Commit db422318 authored by Menny Hamburger's avatar Menny Hamburger Committed by James Bottomley

[SCSI] scsi_dh: propagate SCSI device deletion

Currently, when scsi_dh_activate() returns with an error
(e.g. SCSI_DH_NOSYS) the activate_complete callback is not called and
the error is not propagated to DM mpath.

When a SCSI device attached to a device handler is deleted, userland
processes currently performing I/O on the device will have their I/O
hang forever.

- Set SCSI_DH_NOSYS error when the handler is in the process of being
  deleted (e.g. the SCSI device is in a SDEV_CANCEL or SDEV_DEL state).

- Set SCSI_DH_DEV_OFFLINED error when device is in SDEV_OFFLINE state.

- Call the activate_complete callback function directly from
  scsi_dh_activate if an error has been set (when either the scsi_dh
  internal data has already been deleted or is in the process of being
  deleted).

The patch was tested in an iSCSI environment, RDAC H/W handler and
multipath.  In the following reproduction process, dd will I/O hang
forever and the only way to release it will be to reboot the machine:
1) Perform I/O on a multipath device:
    dd if=/dev/dm-0 of=/dev/zero bs=8k count=1000000 &
2) Delete all slave SCSI devices contained in the mpath device:
   I)  In an iSCSI environment, the easiest way to do this is by
   stopping iSCSI:
       /etc/init.d/iscsi stop
   II) Another way to delete the devices is by applying the following
   bash scriptlet:
       dm_devs=$(ls /sys/block/ | grep dm- | xargs)
       for dm_dev in $dm_devs; do
         devices=$(ls /sys/block/$dm_dev/slaves)
         for device in $devices; do
            echo 1 > /sys/block/$device/device/delete
         done
       done

NOTE: when DM mpath's fail_path uses blk_abort_queue this scsi_dh change
isn't strictly required.  However, DM mpath's call to blk_abort_queue
will soon be reverted because it has proven to be unsafe due to a race
(between blk_abort_queue and scsi_request_fn) that can lead to list
corruption.  Therefore we cannot rely on blk_abort_queue via fail_path,
but even if we could this scsi_dh change is still preferrable.
Signed-off-by: default avatarMenny Hamburger <Menny_Hamburger@Dell.com>
Signed-off-by: default avatarMike Snitzer <snitzer@redhat.com>
Reviewed-by: default avatarBabu Moger <babu.moger@lsi.com>
Signed-off-by: default avatarJames Bottomley <James.Bottomley@suse.de>
parent 35dd3039
......@@ -442,12 +442,19 @@ int scsi_dh_activate(struct request_queue *q, activate_complete fn, void *data)
sdev = q->queuedata;
if (sdev && sdev->scsi_dh_data)
scsi_dh = sdev->scsi_dh_data->scsi_dh;
if (!scsi_dh || !get_device(&sdev->sdev_gendev))
if (!scsi_dh || !get_device(&sdev->sdev_gendev) ||
sdev->sdev_state == SDEV_CANCEL ||
sdev->sdev_state == SDEV_DEL)
err = SCSI_DH_NOSYS;
if (sdev->sdev_state == SDEV_OFFLINE)
err = SCSI_DH_DEV_OFFLINED;
spin_unlock_irqrestore(q->queue_lock, flags);
if (err)
if (err) {
if (fn)
fn(data, err);
return err;
}
if (scsi_dh->activate)
err = scsi_dh->activate(sdev, fn, data);
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment