Re: 2.6.14-mm1

[James Bottomley <James.Bottomley@SteelEye.com>]

On Mon, 2005-11-07 at 10:52 -0800, Andrew Morton wrote:
> sd_issue_flush() has been altered to run scsi_disk_get_from_dev(), which
> takes a semaphore.  It does this from within spinlock and, as we see here,
> from within softirq.
> 
> Methinks the people who developed and tested that patch forgot to enable
> CONFIG_PREEMPT, CONFIG_DEBUG_KERNEL, CONFIG_DEBUG_SLAB,
> CONFIG_DEBUG_SPINLOCK and CONFIG_DEBUG_SPINLOCK_SLEEP.

Actually, I do too (as far as I can on non-x86).  I assume you also need
a filesystem that excites this, though.

Try the attached: We can probably rely on the block device having opened
the sd device, so there should already be a reference held on the
scsi_disk ... well that's my theory and I'm sticking to it.

James

diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c
--- a/drivers/scsi/sd.c
+++ b/drivers/scsi/sd.c
@@ -769,20 +769,16 @@ static void sd_end_flush(request_queue_t
 static int sd_prepare_flush(request_queue_t *q, struct request *rq)
 {
 	struct scsi_device *sdev = q->queuedata;
-	struct scsi_disk *sdkp = scsi_disk_get_from_dev(&sdev->sdev_gendev);
-	int ret = 0;
+	struct scsi_disk *sdkp = dev_get_drvdata(&sdev->sdev_gendev);
 
-	if (sdkp) {
-		if (sdkp->WCE) {
-			memset(rq->cmd, 0, sizeof(rq->cmd));
-			rq->flags |= REQ_BLOCK_PC | REQ_SOFTBARRIER;
-			rq->timeout = SD_TIMEOUT;
-			rq->cmd[0] = SYNCHRONIZE_CACHE;
-			ret = 1;
-		}
-		scsi_disk_put(sdkp);
-	}
-	return ret;
+	if (!sdkp || !sdkp->WCE)
+		return 0;
+
+	memset(rq->cmd, 0, sizeof(rq->cmd));
+	rq->flags |= REQ_BLOCK_PC | REQ_SOFTBARRIER;
+	rq->timeout = SD_TIMEOUT;
+	rq->cmd[0] = SYNCHRONIZE_CACHE;
+	return 1;
 }
 
 static void sd_rescan(struct device *dev)