Commit 8798a694 authored by Michael Reed's avatar Michael Reed Committed by James Bottomley
Browse files

[SCSI] scsi_transport_fc: remove invalid BUG_ON 



I was doing some large lun count testing with 2.6.31 and hit
a BUG_ON() in fc_timeout_deleted_rport(), and it seems like it
should have been just a matter of time before someone did.

It seems invalid to set port_state under lock, then expect it to
remain set after releasing the lock.  Another thread called
fc_remote_port_add() when the lock was released, changing the
port_state.

This patch removes the BUG_ON and moves the test of the
port_state to inside the host_lock.  It's been running for
several weeks now with no ill effect.

Signed-off-by: default avatarMichael Reed <mdr@sgi.com>
Acked-by: default avatarJames Smart <james.smart@emulex.com>
Signed-off-by: default avatarJames Bottomley <James.Bottomley@suse.de>
parent ad630826

drivers/scsi/scsi_transport_fc.c

+42 −26
Original line number Diff line number Diff line
@@ -2384,6 +2384,7 @@ fc_rport_final_delete(struct work_struct *work) 
	struct Scsi_Host *shost = rport_to_shost(rport);

	struct fc_internal *i = to_fc_internal(shost->transportt);

	unsigned long flags;

	int do_callback = 0;



	/*

	 * if a scan is pending, flush the SCSI Host work_q so that
@@ -2422,8 +2423,15 @@ fc_rport_final_delete(struct work_struct *work) 
	 * Avoid this call if we already called it when we preserved the

	 * rport for the binding.

	 */

	spin_lock_irqsave(shost->host_lock, flags);

	if (!(rport->flags & FC_RPORT_DEVLOSS_CALLBK_DONE) &&

	    (i->f->dev_loss_tmo_callbk))

	    (i->f->dev_loss_tmo_callbk)) {

		rport->flags |= FC_RPORT_DEVLOSS_CALLBK_DONE;

		do_callback = 1;

	}

	spin_unlock_irqrestore(shost->host_lock, flags);



	if (do_callback)

		i->f->dev_loss_tmo_callbk(rport);



	fc_bsg_remove(rport->rqst_q);
@@ -2970,6 +2978,7 @@ fc_timeout_deleted_rport(struct work_struct *work) 
	struct fc_internal *i = to_fc_internal(shost->transportt);

	struct fc_host_attrs *fc_host = shost_to_fc_host(shost);

	unsigned long flags;

	int do_callback = 0;



	spin_lock_irqsave(shost->host_lock, flags);
@@ -3035,7 +3044,6 @@ fc_timeout_deleted_rport(struct work_struct *work) 
	rport->roles = FC_PORT_ROLE_UNKNOWN;

	rport->port_state = FC_PORTSTATE_NOTPRESENT;

	rport->flags &= ~FC_RPORT_FAST_FAIL_TIMEDOUT;

	rport->flags |= FC_RPORT_DEVLOSS_CALLBK_DONE;



	/*

	 * Pre-emptively kill I/O rather than waiting for the work queue
@@ -3045,7 +3053,9 @@ fc_timeout_deleted_rport(struct work_struct *work) 
	spin_unlock_irqrestore(shost->host_lock, flags);

	fc_terminate_rport_io(rport);



	BUG_ON(rport->port_state != FC_PORTSTATE_NOTPRESENT);

	spin_lock_irqsave(shost->host_lock, flags);



	if (rport->port_state == FC_PORTSTATE_NOTPRESENT) {	/* still missing */



		/* remove the identifiers that aren't used in the consisting binding */

		switch (fc_host->tgtid_bind_type) {
@@ -3070,15 +3080,21 @@ fc_timeout_deleted_rport(struct work_struct *work) 
		 * went away and didn't come back - we'll remove

		 * all attached scsi devices.

		 */

		rport->flags |= FC_RPORT_DEVLOSS_CALLBK_DONE;

		fc_queue_work(shost, &rport->stgt_delete_work);



		do_callback = 1;

	}



	spin_unlock_irqrestore(shost->host_lock, flags);



	/*

	 * Notify the driver that the rport is now dead. The LLDD will

	 * also guarantee that any communication to the rport is terminated

	 *

	 * Note: we set the CALLBK_DONE flag above to correspond

	 */

	if (i->f->dev_loss_tmo_callbk)

	if (do_callback && i->f->dev_loss_tmo_callbk)

		i->f->dev_loss_tmo_callbk(rport);

}

CRA Git | Maintained and supported by SUSTech CRA and CCSE