After calling the queuecommand callback, say fc_queuecommand in libfc, scsi cmnd could be not delivered to LLD, and done function could also be not issued. Thus there could be nobody responsible for completing the waiter later.
What is more, the subsequent computation based on the timeleft after wait is unreasonable, same for the abort operation since scmd is not sent to underlying device. Fix is to jump wait if no completion will occur. The patch is prepared according to the implementation of fc_queuecommand, and there is room for fix for other cases. Signed-off-by: Hillf Danton <[email protected]> --- --- a/drivers/scsi/scsi_error.c 2010-11-01 19:54:12.000000000 +0800 +++ b/drivers/scsi/scsi_error.c 2010-11-30 21:31:46.000000000 +0800 @@ -790,9 +790,19 @@ static int scsi_send_eh_cmnd(struct scsi spin_lock_irqsave(shost->host_lock, flags); scsi_log_send(scmd); - shost->hostt->queuecommand(scmd, scsi_eh_done); + rtn = shost->hostt->queuecommand(scmd, scsi_eh_done); spin_unlock_irqrestore(shost->host_lock, flags); + if (rtn != 0) { + /* + * scmd could not send to sdev, and + * scsi_eh_done() not called, so + * skip wait for completing. + */ + rtn = NEEDS_RETRY; + goto out; + } + timeleft = wait_for_completion_timeout(&done, timeout); shost->eh_action = NULL; @@ -831,7 +841,7 @@ static int scsi_send_eh_cmnd(struct scsi scsi_abort_eh_cmnd(scmd); rtn = FAILED; } - + out: scsi_eh_restore_cmnd(scmd, &ses); return rtn; } _______________________________________________ devel mailing list [email protected] http://www.open-fcoe.org/mailman/listinfo/devel
