Re: [linux-iscsi-devel] [PATCH] fix irq trace problem reported

[Mike C. <mi...@us...>]

I think I may have been smoking crack when I debugged this. The fix is 
no good - although removing the times_out handler is probably good in 
the long run.




Mike Christie wrote:
> This is to fix the irq problem someone reported. Basically we are 
> running from a soft irq but when we return EH_HANDLED we hit the code in 
> __scsi_done which does not like that we are running from one.
> 
> This code is actually not buying us anything. In some cases we do not 
> have to fire the eh. the eh detects the same situation so it does not 
> end up sending TMFs so it saves the eh thread from having to run and the 
> host/target from being quiesced. In the end the host block/unblock code 
> would catch the case where this is useful best, but I do not think 
> anyone wants to port that stuff around.
> 
> Did not fix the locking comments as they apply to the rpelacement 
> timeout code still. This patch is for the stable distro release btw. In 
> the other version I will fix the locks since it will use the 
> block/unblock code when that patch is updated. I am still trying to see 
> how to use the block and unblock code in open-iscsi to fix some problems 
> though.
> 
> 
> ------------------------------------------------------------------------
> 
> --- linux-iscsi/driver/iscsi-initiator.c	2005-03-14 22:33:41.000000000 -0800
> +++ linux-iscsi.work/driver/iscsi-initiator.c	2005-03-15 14:27:45.000000000 -0800
> @@ -134,55 +134,6 @@ iscsi_slave_configure(struct scsi_device
>  	return 0;
>  }
>  
> -/**
> - * iscsi_eh_timed_out - handle scsi command timeouts
> - * sc: scsi command
> - *
> - * Note:
> - *    In the normal case where the session is ok or the command was queued
> - *    when the session was down (and it never came up but did not yet timeout
> - *    - TODO use device block/unblock), we will fail the command here if
> - *    we can. For all other race cases where for example the session changed
> - *    state at the same time the command timed out but we had the task lock
> - *    we just ask for more time since those events will clean themselves up.
> - **/
> -static enum scsi_eh_timer_return
> -iscsi_eh_timed_out(struct scsi_cmnd *sc)
> -{
> -	struct Scsi_Host *shost = sc->device->host;
> -	struct iscsi_session *session = (struct iscsi_session *)shost->hostdata;
> -	struct iscsi_task *task;
> -	enum scsi_eh_timer_return ret = EH_RESET_TIMER;
> -	int result = DID_BUS_BUSY << 16;
> -
> -	spin_lock(&session->task_lock);
> -
> -	task = (struct iscsi_task *)sc->SCp.ptr;
> -	if (!task) {
> -		/*
> -		 * completed while timer was firing
> -		 */
> -		ret = EH_HANDLED;
> -		goto done;
> -	}
> -
> -	if (test_bit(SESSION_ESTABLISHED, &session->control_bits))
> -		ret = EH_NOT_HANDLED;
> -	else if (test_bit(SESSION_REPLACEMENT_TIMEDOUT, &session->control_bits))
> -		result = DID_NO_CONNECT << 16;
> -
> -	if (task->itt == ISCSI_RSVD_TASK_TAG) {
> -		sc->result = result;
> -		__iscsi_complete_task(task);
> -		ret = EH_HANDLED;
> -	}
> -
> - done:
> -	spin_unlock(&session->task_lock);
> -
> -	return ret;
> -}
> -
>  static int
>  iscsi_eh_abort(struct scsi_cmnd *sc)
>  {
> @@ -451,7 +402,6 @@ static struct scsi_host_template iscsi_d
>  	.proc_name = ISCSI_PROC_NAME,
>  	.module = THIS_MODULE,
>  	.queuecommand = iscsi_queuecommand,
> -	.eh_timed_out = iscsi_eh_timed_out,
>  	.eh_abort_handler = iscsi_eh_abort,
>  	.eh_device_reset_handler = iscsi_eh_device_reset,
>  	.eh_host_reset_handler = iscsi_eh_host_reset,