2008-10-23 16:13:58

by Miquel van Smoorenburg

Subject: [PATCH] elevator.c: prevent flushing small requests to device

While tracing I/O patterns with blktrace (a great tool) a few
weeks ago I identified a minor issue in elevator.c

[PATCH] elevator.c: prevent flushing small requests to device

When elv_insert() is called to insert a new request, and the
device is plugged, bit of code at the end of the function
will unplug the device if the number of pending requests >= q->unplug_thresh.

This means the current request is also sent to the device immidiately
while it is potentially mergeable with the next request. This has been
observed when a lot of small sequential requests are made.

By unplugging the device before we add the new request to the
queue this can be prevented.

Signed-off-by: Miquel van Smoorenburg <[email protected]>

diff -ruN linux-2.6.26.5.orig/block/elevator.c linux-2.6.26.5/block/elevator.c
--- linux-2.6.26.5.orig/block/elevator.c 2008-09-08 19:40:20.000000000 +0200
+++ linux-2.6.26.5/block/elevator.c 2008-10-23 00:33:21.000000000 +0200
@@ -602,6 +602,20 @@

case ELEVATOR_INSERT_SORT:
BUG_ON(!blk_fs_request(rq));
+ /*
+ * If we're going to unplug the device, do it now before
+ * we put a potentially small and mergeable new
+ * request on the queue, instead of just after it.
+ */
+ if (blk_queue_plugged(q)) {
+ int nrq = q->rq.count[READ] + q->rq.count[WRITE]
+ - q->in_flight;
+ if (nrq >= q->unplug_thresh)
+ __generic_unplug_device(q);
+ if (elv_queue_empty(q))
+ blk_plug_device(q);
+ unplug_it = 0;
+ }
rq->cmd_flags |= REQ_SORTED;
q->nr_sorted++;
if (rq_mergeable(rq)) {

2008-10-31 07:35:04

by Jens Axboe

Subject: Re: [PATCH] elevator.c: prevent flushing small requests to device

On Thu, Oct 23 2008, Miquel van Smoorenburg wrote:
> While tracing I/O patterns with blktrace (a great tool) a few
> weeks ago I identified a minor issue in elevator.c
>
> [PATCH] elevator.c: prevent flushing small requests to device
>
> When elv_insert() is called to insert a new request, and the
> device is plugged, bit of code at the end of the function
> will unplug the device if the number of pending requests >= q->unplug_thresh.
>
> This means the current request is also sent to the device immidiately
> while it is potentially mergeable with the next request. This has been
> observed when a lot of small sequential requests are made.
>
> By unplugging the device before we add the new request to the
> queue this can be prevented.
>
> Signed-off-by: Miquel van Smoorenburg <[email protected]>
>
> diff -ruN linux-2.6.26.5.orig/block/elevator.c linux-2.6.26.5/block/elevator.c
> --- linux-2.6.26.5.orig/block/elevator.c 2008-09-08 19:40:20.000000000 +0200
> +++ linux-2.6.26.5/block/elevator.c 2008-10-23 00:33:21.000000000 +0200
> @@ -602,6 +602,20 @@
>
> case ELEVATOR_INSERT_SORT:
> BUG_ON(!blk_fs_request(rq));
> + /*
> + * If we're going to unplug the device, do it now before
> + * we put a potentially small and mergeable new
> + * request on the queue, instead of just after it.
> + */
> + if (blk_queue_plugged(q)) {
> + int nrq = q->rq.count[READ] + q->rq.count[WRITE]
> + - q->in_flight;
> + if (nrq >= q->unplug_thresh)
> + __generic_unplug_device(q);
> + if (elv_queue_empty(q))
> + blk_plug_device(q);
> + unplug_it = 0;
> + }
> rq->cmd_flags |= REQ_SORTED;
> q->nr_sorted++;
> if (rq_mergeable(rq)) {

This makes sense, I'll throw it into the 2.6.29 mix for some testing.
I've slightly modified your patch, there's no need to check for queue
empty and plug on your own, __make_request() will do that.

--
Jens Axboe