Re: RFC on io-stalls patch

[Chris Mason <mason@suse.com>]

On Mon, 2003-07-14 at 15:51, Jens Axboe wrote:

> Some initial results with the attached patch, I'll try and do some more
> fine grained tomorrow. Base kernel was 2.4.22-pre5 (obviously), drive
> tested is a SCSI drive (on aic7xxx, tcq fixed at 4), fs is ext3. I would
> have done ide testing actually, but the drive in that machine appears to
> have gone dead. I'll pop in a new one tomorrow and test on that too.
> 

Thanks Jens, the results so far are very interesting (although I'm
curious to hear how 2.4.21 did).

> --- 1.47/drivers/block/ll_rw_blk.c	Fri Jul 11 10:30:54 2003
> +++ edited/drivers/block/ll_rw_blk.c	Mon Jul 14 20:42:36 2003
> @@ -549,10 +549,12 @@
>  static struct request *get_request(request_queue_t *q, int rw)
>  {
>  	struct request *rq = NULL;
> -	struct request_list *rl;
> +	struct request_list *rl = &q->rq;
>  
> -	rl = &q->rq;
> -	if (!list_empty(&rl->free) && !blk_oversized_queue(q)) {
> +	if ((rw == WRITE) && (blk_oversized_queue(q) || (rl->count < 4)))
> +		return NULL;
> +
> +	if (!list_empty(&rl->free)) {
>  		rq = blkdev_free_rq(&rl->free);
>  		list_del(&rq->queue);
>  		rl->count--;
> @@ -947,7 +949,7 @@

Could I talk you into trying a form of this patch that honors
blk_oversized_queue for everything except the BH_sync requests?

-chris