Kill SCSI_BH

Kill SCSI_BH

[Matthew Wilcox]

I'm not exactly clueful about the linux SCSI layer, but I'd really like
to see old-style bottom halves die before 2.6.  Is there any reason why
SCSI_BH can't simply be comverted to a tasklet?

Andrew Morton kindly tested this patch for me and reports it works fine.
I'd expect the Enterprise people to be thrilled by this as it increases
scalability a touch (eg, SCSI handling is now not serialised against
the serial port).

--- linux-2.5.21/drivers/scsi/scsi.c	Sun Jun  2 18:44:50 2002
+++ linux-2.5.21-flock/drivers/scsi/scsi.c	Sat Jun 15 19:45:45 2002
@@ -1090,6 +1090,9 @@
 	SCSI_LOG_MLQUEUE(3, printk("Leaving scsi_do_cmd()\n"));
 }
 
+void scsi_tasklet_func(unsigned long);
+static DECLARE_TASKLET(scsi_tasklet, scsi_tasklet_func, 0);
+
 /*
  * This function is the mid-level interrupt routine, which decides how
  *  to handle error conditions.  Each invocation of this function must
@@ -1187,7 +1190,7 @@
 	/*
 	 * Mark the bottom half handler to be run.
 	 */
-	mark_bh(SCSI_BH);
+	tasklet_hi_schedule(&scsi_tasklet);
 }
 
 /*
@@ -1213,7 +1216,7 @@
  * half queue.  Thus the only time we hold the lock here is when
  * we wish to atomically remove the contents of the queue.
  */
-void scsi_bottom_half_handler(void)
+void scsi_tasklet_func(unsigned long ignore)
 {
 	Scsi_Cmnd *SCpnt;
 	Scsi_Cmnd *SCnext;
@@ -2545,11 +2548,6 @@
         if (scsihosts)
 		printk(KERN_INFO "scsi: host order: %s\n", scsihosts);	
 	scsi_host_no_init (scsihosts);
-	/*
-	 * This is where the processing takes place for most everything
-	 * when commands are completed.
-	 */
-	init_bh(SCSI_BH, scsi_bottom_half_handler);
 
 	return 0;
 }
@@ -2559,7 +2557,7 @@
 	Scsi_Host_Name *shn, *shn2 = NULL;
 	int i;
 
-	remove_bh(SCSI_BH);
+	tasklet_kill(&scsi_tasklet);
 
         devfs_unregister (scsi_devfs_handle);
         for (shn = scsi_host_no_list;shn;shn = shn->next) {

-- 
Revolutions do not require corporate support.

Re: Kill SCSI_BH

[Douglas Gilbert]

Matthew Wilcox wrote:
> 
> I'm not exactly clueful about the linux SCSI layer, but I'd really like
> to see old-style bottom halves die before 2.6.  Is there any reason why
> SCSI_BH can't simply be comverted to a tasklet?

Probably just another linux scsi subsystem anachronism.

> Andrew Morton kindly tested this patch for me and reports it works fine.
> I'd expect the Enterprise people to be thrilled by this as it increases
> scalability a touch (eg, SCSI handling is now not serialised against
> the serial port).

I applied the patch and it's running fine on this box.
See how it goes with some stress testing. So far so go...

Doug Gilbert

Re: Kill SCSI_BH

[James Bottomley]

willy@debian.org said:
> Andrew Morton kindly tested this patch for me and reports it works
> fine. I'd expect the Enterprise people to be thrilled by this as it
> increases scalability a touch (eg, SCSI handling is now not serialised
> against the serial port). 

It works fine for me on an 8-way box with SCSI root discs.  Go ahead and send 
it to Linus as an obvious fix (given that the bottom half handlers are 
supposed to be deprecated).  I've also stuck it in a public bitkeeper 
repository at:

http://linux-scsi.bkbits.net/scsi-mid-layer-2.5

James

Re: Kill SCSI_BH

[Oliver Xymoron]

On Sun, 16 Jun 2002, Matthew Wilcox wrote:

> I'm not exactly clueful about the linux SCSI layer, but I'd really like
> to see old-style bottom halves die before 2.6.  Is there any reason why
> SCSI_BH can't simply be comverted to a tasklet?

Very nice.

-- 
 "Love the dolphins," she advised him. "Write by W.A.S.T.E.."