger
> channel renegotiation.
>
This should of had the same fixes tag as patch 2.
fixes: faacf8c5f1d5 ("ibmvfc: add alloc/dealloc routines for SCSI Sub-CRQ Channels")
> Signed-off-by: Tyrel Datwyler <[email protected]>
> ---
> drivers/scsi/ibmvscsi/ibmvfc.c | 13 +++++++++++++
> 1 file changed, 13 insertions(+)
>
> diff --git a/drivers/scsi/ibmvscsi/ibmvfc.c b/drivers/scsi/ibmvscsi/ibmvfc.c
> index 4ac2c442e1e2..9ae6be56e375 100644
> --- a/drivers/scsi/ibmvscsi/ibmvfc.c
> +++ b/drivers/scsi/ibmvscsi/ibmvfc.c
> @@ -903,6 +903,9 @@ static int ibmvfc_reenable_crq_queue(struct ibmvfc_host *vhost)
> {
> int rc = 0;
> struct vio_dev *vdev = to_vio_dev(vhost->dev);
> + unsigned long flags;
> +
> + ibmvfc_release_sub_crqs(vhost);
>
> /* Re-enable the CRQ */
> do {
> @@ -914,6 +917,16 @@ static int ibmvfc_reenable_crq_queue(struct ibmvfc_host *vhost)
> if (rc)
> dev_err(vhost->dev, "Error enabling adapter (rc=%d)\n", rc);
>
> + spin_lock_irqsave(vhost->host->host_lock, flags);
> + spin_lock(vhost->crq.q_lock);
> + vhost->do_enquiry = 1;
> + vhost->using_channels = 0;
> +
> + ibmvfc_init_sub_crqs(vhost);
> +
> + spin_unlock(vhost->crq.q_lock);
> + spin_unlock_irqrestore(vhost->host->host_lock, flags);
> +
> return rc;
> }
>
>
On 2/25/21 2:48 PM, Tyrel Datwyler wrote:
> A live partition migration (LPM) results in a CRQ disconnect similar to
> a hard reset. In this LPM case the hypervisor moslty perserves the CRQ
> transport such that it simply needs to be reenabled. However, the
> capabilities may have changed such as fewer channels, or no channels at
> all. Further, its possible that there may be sub-CRQ support, but no
> channel support. The CRQ reenable path currently doesn't take any of
> this into consideration.
>
> For simpilicty release and reinitialize sub-CRQs during reenable, and
> set do_enquiry and using_channels with the appropriate values to trigger
> channel renegotiation.
>
> Signed-off-by: Tyrel Datwyler <[email protected]>
> ---
> drivers/scsi/ibmvscsi/ibmvfc.c | 13 +++++++++++++
> 1 file changed, 13 insertions(+)
>
> diff --git a/drivers/scsi/ibmvscsi/ibmvfc.c b/drivers/scsi/ibmvscsi/ibmvfc.c
> index 4ac2c442e1e2..9ae6be56e375 100644
> --- a/drivers/scsi/ibmvscsi/ibmvfc.c
> +++ b/drivers/scsi/ibmvscsi/ibmvfc.c
> @@ -903,6 +903,9 @@ static int ibmvfc_reenable_crq_queue(struct ibmvfc_host *vhost)
> {
> int rc = 0;
> struct vio_dev *vdev = to_vio_dev(vhost->dev);
> + unsigned long flags;
> +
> + ibmvfc_release_sub_crqs(vhost);
>
> /* Re-enable the CRQ */
> do {
> @@ -914,6 +917,16 @@ static int ibmvfc_reenable_crq_queue(struct ibmvfc_host *vhost)
> if (rc)
> dev_err(vhost->dev, "Error enabling adapter (rc=%d)\n", rc);
>
> + spin_lock_irqsave(vhost->host->host_lock, flags);
> + spin_lock(vhost->crq.q_lock);
> + vhost->do_enquiry = 1;
> + vhost->using_channels = 0;
> +
> + ibmvfc_init_sub_crqs(vhost);
> +
> + spin_unlock(vhost->crq.q_lock);
> + spin_unlock_irqrestore(vhost->host->host_lock, flags);
ibmvfc_init_sub_crqs can sleep, for multiple reasons, so you can't hold
a lock when you call it. There is a GFP_KERNEL allocation in it, and the
patch before this one adds an msleep in an error path.
Thanks,
Brian
--
Brian King
Power Linux I/O
IBM Linux Technology Center