Re: [Bug 11898] mke2fs hang on AIC79 device.

[Mike Christie <michaelc@cs.wisc.edu>]

[-- Attachment #1: Type: text/plain, Size: 1732 bytes --]

James Bottomley wrote:
> The reason for doing it like this is so that if someone slices the loop
> apart again (which is how this crept in) they won't get a continue or
> something which allows this to happen.
> 
> It shouldn't be conditional on the starved list (or anything else)
> because it's probably a register and should happen at the same point as
> the list deletion but before we drop the problem lock (because once we
> drop that lock we'll need to recompute starvation).
> 
> James
> 
> ---
> 
> diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
> index f5d3b96..f9a531f 100644
> --- a/drivers/scsi/scsi_lib.c
> +++ b/drivers/scsi/scsi_lib.c
> @@ -606,6 +606,7 @@ static void scsi_run_queue(struct request_queue *q)
>  		}
>  
>  		list_del_init(&sdev->starved_entry);
> +		starved_entry = NULL;

Should this be starved_head?

>  		spin_unlock(shost->host_lock);
>  
>  		spin_lock(sdev->request_queue->queue_lock);
> 

Do you think we can just splice the list like the attached patch (patch 
is example only and is not tested)?

I thought the code is clearer, but I think it may be less efficient. If 
scsi_run_queue is run on multiple processors then with the attached 
patch one processor would splice the list and possibly have to execute 
__blk_run_queue for all the devices on the list serially.

Currently we can at least prep the devices in parallel. One processor 
would grab one entry on the list and drop the host lock, so then another 
processor could grab another entry on the list and start the execution 
process (I wrote start the process because it might turn out that this 
second entry execution might have to wait on the first one when the scsi 
layer has to grab the queue lock again).

[-- Attachment #2: splice.patch --]
[-- Type: text/x-patch, Size: 1666 bytes --]

diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
index f5d3b96..21a436b 100644
--- a/drivers/scsi/scsi_lib.c
+++ b/drivers/scsi/scsi_lib.c
@@ -567,15 +567,23 @@ static inline int scsi_host_is_busy(struct Scsi_Host *shost)
  */
 static void scsi_run_queue(struct request_queue *q)
 {
-	struct scsi_device *starved_head = NULL, *sdev = q->queuedata;
+	struct scsi_device *sdev = q->queuedata;
 	struct Scsi_Host *shost = sdev->host;
+	LIST_HEAD(starved_list);
 	unsigned long flags;
 
 	if (scsi_target(sdev)->single_lun)
 		scsi_single_lun_run(sdev);
 
 	spin_lock_irqsave(shost->host_lock, flags);
-	while (!list_empty(&shost->starved_list) && !scsi_host_is_busy(shost)) {
+
+	/*
+	 * splice the list in case the target busy check or the
+	 * request_fn's busy checks want to readd the sdev onto
+	 * the starved list.
+	 */
+	list_splice_init(&shost->starved_list, &starved_list);
+	while (!list_empty(&starved_list) && !scsi_host_is_busy(shost)) {
 		int flagset;
 
 		/*
@@ -588,17 +596,8 @@ static void scsi_run_queue(struct request_queue *q)
 		 * scsi_request_fn must get the host_lock before checking
 		 * or modifying starved_list or starved_entry.
 		 */
-		sdev = list_entry(shost->starved_list.next,
+		sdev = list_entry(starved_list.next,
 					  struct scsi_device, starved_entry);
-		/*
-		 * The *queue_ready functions can add a device back onto the
-		 * starved list's tail, so we must check for a infinite loop.
-		 */
-		if (sdev == starved_head)
-			break;
-		if (!starved_head)
-			starved_head = sdev;
-
 		if (scsi_target_is_busy(scsi_target(sdev))) {
 			list_move_tail(&sdev->starved_entry,
 				       &shost->starved_list);