On Tue, Dec 08, 2015 at 08:37:33AM +0100, Hannes Reinecke wrote:
> When we receive a unit attention code of 'ALUA state changed'
> we should recheck the state, as it might be due to an implicit
> ALUA state transition. This allows us to return NEEDS_RETRY
> instead of ADD_TO_MLQUEUE, allowing to terminate the retries
> after a certain time.
> At the same time a workqueue item might already be queued, which
> should be started immediately to avoid any delays.
>
> Signed-off-by: Hannes Reinecke <hare@xxxxxxx>
> ---
> drivers/scsi/device_handler/scsi_dh_alua.c | 58 ++++++++++++++++++++++++------
> 1 file changed, 47 insertions(+), 11 deletions(-)
>
> diff --git a/drivers/scsi/device_handler/scsi_dh_alua.c b/drivers/scsi/device_handler/scsi_dh_alua.c
> index 525449f..04a3a543 100644
> --- a/drivers/scsi/device_handler/scsi_dh_alua.c
> +++ b/drivers/scsi/device_handler/scsi_dh_alua.c
> @@ -121,7 +121,8 @@ struct alua_queue_data {
> static void alua_rtpg_work(struct work_struct *work);
> static void alua_rtpg_queue(struct alua_port_group *pg,
> struct scsi_device *sdev,
> - struct alua_queue_data *qdata);
> + struct alua_queue_data *qdata, bool force);
> +static void alua_check(struct scsi_device *sdev, bool force);
>
> static void release_port_group(struct kref *kref)
> {
> @@ -386,7 +387,7 @@ static int alua_check_vpd(struct scsi_device *sdev, struct alua_dh_data *h,
> rcu_assign_pointer(h->pg, pg);
> pg_found = true;
> }
> - alua_rtpg_queue(h->pg, sdev, NULL);
> + alua_rtpg_queue(h->pg, sdev, NULL, true);
> spin_unlock(&h->pg_lock);
>
> if (pg_found)
> @@ -427,18 +428,24 @@ static int alua_check_sense(struct scsi_device *sdev,
> {
> switch (sense_hdr->sense_key) {
> case NOT_READY:
> - if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0a)
> + if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0a) {
> /*
> * LUN Not Accessible - ALUA state transition
> */
> - return ADD_TO_MLQUEUE;
> + alua_check(sdev, false);
> + return NEEDS_RETRY;
> + }
> break;
> case UNIT_ATTENTION:
> - if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00)
> + if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00) {
> /*
> - * Power On, Reset, or Bus Device Reset, just retry.
> + * Power On, Reset, or Bus Device Reset.
> + * Might have obscured a state transition,
> + * so schedule a recheck.
> */
> + alua_check(sdev, true);
> return ADD_TO_MLQUEUE;
> + }
> if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x04)
> /*
> * Device internal reset
> @@ -449,16 +456,20 @@ static int alua_check_sense(struct scsi_device *sdev,
> * Mode Parameters Changed
> */
> return ADD_TO_MLQUEUE;
> - if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x06)
> + if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x06) {
> /*
> * ALUA state changed
> */
> + alua_check(sdev, true);
> return ADD_TO_MLQUEUE;
> - if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x07)
> + }
> + if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x07) {
> /*
> * Implicit ALUA state transition failed
> */
> + alua_check(sdev, true);
> return ADD_TO_MLQUEUE;
> + }
> if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x03)
> /*
> * Inquiry data has changed
> @@ -777,7 +788,7 @@ static void alua_rtpg_work(struct work_struct *work)
>
> static void alua_rtpg_queue(struct alua_port_group *pg,
> struct scsi_device *sdev,
> - struct alua_queue_data *qdata)
> + struct alua_queue_data *qdata, bool force)
> {
> int start_queue = 0;
> unsigned long flags;
> @@ -797,7 +808,9 @@ static void alua_rtpg_queue(struct alua_port_group *pg,
> pg->rtpg_sdev = sdev;
> scsi_device_get(sdev);
> start_queue = 1;
> - }
> + } else if (!(pg->flags & ALUA_PG_RUN_RTPG) && force)
> + start_queue = 1;
> +
> spin_unlock_irqrestore(&pg->lock, flags);
>
> if (start_queue &&
> @@ -912,7 +925,7 @@ static int alua_activate(struct scsi_device *sdev,
> kref_get(&pg->kref);
> rcu_read_unlock();
>
> - alua_rtpg_queue(pg, sdev, qdata);
> + alua_rtpg_queue(pg, sdev, qdata, true);
> kref_put(&pg->kref, release_port_group);
> out:
> if (fn)
> @@ -921,6 +934,29 @@ out:
> }
>
> /*
> + * alua_check - check path status
> + * @sdev: device on the path to be checked
> + *
> + * Check the device status
> + */
> +static void alua_check(struct scsi_device *sdev, bool force)
> +{
> + struct alua_dh_data *h = sdev->handler_data;
> + struct alua_port_group *pg;
> +
> + rcu_read_lock();
> + pg = rcu_dereference(h->pg);
> + if (!pg) {
> + rcu_read_unlock();
> + return;
> + }
> + kref_get(&pg->kref);
What protects us from pg->kref beeing released? I think the whole
refcounting scheme needs an audit to see where kref_get is called
without synchronization and use kref_get_unless_zero where needed.
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
