Re: [PATCH] dm mpath: delay retry activate_path on SCSI_DH_RETRY

[Chandra Seetharaman <sekharan@us.ibm.com>]

On Fri, 2009-02-20 at 10:33 +0530, Nikanth Karthikesan wrote:
> On Friday 20 February 2009 06:15:29 Chandra Seetharaman wrote:
> > On Thu, 2009-02-19 at 12:40 +0530, Nikanth Karthikesan wrote:
> >
> > <snip>
> >
> > > @@ -431,6 +432,7 @@ static void process_queued_ios(struct work_struct
> > > *work) struct pgpath *pgpath = NULL;
> > >  	unsigned init_required = 0, must_queue = 1;
> > >  	unsigned long flags;
> > > +	unsigned long delay = 0;
> >
> > I do not see the reason for this variable, you can as well put the delay
> > in pg_init_delay and use it directly (and set it to zero after using
> > it) ?
> >
> 
> I missed resetting pg_init_delay to zero after using it. I have attached the
> corrected patch with this. This variable keeps the code cleaner(avoids taking
> m->lock). Also having only a boolean in struct multipath keeps it a bit
> smaller.

Ok. I am fine with your justification.

> 
> off-topic:
> I think struct multipath can be shrunk even further by making various flags
> like pg_init_required, pg_init_in_progress, queue_io, queue_if_no_path,
> saved_queue_if_no_path in to a single variable. Thoughts?

As Konrad mentioned, it is worth the cost of readability ?


<snip>

> SCSI Device Handlers return SCSI_DH_IMM_RETRY if we could retry
> immediately and SCSI_DH_RETRY in cases where it is better to retry
> after some delay.
> 
> Currently we retry immediately irrespective of SCSI_DH_IMM_RETRY and
> SCSI_DH_RETRY.  This patch adds a 2 second delay before retrying to
> activate a device, if it returns SCSI_DH_RETRY.
> 
> Signed-off-by: Nikanth Karthikesan <knikanth@suse.de>

Acked-by: Chandra Seetharaman <sekharan@us.ibm.com>
> 
> ---
> 
> diff --git a/drivers/md/dm-mpath.c b/drivers/md/dm-mpath.c
> index 095f77b..7ddf775 100644
> --- a/drivers/md/dm-mpath.c
> +++ b/drivers/md/dm-mpath.c
> @@ -65,12 +65,13 @@ struct multipath {
>  	spinlock_t lock;
>  
>  	const char *hw_handler_name;
> -	struct work_struct activate_path;
> +	struct delayed_work activate_path;
>  	struct pgpath *pgpath_to_activate;
>  	unsigned nr_priority_groups;
>  	struct list_head priority_groups;
>  	unsigned pg_init_required;	/* pg_init needs calling? */
>  	unsigned pg_init_in_progress;	/* Only one pg_init allowed at once */
> +	unsigned pg_init_delay;		/* delay required before retry? */
>  
>  	unsigned nr_valid_paths;	/* Total number of usable paths */
>  	struct pgpath *current_pgpath;
> @@ -203,7 +204,7 @@ static struct multipath *alloc_multipath(struct dm_target *ti)
>  		m->queue_io = 1;
>  		INIT_WORK(&m->process_queued_ios, process_queued_ios);
>  		INIT_WORK(&m->trigger_event, trigger_event);
> -		INIT_WORK(&m->activate_path, activate_path);
> +		INIT_DELAYED_WORK(&m->activate_path, activate_path);
>  		m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
>  		if (!m->mpio_pool) {
>  			kfree(m);
> @@ -431,6 +432,7 @@ static void process_queued_ios(struct work_struct *work)
>  	struct pgpath *pgpath = NULL;
>  	unsigned init_required = 0, must_queue = 1;
>  	unsigned long flags;
> +	unsigned long delay = 0;
>  
>  	spin_lock_irqsave(&m->lock, flags);
>  
> @@ -452,13 +454,17 @@ static void process_queued_ios(struct work_struct *work)
>  		m->pg_init_required = 0;
>  		m->pg_init_in_progress = 1;
>  		init_required = 1;
> +		if (m->pg_init_delay) {
> +			delay = SCSI_DH_RETRY_DELAY;
> +			m->pg_init_delay = 0;
> +		}
>  	}
>  
>  out:
>  	spin_unlock_irqrestore(&m->lock, flags);
>  
>  	if (init_required)
> -		queue_work(kmpath_handlerd, &m->activate_path);
> +		queue_delayed_work(kmpath_handlerd, &m->activate_path, delay);
>  
>  	if (!must_queue)
>  		dispatch_queued_ios(m);
> @@ -1060,6 +1066,7 @@ static void pg_init_done(struct dm_path *path, int errors)
>  	struct priority_group *pg = pgpath->pg;
>  	struct multipath *m = pg->m;
>  	unsigned long flags;
> +	unsigned delay = 0;
>  
>  	/* device or driver problems */
>  	switch (errors) {
> @@ -1084,8 +1091,11 @@ static void pg_init_done(struct dm_path *path, int errors)
>  		 */
>  		bypass_pg(m, pg, 1);
>  		break;
> -	/* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
> +	/*
> +	 * For SCSI_DH_RETRY we wait before retrying.
> +	 */
>  	case SCSI_DH_RETRY:
> +		delay = 1;
>  	case SCSI_DH_IMM_RETRY:
>  	case SCSI_DH_RES_TEMP_UNAVAIL:
>  		if (pg_init_limit_reached(m, pgpath))
> @@ -1112,6 +1122,7 @@ static void pg_init_done(struct dm_path *path, int errors)
>  	}
>  
>  	m->pg_init_in_progress = 0;
> +	m->pg_init_delay = delay;
>  	queue_work(kmultipathd, &m->process_queued_ios);
>  	spin_unlock_irqrestore(&m->lock, flags);
>  }
> @@ -1120,7 +1131,7 @@ static void activate_path(struct work_struct *work)
>  {
>  	int ret;
>  	struct multipath *m =
> -		container_of(work, struct multipath, activate_path);
> +		container_of(work, struct multipath, activate_path.work);
>  	struct dm_path *path;
>  	unsigned long flags;
>  
> diff --git a/include/scsi/scsi_dh.h b/include/scsi/scsi_dh.h
> index 33efce2..f099d86 100644
> --- a/include/scsi/scsi_dh.h
> +++ b/include/scsi/scsi_dh.h
> @@ -55,6 +55,10 @@ enum {
>  	SCSI_DH_NOSYS,
>  	SCSI_DH_DRIVER_MAX,
>  };
> +
> +/* Time to wait before retry in case of SCSI_DH_RETRY */
> +#define SCSI_DH_RETRY_DELAY ((HZ * 2))
> +
>  #if defined(CONFIG_SCSI_DH) || defined(CONFIG_SCSI_DH_MODULE)
>  extern int scsi_dh_activate(struct request_queue *);
>  extern int scsi_dh_handler_exist(const char *);
> &#0;