In some drives, flush requests are non-queueable. When flush request is running,
normal read/write requests can't run. If block layer dispatches such request,
driver can't handle it and requeue it.
Tejun suggested we can hold the queue when flush is running. This can avoid
unnecessary requeue.
Also this can improve performance. Say we have requests f1, w1, f2 (f is flush
request, and w is write request). When f1 is running, queue will be hold, so w1
will not be added to queue list. Just after f1 is finished, f2 will be
dispatched. Since f1 already flushs cache out, f2 can be finished very quickly.
In my test, the queue holding completely solves a regression introduced by
commit 53d63e6b0dfb9588, which is about 20% regression running a sysbench fileio
workload.
Signed-off-by: Shaohua Li <[email protected]>
---
block/blk-flush.c | 3 +++
block/blk.h | 12 +++++++++++-
include/linux/blkdev.h | 1 +
3 files changed, 15 insertions(+), 1 deletion(-)
Index: linux/block/blk-flush.c
===================================================================
--- linux.orig/block/blk-flush.c 2011-05-04 14:20:33.000000000 +0800
+++ linux/block/blk-flush.c 2011-05-04 15:23:50.000000000 +0800
@@ -199,6 +199,9 @@ static void flush_end_io(struct request
BUG_ON(q->flush_pending_idx == q->flush_running_idx);
+ queued |= q->flush_queue_delayed;
+ q->flush_queue_delayed = 0;
+
/* account completion of the flush request */
q->flush_running_idx ^= 1;
elv_completed_request(q, flush_rq);
Index: linux/include/linux/blkdev.h
===================================================================
--- linux.orig/include/linux/blkdev.h 2011-05-04 14:24:40.000000000 +0800
+++ linux/include/linux/blkdev.h 2011-05-04 14:29:29.000000000 +0800
@@ -365,6 +365,7 @@ struct request_queue
*/
unsigned int flush_flags;
unsigned int flush_not_queueable:1;
+ unsigned int flush_queue_delayed:1;
unsigned int flush_pending_idx:1;
unsigned int flush_running_idx:1;
unsigned long flush_pending_since;
Index: linux/block/blk.h
===================================================================
--- linux.orig/block/blk.h 2011-05-04 14:20:33.000000000 +0800
+++ linux/block/blk.h 2011-05-04 16:09:42.000000000 +0800
@@ -61,7 +61,17 @@ static inline struct request *__elv_next
rq = list_entry_rq(q->queue_head.next);
return rq;
}
-
+ /*
+ * Flush request is running and flush request isn't queeueable
+ * in the drive, we can hold the queue till flush request is
+ * finished. Even we don't do this, driver can't dispatch next
+ * requests and will requeue them.
+ */
+ if (q->flush_pending_idx != q->flush_running_idx &&
+ !blk_queue_flush_queueable(q)) {
+ q->flush_queue_delayed = 1;
+ return NULL;
+ }
if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
return NULL;
}
Hello,
On Wed, May 04, 2011 at 04:17:27PM +0800, [email protected] wrote:
> In some drives, flush requests are non-queueable. When flush request is running,
> normal read/write requests can't run. If block layer dispatches such request,
> driver can't handle it and requeue it.
> Tejun suggested we can hold the queue when flush is running. This can avoid
> unnecessary requeue.
> Also this can improve performance. Say we have requests f1, w1, f2 (f is flush
> request, and w is write request). When f1 is running, queue will be hold, so w1
> will not be added to queue list. Just after f1 is finished, f2 will be
> dispatched. Since f1 already flushs cache out, f2 can be finished very quickly.
> In my test, the queue holding completely solves a regression introduced by
> commit 53d63e6b0dfb9588, which is about 20% regression running a sysbench fileio
> workload.
>
> Signed-off-by: Shaohua Li <[email protected]>
It looks good to me now, but some nitpicks.
> Index: linux/block/blk-flush.c
> ===================================================================
> --- linux.orig/block/blk-flush.c 2011-05-04 14:20:33.000000000 +0800
> +++ linux/block/blk-flush.c 2011-05-04 15:23:50.000000000 +0800
> @@ -199,6 +199,9 @@ static void flush_end_io(struct request
>
> BUG_ON(q->flush_pending_idx == q->flush_running_idx);
>
> + queued |= q->flush_queue_delayed;
> + q->flush_queue_delayed = 0;
> +
> /* account completion of the flush request */
> q->flush_running_idx ^= 1;
> elv_completed_request(q, flush_rq);
Can you please do if (queued || q->flush_queue_delayed) instead of
setting queued? And please also update the comment above the if
statement so that it explains the flush_queue_delayed case too.
> Index: linux/block/blk.h
> ===================================================================
> --- linux.orig/block/blk.h 2011-05-04 14:20:33.000000000 +0800
> +++ linux/block/blk.h 2011-05-04 16:09:42.000000000 +0800
> @@ -61,7 +61,17 @@ static inline struct request *__elv_next
> rq = list_entry_rq(q->queue_head.next);
> return rq;
> }
> -
> + /*
> + * Flush request is running and flush request isn't queeueable
> + * in the drive, we can hold the queue till flush request is
> + * finished. Even we don't do this, driver can't dispatch next
> + * requests and will requeue them.
> + */
Please explain the f1, w1, f2 case here as that's the biggest reason
this optimization is implemented and also explain the use of
flush_queue_delayed (just explain briefly and refer to
flush_end_io()).
Thank you.
--
tejun
Hello.
On 04-05-2011 12:17, [email protected] wrote:
> In some drives, flush requests are non-queueable. When flush request is running,
> normal read/write requests can't run. If block layer dispatches such request,
> driver can't handle it and requeue it.
> Tejun suggested we can hold the queue when flush is running. This can avoid
> unnecessary requeue.
> Also this can improve performance. Say we have requests f1, w1, f2 (f is flush
> request, and w is write request). When f1 is running, queue will be hold, so w1
> will not be added to queue list. Just after f1 is finished, f2 will be
> dispatched. Since f1 already flushs cache out, f2 can be finished very quickly.
> In my test, the queue holding completely solves a regression introduced by
> commit 53d63e6b0dfb9588, which is about 20% regression running a sysbench fileio
Please specify that commit's summary -- for human readers. The ID is only
immediately usable to gitweb.
> workload.
> Signed-off-by: Shaohua Li <[email protected]>
[...]
> Index: linux/block/blk.h
> ===================================================================
> --- linux.orig/block/blk.h 2011-05-04 14:20:33.000000000 +0800
> +++ linux/block/blk.h 2011-05-04 16:09:42.000000000 +0800
> @@ -61,7 +61,17 @@ static inline struct request *__elv_next
> rq = list_entry_rq(q->queue_head.next);
> return rq;
> }
> -
> + /*
> + * Flush request is running and flush request isn't queeueable
Queueable.
WBR, Sergei