Re: [PATCH v4] SCSI host lock push-down

[Jeff Garzik <jeff@garzik.org>]

On 11/13/2010 12:16 AM, Jeff Garzik wrote:
> diff --git a/drivers/scsi/arm/fas216.c b/drivers/scsi/arm/fas216.c
> index 9e71ac6..f9581a3 100644
> --- a/drivers/scsi/arm/fas216.c
> +++ b/drivers/scsi/arm/fas216.c
> @@ -2198,7 +2198,7 @@ no_command:
>    * Returns: 0 on success, else error.
>    * Notes: io_request_lock is held, interrupts are disabled.
>    */
> -int fas216_queue_command(struct scsi_cmnd *SCpnt,
> +static int fas216_queue_command_lck(struct scsi_cmnd *SCpnt,
>   			 void (*done)(struct scsi_cmnd *))
>   {
>   	FAS216_Info *info = (FAS216_Info *)SCpnt->device->host->hostdata;
> @@ -2240,6 +2240,8 @@ int fas216_queue_command(struct scsi_cmnd *SCpnt,
>   	return result;
>   }
>
> +DEF_SCSI_QCMD(fas216_queue_command)
> +
>   /**
>    * fas216_internal_done - trigger restart of a waiting thread in fas216_noqueue_command
>    * @SCpnt: Command to wake
> @@ -2263,7 +2265,7 @@ static void fas216_internal_done(struct scsi_cmnd *SCpnt)
>    * Returns: scsi result code.
>    * Notes: io_request_lock is held, interrupts are disabled.
>    */
> -int fas216_noqueue_command(struct scsi_cmnd *SCpnt,
> +static int fas216_noqueue_command_lck(struct scsi_cmnd *SCpnt,
>   			   void (*done)(struct scsi_cmnd *))
>   {
>   	FAS216_Info *info = (FAS216_Info *)SCpnt->device->host->hostdata;
> @@ -2277,7 +2279,7 @@ int fas216_noqueue_command(struct scsi_cmnd *SCpnt,
>   	BUG_ON(info->scsi.irq != NO_IRQ);
>
>   	info->internal_done = 0;
> -	fas216_queue_command(SCpnt, fas216_internal_done);
> +	fas216_queue_command(SCpnt->device->host, SCpnt, fas216_internal_done);
>
>   	/*
>   	 * This wastes time, since we can't return until the command is


hmmmm.

Linus's suggestion to add Scsi_Host to queuecommand seems to have had a 
helpful side effect, exposing a minor bug present in nab's (v1), Andi's 
(v2), my v3 and v4 push-down patches:

A few odd ducks, such as fas16 or ancient megaraid, call their own 
->queuecommand directly.  None of the major SCSI drivers do this, 
thankfully.

The host lock push-down patch should be calling the _lck function in 
those rare cases, otherwise we've created a deadlock.

Time for v5...

	Jeff